text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'The task of maneuvering a multi-steered articulated vehicle in confined environments is difficult even for experienced drivers. In this work, we present an optimization-based trajectory planner targeting low-speed maneuvers in unstructured environments for multi-steered N-trailer vehicles, which are comprised of a car-like tractor and an arbitrary number of interconnected trailers with fixed or steerable wheels. The proposed trajectory planning framework is divided into two steps, where a lattice-based trajectory planner is used in a first step to compute a resolution optimal solution to a discretized version of the trajectory planning problem. The output from the lattice planner is then used in a second step to initialize an optimal control problem solver, which enables the framework to compute locally optimal trajectories that start at the vehicle’s initial state and reaches the goal state exactly. The performance of the proposed optimization-based trajectory planner is evaluated in a set of practically relevant scenarios for a multi-steered 3-trailer vehicle with a car-like tractor where the last trailer is steerable.'
author:
- |
Oskar Ljungqvist, Kristoffer Bergman and Daniel Axehill\
\
\
bibliography:
- 'root.bib'
title: 'Optimization-based motion planning for multi-steered articulated vehicles'
---
Introduction
============
In recent years, there has been a growing demand within the transportation sector to increase efficiency and reduce environmental impact related to transportation of both people and goods. This has lead to an increased interest for large capacity (multi-) articulated buses [@michalek2019modular] and long tractor-trailer vehicle combinations [@islam2015comparative]. In order to improve these long vehicles ability to maneuver in confined environments, some trailers (or wagons) can be equipped with steerable wheels. In the literature, these vehicles are commonly referred to as multi-steered N-trailer (MSNT) vehicles [@orosco2002modeling], which is a generalization of single-steered N-trailer (SSNT) vehicles. Compared to a SSNT vehicle, the additional steering capability enables an MSNT vehicle to be more flexible and agile, on the expense of increased difficulty in manually maneuvering the vehicle by a human driver. These difficulties partially arise due to the vehicle’s increased degrees of freedom which are hard to successfully cope with for a human operator, and because of the specific kinematic and dynamic properties of an MSNT vehicle (see, e.g., [@tilbury1995multisteering; @michalek2019modular; @orosco2002modeling; @islam2015comparative]). To aid the driver, several advanced driver-assistance system concepts have been proposed to automatically steer the extra steerable wheel(s) in order to increase low-speed maneuverability or to reduce the so called off-tracking during tight cornering [@odhams2011active; @van2015active; @michalek2019modular; @varga2018robust].
Although a large amount of different motion planning techniques has been proposed in the literature for SSNT vehicles (see, e.g., [@sekhavat1997multi; @hillary; @evestedtLjungqvist2016planning; @li2019trajectory; @LjungqvistJFR2019]), there only exists a limited amount of work that consider the trajectory planning problem for special classes of MSNT vehicles (see e.g. [@bushnell1995steering; @tilbury1995multisteering; @vidal2002real; @Beyersdorfer2013tractortrailer; @Yuan2017]). As a consequence, there is still a need to develop a trajectory planner that is able to solve the trajectory planning problem for a generic MSNT vehicle with car-like tractor that: i) can handle various state and input constraints, ii) allows a mixture of on-axle/off-axle hitched and steerable/non-steerable trailers, and iii) computes locally optimal trajectories by combining forward and backward motion.
The contribution of this work is a trajectory planning framework for an MSNT vehicle with car-like tractor targeting low-speed maneuvers in confined and unstructured environments. The framework extends some techniques presented in our previous work in [@LjungqvistJFR2019] and is inspired by [@bergman2019improved; @bergman2019bimproved; @bergman2018combining]. Here, a lattice-based trajectory planner is developed and used in a first step to compute a resolution optimal solution to a discretized version of the trajectory planning problem. The lattice planner uses a finite library of precomputed optimized maneuvers restricted to move the MSNT vehicle within a specified state-space discretization. In a second step, the output from the lattice planner is used to initialize a homotopy-based optimization step enabling the framework to compute locally optimal trajectories that starts at the vehicle’s initial state and reaches desired goal states exactly. To the best of the authors knowledge, this paper presents the first trajectory planning framework for a generic MSNT vehicle with car-like tractor.
The reminder of the paper is organized as follows. In Section \[p8:sec:model\], the kinematic vehicle model and the trajectory planning problem for the considered MSNT vehicle are presented, as well as an overview of the proposed trajectory planning framework. In Section \[p8:sec:lattice\_planner\] and Section \[p8:sec:homo\_optimization\], the lattice-based trajectory planner and the homotopy-based optimization step are presented, respectively. Simulation results for an MS3T vehicle with car-like tractor are presented in Section \[p8:sec:results\] and the paper is concluded in Section \[p8:sec:conclusions\] by summarizing the contributions and discussing the directions for future work.
Kinematic vehicle model and problem formulation {#p8:sec:model}
===============================================
The MSNT vehicle with car-like tractor considered in this work is illustrated in Figure \[c8:fig:vehicle\_model\]. The multi-body vehicle is composed of $N+1$ vehicle segments, including a car-like tractor and $N$ number of trailers that are equipped with steerable or non-steerable wheels. Each vehicle segment is characterized by a segment length $L_i>0$ and a signed hitching offset $M_i$. Since low-speed maneuvers are considered in this work, a kinematic vehicle model is used. The model is based on the work in [@michalek2019modular] and is derived based on various assumptions such as the wheels are rolling without slipping and that the vehicle is operating on a flat surface. Moreover, it is assumed that the front wheel of the tractor is steerable and its rear wheel is non-steerable. The vehicle configuration consists of variables [@michalek2019modular] where $S\in\{1,\hdots,N\}$ denotes the number of steerable trailers:
- the steering angle of the tractor’s front wheel $$\begin{aligned}
\beta_0 \in \pazocal Q_0=[-\bar\beta_0,\bar\beta_0], \quad \bar \beta_0 \in (0,\pi/2),
\label{c8:eq:front_steering_angle}
\end{aligned}$$
- the global position $(x_{N},y_{N})$ and orientation $\theta_{N}$ of the $N$th trailer in a fixed coordinate frame $$\begin{aligned}
q_{N} = \begin{bmatrix}
\theta_{N} & x_{N} & y_{N}
\end{bmatrix}^T \in \mathbb S \times \mathbb R^2, \quad \mathbb S=(-\pi,\pi],
\label{c8:eq:poes_trailer}
\end{aligned}$$
- for $i=1,\hdots, N,$ a number of $N$ constrained joint angles $$\begin{aligned}
\beta_i = \theta_{i-1} -\theta_{i} \in\pazocal B_i=[-\bar\beta_i,\bar\beta_i], \quad \bar \beta_i \in (0,\pi),
\label{c8:eq:joint_angles}
\end{aligned}$$
- and $S\in\{1,\hdots,N\}$ number of steering angles associated with steerable trailer wheels $$\begin{aligned}
\gamma_{s} \in \pazocal Q_{s}=[-\bar\gamma_{s},\bar\gamma_{s}], \quad \bar \gamma_{s} \in (0,\pi/2),
\label{c8:eq:trailer_steering_angles}
\end{aligned}$$ where index $s\in \pazocal I_s \subseteq \{1,\hdots,N\}$ specifies which trailers that have steerable wheels. The configuration vector for the MSNT with car-like tractor will be defined as $$\begin{aligned}
q = \begin{bmatrix}
\beta_0 & \beta_1 & \hdots & \beta_{N} & \bm \gamma_s^T & q_N^T
\end{bmatrix}^T \in \pazocal Q,
\label{c8:eq:configuration_vector}
\end{aligned}$$ where $\bm \gamma_s$ represents a vector of trailer steering angles, and $\pazocal Q =\pazocal Q_0\times \pazocal B_1 \hdots \times\pazocal B_N\times \underbrace{\pazocal Q_s\times \hdots\times \pazocal Q_s}_{S\text{-times}}\times \mathbb S\times \mathbb R^2$.
The leading car-like tractor is described by a kinematic single-track vehicle model and its orientation $\theta_0$ evolves as $$\begin{aligned}
\dot \theta_0 = v_0\kappa_0(\beta_0),
\label{c8:eq:model_orientation_tractor}\end{aligned}$$ where $\kappa_0(\beta_0) =\frac{\tan\beta_0}{L_0}$ is the curvature of the tractor and $v_0$ is the longitudinal velocity of its rear axle. The recursive formula for the transformation of the angular $\dot\theta_i$ and longitudinal $v_i$ velocities between any two neighboring vehicle segments are given by [@michalek2019modular; @orosco2002modeling]: $$\begin{aligned}
\begin{bmatrix}
\dot \theta_{i} \\ v_{i}
\end{bmatrix} &=
\underbrace{\begin{bmatrix}
-\frac{M_{i}}{L_{i}}\frac{\cos{(\beta_{i}-\gamma_{i})}}{\cos\gamma_{i}} & \frac{\sin{(\beta_{i}-\gamma_{i}+\gamma_{i-1})}}{L_{i}\cos\gamma_{i}} \\
M_{i}\frac{\sin{\beta_{i}}}{\cos\gamma_{i}} & \frac{\cos{(\beta_{i}+\gamma_{i-1})}}{\cos\gamma_{i}}
\end{bmatrix}}_{\triangleq J_i(\beta_{i},\gamma_i,\gamma_{i-1})}
\begin{bmatrix}
\dot \theta_{i-1} \\ v_{i-1}
\end{bmatrix},
\label{c8:eq:velocity_transformation}\end{aligned}$$ where $\gamma_i$ denotes the steering angle of the $i$th trailer. For the car-like tractor, we have that $\gamma_0\equiv 0$ since its rear axle is non-steerable. Note that if the $j$th trailer is non-steerable, it suffices to take $\gamma_j=0$ in .
![A schematic description of the geometric lengths and the vehicle configuration for a MSNT vehicle with car-like tractor in a global coordinate system (inspired and adapted from [@michalek2019modular]).[]{data-label="c8:fig:vehicle_model"}](Vehicle_model.pdf){width="0.7\linewidth"}
Each trailer steering angle $\gamma_s$, $s\in\pazocal I_s$ and the tractors steering angle $\beta_0$ are modeled as double integrator systems $$\begin{aligned}
\begin{split}
\dot \gamma_s &= \omega_s, \quad
\dot \omega_s= u_{\omega_s}, \quad s\in\pazocal I_s, \\
\dot \beta_0 &= \omega_0, \quad
\dot \omega_0= u_{\omega_0},
\label{c8:eq:augumented_states}
\end{split}\end{aligned}$$ where $\omega_0$, $\omega_s$, $s\in\pazocal I_s$ and $u_{\omega_0}$, $u_{\omega_s}$, $s\in\pazocal I_s$ denote steering angle rates and accelerations, respectively. This modeling is used to be able to penalize large rates and accelerations, and to enforce constraints in the form $$\begin{aligned}
\begin{split}
\omega_s &\in \Omega_s =[-\bar \omega_s,\bar \omega_s], \quad u_{\omega_s} \in \pazocal U_s =[-\bar\Omega_s,\bar\Omega_s], \quad s\in\pazocal I_s, \\
\omega_0 &\in \Omega_0 =[-\bar \omega_0,\bar \omega_0], \quad u_{\omega_0} \in \pazocal U_0 =[-\bar\Omega_0,\bar\Omega_0],
\end{split}
\label{c8:eq:rate_acc_constraints_on_steering}\end{aligned}$$ where the steering angle accelerations $u_{\omega_0}$ and are treated as control signals. Similarly, the longitudinal velocity of the tractor $v_0$ is constrained as and its dynamics is modeled as a double integrator system $$\begin{aligned}
\label{c8:eq:long_dynamics}
\dot v_0 = a_0, \quad \dot a_0 = u_v\end{aligned}$$ in order to respect constraints on the longitudinal acceleration $a_0\in \pazocal A = [-\bar a,\bar a]$ and jerk $u_v\in \pazocal U_v = [-\bar u_{v},\bar u_{v}]$. During the planning phase, the longitudinal jerk $u_{v}$ is treated as a control signal.
Since rolling without slipping of the wheels is assumed, the position of the last trailer evolves according to standard unicycle kinematics $$\begin{aligned}
\begin{split}
\dot x_{N} &= v_{N}\cos(\theta_{N}+\gamma_N), \\
\dot y_{N} &= v_{N}\cos(\theta_{N}+\gamma_N),
\end{split}
\label{c8:eq:position_Ntrailer}\end{aligned}$$ where its angular rate $\dot \theta_{N}$ and longitudinal velocity $v_{N}$ are given by $$\begin{aligned}
\label{c8:eq:rate_speed_Ntrailer}
\begin{bmatrix}
\dot \theta_{N}\\
v_{N}
\end{bmatrix}=\prod_{i=N}^{1} J_i(\beta_{i},\gamma_i,\gamma_{i-1})\begin{bmatrix} v_0\kappa_0(\beta_0)\\
v_0
\end{bmatrix},\end{aligned}$$ which is derived by recursive usage of $N$ times together with . Combining and , it is possible to compactly represent the model for the pose of the $N$th trailer $q_N =[
\theta_{N} \hspace{5pt} x_{N} \hspace{5pt} y_{N}]^T$ as $\dot{ q}_N = v_0f_{ q_N}( q)$. In analogy, by introducing the vector $ c=[1\hspace{5pt} 0]^T$ together with and , the time derivative of yields the joint-angle kinematics $$\begin{aligned}
\dot\beta_i = \dot\theta_{i-1} - \dot\theta_i &= c^T \prod_{j=i-1}^{1} J_j(\beta_{j},\gamma_j,\gamma_{j-1})\begin{bmatrix}
v_0\kappa_0(\beta_0)\\
v_0
\end{bmatrix} \nonumber \\
- c^T\prod_{j=i}^{1} J_j(&\beta_{j},\gamma_j,\gamma_{j-1})\begin{bmatrix} v_0\kappa_0(\beta_0)\\
v_0
\end{bmatrix} \triangleq v_0f_{\beta_i}( q), \quad i=1,\hdots,N,
\label{c8:eq:joint_angle_kinematics}\end{aligned}$$ where $\prod_{j=0}^{1}(\cdotp)\triangleq I$. Introduce the state vector as $ x = [
q^T \hspace{5pt} \omega_0 \hspace{5pt} \bm\omega_s^T \hspace{5pt} v_0 \hspace{5pt} a_0
]^T\in\pazocal X$ and denote the control signal vector as $ u = [ u_{\omega_0} \hspace{5pt} \bm u_{\omega_s}^T \hspace{5pt} u_v
]^T\in\pazocal U$, where $\bm \omega_s$ and $\bm u_{\omega_s}$ represent vectors of trailer steering angle rates and accelerations, respectively, and $$\begin{aligned}
\begin{split}
\pazocal X &= \pazocal Q\times\Omega_0\times\underbrace{\Omega_s\times\hdots\times\Omega_s}_{S\text{-times}}\times \Omega_v\times \pazocal A, \\
\pazocal U &= \pazocal U_0\times \underbrace{\pazocal U_s\times\hdots\times\pazocal U_s}_{S\text{-times}}\times \pazocal U_v,
\end{split}
\label{c8:eq:pysicalConstraints}\end{aligned}$$ where $\text{dim}(\pazocal X) = 7+N+2S\triangleq n$ and . The constraints in will be referred to the vehicle’s physical constraints arising from, e.g., actuator, mechanical or sensing limitations. Finally, the kinematic model of the MSNT vehicle with car-like tractor is given in , and –, which can compactly be represented as $$\begin{aligned}
\dot{x}=f( x, u), \label{c8:vehicle_model}\end{aligned}$$ where $ f:\mathbb R^n\times \mathbb R^m\rightarrow \mathbb R^n$ is continuous and continuously differentiable with respect to $ x\in\pazocal X$ and $ u\in\pazocal U$.
Problem formulation
-------------------
The MSNT vehicle with car-like tractor is assumed to operate in a closed environment with only static obstacles $\pazocal X_{\text{obs}}$. The free-space where the vehicle is not in collision with any obstacle is defined as $\pazocal X_{\text{free}} = \pazocal X \setminus \pazocal X_{\text{obs}}$. Here, it is assumed that the obstacle-occupied region $\pazocal X_{\text{obs}}$ (hence also $\pazocal X_{\text{free}}$) can be described analytically, e.g., using different bounding regions [@lavalle2006planning]. Since the free-space $\pazocal X_{\text{free}}$ is defined as the complement set of $\pazocal X_{\text{obs}}$, it is in general a non-convex set.
The trajectory planning problem considered in this work is defined as follows: Compute a feasible and collision-free state and control signal trajectory $( x(t), u(t))$, $t\in[0,t_G]$ that moves the vehicle from its initial state $ x_I\in\pazocal X_{\text{free}}$ to a desired goal state $x_G\in\pazocal X_{\text{free}}$, while minimizing the cost functional $J$. This problem can be posed as a continuous-time optimal control problem (OCP) in the following general form $$\begin{aligned}
\begin{split}
\operatorname*{minimize}_{u(\cdotp), \hspace{0.5ex}t_G }\hspace{3.7ex}
& J = t_G + \int_{0}^{t_G}l( x(t), u(t))\,\text{d}t \label{c8:eq:MotionPlanningOCP}\\
\operatorname*{subject\:to}\hspace{3ex}
& \dot{ x}(t) = f( x(t), u(t)), \\
& x(0) = x_I, \quad x(t_G) = x_G, \\
& x(t) \in \pazocal X_{\text{free}}, \quad
u(t) \in \pazocal U,
\end{split}\end{aligned}$$ where the final time $t_G$ and $ u$ are optimization variables, and $l:\mathbb R^n\times \mathbb R^m\rightarrow \mathbb R_+$ is the cost function. In this work, the cost function is defined as $$\begin{aligned}
l( x, u)=\lvert\lvert x \rvert\rvert_{ Q}^2 + \lvert\lvert u \rvert\rvert_{ R}^2,\label{c8:eq:costFunction}\end{aligned}$$ where the weight matrices $Q\succeq 0$ and $R \succeq 0$. It is well-known that the OCP in is in general hard to solve by directly invoking a state-of-the-art OCP solver [@casadi; @ipopt]. This is mainly because the vehicle model is nonlinear and the free-space $\pazocal X_{\text{free}}$ is in many applications non-convex. Hence, a proper initialization strategy for any OCP solver is often a necessity in order to converge to a good locally optimal (or even feasible) solution [@bergman2019bimproved; @zhang2018optimization].
Trajectory planning framework {#p8:sec:framework}
-----------------------------
To efficiently and reliably solve the trajectory planning problem for an MSNT vehicle with car-like tractor, we propose a framework that combines a lattice-based trajectory planner and an online optimization step. The framework is based on and extends the previous works in [@LjungqvistJFR2019; @bergman2019improved; @bergman2019bimproved; @bergman2018combining]. The extensions are made to account for the specific properties of an MSNT vehicle with car-like tractor. The general idea is that a lattice-based trajectory planner is used in a first step to compute an optimal solution to a discrete version of the trajectory planning problem using a discretized state-space and a library of precomputed trajectories. The lattice planner is responsible for solving the combinatorial aspects of the trajectory planning problem, e.g., taking left or right around obstacles, and thus provides the latter optimization step with a proper initial guess. While keeping the combinatorial parts fixed, the objective of the optimization step is to further improve the initial guess computed by the lattice planner such that the resulting trajectory is a locally optimal solution to the original trajectory planning problem . However, since the lattice planner uses a discretized state space, in general its computed trajectory does not satisfies the initial and goal state constraints in . Therefore, the optimization step is also responsible for modifying the initial guess computed by the lattice planner such that the final optimized trajectory starts at the vehicle’s initial state and reaches the goal state exactly. To handle this in a structured and numerically stable way, a homotopy-based optimization strategy is proposed that is inspired by the work in [@bergman2018combining].
The main steps of the trajectory planning framework is summarized in Workflow \[c8:alg1\], where the lattice planner and the optimization step (Step 2) are explained in detail in Section \[p8:sec:lattice\_planner\] and Section \[p8:sec:homo\_optimization\], respectively.
**Step 0 (offline) – State lattice construction:**
1. ***State-space discretization:*** Specify the resolution of the discretized state-space $\pazocal X_d$
2. ***Motion primitive generation:*** Compute the set of motion primitives $\pazocal P$ by specifying a set of desired maneuvers and solve using an OCP solver
3. ***Heuristic function:*** Precompute a HLUT by calculating the optimal cost-to-go on a grid in an obstacle-free environment
**Step 1 – Online planning:**
1. ***Initialization:*** Project the vehicle’s initial state $ x_I$ and desired goal state $ x_G$ to $\pazocal X_d$
2. ***Graph search:*** Solve the discrete graph search problem using a graph search algorithm
**Step 2 – Homotopy-based optimization step:**
1. ***Initialization:*** Initialize the homotopy-based OCP solver with the solution computed by the lattice planner
2. ***Optimization:*** Solve the relaxed trajectory planning problem using an OCP solver
3. ***Return:*** Send the computed solution to a trajectory-tracking controller or report failure
Lattice-based trajectory planner {#p8:sec:lattice_planner}
================================
The idea with lattice-based trajectory planning is to restrict the solutions of the trajectory planning problem to a lattice graph $\pazocal G = \langle \pazocal V, \pazocal E \rangle$, which is a graph embedded in an Euclidean space that forms a regular and repeated pattern [@pivtoraiko2009differentially]. The lattice graph is constructed offline by discretizing the vehicle’s state space $\pazocal X_d\subset\pazocal X$ and precompute a set of motion primitives $\pazocal P$. Each vertex $v[k] \in \pazocal V$ is a vehicle state $ x[k]\in\pazocal X_d$ and each edge $ e_i\in \pazocal E$ represents of a motion primitive $ m_i\in\pazocal P$. A motion primitive is a feasible trajectory that moves the vehicle from an initial state $x[k]\in\pazocal X_d$ to a final state $x[k+1]\in\pazocal X_d$ while satisfying $x^i(\cdotp)\in\pazocal X$ and $u^i(\cdotp)\in\pazocal U$. A motion primitive is in this way designed to connect two vertices in the graph and the kinematic constraints and the physical constraints are satisfied offline. The cardinality of the motion primitive set is $|\pazocal P|=M$ and the motion primitives that can be used from $x[k]$ is denoted $\pazocal P(x[k])\subseteq \pazocal{P}$. Moreover, since the MSNT vehicle is position invariant, the motion primitive set $\pazocal P$ can be computed from the position of the $N$th trailer at the origin. Each motion primitive $m_i$ can then be translated and reused for all other positions on the grid.
Let $x[k+1] = f_{p}(x[k],m_{i})$ denote the successor state when motion primitive $m_{i}\in\pazocal P$ is applied from $x[k]$ and denote $J_p(m_{i})$ as the stage-cost associated to this transition, which is given by $$\begin{aligned}
J_p(m_i) = t^i_f + \int_{0}^{t^i_f}l(x^i(t),u^i(t))\text{d}t,\end{aligned}$$ where $l(x^i,u^i)$ is defined in . The resulting trajectory taken by the vehicle when motion primitive is applied from $x[k]$, is collision-free if it does not collide with any obstacle $c(m_i,x[k])\in\pazocal X_{\text{free}}$. Define as a discrete and integer-valued decision variable that is selected by the lattice planner, where $u_p[k]$ specifies which motion primitive that is applied at stage $k$. Now, the continuous-time trajectory planning problem is approximated by the following discrete-time OCP [@LjungqvistJFR2019]: $$\begin{aligned}
\begin{split}
\operatorname*{minimize}_{\{u_p[k]\}^{N_p-1}_{k=0}, \hspace{0.5ex} N_p}\hspace{3.7ex}
& J_{D} = \sum_{k=0}^{N_p-1}J_p( m_{u_p[k]}) \label{c8:eq:OCP_discrete} \\
\operatorname*{subject\:to}\hspace{3ex}
& x[k+1] = f_{p}( x[k], m_{u_p[k]}), \\
& x[0] = \bar{x}_I, \quad x[N_p] = \bar{ x}_G, \\
& m_{u_p[k]} \in \pazocal P( x[k]), \\
& c( m_{u_p[k]}, x[k]) \in \pazocal X_{\text{free}}, \quad k=0,\hdots,N_p-1,
\end{split}\end{aligned}$$ where $\bar{x}_I$ and $\bar{x}_G$ are obtained by projecting the actual initial state $ x_I$ and desired goal state $ x_G$ to the their closest neighboring state in $\pazocal X_d$. The decision variables to the problem in are the motion primitive sequence $\{u_p[k]\}^{N_p-1}_{k=0}$ and its length $N_p\in\mathbb Z_+$. A feasible solution is an ordered sequence of collision-free motion primitives $\{m_{u_p[k]}\}^{N_p-1}_{k=0}$, i.e., a trajectory $(x(t),u(t))$, $t\in [0, t_G]$ that connects the projected initial state $x(0)=\bar{x}_I$ with the projected goal state $ x(t_G)= \bar{x}_G$. Given the set of all feasible solutions to , an optimal solution is one that minimizes the cost function $J_{\text{D}}$. During online planning, the discrete-time OCP in can be solved using classical graph-search algorithms such as A$^*$ together with an informative precomputed free-space heuristic look-up table (HLUT) as heuristic function [@knepper2006high]. A HLUT significantly reduces the online planning time, as it takes the vehicle’s nonholonomic constraints into account and enables perfect estimation of cost-to-go in free-space scenarios with no obstacles.
State-space discretization
--------------------------
It is important that the resolution of the discretized state space $\pazocal X_d$ and the cardinality of the motion primitive set $\pazocal P$ are chosen such that the vehicle is sufficiently agile to maneuver in confined environments. However, as they also define the size of the lattice graph $\pazocal G$, both the resolution of $\pazocal X_d$ and the cardinality of $\pazocal P$ have to be chosen carefully in order to maintain a reasonable search time during online planning [@pivtoraiko2009differentially]. Motivated by this, the position of the $N$th trailer $(x_N[k],y_N[k])$ is discretized to a uniform grid with resolution $r$ and its orientation is irregularly[^1] discretized $\theta_N[k]\in\Theta$ into different orientations as proposed in [@pivtoraiko2009differentially]. Additionally, the longitudinal velocity of the tractor is discretized as $v_0[k]\in V=\{-\bar v,0,\bar v\}$, where $\bar v$ is the tractor’s maximal allowed speed. All remaining vehicle states are constrained to zero at each vertex in the graph, which implies that the MSNT vehicle is arranged in a straight configuration. This means that the joint angles $\beta_i[k]$, $i=1,\hdots, N$, the steering angles $\beta_0[k]$, $\bm\gamma_s[k]$, the steering angle rates $\omega_0[k]$, $\bm\omega_s[k]$ as well as the longitudinal acceleration of the tractor $a_0[k]$ are all constrained to zero at each $x[k]\in\pazocal X_d$. As a consequence, $\text{dim}(\pazocal X_d)=4$, since only the pose $ p_N[k]$ and the velocity of the tractor $v_0[k]$ are allowed to vary between different vertices in the graph. The proposed discretization will impose restrictions, but is motivated to enable fast online planning. Moreover, since the output from the lattice planner will be used to warm start a second optimization step, it will improve the initial guess computed by the lattice planner such that the finally computed trajectory is a locally optimal solution to the original trajectory planning problem .
Motion primitive generation
---------------------------
The set of motion primitives $\pazocal P$ is computed offline by solving a finite set of OCPs from all initial states $x_s^i\in\pazocal X_d$ with the position of the $N$th trailer at the origin, to a set of final states $x_f^i\in\pazocal X_d$ in a bounded neighborhood in an obstacle-free environment. This procedure can be performed manually as in [@LjungqvistJFR2019] or using exhaustive search together with pruning strategies as proposed in [@pivtoraiko2009differentially; @CirilloIROS2014]. In both cases, the motion primitive generation procedure will become time consuming or requires a designer with high system knowledge. Therefore, here we use the maneuver-based motion primitive generation framework introduced in [@bergman2019improved]. Instead of selecting pairs of discrete vehicle state to connect, a set of desired maneuvers from each initial state $x_s^i\in\pazocal X_d$ is selected and an OCP solver together with a rounding heuristic are used to automatically select the optimal final state $x_f^i\in\pazocal X_d$. Each maneuver is defined with a terminal manifold in the form $g^i(x^i(t_f^i)) = 0$ where $g:\mathbb R^n\rightarrow \mathbb R^l$ and $l<n$, where $n-l$ is the degrees of freedom for the terminal state constraint. To compute a maneuver-specified motion primitive $m_i\in\pazocal P$, the following continuous-time OCP is first solved $$\begin{aligned}
\begin{split}
\operatorname*{minimize}_{u^i(\cdotp), \hspace{0.5ex}t^i_p }\hspace{3.7ex}
& J_p(m_i) \label{c8:eq:MotionPrimitiveGenOCP}\\
\operatorname*{subject\:to}\hspace{3ex}
& \dot{x}^i(t) = f(x^i(t),u^i(t)), \\
&x^i(0) =x^i_s, \quad g^i(x^i(t_f^i)) = 0, \\
&x^i(t) \in \pazocal X^i, \quad
u^i(t) \in \pazocal U^i.
\end{split}\end{aligned}$$ Here it is not required that $x^i(t_f^i)\in\pazocal X_d$. To ensure that the final state $x_f^i=x^i(t_f^i)\in\pazocal X_d$, a rounding heuristic is used and the closest neighboring states represented in the discretized state-space $\pazocal X_d$ from $x^i(t_f^i)$ are evaluated and the solution with lowest objective functional value is selected as the resulting motion primitive $m_i$. Finally, since the MSNT with car-like tractor is orientation invariant, rotational symmetries are exploited to reduce the number of OCPs needed to be solved [@pivtoraiko2009differentially]. For more details of the motion primitive generation framework, the reader is referred to [@bergman2019improved]. Note that the vehicle’s physical constraints $\pazocal X^i\subseteq \pazocal X $ and $\pazocal U^i \subseteq \pazocal U $ in are defined to be maneuver dependent, which is not the case in [@bergman2019improved]. This extension is made to automatically generate similar maneuvers, i.e., same terminal manifold $g^i(x^i(t_f^i))=0$, but resulting in different optimal final states $x^i_f$ and final times $t_f^i$. This additional freedom can be used to design a more flexible lattice planner or, e.g., to adapt to a change in available trailer steering angles $\gamma_s$, $s\in\pazocal I_s$ during different maneuvers.
As proposed in [@bergman2019improved], the motion primitive set is build upon optimized straight, heading change and parallel maneuvers. The heading change and parallel maneuvers are only possible to use from states where the tractor has nonzero velocity, i.e. $v_{0,s}=\pm \bar v$, and are designed to end up in the same final velocity $v_{0,f}=v_{0,s}$. Additionally, short straight maneuvers from $v_{0,s}\in V$ to some $v_{0,f}\in V$ are also optimized to enable the tractor to reduce, increase and keep a constant longitudinal velocity. The heading change and parallel maneuvers are computed using the following terminal manifolds.
**Heading change maneuvers:** By specifying the vehicle’s physical constraints $\pazocal X^i\subseteq \pazocal X$ and $\pazocal U^i \subseteq \pazocal U $, a heading change maneuver from an initial state $x^i_s\in\pazocal X_d$ with pose and $v^i_{0,s} = \pm \bar v$, to a user-defined adjacent orientation $\theta^i_{N,f}\in\Theta \setminus \theta^i_{N,s}$ is optimized by solving using the following terminal constraint
$$\begin{aligned}
g^i\left(p^i_N(t^i_f),v^i_{0,f}\right) = \begin{bmatrix}
\theta^i_N(t^i_f) - \theta^i_{N,f} \\
v^i_{0}(t^i_f) - v^i_{0,s}
\end{bmatrix}=0,
\label{c8:eq:heading_change_constraint}\end{aligned}$$
which implies that $x^i_N(t_f)$ and $y^i_N(t_f)$ are free variables for the OCP solver to select. Note that the vehicle states that are left out from the argument to $g^i$ are all constrained to zero to guarantee that $x^i_f\in\pazocal X_d$. Examples of computed heading change maneuvers from $(\theta_{3,s},v_{0,s})=(\pi/2,\pm 1)$ are depicted in Figure \[c8:fig:primitives\] for an MS3T vehicle with car-like tractor where the last trailer has steerable wheels, i.e., $\pazocal I_s = \{3\}$. Here, the allowed trailer steering angle $|\gamma_3|\leq \bar\gamma_3$ is alternated using $\bar \gamma_3 = 0$, $0.175$ and $0.35$ rad, resulting in different types of optimal trajectories.
![A subset of the motion primitives in $\pazocal P$ for an MS3T vehicle with a car-like tractor from initial position at the origin $(x_{3,s},y_{3,s})=(0,0)$ to different final states $x_f\in\pazocal X_d$. The colored lines are the trajectories in $(x_3(\cdotp),y_3(\cdotp))$ for the different maneuvers. The set of heading change maneuvers from $(\theta_{3,s},v_{0,s})=(\pi/2,1)$ (black) and from $(\theta_{3,s},v_{0,s})=(\pi/2,-1)$ (green). The set of parallel maneuvers from $(\theta_{3,s},v_{0,s})=(0,1)$ (blue) and from $(\theta_{3,s},v_{0,s})=(0,-1)$ (red).[]{data-label="c8:fig:primitives"}](primitives_MS3T.pdf){width="0.55\linewidth"}
**Parallel maneuvers:** A parallel maneuver from initial state $x^i_s\in\pazocal X_d$ with pose $p^i_{N,s}=[
0\hspace{5pt} 0\hspace{5pt} \theta^i_{N,s}]^T$ and , is defined with a user-defined lateral displacement $z^i_{\text{lat}}$ in $(x^i_{N,f},y^i_{N,f})$ with respect to the initial orientation $\theta^i_{N,s}$. This maneuver can be optimized by solving using the following terminal constraint $$\begin{aligned}
g^i\left(p^i_N(t^i_f),v^i_{0,f}\right) = \begin{bmatrix}
y^i_N(t^i_f)\cos\theta^i_{N,s} + x^i_N(t^i_f)\sin\theta^i_{N,s} - z^i_{\text{lat}}\\
\theta^i_N(t^i_f) - \theta^i_{N,s} \\
v^i_{0}(t^i_f) - v^i_{0,s}
\end{bmatrix}=0.
\label{c8:eq:parallel_constraint}\end{aligned}$$ Here, the final position of the $N$th trailer $(x_N(t_f^i),y_N(t_f^i))$ is restricted to a line defined by the first row in $\eqref{c8:eq:parallel_constraint}$. Examples of computed parallel maneuvers for an MS3T vehicle with car-like tractor using from can be seen in Figure \[c8:fig:primitives\].
When the motion primitive set $\pazocal P$ is computed, a free-space heuristic look-up table (HLUT) is computed using techniques presented in [@knepper2006high]. The HLUT is computed offline by solving several obstacle-free graph-search problems from all initial states $\bar {x}_I\in\pazocal X_d$ with $(x_{N,I},y_{N,I})=(0,0)$, to final states with positions on a bounded grid around the origin. This computation can be done efficiently using Dijkstra’s search, as the optimal cost-to-come is simply recorded and stored in the HLUT [@knepper2006high]. Moreover, in analogy to the motion primitive generation, the size of the HLUT is kept small by exploiting the position and orientation invariance properties of $\pazocal P$ [@CirilloIROS2014].
Homotopy-based optimization step {#p8:sec:homo_optimization}
================================
Similar to [@bergman2019bimproved], the optimization step is used to improve the initial guess $(x(t), u(t))$, computed by the lattice planner such that the final trajectory $( x^*(t), u^*(t))$, is a locally optimal solution to . Since the lattice planner uses a discretized state space $\pazocal X_d$, in general its computed state trajectory does not satisfy the initial and goal state constraints in . Thus, the optimization step should not only improve the initial guess but also make it feasible in the original problem formulation . To handle this in a structured way, a homotopy-based initialization strategy is used that is inspired by the work in [@bergman2018combining]. The idea is to start from a relaxed problem that is easy to solve and then gradually transform the relaxed problem to the original one. Here, these ideas are applied on the initial and goal state constraints in such that the solution obtained from the lattice planner is feasible to the relaxed problem. By letting $\epsilon_p^T = [\epsilon_{p,I}\hspace{5pt}\epsilon_{p,G}]\in[0,1]^2$ denote the homotopy parameters [@bergman2018combining], the initial and goal state constraints in are relaxed to $$\begin{aligned}
\label{c8:eq:homoInitGoal}
\begin{split}
& x(0) = \epsilon_{p,I}{\bar x}_I + (1-\epsilon_{p,I}){ x}_I, \\
& x(t_G) = \epsilon_{p,G}{\bar x}_G + (1-\epsilon_{p,G}){ x}_G.
\end{split}\end{aligned}$$ When $\epsilon_p^T = [1\hspace{5pt} 1]$, the initial guess from the lattice planner is feasible to the relaxed version of and when $\epsilon_p^T = [0\hspace{5pt} 0]$, the original problem in is obtained. One possibility is to start with $\epsilon_p^0 = [1\hspace{5pt} 1]^T$ and repeatedly solve the relaxed version of using an OCP solver where $\epsilon_p^k$ is gradually decreased using a fixed step-size $\Delta\epsilon_p$ until $ \epsilon_p^{k} = [0\hspace{5pt} 0]^T$ is reached [@bergman2018combining]. In this work, the idea is instead to let the OCP solver automatically modify the homotopy parameters using a penalty method [@nocedal2006numerical]. Define the linear penalty as $ c_p^T\epsilon_p$, where $ c_p\in\mathbb R^2_{++}$ and define the relaxed version of the trajectory-planning problem as $$\begin{aligned}
\begin{split}
\operatorname*{minimize}_{ u(\cdotp), \hspace{0.5ex}t_G,\hspace{0.5ex}\epsilon_p }\hspace{3.7ex}
& J_{H} = J + c_p^T\epsilon_p \label{c8:eq:MotionPlanningOCPhomo}\\
\operatorname*{subject\:to}\hspace{3ex}
& \dot{ x}(t) = f( x(t), u(t)), \\
& x(0) = \epsilon_{p,I}{\bar x}_I + (1-\epsilon_{p,I}){ x}_I, \\
& x(t_G) = \epsilon_{p,G}{\bar x}_G + (1-\epsilon_{p,G}){ x}_G, \\
& x(t) \in \pazocal X_{\text{free}}, \quad
u(t) \in \pazocal U, \quad \epsilon_p\in[0,1]^2,
\end{split}\end{aligned}$$ which is initialized with the solution from the lattice planner and $\epsilon_p^T = [1\hspace{5pt} 1]$. By choosing $c_p$ sufficiently large, the OCP solver will automatically adjust the step size of $\epsilon_p$ and converge to if a feasible solution to exists in the homotopy class selected by the lattice planner [@bergman2018combining]. As previously mentioned, if is obtained, a locally optimal solution $( x^*(t), u^*(t))$, $t\in[0,t^*_G]$ to the original trajectory planning problem is obtained which can then be sent to a trajectory-tracking controller for plan execution.
Note that if $c_p$ is not chosen sufficiently large, a solution with may not be obtained even though one exists [@nocedal2006numerical]. In that case, one may need to increase $c_p$ and continue the solution process. However, in extensive simulation trials presented in Section \[p8:sec:results\], it is shown that the proposed homotopy-based optimization step is able to reliably compute locally optimal solutions to with without modifying $c_p$ in all problem instances.
Simulation results {#p8:sec:results}
==================
In this section, the proposed trajectory planning framework is evaluated in two complicated parking problem scenarios for an MS3T with car-like tractor where only trailer $N=3$ is steerable, i.e., $\pazocal I_s = \{3\}$, and a mixture of off-axle ($M_1\neq 0$) and on-axle hitching ($M_2=M_3=0$). Using the recursive model presented in Section \[p8:sec:model\] it is now straightforward to derive the vehicle model for this specific MS3T vehicle with configuration $ q=[\beta_0\hspace{5pt}\beta_1\hspace{5pt}\beta_2\hspace{5pt}\beta_3\hspace{5pt}\gamma_3\hspace{5pt}\theta_3\hspace{5pt}x_3\hspace{5pt}y_3]^T$, augmented state vector $ x = [ q^T\hspace{5pt}\omega_0\hspace{5pt}\omega_3\hspace{5pt}v_0\hspace{5pt}a_0]^T$ and control signals $ u = [u_{\omega_0}\hspace{5pt}u_{\omega_3}\hspace{5pt}u_{v}]^T$. The vehicle model $\dot{ x}= f( x, u)$ is presented in Appendix A and the values of the vehicle’s parameters used in this section are summarized in Table \[c8:tab:vehicle\_parameters\].
[l l]{} Vehicle parameter & Value\
Tractor’s wheelbase $L_0$ & 4.6 m\
Length of the off-hitch $M_1$ & 1.6 m\
Length of trailer 1 $L_1$ & 2.5 m\
Length of trailer 2 $L_2$ & 7.0 m\
Length of trailer 3 $L_3$ & 7.0 m\
Maximum joint angles $\bar\beta_i$, $i=1,2,3$ & 0.87 rad\
Maximum steering angle tractor $\bar\beta_0$ & 0.73 rad\
Maximum steering-angle rate tractor $\bar\omega_0$ & $0.8$ rad/s\
Maximum steering-angle acceleration tractor $\bar\Omega_0$ & $10$ rad/s$^2$\
Maximum steering angle trailer 3 $\bar\gamma_3$ & 0.35 rad\
Maximum steering-angle rate trailer 3 $\bar\omega_3$ & $0.4$ rad/s\
Maximum steering-angle acceleration trailer 3 $\bar\Omega_3$ & $10$ rad/s$^2$\
Maximum longitudinal speed tractor $\bar v$ &1 m/s\
Maximum longitudinal acceleration tractor $\bar a$ &$1$ m/s$^2$\
Maximum longitudinal jerk tractor $\bar u_{v}$ &40 m/s$^2$\
\[c8:tab:vehicle\_parameters\]
The cost function is chosen as $$\begin{aligned}
l( x, u) = \frac{1}{2}\left(\beta_0^2+\gamma_s^2+10\omega_0^2+10\omega_3^2 + a_0^2 + u^T u\right),\end{aligned}$$ which is used in all steps of the trajectory planning framework. The linear cost on the homotopy parameters in the optimization step is chosen as $c_p^T = [1000\hspace{5pt}1000]$. The lattice planner is implemented in `C++`, whereas the motion primitive generation and the homotopy-based optimization step are both implemented in Python using [@casadi], where is used as nonlinear programming problem solver. All simulations are performed on a laptop computer with an Intel Core [email protected] CPU.
The motion primitive set consists of heading change, parallel and straight trajectories where and a subset of the motion primitive set can be seen in Figure \[c8:fig:primitives\]. From each initial heading with nonzero longitudinal velocity, there are $20$ parallel and $24$ heading change maneuvers. The heading change maneuvers are computed using three different limits on the trailer steering angle $\bar \gamma_3=0$, $0.175$ and $0.35$ rad, respectively, and the parallel maneuvers using . After the motion primitive set is computed, a free-space HLUT is computed on a square grid $80\times80$ m centered around the origin. To evaluate if the trajectory planner is able to exploit the additional trailer steering, it is compared with an SS3T vehicle, i.e., $\bar \gamma_3=0$, with the same vehicle parameters and the difference that only $8$ heading change maneuvers exist in the motion primitive set $\pazocal P_{\text{SS3T}}$ ($|\pazocal P_{\text{SS3T}}|=1124$).
![Loading-bay parking scenario using the proposed trajectory planner for SS3T and MS3T from two initial positions ($ p_I^A$ and $ p_I^B$), symmetric with respect to the loading bay, initial orientations $\theta_{3,I}\in\Theta$ and zero initial joint angles, to the goal state $ x_G$. The solutions for the position ($x^*_3(\cdotp),y^*_3(\cdotp)$) from 1029 perturbed initial states for SS3T ($ p_I^A$) and MS3T ($p_I^B$) are displayed by blue (red) solid lines for MS3T (SS3T). The initial guess computed by the lattice planner for ($ x_3(\cdotp), y_3(\cdotp)$) and optimization step for ($x^*_3(\cdotp),y^*_3(\cdotp)$) from straight configuration is marked with green (orange) dashed line and green (orange) solid line, respectively, for MS3T (SS3T).[]{data-label="c8:fig:Loading_bay"}](loading_bay.pdf){width="0.8\linewidth"}
[l l l l l l l]{} Vehicle & $\bar t_{\text{lat}}$\[s\] & $\bar t_{\text{ocp}}$\[s\] & $\bar J_D$ & $\bar J_H$ & $\bar r_{\text{imp}}$ & $\bar t_{G,\text{imp}}$\[s\]\
SS3T & 0.11 & 9.3 & 170.8 & 128.4 & -25% & -26.8\
MS3T & 0.09 & 3.0 & 126.8 & 100.8 & -21% & -14.4\
\[c8:tab:loading\_bay\_results\]
The first planning scenario is a loading-bay parking problem that is illustrated in Figure \[c8:fig:Loading\_bay\]. The obstacles and vehicle bodies are described by bounding circles [@lavalle2006planning], where in total, the vehicle bodies are described using 8 bounding circles of radius 2 m. The objective of the trajectory planner is to plan a trajectory from 32 different initial states $x_I\in\pazocal X_d$ (see Figure \[c8:fig:Loading\_bay\]) to the goal state $x_G$. One solution example is provided for the lattice planner (dashed line) and optimization step (solid line) for SS3T (orange) and MS3T (green), respectively, for symmetric planning problems. The results show that the planned trajectory for MS3T is purely in backward motion, as apposed to SS3T which needs to combine forward and backward motion due to less steering capability. A summary of the simulation results are provided in Table \[c8:tab:loading\_bay\_results\]. The average computation time for the lattice planner is only [$0.1$ s]{} for both SS3T and MS3T, whereas the optimization step takes in average [$3.0$ s]{} for MS3T and [$9.0$ s]{} for SS3T. However, a signification reduction of both average cost and time improvement of the solutions are obtained. Note that the average time improvement of the solutions computed by the optimization step $\bar t_{G,\text{imp}}$ is significantly larger compared to the optimization step’s average computation time $\bar t_{G,\text{ocp}}$. Thus, when the optimization step is added, the combined average computation and execution time becomes significantly lower.
![Parallel parking scenario using the proposed trajectory planner for SS3T and MS3T from three different initial positions ($ p^1_I$, $ p^2_I$ and $ p^3_I$), six different initial orientations $\theta_{3,I}$ and zero initial joint angles, to the goal state $ x_G$. The blue (red) lines illustrate the planned trajectories after the optimization step for the position of trailer 3 $(x^*_3(\cdotp),y^*_3(\cdotp))$ for MS3T (SS3T) from three selected initial states.[]{data-label="c8:fig:parallel_parking"}](parallel_parking.pdf){width="1\linewidth"}
[l l l l l l l]{} Vehicle & $\bar t_{\text{lat}}$\[s\] & $\bar t_{\text{ocp}}$\[s\] & $\bar J_D$ & $\bar J_H$ & $\bar r_{\text{imp}}$ & $\bar t_{G,\text{imp}}$\[s\]\
SS3T & 5.6 & 8.4 & 270.6 & 122.4 & -55% & -109.6\
MS3T & 11.3 & 2.4 & 207.9 & 98.1 & -52% & -80.4\
\[c8:tab:parallel\_parking\_results\]
To evaluate how the homotopy-based optimization step handles an infeasible initialization, 1029 perturbed initial states for both SS3T and MS3T are evaluated (see Figure \[c8:fig:Loading\_bay\]), where blue and red trajectories are related to MS3T and SS3T, respectively. The initial joint angles are perturbed with , $i=1,2,3$ and initial orientation with $\theta_{3,I}=-10^\circ,0^\circ,10^\circ$. In all cases, the used optimization step is able to handle the infeasible initial guess obtained from the lattice planner, i.e., the value of homotopy parameter $\epsilon_{p,I}^*=0$ in all cases. That is, in all cases, a solution to the original trajectory planning problem is obtained. Moreover, the average computation time of the optimization step is for MS3T and for SS3T. Hence, the active trailer steering also reduces the computation load of the OCP solver.
The second planning scenario is a parallel parking problem with 18 different problems that is illustrated in Figure \[c8:fig:parallel\_parking\] and the results are summarized in Table \[c8:tab:parallel\_parking\_results\]. This scenario is a confined environment which affects the average computation time of the lattice planner $\bar t_{\text{lat}}$, which is for MS3T and for SS3T. This is because the HLUT is drastically underestimating the cost-to-go in this confined environment. Therefore, both the average cost improvement $\bar r_{\text{imp}}$ (MS3T: $-52\%$, SS3T: $-55\%$) and time improvement $\bar t_{G,\text{imp}}$ (MS3T: $-80.4$ s, SS3T: $-109.6$ s) of the optimization step are significant. The confined environment does however not affect the average computation time of the optimization step which is 2.4 s for MS3T and for SS3T. Moreover, as can be seen in the three highlighted planning problems in Figure \[c8:fig:parallel\_parking\], the final optimized solutions for the MS3T only needs to reverse, as opposed to the SS3T which needs to combine forward and backward motion in two cases. Finally, Figure \[p8:fig:parallel\_parking\_state\_traj\] shows the difference between the trajectories from the lattice planner and the optimization step for the selected planning problem from position $p_I^3$ in Figure \[c8:fig:parallel\_parking\]. As can be seen, the trajectories for the two steering angles, the longitudinal velocity and the joint angles are significantly smoother after the optimization step, at the same time as the final time is decreased from $96$ s to $70$ s.
Conclusions {#p8:sec:conclusions}
===========
An optimization-based trajectory planner for multi-steered articulated vehicles is proposed that targets low-speed maneuvers in unstructured environments. The proposed trajectory planner is divided into two steps, where a lattice planner is used in a first step to compute an optimal solution to a discretized version of the trajectory planning problem using a library of precomputed trajectories. In a second step, the output from the lattice planner is then used to initialize a homotopy-based optimization step, which enables the framework to compute a locally optimal solution that starts at the vehicle’s initial state and reaches the goal state exactly. The performance of the proposed optimization-based trajectory planner is evaluated in a set of practically relevant scenarios for a multi-steered 3-trailer vehicle where the last trailer is steerable. In the simulations, it is shown that the framework can solve challenging trajectory planning problems and that the proposed optimization step provides a significant improvement in terms of reduced objective functional value and final time, at the same time as it enables the framework to plan from and to a larger set of different vehicle states.
As future work we would like to develop a trajectory-tracking controller and evaluate the framework in real-world experiments on a full-scale test vehicle.
Appendix A {#appendix-a .unnumbered}
==========
In this section, the model for the specific MS3T vehicle used in the simulation trails is presented. The matrices $ J_1$, $ J_2$ and $ J_3$ describing the longitudinal and angular velocity transformations between neighboring vehicle segments are $$\begin{aligned}
\label{c8:MS3T_vel_matrices}
\begin{split}
J_1(\beta_1,0,0)&=\begin{bmatrix}
-\frac{M_{1}}{L_{1}}\cos\beta_{1} & \frac{\sin\beta_1}{L_{1}} \\[10pt]
M_{1}\sin\beta_1 & \cos\beta_{1}
\end{bmatrix},
\quad
J_2(\beta_2,0,0)=\begin{bmatrix}
0 & \frac{\sin\beta_2}{L_2} \\[10pt]
0 & \cos\beta_2
\end{bmatrix}, \\
J_3(\beta_3,\gamma_3,0)&=\begin{bmatrix}
0 & \frac{\sin(\beta_3-\gamma_3)}{L_3\cos\gamma_3} \\[10pt]
0 & \frac{\cos\beta_3}{\cos\gamma_3}
\end{bmatrix},
\end{split}\end{aligned}$$ since $M_2=M_3=0$ and $\gamma_0=\gamma_1=\gamma_2=0$. Thus, by inserting in the longitudinal velocity $v_{3}$ of trailer 3 can be written as $$\begin{aligned}
v_{3} = v_0\frac{\cos\beta_3}{\cos\gamma_3}\cos\beta_2\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right).
\label{c8:MS3T_vel_trailer3}\end{aligned}$$ Finally, using and in –, the complete model for the MS3T vehicle becomes @size[10]{}@mathfonts $$\begin{aligned}
\begin{split}
\label{c8:vehicle_model_MS3T}
\dot\beta_1 =& v_0\left(\frac{\tan\beta_0}{L_0} -\frac{\sin\beta_1}{L_1} + \frac{M_1}{L_0L_1}\cos\beta_1\tan\beta_0\right),\\
\dot\beta_2 =& v_0\left[\left(\frac{\sin\beta_1}{L_1} - \frac{M_1}{L_0L_1}\cos\beta_1\tan\beta_0\right) \right. \\ &\left.
-\frac{\sin{\beta_{2}}}{L_{2}}\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right)\right], \\
\dot\beta_3 =& v_0\left[\frac{\sin{\beta_{2}}}{L_{2}}\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right) \right. \\ &\left. - \frac{\sin{(\beta_{3}-\gamma_{3})}}{L_{3}\cos\gamma_{3}}\cos\beta_2\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right)\right], \\
\dot \theta_{3} =&
v_0\frac{\sin{(\beta_{3}-\gamma_{3})}}{L_{3}\cos\gamma_{3}}\cos\beta_2\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right), \\
\dot x_3 =&v_0\frac{\cos\beta_3}{\cos\gamma_3}\cos\beta_2\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right) \cos(\theta_{3}+\gamma_3), \\
\dot y_3 =& v_0\frac{\cos\beta_3}{\cos\gamma_3}\cos\beta_2\left(\frac{M_1}{L_0}\sin\beta_1\tan\beta_0 + \cos\beta_1\right) \sin(\theta_{3}+\gamma_3), \\
\dot \beta_0 =& \omega_0, \quad \dot \omega_0 = u_{\omega_0}, \\
\dot \gamma_3 =& \omega_3,\quad \dot \omega_3 = u_{\omega_3}, \\
\dot v_0 =& a_0, \quad \dot a_0 = u_v.
\end{split}\end{aligned}$$
[^1]: $\Theta$ is the the set of unique angles $-\pi<\theta_{N}\leq \pi$ that can be generated by $\theta_{N} = \arctan2(i,j)$ for two integers $i,j\in\{-2,-1,0,1,2\}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this conference proceedings, some comments on the present status and recent growth of efforts to find Lorentz and CPT violation are given by extracting metrics from the annually updated [*Data Tables for Lorentz and CPT Violation.*]{} They reveal that tests span all the sectors of particle and gravitational physics, and have shown remarkable and consistent growth. Through numerous innovations and refinements in experiments, a large body of data has been amassed with ever-increasing precisions. The Tables are available through the Cornell preprint archive at arXiv:0801.0287.'
address: |
Physics Department, Northern Michigan University,\
Marquette, Michigan 49855, USA
author:
- 'N.E. Russell'
title: Mining the Data Tables for Lorentz and CPT Violation
---
The first edition of the [*Data Tables for Lorentz and CPT Violation*]{} appeared in the Proceedings of the Fourth Meeting on Lorentz and CPT Symmetry and on the Cornell preprint server arXiv.org, where it has been updated each year.[@datatables] A July 2010 edition appeared in Reviews of Modern Physics.[@rmp2011] The purpose of the Tables is to maintain a complete list of measurements of coefficients for Lorentz and CPT violation, to summarize the maximal attained sensitivities in high-activity sectors, and to provide information about properties and definitions relevant to the study of these fundamental symmetries.
With the Tables being in existence for more than ten years, and on the occasion of this Eighth Meeting on Lorentz and CPT Symmetry being held more than 20 years since the first such meeting in 1998, it is a good time to take a retrospective look at the extraordinary theoretical and experimental efforts that have been made by hundreds of researchers since the 1990s to find evidence of Lorentz violation in nature.
The burgeoning number of publications placing limits on Lorentz and CPT symmetry can be traced in the annual updates to the Tables. The upper plot in Fig. \[growth\] shows the number of references in each edition of the Tables growing five-fold from 62 in 2008 to 292 in 2019. Another measure is the page count, which has grown about ten-fold from twelve to 115 over the same period.
![The number of bibliographic references and the page count for the twelve editions of the [*Data Tables for Lorentz and CPT Violation*]{} up to January 2019. []{data-label="growth"}](Metrics3.eps){width="180pt"}
The Standard-Model Extension, or SME, categorizes Lorentz violation in the behavior of all known particles and their interactions,[@dcak] including gravitational ones.[@akgravity] It is the result of considering effective field theory with Lorentz and CPT violation, leading to a general framework incorporating known physics but also admitting violations of these symmetries. The Tables list measurements of the coefficients controlling Lorentz- and CPT-breaking operators of all mass dimensions in the framework.
![The number of Data Table references presenting limits in each of the sectors of the Standard-Model Extension.[]{data-label="sectors"}](Sectors3.eps){width="170pt"}
Figure \[sectors\] indicates the distribution of efforts across the broad scope of the field. The size of each slice of the pie chart is set by the number of publications placing limits on coefficients for Lorentz and CPT violation. The QED sectors account for more than half of the chart. Limits in the photon sector, the largest one, have been reported in 77 publications to date, with the electron, proton, and neutron sectors represented by 48, 34, and 27 publications, respectively. Rounding out the QED sector, eleven papers have presented limits on Lorentz-breaking couplings between matter and photons.
In the neutrino sector, Fig. \[sectors\] shows 40 papers are referenced. In fact, the most recent edition of the Tables has 27 pages of neutrino-sector limits. The quark sector is also well represented, with 23 publications. In comparison, the lepton, electroweak, and gluon sectors have fewer publications so far, reflecting the corresponding experimental challenges and, possibly, revealing some of the many areas where Lorentz and CPT searches could be very profitable. Moving into curved spacetime, the gravitational sector is well represented with 35 publications to date.
![Evolution of sensitivities to $({\tilde\kappa}_{o+})^{YZ}$, one of the 19 coefficients governing Lorentz-violating operators of mass dimension four in the photon sector. The scale is logarithmic, with the first limit placed at a few parts in $10^9$. The first-named author and the year of publication is listed for each result.[@photonexpts] []{data-label="photon"}](PhotonSensitivity3.eps){width="180pt"}
Experimentalists have risen to the challenge of finding the elusive effects of Lorentz violation by steadily improving the technology over a period of years. As an example of this, Fig. \[photon\] shows the evolution in sensitivity to the coefficient $({\tilde\kappa}_{o+})^{YZ}$. The first limit on this dimensionless coefficient was placed in 2003, and up to the present it has been improved 13 times. The sequence of results, originating in labs on three continents, spans more than a decade and shows the experimental reach improving by six orders of magnitude.
At present, the Tables show no evidence of Lorentz or CPT violation. They also show that sectors with unmeasured or very weakly constrained coefficients exist. The potential for finding evidence of Plank-scale physics is of course a central motivator for the field, and phenomenological and experimental efforts continue to grow. As experimental technology improves, the prospect of violations being revealed remains tantalizing.
[x]{}
V.A. Kostelecký and N. Russell, arXiv:0801.0287v1 (2008 edition) to arXiv:0801.0287v12 (2019 edition); [*id.*]{}, in V.A. Kostelecký, ed., [*CPT and Lorentz Symmetry IV*]{}, World Scientific, Singapore, 2008.
V.A. Kostelecký and N. Russell, Rev. Mod. Phys. [**83**]{}, 11 (2011).
D. Colladay and V.A. Kostelecký, Phys. Rev. D [**55**]{}, 6760 (1997); Phys. Rev. D [**58**]{}, 116002 (1998).
V.A. Kostelecký, Phys. Rev. D [**69**]{}, 105009 (2004).
J.A. Lipa , Phys. Rev. Lett. [**90**]{}, 060403 (2003); H. Müller , Phys. Rev. Lett. [**91**]{} 020401 (2003); P. Wolf , Phys. Rev. D [**70**]{}, 051902 (2004); P.L. Stanwix , Phys. Rev. Lett. [**95**]{}, 040404 (2005); S. Herrmann , Phys. Rev. Lett. [**95**]{}, 150401 (2005); P.L. Stanwix , Phys. Rev. D [**74**]{}, 081101(R) (2006); H. Müller , Phys. Rev. Lett. [**100**]{}, 031101 (2008); S. Herrmann , in V.A. Kostelecký, ed., [*CPT and Lorentz Symmetry IV*]{}, World Scientific, Singapore, 2008; S. Herrmann , Phys. Rev. D [**80**]{}, 105011 (2009); Ch. Eisele, A.Yu. Nevsky, and S. Schiller, Phys. Rev. Lett. [**103**]{}, 090401 (2009); M.A. Hohensee , Phys. Rev. D [**82**]{}, 076001 (2010); J.-P. Bocquet , Phys. Rev. Lett. [**104**]{}, 241601 (2010); Y. Michimura , Phys. Rev. Lett. [**110**]{}, 200401 (2013); M. Nagel , Nature Commun. [**6**]{}, 8174 (2015); V.A. Kostelecký, A.C. Melissinos, and M. Mewes, Phys. Lett. B [**761**]{}, 1 (2016).
|
{
"pile_set_name": "ArXiv"
}
|
1.5 cm
[**BOUNDARY DYNAMICS IN DILATON GRAVITY**]{}
.5 cm
[**Sumit R. Das**]{}
[*Tata Institute of Fundamental Research Homi Bhabha Road, Bombay 400005, INDIA e-mail: [email protected]*]{}
and .3 cm [**Sudipta Mukherji**]{}
[*International Centre for Theoretical Physics I-34100 Trieste, ITALYe-mail: [email protected]*]{}
.5 cm
[**ABSTRACT**]{}
.2 cm We study the dynamics of the boundary in two dimensional dilaton gravity coupled to $N$ massless scalars. We rederive the boundary conditions of [@DAM1] and [@CHVER] in a way which makes the requirement of reparametrization invariance and the role of conformal anomaly explicit. We then study the semiclassical behaviour of the boundary in the $N=24$ theory in the presence of an incoming matter wave with a constant energy flux spread over a finite interval. There is a critical value of the matter energy density below which the boundary is stable and all the matter is reflected back. For energy densities greater than this critical value there is a similar behaviour for small values of the [*total*]{} energy thrown in. However, when the total energy exceeds another critical value the boundary exhibits a runaway behaviour and the spacetime develops singularities and horizons.
Introduction
============
In a recent paper [@DAM1], we have studied a model of two dimensional gravity and massless matter field with a boundary in space time [@SVV; @CHVER]. The reason for introducing such a boundary in the model is many-fold. In the two dimensional reduction of four dimensional spherically symmetric space time such a boundary may be thought as the origin of the radial coordinate. Moreover semiclassical studies of dilaton gravity models in two dimensions inevitably lead to a boundary beyond which the theory becomes strongly coupled. In a ficucial coordinate system $(u, v)$ the boundary may be chosen to be the line $u = v$. We will refer to this coordinate system as the “fixed boundary gauge”. In these coordinates the conformal factor of the metric and the dilaton are independent fields. However, $u$ and $v$ are not asymptotically minkowskian. Alternatively we can work in Kruskal coordinates which are simply related to the asymptotically minkowskian coordinates. In these coordinates the dilaton is equal to the liouville mode and the boundary is [*dynamical*]{}. Its world line is determined by the matter energy momentum tensor, through a [*boundary equation*]{} derived in [@DAM1; @CHVER]. In the fixed boundary gauge the same equation determines the relationship between the dilaton and the liouville mode.
In the full quantum theory the boundary undergoes quantum fluctuations which are determined by the boundary equation, which has to be now regarded as an operator equation. (In the fixed boundary gauge, these are the dilaton and liouville fields which fluctuate.) The dilaton-gravity sector of the theory has no physical propagating degrees of freedom, but the degree of freedom corresponding to the boundary survives. Thus we get an idea about the strength of quantum gravity in such simple model by studying boundary fluctuations, as done in [@DAM1].
In some sense the boundary acts as a model for the high energy, strong coupling physics of the problem. Just as in the Callan-Rubakov effect we would like to model the high energy physics by suitable boundary conditions. The boundary conditions must respect some minimum requirements, like conservation of charges which should be conserved. In general such minimal requirements may not be restrictive enough to enable us to get insight into the real problem. However, it is also common that minimal requirements turn out to be rather restrictive and allow one to derive “model indpendent” consequences. The boundary conditions which arise in the Callan-Rubakov effect is such an example [@DCAL].
In the present problem, the boundary conditions must respect conservation of energy momentum and the general covariance of the underlying theory. In a gauge fixed theory the latter means that the reparametrization constraints must be obeyed (in a conformal gauge this is the same as maintaining conformal invariance).
The most general boundary conditions which respect these requirements were derived in [@CHVER; @DAM1]. The conditions of general covariance and energy-momentum conservation turn out to be rather restrictive. Assuming that the matter is reflected off the boundary separately, the boundary conditions in the dilaton gravity sector is characterized by a single parameter $\beta$, which is related to the value of the dilaton field at the boundary. In [@DAM1] a “boundary equation” was derived starting from the fixed boundary gauge. The behavior of the boundary and the resulting space-time structure is similar for all non-zero values of $\beta$, while $\beta =0$ (which are the boundary conditions of [@RST; @SVV]) is a rather singular limit. In fact the boundary conditions arise dynamically when one adds a boundary cosmological constant term proportional to $\beta$ to the two dimensional action [@STHOR; @STROM], which immediately explains why $\beta = 0$ is rather special. Our model is in fact similar to the boundary conformal field theories considered recently in [@BCON].
In [@DAM1] the semiclassical boundary equation was solved in the $N=24$ theory for an incident shock wave. For $\beta \ne 0$ there is a critical value of the strength of the shock wave, i.e. the [*total*]{} energy. Below this critical value the boundary has a stable timelike evolution and no black hole is formed. When the energy exceeds the critical value the boundary runs away with ever increasing acceleration, approaching the speed of light. At the same time a black hole is formed which subsequently evaporates. The stability of the boundary for small disturbances have been recently analyzed in [@STHOR] for arbitrary $N$ and an intriguing proposal based on the similarity of the model with open string theory has been put forward for the $N = \infty$ theory in [@STROM].
Note that a limit on the strength of the shock wave does not set any limit on the [*energy density*]{} for an aribitrary pulse. The $\beta = 0$ theory, which is essentially the RST model, can be solved for arbitrary distribution of incoming matter energy and a black hole is formed when the [*energy density*]{} exceeds a critical value. However, in this model the boundary becomes space-like whenever this critical value is exceeded. It is of interest to know the behavior of the $\beta \ne 0$ model for more general incoming matter distribution and ask whether there is some critical value for the [*energy density*]{} as well.
In this letter we first rederive the boundary condition and the boundary equation for general $N$ in the Kruskal gauge. This derivation makes the role of reparametrization invariance and the anomaly in the boundary conditions explicit. We then solve the semiclassical boundary equation for $N = 24$ for an incoming matter distribution which consists of a constant energy density in a finite interval of retarded time. We find that for the boundary to become unstable and a black hole to form, there is a critical value of the energy density as well as the total energy.
The Boundary Equation
=====================
The classical action of the model is given by S = [1]{}du dv [g]{}\[e\^[-2 ]{}(R - 4 ()\^2 - 4 \^2 + [12]{} (f)\^2 \]\[eq: one\] where $g_{ab}(u, v)$, $\phi (u, v)$ and $f^{i}(u, v)$ are the two dimensional metric, dilaton and the matter fields. $R$ is the curvature scalar and $\lambda$ is the comological constant which we set to one by properly choosing the scale.
We will use the conformal gauge where $g_{ab} = e^{2\rho} \eta_{ab}$. Since the matter is conformally coupled its quantum effects are entirely contained in the Weyl anomaly which means that there is a liouville term in the action. There are additional liouville terms coming from the measures of integration of the conformal mode $\rho$, the dilaton and the ghosts. These measures are chosen so as to ensure that the Hawking radiation consists of only physical propagating particles, which are the matter fields $f^i$. Following RST one can use the freedom to add a local counterterm to make the semiclassical model solvable. The resulting action may be written as S = du dv\[2 - e\^[2]{} - 2(1 - [1 ]{})+ f\^i f\^i + ([ghosts]{})\] \[eq:two\] where we have defined the fields = [1 ]{}\[e\^[-2]{}+\] =-\[eq:three\] and $\kappa = {N \over 24}$.
It is also possible to write down models which are exactly conformally invariant [@BCAL; @DEAL]. These models have an action which is identical to (\[eq:two\]). However the expressions for $\Omega$ and $\rh$ in terms of $\phi$ and $\rho$ are different. The boundary conditions and the boundary equation which we will derive are valid for these theories as well. However the interpretation of the solution in terms of the space-time structure would be different.
The equations of motion following from (\[eq:two\]) are very simple = f\^i = 0 = -e\^[2]{} \[eq:four\] Following [@SVV] we introduce two chiral fields $\xp(u)$ and $\xm(v)$ and solve the equation for $\rh$ e\^[2(-)]{} = \_u (u) \_v (v)\[eq:five\] The equation for $\Omega$ has the general solution = - (u) (v) + g\^+(u) + g\^-(v) + K\[eq:six\] where $K$ is a constant and $g^\pm$ are arbitrary chiral functions. We can trade these functions for two other chiral fields $\yp (v)$ and $\ym (u)$ which are defined as \_u g\^+(u) = Y\^-(u)\_u(u), \_v g\^-(v) = (v) \_v (v), \[eq:seven\]
In the gauge fixed theory the requirement of general covariance of the underlying model is imposed by setting the total energy momentum tensor to be zero T\^g\_[uu]{}+T\^m\_[uu]{}+T\^[gh]{}\_[uu]{}=0 T\^g\_[vv]{}+T\^m\_[vv]{}+T\^[gh]{}\_[vv]{}=0 \[eq:eight\] where $T^g, T^m, T^{gh}$ stand for the graviton-dilaton, matter and the ghost energy momentum tensors respectively. In what follows, the ghosts completely decouple. The matter part $T^m$ is standard, while the graviton-dilaton part has the following expressions T\^g\_[uu]{} & = & \_u \_u + [2]{} \[(\_u[log]{}\_u)\^2 - 2\_u\^2[log]{}\_u \]\
T\^g\_[vv]{} & = & \_v \_v + [2]{} \[(\_v[log]{}\_v )\^2 - 2\_v\^2[log]{}\_v\]\[eq:nine\] where $\kh = {N - 24\over{24}}$. Note that in (\[eq:eight\]) above, the matter part $T^m$ contains a vacuum energy term coming from normal ordering effects. Alternatively in the semiclassical theory the vacuum has to be defined in terms of modes which are positive frequency with respect to asymptotically minkowskian coordinates and there is a term which comes from the anomalous transformation of the energy momentum tensor to some other globally valid coordinate system.
We now derive the general boundary condition. The minimal requirements which the theory with these boundary conditions must satisfy are
1. The dilaton field must be a constant along the boundary
2. The matter must be perfectly reflected from the boundary
3. The reparametrization constraints (\[eq:eight\]) must be obeyed.
The last condition guarantees that the boundary conditions are consistent with the general covariance of the model.
Consider a timelike boundary in the theory specified by the equation v = f\_B (u) \[eq:ten\] or its inverse equation u = g\_B (v) \[eq:eleven\] The condition (2) above and the constraint (\[eq:eight\]) imply that the total energy momentum tensor must be perfectly reflected. This means that the right moving energy momentum tensor must be set to be equal to the left moving energy momentum tensor after performing a coordinate transformation by the function (\[eq:eleven\]). In doing so one must remember that energy momentum tensors transform anomalously with a Schwarzian derivative term.
We will impose Dirichlet boundary conditions on the matter fields f\^i(g\_B(v),v)=0 \[eq:twelve\] We will also impose similar boundary conditions on the ghosts. This means that general covariance requires that the gravity energy momentum tensor must obey T\^g\_[vv]{}(v) = (g\_B(v))\^2 T\^g\_[uu]{}\[g\_B(v)\] + { g\_B(v),v } \[eq:thirteen\] where $\{ .,. \}$ denotes a Schwarzian derivative { f,x } = [f”’ f’]{}- [3 2]{}([f” f’]{})\^2
In the bulk theory independent reparametrizations of $u$ and $v$ are still symmetries in the conformal gauge. This is no longer true if we choose the boundary to be a fixed line $u = v$ as in the fixed boundary gauge. However, [*one*]{} of the conformal reparametrizations still remain a symmetry and we can use this to choose $\xp (u) = u$. This leaves $\xm (v)$ as the dynamical degreee of freedom, which is determined by the constraints. This is the gauge used in [@SVV; @DAM1]. In the semiclassical theory, the constraints determine the function $\xm (v)$ in terms of the energy momentum tensor of the matter. In the quantum theory the fluctuations of $\xm (v)$ are determined in terms of the matter fluctuations via the operator constraints. Since in this gauge $g_B(v) = 1$ the Schwarzian derivative term vanishes. However the nontriviality of the transformation properties of the energy momentum tenaor are contained in the expression for $T^g_{vv}$ in (\[eq:nine\]).
Alternatively one may fix the Kruskal gauge $\xp (u) = u$ and $\xm (v) = v$. The boundary cannot be fixed to be some predetermined line any more. Rather the boundary curve is now dynamically determined, as we shall see soon. In the following we will use the Kruskal gauge and accordingly replace the indices $u$ and $v$ by $+$ and $-$ respectively.
The condition that the dilaton is constant along the boundary follows from the solution for $\Omega$ in (\[eq:six\]) and (\[eq:seven\]) (()-)+(f\_B)(()-) \_[=f\_B()]{}=0 \[eq:fourteen\] Let us introduce a new quantity h\_B() = \[eq:fifteen\] The condition (\[eq:fourteen\]) may be satisfied by setting () & = & g\_B() + F\[h\_B(g\_B()),\]\
() & = & f\_B() - [F\[h\_B(),\] h\_B\^2]{} \[eq:sixteen\] where $F[h_B(x),x]$ is some general functional of $h_B(x)$ and a function of $x$.
We now use these expressions for $Y^\pm$ to evaluate $T^g_{++},T^g_{--}$ in (\[eq:thirteen\]) and obtain a functional differential equation for $F$. In the Kruskal gauge one simply has $T^g_{++}=\pp \yp$ etc. However since the boundary curve is nontrivial, the term in (\[eq:thirteen\]) which involves $\kh$ is nonvanishing. One gets the equation F + h\_B [F h\_B]{} = ( h\_B)F +\[( h\_B)\^2 - \^2 h\_B\] \[eq:seventeen\] This has the most general solution F\[h\_B(),\] = h\_B() + h\_B \[eq:eighteen\] where $\beta$ is an arbitrary parameter. This completes the derivation of the most general boundary condition which satsifies the requirements stated above. The term involving $\kh$ is clearly the result of the anomalous transformation law of the energy momentum tensor.
Finally, using the above expression for $F$ one may derive the equation of motion of the boundary. This is simply the statement that the total $T_{++}$ vanishes and is given by h\_B\^2 + h\_B + \^2 h\_B +T\^m\_[++]{}-[2 ()\^2]{} = 0 \[eq:nineteen\] The last term is the standard vacuum energy which appears in the Kruskal gauge.
So far we have not specified the value of the dilaton field on the boundary. At this stage this may appear as a second parameter specifying the boundary. However, as we shall se below, this is actually determined by the condition that the vacuum solution of the boundary equation corresponds to flat space with a linear dilaton.
The boundary equation (\[eq:nineteen\]) determines the shape of the boundary curve in terms of the matter energy momentum tensor. Once this is known, we can use the expressions for $\ym,\yp$ to write the solution for $\Omega$ and hence derive the space-time structure. In the following we will concentrate on the behaviour of the boundary itself.
It is useful to rewrite this equation in terms of the quantity () = () \[eq:twenty\] In terms of $\qb$ the equation (\[eq:nineteen\]) becomes ()\^2 & + (- )() &\
&+(\^2 - + - )& = - [()\^2 ]{}T\_[++]{}\^M \[eq:twoone\] As noted in [@STHOR] this represents the equation of motion for a particle driven with some external force and subjected to some damping. In vacuum $T^M_{++}=0$ and the external force is the gradient of a potential. Consequently one may expect that the ground state is characterized by a constant $\gb$ which is the minimum of this potential. However, for this to occur for real $\gb$ one must have $\beta^2 + 2 > 4\kk$. For general $N$ one cannot, therefore, choose $\beta = 0$ [@CHVER]. Furthermore for the vacuum solution to be stable under small deviations, one must have $\kh \geq 0$ and $\beta > {{\sqrt{2}}\kh \over \kappa}$.
In the rest of the paper we shall limit ourselves to $N=24$ for which the vacuum is perturbatively stable.
Semiclassical Solution with a finite duration wave
==================================================
We now investigate the stability of the boundary for arbitrary incoming energy and energy density for $N = 24$. Even in this case, we have not been able to solve the equation (\[eq:twoone\]) for arbitrary $T_{++}$. However to understand the nature of threshold of instability and black hole formation we only need to have a situation where we can tune the energy density and the total energy independently. We will thus solve the above equation for an incoming matter wave which has a constant energy density for a finite interval of retarded time. Thus we have a matter energy momentum tensor of the form T\^M\_[++]{}=[2 ()\^2]{}\[(- 1 + )-(- 1 - )\] \[eq:twotwo\] which corresponds to a constant energy density ${\alpha \over 2}$ in the interval $1-\ep < \xp < 1 + \ep$. The total energy is thus $M = \alpha\ep$.
In the following we will name the various regions of the $\xp$ space as follows (i) Region I : $\xp < 1 - \ep$ (ii) Region II $1-\ep < \xp <
1 + \ep$ (iii) Region III $\xp > 1 + \ep$.
In Regions I and II the solution is given by () = \[eq:twothree\] where $d$ is an integration constant. In (\[eq:twothree\]) we have made the following definitions = \^2 + 2 = \[eq:twofour\]
We want to have initial conditions such that the space time is in the vacuum, and the boundary has a vacuum hyperbolic solution discussed above, before any matter has come in. This means that in region I we have to choose $d = 0$. This determines the value of the dilaton field on the boundary to be the quantity [@DAM1] e\^[-2]{} \_[boundary]{} = \[eq:twofoura\] Thus the value of the dilaton field is also determined in terms of the parameter $\beta$ as mentioned above.
The nature of the solution in region II is qualitatively different for small and large $\alpha$ regardless of the value of $\ep$. Let us define $\alpha_0 = {\Delta \over 2}$.
$\alpha < \alpha_0$
-------------------
For $\alpha < \alpha_0$ the solution in region II is given by = \[eq:twofive\] where we have defined = - 2 = \[eq:twosix\] The constant of integration $d_{II}$ is determined by requiring that $\qb$ is continuous across $\xp = 1- \ep$. This gives d\_[II]{} = (1-)\^ [- + ]{} \[eq:twoseven\] From the solution (\[eq:twofive\]) it is clear that $\qb$ and hence $\fb$ will become infinite and the boundary will run away if $(\xp)^{\eta}-d_{II}=0$ for some value of $\xp >
(1 - \ep)$. This requires $d_{II} > (1-\ep)^\eta$. However, using the definitions of $\Delta$ and $\Upsilon$ it is clear from (\[eq:twoseven\]) that $d_{II}$ is always less than $(1-\ep)^\eta$. Thus there cannot be any runaway of the boundary in the region II.
The solution in region III is given by (\[eq:twothree\]) with the integration constant denoted by $d_{III}$. This is in turn determined by the solution in region II by matching across $\xp = 1+ \ep$, which yields the following equation for $d_{III}$ = [(1+)\^+ d\_[II]{} (1+)\^- d\_[II]{}]{} \[eq:twoeight\] Since $d_{II} < (1-\ep)^\eta < (1+\ep)^\eta$ it is clear from (\[eq:twoeight\]) that $d_{III} < (1+\ep)^\delta$. The solution (\[eq:twothree\]) then shows that $\qb$ cannot blow up anywhere in region III as well.
We thus conclude that for $\alpha < {\Delta \over 2}$ the boundary is stable and remains timelike throughout.
$\alpha > \alpha_0$
-------------------
For $\alpha > \alpha_0$ the solution in Region II is given by = \[eq:twonine\] where we have defined = 2- = [2]{} \[eq:thirty\] The intgration constant $c_{II}$ is once again determined by matching the solution with the vacuum solution at $\xp = 1-\ep$. This gives c\_[II]{} = [ + - 1]{} \[eq:thone\] where we have defined = \[eq:thonea\] The solution in region III is still given by (\[eq:twothree\]) which we rewrite as () = \[eq:twothreea\] $c_{III}$ is determined by matching across $\xp = 1+ \ep$ which yields c\_[III]{} = (1+)\^ [(1+\^2)(1-\^2) - 2]{} \[eq:thtwo\]
Let us first look for runaways of the boundary in the region III. The solution shows that this can happen if the value of $c_{III}$ is such that the denominator in (\[eq:twothreea\]) vanishes for a value of $\xp > (1 + \ep)$. This means that one must have $c_{III} > (1+\ep)^\delta$. The solution for $c_{III}$ then shows that this would require ,for $\gamma \geq 1$, -[2\^2 - 1]{} < < - [1 ]{} \[eq:thtwoa\] whereas for $\gamma < 1$ one must have either > [2 1- \^2]{} \[eq:thtwob\] or < -[1 ]{} \[eq:thtwoc\]
It may be easily checked that in the limiting case of a shock wave this reproduces the limit on the total energy derived in [@DAM1]. This limit corresponds to $\alpha \rightarrow \infty$ and $\ep \rightarrow 0$ with the total mass $M = \alpha\ep$ kept fixed. Then (35) reduces to the condition $M > \beta\sqd$.
For finite values of $\alpha$ and $\ep$ one has a more complex behaviour, since the various integration constants are periodic functions of $\alpha$ and $\epsilon$. Let us study the behaviour of the boundary for some given value of $\alpha > \alpha_0$ and $\beta$ and the extent of the pulse $\ep$ increasing from zero. It follows from (\[eq:thtwoa\])-(\[eq:thtwoc\]) that the conditions for runaway in region III are satisfied for $\ep$ lying in the range $\ep^{(1)}_n < \ep < \ep^{(2)}_n$ where \^[(1)]{}\_n & = &[tanh]{} { [1 2]{}\[\_n + \^[-1]{}([21-\^2]{})\] }\
\^[(2)]{}\_n & = & [tanh]{} { [1 2]{}\[n- \^[-1]{}([1 ]{})\] } \[eq:thtwod\] where $n = 1,2, \cdots$ and \_n = (n-1) (1 - ) + n (- 1) In the above equation the branch of the inverse tangent function has been chosen to be the interval $[-{\pi \over 2}, {\pi \over 2}]$. It is easy to check that $\aep_{n} < \bep_n < \aep_{n+1}$
The behaviour of the boundary may be now summarized as follows. For $\ep < \aep_1$, $\qb$ is finite and the boundary is stable. For $\aep_1 < \ep < \bep_1$ the boundary runs away in region III. For $\bep_1 < \ep < \aep_2$ the quantity $\qb$ diverges in region II and there is no divergence in region III. When $\aep_2 < \ep < \bep_2$ a new divergence of $\qb$ appears in region III, and the pattern continues.
In the physical problem a divergence of $\qb$ results in a runaway of the boundary. Just as in the case of the shock wave analyzed in [@DAM1] a singularity develops which is asymptotic to the boundary where it diverges and there is a resulting event horizon. Thus for the physical situation at hand, the critical value of $\ep$ for which the boundary starts running away and a black hole is formed is the lowest value of $\ep$ for which any of the conditions (\[eq:thtwoa\])-(\[eq:thtwob\]) are satisfied, i.e. $\ep_{cr} = \aep_1$. When $\ep$ exceeds a higher value $\bep_1$ the runaway occurs inside the region of the incident pulse. The other runaways for higher values of $\ep$ are irrelevant because they occur in a region which is already behind the singularity and hence not contained in the semiclassical space-time. Profiles of the boundary for fixed value of $\alpha$ and different $\ep$ (and hence different mass $M$) are shown in the Fig.1 - 3.
Note that we have kept $\alpha,\beta$ fixed in the above discussion. It may be seen from (\[eq:thtwod\]) that $\ep_{cr}$ decreases monotonically from $1$ at $\alpha = \alpha_0$ to zero for infinite $\alpha$, with $\ep_{cr} \rightarrow {\beta\sqd \over \alpha}$ as $\alpha \rightarrow \infty$. Thus a limit of $\ep$ for some given $\alpha$ means that there is a lower limit on the total mass of matter $M_{cr}(\alpha) = \alpha \ep_{cr}$ which has been thrown in. For $\alpha = \alpha_0$ one has $M_{cr} = \alpha_0$. As $\alpha$ increases, $M_{cr}(\alpha)$ first rises to a maximum value and then decreases, eventually becoming asympotic to the constant shock wave value $\beta\sqd$ for large $\alpha$.
We thus conclude that for the boundary to run away and a black hole to form there is a minimum value of the incoming energy density [*as well*]{} as a minimum value of the total energy.
Outlook
=======
The fact that the runaway of the boundary occurs at the same threshold at which a black hole singularity is formed in the semiclassical model is significant. While we have demonstrated this in the $N=24$ model, we expect that a similar behaviour would hold for $N \neq 24$ as well. However the differential euqtion in the latter case is higher order and the stability analysis would be significantly different. It would be interesting to get an idea of this by using numerical integration techniques. Some work in this direction has been done in [@STHOR].
For the shock wave solution we found in [@DAM1] that the runaway of the boundary in the semiclassical theory also signals the threshold beyond which an aypmtotic observer measures large quantum fluctuations of the boundary near the horizon. It would be interesting to understand the nature and implications of the boundary flcutuations in a more general setting. We leave that for future investigation.
.5cm [**Acknowledgements:**]{} We would like to thank H. Verlinde for discussions. One of us (S.M.) would like to thank Professor Abdus Salam, the International Atomic Energy and UNESCO for hospitality at the International Centre for Theoretical Physics, Trieste. S.R.D. would like to thank the Theoretical Physics Groups of Brown University, Princeton University, Enrico Fermi Institute and Washington University for hospitality during the final stages of this work.
[99]{}
S.R. Das and S. Mukherji, hep-th/9401102, to appear in , [**D50**]{}.
E. Verlinde and H. Verlinde, , [**B406**]{} (1993) 43; K. Schoutens, E. Verlinde and H. Verlinde, (1993) 2690.
T. Chung and H. Verlinde, , [**B418**]{} (1994) 305.
C.G. Callan and S.R. Das, , [**51**]{} (1983) 1155.
C.G. Callan and I. Klebanov, hep-th/9311092; I. Affleck and J. Sagi, hep-th/9311056; C.G. Callan, I. Klebanov, A.W.W. Ludwig and J.M. Maldacena, hep-th/9402113; J. Polchinski and L. Thorlacius, hep-th/9404008.
J.G. Russo, L. Susskind and L. Thorlacius, , [**D46**]{} (1992) 3444; , [**D47**]{} (1993) 533.
A. Strominger and L. Thorlacius, hep-th/9405084.
A. Strominger, hep-th/9405094.
A. Bilal and C.G. Callan, , [**B394**]{} (1993) 73.
S. de Alwis, , [**B289**]{} 278; , [**B300**]{} (1993) 330; 5429.
FIGURES
Fig.1: Typical boundary curve for $\alpha < \alpha_0$ for any $\epsilon$. Even for $\alpha > \alpha_0$ and $M < M_{cr}(\alpha)$ the boundary curve is similar. No singularities are formed.
.5cm Fig. 2: Boundary curve for $\alpha > \alpha_0$ and $M
> M_{cr}(\alpha)$. Boundary runs away after the matter infall ends. Cosequently space-time singularity is formed. .5cm
Fig. 3: For yet higher value of $M$, i.e. for the corresponding $\ep > \ep_{cr}^{(2)}$, the boundary runaway occurs in the region of incident pulse.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the temperature-magnetic field phase diagram, the binary metallic compound MnSi exhibits three magnetic phases below $T_{\rm c} \approx 29$ K. An unconventional helicoidal phase is observed in zero field. At moderate field intensity a conical phase sets in. Near $T_{\rm c}$, in an intermediate field range, a skyrmion lattice phase appears. Here we show the magnetic structure in the conical phase to strongly depend on the field direction and to deviate substantially from a conventional conical structure.'
author:
- 'P. Dalmas de Réotier'
- 'A. Maisuradze'
- 'A. Yaouanc'
- 'B. Roessli'
- 'A. Amato'
- 'D. Andreica'
- 'G. Lapertot'
bibliography:
- 'reference.bib'
title: Unconventional magnetic order in the conical state of MnSi
---
[*Introduction*]{} — Since their discovery in the cubic binary metallic compound MnSi (cubic space group P2$_1$3) in 2009 [@Muhlbauer09a], the interest in magnetic skyrmions has been tremendous, mainly explained by their potential applications as information carriers for future magnetic memories [@Fert13; @Nagaosa13]. In bulk material a skyrmion lattice phase is observed when an external magnetic field ${\bf B}_{\rm ext}$ is applied to the system. This phase is surrounded by a conical magnetic phase in which the magnetic moments are believed to have a component parallel to ${\bf B}_{\rm ext}$ and a transverse component akin to the conventional helical structure assigned to the zero-field phase. Interestingly, a skyrmion lattice is described as a coherent superposition of three such helices. The helical structure is theoretically interpreted by a competition between a dominant ferromagnetic exchange interaction and a weaker Dzyaloshinski-Moriya interaction authorized by the absence of inversion symmetry in the space group P2$_1$3 [@Bak80; @Nakanishi80].
Recent results show the magnetic structure of chiral magnets in zero field to be only partly understood. While the magnetic moments in planes perpendicular to the magnetic propagation wavevector ${\bf k}$ are ferromagnetically aligned, their in-plane orientation is not merely given by the scalar product ${\bf k}\cdot{\bf r}$ where ${\bf r}$ defines the position of a magnetic site. An additional phase exists for some position, which was deduced from an analysis of muon spin rotation ($\mu$SR) data in the case of MnSi [@Dalmas16] for which ${\bf k}$ is parallel to the \[111\] crystal axis. In the case where ${\bf k}\parallel [001]$, the symmetry of the crystal also authorizes such a phase [@Yaouanc17].
In the context where the current theories accounting for the wealth of textures found in chiral magnets lead to the conventional helical and conical states, a deeper study of the conical phase is timely. Here we present a detailed refinement of the magnetic structure in the prototypal system MnSi for ${\bf B}_{\rm ext}$ applied along the \[111\] or \[001\] crystal axes. According to neutron diffraction, and for high enough fields, we always have ${\bf k} \parallel {\bf B}_{\rm ext}$. For the former direction, we find the additional phase between the moments to be amplified when compared to the zero-field case. For the \[001\] direction, not only may an additional phase be present but the moments of the helical component are found to rotate in planes not perpendicular to ${\bf B}_{\rm ext}$.
The experiments were carried with $\mu$SR; for an introduction to the method, see, e.g. Ref. . With a probe — the muon — sitting and measuring the magnetic field at interstitial sites, it is ideally suited for the determination of possible deviations relative to the conventional helices. Figure \[sensitivity\] illustrates the sensitivity of $\mu$SR to such small angular deviations for ${\bf B}_{\rm ext}\parallel [111]$.
[![(color online). Fourier amplitudes of MnSi $\mu$SR spectra simulated for three values of $\psi$ for ${\bf B}_{\rm ext} \parallel [111]$ with $B_{\rm ext}$ = 200 mT. The angle $\psi$ is defined in Table \[phase\]. The other parameters entering the simulations are those obtained from this report except $\lambda_X$ which was set to zero for the sake of clarity.[]{data-label="sensitivity"}](MnSi_conical_111_comparison.eps "fig:"){width="\linewidth"}]{}
The measurements were performed relatively close to the magnetic ordering temperature $T_{\rm c} \approx 29$ K, with field magnitudes chosen for a single magnetic domain to be present in the crystal [@Grigoriev06].
The Mn atoms in MnSi occupy $4a$ Wyckoff positions. The coordinates of the four positions, labeled by $\gamma \in \{{\rm I},{\rm II},{\rm III},{\rm IV} \}$, depend on a single parameter $x_{\rm Mn}$ = 0.138. They are $(x_{\rm Mn},x_{\rm Mn},x_{\rm Mn})$, $(\bar{x}_{\rm Mn}+\frac{1}{2},\bar{x}_{\rm Mn},x_{\rm Mn}+\frac{1}{2})$, $(\bar{x}_{\rm Mn},x_{\rm Mn}+\frac{1}{2},\bar{x}_{\rm Mn}+\frac{1}{2})$, and $(x_{\rm Mn}+\frac{1}{2},\bar{x}_{\rm Mn}+\frac{1}{2},\bar{x}_{\rm Mn})$. The lattice parameter is $a_{\rm lat}$ = 4.558 Å.
[*Possible magnetic structures in the conical phase*]{} — The conical phase is characterized by an incommensurate propagation wavevector $k \approx 0.36$ nm$^{-1}$ [@Muhlbauer09a] describing the helical component and a second vanishing wavevector responsible for the macroscopic magnetization. We are left with the determination of the magnetic structure compatible with symmetry for the first component.
We shall specify the position of a unit cell by the cubic lattice vector $ {\bf i }$ and that of a Mn atom within a cell by $ {\bf d}_\gamma$. For a magnetic moment at position $ {\bf i + d_\gamma}$ we write $$\begin{aligned}
{\bf m}_{i + d_\gamma} & = & {\bf m}_{\rm u} + m_{\rm h} \left ( \cos \alpha_{i, \gamma}\, {\bf a}_{d_\gamma} - \sin \alpha_{i, \gamma}\, {\bf b}_{d_\gamma} \right )
\label{moment_muon_general}\end{aligned}$$ setting $\alpha_{i, \gamma}$ = ${\bf k}\cdot ({\bf i}+{\bf d}_\gamma)$. Vector ${\bf m}_{\rm u}$ denotes the uniform component parallel to ${\bf B}_{\rm ext}$ and vectors ${\bf a}_{d_\gamma}$ and ${\bf b}_{d_\gamma}$ together with ${\bf n}_{d_{\gamma}} \equiv {\bf a}_{d_\gamma} \times {\bf b}_{d_\gamma}$ form a direct orthonormal basis. For ${\bf B}_{\rm ext} \parallel [111]$, the simple solution where ${\bf a}_{d_\gamma}$ and ${\bf b}_{d_\gamma}$ are orthogonal to \[111\] will be found sufficient, as in zero-field [@Dalmas16]. This is not the case for ${\bf B}_{\rm ext} \parallel [001]$. The Euler angles defining ${\bf a}_{d_\gamma}$, ${\bf b}_{d_\gamma}$, and ${\bf n}_{d_{\gamma}}$ are compiled in Table \[phase\] for each sublattice $\gamma$. As an example, the coordinates of ${\bf n}_{d_{\gamma}}$ are $\left( \cos\varphi_{d_{\gamma}}\sin\theta_{d_{\gamma}}, \sin\varphi_{d_{\gamma}}\sin\theta_{d_{\gamma}}, \cos\theta_{d_{\gamma}} \right )$.
----------------------- ------------ ------------ ------------- ------------ ------------ ------------- -------------- -------------
$\gamma$ ${\rm I}$ ${\rm II}$ ${\rm III}$ ${\rm IV}$ ${\rm I}$ ${\rm II}$ ${\rm III}$ ${\rm IV}$
$\varphi_{d_\gamma} $ 45$^\circ$ 45$^\circ$ 45$^\circ$ 45$^\circ$ $0$ 0 $\varphi_2$ $\varphi_2$
$\theta_{d_\gamma} $ $\theta_0$ $\theta_0$ $ \theta_0$ $\theta_0$ $\theta_1$ $-\theta_1$ $ -\theta_2$ $\theta_2$
$\psi_{d_\gamma} $ $0$ $\psi$ $\psi$ $\psi$ $0$ 0 $\psi_2$ $\psi_2$
----------------------- ------------ ------------ ------------- ------------ ------------ ------------- -------------- -------------
: Parameters for the description of the MnSi magnetic structure in the conical phase for ${\bf k} \parallel {\bf B}_{\rm ext} \parallel [111]$ or $[001]$. The results are obtained from representation analysis; see Refs. [@Dalmas16; @Yaouanc17]. The table gives the Euler angles $\varphi_{d_\gamma}$, $\theta_{d_\gamma} $, and $\psi_{d_\gamma} $ characterizing the $({\bf a}_{d_\gamma}, {\bf b}_{d_\gamma}, {\bf n}_{d_\gamma}$) basis in the crystal cubic axes. While the value of $\theta_0$ is fixed ($\cos^2\theta_0 = 1/3$; $\theta_0\approx 54.7^\circ$), the angles $\psi$, $\varphi_2$, $\theta_1$, $\theta_2$ and $\psi_2$ are free parameters of this study.
\[phase\]
[*The magnetic field at the muon sites*]{} — The crystallographic muon site in MnSi has been determined from earlier $\mu$SR measurements [@Amato14; @Dalmas16]. A density functional theory computation confirms this result [@Bonfa15]. The muon sitting at a $4a$ Wyckoff position, four different magnetic sites exist in the cubic unit cell. They are identified with the index $\eta\in\{1,2,3,4\}$. We will denote ${\bf r}_{0,s_\eta}$ the vector distance between a muon position $s_\eta$ and the origin of the cubic lattice.
The local magnetic field ${\bf B}_{{\rm loc},{s_\eta}}$ at position $s_\eta$ comprises ${\bf B}_{\rm ext}$ and the dipolar and contact fields associated with the Mn magnetic moments. Traditionally the dipolar field is split into three terms and accordingly [@Schenck95; @Dalmas97; @Kalvius01]: $$\begin{aligned}
{\bf B}_{{\rm loc},{s_\eta}} = {\bf B}_{\rm ext} + {\bf B}^\prime_{{\rm dip},{s_\eta}} + {\bf B}_{\rm Lor} +{\bf B}_{\rm dem} + {\bf B}_{{\rm con},{s_\eta}}.
\label{field_muon_1}\end{aligned}$$ Here ${\bf B}^\prime_{{\rm dip},{s_\eta}}$ results from the dipolar interaction between the muon magnetic moment and the localized Mn magnetic moments inside the Lorentz sphere and ${\bf B}_{{\rm con},{s_\eta}}$ is the contact field which originates from the polarized conduction electron density at the muon site. Finally, ${\bf B}_{\rm Lor}$ and ${\bf B}_{\rm dem}$ are the macroscopic Lorentz and demagnetization fields. As usual ${\bf B}_{\rm dem} = -N \mu_0 {\bf M}$, where $N$ ($ 0 \le N \le1$) is the demagnetization field factor, $\mu_0$ is the permeability of free space and ${\bf M}$ = $4\,{\bf m}_{\rm u}/a_{\rm lat}^3$ is the macroscopic magnetization.
Rather than working in direct space, it is convenient to proceed using the reciprocal space [@Dalmas16] with [@Yaouanc93; @Yaouanc93a; @Yaouanc11]: $$\begin{aligned}
& & {\bf B}^\prime_{{\rm dip},{s_\eta}} + {\bf B}_{\rm Lor} + {\bf B}_{{\rm con},{s_\eta}} \label{field_muon_3}\\
& = & \frac{\mu_0}{4\pi} \frac{1}{\sqrt{n_{\rm c}}v_{\rm c}}
\sum_\gamma \sum_{{\bf q} \in {\rm BZ}} \boldsymbol{J}_{d_\gamma,{\bf q}, s_\eta} {\bf m}_{d_\gamma, {\bf q}} \exp (-i {\bf q} \cdot {\bf r}_{0,s_\eta} ),\nonumber\end{aligned}$$ where ${\bf m}_{d_\gamma, {\bf q}}$ is the Fourier component of the sublattice magnetic moment. Here $n_{\rm c}$ is the number of unit cells in the crystal under study and $v_{\rm c}$ = $a_{\rm lat}^3$ is their volume. The sum in Eq. \[field\_muon\_3\] is performed over the ${\bf q}$ vectors of the first Brillouin zone (BZ). The unitless tensor ${\boldsymbol J}_{d_\gamma, {\bf q},s_\eta} = {\boldsymbol F}_{d_\gamma, {\bf q},s_\eta} + {\boldsymbol H}_{d_\gamma, {\bf q},s_\eta}$ is the sum of two unitless tensors: ${\boldsymbol H}_{d_\gamma, {\bf q},s_\eta}$ describes the contact interaction parametrized by the quantity $r_\mu H/4\pi$ [@Amato14; @Dalmas16], where $r_\mu$ is the number of nearest neighbors used to model the contact interaction, and ${\boldsymbol F}_{d_\gamma, {\bf q},s_\eta}$ is expressed as a function of a tensor ${\boldsymbol C}_{d_\gamma, {\bf q},s_\eta}$. In recognition of the different nature of the ferro- and antiferromagnetic components of the magnetization the relation between the latter two tensors takes two distinct, albeit related, forms. In terms of the tensor Cartesian components $\alpha\beta$, $F^{\alpha\beta}_{d_\gamma, {\bf q},s_\eta} = -4\pi\left[ \frac{q^\alpha q^\beta}{q^2} - C^{\alpha\beta}_{d_\gamma, {\bf q},s_\eta}\right]$ for ${\bf q} \ne 0$, and $F_{d_\gamma, {\bf q}=0,s_\eta} = 4\pi C^{\alpha\beta}_{d_\gamma, {\bf q}=0,s_\eta}$ [@Yaouanc11]. The ${\boldsymbol C}_{d_\gamma, {\bf q},s_\eta}$ tensor components are computed following the Ewald summation technique [@Ewald21; @Born54; @Yaouanc93; @Yaouanc93a]. This ensures a fast and exact evaluation of the lattice sum which otherwise converges slowly.
The dependence of ${\bf B}_{{\rm loc},{s_\eta}}$ on the magnetic structure is solely described by ${\bf m}_{d_\gamma, {\bf q}}$: $$\begin{aligned}
{\bf m}_{d_\gamma, {\bf q}} & = & \sqrt{n_{\rm c}} \left(
\delta_{{\bf q}, {\bf k} }\,{\tilde {\bf m}}_{d_\gamma,+} +
\delta_{{\bf q}, -{\bf k} }\,{\tilde {\bf m}}_{d_\gamma,-} +
\delta_{{\bf q}, {\bf 0} }\, {\bf m}_{\rm u}
\right),\cr
& &
\label{Moment_Fourier_component_1}\end{aligned}$$ with $$\begin{aligned}
{\tilde {\bf m}}_{d_\gamma, \pm} = \frac{m_{\rm h}}{2} \left ( {\bf a}_{d_\gamma} \pm i {\bf b}_{d_\gamma} \right ).
\label{Moment_Fourier_component_2}\end{aligned}$$ This implies that ${\bf m}_{d_\gamma, {\bf q}}$ vanishes unless ${\bf q} = \pm {\bf k}$ or ${\bf q} = 0$.
Until now we have considered the muon to probe a unique mean magnetic field. However in a typical $\mu$SR experiment millions of muons are implanted in the specimen under study, which localize in different unit cells of the crystal. Therefore the different muons probe the incommensurate magnetic structure for different phases and accordingly we must consider a distribution of local fields rather than a single field. The following field vector distribution is relevant [@Dalmas16]: $$\begin{aligned}
D_{\rm v}({\bf B}) & = & \int_{0}^{2\pi} \delta\left[ {\bf B}
- \boldsymbol{\mathfrak{B}}_{{\bf q}=0,{\rm s}_\eta}(0)
- \boldsymbol{\mathfrak{B}}_{{\bf k},{\rm s}_\eta}(-{\bf k}\cdot {\bf r}_{0,s_\eta}-\zeta) \right. \cr
& & - \boldsymbol{\mathfrak{B}}_{{\bf -k},{\rm s}_\eta}({\bf k}\cdot {\bf r}_{0,s_\eta}+\zeta) - \left. {\bf B}_{\rm ext} - {\bf B}_{\rm dem} \right] \,{\rm d}\zeta,
\label{distribution_zeta}\end{aligned}$$ where $\zeta$ is the magnetic structure phase just mentioned. From Eq. \[field\_muon\_3\] we derive $\boldsymbol{\mathfrak{B}}_{{\bf q},s_\eta}(\Psi) = {\mu_0 \over 4 \pi} \frac{1}{\sqrt{n_{\rm c}} v_{\rm c}} \sum_{\gamma} {\boldsymbol J}_{d_\gamma, {\bf q},{\rm s}_\eta} {\bf m}_{{d_\gamma}, {\bf q}} \exp (i \Psi )$.
[*The transverse-field (TF) polarization function*]{} — The measurements were performed with a standard TF setup [@Yaouanc11] in which the initial muon polarization has a component perpendicular to ${\bf B}_{\rm ext}$ [^1]. Conventionally the direction of ${\bf B}_{\rm ext}$ defines the Cartesian $Z$ axis of the spectrometer. The positrons resulting from the anisotropic decay of the muons are detected in counters set along the $X$ axis, perpendicular to $Z$. We denote as $\theta_\mu$ and $\varphi_\mu$ the polar and azimuthal angles of the muon spin at the instant of implantation. Due to the in-flight precession of the muon spins in ${\bf B}_{\rm ext}$ prior to their implantation, $\varphi_\mu$ depends on $B_{\rm ext}$. The evolution of the projection $S_X(t)$ of the muon spin in the local field is given by the solution of the Larmor equation [^2]. The polarization function associated with muons stopped at position $s_\eta$ is obtained after averaging over $D_{\rm v}({\bf B})$: $$\begin{aligned}
P_{X,s_\eta}(t) = \int \frac{S_X(t)}{S} D_{\rm v}({\bf B})\, {\rm d}^3{\bf B}.
\label{Polarization_th_2}\end{aligned}$$ Here $P_{X,s_\eta}(t)$ is written in the spectrometer reference frame, while in the previous sections the magnetic structure was expressed in the crystal reference frame. Obviously, proper geometrical transformations of ${\bf B}$ in the spectrometer frame are required before evaluating Eq. \[Polarization\_th\_2\].
We still need to include the effect of three physical phenomena. We shall do it phenomenologically. As usual, we account for the longitudinal dynamical relaxation with an extra $\exp(-\lambda_Z t)$ factor in the non oscillating components of $S_X(t)/S$. The oscillations are damped by the field distribution produced by the nuclear spins of the compound. Assuming a Gaussian field distribution with a root-mean-square $\Delta_{\rm N}$, we include an $\exp(-\gamma_\mu^2\Delta_{\rm N}^2 t^2/2)$ factor to the oscillating terms in $S_X(t)/S$. Two damping sources of electronic origin are also possible. The first arises from the imperfection of the magnetic structure [@Dalmas16]. However, since our data are recorded relatively close to $T_{\rm c}$, we only need to account for the second overwhelming damping induced by magnetic fluctuations. This is achieved by including an $\exp(-\lambda_X t)$ factor to each of the oscillating terms in $S_X(t)/S$.
We cannot distinguish the contribution of the four $s_\eta$ positions in the crystallographic unit cell of MnSi, hence the measured polarization results from the average $P_X(t) = \langle P_{X,s_\eta}(t) \rangle_{\eta}$.
[*Results*]{} — Before proceeding to the analysis of the TF $\mu$SR spectra, we expose a few experimental details.
The platelet-shaped samples were cut from crystals grown by Czochralski pulling and already used in previous measurements [@Yaouanc05; @Amato14; @Dalmas16]. The field was applied perpendicular to the platelets. With this geometry we expect $N \lesssim 1$. The measurements were performed at 28.14(1)K for ${\bf B}_{\rm ext} \parallel [111]$, and at 27.70(1) and 27.85(1)K for ${\bf B}_{\rm ext} \parallel [001]$. For each of them, the field was systematically applied at 35 K and the sample subsequently cooled down to the desired temperature.
TF asymmetry spectra are displayed in Figs. \[Fig\_asymmetry\_111\] and \[Fig\_asymmetry\_001\], together with the Fourier transforms of the precessing component.
(255,145) (-5,0) [![(color online). (a) TF asymmetry spectra recorded at 28.14(1) K in the conical phase of MnSi for different values of $B_{\rm ext}$ with ${\bf B}_{\rm ext} \parallel [111]$. Circles are experimental data while solid lines represent fits as explained in the main text. (b) Real part of the Fourier transforms of the asymmetry spectra precessing component. The solid lines derive from the fits of the asymmetry spectra, i.e. we did not fit the Fourier transforms. In line with the large sample size, the contribution of muons stopped in the sample surrounding which can be distinguished for the 180 mT data is negligible. In both panels, the data for consecutive fields are vertically shifted for better visualization.[]{data-label="Fig_asymmetry_111"}](MnSi_conical_111_asymmetry.eps "fig:"){width="0.50\linewidth"}]{} (0,5)[(a)]{} (123,0) [![(color online). (a) TF asymmetry spectra recorded at 28.14(1) K in the conical phase of MnSi for different values of $B_{\rm ext}$ with ${\bf B}_{\rm ext} \parallel [111]$. Circles are experimental data while solid lines represent fits as explained in the main text. (b) Real part of the Fourier transforms of the asymmetry spectra precessing component. The solid lines derive from the fits of the asymmetry spectra, i.e. we did not fit the Fourier transforms. In line with the large sample size, the contribution of muons stopped in the sample surrounding which can be distinguished for the 180 mT data is negligible. In both panels, the data for consecutive fields are vertically shifted for better visualization.[]{data-label="Fig_asymmetry_111"}](MnSi_conical_111_Fourier.eps "fig:"){width="0.50\linewidth"}]{} (128,5)[(b)]{}
(255,145) (-5,0) [![(color online). Same caption as for Fig. \[Fig\_asymmetry\_111\], but here the data concern measurements recorded for $T = 27.70\,(1)$ K and ${\bf B}_{\rm ext} \parallel [001]$. The contribution of muons implanted in the sample surroundings is more important than for ${\bf B}_{\rm ext} \parallel [111]$ owing to the smaller sample size.[]{data-label="Fig_asymmetry_001"}](MnSi_conical_001_asymmetry.eps "fig:"){width="0.50\linewidth"}]{} (0,5)[(a)]{} (123,0) (128,5)[(b)]{}
For the fits, the following parameters entering in the computation of $P_X(t)$ were fixed to values obtained in previous works [@Amato14; @Dalmas16]: the muon $4a$ position parameter $x_{\mu^+} = 0.532$, $r_\mu H/ 4 \pi = -1.04$, and $\Delta_{\rm N} = 1.11$ mT. The initial asymmetry $a_0$, a parameter strongly correlated to $\theta_\mu$, was fixed to its theoretical value 0.28. In a first instance the model was fit to individual spectra: some free parameters varied with $B_{\rm ext}$ and others were independent of it. Among the latter parameters, for ${\bf B}_{\rm ext} \parallel [001]$, are $\theta_1$, $\theta_2$, and $\psi_2$. Moreover we also found $\theta_1 \approx -\theta_2$ and $\psi_2\approx 0$ within error bars. In a second step, new fits were performed with field-independent parameters common to all spectra recorded for a given ${\bf B}_{\rm ext}$ direction. The results are shown as solid lines in Figs. \[Fig\_asymmetry\_111\] and \[Fig\_asymmetry\_001\]. Figure \[parameters\_B\](a) and Table \[parameters\] display the parameters.
(255,186) (-5,0) [![(color online). (a) Parameters characterizing the conical phase of MnSi for ${\bf B}_{\rm ext} \parallel [001]$: $\varphi_\mu$, $\theta \equiv \arctan (m_{\rm h}/m_{\rm u})$, and $m_0 \equiv \left(m^2_{\rm u} + m^2_{\rm h}\right)^{1/2}$. The horizontal scale is the internal field $B_{\rm int} \equiv B_{\rm ext} - \mu_0 N M$. The helical components at each of the sublattices is pictured in (b). The normal to the rotation plane is at angle $\pm \theta_1$ or $\pm\theta_2$ from the \[001\] direction.[]{data-label="parameters_B"}](MnSi_conical_parameters.eps "fig:"){width="0.50\linewidth"}]{} (0,5)[(a)]{} (120,40) [![(color online). (a) Parameters characterizing the conical phase of MnSi for ${\bf B}_{\rm ext} \parallel [001]$: $\varphi_\mu$, $\theta \equiv \arctan (m_{\rm h}/m_{\rm u})$, and $m_0 \equiv \left(m^2_{\rm u} + m^2_{\rm h}\right)^{1/2}$. The horizontal scale is the internal field $B_{\rm int} \equiv B_{\rm ext} - \mu_0 N M$. The helical components at each of the sublattices is pictured in (b). The normal to the rotation plane is at angle $\pm \theta_1$ or $\pm\theta_2$ from the \[001\] direction.[]{data-label="parameters_B"}](MnSi_conical_001_magnetic_structure_schematics.ps "fig:"){width="0.50\linewidth"}]{} (125,30)[(b)]{}
--------------------- ---------- ----------------- ----------------- ------------- ------------------------------ --
${\bf B}_{\rm ext}$ $T$ $\lambda_Z$ $\lambda_X$ $\psi$ $\theta_1 \approx -\theta_2$
(K) ($\mu$s$^{-1}$) ($\mu$s$^{-1}$) ($^\circ$) ($^\circ$)
$[111]$ 28.14(1) $\le 0.21$ $\approx 2.3$ $-$6.24(15) –
$[001]$ 27.70(1) $\le 0.1$ $\approx 3.8$ – \[20, 35\]
$[001]$ 27.85(1) $\le 0.56$ $\approx 4.7$ – \[20, 35\]
--------------------- ---------- ----------------- ----------------- ------------- ------------------------------ --
: Model parameters obtained from refinements. While a value common to all field intensities is used for the angles, the relaxation rates slightly depend on $B_{\rm ext}$. Ranges for $\theta_1$ and $\theta_2$ are given. The demagnetization field factor is $N$ = 0.96.
\[parameters\]
As expected, $\varphi_\mu$ increases linearly with $B_{\rm ext}$; its magnitude is in accord with the spectrometer characteristics [^3]. This is an independent test of the validity of our analysis. Considering $\theta$ \[Fig. \[parameters\_B\](a)\], which essentially characterizes the opening angle of the conical structure, its decrease with $B_{\rm ext}$ for the two orientations is anticipated since at high field the magnetic structure tends to be collinear ferromagnetic, i.e. $m_{\rm h} \rightarrow 0$. Remarkably, $\theta$ decreases almost linearly with $B_{\rm ext}$. Interestingly, the magnetic moment is found smaller for ${\bf B}_{\rm ext} \parallel [001]$ than for \[111\] despite being measured at a lower temperature. The angle $\psi$ [^4] characterizing the shift between the two orbits for ${\bf B}_{\rm ext} \parallel [111]$ is about three times as large as at low temperature in zero field [@Dalmas16]. For ${\bf B}_{\rm ext} \parallel [001]$, several solutions with similar confidence parameter $\chi^2$ are found to fit the data equally well for $20^\circ \lesssim \theta_1 \lesssim 35^\circ$, always with $\theta_1\approx -\theta_2$. This substantial value shows that the helical component of the moment is not perpendicular to ${\bf B}_{\rm ext}$ \[Fig. \[parameters\_B\](b)\]. This is the main result for this field orientation. The phase shift $\varphi_2$ resulting from the fit is correlated to $\theta_1$ and $\theta_2$ and can be as large as $\approx 20^\circ$.
[*Discussion*]{} — While the equation for a magnetic moment, i.e. Eq. \[moment\_muon\_general\], is generic, the angles are widely different between the two field orientations. The difference in the moment values is also noticeable. Both differences reflect the magnetic anisotropy of the system. Relevant to this discussion, we previously suggested an additional term to the symmetric and antisymmetric ferromagnetic exchanges and weak anisotropic exchanges to contribute to the microscopic spin Hamiltonian [@Dalmas16]. Thinking in terms of free energy, it would be worthwhile to investigate the effect of anisotropy energy as also proposed for the skyrmion lattice [@Everschor11]. The two types of deviations from the regular conical structure consistent with representation analysis which are required for the interpretation of experimental data were also derived in a microscopic level theoretical study [@Chizhikov12; @Chizhikov13]. For ${\bf B}_{\rm ext}\parallel [001]$, the magnetic moment magnitude depends slightly on the Mn position since the (${\bf a}_{d_\gamma}, {\bf b}_{d_\gamma}$) plane is not normal to \[001\]. An alternative fit was performed replacing Eq. \[moment\_muon\_general\] with ${\bf m}_{i + d_\gamma} = m_{\rm M} {\bf n}_{\rm d_\gamma} + m_{\rm h}\left ( \cos \alpha_{i, \gamma}\,{\bf a}_{d_\gamma} - \sin \alpha_{i, \gamma} {\bf b}_{d_\gamma} \right )$, i.e. with a conserved moment. The fit quality is equivalent to that of Fig. \[Fig\_asymmetry\_001\] and the fit parameters are similar to those of Fig. \[parameters\_B\](a) and Table \[parameters\]. Therefore we cannot definitively decide between these two related models.
[*Conclusions*]{} — A $\mu$SR study of the conical phase in bulk MnSi is reported. A quantitative analysis is performed using symmetry analysis. For ${\bf B}_{\rm ext}\parallel [111]$, the phase of the helical component is not solely given by the scalar product ${\bf k}\cdot{\bf r}$. The deviation is enhanced compared to the zero-field case. For ${\bf B}_{\rm ext}\parallel [001]$ the magnetic components associated with the helix rotate in planes which are not perpendicular to ${\bf B}_{\rm ext}$. This information should be helpful for the determination of the microscopic magnetic Hamiltonian. This work suggests the use of $\mu$SR as a three-dimensional microscopic magnetometry tool, in particular for the helimagnets. Applying this method to the characterization of the skyrmion lattice in MnSi is obviously of great interest.
We acknowledge discussions with V.P. Mineev. This research project has been partially supported by the European Commission under the 7th Framework Programme through the ‘Research Infrastructures’ action of the ‘Capacities’ Programme, NMI3-II Grant number 283883. Part of this work was performed at the GPS and Dolly spectrometers of the Swiss Muon Source (Paul Scherrer Institute, Villigen, Switzerland).
[^1]: At the Swiss Muon Source, as a trade-off between amplitude of spin rotation and counting rate, the angle $\theta_\mu$ between ${\bf B}_{\rm ext}$ and the initial muon beam polarization is lower than 90$^\circ$. While it is known to be around 40 to 50$^\circ$, its actual value varies from a setting to another due to hysteresis in the electromagnet. We take it as a free parameter.
[^2]: We have (see, e.g. Eqs. 3.10–13 of Ref. ): $$\begin{aligned}
& & S_X(t)/S & \cr
& = & \left\{ \frac{B_X^2}{B^2}+\left(1-\frac{B_X^2}{B^2}\right)\cos(\omega_\mu t)\right\} \cos \varphi_\mu\sin\theta_\mu\cr
& + & \left\{ \frac{B_X B_Y}{B^2}\left[1-\cos(\omega_\mu t) \right] + \frac{B_Z}{B}\sin(\omega_\mu t)\right\}\sin\varphi_\mu\sin\theta_\mu\cr
& + & \left\{ \frac{B_X B_Z}{B^2}\left[1-\cos(\omega_\mu t) \right] - \frac{B_Y}{B}\sin(\omega_\mu t)\right\}\cos\theta_\mu.
%\nonumber
%\label{P_X_generalized}\end{aligned}$$ Here $\omega_\mu = \gamma_\mu B$ where $\gamma_\mu= 851.616$ Mrads$^{-1}$T$^{-1}$ is the muon gyromagnetic ratio and ${\bf B}$ is the local magnetic field at the muon site.
[^3]: See the user manual of the GPS spectrometer at https://www.psi.ch/smus/gps.
[^4]: In Ref. [@Dalmas16], the angular shift between the two orbits was denoted $\phi$. Here it is denoted $\psi$ for consistency with the notation used for ${\bf B}_{\rm ext} \parallel [001]$. $\phi$ and $\psi$ have the same physical significance.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the construction of exponential frames and Riesz sequences for a class of fractal measures on ${\mathbb{R}}^d$ generated by infinite convolution of discrete measures using the idea of frame towers and Riesz-sequence towers. The exactness and overcompleteness of the constructed exponential frame or Riesz sequence is completely classified in terms of the cardinality at each level of the tower. Using a version of the solution of the Kadison-Singer problem, known as the $R_{\epsilon}$-conjecture, we show that all these measures contain exponential Riesz sequences of infinite cardinality. Furthermore, when the measure is the middle-third Cantor measure, or more generally for self-similar measures with no-overlap condition, there are always exponential Riesz sequences of maximal possible Beurling dimension.'
address:
- |
\[Dorin Ervin Dutkay\] University of Central Florida\
Department of Mathematics\
4000 Central Florida Blvd.\
P.O. Box 161364\
Orlando, FL 32816-1364\
U.S.A.\
- '\[Shahram Emami\] Department of Mathematics, San Francisco State University, 1600 Holloway Avenue, San Francisco, CA 94132.'
- '\[Chun-Kit Lai\] Department of Mathematics, San Francisco State University, 1600 Holloway Avenue, San Francisco, CA 94132.'
author:
- Dorin Ervin Dutkay
- Shahram Emami
- 'Chun-Kit Lai'
bibliography:
- 'eframes.bib'
title: Existence and exactness of exponential Riesz sequences and frames for fractal measures
---
[^1]
Introduction
============
A finite Borel measure $\mu$ is called a [*spectral measure*]{} if there exists a set $\Lambda$ such that the family of exponential functions $E(\Lambda): = \{e^{2\pi i \langle\lambda, x\rangle} :\lambda\in\Lambda\}$ is an orthonormal basis for $L^2(\mu)$. A set $\Omega$ is called a [*spectral set*]{} if $\chi_{\Omega}dx$ is a spectral measure. If the family of exponentials $E(\Lambda)$ forms a frame/Riesz sequence for $L^2(\mu)$ (See Section 2 for the definition of frames and Riesz sequences), we say that the measure $\mu$ is frame-spectral/RS-spectral. If $E(\Lambda)$ is both a frame and a Riesz sequence, then $E(\Lambda)$ is a Riesz basis for $L^2(\mu)$ and $\mu$ is called [*Riesz-spectral*]{}.
The study of spectral measures was initiated in [@Fug74] motivated by analysis of commuting self-adjoint extensions of partial differential operators. Fuglede asked which subsets of the Euclidean space are spectral sets and proposed his famous conjecture which states that these sets are precisely those that tile the Euclidean space by translations. In [@JP98], Jorgensen and Pedersen widened the scope of Fuglede’s question and asked which Borel measures on ${\mathbb{R}}^d$ admit orthogonal Fourier series. They constructed the first example of a [*singular, non-atomic*]{} spectral measure. It is based on a Cantor-type construction, where the unit interval is divided into four pieces and the second and fourth piece are discarded. Many more classes of examples of singular spectral measures have been constructed since, see, e.g, [@Str00; @LaWa02; @DJ06; @DHL18 and the references therein]. Strichartz proved in [@MR2279556] that, in some cases, the Fourier series associated to such singular spectral measures have much better convergence properties than their classical counterparts, see also [@MR3152727].
In their original paper, Jorgensen and Pedersen also proved that the more familiar middle-third Cantor set, with the measure $\mu$ being the standard Hausdorff measure, is not a spectral measure, so it does not admit orthogonal bases of exponential functions. This motivated Strichartz [@Str00], to ask if this middle-third Cantor measure $\mu$ can be frame-spectral or even Riesz-spectral. Very little progress towards an answer for this question has been made since then. In [@MR2826404], some Bessel sequences of exponential functions were constructed with positive Beurling dimension for $\mu$. In [@LW17], the first examples of frame-spectral fractal measures with only finitely many mutually orthogonal exponential were constructed.
In this paper, we generalize the study in [@LW17] on ${\mathbb{R}}^d$ and consider the frame-spectrality and the RS-spectrality of the measures obtained as infinite convolutions of atomic measures, of the form $$\label{measure_conv}
\mu = \mu(\{R_j,B_j\}) = \delta_{{\bf R}_1^{-1}B_1}\ast\delta_{{\bf R}_2^{-1}B_2}\ast...\ast\delta_{{\bf R}_n^{-1}B_n}\ast....,$$ where ${\bf R}_j=R_jR_{j-1}\dots R_1$, with $R_i$ being some expanding matrices with integer entries in ${\mathbb{R}}^d$, $B_j$ are some finite sets of digits in ${\mathbb Z}^d$, and for a finite subset $A$ of ${\mathbb{R}}^d$, $$\delta_A=\frac{1}{\# A}\sum_{a\in A}\delta_a,$$ where $\delta_a$ is the Dirac measure at the point $a$. This class of measures contains self-affine measures defined by affine iterated function systems as well as the middle-third Cantor measure.
[**Main Result and organization of the paper.**]{} The main tool for constructing frames and Riesz sequences for our class of measures is based on the [*frame/Riesz sequence towers*]{} (Definition \[def2.1\]). Originally, the idea of the tower construction is due to Strichartz [@Str00], who considered [*compatible towers*]{} for constructing orthogonal exponential basis. Special cases of the tower constructions for frames were considered previously by the authors [@LW17; @DHL18].
In Section 2, we will present the most general setting for towers to generate frames and, the first examples of Riesz sequences. Basically, we will need to have a frame/Riesz sequence condition at each finite dimensional level $\{{\bf R}_j, B_j\}$ and then concatenate, or convolute these sets to obtain frames or Riesz sequences for the resulting measure. In Section 3, we show how these towers generate frames and Riesz sequences of exponential functions for the infinite convolution measure in (\[measure\_conv\]) (Theorem \[thm1.3\]). We notice that, similar to all previous results in literature, a tail-term estimate (See (\[eq1.5\])) is required for the infinite convolution measure to have a frame. However, no such estimate is required for exponential Riesz sequences.
In Section 4, we investigate the exactness and completeness of the resulted frame and Riesz sequences (Theorem \[theorem2.1\] and Theorem \[th4.4\]) generated by the frame/Riesz sequence towers. In brief, these towers show a rigid structure. Under the tail-estimate (\[eq1.5\]), we get a Riesz basis of exponential functions if and only if we have a square matrix of finite frames at all levels. In particular, this shows that all frame-spectral measures constructed by the third-named author and Wang in [@LW17] are indeed Riesz-spectral.
The recent solution of the Kadison-Singer problem by [@MSS] provides an elegant proof that it is possible to partition a highly redundant tight frame into two frames with roughly the same frame bounds. This has led to important advances in frame theory. A recent survey of the Kadison-Singer problems and its equivalent statements can be found in [@B18]. In Section 5, using one of the consequences of the Kadison-Singer theorem, we show that all infinite convolution measures (\[measure\_conv\]) admit Riesz sequences of exponentials of infinite cardinality (Theorem \[theorem\_Riesz\]). In the more particular case of self-similar measures, we show that there exist Riesz sequences of maximal Beurling dimension (Theorem \[th5.7\]).
Theorem \[th5.7\] tells us that the middle-third Cantor measure has an exponential Riesz sequence of Beurling dimension $\log2/\log3.$ All spectral self-similar measures we know admit a spectrum of maximal Beurling dimension. Our result here leads to some evidence that an exponential Riesz basis may exist for the middle-third Cantor measure and that would lead to a solution to Strichartz’s question. On the other hand, in contrast to the existence of exponential Riesz sequences of maximal Beurling dimension, it is also known that one can also construct exponential orthonormal basis of zero Beurling dimension for some spectral measures [@MR3055992]. This tells us that an exponential frame of the middle-third Cantor measure may exist even if it does not have maximal Beurling dimension.
Frame and Riesz sequence towers
===============================
Recall that a sequence of vectors $\{f_k\}_{k=1}^{\infty}$ is called a [*frame*]{} for a Hilbert space $H$ if there exists $C,D>0$ such that, for all $x\in H$, $$C \|x\|^2 \le \sum_{k=1}^{\infty} |\langle x,f_k\rangle|^2 \le D\|x\|^2.$$ A sequence of vectors $\{f_k\}_{k=1}^{\infty}$ is called a [*Riesz sequence*]{} for a Hilbert space $H$ if there exists $C,D>0$ such that for any finite scalar sequence $(a_k)$ (i.e., there exists $N$ such that $a_k = 0$ for all $k>N$), $$C \sum_{k\le N} |a_{k}|^2\le \left\|\sum_{k\le N} a_kf_k\right\|^2 \le D\sum_{k\le N} |a_{k}|^2.$$ (see [@Chr03] for a comprehensive theory of frames and Riesz sequences). Our goal is to build a frame/Riesz sequence for Cantor-type fractal measures defined by rescaling. This section will be devoted to studying the finite dimensional preparation of such construction. A matrix $R$ is called [*expanding*]{} if all of its eigenvalues have moduli strictly greater than 1. Throughout the paper, $A^{\mathtt T}$ denote the transpose of $A$.
Finite dimensional preliminaries
--------------------------------
\[def2.1\] Let $R$ be an $d\times d$ expanding matrix of integer entries and let $B,L$ be a finite subset of ${\mathbb Z}^d$ and $0\in B\cap L$ (by a simple translation, there is no loss of generality to assume this). Define the vector $${\bf e}_{R,\lambda} = \frac{1}{\sqrt{\#B}}\left( e^{2\pi i \langle R^{-1}b,\lambda\rangle}\right)_{b\in B}^{\mathtt T} \in {\mathbb C}^{\#B}$$ We say that $(R,B,L)$ forms a [*frame triple with bounds $C\le D$*]{} if $$C\|{\bf x}\|^2 \le \sum_{\lambda\in L} |\langle {\bf x},{\bf e}_{R,\lambda}\rangle|^2\le D\|{\bf x}\|^2, \ \forall \ {\bf x}\in {\mathbb C}^{\#B}.$$ We say that $(R,B,L)$ forms a [*Riesz sequence triple with bounds $C\le D$*]{} if $$C\sum_{\lambda\in L}|a_{\lambda}|^2 \le \left\| \sum_{\lambda\in L} a_{\lambda} {\bf e}_{R,\lambda}\right\|^2\le D\sum_{\lambda\in L}|a_{\lambda}|^2, \ \forall (a_{\lambda})\in {\mathbb C}^{\#L}.$$
For $j=1,2,...$, let $0<C_j\le D_j<\infty$ be a sequence of positive numbers such that $\prod_{j=1}^{\infty}C_j>0$ and $\prod_{j=1}^{\infty}D_j<\infty$. Let $R_j$ be a sequence of expanding integer matrices on ${\mathbb R}^d$ and $B_j,L_j$ are finite subsets of ${\mathbb Z}^d$. We say that $\{(R_j,B_j,L_j): j=1,2,...\}$ forms a [*frame tower*]{} (respectively a [*Riesz sequence tower*]{}) with respect to the bounds $C_j,D_j$ if for each $j=1,2,...$, $(R_j,B_j,L_j)$ forms a frame triple (respectively a Riesz sequence triple) with bounds $C_j,D_j$.
We notice that if $C = D = 1$ and $\#B = \#L$, then $\{{\bf e}_{R,\lambda}: \lambda\in\Lambda\}$ forms an orthonormal basis on ${\mathbb C}^{\#B}$. In this case $(R,B,L)$ is called the [*Hadamard triple*]{}, which is known to be the key condition for generating exponential orthonormal basis of fractal measures [@Str00; @DHL18].
The following lemma establishes the duality relation between these two triples.
\[lemma1\] $(R,B,L)$ forms a frame triple with bounds $C,D$ if and only if $(R^{\mathtt T}, L,B)$ forms a Riesz sequence triple with bounds $\left(\frac{\#B}{\#L}C,\frac{\#B}{\#L}D\right)$.
As $\langle R^{-1}b,\lambda\rangle = \langle (R^{\mathtt T})^{-1}\lambda,b\rangle$, we have $$\sum_{\lambda\in L} |\langle {\bf x},{\bf e}_{R,\lambda}\rangle|^2 = \sum_{\lambda\in L} \left| \sum_{b\in B} x_b\frac{1}{\sqrt{\#B}} e^{-2\pi i \langle R^{-1}b,\lambda\rangle}\right|^2$$$$= \sum_{\lambda\in L} \left| \sum_{b\in B} x_b\frac{1}{\sqrt{\#B}} e^{-2\pi i \langle (R^{\mathtt T})^{-1}\lambda,b\rangle}\right|^2 = \frac{\#L}{\#B}\left\| \sum_{b\in B} x_b {\bf e}_{R^{\mathtt T},b}\right\|^2$$ The lemma follows from this.
Given finite set of integers $B,L\subset{\mathbb Z}^d$ and an integral expanding matrix $R$, we define the $(\#L) \times (\#B)$ matrix $${\mathcal F}_{L,B} = \frac{1}{\sqrt{\#B}} \left( e^{2\pi i \langle R^{-1}b,\lambda\rangle}\right)_{\lambda \in L, b\in B}$$ (Rows are indexed by $L$ and columns are indexed by $B$). Then ${\bf e}_{R,\lambda}$, $\lambda\in L$ are all the row vectors in ${\mathcal F}_{L,B}$. With some simple linear algebra, we have the following proposition. Note that this proposition is well-known if $(R,B,L)$ forms a Hadamard triple [@LaWa02].
\[Prop\_distinct\] Suppose that $(R,B,L)$ forms a frame triple. Then each element in $B$ must be a distinct representative in ${\mathbb Z}^d/R({\mathbb Z}^d)$.
Since $\{{\bf e}_{R,\lambda}:\lambda\in\Lambda\}$ forms a frame for ${\mathbb C}^{\#B}$ if $(R,B,L)$ forms a frame triple, we have that the vectors ${\bf e}_{R,\lambda}$ span ${\mathbb C}^{\#B}$. Hence, the rank of the matrix ${\mathcal F}_{L,B}=\#B = $ number of columns. However, if there exists $b,b'$ such that $b = b'+ Rk$ for some $k\in{\mathbb Z}^d$, then, for all $\lambda\in L$, $$e^{2\pi i \langle R^{-1}b,\lambda\rangle} = e^{2\pi i \langle R^{-1}(b'+Rk),\lambda\rangle} = e^{2\pi i \langle R^{-1}b',\lambda\rangle}$$ This means that ${\mathcal F}_{L,B}$ has two identical columns. Hence, the rank of ${\mathcal F}_{L,B}$ is strictly less than $\#B$, a contradiction. Hence, each element in $B$ must be a distinct representative in ${\mathbb Z}^d/R({\mathbb Z}^d)$, completing the proof.
Using Lemma \[lemma1\], it also follows easily that if $(R,B,L)$ forms a Riesz sequence triple, then each element in $L$ must be a distinct representative in ${\mathbb Z}^d/R^{\mathtt T}({\mathbb Z}^d)$. However, it is not true that $B$ is from distinct representative in ${\mathbb Z}^d/R({\mathbb Z}^d)$ if $(R,B,L)$ forms a Riesz sequence triple, as we will see in Example \[example013\]. Since we wish to construct frame-spectral measures from frame towers, we will from now on assume that [*the elements in $B$ are distinct representatives in ${\mathbb Z}^d/R({\mathbb Z}^d)$*]{}.
\[lemma\_tight\] Let $R$ be an $d\times d$ integral expanding matrix and let $B$ be a set containing some distinct representatives in ${\mathbb Z}^d/R({\mathbb Z}^d)$. Suppose that $\overline{L}$ is a complete set of distinct representatives of ${\mathbb Z}^d/R^{\mathtt T}({\mathbb Z}^d)$. Then $(R,B,\overline{L})$ forms a (tight) frame triple with constant $C = D = \frac{|\det (R)|}{\#B}$
Given an integer expanding matrix $R$, let $\overline{B}$ be a complete set of distinct representative of the group ${\mathbb Z}^d/R({\mathbb Z}^d)$. Then it is well-known that the matrix $$\frac{1}{\sqrt{|\det (R)|}} \left( e^{2\pi i \langle R^{-1}b,\lambda\rangle}\right)_{\lambda\in\overline{L},b\in\overline{B}}$$ is a unitary matrix. Hence, the rows form an orthonormal basis for ${\mathbb C}^{\#\overline{B}}$. Let $$\tilde{\bf e}_{R,\lambda} = \frac{1}{\sqrt{\#\overline{ B}}}\left( e^{2\pi i \langle R^{-1}b,\lambda\rangle}\right)_{b\in {\overline{B}}}^{\mathtt T} \in {\mathbb C}^{\#{\overline{B}}}$$ in ${\mathbb{C}}^{\#\overline{ B}}$. Then we have $$\sum_{\lambda\in\overline{L}} \left|\langle{\bf x},\tilde{\bf e}_{R,\lambda} \rangle \right|^2 = \frac{|\det (R)|}{\#B} \|{\bf x}\|^2.$$ Taking ${\bf x}$ to be a vector ${\bf w}\in{\mathbb C}^{\#B}$ and zero on entries located at $\overline{B}\setminus B$, we have that $$\sum_{\lambda\in\overline{L}} \left|\langle{\bf w},{\bf e}_{R,\lambda} \rangle \right|^2 = \frac{|\det (R)|}{\#B} \|{\bf w}\|^2$$ for all ${\bf w}\in{\mathbb C}^{\#B}$. Hence, $\{{\bf e}_{R,\lambda}:\lambda\in \overline{L}\}$ forms a unit norm tight frame on ${\mathbb C}^{\#B}$ with its tight frame constant $D = \frac{|\det (R)|}{\#B}$.
Concatenation of frame/Riesz sequence triples.
----------------------------------------------
Given frame/Riesz sequence towers, we can concatenate finitely many factors to form a larger frame/Riesz sequence triple. Define $${\bf R}_n = R_n...R_1$$ and let $$\label{eqB_n}
{\bf B}_n ={\bf R}_{n} \cdot \left\{\sum_{k=1}^{n}{\bf R}_k^{-1}b_k: b_k\in B_k \right\} = R_n...R_2(B_1)+R_n...R_3(B_2)+...+B_n.$$ $$\label{Lambda_n}
\Lambda_n = L_1+R_1^TL_2+...+(R_1^TR_2^T...R_{n-1}^T)L_n$$
\[proposition1.3\] With the notations above, the following statements hold:
1. Suppose that $\{(R_j,B_j,L_j): j=1,2,...\}$ forms a frame tower. Then $({\bf R}_n, {\bf B}_n,\Lambda_n)$ forms a frame triple with bounds $\prod_{j=1}^nC_j$, $\prod_{j=1}^n D_j$.
2. Suppose that $(R_j,B_j,L_j)$ forms a Riesz sequence tower. Then $({\bf R}_n, {\bf B}_n,\Lambda_n)$ forms a Riesz sequence triple with bounds $\prod_{j=1}^nC_j$, $\prod_{j=1}^n D_j$.
(i). We prove it by mathematical induction. When $n=1$, it is the frame triple for $(R_1,B_1,L_1)$, so the statement is true trivially. Assume now the inequality is true for $n-1$. Then we decompose ${\bf b}\in {\bf B}_n$ and $\lambda\in {\Lambda}_n$ by $${\bf b} = b_n+ R_n {\bf b}_{n-1}, \ \lambda = \lambda_{n-1}+ {\bf R}_{n-1}^{\mathtt T} l_n,$$ where $b_n\in B_{n}$, ${\bf b}_{n-1}\in{\bf B}_{n-1}$, $\lambda_{n-1}\in {\Lambda}_{n-1}$ and $l_n\in L_n$. Let also ${\bf M}_n = \prod_{j=1}^{n}(\#B_j)$ we have $$\begin{aligned}
&\sum_{\lambda\in{\Lambda}_n}\left|\sum_{{\bf b}\in{\bf B}_{n}}w_{\bf b}\frac{1}{\sqrt{{\bf M}_n}}e^{-2\pi i \langle{\bf R}_n^{-1}{\bf b},\lambda\rangle}\right|^2\\
=&\sum_{\lambda_{n-1}\in{\Lambda}_{n-1}}\sum_{l_n\in L_n}\left|\sum_{{\bf b}_{n-1}\in{\bf B}_{n-1}}\sum_{b_n\in B_n}\frac{1}{\sqrt{{\bf M}_{n}}}w_{R_n{\bf b}_{n-1}+b_n}e^{-2\pi i \left\langle {\bf R}_{n}^{-1}(b_n+ R_n {\bf b}_{n-1}),\lambda_{n-1}+ {\bf R}_{n-1}^{\mathtt T}l_n\right\rangle}\right|^2.\\
\end{aligned}$$ Note that $\langle{\bf R}_{n}^{-1} (R_n{\bf b}_{n-1}), {\bf R}_{n-1}^{\mathtt T}l_n\rangle = \langle{\bf b}_{n-1},l_n\rangle$ is always an integer, so the term above can be written as $$\sum_{\lambda_{n-1}\in{\Lambda}_{n-1}}\sum_{l_n\in L_n}\left|\sum_{b_n\in B_n}\frac{1}{\sqrt{\#B_n}}\left(\sum_{{\bf b}_{n-1}\in{\bf B}_{n-1}}\frac{1}{\sqrt{{\bf M}_{n-1}}}w_{R_n{\bf b}_{n-1}+b_n}e^{-2\pi i \langle {\bf R}_{n}^{-1}(b_n+ R_n {\bf b}_{n-1}),\lambda_{n-1}\rangle}\right)e^{-2\pi i \langle R_n^{-1}b_n,l_n\rangle}\right|^2$$ Using the frame triple assumption for $(R_n,B_n, L_n)$ and also the induction hypothesis, we further get $$\begin{aligned}
\leq&D_n\cdot\sum_{\lambda_{n-1}\in{\Lambda}_{n-1}}\sum_{b_n\in B_n}\left|\sum_{{\bf b}_{n-1}\in{\bf B}_{n-1}}\frac{1}{\sqrt{{\bf M}_{n-1}}}w_{R_n{\bf b}_{n-1}+b_n}e^{-2\pi i \langle {\bf R}_{n}^{-1}(b_n+ R_n {\bf b}_{n-1}),\lambda_{n-1}\rangle}\right|^2\\
=&D_n \cdot \sum_{b_n\in B_{n}}\sum_{\lambda_{n-1}\in{\Lambda}_{n-1}}\left|\sum_{{\bf b}_{n-1}\in{\bf B}_{n-1}}\frac{1}{\sqrt{{\bf M}_{n-1}}}w_{R_n{\bf b}_{n-1}+b_n}e^{-2\pi i \langle {\bf R}_{n-1}^{-1} {\bf b}_{n-1},\lambda_{n-1}\rangle}\right|^2\\
\leq &\prod_{j=1}^{n}D_j \cdot\sum_{b_n\in B_{n}}\sum_{{\bf b}_{n-1}\in{\bf B}_{n-1}}|w_{R_n{\bf b}_{n-1}+b_n}|^2\\
=&\prod_{j=1}^{n}D_j \cdot\|{\bf w}\|^2.\\
\end{aligned}$$ This completes the proof of the upper bound and the proof of the lower bound is analogous.
For (ii), by Lemma \[lemma1\], we note that $(R_j^{\mathtt T}, L_j,B_j)$ with $j=n,n-1,...,1$ (in reverse order) now forms a frame tower with frame bound $\left(\frac{\#B_j}{\#L_j}C,\frac{\#B_j}{\#L_j}D_j\right)$. Then we know that $(R_1^{\mathtt T}...R_n^{\mathtt T}, {\bf L}_n, \widetilde{\bf B}_n)$ forms a frame triple with bounds $\left(\prod_{j=1}^n\frac{\#B_j}{\#L_j}C_j,\prod_{j=1}^n\frac{\#B_j}{\#L_j}D_j\right)$, where $${\bf L_n} = R^{\mathtt T}_1...R^{\mathtt T}_{n-1}(L_n)+R^{\mathtt T}_1...R^{\mathtt T}_{n-2}(L_{n-1})+...+L_1 = \Lambda_n,$$ by replacing $R_k$ with $R_{n-k+1}^{\mathtt T}$ and $B_k$ with $L_{n-k+1}$ in (\[eqB\_n\]), and similarly $$\widetilde{{\bf B}}_n = B_n+...+R_n...R_2(B_1) = {\bf B}_n.$$
Hence, $(R_1^{\mathtt T}...R_n^{\mathtt T}, \Lambda_n, {\bf B}_n)$ forms a frame triple with bounds $\left(\prod_{j=1}^n\frac{\#B_j}{\#L_j}C_j,\prod_{j=1}^n\frac{\#B_j}{\#L_j}D_j\right)$. Using Lemma \[lemma1\] again, $({\bf R}_n, {\bf B}_n, \Lambda_n)$ forms a Riesz sequence triple.
Frame-spectral/RS-spectral Cantor measures
===========================================
In this section, we will use the frame/Riesz-sequence towers to generate Cantor measures with Fourier frames and Riesz sequences. Given a sequence of expanding matrices $R_n$ with integer entries and a finite collection of integer digit sets $B_j$, a natural probability measure is induced $$\label{mu}
\mu = \mu(\{R_j,B_j\}) = \delta_{{\bf R}_1^{-1}B_1}\ast\delta_{{\bf R}_2^{-1}B_2}\ast...\ast\delta_{{\bf R}_n^{-1}B_n}\ast....,$$ and we assume that the infinite convolution product is weakly convergent to a Borel probability measure. We let $$\mu_n = \delta_{{\bf R}_1^{-1}B_1}\ast\delta_{{\bf R}_2^{-1}B_2}\ast...\ast\delta_{{\bf R}_n^{-1}B_n}, \ \mu_{>n} = \delta_{{\bf R}_{n+1}^{-1}B_{n+1}}\ast\delta_{{\bf R}_{n+2}^{-1}B_{n+2}}\ast...$$ so that $\mu = \mu_n\ast\mu_{>n}$. Let also $$K_{n} = \left\{\sum_{k=n+1}^{\infty}{\bf R}_k^{-1}b_k: b_k\in B_k \right\}, {\bf B}_n = {\bf R}_n\left\{\sum_{k=1}^{n}{\bf R}_k^{-1}b_k: b_k\in B_k \right\}.$$ Hence, $K_0 = \bigcup_{{\bf b}\in {\bf B}_n}({\bf R}_n^{-1}{\bf b}_n+K_n)$ and $K_0, {\bf R}_n^{-1}{\bf B}_n, K_n$ are respectively the support of $\mu,\mu_n$ and $\mu_{>n}.$ We say that $\mu$ satisfies the [*no overlap condition*]{} if $$\mu (({\bf R}_n^{-1}{\bf b}_n+ K_n)\cap({\bf R}_n^{-1}{\bf b}_n'+K_n))=0, \ \mbox{for all} \ {\bf b}_n\neq{\bf b}_n'\in{\bf B}_n, \ \mbox{for all} \ n\in{\mathbb N}.$$ Let also $K_{{\bf b},n} = {\bf R}_n^{-1}{\bf b}+ K_n$ if ${\bf b}\in {\bf B}_n$.
Recall that the Fourier transform of a finite Borel measure $\mu$ on ${\mathbb{R}}^d$ is defined as $$\widehat\mu(y)=\int e^{-2\pi i{\left\langle y\, , \,x\right\rangle}}\,d\mu(x),\quad(x\in{\mathbb{R}}^d).$$
[*We will assume throughout the paper that the no-overlap condition holds*]{}. If ${\bf R}_j = N_j$ are integers in dimension one and $B_j$ are chosen from $\{0,1,...,N_j-1\}$, then $\mu$ are the Moran-type measure studied in [@LW17]. On the other hand, if all $R_j $ are the same matrix and $B_j$ is a subset of distinct representatives in the group ${\mathbb Z}^d/R({\mathbb Z}^d)$, then the resulting measure is a self-affine measure. They all satisfy the no-overlap condition. For the latter case, the no-overlap condition was proved in [@DHL18 Section 2].
\[lem3.2\] Under the no-overlap condition for the measure $\mu$, suppose that $f=\sum_{{\bf b}\in{\bf B}_n}w_{\bf b}{\bf 1}_{K_{{\bf b},n}}$. Then, $$\label{eq_FT_mu}
\int f(x)e^{-2\pi i{\left\langle \lambda\, , \,x\right\rangle}}\,d\mu(x)=\frac{1}{\#{\bf B}_n}\widehat{\mu_{>n}}(\lambda)\sum_{b\in{\bf B}_n}w_b e^{-2\pi i{\left\langle {\bf R}_n^{-1}b\, , \,\lambda\right\rangle}}.$$ $$\label{eq_norm}
\int|f|^2d\mu = \frac{1}{\#{\bf B}_n} \sum_{{\bf b}\in{\bf B}_n} |w_{\bf b}|^2.$$
We have $$\int f(x)e^{-2\pi i{\left\langle \lambda\, , \,x\right\rangle}}\,d\mu(x)=\sum_{{\bf b}\in{\bf B}_n}w_{\bf b}\int {\bf 1}_{K_{{\bf b},n}}(x)e^{-2\pi i {\left\langle \lambda\, , \,x\right\rangle}}\,d(\mu_n*\mu_{>n})(x)$$$$=\sum_{{\bf b}\in{\bf B}_n}w_b\int {\bf 1}_{{\bf R}_n^{-1}{\bf b}+K_{0,n}}(x+y)e^{-2\pi i {\left\langle \lambda\, , \,x+y\right\rangle}}\,d\mu_n(x)\,d\mu_{>n}(y).$$ Note that $\mu_{>n}$ is supported on $K_{0,n}$ and $\mu_n$ is supported on ${\bf R}_n^{-1}{\bf B}_n$, and, due to the non-overlap condition, $x$ has to be equal to ${\bf R}_n^{-1}b$ to get non-zero contribution. Thus, the quantity above is equal to $$=\sum_{{\bf b}\in{\bf B}_n}\frac{1}{\#{\bf B}_n}\int {\bf 1}_{{\bf R}_n^{-1}{\bf b}+K_{0,n}}({\bf R}_n^{-1}{\bf b}+y)e^{-2\pi i {\left\langle \lambda\, , \,{\bf R}_n^{-1}{\bf b}+y\right\rangle}}\,d\mu_{>n}(y)$$ $$=\sum_{{\bf b}\in{\bf B}_n}\frac{1}{\#{\bf B}_n}e^{-2\pi i {\left\langle \lambda\, , \,{\bf R}_n^{-1}{\bf b}\right\rangle}}\int e^{-2\pi i{\left\langle \lambda\, , \,y\right\rangle}}\,d\mu_{>n}(y)$$ $$=\frac{1}{\#{\bf B}_n}\widehat{\mu_{>n}}(\lambda)\sum_{{\bf b}\in{\bf B}_n}w_{\bf b} e^{-2\pi i{\left\langle {\bf R}_n^{-1}{\bf b}\, , \,\lambda\right\rangle}}.$$ (\[eq\_norm\]) follows from a standard computation.
Define $\Lambda_n$ as in and let $$\Lambda=\bigcup_{n=1}^\infty\Lambda_n.
\label{eqLambda}$$
\[thm1.3\]
1. Suppose that $(R_j,B_j,L_j)$ forms a frame tower with the associated measure $\mu = \mu(R_j,B_j)$ in (\[mu\]) and that the no-overlap condition is satisfied. Suppose furthermore that $$\label{eq1.5}
\delta(\Lambda) = \inf_{n\ge1}\inf_{\lambda\in\Lambda_n} |\widehat{\mu_{>n}}(\lambda)|^2 >0$$ Then $\{e^{2\pi i \langle \lambda,x\rangle}: \lambda\in\Lambda\}$ forms a frame for $L^2(\mu)$ with bounds $ \prod_{j=1}^{\infty}C_j$, $\prod_{j=1}^{\infty} D_j$.
2. Suppose that $(R_j,B_j,L_j)$ forms a Riesz sequence tower and that the associated measure satisfies the no-overlap condition. Then $\{e^{2\pi i \langle \lambda,x\rangle}: \lambda\in\Lambda\}$ forms a Riesz sequence for $L^2(\mu)$ with bounds $\prod_{j=1}^{\infty}C_j$, $\prod_{j=1}^{\infty} D_j$.
Let ${\mathcal S}_n= \{\sum_{{\bf b}\in {\bf B}_n}w_{\bf b} {\bf 1}_{K_{{\bf b},n}}: w_{\bf b}\in {\mathbb C}\}$ and ${\mathcal S} = \bigcup_{n=1}^{\infty} {\mathcal S}_n$. To prove (i), we notice that it suffices to check the frame inequality holds for every function $f\in {\mathcal S}$ as they are dense in $L^2(\mu)$. Since $0\in B_n$ for all $n$, the collection ${\mathcal S}_n$ is an increasing union. Given any $f\in {\mathcal S}_{n_0}$ and write it as $f = \sum_{{\bf b}\in {\bf B}_{n_0}}w_{\bf b} {\bf 1}_{K_{{\bf b},{n_0}}}$, for any $n\ge n_0$, using (\[eq\_FT\_mu\]) in Lemma \[lem3.2\] (as the no-overlap condition is satisfied), $$\sum_{\lambda\in \Lambda_n} \left|\int f(x)e^{-2\pi i \langle\lambda,x\rangle}d\mu(x)\right|^2 = \frac{1}{\#{\bf B}_n}\sum_{\lambda\in\Lambda_n} |\widehat{\mu_{>n}}(\lambda)|^2\left|\sum_{b\in{\bf B}_n}w_b e^{-2\pi i{\left\langle {\bf R}_n^{-1}b\, , \,\lambda\right\rangle}} \right|^2.$$ With the assumption that $\delta(\Lambda)>0$, using that $({\bf R}_n, {\bf B}_n, \Lambda)$ forms a frame triple (Proposition \[proposition1.3\]) and with (\[eq\_norm\]) in Lemma \[lem3.2\], we have $$\delta(\Lambda)\cdot\left(\prod_{j=1}^n C_j \right)\cdot \int|f|^2d\mu \le \sum_{\lambda\in \Lambda_n} \left|\int f(x)e^{-2\pi i \langle\lambda,x\rangle}d\mu(x)\right|^2\le \left(\prod_{j=1}^n D_j \right)\cdot \int|f|^2d\mu$$ Taking $n\rightarrow \infty$, we show that $\{e^{2\pi i \langle \lambda,x\rangle}: \lambda\in\Lambda\}$ forms a frame for $L^2(\mu)$ with a less sharp frame bound $\delta(\Lambda)\cdot\left(\prod_{j=1}^\infty C_j \right)$.
We now adopt the idea of the proof from [@MR3055992] and [@DHL18 Appendix] to show that $\delta(\Lambda)$ does not appear in the lower frame bound. It suffices to show that the lower bound of frame inequality holds for a dense set of functions in $L^2(\mu)$. We will check it for step functions in ${\mathcal S}$. Let $f = \sum_{{\bf b}\in {\bf B}_n}w_{\bf b}{\bf 1}_{K_{\bf b},n}\in {\mathcal S}_n$ and note that $f\in{\mathcal S}_m$, for all $m\ge n$; we use this to define the coefficients $w_{\bf b}$, for $b\in {\bf B}_m$, so $f$ can be written also as a function in $\mathcal S_m$ as $$f = \sum_{{\bf b}\in {\bf B}_m}w_{\bf b}{\bf 1}_{K_{\bf b},m}.$$ Define $$Q_{\infty} (f) = \sum_{\lambda\in\Lambda}\left|\int f(x)e^{-2\pi i \lambda x}d\mu(x)\right|^2 =\lim_{n\rightarrow\infty}Q_n(f),$$ where, with Lemma \[lem3.2\], $$Q_n(f): = \sum_{\lambda\in \Lambda_n}\left|\int f(x)e^{-2\pi i \lambda x}d\mu(x)\right|^2 = \frac{1}{{\bf M}_n}\sum_{\lambda\in\Lambda_n}|\widehat{\mu_{>n}}(\lambda)|^2\left|\sum_{{\bf b}\in B_n}w_{\bf b} \frac{1}{\sqrt{{\bf M}_n}}e^{-2\pi i \langle {\bf R}_n^{-1}{\bf b}, \lambda\rangle}\right|^2.$$ Let ${\mathbf C}_n = \prod_{j=1}^{n}C_j$, ${\mathbf D}_n = \prod_{j=1}^{n} D_j$ for $n=1,2...$ and $n=\infty$ and ${\mathbf M}_n = \prod_{j=1}^n (\#B_j)$. We are going to establish the lower bound. Note that $$\begin{aligned}
Q_m(f) =& Q_n(f)+ \sum_{\lambda\in\Lambda_m\setminus \Lambda_{n}}\left|\int f(x)e^{-2\pi i \langle\lambda, x\rangle}d\mu(x)\right|^2\\
=&Q_n(f)+\sum_{\lambda\in\Lambda_m\setminus \Lambda_{n}}\frac{1}{{\mathbf M}_m}|\widehat{\mu_{>m}}(\lambda)|^2\left|\sum_{{\bf b}\in {\bf B}_m}w_{\bf b} \frac{1}{\sqrt{\bf M}_m}e^{-2\pi i \langle {\bf R}_m^{-1}{\bf b}, \lambda\rangle}\right|^2.\\
\ge& Q_n(f)+\delta(\Lambda)\cdot\sum_{\lambda\in\Lambda_m\setminus \Lambda_{n}}\frac{1}{{\mathbf M}_m}\left|\sum_{{\bf b}\in {\bf B}_m}w_{\bf b} \frac{1}{\sqrt{\bf M}_m}e^{-2\pi i \langle {\bf R}_m^{-1}{\bf b}, \lambda\rangle}\right|^2.\\
\end{aligned}$$ Note that, by Proposition \[proposition1.3\](i) and Lemma \[lem3.2\] $$\sum_{\lambda\in \Lambda_m}\frac{1}{{\bf M}_m}\left|\sum_{{\bf b}\in B_m}w_{\bf b} \frac{1}{\sqrt{{\bf M}_m}}e^{-2\pi i \langle {\bf R}_m^{-1}{\bf b}, \lambda\rangle}\right|^2\ge {\bf C}_m\cdot\frac{1}{{\bf M}_m} \sum_{{\bf b}\in{\bf B}_m} |w_{\bf b}|^2 ={\bf C}_m\int|f|^2d\mu.$$ We further have $$\begin{aligned}
Q_m(f) \ge& Q_n(f)+\delta(\Lambda)\cdot \left({\bf C}_m\int|f|^2d\mu-\sum_{\lambda\in \Lambda_n}\frac{1}{{\bf M}_m}\left|\sum_{{\bf b}\in B_m}w_{\bf b} \frac{1}{\sqrt{{\bf M}_m}}e^{-2\pi i \langle {\bf R}_m^{-1}{\bf b}, \lambda\rangle}\right|^2\right)\\
= &Q_n(f)+\delta(\Lambda)\cdot \left({\bf C}_m\int|f|^2d\mu-\sum_{\lambda\in \Lambda_n}\left|\int f(x)e^{-2\pi i \lambda x}d\mu_m(x)\right|^2\right)
\end{aligned}$$ For a fixed $n$, we let $m$ go to infinity. By the fact that $Q_{m}(f)$ converges to $Q_{\infty}(f)$ and $\mu_m$ converges weakly to $\mu$, we have $$Q_{\infty}(f)\ge Q_n(f) +\delta(\Lambda)\cdot \left({\bf C}_{\infty}\int|f|^2d\mu- \sum_{\lambda\in \Lambda_n}\left|\int f(x)e^{-2 \pi i \lambda x}d\mu(x)\right|^2\right).$$ We then let $n$ go to infinity and obtain $$Q_{\infty}(f)\ge Q_{\infty}(f) +\delta(\Lambda)\cdot \left({\bf C}_{\infty}\int|f|^2d\mu-\sum_{\lambda\in \Lambda}\left|\int f(x)e^{-2\pi i \lambda x}d\mu(x)\right|^2\right).$$ and thus $$\delta(\Lambda)\cdot \left({\bf C}_{\infty}\int|f|^2d\mu-\sum_{\lambda\in \Lambda}\left|\int f(x)e^{-2\pi i \lambda x}d\mu(x)\right|^2\right)\leq 0.$$ However, $\delta(\Lambda)>0$ and we have $${\bf C}_{\infty}\int|f|^2d\mu\le \sum_{\lambda\in \Lambda}\left|\int f(x)e^{-2\pi i \lambda x}d\mu(x)\right|^2$$ This establishes the lower bound.
We now prove (ii). Take any finite subset $\Lambda_0$ of ${\Lambda}$. As $\Lambda_n$ is an increasing union, we have that $\Lambda_0\subset\Lambda_n$ for $n$ large. Note that $$\left\| \sum_{\lambda\in\Lambda_0}a_{\lambda}e^{2\pi i \langle\lambda,x\rangle}\right\|_{L^2(\mu_n)}^2= \frac{1}{\#{\bf B}_n}\sum_{{\bf b}\in {\bf B}_n} \left| \sum_{\lambda\in \Lambda_0} a_{\lambda} e^{2\pi i \langle {\bf R}_n^{-1}b,\lambda\rangle} \right|^2 =\left\| \sum_{\lambda\in \Lambda_0} a_{\lambda} {\bf e}_{R,\lambda}\right\|^2.$$ Hence, by Proposition \[proposition1.3\], $$\label{eq1.2}
\left(\prod_{j=1}^nC_j\right) \cdot \sum_{\lambda\in\Lambda_0}|a_{\lambda}|^2 \le \left\| \sum_{\lambda\in\Lambda_0}a_{\lambda}e^{2\pi i \langle\lambda,x\rangle}\right\|_{L^2(\mu_n)}^2\le \left(\prod_{j=1}^nD_j\right) \cdot \sum_{\lambda\in\Lambda_0}|a_{\lambda}|^2$$ As $\mu_n$ converges weakly to $\mu$ and $\sum_{\lambda\in\Lambda_0}a_{\lambda}e^{2\pi i \langle\lambda,x\rangle}$ is a continuous function, $$\lim_{n\rightarrow\infty}\left\| \sum_{\lambda\in\Lambda_0}a_{\lambda}e^{2\pi i \langle\lambda,x\rangle}\right\|_{L^2(\mu_n)}^2 = \left\| \sum_{\lambda\in\Lambda_0}a_{\lambda}e^{2\pi i \langle\lambda,x\rangle}\right\|^2_{L^2(\mu)}.$$ Hence, our conclusion follows by taking limit in (\[eq1.2\]).
1. The inequality $\delta(\Lambda)>0$ may be regarded as a condition guaranteeing that the tail-term of $\widehat{\mu}$ does not become too small. It is a sufficient condition for the canonical spectrum $\Lambda$ to be a frame-spectrum. But it is in general not necessary (See [@MR3055992]).
2. Theorem \[thm1.3\] (ii) shows that there is no extra condition for $\Lambda$ to be a Riesz sequence once we have formed our Riesz sequence tower. However, it may happen that all the sets $L_j$ have only one element, then trivially, the set $\Lambda$ has only one element which must form a Riesz sequence. Hence, to construct an infinite Riesz sequence, we need to make sure $\#L_j\ge 2$. We will show, using a version of the Kadison-Singer theorem, that such a Riesz sequence always exists (Section \[Section\_KS\]).
Exactness and Overcompleteness
==============================
In this section, we will study the exactness and overcompleteness of the Fourier frame generated by the frame and Riesz-sequence towers.
Exactness and overcompleteness of the frame tower
-------------------------------------------------
\[theorem2.1\] Let $\{(R_j, B_j, L_j): j\ge 1\}$ be a frame tower with no-overlap. Suppose that $\delta(\Lambda)>0$ (see ). Then
\(i) Suppose that all $\#B_j = \#L_j$. Then $E(\Lambda) = \{e^{2\pi i \langle\lambda,x\rangle}:\lambda\in\Lambda\}$ is a Riesz basis for $L^2(\mu)$.
\(ii) Suppose that there exists $j$ such that $\#B_j<\#L_j$. Then $\{e^{2\pi i \langle\lambda,x\rangle}:\lambda\in\Lambda\}$ is a Fourier frame for $L^2(\mu)$ with infinite redundancies (i.e., there exists an infinite subset $\Lambda_0$ of $\Lambda$ such that $\Lambda\setminus\Lambda_0$ is still a frame spectrum for $L^2(\mu)$).
\(i) Since $\delta(\Lambda)>0$, we know that $E(\Lambda)$ is a frame. We now show that it is a Riesz basis by showing that $E(\Lambda)$ is an exact frame. Let $\lambda_0$ be an element in $\Lambda$ and we need to show that $E(\Lambda\setminus\{\lambda_0\})$ is incomplete.
To show this, we note that there exists $n_0$ such that $\lambda_0\in\Lambda_n$ for all $n\ge n_0$. From the assumption, $\#\Lambda_{n}= \#{\bf B}_n = \prod_{j=1}^n\#B_j$. Since $L^2(\mu_n)$ is finite dimensional with dimension $\#{\bf B}_n$, we have that $E(\Lambda_n)$ forms a Riesz basis for $L^2(\mu_n)$. In particular, there exists $f_n\in L^2(\mu_n)$ such that $\|f_n\|_{L^2(\mu_n)} = 1$ and $$\label{eq2.1}
\frac{1}{\sqrt{\#{\bf B}_n}}\langle {\bf w}^n, {\bf e}_{{\bf R}_n,\lambda}\rangle_{{\mathbb C}^{\#{\bf B}_n}} = \langle f_n, e^{2\pi i \langle \lambda, x\rangle}\rangle_{L^2(\mu_n)} = 0, \ \forall \lambda\in \Lambda_n\setminus\{\lambda_0\}.$$ where $w^n_b = f_n(b)$ and ${\bf w}^n = (w^n_b)_{b\in{\bf B}_n}$. As $\|f_n\|_{L^2(\mu_n)} = 1$, $\|{\bf w}^n\|^2 = \#{\bf B}_n$.
The vector $f_n$ can be identified naturally with $f_n = \sum_{b\in{\bf B}_n} w^n_b {\bf 1}_{K_{b,n}}\in L^2(\mu)$. $$\|f_n\|_{L^2(\mu)}^2 = \frac{1}{\#{\bf B}_n}\sum_{b\in {\bf B}_n} |w_{b,n}|^2=\|f_n\|^2_{L^2(\mu_n)}=1$$ By the Banach-Alaoglu theorem, there is a subsequence $f_{n_k}$ that converges weakly to some function $f\in L^2(\mu)$. Note that for all $\lambda\in\Lambda\setminus\{\lambda_0\}$, $\lambda\in \Lambda_{n_k}$ for all $k$ sufficiently large, and we have, with Lemma \[lem3.2\], $$\label{eq2.2}
\langle f_{n_k}, e^{2\pi i \langle \lambda, x\rangle}\rangle_{L^2(\mu)} = \widehat{\mu_{>n_k}}(\lambda) \langle f_{n_k}, e^{2\pi i \langle\lambda,x\rangle}\rangle_{L^2(\mu_{n_k})} = 0.$$ By taking $k\rightarrow\infty$, we obtain that $$\langle f, e^{2\pi i \langle \lambda, x\rangle}\rangle_{L^2(\mu)} = 0, \ \forall \lambda\in\Lambda\setminus\{\lambda_0\}.$$ Therefore, $f$ is orthogonal to all $e^{2\pi i \langle\lambda,x\rangle}$, $\lambda\in\Lambda\setminus\{\lambda_0\}$. We now show that $f\ne 0$ in $L^2(\mu)$, then this implies that $E(\Lambda\setminus\{\lambda_0\})$ is not complete.
Indeed, if we take ${\bf w}^n$ into the definition of the frame triple of $({\bf R}_n,{\bf B}_n,\Lambda_n)$ and use (\[eq2.1\]), we obtain $$\left(\prod_{j=1}^nC_j\right)\|{\bf w}^n\|^2 \le \sum_{\lambda\in\Lambda_n} |\langle {\bf w}^n,{\bf e}_{{\bf R}_n,\lambda}\rangle|^2 = |\langle {\bf w}^n,{\bf e}_{{\bf R}_n,\lambda_0}\rangle|^2 = \#{\bf B}_n |\langle f_n, e^{2\pi i \langle \lambda_0, x\rangle}\rangle_{L^2(\mu_n)}|^2.$$ Note that $ \|{\bf w}^n\|^2=\#{\bf B}_n$ and $|\widehat{\mu_{>n}}(\lambda)|^2\ge \delta(\Lambda)>0$. By (\[eq2.2\]), we have $$|\langle f_n, e^{2\pi i \langle \lambda_0, x\rangle}\rangle_{L^2(\mu)}|^2 \ge \left(\prod_{j=1}^nC_j\right) \cdot \delta(\Lambda).$$ Take $n=n_k$ to infinity, we obtain $$|(fd\mu)^{\widehat{}}(\lambda_0)|^2 \ge \left(\prod_{j=1}^{\infty}C_j\right) \cdot \delta(\Lambda)>0.$$ Since $(fd\mu)^{\widehat{}}(\lambda_0)\ne 0$, $f$ cannot be a zero function in $L^2(\mu)$ and this completes the proof.
\(ii) Suppose that at the $j_0$th level, we have that $\#B_{j_0}<\#L_{j_0}$. Then the collection of vectors $\{{\bf e}_{R_{j_0},\lambda}: \lambda\in L_{j_0}\}$ is a linearly dependent set in ${\mathbb C}^{\#B_{j_0}}$. Since we have a finite dimensional vector space, there exists $\lambda_{j_0}\in L_{j_0}$ such that $\{{\bf e}_{R_{j_0},\lambda}: \lambda\in L_{j_0}\setminus\{\lambda_{j_0}\}\}$ is still a frame for ${\mathbb C}^{\#B_{j_0}}$ with frame bounds $\widetilde{C_j}$ and $\widetilde{D_j}$. Let $\widetilde{L_{j_0}} = L_{j_0}\setminus\{\lambda_{j_0}\}$.
Then it is clear that $\left(\{(R_j,B_j,L_j): j\ge 1\}\setminus \{(R_{j_0},B_{j_0},L_{j_0})\} \right)\cup \{(R_{j_0},B_{j_0},\widetilde{L_{j_0}})\}$ still forms a frame tower with bounds $$\frac{\widetilde{C_j}}{C_j}\cdot\prod_{j=1}^{\infty}C_j \ \mbox{and} \ \frac{\widetilde{D_j}}{D_j}\cdot\prod_{j=1}^{\infty}D_j$$ which are still positive and finite. Moreover, the corresponding $
\widetilde{\Lambda}$ formed with $L_{j_0}$ replaced by $\widetilde{L_{j_0}}$ in (\[Lambda\_n\]) is a subset of $\Lambda$. This implies that $\delta(\widetilde{\Lambda})\ge \delta(\Lambda)>0$. By Theorem \[thm1.3\](i), $E(\widetilde{\Lambda})$ is a frame for $L^2(\mu)$. Note that the removed elements are $$\Lambda\setminus\widetilde{\Lambda} = R_1^{\mathtt T}..R_{j_0-1}^{\mathtt T}\lambda_{j_0}+ \left\{ \sum_{finite,j\ne j_0} (R_1^{\mathtt T}...R_{j-1}^{\mathtt T})\ell_{j}: \ell_j\in L_j\right\},$$ which is an infinite set. Hence, $E(\Lambda)$ is a Fourier frame for $L^2(\mu)$ with infinite redundancies.
In [@LW17], we considered on ${\mathbb R}^1$, $R_j = M_jK_j+\alpha_j$, where $M_j,K_j$ are integers and $0\le \alpha_j<M_j$ and they satisfy $$\sum_{j=1}^{\infty} \frac{\alpha_j\sqrt{M_j}}{K_j}<\infty.$$ Then letting $B_j = \{0,K_j,...,(M_j-1)K_j\}$ and $L_j = \{0,1,...,M_j-1\}$, the resulting $(R_j,B_j,L_j)$ forms a frame tower and the associated fractal measure admits a Fourier frame. Using Theorem \[theorem2.1\], we can actually conclude that the Fourier frame we constructed is actually a Riesz basis.
Incompleteness of Riesz sequence tower
--------------------------------------
In this section we consider the completeness of the Riesz basis obtained from a Riesz-sequence tower. We first recall a proposition about Riesz sequences, whose statements can be found in Young’s book ([@Young], Proposition 2 and Theorem 3 in Chapter 4, p.129).
\[proposition\_Young\] Let $H$ be a Hilbert space and let $\{f_n: n\ge 1\}$ be a Riesz sequence for $H$. Let $C$ be its lower bound. Then for any $\ell^2$ sequence $\{c_n\}$, there exists $f\in H$ such that $$\|f\|\le \frac{1}{C} \sum_{n=1}^{\infty}|c_n|^2$$ and $$\langle f,f_n\rangle =c_n, \forall n\ge1.$$
\[th4.4\] Let $\{(R_j, B_j, L_j): j\ge 1\}$ be a Riesz sequence tower and assume the associated fractal measure satisfies the no-overlap condition and elements in $B_j$ are distinct representatives in ${\mathbb Z}^d/R({\mathbb Z}^d)$. Suppose that there exists $j$ such that $\#L_j<\#B_j$. Then $\{e^{2\pi i \langle\lambda,x\rangle}:\lambda\in\Lambda\}$ is not complete in $L^2(\mu)$.
Without loss of generality, we may assume that, at the first level, we have $\#L_1<\#B_1$. From Proposition \[Prop\_distinct\], elements in $L_1$ are distinct representative in ${\mathbb Z}^d/R^{\mathtt T}({\mathbb Z}^d)$. Let $\overline{L}_1$ be a complete representative of ${\mathbb Z}^d/R^{\mathtt T}({\mathbb Z}^d)$ containing $L_1$. As elements in $B_1$ are distinct representative in ${\mathbb Z}^d/R({\mathbb Z}^d)$, $\{{\bf e}_{R,\lambda}: \lambda\in \overline{L}_1\}$ forms a tight frame in ${\mathbb C}^{\#B}$ by Lemma \[lemma\_tight\]. On the other hand, as $\{{\bf e}_{R,\lambda}: \lambda\in L_1\}$ forms a Riesz sequence, so they must be linearly independent. Hence, we can find $\lambda_1\in \overline{L_1}$ such that $(R_1,B_1,\widetilde{L_1})$ with $\widetilde{L_1} = L_1\cup\{\lambda_1\}$ forms a Riesz sequence triple and $(R_1,B_1,\widetilde{L_1}) \cup \{(R_j, B_j, L_j): j\ge 2\}$ forms a Riesz sequence tower.
We know that $({\bf R}_n, {\bf B}_n, \widetilde{\Lambda_n})$ are Riesz sequence triples for all $n\ge1$, and $$\widetilde{\Lambda_n} = \widetilde{L_1}+ R_1^{\mathtt T} L_2+...+R_1^{\mathtt T}...R_{n-1}^{\mathtt T} L_n.$$ Define $\widetilde{\Lambda}$ analogously. Then $\Lambda_n\subset\widetilde{\Lambda_n}$ and $\Lambda\subset\widetilde{\Lambda}$. Let $$c_{\lambda_1} = 1 \ \mbox{and} \ c_{\lambda} = 0 , \mbox{ for all } \ \lambda\in\widetilde{\Lambda_n}\setminus\{\lambda_1\}.$$ Then $\sum_{\lambda\in \widetilde{\Lambda_n}}|c_{\lambda}|^2 = 1$. By Proposition \[proposition1.3\] and \[proposition\_Young\], we can find $f_n\in L^2(\mu_n)$ such that $$\|f_n\|_{L^2(\mu_n)} \le \frac{1}{\prod_{j=1}^{n}C_j}, \ \mbox{and} \ \langle f_n\, ,\, e^{2\pi i \langle\lambda, x\rangle}\rangle = c_{\lambda}$$ for all $\lambda\in\Lambda_n$. Since $0<\prod_{j=1}^{\infty}C_j<\infty$ and all $C_j>0$, $C: = \inf\{\prod_{j=1}^nC_j: n\ge 1\}>0$. Identifying $f_n$ naturally in $L^2(\mu)$ with a step function, we have that $$\|f_n\|_{L^2(\mu)}\le \frac{1}{ C}.$$ By the Banach-Alaoglu theorem, we have a subsequence $f_{n_k}$ that converges weakly to some function $f\in L^2(\mu)$. Then, with Lemma \[lem3.2\], $$\langle f_{n_k}, e^{2\pi i \langle \lambda,x\rangle}\rangle_{L^2(\mu)} = \widehat{\mu_{>n_k}}(\lambda) \cdot \langle f_{n_k}, e^{2\pi i \langle \lambda,x\rangle}\rangle_{L^2(\mu_{n_k})} .$$ In particular, by taking limit and since $\Lambda \subset\widetilde{\Lambda}$, we have $$\langle f, e^{2\pi i \langle \lambda,x\rangle}\rangle_{L^2(\mu)} =0, \mbox{ for all } \lambda\in \Lambda.$$ If we can show that $f\ne 0$, then $f$ is orthogonal to the the closure of the span of $e^{2\pi i \langle\lambda,x\rangle}$, $\lambda\in\Lambda$. This will show that $E(\Lambda)$ is not complete.
To show that $f\ne 0$, we note that by our choice of $c_{\lambda},$ $$\langle f_n, e^{2\pi i \langle\lambda_1,x\rangle}\rangle_{L^2(\mu_n)} = 1.$$ This implies that $$\langle f_n, e^{2\pi i \langle\lambda_1,x\rangle}\rangle_{L^2(\mu)} = \widehat{\mu_{>n}}(\lambda_1)\langle f_n, e^{2\pi i \langle\lambda_1,x\rangle}\rangle_{L^2(\mu_n)} = \widehat{\mu_{>n}}(\lambda_1).$$ We note that the measure $\mu_{>n}$ converges weakly to $\delta_0$. This means that $\widehat{\mu_{>n}}(\cdot)$ converges to $1$ uniformly on all compact subsets of ${\mathbb R}^d$. Hence, $\widehat{\mu_{>n}}(\lambda_1)$ converges to $1$ and we have $$\langle f, e^{2\pi i \langle\lambda_1,x\rangle}\rangle_{L^2(\mu)} = 1.$$ This shows that $f$ cannot be a zero function in $L^2(\mu)$.
\[example013\] Consider $R = 3$, $B = \{0,1,3\}$ and $L = \{0,1\}$. Then $(R,B,L)$ forms a Riesz sequence triple. However, we can never add another $\lambda\subset {\mathbb Z}$ into $L$ so that $(R,B,L\cup\{\lambda\})$ forms a Riesz sequence triple.
Note that the matrix $${\mathcal F}_{L,B} = \begin{pmatrix} - & {\bf e}_{R,0}& -\\ - & {\bf e}_{R,1}&-\end{pmatrix} = \frac{1}{\sqrt{3}} \begin{pmatrix} 1 & 1& 1\\ 1 & \omega &1\end{pmatrix}$$ where $\omega$ is the cubic root of unity. The two rows are linearly independent. Hence, $(R,B,L)$ forms a Riesz sequence triple. However, to add one more element, we see that we can only add $\lambda = 2 $ (mod 3) by Proposition \[Prop\_distinct\]. However, if we add this, $${\mathcal F}_{L\cup\{2\},B} = \frac{1}{\sqrt{3}} \begin{pmatrix} 1 & 1& 1\\ 1 & \omega &1 \\ 1& \omega^2 & 1\end{pmatrix}.$$ The rows are not linearly independent and do not span ${\mathbb C}^{\#B}$. This example shows that the assumption that elements of $B$ is from a distinct representative in ${\mathbb Z}^d/R({\mathbb Z}^d)$ is required in Theorem \[th4.4\]. Otherwise we cannot add another element that preserve the Riesz sequence property.
Existence of Riesz sequence towers {#Section_KS}
==================================
In this section, we will show that under some simple assumptions on $R_j,B_j$, one can construct a Riesz sequence tower easily. The key theorem is the following version of the solution to the Kadison-Singer problem, given in [@BCMS16 Theorem 6.12].
\[thR\][**\[$R_{\epsilon}$ conjecture\]**]{} Suppose that $\{u_i\}_{i\in I}$ is a unit norm Bessel sequence with Bessel bound $D$ in a separable Hilbert space $H$. Then there exists a universal constant $C_0>0$ such that for any $\epsilon>0$, one can find a partition $\{I_1,...,I_r\}$ of $I$ of size $r \le C_0(D/\epsilon^4)$ such that each $\{u_i\}_{i\in I_j}$, $j=1,...,r$ is a Riesz sequence with bounds $1-\epsilon$, $1+\epsilon$.
Focusing on the finite dimensional Hilbert space in the above theorem, we notice that if $\epsilon$ is very small, then $r$ will most likely be very large and each $I_j$ may possibly be containing only one element, in which case the theorem would be trivially true. In our application, we will need some $I_j$ to contain more than one element.
\[lemma3.2\] Let $R$ be an integral expanding matrix and let $B$ be a finite subset of ${\mathbb{Z}}^d$ containing some distinct representatives (mod $R({\mathbb Z}^d)$). Suppose that $0<\epsilon<1$. Then there exists $L$ with $\#L\ge \frac{\#B\epsilon^4}{C_0}$ such that $(R,B,L)$ forms a Riesz sequence triple with bounds $1-\epsilon,1+\epsilon$.
Given an integer expanding matrix $R$, let $\overline{L}$ be a complete set of distinct representatives of ${\mathbb Z}^d/R^{\mathtt T}({\mathbb Z}^d)$. By Lemma \[lemma\_tight\], we have that $$\sum_{\lambda\in\overline{L}} \left|\langle{\bf w},{\bf e}_{R,\lambda} \rangle \right|^2 = \frac{|\det (R)|}{\#B} \|{\bf w}\|^2$$ for all ${\bf w}\in{\mathbb C}^{\#B}$. By Theorem \[thR\], one can find a partition $\{I_1,...,I_r\}$ such that each $\{{\bf e}_{R,\lambda}\}_{\lambda\in I_j}$, $j=1,...,r$ is a Riesz sequence with bounds $1-\epsilon$, $1+\epsilon$. Note that at least one of the $I_j$ must contain at least $|\det R|/r$ elements. As $r\le C_0(D/\epsilon^4)$ where $D = |\det(R)|/\#B$, there exists $j$ such that $$\#I_j\ge \frac{\#B\epsilon^4}{C_0}.$$ Take $L = I_j$ and this completes the proof.
Indeed, we do not need the full strength of the Kadison-Singer problem (as in the $R_\epsilon$ conjecture above) to prove Lemma \[lemma3.2\]. It can be also obtained as a consequence of the Bourgain-Tzafriri restricted invertibility theorem [@MR890420], see also [@MR2500595 Proposition 4.4] . We would like to thank Peter Casazza for pointing this fact to us.
\[theorem\_Riesz\] Let $\{(R_j,B_j)\}$ be a sequence of pairs with integral expanding matrix $R_j$ and $B_j$ a finite subset containing some distinct representatives (mod $R_j({\mathbb Z}^d)$). Suppose also that the associated fractal measure $$\mu = \delta_{{\bf R}_1^{-1}B_1}\ast\delta_{{\bf R}_2^{-1}B_2}\ast...$$ has the no-overlap condition. Then $\mu$ admits an exponential Riesz sequence of infinite cardinality.
First, we note that we can group any $n_j$ consecutive factors without changing the resulting measure $\mu$. $$\begin{aligned}
\mu =& \left(\delta_{{\bf R}_1^{-1}B_1}\ast...\ast\delta_{{\bf R}_{n_1}^{-1}B_{n_1}}\right)\ast\left(\delta_{{\bf R}_{n_1+1}^{-1}B_{n_1+1}}\ast...\ast\delta_{{\bf R}_{n_1+n_2}^{-1}B_{n_2}}\right)\ast...\\
=&\delta_{{\bf R}_{n_1}^{-1}{\bf B}_{n_1}}\ast\delta_{{\bf R}_{n_1+n_2}^{-1}{\bf B}_{n_2}}\ast....
\end{aligned}$$ As each $\#B_n\ge 2$, $\#{\bf B}_n\ge 2^n$. Using this regrouping, we can assume that each $\#B_n\ge 2^n$. For $n$ large enough, we let $1>\epsilon_n> \left(\frac{C_0}{\#B_n}\right)^{1/4}$. As $\#B_n\ge 2^n$, one can also find sequence $\epsilon_n$ that is summable. By Lemma \[lemma3.2\], we can find $L_n$ such that $(R_n,B_n,L_n)$ forms a Riesz-sequence triple and $\#L_n \ge \frac{\#B_n\epsilon_n^4}{C_0}>1$. Since the sequence $\epsilon_n$ is summable, we can use Theorem \[thm1.3\](ii) and conclude that $\Lambda$ forms a Riesz sequence for $\mu$. Since each $\#L_n>1$, $\Lambda$ is an infinite set.
Riesz sequence with optimal Beurling dimension
----------------------------------------------
In this section, we focus on self-affine measures.
\[defifs\] For a given expansive $d\times d$ integer matrix $R$ and a finite set of integer vectors $B$ with $\#B$, we define the [*affine iterated function system*]{} (IFS) $$\tau_b(x) = R^{-1}(x+b),\quad ( x\in {\mathbb{R}}^d, b\in B).$$ The [*self-affine measure*]{} associated to $R$ and $B$ is the unique probability measure $\mu = \mu(R,B)$ satisfying $$\label{self-affine}
\mu(E) =\frac{1}{\#B} \sum_{b\in B} \mu (\tau_b^{-1} (E)),\mbox{ for all Borel subsets $E$ of ${\mathbb{R}}^d$.}$$
This measure is supported on the [*attractor*]{} $T(R,B)$ which is the unique compact set that satisfies $$T(R,B)= \bigcup_{b\in B} \tau_b(T(R,B)).$$ The set $T(R,B)$ is also called the [*self-affine set*]{} associated with the IFS. It can also be described as $$T(R,B)=\left\{\sum_{k=1}^\infty R^{-k}b_k : b_k\in B\right\}.$$ If $R=\rho O$ where $|\rho|<1$ and $O$ is an orthogonal matrix, then $\mu(R,B)$ is called a self-similar measure. One can refer to [@Hut81] and [@Fal97] for a detailed exposition of the theory of iterated function systems.
It is known that self-affine measures can be realized as an infinite convolution product $$\mu(R,B) = \delta_{R^{-1}B}\ast\delta_{R^{-2}B}\ast.....$$ Therefore, they fit into the category of measures that we have considered in the previous sections. By factorization of any $n_k$ consecutive factors, we can also write it as $$\mu(R,B) = \delta_{R^{-{n_1}}B_{n_1}}\ast\delta_{R^{-{(n_1+n_2)}}B_{n_2}}\ast.....$$ It is known [@DHL18 Section 2] that $\mu(R,B)$ satisfies the no-overlap condition if the digits in $B$ are chosen from distinct coset representatives in ${\mathbb Z}^d/R({\mathbb Z}^d)$.
Let $\Lambda$ be a countable set on ${\mathbb R}^d$. The $\alpha$-Beurling density is defined to be $$D_{\alpha}^{+}(\Lambda) = \limsup_{h\rightarrow\infty}\sup_{x \in {\mathbb R}^d} \frac{\#(\Lambda\cap B(x,h))}{h^{\alpha}}, \ B(x,h)= \{y: |y-x|<h\}$$ and the Beurling dimension of $\Lambda$ is defined to be $$\dim_B(\Lambda) = \sup \{\alpha: D^{\alpha}(\Lambda)>0\}.$$
\[thmDHSW\] [@DHSW11; @MR2826404] [(i)]{} Let $\mu(R,B)$ be the self-similar measure with $R = \rho O$ as in Definition \[defifs\], and let $E(\Lambda)$ be a Bessel sequence of $\mu(R,B)$. Then the Beurling dimension of $\Lambda$ is at most $\log_{\rho}(\#B)$.
[(ii)]{} Let $\mu(R,B)$ be the self-similar measure. Then there exists a Bessel sequence $E(\Lambda)$ of positive Beurling dimension.
We are going to prove the following theorem:
\[th5.7\] Let $\mu(R,B)$ be the self-similar measure with $R = \rho O$ and assume that $\mu(R,B)$ satisfies the no-overlap condition. Then there exists a Riesz sequence $E(\Lambda)$ of Beurling dimension $\log_{\rho}(\#B)$ for $\mu(R,B)$.
By regrouping, $\mu(R,B)$ is generated by the tower $(R^{n_k}, B_{n_k})$, where $\#B_{n_k} = (\#B)^{n_k}$. Let $\epsilon_k = \left(\frac{C_0}{n_k^5}\right)^{1/4}$ and by Lemma \[lemma3.2\], we can find $L_{n_k}$ such that $(R^{n_k}, B_{n_k}, L_{n_k})$ forms a Riesz sequence triple and $$\#L_{n_k}\ge \frac{(\#B)^{n_k}}{n_k^5}.
\label{eqdl}$$
We now take $n_k = k$ so that $\sum \epsilon_k<\infty$ and then we obtain a Riesz sequence tower $(R^{n_k}, B_{n_k}, L_{n_k})$ and hence $\mu(R,B)$ admits a Riesz sequence $E(\Lambda)$, with $$\Lambda = \bigcup_{k=1}^{\infty}{\bf L_k}, \ {\bf L}_{k}: = L_{n_1}+(R^{\mathtt T})^{n_1} L_{n_2}+(R^{\mathtt T})^{n_1+n_2} L_{n_2}+...+(R^{\mathtt T})^{n_1+...+n_{k-1}} L_{n_k}$$ by Theorem \[thm1.3\](ii). It remains to show that the Beurling dimension of $\Lambda$ is $\log_{\rho}(\#B)$.
Let $Q_n = (R^{\mathtt T})^n({\overline{B}}_{\sqrt{d}}(0))={\overline{B}}_{\rho^n\sqrt{d}}(0)$ for $n=1,2,...$, note that $B_{\sqrt d}(0)$ contains the cube $[-1/2,1/2]^d$. We note also that $L_{n_i}$ can be chosen to be inside $Q_{n_i}$, by reducing $({\operatorname{mod}}R^{\mathtt T})^{n_i}$, and then ${\bf L}_{k}$ is inside $Q_{n_1+...+n_k+1}$. For any ${\bf x}\in {\bf L}_k$, $$|{\bf x}|\le \sqrt{d}\rho^{n_1+...+n_k+1}.$$ Therefore, letting $h_k = \sqrt{d}\rho^{n_1+...+n_k+1} =\sqrt{d}\rho^{k(k+1)/2+1}$ (since $n_k = k$), and using , we have $$\#(\Lambda\cap B_{h_k}(0)) \ge \prod_{i=1}^k\#L_{n_i}\ge \prod_{i=1}^k\frac{(\#B)^{n_i}}{n_i^5} = \frac{(\#B)^{k(k+1)/2}}{(k!)^5}$$ Let $\alpha = \log_{\rho}(\#B)$. For any $\eta>0$, $$\begin{aligned}
\frac{\#(\Lambda\cap B_{h_k}(0))}{h_k^{\alpha-\eta}} \ge& \frac{\frac{(\#B)^{k(k+1)/2}}{(k!)^5}}{d^{(\alpha-\eta)/2}(\#B)^{k(k+1)/2+1}\rho^{-\eta (k(k+1)/2+1)}}\\
\ge&\frac{1}{d^{(\alpha-\eta)/2}} \cdot \frac{\rho^{\eta k(k+1)/2}}{k^{5k}}, \ \\
\ge&\frac{1}{d^{(\alpha-\eta)/2}} \cdot e^{ck^2-5k\ln k} \rightarrow +\infty, \ (\mbox{let} \ c = \ln(\rho^{\eta/2})>0)\\
\end{aligned}$$ since $\lim_{k\rightarrow \infty}(ck^2-5k\ln k) = +\infty$. This shows that $D^{+}_{\alpha-\eta}(\Lambda) = \infty$ for all $\eta>0$. Hence, dim$_B(\Lambda)\ge \log_{\rho}(\#B)$. Hence, together with Theorem \[thmDHSW\], we have dim$_B(\Lambda)= \log_{\rho}(\#B)$. This completes the proof.
The authors would like to thank the referee for his/her valuable suggestion. They would also like to thank the referee and professor Peter Casazza for pointing out how Lemma \[lemma3.2\] can also be deduced from a weaker version of the Bourgain-Tzafriri theorem. This work was partially supported by a grant from the Simons Foundation (\#228539 to Dorin Dutkay).
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Quantum and private communications are affected by a fundamental limitation which severely restricts the optimal rates that are achievable by two distant parties. To overcome this problem, one needs to introduce quantum repeaters and, more generally, quantum communication networks. Within a quantum network, other problems and features may appear when we move from the basic unicast setting of single-sender/single-receiver to more complex multiend scenarios, where multiple senders and multiple receivers simultaneously use the network to communicate. Assuming various configurations, including multiple-unicast, multicast, and multiple-multicast communication, we bound the optimal rates for transmitting quantum information, distributing entanglement, or generating secret keys in quantum networks connected by arbitrary quantum channels. These bounds cannot be surpassed by the most general adaptive protocols of quantum network communication.'
author:
- Stefano Pirandola
title: 'Bounds for multi-end communication over quantum networks'
---
Introduction
============
Quantum and private communications represent some of the most advanced areas of quantum information [@NiCh; @first; @HolevoBOOK; @review; @BraRMP]. In particular, quantum key distribution (QKD) [@BB84; @Ekert] has been already developed into several commercial prototypes, besides the fact that quantum-secured networks and satellite quantum communications are being developed by various countries [@QKDadvance]. A more ambitious and long-term goal is that of the quantum internet [@Kimble; @HybridINTERNET; @Whener] where remote quantum computers are suitable connected by optical links so as to ultimately create a worldwide architecture for distributed quantum computing.
In terms of quantum communications, one of the basic reasons to build quantum networks [@Rod] is to overcome the rate limitations of point-to-point protocols. As shown in Ref. [@QKDpaper], the maximum rates at which two remote parties can transmit quantum information, distribute entanglement or secret correlations over a lossy channel of transmissivity $\eta$ are all equal to $-\log_{2}(1-\eta)$, also known as the Pirandola-Laurenza-Ottaviani-Banchi (PLOB) bound. For the specific case of QKD, this ultimate point-to-point rate can be achieved by employing a quantum memory at the sender side, as shown back in 2009 when the notion of reverse coherent information was introduced for bosonic channels [@ReverseCAP; @RevCohINFO]. On the other hand, if a middle node is inserted between the remote parties, the PLOB bound can be practically beaten, as shown by the recent twin-field QKD protocol [@Marco].
Once understood that the use of relays or repeaters [@Briegel; @Rep2; @Rep3] can overcome the PLOB bound, it is also important to understand the ultimate limits achievable by repeater-assisted quantum communications [@netpaper; @longVersion]. Using techniques from network information theory [@Slepian; @Schrijver; @Gamal; @Cover; @netflow] and very recent channel simulation tools developed in quantum information [@QKDpaper] (see also Refs. [@Metro; @nonPauli; @TQC; @BK2; @Qmetro; @revSENS]), one can bound or exactly derive the capacities for quantum and private communication between two end-points of a repeater chain or a quantum network. This was shown in Ref. [@netpaper] which reports the end-to-end (unicast) results originally established in the 2016 unpublished work [@longVersion].
The present paper reports and refines the other (multi-end) results of unpublished [@longVersion], thus providing a generalization of Ref. [@netpaper] from the unicast setting of single-sender/single-receiver to more complex scenarios where multiple senders and receivers are involved. In these scenarios, the remote parties compete with the others in order to make an optimal use of the quantum network. We assume different configurations, including multiple-unicast (where there are many single-sender/single-receiver pairs trying to communicate in a simultaneous fashion), multicast (where a single-sender tries to communicate with multiple receivers), and multiple-multicast (where different senders try to communicate with the same set of multiple receivers). In all these communication configurations, we derive single-letter upper bounds for the maximum rates at which the parties can transmit quantum information, distribute entanglement or secret keys. These bounds are valid for networks connected by arbitrary quantum channels and are expressed in terms of the relative entropy of entanglement (REE) [@RMPrelent; @VedFORMm; @Pleniom].
It is important to remark that the results apply to both discrete- and continuous-variable systems, i.e., quantum networks connected by quantum channels acting over finite- or infinite-dimensional spaces. In fact, as discussed afterwards, the theory presented for finite dimension $d$ can be extended to $d = +\infty$ by generalizing the notion of channel simulation to an asymptotic formulation which is based on a sequence of finite-energy resource states. Thanks to this tool, we can compute the relevant functionals on the sequence and then take the infinite-energy limit of their values over the sequence. In this way, we automatically and rigorously prove the results for bosonic channels, following the methods that were originally designed in Refs. [@QKDpaper; @netpaper; @longVersion].
The paper is organized as follows. In Sec. \[GeneralSECTION\] we present preliminary notions on adaptive protocols, besides the tools of simulation and stretching for channels [@QKDpaper] and networks [@netpaper]. The expert reader can skip this part and directly start from Sec. \[SECmultipleNETs\] which provides general description of the various configurations considered in this work. Secs. \[SecMULTIunicast\] and \[SecMULTIunicast2\] consider multiple-unicast settings under single- and multi-path routing strategies for the quantum systems. Sec. \[SECmulticastSINGLE\] investigates the case of multicast communication from a sender to multiple receivers. Following Sec. \[SECmulticastMANY\] considers multiple-multicast communication between many senders and many receivers. Finally, Sec. \[SECconclusions\] is for conclusions.
Preliminaries\[GeneralSECTION\]
===============================
Adaptive point-to-point protocols
---------------------------------
Let us first discuss the general structure of an adaptive point-to-point protocol $\mathcal{P}$ through a quantum channel $\mathcal{E}$, following the notation from Ref. [@QKDpaper]. Alice has a local register of quantum systems $\mathbf{a}$ and Bob has another local register $\mathbf{b}$; these are prepared in a state $\rho_{\mathbf{ab}}^{0}$ by means of local operations (LOs) assisted by two-way classical communication (CC), also known as adaptive LOCCs. After the first adaptive LOCC $\Lambda_{0}$, Alice selects a system $a_{1}\in\mathbf{a}$ and sends it to Bob through the quantum channel $\mathcal{E}$. Once Bob receives the output $b_{1}$, this is included in his register $b_{1}\mathbf{b}\rightarrow\mathbf{b}$ and another adaptive LOCC $\Lambda_{1}$ is performed by the parties. The second transmission starts by selecting another system $a_{2}\in\mathbf{a}$ which is sent through $\mathcal{E}$ whose output $a_{2}$ is received by Bob. Bob updates his register $b_{2}\mathbf{b}\rightarrow\mathbf{b}$ and another adaptive LOCC $\Lambda_{2}$ is performed. The generic $i$-th transmission is shown in Fig. \[longPIC\]. After $n$ uses, Alice and Bob have implemented an adaptive protocol $\mathcal{P}$ defined by the sequence of LOCCs $\{\Lambda_{0},\Lambda_{1}\ldots\}$ and providing an output state $\rho_{\mathbf{ab}}^{n}$ close in trace norm to a target state $\phi^{n}$ with $nR_{n}^{\varepsilon}$ (target) bits, i.e., such that $\left\Vert \rho_{\mathbf{ab}}^{n}-\phi
^{n}\right\Vert \leq\varepsilon$.
![Generic $i$th transmission through channel $\mathcal{E}$ in a point-to-point adaptive protocol. The transmission $a_{i}\rightarrow b_{i}$ is interleaved by two adaptive LOCCs, $\Lambda_{i-1}$ and $\Lambda_{i}$, performed by Alice and Bob on their local registers $\mathbf{a}$ and $\mathbf{b}$.[]{data-label="longPIC"}](protocol.eps){width="50.00000%"}
If we now consider the limit for large $n$ (asymptotic rate) and small $\varepsilon$ (weak converse), and then we optimize over the protocols $\mathcal{P}$, we define the two-way assisted capacity of $\mathcal{E}$, i.e.,$$\mathcal{C}(\mathcal{E}):=\sup_{\mathcal{P}}\lim_{\varepsilon,n}R_{n}^{\varepsilon}.$$ Assume that the target $\phi^{n}$ is a maximally-entangled state, so that the target bits are entanglement bits (ebits). In this case, $\mathcal{C}(\mathcal{E})$ corresponds to the two-way entanglement-distribution capacity $D_{2}(\mathcal{E})$, which is also equal to the two-way quantum capacity $Q_{2}(\mathcal{E})$. Assume instead that $\phi^{n}$ is a private state [@KD1], so that the target bits are private bits, then $\mathcal{C}(\mathcal{E})$ corresponds to the secret key capacity $K(\mathcal{E})\geq D_{2}(\mathcal{E})$.
LOCC simulation of quantum channels
-----------------------------------
One may follow the general recipe of Ref. [@QKDpaper] to simplify an adaptive protocol over $\mathcal{E}$ and write a single-letter upper bound for the two-way assisted capacity $\mathcal{C}(\mathcal{E})$. The first step is the simulation of the channel $\mathcal{E}$ by means of an LOCC $\mathcal{T}$ and some resource state $\sigma$. For any channel $\mathcal{E}$, we may always find a simulation such that $$\mathcal{E}(\rho)=\mathcal{T}(\rho\otimes\sigma). \label{sigma}$$ A channel simulable with resource state $\sigma$ may also be called $\sigma$-stretchable. It is important to note that the simulation may also be asymptotic, so that we have a sequence of LOCCs $\mathcal{T}^{\mu}$ and resource states $\sigma^{\mu}$ such that we may write the point-wise limit [@QKDpaper]$$\mathcal{E}(\rho)=\lim_{\mu}\mathcal{T}^{\mu}(\rho\otimes\sigma^{\mu}).
\label{asymptotic}$$
A very convenient simulation holds for those channels commuting with the teleportation correction unitaries, which are (generalized) Pauli operators in finite dimension and phase-space displacements in continuous variable systems [@telereview]. By definition, a quantum channel $\mathcal{E}$ is called teleportation-covariant, or just telecovariant when, for any teleportation unitary $U$, we may write$$\mathcal{E}(U\rho U^{\dagger})=V\mathcal{E}(\rho)V^{\dagger}~,
\label{stretchability}$$ for another (generally-different) unitary $V$. Note that Pauli channels [@NiCh], erasure channels and bosonic Gaussian channels [@review] are all telecovariant.
For a telecovariant channel $\mathcal{E}$, we write the simulation $$\mathcal{E}(\rho)=\mathcal{T}_{\text{tele}}(\rho\otimes\sigma_{\mathcal{E}}),
\label{kkkll}$$ where $\mathcal{T}_{\text{tele}}$ is a teleportation protocol [@teleBENNETT; @Samtele; @Samtele2; @telereview] and $\sigma
_{\mathcal{E}}:=\mathcal{I}\otimes\mathcal{E}(\Phi)$ is the Choi matrix of the channel, with $\Phi$ being a maximally entangled state. For bosonic Gaussian channels, the Choi matrix is asymptotic, i.e., defined by the $\mu$-limit of the sequence $\sigma_{\mathcal{E}}^{\mu}:=\mathcal{I}\otimes\mathcal{E}(\Phi^{\mu})$, where $\Phi^{\mu}$ is a two-mode squeezed vacuum (TMSV) state with variance parameter $\mu$ [@review]. Thus, we may write the asymptotic simulation$$\mathcal{E}(\rho)=\lim_{\mu}\mathcal{T}_{\text{tele}}^{\mu}(\rho\otimes
\sigma_{\mathcal{E}}^{\mu}), \label{asymptotics}$$ where $\mathcal{T}_{\text{tele}}^{\mu}$ is a sequence of teleportation-LOCCs.
Stretching and single-letter bound
----------------------------------
In an adaptive protocol, we may replace each transmission through the channel $\mathcal{E}$ with its simulation $(\mathcal{T},\sigma)$. Then, as shown in Ref. [@QKDpaper], we may collapse all the simulation LOCCs $\mathcal{T}$ and the adaptive LOCCs of the protocol $\{\Lambda_{0},\Lambda_{1}\ldots\}$ into a single trace-preserving LOCC $\bar{\Lambda}$. In this way, the $n$-use output state of the protocol can be decomposed as$$\rho_{\mathbf{ab}}^{n}=\bar{\Lambda}\left( \sigma^{\otimes n}\right) .
\label{StretchingMAIN}$$ If the simulation of the channel is asymptotic, the stretching takes the form $\rho_{\mathbf{ab}}^{n}=\lim_{\mu}\bar{\Lambda}_{\mu}(\sigma^{\mu\otimes n})$ for a sequence of trace-preserving LOCC $\bar{\Lambda}_{\mu}$ and resource states $\sigma^{\mu}$. See Ref. [@QKDpaper; @TQC; @BK2] for a precise formulation of this limit where the simulation error is explicitly taken into account.
Suppose that we want to compute an entanglement measure over the output state. In particular, let us consider the REE. For a quantum state $\rho$, this is [@RMPrelent] $$E_{\mathrm{R}}(\rho)=\inf_{\gamma\in\text{\textrm{SEP}}}S(\rho||\gamma),
\label{REEbona}$$ where $\gamma$ is an arbitrary separable state and $S$ is the quantum relative entropy $S(\rho||\gamma):=\mathrm{Tr}\left[ \rho(\log_{2}\rho-\log_{2}\gamma)\right] $. More weakly, if we consider an asymptotic state $\sigma:=\lim_{\mu}\sigma^{\mu}$, the previous definition can be extended as$$E_{\text{\textrm{R}}}(\sigma)=\underset{\mu\rightarrow+\infty}{\lim\inf
}~E_{\text{\textrm{R}}}(\sigma^{\mu})=\inf_{\gamma^{\mu}}~\underset
{\mu\rightarrow+\infty}{\lim\inf}~S(\sigma^{\mu}||\gamma^{\mu}),
\label{extendedREE}$$ where $\gamma^{\mu}$ is a converging sequence of separable states [@QKDpaper].
Because the REE is monotonic under trace-preserving LOCCs (data processing) and sub-additive over tensor products, we may write $$E_{\mathrm{R}}(\rho_{\mathbf{ab}}^{n})=E_{\mathrm{R}}\left[ \bar{\Lambda
}\left( \sigma^{\otimes n}\right) \right] \leq E_{\mathrm{R}}\left(
\sigma^{\otimes n}\right) \leq nE_{\mathrm{R}}\left( \sigma\right) .$$ Now recall that the REE is also asymptotically continuous. This means that for $\rho_{\mathbf{ab}}^{n}$ and $\phi^{n}$ such that $\left\Vert \rho
_{\mathbf{ab}}^{n}-\phi^{n}\right\Vert \leq\varepsilon$, we may write $$|E_{\mathrm{R}}(\phi^{n})-E_{\mathrm{R}}(\rho_{\mathbf{ab}}^{n})|\leq
\delta(d,\varepsilon), \label{deltaTERM}$$ where $\delta(\varepsilon,d)$ depends the $\varepsilon$-closeness, and the dimension $d$ of the total Hilbert space. In the limit of large $n$ and small $\varepsilon$ (weak converse), we can neglect $\delta(\varepsilon,d)/n$. This is a straightforward application of the exponential scaling of the dimension $d$ shown in Refs. [@Matthias1a; @Matthias2a] for DV systems and extended to CV systems in Ref. [@QKDpaper]. For a simplified treatment for CV systems see also Ref. [@TQC]. Because $E_{\mathrm{R}}(\phi^{n})\geq nR_{n}^{\varepsilon}$ [@KD1], we therefore have$$R_{n}^{\varepsilon}\leq E_{\mathrm{R}}\left( \sigma\right) +n^{-1}\delta(d,\varepsilon),$$ which leads to the single-letter upper bound [@QKDpaper] $$\mathcal{C}(\mathcal{E})\leq E_{\mathrm{R}}(\sigma). \label{UB1}$$ In particular, for telecovariant $\mathcal{E}$, we may write $$\mathcal{C}(\mathcal{E})\leq E_{\mathrm{R}}(\sigma_{\mathcal{E}}), \label{UB2}$$ with a suitable asymptotic formulation for bosonic Gaussian channels based on Eq. (\[extendedREE\]).
Among telecovariant channels, the distillable ones are those for which we may write $E_{\mathrm{R}}(\sigma_{\mathcal{E}})=D_{1}(\sigma_{\mathcal{E}})$, where $D_{1}(\sigma_{\mathcal{E}})$ is the distillable entanglement of the (possibly-asymptotic) channel’s Choi matrix $\sigma_{\mathcal{E}}$ via one-way CCs, forward or backward. This is lower-bounded by both the coherent [@Schu96; @Lloyd97] and reverse coherent [@RevCohINFO; @ReverseCAP] information of the Choi matrix. For a distillable channel, the two-way capacities coincide and are given by$$\mathcal{C}(\mathcal{E})=E_{\mathrm{R}}(\sigma_{\mathcal{E}})=D_{1}(\sigma_{\mathcal{E}}). \label{coincidence}$$ This is the case for a number of channels, including the dephasing channel, the erasure channel, the pure-loss channel and the quantum-limited amplifier. For a pure-loss channel with transmissivity $\eta$, the two-way capacity is simply given by the PLOB bound [@QKDpaper Eq. (19)]$$\mathcal{C}(\eta)=-\log_{2}(1-\eta)~. \label{formCloss}$$ As secret-key capacity, this bounds the maximum rate of any point-to-point QKD protocol.
Notation for quantum networks
-----------------------------
We model a quantum communication network $\mathcal{N}$ as an undirected finite graph [@Slepian] $\mathcal{N}=(P,E)$, where $P$ is the set of points or nodes, and $E$ is the set of undirected edges. Every point can be identified with a corresponding local register $\mathbf{p}$ of quantum systems. The existence of an edge $(\mathbf{x},\mathbf{y})$, between two points $\mathbf{x}$ and $\mathbf{y}$, means that there is a physical quantum channel $\mathcal{E}_{\mathbf{xy}}$ between them (which can be forward $\mathcal{E}_{\mathbf{x\rightarrow y}}$ or backward $\mathcal{E}_{\mathbf{y}\rightarrow\mathbf{x}}$). For points $\mathbf{p}_{i}$ and $\mathbf{p}_{j}$, we also use the short-hand notation $\mathcal{E}_{ij}:=\mathcal{E}_{\mathbf{p}_{i}\mathbf{p}_{j}}$. We use the notation $\mathbf{a}$ and $\mathbf{b}$ for Alice and Bob, respectively. A path or route between these two end-points is a sequence of edges $(\mathbf{a},\mathbf{p}_{i}),\cdots,(\mathbf{p}_{j},\mathbf{b})$, that we may simply denote as $\mathbf{a}-\mathbf{p}_{i}-\cdots-\mathbf{p}_{j}-\mathbf{b}$. For a route with $N$ middle points (or repeaters), we have $N+1$ edges and, therefore, a corresponding sequence of $N+1$ channels $\{\mathcal{E}_{0},\cdots,\mathcal{E}_{k},\cdots,\mathcal{E}_{N}\}$ from Alice to Bob.
There are different possible routes between two end-points. For this reason, they may also use different routing strategies. In single-path routing, Alice and Bob exploit a single route in each use of the network, and this route can be changed for the different network uses. In multi-path routing, Alice and Bob exploit many routes in parallel in each use of the network. In particular, they may adopt a flooding strategy [@flooding] where each edge of the network is used exactly once in each end-to-end transmission. In both cases, we assume that the quantum protocols are adaptive, meaning that each transmission through each channel is interleaved with a network adaptive LOCCs, where all points of the network apply LOs on their local registers assisted by unlimited two-way CC with all the other points of the network.
An entanglement cut $C$ of the quantum network $\mathcal{N}$ is a bipartition $(\mathbf{A},\mathbf{B})$ of all the points $P$ such that $\mathbf{a}\in\mathbf{A}$ and $\mathbf{b}\in\mathbf{B}$. Correspondingly, the cut-set $\tilde{C}$ of $C$ is the ensemble of edges across the bipartition, i.e., $\tilde{C}=\{(\mathbf{x},\mathbf{y})\in E:\mathbf{x}\in\mathbf{A},\mathbf{y}\in\mathbf{B}\}$. It is clear that $\tilde{C}$ also identifies an ensemble of channels $\{\mathcal{E}_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in\tilde{C}}$. Given a cut, we may also consider the complementary sets $$\begin{aligned}
\tilde{A} & =\{(\mathbf{x},\mathbf{y})\in E:\mathbf{x,y}\in\mathbf{A}\},\\
\tilde{B} & =\{(\mathbf{x},\mathbf{y})\in E:\mathbf{x,y}\in\mathbf{B}\},\end{aligned}$$ so that $\tilde{A}\cup\tilde{B}\cup\tilde{C}=E$.
Given an undirected network $\mathcal{N}=(P,E)$ we can introduce an orientation by transforming it in a directed graph. One can choose a direction for all edges so that a generic edge $(\mathbf{x},\mathbf{y})$ becomes an ordered pair where $\mathbf{x}$ is the tail and $\mathbf{y}$ is the head. In choosing these directions, we keep Alice as tail and Bob as head, so that the quantum network can be represented as a flow network where Alice is the* source* and Bob is the* sink* [@Harris; @Ford; @ShannonFLOW; @Karp; @Dinic]. There are $\mathcal{O}(2^{|E|})$ possible orientations. Given an orientation of $\mathcal{N}$, there is a corresponding flow network $\mathcal{N}_{D}=(P,E_{D})$, where $E_{D}$ is the set of directed edges. Then, for arbitrary point $\mathbf{p}$, we define its out-neighborhood as the set of heads going from $\mathbf{p}$$$N^{\text{out}}(\mathbf{p})=\{\mathbf{x}\in P:(\mathbf{p},\mathbf{x})\in
E_{D}\},$$ and its in-neighborhood as the set of tails going into $\mathbf{p}$$$N^{\text{in}}(\mathbf{p})=\{\mathbf{x}\in P:(\mathbf{x},\mathbf{p})\in
E_{D}\}.$$ A multi-message quantum multicast from point $\mathbf{p}$ is a point-to-multipoint connection from $\mathbf{p}$ to part of its out-neighborhood $N^{\text{out}}(\mathbf{p})$, so that different messages (quantum states or keys) are received by the receiving points. It is a single-message multicast if the messages coincide.
Simulation and stretching of a network
--------------------------------------
Given an arbitrary network $\mathcal{N}$, we may replace it with its simulation [@netpaper; @longVersion]. In fact, for any edge $(\mathbf{x},\mathbf{y})$, we may replace the quantum channel $\mathcal{E}_{\mathbf{xy}}$ with its simulation $S_{\mathbf{xy}}=(\mathcal{T}_{\mathbf{xy}},\sigma
_{\mathbf{xy}})$ for some LOCC $\mathcal{T}_{\mathbf{xy}}$ and resource state $\sigma_{\mathbf{xy}}$. Repeating this process for all the edges defines an LOCC simulation of the network $S(\mathcal{N})=\{S_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$ where all channels $\mathcal{E}_{\mathbf{xy}}$ are replaced by resource states $\sigma_{\mathbf{xy}}$. There is a corresponding resource representation of the network $\sigma
(\mathcal{N})=\{\sigma_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$. See also Fig. \[diamond\] for a simple example. In particular, for a telecovariant network, where all channels are telecovariant, then the simulation involves teleportation LOCCs and the network can be replaced by its Choi representation $\sigma(\mathcal{N})=\{\sigma_{\mathcal{E}_{\mathbf{xy}}}\}_{(\mathbf{x},\mathbf{y})\in E}$. Here each channel $\mathcal{E}_{\mathbf{xy}}$ is replaced by its (possibly-asymptotic) Choi matrix $\sigma_{\mathcal{E}_{\mathbf{xy}}}$. A quantum network is said to be distillable if it is connected by distillable channels.
![Network simulation. Consider a simple four-point quantum network $\mathcal{N}=(\{\mathbf{p}_{0},\mathbf{p}_{1},\mathbf{p}_{2},\mathbf{p}_{3}\},E)$ with end points $\mathbf{p}_{0}=\mathbf{a}$ (Alice) $\mathbf{p}_{3}=\mathbf{b}$ (Bob). Edge $(\mathbf{p}_{i},\mathbf{p}_{j})$ has an associated quantum channel $\mathcal{E}_{ij}$. By simulating each channel $\mathcal{E}_{ij}$ with a corresponding resource state $\sigma_{ij}$, we define a resource representation of the network $\sigma(\mathcal{N})=\{\sigma_{01},\sigma_{02},\sigma_{13},\sigma_{23}\}$.[]{data-label="diamond"}](diamond.eps){width="50.00000%"}
Given an arbitrary adaptive protocol implemented over the quantum network $\mathcal{N}$, we can use the network simulation $\sigma(\mathcal{N})$ to stretch the protocol and decompose the total output state $\rho
_{\mathbf{a\ldots b}}^{n}$ of network after $n$ uses as follows$$\rho_{\mathbf{a\ldots b}}^{n}=\bar{\Lambda}\left[ \underset{(\mathbf{x},\mathbf{y})\in E}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n_{\mathbf{xy}}}\right] ,
\label{LemmaNETstretching}$$ where $\bar{\Lambda}$ is a trace-preserving LOCC and $n_{\mathbf{xy}}$ is the number of uses of the edge $(\mathbf{x},\mathbf{y})$. In particular, we have $n_{\mathbf{xy}}\leq n$ ($n_{\mathbf{xy}}=n$) for a single-path (flooding) protocol. In other words, introducing the probabilities $p_{\mathbf{xy}}:=n_{\mathbf{xy}}/n$ we have $p_{\mathbf{xy}}\leq1$ ($=1$) for a single-path (flooding) protocol [@netpaper; @longVersion].
Tracing out all the network points except the two end-points, from Eq. (\[LemmaNETstretching\]) we get Alice and Bob’s shared state $\rho_{\mathbf{ab}}^{n}$. For any entanglement cut $C$ and corresponding cut-set $\tilde{C}$, we may write a better decomposition for Alice and Bob’s output state. This is given by$$\rho_{\mathbf{ab}}^{n}(C)=\bar{\Lambda}_{\mathbf{ab}}\left[ \underset
{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n_{\mathbf{xy}}}\right] , \label{cutEQ}$$ where $\bar{\Lambda}_{\mathbf{ab}}$ is a trace-preserving LOCC with respect to Alice and Bob. Previous Eqs. (\[LemmaNETstretching\]) and (\[cutEQ\]) can be extended to asymptotic resource states by introducing suitable limits. See Ref. [@netpaper; @longVersion] for more details on these methods.
Multiple senders and receivers\[SECmultipleNETs\]
=================================================
One of the basic working mechanisms in a quantum communication network is the unicast setting, based on a single sender $\mathbf{a}$ and a single receiver $\mathbf{b}$. However, in general, we may consider multiple senders $\{\mathbf{a}_{i}\}$ and receivers $\{\mathbf{b}_{j}\}$, which may simultaneously communicate according to various configurations. For simplicity, these sets are intended to be disjoint $\{\mathbf{a}_{i}\}\cap\{\mathbf{b}_{j}\}=\emptyset$, so that an end-point cannot be sender and receiver at the same time. It is clear that all the results from Ref. [@netpaper; @longVersion], derived for the two basic routing strategies, provide general upper bounds which are still valid for the individual end-to-end capacities associated with each sender-receiver pair $(\mathbf{a}_{i},\mathbf{b}_{i})$ in the various settings with multiple end-points.
In the following sections, we start with the multiple-unicast quantum network. This consists of $M$ Alices $\{\mathbf{a}_{1},\ldots,\mathbf{a}_{M}\}$ and $M$ Bobs $\{\mathbf{b}_{1},\ldots,\mathbf{b}_{M}\}$, with the generic $i$th Alice $\mathbf{a}_{i}$ communicating with a corresponding $i$th Bob $\mathbf{b}_{i}$. This case can be studied by assuming single-path routing (Sec. \[SecMULTIunicast\]) or multipath routing (Sec. \[SecMULTIunicast2\]). Besides the general bounds inherited from the unicast scenario, we derive a specific set of upper bounds for the rates that are simultaneously achievable by all parties.
Another important case is the multicast (multi-message) quantum network, where a single sender simultaneously communicates with $M\geq1$ receivers, e.g., for distributing $M$ different states or keys. By its nature, this is studied under multipath routing (Sec. \[SECmulticastSINGLE\]). In this setting, an interesting variant is the distribution of the same key to all receivers (single-message multicast).
More generally, we may consider a multiple-multicast (multi-message) quantum network. Here we have $M_{A}\geq1$ senders and $M_{B}\geq1$ receivers, and each sender communicates simultaneously with the entire set of receivers communicating different states or keys (Sec. \[SECmulticastMANY\]). In a private communication scenario, this corresponds to the distribution of $M_{A}M_{B}$ different keys. For a description of these configurations, see the simple example of the butterfly quantum network in Fig. \[butterfly\].
![Butterfly quantum network. (i) An example of multiple-unicast is considering two sender-receiver pairs, e.g., Alice $\mathbf{a}_{1}$ communicating with Bob $\mathbf{b}_{1}$, and Alice $\mathbf{a}_{2}$ with Bob $\mathbf{b}_{2}$. Single-path routing corresponds to the simultaneous use of two end-to-end routes, e.g., $(\mathbf{a}_{1})\mathbf{p}_{0}-\mathbf{p}_{2}-\mathbf{p}_{3}-\mathbf{p}_{5}(\mathbf{b}_{1})$ and $(\mathbf{a}_{2})\mathbf{p}_{1}-\mathbf{p}_{2}-\mathbf{p}_{3}-\mathbf{p}_{4}(\mathbf{b}_{2})$. Multipath routing corresponds to choosing a network orientation, where the end-points may also act as relays. Each point of the network multicasts multiple messages to its out-neighborhood. For instance, we may have the point-to-multipoint multicasts: $\mathbf{p}_{0}\rightarrow
\{\mathbf{p}_{2},\mathbf{p}_{4}\}$, $\mathbf{p}_{1}\rightarrow\{\mathbf{p}_{2},\mathbf{p}_{5}\}$, $\mathbf{p}_{2}\rightarrow\mathbf{p}_{3}$, and $\mathbf{p}_{3}\rightarrow\{\mathbf{p}_{4},\mathbf{p}_{5}\}$. (ii) An example of network multicast is Alice $\mathbf{a}_{1}$ communicating with the two Bobs $\{\mathbf{b}_{1},\mathbf{b}_{2}\}$ via multipath routing. In general, the messages (states, keys) can be different. (iii) In a multiple-multicast, Alice $\mathbf{a}_{1}$ communicates with $\{\mathbf{b}_{1},\mathbf{b}_{2}\}$, and Alice $\mathbf{a}_{2}$ communicates with the same destination set $\{\mathbf{b}_{1},\mathbf{b}_{2}\}$. In general, the messages (states, keys) can be different.[]{data-label="butterfly"}](20_butterfly.eps){width="37.00000%"}
Multiple-unicast quantum networks with single-path routing\[SecMULTIunicast\]
=============================================================================
Let us start by considering two Alice-Bob pairs $(\mathbf{a}_{1},\mathbf{b}_{1})$ and $(\mathbf{a}_{2},\mathbf{b}_{2})$, since the extension to arbitrary number of pairs is immediate. We may easily formulate network protocols which are based on single-path routing. In this case, each sequential use of the network involves the transmission of quantum systems along two (potentially-overlapping) routes$$\omega_{1}:\mathbf{a}_{1}-\cdots-\mathbf{b}_{1},~~~\omega_{2}:\mathbf{a}_{2}-\cdots-\mathbf{b}_{2},$$ where each transmission through an edge is assisted by network LOCCs. The routes are updated use after use.
After $n$ uses, the output of the double-unicast network protocol $\mathcal{P}_{\text{2-unicast}}$ is a state $\rho_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}^{n}$ which is $\varepsilon$-close in trace norm to a target state $$\phi:=\phi_{\mathbf{a}_{1}\mathbf{b}_{1}}^{\otimes nR_{1}^{\varepsilon,n}}\otimes\phi_{\mathbf{a}_{2}\mathbf{b}_{2}}^{\otimes nR_{2}^{\varepsilon,n}},
\label{targettt}$$ where $\phi_{\mathbf{a}_{i}\mathbf{b}_{i}}$ is a one-bit state (private bit or ebit) for the pair $(\mathbf{a}_{i},\mathbf{b}_{i})$ and $nR_{i}^{\varepsilon,n}$ the number of its copies. Taking the limit of large $n$, small $\varepsilon$ (weak converse) and optimizing over all protocols $\mathcal{P}_{\text{2-unicast}}$, we define the capacity region as the closure of the set of the achievable asymptotic rates $(R_{1},R_{2})$. In general, for $M$ sender-receiver pairs, we have an $M$-tuple of achievable rates $(R_{1},\ldots,R_{M})$. Depending on the task of the protocol (i.e., the target state), these rates refer to end-to-end entanglement distillation (equivalently, error-free quantum communication) or secret-key generation.
Before proceeding, let us first introduce more general types of entanglement cuts of the quantum network. Given two sets of senders $\{\mathbf{a}_{i}\}$ and receivers $\{\mathbf{b}_{i}\}$, we adopt the notation $C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}$ for a cut $C=(\mathbf{A},\mathbf{B})$ such that $\{\mathbf{a}_{i}\}\subset\mathbf{A}$ and $\{\mathbf{b}_{i}\}\subset
\mathbf{B}$. Similarly, we write $C:\mathbf{a}_{i}|\mathbf{b}_{i}$ for a cut with $\mathbf{a}_{i}\in\mathbf{A}$ and $\mathbf{b}_{i}\in\mathbf{B}$, and $C:\mathbf{a}_{i}\mathbf{a}_{j}|\mathbf{b}_{i}\mathbf{b}_{j}$ for a cut with $\{\mathbf{a}_{i},\mathbf{a}_{j}\}\subset\mathbf{A}$ and $\{\mathbf{b}_{i},\mathbf{b}_{j}\}\subset\mathbf{B}$. Define also the single-edge flow of entanglement (REE) trough a cut as$$E_{\mathrm{R}}(C):=\max_{(\mathbf{x},\mathbf{y})\in\tilde{C}}E_{R}(\sigma_{\mathbf{xy}}),$$ where $\sigma_{\mathbf{xy}}$ is a resource state associated with an edge $(\mathbf{x},\mathbf{y})$ in the cut-set $\tilde{C}$, under some simulation of the network. We can then state the following result.
\[Multi-unicast with single paths\]\[TheomultipleUNICAST\]Let us consider a multiple-unicast quantum network $\mathcal{N}=(P,E)$ with $M$ sender-receiver pairs $(\mathbf{a}_{i},\mathbf{b}_{i})$ communicating by means of single-path routing. Adopt a simulation of the network with a resource representation $\sigma(\mathcal{N})=\{\sigma_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$. In particular, $\sigma(\mathcal{N})$ can be a Choi-representation for a teleportation-covariant $\mathcal{N}$. We have the following outer bounds for the capacity region$$\begin{aligned}
R_{i} & \leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{i}}E_{\mathrm{R}}(C)~~\text{for any }i,\label{unicvvv}\\
R_{i}+R_{j} & \leq\min_{C:\mathbf{a}_{i}\mathbf{a}_{j}|\mathbf{b}_{i}\mathbf{b}_{j}}E_{\mathrm{R}}(C)~~\text{for any }i\neq j
\label{doublehhhh}\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i} & \leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}(C),\end{aligned}$$ where $E_{\mathrm{R}}(C)$ is the single-edge flow of REE through cut $C$. It is understood that formulations may be asymptotic for quantum networks with bosonic channels.
**Proof.** For simplicity, first consider the case $M=2$, since the generalization to arbitrary $M$ is straightforward. Let us also consider key generation, since it automatically provides an upper bound for all the other tasks. Considering the bipartition $\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2}$, the distillable key of the target state $\phi$ in Eq. (\[targettt\]) is equal to$$K_{\mathrm{D}}(\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2})_{\phi}=n(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n}).$$ Using the REE with respect to the same bipartition, we may write the upper bound$$\begin{aligned}
n(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n}) & \leq E_{\mathrm{R}}(\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2})_{\phi}\nonumber\\
& \leq E_{\mathrm{R}}(\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2})_{\rho^{n}}+\delta(\varepsilon,d),\end{aligned}$$ where the latter inequality comes from the fact that $\rho^{n}:=\rho
_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}^{n}$ is $\varepsilon$-close to $\phi$. The extra term $\delta(\varepsilon,d)$ depends on the $\varepsilon$-closeness and the dimension $d$ of the total Hilbert space, as already discussed in relation to Eq. (\[deltaTERM\]). The term $n^{-1}\delta(d,\varepsilon)$ goes to zero for large $n$ and small $\varepsilon$. As a result we may write $$\lim_{\varepsilon,n}(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n})\leq
\underset{n\rightarrow+\infty}{\lim}n^{-1}E_{\mathrm{R}}(\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2})_{\rho^{n}}~. \label{todecppp}$$
By simulating and stretching the network, we may write the following decomposition of the output state$$\rho_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}^{n}=\bar{\Lambda}_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}\left[ \underset{(\mathbf{x},\mathbf{y})\in E}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n_{\mathbf{xy}}}\right] , \label{cmtoo}$$ where $n_{\mathbf{xy}}=np_{\mathbf{xy}}$ is the number of uses of edge $(\mathbf{x},\mathbf{y})$ and $\bar{\Lambda}_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}$ is a trace-preserving LOCC, which is local with respect to the bipartition $\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2}$. By inserting entanglement cuts which disconnect the senders and receivers, we reduce the number of resource states appearing in Eq. (\[cmtoo\]) while preserving the locality of the LOCC with respect to the bipartition of the end-points. In other words, for any cut $C:\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2}$ we may write$$\rho_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}^{n}(C)=\bar{\Lambda}_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}^{C}\left[ \underset{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n_{\mathbf{xy}}}\right] .$$
Using the latter decomposition in Eq. (\[todecppp\]), we obtain$$\begin{aligned}
\lim_{\varepsilon,n}(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n}) &
\leq\underset{n\rightarrow+\infty}{\lim}~n^{-1}E_{\mathrm{R}}(\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2})_{\rho^{n}(C)}\nonumber\\
& \leq\underset{n\rightarrow+\infty}{\lim}~n^{-1}\sum\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}n_{\mathbf{xy}}E_{\mathrm{R}}(\sigma_{\mathbf{xy}})\nonumber\\
& =\sum\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}p_{\mathbf{xy}}E_{\mathrm{R}}(\sigma_{\mathbf{xy}})\nonumber\\
& \leq\max_{(\mathbf{x},\mathbf{y})\in\tilde{C}}E_{\mathrm{R}}(\sigma
_{\mathbf{xy}}):=E_{\mathrm{R}}(C).\end{aligned}$$
By minimizing over the cuts, we derive$$\lim_{\varepsilon,n}(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n})\leq
\min_{C:\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2}}E_{\mathrm{R}}(C). \label{outerCCC}$$ It is important to note that this bound holds for any protocol $\mathcal{P}_{\text{2-unicast}}$, whose details are all collapsed in the LOCC $\bar{\Lambda}_{\mathbf{a}_{1}\mathbf{a}_{2}\mathbf{b}_{1}\mathbf{b}_{2}}$ and therefore discarded. Thus, the same bound applies if we optimize over all protocols, which means that Eq. (\[outerCCC\]) provides the following outer bound for the capacity region$$\begin{aligned}
R_{1}+R_{2} & =\sup_{\mathcal{P}_{\text{2-unicast}}}\lim_{\varepsilon
,n}(R_{1}^{\varepsilon,n}+R_{2}^{\varepsilon,n})\nonumber\\
& \leq\min_{C:\mathbf{a}_{1}\mathbf{a}_{2}|\mathbf{b}_{1}\mathbf{b}_{2}}E_{\mathrm{R}}(C).\end{aligned}$$
Note that, besides this bound, we also have the following unicast bounds for the individual rates $$R_{1}\leq\min_{C:\mathbf{a}_{1}|\mathbf{b}_{1}}E_{\mathrm{R}}(C),~~R_{2}\leq\min_{C:\mathbf{a}_{2}|\mathbf{b}_{2}}E_{\mathrm{R}}(C).$$ These follows directly from the results of Ref. [@netpaper; @longVersion] on the converse bounds for unicast quantum networks. Equivalently, we may re-derive these bounds here, by setting $R_{2}=0$ or $R_{1}=0$ in the target state of Eq. (\[targettt\]) and repeating the previous derivation. For instance, for $R_{2}=0$, we have $\phi:=\phi_{\mathbf{a}_{1}\mathbf{b}_{1}}^{\otimes nR_{1}^{\varepsilon,n}}\otimes\sigma_{\mathbf{a}_{2}\mathbf{b}_{2}}$, where $\sigma_{\mathbf{a}_{2}\mathbf{b}_{2}}$ does not contain target bits and may be taken to be separable. Therefore, we start from $K_{\mathrm{D}}(\mathbf{a}_{1}|\mathbf{b}_{1})_{\phi}=nR_{1}^{\varepsilon,n}$ and repeat the derivation with respect to $\mathbf{a}_{1}|\mathbf{b}_{1}$.
It is easy to generalize from $M=2$ to arbitrary $M$. For any integer $M$, we have the target state$$\phi:={\textstyle\bigotimes_{i=1}^{M}}
\phi_{\mathbf{a}_{i}\mathbf{b}_{i}}^{\otimes nR_{i}^{\varepsilon,n}}.$$ Considering the bipartition $\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}$ and the corresponding cuts of the network leads to $$\sum\limits_{i=1}^{M}R_{i}\leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}(C),$$ where we note that increasing the number of rates reduces the number of possible cuts in the minimization. To get all the remaining inequalities of the theorem, we just need to set some of the rates to zero. For instance, for $R_{i}\neq0$ and $R_{j\neq i}=0$, we get the unicast bounds of Eq. (\[unicvvv\]). For $R_{i}\neq0$, $R_{j\neq i}\neq0$ and $R_{k\neq
i,j}=0$ we get the double-unicast bounds of Eq. (\[doublehhhh\]), and so on. The extension to asymptotic simulations of bosonic channels is achieved via the weaker definition of REE in Eq. (\[extendedREE\]). $\blacksquare$
Once we have Theorem \[TheomultipleUNICAST\], it is immediate to specify the results for the case of multiple-unicast distillable networks, for which we may write $E_{\mathrm{R}}(\sigma_{\mathbf{xy}})=E_{\mathrm{R}}(\sigma
_{\mathcal{E}_{\mathbf{xy}}})=\mathcal{C}_{\mathbf{xy}}$ for each edge $(\mathbf{x,y})\in E$, where $\mathcal{C}_{\mathbf{xy}}$ is the two-way capacity of the associated quantum channel $\mathcal{E}_{\mathbf{xy}}$. In this case, we may directly write $$E_{\mathrm{R}}(C)=\mathcal{C}(C):=\max_{(\mathbf{x},\mathbf{y})\in\tilde{C}}\mathcal{C}_{\mathbf{xy}},$$ where $\mathcal{C}(C)$ is the single-edge capacity of cut $C$. Thus, we can express the bounds of Theorem \[TheomultipleUNICAST\] in terms of the capacities of the cuts, immediately proving the following.
Consider a multiple-unicast quantum network $\mathcal{N}$ with $M$ sender-receiver pairs $(\mathbf{a}_{i},\mathbf{b}_{i})$ communicating by means of single-path routing. If the network is distillable, then we may write the following outer bounds for the capacity region$$\begin{aligned}
R_{i} & \leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{i}}\mathcal{C}(C)~~\text{for
any }i,\label{distooo}\\
R_{i}+R_{j} & \leq\min_{C:\mathbf{a}_{i}\mathbf{a}_{j}|\mathbf{b}_{i}\mathbf{b}_{j}}\mathcal{C}(C)~~\text{for any }i\neq j\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i} & \leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}\mathcal{C}(C), \label{distoooo}$$ where $\mathcal{C}(C)$ is the single-edge capacity of cut $C$.
Note that we cannot establish the achievability of the outer bounds in Eqs. (\[distooo\])-(\[distoooo\]), apart from the case $M=1$. This case in fact corresponds to a unicast distillable network for which the bound is achievable by solving the widest path problem [@netpaper; @longVersion]. In general, for $M>1$, achievable lower bounds can be established by combining the point-to-point composition strategies with classical routing algorithms that solve the multiple-version of the widest path problem.
Multiple-unicast quantum networks with multipath routing\[SecMULTIunicast2\]
============================================================================
Here we consider a quantum network where $M$ senders $\{\mathbf{a}_{i}\}$ and $M$ receivers $\{\mathbf{b}_{i}\}$ communicate in a pairwise fashion $(\mathbf{a}_{i},\mathbf{b}_{i})$ by means of multipath routing. In a multipath protocol, the points first agree an orientation for the quantum network. For multiple-unicasts note that both the senders and receivers may assists one with each other as relays of the network. This means that $\{\mathbf{a}_{i}\}$ are not necessarily sources and $\{\mathbf{b}_{i}\}$ are not necessarily sinks, i.e., these sets may have both incoming and outgoing edges. Given an orientation, each point multicasts to its out-neighborhood with the assistance of network LOCCs. This flooding process ends when each edge of the network has been exploited. For the next use, the points may agree a different orientation, and so on.
The sequence of the orientations together with the sequence of all network LOCCs (exploited in each orientation) define a multiple-unicast flooding protocol $\mathcal{P}_{\text{M-unicast}}^{\text{flood}}$. Its output will be a shared state $\rho_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{i}\}}^{n}$ which is $\varepsilon$-close to a target state $$\phi:={\textstyle\bigotimes_{i=1}^{M}}
\phi_{\mathbf{a}_{i}\mathbf{b}_{i}}^{\otimes nR_{i}^{\varepsilon,n}},$$ where $\phi_{\mathbf{a}_{i}\mathbf{b}_{i}}$ is a one-bit state (private bit or ebit) for the pair $(\mathbf{a}_{i},\mathbf{b}_{i})$ and $nR_{i}^{\varepsilon,n}$ the number of its copies. By taking the limit of large $n$, small $\varepsilon$, and optimizing over $\mathcal{P}_{\text{M-unicast}}^{\text{flood}}$, we define the capacity region associated with the achievable rates $(R_{1}^{\text{m}},\ldots,R_{M}^{\text{m}})$ for the various quantum tasks. We can then state the following result.
\[Multi-unicast with multipaths\]\[TheomultipleUNICAST2\]Let us consider a multiple-unicast quantum network $\mathcal{N}=(P,E)$ with $M$ sender-receiver pairs $(\mathbf{a}_{i},\mathbf{b}_{i})$ communicating via multipath routing. Adopt a simulation of the network with a resource representation $\sigma(\mathcal{N})=\{\sigma_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$. In particular, $\sigma(\mathcal{N})$ can be a Choi-representation for a teleportation-covariant $\mathcal{N}$. We have the following outer bounds for the capacity region$$\begin{aligned}
R_{i}^{\text{m}} & \leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{i}}E_{\mathrm{R}}^{\text{m}}(C)~~\text{for any }i,\label{rtg}\\
R_{i}^{\text{m}}+R_{j}^{\text{m}} & \leq\min_{C:\mathbf{a}_{i}\mathbf{a}_{j}|\mathbf{b}_{i}\mathbf{b}_{j}}E_{\mathrm{R}}^{\text{m}}(C)~~\text{for any
}i\neq j\label{rtg2}\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i}^{\text{m}} & \leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}^{\text{m}}(C), \label{rtg3}$$ where $E_{\mathrm{R}}^{\text{m}}(C):=\sum_{(\mathbf{x},\mathbf{y})\in\tilde
{C}}E_{\mathrm{R}}(\sigma_{\mathbf{xy}})$ is the multi-edge flow of REE across cut $C$. It is understood that formulations may be asymptotic for quantum networks with bosonic channels.
**Proof.** The proof follows the main steps of the one of Theorem \[TheomultipleUNICAST\]. As before, consider key generation. For the bipartition $\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}$, the distillable key of the target state $\phi$ is equal to$$\begin{aligned}
K_{\mathrm{D}}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\})_{\phi} &
=n\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\\
& \leq E_{R}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\})_{\phi}\\
& \leq E_{R}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\})_{\rho^{n}}+\delta
(\varepsilon,d),\end{aligned}$$ which leads to the inequality$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\leq
\underset{n\rightarrow+\infty}{\lim~}n^{-1}E_{R}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\})_{\rho^{n}}~. \label{repkkkj}$$
For any cut $C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}$ of the (simulated) network, we may write the following decomposition of the output state$$\rho_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{i}\}}^{n}(C)=\bar{\Lambda}_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{i}\}}^{C}\left[ \underset{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n}\right] ,$$ for some trace-preserving LOCC $\bar{\Lambda}_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{i}\}}^{C}$. Note that here we have $n_{\mathbf{xy}}=n$. By replacing $\rho^{n}=\rho_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{i}\}}^{n}(C)$ in Eq. (\[repkkkj\]), we therefore get$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\leq
\sum\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}E_{\mathrm{R}}(\sigma
_{\mathbf{xy}}):=E_{\mathrm{R}}^{\text{m}}(C).$$
The next step is to minimize over the cuts, leading to$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\leq
\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}^{\text{m}}(C).$$ Since the latter inequality holds for any protocol $\mathcal{P}_{\text{M-unicast}}^{\text{flood}}$, it can be extended to the achievable rates $$\begin{aligned}
\sum\limits_{i=1}^{M}R_{i}^{\text{m}} & =\sup_{\mathcal{P}_{\text{M-unicast}}^{\text{flood}}}\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon
,n}\nonumber\\
& \leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}^{\text{m}}(C).\end{aligned}$$
Finally, by setting some of the rates equal to zero in the target state, we may repeat the procedure with respect to different bipartitions and derive all the remaining conditions in Eqs. (\[rtg\])-(\[rtg3\]). The extension to asymptotic simulations of bosonic channels is achieved by adopting the weaker definition of the REE. $\blacksquare$
It is immediate to specify the result for distillable networks for which we may directly write $$E_{\mathrm{R}}^{\text{m}}(C)=\mathcal{C}^{\text{m}}(C):=\sum
\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}\mathcal{C}_{\mathbf{xy}},$$ where $\mathcal{C}^{\text{m}}(C)$ is the multi-edge capacity of cut $C$. We may write the following immediate consequence.
Consider a multiple-unicast quantum network $\mathcal{N}$ with $M$ sender-receiver pairs $(\mathbf{a}_{i},\mathbf{b}_{i})$ communicating via multipath routing. If the network is distillable, then we may write the following outer bounds for the capacity region$$\begin{aligned}
R_{i}^{\text{m}} & \leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{i}}\mathcal{C}^{\text{m}}(C)~~\text{for any }i,\\
R_{i}^{\text{m}}+R_{j}^{\text{m}} & \leq\min_{C:\mathbf{a}_{i}\mathbf{a}_{j}|\mathbf{b}_{i}\mathbf{b}_{j}}\mathcal{C}^{\text{m}}(C)~~\text{for any
}i\neq j\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i}^{\text{m}} & \leq\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{i}\}}\mathcal{C}^{\text{m}}(C),\end{aligned}$$ where $\mathcal{C}^{\text{m}}(C)$ is the multi-edge capacity of cut $C$.
Achievable lower bounds may be determined by combining the point-to-point composition strategy with classical routing algorithms based on the maximization of multiple flows. For the specific case $M=1$, the outer bound is achievable and we retrieve the max-flow min-cut theorem for quantum communications [@netpaper; @longVersion]. For $M>2$, achievable lower bounds may be found by exploiting classical literature on multicommodity flow algorithms, e.g., Ref. [@TCHu] which showed a version of the max-flow min-cut theorem for undirected networks with two commodities, and Ref. [@Schrijver] which discusses extensions to more than two commodities.
Multicast quantum networks\[SECmulticastSINGLE\]
================================================
Let us now consider a multicast scenario, where Alice $\mathbf{a}$ aims at simultaneously communicate generally-different messages to a set of $M$ receivers, i.e., a set of Bobs $\mathbf{\{b}_{i}\}$. Because of the implicit parallel nature of this communication process, it is directly formulated under the assumption of multipath routing. We can easily generalize the description of the one-sender one-receiver flooding protocol to the present case of multiple receivers.
In a $1$-to-$M$ multicast network protocol, the quantum network $\mathcal{N}$ is subject to an orientation where Alice is treated as a source, while the various Bobs are destination points, each one being a receiver but also a potential relay for another receiver (so that they are not necessarily sinks in the general case). Each end-to-end simultaneous communication between Alice and the Bobs consists of a sequence of multicasts from each point of the network to its out-neighborhood, assisted by network LOCCs. This is done in a flooding fashion so that each edge of the network is exploited. The orientation of the network may be updated and optimized at each round of the protocol.
The sequence of orientations and the network LOCCs define the multicast flooding protocol $\mathcal{P}_{\text{multicast}}^{\text{flood}}$. After $n$ uses of the network, Alice and the $M$ Bobs will share an output state $\rho_{\mathbf{a\{b}_{i}\}}^{n}$ which is $\varepsilon$-close to a target state $$\phi:={\textstyle\bigotimes_{i=1}^{M}}
\phi_{\mathbf{ab}_{i}}^{\otimes nR_{i}^{\varepsilon,n}}.$$ where $\phi_{\mathbf{ab}_{i}}$ is a one-bit state (private bit or ebit) for the pair of points $(\mathbf{a},\mathbf{b}_{i})$ and $nR_{i}^{\varepsilon,n}$ the number of its copies. Note that this is a compact notation which involves countable sets of systems $\mathbf{a}=(a,a^{\prime},a^{\prime\prime},\ldots)$ and $\mathbf{b}_{i}=(b_{i},b_{i}^{\prime},b_{i}^{\prime\prime},\ldots)$. Therefore, the tensor product $\phi_{\mathbf{ab}_{1}}^{\otimes nR_{1}^{\varepsilon,n}}\otimes\phi_{\mathbf{ab}_{2}}^{\otimes nR_{2}^{\varepsilon
,n}}$ explicitly means $\phi_{ab_{1}}^{\otimes nR_{1}^{\varepsilon,n}}\otimes\phi_{a^{\prime}b_{2}^{\prime}}^{\otimes nR_{2}^{\varepsilon,n}}$, so that there are different systems involved in Alice’s side.
By taking the limit of large $n$, small $\varepsilon$, and optimizing over $\mathcal{P}_{\text{multicast}}^{\text{flood}}$, we define the capacity region associated with the achievable rates $(R_{1},\ldots,R_{M})$. In particular, we may define a unique capacity which is associated with the symmetric condition $R_{1}=\ldots=R_{M}$ (or, more precisely, with a guaranteed common rate $R$ with each Bob). In fact, we may consider a symmetric type of protocol $\mathcal{\tilde{P}}_{\text{multicast}}^{\text{flood}}$ whose target state $\phi$ must have $nR_{i}^{\varepsilon,n}\geq nR_{\varepsilon,n}$ bits for any $i$. Then, by taking the asymptotic limit of large $n$ small $\varepsilon$, and maximizing over all such protocols, we may define the multicast network capacity$$\mathcal{C}^{M}(\mathcal{N})=\sup_{\mathcal{\tilde{P}}_{\text{multicast}}^{\text{flood}}}\lim_{\varepsilon,n}R_{\varepsilon,n}. \label{multiCC}$$ This rate quantifies the maximum number of target bits per network use (multipath transmission) that Alice may simultaneously share with each Bob in the destination set $\mathbf{\{b}_{i}\}$. We have the usual hierarchy $Q_{2}^{M}(\mathcal{N})=D_{2}^{M}(\mathcal{N})\leq K^{M}(\mathcal{N})$ when we specify the target state. We can now state the following general bound.
\[Quantum multicast\]\[TheomultiCAST\]Let us consider a multicast quantum network $\mathcal{N}$ with one sender and $M$ receivers $\mathbf{\{b}_{i}\}$. Adopt a simulation of the network with a resource representation $\sigma(\mathcal{N})=\{\sigma_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$. In particular, $\sigma(\mathcal{N})$ can be a Choi-representation for a teleportation-covariant $\mathcal{N}$. Then we have the following outer bounds for the capacity region $$\begin{aligned}
R_{i} & \leq E_{\mathrm{R}}^{\text{m}}(i):=\min_{C:\mathbf{a}|\mathbf{b}_{i}}E_{\mathrm{R}}^{\text{m}}(C)~~\text{for any }i,\label{cvcc}\\
R_{i}+R_{j} & \leq\min_{C:\mathbf{a}|\mathbf{b}_{i}\mathbf{b}_{j}}E_{\mathrm{R}}^{\text{m}}(C)~~\text{for any }i\neq j\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i} & \leq\min_{C:\mathbf{a}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}^{\text{m}}(C), \label{cvcc3}$$ where $E_{\mathrm{R}}^{\text{m}}(C)$ is the multi-edge flow of REE through cut $C$. In particular, the multicast network capacity satisfies$$\mathcal{C}^{M}(\mathcal{N})\leq\min_{i\in\{1,M\}}E_{\mathrm{R}}^{\text{m}}(i). \label{CMNmultics}$$ It is understood that formulations may be asymptotic for quantum networks with bosonic channels.
**Proof.** Consider the upper bound given by secret-key generation. With respect to the bipartition $\mathbf{a}|\{\mathbf{b}_{i}\}$, we may write the usual steps starting form the distillable key of the target state $$\begin{aligned}
K_{\mathrm{D}}(\mathbf{a}|\{\mathbf{b}_{i}\})_{\phi} & =n\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\\
& \leq E_{\mathrm{R}}(\mathbf{a}|\{\mathbf{b}_{i}\})_{\phi}\\
& \leq E_{\mathrm{R}}(\mathbf{a}|\{\mathbf{b}_{i}\})_{\rho^{n}}+\delta(\varepsilon,d),\end{aligned}$$ leading to the asymptotic limit$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\leq
\underset{n\rightarrow+\infty}{\lim}~n^{-1}E_{\mathrm{R}}(\mathbf{a}|\{\mathbf{b}_{i}\})_{\rho^{n}}. \label{repkkkj2}$$
For any cut $C:\mathbf{a}|\{\mathbf{b}_{i}\}$ of the (simulated) network, we may write the decomposition$$\rho_{\mathbf{a}\{\mathbf{b}_{i}\}}^{n}(C)=\bar{\Lambda}_{\mathbf{a}\{\mathbf{b}_{i}\}}^{C}\left[ \underset{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n}\right] ,$$ for some trace-preserving LOCC $\bar{\Lambda}_{\mathbf{a}\{\mathbf{b}_{i}\}}^{C}$. By replacing $\rho^{n}=\rho_{\mathbf{a}\{\mathbf{b}_{i}\}}^{n}(C)$ in Eq. (\[repkkkj2\]), we therefore get$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M}R_{i}^{\varepsilon,n}\leq
\sum\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}E_{\mathrm{R}}(\sigma
_{\mathbf{xy}}):=E_{\mathrm{R}}^{\text{m}}(C).$$ By minimizing over the cuts and maximizing over the protocols, we may write$$\sum\limits_{i=1}^{M}R_{i}\leq\min_{C:\mathbf{a}|\{\mathbf{b}_{i}\}}E_{\mathrm{R}}^{\text{m}}(C).$$
The other conditions in Eqs. (\[cvcc\])-(\[cvcc3\]) are obtained by setting part of the rates $R_{i}^{\varepsilon,n}$ to zero in the target state (as in the previous proofs). In particular, set $R_{i}^{\varepsilon,n}\neq0$ for some $i$, while $R_{j}^{\varepsilon,n}=0$ for any $j\neq i$. The target state becomes $\phi:=\phi_{\mathbf{ab}_{i}}^{\otimes nR_{i}^{\varepsilon,n}}\otimes\sigma_{\text{sep}}$ and we repeat the derivation with respect to the bipartition $\mathbf{a|b}_{i}$. This leads to $$\lim_{\varepsilon,n}R_{i}^{\varepsilon,n}\leq\underset{n\rightarrow+\infty
}{\lim}~n^{-1}E_{\mathrm{R}}(\mathbf{a}|\mathbf{b}_{i})_{\rho^{n}},
\label{multi11}$$ where we may directly consider the reduced state $$\rho^{n}=\rho_{\mathbf{ab}_{i}}^{n}=\mathrm{Tr}_{\{\mathbf{b}_{j\neq i}\}}\left[ \rho_{\mathbf{a}\{\mathbf{b}_{1},\ldots,\mathbf{b}_{M}\}}^{n}\right] .$$
For any cut $C:\mathbf{a}|\mathbf{b}_{i}$, we therefore have$$\rho_{\mathbf{ab}_{i}}^{n}(C)=\bar{\Lambda}_{\mathbf{ab}_{i}}^{C}\left[
\underset{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}~\sigma_{\mathbf{xy}}^{\otimes n}\right] , \label{multi22}$$ which leads to $\lim_{\varepsilon,n}R_{i}^{\varepsilon,n}\leq E_{\mathrm{R}}^{\text{m}}(C)$. By minimizing over the cuts, one gets $$\lim_{\varepsilon,n}R_{i}^{\varepsilon,n}\leq E_{\mathrm{R}}^{\text{m}}(i):=\min_{C:\mathbf{a}|\mathbf{b}_{i}}E_{\mathrm{R}}^{\text{m}}(C).
\label{bbbmmm}$$ Since this is true for any protocol $\mathcal{P}^{M}$, it can be extended to the achievable rates, i.e., we get Eq. (\[cvcc\]).
For the multicast network capacity, just note that $$\lim_{\varepsilon,n}R^{\varepsilon,n}\leq\min_{i}\{\lim_{\varepsilon,n}R_{i}^{\varepsilon,n}\}.$$ Therefore, from Eq. (\[bbbmmm\]), we may write$$\lim_{\varepsilon,n}R^{\varepsilon,n}\leq\min_{i}E_{\mathrm{R}}^{\text{m}}(i).$$ This is true for any symmetric protocol $\mathcal{P}_{\text{sym}}^{M}$ which leads to the result of Eq. (\[CMNmultics\]). Results are extended to asymptotic simulations of bosonic channels in the usual way. $\blacksquare$
As usual, in the case of distillable networks, we may prove stronger results. As a direct consequence of Theorem \[TheomultiCAST\], we may write the following cutset bound.
Consider a multicast quantum network $\mathcal{N}$ with one sender and $M$ receivers $\mathbf{\{b}_{i}\}$. If the network is distillable, then we have the following outer bounds for the capacity region $$\begin{aligned}
R_{i} & \leq\mathcal{C}^{\text{m}}(i)=\min_{C:\mathbf{a}|\mathbf{b}_{i}}\mathcal{C}^{\text{m}}(C)~~\text{for any }i,\\
R_{i}+R_{j} & \leq\min_{C:\mathbf{a}|\mathbf{b}_{i}\mathbf{b}_{j}}\mathcal{C}^{\text{m}}(C)~~\text{for any }i\neq j\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M}R_{i} & \leq\min_{C:\mathbf{a}|\{\mathbf{b}_{i}\}}\mathcal{C}^{\text{m}}(C),\end{aligned}$$ where $\mathcal{C}^{\text{m}}(C)$ is the multi-edge capacity of cut $C$ and $\mathcal{C}^{\text{m}}(i)$ is the multipath capacity between the sender and the $i$th receiver (in a unicast setting). In particular, the multicast network capacity must satisfy the bound$$\mathcal{C}^{M}(\mathcal{N})\leq\min_{i\in\{1,M\}}\mathcal{C}^{\text{m}}(i)~.
\label{cutsetDIST}$$
Our previous results refer to the general case of multiple independent messages. In a multicast quantum network, this means that Alice distributes $M$ different sequences of target bits to the $M$ Bobs $\{\mathbf{b}_{i}\}$. For instance, these may represent $M$ different secret keys, one for each Bob in the destination set. For this specific task (key distribution), the multicast capacity of the network $\mathcal{C}^{M}(\mathcal{N})$ becomes a multicast secret-key capacity $\mathcal{K}^{M}(\mathcal{N})$.
In QKD, it is interesting to consider the variant scenario where Alice distributes exactly the same secret key to all Bobs $\{\mathbf{b}_{i}\}$, e.g., to enable a quantum-secured conference among these parties. For this particular task, we may define a single-key version of the multicast secret-key capacity, that we denote as $\mathcal{K}_{\text{1-key}}^{M}(\mathcal{N})$. This represents the maximum rate at which Alice may distribute the same secret key to all Bobs in each parallel use of the network. It is clear that we have $\mathcal{K}_{\text{1-key}}^{M}(\mathcal{N})\geq\mathcal{K}^{M}(\mathcal{N})$, just because Alice may use $M-1$ distributed keys to encrypt and share the shortest key with all Bobs.
Multiple-multicast quantum networks\[SECmulticastMANY\]
=======================================================
In the multiple-multicast quantum network, we have $M_{A}$ Alices $\{\mathbf{a}_{1},\ldots,\mathbf{a}_{i},\ldots,\mathbf{a}_{M_{A}}\}$, each of them communicating generally-different messages with a destination set of $M_{B}$ Bobs $\{\mathbf{b}_{1},\ldots,\mathbf{b}_{j},\ldots,\mathbf{b}_{M_{B}}\}$ by means of multipath routing. Each end-to-multiend multicast $\mathbf{a}_{i}\rightarrow\{\mathbf{b}_{j}\}$ is associated with the distribution of $M_{B}$ independent sequences of target bits (e.g., secret keys) between the $i$th Alice $\mathbf{a}_{i}$ and each Bob $\mathbf{b}_{j}$ in the destination set. The description of a multiple-multicast protocol for a quantum network follows the same main features discussed for the case of a single-multicast network ($M_{A}=1$). Because we have multiple senders and receivers, here we need to consider all possible orientations of the network. Each use of the quantum network is performed under some orientation which is adopted by the points for their out-neighborhood multicasts, suitably assisted by network LOCCs. Use after use, these steps define a multiple-multicast flooding protocol $\mathcal{P}_{\text{M-multicast}}^{\text{flood}}$.
After $n$ uses, the ensembles of Alices and Bobs share an output state $\rho_{\{\mathbf{a}_{i}\}\mathbf{\{b}_{j}\}}^{n}$ which is $\varepsilon$-close to a target state $$\phi:={\textstyle\bigotimes_{i=1}^{M_{A}}}
{\textstyle\bigotimes_{j=1}^{M_{B}}}
\phi_{\mathbf{a}_{i}\mathbf{b}_{j}}^{\otimes nR_{ij}^{\varepsilon,n}}.$$ where $\phi_{\mathbf{a}_{i}\mathbf{b}_{j}}$ is a one-bit state (private bit or ebit) for the pair $(\mathbf{a}_{i},\mathbf{b}_{j})$ and $nR_{ij}^{\varepsilon,n}$ the number of its copies. By taking the limit of large $n$, small $\varepsilon$, and optimizing over $\mathcal{P}_{\text{M-multicast}}^{\text{flood}}$, we define the capacity region for the achievable rates $\{R_{ij}\}$. Assume the symmetric case where the $i$th Alice $\mathbf{a}_{i}$ achieves the same rate $R_{i1}=\ldots=R_{iM_{B}}$ with all Bobs $\{\mathbf{b}_{j}\}$ (or, more precisely, a guaranteed common rate $R_{i}$). This means to consider symmetric protocols whose target state $\phi$ must have $\min
_{j}R_{ij}^{\varepsilon,n}\geq R_{i}^{\varepsilon,n}$ bits for any $i$. By taking the asymptotic limit of $R_{i}^{\varepsilon,n}$ for large $n$, small $\varepsilon$, and maximizing over all these symmetric protocols, we may define the capacity region for the achievable multicast rates $(R_{1},\ldots,R_{M_{A}})$. In the latter set, rate $R_{i}$ provides the minimum number of target bits per use that the $i$th Alice may share with each Bob in the destination set $\mathbf{\{b}_{j}\}$ (in the multi-message setting, i.e., assuming independent sequences shared with the various Bobs). We have the following outer bounds to the capacity region.
\[Quantum multiple-multicast\]\[Theomultimulti\]Let us consider a multiple-multicast quantum network $\mathcal{N}=(P,E)$ where each of the $M_{A}$ senders $\mathbf{\{a}_{i}\}$ communicates with $M_{B}$ receivers $\mathbf{\{b}_{j}\}$ at the multicast rate $R_{i}$. Adopt a simulation of $\mathcal{N}$ with some resource representation $\sigma(\mathcal{N})=\{\sigma_{\mathbf{xy}}\}_{(\mathbf{x},\mathbf{y})\in E}$, which may be a Choi-representation for a teleportation-covariant $\mathcal{N}$. Then, we have the following outer bounds for the capacity region$$\begin{aligned}
R_{i} & \leq\min_{\substack{C:\mathbf{a}_{i}\in\mathbf{A}\\\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset}}E_{\mathrm{R}}^{\text{m}}(C),\label{setCVB0}\\
R_{i}+R_{j} & \leq\min_{\substack{C:\mathbf{a}_{i},\mathbf{a}_{j}\in\mathbf{A}\\\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset}}E_{\mathrm{R}}^{\text{m}}(C),\\
& \vdots\nonumber\\
\sum\limits_{i=1}^{M_{A}}R_{i} & \leq\min_{\substack{C:\{\mathbf{a}_{i}\}\subseteq\mathbf{A}\\\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset
}}E_{\mathrm{R}}^{\text{m}}(C), \label{setCVB}$$ where $E_{\mathrm{R}}^{\text{m}}(C)$ is the multi-edge flow of REE through cut $C$. For a distillable network, we may write the bounds in Eqs. (\[setCVB0\])-(\[setCVB\]) with $E_{\mathrm{R}}^{\text{m}}(C)=\mathcal{C}^{\text{m}}(C)$, i.e., in terms of the multi-edge capacity of the cuts.
**Proof.** Consider the upper bound given by secret-key generation. With respect to the bipartition $\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\}$, we can manipulate the distillable key $K_{\mathrm{D}}$ of the target state $\phi$ as follows $$\begin{aligned}
K_{\mathrm{D}}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\})_{\phi} &
=n\sum\limits_{i=1}^{M_{A}}\sum\limits_{j=1}^{M_{B}}R_{ij}^{\varepsilon,n}\\
& \leq E_{\mathrm{R}}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\})_{\phi}\\
& \leq E_{\mathrm{R}}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\})_{\rho^{n}}+\delta(\varepsilon,d),\end{aligned}$$ leading to the asymptotic limit$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M_{A}}\sum\limits_{j=1}^{M_{B}}R_{ij}^{\varepsilon,n}\leq\underset{n\rightarrow+\infty}{\lim}~n^{-1}E_{\mathrm{R}}(\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\})_{\rho^{n}}.
\label{repkkkj3}$$
For any cut $C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\}$ of the (simulated) network, we may write the decomposition$$\rho_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{j}\}}^{n}(C)=\bar{\Lambda}_{\{\mathbf{a}_{i}\}\{\mathbf{b}_{j}\}}^{C}\left[ \underset{(\mathbf{x},\mathbf{y})\in\tilde{C}}{{\textstyle\bigotimes}
}\sigma_{\mathbf{xy}}^{\otimes n}\right] ,$$ and manipulate Eq. (\[repkkkj3\]) into the following$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M_{A}}\sum\limits_{j=1}^{M_{B}}R_{ij}^{\varepsilon,n}\leq\sum\limits_{(\mathbf{x},\mathbf{y})\in\tilde{C}}E_{\mathrm{R}}(\sigma_{\mathbf{xy}}):=E_{\mathrm{R}}^{\text{m}}(C).$$ By minimizing over the cuts and maximizing over the protocols, we may write$$\sum\limits_{i=1}^{M_{A}}\sum\limits_{j=1}^{M_{B}}R_{ij}\leq\min
_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\}}E_{\mathrm{R}}^{\text{m}}(C)~.$$
By setting part of the rates $R_{ij}^{\varepsilon,n}$ to zero in the target state, we derive the full set of conditions $$\begin{aligned}
\sum\limits_{i=1}^{M_{A}}\sum\limits_{j=1}^{M_{B}}R_{ij} & \leq
\min_{C:\{\mathbf{a}_{i}\}|\{\mathbf{b}_{j}\}}E_{\mathrm{R}}^{\text{m}}(C),\label{xzx1}\\
& \vdots\nonumber\\
R_{ij}+R_{kl} & \leq\min_{C:\mathbf{a}_{i}\mathbf{a}_{k}|\mathbf{b}_{j}\mathbf{b}_{l}}E_{\mathrm{R}}^{\text{m}}(C),\label{xzx2}\\
R_{ij} & \leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{j}}E_{\mathrm{R}}^{\text{m}}(C). \label{xzx3}$$
The latter conditions are valid for the end-to-end rates $R_{ij}$ achievable between each pair $(\mathbf{a}_{i},\mathbf{b}_{j})$. We are interested in the achievable multicast rates $\{R_{i}\}$ between each sender $\mathbf{a}_{i}$ and all receivers $\{\mathbf{b}_{j}\}$. Corresponding conditions can be derived by considering a subset of protocols with target state of the type $$\phi_{k}:={\textstyle\bigotimes_{i=1}^{M_{A}}}
\phi_{\mathbf{a}_{i}\mathbf{b}_{k}}^{\otimes nR_{ik}^{\varepsilon,n}}\otimes\sigma_{\text{sep}},$$ for some $k$, where all Alices $\{\mathbf{a}_{i}\}$ aim to optimize their rates $\{R_{ik}^{\varepsilon,n}\}$ with some fixed Bob $\mathbf{b}_{k}$, so that $R_{ij}^{\varepsilon,n}=0$ for any $j\neq k$. By repeating the previous steps with respect to the bipartition $\{\mathbf{a}_{i}\}|\mathbf{b}_{k}$, we obtain$$\lim_{\varepsilon,n}\sum\limits_{i=1}^{M_{A}}R_{ik}^{\varepsilon,n}\leq
\min_{C:\{\mathbf{a}_{i}\}|\mathbf{b}_{k}}E_{\mathrm{R}}^{\text{m}}(C).$$ Since we have $R_{i}^{\varepsilon,n}\leq\min_{j}R_{ij}^{\varepsilon,n}\leq
R_{ik}^{\varepsilon,n}$ for any $i$, we can then write the same inequality for $\lim_{\varepsilon,n}\sum\nolimits_{i=1}^{M_{A}}R_{i}^{\varepsilon,n}$. Then, by optimizing over the protocols, we get$$\sum\limits_{i=1}^{M_{A}}R_{i}\leq\min_{C:\{\mathbf{a}_{i}\}|\mathbf{b}_{k}}E_{\mathrm{R}}^{\text{m}}(C).$$ Because we may repeat the previous reasoning for any $k$, we may write$$\sum\limits_{i=1}^{M_{A}}R_{i}\leq\min_{C}E_{\mathrm{R}}^{\text{m}}(C),
\label{yyu1}$$ with $C=(\mathbf{A},\mathbf{B})$ such that $\{\mathbf{a}_{i}\}\subseteq
\mathbf{A}$ and $\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset$.
Now, for any fixed $k$, impose that the rates $\{R_{ik}^{\varepsilon,n}\}$ are zero for some of the Alices $\{\mathbf{a}_{i}\}$. If we set $R_{ik}^{\varepsilon,n}\neq0$ for a pair $(\mathbf{a}_{i},\mathbf{b}_{k})$, then the condition $R_{i}^{\varepsilon,n}\leq R_{ik}^{\varepsilon,n}$ leads to$$R_{i}\leq\min_{C:\mathbf{a}_{i}|\mathbf{b}_{k}}E_{\mathrm{R}}^{\text{m}}(C).$$ Because the reasoning can be repeated for any $k$, we may then write$$R_{i}\leq\min_{C}E_{\mathrm{R}}^{\text{m}}(C), \label{yyu2}$$ with $C=(\mathbf{A},\mathbf{B})$ such that $\mathbf{a}_{i}\in\mathbf{A}$ and $\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset$. Extending the previous reasoning to two non-zero rates $R_{ik}^{\varepsilon,n}\neq0$ and $R_{jk}^{\varepsilon,n}\neq0$ leads to$$R_{i}+R_{j}\leq\min_{C}E_{\mathrm{R}}^{\text{m}}(C), \label{yyu3}$$ with $C=(\mathbf{A},\mathbf{B})$ such that $\mathbf{a}_{i},\mathbf{a}_{j}\in\mathbf{A}$ and $\{\mathbf{b}_{j}\}\cap\mathbf{B}\neq\emptyset$. Other similar conditions can be derived for the multicast rates, so that we get the result of Eqs. (\[setCVB0\])-(\[setCVB\]). Finally, for a distillable network we have $E_{\mathrm{R}}^{\text{m}}(C)=\mathcal{C}^{\text{m}}(C)$ and, therefore, it is immediate to express these results in terms of the multi-edge capacities of the cuts. $\blacksquare$
Conclusions\[SECconclusions\]
=============================
In this information-theoretic work, we have investigated the ultimate rates for transmitting quantum information, distributing entanglement and generating secret keys between multiple senders and receivers in an arbitrary quantum network, assuming single- or multi-path routing strategies. We have established general single-letter REE upper bounds for the various multiend capacities associated to the various configurations of multiple-unicast, multicast, and multiple-multicast quantum networks. The bounds apply to networks connected by arbitrary quantum channels (at any dimension) with more specific formulations in the case of teleportation-covariant and distillable channels. In particular, the case of quantum networks connected by bosonic channels is implicitly treated by using asymptotic LOCC simulations, so that the results are automatically proven for fundamental noise models at the optical and telecom regimes, such as the pure-loss channels.
The present paper provides a multiend generalization of the results of Ref. [@netpaper] (first appeared in Ref. [@longVersion]) for the basic end-to-end (unicast) scenario. It also extends the results presented in Ref. [@Ric1] from single-hop to multi-hop quantum networks. Due to the much more complex scenarios associated with the simultaneous multi-hop quantum communication among multiple senders and receivers, we could only bound the capacity regions in the various configurations analyzed, so that their full characterization remains an open question for further investigations.
**Acknowledgments**. This work has been supported by the EPSRC via the ‘UK Quantum Communications HUB’ (EP/M013472/1) and by the European Union via the project ‘Continuous Variable Quantum Communications’ (CiViQ, no 820466).
[99]{}
M. A. Nielsen, and I. L. Chuang, *Quantum computation and quantum information* (Cambridge University Press, Cambridge, 2000).
S. L. Braunstein and P. van Loock, Rev. Mod. Phys. **77**, 513 (2005).
A. Holevo, *Quantum Systems, Channels, Information: A Mathematical Introduction* (De Gruyter, Berlin-Boston, 2012).
C. Weedbrook *et al.*, Rev. Mod. Phys. **84**, 621 (2012).
J. Watrous, *The theory of quantum information* (Cambridge University Press, Cambridge, 2018).
C. H. Bennett and G. Brassard, Proc. IEEE International Conf. on Computers, Systems, and Signal Processing, Bangalaore, pp. 175–179 (1984).
A. K. Ekert, Phys. Rev. Lett. **67**, 661-663 (1991).
S. Pirandola, U. L. Andersen, L. Banchi, M. Berta, D. Bunan-dar, R. Colbeck, D. Englund, T. Gehring, C. Lupo, C. Ottaviani, J. Pereira, M. Razavi, J. S. Shaari, M. Tomamichel, V. C. Usenko, G. Vallone, P. Villoresi, and P. Wallden, “Advances in quantum cryptography”, preprint arXiv:1906.01645 (2019).
H. J. Kimble, Nature **453**, 1023-1030 (2008).
S. Pirandola, and S. L. Braunstein, Nature **532**, 169-171 (2016).
S. Wehner, D. Elkouss, and R. Hanson, Science **362**, 303 (2018).
R. Van Meter, *Quantum Networking* (Wiley, 2014).
S. Pirandola, R. Laurenza, C. Ottaviani and L. Banchi, Nat. Commun. **8**, 15043 (2017). See also arXiv:1510.08863 (2015).
S. Pirandola, R. García-Patrón, S. L. Braunstein, and S. Lloyd, Phys. Rev. Lett. **102**, 050503 (2009).
R. García-Patrón, S. Pirandola, S. Lloyd, and J. H. Shapiro, Phys. Rev. Lett. **102**, 210501 (2009).
M. Lucamarini, Z. L. Yuan, J. F. Dynes, and A. J. Shields, Nature (London) **557**, 400-403 (2018).
H.-J. Briegel, W. Dür, J. I. Cirac, and P. Zoller, Phys. Rev. Lett. **81**, 5932-5935 (1998).
W. Dür, H.-J. Briegel, J. I. Cirac, and P. Zoller, Phys. Rev. A **59**, 169 (1999).
L. M. Duan, M. D. Lukin, J. I. Cirac, and P. Zoller, Nature (London) **414**, 413 (2001).
S. Pirandola, *End-to-end capacities of a quantum communication network*, Commun. Phys. **2**, 51 (2019).
S. Pirandola, *Capacities of repeater-assisted quantum communications*, arXiv:1601.00966 (2016).
P. Slepian, *Mathematical Foundations of Network Analysis* (Springer-Verlag, New York, 1968).
A. Schrijver, *Combinatorial Optimization* (Springer-Verlag, Berlin, 2003).
A. El Gamal and Y.-H. Kim, *Network Information Theory*, (Cambridge Univ. Press, 2011).
T. M. Cover and J. A. Thomas, *Elements of Information Theory*, (Wiley, New Jersey, 2006).
R. K. Ahuja, T. L. Magnanti, and J. B. Orlin, *Network Flows: Theory, Algorithms and Applications* (Prentice Hall, 1993).
S. Pirandola, and C. Lupo, Phys. Rev. Lett. **118**, 100502 (2017).
T. P. W. Cope, L. Hetzel, L. Banchi, and S. Pirandola, Phys. Rev. A **96**, 022323 (2017).
S. Pirandola, S. L. Braunstein, R. Laurenza, C. Ottaviani, and L. Banchi, Quant. Sci. Tech. **3**, 035009 (2018).
S. Pirandola, R. Laurenza, and S. L. Braunstein, Eur. Phys. J. D **72**, 162 (2018).
R. Laurenza, C. Lupo, G. Spedalieri, S. L. Braunstein, and S. Pirandola, Quantum Meas. Quantum Metrol. **5**, 1-12 (2018).
S. Pirandola, B. Roy Bardhan, T. Gehring, C. Weedbrook, and S. Lloyd, Nat. Photon. **12**, 724-733 (2018).
V. Vedral, Rev. Mod. Phys. **74**, 197 (2002).
V. Vedral, M. B. Plenio, M. A. Rippin, and P. L. Knight, Phys. Rev. Lett. **78**, 2275-2279 (1997).
V. Vedral, and M. B. Plenio, Phys. Rev. A **57**, 1619 (1998).
K. Horodecki, M. Horodecki, P. Horodecki, and J. Oppenheim, Phys. Rev. Lett. **94**, 160502 (2005).
C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, and W. K. Wootters, Phys. Rev. Lett. **70**, 1895-1899 (1993).
S. L. Braunstein, and H. J. Kimble, ** Phys. Rev. Lett. **80**, 869–872 (1998).
S. L. Braunstein, G. M. D’Ariano, G. J. Milburn, and M. F. Sacchi, Phys. Rev. Lett. **84**, 3486–3489 (2000).
S. Pirandola, J. Eisert, C. Weedbrook, A. Furusawa, and S. L. Braunstein, Nature Photon. **9**, 641-652 (2015).
M. Christiandl, A. Ekert, M. Horodecki, P. Horodecki, J. Oppenheim, and R. Renner, Lecture Notes in Computer Science **4392**, 456-478 (2007). See also arXiv:quant-ph/0608199v3 for a more extended version.
M. Christiandl, N. Schuch, and A. Winter, Comm. Math. Phys. **311**, 397-422 (2012).
B. Schumacher and M. A. Nielsen, Phys. Rev. A **54**, 2629 (1996).
S. Lloyd, Phys. Rev. A **55,** 1613 (1997).
A. S. Tanenbaum and D. J. Wetherall, *Computer Networks* (5th Edition, Pearson, 2010).
T. E. Harris, and F. S. Ross, Research Memorandum, Rand Corporation (1955).
L. R. Ford, and D. R. Fulkerson, Canadian Journal of Mathematics **8**, 399 (1956).
P. Elias, A. Feinstein, and C. E. Shannon, IRE Trans. Inf. Theory **2**, 117–119 (1956).
J. Edmonds and R. M. Karp, Journal of the ACM **19**, 248–264 (1972).
E. A. Dinic, Soviet Math. Doklady (Doklady) **11**, 1277–1280 (1970).
T. C. Hu, Oper. Res. **11**, 344-360 (1963).
R. Laurenza, and S. Pirandola, Phys. Rev. A **96**, 032318 (2017).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We make a model-independent measurement of the moment of inertia of a rotating, expanding strongly-interacting Fermi gas. Quenching of the moment of inertia is observed for energies both below and above the superfluid transition. This shows that a strongly interacting Fermi gas with angular momentum can support irrotational flow in both the superfluid and collisional normal fluid regimes.'
author:
- 'Bason Clancy, Le Luo, and John E. Thomas'
title: |
Observation of a quenched moment of inertia\
in a rotating strongly interacting Fermi gas
---
Strongly interacting Fermi gases [@OHaraScience] provide a unique paradigm for exploring strongly interacting nearly perfect fluids in nature, from high temperature superfluids and superconductors to exotic normal fluids, such as the quark-gluon plasma of the Big Bang [@Heinz; @Shuryak], or low viscosity quantum fields [@Son]. Rotating superfluids require irrotational flow, which quenches the moment of inertia or produces a vortex lattice [@LondonRot]. Recently, the observation of vortices has directly demonstrated superfluidity in a strongly interacting Fermi gas [@KetterleVortices]. However, the moment of inertia has not been measured and the rotational properties have not been characterized in the normal regime, where irrotational flow is not required.
We report a measurement of the moment of inertia $I$ of a strongly interacting Fermi gas of $^6$Li atoms in both the superfluid and normal fluid regimes, by releasing rotating clouds from a cigar-shaped optical trap. In the superfluid regime, a rapid increase in the angular velocity is observed as the cloud expands, indicating a quenching of $I$ to values as low as 0.05 of the rigid body value. However, quenching persists for energies far above the superfluid transition, in contrast to previous measurements spanning half a century, in nuclei [@Migdal; @WintherNuclei], liquid helium [@FairbanksHe] and Bose-Einstein condensates (BEC’s) [@Foote; @Inguscio], where irrotational behavior was attributed only to superfluidity. Our results demonstrate that a strongly interacting Fermi gas with angular momentum can support irrotational flow not only in the superfluid regime, but also in the normal fluid regime, which we attribute to nearly perfect collisional hydrodynamics.
Strongly interacting Fermi gases exhibit hydrodynamic behavior in both the superfluid and normal fluid regimes. While superfluidity explains the low temperature hydrodynamics [@Kinast], the origin of nearly ideal flow in the normal fluid remains an open question [@BruunViscous]. The hydrodynamics of the gas has been studied in expansion dynamics [@OHaraScience; @SalomonExpInt; @JinExpansion; @MITvortexexpansion], and in collective modes [@Kinast; @KinastDampTemp; @Bartenstein; @GrimmPrecmeas], but is also predicted to dramatically affect the rotational properties [@StingariSlowFermi], which have been investigated only in vortices [@KetterleVortices; @MITvortexexpansion]. Vortices demonstrate irrotational hydrodynamic flow, but only form at temperatures well below the superfluid transition [@KetterleVortices]. In contrast, measurement of the moment of inertia provides a probe of the rotational properties in both the normal and superfluid regimes.
In our experiments, a strongly interacting Fermi gas is prepared using a 50:50 mixture of the two lowest hyperfine states of $^6$Li atoms in an ultrastable CO$_2$ laser trap with a bias magnetic field tuned to a broad Feshbach resonance at $B = 834$ G [@BartensteinFeshbach]. At 834 G, the gas is cooled to quantum degeneracy through forced evaporation by lowering the trap depth $U$ [@OHaraScience]. Then $U$ is recompressed to $U_0/k_B=100\,\mu$K, which is large compared to the energy per particle of the gas.
At the final trap depth $U_0$, the measured oscillation frequencies in the transverse directions are $\omega_x=2\pi\times
2354(4)$ Hz and $\omega_y=2\pi\times 1992(2)$ Hz, while the axial frequency is $\omega_z=2\pi\times 71.1(.3)$ Hz, producing a cigar-shaped trap with $\lambda=\omega_z/\omega_x=0.030$. The total number of atoms at is typically $N = 1.3\times 10^5$. The corresponding Fermi energy $E_F$ and Fermi temperature $T_F$ for an ideal (noninteracting) harmonically trapped gas at the trap center are $E_F=k_B T_F=\hbar\,\bar{\omega}(3N)^{1/3}$, where $\bar{\omega}=(\omega_x\omega_y\omega_z)^{1/3}$. For our trap conditions, $T_F = 2.4\,\mu$K.
After evaporation and recompression, we typically achieve energies of $E = 0.56\,E_F$, which is close to the ground state energy of $E
= 0.50\,E_F$ [@Entropy]. To heat the sample, energy is added by releasing and then recapturing the gas, after which the gas is allowed to reach equilibrium for 0.5 s. The total energy $E$ of the cloud is determined in the universal, strongly interacting regime from the mean square axial (z) cloud size, using $E=3
m\omega_z^2\,\langle z^2\rangle$, where $m$ is the atom mass [@Entropy; @ThomasUniversal].
Once the gas has been prepared in the desired energy state, it is rotated by suddenly changing the direction of CO$_2$ laser beam as shown in Fig. \[fig:fig1\]. Rotation of the cigar-shaped CO$_2$ laser trap is accomplished by changing the frequency of the acousto-optic modulator (AOM) that controls the trap laser intensity, using a radiofrequency (rf) switch. When the frequency is changed, the direction of the trapping laser beam changes, causing the position of the beam on the final focusing lens to translate. This translation causes primarily a rotation of the cigar-shaped trap at the focal point, about an axis (y) perpendicular to the plane of the cigar-shaped trap. A scissors mode [@StringariScissors] is excited by the rotation. Then the cloud is permitted to oscillate in the trap for a chosen period that determines the initial angular velocity of the cloud before release.
![Scheme to rotate the optical trap. The optical trap is rotated by changing the frequency of an acoustooptic modulator(AOM). The rotation adds angular momentum to the trapped cloud before release. []{data-label="fig:fig1"}](Fig10514.eps){width="3.5in"}
Fig. \[fig:fig2\] shows cloud images as a function of expansion time for the coldest samples, with a typical energy $E = 0.56
E_F$ near the ground state. When the gas is released without rotation of the trap, Fig. \[fig:fig2\] (top), the Fermi cloud expands anisotropically, as previously predicted [@Menotti] and observed [@OHaraScience; @SalomonExpInt]. In that case, the gas expands rapidly in the narrow (x,y) directions of the cigar, while remaining nearly stationary in the long (z) direction, inverting the aspect ratio $\sigma_x/\sigma_z$ as the cloud becomes elliptical in shape.
{width="7.0in"}
Quite different expansion dynamics occurs when the cloud is rotating prior to release, Fig. \[fig:fig2\] (middle) and (bottom), which demonstrates irrotational flow in a nearly perfect hydrodynamic regime. In this case, the aspect ratio $\sigma_x/\sigma_z$ initially increases toward unity. However, as the aspect ratio approaches unity, the moment of inertia decreases and the angular velocity of the principal axes increases to conserve angular momentum as previously predicted [@StringariIrrot] and observed [@Foote; @Inguscio] in a superfluid BEC. After the aspect ratio reaches a maximum less than unity [@StringariIrrot], it and the angular velocity begin to decrease as the angle of the cigar shaped cloud approaches a maximum value less than 90$^\circ$.
Fig. \[fig:fig3\] shows the measured aspect ratio and the angle of the principal axes versus expansion time, which are determined from the cloud images. The measured density profiles are fit with a two-dimensional gaussian distribution, which takes the form $A\,
exp[-a z^2- b z x-c x^2]$. From the values of a, b, and c, the aspect ratio of the rotated cloud and the angle of the long axis of the cloud with respect to the laboratory z-axis are determined.
{height="2.5"}
We attempt to model the data for measurements near the ground state (blue solid circles and green triangles of Fig. \[fig:fig3\]), by using a zero temperature hydrodynamic theory for the expansion of a rotating strongly-interacting Fermi gas in the superfluid regime. A theory of this type was first used to describe the rotation and expansion of a weakly interacting BEC [@StringariCritical; @StringariIrrot]. The model consists of the Euler and continuity equations for a superfluid, where the velocity field $\mathbf{v}$ is irrotational, i.e., $\nabla \times
\mathbf{v}=0$. The driving force for the expansion arises from the gradient of the chemical potential, which we take to be that of a strongly interacting Fermi gas [@OHaraScience; @StingariSlowFermi]. We also include the force arising from magnet field curvature, which changes the angular momentum at the point of maximum aspect ratio by 10$\%$ and the angle and aspect ratio at the longest release times by a few percent. To determine the initial conditions for our model, we directly measure the initial angular velocity and axial cloud radius just after release, while assuming the transverse radii are given by zero temperature values for our trap frequencies. The results yield excellent agreement with all of the Fermi gas angle and aspect ratio data, with no free parameters, as shown in Fig. \[fig:fig3\].
We make a model-independent measurement of the effective moment of inertia $I\equiv L/\Omega$, where $\Omega$ is the angular velocity of the principal axes of the cloud after release and $L=\Omega_0\,
I_0$ is the angular momentum, which is conserved during the expansion (we neglect the small change arising from the magnetic potential). The angular velocity $\Omega$ is calculated from the time derivative of a polynomial fit to the angle versus time data. To determine the initial moment of inertia $I_0$, we note that for a cigar-shaped cloud with a small aspect ratio $\sigma_x/\sigma_z$, the moment of inertia for the irrotational fluid is nearly equal to the rigid body value [@Inguscio; @StringariIrrot]. For our parameters $I_0\simeq N m \langle z^2 \rangle_0$, within 0.3% accuracy, where $\langle z^2 \rangle_0$ is measured with respect to the principal axes of the cloud. Correspondingly, the initial angular momentum of the cloud is essentially equal to the rigid body value just after release. The measured effective moment of inertia after release is then $I=I_0\, \Omega_0/\Omega $. The corresponding rigid body moment of inertia is determined from the fit to the cloud profile, $I_{rig}=N m \langle x^2+z^2 \rangle$. Hence, we obtain $I/I_{rig}=(\Omega_0/\Omega)I_0/I_{rig}$.
![Quenching of the moment of inertia versus initial angular velocity $\Omega_0$. $I_{min}$ is the minimum moment of inertia measured during expansion. $I_{rig}$ is the rigid body moment of inertia corresponding to the cloud profile. Blue solid circles- initial energy before rotation below the superfluid transition energy $E_c =0.94\,E_F$. Red open circles- initial energy before rotation above the superfluid transition energy. Insert shows the energy for each data point. The dashed line shows the superfluid transition energy $E_c$.[]{data-label="fig:fig4"}](Fig40514.eps){width="3.5in"}
Fig. \[fig:fig4\] shows the measured minimum value of $I/I_{rig}$ as a function of initial angular velocity $\Omega_0$. The smallest values of $I/I_{rig}$ occur for the smallest $\Omega_0$. For the coldest clouds (blue solid circles), where the energy of the gas is close to that of the ground state, the gas is believed to be in the superfluid regime [@KetterleVortices; @Kinast; @Entropy]. In this case, we observe values of $I/I_{rig}$ as small as 0.05, smaller than those obtained from the scissors mode of a BEC of atoms [@StringariMI; @FootMI]. The solid line shows $I/I_{rig}$ as predicted by the superfluid hydrodynamic theory, which is in very good agreement with the measurements.
Such nearly perfect irrotational flow usually arises only in the superfluid regime. For example, normal weakly interacting Bose gases expand ballistically above the critical temperature. We observe ballistic expansion of the Fermi gas at 528 G, where the scattering length vanishes. In this case, after release of the rotating cloud, there is no evidence of irrotational hydrodynamics. The aspect ratio asymptotically approaches unity, and there is no increase in angular velocity.
In contrast, for a normal strongly interacting Fermi gas, we observe quenching of the moment of inertia. To investigate the normal fluid regime, we increase $E$ above the transition energy, which we estimate to be $E_c = 0.94\,E_F$ [@Entropy]. The open red circles in Fig. \[fig:fig3\], show the aspect ratio and angle versus time for $E = 2.1\,E_F$ and an initial angular velocity $\Omega_0/\omega_z=0.4$. The results are nearly identical to those obtained for $\Omega_0/\omega_z=0.4$ in the superfluid regime (blue solid circles). We attribute this result to nearly perfect collisional hydrodynamics in the normal strongly interacting fluid, although a complete many-body microscopic description of this regime does not yet exist.
We see from Fig. \[fig:fig4\] that the moment of inertia is quenched for energies both above and below the superfluid transition. Irrotational hydrodynamics generally requires the quenched moment of inertia to be given by [@StringariIrrot] $$I/I_{rig}=\delta^2\equiv\langle z^2-x^2\rangle^2/\langle
z^2+x^2\rangle^2,
\label{eq:quenching}$$ where $\delta^2$ is computed with respect to the principal axes. Fig. \[fig:fig5\] compares the measured minimum values of $I/I_{rigid}$ with the values of $\delta^2$ obtained from the measured cloud aspect ratios, which directly verifies this prediction.
![Quenching of the moment of inertia versus the measured cloud deformation factor $\delta^2$. Blue solid circles- initial energy before rotation below the superfluid transition energy $E_c
=0.94\,E_F$. Red open circles - initial energy before rotation above the superfluid transition energy.[]{data-label="fig:fig5"}](Fig50514.eps){height="2.5in"}
We have observed quenching of the moment of inertia of a strongly interacting Fermi gas over a wide range of energies, demonstrating nearly perfect irrotational flow and low viscosity hydrodynamics in both the superfluid and normal regimes. Our observations show that these properties, which have been suggested as signatures of superfluidity in the past [@KersonTransport], also appear in the normal strongly interacting fluid. These results have important implications for other strongly interacting systems in nature. It is known that quark-gluon plasmas, as created recently in heavy ion accelerators, exhibit minimum viscosity hydrodynamics and elliptic flow [@Heinz; @Shuryak]. It is therefore possible that irrotational flow will alter the signature of quark-gluon plasmas with finite angular momentum.
This research was supported by the Chemical Sciences, Geosciences and Biosciences Division of the Office of Basic Energy Sciences, Office of Science, U. S. Department of Energy, the Physics Divisions of the Army Research Office and the National Science Foundation.
[30]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , ****, ().
, ** (, ), p. .
, ****, ().
, , , ****, ().
, ** (, , ), vol. , p. .
, , ****, ().
, ****, ().
, , ****, ().
, ****, ().
, , ****, ().
, , ****, ().
, , ****, ().
, , , ****, ().
, , ****, ().
, ****, ().
, , ****, ().
, , , ****, ().
, , ****, ().
, , ****, ().
, ****, ().
, , ****, ().
, , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , ****, ().
, , , ****, ().
, ****, ().
, , ****, ().
, in **, edited by , , (, ), p. .
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'P. M. Saz Parkinson'
title: 'Latest results on gamma-ray pulsars with [*Fermi*]{}'
---
Gamma-ray pulsars pre-[*Fermi*]{}
=================================
Gamma-ray astronomy has a long history, going back to the 1960s. In the 1990s, the Energetic Gamma Ray Experiment Telescope (EGRET), on board the [*Compton Gamma-Ray Observatory*]{} (CGRO[^1], 1991–2000), detected almost 300 gamma-ray sources, a majority of which were unidentified [@3EG]. The seven gamma-ray pulsars detected by CGRO (6 by EGRET, and PSR B1509–58 by COMPTEL), shared many characteristics (e.g. young and highly energetic, mostly double-peaked) but also covered various categories: [*radio-loud*]{}, [*radio-quiet*]{} (Geminga), [*soft*]{} (MeV) , but the somewhat limited statistics (particularly above 5 GeV), made it challenging to discriminate between the leading pulsar emission models [@Thompson04]. For a review of the EGRET era results, immediately preceding the launch of [*Fermi*]{}, see Thompson (2008).
The [*Fermi*]{} era
===================
The [*Fermi*]{} gamma-ray space telescope, launched on 11 June 2008, is a giant leap forward for gamma-ray astronomy. The Large Area Telescope [@Atwood09] (LAT), the main instrument on [*Fermi*]{}, uses silicon strip detectors (far superior to the old gaseous detectors), making it the most sensitive instrument in the $\sim$0.5–300 GeV energy range for the foreseeable future. Indeed, the LAT recently detected its billionth gamma ray ($\sim$1000 times the number of gamma rays detected by EGRET in its lifetime) and is showing no signs of aging. Not surprisingly, [*Fermi*]{} quickly made a big impact in many areas, and pulsars in particular, for example uncovering a large population of [*radio-quiet*]{} gamma-ray pulsars hiding among the [*previously unidentified*]{} EGRET sources [@BSP1; @BSP2]. The most recent catalog released by the LAT Collaboration, the Third LAT source catalog (3FGL), contains over 3,000 sources, of which approximately one third are [*unassociated*]{} [@3FGL]. Uncovering the nature of LAT [*unassociated*]{} sources is (and will remain for many years) a key pursuit for the gamma-ray (and broader) astrophysics community. In this context, a number of statistical methods (e.g. machine learning techniques, neural networks), in combination with multi-wavelength follow-up observations are helping to identify the likely nature of many of these sources [@Chiaro16; @Saz16]. For a detailed review of the “Gamma-ray Pulsar Revolution”, see [@Caraveo14].
Recent pulsar results with Pass 8
---------------------------------
The event selection algorithms developed for the LAT are the result of a long, iterative process, with the various releases known as [ *Passes*]{}. [*Pass 6*]{} data were publicly released after launch but based only on [*pre-launch*]{} information. [*Pass 7*]{} data, released in August 2011, incorporated knowledge gained from the first few years in orbit. The [*Pass 8*]{} release represents a complete redesign of every aspect of the event selection, leading to a significant increase in effective area, an improvement in the point-spread function, and a reduction in background contamination [@Pass8]. Because every [*Pass*]{} results in the entire [*Fermi*]{} data (from the beginning of the mission) being reprocessed, the release of [*Pass 8*]{} produced scientific results immediately after its release, without the need to wait for [*additional*]{} data.
A significant number of known pulsars suddenly showed gamma-ray pulsations with [*Pass 8*]{}, despite being previously undetected [@Laffon15]. The [*Pass 8*]{} data also improved significantly the sensitivity of LAT blind searches for pulsars. The Einstein@Home survey, for example, recently reported 17 new (mostly radio-quiet) gamma-ray pulsars [@Einstein]. The number of gamma-ray pulsars detected by [*Fermi*]{} (now over 200) continues to increase, with the rate of discovery showing no signs of tapering off[^2]. Interestingly, millisecond pulsars (MSPs) represent roughly half the entire gamma-ray pulsar population, with some of them meeting the stringent criteria to be added to the pulsar timing arrays, thus aiding in the search for graviational waves [@Ray12]. One of the most interesting new gamma-ray pulsars detected by the LAT is PSR J0540–6919, in the Large Magellanic Cloud (LMC), located at $\sim$50 kpc, making it the first extra-Galactic gamma-ray pulsar (and hence the most distant) ever detected [@LMCPulsar]. Curiously, PSR J0537–6910, also in the LMC and with very similar characteristics, still shows no gamma-ray pulsations.
Gamma-ray binaries with Fermi
-----------------------------
Another gamma-ray source in the LMC that has recently attracted a great deal of attention was first identified, rather mundanely, as P3 [@LMC]. This source turns out to be a gamma-ray binary with a 10.3 day orbital period, as confirmed also by radio and X-ray observations [@Corbet16]. Coming over four years after the discovery of 1FGL J1018.6–5856 (J1018), the first gamma-ray binary discovered by [*Fermi*]{} [@Corbet11; @J1018], this new gamma-ray binary broke several records (most luminous gamma-ray binary, first extra-Galactic gamma-ray binary), and like J1018, is likely powered by an energetic pulsar [@Corbet16].
While many (if not most) gamma-ray binaries are thought to contain pulsars, in most cases the pulsar has eluded detection (e.g. LSI+61$^\circ$303, LS 5039). In one recent case, however, the pulsar (J2032+4127) was discovered [*first*]{}, while the binary nature of the system was uncovered subsequently. When first discovered in a blind search by [ *Fermi*]{}, PSR J2032+4127 was thought to be an isolated gamma-ray pulsar [@BSP1]. Long-term timing in radio, however, reveals it to be in a binary system with a very long ($\sim$decades) orbital period [@Lyne15]. Recent multi-wavelength monitoring observations report an increase in X-ray emission from the system (by a factor of $\sim$20 since 2010 and a factor of $\sim$70 since 2002) and refined its orbital period to be 45–50 yr, with its time of periastron predicted to be in November 2017 [@Ho17].
The LAT has also been very successful at finding so-called “black widow” or “redback” systems: eclipsing binary millisecond pulsars [*eating*]{} away their low-mass companion star, with their radiation beams. Some of these systems are first identified through their multi-wavelength emission, such as the case of 0FGL J2339.8–0530 [@Romani11]. Radio follow-up searches in this case revealed a pulsar [@Ray14] and gamma-ray pulsations were also detected [^3]. Long term gamma-ray timing of PSR J2339–0533 recently revealed dramatic orbital-period modulations ascribed to a change in the gravitational quadrupole moment [@Pletsch15]. Due to the eclipsing nature of these systems, radio non-detections are frequent, making gamma-ray searches complementary. Indeed, in one case, the pulsar was discovered in gamma rays first [@Pletsch12], with radio pulsations coming later [@Ray13]. A number of redback candidates have been identified (e.g. 3FGL J2039.6–5618 [@Romani15; @Salvetti15], 3FGL J0212.1+5320 [@Li16; @Linares17]) and searches for these pulsars are ongoing.
Variable and transition gamma-ray pulsars
-----------------------------------------
Until recently, gamma-ray pulsars were thought to be [*steady*]{} sources[^4]. The long-term monitoring of large numberes of pulsars over a period of years, however, has started to reveal more complicated behavior in some sources. PSR J2021+4026, a bright, [ *radio-quiet*]{} gamma-ray pulsar discovered by [*Fermi*]{} early on in the mission [@BSP1] experienced an abrupt drop in flux of $\sim$20%, associated with a $\sim$4% increase in spindown rate, also accompanied by changes in the pulse profile, making this the first known variable gamma-ray pulsar [@Allafort13]. The most recent observations appear to show that the flux of J2021+4026 has now gone back to its original values [@Ng16].
An even more dramatic transition was detected in PSR J1023+0038, the so-called “missing link” pulsar known to have previously been in a Low Mass X-ray Binary state, subsequently switching to a rotation-powered state. Recently, this system experienced new state transition, with a five-fold increase in gamma-ray flux accompanying the disappearance of the radio pulsations [@Stappers14].
Another pulsar that has benefitted from the long-term monitoring capabilities of the LAT is PSR J1119–6127 [@Camilo00]. This young, energetic pulsar associated with supernova remnant G292.2–0.5 has an extremely large inferred surface magnentic field ($\sim4\times10^{13}$G), and was detected as a gamma-ray pulsar early on by the LAT [@Parent11]. Recently, the [*Fermi*]{} GBM [@GCN19736] and [*Swift*]{} [@ATel9274] detected a series of strong SGR-like bursts, followed by hard X-ray pulsations [@ATel9282], in conjunction with a large spin-up glitch [@ATel9284]. Radio pulsations disappeared [@ATel9286], reappearing two weeks later [@ATel9366]. Unfortunately, despite a one-week LAT Target of Opportunity (TOO) pointed observation (increasing the exposure by a factor of $\sim$2.4), no significant changes in gamma-ray flux were detected [@ATel9365], and no significant pulsations were detected post-burst [@ATel9378].
Finally, the recent possible detection of pulsed [*soft*]{} gamma-ray emission from PSR J1846–0258 (up to 100 MeV) is of great interest [@ATel9077]. This pulsar shares many similarities with PSR J1119–6127: large magnetic field and past [*magnetar-like*]{} bursts following a large glitch. Thus, it represents another possible “transition” pulsar, making it a worthwhile target to monitor, going forward.
Conclusions
===========
Since its launch, almost nine years ago, [*Fermi*]{} has produced a long list of discoveries in the field of gamma-ray pulsars. More surprisingly, the rate of these discoveries does not appear to be slowing down. [*Fermi*]{} continues to detect new pulsars in every category: young, MSPs, radio-loud, radio-quiet, etc. Finally, the longer data sets and the development of [ *Pass 8*]{} are now enabling [*Fermi*]{} to delve deeper into new parameter space, revealing a range of [*variability*]{} in gamma-ray pulsars that was hitherto unknown.
I thank the organizers of SciNeGHE 2016 for putting together such an interesting and timely conference. In particular, I am very grateful to Max Razzano, for his hospitality during my visit to Pisa. The *Fermi*-LAT Collaboration acknowledges support for LAT development, operation and data analysis from NASA and DOE (United States), CEA/Irfu and IN2P3/CNRS (France), ASI and INFN (Italy), MEXT, KEK, and JAXA (Japan), and the K.A. Wallenberg Foundation, the Swedish Research Council and the National Space Board (Sweden). Science analysis support in the operations phase from INAF (Italy) and CNES (France) is also gratefully acknowledged.
[0]{} Abdo, A. A., Ackermann, M., Ajello, M., et al. 2009, [*Science*]{}, [**325**]{}, 840 Acero, F., Ackermann, M., Ajello, M., et al. 2015, [*ApJS*]{}, [**218**]{}, 23 Ackermann, M., Ajello, M., et al. 2012, [*Science*]{}, [**335**]{}, 189 Ackermann, M., Albert, A., et al. 2015, [*Science*]{}, [**350**]{}, 801 Ackermann, M., Albert, A., et al. 2016, [*A&A*]{}, [**586**]{}, A71 Allafort, A., Baldini, L., Ballet, J., et al. 2013, [*ApJL*]{}, [**777**]{}, L2 , 9282 , 9284 Archibald, R. F., Kaspi, V. M., et al. 2016, [ *ApJL*]{}, [**829**]{}, L21 Atwood, W. B., et al. 2009, [*ApJ*]{}, [**697**]{}, 1071 Atwood, W., Albert, A., Baldini, L., et al. 2013, arXiv:1303.3514 , 9286 , 9366 Camilo, F., Kaspi, V. M., Lyne, A. G., et al. 2000, [*ApJ*]{}, [**541**]{}, 367 Caraveo, P. A. 2014, [*ARA&A*]{}, [**52**]{}, 211 Chiaro, G., Salvetti, D., et al. 2016, [*MNRAS*]{}, [**462**]{}, 3180 Clark, C. J., Wu, J., et al. 2017, [*ApJ*]{}, [**834**]{}, 106 , 3221 Corbet, R. H. D., Chomiuk, L., Coe, M. J., et al. 2016, [*ApJ*]{}, [**829**]{}, 105 Hartman, R. C., Bertsch, D. L., et al. 1999, [*ApJS*]{}, [**123**]{}, 79 Ho, W. C. G., Ng, C.-Y., Lyne, A. G., et al. 2017, [*MNRAS*]{}, [**464**]{}, 1211 , 9274 , 9077 Laffon, H., Smith, D. A., Guillemot, L., et al. , arXiv:1502.03251 Li, K.-L., Kong, A. K. H., Hou, X., et al. 2016, [*ApJ*]{}, [**833**]{}, 143 Linares, M., Miles-P[á]{}ez, et al. 2017, [*MNRAS*]{}, [**465**]{}, 4602 Lyne, A. G., Stappers, B. W., Keith, M. J., et al. 2015, [*MNRAS*]{}, [**451**]{}, 581 Majid, W. A., Pearlman, A. B., Dobreva, T., et al. 2017, [*ApJL*]{}, [**834**]{}, L2 Ng, C. W., Takata, J., & Cheng, K. S. 2016, [*ApJ*]{}, [**825**]{}, 18 Parent, D., Kerr, M., den Hartog, P. R., et al. 2011, [*ApJ*]{}, [**743**]{}, 170 Pletsch, H. J., Guillemot, et al. 2012, [*Science*]{}, [**338**]{}, 1314 Pletsch, H. J., & Clark, C. J. 2015, [*ApJ*]{}, [**807**]{}, 18 Ray, P. S., et al. 2012, arXiv:1205.3089 Ray, P. S., et al. 2013, [ *ApJL*]{}, [**763**]{}, L13 Ray, P., Belfiore, A, Saz Parkinson, P., et al. 2014, AAS \#223, 223, 140.07 Romani, R. W., & Shaw, M. S. 2011, [*ApJL*]{}, [**743**]{}, L26 Romani, R. W. 2012, [*ApJL*]{}, [ **754**]{}, L25 Romani, R. W. 2015, [*ApJL*]{}, [ **812**]{}, L24 Salvetti, D., Mignani, R. P., De Luca, A., et al. 2015, [*ApJ*]{}, [**814**]{}, 88 Saz Parkinson, P. M., Dormody, et al. 2010, [*ApJ*]{}, [**725**]{}, 571 Saz Parkinson, P. M., Xu, H., Yu, P., et al. 2016, [*ApJ*]{}, [**820**]{}, 8 Stappers, B. W., Archibald, A. M., et al. 2014, [*ApJ*]{}, [**790**]{}, 39 , 9365 Thompson, D. J. 2004, Cosmic Gamma-Ray Sources, [****]{}s304, 149 Thompson, D. J. 2008, [ *Reports on Progress in Physics*]{}, [**71**]{}, 116901 , 19736 , 9378
[^1]: The second of NASA’s great observatories.
[^2]:
[^3]: See talk by A. Belfiore at the 2013 Aspen Meeting on Physical Applications of Millisecond Pulsars, [ http://aspen13.phys.wvu.edu/aspen\_talks/Belfiore\_Gamma\_Ray\_Searches.pdf]{}
[^4]: In fact, a key characteristic distinguishing pulsars from AGN is precisely the [*low variability*]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present coefficients for the calculation of the continuous emission spectra of H [i]{}, He [i]{} and He [ii]{} due to electron-ion recombination. Coefficients are given for photon energies from the first ionization threshold for each ion to the $n=20$ threshold of hydrogen ($36.5\mu$), and for temperatures 100 K$~\le~T~\le~10^5$ K. The emission coefficients for He [i]{} are derived from accurate [*ab initio*]{} photoionization data. The coefficients are scaled in such a way that they may be interpolated by a simple scheme with uncertainties less than 1% in the whole temperature and wavelength domain. The data are suitable for incorporation into photoionisation/plasma codes and should aid with the interpretation of spectra from the very cold ionised gas phase inferred to exist in a number of gaseous clouds.'
author:
- |
B. Ercolano$^1$, P. J. Storey$^1$\
$^1$Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, UK\
bibliography:
- 'references.bib'
date: 'Received:'
title: 'Theoretical calculations of the H [i]{}, He [i]{} and He [ii]{} free-bound continuous emission spectra'
---
atomic data
Introduction
============
Analyses of optical recombination line (ORL) spectra of photoionised regions have suggested the existence of cold ionised gas (100 K–2000K), mixed within a warmer component at more typical nebular electron temperatures (8000 K–10000 K). The pockets of cold ionised gas of high metal content are invoked as a possible explanation of the long-standing problem of the discrepancy between elemental abundances derived from ORLs and those derived from collisionally excited lines (CELs). In this scenario, ORLs and CELs are preferentially emitted by the cold and warm phases, respectively (for a recent review see Liu, 2002); this problem is closely linked to the observation that Balmer jump temperatures of H [ii]{} regions and Planetary Nebulae are systematically lower than those derived from the $[$O [iii]{}$]$ nebular to Auroral line ratio (Liu & Danziger 1993).
Understanding the complex spectra arising from such regions relies on the construction of detailed photoionisation models able to account for all gas phases that may be present. Currently, one of the major limitations in such modelling is the lack of an accurate atomic data set extending to such low temperatures. The need of low-temperature effective recombination coefficient for the calculation of recombination lines for metals is apparent. However, in addition to the discrete emission line spectrum, a continuous emission is also produced by the ionised gas, mainly due to free-bound recombination processes of hydrogen and helium ions, free-free transitions in the Coulomb fields of H$^+$, He$^+$ and He$^{2+}$ and two-photon decay of the 2 $^2$S$_{1/2}$ level of H [i]{} and He [ii]{}, and, less importantly of the 2 $^1$S$_{0}$ level of He [i]{}. Accurate continuous emission coefficients are essential for the correct prediction of the Balmer jump by photoionisation codes.
The importance of the continuum processes listed above has long been known and emission coefficients have been tabulated (see e.g. Seaton 1955, 1960; Brown & Mathews 1970), for a range of temperatures and wavelengths mainly aimed at the study of optical data for classical H [ii]{} regions. Ferland (1980) derived H [i]{} and He [ii]{} continuous emission and recombination coefficients for a wider range of temperatures (500 K–2$\cdot$10$^6$ K) and wavelengths to aid the interpretation of ultraviolet (UV) and infrared (IR) as well as optical data from nova ejecta.
In this work we present new calculations of the H [i]{}, He [i]{} and He [ii]{} continuous emission coefficients due to free-bound recombination, over temperatures ranging from 100 K to 10$^{5}$ K, for the physical conditions thought likely to occur in chemically inhomogeneous regions, which may include pockets of cold ionised material intermixed within typical nebular gas.
Calculations Method
===================
{width="8.5cm"}
{width="8.5cm"}
Using the Saha-Boltzmann equation and the Milne relation, we may express the continuous emission coefficient $\gamma(\nu)$ corresponding to the recombination process $$X^+ + {\rm e}^-(\epsilon) \rightarrow X^* + h\nu$$ in terms of the photoionization cross-section $\sigma_{\nu}(X^*)$ $$\gamma(\nu) = \frac{4\pi h}{c^2} (\frac{h^2}{2\pi m k T})^{3/2} {\rm e}^{-\epsilon/kT} \frac{\omega^*}{\omega^+} \nu^3 \sigma_{\nu}(X^*)$$ where $\omega^+$ and $\omega^*$ are the statistical weights of the recombining ion initial state and final state respectively, and $\epsilon$ is the free electron energy. In terms of $\gamma(\nu)$, the energy emitted per unit volume per unit time in frequency interval $\nu$ to $\nu + {\rm d\nu}$ is $N_e N(X^+) \gamma(\nu) {\rm d}\nu$.
For the contribution $\gamma_n(\nu)$ at frequency $\nu$ from recombinations to states of principal quantum number $n$ of H$^0$ and He$^+$ we compute the necessary energy-dependent photoionization cross-sections using the hydrogenic codes described by Storey & Hummer (1991). The total emission coefficient $\gamma(\nu)$ is then $$\gamma(\nu) = \sum_{n_0}^{\infty} \gamma_n(\nu)$$ where in practice we truncate the sum at $n=200$ for H [i]{} and $n=350$ for He [ii]{}, which is sufficient to ensure convergence to all figures given in the Tables.
For recombination to atomic helium we use the [*ab initio*]{} calculated photoionization cross-sections described by Hummer & Storey (1998). Cross-section data are available for $n^1$S, $n^3$S, $n^1$P$^{\rm o}$, $n^3$P$^{\rm o}$, $n^1$D, $n^3$D, $n^1$F$^{\rm o}$ and $n^3$F$^{\rm o}$states with $l+1\leq n \leq 20$. For higher values of $n$ and for $l>3$ we use hydrogenic data. Thus for $n\geq 5$ the emission coefficient has nine distinct thresholds for each $n$ corresponding to the eight separate terms plus a threshold at the hydrogenic energy. Hummer & Storey (1998) showed that the results of their [*ab initio*]{} calculation of the photoionization cross-sections are in better agreement at threshold with the highly accurate bound-bound calculations of Drake (1996) than any of the other methods used to compute helium recombination processes.
Contributions due to bremsstrahlung emission of a Maxwellian distribution of electrons in the Coulomb fields of hydrogen and helium ions are not included in our results, but can be obtained readily using (e.g.) equation 2 of Brown & Mathews (1970) or the free-free computer code published by Storey & Hummer (1991).
Contributions to the continuum emission from two-photon emission are also not included but can be computed using the formulae of Nussbaumer & Schmutz (1984).
[ccccccccc]{} $\nu~[{\rm Hz}]$ & E $[{\rm Ryd}]$ & $\lambda_{\rm vac}[$Å$]$ &\
& & & H[i]{}(exact) & H[i]{}(interp) & He[i]{}(exact) & He[i]{}(interp) & He[ii]{}(exact) & He[ii]{}(interp)\
1.000($+$14) & 0.030397 & 29979.246 & 2.415($+$01) & 2.418($+$01) & 2.408($+$01) & 2.408($+$01) & 7.393($+$01) & 7.412($+$01)\
2.000($+$14) & 0.060793 & 14989.623 & 2.852($+$00) & 2.858($+$00) & 3.026($+$00) & 3.028($+$00) & 3.222($+$01) & 3.233($+$01)\
3.000($+$14) & 0.091190 & 9993.082 & 1.653($+$00) & 1.655($+$00) & 2.067($+$00) & 2.068($+$00) & 7.034($+$01) & 7.047($+$01)\
4.000($+$14) & 0.121586 & 7494.811 & 4.442($+$01) & 4.442($+$01) & 5.948($+$01) & 5.947($+$01) & 9.356($+$01) & 9.366($+$01)\
5.000($+$14) & 0.151983 & 5995.849 & 6.996($-$01) & 6.988($-$01) & 2.848($+$00) & 2.856($+$00) & 1.457($+$00) & 1.458($+$00)\
6.000($+$14) & 0.182380 & 4996.541 & 1.093($-$02) & 1.092($-$02) & 5.478($-$02) & 5.495($-$02) & 3.098($+$01) & 3.098($+$01)\
7.000($+$14) & 0.212776 & 4282.749 & 1.698($-$04) & 1.694($-$04) & 1.015($-$03) & 1.017($-$03) & 4.807($-$01) & 4.805($-$01)\
8.000($+$14) & 0.243173 & 3747.406 & 2.629($-$06) & 2.628($-$06) & 1.828($-$05) & 1.831($-$05) & 7.437($-$03) & 7.438($-$03)\
9.000($+$14) & 0.273569 & 3331.027 & 2.376($+$01) & 2.376($+$00) & 1.795($+$02) & 1.795($+$02) & 5.054($+$01) & 5.054($+$01)\
1.000($+$15) & 0.303966 & 2997.925 & 3.704($-$01) & 3.702($-$01) & 1.097($+$01) & 1.100($+$01 & 7.823($-$01) & 7.821($-$01)\
{width="20cm"} \[t\]
Results
=======
The full tables of continuous emission coefficients, Tables 3, 4 and 5 for H [i]{}, He [ii]{} and He [i]{} respectively are available in electronic form only. In Table 1 we show an extract from the table for H [i]{}. Coefficients are tabulated for log $T$\[K\] = 2.0(0.1)5.0 and for photon energies (in Rydbergs) from just less than the ground state threshold energy to just above the energy of the threshold at $n=20$. Values are tabulated on either side of each threshold and at some additional nodal points inserted to make interpolation sufficiently accurate. At low temperatures the coefficients fall rapidly and exponentially at energies above each threshold making interpolation difficult so rather than $\gamma(\nu)$ we tabulate $\gamma^{\dagger}(\nu)$ defined by
$$\begin{aligned}
\gamma^{\dagger}(\nu)& = & \gamma(\nu)\ 10^{34}\ T^{3/2} {\rm e}^{{\Delta E}/kT} \\
& = & \gamma(\nu)\ 10^{40}\ t^{3/2} {\rm e}^{{15.7887\Delta E_R}/t} \end{aligned}$$
where $t~=~T[K]/10^{4}$, $\Delta E$ is the difference between the photon energy, h$\nu$, and the energy of the nearest threshold of lower energy and $\Delta E_R$ is the same energy in Rydberg units. Thresholds are indicated by index 1 and additional nodal points by index 0.
The recommended procedure for deriving the emission coefficient at a given temperature and photon energy is to interpolate linearly in the appropriate table in the variables log$~T$ and photon energy to obtain the scaled coefficient $\gamma^{\dagger}(\nu)$. Equation 1 can then be applied to obtain $\gamma(\nu)$. Table 2 gives exact values of $\gamma(\nu)$ and values derived from the recommended interpolation scheme for each ion at a range of photon energies.
Comparison of the tabulated values with calculations performed on a finer frequency grid and temperature grid show that linear interpolation in log $T$ and photon energy yields accurate results with maximum deviations of 1% and average deviations much smaller than this.
Figure 1 shows the non-scaled continuum emission coefficients, $\gamma(\nu)$ for H [i]{}, He [i]{} and He [ii]{} in the optical wavelength range for a temperature of 10000 K (left panel) and 100 K (right panel). The left panel of figure 1 is directly comparable to figure 1 of Brown & Mathews (1970). We can only compare the magnitude of the discontinuity at each threshold directly with the results of Brown & Matthews (1970), since their results incorporate all continuum processes while ours only deal with free-bound processes. Comparing results for H [i]{} and He [ii]{} we find a maximum difference of 1% for H [i]{} at the Balmer threshold and at the lowest temperature tabulated by them of 4000K. The difference is attributable to Brown & Matthews using an approximate expression for the hydrogenic threshold photoionization cross-sections while we use the exact expressions incorporated in the codes of Storey & Hummer (1991). A similar comparison for He [i]{} shows that for all thresholds except that corresponding to the 3 $^3$P$^{\rm o}$ state at an air wavelength of 7849Å, our results differ by no more than 2.1% from those of Brown & Matthews (1970). The differences that do exist are due to the approximate method used by Brown & Matthews to calculate the helium photoionization cross-sections. In the case of the 3 $^3$P$^{\rm o}$ threshold, we find much larger differences reaching 21% at 4000K. This is almost certainly due to a numerical error in the work of Brown & Matthews, since the magnitude of the discontinuity at the 3 $^3$P$^{\rm o}$ threshold does not obey the correct scaling with temperature in their work.
We also compared our values of $\gamma(\nu)$ for H [i]{} and He [ii]{} with those published by Ferland (1980) and found, in general, good agreement, with typical deviations of the order of 2-5% in the overlapping temperature range.
The electronic tables are structured as follows: node/threshold (0/1) indices are given in column 1, the photon energies in Ryd are given in column 2, the scaled free-bound emission coefficients for temperatures in the 100K-100000K range are given in the remaining columns. For He [i]{} and He [ii]{} coefficients are tabulated for log $T$\[K\] = 2.0(0.1)5.0 while for He [i]{}, log $T$\[K\] = 2.0(0.04)5.0. The maximum photon energy is slightly less than the ground state ionization energy for each ion and the minimum photon energy corresponds to the $n=20$ threshold in H [i]{}.
Conclusions
===========
We have presented new calculations of the free-bound continuous emission coefficients for the hydrogen and helium ions. The results for He [i]{} are derived from accurate [*ab initio*]{} photoionization cross-section data. The coefficients are given for a wide range of temperatures and frequencies, extending to the previously unexplored very low-temperatures regime. This is needed for the interpretation of spectra from very cold ionised gas, which has been inferred to exist, possibly in the form of density/chemical inhomogeneities, from observations of ORLs in H [ii]{} regions and planetary nebulae. The data are presented under an interpolation scheme that allows estimates to be obtained over the entire temperature and energy range presented with less than 1% uncertainty.
[8.]{}
Brown, R. L. & Mathews, W. G., 1970, ApJ, 160, 939. Drake, G.W.F., 1996, in Drake G.W., ed., Atomic, Molecular and Optical Physics Handbook, American Institute of Physics, Woodbury, p154. Ferland, G. J., 1980, PASP, 92, 596. Hummer, D. G., Storey, P. J., 1998, MNRAS, 297, 1073. Liu, X.-W., 2002, Rev. Mex. Astron. Astrof. Ser. Conf.. 12. 70 Liu, X.-W., Danziger, J., 1993, MNRAS, 263, 256 Nussbaumer, H., Schmutz, W., 1984, A&A, 138, 495. Seaton, M. J., 1955, MNRAS, 115, 279. Seaton, M. J., 1960, Rep. Prog. Phys, 23, 313. Storey, P. J., Hummer, D. G., 1991, CoPhC, 66, 129.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the highly magnified arc SGAS J122651.3+215220 caused by a star-forming galaxy source at $z_s=2.93$ crossing the lensing caustic cast by the galaxy cluster SDSS J1226+2152 ($z_l=0.43$), using archival Hubble Space Telescope observations. We report in the arc a number of asymmetric surface brightness features consistently seen across four filters, lying at a fraction of an arcsecond from the lensing critical curve and likely to be highly but unequally magnified image pairs of underlying compact sources. One explanation of unequal magnification is strongly coupled microlensing by intracluster stars, which induces independent flux variations in the images of individual or groups of source stars in the lensed galaxy. For a second possibility, intracluster dark matter subhalos invisible to the telescope can effectively perturb lensing magnifications near the critical curve and give rise to persistently unequal image pairs. Our modeling suggests that the microlensing hypothesis is in tension with the absence of notable temporal variation in the asymmetries over a six-year baseline, while subhalos of $\sim 10^6$–$10^8\,M_\odot$ anticipated from structure formation with Cold Dark Matter typically produce stationary and sizable asymmetries. We judge that observations at additional times and more precise lens models are necessary to stringently constrain temporal variability and robustly distinguish between the two possible explanations. The arc under our study is a scheduled target of a Director’s Discretionary Early Release Science program of the James Webb Space Telescope, which will provide deep images as well as a high-resolution view with integral field spectroscopy.'
author:
- |
Liang Dai,$^{1}$[^1] Alexander A. Kaurov,$^{1}$ Keren Sharon,$^{2}$ Michael K. Florian,$^{3}$\
$^{1}$Institute for Advanced Study, 1 Einstein Drive, Princeton, NJ 08540, USA\
$^{2}$Department of Astronomy, University of Michigan, 1085 S. University Ave, Ann Arbor, MI 48109, USA\
$^{3}$Observational Cosmology Lab, NASA Goddard Space Flight Center, 8800 Greenbelt Rd., Greenbelt, MD 20771, USA\
$^{4}$Institució Catalana de Recerca i Estudis Avançats, Barcelona, Catalonia\
$^{5}$Institut de Ciències del Cosmos, Universitat de Barcelona (IEEC-UB), Barcelona, Catalonia\
$^{6}$Department of Astronomy/Steward Observatory, University of Arizona, 933 N. Cherry Ave, Tucson, AZ 85721, USA\
$^{7}$Department of Physics, University of Cincinnati, Cincinnati, OH 45221, USA\
$^{8}$MIT Kavli Institute for Astrophysics and Space Research, 77 Massachusetts Ave., Cambridge, MA 02139, USA
bibliography:
- 'refs.bib'
- 'refs2.bib'
- 'refs3.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: 'Asymmetric Surface Brightness Structure of Lensed Arc in SDSS J1226+2152: A Case for Dark Matter Substructure'
---
\[firstpage\]
Introduction
============
Massive clusters of galaxies are the most powerful cosmic gravitational lenses that often make outstanding images of magnified background galaxies, usually referred to as “arcs”. The most dramatic phenomenon takes place when a background galaxy happens to straddle the caustic cast by the foreground gravitational lens, offering an opportunity to closely study an elongated image of the galaxy with a zoomed-in view near the caustic. Smoothly distributed mass in the foreground lens universally produces symmetric pairs of merging images in arcs crossing the lensing critical curve. However, intracluster microlensing of individual superluminous stars in the source galaxy [@1991ApJ...379...94M; @2017ApJ...850...49V; @2018ApJ...857...25D; @Oguri:2017ock] or sub-galactic dark matter (DM) substructure predicted to exist in cluster halos [@press1974formation; @white1978core; @1985ApJ...292..371D] can cause temporary or persistent departure from the symmetric appearance of the arc.
By analysing this image asymmetry one can study distant extragalactic objects that are otherwise prohibitively difficult to probe. Firstly, microlensing allows one to scrutinize individual bright stars at $z \gtrsim 1$ [@2018NatAs...2..334K; @Chen:2019ncy; @2019ApJ...880...58K], or reveal stellar transients from the distant Universe [@2018NatAs...2..324R; @Diego:2018fzr; @windhorst2018observability]. The upcoming 192-orbit HST program (GO-15936, PI: Kelly) will potentially dramatically increase the number of detections of extremely bright extragalactic stars [@kelly2019flashlights]. On the other hand, one is able to probe the intracluster population of stellar microlenses inside the cluster halo, or even a multitude of (sub-)planetary mass self-gravitating DM minihalos if they form in large abundance during the early universe [@Dai:2019lud; @Arvanitaki:2019rax; @Blinov:2019jqc]. Second, perturbations in the shape of the critical curve on scales $\sim 0.1\arcsec$ will inform us about the abundance of $10^6-10^8\,M_\odot$ DM subhalos within the cluster-sized $10^{13}-10^{15}\,M_\odot$ lens halo [@2018ApJ...867...24D], thereby probing a dynamic range too large to be easily accessible with other methods.
Previously, multi-epoch Hubble Space Telescope (HST) observations in the field of the galaxy cluster MACS J0416.1-2403 allowed [@Chen:2019ncy] and [@2019ApJ...880...58K] to carry out independent studies of image asymmetries in a $z=0.9397$ arc across the cluster critical curve and their temporal variability. Both groups suggested that the most significant asymmetric feature is at least partially caused by microlensing induced random brightening of macro images of an underlying supergiant star. Moreover, [@2019ApJ...880...58K] mentioned that the arc may have several other less significant asymmetric features.
In this paper, we examine the spectacular lensed arc SGAS J122651.3+215220 (hereafter S1226) with a spectroscopic redshift $z_s=2.93$ in the field of the $z = 0.43$ galaxy cluster SDSS J1226+2152. This arc, resulting from two images of a background galaxy merging along a critical line, was first discovered from the SDSS Giant Arcs Survey (SGAS) as a Lyman Break Galaxy during intense star formation [@koester2010two]. Concerning the study of image asymmetries, S1226 is particularly advantageous because it exhibits multiple bright compact star-forming clumps, whose image pairs across the critical curve have small angular separations $\lesssim 0.3\arcsec$ and hence are hugely magnified. We examine one HST epoch of exposure in 2011 in two optical wide filters and one epoch of exposure in 2017 in two infrared (IR) wide filters. We perform an image asymmetry analysis similar to the one presented in [@2019ApJ...880...58K] for the arc in the field of MACS J0416.1-2403.
We report significant surface brightness asymmetries in the arc S1226, which is unexpected for a smooth mass distribution in the lens. Located within a fraction of an arcsecond from the critical curve, these asymmetric features must be highly magnified images of underlying bright sources in the star-forming galaxy smaller than a few parsecs, which can be either individual bright stars or clusters of many stars. Appearing at several places across the arc, these flux asymmetries are consistency in all four HST filters, with tentative evidence for persistency over a six-year observing baseline. We find that the absence of significant variability in asymmetric features implies tension with stochastic intracluster microlensing as the explanation for the asymmetry. On the other hand, we suggest that lensing by sub-galactic DM subhalos of masses $\sim 10^6$–$10^8\,M_\odot$ as predicted by the standard theory of hierarchical structure assembly with Cold Dark Matter (CDM) [@moore1999dark; @2008MNRAS.391.1685S; @2016ApJ...824..144F] can explain persistent asymmetries at the observed level. This derives from the idea of using image flux ratios as diagnostic of invisible substructure [@1998MNRAS.295..587M; @2001ApJ...563....9M], but extends from galactic host halos [@2019MNRAS.487.5721G] to more massive ones $M_{\rm host} \gtrsim 10^{14}\,M_\odot$ and into the regime of huge magnification factors $\gtrsim 100$, by examining just one caustic straddling system. Lack of existing multi-epoch imaging in any single HST filter, however, prevents us from decisively ruling out microlensing variability in favor of substructure lensing. Promisingly, S1226 has been selected as a target of the approved Director’s Discretionary Early Release Science (DD-ERS) program with the James Webb Space Telescope (JWST) (\# 1335, PI: Rigby) [@rigby2017templates]. Forthcoming data will make this highly magnified arc a particularly interesting system. This paper is organized as follows: in [Sec. \[sec:data\]]{} we briefly summarize existing HST and the upcoming JWST data. Then in [Sec. \[sec:asym\]]{}, we show that the arc hosts underresolved highly magnified sources which exhibit significant flux asymmetry between their image pairs across the cluster critical curve. In [Sec. \[sec:macrolens\]]{}, we use the macro lens model describing the galaxy cluster SDSS J1226+2152 to infer local properties at the location of the magnified arc. In [Sec. \[sec:icstars\]]{}, we estimate the abundance of foreground intracluster stars in the arc projected vicinity, which is useful information for quantitatively understanding microlensing. We then consider two possible physical explanations for the observed image asymmetry: time-varying intracluster microlensing in [Sec. \[sec:microlensing\]]{}, and essentially persistent DM substructure lensing in [Sec. \[sec:subhalolensing\]]{}. We conclude that intracluster microlensing is not conclusively ruled out but is unfavorable in light of the roughly consistent asymmetries across a six-year observational baseline, while DM subhalo lensing is theoretically expected and naturally explains non-variability. Finally, we give concluding remarks in [Sec. \[sec:disc\]]{}. Furthermore, we include [App. \[app:nonrig\]]{} presenting non-rigid registration as an alternative method to corroborate the existence of flux asymmetries, and [App. \[app:magasym\]]{} discussing magnification asymmetry induced by perturber lenses in the proximity of the critical curve.
 
Data {#sec:data}
====
The caustic-straddling lensed arc SGAS 122651.3+215220 was discovered in December 2007 at the 2.5m Nordic Optical Telescope [@2010ApJ...723L..73K] and is in the catalog of SDSS Giant Arcs Survey (SGAS; [@2008AJ....135..664H]). The arc redshift is determined to be $z_s=2.9260\pm0.0002$ from nebular lines [@rigby2018magellan]. The spectroscopic redshift of the galaxy cluster lens SDSS J1226+2152 $z_l=0.43$ was reported by [@2014ApJ...783...41B]. This data was used to perform lens modeling with the [`L`ENSTOOL]{} software package [@2007NJPh....9..447J]. It approximates the cluster mass distribution as a linear superposition of mass halos and then uses Markov Chain Monte Carlo sampling to determine the best-fit set of parameters for the lens model. The fitting procedure for analogous clusters is described in [@2019arXiv190405940S].
The arc S1226 was imaged in multiple HST bands. We use for our analysis the aligned and drizzled images taken with the Advanced Camera for Surveys (ACS) in two optical wide filters, F606W and F814W, and with the Wide Field Camera 3 (WFC3) in two infrared (IR) wide filters, F110W and F160W. The images in the optical filters were obtained 6 years prior to the images obtained in the IR filters. [Tab. \[tab:obs\]]{} summarizes the observations. In [Fig. \[fig:image\]]{}, we show a composite false-color image of the cluster and a zoomed-in view of the arc.
The arc will be observed with the Near Infrared Spectrograph (NIRSpec) and the Mid-Infrared Instrument (MIRI) integral field unit (IFU) and imaged with the Near Infrared Camera (NIRCam) in multiple filters aboard JWST, as part of the DD-ERS program (see [Tab. \[tab:obsj\]]{} for details).
Instrument Filter Exposure \[ks\] Date Program \# PI
------------ -------- ----------------- ------------ ------------ ---------
ACS F606W 2.0 2011-04-14 12368 Morris
ACS F814W 2.0 2011-04-14 12368 Morris
WCF3 F110W 1.2 2017-12-18 15378 Bayliss
WCF3 F160W 1.3 2017-12-18 15378 Bayliss
Instrument Filters/methods Exposure \[ks\]
------------ --------------------- -----------------
NIRSPEC F170LP, G235H (IFU) 8.3
MIRI F560W, MRS (IFU) 3.6
NIRCAM F115W, F277W 0.3
NIRCAM F150W, F356W 0.3
NIRCAM F200W, F444W 0.3
Asymmetries in the arc {#sec:asym}
======================
According to our macro lens model, the portion of the critical line that passes through the arc is nearly a straight line and is approximately perpendicular to the direction of arc elongation (shown in [Fig. \[fig:image\]]{}), except that a small cluster member galaxy to the northwest mildly disturbs the surface brightness profile of the arc. We focus on the brightest part of the arc on the south which is not significantly influenced by the galaxy perturber (also see discussion in [Sec. \[sec:galperturber\]]{}). If the lens surface mass distribution is locally smooth, symmetric appearance across the critical curve is expected. We look for departure from symmetric appearance using two methods which were previously applied to a similar caustic-straddling arc in MACS J0416.1-2403 [@2019ApJ...880...58K]. In [Sec. \[sec:slits\]]{}, we present the results from the first method which is based on aligning multiple slits with the direction of arc elongation, and identifying bright features on both sides of the critical curve that are likely to be image pairs of compact sources. Then, in [Sec. \[sec:photometry\]]{}, we perform photometry to estimate the flux differences between the two images of each image pair. For a comparison, in [App. \[app:nonrig\]]{} we apply a non-physical technique to indicate asymmetric features, which is flexible to account for small deviation from perfect fold symmetry.
Slits {#sec:slits}
-----
We define five $1\arcsec \times 0.17\arcsec$ slits along the critical curve, each of which is positioned to align with the direction of arc elongation and to include a pair of major surface brightness feature. For every HST filter, we sum the flux along the perpendicular direction, and plot the summed flux as a function of the position along the slit direction in [Fig. \[fig:profiles\]]{}. In Slits B, C, D and E, we clearly identify a pair of bright features, whose flux asymmetries are significant and are quite consistent in all 4 filters. Slit A shows a complex situation of many brightness peaks, which prevents us from unambiguously defining an image pair. Thus, surface brightness peaks near the critical curve appear inconsistent with symmetry anticipated from a smooth lens.
The diffuse intracluster light (ICL) [@zwicky1951coma; @lin2004k; @zibetti2005intergalactic] at the location of the arc is moderate as can be seen in [Fig. \[fig:image\]]{}. Since the observed asymmetric image pairs have small separations, ICL offsets the flux zero points but cannot induce significant flux asymmetries, especially in short-wavelength filters. The same should be true for the diffuse light extending from the small galaxy perturber on the northwest, which is nearer to the arc but is fainter. Moreover, asymmetries in Slits B, D and E are opposite to that in Slit C, inconsistent with a large-scale ICL gradient. When reporting flux asymmetries in [Fig. \[fig:flux\_diff\]]{}, we model and subtract an ICL ellipsoid from around the Brightest Cluster Galaxy (BCG).
![Asymmetric surface brightness features in the arc. Top four panels show cutouts in four HST filters. The bottom panel shows a false-color image, where we define multiple $1\arcsec \times 0.17\arcsec$ slits along the (roughly) expected position of the cluster critical curve. The brightness profiles along these slits are shown in [Fig. \[fig:profiles\]]{}. The images are oriented such that the cluster critical curve is nearly vertical and the slits are aligned with the direction of arc elongation.[]{data-label="fig:slits"}](hor.pdf){width="\linewidth"}
![Brightness profiles along several $1\arcsec \times 0.17\arcsec$ slits as defined in [Fig. \[fig:slits\]]{}. Slits B, C and D show clear asymmetries in all four HST filters. []{data-label="fig:profiles"}](slits.pdf)
Photometry {#sec:photometry}
----------
Using two different methods, we estimate the flux differences between prominent image pairs in Slits B, C and D, provided that the underlying sources appear unresolved. For the first method, which we refer to as PSF photometry, we locate the central pixels of maximal flux of each tentative image pair and estimate the flux difference based on these single pixels, accounting for their flux fraction given the expected point-spread function. Results obtained with this method are presented in [Fig. \[fig:flux\_diff\]]{}, with uncertainties indicated as the shaded bands. For the second method, we perform aperture photometry. We choose appropriate top-hat apertures that best approximate the expected PSF for various filters. The results, shown in [Fig. \[fig:flux\_diff\]]{} as dashed lines, are broadly consistent with those derived from PSF photometry. From both the slit analysis in [Sec. \[sec:slits\]]{} and the two photometric analyses in this Section, we substantiate the flux asymmetries of image pairs in the arc. We note from [Fig. \[fig:slits\]]{} that several clear image pairs outside our defined slits also show asymmetric fluxes. Having larger separations from the cluster critical curve $\gtrsim 1\arcsec$, those are however not expected to be highly symmetric even in the case of a locally smooth lens, and are more prone to differential ICL contamination.
![Absolute value of flux difference between the pairs of images in four HST filters as measured from PSF photometry (shaded bands correspond to $1\sigma$ error) and aperture photometry (crosses). We show results for Slits B, C and D, whose peak pixels of maximal flux are relatively easy to determine. Colored regions at the bottom are filled under the throughput curves of the various filters.[]{data-label="fig:flux_diff"}](flux_diff.pdf){width="\linewidth"}
Macro lens model near critical curve {#sec:macrolens}
====================================
At the location of arc, we infer from the cluster lens model a ratio $\kappa_0 = \Sigma_0/\Sigma_{\rm crit} = 0.8$ between the coarse-grained total surface mass density $\Sigma_0$ of the lens at the line of sight and the critical surface mass density [@blandford1986fermat] $\Sigma_{\rm crit}=(c^2/4\pi\,G)\,(D_S/D_L\,D_{LS})\approx 2\times 10^9\,M_\odot/{\rm kpc}^2$, where $D_L$, $D_S$ and $D_{LS}$ are the angular diameter distances to the cluster lens at $z_l=0.43$, to the source galaxy at $z_s=2.93$, and from the lens to the source. This is the local value for the lensing convergence of the macro lens. Near the critical curve, a compact source appears as a pair of macro images, each of which has an unsigned magnification factor $$\begin{aligned}
\label{eq:barmu}
\bar{\mu} = \frac{1}{2\,|\Delta\theta|\,|1-\kappa_0|\,d\,|\sin\alpha|}\end{aligned}$$ where $\Delta\theta$ (signed) equals to half the angular separation between the image pair, $d$ is the gradient of inverse magnification in the critical curve vicinity, and $\alpha$ is the angle between the critical curve and the direction of arc elongation. Numerically, we measure from our cluster lens model $d \approx 7.5\,{\rm arcmin}^{-1}$ and $\alpha \approx 90^\circ$, which gives $\bar{\mu} \approx 140\,\left(|\Delta \theta|/0.14\,\arcsec\right)^{-1}$. While the lensed arc considered here at $z_s=2.92$ is substantially more distant than the caustic straddling lensed galaxies studied in MACS J1149.5+2223 [@2018NatAs...2..334K] and MACS J0416.1-2403, the magnification factor as a function of image separation across the critical curve is similar in order of magnitude to the values found in those systems.
Intracluster stars {#sec:icstars}
==================
The line of sight to the arc is at a projected distance of $B \sim 50\,{\rm kpc}$ to the BCG. Intervening intracluster stars can originate from several sources: (i) diffuse light extending from the BCG; (ii) diffuse light extending from a major cluster member galaxy $\sim 4 \arcsec$ south of the arc; (iii) diffuse light extending from a minor cluster member galaxy just $\sim 2\arcsec$ away from the arc to the northwest. The minor member galaxy acts as a substructure lens, causing a compact and bright star-forming source to appear in multiple images, as can be seen in the right panel of [Fig. \[fig:image\]]{}. The colors measured from the 4 HST filters are consistent between components (i) and (ii), while it is difficult to reliably determine the colors of component (iii) due to arc contamination. Inferred from the image taken in the reddest filter F160W, component (iii) can be safely neglected, and component (ii) is subdominant compared to the diffuse light halo of the BCG, if not entirely negligible.
The major axis of the BCG diffuse light nearly intersects the lensed arc. We fit the isophotes in each HST filter to ellipses centered at the BCG and adopt the same ellipticity for all four filters. We find consistent colors wherever contamination from other sources are negligible. The surface brightness at the location of the arc calculated from fitting isophote ellipses is within a factor of two when compared to a few nearby places that are not on top of the arc.
We use the population synthesis code Flexible Stellar Population Synthesis ([`F`SPS]{}) to model intracluster stars [@conroy2009propagation; @conroy2010propagation] assuming the initial mass function of [@kroupa2001variation]. We find that the colors are consistent with an old simple stellar population, with a single age and metallicity. The age ranges from $t_{\rm ssp} = 7\,$Gyr for low metallicity $\log(Z/Z_\odot) = -1.0$ to $ t_{\rm ssp} = 1.5\,$Gyr for high metallicity $\log(Z/Z_\odot) = 0.3$. Varying $\log(Z/Z_\odot)$ from $-1.0$ to $0.3$ correspond to a conservative range $0.002 < \kappa_\star < 0.009$. While breaking the age-metallicity degeneracy necessitates spectroscopic analysis, we think the most likely age is $t_{\rm ssp} \approx 3\,$Gyr with metallicity $\log(Z/Z_\odot) \simeq -0.3$, a value found typical for ICL at similar cluster-centric radii in many cluster lenses of similar redshifts [@Montes:2017yct]. This translates into a local surface density of intracluster stars $\Sigma_{\rm ICL} \approx 10^7\,M_\odot/{\rm kpc}^2$ and hence a convergence $\kappa_\star = \Sigma_{\rm ICL}/\Sigma_{\rm crit} \approx 0.005$, which we use hereafter as our fiducial value. This value is comparable to previous estimates carried out for caustic straddling lensed galaxies in MACS J1149.5+2223 [@PhysRevD.97.023518] and in MACS J0416.1-2403 [@2019ApJ...880...58K].
In calculating $\kappa_\star$, we have included into the synthesized stellar population white dwarfs in addition to main sequences and evolved stars, but have neglected neutron stars and black holes. The latter group are in any case remnants of massive stars $M \gtrsim 8\,M_\odot$, which account for only a sub-dominant fraction $\sim 14\%$ of the initial star-forming mass. In our model, intracluster stars have masses in the range $0.08$–$1.4\,M_\odot$. Insensitive to modeling uncertainty, the most abundant microlenses are main sequence dwarfs whose masses are around $\sim 0.2\,M_\odot$.
Intracluster Microlensing {#sec:microlensing}
=========================
The observed flux of a compact source is affected by stellar microlenses near the line of sight, which render the two macro images unequally bright at a given epoch. One possibility is that each observed asymmetric pair of images reflects the microlensing phenomenon acting on either a [*single*]{} super-luminous source star or a compact [*group*]{} of bright source stars in the lensed galaxy. Under this hypothesis, no substructure in the lens mass distribution other than that in individual stellar microlenses needs to be invoked to explain the asymmetries.
Intracluster stars should break the smooth macro critical curve into an interconnected network of micro critical curves, whose full width is $2\,\kappa_\star/d = 0.03$–$0.14\arcsec$, with our best estimate being $0.08\arcsec$. This network band has one edge at the expected location of the macro critical curve, and the other edge lies on its interior [@2017ApJ...850...49V]. Under the assumption that the macro critical curve roughly bisects the asymmetric pairs of surface brightness features in [Fig. \[fig:nonrig\]]{}, none of the features lie within this band. This implies that microlensing variability is infrequent and occurs over a long timescale for any single lensed star.
Microlensing of a single source star
------------------------------------
To corroborate this, in [Fig. \[fig:mumaps\]]{} we simulate random realizations of the total microlensing magnification factor for [*one*]{} of the macro images across a small region on the source plane, assuming a macro lens model appropriate for S1226, and stellar microlens masses randomly drawn from a Simple Stellar Population model with age $t_{\rm ssp}=3\,$Gyr and $\log(Z/Z_\odot)=-0.3$. The resolution of these magnification maps is $\sim 1\,{\rm AU}$. For $\kappa_\star=0.002$–$0.009$ and $\Delta\theta=\pm 0.14\arcsec$, appropriate e.g. for the prominent asymmetries in Slit C and D, the range of source-plane length scales indicates that the magnification may vary by order unity on a timescale of several years for a typical effective source-lens relative velocity $v_t$ (see Eq.(12) of [@2017ApJ...850...49V]) of $\sim 100\, {\rm AU/yr}$.
If the asymmetric fluxes observed in several slits (e.g. Slits B, C and D) are due to microlensing of a [*single*]{} source star, the pairs of macro images must have large differences in their magnification factor. At $z_s = 2.93$, only the most luminous source stars in the rest-frame UV under unusually high magnification can explain the observed flux asymmetries reported in [Fig. \[fig:flux\_diff\]]{}. The best constraints are obtained from the F814W filter because the asymmetric features are brighter than in F606W, and sharper than in F110W and F160W. In this filter, the image pair flux asymmetries correspond to AB magnitudes of $m \approx 28.5$ in slits B and C, and $m \approx 27.5$ in slit D. We list in [Tab. \[tab:srcstarmag\]]{} the required minimum magnification differences, $|\Delta\mu_{\rm min}|$, needed to explain the measured flux asymmetries for several candidate types of stars, for slits B and C. For slit D the required magnification is even higher, but we conservatively use the values in slits B and C.
The best candidate source stars are low-metallicity blue supergiants (BSG), but even these require a magnification asymmetry $|\Delta\mu| \gtrsim 600$. The most luminous main-sequence (MS) stars require much higher magnifications of $|\Delta\mu| \gtrsim 3000$, which are much less likely. This is under the optimistic assumption of zero dust reddening, while [@2019ApJ...882..182C] report E(B-V)$=0.13\,$mag measured for S1226. The magnification difference needs to be a factor of two larger to explain the asymmetry in Slit D.
There is no evidence for rare stars (e.g. zero metallicity, extremely massive stars) that might be much brighter than those listed in [Tab. \[tab:srcstarmag\]]{}. The rest-frame UV spectrum of the arc analyzed by [@2019ApJ...882..182C] suggests that S1226 has a moderately sub-solar metallicity $-0.9 <\log(Z/Z_\odot)<-0.4$ and a stellar age $\sim 20$–$26\,$Myr, and is dominated by B-type giants with no trace of O stars detected. The best stellar candidates for the image pairs that can produce asymmetries through microlensing are therefore BSG as listed in [Tab. \[tab:srcstarmag\]]{}.
The required values of $|\Delta\mu|$ in [Tab. \[tab:srcstarmag\]]{} clearly exceed the expected macro magnification $\bar{\mu} \approx 140$ by a large factor, and are rarely reached as seen in [Fig. \[fig:mumaps\]]{}. This large magnification difference is only possible when the source star undergoes a micro caustic transit. [Fig. \[fig:mu\_CDF\_N1\]]{} indicates that $|\Delta\mu| \gtrsim 600$ randomly occurs for less than a few percent of the time. Microlensing of a single source star is therefore unlikely to be the cause of the flux asymmetries seen in the arc. The main difference between the case of S1226 and the caustic straddling galaxies lensed by MACS J1149.5+2223 and MACS J0416.1-2403, which have confirmed microlensed individual stars, is the lower redshift of the latter sources, $z=1.49$ and $0.94$, making it much more likely for a microlensed luminous star to reach the faintest observable fluxes.
stellar type $t_{\rm age}\,[{\rm Myr}]$ $\log\left(Z/Z_\odot\right)$ $M_\star\,[M_\odot]$ $T_{\rm eff}\,[{\rm K}]$ $R_\star\,[R_\odot]$ $L_{\rm bol}\,[10^6\,L_\odot]$ $|\Delta\mu|_{\rm min}$
----------------- ---------------------------- ------------------------------ ---------------------- -------------------------- ---------------------- -------------------------------- -------------------------
blue supergiant 3 $0.0$ $42$ $22000$ $64$ $0.88$ $1869$
blue supergiant 3 $-0.3$ $47$ $13000$ $274$ $1.9$ $1064$
blue supergiant 3 $-1.0$ $66$ $16000$ $205$ $2.4$ $655$
blue supergiant 3 $-2.0$ $66$ $16000$ $200$ $2.4$ $645$
main sequence 0.3 $0.0$ $115$ $50000$ $17.5$ $1.7$ $3317$
main sequence 0.3 $-0.3$ $117$ $52000$ $16$ $1.7$ $3565$
main sequence 1 $-1.0$ $107$ $53000$ $15$ $1.6$ $4127$
main sequence 0.3 $-2.0$ $118$ $60000$ $12$ $1.7$ $4085$
![Numerically derived cumulative distribution for the magnification difference $\Delta\mu$ between the two macro images of a single source star. The assumption of a point source is valid for either MS or giant star across the range of $|\Delta\mu|$ shown here. Curves are plotted for $\kappa_\star=0.002$, $0.005$, and $0.009$. The grey vertical line marks the required minimum magnification asymmetry $|\Delta\mu| \simeq 600$ for the example BSGs considered in [Tab. \[tab:srcstarmag\]]{}. The tail of the distribution at $|\Delta\mu| \gtrsim {\rm few} \times 10^3$ artificially steepens due to finite pixels in numerical ray shooting. The choice $\Delta\theta=\pm 0.14\arcsec$ is appropriate for Slits C and D. The grey vertical line indicates the required magnification difference for the most luminous blue supergiants available in our model stellar population, and suggests that a single source star is unlikely to account for the observed flux asymmetry.[]{data-label="fig:mu_CDF_N1"}](mu_CDF_N1_0D14.pdf)
![Cumulative distribution for the magnification asymmetry $\Delta\mu_g$ for $\kappa_\star=0.005$. This is similar to [Fig. \[fig:mu\_CDF\_N1\]]{} but for a group of $N_\star$ identical source stars subject to uncorrelated microlensing. The grey shaded band corresponds to $10$%–$50$% magnification asymmetry estimated for the asymmetric features shown in [Fig. \[fig:profiles\]]{}. For each value of $N_\star$ (color coded), a vertical line is drawn to indicate $|\Delta\mu_g|=600/N_\star$. The location of the grey shaded band in this plot favors $N_\star\sim 10$–$30$ brightest stars.[]{data-label="fig:mu_CDF_multi_stars"}](mu_CDF_multi_stars_0D14.pdf)
Microlensing of a cluster of source stars
-----------------------------------------
The large asymmetric fluxes are more naturally explained by a group of luminous stars (possibly co-existing with many more fainter stars). Indeed, hot luminous stars commonly cluster in compact star-forming regions or open clusters. The age of the arc stellar population, estimated to be $\sim 20$–$26\,$Myr from rest-frame UV spectroscopy [@2019ApJ...882..182C], is longer than that of the BSGs in [Tab. \[tab:srcstarmag\]]{}, which implies that the stellar light is dominated by giant B stars that are individually less massive and luminous than the BSGs in [Tab. \[tab:srcstarmag\]]{}.
The asymmetric image pairs, e.g. in Slits C and D, appear barely resolved in the F814W image. Given the angular distance from the cluster critical curve, our macro lens model constrains their physical sizes to be $\lesssim 5\,$pc. Neglecting closely bound multiple stars for the time being, member stars within the group should have mutual separations on the order of parsecs, which is much greater than the source-plane correlation lengths $\sim 10^2$–$10^3\,{\rm AU}$ of microlensing magnifications as shown in [Fig. \[fig:mumaps\]]{}.
This implies that at any given time individual luminous stars have statistically independent microlensing magnifications. Under this assumption, we show in [Fig. \[fig:mu\_CDF\_multi\_stars\]]{} the probability distribution of $\Delta\mu_g$ for a star group, where $\mu_g$ is the (microlensing affected) magnification averaged over $N_\star$ stars for one (unresolved) [*macro*]{} image of the star group on one side of the macro critical curve. As $N_\star$ increases, the distribution of $\Delta\mu_g$ approaches a Gaussian centered at zero with decreasing width. For simplicity, we assume that all $N_\star$ stars are identical. A small number $N_\star$ is disfavored for the same reason as in the previous subsection (typical values of $|\Delta\mu_g|$ are insufficient to explain the flux asymmetries). The minimum $|\Delta\mu_g|$ to account for the flux asymmetries is reduced as $\propto 1/N_\star$. We find that a value $N_\star \sim 10$–$30$ is most consistent with the level of macro image asymmetry seen in [Fig. \[fig:profiles\]]{}. This number yields a typical magnification difference $|\Delta\mu_g|$ induced by uncorrelated microlensing of $\sim 10\%$–$50\%$ of $\bar{\mu}\approx 140$, shown as the shaded vertical band in [Fig. \[fig:mu\_CDF\_multi\_stars\]]{}. A number $N_\star \sim 100$–$500$ of stars of comparable brightness might still produce a ratio $|\Delta\mu_g|/\bar{\mu}$ compatible with [Fig. \[fig:mu\_CDF\_multi\_stars\]]{}, but each star should then be fainter in order not to overproduce the observed flux of either macro image. For $N_\star \gtrsim 500$, $\Delta\mu_g$ would be highly diluted and a flux asymmetry as large as $\sim 10\%$–$50\%$ would be implausible. For a more accurate analysis, one could replace the simplifying assumption of identical source stars with a model of a continuous luminosity function. We leave such analysis for future investigation.
If $N_\star$ ultra-luminous stars form a system more compact than $\sim 10^2\,$AU, their microlensing magnifications will be highly correlated. They may simultaneously have a magnification asymmetry $|\Delta\mu| \simeq 600/N_\star$. For $N_\star \sim 3$–$6$, the value for $|\Delta\mu_g| \approx |\Delta\mu|$ would be typical of the distribution shown in [Fig. \[fig:mu\_CDF\_N1\]]{}. The lower required magnification may then be more likely to occur, but the tight star clusters required may not be abundant enough to explain the multiple observed asymmetric image pairs.
To summarize, by measuring unequal image pairs at just one random epoch per HST filter, we conclude that while the hypothesis of microlensing of a single ultra-luminous source star is hardly viable, the asymmetries might be explained by a group of stars, under one of the following three situations: uniform microlensing of $N_\star \sim 3$–$6$ extremely luminous BSGs tightly bound within $\lesssim 10^2\,$AU, or uncorrelated microlensing of a group of $N_\star \sim 10$–$30$ such BSGs within a region of size $\lesssim \mathcal{O}({\rm pc})$, or uncorrelated microlensing of a cluster of $N_\star \sim 100$–$500$ less luminous stars (most likely B-type giants) within a region of size $\lesssim \mathcal{O}({\rm pc})$. As we will see next, frequent temporal variability of the flux asymmetry is generally expected from these microlensing scenarios.
Temporal variability of flux asymmetry
--------------------------------------
The flux asymmetry between the two macro images induced by microlensing is expected to vary with time as each source star slowly traverses the microlensing magnification pattern on the source plane (as in [Fig. \[fig:mumaps\]]{}). As a result, the asymmetry can frequently change its sign.
As the star number $N_\star$ of a group increases, the overall flux asymmetry should vary more rapidly because the flux asymmetry from each member star varies independently and there are more frequent micro caustic crossings, but the fractional size of the asymmetry is more diluted. This is clearly reflected in [Fig. \[fig:lc\_examples\]]{}, where we show numerical examples of flux asymmetry variation from randomly drawn microlens realizations, up to the uncertainty in the timescale that scales with the unknown effective transverse velocity parameter $v_t$.
Since the arc has not been imaged at more than one epoch in any single filter, we cannot rule out temporal variation in the flux asymmetries with complete certainty. Still, the asymmetries seen in F606W and F814W are similar to those seen in F110W and F160W, if not identical, even though the two optical filters and the two IR filters were used at two different epochs separated by more than 6 years. This seems to hint that the observed flux asymmetries are persistent, or at least the variable component is relatively small.
Assuming a fiducial velocity $v_t \sim 400\,{\rm km/s}$, up to a factor of few uncertainty unless fine-tuned, the apparent non-variability over $\sim 6\,$yr can be compatible with $N_\star\sim 1$–$3$, but we have seen that such few number of dominant bright stars are probably incompatible with the large size of the asymmetries in absolute flux units. On the other hand, [Fig. \[fig:lc\_examples\]]{} suggests that $N_\star \gtrsim 10$ would lead to changes in the sign and magnitude of $\Delta\mu_g$ that occur too frequently over the $\sim 6\,$yr timescale, unless the relative transverse velocity is surprisingly low, $v_t \lesssim 100\,{\rm km/s}$. Extremely tight systems of $N_\star \simeq 3$–$6$ supergiant stars with nearly uniform microlensing might still explain some of the asymmetries, but taking into account all previous considerations, we believe the observed persistent flux asymmetries are not in good agreement with microlensing models of either single or under-resolved groups of source stars. Ultimately, imaging in the same filters at additional epochs is necessary to robustly establish or falsify microlensing variability over a timescale of years.
Substructure lensing {#sec:subhalolensing}
====================
In [Sec. \[sec:microlensing\]]{}, we have studied flux asymmetries from intracluster microlensing operating on minuscule angular scales far beyond telescope resolution. In this Section, we consider substructure lensing from either small galaxies or star-free DM subhalos inside the cluster halo, which also cause flux asymmetries as we show below. The characteristic lensing angular scales of $\sim 0.01$ to $0.1\,{\rm arcsec}$ are marginally resolved in diffraction limited exposures. Unlike microlensing, asymmetric patterns resultant from substructure lensing do not exhibit noticeable variability over timescales of decades to centuries.
Galaxy perturbers {#sec:galperturber}
-----------------
The smooth lens model inevitably deviates from the idealized fold model [@1992grle.book.....S] away from the macro critical curve. Large perturber lenses far away in projection can induce curvature in the macro critical curve on arc second scales. Indeed, the right panel of [Fig. \[fig:image\]]{} shows that our smooth lens model predicts such an example of curvature due to a minor foreground galaxy $\sim 2\arcsec$ to the northwest. However, these large and distant perturbers are usually unable to generate magnification asymmetries larger than $\sim 10\%$ between image pairs with small separation $\Delta\theta$.
Consider an image of a point source on one side of the critical curve, with magnification $\mu$, and its counter image with magnification $\mu'$. Define the signed fractional magnification asymmetry $$\begin{aligned}
\label{eq:amu}
a_\mu \equiv 2\,\left( \left|\mu \right| - \left|\mu'\right| \right) /\left( \left|\mu \right| + \left|\mu' \right| \right).\end{aligned}$$ Let $M_p$ be the mass of a faraway perturber, $\theta_p$ its Einstein angular scale, and $b$ the angular impact parameter from the image pair. We estimate that $|a_\mu|$ is of the order (derived in [App. \[app:magasym\]]{}) $$\begin{aligned}
\label{eq:amuest}
\left|a_\mu\right| & \sim \left( \theta^2_p\,\Delta\theta \right)/\left( d\,b^4 \right) \nonumber\\
& \sim 0.04\,\left( \frac{M_p}{10^{11}\,M_\odot} \right)\,\left( \frac{7.5\,{\rm arcmin}^{-1}}{d} \right)\,\left( \frac{2\,\arcsec}{b} \right)^{4}\,\left( \frac{\Delta\theta}{0.14\,\arcsec}\right),\end{aligned}$$ where we assume $b \gg \Delta\theta$ and that the perturber is at the cluster redshift $z_l=0.43$. The perturber galaxy would therefore have to enclose a mass $\sim 10^{12}\,M_\odot$ within $\lesssim 10\,{\rm kpc}$ to produce tens of percent asymmetry. In fact, a blue compact source, probably a young star cluster associated with the arc at $z_s=2.93$, is lensed by this foreground galaxy into multiple images separated by a critical curve loop of size $\sim 1\,\arcsec$. Taking into account that a macro magnification $\sim 30$ should be acting to enhance the effect of the perturber, we find the small size of the critical curve loop inconsistent with the central part of the foreground galaxy (where star light is detected) enclosing $\gtrsim 10^{11}\,M_\odot$.
In [Fig. \[fig:mag\_asym\]]{}, we further confirm that the galaxy perturber alone could not have caused the magnification asymmetry. For every image point, we locate the counter image point and calculate $a_\mu$ as defined in [Eq. (\[eq:amu\])]{}. In our smooth lens model, $a_\mu$ only reaches a few percent within $\sim 0.2\,\arcsec$ of the critical curve, which is compatible with our order of magnitude estimate in [Eq. (\[eq:amuest\])]{}.
![Fractional magnification asymmetry $a_\mu$ ([Eq. (\[eq:amu\])]{}) between any image point and its counter image point, according to our smooth lens model and within a $2\arcsec\times 2\arcsec$ FoV centered at the caustic straddling arc. Black and magenta contours, respectively, correspond to $a_\mu=\pm 0.05$ and $\pm 0.1$, with solid ones for $a_\mu>0$ and dashed ones for $a_\mu<0$. Brown (solid and dashed) curves are contours of constant signed macro magnification factor. The region having $|a_\mu|<10^{-2}$ in the proximity of the critical curve is artificially set to have $a_\mu=\pm 10^{-2}$. For visual guidance, shaded pixels indicate bright features in the F814W filter. We have artificially applied a small, uniform shift of the F814W image relative to the lens model such that the smooth critical curve crudely bisects pairs of surface brightness features. Large magnification asymmetries $|a_\mu|\gtrsim 10\%$ are not expected from a smooth macro lens within $\lesssim 0.2\arcsec$ of the critical curve.[]{data-label="fig:mag_asym"}](mag_asym_smooth_model.pdf)
Dark matter subhalos
--------------------
While no evidence supports that any minor foreground galaxy significantly breaks the symmetry near the cluster critical curve, we hypothesize that a population of abundant and non-luminous DM subhalos may be the reason. Unlike the larger but rarer perturbers, sub-galactic subhalos are predicted to be numerous enough to be frequently found close to the critical curves, where their perturbing effects are greatly enhanced [@2017ApJ...845..118M; @2018ApJ...867...24D].
To numerically assess the effect of subhalos expected from the standard Cold Dark Matter (CDM) theory, we populate the cluster DM halo with randomly generated subhalos, which are superimposed on top of an ideal fold model with parameters appropriate for the arc S1226. We follow the method outlined in [@2018ApJ...867...24D] using an extrapolation of the semi-analytic model of [@han2016unified] down to subhalo-host mass ratios as small as $m_{\rm sh}/M_{\rm host} \sim 10^{-8}$. For simplicity, we model the host DM halo of SDSS J1226+2152 using a spherical Navarro-Frenk-White (NFW) profile [@1996ApJ...462..563N; @navarro1997universal], with a characteristic mass $M_{\rm 200}=1.5\times 10^{14}\,M_\odot$ and a concentration parameter $C_{\rm 200} \simeq 13$. These broadly reproduce a measured Einstein radius $\theta_E \simeq 10\arcsec$ [@2012MNRAS.420.3213O], a local convergence $\kappa_0=0.8$ at the line of sight as inferred from our macro lens model, as well as an empirical mass-concentration relation found for a catalog of strong and weak lensing SGAS clusters [@bayliss2011gemini; @2012MNRAS.420.3213O]. The host DM halo has an NFW scale radius $R_s = 75\,{\rm kpc}$. Subhalos are assumed to follow a power-law mass function $\rmd n(m_{\rm sh})/\rmd \log m_{\rm sh} \propto m_{\rm sh}^{-0.9}$ [@mo2010galaxy], and are approximated as spherical but tidally truncated NFW profiles [@baltz2009analytic; @cyr2016dark]. We refer interested readers to Appendix B of [@2018ApJ...867...24D] for technical details.
Similarly to [Fig. \[fig:mag\_asym\]]{}, we calculate the magnification asymmetry $a_\mu$ between a pair of images by ray shooting: for every chosen image position, we first find the corresponding source position, and then locate the counter images. When more than one counter image is present, we pick the brightest counter image and evaluate $a_\mu$ according to [Eq. (\[eq:amu\])]{}.
In [Fig. \[fig:amusubhalo\]]{}, we show the pattern of $a_\mu$ influenced by one random realization of subhalos for a compact source, within a $1\arcsec \times 1\arcsec$ vicinity of the cluster critical curve. The pattern of signed magnification is also shown. The influence of a subhalo much more massive than $\sim 10^{10}\,M_\odot$ lying on top of the lensed arc should be large enough to be noticeable. But the expected number density of such massive subhalos is low enough that having any one of them within $\sim 1\arcsec$ of the cluster critical curve is improbable. We generate smaller subhalos with $10^6\,M_\odot < m_{\rm sh} < 10^{10}\,M_\odot$. The model of [@han2016unified] we adopt here predicts about one subhalo with $m_{\rm sh} \gtrsim 10^8\,M_\odot$ within a disk of $1\arcsec$ radius centered on the cluster critical curve. On the other hand, typically many subhalos with $m_{\rm sh} \sim 10^6\,M_\odot$ are present just a fraction of an arcsecond away from the cluster critical curve. Unable to host star formation, these halos cannot be observed except for gravitational probes sensitive to small scale structures, like the lensing perturbations causing asymmetries in highly magnified image pairs.
As shown in [Fig. \[fig:amusubhalo\]]{}, even one subhalo can induce substantial magnification asymmetry if it is sufficiently close to one of the image pair. A subhalo of mass $m_{\rm sh} \simeq 10^8\,M_\odot$ located $\sim 0.4\arcsec$ from the cluster critical curve can cause $a_\mu \gtrsim 0.1$ within $\sim 0.1\arcsec$–$0.3\arcsec$ from its center. At $\lesssim 0.1\arcsec$ from the cluster critical curve, small subhalos with $m_{\rm sh} \sim 10^6\,M_\odot$ appear common enough to produce $a_\mu \gtrsim 0.1$ across a significant portion of the image plane. Even smaller subhalos should be more numerous and may contribute to sizable $a_\mu$ very close to the critical curve. Our simulation of tiny subhalos $m_{\rm sh} < 10^6\,M_\odot$ is limited by computational cost. Their effects may be important for extremely compact sources such as individual bright stars, but are likely to dilute away for sources of size larger than $\sim 1\,$pc.
The precise distribution of magnification asymmetry naturally fluctuates among realizations of subhalos, and mildly depends on host halo structural parameters $M_{\rm 200}$ and $C_{\rm 200}$. The typical magnification asymmetry scales with the normalization of the subhalo mass function. Despite these uncertainties, it seems feasible from our simulations that a population of DM subhalos in the mass range $\sim 10^6$–$10^8\,M_\odot$ induces ubiquitous asymmetry at the level $a_\mu \gtrsim 10\%$ for pairs of highly magnified images within $\sim 0.2\arcsec$ from the cluster critical curve. Subhalos in this mass range are also capable of imprinting astrometric effects detectable with future imaging efforts [@2018ApJ...867...24D]. It is therefore theoretically possible that the flux asymmetries in S1226, such as those in Slits C, D and E, are entirely or partially caused by subhalo lensing. Regarding the ambiguous identification of any image pair in Slit A, for example, we surmise that it could be due to a subhalo creating additional lensed images of the same source.
Satellite galaxies and globular clusters {#sec:}
----------------------------------------
If sub-galactic perturber lenses are in fact responsible for the observed image pair asymmetries, could DM subhalos be confused by dwarf galaxies in the cluster or along the line of sight, globular clusters or other known stellar systems that are too faint to be detected by HST? Such star clusters or dwarf galaxies would have to be much smaller than, e.g., the minor perturber galaxy to the north of S1226 so as not to induce obvious deformation of the arc, and they would have to be numerous enough to be found near the critical curve with reasonable likelihood. Our ICL estimate at the location of the arc $\kappa_\star \simeq 0.005$ implies that the $1\arcsec\times 1 \arcsec$ FoV in [Fig. \[fig:amusubhalo\]]{} on average only encloses $\sim 10^6\,M_\odot$ of stars, leaving little mass budget to have many stellar mass clumps comparable to the DM subhalos we consider. It is also well known that the general dwarf galaxy luminosity function is much less steep at the faint end than the predicted CDM subhalo mass function at the low-mass end, so dwarf galaxy abundances should be much smaller than required for producing the perturbations illustrated in [Fig. \[fig:amusubhalo\]]{}.
Globular clusters (GCs) populate the intracluster medium due to tidal stripping from their host galaxies [@white1987globular; @west1995intracluster]. In recent years, intracluster GCs have been studied in several galaxy clusters up to $z \simeq 0.3$ [@2010Sci...328..334L; @peng2011hst; @west2011globular; @Alamo-Martinez:2013aaa; @d2016extended; @lee2016globular]. In the core of rich clusters, the GC surface number density is found to be around ${\rm few} \times 10^{-2}\,{\rm kpc}^{-2}$ [@2010Sci...328..334L; @Alamo-Martinez:2013aaa; @d2016extended; @lee2016globular], which would correspond to no more than a few GCs within the same FoV in [Fig. \[fig:amusubhalo\]]{}. Even in the extreme case of the core of the Coma Cluster, for which a GC surface density of $\sim 1\,{\rm kpc}^{-2}$ was reported by [@peng2011hst], this number would be $\sim 30$, an order of magnitude smaller than DM subhalos with $m_{\rm sh} \gtrsim 10^6\,M_\odot$. Furthermore, these GCs have masses $\sim 10^5$–$10^6\,M_\odot$ (probably without any significant DM), smaller than those of the DM subhalos we have considered. Hence, interloper GCs are also not abundant enough to be the major cause of the observed image asymmetries.
Satellite galaxies are thought to reside in DM halos. Therefore, satellites should be no more numerous than the DM subhalos that match their host DM halo size, which are already included in the DM subhalo population model we adopt. DM subhalos small enough to be common within the FoV of [Fig. \[fig:amusubhalo\]]{} probably lack any significant stellar component. In the cores of nearby rich galaxy clusters, an extrapolation of the observed satellite luminosity function to the luminosity of GCs [@1995AJ....110.1507B; @mobasher2003photometric; @2016ApJ...824...10F] still falls short of the DM subhalo number density seen in [Fig. \[fig:amusubhalo\]]{} by more than two orders of magnitude. Averaging over the entire cluster, we suggest that intracluster satellite galaxies and GCs are less abundant than $m_{\rm sh} \gtrsim 10^6$ DM subhalos to ubiquitously perturb the proximity of the smooth critical curve and induce magnification asymmetries. In the case of S1226, this abundance may be enhanced by satellite galaxies and GCs bound to nearby cluster member galaxies in projection. Deep imaging of the nearby ICL with JWST will set stringent limits on the abundance of faint satellites. Spectroscopic analysis will help rule out foreground stellar systems superimposed on top of the arc.
Conclusions {#sec:disc}
===========
We have investigated archival HST data showing asymmetric surface brightness features within $\lesssim 0.3\arcsec$ from the lensing critical curve in a $z_s=2.93$ magnified arc behind the galaxy cluster SDSS J1226+2152. We have identified highly magnified image pairs that must correspond to several sources more compact than just a few parsecs, all of which have flux differences of more than $\sim 10\%$ which are confirmed in multiple HST filters.
Intracluster microlensing inevitably introduces uncorrelated magnification fluctuations in different images of the same source, causing unequal fluxes at any given epoch. Due to the high source redshift $z_s=2.93$, we have found that the sources of the asymmetric image pairs are too luminous to be just one or a few supergiants even if each star is as bright as $L_{\rm bol}\simeq 10^6\,L_\odot$, but are more likely to be clusters of more than $\sim 10$ bright stars. Our modeling suggests that having $N_\star \sim 10$–$30$ comparably luminous stars shortens the timescale of asymmetry variability to just $\sim 1$ or $2$ yrs. In this case, the asymmetry should even flip its sign several times over six years. This appears at odds with the largely consistent sign and degree of asymmetries between optical and IR exposures taken six years apart. For a cluster of $N_\star \gtrsim 100$ comparable stars, microlensing induced asymmetry should be severely diluted, while variability should occur even more rapidly. Since HST images in 2011 and in 2017 were not taken in the same filters, we have not been able to derive a tight constraint on variability, and in fact, variability at some level, however small, is unavoidable. An additional HST epoch in both UV and IR filters would tighten the constraint on any asymmetry component of microlensing origin, giving helpful guidance to upcoming JWST observations.
A population of $\sim 10^6$–$10^8\,M_\odot$ intervening DM subhalos perturbing the magnification symmetry in the proximity of the critical curve is the other viable hypothesis we have considered, for which persistent asymmetries are natural outcomes. Using a reasonable model for subhalo abundance and mass function derived for the CDM paradigm, we have found it plausible that image pairs having more than a $\sim 10\%$ fractional difference in magnifications are common within $\lesssim 0.2\arcsec$ of the cluster critical curve in the case of S1226.
Precise characterization of the cluster halo of SDSS J1226+2152 and improved substructure modeling will enable us to draw a more robust conclusion on whether asymmetric image pairs neighboring the critical curve in S1226 are caused by substructure lensing. Deeper images taken at many epochs with HST and JWST will help reduce the uncertainty in the flux asymmetry measurement, separate any time-varying component from the persistent component, and rule out possible interloper satellite galaxies or GCs. Nebular lines from HII regions surrounding hot stars, if spatially compact enough, might provide additional evidence for asymmetries insensitive to microlensing thanks to large source sizes. More generally, we expect an increased number of caustic straddling arcs with clumpy star-forming features in the proximity of lensing critical curves to be discovered in the future. Imaging and spectroscopy follow-ups of the source-lens systems will allow us to examine many such magnified image pairs and establish whether asymmetric magnifications are ubiquitous.
Detecting the population of small-scale subhalos through lensing near caustics will be unprecedented and will complement other efforts to uncover the invisible DM structure on sub-galactic scales, including flux ratio measurements of multiply imaged quasars [@nierenberg2014detection; @nierenberg2017probing; @2019MNRAS.487.5721G] and dynamic modeling of Milky Way stellar streams [@johnston2002lumpy; @ibata2002substructure; @carlberg2009star; @2019arXiv191102663B; @2019ApJ...880...38B]. Thereby, improved constraints will be derived for alternative DM models such as Warm Dark Matter [@kusenko2009sterile; @shoemaker2009gravitino; @abazajian2017sterile] or “fuzzy” DM [@press1990single; @PhysRevD.50.3650; @goodman2000repulsive; @Peebles:2000yy; @hu2000fuzzy; @2006PhLB..642..192A; @2014NatPh..10..496S; @hui2017ultralight].
Acknowledgements {#acknowledgements .unnumbered}
================
This work is based on observations made with the NASA/ESA Hubble Space Telescope, obtained from the data archive at the Space Telescope Science Institute. STScI is operated by the Association of Universities for Research in Astronomy, Inc. under NASA contract NAS 5-26555. LD and TV acknowledge the support of John Bahcall Fellowships at the Institute for Advanced Study. AK is supported by the IBM Einstein fellowship. JM has been supported by Spanish Fellowship PRX18/00444, and by the Corning Glass Works Foundation Fellowship Fund. MB acknowledges support by NASA through grant number HST-GO-15378.006-A from the Space Telescope Science Institute, which is operated by AURA, Inc., under NASA contract NAS 5-26555. We acknowledge the use of the publicly available Python package `Colossus` to carry out cosmological calculations [@diemer2018colossus].
Non-rigid registration {#app:nonrig}
======================
The method of non-rigid registration allows to smoothly deform one surface brightness pattern to match another. In [@2019ApJ...880...58K], this method was applied to the arc in MACS J0416.1-2403. In that case, asymmetries were marginally observed in several slits using deeper co-added HST exposures. Thus, in that case non-rigid registration relied on matching arc surface brightness features seen on both sides of the critical curve to account for imperfections of the ideal fold lens model. The main finding from this method in [@2019ApJ...880...58K] was that asymmetries appear more prominent as one examines the portion of the image plane closer to the critical curve, a trend broadly expected from both intracluster microlensing and DM substructure lensing. For S1226, asymmetries are significant in many of the slits we define in [Sec. \[sec:slits\]]{}. Thus, the method of non-rigid registration is naively less reliable. However, we apply it here to see what we can recover.
The resulting non-rigid transformation is presented in [Fig. \[fig:nonrig\]]{}, where we plot the speculated morphology of the (perhaps perturbed) critical curve as the white dashed curve. The figure highlights many inconsistencies along the critical line, which is in line with what we have found in [Sec. \[sec:slits\]]{}. In [Fig. \[fig:nonrig\_prof\]]{}, we show the average amplitude of match residuals as a function of distance to the critical curve, similar to what was done in [@2019ApJ...880...58K]. While some surface brightness mismatches are mitigated or eliminated by the non-rigid transformation, others cannot be perfectly cured, especially at smaller distances to the critical curve.
![The left column presents the same HST images as in [Fig. \[fig:slits\]]{} for F606W, F814W, F110W and F160W filters. The middle column shows the left part of the arc flipped under non-rigid transformation. The right column shows the surface brightness difference between the left and the middle columns. The white dashed curve indicates the estimated position of the critical curve. []{data-label="fig:nonrig"}](nonrig.pdf)
![Standard deviation of residuals in the images of the arc as a function of the angular distance to the critical curve. The standard deviations are calculated in strips parallel to the critical curve using residuals presented in the rightmost column of [Fig. \[fig:nonrig\]]{}. The rise in the standard deviation within $\sim 0.2\arcsec$ may be a sign of intracluster microlensing of individual stars or compact star forming complexes, or substructure lensing from a population of DM subhalos.[]{data-label="fig:nonrig_prof"}](nonrig_prof.pdf)
Magnification asymmetry across critical curve {#app:magasym}
=============================================
In this Appendix, we estimate the degree of uneven fluxes between a close pair of highly magnified images on both sides of the critical curve. We aim to derive the parametric dependence on the perturber parameters.
Consider a pair of images at image plane positions $x_{1,i}$ and $x_{2, i}$. Demanding that they map to the same source point, we obtain from the lens equation $$\begin{aligned}
x_{2, i} - x_{1, i} = \alpha_i(x_2) - \alpha_i(x_1),\end{aligned}$$ where $\alpha_i(x)$ is the deflection field as a function of the two-dimensional image plane vector $x$. Let us write $x_{1,i} = x_{0, i} - \Delta x_i$ and $x_{2,i} = x_{0, i} + \Delta x_i$, where $x_{0, i}$ is the midpoint between the image pair.
Assuming that $\alpha_i(x_j)$ is smooth, we Taylor expand in powers of $\Delta x_i$: $$\begin{aligned}
\Delta x_i & = \left.\left[ \nabla_j\,\alpha_i \right]\right|_{x_0}\,\Delta x_j + \frac{1}{6}\,\left.\left[ \nabla_l\,\nabla_k\,\nabla_j\,\alpha_i \right]\right|_{x_0}\,\Delta x_l\,\Delta x_k\,\Delta x_j \nonumber\\
& + \mathcal{O}\left[\Delta x^{5}\right],\end{aligned}$$ where $\nabla_i$ denotes image-plane derivative with respect to $x_i$, and $\left.[\cdots]\right|_{x_0}$ stands for evaluation at $x_0$. Ignoring terms at cubic and higher orders, the Jacobian matrix $J_{ij}(x)=\delta_{ij} - \nabla_i\,\alpha_j(x)$ at $x_0$ satisfies $\left.[J_{ij}]\right|_{x_0}\,\Delta x_j \approx 0$. This means that $\Delta x_i$ is an eigenvector with a vanishing eigenvalue.
The inverse of the signed magnification is $$\begin{aligned}
1/\mu(x) = {\rm det}\left[ \delta_{ij} - \nabla_i\,\alpha_j(x)\right],\end{aligned}$$ where we again Taylor expand around $x=x_0$: $$\begin{aligned}
\label{eq:JijTaylorExpand}
\delta_{ij} - \nabla_i\,\alpha_j(x_{1,2}) & = \left.[J_{ij}]\right|_{x_0} \pm \frac12\,\left.\left[ \nabla_k\,\nabla_j\,\alpha_i \right]\right|_{x_0}\,\Delta x_k \\
& - \frac16\,\left.\left[ \nabla_l\,\nabla_k\,\nabla_j\,\alpha_i \right]\right|_{x_0}\,\Delta x_l\,\Delta x_k + \mathcal{O}\left[\Delta x^{3}\right]. \nonumber\end{aligned}$$ In a suitably oriented coordinate system, $\Delta x_i$ aligns with the first axis, and the only nonzero element of $\left.[J_{ij}]\right|_{x_0}$ is the 22 element being order unity.
If we neglect terms of $\mathcal{O}\left[\Delta x^2\right]$ in [Eq. (\[eq:JijTaylorExpand\])]{}, the determinant is at the leading order proportional to the $\mathcal{O}\left[\Delta x\right]$ term in the 11 element of the full Jacobian matrix. Since this term has equal magnitudes but opposite signs at both image positions $x_1$ and $x_2$, we recover the familiar result that a pair of images across an ideal fold are equally magnified.
Departure from perfect magnification symmetry is induced by two types of correction in [Eq. (\[eq:JijTaylorExpand\])]{}: (i) the $\mathcal{O}\left[\Delta x^2\right]$ term in the 11 element; (ii) the product of $\mathcal{O}\left[\Delta x\right]$ terms from two matrix elements. Having the same signs at $x_1$ and at $x_2$, these correct the leading magnification factor in opposite directions at the two images and hence generate asymmetry.
The lowest order derivative of the deflection field is usually determined by the large-scale lens, while the higher order derivatives can be dominated by small perturber lenses. By definition, $\Delta x$ is on the order of the image separation $\Delta\theta$. Neglecting numerical prefactors, we assume $\left.\left[ \nabla_k\,\nabla_j\,\alpha_i \right]\right|_{x_0}$ is of order $d$ from the large-scale mass distribution. We assume that a perturber lens dominates the derivative at the next order, $$\begin{aligned}
\left.\left[ \nabla_l\,\nabla_k\,\nabla_j\,\alpha_i \right]\right|_{x_0} \sim \theta^2_p/b^4,\end{aligned}$$ as long as the angular impact parameter $b$ is larger than the characteristic Einstein radius $\theta_p$ of the perturber. We also require $b \gtrsim \Delta \theta$ so that Taylor expansion in $\Delta x$ is justified. Fractional correction to magnification from terms of type (i) is therefore $$\begin{aligned}
|a_\mu| \sim \left(\theta^2_p\,\Delta\theta\right)/\left(d\,b^4\right).\end{aligned}$$ The fractional correction to magnification from terms of type (ii) is of order $d\,\Delta\theta$, independent of the perturber. This is comparable to the inverse of the leading order magnification factor. For image pairs of interest in this work, the magnification factor is as large as $\gtrsim 100$. Hence, these terms only induce insignificant fractional magnification asymmetry at (sub-)percent levels.
\[lastpage\]
[^1]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The presence of the triplet $\Delta_{L}$ in left-right symmetric theories leads to type-II see-saw mechanism for the neutrino masses. In these models, assuming a normal mass hierarchy for the heavy Majorana neutrinos, we derive a lower bound on the mass of the lightest of heavy Majorana neutrino from the leptogenesis constraint. From this bound we establish a consistent picture for the hierarchy of heavy Majorana neutrinos in a class of left right symmetric models in which we identify the neutrino Dirac mass matrix with that of Fritzsch type charged lepton mass matrix. It is shown that these values are compatible with the current neutrino oscillation data.'
author:
- Narendra Sahu
- 'S. Uma Sankar'
title: '**Bounds on neutrino masses from leptogenesis in type-II see-saw models**'
---
Introduction
============
A plausible explanation of the observed baryon ($B$) asymmetry of the Universe is that it arose from a lepton ($L$) asymmetry [@fukugita.86; @luty.92; @mohapatra.92; @plumacher.96]. The conversion of the $L$-asymmetry to the $B$-asymmetry then occurs via the high temperature behavior of the $B+L$ anomaly of the Standard Model ($SM$) [@krs.86]. This is an appealing route for several reasons. The extremely small neutrino masses suggested by the solar [@solar] and atmospheric [@atmos] neutrino anomalies and the KamLAND experiment [@kamland], point to the possibility of Majorana masses for the neutrinos generated by the see-saw mechanism [@ge-ra-sl-ya]. This suggests the existence of new physics at a predictable high energy scale. Since the Majorana mass terms violate lepton number they can generate $L$-asymmetry.
Early proposals along these lines relied on out-of-equilibrium decay of the heavy Majorana neutrinos to generate the $L$-asymmetry. In the simplest scenario a right-handed neutrino per generation is added to the $SM$ [@fukugita.86; @luty.92; @plumacher.96]. They are coupled to left-handed neutrinos via Dirac mass matrix ($m_{D}$) which is assumed to be similar to charged lepton mass matrix [@ge-ra-sl-ya]. Since the right handed neutrino is a singlet under $SM$ gauge group a Majorana mass term ($M_{R}$) for it can be added to the Lagrangian. Diagonalisation of neutrino mass matrix leads to two Majorana neutrino states per generation: a light neutrino state (mass $\sim m_{D}^{2}/M_{R}$) which is almost left handed and a heavy neutrino state (mass $\sim M_{R}$) which is almost right handed. This is called type-I see-saw mechanism in which the left handed fields do not have Majorana mass terms in the Lagrangian.
It is desirable to consider neutrino masses in the context of grand unified theories ($GUTs$). The gauge groups of most of the GUTs contain the left-right symmetry group $SU(2)_L\times SU(2)_R$ as a subgroup [@slansky_rep]. In such models Majorana masses, $M_L$, for left handed neutrinos occur in general, through the vacuum expectation value (VEV) of the triplet $\Delta_L$ [@magg-wet.80; @wett.81; @moh-senj.81; @laz-shf-wet.81; @moha-susy-book.92]. In these models also there is a light and a heavy neutrino state per generation. The heavy neutrino state has mass $\sim M_{R}$ but the light neutrino mass is $\sim(M_{L}-m_{D}^{2}/M_{R})$. The presence of new scalars and their couplings, which give rise to $M_L$, can contain adequate $CP$-violation to accommodate $L$-asymmetry. The two contributions to the light neutrino mass, $m_D^2/M_R$ and $M_L$ are called type-I and type-II terms respectively.
An additional grace of left-right symmetric models [@leftright_group] is that $B-L$ is a gauge symmetry in contrast to type-I models where $B-L$ conservation is ad-hoc. Because of $B-L$ is a gauge charge of this model, no primordial $B-L$ can exist. Further, the rapid violation of the $B+L$ conservation by the anomaly due to the high temperature sphaleron fields erases any $B+L$ generated earlier. Thus the lepton asymmetry must be produced entirely during or after the $B-L$ symmetry breaking phase transition.
Several authors have recently dealt with the bound on the mass scale of lightest right handed neutrino $(M_1)$ in type-I see-saw models from the leptogenesis constraint. With the assumption of hierarchical mass spectrum of the heavy Majorana neutrinos, a common outcome was that $M_{1}\geq 10^{9}$ GeV. With the same assumption, a bound on $M_1$ was obtained in type-II see-saw models in [@antusch], which is an order of magnitude less than that in type-I case.
In this paper we revisit the lower bound on the mass scale of lightest heavy Majorana neutrino due to its CP-violating decay to SM particles in generic type-II see-saw models. Recently it is claimed that the large atmospheric neutrino mixing can be achieved naturally in case of renormalisable $SO(10)$ theories if the type-II term dominates [@baj-sen-vis.03; @bajcetal.04]. Therefore, it is interesting to extend our formalism to these models and derive an upper bound on the CP-asymmetry. In both cases it is shown that the mass scale of lightest right handed neutrino satisfies the constraint, $M_1\geq 2.5\times 10^{8}$GeV in order to produce the present baryon asymmetry of the Universe. On the other hand leptogenesis in models, where the type-II term was included in the neutrino mass matrix, was considered in the literature [@joshipura_npb.01; @joshipura_jhep.01]. However, the contribution to the CP-violating parameter, $\epsilon_1$, due to the triplet $\Delta_L$ in the loop, are not taken into account.
Rest of our paper is organised as follows. In section II, we derive an upper bound on the $CP$-asymmetry in type-II see-saw models assuming that type-I and type-II terms are similar in magnitude. In section III, we discuss the upper bound on the mass scale of lightest right handed neutrino from the leptogenesis constraint. In light of current neutrino oscillation data in section IV a consistent picture for the heavy Majorana neutrino mass hierarchies are obtained in a class of left-right symmetric models in which we identify the neutrino Dirac mass matrix with that of charged lepton mass matrix [@ge-ra-sl-ya]. Further we choose this matrix to be of Fritzsch type [@fritzsch.79]. Finally in Section V, we put our summary and conclusions.
Upper bound on $CP$-asymmetry in type-II models
===============================================
Before proceeding with our analysis, we review the breaking scheme of $SO(10)$ grand unified theory through the left-right symmetric path [@moh-senj.81; @moha-susy-book.92]. This breaking can be accomplished by using a $\{126\}$ of $SO(10)$ as an intermediate. Under $SU(2)_{L}\otimes SU(2)_{R}\otimes SU(4)_{C}(= SU(2)_{L}\otimes
SU(2)_{R}\otimes SU(3)_{C}\otimes U(1)_{B-L})$ its decomposition can be written as {126}=\_[L]{}(3,1,10)+\_[R]{}(1,3,10)+(2,2,15) +(1,1,6), where $\sigma(1,1,6)$ is a singlet under $SU(2)_{L}\otimes
SU(2)_{R}$ and has no role in neutrino mass generation. Since $126$ of $SO(10)$ contains a pair of triplets $\Delta_{L}$ and $\Delta_{R}$ and the bidoublet $\Phi$, left-right symmetry can be preserved at the intermediate level. As the right handed triplet $\Delta_{R}(1,3,10)$ gets a VEV $v_R$, left-right symmetry is broken to the SM symmetry. The scalars in both the triplets $\Delta_L$ and $\Delta_R$ acquire masses of the order $v_R$. At a lower scale, $\Phi$ gets a VEV $v$ breaking the SM symmetry to $U(1)_{em}$. This induces a small VEV $v_L$ for the neutral component of the triplet $\Delta_L$ [@moh-senj.81; @laz-shf-wet.81; @hamb-senj.03]. The three VEVs are related by $v_L = \gamma v^2/v_R$, where $\gamma$ is a model dependent parameter which depends on the quartic couplings of the Higgs and can be as small as $10^{-4}$.
After the final symmetry breaking, the effective neutrino mass matrix is M\_=|\_[Li]{}\_[D\_[ij]{}]{}’\_[Rj]{}+ +H.C. \[Yukawa\] Because of $L \leftrightarrow R$ symmetry, the same symmetric Yukawa matrix $f_{ij}$ gives rise to Majorana masses for both left and right handed fields. The Dirac mass matrix is given by $\tilde{m}_{D}=\tilde{h}v$, where $\tilde{h}$ is the Yukawa matrix for neutrino Dirac masses.
The Majorana mass matrix for the right handed neutrinos can be diagonalized by making the following rotation on $\nu'_R$ \_[R]{} = U\_[R]{}\^\_[R]{}\^[’]{}. In this basis, we have U\_[R]{}\^[T]{} f U\_[R]{} & = & f\_[dia]{}, \[fdia\]\
h &=& U\_[R]{}. In this rotated basis we get the mass matrix for the neutrinos
fv\_[L]{} & m\_[D]{}\
m\_[D]{}\^[T]{} & M\_[R]{}
, \[massmatrix\] where $M_{R}=f_{dia}v_{R}$ and $m_D = h v$. Diagonalising the mass matrix (\[massmatrix\]) into $3\times3$ blocks we get the mass matrix for the light neutrinos to be m\_ &=& fv\_[L]{}- h f\_[dia]{}\^[-1]{}h\^T\
&=&m\_\^[II]{}+m\_\^[I]{}. \[see-saw\] Note that in contrast to the present case in type-I models, $m_{\nu}^{II}$ is absent. Diagonalization of the above light neutrino mass matrix $m_{\nu}$, through lepton flavour mixing PMNS matrix $U_{L}$, gives us three light Majorana neutrinos. Its eigenvalues are U\_L\^ m\_ U\_L\^\*=dia(m\_[1]{}, m\_[2]{}, m\_[3]{})D\_m, \[diag\] where the masses are real.
We assume a normal mass hierarchy for heavy Majorana neutrinos. In this scenario while the heavier neutrinos, $N_2$ and $N_3$, decay yet the lightest of heavy Majorana neutrinos is still in thermal equilibrium. Any asymmetry produced by the decay of $N_2$ and $N_3$ will be washed out by the lepton number violating interactions mediated by $N_1$. Therefore, the final lepton asymmetry is given only by the CP-violating decay of $N_1$ to standard model leptons ($l$) and Higgs ($\phi$). The CP-asymmetry in this scenario is given by \_1=\_1\^I+\_1\^[II]{}, \[epsilon\] where the contribution to $\epsilon_1^I$ comes from the interference of tree level, self-energy correction and the one loop radiative correction diagrams involving the heavier Majorana neutrinos $N_2$ and $N_3$. This contribution is the same as in type-I models and is given by \_1\^[I]{}=. \[epsilonI\] On the other hand the contribution to $\epsilon_1^{II}$ in equation (\[epsilon\]) comes from the interference of tree level diagram and the one loop radiative correction diagram involving the triplet $\Delta_L$. It is given by [@hamb-senj.03; @antusch] \_1\^[II]{}=. \[epsilonII\] Substituting (\[epsilonI\]) and (\[epsilonII\]) in equation (\[epsilon\]) we get the total CP-asymmetry \_1 =. \[cpasym-1\] Using (\[diag\]) in the above equation (\[cpasym-1\]) we get \_1 &=&\
&=& . \[cpasym-2\] With an assumption of normal mass hierarchy for the light Majorana neutrinos the maximum value of CP-asymmetry (\[cpasym-2\]) can be given by \_[1,max]{} = m\_3. \[cpasym-3\] Note that the above upper bound (\[cpasym-3\]) for $\epsilon_1$ as a function of $M_1$ and $m_3$ was first obtained for the case of type-I see-saw models . However, the same relation holds in the case of type-II see-saw models also [@antusch] [*independent of the relative magnitudes of $m_{\nu}^{I}$ and $m_{\nu}^{II}$*]{}.
Estimation of Baryon Asymmetry
==============================
A net $B-L$ asymmetry is generated when left-right symmetry breaks. A part of this $B-L$ asymmetry then gets converted to $B$-asymmetry by the high temperature sphaleron transitions. However these sphaleron fields conserve $B-L$. Therefore, the produced $B-L$ will not be washed out, rather they will keep on changing it to $B$-asymmetry. In a comoving volume a net $B$-asymmetry is given by Y\_B &=& = \_1 Y\_[N\_1]{}, \[B-asym\] where the factor $\frac{28}{79}$ in front [@harvey.90] is the fraction of $B-L$ asymmetry that gets converted to $B$-asymmetry. Further $Y_{N_1}$ is density of lightest right handed neutrino in a comoving volume given by $Y_{N_1}=n_{N_1}/s$, where $s=(2\pi^{2}/45)g_{*}T^{3}$ is the entropy density of the Universe at any epoch of temperature $T$. Finally $\delta$ is the wash out factor at a temperature just below the mass scale of $N_1$. The value of $Y_{N_1}$ depends on the source of $N_1$. For example, the value of $Y_{N_1}$ estimated from topological defects [@sahuetal.04] can be different from thermal scenario [@plumacher.96; @buch-bari-plumcher-03]. In the present case we will restrict ourselves to thermal scenario only.
Recent observations from WMAP show that the matter-antimatter asymmetry in the present Universe measured in terms of $(n_B/n_\gamma)$ is [@spergel.03] ()\_[0]{}(6.1\^[+0.3]{}\_[-0.2]{})10\^[-10]{}, \[baryon-asym\] where the subscript $0$ presents the asymmetry today. Therefore, we recast equation (\[B-asym\]) in terms of $(n_B/n_\gamma)$ and is given by ()\_[0]{} &=& 7 (Y\_[B]{})\_[0]{}\
&=& 2.48 \_[1]{} Y\_[N\_1]{}. \[baryon-asym1\] Substituting equation (\[cpasym-3\]) in (\[baryon-asym1\]) we get ()\_[0]{}2.48( ) m\_3 Y\_[N\_1]{}. \[baryon-asym2\] The present neutrino oscillation data favours the bilarge neutrino mixing with the mass squared differences $\Delta m^2_{\rm atm}\equiv |m_3^2-m_1^2|\approx
2.6\times10^{-3} eV^2$ and $\Delta m^2_{\rm sol}\equiv
|m_2^2-m_1^2|\approx 7.1\times 10^{-5} eV^2$. Assuming a normal mass hierarchy ($m_3^2\gg m_2^2\gg m_1^2$) for the light Majorana neutrinos the above number gives $m_3\simeq
\left(\Delta m^2_{\rm atm}\right)^{1/2}\simeq 0.05 eV$. With this approximation we get from equation (\[baryon-asym2\]) M\_12.510\^[8]{} GeV ( )().
For the above lower limit on the mass scale of lightest right handed neutrino, $M_1\geq 10^{8}$ GeV, we now proceed to find the plausible hierarchies $(M_2/M_1)$ and $(M_3/M_1)$ of the massive Majorana neutrinos in a model that are compatible with the current neutrino oscillation data. We check that the values we obtained are consistent with our assumptions.
Examining the consistency of f-matrix eigenvalues
=================================================
The solar and atmospheric evidences of neutrino oscillations are nicely accommodated in the minimal framework of the three-neutrino mixing, in which the three neutrino flavours $\nu_{e}$, $\nu_{\mu}$, $\nu_{\tau}$ are unitary linear combinations of three neutrino mass eigenstates $\nu_{1}$, $\nu_{2}$, $\nu_{3}$ with masses $m_{1}$, $m_{2}$, $m_{3}$ respectively. The mixing among these three neutrinos determines the structure of the lepton mixing matrix [@mns-matrix] which can be parameterized as U\_[L]{}=
c\_[1]{}c\_[3]{} & s\_[1]{}c\_[3]{} & s\_[3]{}e\^[i]{}\
-s\_[1]{}c\_[2]{}-c\_[1]{}s\_[2]{}s\_[3]{}e\^[i]{} & c\_[1]{}c\_[2]{}-s\_[1]{}s\_[2]{}s\_[3]{} e\^[i]{} & s\_[2]{}c\_[3]{}\
s\_[1]{}s\_[2]{}-c\_[1]{}c\_[2]{}s\_[3]{} & -c\_[1]{}s\_[2]{}-s\_[1]{}c\_[2]{}s\_[3]{}e\^[i]{} & c\_[2]{}c\_[3]{}
dia(1, e\^[i]{}, e\^[i(+)]{}), \[mns-matrix\] where $c_{j}$ and $s_{j}$ stands for $cos \theta_{j}$ and $sin \theta_{j}$. The two physical phases $\alpha$ and $\beta$ associated with the Majorana character of neutrinos are not relevant for neutrino oscillations [@bilenkyetal.80] and will be set to zero here onwards. While the Majorana phases can be investigated in neutrinoless double beta decay experiments [@rodejohan_npb.01], the CKM-phase $\delta \in [-\pi, \pi]$ can be investigated in long base line neutrino oscillation experiments. For simplicity we set it to zero, since we are interested only in the magnitudes of elements of $U_{L}$. The best fit values of the neutrino masses and mixings from a global three neutrino flavors oscillation analysis are [@gonzalez-garcia_prd.03] \_[1]{}=\_34\^, \_[2]{}=\_[atm]{} =45\^, \_3 13\^, \[bestfit-theta\] and m\_\^[2]{}= m\_2\^2 - m\_1\^2 & & m\_2\^2 = 7.110\^[-5]{} [eV]{}\^[2]{}\
m\_[atm]{}\^[2]{}= m\_3\^2 - m\_1\^2 & & m\_3\^2 = 2.610\^[-3]{} [eV]{}\^[2]{}. From equation (\[see-saw\]) and (\[diag\]) we have f=. \[f-matrix\] We now assume a hierarchical texture for the Majorana Yukawa coupling to be f\_[dia]{}=
1 & 0 & 0\
0 & \_[A]{} & 0\
0 & 0 & \_[B]{}
, \[fdia-texture\] where $1 \ll \alpha_{A}=(M_{2}/M_{1}) < \alpha_{B}=
(M_{3}/M_{1})$. We identify the neutrino Dirac Yukawa coupling $h$ with that of charged leptons [@ge-ra-sl-ya]. Further we assume it to be of Fritzsch type [@fritzsch.79] h=
0 & a & 0\
a & 0 & b\
0 & b & c
. \[h-texture\] By choosing the values of a,b and c suitably one can get the hierarchy for charged leptons. In particular, the values a=0.004, b=0.24 [and]{} c=1 \[abcvalues\] give the mass hierarchy of charged leptons. For this set of values the mass matrix h is normalized with respect to the $\tau$-lepton mass. The values of $a, b$ and $c$ are rougly in geometric progression. They can be expressed in terms of the electro-weak gauge coupling $\alpha_{w}=
\frac{g^{2}}{4\pi}=\frac{\alpha}{sin^{2}\theta_{w}}$. In particular $a=2.9 \alpha_{w}^{2}$, $b=6.5 \alpha_{w}$ and $c=1$. Here onwards we will use this set of values for the parameters of $h$.
Substituting (\[fdia-texture\]) and (\[h-texture\]) in equation (\[f-matrix\]) we get f=(), \[maj-yukawa\] For the demonstration purpose we choose a typical value of $M_1=2.5\times 10^{8}$ GeV. Now by suitably choosing the parameters $m_{1}=1.0 \times 10^{-4}eV$, $\alpha_{A}=51$, $\alpha_{B}=191$, $\theta_{13}=10^\circ$ we get f\_[dia]{}=
1 & 0 & 0\
0 & 50.98 & 0\
0 & 0 & 190.47
. \[fdia-cal\] Thus, for the above values of $m_1$ and $M_1$, the assumed hierarchies of right-handed neutrino masses are consistent with global low energy neutrino data. Further we emhasize that the value of $m_1$ for which the consistency is obtained is compatible with our assumption that m\_1\^2m\_3\^2m\^2\_[atm]{}.
Comparing equation (\[fdia-cal\]) with (\[fdia-texture\]) one can get =. This implies that $v_{R}=O(10^{11})$ GeV for $v_{L}=0.1$ eV. These values of $v_{L}$ and $v_{R}$ are compatible with genuine see-saw $v_{L}v_{R}=\gamma v^{2}$ for a small value of $\gamma\simeq O(10^{-3})$. Although in the literature frequently it is stated $\gamma$ is to be order of unity, we see, however, here that it is required to be $O(10^{-3})$ for the consistency. See, for example, the recently proposed type-II see-saw models in which $\gamma=v_L v_R/v^2 \geq O(10^{-4})$ [@nimai.04].
Here we demonstrated the consistency of our choice of the matrix $f$ with the current neutrino data for the minimum value of $M_1$. For higher values of $M_1$, to be consistent with $1<<\alpha_A <\alpha_B$, one can choose appropriate values of $m_1\leq 10^{-3}$ eV and $\theta_{13}\leq 13^{\circ}$ in equation (\[maj-yukawa\]) which will reproduce the correct eigenvalues of the matrix $f$.
Conclusion
==========
In this work we proved the universality of the upper bound on the $CP$-violating parameter $\epsilon_{1}$ in general type-II see-saw models. Irrespective of any assumption regarding the magnitude of type-I and type-II terms we found that same bound holds for all cases. Assuming a normal mass hierarchy for the massive Majorana neutrinos we demonstrated that in a thermal scenario the present baryon asymetry can be explained for all values of $M_1\geq
O(10^8)GeV$ in case of left-right symetric models. We also consider a class of left-right symmetric models in which we assume the neutrino Dirac masses are of the Fritzsch type. In such models we found that plausible hierarchies for the massive Majorana neutrinos can be obtained which are compatible with the current neutrino oscillation data.
Acknowledgment {#acknowledgment .unnumbered}
==============
It is our pleasure to thank Prof. U. A. Yajnik for his critical comments on the draft. We also thank Prof. R. N. Mohapatra and Prof. M. K. Parida for their interest and valuable discussions during the WHEPP-8 held at IIT Bombay.
[99]{} M. Fukugita and T. Yanagida, Phys. Lett. [**B174**]{}, 45 (1986).
M. A. Luty, Phys. Rev.D[**45**]{}, 455 (1992).
R. N. Mohapatra and X. Zhang; Phys. Rev. D[**46**]{}, 5331 (1992).
M. Plumacher, Z. Phys. [**74**]{}, 549 (1997).
V. A. Kuzmin, V. A. Rubakov and M. E. Shaposhnikov, Phys. Lett. B [**155**]{}, 36 (1985).
Q. R. Ahmed [*et al*]{} (SNO Collaboration), Phys.Rev.Lett. [**89**]{}, 011301-011302 (2002); J. N. Bahcall and C. Pena-Garay, hep-ph/0404061.
S. Fukuda [*et al*]{} (Super-Kamiokande Collaboration), Phys. Rev. Lett. [**86**]{}, 5656 (2001).
K. Eguchi [*et al*]{} (KamLAND collaboration), Phys. Rev. Lett. [**90**]{}, 021802 (2003).
M. Gell-Mann, P. Ramond and R. Slansky in [*Supergravity*]{} (P. van Niewenhuizen and D. Freedman, eds), (Amsterdam), North Holland, 1979; T. Yanagida in [*Workshop on Unified Theory and Baryon number in the Universe*]{} (O. Sawada and A. Sugamoto, eds), (Japan), KEK 1979; R. N. Mohapatra and G. Senjanovic, Phys. Rev. Lett. [**44**]{}, 912 (1980).
R. Slansky, Phys. Rep. [**79**]{}, 1(1981).
M. Magg and C. Wetterich, Phys. Lett. B [**94**]{}, 61 (1980).
C. Wetterich, Nucl. Phys. B [**187**]{}, 343 (1981).
R. N. Mohapatra and G. Senjanovic Phys. Rev. D[**23**]{}, 165 (1981).
G. Lazarides, Q. Shafi and C. Wetterich, Nucl. Phys. B[**181**]{}, 287 (1981).
R. N. Mohapatra,[*Unification And Supersymmetry*]{}, (Springer-Verlag, New-York, 1992).
J. C. Pati and A. Salam, Phys. Rev. D [**10**]{}, 275 (1974); R. N. Mohapatra and J. C. Pati, Phys. Rev. D[**11**]{}, 566(1975); Phys. Rev. D [**11**]{}, 2558 (1975); R. N. Mohapatra and G. Senjanovic, Phys. Rev. D[**12**]{}, 1502 (1975); See also ref.\[13,15\]
S. Davidson and A. Ibarra, Phys. Lett. B [**535**]{}, 25 (2002). \[arXiv:hep-ph/0202239\]
W. Buchmuller, P. Di Bari and M. Plumacher, Nucl. Phys. B [**643**]{}, 367 (2002). \[arXiv:hep-ph/0205349\].
K. Hamaguchi, H. Murayama and T. Yanagida, Phys. Rev. D[**65**]{}, 043512 (2002).
T. Hambye [*et al*]{}, arXiv:hep-ph/0312203
S. Antusch and S. F. King, Phys. Lett. B [**597**]{}, 199 (2004). \[arXiv:hep-ph/0405093\]
B. Bajc, G. Senjanovic, F. Vissani, Phys. Rev. Lett. [**90**]{}: 051802, 2003
B. Bajc, G. Senjanovic and F. Vissani arXiv:hep-ph/0402140
A. Joshipura, E. A. Paschos and W. Rodejohann, Nucl. Phys. B [**611**]{}, 227, 2001.
A. Joshipura, E. A. Paschos, and W. Rodejohann, JHEP[**08**]{} 029 (2001).
H. Fritzsch, Nucl. Phys. B [**155**]{}, 189 (1979).
T. Hambye and G. Senjanovic, Phys. Lett. B [**582**]{}, 73 (2004).
J. A. Harvey and M. S. Turner, Phys. Rev. D[**42**]{}, 3344 (1990).
P. Bhattacharjee, Narendra Sahu and U. A. Yajnik, Phys. Rev. D, [**70**]{}, 083534, 2004 \[arXiv:hep-ph/0406054\].
D. N. Spergel [*et al*]{} Astrophys. J. Suppl. [**148**]{} 175 (2003) \[astro-ph/0302209\]
Z. Maki, M. Nakagawa and S. Sakata, Prog. Theor. Phys. [**28**]{}, 870 (1962).
S. M. Bilenky, J. Hosek and S. T. Petcov, Phys. Lett. B [**94**]{}, 495 (1980).
W. Rodejohann, Nucl. Phys. B [**597**]{}, 110 (2001).
M. C. Gonzalez-Garcia and C. Pena-Garay, Phys. Rev. D[**68**]{}, 093003 (2003).
N. Nimai Singh, M. Patagiri, Mrinal Das, \[arXiv:hep-ph/0406075\].
|
{
"pile_set_name": "ArXiv"
}
|
---
bibliography:
- 'auto\_generated.bib'
title: 'Search for resonant pair production of Higgs bosons in the bbZZ channel in proton-proton collisions at sqrt(s) = 13 TeV'
---
=1
Introduction
============
The discovery of the Higgs boson ($\Ph$) in 2012 [@Chatrchyan:2012xdj; @CMS:2012nga; @Aad:2012tfa] has led to a detailed program of studies of the Higgs field couplings to the elementary particles of the standard model (SM) of particle physics: leptons, quarks, and gauge bosons. To fully understand the form of the Higgs field potential, which is a key element in the formulation of the SM, it is important to also study the self-interaction of the Higgs boson. The self-interaction can be investigated through measurements of the production of a pair of Higgs bosons ($\Ph\Ph$). In the SM, $\Ph\Ph$ production is a rare, nonresonant process, with a small production rate [@HiggsXS] that will require the future data sets of the high-luminosity LHC to be observed [@HiggsXS]. Hence, an early observation of $\Ph\Ph$ production, a resonant production in particular, would be a spectacular signature of physics beyond the standard model (BSM). The production of gravitons, radions, or stoponium [@Tang:2012pv; @PhysRevD.63.056007; @PhysRevD.90.055007], for example, could lead to $s$-channel $\Ph\Ph$ production via narrow-width resonances. The breadth of the Higgs boson decay channels provides a unique opportunity to test the self-consistency of an $\Ph\Ph$ signal with the SM or models with extended electroweak sectors, such as two-Higgs doublet models (2HDM) [@2HDM1; @2HDM2] or extensions of the minimal supersymmetric standard model [@hH2; @hH3; @hH5].
This paper reports a search for resonant $\Pp\Pp\to\mathrm{X}\to\PH\PH$ production in the $\PH\PH\to{\ensuremath{\PQb\PQb\PZ\PZ}\xspace}$ decay channel, where X is a narrow-width resonance of spin-0 or spin-2, and $\PH$ can represent either $\Ph$ or an additional Higgs boson from an extended electroweak sector. The search uses proton-proton ($\Pp\Pp$) collision data at $\sqrt{s}=13\TeV$, recorded with the CMS detector at the LHC in 2016, and corresponding to an integrated luminosity of 35.9. It covers a range of resonance masses between 260 and 1000. The final state consists of two jets from one Higgs boson decay and two distinct $\PZ$ boson decay signatures from the other $\PH\to\PZ\PZ$ decay: two same-flavor, opposite-sign (OS) leptons from a decay of one of the bosons, and either two jets of any flavor (the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel) or significant missing transverse momentum (the [$\PQb\PQb\ell\ell\nu\nu$]{}channel) from the decay of the second boson to neutrinos. In both cases, the selected charged leptons are either electrons or muons. In the SM, the branching fractions of these signatures represent 0.43 (0.12)% of the full $\Ph\Ph$ decay through the [$\PQb\PQb\PZ\PZ$]{}intermediate state in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}([$\PQb\PQb\ell\ell\nu\nu$]{}) channel. The challenging aspect of the search in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel is the ability to discriminate the signal containing two jets and two additional jets from multijet background events. For a search in the [$\PQb\PQb\ell\ell\nu\nu$]{}channel, the challenge resides in discriminating the signal against top quark anti-top quark ($\ttbar$) events and instrumental background sources of large missing transverse momentum arising from the mismeasurement of the energies of jets in the final state. The two channels are kept independent by applying orthogonal selections on the missing transverse momentum of the event. Signal yields are calculated for each individual channel and are then combined. Having multiple decay channels with complementary background compositions and sensitivities over a large resonance mass ($m_X$) range makes this combination of the [$\PQb\PQb\ell\ell\nu\nu$]{}and [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channels highly efficient for covering the [$\PQb\PQb\PZ\PZ$]{}final state. This is the first search for Higgs boson resonant pair production in the [$\PQb\PQb\PZ\PZ$]{}channel.
Previous searches for resonant $\Ph\Ph$ production have been performed by the CMS and ATLAS Collaborations in the ${\ensuremath{\PQb\PQb}\xspace}{\ensuremath{\PQb\PQb}\xspace}$ [@hh4b; @Aaboud:2018knk], ${\ensuremath{\PQb\PQb}\xspace}\tau\tau$ [@bbtautau; @Aaboud:2018bun], ${\ensuremath{\PQb\PQb}\xspace}\gamma\gamma$ [@bbgammagamma], and ${\ensuremath{\PQb\PQb}\xspace}\ell\nu\ell\nu$ [@bbWW; @Aaboud:2018bun] channels. While coverage of as many $\Ph\Ph$ decay channels as possible remains necessary to understand the exact nature of the Higgs boson self-coupling and the electroweak symmetry breaking mechanism, a [$\PQb\PQb\PZ\PZ$]{}search is particularly interesting in models with extended electroweak sectors, where the phenomenology of additional Higgs bosons can lead to significantly enhanced [$\PQb\PQb\PZ\PZ$]{}production, while suppressing the BSM production of ${\ensuremath{\PQb\PQb}\xspace}{\ensuremath{\PQb\PQb}\xspace}$, ${\ensuremath{\PQb\PQb}\xspace}\tau\tau$, or ${\ensuremath{\PQb\PQb}\xspace}\gamma\gamma$ final states.
Benchmark models {#benchmarks}
================
As in the previous searches, a class of narrow width resonance models arising from the Randall–Sundrum (RS) model [@RandallSundrum] in warped extra dimensions [@radion1; @radion2; @radion3; @radion4] are considered. This scenario introduces one small spatial extra dimension with a nonfactorizable geometry, where the SM particles are not allowed to propagate along that extra dimension, and is referred to in this search as RS1. The resonant particle can be a radion (spin-0) or the first Kaluza–Klein (KK) excitation of a graviton (spin-2). The production cross section of the radion is proportional to $1/\lambda_{\mathrm{R}}^{2}$ where $\lambda_{\mathrm{R}}$ is the interaction scale parameter of the theory. In this analysis, we consider the cases where $\lambda_{\mathrm{R}}=1\TeV$ with $kL = 35$, where $k$ is the constant in the warp factor ($e^{-kL}$) appearing in the space-time metric of the theory and $L$ is the size of the extra dimension. The free parameter of the model for the graviton case is $\tilde{k} = k/\overline{M}_{\mathrm{Pl}}$, where $\overline{M}_{\mathrm{Pl}}$ is the reduced Planck scale, and we consider $\tilde{k}=0.1$ in this analysis [@Oliveira:2014kla]. We further scan the model parameter space in the $\lambda_{\mathrm{R}}$ and $\tilde{k}$ parameters for their respective models. Production at hadron colliders is expected to be dominated by gluon-gluon fusion, and we assume that the radion or graviton is produced exclusively via this process. Due to the small branching fraction of $\Ph\Ph\to{\ensuremath{\PQb\PQb\PZ\PZ}\xspace}$ and the high multiplicities of the final states, the analyses presented in this paper are less sensitive to these models compared to the previous searches. As noted in Section \[introduction\], however, certain models with extended electroweak sectors can produce significantly enhanced [$\PQb\PQb\PZ\PZ$]{}production, while suppressing final states with Higgs boson decays to fermions and scalar bosons.
Such an enhancement can be produced for example in the next-to-minimal 2HDM (N2HDM) extended Higgs sector [@N2HDM; @N2HDM1], where an additional real singlet is introduced in addition to the usual two doublet Higgs bosons of the 2HDM. This analysis is further interpreted in this context. The so-called broken phase is considered, wherein both the Higgs doublets and the singlet acquire vacuum expectation values (vev) [@N2HDM1]. Mixing between these states produces 3 $CP$-even Higgs bosons ${\PH}_1$, ${\PH}_2$, and ${\PH}_3$, with masses that are free parameters of the model. This search considers the nearly mass-degenerate case where the masses of the two bosons ${\PH}_1$ and ${\PH}_2$ are constrained to the experimental measurements of the $\Ph$ mass, which would be indistinguishable from $\Ph$ production with current LHC data sets [@hH1; @hH4; @hH2], but may give rise to manifestly non-SM-like rates in the case of $\Ph\Ph$ production. In what is commonly referred to as Higgs boson cascade decays, ${\PH}_3$ can decay to any combination of bosons ${\PH}_1$ and ${\PH}_2$, which then both can have different decay branching fractions compared to the SM Higgs boson. The model spectrum depends on the ratio of the vevs of the two Higgs doublets $\tan\beta$, low values of which enhance ${\PH}_3$ production; the vev of the singlet, which affects the decay branching fractions of ${\PH}_3$ to ${\PH}_1$ and ${\PH}_2$; and three mixing angles, which determine the decay branching fractions of ${\PH}_1$ and ${\PH}_2$ [@N2HDM1]. The model spectra described below are determined using <span style="font-variant:small-caps;">N2HDECAY</span> [@N2HDM2], and are chosen to enhance production of the [$\PQb\PQb\PZ\PZ$]{}final state while roughly respecting current LHC measurements of the SM $\Ph$ branching fractions [@HiggsXS]. The gluon-gluon fusion production cross sections of ${\PH}_3$ are determined from the BSM Higgs boson predictions of the LHC Higgs Cross Section Working Group [@HiggsXS]. These cross sections assume SM decay branching fractions of the Higgs boson, and changing these branching fractions affects the production cross section. The cross sections are corrected at leading order (LO) by the ratio of the relative partial width of ${\PH}_3$ in the decay to two gluons compared to the BSM Higgs boson prediction. Enhanced (reduced) coupling of ${\PH}_3$ to gluons will enhance (reduce) the production cross section of ${\PH}_3$. The mass of the Higgs bosons ${\PH}_1$ and ${\PH}_2$ are set to 125, and the mass of ${\PH}_3$ is generated in the range $260\leq m_\mathrm{X}\leq 1000\GeV$. Two benchmark points are chosen for this analysis, corresponding to $\tan\beta=0.5$ and 2.0. In both cases, the scalar vev is set to 45, and the mixing angles $\alpha_1$, $\alpha_2$, $\alpha_3$ are set to 0.76, 0.48, and 1.00, respectively. For $\tan\beta=0.5$, this results in branching fractions of ${\PH}_3$ to ${\PH}_1{\PH}_1$, ${\PH}_1{\PH}_2$, and ${\PH}_2{\PH}_2$ around 0.02, 0.29, and 0.64 respectively, branching fractions of ${\PH}_1\to\PQb\PQb$ (${\PH}_1\to\PZ\PZ$) of 0.70 (0.01), and branching fractions of ${\PH}_2\to\PQb\PQb$ (${\PH}_2\to\PZ\PZ$) of 0.42 (0.05). This represents a 33% increase in the branching fraction to [$\PQb\PQb\PZ\PZ$]{}compared to SM $\Ph\Ph$ decays. The correction factor based on the relative partial width of ${\PH}_3$ to two gluons is around 3.0. For $\tan\beta=2.0$, this results in branching fractions of ${\PH}_3$ to ${\PH}_1{\PH}_1$, ${\PH}_1{\PH}_2$, and ${\PH}_2{\PH}_2$ around 0.07, 0.22, and 0.67 respectively, branching fractions of ${\PH}_1\to\PQb\PQb$ (${\PH}_1\to\PZ\PZ$) of 0.53 (0.03), and branching fractions of ${\PH}_2\to\PQb\PQb$ (${\PH}_2\to\PZ\PZ$) of 0.58 (0.03). This represents a 5% increase in the branching fraction to [$\PQb\PQb\PZ\PZ$]{}compared to SM $\Ph\Ph$ decays. The correction factor based on the relative partial width of ${\PH}_3$ to two gluons is around 0.7. These corrections and branching fractions produce significant differences in the production rates of the [$\PQb\PQb\PZ\PZ$]{}system compared to $\Ph\Ph$ production both in the SM and through resonant production of radions or gravitons.
The CMS detector {#cms}
================
The central feature of the CMS apparatus is a superconducting solenoid of 6 internal diameter, providing a magnetic field of 3.8. Within the solenoid volume are a silicon pixel and strip tracker, a lead tungstate crystal electromagnetic calorimeter (ECAL), and a brass and scintillator hadron calorimeter (HCAL), each composed of a barrel and two endcap sections. Forward calorimeters extend the pseudorapidity coverage provided by the barrel and endcap detectors, where pseudorapidity is defined as $\eta = -\ln[\tan(\theta /2)]$, and $\theta$ is the polar angle. Muons are measured in gas-ionization detectors embedded in the steel flux-return yoke outside the solenoid. CMS uses a two-level trigger system [@Khachatryan:2016bia]. The first level of the CMS trigger system, composed of custom hardware processors, uses information from the calorimeters and muon detectors to select the most interesting events. The high-level trigger (HLT) processor farm further decreases the event rate from around 100 to a rate of around 1, before data storage. A more detailed description of the CMS detector, together with a definition of the coordinate system used and the relevant kinematic variables, can be found in Ref. [@Chatrchyan:2008aa].
Event simulation {#samples}
================
The signal samples of RS1 spin-0 radion and RS1 KK spin-2 graviton narrow resonances decaying to a pair of Higgs bosons ($\mathrm{X}\to\Ph\Ph$) are generated at LO using . The $\Ph$ mass is set to 125, and the $m_X$is generated in the range of 260–1000. In the [$\PQb\PQb\ell\ell\nu\nu$]{}channel the final state can be produced via either the [$\PQb\PQb\PZ\PZ$]{}or $\PQb\PQb\PWp\PWm$ intermediate states.
The main background processes to production of a pair of Higgs bosons in the ${\ensuremath{\PQb\PQb\PZ\PZ}\xspace}\to$ [$\PQb\PQb\ell\ell\mathrm{jj}$]{}or [$\PQb\PQb\ell\ell\nu\nu$]{}final states are [$\PZ/\gamma^{*}\text{+jets}$]{}and $\ttbar$ processes. Less significant backgrounds arise from single top quark, $\PW$+jets, diboson+jets, SM Higgs boson production, and quantum chromodynamics (QCD) multijet production. Signal and background processes are modeled with simulations, with the exception of the QCD multijet background that is estimated using data control regions.
In the analysis using the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel, the [$\PZ/\gamma^{*}\text{+jets}$]{}and $\PW$+jets processes are generated with 2.4.2 [@Alwall:2014hca] at next-to-leading order (NLO). In this case, the generator uses the <span style="font-variant:small-caps;">FxFx</span> jet merging scheme [@Frederix:2012ps]. The analysis of the [$\PQb\PQb\ell\ell\nu\nu$]{}channel uses samples of [$\PZ/\gamma^{*}\text{+jets}$]{}generated with at LO, with the MLM matching scheme [@Alwall:2007fs], and reweighted to account for higher order QCD and electroweak effects [@DY_QCDnEWK].
=800 The process is generated at NLO with 2.0 [@Nason:2004rx; @Frixione:2007vw; @Alioli:2010xd; @Alioli:2009je; @Re:2010bp; @Alioli:2008tz]. Single top processes and SM Higgs boson production processes are simulated at NLO either with or , depending on the particular channel. The diboson processes ($\PW\PW$+jets, $\PW\PZ$+jets, $\PZ\PZ$+jets) are simulated at NLO with .
The simulated samples are normalized to their best-known highest-order-QCD cross sections, either evaluated at NLO with [@Campbell:2010ff] (diboson+jets) or at next-to-next-to-leading order with 3.1 [@Li:2012wna] (single top quark, $\PW$+jets, SM Higgs boson), with the exception of $\ttbar$ and [$\PZ/\gamma^{*}\text{+jets}$]{}processes, which are normalized using data.
The simulated samples are interfaced with 8.212 [@pythia8p2] for parton showering and hadronization. The generator uses the CUETP8M1 underlying event tune [@Khachatryan:2015pea]. The <span style="font-variant:small-caps;">nnpdf3.0</span> NLO and LO parton distribution functions (PDFs) [@Ball:2014uwa] are used for the various processes, with the precision matching that in the matrix element calculations.
For all the simulated samples used in this analysis, a simulation of CMS detector response based on [@GEANT4] is applied. The presence of additional interactions in the same bunch crossing (pileup, or PU), both in-time and out-of-time with respect to the primary interaction, is simulated and corrected to agree with a multiplicity corresponding to the distribution measured in data.
Event reconstruction and background estimation {#selection}
==============================================
Event reconstruction {#reconstruction}
--------------------
Events are selected using triggers that require two muons with transverse momentum $\pt >17$ (8)or two electrons with $\pt >23$ (12)for the leading (sub-leading) lepton.
The particle-flow (PF) algorithm [@PFlast], which combines information from various elements of the CMS detector, is used to reconstruct and identify final-state particles, such as photons, electrons, muons, and charged and neutral hadrons, as individual PF objects. Combinations of PF objects are then used to reconstruct higher-level objects such as jets and missing transverse momentum.
Jets are reconstructed from the PF objects, using the anti- [@antikt; @fastjet] algorithm with a distance parameter of $R = 0.4$. In order to reduce instrumental backgrounds and the contamination from PU, selected jets are required to satisfy loose identification criteria [@CMS:2017wyc] based on the multiplicities and energy fractions carried by charged and neutral hadrons. The energy of reconstructed jets is calibrated using $\pt $- and $\eta$-dependent corrections to account for nonuniformity and nonlinearity effects of the ECAL and HCAL energy response to neutral hadrons, for the presence of extra particles from PU, for the thresholds used in jet constituent selection, reconstruction inefficiencies, and possible biases introduced by the clustering algorithm. These jet energy corrections are extracted from the measurement of the momentum balance in dijet, $\text{photon}+\text{jet}$, [$\PZ/\gamma^{*}\text{+jets}$]{}, and multijet events [@Khachatryan:2016kdb]. A residual $\eta$- and $\pt $-dependent calibration is applied to correct for the small differences between data and simulated jets. The jets that are candidates to be from the decay of one of the Higgs bosons and of one of the bosons are required to have $\pt >20\GeV$. Furthermore, jets are required to have a spatial separation of $\Delta R > 0.3$ from lepton candidates.
Jets originating from quarks are identified with the combined multivariate analysis (cMVA) algorithm [@Sirunyan:2017ezt]. A jet is tagged as a jet if the cMVA discriminant is above a certain threshold, chosen such that the misidentification rate is about 1% for light-flavor quark and gluon jets, and about 13% for charm quark jets. The jet tagging efficiency for this working point is about 66%.
The missing transverse momentum vector is computed as the negative vector sum of the transverse momenta of all the PF candidates in an event, and its magnitude is denoted as [@Sirunyan:2019kia]. The is modified to account for corrections to the energy scale of the reconstructed jets in the event.
The candidate vertex with the largest value of summed physics-object $\pt^2$ is taken to be the primary $\Pp\Pp$ interaction vertex. The physics objects are the jets, clustered using the jet finding algorithm [@antikt; @fastjet] with the tracks assigned to candidate vertices as inputs, and the associated missing transverse momentum, taken as the negative vector sum of the of those jets.
Muons are reconstructed as tracks in the muon system that are matched to the tracks reconstructed in the inner silicon tracking system [@MuId]. The leading muon is required to have $\pt > 20\GeV$, while the subleading muon must have $\pt > 15$ (10)in the [$\PQb\PQb\ell\ell\nu\nu$]{}([$\PQb\PQb\ell\ell\mathrm{jj}$]{}) channel. Muons are required to be reconstructed in the HLT fiducial volume, , with $\abs{\eta} < 2.4$, to ensure that the offline selection is at least as restrictive as the HLT requirements. The selected muons are required to satisfy a set of identification requirements based on the number of spatial measurements in the silicon tracker and in the muon system and the fit quality of the combined muon track [@Khachatryan:2015pea], and are required to be consistent with originating from the primary vertex.
Electrons are reconstructed by matching tracks in the silicon tracker to the clusters of energy deposited in the ECAL [@EMId]. The leading (subleading) electron is required to have $\pt > 25$ (15)and $\abs{\eta} < 2.5$ to be within the geometrical acceptance, excluding candidates in the range $1.4442 < \abs{\eta} < 1.5660$, which is the transition region between the ECAL barrel and endcaps, because the reconstruction of an electron in this region is poor compared to other regions. Electrons are required to pass an identification requirement based on an MVA [@TMVA] technique that combines information from various observables related to the shower shape in the ECAL and the quality of the matching between the tracks and the associated ECAL clusters [@EMId]. They are further required to be consistent with originating from the primary vertex. Candidates that are identified as originating from photon conversions in the material of the detector are removed.
Both muons and electrons have a requirement that the lepton relative isolation, defined in Eq.(\[isoeq\]), be less than 0.25 (0.15) and 0.15 (0.06), respectively, for the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}([$\PQb\PQb\ell\ell\nu\nu$]{}) channel. In Eq.(\[isoeq\]), the sums run over charged hadrons originating from the primary vertex of the event, neutral hadrons, and photons inside a cone of radius $\Delta R = \sqrt{\smash[b]{(\Delta\phi )^2 + (\Delta\eta) ^2}} < 0.4$ (0.3) around the direction of the muon (electron), where $\phi$ is the azimuthal angle in radians.
Simulated background and signal events are corrected with scale factors for differences observed between data and simulation, in trigger efficiencies, in lepton $\pt $- and $\eta$-dependent identification and isolation efficiencies, and in tagging efficiencies.
Event selection in the bblljj channel {#lljjselection}
-------------------------------------
After selection of the candidate physics objects, an initial event selection is performed by requiring at least two same-flavor leptons (muons or electrons) in each event. The two leptons are required to be oppositely charged. The invariant mass of the two leptons, $m_{\ell\ell}$, is required to be larger than 15. Four of the jets in an event are designated as the $\Ph$ and boson decay products. These jets are required to have $\pt >20\GeV$ and at least one of those must be tagged with a minimum requirement on the tagging discriminant, that is looser than the requirement in the final selection. We refer to this selection as the preselection.
Since the signal contains two jets from the decay of a Higgs boson, and two jets of any flavor from the decay of a boson, it is important to carefully categorize the jets in the event. Starting from a collection of jets identified as described above, the information from the tagging discriminant, as well as the kinematic properties of the jets, are taken into account when assigning jets as each particle’s decay products.
The following selection is applied to identify the jets originating from the decay of the Higgs boson. The two jets with the highest tagging scores above a certain threshold are assigned to the decay of the Higgs boson. If only one jet is found that meets the minimum tagging score value, a second jet that leads to an invariant mass closest to 125 is selected. If no jets with tagging scores above threshold are found, the two jets whose invariant mass is closest to 125 are chosen.
After jets are assigned to the decay of [$\Ph\to\PQb\PQb$]{}, from the remaining jets the two jets with four-object invariant mass $M(\ell\ell\text{jj})$ closest to 125are assigned to the decay of the boson.
After preselection, additional requirements are imposed. At least one of the four jets assigned as the decay products of the $\Ph$ or boson must satisfy the tagging requirement, to increase the signal-to-background ratio. To impose orthogonality with the [$\PQb\PQb\ell\ell\nu\nu$]{}decay channel, upper limits on the $\ptmiss$ are imposed as follows: $\ptmiss < 40,\ 75,\ \mathrm{and}\ 100\GeV$ for the $m_X$ of 260–350, 350–650, and ${\geq}$650, respectively. We refer to this selection as the final selection in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel.
After the final selection, twenty-two variables that exploit the differences in kinematic and angular distributions between the signal and background processes are combined into a boosted decision tree (BDT) discriminant [@bdt]. In the $m_X$ range of 260–300, the most important variables are $m_{\ell\ell}$, the separation between the leading lepton and leading tagged jet $\Delta R_{\ell1\PQb1}$, and the invariant mass of the pair of tagged jets $m^{\Ph}_{\PQb\PQb}$. In the $m_X$ range of 350–550, $m^{\Ph}_{\PQb\PQb}$ is the most important variable, while $m_{\ell\ell}$ becomes less important, and the separation between the pair of leptons $\Delta R_{\ell\ell}$ gradually becomes more important when the $m_X$ increases. For the $m_X$ higher than 550, $\Delta R_{\ell\ell}$ becomes the most important variable followed by $m^{\Ph}_{\PQb\PQb}$ and the separation between the pair of tagged jets $\Delta R^{\Ph}_{\PQb\PQb}$. The BDTs are configured to use stochastic gradient boosting with the binomial log-likelihood loss function. The software package TMVA [@TMVA] is used for BDT implementation, training, and application.
The BDT is trained using all background processes described in Section \[samples\], excluding the multijet background. In each lepton channel and for each spin hypothesis, one BDT is trained for each simulated signal $m_X$. In the training, signal events include samples from the two neighboring mass points, in addition to the targeted mass point. In total, 48 BDTs are trained. These BDT distributions for data and expected backgrounds are used as the final discriminating variable in the analysis.
Background estimation in the bblljj channel {#lljjbackground}
-------------------------------------------
The main processes that can mimic the signature of the signal in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel are [$\PZ/\gamma^{*}\text{+jets}$]{}and $\ttbar$, with smaller contributions from QCD multijets, diboson+jets, $\PW$+jets, and SM Higgs boson production.
The contribution from the principal background, [$\PZ/\gamma^{*}\text{+jets}$]{}, is estimated with simulated events normalized to the data at the preselection level in the $\PZ$ boson-enriched control region $80 < m_{\ell\ell} < 100\GeV$. The contribution from $\ttbar$ is estimated in a similar manner, with the $\ttbar$-enriched control region defined by $m_{\ell\ell} > 100\GeV$, and $\ptmiss>100\GeV$. The data-to-simulation normalization factors derived from the two control regions are $R_{\PZ} = 1.14 \pm 0.01\stat$ and $R_{\ttbar} = 0.91 \pm 0.01\stat$ in the muon channel and $R_{\PZ} = 1.24 \pm 0.01\stat$ and $R_{\ttbar} = 0.97 \pm 0.02\stat$ in the electron channel. These normalization factors are found to be consistent between lepton flavors when applying lepton-specific systematic variations.
The contribution from QCD multijet processes is determined from data with a method that exploits the fact that neither signal events nor events from other backgrounds produce final states with same-sign leptons at any significant level. Data events with same-sign isolated leptons are used to model the shape of the multijet background, after all non-QCD sources of background contributing to this selection are subtracted using simulation. The yield in this region is normalized with the ratio of the number of events with nonisolated OS leptons to the number of events with nonisolated same-sign leptons. Here, nonisolated leptons are those muons (electrons) that fail the relative isolation requirements described in Section \[reconstruction\]. All non-QCD sources of background, estimated with simulated events, are subtracted from the numerator and the denominator before computing the ratio.
The contributions from diboson+jets, $\PW$+jets, and SM Higgs boson production are estimated from simulation.
Event selection in the bbllnunu channel {#llnunuselection}
---------------------------------------
Candidate events in the [$\PQb\PQb\ell\ell\nu\nu$]{}channel are reconstructed from the physics objects, as described above. The two leptons (muons or electrons) are required to have OS, and the invariant mass of the two leptons, $m_{\ell\ell}$, is required to exceed 76. One of the Higgs bosons is formed from the pair of jets with the highest output value of the tagging discriminant, and the second Higgs boson is reconstructed as a combination of the two charged leptons and the , representing the visible and invisible decays products, respectively, of the pair of bosons. The requirement on $m_{\ell\ell}$ reduces the contribution from resonant $\mathrm{X}\to\Ph\Ph$ production in the $\PQb\PQb\PW\PW$ final state, and makes this measurement orthogonal to a previous $\PQb\PQb\PW\PW$ search [@bbWW], where only events with $m_{\ell\ell}$ below 76were considered.
For the Higgs boson decaying to a pair of bosons, the two neutrinos are not reconstructed in the detector, and a pseudo invariant mass of the Higgs boson is used to approximate the incomplete momentum four-vector of the $\PH$. The pseudo invariant mass is formed from the momenta of the two charged leptons coming from one of the bosons and the four-vector $(\ptmiss, \ptvecmiss)$ approximating that of the two-neutrino system coming from the other of the bosons, where the $z$ component of is zero. While the true invariant mass of the pair of neutrinos is not zero but is equal to the invariant mass of the parent boson, that boson is off the mass shell and has relatively low mass.
In order to suppress the backgrounds from the [$\PZ/\gamma^{*}\text{+jets}$]{}and QCD multijet processes as well as from the SM Higgs boson production via the $\PZ\Ph$ process, a requirement is imposed on the minimum , which is 40 (75)for the $m_X$ of 260–300 (350–600), and 100for higher $m_X$.
Three regions, a signal region (SR) and two control regions (CRs), are further defined using $m_{\ell\ell}$ and the invariant mass $m^{\Ph}_{\PQb\PQb}$ of the two jets. The SR is defined by the requirements $76 < m_{\ell\ell} <106\GeV$ and $90< m^{\Ph}_{\PQb\PQb} <150\GeV$. A first CR, dominated by events, is defined by $m_{\ell\ell} >106\GeV$ and $90< m^{\Ph}_{\PQb\PQb} <150\GeV$. A second CR, enriched in [$\PZ/\gamma^{*}\text{+jets}$]{}events, is defined by requiring $20< m^{\Ph}_{\PQb\PQb} <90\GeV$ or $m^{\Ph}_{\PQb\PQb} >150\GeV$, and $76 < m_{\ell\ell} <106\GeV$. The two CRs and the SR are used to estimate the backgrounds in the SR via a simultaneous fit.
To further differentiate signal from backgrounds in the SR, a BDT discriminant is trained using all simulated signal and background processes described in Section \[samples\]. Of the nine input distributions to the BDT, the most important variables in the low-mass range are the separation between the pair of $\PQb$ tagged jets $\Delta R^{\Ph}_{\PQb\PQb}$, , and $m^{\Ph}_{\PQb\PQb}$. In the high-mass region, $m^{\Ph}_{\PQb\PQb}$ and $\Delta R^{\Ph}_{\PQb\PQb}$ are also the most significant, together with the separation between the pair of charged leptons $\Delta R_{\ell\ell}$, which becomes more important as the resonance mass increases. Two BDTs are trained for each lepton channel and each resonance spin hypothesis, one for $m_X$ in the range of 250–450, and another one for the $m_X$ above 450. A minimum BDT value is required for candidates in the SR, optimized for each narrow $m_X$ hypothesis to yield the best 95% confidence level () expected upper limit on resonant production. The BDTs are configured with the same classification and loss function parameters described in Section \[lljjselection\].
Finally, a quantity closely correlated with the energy-momentum four-vector of the $\Ph\Ph$ system is constructed as the vector sum of the of the two leptons, two $\PQb$ jets, and the four-vector formed as $(\ptmiss, \ptvecmiss)$ for the neutrinos, as described above. Subsequently, the pseudo transverse mass of the $\Ph\Ph$ system is defined as $\widetilde{M}_{\mathrm{T}}(\Ph\Ph) = \sqrt{\smash[b]{E^2 - p_{z}^2}}$, where $E$ and $p_{z}$ are the energy and the $z$ component of the combined four-vector.
The $\widetilde{M}_{\mathrm{T}}(\Ph\Ph)$ distributions for data and expected backgrounds, in the combined signal and CRs, will be used as the final discriminating variable in the analysis.
After the event selection in this channel is applied, the signal $\Ph\Ph$ events in the SR come predominantly from the decays with the [$\PQb\PQb\PZ\PZ$]{}intermediate state (80%) with a smaller contribution from the $\PQb\PQb\PWp\PWm$ intermediate state (20%).
Background estimation in the bbllnunu channel {#llnunubackground}
---------------------------------------------
The dominant sources of background in the [$\PQb\PQb\ell\ell\nu\nu$]{}channel are and [$\PZ/\gamma^{*}\text{+jets}$]{}production. Several other background processes contribute, including single top quark and diboson production, and SM Higgs boson production in association with a boson. While these are typically minor backgrounds, their contribution can vary over the $m_X$ range. The QCD multijet background is negligible across the full mass range because of the stringent selection on $m_{\ell\ell}$.
The event yields in the signal and two CRs, which are dominated by and [$\PZ/\gamma^{*}\text{+jets}$]{}events, are determined from data. The corresponding normalizations of the simulated $\widetilde{M}_{\mathrm{T}}(\Ph\Ph)$ distributions are free parameters in the simultaneous fit of all three regions. The remaining backgrounds are estimated from simulation and normalized according to their theoretical cross sections.
Systematic uncertainties {#systematics}
========================
The dominant source of systematic uncertainty in this analysis is the jet energy scale (JES) uncertainty, which is of the order of a few percent and is estimated as a function of jet and $\eta$ [@Khachatryan:2016kdb]. The $\eta$-dependent jet energy resolution (JER) correction factors are varied by $\pm$1 standard deviation in order to estimate the effect of the uncertainty. Uncertainties in the JES are propagated to the calculation of $\ptmiss$. A residual $\ptmiss$ uncertainty of 3% is applied in the [$\PQb\PQb\ell\ell\nu\nu$]{}channel to take into account the effect, at low $\ptmiss$, of the unclustered energy from neutral hadrons and photons that do not belong to any jet, and from jets with $\pt <10\GeV$.
An uncertainty of 2% per muon in the muon reconstruction, identification, and isolation requirements, as well as a 1% per muon uncertainty in the muon HLT efficiency are assigned [@MuId]. A per-muon uncertainty due to measured differences of tracking efficiency in data and simulation is estimated to be 0.5% for muon $\pt <300\GeV$ and 1.0% for muon $\pt >300\GeV$ [@wprime]. Per-electron uncertainties in the efficiency for electron trigger, identification and isolation requirements, estimated by varying the scale factors within their uncertainties, are applied. The uncertainties in the efficiency scale factors are generally $<$2% for trigger and $<$3% for identification and isolation [@EMId]. The effect of the variations on the yield of the total background is $<$1%. Uncertainties in the data-to-simulation correction factors of the tagging and of light flavor mis-tagging efficiencies are included.
Normalization and shape uncertainties are assigned to the modeling of the backgrounds. An uncertainty in the shape of the signal and background models is determined by varying the factorization and the renormalization scales between their nominal values and 0.5 to 2.0 times the nominal values in the simulated signal and background samples. The variations where one scale increases and the other decreases are not considered. Each of the remaining variations of the renormalization and the factorization scales are considered, and the maximum variation among all the samples with respect to the nominal sample used in the analysis is taken as the systematic uncertainty, which is found to be 5–7% depending on the process. An uncertainty in the signal acceptance and background acceptance and cross section due to PDF uncertainties and to the value chosen for the strong coupling constant is estimated by varying the NNPDF set of eigenvectors within their uncertainties, following the PDF4LHC prescription [@PDF4LHC]. Statistical uncertainties in the simulated samples for [$\PZ/\gamma^{*}\text{+jets}$]{}and $\ttbar$ background estimates result in uncertainties on the data-derived normalization factors in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel.
An uncertainty of 2.5% is assigned to the determination of the integrated luminosity [@lumi2016]. The uncertainty in the PU condition and modeling is assessed by varying the inelastic $\Pp\Pp$ cross section from its central value by $\pm$4.6% [@pu2016].
All the uncertainties discussed are applied to all background and signal simulated samples. The sensitivity of the presented search is limited by the statistical uncertainties.
Results
=======
Results are obtained by performing a binned maximum likelihood fit of the BDT distributions for the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel, and of the $\Ph\Ph$ pseudo transverse mass simultaneously in the SR and two CRs for the [$\PQb\PQb\ell\ell\nu\nu$]{}channel.
The data and background predictions at final selection level in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}channel are shown in Fig. \[figapp:bdt\], for the distributions of the BDT discriminant for signal masses of 500 and 1000, in the muon and electron final states. Studies performed on all 48 BDT discriminants show stability of the trainings with no evidence of bias or overtraining.
Figure \[MCcomparisons\] shows the $\Ph\Ph$ pseudo transverse mass distributions in the data, background estimates, and spin-2 RS1 graviton for the 300mass hypothesis, after the final selection in the [$\PQb\PQb\ell\ell\nu\nu$]{}channel.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
The systematic uncertainties are represented by nuisance parameters that are varied in the fit according to their probability density functions, prescribed as follows. A log-normal probability density function is assumed for the nuisance parameters affecting the event yields of the various background contributions, whereas systematic uncertainties that affect the distributions are represented by nuisance parameters whose variation is a vertical interpolation in each bin with a sixth-order polynomial for upward and downward shifts of one standard deviation, and linearly outside of that [@ATLAS:2011tau].
The statistical uncertainty from the limited number of events in the simulated samples is taken into account, for each bin of the discriminant distributions, by assigning a nuisance parameter to scale the sum of the process yields in that bin according to the statistical uncertainty using the Barlow–Beeston “lite” prescription [@BARLOW1993219; @Conway:2011in].
In both channels the data distributions are well reproduced by the SM background processes. Upper limits on the resonance production cross section are set, using the asymptotic modified frequentist approach [@Junk:1999kv; @Read_2002; @AsympOpt].
The observed and expected 95% upper limits on $\sigma(\Pp\Pp\to\mathrm{X}\to\PH\PH\to{\ensuremath{\PQb\PQb\PZ\PZ}\xspace})$ in the [$\PQb\PQb\ell\ell\mathrm{jj}$]{}and [$\PQb\PQb\ell\ell\nu\nu$]{}channels as a function of $m_X$ are shown in Fig. \[fig:limit\_plots\_channel\], together with the NLO predictions for the RS1 radion, RS1 KK graviton, and N2HDM resonance production cross sections, where $\PH$ can represent either the SM Higgs boson or an additional Higgs boson from an extended electroweak sector. As two different BDTs are defined for the search in the low- and high-mass ranges of the [$\PQb\PQb\ell\ell\nu\nu$]{}channel, the limit calculation is performed with both of the BDTs at the boundary of the two ranges, around 450, where a discontinuity is seen.
Combined 95% upper limits from both channels on $\sigma(\Pp\Pp\to\mathrm{X}\to\PH\PH\to{\ensuremath{\PQb\PQb\PZ\PZ}\xspace}$) as a function of $m_X$, are shown in Fig. \[fig:limit\_plots\_combined\], together with the theoretical predictions for the RS1 radion and RS1 KK graviton. In the $m_X$ range between 260 and 1000, limits on the production cross section times branching fraction of RS1 radion and RS1 KK graviton range from 0.1 to 5.0 and 0.1 to 3.6, respectively. In the spin-0 case, the predictions of the N2HDM model with $\tan\beta=0.5\mathrm{\ and\ }2.0$ are shown, for all ${\PH}_3\to{\PH}_1{\PH}_1/{\PH}_1{\PH}_2/{\PH}_2{\PH}_2\to{\ensuremath{\PQb\PQb\PZ\PZ}\xspace}$ decays. In the $\tan\beta=0.5$ case, the model can be excluded with ${\PH}_3$ in the $m_X$ range of 360–620. In comparison to previous searches, we achieve a sensitivity to the RS1 radion and RS1 KK graviton models comparable to the difference in $\Ph\Ph$ decay branching ratios.
Finally, the results are also interpreted as a function of both the $m_X$ and $\lambda_{\mathrm{R}}$ ($\tilde{k}$) for the radion (graviton) case, with $\lambda_{\mathrm{R}}<0.3\TeV$ ($\tilde{k}>0.6$) excluded for all of the $m_X$ considered, as shown in Fig. \[fig:limit\_plots\_combined\_2D\].
Summary
=======
A search for the production of a narrow-width resonance decaying into a pair of Higgs bosons decaying into the [$\PQb\PQb\PZ\PZ$]{}channel is presented. The analysis is based on data collected with the CMS detector during 2016, in proton-proton collisions at the LHC, corresponding to an integrated luminosity of 35.9. The final states considered are the ones where one of the bosons decays into a pair of muons or electrons, and the other boson decays either to a pair of quarks or a pair of neutrinos. Upper limits at 95% confidence level are placed on the production of narrow-width spin-0 or spin-2 particles decaying to a pair of Higgs bosons, in models with and without an extended Higgs sector. For a resonance mass range between 260 and 1000, limits on the production cross section times branching fraction of a spin-0 and spin-2 resonance range from 0.1 to 5.0 and 0.1 to 3.6, respectively. These results set limits in parameter space in bulk Randall–Sundrum radion, Kaluza–Klein excitation of the graviton, and N2HDM models. For specific choices of parameters the N2HDM can be excluded in a mass range between 360 and 620for a resonance decaying to two Higgs bosons. This is the first search for Higgs boson resonant pair production in the [$\PQb\PQb\PZ\PZ$]{}channel.
We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centers and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: BMBWF and FWF (Austria); FNRS and FWO (Belgium); CNPq, CAPES, FAPERJ, FAPERGS, and FAPESP (Brazil); MES (Bulgaria); CERN; CAS, MoST, and NSFC (China); COLCIENCIAS (Colombia); MSES and CSF (Croatia); RPF (Cyprus); SENESCYT (Ecuador); MoER, ERC IUT, PUT and ERDF (Estonia); Academy of Finland, MEC, and HIP (Finland); CEA and CNRS/IN2P3 (France); BMBF, DFG, and HGF (Germany); GSRT (Greece); NKFIA (Hungary); DAE and DST (India); IPM (Iran); SFI (Ireland); INFN (Italy); MSIP and NRF (Republic of Korea); MES (Latvia); LAS (Lithuania); MOE and UM (Malaysia); BUAP, CINVESTAV, CONACYT, LNS, SEP, and UASLP-FAI (Mexico); MOS (Montenegro); MBIE (New Zealand); PAEC (Pakistan); MSHE and NSC (Poland); FCT (Portugal); JINR (Dubna); MON, RosAtom, RAS, RFBR, and NRC KI (Russia); MESTD (Serbia); SEIDI, CPAN, PCTI, and FEDER (Spain); MOSTR (Sri Lanka); Swiss Funding Agencies (Switzerland); MST (Taipei); ThEPCenter, IPST, STAR, and NSTDA (Thailand); TUBITAK and TAEK (Turkey); NASU (Ukraine); STFC (United Kingdom); DOE and NSF (USA).
Individuals have received support from the Marie-Curie program and the European Research Council and Horizon 2020 Grant, contract Nos. 675440, 752730, and 765710 (European Union); the Leventis Foundation; the A.P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation à la Recherche dans l’Industrie et dans l’Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the F.R.S.-FNRS and FWO (Belgium) under the “Excellence of Science – EOS" – be.h project n. 30820817; the Beijing Municipal Science & Technology Commission, No. Z191100007219010; the Ministry of Education, Youth and Sports (MEYS) of the Czech Republic; the Deutsche Forschungsgemeinschaft (DFG) under Germany’s Excellence Strategy – EXC 2121 “Quantum Universe" – 390833306; the Lendület (“Momentum") Program and the János Bolyai Research Scholarship of the Hungarian Academy of Sciences, the New National Excellence Program ÚNKP, the NKFIA research grants 123842, 123959, 124845, 124850, 125105, 128713, 128786, and 129058 (Hungary); the Council of Science and Industrial Research, India; the HOMING PLUS program of the Foundation for Polish Science, cofinanced from European Union, Regional Development Fund, the Mobility Plus program of the Ministry of Science and Higher Education, the National Science Center (Poland), contracts Harmonia 2014/14/M/ST2/00428, Opus 2014/13/B/ST2/02543, 2014/15/B/ST2/03998, and 2015/19/B/ST2/02861, Sonata-bis 2012/07/E/ST2/01406; the National Priorities Research Program by Qatar National Research Fund; the Ministry of Science and Higher Education, project no. 02.a03.21.0005 (Russia); the Tomsk Polytechnic University Competitiveness Enhancement Program and “Nauka" Project FSWW-2020-0008 (Russia); the Programa Estatal de Fomento de la Investigaci[ó]{}n Cient[í]{}fica y T[é]{}cnica de Excelencia María de Maeztu, grant MDM-2015-0509 and the Programa Severo Ochoa del Principado de Asturias; the Thalis and Aristeia programs cofinanced by EU-ESF and the Greek NSRF; the Rachadapisek Sompot Fund for Postdoctoral Fellowship, Chulalongkorn University and the Chulalongkorn Academic into Its 2nd Century Project Advancement Project (Thailand); the Kavli Foundation; the Nvidia Corporation; the SuperMicro Corporation; the Welch Foundation, contract C-1845; and the Weston Havens Foundation (USA).
The CMS Collaboration \[app:collab\]
====================================
=5000=500=5000
|
{
"pile_set_name": "ArXiv"
}
|
---
title: Growth for the central polynomials
---
Amitai Regev
Dept. Math and CS
The Weizmann Institute,
Rehovot 76100, Israel
email: amitai.regev at weizmann.ac.il
remark6.tex
: We study the growth of the central polynomials for the algebras $G$ and $M_k(F)$, the infinite dimensional Grassmann algebra and the $k\times k$ matrices over a field $F$ of characteristic zero. In particular it follows that $M_k(F)$ satisfy many proper central polynomials.
[**Key words:**]{} PI algebras, central polynomials, Young tableaux.
Introduction and main result
============================
Let $F$ be a field of characteristic zero, $A$ an $F$ algebra, $Id(A)\subseteq F\langle x \rangle$ the polynomial identities of $A$, and let $Id^z(A)\subseteq F\langle x \rangle$ denote the polynomials which are central on $A$: $g(x_1,\ldots,x_n)\in Id^z(A)$ if for any $a_1,\ldots ,a_n\in A,$ $g(a_1,\ldots , a_n)\in center(A)$. Such $g$ is [*proper central*]{} if $g$ is central and non-identity of $A$. For example let $A=M_2(F)$, then $[x,y]^2\not \in Id(A)$ but $[x,y]^2 \in Id^z(A)$. Note that polynomial identities are considered here as central polynomials, and in particular $Id(A)\subseteq Id^z(A)$. Clearly, if $g\in Id^z(A)$ then $g+Id(A)\subseteq Id^z(A)$. Thus the proper central polynomials of $A$ correspond to the space $Id^z(A)/Id(A)$.
The existence of proper central polynomials for matrix algebras was an important open problem in PI theory. It was solved independently by Formanek [@formanek.central] and by Razmyslov [@raz], who constructed proper central polynomials for any matrix algebra $M_k(F)$. However, only very little is known about the question “how many proper central polynomials there are?” We show here that this problem is related to the growth of the central cocharacters and codimensions of the given algebra.
In this paper we study that question in the cases $A=G$ and $A=M_k(F)$; here $G$ is the infinite dimensional Grassmann algebra. To study that “central growth” for a given algebra $A$ we intersect these spaces with the multilinear polynomials $V_n=V_n(x_1,\ldots,x_n)$, then the proper multilinear central polynomials correspond to the quotient space $D_n(A)=(V_n\cap Id^z(A))/(V_n\cap Id(A))$, with $\dim D_n(A)=\delta_n(A)$. The sequence $\delta_n(A)$ determines the growth of the proper central polynomials. The $S_n$ character of $D_n(A)$ is the difference of cocharacters (with corresponding codimensions) $$\chi_{S_n}(D_n(A))=\chi_n(A)-\chi_n^z(A)\qquad\mbox{hence}\qquad \delta_n(A)=c_n(A)-c_n^z(A).$$ In the first part of the paper we analyze the case $A=G$ and determine, [*precisely*]{}, these cocharacters, thus we determine precisely $\chi_{S_n}(D_n(A))$. In the second part we study the case $A=M_k(F)$ and prove that $$\lim_{n\to\infty} (\delta_n(M_k(F)))^{1/n}=k^2.$$ This indicate that $M_k(F)$ satisfies a large (exponential) amount of proper central polynomials.
Some generalities
=================
Ordinary representations of $S_n$
---------------------------------
The representation theory of $S_n$ is a basic tool in what follows. Here is a brief review of that theory. The irreducible characters of $S_n$ are indexed by the partitions $\lambda\vdash n$, denoted $\chi^\lambda$, with $f^\lambda=\deg \chi^\lambda$ the corresponding degree. To $\lambda\vdash n$ corresponds the (unique) Young diagram $D_\lambda$ (of shape $\lambda$) and (many) Young tableaux $T_\lambda$, again of shape $\lambda$. To such tableaux $T_\lambda$ corresponds the semi-idempotent $e_{T_\lambda}$ in the group algebra $FS_n$; the left ideal $FS_ne_{T_\lambda}$ is an irreducible $S_n$-representation with corresponding $S_n$ character $\chi^\lambda$: $\chi_{S_n}(FS_ne_{T_\lambda})=\chi^\lambda$. For a detailed account of that theory see for example [@sagan].
Proper central polynomials
--------------------------
Recall that the elements of $Id^z(A)/ Id(A)$ correspond to the [*proper central polynomials of $A$*]{}. We would like to estimate the amount of the proper central polynomials of $A$, and we do that by restricting to multilinear polynomials (see Definition \[new1\] below), since as in the case of identities, the multilinear central polynomials generate all such polynomials.
It is well known that $Id(A)$ is a $T$-ideal in $F\langle x \rangle$, namely it is closed under substitutions and under ideal operations in $F\langle x \rangle$. On the other hand $Id^z(A)$ is only a $T$-subalgebra of $F\langle x \rangle$, it is closed under substitutions and under algebra – but not ideal – operations in $F\langle x \rangle$. Since $Id^z(A)$ and $Id(A)$ are closed under substitutions, the following definition makes sense as it is independent of the particular variables $x_1,\ldots,x_n$.
\[new1\]
Let $V_n=V_n(x_1,\ldots,x_n)$ be the multilinear polynomials of degree $n$ in $x_1,\ldots,x_n$. To study $Id^z(A)/ Id(A)$ we intersect with $V_n$ since, as in the case of identities, the central polynomials are generated be the multilinear central polynomials. We define
1. $$D_n(A):=\frac{V_n\cap Id^z(A)}{ V_n\cap Id(A)}\quad\mbox{and}\quad \delta_n(A):=\dim(D_n(A))=
\dim \left(\frac{(V_n\cap Id^z(A))}{ (V_n\cap Id(A))}\right)$$ Similarly
2. $$c_n(A):=\dim \left(\frac{V_n}{V_n\cap Id(A)}\right) \qquad \chi_n(A):= \chi_{S_n}\left(\frac{V_n}{V_n\cap Id(A)}\right),$$ $$c^z_n(A):=\dim \left(\frac{V_n}{V_n\cap Id^z(A)}\right) \qquad\mbox{and}\qquad \chi_n^z(A):= \chi_{S_n}\left(\frac{V_n}{V_n\cap Id^z(A)}\right),$$
the ordinary and the central cocharacters with their corresponding codimensions.
\[new2\] $D_n(A)$ is a left $S_n$ module, its $S_n$ character is $$\begin{aligned}
\label{difference2}
\chi_{S_n}(D_n(A))=\chi_n(A)-\chi^z_n(A)\end{aligned}$$ and by taking degrees we get $$\begin{aligned}
\label{difference3}
\delta_n(A)=c_n(A)-c^z_n(A).\end{aligned}$$
Note that $$D_n(A)=\frac{V_n\cap Id^z(A)}{ V_n\cap Id(A)}\cong \frac{V_n/(V_n\cap Id(A))}{V_n/(V_n\cap Id^z(A))},$$ isomorphism of quotient $S_n$ modules. The proof follows by computing the corresponding $S_n$ characters – and their degrees.
The case $A=G$
==============
We now calculate the above invariants $c_n^z(G),$ $\chi_n^z(G)$ and $\delta_n(G)$ for $A=G$, the infinite dimensional Grassmann algebra.
The results for $G$
-------------------
Denote $\chi_n^z(G)=\sum_{\lambda\vdash n}a_\lambda\chi^\lambda\quad\mbox{and}\quad
\chi_n(G)=\sum_{\lambda\vdash n}b_\lambda\chi^\lambda.$ Since $\chi_n^z(G)\le \chi_n(G)$, all $a_\lambda\le b_\lambda.$ By [@olsson] $ b_\lambda=1$ if $\lambda$ is a $(1,1)$ hook partition $\lambda=(n-j,1^j),$ and $b_\lambda=0$ otherwise. It follows that $a_\lambda\in\{0,1\}$ if $\lambda$ is a $(1,1)$ hook partition $\lambda=(n-j,1^j)$, and $a_\lambda=0$ otherwise. We prove here the following theorem.
\[main\] Let $\chi_n^z(G)=\sum_{\lambda\vdash n}a_\lambda\chi^\lambda$, then $ a_{(n-2j,1^{2j})}=1$ for $0\le j\le\lfloor n/2\rfloor$, and $a_\lambda=0$ for all other partitions $\lambda\vdash n$.
See also Theorem \[1.9\] below.
It is known [@krakowsky] that $Id(G)=T([[x,y],u])$, the $T$-ideal of the triple commutator. An obvious central polynomial for $G$ is the commutator $[x,y]\in Id^z(G)$. Theorem \[corollary2\] below is of interest on its own; it says that the algebra of central polynomials of $G$ is $T$ generated by the commutator $[x,y]$ and by the triple commmutator $[[x,y],u]$.
We remind here that the central polynomials form a $T$-subalgebra, but not an ideal, hence they behave rather differently from identities. For example, modulo $Id^z(G)$ the monomials $x_1x_2$ and $x_2x_1$ are linearly dependent (since $x_1x_2-x_2x_1\in Id^z(G)$) but $x_3x_1x_2$ and $x_3x_2x_1$ are linearly independent modulo $Id^z(G)$, see Lemma \[cancellation2\] below.
Some preparations
-----------------
Our aim here is to calculate the $S_n$ character $\chi^z_n(G)$, the $n$-th central cocharacter of $G$. We first calculate the restriction $\chi^z_n(G)\downarrow_{S_{n-1}}$, proving it in fact is the ordinary $n-1$ cocharacter of $G$, see Theorem \[corollary1\]. A main tool here is the following isomorphism.
There is a canonical isomorphism of $FS_{n-1}$ modules $$\frac{V_n}{V_n\cap Id^z(G)}\cong \frac{V_{n-1}}{V_{n-1}\cap Id(G)}.$$ It follows that the cocharacters satisfy $$\chi_n^z(G)\downarrow_{S_{n-1}}=\chi_{n-1}(G),$$ [and in particular the codimensions satisfy $c_n^z(G)=c_{n-1}(G).$]{} Since $c_n(G)=2^{n-1}$ [@krakowsky], hence $$c_n^z(G)=2^{n-2}=\frac{1}{2}c_n(G).$$
The proof is given in Section \[isomorphism\] below.
\[cancellation\] Given $h=h_n\in V_n$, there exist $h_{n-1}\in V_{n-1}$ (given canonically, so the map $h_n\to h_{n-1}$ is well defined) such that $h_n=x_nh_{n-1}+q_n$ where $q_n\in Id^z(G)\cap V_n$.
Suffices to prove when $h\in V_n$ is a monomial $h=M$. Then $M=ax_nb$, so $M=x_nba+[a,x_nb]$, and $q_n=[a,x_nb]\in Id^z(G)\cap V_n. $
[**Note**]{} that this presentation is not unique, since if $h'_{n-1}\in V_{n-1}\cap Id(G)$, we have $h_n=x_n(h_{n-1}+h'_{n-1})+(q_n-x_nh'_{n-1}),$ and $q_n-x_nh'_{n-1}\in Id^z(G)\cap V_n.$
Let $B=\{e_{i_1}\cdots e_{i_r}\mid r=1,2,\ldots,~~i_1<\cdots< i_r \}$ the canonical linear basis of $G$. We consider [*disjoint*]{} substitutions, namely $x_j\to \bar x_j\in B$, $j=1,\ldots, n$ such that $\bar x_1\cdots \bar x_n\ne 0$. Given $h_n\in V_n$, such a substitution induces $h_n=h_n(x_1,\ldots,x_n)\to \bar h_n=h_n(\bar x_1,\ldots ,\bar x_n)$. If $e_{i_1}\cdots e_{i_r}\ne 0$, denote $\ell(e_{i_1}\cdots e_{i_r})=r$ (length). Such a substitution is even (odd) if $\ell(\bar x_1\cdots\bar x_n)=\sum _{i=1}^n\ell(\bar x_i)$ is even (odd).
[**Note**]{} that such a disjoint substitution is even if and only if $\bar x_1\cdots\bar x_n$ is central in $G$.
\[cancellation2\] Let $h_{n-1}\in V_{n-1}$ then $x_nh_{n-1}\in Id^z(G)$ if and only if $h_{n-1}\in Id(G)$.
1\. Assume $h_{n-1}\in Id(G)$ then $x_nh_{n-1}\in Id(G)\subseteq Id^z(G).$ Conversely,
2\. Let $x_nh_{n-1}\in Id^z(G).$ Make a disjoint substitution $x_j\to \bar x_j$, $j=1,\ldots,n$, then $\bar x_n\bar h_{n-1}$ is central, hence even. Now slightly change that substitution: leave $x_j\to \bar x_j$, $j=1,\ldots,n-1$ unchanged, and let $x_n\to \hat x_n $ with the parity of $\hat x_n $ opposite that of $\bar x_n .$ Since $x_nh_{n-1}\in Id^z(G),$ the result is still even, hence we must have $\bar h_{n-1}=0$, namely $h_{n-1}\in Id(G)$.
### An isomorphism {#isomorphism}
Let $$\varphi:V_{n-1}\longrightarrow \frac{V_n}{V_n\cap Id^z(G)}$$ be given as follows: Given $h_{n-1}\in V_{n-1}$, then $\varphi (h_{n-1})=x_nh_{n-1} + (V_n\cap Id^z(G)).$ Then
1\. $\varphi$ is onto ${V_n}/({V_n\cap Id^z(G)})$, and
2\. $\ker\varphi = V_{n-1}\cap Id(G)$.
1\. Let $y_n\in V_n$ then by Lemma \[cancellation\] we can write $y_n=x_ny_{n-1}+q_n$, $q_n\in Id^z(G)\cap V_n.$ Thus $\varphi( y_{n-1})= x_ny_{n-1}+(V_n\cap Id^z(G))= x_ny_{n-1}+q_n+(V_n\cap Id^z(G))=y_n+(V_n\cap Id^z(G)).$
2\. Let $h_{n-1}\in V_{n-1}$ then $h_{n-1}\in ker \varphi$ if and only if
$x_n h_{n-1}\in V_n\cap Id^z(G)$, so if and only if
$x_n h_{n-1}\in Id^z(G),$ by Lemma \[cancellation2\] if and only if
$ h_{n-1}\in Id(G)$ if and only if
$ h_{n-1}\in V_{n-1}\cap Id(G).$
As a corollary we have
\[corollary1\] $\varphi$ induces the isomorphism $$\bar\varphi: \frac{V_{n-1}}{V_{n-1}\cap Id(G)}\cong\frac{V_{n}}{V_{n}\cap Id^z(G)}$$ an isomorphism of $FS_{n-1}$ modules. Therefore $$\begin{aligned}
\label{ab2}
\chi_{n-1}(G)=\chi_n^z(G)\downarrow_{S_{n-1}}.\end{aligned}$$ Recall [@krakowsky] that $c_n(G)=2^{n-1}$. Thus in particular, the codimensions satisfy $$c_n^z(G)=c_{n-1}(G)=2^{n-2}=\frac{1}{2}2^{n-1}=\frac{1}{2}c_n(G).$$
### The central cocharacter $\chi^z_n(G)$
We can now prove our main result of this section.
\[1.9\] $$\chi^z_n(G)=\sum_{j=0}^{\lfloor n/2\rfloor} \chi^{(n-2j,1^{2j})}
=\chi^{(n)}+\chi^{(n-2,1^2)}+\chi^{(n-4,1^4)}+\cdots$$ then by Equation it follows that $$\chi_{S_n}(D_n(G))=
%\chi^z_n(G)=
\sum_{j=0}^{\lfloor n/2\rfloor} \chi^{(n-2j+1,1^{2j-1})}
=\chi^{(n-1,1)}+\chi^{(n-3,1^3)}+\chi^{(n-5,1^5)}+\cdots.$$
We know [@olsson] that $$\begin{aligned}
\label{equation}\chi_n(G)=\sum_{k}\chi^{(n-k,1^k)}.\end{aligned}$$ and $$\begin{aligned}
\label{end}
\chi^z_n(G)=\sum_{k\ge 0} a_k\cdot \chi^{(n-k,1^k)}\end{aligned}$$ where each multiplicity $a_k$ is either $=0$ or $=1$. The trivial character $\chi^{(n)}$ corresponds to the semi-idempotent $e_{(n)}=\sum_{\sigma\in S_n}\sigma$ and clearly $e_{(n)}$ is not a central identity of $G$, hence $a_0=1$. By Equations and , $$\chi^z_n(G)\downarrow_{S_{n-1}}=\chi_{n-1}(G)=\sum_{j\ge 0}\chi^{(n-1-j,1^j)}.$$ Restrict now the $S_n$ character $\chi^z_n(G)$ down to $S_{n-1}$. By “branching”, $\chi^{(n)}\downarrow _{S_{n-1}}=\chi^{(n-1)}$, while for $k>0$ $$\chi^{(n-k,1^k)}\downarrow _{S_{n-1}}=\chi^{(n-k-1,1^k)}+\chi^{(n-k,1^{k-1})}.$$ Together with we obtain the following trivial system of equations for the multiplicities $a_k$: $$a_0=1, ~~a_0+a_1=1, ~~a_1+a_2=1, ~~a_2+a_3=1 \ldots$$ and the proof follows.
### $T$ generation of $Id^z(G)$
We assume now that the characteristic of the base field is zero.
\[corollary2\] The algebra $Id^z(G)$ of the central polynomials of $G$ contains $Id(G)$, and is $T$-generated as an $F$-algebra over $Id(G)$ by the commutator $[x,y]$. In other words, the algebra $Id^z(G)$ of the central polynomials of $G$ is $T$-generated by $[[x,y],u]$ and by $[x,y]$.
Let $H\subseteq F\langle x \rangle$ be the subalgebra containing $Id(G)$, and $H$ is $T$-generated, as an $F$-algebra over $Id(G)$, by the commutator $[x,y]$. Thus $H\subseteq Id^z(G)$. Now the previous arguments all work with $H$ replacing $Id^z(G)$. In particular $$\frac{V_{n-1}}{V_{n-1}\cap Id(G)}\cong\frac{V_{n}}{V_{n}\cap H}$$ for all $n$. This implies that for all $n$, $\dim (V_n\cap H)=\dim (V_n\cap Id^z(G))$ and since $H\subseteq Id^z(G)$ it follows that for all $n$, $V_n\cap H=V_n\cap Id^z(G)$. Since we assume that $char(F)=0$, this implies that $H=Id^z(G)$.
In fact, the same argument shows that in any characteristic, the multilinear central polynomials of $G$ are $T$-generated by $[[x,y],u]$ and by $[x,y]$.
Motivated by Theorem \[corollary2\] we make the following conjecture.
In characteristic zero the central polynomials of a PI algebra $A$ satisfy the Specht property – in the following sense: the sub algebra $Id^z(A)\subseteq F\langle X\rangle$ of the central polynomials of $A$ is $T$-generated [*over*]{} $Id(A)$ by a finite set of central polynomials.
The case $A=M_k(F)$
===================
Further introductory remarks
----------------------------
We are interested in the question how many proper central polynomials the algebra $M_k(F)$ satisfies, namely, how large is $\delta_n(M_k(F))$, and what is the asymptotics of $\delta_n(M_k(F))$ as $n$ goes to infinity. The case $k=1$ being trivial, we assume w.l.o.g here that $k\ge 2$.
Lots of work was done on the invariants $c_n(M_k(F))$ and $\chi_n(M_k(F))$. In particular it was proved that $$\begin{aligned}
\label{bliz}
\lim_{n\to\infty} c_n(M_k(F))^{1/n} = k^2,\end{aligned}$$ see [@regev3], [@G.Z Theorem 5.10.2].
Turn now to central polynomials. By [@formanek.central] [@raz] (see also [@formanek.regev]) $M_k(F)$ satisfies proper central polynomials, hence $c_n^z(M_k(F))$ is strictly smaller than $ c_n(M_k(F))$. Recall that $$D_n=\frac{V_n\cap Id^z(M_k(F)))}{ V_n\cap Id(M_k(F)))}\quad\mbox{and}\quad
\delta_n(M_k(F))=\dim (V_n\cap Id^z(M_k(F))) -\dim (V_n\cap Id(M_k(F))).$$ The next theorem is the main result of this section. It indicates that $M_k(F)$ satisfies many proper central polynomials.
\[main2\] Let $k\ge 2$ then (compare with ) $$\lim_{n\to\infty}(\delta_n(M_k(F)))^{1/n}=k^2.$$
Towards the proof of Theorem \[main2\]
--------------------------------------
### A sandwich for $\delta_n(A)$
Lemma \[new2\] already implies the upper bound $$\begin{aligned}
\label{upper.bound}
\delta_n(A)\le c_n(A).\end{aligned}$$ For the lower bound we need Proposition \[shem.1\] below, which we now prove.
\[sham\] Let $H_n\subseteq V_n$ be a subspace satisfying $H_n\cap Id^z(A)=0$. Then $H_n$ imbeds in $D_n(A)$, and in particular $\dim H_n\le dim D_n(A)=\delta_n(A)$.
Map $\varphi:H_n\to D_n$ as follows:
$$\varphi:V_n\longrightarrow ~\frac{V_n}{V_n\cap Id(A)}~\longrightarrow
\frac{V_n/V_n\cap Id(A)}{V_n/V_n\cap Id^z(A)}\cong D_n(A)$$ via $x\in V_n$, $~~\varphi:x\longrightarrow x+{V_n\cap Id(A)} \longrightarrow x+{V_n\cap Id^z(A)}=\varphi(x),$ well defined maps. If $x\in H_n$ and $\varphi(x)=0$ then $x\in {V_n\cap Id^z(A)}$ so $x=0$ since then $x\in H_n\cap Id^z(A)=0$.
Recall the identification $V_n=FS_n$.
\[18\] Let $g\in V_n,$ $g\not\in Id^z(A)$ and assume $FS_ng$ is an irreducible left $S_n$ module. Then $FS_ng\cap Id^z(A)=0$.
Let $h\in FS_ng\cap Id^z(A)$. If $h\not = 0$ then $FS_nh=FS_ng$ (by irreducibility). Since by assumption $h\in Id^z(A)$ we get $g\in FS_nh\subseteq Id^z(A)$ so $g\in Id^z(A)$, contradiction.
The following proposition is one of the main tools for estimating $\delta_n(A)$. Of course, for the upper bound we already have . and as usual, the difficulties are with the lower bound.
\[shem.1\] Let $\lambda\vdash n$ with a corresponding tableau $T_\lambda$ and semi-idempotent $e_{T_\lambda}$, and assume that $e_{T_\lambda}\not\in Id^z(A)$, then $f^\lambda\le \delta_n(A)$.
Let $FS_ne_{T_\lambda}=H_n,$ then $H_n\cap Id^z(A)=0$, so by Lemma \[sham\] $f^\lambda =\dim H_n\le dim D_n(A)$.
Recall from Lemma \[new2\] that $\delta_n(M_k(F))=c_n(M_k(F))-c_n^z(M_k(F))$ and in particular\
$\delta_n(M_k(F))\le c_n(M_k(F)).$ Together with Equation this proves
$$\lim_{n\to\infty} \delta_n(M_k(F))^{1/n} \le\lim_{n\to\infty} c_n(M_k(F))^{1/n} = k^2.$$
A lower bound is given in the following Lemma.
\[lower\] We also have $$k^2\le \lim_{n\to\infty} \delta_n(M_k(F))^{1/n}.$$
[**Idea of the proof:**]{} For each $n$ we construct a partition $\lambda\vdash n$ with a tableau $T_\lambda$ such that $e_{T_\lambda}$ is non central; $\lambda$ is constructed such that $k^2\le \lim_{n\to\infty}(f^\lambda)^{1/n}$. Let $H_n=FS_ne_{T_\lambda}$, then $H_n\cap Id^z(M_k(F))=0$. By Lemma \[sham\] $f^\lambda \le \delta_n(M_k(F))$, and the proof is complete. That proof takes the rest of this paper.
### Gluing tableaux
The proof of Lemma \[lower\] applies several ingredients which we briefly review. We start with “gluing together” Young tableaux, see [@regev1 Theorem 1.6]. Let $\lambda=(a_1,\ldots, a_r)$ be a partition, with $a_1\ge\cdots\ge a_r \ge 1$, so $\ell(\lambda)=r$. Recall that $\lambda$ is identified with its Young diagram $D_\lambda$. Let $h_j(\lambda)=\lambda'_j$ denote the length of the $j$-th column of $\lambda$, so $h_1(\lambda)=\ell(\lambda)$, and $h_{a_1}(\lambda)$ is the length of the last (rightmost) column of $\lambda$. Let $\mu=(b_1,\ldots,b_s)$ be a second partition, then denote $\lambda*\mu:=(\lambda_1+\mu_1,\lambda_2+\mu_2,\ldots )\vdash |\lambda|+|\mu|$. If $h_{a_1}(\lambda)\ge h_1(\mu)$ then $D_\mu$ can be glued to the right of $D_\lambda$ and in that case we denote $D_{\lambda*\mu}=D_\lambda | D_\mu$.
A tableau $T_\lambda$ of shape $\lambda$ and with entries $t_{i,j}$ is denoted $T_\lambda=D_\lambda(t_{i,j})$. Also let $T_\mu=D_\mu(u_{i,j})$ be a second tableau of shape $\mu$, and assume $D_{\lambda*\mu}=D_\lambda | D_\mu$. Then glue $T_\lambda$ with $T_{\mu}$ by constructing $T_{\lambda*\mu}=D_\lambda(t_{i,j})\mid D_\mu(u_{i,j}+|\lambda|)$.
As usual the tableau $T_\lambda$ corresponds to the semi-idempotent $e_{T_\lambda}$ in $FS_{|\lambda|}$. We have
\[1.6\] [@regev1 Theorem 1.6] Let $\lambda=(a_1,\ldots,a_r)\vdash m$, $\mu\vdash n$ with corresponding tableaux $T_\lambda$ and $ T_\mu$ and assume $h_{a_1}(\lambda)\ge h_1(\mu)$, namely $D_\lambda$ and $ D_\mu$ can be glued to form $D_{\lambda*\mu}=D_\lambda\mid D_\mu$. Let $T_{\lambda*\mu}=D_\lambda(t_{i,j})\mid D_\mu(u_{i,j}+m)=
%=T_{\lambda*\mu}
D_{\lambda\mid \mu}(w_{i,j})$ and substitute $w_{i,j}\to y_i$. Then there exists $d\in\mathbb N$ such that $$e_{T_{\lambda*\mu}}(y)=d\cdot e_{T_\lambda}(y)\cdot e_{T_\mu}(y).$$
Capelli-type polynomials {#regev}
------------------------
Another main ingredient in the proof of Lemma \[lower\] are the Capelli-type polynomials, called sometime Regev-polynomials [@formanek.regev], [@GV]. These polynomials correspond to rectangles of height $k^2$. The case of the $k^2\times 2$ rectangle (namely the partition $(2^{(k^2)})$) is done in details in [@formanek.regev].
### Two sets of variables {#two}
In the case of two sets of variables the polynomial is constructed as follows: first construct the monomial $$\begin{aligned}
\label{monomial2}
M(x,y)=(x)(y)(xxx)(yyy)(xxxxx)(yyyyy)\cdots\end{aligned}$$ of pairs of blocks of the odd lengths $1,3,5,\ldots ,2k-1$, then alternate the $x'$s and independently alternate the $y'$s. Thus the corresponding polynomial is $$L(x,y)=\sum_{\sigma\in S_{k^2}}\sum_{\tau\in S_{k^2}}sgn (\sigma)sgn(\tau)M(x_{\sigma},y_{\tau}),$$ in two sets of $k^2$ alternating variables. It corresponds to a particular tableau on the $k^2\times 2$ rectangle, namely the partition $(2^{(k^2)})$. By [@formanek.regev] $L(x,y)$ is a proper central polynomial of $M_k(F)$. It also satisfies
[**Property L**]{}: For any bijection $x_u\leftrightarrow \bar x_u =e_{i,j}$ between the set $\{x_1,\ldots,x_{k^2}\}$ and the set $\{e_{i,j}\mid 1\le i,j\le k\},$ and similarly $y_v\leftrightarrow \bar y_v= e_{i,j}$ we get, up to a $\pm$ sign, the same [*non zero*]{} scalar value $L(\bar x, \bar y)$.
Equating $x_i=y_i, ~i=1,\ldots,k^2$ we obtain $g_2(x)=L(x,x)=g_2(x_1,\ldots,x_{k^2})$, and [*Property L*]{} implies [*Property G:*]{}
For any bijection $x_u\leftrightarrow \bar x_u =e_{i,j}$ we get, up to a $\pm$ sign, the same non zero scalar value $g_2(\bar x)$.
that $g_2(x)$ is of degree 2 in each $x_i$. Its multilinearization yields $d\cdot e_{T_\nu}\in S_{2k^2}$, where $d\ne 0$, $\nu = (2^{(k^2)})$ and $T_\nu$ is is the tableau of shape $\nu$ corresponding to the monomial . $g_2(x)$ is proper central, therefore so is $e_{T_\nu}$.
### Three sets of variables
Instead of two sets of variables we can repeat that construction with, say, three sets of $k^2$ alternating variables, starting with the monomial $$\begin{aligned}
\label{monomial3}
M(x,y,z)=(x)(y)(z)(xxx)(yyy)(zzz)(xxxxx)(yyyyy)(zzzzz)\cdots\end{aligned}$$ It corresponds to a particular tableau on the $k^2\times 3$ rectangle, namely on the partition $(3^{(k^2)})$. Similar to Section \[two\], the resulting polynomial $L(x,y,z)$, in three sets of $k^2$ alternating variables is proper central for $M_k(F)$. Equating $x_i=y_i=z_i, ~i=1,\ldots,k^2$ we obtain $g_3(x)=L(x,x,x)=g_3(x_1,\ldots,x_{k^2})$, and it satisfies the analogue properties L and G in three sets of $k^2$ alternating variables.
The proof of Lemma \[lower\] and Theorem \[main2\]
---------------------------------------------------
We need to show that $$k^2\le \lim_{n\to\infty} \delta_n(M_k(F))^{1/n}.$$ Let $n\ge 2k^2$ (this is not a restriction since we later send $n$ to infinity) and write it as $n=k^2m+r$ where $0\le r<k^2$, so $m\ge 2$.
Case 1: $m$ is odd. In that case $m\ge 3$ so $m=2q+3$ where $q\ge 0$.
Case 2: $m=2q$ is even, $q\ge 1$.
The summand $k^2m$ of $n$ determines the $k^2\times m$ rectangle. In case 1, $k^2m=k^22q+k^23$ so the $k^2\times m$ rectangle is obtained by gluing $q$ $~k^2\times 2 $ rectangles, plus the single $k^2\times 3 $ rectangle. In case 2, $k^2m=k^22q$ so the $k^2\times m$ rectangle is obtained by gluing $q$ $~k^2\times 2$ rectangles. Each $k^2\times 2$ rectangle, with its particular tableau, corresponds to $g_2(x)$ of Section \[regev\]. Similarly the $k^2\times 3 $ rectangle with its particular tableau corresponds to $g_3(x)$.
In case 1, by the gluing technique we glue together $q$ such tableaux each of shape $(2^{(k^2)})$, together with the single tableau os shape $(3^{(k^2)})$. The corresponding shape is $(m^{(k^2)})$ – with the corresponding polynomial $g_2(x)^q\cdot g_3(x)$, which is proper central of $M_k(F)$, since both $g_2(x)$ and $g_3(x)$ are such.
In case 2 we glue together $q$ such tableaux each of shape $(2^{(k^2)})$. The corresponding shape is $((2q)^{(k^2)})$ – with the corresponding polynomial $g_2(x)^q$, which again is proper central of $M_k(F)$.
Turn now to the integer $0\le r< k^2$. Construct a partition $\pi\vdash r$ together with a tableau $T_\pi$ of shape $\pi$ and with a corresponding polynomial $h_\pi=h_{T_\pi}$, such that $h_\pi(x)$ is [*non*]{} central for $M_k(F)$. Indeed if $r=0$ then we are already done, so assume $r\ge 1$. Then, for example, choose $\pi=(r)$, $T_\pi$ is the one-row tableau filled with $1,\ldots,r$. Then $h_\pi(x)=\sum_{\sigma\in S_r}x_{\sigma (1)}\cdots x_{\sigma(r)}$ which is non central for $M_k(F)$: for example for $i=1,\ldots,r$ substitute $x_i\to\tilde x_i=diagonal(1,2,\ldots,k)$, then $h_\pi(\tilde x)=r!\cdot diagonal(1^r,2^r,\ldots,k^r),$ which is non central since $k\ge 2$.
Since $h_\pi(x)$ is multilinear and of degree $r<k^2$, there exist an injection $$x_d\to \bar x_d \in \{e_{i,j}\mid 1\le i,j\le k\}\quad d=1,\ldots,r$$ such that $h_\pi(\bar x)$ is non central. Complete this substitution to a bijection $x_d\longleftrightarrow\bar x_d=e_{i,j}$ of $x_1,\ldots,x_{k^2}$ with the $k^2$ elements $\{e_{i,j}\mid 1\le i,j\le k\}$, so $g_2(\bar x)$ and $g_3(\bar x)$ are proper central and $h_\pi(\bar x)$ is non-central – of $M_k(F)$.
Glue now $T_\pi$ to the right of $T_\mu$; this can be done since $\ell(\pi)\le k^2$, and we obtain the tableau $T_\mu *T_\pi$. In case 1 that tableau $ T_\mu *T_\pi$ corresponds to the polynomial $g_2(x)^q\cdot g_3(x)\cdot h_\pi(x)$. The shape of $ T_\mu *T_\pi$ is the partition $$\lambda=\mu *\pi=((2q+3)^{(k^2)}) *\pi =(2q+3+\pi_1,\ldots,2q+3+\pi_{k^2})\vdash (2q+3)k^2 +r=n.$$ Similarly in case 2 $$\lambda=\mu *\pi=((2q)^{(k^2)}) *\pi =(2q+\pi_1,\ldots,2q+\pi_{k^2})\vdash 2qk^2 +r=n.$$
Since [*proper-central*]{} multiplied by [*non-central*]{} is [*non-central*]{}, hence in case 1\
$g_2(x)^q\cdot g_3(x)\cdot h_\pi(x)$ is a [*non central*]{} polynomial of $M_k(F)$; it corresponds to $\lambda$, with the corresponding tableau $T_\lambda$, therefore $e_{T_\lambda}$ is non central. By Proposition \[shem.1\] this implies that $\delta _n(M_k(F))\ge f^\lambda $. Similarly in case 2.
Now send $n$ (hence also $m$ and $q$) to infinity, then $\mu$ becomes a large rectangle (of fixed height $k^2$) and $\pi\vdash r$, $0\le r<k^2$, is a small extra part of $\lambda=\mu*\pi$. For the $k^2\times m$ rectangle $\mu=(m^{(k^2)})$ it is known that $$\lim_{m\to \infty}(f^\mu)^{1/|\mu|}= \lim_{|\mu|\to \infty}(f^\mu)^{1/|\mu|}=k^2$$ see for example [@explicit Section 3]. Also, by “Branching”, $f^\lambda\ge f^\mu$ since $\lambda\supseteq \mu$. Hence as $n,$ $m$ and $q$ go to infinity, $$%(c_n^z(M_k(F)))^{1/n}
(\delta_n(M_k(F)))^{1/n} \ge (f^\lambda)^{1/n}\ge (f^\mu)^{1/n}=((f^\mu)^{1/|\mu|})^{|\mu| /n }\to (k^2)^{|\mu| /n}\to k^2$$ since $|\mu|/n\to 1.$ The proof of Lemma \[lower\] now follows.
Some conjectures
----------------
\[C1\] For any integer $k\in\mathbb N$ there exist a real number $0\le \alpha_k \le 1$ such that $$\begin{aligned}
\label{ratio}
\lim_{n\to\infty} \frac{c_n^z(M_k(F))}{c_n(M_k(F))} =\alpha_k.\end{aligned}$$
More generally, an analogue of holds when $M_k(F)$ is replaced by a more general algebra $A$.
Giambruno and Zaicev [@G.Z2], [@G.Z] proved that for any PI algebra $A$, when $n$ goes to infinity $$c_n(A)\sim \alpha\cdot n^g\cdot d^n$$ where $d$ is an integer. In addition Berele [@berele] proved that $g\in \frac{1}{2}\mathbb Z$.
We tend to conjecture that the analogue theorems hold for $\delta_n(M_k(F))$, and probably for other algebras $A$.
[99]{}
A. Berele, Properties of hook Schur functions with applications to p.i. algebras, Adv. in Appl. Math. [**41**]{} (2008) 52-75.
A. P. Brandao, P. Koshlukov, A. Krasilnikov and E. A. Da Silva, the central polynomials for the Grassmann algebra, Israel J. Math. [**170**]{} (2010),127-144.
V. Drensky, Free Algebras and PI-Algebras, Graduate Course in Algebra, Springer (1999).
E. Formanek, Central polynomials for matrix rings, J. Algebra [**23**]{} (1972) 129-132.
E. Formaneck, A conjecture of Regev about the Capelli polynomials, J. Algebra [**109**]{}, (1987) 93-114.
A. Giambruno and A. Valenty, Central polynomials and matrix invariants, Israel J. Math. [**96**]{} (1996), 281-297.
A. Giambruno and M. Zaicev, Exponential codimension growth of P.I. algebras: an exact estimate, Adv. Math. [**142**]{} (1999) 221-243.
A. Giambruno and M. Zaicev, Polynomial Identities and Asymptotic Methods, A.M.S. Mathematical Surveys and Monographs, [**Vol. 122**]{} (2005).
D. Krakowski and A. Regev, The polynomial identities of the Grassmann algebra, Trans. A.M.S. [**181**]{} (1973), 429-438.
J. Olsson and A. Regev, Colength sequence of some $T$ ideals, J. Algebra [**38**]{} (1976) 100-111.
Yu. P. Razmyslov, A certain problem of Kaplansky, Izvestiya Akademii Nauk SSSR. Seriya Matematicheskaya [**37**]{} (1973), 483-501; English translation: Math. USSR Izv. [**7**]{} (1973) 479-496.
A. Regev,The representations of $S_n$ and explicit identities for P.I. algebras, [**51**]{} (1978) 25-40.
A. Regev, The polynomial identities of matrices in characteristic zero, Comm. in Algebra, [**8 (15)**]{}, (1980) 1417-1467.
A. Regev, Codimensions and trace codimensions of matrices are asymptotically equal, Israel J. Math. [**47**]{} (1984), 246-250.
B. E. Sagan, The Symmetric Group, Springer Graduate Texts in Mathematics, Second Edition, (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A particular case of initial data for the two-dimensional Euler equations is studied numerically. The results show that the Godunov method does not always converge to the physical solution, at least not on feasible grids. Moreover, they suggest that entropy solutions (in the weak entropy inequality sense) are not well-posed.'
address: |
Brown University\
Division of Applied Mathematics\
182 George Street\
Providence, RI 02912
author:
- Volker Elling
bibliography:
- '../../mypapers/thesis/elling.bib'
title: A possible counterexample to wellposedness of entropy solutions and to Godunov scheme convergence
---
[^1]
Introduction
============
Consider the Cauchy problem for a system of hyperbolic conservation laws, $$\begin{aligned}
{1}
\frac{\partial u}{\partial t} + \nabla\cdot(\Vec f(u)) &= 0, \label{eq:system} \\
u(0,\cdot) &= u_0, \label{eq:system-inicond}\end{aligned}$$ where $u=u(t,\Vec x):{{\mathbb{R}}^{d+1}_+}:=(0,\infty)\times{\mathbb{R}}^d\rightarrow P\subset{\mathbb{R}}^m$ is the desired solution ($P$ the set of physically reasonable values), $\Vec f=(f^i)$, $f^i:P\rightarrow{\mathbb{R}}^m$, the (smooth) [[*flux*]{}]{} function, $u_0:{\mathbb{R}}^d\rightarrow P$ initial data. Here and in the sequel “$\nabla,\Delta,\cdot$” are meant with respect to $\Vec x$.
An important example of hyperbolic systems of conservation laws are the (nonisentropic) compressible Euler equations: $$\begin{aligned}
{1}
\rho_t + \nabla\cdot(\rho\Vec v) &= 0, \notag \\
(\rho v^i)_t + \nabla\cdot(\rho v^i\Vec v) + p_{x_i} &= 0 \qquad (i=1,\dotsc,d), \notag \\
(\rho e)_t + \nabla\cdot((\rho e+p)\Vec v) &= 0. \label{eq:nonisentropic-euler}\end{aligned}$$ Here, $\rho$ is density, $\Vec v=(v^i)$ velocity, $e$ specific energy, which decomposes into $$\begin{aligned}
{1}
e &= \frac{|\Vec v|^2}{2}+q; \label{eq:specific-energy}\end{aligned}$$ the first summand is specific kinetic energy, $q$ is specific internal energy. The pressure is a function of $\rho,q$; a common choice is the polytropic pressure law $$\begin{aligned}
{1}
p &= (\gamma-1)\rho q \label{eq:nonisentropic-polytropic-pressure}\end{aligned}$$ ($1<\gamma\leq\frac{5}{3}$; for air, $\gamma=\frac{7}{5}$). The set of admissible values is $$P=\{q>0,\ \rho>0\}.$$
It is well-known that and need not have a global smooth solution, even if the initial data $u_0$ is smooth. For this reason, one has to study [[*weak solutions*]{}]{}, defined as functions $u\in L^1_{{\operatorname{loc}}}({{\mathbb{R}}^{d+1}_+};P)$ that satisfy $$\begin{aligned}
{1}
-\int_0^\infty\int_{{\mathbb{R}}^d} u\frac{\partial\phi}{\partial t}+\Vec f(u)\cdot\nabla\phi~d\Vec x~dt &= \int_{{\mathbb{R}}^d}u_0(\Vec x)\phi(0,\Vec x)~d\Vec x, \label{eq:weak}\end{aligned}$$ for all test functions $\phi\in C^\infty_c(\overline{{{\mathbb{R}}^{d+1}_+}})$. Moreover, there can be more than one weak solution, so it is necessary to impose an additional condition, called [[*entropy condition*]{}]{}, to single out a unique weak solution (the [[*entropy solution*]{}]{}).
One definition of entropy solutions is the [[*vanishing viscosity*]{}]{} (VV) definition; it requires that $u$ is the limit of the sequence $(u^\epsilon)_{\epsilon>0}$ of solutions of $$\begin{aligned}
{1}
\frac{\partial u^\epsilon}{\partial t} + \nabla\cdot(\Vec f(u^\epsilon)) &= \epsilon\Delta u^\epsilon\qquad\text{in ${{\mathbb{R}}^{d+1}_+}$,}\label{eq:viscous-system} \\
u^\epsilon(0,\cdot) &= u_0\qquad\text{on $\{0\}\times{\mathbb{R}}^d$}. \label{eq:viscous-system-inicond}\end{aligned}$$ The limit is taken in some suitable topology, usually as a boundedly almost everywhere limit. We call such a function $u$ a [[*VV solution*]{}]{}.
Another definition uses [[*entropy/entropy flux*]{}]{} (EEF) pairs $(\eta,\Vec\psi)$, where $\eta:P\rightarrow{\mathbb{R}}$ is a smooth strictly convex function, called [[*entropy*]{}]{}, whereas $\Vec\psi=(\psi^1,\dotsc,\psi^d)'$ with smooth $\psi^i:P\rightarrow{\mathbb{R}}$ is called [[*entropy flux*]{}]{}; $\eta$ and $\Vec\psi$ are required to satisfy $$\begin{aligned}
{1}
\frac{\partial\psi^i}{\partial u^\alpha}&=\sum_{\beta=1}^m\frac{\partial\eta}{\partial u^\beta}\frac{\partial f^{i\beta}}{\partial u^\alpha}
\qquad(i=1,\dotsc,d,\ \alpha=1,\dotsc,m).
\label{eq:eef}\end{aligned}$$ By multiplying from the left with $\eta'(u^\epsilon)$ and using , one obtains $$\begin{aligned}
{1}
\frac{\partial(\eta\circ u^\epsilon)}{\partial t} + \sum_{i=1}^d\frac{\partial(\psi^i\circ u^\epsilon)}{\partial x^i} &= \epsilon\Delta(\eta\circ u^\epsilon) - \epsilon\sum_{i=1}^d\eta''(u^\epsilon)\frac{\partial u^\epsilon}{\partial x^i}\frac{u^\epsilon}{\partial x^i}
\leq \epsilon\Delta(\eta\circ u^\epsilon). \label{eq:weak-viscous}\end{aligned}$$ (here, we used that $\eta$ is convex). Upon multiplying the last equation with a *nonnegative* test function $\phi$ and integrating by parts, this yields $$\begin{aligned}
{1}
-\int_0^\infty\int_{{\mathbb{R}}^d}\eta(u^\epsilon)\frac{\partial\phi}{\partial t} + \Vec\psi(u^\epsilon)\cdot\nabla\phi~d\Vec x~dt &\leq \epsilon\int_0^\infty\int_{{\mathbb{R}}^d}\eta(u^\epsilon)\Delta\phi~d\Vec x~dt + \int_{{{\mathbb{R}}^d}}\eta(u_0)~d\Vec x. \label{eq:wwv}\end{aligned}$$ If, as assumed above, $(u^\epsilon)\rightarrow u$ boundedly almost everywhere, then implies $$\begin{aligned}
{1}
-\int_0^\infty\int_{{\mathbb{R}}^d}\eta(u)\frac{\partial\phi}{\partial t} + \Vec\psi(u)\cdot\nabla\phi~d\Vec x~dt & \leq \int_{{\mathbb{R}}^d}\eta(u_0)~d\Vec x. \label{eq:eef-entropy}\end{aligned}$$ Functions $u$ that satisfy for *all* EEF flux pairs are called [[*EEF solutions*]{}]{} (of ). As we have shown, VV solutions are necessarily EEF solutions.
In the literature, the term [[*entropy solution*]{}]{} is used to refer either to EEF or to VV solutions, often without explicit mention, because it has been assumed that the two definitions are equivalent for the Euler equations and many other physically relevant systems (see [@serre-1] p. 101, [@dafermos-book] p. 49, [@godlewski-raviart] p. 32; see the discussion in Section \[section:related-work\] for verified special cases). However, for the purposes of this paper it is necessary to distinguish the two notions, as we will discuss a possible numerical counterexample to their equivalence.
The (gas-dynamic) specific entropy $s$ is defined as $$\begin{aligned}
{1}
s &= \log q+(1-\gamma)\log\rho; \label{eq:gas-dynamic-entropy}\end{aligned}$$ $$\begin{aligned}
{1}
\eta &:= -\rho s,\quad \psi^i:=-\rho sv^i\end{aligned}$$ provides an EEF pair for the Euler equations.
A common simplification is to assume that $s$ is constant in space and time. This yields the *isentropic* Euler equations $$\begin{aligned}
{1}
\rho_t + \nabla\cdot(\rho\Vec v) &= 0, \notag \\
(\rho v^i)_t + \nabla\cdot(\rho v^i\Vec v) + p_{x_i} &= 0 \qquad (i=1,\dotsc,d) \label{eq:isentropic-euler}\end{aligned}$$ with $$\begin{aligned}
{1}
p(\rho) &= \rho^\gamma. \label{eq:isentropic-polytropic-pressure}\end{aligned}$$ In this case, $P=\{\rho>0\}$. An EEF pair is provided by the specific energy $e$, $$\begin{aligned}
{1}
e &= \frac{|\Vec v|^2}{2}+\frac{\rho^{\gamma-1}}{\gamma-1},\end{aligned}$$ with $$\begin{aligned}
{1}
\eta &:= \rho e, \quad \psi^i:=(\rho e+p)v^i.\end{aligned}$$
It is cumbersome to verify the EEF condition directly, not to mention the VV condition. There are easier criteria for piecewise smooth functions, which we define in the following customized way:
1. A point $(t,\Vec x)\in{{\mathbb{R}}^{d+1}_+}$ is called [[*point of smoothness*]{}]{} if $u$ is $C^\infty$ in a small neighbourhood of $(t,\Vec x)$.
2. A point $(t,\Vec x)\in{{\mathbb{R}}^{d+1}_+}$ is called [[*point of piecewise smoothness*]{}]{} of $u$ if there is a $C^\infty$ diffeomorphism $\Phi$ of a ball $V$ around $0$ in ${\mathbb{R}}^{d+1}$ onto a neighbourhood $B$ of $(t,\Vec x)=\Phi(0)$ so that $u\circ\Phi$ is $C^\infty$ on $B_-$ and on $B_+$ (where $B_\pm:=\Phi(V_\pm)$, $V_\pm:=\{y\in V:y_1\gtrless 0\})$; for later use, let $S$ be the surface $\Phi(V\cap(\{0\}\times{\mathbb{R}}^d))$, $n=(n^t,\Vec n)\in{\mathbb{R}}^{d+1}$ a unit normal to $S$ in $(t,\Vec x)$ pointing into $B_+$; let $u_+,u_-$ be the one-sided limits of $u$ in $(t,\Vec x)$ within $B_-$ resp. $B_+$). We also require $\vec n\neq 0$.
3. $u$ is called [[*piecewise smooth*]{}]{} if there is a set $N$ of $d$-dimensional Hausdorff measure $0$ so that all points in ${{\mathbb{R}}^{d+1}_+}-N$ are points of piecewise smoothness.
Let $u$ be piecewise smooth. $u$ is an EEF solution of if and only if
1. it is a (classical) solution of in each point of smoothness,
2. $u(t,\cdot)\rightarrow u_0$ in $L^1_{{\operatorname{loc}}}$ as $t\downarrow 0$, and
3. in each point $(t,x)$ of piecewise smoothness it satisfies the Rankine-Hugoniot conditions $$\begin{aligned}
{1}
(u_+-u_-)n^t + (\Vec f(u_+)-\Vec f(u_-))\cdot\Vec n &= 0 \label{RH}
\end{aligned}$$ and (for all EEF pairs $(\eta,\Vec\psi)$) $$\begin{aligned}
{1}
(\eta(u_+)-\eta(u_-))n^t + (\Vec\psi(u_+)-\Vec\psi(u_-))\cdot\Vec n &\leq 0. \label{RH-EEF}
\end{aligned}$$
\[th:piecewise\]
Proposition \[th:piecewise\] is well-known (see, for example, Section 11.1.1 in [@evans]), as is the following property:
\[rem:shock-vn\] For the Euler equations resp. (with polytropic gas law resp. ), is equivalent to the simpler condition that the normal velocity does not increase across discontinuities: $$\begin{aligned}
{1}
(\vec v_+-\vec v_-)\cdot\vec n &\leq 0.
\end{aligned}$$
The Cauchy problem for the Euler equations has several important symmetry properties, including the following:
\[th:symmetries\] Let $u=(\rho,\Vec v',q)'$ be a weak solution for initial data $u_0=(\rho_0,\Vec v_0',\Vec q_0)'$.
1. Change of inertial frame: For all $\Vec w\in{\mathbb{R}}^d$, $(\rho(x+\Vec wt),(\Vec v(x+\Vec wt,t)-\Vec w)',q(x+\Vec wt))'$ is a weak solution for the same initial data $u_0$.
2. Self-similarity: a function $f:{{\mathbb{R}}^{d+1}_+}\rightarrow{\mathbb{R}}^m$ is called [[*self-similar*]{}]{} if $f(rt,r\Vec x)=f(t,\Vec x)$ for all $r>0$; same for functions on ${{\mathbb{R}}^d}$. If the initial data is self-similar, then for any $r>0$, $u(r\Vec x,rt)$ is a weak solution for the same initial data $u_0$.
These symmetries remain true after replacing “weak” by “VV” or “EEF”. Analogous symmetries hold for the isentropic case.
Example and numerical results {#section:numerical-results}
=============================
Consider the following set $u_0$ of initial data for with $d=2$ (see Figure \[fig:example\]): the data is symmetric under reflection across the $x$-axis and constant in each of four cones centered in the origin (in particular, constant along rays starting in the origin). In the origin, two shocks emanate into the first and fourth quadrant; the area on the left is supersonic inflow (parallel to the $x$-axis); the two areas on the other side of the shocks are denser and hotter gas, moving parallel to the contact discontinuities (see [@courant-friedrichs] Chapter IV C on choosing pre- and post-shock values that satisfy the Rankine-Hugoniot conditions; we choose the ones that yield the weaker shock). The gas in the stagnation area (enclosed by the contact discontinuities) has the same pressure as the post-shock gas on the other side, but velocity $\Vec v=0$. It is easy to check, using Propositions \[th:piecewise\] and \[rem:shock-vn\], that the steady solution $u(t,\Vec x)=u_0(\Vec x)$ is an EEF solution of resp. . Henceforth we refer to it as [[*Solution T*]{}]{} (for [[*theoretical*]{}]{}).
![Solution N: (Rotate clockwise by $90^o$ to align with dotted area in Figure \[fig:example\].) Each square in the coordinate grid corresponds to a $100~m/s\times100~m/s$ square in the $\vec\xi$ plane. Origin marked by diamond (lower left corner). Plotted: horizontal velocity. Godunov scheme for isentropic Euler equations; data: $\gamma=1.4$, $\alpha=10^o$; inflow: $\rho=1.19~kg/m^3$, $v=1000~m/s$, $T=20^oC$. The solution differs significantly from Figure \[fig:example\]. Results for nonisentropic Euler equations or other numerical schemes are similar. []{data-label="fig:second-vx"}](godunov-vx.eps){width="\textwidth"}
[However, instead of Solution T, numerical calculations produce the markedly different result in Figure \[fig:second-vx\] which we call [[*Solution N*]{}]{} (for [[*numerical*]{}]{}) in the sequel (of course it is not known to be an exact solution). The numerical domain in Figure \[fig:second-vx\] is indicated as the dotted quadrilateral in Figure \[fig:example\]. Figure \[fig:second-vx\] was computed as follows:]{} adaptive refinement was used to achieve better resolution at same computational cost. To reduce numerical viscosity the grid was chosen so that near the right domain boundary the edges are aligned with the contact discontinuity and the shock. In order to capture self-similarity, the computations were done for a grid with moving vertices with coordinates $\Vec x=t\Vec\xi$ ($\Vec\xi$ has the dimension of a velocity; its components are called [[*similarity coordinates*]{}]{}). The moving-edge modifications discussed in [@elling-diplom Section 2.1.6] and [@phd-thesis Chapter 4] were used (the essential idea is to compute numerical fluxes across a moving edge by transforming to a steady edge, using invariance under change of inertial frame (Proposition \[th:symmetries\]), and to apply an arbitrary approximate Riemann solver to the transformed problem). The domain boundaries were chosen so that small perturbations on them propagate into the domain ($\Vec\xi\cdot\Vec n$ ($\Vec n$ outer unit normal) in each boundary point is larger than the maximum of $|\Vec v|+c$ in the domain). This allows to prescribe all components of the fluxes on the boundary.
Experiments with various modifications were made: changing the numerical scheme (the experiments were repeated for the Godunov scheme [@godunov], the Osher-Solomon scheme [@solomon-osher], the ENO-RF scheme [@shu-osher-llf], and a second-order MUSCL code based on the first-order ENO-RF scheme), adding more numerical dissipation, refining uniformly rather than adaptively, using a Cartesian grid including origin and lower half-plane, or calculating in space rather than similarity coordinates. None of these modifications change the numerical results significantly; in all cases, the numerical results converge to the same Solution N.
[Solution N appears to be self-similar (i.e. steady in similarity coordinates), but it is strongly *unsteady*, so it is clearly different from Solution T.]{}
Conclusions about numerical methods {#section:numfail}
===================================
[While the discrepancy between Solutions N and T opens many new problems, we can already draw one definite conclusion.]{}
If we assume that [Solution T is the correct solution]{}, many[^2] popular numerical schemes fail to converge to physical solutions. Although it cannot be ruled out that they ultimately converge to Solution T as the numerical grid becomes infinitely fine, they approach Solution N for computationally accessible grids — which is all that matters for practical purposes.
On the other hand, if we assume that [Solution N]{} is the correct solution, there is a trivial *theoretical* example of misconvergence: consider the (semidiscrete) Godunov scheme on grids whose edges are exactly aligned with the discontinuities of [Solution T]{} (see Figure \[fig:example\]): in exact arithmetic it would have [Solution T]{} as steady state on *every* grid.
[In either case — even if Solution N is correct, which would be less catastrophic for numerical analysis —]{} we have to conclude that discrete entropy inequalities are not sufficient to avoid convergence to unphysical solutions on feasible grids. Hence they lose a bit of their value as design principles for numerical schemes, although they are still useful as easy-to-check necessary conditions that are sufficient for scalar conservation laws and (probably) 1D systems (as supported by the recent work on small total variation solutions described in Section \[section:related-work\]).
[Although many reports of deficiencies of various numerical schemes have been published, the clear case of failure observed here has no precedent.]{}
01 The Godunov setup above appears unstable, especially if any amount of artificial viscosity is added, and may be rare in practice. However, the well-known “carbuncle phenomenon” (see [@quirk]), which is rather persistent, is another example of a phenomenon that affects especially numerical methods based on Riemann solvers and vanishes if a modicum of artificial viscosity is added to the numerical scheme. It is possible that the carbuncle phenomenon and the numerical phenomena discussed in this paper are related; in fact Figure \[fig:example\] has a certain similarity to the numerically observed “carbuncles”. [@robinet-gressier-casalis-moschetta] discovered a new unstable mode in the linearized Euler equations for a plane shock wave and conjecture that carbuncles are an inherent instability of the Euler equations (TODO: need to be very precise in this quotation), i.e. in some way reflect a particular exact solution. Based on the findings in this paper, we propose to modify the conjecture in [@robinet-gressier-casalis-moschetta] as follows: carbuncles reflect an inherent instability or even nonuniqueness in the Euler equations with EEF condition that does not appear if the VV condition is used, in correspondence to the disappearance of carbuncle phenomena if artificial viscosity is added to numerical schemes.
Theoretical interpretation
==========================
It remains to discuss which of Solution T and Solution N is the physical one and, if Solution T is correct, what causes numerical schemes to produce Solution N. There are three possible explanations (which are not mutually exclusive):
1. either Solution N is an example of failure of numerical methods, or
2. EEF solutions are not stable (in the sense of continuous dependence on initial data), or
3. EEF solutions are not unique.
Breakdown of numerical methods
------------------------------
[It has already been shown in Section \[section:numfail\] that the Godunov scheme is flawed, in the sense that it can fail to converge to the physical solution on feasible grids. Hence it is natural to suspect that Solution N is a numerical artifact that does not correspond to a seccond EEF solution (or any other type of solution of the Euler equations).]{}
[Solution T is steady and self-similar.]{} “Steadyness” is a non-generic property that is usually not inherited by finite-accuracy numerical solutions (for example for a Riemann problem that is solved exactly by a single shock, most numerical schemes produce small additional waves and a slightly different shock). However, one would expect numerical approximations to be at least *almost* steady, unlike Solution N.
[In a single space dimension, the conservation property of numerical schemes often guarantees accurate shock locations, even if the overall accuracy of the scheme is poor.]{} [On the other hand,]{} in two or more dimensions numerical imprecision can significantly change the shape and location of shocks. This may be the cause of Solution N. For example, the upwards deflection of the incoming flow by the high-pressure area in front of the stagnation region could be weaker in numerical calculations than in Solution T; the additional pressure would cause the stagnation region to collapse.
However, in this case the numerical results would depend strongly on the choice of [numerical method, mesh width and other parameters.]{} This is not observed; rather, all choices produce essentially the same results.
Instability
-----------
[A second explanation is instability (in the sense of lack of continuous dependence on the initial data). It is possible that Solution N is an approximation to an unsteady EEF solution T’ that results from a slight perturbation of solution T at initial time]{} (such perturbations are inevitable in most numerical computations due to inexact arithmetic, discretization error, artificial viscosity etc.) Since [Solution N]{} is produced (up to minor differences) for any “perturbation” (i.e. for any choice of mesh, numerical method and parameters), it would indicate that [Solution T]{} constitutes a set of initial values for which the Euler equations are not stable.
On inspection in similarity coordinates, it appears that the numerical solutions are bounded, converge quickly to Solution N and remain steady (many orders of magnitude of time have been observed), so Solution T’ would have to be at least approximately self-similar. If we assume it is *asymptotically* self-similar for large time, then the following theorem implies that the asymptote is an EEF solution:
\[th:asymptotic\] Let $u\in L^\infty({{\mathbb{R}}^{d+1}_+})$ be an EEF solution of . Assume that $u$ is [[*asymptotically self-similar*]{}]{} (see Definition \[def:asymptotic\]), then its asymptotic limit $w$ (a self-similar function) is an EEF solution as well.
(The proof of Theorem \[th:asymptotic\] and an analogous result for steady solutions are presented in the Appendix.) The asymptote would assume the same initial data as Solution T, but would have to be different from it (by closeness to Solution N). This would already imply the third explanation (nonuniqueness of EEF solution).
[To avoid that, it is necessary to assume either that Solution T’ is approximately, but *not asymptotically* self-similar — for example it might oscillate periodically around some self-similar function without approaching it —, or that the self-similar asymptote has data at infinity that does not match the initial data (Solution T). Either of these cases would be revealed by a sufficiently fine numerical grid: the numerical computation on that grid would refuse to converge to a steady state for the given boundary data. But the mesh used to compute Solution N is already rather fine, as can be seen from the curved shocks in Figure \[fig:second-vx\]; there is no obvious reason why an even better grid is required.]{}
Nonuniqueness of EEF solutions
------------------------------
[The third explanation is nonuniqueness: Solution N corresponds to an EEF solution that assumes the same initial data as Solution T.]{}
In this context, the following peculiarity is important: according to Proposition \[th:piecewise\], the EEF condition is “insensitive” to sets with $(d-1)$-dimensional Hausdorff measure $0$ (such as a single point, for $d=2$); e.g. if is satisfied for $\phi\in C_c^\infty(\overline{{{\mathbb{R}}^{d+1}_+}}-\{0\})$, it is satisfied for *all* $\phi\in C_c^\infty(\overline{{{\mathbb{R}}^{d+1}_+}})$. [In verifying that Solution T is an EEF solution, we may ignore the rather singular wedge tip in the origin of Figure \[fig:example\]. It is counterintuitive that conditions for the physical correctness of solutions may ignore such singularities.]{}
It seems unlikely that the solution in Figure \[fig:example\] is stable under small perturbations at the origin (such as perturbations from viscous terms in the VV limit). Note that a planar shock, with inflow state on one side and stagnation area state on the other side, would *not* be steady but move into the stagnation zone quickly — it seems unlikely that the example data, which has less mass and energy and more $x$-momentum in the $\{x>0\}$ halfplane, would yield a steady pattern in the origin (however, the “maximum principle” implicit in this argument is merely heuristic and may be wrong in some instances). Figure \[fig:example\] is closely related to the well-known Prandtl-Meyer problem of flow along solid walls with corners. The physically observed pattern [(which is at the same time]{} an EEF solution of the Euler equations) is on the left of Figure \[fig:bend\] (see Section 111 in [@courant-friedrichs] for discussion). On the right is another steady EEF solution; as in Figure \[fig:example\], there is flow along a contact discontinuity with a stagnation area on the other side. This solution is not observed in physical[^3] experiments. This analogy is further evidence that Figure \[fig:second-vx\] is indeed an approximation to the physically correct solution for the initial data in Figure \[fig:example\].
But it is the following observation that provides the strongest argument for nonuniqueness: the Lax-Wendroff theorem (see [@lax-wendroff]; see also [@godlewski-raviart], [@kroener-rokyta-wierse] and most generally [@elling-lax-wendroff] for Lax-Wendroff-type theorems for irregular grids) states that if a numerical scheme is consistent and satisfies a discrete entropy inequality (see [@harten-hyman-lax; @majda-osher-discrete-entropy; @osher-chakravarthy; @tadmor-entropy-mathcomp; @tadmor-entropy-stable; @osher-tadmor]), then the limit of a boundedly almost everywhere converging sequence of numerical solutions is an EEF solution. [(Note that we prescribe the full numerical flux on the boundaries, so the boundary conditions are analogous to an initial condition and can be treated with a straightforward modification of the Lax-Wendroff theorem for initial-value problems.)]{} The Godunov scheme, used to compute Figure \[fig:second-vx\], is the standard example for a consistent scheme that satisfies all discrete entropy inequalities. Our numerical solutions do, on inspection, appear to converge quickly; this would imply that Solution N corresponds to an EEF solution.
Related work {#section:related-work}
============
For multidimensional scalar ($m=1$) conservation laws with arbitrary $f$, [@kruzkov] (generalizing earlier work) shows that a global EEF solution exists, is unique, satisfies the VV condition as well, and is stable under $L^1$ perturbations of the initial data.
[@glimm] provides a famous existence proof for strictly hyperbolic *systems* with genuinely nonlinear fields and initial data with small total variation; the interaction functionals constructed in this paper are a crucial ingredient for all subsequent work. [@liu-admissibility-memoir] extends the result to systems with linearly and some nonlinearly degenerate fields. [@bressan-crasta-piccoli] constructed the Standard Riemann Semigroup (SRS), an $L^1$-stable semigroup of EEF solutions for initial data with small total variation, for strictly hyperbolic systems with genuinely nonlinear or linearly degenerate fields (see also [@liu-yang]). [@bressan-lefloch] showed that EEF solutions to 1D systems are unique and coincide with the SRS solutions, under certain smoothness assumptions including small total variation (see also [@bressan-goatin]). [@bianchini-bressan] prove that for small $TV$ initial data and strictly hyperbolic (but otherwise arbitrary) systems VV solutions exist and are stable under $L^1$ perturbations of the initial data, so for the class of solutions that are subject both to [@bressan-lefloch] and to [@bianchini-bressan], EEF and VV solutions are equivalent.
On the other hand, an EEF pair $(\eta,\Vec\psi)$ has to satisfy the condition which is an overdetermined problem for $m\geq 3$, so for some systems no EEF pairs exist and the EEF condition is void. However, EEF pairs do exist for most physically relevant systems, even those with $m\geq 3$. More seriously, for certain $2\times 2$ systems (with nonlinear degenerate fields) [@conlon-liu] construct a single weak shock that is an EEF solution but does not satisfy the Liu entropy condition (see [@liu-condition-uniqueness-twobytwo; @liu-condition-uniqueness-general]). By [@bianchini-bressan], there must be a VV solution (for the same initial data) that satisfies the Liu entropy condition as well — so it cannot be the aforementioned weak shock. Therefore the example in [@conlon-liu] also constitutes an example of a nonunique EEF solution, albeit for an “artificial” system with nonlinear degeneracy.
[@hopf-eef] proposes the EEF condition for scalar conservation laws ($m=1$), proves that it is implied by the VV condition under some circumstances and notes that there is a large set of convex entropies. Apparently independently, [@kruzkov] obtained analogous results for systems. [@lax-zarantonello] contains the first use of the term “entropy condition” for the EEF condition. Various forms of the EEF condition had been known and in use for special systems such as the Euler equations for a long time (e.g. by the name of [[*Clausius-Duhem inequality*]{}]{}), especially as shock relations; however, the above references seem to be the first to define the general notion of strictly convex EEF pairs, to propose the EEF condition as a mathematical tool for arbitrary systems of conservation laws and to formulate it in the weak form rather than the special case .
[@li-zhang-yang] provide an analytical and numerical discussion of 2D Riemann problems for various systems including the Euler equations. However, they focus on data constant in each of the four quadrants, so Solution T is not covered.
01 (TODO: what about viscous traveling wave solutions for large TV data riemann problems for Euler equations?) (TODO: what about $2\times 2$ conservation laws with large data?) (TODO: look at diperna paper cited in goatin/lefloch article.)
(TODO: mention that [@courant-friedrichs] fig 30 p. 296 is our example if the wedge is replaced by stagnation air; see p. 345 Fig 69. Quote: “All these and other mathematically possible flow patterns with a singular center Z are at our disposal for interpreting experimental evidence. Which, if any, of these possibilities occurs under given circumstances is a question that cannot possibly be decided within the framework of a theory with such a high degree of indeterminacy. Here we have a typical instance of a theory incomplete and oversimplified in its basic assumptions; only by going more deeply into the physical basis of our theory, i.e. by accounting for heat conduction and viscosity, can we hope to clarify completely the phenomena at a three-shock singularity. It may well be that the boundary layer which develops along the constant discontinuity line modifies the flow pattern sufficiently to account for the observed deviation; \[...quote Liepmann paper\]”
Conjectures and final remarks
=============================
The total variation of the initial data cannot be made arbitrarily small because the oblique shock relations can be solved only for inflow velocity above some supersonic limit (which depends on $\alpha$). It would be interesting to find modified examples with arbitrarily small total variation.
The results demonstrate that
*the Godunov method does not always converge to the physical solution on feasible grids.*
Moreover, they suggest the following conjecture:
*EEF solutions to the multidimensional Euler equations are not always unique.*
If this conjecture is true, it would have far-reaching consequences. The EEF condition would not be sufficient as a selection principle for physical/unique solutions, except in special cases like the ones described in Section \[section:related-work\]. It would be necessary to find ways to use the cumbersome VV condition or to discover new entropy conditions.
Although the numerical results support the conjecture unambiguously, the question is so important that a rigorous proof is highly desirable. However, since the initial data has large vorticity at the contact discontinuity, it seems difficult to construct (or to prove results about) exact solutions. One possible line of attack is to derive novel entropy conditions from the VV condition and to check whether they are violated by the steady solution in Figure \[fig:example\].
In any case, this paper motivates the investigation of multidimensional Riemann problems for systems; these appear to be very difficult and exhibit a large variety of phenomena (see [@lax-xdliu; @li-zhang-yang]). This goal requires techniques for proving existence of smooth steady or self-similar solutions to boundary-value problems for systems of nonlinear hyperbolic conservation laws; while there are classical methods for smooth solutions in hyperbolic regions, work on tools for the elliptic and mixed case has begun only recently (see e.g. [@elling-liu-ellipticity-journal]). Finally, one could wonder whether Figure \[fig:example\] is an example of a “generic” phenomenon or whether EEF solutions are unique for all initial data outside a “small” complement. This question is being investigated; we suspect that nonuniqueness is indeed generic.
Appendix: asymptotically steady and self-similar weak solutions {#appendix-asymptotically-steady-and-self-similar-weak-solutions .unnumbered}
===============================================================
Remark: in Theorem \[th:asymptotic\] and in the following statements, $$f(t,\cdot) \rightarrow g \qquad\text{in $L^1_{{\operatorname{loc}}}(\Omega)$}$$ as $t\downarrow 0$ resp. $t\uparrow\infty$ is to be understood as: for all $\epsilon>0$ and $K\Subset\Omega$ there is a $T=T(\epsilon)>0$ so that for almost all $0<t\leq T$ resp. $t\geq T$, $$\| f(t,\cdot)-g \|_{L^1(K)} \leq\epsilon.$$
\[lemma:weak\] Let $u\in L^\infty({{\mathbb{R}}^{d+1}_+})$, $u_0\in L^\infty({{\mathbb{R}}^d})$. If
1. $u(t,\cdot)\rightarrow u_0$ in $L^1_{{\operatorname{loc}}}({\mathbb{R}}^d)$ as $t\downarrow 0$, and
2. $u$ satisfies for all $\phi\in C^\infty_c(\underline{(}0,\infty)\times{\mathbb{R}}^d)$,
then $u$ satisfies for all $\phi\in C^\infty_c(\underline{[}0,\infty)\times{\mathbb{R}}^d)$.
Let $\theta\in C^\infty[0,\infty)$ so that $\theta=1$ on $[0,1]$, $\theta=0$ on $[2,\infty)$. For any $T>0$, define $\theta_T(t):=\theta(T^{-1}t)$. Note that $|\theta_T|=O(1)$, $|\theta_T'|=O(T^{-1})$ (as $T\downarrow 0$). For given $\phi\in C_c^\infty([0,\infty)\times{{\mathbb{R}}^d})$, split $\phi_1(t,x):=\theta_T(t)\phi(t,x)$ and $\phi_2=\phi-\phi_1$. $$\begin{aligned}
{1}
&\int_0^\infty\int_{{{\mathbb{R}}^d}}u\phi_t+\Vec f(u)\cdot\nabla\phi~dx~dt \\
&= \int_0^{2T}\int_{{{\mathbb{R}}^d}}u\phi_{1t}+\Vec f(u)\cdot\nabla\phi_1~dx~dt + \int_T^\infty\int_{{{\mathbb{R}}^d}}u\phi_{2t}+\Vec f(u)\cdot\nabla\phi_2~dx~dt.
\end{aligned}$$ Since $\phi_2\in C_c([T,\infty)\times{{\mathbb{R}}^d})$, the second summand vanishes by assumption. The first summand equals $$\begin{aligned}
{1}
&= O\left(T\sup_{0<t\leq 2T}\|u(t,\cdot)-u_0\|_1\cdot(T^{-1}+1)\right)
+ \int_0^{2T}\int_{{{\mathbb{R}}^d}}u_0\phi_{1t}~dx~dt\\
&+O\left(T(\|u_0\|_1+\sup_{0<t\leq 2T}\|u(t,\cdot)-u_0\|_1\right) \\
&= O\left(\sup_{0<t\leq 2T}\|u(t,\cdot)-u_0\|_1\right) - \int_{{{\mathbb{R}}^d}}u_0(x)\phi(0,x)~dt \\
&+ O\left(T(\|u_0\|_1+\sup_{0<t\leq 2T}\|u(t,\cdot)-u_0\|_1\right).
\end{aligned}$$ On taking $T\downarrow 0$, all $O$ terms vanish; hence $u$ satisfies .
Remark: the converse of Lemma \[lemma:weak\] (which is not needed) is not immediate because $u(t,\cdot)-u_0$ may be large for some $t$ as long as the set of such $t$ has small measure near $0$.
\[def:asymptotic\]
1. A function $u\in L^1_{{\operatorname{loc}}}({{\mathbb{R}}^{d+1}_+};{\mathbb{R}}^m)$ is called [[*asymptotically self-similar*]{}]{} if there is a function $w:{\mathbb{R}}^d\rightarrow{\mathbb{R}}^m$ so that $$u(t,t^{-1}\cdot) \rightarrow w\qquad\text{in $L^1_{{\operatorname{loc}}}(R^d)$}.$$
2. $u$ is called [[*self-similar*]{}]{} if, for some $w$, $u(t,t^{-1}\cdot)=w$ for almost all $t>0$.
(of Theorem \[th:asymptotic\]) By Lemma \[lemma:weak\], to show that $w$ is a weak solution it is sufficient to check that $$\begin{aligned}
{1}
\int_0^\infty\int_{{\mathbb{R}}^d}w\left(\frac{x}{t}\right)\phi_t(t,x)+\Vec f\left(w\left(\frac{x}{t}\right)\right)\cdot\nabla\phi(t,x)~dx~dt
&= 0 \label{eq:asymptotic-1}
\end{aligned}$$ for all $\phi\in C^\infty_c(]0,\infty))$. The essential idea is to scale coordinates to shift the support of $\phi$ into a large-$t$ region and to use asymptotic convergence.
Let $0<t_1<t_2$ be such that ${\operatorname{supp}}\phi\subset[t_1,t_2]\times{\mathbb{R}}^d$. Let $\epsilon>0$ be arbitrary, set $T=T(\epsilon)$ as in Definition \[def:asymptotic\]. The change of coordinates $t=\frac{t_1}{T}\tau$, $x=\frac{t_1}{T}\xi$ changes the left-hand side of into $$\begin{aligned}
{1}
& \left(\frac{t_1}{T}\right)^{d+1}\int_T^{\frac{t_2T}{t_1}}\int_{{{\mathbb{R}}^d}}w\left(\frac{\xi}{\tau}\right)\phi_t\left(\frac{t_1}{T}\tau,\frac{t_1}{T}\xi\right)
+ \Vec f\left(w\left(\frac{\xi}{\tau}\right)\right)\cdot\nabla\phi(\tau,\xi)~d\xi~d\tau \notag \\
& = \left(\frac{t_1}{T}\right)^{d+1}\int_T^{\frac{t_2T}{t_1}}\int_{{{\mathbb{R}}^d}}u(\tau,\xi)\phi_t\left(\frac{t_1}{T}\tau,\frac{t_1}{T}\xi\right)
+ \Vec f(u(\tau,\xi))\cdot\nabla_x\phi\left(\frac{t_1}{T}\tau,\frac{t_1}{T}\xi\right)~d\xi~d\tau \notag \\
& + O\left(\left(\frac{t_1}{T}\right)^{d+1}\cdot T\cdot\epsilon T^d\right) \label{eq:asymptotic-2}
\end{aligned}$$ where $O$ is with respect to $\epsilon\rightarrow\infty$. Note that the support of the scaled $\phi$ is in $[T,\infty)\times{\mathbb{R}}^d$. Also, the assumption that $u$ is bounded is essential here. The second summand on the right-hand side equals $$\begin{aligned}
{1}
& \int_T^{\frac{t_2T}{t_1}}\int_{{{\mathbb{R}}^d}}u(\tau,\xi)\frac{T}{t_1}\phi\left(\frac{t_1}{T}\tau,\frac{t_1}{T}\xi\right)_\tau +
\Vec f(u(\tau,\xi))\cdot\nabla_\xi(\phi(\tau,\xi))~d\xi~d\tau.
\end{aligned}$$ Since $u$ is assumed to be a weak solution, this term vanishes. Taking $\epsilon\downarrow 0$ in yields .
For the proof of the EEF part, replace $u,w$ by $\eta(u),\eta(w)$ and $f(u),f(w)$ by $\psi(u),\psi(w)$ above.
The same results as for self-similar weak solutions can be obtained for steady solutions:
\[def:asymptotically-steady\]
1. $u\in L^1_{{\operatorname{loc}}}({{\mathbb{R}}^{d+1}_+};{\mathbb{R}}^m)$ is called [[*steady*]{}]{} if, for some $w:{\mathbb{R}}^d\rightarrow{\mathbb{R}}^m$, $u(t,\cdot)=w$ for almost all $t>0$.
2. $u\in L^1_{{\operatorname{loc}}}({{\mathbb{R}}^{d+1}_+};{\mathbb{R}}^m)$ is called [[*asymptotically steady*]{}]{} if there is a $w:{\mathbb{R}}^d\rightarrow{\mathbb{R}}^m$, so that $$u(t,\cdot) \rightarrow w\qquad\text{in $L^1_{{\operatorname{loc}}}({\mathbb{R}}^d)$}$$ for almost all $t\geq T$.
If $u\in L^\infty({{\mathbb{R}}^{d+1}_+};{\mathbb{R}}^m)$ is an asymptotically steady and bounded weak solution, then $w$ (as in Definition \[def:asymptotically-steady\]) is a weak solution as well. If $u$ is an EEF solution, so is $w$.
Let $\phi\in C^\infty_c((0,\infty)\times{{\mathbb{R}}^d})$ be arbitrary. Let ${\operatorname{supp}}\phi\subset[0,\tau]$. For any $\epsilon>0$, $$\begin{aligned}
{1}
& \int_0^\infty\int_{{{\mathbb{R}}^d}}w(x)\phi_t(t,x)+\Vec f(w(x))\cdot\nabla\phi(t,x)~dx~dt \\
&= \int_0^\infty\int_{{{\mathbb{R}}^d}}u(T+t,x)\phi_t(t,x)+\Vec f(u(T+t,x))\cdot\nabla\phi(t,x)~dx~dt + O(\tau\epsilon\|D\phi\|_\infty) \\
&= O(\tau\epsilon\|D\phi\|_\infty)
\end{aligned}$$ because we can extend $\phi(\cdot-T,\cdot)\in C_c((T,\infty)\times{{\mathbb{R}}^d})$ smoothly by $0$ to a map $\tilde\phi\in C_c((0,\infty)\times{{\mathbb{R}}^d})$ and use that $u$ is a weak solution. Lemma \[lemma:weak\] shows that $w$ is a weak solution.
For the proof of the EEF part, replace $u,w$ by $\eta(u),\eta(w)$ and $f(u),f(w)$ by $\psi(u),\psi(w)$ above.
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank Tai-Ping Liu for his support and comments and Ron Fedkiw, Doron Levy, Wolfgang Dahmen and Ralf Massjung for sharing their insight about numerical schemes.
01
More speculation {#more-speculation .unnumbered}
================
Colella and Woodward say that Godunov produces an unphysical shock at the forward-facing step. But then we know that Godunov should not produce unphysical things. The truth is that it does, it’s just that physical is not the same as entropy-admissible: their example is precisely the Prandtl-Meyer corner where one expects a rarefaction wave, but a contact (or shock here, since waves impinge on the contact) solution is also possible.
[^1]: This material is based upon work supported by an SAP/Stanford Graduate Fellowship and by the National Science Foundation under Grant no. DMS 0104019. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation.
[^2]: [in the sense of: every scheme that was tested]{}
[^3]: [It is not observed in sophisticated numerical computations either. However, some classical schemes — particularly the Godunov scheme — are prone to produce the solution on the right. This adds another example for numerical failure (similar to Section \[section:numfail\] and for insufficiency of the EEF condition. However, this example is not as strong as the one in Figures \[fig:example\] and F\[fig:second-vx\] because it has a solid boundary.]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we give an upper bound for the number of SRB measures of saddle type of local diffeomorphisms of boundaryless manifolds in terms of maximal cardinality of set of periodic points without any homoclinic relation.'
address:
- 'ICMC-USP São Carlos, Caixa Postal 668, 13560-970 São Carlos-SP, Brazil.'
- 'Departamento de Matemática, ICMC-USP São Carlos, Caixa Postal 668, 13560-970 São Carlos-SP, Brazil.'
author:
- 'P. Mehdipour'
- 'A. Tahzibi'
title: SRB measures and Homoclinic relation for Endomorphisms
---
[^1]
[^2]
[^3]
**Introduction**
================
The contrast between topological and measure theoretical properties is an interesting subject which frequently appears in the study of dynamics.
In a beautiful simple construction, I. Kan [@5] gave an example of a local diffeomorphism $f$ defined on the cylinder $\mathbb{S}^1 \times [0,1]$ such that $f$ is topologically transitive and moreover, $f$ admits two SRB measures with intermingled basins. Besides the richness of intermingled basins property, the non-uniqueness of SRB measures joint with topological transitivity is amazing.
In [@13], the authors proved that the above phenomenon can not exist for surface diffeomorphism. More precisely they proved
[@13] Let $f: M \rightarrow M$ be a $C^{1+\alpha}, \alpha > 0$ diffeomorphism over a compact surface $M.$ If $f$ is topologically transitive then there exists at most one (hyperbolic) SRB measure.
In this paper we deal with endomorphisms and by an endomorphism we refer to local diffeomorphism of a closed Riemannian manifold (compact and boundaryless). We remark that the endomorphism setting brings many surprises and distinctions with respect to the diffeomorphism context. Recall that I. Kan’s example is made on two dimensional cylinder. His construction was extended by Ilyashenko, Kleptsy, Saltykov [@IKS]. See also [@BDV 11.1.1]. By the way, it is possible to use Kan’s example and construct a transitive [^4] endomorphism on $\mathbb{T}^2$ with two SRB measures of intermingled basins contrasting the above theorem for the case of non-invertible dynamics.
Here we find an upper bound for the number of ergodic SRB measures of [*saddle type*]{} in terms of the maximial cardinality of set of periodic points without any homoclinic relation. By a SRB measure of saddle type we mean a SRB measure whitout zero Lyapunov exponents and having both positive and negative ones.
For any endomorphism $f: M \rightarrow M$ we denote by $ \tilde{f}: M^f \rightarrow M^f$ the natural extension of $f$ where $\tilde{f}$ is the shift map. The natural projection $\pi: M^f \rightarrow M$ is a semi conjugacy between $f$ and $\tilde{f}.$ Given any periodic point $p$ for an endomorphism $f$ we denote by $\bar{p}$ the unique point such that $\pi(\bar{p}) = p$ and $\bar{p}$ is periodic for $\tilde{f}.$
For any two hyperbolic periodic points $p$ and $q$ we say that $[p, q] \neq \emptyset$ iff $ W^u(\bar{p}) \pitchfork W_{loc}^s(\mathcal{O}(q)) \neq \emptyset.$ If $z \in W^u(\bar{p}) \pitchfork W_{loc}^s(\mathcal{O}(q))$ then $T_z W^u(\bar{p}) \oplus T_z(W_{loc}^s(\mathcal{O}(q))) = T_z(M).$ See Section \[stable-unstable\] for more precise definitions.
To give a concrete bound for the number of SRB measures we define [*skeleton*]{} inside hyperbolic periodic points of a fixed stable index.
A $k-$skeleton $( 0 < k < n=\dim(M))$ of $f$ is a subset of hyperbolic periodic points $\{p_i\}_{i \in \mathcal{I}}$ of stable index $k$ such that:
- For any hyperbolic periodic point $p \in M$ of index $k$, there is $i \in \mathcal{I}$ such that either $[p, p_i] \neq \emptyset$ or $[p_i, p] \neq \emptyset.$
- For every $i \neq j, [p_i, p_j] = \emptyset.$
Let us denote by $\mathcal{E}_k,$ the maximal cardinality of $k-$skeletons inside $Per_k(f)$ (hyperbolic periodic points of stable index $k$).
\[main\] Let $f: M \rightarrow M$ be a $C^2-$endomorphism of a closed $n-$dimensional manifold. Then for any $ 0 < k < n$ $$\sharp \{ \text{Ergodic hyperbolic SRB measures of index k} \} \leq \mathcal{E}_k.$$
The idea of using hyperbolic periodic points to analyze the number of SRB measures appear in [@13] and [@VY1]. In the context of partially hyperbolic diffeomorphisms with mostly contracting center, Viana and Yang [@VY] exhibited skeleton (defined by them with some similar properties) determining the number of basins of physical measures and concluded continuity results about the number of physical measures. In this paper we are not assuming any partial hyperbolicity assumption and invertibility of dynamics.
Although the upper bound in the above theorem may be far from the number of SRB measures for a general endomorphism, in some cases we can obtain sharp number of SRB measures. For instance in the case of Kan example using the proof of the above theorem we conclude that there are at most two SRB measures (which is a known fact), See \[kan\]. Indeed, in the proof of theorem \[main\] we correspond to each ergodic hyperbolic SRB measure $\mu$ of index $k$ a hyperbolic periodic point $P_{\mu}$ in $Per_k$ whose [*ergodic homoclinic class*]{} has full measure. Then, the key point is that given any two hyperbolic SRB measures $\mu$ and $\nu$ of index $k,$ if for the corresponding periodic points $P_{\mu}, P_{\nu}$ one of the conditions: $[P_{\mu}, P_{\nu}] \neq \emptyset$ or $[P_{\nu}, P_{\mu}] \neq \emptyset$ is satisfied, then $\mu=\nu.$
We also mention a result of Hirayama-Sumi [@HS] where they prove the ergodicity of hyperbolic smooth (SRB) measures under the condition of constancy of the dimension of unstable bundle and intersection property of stable and unstable manifolds of almost every pair of regular points. We emphasize that all the referred previous known results have been proved in the setting of invertible dynamical systems.
Preliminaries on Endomorphisms {#preliminaries}
==============================
Let $M$ be a closed Riemannian surface. By a $C^2-$endomorphisms $f:M\rightarrow M$ we mean a local $C^{2}-$diffeomorphism and $\mathcal{M}_{f}(M)$ denotes the set of all $f-$invariant Borel probability measures. Note that $f$ satisfies the following integrability condition. $$\log \vert det\,d_{x}f\vert\in \mathcal{L}^{1}(M,\mu).$$
For such $f$, consider the compact metric space $$M^{f}:=\{\tilde{x}=(x_{n})\in\prod_{-\infty}^{\infty}M : f(x_{n})=x_{n+1} \quad \text{for all} \quad n\in \mathbb{Z}\},$$ equipped with the distance $\tilde{d},$ between $\tilde{x}=(x_{n})$ and $\tilde{y}=(y_{n})\in M^{f}$ defined by $$\tilde{d}(\tilde{x},\tilde{y}):=\sum_{n=-\infty}^{\infty}2^{-|n|}d(x_{n}, y_{n}).$$ where $d$ is the distance on $M$ induced by the Riemannian metric. Let $\pi$ be the natural projection from $M^{f}$ to $M$ i.e, $\pi((x_{n}))=x_{0}, \forall \tilde{x}\in M^{f}$ and $\tilde{f}:M^{f}\rightarrow M^{f}$ be the shift homeomorphism. It is clear that $\pi\circ \tilde{f}=f\circ \pi.$ The map $\tilde{f}: M^{f}\rightarrow M^{f}$ is called the *Inverse Limit* of $f$ or the *Natural Extension* of the system $(M,f)$ and $M^{f}$ is the *Inverse Limit Space*.
Any periodic point $p,$ i.e $f^n(p)=p$ has an special pre-image (under $\pi$) in $M^f,$ $\bar{p} = (\cdots, p, f(p), \cdots, f^{n-1}(p), \cdots)$ which is $\tilde{f}-$periodic. We work with this special pre-image in many instances.
The map $\pi$ induces a continuous map from $\mathcal{M}_{\tilde{f}}(M^{f})$ to $\mathcal{M}_{f}(M),$ usually denoted by $\pi_{*}$ i.e. for any $\tilde{f}-$invariant Borel probability measures $\tilde{\mu}$ on $M^{f}$, $\pi_*$ maps it to an $f-$invariant Borel probability measure $\pi_{*}\tilde{\mu}$ on $M$ defined as $$\pi_{*}\tilde{\mu}(\phi)=\tilde{\mu}(\phi\circ \pi),\,\,\,\,\,\,\forall \phi \in C(M).$$ The following proposition I.3.1 of [@10] guarantees that $\pi$ is a bijection between $\mathcal{M}_{\tilde{f}}(M^{f})$ and $\mathcal{M}_{f}(M).$
\[1\] Let $f$ be a continuous map on $M$. For any $f-$invariant Borel probability measure $\mu$ on $M$, there exists a unique $\tilde{f}-$invariant Borel probability measure $\tilde{\mu}$ on $M^{f}$ such that $\pi_{*} \tilde{\mu}=\mu$. Moreover, $\mu$ is ergodic if and only if $\tilde{\mu}$ is ergodic.
The above proposition is standard and we just recall the proof of correspondence between ergodic measures. Consider the following diagram which permutes $\tilde f$ and $f$. **$$\begin{CD}
M^f @>\tilde f>> M^f\\
@VV\pi V @VV\pi V\\
M @>f>> M
\end{CD}$$** Suppose $\tilde{\mu}$ is ergodic. For each $f$-invariant subset $A\subset M$ i.e, $f^ {-1}(A)=A$, we can easily observe that $\pi^{-1}(A)$ satisfies $\tilde{f}^{-1}(\pi^{-1}( A)) = \pi^{-1}( A)$ and by ergodicity of $\tilde{\mu}$ then $\tilde{\mu}(\pi^{-1}(A))=\pi_{*}\tilde{\mu}=\mu(A)$ is either zero or one. Now let prove the reciprocal claim. Consider $\mathcal{\tilde{B}}_n := \tilde{f}^n (\pi^{-1}\mathcal{B})$ where $\mathcal{\tilde{B}}$ is the Borel $\sigma-$algebra of $M.$ It is easy to see that $(M, \mathcal{B}, \mu, f)$ is isomorphic to $(\tilde{M}, \mathcal{\tilde{B}}_n, \tilde{\mu}, \tilde{f}).$ Observe that by a general statement for conditional expectations, for any $\tilde{\phi}\in L^{1}(\tilde{\mu})$ we have: $$E(\tilde{\phi} \circ \tilde{f}| \tilde{f}^{-1}(\mathcal{\tilde{B}}_n)) = E(\tilde{\phi}|\mathcal{\tilde{B}}_n) \circ \tilde{f}.$$ By invariance property of $\mathcal{\tilde{B}}_n, i.e, \tilde{f}^{-1}(\mathcal{\tilde{B}}_n) = \mathcal{\tilde{B}}_n$ we conclude that $$E(\tilde{\phi} \circ \tilde{f}| \mathcal{\tilde{B}}_n) = E(\tilde{\phi}| \mathcal{\tilde{B}}_n) \circ \tilde{f}.$$ Now, take any $\tilde{\phi} \in L^1(\tilde{\mu})$ which is $\tilde{f}-$invariant. By the above relation we have that $E(\tilde{\phi} | \mathcal{\tilde{B}}_n)$ is $\tilde{f}-$invariant. The $E(\tilde{\phi} | \mathcal{\tilde{B}}_n)$ can be considered as $\mathcal{B}$ measurable by isomorphism and ergodicity of $\mu$ implies that $E(\tilde{\phi} | \mathcal{\tilde{B}}_n)$ is constant. Finally $\mathcal{\tilde{B}}_n$ converge to the Borel $\sigma-$algebra of $\tilde{M}$ and this implies that $\tilde{\phi} = \lim E(\tilde{\phi} |\mathcal{\tilde{B}}_n )$ is an almost everywhere constant function.
Multiplicative Ergodic Theorem on Natural Extension
---------------------------------------------------
Let $\mu$ be an $f-$invariant Borel probability measure on $M$. We denote by $\tilde{\mu}$ the $\tilde f-$invariant Borel probability measure on $M^{f}$ such that $\pi_{*}\tilde{\mu}=\mu$. There exists a full measure subset $\tilde{\mathcal{R}}$ called set of [*regular points*]{} such that for all $\tilde{x}=(x_{n})\in \tilde{\mathcal{R}}$ and $n\in \mathbb{Z}$ the tangent space $T_{x_{n}}M$ splits into a direct sum $$T_{x_{n}}M=E_{1}(\tilde{x}, n)\oplus\cdots\oplus E_{r(x_{0})}(\tilde{x}, n)$$ and there exists $-\infty<\lambda_{1}(\tilde x)<\cdots<\lambda_{r(\tilde x)}<\infty$ and $m_{i}(\tilde x)$ ($i=0,1,...,r(\tilde{x})$) such that:
1. dim $E_{i}(\tilde{x}, n)=m_{i}(\tilde x)$;
2. $D_{x_{n}}f(E_{i}(\tilde{x}, n))= E_{i}(\tilde{x}, n+1)$, and $D_{x_{n}}f|_{E_{i}(\tilde{x},n)}:E_{i}(\tilde{x}, n)\rightarrow E_{i}(\tilde{x}, n+1)$ is an isomorphism. For $v\in E_{i}(\tilde{x}, n)\backslash \{0\}$,
$\begin{cases}
\lim_{m\rightarrow\infty}\frac{1}{m}\log\Vert D_{x_{n}}f^{m}(v)\Vert=\lambda_{i}(\tilde{x});\\
\lim_{m\rightarrow\infty}-\frac{1}{m}\log\Vert(D_{x_{n- m}}f^{m}|_{E_{i}(\tilde{x},n-m)})^{-1}(v)\Vert=\lambda_{i}(\tilde{x});
\end{cases}$
3. if $i\neq j$ then $$\lim_{n\rightarrow\pm\infty} \frac{1}{n}\log\sin\angle(E_{i}(\tilde{x}, n), E_{j}(\tilde{x}, n))=0,$$ where $\angle(V, W)$ denotes the angle between sub-spaces $V$ and $W$.
4. $r(.), \,\lambda_{i}(.)\,\,and\,\,m_{i}(.)$ are measurable and $\tilde f-$invariant. Moreover $r(\tilde x)=r(x_{0}),\,\lambda_{i}(\tilde x)=\lambda_{i}(x_{0})$ and $m_{i}(\tilde{x})=m_{i}(x_{0})$ for all $i=1,2,...,r(\tilde x)$.
From now on we work with ergodic measures and the Lyapunov exponents are constant almost everywhere with respect to the reference measure. The celebrated Pesin’s blocks are defined naturally in the non-invertible case in the limit inverse space. We use a simple definition which is enough for our purpose. Let $\mu$ be an ergodic invariant measure and $\lambda$ (resp. $\theta$) the least in modulus positive (resp. negative) Lyapunov exponent. Suppose that $\mu$ has $k$ negative Lyapunov exponents.
\[1-08\] Fix $0 < \epsilon \ll 1.$ For any $l>1$, we define a Pesin block $\tilde{\Delta}_{l}$ of $M^{f}$ consisting of $\tilde{x}=(x_{n})\in M^{f}$ for which there exists a sequence of splittings $T_{x_{n}}M=E^{s}(\tilde{x}, n)\oplus E^{u}(\tilde{x}, n)$, $n\in \ZZ$, satisfying:
- $\dim E^{s}(\tilde{x}, n)=k$ ;
- $D_{x_{n}}f(E^{s}(\tilde{x},n))= E^{s}(\tilde{x}, n+1),$ $D_{x_{n}}f(E^{u}(\tilde{x}, n))=E^{u}(\tilde{x}, n+1)$;
- for $m\geq 0, \, v\in E^{s}(\tilde{x}, n)$ and $w\in E^{u}(\tilde{x}, n);$\
$\begin{cases}
\Vert D_{x_{n}}f^{m}(v)\Vert\leq e^{l} e^{-(\theta-\epsilon)m } e^{(\epsilon |n|)}\Vert v \Vert,\forall n\in\ZZ,n\geq 1\\
\Vert (D_{x_{n-m}}f^{m}|_{E^{u}(\tilde{x},n-m)})^{-1}(w) \Vert\leq e^{l}e^{-(\lambda-\epsilon ) m} e^{(\epsilon |n-m|)}\Vert w \Vert,\forall n\in\ZZ,n\geq 1;
\end{cases}$\
- sin $\angle(E^{s}(\tilde{x}, n),$ $E^{u}(\tilde{x}, n))\geq e^{-l}e^{-\epsilon |n|}$.
In the above definition it is enough to take $\epsilon$ less than $\frac{1}{2} \min \{ \lambda, \theta \}.$
Pesin blocks are compact subsets of $M^{f}$ where the subspaces $E^{s}(\tilde{x}, n)$ and $E^{u}(\tilde{x}, n)$ of $T_{x_{n}}M$ depend continuously on $\tilde{x}$ and $\tilde{f}^{\pm}(\tilde{\Delta}_{l})\subset\tilde{\Delta}_{l+1}$.
Stable, Unstable Sets, SRB Property {#stable-unstable}
===================================
After the works of Pesin on general theory of stable and unstable manifolds for non-uniformly hyperbolic diffeomorphisms (see [@7]), P.-D Liu and M. Qian [@36] developed a rigorous related theory for random diffeomorphisms. Using similar techniques, Sh. Zhu proved an unstable manifold theorem for non-invertible differentiable maps of finite dimension [@21] (see [@10] for more details.) Here we would like to emphasize the differences between unstable and stable sets (and manifolds).
\[1-16\]
Let $\tilde x\in\tilde{\mathcal{R}}$ and $\lambda$ the least positive Lyapunov exponent of $\mu$. We call $W_{loc}^{u}(\tilde x)$ a *local unstable manifold* of $f$ at $\tilde x$ when exists a $u-$dimensional $C^2-$embedded sub-manifold of $M$ ($u$ is the number of positive Lyapunov exponents.) such that there are $\epsilon, C >0$, and for any $y_{0}\in W^{u}_{loc}(\x)$, there exists a unique $\tilde{y}=\{y_{n}\}_{n\in \mathbb{Z}}\in M^{f}$ such that $\pi(\tilde y)=y_0$ and $\forall n\in \mathbb{N},$ $$d(y_{-n},x_{-n})\leq C\,e^{-n(\lambda-\epsilon)}\,d(x_0,y_0)$$ Moreover we define the **local unstable set** of $\tilde{f}$ at $\tilde x=(x_n)$ as $$\widetilde{W}^{u}_{loc}(\tilde x):=\{\tilde y \in M^{f}: y_0\in W^{u}_{loc}(\tilde x),d(y_{-n},x_{-n})\leq C\,e^{-n(\lambda-\epsilon)}\,d(x_0,y_0)\}.$$
It comes out that $\pi(\widetilde{W^{u}_{loc}}(\tilde{x}))=W^{u}_{loc}(\tilde{x})$ is the local unstable manifold of $\x.$
\[1-17\] The unstable manifold of $f$ corresponding to $\tilde x\in \tilde{\mathcal{R}}$ is defined as
$$\label{1-10}
W^{u}(\tilde x)=\{y_0\in M|\,\exists\tilde y\in M^{f}\,\,with\,\,\pi\tilde y=y_0,\,\,and\,\,\overline\lim_{n\rightarrow+\infty}\,\frac{1}{n}log\,d(x_{-n},y_{-n})<0\}\}$$
and we will write
$$\label{1-11}
\widetilde{W}^{u}(\tilde x)=\{\tilde y\in M^{f}|\overline\lim_{n\rightarrow+\infty}\,\frac{1}{n}log\,d(x_{-n},y_{-n})<0\}.$$
Notice that $\pi(\widetilde{W}^{u}(\tilde x))= W^{u}(\tilde x)$ and the global unstable manifold $W^{u}(\tilde x)$ is the union of forward images of local unstable manifolds at $x_{-n}.$
\[Fig:US\] {width="0.9\linewidth"}
{width="1\linewidth"}
Observe that the global unstable set is not necessarily a manifold and it is defined for $x \in M$ and not $\x \in M^f.$
**Local Stable Manifolds:** Besides the case of unstable manifolds, the local stable manifolds are defined uniquely for $x \in M$, when they exist. In fact $ W^s_{loc}(\x)$ is defined exactly as in the case of invertible dynamics. So, we may use the notation $W^s_{loc}(x)$ or $W^s_{loc}(\x)$ for the same object. In this text, we use $W^s_{loc}(\x).$ However, the global stable [**set**]{} is defined as follow: $$W^{s}(x)=\bigcup^{+\infty}_{n=0}f^{-n}(W^s_{loc}({f^n(x)}))$$ Observe that the global stable set is not necessarily a connected [**manifold**]{} (see figure 1).
Using definitions and local stable-unstable manifolds theorems from [@10], one concludes the following invariance properties :
- $f(W^{s}(\x))=W^{s}(\tilde{f}(\x))$;
- $f(W^{u}(\x))=W^{u}(\tilde{f}(\x)).$
Heteroclinic Relation and Incliniation Lemma {#homoclinici}
--------------------------------------------
Let $p, q$ be two hyperbolic periodic points. We write $[p, q] \neq \emptyset$ iff $W^u(\bar{p}) \pitchfork W^s_{loc}(\mathcal{O}(q)) \neq \emptyset.$
In what follows we recall the well known inclination lemma. In typical texts in dynamics, the inclination lemma is proved for invertible dynamics. However, there is a small subtle difference between the invertible and non-invertible case.
\[4-7\] Let $p$ be a hyperbolic fixed point of map $f$ and $D \subset W^u (\bar p)$ be a compact disk. If $\Sigma$ is an embedded $C^1$ sub-manifold of $M$ intersecting $W^s_{loc}(p)$ transversally, then for any large $n,$ $f^n(\Sigma)$ contains an embedded manifold $\Sigma_n$ which is $C^1-$close to $D$ where $\bar p=(...pppp...)\in M^f$.
The proof is similar to the diffeomorphisms case. We just observe that $E^{u}(\p):=\bigcap_{k=0}^{\infty}Df_{p}^{k}(C^u(p))$ where $C^u(p)$ is the complement of a thin cone around $E^s(p).$ As for any point in a small neighborhood of $\Sigma \cap W^s_{loc}(p)$ the tangent space is outside the stable cone, the same argument of usual $\lambda-$lemma yields the proof.
SRB property
------------
In this subsection we review the definitions and basic properties of SRB measures.
\[def: unstable partition\] A measurable partition $\zeta$ of $M^f$ is said to be *subordinate to $W^u$ manifolds of $(f,\mu)$* if for $\tilde\mu$-a.e. $\tilde x$, $\zeta(\tilde x)$ has the following properties:
- $\pi_|\zeta(\tilde x): \zeta(\tilde x)\rightarrow \pi(\zeta(\tilde x))$ is bijective
- There exists a $k(\tilde x)$ dimensional $C^2-$embedded sub-manifold $W_{\tilde x}$ of $M$, such that $W_{\tilde x}\subset W^u(\tilde x),\,\,\pi(\zeta(\tilde x))\subset W_{\tilde x}$ and $\pi(\zeta(\tilde x))$ contains an open neighborhood of $x$ in $W_{\tilde x}$, this neighborhood being taken in the topology of $W_{\tilde x}$ as a sub-manifold of $M$.
It is always possible to construct measurable partitions sub-ordinated to unstable manifolds (see [@10], [@6].)
\[def: SRB property\] An $f-$invariant measure $\mu$ has the SRB property, if for every measurable partition $\zeta$ of ${M^f}$ sub-ordinate to ${W^u}-$manifolds of $(f,\mu)$ we have, for $\tilde \mu$-a.e. $\tilde x \in M^f$, $$\pi(\tilde {\mu}_{\tilde x}^{\zeta})\prec\prec m_{\tilde x}^{u},$$ where $\{\tilde {\mu}_{\tilde x}^{\zeta}\}_{\tilde x\in M^f}$ is a canonical system of conditional measures of $\tilde\mu$ associated with $\zeta$, $\pi(\tilde {\mu}_{\tilde x}^{\zeta})$ is the projection of $\tilde {\mu}_{\tilde x}^{\zeta}$ under $\pi_|\zeta(\tilde x): \zeta(\tilde x)\rightarrow \pi(\zeta(\tilde x))$ and $m_{\tilde x}^{u}$ denotes the Lebesgue measure on $W_{loc}^{u}(\tilde x)$ induced by its inherited Riemannian structure.
In the diffeomorphisms context, F. Ledrappier and L.-S. Young [@4] proved that a measure satisfies Pesin entropy formula, if and only if it is absolutely continuous with respect to (Pesin) unstable manifolds. Moreover, they showed that the densities $d\mu^{u}_{x}/dm^{u}_{x}$ are given by **strictly positive** functions that are $C^{1}$ along unstable manifolds. ($\mu^{u}_{x}$ is the unstable conditional measure and $m^{u}_{x}$ the induced Lebesgue inherited from Riemannian structure.)
The same results (adapted to the non-invertible case) hold for endomorphisms. This implies that $\pi(\tilde {\mu}_{\tilde x}^{\zeta})\prec\prec m_{\tilde x}^{u}$ in the above definition can be substituted by $\pi(\tilde {\mu}_{\tilde x}^{\zeta}) \approx m_{\tilde x}^{u}$ for $\tilde{\mu}$ almost every $\tilde{x}.$
Admissible Manifolds and Katok Closing Lemma {#4}
============================================
In this section we try to give simplified definition of admissible manifolds which are tools in the proof of the main theorem. To define admissible manifolds we recall the definition of Lyapunov charts in the non-invertible case of Pesin theory.
Let $f$ be a $C^{2}-$endomorphism of a closed Riemannian surface $M$. Assume that we have a non-empty Pesin block $\tilde{\Delta}_{l}$ for $l>1$ for an ergodic measure with $k-$negative Lyapunov exponents and $(n-k)-$positive Lyapunov exponents. We can change the metric on $\tilde{\Delta}_{l}$ so that $f|_{\tilde{\Delta}_{l}}$“looks uniformly hyperbolic". This happens by replacing the induced Riemannian metric on $T_{x_{n}}M$ ($\tilde{x}=(x_{n})\in \tilde{\Delta}_{l}$) by a new metric which is called Lyapunov metric.
Let $0<\lambda^ {'}<\mu^{'}<\infty$ such that $$\label{2-00}
\lambda^{'}=\lambda-2\epsilon, \quad \mu^{'}=\mu-\, 2\epsilon.$$ We can define the new metric $<.,.>_{\tilde{x}}^{\prime}$ as: $$\label{2-1-00}
< v_{s},w_{s}>_{\tilde{x}}^{\prime}:=\sum_{m=0}^{\infty}< df^{m}(v_{s}),df^{m}(w_{s})>_{x_n}e^{2\lambda^{\prime}m};$$ where $v_s,w_s\in E^{s}(\tilde{x},0)$, and $$\label{2-000}
< v_{u},w_{u}>_{\tilde{x}}^{\prime}:=\sum_{m=0}^{\infty}< df^{-m}(v_{u}),df^{-m}(w_{u})>_{x_{-n}}e^{2\lambda^{\prime}m};$$ where $v_u,w_u\in E^{u}(\tilde{x},0)$. Now for $(v,w)\in T_{x_{0}}M$ that $v=v_s+v_u,w=w_s+w_u$, define $$\label{Lyp metric}
<v,w>_{\tilde{x}}^{\prime}:= \max\{< v_{s},w_{s}>_{\tilde{x}}^{\prime},< v_{u},w_{u}>_{\tilde{x}}^{\prime}\}.$$ The new metric induces a new norm $\Vert . \Vert^{\prime}_{\tilde x}$ on $T_{x}M$: $$\begin{aligned}
&\Vert v_{s}\Vert_{\tilde{x}}^{\prime}=(\sum_{m=0}^{\infty}e^{2\lambda^{\prime}m}\Vert d_{x}f^{m}(v_{s})\Vert^{2})^{1/2}
,\\
&\Vert v_{u}\Vert_{\tilde{x}}^{\prime}=(\sum_{m=0}^{\infty}e^{\mu^{\prime}m}\Vert(d_{x}f^{m}|_{E^{u}(\tilde{x},0)})^{-1}(v_{u})\Vert^{2})^{1/2}
,\\
&\Vert v\Vert_{\tilde{x}}^{\prime}=\max\{\Vert v_{s}\Vert^{\prime}_{\tilde x},\, \Vert v_{u}\Vert_{\tilde{x}}^{\prime}\}.\end{aligned}$$
One can verify that for $v_{s}\in E^{s}(\tilde x,0),\,v_{u}\in E^{u}(\tilde x,0)$: $$\Vert d_{x_{0}}f(v_{s})\Vert^{\prime}_{\tilde{f}(\tilde x)}\leq e^{-\lambda^{'}}\Vert v_{s} \Vert^{\prime}_{\tilde x}
,$$ $$\Vert d_{x_{0}}f(v_{u})\Vert^{\prime}_{\tilde{f}(\tilde x)}\geq e^{\mu^{'}}\Vert v_{u} \Vert^{\prime}_{\tilde x}.$$
There exist also the following estimate on the norms ($\Vert.\Vert$ is the induced Riemannian norm on $T_{x}M$). $$\label{2-1}
\frac{1}{2}\Vert v \Vert\leq\Vert v \Vert_{\tilde{x}}^{\prime}\leq \textit{a}_{l} \Vert v \Vert\,\,\,\,\,\,\,\,\,\,\forall\tilde{x}=(x_{n})\in\tilde{\Delta}_{l}$$
Usually $\Vert . \Vert^{\prime}_{\tilde{x}}$ is called **Lyapunov norm**. The following proposition is about the existence of Lyapunov charts in the context of local diffeomorphisms and its proof is a simple adaptation of proposition (2.3) of [@8].
\[2-01\] There exists a number $r>0$ so that for every point $\tilde{x}\in \tilde{\Delta_{l}}$ we can find a neighborhood $B(\tilde{x})$ around the point $x=\pi(\tilde{x})$ and a diffeomorphism $\Phi_{\tilde{x}}:B^k_{r}\times B^{\dim(M)-k}_{r}\rightarrow B(\tilde{x})$($B^{d}_{r}$ is Euclidean closed disc of radius $r$ around the origin in $\mathbb{R}^d$). Also there exists a family of $C^{1}-$maps $F_{\tilde{x}}: B^k_{r}\times B^{\dim(M)-k}_{r}\rightarrow \mathbb{R}^k\times \mathbb{R}^{\dim(M)-k}$ satisfying the following properties:
1. $\Phi_{\tilde{x}}(0)=\pi(\tilde{x});\,\,\,\,\, $
2. $F_{\tilde{x}}(z)=\Phi_{\tilde{f}(\tilde{x})}^{-1}\circ f\circ\Phi_{\tilde{x}}(z)\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ $ $z=(u,v);$
3. $F_{\tilde{x}}$ has the form: $$F_{\tilde{x}}(u,v)=(A_{\tilde{x}}\,u+h^{1}_{\tilde{x}}(u,v),\,B_{\tilde{x}}\,v+h^{2}_{\tilde{x}}(u,v)),$$ such that: $$h^{2}_{\tilde{x}}(0,0)=h^{2}_{\tilde{x}}(0,0)=0,\,\,\,dh^{1}_{\tilde{x}}(0,0)=dh^{2}_{\tilde{x}}(0,0)=0$$\
and $$\Vert A_{\tilde{x}}\Vert\leq\,e^{- \lambda^{'}}, \,\,\,\,\,\,\,\,\,\,\Vert B_{\tilde{x}}\Vert\geq\, e^{\mu^{'}}.$$ (all the norms are considered as Euclidean.)
For $z\in B^k_{r}\times B^{\dim(M)-k}_{r}$ let $ h_{\tilde{x}}(z)=(h^{1}_{\tilde{x}}(z),h^{2}_{\tilde{x}}(z))$, then $$\Vert (dh_{\tilde{x}})_{z_{1}}-(dh_{\tilde{x}})_{z_{2}} \Vert\leq \Upsilon\,\textit{a}_{l}\, \Vert z_{1}-z_{2}\Vert$$ where $\Upsilon$ is an absolute constant.
4. the metric $\Vert . \Vert ^{\prime}_{\tilde{x}}$ depends continuously on $\tilde{x}$ over any Pesin block $\tilde{\Delta}_{l}$.
![unstable manifolds for $\tilde{x}_1\neq \tilde{x}_2$ such that $\pi(\tilde{x}_1)=\pi(\tilde{x}_2)=x$.[]{data-label="fig:Lyp chart"}](chart01.pdf){width="0.45\linewidth"}
The set of admissible manifolds is just the set of graph of $C^1-$functions defined locally and with bounded Lipschitz constant. The difference with the invertible setting is that they depend on $\x$ and not just $\pi(\x) = x.$
To be more precise, we define the class of $(\gamma, \delta, h)$ stable-admissible and $(\gamma, \delta, h)$ unstable-admissible manifolds close to $x$ as follow: $$S_{\tilde{x}}^{\gamma,\delta,h}=\{\Phi_{\tilde{x}}(graph\, \phi) | \phi\in C^{1}(B^{k}_{h},\,B^{n-k}_{h}),\Vert\phi(0)\Vert\leq\delta,\,\Vert d\,\phi\Vert\leq\gamma\}$$ $$\label{2-002}
U_{\tilde{x}}^{\gamma,\delta,h}=\{\Phi_{\tilde{x}}(graph\, \phi)| \phi\in C^{1}(B^{n-k}_{h},\,B^{k}_{h}),\Vert\phi(0)\Vert\leq\delta,\,\Vert d\,\phi\Vert\leq\gamma\},$$ where by $B^k_h, B^{n-k}_h$ we denote the ball of radius $h$ around origin respectively in the stable and unstable bundle. A simple transversality argument implies:
\[2-2\] Let $\tilde{x}\in \tilde{\Delta}_{l}$ and $h > 0$ small. Then for small constants $\gamma, \delta$ any stable-admissible manifold in $S_{\tilde{x}}^{\gamma,\delta,h}$ intersects any unstable-admissible manifold in $U_{\tilde{x}}^{\gamma,\delta,h}$ at exactly one point and the intersection is transversal. The constants can be chosen universal in a Pesin block.
Katok Closing Lemma for Endomorphisms
-------------------------------------
Let $\mu$ be an ergodic invariant measure for an endomorphism $f: M \rightarrow M$ with non-zero Lyapunov exponents. The following lemma is an endomorphism version of Katok closing lemma for diffeomorphisms. We emphasize that in our context we deal with local diffeomorphisms and we use Lyapunov charts which depend on trajectories. So, a messy adaptation of the same proof of Katok implies the following closing lemma.
(Katok Closing Lemma)\[2-7\] Let $f$ be a $C^2-$endomorphism of a compact Riemannian surface $M$. For any positive numbers $l,\delta$ there exists a number $\varrho=\varrho(l,\delta)>0$ such that if for some point $\tilde{x}\in \tilde{\Delta}_{l}$ (Pesin block) and some integer $m$ one has $$\label{2-07}
\tilde{f}^{m}(\tilde{x})\in \tilde{\Delta}_{l}\,\,\,\,\,\,\,and\,\,\,\,\,\,\,\tilde{d}(\tilde{x},\tilde{f}^{m}(\tilde{x}))<\varrho,$$ then there exists a point $z\in M$ and $\bar z\in M^{f}$ such that $z=\pi(\bar z)$ and
- $f^{m}( z)= z$ and $\tilde{f}^{m}(\bar z)=\bar z$;
- $d_{n}^{f}(x,z)<\delta$;($\,\,\,d_{n}^{f}$ is defined as $d_{n}^{f}(x,z)= \max_{0\leq i\leq n-1}\,d(f^{i}x,f^{i}z).$
- the point $z$ is a hyperbolic periodic point for $f$ and its $W^{s}_{loc}(z)$ and $W^{u}_{loc}(\bar z)$ are admissible manifolds close to $x$, in the chart of $\x$.
Ergodic Homoclinic Classes and Proof of Theorems
================================================
The notion of ergodic homoclinic classes comes from the work of [@13]. The authors defined this notion proving the (at most) uniqueness of SRB measures for surface transitive diffeomorphisms. Here we define a similar notion for endomorphisms.
Ergodic Homoclinic Classes. {#4-0}
---------------------------
Let $p \in M$ be a hyperbolic periodic point with period $n$. We define the *Ergodic Homoclinic Class* of $p$ both in limit inverse and in the manifold $M.$ Recall that $\bar{p}$ is the unique periodic point of $\tilde{f}$ such that $\pi(\bar{p}) =p.$ That is, $\bar{p}= (\cdots, p, f(p), \cdots f^{n-1}(p), p, \cdots).$
The inverse limit Ergodic Homoclinic Class ($\widetilde{EHC}$) is defined as $\tilde{\Lambda}(\bar{p}):=\tilde{\Lambda}^{s}(\bar{p})\cap\tilde{\Lambda}^{u}(\bar{p})$ where,
$$\tilde{\Lambda}^s(\bar{p}):=\{\x \in \tilde{\mathcal{R}}\big{|} \exists n \geq 0 , W_{loc}^{s}(\tilde{f}^n(\x)) \pitchfork W^{u}(\mathcal{O}(\bar{p}))\neq \emptyset \}$$
and $$\tilde{\Lambda}^u(\bar{p}):=\{ \x \in\tilde{\mathcal{R}}\big{|} \exists \tilde{y} \in \tilde{\mathcal{R}}, \pi(\tilde{y}) = \pi(\x), \exists n \geq 0, \, \, f^n(W^{u}(\y)) \pitchfork W_{loc}^{s}(\mathcal{O}(\bar{p}))\neq \emptyset\}$$ Here $\tilde{\mathcal{R}}$ denotes regular points in $M^f.$ Observe that $\pi^{-1}(\pi (\tilde{\Lambda}^{*}(\bar{p}))) = \tilde{\Lambda}^{*}(\bar{p}) $ for $* \in \{s, u\}.$ We denote by $\Lambda^s(p) := \pi(\tilde{\Lambda}^s(\bar{p}))$, $\Lambda^u(p) := \pi(\tilde{\Lambda}^u(\bar{p})) $ and $ \Lambda(p) := \pi(\tilde{\Lambda}(\bar{p}))$.
Suppose that $\mu$ is a hyperbolic $f-$invariant Borel probability measure. Take $\x \in \tilde{\Delta}_l$ a recurrent point in the support of $\tilde{\mu}$ restricted to the Pesin block $\tilde{\Delta}_l$. Using closing lemma \[2-7\] we find a hyperbolic periodic point $\bar{p}$ and we prove two following crucial lemmas about the ergodic homoclinic class of $\bar{p}.$
\[4-5\] Let $p$ be a periodic point obtained as above, then $\tilde{\mu}(\tilde{\Lambda}(\bar p))>0.$
Let $\tilde{B}$ be a small ball around $\x$ such that $\tilde{\mu}(\tilde{B} \cap \tilde{\Delta}_l) > 0$ . By the item (c) in the closing lemma \[2-7\], $W^u_{loc}(\bar{p})$ and $W^s_{loc}(\bar{p})$ are respectively close to $W^u_{loc}(\tilde{x})$ and $W^s_{loc}(\x).$ By continuity of stable and unstable manifolds in the Pesin blocks, for any point $\tilde{y}\in \tilde{B} \cap \tilde{\Delta}_l$ we have that $W^u_{loc}(\y)$ and $W^s_{loc}(\y)$ are respectively close to $W^u_{loc}(\tilde{x})$ and $W^s_{loc}(\x)$ in the $C^1-$topology. Consequently by transversality arguments we conclude that $\tilde{y} \in \tilde{\Lambda}(\bar p).$
\[4-6\] $\tilde{\Lambda}(\bar p)$ is $\tilde{f}-$invariant.
In fact we prove that both $\tilde{\Lambda}^s(\bar{p})$ and $\tilde{\Lambda}^u(\bar{p})$ are invariant. Firstly let us prove that $\tilde{f}(\tilde{\Lambda}(\bar p))\subset \tilde{\Lambda}(\bar p)$. Recall that $\tilde{\Lambda}(\bar p)= \tilde{\Lambda}^{s}(\bar p)\cap\tilde{\Lambda}^{u}(\bar p).$ Without loss of generality let suppose that $p$ is a hyperbolic fixed point, then: $$\begin{aligned}
\tilde{x}\in \tilde{\Lambda}^{u}(\bar p)&\Rightarrow f^n (W^{u}(\tilde{y}))\pitchfork W_{loc}^{s}(\bar p)\neq\emptyset, \pi(\tilde{y}) = \pi(\x)\\
&\Rightarrow f^{n+1}(W^{u}(\tilde{y}))\pitchfork W_{loc}^{s}(\bar p)\neq\emptyset\\
&\Rightarrow f^n(W^{u}(\tilde{f}(\tilde{y}))\pitchfork W_{loc}^{s}(\bar p)\neq\emptyset\\
&\Rightarrow \tilde{f}(\tilde{x})\in \tilde{\Lambda}^{u}(\bar p).\end{aligned}$$
On the other hand, if $\tilde{x}\in \tilde{\Lambda}^{s}(\bar p) \Rightarrow W_{loc}^{s}(\tilde{f}^n(\x))\pitchfork W^{u}(\bar p)\neq\emptyset$ for some $n \geq 0.$ This implies:
$\begin{cases}
W_{loc}^{s}(\tilde{f}^{n-1} (\tilde{f}(\x))))\pitchfork W^{u}(\bar p)\neq\emptyset, \quad \text{if} \quad n > 0;\\
W^s_{loc} (\tilde{f}(\x)) \pitchfork W^{u}(\bar p)\neq\emptyset \quad \text{if} \quad n = 0
\end{cases}$
So, $\tilde{f}(\tilde{x})\in \tilde{\Lambda}^{s}(\bar p).$
Now we prove that $\tilde{\Lambda}(\bar p)\subset\tilde{f}(\tilde{\Lambda}(\bar p))$. We divide the proof into two steps again:
- $\tilde{\Lambda}^{u}(\bar{p})\subset \tilde{f}(\tilde{\Lambda}^{u}(\bar{p}))$: Take $\tilde{x}=\tilde{f}(\tilde{f}^{-1}(\tilde{x})) \in \tilde{\Lambda}^{u}(\bar{p})$. So by definition there exist $\tilde{y}, \pi(\tilde{y})= \pi(\x)$ and $n \geq 0$ such that $ W^{u}(\tilde{f}^n(\y))) \pitchfork W_{loc}^{s}(\bar{p})\neq \emptyset$ which implies that (by applying $f$)
$$W^{u}(\tilde{f}^{n+1}(\tilde{f}^{-1}(\y)))) \pitchfork W_{loc}^s(\bar{p})\neq \emptyset \Rightarrow \tilde{f}^{-1}(\x) \in \tilde{\Lambda}^{u}(\bar{p}).$$
- $\tilde{\Lambda}^{s}(\bar p)\subset \tilde{f}(\tilde{\Lambda}^{s}(\bar{p}))$: Once again taking $\tilde{x}=\tilde{f}(\tilde{f}^{-1}(\tilde{x}))$; by definition of $\tilde{\Lambda}^{s}(\bar p)$ we have:
$$W^{s}(\tilde{f}^{n+1}(\tilde{f}^{-1}(\x))) \pitchfork W^{u}(\bar{p}) = \emptyset$$
which implies $\tilde{f}^{-1}(\x) \in \tilde{\Lambda}^s(\bar{p})$.
Ergodic Criterion.
------------------
In this subsection we give the most important part of the proof and from now on $\mu$ is a measure with SRB property.
\[4-10\] Let $\tilde f: M^f\to M^f$ be the lift homeomorphism on inverse limit space of $f$ and $\tilde \mu$ the unique $\tilde f-$invariant lift measure of a Borel probability $f-$invariant measure $\mu$ on $M$. Let $\tilde{\phi} \in C(M^{f})$ a continuous function on $M^{f}$. For $\tilde \mu$ almost every point $\tilde x \in M^f$ the following two limits exist $$\tilde{\phi}^{\pm}(\tilde x)=\lim_{n \rightarrow \pm\infty} \frac{1}{n}\sum_{j=0}^{n-1} \tilde{\phi}(\tilde{f}^{j}(\tilde x)).$$
Both $\tilde{\phi}^{+}$ and $\tilde{\phi}^{-}$ are $\tilde \mu-$integrable $\tilde{f}-$invariant function with $\int \tilde{\phi}^{\pm} d \tilde{\mu} = \int \tilde{\phi} d\tilde{\mu}.$ In particular if $\tilde{\mu}$ is ergodic then $\tilde{\phi}^{\pm}$ are constant functions.
It comes out from the proof of Birkhoff ergodic theorem that for $\tilde \mu-$a.e $\x\in M^f$, $\tilde{\phi}^{+}(\tilde x)=\tilde{\phi}^{-}(\tilde x)$. Such points are called **typical points**.
\[4-11\] There exists an invariant set $\tilde{S}$ of typical points with $\tilde{\mu}(\tilde{S})=1$ such that for all $\tilde{\phi}\in C(M^f)$, if $\tilde x\in\tilde{S}$ then for all $\tilde{\omega}\in \widetilde{W^{s}}(\tilde x)$ and $\tilde{\mu}^{u}_{\tilde x}-$a.e $\tilde{\zeta}\in \widetilde{W^{u}}(\tilde x)$, $$\tilde{\phi}^{+}(\tilde x)= \tilde{\phi}^{+}(\tilde{\zeta}) = \tilde{\phi}^{+}(\tilde{\omega}).$$
We claim that for almost all typical points $\x$, we have $\tilde{\mu}^{u}_{\tilde x}-$a.e $\tilde{\zeta}\in\widetilde{W^{u}_{loc}}(\tilde x)$ is typical. The proof of this claim is mutatis mutandis to the lemma 3.1 in [@13]. Indeed, if there exists a subset of $\tilde{\mu}-$positive measure which does not satisfy the above claim, then using the definition of conditional measures we get a contradiction to the fact that typical points has full $\tilde{\mu}-$measure.
We take $\tilde{S}$ as the full $\tilde{\mu}-$measure subset of points $\x$ obtained above.
Observe that by definition of typical points: $$\begin{aligned}
\label{4-11-1}
\bar{\phi}^{+}(\tilde{x})=\bar{\phi}^{-}(\tilde{x})\,\,\,\,\, and \,\,\,\,\, \bar{\phi}^{+}(\tilde{\zeta})=\bar{\phi}^{-}(\tilde{\zeta}),
\end{aligned}$$ and from continuity of $\tilde{\phi}$: $$\begin{aligned}
\label{4-11-2}
\tilde{\phi}^{-}(\tilde{\zeta})=\tilde{\phi}^{-}(\tilde{x}) \, \text{for all} \,\,\,\,\, \tilde{\zeta}\in\widetilde{W^{u}}(\tilde x).
\end{aligned}$$
Using \[4-11-2\], \[4-11-1\] we conclude that $\tilde{\phi}^{+}(\tilde x)=\tilde{\phi}^{+}(\tilde{\zeta})$.
Again by continuity of $\tilde{\phi}$ and using the definition of stable sets, we conclude that $\tilde{\phi}^{+}(\tilde x) = \tilde{\phi}^{+}(\tilde{\omega})$ for every $\tilde{\omega} \in \widetilde{W^{s}}(\tilde x).$
By the definition of ergodic sum it is clear that $\tilde{S}$ is an invariant set.
\[4-12\] Given $\tilde{\phi}\in \mathcal{L}^{1}(M^{f})$, there exists an invariant set $\tilde{S}_{\tilde{\phi}}\subset M^{f}$ with $\tilde{\mu}(\tilde{S}_{\tilde{\phi}})=1$ such that if $\tilde{x}\in \tilde{S}_{\tilde{\phi}}$ then $\tilde{\mu}^{u}_{\tilde{x}}-$a.e $\tilde{y}\in\widetilde{W}^{u}(\tilde{x})$ satisfies $$\tilde{\phi}^{+}(\tilde{y})=\tilde{\phi}^{+}(\tilde{x}).$$
Given $\tilde{\phi}\in\mathcal{L}^{1}(M^{f})$, as $C(M^{f})$ is dense in $\mathcal{L}^{1}(M^{f})$ therefore we can take a sequence of continuous functions $\tilde{\phi}_{n}$ converging to $\tilde{\phi}$ in $L^1-$topology. By Birkhoff Ergodic Theorem for natural extensions, $\tilde{\mu}-$a.e $\tilde{x}$, $\tilde{\phi}_{n}^{+}(\tilde{x})$ exists and $\tilde{\phi}_{n}^{+}$ converges to $\tilde{\phi}^{+}$ in $L^1-$topology.
As $M^{f}$ is a compact metric space, there exists a sub-sequence $\tilde{\phi}^{+}_{n_{k}}$ and a full $\tilde{\mu}$ measure subset $\tilde{J}$ such that for every $\tilde{x} \in \tilde{J}$ we have $\tilde{\phi}^{+}_{n_{k}} (\x)$ converge to $\tilde{\phi}^{+}(\tilde{x})$.
Now take $\tilde{S}_{\tilde{\phi}}:= \tilde{J} \cap \tilde{S}$ and the proof is complete.
\[4-13\] Let $f:M\rightarrow M$ be a $C^2-$endomorphism over a compact manifold $M$ equipped with a hyperbolic measure $\mu$ with SRB property. If $\tilde{\mu}(\tilde{\Lambda}(\bar p))>0$ then $$\tilde{\Lambda}^{u}(\bar p)\subset^{\circ}\tilde{\Lambda}^{s}(\bar p).$$
First remember that by definition $\tilde{\Lambda}(\bar p)=\tilde{\Lambda}^{u}(\bar p)\cap \tilde{\Lambda}^{s}(\bar p).$ By lemma \[4-12\] it is enough to prove that $ \tilde{\Lambda}^{u}(\bar p)\cap\tilde{S}_{1_{\tilde{\Lambda}^{s}}} \subset \tilde{\Lambda}^{s}(\bar p)$.\
The following two lemmas are useful in the proof.
\[4-13-1\] If exists a $\tilde{\mu}^{u}_{\tilde{x}}-$positive measure subset of $\widetilde{W}^{u}(\tilde{x})$ belonging to $\tilde{\Lambda}^{s}(\bar p)$, then $\tilde{x}\in \tilde{\Lambda}^{s}(\bar p)$.
The $\tilde{\Lambda}^{s}(\bar p)$ is $\tilde{f}-$invariant and $1_{\tilde{\Lambda}^{s}(\bar p)}=\tilde{f}(1_{\tilde{\Lambda}^{s}(\bar p)})$. This implies that if$\tilde{x}\notin \tilde{\Lambda}^{s}(\bar p)$ then $\tilde{\mu}^{u}_{\tilde{x}}-a.e. \tilde{y}\in \widetilde{W}^{u}_{loc}(\tilde{x})$ does not belong to $\tilde{\Lambda}^{s}(\bar p)$.
\[4-13-2\] If exists some $m\in \mathbb{N}$ such that a $\tilde{\mu}^{u}_{\tilde{f}^{m}(\tilde{x})}-$positive measure subset, of $\widetilde{W}^{u}(\tilde{f}^{m}(\tilde{x}))$ belong to $\tilde{\Lambda}^{s}(\bar p)$, then $\tilde{x}\in \tilde{\Lambda}^{s}(\bar p)$.
This comes from the fact that $\tilde{x}\in\tilde{S}_{1_{\tilde{\Lambda}^{s}}}$ and $\tilde{S}_{1_{\tilde{\Lambda}^{s}}}$ is an $\tilde{f}-$invariant set. The rest will be a corollary of last lemma.
Take $\tilde{y}\in\tilde{\Lambda}^{s}(\bar p)$ an auxiliary point in a way that for some $ l>0$ both $\tilde{x},\tilde{y}$ lies in the same Pesin block $\tilde{\Delta}_{l}$ and $\tilde{y}\in supp(\tilde{\mu}|_{\tilde{\Delta}_{l}\cap\tilde{\Lambda}^{s}(\bar p)})$. We additionally suppose that $\tilde{y}$ returns back to $\tilde{\Delta}_{l}\cap\tilde{\Lambda}^{s}(\bar p)$ infinitely many times.
As $\tilde{y}\in\tilde{\Lambda}^{s}(\bar p)$ then by definition there exists $n \geq 0$ such that $W_{loc}^{s}(\tilde{f}^n(\y))\pitchfork W^{u}(\bar p)\neq\emptyset$. Without loss of generality we suppose that $n=0.$ As $ W^{s}_{loc}(\y)\pitchfork W^{u}(\bar p)\neq\emptyset$ for large enough $n$ we have that $f^{n}(y)$ is very close to $W^{u}(\bar p).$ Using Poincaré recurrence theorem, we could choose $n$ in such a way that $\tilde{f}^{n}(\tilde{y})\in \tilde{\Delta}_{l}$ and put $\tilde{\alpha}:= \tilde{f}^{n}(\y).$
By definition $\x \in \tilde{\Lambda}^{u}(\bar p)$. So again without loss of generality we suppose $W^{u}(\tilde x)\pitchfork W^{s}_{loc}(\bar p)\neq\emptyset$ and $\tilde{\alpha}$ is such that $W^{s}_{\delta}(\tilde{\alpha})\pitchfork W^{u}(\bar p)\neq\emptyset$. Using $\lambda-$lemma we find some large $m$ that $\tilde{f}^{m}(\x)\in\tilde{\Delta}_{l}$ and $$W^{u}(\tilde{f}^{m}(\tilde{x}))\pitchfork W^{s}_{\delta}(\tilde{\alpha})\neq \emptyset.\,\,\,\,\,\,\,\,(*)$$ By hypothesis $\mu$ is hyperbolic with SRB property and $\pi_{*}(\tilde{\mu}^{u}_{\tilde{x}})\approx m^{u}_{\tilde{x}}$. Let call $\pi_{*}(\tilde{\mu}^{u}_{\tilde{x}})=\mu^{u}_{\tilde{x}}$. We are going to find a positive $\tilde{\mu}^{u}_{\tilde{x}}-$subset of $\widetilde{W}^{u}(\tilde{x})$ belonging to $\tilde{\Lambda}^{s}(\bar p)$. Using lemma \[4-13-2\] it is enough to find it on local unstable manifold of some iterate of $\x$.
![Ergodic Criterion[]{data-label="fig:ergodic criteria"}](crit01.pdf)
For this purpose, consider a very small ball $\tilde{B}_{\delta}(\tilde{\alpha})$ around $\tilde{\alpha}$. Covering $\tilde{B}_{\delta}(\tilde{\alpha})$ with a measurable partition sub-ordinate to unstable manifolds, there exists some point $\tilde{z}\in \tilde{\Delta}_{l}$ such that $$\tilde{\mu}^{u}_{\tilde{z}}(\tilde{S}|_{\tilde{\Lambda}^{s}(\bar p)}\cap\tilde{\Delta}_{l}\cap \tilde{B}_{\delta}(\tilde{\alpha}))>0.$$ Let us call this subset of $\widetilde{W}_{loc}^{u}(\tilde{z})$ by $\tilde{Z}:= \tilde{S}|_{\tilde{\Lambda}^{s}(\bar p)}\cap\tilde{\Delta}_{l}\cap \tilde{B}_{\delta}(\tilde{\alpha}))$. The projection $Z:= \pi(\tilde{Z})$ is inside $W^{u}(\tilde{z})\cap \pi( B(\tilde{\alpha}))$.
By definition of metric in the orbit space $M^f$, it is clear that $\pi (B(\tilde{\alpha}))$ is a ball of smaller radius (than the radius of $B(\tilde{\alpha})$) around $\alpha:= \pi(\tilde{\alpha}).$ So, $z :=\pi(\tilde{z})$ is close enough to $\alpha.$
Observe that stable lamination varies continuously in a Pesin block and as a consequence of transversality of $W^s_{\delta}(\tilde{\alpha})$ and $W^u(\bar{p})$ and $C^1-$closeness of $W^u(\tilde{f}^m(\x))$ to $W^u(\bar{p})$ one can define the stable holonomy map from $W^u(\tilde{z})$ into both $W^u(\tilde{f}^m(\x))$ and $W^u(\bar{p}).$ The domain of the holonomy map at least contains $Z.$
By SRB property we know that $m_{\tilde{z}}^u (Z) > 0$ and using absolute continuity of stable holonomy into $W^u(\tilde{f}^m(\x))$, we have $m^{u}_{\tilde{f}^{m}(\tilde{x})}(Z^*) > 0$ where $Z^*$ is the image of $Z$ by stable holonomy. Again using SRB property (equivalence of conditional measures and Lebesgue measure) it comes out that $\mu^{u}_{\tilde{f}^{m}(\tilde{x})}(Z^{*})>0.$ Now using definition of $\tilde \mu^{u}$ we have: $$\tilde\mu^{u}_{\tilde{f}^{m}(\tilde{x})}(\pi^{-1}(Z^{*}))=\mu^{u}_{\tilde{f}^{m}(\tilde{x})}(Z^{*})>0.$$ Finally observe that any point in $\pi^{-1}(Z^*)$ belongs to $\tilde{\Lambda}^s (\bar{p})$ as its stable manifold intersects $W^u(\bar{p})$ and this completes the proof.
\[4-14\] Let $f:M\rightarrow M$ be a $C^2-$endomorphism over a compact manifold $M$ and $\mu$ any measure with SRB property. If $\tilde{\mu}(\tilde{\Lambda}(\bar p))>0$ for a hyperbolic periodic point $\bar p$, then $\tilde{\mu}|_{\tilde{\Lambda}(\bar p)}$ is an ergodic component of $\tilde{\mu}$.
For simplicity once again let assume that $\bar p$ is a hyperbolic fixed point. As $\tilde{\mu}(\tilde{\Lambda}(\bar p))>0$, taking any $\tilde{f}-$invariant continuous function $\tilde{\phi}: M^{f}\rightarrow \mathbb R$, we show that $\tilde{\mu}$-a.e. points in $\tilde{\Lambda}(\bar p)\cap \tilde{J}$ ($\tilde{J}$ is the set of typical points from \[4-11\]) are $\bar{\phi}^{+}-$constant. Which implies the ergodicity of $\tilde{\mu}|_{\tilde{\Lambda}(\bar p)}$. Let choose arbitrary $\tilde{x},\tilde{y}\in \tilde{\Delta}_{l}\cap\tilde{\Lambda}(\bar p)\cap \tilde{S_1}:=\tilde{\Gamma}$ for some $l>0$. Without loose of generality we may assume that such $\tilde{x},\tilde{y}$ are in the support of $\tilde{\mu}|_{\tilde{\Lambda}(\bar p)}$. Using Poincaré recurrence theorem these points come back infinitely many times to $\tilde{\Gamma}$. Following a similar argument to theorem \[4-13\] we prove that $\tilde{\phi}^+(\x) = \tilde{\phi}^ +({\tilde{y}}).$ As the argument is exactly the same, just substituting $1_{\tilde{\Lambda}^s(\bar p)}$ to $\tilde{\phi}$ we do not repeat it here.
\[keypoint\] Let $f: M \rightarrow M$ be a $C^2-$endomorphism and $\mu$ any ergodic SRB measure of index $0 < k < n.$ Then there exists a hyperbolic periodic point $p_{\mu}$ in $Per_{k}$ such that $\tilde{\mu}(\tilde{\Lambda}(\bar{p}_{\mu})) =1.$
Observe that by lemma \[4-5\] we have a periodic point with ergodic homoclinic class of positive measure. The above theorem and the ergodicity of $\mu$ imply that $\tilde{\mu}(\tilde{\Lambda} (\bar{p}_{\mu}))=1.$
Proof of the Main Theorem
=========================
We have proved that ergodic homoclinic classes are in a close relationship with ergodic components of a measure. In fact for any two hyperbolic ergodic measure $\mu$ and $\nu$ with SRB property, we show that they are supported on the ergodic homoclinic class of some periodic point respectively $p_{\mu}$ and $p_{\nu}$. Then we prove that if either $[p_{\mu}, p_{\nu}] \neq \emptyset$ or $[p_{\nu}, p_{\mu}] \neq \emptyset$, then the measures are the same.
Although the base of this work is settled on the assumption of ergodic hyperbolic measures, using ergodic decomposition theorem [@23], theorem \[5-2\] and proposition \[5-3\] we can reduce the proof in the ergodic case. Theorem II.1.1 of [@10] gives a version of Margulis-Ruelle inequality for $C^{1}-$maps.
Lef $f$ be a $C^{1}-$map of a compact, smooth Riemannian manifold $M$. If $\mu$ is an $f-$invariant Borel probability measure on $M$, then $$h_{\mu}(f)\leq \int_{M}\,\sum_{i}\,\lambda_{i}^{+}(x)\,m_{i}(x)\,d\mu(x),$$ where $-\infty<\lambda_{1}(x)<\cdots<\lambda_{r(x)}<\infty$ are Lyapunov exponents of $f$ at $x$ and $m_{i}(x)$ is the multiplicity of $\lambda_{i}(x)$ for each $i=1,2,..,r(x).$
\[5-3\] Almost all ergodic components of $\mu$ are hyperbolic and SRB.
The hyperbolicity is easy to see because if not it would be possible to find a positive measure set of points with zero Lyapunov exponent and this contradicts the fact that $\mu$ is hyperbolic. For SRB property, we know that by ergodic decomposition and Margulis-Ruelle inequality, there exists a probability measure $\hat \mu$ in the space of all probability measures supported on ergodic measures $\mathcal{M}(f)$, such that $h_{\mu}=\int_{\mathcal{M}(f)}h_{\nu}\,d\hat{\mu}(\nu)\leq \int\sum_{i} \lambda^{+}_{i}(x)\,d\mu$, counting multiplicities. By theorem VII.1.1 of [@10] $\mu$ has SRB property if and only if $$h_{\mu}(f)=\int_{M}\sum_{i}\lambda_{i}^{+}(x)\,d\mu(x).$$ These clearly imply that $\hat{\mu}-$almost every $\nu$ will satisfy the entropy formula and so has SRB property.
\[5-2\] $f:M\rightarrow M$ a $C^2-$endomorphism over a compact manifold $M$ equipped with a hyperblic measure $\mu$ with SRB property and $\tilde{\mu}$ its lift. Then for any ergodic component $\tilde{\nu}$ of it, there exists a hyperbolic periodic point $p$ such that $\tilde{\nu}(\tilde{\Lambda}(\bar{p}))=1$.
Suppose that $\nu$ is an ergodic component of a hyperbolic measure $\mu$ with SRB property. By proposition \[5-3\] the ergodic components are also hyperbolic with SRB property. By proposition \[1\] we know the existence of the unique ergodic $\tilde\nu$ such that $\pi_{*} (\tilde\nu) = \nu$. By corollary \[keypoint\] we get the desired periodic point.
Let $\mu$ and $\nu$ be ergodic hyperbolic measures with SRB properties with respective periodic points $p_{\mu}$ and $p_{\nu}.$ Assume that $[p_{\mu}, p_{\nu}] \neq \emptyset.$
Let $B(\tilde \mu)$ and $B(\tilde \nu)$ be respectively the basins of $\tilde \mu$ and $\tilde \nu:$ $$B(\tilde\mu)=\{\tilde x:\lim_{n\rightarrow \infty}\frac{1}{n}\sum^{n-1}_{0}\tilde{\phi}(\tilde{f}^{i}(\tilde x))= \int\tilde \phi d\tilde \mu\,\,\,\,\,\,\,\; \forall\,\tilde\phi\in C(M^{f})\};$$
$$B(\tilde\nu)=\{\tilde x: \lim_{n\rightarrow \infty}\frac{1}{n}\sum^{n-1}_{0}\tilde{\phi}(\tilde{f}^{i}(\tilde x))= \int\tilde \phi d\tilde \nu\,\,\,\,\,\,\,\; \forall\,\tilde\phi\in C(M^{f})\};$$
By ergodicity $\tilde{\mu}(\tilde{\Lambda}(\bar p_{\mu}))=\tilde{\nu}(\tilde{\Lambda}(\bar p_{\nu}))=1$ and by $\widetilde{BET}$ \[4-10\], we can define $B_{\tilde{\mu}}$ and $B_{\tilde{\nu}}$ with $\tilde{\mu}(B_{\tilde{\mu}})=\tilde{\nu}(B_{\tilde{\nu}})=1$ as follows:
$$B_{\tilde{\mu}}=\{\tilde x:lim_{n\rightarrow \pm\infty}\frac{1}{n}\sum^{n-1}_{0}\tilde{\phi}(\tilde{f}^{i}(\tilde{x}))= \int\tilde{\phi} d\mu\,\,\,\,\,\,\,\; \forall\,\tilde{\phi}\in C(M^{f})\};$$
$$B_{\tilde{\nu}}=\{\tilde x:lim_{n\rightarrow \pm\infty}\frac{1}{n}\sum^{n-1}_{0}\tilde{\phi}(\tilde{f}^{i}(\tilde{x}))= \int\tilde{\phi} d\nu\,\,\,\,\,\,\,\; \forall\,\tilde{\phi}\in C(M^{f})\};$$
It means that $\tilde{\mu}$ (resp. $\tilde{\nu}$)-a.e. point $\tilde{x}\in\tilde{\Lambda}(\bar p_{\mu})$ (resp. $\tilde{\Lambda}(\bar{p}_{\nu})$) belongs to $B_{\tilde{\mu}}$ (resp. $B_{\tilde{\nu}}$). If we show that $B_{\tilde{\mu}}\cap B_{\tilde{\nu}}\neq\emptyset$ then we are done.
Let us take $\tilde{x}\in\tilde{\Lambda}(\bar{p}_{\mu})$ a point for which $\tilde{\mu}^{u}_{\tilde{x}}((B_{\tilde{\mu}}\cap \tilde{\Lambda}(\bar{p}_{\mu}))^ c)=0$. By SRB property of $\mu$ we will have $m^{u}_{\tilde{x}}((B_{\mu}\cap \Lambda(p_{\mu}))^ c)=0$ where $B_{\mu}:=\pi(B_{\tilde{\mu}})$. There exists some large Pesin block $\tilde{\Delta}_{l}, l\geq 1$ and $\tilde{y}\in\tilde{\Delta}_{l} \cap \tilde{\Lambda}(\bar{p}_{\nu})$, a density point of $\tilde{\nu}$ such that $m^{u}_{\tilde{y}}(B_{\nu}\cap \Delta_{l})>0$. Here $B_{\nu}:=\pi(B_{\tilde{\nu}}),\Delta_{l}=\pi(\tilde{\Delta}_{l})$. By Poincaré recurrence theorem $\tilde{y}$ returns back infinitely many times to $\tilde{\Delta}_{l}\cap B_{\tilde\nu}$ and consequently $y$ to $\Delta_{l}\cap B_{\nu}$.
As $\tilde{y}\in \tilde{\Lambda}(\bar{p}_{\nu})$, there exists large iterate $\tilde{\alpha}=\tilde{f}^{n}(\tilde{y})$ such that $W^{u}(\tilde\alpha)$ becomes very close to $W^{u}(\bar{p}_{\nu})$ in a similar way that has been explained in ergodic criteria section. Figure \[fig:ergodic criteria\] We have $\tilde{x}\in \tilde{\Lambda}(\bar{p}_{\mu})$ and $[p_{\mu}, p_{\nu}] \neq \emptyset.$ So, using $\lambda-$lemma we may find some large iterate $\tilde{f}^{m}(\tilde{x})$ such that $W^{u}(\tilde{f}^{m}(\tilde{x}))$ becomes close enough to $W^{u}(\bar{p}_{\nu})$ in a way that for a positive ${\nu}^{u}_{\tilde{\alpha}}-$measure $z \in \Delta_{l}\cap B_{\nu}$ we have $W^{s}_{loc}(z)\pitchfork W^{u}(\tilde{f}^{m}(\tilde{x}))\neq\emptyset$.
The stable lamination on a Pesin block is absolutely continuous [@10] and this implies that $m^{u}_{\tilde{f}^{m}(\tilde{x})}(B_{\nu}\cap \Delta_{l})>0$. We also have $m^{u}_{\tilde{f}^{m}(\tilde{x})}((B_\mu)^ c)=0$ which implies $B_{\mu}\cap B_{\nu}\neq\emptyset$ and finishes the proof.
Kan example {#kan}
===========
The Kan example is a local diffeomorphism $F$ defined on the cylinder $\mathbb{S}^1 \times [0,1]$ as a skew product: $$F(z, t):= (z^d, f_z(t)),$$ where $z \in \mathbb{S}^1$ is a complex number of norm one and $z^d$ is the expanding covering of the circle of degree $d > 1$. For each $z \in \mathbb{S}^1$ the function $f_z: [0, 1] \rightarrow [0,1]$ is a diffeomorphism fixing the boundary of $[0, 1].$ Take two fixed points of $z^d$ called $p, q.$ We require that $f_p$ and $f_q$ have exactly two fixed ponits each, a source at $t=1$ (respectively $t=0$) and a sink at $t=0$ (respectively $t=1$). Furthermore, $|f_z^{'} (t)| < d$ and $$\int \log f_z^{'}(0) dz < 0 \quad \text{and} \quad \int \log f_z^{'}(1) dz < 0.$$ Under these conditions $F$ has two intermingles SRB measures which are normalized Lebesgue measure of each boundary circle. Under some more conditions $F$ is also transitive (see [@BDV].) We consider two such examples and glue them to find a local diffeomorphism of $\mathbb{T}^2$ admitting two SRB measures and topologically transitive. Take $G : \mathbb{S}^1 \times [0,1] \rightarrow \mathbb{S}^1 \times [0,1] $ as follows: $$\label{doublekan}
G(z, t)= \left\{
\begin{array}{ll}
(z^d, 1 - \frac{1}{2} f_z(2t)) \qquad & 0 \leq t \leq \frac{1}{2}\\
(z^d, \frac{1}{2}f_z(2(1-t)) ) & \frac{1}{2} \leq t \leq 1.
\end{array} \right.$$
Observe that the two circles $\{t=0\}, \{t=1/2\}$ are invariant and support SRB measures with intermingles basin on the $2-$torus. We can see also that both $F$ and $G$ are transitive. However, $G^2$ lets invariant each half trous and consequently $G$ is not mixing. We are not aware of topologically mixing example of systems with intermingled basins of SRB measures.
Using the proof of our main theorem and the fact that the Lebesgue measures on each invariant circle $\{t=0\}$ and $\{t=1/2\}$ are hyperbolic SRB measures we can conclude that the number of SRB measures of $G$ is precisely two (without much geometric information about the volume of their basins).
Let $\mu_1$ and $\mu_2$ be respectively the normalized Lebesgue measure on the two invariant circles. It is easy to see that each of these circles is the ergodic homoclinic class of fixed points $p_{\mu_1}:= p, p_{\mu_2} := q$ corresponding to each SRB measure ($p, q$ defined above). Suppose that there exists another hyperbolic ergodic SRB measure $\nu.$ By corollary \[keypoint\] there exits a hyperbolic periodic point $P_{\nu}$ such that $\mu(\Lambda(p_{\nu}))= 1.$ Observe that $p_{\nu}$ is a periodic point in the torus minus two invariant circles. Taking large iterates of the local unstable manifold of $p_{\nu}$ we get a large curve transversal to the stable manifold of $p$ and $q$. That is $[p_{\nu}, p_{\mu_1}] \neq \emptyset$ and $[p_{\nu}, p_{\mu_2}] \neq \emptyset.$ By the proof of the main result $\nu=\mu_1 = \mu_2$ which is an absurd.
[RRRRRR]{}
Luis Barreira; Yakov B. Pesin, *Lyapunov exponents and smooth ergodic theory.* University Lecture Series, 23. American Mathematical Society, Providence, RI, 2002.
L. Barreira; Y. Pesin, *Nonuniform hyperbolicity. Dynamics of systems with nonzero Lyapunov exponents*. Encyclopedia of Mathematics and its Applications, 115. Cambridge University Press, Cambridge, 2007.
C. Bonatti, L. J. Díaz, and M. Viana. Dynamics beyound uniform hyperbolicity, *Volume 102 of Encyclopedia of Mathematical Sciences.* Springer-Verlag, 2005.
Yong Moo, Chung, Shadowing property of non-invertible maps with hyperbolic measures. (English summary) *Tokyo J. Math. 22*, (1999), no. 1, 145-166.
F. Rodriguez Hertz; M. Rodriguez Hertz; A. Tahzibi; R. Ures, Uniqueness of SRB measures for transitive diffeomorphisms on surfaces. *Comm. Math. Phys.* 306 (2011), no. 1, 35-49. Hirayama, M. and Sumi, N, On the ergodicity of hyperbolic Sinaĭ-Ruelle-Bowen measures: the constant unstable dimension case, *Ergodic Theory and Dynamical Systems*, FirstView, month 2, (2015), pages = [1–22]{}, doi: 10.1017/etds.2014.124. Anatole Katok; Boris Hasselblatt; *Introduction to the modern theory of dynamical systems*. (English summary) With a supplementary chapter by Katok and Leonardo Mendoza. Encyclopedia of Mathematics and its Applications, 54. Cambridge University Press, Cambridge, 1995.
A. Katok, Lyapunov exponents, entropy and periodic orbits for diffeomorphisms. *Inst. Hautes Études Sci. Publ. Math.*, No. 51 (1980), 137-173.
Yu Ilyashenko, V. Kleptsyn, and P. Saltykov. Openness of the set of boundary preserving maps of an annulus with intermingled attracting basins. *J. of Fixed Point Theory Appl.* no. 3 (2008), 449-463.
Ittai Kan, Open sets of diffeomorphisms having two attractors each with an everywhere dense basin, *Bull.Amer. Math. Soc 31*, (1994), no. 1, 68-74.
P.-D. Liu, M. Qian, Smooth Ergodic Theory of Random Dynamical Systems, Lect. Notes Math., *Vol. 1606,Springer-Verlag, New York*, 1995.
F. Ledrappier and J.-M. Strelcyn, A proof of the estimation from below in Pesin entropy formula,*Ergod. Th. & Dynam. Sys.* 2, (1982), 203-219.
F. Ledrappier and L.-S. Young, The metric entropy of diffeomorphisms, *Ann. Math.*122, (1985), 509-574.
Feliks Przytycki; Anosov endomorphisms. *Studia Math.* 58 (1976), no. 3, 249-285.
Min, Qian, Jian-sheng, Xie; Shu, Zhu *Smooth ergodic theory for endomorphisms*. Lecture notes in mathematics, Vol. 1978 . Springer-Verlag, Berlin Heidelberg, 2009.
Min Qian; Shu Zhu, SRB measures and Pesin’s entropy formula for endomorphisms. (English summary)*Trans. Amer. Math. Soc.*,354 (2002), no. 4, 1453-1471.
Naoya Sumi, A class of diferentiable toral maps which are topologically mixing, *Proc. Amer. Math. Soc.* 127 (1999), 915-924.
M. Viana and J Yang, Physical measures and absolute continuity for one-dimensional center direction. *Annales Inst. Henri Poincaré - Analyse Non-Linéaire 30* (2013), 845-877.
M. Viana and J Yang, Geomteric and measure-theoretical structures of maps with mostly contracting center, Preprint, 2015.
Zhu, Shu, Unstable manifolds for endomorphisms. (English summary) *Sci. China Ser.* A 41 (1998), no. 2, 147-157.
[^1]:
[^2]:
[^3]: P.M was supported by Capes. A.T was partially supported by CNPq and Fapesp.
[^4]: We would like to thank M. Andersson and J. Yang for pointing to us such construction of Kan example on $\mathbb{T}^2.$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use [*Kepler*]{} [*K2*]{} Campaign 4 short-cadence (one-minute) photometry to measure white light flares in the young, moving group brown dwarfs 2MASS J03350208+2342356 (2M0335+23) and 2MASS J03552337+1133437 (2M0355+11), and report on long-cadence (thirty-minute) photometry of a superflare in the Pleiades M8 brown dwarf CFHT-PL-17. The rotation period (5.24 hr) and projected rotational velocity ($45$ km s$^{-1}$) confirm 2M0335+23 is inflated ($R \ge 0.20 R_\odot$) as predicted for a $0.06M_\odot$, 26-Myr old brown dwarf $\beta $Pic moving group member. We detect 22 white light flares on 2M0335+23. The flare frequency distribution follows a power-law distribution with slope $-\alpha = -1.8 \pm 0.2$ over the range $10^{31}$ to $10^{33}$ erg. This slope is similar to that observed in the Sun and warmer flare stars, and is consistent with lower energy flares in previous work on M6-M8 very-low-mass stars; taken the two datasets together, the flare frequency distribution for ultracool dwarfs is a power law over 4.3 orders of magnitude. The superflare ($2.6\times10^{34}$ erg) on CFHT-PL-17 shows higher energy flares are possible. We detect no flares down to a limit of $2 \times 10^{30}$ erg in the nearby L$5\gamma$ AB Dor Moving Group brown dwarf 2M0355+11, consistent with the view that fast magnetic reconnection is suppressed in cool atmospheres. We discuss two multi-peaked flares observed in 2M0335+23, and argue that these complex flares can be understood as sympathetic flares, in which a fast-mode MHD waves similar to EUV waves in the Sun trigger magnetic reconnection in different active regions.'
author:
- 'John E. Gizis'
- 'Rishi R. Paudel'
- Dermott Mullan
- 'Sarah J. Schmidt'
- 'Adam J. Burgasser'
- 'Peter K. G. Williams'
bibliography:
- '../astrobib.bib'
title: '[*K2*]{} Ultracool Dwarfs Survey II: The White Light Flare Rate of Young Brown Dwarfs'
---
Introduction\[intro\]
=====================
The paradigm for the evolution of magnetic activity in low-mass main sequence stars is that magnetic braking causes the initially rapid rotation from pre-main sequence contraction to gradually decline, and this in turn causes the magnetic fields generated by the dynamo to weaken . As a result, both the rotation rate and magnetic activity such as flaring decrease with age [@Gershberg:2005lr; @2005ApJ...622..653T]. For fully convective $0.3 M_\odot$ stars, half the angular momentum is shed between 3 Myr [**and the Pleiades age**]{} [@2016AJ....152..115S]. The rotation and magnetic activity evolution of brown dwarfs is quite different. Measurements of $v \sin i$ for field brown dwarfs [@2006ApJ...647.1405Z; @2008ApJ...684.1390R] imply a mean rotation period of 4.1 hours [@2014ApJ...793...75R], and a large sample of mid-infrared photometric periods confirm this view [@2015ApJ...799..154M]. All of these are rapid rotators compared to field stars. @2009Natur.457..167C show that turbulent dynamos can generate magnetic fields in stars, brown dwarfs and planets, and that provided the object is rapidly rotating, the strength of the magnetic fields is determined by the energy flux. show this theory implies that massive brown dwarfs have fields of a few kilogauss in their first few hundred million years, weakening to fields of 100-1000G by an age of $10^{10}$ years. Simulations of the turbulent dynamo in fully convective stars show that both large-scale dipole and small-scale magnetic fields are generated [@2015ApJ...813L..31Y]. Overall, the expectation is that all brown dwarfs have significant magnetic fields, and indeed radio observations support the existence of magnetic fields even in cool T-type brown dwarfs [@2012ApJ...747L..22R; @2015ApJ...808..189W; @2016ApJ...830...85R].
Despite the presence of strong magnetic fields, the fraction of a star or brown dwarf’s energy converted into chromospheric activity weakens for “ultracool dwarfs" with temperatures below the M6 spectral type [@2000AJ....120.1085G; @2015AJ....149..158S]. @Mohanty:2002lr have shown this can be understood as a consequence of the increasingly neutral atmospheres: As the ionization fraction drops and the resistance increases, the magnetic fields become decoupled from the matter. These were equilibrium calculations, and as @Mohanty:2002lr noted, the existence of flares implies transient, time-dependent processes are important. A transition from fast magnetic reconnection at high temperatures to a high resistivity regime where only slow magnetic reconnection is allowed may explain the decline in chromospheric and coronal activity but continued radio emission [@2010ApJ...721.1034M]. This scenario sees the fast reconnection events resulting in a range of energy release events, from many nanoflares that heat the chromosphere and corona to rarer but more powerful white light flares that can be individually observed. Additional parameters seem to be important in magnetic activity: Magnetic topology may explain the difference between radio-quiet, X-ray bright dwarfs and the radio-loud, X-ray faint dwarfs [@2014ApJ...785...10C; @2014ApJ...785....9W].
Even setting aside the numerous radio-only bursts, it is well established that X-ray and optical flares do occur in stellar late-M and early-L dwarfs – some notable examples include the very first optical spectrum of , the first known M8 dwarf [@1956PASP...68..531H], the discovery of a nearby M9 dwarf due to a huge X-Ray flare with $L_X/L_{\rm bol} =0.1$ [@Hambaryan:2004mz], and an L0 dwarf with a $\Delta V<-11$ white light flare [@2016ApJ...828L..22S]. [**The serendipitous optical spectra of the M7-M9 dwarf flares reported by @Bessell:1991rw, @1999MNRAS.302...59M, @1999ApJ...519..345L, @1999ApJ...527L.105R, @2000AJ....120.1085G, and @Martin:2001qy all showed strong atomic emission lines and many included veiling or a blue continuum.**]{} A difficulty, however, with flare studies is that detectable flares are rare enough that it is difficult to assess their frequency [**as a function of energy**]{}. @2011PhDT.......144H monitored four field M6-M8 stars in U-band and found 39 flares over 59 hours. These flares followed a power-law frequency distribution, as seen in hotter flare stars, but with a rate comparable to “inactive" but more luminous M0-M2.5 dwarfs. Similarly, [*Kepler*]{} optical monitoring of a L1 dwarf star also found a power-law flare frequency distribution [@2013ApJ...779..172G]. Young M-type brown dwarfs with similar $T_{\rm eff}$ also exhibit flares, such as the $\sim500$-Myr M9 lithium brown dwarf [@2000ApJ...538L.141R]. X-Ray flares, as well as quiescent emission, have also been reported from very young ($<5$ Myr) M6-M9 brown dwarfs in Orion [@2005ApJS..160..582P] and Taurus [@Grosso:2007vn].
By monitoring over 100,000 stars over four years, [*Kepler*]{} mission [@2010ApJ...713L..79K] is known to have detected over 800,000 flares in 4041 stars [@2016ApJ...829...23D]. These include superflares in A and F stars with thin convective zones [@2012MNRAS.423.3420B; @2015MNRAS.447.2714B] and solar-like G dwarfs [@2012Natur.485..478M] as well as fully-convective M dwarfs [****]{} and even an L1 dwarf [@2013ApJ...779..172G]. These flares are detected by white light emission enhancing the normal stellar photospheric emission through [*Kepler*]{}’s broad (430nm - 900nm) filter; thus, only extremely energetic events ($>10^{34}$ erg) are seen in the warmer stars but weaker flares can be seen in the coolest stars. In the models of @2015SoPh..290.3487K, a beam of non-thermal electrons with a energy flux of $10^{13}$ erg cm$^{-2}$ s$^{-1}$ can produce a dense, hot chromospheric condensation that emits white light like a $\sim10,000$K blackbody. Solar flares with energies $\sim10^{31}$ erg also emit most of their energy in white light like $\sim9000$K blackbodies — flares of this energy would not be detectable by [*Kepler*]{} on solar-type stars but were detected on the L1 dwarf and the M4 dwarf GJ 1243 [@2016ApJ...829..129S].
The extended [*Kepler*]{} [*K2*]{} mission [@2014PASP..126..398H] allows many new targets to be observed. We are using [*K2*]{} to monitor ultracool dwarfs for white light flares as well as measuring rotational periods and searching for transits. Targets that happen to lie within each [*K2*]{} field of view are monitored for a $\sim2.5$ month-long campaign. The overall aim of our survey is to measure quantities such as the flare frequency, maximum flare energy, flare light curve morphology in order to understand their dependence on parameters such as temperature, mass, and age. In this paper, we present [*K2*]{} Campaign 4 observations of three brown dwarfs which are confirmed members of nearby moving groups and clusters, so that unlike most field dwarfs, their age, mass, radius and other parameters are well determined. We present the target properties in Section \[s:targets\], the [*K2*]{} observations in in Section \[s:K2data\], and discussion of the magnetic activity in Section \[s:discussion\].
Targets and Spectroscopy {#s:targets}
========================
Target Characteristics
----------------------
In Table \[tab:prop\], we list the key properties of our targets.[^1] (hereafter 2M0335+23) was observed by [*K2*]{} as source EPIC 211046195. This field ultracool dwarf was discovered by @2000AJ....120.1085G, who classified it as M8.5 in the optical and noted H$\alpha$ emission. [**It is apparently single in [*Hubble Space Telescope*]{} imaging [@2003AJ....125.3302G].**]{} @hiresmdwarfs detected significant rotational broadening ($v \sin i \approx$ 30 [km s$^{-1}$]{}), lithium in absorption, and again detected H$\alpha$ emission. The presence of lithium identified this object unambiguously as a brown dwarf [**[@1992ApJ...389L..83R].**]{} @2012ApJ...758...56S measured its trigonometric parallax distance to be $42.4 \pm 2.3$ pc and showed that its distance and space velocity identified it as a member of the $\beta$ Pic Moving Group (BPMG). The latest parallax from @2016ApJ...833...96L places it at $46\pm 4$ pc; for consistency with the literature we adopt the nominally more precise distance of 42.4pc for our analysis. The age of this group is $24 \pm 3$ Myr [@2014MNRAS.445.2169M; @2015MNRAS.454..593B]. @2015ApJS..219...33G have analyzed 2M0335+23 in detail to find that it is $60.9^{+4.0}_{-4.4}$ jupiter masses (i.e, $0.058\pm0.004 M_\odot$) with a radius of $2.40 \pm 0.04$ jupiter radius according to models [@2015ApJS..219...33G]. Adopting BC$_J = 2.0$ [@2015ApJ...810..158F], the luminosity is $10^{-2.55} L_\odot$ and $T_{\rm eff} = 2700$K. measured it to have apparent (AB) magnitude $i=15.601$ (as source DANCe 5121623). Infrared spectroscopy confirms that it is lower surface gravity than ordinary field dwarfs, with a classification of M7 VL-G [@2013ApJ...772...79A] and M7.5$\beta$ [@2015ApJS..219...33G]. In the optical, low surface gravity leads to enhanced VO features [@2008ApJ...689.1295K]; this would bias @2000AJ....120.1085G’s classification to a later type. We re-classify the spectrum as M7$\beta$ in the optical. In particular, the optical spectrum of 2M0335+23 is definitely “earlier" (warmer) than young M9 brown dwarf 2MASS J06085283-2753583 [@2003AJ....126.2421C] [**which may be a BPMG member [@2010ApJ...715L.165R] or more likely a 40-Myr-old Columba member [@2016ApJ...833...96L].**]{} The observed H$\alpha$ emission line strength of EW$\approx 5\AA$ imply $\log H\alpha/L_{bol} \approx -5.5$ [@2014PASP..126..642S]. Despite 2M0335+23’s rapid rotation and youth, this places it in the bottom half of the M7 activity range [@2014arXiv1410.0014S]. Finally, we use the mass, luminosity and radius to predict theoretical mean surface magnetic fields of our targets using Equation 1 of .
(hereafter 2M0355+11) was observed by [*K2*]{} as source EPIC 210327027. Discovered by @Reid:2006fj and classified as low-surface gravity (L5$\gamma$) [**with lithium**]{} by @2009AJ....137.3345C, this brown dwarf is now recognized as a dusty, low-surface-gravity member of the AB Doradus Moving Group (ABDMG) which shares many spectral characteristics with directly imaged exoplanets [@Faherty:2013qy; @Liu:2013lr]. @2015MNRAS.454..593B derive an ABDMG age of $149^{+51}_{-19}$ [Myr]{}; note that this age is tied to the Pleiades age of $130 \pm 20$ Myr from lithium depletion [@2004ApJ...614..386B] [**that has been updated to $112 \pm 5$ Myr [@2015ApJ...813..108D]**]{}. @2016ApJ...833...96L measure a trigonometric parallax of [**$109.5\pm 1.4$ mas**]{}, and we use a distance of 9.1 pc for the rest of this paper. For discussion purposes, we adopt the values derived by @2015ApJ...810..158F: $\log L/L_{bol} = -4.10 \pm 0.03$, radius $R=1.32\pm0.09 R_J$, surface gravity $\log g = 4.45 \pm 0.21$, $T_{\rm eff} = 1478 \pm 57$K, and mass $M=19.98 \pm 7.76 M_J$ (i.e., $\sim0.02 M_\odot$). @2010ApJ...723..684B measured $v \sin i = 12.31 \pm 0.15$ [km s$^{-1}$]{}, noting that it is an unusually slow rotator for a L dwarf. This, however, still implies a minimum rotation period of 13 hr, a rapid rotator compared to M dwarf stars. No H$\alpha$ emission has been detected (EW$<24.29$Å, @2007AJ....133.2258S) though the upper limit is above the emission level of most L dwarfs.
, a brown dwarf member of the Pleiades discovered by ), was observed by [*K2*]{} as source EPIC 211110493. @2000ApJ...543..299M classified it as optical spectral type M7.9 (which we will hereafter round off to M8.) and found H$\alpha$ emission (EW$ \approx 7$Å). confirm it has a 100% chance of being a cluster member and measure $i=19.745$ (AB). We adopt the VLBI Pleiades distance of $136.2 \pm 1.2$ pc [@2014Sci...345.1029M]; , using the same distance, find a luminosity of $0.0008456 L_\odot$ ($T_{\rm eff} = 2500$K) which implies the mass is $0.06 M_\odot$. Thus, CFHT-PL-17 is very similar in mass to 2M0335+23, but older and one spectral type later ($\sim200$K cooler), and it is a similar age to 2M0355+11, but more massive and warmer.[^2]
[llllrrrcrr]{} 2M0335+23 & 211046195 & 16.7 & M7$\beta$ & 42.4 pc & 24 Myr & $0.06 M_\odot$ & -2.55 &2700K & 2.2 kG\
2M0355+11 & 210327027 & 20.4 & L5$\gamma$ & 9.1 pc & 150 Myr & $0.02 M_\odot$ &-4.10 &1480K & 1.1 kG\
CFHT-PL-17 & 211110493 & 20.8 & M8$\beta$ & 136 pc & [**112**]{} Myr & $0.06 M_\odot$ & -3.07 &2500K & 2.5 kG\
New Spectroscopy
----------------
We observed 2M0335+23 on UT Date 2016 February 3 with the Keck NIRSPEC spectrograph [@McLean:2000lr] to obtain spectra with $\lambda / \Delta \lambda = $20,000 in the 2.3 $\mu$m region dominated by CO bands. Conditions were clear with 1 seeing. We obtained two exposures of 750 sec each, following observation of the A0V star HD 19600 for telluric calibration.
The setup and data analysis were as described in @2013ApJ...779..172G. We achieved a typical signal-to-noise of $>50$ for these observations. We find $v_{rad} = 12.6 \pm 1.0$ [km s$^{-1}$]{}and $v \sin i = 45.4 \pm 3.4$ [km s$^{-1}$]{}. This radial velocity increases 2M0335+23’s probability of BPMG membership to 96.5% using the @2012ApJ...758...56S astrometry in the BANYAN II model [@2014ApJ...783..121G; @2015ApJ...798...73G].
[*K2*]{} Photometry {#s:K2data}
===================
[*Kepler*]{} records the pixels for every target as averages over “long" (30 minute, @Jenkins:2010fk) cadences; for 2M0335+23 and 2M0355+11, it also recorded “short" (1 minute, @2010ApJ...713L.160G) cadence data. We report [*Kepler*]{} mission times, which are equal to BJD - 2454833.0.
The brightnesses of Kepler and [*K2*]{} targets are described on the $K_p$ magnitude system [@2011AJ....142..112B] tied to ground-based photometry; this system was not designed for ultracool dwarfs and the EPIC catalog [@2016ApJS..224....2H] magnitudes for our targets are not useful. @2015ApJ...806...30L defined $\widetilde{K}_p \equiv 25.3 - 2.5 \log({\rm flux})$, where $\widetilde{K}_p \approx K_p$ for most (e.g., AFGK-type) stars and “flux" is the count rate measured through a 3-pixel radius aperture. We find that the apparent $\widetilde{K}_p$ magnitudes of 2M0335+23, 2M0355+11, and CHFT-PL-17 are 16.7, 20.4, and 20.8. By using $\widetilde{K}_p$ we can discuss both extremely red sources and time-dependent (blue) flares in a terms of the well established $K_p$ system.
2M0335+23 is bright enough that photospheric variability is detectable. We use the [*K2*]{} mission pipeline photometry corrected for the effects of pointing drift and other systematic errors. For 2M0335+23, the Lombe-Scargle periodogram shows a strong signal at $P=0.2185$ day (5.244 hour), which we identify as the rotation period of the brown dwarf. The phased data are shown in Figure \[fig-phased0335\] normalized to the median. We have verified that other [*K2*]{} pipelines [@2014PASP..126..948V; @2016MNRAS.456.2260A; @2016MNRAS.459.2408A] give consistent results for this source. We do not detect periodic photometric variations in 2M0355+11 or CFHT-PL-17. We note that in the case of the [*K2*]{} mission corrected photometry for 2M0355+11, there is a periodic signal of 1.11 days, but we believe this is a spurious signal, and it is not present in the other reduction pipelines.
We measured short cadence photometry using the [*K2*]{} Release 10 pixel files. (There is not yet any [*K2*]{} mission official light curve product for short cadence data.) We used circular aperture photometry ([photutils]{}) with radius of 2 pixels centered on the target; but we verified that our results are qualitatively unchanged with circular apertures of radius 3 or rectangular fixed apertures. The sources, especially 2M0355+11, are faint enough that centroiding introduces photometric noise; we adopt a best position based on the median of all centroid measurements, and then adjust it for each observation using the spacecraft motion estimate calculated by the mission (recorded as POS\_CORR1 and POSS\_CORR2 in the FITS file headers). The photometry shows the usual systematic drifts, but these have little effect on measurements of flares which have timescales of a few minutes.
We lack short cadence photometry for CFHT-PL-17, but motivated by strong flare at mission day 2261.94 which we noticed in the mission pipeline photometry, we measure our own 2 pixel radius long cadence photometry in the same way to analyze its flare. We identified 22 flares in 2M0335+23 by visually examining the lightcurves; they are shown in Figure \[fig-flares\] and listed in Table \[tab:flares\]. We reject all events that brighten in only a single observation or are not centered on the target’s position. No flares are detected in 2M0355+11; we note however that a passing asteroid creates a spurious brightening in aperture photometry at mission time 2271.13.
For each 2M0335+23 flare, we fit the M dwarf flare light curve described by @2014ApJ...797..122D, hereafter D14, who found that most flares on the M4 dwarf GJ 1243 could be described by a fast rise (a 4th order polynomial) and a slower double exponential decay:
$$\Delta F = A (\alpha_i e^{-\gamma_i \Delta t/t_{1/2}} + \alpha_g e^{-\gamma_g \Delta t/t_{1/2}})
\label{D14equation}$$
All flare light curves are then described by nine universal parameters: $\alpha_i=0.6890 (\pm 0.0008)$, $\alpha_g = 0.3030 (\pm 0.0009)$, $\gamma_i=1.600(\pm0.003)$, $\gamma_g = 0.2783(\pm0.0007)$ plus the five polynomial parameters given in D14. Each flare also has three unique parameters: the peak amplitude of the flare, the full-width time at half-max ($t_{1/2}$), and the time of the flare peak. We fit four free parameters to each flare using [emcee]{} [@2013PASP..125..306F]: the non-flaring photosphere, the peak amplitude of the flare, the full-width time at half-max ($t_{1/2}$), and the time of the flare peak. The fits are shown in red in Figure \[fig-flares\]. Two of the brighter flares (on mission days 2240 and 2287) are complex flares with two peaks in their light curves: We have fit them as the superposition of two flares. The equivalent duration listed in Table \[tab:flares\] is a measure of the flare energy compared to the quiescent luminosity; it is obtained by integrating the observed flare count rate and dividing by the photosphere count rate. This is a distance-independent measure of the flare, but it is dependent on the filter and the photospheric properties of the stars: Our durations will be much longer than otherwise identical flares observed on a G dwarf due to the lower photospheric flux. Our detection of weak flares is limited by noise, and it is clear that we cannot reliably detect flares below equivalent durations of 20s. The flares on mission days 2238, 2249, and 2251 in Figure \[fig-flares\] are examples of marginal detections that may be some form of correlated noise rather than real flares. The flare on mission day 2299 is also questionable because most of the flux occurs in a single time period; its exclusion would have negligible effects on the remaining analysis.
The strongest 2M0335+23 flare (mission time 2253.65107) has noticeable deviations from the D14 template: The best fitting model (red) under-predicts the peak and over-predicts the gradual phase, with a $t_{1/2}$ that is too short. We therefore use a new model template in which we keep the polynomial rise parameters fixed but allow the decay parameters to be fitted. This adds three free parameters, since we require ($\alpha_i + \alpha_g = 1$). The results are: $\alpha_i=0.9233\pm 0.0055$ ($\alpha_g=0.0767$), $\gamma_i = 1.3722 \pm 0.054$, and $\gamma_g=0.1163 \pm 0.011$. @2014ARep...58...98G argues during the impulsive decay phase cooling is by blackbody emission, which suggests the relative contribution of this component of radiative cooling was different in this flare.
To calibrate equivalent duration in terms of energy, we follow the procedures described in @2013ApJ...779..172G. Because flares are much hotter than the brown dwarf targets, white light flares have a higher average energy per detected [*K2*]{} photon than the photosphere. The photosphere is modeled with an active M7 dwarf template [@Bochanski:2007ys] scaled to the measured $i$ photometry and known trigonometric parallax distance. The flare is modeled as an 10,000K blackbody, which gives good agreement with the flare measurements in @Hawley:1991uq. Figure \[fig-calib\] shows the optical and near-infrared spectral energy distribution of 2M0335+23 and flare with the same count rate through the [*Kepler*]{} filter. We find that the a flare with equivalent duration of 1 second has a total (bolometric UV/Vis/IR) energy of $2.0 \times 10^{30}$ erg. We emphasize that we have extrapolated to wavelengths not detected by [*K2*]{} and that our analysis includes atomic emission features between 430nm and 900nm in the observed “white light" photometry. Finally, we also report the peak (short cadence) absolute $\widetilde{K}_p$ magnitude of the flare. For these, we have applied an aperture correction of 1.08 to correct the $r=2$ pixel aperture to the $r=3$ pixel aperture. We note that the total equivalent duration of detected flares is 0.030 days, so that just 0.04% of the 2M0335+23’s optical light over the course of the campaign is due to white light flares.
[rrrrrl]{} 2253.65107 & 1.83 & 518 & 103.5 &\
2240.04075 & 0.68 & 494 & 98.9 & 2240.0511 & Complex\
2281.57290 & 1.47 & 339 & 67.8 &\
2284.68258 & 1.69 & 267 & 53.4 &\
2287.91009 & 0.22 & 203 & 40.5 & 2287.9161 & Complex\
2268.87127 & 0.20 & 145 & 29.0 &\
2295.39965 & 0.75 & 110 & 22.0 &\
2299.37770 & 0.66 & 56 & 11.2 & & Questionable\
2291.45770 & 0.38 & 53 & 10.7 &\
2248.59421 & 0.24 & 48 & 9.6 &\
2261.51045 & 0.23 & 47 & 9.4 &\
2269.02792 & 0.15 & 35 & 7.0 &\
2249.17243 & 0.11 & 35 & 6.9 &\
2276.02709 & 0.17 & 32 & 6.4 &\
2229.08574 & 0.07 & 28 & 5.5 &\
2293.30131 & 0.13 & 26 & 5.2 &\
2295.02575 & 0.15 & 25 & 5.0 &\
2258.54786 & 0.04 & 24 & 4.8 & & Complex?\
2284.92844 & 0.09 & 23 & 4.7 &\
2238.85706 & 0.11 & 18 & 3.5 & & Questionable\
2251.42265 & 0.05 & 13 & 2.6 && Questionable\
2249.41148 & 0.05 & 10 & 1.9 & & Questionable\
We detect no flares in 2M0355+11. We verified that we could have detected flares by adding our observed 2M0335+24 flare data back into the 2M0355+11 at random times, and recovered them all. Because 2M0355+11 is 4.7 times closer than 2M0335+24, a flare with 22 times less energy would produce the same count rate. We conclude that we could have detected flares with $E>2.0\times10^{30}$ ergs, and place a 95% confidence upper limit of 3 such flares over 70.7 days. (The effect of 2M0355+11 photosphere’s much fainter apparent magnitude would simply to be increase the equivalent duration or relative amplitude of the flare.) If the timescale for these flares, however, was less than one minute than we could not distinguish them from cosmic rays or other noise sources. However, because the late-M dwarf flares (@2011PhDT.......144H, Table 4.1) with energies at or above our limit have timescales of several minutes, we conclude that this effect is not a concern.
We measure the flare on CFHT-PL-17 using 3-pixel radius aperture photometry (Figure \[fig-superflare\]). The flare is first detectable at mission time 2261.9407 where it has brightened to 9.0 times the original photosphere. In the 2261.9612 exposure, it has reached 77 times the photosphere to achieve $\widetilde{K}_p= 16.0$. The flare then declines, with the last detectable excess of 24% at 2262.1246, for a total observed duration of five hours. The equivalent duration is 170,000 s (2.0 days). Using the same calibration procedure with an active M8 template [@Bochanski:2007ys], we calculate the flare energy is $2.6 \times 10^{34}$ erg. Given the sharply peaked light curve, we conclude that $t_{1/2} < 30$ min.
We can fit the D14 template by computing it on one-minute timesteps but comparing to the long-cadence data, as in our analysis of an L dwarf super flare in Paper I [@2017ApJ...838...22G]. We find that the peak is 380 times the photosphere ($\widetilde{K}_p \approx 14.3$), with $t_{1/2}$ of 3.9 minutes. This should be viewed with caution because we do not know if the D14 template applies to this flare, or if the flare was complex and multi-peaked. On the other hand, [*Kepler*]{} short cadence photometry of comparable energy flares in F stars shown by @2012MNRAS.423.3420B are sharply peaked.
Discussion {#s:discussion}
==========
The Radius of a 24 Myr Brown Dwarf
----------------------------------
Young brown dwarfs should have inflated radii compared to field stars of the same spectral type. Our measured rotation period and measured $v \sin i$ together imply that $R \sin i = 0.196 \pm 0.015 R_\odot$ for 2M0335+23, whose age of $24 \pm 3$ Myr is independently known. This is much larger than the radii of field M7 dwarfs, which have have $R = 0.12 R_\odot$ [@2014AJ....147...94D]. This result can be seen as either an independent confirmation of the evolutionary model prediction that young brown dwarfs are larger, or if the models are trusted, as independent support for BPMG group membership of 2M0335+23. Using the previously estimated radius of @2015ApJS..219...33G, we find that the inclination is $i = 54.4 \pm 6.6^\circ$.
Flare Frequency Distribution
----------------------------
Studies of flare stars have found that the cumulative flare frequency distribution (FFD) follows a power-law trend . We compute the cumulative frequency ($\nu$) of 2M0335+23 flares as the number ($N$) observed with a given energy or greater divided by the total time of observation (70.7 days) and plot the results in in Figure \[fig-ffd\]. [**The statistical properties of power-law (“Pareto") distributions are reviewed by @Arnold2015. We consider both the graphical technique of fitting a line to Figure \[fig-ffd\] and maximum likelihood estimation: @Arnold2015 notes that the two techniques are consistent and that the traditional graphical techinique is “only slightly inferior" to the maximum likelihood estimates.** ]{} Considering flares in the energy range $4\times10^{31}$ erg to $1.1 \times 10^{33}$ erg, using frequency units of day$^{-1}$, and weighting each point by $\sqrt{N}$, we fit a linear relationship.
$$\log \nu = a +\beta \log(E/{10^{32} \rm{erg}})$$
We find $\beta= -0.66 \pm 0.04$ and $a= -0.83 \pm 0.01$ for 2M0335+23. An alternative expression of this power law dependence is:
$$dN \propto E^{-\alpha} dE$$
Here $dN$ is the number of events between energy $E$ and $E+dE$. As often discussed for flares, if $\alpha>2$ the total energy in small events (nanoflares) would diverge. Because $\beta = -\alpha +1$ (see @2014ApJ...797..121H for helpful discussion), $\alpha=1.7$ for 2M0335+23. We also plot limits on the FFD for the L5$\gamma$ brown dwarf 2M0355+11. We also show the [*Kepler*]{} L1 dwarf W1906+40 [@2013ApJ...779..172G] FFD but over the energy range $10^{31}$ erg to $2 \times 10^{32}$ erg. For this star, $a=-1.35\pm0.06$ and $\beta= -0.59 \pm 0.09$, but we caution that the slope depends sensitively on the energy range chosen.
**An alternative approach is to use the maximum likelihood estimator for $\alpha$ [@Arnold2015]:**
$$(\alpha -1) = n \left[ \sum_{i=1}^{n} \ln \frac{E_i}{E_{min}}\right]^{-1}$$
where for 2M0335+23 case $n=19$ and $E_{min} = 4.66\times10^{31}$ erg. The uncertainty in this estimator can also be calculated (we used the software from @2014PLoSO...985777A). Because our sample is relatively small the uncertainty is large ($\pm 0.2$), and importantly, the estimator is biased. @Arnold2015 shows it can be made unbiased by multiplying it by a factor of $\frac{n-2}{n}$, giving $\alpha = 1.79 \pm 0.21$. We conclude that the maximum likelihood estimate of $\alpha$ is consistent with the linear fit value. We adopt the (larger) uncertainty of $\pm 0.2$ from the maximum likelihood estimator. For W1906+40, the maximum likelihood estimator is $\alpha=1.6 \pm 0.2$.
The 2M0355+11 limit is shown in red in Figure \[fig-ffd\]. If one assumes that 2M0355+11 would follow the same power law slope as 2M0335+23 or W1906+40, then the red dashed-line upper limit shown in Figure \[fig-ffd\] applies. In any case, 2M0355+11’s incidence of flares of $2 \times 10^{32}$ erg or greater is less than 13% that of 2M0335+23 and less than 40% that of the much older star W1906+40.
@2011PhDT.......144H found that for four M6-M8 dwarfs, $a=20.49 \pm 3.3$ and $\beta = -0.73 \pm 0.1$ over the range $10^{27.94} \le E_U \le 10^{30.60}$ erg, where the frequency is in hour$^{-1}$, $E_U$ is the energy in U-band, and the reported intercept $a$ is at zero energy, not $10^{32}$ erg. In order to compare this FFD, we can make a crude energy correction by noting that @Hawley:1991uq found that the total energy of a flare was $806/145=5.55$ times greater than the U-band energy (their Table 6). (The 10,000K blackbody agrees well with this flare, see @2013ApJ...779..172G). Applying this correction, the @2011PhDT.......144H relation agrees remarkably well not only in slope but also normalization (Figure \[fig-ffd\]). The agreement in normalization seems rather fortuitous given the uncertain energy corrections applied, the combination of multiple stars, and the fact that as a young brown dwarf, 2M0335+23 is both larger and more luminous than a field M7 dwarf.
We conclude that the flares for ultracool dwarfs follow a power law over the range $5 \times 10^{28}$ to $10^{33}$ erg, a range of 4.3 orders of magnitude. The existence of the Pleiades flare shows that this distribution must continue to at least $3 \times 10^{34}$, although we cannot determine whether it follows the same power-law slope or turns over. The predicted rate using the extrapolated 2M0335+23 FFD for the Pleiades superflare is 1.4 per year. The striking aspect of this power-law slope is that it agrees so well not only with stellar M6-M8 dwarfs, but also “inactive" M dwarf stars. It is also in excellent agreement with solar flares. @1993SoPh..143..275C found that $\alpha \approx 1.5$ for solar hard-X ray flares over three orders of magnitude. @2000ApJ...535.1047A argue that solar flares follow a power low slope with $\alpha=1.8$ over the range $10^{24}$ to $10^{32}$ erg. Power-law slopes near this value are explained by self-organized criticality models [@1991ApJ...380L..89L; @2016SSRv..198...47A] where flares are due to “avalanches" of many small magnetic reconnection events. 2M0355+11 may be at a temperature (1480K) below which fast magnetic reconnection events are no longer possible.
While the slopes ($\alpha$) are consistent between the brown dwarfs, the Sun, and other stars, the overall normalization of the flare rate ($a$) is much different. Flares on 2M0335+23 are much less frequent than in the rapidly rotating M4 dwarf GJ 1243 [@2016ApJ...829..129S], comparable to “inactive" M dwarfs studied by @2014ApJ...797..121H, and more frequent than in the Sun. It is perhaps most interesting to compare to the L1 dwarf W1906+40. At an energy of $10^{32}$ ergs, flares are 3-4 times as frequent on 2M0335+23 as on W1906+40. However, 2M0335+23 is 13 times more luminous than W1906+40, so despite its higher temperature, it is less efficient at converting its energy into flares; it also has seven times the surface area of W1906+40 so the flare rate per unit area is lower. It is intriguing that in the theory, W1906+40’s predicted magnetic field is $3.1$ kG, 40% stronger than 2M0335+23.
If we compute the total power in 2M0335+24 white light flares as the integral of $EdN$ from the solar microflare energy of $10^{24}$ erg to the observed superflare energy of $2.6\times10^{34}$ erg using the 2M0335+23 FFD fit, we find $\log L_{\rm WLF} / L_{\rm bol} = -4.2$ . A more conservative upper limit $10^{33}$, and lower limit of $5 \times 10^{28}$ erg still gives $\log L_{\rm WLF} / L_{\rm bol} = -4.7$.
The Pleiades Brown Dwarf Superflare
-----------------------------------
With an energy of $2.6 \times 10^{34}$ erg, the CFHT-PL-17 flare event is comparable to superflares observed in [*Kepler*]{} G, K, and M stars @2014ApJ...792...67C, though below the mean observed superflare energy [@2016ApJ...829...23D]. It is helpful to consider the flare in terms of the observed peak absolute $M_{Kp} = 10.3$ (long cadence) or $\sim8.7$ (short cadence): If this brown dwarf were an unresolved companion to an A star ($M_{Kp} \approx 0$) or F dwarf ($M_{Kp} \approx 2.5$) the flare would be a detectable event with [*Kepler*]{}. However, the flare rates seen in some A and F stars by @2012MNRAS.423.3420B seem to be too high to be explained by brown dwarf companions, since they re-occur on timescales of 1-120 days whereas we expect less than one per year due to a brown dwarf. X-ray triggered events have revealed that even more energetic superflares occur in young M dwarfs such as the $\sim30$-Myr old M dwarf binary [@2016arXiv160904674O]. For comparison to other ultracool dwarfs, this superflare has more energy than the L1 dwarf superflare we reported in Paper I of this series [@2017ApJ...838...22G] but less than the ASASAN-16ae L0 dwarf superflare [@2016ApJ...828L..22S].
As noted above, the extrapolated flare rate of 2M0335+23 suggests a superflare only $\sim1.4$ times per year. We find that there are nine well-resolved Pleaides brown dwarfs in the spectral type range M6-M9 Campaign 4 for which we would have been detected a similar superflare. The combined superflare rate of these nine brown dwarfs is $\sim1.7$ times per year. This suggests that the superflare may be understood as the high energy tail of the white light flare power-law distribution. However, @Rubenstein:2000nx suggested superflares in solar-type stars may be the result of interactions with a planetary companion, and we have no information about whether CFHT-PL-17 has a lower-mass brown dwarf or exoplanet companion. [**We see no reason to invoke interactions with a substellar companion to explain the white light flares in either 2M0335+23 or CFHT-PL-17.**]{}
Complex flares: Is sympathetic flaring at work?
-----------------------------------------------
The two 2M0335+23 complex flares (Fig \[fig-flares\], Table whatever) are very well described as the sum of two individual flares that follow the template, [**and we assume through this section that they are occurring on 2M0335+23 rather than an unknown companion.**]{} The time separation is 14.6 minutes for the flare on mission day 2240 and 9.2 minutes for the flare on mission day 2287. A third possible example of a complex flare occurs on mission day 2258, with a time separation of about 8 minutes. However, in the third case, the noise level is large enough that it is not altogether clear that two individual flares can be reliably identified. Thus, among the 22 flares illustrated in Fig. \[fig-flares\] , we can confidently state that about 10% exhibit the occurrence of two flares within a time interval shorter than 20 minutes.
Flares which occur within a short time interval of each other may belong to the class of “sympathetic flares" (SF). By definition, SF are related to each other in the sense that a disturbance (of a kind that we will discuss below) generated by the first flare propagates to another active region and triggers a flare there. However, a “short" time interval between flares is not necessarily an indication of SF. In fact, it may be difficult to identify with confidence a bona fide SF in certain stars. For example, a very active flare star may have multiple flares occurring randomly in multiple active regions within short time intervals, and these flares may have little or physical relationship to one another. Is there a way to distinguish between unrelated flares and SF? We suggest that one possible approach may be to consider the ratio of (i) the time interval T between two particular flares, and (ii) the mean time interval $T(m)$ between flares averaged over the length of the entire observing period.
For example, the M4 flare star GJ 1243 [@2014ApJ...797..122D] was observed by Kepler for a period of 11 months, during which 6107 “unique events" were identified as flares. This star has an average of 18-19 flares per day, i.e. $T(m) = 75-80$ minutes. An example of a “complex flare" is illustrated by @2014ApJ...797..122D in their Fig. 6, which spans an interval of 3.6 hours: 7 template flares are required to produce a good fit to the light curve. The average time interval between the template flares in this case is T = 31 minutes. This is already shorter than T(m) by a factor of 2, and might therefore suggest that SF could be at work. A fortiori, if we exclude one outlier flare at late times (with a peak at abscissa 549.865 days), inspection of their Fig. 6 suggests that 5 template flares occurred on GJ 1243 within a time interval of only 75 minutes, i.e. $T = 15$ minutes. This is 5-6 times shorter than $T(m)$, again suggestive of SF.
In contrast to the active flare star GJ 1243, when we return to considering the object of interest to us here (2M0335+23), we find that the flares in Fig. \[fig-flares\] were observed over a 66-day interval. This means that, with 21 flares in our sample, the mean interval between flares is $T(m) = 3$ days. The fact that we have identified two (or possibly 3) pairs of flares separated by only 20 minutes means that our pairs of flares have time separations T which are shorter than $0.01T(m)$. As a result, while we may assert that neighboring flares on 2M0335 are probably randomly related if they are separated by 3 days or more, it is much more difficult to make such an assertion for flares which are separated by less than 1% of T(m). It seems more likely to consider the possibility the two flares which are separated by only 1% of T(m) are physically related to each other. Specifically, is it possible that we are observing pairs of SF in 2M0335+23?
In the Sun, the SF possibility was subject to opposing claims in the 1930Õs based on optical data [@1937PASP...49..233R; @1938ZA.....16..276W]. Conflicting claims for the existence (or non-existence) of SF surfaced again in the 1970Õs, based on radio and X-ray data (@1976SoPh...48..275F: hereafter FCS). Based on X-ray data, FCS reported on the absence of significant evidence for SF except in one sub-set of their data: active regions which were closer to each other than a critical distance exhibited a 3.4$\sigma$ increase in the occurrence of “short" time intervals between flares. Coincidentally, FCS defined “short" as being $<20$ minutes, i.e. the same interval as we mentioned above in connection with flares on 2M0335+23. However, FCS seemed suspicious about even the one SF case they had detected because they could not identify “any mode of propagation of a triggering agent in the solar atmosphere."
In the case of stars, @1971Ap......7...48O\[OT\] analyzed the time intervals between successive flares in and and found that the intervals in general followed a Poisson distribution. However, in UV Cet, there exist certain “sequences of closely spaced flares whose probability of occurrence is very small in the case of a Poisson process." In one case, 7 flares occurred in 5.4 minutes, and on each of 2 separate occasions, 3 flares were observed within a 2-minute interval. OT demonstrated that such small intervals of time between flares are highly improbable in the context of the overall Poisson distribution.
Also in the case of the same flare star as that discussed by OT, \[HS\] reported on an independent study of the time intervals between flares on UV Cet. In the course of 26 hours of observing, they detected 94 flares. Thus, they obtained T(m) = 17 min as the average time between flares. However, when they examined the distribution of time intervals between individual flares, the intervals ranged up to as long as 110 min. A Poisson distribution was found to fit the flare interval data at the 98% confidence level with one proviso: only intervals larger than 4 min were included. At intervals shorter than 4 min, there is a large spike in the distribution: 38 of the 94 flares were found to have $T\le 4$ min. This excess at short times is far above what the Poisson distribution predicts. HS cited OT as having also “noted" this excess at short times. But then, with regards to the excess at short times, HS make the following statement (which has no explicit analog in OT): “This might be due to triggering of the second flare by the first, like sympathetic flares on the sun."
A possible triggering agent for SF in the Sun was suggested by @1968SoPh....4...30U: a fast-mode MHD wave/shock which is launched into the corona by certain flares. The idea is that as the wave/shock propagates through the corona, it may encounter a second active region: in that case, the wave/shock may perturb the second active region in such a way that a “sympathetic" flare occurs in that active region. Evidence for disturbances propagating away from certain flare sites was at first based solely on chromospheric data, where a “Moreton wave" was observed sweeping across the chromosphere. An observational break-through as regards a triggering agent for SF occurred with the launch of SOHO in 1995, when the Extreme-ultraviolet \[EUV\] Imaging Telescope (EIT) detected waves which propagate through large distances in the solar corona following certain events. The waves are referred to variously as “EIT waves" or “EUV waves." The waves were first interpreted as fast-mode MHD waves driven either by an erupting coronal mass ejection (CME) or as a blast wave driven by the energy release in a flare. The earliest data indicated that EUV waves propagate at speeds of 200-500 km/sec in the solar corona , but speeds as large as 1400 km/sec have been reported [@2013ApJ...776...58N]. An extensive survey of multiple theories which have been proposed to explain EUV waves [@2016arXiv161105505L] has concluded that the waves in the Sun are “best described as fast-mode large-amplitude waves or shocks that are initially driven by the impulsive expansion of an erupting CME in the low corona."
In the case of stellar (and brown dwarf) flares, is it possible that we might rely on solar-like phenomena to understand SF? Flares on stars involve magnetic energy release, so to the extent that a stellar flare contributes to the launch of a blast wave in the corona (analogous to the Sun), we may expect that flare-induced EUV waves could contribute to stellar SF. What about EUV waves generated by CMEÕs? Can we count on those to occur in flare stars and serve to launch EUV waves to help generate SF? Although flares and CMEÕs in the Sun both involve release of magnetic energy, they do not always occur together: one can occur without the other, depending on local details of the parent active region. We note that at least one detection of a stellar CME has been reported from an active K dwarf which is known to be a flare star [@2001ApJ...560..919B].
Let us examine the hypothesis that the 2 complex flares which we have detected in 2M0335+23 involve SF which are triggered by the equivalent of an EUV wave. In this context, the maximum speed of the wave would be obtained if the two active regions which are involved in the individual flares were located at antipodal points on the surface of 2M0335+23, at a distance $\pi R$ from each other. Inserting the radius $R$ of 2M0335+23, and inserting time delays of 14.6 and 9.2 minutes, the SF hypothesis leads to $v(EUV) < 600$ and $950$ km s$^{-1}$. Such values are well within the range of EUV wave speeds which have been reported in the solar corona. With fast-mode speeds determined mainly by the Alfven speed (which greatly exceeds the thermal speed of order 100-200 km/sec in a 1-2 MK corona), our SF interpretation suggests that Alfven speeds in the corona of 2M0335+23 may not differ greatly from those in the solar corona. In the latter, a map of coronal Alfven speeds reported by @2015ApJ...812..119S spans a range from 500 to 900 km/sec at essentially all latitudes within a radial distance of 5 solar radii.
Complex flares: Does the weakness of the second flare contain physical information?
-----------------------------------------------------------------------------------
We note that for the complex flares in Figure \[fig-flares\], when the tail of the first light curve is subtracted from the second flare, the amplitude at the peak of the second flare is smaller than the amplitude at the peak of the first flare. We ask: Is this “weaker secondary" a common feature of complex flares?
To address this, we consider some flare data which were recorded in different settings.
1. A large homogeneous sample of X-ray flares which were recorded by Chandra in the Orion Ultradeep Project [@2008ApJ...688..418G]. In their study of the 216 brightest flares from 161 pre-main sequence stars, 8 events were classified as ÒdoubleÓ flares, i.e. they Òlook like two overlapped typical flaresÓ. By subtracting off the tail of the first flare in each case, we evaluated the ratio of peak 2 to peak 1, and we found the following values: 0.3, 0.4, 0.2, 0.5, 1.1, 0.2, 0.3, and 0.1. Thus, in 7 out of 8 cases, the Òsympathetic flareÓ has a smaller amplitude than the original flare.
2. Optical data (in the $r$ band) for stars in the intermediate age cluster M37 (0.55 Gyr) resulted in detection of several hundred flares from cluster members [@2015AJ....150...27C; @2015ApJ...814...35C]. @2015ApJ...814...35C drew attention to the result that their Òalgorithm often detects secondary flares which occur during the decay of a much larger flareÓ. Visual inspection of Fig. 2 in @2015ApJ...814...35C and Figs. 17, 18 in @2015AJ....150...27C suggests that as many as 8 or 9 secondary flares can be identified among the plotted light curves of 23 stars: in all cases but one, the flare which occurred later in time was smaller in amplitude than the first flare.
3. A small sample (15-20) of light curves in a variety of visible and near UV wavelengths from a number of solar neighborhood flare stars has been presented by @Gershberg:2005lr\[,pp. 195-205\] Secondary flares can be identified in at least 5 cases, and in all cases, the flare which occurred later in time had a smaller peak intensity. These examples suggest that, no matter which wavelength range we examine, the flare which occurs later in time (the “secondary") in a “close double" has (in most cases) a smaller amplitude than the flare which occurs earlier. We believe that this is a feature which contains information related to one of the key physical processes involved in sympathetic flaring.
How might this “weaker secondary" behavior be understood in the context of the SF explanation proposed above (i.e. SF is triggered by a fast-mode MHD wave)? We suggest that it may be understood in terms of wave refraction. When fast-mode MHD waves propagate through an inhomogeneous medium, the compressive nature of the waves has the effect that the waves are refracted away from regions of high Alfven speed ($v_A$), and are refracted into regions where $v_A$ is small [@1968SoPh....4...30U]. (Alfven waves, lacking compression, refract differently.) Tests of UchidaÕs predictions have been provided by observations (e.g. Shen et al. 2013) and modelling (e.g, @1980ApJ...241.1186S,@2016ApJ...820...16J) of fast-mode waves propagating through a variety of structures in the solar corona. However, the tests mentioned in the preceding sentence were based on indirect inferences of the $v_A$ value in different regions of the Sun. For more reliable tests of UchidaÕs theory, it is preferable to consider a medium where in situ measurements of field strength and ion density can be made directly: in such a medium, the local value of $v_A$ can be calculated, various wave modes can be distinguished (fast MHD, slow MHD, Alfvenic), and UchidaÕs predictions can be tested directly. The solar wind is one such medium. In the solar wind at a radial distance of about 1 AU, data from the ACE satellite have been used to demonstrate that fast mode waves are indeed depleted in [**high-$v_A$ regions [@2001JGR...10618625S]**]{}, and fast mode waves are indeed enhanced in regions of low-$v_A$ [@2003ApJ...583..496M].
Now let us consider how these results might find an application in SF in low-mass stars. Suppose an initial flare is triggered (somehow) in a certain active region (AR), thereby launching a fast-mode EUV wave with a certain speed: the speed will be related to the $v_A$ value in the AR where the flare was initiated. Suppose there are two other ARÕs on the surface of the star, AR-A with large $v_A$, and AR-B with small $v_A$. What happens when the EUV wave approaches AR-A? The fast-mode wave will be refracted away from AR-A because of the locally large $v_A$ : the wave will be unable to penetrate into AR-A. Therefore, a SF is unlikely to occur in AR-A. But in the case of AR-B, the fast-mode wave will be refracted into the AR, thereby having a chance to perturb the plasma inside AR-B, perhaps enough to initiate a flare. In this scenario, a SF is more likely to occur in an AR with a small value $v_A$: or, in the terminology of @2016ApJ...820...16J, the “impact" of the wave on AR-B (with its lower $v_A$ ) would be larger. What might cause the $v_A$ value in AR-B to be smaller? It could be either weaker field or higher density, or both. In cases where the field is weaker, we expect that (other things being equal) a flare in such an AR will have (in general) a smaller total energy. Thus the flare originating in AR-B (where the “impact" of the fast-mode wave is largest) will be “weaker," and is expected to have a smaller peak amplitude. If this is a correct interpretation of SF in stars, then the phenomenon of the “weaker secondary" may provide an observational signature of the physics of refraction of fast-mode MHD waves in the inhomogeneous corona of a flare star.
Conclusions
===========
White light flares on a young (24 Myr) M7-type brown dwarf are similar in most respects to stellar M dwarf flares, including their light curves, power-law flare frequency distribution, and sympathetic flaring. Adding a flare on a Pleiades brown dwarf, we see that these flares extend up to at least $2.6 \times 10^{34}$ erg. However, we observe no white light flares on the L$5\gamma$ brown dwarf despite its known young age and rapid rotation. Since there is overwhelming observational and theoretical evidence that magnetic fields exist on L and T-type brown dwarfs, we conclude that the change in flare rates is direct evidence that fast magnetic reconnection is suppressed or forbidden at temperatures $\sim1500$K. In this work, we have studied brown dwarfs of known age. In our next paper, we will measure the white light flare rates of a sample of field late-M and L dwarfs and investigate their dependence on age, rotation, effective temperature, and other observable properties.
We thank James Davenport and Rachel Osten for discussions of stellar flares, Jonathan [Gagn[é]{}]{} and Jackie Faherty for discussions of moving groups, Mike Liu and Conard Dahn for comments on the preprint, and the anonymous referee and statistical consultant for suggestions. This paper includes data collected by the Kepler mission. Funding for the Kepler mission is provided by the NASA Science Mission directorate. The material is based upon work supported by NASA under award Nos. NNX15AV64G, NNX16AE55G, and NNX16AJ22G. A.J.B. acknowledges funding support from the National Science Foundation under award No. AST-1517177. Some of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX09AF08G and by other grants and contracts. Some of the data presented herein were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation. This research has made use of NASA’s Astrophysics Data System, the VizieR catalogue access tool, CDS, Strasbourg, France, and the NASA/ IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA. We have also made use of the maintained by Jonathan [Gagn[é]{}]{}.
[^1]: EPIC 211046195 and EPIC 210327027 were observed for K2 Guest Observer Program 4036 (PI Gizis); EPIC 211110493 was observed for GO Programs 4024 (PI Lodieu), 4026 (PI Scholz), and 4081 (PI Demory).
[^2]: We also identify a flare in the candidate Pleiades brown dwarf [@2000MNRAS.313..347P], observed as EPIC 211000317. , however, have measured its proper motion and assign it a membership probability of zero.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Markus Rösch
- Giacomo Scalari
- Gustavo Villares
- Lorenzo Bosco
- Mattias Beck
- Jérôme Faist
title: 'Supplementary Data: On-chip, self-detected THz dual-comb spectrometer'
---
Both lasers have been characterized individually. Supplementary figure \[fig:BNmap\](a) and (b) show a beatnote map for the two lasers. Similar to previous work with such lasers, the comb regime is limited to certain current ranges which are slightly different for different lasers [@Roesch2014; @li2015]. The beatnote maps (supplementary figure \[fig:BNmap\](a), (b)) were measured in continuous wave (CW) at a fixed temperature of 23 Kelvin. Only the laser under test was switched on during the measurement. The laser was left free running without any active stabilization. A commercial power supply (Keithley instruments) has been used as a current source. As the current stability of this power supply is limited, a homemade low-pass filter has been employed to reduce fast current fluctuations on the laser bias. The beatnote was measured by picking up the RF signal from the current line. It is extracted with a bias-tee and the RF part is sent to a microwave spectrum analyzer (Rohde & Schwarz FSU50). To acquire the beatnote maps (beatnote as a function of driving current) the spectrum analyzer was set with a resolution bandwidth (RBW) of 10 kHz and a video bandwidth (VBW) of 1MHz. The traces were recorded in a single sweep mode for every current step (5mA). The signal was amplified with a 35dB RF amplifier (Miteq AFS43-01002600-38-8P-44) and by another 35 dB using the internal preamplifier of the spectrum analyzer.\
The operation conditions of the two lasers are not independent. This also affects the comb regimes of the lasers. When we are operating both lasers at the same time, the comb regimes eventually change. In supplementary figure \[fig:BNmap\](c) one laser was kept at a fixed current (316 mA) while the other laser was swept over the full current range. For low currents of the swept laser, the fixed laser is in a comb regime (Beatnote at 21.75 GHz in supplementary figure \[fig:BNmap\](c)). It leaves this regime when the swept laser is driven at a higher current ($>275$mA), although the temperature is fixed at a constant value of 23 Kelvin. But also the swept laser changes its behavior. For example the comb regime at 340mA is no longer there in the presence of the second laser.\
[2]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} [****, ()](\doibase 10.1364/OE.23.033270)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Quantum walks are considered in a one-dimensional random medium characterized by static or dynamic disorder. Quantum interference for static disorder can lead to Anderson localization which completely hinders the quantum walk and it is contrasted with the decoherence effect of dynamic disorder having strength $W$, where a quantum to classical crossover at time $t_{c}\propto W^{-2}$ transforms the quantum walk into an ordinary random walk with diffusive spreading. We demonstrate these localization and decoherence phenomena in quantum carpets of the observed time evolution and examine in detail a dimer lattice which corresponds to a single qubit subject to randomness.'
author:
- 'Yue Yin, D.E. Katsanos and S.N. Evangelou[^1]'
title: Quantum Walks on a Random Environment
---
I. INTRODUCTION
===============
During last decade quantum algorithms were proposed, such as Grover’s search[@r1] and Shor’s factorization[@r2], which can in principle perform certain computational tasks quantum-mechanically, much more efficiently than their classical counterparts. The related idea of quantum walks was also introduced[@r3; @r4; @r5; @r6] which generalize the classical random walks widely used in various computations as the basis of classical algorithms. The quantum walks are similar to classical random walks but with a “quantum coin” operation which replaces the coin-flip randomness in between each moving step on a lattice or a graph. The state of the quantum coin which uniquely determines the subsequent movement direction can also exist in quantum superpositions, something impossible in the classical domain where the coin has a specific outcome. In analogy with classical random walks the quantum walks are expected to be useful for designing quantum algorithms. For example, Grover’s algorithm can be combined with quantum walks in a quantum algorithm for “glued trees” which provides even an exponential speed up over classical methods[@r7].
The main advantage of quantum walks is a highly improved behavior over their classical counterparts since quantum wave propagation is superior than classical diffusion. For example, the ballistic mean square variance $\sigma^{2}(t)\propto t^{2}$ in the quantum case can be compared to the linear spread law $\sigma^{2}(t)\propto t$ of classical diffusion. This quadratic speed-up is a general feature of quantum search algorithms[@r1] and is also familiar from standard quantum evolution of tight-binding electron waves on a periodic lattice[@r8]. In quantum walks the classical probabilities $P(x,t)$ are replaced by complex probability amplitudes $\Psi(x,t)$ computed from the unitary dynamics of the Schr$\ddot{o}$dinger’s equation. The corresponding probability amplitudes are determined by summing up over all possible paths of propagation. Furthermore, to describe wave propagation in lattices or graphs one does not need a “quantum coin”[@r9] and a related continuous-time versions of quantum walks have been introduced[@r10]. The discrete and continuous-time quantum walks have recently been related to Dirac and Schr$\ddot{o}$dinger’s equations, respectively[@r11].
We consider a continuous-time quantum walk via the equivalent problem of a quantum particle initially localized in one-dimensional lattice in the presence of static or dynamic disorder. For a tight-binding electron at an integer lattice site labeled by $x$ in one dimension at time $t$ the wave function $\Psi(x,t)$ obeys the linear wave equation ($\hbar=1$) $$\imath \partial {\Psi(x,t)}/\partial{t} =
\epsilon(x,t) \Psi(x,t)+\Psi(x-1,t)+\Psi(x+1,t),$$ with $\epsilon(x,t)$ an $x$-dependent random variable for static disorder which is also $t$-dependent for dynamic disorder, where lengths are measured in units of the lattice spacing and energies or inverse times in units of the hopping integral. In order to study quantum walks via Eq. (1) we choose the initial condition of a particle at the origin $x=0$ with $\Psi(x,t=0)= \delta_{x,0}$ and characterize the quantum motion by the second moment for its position $\sigma^{2}(t) = \sum_{x}|x|^{2} P(x,t)$, where $P(x,t)=|\Psi(x,t)|^{2}$ is the probability density. In the absence of disorder ($\epsilon(x,t)=0$) the amplitude is given by the Bessel function so that $P(x,t)= (J_{x}(2 t))^{2}$, where the order $x$ of the Bessel function measures the distance travelled from the origin while its argument is proportional to time $t$. The evolution of the wave-packet at time $t$ was shown[@r8] to display two sharp ballistic fronts at $x=\pm2t$. From properties of the Bessel functions inside the spatial region $[-2t,2t]$ the probability density $P(x,t)$ is an oscillating function multiplied by $1/t$ while outside this region, denoted by the two ballistic peaks, $P(x,t)$ decays exponentially. The ballistic mean-square displacement is $\sigma^{2}(t) = 2 t^{2}$.
In the absence of disorder the quantum evolution of wave-packets which occurs via the evolution operator $\exp(-iHt)$ with Hamiltonian $H$ is very different from classical diffusion where any initial state converges to a Gaussian steady state. An initially squeezed $\delta$-function spatial wave-packet has reduced spatial uncertainty and behaves like a quantum particle consisting of all the eigenstates of $H$. Alternatively, a spatially uniformly distributed initial wave-packet $\Psi(x,0)=1/\sqrt{N}$, for every $x$ in an $N$-site chain, consists of few eigenstates of $H$ near the lower band edge only. Since the latter choice emphasizes states from the band edge the semi-classical asymptotics is relevant[@r12; @r13; @r14]. The wave-packets which initially consist of many eigenstates can be related via the uncertainty principle to ultrashort laser pulses of femtosecond duration. The evolution of such coherent superpositions of quantum states is realized in the physics of trapped atoms in optical lattices[@r15], trapped ions[@r16], etc.
Classical random walks in perfect one-dimensional lattices are defined by the probabilities $p_{x}=q_{x}=1/2$ which determine the random left and right motion from site $x$. This externally induced randomization leads to classical diffusion for long $t$ with a Gaussian $P(x,t)$ and $\sigma^{2}(t) = 2D t$, $D$ the diffusion coefficient. For random walks in random media the probabilities $p_{x},q_{x}=1-p_{x}$ become random variables themselves, for example, they could be chosen from a flat distribution within $(0,1)$. This is the so-called random random or Sinai walk[@r17; @r18] which leads to ultra slow classical evolution $\sigma^{2}(t)\propto \ln t^{4}$, very close to a complete cease.
The problem addressed in this paper concerns the fate of quantum walks in a random environment, with both static and dynamic disorder. To answer this question for static disorder we shall combine previous knowledge from the field of wave propagation in the presence of randomness where the quantum phenomenon of Anderson localization[@r19] takes place (for its consequences for quantum walks see[@r20]). We shall show that static disorder is responsible for exponentially suppressed quantum evolution with variance $\sigma^{2}(t)$ reaching a time-independent limit for long $t$, depending on the strength of static disorder and space dimensionality. Surprisingly, classical random walks for static disorder are still propagating, although ultraslowly[@r17; @r18]. Other generalizations of discrete-time quantum walks in aperiodic or fractal media by using biased quantum coins have given slower (sub-ballistic) quantum evolution[@r21]. For dynamic disorder by coupling the quantum system to a random environment decoherence occurs[@r22] and quantum physics becomes classical so that a quantum walk is still propagating but only diffusively.
The main reason for examining the robustness of quantum walks in the presence of noise is because disorder is unavoidable in most quantum systems. Static disorder also appears for electrons in lattices with permanent modifications due to impurities. The dynamic disorder in this case could be driven by time-dependent vibrations of the lattice atoms which have an impact on the electronic site-energies. Apart from describing such electron-phonon interactions, dynamic disorder also addresses the presence of time-dependent noise in the memory qubits of quantum computers. In this paper we demonstrate that although static disorder hinders the motion of quantum walks due to negative quantum interference from Anderson localization via multiple scattering from impurities, instead, for a dynamically random environment the time-dependent disorder acts as a decoherence mechanism at a crossover time $t_{c}$ which randomizes the quantum walks and turns the quantum motion into classical. The discrete finite space chosen in the simulations could be also used on a finite computer. This general scheme for discretizing space is not only suggested by solid state applications but it is, somehow, related to the discreteness of the quantum information itself.
The paper is organized as follows: In Sec. I we introduced the reader to the subject of quantum communication by setting the aims of our quantum walk simulation in random media. In Sec. II we briefly review the properties of quantum walks in ballistic and disordered one-dimensional media by showing quantum carpets which demonstrate the difference between static and dynamic disorder. Static disorder is shown to be responsible for negative quantum interference of Anderson localization which stops completely the quantum walk while dynamic disorder permits only diffusive evolution of classical random walks. In Sec. III we display the quantum to classical crossover for dynamic disorder and consider a qubit subject to dynamic disorder. Finally, in Sec. IV we summarize our main conclusions.
II. QUANTUM CARPETS
====================
We have created space-time $x-t$ structures for the probability density $P(x,t)$ on a one-dimensional finite $N$-site orthonormal lattice space without disorder, with $t$-independent static disorder and also rapidly varying dynamic disorder. The white color in the figures denotes high probability density and the darker colors lower values. In Fig. 1 a state is initially released in the middle of the chain and the probability density $P(x,t)$ is obtained by solving Eq. (1). In Fig. 2 the same is done for a spatially uniform initial state. For static disorder Eq. (1) could be alternatively solved by considering the time-evolution of a state vector $|\Psi(0)\rangle$ expressing the probability density via the stationary eigen-solutions $H|j\rangle=E_{j}|j\rangle$ with space-time wave function $$\Psi(x,t) = \sum_{j=1}^{N}e^{-iE_{j}t} \psi_{j}(x)
\langle j|\Psi(0)\rangle ,$$ with the amplitude on site $x$ denoted by $\psi_{j}(x) = \langle
x|j\rangle$. For time-dependent disorder Eq. (1) was solved via a fourth-order Runge-Kutta algorithm.
In Fig. 1 we present our results for the initial choice of a $\delta$-function in the middle of the chain with $|\Psi(0)\rangle=|0\rangle$ and in Fig. 2 for a uniform initial state $|\Psi(0)\rangle=(1/\sqrt{N}) \sum_{x=1}^{N}|x\rangle$ is computed on finite chains where the wave-packet scatters from their hard ends. In the rest of Figs. 3-8 our results are obtained for a self-expanding chain with a $\delta$-function initial choice to make sure that the wave-packet does not reach the boundaries. This allows to study the quantum to classical crossover by computing the mean-square-variance and the autocorrelation function vs. time.
![Quantum carpets which show the probability density $P(x,t)$, for $x$ in the vertical axis and $t$ in the horizontal axis, are generated by an [**initial $\delta$-type**]{} spatial state for $t=0$ $\Psi(x,0)=\delta_{x,0}$ in the middle (left of the figure) with chain length $N=101$. [**(a)**]{} The ballistic case for the absence of disorder where perfect quantum revivals can be clearly seen. [**(b)**]{} For static disorder of strength $W=1.5$ quantum interference causes Anderson localization which stops the quantum motion and the probability to stay in the initial site remains high. For stronger static disorder this probability becomes even higher. [**(c,d)**]{} For dynamic disorder with values $W=1$ and $W=5$, one can see quantum interference only initially for small $t$ on the left hand side of the figure where the $\delta$-type wave-packet moves ballistically. After certain time (for $W=1$ is estimated $t_{c}\approx 60$ and for $W=5$ about $t_{c}\approx 3$) the quantum interference is lost while the particle still moves but classically.](fig2a.eps "fig:"){width="8.0cm"} ![Quantum carpets which show the probability density $P(x,t)$, for $x$ in the vertical axis and $t$ in the horizontal axis, are generated by an [**initial $\delta$-type**]{} spatial state for $t=0$ $\Psi(x,0)=\delta_{x,0}$ in the middle (left of the figure) with chain length $N=101$. [**(a)**]{} The ballistic case for the absence of disorder where perfect quantum revivals can be clearly seen. [**(b)**]{} For static disorder of strength $W=1.5$ quantum interference causes Anderson localization which stops the quantum motion and the probability to stay in the initial site remains high. For stronger static disorder this probability becomes even higher. [**(c,d)**]{} For dynamic disorder with values $W=1$ and $W=5$, one can see quantum interference only initially for small $t$ on the left hand side of the figure where the $\delta$-type wave-packet moves ballistically. After certain time (for $W=1$ is estimated $t_{c}\approx 60$ and for $W=5$ about $t_{c}\approx 3$) the quantum interference is lost while the particle still moves but classically.](fig2b.eps "fig:"){width="8.0cm"} ![Quantum carpets which show the probability density $P(x,t)$, for $x$ in the vertical axis and $t$ in the horizontal axis, are generated by an [**initial $\delta$-type**]{} spatial state for $t=0$ $\Psi(x,0)=\delta_{x,0}$ in the middle (left of the figure) with chain length $N=101$. [**(a)**]{} The ballistic case for the absence of disorder where perfect quantum revivals can be clearly seen. [**(b)**]{} For static disorder of strength $W=1.5$ quantum interference causes Anderson localization which stops the quantum motion and the probability to stay in the initial site remains high. For stronger static disorder this probability becomes even higher. [**(c,d)**]{} For dynamic disorder with values $W=1$ and $W=5$, one can see quantum interference only initially for small $t$ on the left hand side of the figure where the $\delta$-type wave-packet moves ballistically. After certain time (for $W=1$ is estimated $t_{c}\approx 60$ and for $W=5$ about $t_{c}\approx 3$) the quantum interference is lost while the particle still moves but classically.](fig2c.eps "fig:"){width="8.0cm"} ![Quantum carpets which show the probability density $P(x,t)$, for $x$ in the vertical axis and $t$ in the horizontal axis, are generated by an [**initial $\delta$-type**]{} spatial state for $t=0$ $\Psi(x,0)=\delta_{x,0}$ in the middle (left of the figure) with chain length $N=101$. [**(a)**]{} The ballistic case for the absence of disorder where perfect quantum revivals can be clearly seen. [**(b)**]{} For static disorder of strength $W=1.5$ quantum interference causes Anderson localization which stops the quantum motion and the probability to stay in the initial site remains high. For stronger static disorder this probability becomes even higher. [**(c,d)**]{} For dynamic disorder with values $W=1$ and $W=5$, one can see quantum interference only initially for small $t$ on the left hand side of the figure where the $\delta$-type wave-packet moves ballistically. After certain time (for $W=1$ is estimated $t_{c}\approx 60$ and for $W=5$ about $t_{c}\approx 3$) the quantum interference is lost while the particle still moves but classically.](fig2d.eps "fig:"){width="8.0cm"}
![Quantum carpets showing the probability density $P(x,t)$ for a [**spatially uniform**]{} initial state $\Psi(x,0)=1/\sqrt{N}$ on a chain of length $N=101$. [**(a)**]{} For the ballistic case in the absence of disorder one can see recurrences for short times which disappear for longer times where quantum interference effects become apparent. [**(b)**]{} For static disorder of strength $W=5$ Anderson localization occurs with localization length $\xi\sim 100W^{-2}\sim 4$ much less than the system size $N=101$. The displayed regions of high amplitudes show the positions where the particle localizes. [**(c)**]{} In the presence of dynamic disorder $W=5$ the quantum interference effects vanish. The main difference between dynamic disorder (c) and static disorder (b) is that in (c) the regions with high values of $P(x,t)$ keep changing leading to randomization so that the particle can still move but in a “classical” fashion.](fig1a.eps "fig:"){width="8.0cm"} ![Quantum carpets showing the probability density $P(x,t)$ for a [**spatially uniform**]{} initial state $\Psi(x,0)=1/\sqrt{N}$ on a chain of length $N=101$. [**(a)**]{} For the ballistic case in the absence of disorder one can see recurrences for short times which disappear for longer times where quantum interference effects become apparent. [**(b)**]{} For static disorder of strength $W=5$ Anderson localization occurs with localization length $\xi\sim 100W^{-2}\sim 4$ much less than the system size $N=101$. The displayed regions of high amplitudes show the positions where the particle localizes. [**(c)**]{} In the presence of dynamic disorder $W=5$ the quantum interference effects vanish. The main difference between dynamic disorder (c) and static disorder (b) is that in (c) the regions with high values of $P(x,t)$ keep changing leading to randomization so that the particle can still move but in a “classical” fashion.](fig1b.eps "fig:"){width="8.0cm"} ![Quantum carpets showing the probability density $P(x,t)$ for a [**spatially uniform**]{} initial state $\Psi(x,0)=1/\sqrt{N}$ on a chain of length $N=101$. [**(a)**]{} For the ballistic case in the absence of disorder one can see recurrences for short times which disappear for longer times where quantum interference effects become apparent. [**(b)**]{} For static disorder of strength $W=5$ Anderson localization occurs with localization length $\xi\sim 100W^{-2}\sim 4$ much less than the system size $N=101$. The displayed regions of high amplitudes show the positions where the particle localizes. [**(c)**]{} In the presence of dynamic disorder $W=5$ the quantum interference effects vanish. The main difference between dynamic disorder (c) and static disorder (b) is that in (c) the regions with high values of $P(x,t)$ keep changing leading to randomization so that the particle can still move but in a “classical” fashion.](fig1c.eps "fig:"){width="8.0cm"}
ballistic motion
----------------
In this case, obviously, quantum walks perform at their best. The ballistic description is valid for solid state systems in the absence of disorder which refers to the motion of a point particle in an $N$-site chain with $\epsilon=0$ in Eq. (1) which gives $E_{j}=2\cos({\frac{\pi j}{N+1}})$ and $ \psi_{j}(x) =\sqrt{{\frac
{2}{N+1}}} \sin({{\frac {{j\pi}{x}}{N+1}}})$, $j=1,2,\ldots,N$. The corresponding space-time pictures are shown in Figs. 1(a) and 2(a). Quantum revivals can be seen where the particle returns to its initial position and reconstructs like a classical particle which moves with constant velocity reflecting at the boundaries of the chain[@r14]. We observe that the quantum revivals of Fig. 1(a) and Fig. 2(a) do not repeat indefinitely but become less and less accurately as time progresses. This is due to effects from boundary scattering which become more prominent for broad wave-packets in the right hand-side of Fig. 2(a) where noisy evolution is established. The obtained fractal pattern is a result of peculiar quantum interference effects due to scattering from the hard walls at the ends of the chain[@r14].
static disorder
---------------
It can have dramatic consequences for quantum walks, particularly in low dimensions, since for static disorder they can stop completely due to Anderson localization. From Figs. 1(b) and 2(b) we can see how strong static disorder with $\epsilon(x)$ chosen from a uniform probability distribution within $[-W/2,+W/2]$ causes destructive interference with Anderson localization. This crossover from ballistic motion to localization has dramatic consequences for quantum walks in one-dimension in the presence of static disorder. In higher dimensions the Anderson transition from extended to localized states is expected, via an intermediate chaotic regime which is rather better for quantum propagation. The probability density of Fig. 1(b) is shown to stay around the middle site where the initial wave-packet has maximum amplitude and it remains there for longer times. For the uniform initial state of Fig. 2(b) the larger amplitudes remain on many sites indefinitely.
dynamic disorder
----------------
A quantum walk can operate in the presence of dynamic disorder but only for short times since for longer times its motion becomes entirely classical, indistinguishable from an ordinary random walk. The effect of dynamic disorder is equivalent to introducing coin chaos which makes the quantum coherence disappear[@r22]. In order to see this decoherence effect we have chosen a random $\epsilon(x,t)$ rapidly varying with $t$ by:
\(i) Updating at random the diagonal site energies $\epsilon(x,t)$ at a time length comparable and often much smaller to the time step of the numerical method.
\(ii) Varying the diagonal energies by $$\epsilon(x,t)=amp* \cos(\omega_{x}t+\phi_{x})$$ where $amp$, $\omega$ and $\phi$ are the amplitude, frequency and phase for the motion of levels, respectively. We chose to vary the frequency $\omega$ at random uniformly within the interval $[0,2\pi]$ and fixed the phase $\phi_{x}$ to zero.
From Figs. 1(c), (d) and 2(c), we can see the effect of dynamic disorder for the two initial wave-packets, $\delta$- function and broad, respectively. In Fig. 1(c) the quantum motion seen on the left hand side of the figure quickly disappears and this also happens in Fig. 1(d) where classical diffusive motion is seen to arise. In Fig. 2(c) the randomization effects of dynamic disorder become more obvious and could be contrasted with quantum localization for static disorder (Fig. 2(b)). The quantum-to-classical crossover takes place after a characteristic time $t_{c}\propto W^{-2}$.
III. QUANTUM TO CLASSICAL CROSSOVER FOR DYNAMIC DISORDER
========================================================
decoherence in an $N$-site chain
--------------------------------
{width="6.0cm"}
The decoherence effect of dynamical disorder which turns the quantum wave propagation into classical diffusion is shown in Fig. 3 for a self-expanding chain. The probability density $P(x,t)$ gradually changes from a shape displaying two ballistic peaks of the quantum wave for small $t$ towards a Gaussian for large $t$. In Fig. 4 the quantum to classical crossover is shown for the mean-square-variance $\sigma^{2}(t)$ and the autocorrelation function or return probability $C(t)={\frac
{1}{t}}\int_{0}^{t}P(0,t')dt'$. Eventually, the classical asymptotic laws $\sigma^{2}(t)\propto t$ and $C(t)\propto
t^{-1/2}$ set in after an initial period of quantum ballistic motion where $\sigma^{2}(t)\propto t^{2}$ and $C(t)\propto
t^{-1}$. In Fig. 5 the effect of sinusoidal dynamic disorder is considered with constant amplitude and randomly varying the phase. The results are similar to Fig. 4 with the approach to the classical limit even faster in this case. From Figs. 4 and 5, except for no difference between the two types of dynamic disorder, we find that the crossover region between the ballistic law for small times and the diffusive law for long times is smooth having a mixed quantum and classical character.
![[**(a)**]{} Log-log plot of the mean-square -displacement or variance $\sigma^{2}(t)$ vs. time $t$ for various values for the dynamic disorder $W$. The crossover from ballistic quantum motion to diffusive classical motion occurs at earlier times as $W$ increases. [**(b)**]{} The autocorrelation function or “return to the origin” probability $C(t)$ vs. $t$ for $W=20$ displays classical diffusive behavior.](tdd-w-all-msd.eps "fig:"){width="6.0cm"} ![[**(a)**]{} Log-log plot of the mean-square -displacement or variance $\sigma^{2}(t)$ vs. time $t$ for various values for the dynamic disorder $W$. The crossover from ballistic quantum motion to diffusive classical motion occurs at earlier times as $W$ increases. [**(b)**]{} The autocorrelation function or “return to the origin” probability $C(t)$ vs. $t$ for $W=20$ displays classical diffusive behavior.](acf-w20.eps "fig:"){width="6.0cm"}
![ [**(a)**]{} The $\sigma^{2}(t)$ vs. time $t$ same as in Fig. 4(a) but for the sinusoidal dynamic disorder of Eq. (4). [**(b)**]{} The $C(t))$ vs. time $t$ same as in Fig. 4(b) but for the sinusoidal dynamic disorder of Eq. (4).](fig-3a.eps "fig:"){width="6.0cm"} ![ [**(a)**]{} The $\sigma^{2}(t)$ vs. time $t$ same as in Fig. 4(a) but for the sinusoidal dynamic disorder of Eq. (4). [**(b)**]{} The $C(t))$ vs. time $t$ same as in Fig. 4(b) but for the sinusoidal dynamic disorder of Eq. (4).](fig-3b.eps "fig:"){width="6.0cm"}
The effect of dynamic disorder on the quantum evolution is displayed in the linear and log plots of Fig. 6 which show the snapshots of the evolving spatial wave for an initial $\delta$-function in a self-expanding chain. The decoherence effect of dynamical disorder is seen from the rapid approach to a Gaussian shape. The complete phase diagram is summarized in Fig. 7 with $t_{c}$ vs the strength of dynamic disorder $W$ which displays a wide crossover grey color region where the law $t_{c}\propto W^{-2}$ is approximately obeyed.
![The snapshots for dynamic disorder of strength W=20 in a self-expanding chain where the approach to a Gaussian is seen. [**(a)**]{} The probability density $P(x,t)=|\Psi(x,t)|^{2}$ as a function of space $x$ at fixed times $t=5000$, $t=20000$, $t=60000$ and $t=80000$. [**(b)**]{} The same as in (a) but for the log of the amplitude. ](snap-w20.eps "fig:"){width="6.0cm"} ![The snapshots for dynamic disorder of strength W=20 in a self-expanding chain where the approach to a Gaussian is seen. [**(a)**]{} The probability density $P(x,t)=|\Psi(x,t)|^{2}$ as a function of space $x$ at fixed times $t=5000$, $t=20000$, $t=60000$ and $t=80000$. [**(b)**]{} The same as in (a) but for the log of the amplitude. ](snap-log-w20.eps "fig:"){width="6.0cm"}
{width="6.0cm"}
decoherence in a qubit
----------------------
We have examined in detail the quantum walk in a random two-level system ($N=2$) which has recently attracted attention in the context of quantum information processing. The operation of qubit and logical gates in the presence of a noisy environment is important for understanding quantum computers. The usual noise for such two-level system is usually due to various sources while non-Gaussian randomness arises from hopping background charges for different statistically independent fluctuators.
The time-dependent Hamiltonian is $$H= \left( \begin{array}{cc}
\epsilon_{1}(t) & \gamma \\
\gamma & \epsilon_{2}(t) \\
\end{array}
\right)$$ with random diagonal terms defined by $$\langle \epsilon_{i}(t) \rangle =0,
\langle \epsilon_{i}(t)\epsilon_{j}(t') \rangle =\delta
\delta_{i,j}\delta(t-t'),i,j=1,2,$$ where $\delta=W^{2}/12$ measures the disorder chosen from a box distribution within $[-W/2,+W/2]$. The averaged matrix elements of the density matrix $\bf{\rho}$ can be obtained from a decoupling suggested in[@r23] $$\imath\overline{\dot{\rho_{11}}}=
-\imath\overline{\dot{\rho_{22}}}=\gamma
(\overline{\rho_{21}}-\overline{\rho_{12}})$$ $$\imath\overline{\dot{\rho_{12}}}=-2\imath \delta
\overline{\rho_{12}}+\gamma(\overline{\rho_{22}}-\overline{\rho_{11}}))$$ and if we define $$\overline{\rho_{11}}={\frac {1}{2}} +\rho,
\overline{\rho_{22}}={\frac {1}{2}} -\rho,
\overline{\rho_{12}}=R+iJ,
\overline{\rho_{21}}=R-iJ,$$ the corresponding equations become $$\dot\rho=-2\gamma J,
\dot R=-2\delta R,
\dot J=-2\delta J + 2\gamma \rho.$$ Their general solutions (with appropriate constants) $$\left(
\begin{array}{c}
\rho \\
J \\
\end{array}
\right)
=C_{+} \left(
\begin{array}{c}
-2\gamma \\
\Lambda_{+} \\
\end{array}
\right) e^{\Lambda_{+}t} +C_{-} \left(
\begin{array}{c}
-2\gamma \\
\Lambda_{-} \\
\end{array}
\right) e^{\Lambda_{-}t},$$ $$R=C_{R} e^{-2\delta t},
\Lambda_{\pm}=-\delta\pm\sqrt{\delta^{2}-4\gamma^{2}}$$ by choosing as initial state one of the two levels with $
\rho(0)=1/2, R(0)=J(0)=0$.
Finally, the averaged off-diagonal matrix element of the density matrix is easily shown to be $$\overline{\rho_{12}(t)}=\imath J(t)= {\frac {\imath \gamma}
{\sqrt{4\gamma^{2}}-\delta^{2}-}
\sin(\sqrt{4\gamma^{2}-\delta^{2}}t)e^{-\delta t}}$$ for $\delta^{2}<4\gamma^{2}$. If $\delta^{2}>4\gamma^{2}$ in Eq. (12) the quantity under the square root changes sign and $\sin$ is replaced by $\sinh$. Thus, for $t\to \infty$ the averaged density matrix approaches half the unit matrix with only diagonal matrix elements and the quantum coherences described by $\overline{\rho_{12}}$ becoming zero, oscillating for the quantum case $\delta<2\gamma$ and monotonically for the classical case $\delta>2\gamma$.
In order to examine the dephasing effect of dynamic disorder we have plotted in Fig. 8 the phase $\theta$ of $ \rho_{1,2}$ vs. $t$ obtained from numerical computations. Our results are presented for $\gamma=1$ and different values of disorder $W$ which verify the critical value $W_{c}=\sqrt{24}$ of the previous analysis based on averages. The quantum coherence remains for weak dynamic disorder $W< W_{c}$ while for higher dynamic disorder $W> W_{c}$ the phase randomizes and the system becomes classical.
{width="6.0cm"}
IV. DISCUSSION
==============
Quantum walks are quantum analogues of classical random walks which have been proposed for quantum computation purposes to create quantum algorithms which run faster in quantum computers. They can also arise from mapping various physical problems (e.g. see [@r24]). Some quantum algorithms which speed-up classical methods have already been successfully employed, such as for search problems on graphs. These algorithms which show amplitude amplification during the evolution could be efficiently implemented in a quantum computer. However, since they are often confronted with disorder we have examined how quantum wave-packets move in the presence of disorder, by computing the probability density $P(x,t)=|\Psi(x,t)|^{2}$ from solving the time-dependent Schr$\ddot{o}$dinger equation in the discrete space $x$ of one-dimensional lattice. The static disorder due to imperfections and the dynamic disorder due to the environment could become obvious in scattering from nanostructures or can appear from environmental noise, averaging over measurements, etc.
Our main conclusions from the quantum evolution of $\delta$-function and very broad initial spatial wave-packets are: (1) In random media quantum walks can perform even worse than their classical counterparts since Anderson localization from negative quantum interference completely stops the quantum walk although the corresponding ordinary random walks in the presence of disorder can still move despite infinitely slowly. (2) For dynamic disorder we have no benefit from quantum walks either, since for longer times the ballistic evolution for small-$t$ crosses over to classical diffusion for long-$t$ and the quantum walks become classical via a quantum to classical crossover. (3) The answer to the question “what slows down the quantum walk?” is, on one hand, “static disorder via negative quantum interference” and, on the other hand, “dynamic disorder at long enough times which slows down the quantum walk and makes it no different from ordinary random walk”. Therefore, quantum interference in random media can hold surprises for quantum walks and their advantages should appear for weak disorder or short times only. In higher dimensions quantum walks are also expected to operate for weak disorder to avoid quantum localization. In conclusion, our computations show Anderson localization or decoherence as the main enemies of quantum walks in the presence of static and dynamic disorder, respectively, which destroy their well-known quadratic or exponential speed-up. Our study could be useful towards creating better quantum search algorithms in the presence of disorder[@r25].
[20]{}
L.K. Grover, Phys. Rev. Lett. [**79**]{}, 325 (1997).
P.W. Shor, SIAM J. Comp., [**26(5)**]{}, 1184 (1997).
Y. Aharonov, L. Davidovich and N. Zagury, Phys. Rev. A [**48(2)**]{}, 1687 (1993).
D.A. Meyer, J. Stat. Phys.[**85**]{}, 551 (1996).
E. Farhi and S. Gutmann, Phys. Rev. [**A**]{} [**58**]{}, 915 (1998).
For an introductory review to [*quantum random walks*]{} see: J. Kempe, Contemp. Phys. [**44**]{}, 307 (2003).
A.M. Childs, R. Cleve, E. Deotto, E. Farhi, S. Gutmann and D.A. Spielman, Proc. 35th ACM Symposium on Theory of Computing, ACM Press, 59 (2003).
D.E. Katsanos, S.N. Evangelou and Xiong, Phys. Rev. [**B**]{} [**51**]{}, 895 (1995).
A. Patel, K.S. Raghunathan and P. Rungta, Phys. Rev. [**A**]{} [**71**]{}, 032347 (2005).
A.M. Childs, E. Farhi and S. Gutmann, Quantum Information Processing [**1**]{}, 35 (2002).
F.W. Strauch, quant-ph/0606050; quant-ph/0508096.
M.V. Berry, J. Phys. [**A**]{} [**26**]{}, 6617 (1996).
M. Berry, I. Marzoli and W. Schleich, [*“Quantum Carpets, carpets of light"*]{}, Physics World, pg. 39, June (2001).
I. Amanatidis, D.E. Katsanos and S.N. Evangelou, Phys. Rev. [**B**]{} [**69**]{}, 190503 (2004).
W. Dur, R. Raussendorf, V.M. Kendon and H.-J. Briegel, Phys. Rev. [**A**]{} [**66**]{}, 052319 (2002).
B.C. Travaglione and G.J. Milburn, Phys. Rev. [**A**]{} [**65**]{}, 032310 (2002).
Ya.G. Sinai, Theory Prob. Appl. [**27**]{}, 247 (1982).
J.P. Bouchaud and A. Georges, Phys. Rep. [**195**]{}, 127 (1990).
P.W. Anderson, Phys. Rev. [**109**]{}, 1492 (1958).
J.P. Keating, N. Linden, J.C.F. Matthews and A. Winter, quant-ph/0606205.
P. Ribeiro, P. Milmann and R. Mosseri, Phys. Rev. Lett. [**93**]{}, 195107 (2004).
V. Kendon, quant-ph/0606016.
A.A. Ovchinikov and N.S. Erikhman, Sov. Phys. JETP, Vol. 40, 733 (1975).
T. Oka, N. Konno, R. Arita and H. Aoki, Phys. Rev. Lett. 94, 100602 (2005).
P. Shor, Quantum Information Processing [**3**]{}, 5 (2004).
[^1]: e-mail:[email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Fluctuations of conserved quantities in heavy-ion collisions are used to probe the phase transition and the QCD critical point for the strongly interacting hot and dense nuclear matter. The STAR experiment has carried out moment analysis of net-proton (proxy for net-baryon (B)), net-kaon (proxy for net-strangeness (S)), and net-charge (Q). These measurements are important for understanding the quantum chromodynamics phase diagram. We present the analysis techniques used in the moment analysis by the STAR experiment and discuss the moments of net-proton and net-charge distributions from the first phase of the Beam Energy Scan program at the Relativistic Heavy Ion Collider.'
address: ' Institute of Particle Physics and Key Laboratory of Quark & Lepton Physics (MOE), Central China Normal University, Wuhan, 430079, China.'
author:
- Xiaofeng Luo
bibliography:
- 'FAIRNESS2014.bib'
title: Fluctuations of Conserved Quantities in High Energy Nuclear Collisions at RHIC
---
Introduction
============
One of the main goal of heavy-ion collisions is to explore the phase structure of the hot and dense nuclear matter. The phase structure can be displayed in the two dimensional Quantum Chromodynamics (QCD) phase diagram (temperature, $T$ vs. baryon chemical potential, $\mu_{B}$). It is conjectured on the basis of theoretical calculation, at finite temperature and vanishing $\mu_B$ ($\mu_B=0$) region, there is a crossover transition between hadronic phase and partonic phase, while at large $\mu_B$ region, the phase transition is of the first order. Thus, there should be the so called QCD Critical Point as the endpoint of the first order phase boundary towards the crossover region, which is a second order phase transition point [@QCP_Prediction]. Since the [*ab initio*]{} Lattice QCD calculation encounters a serious so called sign problem at finite $\mu_B$ region, there are large uncertainties in determining the location of the CP or even its existence [@qcp; @qcp_Rajiv]. Experimental confirmation of the existence of the CP will be an excellent verification of QCD theory in the non-perturbative region and a milestone of exploring the QCD phase structure. It is one of the main goals of the Beam Energy Scan (BES) program at the Relativistic Heavy Ion Collider (RHIC). By tuning the colliding energy of the gold nuclei from high to low values, we can vary the $T$ and $\mu_B$ of the nuclear matter created in heavy-ion collisions [@bes]. This enables us to probe a broad region of the QCD phase diagram. On the other hand, fluctuations of conserved quantities, such as net-baryon (B), net-charge (Q) and net-strangeness (S), are predicted to be sensitive to the correlation length of the system [@qcp_signal; @ratioCumulant; @Neg_Kurtosis] and directly connected to the susceptibilities computed in the theoretical calculations [@science; @Lattice]. Thus, it can serve as a powerful tool to probe the phase transition and CP signal in heavy-ion collisions. Experimentally, the STAR experiment has measured the energy dependence of moments (up to the fourth order) of event-by-event net-proton (proton minus anti-proton number, proxy of net-baryon [@Hatta]) and net-charge multiplicity distributions in Au+Au collisions at [[[[$\sqrt{s_{_{\mathrm{NN}}}}$]{}]{}]{}]{}= 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV. Those data are collected from the first phase of the RHIC BES in the years 2010 and 2011. In the year 2014, another energy point 14.5 GeV is successfully taken and can fill in a large $\mu_B$ gap between 11.5 and 19.6 GeV. In this paper, we will present recent experimental results from RHIC/STAR experiment for fluctuations of conserved quantities as well as the analysis techniques. The physics implication of the results will then be discussed.
The paper is organized as follows: In the second section, we will give a brief discussion of the techniques used in the moment analysis. The experimental results and discussion will be presented in the third section. Finally, we will give a summary and outlook.
Analysis Techniques
===================
During the last five years, a series of analysis techniques have been applied to the moments of conserved quantities distributions in heavy-ion collisions. Those include : (1) Centrality bin width correction [@WWND2011; @technique]. (2) Novel centrality determination to account for the effects of centrality resolution and auto-correlation [@technique]. (3) Efficiency correction and error estimation [@Delta_theory; @voker_eff1; @voker_eff2; @Unified_Errors]. Those techniques are very important to address the background effects and extract the dynamical fluctuation signals from the observables. In heavy-ion collisions, we cannot directly measure the collision centrality and/or initial collision geometry of the system of two nuclei. The centrality in heavy-ion collisions is generally determined by comparing the measured particle multiplicity with the Glauber Monte Carlo simulations. It is denoted as a percentage value (for e.g. 0-5%, 5-10%,...) for a collection of events to represent the fraction of the total cross section. This in general can cause two undesirable effects in the moment analysis of particle multiplicity distributions within finite centrality bin. One is the so called centrality bin width effect, which is caused by volume variation within a finite centrality bin size and the other one is centrality resolution effect, which is due to the initial volume fluctuations. Let’s review those techniques one by one.
Centrality Bin Width Correction
-------------------------------
The centrality bin width correction is to address the volume fluctuations effects on the higher order moments of conserved quantities distributions within a finite width centrality bin. It is also called centrality bin width effect. This effect needs to be eliminated, as an artificial centrality dependence could be introduced by it. To do this, the centrality bin width correction is applied to calculate the various moments of particle multiplicities distributions in one wide centrality bin. Experimentally, the smallest centrality bin is determined by a single value of particle multiplicity. Experimental results are usually reported for a wider centrality bin (a range of particle multiplicity), such as $0-5\%$,$5-10\%$,...etc., to reduce statistical errors. To eliminate the centrality bin width effect, we calculate the various order cumulants ($C_{n}$) for each single particle multiplicity within one wider centrality bin and weighted averaged by the corresponding number of events in that multiplicity. $$\label{eq:cbwc}
{C_n} = \frac{{\sum\limits_{r = {N_1}}^{{N_2}} {{n_r}C_n^r} }}{{\sum\limits_{r = {N_1}}^{{N_2}} {{n_r}} }} = \sum\limits_{r = {N_1}}^{{N_2}} {{\omega _r}C_n^r}$$ where the $n_r$ is the number of events for multiplicity value $r$ and the corresponding weight for the multiplicity $r$, ${\omega _r} = {n_r}/\sum\limits_{r = {N_1}}^{{N_2}} {{n_r}}$. $N_1$ and $N_2$ are the lowest and highest multiplicity values for one centrality bin. Once getting the centrality bin width corrected cumulants via Eq. (\[eq:cbwc\]), we can calculate the various order moments, for e.g. $=C_{4}/C_{2}$ and $=C_{3}/C_{2}$, where the $\kappa$ and $S$ are kurtosis and skewness, respectively. The final statistical error of moments for one centrality can be evaluated by standard error propagation based on Eq. (\[eq:cbwc\]). For more details, one can see [@technique].
Centrality Resolution and Auto-correlation Effects
--------------------------------------------------
Particle multiplicity are usually used in the centrality determination, as it can reflect the initial geometry of heavy-ion collision. However, the relation between measured particle multiplicities and collision geometry is not one-to-one correspondence and there are fluctuations in the particle multiplicity even for a fixed collision geometry. Thus, one could obtain a finite resolution of initial collision geometry (centrality resolution) by using particle multiplicity to determine the centrality. The more particles are used in the centrality determination, the better centrality resolution and smaller fluctuation of the initial geometry (volume fluctuation) we get. This may affect moments of the event-by-event multiplicity distributions. On the other hand, we use the multiplicity of charged kaon and pion to define the collision centrality in Au+Au collisions. This is to prevent the effect of auto-correlation between protons/antiprotons involved in our moments analysis and in the centrality definition. The auto-correlation effect will results in the suppressing values of the moments. To avoid the auto-correlation effect, we should exclude the corresponding proton/antiproton from the centrality definition. The two background effects can be well avoided by a novel centrality definition using particle multiplicities. For more details, one can see [@technique; @CPOD2013].
Efficiency Correction and Error Estimation
------------------------------------------
Finite detecting efficiency for particle yield measurement in heavy-ion collisions can be easily corrected. However, it is not straightforward to get the efficiency corrected results for higher moments of particle multiplicity distributions. Since the fluctuation analysis is statistics hungry, it is crucial to get the correct statistical errors with limited statistics. We provide a unified description of efficiency correction and error estimation for various order moments of multiplicity distributions [@Unified_Errors]. The basic idea is to express the moments and cumulants in terms of the factorial moments, which can be easily corrected for efficiency effects. By knowing the covariance between multivariate factorial moments, we use the standard error propagation based on Delta theorem to obtain the error formula for efficiency corrected moments. This method can also be applied to the phase space efficiency case, where the efficiency of proton or anti-proton are not constant within studied phase space. One needs to note that the efficiency correction and error estimation should be done for events with the same particle multiplicity used for centrality determination and just before the centrality bin width correction.
Experimental Results
====================
Recently, the STAR experiment has published the beam energy dependence of moments (up to forth order) of event-by-event net-proton [@STAR_BES_PRL] and net-charge [@netcharge_PRL] multiplicity distributions in Au+Au collisions at [[[[$\sqrt{s_{_{\mathrm{NN}}}}$]{}]{}]{}]{}= 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV. For net-proton analysis, the protons and anti-protons are identified with ionization energy loss in the Time Projection Chamber (TPC) of the STAR detector within the transverse momentum range $0.4<p_{T}<0.8$ GeV/c and at mid-rapidity $|y|<0.5$. In the net-charge case, the charged particles are measured within transverse momentum range $0.2<p_{T}<2$ GeV/c and pseudo-rapidity range $|\eta|<0.5$.
Figure 1 shows the energy dependence of , and of net-proton and net-charge distributions of Au+Au collisions for two centralities (0-5% and 70%-80%) at [[[[$\sqrt{s_{_{\mathrm{NN}}}}$]{}]{}]{}]{}= 7.7, 11.5, 19.6, 27, 39, 62.4 and 200 GeV. The Skellam (Poisson) expectations shown in the figure reflect a system of totally uncorrelated, statistically random particle production. It predicts the and /Skellam to be unity for Skellam expectations as well as in the hadron resonance gas model. For the net-proton results, the most significant deviation of and from Skellam distribution is observed at 19.6 and 27 GeV for 0-5% Au+Au collisions. At energies above 39 GeV, the results are close to Skellam expectation. As the statistical errors are large at low energies (7.7 and 11.5 GeV), more statistics is necessary to quantitatively understand the energy dependence of and . To understand the effects of baryon number conservation etc., UrQMD model calculations (a transport model which does not include a CP) for 0-5% are presented and the results show a monotonic decrease with decreasing beam energy. For more details on baseline comparison, one can see [@QM2014_baseline]. For the net-charge results, we did not observe non-monotonic behavior for and within current statistics. The expectations from negative binomial distribution can better describe the net-charge data than the Poisson (Skellam) distribution. More statistics is needed for net-charge moment calculation.
![(Color online) Energy dependence of moments of net-proton (left) and net-charge (right) distributions for Au+Au collisions at RHIC BES energies. The statistical and systematical error are shown in bars and brackets, respectively. []{data-label="fig:cumulants_energy"}](netproton.eps)
![(Color online) Energy dependence of moments of net-proton (left) and net-charge (right) distributions for Au+Au collisions at RHIC BES energies. The statistical and systematical error are shown in bars and brackets, respectively. []{data-label="fig:cumulants_energy"}](netcharge.eps)
We would like to point out: (1) The STAR experiment has carried out moment analysis for net-proton, net-kaon, and net-charge. Different measurements are affected by kinematic cuts, resonance decays, and other dynamical effects differently. In search for the QCD critical point, careful studies are called for. (2) So far, the resonance decay effects are contained in the experimental results of net-proton and net-charge moments. One can estimate the decay effects by theoretical calculations and/or models. Based on the hadron resonance gas model calculation [@HRG_baseline], the decay effects for net-proton is small and at 2% level. But for the net-charge, the decay effects are large. (3) Based on the Delta theorem, statistical error of cumulants ($\Delta(C_n$)) are related to the width of the distribution as $\Delta(C_n)$ $\sim$ O($\sigma^{n}$) [@Delta_theory; @Unified_Errors]. Thus, the wider is the distribution, the larger are statistical errors for the same number of events. That’s explains why we got larger statistical errors for net-charge moments than that of net-proton, as the former has much wider distribution. (4) It is predicted that the coupling strength of pions to the critical fluctuation is smaller than protons/anti-protons. Thus, the fluctuation of net-protons will be more sensitive to the critical fluctuations than the fluctuation of net-charges (dominated by pions) [@ratioCumulant; @privateCom].
In the year 2018, the second phase of the beam energy scan program (BES-II) at RHIC will get started. During the BES-II, we will fine tune the beam energies below 20 GeV and accumulate more events due to the increase of the luminosity of 3-10 times by using stochastic electron cooling technique. Several sub-detector upgrades for the STAR experiment are ongoing and will be ready when BES-II starts. Two of the upgrades are important for moment analysis. The inner TPC (iTPC) upgrade will improve the tracking efficiency and enlarge the TPC acceptance. The Event Plane Detector (EPD) is a forward detector used to determine the event plane for Au+Au collisions. It can provide centrality determination with particles far away from the central collision region.
Summary
=======
In this paper, we discussed the analysis techniques used in the moment analysis and the energy dependence of moments of net-proton and net-charge distributions published by the STAR experiment. For the net-charge results, we didn’t observe non-monotonic behavior for and within current statistics. For the net-proton results, the most significant deviation of and from Skellam distribution is observed at 19.6 and 27 GeV for 0-5% Au+Au collisions. More statistics are needed at low energies. In principle, one can extend the transverse momentum coverage for proton/anti-proton up to about 2$\sim$3 GeV/c with time of flight (ToF) detector at STAR for particle identification. This would allows us to have larger acceptance for protons and anti-protons in the analysis. Large acceptance is crucial for fluctuation of conserved quantities in heavy-ion collisions to probe the signals of the phase transition and the QCD critical point.
Acknowledgement {#acknowledgement .unnumbered}
===============
The work was supported in part by the MoST of China 973-Project No. 2015CB856901, NSFC under grant No. 11205067, 11221504 and 11228513.
References {#references .unnumbered}
==========
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using total variation based energy minimisation we address the recovery of a blurred (convoluted) one dimensional (1D) bar code. We consider functionals defined over all possible bar codes with fidelity to a convoluted signal of a bar code, and regularised by total variation. Our fidelity terms consist of the $L^2$ distance either directly to the measured signal or preceded by deconvolution. Key length scales and parameters are the $X$-dimension of the underlying bar code, the size of the supports of the convolution and deconvolution kernels, and the fidelity parameter. For all functionals, we establish parameter regimes (sufficient conditions) wherein the underlying bar code is the unique minimiser. We also present some numerical experiments suggesting that these sufficient conditions are not optimal and the energy methods are quite robust for significant blurring.'
author:
- 'Rustum Choksi[^1] Yves van Gennip [^2]'
bibliography:
- 'bibliography.bib'
title: Deblurring of One Dimensional Bar Codes via Total Variation Energy Minimisation
---
**Key words: bar code, deblurring, total variation, energy minimisation**
**MSC2010:** 49N45, 94A08
Introduction and notation {#sec:intro}
=========================
A one-dimensional bar code is a finite series of alternating black bars and white spaces with varying widths (Figure \[Fig1\]). The so-called [*$X$-dimension*]{} of the bar code is the width of the narrowest bar or space.
[![Left: A standard 1D 12-digit Universal Product Code (UPC) bar code. The 12 numerical digits readable to the human eye are encoded in the bar spacings. Right: A bar code scanner in which the black bars absorb light, while the white bars reflect it. Photo diodes turn the reflected light into an electrical signal, which may be both blurred and noisy. This signal is then converted into digital pulses. []{data-label="Fig1"}](barcode.pdf "fig:"){width="2.3in"}]{}
A bar code scanner can use, for example, light detectors or photoreceptors similar to those used by a digital camera to pick up light and dark areas of the bar code, and produce a continuous signal associated with the darkness distribution across the span of the bar code ([c.f.]{} [@Palmer07]). One is left with an approximation to the bar code which depending on the scanner and the way in which the scan was taken (distance, vibrations etc.) can be blurred and noisy. (see Figure \[Fig1\]). Thus in this article we deal with the general question: [*Given a blurred and possibly noisy [*signal*]{} $f$ associated with a bar code, how can we deblur and denoise effectively to reconstruct the bar code?*]{}
Standard commercial deblurring techniques are often based upon edge detection, for example, finding local extrema of $f'$ which hopefully correspond to the discontinuities (i.e. interfaces) of the original bar code. As noted in [@Esedoglu04], this presents several difficulties: (i) the process is highly unstable to small changes in the signal, for example to the presence of noise; (ii) points associated with local extrema of $f'$ are only crude approximations to the true locations of the bars with convolution tending to distort these points if the standard deviation of the convolution kernel —or if the kernel is compactly supported the size of its support— is comparable to the $X$-dimension of the bar code; (iii) if the standard deviation or support size of the kernel is very large compared to the $X$-dimension, some edges in the bar code may have no corresponding extrema of $f'$ at all.
In this article, we take a different approach based upon energy minimisation involving a total variation (TV) regularisation – a method introduced by Rudin, Osher and Fatemi in [@RudinOsherFatemi92-2]. One of the many advantages of this approach is that energy minimisation is a process [*stable*]{} with respect to small changes in the input signal. In the context of bar code reconstruction, this type of energy minimisation was first proposed by Fadil Santosa, and analysed by Selim Esedoglu [@Esedoglu04] (see also [@Wittman04]). Here we take a similar approach but with an important difference: We are interested in directly testing the merits of energy minimisation for TV-based functionals by considering *ansätze* for $f$ which involve convolutions of a bar code with certain blurring kernels. The functionals, blurring kernels, and admissible classes of bar codes possess certain length scale parameters. In [@Esedoglu04] Esedoglu shows existence of solutions for a variety of functionals and then proceeds to numerically test these functionals via an algorithm which approximates the actual length scale parameter of the blurring kernel and uses this information to reconstruct the clean bar code signal. Here we fix these length scales as parameters and ask under what conditions can we insure that energy minimisation gives back the underlying bar code.
We begin by introducing some notation. A bar code is given by a function $u$ of bounded variation ([*c.f.*]{} [@Giusti84]) with $\operatorname{supp}u \subset [0,1]$ taking on the values $0$ and $1$ a.e., i.e. we consider a subset of the space $BV({{{\ensuremath{\mathbb{R}}}}}; \{0, 1\})$. In particular the [*general admissible set for bar codes*]{} is $${{\cal B}}:= \bigg\{ u \in BV({ {{{\ensuremath{\mathbb{R}}}}}}; \{0, 1\}) \,
: \, u=0 \text{ a.e. in } [0,1]^c \bigg\}.$$ Modulo a set of measure zero, the set $\{x\in[0,1]: u(x)=1\}$ consists of a finite number of disjoint non-empty intervals called bars. We denote these non-empty bar intervals by $b_i$ with length $|b_i|$. Similarly the intervals of nonempty [*white*]{} spaces (i.e. intervals in $[0,1]$ where $u = 0$) are denoted by $w_i$. In this paper, signals will always be generated by bar codes whose $X$-dimension is bounded below. That is, we assume there exists a constant $\omega >0$ such that the minimum width of these bars and spaces is [*a priori*]{} bounded below by $\omega $. Our space of [*generating bar codes*]{} is thus $${{\cal B}}_{\omega} := \bigg\{ u \in \,{{\cal B}}: \,\,\, \forall i \,\,\, \min\{ |b_i|, |w_i| \} \ge
\omega \bigg\}.$$
In some calculations it is useful, though harmless, to assume that we know whether the bar code starts (and/or ends) with a bar or a space. To this end, we define for $i, j\in\{0,1\}$, the sets $${{\cal B}}^{ij} := \Bigl\{ u\in {{\cal B}}: \text{there exists } x_1, x_2 \text{ such that } u=i \text{ on } [0,x_1] \text{ and } u=j \text{ on } [x_2, 1]\Bigr\},$$ $${{\cal B}}_{\omega}^{ij} := \Bigl\{ u\in {{\cal B}}_{\omega}: \text{there exists } x_1, x_2 \text{ such that } u=i \text{ on } [0,x_1] \text{ and } u=j \text{ on } [x_2, 1]\Bigr\}.$$ As a useful addition to our terminology we will call the transition from a bar to a space or vice versa an *interface*. This means that $\int_{{{\ensuremath{\mathbb{R}}}}}|u'|$ is equal to the number of interfaces in $u$ (where interfaces at $x=0$ and $x=1$ are included). The notation $\chi_S$ will be used for the characteristic function of a given set $S$.
We approach the blurring via convolution on a length scale $\sigma >0$ (c.f. [@JosephPavlidis93; @JosephPavlidis94; @ShellhammerGorenPavlidis99; @Esedoglu04]) with a symmetric unimodal kernel of unit mass. Different results hold for kernels of varying generality and we will discuss these kernels shortly. For the moment, let the kernel $\phi_\sigma$ denote a symmetric probability distribution on ${{{\ensuremath{\mathbb{R}}}}}$ with [*size*]{} (for example, its standard deviation or if compactly supported, half the size of its support) $\sigma$. Given a bar code $z \in {{\cal B}}_{\omega}$, we will assume observed signals of the form: $$f_\sigma (x) \, := \, (\phi_\sigma \ast z) (x) \, = \, \int_{-\infty}^\infty \phi_\sigma ( x - y) \, z (y) \, dy.$$ where $\sigma \ge 0$ is fixed (note that $\phi_0=\delta$, the Dirac delta distribution, and thus $f_0=z$).
For $u \in {{\cal B}}$:
- We consider a fidelity term which compares $u$ directly with the observed signal: $$F_1(u) \, := \, \int_{{{{\ensuremath{\mathbb{R}}}}}} |u'| \, + \, \lambda \|u \, -\, f_\sigma \|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2.$$
- ([*deconvolution/deblurring functional*]{}) Under the belief that the signal involves a convolution with a known kernel, we may incorporate this convolution into the structure of the fidelity term and consider: $$F_2 (u) \, := \, \int_{{{\ensuremath{\mathbb{R}}}}}|u'| \, + \, \lambda \|\phi_\sigma * u \, - \, f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2.$$ If $\phi_\sigma$ is known, one may ask as to the merits of energy minimisation as we could simply Fourier transform the observed signal $f_\sigma$ and divide by the Fourier transform of $\phi_\sigma$ to recover the Fourier transform of $z$. However, this process is highly unstable with respect to small perturbations and in practice, there is always some noise associated with the observed signal $ f_\sigma$. Directly solving back for $z$ is analogous to solving the heat equation backwards. Energy minimisation provides a stable numerical approach to deblurring and indeed denoising.
- ([*blind deconvolution/deblurring functional*]{}) Assuming no knowledge of $\sigma$, we may deconvolute with a kernel of similar type but with [*size*]{} $\rho$: $$F_3 (u)\, := \,
\int_{{{\ensuremath{\mathbb{R}}}}}|u'| \, + \, \lambda \|\phi_\rho * u \, - \, f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2.$$
Note that $F_1$ and $F_2$ are special cases of $F_3$ with $\rho$ equal to $0$ and $\sigma$ respectively, and that in our notation, the dependence of $F_i$ on $\lambda$, $z$, $\phi_\sigma$ and $\phi_\rho$ is suppressed. It is straightforward to see (cf. Lemma \[lemma0\]), that for all parameters and all generating bar codes $z$, a minimiser of $F_i(u)$ over $u\in{{\cal B}}$ exists. In this article, we examine the following questions: [*For what values of the parameters $\lambda, \omega, \sigma$ and $\rho$, is the minimiser of $F_i$ known, and in particular, when is it the underlying bar code $z$*]{}? Our results consist of two parts: First off we present a simple result that if $\lambda$ is sufficiently small, the unique minimiser is simply $u \equiv 0$. Part 2 deals with sufficient conditions for when the unique minimiser is the underlying bar code $z$.
Note that if $\lambda$ is sufficiently small, $u \equiv 0$ is the unique minimiser in ${{\cal B}}$ (i.e. an empty bar code). This is clearly the case if $$\lambda < \lambda_0 := 2 / \|f_\sigma \|_{L^2}^2,$$ since any nontrivial bar code has a minimum total variation of $2$. Sufficient and necessary conditions on $\lambda$ for $u=0$ to be the unique minimiser in $BV({{{\ensuremath{\mathbb{R}}}}})$ are given in the first parts of the Theorems \[thm:minimofF2\], \[thm:F4\] and Corollary \[thm:F3\] by exploiting a method adopted from [@Meyer01] wherein the following dual norm associated with $BV$ is used: $$\label{def:starnorm}
\|f\|_\ast := \sup \left\{ \left|\int_{{{\ensuremath{\mathbb{R}}}}}f v\right| : v \in L^1({{{\ensuremath{\mathbb{R}}}}}) \text{ and } \int_{{{\ensuremath{\mathbb{R}}}}}|v'| \leq 1\right\}.$$ This threshold for a trivial minimiser is given by $\lambda \leq \frac1{2 \|\phi_\rho*f_\sigma\|_\ast}$, $\rho \geq 0$ (parts 1 of the theorems below) and is lower than $\lambda_0$. To see this note that for any $f\in L^2({{{\ensuremath{\mathbb{R}}}}}; [0, 1])$ with compact support we may take $I$ to be a bounded subset of ${{{\ensuremath{\mathbb{R}}}}}$ such that $\operatorname{supp}\phi_\rho*f \subset I$. Since $v=\frac12 \chi_I$ is an admissible function in the definition of $\|\phi_\rho*f\|_*$ we have $$\|\phi_\rho*f\|_* \geq \frac12 \int_I \phi_\rho*f = \frac12 \int_{{{\ensuremath{\mathbb{R}}}}}f \geq \frac12 \int_{{{\ensuremath{\mathbb{R}}}}}f^2.$$ Therefore $$\frac1{2 \|\phi_\rho*f\|_*}\leq \left(\int_{{{\ensuremath{\mathbb{R}}}}}f^2\right)^{-1} < \frac2{\|f\|_{L^2}^2}.$$ For $\lambda$ between this threshold and $\lambda_0$, $u=0$ is the unique minimiser in ${{\cal B}}$ but not in $BV({{{\ensuremath{\mathbb{R}}}}})$.
The case where $z$ is the unique minimiser is more subtle and depends critically on both the size and particular nature of the blurring kernel. We make two assumptions here:
1. \[item:Xdimassump\] We restrict our attention to kernels with compact and [*small*]{} (with respect to the $X$-dimension $\omega$) support.
2. We further consider unimodal, symmetric kernels in $\mathcal{K}$ defined below and explicit regimes are computed using a prototype of such a kernel, the hat function defined by $$\hat \phi_{\sigma}(x) := \left\{ \begin{array}{ll} (1\, - \, {x}/{\sigma})/{\sigma} & \text{if } 0\leq x<\sigma,\\
(1\,+ \, {x}/{\sigma})/{\sigma} & \text{if } -\sigma < x \leq 0,\\ 0 & \text{if }|x| \geq \sigma. \end{array}\right.$$ The class $\mathcal{K}$ is defined by $$\begin{aligned}
\label{eq:Ksigma}
\mathcal{K} &:= \left\{ \phi_\sigma \in L^1({{{\ensuremath{\mathbb{R}}}}}): \exists \sigma>0 \,\,\, \phi_\sigma (x) = p(-x, \sigma) \chi_{[-\sigma, 0]}(x) + p(x, \sigma) \chi_{[0, \sigma]}(x)\right. \notag\\&\hspace{0.7cm} \left. \text{ for a non-negative function } p: [0, \sigma]\times(0, \infty) \text{ monotonically decreasing} \right.\notag \\ &\hspace{0.7cm} \left. \text{ in } x, \text{ and } \int_0^\sigma p(x, \sigma)\, dx = \frac12\right\}. \end{aligned}$$ We will consistently use a subscript as in $\phi_\sigma$ to indicate the value of the parameter $\sigma$ in the definition of $\mathcal{K}$, i.e. $\phi_\rho \in \mathcal{K}$ is in the subset of $\mathcal{K}$ where $\sigma=\rho$.
Note in particular that if $\phi_\sigma\in L^1({{{\ensuremath{\mathbb{R}}}}})$, we have $\phi_\sigma*u\in L^2({{{\ensuremath{\mathbb{R}}}}})$ for all $u\in \mathcal{B}$.
It is also convenient to consider a subclass of $\mathcal{K}$ $$\begin{aligned}
\mathcal{K}_3 &:= \left\{ \phi_\sigma \in \mathcal{K}: \phi_\sigma \in C_c({{{\ensuremath{\mathbb{R}}}}}), p \text{ continuously differentiable in } \sigma \text{ and } (\ref{eq:condpsit}) \text{ holds}\right\},\notag \end{aligned}$$ where the non-obvious condition (\[eq:condpsit\]) is $$\begin{aligned}
\label{eq:condpsit}
&\hspace{1cm} \forall \tau \in (0, \sigma], \,\, \forall c \geq 2 \sigma, \,\, \forall x\in [0,c]: \notag\\
&\mathcal{J}(\sigma, \tau, x, c) := \int_0^\tau \int_{x-c}^x \frac{\partial}{\partial \tau} p(y, \tau) \left[ \phi_\sigma(y-w)+\phi_\sigma(y+w)\right]\, dw\, dy \leq 0.\end{aligned}$$
This condition insures a certain monotonicity property (c.f. Lemma \[lem:decreasing\]) of double convolutions with bar codes. As we show in Appendix \[app:psitaufrhononpos\] a sufficient condition for (\[eq:condpsit\]) to be satisfied is if for each $\tau \in (0, \sigma]$
1. either $\displaystyle \frac{\partial}{\partial \tau} p(x, \tau)$ is monotonically increasing in $x$ and $\mathcal{J}(\sigma, \tau, 0, c) \leq 0$ for all $c\geq 2\sigma$,
2. or $\displaystyle \frac{\partial}{\partial \tau} p(x, \tau)$ is monotonically decreasing in $x$ and $\mathcal{J}(\sigma, \tau, \frac{c}2, c) \leq 0$ for all $c\geq 2\sigma$.
In the same appendix we show that the hat function $\hat \phi_\sigma$ satisfies condition (a) above for each $\tau\in (0, \sigma]$. In practical applications using kernels that do not satisfy either (a) or (b), condition (\[eq:condpsit\]) can be tested numerically.
Assumption \[item:Xdimassump\] is rather important and indeed restrictive as it limits the possible effect of blurring. For $F_2$ and $F_3$ we require $\sigma, \rho \leq \omega/2$, insuring no interactions between neighbouring bars. For $F_1$, the condition is slightly less restrictive, namely $\sigma \leq \omega$. The second assumption, particularly, the specification of the hat function is more for convenience. A crucial step in our direct and rather [*brute-force*]{} approach is to assume a minimiser with a certain structure and directly construct competitors which differ on a set bar or space. For this step, one can explicitly calculate a regime wherein such a competitor exists, and for simplicity we have performed the calculations for the hat function (which were greatly simplified by the use of Maple). We discuss modifications for other kernels in Remark \[other\] below. Let us now state our results.
\[thm:minimofF2\] The following hold:
1. \[item:F1trivminoverBV\] Let $\phi_\sigma \in \mathcal{K}$. $ u \equiv 0$ is the unique minimiser for $F_1$ over $BV({{{\ensuremath{\mathbb{R}}}}})$ iff $\|f_\sigma\|_\ast \leq \frac1{2\lambda}$.
2. \[item:F1zminoverB\] Let $\phi_\sigma = \hat \phi_\sigma$ and $z\in {{\cal B}}_{\omega}$. If $\sigma$ and $\lambda$ satisfy $$\label{eq:conditionsigmalambda}
\sigma \leq \omega \qquad {\rm and} \qquad \frac23 \sigma + \frac2{\lambda} < \omega,$$ then $u = z$ is the unique minimiser of $F_1$ over ${{\cal B}}$.
\[thm:F4\] The following hold:
1. \[item:F3trivminoverBV\] Let $\phi_\sigma \in \mathcal{K}$. $ u \equiv 0$ is the unique minimiser for $F_3$ over $BV({{{\ensuremath{\mathbb{R}}}}})$ iff $\|\phi_\rho*f_\sigma\|_\ast \leq \frac1{2\lambda}$.
2. \[item:F3zminoverB\] Let $\phi_\sigma = \hat \phi_\sigma$ and $z
$ implicit in the definition of $ F_3$, be taken from $
\in {{\cal B}}_{\omega}^{ij}$ for some $i, j\in \{0,1\}$. Let $\sigma \leq \rho \leq \frac{\omega}2$. If $\lambda$, $\rho$, and $\sigma$ satisfy $$\label{eq:conditiononrhosigmalambda}
\frac2\lambda + \frac1{15 \rho^2} \Bigl(-\sigma^3 + 5 \rho \sigma^2 + 17 \rho^3\Bigr) < \omega,$$ then $u=z$ is the unique minimiser of $F_3$ over ${{\cal B}}^{ij}$.
Note that the left hand side of (\[eq:conditiononrhosigmalambda\]) is increasing as a function of (real and positive) $\rho$ and $\sigma$. By taking $\rho = \sigma$ in Theorem \[thm:F4\], we obtain
\[thm:F3\] The following hold:
1. \[item:F2trivminoverBV\] Let $\phi_\sigma \in \mathcal{K}$. $ u \equiv 0$ is the unique minimiser for $F_2$ over $BV({{{\ensuremath{\mathbb{R}}}}})$ iff\
$\|\phi_\sigma*f_\sigma\|_* \leq \frac1{2\lambda}$.
2. \[item:F2zminoverB\] Let $\phi_\sigma = \hat \phi_\sigma$ and $z
$, implicit in the definition of $ F_2$, be taken from $
\in {{\cal B}}_{\omega}^{ij}$ for some $i, j\in \{0,1\}$. If $\sigma \leq \frac{\omega}2$ and $\lambda>0$ satisfy $$\label{eq:condsigmalambdaconvolutions}
\frac2\lambda + \frac{21}{15} \sigma < \omega,$$ then $u=z$ is the unique minimiser of $ F_2$ over ${{\cal B}}^{ij}$.
\[other\] [**Extensions to other kernels**]{} We remark on extensions of parts 2 of Theorem \[thm:minimofF2\], Theorem \[thm:F4\], and Corollary \[thm:F3\] to more general kernels in $\mathcal{K}$. Their proofs consist of two steps: (i) First is to establish that any minimiser of $F_1$ or $F_3$ distinct from $z$ must have strictly less interfaces than $z$ (note that this is trivially satisfied for $F_2$ since $z$ uniquely minimises the fidelity term). For $F_1$, this is a simple consequence of vanishing first variation (Lemma \[lem:firstvar-F1\]), and holds for any kernel in $\mathcal{K}$. For $F_3$, the consequences of vanishing first variation are more involved (c.f. Lemmas \[lem:levelhalfrhosigmaz\] - \[lem:F3firstvar\]), and an important ingredient is a monotonicity property of double convolutions (c.f. Lemma \[lem:decreasing\]) which is responsible for condition (\[eq:condpsit\]). Thus, this step holds for any kernel in $\mathcal{K}_3$. (ii) The second step involves the explicit parameter regimes and is the reason why we have conveniently adopted the hat function. Here we show that there cannot exist a minimiser for $F_1$, $F_2$, or $F_3$ with fewer interfaces than $z$ if (\[eq:conditionsigmalambda\]), (\[eq:condsigmalambdaconvolutions\]), or (\[eq:conditiononrhosigmalambda\]) holds respectively: If a minimiser $u_0$ does have fewer interfaces than $z$, then there exists an interval (bounded below in length by $\omega$) on which $z$ has a bar and $u_0$ a space or vice versa. We then contradict our assumption by explicitly modifying $u_0$ on this interval to achieve lower energy. This last step requires some straightforward but tedious calculations. Maple has been a great help in performing the many integrations necessary involving the hat function $\hat \phi_\sigma$. This step can be reproduced for any $\phi_\sigma \in \mathcal{K}$ with different parameter regimes for each specific choice of kernel $\phi_\sigma$ as a result. Specifically, the calculations in the proof of Theorem \[thm:minimofF2\], part \[item:F1zminoverB\] after (\[eq:twointegrals\]) or the calculations in the proof of Lemma \[lem:F3partresult\] after (\[eq:F4inequality\]) respectively need to be redone for the new kernel.
Collecting the conditions necessary for steps (i) and (ii) we find that it is in principle possible to get results of the form of those in Theorem \[thm:minimofF2\], part \[item:F1zminoverB\] and Corollary \[thm:F3\], part \[item:F2zminoverB\] for $F_1$ and $F_2$ respectively for any $\phi_\sigma \in \mathcal{K}$. Similarly a result for $F_3$ as the one in Theorem \[thm:F4\], part \[item:F3zminoverB\] can be obtained for any $\phi_\sigma\in \mathcal{K}_3$. (Note that Corollary \[thm:F3\], part \[item:F2zminoverB\] can be obtained either as a consequence of Theorem \[thm:F4\], part \[item:F3zminoverB\] or using Lemma \[lem:F3partresult\]. The latter option allows for less restrictions on $\phi_\sigma$ in Corollary \[thm:F3\] than in Theorem \[thm:F4\].)
We conclude this section with a few comments on the results, their interpretations and limitations. Our [*brute force*]{} arguments are based upon explicit calculations and are as such limited to a blurring kernel with [*small*]{} (with respect to $\omega$) support. Indeed, this is most discouraging for the deconvolution functionals $F_2$ and $F_3$ where one one would expect the regime of acceptable $\sigma$ to extend far beyond the $X$-dimension of the underlying bar code. Numerical experiments (see Section \[numerics\]) support this conjecture. The conditions that $\sigma, \rho \leq w/2$ for $F_2$ and $F_3$ may seem particularly alarming since the analogous condition for $F_1$ is simply $\sigma \leq \omega$. However, note that the other condition (\[eq:conditiononrhosigmalambda\]) also puts a restriction on the size of $\sigma$ which is essentially of the same form. For $F_2$, one could leave out the condition $\sigma \leq \frac\omega2$ and not change the principles of the proof, but many more orderings in the computation of the integrals become possible (see Remark \[rem:roleofconditions\]) and many more calculations need to be done in the proof of Lemma \[lem:F3partresult\]. Since we still have condition (\[eq:condsigmalambdaconvolutions\]) on $\sigma$ in place (these extra calculations can only replace (\[eq:condsigmalambdaconvolutions\]) by a stricter condition on $\sigma$, if anything changes at all) it is doubtful that much can be won by leaving out the condition $\sigma \leq \frac\omega2$. For $F_3$ the conditions $\rho, \sigma \leq \frac\omega2$ are vital to our proofs via Lemmas \[lem:levelhalfrhosigmaz\], \[lem:decreasing\], \[lem:F3firstvar\].
The numbers in condition (\[eq:conditiononrhosigmalambda\]) may seem a little strange. They are simply a consequence of the direct calculations with the hat function. As we have remarked, these calculations can be repeated for other kernels in $\mathcal{K}$. Note that, taking $\rho = \omega/2$, the condition implies that the bar code is always recoverable for any $\sigma \leq \omega/2$, provided $\lambda > 20 / 3\omega$.
Another surprising condition might be $\sigma \leq \rho$ in Theorem \[thm:F4\], part \[item:F3zminoverB\]. In general if $\rho < \sigma$ we do not expect $z$ to be a minimiser of $F_3$ over ${{\cal B}}$ (or ${{\cal B}}^{ij}$), even if $\rho, \sigma \leq \frac\omega2$ and (\[eq:conditiononrhosigmalambda\]) are satisfied. A counter example in this case is given by $z=\chi_{[0.425, 0.575]}$ and $u = \chi_{[0.425, 0.4999]} + \chi_{[0.5001, 0.575]}$ with $\rho=0.05$ and $\sigma=0.06$. The fidelity term in $F_3(u)$ is smaller than the fidelity term in $F_3(z)$ ($\|\phi_\rho*u-\phi_\sigma*z\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 \approx 2.378 \cdot 10^{-4}$ and $\|\phi_\rho*z-\phi_\sigma*z\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 \approx 2.407 \cdot 10^{-4}$) and thus for $\lambda$ large enough $z$ will not be the minimiser of $F_3$ in ${{\cal B}}$. In this particular case the difference is small and so in practical applications where $\lambda$ is not too large it might not cause problems, since then the energetic cost $2$ for two extra interfaces will be much higher than the gain in the fidelity term. Extra conditions on the parameters in the case $\rho < \sigma$ might ensure that $u=z$ is the minimiser for $F_3$. The above example suggests that an upper bound for $\lambda$ may be in order. In fact, numerical simulations in Section \[numerics\] show that the regime $\rho < \sigma$ poses no problem for suitable [*midrange*]{} choices of $\lambda$. Indeed, they suggest that fixing $\rho$ comparable with the $X$-dimension and minimising $F_3$ works well for $\sigma$ up to twice the $X$-dimension.
We know however that in the degenerate case $0=\rho < \sigma$, Theorem \[thm:minimofF2\] assures that we have $u=z$ as minimiser if both conditions in (\[eq:conditionsigmalambda\]) are satisfied, without an upper bound on $\lambda$. Why the second condition in (\[eq:conditionsigmalambda\]) is the correct degenerate form of (\[eq:conditiononrhosigmalambda\]) can be seen by recognizing their common source (\[eq:conditiononrhosigmalambda2\]).
Finally, we note that there is a wealth of work on total variation energy minimisation for image analysis. While our results are for rather simple one dimensional images, we feel they are novel in that the 1D bar code setting entails an image deblurring problem of contemporary interest yielding very precise results, and we are unaware of any general method for analogous deblurring functionals which would yield similar results. In addition to geometric simplicity due to its binary nature (the simplest case of what is called *Quantum TV* in [@ShenKang07]), the bar code problem is different from many other imaging problems in that there is a known [*a priori*]{} lower bound on the length scale of the structures in the image (via the $X$-dimension). An analytically deeper study entails deblurring of 2D bar codes [@ChoksiGennipOberman10].
Proofs of the Theorems {#sec:proofs}
======================
Existence and the trivial minimiser
-----------------------------------
\[lemma0\] Let $\phi_\sigma \in \mathcal{K}$ and fix $z\in {{\cal B}}_\omega$ (${{\cal B}}_\omega^{ij}$) and $\lambda, \sigma, \rho > 0$. Then minimisers for $ F_1, F_2$ and $ F_3$ over ${{\cal B}}$ (${{\cal B}}^{ij}$) exist.
The proof is a simple application of the [*direct method in the calculus of variations*]{} and follows along the same lines for all these functionals. For completeness, we present it for $F_1$ and $z\in {{\cal B}}_\omega$. Let $\{u_n\}$ be a minimising sequence for $F_1$ in ${{\cal B}}$, then we can assume every $u_n$ has bounded $L^1$-norm and bounded BV measure. Therefore, by , there exists $u\in BV([0,1])$ such that $u_n \to u$ in $L^1([0,1])$. Since the $u_n$ only take values $0$ and $1$ (and $0$ a.e. in $[0,1]^c$), so does $u$. Thus $u \in {{\cal B}}$.
The total variation is lower semicontinuous under $L^1$ convergence and under the special conditions that the functions only take values $0$ and $1$, so is the $L^2$ norm, therefore we conclude via the direct method in the calculus of variations that $u$ is a minimiser for $F_1$. For the functionals $F_2$ and $F_3$ we use in addition, that the functional $u\mapsto \phi_\sigma * u$ is continuous under $L^1$ convergence, for any $\sigma>0$. If we replace ${{\cal B}}_\omega$ and ${{\cal B}}$ by ${{\cal B}}_\omega^{ij}$ and ${{\cal B}}^{ij}$ respectively the proof does not change.
Note that we have not used the fact that $\phi_\sigma$ is symmetric and unimodal with compact support in the above. We only need continuity of $u\mapsto \phi_\sigma * u$ under $L^1$ convergence.
Next we recall a result about convolutions whose proof follows directly from Fubini’s Theorem.
\[eq:intconv\] Let $f, g, h \in L^2({{{\ensuremath{\mathbb{R}}}}})$ such that $f(-x)=f(x)$, then $$\int_{{{\ensuremath{\mathbb{R}}}}}\big[(f*g)\cdot h\big] = \int_{{{\ensuremath{\mathbb{R}}}}}\big[g \cdot(f*h)\big].$$
Parts 1 of Theorems \[thm:minimofF2\], \[thm:F4\] and Corollary \[thm:F3\] follow directly from the following lemma.
\[lem:trivialminimiser\] Let $i\in\{1, 2, 3\}$, $\phi_\sigma \in \mathcal{K}$, and $\lambda>0$, then the following two statements are equivalent:
1. \[item:trivialminimiser\] $u=0$ is the unique minimiser of $F_i$ over $BV({{{\ensuremath{\mathbb{R}}}}})$.
2. \[item:starnorms\]
1. If $i=1$, $\|f_\sigma\|_* \leq \frac1{2\lambda}$.
2. If $i=2$, $\|\phi_\sigma*f_\sigma\|_* \leq \frac1{2\lambda}$.
3. If $i=3$, $\|\phi_\rho*f_\sigma\|_* \leq \frac1{2\lambda}$.
The idea of the proof is similar to that in [@Meyer01 §1.14, Lemma 4]. Note that for general $u\in BV({{{\ensuremath{\mathbb{R}}}}})$ we cannot conclude that $\phi_\sigma*u \in L^2({{{\ensuremath{\mathbb{R}}}}})$. For functions $u$ and parameters $\sigma$ or $\rho$ for which this fails we set $F_i(u)=\infty$, $i\in \{2, 3\}$.
First let $i=1$. We first prove \[item:trivialminimiser\] $\Longrightarrow$ \[item:starnorms\]. Assume $u=0$ is the unique minimiser of $F_1$ in $BV({{{\ensuremath{\mathbb{R}}}}})$. This is equivalent to, for all $h\in BV({{{\ensuremath{\mathbb{R}}}}})$ with $h\neq 0$, $$\label{eq:uisatrivialminimiser}
\lambda \|f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 < \int_{{{\ensuremath{\mathbb{R}}}}}|h'| + \lambda \|h-f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 = \int_{{{\ensuremath{\mathbb{R}}}}}|h'| + \lambda \|f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + \lambda \|h\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 - 2 \lambda \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma h.$$ Because this holds for all $h\in BV({{{\ensuremath{\mathbb{R}}}}})$, by rescaling $h$ we can rewrite this as $$\label{eq:uisatrivialminimiser2}
2 \lambda \e \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma h < |\e| \int_{{{\ensuremath{\mathbb{R}}}}}|h'| + \lambda \e^2 \|h\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2,$$ for all $\e\in{{{\ensuremath{\mathbb{R}}}}}$ and all $h\in BV({{{\ensuremath{\mathbb{R}}}}})$. Dividing by $\epsilon$, taking the limit $\e\to 0$, and recognizing that $\e$ can be positive and negative, we find that (\[eq:uisatrivialminimiser2\]) implies $$\label{eq:equivstatement}
\left| \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma h \right| \leq \frac1{2\lambda} \int_{{{\ensuremath{\mathbb{R}}}}}|h'|, \qquad \text{for all } h\in BV({{{\ensuremath{\mathbb{R}}}}}).$$ Now per definition $$\|f_\sigma\|_* = \underset{v\in L^1({{{\ensuremath{\mathbb{R}}}}}), \int_{{{\ensuremath{\mathbb{R}}}}}|v'| \leq 1}{\sup}\, \left|\int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma v\right|
\leq \frac1{2\lambda},$$ where the inequality follows by taking the supremum in (\[eq:equivstatement\]) over all\
$h\in \left\{v\in L^1({{{\ensuremath{\mathbb{R}}}}}): \int_{{{{\ensuremath{\mathbb{R}}}}}} |v'| \leq 1\right\} \subset BV({{{\ensuremath{\mathbb{R}}}}})$.
To prove \[item:starnorms\] $\Longrightarrow$ \[item:trivialminimiser\] let $\|f_\sigma\|_* \leq \frac1{2\lambda}$. Then for all $v \in L^1({{{\ensuremath{\mathbb{R}}}}})$ satisfying $\int_{{{\ensuremath{\mathbb{R}}}}}|v'| \leq 1$ we have $$\left|\int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma v \right| \leq \frac1{2\lambda},$$ from which it follows that for all $h\in BV({{{\ensuremath{\mathbb{R}}}}})$ $$\left| \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma \frac{h}{\int_{{{\ensuremath{\mathbb{R}}}}}|h'|}\right| \leq \frac1{2\lambda}.$$ Inequality (\[eq:equivstatement\]) now follows.
We proved above that (\[eq:uisatrivialminimiser\]) implies (\[eq:equivstatement\]). On the other hand we see that inequality (\[eq:equivstatement\]) implies for $h\neq 0$ $$\int_{{{\ensuremath{\mathbb{R}}}}}|h'| + \lambda \|f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + \lambda \|h\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 - 2 \lambda \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma h \geq \lambda \|f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + \lambda \|h\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 > \lambda \|f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2,$$ and thus inequality (\[eq:equivstatement\]) is equivalent to (\[eq:uisatrivialminimiser\]). This proves the result for $i=1$.
$F_2$ is a special case of $F_3$ (with $\rho=\sigma$). For $i=3$ we can derive a statement analogous to inequality (\[eq:equivstatement\]), with $h$ on the left hand side replaced by $\phi_\rho*h$. Having $u=0$ as unique minimiser of $F_3$ in $BV({{{\ensuremath{\mathbb{R}}}}})$ is equivalent to $$\left| \int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma \cdot \phi_\rho*h \right| \leq \frac1{2\lambda} \int_{{{\ensuremath{\mathbb{R}}}}}|h'|, \qquad \text{for all } h\in BV({{{\ensuremath{\mathbb{R}}}}}).$$ By Lemma \[eq:intconv\] we recognise that $$\int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma \cdot \phi_\rho*h = \int_{{{\ensuremath{\mathbb{R}}}}}\phi_\rho * f_\sigma \cdot h$$ and the result follows as before.
Although we assume $\phi_\sigma\in \mathcal{K}$ in the proof above because that is the most general class of kernels we consider, we only use the symmetry and integrability of $\phi_\sigma$.
Proof of Theorem \[thm:minimofF2\], part \[item:F1zminoverB\]
-------------------------------------------------------------
We now turn our attention from the trivial minimiser $u=0$ to $u=z$ as minimiser. First, we present some elementary results.
\[lem:onebarconvolution\] Let $\phi_\sigma \in \mathcal{K}$, $a<b$, $\sigma \leq b-a$ and $z=\chi_{[a, b]}$, then $$\left\{ x\in {{{\ensuremath{\mathbb{R}}}}}: f_\sigma(x) = \frac12\right\} = \{a, b\} \quad \text{and} \quad \left\{ x\in {{{\ensuremath{\mathbb{R}}}}}: f_\sigma(x) \geq \frac12\right\} = [a, b].$$
We compute $$f_\sigma(a) = \int_a^b \phi_\sigma(a-y)\,dy = \int_a^{a+\sigma} \phi_\sigma(a-y)\,dy = \int_{-\sigma}^0\phi_\sigma(y)\,dy = \frac12$$ and by symmetry $f_\sigma(b)=\frac12$.
Furthermore for $x\in (a-\sigma, b+\sigma)$ we compute $$f_\sigma'(x) = \int_a^b \phi_\sigma'(x-y)\,dy = \int_{x-b}^0 \phi_\sigma'(y)\,dy + \int_0^{x-a} \phi_\sigma'(y)\,dy.$$ The first term on the right is nonnegative if $x \leq b$ and nonpositive if $x\geq b$ and the second term is nonpositive if $x\geq a$ and nonnegative if $x\leq a$. By symmetry of $\phi_\sigma$ we then conclude that $f_\sigma'(x) \geq 0$ if $x\leq \frac{a+b}2$ and $f_\sigma'(x) \leq 0$ if $x\geq \frac{a+b}2$. Moreover we have $f_\sigma'(a) >0$ and $f_\sigma'(b)<0$.
\[lem:lessthanahalf\] Let $\phi_\sigma \in \mathcal{K}$, $z\in {{\cal B}}_\omega$, and $\sigma \leq \omega$, then for every $x\in {{{\ensuremath{\mathbb{R}}}}}\setminus \operatorname{supp}z$ we have $\phi_\sigma*z(x) < \frac12$.
Let $x \in {{{\ensuremath{\mathbb{R}}}}}\setminus \operatorname{supp}z$, then there exist $a<b$ such that $b-a \geq \omega$, $x \in (a,b)$, and $z(y)=0$ for all $y\in (a,b)$. Define $z_0 := \chi_{(-\infty, a)} + \chi_{(b, \infty)}$, then $$\phi_\sigma * z(x) \leq \phi_\sigma * z_0(x)
= \int_{{{\ensuremath{\mathbb{R}}}}}\phi_\sigma(x-y)\, dy - \int_a^b \phi_\sigma(x-y)\,dy
= 1- \phi_\sigma*\chi_{[a,b]}(x)
< \frac12.$$ The final inequality follows since $\sigma \leq \omega \leq b-a$ and thus by Lemma \[lem:onebarconvolution\] $\phi_\sigma*\chi_{[a,b]} > \frac12$ on $(a, b)$.
The following lemma allows us to consider only minimisers of $F_1$ that have less interfaces than $z$ or are equal to $z$.
\[lem:firstvar-F1\] Let $\phi_\sigma \in \mathcal{K}$, $z\in {{\cal B}}_{\omega}$ and let $u$ be a minimiser of $F_1$ over ${{\cal B}}$. Denote by $x_i$ the locations of the interfaces of $u$, then we have for every $i$ $$f_\sigma(x_i)=\frac12.$$ Consequently if $\sigma \leq \omega$, $x_i$ is the location of an interface of $z$ for every $i$.
Assume without loss of generality that $z\neq 0$. Let $u$ minimise $F_1$ over ${{\cal B}}$. We show that vanishing first variation of $u$ implies that at any interface $x_i$, we must have $f_\sigma (x_i) = \frac12$. To this end, consider an interface of transition from $u=1$ to $u=0$ at $x_i$ (the other case is treated similarly). By considering a perturbation consisting of extending the $u = 1$ bar up to $x_i + t$ for $t$ small, one obtains no change in the total variation and a change in the fidelity term of $$\int_{x_i}^{x_i + t} \Bigl((1 - f_\sigma)^2 - f_\sigma^2\Bigr)\, dx \, = \,
\int_{x_i}^{x_i + t} (1 - 2f_\sigma) \, dx.$$ Differentiating with respect to $t$ and setting $t = 0$ gives $1 - 2 f_\sigma (x_i) = 0$.
Let now $\sigma \leq \omega$. By Lemma \[lem:onebarconvolution\] if $z$ consists of one bar only the $\frac12$-level set of $f_\sigma$ is exactly the set of locations of the interfaces of $z$. If $z$ has more bars Lemma \[lem:lessthanahalf\] assures that the $\frac12$-lower level set is not affected by the convolutions of different bars interacting.
Lemma \[lem:firstvar-F1\] tells us that, if $\sigma\leq \omega$, any candidate for minimising $F_1$ over ${{\cal B}}$ not equal to $z$, should have less interfaces than $z$ which are located at places where $z$ also has an interface. We use this to complete the proof of Theorem \[thm:minimofF2\]. First, we introduce a notation that will be used frequently in what follows. For $\sigma>0$, $a<b$, and $x\in{{{\ensuremath{\mathbb{R}}}}}$ define the functions $$\label{eq:Ipm}
I_\pm^\sigma(x, a, b) := \frac1\sigma \int_a^b \left(1\pm \frac{x-y}\sigma\right)\, dy = \frac1\sigma \left( (b-a) \left( 1\pm\frac{x}\sigma\right) \mp \frac{b^2-a^2}{2\sigma} \right).$$ Note that this definition is tailored to the needs of the hat function $\hat \phi_\sigma$. If we want to reproduce the calculations that follow for a general convolution kernel $\phi_\sigma\in \mathcal{K}$ we can write $\phi_\sigma$ as in (\[eq:Ksigma\]) and define $I_\pm^\sigma$ as $$I_+^\sigma(x, a, b) := \int_a^b p(y-x, \sigma)\, dy \quad \text{and} \quad I_-^\sigma(x, a, x) := \int_a^b p(x-y, \sigma)\, dy.$$
We will only state the results for the hat function and hence use the definitions in (\[eq:Ipm\]).
[**Proof of Theorem \[thm:minimofF2\], part \[item:F1zminoverB\].**]{} Let ${{\cal B}}\ni u_0\neq z$ be a minimiser of $F_1$. By Lemma \[lem:firstvar-F1\], the number of interfaces of $u_0$ must be less than the number of interfaces of $z$ and the location of every interface of $u_0$ coincides with the location of an interface of $z$. Therefore there exists a connected interval $N\subset[0,1]$ such that $|N| \ge \omega$ and either
- $z=0$ and $u_0=1$ on $N$, or
- $z=1$ and $u_0=0$ on $N$.
First assume the former case, and let $$\hat u := \left\{ \begin{array}{ll} u_0 & \text{on } N^c \\ u_0-1=z & \text{on } N. \end{array}\right.$$ We compute $$\label{eq:jumpestimate}
\int_{{{{\ensuremath{\mathbb{R}}}}}} |\hat u'| \leq 2+\int_{{{{\ensuremath{\mathbb{R}}}}}} |u_0'|$$ and $$\begin{aligned}
\|u_0 - f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 &= \|u_0 - \hat u + \hat u - f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2\notag\\
&= \|u_0-\hat u\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + \|\hat u - f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + 2 \int_{{{{\ensuremath{\mathbb{R}}}}}} (u_0-\hat u) (\hat u - f_\sigma)\notag\\
&= \|\hat u - f_\sigma\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + |N| + 2 \int_{{{{\ensuremath{\mathbb{R}}}}}} (u_0-\hat u) (\hat u - f_{\sigma}),\label{eq:L2normestimate}\end{aligned}$$ where we have used that $$\|u_0-\hat u\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 = |N|.$$
Let $a\in [0,1]$ be such that $N=[a, a+|N|]$ and thus $N^c\cap[0,1]=[0,a)\cup(a+|N|,1]$, then we compute $$\begin{aligned}
\label{eq:twointegrals}
-2 \int_{{{{\ensuremath{\mathbb{R}}}}}} (u_0-\hat u) (\hat u - f_\sigma) &= 2 \int_N \int_{{{{\ensuremath{\mathbb{R}}}}}} \hat\phi_{\sigma}(x-y) z(y) \, dy \, dx \leq 2 \int_N \int_{N^c\cap[0,1]} \hat\phi_{\sigma}(x-y)\, dy \, dx \nonumber\\
&= 2 \int_a^{a+|N|} \left\{\int_0^a \hat\phi_{\sigma}(x-y)\,dy + \int_{a+|N|}^1 \hat\phi_{\sigma}(x-y)\, dy\right\}\, dx.\end{aligned}$$ Integrals as those in the right hand side of (\[eq:twointegrals\]) are commonplace in the proofs of this paper. It is therefore very illustrative to work out one of them in detail. Let us consider $$\int_a^{a+|N|} \int_0^a \hat\phi_{\sigma}(x-y)\,dy \,dx.$$ Per definition $\hat\phi_\sigma(x-y)$ is zero if $|x-y|\geq\sigma$ and on its support its value is given by $\frac1\sigma \left(1+\frac{x-y}\sigma\right)$ if $x-y\in (-\sigma, 0)$ and by $\frac1\sigma \left(1-\frac{x-y}\sigma\right)$ if $x-y\in (0,\sigma)$. Let us fix $x\in[a, a+|N|]$ for the moment and remember that $y\in(0,a)$ in the integral, then $\phi_\sigma(x-y) = \frac1\sigma \left(1-\frac{x-y}\sigma\right)$ if $$y \in (x-\sigma, x) \cap (0, a) = \left\{ \begin{array}{ll}
\emptyset & \text{if } x-\sigma < x < 0 < a,\\
(0, x) & \text{if } x-\sigma < 0 < x < a,\\
(0, a) & \text{if } x-\sigma < 0 < a < x,\\
(x-\sigma, x) & \text{if } 0 < x-\sigma < x < a,\\
(x-\sigma, a) & \text{if } 0< x-\sigma< a < x,\\
\emptyset & \text{if } 0 < a < x-\sigma < x.
\end{array}\right.$$ Because $x\in[a, a+|N|]$ we can rule out some of these cases[^3] and end up with $$y \in \left\{ \begin{array}{ll}
(0, a) & \text{if } x \in (-\infty, \sigma) \cap (a, \infty) = \left\{ \begin{array}{ll} \emptyset & \text{if } \sigma \leq a,\\ (a, \sigma) & \text{if } a < \sigma,\end{array}\right.\\
(x-\sigma, a) &\text{if } x \in (\sigma, a+\sigma) \cap (a, \infty) = \left\{ \begin{array}{ll} (a, a+\sigma) & \text{if } \sigma \leq a,\\ (\sigma, a+\sigma) & \text{if } a < \sigma,\end{array}\right.\\
\emptyset & \text{if } x\in (a+\sigma, a+|N|).
\end{array}\right.$$ We see that we have to distinguish between the cases $\sigma \leq a$ and $a < \sigma$. Similarly we find that $\hat\phi_\sigma(x-y) = \frac1\sigma \left(1+\frac{x-y}\sigma\right)$ if $y \in (x, x+\sigma) \cap (0, a)$, which is the empty set because of the restrictions on $x$.
This now leads us to the computation $$\int_a^{a+|N|} \int_0^a \hat\phi_{\sigma}(x-y)\,dy \,dx = \left\{ \begin{array}{ll}
\frac1\sigma \int_a^{a+\sigma} \int_{x-\sigma}^a \left(1-\frac{x-y}\sigma\right)\,dy\,dx & \text{if } \sigma \leq a,\\
\frac1\sigma \left[ \int_a^\sigma \int_0^a \left(1-\frac{x-y}\sigma\right)\,dy\,dx + \int_\sigma^{a+\sigma} \int_{x-\sigma}^a \left(1-\frac{x-y}\sigma\right)\,dy\,dx\right] & \text{if } a < \sigma.
\end{array}\right.$$ Because all the integrands are positive, in the case $a<\sigma$ we can estimate $$\int_a^\sigma \int_0^a \left(1-\frac{x-y}\sigma\right)\,dy\,dx \leq \int_a^\sigma \int_{x-\sigma}^a \left(1-\frac{x-y}\sigma\right)\,dy\,dx.$$ Therefore we conclude that for both $\sigma \leq a$ and $a<\sigma$ $$\int_a^{a+|N|} \int_0^a \hat\phi_{\sigma}(x-y)\,dy \,dx \leq \int_a^{a+\sigma} I_-^\sigma(x, x-\sigma, a)\, dx.$$
In a similar fashion we compute $$\int_a^{a+|N|} \int_{a+|N|}^1 \hat\phi_{\sigma}(x-y)\, dy\, dx \leq \int_{a+|N|-\sigma}^{a+|N|} I_+^\sigma(x, a+|N|, x+\sigma)\,dx.$$ During this computation we need to distinguish between the cases $a+|N| \leq 1-\sigma$ and $a+|N| > 1-\sigma$, but as before this distinction doesn’t play a role in the final estimate.
Continuing from (\[eq:twointegrals\]) we now find $$\begin{aligned}
-2 \int_{{{{\ensuremath{\mathbb{R}}}}}} (u_0-\hat u) (\hat u - \hat\phi_{\sigma} * z) &\leq 2 \left\{ \int_a^{a+\sigma} I_-^\sigma(x, x-\sigma, a)\, dx + \int_{a+|N|-\sigma}^{a+|N|} I_+^\sigma(x, a+|N|, x+\sigma)\,dx\right\}\\
&= \frac23 \sigma.\end{aligned}$$ Using this in (\[eq:jumpestimate\]–\[eq:L2normestimate\]) together with $|N| \geq \omega$ we find $$F_1 (\hat u) \leq F_1 (u_0) + 2 + \lambda \left(\frac23 \sigma - \omega \right) < F_1(u_0),$$ where the second inequality follows by condition (\[eq:conditionsigmalambda\]). This contradicts $u_0$ being a minimiser of $F_1$.
Next we consider the second case, i.e. $z=1$ and $u_0=0$ on $N$. We define $$\bar u := \left\{ \begin{array}{ll} u_0 & \text{on } N^c \\ u_0+1=z & \text{on } N. \end{array}\right.$$ As in the first case, we will find an estimate for the integral in the brackets in the right hand side of (\[eq:L2normestimate\]), but for $\bar u$ instead of $\hat u$: $$-2 \int_{{{{\ensuremath{\mathbb{R}}}}}} (u_0-\bar u) (\bar u - \hat\phi_{\sigma} * z) = 2 \int_N (1-\hat\phi_{\sigma} * z) = 2 |N| - 2 \int_N \hat\phi_{\sigma} * z.$$ Again we write $N=[a, a+|N|]$ and we compute $$\begin{aligned}
\int_N \phi_{\sigma} * z &= \int_N \int_0^1 \hat\phi_{\sigma}(x-y) z(y) \, dy \,dx \geq \int_N \int_N \hat\phi_{\sigma}(x-y) z(y)\,dy\,dx\\
&= \int_a^{a+|N|}\int_a^{a+|N|} \hat\phi_{\sigma}(x-y)\,dy\,dx\\
&= \int_a^{a+\sigma} I_-^\sigma(x, a, x)\, dx + \int_{a+\sigma}^{a+|N|} I_-^\sigma(x, x-\sigma, x) \,dx\\
&\hspace{0.4cm} + \int_a^{a+|N|-\sigma} I_+^\sigma(x, x, x+\sigma) \,dx + \int_{a+|N|-\sigma}^{a+|N|} I_+^\sigma(x, x, a+|N|)\,dx\\
&= |N| - \frac13\sigma.\end{aligned}$$ As in the first case we now find $$F_1(\bar u) \leq F_1(u_0) + 2 + \lambda \left(\frac23 \sigma - \omega \right) < F_1 (u_0),$$ which is again a contradiction with $u_0$ being a minimiser. Therefore the only candidate for a minimiser is $u=z$ and hence by Lemma \[lemma0\] $u=z$ is the unique minimiser.
In the above proof everything up to and including (\[eq:twointegrals\]) is independent of the choice of specific blurring kernel and we could have used any $\phi_\sigma \in \mathcal{K}$. The explicit calculations that follow in the remainder of the proof depend on our choice $\phi_\sigma=\hat\phi_\sigma$, but can be redone for a different choice of kernel as explained in the paragraphs preceding the proof.
Proofs of Theorem \[thm:F4\], part \[item:F3zminoverB\] and Corollary \[thm:F3\], part \[item:F2zminoverB\]
-----------------------------------------------------------------------------------------------------------
We now turn to $F_3$.
\[lemma1\] Let $z_1, z_2 \in {{\cal B}}$. Both $z_1$ and $z_2$ on $[0,1]$ consist of a finite collection of subintervals of $[0,1]$, i.e. alternating bars and spaces. Let $t_i$, $i = 1 \dots n$ and $t_i'$, $i = 1 \dots n'$ denote the right hand sides of the intervals of $z_1$ and $z_2$ respectively. In particular $t_n=t_{n'}'=1$. If $n > n'$, then there exists an interval $N\subset [0,1]$ such that $[t_i, t_{i+1}] \subset N$ for some $i$; and for all $ x \in N$, either $$\label{def-N}
z_1(x) =0 \,\,\, {\rm and} \,\,\, z_2 (x) =1 \qquad {\rm or} \qquad z_1 (x) =1 \,\,\, {\rm and} \,\,\, z_2 (x) =0.$$ In particular, if $z_1\in {{\cal B}}_{\omega}$, then $|N|\geq \omega$.
First assume that $z_1$ starts with a bar and $z_2$ starts with a space, i.e. $z_1=1$ on $[0, t_1]$ and $z_2=0$ on $[0, t_1']$. If $t_1\leq t_1'$ then $[0, t_1]\subset N$. Suppose $t_1' \leq t_1$. If the conclusion of the lemma is false, then for all $i \leq n'$, $t_i' < t_i$. This is a contradiction since $t'_{n'} = t_n = 1$. If $z_1$ starts with a space and $z_2$ starts with a bar we arrive at a similar conclusion.
Now assume that $z_1$ and $z_2$ both start with a bar (the situation in which both start with a space is similar). Note that $z_1 = 1$ on $[0, t_1]$ and $z_2 = 1$ on $[0, t_1']$. Suppose $t_1 \le t_1'$. Then if the conclusion of the lemma is false, we must have $t_i' < t_{i+1}$ for $i=1 \dots n'$ which implies $1 = t'_{n'} < 1$. Suppose $t_1 > t_1'$. If for some $i>1$, we have $t_i' \ge t_i$, then the previous argument again gives a contradiction. Thus we must have $t_i > t_i'$ for all $i = 2 \dots n' -1$. But then (\[def-N\]) must hold on one of the intervals $[t_i, t_{i+1}]$, for $i \ge n'$.
\[lem:F3partresult\] Let $z
$ implicit in the definition of $ F_3$, be taken from $
\in {{\cal B}}_{\omega}^{ij}$ for some $i, j\in \{0,1\}$, $\rho, \sigma \leq \frac{\omega}2$, $\phi_\sigma = \hat\phi_\sigma$ and define $$\label{eq:thisfunctionofrhoandsigma}
f(\rho, \sigma) := \left\{ \begin{array}{ll} \frac1{\rho^2} \Bigl(-\sigma^3 + 5 \rho \sigma^2 + 10 \rho^3\Bigr) & \text{if } \sigma\leq \rho,\\ \frac1{\sigma^2} \Bigl( -\rho^3+5 \sigma \rho^2 + 10 \sigma^3\Bigr) & \text{if } \rho\leq\sigma.\end{array} \right.$$ Let $\lambda$, $\rho$, and $\sigma$ satisfy in addition $$\label{eq:conditiononrhosigmalambda2}
\frac2\lambda + \frac1{15} \Bigl(7 \rho + f(\rho, \sigma)\Bigr) < \omega.$$ If $u \in {{\cal B}}^{ij}$ is a minimiser of $F_3$ over ${{\cal B}}^{ij}$, then $$\int_{{{\ensuremath{\mathbb{R}}}}}|u'| \, \geq \, \int_{{{\ensuremath{\mathbb{R}}}}}|z'|.$$
We prove this by contradiction. Let ${{\cal B}}^{ij}\ni u_0 \neq z$ be a minimiser of $F_3$ in ${{\cal B}}^{ij}$ and assume that $u_0$ has less interfaces than $z$, i.e. $\int_{{{\ensuremath{\mathbb{R}}}}}|u_0'| \, < \, \int_{{{\ensuremath{\mathbb{R}}}}}|z'|$. By Lemma \[lemma1\], there exists a connected interval $N\subset[0,1]$ such that $|N| \ge \omega$ and either
- $z=0$ and $u_0=1$ on $N$, or
- $z=1$ and $u_0=0$ on $N$.
Define $$\hat u := \left\{\begin{array}{ll} u_0 & \text{on } N^c,\\ z & \text{on } N, \end{array}\right.$$ then $$\int |\hat u'| \leq \int |u_0'| + 2$$ and $$\begin{aligned}
\|\hat\phi_\rho * u_0 - \hat\phi_\sigma*z\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 &= \|\hat\phi_\rho * \hat u -\hat\phi_\sigma*z\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2 + \|\hat\phi_\rho * (u_0-\hat u)\|_{L^2({{{\ensuremath{\mathbb{R}}}}})}^2\\ &\hspace{0.4cm}+ 2 \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho * (u_0-\hat u)\Bigr) \cdot \Big(\hat\phi_\rho * \hat u - \hat\phi_\sigma* z\Big),\end{aligned}$$ from which we conclude that $$\label{eq:F4inequality}
F_3(\hat u) \leq F_3(u_0) + 2 - \lambda \Biggl( \|\hat\phi_{\rho} * (u_0-\hat u)\|_{L^2}^2 + 2 \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_{\rho} * (u_0-\hat u)\Big) \cdot \Bigl(\hat\phi_{\rho}*\hat u - \hat\phi_{\sigma}*z\Bigr)\Biggr).$$ Because $u_0-\hat u = \pm \chi_N$, Lemma \[lem:fsigmasquared\] gives $$\|\hat\phi_{\rho} * (u_0-\hat u)\|_{L^2}^2 = |N| - \frac7{15} \rho.$$ Next we again distinguish two cases: Case I in which $u_0=1$ and $\hat u = z = 0$ on $N$ and Case II in which $u_0=0$ and $\hat u = z = 1$ on $N$. We first treat Case I: $$\begin{aligned}
&\hspace{0.4cm}2 \int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho * (u_0-\hat u) \Bigl(\hat\phi_\rho * \hat u - \hat\phi_\sigma * z \Bigr)\\
&= 2 \int_R \int_N \hat\phi_\rho(x-y)\,dy \biggl(\int_{N^c\cap[0,1]} \hat\phi_\rho(x-w) \hat u(w)\,dw - \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w) z(w)\,dw\biggr)\,dx\\
&\geq -2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx.\end{aligned}$$ Now we subdivide Case I into two subclasses: Case Ia in which $\sigma\leq\rho$ and Case Ib in which $\rho\leq\sigma$. For Case Ia we compute $$-2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx = \frac1{15 \rho^2} \Bigl(\sigma^3 - 5 \rho \sigma^2 - 10 \rho^3\Bigr).$$ For details of this computation we refer to (\[eq:lemCaseIa\]) in Appendix \[sec:moredetails\].
In Case Ib the computation is $$-2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx = \frac1{15\sigma^2} \Bigl(\rho^3 - 5 \sigma \rho^2 - 10 \sigma^3\Bigr),$$ the details of which can be found in (\[eq:lemCaseIb\]) in Appendix \[sec:moredetails\].
In Case II we compute $$\begin{aligned}
&\hspace{0.4cm}2 \int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho * (u_0-\hat u) \Bigl(\hat\phi_\rho * \hat u - \hat\phi_\sigma * z \Bigr)\\
&= -2\int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \biggl(\int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho(x-w)\hat u(w)\,dw - \int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\sigma(x-w) z(w)\,dw\biggr)\,dx\\
&\geq -2\int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \biggl(\int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho(x-w)\,dw - \int_N\hat\phi_\sigma(x-w)\,dw\biggr)\,dx.\end{aligned}$$
For the first term we find $$-2\int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho(u-w)\,dw\,dx =-2N.$$ Details of this calculation are given in (\[eq:lemCaseIIfirstterm\]) in Appendix \[sec:moredetails\].
For the second term again we need to subdivide into Case IIa in which $\sigma\leq\rho$ and Case IIb in which $\rho\leq\sigma$. For Case IIa we compute $$2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y) \,dy \int_N \hat\phi_\sigma(x-w)\,dw\,dx = 2N + \frac1{15 \rho^2} \Bigl(\sigma^3 - 5 \rho \sigma^2 - 10 \rho^3\Bigr).$$ For more details of this computation see (\[eq:lemCaseIIa\]) in Appendix \[sec:moredetails\].
In Case IIb we can repeat the calculation with $\rho$ and $\sigma$ interchanged to get $$2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y) \,dy \int_N \hat\phi_\sigma(x-w)\,dw\,dx = 2N + \frac1{15 \sigma^2} \Bigl(\rho^3 - 5 \rho \sigma^2 - 10 \sigma^3\Bigr).$$
Using the combined results of Cases I and II in inequality (\[eq:F4inequality\]) leads to $$\begin{aligned}
F_3(\hat u) &\leq F_3(u_0) + 2 - \lambda \biggl( |N| -\frac1{15} \Big(7 \rho + f(\rho, \sigma)\Big) \biggr)\\
&\leq F_3(u_0) + 2 - \lambda \biggl( \omega -\frac1{15} \Big(7 \rho + f(\rho, \sigma)\Big) \biggr)\\
&< F_3(u_0),\end{aligned}$$ where the final inequality follows from (\[eq:thisfunctionofrhoandsigma\]) - (\[eq:conditiononrhosigmalambda2\]). This contradicts the fact that $u_0$ is a minimiser of $F_3$.
\[rem:roleofconditions\] In the proof of Lemma \[lem:F3partresult\] we have used the conditions $z\in {{\cal B}}_\omega^{ij}$, $u_0\in {{\cal B}}^{ij}$, and $\rho, \sigma \leq \frac\omega2$ but it might not be immediately clear where. They allow us to order the endpoints of the intervals of integration that occur in the integrals in Appendix \[sec:moredetails\]. In particular $z\in {{\cal B}}_\omega^{ij}$ and $u_0\in {{\cal B}}^{ij}$ imply that the interval $N$ on which $z$ and $u_0$ differ is located at least a distance $\omega$ away from the endpoints of the interval $[0, 1]$, i.e. $a \geq \omega$ and $a+|N| \leq 1-\omega$. If we also take into account the conditions $\rho, \sigma \leq \frac\omega2$ we have the ordering, for $\sigma\leq\rho$, $$\begin{aligned}
&-\rho \leq -\sigma \leq 0 \leq a-\rho-\sigma \leq a-\sigma \leq a \leq a+\sigma \leq a+\rho \leq a+|N|-\rho \leq a+|N|-\sigma\\
&\hspace{0.65cm}\leq a+|N| \leq a+|N|+\sigma \leq a+|N|+\rho \leq 1-\rho \leq 1-\sigma \leq 1 \leq 1+\sigma \leq 1+\rho\end{aligned}$$ and an analogous one for $\rho\leq \sigma$. These orderings are important when determining exactly which $I_\pm^\sigma(x, a, b)$ contribute over which $x$-intervals to integrals like $$\int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx.$$
Loosening the condition $z \in {{\cal B}}_\omega^{ij}$ to $z \in {{\cal B}}_\omega$ and consequently $u_0\in {{\cal B}}^{ij}$ to $u_0\in {{\cal B}}$ is possible in principle, but will give rise to more possible orderings of the kind above and separate calculations of all the integrals involved need to be done for each possible ordering. It is not expected however that this will influence the end result by much if at all.
Up to and including (\[eq:F4inequality\]) the steps in the proof of Lemma \[lem:F3partresult\] are independent of the specific choice of kernels $\phi_\sigma$ and $\phi_\rho$, but the calculations that make up the remainder of the proof do depend on the explicit choice $\phi_\sigma=\hat\phi_\sigma$. In order to derive similar results for other kernels we need to redo those computations with an explicitly given alternative choice.
The result for $F_2$ in Corollary \[thm:F3\], part \[item:F2zminoverB\] follows as a direct consequence of Theorem \[thm:F4\], part \[item:F3zminoverB\] for $F_3$ by choosing $\rho=\sigma$. However, the fact that the fidelity term in $F_2$ vanishes if and only if $u=z$ allows for a direct proof as well.\
[**Proof of Corollary \[thm:F3\], part \[item:F2zminoverB\]**]{}: Since for $F_2$, the fidelity term vanishes at $u = z$ any potential competitor must have strictly less interfaces than $z$. The result follows then immediately from Lemma \[lem:F3partresult\] with $\rho=\sigma$ and Lemma \[lemma0\].
To complete the proof of Theorem \[thm:F4\] we need a result that tells us that, if $\sigma \leq \rho$, a minimiser of $F_3$ is either equal to $z$ or has strictly less interfaces. Lemma \[lem:F3firstvar\] will provide exactly this. First we need some preparatory lemmas.
\[lem:levelhalfrhosigmaz\] Let $z\in {{\cal B}}_\omega$, $\rho, \sigma \leq \frac\omega2$, and $\phi_\sigma \in \mathcal{K}\cap C({{{\ensuremath{\mathbb{R}}}}})$, then the level-$\frac12$ set of $\phi_\rho*f_\sigma$ consists of exactly the locations of the interfaces of $z$. Furthermore the upper level-$\frac12$ set where $\phi_\rho*\phi_\sigma*z \geq \frac12$ is $\operatorname{supp}z$.
If $z=0$ the results follow trivially. We assume now $z\neq 0$. First we consider the case of a bar code with only one bar. Let $a<b$ be such that $b-a \geq \omega$ and define $z:= \chi_{[a, b]}$. Since the convolution of two symmetric unimodal functions is again a symmetric unimodal function (see [@Uhrin84; @EatonPerlman91] and references therein) we find that $\phi_\sigma*z$ is a unimodal function with mode at $x_0:=\frac{a+b}2$, i.e. $\phi_\sigma*z$ is non-decreasing for $x\geq x_0$ and non-increasing for $x\leq x_0$, and symmetric around $x=x_0$. Since $\phi_\rho$ is unimodal with mode at $x=0$ and symmetric around $x=0$ we conclude that $\phi_\rho*\phi_\sigma*z$ is unimodal with mode at $x=x_0$ and symmetric around $x=x_0$. Therefore for all $x \leq a$ and all $x\geq b$ $$\label{eq:nonstrictineq1}
\phi_\rho*\phi_\sigma*z(x) \leq \phi_\rho*\phi_\sigma*z(a)=\phi_\rho*\phi_\sigma*z(b)$$ and for all $x\in [a,b]$ $$\label{eq:nonstrictineq2}
\phi_\rho*\phi_\sigma*z(x) \geq \phi_\rho*\phi_\sigma*z(a)=\phi_\rho*\phi_\sigma*z(b).$$ In the sense of distributions we have $$z' = \delta_a-\delta_b$$ where $\delta_x$ is the Dirac delta measure at $x$. Hence $$\phi_\rho*\phi_\sigma*z'(x) = \phi_\rho*\phi_\sigma(x-a) - \phi_\rho*\phi_\sigma(x-b).$$ Because $\phi_\rho*\phi_\sigma$ is unimodal with maximum at $0$ we deduce that $(\phi_\rho * \phi_\sigma * z)'(a)>0$ and $(\phi_\rho * \phi_\sigma * z)'(b)<0$. Combined with (\[eq:nonstrictineq1\]) and (\[eq:nonstrictineq2\]) this implies that for all $x < a$ and all $x > b$ $$\phi_\rho*\phi_\sigma*z(x) < \phi_\rho*\phi_\sigma*z(a)=\phi_\rho*\phi_\sigma*z(b)$$ and for all $x\in (a,b)$ $$\phi_\rho*\phi_\sigma*z(x) > \phi_\rho*\phi_\sigma*z(a)=\phi_\rho*\phi_\sigma*z(b).$$ We now explicitly compute the value $\phi_\rho*\phi_\sigma*z(a)$. $$\begin{aligned}
\phi_\rho*\phi_\sigma*z(a) &= \int_{{{\ensuremath{\mathbb{R}}}}}\int_{{{\ensuremath{\mathbb{R}}}}}\phi_\rho(a-x) \phi_\sigma(x-y) \chi_{[a,b]}(y)\, dy\, dx\notag\\
&= \int_{a-\rho}^{a+\rho} \int_a^{x+\sigma} \phi_\rho(a-x) \phi_\sigma(x-y)\, dy\, dx\notag = \int_{-\rho}^\rho \int_{-\sigma}^{-z} \phi_\rho(z) \phi_\sigma(q)\, dq\, dz\notag\\
&= \int_{-\rho}^\rho \int_{-\sigma}^{0} \phi_\rho(z) \phi_\sigma(q)\, dq\, dz - \int_{-\rho}^\rho \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq\, dz\notag\\
&= \frac12 - \int_{-\rho}^\rho \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq\, dz.\label{eq:1/2}\end{aligned}$$ In the third equality we have used the change of variables $$\left(\begin{array}{c} z\\ q\end{array}\right) = \left(\begin{array}{c} a\\ 0\end{array}\right) + \left(\begin{array}{cc} -1&0\\ 1&-1\end{array}\right)\left(\begin{array}{c} x\\ y\end{array}\right).$$ The last equality follows by symmetry of $\phi_\sigma$ and the fact that $\phi_\rho$ and $\phi_\sigma$ have unit mass.
Because $$\begin{aligned}
\int_{-\rho}^0 \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq \, dz &= \int_{\rho}^0 \int_z^0 \phi_\rho(-z) \phi_\sigma(q)\, dq \, d(-z) = -\int_0^\rho \int_0^z \phi_\rho(z) \phi_\sigma(q)\, dq \, dz\\
&= -\int_0^\rho \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq \, dz\end{aligned}$$ we have $$\int_{-\rho}^\rho \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq\, dz = \int_{-\rho}^0 \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq\, dz + \int_0^\rho \int_{-z}^0 \phi_\rho(z) \phi_\sigma(q)\, dq\, dz
= 0$$ and hence by (\[eq:1/2\]) $$\label{eq:equaltoahalf}
\phi_\rho*\phi_\sigma*z(a) = \frac12.$$
This proves the result if $z$ has only one bar. If $z$ has more bars then we prove that the $\frac12$-lower level set of $\phi_\rho*f_\sigma$ is the same as the $\frac12$-lower level set of $f_\sigma$ in a similar fashion as the $\frac12$-lower level set was identified in the proof of Lemma \[lem:lessthanahalf\]. Let $x\in {{{\ensuremath{\mathbb{R}}}}}\setminus \operatorname{supp}z$, then there exist $c<d$ such that $d-c \geq \omega$, $x\in (c, d)$, and $z(y)=0$ for all $y\in (c, d)$. Define $z_0 := \chi_{(-\infty, c)} + \chi_{(d, \infty)}$, then $$\begin{aligned}
\phi_\sigma * f_\sigma(x) &\leq \phi_\rho*\phi_\sigma * z_0(x)
= \int_{{{\ensuremath{\mathbb{R}}}}}\phi_\rho*\phi_\sigma(x-y)\, dy - \int_c^d \phi_\rho*\phi_\sigma(x-y)\,dy\\
&= 1- \phi_\rho*\phi_\sigma*\chi_{[c,d]}(x)
< \frac12.\end{aligned}$$ The last inequality follows from $\phi_\rho*\phi_\sigma*\chi_{[c, d]} > \frac12$ on $(c, d)$ as proven above.
\[lem:decreasing\] Let $z\in {{\cal B}}_\omega$, $\phi_\rho\in \mathcal{K}_3$ and $\rho \leq \frac\omega2$. Fix $x\in \operatorname{supp}z$, then the function $$(0, \rho] \to {{{\ensuremath{\mathbb{R}}}}}: \tau \mapsto \phi_\tau * \phi_\rho*z(x)$$ is non-increasing.
First assume that $z =\chi_{[a, b]}$ for some $a<b$ satisfying $b-a \geq \omega$. We compute $$\frac{\partial}{\partial \tau} \phi_\tau*\phi_\rho*z(x) = \int_{{{\ensuremath{\mathbb{R}}}}}\int_a^b \phi_\rho(y-w) \frac{\partial}{\partial \tau} \phi_\tau(x-y) \, dw\, dy.$$ As in (\[eq:Ksigma\]) we write $
\phi_\tau(x) = p(-x, \tau) \chi_{[-\tau, 0]}(x) + p(x, \tau) \chi_{[0, \tau]}(x)$ and thus, by continuity of $\phi_\tau$ in $\tau$, $$\frac{\partial}{\partial \tau} \phi_\tau(x) = \chi_{[-\tau, 0]}(x) \frac{\partial}{\partial \tau} p(-x, \tau) + \chi_{[0, \tau]}(x)\frac{\partial}{\partial \tau}p(x, \tau)$$ and thus $$\begin{aligned}
\frac{\partial}{\partial \tau} \phi_\tau*\phi_\rho*z(x) &= \int_x^{x+\tau} \int_a^b \phi_\rho(y-w) \frac{\partial}{\partial \tau} p(y-x, \tau)\, dw\, dy\\ &\hspace{0.7cm}+ \int_{x-\tau}^x \int_a^b \phi_\rho(y-w) \frac{\partial}{\partial \tau} p(x-y, \tau)\, dw\, dy.\end{aligned}$$ Using the substitution of variables $\displaystyle \left(\begin{array}{c} \tilde y\\ \tilde w\end{array}\right) = \left(\begin{array}{c} -x\\ x\end{array}\right) + \left(\begin{array}{cc} 1&0\\ 0&-1\end{array}\right) \left(\begin{array}{c} y\\ w\end{array}\right)$, using the symmetry of $\phi_\sigma$, then writing $\tilde x = x-a$ and $c=b-a$ and finally dropping the tildes, allows us to rewrite the integrals above as the integral in (\[eq:condpsit\]) with $\rho$ instead of $\sigma$. We can thus conclude that $$\frac{\partial}{\partial \tau} \phi_\tau*\phi_\rho*z(x) \leq 0.$$
If $z$ has more bars we note that the double convolution of a single bar of $z$ extends a distance of $\tau+\rho \leq 2 \rho \leq \omega$ outside of the bar and thus will not influence the value of $\phi_\tau*\phi_\rho*z$ inside other bars of $z$.
\[lem:F3firstvar\] Let $z\in {{\cal B}}_{\omega}$, $\phi_\sigma \in \mathcal{K}$, and $u$ a minimiser of $F_3$ over ${{\cal B}}$. Denote by $x_i$ the locations of the interfaces of $u$, with $x_0 < x_1 < \ldots$, then we have for every $i$ $$\label{eq:F3firstvar}
\phi_\rho*f_\sigma(x_i) = \frac12 +\phi_\rho*\phi_\rho*\overline u_i(x_i),$$ where $$\overline u_i := \left\{ \begin{array}{ll} u-\chi_{[x_i, x_{i+1}]} & \text{ if } i \text{ is even, i.e. } x_i \text{ is the left interface of a bar of } u,\\ u-\chi_{[x_{i-1}, x_i]} & \text{ if } i \text{ is odd, i.e. } x_i \text{ is the right interface of a bar of } u.\end{array}\right.$$ (Note that for $i$ even, $\overline u_i=\overline u_{i+1}$.)
Consequently if $\phi_\sigma = \hat \phi_\sigma$, $\sigma \leq \rho \leq \frac{\omega}2$, and $\lambda$, $\rho$ and $\sigma$ satisfy in addition $$\label{eq:sigmarhoomegacondition}
\frac2\lambda + \frac1{15 \rho^2} \Bigl(-\sigma^3 + 5 \rho \sigma^2 + 17 \rho^3\Bigr) < \omega,$$ then for every $i$, $x_i$ is the location of an interface of $z$.
An example of a bar code $u$ and its accompanying bar codes $\overline u_0$, $\overline u_1$, and $\overline u_2$ is shown in Figure \[fig:uubar\].
Let $\phi_\sigma\in \mathcal{K}$. Let $u$ minimise $F_3$ over ${{\cal B}}$, then $F_3$ has vanishing first variation in $u$ with respect to small perturbations in the locations of the interfaces of $u$. Let $x_0$ be the location of an interface of $u$ where the value of $u$ jumps from $0$ to $1$, in other words, it is the left interface of a bar. The argument is analogous for a right interface. We consider a perturbed $u(t):= \overline u_0 + \chi_{[x_0+t, x_1]} \in {{\cal B}}$, where $|t|$ is small enough such that no interfaces are created or annihilated. The number of interfaces of $u(t)$ is equal to that of $u$ and hence we compute (integration is with respect to $x$) $$\begin{aligned}
&\hspace{0.7cm}\lambda^{-1} \Big(F_3(u(t)) - F_3(u)\Big) \\
&= \int_{{{\ensuremath{\mathbb{R}}}}}\bigg[ \Big(\phi_\rho*u(t)-f_\sigma\Big)^2 - \Big(\phi_\rho*u-f_\sigma\Big)^2\bigg]\\
&= \int_{{{\ensuremath{\mathbb{R}}}}}\bigg[ \Big(\phi_\rho*u(t)\Big)^2 - \Big(\phi_\rho*u\Big)^2 + 2 f_\sigma \cdot \phi_\rho*\Big(u-u(t)\Big)\bigg]\\
&= \left\{ \begin{array}{ll}
\mathlarger\int_{{{\ensuremath{\mathbb{R}}}}}\bigg[
\Big(\phi_\rho*\chi_{[x_0, x_0+t]}\Big)^2 - 2 \phi_\rho*u \cdot \phi_\rho*\chi_{[x_0, x_0+t]}
+ 2 \phi_\rho*f_\sigma \cdot\Big(u-u(t)\Big)\bigg] & \text{if } t>0,\\
\mathlarger\int_{{{\ensuremath{\mathbb{R}}}}}\bigg[
\Big(\phi_\rho*\chi_{[x_0+t, x_0]}\Big)^2 + 2 \phi_\rho*u \cdot \phi_\rho*\chi_{[x_0+t, x_0]}
+ 2 \phi_\rho*f_\sigma \cdot\Big(u-u(t)\Big)\bigg] & \text{if } t<0,
\end{array}\right.\end{aligned}$$ where we have used that $u(t)=u-\chi_{[x_0, x_0+t]}$ if $t>0$ and $u(t)=u+\chi_{[x_0+t, x_0]}$ if $t<0$ in the last line as well as using Lemma \[eq:intconv\].
Assume for now that $t>0$. The case for $t<0$ is analogous. Then, using $u(t)=\overline u_0+\chi_{[x_0+t, x_1]}$, $$\begin{aligned}
\left.\frac{d}{dt} \lambda^{-1} \Big(F_3(u(t)) - F_3(u)\Big)\right|_{t=0^+} &= \left[2 \mathlarger\int_{{{\ensuremath{\mathbb{R}}}}}\left\{\left(\frac{d}{dt} \int_{x_0}^{x_0+t} \phi_{\rho}(x-y)\,dy\right) \cdot \int_{x_0}^{x_0+t} \phi_\rho(x-y)\, dy\right.\right.\notag\\&\hspace{0.7cm}\left.\left. - 2 \frac{d}{dt} \int_{x_0}^{x_0+t} \phi_{\rho}*\phi_{\rho}*u + 2 \frac{d}{dt} \int_{x_0}^{x_0+t} \phi_\rho*f_\sigma\right\}\,dx \right]_{t=0}\notag\\
&= -2 \phi_\rho*\phi_\rho*u(x_0) - 2 \phi_\rho*f_\sigma(x_0).\label{eq:derivativewrtt}\end{aligned}$$ We can rewrite the first terms as follows: $$\phi_\rho*\phi_\rho*u(x_0) = \phi_\rho*\phi_\rho*(\chi_{[x_0, x_1]} + \overline u_0)(x_0) = \frac12 + \phi_\rho*\phi_\rho*\overline u_0(x_0),$$ where we have used (\[eq:equaltoahalf\]) to compute $\displaystyle \phi_\rho*\phi_\rho*\chi_{[x_0, x_1]}(x_0) = \frac12$.
Vanishing of the first variation tells us that the right hand side in (\[eq:derivativewrtt\]) is zero and hence $$1 + 2 \phi_\rho*\phi_\rho*\overline u_0(x_0) - 2 \phi_\rho*f_\sigma(x_0) = 0,$$ which gives equation (\[eq:F3firstvar\]) for $x_i=x_0$.
Now assume $\sigma \leq \rho \leq \frac{\omega}2$ and $\phi_\sigma=\hat\phi_\sigma$. If $u$ is such that the white spaces between every two subsequent black bars have widths of at least $2\rho$ then it follows that $\hat\phi_\rho*\hat\phi_\rho*\overline u_0(x_i)=0$ for every $i$ and equation (\[eq:F3firstvar\]) reduces to $$\hat\phi_\rho*f_\sigma(x_i) = \frac12.$$ Lemma \[lem:levelhalfrhosigmaz\] then completes the argument. Note that in this case condition (\[eq:sigmarhoomegacondition\]) is not necessary.
Now assume that $u$ is not as above, i.e. there exist two bars in $u$ separated by a white space of width strictly less than $2\rho$. We will show a contradiction. Let $x_1$ be the right interface of a bar of $u$ and let $x_2$ be the left interface of the next bar, such that $x_2-x_1 < 2\rho \leq \omega$. Then the following inequalities should be satisfied $$\hat\phi_\rho*f_\sigma(x_i) = \frac12 + \hat\phi_\rho*\hat\phi_\rho*\overline u_i(x_i) \geq \frac12, \quad \text{for } i\in\{1, 2\}.$$ According to Lemma \[lem:levelhalfrhosigmaz\] this means that $x_1, x_2 \in \operatorname{supp}z$. Now there are two possibilities. The first is that $x_1$ and $x_2$ are located in different bars of $z$. Since $z\in {{\cal B}}_{\omega}$ this means that $x_2-x_1 \geq \omega$ which contradicts our assumption. The second possibility is that $x_1$ and $x_2$ are in the same bar of $z$. Assume the latter now.
By the same arguments the right interface of the second bar, i.e. $x_3$ also lies in $\operatorname{supp}z$. It can lie either in a different bar of $z$ than $x_2$ or in the same one. In the former case we have that there exists an interval $N$ with $|N| \geq \omega$ such that $z=0$ and $u=1$ on $N$ and using (\[eq:sigmarhoomegacondition\]) we can use the arguments as in Lemma \[lem:F3partresult\] to arrive at a contradiction with the fact that $u$ is a minimiser of $F_3$.[^4] We conclude that $x_2$ and $x_3$ must lie in the same bar of $z$. In a similar way we find that $x_0$ lies in the same bar. If $z$ has more than two bars, via induction on the interfaces we find that for every even $i$, $[x_i, x_{i+1}] \subset \operatorname{supp}z$. In words, every bar of $u$ is contained in a bar of $z$.
From the foregoing we deduce that $(u-z)(x) \in \{-1, 0\}$ a.e. and $$\label{eq:uminzestimate}
u-z \leq -\chi_{[x_1, x_2]}.$$
Define $\hat u := u + \chi_{[x_1, x_2]}$, then $$\begin{aligned}
&\hspace{0.6cm} \int_{{{\ensuremath{\mathbb{R}}}}}\bigg( \Big(\hat\phi_\rho*\hat u-\hat\phi_\sigma*z\Big)^2 - \Big(\hat\phi_\rho*u-\hat\phi_\sigma*z\Big)^2 \bigg)\notag\\
&= \int_{{{\ensuremath{\mathbb{R}}}}}\bigg( \Big(\hat\phi_\rho*\hat u\Big)^2 + 2 \hat\phi_\sigma*z \cdot \hat\phi_\rho*(u-\hat u) - \Big(\hat\phi_\rho*u\Big)^2 \bigg)\notag\\
&= \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho * \chi_{[x_1, x_2]}\Big)^2 + 2 \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho*u\cdot \hat\phi_\rho*\chi_{[x_1, x_2]} - \hat\phi_\sigma*z \cdot \hat\phi_\rho*\chi_{[x_1, x_2]}\Big)\notag\\
&= \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho * \chi_{[x_1, x_2]}\Big)^2 + 2 \int_{x_1}^{x_2} \Big(\hat\phi_\rho*\hat\phi_\rho*u - \hat\phi_\rho*\hat\phi_\sigma*z\Big),\label{eq:F3L2uhatucompare}\end{aligned}$$ where the last equality follows by Lemma \[eq:intconv\].
We now use Lemma \[lem:decreasing\] and inequality (\[eq:uminzestimate\]) to estimate $$\begin{aligned}
\int_{x_1}^{x_2} \Big(\hat\phi_\rho*\hat\phi_\rho*u - \hat\phi_\rho*\hat\phi_\sigma*z \Big) &\leq \int_{x_1}^{x_2} \hat\phi_\rho*\hat\phi_\rho*(u-z) \leq -\int_{x_1}^{x_2} \hat\phi_\rho*\hat\phi_\rho*\chi_{[x_1, x_2]}\\
&=-\int_{x_1}^{x_2} \int_{{{\ensuremath{\mathbb{R}}}}}\int_{x_1}^{x_2} \hat\phi_\rho(x-y) \hat\phi_\rho(y-q) \,dq \,dy \,dx\\
&= -\int_{{{\ensuremath{\mathbb{R}}}}}\bigg(\int_{x_1}^{x_2} \hat\phi_\rho(y-x) \,dx\bigg)^2 \,dy = -\int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho*\chi_{[x_1, x_2]}\Big)^2.\end{aligned}$$
Using this in (\[eq:F3L2uhatucompare\]) we find $$F_3(\hat u) - F_3(u) \leq -2 - \lambda \int_{{{\ensuremath{\mathbb{R}}}}}\Big(\hat\phi_\rho*\chi_{[x_1, x_2]}\Big)^2 < 0$$ which contradicts $u$ being a minimiser.
The result of Lemma \[lem:F3firstvar\] doesn’t change if $z\in B_\omega^{ij}$ and we minimise $F_3$ over $B^{ij}$ for $i, j \in \{0, 1\}$. Also note that the result can be obtained for any $\phi_\sigma \in \mathcal{K}_3$ if we replace condition (\[eq:sigmarhoomegacondition\]) by the corresponding parameter range for that choice of kernel, which we can obtain be redoing the calculations in the proof of Lemma \[lem:F3partresult\] after (\[eq:F4inequality\]) for the new kernel.
Note that in the case where $\sigma \leq \rho$ we could have used Lemma \[lem:F3firstvar\] in the proof of Lemma \[lem:F3partresult\] instead of Lemma \[lemma1\].
[**Proof of Theorem \[thm:F4\], part \[item:F3zminoverB\]**]{}: From Lemma \[lem:F3firstvar\] it follows that under the stated conditions the only possible minimisers of $F_3$ over ${{\cal B}}^{ij}$ are $u=z$ or a $u$ with strictly less interfaces than $z$. By Lemma \[lem:F3partresult\] however such a minimiser cannot have less interfaces than $z$ and hence by Lemma \[lemma0\] $u=z$ is the unique minimiser of $F_3$ over ${{\cal B}}^{ij}$.
Numerical simulations {#numerics}
=====================
We present a few test simulations for the minimisation problems $F_2$ and $F_3$. To this end, there exists an increasing number of state of the art techniques concerning TV-based minimisation. However here we are not attempting to write the most efficient algorithm, we only aim to test whether the parameter regimes we found theoretically are close to optimal or not. Hence we take the naive approach of using a phase field to approximate the total variation: That is, choosing $\epsilon$ small, we replace the total variation with $$\mathlarger\int_0^1 \left(\epsilon \, |u'|^2 \,\, + \,\, \frac{ u^2(1 -u)^2 }{2\epsilon}\,\, dx \right),$$ and consider the $L^2$ gradient descent of the resulting functional. While this technique brings in diffuse interfaces[^5] (i.e. minimisers will no longer be bar codes), it is well-justified for small $\epsilon$ (c.f. [@Braides02]) in that minimisers will be close to minimisers of the original sharp interface problem. One problem with this method in higher dimensions is that one tends to get stuck in metastable states, and hence this method would not work well for 2D bar codes. However, our 1D problem is sufficiently rigid so that the method works well and fairly quickly. It takes seconds to run our Python code, and while more direct state of the art methods would be substantially faster (as would be needed in a practical application), our limited goals are well served by the phase field approach.
For $F_3$ the $L^2$ gradient flow gives the equation $$\label{pde}
u_t \, = \, 2\epsilon u_{xx} \, - \, \frac{1}{\epsilon} W'(u) \, - \, 2 \lambda \, \phi_\rho \, * \, (\phi_\rho * u - f_\sigma ),$$ where[^6] $W(u) = \, \frac{u^2(1 -u)^2}{2}$. In all of our experiments, a bar code is generated with $X$-dimension $\omega \approx 0.0133$. Except for Figure \[plot3\] (bottom right), convolution with the hat function $\hat \phi_\sigma$ is followed by the addition of noise with amplitude $a = 0.1$[^7]. We used $\e=0.0004$ and initial data was always taken to be either $u\equiv 0$ or $u \equiv 1/2$.
The algorithm works well for $\sigma$ far beyond the regime of Theorem \[thm:F4\]. We give a few sample results. In Figure \[plot1\] (left) we see that choosing $\rho = \sigma$ (i.e. using $F_2$), one obtains good results for $\sigma$ larger than twice $\omega$. Figure \[plot1\] (right) shows that even for $\sigma \approx 3 \omega$, the results are not bad, however they begin to loose accuracy. In Figure \[plot3\] we note that choosing $\rho$ to be the $X$-dimension works well for blurring with $\sigma$ up to twice $\omega$. Note that here we are in the regime $\rho < \sigma$, which we avoided in Theorem \[thm:F4\]. In fact, our counter example suggested that in this regime an upper bound on $\lambda$ is necessary. Figure \[plot3\] (bottom left) indeed supports this observation by taking $\lambda$ much larger than in Figure \[plot3\] (top right).
[![ Here we look at minimisers of $F_2$ ($\rho = \sigma$) to find that the algorithm works for blurring far past $\omega$. In all simulations, the three rows are as follows: A bar code is generated with $X$-dimension exactly $\omega = 0.0133$; convolution $f_\sigma$ of the bar code with $\hat \phi_\sigma$ with added noise of amplitude $a$; final steady state for (\[pde\]) superimposed with the generating bar code. []{data-label="plot1"}](sim2.png "fig:"){height="2.8in"}]{}
[![Minimisers for $F_3$. Here we take the deconvolution kernel size to be the $X$-dimension. Top row: we convolute the data with $X$-dimension exactly $\omega$ for two choices of $\sigma$. The algorithm works well with $\lambda = 1000$. However, as noted in the bottom left, for larger values of $\lambda$ it loses information. Bottom right: a bar code is convoluted with a Gaussian with standard deviation $\sigma$ but deconvoluted with $\hat\phi_\sigma$. []{data-label="plot3"}](sim4.png "fig:"){height="2.8in"}]{}
[![Minimisers for $F_3$. Here we take the deconvolution kernel size to be the $X$-dimension. Top row: we convolute the data with $X$-dimension exactly $\omega$ for two choices of $\sigma$. The algorithm works well with $\lambda = 1000$. However, as noted in the bottom left, for larger values of $\lambda$ it loses information. Bottom right: a bar code is convoluted with a Gaussian with standard deviation $\sigma$ but deconvoluted with $\hat\phi_\sigma$. []{data-label="plot3"}](sim3.png "fig:"){height="2.8in"}]{}
We also performed tests where we [*convolute/blur*]{} the bar code with a Gaussian kernel with standard deviation $\sigma$ but [*deconvolute/deblur*]{} with the hat function $\hat \phi_\rho$. To obtain satisfactory results, one must choose $\sigma$ and $\rho$ very close to each other and no larger than $\omega$, and use a suitably tuned midrange $\lambda$. We give one example in Figure \[plot3\] (bottom right).
Simulations were also performed for $F_1$ (no deconvolution/deblurring kernel) but we always found that using $F_2$ or $F_3$ (with even a small deblurring kernel) was preferable.
Discussion {#disc}
==========
We have presented results on the accuracy of TV-based energy minimisation methods for bar code deblurring in certain parameter regimes. Numerical simulations, which included the effects of noise, show that these methods are valid in much larger regimes and in particular, allow for significantly more blurring. While our analytical results did not showcase the benefits of using a deconvolution/deblurring kernel (i.e. the merits of $F_2, F_3$ versus $F_1$), numerical experiments showed clearly that the presence of a deblurring kernel in $F_2$ or $F_3$ always gave better results over no deblurring ($F_1$).
In practice, the size of the blurring kernel pertains to the so-called [*spot diameter*]{} of the laser beam at impact with the bar code. This [*spot diameter*]{} is a function of the laser beam and the distance from the scanner to the bar code. According to Palmer [@Palmer07] (p.127), most scanners can successfully read a bar code if the spot diameter is no greater than $\sqrt{2}$ times the $X-$dimension (i.e. for $2 \sigma < \sqrt{2} \omega$). This suggests that the even the conditions we have imposed on $\sigma$ in our results are not completely unreasonable. However, as suggested by the numerics, one might be able to prove results for $\sigma$ past the $X-$dimension.
Realistically neither the spot diameter (size of the blurring kernel) nor the distribution of the beam intensity (shape of the blurring kernel) is exactly known, and inferring this information from signals is an ill-posed problem. In [@Esedoglu04], the author considers a Gaussian *ansatz* for all kernels but introduces a novel optimization scheme for determining the standard deviation of the blurring kernel. In terms of the shape of the kernel, our last simulation in Figure \[plot3\] (bottom right) is suggestive. We note that if the convolution in the measured signal is done with an infinitely supported Gaussian with standard deviation $\sigma$, then deconvolution with a hat function of approximate size $\sigma$ works reasonably well. Thus if one could determine certain statistics of the blurring kernel, one could then deconvolute with a set kernel possessing similar statistics. Determining such statistics should in principle be possible as some standard bar code symbologies have a fixed structure at their left and right boundaries (the [*left and right guards*]{}, c.f. [@Palmer07]).
[**Acknowledgments:**]{} This work was completed while both authors were at Simon Fraser University. We are grateful to Fadil Santosa for bringing this problem to our attention and for many interesting conversations. We also thank Selim Esedoglu for useful discussions and for the use of his original code which was the basis for our numerical experiments. This code was modified and tested in *Python* with the *NumPy* package by Simon Fraser undergraduate student Jacob Groundwater, who we would also like to thank. This research was partially supported by an NSERC (Canada) Discovery Grant. YvG was also supported by a PIMS postdoctoral fellowship.
Calculations in the proof of Lemma \[lem:F3partresult\] {#sec:moredetails}
=======================================================
In this appendix we collect some of the longer calculations in the proof of Lemma \[lem:F3partresult\]. We start with a lemma.
\[lem:fsigmasquared\] Let $z:= \chi_{[a, b]}$ for some $a < b$, $\sigma \leq \frac{b-a}2$, and $f_\sigma = \hat\phi_\sigma*z$, then $$\int_{{{\ensuremath{\mathbb{R}}}}}f_\sigma^2 = b-a-\frac7{15} \sigma.$$
We compute $$\begin{aligned}
\label{eq:fsigmaexplicit1}
f_\sigma(x) &=
\left\{ \begin{array}{ll}
0 & \text{if } x \in (-\infty, a-\sigma],\\
I_+^\sigma(x, a, x+\sigma) & \text{if } x\in [a-\sigma, a],\\
I_-^\sigma(x, a, x) + I_+^\sigma(x, x, x+\sigma) & \text{if } x\in [a, a+\sigma],\\
I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) & \text{if } x\in[a+\sigma, b-\sigma],\\
I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, b) & \text{if } x\in[b-\sigma, b],\\
I_-^\sigma(x, x-\sigma, b) &\text{if } x\in[b, b+\sigma],\\
0 & \text{if } x\in[b+\sigma, \infty).
\end{array}\right. \notag\\
&= \left\{ \begin{array}{ll}
0 & \text{if } x \in (-\infty, a-\sigma],\\
\frac1{2\sigma^2} (x+\sigma-a)^2 & \text{if } x\in [a-\sigma, a],\\
-\frac1{2\sigma^2} \Bigl(x-a-\bigl(1+\sqrt{2}\bigr) \sigma\Bigr) \Bigl(x-a-\bigl(1-\sqrt{2}\bigr) \sigma\Bigr) & \text{if } x\in [a, a+\sigma],\\
1 & \text{if } x\in[a+\sigma, b-\sigma],\\
-\frac1{2\sigma^2} \Bigl(x-b+\bigl(1-\sqrt{2}\bigr)\sigma\Bigr) \Bigl(x-b+\bigl(1+\sqrt{2}\bigr)\sigma\Bigr) & \text{if } x\in[b-\sigma, b],\\
\frac1{2\sigma^2} (x-\sigma-b)^2 &\text{if } x\in[b, b+\sigma],\\
0 & \text{if } x\in[b+\sigma, \infty).
\end{array}\right.\end{aligned}$$ An explicit computation of the integral we are interested in leads to the result.
Next we give the calculations for the different cases described in the proof of Lemma \[lem:F3partresult\].
**Case Ia:** $$\begin{aligned}
&\hspace{0.4cm}-2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx\notag\\
&=-2\Biggl\{ \int_{a-\rho}^{a-\sigma} I_+^\rho(x, a, x+\rho) \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\notag\\
&\hspace{1.6cm} +\int_{a-\sigma}^a I_+^\rho(x, a, x+\rho) \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, a)\biggr)\,dx\notag\\
&\hspace{1.6cm} +\int_a^{a+\sigma} \biggl(I_-^\rho(x, a, x) + I_+^\rho(x, x, x+\rho)\biggr) I_-^\sigma(x, x-\sigma, a) \,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|-\sigma}^{a+|N|} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, a+|N|) \biggr) I_+^\sigma(x, a+|N|, x+\sigma) \,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|}^{a+|N|+\sigma} I_-(x, x-\rho, a+|N|) \biggl(I_-^\sigma(x, a+|N|, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|+\sigma}^{a+|N|+\rho} I_-^\rho(x, x-\rho, a+|N|) \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\Biggr\}\notag\\
&= \frac1{15 \rho^2} \Bigl(\sigma^3 - 5 \rho \sigma^2 - 10 \rho^3\Bigr).\label{eq:lemCaseIa}\end{aligned}$$
The way to find the specific intervals of integration in the integrals above (and those that follow below) is similar in spirit to what was done in the proof of Theorem \[thm:minimofF2\]. We do not give all the details here, but it is important to reflect on the role of the conditions $z\in {{\cal B}}_\omega^{ij}$, $u_0 \in {{\cal B}}^{ij}$ (instead of $z\in {{\cal B}}_\omega$ and $u\in{{\cal B}}$) and $\rho, \sigma\leq \frac\omega2$. Such considerations are addressed in Remark \[rem:roleofconditions\].
**Case Ib:** $$\begin{aligned}
&\hspace{0.4cm}-2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{N^c\cap[0,1]} \hat\phi_\sigma(x-w)\,dw\,dx\notag\\
&=-2 \Biggl\{ \int_{a-\rho}^a I_+^\rho(x, a, x+\rho) \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, a) \biggr)\,dx\notag\\
&\hspace{1.6cm} +\int_a^{a+\rho} \biggl(I_-^\rho(x, a, x) + I_+^\rho(x, x, x+\rho) \biggr) I_-^\sigma(x, x-\sigma, a)\,dx\notag\\
&\hspace{1.6cm} +\int_{a+\rho}^{a+\sigma} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho) \biggr) I_-^\sigma(x, x-\sigma, a) \,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|-\sigma}^{a+|N|-\rho} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho) \biggr) I_+^\sigma(x, a+|N|, x+\sigma) \,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|-\rho}^{a+|N|} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, a+|N|) \biggr) I_+^\sigma(x, a+|N|, x+\sigma) \,dx\notag\\
&\hspace{1.6cm} +\int_{a+|N|}^{a+|N|+\rho} I_-^\rho(x, x-\rho, a+|N|) \biggl(I_-^\sigma(x, a+|N|, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\Biggr\}\notag\\
&= \frac1{15\sigma^2} \Bigl(\rho^3 - 5 \sigma \rho^2 - 10 \sigma^3\Bigr).\label{eq:lemCaseIb}\end{aligned}$$
**Case II, first term:** $$\begin{aligned}
&-2\int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y)\,dy \int_{{{\ensuremath{\mathbb{R}}}}}\hat\phi_\rho(u-w)\,dw\,dx\notag\\
&=-2 \Biggl\{ \int_{a-\rho}^a I_+^\rho(x, a, x+\rho) \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho)\biggr)\,dx \notag\\
& +\int_a^{a+\rho} \biggl(I_-^\rho(x, a, x) + I_+^\rho(x, x, x+\rho)\,dy\biggr)
\cdot
\biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho)\biggr)\,dx \notag\\
& +\int_{a+\rho}^{a+|N|-\rho} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho)\biggr)
\cdot
\biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho) \biggr)\,dx \notag\\
& +\int_{a+|N|-\rho}^{a+|N|} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, a+|N|)\biggr)
\cdot
\biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho)\biggr) \,dx \notag\\
& +\int_{a+|N|}^{a+|N|+\rho} I_-^\rho(x, x-\rho, a+|N|) \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho) \biggr)\,dx\Biggr\} \notag\\
&=-2N.\label{eq:lemCaseIIfirstterm}\end{aligned}$$
**Case II, second term, IIa:** $$\begin{aligned}
& 2 \int_{{{\ensuremath{\mathbb{R}}}}}\int_N \hat\phi_\rho(x-y) \,dy \int_N \hat\phi_\sigma(x-w)\,dw\,dx\notag\\
&= 2 \Biggl\{ \int_{a-\sigma}^a I_+^\rho(x, a, x+\rho) I_+^\sigma(x, a, x+\sigma)\,dx\notag\\
& +\int_a^{a+\sigma} \biggl(I_-^\rho(x, a, x) + I_+^\rho(x, x, x+\rho)\biggr)
\cdot \biggl(I_-^\sigma(x, a, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\notag\\
&+\int_{a+\sigma}^{a+\rho} \biggl(I_-^\rho(x, a, x) + I_+^\rho(x, x, x+\rho) \biggr)
\cdot \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\notag\\
& +\int_{a+\rho}^{a+|N|-\rho} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, x+\rho)\biggr)
\cdot \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) \biggr) \,dx\notag\\
& +\int_{a+|N|-\rho}^{a+|N|-\sigma} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, a+|N|)\biggr)
\cdot \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, x+\sigma) \biggr)\,dx\notag\\
& +\int_{a+|N|-\sigma}^{a+|N|} \biggl(I_-^\rho(x, x-\rho, x) + I_+^\rho(x, x, a+|N|) \biggr)
\cdot \biggl(I_-^\sigma(x, x-\sigma, x) + I_+^\sigma(x, x, a+|N|)\biggr)\,dx\notag\\
& +\int_{a+|N|}^{a+|N|+\sigma} I_-^\rho(x, x-\rho, a+|N|) I_-^\sigma(x, x-\sigma, a+|N|) \,dx\notag\\ \notag\\
&= \,\,\, 2N + \frac1{15 \rho^2} \Bigl(\sigma^3 - 5 \rho \sigma^2 - 10 \rho^3\Bigr).\label{eq:lemCaseIIa}\end{aligned}$$
Proof that $\hat\phi_\sigma$ satisfies condition (\[eq:condpsit\]) {#app:psitaufrhononpos}
==================================================================
In this Appendix we prove that the hat function $\hat\phi_\sigma$ satisfies condition (\[eq:condpsit\]). We do this by proving a more general result first and then showing that this holds for the hat function in particular.
We use the notation as in (\[eq:Ksigma\]) and introduce
\[lem:B1\] Use the notation as in (\[eq:Ksigma\]). If for each $\tau \in (0, \sigma]$
1. \[item:B1case1\] either $\displaystyle \frac{\partial}{\partial \tau} p(x, \tau)$ is monotonically increasing in $x$ and $\mathcal{J}(\sigma, \tau, 0, c) \leq 0$ for all $c\geq 2\sigma$,
2. \[item:B1case2\] or $\displaystyle \frac{\partial}{\partial \tau} p(x, \tau)$ is monotonically decreasing in $x$ and $\mathcal{J}(\sigma, \tau, \frac{c}2, c) \leq 0$ for all $c\geq 2\sigma$,
then $\mathcal{J}(\sigma, \tau, x, c) \leq 0$ for all $\tau \in (0, \sigma]$, for all $c \geq 2 \sigma$ and all $x\in [0, c]$, i.e condition (\[eq:condpsit\]) holds.
Let $\sigma>0$, $c\geq 2\sigma$, and $\tau\in (0, \sigma]$.
Define $f_\sigma:=\phi_\sigma*\chi_{[0,c]}$ and $$\psi_\tau(x):= \left\{ \begin{array}{ll} \frac{\partial}{\partial \tau} p(-x, \tau)& \text{ if } -\tau\leq x \leq 0,\\ \frac{\partial}{\partial \tau} p(x, \tau) & \text{ if } 0 \leq x \leq \tau, \\ 0 & \text{ otherwise}. \end{array}\right.$$ This allows us to rewrite $$\mathcal{J}(\sigma, \tau, x, c) = \psi_\tau*f_\sigma(x).$$
We first consider case \[item:B1case1\]. Since $\displaystyle \frac{\partial}{\partial \tau} p(\cdot, \tau)$ is monotonically increasing, the function $-\psi_\tau$ is symmetric unimodal. Since the convolution of two symmetric unimodal functions is again a symmetric unimodal function (see [@Uhrin84; @EatonPerlman91] and references therein) we find that $-\psi_\tau*f_\sigma$ is a unimodal function with mode at $\frac{c}2$. Hence $\psi_\tau*f_\sigma(0) = \psi_\tau*f_\sigma(c)$ and $\psi_\tau*f_\sigma(x) \leq \psi_\tau*f_\sigma(0) \leq 0$ for all $x\in[0, c]$.
Next we consider case \[item:B1case2\]. In this case $\psi_\tau$ is symmetric unimodal and hence $\psi_\tau*f_\sigma$ is unimodal with mode at $\frac{c}2$ and hence for all $x\in [0, c]$ we have $\psi_\tau*f_\sigma(x) \leq \psi_\tau*f_\sigma(c/2) \leq 0$.
We complete the proof that $\hat \phi_\sigma$ satisfies condition (\[eq:condpsit\]) by showing that $\hat\phi_\sigma$ satisfies condition \[item:B1case1\] in Lemma \[lem:B1\]. Let $\sigma$, $\tau$, $c$ and $x$ satisfy the conditions in (\[eq:condpsit\]). For the hat function we have $\displaystyle p(x, \tau) = \frac1\tau \left(1-\frac{x}\tau\right)$ and hence $\displaystyle \frac{\partial}{\partial \tau} p(x, \tau) = \frac1{\tau^2}\left(-1+\frac{2x}\tau\right)$ is monotonically increasing in $x$. Furthermore by (\[eq:fsigmaexplicit1\]) —with $a=0$, $b=c$— we find that for $y\in[0, \tau]$ we have $f_\sigma(y)=1-f_\sigma(-y)$. We then compute $$\begin{aligned}
\mathcal{J}(\sigma, \tau, 0, c) &= \psi_\tau*f_\sigma(0) = \frac1{\tau^2} \int_0^\tau \left(-1+\frac{2y}\tau\right) \big(f_\sigma(-y)+1-f_\sigma(-y)\big)\,dy\\
&= \frac1{\tau^2} \int_0^\tau \left(-1+\frac{2y}\tau\right)\,dy = 0.\end{aligned}$$
[^1]: Department of Mathematics and Statistics, McGill University, Burnside Hall, 805 Sherbrooke Street West, Montreal, Quebec, H3A 2K6, Canada, [email protected]
[^2]: Department of Mathematics, University of California Los Angeles, 520 Portola Plaza, Math Sciences Building 6363, Los Angeles, California, 90095, USA, [email protected]
[^3]: For many of the similar calculations in the rest of this paper, $x\in {{{\ensuremath{\mathbb{R}}}}}$ and this kind of simplification will not be possible.
[^4]: We don’t need $u\in{{\cal B}}^{ij}$ here, because we know that in this construction $N$ has a distance of at least $\omega$ to $x=0$ and to $x=1$.
[^5]: For actual implementation, one would need to threshold the output of the minimisation process in order to generate a bar code
[^6]: This choice of constant prefactor in $W$ does not lead to unit surface tension in the sharp interface limit, hence $\lambda$ in the simulations differs by an $\mathcal{O}(1)$ multiplicative factor from the $\lambda$ in the analytical results in this paper.
[^7]: The added noise was determined as follows: The 400 grid points making up each interval of length $\omega$ were divided into 16 equal groups, each of which was assigned a random number between $-a$ and $a$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We describe the properties of the blazars detected by EGRET and summarize the results on the calculations of the evolution and luminosity function of these sources. Of the large number of possible origins of extragalactic diffuse $\gamma$-ray emission, it has been postulated that active galaxies might be one of the most likely candidates. However, some of our recent analysis indicate that only 25% of the diffuse extragalactic emission measured by SAS-2 and EGRET can be attributed to unresolved $\gamma$-ray blazars. Therefore, other sources of diffuse extragalactic $\gamma$-ray emission must exist. We present a summary of these results in this article.'
address:
- 'Barnard College & Columbia University, Dept. of Physics & Astronomy, New York, NY 10027'
- 'JILA, University of Colorado, Boulder CO 80309-0440'
author:
- 'R. Mukherjee'
- '& J. Chiang'
title: 'EGRET Gamma-Ray Blazars: Luminosity Function and Contribution to the Extragalactic Gamma-Ray Background'
---
galaxies: active - galaxies: luminosity function - gamma rays: observations - quasars: general
Introduction
============
EGRET has detected a total of 66 active galactic nuclei (AGN) in high energy ($> 100$ MeV) gamma rays since the launch of CGRO in April 1991 (Hartman et al. 1999). These sources all appear to be members of the blazar class of AGN (BL Lac objects, highly polarized ($ >$ 3%) quasars (HPQ), and optically violently variable (OVV) quasars) and are radio-loud sources with flat-spectrum at radio bands. Many of the blazars exhibit variability in their $\gamma$-ray flux on timescales of several days to months (McLaughlin et al. 1996, Mukherjee et al. 1997). The photon spectra of the blazars in the energy range 30 MeV to 30 GeV are generally well represented by power laws in energy with photon spectral indices in the range 1.4 to 3.0. The sources have non-thermal continuum spectra with the $\gamma$-ray luminosity exceeding those at other frequencies in most cases. The high $\gamma$-ray luminosities of the blazars suggest that the emission is likely to be beamed and, therefore, Doppler-boosted along the line of sight. The spectral energy distribution of blazars can be modeled as follows: the radio to UV emission can be explained as synchrotron emission from relativistic electrons in a uniform relativistically moving plasma. The high energy emission is due to the inverse Compton scattering of seed photons off the relativistic electrons, although the source of the soft photons still remains unresolved (see Hartman et al. 1997 for a review).
In this article we summarize the luminosity function and evolution properties of the EGRET blazars and use the results to examine the contribution of the $\gamma$-ray-loud AGN to the diffuse extragalactic background.
Luminosity function of EGRET blazars
====================================
The evolution and luminosity function of the EGRET blazars was calculated by Chiang & Mukherjee (1998) using data from the Phase 1 through Cycle 4 CGRO observations. Inclusion in the 1 Jy catalog of Kühr et al. (1981) of the EGRET blazars was used to account for possible biases introduced by missing optical identifications. A $V/V_{\rm max}$ test was used to show evidence of evolution. Here $V$ is the minimum volume that contains an object with redshift $z$; $V_{\rm max}$ is the largest volume that could contain an object with the same luminosity, and still be detected at the given flux limit. For a limiting significance of detection of $4\sigma$, a value of $\langle V/V_{\rm max}\rangle = 0.7$ was obtained, which means that we are preferentially detecting more sources at larger redshifts. No evidence of a density evolution of EGRET blazars was found. The evolution is consistent with pure luminosity evolution. (That is, the luminosity of the object is changing with time (i.e. redshift), while the co-moving number density remains the same).
The luminosity of a given object as a function of redshift $z$ can be described by $L(z) = L_0 f(z)$ where $L_0=L(z=0)$. Chiang & Mukherjee (1998) have discussed several different forms for the luminosity evolution function, including the power-law and exponential forms.
The redshift distribution of EGRET blazars was used to characterize the low end of the luminosity function better. The high end of the luminosity function was fixed by the non-parametric estimate mentioned above. The redshift distribution of the EGRET data was used to fit both the break luminosity and power law index of the low end of the luminosity function. A likelihood function of the redshift distribution was constructed. The probability density for the redshift of a given blazar was computed and normalized assuming the flux limit derived for that blazar. The data were best fit with a single power law at high luminosities and a luminosity cutoff of $1.1\times 10^{46}$ ergs s$^{-1}$.
Using the lower limit of the de-evolved luminosity function, the $\gamma$-ray loud AGN contribution to the extragalactic $\gamma$-ray flux is estimated to be $4.0^{+1.0}_{-0.9}\times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$ sr$^{-1}$. The sky-averaged flux contribution of identified EGRET blazars is $\simeq 1\times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$ sr$^{-1}$. Contribution to the diffuse background by unresolved blazars, therefore, is $\sim 3.0^{+1.0}_{-0.9}\times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$ sr$^{-1}$. The extragalactic diffuse flux for $E>100$ MeV estimated by Sreekumar et al. (1998) is $\simeq 1.36\times 10^{-5}$ photons cm$^{-2}$ s$^{-1}$ sr$^{-1}$. We therefore find that blazars cannot account for all of the diffuse extragalactic $\gamma$-ray background at the energies considered.
Summary
=======
The luminosity function and evolution properties of $\gamma$-ray-loud blazars imply that only $\sim$ 25% of the diffuse extragalactic emission measured by SAS-2 and EGRET can be attributed to unresolved $\gamma$-ray blazars. This is contrary to other estimates which assume a linear correlation between the measured radio and $\gamma$-ray fluxes (e.g., Stecker & Salamon 1996). However, we note that our result is consistent with recent work by Mücke & Pohl (1998) where the extragalactic diffuse contribution from blazars is synthesized using a specific blazar emission model (Dermer & Schlickeiser 1993) and an extrapolation of the observed log$N$–log$S$ distribution of EGRET blazars. As in our study, Mücke & Pohl make no assumptions regarding supposed correlations between the $\gamma$-ray fluxes with any other spectral band.
Our results lead to the exciting conclusion that other sources of diffuse extragalactic $\gamma$-ray emission must exist. The spectrum of the measured extragalactic emission implies that the average quiescent energy spectra of these sources extend to at least 50 GeV and maybe up to 100 GeV, without a significant change in the slope. If gamma-ray blazars continue to make a significant contribution to the diffuse emission at these energies, then the spectra of the parent relativistic particles in blazars which produce the gamma-rays must also remain hard to even higher energies.
[999]{} Caditz, D., & Petrosian, V., [*ApJ*]{} [**441**]{} (1993) 450. Chiang, J., & Mukherjee, R., [*ApJ*]{} [**496**]{} (1998) 752. Dermer, C. D., & Schlickeiser, R., [*A&A*]{} [**416**]{} (1993) 458 Hartman, R. C., et al., [*ApJ*]{} (1999) in press. Hartman, R. C., et al., [*Proc. of the Fourth Compton Symposium, Part 1*]{} eds. C. D. Dermer, M. S. Strickman, & J. D. Kurfess [**CP410**]{} (1997) 307. Kühr, H., et al., [*A&AS*]{} [**45**]{} (1981) 367. McLaughlin, M. A., et al., [*ApJ*]{} [**473**]{} (1996) 763. Mücke, A., & Pohl, M., [*Proc. of the Fourth Compton Symposium, Part 2*]{} eds. C. D. Dermer, M. S. Strickman, & J. D. Kurfess [**CP410**]{} (1997) 1233. Mukherjee, R., et al., [*ApJ*]{} [**490**]{} (1997) 116. Sreekumar, P., et al., [*ApJ*]{} [**494**]{} (1998) 523. Stecker, F. W., & Salamon, M. H., [*ApJ*]{} [**464**]{} (1996) 600.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Several ideas for solving the problem of fermion mass hierarchy and mixing and specific supersymmetric models that realize it are reviewed. In particular, we discuss many models based on $SO(10)$ in four dimensions combined with a family symmetry to accommodate fermion mass hierarchy and mixing, including the case of neutrinos. These models are compared and various tests that can be used to distinguish these models are suggested. We also include a discussion of a few $SO(10)$ models in higher space-time dimensions.'
address:
- |
High Energy Theory Group, Department of Physics\
Brookhaven National Laboratory, Upton, NY 11973-5000, U.S.A.\
[email protected]
- |
Department of Physics, University of Colorado\
Boulder, CO 80309-0390, U.S.A.\
[email protected]
author:
- 'Mu-Chun Chen'
- 'K.T. Mahanthappa'
title: 'Fermion Masses and Mixing and CP-Violation in $SO(10)$ Models with Family Symmetries'
---
Introduction {#intro}
============
The origin of fermion masses and mixing and CP violation is the least understood aspect of the Standard Model (SM) of particle physics. In the SM, the fermion masses and mixing angles are completely arbitrary. In order to accommodate their diverse values (see Table \[exp\]),[@Hagiwara:fs; @Hocker:2001xe; @Fusaoka:1998vc] the Yukawa couplings must range over five orders of magnitude.
Another relevant aspect is that neutrinos are massless in the framework of SM, but recent experiments strongly indicate that they do have small but non-vanishing masses. Incorporating these into theory leads to an increase in the number of parameters. The [*flavor problem*]{} thus consists the following aspects: how to reduce the number of parameters in the Yukawa sector, how to obtain an explanation of the mass hierarchy, how to obtain small neutrino masses and large leptonic mixing angles.
As we extend the SM to the Minimal Supersymmetric Standard Model (MSSM), the particle spectrum is doubled and many more parameters are introduced into the model. One thus expects large flavor changing neutral current (FCNC) due to the presence of squarks. The strongest constraints come from the lighter two generations due to the type of diagram given in Fig.\[sflavor\]:
$$\begin{aligned}
\frac{m_{\tilde{d}}^{2}-m_{\tilde{s}}^{2}}{m_{\tilde{d}}^{2}}
& \lesssim & 6 \times 10^{-3}
(\frac{m_{\tilde{s}}}{TeV})
(\sin\tilde{\theta})^{2}
\\
\frac{m_{\tilde{e}}^{2}-m_{\tilde{\mu}}^{2}}{m_{\tilde{e}}^{2}}
& < & 10^{-1}
(\frac{m_{\tilde{e}}}{TeV})
(\sin\tilde{\theta^{'}})^{2}\end{aligned}$$
where $\tilde{\theta}$ and $\tilde{\theta^{'}}$ refer to the mixing angles in the squark and slepton sectors in the bases where the mass matrices of ($d$ $s$) and ($e$ $\mu$) are diagonal.[@Gabbiani:1996hi; @Rosner:fg] Thus, in order to suppress the supersymmetric (SUSY) FCNC, a near degeneracy between the first and the second generations of squarks and sleptons is required. This is sometimes called [*SUSY flavor problem*]{}.
There have been many SUSY models proposed to accommodate the observed masses and mixing angles. These models can be classified according to the family symmetry implemented in the model. We also discuss other mechanisms that have been proposed to solve the problem of the fermion mass hierarchy and mixing.
In the following two sections, we introduce various tools that have been used to solve the flavor problem in quark sector and lepton sector, respectively; in Sec. \[so10\], $SO(10)$ is reviewed. This is then followed by a review in Sec. \[modelt\] and Sec. \[models\] on various models based on $SO(10)$ in $4D$, combined with mass texture ansatz and family symmetry, respectively; a comparison of these models is given at the end of Sec. \[models\]; Sec. \[modeled\] is devoted to models based on $SO(10)$ in higher space-time dimensions. Sec. \[conclude\] concludes this review.
Quark Masses and Mixing {#qmass}
=======================
Because the strong interaction eigenstates (same as the mass eigenstates) and the weak interaction eigenstates do not match, flavor mixing arises. However, currently we do not have any fundamental understanding for such a mismatch between the strong eigenstates and the weak eigenstates. In the weak basis, the quark mass terms are $$\mathcal{L}_{mass} =
-Y_{u} \overline{U}_{R} Q_{L} H_{u} - Y_{d} \overline{D}_{R} Q_{L} H_{d} +
h.c.$$ where $Q$ stands for the $SU(2)$ doublet quark; $U$ and $D$ are up- and down-type fermion $SU(2)$ singlet; $H_{u}$ and $H_{d}$ are Higgs fields giving masses to up- and down-type quarks. The charged current interaction is given by $$\mathcal{L}_{cc} = \frac{g}{\sqrt{2}} (W_{\mu}^{+}
\overline{U}_{L} \gamma_{\mu} D_{L}) + h.c..$$ The Yukawa couplings $Y_{i}, \; (i=u,d,e,\nu_{LR})$ are in general non-diagonal. They are diagonalized by the bi-unitary transformations $$\begin{aligned}
Y_{u}^{diag} = V_{u_{R}} Y_{u} V_{u_{L}}^{\dagger}
= diag(y_{u},y_{c},y_{t})
\\
Y_{d}^{diag} = V_{d_{R}} Y_{d} V_{d_{L}}^{\dagger}
= diag(y_{d},y_{s},y_{b})\end{aligned}$$ where $V_{R}$ and $V_{L}$ are the right-handed and left-handed rotations respectively, and all the eigenvalues $y_{i}$’s are real and non-negative, and are obtained by diagonalizing the hermitian quantity $Y^{\dagger}Y$ and $Y Y^{\dagger}$. The Cabbibo-Kobayashi-Maskawa (CKM) matrix is then given by[@Kobayashi:fv] $$V_{CKM} = V_{u_{L}} V_{d_{L}}^{\dagger}
= \left(
\begin{array}{ccc}
V_{ud} & V_{us} & V_{ub}\\
V_{cd} & V_{cs} & V_{cb}\\
V_{td} & V_{ts} & V_{tb}
\end{array}
\right).$$ The unitary matrix $V_{CKM}$ has in general $6$ phases. By phase redefinition of various quark fields, one can remove $5$ of the $6$ phases. The remaining one phase is one of the sources for CP violation in the quark sector. There are many ways to parameterize the CKM matrix, for example, $$V_{CKM} =
\left(
\begin{array}{ccc}
c_{12}^{q} c_{13}^{q} & s_{12}^{q} c_{13}^{q} & s_{13}^{q} e^{-i\delta_{q}}\\
-s_{12}^{q} c_{23}^{q}-c_{12}^{q}s_{23}^{q}s_{13}^{q}e^{i\delta_{q}} &
c_{12}^{q}c_{23}^{q}-s_{12}^{q}s_{23}^{q}s_{13}^{q}e^{i\delta_{q}} &
s_{23}^{q}c_{13}^{q}\\
s_{12}^{q}s_{23}^{q}-c_{12}^{q}c_{23}^{q}s_{13}^{q}e^{i\delta_{q}} &
-c_{12}^{q}s_{23}^{q}-s_{12}^{q}c_{23}^{q}s_{13}^{q}e^{i\delta_{q}} &
c_{23}^{q}c_{13}^{q}
\end{array}
\right).$$ Defining $\lambda=s_{12}^{q}, \; A\lambda^{2} = s_{23}^{q}$ and $A\lambda^{3}(\rho-i\eta)=s_{13}^{q}e^{-i\delta_{q}}$, we obtain an alternative parameterization, the Wolfenstein parameterization,[@Wolfenstein:1983yz] $$V_{CKM} =
\left(
\begin{array}{ccc}
1-\frac{1}{2}\lambda^{2} & \lambda & A\lambda^{3}(\rho-i\eta)\\
-\lambda & 1-\frac{1}{2}\lambda^{2} & A\lambda^{2}\\
A\lambda^{3}(1-\rho-i\eta) & -A\lambda^{2} & 1
\end{array}
\right).$$ Here the parameters $A, \; \rho, \; \eta$ are of order $1$, and $\lambda$ is the sine of the Cabbibo angle which is about $0.22$, and thus a good choice as an expansion parameter. A parameterization independent measure for the CP violation is the Jarlskog invariant,[@Jarlskog:1985cw] defined as $$J_{CP}^{q} \equiv Im \{ V_{ud}V_{us}^{\ast}V_{cd}^{\ast}V_{cs} \}.$$ Unitarity of $V_{CKM}$ applied to the first and the third columns leads to the following condition[^1] $$V_{td}V_{tb}^{\ast} + V_{ud}V_{ub}^{\ast} + V_{cd}V_{cb}^{\ast} = 0.$$ A geometrical representation of this equation on the complex plane gives rise to the CKM “unitarity triangle” shown in Fig.\[triangle\].
\[\]\[\][$\alpha$]{} \[\]\[\][$\beta$]{} \[\]\[\][$\gamma$]{} \[\]\[\][$V_{ud}V_{ub}^{\ast}$]{} \[\]\[\][$V_{td}V_{tb}^{\ast}$]{} \[\]\[\][$V_{cd}V_{cb}^{\ast}$]{}
The three angles of the CKM unitarity triangle are $$\alpha \equiv
Arg(-\frac{V_{td}V_{tb}^{\ast}}{V_{ud}V_{ub}^{\ast}}), \quad
\beta \equiv
Arg(-\frac{V_{cd}V_{cb}^{\ast}}{V_{td}V_{tb}^{\ast}}), \quad
\gamma \equiv
Arg(-\frac{V_{ud}V_{ub}^{\ast}}{V_{cd}V_{cb}^{\ast}}).$$ The Jarlskog invariant $J_{CP}^{q}$ is proportional to the area of the unitarity triangle. A non-vanishing value for $J_{CP}^{q}$ thus implies non-vanishing values for $(\alpha,\beta,\gamma)$ which in turn indicates CP violation.
Textures of Mass Matrices
-------------------------
There have been many mass textures, with different elements having zeros in the mass matrices, proposed in order to accommodate the observed fermion mass hierarchy and mixing pattern. Imposing texture ansatz on mass matrices reduces the number of parameters in the Yukawa sector; as a consequence, masses and mixing angles may be related in some simple ways. This is illustrated in a simplified two family example in the up- and down-quark sectors, in which the mass matrices of the up- and down-type quarks are assumed to be symmetric and each contains one zero entry, $$M_{U}=
\left(\begin{array}{cc}
0 & a\\
a & b
\end{array}
\right), \qquad
M_{D}=
\left(\begin{array}{cc}
0 & c\\
c & d
\end{array}
\right).$$ In this case, there is a very simple relation between the Cabbibo angle and the quark masses, $$|V_{us}| = |\sqrt{\frac{m_{d}}{m_{s}}} - e^{i\alpha}\sqrt{\frac{m_{u}}{m_{c}}}|$$ where the CP violating phase $\alpha$ arises as the relative phase that enters when combining the up- and down-quark rotaion matrices. This relation is in good agreement with experiment. Note that because $m_{u}/m_{c}$ is quite small compared to $m_{d}/m_{s}$, the value of $|V_{us}|$ is not very sensitive to the complex phase $\alpha$. One can then generalize this to consider the three family case. The ultimate goal of studying texture zero ansatz is that it may help us to understand the underlying theory of flavor, if the zeros are protected by some family symmetry. In the quark sector, assuming symmetric mass matrices, the total number of texture zeros is at most six, because there are six different quark masses. Nevertheless, Ramond, Robert and Ross[@Ramond:1993kv] found that the observed masses and mixing angles cannot be accommodated with six texture zeros. They found five combinations of five-zero texture for up- and down-type quark mass matrices which give rise to predictions that are consistent with current observations for fermion masses and mixing angles. One should note that in the context of a grand unified theory (GUT), the texture ansatz is valid only at the GUT scale. The vanishing entries in the mass matrices at the GUT scale will be filled in by radiactive corrections at lower energy scales. These five solutions are summarized in Table.\[5zero\].
Lop-sided textures have also been considered in model building. In section $V$, we classify various $SO(10)$ models according to whether the mass textures in the models are symmetric or lop-sided. Symmetric textures arise if $SO(10)$ breaks down to the SM group with the left-right symmetric group as the intermediate symmetry, while lop-sided textures arise if the intermediate symmetry is $SU(5)$.
Froggatt-Nielsen Mechanism and Family Symmetry
----------------------------------------------
A prototype scenario which produces hierarchy in the fermion mass matrices is the Froggatt-Nielsen mechanism.[@Froggatt:1978nt] The idea is that the heaviest matter fields acquire their masses through tree level interactions with the Higgs fields while masses of lighter matter fields are produced by higher dimensional interactions involving, in addition to the regular Higgs fields, exotic vector-like pairs of matter fields and the so-called flavons (flavor Higgs fields). Schematic diagrams for these interactions are shown in Fig.\[fn\].
After integrating out superheavy vector-like matter fields of mass $M$, the mass terms of the light matter fields get suppressed by a factor of $\frac{<\theta>}{M}$, where $<\theta>$ is the VEVs of the flavons and $M$ is the UV-cutoff of the effective theory above which the flavor symmetry is exact. When the family symmetry is exact, only the $(33)$ entry is non-zero. When the family symmetry is spontaneously broken, the zero entries will be filled in at some order $O(\frac{<\theta>}{M})$. Suppose the family symmetry allows only the $(23)$ and $(32)$ elements at order $O(\frac{<\theta>}{M})$, $$\left(\begin{array}{ccc} 0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 1\\
\end{array}\right)
\qquad
{\mbox{SSB} \atop \longrightarrow} \qquad
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & \frac{<\theta>}{M}\\
0 & \frac{<\theta>}{M} & 1\\
\end{array}\right).$$ Then a second fermion mass is generated at order $O((\frac{<\theta>}{M})^{2})$ after the family symmetry is spontaneous broken. The fermion mass hierarchy thus arises.
To illustrate how the Froggatt-Nielsen mechanism works, suppose there is a vector-like pair of matter fields $(\chi \oplus \overline{\chi})$ with mass $M$ and carrying the same quantum numbers as $\psi_{R}$ under the vertical gauge group (e.g. SM or $SO(10)$), but different quantum numbers under the family symmetry. It is therefore possible to have a Yukawa coupling $y\overline{\chi}\psi_{L} H$ where $H$ is the SM doublet Higgs if the family symmetry permits such a coupling. In addition, there is a gauge singlet $\theta$ which transforms non-trivially under the family symmetry. Suppose the coupling $y^{'}\overline{\psi}_{R}\chi\theta$ is allowed by the family symmetry, we then obtain the following seesaw mass matrix, upon $H$ and $\theta$ acquiring VEV’s $$\left(\overline{\psi}_{R} \; \overline{\chi}\right)
\left(\begin{array}{cc}
0 & y^{'}<\theta>\\
y<H> & M
\end{array}
\right)
\left(\begin{array}{c}
\psi_{L}\\
\chi
\end{array}\right).$$ Diagonalizing this matrix gives the following mass term for $\psi$ $$m_{\psi} \simeq \frac{y y^{'}<H><\theta>}{M}.$$ So the suppression factor $\frac{<\theta>}{M}$ is due to the mixture between the light states and the heavy states. This is very similar to how the light neutrino masses are generated in the seesaw mechanism.[@Gell-Mann:fmass1979a; @Yanagida:1979; @Mohapatra:1979ia]
So what are the possible family symmetries that can be incorporated with the Froggatt-Nielsen mechanism? The kinetic terms and gauge interactions of the SM have a very large family symmetry, $[U(3)]^{5}$, where the $U(3)$ factors act on the right- and left-handed multiplets of quarks and leptons, respectively. If right-handed neutrinos are included, the family symmetry becomes $[U(3)]^{6}$. Therefore, any family symmetry group proposed to incorporate the Froggatt-Nielsen mechanism must be contained in $[U(3)]^{6}$. If all the particles in each family are unified into one single multiplet, as in the case of $SO(10)$, the maximal possible family symmetry group is reduced to $U(3)$. The family symmetry can either be global or gauged. However, in the case of a global symmetry, there are problems associated with the massless Goldstone bosons when the symmetry is broken, and with possibly large gravitational quantum corrections. These problems do not arise in the case of gauged symmetries. In what follows, we discuss separately the abelian group and non-abelian group as a family symmetry.
### Abelian Case
In compactified string theories, one usually obtains aboundant Abelian symmetries below the compactification scale, in additional to the SM gauge group (or a GUT gauge group). Suppose that the flavon field $\theta$ and $\overline{\theta}$ carry $+1$ and $-1$ charges under $U(1)$, and the ratios $\frac{<\theta>}{M}$ and $\frac{<\overline{\theta}>}{M}$ have approximately the value of the Cabbibo angle $0.22$ (or equivalently, the parameter $\lambda$ in the Wolfenstein parameterization). By assigning different $U(1)$ charges to different family, one in general, obtains a mass matrix of the form $$\left(\begin{array}{ccc}
\lambda^{n_{11}} & \lambda^{n_{12}} & \lambda^{n_{13}}\\
\lambda^{n_{21}} & \lambda^{n_{22}} & \lambda^{n_{23}}\\
\lambda^{n_{31}} & \lambda^{n_{32}} & \lambda^{n_{33}}\\
\end{array}
\right).$$ The exponent $n_{ij}$ is given by $Q_{L,i}+Q_{R,j}$, where $Q_{L,i} \; (Q_{R,i})$ is the $U(1)$ charge of the left-handed (right-handed) field of the $i$-th family. It is usually not difficult to find solutions for the charge assignments that are consistent with experiments, and the solutions are not unique. An interesting model for both quarks and leptons based on anomalous $U(1)_{H}$ was proposed by Irges, Lavigna and Ramond which makes use of the Green-Schwartz anomaly cancellation condition to constrain the $U(1)$ charge assignments.[@Irges:1998ax] Due to the Abelian nature, models of this type have limited predictive power, because only the powers of the small expansion parameter, $\lambda$, are determined while the relative strengths between different entries are left un-determined. As a consequence, every entry in the above mass matrix has an un-known $\mathcal{O}(1)$ coefficient associated with it.
Non-anomalous $U(1)_{H}$ has also been considered in a model constructed by Mira, Nardi and Restrepo.[@Mira:1999fx] In this model, anomaly cancellation condition leads to a massless up-quark, thus solving the strong CP problem.
### Non-Abelian Case Based on $SU(2)$
Models with non-Abelian family symmetry generally have more predictive power because the relative strengths between different matrix elements can be determined. The original motivation of using non-Abelian group as the family symmetry is to solve the SUSY flavor problem. $SU(2)$ was proposed by Barbieri [*et al*]{} [@Barbieri:1996ww] as a family symmetry. It has two attractive features: (i) As we have seen previously, the constraints from the SUSY FCNC requires that $$\frac{m_{2}^{2}-m_{1}^{2}}{m^{2}} \le
\frac{10^{-3}}{\sin\phi} (\frac{m}{300 \mbox{GeV}})$$ where $(m_{3}^{2}-m_{1,2}^{2}) \sim m^{2}$ is the average scalar mass squared and $\phi$ is some relevant CP phase. $SU(2)$ gives rise to the degeneracies between 1-2 families needed to suppress the supersymmetric FCNC in the squark sector; (ii) A multi-step breaking of $SU(2)$ gives rise to the observed inter-family hierarchy naturally. Unlike models based on the $U(1)$ family symmetry, in which one has the freedom in choosing $U(1)$ charges for various matter fields, a $SU(2)$ family symmetry appears to be a much more constrained framework for constructing realistic models.
The heaviness of the top-quark and suppression of the SUSY FCNC together suggest that the three families of matter fields transform under a $SU(2)$ family symmetry as $$\psi_{a} \oplus \psi_{3} = 2 \oplus 1$$ where $a =1,2$ and the subscripts refer to family indices. In the symmetric limit, only the third family of matter fields have non-vanishing masses. This can be understood easily since the third family of matter fields have much higher masses compared to the other two families of matter fields. $SU(2)$ breaks down in two steps: $$\label{eq:steps} SU(2) \stackrel{\epsilon M}{\longrightarrow}
U(1) \stackrel{\epsilon' M}{\longrightarrow}
nothing$$ with $\epsilon' \ll \epsilon \ll 1$ and $M$ is the UV cut-off of the effective theory mentioned before. These small parameters $\epsilon$ and $\epsilon'$ are the ratios of the vacuum expectation values of the flavon fields to the cut-off scale. Note that because $$\psi_{3}\psi_{3} \sim 1_{S},\qquad \psi_{3}\psi_{a} \sim 2, \qquad
\psi_{a}\psi_{b} \sim 2 \otimes 2 = 1_{A} \oplus 3$$ the only relevant flavon fields are in the $1_{A}, 2$ and $3$ dimensional representations of $SU(2)$, namely, $$A^{ab} \sim 1_{A}, \qquad \phi^{a} \sim 2, \qquad S^{ab} \sim 3.$$ So a generic mass matrix constrained by $SU(2)$ family symmetry is of the following form: $$\left(\begin{tabular}{c|c}
&\\
$<S>$, $<1_{A}>$ & $<\phi>$\\
&\\
\hline
$<\phi>^{T}$ & 1
\end{tabular}
\right).$$
To see how the vacuum alignment in the flavon sector is achieved, let us first consider the supersymmetric limit with only one conjugate pair of doublets $(\phi \oplus \overline{\phi})$, anti-symmetric singlets $(A \oplus \overline{A})$ and triplets $(S \oplus
\overline{S})$ of $SU(2)$. The most general renormalizable superpotential is then given by[@Barbieri:1998em] $$W_{\mbox{flavon}}(\phi,\overline{\phi},S,\overline{S}) =
\phi \overline{S}\phi + \overline{\phi}S\overline{\phi}
+X_{\phi} \phi\overline{\phi} + X_{s}S\overline{S}
+ X_{A} A \overline{A}$$ where $X_{s}$, $X_{\phi}$ and $X_{A}$ are dimensionful parameters in the superpotential. Note that the anti-symmetric singlet fields $A$ and $\overline{A}$ are decoupled from other fields. From the F-flat conditions, one obtains the following solutions, $$\begin{aligned}
<S_{ab}> & = & -<\phi_{a}><\phi_{b}>/X_{s}
\\
\sum_{a} \; <\phi_{a}\overline{\phi}_{a}> & = & \frac{1}{2}(X_{s}X_{\phi})
\\
X_{A} A^{ab} & = & 0.\end{aligned}$$ Thus the relative strengths of $<S>$ and $<\phi>$ are determined. For $X_{A} \ne 0$, the F-flat conditions imply $<A> = <\overline{A}> = 0$. Non-vanishing $<A>$ and $<\overline{A}>$ can be obtained if non-renormalizable operators are introduced.[@Barbieri:1998em]
If all the $16$ observed matter fields in one family form a single representation as in the case of $SO(10)$, the most general effective superpotential, after integrating out all the heavy Froggatt-Nielsen fields, that generates fermion masses for a $SO(10) \times
SU(2)$ model has the following very simple form $$W = H(\psi_{3}
\psi_{3} + \psi_{3} \frac{\phi^{a}}{M} \psi_{a}
+ \psi_{a}\frac{S^{ab}}{M} \psi_{b} + \psi_{a}\frac{A^{ab}}{M} \psi_{b}).$$ In a specific $SU(2)$ basis, $$\label{genU2}
\frac{ \left< \phi \right> }{M} \sim O \left( \begin{array}{c}
\epsilon'\epsilon\\
\epsilon
\end{array} \right),
\qquad
\frac{ \left< A^{ab} \right> }{M} \sim O \left( \begin{array}{cc}
0 & -\epsilon' \\
\epsilon' & 0
\end{array} \right),
\qquad
\frac{ \left< S^{ab} \right> }{M} \sim O \left( \begin{array}{cc}
\epsilon'^{2} & \epsilon'\epsilon\\
\epsilon'\epsilon & \epsilon
\end{array} \right).$$ Here we have indicated the VEVs *all* the flavon fields could acquire for symmetry breaking in Eq.(\[eq:steps\]). The mass matrix $\widetilde{M}$ takes the following form $$\widetilde{M} \sim \mathcal{O} \left( \begin{array}{ccc}
\epsilon'^{2} & \epsilon' & \epsilon'\epsilon \\
-\epsilon' & \epsilon & \epsilon \\
\epsilon'\epsilon & \epsilon & 1
\end{array} \right).$$ The hierarchy is thus built into this mass matrix.
### Non-Abelian Case Based on $SU(3)$
Ultimately, the $SU(2)$ family symmetry can be embedded into $SU(3)$, under which the three families form a triplet. A model based on $SU(3)$ is presented in Ref. 18. The three families form a $SU(3)$ triplet before the symmetry is broken. The symmetry breaking takes place at two steps $$SU(3) \longrightarrow SU(2) \longrightarrow {\mbox nothing}.$$ The $SU(3)$ anti-triplet flavon fields, $\overline{\phi}_{3}$ and $\overline{\phi}_{23}$, acquire VEV’s along the following directions, breaking the $SU(3)$ symmetry $$\label{vev1}
<\overline{\phi}_{3}> = \left(\begin{array}{c}
0 \\ 0 \\ a_{3}
\end{array}\right), \qquad
<\overline{\phi}_{23}> = \left(\begin{array}{c}
0 \\ b \\ b
\end{array}\right)$$ where $<\overline{\phi}_{3}>$ triggers the first stage of breaking, and $<\overline{\phi}_{23}>$ triggers the second stage of breaking. The leading order Yukawa couplings of the matter fields are $$\label{leading}
H(\frac{1}{M_{3}^{2}} \psi_{i}\overline{\phi}_{3}^{i}\psi_{j}^{c}
\overline{\phi}_{3}^{j}
+ \frac{1}{M_{23}^{2}} \psi_{i}\overline{\phi}_{23}^{i}\psi_{j}^{c}
\overline{\phi}_{23}^{j}).$$ This gives rise to a mass matrix of the following form $$\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & \frac{b^{2}}{M_{23}^{2}} & \frac{b^{2}}{M_{23}^{2}}\\
0 & \frac{b^{2}}{M_{23}^{2}} & \frac{a_{3}^{2}}{M_{3}^{2}}
+ \frac{b^{2}}{M_{23}^{2}}
\end{array}
\right) \sim
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & \epsilon^{2} & \epsilon^{2}\\
0 & \epsilon^{2} & 1
\end{array}\right).$$ Assuming $M_{3} \simeq a_{3} \gg M_{23} \gg b$ and $\epsilon=b/M_{23}$, one thus obtains a hierarchical structure in the $(2,3)$ block. There are also operators that mix $\overline{\phi}_{3}$ with $\overline{\phi}_{23}$, $$\label{mix}
\frac{\epsilon^{2}}{M_{23}M_{3}} \; H
(\psi_{i}\overline{\phi}_{23}^{i}\psi_{j}^{c}\overline{\phi}_{3}^{j}
+\psi_{i}\overline{\phi}_{3}^{i}\psi_{j}^{c}\overline{\phi}_{23}^{j}).$$ These operators give rise to contributions to the $(23)$ and $(32)$ matrix elements of order $\mathcal{O}(\epsilon)$, which is larger than the $(22)$ element of order $\mathcal{O}(\epsilon^{2})$, leading to a wrong prediction for $V_{cb}$. One way to suppress these operators is to impose a discrete $Z_{2}$ symmetry, under which $\overline{\phi}_{3}$ and $\overline{\phi}_{23}$ have opposite parity. Thus the operators given in Eq.(\[leading\]) are allowed by the $Z_{2}$ symmetry while the operators given in Eq.(\[mix\]) only arise at higher order with a suppression factor $\epsilon^{2}$. To fill in the first row and column, one has to introduce additional $SU(3)$ triplet flavon fields, $\phi_{3}$ and $\phi_{23}$, which acquire VEV’s along the following directions $$\label{vev2}
<\phi_{3}> = \left(\begin{array}{c}
0 \\ 0 \\ a_{3}
\end{array}\right), \qquad
<\phi_{23}> = \left(\begin{array}{c}
0 \\ b \\ -b
\end{array}\right)$$ and consider the following operators $$\begin{aligned}
& \frac{\epsilon^{2}}{M_{23}} (\epsilon^{ijk} \psi_{i}
\phi_{23,j} \psi_{k}^{c}) H \hspace{3.3cm}
\label{13op}\\
& \frac{\epsilon^{6}}{M_{3}^{2}M_{23}^{2}}
(\epsilon^{ijk} \psi_{i}\phi_{3,j}\phi_{23,k})
(\epsilon^{lmn}\psi_{l}^{c}\phi_{3,m}\phi_{23,n}) H.
\label{11op}\end{aligned}$$ The operators given in Eq.(\[13op\]) generate the entries $(12) = (13) = -(21) = -(31)$ of order $\mathcal{O}(\epsilon^{3})$. These entries are anti-symmetric due to the presence of the anti-symmetric tensor $\epsilon^{ijk}$ in the couplings. The operators given in Eq.(\[11op\]) generate the $(11)$ matrix element of order $\mathcal{O}(\epsilon^{8})$. The suppression in this operator is due to the presence of the $Z_{2}$ symmetry discussed above and an additional $R$-symmetry; these symmetries also forbid all the operators which could lead to un-realistic predictions. The operators given in Eq.(\[leading\]), (\[mix\]), (\[13op\]) and (\[11op\]) together give rise to a Yukawa matrix of the form, in the leading order of the expansion parameter, $\epsilon$, $$\left(\begin{array}{ccc}
\mathcal{O}(\epsilon^{8}) & \lambda \epsilon^{3} & \lambda \epsilon^{3}\\
-\lambda \epsilon^{3} & \epsilon^{2} & \epsilon^{2}\\
-\lambda \epsilon^{3} & \epsilon^{2} & 1
\end{array}\right)$$ which can accommodate realistic fermion masses and mixing angles. The vacuum alignment leading to Eq.(\[vev1\]) and (\[vev2\]) is discussed in detail in Ref.18.
We comment that the absolute mass scale of the family symmetry, $M$, cannot be determined in the Froggatt-Nielsen type of scenario, because it is the ratio, $(\epsilon, \epsilon^{'})$, rather than the absolute mass scale, $M$, that is phenomenologically relevant. Some attempts have been made by having the SUSY breaking messenger fields play also the role of Froggatt-Nielsen fields such that the family symmetry scale is linked to the SUSY breaking scale. Though attractive, these models have difficulties getting low SUSY breaking scale.[@Arkani-Hamed:1996xm]
Ideas from Extra Dimensions
---------------------------
### Split Fermion Scenario in Factorizable Geometry
It has been proposed that the existence of extra compact spatial dimensions could account for the large hierarchy between the Planck scale and the electroweak scale. Arkani-Hamed, Dimopoulos and Dvali[@Arkani-Hamed:1998rs] showed that extra dimensions of size $\sim 1/TeV$ may provides a solution to the gauge hierarchy problem.
Based on this framework, Arkani-Hamed and Schmaltz[@Arkani-Hamed:1999dc] proposed a non-supersymmetric model in which the mass hierarchy is generated by localizing zero modes of the weak doublet and singlet fermions at different locations. Note that this mechanism works despite of the size of the extra dimensions. Consider a chiral fermion $\Psi$ in $5D$. The action for $\Psi$ coupled to a background scalar field $\Phi$ is given by $$\label{split1}
S = \int d^{4} dy \overline{\Psi} \left[
i \gamma^{\mu} \partial_{\mu} + i\gamma^{5} \partial_{5} + \Phi(y) \right] \Psi.$$ The chiral fermion $\Psi(x,y)$ can be expanded in the product basis $$\Psi(x,y) = \sum_{n} <y|L,n> P_{L} \psi_{n}(x) + <y|R,n> P_{R} \psi_{n}(x).$$ Here $|L,n>$ and $|R,n>$ satisfy the following equations $$\begin{aligned}
aa^{\dagger} |L,n> & = & (-\partial_{5}^{2} + \Phi^{2} + \dot{\Phi})|L,n>
= \mu^{2}_{n} |L,n>
\\
a^{\dagger}a |R,n> & = & (-\partial_{5}^{2} + \Phi^{2} - \dot{\Phi})|R,n>
= \mu_{n}^{2} |R,n>,\end{aligned}$$ where $\dot{\Phi} \equiv \partial_{5}\Phi$ and $$a = \partial_{5} + \Phi(y), \qquad a^{\dagger} = -\partial_{5} + \Phi(y).$$ For a special choice of linearized background field, $\Phi(y) = 2\mu^{2}y$ where $\mu$ is a constant of mass dimension one, the operator $a$ and $a^{\dagger}$ become the usual creation and annihilation operators of a simple harmonic oscillator. In what follows, we will concentrate on this special case. Expanding the $5D$ action in terms of $|L,n>$ and $|R,n>$, and then integrating out $y$, we obtain the $4D$ effective action $$S = \int d^{4}[\overline{\psi}_{L} i \gamma^{\mu} \partial_{\mu} P_{L} \psi_{L}
+ \overline{\psi}_{R} i \gamma^{\mu} \partial_{\mu} P_{R} \psi_{R}
+ \sum_{n=1}^{\infty} \overline{\psi}_{n} (i\gamma^{\mu}\partial_{\mu} + \mu_{n})
\psi_{n} ],$$ where the first two terms are kinetic terms of the chiral zero modes whose wave functions are Gaussian centered at $y=0$, $$<y|L,0> = \frac{\mu^{1/2}}{(\pi/2)^{1/4}}e^{-\mu^{2}y^{2}}.$$ In general, there is a bulk mass term, $m\overline{\Psi}\Psi$, in the $5D$ action, Eq.(\[split1\]), because mass terms for $5$-dimensional field are allowed by all symmetries. In this case, instead of being centered at $y=0$, the Gaussian wave function is centered at $y=m/2\mu$. Thus by having different bulk mass term, $m$, different bulk fields can be localized at different $4D$ slices in the $5D$ bulk.
For simplicity, the wave function of the Higgs doublet is assumed to have constant spread along the fifth dimension. In this case, the Yukawa coupling is $$S_{yuk} = y_{ij} \int d^{4} x H(x) f_{i}(x) \tilde{f}_{j}^{c}(x)
\int dy \phi_{f_{i}}(y) \phi_{\tilde{f}^{c}_{j}}(y),$$ where $\phi_{f_{i}}(y)$ and $\phi_{\tilde{f}^{c}_{j}}(y)$ are the zero mode wave functions of the $SU(2)$ doublet, $f_{i}$, and singlet, $\tilde{f}_{j}^{c}$, respectively. Integrating out the $y$ coordinate, the effective Yukawa coupling is then given by $$y_{ij} \int dy \phi_{f_{i}} (y) \phi_{\tilde{f_{j}}^{c}} (y) = y_{ij}
\frac{\sqrt{2}\mu}{\sqrt{\pi}}
\int dy e^{-\mu^{2}(y-r_{f_{i}})^{2}} e^{-\mu^{2}(y-r_{\tilde{f}^{c}_{j}})^{2}}
= e^{-\mu^{2}r_{ij}^{2}/2},$$ where $r_{f_{i}}=m_{f_{i}}/2\mu$ and $r_{\tilde{f}^{c}_{j}}
=m_{\tilde{f}^{c}_{j}}/2\mu$ are the locations at which the Gaussian wave functions $\phi_{f_{i}}$ and $\phi_{\tilde{f}^{c}_{j}} (y)$ are centered, and $r_{ij} = |r_{\tilde{f}^{c}_{j}}-r_{f_{i}}|$ is the “distance” between the peaks of these two wave functions. Thus the large hierarchy among fermion masses can be generated by having different $m_{i}$.
Variations based on this mechanism have been investigated in Ref. 22, 23, 24.
### Non-Factorizable Geometry
Randall and Sundrum[@Randall:1999ee] proposed an alternative based on non-factorizable geometry from which the gauge hierarchy can be derived. In their original proposal, only gravity propagates in the bulk. If the SM Higgs doublet is confined to the TeV brane, which is required if the gauge hierarchy is assumed to arise from the warped geometry, while all other SM particles are allowed to propagate in the bulk, it is possible to understand the fermion mass hierarchy in this setup. The equations of motion for various bulk fields are given in the following compact form[@Grossman:1999ra; @Gherghetta:2000qt] $$\label{geneom}
(e^{2\sigma}\eta^{\mu\nu}\partial_{\mu}\partial_{\nu}
+ e^{s\sigma}\partial_{5}(e^{-s\sigma}\partial_{5}) - M_{\Phi}^{2})
\Phi(x^{\mu},y) = 0,$$ where for $\Phi=(\phi,e^{-2\sigma}\Psi_{L,R})$ we have $M_{\Phi}^{2}=(ak^{2}+b\sigma^{''}(y),C(C \pm 1)k^{2}
\pm C\sigma^{''}(y))$ and $s=(4,1)$, where $a$ and $C$ are bulk mass terms of the scalar and fermionic fields, and $b$ is a boundary mass term for the scalar field. The field $\Phi(x^{\mu},y)$ is decomposed into an infinite sum of Kaluza-Kline (KK) modes as follows, $$\label{KKdecompose}
\Phi(x^{\mu},y)=\frac{1}{\sqrt{2\pi R}} \sum_{n}
\Phi_{(n)} (x^{\mu}) f_{n}(y).$$ The profile of the n-th mode, $f_{n}(y)$, satisfies $$\label{eomfn}
(-e^{s\sigma}\partial_{5}(e^{-s\sigma} \partial_{5})+\hat{M}_{\Phi}^{2})
f_{n}(y) = e^{2\sigma} m_{n}^{2} f_{n}(y),$$ where $\hat{M}_{\Phi}^{2} = (ak^{2}, C(C \pm 1)k^{2})$. $m_{n}$ is the mass of the $n$-th KK mode. The solution for the zero modes of the bulk spin-$1/2$ fields are found to be [@Grossman:1999ra; @Gherghetta:2000qt] $$\label{LH}
f_{0}(y) = \frac{1}{N_{0}} e^{-c\sigma},$$ where $c=C$ for left-handed fermions and $c=-C$ for right-handed fermions. The normalization constant $1/N_{0}$ is given by $$\frac{1}{N_{0}^{2}} = \frac{(1-2c) \pi k R}{e^{(1-2c)\pi k R} -1}.$$ Thus the bulk fermion can be decomposed into $$\Psi_{L}(x^{\mu},y) = e^{2\sigma} \Phi(x^{\mu},y)
= \frac{1}{\sqrt{2\pi R}} \frac{1}{N_{0}} e^{(2-c)\sigma}
\Phi_{L(0)}(x^{\mu}) + \cdots.$$ The Yukawa coupling for the charged fermions to the Higgs doublet reads $$\label{yuk}
\frac{Y_{ij}}{M_{pl}} \int d^{4}x \int dy \sqrt{-g}
\overline{\Psi}_{R,i}(x,y) \Psi_{L,j}(x,y) H(x) \delta(y-\pi R),$$ where $Y_{ij}$ are dimensionless $\mathcal{O}(1)$ coefficients. The effective Yukawa coupling in four dimensions is obtained after integrating out the fifth coordinate, $y$: $$\begin{aligned}
\tilde{Y}_{ij} & = & \frac{Y_{ij}}{M_{pl}}
(\frac{1}{2\pi R})
\left[
\frac{(1-2c_{R,i})\pi kR}{e^{(1-2c_{R,i})\pi kR} -1 }
\right]^{1/2}
\left[
\frac{(1-2c_{L,j})\pi kR}{e^{(1-2c_{L,j})\pi kR} -1 }
\right]^{1/2}
\nonumber\\
& & \qquad \quad \cdot
\int_{-\pi R}^{\pi R} dy \sqrt{-g} \;
e^{(2-c_{R,i})\sigma} e^{(2-c_{L,j})\sigma}
e^{\sigma} \delta(y-\pi R)
\nonumber\\
& = & \frac{Y_{ij}}{2} \frac{k}{M_{pl}}
\frac{\sqrt{(1-2c_{R,i})(1-2c_{L,j})}}
{\sqrt{(e^{(1-2c_{R,i})\pi kR}-1)(e^{(1-2c_{L,i})\pi kR}-1)}}
e^{(1-c_{R,i} - c_{L,j})\pi k R}.\end{aligned}$$ Thus by choosing $c_{L,i}$ and $c_{R,i}$ all of $\mathcal{O}(1)$, we can reproduce the observed mass hierarchy and mixing. We note that this mechanism is not predictive in the sense that it does not reduce the number of parameters; the virtue of this mechanism is that the large hierarchy observed in fermion masses arises from parameters all of $\mathcal{O}(1)$. A configuration that reproduce the observed mass hierarchy has been found by Huber and Shafi.[@Huber:2000ie]
Lepton Masses and Mixing {#lmass}
========================
If neutrinos are massive, the mixing arises in the leptonic charged current interaction, $$\mathcal{L}_{cc} = \frac{g}{\sqrt{2}} U_{LM}^{\dagger} (W_{\mu}^{+}
\overline{\nu}_{L} \gamma_{\mu} E_{L}) + h.c..$$ The leptonic mixing (LM) matrix,[@pontecorvo:1967; @Maki:mu; @Lee:1977qz; @Lee:1977ti] $U_{LM}$,[^2] is obtained by diagonalizing the Yukawa matrix of the charged leptons and the effective neutrino mass matrix, assuming neutrinos are Majorana particles,[^3] $$\begin{aligned}
Y_{e}^{diag} & = & V_{e_{R}} Y_{e} V_{e_{L}}^{\dagger}
= diag(y_{e},y_{\mu},y_{\tau})\\
M_{\nu_{LL}}^{diag} & = & V_{\nu_{LL}} M_{\nu_{LL}}^{eff} V_{\nu_{LL}}^{T}
= diag(m_{\nu_{1}},m_{\nu_{2}},m_{\nu_{3}}),\label{nudiag}\end{aligned}$$ where $V_{\nu_{LL}}$ is an orthogonal matrix, and it can be parameterized as a product of a CKM-like mixing matrix, which has three mixing angles and one CP violating phase, with a diagonal phase matrix, $$\begin{aligned}
& U_{LM} \equiv V_{e_{L}}V_{\nu_{LL}}^{\dagger} \hspace{10cm}\nonumber\\
& \simeq
\left(
\begin{array}{ccc}
c_{12}^{l} c_{13}^{l} & s_{12}^{l} c_{13}^{l} & s_{13}^{l} e^{-i\delta_{l}}\\
-s_{12}^{l} c_{23}^{l}-c_{12}^{l}s_{23}^{l}s_{13}^{l}e^{i\delta_{l}} &
c_{12}^{l}c_{23}^{l}-s_{12}^{l}s_{23}^{l}s_{13}^{l}e^{i\delta_{l}} &
s_{23}^{l}c_{13}^{l}\\
s_{12}^{l}s_{23}^{l}-c_{12}^{l}c_{23}^{l}s_{13}^{l}e^{i\delta_{l}} &
-c_{12}^{l}s_{23}^{l}-s_{12}^{l}c_{23}^{l}s_{13}^{l}e^{i\delta_{l}} &
c_{23}^{l}c_{13}^{l}
\end{array}
\right)
\cdot \left(
\begin{array}{ccc}
1 & &\\
& e^{i \frac{\alpha_{21}}{2}} & \\
& & e^{i\frac{\alpha_{31}}{2}}
\end{array}
\right),\nonumber\\\end{aligned}$$ which relates the neutrino mass eigen states to the flavor eigenstates by $$\begin{aligned}
|\nu_{e}> & = &
U_{e\nu_{1}} |\nu_{1}> + U_{e\nu_{2}} |\nu_{2}> + U_{e\nu_{3}} |\nu_{3}>
\\
|\nu_{\mu}> & = &
U_{\mu\nu_{1}} |\nu_{1}> + U_{\mu\nu_{2}} |\nu_{2}> + U_{\mu\nu_{3}} |\nu_{3}>
\\
|\nu_{\tau}> & = &
U_{\tau\nu_{1}} |\nu_{1}> + U_{\tau\nu_{2}} |\nu_{2}> + U_{\tau\nu_{3}} |\nu_{3}>.\end{aligned}$$ Note that the Majorana condition, $$C(\overline{\nu}_{j})^{T} = \nu_{j},$$ where $C$ is the charge conjugate operator, forbids the rephasing of the Majorana fields. Therefore, we can only remove $3$ of the $6$ phases present in the unitary matrix $U_{LM}$ by redefining the charged lepton fields and are left with three CP violating phases in the leptonic sector, if neutrinos are Majorana particles.[@Bilenky:1980cx; @Schechter:1980gr; @Doi:1980yb] Thus $U_{LM}$ can be parameterized as a product of a unitary matrix, analogous to the CKM matrix which has one phase (the so-called universal phase), $\delta_{l}$, and a diagonal phase matrix which contains two phases (the so-called Majorana phases), $\alpha_{21}$ and $\alpha_{31}$. The leptonic analog of the Jarlskog invariant, which measures the CP violation due to the universal phase, is given by $$J_{CP}^{l} \equiv Im\{ U_{\mu \nu_{2}} U_{e\nu_{3}}
U_{\mu \nu_{3}}^{\ast} U_{e\nu_{2}}^{\ast} \}.$$ For the Majorana phases, the rephasing invariant CP violation measures are [@Nieves:1987pp] $$S_{1} \equiv Im\{ U_{e\nu_{1}} U_{e\nu_{3}}^{\ast} \}, \qquad
S_{2} \equiv Im\{ U_{e\nu_{2}} U_{e\nu_{3}}^{\ast} \}.$$ From $S_{1}$ and $S_{2}$, one can then determine the Majorana phases $$\begin{aligned}
\cos \alpha_{31} = 1 - 2 \frac{S_{1}^{2}}{\vert U_{e\nu_{1}} \vert^{2}
\vert U_{e\nu_{3}} \vert^{2}}
\\
\cos (\alpha_{31} - \alpha_{21})
= 1 - 2 \frac{S_{2}^{2}}{\vert U_{e\nu_{2}} \vert^{2}
\vert U_{e\nu_{3}} \vert^{2}}.\end{aligned}$$
The recently reported measurements from KamLAND reactor experiment[@Eguchi:2002dm] confirm the large mixing angle (LMA) solution to be the unique oscillation solution to the solar neutrino problem at $4.7 \; \sigma$ level.[@Maltoni:2002aw; @Bahcall:2002ij; @Fogli:2002au] The global analysis including Solar + KamLAND + CHOOZ[@Apollonio:1999ae] indicate the following allowed region at $3\sigma$,[@Maltoni:2002aw] $$\begin{aligned}
5.1 \times 10^{-5} < & \Delta m_{21}^2 & < 9.7 \times 10^{-5} eV^{2}
\\
0.29 \le & \tan^{2}\theta_{12} & \le 0.86.\end{aligned}$$ The allowed regions at $3\sigma$ level based on a global fit including SK[@Fukuda:2000np] + Solar + CHOOZ for the atmospheric parameters and the CHOOZ angle are[@Gonzalez-Garcia:2002dz] $$\begin{aligned}
1.4 \times 10^{-3} < & \Delta m_{32}^2 & < 6.0 \times 10^{-3} eV^{2}
\\
0.4 \le & tan^{2} \theta_{23} & \le 3.0\\
& \sin^{2} \theta_{13} & < 0.06.\end{aligned}$$ And the magnitudes of $U_{LM}$ elements at $1 \sigma$ ($3 \sigma$) are given by[@Gonzalez-Garcia:2003qf] $$|U_{LM}| = \left(
\begin{array}{ccc}
(0.73) 0.79 - 0.86 (0.88) & (0.47) 0.50 - 0.61 (0.67) & 0 - 0.16 (0.23)\\
(0.17) 0.24 - 0.52 (0.57) & (0.37) 0.44 - 0.69 (0.73) & (0.56) 0.63 - 0.79 (0.84)\\
(0.20) 0.26 - 0.52 (0.58) & (0.40) 0.47 - 0.71 (0.75) & (0.54) 0.60 - 0.77 (0.82)
\end{array}
\right)$$
To see what a bi-large mixing means, let us assume, to a good approximation, that $\theta_{13}=\eta \ll 1$. In this case, the LM matrix reads, $$U_{LM} = \left(\begin{array}{ccc}
c_{12} & s_{12} & \eta \\
-\frac{(s_{12}+\eta c_{12})}{\sqrt{2}}
& \frac{(c_{12}-\eta s_{12})}{\sqrt{2}} & 1/\sqrt{2}\\
\frac{(s_{12}-\eta c_{12})}{\sqrt{2}} & -\frac{(c_{12}+\eta s_{12})}{\sqrt{2}}
& 1/\sqrt{2}
\end{array}
\right),$$ which in turn implies the neutrino mass eigenstates given in terms of flavor eigenstates as $$\begin{aligned}
|\nu_{1}> & = &
c_{12} |\nu_{e}> - \frac{1}{\sqrt{2}}s_{12}
( \; |\nu_{\mu}> \; - \; |\nu_{\tau}> \; )
- \frac{\eta}{\sqrt{2}} c_{12} ( \; |\nu_{\mu}> \; + \; |\nu_{\tau}> \; )
\\
|\nu_{2}> & = &
s_{12} |\nu_{e}> + \frac{1}{\sqrt{2}} c_{12}
( \; |\nu_{\mu}> \; - \; |\nu_{\tau}> \; )
-\frac{\eta}{\sqrt{2}} s_{12} ( \; |\nu_{\mu}> \; + \; |\nu_{\tau}> \; )
\\
|\nu_{3}> & = & \frac{1}{\sqrt{2}} ( \; |\nu_{\mu}> \; + \; |\nu_{\tau}> \; )
+ \eta |\nu_{e}>.\end{aligned}$$
Generation of Small Neutrino Masses
-----------------------------------
### Small Neutrino Masses from See-saw Mechanism
The observation of neutrino oscillations provides the first indication of beyond the Standard Model physics. It has two implications: neutrinos have non-zero masses, and lepton family numbers are violated. In the SM, neutrinos are massless because there are no $SU(2)$ singlet neutrinos nor are there $SU(2)$ triplet Higgss. Adding one $SU(2)$ singlet neutrino for each family is the simplest way to introduce neutrino masses. Because the right-handed neutrinos are SM singlets, the Majorana mass terms for the $SU(2)$ singlet neutrinos are not forbidden by the symmetry. In the Lagrangian, there are Dirac mass term for the neutrinos and the right-handed Majorana mass terms,[@Gell-Mann:fmass1979a; @Yanagida:1979; @Mohapatra:1979ia] $$\begin{aligned}
\mathcal{L}_{seesaw} & = &
-M_{\nu_{LR}}\overline{\nu}_{R}\nu_{L}
- \frac{1}{2}M_{\nu_{RR}}\overline{\nu}_{R}^{T}\overline{\nu}_{R} + h.c.
\nonumber\\
& = & -\frac{1}{2}
\left(\begin{array}{cc}\nu_{L} & \overline{\nu}_{R}
\end{array}\right)
\left(\begin{array}{cc}
0 & M_{LR}^{T}\\
M_{LR} & M_{RR}
\end{array}\right)
\left(\begin{array}{c}
\nu_{L} \\ \overline{\nu}_{R}
\end{array}\right).\end{aligned}$$ (In general, there could be a non-vanishing mass term for $(\nu_{L}\nu_{L})$; this is the Type II see-saw mechanism.[@Mohapatra:1979ia] Such a mass term can be obtained in $SO(10)$ from the Yukawa coupling given in Eq.(\[typeII\])). After integrating out the heavy right-handed neutrinos, $$\begin{aligned}
-\frac{\partial\mathcal{L}}{\partial \overline{\nu}_{R}}
& = & M_{LR}\nu_{L} + M_{RR} \overline{\nu}_{R}^{T}
\nonumber\\
\overline{\nu}_{R} & = & -\nu^{T}M_{LR}^{T}M_{RR}^{-1},\end{aligned}$$ one then obtains the effective light neutrino Majorana mass matrix $$\label{seesaw}
M_{LL} = M_{\nu_{LR}}^{T} M_{RR}^{-1} M_{\nu_{LR}}.$$ For $M_{LR}$ of the order of the weak scale as the mass scale of other charged fermions, the Majorana mass term $M_{RR}$ must be around $10^{12-14} GeV$ to give rise to neutrino masses of the order of $0.1 \; eV$. As we will see later, most grand unified theories predict the existence of the right-handed neutrinos. Furthermore, the GUT scale provides an understanding why $M_{RR}$ is large.
### Small Neutrino Masses from Large Extra Dimension
An interesting way to generate small [*Dirac*]{} neutrino masses arises in models with large extra dimensions of size $\sim (1/TeV)$.[@Dienes:1998sb; @Dienes:2000ph] Consider the case in which only gravity can propagate in the bulk, while the SM particles and interactions are confined to the brane. Because the right-handed neutrinos are SM singlets, they are the only particles that can propagate in the bulk. Its Yukawa coupling to the charged lepton $L(x)$, and the Higgs doublet, $H(x)$, which are confined to the brane, is given by $$S = y \int d^{4}x \int dy L(x) H(x) \nu_{R}(x,y) \delta(y).$$ Compactified on a circle $S^{1}$, $\nu_{R}(x,y)$ can be decomposed into $$\nu_{R}(x,y) = \frac{1}{\sqrt{2\pi R M_{pl}^{5D}}} \sum_{-\infty}^{\infty}
\nu_{R}^{(n)} (x) e^{i n y /R}.$$ Below the compactification scale, we thus obtain a Dirac neutrino mass $$m_{\nu} = \frac{y <H>}{\sqrt{2 \pi R M_{pl}^{5D}}}
= \frac{y <H> M_{pl}^{5D}}{M_{pl}^{4D}},$$ where tha last steps follows from the relation between the $4D$ Planck scale and $5D$ Planck scale, $2 \pi R M_{pl}^{5D} = (M_{pl}^{4D}/M_{pl}^{5D})^{2}$. With $M_{pl}^{5D} \sim (1-10) TeV$ which could avoid the gauge hierarchy problem, one obtains a highly suppressed $m_{\nu}$ which is consistent with the experimental observations.
### Small Neutrino Masses from Warped Geometry
Small neutrino masses of the Dirac type are possible if right-handed neutrinos are localized toward the Planck brane while the lepton doublets are localized toward the TeV brane. This results in a small overlap between the zero mode profiles of the lepton doublets and the right-handed neutrinos based on the formulation given in Sec. II. Models of this type have been constructed in Ref. 26, 28, 43.
Bi-Large Neutrino Mixing Angles
-------------------------------
To obtain the bi-large mixing pattern for the neutrinos, in additional to having the hierarchical mass pattern, let us consider, for example, the following mass texture,[@Chen:2002pa] $$\label{texture}
\left(\begin{array}{ccc}
0 & 0 & t\\
0 & 1 & 1+t^{n}\\
t & 1+t^{n} & 1
\end{array}
\right),$$ with $t < 1$ which is a special case of the following texture $$\label{cmtexture}
\left(\begin{array}{ccc}
0 & 0 & \ast\\
0 & \ast & \ast\\
\ast & \ast & \ast
\end{array}
\right)$$ first proposed in Ref. 49 in which the elements in $(2,3)$ block are taken to have equal strengths to accommodate near bi-maximal mixing. The modification of adding the term $t^{n}$ in the $(23)$ and $(32)$ entries is needed in order to accommodate a large, but non-maximal solar angle in the so-called “light side” region ($0 < \theta < \pi/4$).[@deGouvea:2000cq] It is possible to obtain the LMA solution at $3\sigma$ level with $n$ ranging from $1$ to $2$.
An interesting alternative in which a $3 \times 2$ neutrino Dirac mass matrix is considered was proposed recently by Kuchimanchi and Mohapatra[@Kuchimanchi:2002yu; @Kuchimanchi:2002fi]. A $3 \times 2$ neutrino Dirac mass matrix arises if there are only two right-handed neurtinos, instead of three. The existence of two right-handed neutrinos is required by the cancellation of Witten anomaly, if a global leptonic $SU(2)$ family symmetry is imposed[@Kuchimanchi:2002yu; @Kuchimanchi:2002fi]. Along this line, Frampton, Glashow and Yanagida proposed a model, which has the following Lagrangian,[@Frampton:2002qc] $$\label{FGY}
\mathcal{L} = \frac{1}{2} (N_{1} N_{2})
\left(\begin{array}{ccc}
M_{1} & 0\\
0 & M_{2}
\end{array}\right)
\left(\begin{array}{c} N_{1} \\ N_{2} \end{array}\right)
+ (N_{1} N_{2}) \left(\begin{array}{ccc}
a & a^{'} & 0\\
0 & b & b^{'}
\end{array}\right)
\left(\begin{array}{c} l_{1} \\ l_{2} \\ l_{3}
\end{array}\right) H + h.c..$$ The effective neutrino mass matrix due to this Lagrangian is obtained, using the see-saw formula, $$\left(\begin{array}{ccc}
\frac{a^{2}}{M_{1}} & \frac{aa^{'}}{M_{1}} & 0
\\
\frac{aa^{'}}{M_{1}} & \frac{a^{'2}}{M_{1}} + \frac{b^{2}}{M_{2}} & \frac{bb^{'}}{M_{2}}
\\
0 & \frac{bb^{'}}{M_{2}} & \frac{b^{'2}}{M_{2}}
\end{array}\right),$$ where $a, b, b^{'}$ are real and $a^{'} = |a^{'}| e^{i\delta}$. By takinging all of them to be real, with the choice $a^{'} = \sqrt{2} a$ and $b=b^{'}$, and assuming $a^{2}/M_{1} \ll b^{2}/M_{2}$, the effective neutrino masses and mixing matrix are obtained $$m_{\nu_{1}} = 0, \quad m_{\nu_{2}} = \frac{2a^{2}}{M_{1}}, \quad
m_{\nu_{3}} = \frac{2b^{2}}{M_{2}}$$ $$U = \left(\begin{array}{ccc}
1/\sqrt{2} & 1/\sqrt{2} & 0\\
-1/2 & 1/2 & 1/\sqrt{2}\\
1/2 & -1/2 & 1/\sqrt{2}
\end{array}\right) \times
\left(\begin{array}{ccc}
1 & 0 & 0\\
0 & \cos\theta & \sin\theta \\
0 & -sin\theta & \cos\theta
\end{array}\right),$$ where $\theta \simeq m_{\nu_{2}}/\sqrt{2}m_{\nu_{3}}$, and the observed bi-large mixing angles and $\Delta m_{atm}^{2}$ and $\Delta m_{\odot}^{2}$ can be accommodated. An interesting feature of this model is that the sign of the baryon number asymmetry ($B \propto \xi_{B} =
Y^{2} a^{2} b^{2} \sin2\delta$) is related to the sign of the CP violation in neutrino oscillation ($\xi_{osc}$) in the following way $$\xi_{osc} = -\frac{a^{4}b^{4}}{M_{1}^{3}M_{2}^{3}} (2 + Y^{2}) \xi_{B} \propto - B$$ assuming the baryon number asymmetry is resulting from leptogenesis due to the decay of the lighter one of the two heavy neutrinos, $N_{1}$, whose mass is of $\mathcal{O}(10^{10} \; GeV)$. A $SO(10)$ model which gives rise to the neutrino mass ansatz, Eq.(\[FGY\]), has been constructed.[@Raby:2003ay] A more detailed discussion on this model is given in Sec. 6.
Other phenomenologically viable textures for neutrino mass matrix are analyzed in Ref. 55.
### $SO(10)$ GUT realization
In $SO(10)$ models in $4D$, the bi-large mixing in the leptonic sector arises in two ways (A detailed classification according to how the maximal $\nu_{\mu}-\nu_{\tau}$ mixing arises is given by Barr and Dorsner[@Barr:2000ka]):\
[(i) Symmetric mass textures for the charged fermions:]{}
[This scenario is realized in symmetric textures arising from left-right symmetric breaking chain of $SO(10)$. In this case, both the large solar mixing angle and the maximal atmospheric mixing angle come from the diagonalization of the effective neutrino mass matrix. A characteristic of this class of models is that the predicted value for $|U_{e\nu_{3}}|$ element tends to be larger than the value predicted by models in class (ii) below.]{}\
[(ii) Lop-sided mass textures for charged fermions:]{}
[In this scenario, the large atmospheric mixing angle comes from charged lepton mixing matrix. This scenario is realized in models with $SU(5)$ as the intermediate symmetry which gives rise to the so-called “lop-sided” mass textures, due to the $SU(5)$ relation, $$M_{e} = M_{d}^{T}.$$ Due to the lop-sided nature of $M_{e}$ and $M_{d}$, the large atmospheric neutrino mixing is related to the large mixing in the $(23)$ sector of the RH charged lepton diagonalization matrix, instead of $V_{cb}$. It thus provides an explanation why the small value of $V_{cb}$ and the large value of $U_{\mu\nu_{3}}$ exist simultaneously. The large solar mixing angles comes from the diagonalization matrix for the neutrino mass matrix. Because the two large mixing angles come from different sources, the constraint on $U_{e\nu_{3}}$ is not as strong as in class (i). In fact, the prediction for $U_{e\nu_{3}}$ in this class of models tend to be quite small. On the other hand, this mechanism also predicts an enhanced decay rate for the flavor-violating process, $\mu \rightarrow e \; \gamma$, which is close to current experimental limit.]{}
We will discuss these two classes of models in detail in Sec.\[models\].
### Large Mixing from Renormalization Group Evolution
It has been shown that it is possible to obtain large neutrino mixing angles through the renormalization group evolution.[@Balaji:2000gd; @Balaji:2000au; @Balaji:2000ma; @Antusch:2002hy; @Antusch:2002fr] Recently, Mohapatra, Parida and Rajasekaran observed in Ref. 62 that bi-large mixing angles can be driven by the renormalization group evolution, assuming that the CKM matrix and the LM matrix are identical at the GUT scale, which is a natural consequence of quark-lepton unification. The only requirement for this mechanism to work is that the masses of the three neutrinos are nearly degenerate of the form $m_{3} \gtrsim m_{2} \gtrsim
m_{1}$ and have same CP parity. The one-loop RGE of the effective left-handed Majorana neutrino mass operator is given by[@Chankowski:1993tx; @Babu:qv; @Antusch:2001ck; @Antusch:2001vn] [^4] $$\label{rgem}
\frac{d m_{\nu}}{dt}=-\{\kappa_{u}m_{\nu}+m_{\nu}P+P^{T}m_{\nu}\},$$ where $t \equiv \ln \mu$. In the MSSM, $P$ and $\kappa_{u}$ are given by, $$\begin{aligned}
P & = & -\frac{1}{32\pi^{2}} \frac{Y_{e}^{\dagger}Y_{e}}{\cos^{2} \beta}
\simeq -\frac{1}{32\pi^{2}} \frac{h_{\tau}^{2}}{\cos^{2}\beta} diag(0,0,1)
\equiv diag(0,0,P_{\tau})\\
\kappa_{u} & = &
\frac{1}{16\pi^{2}}[\frac{6}{5}g_{1}^{2} + 6g_{2}^{2}
- 6 \frac{Tr(Y_{u}^{\dagger}Y_{u})}{\sin^{2}\beta}]
\simeq
\frac{1}{16\pi^{2}}[\frac{6}{5}g_{1}^{2} + 6g_{2}^{2}
- 6 \frac{h_{t}^{2}}{\sin^{2}\beta}], \end{aligned}$$ where $g_{1}^{2}=\frac{5}{3}g_{Y}^{2}$ is the $U(1)$ gauge coupling constant, $Y_{u}$ and $Y_{e}$ are the $3 \times 3$ Yukawa coupling matrices for the up-quarks and charged leptons respectively, and $h_{t}$ and $h_{\tau}$ are the $t$- and $\tau$-Yukawa couplings. One can then follow the “diagonal-and-run” procedure, and obtain the RGE’s at scales between $M_{R} \ge \mu \ge M_{SUSY}$ for the mass eigenvalues and the three mixing angles, assuming CP violating phases vanish, $$\begin{aligned}
\frac{d \; m_{i}}{d t} & = &
-4 P_{\tau} m_{i} U_{\tau \nu_{i}}^{2} - m_{i} \kappa_{u}, \quad (i=1,2,3)
\label{nurge1}\\
\frac{d \; s_{23}}{d t} & = & -2P_{\tau} c_{23}^{2}
(-s_{12}U_{\tau\nu_{1}}\nabla_{31} + c_{12}U_{\tau\nu_{2}}\nabla_{32})
\\
\frac{d \; s_{13}}{dt} & = & -2P_{\tau}c_{23}c_{13}^{2}
(c_{12}U_{\tau\nu_{1}}\nabla_{31}+s_{12}U_{\tau\nu_{2}}\nabla_{32})
\\
\frac{d \; s_{12}}{dt} & = & -2P_{\tau}c_{12}
(c_{23}s_{13}s_{12}U_{\tau\nu_{1}}\nabla_{31}
-c_{23}s_{13}c_{12}U_{\tau\nu_{2}}\nabla_{32}
+U_{\tau\nu_{1}}U_{\tau\nu_{2}}\nabla_{21}),
\label{nurge4}\end{aligned}$$ where $\nabla_{ij} \equiv (m_{i}+m_{j})/(m_{i}-m_{j})$. Because the LM matrix is identical to the CKM matrix, we have, at the GUT scale, the following initial conditions, $s_{12}^{0} \simeq \lambda, \quad
s_{23}^{0} \simeq \mathcal{O}(\lambda^{2})$ and $s_{13}^{0} \simeq \mathcal{O}(\lambda^{3})$, where $\lambda$ is the Wolfenstein parameter. (Note that the RG evolution has negligible effect on the Wolfenstein parameter, see Eq.(\[rgeV\])). When the masses $m_{i}$ and $m_{j}$ are nearly degenerate, $\nabla_{ij}$ approaches infinity. Thus it drives the mixing angles to become large. Starting with the values of $(m_{1}^{0},m_{2}^{0},m_{3}^{0})
= (0.2983,0.2997,0.3383) \; eV$ at the GUT scale, the solutions at the weak scale for the masses are $(m_{1},m_{2},m_{3})
= (0.2410,0.2411,0.2435) \; eV$, which correspond to $\Delta m_{atm}^{2} =
1.1 \times 10^{-3} eV^{2}$ and $\Delta m_{\odot}^{2} = 4.8 \times 10^{-5} eV^{2}$. The mixing angles predicted at the weak scale are $\sin\theta_{23} = 0.68$, $\sin\theta_{12} = 0.568$ and $\sin\theta_{13} = 0.08$. Because the masses are larger than $0.1 \; eV$, they are testable at the present searches for the neutrinoless double beta decay.
Models based on GUT and horizontal symmetry often suffer from fine-tunning or the difficulty of constructing a viable scalar potential that gives rise to the required vacua. Along the line discussed in the above paragraph, some attempts have been made to show that the maximal mixing angle[@Chankowski:1999xc; @Casas:1999tp; @Casas:1999ac] and nearly degenerate neutrino masses[@Ellis:1999my; @Ma:1999xq; @Haba:1999xz; @Chankowski:2000fp; @Chen:2001gk; @Miura:2002nz; @Bhattacharyya:2002aq; @Frigerio:2002in] are manifestations of infrared fixed points (IRFP) of the RGEs given above in Eq.(\[nurge1\])-(\[nurge4\]), under certain assumptions.
### Bi-large Mixing and $b-\tau$ Unification
In the minimal $SO(10)$ model utilizing Type II see-saw mechanism with one $10$ and one $\overline{126}$, we have the following relations: $$\begin{aligned}
M_{u} & = & f <10> + h <\overline{126}^{+}>
\\
M_{d} & = & f <10> + h <\overline{126}^{-}>
\\
M_{e} & = & f <10> -3 h <\overline{126}^{-}>
\\
M_{\nu_{LR}} & = & f <10> -3 h <\overline{126}^{+}>\end{aligned}$$ where $f$ and $h$ are Yukawa matrices; the mass terms $M_{\nu,LL}$ and $M_{\nu,RR}$ are both due to the coupling to $\overline{126}$, $$\begin{aligned}
M_{\nu,LL} & = & h <\overline{126}^{'+}>
\\
M_{\nu,RR} & = & h <\overline{126}^{'0}>\end{aligned}$$ where the superscripts $+/-/0$ refer to the sign of the hypercharge $Y$ (see Table 3). The small neutrino masses are explained by the Type II see-saw mechanism with the [*assumption*]{} that the LH Majorana mass term dominates over the usual Type I see-saw term, thus it is proportional to the Yukawa matrix $h$, which can be determined by calculating the difference between $M_{d}$ and $M_{e}$. Using down-type quark masses, charged lepton masses, and CKM matrix elements, which have roughly the form $$M_{b,\tau} \sim
\left(
\begin{array}{ccc}
\lambda^{3} & \lambda^{3} & \lambda^{3}\\
\lambda^{3} & \lambda^{2} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}
\right) m_{b,\tau} \quad,$$ predictions for neutrino masses and LM matrix elements have been made.[@Bajc:2002iw; @Goh:2003sy] The large atmospheric mixing results from a small deviation of $\mathcal{O}(\lambda^{2})$ from $b-\tau$ unification.
Predictions from a detailed numerical study made in Ref. 79 are $\sin^{2}2\theta_{23} < 0.9$ and $\sin^{2}2\theta_{12} > 0.9$, which are experimentally allowed only at $3 \sigma$ level; these unique predictions can thus be used to test this type of models. Note that the best fit value of $\sin^{2}2\theta_{23}$ [*cannot*]{} be accommodated in these models. The prediction for $U_{e3}$ is about $0.16$, very close to the sensitivity of current experiments. We also note that as this type of models do not address the origin of the flavor structure, they are not as predictive as $SO(10)$ models combined with family symmetry, in which as non-minimal Higgs content is usually present (see Sec. 6).
CP violation in Leptonic Sector
-------------------------------
As we mentioned previously, if neutrinos are Dirac particles, there is only one phase in the LM matrix. On the other hand, if neutrinos are Majorana particles, which is the case if see-saw mechanism is implemented to give small neutrino masses, there are two additional phases. These two types of phases have very different impacts on phenomenology. The universal phase, $\delta_{l}$, affects the transition probability in the neutrino oscillation $$\begin{aligned}
P(\nu_{\alpha}\rightarrow \nu_{\beta}) & = &
\delta_{\alpha\beta} - 4 \sum_{i>j}
Re\{ U_{\alpha i} U_{\beta j} U_{\alpha j}^{\ast}
U_{\beta i}^{\ast} \} \sin^{2} (\Delta m_{ij}^{2} \frac{L}{4E})\nonumber\\
& &
+ 2 \sum_{i>j}
Im\{ U_{\alpha i} U_{\beta j} U_{\alpha j}^{\ast}
U_{\beta i}^{\ast} \} \sin^{2} (\Delta m_{ij}^{2} \frac{L}{2E}).\end{aligned}$$ Note that the only chance that one might observe CP violation in neutrino oscillation is to have LMA solution in the solar sector, and to have large value for $\theta_{13}$. The Majorana phases affect the matrix element for the neutrinoless double beta $(\beta\beta_{0\nu})$ decay, $\vert < m > \vert$, given in terms of the rephasing invariant quantities by $$\begin{aligned}
\vert < m > \vert ^{2} & = &
m_{1}^{2} \vert U_{e1} \vert^{4} + m_{2}^{2} \vert U_{e2} \vert^{4}
+ m_{3}^{2} \vert U_{e3} \vert^{4}
\nonumber\\
& & +
2m_{1}m_{2} \vert U_{e1} \vert^{2} \vert U_{e2} \vert^{2} \cos\alpha_{21}
\nonumber\\
& & +
2m_{1}m_{3} \vert U_{e1} \vert^{2} \vert U_{e3} \vert^{2} \cos\alpha_{31}
\nonumber\\
& & +
2m_{2}m_{3} \vert U_{e2} \vert^{2} \vert U_{e3} \vert^{2}
\cos(\alpha_{31}-\alpha_{21}).\end{aligned}$$ The current bound on $|<m>|$ from Heidelberg-Moscow experiment is $0.11-0.56 \; eV$ at $95\%$ confidence level.[@Klapdor-Kleingrothaus:2001ke]
Grand Unified Theories Based on $SO(10)$ {#so10}
========================================
The smallest GUT group $SU(5)$ in its minimal form is very strongly constrained due to the non-observation of proton decay.[@Murayama:2001ur; @Bajc:2002bv; @Emmanuel-Costa:2003pu] The next candidate is the rank-$5$ $SO(10)$, which is a very attractive candidate as a GUT group for many reasons: First of all, all of its irreducible representations are free of anomaly, unlike in the case of $SU(5)$ where the representations of the matter fields, $\overline{5} \oplus 10$, are carefully chosen to cancel the anomaly. It unifies all the $15$ known fermions with the right-handed neutrino for each family into one $16$-dimensional spinor representation. The seesaw mechanism then arises very naturally, and the small but non-zero neutrino masses can thus be explained, as evidenced by recent atmospheric neutrino oscillation data from Super-Kamiokande indicating small non-zero neutrino masses. Because a complete quark-lepton symmetry is achieved, it has the promise for explaining the pattern of fermion masses and mixing. In some $SO(10)$ models, R-parity is conserved automatically at all energy scales. This is to be contrasted with MSSM and SUSY $SU(5)$ where R-parity must be imposed by hand. Because $(B-L)$ is a gauge symmetry contained in $SO(10)$, it has the promises of baryogenesis. In what follows, we briefly review the structure of $SO(10)$ models. A detail discussion can be found in Ref. 84.
The Algebra of $SO(2n)$ {#susygut-algebra}
-----------------------
It is convenient to discuss the $SO(2n)$ algebra in the $SU(n)$ basis.[@Mohapatra:1979nn] Consider a set of $n$ operators $\; \xi_{i} \; (i = 1, ..., n)$, and their hermitian conjugates, $\; \xi_{i}^{\dagger}$, satisfying $$\{\xi_{i},\xi_{j}^{\dagger}\} = \delta_{ij},
\qquad
\{\xi_{i},\xi_{j}\} = 0,$$ where $\{ \; , \; \}$ denotes an anti-commutator and $[ \; , \; ]$ denotes a commutator. The operators $K^{i}_{j}$ defined as $$K^{i}_{j} \equiv \xi_{i}^{\dagger}\xi_{j}$$ satisfy the algebra of the $U(n)$ group $$[K^{i}_{j},K^{m}_{n}] = \delta^{m}_{j} K^{i}_{n}
- \delta^{i}_{n} K^{m}_{j}.$$ We can then define the following $2n$ operators, $\Gamma_{\mu}\;
(\mu=1,...,2n)$ $$\begin{aligned}
\Gamma_{2j-1} & = & -i \; ( \; \xi_{j} - \xi_{j}^{\dagger} \; )
\nonumber\\
\Gamma_{2j} & = & ( \; \xi_{j} + \xi_{j}^{\dagger}\; ), \qquad j = 1,...,n.\end{aligned}$$ The $\Gamma_{\mu}$ form the Clifford algebra of rank $2n$ $$\{ \Gamma_{\mu}, \Gamma_{\nu} \} = 2 \delta_{\mu\nu}$$ and they can then be used to construct the generators of $SO(2n)$ as follows: $$\Sigma_{\mu\nu} = \frac{1}{2i} \; [\Gamma_{\mu}, \Gamma_{\nu}].$$ The dimensionality of the spinor representation of $SO(2n)$ is $2^{n}$. In terms of the $SU(n)$ basis, the spinor representation of $SO(2n)$ can then be constructed by, $$\begin{aligned}
|0> & \sim & 1\\
\xi_{i}^{\dagger} \; |0> & \sim & n\\
\xi_{i}^{\dagger} \; \xi_{j}^{\dagger} \; |0> & \sim & \frac{n(n-1)}{2}\\
\xi_{i}^{\dagger} \; \xi_{j}^{\dagger} \; \xi_{l}^{\dagger} \; |0>
& \sim & \frac{n(n-1)(n-2)}{6}\\
.....\\
\xi_{1}^{\dagger} \; ... \; \xi_{n}^{\dagger} \; |0> & \sim & n\end{aligned}$$ where $|0>$ is the $SU(n)$ invariant vacuum state. The spinor representation can then be split into two $2^{n-1}$-dimensional representations by a chiral projection operator. Let us define $$\Gamma_{0} \equiv i^{n} \Gamma_{1} \; \Gamma_{2} \; ... \; \Gamma_{2n}$$ and the number operator $$N_{i} \equiv \xi_{i}^{\dagger} \xi_{i}.$$ $\Gamma_{0}$ then can be written as $$\begin{aligned}
\Gamma_{0} & = & [\xi_{1},\xi_{1}^{\dagger}] \; [\xi_{2},\xi_{2}^{\dagger}] \;
... \; [\xi_{n},\xi_{n}^{\dagger}]
\nonumber\\
& = & \prod_{i-1}^{n} \; (1-2n_{i})
\nonumber\\
& = & (-1)^{n}.\end{aligned}$$ To arrive at the last step, we have used the property of the number operator $n_{i}^{2} \; = \; n_{i}$ to get $1-2n_{i} = (-1)^{n_{i}}$ and $n = \sum_{i}
\; n_{i}$. One can then check that $$[\Sigma_{\mu\nu},\Gamma_{0}]=0.$$ The chirality projection operator is therefore defined by $$\frac{1}{2} (1 \pm \Gamma_{0}).$$ Consider the case $n=5$ and define a column vector $|\psi>$: $$\begin{aligned}
|\psi> & = & |0> \psi_{0}
+ \xi_{j}^{\dagger} |0> \psi_{j}
+ \frac{1}{2}\xi_{j}^{\dagger}\xi_{k}^{\dagger} |0> \psi_{jk}
+ \frac{1}{12}
\epsilon^{ijklm} \xi_{k}^{\dagger}
\xi_{l}^{\dagger}\xi_{m}^{\dagger}|0>\overline{\psi}_{ji} \nonumber\\
& & + \frac{1}{24}
\epsilon^{jklmn} \xi_{k}^{\dagger}
\xi_{l}^{\dagger}\xi_{m}^{\dagger}\xi_{n}^{\dagger}|0>\overline{\psi}_{j}
+ \xi_{1}^{\dagger}
\xi_{2}^{\dagger}\xi_{3}^{\dagger}\xi_{4}^{\dagger}\xi_{5}^{\dagger}|0>
\overline{\psi}_{0}\end{aligned}$$ where $\overline{\psi}$ is not the complex conjugate of $\psi$ but an independent vector. This can be generalized to any $n$ if we write $$\psi = \left(
\begin{array}{cccccc}
\psi_{0} &
\psi_{i} &
\psi_{ij} &
\overline{\psi}_{ij} &
\overline{\psi}_{i} &
\overline{\psi}_{0}
\end{array}
\right)^{T}.$$ The spinor representation is then split under the chirality projection operator as $$\psi = \left(
\begin{array}{c}
\psi_{+}\\
\psi_{-}
\end{array}
\right)$$ where $$\psi_{\pm} = \frac{1}{2} (1 \pm \Gamma_{0}) \; \psi$$ and $$\psi_{+} =
\left(
\begin{array}{c}
\psi_{0}\\
\psi_{ij}\\
\overline{\psi}_{j}
\end{array}
\right), \qquad
\psi_{-} =
\left(
\begin{array}{c}
\overline{\psi}_{0}\\
\overline{\psi}_{ij}\\
\psi_{j}
\end{array}
\right).$$ In the case of $n=5$, $\overline{\psi}_{i}$ and $\psi_{ij}$ are $\overline{5}$ and $10$-dimensional representations of $SU(5)$ and $\psi_{0}$ is the singlet. All the SM fermions are assigned to $\psi_{+}$. The electric charge formula for $SO(10)$ is given by $$Q = \frac{1}{2}\Sigma_{78} - \frac{1}{6}(\Sigma_{12} + \Sigma_{34} +
\Sigma_{56}).$$
The dimensionality of the adjoint representation of $SO(2n)$ is $\frac{(2n)(2n-1)}{2}$. For $SO(10)$, it is $45$-dimensional. Under the decomposition with respect to $SU(3)
\times SU(2)_{L} \times SU(2)_{R}$ these $45$ gauge bosons are: $$\begin{aligned}
45 & = & (8,1,1) + (1,3,1) + (1,1,1)\nonumber\\
& & + (1,1,3) + (\overline{3},2,2) + (3,2,2) + (3,1,1) + (\overline{3},1,1).\end{aligned}$$ In this basis, the $12$ Standard Model gauge fields are in the $(8,1,1), \;
(1,3,1)$ and $(1,1,1)$ multiplets. The rest are $33$ new gauge bosons which could potentially mediate proton decay.
Symmetry Breaking
-----------------
Because $SO(10)$ is a rank-$5$ group while SM is a rank-$4$ group, there exist several intermediate symmetries through which $SO(10)$ can descend to $SU(3)\times SU(2)_{L} \times U(1)_{Y}$. There are four maximal subgroups of $SO(10)$: $SU(4)\times SU(2)_{L} \times SU(2)_{R}$, $SU(5)\times U(1)$, $SO(9)$, and $SO(7)\times SU(2)$. Only through the breaking chains of $SU(4)\times SU(2)_{L} \times SU(2)_{R}$ and $SU(5)\times U(1)$ can one obtain the correct quantum numbers for the SM particle content. Due to the presence of these intermediate scales, the predictions for proton lifetime and $\sin^{2}\theta_{w}$ are much less certain, compared to the case of $SU(5)$. Details of breaking chains giving the SM are as follows:
[(i) The left-right symmetry breaking chain is]{} $$\begin{aligned}
\label{sblr}
SO(10)
& {<54> \atop \longrightarrow}
& SU(4) \times SU(2)_{L} \times SU(2)_{R} \nonumber\\
& {<45> \atop \longrightarrow}
& SU(3) \times SU(2)_{L} \times SU(2)_{R} \times U(1)_{B-L}\nonumber\\
& {<126 \oplus \overline{126}> \atop \longrightarrow}
& SU(3) \times SU(2)_{L} \times U(1)_{Y} \nonumber\\
& {<10> \atop \longrightarrow}
& SU(3) \times U(1)_{EM}.\end{aligned}$$ In this case, the hypercharge $Y$ is given by $Y=2T_{3R}+(B-L)$. The first step of breaking to the left-right symmetry group is achieved by a symmetric two-index tensor, $<54>$. $SU(4)$ in the left-right symmetry group is then broken to $SU(3)\times U(1)_{B-L}$ by the adjoint $<45>$. The subsequent breaking to the SM gauge group is achieved by the anti-symmetric $5$-index tensor, $<126>$ and $<\overline{126}>$; the electroweak symmetry breaking is then achieved by $<10>$. In realistic models, the two Higgs doublets in MSSM are linear combinations of the $SU(2)$ doublet components from different $SO(10)$ representations of Higgses, for example, $10$ and $\overline{126}$. Thus all fields in the linear combinations contribute to electroweak symmetry breaking.
[(ii)]{} For the $SU(5)$ breaking chain, $$\begin{aligned}
SO(10)
& \rightarrow
& SU(5) \times U(1)_{x}
\nonumber\\
& \rightarrow
& SU(3) \times SU(2) \times U(1)_{z} \times U(1)_{x}
\nonumber\\
& \rightarrow
& SU(3) \times SU(2) \times U(1)_{Y}
\nonumber\\
& \rightarrow
& SU(3) \times U(1)_{EM}\end{aligned}$$ the hypercharge $Y$ is given by $\frac{1}{2} Y = \alpha z + \beta x$ where $z$ and $x$ are the charges under $U(1)_{z}$ and $U(1)_{x}$ respectively. There are two possible ways to embed the SM under this route: $(\alpha,\beta) =
(1/6,0)$ or $(-1/15,-2/5)$. In the case of $(\alpha,\beta) = (1/6,0)$, we obtain $Y = \frac{1}{6}(2z)$. This corresponds to the $SU(5)$ breaking chain $$\begin{aligned}
SO(10)
& {<16 \oplus \overline{16}> \atop \rightarrow}
& SU(5)
\nonumber\\
& {<45> \atop \rightarrow}
& SU(3) \times SU(2) \times U(1)_{Y}
\nonumber\\
& {<10> \atop \rightarrow}
& SU(3) \times U(1)_{EM}.\end{aligned}$$ In this case, the spinors $<16 \oplus \overline{16}>$ break the symmetry down to $SU(5)$; $<45>$ then breaks $SU(5)$ down to the SM. In the case of $(\alpha,\beta) = (-1/15,-2/5)$, we have $Y = \frac{-1}{15}(z+6x)$. This corresponds to flipped $SU(5)$ (that is, $SU(5) \times U(1)$) breaking chain $$\begin{aligned}
SO(10)
& {<45> \atop \rightarrow}
& SU(5) \times U(1)_{x}
\nonumber\\
& {<16 \oplus \overline{16}> \atop \rightarrow}
& SU(3) \times SU(2) \times U(1)_{Y}
\nonumber\\
& {<10> \atop \rightarrow}
& SU(3) \times U(1)_{EM}.\end{aligned}$$ For this breaking to occur, again we need $<16 \oplus \overline{16}>$ and $<45>$. Each breaking chain has its characteristic set of Higgs fields and symmetry breaking superpotential. The electroweak symmetry breaking is then achieved by $<10>$. In realistic models, the two Higgs doublets in MSSM are linear combinations of the $SU(2)$ doublet components from different $SO(10)$ representations of Higgses, for example, $10$ and $16$. As we will see below, different symmetry breaking chains give rise to different mass relations between various sectors.
Renormalization Group Equation and the Georgi-Jarlskog (GJ) relations
---------------------------------------------------------------------
Before describing the Yukawa sector of the $SO(10)$ models, we discuss how to relate the weak scale observables to the GUT scale parameters. We use the expressions derived from 1-loop RGEs given by:[@Arason:1991ic; @Castano:1993ri; @Barger:1992ac; @Berezhiani:1998vn] $$\begin{aligned}
m_{u} = Y_{u}^{0}R_{u}\eta_{u}B_{t}^{3}v_{u}, \quad &
m_{c} = Y_{c}^{0}R_{u}\eta_{c}B_{t}^{3}v_{u}, \quad &
m_{t} = Y_{c}^{0}R_{u}B_{t}^{6}v_{u} \nonumber\\
m_{d} = Y_{d}^{0}R_{d}\eta_{d}v_{d}, \quad &
m_{s} = Y_{s}^{0}R_{d}\eta_{s}v_{d}, \quad &
m_{b} = Y_{b}^{0}R_{d}\eta_{b}B_{t}v_{d} \nonumber\\
m_{e} = Y_{e}^{0}R_{e}v_{d}, \quad &
m_{\mu} = Y_{\mu}^{0}R_{e}v_{d}, \quad &
m_{\tau} = Y_{\tau}^{0}R_{e}v_{d} \end{aligned}$$ and $$\label{rgeV}
V_{ij} = \bigg\{
\begin{array}{lll}
V_{ij}^{0}, & & ij = ud, us, cd, cs, tb \\
V_{ij}^{0} B_{t}^{-1}, & & ij = ub, cb, td, ts
\end{array}$$ where $V_{ij}$ are CKM matrix elements; quantities with superscript $0$ are evaluated at GUT scale, and all the $m_{f}$ and $V_{ij}$ are the experimental values at $M_{Z}$;[@Hagiwara:fs; @Hocker:2001xe; @Abe:2001xe; @Aubert:2001nu] $Y_{f}^{0}$ are Yukawa couplings, and $v_{u}$ and $v_{d}$ are VEV of the Higgs fields $H_{u}$ and $H_{d}$. The SM Higgs VEV is $v=\sqrt{v_{u}^{2}+v_{d}^{2}}=246/\sqrt{2} \; GeV.$ The running factor $\eta_{f}$ includes QCD + QED contributions: For $f=b,c$, $\eta_{f}$ is for the range $m_{f}$ to $m_{t}$, and for $f=u,d,s$, $\eta_{f}$ is for the range $1 GeV$ to $m_{t}$; $$\begin{aligned}
\eta_{u}=\eta_{d}=\eta_{s}=2.38_{-0.19}^{+0.24}\nonumber\\
\eta_{c}=2.05_{-0.11}^{+0.13}\nonumber\\
\eta_{b}=1.53_{-0.04}^{+0.03}.\end{aligned}$$ $R_{u,d,e}$ are contributions of the gauge-coupling constants running from weak scale $M_{z}$ to the SUSY breaking scale, taken to be $m_{t}$, with the SM spectrum, and from $m_{t}$ to the GUT scale with MSSM spectrum; $$R_{u}=3.53_{-0.07}^{+0.06}, \qquad
R_{d}=3.43_{-0.06}^{+0.07}, \qquad
R_{e}=1.50.$$ $B_{t}$ is the running induced by large top-quark Yukawa coupling defined by $$B_{t}=\exp\left[\frac{-1}{16 \pi^{2}} \int_{\ln M_{SUSY}}^{\ln M_{GUT}}
Y_{t}^{2}(\mu) d(\ln \mu)\right]$$ which varies from $0.7$ to $0.9$ corresponding to the perturbative limit $Y_{t}^{0}\approx 3$ and the lower limit $Y_{t}^{0} \approx 0.5$ imposed by the top-pole mass.
Naively, one would expect that masses of the down type quarks are identical to masses of the charged leptons, because of the quark-lepton unification. This turns out to be not true. Taking the experimentally measured values for masses of the down-type quarks and those of the charged leptons, evolving these values using the RGE’s from the weak scale to the GUT scale, however, one finds the following approximate relations[@Georgi:1979df; @Georgi:1979ga] $$m_{d} \simeq 3 m_{e}, \qquad m_{s} \simeq \frac{1}{3}m_{\mu}, \qquad
m_{b} \simeq m_{\tau}.$$ These are known as the [*Georgi-Jarlskog relations*]{}. As it will become apparent later, one way to satisfy these relations is by having a relative factor of $-3$ in the $(22)$ entry of the charged lepton mass matrix with respect to that of the down-type quarks, and all other elements are identical in these two mass matrices. An example suggested by Georgia and Jarlskog is: $$M_{u} = \left(
\begin{array}{ccc}
0 & A & 0\\
A & 0 & B\\
0 & B & C
\end{array}\right),
M_{d} = \left(
\begin{array}{ccc}
0 & E & 0\\
E & F & 0\\
0 & 0 & G
\end{array}\right),
M_{e} = \left(
\begin{array}{ccc}
0 & E & 0\\
E & -3F & 0\\
0 & 0 & G
\end{array}\right).$$ This factor arises naturally as a Clebsch-Gordon (CG) coefficient in some models of $SO(10)$.
Yukawa sector in $SO(10)$
-------------------------
In $SO(10)$, at the renormalizable level, only three types of Higgs fields can couple to fermions, $$16 \otimes 16 = 10_{S} \oplus 120_{A} \oplus 126_{S}$$ namely, $10$, $120_{A}$, and $\overline{126}_{S}$, where the subscripts $S$ and $A$ refer to the symmetry property under interchanging two family indices in the Yukawa couplings $\mathcal{Y}_{ab}$. That is, $$\mathcal{Y}_{ab}^{10} = \mathcal{Y}_{ba}^{10},\quad
\mathcal{Y}_{ab}^{120} = -\mathcal{Y}_{ba}^{120},\quad
\mathcal{Y}_{ab}^{\overline{126}} = \mathcal{Y}_{ba}^{\overline{126}}.$$ The gauge invariant Yukawa couplings are then given by $$\begin{aligned}
\mathcal{Y}_{ab}^{10} (16)_{a} (16)_{b} (10) & = &
\mathcal{Y}_{ab}^{10} \Psi_{a}^{T} B C^{-1} \Gamma_{\alpha} \Psi_{b}
H_{\alpha}
\\
\mathcal{Y}_{ab}^{120} (16)_{a} (16)_{b} (120) & = &
\mathcal{Y}_{ab}^{120} \Psi_{a}^{T}BC^{-1} \Gamma_{\alpha}
\Gamma_{\beta} \Gamma_{\gamma} \Psi_{b} \Lambda_{\alpha\beta\gamma}
\\
\mathcal{Y}_{ab}^{126} (16)_{a} (16)_{b} (\overline{126}) & = &
\mathcal{Y}_{ab}^{126} \Psi_{a}^{T}BC^{-1} \Gamma_{\alpha}
\Gamma_{\beta} \Gamma_{\gamma} \Gamma_{\delta} \Gamma_{\xi} \Psi_{b}
\overline{\Delta}_{\alpha\beta\gamma\delta\xi}\end{aligned}$$ where $\Psi$ denotes the matter fields; $H_{\alpha}, \;
\Lambda_{\alpha\beta\gamma}$ and $\overline{\Delta}_{\alpha\beta\gamma\delta\xi}$ denote the $10$-, $120$- and $\overline{126}$-dim Higgs fields, respectively. $C$ is the usual Dirac charge conjugate operator and $B$ is the charge conjugate operator for $SO(10)$ defined by $$B^{-1} \Gamma_{\mu}^{T}B=-\Gamma_{\mu}$$ and we can choose $$B=\prod_{\mu=\mbox{odd}} \Gamma_{\mu}.$$ Under this charge conjugation $$B\left(
\begin{array}{c}
\psi_{0}\\
\psi_{ij}\\
\overline{\psi}_{i}\\
\psi_{i}\\
\overline{\psi}_{ij}\\
\overline{\psi}_{0}
\end{array}\right)
=\left(\begin{array}{c}
\overline{\psi}_{0}\\
-\overline{\psi}_{ij}\\
\psi_{i}\\
-\overline{\psi}_{i}\\
\psi_{ij}\\
\psi_{0}
\end{array}\right).$$
Note that $SO(10)$ can break down to SM through many different breaking chains. Different breaking chains give rise to different mass relations among the up-quark, down-quark, charged lepton and neutrino sectors. In what follows, we discuss the two symmetry breaking separately.\
[(i) $SU(5)$ breaking chain:]{}\
Under the $SU(5)$ decomposition, we have $$\begin{aligned}
16 & = & 1 + \overline{5} + 10\\
10 & = & 5 + \overline{5}\\
120 & = & 5 + \overline{5} + 10 + \overline{10} + 45 + \overline{45}\\
126 & = & 1 + \overline{5} + 10 + \overline{15} + 45 + \overline{50}\end{aligned}$$ where the $SU(5)$ component $5$ and $45$ contain the $Y = +1$ $SU(2)_{L}$ Higgs doublet, and the $SU(5)$ component $\overline{5}$ and $\overline{45}$ contain the $Y = -1$ $SU(2)_{L}$ Higgs doublet. The $SU(5)$ singlet contained in $\overline{126}$ gives masses to the RH neutrinos through the coupling $(16_{i})(16_{j})(\overline{126}_{H})$. (As we will discuss later, some models utilize non-renormalizable operators $(16_{i})(16_{j})(16_{H})(16_{H})$ to generate RH neutrino masses. This can be achieved because $16_{H}$ also contains a $SU(5)$ singlet component.) The $\overline{15}$ of $SU(5)$ contained in $\overline{126}$ has a $(1,3,2)$ component under $SU(3) \times SU(2)_{L}
\times U(1)_{Y}$ which couples to two lepton doublets as $(1,2,-1)(1,2,-1)(1,3,2)_{H}$ under $G_{SM}$ and gives the LH neutrino Majorana masses in the Type II see-saw mechanism.
As the neutral components in these $SU(2)_{L}$ doublets acquires VEV’s of the electroweak scale, the following mass matrices are obtained $$\begin{aligned}
\label{su5mass}
& M_{u} = \mathcal{Y}^{10}_{ab} \left< 5(10) \right>
+ \mathcal{Y}^{120}_{ab} \left< 45(120) \right>
+ \mathcal{Y}^{\overline{126}}_{ab} \left< 5(\overline{126}) \right>
\equiv Y_{u} v_{u} \hspace{2.9cm}
\\
& M_{d} = \mathcal{Y}^{10}_{ab} \left< \overline{5}(10) \right>
+ \mathcal{Y}^{120}_{ab} ( \left< \overline{5}(120) \right>
+ \left< \overline{45}(120) \right> )
+ \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{45}(\overline{126}) \right>
\equiv Y_{d} v_{d} \qquad
\\
& M_{e} = \mathcal{Y}^{10}_{ab} \left< \overline{5}(10) \right>
+ \mathcal{Y}^{120}_{ab}
(\left< \overline{5}(120) \right>
-3 \left< \overline{45}(120) \right> )
-3 \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{45}(\overline{126})\right>
\equiv Y_{e} v_{d}
\qquad \\
& M_{\nu_{LR}} = \mathcal{Y}^{10}_{ab} \left< 5(10) \right>
+ \mathcal{Y}^{120}_{ab} \left< 5(120) \right>
-3 \mathcal{Y}^{\overline{126}}_{ab} \left< 5(\overline{126}) \right>
\equiv Y_{\nu_{LR}} v_{u} \hspace{2.6cm}\end{aligned}$$ where we denote the $m$-dim $SU(5)$ component of the $n$-dim representation of $SO(10)$ by $m(n)$; $v_{u}$ and $v_{d}$ are the vacuum expectation values of the two Higgs doublets in MSSM. A Clebsch-Gordon coefficient $(-3)$ is generated in the lepton sectors when the $SU(5)$ component $\overline{45}$ are involved in the Yukawa couplings. This factor of $(-3)$ is very crucial for obtaining the Georgi-Jarlskog relations as we have seen in the previous section. The neutrino Majorana mass matrices are given by $$\begin{aligned}
M_{\nu,RR} & = & \mathcal{Y}^{\overline{126}}_{ab} \left< 1(\overline{126})
\right>
\\
M_{\nu,LL} & = & \mathcal{Y}^{\overline{126}}_{ab} \left<
15(\overline{126}) \right>.\label{typeIIsu5}\end{aligned}$$
To see how these CG coefficients $(-3)$ come about, let us decompose the following representations under $SU(5)$. The $120$ and $\overline{126}$-dimensional representation of $SO(10)$ both contain a component which transforms as $\overline{45}$ under $SU(5)$. The Yukawa interactions then can be written as $$(16)(16)(120), \; (16)(16)(\overline{126})
\supset (10)(\overline{5})(\overline{45})
= \psi^{\alpha\beta}\psi_{\gamma}<H_{\alpha\beta}^{\gamma}>.$$ We then write out all the terms in the summation $$\psi^{\alpha\beta}\psi_{\gamma} <H_{\alpha\beta}^{\gamma}>
\supset
(\psi^{45}\psi_{4} <H_{45}^{4}> + \psi^{a5} \psi_{a} <H_{a5}^{a}>)
=(-3 e^{+}e^{-} + d d^{c}) v^{'}.$$ Here the index $a = 1, 2, 3$. Note that $H_{\alpha\beta}^{\gamma}$ is anti-symmetric under inter-changing $\alpha \leftrightarrow \beta$ and it is traceless $$H_{\alpha\beta}^{\gamma} = - H_{\beta\alpha}^{\gamma}, \qquad
\sum_{\beta} H_{\beta 5}^{\beta} = 0.$$ This implies that $$3H_{a5}^{a} + H_{45}^{4} = 0, \qquad
<H_{45}^{4}> = - 3<H_{a5}^{a}> \equiv -3 v^{'}.$$ Essentially, the CG factor of $(-3)$ is related to the fact that there are three colors.\
We note that if $10$ and $\overline{126}$ are the only fields utilized in the Yukawa sector, we have the up-quark mass matrix related to the Dirac neutrino mass matrix, and the down-quark mass matrix related to the charged lepton mass matrix. When $120$ is introduced, the relation between the up-quark sector and the Dirac neutrino sector is lost because these two sectors receive contributions from different components of $120$.\
[(ii) $SU(4) \times SU(2)_{L} \times SU(2)_{R}$ breaking chain:]{}\
Under $SU(4) \times SU(2)_{L} \times SU(2)_{R}$, the relevant $SO(10)$ representations have the following decomposition $$\begin{aligned}
16 & = & (4,2,1) + (\overline{4},1,2)\\
10 & = & (6,1,1) + (1,2,2)\\
120 & = & (15,2,2) + (6,3,1) + (6,1,3) + (1,2,2) + (10,1,1) + (\overline{10},1,1)\\
126 & = & (10,1,3) + (\overline{10},3,1) + (15,2,2) + (6,1,1)\end{aligned}$$ where the components $(15,2,2)$ and $(1,2,2)$ both contain a pair of the $Y = \pm 1$ $SU(2)_{L}$ Higgs doublets, whose neutral components give masses to the fermions. The component $(10,1,3)$ contained in $\overline{126}$ gives masses to the RH neutrinos through the coupling $(16_{i})(16_{j})(\overline{126}_{H})$. The LH neutrino Majorana masses are generated due to the $(\overline{10},3,1)$ component of $\overline{126}$. As these $SU(2)_{L}$ doublets acquire VEV’s, the following mass matrices are generated, $$\begin{aligned}
\label{so10mass}
M_{u} & = & \mathcal{Y}^{10}_{ab} \left< 10^{+} \right>
+ \mathcal{Y}^{120}_{ab} ( \left< 120^{+} \right> +
\frac{1}{3} \left< 120^{'+} \right> )
+ \frac{1}{3} \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{126}^{+} \right>
\equiv Y_{u} v_{u}
\\
M_{d} & = & \mathcal{Y}^{10}_{ab} \left< 10^{-} \right>
+ \mathcal{Y}^{120}_{ab} (- \left< 120^{-} \right>
+\frac{1}{3} \left< 120^{'-} \right> )
-\frac{1}{3} \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{126}^{-} \right>
\equiv Y_{d} v_{d}
\qquad \\
M_{e} & = & \mathcal{Y}^{10}_{ab} \left< 10^{-} \right>
+ \mathcal{Y}^{120}_{ab} (-\left< 120^{-} \right>
-\left< 120^{'-} \right>)
+ \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{126}^{-} \right>
\equiv Y_{e} v_{d}
\qquad \\
M_{\nu_{LR}} & = & \mathcal{Y}^{10}_{ab} \left< 10^{+} \right>
+ \mathcal{Y}^{120}_{ab} (\left< 120^{+} \right>
-\left< 120^{'+} \right>)
+ \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{126}^{+}\right>
\equiv Y_{\nu_{LR}} v_{u}.\end{aligned}$$ Note that a Clebsch-Gordon coefficient $(-3)$ is generated in the lepton sectors when the $SU(4) \times SU(2)_{L} \times SU(2)_{R}$ components $(15,2,2)$ are involved in the Yukawa couplings. The neutrino Majorana mass matrices are given by $$\begin{aligned}
M_{\nu,RR} & = & \mathcal{Y}^{\overline{126}}_{ab} \left< \overline{126}^{'0}
\right>
\\
M_{\nu,LL} & = & \mathcal{Y}^{\overline{126}}_{ab} \left<
\overline{126}^{'+} \right>\label{typeII} \end{aligned}$$ where various VEVs are those of the neutral components of $SO(10)$ representations as indicated in Table \[notation\].
\
\
If $SO(10)$ breaks down through the $SU(5)$ breaking chain, the Higgs fields needed are in the $(16 \oplus 16)$, $45$ and $54$-dim representaions. A lop-sided texture can arise, when non-renormalizable operators involing the $16$’s are utilized to generate fermion masses. The $16$-dim Higgs fields are needed in this case to break $SO(10)$ down to $SU(5)$. A lop-sided texture is generated by the operator[@Albright:1998vf] $$\label{lopsided}
\lambda (16_{i}16_{H_{1}})(16_{j}16_{H_{2}}).$$ If $16_{H_{1}}$ acquires a VEV along the $SU(5)$ singlet direction which breaks $SO(10)$ down to $SU(5)$, and $16_{H_{2}}$ acquires a VEV along the $\overline{5}$ direction of $SU(5)$ which breaks the electroweak symmetry $SU(2) \times U(1)$, we obtain $$(\overline{5}_{i})(10_{j})<1_{H_{1}}> <\overline{5}_{H_{2}}>.$$ Inside the first parenthesis in Eq.(\[lopsided\]), the two $16$’s contract to from a $\overline{5}$ of $SU(5)$, while inside the second parenthesis, the two $16$’s contract to form a $5$ of $SU(5)$. This contraction arises by integrating out a pair of $5$ and $\overline{5}$ of $SU(5)$ from the $10$’s of $SO(10)$, as shown in Fig.\[lopsideddiagram\].
Because the $\overline{5}$ contains the $SU(2)$ lepton doublet and the singlet down-type quarks, the resulting mass terms $$\lambda (d_{L,i}^{c}d_{L,j} + e_{L,i}e_{L,j}^{c}) v_{d}$$ are related by $M_{d} = M_{e}^{T}$ and the lop-sided mass texture arises, $$(\overline{d}_{R,2} \; \overline{d}_{R,3})
\left(\begin{array}{cc}
0 & \lambda\\
0 & 0
\end{array}\right)
\left(\begin{array}{c}
d_{L,2}\\
d_{L,3}
\end{array}\right)v_{d}
+
(\overline{e}_{R,2} \; \overline{e}_{R,3})
\left(\begin{array}{cc}
0 & 0\\
\lambda & 0
\end{array}\right)
\left(\begin{array}{c}
e_{L,2}\\
e_{L,3}
\end{array}\right)v_{d},$$ if $(i,j)$ is chosen to be $(2,3)$. The $(33)$ entry which is expected to be of order $\mathcal{O}(1)$ is generated by tree level diagram involving a $10$. The $(23)$ and $(32)$ entries of $M_{d}$ and $M_{e}$ also receive contributions from other non-renormalizable operators, for example, $16_{i} 16_{j} 45_{H} 10_{H}$. If $\lambda$ is of order $\mathcal{O}(1)$ and other contributions to the $(23)$ and $(32)$ entries are much smaller than one, a large mixing angle is then generated in the right-handed down quark sector, and the left-handed charged lepton sector, while the corresponding mixing angle in the left-handed down quark sector is small. This thus provides a way to explain the large mixing angle in atmospheric neutrinos while the quark mixing $V_{cb}$ is small.
As we have seen above, when additional matter fields are introduced into the model and non-renormalizable operators that generate fermion masses are taken into account, as in the case of the Froggatt-Nielsen mechanism, other Higgs representations can play a role in the mass generation. An interesting case is the $45$-dimensional Higgs representation which has Yukawa couplings to a $16$- and a $10$-dim matter fields. These non-renormalizable operators can be expressed generically by[@Anderson:1993fe] $$\mathcal{O}_{ij} = 16_{i} \cdot \mathcal{O}_{n} \cdot 10 \cdot \mathcal{O}_{m}
\cdot 16_{j}$$ where the operator $\mathcal{O}_{n}$ is given by $$\mathcal{O}_{n} =
\frac{M_{G}^{p} \; 45_{p+1} ... 45_{q}}{M_{Pl}^{q} \; 45_{X}^{n-q}}, \qquad
n, \; p, \; q = {\mbox integer}.$$ Here $M_{G}$ and $M_{Pl}$ refer to the GUT scale and the Planck scale respectively. The $45$-dimensional representation Higgs can acquire VEV along the following four directions: $X, Y, B-L, T_{3R}$, where $X$ and $Y$ are defined as $$\begin{aligned}
X & = & -(\Sigma_{12} + \Sigma_{34} + \Sigma_{56} + \Sigma_{78} + \Sigma_{9 \; 10})
\\
Y & = & y \otimes \zeta
\nonumber\\
& & {\mbox where} \qquad
\zeta = \left(\begin{array}{cc}
0 & -i\\
i & 0
\end{array}\right), \qquad
y = diag(1/3,1/3,1/3,-1/2,-1/2).\end{aligned}$$ A systematical way to search for these effective operators is discussed in Ref. 95. The coefficient $-3$ is obtained in this case when the $45$-dimensional Higgs acquires a VEV along the $B-L$ direction. We summarize in Table. \[CG\] all possible CG coefficients for the four possible directions of $<45>$.
Automatic R-parity conservation and $\overline{126}$ v.s. $(16 \times 16)$ for neutrino masses
----------------------------------------------------------------------------------------------
One of the salient features of $SO(10)$ is that in some classes of models, R-parity is conserved automatically. Group theoretically, the congruence number is defined as follows: For an irreducible representation of $SO(2n)$ whose Dynkin index reads $(a_{1} \; a_{2} \; ... a_{n})$, this representation has congruence number $$\begin{aligned}
(c_{1},c_{2}) & \equiv & (a_{n-1}+a_{n},2a_{1}+2a_{3} + ... +
2a_{n-2}+(n-2)a_{n-1}+na_{n})
\nonumber\\
&& \mbox{mod} \; (2,4) \qquad \mbox{for} \; n = \; \mbox{odd}.\end{aligned}$$ There are four possible classes of $(c_{1},c_{2})$: $(0,0), (0,2), (1,1),
(1,3)$. The first two classes are tensor-like while the later two classes are spinor-like which is troublesome. To see how $(c_{1},c_{2})$ relates to R-parity, alternatively we can define, in $SO(10)$, $$c = 3 (B-L) \; \mbox{mod} \; 4$$ and $c = \mbox{Max} \; (c_{1},c_{2})$. It has been shown that if all the Higgs representations that break $SO(10)$ down to the SM are chosen to have congruence number $c = 0 \; \mbox{or} \; 2$, then R-parity is preserved at all energies.[@Mohapatra:su; @Martin:1992mq] Representations having $c=0$ are: $45,\; 54, \; 210, ...$; those having $c=2$ are: $10, \; 126, \; \overline{126}, ...$. Note that the spinor representations $16$ and $\overline{16}$ have $c=1$ and $3$ respectively.
Some models avoid the use of $\overline{126}$-dim Higgses by introducing non-renormalizable operators of the form $\frac{1}{M} \psi_{a}\psi_{b}(\overline{16}_{H})(\overline{16}_{H})$ instead of a renormalizable $\psi_{a}\psi_{b} \overline{126}_{H}$. Models utilizing the spinor representation $16$ to construct the neutrino mass operators generally have R-parity broken at some high energy scale. Such models also appear to be less constrained due to the inclusion of non-renormalizable operators. Also, a discrete symmetry, the R-parity symmetry, must be imposed by hand to avoid dangerous dim-$4$ baryon number violating operators in the effective potential at low energies which otherwise could lead to fast proton decay rate. Utilizing $\overline{126}$-dim representation of Higgses has the advantage that R-parity symmetry is automatic. The $\overline{126}$ representation has been used in model building before.[@Lee:1994je; @Brahmachari:1997cq; @Aulakh:1999cd; @Aulakh:2000sn; @Chen:2000fp; @Chen:2001pr; @Chen:2002pa; @Nath:2001uw; @Nath:2001yj] It is to be noted that the contribution of the $\overline{126}$-dimensional representation to the $\beta$-function makes the model nonperturbative (with the onset of the Landau pole) above the unification scale $M_{GUT}$. One could view these models as an effective theories valid below this scale where coupling constants are perturbative. One may also argue against the use of $\overline{126}$ with the fact that it is not possible to obtain such a large representation from string theory. It has been shown that in heterotic string theory it is not possible to get $126$ of $SO(10)$ up to Kac-Moody level-$5$.[@Dienes:1996wx] Nevertheless, no such constraints have been found in other types of string theories.[@Dienes:comm]
Some Related Issues
-------------------
### Proton Decay
One of the signatures of any GUT model is baryon number violating processes. These processes include proton decay ($\Delta(B-L)=0$), $N-\overline{N}$ oscillation ($\Delta(B-L)\ne0$), etc. The theoretical prediction for oscillation time in $N-\overline{N}$ oscillation can naturally satisfy the experimental lower limit $\tau_{N-\overline{N}} \ge 0.86 \times 10^{8}$ sec,[@Baldo-Ceolin:1994jz] if a high $B-L$ breaking scale is assumed. On the other hand, the non-observation of proton decay has put many GUT models under siege. There are three kinds of operators leading to proton decays in SUSY GUT’s:\
[(i) Dimension-$6$ operators:]{}
[As we have mentioned previously, the extra gauge bosons in GUT models, the $X$ and $Y$ gauge bosons, can lead to proton decay. The terms in the Lagrangian containing the $X$ $Y$ gauge bosons are $$\begin{aligned}
\mathcal{L}_{X,Y} & = &
\frac{ig_{X}}{\sqrt{2}}X_{\mu,i}
(\epsilon_{ijk}\overline{u}^{c}_{kL}\gamma_{\mu}u_{jL}
+ \overline{d}_{i}\gamma_{\mu}e^{+}) \nonumber\\
& & + \frac{ig_{Y}}{\sqrt{2}}Y_{\mu,i}
(\epsilon_{ijk}\overline{u}^{c}_{kL}\gamma_{\mu}d_{jL}
- \overline{u}_{iL}\gamma_{\mu}e^{+}_{L}
+ \overline{d}_{iR}\gamma_{\mu}\nu^{c}_{R}) + h.c.\end{aligned}$$ and the following vertices are allowed $$\overline{X} \rightarrow uu, \; \overline{d}e^{+}, \qquad
\overline{Y} \rightarrow ud, \; \overline{u} e^{+}.$$ These can thus lead to proton decay via the dim-$6$ operators. Note that these type of operators exist in both non-SUSY and SUSY GUT’s. The dominant decay mode is $p \rightarrow e^{+}\pi^{0}$, and the decay amplitude associated with this mode is[@Mohapatra:1999vv] $$\mathcal{M}_{p \rightarrow e^{+} \pi^{0}} \simeq
4\pi \alpha_{GUT} / M_{GUT}^{2}$$ leading to a life-time of[@Mohapatra:1999vv] $$\tau_{p} \simeq
\frac{1}{\mathcal{M}_{p \rightarrow e^{+} \pi^{0}}^{2}m_{p}^{5}}
\simeq 4.5 \times 10^{29{+\atop-}0.7} (\frac{M_{GUT}}{2.1\times 10^{14}
GeV})^{4}$$ where $m_{p}$ is the mass of the proton. For $M_{GUT} \simeq 2 \times 10^{16} GeV$, we get $\tau_{p} \sim 4.5 \times
10^{37{+\atop-}0.7} \; \mbox{years}$ which is far above the current capability of SuperKamiokande experiments whose limit is $\sim 10^{34}$ years.]{}\
[(ii) Dimension-$5$ operators]{}
[In SUSY GUT’s, a new channel for proton decay is possible via the dim-$5$ operator through the exchange of the color triplet Higgsinos where $QQH$ and $QL\overline{H}$ via $H\overline{H}$ mixing generate an effective operator $$QQQL/M_{H}.$$ In order to suppress this operator, the masses of the color triplet Higgsinos must have superheavy masses. From the point of view of unification, we would like to have the spectrum of MSSM below $M_{GUT}$. This requires that all the Higgs fields, including the color triplet Higgsinos, to be very heavy, with masses of the order of $M_{GUT}$, except the pair of $SU(2)_{L}$ doublet Higgses remaining light which are then identified as the pair of Higgs doublets in MSSM. How to achieve such a mass splitting is referred to as the [*doublet-triplet splitting*]{} (DTS) and [*doublet-doublet splitting*]{} (DDS) problems. The dominant decay mode of these operator is $\tau(p\rightarrow K^{+}\tilde{\nu}) \sim
m_{\tilde{h}}^{2} \sim M_{GUT}^{2}$. Its decay amplitude is[@Mohapatra:1999vv] $$\mathcal{M}_{p\rightarrow K^{+}\tilde{\nu}} \simeq
\frac{h_{u}h_{d}}{M_{H}}
\frac{m_{gaugino}g_{GUT}^{2}}{16\pi^{2}M_{\tilde{Q}}^{2}}.$$ Dimopolous and Wilczek proposed a mechanism[@Dimopoulos:zu] to achieve such mass splittings using $<45_{H}>$ along the $(B-L)$ direction $$<45_{H}> = i\tau_{2}\otimes {\mbox diag}(a,a,a,0,0).$$ Chacko and Mohapatra[@Chacko:1998zn] found that with a complimentary VEV pattern to that of the Dimopoulos-Wilczek type, that is, $$<45_{H}> = i\tau_{2}\otimes {\mbox diag}(0,0,0,b,b)$$ the same goals can be achieved. A solution utilizing $126_{H}$ instead of $45_{H}$ to achieve the DTS and DDS was proposed by Lee and Mohapatra.[@Lee:1994je] Detailed calculations have shown, nevertheless, that even with these mechanisms in place, in order for the prediction of $\tau_{p}$ to be consistent with the experimental limit, the effective $M_{H}$ must be larger than $M_{GUT}$ by at least a factor of $10$. This in turn requires some couplings to be much larger than $1$ which is somewhat unnatural.]{}\
[(iii) Dimension-$4$ operators:]{}
[As we have seen, the dim-$4$ operators are forbidden if there is R-parity in the model. In SUSY $SU(5)$ one has to impose R-parity by hand, while in some $SO(10)$ models, R-parity is automatically conserved if certain type of Higgs fields are chosen to construct the model, as we have discussed in the previous section.]{}
### Baryogenesis
The three Sakharov conditions,[@Sakharov:1967dj] ([*i*]{}) baryon number ($B$) violating processes, ([*ii*]{}) $C$ and $CP$ violation, and ([*iii*]{}) the deviation from thermal equilibrium, for the generation of the cosmological matter anti-matter asymmetry can be naturally satisfied in the $SO(10)$ model. Due to the presence of sphaleron effects, the only chance for GUT baryogenesis to work is to produce an asymmetry in $B-L$ at a high scale. To see this, let us first write the baryon number $B$ as $$B=\frac{1}{2} (B+L) + \frac{1}{2} (B-L).$$ The electroweak sphaleron transitions rapidly erase the asymmetry $B+L$ as soon as the temperature drops down to about $10^{12} \; GeV$. Therefore, to have a non-vanishing baryonic asymmetry requires a non-vanishing $B-L$ asymmetry. The crucial point to note is that even though the $B+L$ asymmetry is erased by the sphaleron transitions, the orthogonal combination $B-L$ is left untouched, and it opens up the possibility that the baryonic asymmetry can be generated through leptogenesis.[@Fukugita:1986hr; @Luty:1992un; @Buchmuller:1996pa] The basic idea is that, since $B+L$ must vanish at all times due to the sphaleron transitions, the asymmetry in the lepton number will consequently be converted into the asymmetry in the baryon number (with an opposite sign). The primordial leptonic asymmetry is generated by the out-of-equilibrium decay of the heavy right-handed Majorana neutrinos and their scalar partners in the supersymmetric case. The relevant superpotential is $$W_{Leptogenesis} =
(Y_{e})_{ij} E_{i}^{c}L_{j}H_{1}
+ (Y_{\nu_{LR}})_{ij} N_{i}^{c}L_{j}H_{2}
+ \frac{1}{2} (M_{RR})_{ij} N_{i}^{c}N_{j}^{c}.$$ The heavy right-handed neutrinos and their scalar partners can decay through the following four decay modes: $$\begin{aligned}
N_{1} & \longrightarrow & \tilde{l} \quad + \quad \tilde{h^{c}}\\
N_{1} & \longrightarrow & l \quad + \quad H_{2}\\
\tilde{N_{1}^{c}} & \longrightarrow & \tilde{l} \quad + \quad H_{2}\\
\tilde{N_{1}^{c}} & \longrightarrow & l \quad + \quad \tilde{h^{c}}.\end{aligned}$$ The interference between the tree-level and one-loop diagrams thus gives rise to the CP asymmetry.
In the basis where both charged lepton Yukawa couplings and the right-handed neutrino mass matrix are diagonal, the amount of CP asymmetry due to the interference between the tree level and one-loop diagrams for [*each*]{} decay mode is given by[@Buchmuller:1996pa] $$\epsilon_{1} =
-\frac{1}{8\pi}
\frac{1}{(h_{\nu}h_{\nu}^{\dagger})_{11}}
\sum_{i=2,3} Im \{ (h_{\nu}h_{\nu}^{\dagger})_{1i}^{2} \}
f(\frac{M_{i}^{2}}{M_{1}^{2}})$$ where $$f(x)=\sqrt{x} \; [ \; \ln(\frac{1+x}{x}) + \frac{2}{x-1} \;],$$ and $h_{\nu}$ is the Dirac neutrino Yukawa matrix in the new basis. The right-handed Majorana neutrino mass matrix is diagonalized by $$P_{M} O_{R} M_{RR} O_{R}^{T} P_{M} = diag(M_{1},M_{2},M_{3})$$ where $M_{i}$’s are real and non-negative, and $P_{M}$ is the diagonal Majorana phase matrix. In terms of the diagonalization matrices and the original Dirac neutrino Yukawa coupling, $(h_{\nu}h_{\nu}^{\dagger})$ can then be rewritten as $$h_{\nu}h_{\nu}^{\dagger}
= P_{M_{RR}} O_{R}Y_{\nu_{LR}}U_{e_{L}}^{\dagger}U_{e_{L}}
Y_{\nu_{LR}}^{\dagger}O_{R}^{\dagger} P_{M_{RR}}^{-1}
=P_{M_{RR}} O_{R}Y_{\nu_{LR}}Y_{\nu_{LR}}^{\dagger}
O_{R}^{\dagger} P_{M_{RR}}^{-1}.$$ We see that the phases in the right-handed neutrino mass matrix $M_{RR}$ and the Dirac neutrino mass matrix $Y_{\nu_{LR}}$ are responsible for the CP asymmetry needed for the leptogenesis. For a hierarchical heavy right-handed neutrino mass spectrum, $M_{3} \gg M_{2} \gg M_{1}$, the argument of the function $f(x)$, $x \equiv \frac{M_{i}^{2}}{M_{1}^{2}}$, is much greater than $1$. We can then approximate $f(x)$ as $$f(x) = \sqrt{x} \;
[ \; (\frac{1}{x} - \frac{1}{2 x^{2}} + \ldots) +\frac{2}{x-1} \;]
\simeq \frac{3}{\sqrt{x}}
= 3 \frac{M_{1}}{M_{i}}.$$ The asymmetry $\epsilon_{1}$ can thus be rewritten as $$\label{epsilon1}
\epsilon_{1} \simeq -\frac{4}{8\pi}
\frac{1}{(h_{\nu}h_{\nu}^{\dagger})_{11}}
\sum_{i=2,3} Im \{ (h_{\nu}h_{\nu}^{\dagger})_{1i}^{2} \}
(\frac{3M_{1}}{M_{i}})$$ where the factor $4$ accounts for the fact that there are four decay modes. Using the fact that the mixing in $Y_{\nu_{LR}}$ is small and that $(Y_{\nu_{LR}})_{33}$ dominates other elements, we can further approximate $$\begin{aligned}
(h_{\nu}h_{\nu}^{\dagger})_{1i} & \simeq &
(P_{11}P_{ii}^{-1}) \vert y_{\nu_{\tau}} \vert^{2}
(O_{R})_{13}(O_{R}^{\ast})_{i3}
\\
(h_{\nu}h_{\nu}^{\dagger})_{11} & \simeq &
\vert y_{\nu_{\tau}} \vert^{2}
\vert (O_{R})_{13} \vert^{2}\end{aligned}$$ and $$\frac{Im \{ (h_{\nu}h_{\nu}^{\dagger})_{1i}^{2} \} }
{(h_{\nu}h_{\nu}^{\dagger})_{11}}
\simeq
\vert y_{\nu_{\tau}} \vert^{2}
Im \{ (P_{11}P_{1i}^{-1})^{2} \frac{(O_{R})_{13}}{(O_{R})_{13}^{\ast}}
(O_{R}^{\ast})_{i3}^{2} \}.$$ To have a large amount of CP asymmetry, $\epsilon_{1}$, thus requires that the hierarchy among the three right-handed neutrino masses cannot be too large (that is, $\frac{M_{1}}{M_{2,3}}$ cannot be too small), and that the imaginary part of $\{ (P_{11}P_{1i}^{-1})^{2} \frac{(O_{R})_{13}}{(O_{R})_{13}^{\ast}}
(O_{R}^{\ast})_{i3}^{2} \}$ together with the neutrino Dirac Yukawa couplings cannot be too small.
The amount of the lepton asymmetry generated is given by $$Y_{L} \equiv \frac{n_{L}-\overline{n}_{L}}{s} =
\kappa \frac{\epsilon_{1}}{g_{\ast}}.$$ Here $g_{\ast}$ is the number of relativistic degrees of freedom. For MSSM, it is $g_{\ast}=228.75$. The out-of-equilibrium decay of the heavy Majorana neutrinos requires the decay width of the lightest neutrino, $\Gamma_{1}$, smaller than the Hubble constant at the temperature of the decay. That is, $$r \equiv \frac{\Gamma_{1}}{H \vert_{T=M_{1}}}
= \frac{M_{pl}}{(1.7)(32\pi)\sqrt{g_{\ast}}}
\frac{( h_{\nu} h_{\nu}^{\dagger} )_{11}}{M_{1}}
< 1$$ where $M_{Pl}$ is the Planck scale taken to be $1.2 \times 10^{19} \; GeV$. In general, one can still have a sizable CP asymmetry remains even for $1 < r < 10$. The wash-out effects due to inverse decays and lepton number violating scattering processes together with the time evolution of the system is then accounted for by the factor $\kappa$. It is obtained by solving the Boltzmann equation for the system. An approximation is given by[@Kolb:vq] $$\begin{aligned}
10^{6} \le r: & \quad
\kappa = (0.1 r)^{1/2} e^{-(\frac{4}{3})(0.1r)^{1/4}} \quad &
( < 10^{-7} )\\
10 \le r \le 10^{6}: & \quad \kappa = \frac{0.3}{r (\ln r)^{0.6}} \quad &
(10^{-2} - 10^{-7})\\
0 \le r \le 10: & \quad \kappa = \frac{1}{2\sqrt{r^{2}+9}} \quad &
(10^{-1} - 10^{-2}) \end{aligned}$$ where inside the parentheses we give the order of magnitude of $\kappa$ for each corresponding $r$. We note that in order to have a small dilution factor, the lightest right-handed neutrino cannot be too light. The electroweak sphaleron effect will convert the lepton asymmetry $Y_{L}$ into baryon asymmetry $Y_{B}$, and they are related by $$Y_{B} \equiv \frac{n_{B}-\overline{n}_{B}}{s} = c Y_{B-L} = \frac{c}{c-1} Y_{L}$$ with $$c = \frac{8N_{F}+4N_{H}}{22N_{F}+13N_{H}}$$ where $N_{F}$ is the number of families and $N_{H}$ is the number of $SU(2)$ Higgs doublets. For the MSSM spectrum, $(N_{F},N_{H})=(3,2)$, we have the conversion factor $(\frac{c}{c-1}) \simeq -0.53$.
Models with only two RH neutrinos have been constructed by Frampton [*et al*]{} [@Frampton:2002qc] and by Raby[@Raby:2003ay] (see Sec. 3 and 6), which give rise to bi-large mixing pattern and a correlation between the sign of the baryon number asymmetry and the sign of the CP violation in neutrino oscillation. General analyses on the consistency between constraints from simplest $SO(10)$ models and leptogenesis can be found in Ref. 117, 118.
SUSY Breaking
-------------
SUSY breaking can be incorporated into models by including explicitly the soft SUSY breaking terms. Since the RGE’s for the Yukawa coupling constants and gauge coupling constants do not have any dependence on the soft breaking parameters up to two-loop level, the presence of these soft SUSY breaking terms does not affect the predictions for fermion masses and mixing angles. On the other hand, since the evolutions of soft SUSY breaking parameters does depend on the Yukawa coupling constants and gauge coupling constants, whether the EW symmetry is broken (that is, the mass-square of the light Higgs doublet is driven to be negative) may depend on the Yukawa sector, which can be used as a test of the validity of the model.
Even though SUSY breaking does not affect the running of the gauge coupling constants and that of the Yukawa coupling constants, it could have a large contribution to the threshold corrections, which is the subject of the next section.
Threshold Corrections
---------------------
When the RGE analysis is performed, we usually consider $\beta$ function coefficient as a constant for each coupling constant between the two relevant scales. This is done under the assumption that all the heavy modes decouple at the same scale, the symmetry breaking scale. Threshold corrections are the corrections due to the differences between the symmetry breaking scale and the masses of the heavy particles decoupled from the spectrum after symmetry breaking takes place. If all the heavy particles acquire masses exactly the same as the symmetry breaking scale, there are no threshold corrections. In practice, this is not the case. There are two possible sources of threshold corrections:
[(i) GUT scale threshold corrections:]{} Due to the presence of many large Higgs representations in $SO(10)$, the GUT scale threshold corrections could be large.
[(ii) SUSY threshold corrections:]{} Large threshold corrections to $m_{b}$, $\delta m_{b} / m_{b} \simeq -(0.15
\sim 0.2)$, are needed in most $SO(10)$ models in order to have a prediction for $m_{b}$ consistent with experiment. The dominant contributions are from the diagrams shown in Fig. \[threshold\].
They give a correction $$\delta m_{b} \; / \; m_{b} \simeq (\tan\beta/50) I$$ and $I$ is given by,[@Hall:1993gn; @Rattazzi:1995gk; @Babu:1999hn] $$I \simeq
\frac{50}{16\pi^{2}}\frac{\alpha_{G}}{\alpha_{2}}
\frac{\mu m_{\tilde{W}}}{m_{eff}^{2}}
[\frac{8}{3}\frac{\alpha_{3}}{\alpha_{G}}g_{3}^{2}
f(\frac{m_{\tilde{g}}^{2}}{m_{eff}^{2}})
-2\lambda_{t}^{2}f(\frac{\mu^{2}}{m_{eff}^{2}})]$$ where $f(x)=(1-x+x\ln x)/(1-x)^{2}$; $m_{\tilde{g}}$ and $\mu$ are the gluino mass and the $\mu$ term evaluated at the weak scale.; $m_{eff}^{2} \equiv \frac{1}{2}(m_{\tilde{b}}^{2}+ m_{\tilde{Q}}^{2})$ is the average of the squared masses of the $SU(2)$-singlet bottom squark and the $SU(2)$-doublet third generation squarks. A large soft SUSY breaking parameter space can give rise to such a correction. With the typical values $\alpha_{G}=0.75$ and $(\alpha_{3},\alpha_{2},y_{33}) \simeq (0.124,0.034,1)$ at $M_{weak}$, $\delta_{b}$ can then be approximated as, with the assumption $2 m_{1/2} = 2 \frac{\alpha_{G}}{\alpha_{w}} m_{\tilde{W}}$, $$I = (0.315) \; x \; t \; ( 0.69 f(t^2) - f(x^2) )$$ where $t=(\mu/m_{0})$ and $x = m_{1/2}/m_{0}$ with $m_{1/2}$ and $m_{0}$ being the gluino mass and the common scalar mass respectively in the constrained MSSM (CMSSM). With $(t,x)=(5.2,2)$ which are typical values in CMSSM,[@Castano:1993ri] a correction $(\delta m_{b} / m_{b} ) = -0.15$ for $\tan\beta=10$ is obtained.
$SO(10)$ Models with Texture Assumptions {#modelt}
========================================
In what follows we discuss a few selected $SO(10)$ models combined with some texture ansatz for the mass matrices.
Buchmuller and Wyler
--------------------
Buchmuller and Wyler[@Buchmuller:2001dc] assume symmetric mass textures for the up- and down-type quarks $$M_{u,d} \sim \left(\begin{array}{ccc}
0 & \epsilon^{3}e^{i\phi} & 0\\
\epsilon^{3}e^{i\phi} & \rho \epsilon^{2} & \eta \epsilon^{2}\\
0 & \eta \epsilon^{2} & e^{i\psi}
\end{array}\right).$$ Here $\rho = |\rho|e^{i\alpha}$ and $\eta = |\eta|e^{i\beta}$ are complex parameters of $\mathcal{O}(1)$. Using the $SO(10)$ relations, they have $$m_{\nu}^{Dirac} = m_{u}, \qquad m_{e} = m_{d}$$ [*assuming*]{} the incorrect mass relations in the lighter two generations are lifted when higher dimensional Higgs representations are introduced. The right-handed neutrino Majorana mass matrix is generated by $\overline{126}_{H}$, and is assumed to have the following form $$M_{\nu,RR} = \left(\begin{array}{ccc}
0 & M_{12} & 0\\
M_{12} & M_{22} & M_{23}\\
0 & M_{23} & M_{33}
\end{array}\right).$$ With the relations $$M_{12} : M_{22} : M_{33} = \epsilon^{5}:\epsilon^{4}:1, \qquad
M_{23} \sim M_{22},$$ the resulting effective neutrino mass matrix has the following form $$M_{\nu}^{eff} = \left(\begin{array}{ccc}
0 & \epsilon e^{2i\phi} & 0\\
\epsilon e^{2i\phi} & -\sigma e^{2i\phi}+2\rho e^{i\phi} & \eta e^{i\phi}\\
0 & \eta e^{i\phi} & e^{2i\phi}
\end{array}\right) \cdot \frac{v_{1}^{2}}{M_{3}}$$ where $M_{3}$ is the heaviest eigenvalue of $M_{\nu, RR}$ which is of the same order as $M_{33}$, and $\sigma$ is defined as $\sigma \epsilon^{4} = M_{22}/M_{3}$. The parameters $\eta, \; \rho$ and $\sigma$ are all of $\mathcal{O}(1)$, and $\eta$ and $\rho$ are determined using the quark masses. The effective neutrino masses form the following pattern, $$m_{\nu_{1}} : m_{\nu_{2}} : m_{\nu_{3}} = \epsilon : \epsilon : 1$$ with $\epsilon \sim 0.1$, we have $(m_{\nu_{2}}^{2}-m_{\nu_{1}}^{2})/(m_{\nu_{3}}^{2}-m_{\nu_{2}}^{2}) \sim 10^{-2}$, which is consistent with the LMA solution. Nevertheless, it is not clear whether the predicted solar angle is consistent with experiment or not. The prediction for $U_{e\nu_{3}}$ is of order $\mathcal{O}(\epsilon)$.
An interesting feature of this model is that, with $\epsilon \sim 0.1$, the scale of $M_{1}$ is about $10^{9} \; GeV$. The amount of baryonic asymmetry which is very sensitive to this scale is given by $$Y_{B} \sim -\kappa \; sign(\sigma) \; \sin(\phi-\alpha) \times 10^{-9}.$$ The result has the promise to be consistent with the observed value, once the parameters $\kappa$, $\sigma$, $\phi$ and $\alpha$ are fixed.
Matsuda, Fukuyama and Nishiura
------------------------------
Matsuda, Fukuyama and Nishiura[@Matsuda:1999yx] proposed Hermitian textures with four zeros in the context of $SO(10)$ $$\begin{aligned}
\label{4zero}
M_{u,\nu} & = & \left(\begin{array}{ccc}
0 & A_{u,\nu} & 0\\
A_{u,\nu} & B_{u,\nu} & C_{u,\nu}\\
0 & C_{u,\nu} & D_{u,\nu}
\end{array}\right)
\\
M_{d,e} & = & \left(\begin{array}{ccc}
0 & A_{d,e}e^{i\alpha_{12}^{d,e}} & 0\\
A_{d,e}e^{-i\alpha_{12}^{d,e}} & B_{d,e} & C_{d,e}e^{i\alpha_{23}^{d,e}}\\
0 & C_{d,e} e^{-i\alpha_{23}^{d,e}} & D_{d,e}
\end{array}\right).\end{aligned}$$ For a general matrix of the form $$M = \left(\begin{array}{ccc}
0 & A & 0\\
A & B & C\\
0 & C & D
\end{array}\right),$$ one can relate its three eigenvalues, $m_{1}$, $m_{2}$, and $m_{3}$, to the matrix elements by $$\begin{aligned}
DA^{2} & = & -m_{1}m_{2}m_{3}\\
BD-A^{2}-C^{2} & = & m_{1}m_{2}+m_{2}m_{3}+m_{3}m_{1}\\
B+D & = & m_{1}+m_{2}+m_{3}.\end{aligned}$$ Because of the observed fermion mass hierarchy, $|m_{3}| \gg |m_{2}| \gg |m_{1}|$, it is a good approximation to write $B=m_{2}$ and $D=m_{3}-m_{1}$ and express the mass matrix in terms of its three eigenvalues as $$M \simeq \left(\begin{array}{ccc}
0 & \sqrt{-m_{1}m_{2}} & 0\\
\sqrt{-m_{1}m_{2}} & m_{2} & \sqrt{-m_{1}m_{3}}\\
0 & \sqrt{-m_{1}m_{3}} & m_{3}+m_{1}
\end{array}\right).$$ To put this idea to work in the context of $SO(10)$, we first note that the set of equations given in Eq.(160)-(163) can be re-written as, assuming $<120>$ is small and contribute to $M_{e}$ and $M_{d}$ only, $$\begin{aligned}
M_{u} & = & Y^{10} <10> + \epsilon Y^{126} <\overline{126}>
\\
M_{d} & = & \alpha Y^{10} <10> + Y^{126} <\overline{126}>
+ Y^{120} <120>
\nonumber\\
& = & \alpha M_{u} - (\alpha\epsilon-1) Y^{126} <\overline{126}>
+ Y^{120} <120>
\\
r M_{e} & = & \alpha Y^{10} <10> -3 Y^{126} <\overline{126}>
+ \delta Y^{120} <120>
\nonumber\\
& = & \alpha M_{u} - (\alpha\epsilon+3) Y^{126} <\overline{126}>
+ \delta Y^{120} <120>
\\
r^{'} M_{\nu}^{Dirac} & = & Y^{10} <10> - 3 \epsilon Y^{126} <\overline{126}>
\\
s M_{\nu,LL} & = & \beta Y^{126} <\overline{126}>
\\
s^{'} M_{\nu,RR} & = & \gamma Y^{126} <\overline{126}>\end{aligned}$$ where parameters $\alpha,\beta, \gamma$ are ratios of SM Higgs doublet VEVs from different $SO(10)$ representations. The symmetric (anti-symmetric) matrices $Y^{10,126}$ ($Y^{120}$) can be expressed in terms of the symmetric (anti-symmetric) part of the mass matrices $M_{u,d,e,\nu}$ as $$\begin{aligned}
(1-\alpha \epsilon) Y^{10} <10> & = & (M_{u})_{s} - \epsilon (M_{d})_{s}\\
Y^{126} <\overline{126}> & = & \frac{1}{4} (M_{d})_{s} - \frac{1}{4} r (M_{e})_{s}\\
Y^{120} <120> & = & (M_{d})_{a}\end{aligned}$$ where the subscripts $s$ and $a$ refer to symmetric part and anti-symmetric part, respectively. One can then solve for the elements in matrices $Y$’s in terms of quark masses. As the simple approximations, $B=m_{2}$ and $D=m_{3}+m_{1}$, work well for quark masses and CKM matrix elements, in order to obtain viable neutrino masses and mixing angles, a deviation must be made in the charged lepton mass matrix, $B_{e} = m_{\mu} (1+\xi)$ and $D_{e} =m_{\tau}+m_{e}-\xi m_{\mu}$. With $\xi \sim 0.01$, maximal $\nu_{\mu}-\nu_{\tau}$ mixing angle and the LMA solution can be accommodated. In this model, the allowed region for the leptonic CP violating Dirac phase can be obtained.
Bando and Obara
---------------
Bando and Obara[@Bando:2002tu; @Bando:2003ei] pursue along the line of Matsuda, Fukuyama and Nishiura[@Matsuda:1999yx] to analyze mass matrices of the type given in Eq.(\[4zero\]), and have a detailed analysis on all possible combinations of contributions from either $<10_{H}>$ or $<\overline{126}_{H}>$ for each non-vanishing entry. In other words, all possible ways the CG factor $(-3)$ due to $<\overline{126}_{H}>$ can appear in the neutrino Dirac mass matrix $$\left(\begin{array}{ccc}
0 & \ast a_{\nu} & 0\\
\ast a_{\nu} & \ast b_{\nu} & \ast c_{\nu}\\
0 & \ast c_{\nu} & 1
\end{array}\right)$$ where $\ast$ is either $1$ or $-3$ depending upon whether the coupling is due to $<10_{H}>$ or $<\overline{126}_{H}>$. They found the following texture has best agreement with experiments $$M_{u} = \left(\begin{array}{ccc}
0 & 126 & 0\\
126 & 10 & 10\\
0 & 10 & 126
\end{array}\right)
=\left(\begin{array}{ccc}
0 & a_{u} & 0\\
a_{u} & b_{u} & c_{u}\\
0 & c_{u} & 1
\end{array}\right).$$ In this case, the Dirac neutrino mass matrix is $$M_{\nu}^{Dirac} = \left(\begin{array}{ccc}
0 & -3a_{u} & 0\\
-3a_{u} & b_{u} & c_{u}\\
0 & c_{u} & -3
\end{array}\right) m_{t}$$ and the right-handed neutrino Majorana mass matrix is generated by the coupling to $\overline{126}$ Higgs representation, and is of the form $$M_{\nu,RR} = \left(
\begin{array}{ccc}
0 & r & 0\\
r & 0 & 0\\
0 & 0 & 1
\end{array}\right) m_{R}.$$ The effective neutrino mass matrix is thus given by $$M_{\nu}^{eff} =
\left(\begin{array}{ccc}
0 & \frac{a^{2}}{r} & 0\\
\frac{a^{2}}{r} & \frac{2ab}{r}+c^{2} & c(\frac{a}{r}+1)\\
0 & c(\frac{a}{r}+1) & d^{2}
\end{array}\right)
\frac{m_{t}^{2}}{m_{R}}$$ where $a=-3 a_{u}$, $b=b_{u}$, $c=-3c_{u}$, and $d=-3$. The typical predictions for this type of mass matrices are $$\begin{aligned}
\sin^{2} 2\theta_{23} & \sim & 0.98 - 1\\
\tan^{2} \theta_{12} & \sim & 0.29 - 0.46\\
|\theta_{13}| & \sim & 0.037 - 0.038\\
|m_{\nu_{3}}| & \sim & 0.053-0.059 \; eV\\
|m_{\nu_{2}}| & \sim & 0.003-0.008 \; eV\\
|m_{\nu_{1}}| & \sim & 0.0006-0.001 \; eV.\end{aligned}$$
$SO(10)$ Models with Family Symmetry in $4$-dimensions {#models}
======================================================
A natural framework to accommodate small neutrino masses is a grand unified theory based on SO(10) in which a right-handed neutrino in each family is predicted and the see-saw mechanism can be implemented naturally. Many other models based on $SO(10)$ besides those mentioned in Section 5 have been proposed to accommodate the observed fermion masses and mixing angles. Here we concentrate only on models which utilize family symmetry. We classify these models according to their intermediate symmetry below $SO(10)$ breaking scale and the family symmetry. Different symmetry breaking pattern of $SO(10)$ have different mass relations among the quark and lepton sectors, resulting in different way to generate large leptonic mixing angles.
Models with Symmetric Mass Textures
-----------------------------------
Symmetric mass textures naturally arise if the $SO(10)$ is broken down to the SM gauge group through the left-right symmetry breaking chain. Due to its symmetric nature, this type of models tend to be more predictive compared to models with lop-sided/asymmetric mass texture.\
### Chen and Mahanthappa
The model proposed by Chen and Mahanthappa[@Chen:2000fp; @Chen:2001pr; @Chen:2002pa] has $SU(2)$ family symmetry. Because $SO(10)$ breaks down through the left-right symmetry breaking chain, symmetric mass matrices arise. The field content of this model is given by, in terms of $SO(10) \times SU(2)$ quantum numbers, $$\begin{aligned}
{\mbox Matter fields:} & 1(16,2), \; 1(16,1)
\nonumber\\
{\mbox Higgs fields:} & 5(10,1), \; 3(\overline{126},1)
\nonumber\\
{\mbox Flavon fields:} & 3(1,2), \; 3(1,3).
\nonumber\end{aligned}$$ After the symmetry is broken, the following mass matrices are generated $$M_{u,\nu_{LR}} =
\left( \begin{array}{ccc}
0 & 0 & \left<10_{2}^{+} \right> \epsilon'\\
0 & \left<10_{4}^{+} \right> \epsilon & \left<10_{3}^{+} \right> \epsilon \\
\left<10_{2}^{+} \right> \epsilon' & \left<10_{3}^{+} \right> \epsilon &
\left<10_{1}^{+} \right>
\end{array} \right)
=
\left( \begin{array}{ccc}
0 & 0 & r_{2} \epsilon'\\
0 & r_{4} \epsilon & \epsilon \\
r_{2} \epsilon' & \epsilon & 1
\end{array} \right) M_{U}$$ $$M_{d,e} =
\left(\begin{array}{ccc}
0 & \left<10_{5}^{-} \right> \epsilon' & 0 \\
\left<10_{5}^{-} \right> \epsilon' & (1,-3)\left<\overline{126}^{-} \right>
\epsilon & 0\\ 0 & 0 & \left<10_{1}^{-} \right>
\end{array} \right)
= \left(\begin{array}{ccc}
0 & \epsilon' & 0 \\
\epsilon' & (1,-3) p \epsilon & 0\\
0 & 0 & 1
\end{array} \right) M_{D}$$ where $$\label{eq:higgsvev}
M_{U} \equiv \left<10_{1}^{+} \right>,
\qquad
M_{D} \equiv \left<10_{1}^{-} \right>$$ $$r_{2} \equiv \left<10_{2}^{+} \right> / \left<10_{1}^{+} \right>,
\quad
r_{4} \equiv \left<10_{4}^{+} \right> / \left<10_{1}^{+} \right>
\quad
p \equiv \left<\overline{126}^{-} \right> / \left<10_{1}^{-} \right>.$$ $\epsilon M$ and $\epsilon^{'} M$ are the VEV’s accompanying the flavon fields. The mass hierarchy arises due to the Froggatt-Nielsen mechanism. The right-handed neutrino masses are generated due to $\overline{126}_{H}$, thus R-parity is preserved at all energy scales, and no additional proton decay modes are allowed, which is to be contrasted to the case when $\overline{16}_{H}$’s are implemented to generate right-handed neutrino Majorana masses (see the model of Babu, Pati and Wilczek discussed in the next section.) The right-handed neutrino mass matrix is given by $$M_{\nu_{RR}} =
\left( \begin{array}{ccc}
0 & 0 & \left<\overline{126}_{2}^{'0} \right> \delta_{1}\\
0 & \left<\overline{126}_{2}^{'0} \right> \delta_{2}
& \left<\overline{126}_{2}^{'0} \right> \delta_{3} \\
\left<\overline{126}_{2}^{'0} \right> \delta_{1}
& \left<\overline{126}_{2}^{'0} \right> \delta_{3} &
\left<\overline{126}_{1}^{'0} \right> \end{array} \right)
=
\left( \begin{array}{ccc}
0 & 0 & \delta_{1}\\
0 & \delta_{2} & \delta_{3} \\
\delta_{1} & \delta_{3} & 1
\end{array} \right) M_{R}$$ with $M_{R} \equiv \left<\overline{126}^{'0}_{1}\right>$. The effective neutrino mass matrix is of the following form $$\label{mll}
M_{\nu_{LL}}^{eff} = \left(
\begin{array}{ccc}
0 & 0 & t\\
0 & 1 & 1+t^{3/2}\\
t & 1+t^{3/2} & 1
\end{array}
\right)\frac{d^{2}v_{u}^{2}}{M_{R}}$$ where $t < 1$. This model can accommodate the LMA solution, in addition to LOW and “Just SO” VO solutions. Its prediction for $U_{e\nu3}$ is about $0.15$. With $11$ input parameters, this model predicts $22(+9)$ masses, mixing angles, and CP violating phases for quarks and leptons (and right-handed neutrinos).
We note that in this model, the mass matrices, $M_{\nu_{LR}}$, $M_{\nu_{RR}}$ and $M_{\nu_{LL}^{eff}}$, have identical form. In other words, the texture considered in Eq.(\[cmtexture\]) is invariant under the see-saw mechanism. The form invariance also occurs in a model of neutrino mixing[@Fritzsch:1999ee] which uses different texture.
Models with Lop-sided/Asymmetric Mass Textures
----------------------------------------------
Models having $SU(5)$ as the intermediate symmetry have lop-sided Yukawa matrices. This is due to the $SU(5)$ relation, $M_{d} = M_{e}^{T}$. This opens up the possibility of large leptonic mixing angles due to the large mixing angle in the charged lepton mixing matrix.\
### Babu, Pati and Wilczek
The model proposed by Babu, Pati and Wilczek[@Babu:1998wi] utilizes the Abelian $U(1)$ as its family symmetry. Because it is based on dimension-$5$ operators to generate fermion masses, its field content, in terms of $SO(10)$ quantum numbers, is somewhat simple $$\begin{aligned}
{\mbox Matter fields:} & 16_{1}, \; 16_{2}, \; 16_{3}
\nonumber\\
{\mbox Higgs fields:} & 1(10),\; 1 (16+\overline{16}), \; 45.
\nonumber\end{aligned}$$ The mass matrices generated are given by $$\begin{aligned}
M_{u} & = & \left(\begin{array}{ccc}
0 & \epsilon^{'} & 0\\
-\epsilon^{'} & 0 & \epsilon+\sigma\\
0 & -\epsilon+\sigma & 1
\end{array}\right) \cdot m_{u}
\\
M_{d} & = & \left(\begin{array}{ccc}
0 & \epsilon^{'} + \eta^{'} & 0\\
-\epsilon^{'}+\eta^{'} & 0 & \epsilon+\eta\\
0 & -\epsilon+\eta & 1
\end{array}\right) \cdot m_{d}
\\
M_{\nu_{LR}} & = & \left(\begin{array}{ccc}
0 & -3\epsilon^{'} & 0\\
3\epsilon^{'} & 0 & -3\epsilon+\sigma\\
0 & 3\epsilon+\sigma & 1
\end{array}\right) \cdot m_{u}
\\
M_{e} & = & \left(\begin{array}{ccc}
0 & -3\epsilon^{'} + \eta^{'} & 0\\
3\epsilon^{'}+\eta^{'} & 0 & -3\epsilon+\eta\\
0 & 3\epsilon+\eta & 1
\end{array}\right) \cdot m_{d}.\end{aligned}$$ The right-handed neutrino Majorana mass matrix is generated by the effective operator, $\frac{1}{M}16_{i}16_{j}\overline{16}_{H}\overline{16}_{H}$ and is given by $$M_{\nu_{RR}} = \left(\begin{array}{ccc}
x & 0 & z\\
0 & 0 & y\\
z & y & 1
\end{array}\right) \cdot M_{R}$$ and the resulting effective neutrino mass matrix is given by $$\begin{aligned}
& M_{\nu}^{eff} = \hspace{10cm}\nonumber\\
& \left(\begin{array}{ccc}
{\scriptstyle 9\epsilon^{'2}(x-z^{2})} \; &
{\scriptstyle 3\epsilon^{'}y(-3\epsilon^{'}z+(-3\epsilon+\sigma)x)} \; &
{\scriptstyle 3\epsilon(xy-(3\epsilon+\sigma)(x-z^{2})}
\\
{\scriptstyle 3\epsilon^{'}y(-3\epsilon^{'}z+(-3\epsilon+\sigma)x)} \; &
{\scriptstyle -9\epsilon^{'2}y^{2}} \; &
{\scriptstyle (3\epsilon+\sigma)y(3\epsilon^{'}z-(-3\epsilon+\sigma)x)}
\\
{\scriptstyle 3\epsilon(xy-(3\epsilon+\sigma)(x-z^{2})} \; &
{\scriptstyle (3\epsilon+\sigma)y(3\epsilon^{'}z-(-3\epsilon+\sigma)x)} \; &
{\scriptstyle (3\epsilon+\sigma)(-2xy+(3\epsilon+\sigma)(x-z^{2}))}
\end{array}\right) \cdot m_{\nu}^{\mbox{eff}}.\nonumber\\\end{aligned}$$ The large atmospheric mixing comes from the effective neutrino mixing matrix, by choosing the value of parameter $y$. As $y$ also gives rise to small $V_{cb}$ value, the smallness of $V_{cb}$ and the maximality of the atmospheric mixing angle are thus related. This model can only accommodate SMA solution for the solar neutrinos. The LMA solution can be obtained if an intrinsic LH neutrino Majorana mass term arising from a dim-7 operator is included, assuming it dominates over the regular Type I seesaw term.[@Pati:2002ig; @Pati:2002pe] A characteristic of this model is the presence of a new prominent dim-$5$ proton decay mode, $p \rightarrow \mu^{+}K^{0}$, in addition to the $p \rightarrow \overline{\nu}K^{+}$ mode. This is a consequence of utilizing the $16_{H}$ to generate neutrino masses.[@Babu:1997js] In this model, the $\theta_{13}$ angle is predicted to be about $5.5 \times 10^{-4}$. As all parameters are taken to be real, CP is conserved in this model. With $11$ parameters, this model accommodates $18(+6)$ masses and mixing angles for quarks and leptons (and RH neutrinos), yielding $7$ predictions in accord with the data.
### Albright, Babu and Barr
The model proposed by Albright, Babu and Barr[@Albright:1998vf; @Albright:2000sz; @Albright:2000dk; @Albright:2001uh] has $U(1) \times Z_{4} \times Z_{4}$ as family symmetry. The model has the following particle content, in terms of $SO(10)$ representations, $$\begin{aligned}
{\mbox Matter fields:} & 16_{1}, \; 16_{2}, \; 16_{3}, \; 2(16 \oplus \overline{16}), \;
2(10),\; 6(1)
\nonumber\\
{\mbox Higgs fields:} &
4 (10), \; 2 (16 \oplus \overline{16}), \; 1 (45), \; 5(1).
\nonumber\end{aligned}$$ The mass matrices in this model are generated by the Froggatt-Nielsen mechanism, and are given by $$\begin{aligned}
M_{u} = & \left(\begin{array}{ccc}
\eta & 0 & 0\\
0 & 0 & \epsilon/3\\
0 & -\epsilon/3 & 1
\end{array}\right)\cdot m_{u}, \quad
M_{d} & = \left(\begin{array}{ccc}
\eta & \delta & \delta^{'}e^{i\phi}\\
\delta & 0 & \sigma+\epsilon/3\\
\delta^{'}e^{i\phi} & -\epsilon/3 & 1
\end{array}\right)\cdot m_{d}
\nonumber\\
M_{\nu_{LR}} = & \left(\begin{array}{ccc}
\eta & 0 & 0\\
0 & 0 & \epsilon\\
0 & -\epsilon & 1
\end{array}\right)\cdot m_{u}, \quad
M_{e} & = \left(\begin{array}{ccc}
\eta & \delta & \delta^{'}e^{i\phi}\\
\delta & 0 & -\epsilon\\
\delta^{'}e^{i\phi} & \sigma+\epsilon & 1
\end{array}\right)\cdot m_{d}.\end{aligned}$$ The right-handed neutrino Majorana mass matrix is generated by the effective operator of the type, $\frac{1}{M}16_{i}16_{j}16_{H}16_{H}$, and is given by $$M_{\nu_{RR}} = \left(\begin{array}{ccc}
c^{2}\eta^{2} & -b\epsilon\eta & a \eta\\
-b\epsilon\eta & \epsilon^{2} & -\epsilon\\
a \eta & -\epsilon & 1
\end{array}\right)\cdot \Lambda_{R}.$$ And the effective neutrino mass matrix that accommodates the LMA solution is given by $$\begin{aligned}
M_{\nu}^{eff} & \sim & \left(\begin{array}{ccc}
0 & \epsilon/(a-b) & 0\\
\epsilon/(a-b) & -\epsilon^{2}(c^{2}-b^{2})/(a-b)^{2} & -b\epsilon/(a-b)\\
0 & -b\epsilon/(a-b) & 1
\end{array}\right)
m_{u}^{2}/\lambda_{R}
\nonumber\\
& = & \left(\begin{array}{ccc}
0 & -\epsilon & 0\\
-\epsilon & 0 & 2\epsilon\\
0 & 2\epsilon & 1
\end{array}\right)
m_{u}^{2}/\lambda_{R}, \qquad ({\mbox choosing} \; a=1, \; b = c = 2).\end{aligned}$$ This model can also accommodate the SMA, and “Just So” VO solutions. An interesting property of this model is that the large mixing angle in atmospheric neutrinos is due to the lop-sided structure of $M_{e}$, in which $\epsilon \sim 0.1$ and $\sigma \sim 1$, giving rise to a large left-handed mixing angle in the $(2,3)$ block of $V_{e,L}$. The matrix $M_{e}$ is related to $M_{d}$ by the $SU(5)$ relation, $M_{e} = M_{d}^{T}$, thus such a lop-sided structure gives rise to a large mixing angle in the right-handed rotation matrix for the down-type quarks, $V_{d,R}$, making it un-observable. The large solar mixing angle is due to the structure in the $(1,2)$ block of $M_{\nu}^{eff}$, $$\left(\begin{array}{cc}
0 & -\epsilon\\
-\epsilon & 0
\end{array}\right).$$ Because the large mixing angles in the atmospheric and solar neutrinos are due to different mass matrices, the prediction for $U_{e\nu_{3}}$ can be made to be extremely small. In this model, $|U_{e\nu_{3}}|$ is predicted to be $0.014$. With $10$ parameters, this model accommodates all the $22$ masses, mixing angles, and CP violating phases at low energies.
### Maekawa
Maekawa[@Maekawa:2001uk] has proposed a $SO(10)$ model combined with an anomalous $U(1)_{A}$ symmetry. This anomalous $U(1)$ symmetry is important for achieving DTS via Dimopoulos-Wilczek mechanism and for generating fermion mass hierarchy. The $U(1)_{A}$ charge assignments project out terms that destabilize the VEV of $45_{H}$, and thus guarantee that $<45_{H}>$ is along the $B-L$ direction, which otherwise can only be achieved with fine-tuning and introducing additional Higgs multiplets. These charge assignments also give rise to the fermion mass matrices of the following form: $$M_{u} = \left(\begin{array}{ccc}
\lambda^{6} & \lambda^{5} & \lambda^{3}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right) <H_{u}>, \quad
M_{d} = \left(\begin{array}{ccc}
\lambda^{4} & \lambda^{7/2} & \lambda^{3}\\
\lambda^{3} & \lambda^{5/2} & \lambda^{2}\\
\lambda & \lambda^{1/2} & 1
\end{array}\right) <H_{d}>.$$ The charged lepton mass matrix is $M_{e} = M_{d}^{T} \cdot \eta$, where $\eta$ characterizes the renormalization effects, and the Dirac neutrino mass matrix is given by $$M_{\nu}^{Dirac} = \left(\begin{array}{ccc}
\lambda^{4} & \lambda^{3} & \lambda\\
\lambda^{7/2} & \lambda^{5/2} & \lambda^{1/2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right)\cdot \lambda^{2} \eta <H_{u}>.$$ The right-handed neutrino Majorana masses are generated by couplings to $16$-dim Higgs representations $$M_{\nu,RR} = \left(\begin{array}{ccc}
\lambda^{6} & \lambda^{5} & \lambda^{3}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right) \cdot \lambda^{9}$$ and the resulting effective neutrino mass matrix is $$M_{\nu}^{eff} = \left(\begin{array}{ccc}
\lambda^{2} & \lambda^{3/2} & \lambda\\
\lambda^{3/2} & \lambda & \lambda^{1/2}\\
\lambda & \lambda^{1/2} & 1
\end{array}\right) \cdot \lambda^{-5} \eta^{2} <H_{u}^{2}>.$$ The LM matrix is then given by $$U_{LM} = \left(\begin{array}{ccc}
1 & \lambda^{1/2} & \lambda\\
\lambda^{1/2} & 1 & \lambda^{1/2}\\
\lambda & \lambda^{1/2} & 1
\end{array}\right).$$ The LMA solution can be accommodated in this model. The prediction for $U_{e\nu_{3}}$ is about $\lambda$ which is very close to the current bound from experiment. All order $\mathcal{O}(1)$ coefficients in the mass matrices are not specified in this model, thus the validity of this model is unclear.
### Shafi and Tavartkiladze
Shafi and Tavartkiladze proposed a SUSY SO(10) model combined with an anomalous $U(1)_{H}$ symmetry[@Shafi:1999au] By extending matter content, the $U(1)_{H}$ can account for the observed mass hierarchy and mixing angles in the charged fermion sector. In this model, the three light lepton families are contained in the additional matter fields, $10_{i}$, rather than the three usual $16_{i}$ in which quarks reside. From this point of view, the quark-lepton unification is lost. The bi-large neutrino mixing is achieved by introducing three additional SO(10) singlet states. Due to the $U(1)_{H}$ charge assignment, only two of these singlets interact with the lepton doublets giving a $3 \times 2$ neutrino Dirac mass matrix, and the neutrino RH Majorana mass matrix is approximately $2 \times 2$. The LMA solution can be accommodated in this mdel.\
### Blazek, Raby and Tobe
The model proposed by Blazek, Raby and Tobe[@Blazek:1999ue; @Blazek:1999hz] has $U(2) \times U(1)^{n}$ as its family symmetry. It has the following field content, in terms of $SO(10) \times U(2)$: $$\begin{aligned}
{\mbox Matter fields:} & 1(16,2), \; 1(16,1), \; 1(1,2), \; 1(1,1)
\nonumber\\
{\mbox Higgs fields:} & 1 (10,1), \; 1(45,1)
\nonumber\\
{\mbox Flavon fields:} & 2(1,2), \; 1(1,3), \; 2(1,1_{A}).
\nonumber\end{aligned}$$ The mass matrices of this model have the following form $$\begin{aligned}
Y_{u} & = &
\left(\begin{array}{ccc}
\kappa_{1} \epsilon \rho &
(\epsilon^{'}+\kappa_{2}\epsilon)\rho &
0\\
-(\epsilon^{'}-\kappa_{2}\epsilon)\rho &
\epsilon\rho &
\epsilon r T_{\overline{u}}\\
0 &
\epsilon r T_{\overline{Q}} &
1
\end{array}\right) \lambda
\\
Y_{d} & = &
\left(\begin{array}{ccc}
\kappa_{1} \epsilon \rho &
\epsilon^{'}+\kappa_{2}\epsilon &
0\\
-(\epsilon^{'}-\kappa_{2}\epsilon) &
\epsilon\rho &
\epsilon r T_{\overline{d}}\\
0 &
\epsilon r T_{\overline{Q}} &
1
\end{array}\right) \lambda
\\
Y_{\nu_{LR}} & = &
\left(\begin{array}{ccc}
3 \kappa_{1} \epsilon \omega &
-(\epsilon^{'}-3\kappa_{2}\epsilon)\omega &
0\\
(\epsilon^{'}+3\kappa_{2}\epsilon)\omega &
3\epsilon\omega &
\frac{1}{2}\epsilon r T_{\overline{\nu}}\omega\\
0 &
\epsilon r T_{\overline{L}} \sigma &
1
\end{array}\right) \lambda
\\
Y_{e} & = &
\left(\begin{array}{ccc}
3\kappa_{1} \epsilon &
-(\epsilon^{'}-3\kappa_{2}\epsilon) &
0\\
(\epsilon^{'}+3\kappa_{2}\epsilon) &
3\epsilon &
\epsilon r T_{\overline{e}} \sigma\\
0 &
\epsilon r T_{\overline{L}} \sigma &
1
\end{array}\right) \lambda\end{aligned}$$ where $\omega = 2\sigma/(2\sigma-1)$ and $T_{f} =$ (Baryon number - Lepton number) of multiplet $f$. The right-handed neutrino Majorana mass matrix is generated by the effective operator of the type, $\frac{1}{M}16_{i}16_{j}\overline{16}_{H}\overline{16}_{H}$, and is given by $$M_{\nu_{RR}} = \left(\begin{array}{ccc}
\kappa_{1} S &
\kappa_{2} S &
0\\
\kappa_{2} S &
S &
\phi\\
0 & \phi & 0
\end{array}\right).$$ This model can accommodate LMA solution, in addition to SMA and “Just So” VO solutions. In the case of LMA solution, its prediction for $U_{e\nu_{3}}$ is $0.049$. It has $16$ input parameters which yield $22$ masses and mixing angles.
### Raby
Raby has proposed a $SO(10)$ model combined with $SU(2) \times U(1)$ family symmetry,[@Raby:2003ay] in which the ansatz proposed by Frampton, Glashow and Yanagida[@Frampton:2002qc] discussed in Sec. 3 naturally arises. The representations utilized in this model, in terms of $SO(10) \times SU(2)$, are given as follows: $$\begin{aligned}
{\mbox Matter fields:} &
1(16,2), \; 1(16,1), \; 3(1,1)
\nonumber\\
{\mbox Higgs fields:} & (10,1), \; (\overline{16},1), \; (45,1)
\nonumber\\
{\mbox Flavon fields:} & (1,1), \; (1,1_{A}), \; (1,2), \; (1,\overline{2}),
\; (1,3),
\nonumber\end{aligned}$$ in addition to several vector-like Froggatt-Nielsen fields. For the purpose of discussion, we denote the three $SO(10)\times SU(2)$ singlets by $N_{i}$ ($i=1,2,3$) which play an important role in the neutrino sector, the $SU(2)$ (anti-)doublet flavon fields by ($\tilde{\phi}$) $\phi$, and the $SU(2)$ singlet flavon fields by $\theta$ and $S_{i}$ ($i=1,2$). In the charged fermion sector, the Yukawa matrices are given by $$\begin{aligned}
Y_{u} & = &
\left(\begin{array}{ccc}
0 &
\epsilon^{'}\rho &
-\epsilon \xi\\
-\epsilon^{'}\rho &
\tilde{\epsilon}\rho &
-\epsilon\\
\epsilon\xi &
\epsilon &
1
\end{array}\right) \lambda
\\
Y_{d} & = &
\left(\begin{array}{ccc}
0 &
\epsilon^{'} &
-\epsilon\xi\sigma\\
-\epsilon^{'} &
\tilde{\epsilon} &
- \epsilon \sigma\\
\epsilon \xi &
\epsilon &
1
\end{array}\right) \lambda
\\
Y_{\nu_{LR}} & = &
\left(\begin{array}{ccc}
0 &
-\epsilon^{'} \omega &
\frac{3}{2} \epsilon\xi\omega\\
\epsilon^{'} \omega &
3\tilde{\epsilon}\omega &
\frac{3}{2}\epsilon\omega\\
-3 \epsilon \xi \sigma &
-3 \epsilon \sigma &
1
\end{array}\right) \lambda
\\
Y_{e} & = &
\left(\begin{array}{ccc}
0 &
-\epsilon^{'} &
3\epsilon \xi \\
\epsilon^{'} &
3 \tilde{\epsilon} &
3 \epsilon \\
-3 \epsilon \xi \sigma &
-3 \epsilon \sigma &
1
\end{array}\right) \lambda.\end{aligned}$$ The ansatz of Frampton, Glashow and Yanagida is obtained by considering the following superpotential in which the three $SO(10) \times SU(2)$ singlets $N_{i}$ mix with the SM singlets in the three families, via the following superpotential $$\label{numix}
W_{\nu}=
\frac{\overline{16}}{\hat{M}}(N_{1}\tilde{\phi}^{a}16_{a} + N_{2}\phi^{a}16_{a}
+ N_{3}\theta 16_{3}) + \frac{1}{2}(S_{1}N_{1}^{2}+S_{2}N_{2}^{2}).$$ The complete neutrino mass terms are thus given by $$\nu m_{\nu} \overline{\nu} + \overline{\nu} V N + \frac{1}{2}N M_{N} N$$ where $m_{\nu}$ ($\propto Y_{\nu_{LR}}$) is the Dirac mass matrix due to the coupling between $\nu$ and $\overline{\nu}$, which is related to the up-quark mass matrix, and $V$ and $M_{N}$ are the Majorana mass matrices due to the coupling between $N_{i}$ and $\overline{\nu_{i}}$, and the self coupling of $N_{i}$ in Eq.(\[numix\]), $$V = \frac{v_{16}}{\hat{M}} \left(\begin{array}{ccc}
0 & \phi^{1} & 0\\
\tilde{\phi}^{2} & \phi^{2} & 0\\
0 & 0 & \theta
\end{array}\right), \qquad M_{N} = {\mbox diag}(M_{1},M_{2},0).$$ After integrating out the heavy SM singlet neutrinos, $\overline{\nu}_{i}$ and $N_{i}$, the neutrino effective mass matrix is obtained, $$M_{\nu}^{eff} = m_{\nu} (V^{T})^{-1}M_{N} V^{-1} m_{\nu}^{T}.$$ The key observation is that if one defines $$D^{T} \equiv m_{\nu} (V^{T})^{-1} M_{N} \mathcal{P} = \left(
\begin{array}{cc}
a & 0\\
a^{'} & b\\
0 & b^{'}
\end{array}\right), \qquad
{\mbox with} \quad
\mathcal{P} = \left(\begin{array}{cc}
1 & 0 \\
0 & 1 \\
0 & 0
\end{array}\right),$$ the effective neutrino mass matrix can be rewritten in the following form $$D^{T} \hat{M}_{N}^{-1} D$$ where $$\hat{M}_{N} \equiv \left(\begin{array}{cc}
M_{1} & 0\\
0 & M_{2}
\end{array}\right)$$ which is the ansatz proposed by Frampton [*et al*]{}. Note that in this model, because the mixing in the charged lepton sector is small, the LM matrix is approximately the diagonalization matrix of $M_{\nu}^{eff}$, and the prediction of a bi-large mixing pattern is not affected. Even though the bi-large mixing pattern can naturally arise in this model, the connection between the CP violation in neutrino oscillation and the sign of the baryogenesis, which exists in the model of Frampton, Glashow and Yanagida, is lost, due to additional CP phases and more complicated structure in this model.\
### Berezhiani and Rossi
The model proposed by Berezhiani and Rossi[@Berezhiani:1998vn] has $SU(3)$ as family symmetry with some unspecified discrete symmetries imposed. It has the following field content, in terms of $SO(10) \times SU(3)$: $$\begin{aligned}
{\mbox Higgs fields:} &
2(10,1), \; 1 (16,1),\; (\overline{16},1), \; 2(45,1), \; 1 (54,1)
\nonumber\\
{\mbox Flavon fields:} &
1(1,\overline{6}), \; 3 (1,3),\; 1(1,8).
\nonumber \end{aligned}$$
The mass matrices of this model are given as follows: $$\begin{aligned}
Y_{u} & = \left(\begin{array}{ccc}
y_{u} & 0 & 0\\
0 & y_{c} & 0\\
0 & 0 & y_{t}\\
\end{array}\right), \qquad
Y_{d} & = \left(\begin{array}{ccc}
y_{ut} D e^{i\xi} & Ae^{i\sigma} & C\\
Ae^{i\sigma} & y_{ct}D e^{i\xi} & B\\
\frac{1}{b}C & \frac{1}{b}B & D\\
\end{array}\right)
\\
Y_{\nu_{LR}} & = \left(\begin{array}{ccc}
y_{u} & 0 & 0\\
0 & y_{c} & 0\\
0 & 0 & \frac{1}{b}y_{t}\\
\end{array}\right), \qquad
Y_{e} & = \left(\begin{array}{ccc}
y_{ut} D e^{i\xi} & k_{3} Ae^{i\sigma} & \frac{1}{b}k_{2}C\\
k_{3}Ae^{i\sigma} & y_{ct}D e^{i\xi} & \frac{1}{b}k_{1}B\\
k_{2} C & k_{1} B & D\\
\end{array}\right).\end{aligned}$$ The right-handed neutrino Majorana mass matrix is taken to be diagonal. Because both $Y_{\nu_{LR}}$ and $M_{\nu_{RR}}$ are diagonal, the effective neutrino mass matrix is also diagonal. Thus the leptonic mixing matrix is proportional to $V_{e,L}$. Due to the lop-sided structure in the $(2,3)$ block of $Y_{e}$ arising from the $SU(5)$ breaking chain, the maximal mixing angle in atmospheric neutrinos is obtained. Nevertheless, the solar mixing angle in this model is very small, which is in the range of SMA solution. This model has $14$ input parameters in the Yukawa sector. The value of $\sin2\theta_{13}$ is predicted to be $\mathcal{O}(10^{-2})$.
### Kitano and Mimura
Kitano and Mimura[@Kitano:2000xk] propose a $SO(10)$ model combined with $SU(3) \times U(1)_{H}$ family symmetry. The three families of matter fields form a triplet of $SU(3)$. The Yukawa couplings in this model have the form $$\sum_{i,j=1}^{3} (\frac{\Phi}{M_{pl}})^{x_{i}+x_{j}}
\frac{\xi_{j} \xi_{j}}{M_{\ast}^{2}}.$$ The field $\Phi$ is a singlet of $SU(3)$, but it has non-vanishing $U(1)_{H}$ charges. Thus its VEV $\lambda M_{pl}$ provides the mass hierarchy, if different generations, $i$, have different $U(1)_{H}$ charge, proportional to $-x_{i}$. The field $\xi_{i}$’s are $\overline{3}$ representation of $SU(3)$, whose VEV break the $SU(3)$ family symmetry $$<\xi_{1}> \sim \left(\begin{array}{c}
0 \\ 0 \\ 1 \end{array}\right)M_{\ast}, \quad
<\xi_{2}> \sim \left(\begin{array}{c}
0 \\ 1 \\ 1 \end{array}\right)M_{\ast}, \quad
<\xi_{3}> \sim \left(\begin{array}{c}
1 \\ 1 \\ 1\end{array}\right)M_{\ast}$$ where $M_{\ast}$ is the family symmetry breaking scale. With the $U(1)_H$ charge assignment, the mass matrices of the up- and down-type quarks are $$\begin{aligned}
Y_{u} & \sim & \frac{1}{M_{\ast}^{2}}
(<\xi_{3}> \; <\xi_{2}> \; <\xi_{1}>)
\; \left(\begin{array}{ccc}
\lambda^{6} & \lambda^{5} & \lambda^{3}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right)
\left(\begin{array}{c}
<\xi_{3}^{T}> \\ <\xi_{2}^{T}> \\ <\xi_{1}^{T}>
\end{array}\right)
\nonumber\\
& \sim &
\left(\begin{array}{ccc}
\lambda^{6} & \lambda^{5} & \lambda^{3}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right)
\\
Y_{d} & = & Y_{e}^{T} \sim \frac{1}{M_{\ast}^{2}}
(<\xi_{3}> \; <\xi_{2}> \; <\xi_{1}>)
\; \left(\begin{array}{ccc}
\lambda^{5} & \lambda^{4} & \lambda^{4}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right)
\left(\begin{array}{c}
<\xi_{3}^{T}> \\ <\xi_{2}^{T}> \\ <\xi_{1}^{T}>
\end{array}\right)
\nonumber\\
& \sim &
\left(\begin{array}{ccc}
\lambda^{5} & \lambda^{4} & \lambda^{4}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right).\end{aligned}$$ These Yukawa matrices give rise to the observed hierarchical masses $$\begin{aligned}
m_{u} : m_{c} : m_{t} & \sim & \lambda^{6} : \lambda^{4} : 1
\\
m_{d} : m_{s} : m_{b} & = & m_{e} : m_{\mu} : m_{\tau}
\sim \lambda^{4} : \lambda^{2} : 1\end{aligned}$$ and the CKM matrix $$V_{CKM} \sim \left(\begin{array}{ccc}
1 & \lambda & \lambda^{3}\\
-\lambda & 1 & \lambda^{2}\\
-\lambda^{3} & -\lambda^{2} & 1
\end{array}\right).$$ The role of these $\xi$ fields is that it gives rise to the bi-maximal mixing pattern in the neutrino sector. With the $U(1)$ charge assignment, the neutrino mass matrix is $$(<\xi_{3}> \; <\xi_{2}> \; <\xi_{1}>)
\; \left(\begin{array}{ccc}
\lambda^{2} & \lambda & \lambda^{3}\\
\lambda & 1 & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & \lambda^{4}
\end{array}\right)
\left(\begin{array}{c}
<\xi_{3}^{T}> \\ <\xi_{2}^{T}> \\ <\xi_{1}^{T}>
\end{array}\right)
\sim
\left(\begin{array}{ccc}
\lambda^{2} & \lambda & \lambda\\
\lambda & 1 & 1\\
\lambda & 1 & 1
\end{array}\right).$$ To implement the above idea in $SO(10)$ is not trivial, because all the three families must have the same charge assignment under $U(1)_{H}$. This is overcome by introducing many Froggatt-Nielsen fields and diagrams so that the phenomenologically viable mass matrices discussed above are reproduced. The neutrino masses are generated by a complicated mechanism in which each SM singlet in the $16$-dim matter field mixes with a Froggatt-Nielsen field which transforms as singlet under $SO(10)$. Thus, instead of a $6 \times 6$ neutrino mass matrix, this model has a $9 \times 9$ neutrino mass matrix. The Dirac neutrino mass matrix is different from that of the up-type quark because they are due to different Froggatt-Nielsen diagrams. Large mixing angle in the atmospheric sector can be accommodated, and the LMA solution can also be accommodated.
### Ross and Velasco-Sevilla
Ross and Velasco-Sevilla[@Ross:2002fb] utilize $SU(3)$ as the family symmetry in combination with $SO(10)$. Based on the formulation given in Sec. 2,[@King:2001uz] the up- and down-type quark mass matrices are of the form $$M = \left(\begin{array}{ccc}
\lambda^{'} \epsilon^{8} & \lambda \epsilon^{3} & \lambda \epsilon^{3}\\
-\lambda \epsilon^{3} & \lambda^{''} \epsilon^{2} & \lambda^{''} \epsilon^{2}\\
-\lambda \epsilon^{3} & \lambda^{''} \epsilon^{2} & 1 + \lambda^{''} \epsilon^{2}
\end{array}\right) M_{33}.$$ Note the expansion parameter $\epsilon$ is different for up- and down-quark sectors. The Dirac neutrino mass matrix is of the form $$M_{\nu}^{Dirac} = \left(\begin{array}{ccc}
\mathcal{O}(\epsilon^{8})
& \epsilon^{3} (z + \epsilon (x+y) )
& \epsilon^{3} (x + \epsilon (x-y) )\\
\epsilon^{3} (z + \epsilon (x+y) )
& \epsilon^{2} (a w + \epsilon u)
& \epsilon^{2} (a w - \epsilon u)\\
\epsilon^{3} (x + \epsilon (x-y) )
& \epsilon^{2} (a w - \epsilon u)
& 1
\end{array}\right) \cdot M_{\nu, \; 33}^{Dirac}$$ where $M_{\nu, \; 33}^{Dirac}$ is the $(33)$ component of the Dirac neutrino mass matrix, and the real parameters $z, a, w$ and complex parameters $x, y, u$ are of $\mathcal{O}(1)$. Note that the higher order terms in symmetry breaking parameters have been included in $M_{\nu}^{Dirac}$ given above, as they are crucial for getting near bi-maximal mixing pattern. The right-handed neutrino Majorana mass matrix is assumed to be diagonal $$M_{\nu,RR} = \left(\begin{array}{ccc}
m_{1} & 0 & 0\\
0 & m_{2} & 0\\
0 & 0 & m_{3}
\end{array}\right).$$ The mass eigenstates of the effective neutrino mass matrix are given by $$\begin{aligned}
\nu_{1} & \simeq & \frac{|r|\nu_{e}-|z|e^{-i\xi}\nu_{b}}
{\sqrt{z^{2}+r^{2}}}
\\
\nu_{2} & \simeq & \frac{|r|e^{i\xi}\nu_{e}+|z|\nu_{b}}
{\sqrt{z^{2}+r^{2}}}
\\
\nu_{3} & \simeq & \nu_{a}\end{aligned}$$ where $\xi = Arg(z) - Arg(r)$ and $r = \sqrt{2}(z u - a_{\nu} w y) / z$. Here $\nu_{a}$ is the heaviest mass eigen states given by $$\nu_{a} =
\frac{(z+\epsilon x)(\nu_{\tau} + \nu_{\mu}) - \epsilon y (\nu_{\tau}-\nu_{\mu})}
{\sqrt{(z+\epsilon(x+y))^{2}+(z+\epsilon(x-y))^{2}}}$$ and $\nu_{b}$ is orthogonal to $\nu_{a}$. As one can see in $\nu_{3}$, $\nu_{\mu}$ contributes equally as $\nu_{\tau}$, thus the mixing angle between $\nu_{\mu}$ and $\nu_{\tau}$ is maximal. A large solar mixing angle is also obtained if $\tan^{2}\theta_{12} \simeq |z / r|^{2}$ is in the right parameter space. Solutions for these parameters that are consistent with other charged fermion masses and CKM matrix elements have been found; it can accommodate LMA, LOW and “Just So” VO solutions. The CHOOZ angel in this model is predicted to be about $U_{e\nu_{3}} \simeq 0.07$.
Comparisons of Models and Other Issues
--------------------------------------
### Distinguishing Models using CKM Unitarity triangle
To distinguish various existing models, clearly we need more precise results from the experiments. The most sensitive test are the three angles in the CKM unitarity triangle, as suggested in Ref. 141. To illustrate this point, in Table \[compare\],
predictions for the three angles of the CKM unitarity triangle from various models are given. Clearly, the predictions from these models are very different. Many models will be ruled out as soon as one can pin down the values for these three angles more accurately.
### Distinguish Models Using $\sin^{2}\theta_{13}$
The two classes of models discussed in this section predict very different relations between $U_{e\nu_{3}}$ and $\Delta m_{\odot}^{2} / \Delta m_{atm}^{2}$, as discussed in Ref. 56, 142, 143, 48. The predictions for $\sin\theta_{13}$ of various $SO(10)$ models are summarized in Table.\[theta13\].
In the model of Chen and Mahanthappa,[@Chen:2002pa] the typical value for $\sin\theta_{13}$ is very close to the sensitivity of current experiments. In this model, the value of $U_{e\nu_{3}}=\sin\theta_{13}$ is related to the ratio $\Delta m_{\odot}^{2} / \Delta m_{atm}^{2}$ as $$U_{e\nu_{3}} \sim (\Delta m_{\odot}^{2} / \Delta m_{atm}^{2})^{1/3}.$$ Thus as this ratio increases, the value of the angle $\theta_{13}$ increases. As the LMA solution is the most favored solution, the angle $\theta_{13}$ in this model is predicted to be very close to the current sensitivity of experiments. We note that $\theta_{13}$ of this order of magnitude leads to observable CP violation in neutrino oscillation.
In the model of Albright and Barr,[@Albright:2002if; @Albright:2001xq] the relation between $\Delta m_{\odot}^{2} / \Delta m_{atm}^{2}$ and $\sin\theta_{13}$ is quite different: as the ratio $\Delta m_{\odot}^{2} / \Delta m_{atm}^{2}$ increases, the prediction for $|U_{e\nu_{3}}|$ decreases. This is due to the fact that the large mixing angles in the atmospheric sector and solar sector have different origins. The most favored LMA solution thus implies that the value of $U_{e\nu_{3}}$ is extremely small; a neutrino factory is needed in this case in order to pin down its value.
### $b-\tau$ unification
In most models, the prediction for $m_{b}$ at the weak scale tends to be higher than the experimental observed value, and thus a threshold correction of the size $-(15-20) \; \%$ is needed to bring down $m_{b}$. Such a large threshold corrections for $m_{b}$ are expected due to loop diagrams of $SU(2)$-singlet bottom squark, the $SU(2)$-doublet third generation squark, gluinos and charginos, as discussed in Sec. 4. Barr and Dorsner[@Barr:2002mw] suggested that, instead of these threshold corrections to $m_{b}$ being large and negative, $m_{b}/m_{\tau}$ may indeed be smaller than one at the GUT scale, and the deviation from the naive $b-\tau$ unification in $SU(5)$ is due to the large off-diagonal element of the charged lepton mixing matrix which also explain the large mixing in atmospheric neutrinos in models with lop-sided mass textures.
### CP Violation
CP violation arises in these models from the complex phases in the VEV’s of the scalar fields, and from the complex phases of the Yukawa coupling constants. Thus they are free parameters in the models. In the quark sector, the complex phase is constrained by the masses and the mixing angles, thus definite predictions for the three angles $(\alpha, \beta, \gamma)$ in the CKM unitarity triangle can be obtained. In the leptonic sector, on the contrary, it is not possible at this moment to obtain definite predictions for the three CP violating phases. The situation will be improved once the absolute scales of neutrino masses and the three mixing angles are known to much better precision.
SUSY $SO(10)$ in Higher Dimensions {#modeled}
==================================
The idea of orbifold (SUSY) GUT’s was first proposed by Kawamura to solve the doublet-triplet splitting problem,[@Kawamura:1999nj; @Kawamura:2000ev; @Kawamura:2000ir] and later developed by Altarelli and Ferugio[@Altarelli:2001qj] and Hall and Nomura.[@Hall:2001pg] The size of this type of extra dimensions are small, being inverse of the GUT scale, $R \sim 1/M_{GUT}$. To see how it works, let us consider the case with only one extra dimension, which is compactified on a $S^{1}/Z_{2}$ orbifold. The circle $S^{1}$ has radius $R$ and is defined by $y=y + 2\pi R$. Under $Z_{2}$, $y$ is mapped to $-y$. Thus the physical region can be taken as $0 \le y \le \pi R$. Various components transform under the $Z_{2}$ symmetry as follows $$\begin{aligned}
A_{\mu}(x,y) & \rightarrow & A_{\mu}(x,-y) = P A_{\mu}(x,y) P^{-1}
\\
A_{5}(x,y) & \rightarrow & A_{5}(x,-y) = -P A_{5}(x,y) P^{-1}
\\
\Phi(x,y) & \rightarrow & \Phi(x,-y) = \pm P \Phi(x,y)\end{aligned}$$ where $A_{\mu}(x,y)$ and $A_{5}(x,y)$ are the components of the gauge fields along the usual $4D$ and the $5th$-dimension, respectively; $\Phi(x,y)$ is a generic matter or Higgs field. The generators transform according to the following transformation rules, $$\label{sbo}
P T^{a} P^{-1} = T^{a}, \quad P T^{\hat{a}} P^{-1} = -T^{\hat{a}}.$$ Here $T^{a}$ are generators of the residual symmetry group while $T^{\hat{a}}$ are the broken generators. The $5D$ bulk field can be decomposed into a infinite tower of $KK$ states $$\begin{aligned}
\phi_{+}(x^{\mu},y) & = & \frac{1}{\sqrt{\pi R}} \sum_{n=0}^{\infty} \phi_{+}^{(n)}
\cos \frac{ny}{R}\\
\phi_{-}(x^{\mu},y) & = & \frac{1}{\sqrt{\pi R}} \sum_{n=1}^{\infty} \phi_{-}^{(n)}
\sin \frac{ny}{R}.\end{aligned}$$ The mode $\phi_{+}$ is even under the $Z_{2}$ symmetry, $$P \phi_{+}(x,y) = \phi_{+}(x,-y) = + \phi_{+}(x,y).$$ After compactification, it has a zero mode $\phi_{+}^{(0)}$. The $\phi_{-}$ is odd under $Z_{2}$ transformation, $$P \phi_{-} (x,y) = \phi_{-}(x,-y) = - \phi_{-}(x,y),$$ thus it does not have a zero mode, and its $n$-th KK mode has a GUT scale mass, $(2n+1)/\pi R$. For the broken generators, the corresponding gauge bosons thus acquire GUT scale masses; only those corresponding to the un-broken generators have $4D$ zero modes, which are then identified as the gauge fields of the little group. From the conditions given in Eq.(\[sbo\]), one sees that this symmetry breaking mechanism only works for non-Abelian symmetry, because the generators of an Abelian symmetry always commutes with the parity operator, $P$. In other words, it is not possible to reduce the rank of a group by Abelian orbifolding,[^5] and additional $U(1)$ symmetries survive, along with the SM gauge group, if the GUT symmetry has rank larger than $4$. Breaking these $U(1)$ symmetries can be achieved by the usual Higgs mechanism. In orbifold GUT models, because the GUT symmetry is broken by the orbifold boundary conditions, one does avoid the task of constructing symmetry breaking scalar potential which is usually non-trivial.
One should also note that in higher space-time dimensions, supersymmetry is enlarged: In $5D$, $N=1$ SUSY has $8$ super charges; it corresponds to $N=2$ SUSY from the $4D$ point of view. By having different orbifold boundary conditions for different components in a “super” multiplet, one can reduce supersymmetry to $4D$ $N=1$, similar to the case of GUT breaking. To break both SUSY and the gauge symmetry by orbifolding, a larger discrete Abelian orbifold is thus needed when building a realistic model. A $N=2$ hypermultiplet can be decomposed into two $N=1$ chiral multiplets; and a $N=2$ vector multiplet can be decomposed into a $N=1$ vector multiplet (denoted by “$V$”) and a $N=1$ chiral multiplet (denoted by “$\Sigma$”). (See, for example, pp. 348-351 of Ref. 84.) To break $N=2$ SUSY down to $N=1$, we thus require $V$ to be even under the parity, and $\Sigma$ to be odd.
An immediate question one might ask is that how does the Georgi-Jarlskog $(-3)$ factor arise? Recall that in $4D$ GUT models, the GJ factor arises as the Clebsch-Gorden coefficients associated with the VEV’s of scalar fields along certain symmetry breaking directions. In orbifold GUT scenario, because the GUT symmetry is broken by orbifold boundary conditions, the GJ factor must arise in some other way.
Breaking SUSY $SO(10)$ by Orbifolding
-------------------------------------
Several orbifoldings have been found to break $SO(10)$. The number of extra dimensions that has been considered is either one or two. Some orbifoldings break $SO(10)$ to only its maximal subgroups; others break $SO(10)$ fully down to $SU(3) \times SU(2)_{L}
\times U(1)_{Y} \times U(1)^{'}$. Here we summarize various orbifolding that have been constructed.
### SUSY $SO(10)$ in $5D$
Dermisek and Mafi[@Dermisek:2001hp] consider $SO(10)$ in $5D$ and the extra dimension is compactified on $S^{1}/(Z_{2}\times Z_{2}^{'})$ orbifold. The parities are chosen to be $$\begin{aligned}
P & = & I_{5 \times 5} \otimes I_{2 \times 2}\\
P^{'} & = & diag(-1,-1,-1,1,1) \otimes I_{2 \times 2}\end{aligned}$$ and the orbifold boundary conditions are given by $$\begin{array}{lllllllll}
45_{{\mbox v}}: & V_{(15,1,1)}^{++} & V_{(1,3,1)}^{++}
& V_{(1,1,3)}^{++} & V_{(6,2,2)}^{+-}
&
\Sigma_{(15,1,1)}^{-+} & \Sigma_{(1,3,1)}^{-+} & \Sigma_{(1,1,3)}^{-+}
& \Sigma_{(6,2,2)}^{--}
\\
&&&&&&&&
\\
10_{{\mbox H}}: &
H_{(1,2,2)}^{++} & H_{(6,1,1)}^{+-} &
H_{(1,2,2)}^{c \; --} & H_{(6,1,1)}^{c \; -+}.&&&&
\end{array}$$ After compactification, the parity $Z_{2}$ reduces $N=2$ SUSY to $N=1$ SUSY in $4D$, and the parity $Z_{2}^{'}$ reduces $SO(10)$ to $G_{PS}$. The residual symmetry below the compactification scale is the Pati-Salam group $SU(4)_{c} \times SU(2)_{L} \times SU(2)_{R}$ on the 4D “hidden” brane at fixed point $y=\pi R/2$, which is then broken down to $G_{SM}$ by the usually Higgs mechanism, and the symmetry on the “visible” brane at the fixed point $y=0$ is $SO(10)$.
Kim and Raby[@Kim:2002im] pursue along this line, and analyze the renormalization group evolution in this model: The $5D$ gauge coupling constant unification scale $M_{\ast}$ (to be contrasted with the $4D$ unification scale $M_{GUT}$) is found to be $\sim 3 \times 10^{17} \; GeV$ and the compactification scale is found to be $\sim 1.5 \times 10^{14} \; GeV$.
Kyae and Shafi[@Kyae:2002ss; @Kyae:2002hu] consider different parities, $$\begin{aligned}
P & = & diag(I_{3 \times 3}, I_{2 \times 2}, -I_{3 \times 3},
-I_{2 \times 2})\\
P^{'} & = & diag(-I_{3 \times 3}, I_{2 \times 2}, I_{3 \times 3},
-I_{2 \times 2})\end{aligned}$$ and the $Z_{2} \times Z_{2}^{'}$ charge assignments for the components of the $SO(10)$ gauge field is $$\begin{array}{lllllllll}
45_{{\mbox v}}: & V_{(8,1,0)}^{++} & V_{(1,3,0)}^{++}
& V_{(1,1,0)}^{++} & V_{(3,\overline{2},-5/6)}^{+-} &
V_{(\overline{3},2,5/6)}^{+-} & V_{(\overline{3},1,-2/3)}^{--}
& 2 V_{(3,2,1/6)}^{-+}
& V_{(1,1,1)}^{--}
\end{array}.$$ With these boundary conditions, they are able to break $SO(10)$ down to $SU(3) \times SU(2)_{L} \times U(1)_{Y} \times U(1)^{'}$.
### SUSY $SO(10)$ in $6D$
Asaka, Buchmuller and Covi [*et al*]{} [@Asaka:2001eh] consider $SO(10)$ in $6D$ and the extra dimensions are compactified on a $T^{2}/(Z_{2}\times Z_{2}^{GG}
\times Z_{2}^{PS})$ orbifold. The idea is based on the observation that a simple extension of the SM gauge group to $SU(3) \times SU(2)_{L} \times U(1)_{Y} \times U(1)'$ (which has the same rank as $SO(10)$) is the common symmetry subgroup of the Pati-Salam gauge group, $SU(4) \times SU(2)_{L} \times SU(2)_{R}$ ($G_{PS}$) and the Georgi-Glashow gauge group, $SU(5) \times U(1)$, ($G_{GG})$. The first parity $P$ breaks supersymmetris down to $N=1$ in $4D$, upon compactification on $T^{2}/Z_{2}$. The other two parities break the $SO(10)$ gauge symmetry, and can be taken to be $$P_{GG} = \left(\begin{array}{ccccc}
\sigma_{2} &&&&\\
& \sigma_{2} &&&\\
&& \sigma_{2} &&\\
&&& \sigma_{2} &\\
&&&& \sigma_{2}
\end{array}\right), \quad
P_{PS} =
\left(\begin{array}{ccccc}
-\sigma_{0} &&&&\\
& -\sigma_{0} &&&\\
&& -\sigma_{0} &&\\
&&& -\sigma_{0} &\\
&&&& -\sigma_{0}
\end{array}\right)$$ in the vector representation of $SO(10)$. At the fixed point of $Z_{2}^{GG}$, SUSY $G_{GG}$ is respected; at the fixed point of $Z_{2}^{PS}$, SUSY $G_{PS}$ is respected. The charge assignments for the gauge fields are chosen such that component fields belonging to the symmetric subgroup have positive parity and those belonging to the coset space have negative parity. At the intersection of two $5D$ subspaces of the $6D$ bulk, in which $G_{PS}$ and $G_{GG}$ are un-broken, respectively, the extended SM gauge group, $SU(3) \times SU(2)_{L} \times [U(1)]^{2}$ is realized. At this intersection (which is also one of the fixed points of the orbifold transformations), the electroweak symmetry and the additional $U(1)^{'}$ are broken by the usual Higgs mechanism.
Fermion Mass Hierarchy in SUSY $SO(10)$ Models in Higher Dimensions
-------------------------------------------------------------------
A few mechanisms have been proposed to solve the fermion mass hierarchy problem in orbifold GUTs. Some models address the realistic mass relations; the gauge symmetry breaking in this type of models is due to both orbifolding and the usual Higgs mechanism from which the GJ factor $-3$ needed to achieve the realistic mass relations arises as the CG coefficient associated with the VEVs of some Higgs multiplets. Other models make use of mechanisms that are purely higher dimensional, [*e.g.*]{} overlap wave function, to generate the mass hierarchy. No mechanisms have been found to generate the CG factor by compactification.
### Hall et al
Hall [*et al*]{} [@Hall:2001xr] proposed three $SO(10)$ models in $6D$, and the two extra dimensions are compactified on $T^{2}/Z_{2}$, $T^{2}/Z_{6}$, and $T^{2}/(Z_{2}\times Z_{2}^{'})$ tori, respectively, in each of these three models. In the first model there is $N=1$ SUSY in the bulk, and the other two models have $N=2$ SUSY. These models incorporate a mechanism proposed by Hall [*et al*]{} [@Hall:2001pg; @Hall:2001rz] to solve the fermion mass hierarchy problem in which the correct mass relations are generated by mixing the brane localized matter fields with additional matter fields that propagate along the fixed line. The generic Yukawa interactions for a brane-confined field, $\psi(x)$, and a bulk field, $\Phi(x,y)$, are given as follows: $$\mathcal{L} \sim \int dy
\{ \lambda_{0} \delta(y) \psi^{3} + \lambda_{1} \delta(y) \psi^{2}\Phi(y)
+ \lambda_{2} \delta(y) \psi \Phi(y)^{2} + \lambda_{3} \delta(y)\Phi(y)^{3}
+ \lambda_{4} \Phi(y)^{3}\}.$$ The effective Lagrangian below the compactification scale is $$\lambda_{0} \psi^{3} + \frac{\lambda_{1}}{V^{1/2}} \psi^{2} \Phi_{(0)}
+ \frac{\lambda_{2}}{V} \psi \Phi_{(0)}^{2}
+ \frac{\lambda_{3}}{V^{3/2}} \Phi_{(0)}^{3}
+ \frac{\lambda_{4}}{V^{1/2}} \Phi_{(0)}^{3}$$ where $V=M_{string} \cdot R$ is the volume factor. Thus different Yukawa coupling constants in the $4D$ effective Lagrangian have different volume suppression factors, depending upon how many bulk fields are involved in the interactions. Therefore by having different matter multiplets locate at different locations in the bulk, the mass hierarchy can be generated.
### Albright and Barr
Albright and Barr[@Albright:2002pt] consider $SO(10)$ in $5D$ and the fifth dimension is compactified on a $S^{1}/(Z_{2}\times Z_{2}^{'})$ orbifold. All quarks and leptons and most of the Higgs fields employed in Ref. 131 are confined to the $SO(10)$ 3-brane; only the $SO(10)$ gauge fields and a $10$- and a $45$-dimensional Higgs fields are placed in the $5D$ bulk. As a consequence, most features of the $4D$ model[@Albright:2000sz] also exist here; the only exception is that because the DTS problem is solved by orbifolding, the Higgs superpotential in this case is simpler as the terms needed for DTS are absent. The parity $Z_{2}$ breaks $N=2$ SUSY (from the $4D$ point of view) down to $N=1$ SUSY; $Z_{2}^{'}$ breaks $SO(10)$ down to the Pati-Salam group. The orbifold boundary conditions under $Z_{2} \times Z_{2}^{'}$ are given by $$\begin{array}{lllllllll}
45_{{\mbox v}}: & V_{(15,1,1)}^{++} & V_{(1,3,1)}^{++}
& V_{(1,1,3)}^{++} & V_{(6,2,2)}^{+-}
&
\Sigma_{(15,1,1)}^{-+} & \Sigma_{(1,3,1)}^{-+} & \Sigma_{(1,1,3)}^{-+}
& \Sigma_{(6,2,2)}^{--}
\\
&&&&&&&&
\\
45_{{\mbox H}}: &
H_{(15,1,1)}^{++} & H_{(1,3,1)}^{++}
& H_{(1,1,3)}^{++} & H_{(6,2,2)}^{+-}
&
H_{(15,1,1)}^{c \; --} & H_{(1,3,1)}^{c \; -+}
& H_{(1,1,3)}^{c \; --} & H_{(6,2,2)}^{c \; -+}
\\
&&&&&&&&
\\
10_{{\mbox H}}: &
H_{(1,2,2)}^{++} & H_{(6,1,1)}^{+-} &
H_{(1,2,2)}^{c \; --} & H_{(6,1,1)}^{c \; -+}.&&&&
\end{array}$$ The VEV $<45_{H}>$ arising from the complete Higgs superpotential on the visible $SO(10)$ brane is along the $(B-L)$ direction, thus the GJ factor of $-3$ remain in this model. Because all the matter fields are confined to the $4D$ brane, the Yukawa sector of this $5D$ model is essentially the same as that given in Ref. 131.
### Kitano and Li
Kitano and Li[@Kitano:2003cn] propose a supersymmetric $SO(10)$ model in $5D$; the extra dimension is compactified on $S_{1}/Z_{2}$ orbifold, which breaks $N=2$ SUSY down to $N=1$, from the $4D$ point of view. The gauge symmetry $SO(10)$ can be broken either by the orbifold boundary conditions or by the usual Higgs mechanism. All three families of matter fields along with the gauge fields propagate in the bulk; a $45_{H}$, a pair of $16 \oplus \overline{16}$ which are needed to break the rank of the symmetry, and a $10_{H}$ are confined to the visible brane.
In this model, the fermion mass hierarchy is accommodated utilizing the overlap between zero mode profiles along the fifth dimension, as discussed in Sec.\[qmass\]. In a $5D$ SUSY theory compactified on $S^{1}/Z_{2}$ orbifold, the zero mode wave function of a $5D$ bulk field with a bulk mass term $m$ is localized exponentially as $$f_{0}(y) \sim e^{-my}.$$ In the $SO(10)$ symmetric limit, all fields in one family must have the same bulk mass term, $m_{i}$, resulting in unrealistic mass spectrum. When the $U(1)_{X}$ subgroup of $SO(10)$ is broken by the Higgs mechanism, the VEV of the scalar field $<\phi>$ which triggers this breaking also contributes to the bulk mass terms of the matter fields. The resulting bulk mass terms are of the form $$m_{i} \rightarrow m_{i} - \sqrt{2} g_{X} Q_{X}^{i} <\phi>.$$ Because different $SU(5)$ components of $SO(10)$, $1$, $\overline{5}$ and $10$, have different $U(1)_{X}$ charges, $Q_{X}^{i}$, a realistic mass spectrum can be obtained $$Y_{u} \sim \left(\begin{array}{ccc}
\lambda^{6} & \lambda^{5} & \lambda^{3}\\
\lambda^{5} & \lambda^{4} & \lambda^{2}\\
\lambda^{3} & \lambda^{2} & 1
\end{array}\right), \;
Y_{d} = Y_{e}^{T} \sim \left(\begin{array}{ccc}
\lambda^{4} & \lambda^{3} & \lambda^{3}\\
\lambda^{3} & \lambda^{2} & \lambda^{2}\\
\lambda & 1 & 1
\end{array}\right), \;
m_{\nu}^{eff} \sim \left(\begin{array}{ccc}
\lambda^{2} & \lambda & \lambda\\
\lambda & 1 & 1\\
\lambda & 1 & 1
\end{array}\right).$$ The lop-sidedness of $Y_{e}$ thus gives the maximal atmospheric mixing angle and the LMA solution to the solar neutrino problem is accommodated.
### Shafi and Tavartkiladze
Shafi and Tavartkiladze considered SO(10) in 5D compactified on a $S^{1}/(Z_{2} \times Z_{2}^{'})$ orbifold.[@Shafi:2003ie] The flavor structure of this model arises at the fixed point which has $G_{PS}$ symmetry, thus the mechanism which generates fermion mass hierarchy and mixing angles is purely $4$-dimensional. By extending the matter content of their model and by imposing $U(1)_{H}$ symmetry, the charged fermion masses and mixing angles can be accommodated. Bi-large neutrino mixing pattern is achieved by imposing “flavor democracy” in the neutrino sector, which has been made possible due to the extension of matter content.
Conclusion {#conclude}
==========
SUSY GUT is one of the promising candidates for physics beyond the standard model: the hierarchy problem is solved, charge quantization is explained, gauge coupling constants unification is achieved; as a consequence, a prediction for the weak mixing angle $\sin^{2}\theta_{w}$ is obtained. It provides a natural framework for small neutrino masses to arise, and it has the promise for baryogenesis. We have seen in this review how the fermion mass hierarchy can arise from a very contrained framework of $SO(10)$; this is achieved by imposing family symmetries. As proton decay has not been observed, SUSY GUT’s in $4D$ are under siege. This situation can be alleviated if the SUSY GUT model is constructed in higher dimensions. The presence of extra dimensions also provides new ways to understand fermion mass hierarchy. On the experimental side, one hopes that more precise measurements for the masses and CKM matrix elements will enable us to distinguish these models, thus pointing out the right direction for model building. On the theory side, one hopes to obtain an understanding of the complicated symmetry breaking patterns and charge assignments that are needed in many of these models, which will then shed some light on Physics beyond the Standard Model.
Acknowledgements {#acknowledgements .unnumbered}
================
M.-C.C. and K.T.M. are supported in part by the US Department of Energy Grant No. DE-AC02-98CH10886 and DE-FG03-95ER40894, respectively.
[0]{}
bio.tex
[^1]: There are five other similar conditions one can write down.
[^2]: The LM matrix is sometimes called Pontecorvo-Maki-Nakagawa-Sakata (PMNS) or Maki-Nakagawa-Sakata (MNS) matrix. It was first discussed, in a two flavor case, by Pontecorvo[@pontecorvo:1967] and by Maki, Nakagawa and Sakata.[@Maki:mu] The mixing matrix with $3$ flavors was first discussed by Lee, Pakvasa, Shrock and Sugawara.[@Lee:1977qz; @Lee:1977ti]
[^3]: If neutrinos are Dirac particles, Eq.(\[nudiag\]) becomes $$M_{\nu}^{diag} = V_{e_{L}} M_{\nu}^{eff} V_{\nu_{R}}^{\dagger}
= diag(m_{\nu_{1}},m_{\nu_{2}},m_{\nu_{3}}).$$
[^4]: Note that some of the earlier results were not entirely correct; re-derivation of these results has been done in Ref. 65, 66.
[^5]: In very limited cases of orbifold breaking by outer automorphism, the rank reduction may be possible.[@Hebecker:2001jb; @Quiros:2003gg]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The present study deals with dissipative future universe without big rip in context of Eckart formalism. The generalized chaplygin gas, characterized by equation of state $p=-\frac{A}{\rho^\frac{1}{\alpha}}$, has been considered as a model for dark energy due to its dark-energy-like evolution at late time. It is demonstrated that, if the cosmic dark energy behaves like a fluid with equation of state $p=\omega\rho$; $\omega < -1$, as well as chaplygin gas simultaneously then the big rip problem does not arises and the scale factor is found to be regular for all time.'
---
\
\
\
PACS: 98.80Cq, 98.80 JK\
Keywords : Dissipative effect; Phantom fluid; Big rip; Accelerated universe\
Recent observations like CMB anisotropy, supernova and galaxy clustering strongly indicate that our universe is spatially flat and there exists an exotic cosmic fluid called dark energy with negative pressure, which constitutes about 70 percent of the total energy of universe. The dark energy is usually described by an equation of state (EoS) parameter $\omega\equiv\frac{p}{\rho}$, the ratio of spatially homogeneous dark-energy pressure p to its energy density $\rho$. A value $\omega<-\frac{1}{3}$ is required for cosmic acceleration. The simplest explanation for dark energy is cosmological constant, for which $\omega=-1$. The increasing evidence from observational data indicates that $\omega$ lies in a narrow strip around $\omega=-1$ quite likely being less than this value [@ref1]$-$[@ref3]. The region where EoS parameter $\omega<-1$, is typically referred to as a phantom dark energy universe. The existence of the region with $\omega<-1$ opens up a number of fundamental questions. For instance, the entropy of such universe is negative. The dominant energy condition (DEC) for phantom fluid is violated, as a rule. The phantom dominated universe end up with a finite time future singularity called big rip or cosmic doomsday [@ref4; @ref5]. The last property attracted much attention and brought the number of speculations upto the explicit calculation of the rest of the life-time of our universe.\
Soon after Caldwell [@ref4] proposed phantom dark energy model with cosmic doomsday of future universe, cosmologists started making efforts to avoid this problem using $\omega<-1$ [@ref6]$-$[@ref8]. In the braneworld scenario, Sahni and Shtanov has obtained well-behaved expansion for the future universe without big rip problem with $\omega<-1$. They have shown that acceleration is a transient phenomenon in the current universe and the future universe will re-enter matter dominated decelerated phase [@ref9]. It is found that general relativity (GR) based phantom model encounters “sudden future singularity” leading a divergent scale factor, energy density and pressure at finite time $t=t_{s}$. Thus the classical approach to phantom model exhibits big rip problem. For future singularity model, curvature invariant becomes very strong and energy density is very high near $t=t_{s}$ [@ref10]. So, quantum effects should be dominated for $|t=t_{s}|<$ one unit of time (Early universe) [@ref11]$-$[@ref13] and it is shown that the an escape from the big rip is possible on making quantum corrections to the energy density and pressure in Friedmann equations.\
In the framework of Robertson-Walker cosmology, Chaplygin gas (CG) is also considered as a good source of dark energy for having negative pressure, given as $$\label{eq1}
p=-\frac{A}{\rho}$$ where p and $\rho$ are, respectively, pressure and energy density in a co moving reference frame, with $\rho>0$; A is a positive constant.\
Moreover, it is only gas having super-symmetry generalization [@ref14]$-$[@ref16]. Bertolami et al [@ref17] have found that generalized Chaplaincy gas (CG) is better fit for latest Supernova data. In case of CG, equation (\[eq1\]) is modified as $$\label{eq2}
p=-\frac{A}{\rho^\frac{1}{\alpha}}$$ where $1\leq\alpha<\infty$.\
For $\alpha=1$, equation (\[eq2\]) corresponds to equation (\[eq1\]).\
Other approaches have considered dissipative effects in CG models, using the framework of Elkhart theory [@ref18]. Thai et al [@ref19], have investigated a viscous GCG, assuming that there is a bulk viscosity in a linear borotropic fluid and GCG. It is found that the equation of state of GCG can cross the boundary $\omega=-1$. Also in Ref. [@ref20], it is found that a dissipative chaplygin gas can give rise to structurally stable evaluational scenarios. It is interesting to note that the GCG itself can behave like a fluid with viscosity in the context of Eckart formalism [@ref18]. Fabris et al [@ref21], have investigated an equivalence GCG and dust like fluid. Recently Cruz et al [@ref22], have studied dissipative generalized chaplygin gas as phantom dark energy and found the cosmological solutions for GCG with bulk viscosity.\
The FRW metric for an homogeneous and isotropic flat universe is given by $$\label{eq3}
ds^2=-dt^2+a(t)^2\left(dx^2+dy^2+dz^2\right)$$ where a(t) is the scale factor and t represents the cosmic time.\
The field equations in the presence bulk viscous stresses are $$\label{eq4}
\frac{a_{4}^2}{a^2}=H^2=\frac{\rho}{3}$$ $$\label{eq5}
\frac{a_{44}}{a}=-\frac{1}{6}\left(\rho+3\bar{p}\right)$$ where $\bar{p}$ is the effective pressure given by $$\label{eq6}
\bar{p}=p-3H\xi$$ Here p, $\xi$ are the isotropic pressure and bulk viscous coefficient respectively.\
The energy conservation equation is given by $$\label{eq7}
\rho_{4}+3H\left(\rho+\bar{p}\right)=0$$ Here, and in what follows the sub indices 4 on $a$, $\rho$ and elsewhere denote differentiation with respect to t.\
Using equations (\[eq2\]), (\[eq4\]) and (\[eq6\]), equation (\[eq7\]) leads to $$\label{eq8}
\rho_{4}+3\frac{a_{4}}{a}\left(\rho-\frac{A}{\rho^\frac{1}{\alpha}}-3H\xi\right)=0$$ In order to obtain solution of equation (\[eq8\]), we will assume that the viscosity has a power-law dependence upon the density $$\label{eq9}
\xi=\xi_{0}\rho^n$$ where $\xi_{0}$ and n are constant.\
On using equation (\[eq9\]) in equation (\[eq8\]), we obtain $$\label{eq10}
\frac{d\rho}{dt}+\frac{3}{a}\frac{da}{dt}\frac{\left(\rho^\frac{1+\alpha}{\alpha}-A\right)}{\rho^\frac{1}{\alpha}}=3\rho^{n+1}\xi_{0}$$ To solve equation (\[eq10\]), we use the transformation $\rho_{4}=f(\rho)=\rho^{n+1}$, accordingly equation (\[eq10\]) leads to $$\label{eq11}
\rho^\frac{1+\alpha}{\alpha}(t)=A+\left(\rho_{0}^\frac{1+\alpha}{\alpha}-A\right)\left[\frac{a_{0}}{a(t)}\right]^\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}$$ where $\rho_{0}=\rho(t_{0})$ and $a_{0}=a(t_{0})$; $t_{0}$ is the present time.\
In the present model, it is assumed that the dark energy behaves like GCG, obeying equation (\[eq2\]) as well as fluid with equation of state $$\label{12}
p=\omega\rho$$ with $\omega<-1$ simultaneously.\
From equations (\[eq2\]) and (12), we obtain $$\label{eq13}
\omega(t)=-\frac{A}{\rho(t)^\frac{1+\alpha}{\alpha}}$$ So, evolution of equation (\[eq13\]) at $t=t_{0}$ leads to $$\label{eq14}
A=-\omega_{0}\rho_{0}^\frac{1+\alpha}{\alpha}$$ with $\omega_{0}=\omega(t_{0})$.\
Using equation (\[eq14\]), equation (\[eq11\]) leads to $$\label{eq15}
\rho=\rho_{0}\left[-\omega_{0}+(1+\omega_{0})\left(\frac{a_{0}}{a(t)}\right)^\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]^\frac{\alpha}{1+\alpha}$$ In the homogeneous model of universe, a scalar field $\phi(t)$ with potential $V(\phi)$ has energy density $$\label{eq16}
\rho_{\phi}=\frac{1}{2}\phi_{4}^2 + V(\phi)$$ and pressure $$\label{eq17}
p_{\phi}= \frac{1}{2}\phi_{4}^2 - V(\phi)$$ Equation (\[eq16\]) and (\[eq17\]) lead to $$\label{eq18}
\phi_{4}^2=\rho_{\phi}+p_{\phi}$$ Using equations (\[eq2\]), (12) and (\[eq14\]), equation (\[eq18\]) reduces to $$\label{eq19}
\phi_{4}^2=\frac{\rho^\frac{(1+\alpha)}{\alpha}+\rho_{0}^\frac{(1+\alpha)}{\alpha}\omega_{0}}
{\rho^\frac{1}{\alpha}}$$ Equation (\[eq15\]) and (\[eq19\]) lead to $$\label{eq20}
\phi_{4}^2=\frac{(1+\omega_{0})\rho_{0}\left(\frac{a_{0}}{a(t)}\right)^\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}}{\left[-\omega_{0}+(1+\omega_{0})\left(\frac{a_{0}}{a}\right)^\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]^\frac{1}{1+\alpha}}$$ From equation (\[eq20\]), it is clear that for $1+\omega_{0} > 0$, $\phi_{4}^2 > 0$, giving positive kinetic energy and for $1+\omega_{0} < 0$, $\phi_{4}^2 < 0$, giving negative kinetic energy. $1+\omega_{0} > 0$ and $1 + \omega_{0} < 0$ are representing the case of quintessence and phantom fluid dominated universe respectively. Similar result are obtained by Hoyle and Narlikar in C -field with negative kinetic energy for steady state theory of universe [@ref23].\
Now, from equation (\[eq4\]) and (\[eq15\]), we obtain $$\label{eq21}
\frac{a_{4}^2}{a^2}=\Omega_{0}H_{0}^2\left[|\omega_{0}|+(1-|\omega_{0}|)\left(\frac{a_{0}}{a(t)}\right)^
\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]^\frac{\alpha}{1+\alpha}$$ where $|\omega_{0}|=-\omega$, $H_{0}=100$hkm/s Mpc, present value of the Hubble’s parameter and $\Omega_{0}=\frac{\rho_{0}}{\rho_{cr,0}}$ with $\rho_{cr,0}=\frac{3H_{0}^2}{8\pi G}$.\
Equation (\[eq21\]) may be written as $$\label{eq22}
\frac{a_{4}}{a}=\sqrt{\Omega_{0}}H_{0}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}\left[1+\frac{(1-|\omega_{0}|)}{|\omega_{0}|}\left(\frac{a_{0}}{a(t)}\right)^
\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]^\frac{\alpha}{2(1+\alpha)}$$ Neglecting the higher powers of $\frac{(1-|\omega_{0}|)}{|\omega_{0}|}\left(\frac{a_{0}}{a(t)}\right)^
\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}$, equation (\[eq22\]) leads to $$\label{eq23}
\frac{a_{4}}{a}=\sqrt{\Omega_{0}}H_{0}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}\left[1+\frac{\alpha(1-|\omega_{0}|)}{2(1+\alpha)|\omega_{0}|}\left(\frac{a_{0}}{a(t)}\right)^
\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]$$
![The plot of scale factor $(a)$ and horizon distance $(d_{H})$ versus time $(t)$.[]{data-label="fg:anil29fig1.eps"}](anil29fig1.eps){height="8"}
![The plot of Hubble distance $(H^{-1})$ versus time $(t)$.[]{data-label="fg:anil29fig2.eps"}](anil29fig2.eps){height="8"}
Integrating equation (\[eq23\]), we obtain $$a(t)=\left(\frac{a_{0}}{\left(2(1+\alpha)|\omega_{0}|\right)}\right)^\frac{\alpha(1-3\xi_{0})}{3(1+\alpha)}\times$$ $$\label{eq24}
\left[
(\alpha+2(1+\alpha)|\omega_{0}|)e^{6H_{0}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}\sqrt{\Omega_{0}}(t-t_{0})}
-\alpha(1-|\omega_{0}|)\right]^\frac{\alpha(1-3\xi_{0})}{3(1+\alpha)}$$ From equation (\[eq24\]), it is clear that as $t\rightarrow \infty$, $a(t)\rightarrow\infty $ which is supported by recent observation of Supernova Ia [@ref24; @ref25] and WMAP [@ref26; @ref27]. Therefore the present model is free from finite time future singularity.\
Now the horizon distance is obtained as $$d_{H}(t)=\frac{3(1+\alpha)a(t)}{\alpha(1-3\xi_{0})a_{0}}\left(\frac{2(1+\alpha)|\omega_{0}|}{\alpha+(\alpha+2)|\omega_{0}|}\right)^\frac{\alpha(1-3\xi_{0})}{3(1+\alpha)}\times$$ $$\label{eq25}
e^{\left[6H_{0}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}\sqrt{\Omega_{0}}\frac{\alpha(1-3\xi_{0})t}{3(1+\alpha)}\right]}$$ **Fig. 1** depicts the variation of scale factor $(a)$ and horizon distance $(d_{H})$ versus cosmic time, as a representative case with appropriate choice of constants and other physical parameters. From equation (\[eq24\]) and (\[eq25\]) it is clear that $d_{H}(t)>a(t)$ i. e. horizon grows more rapidly than scale factor which is clearly shown in **Fig. 1**.\
In this case, the Hubble distance is given by $$\label{eq26}
H^{-1}=\frac{1}{\sqrt{\Omega_{0}}H_{0}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}}\left[1-\frac{\alpha(1-|\omega_{0}|)}{2(1+\alpha)|\omega_{0}|}\left(\frac{a_{0}}{a(t)}\right)^
\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]$$ Equation (\[eq26\]) is showing the growth of Hubble distance $(H^{-1})$ with time such that $H^{-1}\rightarrow \frac{1}{H_{0}\sqrt{\Omega_{0}}|\omega_{0}|^\frac{\alpha}{2(1+\alpha)}}\neq 0$ as $t\rightarrow\infty$. This behaviour of $H^{-1}$ is clearly depicted in **Fig. 2**. Thus in present case, the galaxies will not disappear when $t\rightarrow\infty$, avoiding big rip singularity. Therefore, one can conclude that if phantom fluid behaves like GCG and fluid with $p=\omega\rho$ simultaneously then the future accelerated expansion of universe will free from catastropic situation like big rip. Equation (\[eq15\]) may be written as $$\label{eq27}
\rho=\rho_{0}\left[|\omega_{0}|+(1-|\omega_{0}|)\left(\frac{a_{0}}{a(t)}\right)^\frac{3(1+\alpha)}{\alpha(1-3\xi_{0})}\right]^\frac{\alpha}{1+\alpha}$$ From equation (\[eq27\]), it is clear that as $t\rightarrow \infty$, $\rho\rightarrow\rho_{0}|\omega_{0}|^\frac{\alpha}{1+\alpha}>\rho_{0}$ (since $t\rightarrow \infty$, $a(t)\rightarrow\infty $). Thus one can conclude that energy density increases with time, contrary to other phantom models having future singularity at $t=t_{s}$ [@ref4; @ref11]. Based on Ia Supernova data, singh et al [@ref28] have estimated $\omega_{0}$ for model in the range $-2.4<\omega_{0}<-1.74$ up to $95$ percent confidence level. Taking this estimate as an example, with $\alpha=1$, $\rho_{\infty}$ is found in the range $1.31\rho_{0}<\rho_{\infty}<1.54\rho_{0}$ and with $\alpha=2$, $\rho_{\infty}$ is found in the range $1.44\rho_{0}<\rho_{\infty}<1.78\rho_{0}$ and so on. This does not yields much increase in energy density as $t\rightarrow\infty$ but if the future experiments supports large value of $|\omega_{0}|$ then $\rho_{\infty}$ will be high.\
It is interesting to see here that present model is derived by Bulk viscosity in the context of Eckart formalism [@ref18]. We see that big rip problem does not arises in the present model. In Refs [@ref10]$-$[@ref13], for models with future singularity escape from big rip, is demonstrated using quantum corrections in the field equations near $t=t_{s}$. However the present study deals phantom cosmology with accelerated expansion without catastrophic situations using classical approach. It is also seen that the model of future universe presented by Srivastava [@ref29] is a particular case of the model presented in this letter.
Acknowledgements {#acknowledgements .unnumbered}
================
Author would like to thanks The Institute of Mathematical Science (IMSc), Chennai, India for providing facility and support where part of this work was carried out.
[000]{} Hannestad, S., Mortsell, E.: Phys. Rev. D [**66**]{}, 063508 (2002). Weller, J., Lewis, A. M.: Mon. Not. Astron. Soc. [**346**]{}, 987 (2003). Corasaniti, P. S. et al: Phys. Rev. D [**70**]{}, 083006 (2004). Caldwell, R. R.: Phys. Lett. B [**545**]{}, 23 (2002);\
Caldwell, R. R. et al: Phys. Rev. Lett. [**71**]{}, 071301 (2003). Nojiri, S., Odintsov, S. D.: Phys. Rev. D [**70**]{}, 103522 (2004). McInnes, B.: JHEP [**0208**]{}, 029 (2002). Gonzalez-Diaz, P. F.: Phys. Rev. D [**68**]{}, 021303(R) (2003). Onemli, V. K. et al: Class. Quant. Grav. [**19**]{}, 4607 (2002);\
Onemli, V. K. et al: Phys. Rev. D [**70**]{}, 107301 (2004);\
Kahya, E. O., Onemli, V. K., Woodard, R. P.: Phys. Rev. D [**81**]{}, 023508 (2010). Sahni, V., Shtanov, Yu. V.: JCAP [**0311**]{}, 014 (2003). Elizalde, E., Nojiri, S., Odintsov, S. D.: Phys. Rev. D [**70**]{}, 043539 (2004). Nojiri, S., Odintsov, S. D.: Phys. Lett. B [**595**]{}, 1 (2004);\
Nojiri, S., Odintsov, S. D.: Phys. Rev. D [**70**]{}, 103522 (2004). Srivastava, S. K.: Gen. Rel. Grav. [**39**]{}, 241 (2007). Nojiri, S., Odintsov, S. D., Tsujikawa, S.: Phys. Rev. D [**71**]{}, 063004 (2005). Jackiw, R.: physics/0010042. Bento, M. C., Betrolami, O., Sen, A. A.: Phys. Rev. D [**66**]{}, 043507 (2002). Gorini,V.,Kamenshchik, A., Moschella, U.: Phys. Rev. D [**67**]{}, 063509 (2003). Bertolami, O., et al: Mon. Not. R. Astron. Soc. [**353**]{}, 329 (2004). Eckart, C.: Phys. Rev. [**58**]{}, 919 (1940). Zhai, X. H., Xu, Y. D., Li, X. Z.: Int. J. Mod. Phys. D [**15**]{}, 1151 (2006). Szydlowsk, M., Hrycyna, O.: Annals Phys. [**322**]{}, 2745 (2007). Fabris, J. C., Goncalves, S. V. B., Ribeiro, de Sa R.: astro-ph/0503362. Cruz, N., Lepe, S., Pena, F.: Phys. Lett. B [**646**]{}, 177 (2007). Hoyle, F., Narlikar, J. V.: Mon. Not. R. Astron. Soc. [**108**]{}, 372 (1948)\
Hoyle, F., Narlikar, J. V.: Mon. Not. R. Astron. Soc. [**155**]{}, (1972) 305\
Hoyle, F., Narlikar, J. V.: A Different Approach to Cosmology, Cambridge University Press, Cambridge, 2000. Perlmutter, S., et al: Astrophys. J. [**517**]{}, 565. (1999) Riess, A. G., et al: Astron. J. [**116**]{}, 1009 (1998). Spergel, D. N., et al: Astrophys. J. Suppl. [**148**]{}, 175 (2003). Page, L., et al: Astrophys. J. Suppl. [**148**]{}, 233 (2003). Singh, P., et al: Phys. Rev. D [**68**]{}, 023522 (2003). Srivastava, S. K.: Phys. Lett. B [**619**]{}, 1 (2005).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the context of the LENS R&D solar neutrino project, the $\gamma$ spectrum of a sample of metallic indium was measured using a single experimental setup of 4 HP-Ge detectors located underground at the Gran Sasso National Laboratories (LNGS), Italy. A $\gamma$ line at the energy (497.48$\pm$0.21) keV was found that is not present in the background spectrum and that can be identified as a $\gamma$ quantum following the $\beta$ decay of [$^{115}$In ]{}to the first excited state of $^{115}$Sn ($\frac{9}{2}^{+} \rightarrow \frac{3}{2}^{+}$). This decay channel of $^{115}$In, which is reported here for the first time, has an extremely low $Q_{\beta}$-value, keV, and has a much lower probability than the well-known ground state-ground state transition, being the branching ratio $b = (1.18 \pm 0.31) \cdot 10^{-6}$. This could be the $\beta$ decay with the lowest known $Q_{\beta}$-value. The limit on charge non-conserving $\beta$ decay of $^{115}$In is set at 90% C.L. as $\tau_{CNC} > 4.1\cdot10^{20}$ y.'
address:
- 'INFN, Laboratori Nazionali del Gran Sasso, S.S. 17 bis km 18+910, I-67010 L’Aquila, Italy'
- 'INFN Milano, Via Celoria 16, I-20133 Milano, Italy'
- 'Università dell’Aquila, Dipartimento di Fisica, Via Vetoio 1, I-67010 L’Aquila, Italy'
- 'Institute for Nuclear Research, MSP 03680, Kiev, Ukraine'
author:
- 'C. M. Cattadori'
- 'M. De Deo'
- 'M. Laubenstein'
- 'L. Pandola'
- 'V. I. Tretyak'
title: 'Observation of $\beta$ decay of [$^{115}$In ]{}to the first excited level of [$^{115}$Sn ]{}'
---
,
,
,
,
Indium ,$\beta$ decay ,$\gamma$ spectroscopy 23.20.Lv ,23.40.-s ,27.60.+j
Introduction
============
[$^{115}$In ]{}has been envisaged in the LENS project (Low Energy Neutrino Spectroscopy) as a possible target for the real-time measurement of low energy solar neutrinos. The detection principle is based on the inverse electron capture (EC) reaction $^{115}$In($\nu_{e}$,e$^{-}$)$^{115}$Sn$^{*}$ on $^{115}$In, which has a threshold energy of 114 keV and populates the second excited state of [$^{115}$Sn ]{}at 613 keV (see Fig. \[levels\] later). This state is metastable, with a life time $\tau$ = 4.76 $\mu$s, and its subsequent two-step decay to the ground state provides an highly specific $\nu$ signature [@LENS]. Namely, the prompt electron emitted in the inverse-EC reaction ($e_{1}$) is followed, with a typical delay of $\tau$=4.76 $\mu$s, by a localized deposition of 116 keV ($e/\gamma$)$_{2}$, in spatial coincidence, and by a $\gamma$-ray ($\gamma_{3}$) of 497 keV (see Fig. \[levels\]). Though [$^{115}$In ]{}has several favorable features as a target for low energy solar neutrinos (high isotopic abundance, low threshold, strong $\nu$-tag), the detection technique is extremely challenging because [$^{115}$In ]{}is unstable and can $\beta$-decay directly to the [$^{115}$Sn ]{}ground state. The specific activity of natural indium is 0.25 Bq/g and thus indium itself is the major irremovable source of background. The LENS R&D project has demonstrated that the 10$^{6}$ background suppression factor needed for the $\emph{pp}$ solar neutrino measurement can be achieved, however at the cost of a very high segmentation ($\sim 10^{5}$ cells) and of an one-order-of-magnitude increase of the overall detector mass [@LENS].\
The most serious background that has to be faced is related to the coincidence of two spatially-close indium decays. The coincidence of two [$^{115}$In ]{}decays can mimic the $\nu$-tag if the second $\beta$ has energy close to the end point and emits a hard bremsstrahlung $\gamma$. The importance of this source of background depends on the granularity of the detector (because of the requirement of spatial coincidence), on the energy resolution and on the indium bremsstrahlung spectrum. In order to have a better comprehension and characterization of the bremsstrahlung, a measurement of the $\gamma$ spectrum of an indium sample was performed with HP-Ge detectors in the Low-Level Background Facility underground in the Gran Sasso National Laboratories, Italy. During the analysis of the data we found evidence for a previously unknown decay of [$^{115}$In ]{}to the first excited level of $^{115}$Sn, at the energy 497.4 keV.\
In Section \[sectuno\] we describe the indium sample and the experimental details of the measurement. In Section \[sectdue\] we present the spectrum and the evidence for the $\gamma$ line at 497 keV. In Section \[secttre\] we discuss the interpretation of this line in terms of decay of [$^{115}$In ]{}to the first excited state of [$^{115}$Sn ]{}and the consequent limit set on the [$^{115}$In ]{} charge non-conserving decay. Finally, in Section \[sectquattro\] we summarize our results and briefly describe some possible future perspectives.
The setup {#sectuno}
=========
The sample and the detector
---------------------------
The sample used for the measurement consists of an ingot of metallic $^{nat}$In of mass (928.7$\pm$0.1) g. It has the shape of a cylindrical shell, with the approximate dimensions of 2.0 cm (internal diameter), 5.5 cm (external diameter) and 6.5 cm (height). The high-purity indium (6N5 grade) used for the production of the sample has been provided by the Indium Corporation of America in May 2002.\
The $\gamma$ spectrum of the indium source was measured using a set of 4 HP-Ge detectors installed underground at LNGS. The detectors are very similar coaxial germanium crystals mounted altogether in one cryostat made by Canberra; their main parameters are displayed in Table \[detectors\]. They are arranged as shown in Fig. \[setupplot\] and surround the indium sample, which is placed in the central well. The experimental setup is enclosed in a lead/copper passive shielding and has a nitrogen ventilation system against radon.\
One measurement of the indium ingot (2762.3 h of counting time for each detector) and one of the background (1601.0 h) were carried out, both with the complete shielding around the detectors.
-------------------------------------- ------- ------- ------- -------
ge178 ge179 ge180 ge188
Volume (cm$^{3}$) 225.2 225.0 225.0 220.7
Endcap and holder material
Energy resolution (FWHM) at 1332 keV 2.1 2.0 2.0 2.0
-------------------------------------- ------- ------- ------- -------
: Germanium detectors parameters[]{data-label="detectors"}
Detector performances {#MC}
---------------------
The detectors show a very good energy resolution and linearity. The efficiency of the configuration (detectors and ingot) has been evaluated using the <span style="font-variant:small-caps;">Geant4</span>-based code <span style="font-variant:small-caps;">Jazzy</span>, developed by O. Cremonesi[^1]. The computed full energy efficiency, i.e. the probability that a $\gamma$ ray produced in a random position of the indium ingot deposits its full energy in one of the four Ge detectors, is displayed in Fig. \[effplot\] as a function of energy.
The Monte Carlo code has been checked, with the procedure described in Ref. [@neodimio], using a previous measurement performed in the same setup with a $^{60}$Co $\gamma$ source (1173 keV and 1332 keV). Taking into account the 2505 keV sum peak in the experimental spectrum and neglecting the angular correlations of the two photons, the measured absolute efficiencies in this configuration agree with the computed ones at 12$\%$ and are consistent within their statistical uncertainities. On the basis of our previous experience with similar simulations of analogous experimental setups we estimate the systematic uncertainty on the Monte Carlo efficiencies to be 10%. In any case, the knowledge of the absolute efficiency at the few $\%$ level is not needed for the present analysis, as it does not represent the main contribution to the uncertainty on the measured [$^{115}$In ]{}$\rightarrow$ $^{115}$Sn$^{*}$ decay rate (see Sect. \[secttre\]). For this reason, we can conclude that our knowledge of the detector and of the absolute efficiency is satisfactory for our purposes.
Analysis of the measured spectra {#sectdue}
================================
Comparison of indium and background spectra {#spectrum_des}
-------------------------------------------
The spectra obtained by the four detectors for the indium sample and for the background have been added (see Fig. \[brem\]) and analyzed using both an automatic program and visual inspection. In the indium measurement, besides the continuous component (see Fig. \[brem\]a) due to the bremsstrahlung of the electrons from the $\beta$ decay of [$^{115}$In ]{}(end point: 499 keV), 42 $\gamma$ lines with energy above 200 keV could be identified: they are listed, together with their interpretation, in Table \[gammatable\]. The same $\gamma$ lines, except the one at (497.48$\pm$0.21) keV (which represents the main point of this paper and will be discussed in detail in the following sections), are also found in the background spectrum, though some of them (e.g. 795, 1588 and 2447 keV) have a poor statistical significance because of the shorter measurement time. As shown in Table \[gammatable\], the observed $\gamma$ lines come from the natural radionuclides and radioactive series ($^{40}$K, $^{238}$U, $^{235}$U, $^{232}$Th) and from cosmogenic or antropogenic nuclides ($^{60}$Co, $^{137}$Cs, $^{207}$Bi, $^{26}$Al) that are usually present as contaminations in normal copper and lead. The same Table shows the counting rates of the $\gamma$ lines for the indium sample and the background, as well as their difference (statistical errors only). For each line (except the one at 497 keV) the difference turns out to be statistically consistent with zero. Hence there is no statistical evidence of radioactive contamination of the indium sample, since the data are consistent with contaminations of the experimental setup only (germanium crystals and passive lead/copper shielding).\
The efficiencies quoted in Table \[gammatable\] (full energy efficiencies) have been calculated in the hypothesis that the $\gamma$ rays are generated inside the indium sample so they include the effect of self-absorption in the ingot itself. The fact that in some cases the indium-background difference rate is negative (though statistically consistent with zero) is explained because indium is an effective $\gamma$ ray absorber (the atomic number $Z$ is 49) and can hence act as an additional shielding for the Ge detectors with respect to the background measurement, where the well is filled with a plexiglas plug.
The 497 keV line {#luciano}
----------------
As anticipated in Sect. \[spectrum\_des\], the only $\gamma$ line of the indium spectrum which is not present in the background measurement and cannot be ascribed to the usual radioactive contaminants is located at the energy of (497.48$\pm$0.21) keV. The interesting region of the spectrum is shown in Fig. \[lineplot\].
From the fit of the indium spectrum in the energy region (487-508) keV with a Gaussian peak and linear background assumption, we get a net area of (90$\pm$22) counts, inconsistent with zero at more than 4$\sigma$. This corresponds to a counting rate of (0.78$\pm$0.19) counts/day. Variations of the energy interval for the fit result in changes of the area inside the quoted uncertainty. With the same procedure applied to the background spectrum, no Gaussian peak could be found and the resulting area is (0$\pm$14) counts; the corresponding upper limit derived with the Feldman-Cousins method [@feld] is 23.0 counts (90$\%$ CL) or 0.34 counts/day (90$\%$ CL).\
It can hence be concluded that the peak under examination is statistically significant and related with the indium sample, being absent in the background measurement.
Interpretation {#secttre}
==============
Decay of $^{115}$ to the first excited state of $^{115}$ {#betaexc}
--------------------------------------------------------
The peak with the energy (497.48$\pm$0.21) keV found in the indium measurement and absent in the background one can be explained with the $\beta$ decay of [$^{115}$In ]{}to the first excited level of $^{115}$Sn, whose excitation energy is 497.4 keV. Such a process has never been observed previously and the $\beta$ decay of [$^{115}$In ]{}was considered up-to-date as going exclusively to the ground state of [$^{115}$Sn ]{}[@isotopes; @blachot; @audi1]. Because of the large change in the nuclear angular momentum ($\frac{9}{2}^{+} \rightarrow \frac{1}{2}^{+}$), the ground state to ground state $\beta$ decay of [$^{115}$In ]{}is a 4$^{th}$ forbidden transition and has one of the largest known $\log ft$ values ($\log ft =$ 22.5). The measured half life of [$^{115}$In ]{}is $t_{1/2} = 4.41 \cdot 10^{14}$ y [@isotopes; @blachot; @audi1; @Pfe79].\
According to the most recent table of atomic masses [@audi2], the mass difference between [$^{115}$In ]{}and [$^{115}$Sn ]{}is (499$\pm$4) keV. The decay [$^{115}$In ]{}$\rightarrow ^{115}$Sn$^{*}$ to the first excited level of [$^{115}$Sn ]{}is hence kinematically allowed, though with an extremely small $Q_{\beta}$ value, keV.\
Using the area of the 497 keV peak observed in the indium spectrum, the decay rate for the transition to the first excited level of [$^{115}$Sn ]{}can be evaluated through the relation $$\Gamma (^{115}\mbox{In} \rightarrow ^{115}\mbox{Sn}^{*}) \ = \ \frac{S (1+\alpha)}{N \
\varepsilon \ t_{M}},$$ where $\varepsilon$ is the efficiency to detect the full energy $\gamma$ with the 4 Ge detectors, $N$ is the number of [$^{115}$In ]{}nuclei in the sample, $t_{M}$ is the measurement time, $S$ is the area of the peak and $\alpha$ is the coefficient of conversion of $\gamma$ quanta to electrons for the given nuclear transition.\
The full peak efficiency at 497 keV is $\varepsilon$ = (3.32$\pm$0.33)% and it was evaluated using the Monte Carlo simulation described in Sect. \[MC\]. Taking into account the total mass of the indium sample (928.7 g), the atomic weight of indium (114.8 g$\cdot$mol$^{-1}$) [@weights] and the isotopic abundance of [$^{115}$In ]{}(95.7%) [@chem], the number of [$^{115}$In ]{}nuclei in our sample results to be $N$=4.66$\cdot 10^{24}$. The area of the peak is (90$\pm$22) counts (see Sect. \[luciano\]) and the electron conversion coefficient for the transition is $\alpha$=8.1$\cdot 10^{-3}$ [@blachot]. With these values and $t_{M}$=2762.3 h, the decay rate for the [$^{115}$In ]{}$\beta$ decay to the first excited level of $^{115}$Sn is estimated to be $$\Gamma (^{115}\mbox{In} \rightarrow ^{115}\mbox{Sn}^{*}) \ =
\ (1.86 \pm 0.49) \cdot 10^{-21} \ \ \textrm{y}^{-1},$$ that corresponds to a partial half life of $$t_{1/2} (^{115}\mbox{In} \rightarrow ^{115}\mbox{Sn}^{*}) \ =
\ \frac{\ln 2}{\Gamma} \ = \ (3.73 \pm 0.98) \cdot 10^{20} \ \ \textrm{y}.$$ The probability of this process is thus near one million times lower than for the transition to the ground state of [$^{115}$Sn ]{}(see Fig. \[levels\]); the experimental branching ratio is $$b \ = \ (1.18 \pm 0.31) \cdot 10^{-6}.$$ The uncertainty on the decay rate and on the branching ratio mainly comes from the statistical error on the net area of the 497 keV peak.
Nuclear spin and parity are changed in the observed transition from the initial $\frac{9}{2}^{+}$ of the [$^{115}$In ]{}ground state to $\frac{3}{2}^{+}$ of $^{115}$Sn$^*$; this is therefore a 2$^{nd}$ forbidden unique $\beta$ decay. The recommended value of $\log ft$ for such a decay is equal to (11.7$\pm$0.9) [@Akh88]. With the measured value of the half life of $(3.73 \pm 0.98) \cdot 10^{20}$ y, the “experimental” $\log f$ value can be calculated as $\log f = (-16.37 \pm 0.91)$. On the other hand, the $\log f$ value can be estimated using the formalism described in [@Dzh72]: $$f = f_0 \ s_2,$$ where $f_0$ is the value for the allowed transition and $s_2$ is the correction for the 2$^{nd}$ forbidden unique decay. The $f_0$ factor can be written as $$f_0 = \phi_0 s_0,$$ where $\phi_0$ is the phase space integral $$\phi_0 = \int_1^{W_0} W \sqrt{W^2-1} (W_0-W)^2 dW,$$ $W_0$ is the energy release in the $\beta$-decay in units of the electron mass and $s_0$ is the correction for a non-zero $Z$ value of the daughter nucleus. The phase space integral can be analytically calculated as $$\phi_0 = \sqrt{W_0^2-1}\frac{2W_0^4-9W_0^2-8}{60} + \frac{W_0}{4}
\ln \Big( W_0+\sqrt{W_0^2-1} \Big),$$ while the $s_0$ values are tabulated in [@Dzh72]. The correction $s_2$ can be estimated following Ref. [@Dav51] as: $$s_2 = \frac{1}{1080}
\Big[ \frac{3}{7}(W_0^2-1)^2 - \frac{26}{105}(W_0^2-1)(W_0-1)
- \frac{2}{105}(W_0-1)^2 \Big].$$ In this way the $\log f$ value for $Q_\beta=1.6$ keV is equal to $-15.78$, which is in agreement with the experimental value of $\log f = (-16.37 \pm 0.91)$.\
The $\log f$ value can also be evaluated with the help of the LOGFT tool at the National Nuclear Data Center, USA [@logft] which is based on procedures described in [@Gov71]. The value of $\log f$ for $Q_\beta=1.6$ keV calculated with the LOGFT code is however very different from the previous estimation: $\log f = -10.8$; this means that with such a $Q_\beta$ the beta decay should go near 6 orders of magnitude faster. One can solve the inverse problem and use the LOGFT code to adjust the $Q_\beta$ value corresponding to the measured $\log f = (-16.37 \pm 0.91)$. Such a procedure gives an extremely low value of $Q_\beta = (23^{+23}_{-12})$ eV.\
While both these calculations can be considered as very tentative and not being intended to be used for such low energies, the last estimation gives a hint that the $Q_\beta$ value in the beta decay $^{115}$In $\rightarrow$ $^{115}$Sn$^*$ can be very close to zero[^2]. Such a unique situation could be used to establish a limit on the antineutrino mass, in addition to the experiments with $^{3}$H and $^{187}$Re, where up-to-date limits are in the range of $\simeq$ 2 eV [@Lob03] and $\simeq$ 15 eV [@Sis04], respectively. To do this in a competitive way both the [$^{115}$In ]{}$Q_\beta$-value and the energy of the [$^{115}$Sn ]{}497.4 keV level should be measured with an accuracy of $\sim$1 eV or better. The uncertainty in the energy of the 497.4 keV level is equal now to 22 eV [@blachot] and it could be further reduced doing an accurate investigation of the $^{115}$Sb decay[^3]. The atomic mass difference [$^{115}$In ]{}– [$^{115}$Sn ]{}can possibly also be measured with an accuracy of $\sim$1 eV. For example, the mass difference of $^{76}$Ge – $^{76}$Se was determined with 50 eV uncertainty in Ref. [@Dou01], where it was stated that the Penning traps technique is able to deliver even more accurate results. Both such measurements require strong experimental efforts but the physical result could be very interesting and important.
Possible imitation of the effect
--------------------------------
In some nuclear processes $\gamma$ rays with energies close to 497 keV are emitted. This could give an alternative explanation of the peak observed in the experimental spectrum. Luckily, additional $\gamma$ rays are also emitted in such decays, allowing to tag those mimicking effects.\
The [$^{115}$In ]{}nucleus has an isomeric state [$^{115m}$In ]{}with the energy $E_{iso}$ = 336.2 keV and a half life of 4.5 h [@isotopes]. With the probability of 0.047% the [$^{115m}$In ]{}nucleus $\beta$-decays to the first excited level of $^{115}$Sn, with the subsequent emission of a 497 keV $\gamma$ ray [@isotopes; @blachot; @audi1]. However, in this case a $\gamma$ ray with the energy $E_{iso}$ = 336.2 keV is emitted with much higher probability (45.84$\%$ [@isotopes]) because of the electromagnetic transition from the isomeric [$^{115m}$In ]{}to the ground [$^{115}$In ]{}state. This huge peak at 336.2 keV, whose area should be $\sim$ 10$^{3}$ times bigger than that of the observed 497.4 keV peak, is absent in the experimental spectrum; only a peak at 338.3 keV is observed, with the net area of (138$\pm$50) counts, which corresponds to the decay of $^{228}$Ac from the $^{232}$Th natural chain (see Table \[gammatable\]). Therefore the decay of the isomeric state [$^{115m}$In ]{}is absolutely negligible and the 497 keV peak cannot be ascribed to it, not even in part. Similarly, given also the underground location of the experimental setup and the low flux of neutrons [@belli], we conclude that $(n,\gamma)$ reactions cannot contribute to the peak under analysis.\
Protons produced by fast neutron or cosmic ray muons can populate the second excited level of [$^{115}$Sn ]{}(Fig. \[levels\]) via the $(p,n)$ reaction on [$^{115}$In ]{}($E_{thr}$=0.9 MeV); the [$^{115}$Sn ]{}nucleus quickly returns to the ground state with the emission of two $\gamma$ rays of energy 115.4 and 497.4 keV. The contribution originated by fast neutrons is practically zero (see f.i. [@gallex1]) because of the deep underground location and the lack of hydrogenous materials in the setup. On the other hand, since the muon flux in the laboratory is extremely low (1 $\mu$/m$^{2}$h [@macro]), also the contribution of $(p,n)$ reactions induced by cosmic rays (see also [@gallex2]) to the 497 keV peak is absolutely negligible ($<$ 10$^{-3}$ counts).\
Some decays from the natural $^{238}$U and $^{232}$Th chains can also give $\gamma$ rays in the energy region of interest, though with very low intensity. They are in particular [@isotopes] $^{214}$Bi ($E$ = 496.7 keV, $I$ = 0.0069%), $^{228}$Ac ($E$ = 497.5 keV, $I$ = 0.0059%) and $^{234m}$Pa ($E$ = 498.0 keV, $I$ = 0.062%). However, the sum contribution of these decays to the 497 keV peak is less than 1 count and can be easily estimated using their stronger associated $\gamma$ lines.[^4]\
We could not figure out other sources than can mimic the experimentally observed 497 keV peak.
Charge non-conserving $\beta$ decay of [$^{115}$In ]{}
------------------------------------------------------
The present measurements give also the possibility to set a limit on the charge non-conserving (CNC) $\beta$ decay of $^{115}$In, a process in which the $(A,Z) \to (A,Z+1)$ transformation is not accompanied by the emission of an electron [@Fei59; @Oku92]. It is supposed that instead of an $e^-$, a massless particle is emitted (for example a $\nu_e$, a $\gamma$ quantum, or a Majoron): $(A,Z) \to (A,Z+1) + (\nu_e$ or $\gamma$ or M) + $\overline{\nu}_e$. Up-to-date, the CNC $\beta$ decay was searched for with four nuclides only: $^{71}$Ga, $^{73}$Ge, $^{87}$Rb and $^{113}$Cd. Only lower limits on corresponding life times $\tau_{CNC}$ were established, in the range of ($10^{18}-10^{26}$) y (we refer to the last work on this subject [@Kli02], where all previous attempts are reviewed). Measuring only the $\gamma$ quanta from the deexcitation of the 497 keV level of $^{115}$Sn, we cannot distinguish, in principle, which mechanism leads to the detected 497 keV peak: the CNC beta decay or the usual charge-conserving beta decay to the excited level of $^{115}$Sn. If the whole area of the observed 497 keV peak is considered as belonging to the CNC beta decay (instead of the much less exotic usual $\beta$-decay of $^{115}$In), and substituting in the formula for the life time $\tau_{CNC} = \varepsilon \cdot N \cdot t_M / [S_{lim}(1+\alpha)]$ the values of efficiency and other parameters described in section \[betaexc\] together with $S_{lim}=118$, the Feldman-Cousins 90% CL limit on the life time of this process is $$\tau_{CNC} \ > \ 4.1 \cdot 10^{20} \ \textrm{y}.$$ Though this value is relatively low, it is determined for the first time for $^{115}$In, expanding the scarce information on the CNC processes.
Conclusions {#sectquattro}
===========
From the measurement of the $\gamma$ spectrum of a sample of metallic indium performed with HP-Ge detectors in the Low-Level Background Facility of the Gran Sasso Laboratory, we have found evidence for the previously unknown $\beta$ decay of [$^{115}$In ]{}to the first excited state of $^{115}$Sn at 497.4 keV ($\frac{9}{2}^{+} \rightarrow \frac{3}{2}^{+}$). The $Q_{\beta}$-value for this channel is [$Q_{\beta}$ = (2$\pm$4)]{} keV; this could be lower than the $Q_{\beta}$-value of $^{163}$Ho, $Q_{\beta}$=2.565 keV [@isotopes; @olmio], and hence be the lowest of all the known $\beta$ decays. The branching ratio is found to be $b = (1.18 \pm 0.31) \cdot 10^{-6}$. We also set a limit on the charge non-conserving $\beta$ decay of [$^{115}$In ]{} $\tau_{CNC} > 4.1\cdot10^{20}$ y (90% CL).\
This measurement was carried out in the context of the LENS R&D project, with the aim of better characterize the [$^{115}$In ]{}bremsstrahlung spectrum, which is poorly known near its end-point. The discovered decay of [$^{115}$In ]{}on the first excited state of [$^{115}$Sn ]{}could in principle be dangerous for the LENS solar neutrino measurement, since it is an irremovable background source of 497 keV $\gamma$ rays. The coincidence of [$^{115}$In ]{}decay on the ground state of [$^{115}$Sn ]{}and one [$^{115}$In ]{}decay on the first excited state can in fact mimic the $\nu$-tag and the possible impact is currently under investigation.\
In a future work we plan to study the possible atomic effects on the half life of the [$^{115}$In ]{}$\rightarrow$ $^{115}$Sn$^{*}$ decay by measuring a new sample where indium is present in a different chemical form (e.g. InCl$_{3}$ solution instead of metallic indium).\
We also point out that, given the extremely low $Q_{\beta}$ value, the decay reported in this work could be used in a future experiment to directly measure the neutrino mass. In order to reduce the background due to the $\beta$ decay of [$^{115}$In ]{} to the ground state of $^{115}$Sn, such experiment would need a rejection power of $10^{6}$, that could be achieved tagging the 497 keV $\gamma$ ray emitted in coincidence with the $\beta$ particle. New In-based semiconductor detectors or fast bolometers could be used for the purpose.
Acknowledgments
===============
The authors would like to thank O. Cremonesi, for providing his Monte Carlo code to estimate the efficiencies, E. Bellotti and Yu.G. Zdesenko, for their continuous support and useful advices, and A. di Vacri, who collaborated in the early stages of this work. We would also like to express our gratitude to D. Motta, for his valuable advices and feedback, and to the LNGS Mechanical Workshop, especially to E. Tatananni and B. Romualdi, for the melting of the indium ingot.
[99]{} R.S. Raghavan, Phys. Rev. Lett. 37 (1976) 359; R.S. Raghavan, hep-ex/0106054; T. Bowles, Talk at VIII International Workshop on Topics in Astroparticle and Underground Physics (TAUP03), Seattle (2003), to appear in the Proceedings C. Arpesella *et al.*, Nucl. Instr. Meth. A 372 (1996) 415 G.J. Feldman and R.D. Cousins, Phys. Rev. D 57 (1998) 3873 R.B. Firestone *et al.*, *Table of Isotopes*, John Wiley & Sons, New York, 8th Edition (1996) and CD update (1998) J. Blachot, Nucl. Data Sheets 86 (1999) 151 G. Audi *et al.*, Nucl. Phys. A 729 (2003) 3 L. Pfeiffer *et al.*, Phys. Rev. C 19 (1979) 1035 G. Audi *et al.*, Nucl. Phys. A 729 (2003) 337 J.R. De Laeter *et al.*, Pure and Appl. Chem. 75 (2003) 683 *Handbook of Chem. Phys.*, 84th Edition, CRC Press (2003-2004) E.Kh. Akhmedov, *Beta decay*, in: A.M. Prokhorov (Ed.), *Physical Encyclopedia*, Vol. 1 (Moscow, Sovetskaya Enciklopediya, 1988), p. 190, in Russian B.S. Dzhelepov *et al.*, *Beta Processes. Functions for analysis of beta-spectra and electron capture* (Leningrad, Nauka, 1972), in Russian J.P. Davidson, Jr., Phys. Rev. 82 (1951) 48 http://www.nndc.bnl.gov/nndc/physco N.B. Gove, M.J. Martin, Nucl. Data Tables 10 (1971) 205 G. Audi, A.H. Wapstra, Nucl. Phys. A 565 (1993) 66 G. Audi, A.H. Wapstra, Nucl. Phys. A 595 (1995) 409 V.M. Lobashev, Nucl. Phys. A 719 (2003) 153 M. Sisti *et al.*, Nucl. Instrum. Meth. A 520 (2004) 125 G. Douysset *et al.*, Phys. Rev. Lett. 86 (2001) 4259 P. Belli *et al.*, Nuovo Cimento 101 A (1989) 959 M. Cribier *et al.*, Astrop. Phys. 4 (1995) 23 MACRO Collaboration, M. Ambrosio *et al.*, Phys. Rev. D 52 (1995) 3793 M. Cribier *et al.*, Astrop. Phys. 6 (1997) 129
G. Feinberg, M. Goldhaber, Proc. Nat. Acad. Sci. U.S.A. 45 (1959) 1301 L.B. Okun in: Particle Data Group, K. Hikasa *et al.*, Phys. Rev. D 45 (1992) VI.10 A.A. Klimenko *et al.*, Phys. Lett. B 535 (2002) 77 F.X. Hartmann and R.A. Naumann, Phys. Rev. C 31 (1985) 1594
[^1]: INFN Milano and Milano Bicocca University.
[^2]: Even the history of the $Q_\beta$ evaluation for $^{115}$In gives some indication for this: the $Q_\beta$ value was slightly lower than energy of the first excited 497.4 keV $^{115}$Sn state in accordance with older tables of atomic masses, $Q_\beta=(495 \pm 4)$ keV [@Aud93] and ($496 \pm 4$) keV [@Aud95], while it is slightly higher in the last evaluation, ($499 \pm 4$) keV [@audi2].
[^3]: It can be noted that energies of many $\gamma$ lines of calibration sources are known with an accuracy of $(0.1-0.3)$ eV, also in the $\sim$500 keV region of our interest [@isotopes].
[^4]: For instance, the area of the 338.3 keV line of $^{228}$Ac, whose relative intensity is 11.27%, is only (138$\pm$50) counts. Therefore, if the contamination were located in the indium ingot, the estimated contribution to the 497 keV peak, taking also into account the different full peak efficiency, would be (7.3$\pm$2.6)$\cdot$10$^{-2}$ counts.
|
{
"pile_set_name": "ArXiv"
}
|
Automorphism groups of simplicial complexes and rigidity for uniformly bounded representations {#automorphism-groups-of-simplicial-complexes-and-rigidity-for-uniformly-bounded-representations .unnumbered}
==============================================================================================
[Juhani Koivisto]{} [^1]
\
*Department of Mathematics and Statistics, University of Helsinki, Finland, email: [email protected]*
#### Abstract.
We consider $L^p$-cohomology of reflexive Banach spaces and give a spectral condition implying the vanishing of 1-cohomology with coefficients in uniformly bounded representations on a Hilbert space.
#### Mathematics Subject Classification (2000): {#mathematics-subject-classification-2000 .unnumbered}
20F65
#### Key words {#key-words .unnumbered}
Fixed point property, cohomology, Banach space, uniformly bounded representation, spectral criterion.
Introduction
============
Since its introduction by David Kazhdan in [@Kazhdan], property $(T)$ and its generalizations as cohomological vanishing has become a fundamental concept in mathematics [@BHV]. The aim of this paper is to extend the framework of W. Ballman and J. Światkowski [@BS] to reflexive Banach spaces and as an application, to give a spectral condition implying vanishing of cohomology for uniformly bounded representations on a Hilbert space. Along with W. Ballman and J. Światkowski, A. Żuk [@Zuk96] was among the first to also consider such criteria for unitary representations, both following fundamental work by H. Garland \[G\]. Since then, extending the spectral method beyond Hilbert spaces has been considered in [@Monod; @Cha; @Dymara; @Ers; @Fish; @L] and by Piotr W. Nowak [@Nowak] extending the spectral method in [@Zuk03] to reflexive Banach spaces. Appropriately extending the scheme in [@Nowak] we similarly extend the spectral condition of [@BS] to uniformly bounded representations on a Hilbert space. Motivation for such generalizations arises, among others, from Shalom’s conjecture [@OWR] stating that any hyperbolic group $\Gamma$ admits a uniformly bounded representation $\pi$ with ${H^1}(\Gamma, \pi) \neq 0$ together with a proper cocycle in $Z^1(\Gamma, \pi)$.\
For a finite graph $K$ with vertices $\mathcal{V}_K$, consider the graph Laplacian $\triangle_+$ on the space of real valued functions on $\mathcal{V}_K$ defined by $$\triangle_+f(v) = f(v)-Mf(v),$$ where $Mf(v)$ is the mean value of $f$ on the vertices adjacent to $v$. Denote by $\lambda_1(K)$ the spectral gap of $\triangle_+$ and its associated Poincaré constant by $\kappa_2(K, \mathbb{R})=\lambda_1(K)^{-1/2}$. More generally $\kappa_2(K, \mathcal{H})= \lambda_1(K)^{-1/2}$ for any separable infinite-dimensional Hilbert space $\mathcal{H}$, [@Nowak].
Let $X$ be a locally finite $2$-dimensional simplicial complex, $\Gamma$ a discrete properly discontinuous group of automorphisms of $X$ and $\pi : \Gamma \rightarrow \mathrm{B}(\mathcal{H})$ a uniformly bounded representation of $\Gamma$ on a separable infinite-dimensional Hilbert space $\mathcal{H}$. If for any vertex $\tau$ of $X$ the link $X_\tau$ is connected and $$\sup_{g \in \Gamma} \Vert \pi_g\Vert < \dfrac{\sqrt{2}}{ \kappa_2(X_\tau, \mathcal{H})},$$ then $L^2H^1(X,\pi) = 0$.
##### Structure of the paper {#structure-of-the-paper .unnumbered}
In Sections \[SU\] to \[PI\] the framework of [@BS] for unitary representations on Hilbert spaces is extended to reflexive Bananch spaces and isometric representations: Section \[SU\] introduces the generalized set up; Section \[projsec\] and \[L\] deal with the dual of the twisted cochains; Section \[DC\] introduces differentials and codifferentials; Section \[secloc\] discusses localization of the problem and Section \[PI\] introduces the spectral set up in terms of Poincaré inequalities and constants on the links. Section \[LH\] introduces $L^p$-cohomology as a natural extension to $L^2$-cohomology, which is then applied to uniformly bounded representations using the fact that they correspond to isometric representations on some reflexive Banach space.
##### Acknowledgements {#acknowledgements .unnumbered}
I would like to thank Piotr W. Nowak for suggesting this topic, invaluable advice, and devotion without which this project would not have been possible. I would also like to thank Väisälä foundation, my advisor Ilkka Holopainen and the ”Analysis, metric geometry and differential and metirc topology” project for financial support, Pekka Pankka and Izhar Oppenheim for helpful discussions and correspondence, and Antti Perälä for many enjoyable conversations on related topics.
Set up {#SU}
======
In this chapter notation is fixed. We recall the notation and some basic facts used by [@BS] for weighted simplicial complexes and extend the notion of square integrable cochains to reflexive Banach spaces and $p > 1$.
Weighted complexes
------------------
Throughout, let $X$ denote an $n$-dimensional locally finite simplicial complex. Following [@BS] we use the following notation: $X(k)$ is the set of (unordered) $k$-simplexes of $X$; $\Sigma(k)$ is the set of ordered $k$-simplexes of $X$. As usual we write $\sigma = \lbrace v_0, \dots, v_k \rbrace$ for a $k$-simplex and $\sigma =(v_0, \dots, v_k)$ for an ordered $k$-simplex. If the vertices of $\tau \in \Sigma(l)$ are vertices of $\sigma \in \Sigma(k)$, we say that $\tau \subset \sigma$, and for $\tau = (v_0, \dots, \hat{v}_i, \dots, v_k)$, i.e. $v_i \notin \tau$, we denote by $[\sigma : \tau] = (-1)^i$ the sign of $\tau$ in $\sigma = (v_0, \dots, v_k)$. As customary, we write $\sigma_i$ for $(v_0, \dots, \hat{v}_i, \dots, v_k)$. In addition to orientation we consider $X$ to be equipped with a weight $\omega$, by which we mean a map from the oriented simplexes of $X$ to the integers such that for $\sigma = (v_0, \dots, v_k) \in \Sigma(k)$, $$\omega(\sigma) = \omega(\lbrace v_0, \dots, v_k \rbrace),$$ where $\omega(\lbrace v_0, \dots, v_k \rbrace)$ denotes the number of $n$-simplexes containing $\lbrace v_0, \dots, v_k\rbrace$. In addition, we assume that $\omega(\sigma) \geq 1$ for every simplex of $X$. Beginning from Section \[secloc\] and onwards, we consider $X$ locally through its links, where, by the link of $\tau = (v_0, \dots, v_l) \in \Sigma(l)$ denoted by $X_\tau$, we mean the $(n-l-1)$-dimensional subcomplex consisting of all simplexes $\lbrace w_0, \dots, w_j\rbrace$ disjoint from $\tau$ such that $\lbrace v_0, \dots, v_l\rbrace \cup \lbrace w_0, \dots, w_j\rbrace$ is a simplex of $X$. Since $X$ is locally finite, $X_\tau$ is finite. Here as previously, $X_\tau(j)$ denotes the $j$-simplexes of $X_\tau$, $\Sigma_\tau(j)$ its oriented $j$-simplexes and so on. In particular, for $\sigma \in \Sigma_\tau(j)$ and $\tau \in \Sigma(l)$ we denote by $\sigma * \tau \in \Sigma(j+l+1)$ the join of $\sigma$ and $\tau$ obtained by juxtaposing the two in that order.\
In addition to the above, we assume throughout that $X$ is a $\Gamma$-space where $\Gamma$ is a discrete topological group acting properly and discontinuously by simplicial automorphisms on $X$. In other words, $\Gamma$ permutes the simplexes of $X$ preserving their order and weights: that is for $\sigma = (v_0, \dots, v_k) \in \Sigma(k)$, $g \cdot \sigma = (g(v_0), \dots, g(v_k)) \in \Sigma(k)$ and $\omega (\sigma) = \omega(g \cdot \sigma)$. As usual, we denote by $\Gamma \sigma$ and $\Gamma_\sigma$ the $\Gamma$-orbit and stabilizer of $\sigma \in \Sigma(k)$, respectively, by $\Sigma(k, \Gamma) \subset \Sigma(k)$ some chosen set of representatives of $\Gamma$-orbits in $\Sigma(k)$, and by $\vert \cdot \vert$ the counting measure on $\Gamma$. In particular since $\Gamma$ is discrete, stabilizers are finite and the Haar measure on $\Gamma$ is $\vert \cdot \vert$. Although the discreteness assumption can be avoided, it will be used when constructing projections in Section \[projsec\]. For the following frequently used facts we refer to [@BS]:
\[combinatorial\] [@BS] Let $n$ be the dimension of $X$. Then, for $\tau \in \Sigma(k)$ $$\sum_{\substack{\sigma \in \Sigma(k+1) \\ \tau \subset \sigma}}\omega(\sigma) = (n-k)(k+2)!\omega(\tau).$$
\[switchingsums\] [@BS] For $0 \leq l < k \leq n$, let $f=f(\tau, \sigma)$ be a $\Gamma$-invariant function on the set of pairs $(\tau, \sigma)$, $\tau \in \Sigma(l)$, $\sigma \in \Sigma(k)$, such that $\tau \subset \sigma$. Then $$\sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{\substack{\tau \in \Sigma(l) \\ \tau \subset \sigma}} \dfrac{f(\tau, \sigma)}{\vert \Gamma_\sigma \vert} = \sum_{\tau \in \Sigma(l, \Gamma)}\sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}}\dfrac{f(\tau, \sigma)}{\vert \Gamma_\tau \vert},$$ whenever either side is absolutely convergent.
More generally, Proposition \[switchingsums\] holds for locally compact unimodular groups [@DJ00] replacing the counting measure with the Haar measure.
Banach space setting, isometric representations and $p$-integrable cochains
---------------------------------------------------------------------------
Throughout, let $(E, \Vert \cdot \Vert_E)$ denote a reflexive Banach space, $\langle \cdot, \cdot \rangle_E$ the natural pairing between $E$ and its continuous dual $E^*$, $\simeq$ isomorphism, $\cong$ isometric isomorphism and $p^*$ the adjoint index of $p$ such that $1/p+1/p^* = 1$. Moreover, let $\pi: \Gamma \rightarrow \mathrm{Iso}(E)$ denote an isometric representation of $\Gamma$ on $E$ where $\mathrm{Iso}(E)$ denotes the group of isometric linear automorphisms on $E$ and by $\bar{\pi}: \Gamma \rightarrow \mathrm{Iso}(E^*)$ its corresponding contragradient representation given by $\bar{\pi}_g = \pi^*_{g^{-1}}$ where $\pi^*$ is the transpose of $\pi$. For combinatorial purposes we also introduce antisymmetrization:
For $n \geq 1$ we denote by $S_n$ the symmetric group of $n$ elements and by $\mathrm{sign} \colon S_n \rightarrow \lbrace -1, 1 \rbrace$ the signature of the permutation: $1$ if $\alpha \in S_n$ is an even permutation of the $n$ elements and otherwise $-1$. For $f: \Sigma(k) \rightarrow E$ define its alternation point-wise as the linear idempotent map $$\mathrm{Alt} f(\sigma) = \dfrac{1}{(k+1)!} \sum_{\alpha \in S_{k+1}} \mathrm{sign}(\alpha) \alpha^*f(\sigma),$$ where $\alpha^*f(\sigma) = f(v_{\alpha(0)}, \dots, v_{\alpha(k)})$ for $\sigma = (v_0, \dots, v_k) \in \Sigma(k).$ As usual, we say that $f$ is alternating if $\mathrm{Alt} f = f$, and symmetric if $\mathrm{Alt} f = 0$.
Replacing inner product with dual pairing and unitary representations by isometric representations, we next introduce twising and cochains as in [@BS].
Let $ \mathcal{E}^{(k,p)}(X,E)$ denote the semi-normed vector space of $k$-cochains $f \colon \Sigma(k) \rightarrow E $ for which the semi norm given by $$\Vert f \Vert_{(k,p)} = \left( \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert f(\sigma) \Vert^p_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma\vert}\right)^{1/p},$$ is finite.
For $f \in {\mathcal{E}^{(k,p)}}(X,E)^*$, we denote by $$\langle \phi, f \rangle_k = \sum_{\sigma \in \Sigma(k, \Gamma)} \langle \phi(\sigma), f(\sigma) \rangle_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma \vert}$$ the dual pairing between $\mathcal{E}^{(k,p)}(X,E)$ and ${\mathcal{E}^{(k,p)}}(X,E)^*$.
\[cochaindual\] ${\mathcal{E}^{(k,p)}}(X,E)^*\cong \mathcal{E}^{(k,p^*)}(X,E^*)$.
Let $f: \Sigma(k) \rightarrow E$. If for every $g \in \Gamma$ and every $\sigma \in \Sigma(k)$ $$f(g \cdot \sigma) = \pi_g \cdot f(\sigma),$$ then we say that $f$ is twisted by $\pi$, or for short just twisted.
Let $C^{(k,p)}(X,E)$ denote the vector space of all alternating maps $f \colon \Sigma(k) \rightarrow E$ twisted by $\pi$.
Those alternating maps twisted by $\pi$ whose $\Vert \cdot \Vert_{(k,p)}$ norm is finite are called $p$-integrable mod $\Gamma$ and we use the following notation:
Let $L^{(k,p)}(X, E) = \lbrace f \in C^{(k,p)}(X,E) \colon \Vert f\Vert_{(k,p)} < \infty \rbrace$ denote the vector subspace of all alternating $k$-cochains of $X$ twisted by $\pi$.
In particular, if $\Gamma$ acts cocompactly on $X$, then $L^{(k,p)}(X,E) = C^{(k,p)}(X,E)$ since then $X/\Gamma$ is compact, the set of representatives $\Sigma(k, \Gamma)$ is finite, and $\Vert f \Vert_{(k,p)} < \infty$ for all $f \in C^{(k,p)}(X,E)$.\
We end this section by proving that $L^{(k,p)}(X, E)$ is a normed space with respect to $\Vert \cdot \Vert_{(k,p)}$. Towards this end we first show that $\Vert \cdot \Vert_{(k,p)}$ is independent of the set of representatives when $f \in L^{(k,p)}(X,E)$.
\[independentofrepresentatives\] If $f \in L^{(k,p)}(X, E)$, then $\Vert f \Vert_{(k,p)}$ is independent of the choice of $\Sigma(k, \Gamma)$.
Let $\Sigma'(k, \Gamma)$ be another set of representatives. Then, $$\begin{aligned}
\Vert f \Vert_{(k,p)}^p &= \sum_{\sigma' \in \Sigma'(k, \Gamma)} \Vert f(\sigma') \Vert^p_E \dfrac{\omega (\sigma')}{(k+1)!\vert \Gamma_{\sigma'} \vert} = \sum_{\sigma' \in \Sigma'(k, \Gamma)} \Vert f(g' \cdot \sigma') \Vert^p_E \dfrac{\omega (g' \cdot \sigma')}{(k+1)!\vert \Gamma_{g' \cdot \sigma'} \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert f(\sigma) \Vert^p_E \dfrac{\omega (\sigma)}{(k+1)!\vert \Gamma_{\sigma} \vert}, \nonumber \end{aligned}$$ choosing for each $\sigma' \in \Sigma'(k, \Gamma)$ a $g' \in \Gamma$ such that $g' \cdot \sigma' = \sigma \in \Sigma(k, \Gamma)$ and observing that $f$ is twisted by $\pi$ and both the norm and $\omega$ are $\Gamma$-invariant.
$L^{(k,p)}(X,E) \subseteq \mathcal{E}^{(k,p)}(X,E)$ is a normed vector space.
It suffices to show that the seminorm $\Vert \cdot \Vert_{(k,p)}$ on $\mathcal{E}^{(k,p)}(X,E)$ restricted to $L^{(k,p)}(X,E)$ is a norm. To this end, suppose $\Vert f \Vert_{(k,p)} = 0$ for $f \in L^{(k,p)}(X,E)$. By Lemma \[independentofrepresentatives\] we may assume $f(\sigma) = 0$ for all $\sigma \in \Sigma(k, \Gamma)$. Since $f(g \cdot \sigma) = \pi_g f(\sigma)$ and the action of $\Gamma$ is transitive on the orbits it follows that $f(\sigma)=0$ for all $\sigma \in \Sigma(k)$.
Projecting $k$-cochains onto $L^{(k,p)}(X,E)$ {#projsec}
=============================================
In order to extend the framework of [@BS], the dual space of the alternating and twisted cochains has at first to be identified up to isometric isomorphism. Following the scheme presented in [@Nowak], we begin by stepwise constructing a continuous projection $P_L$ from $\mathcal{E}^{(k,p)}(X,E)$ onto $L^{(k,p)}(X,E)$.
\[projection1\] Define $\widetilde{P}: \mathcal{E}^{(k,p)}(X,E) \rightarrow \mathcal{E}^{(k,p)}(X,E)$ by $$\widetilde{P}f (\sigma) = \left\{ \begin{array}{ll}
\displaystyle \sum_{s \in \Gamma_\sigma} \pi_s f'(\sigma) & \textrm{if $\sigma \in \Sigma(k, \Gamma)$}\\
\displaystyle \sum_{\substack{h \in \Gamma \\ h \cdot \tau = \sigma}} \pi_{h} f'(\tau) & \textrm{if $\sigma \notin \Sigma(k, \Gamma)$ for $\tau \in \Sigma(k, \Gamma)$},\\
\end{array} \right.$$ where $f': \Sigma(k, \Gamma) \rightarrow E$ is the restriction of $f: \Sigma(k) \rightarrow E$ to $\Sigma(k, \Gamma)$.
This map is well defined, in particular we note that $\lbrace h \in \Gamma \colon h \cdot \tau = \sigma \rbrace = h\Gamma_\tau$. As the following proposition shows, $\widetilde{P}$ maps $k$-cochains to $k$-cochains twisted by $\pi$.
For $f \in \mathcal{E}^{(k,p)}(X,E)$, the $k$-cochain $\widetilde{P}f: \Sigma(k) \rightarrow E$ is twisted by $\pi$.
Let $\sigma \in \Sigma(k)$. Then either $\sigma \in \Sigma(k, \Gamma)$ or $\sigma \notin \Sigma(k, \Gamma)$. Suppose at first $\sigma \in \Sigma(k, \Gamma)$. If $g \in \Gamma_\sigma$, then clearly $\pi_g \widetilde{P}f(\sigma) = \widetilde{P}f(g \cdot \sigma)$. On the other hand, if $g \notin \Gamma_\sigma$ we get $$\widetilde{P}f(g \cdot \sigma) = \sum_{\substack{h \in \Gamma \\ h \cdot \sigma = g \cdot \sigma}} \pi_hf'(\sigma),$$ and $$\pi_g\widetilde{P}f(\sigma) = \sum_{h \in g \Gamma_\sigma} \pi_h f'(\sigma).$$ But $\lbrace h \in \Gamma \colon h \cdot \sigma = g \cdot \sigma \rbrace = \lbrace h \in \Gamma \colon h \in g \Gamma_\sigma \rbrace$, so the claim holds for $\sigma \in \Sigma(k, \Gamma)$. Suppose $\sigma \notin \Sigma(k, \Gamma)$. If $g \cdot \sigma \in \Sigma(k, \Gamma)$, then $$\begin{aligned}
\widetilde{P}f(g \cdot \sigma) &= \sum_{s \in \Gamma_{g \cdot \sigma}} \pi_s f'(g \cdot \sigma) = \sum_{s \in g\Gamma_{\sigma}g^{-1}} \pi_s f'(g \cdot \sigma) = \sum_{h \in \Gamma_\sigma} \pi_{ghg^{-1}}f'(g \cdot \sigma) \nonumber, \end{aligned}$$ as $\Gamma_{g \cdot \sigma} = g \Gamma_\sigma g^{-1}$, and so $$\begin{aligned}
\pi_g\widetilde{P}f(\sigma) &= \sum_{\substack{h \in \Gamma \\ hg \cdot \sigma = \sigma}} \pi_{gh}f'(g \cdot \sigma) = \sum_{\substack{h \in \Gamma \\ hg \in \Gamma_\sigma}} \pi_{gh}f'(g \cdot \sigma) = \sum_{h \in \Gamma_\sigma g^{-1}} \pi_{gh}f'(g \cdot \sigma) \nonumber \\ &= \sum_{h \in \Gamma_\sigma} \pi_{ghg^{-1}}f'(g \cdot \sigma) = \widetilde{P}f(g \cdot \sigma). \nonumber \end{aligned}$$ On the other hand, if $g \cdot \sigma \notin \Sigma(k, \Gamma)$, write $\widetilde{P}f(g \cdot \sigma) = \sum_{h \in A} \pi_h f'(\tau)$ where $A = \lbrace h \in \Gamma \colon h \cdot \tau = g \cdot \sigma \rbrace = gB$ for $B = \lbrace s \in \Gamma \colon s \cdot \tau = \sigma \rbrace$ and $\tau \in \Sigma(k, \Gamma)$. Hence, $$\begin{aligned}
\pi_g \widetilde{P}f(\sigma) &= \sum_{\substack{s \in \Gamma \\ s \cdot \tau = \sigma}} \pi_{gs}f'(\tau) = \sum_{h \in g \lbrace s \in \Gamma \colon s \cdot \tau = \sigma \rbrace} \pi_h f'(\tau)= \sum_{h \in gB} \pi_h f'(\tau) \nonumber \\ &= \sum_{h \in A} \pi_h f'(\tau) = \widetilde{P}f(g \cdot \sigma), \nonumber\end{aligned}$$ so the claim holds for $\sigma \notin \Sigma(k, \Gamma)$ as well.
Recalling that $\Gamma$ is discrete, normalizing $\widetilde{P}$ as below gives a projection onto the twisted cochains.
\[projection2\] Define ${P}: \mathcal{E}^{(k,p)}(X,E) \rightarrow \mathcal{E}^{(k,p)}(X,E)$ by $$Pf(\sigma) = \dfrac{1}{\vert \Gamma_\sigma \vert}\widetilde{P}f(\sigma).$$
\[projection11\] ${P}$ is a projection onto the twisted cochains.
Clearly ${P}^2 = {P}$ and onto. Now, suppose $f$ is twisted and $\sigma \in \Sigma(k)$. If $\sigma \in \Sigma(k, \Gamma)$, then, recalling the discreteness assumption $$\begin{aligned}
{P}f(\sigma) &= \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \pi_sf'(\sigma)= \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} f'(s \cdot \sigma) \nonumber = \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} f'(\sigma) = f'(\sigma) =f(\sigma). \end{aligned}$$ Similarly, for $\sigma \notin \Sigma(k, \Gamma)$ $$\begin{aligned}
{P}f(\sigma) &= \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{\substack{h \in \Gamma \\ h \cdot \tau = \sigma}} \pi_{h} f'(\tau) = \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{\substack{h \in \Gamma \\ h \cdot \tau = \sigma}} f(h \cdot \tau) = \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{\substack{h \in \Gamma \\ h \cdot \tau = \sigma}} f(\sigma) = f(\sigma), \nonumber\end{aligned}$$ as $\vert \lbrace h \in \Gamma \colon h \cdot \tau = \sigma \rbrace \vert = \vert h \Gamma_\tau \vert = \vert \Gamma_\tau \vert$ and $\vert \Gamma_\sigma \vert = \vert h \Gamma_\tau h^{-1}\vert = \vert \Gamma_\tau \vert$.
$P$ is continuous with $\Vert {P}f \Vert_{(k,p)} \leq \Vert f \Vert_{(k,p)}$ for $f \in \mathcal{E}^{(k,p)}(X,E)$ with equality for $f \in L^{(k,p)}(X,E)$.
A straightforward consequence of Definition \[projection2\], and the observation that ${P}f=f$ for $f \in L^{(k,p)}(X,E)$.
Thus, we have constructed a projection $P$ onto the cochains twisted by $\pi$. However, cochains in the image are not necessarily alternating and hence not necessarily in $L^{(k,p)}(X,E)$. Antisymmetrizing $P$ fixes this. We begin with the following observation:
\[remainstwisted\] \[propAlt\] If $f$ is twisted, then $\mathrm{Alt} f$ is twisted.
Suppose $f: \Sigma(k) \rightarrow E$ is twisted. Then, $\mathrm{Alt}f(g \cdot \sigma) = \mathrm{Alt}(\pi_g f(\sigma)) = \pi_g (\mathrm{Alt}f(\sigma))$ for all $g \in \Gamma$ and $\sigma \in \Sigma(k)$, where we used the fact that $f$ is twisted in the first equality and linearity of $\pi_g$ in the last equality. Hence, $\mathrm{Alt}f$ is twisted as well.
\[althelp\] Suppose $f \in \mathcal{E}^{(k,p)}(X,E)$, then $$\Vert \mathrm{Alt}f \Vert_{(k,p)}^p \leq (k+1)! \Vert f \Vert_{(k,p)}^p$$
Since $$\begin{aligned}
\Vert \mathrm{Alt}f(\sigma) \Vert^p_E &= \dfrac{1}{(k+1)!^p} \Vert \sum_{\alpha \in S_{k+1}} \mathrm{sign}(\alpha)\alpha^*f(\sigma) \Vert^p_E \leq \dfrac{(k+1)!^p}{(k+1)!^p} \sum_{\alpha \in S_{k+1}} \Vert \alpha^*f(\sigma) \Vert^p_E \nonumber \\ &= \sum_{\alpha \in S_{k+1}}\Vert \alpha^*f(\sigma) \Vert^p_E, \nonumber \end{aligned}$$ it follows that $$\begin{aligned}
\Vert \mathrm{Alt}f \Vert^p_{(k,p)} &= \sum_{\sigma \in \Sigma(k,\Gamma)} \Vert \mathrm{Alt} f(\sigma) \Vert^p \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\leq \sum_{\alpha \in S_{k+1}} \sum_{\sigma \in \Sigma(k,\Gamma)} \Vert \alpha^* f(\sigma) \Vert^p \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &= (k+1)! \sum_{\sigma \in \Sigma(k,\Gamma)} \Vert f(\sigma) \Vert^p \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \end{aligned}$$ since we sum over all representatives in the last equality, and for $\sigma = (v_0, \dots, v_k)$, $\omega(v_{\alpha(0)}, \dots, v_{\alpha(k)}) = \omega(\sigma)$ and $\Gamma_{(v_{\alpha(0)}, \dots v_{\alpha(k)})} = \Gamma_\sigma$ for all $\alpha \in S_{k+1}$.
The map $P_L : \mathcal{E}^{(k,p)}(X,E) \rightarrow \mathcal{E}^{(k,p)}(X,E)$, given by $$P_L= \mathrm{Alt} \circ {P}$$ defines a projection onto $L^{(k,p)}(X,E)$. In other words, the diagram $$\xymatrix{
{\mathcal{E}^{(k,p)}(X,E)} \ar@{->>}[r]^P \ar@{->>}[dr]_{{P_L}} &
{\lbrace \mathrm{k-cochain}_{\pi}\rbrace} \ar@{->>}[d]^{{\mathrm{Alt}}} \\ {} & {{L^{(k,p)}(X,E)}}
}$$ commutes.
Clearly $P_L^2 = P_L$. By Proposition \[projection11\] $P$ is a projection onto the twisted cochains, and since $\mathrm{Alt}$ preserves twisting by Corollary \[remainstwisted\], $P_L$ is a projection onto $L^{(k,p)}(X,E)$.
\[P\_Lcontinuous\] $P_L$ is continuous with $\Vert P_Lf \Vert_{(k,p)}^p \leq (k+1)! \Vert f \Vert_{(k,p)}^p$.
$$\begin{aligned}
\Vert P_Lf \Vert_{(k,p)}^p &= \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert P_Lf(\sigma) \Vert^p_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\
&= \sum_{\sigma \in \Sigma(k, \Gamma)} \left\Vert \mathrm{Alt}\left( \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \pi_s f'(\sigma) \right)\right\Vert^p_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \left\Vert \dfrac{1}{\vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \pi_s \mathrm{Alt} f'(\sigma) \right\Vert^p_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ & \leq \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{1}{\vert \Gamma_\sigma \vert^p} \vert \Gamma_\sigma \vert^p \max_{s \in \Gamma_s} \left\lbrace \Vert \pi_s \mathrm{Alt}f'(\sigma)\Vert^p_E \right\rbrace \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert \mathrm{Alt}f'(\sigma)\Vert^p_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ \leq} \sum_{\sigma \in \Sigma(k, \Gamma)} (k+1)! \Vert f'(\sigma)\Vert^p_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &= (k+1)! \Vert f \Vert_{(k,p)}^p, \nonumber\end{aligned}$$
where we have used Corollary \[althelp\] in the last inequality.
\[Lclosed\] $\displaystyle L^{(k,p)}(X,E) \subseteq \mathcal{E}^{(k,p)}(X,E)$ is closed.
$L^{(k,p)}(X,E)^* \cong L^{(k,p^*)}(X,E^*)$ {#L}
===========================================
Having constructed a continuous projection from $\mathcal{E}^{(k,p)}(X,E)$ onto $L^{(k,p)}(X,E)$ we show that the dual of $L^{(k,p)}(X,E)$ can be identified up to isometric isomorphism with $\mathcal{E}^{(k,p^*)}(X,E^*) / \mathrm{Ann}(L^{(k,p)}(X,E))$, cf. Corollary \[fin2\], and finally that the latter is isometrically isomorphic to $L^{(k,p^*)}(X,E^*)$, cf. Proposition \[isometric1\] and \[fin4\] below. Towards this end, recall that by the annihilator of a subspace $M \subseteq E$ we mean the vector space $\mathrm{Ann}(M) = \lbrace x \in E^* \colon \langle y,x \rangle_E = 0 \,\, \forall y \in M \rbrace$ of all bounded linear functionals on $E$ that vanish on $M$. The following fact contains the idea of the proof:
\[dualcomposition1\] \[firstdual\] [@Douglas] Suppose $E$ is a Banach space such that $E= M \oplus N$ and denote by $P$ the corresponding projection onto $M$. Then,
1. $\ker P^* = \mathrm{Ann}(M)$ and $\mathop{\mathrm{im}}P^* = \mathrm{Ann}(N)$;
2. $E^* \simeq \mathrm{Ann}(N) \oplus \mathrm{Ann}(M)$;
3. if $M$ is closed $M^* \cong E^* / \mathrm{Ann}(M).$
Let $L^{(k,p)}_-(X,E)$ denote the closed complement of $L^{(k,p)}(X,E)$ in $\mathcal{E}^{(k,p)}(X,E)$. That is $ L^{(k,p)}_-(X,E) = \ker \, {P}_{L}$, or in other words:
\[symmetry\] \[fin1\] $L_-^{(k,p)}(X,E) = \lbrace f \in \mathcal{E}^{(k,p)}(X,E) \colon \mathrm{Alt}f(\sigma) = 0 \,\, \forall \sigma \in \Sigma(k, \Gamma) \rbrace$ is a closed subspace of $\mathcal{E}^{(k,p)}(X,E).$
Given $f \in L^{(k,p)}_-(X,E)$, $(I-P_L)f(\sigma) = f(\sigma)$ for all $\sigma \in \Sigma(k)$, and hence for all $\sigma \in \Sigma(k, \Gamma)$ $$\begin{aligned}
(I-P_L)f'(\sigma) &= f'(\sigma) - P_Lf'(\sigma) = f'(\sigma) - \dfrac{1}{\vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \pi_s \mathrm{Alt}f'(\sigma) = f'(\sigma), \nonumber \end{aligned}$$ implying by linearity that $\mathrm{Alt}f'(\sigma) = 0$. Hence, $f$ is symmetric on representatives.
\[fin2\] $\displaystyle L^{(k,p)}(X,E)^* \cong \mathcal{E}^{(k,p^*)}(X,E^*) / \mathrm{Ann}(L^{(k,p)}(X,E)).$
Since $L^{(k,p)}(X,E)$ is a closed subspace of $\mathcal{E}^{(k,p)}(X,E)$ by Corollary \[Lclosed\], the claim now follows from Proposition \[firstdual\](3) and the fact that ${\mathcal{E}^{(k,p)}(X,E)}^* \cong \mathcal{E}^{(k,p^*)}(X,E^*)$.
It now remains to identify the annihilators, cf. Proposition \[isometric1\], to prove isomorphism and finally isometry. As indicated by Proposition \[dualcomposition1\] this requires knowing $P_L^*$.
\[P\_Ladjoint\] Let $\overline{P}_L : \mathcal{E}^{(k,p^*)}(X,E^*) \rightarrow \mathcal{E}^{(k,p^*)}(X,E^*)$ be a projection as above. Then $\overline{P}_L = P_L^*$.
Assume first $k=1$, let $f \in \mathcal{E}^{(1,p)}(X,E)$ and $\phi \in \mathcal{E}^{(1,p^*)}(X,E^*)$. For $\sigma = (v_0, v_1) \in \Sigma(1, \Gamma)$ we denote by $-\sigma$ the simplex $(v_1,v_0)$. Now, $$\begin{aligned}
&\langle P_Lf, \phi \rangle_1 = \sum_{\sigma \in \Sigma(1, \Gamma)} \langle P_Lf(\sigma), \phi(\sigma) \rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &\substack{(*) \\ =} \sum_{\sigma \in \Sigma(1, \Gamma)} \left\langle \dfrac{1}{\vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_s} \pi_s \left( \dfrac{1}{2} f'(\sigma) - \dfrac{1}{2}f'(-\sigma)\right), \phi'(\sigma)\right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \left\langle \pi_s \left(f'(\sigma) - f'(-\sigma)\right), \phi'(\sigma)\right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \left\langle \pi_s f'(\sigma), \phi'(\sigma) \right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &\phantom{=} -\sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \left\langle \pi_s f'(-\sigma), \phi'(\sigma)\right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &\substack{(**) \\ =} \sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \left\langle \pi_s f'(\sigma), \phi'(\sigma) \right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &\phantom{=} -\sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert}\sum_{s \in \Gamma_\sigma} \left\langle \pi_s f'(\sigma), \phi'(-\sigma)\right\rangle_E
\dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \end{aligned}$$ $$\begin{aligned}
&= \sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \left\langle f'(\sigma), \overline{\pi}_s \phi'(\sigma) \right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &\phantom{=} -\sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \left\langle f'(\sigma), \overline{\pi}_s\phi'(-\sigma)\right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(1, \Gamma)} \dfrac{1}{2 \vert \Gamma_\sigma \vert} \sum_{s \in \Gamma_\sigma} \left\langle f'(\sigma), \overline{\pi}_s \left(\phi'(\sigma)-\phi'(-\sigma) \right) \right\rangle_E \dfrac{\omega(\sigma)}{2! \vert \Gamma_\sigma \vert} \nonumber \\ &= \langle f, \overline{P}_L \phi \rangle_1,\nonumber \end{aligned}$$ where $(*)$ and the last equality follow from the definition of $P_L$ and $\overline{P}_L$, respectively when $k=1$. $(**)$ follows as we sum over all $\sigma \in \Sigma(1, \Gamma)$, so the sums where we switch the summation variable $\sigma$ with $-\sigma$ agree as $\omega(\sigma) = \omega(-\sigma)$. For $k>1$ the calculation goes similarly, denoting $\sigma = (v_0, \dots, v_k) \in \Sigma(k, \Gamma)$ and arguing similarly, $$\begin{aligned}
&\langle P_Lf, \phi \rangle_k = \sum_{\sigma \in \Sigma(k, \Gamma)} \langle P_Lf(\sigma), \phi(\sigma) \rangle_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =}\sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{s \in \Gamma_s} \sum_{\alpha \in S_{k+1}} \dfrac{1}{\vert \Gamma_\sigma \vert} \dfrac{1}{(k+1)!} (-1)^{\mathrm{sgn}(\alpha)} \left\langle \pi_s f'((v_{\alpha(0)}, \dots, v_{\alpha(k)})), \phi'(\sigma)\right\rangle_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =}\sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{s \in \Gamma_s} \sum_{\alpha \in S_{k+1}} \dfrac{1}{\vert \Gamma_\sigma \vert} \dfrac{1}{(k+1)!} (-1)^{\mathrm{sgn}(\alpha)} \left\langle \pi_s f'(\sigma), \phi'((v_{\alpha(0)}, \dots, v_{\alpha(k)}))\right\rangle_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma \vert} \nonumber \\ &=\sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{s \in \Gamma_s} \sum_{\alpha \in S_{k+1}} \dfrac{1}{\vert \Gamma_\sigma \vert} \dfrac{1}{(k+1)!} (-1)^{\mathrm{sgn}(\alpha)} \left\langle f'(\sigma), \overline{\pi}_s \phi'((v_{\alpha(0)}, \dots, v_{\alpha(k)}))\right\rangle_E \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma \vert} \nonumber \\ &= \langle f, \overline{P}_L \phi \rangle_k, \nonumber\end{aligned}$$ where the first and last equality follows by the definition of $P_L$ and linearity of the dual pairing, and the third similarly as in the case $k=1$ above.
\[isometric1\] The following are equal as sets:
1. $\mathrm{Ann}(L_-^{(k,p)}(X,E)) = L^{(k,p^*)}(X,E^*);$
2. $\mathrm{Ann}(L^{(k,p)}(X,E)) = L_-^{(k,p^*)}(X,E^*).$
Suppose $f \in L_-^{(k,p)}(X,E)$ and $\phi \in \mathcal{E}^{(k,p^*)}(X,E^*)$. Then, $$\begin{aligned}
\langle f, \phi \rangle_k &= \sum_{\sigma \in \Sigma(k, \Gamma)} \langle f(\sigma), \phi(\sigma) \rangle_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \langle (I-P_L)f(\sigma), \phi(\sigma) \rangle_E \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &= \langle f,\phi \rangle_k - \langle P_L f,\phi \rangle_k = \langle f,\phi \rangle_k - \langle f,\overline{P}_L \phi \rangle_k \nonumber.\end{aligned}$$ Hence, $\langle f, \overline{P}_L \phi \rangle_E = 0$ for all $f \in L_-^{(k,p)}(X,E)$ where $\overline{P}_L \phi \in L^{(k,p^*)}(X,E^*)$. Thus, $L^{(k,p^*)}(X,E^*) \subseteq \mathrm{Ann}(L_-^{(k,p)}(X,E))$. On the other hand, suppose $\phi \in \mathrm{Ann}(L_-^{(k,p)}(X,E))$, then $\langle f, \phi \rangle_k = 0$ for all $f \in L_-^{(k,p)}(X,E)$. Hence, for all $f \in \mathcal{E}^{(k,p^*)}(X,E^*)$, so $0= \langle (I-P_L)f, \phi \rangle_k$ if and only if $\langle f, \phi \rangle_k = \langle P_Lf, \phi \rangle_k = \langle f, \overline{P}_L \phi \rangle_k$ for all $f \in \mathcal{E}^{(k,p^*)}(X,E^*)$. Thus, $\phi = \overline{P}_L \phi$ implies that $\phi \in L^{(k,p^*)}(X,E^*)$, so $\mathrm{Ann}(L_-^{(k,p)}(X,E)) \subseteq L^{(k,p^*)}(X,E^*)$ proving the first claim. The proof of the second claim goes similarly.
\[isometry1\] The following are isomorphic:
1. $ L^{(k,p)}(X,E)^* \cong \mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}_-(X,E^*) \simeq L^{(k,p^*)}(X,E^*);$
2. $ L^{(k,p)}_-(X,E)^* \cong \mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}(X,E^*) \simeq L^{(k,p^*)}_-(X,E^*).$
The first isomorphic isomorphisms follow immediately combining Propositions \[firstdual\] and \[isometric1\], and the latter isomorphisms by Proposition \[dualcomposition1\].
\[fin4\] The following are isometrically isomorphic:
1. $L^{(k,p)}(X,E)^* \cong L^{(k,p^*)}(X,E^*);$
2. $L^{(k,p)}_-(X,E)^* \cong L^{(k,p^*)}_-(X,E^*).$
Proof. Consider the second claim. Consider $\mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}(X,E^*)$ consisting of the cosets $[\phi] = \phi + L^{(k,p^*)}(X,E^*)$ for $\phi \in \mathcal{E}^{(k,p^*)}(X,E^*)$. We claim that if $\phi \in L^{(k,p^*)}_-(X,E^*)$, then $\Vert [\phi]\Vert = \Vert \phi \Vert_{(k,p^*)}$ where $\Vert \cdot \Vert = \inf \lbrace \Vert \xi \Vert_ {(k,p^*)} \colon \xi \in [\phi] \rbrace$ is the quotient norm. On the other hand, $\mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}(X,E^*) \simeq L^{(k,p^*)}_-(X,E^*)$ by Corollary \[isometry1\], so $\mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}(X,E^*) \cong L^{(k,p^*)}_-(X,E^*)$ and consequently $L^{(k,p)}_-(X,E)^* \cong L^{(k,p^*)}_-(X,E^*)$. Towards this end, fix $\phi \in L^{(k,p^*)}_-(X,E^*)$. Thus, for $\psi \in L^{(k,p^*)}(X,E^*)$ we have $$\begin{aligned}
\Vert \phi + \psi \Vert_{(k,p^*)}^{p^*} &= \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert \phi(\sigma) + \psi(\sigma) \Vert^{p^*}_{E^*} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert \phi(\sigma) - \psi(-\sigma) \Vert^{p^*}_{E^*} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert \phi(-\sigma) - \psi(-\sigma) \Vert^{p^*}_{E^*} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \Vert \phi(-\sigma) - \psi(-\sigma) \Vert^{p^*}_{E^*} \dfrac{\omega(-\sigma)}{(k+1)!\vert \Gamma_{-\sigma} \vert} \nonumber \\ &\substack{{} \\ =} \Vert \phi - \psi \Vert^{p^*}_{(k,p^*)}, \nonumber\end{aligned}$$ where $-\sigma = (v_1,v_0, v_2, \dots, v_k)$ for $\sigma = (v_0, \dots, v_k)$. The second equality follows since $\psi$ is alternating, the third equality since $\phi$ is symmetric on representatives, the fourth since $\omega$ is symmetric and $\Gamma_\sigma = \Gamma_{- \sigma}$, and the last equality holds since we sum over all $\sigma \in \Sigma(k, \Gamma)$, which contains all the oriented simplexes with the vertices of $\sigma$. Thus, $\Vert \phi + \psi \Vert_{(k,p^*)}^{p^*} = \Vert \phi - \psi \Vert_{(k,p^*)}^{p^*}$ and consequently by the triangle inequality $$2\Vert \phi \Vert_{(k,p^*)} = \Vert 2 \phi + \psi - \psi \Vert_{(k,p^*)} \leq \Vert \phi + \psi \Vert_{(k,p^*)} + \Vert \phi - \psi \Vert_{(k,p^*)},$$ and so $$\Vert \phi \Vert_{(k,p^*)} \leq \dfrac{1}{2}(\Vert \phi + \psi \Vert_{(k,p^*)} + \Vert \phi - \psi \Vert_{(k,p^*)}) = \Vert \phi + \psi \Vert_{(k,p)}.$$ Now, taking the infimum over $\psi \in L^{(k,p^*)}(X,E^*)$ thus shows that the quotient norm of $[\phi]$ is $\Vert \phi\Vert_{(k,p^*)}$, proving the second claim. The first claim is proven similarly by considering the cosets in $\mathcal{E}^{(k,p^*)}(X,E^*) \, / L^{(k,p^*)}_-(X,E^*)$.
$\Box$
Differentials and codifferentials {#DC}
=================================
Having identified the dual of $L^{(k,p)}(X,E)$ up to isometric isomorphism we extend the notion of differentials and codifferentials as presented in [@BS] to reflexive Banach spaces.
Codifferentials and differentials. Define the codifferential $$d_k: \mathcal{E}^{(k,p)}(X,E) \rightarrow \mathcal{E}^{(k+1,p)}(X,E)$$ point-wise by $$d\phi(\sigma) = \sum_{i=0}^{k+1}(-1)^i\phi(\sigma_i),$$ and the differential $$\delta_{k+1}: L^{(k+1,p^*)}(X,E^*) \rightarrow L^{(k,p^*)}(X,E^*),$$ as the adjoint of $d$ given by $\langle \phi , d \psi \rangle_{k+1} = \langle \delta \phi, \psi \rangle_k$ for all $\psi \in L^{(k,p)}(X,E)$ and $\phi \in L^{(k+1,p^*)}(X,E^*)$.
Similarly, we denote by $\bar{d}_k: \mathcal{E}^{(k,p^*)}(X,E^*) \rightarrow \mathcal{E}^{(k+1,p^*)}(X,E^*)$ the map given by $\bar{d}\psi(\sigma) = \sum_{i=0}^{k}(-1)^i\psi(\sigma_i)$ for $\psi \in \mathcal{E}^{(k,p^*)}(X,E^*)$, and likewise for the differential.
$$\xymatrix@1{{\cdots} \ar[r]^-{\delta_{k+2}} & L^{(k+1,p^*)}(X,E^*) \ar[r]^-{\delta_{k+1}} & L^{(k,p^*)}(X,E^*) \ar[r]^-{\delta_k} & {\cdots}
}$$ is a chain complex over $\mathbb{R}$ dual to the cochain complex $$\xymatrix@1{{\cdots\,} & L^{(k+1,p)}(X,E) \ar[l]_-{d_{k+1}} & L^{(k,p)}(X,E) \ar[l]_-{d_k} & {\cdots} \ar[l]_-{d_{k-1}}
}$$
As the following shows, both $d$ and $\delta$ are bounded operators.
Let $\phi \in L^{(k,p)}(X,E)$. Then $d: L^{(k,p)}(X,E) \rightarrow L^{(k+1,p)}(X,E)$ is a bounded operator and $$\Vert d\phi \Vert^p_{(k+1,p)} \leq (n-k)(k+2)^p\Vert \phi \Vert^p_{(k,p)}.$$
Using the estimate $$\begin{aligned}
\left\Vert \sum_{i=0}^{k+1} (-1)^i \phi(s_i) \right\Vert_E^p &\leq \left( \Vert \phi(s_0) \Vert_E + \left\Vert \sum_{i=1}^{k+1} (-1)^i \phi(s_i) \right\Vert_E \right)^p \leq \left( \sum_{i=0}^{k+1} \Vert \phi(s_i)\Vert_E \right)^p \nonumber \\ &\leq ((k+2) \max \lbrace \Vert \phi(s_0) \Vert, \dots, \Vert \phi(s_{k+1})\Vert \rbrace )^p \nonumber \\ &\leq (k+2)^p \sum_{i=0}^{k+1} \Vert \phi(s_i) \Vert_E^p , \nonumber \end{aligned}$$ it follows that $$\begin{aligned}
\Vert d \phi \Vert^p_{(k+1,p)} &= \sum_{s \in \Sigma(k+1, \Gamma)} \Vert d \phi (s) \Vert_E^p \dfrac{\omega(s)}{(k+2)! \vert \Gamma_s \vert} \nonumber \\ &= \sum_{s \in \Sigma(k+1, \Gamma)} \left\Vert \sum_{i=0}^{k+1}(-1)^i \phi(s_i) \right\Vert_E^p \dfrac{\omega(s)}{(k+2)! \vert \Gamma_s \vert} \nonumber \\ &\substack{{} \\ \leq} \sum_{s \in \Sigma(k+1, \Gamma)} \left( \dfrac{(k+2)^p\omega(s)}{(k+2)! \vert \Gamma_s \vert} \sum_{i=0}^{k+1}\Vert \phi(s_i)\Vert_E^p \right) \nonumber \end{aligned}$$ $$\begin{aligned}
&\substack{{} \\ =} \sum_{s \in \Sigma(k+1, \Gamma)} \left( \dfrac{(k+2)^p\omega(s)}{(k+2)!(k+1)! \vert \Gamma_s \vert} \sum_{\substack{t \in \Sigma(k) \\ t \subset s}} \Vert [s:t] \phi(t)\Vert_E^p \right) \nonumber \\ &\substack{{} \\ =} \sum_{s \in \Sigma(k+1, \Gamma)} \left( \dfrac{(k+2)^p\omega(s)}{(k+2)!(k+1)! \vert \Gamma_s \vert} \sum_{\substack{t \in \Sigma(k) \\ t \subset s}} \Vert \phi(t)\Vert_E^p \right) \nonumber \\ %&\substack{(3) \\ =} \sum_{\substack{s \in \Sigma(k+1, \Gamma) \\ t \in \Sigma(k, \Gamma)}} \sum_{\substack{g \in \Gamma \\ g \cdot t \subset s}} \dfrac{(k+2)^p\omega(s)}{(k+2)!(k+1)! \vert \Gamma_s \vert \vert \Gamma_t \vert} \Vert \phi(g \cdot t)\Vert_E^p \nonumber \\
%Here, I figure I could look for the possibility of $\Gamma$-invariance of $\Vert \phi(t)\Vert^p$, or some inequality making it independent of $g$; if I suppose that twisting holds for those $g \in \Gamma$ such that $g \cdot t \subset s$, then I get $\Vert \phi(g \cdot t) \Vert_E^p = \Vert \pi_g \phi(t) \Vert^p \leq \Vert \pi_g\Vert_E^p \Vert \phi(t)\Vert_E^p$, and what I would need is that $\Vert \pi_g\Vert_E^p \leq C$ for all such $g$. Another alternative would be to choose the $D^{(k,p)}(X,E)$ somehow as in your paper? If I assume I can get past this, say by finding such a $C$, I can continue as follows:
%&\substack{(4) \\ =} \sum_{\substack{s \in \Sigma(k+1, \Gamma) \\ t \in \Sigma(k, \Gamma)}} \dfrac{(k+2)^p \omega(s)}{(k+2)!(k+1)! \vert \Gamma_s \vert \vert \Gamma_t \vert} \sum_{\substack{g \in \Gamma \\ g \cdot t \subset s}} \Vert \phi(t) \Vert^p_E \nonumber \\ &\substack{(5) \\ =} \sum_{t \in \Sigma(k, \Gamma)} \dfrac{(k+2)^p \Vert \phi(t)\Vert^p_E}{(k+2)!(k+1)!\vert \Gamma_t \vert} \sum_{s \in \Sigma(k+1, \Gamma)} \sum_{\substack{g \in \Gamma \\ t \subset g^{-1} \cdot s}} \dfrac{ \omega(s)}{\vert \Gamma_s \vert} \nonumber \\ &\substack{(6) \\ =} \sum_{t \in \Sigma(k, \Gamma)} \dfrac{(k+2)^p \Vert \phi(t)\Vert^p_E}{(k+2)!(k+1)!\vert \Gamma_t \vert} \sum_{s \in \Sigma(k+1, \Gamma)} \sum_{\substack{g \in \Gamma \\ t \subset g \cdot s}} \dfrac{ \omega(s)}{\vert \Gamma_s \vert} \nonumber \\
&\substack{(*) \\ =}\sum_{t \in \Sigma(k, \Gamma)} \left( \dfrac{(k+2)^p \Vert \phi(t)\Vert^p_E}{(k+2)!(k+1)!\vert \Gamma_t \vert} \sum_{\substack{s \in \Sigma(k+1) \\ t \subset s}} \omega(s) \right) \nonumber \\
%&= \sum_{t \in \Sigma(k, \Gamma)} \dfrac{C(k+2)^p}{(k+2)!(k+1)!\vert \Gamma_t \vert} \sum_{s \in \Sigma(k+1, \Gamma)} \sum_{\substack{g \in \Gamma \\ t \subset g^ \cdot s}} \dfrac{\Vert \phi(t)\Vert^p_E \omega(s)}{\vert \Gamma_s \vert} \nonumber \\
%&= \sum_{\substack{s \in \Sigma(k+1, \Gamma) \\ t \in \Sigma(k, \Gamma)}} \sum_{\substack{g \in \Gamma \\ t \subset g \cdot s}} \dfrac{(k+2)^p\omega(s)}{(k+2)!(k+1)! \vert \Gamma_s \vert \vert \Gamma_t \vert} \Vert \phi(g \cdot t)\Vert_E^p \nonumber \\ &\leq \sum_{t \in \Sigma(k, \Gamma)} \sum_{\substack{s \in \Sigma(k+1) \\ t \subset s}} \dfrac{C(k+2)^p\omega(s)}{(k+2)!(k+1)! \vert \Gamma_t \vert} \Vert \phi(t)\Vert_E^p \nonumber \\
&\substack{(**) \\ =} \sum_{t \in \Sigma(k, \Gamma)} \dfrac{(n-k)(k+2)!(k+2)^p\omega(t)}{(k+2)!(k+1)! \vert \Gamma_t \vert} \Vert \phi(t)\Vert_E^p \nonumber \\ &= (n-k)(k+2)^p\Vert \phi \Vert^p_{(k,p)}, \nonumber\end{aligned}$$ where we used Proposition \[switchingsums\] in $(*)$ followed by Proposition \[combinatorial\] in $(**)$.
Similarly to [@BS], we also have the following useful point-wise expression for the differential.
\[codifferential\] For $\phi \in L^{(k,p^*)}(X, E^*)$ and $\tau \in \Sigma(k-1)$ $$\delta \phi(\tau) = \sum_{\substack{v \in \Sigma(0) \\ v * \tau \in \Sigma(k)}} \dfrac{\omega(v * \tau)}{\omega(\tau)} \phi(v * \tau).$$
Let $\psi \in L^{(k-1,p)}(X,E)$. The claim follows by a straightforward computation, $$\begin{aligned}
\langle \phi, d \psi\rangle_{k} &= \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma\vert}\langle \phi(\sigma), d \psi(\sigma)\rangle_E \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma\vert} \langle \phi(\sigma), \sum_{i=0}^{k-1}(-1)^i \psi(\sigma_i)\rangle_E \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{\omega(\sigma)}{(k+1)! \vert \Gamma_\sigma\vert} \langle \phi(\sigma), \dfrac{1}{(k-1+1)!} \sum_{\substack{\tau \in \Sigma(k-1) \\ \tau \subset \sigma}} [\sigma: \tau] \psi(\tau)\rangle_E \nonumber \\ &= \sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{\substack{\tau \in \Sigma(k-1) \\ \tau \subset \sigma}} [\sigma :\tau]\dfrac{\omega(\sigma)\langle \phi(\sigma), \psi(\tau)\rangle_E}{(k+1)! \vert \Gamma_\sigma\vert k!} \nonumber
% \\ &= \sum_{\substack{\sigma \in \Sigma(k,\Gamma) \\ \tau \in \Sigma(k-1, \Gamma)}} \sum_{\substack{g \in \Gamma \\ g \cdot \tau \subset \sigma}} [\sigma :g \cdot \tau]\dfrac{\omega(\sigma)\langle \phi(\sigma), \psi(g \cdot \tau)\rangle_E}{(k+1)! k! \vert \Gamma_\sigma \vert \vert \Gamma_\tau \vert}
\nonumber \end{aligned}$$ $$\begin{aligned}
%&=\sum_{\substack{\sigma \in \Sigma(k,\Gamma) \\ \tau \in \Sigma(k-1, \Gamma)}} \sum_{\substack{g \in \Gamma \\ g \cdot \tau \subset \sigma}} [\sigma : g \cdot \tau] \dfrac{\omega(\sigma)\langle \phi(g^{-1} \cdot \sigma), \psi(\tau)\rangle_E}{(k+1)! k! \vert \Gamma_\sigma \vert \vert \Gamma_\tau \vert} \nonumber \\ &= \sum_{\substack{\sigma \in \Sigma(k,\Gamma) \\ \tau \in \Sigma(k-1, \Gamma)}} \sum_{\substack{g \in \Gamma \\ \tau \subset g^{-1} \cdot \sigma}} [g^{-1} \cdot \sigma : \tau] \dfrac{\omega(\sigma)\langle \phi(g^{-1} \cdot \sigma), \psi(\tau)\rangle_E}{(k+1)! k! \vert \Gamma_\sigma \vert \vert \Gamma_\tau \vert} \nonumber \\ &= \sum_{\substack{\sigma \in \Sigma(k,\Gamma) \\ \tau \in \Sigma(k-1, \Gamma)}} \sum_{\substack{g \in \Gamma \\ \tau \subset g \cdot \sigma}} [g \cdot \sigma : \tau] \dfrac{\omega(\sigma)\langle \phi(g \cdot \sigma), \psi(\tau)\rangle_E}{(k+1)! k! \vert \Gamma_\sigma \vert \vert \Gamma_\tau \vert} \nonumber \\
\phantom{\langle \phi, d \psi\rangle_{k} = \,} &\substack{(*) \\ =} \sum_{\tau \in \Sigma(k-1, \Gamma)} \sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}} [\sigma :\tau] \dfrac{\omega(\sigma)\langle \phi(\sigma), \psi(\tau)\rangle_E}{(k+1)! \vert \Gamma_\tau\vert k!} \nonumber \\ &= \sum_{\tau \in \Sigma(k-1, \Gamma)} \dfrac{\omega(\tau)}{\omega(\tau)} \sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}} [\sigma :\tau] \dfrac{\omega(\sigma)\langle \phi(\sigma), \psi(\tau)\rangle_E}{(k+1)! \vert \Gamma_\tau\vert k!} \nonumber \\ &= \sum_{\tau \in \Sigma(k-1, \Gamma)} \dfrac{\omega(\tau)}{k! \vert \Gamma_\tau \vert} \sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}} [\sigma :\tau] \dfrac{\omega(\sigma)\langle \phi(\sigma), \psi(\tau)\rangle_E}{(k+1)! \omega(\tau)} \nonumber \\ &= \sum_{\tau \in \Sigma(k-1, \Gamma)} \dfrac{\omega(\tau)}{k! \vert \Gamma_\tau \vert} \sum_{\substack{v \in \Sigma(0) \\ v * \tau \in \Sigma(k)}} [v\tau : \tau] \dfrac{\omega(v * \tau)}{\omega(\tau)}\langle \phi(v * \tau), \psi(\tau)\rangle_E \nonumber \\ &= \sum_{\tau \in \Sigma(k-1, \Gamma)} \dfrac{\omega(\tau)}{k! \vert \Gamma_\tau \vert} \left\langle \sum_{\substack{v \in \Sigma(0) \\ v * \tau \in \Sigma(k)}} \dfrac{\omega(v * \tau)}{\omega(\tau)} \phi(v * \tau), \psi(\tau) \right\rangle_E = \langle \delta \phi, \psi\rangle_{k-1}, \nonumber\end{aligned}$$ where we used Proposition \[switchingsums\] in $(*)$ above.
Localization and restriction {#secloc}
============================
In this section we recall the concept of localization following [@BS] and develop the notion in the setting of reflexive Banach spaces. We also consider the concept of restriction, recently considered by I. Oppenheim in the context of $L^2$-cohomology [@IO]. Proposition \[averagenorms\] and \[propQ\] are the key results. The former relates the norm of the average to the norm of the differential, whereas the latter gives a global vanishing condition in the kernel of the full codifferential.
For a weight $\omega$, define the localized weight as $$\omega_\tau(\sigma) = \omega(\tau * \sigma)$$ for $\sigma \in \Sigma_\tau(j)$ and $\tau \in \Sigma(l)$ such that $\tau * \sigma \in \Sigma(j+l+1)$.
In other words, for $\tau \in \Sigma(l)$, $\omega_\tau(\sigma)$ is the number of $(n-l-1)$-simplexes in $X_\tau$ containing $\sigma \in \Sigma_\tau(j)$.
$\Gamma_\tau$ acts by simplicial automorphisms on $X_\tau$.
Let $\sigma \in X_\tau$ and suppose $g \in \Gamma_\tau$. Since, $\sigma \subset \tau * \sigma$, the join of $\sigma$ and $\tau$, it follows that $g \cdot \sigma \subset g \cdot \tau * g \cdot \sigma = \tau * g \cdot \sigma$ as $\Gamma$ acts by simplicial automorphisms on $X$ and $g \in \Gamma_\tau$. Thus, $g \cdot \sigma$ is a simplex in $\tau * g \cdot \sigma$, and so $g \cdot \sigma \in X_\tau$ since it is disjoint from $\tau$. Hence, $\Gamma_\tau$ act by simplicial automorphisms on $X_\tau$.
For $\eta \in X_\tau$, $\Gamma_{\tau \eta} = \Gamma_\tau \cap \Gamma_\eta$.
$\Gamma_{\tau \eta} = \lbrace g \in \Gamma_\tau \colon g \cdot \eta = \eta \rbrace = \Gamma_\tau \cap \Gamma_\eta.$
We denote by
1. $\pi_\tau$ the restriction of $\pi$ to $\Gamma_\tau$, that is $\pi_\tau = \pi \vert_{\Gamma_\tau}$;
2. $d_\tau$ the restriction of $d$ to $\mathcal{E}^{(k,p)}(X_\tau, E)$, that is $d_\tau = d \vert_{\mathcal{E}^{(k,p)}(X_\tau, E)}$;
3. $\delta_\tau$ the restriction of $\delta$ to $L^{(k+1,p^*)}(X_\tau, E^*)$, that is $\delta_\tau = \delta \vert_{L^{(k+1,p^*)}(X_\tau, E^*)}$.
Let $$\mathcal{E}^{(k,p)}(X_\tau,E) = \left\lbrace f \colon \Sigma_\tau(k) \rightarrow E \colon \Vert f \Vert^p_{(k,p)} < \infty \right\rbrace$$ for $X_\tau \subset X$ denote the vector space of $p$-summable functions with semi-norm $$\Vert f \Vert_{(k,p)} = \left( \sum_{\sigma \in \Sigma_\tau(k, \Gamma_\tau)} \Vert f(\sigma) \Vert^p_E \dfrac{\omega_\tau(\sigma)}{(k+1)! \vert \Gamma_{\tau \sigma}\vert}\right)^{\dfrac{1}{p}}.$$
Let $L^{(k,p)}(X_\tau, E)$ denote the subspace $$\left\lbrace f \in \mathcal{E}^{(k,p)}(X_\tau,E)\, \colon \, f \, \mathrm{alternating} \, \mathrm{and} \, \forall g \in \Gamma_\tau, \forall \sigma \in \Sigma_\tau(k), f(g \cdot \sigma) = \pi_{\tau g} \cdot f(\sigma) \right\rbrace$$ of simplicial $k$-cochains of $X_\tau$ twisted by $\pi_\tau$.
For $f \in \mathcal{E}^{(k,p)}(X,E)$ and $\tau \in \Sigma(j)$ such that $k-j-1 \geq 0$, the localization of $f$ to $X_\tau$ is the function $f_\tau \in \mathcal{E}^{(k,p)}(X_\tau,E) \in \mathcal{E}^{(k-j-1, p)}(X_\tau, E)$ defined by the localization map $${}_\tau \colon \mathcal{E}^{(k,p)}(X,E) \longrightarrow \mathcal{E}^{(k-j-1,p)}(X_\tau,E)$$ where for all $\sigma \in \Sigma_\tau(k-j-1)$, $f_\tau(\sigma) = f (\tau * \sigma)$. Similarly we define its dual $\mathcal{E}^{(k,p^*)}(X,E^*) \longrightarrow \mathcal{E}^{(k,p^*)}(X_\tau,E^*)$, also denoted by ${}_\tau$.
For $f \in \mathcal{E}^{(k,p)}(X,E)$ and $\tau \in \Sigma(j)$ such that $k+j+1 \leq n$, the restriction of $f$ to $X_\tau$ is the function $f^\tau \in \mathcal{E}^{(k,p)}(X_\tau,E)$ defined by the restriction map $${}^\tau \colon \mathcal{E}^{(k,p)}(X,E) \rightarrow \mathcal{E}^{(k,p)}(X_\tau,E)$$ where for all $\sigma \in \Sigma_\tau(k)$, $f^\tau (\sigma) = f(\sigma)$. Similarly we define its dual $\mathcal{E}^{(k,p^*)}(X,E^*) \longrightarrow \mathcal{E}^{(k-j-1,p^*)}(X_\tau,E^*)$, also denoted by ${}^\tau$.
Next, we consider a number of local to global equalities that will be of use. We begin by the following useful local relation:
\[BS1.10\] For $f_\tau \in L^{(k,p)}(X_\tau, E)$ $$\Vert f_\tau \Vert^p_{(k,p)} = \dfrac{1}{(k+1)!\vert \Gamma_\tau \vert} \sum_{\sigma \in \Sigma_\tau(k)} \Vert f_\tau(\sigma) \Vert^p_E \omega(\tau * \sigma).$$
Proof. $$\begin{aligned}
\Vert f_\tau \Vert^p_{(k,p)} &= \sum_{\sigma \in \Sigma_{\tau}(k, \Gamma_\tau)} \Vert f_\tau(\sigma) \Vert^p_E \dfrac{\omega_\tau(\sigma)}{\vert \Gamma_{\tau \sigma}\vert(k+1)!} \substack{{} \\ =} \sum_{\sigma \in \Sigma_{\tau}(k)} \dfrac{\Vert f_\tau(\sigma) \Vert^p_E}{\vert \Gamma_\tau \sigma \vert} \dfrac{\omega_\tau(\sigma)}{\vert \Gamma_{\tau \sigma}\vert(k+1)!} \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma_{\tau}(k)} \Vert f_\tau(\sigma) \Vert^p_E \dfrac{\vert \Gamma_{\tau\sigma}\vert}{\vert\Gamma_\tau\vert} \dfrac{\omega_\tau(\sigma)}{\vert \Gamma_{\tau \sigma}\vert(k+1)!} = \sum_{\sigma \in \Sigma_{\tau}(k)} \Vert f_\tau(\sigma) \Vert^p_E \dfrac{\omega_\tau(\sigma)}{\vert \Gamma_{\tau}\vert(k+1)!}, \nonumber \\ &= \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert}\sum_{\sigma \in \Sigma_{\tau}(k)} \Vert f_\tau(\sigma) \Vert^p_E {\omega}(\tau *\sigma) \nonumber\end{aligned}$$ where the second equality in terms of the $k$-simplexes of $X_\tau$ follows from the $\Gamma_\tau$ invariance of the norm, and the third by the fact that $\vert \Gamma_\tau \sigma \vert$, the size of the $\Gamma_\tau$ orbit of $\sigma$, is $\vert \Gamma_\tau \vert / \vert \Gamma_{\tau \sigma} \vert$.
$\Box$
\[oppenheim\] Let $f \in L^{(k,p)}(X,E)$. If $k + 1 \leq n$, then $$(n-k) \Vert f \Vert_{(k,p)}^p = \sum_{\tau \in \Sigma(0, \Gamma)} \Vert f^\tau \Vert_{(k,p)}^p.$$
For $\xi \in \Sigma(k+1)$ such that $\tau \subset \xi$, denote by $\xi-\tau$ the $k$-simplex in $X_\tau$ obtained by removing the vertex $\tau$ from $\xi$. Now, $$\begin{aligned}
\sum_{\tau \in \Sigma(0, \Gamma)} \Vert f^\tau \Vert_{(k,p)}^p &\substack{{} \\ =} \sum_{\tau \in \Sigma(0, \Gamma)} \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert} \sum_{\sigma \in \Sigma_\tau(k)} \Vert f (\sigma) \Vert_E^p \omega (\tau * \sigma) \nonumber \\ &\substack{{} \\ =} \sum_{\tau \in \Sigma(0, \Gamma)} \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert} \sum_{\substack{\xi \in \Sigma(k+1) \\ \tau \subset \xi}} \dfrac{1}{(k+2)} \Vert f (\xi - \tau) \Vert_E^p \omega (\xi) \nonumber \\ &= \sum_{\tau \in \Sigma(0, \Gamma)} \sum_{\substack{\xi \in \Sigma(k+1) \\ \tau \subset \xi}} \dfrac{1}{(k+2)! \vert \Gamma_\tau \vert} \Vert f (\xi - \tau) \Vert_E^p \omega (\xi) \nonumber \\ &\substack{(*) \\ =} \sum_{\xi \in \Sigma(k+1, \Gamma)} \sum_{\substack{\tau \in \Sigma(0) \\ \tau \subset \xi}} \dfrac{1}{(k+2)! \vert \Gamma_\xi \vert} \Vert f (\xi - \tau) \Vert_E^p \omega (\xi) \nonumber \\ &\substack{{} \\ =} \sum_{\xi \in \Sigma(k+1, \Gamma)} \sum_{\substack{\sigma \in \Sigma(k) \\ \sigma \subset \xi}} \dfrac{1}{(k+2)! \vert \Gamma_\xi \vert} \dfrac{1}{(k+1)!} \Vert f (\sigma) \Vert_E^p \omega (\xi) \nonumber \\ &\substack{(**) \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{\substack{\xi \in \Sigma(k+1) \\ \sigma \subset \xi}} \dfrac{1}{(k+2)! \vert \Gamma_\sigma \vert} \dfrac{1}{(k+1)!} \Vert f (\sigma) \Vert_E^p \omega (\xi) \nonumber \\ &\substack{{} \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{1}{(k+2)! \vert \Gamma_\sigma \vert} \dfrac{1}{(k+1)!} \Vert f (\sigma) \Vert_E^p (n-k)(k+2)!\omega (\sigma) \nonumber \\ &= (n-k) \Vert f \Vert_{(k,p)}^p, \nonumber\end{aligned}$$ where the first and second equality follow by Proposition \[BS1.10\]. noting that $f^\tau (\sigma) = f(\sigma)$, writing $\xi - \tau$ as $\sigma$ and accounting for ordering. $(*)$ and $(**)$ follows by switching sums by Proposition \[switchingsums\] and the second last equality follows by Proposition \[combinatorial\].
\[summing\] Let $f \in L^{(k,p)}(X, E)$ and $0 \leq j < k$. Then, $$(k+1)! \Vert f \Vert^p_{(k,p)} = (k-j)! \sum_{\tau \in \Sigma(j, \Gamma)} \Vert f_\tau \Vert^p_{(k-j-1,p)}.$$
$$\begin{aligned}
\sum_{\tau \in \Sigma(j, \Gamma)} \Vert f_\tau \Vert^p_{(k-j-1,p)} &\substack{{} \\ =} \sum_{\tau \in \Sigma(j, \Gamma)} \sum_{\eta \in \Sigma_\tau(k-j-1)} \dfrac{\omega(\tau *\eta)}{(k-j)! \vert \Gamma_\tau\vert} \Vert f_\tau(\eta)\Vert^p_E \nonumber \\ &\substack{{} \\ =} \sum_{\tau \in \Sigma(j, \Gamma)} \sum_{\substack{\sigma \in \Sigma(k) \\ \sigma = \tau * \eta}} \dfrac{\omega(\sigma)}{(k-j)!\vert \Gamma_\tau \vert} \Vert f(\sigma) \Vert^p_E \nonumber \\ &\substack{(*) \\ =} \sum_{\tau \in \Sigma(j, \Gamma)} \sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}} \dfrac{(k-j)!}{(k+1)!} \dfrac{\omega(\sigma)}{(k-j)!\vert \Gamma_\tau \vert} \Vert f(\sigma) \Vert^p_E \nonumber \\ &= \sum_{\tau \in \Sigma(j, \Gamma)} \sum_{\substack{\sigma \in \Sigma(k) \\ \tau \subset \sigma}} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\tau \vert} \Vert f(\sigma) \Vert^p_E \nonumber \\ &\substack{(**) \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \sum_{\substack{\tau \in \Sigma(j) \\ \tau \subset \sigma}} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \Vert f(\sigma) \Vert^p_E \nonumber \\ &\substack{(***) \\ =} \sum_{\sigma \in \Sigma(k, \Gamma)} \dfrac{(k+1)!}{(k-j)!} \dfrac{\omega(\sigma)}{(k+1)!\vert \Gamma_\sigma \vert} \Vert f(\sigma) \Vert^p_E \nonumber \\ &= \dfrac{(k+1)!}{(k-j)!} \Vert f \Vert_{(k,p)}^p, \nonumber\end{aligned}$$
where the first and second equality follow by Proposition \[BS1.10\] and writing $\tau * \eta$ as $\sigma \in \Sigma(k)$, respectively. On the other hand, $(*)$ follows since summing over all $\sigma \in \Sigma(k)$ such that $\tau \subset \sigma$ amounts to summing over each term in the previous sum $(k+1)! / ((k+1)-(j+1))! = (k+1)!/(k-j)!$ times recalling that $\omega$ is symmetric and $f$ alternating. $(**)$ follows by Proposition \[switchingsums\], and finally $(***)$ follows since there are $(k+1)!/(k-j)!$ terms independent of $\tau$ in the sum over all $\tau \in \Sigma(j)$ with vertices in $\sigma$.
\[combinedsumming\] Suppose $f \in L^{(k,p)}(X,E)$. If $1 < k + 1 \leq n$, then $$\sum_{\tau \in \Sigma(0, \Gamma)} \Vert f^\tau \Vert_{(k,p)}^p = \dfrac{n-k}{k+1} \sum_{\tau \in \Sigma(0, \Gamma)} \Vert f_\tau \Vert_{(k-1,p)}^p.$$
Follows immediately by Proposition \[oppenheim\] and Proposition \[summing\] above in the case $j=0$, $$\begin{aligned}
\dfrac{k!}{(k+1)!} \sum_{\tau \in \Sigma(0, \Gamma)} \Vert f_\tau \Vert_{(k-1,p)}^p = \Vert f \Vert^p_{(k,p)} = \dfrac{1}{(n-k)} \sum_{\tau \in \Sigma(0, \Gamma)} \Vert f^\tau \Vert^p_{(k,p)}. \nonumber\end{aligned}$$
Define the localized average over a cochain $\phi$ by the map $$M : L^{(k,p)}(X_\tau,E) \rightarrow L^{(k,p)}(X_\tau,E)$$ $\phi_\tau \mapsto M \phi_\tau = \phi_\tau^0$ where $$\phi_\tau^0 (\sigma) = \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{-1} \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \phi_\tau(\sigma).$$ Similarly, we define its dual as $\overline{M} : L^{(k,p^*)}(X_\tau,E^*) \rightarrow L^{(k,p^*)}(X_\tau,E^*)$.
\[Mbound\] The map $M : L^{(k,p)}(X_\tau,E) \rightarrow L^{(k,p)}(X_\tau,E)$ and its dual $\overline{M}$ are bounded projections onto the space of constant maps.
$M$ is well-defined. Towards this end, let $\phi_\tau \in L^{(k,p)}(X_\tau,E)$. Since $\omega_\tau$ is symmetric and $\Gamma_\tau$-invariant, and $\phi_\tau$ is alternating and twisted by $\pi_\tau$, $M \phi_\tau$ is alternating and twisted by $\pi_\tau$ as a finite weighted sum of such functions. Moreover, $$\begin{aligned}
\Vert M \phi_\tau \Vert^p_{(k,p)} &= \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert} \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \Vert \phi^0_\tau (\eta) \Vert^p_E \nonumber \\ &= \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{-p} \left\Vert \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \phi_\tau(\sigma) \right\Vert_E^p \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \nonumber \\ &= \dfrac{1}{(k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{1-p} \left\Vert \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \phi_\tau(\sigma) \right\Vert_E^p \nonumber \\ &\leq \dfrac{C^p}{(k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{1-p} \left\Vert \sum_{\sigma \in \Sigma_\tau(k)} \phi_\tau(\sigma) \right\Vert_E^p \nonumber \\ &\leq \dfrac{C^p \vert \Sigma_\tau(k) \vert^p}{(k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{1-p} \sum_{\sigma \in \Sigma_\tau(k)} \left\Vert \phi_\tau(\sigma) \right\Vert_E^p \nonumber \\ &= \dfrac{C^p \vert \Sigma_\tau(k) \vert^p}{(k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{1-p} \sum_{\sigma \in \Sigma_\tau(k)} \dfrac{\omega_\tau(\sigma)}{\omega_\tau(\sigma)}\left\Vert \phi_\tau(\sigma) \right\Vert_E^p \nonumber \\ &\leq \dfrac{C^p \vert \Sigma_\tau(k) \vert^p}{D (k+1)! \vert \Gamma_\tau \vert} \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{1-p} \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma)\left\Vert \phi_\tau(\sigma) \right\Vert_E^p \nonumber \\ &\leq \dfrac{C^p \vert \Sigma_\tau(k) \vert^p}{D} \vert \Sigma_\tau(k) \vert^{1-p} D^{1-p} \Vert \phi_\tau \Vert^p_{(k,p)} \nonumber \\ &= \dfrac{(C/D)^p}{\vert \Sigma_\tau(k) \vert} \Vert \phi_\tau \Vert^p_{(k,p)} \leq (C/D)^p \Vert \phi_\tau \Vert^p_{(k,p)} \nonumber\end{aligned}$$ where
1. $C = \max \lbrace \omega_\tau(\sigma) \colon \sigma \in \Sigma_\tau(k) \rbrace$ which exists as $\Sigma_\tau (k)$ contains only finitely many $k$-simplexes;
2. $D = \min \lbrace \omega_\tau(\sigma) \colon \sigma \in \Sigma_\tau(k) \rbrace$.
Hence, $M \phi_\tau \in L^{(k,p)}(X_\tau,E)$ and $M$ is well-defined and bounded. Clearly $M$ is linear and $$\begin{aligned}
M^2 \phi_\tau &= M \phi_\tau^0 \nonumber \\ &= \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{-1} \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \phi_\tau^0(\sigma) \nonumber \\ &= \left( \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \right)^{-1} \sum_{\sigma \in \Sigma_\tau(k)} \omega_\tau(\sigma) \left( \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \right)^{-1} \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \phi_\tau(\eta) \nonumber \\ &= \left( \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \right)^{-1} \sum_{\eta \in \Sigma_\tau(k)} \omega_\tau(\eta) \phi_\tau(\eta) \nonumber \\ &= M \phi_\tau, \nonumber\end{aligned}$$ so $M$ is a continuous projection onto $ \lbrace f : \Sigma_\tau(k) \rightarrow E \colon f = \, \mathrm{constant} \rbrace \subseteq L^{(k,p)}(X_\tau,E)$. Similarly for $\overline{M}$
\[averagenorms\] Let $0 \leq j < k \leq n$, $\tau \in \Sigma(j)$ and $\phi \in L^{(k,p^*)}(X,E^*)$. Then,
1. if $j < k-1$, then $\delta_\tau \phi_\tau = (-1)^{j+1}(\delta \phi)_\tau$;
2. if $j= k-1$, then $(-1)^k(n-k+1) \phi^0_\tau = \delta \phi(\tau)$ and $$\Vert \phi^0_\tau \Vert^{p^*}_{(0,p^*)} = \dfrac{\omega(\tau)}{(n-k+1)^{p^*-1}\vert \Gamma_\tau \vert} \Vert \delta \phi(\tau) \Vert^{p^*}_{E^*}.$$
$(1)$ As $\phi \in L^{(k,p^*)}(X,E^*)$, it follows that $\phi_\tau \in L^{(k-j-1,p^*)}(X_\tau,E^*)$ where $k-j-1 > 0$ so $\delta_\tau \phi_\tau \in L^{(k-j-2,p^*)}(X_\tau, E^*)$ and by Proposition \[codifferential\], $$\begin{aligned}
\delta_\tau \phi_\tau (\sigma) &\substack{(*) \\ =} \sum_{\substack{v \in \Sigma_\tau(0) \\ v * \sigma \in \Sigma_\tau(k-j-1)}} \dfrac{\omega_\tau(v * \sigma)}{\omega_\tau(\sigma)} \phi_\tau(v * \sigma) \nonumber \\ &= \sum_{\substack{v \in \Sigma_\tau(0) \\ v * \sigma \in \Sigma_\tau(k-j-1)}} \dfrac{{\omega}(\tau * v * \sigma)}{{\omega}(\tau * \sigma)} \phi(\tau * v * \sigma) \nonumber \\ &\substack{(**) \\ =} \sum_{\substack{v \in \Sigma_\tau(0) \\ v * \sigma \in \Sigma_\tau(k-j-1)}} (-1)^{j+1} \dfrac{{\omega}(v * \tau * \sigma)}{{\omega}(\tau * \sigma)} \phi(v * \tau * \sigma) \nonumber \\ &= \sum_{\substack{v \in \Sigma(0) \\ v * \tau * \sigma \in \Sigma(k)}} (-1)^{j+1} \dfrac{{\omega}(v * \tau * \sigma)}{{\omega}(\tau * \sigma)} \phi(v * \tau * \sigma) \nonumber \\ &= (-1)^{j+1} \delta \phi(\tau * \sigma) = (-1)^{j+1} \left( \delta \phi \right)_\tau (\sigma) \nonumber\end{aligned}$$ where $(*)$ follows by Proposition \[codifferential\] and $(**)$ holds since $\omega$ is symmetric, $\phi$ alternating and $\tau \in \Sigma(j)$. As for $(2)$, by Proposition \[codifferential\] together with the fact that $\omega$ is symmetric and $\phi$ antisymmetric it follows for the codifferential that $$\begin{aligned}
\delta\phi(\tau) &= \dfrac{1}{{\omega}(\tau)} \sum_{\sigma \in \Sigma_\tau(0)} \omega(\sigma * \tau) \phi (\sigma * \tau) = \dfrac{1}{{\omega}(\tau)} \sum_{\sigma \in \Sigma_\tau(0)} (-1)^{j+1}\omega(\tau * \sigma) \phi (\tau * \sigma) \nonumber \\ &= \dfrac{1}{{\omega}(\tau)} \sum_{\sigma \in \Sigma_\tau(0)} (-1)^{k}\omega(\tau * \sigma) \phi (\tau * \sigma), \nonumber\end{aligned}$$ since $j=k-1$. Therefore, in terms of $\phi_\tau^0 \in L^{(0,p^*)}(X_\tau,E^*)$ $$\begin{aligned}
\delta \phi(\tau) = \dfrac{(-1)^k}{{\omega}(\tau)} \sum_{\sigma \in \Sigma_\tau(0)} \omega_\tau(\sigma) \phi_\tau (\sigma) = \dfrac{(-1)^k}{{\omega}(\tau)} \left( \sum_{\sigma \in \Sigma_\tau(0)} \omega_\tau(\sigma)\right) \phi_\tau^0. \nonumber\end{aligned}$$ However, $$\begin{aligned}
\dfrac{\displaystyle \sum_{\sigma \in \Sigma_\tau(0)} \omega_\tau(\sigma)}{\omega(\tau)} = \dfrac{\displaystyle \sum_{\tau * \sigma \in \Sigma(j+1)} {\omega}(\tau * \sigma)} {\omega(\tau)} = \dfrac{(n-j)(j+1)!}{(j+1)!} = (n-j) = (n-k+1) \nonumber\end{aligned}$$ by Proposition \[combinatorial\]. The factor $(j+1)!$ in the denominator corresponds to the fact that we sum over one ordering as $\tau$ is fixed. Therefore, $$\delta \phi(\tau) = (-1)^k(n-k+1)\phi_\tau^0,$$ and once again by Proposition \[BS1.10\] this gives $$\begin{aligned}
\Vert \phi^0_\tau \Vert^{p^*}_{(0,p^*)} &= \dfrac{1}{\vert \Gamma_\tau \vert} \sum_{\sigma \in \Sigma_\tau(0)} \Vert \phi^0_\tau(\sigma)\Vert^{p^*}_{E^*} \omega_\tau(\sigma) = \dfrac{1}{\vert \Gamma_\tau \vert} \sum_{\sigma \in \Sigma_\tau(0)} \dfrac{\Vert \delta \phi(\tau) \Vert^{p^*}_{E^*}}{(n-k+1)^{p^*}} \omega_\tau(\sigma) \nonumber \\ &= \dfrac{1}{\vert \Gamma_\tau \vert} \dfrac{\Vert \delta \phi(\tau) \Vert^{p^*}_{E^*}}{(n-k+1)^{p^*}} (n-k+1) \omega(\tau) = \dfrac{\omega(\tau) \Vert \delta \phi(\tau) \Vert^{p^*}_{E^*}}{(n-k+1)^{p^*-1}\vert \Gamma_\tau \vert}. \nonumber\end{aligned}$$
\[try1\] Let $\phi \in L^{(k,p)}(X,E)$ and $\tau \in \Sigma(k-1)$. Then,
1. if $k=1$, $$d_\tau \phi_\tau (\sigma) = -(d \phi)_\tau (\sigma) + \phi (\sigma).$$
2. if $k > 1$, $$d_\tau \phi_\tau (\sigma) = (d \phi)_\tau (\sigma) + \sum_{i=0}^{k-1}(-1)^i\phi (\tau_i * \sigma).$$
Suppose $k=1$. Then, for $\sigma = x * y \in \Sigma_\tau(1)$ $$\begin{aligned}
d_\tau \phi_\tau (\sigma) &= \phi_\tau (y) -\phi_\tau (x) = \phi(\tau * y) - \phi(\tau * x). \nonumber\end{aligned}$$ On the other hand, as $[\tau * x * y : x * y] =1$ $$\begin{aligned}
d \phi (\tau * x * y) &= [\tau * x * y : x * y] \phi(x*y) + [\tau * x * y: \tau * y] \phi (\tau * y) + [\tau * x * y: \tau * x] \phi(\tau * x) \nonumber \\ &= \phi(x*y) - \phi (\tau * y) +\phi(\tau * x), \nonumber\end{aligned}$$ gives together with the expression for $d_\tau \phi_\tau$ $$d_\tau \phi_\tau (\sigma) = -d \phi (\tau * x * y) + \phi (x * y) = -(d \phi)_\tau (\sigma) + \phi (\sigma).$$ Suppose $k>1$ and $\tau \in \Sigma(k-1)$. Then, as previously $$\begin{aligned}
d_\tau \phi_\tau (\sigma) = \phi_\tau (y) - \phi_\tau (x), \nonumber\end{aligned}$$ and the two rightmost terms are as previously the last two terms in $d \phi (\tau * \sigma)$.
\[diffrelnew\] Let $\phi \in L^{(k,p)}(X,E)$ and $\tau \in \Sigma(0)$, then $$(d \phi)_\tau (\sigma) = - d_\tau \phi_\tau (\sigma) + \phi(\sigma).$$
Let $\sigma \in \Sigma(k)$. Then, similarly as in Proposition \[try1\] $$d_\tau \phi_\tau (\sigma) = \sum_{i=0}^{k}(-1)^i \phi_\tau (\sigma_i) = \sum_{i=0}^{k}(-1)^i \phi(\tau * \sigma_i) = - (d \phi)_\tau(\sigma) + \phi(\sigma).$$
\[help1\] Suppose $\phi \in L^{(k,p)}(X,E)$, $\tau \in \Sigma(0)$ and $k+1 \leq n$. Then, if $\phi \in \ker d$, $$\Vert d_\tau \phi_\tau \Vert_{(k,p)} = \Vert \phi^\tau \Vert_{(k,p)}.$$
By Proposition \[diffrelnew\] it follows since $\phi \in \ker d$ that $$\begin{aligned}
\Vert d_\tau \phi_\tau \Vert_{(k,p)}^p &= \sum_{\sigma \in \Sigma_\tau (k, \Gamma_\tau)} \Vert d_\tau \phi_\tau (\sigma)\Vert_E^p \dfrac{\omega_\tau(\sigma)}{(k+1)! \vert \Gamma_{\tau \sigma}\vert} \nonumber \\ &= \sum_{\sigma \in \Sigma_\tau (k, \Gamma_\tau)} \Vert \phi (\sigma)\Vert_E^p \dfrac{\omega_\tau(\sigma)}{(k+1)! \vert \Gamma_{\tau \sigma}\vert} \nonumber \\ &= \Vert \phi^\tau \Vert_{(k,p)}^p. \nonumber\end{aligned}$$
\[try2\] Suppose $\phi \in L^{(k,p)}(X,E)$ and $1 < k+1 \leq n$. If $\phi \in \ker d$, then $$\sum_{\tau \in \Sigma(0, \Gamma)} \Vert d_\tau \phi_\tau \Vert_{(k,p)}^p = \dfrac{n-k}{k+1} \sum_{\tau \in \Sigma(0, \Gamma)} \Vert \phi_\tau \Vert_{(k-1,p)}^p.$$
Follows by Corollary \[help1\] and Proposition \[combinedsumming\].
\[BSspecial\] Suppose $\phi \in L^{(1,p)}(X,E)$. If $\phi \in \ker d$, then $$- (n-1) \Vert \phi \Vert_{(1,p)}^p = \sum_{\tau \in \Sigma(0, \Gamma)} \left( \Vert d_\tau \phi_\tau \Vert_{(1,p)}^p - (n-1) \Vert \phi_\tau \Vert^p_{(0,p)} \right).$$
By a direct computation, $$\begin{aligned}
&\sum_{\tau \in \Sigma(0, \Gamma)} \Vert d_\tau \phi_\tau \Vert_{(1,p)}^p - (n-1) \Vert \phi_\tau \Vert_{(0,p)}^p \substack{(*) \\ =} \sum_{\tau \in \Sigma(0, \Gamma)} \dfrac{(n-1)}{2} \Vert \phi_\tau\Vert_{(0,p)}^p - (n-1) \Vert \phi_\tau\Vert_{(0,p)}^p \nonumber \\ &= -\dfrac{(n-1)}{2} \sum_{\tau \in \Sigma(0, \Gamma)} \Vert \phi_\tau \Vert_{(0,p)}^p \substack{(**) \\ =} - \dfrac{(n-1)}{2} \dfrac{2!}{(1-0)!} \Vert \phi \Vert_{(1,p)}^p = -(n-1)\Vert \phi \Vert_{(1,p)}^p, \nonumber\end{aligned}$$ where in $(*)$ we used Corollary \[try2\] and in $(**)$ Proposition \[summing\].
\[try3\] Suppose $\phi \in L^{(1,p)}(X,E)$. If $\phi \in \ker d$, then $$\sum_{\tau \in \Sigma(0, \Gamma)} \Vert d_\tau \phi_\tau \Vert_{(1,p)}^p = (n-1) \Vert \phi \Vert_{(1,p)}^p.$$
Follows directly from Corollary \[BSspecial\] using Proposition \[summing\] once more.
\[Q\] Let $\phi \in L^{(1,p)}(X,E)$. For $\tau \in \Sigma(0)$ define a $p$-form on $L^{(0,p)}(X_\tau, E)$ by $$Q_\tau (\phi_\tau) = \Vert d_\tau \phi_\tau \Vert^p_{(1,p)} - \dfrac{(n-1)}{2} \Vert \phi_\tau \Vert_{(0,p)}^p.$$ Similarly, we define a $p^*$-form on $L^{(0,p^*)}(X_\tau, E^*)$.
\[propQ\] Suppose $\phi \in L^{(1,p)}(X,E)$. If $\phi \in \ker d$ then $$\sum_{\tau \in \Sigma(0, \Gamma)} Q_\tau (\phi_\tau) = 0.$$
Follows immediately by Corollary \[try3\] and Proposition \[summing\].
Poincaré inequalities on finite weighted graphs {#PI}
===============================================
In this section we recall some basic facts concerning Poincaré inequalities on finite weighted graphs necessary for the spectral method. For details we refer to [@gn; @Nowak; @ny].
[@Nowak] \[poink\] Suppose $\dim X = 2$. Then the link $X_\tau$ of every vertex of $X$ is a finite graph. Hence, for any $p \geq 1$ the $p$-Poincaré inequality $$\begin{aligned}
\sum_{\sigma \in \Sigma(0, \Gamma_\tau)} \Vert f_\tau(\sigma) - Mf_\tau(\sigma)\Vert_E^p &\dfrac{{\omega_\tau}(\sigma)}{\vert \Gamma_{\tau \sigma} \vert} \nonumber \\ &\leq \kappa_p^p \sum_{\eta \in \Sigma_\tau(1, \Gamma_\tau)} \dfrac{1}{2} \Vert f_\tau(\eta_0) - f_\tau(\eta_1) \Vert_E^p \dfrac{\omega_\tau(\eta)}{\vert \Gamma_{\tau \eta} \vert} \nonumber \end{aligned}$$ holds for some $\kappa_p>0$ and all $f \colon \Sigma_\tau(0) \rightarrow E$. Similarly for $f \colon \Sigma_\tau(0) \rightarrow E^*$.
The infimum of the above constants $\kappa_p$ is known as the Poincaré constant of the link $X_\tau$, and denoted by $\kappa_p(X_\tau, E)$. In terms of the norms introduced previously:
Let $X$ be two dimensional. Then, for all $f \in L^{(1,p)}(X,E)$ it holds that $$\Vert f_\tau - Mf_\tau \Vert_{(0,p)} \leq \kappa_{p}(X_\tau,E) \Vert {d}_\tau f_\tau \Vert_{(1,p)},$$ for some $\kappa_{p}(X_\tau,E)$. Similarly for $f \in L^{(1,p^*)}(X,E^*)$, $$\Vert f_\tau - \overline{M}f_\tau \Vert_{(0,p^*)} \leq \kappa_{p^*}(X_\tau,E^*) \Vert \overline{d}_\tau f_\tau \Vert_{(1,p^*)},$$ for some $\kappa_{p^*}(X_\tau,E^*)$.
Sometimes it is useful to know how Poincaré constants change under isomorphisms. The following is immediate:
\[BI\] Let $T: E \rightarrow F$ be a Banach space isomorphism. If $$\Vert x \Vert_E \leq \Vert T(x) \Vert_F \leq C \Vert x \Vert_E,$$ then $\kappa_p(X_\tau, E) \leq C \kappa_p(X_\tau, F)$.
For $1 < p< \infty$ we denote by ${L}^p$ the Banach space ${L}^p(\mu)$ of $p$-integrable functions on a standard Borel space $(Y, \mathcal{B})$ with $\sigma$-finite measure $\mu$. As such, any separable infinite-dimensional Hilbert space $\mathcal{H}$ is isometrically isomorphic to ${L}^2$. In particular, we have the following relation between the Poincaré constant and spectral gap:
\[hilb1\] [@Nowak] Let $\lambda_1$ be the smallest positive eigenvalue of the graph Laplacian $\triangle_+ = (\delta d)_\tau$, defined by $$\triangle_+f(v) = f(v) - \dfrac{1}{\omega_\tau(v)} \sum_{u \in L_v}f(u),$$ where $L_v$ denotes the link of $v$ in $X_\tau$, over the space $C^{(0,p)}(X_\tau, \mathbb{R})$ of real-valued functions on the vertices. For $L^2$ when $X$ is $2$-dimensional, $\kappa_p(X_\tau, L^2) = \lambda_1^{-1/2}$ and more generally $\kappa_p(X_\tau, L^p) = \kappa_p(X_\tau, \mathbb{R})$.
$L^p\,$-cohomology and vanishing for uniformly bounded representations {#LH}
======================================================================
Following [@BS] we introduce $L^p$-cohomology of $X$ with coefficients in $\pi$ as a natural extension of the $L^2$-cohomology for unitary representations. In particular, if $\pi$ is a unitary representation $L^2 H^k(X, \pi)$, as described below, is the cohomology of the complex of $\textrm{mod}$ $\Gamma$ square integrable cochains of $X$ twisted by $\pi$. The connection to property $(T)$ is as follows: if $X$ is a two dimensional contractible simplicial complex and $\Gamma$ acts properly discontinuously and cocompactly by automorphisms on it, then $\Gamma$ has property $(T)$ if and only if $L^2 H^1(X, \pi) = 0$ for any unitary representation [@hv]. As an application we derive a spectral condition for cohomological vanishing for square integrable cochains on a two dimensional simplicial complex twisted by a uniformly bounded representation.
Let $$L^p H^k(X,\pi) = \ker \left( d \vert_{L^{(k,p)}(X,E)} \right) / \mathrm{im} \left( d \vert_{L^{(k-1,p)}(X,E)} \right)$$ denote the $L^p$-cohomology groups of $X$ with coefficients twisted by $\pi$.
As the following shows, cohomological vanishing takes place when $\delta$ is bounded from below:
\[VCH1\] The map $$d_{k-1} \vert_{L^{(k-1,p)}(X,E)} : L^{(k-1,p)}(X,E) \longrightarrow \ker \, d_k \vert_{L^{(k,p)}(X,E)}$$ is onto if its adjoint $$\delta_{k} : (\ker \, d_{k} \vert_{L^{(k,p)}(X,E)})^* \rightarrow L^{(k-1,p^*)}(X,E^*),$$ is bounded from below, that is $\exists \, K > 0$ such that for all $f \in \left( \ker {d_k} \vert_{L^{(k,p)}(X,E)} \right)^*$ $$\Vert \delta f \Vert_{(k-1,p^*)} \ge K \Vert f \Vert_{(k,p^*)}.$$ If in addition $d_{k-1}$ is injective, $d_{k-1}$ is onto if and only if $\delta_k$ is bounded from below.
Since $d_{k} \circ d_{k-1} \vert_{L^{(k-1,p)}(X,E)} = 0$, $\mathrm{im} \, d_{k-1} \subseteq \ker d_k \vert_{L^{(k,p)}(X,E)}$ without further assumptions. Now, assume $\delta_k$ is bounded from below. Then $\delta_k$ is injective; towards a contradiction, suppose $f,g \in (\ker \, d_{k} \vert_{L^{(k,p)}(X,E)})^*$ such that $f \neq g$ and $\delta_k \, f = \delta_k \, g$. Recalling that $\Vert \cdot \Vert_{(k,p^*)}$ is a norm restricted to $L^{(k,p^*)}(X,E)$ leads to a contradiction $$\begin{aligned}
0 = \Vert \delta_k f - \delta_k g \Vert_{(k-1,p^*)} = \Vert \delta_k (f-g)\Vert_{(k-1,p^*)} \geq K \Vert f - g\Vert_{(k,p^*)} > 0. \nonumber\end{aligned}$$ Thus, $\delta$ is injective. In particular, $\ker \, \delta_k = \lbrace 0 \rbrace$ and since $L^{(k,p)}(X,E)$ is reflexive and $\ker \, d_k\vert_{L^{(k,p)}(X,E)}$ is closed, the latter is also reflexive and $$\begin{aligned}
\mathrm{im} \, d_{k-1} &= \mathrm{Ann}(\ker \, \delta_k) \nonumber \\ &= \lbrace f \in {(\ker \, d_{k} \vert_{L^{(k,p)}(X,E)})^*}^* \colon \langle g,f \rangle_{k} = 0, \, \forall g \in \ker \, \delta_k \rbrace \nonumber \\ &\cong \lbrace f \in \ker \, d_{k} \vert_{L^{(k,p)}(X,E)} \colon \langle g,f \rangle_{k} = 0, \, \forall g \in \lbrace 0 \rbrace \rbrace \nonumber \\ &= \ker \, d_{k} \vert_{L^{(k,p)}(X,E)}, \nonumber \end{aligned}$$ so $d_{k-1}$ is onto $\ker d_k \vert_{L^{(k,p)}(X,E)}$. Next, suppose $d_{k-1}$ is onto. Since $d_{k-1}$ is bounded, it is bounded from below by the open mapping theorem if $d_{k-1}$ is injective.
This criteria is in fact related to the Poincaré constants of the links as Proposition \[bedlewoinequalityNOT\] shows. This allows us to formulate a spectral condition for cohomological vanishing.
\[FIX1\] Suppose $X$ is a $2$-dimensional locally finite simplicial complex such that for any vertex $\tau$ of $X$ the link $X_\tau$ is connected and $\mathcal{H}$ a separable infinite-dimensional Hilbert space. Suppose there exists a constant $C$ such that
1. the map $C \mathcal{I}: E \rightarrow \mathcal{H}$ where $\mathcal{I}: E \rightarrow \mathcal{H}$ is the identity map, is a Banach space isomorphism with the property $\Vert x \Vert_E \leq \Vert C\mathcal{I}(x) \Vert_\mathcal{H} \leq C \Vert x \Vert_E$ for all $x \in E$. Then, for $f \in L^{(1,2)}(X,E) \cap \ker {d}$ $$\kappa_2(X_\tau, \mathcal{H})^{-2} \Vert Mf_\tau\Vert^2_{(0,2)} + Q_\tau(f_\tau) \geq \dfrac{1}{C^2}\left( \kappa_2(X_\tau, \mathcal{H})^{-2} - \dfrac{C^2}{2}\right) \Vert f_\tau \Vert^2_{(0,2)};$$
2. the identity map $\overline{\mathcal{I}}: E^* \rightarrow \mathcal{H}$ is a Banach space isomorphism with the property $\Vert x \Vert_{E^*} \leq \Vert \overline{\mathcal{I}}(x) \Vert_{\mathcal{H}} \leq C \Vert x \Vert_{E^*}$ for all $x \in {E^*}$. Then, for $f \in L^{(1,2)}(X,E^*) \cap \ker \overline{d}$ $$\kappa_2(X_\tau, \mathcal{H})^{-2} \Vert \overline{M}f_\tau\Vert^2_{(0,2)} + Q_\tau(f_\tau) \geq \dfrac{1}{C^2}\left( \kappa_2(X_\tau, \mathcal{H})^{-2} - \dfrac{C^2}{2}\right) \Vert f_\tau \Vert^2_{(0,2)}.$$
$(i)$. As usual, write $\Vert \cdot \Vert_{(k,2)}$ for the seminorm on $\mathcal{E}^{(k,2)}(X,E)$ and write $\Vert \cdot \Vert_{(k,2), \mathcal{H}}$ for the seminorm on $\mathcal{E}^{(k,2)}(X,\mathcal{H})$. For $f \in L^{(1,2)}(X,E) \cap \ker {d}$ $$\begin{aligned}
Q_\tau(f_\tau) = \Vert d_\tau f_\tau \Vert_{(1,2)}^2 - \dfrac{1}{2} \Vert f_\tau \Vert^2_{(0,2)} \geq \Vert d_\tau f_\tau \Vert_{(1,2), \mathcal{H}}^2 - \dfrac{C^2}{2} \Vert f_\tau \Vert^2_{(0,2), \mathcal{H}} \nonumber\end{aligned}$$ On the other hand, by the Poincaré inequality and the Pythagorean identity $$\Vert d_\tau f_\tau \Vert_{(1,2), \mathcal{H}}^2 \geq \kappa_2(X_\tau, \mathcal{H})^{-2}\Vert f_\tau\Vert^2_{(0,2), \mathcal{H}} - \kappa_2(X_\tau, \mathcal{H})^{-2}\Vert Mf_\tau\Vert^2_{(0,2), \mathcal{H}}$$ Hence, $$\begin{aligned}
Q_\tau(f_\tau) \geq \kappa_2(X_\tau, \mathcal{H})^{-2}\Vert f_\tau\Vert^2_{(0,2), \mathcal{H}} - \kappa_2(X_\tau, \mathcal{H})^{-2}\Vert Mf_\tau\Vert^2_{(0,2), \mathcal{H}} - \dfrac{C^2}{2} \Vert f_\tau \Vert^2_{(0,2), \mathcal{H}}, \nonumber \end{aligned}$$ and in terms of the $\Vert \cdot \Vert_{(k,p)}$ norm $$\begin{aligned}
\kappa_2(X_\tau, \mathcal{H})^{-2}\Vert Mf_\tau\Vert^2_{(0,2)} + Q_\tau(f_\tau) \geq \dfrac{1}{C^2}\left( \kappa_2(X_\tau, \mathcal{H})^{-2} - \dfrac{C^2}{2} \right) \Vert f_\tau \Vert^2_{(0,2)}. \nonumber\end{aligned}$$ Similarly for $(ii)$.
\[bedlewoinequalityNOT\] Assuming Proposition \[FIX1\] holds such that $\kappa_2(X_\tau,\mathcal{H}) \leq \kappa_2(X,\mathcal{H})$ for every link $X_\tau$ of $X$. Then, for $$\kappa_2(X,\mathcal{H}) < \sqrt{2}C^{-1},$$ $\delta$ and $\overline{\delta}$ are bounded from below.
By Proposition \[FIX1\](i) $$\begin{aligned}
\kappa_2(X_\tau, \mathcal{H})^{-2}\Vert Mf_\tau\Vert^2_{(0,2)} + Q_\tau(f_\tau) \geq \dfrac{1}{C^2}\left( \kappa_2(X_\tau, \mathcal{H})^{-2} - \dfrac{C^2}{2} \right) \Vert f_\tau \Vert^2_{(0,2)}. \nonumber\end{aligned}$$ Thus, summing over the representatives $\tau \in \Sigma(0, \Gamma)$ gives, applying Propositions \[summing\], \[averagenorms\], and \[propQ\] to the three terms respectively, that $$\begin{aligned}
\Vert \delta f\Vert^2_{(0,2)} \geq \left( \dfrac{2 \kappa_2(X,\mathcal{H})}{C}\right)^2 \left( \kappa_2(X_\tau, \mathcal{H})^{-2} - \dfrac{C^2}{2} \right) \Vert f \Vert_{(1,2)}^2. \nonumber\end{aligned}$$ So, $\delta$ is bounded from below for $\kappa_2(X,\mathcal{H}) < {\sqrt{2}}{C^{-1}}$. Similarly for $\overline{\delta}$.
\[theorem\] Let $X$ be a locally finite $2$-dimensional simplicial complex, $\Gamma$ a discrete properly discontinuous group of automorphisms of $X$ and $\pi : \Gamma \rightarrow \mathrm{B}(\mathcal{H})$ a uniformly bounded representation of $\Gamma$ on a separable infinite-dimensional Hilbert space $\mathcal{H}$. Suppose the link $X_\tau$ of every vertex $\tau$ of $X$ is connected and the associated Poincaré constants satisfy $$C < \dfrac{\sqrt{2}}{\kappa_2(X_\tau, \mathcal{H})}$$ for $C = \sup_{g \in \Gamma} \Vert \pi_g \Vert$. Then, $L^2H^1(X,\pi) = 0$.
Let $E$ be the Banach space $(\mathcal{H}, \Vert \cdot \Vert_E)$ where $\Vert \cdot\Vert_E = \sup_{g \in \Gamma}\Vert \pi_g(\cdot)\Vert_\mathcal{H}$. Now, $\pi$ is an isometric representation on $E$ and have the dual diagrams:\
\
We claim that $L^2H^1(X,E) = 0$, that is $d_0$ is onto $\ker d_1$. By Proposition \[VCH1\] it is enough to prove that $d^*_0 = \delta_1$ is bounded from below on $(\ker {d}_1)^*$. Since Proposition \[FIX1\] holds for $C = \sup_{g \in \Gamma} \Vert \pi_g \Vert$, and $ \kappa_2(X_\tau, \mathcal{H}) < \sqrt{2} C^{-1}$, it follows by Corollary \[bedlewoinequalityNOT\] that $\delta_1$ is bounded from below when restricted to $\ker \overline{d}$. Hence, $\delta_1$ is bounded from below on the image of $i^* \circ \overline{i}$. Thus, if $i^* \circ \overline{i}$ is onto $(\ker d_1)^*$, then $\delta_1$ is bounded from below on $(\ker d_1)^*$ and ${\delta_1 \,}^* = d_0$ is onto, by which the claim follows. By a similar argument ${\overline{d^*_0}}$ restricted to $\ker d_1$ is bounded from below, and thus $${\overline{d}_0\,}^* \circ {\overline{i}\,}^* \circ i \colon \ker d_1 \rightarrow L^{(0,p)}(X,E)$$ is bounded from below. In particular, ${\overline{i}\,}^* \circ i$ is bounded from below and hence $({\overline{i}\,}^* \circ i)^* = i^* \circ \overline{i}$ is onto $(\ker d_1)^*$.
[100]{} U. Bader, A. Furman, T. Gelander, N. Monod, Property $(T)$ and rigidity of actions on Banach spaces, Acta Math. 198 (2007), no. 1, 57-105. W. Ballman, J. Światkowski, On $L^2$-cohomology and property $(T)$ for automorphism groups of polyhedral cell complexes, Geom. Funct. Anal. 7 (1997), no. 4, 615-645. B. Becka, P. de la Harpe, A. Valette, Kazhdan Property (T), New Mathematical Monographs, vol. 11, Cambridge University Press, Cambridge (2008). I. Chatterji, C. Druţu, F. Haglund, Kazhdan and Haagerup properties from the median viewpoint, Adv. Math. 225 (2010), no. 2, 882-921. R. G. Douglas, Banach Algebra Techniques in Operator Theory, Academic Press (1972). J. Dymara, T. Januszkiewicz, Cohomology of buildings and their automorphism groups, Invent. Math. 150 (2002), no. 3, 579-627. J. Dymara, T. Januszkiewicz, New Kazhdan groups, Geom. Dedicata, 80 (2000), no. 1-3, 311-317. M. Ershov, A. Jaikin-Zapirain, Property $(T)$ for noncommutative universal lattices, Invent. Math 179 (2010), no. 2, 303-347. D. Fisher, G. Margulis, Almost isometric actions, property $(T)$, and local rigidity, Invent. Math. 162 (2005), no. 1, 19-80. H. Garland, $p$-adic curvature and and the cohomology of discrete subgroups of $p$-adic subgroups, Annals of Math. 97 (1973), 375-423. R. I. Grigorchuk, P. W. Nowak, Diameters, distortion, and eigenvalues, Europ. J. Combin. 33 (2012), no. 7, 1574-1587 P. de la Harpe, A. Valette, La propriété $(T)$ de Kazhdan pour les groupes localement compacts, Astérisque 175, Soc. Math. France, 1989. D. A. Kazhdan, Connection of the dual space of a group with the structure of its closed subgroups, Funct. Anal. and Appl. 1 (1967), 63-65. V. Lafforgue, Un renforcement de la propriété $(T)$, Duke Math. J. 143 (2008), no 3, 559-602. P. W. Nowak, Poincaré inequalities and rigidity for actions on Banach spaces, preprint arXiv:1107.1896v2 \[math.GR\] P. W. Nowak. G. Yu, Large Scale Geometry, European Mathematical Society Publishing House, Textbook in Mathematics series (2012). Oberwolfach Report, no. 29/2001. I. Oppenheim Vanishing of $L^2$-cohomology and property (T) for groups acting on simplicial complexes, preprint arXiv:1110.5724v3. A. Żuk, La propriété de Kazhdan pour les groupes agissant sur les polyèdres, C.R. Acad. Sci. Paris Sér. I Math. 323 (1996), no. 5, 453-458. A. Żuk, Property (T) and Kazhdan constants for Discrete Groups, Geom. Funct. Anal. 13 (2003), no. 3, 643-670.
[^1]: Supported by Väisälä Foundation and the Academy of Finland, project 252293.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $Q(X)$ be any integral primitive positive definite quadratic form in $k$ variables where $k\geq4$ and discriminant $D$. We give an upper bound on the number of integral solutions of $Q(X)=n$ for any integer $n$ in terms of $n$, $k$ and $D$. As a corollary, we give a definite answer to a conjecture of Lester and Rudnick on the small scale equidistribution of orthonormal basis of eigenfunctions restricted to an individual eigenspace on the flat torus $\mathbb{T}^d$ for $d\geq 5$. Another application of our main theorem gives a sharp upper bound on $A_{d}(n,t)$, the number of representation of the positive definite quadratic form $Q(x,y)=nx^2+2txy+ny^2$ as a sum of squares of $d\geq 5$ binary linear forms where $n- n^{\frac{1}{(d-1)}-o(1)}< t < n$. This upper bound allows us to study the local statistics of integral points on sphere.'
author:
- Naser T Sardari
bibliography:
- 'Preprint.bib'
date: Spring 2016
title: Quadratic forms and semiclassical eigenfunction hypothesis for flat tori
---
Introduction
============
Semiclassical eigenfunction hypothesis for flat tori
----------------------------------------------------
We begin by describing the main application of this paper. Let $\mathbb{T}^d:= \mathbb{R}^d/\mathbb{Z}^d$ be the flat torus of dimension $d\geq2$ with the laplacian operator $\Delta:=\frac{1}{(2\pi)^2 }( \frac{{\partial}^2}{\partial \theta_1}+ \dots +\frac{{\partial}^2}{\partial \theta_d})$, and $\{ \psi_i \}$ an orthonormal basis of eigenfunctions. Marklof and Rudnick [@Marklof] showed that for flat torus $\mathbb{T}^d$ there is a density one subsequence of any orthonormal basis $\{\psi_i\}$ such that $|\psi_i|^2$ converges weakly to the uniform distribution in $\mathbb{T}^d$, i.e. for any continuous function $f$ on $\mathbb{T}^d$
$$\int_{\mathbb{T}^d} |\psi_n(x)|^2f(x) d\vol(x) \rightarrow \int_{\mathbb{T}^d} f(x) d\vol(x) \text { as $n \to \infty$. }$$ M.V. Berry [@Berry; @Michael] in his work on the “ Semiclassical Eigenfunction Hypothesis” suggested to go beyond this weak convergence, and study the equidistribution of $|\psi_n|^2$ on small scale. Hezari and Riviere [@Hamid] established the first result on the small scale equidistribution on rational flat tori for balls with a radius shrinking at a polynomial rate. Motivated by their work, Lester and Rudnick [@Rudnick] proved the equidistribution of a density one subsequence of $\{\psi_i\}$ in an optimal small scale on the flat torus $\mathbb{T}^d$ for $d\geq 2$. Density one means if we order the eigenfunctions by their eigenvalue then the subsequence contains a density one subset of the basis. More precisely, they show that along a density one subsequence of the orthonormal basis $\{\psi_n \}$,
$$\label{smallscale} \lim_{n\to \infty} \sup_{B(y,r)}\Big|\frac{1}{\vol (B(y,r))}\int_{B(y,r)} |\psi_n(x)|^2 d \vol(x)-1\Big|=0,$$
where $\psi_n$ is an eigenfunction with eigenvalue $\lambda_n$ and $B(y,r)$ is any ball of radius $r > \lambda_n^{-\frac{1}{2(d-1)}+o(1)}$ centered at an arbitrary point $x \in \mathbb{T}^d $. The exponent $-\frac{1}{2(d-1)}$, for the size of the ball, is optimal. Moreover, for dimension 3 and 4, they prove a stronger result. They prove the small scale equidistribution holds for almost every eigenfunction in individual eigenspace with the optimal exponent $-\frac{1}{2(d-1)} $ and they conjecture that the same result holds for every $d\geq 5$ . The main application of this paper is to resolve this conjecture. In what follows, we explain their result and conjecture in detail and how it is reduced to counting pairs of integral points with small distance on sphere.
We note that if $d\geq 4$ then the set of eigenvalues of the laplacian $\Delta$ on $\mathbb{T}^d$ are given by the set of non-negative integers $\{n\in \mathbb{Z} : n\geq 0 \}$. We denote the associated eigenspace by $E_n$. It is well-known that the dimension of this space is equal to the number of integral points on sphere of radius $\sqrt{n}$ in $\mathbb{R}^d$. Therefore, for $d \geq 5$, the multiplicity of each eigenvalue $n$ is large and grow like a scalar multiple of $n^{\frac{d-2}{2}}$. The intersection of the orthonormal basis of eigenfunctions $\{ \psi_i\}$ with $E_n$ gives us an orthonormal basis of eigenfunctions for $E_n$. We denote this orthonormal basis of $E_n$ by $B_n:= E_n \cap B$. Lester and Rudnick [@Rudnick] prove that there exists a large subset $C_n \subset B_n$, which means $|C_n|=(1-o(1))|B_n|$, such that along any sequence of $\{\psi_n : \psi_n \in C_n \}$ $$\label{smallscale} \lim_{n\to \infty} \sup_{B(y,r)}\Big|\frac{1}{\vol (B(y,r))}\int_{B(y,r)} |\psi_n(x)|^2 d \vol(x)-1\Big|=o(1),$$ where $B(y,r)$ is any ball of radius $r > \lambda_n^{-\frac{1}{2(d-1)}+o(1)}$ centered at an arbitrary point $x \in \mathbb{T}^d $. This means that we can choose our density one subsequence so that it contains a large proportion of the basis restricted to each eigenspace. In this paper we prove the analogue of this result for $d\geq 5$.
\[small\]Let $\mathbb{T}^d:= \mathbb{R}^d/\mathbb{Z}^d$ be a $d\geq5$ dimensional flat torus and $\{ \psi_i \}$ an orthonormal basis of eigenfunctions for the laplacian operator. Let $E_n$ be the associated eigenspace for eigenvalue $n \in\mathbb{Z} \geq 0$ and $B_n:= E_n \cap $ $\{ \psi_i \}$ be the restriction of the orthonormal basis $\{ \psi_i \}$ to $E_n$. Then there exists a large subset $C_n \subset B_n$, which means $|C_n|=(1-o(1))|B_n|$, such that along any sequence of $\{\psi_n : \psi_n \in C_n \}$ $$\label{smallscale} \lim_{n\to \infty} \sup_{B(y,r)}\Big|\frac{1}{\vol (B(y,r))}\int_{B(y,r)} |\psi_n(x)|^2 d \vol(x)-1\Big|=o(1),$$ where $B(y,r)$ is any ball of radius $r > \lambda_n^{-\frac{1}{2(d-1)}+o(1)}$ centered at an arbitrary point $x \in \mathbb{T}^d $. The exponent $-\frac{1}{2(d-1)}$ is optimal. In fact there exists a sequence of balls with radius $r\approx \lambda_n^{-\frac{1}{2(d-1)}-o(1)}$ and an orthonormal basis $\{ \psi_{i}\}$ where (\[smallscale\]) does not hold for a positive proportion of eigenfunctions in $B_n$.
Local statistics of integral points on sphere
---------------------------------------------
In this section we explain Lester and Rudnick’s [@Rudnick]\[Remark 5.4\] observation that the small scale equidistribution for individual eigenspace on flat tori is related to counting pairs of integral points with small distance on sphere.
Let $\psi \in B_n$ be an orthonormal function in our basis then we can write
$$\label{1}
\psi(\theta)=\sum_{\lambda \in \mathbb{Z}^d, \| \lambda\|^2=n} c_{\lambda}(\psi) e(\langle \lambda,\theta \rangle),$$
where $$\sum_{\lambda \in \mathbb{Z}^d, \| \lambda\|^2=n} \|c_{\lambda}(\psi)\|^2=1.$$ Since $B_n$ is an orthonormal basis for $E_n$, similarly if we fix $\lambda \in \mathbb{Z}^d$ we also obtain $$\label{orthogonality}
\sum_{\psi \in B_n} \|c_{\lambda}(\psi)\|^2=1.$$ Lester and Rudnick estimate the indicator function of the ball $B(x,r) $ by the majorant and minorant Beurling-Selberg trigonometric polynomials $a^{\pm} (\theta)$ on the flat torus $\mathbb{T}^d$ such that
$$\begin{aligned}
a(\theta)^{\pm}&=&\sum_{\xi \in \mathbb{Z}^d} a(\xi)e(\langle \theta, \xi \rangle),
\\
a^{\pm}(0)^{\pm}&=& \vol(B(x,r))+O(r^{d-o(1)}),
\\
\hat{a}(\xi)^{\pm}&=&0 \text{ if } |\xi|> n^{\frac{1}{2(d-1)}-o(1)},
\\
\big| \hat{a}(\xi)^{\pm} \big| &\leq & r^d.\end{aligned}$$
Then one can estimate from below and above the local integral (\[smallscale\]) by using the trigonometric polynomials $a^{\pm}$ and reduce the problem into counting pairs of integral points with small distance on the sphere. We include a brief exposition of this reduction in what follows $$\begin{aligned}
\frac{\int_{\mathbb{T}^d} \| \psi \|^2 a^{\pm}(\theta) d\theta }{a^{\pm}(0)}-1&=& \int \big| \sum_{\lambda}c_{\lambda}(\psi)e( \langle \lambda,\theta \rangle ) \big|^2 \big( \sum_{\xi} \frac{\hat{a}^{\pm}(\xi)}{\hat{a}^{\pm}(0)} e( \langle \xi,\theta \rangle ) \big) d\theta
\\
&=&\sum_{\lambda-\lambda^{\prime}=\xi} c_{\lambda}(\psi)\bar{c}_{\lambda^{\prime}} \frac{\hat{a}^{\pm}(\xi)}{\hat{a}^{\pm}(0)}
\\
&\ll& \sum_{0<\|\lambda-\lambda^{\prime}\|\leq n^{\frac{1}{2(d-1)}-o(1) }} \|c_{\lambda}(\psi)\|^2+ \|c_{\lambda^{\prime}}(\psi)\|^2.\end{aligned}$$ Average this inequality over $\psi \in B_n$ and use the identity (\[orthogonality\]) to obtain $$\begin{aligned}
\sum_{\psi \in B_n }\Big[ \sup_{B(y,r)}\Big|\frac{1}{\vol (B(y,r))}\int_{B(y,r)} |\psi(x)|^2 d \vol(x)-1\Big| \Big]
\\
\leq \frac{1}{|B(n)|}\sum_{0<\|\lambda-\lambda^{\prime}\|\leq n^{\frac{1}{2(d-1)}-o(1) }} 1.
\end{aligned}$$ Note that the right hand side of the above inequality is the average over the pair of integral points $\lambda\neq \lambda^{\prime}$ on the sphere of radius $\sqrt{n}$ where their distance is less than $n^{\frac{1}{2(d-1)}-o(1) }$. For a large integer $n$, let
$$E(n)=E_d(n)=\{x\in \mathbb{Z}^d :|x|^2 =n\},$$ be the set of integral lattice points on the sphere $\sqrt{n}S^{d-1}$ of radius $\sqrt{n}$. For any $Y \ll \sqrt{n}$, and $|x|^2 =n$, let
$$\text{cap}(x;n,Y):=\{y \in\sqrt{n}S^{d-1} , |x-y|\leq Y\},$$ be the spherical cap of size $Y$ around the point $x$ on the sphere $\sqrt{n}S^{d-1}$. Given a point $x \in E(n)$, let $$\mu(x;n,Y):=|\text{cap}(x;n,Y)\cap \mathbb{Z}^d|-1,$$ be the number of other lattice points in the cap around $x$. The mean of $\mu(x; n,Y)$, averaged over all lattice points $E(n)$ is: $$\left< \mu(\bullet ;n,Y)\right>:=\frac{1}{|E(n)|}\sum_{x\in E(n)}\big(|\text{cap}(x;n,Y)\cap\mathbb{Z}^d |-1\big).$$ We note that $|E(n)| \approx n^{\frac{d-2}{2}}$ when $d\geq 5$ and the volume of $\sqrt{n}S^{d-1}$ is $\approx n^{\frac{d-1}{2}}$. Heuristically, if we assume that the integral points are uniformly distributed, then we expect to have $\frac{Y^{d-1}}{n^{\frac{1}{2}}}$ integral points inside a cap of size $Y$. So, if $Y \ll n^{\frac{1}{2(d-1)}-o(1)}$ then we expect to have no integral points inside a cap of size $Y$. On the other hand, if $n^{\frac{1}{2(d-1)}+o(1)}\ll Y$ then we expect to have many points. In fact, we have the following corollary of our main theorem that makes this heuristic rigorous.\
\[mainn\] Let $d\geq 5$. Assume that $Y$, the size of the caps, satisfies
$$n^{\frac{1}{2(d-1)}+o(1)}\ll Y \ll n^{1/2}.$$ Then the probability that a cap of size $Y$ centered at integral point has more than $\log(n)$ points is greater than $1/2$ $$\mathbb{P}\big[ \mu(\bullet;n,Y)>\log(n)\big]> 1/2,$$ as a result $$\left<\mu(\bullet;n,Y) \right> \to \infty \text{ as } n\to \infty.$$ On the other hand, if $Y \ll n^{\frac{1}{2(d-1)}-o(1)}$, then $$\left<\mu(\bullet;n,Y) \right> \to 0 \text{ as } n\to \infty.$$\
Lester and Rudnick [@Rudnick]\[Remark 5.4\] remarked that the small scale equidistributin for individual eigenspace as stated in Corollary \[small\] is implied from $$\left<\mu(\bullet;n,Y) \right> \to 0 \text{ as } n\to \infty,$$ where $Y \ll n^{\frac{1}{2(d-1)}-o(1)}$. Bourgain [@Rudnick]\[Theorem 4.1\] show that the exponent $-\frac{1}{2(d-1)}$ is optimal by using the fact that $$\mathbb{P}\big[ \mu(\bullet;n,Y)>2 \big]> 1/2.$$ Therefore corollary \[mainn\] implies the corollary \[small\].
Main theorem {#Main theorem}
------------
We begin by introducing some notations. Let $Q(X)$ be an integral quadratic form where $X=(x_1,\dots,x_k)$ and define $$A:=\Big[\frac{\partial^2 Q}{\partial x_i \partial x_j}\Big],$$ then $$Q(X)=1/2X^T A X.$$ Let $D:=det (A)$ be the discriminant of $Q$. We write $r(Q,n)$ for the number of integral solutions of $$Q(X)=n.$$ We consider the Theta series associated to this quadratic form $$\Theta(z)=\sum_{n}r(Q,n)e(nz).$$ This is a modular form of weight $k/2$ and level $N$, where $N$ is the smallest integer such that $NA^{-1}$ is an even integral matrix . By the theory of modular forms we can write $\Theta(z)$ uniquely as a sum of standard Eisenstein series $E(z,Q)$ (the Eisenstein series associated to $Q$) and a cusp form $F(z,Q)$
$$\Theta(z)=E(z,Q)+F(z,Q).$$ From this decomposition $$r(Q,n)=\rho(n,Q)+\tau(n,Q).$$ where, $\rho(n,Q)$ and $\tau(n,Q)$ are the $n$-th Fourier coefficients of $E(z,Q)$ and $F(z,Q)$ respectively. We use the spectral theory of automorphic forms and bounds on the Fourier coefficients of modular forms to prove Theorem (\[main\]). In our main theorem, we give a uniform upper bound on the number of integral points on a quadric that is defined by any positive definite integral primitive quadratic form in $k\geq 4$ variables and with discriminant $D$ . Similar uniform results in a different context (explicit bounds for representability by a quadratic form) has been extensively studied by various authors. In particular, there is PhD work of Hanke [@Hanke] who uses theta series to get estimates which are uniform in the coefficients and also the work of Schulze-Pillot [@Schulze]. More recently, Browning and Deitmann [@Browning Proposition 1] established a result that recovers our theorem for in the generic situation where the coefficients of the quadratic form is of order $D^{1/k}$. This result is not enough to establish the small scale equidistribution on rational flat tori. We need a uniform result for all quadratic forms with discriminant $D$ that is stated in Theorem 1.4. We explain this in Remark \[remarks\] after stating the our main theorem.
\[main\] Let $n$ be any integer, and let $Q(X)$ be any primitive positive definite integral quadratic form of discriminant $D$ in $k\geq 4$ variables. If $D \ll n^{\frac{k-3}{2(k-2)}}$, then the number of integral solutions of $Q(X)=n$ is bounded from above by $$c_{\epsilon}\frac{n^{\frac{k-2}{2}}}{\sqrt{D}}\gcd(D,n)^{1/2}n^{\epsilon},$$ where $c_{\epsilon}$ is a constant which depends only on $\epsilon$ and not on $Q(X)$ or $n$.\
\[remarks\] Corollary (\[small\]) is a consequence of Theorem (\[main\]) with the discriminant bound $D\ll n^{1/4}$ and no conditions on the height of the quadratic forms. In fact the quadratic forms that we deal with are coming form the lattices given by the hyperplanes orthogonal to integral vectors of square norm $D$. So, the height of the quadratic forms might be as big as $D$. For quadratic forms in 5 or more variables in Theorem (\[main\]), we do not need to appeal to Blomer’s result stated in the appendix.
Outline of the paper
--------------------
We give a brief outline of this paper. In section (\[sec\]), we show that corollary (\[mainn\]) is a consequence of our main Theorem (\[main\]). Next, we give a proof of our main Theorem (\[main\]). In the proof of Theorem (\[main\]), we use an improved version of a lemma in Blomer’s papers or [@VB Lemma 3]. Professor Blomer provided us a proof for this improved version of his previous lemma.We include his proof in our appendix. We are responsible for any gap or typo in the appendix. In Lemma (\[1\]), we use the Siegel product formula (The main term of the Hardy-Littlewood formula) to give an upper bound on $\rho(n,Q)$. In Lemma (\[Cusp\]), we invoke the upper bound of Blomer [@Val Lemma 4.2] on $\|F(z,Q)\|^2$ and then we apply the Petersson trace formula to give an upper bound on $\tau(n,Q)$. The theorem is a consequence of lemma (\[1\]) and lemma (\[Cusp\]).
Acknowledgments
---------------
We are grateful to Professor Valentin Blomer for his comments and letter to us. In the letter, he proves lemma (\[l2\]) which improves his earlier lemma [@Val Lemma 4.2]. This lemma is crucial in our work. We would also like to thank Professor Zeev Rudnick for suggesting this project to us and his comments on the earlier version of this paper. Finally, I would like to thank Masoud Zargar for several comments and remarks on the earlier versions of this work.
Proof of corollary (\[mainn\]) {#sec}
===============================
We begin by proving the first part of Corollary (\[mainn\]) when the size of the caps is large, i.e.
$$n^{\frac{1}{2(d-1)}+o(1)}\ll Y \ll n^{1/2}.$$ This part is elementary; we use a covering argument in combination with a pigeonhole argument. We assume that $n^{\frac{1}{2(d-1)}+o(1)}\ll Y$. Call an integral point $p\in E(n) $ *bad* if $$\mu(p;n,Y)\leq \log(n).$$ We denote the number of bad points by $B$. Assume to the contrary that $B\geq \frac{1}{2}|E(n)|$. Hence, using $|E(n)|\approx n^{\frac{d-2}{2}}$, $B \gg n^{\frac{d-2}{2}}.$ Consider balls of radius $\frac{Y}{2}$ centered at each bad point. Each point of the sphere is covered by at most $\log(n)$ of these balls; otherwise, there are more than $\log(n)$ bad points with distance at most $Y$, contradicting the definition of a bad point. Therefore, we have the following inequality from a covering argument
$$B(Y/2)^{d-1} \ll n^{\frac{d-1}{2}} \log(n).$$ Hence, $$Y^{d-1}\ll n^{1/2}\log(n).$$ This is a contradiction to $Y\gg n^{\frac{1}{2(d-1)}+o(1)}.$ Therefore, $B<\frac{1}{2}|E(n)|$. Consequently, $$\mathbb{P}[\left<\mu(\bullet;n,Y) \right> > \log(n)]>1/2,$$ Therefore, $$\left<\mu(\bullet;n,Y) \right> \to \infty \text{ as }n\to\infty.$$ This concludes the proof of the first part.\
\[ll\]
The second part of the Corollary (\[mainn\]) is a consequence of the inequality $$A_{d}(n,t) \ll n^{\frac{d-3}{2}+\varepsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}},$$ where $n- n^{\frac{1}{(d-1)}-o(1)}<t<n$ and $5\leq d$.
For the second part of Corollary (\[mainn\]), when the size of caps is small, i.e. $$Y \ll n^{\frac{1}{2(d-1)}-o(1)} ,$$ we follow the technique in the argument for [@Rudnick lemma 11 ] to prove lemma (\[ll\]). Let $A_{d}(n,t)$ be the number of ordered pairs of distinct integral lattice points $(p,q)\in\mathbb{Z}^d\times\mathbb{Z}^d$ such that $|p|^2=|q|^2=n$ and $|p-q|^2=2(n-t)$. Note that a change of summation argument gives us the equality $$\left<\mu(\bullet;n,Y)\right>=\frac{1}{|E(n)|}\sum_{n-\frac{Y^2}{2}< t<n}A_d(n,t).$$ Since $Y \ll n^{\frac{1}{2(d-1)}-o(1)} $ and $n-\frac{Y^2}{2}< t<n$, so $n- n^{\frac{1}{(d-1)}-o(1)}<t<n$ and we can apply the inequality $$A_{d}(n,t) \ll n^{\frac{d-3}{2}+\varepsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}},$$ and we obtain $$\begin{aligned}
\left<\mu(\bullet;n,Y)\right>&\ll& \frac{1}{|E(n)|}\sum_{n-\frac{Y^2}{2}< t<n}n^{\frac{d-3}{2}+\varepsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}}\\
&\ll& \frac{n^{\frac{d-3}{2}+\varepsilon}}{|E(n)|}\sum_{n-\frac{Y^2}{2}< t<n}\gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}}\\
&\ll& \frac{n^{\frac{d-3}{2}+\varepsilon}}{|E(n)|}\sum_{1< s<\frac{Y^2}{2}}\gcd(n,s)^{1/2} s^{\frac{d-3}{2}}\ (\text{where } s:=n-t)\\
&\ll& \frac{n^{\frac{d-3}{2}+\varepsilon}Y^{d-1}n^{\varepsilon}}{|E(n)|}.\end{aligned}$$ Using $|E(n)|\approx n^{\frac{d-2}{2}}$ and $Y\ll n^{\frac{1}{2(d-1)}-o(1)}$, we get $$\left<\mu(\bullet;n,Y)\right>\ll n^{\varepsilon-o(1)}\to 0\text{ as }n\to\infty.$$
Lemma (\[lem\]) is devoted to showing that the following inequality in lemma (\[ll\]) is true: $$A_{d}(n,t) \ll n^{\frac{d-3}{2}+\varepsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}},$$ for $n- n^{\frac{1}{(d-1)}-o(1)}<t<n$ and $ 5 \leq d$.\
\
\[lem\] Theorem (\[main\]) implies that $$A_{d}(n,t) \ll n^{\frac{d-3}{2}+\varepsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}},$$ where $n- n^{\frac{1}{(d-1)}-o(1)}<t<n$ and $ 5 \leq d$.\
Recall that $A_d(n,t)$ is the number of ordered pairs of distinct integral lattice points $(p,q)\in\mathbb{Z}^d\times\mathbb{Z}^d$ such that $|p|^2=|q|^2=n$ and $|p-q|^2=2(n-t)$. Let $v:=p-q$. Note that $v$ is an integral vector of length $\sqrt{2(n-t)}$ and so, up to a constant, we have at most $$\label{choice}(n-t)^{\frac{d-2}{2}},$$ choices for $v$, because $d\geq 5$ . Note that $$\left<p+q,v \right>=0,$$ and $$|p+q|^2=2(n+t).$$ Let $L$ be the $(d-1)$-dimensional lattice that is given by the intersection of $\mathbb{Z}^d$ and the hyperplane orthogonal to $v$. Therefore, $p+q$ lies inside the lattice $L$ and the sphere of radius $\sqrt{2(n+t)}$. The fundamental domain of the lattice $L$ has volume $|v^{\prime}|$ where $v^{\prime}$ is the primitive integral vector in the direction of $v$. We take an integral basis for the lattice $L$ and denote it by $$\{e_1,\dots, e_{d-1} \}.$$ We define the symmetric matrix $A$ by $$A:=\big[ \left< e_i,e_j \right>_{1 \leq i,j \leq d-1} \big].$$ We define the quadratic form $Q(X)$ by $$Q(X):=X^{T}AX.$$ Let $k:=(d-1)$. This quadratic from is a primitive quadratic form in $k \geq 4$ variables with discriminant $D:=|v^{\prime}|^2$. We have the following upper bound on the discriminant of $Q$
$$D=|v^{\prime}|^2 \leq n^{\frac{1}{k}-o(1)}.$$ Since $ \frac{1}{k} \leq \frac{k-3}{2(k-2)}$ when $k \geq 4$ then $$D \ll n^{\frac{k-3}{2(k-2)}}.$$ We can apply theorem \[main\], and as a consequence we have the following bound on the number of lattice points of $L$ with norm $2(n+t)$ $$c_{\epsilon}\frac{(2(n+t))^{\frac{d-3}{2}}}{|v^{\prime}|}\gcd(2(n-t),2(n+t))^{1/2}(2(n+t))^{\epsilon}\ll\frac{n^{\frac{d-3}{2}}}{|v^{\prime}|}\gcd(n,t)^{1/2}n^{\epsilon}.$$ Recall that $|v|^2=|p-q|^2=2(n-t)$ and the number of integral points $v$ where $\frac{|v|}{|v^{\prime}|}=l$ is less than $\big(\frac{{n-t}}{l^2}\big)^{\frac{d-2}{2}+\epsilon}$ up to a constant. We give an upper bound on $A_d(n,t)$ by first choosing $v=p-q$ and then $p+q$ $$A_{d}(n,t) \ll \sum_{l^2|2(n-t)}(\frac{n-t}{l^2})^{\frac{d-2}{2}}\frac{n^{\frac{d-3}{2}}}{\frac{\sqrt{n-t}}{l}}\gcd(n,t)^{1/2}n^{\epsilon}.$$ Therefore, $$A_{d}(n,t) \ll n^{\frac{d-3}{2}+\epsilon} \gcd(n,t)^{1/2} (n-t)^{\frac{d-3}{2}}.$$
Proof of the Main theorem
=========================
We give a brief plan of the proof of Theorem (\[main\]) in what follows. Recall the notations that were introduced in section (\[Main theorem\]). In Lemma \[1\], we use the Siegel product formula (the main term of the Hardy-Littlewood formula) to give an upper bound on $\rho(n,Q)$. In Lemma \[Cusp\], we invoke the upper bound of Blomer on $\|F(z,Q)\|^2$ in Lemma (\[Valup\]) that is proved in our appendix. Finally we apply the Petersson trace formula together with a Cauchy inequality to give an upper bound on the $n$-th Fourier coefficient $\tau(n,Q)$ of $F(z,Q)$. Theorem (\[main\]) is a consequence of lemma (\[1\]) and lemma (\[Cusp\]).\
The following lemma proved by Blomer [@VB page 6] for $k=3$. We follow his strategy and give a proof for every $k \geq 4$.
\[1\]\[Val\] We have the following upper bound on the $n$-th Fourier coefficient of the Eisenstein series of the form $Q$ $$\rho(n,Q)\leq c_{\epsilon}\frac{n^{\frac{k-2}{2}}}{\sqrt{D}}\gcd(N,n)^{1/2}(nN)^{\epsilon},$$ where $N$ is the level of the quadratic from $Q$.\
$\rho(n,Q)$, the $n$-th Fourier coefficient of the Eisenstein series, coincides with the main term of the Hardy-Littlewood formula. The main term of the Hardy-Littlewood formula, is given by the product of local densities $$\rho(n,Q):=n^{\frac{k-2}{2}}\sigma_{\infty}\prod_{p}\sigma_{p} .$$ We have the following formula for the local densities: $$\sigma_p=\sum_{t=0}^{\infty} S(p^t),$$ where $$S(p^t):=\frac{1}{p^{tk}} \sum_{a}^* \sum_{b}e\Big(\frac{a(Q(b)-n)}{p^t}\Big).$$ $\sum_{a}^*$ means that $a$ varies mod $p^{t}$ where $\gcd(a,p)=1$, $b$ is a vector that varies mod $p^t\mathbb{Z^d}$ and S(1)=1.
We give an upper bound for each place. First, we start with $\infty$. The density at $\infty$ is given by $$\sigma_{\infty}=\lim_{\epsilon \to 0} \frac{\int_{1<Q(X)<1+\epsilon}1 dx_1 \dots dx_k}{\epsilon}.$$ We diagonalize $Q(X)$ in an orthonormal coordinates $Y:=(y_1, \dots, y_k)$ such that $$Q(Y):=n_1 y_1^2+ \dots +n_k y_k^2.$$ where $n_1, \dots, n_k$ are the eigenvalues of the symmetric matrix $A$. We change the variables to $(y_1, \dots, y_k)$ to get $$\sigma_{\infty}=\frac{1}{\sqrt{det{A}}}\vol({S^{k-1}}).$$ Next, we give an upper bound on the local densities $\sigma_p$ where $p\neq 2$. Since $p$ is an odd prime number, we can diagonalize our quadratic form $Q(X)$ over the local ring $\mathbb{Z}_{p}$. Without loss of generality we assume that $$Q(x_1,\dots,x_k)=\sum_{i=1}^{k}a_ip^{\alpha_i}x_i^2,$$ where $\gcd(a_i,p)=1$ and $a_i\in \mathbb{Z}_p$. We substitute the diagonal expansion of $Q(x_1,\dots,x_k)$ to compute $S(p^t)$ $$\begin{split}
S(p^t)&:=\frac{1}{p^{tk}}\sum_{a}^*\sum_{b\in{(\frac{\mathbb{Z}}{p^t\mathbb{Z}}})^k}e\Big(\frac{a(Q(b)-n)}{p^t}\Big)
\\
&=\frac{1}{p^{tk}}\sum_{a}^*\sum_{b\in{(\frac{\mathbb{Z}}{p^t\mathbb{Z}}})^k}e\Big(\frac{a(\sum_{i=1}^{k} a_{i}p^{\alpha_i}b_i^2-n)}{p^t}\Big)
\\
&=\frac{1}{p^{tk}}\sum_{a}^*e\Big(\frac{-an}{p^t}\Big)\prod_{i=1}^{k}\sum_{b \text{ mod } p^t}e\Big(\frac{aa_{i}p^{\alpha_i}b^2}{p^t}\Big).
\end{split}$$ We note that the last summation is a Gauss sum. Let $G(h,m):=\sum_{x \text{ mod } m} e(\frac{hx^2}{m})$ be the Gauss sum, and let $\epsilon_{m}=1$ if $m\equiv1 (\text{ mod } 4) $ and $\epsilon_{m}=i$ if $m\equiv 3 (\text{ mod } 4) $. Then if $\gcd(h,m)=1$ we have
$$G(h,m):=\begin{cases}
\epsilon_m \Big(\frac{h}{m} \Big) m^{1/2} & \text{ if } m \text{ is odd },\\
(1+\chi_{-4}(h))m^{1/2} &\text{ if } m=4^{\alpha},\\
(\chi_8(h)+i\chi_{-8}(h)) m^{1/2} &\text{ if } m=2.4^{\alpha}, \alpha\geq 1.
\end{cases}$$
where $\Big(\frac{h}{m} \Big)$ is the Jacobi symbol. We have $$S(p^t)=\frac{1}{p^{tk}}\sum_{a}^*e\Big(\frac{-an}{p^t}\Big)\prod_{i=1}^{k}p^{\min(\alpha_i,t)}G(aa_i,p^{t-\alpha_i}).$$ We define $G(\bullet,p^{t-\alpha_i}):=1$ when $t <\alpha_i $. We substitute the values of $G$ and obtain, $$\begin{split}
|S(p^t)|=\frac{\prod_{i=1}^k p^{\min(\frac{\alpha_i+t}{2},t)}}{p^{tk}}\big|\sum_{a}^*e\Big(\frac{-an}{p^t}\Big)\Big(\frac{a}{p} \Big)^{k^{\prime}}\big|,
\end{split}$$ where $k^{\prime}$ is the number of integers $i$ such that $1 \leq i \leq k$ and $t-\alpha_i$ is a positive odd integer. Assume that $n=p^{\beta}n^{\prime}$, where $\gcd(n^{\prime},p)=1$. If $k^{\prime}$ is an odd number then the inner sum is a Gaussian sum, and we obtain $$\big|\sum_{a \text{ mod } p^t}^*e\Big(\frac{-ap^{\beta}n^{\prime}}{p^t}\Big)\Big(\frac{a}{p} \Big)\big|=
\begin{cases}
p^{t-\frac{1}{2}} &\text{ if } \beta=t-1,\\
0 &\text{ otherwise }.
\end{cases}$$ Hence, if $k^{\prime}$ is odd we deduce that $$\label{odd}
|S(p^t)|= \begin{cases}\frac{\prod_{i=1}^k p^{\min(\frac{\alpha_i+t}{2},t)}}{p^{tk}}p^{t-\frac{1}{2}} &\text{ if } \beta=t-1 ,\\
0 &\text{ otherwise. }
\end{cases}
%
% \frac{\prod_{i=1}^d p^{\min(\frac{\alpha_i+k}{2},k)}}{p^{kd}} \big|\sum_{a}^*e\Big(\frac{-ap^{\beta}N^{\prime}}{p^k}\Big)\Big(\frac{a}{p} \Big)\big|
%\\
%\frac{\prod_{i=1}^d p^{\min(\frac{\alpha_i+k}{2},k)}}{p^{kd}}p^{\min(\frac{\beta+k}{2},k)}
%\end{split}$$ On the other hand, if $k^{\prime}$ is even then the inner sum is a Ramanujan sum $c_{p^t}(n)$, $$c_{p^t}(n)=\sum_{a}^*e\Big(\frac{-an}{p^t}\Big)=\begin{cases} 0 &\text{ if } \beta< t-1, \\
-p^{t-1} &\text{ if } \beta=t-1, \\
\phi(p^t) &\text{ if } \beta\geq t.
\end{cases}$$ Hence, if $k^{\prime}$ is even we deduce that
$$\label{even}
|S(p^t)|=\begin{cases} 0 &\text{ if } \beta< t-1, \\
-\frac{\prod_{i=1}^k p^{\min(\frac{\alpha_i+t}{2},t)}}{p^{tk}}p^{t-1} &\text{ if } \beta=t-1, \\
\phi(p^t)\frac{\prod_{i=1}^k p^{\min(\frac{\alpha_i+t}{2},t)}}{p^{tk}} &\text{ if } \beta\geq t.
\end{cases}$$
Without loss of generality suppose that $\alpha_1\leq \alpha_2 \leq \dots \leq \alpha_k$. Since $Q(X)$ is primitive we deduce that $\alpha_1=0$. If $k^{\prime}$ is odd (\[odd\]) and if $k^{\prime}$ is even (\[even\]), we deduce that $$\label{inq}
\begin{split}
|\sigma_p| &\leq \sum_{t=0}^{\beta+1} p^{t/2}\frac{\prod_{i=2}^k p^{\min(\frac{\alpha_i+t}{2},t)}}{p^{tk}}p^{\min(t,\frac{t+\beta}{2})}\\
&\leq \sum_{t=0}^{\beta+1}\frac{p^{t/2}p^{(k-1)\min(\frac{\alpha_k+t}{2},t)}p^{\min(t,\frac{t+\beta}{2})}}{p^{tk}}.
\end{split}$$ From the above inequality and the condition that $k\geq 4$, we give an upper bound on the product of local densities $\prod_{p\nmid 2nN}\sigma_p$ as follows. $$\begin{split}
\prod_{p\nmid 2nN}\sigma_p&\leq \prod_{p\nmid 2nN}(1+p^{-\frac{k-1}{2}})
\\
&\leq \sum_{n=1}^{\infty}\frac{1}{n^{3/2}}=O(1).
\end{split}$$ If $p$ is an odd prime number from the inequality (\[inq\]), we obtain $$\sigma_p\ll p^{\frac{\min(\alpha_k,\beta)}{2}}.$$ Hence, for odd prime numbers where $p|nN$ we have $$\label{rrr}
\prod_{p|nN}\sigma_p\leq (nN)^{\epsilon}\gcd(n,N)^{1/2}.$$ Finally, we assume that $p=2$. One can write any quadratic form $Q(X)$, after a change of variables over the 2-adic integers $\mathbb{Z}_2$, as a direct sum of scalar multiples of $q_0(x):=x^2$, $q_1(x_1,x_2):=x_1x_2$ and $q_2(x_1,x_2):=x_1^2+x_1x_2+x_2^2$. The corresponding Gauss sums can be evaluated for odd $h$:
$$\sum_{b_1, b_2 \text{ mod } 2^t} e\big(\frac{2^{\alpha_1}x_1x_2h}{2^t} \big)=2^{\min(t+\alpha_1,2t)},$$ and $$\sum_{b_1, b_2 \text{ mod } 2^t} e\big(\frac{2^{\alpha_1}(x^2+x_1x_2+x^2)h}{2^t} \big)=2^{\min(t+\alpha_1,2t)}.$$ We substitute the values of these Gauss sums and obtain $$\label{inq22}
\begin{split}
\sigma_2(Q,n)\leq \sum_{t=0}^{\beta+1} 2^{t/2}\frac{\prod_{i=2}^k 2^{\min(\frac{\alpha_i+t}{2},t)}}{2^{tk}}2^{\min(t,\frac{t+\beta}{2})}\\
\leq \sum_{t=0}^{\beta+1} \frac{2^{t/2}2^{(k-1)\min(\frac{\alpha_k+t}{2},t)}2^{\min(t,\frac{t+\beta}{2})}}{2^{tk}}.
\end{split}$$ From the inequality (\[inq22\]), we obtain
$$\sigma_2(Q,n)\ll 2^{\frac{\min(\alpha_k,\beta)}{2}}.$$ This inequality together with inequality (\[rrr\]) for odd primes $p$ implies
$$\prod_{p|\gcd(D,N)}\sigma_p \leq \gcd(N,D)^{1/2}.$$ Therefore, we conclude the lemma.\
In Lemma \[Cusp\], we invoke a result of Blomer, proved in Appendix 1, and then we apply the Petersson trace formula together with a Cauchy inequality to give an upper bound on $\tau(n,Q)$. Theorem (\[main\]) is a consequence of lemma (\[1\]) and lemma (\[Cusp\]).
\[Cusp\] Let $Q(X)$ be a quadratic form in $k\geq 4$ variables and discriminant $D<n^{\frac{k-3}{2(k-2)}}$. Then we have the following upper bound on the $n$-th Fourier coefficient of the cusp form part of the theta series associated to $Q$ $$|\tau(n,Q)|\ll_{d,\epsilon} D^{\frac{(k-3)}{2}}n^{(k-1)/4}\gcd(n,D)^{1/4}n^{\epsilon}.$$ In particular $$|\tau(n,Q)|\ll c_{\epsilon}\frac{n^{\frac{k-2}{2}}}{\sqrt{D}}\gcd(D,n)^{1/4}n^{\epsilon}.$$
We take an orthonormal basis $\{f_i \}$ for the space of cusp forms of weight $k/2$ and level $N$. We write $F(z,Q)$ as a linear combination of them $$F(Q,z)=\sum_{f_i}C(Q,f_i)f_i(z).$$ We obtain $$\tau(n,Q)=\sum_{f_i}C(Q,f)\rho_{f_i}(n).$$ We apply Cauchy inequality to obtain $$\label{kkk}
|\tau(n,Q)|^2\leq\big(\sum_{f_i}|C(Q,f)|^2\big) \big(\sum_{f_i} |\rho_{f_i}(n)|^2 \big).$$ Since we take an orthonormal basis for the space of cusp forms we deduce that $$||F||^2=\sum_{f_i}|C(Q,f_i)|^2.$$\
We invoke a result of Blomer. The proof is included in Appendix 1.
\[l2\]\[Valup\] We have $$\|F\|^2 \ll N^{\epsilon}(N^{k-2} +N^{\frac{k-3}{2}}D^{1/2}+D^{1-1/k} ).$$ for $k\geq 4$. As a corollary, $$\label{kk}
\|F\|^2 \ll N D^{k-3+\epsilon}.$$
Next, we give an upper bound on the sum over the square norm of the $n$-th Fourier coefficients of $\{f_i\}$
$$\big(\sum_{f_i} |\rho_{f_i}(n)|^2 \big).$$ We apply the Petersson trace formula for the modular forms of weight $k/2$ . See [@Hen [Lemma 1]{} ] $$\frac{\Gamma(k/2-1)}{(4\pi n)^{k/2-1}}\sum_{f_i} |\rho_{f_i}(n)|^2= 1+2\pi i^{-k/2}\sum_{c\equiv0 \text{ mod } N} c^{-1}J_{k/2-1}\big(\frac{4\pi n}{c} \big)K(n,n;c).$$ In order to give an upper bound on $K(n,n;c)$, we apply the bound on Salie’s sum for odd dimensions $k$ and Weil’s bound on the Kloosterman’s sum for even dimensions $k$. We invoke the following formula [@Iwan Corollary 14.24 ] $$\frac{\Gamma(l-1)}{(4\pi n)^{l-1}}\sum_{f_i} |\rho_{f_i}(n)|^2 = 1+ O\Big(\tau_3(n)\gcd(n,N)^{1/2}n^{1/2}\frac{\tau(N)}{N\sqrt{l}} \log\big(1+\frac{n^{1/2}}{\sqrt{Nl}} \big) ),$$where $l=k/2$. Since the discriminant $D<n^{\frac{k-3}{2(k-2)}}$ and $N|D$ then $$\label{k}
\sum_{f_i} |\rho_{f_i}(n)|^2 \ll_{\epsilon}\frac{n^{(k-1)/2}}{N}n^{\epsilon}\gcd(n,D)^{1/2}.$$ From the inequalities (\[kkk\]), (\[kk\]) and (\[k\]) we deduce that $$|\tau(n,Q)|^2\ll_{d,\epsilon} D^{k-3}n^{(k-1)/2}\gcd(n,D)^{1/2}n^{\epsilon}.$$ This concludes the first part of the lemma. For the second part, $D \ll n^{\frac{k-3}{2(k-2)}}$ implies that $$D^{\frac{k-3}{2}}n^{\frac{k-1}{4}} \leq \frac{n^{\frac{k-2}{2}}}{\sqrt{D}}.$$ Hence, $$|\tau(n,Q)|\ll c_{\epsilon}\frac{n^{\frac{k-2}{2}}}{\sqrt{D}}\gcd(D,n)^{1/4}n^{\epsilon}.$$ The main theorem (\[main\]) is a consequence of lemma (\[l2\]) and lemma (\[1\]).
Appendix
========
In this appendix, we include Valentin Blomer’s paper to us. We are responsible for any gap or typo in this section.\
Let $$Q(X)=\frac{1}{2} X^{T}AX.$$ be a primitive positive definite integral $k$-dimensional quadratic form $(k\geq 3)$ of discriminant $D$ and level $N$. For $n \in \mathbb{N}$ let $r(n,Q)$ denote the number of representations of $n$ by $Q$ and let $\rho(n,Q)$ be the main term given by a formal application of the circle method. Denote by $\mu_1, \dots \mu_k$ the successive minima (see [@Cassels Chapter 12] ) of $Q(X)$. We can write
$$Q(X):=h_1(x_1+c_{12}x_2+\dots+c_{1n}x_k)^2+\dots+h_kx_k^2,$$ where $c_j,h_j\in \mathbb{Q}$ such that $$h_j\asymp \mu_j.$$ Let $\Theta(Q,z)$ denote the corresponding theta-series and $F(z,Q)$ the orthogonal projection onto the space of cusp forms. The following lemma uses only reduction theory and the Lipschitz principle.
We have $$r(Q,n)\ll \Big(1+\frac{n^{1/2}}{\mu_3}+\frac{n}{(\mu_3\mu_4)^{1/2}}+\dots+\frac{n^{\frac{k-2}{2}}}{(\mu_3\dots \mu_k)^{1/2}} \Big)n^{\epsilon},$$ and $$\sum_{n\leq x}r(Q,n)\ll \max_{j\leq k} \frac{x^{\frac{j}{2}}}{(\mu_1\dots \mu_j)^{1/2}}.$$
We choose $x_k,x_{k-1},\dots, x_3$ in $$\ll \big(1+\frac{n}{\mu_k}\big)^{1/2}\dots \big( 1+\frac{n}{\mu_3} \big)^{1/2},$$ ways. Then we are left with an inhomogeneous binary problem that has $O(n^{\epsilon})$ solutions (see [@Michel Lemma 3a]), uniformly for all choices of $x_k,\dots,x_3$. To prove the second part, we choose $x_k,\dots,x_1$ in $$\ll (1+\frac{x}{\mu_k})^{1/2}\dots (1+\frac{x}{\mu_1})^{1/2},$$ ways.
\[cor1\] We have $$\sum_{n\leq x} r(Q,n)^2 \ll x^{\epsilon} \Big(x^{k-2}+\frac{x^{k-\frac{3}{2}}}{D^{1/2}}+ \frac{x^{k-1}}{D^{1-1/k}} \Big).$$\
Finally, we give a proof of Lemma (\[Valup\]). It suffices to show that $$\|F(z,Q)\|^2 \ll N^{\epsilon} (N^{k-2}+N^{\frac{k-3}{2}}D^{1/2}+D^{1-1/k}).$$ Note that if $k\geq 4$, then the right hand side is $\ll ND^{k-3+\epsilon}$.
We follow the argument of [@Val Lemma 4.2] or [@VB Lemma 3]. This gives $$\|F(z,Q) \|^2 \ll N^{\epsilon}\Big(\frac{1}{N}+\int_{1/2}^{\infty} \sum_{n=1}^{\infty}\big( r(\hat{Q},n)^2+r_{Eis}(\hat{Q},n)^2 \big)e^{\frac{-4\pi ny}{N}} y^{k/2-2} dy \Big),$$ where $\hat{Q}(X)=\frac{1}{2}X^T(NA^{-1})X$ has determinant $N^kD^{-1}$ and level $N$. Inserting Lemma \[1\] and Corollary \[cor1\], we obtain after a short computation the desired bound.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we construct two ternary linear codes $C(SO(3,q))$ and $C(O(3,q))$, respectively associated with the orthogonal groups $SO(3,q)$ and $O(3,q)$. Here $q$ is a power of three. Then we obtain two recursive formulas for the power moments of Kloosterman sums with $\lq\lq$trace nonzero square arguments" in terms of the frequencies of weights in the codes. This is done via Pless power moment identity and by utilizing the explicit expressions of Gauss sums for the orthogonal groups.\
Index terms - power moment, Kloosterman sum, trace nonzero square argument, orthogonal group, Pless power moment identity, weight distribution, Gauss sum.\
MSC 2000: 11T23, 20G40, 94B05.
address: 'Department of Mathematics, Sogang University, Seoul 121-742, Korea'
author:
- dae san kim
title: |
$\begin{array}{c}
\text{Ternary Codes Associated with $O(3,3^r)$ and }\\
\text{Power Moments of Kloosterman Sums with}\\
\text{Trace Nonzero Square Arguments}
\end{array}$
---
[^1]
Introduction
============
Let $\psi$ be a nontrivial additive character of the finite field $\mathbb{F}_q$ with $q=p^r$ elements ($p$ a prime). Then the Kloosterman sum $K(\psi ; a)$([@LN1]) is defined by
$$\begin{aligned}
K(\psi ; a)=\sum_{\alpha \in \mathbb{F}_q^*} \psi(\alpha + a
\alpha^{-1})~(a \in \mathbb{F}_q^*).\end{aligned}$$
The Kloosterman sum was introduced in 1926([@K1]) to give an estimate for the Fourier coefficients of modular forms.
For each nonnegative integer $h$, by $MK(\psi)^h$ we will denote the $h$-th moment of the Kloosterman sum $K(\psi;a)$. Namely, it is given by
$$\begin{aligned}
MK(\psi)^h=\sum_{\alpha \in \mathbb{F}_q^*}K(\psi;a)^h.\end{aligned}$$
If $\psi=\lambda$ is the canonical additive character of $\mathbb{F}_q$, then $MK(\lambda)^h$ will be simply denoted by $MK^h$.
Explicit computations on power moments of Kloosterman sums were begun with the paper [[@S1]]{} of Sali$\acute{e}$ in 1931, where he showed, for any odd prime $q$,
$$\begin{aligned}
MK^h=q^2 M_{h-1}-(q-1)^{h-1}+2(-1)^{h-1}~(h\geq1).\end{aligned}$$
Here $M_0=0$, and, for $h \in \mathbb{Z}_{>0}$, $$\begin{aligned}
M_h=|\{(\alpha_1,\cdots,\alpha_h)\in
(\mathbb{F}_q^*)^h|\sum_{j=1}^{h}\alpha_j=1=\sum_{j=1}^h
\alpha_j^{-1}\}|.\end{aligned}$$ For $q=p$ odd prime, Sali$\acute{e}$ obtained $MK^1$, $MK^2$, $MK^3$, $MK^4$ in [@S1] by determining $M_1$, $M_2$, $M_3$. On the other hand, $MK^5$ can be expressed in terms of the $p$-th eigenvalue for a weight $3$ newform on $\Gamma_0(15)$ (cf. [@L1], [@PV1]). $MK^6$ can be expressed in terms of the $p$-th eigenvalue for a weight $4$ newform on $\Gamma_0(6)$ (cf.[@HS1]). Also, based on numerical evidence, in [@E1] Evans was led to propose a conjecture which expresses $MK^7$ in terms of Hecke eigenvalues for a weight $3$ newform on $\Gamma_0(525)$ with quartic nebentypus of conductor 105.
From now on, let us assume that $q=3^r$. Recently, Moisio was able to find explicit expressions of $MK^h$, for $h\leq10$ (cf. [@M1]). This was done, via Pless power moment identity, by connecting moments of Kloosterman sums and the frequencies of weights in the ternary Melas code of length $q-1$, which were known by the work of Geer, Schoof and Vlugt in [@GS1].
In order to describe our results, we introduce three incomplete power moments of Kloosterman sums. For every nonnegative integer $h$, and $\psi$ as before, we define
$$\label{a}
T_0 SK(\psi)^h=\sum_{a\in\mathbb{F}_{q}^{*},~tra=0}K(\psi;a^2)^h,~
T_{12}
SK(\psi)^h=\sum_{a\in\mathbb{F}_{q}^{*},~tra\neq0}K(\psi;a^2)^h,$$
which will be respectively called the $h$-th moment of Kloosterman sums with $\lq\lq$trace zero square arguments“ and those with $\lq\lq$trace nonzero square arguments.” Then, clearly we have $$\label{b}
2SK(\psi)^h=T_0SK(\psi)^h+T_{12}SK(\psi)^h,$$ where $$\label{c}
SK(\psi)^h=\sum_{a\in\mathbb{F}_{q}^{*},~a~square}K(\psi;a)^h,$$ which is called the $h$-th moment of Kloosterman sums with $\lq\lq$square arguments." If $\psi=\lambda$ is the canonical additive character of $\mathbb{F}_{q}^{}$, then $SK(\lambda)^h$, $T_{0}SK(\lambda)^h$, and $T_{12}SK(\lambda)^h$ will be respectively denoted by $SK^h$, $T_{0}SK^h$ and $T_{12}SK^h$, for brevity.
We derived recursive formulas generating the odd power moments of Kloosterman sums with trace one arguments in [@D2] and [@D5]. To do that we constructed binary linear codes associated with $O(3,2^r)$ and with double cosets with respect to certain maximal parabolic subgroup of $O(2n+1,2^r)$.
In this paper, we will show the main Theorem \[A\] giving recursive formulas for the power moments of Kloosterman sums with $\lq\lq$trace nonzero square arguments.“ To do that, we construct ternary linear codes $C(SO(3,q))$ and $C(O(3,q))$, respectively associated with the orthogonal groups $SO(3,q)$ and $O(3,q)$, and express those power moments in terms of the frequencies of weights in the codes. Then, thanks to our previous results on the explicit expressions of $\lq\lq$Gauss sums” for the orthogonal group $O(2n+1,q)$ [@D1], we can express the weight of each codeword in the duals of the codes in terms of Kloosterman sums. Then our formulas will follow immediately from the Pless power moment identity.
Henceforth, we agree that, for nonnegative integers $a,b,c$, $$\label{d}
{\binom{c}{a,b}}={\frac{c!}{a!~b!~(c-a-b)!}},~if~a+b\leq c,$$ and $$\label{e}
{\binom{c}{a,b}}=0,~if~a+b>c.$$
\[A\] Let $q=3^r$. Then we have the following.
$(1)$ For $h=1,2,3,\cdots,$ $$\begin{aligned}
\label{f}
\begin{split}
&((-1)^{h+1}+2^{-h})T_{12}SK^h\\
&=-\sum_{j=1}^{h-1}((-1)^{j+1}+2^{-j}){\binom{h}{j}}(q^2-1)^{h-j}T_{12}SK^j\\
&\quad+q^{1-h}\sum_{j=0}^{min\{N_1,h\}}(-1)^j(C_{1,j}-\hat{C}_j)\sum_{t=j}^{h}t!
S(h,t)3^{h-t}2^{t-h-j}{\binom{N_1-j}{N_1-t}},
\end{split}\end{aligned}$$ where $N_1=|SO(3,q)|=q(q^2-1)$, and $\{C_{1,j}\}_{j=0}^{N_1}$ and $\{\hat{C}_{j}\}_{j=0}^{N_1}$ are respectively the weight distributions of $C(SO(3,q))$ and $C(Sp(2,q))$ given by: for $j=0,\cdots,N_1$, $$\begin{aligned}
\label{g}
\begin{split}
C_{1,j}=&\sum{\binom{q^2}{\nu_0,\mu_0}}{\binom{q^2}{\nu_2,\mu_2}}\\
& \qquad \times\prod_{\beta^2-2\beta\neq0~square}
{\binom{q^2+q}{\nu_\beta,\mu_\beta}}
\prod_{\beta^2-2\beta~nonsquare}
{\binom{q^2-q}{\nu_\beta,\mu_\beta}},
\end{split}\end{aligned}$$ $$\begin{aligned}
\label{h}
\begin{split}
\hat{C}_{j}=&\sum{\binom{q^2}{\nu_1,\mu_1}}{\binom{q^2}{\nu_{-1},\mu_{-1}}}\\
& \qquad \times\prod_{\beta^2-1\neq0~square}
{\binom{q^2+q}{\nu_\beta,\mu_\beta}} \prod_{\beta^2-1~nonsquare}
{\binom{q^2-q}{\nu_\beta,\mu_\beta}}.
\end{split}\end{aligned}$$ Here the first sum in (\[f\]) is 0 if $h=1$ and the unspecified sums in (\[g\]) and (\[h\]) run over all the sets of nonnegative integers $\{\nu_\beta\}_{\beta\in\mathbb{F}_q}$ and $\{\mu_\beta\}_{\beta\in\mathbb{F}_q}$ satisfying $$\begin{aligned}
\sum_{\beta\in\mathbb{F}_q}\nu_\beta+\sum_{\beta\in\mathbb{F}_q}\mu_\beta=j,~
and~ \sum_{\beta\in\mathbb{F}_q}\nu_\beta
\beta=\sum_{\beta\in\mathbb{F}_q}\mu_\beta \beta.\end{aligned}$$ In addition, $S(h,t)$ is the Stirling number of the second kind defined by $$\begin{aligned}
\label{i}
S(h,t)=\frac{1}{t!}\sum_{j=0}^{t}(-1)^{t-j}{\binom{t}{j}}j^h.\end{aligned}$$ $(2)$ For $h=1,2,3,\cdots,$
$$\begin{aligned}
\label{j}
\begin{split}
&((-1)^{h+1}+2^{-h})T_{12}SK^h\\
&=-\sum_{j=1}^{h-1}((-1)^{j+1}+s^{-j}){\binom{h}{j}}(q^2-1)^{h-j}T_{12}SK^j\\
&\quad +q^{1-h}\sum_{j=0}^{min\{N_2,h\}}(-1)^j
C_{2,j}\sum_{t=j}^{h}t!S(h,t)3^{h-t}2^{t-2h-j}
{\binom{N_2-j}{N_2-t}}\\
&\quad -q^{1-h}\sum_{j=0}^{min\{N_1,h\}}(-1)^j
\hat{C}_{j}\sum_{t=j}^{h}t!S(h,t)3^{h-t}2^{t-h-j}
{\binom{N_1-j}{N_1-t}},
\end{split}\end{aligned}$$
where $N_2=|O(3,q)|=2q(q^2-1)$, and $\{C_{2,j}\}_{j=0}^{N_1}$ is the weight distribution of $C(O(3,q))$ given by: for $j=0,\cdots,N_2$, $$\begin{aligned}
\label{k}
\begin{split}
C_{2,j}=&\sum\prod_{\beta\in\mathbb{F}_q}{\binom{n_2(\beta)}{\nu_\beta,\mu_\beta}}
~(j=0,\cdots,N_2),\\
&\quad with~
n_2(\beta)=2q^2-2q+q\delta(1,q;\beta-1)+q\delta(1,q;\beta+1).
\end{split}\end{aligned}$$ Here the first sum in (\[j\]) is 0 if $h=1$, the unspecified sum in (\[k\]) runs over all the sets of nonnegative integers $\{\nu_\beta\}_{\beta\in\mathbb{F}_q}$ and $\{\mu_\beta\}_{\beta\in\mathbb{F}_q}$ satisfying $$\begin{aligned}
\sum_{\beta\in\mathbb{F}_q}\nu_\beta+\sum_{\beta\in\mathbb{F}_q}\mu_\beta=j,\quad
~and~ \sum_{\beta\in\mathbb{F}_q}\nu_\beta
\beta=\sum_{\beta\in\mathbb{F}_q}\mu_\beta \beta,\end{aligned}$$ $S(h,t)$ indicates the Stirling number of the second as in (\[i\]), $\hat{C}_j$’s are as in (\[h\]), and $$\label{l}
\begin{split}
\delta(1,q;\beta)&=|\{x\in\mathbb{F}_{q}^{}|x^2-\beta x+1=0\}|\\
&=
\begin{cases}
2,& \text {if $\beta^2-1\neq0$ is a square,}\\
1,& \text {if $\beta^2-1=0$,}\\
0,& \text {if $\beta^2-1$ is a nonsquare.}
\end{cases}
\end{split}$$
$O^{}(2n+1,q)$
==============
For more details about the results of this section, one is referred to the paper [@D1]. Throughout this paper, the following notations will be used: $$\begin{aligned}
\begin{split}
q&=3^r~(r\in\mathbb{Z}_{>0}),\\
\mathbb{F}_q&=~the~finite~field~ with~ q~ elements,\\
TrA&=~the~ trace~ of~ A~ for~ a~ square~ matrix~ A,\\
^t B&=~the~transpose~ of~ B~for~any~matrix~B.
\end{split}\end{aligned}$$
The orthogonal group $O(2n+1,q)$ is defined as: $$\begin{aligned}
O(2n+1,q)=\{w\in GL(2n+1,q)|^twJw=J\},\end{aligned}$$ where $$\begin{aligned}
J=\begin{bmatrix}
0 & 1_{n} & 0 \\
1_{n} & 0 & 0 \\
0 & 0 & 1 \\
\end{bmatrix}.\end{aligned}$$ It consists of the matrices
$$\begin{aligned}
\begin{bmatrix}
A & B & e \\
C & D & f \\
g & h & i \\
\end{bmatrix}
(A,B,C,D~n\times n,e,f~n\times1,g,h~1\times n,i~1\times1)\end{aligned}$$
in $GL(2n+1,q)$ satisfying the relations:
$$\begin{aligned}
\begin{split}
&{^tA}C+{^tC}A+{^tg}g=0,~{^tB}D+{^tD}B+{^th}h=0,\\
&{^tA}D+{^tC}B+{^tg}h=1_{n},~{^te}f+{^tf}e+i^2=1,\\
&{^tA}f+{^tC}e+{^tg}i=0,~{^tB}f+{^tD}e+{^th}i=0.\\
\end{split}\end{aligned}$$
Let $P(2n+1,q)$ be the maximal parabolic subgroup of $O(2n+1,q)$ given by $$\begin{aligned}
\begin{split}
P&=P(2n+1,q)\\
&=\left\{\begin{bmatrix}
A & 0 & 0 \\
0 & {^tA}^{-1} & 0 \\
0 & 0 & i \\
\end{bmatrix}
\begin{bmatrix}
1_{n} & B & -{^th} \\
0 & 1_{n} & 0 \\
0 & h & 1 \\
\end{bmatrix}
\Bigg|\begin{array}{c}
A\in GL(n,q),~i=\pm1 \\
B+{^tB}+{^th}h=0
\end{array}
\right\},
\end{split}\end{aligned}$$ and let $Q=Q(2n+1,q)$ be the subgroup of $P(2n+1,q)$ of index 2 defined by $$\begin{aligned}
\begin{split}
Q&=Q(2n+1,q)\\
&=\left\{
\begin{bmatrix}
A & 0 & 0 \\
0 & {^tA}^{-1} & 0 \\
0 & 0 & 1 \\
\end{bmatrix}
\begin{bmatrix}
1_{n} & B & -{^th} \\
0 & 1_{n} & 0 \\
0 & h & 1 \\
\end{bmatrix}
\Bigg|\begin{array}{c}
A\in GL(n,q) \\
B+{^tB}+{^th}h=0
\end{array}
\right\}.
\end{split}\end{aligned}$$ Then we see that $$\begin{aligned}
P(2n+1,q)=Q(2n+1,q)\amalg\rho Q(2n+1,q),\end{aligned}$$ with $$\begin{aligned}
\rho=\begin{bmatrix}
1_{n} & 0 & 0 \\
0 & 1_{n} & 0 \\
0 & 0 & -1 \\
\end{bmatrix}.\end{aligned}$$
Let $\sigma_r$ denote the following matrix in $O(2n+1,q)$ $$\begin{aligned}
\sigma_r=\begin{bmatrix}
0 & 0 & 1_r & 0 & 0 \\
0 & 1_{n-r} & 0 & 0 & 0 \\
1_r & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1_{n-r} & 0 \\
0 & 0 & 0 & 0 & 1 \\
\end{bmatrix}
~(0\leq r\leq n).\end{aligned}$$ Then the Bruhat decomposition of $O(2n+1,q)$ with respect to $P=P(2n+1,q)$ is given by $$\begin{aligned}
O(2n+1,q)=\coprod_{r=0}^{n}P\sigma_r P=\coprod_{r=0}^{n}P\sigma_r Q,\end{aligned}$$ which can further be modified as $$\label{m}
\begin{split}
O(2n+1,q)&=\coprod_{r=0}^{n}P\sigma_r (B_r\setminus Q)\\
&=\coprod_{r=0}^{n}Q\sigma_r (B_r\setminus
Q)\amalg\coprod_{r=0}^{n}\rho Q\sigma_r(B_r\setminus Q),
\end{split}$$ with $$\begin{aligned}
B_r=B_r (q)=\{w\in Q(2n+1,q)|\sigma_r w \sigma_r^{-1}\in
P(2n+1,q)\}.\end{aligned}$$
The special orthogonal group $SO(2n+1,q)$ is defined as $$\begin{aligned}
SO(2n+1,q)=\{w\in O(2n+1,q)|\text{det}w=1\}.\end{aligned}$$ Then we see from (\[m\]) that $$\label{n}
SO(2n+1,q)=\coprod_{0\leq r\leq n,~r~even}Q\sigma_r(B_r\setminus
Q)\amalg\coprod_{0\leq r\leq n,~r~odd} \rho Q\sigma_r(B_r\setminus
Q).$$
The sympletic group $Sp(2n,q)$ is defined as: $$\begin{aligned}
Sp(2n,q)=\{w\in GL(2n,q)|{^tw}\hat{J}{w}=\hat{J}\},\end{aligned}$$ with $$\begin{aligned}
\hat{J}=\begin{bmatrix}
0 & 1_n \\
1_n & 0 \\
\end{bmatrix}.\end{aligned}$$
As is well-known or mentioned in [@D3] and [@D1], $$\label{o}
|O(2n+1,q)|=2q^{n^2}\prod_{j=1}^{n}(q^{2j}-1),\qquad\qquad\qquad$$ $$\label{p}
|SO(2n+1,q)|=|Sp(2n,q)|=q^{n^2}\prod_{j=1}^{n}(q^{2j}-1).$$
For integers $n$, $r$ with $0\leq r\leq n$, the $q$-binomial coefficients are defined as: $$\begin{aligned}
\begin{bmatrix}
n \\
r \\
\end{bmatrix}
_q=\prod_{j=0}^{r-1}(q^{n-j}-1)/(q^{r-j}-1).\end{aligned}$$ It is shown in [@D1] that $$\label{r}
|B_r(q)\setminus Q(2n+1)|=q^{\binom{r+1}{2}}
\begin{bmatrix}
n \\
r \\
\end{bmatrix}
_q.$$
Gauss sums for $O^{}(2n+1,q)$
=============================
The following notations will be employed throughout this paper.
$$\begin{aligned}
\begin{split}
tr(x)&=x+x^{3}+\cdots+x^{{3}^{r-1}} ~the~ trace~function~ \mathbb{F}_{q}~\rightarrow ~\mathbb{F}_{3},\\
\lambda_0(x)&=e^{2\pi ix/3} ~the~ canonical~ additive~ character~
of~\mathbb{F}_{3}^{},\\
\lambda(x)&=e^{2\pi itr(x)/3} ~the~ canonical~ additive~ character~
of~ \mathbb{F}_{q}^{}.
\end{split}\end{aligned}$$
Then any nontrivial additive character $\psi$ of $\mathbb{F}_q$ is given by $\psi(x)=\lambda(ax)$, for a unique $a\in\mathbb{F}_{q}^{*}$. Also, since $\lambda(a)$ for any $a\in\mathbb{F}_q$ is a 3th root of 1, we have
$$\label{s}
\lambda(-a)=\lambda(2a)=\lambda(a)^2=\lambda(a)^{-1}=\overline{\lambda(a)}.$$
For any nontrivial additive character $\psi$ of $\mathbb{F}_q$ and $a\in\mathbb{F}_{q}^{*}$, the Kloosterman sum $K_{GL(t,q)}(\psi;a)$ for $GL(t,q)$ is defined as
$$\begin{aligned}
K_{GL(t,q)}(\psi;a)=\sum_{w\in GL(t,q)}\psi(Trw+aTrw^{-1}).\end{aligned}$$
Observe that, for $t=1$, $K_{GL(1,q)}(\psi;a)$ denotes the Kloosterman sum $K(\psi;a)$.
In [@D3], it is shown that $K_{GL(t,q)}(\psi;a)$ satisfies the following recursive relation: for integers $t\geq2$, $a\in\mathbb{F}_q^*$, $$\begin{aligned}
K_{GL(t,q)}(\psi;a)=q^{t-1}K_{GL(t-1,q)}(\psi;a)K(\psi;a)+q^{2t-2}(q^{t-1}-1)
K_{GL(t-2,q)}(\psi;a),\end{aligned}$$ where we understand that $K_{GL(0,q)}(\psi;a)=1$.\
\[B\]$($[@D1]$)$ Let $\psi$ be a nontrivial additive character of $\mathbb{F}_q$. For each positive integer $r$, let $\Omega_r$ be the set of all $r\times
r$ nonsingular symmetric matrices over $\mathbb{F}_q$. Then we have $$\label{t}
a_r(\psi)=\sum_{B\in\Omega_r}\sum_{h\in\mathbb{F}_q^{r\times
1}}\psi({^th}Bh)=
\begin{cases}
q^{r(r+2)/4}\prod_{j=1}^{r/2}(q^{2j-1}-1), & \text{for $r$ even},\\
0,& \text{for $r$ odd.}
\end{cases}$$
From [@D3] and [@D1], the Gauss sums for $SO(2n+1,q)$ and $O(2n+1,q)$ are respectively equal to $\psi(1)$ times that for $Sp(2n,q)$ and $\psi(1)+\psi(-1)$ times that for $Sp(2n,q)$. Indeed, using the decomposition in (\[n\]), for any nontrivial additive character $\psi$ of $\mathbb{F}_q$, it is shown that
$$\begin{aligned}
\begin{split}
&\sum_{w\in SO(2n+1,q)}\psi(Trw)\\
&=\sum_{\substack{0\leq r \leq n\\r~even}}|B_r\setminus Q|\sum_{w\in
Q}\psi(Trw\sigma_r)
+\sum_{\substack{0\leq r \leq n\\r~odd}}|B_r\setminus Q|\sum_{w\in Q}\psi(Tr\rho w\sigma_r)\\
&=q^{\binom{n+1}{2}}\{\psi(1)\sum_{\substack{0\leq r \leq n\\r~even}}|B_r\setminus Q|
q^{r(n-r-1)}a_r(\psi)K_{GL(n-r,q)}(\psi;1)\\
&\qquad \qquad \qquad \qquad +\psi(-1)\sum_{\substack{0\leq r \leq
n\\r~odd}}|B_r\setminus Q|
q^{r(n-r-1)}a_r(\psi)K_{GL(n-r,q)}(\psi;1)\}\\
&=\psi(1)q^{\binom{n+1}{2}}\sum_{\substack{0\leq r \leq
n\\r~even}}q^{rn-{\frac{1}{4}}r^2}\begin{bmatrix}
n \\
r \\
\end{bmatrix}
_q\prod_{j=1}^{r/2}(q^{2j-1}-1)K_{GL(n-r,q)}(\psi;1)
~(cf.(\ref{r}),~(\ref{t}))\\
&(=\psi(1)\sum_{w\in Sp(2n,q)}\psi(Trw))~(cf. ~[4]).
\end{split}\end{aligned}$$
Similarly, from the decomposition in (\[m\]) it is shown in [@D1] that
$$\begin{aligned}
\begin{split}
&\sum_{w\in O(2n+1,q)}\psi(Trw)\\
&=(\psi(1)+\psi(-1))q^{\binom{n+1}{2}}\sum_{\substack{0\leq r \leq
n\\r~even}}q^{rn-{\frac{1}{4}}r^2}\begin{bmatrix}
n \\
r \\
\end{bmatrix}
_q\prod_{j=1}^{r/2}(q^{2j-1}-1)
K_{GL(n-r,q)}(\psi;1)\\
&(=(\psi(1)+\psi(-1))\sum_{w\in Sp(2n,q)}\psi(Trw)).
\end{split}\end{aligned}$$
For our purposes, we only need the following expressions of Gauss sums for $SO(3,q)$ and $O(3,q)$. So we state them separately as a theorem. Also, for the ease of notations, we introduce
$$\begin{aligned}
G_1(q)=SO(3,q),~G_2(q)=O(3,q).\end{aligned}$$
\[C\] Let $\psi$ be any nontrivial additive character of $\mathbb{F}_q$. Then we have $$\begin{aligned}
\begin{split}
&\sum_{w\in G_1(q)}\psi(Trw)=\psi(1)qK(\psi;1),\\
&\sum_{w\in G_2(q)}\psi(Trw)=(\psi(1)+\psi(-1))qK(\psi;1).
\end{split}\end{aligned}$$
The next corollary follows from Theorem \[B\] and by simple change of variables.
\[D\] Let $\lambda$ be the canonical additive character of $\mathbb{F}_q$, and let $a\in \mathbb{F}_{q}^{*}$. Then we have $$\label{u}
\sum_{w\in G_1(q)}\lambda(aTrw)=\lambda(a)qK(\lambda;a^2),\qquad
\qquad\quad\quad\quad$$ $$\label{v}
\begin{split}
\sum_{w\in G_2(q)}\lambda(aTrw)&=(\lambda(a)+\lambda(-a))qK(\lambda;a^2)\\
&=2(Re\lambda(a))qK(\lambda;a^2)~(cf.~(\ref{s})).
\end{split}$$
\[E\]$($[@D4], $($5.3-5$)$$)$ Let $\lambda$ be the canonical additive character of $F_q$, $m\in
\mathbb{Z}_{\geq0}$, $\beta\in\mathbb{F}_q$. Then $$\label{w}
\sum_{a\in\mathbb{F}_{q}^{*}}\lambda(-a\beta)K(\lambda;a^2)^m=q\delta(m,q;\beta)-(q-1)^m,$$ where, for $m\geq1$, $$\label{x}
\delta(m,q;\beta)=|\{(\alpha_{1}^{},\cdots,\alpha_{m}^{})\in(\mathbb{F}_{q}^{*})^m|\alpha_{1}^{}+\alpha_{1}^{-1}+
\cdots,\alpha_{m}^{}+\alpha_{m}^{-1}=\beta\}|,$$ and $$\begin{aligned}
\delta(0,q;\beta)=
\begin{cases}
1,& \text {$\beta$=0,}\\
0,& \text {otherwise.}
\end{cases}\end{aligned}$$
\[F\] Here one notes that $$\label{y}
\begin{split}
\delta(1,q;\beta)&=|\{x\in\mathbb{F}_{q}^{}|x^2-\beta x+1=0\}|\\
&=
\begin{cases}
2,& \text {if $\beta^2-1\neq0$ is a square,}\\
1,& \text {if $\beta^2-1$=0,}\\
0,& \text {if $\beta^2-1$ is a nonsquare.}
\end{cases}
\end{split}$$
Let $G(q)$ be one of finite classical groups over $\mathbb{F}_q$. Then we put, for each $\beta\in\mathbb{F}_q$,
$$\begin{aligned}
N_{G(q)}(\beta)=|\{w\in G(q)|Tr(w)=\beta\}|.\end{aligned}$$
Then it is easy to see that
$$\label{z}
qN_{G(q)}(\beta)=|G(q)|+\sum_{a\in\mathbb{F}_{q}^{*}}\lambda(-a\beta)
\sum_{w\in G(q)}\lambda(aTrw).$$
For brevity, we write
$$\label{a1}
n_1(\beta)=N_{G_1(q)}(\beta),~n_2(\beta)=N_{G_2(q)}(\beta).$$
Using (\[u\])-(\[w\]), and (\[d1\]), one derives the following.
\[G\] With the notations in (\[x\]), (\[y\]), and (\[a1\]), we have: $$\label{b1}
n_1(\beta)=q^2-q+q\delta(1,q;\beta-1),\qquad \qquad \qquad \qquad$$ $$\label{c1}
n_2(\beta)=2q^2-2q+q\delta(1,q;\beta-1)+q\delta(1,q;\beta+1).$$
\[H\] $Tr:G_1(q)\rightarrow \mathbb{F}_q$, and $Tr:G_2(q)\rightarrow
\mathbb{F}_q$ are surjective.
This is immediate from the above Proposition \[G\].
Construction of codes
=====================
Let $$\label{d1}
N_1=|G_1(q)|=q(q^2-1),~N_2=|G_2(q)|=2q(q^2-1).$$ Here we will construct ternary linear codes $C(G_1(q))$ of length $N_1$ and $C(G_2(q))$ of length $N_2$, respectively associated with the orthogonal groups $G_1(q)$ and $G_2(q)$. By abuse of notations, let $g_1,g_2,\cdots,g_{N_i}$ be a fixed ordering of the elements in the group $G_i(q)$, for $i=1,2$.\
Also, we put $$\begin{aligned}
v_i=(Trg_1,Trg_2,\cdots,Trg_{N_i})\in\mathbb{F}_q^{N_i},~for~i=1,2.\end{aligned}$$ Then the ternary linear code is defined as $$\label{e1}
C(G_i(q))=\{u\in\mathbb{F}_3^{N_i}|u\cdot v_i=0\},~for~i=1,2,$$ where the dot denotes the usual inner product in $\mathbb{F}_q^{N_i}$.
The following theorem of Delsarte is well-known.
\[I\] $($[@MS1]$)$ Let B be a linear code over $\mathbb{F}_q$. Then $$\begin{aligned}
(B|_{\mathbb{F}_3})^\bot=tr(B^\bot).\end{aligned}$$ In view of this theorem, the dual $C(G_i(q))^\bot$ is given by $$\label{f1}
C(G_i(q))^\bot=\{c_i(a)=(tr(aTrg_1),\cdots,tr(aTrg_{N_i}))|a\in\mathbb{F}_q\},~for~i=1,2.$$
\[J\] For every $q=3^r$, the map $\mathbb{F}_q\rightarrow
C(G_i(q))^\bot(a\mapsto c_i(a))$ is an $\mathbb{F}_3$-linear isomorphism, for $i=1,2$.
The maps are clearly $\mathbb{F}_3$-linear and surjective. Let $a$ be in the kernel of either of the map. Then, in view of Corollary \[H\], $tr(a\beta)=0$, for all $\beta\in\mathbb{F}_q$. Since the trace function $\mathbb{F}_q \rightarrow \mathbb{F}_2$ is surjective, $a=0$.
Power moments of Kloosterman sums with trace nonzero square arguments
=====================================================================
In this section, we will be able to find, via Pless power moment identity, recursive formulas for the power moments of Kloosterman sums with trace nonzero square arguments in terms of the frequencies of weights in $C(SO(3,q))$ and $C(O(3,q))$.
\[K\] $($Pless power moment identity, [@MS1]$)$ Let $B$ be an $q$-ary $[n,k]$ code, and let $B_i$(resp. $B_i^\bot$) denote the number of codewords of weight $i$ in $B$(resp. in $B^\bot$). Then, for $h=0,1,2,\cdots,$ $$\label{g1}
\sum_{j=0}^{n}j^hB_j=\sum_{j=0}^{min\{n,h\}}(-1)^jB_j^\bot
\sum_{t=j}^{h}t!S(h,t)q^{k-t}(q-1)^{t-j}{\binom{n-j}{n-t}},$$ where $S(h,t)$ is the Stirling number of the second kind defined in (\[i\]).
\[LL\] Let $c_i(a)=(tr(aTrg_1),\cdots,tr(aTrg_{N_i}))\in C(G_i(q))^\bot$, for $a\in\mathbb{F}_{q}^{*}$, and $i=1,2$. Then the Hamming weight $w(c_i(a))$ can be expressed as follows: $$\label{h1}
w(c_i(a))={\frac{2qi}{3}}(q^2-1-(Re\lambda(a))K(\lambda;a^2)),~for~i=1,2.$$
For $i=1,2,$ $$\begin{aligned}
\begin{split}
w(c_i(a))&=\sum_{j=1}^{N_i}(1-{\frac{1}{3}}\sum_{\alpha\in\mathbb{F}_3}
\lambda_0(\alpha tr(aTrg_j)))\\
&=N_i-{\frac{1}{3}}\sum_{\alpha\in\mathbb{F}_3}\sum_{w\in
G_i(q)}\lambda(\alpha aTrw)\\
&={\frac{2}{3}}N_i-{\frac{1}{3}}\sum_{\alpha\in\mathbb{F}_3^*}\sum_{w\in
G_i(q)}\lambda(\alpha aTrw).
\end{split}\end{aligned}$$ Our results now follow from (\[s\]), (\[u\]), (\[v\]) and (\[d1\]).
Fix $i(i=1,2)$, and let $u=(u_1,\cdots,u_{N_i})\in\mathbb{F}_3^{N_i}$, with $\nu_\beta$ 1’s and $\mu_\beta$ 2’s in the coordinate places where $Tr(g_j)=\beta$, for each $\beta\in\mathbb{F}_{q}^{}$. Then we see from the definition of the code $C(G_i(q))$(cf. (\[e1\])) that $u$ is a codeword with weight $j$ if and only if $\sum_{\beta\in\mathbb{F}_q}\nu_\beta+\sum_{\beta\in\mathbb{F}_q}\mu_\beta=j$ and $\sum_{\beta\in\mathbb{F}_q}\nu_\beta
\beta=\sum_{\beta\in\mathbb{F}_q}\mu_\beta \beta$(an identity in $\mathbb{F}_q$). Note that there are $\prod_{\beta\in\mathbb{F}_q}{\binom{n_i(\beta)}{\nu_\beta,\mu_\beta}}$(cf. (\[d\]), (\[e\])) many such codewords with weight $j$. Now, we get the following formulas in (\[i1\])-(\[j1\]), by using the explicit values of $n_i(\beta)$ in (\[b1\]),(\[c1\])(cf. (\[x\]), (\[y\])).
\[M\] Let $q=3^r$ be as before, and let $\{C_{i,j}\}_{j=0}^{N_i}$ be the weight distribution of $C(G_i(q))$, for $i=1,2.$ Then
$(1)$ $$\label{i1}
C_{1,j}=\sum\prod_{\beta\in\mathbb{F}_q}{\binom{n_1(\beta)}{\nu_\beta,\mu_\beta}}~
(j=0,\cdots,N_1),\qquad \qquad \qquad \qquad$$ with $$\begin{split}
n_1(\beta)&=q^2-q+q\delta(1,q;\beta-1)\\
&=\sum{\binom{q^2}{\nu_0,\mu_0}}{\binom{q^2}{\nu_2,\mu_2}}
\prod_{\beta^2-2\beta\neq0~square}{\binom{q^2+q}{\nu_\beta,\mu_\beta}}\\
&\quad\qquad\qquad\qquad\qquad\qquad\qquad\times\prod_{\beta^2-2\beta~nonsquare}
{\binom{q^2-q}{\nu_\beta,\mu_\beta}},
\end{split}$$
$(2)$ $$\label{j1}
C_{2,j}=\sum\prod_{\beta\in\mathbb{F}_q}{\binom{n_2(\beta)}{\nu_\beta,\mu_\beta}}~
(j=0,\cdots,N_2),\qquad \qquad \qquad \qquad$$ with $$n_2(\beta)=2q^2-2q+q\delta(1,q;\beta-1)+q\delta(1,q;\beta+1).\qquad\qquad\quad$$
Here in both (\[i1\]) and (\[j1\]) the unspecified sums run over all the sets of nonnegative integers $\{\nu_\beta\}_{\beta\in\mathbb{F}_q}$ and $\{\mu_\beta\}_{\beta\in\mathbb{F}_q}$ satisfying $$\begin{aligned}
\sum_{\beta\in\mathbb{F}_q} \nu_\beta
+\sum_{\beta\in\mathbb{F}_q}\mu_\beta=j\quad and~
\sum_{\beta\in\mathbb{F}_q} \nu_\beta
\beta=\sum_{\beta\in\mathbb{F}_q}\mu_\beta \beta,\end{aligned}$$ and, for every $\beta\in\mathbb{F}_q$, $\delta(1,q;\beta)$ is as in (\[y\]).\
The recursive formula in the following theorem follows from the study of ternary linear codes associated with the symplectic group $Sp(2,q)=SL(2,q)$. It is slightly modified from its original version, which makes it more usable in below.\
\[L\]$($[@DJ1]$)$ For $h=1,2,3,\cdots$, $$\label{k1}
\begin{split}
2(\frac{2q}{3})^h\sum_{j=0}^{h}&(-1)^j{\binom{h}{j}}(q^2-1)^{h-j}SK^{j}\\
&=q\sum_{j=0}^{min\{N_1,h\}}(-1)^j
\hat{C}_{j}\sum_{t=j}^{h}t!S(h,t)3^{-t}2^{t-j}
{\binom{N_1-j}{N_1-t}},
\end{split}$$ where $N_1=q(q^2-1)=|Sp(2,q)|=|SO(3,q)|$, $S(h,t)$ indicates the Stirling number of the second kind as in (\[i\]), and $\{\hat{C}_j\}_{j=0}^{N_1}$ denotes the weight distribution of the ternary linear code $C(Sp(2,q))$, given by $$\begin{aligned}
\begin{split}
\hat{C}_j&=\sum_{\beta\in\mathbb{F}_q}\prod_{\beta}{\binom{q\delta(1,q;\beta)+q^2-q}{\nu_\beta,\mu_\beta}}\\
&=\sum{\binom{q^2}{\nu_1,\mu_1}}{\binom{q^2}{\nu_{-1},\mu_{-1}}}
\prod_{\beta^2-1\neq0~square}{\binom{q^2+q}{\nu_\beta,\mu_\beta}}\prod_{\beta^2-1~nonsquare}{\binom{q^2-q}{\nu_\beta,\mu_\beta}}
\end{split}\end{aligned}$$ $(j=0,\cdots,N_1)$.\
Here the sum is over all the sets of nonnegative integers $\{\nu_\beta\}_{\beta\in\mathbb{F}_q}$ and $\{\mu_\beta\}_{\beta\in\mathbb{F}_q}$ satisfying $\sum_{\beta\in\mathbb{F}_q} \nu_\beta
+\sum_{\beta\in\mathbb{F}_q}\mu_\beta=j$ and $\sum_{\beta\in\mathbb{F}_q} \nu_\beta
\beta=\sum_{\beta\in\mathbb{F}_q}\mu_\beta \beta$.
We are now ready to apply the Pless power moment identity in (\[g1\]) to $C(G_i(q))^\bot$ for $i=1,2$, in order to obtain the result in Theorem \[A\](cf. (\[f\])-(\[h\]), (\[j\])-(\[l\])) about recursive formulas. We do this for $i=1,2$ at the same time.\
The left hand side of that identity in (\[g1\]) is equal to $$\label{l1}
\sum_{a\in\mathbb{F}_{q}^{*}}w(c_i(a))^h,$$ with the $w(c_i(a))$ given by (\[h1\]).\
In below, $\lq\lq$the sum over $tra=0$ (resp. $tra\neq0$)“ will mean $\lq\lq$the sum over all $a\in\mathbb{F}_{q}^{*}$ with $tra=0$(resp. $tra\neq0$).”\
(\[l1\]) is given by
$$\begin{aligned}
\begin{split}
({\frac{2qi}{3}})^h&\sum_{a\in\mathbb{F}_{q}^{*}}(q^2-1-(Re\lambda(a))K(\lambda;a^2))^h\\
=&({\frac{2qi}{3}})^h\sum_{tra=0}(q^2-1-K(\lambda;a^2))^h\\
&\qquad+({\frac{2qi}{3}})^h\sum_{tra\neq0}(q^2-1+{\frac{1}{2}}K(\lambda;a^2))^h\\
(noting&~ that~ Re\lambda(a)=1,~ if~ tra=0; Re\lambda(a)=-\frac{1}{2},~ if~ tra\neq0, i.e., tra=1,2)\\
=&({\frac{2qi}{3}})^h\sum_{tra=0}\sum_{j=0}^h(-1)^j{\binom{h}{j}}(q^2-1)^{h-j}K(\lambda;a^2)^j\\
&\qquad+({\frac{2qi}{3}})^h\sum_{tra\neq0}\sum_{j=0}^h{\binom{h}{j}}(q^2-1)^{h-j}2^{-j}K(\lambda;a^2)^j\\
=&({\frac{2qi}{3}})^h\sum_{j=0}^h(-1)^j{\binom{h}{j}}(q^2-1)^{h-j}(2SK^j-T_{12}SK^j)\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\text{($\psi=\lambda$ case of (\ref{a}), (\ref{b}))}\\
&\qquad+({\frac{2qi}{3}})^h\sum_{j=0}^h{\binom{h}{j}}(q^2-1)^{h-j}2^{-j}T_{12}SK^j\\
=&i^h2({\frac{2q}{3}})^h\sum_{j=0}^h(-1)^j{\binom{h}{j}}(q^2-1)^{h-j}SK^j\\
&\qquad+({\frac{2qi}{3}})^h\sum_{j=0}^h((-1)^{j+1}+2^{-j}){\binom{h}{j}}(q^2-1)^{h-j}T_{12}SK^j\\
\end{split}\end{aligned}$$
$$\label{m1}
\begin{split}
=i^hq&\sum_{j=0}^{min\{N_1,h\}}(-1)^j
\hat{C}_{j}\sum_{t=j}^{h}t!S(h,t)3^{-t}2^{t-j}
{\binom{N_1-j}{N_1-t}}~\text{(from (\ref{k1}))}\hspace{8mm}\\
&+({\frac{2qi}{3}})^h\sum_{j=0}^h((-1)^{j+1}+2^{-j}){\binom{h}{j}}(q^2-1)^{h-j}T_{12}SK^j.
\end{split}$$
On the other hand, the right hand side of (\[g1\]) is $$\label{n1}
q\sum_{j=0}^{min\{N_1,h\}}(-1)^j
C_{i,j}\sum_{t=j}^{h}t!S(h,t)3^{-t}2^{t-j} {\binom{N_i-j}{N_i-t}}.$$ Here one has to note that $dim_{\mathbb{F}_2}C(SO(3,q))=dim_{\mathbb{F}_2}C(O(3,q))=r$(cf. Prop. \[J\]) and to separate the term corresponding to $l=h$ of the second sum in (\[m1\]). Our main results in Theorem \[a\] now follow by equating (\[m1\]) and (\[n1\]).\
\[N\] Let $q=3^r$. Then we have the following.
1. \[Na\] $SK={\frac{1}{2}}\{(-1)^rq+1\}$,
2. \[Nb\] $T_0SK={\frac{1}{3}}(-1)^rq+1$,
3. \[Nc\] $T_{12}SK={\frac{2}{3}}(-1)^r q$.
From either (\[f\]) or (\[j\]), we get (\[Nc\]). (\[Na\]) follows from our previous result ([@DJ1], (4)) or can be derived directly as follows. $$\begin{aligned}
\begin{split}
SK&={\frac{1}{2}}\sum_{a\in\mathbb{F}_{q}^{*}}K(\lambda;a^2)\\
&={\frac{1}{2}}\sum_{a\in\mathbb{F}_{q}^{*}}\sum_{\alpha\in\mathbb{F}_{q}^{*}}
\lambda(\alpha_{}^{} +a^2\alpha_{}^{-1})\\
&={\frac{1}{2}}\sum_{a\in\mathbb{F}_{q}^{*}}\sum_{\alpha\in\mathbb{F}_{q}^{*}}
\lambda(a(\alpha_{}^{} +\alpha_{}^{-1}))\\
\end{split}\end{aligned}$$ $$\label{o1}
\qquad\qquad\quad\qquad={\frac{1}{2}}\sum_{\alpha\in\mathbb{F}_{q}^{*}}\sum_{a\in\mathbb{F}_{q}^{}}
\lambda(a(\alpha_{}^{} +\alpha_{}^{-1}))-{\frac{1}{2}}(q-1)$$ $$\hspace{11.5mm}=
\begin{cases}
\frac{1}{2}2q-\frac{1}{2}(q-1),& \text {if $r$ even},\\
-\frac{1}{2}(q-1),& \text {if $r$ odd}.
\end{cases}$$ In (\[o1\]), we note that $\alpha+\alpha^{-1}=0$ has a solution in $\mathbb{F}_{q}^{*}$ if and only if -1 is a square in $\mathbb{F}_q$ if and only if $r$ is even, in which case there are two distinct solutions. Finally, (\[Nb\]) follows from the relation (\[b\]) with $h=1$ and $\psi=\lambda$.
[10]{}
R.J. Evans, *Seventh power moments of Kloosterman sums*, Israel J. Math., **** to appear.
G. van der Geer, R. Schoof and M. van der Vlugt, *Weight formulas for ternary Melas codes*, Math. Comp. Math. **58**(1992), 781–792.
K. Hulek, J. Spandaw, B. van Geemen and D. van Straten, *The modulartiy of the Barth-Nieto quintic and its relatives*, Adv. Geom. **1** (2001), 263–289.
D. S. Kim, *Gauss sums for symplectic groups over a finite field*, Mh. Math. **126** (1998), 55–71.
D. S. Kim, *Exponential sums for symplectic groups and their apllications*, Acta Arith. **88** (1999), 155–171.
D. S. Kim, *Gauss sums for $O^{}(2n+1,q)$*, Finite Fields Appl. **4** (1998), 62–86.
D. S. Kim, *Codes associated with $O(3,2^r)$ and power moments of Kloosterman sums with trace one arguments*, **** submitted.
D. S. Kim, *An infinite family of recursive formulas generating power moments of Kloosterman sums with trace one arguments: $O(2n+1,2^r)$ case*, **** submitted.
D. S. Kim and J. H. Kim, *Ternary codes associated with symplectic groups and power moments of Kloosterman sums with square arguments*, submitted.
H. D. Kloosterman, *On the representation of numbers in the form $ax^2+by^2+cz^2+dt^2$* , Acta Math. **49** (1926), 407-464.
R. Lidl and H. Niederreiter, *Finite Fields*, Encyclopedia Math. Appl.**20**, Cambridge University Pless, Cambridge, 1987.
R. Livn$\acute{e}$, *Motivic orthogonal two-dimensional representations of $Gal(\overline{\mathbb{Q}}/\mathbb{Q})$*, Israel J. Math. **92** (1995), 149-156.
F. J. MacWilliams and N. J. A. Sloane, *The Theory of Error Correcting Codes*, North-Holland, Amsterdam, 1998.
M. Moisio, *On the moments of Kloosterman sums and fibre products of Kloosterman curves*, Finite Fields Appl. **14**(2008), 515–531.
C. Peters, J. Top, and M. van der Vlugt, *The Hasse zeta function of a K3 surface related to the number of words of weight 5 in the Melas codes*, J. Reine Angew. Math. **432** (1992), 151-176.
H. Sali$\acute{e}$, *$\ddot{U}$ber die Kloostermanschen Summen $\mathcal{S}(u,v;q)$*, Math. Z. **34**(1931), 91-109.
[^1]: This work was supported by National Research Foundation of Korea Grant funded by the Korean Government 2009-0072514.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $\Delta =\{ \delta _1,\delta _2,...,\delta _m \} $ be a finite set of 2-connected patterns, i.e. graphs up to vertex relabelling. We study the generating function $D_{\Delta }(z,u_1,u_2,...,u_m),$ which counts polygon dissections and marks subgraph copies of $\delta _i$ with the variable $u_i$. We prove that this is always algebraic, through an explicit combinatorial decomposition depending on $\Delta $. The decomposition also gives a defining system for $D_{\Delta }(z,\mathbf{0})$, which encodes polygon dissections that avoid these patterns as subgraphs. In this way, we are able to extract normal limit laws for the patterns when they are encoded, and perform asymptotic enumeration of the resulting classes when they are avoided. The results can be transfered to the case of labelled outerplanar graphs. We give examples and compute the relevant constants when the patterns are small cycles or dissections.'
author:
- 'Vasiliki Velona[^1] [^2] [^3]'
title: 'Encoding and avoiding 2-connected patterns in polygon dissections and outerplanar graphs'
---
Introduction
============
The study of subgraph appearances in random graph models is a well established line of research, beginning with the classic *Erdős-Rényi* graph and results concerning the distribution of such appearances and threshold phenomena, as in [@karonski1983number],[@rucinski1988small]. In parallel, attention was also drawn on models where, given some well-known graph class, an object is chosen uniformly at random from all the objects of size $n$; see for instance [@kim2007small] and [@gao2008distribution] for regular graphs. In the last decades, techniques using a mixture of generating function theory and analytic tools have evolved significantly and are in the centre of such advances for various other graph classes. A number of graph statistics, such as number of components, edges, cut vertices, triangles, chromatic number and others, have been studied for standard graph classes, such as planar graphs, outerplanar, series-parallel, graphs of fixed genus, and minor-closed families; see for instance [@chapuy2011asymptotic], [@bodirsky2007enumeration], [@gimenez2013graph], [@mcdiarmid2009random].
In [@subcritical], the authors present a normality result for the so-called *subcritical* family of graphs, that contains standard graph classes such as trees, cacti graphs, outerplanar, and series-parallel graphs. In particular, all subgraph parameters in such a class follow a normal limit law, with linear mean and variance. However, no constructive way is given in it, in order to compute the corresponding constants for the mean and variance. One of the results of this work is an explicit way to do so in outerplanar graphs, for any set of 2-connected patterns, i.e. graphs up to vertex relabelling. As a case study, we examine 3 and 4-cycles, but the process by which these constants are obtained can be directly transferred to the case of any set of 2-connected parameters.
\[222\] The number of appearances $X_n$ of 3-cycles and 4-cycles in polygon dissections and outerplanar graphs of size $n$ follows a normal limit law, as in \[quasi\], where the mean and variance are asymptotically linear, i.e. $\mathbb{E}[X_n]=\mu n+\mathcal{O}(1)$ and $ \mathbb{V}\textsl{ar}\hspace{.03cm}[X_n]=\sigma ^2 n+\mathcal{O}(1)$. The constants $\mu $ and $\sigma ^2$ are the following, in their exact values for dissections and in approximation for outerplanar graphs:
-------------------------------------------------------------------------------------------------------------------------------------------------------
Parameter $\mu $ $\sigma ^2 $ $\mu $ $\sigma ^2 $
----------- ------------------------------------------------- ---------------------------------------------------------------- --------- --------------
3-cycles $\frac{1}{2}$ ${\frac {-13+9\,\sqrt {2}}{-12+8\,\sqrt {2}}}\approx 0.39644$ 0.34793 0.40737
4-cycles $ {\frac {-30+21\,\sqrt {2}}{-12+8\,\sqrt {2}}} $\,{\frac {-24216+17123\,\sqrt {2}}{-32 \left( -3+2\,\sqrt {2} 0.33705 0.36145
\right) ^{2}}}
\approx 0.43933$ \approx 0.44710$
-------------------------------------------------------------------------------------------------------------------------------------------------------
A necessary step for the analysis of outerplanar graphs is the analysis of polygon dissections, denoted by $\mathcal{D}$, with some fixed numbering on the vertices. For a finite set of 2-connected patterns $\Delta =\{ \delta _1,\delta _2,...,\delta _k \} $, we prove a combinatorial decomposition of $\mathcal{D}$ that allows the encoding of such patterns, depending on $\Delta $. In this way, we obtain defining systems for the multivariate generating function $D_{\Delta }(z,u_1,u_2,...,u_k)$, where the coefficient of $z^nu_1^{n_1}\cdots u_m^{n_m}$ counts the number of $\alpha \in\mathcal{D}$ that have $n$ vertices and $n_i$ subgraph occurrences of the pattern $\delta _i$.
This task is of independent interest, as it is related to the enumeration problem of polygon dissections, a line of work that is quite old. Starting from the enumeration of polygon triangulations with Euler and Segner in the 18th century, a great amount of work has been devoted up until today to relevant problems. Usually, these problems put restrictions either on the number or the size of the partition’s polygonal components, or even colour restrictions, recently; see for instance [@cayley], [@read1978general], [@birmajer2017colored]. However, the problem where a whole pattern is avoided as subgraph (i.e., cannot be recovered by applying edge and vertex deletions) seems to not have been studied at all, except for the case of triangle freeness in [@birmajer2017colored], where the problem the authors are dealing with does not concern subgraph restrictions, but restrictions on the type and colour of the partition’s polygonal components. With results of this work, it is possible to handle subgraph restrictions of any set $\Delta $ and perform asymptotic enumeration of the resulting classes. We give such examples. In fact, we obtain the following results (corresponding to Corollary \[cor\] and Theorem \[enumm\], respectively):
\[alg\] The generating function $D_{\Delta }(z,\mathbf{u})$ is algebraic and the defining polynomial is computable. The generating function of polygon dissections that avoid all $\Delta $-patterns as subgraphs, $D_{\Delta }(z,\mathbf{0})$, is likewise algebraic.
\[ex\] Let $\mathcal{D}, \mathcal{G}$ be the classes of dissections and outerplanar graphs avoiding a set of 2-connected patterns $\Delta =\{\delta _1 ,...,\delta _m\}$, respectively. Then, $\mathcal{D}$ and $\mathcal{G}$ have asymptotic growth of the form: $$\alpha _n\sim \frac{\alpha}{\Gamma (-\frac{1}{2})}\cdot n^{-3/2}\cdot r ^{-n} \quad \mathrm{and}\quad g_n\sim \frac{g}{\Gamma (-\frac{3}{2})}\cdot n^{-5/2} \cdot \rho ^{-n}\cdot n!,$$ respectively, where both $\alpha ,g$ are computable constants. In Table \[table:3\], there are approximations of $\alpha , g$ for various choices of $\Delta $.
We also prove a multivariate central limit theorem for the number of appearances of 2-connected patterns in polygon dissections (corresponding to Theorem \[random\]):
\[1st\] Let $\Delta =\{\delta _1,...,\delta _m\}$ be a set of 2-connected patterns. Let $\Omega _n$ be the set of polygon dissections of size $n$ and $X_n:\Omega _n \rightarrow \mathbb{Z}_{\geq 0}^m$ be a vector of random variables $X_{\delta _1},...,X_{\delta _m}$ in $\Omega _n$, such that $X_{\delta _i}(\omega )$ is the number of $\delta _i$ patterns in $\omega \in\Omega _n$. Then, $\mathbf{X}_n$ satisfies a central limit theorem $$\frac{1}{\sqrt{n}}(\mathbf{X}_n-\mathbb{E}[\mathbf{X}_n])\xrightarrow[]{d} N(\mathbf{0},\mathbf{\Sigma })$$ with $$\mathbb{E}[\mathbf{X}_n]=\boldsymbol{\mu }n+\mathcal{O}(1) \text{ and } \mathbb{C}\textsl{ov}\hspace{.03cm}[\mathbf{X}_n]=\mathbf{\Sigma}n+\mathcal{O}(1),$$ where $\boldsymbol{\mu }$ and $\mathbf{\Sigma}$ are computable.
There are some natural questions arising from this work. One is whether it is possible to extend the combinatorial construction that is proved for general parameters, with multiple cut vertices, and how. Also, one might wonder in which other combinatorial structures we can apply this reasoning, apart from outerplanar graphs. An example for the latter can be found in the dual class of polygon dissections, planted plane trees with outdegrees in $\mathbb{N}\setminus \{1\}$, denoted by $\mathcal{T}$. Consider as parameter in $T\in\mathcal{T}$ the number of subtrees $T'$ with $k$ leaves, such that $\deg _T(v)=\deg _{T'}(v)$ for each node $v$ that is inner in $T'$. Then, the equivalent parameter for polygon dissections is the number of $k$-cycles. **Plan of the paper.** In Section 2, we mention definitions and theorems that will be used. In Section 3, we prove a combinatorial decomposition of $\mathcal{D}$ depending on $\Delta $ and then Theorem \[alg\]. We also prove Theorem \[1st\]. In section 4, we give applications of the previous and prove Theorem \[222\] and Theorem \[ex\]. In the Appendix, Table \[count\] contains the initial terms of all the counting sequences appearing in Section 4.
Preliminaries
=============
The framework we use is the *symbolic method* and the corresponding analytic techniques, as they were presented in [@flajolet1999analytic].
#### Symbolic methods for counting.
A *combinatorial class* is a set $\mathcal{A}$ with a *size function* $\mathcal{A}\rightarrow \mathbb{Z}_{\geq 0}$, such that the inverse image of any integer is a finite set, denoted $\mathcal{A}_n$. Each $\alpha\in\mathcal{A}_n$ comprises $n$ *atoms* of size 1, and we denote by $\mathcal{Z}$ the *atomic class* that contains exactly one object of size one. In this work, atoms always represent graph vertices. We call a class $\mathcal{A}$ *labelled* if the atoms have labels and $\mathcal{A}$ is closed under atom relabelling. The *ordinary generating function* $A(z)$ of $\mathcal{A}$, referred also as *ogf*, is defined as $\sum _{n=0}^{\infty }|\mathcal{A}_n|z^n$. If $\mathcal{A}$ is labelled, we use the *exponential generating function* $A(z)$, referred also as *egf*, that is defined as $\sum _{n=0}^{\infty }|\mathcal{A}_n|\frac{z^n}{n!}$. We then write $[z^n]A(z)=|\mathcal{A}_n|$ for ogfs and $[z^n]A(z)=\frac{|\mathcal{A}_n|}{n!}$ for egfs.[^4]
In order to create functional equations for the generating functions of interest, we use the so-called *admissible* combinatorial constructions from [@flajolet1999analytic]. The aim is to express a combinatorial class in terms of other ones, itself included, in an *admissible way*. Then, there is a direct translation in terms of generating functions. From unlabelled classes and ogfs, we only need the elementary cases $\mathcal{A}=\mathcal{B}\cup\mathcal{C}\Rightarrow A(z)=B(z)+C(z)$ and $\mathcal{A}=\mathcal{B}\times\mathcal{C}\Rightarrow A(z)=B(z)\cdot C(z)$. In Table \[table:1\], there are all the labelled constructions that are useful to this work, along with their translations to egfs. It is useful to consider *parameters* on the objects of $\mathcal{A},$ i.e functions $\chi _i:\mathcal{A}\rightarrow \mathbb{Z}_{\geq 0}$ that quantify some structure of the objects. Let $\mathbf{j} $ be $(j_1,...,j _{m})$[^5], where $j_i \in \mathbb{Z}_{\geq 0}$ and let us define $\mathcal{A} _{n,\mathbf{j}}$ as the set of elements $\alpha\in\mathcal{A}$ that have size $n$ and $\chi _{i}(\alpha )=j _{i}.$ Then we work with multivariate generating functions, ordinary $\sum _{n,j _i\geq 0}|\mathcal{A}_{n,\mathbf{j} }|z^{n}u_1^{j _1}u_1^{j _2}...u_m^{j _{m}}$ for unlabelled classes and exponential $\sum _{n,j _i\geq 0}|\mathcal{A} _{n,\mathbf{j} }|\frac{z^{n}}{n!}u_1^{j _1}u_2^{j _2}...u_m^{j _{m}}$ for labelled ones. All the mentioned translations are also valid for multivariate generating functions, if the parameters are *compatible*, i.e. $\chi (\alpha ')$ is the same for all order-preserving relabelings $\alpha '$ of $\alpha \in \mathcal{A}$, and *additive*, i.e. $\chi (\alpha )=\sum _i \chi (\beta _i)$ when $\alpha \in \mathcal{A}$ is composed of smaller elements $\beta _i\in \mathcal{B}$.
A generating function $y(z,\mathbf{u})$ is called *algebraic* if it satisfies a polynomial equation $P(z,y,\mathbf{u})=0$.
--------------------- -------------------------------------------- ------------------ --
*Labelled product:* $\mathcal{B}\star \mathcal{C}$ $B(z)\cdot C(z)$
*Set:* $\mathrm{S\scalebox{.9}{ET}}(\mathcal{B})$ $\exp (B(z))$
*Substitution:* $\mathcal{B}\circ \mathcal{C}$ $ B(C(z)) $
--------------------- -------------------------------------------- ------------------ --
: Some labelled constructions and their translations to exponential generating functions.[]{data-label="table:1"}
------------- ------------------------- ---------------------
*Pointing:* $\mathcal{B}^{\bullet}$ $ z\partial _z B(z)
$
*Deriving:* $\mathcal{B}^{\circ}$ $ \partial _z B(z)
$
------------- ------------------------- ---------------------
: Some labelled constructions and their translations to exponential generating functions.[]{data-label="table:1"}
#### Graph theoretic preliminaries.
We now mention some basic graph theoretic language and refer to [@diestel2010graph] for a rigorous exposition.
A *graph* $G(V,E)$ is defined by the vertex set $V$ and the edge set $E$ that is a set of 2-element subsets of $V$. If the elements of $E$ are ordered pairs of vertices, the graph is called *directed*. A graph $G_1$ is a *subgraph* of $G_2$ if it can be obtained by $G_2$ with edge and vertex deletions. In this work, a *pattern* is the equivalence class of a graph, up to vertex relabelling.
A graph is called *2-connected* if at least two vertex deletions are needed in order to disconnect it.
A graph is an $m$-*cycle*, denoted $C_m$, if $E=\{\{v_m,v_1\},\{v_1,v_2\},...,\{v_{m-1},v_m\}\}$ for $m> 2$ and some ordering $v_1,...,v_m$ of $V$. Let $C_m$ be a subgraph of $G$. Any edge $\{v_i,v_j\}\in E_G,$ such that $\{v_i,v_j\}\subset V_{C_m}$ and $\{v_i,v_j\}\not\in E_{C_m}$ is called a *chord* of $C_m$. If $V_{C_m}=V$, $C_m$ is called a *Hamilton cycle* of $G$.
Suppose that $G$ admits a *planar* embedding $\Gamma $ on the plane, i.e. an embedding such that the edges do not cross one another. The closures of the connected components of $\mathbb{R}^2\backslash \Gamma $ are called *faces* of the embedding and there is always a unique face that is unbounded. Edges that lie on this face will be called *outer*; otherwise, they will be called *inner*.
#### Outerplanar Graphs.
Let $P$ be a polygon with vertices numbered in $\{1,...,n\}$, in counterclockwise order. A *polygon dissection* is an arrangement of diagonals on $P$, such that no two of them are intersecting.
*Outerplanar graphs* are graphs that can be embedded on the plane in such a way that all vertices lie on the boundary of the unbounded face. Let $\mathcal{G}$ be the class of labelled outerplanar graphs. In [@bodirsky2007enumeration], the authors bring together a set of combinatorial constructions that define the class $\mathcal{G}$ and involve the classes $\mathcal{D}, \mathcal{B}, \mathcal{C}$, corresponding to polygon dissections, labelled 2-connected, and labelled connected outerplanar graphs, respectively. These constructions translate to the functional equations in Table \[table:2\]. Note that the first one is an ordinary generating function and the rest are exponential.
------------------------- -----------------------------------------------
Polygon dissections $D(z)=z/4+z^2/4-z/4\,\sqrt {{z}^{2}-6\,z+1} $
2-connected outerplanar $B'(z)=\frac{1}{2z}D(z)+\frac{z}{2}$
Connected outerplanar $zC'(z)=z\exp (B'(zC'(z)))$
General outerplanar $G(z)=\exp (C(z))$
------------------------- -----------------------------------------------
: A defining set of equations for labelled outerplanar graphs[]{data-label="table:2"}
We are interested on how the first equation is derived. Each 2-connected outerplanar graph with $|V|>2$ has a Hamilton cycle, so, assuming a numbering on it, counting 2-connected outerplanar graphs of size $n$ is equivalent to counting dissections of the same size. The starting point is [@flajolet1999analytic], where polygon dissections were modelled in a symbolic way, based on the recursive structure of the class, as shown in Figure \[fig:1\]. In short, one designates an edge of the polygon, say the $e=\{v_1,v_2\}$ edge, and then divides the dissections according to the size of the polygon where $e$ lies. The latter will be called *root polygon* and $e$ will be called *root*. On the rest of the edges, other dissections are attached.
![A combinatorial decomposition of fixed-polygon dissections.[]{data-label="fig:1"}](Dissections)
The following translation is then implied, in terms of ogfs: $$D=z^2+\frac{D^2}{z}+\frac{D^3}{z^2}+\frac{D^4}{z^3}+...=z^2+\frac{D^2}{z-D}\Rightarrow 2D^2-D(z+z^2)+z^3=0.\label{eq:1}$$
The second equation in Table \[table:2\] is derived by observing that $n![z^n]B(z)=\frac{(n-1)!}{2}[z^n]D(z)$. The third and fourth correspond to the symbolic constructions: $\mathcal{Z}\star \mathcal{C}^{\circ}=\mathcal{Z}\star \mathrm{S\scalebox{.9}{ET}}(\mathcal{B}^{\circ}\circ \mathcal{C}^{\bullet})$ and $\mathcal{G}=\mathrm{S\scalebox{.9}{ET}}(\mathcal{C})$. The former relation is well-known (see for instance [@harary9graphical p.10],[@gimenez2009asymptotic], [@subcritical]) and is based on the decomposition of a graph into 2-connected components. The latter one is straightforward.
#### Analytic Preliminaries.
We denote by $\mathbf{y}=\mathbf{F}(z,\mathbf{y},\mathbf{u})$ a system of the form: $$\left\{
\begin{array}{lll}
y_1&=& f_1(z,\mathbf{y},\mathbf{u})\\
\vdots & & \vdots \\
y_{m} &=& f_{m}(z,\mathbf{y},\mathbf{u}).
\end{array}
\right.$$ Let $f_1,...,f_m$ be analytic functions with non-negative coefficients, such that $\mathbf{F}(0,\mathbf{0},\mathbf{u}) \equiv \mathbf{0}$, $\mathbf{F}(z,\mathbf{0},\mathbf{u}) \not\equiv \mathbf{0}$ and there exists $j$ with $\mathbf{F}_{y_jy_j}\not\equiv \mathbf{0}$, where $\mathbf{F}_{y_jy_j}$ denotes the second derivative with respect to $y_j$. To any such system, we relate a directed graph with vertices $y_i$ and edges $(y_i,y_j)\in E$ whenever $F_i$ depends on $y_j$, i.e. whenever $\frac{\partial F_i}{\partial y_j}\not\equiv 0$. We call this the *dependency graph* of the system and suppose that it is strongly connected, i.e. one can move from any vertex to any other through a path of directed edges. If such a system has unique analytic solutions with non-negative coefficients $y_i(z,u_1,...,u_{m})$ around $z=0,u_i=1$, it is called *well defined*. Then, Theorem \[drmota\] [@drmota1997systems Prop.3][@drmota2009random Ch.2] holds, adjusted to our purpose: the only missing requirement is $\mathbf{F}(0,\mathbf{y},\mathbf{u})=0$, but the result is still valid when one deals with well-defined systems.
\[drmota\] Let $\mathbf{y}=\mathbf{F}(z,\mathbf{y},\mathbf{u})$ be a well-defined system and let $\mathbf{y}=\mathbf{y}(z,\mathbf{u})=(y_1(z,\mathbf{u}),...,y_N(z,\mathbf{u}))$ denote the analytic solutions of the system. Suppose that the radius of convergence of $\mathbf{F}$ is large enough that there is a positive number $z_0$ of minimum modulus and real numbers $\mathbf{y}_0$ that satisfy the system $$\begin{gathered}
\label{charsystem}\begin{split}\mathbf{y} &= \mathbf{F}(z,\mathbf{y},\mathbf{1}) \\
0 &= \det (\mathbf{I}-\mathbf{F}_{\mathbf{y}}(z,\mathbf{y},\mathbf{1})).\end{split}\end{gathered}$$ Then there exist functions $\rho (\mathbf{u}), g_i(z,\mathbf{u}), h_i(z,\mathbf{u}),$ for $1\leq i\leq N,$ which are analytic around $z=z_0,\mathbf{u}=\mathbf{1},$ and satisfy $\rho (\mathbf{1})=z_0,$ $h_i(z_0,\mathbf{1})<0,$ such that: $$y_i(z,\mathbf{u})=g_i(z,\mathbf{u})-h_i(z,\mathbf{u})\sqrt{1-\frac{z}{\rho (u)}}\label{criticalexp}$$ locally around $z=z_0$, $\mathbf{u}=\mathbf{1}$ with $\arg (z-\rho (\mathbf{u}))\neq 0.$ Assume also that $[z^n]y_j(z,\mathbf{1})>0$ for $1 \leq j \leq N$ and for all large enough $ n$. Then, for $\mathbf{u}$ sufficiently close to $\mathbf{1}$, the radius of convergence of all $y_i$ is $\rho (u)$ and there are no other singularities on the circle of convergence $|z|=|\rho (u)|$ than $z=\rho (u)$. Furthermore, there exists $\epsilon >0$, such that $y_i$ can be analytically continued to the region $|z| < |\rho (u)|+\epsilon $, $|\arg(z-\rho (u))|>\epsilon $. \[eq:expansion\]
Note that, according to Condition , 1 is an eigenvalue of the matrix $\mathbf{F}_{\mathbf{y}}(z_0,\mathbf{y}_0,\mathbf{1})$. Systems like will be called *characteristic*. In expansions of the form , we will call *critical exponent* the first non-integer power of the expansion (in this case, for instance, the critical exponent is equal to $1/2$).
For the asymptotic analysis, we follow the transfer principles of *singularity analysis*, as they are presented in [@flajolet2009analytic]. Let $f(z)$ be an analytic function at zero with a unique smallest singularity at $z=\rho $ and $\rho >0$. We need the fact that, if $f(z)$ has a singular expansion $f(z)=a_0+a_1(1-z/ \rho )^{-\alpha }+\mathcal{O}\big( (1-z/ \rho )^{-\alpha +\delta })\big)$ in a domain $|z| \leq \rho+\epsilon $, $|z-\rho |\geq\epsilon $, where $\delta , \epsilon >0$ and $\alpha \in \mathbb{C}\backslash \mathbb{Z}_{\leq 0}$, then: $$[z^n]f(z)=a_1\frac{n^{\alpha -1}}{\Gamma (\alpha )}\rho ^{-n} \big(1+o(1)\big) ,$$ where $\Gamma (\alpha )$ refers to the *Euler Gamma function*, defined as $\Gamma (x)=\int _0^\infty t^{x-1}e^{-t}dt$. For the extraction of normal limit laws, we use Theorem 2.25 from [@drmota2009random].
Suppose that a sequence of $k$-dimensional random vectors $\mathbf{X}_n$ satisfies $\mathbb{E} [\mathbf{u}^{\mathbf{X}_n}]=\frac{c_n(\mathbf{u})}{c_n(\mathbf{1})},$ where $c_n(\mathbf{u})$ is the coefficient of $z^n$ of an analytic function $y(z,\mathbf{u})=\sum _{n\geq 0}c_n(\mathbf{u})z^n$ around $z=0,\mathbf{u}=\mathbf{1}$ and $c_n(\mathbf{u})>0$ for $n\geq n_0$ and positive real $\mathbf{u}$. Suppose also that $y(z,\mathbf{u})$ has a local singular representation of the form $$y(z,\mathbf{u})=g(z,\mathbf{u})+h(z,\mathbf{u})\Big( 1-\frac{z}{\rho (\mathbf{u})}\Big)^{\alpha}$$ for some real $\alpha \in \mathbb{R}\backslash \mathbb{N}$ and functions $g(z,\mathbf{u}),h(z,\mathbf{u})\neq 0$ and $\rho (\mathbf{u})\neq 0$ that are analytic around $z=z_0>0$ and $\mathbf{ u}=\mathbf{1}$. If $z=\rho (\mathbf{u})$ is the only singularity of $y(z,\mathbf{u})$ on the disk $|z|\leq |\rho (\mathbf{u})| $, when $\mathbf{u}$ is sufficiently close to $\mathbf{1}$, and there exists an analytic continuation of $y(z,\mathbf{u})$ to the region $|z|< |\rho (\mathbf{u})| +\delta$, $|\arg (z-\rho (\mathbf{u}))|>\epsilon $ for some $\delta >0$ and $\epsilon >0$, then $\mathbf{X}_n$ satisfies a central limit theorem $$\frac{1}{\sqrt{n}}(\mathbf{X}_n-\mathbb{E}[\mathbf{X}_n])\xrightarrow[]{d} N(\mathbf{0},\mathbf{\Sigma })$$ with $$\mathbb{E}[\mathbf{X}_n]=\boldsymbol{\mu }n+\mathcal{O}(1)\quad \mathrm{ and }\quad \mathbb{C}\textsl{ov}\hspace{.03cm}[\mathbf{X}_n]=\mathbf{\Sigma}n+\mathcal{O}(1),$$ where $$\boldsymbol{\mu }=-\frac{\rho _{\mathbf{u}}(\mathbf{1})}{\rho (\mathbf{1})}\quad \mathrm{ and } \quad\boldsymbol{\Sigma } =-\frac{\rho _{\mathbf{uu}}(\mathbf{1})}{\rho (\mathbf{1})}+\boldsymbol{\mu}\boldsymbol{\mu}^T +\textsl{diag}( \boldsymbol{\mu } ).$$
\[quasi\]
Finally, a pair of combinatorial classes with generating functions $(y(z),g(z))$ is called *subcritical* if $y(z)=g(y(z))$ and $y(\rho _y )<\rho _g $, where $\rho _y$ and $\rho _g $ are the radius of convergence of $y,g$, respectively.
Encoding 2-connected patterns in polygon dissections {#maain}
====================================================
Let $\Delta =\{ \delta _1,\delta _2,...,\delta _m \} $ be a set of 2-connected patterns and let $D_{\Delta }(z,\mathbf{u})$ be a multivariate generating function, where the coefficient of $z^nu_1^{n_1}\cdots u_m^{n_m}$ is the number of polygon dissections in $\mathcal{D}$ that have $n$ vertices and $n_i$ subgraph occurrences of the pattern $\delta _i$.[^6] In the construction of Figure \[fig:1\] and the corresponding Equation , observe that the encoding of subgraphs of type $\delta _i$ is not straightforward, since they do not behave additively as parameters. The aim in this section is to prove, for any set $\Delta $, an explicit combinatorial construction for $\mathcal{D}$ that allows this encoding. The approach we follow is to partition the class $\mathcal{D}$ into smaller combinatorial classes and build a symbolic system with them and $\mathcal{D}$, in which we can handle uniformly the appearance of such patterns. The resulting system uses only the operations of addition and cartesian product, and thus settles formally the algebraic nature of $D_{\Delta }(z,\mathbf{u})$. For clarity of presentation, we first work with dissections $\bar{\mathcal{D}}$ that miss one of the two vertices of the root-edge, hence $[z^n]D(z,\mathbf{1})=[z^{n-1}]\bar{D}(z,\mathbf{1})$, in order to avoid the denominators of Equation . We proceed by defining the auxiliary combinatorial classes $\bar{\mathcal{D}}_{\circ },\bar{\mathcal{D}}_{\nu _1},...,\bar{\mathcal{D}}_{\nu _k}$ that give us the required partition.
Since $\delta _i$ are subgraphs of polygon dissections and 2-connected, they are themselves isomorphic to polygon dissections. Let $H_{\Delta }$ be the length of the maximum Hamilton cycle over all $\delta _i$. In order to encode the appearances of $\delta _i$ in $\bar{\mathcal{D}}$, we need to control the way the dissections are glued recursively, as suggested in Figure \[fig:1\]. In fact, we need to control the construction until the root polygon is of length at most $H_{\Delta }$: no new copies of $\delta _i$ are created when already made dissections are glued around a big root polygon. Thus, we consider as *small*, respectively *big*, the polygons that are equal to or smaller than, respectively larger than, an $H_{\Delta }$-gon and denote by $\bar{\mathcal{D}}_{\circ }$ the class which contains all dissections in $\bar{\mathcal{D}}$ that have a big root polygon, plus the edge $e=\{v_1,v_2\}$. We are now able to give the following definition:
A polygon dissection is called a *composite root* with respect to a set of 2-connected patterns $\Delta =\{ \delta _1,\delta _2,...,\delta _m \} $ if the following two conditions hold:
1. It consists only of small polygons.
2. Let $F$ be a face of the composite root that shares an edge with the unbounded face and let an edge $e_1\in F$ that is not an outer edge. Then, $e_1$ is connected to the root-edge with a simple path of adjacent polygons that constitutes a dissection of size less than $H_{\Delta }$.
Observe that there is a finite number of composite roots. They are denoted by $\nu _j$, where $j$ refers to some arbitrary ordering among them. Alternatively, we identify a composite root with a tuple of indices $i_1[i_2]$, where the first index $i_1$ is the size of its root polygon and the second index $i_2$ is its ordering among all the other composite roots with the with the same root polygon of size $i_1$, according to some arbitrary ordering among them (see, for instance, Figure \[new444\] or \[fig:2\]).
Let $A,B$ be polygon dissections. $B$ will be called an *extension* of $A$ if $A$ is an induced subgraph of $B$ preserving the root-edge, i.e., one can obtain $A$ from $B$ by a sequence of vertex deletions, excluding the vertices of the root-edge, and renumbering the vertices according to their final position with respect to the root-edge. For instance, the dissections $3[3]$ and $3[8]$ in Figure \[new444\] are extensions of $3[1]$, but $3[9]$ is not an extension of $4[1]$. A composite root is called *maximal* if there is no composite root that is an extension of it.
![The composite roots, when $\Delta =\{\delta _1\}$ and $\delta _1$ is a 4-cycle. The roots $3[4]$, $3[7]$, $3[8]$, $3[9]$, and $4[1]$ are the only maximal ones and an edge is blue if it is outer in some maximal extension.[]{data-label="new444"}](new444)
We associate to each one of the composite roots $\nu _j$ the combinatorial class $\bar{\mathcal{D}}_{\nu _j},$ which corresponds to polygon dissections that are extensions of the composite root $\nu _j$ and satisfy the following condition, called *Condition (I)*:
1. If an outer edge of $\nu _j$ is inner in the maximal extensions of $\nu _j$, then only elements of $\bar{\mathcal{D}}_{\circ }$ are attached on it.
The classes $ \bar{\mathcal{D}}_{\circ},\bar{\mathcal{D}}_{\nu _j}$ partition $\bar{\mathcal{D}}$. Moreover, each of the classes $\bar{ \mathcal{D}}, \bar{\mathcal{D}}_{\circ},\bar{\mathcal{D}}_{\nu _j}$ can be constructed in a non-trivial admissible way by the classes $\bar{ \mathcal{D}}, \bar{\mathcal{D}}_{\circ},\bar{\mathcal{D}}_{\nu _j}$.
Condition (I) forces the $\bar{\mathcal{D}}_{\nu _j}$ classes to be disjoint: if $p _i\in \bar{\mathcal{D}}_{\nu _i}, p _j\in \bar{\mathcal{D}}_{\nu _j}$, $i\neq j$, then $p _i\neq p _j$, since their maximal composite roots differ in at least one small polygon. Moreover, any object in $\bar{\mathcal{D}}$ with small root polygon and maximal composite root $\nu _j$ belongs in $\bar{\mathcal{D}}_{\nu _j}$. Since the class $\bar{\mathcal{D}}_{\circ }$ contains the edge graph and all dissections with big root polygon, the $\bar{\mathcal{D}}_{\circ},\bar{\mathcal{D}}_{\nu _j}$ classes indeed form a partition of $\bar{\mathcal{D}}$. It then holds that
$$\bar{\mathcal{D}}=\bar{\mathcal{D}}_{\circ }\text{ }\bigcup _{j=1 }^m\bar{\mathcal{D}}_{\nu _j } \quad\text{ and }\quad\bar{\mathcal{D}}_{\circ }=\{ \includegraphics[scale=.6]{edge}\}\bigcup _{i > H_{\Delta }}\underbrace{\bar{\mathcal{D}}\times \ldots
\times\bar{\mathcal{D}}}_{i \text{ times}}. \label{union}$$
For $p\in \bar{\mathcal{D}}_{\nu _j}$, $p$ is decomposed uniquely into its maximal composite root $ \nu _j$ and a sequence of objects from the classes $\bar{\mathcal{D}}_{\circ},\bar{\mathcal{D}}_{\nu _j}$ that respects Condition (I). In particular, if an edge of $\nu _j$ is outer in its maximal extensions, then objects from any class are attached. Else, only members of $\bar{\mathcal{D}}_{\circ }$ are attached. Let $t$ be the number of such outer edges in $\nu _{j}$, $s$ be the number of all outer edges, and $\mathbf{c}\in
\{\circ ,\nu _1,\nu _2,...,\nu _m\} ^t $. Then it holds that $$\bar{\mathcal{D}}_{{\nu _j}}=\bigcup _{\mathbf{c}} \underbrace{\bar{\mathcal{D}}_{\circ } \times \ldots \times\bar{\mathcal{D}}_{\circ }}_{s-t-1 \text{ times}}\times \bar{\mathcal{D}}_{c_1}\times ...\times \bar{\mathcal{D}}_{c _t}\label{unionplus}.$$
![The decomposition of an element in $\bar{\mathcal{D}}_{\circ }$ and an element in $\bar{\mathcal{D}}_{3[4]}$, when $\Delta =\{\delta _1\}$ and $\delta _1$ is a 4-cycle. See Figure \[new444\] for the class indices.[]{data-label="dec1"}](dec1)
\[main\] Let $\Delta =\{ \delta _1,\delta _2,...,\delta _m \} $ be a set of 2-connected patterns and $\nu _1 ,..., \nu _k$ the corresponding composite roots. The generating functions $\bar{D}(z,\mathbf{u}), \bar{D}_{\circ }(z,\mathbf{u}),\bar{D}_{\nu_1}(z,\mathbf{u}),...,\bar{D}_{\nu _k}(z,\mathbf{u}),$ where $\mathbf{u}=(u_1,...,u_m),$ satisfy a computable system of the form: $$\left\{
\begin{array}{lll}
y&=&r(z,u_1,...,u_m,y,y_{\circ },y_{\nu _1},...,y_{\nu _k}),\\
y_{\circ }&=&r_{ \circ}(z,u_1,...,u_m,y,y_{\circ },y_{\nu _1},...,y_{\nu _k}),\\
y_{\nu _1} &=& r_{\nu _1}(z,u_1,...,u_m,y,y_{\circ },y_{\nu _1},...,y_{\nu _k}),\\
\vdots & & \vdots \\
y_{\nu _k} &=& r_{\nu _k}(z,u_1,...,u_m,y,y_{\circ },y_{\nu _1},...,y_{\nu _k}).
\end{array}
\right.$$ which is non-linear in $y,y_{\circ },y_{\nu _j}$ and where each $r_j$ is a $\mathbb{Q}$-rational and analytic function around zero, with non-negative coefficients. Moreover, $r_{\circ }(z,\mathbf{0})\neq 0$ and the system is strongly connected.
The parameters $\delta _i$ are additive in the symbolic Equations , so their translation to multivariate generating functions depending on $z$ and $u$ is immediate: $$\bar{D}=\bar{D}_{\circ }+\sum _{i=1}^m \bar{D}_{\nu _i},
\quad\quad \bar{D}_{\circ }=z+\sum _{i> H_{\Delta }}\bar{D}^i\Rightarrow \bar{D}_{\circ }=z+\frac{\bar{D}^{H_{\Delta }}}{1-\bar{D}}.\label{eq:1234}$$
The parameters $\delta _i$ are not additive in the symbolic Equation , since new copies of them might occur after the attachment of objects in $\bar{\mathcal{D}}_{{\nu _j}}$ around the composite root. However, any new copies occur locally, in the interactions between $\nu _j$ and subsets of $\mathbf{c}$. This is a fixed number $p_{\mathbf{c}}^{ji}$ for every $\mathbf{c}$ and $\delta _i$. Thus, Equation is translated in the following way: $$\bar{D}_{\nu _j}=\sum _{\mathbf{c} } \bar{D}_{\circ }^{s-t-1}\bar{D}_{c _1}...\bar{D}_{c _t}u_1^{p_{\mathbf{c}}^{j1}} ...u_m^{p_{\mathbf{c}}^{jm}}.$$
The emerging system is indeed strongly connected: all $\bar{D}_{\nu _j}$ are connected to $\bar{D}_{\circ }$, which connects to $\bar{D}$. The rest of the stated properties are immediate.
By Equation (\[eq:1234\]), we obtain $\bar{D}_{\circ}=\bar{D}^{H_{\Delta }}+\bar{D}\bar{D}_{\circ}-z\bar{D}+z$. Also, one can substitute the $y_{\circ },y_{\nu _j}$ variables in the right part of $y$’s equation with their equivalent expressions. Thus, systems of Theorem \[main\] can be turned to *proper algebraic*, i.e., in the right part there is no constant term or linear term $y,y_i$. Then, we can argue that $\bar{D}(z,\mathbf{u})$ also satisfies some computable polynomial equation $p(z,\mathbf{y},\mathbf{u})=0$ (see [@Panholzer]). Consequently, also $D(z,\mathbf{u})$ is algebraic, as well as $D(z,\mathbf{0})$, i.e., the generating function of polygon dissections that avoid all patterns in $\Delta $ as subgraphs.
The generating function $D(z,\mathbf{u})$ is algebraic and the defining polynomial is computable. The generating function of polygon dissections that avoid all $\delta $-patterns as subgraphs, $D(z,\mathbf{0})$, is likewise algebraic.\[cor\]
Note that the systems resulting from Theorem \[main\] are large with respect to $H_{\Delta }$. In particular, any combination of at most $H_{\Delta }-2$ small polygons around a root polygon of size $H_{\Delta }-1$ will constitute a composite root. These are $(H_{\Delta }-1)^{H_{\Delta }-2}$, since there are $H_{\Delta }-2$ available edges and $H_{\Delta }-1$ choices, when considering also the empty choice. However, when $H_{\Delta }$ is small, one can find ad hoc arguments to make the systems manageable; see for instance Section \[app\].
\[random\] Let $\Delta =\{\delta _1,...,\delta _m\}$ be a set of 2-connected patterns. Let $\Omega _n$ be the set of polygon dissections of size $n$ and $X_n:\Omega _n \rightarrow \mathbb{Z}_{\geq 0}^m$ be a vector of random variables $X_{\delta _1},...,X_{\delta _m}$ in $\Omega _n$, such that $X_{\delta _i}(\omega )$ is the number of $\delta _i$ patterns in $\omega \in\Omega _n$. Then, $\mathbf{X}_n$ satisfies a central limit theorem $$\frac{1}{\sqrt{n}}(\mathbf{X}_n-\mathbb{E}[\mathbf{X}_n])\xrightarrow[]{d} N(\mathbf{0},\mathbf{\Sigma })$$ with $$\mathbb{E}[\mathbf{X}_n]=\boldsymbol{\mu }n+\mathcal{O}(1) \text{ and } \mathbb{C}\textsl{ov}\hspace{.03cm}[\mathbf{X}_n]=\mathbf{\Sigma}n+\mathcal{O}(1),$$ where $\boldsymbol{\mu }$ and $\mathbf{\Sigma}$ are computable.
Any system resulting from Theorem \[main\], $\mathbf{y}-\mathbf{r}(\mathbf{y},\mathbf{z},\mathbf{u})=\mathbf{0}$, admits a non-negative power series solution $\mathbf{y}(z,\mathbf{u})$ around zero and $\mathbf{1}$ by construction. This is also unique by the implicit function theorem, since $$\det (\mathbf{I}-\mathbf{r}_{\mathbf{y}}(\mathbf{y},z,\mathbf{u}))| _{(\mathbf{y},z)=\mathbf{0},\mathbf{u}=\mathbf{1}}=1$$ for every set $\Delta $. Thus, the defining system of $\bar{D}$ is always well defined.
By construction, the system is also strongly connected. Consequently, the matrix $\mathbf{r}_{\mathbf{y}}$ is non-negative and irreducible in $\mathbb{R}^+$. It is known [@minc1988nonnegative] that non-negative irreducible matrices have a unique dominant eigenvalue $\lambda $ that is positive and strictly increasing with respect to the entries of the matrix. Let $\rho $ be the radius of convergence of $\bar{D}(z,\mathbf{1})$. For $z< \rho $, it holds that $ \lambda (\mathbf{r}_{\mathbf{y}} (z, \mathbf{y}(z),\mathbf{1} ))<1$: if this was not the case, then $\bar{D}(z,\mathbf{1})$’s radius of convergence would be smaller, by Theorem \[drmota\]. The value $\bar{D} (\rho ,\mathbf{1})$ is finite, since $\bar{D}$ is algebraic. Consequently, the characteristic system always has the minimal solution $(\rho ,y(\rho ,\mathbf{1}))$. Moreover, it is true that $[z^n]\bar{D}(z,\mathbf{1})>0$.
The result can now be obtained as direct consequence of Theorems \[main\], \[drmota\], and \[quasi\].
Applications {#app}
=============
In this section, we give examples and applications of Theorem \[main\]. The applications concern the combinatorial classes of polygon dissections and outerplanar graphs and they are of two different kinds: computation of limit laws for 2-connected parameters $\delta _i$ and asymptotic enumeration of these classes, when the patterns $\delta _i$ are forbidden as subgraphs. For clarity, we give the defining equations for $\bar{D}$ and not for $D$, but the final computations will be done in terms of $D$. The equation analysis process is similar to the one in [@bodirsky2007enumeration].
Extraction of limit laws
------------------------
### Encoding $3$-cycles {#encoding-3-cycles .unnumbered}
The only composite root is the triangle, denoted by $3[1]$. Thus, the defining system of $\bar{D}$ is the following: $$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3[1]},\\ \bar{D}_{\circ } &=& z+\frac{\bar{D}^{3}}{1-\bar{D}},\\ \bar{D}_{3[1]} &=& u(\bar{D}_{\circ }+\bar{D}_{3[1]})^2.\end{aligned}$$ The latter is equivalent to the following polynomial system (notice that in this form it is not non-negative): $$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3[1]},\\ \bar{D}_{\circ } &=& \bar{D}^3+\bar{D}\bar{D}_{\circ }-z\bar{D}+z,\\ \bar{D}_{3[1]} &=& u(\bar{D}_{\circ }+\bar{D}_{3[1]})^2.\end{aligned}$$ By observing that $\bar{D}_{3[1]}=u\bar{D}^2$ and $\bar{D}\bar{D}_0=\bar{D}(\bar{D}-\bar{D}_{3[1]})=\bar{D}^2(1-u\bar{D})$, we obtain $$\bar{D}=\bar{D}^3(1-u)+\bar{D}^2(1+u)-\bar{D}z+z.$$
### Encoding $4$-cycles {#encoding-4-cycles .unnumbered}
The composite roots are all the dissections in Figure \[new444\]. From now on, we write $\bar{D}_i$ for the sum $\sum _j \bar{D}_{i[j]}$. Also, when $m$ equations are the same and correspond to the same root polygon with $n$ sides, we write $ \bar{D}_{n[i_1,...,i_m]}$ or $ \bar{D}_{n[i_1-i_m]}$, for shortness.
$$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{ 3}+\bar{D}_{ 4},\\ \bar{D}_{\circ } &=& \bar{D}^4+\bar{D}\bar{D}_{\circ }-z\bar{D}+z, \\ \bar{D}_{3[1]} &=& \bar{D}_{\circ }^2 ,\\ \bar{D}_{3[2,3] } &=& u\bar{D}_{\circ }(\bar{D}_{\circ }+u\bar{D}_{3}+\bar{D}_{4[1]})^2 ,\end{aligned}$$
$$\begin{aligned}
\bar{D}_{3[4]} &=& u^2(\bar{D}_{\circ } +u \bar{D}_{3} +\bar{D}_{4[1]})^4 , \\ \bar{D}_{3[5,6]} &=& u\bar{D}_{\circ }\bar{D}^3 , \\ \bar{D}_{3[7]} &=& u^2\bar{D}^6 , \\ \bar{D}_{3[8,9]} &=& u^2\bar{D}^3(\bar{D}_{\circ }+u\bar{D}_3+\bar{D}_{4[1]})^2 , \\ \bar{D}_{4[1]} &=& u\bar{D}^3.\end{aligned}$$
Notice that the term $(\bar{D}_{\circ } +u \bar{D}_{3} +\bar{D}_4)^2$ is equal to $\bar{D}_3$. So, the system is equivalent to:
$$\bar{D}=\bar{D}_{\circ }+\bar{D}_3+\bar{D}_{4[1]} ,\quad \bar{D}_{\circ }= \bar{D}^4+\bar{D}\bar{D}_{\circ }-z\bar{D}+z ,\quad \bar{D}_3=(\bar{D}_{\circ }+u\bar{D}_3+\bar{D}_{4[1]} )^2 \quad \bar{D}_{4[1]} =u\bar{D}^3.$$
We now use the previous systems, encoding $3$ and $4$-cycles, to obtain the following theorem:
\[random2\] The number of appearances $X_n$ of 3-cycles and 4-cycles in polygon dissections and outerplanar graphs of size $n$ follows a central limit theorem as in \[1st\], where the mean and variance are asymptotically linear. The constants are the following, in their exact values for dissections and in approximation for outerplanar graphs:
-------------------------------------------------------------------------------------------------------------------------------------------------------
Parameter $\mu $ $\sigma ^2 $ $\mu $ $\sigma ^2 $
----------- ------------------------------------------------- ---------------------------------------------------------------- --------- --------------
3-cycles $\frac{1}{2}$ ${\frac {-13+9\,\sqrt {2}}{-12+8\,\sqrt {2}}}\approx 0.39644$ 0.34793 0.40737
4-cycles $ {\frac {-30+21\,\sqrt {2}}{-12+8\,\sqrt {2}}} $\,{\frac {-24216+17123\,\sqrt {2}}{-32 \left( -3+2\,\sqrt {2} 0.33705 0.36145
\right) ^{2}}}
\approx 0.43933$ \approx 0.44710$
-------------------------------------------------------------------------------------------------------------------------------------------------------
The central limit theorem is obtained from Theorem \[random\]. We present an outline of how to get the exact constants in both cases, which can be replicated for any system derived from Theorem \[main\]. For specific steps of the computations, see Section \[comp\]
In both cases, $D$ has a singular expansion of the form $$g(z,u)-h(z,u)\sqrt{1-\frac{z}{r (u)}},$$ that satisfies the requirements of Theorem \[quasi\]. This can be obtained by the same reasoning as in Theorem \[random\].
The value $r(1)$ can be computed using the *discriminant* of $D$’s defining polynomial $p(y,z,u)$ (see [@flajolet2009analytic Ch.VII]), $\mathrm{disc}(z,u)$. In this case, $r(1)=3-2\sqrt{2}$, which is known from [@flajolet1999analytic]. Then, we also find the values $r'(1), r''(1)$, by consecutively differentiating $\mathrm{disc}(r(u),u)$ with respect to $u$. With these values, we compute the constants required for the mean and variance according to Theorem \[quasi\] and obtain the indicated numbers.
In order to pass to labelled 2-connected, connected, and then general outerplanar graphs, we use the multivariate analogues of the equations in Table \[table:2\], i.e. $$B'(z,u)=\frac{1}{2z}D(z,u)+\frac{z}{2}, \quad zC'(z,u)=z\exp (B'(zC'(z,u),u)), \quad G(z,u)=\exp (C(z,u)),$$ where the derivatives are taken with respect to $z$. Let $y$ denote $zC'(z,1)$ and consider the characteristic system: $$y-z\exp (B'(y,1))=0$$ $$1-z\exp (B'(y,1))B''(y,1)=0$$ This has indeed a minimal positive solution $(\tau ,z_0)$, since the outerplanar graph class belongs to the subcritical family of graphs [@subcritical], i.e. $z_0 C'(z_0,1)<r(1)$, where $z_0$ is the radius of convergence of $C'(z,1)$ and $r(1)$ the one of $B'(z,1)$. Moreover, the system satisfies $1-yB''(y,1)=0$. Solving for $y$, we find the value $\tau $ and then $z_0=\tau \exp (-B'(\tau ,1))$.
We can now apply Theorem \[drmota\] and get a singular expansion around $z_0$ for $C'$, with critical exponent $1/2$. Moreover, the point $z_0$ is the only singularity on the radius of convergence of $C'$ and there exists an analytic function $\rho (u)$ around $u=1$ that gives the unique smallest singularity of $C'$ when $u$ is close to $1$; in particular, $\rho (1)=z_0$. Then, also $\tau (u)$ is an analytic function close to 1, where $\tau (u)=\rho (u)C'(\rho (u),u)$.
As in [@bodirsky2007enumeration], if $\Psi (y,u)$ is an analytic function such that $\Psi (y,u)=y\exp (-B'(y,u))$, then $\rho (u)=\Psi (\tau (u),u)$ and it holds that $$\rho '(u)=\frac{\partial \Psi }{\partial u}(\tau (u),u)\quad \text{ and }\quad \rho ''(u)=\frac{\partial ^2\Psi }{\partial y\partial u}(\tau (u),u)\tau '(u)+\frac{\partial ^2\Psi}{\partial u^2}(\tau (u),u).\label{coomp}$$
The functions $C$ and $G$ have the same singularity function $\rho (u)$ as $C'$, but the critical exponent of their expansion on $\rho (u)$ is $3/2$ (see the analysis in [@bodirsky2007enumeration] for details). We can thus apply Theorem \[quasi\], after computing the relevant constants. The value $\tau '(1)$ can be computed through the relation $\tau (u)B''(\tau (u),u)=1$.
In the case of outerplanar graphs, the limit laws are expected from [@subcritical]. However, in [@subcritical] there is no constructive way to compute the relevant constants. This is a contribution of this work, that offers specific defining equations for the function $B'$.
### Computations {#comp}
The following were performed in the computational software `Maple`.
Parameter: 3-cycles {#parameter-3-cycles .unnumbered}
-------------------
For dissections, the defining polynomial $p_3(D,z,u)$ is the following:
$$p_3=-u{D}^{3}+u{D}^{2}z-D{z}^{3}+{z}^{4}+{D}^{3}+{D}^{2}z-D{z}^{2}$$
Its discriminant with respect to $D$, $\mathrm{disc}(z,u)$, is equal to $$-{z}^{6} \left( 4\,{u}^{3}z+8\,{u}^{2}{z}^{2}+4\,u
{z}^{3}-8\,{u}^{2}z-44\,u{z}^{2}-4\,{z}^{3}-{u}^{2}+20\,uz+32\,{z}^{2}
+2\,u+8\,z-5 \right).$$ From this, we retrieve the root $r(1)=3-2\sqrt{2}$ . By setting $z=r(u)$ and differentiating with respect to $u$ in $\mathrm{disc}(r(u),u)$, we also retrieve $r'(1)=-\frac{3}{2}+\sqrt{2},$ $r''(1)=\frac{3\sqrt{2}}{4}-1$.
By differentiating $p_3$ with respect to $D$, we obtain exact expressions for the derivatives $\frac{\partial D(z,u)}{\partial u}$ and $\frac{\partial D(z,u)}{\partial z}$: $$\frac{\partial D(z,u)}{\partial u}=-{\frac { \left( D \left( z,u \right) \right) ^{2} \left( D \left( z,
u \right) -z \right) }{3\,u \left( D \left( z,u \right) \right) ^{2}-
2\,D \left( z,u \right) uz+{z}^{3}-3\, \left( D \left( z,u \right)
\right) ^{2}-2\,D \left( z,u \right) z+{z}^{2}}}$$
$$\frac{\partial D(z,u)}{\partial z}={\frac {u \left( D \left( z,u \right) \right) ^{2}-3\,D \left( z,u
\right) {z}^{2}+4\,{z}^{3}+ \left( D \left( z,u \right) \right) ^{2}
-2\,D \left( z,u \right) z}{3\,u \left( D \left( z,u \right) \right)
^{2}-2\,D \left( z,u \right) uz+{z}^{3}-3\, \left( D \left( z,u
\right) \right) ^{2}-2\,D \left( z,u \right) z+{z}^{2}}}$$ Then, we write $1-zB''(z,1)=0$ in terms of $D$, using the previous expressions, i.e. $$1+z \left( \,{\frac {D}{2{z}^{2}}}-\,{\frac {-3\,D{z}^{2}+4\,{z}^
{3}+2\,{D}^{2}-2\,Dz}{2z \left( {z}^{3}-4\,Dz+{z}^{2} \right) }}-\frac{1}{2}
\right) =0\label{tau}$$ and solve the system of Equation (\[tau\]) and $p_3(D,z,1)=0$. The values we obtain are $D \approx 0.04709517290,$ $ \tau \approx 0.1707649868$. Then $\rho (1)=\tau {{\rm \exp}{(-\,{\frac {D \left( \tau,1 \right) }{2\tau }}-\frac{\tau}{2})}}\approx 0.1365937336$.
To compute the derivatives of $\Psi (\tau (u),u)$, we write them in terms of $D(\tau (u),u)$. The value $\tau '(1)$ can be found similarly, after writing the equation $\tau (u)B''(\tau (u),u)=1$ in terms of $D(\tau (u),u)$. In particular, we obtain $\tau '(1)\approx -0.849388502$, $\rho '(1)\approx -0.5564505691$ and $\rho ''(1)\approx 0.3078771691$. The final values are computed as indicated in Theorem \[quasi\].
Parameter: 4-cycles {#parameter-4-cycles .unnumbered}
-------------------
The procedure of the computations is the same as in the previous case. We only note that the defining polynomial $p_4(D,z,u)$ is the following: $$\begin{aligned}
p_4 &=& {u}^{4}{z}^{2}{D}^{6}-2\,{u}^{4}z{D}^{7}+{u}^{4}{D}^
{8}+2{u}^{3}{z}^{6}{D}^{3}-4{u}^{3}{z}^{5}{D}^{4}+2{
u}^{3}{z}^{4}{D}^{5}+{u}^{2}{z}^{10}-2\,{u}^{2}{z}^{9}D +\\ & & +{u
}^{2}{z}^{8}{D}^{2}-2\,{u}^{3}{z}^{4}{D}^{4}+4\,{u}^{3}{z}
^{3}{D}^{5}-4\,{u}^{3}{z}^{2}{D}^{6}+6\,{u}^{3}z{D}^
{7}-4\,{u}^{3}{D}^{8}-2\,{u}^{2}{z}^{8}D+\\ & & +4\,{u}^{2}{z}^{7}
{D}^{2} -6\,{u}^{2}{z}^{6}{D}^{3}+10\,{u}^{2}{z}^{5}{D_{{1}
}}^{4}-6\,{u}^{2}{z}^{4}{D}^{5}-2\,u{z}^{10}+4\,u{z}^{9}D-
2\,u{z}^{8}{D}^{2}+{u}^{2}{z}^{6}{D}^{2}- \\ & & -2\,{u}^{2}{z}^{5}
{D}^{3}+3\,{u}^{2}{z}^{4}{D}^{4}-6\,{u}^{2}{z}^{3}{D
}^{5}+5\,{u}^{2}{z}^{2}{D}^{6}-6\,{u}^{2}z{D}^{7}6\,{u}^{
2}{D}^{8}+2\,u{z}^{8}D-4\,u{z}^{7}{D}^{2}+4\,u{z}^{6
}{D}^{3}-\\ & &-8\,u{z}^{5}{D}^{4}+6\,u{z}^{4}{D}^{5}{z}^{
10}-2\,{z}^{9}D+{z}^{8}{D}^{2} +u{z}^{5}{D}^{3}-2\,u{
z}^{4}{D}^{4}+3\,u{z}^{3}{D}^{5}-2\,u{z}^{2}{D}^{6}+
2\,uz{D}^{7}-\\ & & -4\,u{D}^{8}+{z}^{9}-2\,{z}^{8}D+{z}^{7}
{D}^{2}+2\,{z}^{5}{D}^{4}-2\,{z}^{4}{D}^{5} -{z}^{7}D
_{{1}}+2\,{z}^{6}{D}^{2}-{z}^{5}{D}^{3}+{z}^{4}{D}^{
4}-{z}^{3}{D}^{5}+{D}^{8}\end{aligned}$$ and gives the intermediate values $r'(1)\approx -15/2+21\sqrt{2}{4}$, $r'(1)\approx -413/4+2337\sqrt{2}{32}$, $\tau \approx 0.1707649868$, $D(\tau ,1) \approx 0.4709517290$, $\tau '(1)\approx -0.7427876522$.
Restricted classes {#restrict}
------------------
Now we apply the results of Theorem \[main\] in the context of asymptotic enumeration. We give various examples in restricted classes of polygon dissections and outerplanar graphs.
### Avoiding 3 and 4-cycles {#avoiding-3-and-4-cycles .unnumbered}
Using the equations of the previous section, we obtain immediately sets of equations for polygon dissections avoiding $3$ and $4$-cycles. Setting $u=0$ and substituting for $\bar{D}$, we obtain $$\begin{aligned}
\bar{D} =\bar{D}^3+\bar{D}\bar{D}_{\circ }-z\bar{D}+z\end{aligned}$$ for $3$-cycles and $$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3[1]}\\ \bar{D}_{\circ } &=& \bar{D}^4+\bar{D}\bar{D}_{\circ }-z\bar{D}+z, \\ \bar{D}_{3[1]} &=& \bar{D}_{\circ }^2 \end{aligned}$$ for $4$-cycles.
### Avoiding 5-cycles {#avoiding-5-cycles .unnumbered}
In Figure \[fig:2\], there are all the composite roots when $5$-cycles are avoided. We obtain the following system, after setting $u=0$ where appropriate.
$$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3}+\bar{D}_{4} \\
\bar{D}_{\circ } &=& \bar{D}^5+\bar{D}\bar{D}_{\circ }-z\bar{D}+z\\
\bar{D}_{3[1]} &=& \bar{D}_{\circ }^2\\
\bar{D}_{3[2,3]} &=& \bar{D}_{\circ }^3\\ \end{aligned}$$
$$\begin{aligned}
\bar{D}_{4[1]} &=& \bar{D}_{\circ }^3 \\
\bar{D}_{4[2,3,4]} &=& \bar{D}_{\circ }^2( \bar{D}_{\circ }+ \bar{D}_{4})^3 \\
\bar{D}_{4[5,6,7]} &=& \bar{D}_{\circ } (\bar{D}_{\circ } +\bar{D}_{4})^6 \\
\bar{D}_{4[8]} &=& (\bar{D}_{\circ } +\bar{D}_{4})^9 \\\end{aligned}$$
The system can be simplified, after observing that $\bar{D}_4=(\bar{D}_{\circ } +\bar{D}_{4})^3$ and $\bar{D}_3=\bar{D}_{\circ }^2(1+2\bar{D}_{\circ })$.
![The composite roots when $5$-cycles are avoided. The dissection $4[8]$ is the only maximal one among them.[]{data-label="fig:2"}](5)
### Avoiding 6-cycles {#avoiding-6-cycles .unnumbered}
In Figure \[fig:4\], there are the composite roots when $6$-cycles are excluded, except for the ones including a $5$-gon. For $\bar{D}_5$, we observe immediately that $\bar{D}_5=(\bar{D}_{\circ }+\bar{D}_{4}+\bar{D}_{5})^4$.
$$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3}+\bar{D}_{4}+\bar{D}_{5} \\
\bar{D}_{\circ } &=& \bar{D}^6+\bar{D}\bar{D}_{\circ }-z\bar{D}+z\\
\bar{D}_{3[1]} &=& \bar{D}_{\circ }^2\\
\bar{D}_{3[2,3]} &=& \bar{D}_{\circ }^3\\
\bar{D}_{3[4-8]} &=& \bar{D}_{\circ }^4\\
\bar{D}_{3[9,10]} &=& \bar{D}_{\circ }(\bar{D}_{\circ }+\bar{D}_5)^3\\
\bar{D}_{3[11]} &=& (\bar{D}_{\circ }+\bar{D}_5)^6\\ \end{aligned}$$
$$\begin{aligned}
\bar{D}_{4[1]} &=& (\bar{D}_{\circ }+\bar{D}_5)^3 \\[1.5pt]
\bar{D}_{4[2,3,4]} &=& \bar{D}_{\circ }^2(\bar{D}_{\circ }+\bar{D}_5)^2 \\[1.5pt]
\bar{D}_{4[5-10]} &=& \bar{D}_{\circ }(\bar{D}_{\circ }+\bar{D}_5)^5 \\[1.5pt]
\bar{D}_{4[11,12,13]} &=& (\bar{D}_{\circ } +\bar{D}_{5})^8 \\[1.5pt]
\bar{D}_{5} &=& (\bar{D}_{\circ }+\bar{D}_{4}+\bar{D}_{5})^4 \\[1.5pt]\end{aligned}$$
We can immediately group all the $\bar{D}_{3[i]}$ and $\bar{D}_{4[j]}$ together to form equations for $\bar{D}_3$ and $\bar{D}_4$, respectively.
### Avoiding other patterns
Now we avoid non-cyclic patterns. We analyse the ones induced by the dissections $3[2]$ and $4[2]$ in Figure \[fig:4\], separately and together. We refer to them as *Pattern I* and *Pattern II*, respectively.
For Pattern I, the composite roots are the dissections $3[1,9,10,11] $ and $4[1]$ in Figure \[fig:4\] and the equations are the following:
$$\begin{aligned}
\bar{D} &=& \bar{D}_{\circ }+\bar{D}_{3}+\bar{D}_{4[1]}\\ \bar{D}_{\circ } &=& \bar{D}^4+\bar{D}\bar{D}_{\circ }-z\bar{D}+z\\ \bar{D}_{4[1]} &=& (\bar{D}_{\circ }+\bar{D}_{3}+\bar{D}_{4[1]})^3\\ \end{aligned}$$
$$\begin{aligned}
\bar{D}_{3} &=& \bar{D}_{\circ }^2\\
\bar{D}_{3[9,10]} &=&( \bar{D}_{\circ }+ \bar{D}_{3}+ \bar{D}_{4[1]})^3 \bar{D}_{\circ }\\
\bar{D}_{3[11]} &=& ( \bar{D}_{\circ }+ \bar{D}_{3}+ \bar{D}_{4[1]})^6\\
\end{aligned}$$
For Pattern II, the composite roots are the ones in Figure \[fig:2\] and the ones containing some $5$-gon. The equations are the same as in the $5$-cycle case, with the following differences: Now $\bar{D}= \bar{D}_{\circ }+\bar{D}_{3}+\bar{D}_{4}+\bar{D}_5$. Also, in all the equations apart from the one of $\bar{D}$ and $\bar{D}_{\circ }$, we substitute $\bar{D}_{\circ }$ for $\bar{D}_{\circ }+\bar{D}_5$. The equation $\bar{D}_5=\bar{D}^4$ must be added as well.
![The composite roots when $6$-cycles are excluded, except for the ones including a $5$-gon.[]{data-label="fig:4"}](61)
For Patterns I and II, the composite roots are the dissections $3[1],4[1-8]$ in Figure \[fig:2\] and the ones containing some $5$-gon. The equations are the same as when avoiding Pattern II, only now $\bar{D}_{3[2]}$ and $\bar{D}_{3[3]}$ are omitted.
\[enumm\] Let $\mathcal{D}, \mathcal{G}$ be the classes of dissections and outerplanar graphs avoiding a set of 2-connected patterns $\Delta =\{\delta _1 ,...,\delta _m\}$, respectively. Then, $\mathcal{D}$ has asymptotic growth of the form $$\alpha _n\sim \frac{\alpha}{\Gamma (-\frac{1}{2})}\cdot n^{-3/2} \cdot r ^{-n}$$ and $\mathcal{G}$ has asymptotic growth of the form $$g_n\sim \frac{g}{\Gamma (-\frac{3}{2})} \cdot n^{-5/2} \cdot\rho ^{-n}\cdot n!$$ where both $\alpha ,g$ are computable constants. In Table \[table:3\], there are approximations of $\alpha , g$ for various choices of $\Delta $.
We apply Theorem \[drmota\] with the same reasoning as in Theorem \[random\] and, in the end, obtain singular expansions with singular exponents $1/2$ and $3/2$ for $\mathcal{D}$ and $\mathcal{G}$, respectively. Then, the types of asymptotic growth can be obtained from the transfer principles of singularity analysis.
It is true that $g=\tau (\log \rho -\log (\tau )+1)+B(\tau )$ [@bodirsky2007enumeration]. The value $B(\tau )$ can be approximated from the systems in this work. For details on the computations, see Section \[computations\].
Restriction $r $ $r^{-1}$ $\alpha $ $\rho $ $\rho ^{-1}$ $g$
-------------- --------- ---------- ----------- --------- -------------- ---------
3-cycles 0.29336 3.40869 0.02330 0.20836 4.79916 0.01578
4-cycles 0.26488 3.77515 0.02177 0.18919 5.28562 0.01462
5-cycles 0.25383 3.93949 0.02217 0.18045 5.54143 0.01514
6-cycles 0.24835 4.02657 0.02321 0.17510 5.71082 0.01630
pattern I 0.20867 4.79214 0.01592 0.15895 6.29100 0.01050
pattern II 0.22416 4.46098 0.01856 0.16608 6.02092 0.01195
pattern I&II 0.24332 4.10977 0.01987 0.17751 5.63345 0.01351
: The constants for the asymptotic growth of restricted polygon dissections and outerplanar graphs, respectively.[]{data-label="table:3"}
### Computations {#computations}
We will use the notation from the proof of Theorem \[random\].
In the level of dissections, there is no computational difficulty, since in all cases the constants can be computed either through the defining equation of $\bar{D}$, or using the characteristic system. In all cases, the main singularity can be found by solving directly the characteristic system, while $\alpha $ can be found by substituting the $\bar{D}_i$ variables in the system by their singular expansions and solve the system with respect to the undetermined coefficients.
Moving to the connected and general level, some values are harder to compute. In particular, the computation of $B(\tau )$ is not easily accessible through our implicit function setting, while also $\tau $ is hard to compute when the value $H_{\Delta }$ grows and the defining equation for $d$ becomes either too big or too hard to compute. For instance, the defining polynomial for the 5-cycle case has degree 45 with respect to $D$ and 54 with respect to $z$, while the polynomial for the 6-cycle case was not retrieved in a reasonable amount of time, i.e. in half hour. For our purposes, we computed an approximation for both values $\tau ,B (\tau ) $, using the first 700 terms of the power series expansion of $B$. The expansion was extracted from the one of $\bar{D}$, which was found by iterating the defining system in `Maple`. The results are displayed in Table \[table:3\].
Acknowledgements
================
This research was funded under an FPI grant from the MINECO research project MTM2015-67304-PI. The author was also partially funded by the Barcelona Graduate School of Mathematics, funded by Maria de Maetzu research grant MDM-2014-0445. The author is grateful to Prof. Juanjo Rué for posing the problem and for making valuable comments on the draft. The author is also grateful to Prof. Dimitrios M. Thilikos who made this research possible. The anonymous referees are warmly thanked for their remarks that improved significantly the initial version of this work.
[10]{}
Daniel Birmajer, Juan B Gil, and Michael D Weiner. Colored partitions of a convex polygon by noncrossing diagonals. , 340(4):563–571, 2017.
Manuel Bodirsky, Omer Gim[é]{}nez, Mihyun Kang, and Marc Noy. Enumeration and limit laws for series–parallel graphs. , 28(8):2091–2105, 2007.
Arthur Cayley. On the partition of a polygon. , 22:237–262, 1891.
Guillaume Chapuy, [É]{}ric Fusy, Omer Gim[é]{}nez, Bojan Mohar, and Marc Noy. Asymptotic enumeration and limit laws for graphs of fixed genus. , 118(3):748–777, 2011.
Reinhard Diestel. Graph theory, volume 173 of. , page 5, 2010.
Michael Drmota. Systems of functional equations. , 10(1-2):103–124, 1997.
Michael Drmota. . Springer Science & Business Media, 2009.
Michael Drmota, Lander Ramos, and Juanjo Ru[é]{}. Subgraph statistics in subcritical graph classes. , 51(4):631–673, 2017.
Philippe Flajolet and Marc Noy. Analytic combinatorics of non-crossing configurations. , 204(1-3):203–229, 1999.
Philippe Flajolet and Robert Sedgewick. . cambridge University press, 2009.
Zhicheng Gao and Nicholas C Wormald. Distribution of subgraphs of random regular graphs. , 32(1):38–48, 2008.
Omer Gim[é]{}nez and Marc Noy. Asymptotic enumeration and limit laws of planar graphs. , 22(2):309–329, 2009.
Omer Gim[é]{}nez, Marc Noy, and Juanjo Ru[é]{}. Graph classes with given 3-connected components: Asymptotic enumeration and random graphs. , 42(4):438–479, 2013.
F. Harary and E.M. Palmer. . Academic Press, 1973.
Micha[ł]{} Karo[ń]{}ski and Andrezj Ruci[ń]{}ski. On the number of strictly balanced subgraphs of a random graph. In [*Graph theory*]{}, pages 79–83. Springer, 1983.
Jeong Han Kim, Benny Sudakov, and Van Vu. Small subgraphs of random regular graphs. , 307(15):1961–1967, 2007.
Colin Mcdiarmid. Random graphs from a minor-closed class. , 18(4):583–599, 2009.
Henryk Minc. . Wiley, 1988.
Alois Panholzer. Gröbner bases and the defining polynomial of a context-free grammar generating function. , 10(1):79–97, January 2005.
Ronald C Read. On general dissections of a polygon. , 18(1):370–388, 1978.
Andrzej Ruci[ń]{}ski. When are small subgraphs of a random graph normally distributed? , 78(1):1–10, 1988.
N. J. A. Sloane. The encyclopedia of integer sequences.
Leonard Smiley. Triangle-free polygon dissections. http://www.math.uaa.alaska.edu/\~smiley/schroed.html, 1999.
Vasiliki Velona. [Asymptotic Enumeration of Outerplanar Graphs under Subgraph Obstructions](http://mpla.math.uoa.gr/en/theses/msc/vasiliki.velona/). Master’s thesis, Graduate Program in Logic, Algorithms and Computation, ($\mu\Pi\lambda\forall$), Greece, 2016.
Appendix
========
Counting series
---------------
Here are the first terms of the counting sequences of all the restricted dissection classes that appear in Section \[restrict\]. The count for triangle-free polygon dissections appears also in [@smiley] and is the sequence `A046736` in [@Sloane_theencyclopedia]. The counts for $3,4,5,6$-cycle free dissections appear also in the author’s [master’s thesis]{} [@velonamsc], where they are derived using ad hoc arguments.
\[count\]
n 3-cycles 4-cycles 5-cycles 6-cycles pattern I pattern II patterns I & II
---- ---------- ---------- ----------- ----------- ------------ ------------ -----------------
2 1 1 1 1 1 1 1
3 0 1 1 1 1 1 1
4 1 0 3 3 1 3 1
5 1 1 0 11 6 1 1
6 4 7 4 0 19 10 10
7 8 22 8 15 64 43 29
8 25 49 65 37 251 181 101
9 64 130 229 85 979 643 283
10 191 468 946 651 3888 2233 1023
11 540 1651 2850 2498 15896 8152 3576
12 1616 5240 9367 10556 65871 31523 13143
13 4785 16485 28068 46112 276225 125776 46502
14 14512 55184 97408 167100 1171838 502449 169221
15 44084 190724 339694 621677 5016697 2001773 618807
16 135545 652359 1276467 2215039 21644451 8002279 2301983
17 418609 2213044 4659990 7524303 94033342 32271594 8576756
18 1302340 7584939 17107629 26414280 410990601 131355333 32169753
19 4070124 26346522 61200635 92579458 1805881012 538125069 121134235
20 12785859 91951596 220323189 332018450 7972740040 2213876868 458881370
: The first terms of the restricted dissection classes of Section \[app\].
[^1]: Department of Mathematics, Universitat Politècnica de Catalunya, Barcelona, Spain.
[^2]: Department of Economics, Universitat Pompeu Fabra, Barcelona, Spain.
[^3]: Email: <[email protected]>
[^4]: From now on, generating functions will be denoted by plane letters and combinatorial classes by calligraphic letters.
[^5]: This convention is followed in an analogous way for all bold characters.
[^6]: From now on, we will refer to $\delta _i$ also as *parameters*, in an abuse of terminology, since we are interested in their number in polygon dissections of size $n$.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
\
Max Planck Institute for Intelligent Systems, Tübingen, Germany\
Technical University of Munich, Germany\
Max Planck Institute for Intelligent Systems, Tübingen, Germany\
Technical University of Munich, Germany\
Max Planck Institute for Intelligent Systems, Tübingen, Germany
bibliography:
- 'references.bib'
title: Planning from Images with Deep Latent Gaussian Process Dynamics
---
model-based reinforcement learning, learning-based control, representation learning, dynamics model learning, transfer learning
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Cosmic strings are stable topological defects that may have been created at a phase transition in the early universe. It is a growing belief that, for a wide range of theoretical models, such strings may be superconducting and carry substantial currents which have important astrophysical and cosmological effects. This paper explores the possibility of generation of a primordial magnetic field by a network of charged–current carrying cosmic strings. The field is created by vorticity, generated in the primordial plasma due to the strings’ motion and gravitational pull. In the case of superconducting strings formed at the breaking of grand unification, it is found that strong magnetic fields of high coherence can be generated in that way. Such fields could account for the observed galactic and intergalactic magnetic fields since they suffice to seed magnetic dynamos on galactic scales.'
author:
- |
[^1]\
\
\
title: '\'
---
Introduction
============
The origin of the observed galactic magnetic fields remains elusive. The magnetic field of the Milky Way and the nearby galaxies is of the order of a $\mu Gauss$ and is coherent over $kpc$ scales. In the Milky Way the magnetic field is orientated along the spiral arms alternating its direction from one arm to the other. This strongly suggests that the field is supported by a dynamo mechanism which arranges it along the spiral density waves [@parker][@dwaves]. Similar evidence for a dynamo mechanism comes from galaxies at a stage of rapid star formation (so called star-burst galaxies [@kron]) where the magnetic flux needs to be amplified by a factor of to account for the observations [@parker].
It is, therefore, a wide belief that the galactic magnetic fields are generated through a dynamo mechanism [@dynamo][@zeld]. A number of mean field theory dynamo models exist in the literature, the most popular of which is the well known $\alpha-\Omega$ dynamo. The basic idea of the dynamo mechanism is that a weak seed field could be amplified by the turbulent motion of ionised gas, which follows the differential rotation of the galaxy. The growth of the field is exponential and, thus, its strength can be increased several orders of magnitude in only a few e-foldings of amplification. When the field reaches the equipartition energy () then its growth is suppressed by dynamical back-reaction.
If the time scale of growth of the field is no more than a galactic rotation period , then the field amplification factor since the collapse of the protogalaxy is of the order , given that the total number of galactic rotations is . Thus, the seed field required has to be at least on the comoving scale of a protogalaxy () at the time of gravitational collapse. Since the collapse of the galaxies enhances their frozen-in magnetic field by a factor of (where is the typical mass density of a galaxy and is the current cosmic mass density), the above seed field corresponds to a field of the order of over the comoving scale of . Assuming that the rms field scales as $a^{-2}$ with the expansion of the universe ($a$ is the scale factor of the universe, in the matter era) we find that the magnitude of the seed field at the start of structure formation has to be at least , where is the time of equal matter and radiation energy densities.
The origin and nature of this seed field is still an open question. Many authors have argued that it could be produced by stellar winds and other explosions [@zeld] but such a field would be extremely incoherent over galactic scales. Significant incoherences of the field have been shown [@kuls] to destabilise and destroy the galactic dynamo and, therefore, it seems that coherency is a crucial factor for seeding the galactic magnetic fields. Also, the existing evidence for intergalactic fields [@kron][@rees] led people to believe that the origin of galactic seed fields may be truly primordial.
Most of the attempts to create a primordial magnetic field in the early universe involve either inflation or phase transitions because the creation of a primordial field can occur only in out of thermal equilibrium conditions [@sacha].
At phase transitions primordial magnetic fields can be created on the surface of the bubble walls if the transition is a first order one [@bubbles], or due to stochastic Higgs–field gradients [@higgs]. Unfortunately though, phase transitions occur very early and the causal horizon is of much smaller comoving scale than the protogalactic one (e.g. for the electroweak transition the comoving scale of the horizon is ). Thus, the generated magnetic field is too incoherent to successfully trigger the galactic dynamo.
Due to this fact, inflation has been considered as another option that could increase the coherency of the created magnetic field [@infl]. However, the conformal invariance of electromagnetism suggested that the field would scale as $a^{-2}$ during the inflationary period which, as a result, diminished the field strength to much lower values than the required seed field limit. In order to overcome this problem, additional terms have been introduced in the Lagrangian to break explicitly conformal invariance. Even so, most attempts produced too weak seed fields.
Other attempts to generate an adequate seed field involve string theory cosmology [@stth]. The problem has been pushed even further back in time, into a possible pre-big bang epoch of negative time, where dilaton inflation can take place. These models, although attractive, suffer from lack of understanding of the interface between the dilaton inflation era and the usual, post-big bang radiation era.
Perhaps the most realistic approach to the problem has been the cosmic string scenario, where the magnetic field is generated by vortical motions inside the wakes of cosmic strings [@cosst]. Vortical generation of a magnetic field has been an early idea of Harrison [@hary], who considered the field to be created during the radiation era inside expanding spinning volumes of plasma (eddies). However, Rees has shown that Harrison’s eddies would be unstable and decay with cosmic expansion, whereas irrotational density perturbations (lumps) from curvature fluctuations would grow and dominate [@rees]. Rees suggested a different version of vortical magnetic field growth which is similar to Harrison’s idea but can be applied to a gravitationally bound spinning body.
In the cosmic string scenario vorticity is generated in the wakes of cosmic strings after structure formation begins. Therefore, the vortical eddies are gravitationally bound and do not suffer from instabilities. Matter in the trail of a string is substantially ionised and so the Harrison–Rees mechanism can still operate. The scale of coherency of the generated magnetic field is set by the scale of wiggles on the string and, for wakes created at $t_{eq}$ it can be up to $100\,kpc$. The field strength is of the order of . Thus, cosmic strings are able to generate magnetic fields of enough strength and coherency to seed the galactic dynamo mechanism.
It is not very clear, though, whether stable vortical motions can be generated by the rapid, stochastic, motion of the string wiggles. An alternative mechanism by Avelino and Shellard [@paper] has overcome this problem by considering dynamical friction. In this model, vorticity is generated not by the wiggles but by the strings themselves, which drag matter behind them and introduce circular motions over inter-string scales. The magnetic field obtained though, is weak and can only marginally seed the galactic dynamo. Also its coherency is very high $\sim 100\,M\!pc$, which may be incompatible with intergalactic field observations.
In this paper we employ the mechanism of Avelino and Shellard in the context of superconducting strings. It is an increasing belief that cosmic strings may be generically superconducting and carry substantial currents [@anne]. Charged current–carrying string networks may evolve in a much different way that the usual, non-superconducting case. As shown by Dimopoulos and Davis [@mine], superconducting networks may be more tangled with slower moving strings. In this case we show that the magnetic field generated can be much stronger than in the non-superconducting case and still be coherent over protogalactic scales $\sim\,1\,M\!pc$.
The structure of this paper is as follows. In Section 2 we give a detailed overview of the field theory of string superconductivity in both the bosonic and fermionic case. In Section 3 we deal with the energy–momentum tensor of the superconducting string, which we use to describe the string spacetime and its consequences on particle deflection in Section 4. In Section 5 we calculate the primordial magnetic field generated. First we find the momentum transfer from the string to the plasma and the resulting rotational velocity of the plasma vortical motions. Then, we estimate the generated primordial magnetic field at the time when structure formation begins, taking also into account constraints coming from the observations of the microwave background anisotropy. Finally, in Section 6 we dicuss our results and give our conclusions. Throughout this paper, unless stated otherwise, we use natural units () for which the Planck mass is .
String superconductivity
========================
String superconductivity was initially conceived by Witten [@witt]. There are two types of models that give rise to superconducting strings depending on whether the current carriers are bosons or fermions. However, through a specific formalism, there is exact correspondence between the two cases in most aspects. In what follows, we give a description of bosonic and fermionic superconductivity and develop the general formalism to treat them both.
Bosonic superconductivity
-------------------------
We will describe bosonic superconducting strings in the context of a theory as in [@witt]…[@peter]. This theory involves two scalar fields: The $\overline{U(1)}$ Higgs–field $\phi$, which is responsible for the formation of the vortex and the $U(1)_{em}$ field $\sigma$ which breaks electromagnetism inside the string and turns it superconducting. The Lagrangian density of the theory is,
$${\cal L}=
-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-\frac{1}{4}G_{\mu\nu}G^{\mu\nu}
+ (D_{\mu}\sigma)^{*}D^{\mu}\sigma + (D_{\mu}\phi)^{*}D^{\mu}\phi
- V(\phi,\sigma)
\label{Lbos}$$
where and , and also and with $W_{\mu}$ and $A_{\mu}$ being the gauge fields coupled to the vortex and the photon field respectively ($e$ and $g$ are the relevant gauge couplings). The potential $V(\phi,\sigma)$ is given by,
$$V(\phi,\sigma)=\frac{1}{4}\lambda(\phi^{2}-\eta^{2})^{2}+
\frac{1}{4}\lambda'\sigma^{4}+f\sigma^{2}\phi^{2}-m^{2}\sigma^{2}
\label{V}$$
where are coupling constants and $\eta$ is the energy scale of the string.
Through the above potential it is possible to form a string by breaking $\overline{U(1)}$. For suitable values of parameters the coupling between the scalar fields may force the breaking of electromagnetism inside the string, making it superconducting.
One can rewrite the above potential as,
$$V(\phi,\sigma)=\frac{1}{4}\lambda(\phi^{2}-\eta^{2})^{2}+
\frac{1}{4}\lambda'(\sigma^{2}-\sigma_{0}^{2})^{2}+
f\sigma^{2}\phi^{2}-m^{2}\sigma^{2}$$
where is the expectation value of $\sigma$ inside the string core.
From the above it follows that, in order for $\overline{U(1)}$ to be broken and to form the string we require, which gives the constraint,
$$(\frac{m}{\eta})^{4}<\frac{\lambda\lambda'}{4}
\label{A}$$
Furthermore, in order for electromagnetism not to be also broken outside the string we require that the unbroken state is a global minimum of $V$, i.e. that which suggests that,
$$\frac{m^{2}}{f\eta^{2}}<1
\label{B}$$
In order for electromagnetism to be broken inside the string we require that the minimum of the potential in the core corresponds to a non-zero value of $\sigma$, i.e. which gives,
$$m^{2}>0$$
Finally, in order for the $\sigma$-condensate to be contained inside the string we require,
$$m_{\phi}^{-1}>(\sqrt{\lambda'}\,\sigma_{0})^{-1}\Rightarrow
(\frac{m}{\eta})^{2}>\frac{\lambda}{2}
\label{C}$$
where is the mass of the Higgs particle. The mass of the $\sigma$ particle is easily found as,
$$m_{\sigma}^{2}=f\eta^{2}-m^{2}>0$$
where we have also used (\[B\]).
By combining (\[A\]) and (\[B\]) we obtain,
$$\lambda<\frac{4}{\lambda}(\frac{m}{\eta})^{4}<\lambda'$$
Under the above conditions the theory admits a vortex solution with electromagnetism broken in its core. The $\sigma$ field may be parametrised as,
$$\sigma=\sigma(r)\,e^{i\psi(z,t)}
\label{sigma}$$
where we have assumed that the vortex core lies on the $z$-axis. The phase field $\psi$ may vary randomly along the string and wind up in a non-trivial way. This winding gives rise to the string current.
If one defines the current as, it is straightforward to show that the string current is, . Integrating over the string core gives,
$$J_{a}=2Ke\,(\partial_{a}\psi+eA_{a})
\label{J}$$
where the constant $K$ is given by,
$$K\equiv\int d^{2}r\,\sigma^{2}\simeq\frac{1}{\lambda'}
\label{K}$$
Varying (\[Lbos\]) with respect to $\phi$,$\sigma$,$W_{\mu}$,$A_{\mu}$ and $\psi$ respectively, gives the following equations,
$$\begin{aligned}
& & \Box\phi-gW^{\mu}W_{\mu}\phi-
\frac{1}{2}\lambda(\phi^{2}-\eta^{2})\phi-f\sigma^{2}\phi=0\\
& & \nonumber\\
& & \Box\sigma-(f\phi^{2}-m^{2})\sigma-
(\partial_{\mu}\psi+eA_{\mu})(\partial^{\mu}\psi+eA^{\mu})\sigma-
\frac{1}{2}\lambda'\sigma^{3}=0\\
& & \nonumber\\
& & \partial_{\mu}G^{\mu\nu}=g^{2}\phi^{2}W^{\nu}\\
& & \nonumber\\
& & \partial_{\mu}F^{\mu\nu}=J^{\nu}\\
& & \nonumber\\
& & \partial_{\mu}J^{\mu}=0\label{Jcons}\end{aligned}$$
where .
The last equation of the above expresses the dynamical conservation of the string current.
Fermionic superconductivity
---------------------------
Following Witten [@witt] we introduce two fermion fields $\Psi_{L}$ and $\Psi_{R}$ which couple to the vortex field $\phi$ through a coupling $\lambda$. Then, the Lagrangian density may be written as [@book],
$${\cal L}=\overline{\Psi}_{L}i\!\not\!\!D\Psi_{L}+
\overline{\Psi}_{R}i\!\not\!\!D\Psi_{R}-
\lambda(\phi\overline{\Psi}_{L}\Psi_{R}+h.c.)
\label{Lferm}$$
where and with $\gamma^{\mu}$ being the Dirac matrices and $q$ the gauge coupling.
As shown in [@witt] fermionic superconductivity resembles strongly the bosonic one apart from some minor aspects such as particle production. There is indeed a way to switch from one picture to the other. Let us introduce the scalar field $y$ such that,
$$\overline{\Psi}\gamma^{a}\Psi\equiv
\frac{1}{\sqrt{\pi}}\,\varepsilon^{ab}\partial_{a}y
\label{Y}$$
where $\varepsilon^{ab}$ is the 2-dimensional Levi-Civita tensor.
Then, with the equivalence, the string current is,
$$J^{a}\equiv -q\overline{\Psi}\gamma^{a}\Psi=
\sqrt{2K}e\,\varepsilon^{ab}\partial_{a}y
\label{Jy}$$
If we define $y$ such that,
$$\partial^{a}y=\sqrt{2K}\varepsilon^{ab}(\partial_{b}\psi+eA_{b})
\label{y}$$
then the current in (\[Jy\]) is identified with the one given in (\[J\]). As shown in [@witt] the above definition of $y$ is consistent. In terms of this formalism we can describe both the fermionic and the bosonic superconducting strings.
The effective action
--------------------
Non-superconducting cosmic strings are described by the well–known Goto–Nambu action [@book],
$$S=-\mu\int d^{2}\xi\sqrt{-\gamma}
\label{Sns}$$
where $\mu$ is the energy per unit length of the string, $\xi^{a}$ are the string world-sheet coordinates and $\gamma$ is the determinant of the world-sheet metric . The relation between the spacetime and the world-sheet coordinates may be expressed as,
$$x^{\mu}=x^{\mu}(\xi^{a})+n_{A}^{\mu}r^{A}$$
where $n_{A}^{\mu}$ are vectors perpendicular to the string, that form a basis for the 2-dimensional perpendicular space spanned by the $r^{A}$ coordinates for which, .[^2] The relation between the string metric $\gamma_{ab}$ and the spacetime metric $g_{\mu\nu}$ to first order is [@turok],
$$g_{\mu\nu}\simeq\left(
\begin{array}{lr}
\gamma_{ab} & 0 \\
& \\
0 & \delta_{AB}
\end{array}
\right)$$
In the case of superconducting string the action has to account for the existence of a current. In the bosonic picture this, in fact, is equivalent with considering the addition of the $\sigma$ field kinetic–energy term,[^3]
$$\begin{aligned}
\Delta S_{J} & = & \int d^{2}\xi\,d^{2}r\sqrt{-\gamma}\,|D_{a}\sigma|^{2}=
\int d^{2}r\,\sigma(r)^{2}\int
d^{2}\xi\sqrt{-\gamma}\,|\partial_{a}\psi+eA_{a}|^{2}=\nonumber\\
& & \nonumber\\
& = & -\frac{1}{2}\int d^{2}\xi\sqrt{-\gamma}\,(\partial y)^{2}
\label{DSJ}\end{aligned}$$
where and we have used (\[sigma\]), (\[K\]) and (\[y\]).
Thus, with the addition of the current term the Goto–Nambu action becomes [@turok],
$$\Delta S=-\int d^{2}\xi\sqrt{-\gamma}\,[\mu+\frac{1}{2}(\partial
y)^{2}]=-\int d^{2}\xi\sqrt{-\gamma}\,(\mu-\frac{J^{2}}{4Ke^{2}})
\label{DS}$$
where we have also used (\[Jy\]).
Including the usual Maxwell terms, the total action for a superconducting string is,
$$S=-\frac{1}{4}\int d^{4}x\sqrt{-g}\,F^{\mu\nu}F_{\mu\nu}-
\int d^{2}\xi\,[\sqrt{-\gamma}(\mu-\frac{J^{2}}{4Ke^{2}})+J^{a}A_{a}]
\label{SJ}$$
In terms of $y$ the above action is [@witt][@sper]…[@linet],
$$S=-\frac{1}{4}\int d^{4}x\sqrt{-g}\,F^{\mu\nu}F_{\mu\nu}-
\int d^{2}\xi\,\{\sqrt{-\gamma}\,[\mu+\frac{1}{2}(\partial y)^{2}]-
\sqrt{2K}e\,\varepsilon^{ab}(\partial_{a}y)A_{b}\}
\label{Sy}$$
where $g$ is the determinant of the spacetime metric and we have used (\[Jy\]).
The above action can be recovered equivalently using the fermionic action [@book],
$$\Delta S_{J}=-\int
d^{2}\xi\sqrt{-\gamma}\;(\overline{\Psi}i\!\not\!\!D\Psi)=
-\int d^{2}\xi\sqrt{-\gamma}\,[\frac{1}{2}(\partial y)^{2}-
\sqrt{2K}e\,\varepsilon^{ab}(\partial_{a}y)A_{b}]$$
where we have used (\[Y\]).
Varying the action with respect to $x^{\mu}$,$A_{\mu}$ and $y$ yields the following equations of motion respectively,
$$\begin{aligned}
& & [\mu+\frac{1}{2}(\partial y)^{2}]\Box_{2}x^{\mu}-
(\partial^{b}y)(\partial_{b}x^{\mu})\Box_{2}y-
(\partial^{a}y)(\partial^{b}y)(\partial_{a}\partial_{b}x^{\mu})+
J_{\nu}F^{\mu\nu}=0\\
& & \nonumber\\
& & \Box_{2}y=\sqrt{2K}e\,\varepsilon^{ab}\partial_{a}A_{b}
\label{boxy}\\
& & \nonumber\\
& & \partial_{\mu}F^{\mu\nu}=\Box A^{\nu}=J^{\nu}\end{aligned}$$
where and we have used the Coulomb gauge, for the Maxwell’s equations. Note also that, .
Using (\[J\]) and (\[Jy\]) the field equation (\[boxy\]) reduces to the trivial expression, , which in fact led to the introduction of $y$ in [@witt]. Also, from (\[y\]) one can easily show that,
$$\Box_{2}\psi=-e\,\partial_{a}A^{a}
\label{boxpsi}$$
Using the above and (\[J\]) one can obtain the current conservation equation (\[Jcons\]). Employing the Coulomb gauge we obtain, , i.e. [*the gradient of the phase field $\psi$ remains constant along the string*]{} [@shel].
Finally, from (\[Jy\]) and (\[boxy\]) we obtain, . Taking the time component of this, we find,
$$\frac{\partial J}{\partial t}=2Ke^{2}E
\label{JE}$$
where is the electric field, externally applied on the string.
The above justifies the fact that the strings are considered to be superconducting. However, as shown in [@witt], (\[JE\]) breaks down when the current grows very large. Indeed, there is a maximum current over which the string loses its superconducting properties. This occurs when the energy of the current either permits unwindings of the phase field (bosonic case) or allows current carriers to escape from the string (fermionic case). In both cases the maximum current cannot exceed, [@witt].
Back-reaction and external fields
---------------------------------
The self-inductance of the string may be defined as follows [@hill],
$$L\equiv\frac{1+4Ke^{2}\ln(\Lambda R)}{2Ke^{2}}\simeq 2\ln(\Lambda R)
\label{L}$$
where is the string width and $R$ is a suitable cut-off radius usually taken as the curvature radius of the string or the inter-string distance of the string network. In the above .
Considering self-inductance effects (\[JE\]) is modified as [@book],
$$\frac{dJ}{dt}=2Ke^{2}(E-L\frac{dJ}{dt})=2K\tilde{e}^{2}E$$
where the renormalised charge is,
$$\tilde{e}^{2}\equiv\frac{e^{2}}{1+2Ke^{2}L}\simeq\frac{1}{2KL}
\label{etild}$$
Considering back-reaction effects the photon field may be expressed as [@witt][@amst1],
$$\hat{A}^{\mu}=
2\int d^{2}\xi\sqrt{-\gamma}\,J^{a}(\xi)\,\partial_{a}x^{\mu}(\xi)\,
\Theta[x^{0}-x^{0}(\xi)]\,\delta([x-x(\xi)]^{2})$$
where the step function $\Theta$ signifies the initiation of the current.
After some algebra the above reduces to [@witt][@turok][@book],
$$\hat{A}^{a}=-2\ln(\Lambda R)J^{a}\simeq -LJ^{a}
\label{AL}$$
Using the above, (\[Jy\]) and (\[boxy\]) we obtain [@book],
$$\Box_{2}\tilde{y}=
\frac{1}{2}\sqrt{2K}\tilde{e}\,\varepsilon_{ab}\hat{F}^{ab}
\label{ytild}$$
which is the renormalised version of (\[boxy\]). In the above and $\hat{F}^{ab}$ is the external field strength. The right-hand side of (\[ytild\]) is the Lorentz force on the string since , where $E_{a}$ and $B_{i}$ are the external electric and magnetic fields respectively [@book].
Similarly, for the string current we have [@vil]…[@linet], $$\begin{aligned}
J^{\mu} & = &
\int d^{2}\xi\sqrt{-\gamma}\,J^{a}\partial_{a}x^{\mu}(\xi)\,
\delta^{(4)}(x-x(\xi))=\nonumber\\
& & \nonumber\\
& = & \int d^{2}\xi\sqrt{-\gamma}\,\varepsilon^{ab}
\partial_{b}y\,\partial_{a}x^{\mu}(\xi)\,
\delta^{(4)}(x-x(\xi))\end{aligned}$$
From the above, since the quantity $ey$ is unaltered by renormalisation, we conclude that [*the string current is unaffected by charge renormalisation*]{}. Also, (\[ytild\]) suggests that, in the absence of external fields [@vil]. Using the renormalised version of (\[Jy\]) we can immediately obtain current conservation.
Equivalently, using (\[J\]) and (\[AL\]) one can show that,
$$J_{a}=\frac{1}{\tilde{e}L}\partial_{a}\psi
\label{Jpsi}$$
which again suggests that the current is conserved since .
Finally, we can obtain the renormalised version of the action by considering the inductance energy of the string [@hill], . Thus, the addition to the Goto–Nambu action due to the string current is,
$$\Delta S=\frac{1}{2}\int d^{2}\xi\sqrt{-\gamma}LJ^{2}$$
Adding the above to (\[Sns\]) and using also (\[etild\]) we find,
$$\Delta S=
-\int d^{2}\xi\sqrt{-\gamma}\,(\mu-\frac{J^{2}}{4K\tilde{e}^{2}})$$
which is the renormalised analogue of (\[DS\]).
In a similar way we can regard the total action in (\[SJ\]) and (\[Sy\]) by considering the external fields and the renormalised values of the physical quantities. From now on, unless stated otherwise, we will refer to the renormalised values of the latter. For economy, we will drop the tildes and hats.
The string current
------------------
Suppose that the string is lying along the $z$-axis. Then, . With the use of (\[Jy\]) we find [@aryal],
$$J^{\mu}(z,t)=
\sqrt{2K}e\,(\frac{\partial y}{\partial z},0,0,-\frac{\partial
y}{\partial t})
\label{Jzt}$$
If is the charge density inside the string then, and . In view of (\[Jzt\]) we can identify,
$$\rho_{e}=-\sqrt{2K}e\,y = \frac{q}{\sqrt{\pi}}\,y$$
that is, [*the field $y$ is a measure of the charge density along the string*]{} .
Now, the Maxwell’s equations give,
$$\frac{1}{r}\frac{\partial}{\partial r}(r\partial_{r}A^{\mu})
=4\pi J^{\mu}\delta^{(2)}(r)\Rightarrow
\partial_{r}A^{\mu}=\frac{2J^{\mu}}{r}
\label{Ar}$$
Using the above we can find the electric and magnetic field around the string,
$$\begin{aligned}
E_{r} & = & \frac{2J^{0}}{r}\nonumber\\
& & \label{sfields}\\
B_{\theta} & = & \frac{2J^{z}}{r}\nonumber\end{aligned}$$
The last of the above is the well-known Biot–Savart law for a line current.
Assuming now that, the charge density along the string is uniformly distributed, and . Then, the electric field vanishes and the only non-zero components of the electromagnetic field strength are [@peter][@linet],
$$F_{Az}=\frac{2J}{r}(\frac{r^{A}}{r})
\label{F}$$
The energy–momentum tensor
==========================
The string current component
----------------------------
Using (\[Jy\]) the action (\[DSJ\]) may be written as,
$$\Delta S_{J}=\frac{1}{4Ke^{2}}\int d^{2}\xi\sqrt{-\gamma}\,J^{2}$$
Varying the above action with respect to the world-sheet metric we obtain the energy–momentum tensor of the string current [@turok][@book],
$$\begin{aligned}
\Theta^{ab} & \equiv & \frac{1}{\sqrt{-\gamma}}\frac{\delta\Delta
S_{J}}{\delta\gamma_{ab}}=
\frac{1}{4Ke^{2}}[2J^{a}J^{b}-\gamma^{ab}J^{2}]=\nonumber\\
& & \nonumber\\
& = & -\frac{1}{2}(\partial
y)^{2}\gamma_{ab}+(\partial_{a}y)(\partial_{b}y)
\label{Theta}\end{aligned}$$
It can be easily checked that the above tensor is traceless. Using (\[ytild\]) we find,
$$\partial_{b}\Theta^{b}_{a}=\partial_{a}y\Box_{2}y=
\frac{1}{2}J^{\mu}F_{\mu\nu}\partial_{a}x^{\nu}$$
In the absence of external fields the above becomes, . In view of (\[Theta\]) this gives,
$$\begin{aligned}
{\bf J}\cdot{\bf E}=0 & \mbox{and} & {\bf J}\times{\bf B}=0\end{aligned}$$
which are obviously satisfied by the self-fields (\[sfields\]) of the string.
With no electromagnetic back-reaction
-------------------------------------
From (\[Sy\]) the energy–momentum tensor of the string is given by [@book],
$$\begin{aligned}
T^{ab} & = & \frac{1}{\sqrt{-\gamma}}
\frac{\delta
S}{\delta\gamma^{ab}}=-\gamma^{ab}[\mu+\frac{1}{2}(\partial y)^{2}]+
\partial_{a}y\partial_{b}y\Rightarrow\nonumber\\
& & \label{T1a}\\
T^{\mu\nu} & = & -(\mu\gamma^{ab}-\Theta^{ab})\,
\partial_{a}x^{\mu}\partial_{b}x^{\nu}\nonumber\end{aligned}$$
Using (\[Theta\]) with it is easy to see that [@linet],
$$\Theta_{0}^{0}=-\Theta_{z}^{z}=-\frac{J^{2}}{4Ke^{2}}$$
Thus, we can write (\[T1a\]) as [@turok],
$$T^{\mu}_{\nu}=-\delta^{(2)}(r)\,
\mbox{diag}(\mu+\frac{J^{2}}{4Ke^{2}},0,0,\mu-\frac{J^{2}}{4Ke^{2}})
\label{T1}$$
Note that the above is a core solution which does not include electromagnetic back-reaction effects.
With electromagnetic back-reaction
----------------------------------
The electromagnetic energy–momentum tensor is,
$$T^{\mu\nu}_{em}=F^{\mu}_{\lambda}F^{\nu\lambda}-
\frac{1}{4}g^{\mu\nu}F_{\lambda\rho}F^{\lambda\rho}
\label{Tem}$$
where $F^{\mu\nu}$ refers to the self-fields of the string.
Using (\[F\]) one easily finds,
$$\begin{aligned}
T^{00}_{em} & = & T^{zz}_{em}=\frac{2J^{2}}{r^{2}}\nonumber\\
& & \label{T2}\\
T^{AB}_{em} & = & \frac{2J^{2}}{r^{2}}\,
(\frac{2r^{A}r^{B}}{r^{2}}-\delta^{AB})\nonumber\end{aligned}$$
With a suitable rotation of the $r^{A}$ coordinates (\[T2\]) can be written as [@HK][@babul],
$$[T_{em}(r)]^{\mu}_{\nu}=\frac{2J^{2}}{r^{2}}\,\mbox{diag}(-1,1,-1,1)$$
The above tensor is traceless but not well-defined at . In order to overcome this problem we employ a method introduced by Linet [@linet] and expand the components of $T^{\mu\nu}_{em}$ as distributions around the string core (see also [@peter]). We use the fact that where is the 2-dimensional Laplacian and $r_{0}$ is the radius of the string core.[^4] The energy–momentum tensor (\[T2\]) is, then, written as,
$$\begin{aligned}
T^{00}_{em} & = & T^{zz}_{em}=J^{2}\Delta_{2}
[\ln(r/r_{0})]^{2}\nonumber\\
& & \label{Tln}\\
T^{AB}_{em} & = & -2J^{2}\partial_{A}\partial_{B}[\ln(r/r_{0})]
\nonumber\end{aligned}$$
However, since , one finds, and $T_{em}^{\mu\nu}$ is not conserved. In order to have conservation we need to add the term on $T^{AB}_{em}$. The addition of this term suggests that the core solution (\[T1\]) for the string energy–momentum tensor becomes [@MP][@linet],
$$T^{\mu}_{\nu}=-\delta^{(2)}(r)\,
\mbox{diag}(\mu+\frac{J^{2}}{4Ke^{2}},-\frac{J^{2}}{2},
-\frac{J^{2}}{2},\mu-\frac{J^{2}}{4Ke^{2}})
\label{T}$$
The above is a core solution that takes into account electromagnetic back-reaction. As can be seen in (\[T\]), the string current increases the linear mass density $\mu$ of the string by a factor $\frac{1}{2}LJ^{2}$ which corresponds to the self-inductance energy density of the current. Also, it generates pressure due to the inertia of the charge carriers [@babul], both towards the perpendicular direction ($T^{1}_{1}$ and $T^{2}_{2}$ terms) and along the string. The latter also reduces the total string tension $T^{z}_{z}$.
In fact, (\[T\]) suggests that, for high enough currents, the string tension may be diminished to zero. In that case the strings would not have a driving force to untangle. Also, string loops would not collapse, but remain stable and form the so called “springs” [@turok][@hill][@spring]. However, it has been argued [@shel] that the back-reaction of the current would shrink the $\sigma$-condensate in the string core (in the bosonic case) thus reducing the pressure from the charge carriers that counteracts the string tension. In overall, the string tension may not decrease as rapidly due to the current as implied by (\[T\]). Indeed, numerical simulations [@peter2] have shown that the string tension is little affected by changes of the string current. It would be more accurate, then, to express the core solution as,
$$T^{\mu}_{\nu}=-\delta^{(2)}(r)\,\mbox{diag}(U,W,W,T)
\label{TU}$$
where [@peter]. Then the overall solution for the string energy–momentum tensor may be written as,
$$\begin{aligned}
T^{00}_{em} & = & U\delta^{(2)}(r)+J^{2}\Delta_{2}
[\ln(r/r_{0})]^{2}\nonumber\\
& & \nonumber\\
T^{zz}_{em} & = & -T\delta^{(2)}(r)+J^{2}\Delta_{2}
[\ln(r/r_{0})]^{2}\label{Ts}\\
& & \nonumber\\
T^{AB}_{em} & = & J^{2}\delta^{AB}\delta^{(2)}(r)
-2J^{2}\partial_{A}\partial_{B}[\ln(r/r_{0})]
\nonumber\end{aligned}$$
Thus, the following should be regarded as extreme estimates,
$$\begin{aligned}
U & \simeq & \mu+\frac{J^{2}}{4Ke^{2}}\label{U}\\
& &\nonumber\\
T & \simeq & \mu-\frac{J^{2}}{4Ke^{2}}\label{TT}\end{aligned}$$
The string spacetime
====================
The string metric
-----------------
Using (\[Ts\]) one can solve the Einstein equations,
$$R^{\mu\nu}-\frac{1}{2}g^{\mu\nu}R^{\lambda}_{\lambda}=8\pi GT^{\mu\nu}$$
where is Newton’s gravitational constant.
In first order in $G$ the metric is found to be [@HK][@peter],
$$\begin{aligned}
ds^{2} & = &
\{1+4G[J^{2}+(U-T)]\ln(r/r_{0})+4GJ^{2}[\ln(r/r_{0})]^{2}\}
(-dt^{2}+dr^{2})\nonumber\\
& & \nonumber\\
& + & \{1-8G(U+\frac{J^{2}}{2})-
4G[J^{2}-(U-T)]\ln(r/r_{0})+4GJ^{2}[\ln(r/r_{0})]^{2}\}
r^{2}d\theta^{2}\nonumber\\
& & \nonumber\\
& + &
\{1+4G[J^{2}-(U-T)]\ln(r/r_{0})-4GJ^{2}[\ln(r/r_{0})]^{2}\}dz^{2}
\label{ds}\end{aligned}$$
where .
The metric of the spacetime perpendicular to the string may be found by setting, . Then the above gives,
$$ds^{2}_{\perp}=(1-h_{00})[-dt^{2}+dr^{2}+(1-b)r^{2}d\theta^{2}]
\label{dsT}$$
where,
$$h_{00}=-4G[J^{2}+(U-T)]\ln(r/r_{0})-4GJ^{2}[\ln(r/r_{0})]^{2}
\label{h00}$$
and to first order in $G$,
$$\begin{aligned}
1-b & = & \frac{1-8G(U+J^{2}/2)-
4G[J^{2}-(U-T)]\ln(r/r_{0})+4GJ^{2}[\ln(r/r_{0})]^{2}}{
1+4G[J^{2}+(U-T)]\ln(r/r_{0})+4GJ^{2}[\ln(r/r_{0})]^{2}}\nonumber\\
& & \\
& \simeq & 1-8G(U+\frac{J^{2}}{2})-8GJ^{2}\ln(r/r_{0})\nonumber\end{aligned}$$
From (\[dsT\]) we see that the spacetime around a superconducting string resembles the conical spacetime of a non-superconducting string. Indeed, if we set and , then and , and (\[dsT\]) reduces to the non-superconducting string, conical–spacetime solution with deficit angle . In the case of non-vanishing current, though, the deficit angle is,
$$\delta=b\pi=8\pi G\{U+J^{2}[\frac{1}{2}+\ln(r/r_{0})]\}
\label{delr}$$
Thus, the spacetime is not exactly conical since $\delta$ is dependent on $r$. However, this logarithmic dependence is very weak and, if one is interested in astrophysical effects (such as primordial magnetic field generation) then the logarithmic dependence may well be approximated as,
$$\ln(r/r_{0})\simeq\ln(\Lambda R)
\label{app}$$
Under this approximation (\[delr\]) is in good agreement with the more rigourous calculations of [@HK][@peter]. Now, using (\[U\]) and (\[etild\]) we find,
$$\delta\simeq 8\pi G[\mu+\frac{1}{2}(QJ)^{2}]
\label{d}$$
where,
$$Q\equiv\sqrt{1+4\ln(\Lambda R)}\simeq 20
\label{Q}$$
The gravitational field
-----------------------
From (\[ds\]) it can be seen that the string metric deviations from Minkowski spacetime are of the order of . Thus, we can use linear theory to approximate the gravitational field. The metric, then, can be written as,
$$g_{\mu\nu}=\eta_{\mu\nu}+h_{\mu\nu}$$
where is the Minkowski metric.
The geodesic equation is, where, is the 4-velocity and are the Christoffel symbols in the linear approximation. Since, the geodesic equation becomes,
$$\frac{d^{2}x^{i}}{d\tau^{2}}+\Gamma^{i}_{00}=0
\label{geo}$$
where $i$ denotes the spacial coordinates and is the proper time. Since, we find,
$${\bf f}=\frac{1}{2}\nabla h_{00}$$
Inserting (\[h00\]) into the above we obtain,
$$f(r)=-\frac{2GJ^{2}}{r}\,[1+\frac{U-T}{J^{2}}+2\ln(r/r_{0})]
\label{fF}$$
Using (\[etild\]), (\[U\]), (\[TT\]) and the approximation (\[app\]) we find,
$$f(r)\simeq -\frac{2G(QJ)^{2}}{r}
\label{f}$$
The above is an attractive gravitational force similar to the one of a massive rod of linear mass density .
In (\[f\]) we have used the extreme values (\[U\]) and (\[TT\]) of the string linear mass density and tension. As mentioned, though, the string tension may be larger than suggested by (\[TT\]). In that case $Q$ in the expression (\[f\]) would decrease and the gravitational pull of the string will be weakened since it depends on the difference as shown in (\[fF\]). However, (\[ff\]) suggests that, even in the extreme case, when , the attractive force is decreased only by a factor of 2.
Particle deflection
-------------------
Writing the metric (\[dsT\]) in cartesian coordinates we have,
$$ds^{2}_{\perp}=(1-h_{00})(-dt^{2}+dx_{k}dx^{k})$$
where the and we have assumed that the string lies on the $z$-axis. Note that we need to extract from the above a wedge of angle width $\delta$ [@tvach]. Then,
$$d\tau^{2}=-d^{2}s_{\perp}=(1-h_{00})dt^{2}(1-\dot{x}_{k}\dot{x}^{k})$$
where the dots signify derivation with respect to time.
Using and , we insert the above to the geodesic equation (\[geo\]) and obtain [@tvach],
$$2\ddot{x}^{i}=(1-\dot{x}_{k}\dot{x}^{k})\partial^{i}h_{00}
\label{acl}$$
where .
The above gives the acceleration felt by the particles due to the gravitational pull of the string field, in the frame of the string.
Suppose that the string moves in the $x$-direction with constant velocity . Then for a particle in the position we have initially, and . The velocity boost felt by the particle towards the $y$-direction after its encounter with the string is [@book],
$$v_{y}=\int_{-\infty}^{\infty}\ddot{y}dt=
\frac{1}{2v\gamma^{2}}\int_{-\infty}^{\infty}\partial_{y}h_{00}dx$$
where is the Lorentz factor and we have also used (\[acl\]).
Using (\[h00\]) and (\[app\]) the above gives,
$$v_{y}=-\frac{2\pi G(QJ)^{2}}{v\gamma^{2}}$$
Taking into account the deficit angle, the relative boost between two particles on the opposite sides of the string is, . Switching to the particle frame gives,
$$u=\gamma\Delta v_{y}=8\pi G\mu v\gamma+4\pi
G(QJ)^{2}(v\gamma+\frac{1}{v\gamma})
\label{boost}$$
where we have also used (\[d\]).
Deceleration of the string
--------------------------
Avelino and Shellard [@paper] have realised that the deflection of particles in the string spacetime results into a net drag of the plasma behind the string. This is due to the fact that the magnitude of the particle velocity is unaltered after its interaction with the string. The back-reaction of this effect is a decelerating force on the string. The velocity of plasma dragging may be estimated by Taylor expansion of the particle velocity. In the lowest order in $G$ we find,
$$\delta v_{x}\simeq-\frac{1}{2}\frac{u^{2}}{v}=
-32\pi^{2}G^{2}\mu v\gamma^{2}[\mu+J^{2}(1+\frac{1}{v^{2}\gamma^{2}})]-
8\pi^{2}G^{2}(QJ)^{4}v\gamma^{2}(1+\frac{1}{v^{2}\gamma^{2}})
\label{vx}$$
The momentum change of the string is, , where is the energy density of the universe with . Thus, the drag force per unit length on the string is,
$$f_{x}=\int\frac{d^{2}p}{dtdz}\,dxdy=2R\rho v\,\delta v_{x}$$
where is the inter-string distance of the network [@mine].
Inserting (\[vx\]) into the above we find,
$$f_{x}=-6\pi GHv\{2\mu^{2}v^{2}\gamma^{2}+
\frac{1}{2}(QJ)^{4}(v\gamma+\frac{1}{v\gamma})^{2}+
2\mu J^{2}(v^{2}\gamma^{2}+1)\}
\label{fx}$$
where is the Hubble radius.
The above force should be compared with the plasma friction force on the strings, , [@mine] where $m$ is the particle mass. It is easy to show that throughout the range of $J$ the above drag force is always subdominant and, thus, it does not affect the network evolution.[^5]
Primordial magnetic field generation
====================================
Although the motion of the strings is not significantly affected by the dragging of the plasma, the latter may gain substantial momentum during this process. Such momentum may introduce turbulence which could generate a primordial magnetic field.
Momentum transfer
-----------------
In the one scale model the scale of the string network inter-string distance is comparable to the string curvature radius, . A string segment of length $\sim R$ may transfer momentum to the plasma contained in an inter-string volume $\sim R^{3}$. In that way the network may introduce vortical motions to the plasma on inter-string scales. An estimate of the plasma rotational velocity $v_{rot}$ may be obtained as follows,
$$FR \simeq \frac{1}{2}Mv_{rot}^{2}$$
where is the mass of an inter-string volume and $F$ is the total force on the plasma by a string segment of length $R$. From the above the rotational velocity is estimated as,
$$v_{rot}\sim\frac{\sqrt{GF}}{v}
\label{vrot}$$
The total force may be obtained from (\[fx\]) as follows,
$$\begin{aligned}
F & = & \int_{0}^{R}f_{x}dz=vtf_{x}\Rightarrow\nonumber\\
& & \label{ff}\\
F & = & 3\pi Gv^{2}[2\mu^{2}v^{2}\gamma^{2}+
\frac{1}{2}(QJ)^{4}(v\gamma+\frac{1}{v\gamma})^{2}+
2\mu J^{2}(v^{2}\gamma^{2}+1)]\nonumber\end{aligned}$$
As shown in [@mine], for currents greater than a critical value $J_{c}$ the network can never exit the friction domination era of its evolution. In [@mine] $J_{c}$ is estimated as,
$$J_{c}\sim\sqrt{G}\,\mu
\label{Jc}$$
However, even for such strong currents the network does reach a scaling solution during which the strings move with a terminal velocity given by [@mine],
$$v_{T}^{2}\sim\frac{G\mu}{\sqrt{GJ^{2}}}
\label{vT}$$
As implied in [@mine] the string current is influential on the behaviour of the network only if . Thus, we will consider this case only. Using (\[vT\]) the total force (\[ff\]) becomes,
$$F=3\pi[2(G\mu)^{2}(\frac{\mu}{J^{2}})\mu+
\frac{1}{2}Q^{4}(G\mu)^{2}J^{2}+
\frac{1}{2}Q^{4}(GJ^{2})J^{2}+
Q^{4}(GJ^{2})\sqrt{G\mu}\sqrt{\frac{\mu}{J^{2}}}J^{2}+
2(G\mu)\mu^{2}+2\sqrt{GJ^{2}}\mu^{2}]$$
In the above the 2nd, the 4th and the 5th terms of the right hand side may be disregarded by means of the condition . By comparing the remaining terms with each other it can be shown that the last term remains also always subdominant. Therefore, the above can be written as,
$$F=6\pi J^{2}[(G\mu)^{2}(\frac{\mu}{J^{2}})^{2}+
\frac{1}{2}Q^{4}(GJ^{2})]
\label{FF}$$
Considering the overall range of $J$ we can rewrite the above as,[^6]
$$F\simeq\left\{
\begin{array}{lr}
3\pi G(QJ)^{4} & \;\;\;\;\;J_{c1}<J\leq J_{max}\\
& \\
6\pi(G\mu)^{2}(\mu/J^{2})\mu & \;\;\;\;\;J_{c}<J\leq J_{c1}\\
& \\
6\pi G\mu^{2} & \;\;\;\;\;J\leq J_{c}
\end{array}
\right.
\label{F3}$$
where the critical current $J_{c1}$ is,
$$J_{c1}\equiv Q^{-2/3}(G\mu)^{1/6}\sqrt{\mu}
\label{Jc1}$$
For currents stronger that the above value the current dependent term in (\[boost\]) becomes dominant, i.e. the gravitational pull of the string is actually felt by the particles.
Inserting (\[F3\]) into (\[vrot\]) and using also (\[vT\]) we find,
$$v_{rot}\simeq\left\{
\begin{array}{lr}
\sqrt{3\pi}Q^{2}(G\mu)^{-1/4}(GJ^{2})
(\mu/J^{2})^{-1/4} & \;\;\;\;\;J_{c1}<J\leq J_{max}\\
& \\
\sqrt{6\pi}(G\mu)^{5/4}(\mu/J^{2})^{1/4} & \;\;\;\;\;J_{c}<J\leq J_{c1}\\
& \\
\sqrt{6\pi}(G\mu) & \;\;\;\;\;J\leq J_{c}
\end{array}
\right.
\label{vrot1}$$
The Harrison–Rees magnetic field
--------------------------------
Harrison [@hary] suggested first that turbulence in an expanding universe may generate a magnetic field since the turbulent velocity of the plasma would be different for the ions and the electrons.
Consider a rotating volume $V$ of plasma. Suppose that the angular velocities $\omega_{m}$ and $\omega_{r}$ of the ion and the electron fluid respectively are uniform inside $V$. Then, since , we find that,
$$\begin{aligned}
\rho_{m}V=const.\;\;\;\;\; & \mbox{and} &
\;\;\;\;\;\rho_{r}V^{4/3}=const.\end{aligned}$$
where is the ion density, which scales as pressureless matter, and is the electron density, which scales as radiation due to the strong coupling between the electrons and the photons through Thompson scattering. The angular momentum of each plasma component has to be conserved. This suggests that,
$$\begin{aligned}
\omega_{m}\propto V^{-2/3}\propto a^{-2}\;\;\;\;\; &
\mbox{and} & \;\;\;\;\;\omega_{r}\propto V^{-1/3}\propto a^{-1}\end{aligned}$$
Thus, the ion fluid spins down faster than the electron-photon gas. Consequently, a current is generated which creates a magnetic field in the volume $V$.
Rees, however, has shown that expanding volumes of spinning plasma are unstable and decay with cosmic expansion compared to irrotational density perturbations [@rees]. He suggested instead a different version of vortical magnetic field generation involving the scattering of the electrons on the microwave background radiation. This would tend to damp the vortical motions of the electrons in contrast to the ions which would stay unaffected. The result is again the generation of currents but, this time, it is the electron fluid that slows down. The mechanism applies to gravitationally bound spinning bodies such as those formed when structure formation begins.
In both cases, the Maxwell’s equations suggest [@hary],
$${\bf B}\simeq-\frac{m}{e}\,{\bf w}$$
where is the nucleon mass and [**w**]{} is the vorticity given by,
$${\bf w}=\nabla\times{\bf v}_{rot}$$
From the above the magnetic field generated over inter-string distances is,
$$B\simeq\frac{m}{e}(\frac{v_{rot}}{R})
\label{Bv}$$
The turbulence inside the wake of cosmic strings is expected to ionise the plasma even after decoupling [@book]. Also for superconducting cosmic strings the existence of a magnetocylinder around the core [@mine] is expected to induce further charge separation due to the difference of the inertia of the scattered particles. Since , from (\[Bv\]) it is evident that the stronger field will be generated at early times. Thus, we will calculate the magnetic field generated at , the time of equal matter and radiation densities, since this is the earliest that large scale streaming of the plasma is possible. From (\[vrot1\]) and (\[B\]) we obtain,
$$B_{eq}\simeq\frac{\sqrt{6\pi}\,m}{et_{eq}}\times\left\{
\begin{array}{lr}
\frac{1}{\sqrt{2}}Q^{2}(G\mu)^{-1/2}(GJ^{2})
(\mu/J^{2})^{-1/2} & \;\;\;\;\;J_{c1}<J\leq J_{max}\\
& \\
G\mu & \;\;\;\;\;J\leq J_{c1}
\end{array}
\right.
\label{Beq}$$
For the above gives,
$$B^{max}_{eq}\sim 10^{-13}\sqrt{G\mu}\,Gauss$$
which is coherent over comoving scales,
$$l\sim(v_{T}t_{eq})(\frac{t_{pr}}{t_{eq}})^{2/3}\sim
10^{2}(G\mu)^{1/4}M\!pc$$
where is the present time.
For energy scales corresponding to a grand unified theory (GUT) phase transition, and the above suggest that with a coherency of the order of which is quite sufficient to seed the galactic magnetic fields. However, as we discuss below, this estimate may be over-optimistic.
Temperature anisotropy constraint
---------------------------------
The deficit angle of the string spacetime apart from deflecting the trajectories of particles affects light propagation as well [@book]. From (\[dsT\]) by setting we see that photons are boosted by the deficit angle even though the prefactor is irrelevant. As a result a string moving in front of radiation will blueshift light due to the Doppler effect. Thus, a string network is expected to generate temperature anisotropies on the microwave background radiation. These cannot exceed the observed values,
$$(\frac{\Delta T}{T})_{_{rms}}\leq 10^{-6}$$
The anisotropy generated by a single string is [@book],
$$(\frac{\Delta T}{T})_{_{S}}\sim\delta v\gamma$$
The rms anisotropy due to a network of cosmic strings is estimated as [@mine][@leandros],
$$(\frac{\Delta T}{T})_{_{rms}}\sim\frac{H^{-1}}{R}(\frac{\Delta T}{T})_{S}$$
From the above and (\[d\]) we find that for superconducting cosmic strings,
$$(\frac{\Delta T}{T})_{_{rms}}\sim\delta\sim G[\mu+\frac{1}{2}(QJ)^{2}]
\label{Trms}$$
Thus, for even the maximum current could not challenge the observations. However, for GUT strings one cannot allow the current to reach its maximum value. Indeed, there is some doubt whether superconducting strings may attain the maximum current [@babul] since the energy density of the current should not exceed the string linear mass density $\mu$ [@aryal]. This implies that the current term in (\[Trms\]) should always remain subdominant. Therefore,
$$J\leq J_{c2}\equiv\frac{\sqrt{\mu}}{Q}
\label{Jc2}$$
By evaluating the (\[Beq\]) with we obtain the maximum permissible magnetic field for GUT energy scales,
$$B^{max}_{eq}\sim 10^{-15}Q^{-1}\sqrt{G\mu}\,Gauss\sim 10^{-19}Gauss$$
with coherence length,
$$l\sim 10^{2}\sqrt{Q}(G\mu)^{1/4}M\!pc\sim 1\,M\!pc$$
The above field is of sufficient strength and coherency to seed the galactic dynamo process and generate the observed galactic magnetic fields.
Discussion and conclusions
==========================
We have showen that GUT superconducting cosmic strings are able to generate turbulence on inter-string scales, which gives rise to a primordial magnetic field of enough strength to seed the dynamo process in galaxies and account for the observed galactic magnetic fields. Moreover, the generated field is coherent over very large scales, comparable with the protogalactic ones before gravitational collapse commenced. Turbulence and coherent rotation on these scales may also be related to the fragmentation process of galaxy formation. Furthermore, the existence of a primordial field, coherent over an entire protogalaxy, may have played a crucial role in removing angular momentum in a similar way as in the case of the collapse of protostellar clouds [@zeld][@mestel].
Due to excessive microwave temperature anisotropies, a maximum string current may not be acceptable. However, even under this constraint, the magnetic field generated is still adequate to seed the galactic dynamo. A possibly stronger constraint may be implemented by considering the density inhomogeneities due to the string wakes, since their growth is enhanced not only due to the deficit angle but also, due to the gravitational pull of the strings [@book]. However, the magnitude of the overdensities is strongly dependent on the dark matter model used. Indeed, string wake overdensities may be substantially suppressed in the case of hot dark matter, due to free streaming [@book].
In our estimates of the generated magnetic field we have used the extreme values (\[U\]) and (\[TT\]) for the string linear mass density and tension. As we mentioned though, the string tension may be larger than suggested by (\[TT\]). This however, would relax the temperature anisotropy constraint and the density inhomogeneity constraint even more.
We have implicitly assumed that the string current remains constant during the network evolution. Indeed, the current is dynamically conserved as shown by (\[Jcons\]). Also, in the bosonic case (\[Jpsi\]) suggests that current conservation is ensured on topological grounds. However, most of the field theory of Section 2, concerns a straight and infinite superconducting string. A realistic string may be much more complicated. Still, it can be shown [@mine] that even in this case the string current remains constant when the network reaches a scaling solution.
Our treatment is similar to this of Avelino and Shellard [@paper]. In their work, however, they consider the case of wiggly strings, which also have an attractive gravitational field due to the difference between their renormalised linear mass density and tension. The latter is a result of the tangled shape of wiggly strings, which, in effect, increases the string length (and thus, the linear mass density) between two fixed points on the string while also decreasing the tension due to the random orientation of the string microstructure [@wiggly]. For superconducting strings, microstructure is suppressed by electromagnetic radiation emission [@grays] and the difference between the linear mass density and tension is due to the existence of a current, which increases the energy content of the string while generating pressure and reducing, thus, the string tension [@babul]. Therefore, the gravitational field of superconducting strings arises in a qualitatively different way that the one of wiggly strings. Moreover, there are also quantative differences between the two cases, since, for strong enough currents, a superconducting string network reaches a different scaling solution, due to excessive plasma friction [@mine]. As a result, the inter-string distance is much smaller than the horizon and the strings are slow moving. The above enable superconducting strings to generate much stronger magnetic fields than wiggly strings. In fact, since the string microstructure is suppressed by the existence of a current, wiggly strings may be regarded as the limit of superconducting strings when the current is very small.
The superconducting string model provides a realistic mechanism for primordial magnetic field generation. The field generated by natural (GUT-scale) values of the parameters is coherent over protogalactic scale and strong enough to seed the dynamo process in galaxies. Thus, our mechanism may be considered as a prime candidate to explain the observed galactic magnetic fields.
I would like to thank A.C. Davis for discussions and her support during the first stages of this work and M. Papastathi for her edditing help. This work was supported by the Isaac Newton Fund (Trinity College, Cambridge), and the Greek State Scholarships Foundation (I.K.Y.).
[blabla]{}
E.N. Parker, “Galactic Dynamos and other questions on the origin of Magnetic Fields” in Proceedings of [*Critical Dialogues in Cosmology*]{}. ed. , World Scientific, in press.
L. Mestel and K. Subramanian, Mon. Not. R. Astr. Soc. [**248**]{} (1991) 677; , and , “Galactic dynamo: Modes and models” in [*The Cosmic Dynamo*]{}, ed. , and , Kluwer Academic Publishers, Dordrecht, Holland, .
P.P. Kronberg, Rep. Prog. Phys. (1994) 325.
S.I. Vainstein and A.A.Rutzmaikin, A. Zh. [**48**]{} (1971) 902; , [*Magnetic Field Generation in Electrically Conducting Fluids*]{}, CUP 1978; , [*Cosmological Magnetic Fields*]{}, OUP, Oxford 1979; , and , [*Magnetic Fields of Galaxies*]{}, Kluwer, Dordrecht 1988; and , Sov. Astr. [**15**]{} (1992) 714; , , , and , An. Rev. Astron. Astrophys. [**34**]{} (1996) 155.
Y.B. Zel’dovich, A.A. Ruzmaikin and D.D. Sokoloff, [*Magnetic Fields in Astrophysics*]{}, McGraw-Hill, New York 1983.
R.M. Kulsrud and S.W. Anderson, Ap. J. [**396**]{} (1992) 606.
M.J. Rees, Q. Jl. R. Astr. Soc. [**28**]{} (1987) 197.
S. Davidson, Phys. Lett. B [**380**]{} (1996) 253.
C.J. Hogan, Phys. Rev. Lett. [**51**]{} (1983) 1488; , and , Ap. J. [**344**]{} (1989) L49; and ,Phys. Rev. D [**50**]{} (1994) 2421; , and , Phys. Rev. D [**53**]{} (1996) 662; , and , Phys. Rev. D [**55**]{} (1997) 4582.
T. Vachaspati, Phys. Lett. B [**265**]{} (1991) 258; and , Phys. Lett. B [**319**]{} (1993) 178; and , Phys. Lett. B [**360**]{} (1995) 71; and , Phys. Rev. D [**55**]{} (1997) 7398.
M.S. Turner, L.M. Widrow, Phys. Rev. D [**37**]{} (1988) 2743; , Ap. J. [**391**]{} (1992) L1; , and , Phys. Rev. D [**46**]{} (1992) 5346; , Phys. Rev. D [**48**]{} (1993) 2499; and |mbox[F.M. Spedalieri]{}, Phys. Rev. D [**52**]{} (1995) 6694.
D. Lemoine and M. Lemoine, phys. Rev. D [**52**]{} (1995) 1955; , and , Phys. Rev. Lett. [**75**]{} (1995) 3796.
T. Vachaspati and A. Vilenkin, Phys. Rev. Lett. [**67**]{} (1991) 1057; , Phys. Rev. D [**45**]{} (1992) 3487; , Phys. Rev. D [**48**]{} (1993) 3585.
E.R. Harrison, Nature [**224**]{} (1969) 1089; Phys. Rev. Lett. [**30**]{} (1973) 188.
P.P. Avelino and E.P.S. Shellard, Phys. Rev. D[**51**]{} (1995) 5946.
A.C. Davis and W.B. Perkins, Phys. Lett. B [**390**]{} (1997) 107.
K. Dimopoulos and A.C. Davis, preprint DAMTP-97-8 (hep-ph/9705302).
E. Witten, Nucl. Phys. B [**249**]{} (1985) 557.
E. Copeland, M. Hindmarsh and N. Turok, Phys. Rev. Lett. [**58**]{} (1987) 1910; Nucl. Phys. B [**306**]{} (1988) 908.
C.T. Hill, H.M. Hodges and M.S. Turner, Phys. Rev. Lett. [**59**]{} (1987) 2493; Phys. Rev. D [**37**]{} (1988) 263.
M. Hindmarsh, Phys. Lett. B [**225**]{} (1989) 127.
R.L. Davis and E.P.S. Shellard, Phys. Lett. B [**207**]{} (1988) 404; [*ibid*]{}. [**209**]{} (1988) 485.
T.M. Helliwell and D.A. Konkowski, Phys. Lett. A [**143**]{} (1990) 438.
A. Babul, T. Piran and D.N. Spergel, Phys. Lett. B [**202**]{} (1988) 307; [*ibid*]{}. [**209**]{} (1988) 477.
P. Peter and D. Puy, Phys. Rev. D [**48**]{} (1993) 5546.
A. Vilenkin and E.P.S. Shellard, [*Cosmic Strings and other Topological Defects*]{}, Cambridge Monographs on Mathematical Physics, CUP, Cambridge 1994.
D.N. Spergel, T. Piran and J. Goodman, Nucl. Phys. B [**291**]{} (1987) 847.
A. Vilenkin and T. Vachaspati, Phys. Rev. Lett. [**58**]{} (1987) 1041.
M. Aryal, A. Vilenkin and T. Vachaspati, Phys. Lett. B [**194**]{} (1987) 25.
P. Amsterdamski, Phys. Rev. D [**39**]{} (1989) 1524.
B. Linet, Cl. Quant. Grav. [**6**]{} (1989) 435; Phys. Lett. A [**146**]{} (1990) 159.
I. Moss and S. Poletti, Phys. Lett. B [**199**]{} (1987) 34.
D. Haws, M. Hindmarsh and N. Turok, Phys. Lett. B [**209**]{} (1988) 255.
P. Peter, Phys. Rev. D [**46**]{} (1992) 3335.
T. Vachaspati, Phys. Rev. D [**45**]{} (1992) 3487.
L. Perivolaropoulos, Phys. Lett. B [**298**]{} (1993) 305.
L. Mestel, Physica Scripta [**T11**]{} (1986) 53.
B. Carter, Phys. Rev. D [**41**]{} (1990) 3886; , [*ibid*]{}. [**41**]{} (1990) 3038.
B. Pacyński, Ap. J. [**335**]{} (1988) 525; , [*ibid*]{}.. [**424**]{} (1994) L9.
[^1]: e-mail: [email protected]
[^2]: We will use $a,b,c$ e.t.c. indeces to denote the components on the 2-dimensional world-sheet of the string, whereas $\lambda,\mu,\nu,$ e.t.c. indeces will be used to denote the components in 4-dimensions. The projection of a vector $V^{\mu}$ on the string world-sheet is where $x^{\mu}$ are the 4-dimensional space coordinates. In a similar way we can project higher rank tensors. Finally, the $A,B,C$ e.t.c. indeces will be used for the 2-dimensional space perpendicular to the string.
[^3]: The potential energy due to the $\sigma$ field is included into $\mu$.
[^4]: The core radius is defined as the distance where the string gauge fields assume their long-range logarithmic behaviour whereas the rest of the vortex fields become negligible. The above results are also in agreement with [@HK] if one takes into account that .
[^5]: For example, in the case of non-superconducting strings () the drag force dominates string friction at the temperature, . Comparing this with the temperature when the network reaches the horizon scaling solution [@mine], we see that $f_{x}$ has no effect during the friction era of the network. During horizon scaling we can compare $f_{x}$ with the string tension $\sim(\mu/t)$ and verify that the latter dominates at all times.
[^6]: Note that .
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We review the main research directions ongoing in the development of high-harmonic generation-based extreme ultraviolet sources for the synthesization and application of trains and isolated attosecond pulses to time-resolved spectroscopy. A few experimental and theoretical works will be discussed in connection to well-established attosecond techniques. In this context, we present the unique possibilities offered for time-resolved investigations on the attosecond timescale by the new Extreme Light Infrastructure Attosecond Light Pulse Source, which is currently under construction.'
address:
- 'Dipartimento di Fisica, Politecnico di Milano Piazza Leonardo da Vinci 32, 20133 Milano Italy'
- 'Institute of Photonics and Nanotechnologies, CNR-IFN, Piazza Leonardo da Vinci 32, 20133 Milano Italy'
- 'ELI-ALPS, ELI-Hu Kft., Dugonics ter 13, H-6720 Szeged Hungary'
- 'Foundation for Research and Technology-Hellas, Institute of Electronic Structure and Lasers B.O. Box 1527, GR-711 10 Heraklion Crete, Greece'
- 'Department of Optics and Quantum Electronics, University of Szeged, Dóm tér 9, 6720 Szeged, Hungary'
author:
- Maurizio Reduzzi
- Paolo Carpeggiani
- Sergei Kühn
- Francesca Calegari
- Mauro Nisoli
- Salvatore Stagira
- Caterina Vozzi
- Peter Dombi
- Subhendu Kahaly
- Paris Tzallas
- Katalin Varju
- Karoly Osvay
- Giuseppe Sansone
bibliography:
- 'Rev\_bib.bib'
title: 'Advances in high-order harmonic generation sources for time-resolved investigations'
---
high-order harmonic generation; attosecond spectroscopy; ultrafast time-resolved dynamics
Introduction {#sec1}
============
Since the first demonstration of high-order harmonic generation (HHG) in gases [@JPB-Ferray-1988; @PRL-Lhuillier-1993], the efforts of several research groups, combined with the development of new technologies for the generation of intense, high-repetition rate driving sources in the near (IR) and mid-infrared (mid-IR) spectral range, has led to impressive progresses in the field of ultrafast extreme ultraviolet (XUV) spectroscopy and of attosecond science. After the first pioneering experiments in atoms [@NATURE-Drescher-2002], molecules [@NATURE-Sansone-2010], and condensed phase systems [@NATURE-Cavalieri-2007], time-resolved experiments using high-order harmonics are moving fast towards the investigation of more complex systems such as biomolecules [@SCIENCE-Calegari-2014] and composite materials [@NATURE-Neppl-2015]. All these experiments indicate that the fundamental steps of electronic dynamics evolve on the attosecond timescale, calling for the reproducible generation and characterization of sub-femtoseconds pulses to excite and probe electronic wave packets. While the generation of trains of attosecond pulses can be accomplished by using multi-cycle IR intense femtosecond pulses, the synthesization of isolated attosecond pulses requires the precise control of the electric field of the driving pulse [@NATURE-Baltuska-2003] in combination with techniques for the (spectral or temporal) confinement of the harmonic generation mechanism [@NATPHOT-Chini-2014]. Nowadays, pulse durations are quickly approaching the atomic unit of time (1 a.u.=24 as) [@JPB-Sansone-2009; @OL-Zhao-2012].
Time-resolved studies based on high-order harmonic radiation, however, are affected by limitations that can be traced back to the fundamental characteristics of the HHG process. While the XUV-pump-IR-probe approach is routinely implemented in several laboratories, the conceptually more straightforward XUV-pump-XUV-probe approach still represents a formidable experimental challenge and it has been demonstrated only by a few groups worldwide [@NATPHYS-Tzallas-2011; @NATCOMM-Takahashi-2013]. The main reason resides in the low conversion efficiency of the HHG process (usually in the $10^{-9}-10^{-5}$ range) that calls for high-energy (several tens or hundreds of mJ) driving pulses for reaching the energy level required for multi-photon interaction and nonlinear absorption in the XUV spectral range.
Similarly, HHG by driving pulses in the IR spectral range is limited to photon energies up to, typically, $\simeq$ 100 eV, thus preventing the possibility to address and investigate electronic dynamics initiated by excitation or ionization of core electrons. In this context the development of mid-IR driving source is motivated by the favorable scaling of the high harmonics cut-off energy on the wavelength of the driving pulse, which holds the promise to give access to keV isolated attosecond pulses [@SCIENCE-Popmintchev-2012].
Finally, several experimental techniques, such as photoelectron microscopy, and coincidence photoelectron and photoion spectroscopy, require moderate pulse energy, but at (very) high repetition rates in order to overcome space-charge effects and to improve the signal-to-noise ratio. Recent technological developments indicate that generation of isolated attosecond pulses at MHz repetition rate should be within reach in the next few years [@NATPHOT-Krebs-2013].
The advent of Free Electron Lasers operating in the XUV/X-ray spectral range, however, has already allowed to overcome some of these limitations, and new opportunities for time-resolved studies will be offered by the continuous improvement of their characteristics (for example through the implementation of different seeding schemes that are expected to improve the shot-to-shot reproducibility and the longitudinal coherence [@NATPHOT-Thompson-2010]). Apart from these new developments, the novel European Extreme Light Infrastructure Attosecond Light Pulse Source (ELI-ALPS) holds the promise to overcome several of the limitations currently affecting HHG-based sources, by offering users with an unprecedented combination of attosecond XUV sources in terms of pulse energy, photon spectral range and repetition rate. Moreover, the facility will offer the possibility to combine XUV radiation with pulses extending from the THz region up to the X-ray spectral range and with particle beams.
The manuscript is organized in the following way:\
in section \[sec2\], we provide an overview of the main development directions ongoing in laser technology for the generation of attosecond pulses at high-repetition rate ($>$ 10 kHz) (\[sec21\]), with high energy per pulse (in the $\mathrm{\mu J}$ range) (\[sec22\]), and at high photon frequencies (hundreds of eV) (\[sec23\]). In section \[sec3\], we review the main spectroscopic techniques implemented in the attosecond field, in connection with a few selected experimental and theoretical works. Finally, in the last section, we introduce the new possibilities offered by the new attosecond facility ELI-ALPS for the investigation of atomic, molecular and cluster dynamics under intense and ultrashort XUV light pulses.
Current development directions in HHG {#sec2}
=====================================
Three main research directions can be identified in the development of HHG-based sources as schematically shown in Fig. \[Fig1\]: generation of attosecond pulses at high repetition rates (up to the MHz), generation of intense attosecond pulses for high-peak intensities, and extension of the cut-off energy in the keV range. These three directions call for different driving sources and will be discussed separately in the next sub-sections.
![The three main research directions in HHG sources are aiming for the development of high-repetition rate, high intensity, and high photon energy XUV pulses. The shaded area represents the parameter-space spanned by typical HHG-based XUV sources. State-of-the-art experimental setups will be discussed in the corresponding sections.[]{data-label="Fig1"}](Fig1.pdf){width="10cm"}
HHG at high-repetition rates {#sec21}
----------------------------
The generation of high-order harmonics at high repetition rates (0.1 - 100 MHz) represents a milestone for atomic and molecular dynamics applications, such including photoion and photoelectron spectroscopy [@IEEE-Sansone-2012] and for condensed phase and nanoplasmonics investigations using time-resolved photoelectron microscopy [@RSI-Mikkelsen-2009]. Multi-kHz HHG was realized in 2003 using a titanium:sapphire (Ti:Sa) amplifier which delivered pulses with energies up to 7 $\mathrm{\mu J}$ at 100 kHz repetition rate [@PRA-Lindner-2003], and ,recently, the first full temporal characterization of isolated attosecond pulse up to 10 kHz was reported [@RSI-Sabbar-2014].
{width="12cm"}
For the operation of HHG up to the MHz regime, the efficient conversion from the IR to the XUV spectral range requires high intensities ($\mathrm{I>10^{13} W/cm^2}$), which cannot be easily reached using directly the output of a femtosecond oscillator. A high finesse enhancement cavity offers a strategy for increasing the energy, and thus intensity, per pulse whilst maintaining the high repetition rate of the mode-locked oscillators The coherent addition of pulses leads to intracavity average power and pulse energies on the order of several tens or hundreds of Watts and a few $\mathrm{\mu Js}$, respectively. A gas jet for HHG is introduced in the enhancement cavity leading to harmonic generation at MHz repetition rates [@NATURE-Gohle-2005; @PRL-Jones-2005].
In the first experiments, Ti:Sa-based oscillators were used demonstrating generation of odd harmonics up to the 13th order ($\mathrm{\sim 20~eV}$) from the fundamental radiation. The harmonic radiation was extracted from the cavity using a reflective plate oriented at the Brewster angle for the fundamental radiation. This, however, limits the performances of the enhancement cavity due to the introduction of additional dispersion and induced nonlinearities. At the same time, the low-energy per pulse offered by Ti:Sa oscillator limits the intracavity total power to about 38 W [@NATURE-Gohle-2005].
In order to overcome these limitations, Pupeza and coworkers introduced a new Yb-based experimental setup consisting of an 1040-nm oscillator and a chirped-pulse fibre amplifier (output power of 60 W, 78 MHz repetition rate, and 172 fs pulse duration). The pulse duration could be further reduced to 51 fs by using an additional nonlinear compression stage based on a large mode area fibre and a chirped mirror compressor set. Such pulses were coupled in an enhancement cavity where the XUV radiation, generated in the gas jet, was coupling out of the cavity by a custom-designed pierced mirror. This output coupler reduced the intracavity losses and led to circulating power up to 5.6 kW, which resulted in a high conversion in the XUV spectral range up to a few $\mathrm{\mu W}$ for a single harmonic around 32 eV. At the same time, the coupling efficiency increased with decreasing wavelength which favored the generation of high photon energies. Using neon as generating medium, the harmonic cut-off was extended up to 100 eV.
The effect of the pulse duration on the efficiency of the intracavity HHG was investigated by this experimental setup. In general, it is well known that, for laser intensities below the ionization saturation threshold, HHG efficiency increases by decreasing the duration of the driving pulses; moreover, the use of short femtosecond pulses (down to the few-cycle regime) is a prerequisite for the generation of isolated attosecond pulses.In general, for a fixed peak intensity, shorter pulses reduces the ionization of the gas target due to the reduced duration of the interaction. The use of short pulses is expected to improve the performances of the enhancement cavity as ionization-induced nonlinearities constitutes one of the limiting factors for the average power circulating inside the cavity. This was experimentally demonstrated by comparing cavity peak intensity achieved with and without gas injection for two different pulse durations (57 and 172 fs) (see Fig. 3 in ref. [@NATPHOT-Pupeza-2013]).\
Recent developments in optical parametric amplification (OPA) and optical parametric chirped-pulse amplification (OPCPA) systems based on fiber and thin-disk technology have pushed the frontier of the generation of intense, ultrashort femtosecond pulses at high-repetition rates. They are ideal driving sources for HHG and could determine a breakthrough in the generation and application of ultrashort XUV radiation [@IEEE-Sansone-2012]. Krebs and coworkers have recently demonstrated the generation of XUV continua up to 600 kHz using a CEP-stable OPCPA system [@NATPHOT-Krebs-2013]. The system was based on two OPA stages delivering a smooth output spectrum spanning from 750 nm to 1,250 nm, seeded by the output of Ti:Sa CEP-stable oscillator. The two stages were pumped by a Yb-fibre laser amplifier delivering frequency-doubled 100 $\mathrm{\mu J}$, 500 fs pulses at the wavelength of 515 nm, with a repetition rate up to 600 kHz. The compressed pulse energy was 14 $\mathrm{\mu J}$ with a pulse duration of 6.6 fs. Figure \[Fig2\] shows the XUV spectra generated as a function of the CEP in neon where a clear CEP-dependence can be observed (Fig. \[Fig2\]a) which confirms the short duration of the driving pulses in the generating medium. A CEP change of $\pi/2$ drives a transition between a highly modulated spectrum (Fig. \[Fig2\]b) and a continuous distribution (Fig. \[Fig2\]c), which is a prerequisite for the generation of an isolated attosecond pulse.
Coherent combination techniques are expected to dramatic increase the average power of fiber-based systems in the coming years [@IEEE-Limpert-2014]. Pulses as short as 30 fs at an average power up to 163 W were achieved by combining a four-channel fiber chirped-pulse amplification system [@OL-Klenke-2013] with an external compressor, based on a hollow-fiber capillary, and a set of chirped mirrors [@OL-Hadrich-2013], . Phase-matched HHG in xenon driven by this source led to the demonstration of unprecedented average power in the XUV spectral range up to $\mathrm{143 \mu W}$ at 30 eV [@NATPHOT-Hadrich-2014].
HHG for high XUV intensities {#sec22}
----------------------------
HHG in gas targets is characterized by a low-conversion efficiency on the order, typically, of $10^{-9}-10^{-5}$, depending on the energy range and on the particular experimental conditions [@NATPHOT-Sansone-2011]. Investigation of nonlinear effects on the attosecond timescale in the XUV spectral range (for example two-photon double ionization of helium [@PRA-Ishikawa-2002]) requires intensities of $\mathrm{I=10^{14}-10^{15}~W/cm^2}$, corresponding to pulse energy in the $\mu$J range. This requires driving laser pulses of a few tens or hundreds of mJ for the generation and application of trains and isolated attosecond pulses for the characterization of nonlinear dynamics in atoms and molecules [@PQE-Midorikawa-2008]. The first pioneering experiments focused on the demonstration of attosecond nonlinear metrology [@NATURE-Tzallas-2003] by measuring the interferometric autocorrelation of an attosecond pulse train [@NATURE-Tzallas-2003; @PRL-Nabekawa-2006a], which showed the $\pi$-phase flip occurring between consecutive attosecond pulses [@PRL-Nabekawa-2006b]. Since then, the field has rapidly progressed towards the generation, characterization, and application of intense isolated attosecond pulses to XUV-pump-XUV-probe approaches. While trains of attosecond pulses are usually generated by multicycle driving fields, additional techniques are required for confining in time the HHG process, allowing for the creation of an isolated attosecond pulse [@NATPHOT-Chini-2014]. It is important to point out that while the first demonstration of isolated attosecond pulses was based on the use of few-cycle driving fields [@NATURE-Hentschel-2001], recently the creation of isolated attosecond waveform using multicycle driving pulses was demonstrated [@PRL-FENG-2009]. Tzallas and coworkers showed, for the first time, the time-resolved characterization of electronic dynamics in atoms [@NATPHYS-Tzallas-2011] and molecules [@PRA-Carpeggiani-2014] with a resolution down to 1 fs by implementing an XUV-pump-XUV-probe approach and using the interferometric polarization gating (IPG) technique [@NATPHYS-Tzallas-2007].
A schematic view of the experimental apparatus used in those investigations in shown in Fig. \[Fig3\]a:
![a) Experimental setup for generation and application of intense attosecond XUV pulses. IPG: interferometric polarization gating; GJ1 and GJ2 : gas jets; SM: split mirror. b) XUV continuum generated in Xenon. c) TOF mass spectra generated by the focused XUV pulse. Reprinted by permission from Macmillan Publishers Ltd: Nature Physics ref. [@NATPHYS-Tzallas-2011], copyright (2011).[]{data-label="Fig3"}](Fig3.pdf){width="12cm"}
the high-energy driving pulse was focused in a first gas jet (GJ1) for efficient generation of harmonics. Loose focusing geometries are typically employed for increasing the interaction volume, and hence the total conversion efficiency. Figure \[Fig3\]b shows that IPG implementation results in the generation of a supercontinuum. The fundamental IR radiation was then separated by the co-propagating harmonic radiation by using a reflection off a silica plate at the Brewster angle. In this configuration, an additional metallic filter was needed in order to eliminate the residual IR radiation reflected by the plate. The attosecond pulse was focused into a second gas jet (GJ2) by means of a split mirrors (SM) that operated also as delay stage to delay one part of the XUV beam with respect to the other [@JOSAB-Kolliopoulos-2014]. The ions generated in the interaction region were finally collected by a Time-of-Flight (TOF) mass spectrometer. Due to the high-intensity of the XUV beam, nonlinear ionization occurred as shown in Fig. \[Fig3\]c, which reports the mass ion spectra measured in xenon. The observation of $\mathrm{Xe^{2+}}$ clearly indicates the nonlinear ionization of xenon atoms in the jet.
Finally, the temporal structure of isolated attosecond pulses with 1.3 $\mathrm{\mu J}$ energy has been characterized using a nonlinear autocorrelation method [@NATCOMM-Takahashi-2013]. The peak power of this source (2.6 GW) makes it a feasible alternative to XUV FEL for the investigation of nonlinear dynamics in the XUV spectral range.\
An alternative approach for the generation of intense attosecond pulses is based upon HHG using solid targets. This overcomes the ionization saturation of the generating medium that limits the driving pulse energy of HHG in gas targets. In surface-based HHG, the intense driving pulse creates a plasma at the interface, which is responsible for the generation of the high-order harmonics in the direction of the reflected beam [@NJP-Tsakiris-2006]. There are not any limitations to the maximum intensity of the laser field that can be applied. For relativistic intensities, the properties of the emitted radiation are independent from the target characteristics but solely depend on the properties of the electronic plasma created by the pulse [@PRE-Baeva-2006].
Depending on the intensity of the driving field, surface HHG can be described in terms of coherent wake emission from the plasma waves excited in the electronic plasma ($\mathrm{I<10^{18} W/cm^2}$) [@PRL-Quere-2006], or as the Doppler upshifted radiation created by the relativistic motion of the electronic plasma under the influence of the external laser field ($\mathrm{I>10^{18}~W/cm^2}$) [@NATPHYS-Dromey-2006]. Harmonic radiation using the second process up to 3.8 keV was demonstrated [@PRL-Dromey-2007]. The conversion efficiency associated to the process is also orders of magnitude greater than HHG in gas targets [@NJP-Tsakiris-2006]. In particular, the expected conversion efficiency of a few percent in the low XUV photon energy range creates the possibility to generate mJ-level attosecond trains and isolated attosecond pulses using table-top laser system [@NJP-Rykovanov-2008; @PRL-Heissler-2012].
Temporal characterization of attosecond pulse trains generated on solid targets was recently demonstrated by implementing a second-order autocorrelation [@NATPHYS-Nomura-2009; @NJP-Horlein-2010]. The generation and characterization of isolated attosecond pulses using this mechanism would represent a major breakthrough for nonlinear, keV attosecond spectroscopy [@NJP-Tsakiris-2006].
HHG up to the keV spectral range {#sec23}
--------------------------------
Equation \[Eq1\] describes the maximum photon (cut-off) energy reached in HHG in gas targets: $$E=I_p+3.17U_p=I_p+3.17\frac{e^2}{8\pi^2 \varepsilon_0 m c^3} I\lambda^2
\label{Eq1}$$ where $I_p$ is the ionization potential, $U_p$ the ponderomotive potential, $e$ the electron charge, $m$ the electron mass, $c$ the speed of light, $I$ the pulse intensity, and $\lambda$ the field wavelength. The generation and application of trains and isolated attosecond pulses have, so far, been mainly limited to energies below 150 eV. In order to extend the cut-off energies up to the water window (280-530 eV) and to the keV range, Eq. \[Eq1\] suggests that an increase the intensity or the wavelength of the driving pulse is required but this formula does not incorporate saturation effects. These limit the cut-off photon energy for driving intensities exceeding the saturation intensity which is the intensity required for full ionization of the generating medium. Increasing the intensity of the pulses beyond the saturation intensity leads to the complete ionization of the medium on the leading edge of the pulse. Under this conditions, the medium does not experience the maximum electric field intensity leading to a reduction of the cut-off photon energy with respect to the prediction of Eq. \[Eq1\].
![X-ray spectra generated by a multi-mJ mid-IR driving source operating at 3.9 $\mathrm{\mu m}$. The bandwidth of the coherent radiation extends over more than 700 eV supporting a Fourier limited duration of 2.5 as. Phase-matched generation at this driving wavelength allows one to extends the cut-off energy up to 1.6 keV, much higher with respect to the cut-off achievable with shorter driving wavelengths: the XUV spectra generated by 800-nm, 1.3-$\mathrm{\mu m}$ and 2.0-$\mathrm{\mu m}$ fields are shown for comparison. From ref. [@SCIENCE-Popmintchev-2012]. Reprinted with permission from AAAS.".[]{data-label="Fig4"}](Fig4.pdf){width="12cm"}
The $\lambda^2$-scaling offers more advantage for the extension of the harmonic cut-off, and it has driven the development of mid-IR driving sources during the last years. Currently, driving pulses with central wavelength between 1.4 and 3.9 $\mu m$ have been demonstrated. These sources typically exploit nonlinear effects, such as difference-frequency generation, for the generation of low-energy pulses in the mid-IR, which are then amplified in nonlinear optical parametric amplifiers. Using this approach, CEP-stable pulses can be generated [@OL-Vozzi-2007] and used for harmonic generation up to 200 eV in a two-color scheme [@OL-Calegari-2009] and the CEP-dependence of the HHG spectra in the water window generated by few-cycle mid-IR has recently been reported [@NATCOMM-Ishii-2014].
The possibility to extend the generation of coherent radiation up to the keV spectral range was recently demonstrated [@SCIENCE-Popmintchev-2012]. In this work Popmintchev and coworkers used 3.9 $\mathrm{\mu m}$ with an energy of 10 mJ at a repetition rate of 20 Hz to drive the HHG process in a waveguide filled with helium at pressures as high as 35 atm. Phase-matched harmonic generation was achieved by balancing the dispersion introduced by the neutral gas and by the free-electron plasma generated by the driving field. The geometrical dispersion imposed by the guiding structure can contribute to the achievement of phase-matching condition. Figure. \[Fig4\] reports the experimental spectrum measured after the waveguide using a soft x-ray spectrometer and an x-ray charge coupled device camera. The X-ray spectrum extends up to 1.6 keV indicating the generation of nonlinear harmonic orders larger than 5000. The coherence properties of the emerging beam were confirmed by a Young’s double slit apparatus. In these experimental conditions, the main characteristics of the HHG process differs from IR driving laser induced HHG. Helium, despite its low polarizability, is the most efficient medium for HHG up to the keV range as the lack of core electrons prevents the re-absorption of the generated radiation and does not represent a limitation for the coherent buildup of the harmonic signal. At the same time, the high pressure required for the phase-matching condition and the long driving wavelength, determines a mean interatomic distance (about 10 $\mathrm{\r{A}}$) which is much smaller than the extension of the electronic wave function during its motion in the continuum (about 500 $\mathrm{\r{A}}$). In the case of 800-nm wavelength driving pulses, the typical extension of the electron trajectory in the continuum is 20 $\mathrm{\r{A}}$, while the interatomic separation (at 0.1 atm) is 70 $\mathrm{\r{A}}$. It can be therefore assumed that the electronic wave packet does not encounter any other atom (or ion) during its motion. This assumption is no longer valid in the case of long-driving wavelength and high pressure, but, nevertheless the scattering of the electronic wave packet does not appreciably affect the coherence properties of the X-ray beam. The efficient generation of multicycle mid-IR pulses in waveguide is expected to lead to the confinement of the HHG process to short time windows, eventually leading to an isolated attosecond pulse [@PNAS-Chen-2014]. This property is again the result of the phase-matching conditions occurring in the capillary during the propagation of the driving pulse.
Temporal confinement of the HHG process could lead even to the generation of zeptosecond pulses as recently predicted [@PRL-Garcia-2013]. The zeptosecond temporal structure arises from the interference of high harmonic emission from multiple reecounters of the photoelectron wave packet released in the continuum with the parent ion. In HHG the contributions of quantum paths revisiting more than once the original position can be neglected due to the reduced recombination probability as the result of wave packet spreading. However, on the trailing edge of the laser pulse the smaller recombination probability is compensated for by the higher ionization rate, which occurs closer to the peak of the mid-IR pulse. This leads to comparable amplitudes from trajectories that revisit possibly several times the parent ion. It is important to observe that for even longer driving wavelength ($\lambda > \mathrm{9~\mu m}$), the action of the magnetic field on the electron paths will start playing a role, introducing a new element that could affect the recombination probability with the parent ion.
XUV-driven ultrafast dynamics: a few examples {#sec3}
=============================================
Imaging and control of two-electron wave packet {#Fano}
-----------------------------------------------
The time-resolved observation and control of electronic dynamics driven by electron-electron correlation is one of the main goals of ultrafast XUV science, and, in particular, of attosecond physics [@RMP-Krausz-2009; @CHEMPHYSCHEM-Sansone-2012].Helium is a fundamental benchmark, as it represents the simplest two-electron system and the investigation of its Fano resonances is a fundamental step towards the understanding of correlated electrons dynamics. Fano resonances are peculiar resonant structures with an asymmetric shape which are markedly different from the common lorentzian lineshape. In atoms, Fano resonances appear in the presence of a configuration state lying above the ionization threshold. The asymmetric shape is explained as the result of the interference between two indistinguishable pathways leading to the same final state into the continuum: direct ionization or excitation and subsequent decay from the (metastable) configuration state. These resonances were first explained by Ugo Fano in 1961 in his seminal paper [@PR-Fano-1961]. Several investigations since have focused upon the time-resolved observation of the resonances dynamics and have led to the time-resolved estimation of the autoionization time by measuring the time duration of the photoelectron peak corresponding to the IR-generated sideband of the resonance, as a function of the relative delay between the isolated attosecond and IR pulse [@PRL-Gilbertson-2010]. The broadband attosecond pulses can lead to the excitation of several autoinizing resonances, leading to the formation of an excited two-electron wave packet that evolves on a few femtosecond or even attosecond timescale. This correlated electronic motion has been recently observed by Ott. *et al* [@NATURE-OTT-2014] in a transient absorption experiment, where the observable is provided by the XUV spectrum transmitted through a thin gas cell filled with helium. A moderately intense IR pulse was used to couple the XUV-populated $2s2p$ and $sp_{2,3+}$ states via two IR photons. The difference in the phase evolution between the two components manifests itself in the periodic modulation of the transmitted XUV spectrum, as it is shown in Fig. \[Fig5\]. The measured data (Fig. \[Fig5\]a) are in excellent agreement with the two-electron wave packet evolution simulated using either *ab-initio* TDSE calculations or a reduced model (Fig. \[Fig5\]b,c) [@NATURE-OTT-2014]. From the comparison between the data and the theoretical expectation, it is possible to retrieve the relative phase between the two components of the two-electron wave packet (Fig. \[Fig5\]d,e) and, therefore, reconstruct the correlated electronic dynamics as shown in Fig. \[Fig5\]f. Only the relative phase can be reconstructed using this approach and prior knowledge of the wave function is required in order to visualize the spatial distribution of the wave packet. Nevertheless, there is excellent agreement between theoretical expectation and experiment.
![Time-resolved observation of a two-electron wave packet in helium. a-c: helium XUV light absorbance ($\Delta A(\omega,\tau)$) versus time delay; experiment (a), few-level model simulation (b) and ab initio calculation (c). d: oscillation of $\Delta A(\tau)$ near $sp_{2,3+}$ resonance at 63.67 eV. e: phase of $\Delta A(\tau)$, $\varphi_{A}(\tau)$, and relative phase $\varphi(\tau)$ of the two-electron wave packet involving the $2s2p$ and $sp_{2,3+}$ states. f: visualization of the two-electron wave-packet motion. Left column: experimentally reconstructed wave packet, including only the two measured states $2s2p$ and $sp_{2,3+}$. Right column: *ab initio* TDSE simulation, including all excited states. Reprinted by permission from Macmillan Publishers Ltd: Nature ref. [@NATURE-OTT-2014], copyright (2014).[]{data-label="Fig5"}](Fig5.pdf){width="12cm"}
Time-resolved dynamics of Fano resonances
-----------------------------------------
Direct imaging of a correlated electronic wave function requires the simultaneous collection of information regarding the different degrees of freedom of the two particles, for example the momenta resulting from the ionization of the system. The time-resolved evolution of a Fano resonance profile highlighting the dynamics associated to the formation of the characteristic lineshape was studied in ref. [@PRA-Chu-2010]. In that work, an experimental scheme for the time-resolved investigation of this process was also indicated, as shown in Fig. \[Fig6\]. The proposed example was beryllium (electronic ground state $2s^2$), excited to the doubly excited states $2pns$ (with $n=3$ to $n=9$). These states are embedded in the $2sEp$ continuum and autoionize on a timescale of a few femtoseconds. The wave function of the system can be described as: $$\psi(t)=\sum_n d_n(t)\ket{2pns}+\int d_E(t)\ket{2sEp} dE
\label{wavepacket}$$ where $d_n(t)$ and $d_E(t)$ are time-dependent coefficient of the bound and continuum part of the electronic wave packet. In proposed scheme, the evolution of the absorption profile can be reconstructed by using the combination of two attosecond (or few-femtosecond) pulses for exciting and probing the dynamics. The first pulse creates the wave packet described by Eq. \[wavepacket\] and thus starts the autoionization mechanism (see Fig. \[Fig6\]). The second synchronized probe pulse projects a fraction of the wave packet $\int d_E(t)\ket{2sEp} dE$ onto states of the $\ket{E'_pE_p}$ double ionization continuum, corresponding to the ejection of the $2s$ electron and resulting in the creation of a doubly charged ion $\mathrm{Be^{2+}}$. Since these are stationary states of the dication, the coefficients describing their evolution, $d_{E'E}(\tau)$, depend only on the delay $\tau$ and do not evolve further in time (differently from $d_n(t)$ and $d_{E}(t)$, because of the autoionization process). Thus, sampling the observable $d_{E'E}(\tau)$ at different delays allows for the reconstruction of time-resolved absorption profile, which, for large delays, converges to the expected Fano shape [@PRA-Chu-2010].\
It is necessary to measure in coincidence the two electrons or one electron and the double charged ion $\mathrm{Be^{2+}}$ in order to characterize the $d_{E'E}(\tau)$ function. This experimental approach challenges current experimental possibilities in attosecond science as it requires intense XUV sources (a XUV-pump-XUV-probe approach needs to be implemented) at high repetition rate to generate statistically meaningful data. Suitable attosecond sources are currently not available but these experiments will become feasible with newly developed large scale facilities XUV photon sources such as FELs and ELI-ALPS.
![Concept of the scheme for probing the time-dependent Fano resonance profile proposed in [@PRA-Chu-2010]. Autoionization starts at t=0, when the pump pulse is over. To determine the evolving electron wave packet at time $\tau$, a probe pulse is used to ionize the 2s electron, creating the $E'_p E_p$ states. After the probe is over, the new wave packet consists of a part of the old wave packet which continues to evolve and autoionize (upper row) and the newly generated part made of $E'_p E_p$ states (lower row). The latter does not change with time. By collecting the energy-resolved double-ionized signals, the resonance profile in the autoionization at time $\tau$ is captured. Figure reproduced with permission from ref. [@PRA-Chu-2010], copyright (2006) APS.[]{data-label="Fig6"}](Fig6.pdf){width="12cm"}
Time-resolved dynamics of Auger processes
-----------------------------------------
The interaction of an XUV or X-ray pulse (from several tens of eV up to a few keV) can lead to the formation of highly charged ions, through a mechanism initiated by emission of an inner-valence or core-shell electron leaving the ion in an excited state. The initial hole can be filled through an Auger decay process leading to the emission of a second electron and the formation of doubly charged ions. Depending on the external photon flux, a second photon can be eventually absorbed from the same (or a different) inner-valence or core-shell triggering a second Auger decay process, and so on. As a result the ionization potential increases at each step until it exceeds the XUV/X-ray photon energy. At this point, further ionization can only occur by emission of valence electrons. For intense X-ray pulses in the keV, the interplay between these mechanisms was investigated in ref. [@NATURE-Young-2010].
The core-shell ionization process can also occur according to a non-sequential mechanism for pulse durations comparable to the Auger decay lifetime. In this case, the second XUV (X-ray) photon can be absorbed when the first hole has not yet been filled (before the first Auger decay), leading to the formation of a double core hole. These states decay with the emission of an electron whose energy is displaced with respect to the energy of the single Auger decay. This gives a very specific fingerprint process which could be used for chemical analysis [@PNAS-Berrah-2011]. Moreover, X-ray fluorescence from a double core hole state was recently used to infer information about the duration of an intense X-ray pulse in the few-femtosecond domain [@PRL-Tamasaku-2013].
Time-resolved observation of Auger decay on the few-femtosecond timescale was first reported in ref. [@NATURE-Drescher-2002] where core-level ionization was triggered by the absorption of a single photon from an isolated attosecond pulse and the lifetime of of the relaxation process was determined with an analogous technique used in the determination of Fano resonance lifetime [@PRL-Gilbertson-2010] (see section \[Fano\]). In the XUV/IR pump probe experiment sidebands of the main Auger-electron line are created and the duration of the sidebands as a function of the relative delay between the IR and XUV pulses allows one to estimate the decay time of the core-hole state.
An alternative approach based on the measurement of the ion-charge state and yield was demonstrated in ref. [@NATURE-Uiberacker-2007], by monitoring the ion yield for different charge states as a function of the relative delay. The scheme can be understood considering Fig. \[Fig7\]: after the Auger decay process, $\mathrm{Kr^{3+}}$ can be populated by IR-ionization of $\mathrm{Kr^{2+}}$ ion states. The formation of these ionic states can occur only after the Auger decay process and presents a rise-time that reflects the autoionization dynamics. In the measurement shown in Fig. \[Fig8\], the increase of the signal presents a rise-time of $\tau_{A1}=7.9$ fs, which is consistent with the value obtained from the sidebands duration [@NJP-Uphues-2008].
HHG-based sources offer also the possibility to time-resolve the cascaded Auger decay process. Indeed, after the first Auger decay, a sequential second decay can occur which leads to the emission of a third electron. The associated dynamics were probed in time by measuring the time evolution of xenon ion yield as a function of the relative delay between the initial attosecond pulses and a delayed few-cycle infrared pulse [@NATURE-Uiberacker-2007]. While ion chronoscopy reveals integral information regarding the cascaded Auger dynamics, photoelectron spectroscopy allows the assignment of intermediate states and resonances in the process. This approach was exploited in [@NJP-Verhoef-2011] to investigate the resonant and normal Auger decay in Krypton excited by an isolated attosecond pulse centered at 94 eV. The XUV-pump/IR-probe time-resolved measurement indicated a lifetime of about 70 fs in the relaxation dynamics of the excited ionic states in the cascaded resonant Auger decay. Comparison with theoretical models led to the attribution of this lifetime to the second-step Auger decay of the resonantly excited $3d^{-1}np$ states with $n=6,7$.\
![Schematic representation of energy levels and transitions in neutral, singly, doubly and triply ionized krypton. After the $3d$ excitation by the attosecond XUV pulse (magenta arrows), different relaxation channels (green arrows) open up: resonant and normal Auger decay as well as cascaded Auger decay. Figure reproduced with permission from ref. [@NJP-Uphues-2008], copyright(2008) IOP.[]{data-label="Fig7"}](Fig7.pdf){width="10cm"}
![Ion chronoscopy determination of the lifetime of Auger decay in krypton. $\mathrm{Kr^{3+}}$ signal is plotted against the XUV/IR delay. The rise time allows for the estimation of the lifetime, since the formation of $\mathrm{Kr^{3+}}$ ions is made possible by the building up (and subsequent IR-ionization) of the $\mathrm{Kr^{2+}}$ population resulting from the Auger process. Figure reproduced with permission from ref. [@NJP-Uphues-2008], copyright(2008) IOP.[]{data-label="Fig8"}](Fig8.pdf){width="10cm"}
The Extreme Light Infrastructure Attosecond Light Pulse Source {#sec4}
==============================================================
The Attosecond Light Pulse Source (ELI-ALPS) is one of the three pillars of the Extreme Light Infrastructure (ELI), a new EU research initiative that has cutting edge, future oriented light based fundamental science and technology as its primary mission. ELI-ALPS will provide some of the shortest and most powerful light pulses with the highest repetition frequency, all within XUV/X-ray wavelength range. The short duration of the individual XUV/X-ray light pulses, $<$ 1 fs and the intrinsic attosecond synchronization with other light fields ranging from the THz over the IR/mid-IR into the visible, and bursts of elementary particles (electrons and ions) distinguishes ELI-ALPS from alternative high-brightness X-ray sources such as FELs and accelerators. To the scientist ELI-ALPS will offer an unique possibility to explore non-linear processes at and beyond the XUV, to observe and control ultrafast electronic and structural dynamics of atoms, molecules, clusters, liquids and solids, and to investigate new ideas for biological research, material science and medical applications. ELI-ALPS will maintain state-of-the-art end stations including reaction microscope (REMI), cold-target recoil ion spectrometer (COLTRIMS), electron/ion 3D momentum imaging and ion microscopy, velocity map imaging (VMI), angle-resolved photoemission spectrometer and photoelectron emission microscope. In addition, there will be the possibility of installing specialized apparatus.
The research equipment of ELI-ALPS is schematically shown in Fig. \[Fig9\]. It will feature four main driving laser sources (primary sources), which will drive XUV and X-ray photon beamlines and particles (electron and ions) beamlines (secondary sources). Different combination of primary and secondary sources will be delivered to experimental end stations for time-resolved experiments (experiments/user stations).
Table \[table1\] summarizes the main characteristics of the primary sources. Due to the challenging specifications, which are well-beyond the state-of-the-art, the implementation of these sources has been planned in two different phases (Phase I and Phase II). The characteristics for Phase I have already been consolidated whilst the expected, final performances for the second implementation phase depends on the outcome of the first phase.
The high-repetition rate (HR) primary source will operate at 100 kHz and will drive two XUV photon beamlines based on HHG in gas targets (GHHG) These secondary sources will provide users with trains and isolated attosecond pulses for time-resolved experiments. The expected performances for the XUV photons secondary sources are summarized in Table \[table2\].
The single cycle laser (SYLOS) will favor energy at the expenses of repetition rate, operating at 1 kHz for a pulse energy of 100 mJ. This primary source will drive two GHHG beamlines, one beamline for the generation of harmonics from solid targets (SHHG beamlines), and one beamline for the generation and acceleration of electron beam to be used in pump-probe experiments in combination with other secondary or primary beams.
Finally, the high-field (HF) source will deliver pulses with 2 PW peak-power for the investigation of plasma physics and for exploring the potential of surface harmonic generation through a dedicated SHHG beamline (x-ray).\
The availability of light pulses with duration well below 1 fs has opened the possibility to study the pure electronic dynamics of valence electrons in atoms and molecules. In analogy to the familiar creation of vibrational wave packets in femtospectroscopy, attosecond valence electron spectroscopy aims at creating electronic wave packets on a time scale that is considerably faster than relaxation or nuclear motion in molecules. The evolution of such wave packets can lead to localization of charge at specific locations of the molecules and thus open chemical reaction pathways that would not be favored in a conventional chemical reaction [@NATPHOT-Lepine-2014]. Understanding the dynamics of such wave packets is subject to recent attosecond studies. Experimentally, the problem is approached by either two color pump-probe schemes, where one pulse is of attosecond duration while the other may be femtoseconds long, or directly by attosecond pump-probe schemes consisting of identical XUV replica. In the former case, the challenge is not so much the pulse energy but rather the requirement for the tunability of the photon energies and bandwidths, a stable, well-defined and adjustable attosecond and femtosecond time structure, and the precise synchronization of the two-color field. An IR pulse is frequently used to probe the excitation either by attosecond streaking or by coupling distinct quantum paths in the ionization and dissociation process. Obviously, the synchronization with the attosecond pulse is crucial for the temporal resolution, while the IR/mid-IR photon energy determines the time range that can be accessed or the states which are coupled. All these requirements can be satisfied by the all-optical approach envisaged at ELI-ALPS. Shorter than 5 fs CEP stabilized IR pulses are promised at mJ pulse energies from which state-of-the-art high harmonic pulses in the few 10 nJ range and mid-IR pulses in the $\mathrm{10~\mu J}$ range are derived. What is more, this operation is performed at high repetition rates of up to 100 kHz. This unique feature is indispensable not only to achieve an unprecedented signal quality and thus statistically meaningful data, but will also enable access to higher order, multi-parameter detection schemes [@JPB-Rudenko-2010].\
An illustrative example for the capabilities of ELI-ALPS is the non-linear interaction of atoms with XUV pulses, introduced in section \[sec2\]B. The two-photon double ionization of helium is the most fundamental realization of an attosecond pump-probe experiment [@PRA-Ishikawa-2002]. The theory of this rich and involved three-body problem has matured but still awaits experimental testing [@JPB-Palacios-2010]. Thorough understanding of this system is essential for the design and the interpretation of experiments on more complex systems. Due to the stringent requirements on the pulse parameters [@JPB-Nikolopoulos-2001] it has only recently become possible to detect $\mathrm{He^{2+}}$ from a two-photon process induced by high-order harmonics [@PRL-Nabekawa-2005]. To date there is still no full kinematic data available owing to the very low reaction yield and the low repetition rate (typically less than 10 Hz) of the driving pulses available so far. In this experiment pulses of 10 fs and 24 nJ were used with $\mathrm{I_{peak}=10^{13} W/cm^2}$. For a comprehensive investigation of such processes ELI-ALPS will eventually generate sub-femtosecond pulses reaching intensities well beyond $\mathrm{10^{14}~W/cm^2}$. This will be the starting point for the thorough understanding of the dynamics and interferences of the correlated electrons predicted for delayed pulse experiments [@JPB-Palacios-2010; @PRA-Lambropoulos-2008] as a benchmark for all further experiments.\
![Schematics of the research equipment of the ELI-ALPS facility. The primary sources will be delivered to several beamlines, where secondary sources consisting in attosecond pulses, X-ray radiation, THz pulses, and particle beams are generated. These sources will finally drive experiments and user end stations in three different areas accordingly to the shielding characteristics required (Low-, medium-, and high-shielded target areas)[]{data-label="Fig9"}](Fig9.pdf){width="10cm"}
Typical photon energies for the excitation of small molecules such as $\mathrm{H_2}$ are around 15-35 eV, while the separation of subsequent excited electronic levels is on the order of 3-20 eV corresponding to 100-600 as pulses in the XUV. Even broader phase-locked XUV spectra are required if continuum states are to be included in the wave packet. A recent example that would substantially benefit from ELI-ALPS’s sources is the asymmetry of $\mathrm{D^{+}}$ ejection from two-color ionization of $\mathrm{D_2}$ that results from the steered electron localization by a timed XUV/IR field [@PRL-Fischer-2010; @NATURE-Sansone-2010]. It may be regarded as the original demonstration on how light-induced electron migration influences the outcome of a chemical reaction through the coupling of electronic and nuclear degrees of freedom. Related experiments have been undertaken by utilizing a XUV/XUV pump-probe scheme where the attosecond probe beam ionizes the evolving molecule at a preset delay time yielding detectable fragments [@PRA-Carpeggiani-2014]. In contrast to the XUV/IR scheme this approach is fully perturbative and insights into the intrinsic charge dynamics are obtained. The time scales for the simultaneous but clearly distinct electronic and nuclear wave packet motion reaches into the sub 1 fs region [@PNAS-Palacios-2014].
The interplay between the electronic and nuclear degrees of freedom beyond the Born-Oppenheimer regime is a largely unexplored but highly important topic [@PhysScripta-Muskatel-2009]. In larger molecules, the migration of the hole or electron wave packet following photoionization or photoexcitation can be probed in a similar manner and the yield of specific fragments indicates the charge distribution at the moment of detection [@SCIENCE-Calegari-2014]. In this way the time-dependent redistribution of the electronic density can be inferred. These processes occur on the time scale of 1-10 fs [@ChemPhys-Kuleff-2007]. Just as for the homo-nuclear molecules, correlated detection schemes and simultaneous acquisition of several observables will open new insights into the intramolecular charge dynamics of these larger molecules.
The high repetition rates at high to very high photon energies available at ELI-ALPS will also allow to venture into the study of charge transfer processes [@NATURE-Fohlisch-2005] and the dynamics of molecular orbitals that occur at the interfaces of solids. In some cases reorganization times well below 1 fs are foreseen, while in others no experimental data exists yet. Further examples include light-induced charge migrations and charge density waves (e.g. $\mathrm{TaS_2}$) [@OSA-Cacho-2012] and phase transitions in dielectrics (e.g. SiO2) [@NATURE-Schiffrin-2013] or carrier dynamics in graphene [@NATMAT-Gierz-2013]. Novel techniques may also make solid state phenomena accessible to attosecond spectroscopy [@CHEMPHYS-Ivanov-2013].\
What has been said about valence electron dynamics can directly be transferred to the time-resolved spectroscopy of core electron excitations. For these studies ELI-ALPS provides X-ray sources up to 10 keV similar to those in the XUV. The dynamics following removal of a core electron are decidedly more complex than for valence excitations but can be resolved in two- or one-color pump-probe schemes as has been demonstrated for I2 and $\mathrm{C_2H_2}$, respectively. Optically highly ionized atoms may be counted into this category [@JPB-Schippers-2014]. Experimental shortcomings due to timing jitter, pulse duration and pulse energy can easily be overcome by ELI-ALPS’s all-optical approach. The ultrafast relaxation of highly excited molecular states has important implications for the radiation resistance of biomolecules. An understanding of these mechanisms is still in its infancy but electronic correlations and proton transfer on the fs time scale appear as major factors.\
A highly demanding non-linear process is the two-photon ionization of inner shells. Estimates show that ELI-ALPS will be the first of its kind capable to provide these. The promise is the ability to study fundamental inner shell dynamics with the attosecond time resolution facilitated by optical pulses. Another utilization of core electrons is the holographic imaging of the nuclear configuration with a photoelectron ejected from the molecule itself. This method also relies on the energy selectivity and the timing accuracy attainable with ELI-ALPS’s secondary sources.
In the experiments discussed so far information about the dynamics of the electronic density is obtained only indirectly by watching specific parts of a molecule after cleavage or from spectral changes. This requires some knowledge of the nuclear and electronic structure at the time of observation. Spectroscopy is insensitive to electronic configuration changes that do not involve measurable differences in energy. A direct probing technique for electronic distribution dynamics is time-resolved X-ray diffraction and imaging techniques which are well established in femtosecond spectroscopy. It is clear that for the understanding of light-matter interactions and correlated motions to extend such measurements to the attosecond time scale with a spatial resolution that is compatible with molecular and atomic dimensions. This approach is known as 4-dimensional imaging. ELI-ALPS will be equipped to fill this need by offering reproducible X-ray pulses of 100 nJ down to $\mathrm{1-12~\mathrm{\r{A}}}$ along with synchronized attosecond pulses to initiate the dynamics of interest. Experiments planned at ELI-ALPS will include the light-field induced charge dynamics in ordered dielectrics and crystals [@NATURE-Schultze-2013; @NATURE-Schiffrin-2013], and photonic nanostructures at optical frequencies. Recent studies reveal that coherent motions of the coupled electronic-nuclear system play a hitherto underestimated role in the charge transfer of organic molecules [@SCIENCE-Falke-2014]. Spatial tracking of photo-initiated electron and proton motion in natural and artificial molecules (e.g. light harvesting complexes or photochromic molecules) is another area of broad scientific and technological interest. Their ultrafast time scales and their complex structure make the 4-dimensional approach indispensable. Dynamics in small gas-phase molecules [@PRL-Kupper-2014], larger functional molecules [@SCIENCE-Calegari-2014], molecules in the liquid-phase [@SS-Brown-2013; @PRL-Nordlund-2007], plasma formation [@PRL-Liseykina-2013] and plasma wave evolution [@PRL-Varin-2012] in atomic nano-clusters, and metallic nanostructures [@NATPHOT-Stockman-2007; @NATURE-Aeschlimann-2007] are further examples that make the light based approach a convincing step. ELI-ALPS will also ambitiously pursue the extension of this method towards shorter wavelengths on the order of 0.1 $\mathrm{\r{A}}$ and to true 3-dimensonal spatial imaging. This will allow to follow the charge density dynamics of core electrons and to obtain the full, unambiguous set of coordinates for the electronic and nuclear positions in time.
Acknowledgments
===============
Financial support by the Alexander von Humboldt Foundation (Project “Tirinto”), the Italian Ministry of Research (Project FIRB No. RBID08CRXK). This project has received funding from the European Union$\mathrm{'}$s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 641789 “MEDEA” (Molecular Electron Dynamics investigated by IntensE Fields and Attosecond Pulses) and the Hungarian Scientific Research Fund (OTKA project NN 107235) . We thank Prof. Dimitris Charalambidis for fruitful discussions and continuous support.
e-mail: [email protected]\
References {#references .unnumbered}
==========
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In many practical uses of reinforcement learning (RL) the set of actions available at a given state is a random variable, with realizations governed by an exogenous stochastic process. Somewhat surprisingly, the foundations for such sequential decision processes have been unaddressed. In this work, we formalize and investigate *MDPs with stochastic action sets (SAS-MDPs)* to provide these foundations. We show that optimal policies and value functions in this model have a structure that admits a compact representation. From an RL perspective, we show that Q-learning with sampled action sets is sound. In model-based settings, we consider two important special cases: when individual actions are available with independent probabilities; and a sampling-based model for unknown distributions. We develop poly-time value and policy iteration methods for both cases; and in the first, we offer a poly-time linear programming solution.'
author:
- |
Craig Boutilier, Alon Cohen, Amit Daniely, Avinatan Hassidim,\
**Yishay Mansour, Ofer Meshi, Martin Mladenov, Dale Schuurmans**\
Google Research, Mountain View, CA, USA\
bibliography:
- 'long.bib'
- 'standard.bib'
title: Planning and Learning with Stochastic Action Sets
---
Introduction {#sec:intro}
============
Markov decision processes (MDPs) are the standard model for sequential decision making under uncertainty, and provide the foundations for reinforcement learning (RL). With the recent emergence of RL as a practical AI technology in combination with deep learning [@mnih2013; @mnih2015], new use cases are arising that challenge basic MDP modeling assumptions. One such challenge is that many practical MDP and RL problems have *stochastic sets of feasible actions*; that is, the set $A_s$ of feasible actions at state $s$ *varies stochastically* with each visit to $s$. For instance, in online advertising, the set of available ads differs at distinct occurrences of the same state (e.g., same query, user, contextual features), due to exogenous factors like campaign expiration or budget throttling. In recommender systems with large item spaces, often a set of *candidate* recommendations is first generated, from which top scoring items are chosen; exogenous factors often induce non-trivial changes in the candidate set. With the recent application of MDP and RL models in ad serving and recommendation [@charikar:stoc99; @Li:adkdd2009; @archak-mirrokni-muthu:www10; @mirrokni:wine12; @kearns:uai12; @silver:icml13; @theocharous:ijcai15; @logisticMDPs:ijcai17], understanding how to capture the stochastic nature of available action sets is critical.
Somewhat surprisingly, this problem seems to have been largely unaddressed in the literature. Standard MDP formulations [@puterman] allow each state $s$ to have its own feasible action set $A_s$, and it is not uncommon to allow the set $A_s$ to be non-stationary or time-dependent. However, they do not support the treatment of $A_s$ as a stochastic random variable. In this work, we: (a) introduce the *stochastic action set MDP (SAS-MDP)* and provide its theoretical foundations; (b) describe how to account for stochastic action sets in model-free RL (e.g., Q-learning); and (c) develop tractable algorithms for solving SAS-MDPs in important special cases.
An obvious way to treat this problem is to embed the set of available actions into the state itself. This provides a useful analytical tool, but it does not immediately provide tractable algorithms for learning and optimization, since each state is augmented with all possible *subsets* of actions, incurring an exponential blow up in state space size. To address this issue, we show that SAS-MDPs possess an important property: the Q-value of an available action $a$ is independent of the availability of other actions. This allows us to prove that optimal policies can be represented compactly using (state-specific) decision lists (or orderings) over the action set.
This special structure allows one to solve the SAS RL problem effectively using, for example, Q-learning. We also devise model-based algorithms that exploit this policy structure. We develop value and policy iteration schemes, showing they converge in a polynomial number of iterations (w.r.t. the size of the underlying “base” MDP). We also show that per-iteration complexity is polynomial time for two important special forms of action availability distribution: (a) when action availabilities are independent, both methods are exact; (b) when the distribution over sets $A_s$ is sampleable, we obtain approximation algorithms with polynomial sample complexity. In fact, policy iteration is strongly polynomial under additional assumptions (for a fixed discount factor). We show that a linear program for SAS-MDPs can be solved in polynomial time as well. Finally, we offer a simple empirical demonstration of the importance of accounting for stochastic action availability when computing an MDP policy.
Additional discussion and full proofs of all results can be found in a longer version of this paper [@sasmdps_full:arxiv18].
MDPs with Stochastic Action Sets {#sec:sas}
================================
We first introduce SAS-MDPs and provide a simple example illustrating how action availability impacts optimal decisions. See [@puterman] for more background on MDPs.
The SAS-MDP Model {#sec:sasmodel}
-----------------
Our formulation of *MDPs with Stochastic Action Sets (SAS-MDPs)* derives from a standard, finite-state, finite-action MDP (the *base MDP*) $\calM$, with $n$ states $S$, *base* actions $B_s$ for $s\in S$, and transition and reward functions, $P: S \times B \rightarrow \Delta(S)$ and $r:S\times B \rightarrow \reals$. We use $p^k_{s,s'}$ and $r^k_s$ to denote the probability of transition to $s'$ and the accrued reward, respectively, when action $k$ is taken at state $s$. For notational ease, we assume that feasible action sets for each $s\in S$ are identical, so $B_s = B$ (allowing distinct base sets at different states has no impact on what follows). Let $|B| = m$ and $M = |S \times B | =nm$. We assume an infinite-horizon, discounted objective with fixed discount rate $\gamma$, $0\leq\gamma<1$.
In a SAS-MDP, the set of actions available at state $s$ at any stage $t$ is a random subset $A^{(t)}_s \subseteq B$. We assume a family of *action availability distributions* $P_s\in\Delta(2^B)$ defined over the powerset of $B$. These can depend on $s\in S$ but are otherwise history-independent, hence $\Pr(A^{(t)}_s | s^{(1)},\ldots,s^{(t)}) = \Pr(A^{(t)}_s | s^{(t)})$. Only actions $k\in A^{(t)}_s$ in the realized available action set can be executed at stage $t$. Apart from this, the dynamics of the MDP is unchanged: when an (available) action is taken, state transitions and rewards are prescribed as in the base MDP. In what follows, we assume that some action is always available, i.e., $\Pr(A^{(t)}_s = \emptyset) = 0$ for all $s, t$.[^1] Note that a SAS-MDP does not conform to the usual definition of an MDP.
Example {#sec:sasexample}
-------
The following simple MDP shows the importance of accounting for stochastic action availability when making decisions. The MDP below has two states. Assume the agent starts at state $s_1$, where two actions (indicated by directed edges for their transitions) are always available: one ($\Stay$) stays at $s_1$, and the other ($\Go$) transitions to state $s_2$, both with reward $1/2$. At $s_2$, the action $\Down$ returns to $s_1$, is always available and has reward 0. A second action $\Up$ also returns to $s_1$, but is available with only probability $p$ and has reward 1. {width="0.5\columnwidth"}
A naive solution that ignores action availability is as follows: we first compute the optimal $Q$-function assuming all actions are available (this can be derived from the optimal value function, computed using standard techniques). Then at each stage, we use the best action available at the current state where actions are ranked by Q-value. Unfortunately, this leads to a suboptimal policy when the $\Up$ action has low availability, specifically if $p < 0.5$.
The best naive policy always chooses to move to $s_2$ from $s_1$; at $s_2$, it picks the best action available. This yields a reward of $1/2$ at even stages, and an expected reward of $p$ at odd stages. However, by anticipating the possibility that action $\Up$ is unavailable at $s_2$, the optimal (SAS) policy always stays at $s_1$, obtaining reward $1/2$ at all stages. For $p < 1/2$, the latter policy dominates the former: the plot on the right shows the fraction of the optimal (SAS) value *lost* by the naive policy ($Std$) as a function of the availability probability $p$. This example also illustrates that as action availability probabilities approach $1$, the optimal policy for the base MDP is also optimal for the SAS-MDP.
Related Work {#sec:related}
------------
While a general formulation of MDPs with stochastic action availability does not appear in the literature, there are two strands of closely related work. In the bandits literature, *sleeping bandits* are defined as bandit problems in which the arms available at each stage are determined randomly or adversarially (sleeping experts are similar, with complete feedback being provided rather than bandit feedback) [@kleinbergEtAl:MLJ2010; @kanadeEtAl:aistats09Sleeping]. Best action orderings (analogous to our decision list policies for SAS-MDPs) are often used to define regret in these models. The goal is to develop exploration policies to minimize regret. Since these models have no state, if the action reward distributions are known, the optimal policy is trivial: always take the best *available* action. By contrast, a SAS-MDP, even a known model, induces a difficult optimization problem, since the quality of an action depends not just on its immediate reward, but also on the availability of actions at reachable (future) states. This is our focus.
The second closely related branch of research comes from the field of stochastic routing. The “Canadian Traveller Problem”—the problem of minimizing travel time in a graph with unavailable edges—was introduced by Papadimitriou and Yannakakis [@papadimitriou:shortestpath], who gave intractability results (under much weaker assumptions about edge availability, e.g. adversarial). Poliyhondrous and Tsitsiklis [@polychondrous:recourse] consider a stochastic version of the problem, where edge availabilities are random but static (and any edge observed to be unavailable remains so throughout the scenario). Most similar to our setting is the work of Nikolova and Karger [@nikolova:canadian], who discuss the case of resampling edge costs at each node visit; however, the proposed solution is well-defined only when the edge costs are finite and does not easily extend to unavailable actions/infinite edge costs. Due to the specificity of their modeling assumptions, none of the solutions found in this line of research can be adapted in a straightforward way to SAS-MDPs.
Two Reformulations of SAS-MDPs {#sec:reformulate}
==============================
The randomness of feasible actions means that SAS-MDPs do not conform to the usual definition of an MDP. In this section, we develop two reformulations of SAS-MDPs that transform them into MDPs. We discuss the relative advantages of each, outline key properties and relationships between these models, and describe important special cases of the SAS-MDP model itself.
The Embedded MDP {#sec:embedded}
----------------
We first consider a reformulation of the SAS-MDP in which we embed the (realized) available action set into the state space itself. This is a straightforward way to recover a standard MDP. The *embedded MDP* $\calM_e$ for a SAS-MDP has state space $S_e = \{s\circ A: s\in S, A\subseteq B\}$, with $s\circ A$ having feasible action set $A$.[^2] The history independence of $P_s$ allows transitions to be defined as: $$p^k_{s\circ A,s'\circ A'} =
P(s'\circ A' | s\circ A, k) = p^k_{s,s'} P_{s'}(A')
,
%\quad
\;\;
\forall k\in A.$$ Rewards are defined similarly: $r^k(s\circ A) = r^k(s)$ for $k\in A$.
In our earlier example, the embedded MDP has three states: $s_1\circ\{\Stay,\Go\}, s_2\circ\{\Up,\Down\}, s_2\circ\{\Down\}$ (other action subsets have probability $0$ hence their corresponding embedded states are unreachable). The feasible actions at each state are given by the embedded action set, and the only stochastic transition occurs when $\Go$ is taken at $s_1$: it moves to $s_2\circ\{\Up,\Down\}$ with probability $p$ and $s_2\circ\{\Down\}$ with probability $1-p$.
Clearly, the induced reward process and dynamics are Markovian, hence $\calM_e$ is in fact an MDP under the usual definition. Given the natural translation afforded by the embedded MDP, we view this as providing the basic “semantic” underpinnings of the SAS-MDP model. This translation affords the use of standard MDP analytical tools and methods.
A (stationary, determinstic, Markovian) policy $\pi:S_e \rightarrow B$ for $\calM_e$ is restricted so that $\pi(s\circ A) \in A$. The policy backup operator $T^\pi_e$ and Bellman operator $T^\ast_e$ for $\calM_e$ decompose naturally as follows:
$$\begin{aligned}
&T^\pi_e V_e(s\circ A_s)
= r^{\pi(s\circ A_s)}_s + \nonumber
\\ &\qquad \gamma\sum_{s'}p^{\pi(s\circ A_s)}_{s,s'}
\sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V_e(s'\circ A_{s'})
\label{eq:embeddedpolicybackup},
\\
% V^\pi_e(s\circ A_s)
% &= r^{\pi(s\circ A_s)}_s
% + \gamma\sum_{s'} p^{\pi(s\circ A_s)}_{s,s'}
% \sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V^\pi_e(s'\circ A_{s'})
% \label{eq:embeddedpolicyval}
%,
%\\
&T^\ast_e V_e(s\circ A_s)
= \max_{k\in A_s} r^k_s + \nonumber
\\ &\qquad \gamma\sum_{s'}p^k_{s,s'}
\sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V_e(s'\circ A_{s'})
\label{eq:embeddedbellmanbackup}
%,
%\\
% V^\ast_e(s\circ A_s)
% &= \max_{k\in A_s} r^k_s
% + \gamma\sum_{s'}p^k_{s,s'}
% \sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V^\ast_e(s'\circ A_{s'})
% \label{eq:embeddedoptval}
%.\end{aligned}$$
Their fixed points, $V^\pi_e$ and $V^\ast_e$ respectively, can be expressed similarly. Obtaining an MDP from an SAS-MDP via action-set embedding comes at the expense of a (generally) exponential blow-up in the size of the state space, which can increase by a factor of $2^{|B|}$.
The Compressed MDP {#sec:compressed}
------------------
The embedded MDP provides a natural semantics for SAS-MDPs, but is problematic from an algorithmic and learning perspective given the state space blow-up. Fortunately, the history independence of the availability distributions gives rise to an effective, compressed representation. The *compressed MDP* $\calM_c$ recasts the embedded MDP in terms of the original state space, using expectations to express value functions, policies, and backups over $S$ rather than over the (exponentially larger) $S_e$. As we will see below, the compressed MDP induces a blow-up in action space rather than state space, but offers significant computational benefits.
Formally, the state space for $\calM_c$ is $S$. To capture action availability, the feasible action set for $s\in S$ is the set of *state policies*, or mappings $\mu_s: 2^B \rightarrow B$ satisfying $\mu_s(A_s) \in A_s$. In other words, once we reach $s$, $\mu_s$ dictates what action to take for any realized action set $A_s$. A policy for $\calM_c$ is a family $\mu_c = \{\mu_s :s\in S\}$ of such state policies. Transitions and rewards use expectations over $A_s$:
$$\begin{aligned}
p^{\mu_s}_{s,s'}
= \sum_{A_{s}\subseteq B} P_s(A_{s}) p^{\mu_s(A_s)}_{s,s'}
~~\mbox{and}~~
r^{\mu_s}_{s}
= \sum_{A_{s}\subseteq B} P_s(A_{s}) r^{\mu_s(A_s)}_{s}~.\end{aligned}$$
In our earlier example, the compressed MDP has only two states, $s_1$ and $s_2$. Focusing on $s_2$, its “actions” in the compressed MDP are the set of state policies, or mappings from the realizable available sets $\{\{\Up,\Down\}, \{\Down\}\}$ into action choices (as above, we ignore unrealizable action subsets): in this case, there are two such state policies: the first selects $\Up$ for $\{\Up,\Down\}$ and (obviously) $\Down$ for $\{\Down\}$; the second selects $\Down$ for $\{\Up,\Down\}$ and $\Down$ for $\{\Down\}$.
It is not hard to show that the dynamics and reward process defined above over this compressed state space and expanded action set (i.e., the set of state policies) are Markovian. Hence we can define policies, value functions, optimality conditions, and policy and Bellman backup operators in the usual fashion. For instance, the Bellman and policy backup operators, $T^\star_c$ and $T_{\mu}^c$, on compressed value functions are:
$$\begin{aligned}
T_c^*V_c(s) =&
\E_{A_s\subseteq B}\;
\max_{k\in A_s} r^k_s + \gamma \sum_{s'} p^k_{s,s'} V_c(s'),
\label{eq:compressedBellmanOp}
\\
T^{\mu}_cV_c(s) =&
\E_{A_s\subseteq B}\;
r^{\mu_s(A_s)}_s + \gamma \sum_{s'} p^{\mu_s(A_s)}_{s,s'} V_c(s').
\label{eq:compressedPolicyOp}\end{aligned}$$
It is easy to see that any state policy $\mu$ induces a Markov chain over base states, hence we can define a standard $n\times n$ transition matrix $P^{\mu}$ for such a policy in the compressed MDP, where $p^{\mu}_{s,s'} = \E_{A\subseteq B} p^{\mu(s)(A)}_{s,s'}$. When additional independence assumptions hold, this expectation over subsets can be computed efficiently (see \[sec:pda\]). Critically, we can show that there is a direct “equivalence” between policies and their value functions (including optimal policies and values) in $\calM_c$ and $\calM_e$. Define the action-expectation operator $E: \mathbb{R}^{n2^m} \rightarrow \mathbb{R}^n$ to be a mapping that compresses a value function $V_e$ for $\calM_e$ into a value function $V^e_c$ for $\calM_c$:
$$V^e_c(s)
= EV_e(s)
=\!\!\! \E_{A_s\subseteq B} V_e(s\circ A_s)
=\!\!\! \sum_{A_s \subseteq B} P_s(A_{s}) V_e(s\circ A_s).$$
We emphasize that $E$ transforms an (arbitrary) value function $V_e$ in embedded space into a new value function $V_c^e$ defined in compressed space (hence, $V_c^e$ is *not* defined w.r.t.$\calM_c$).
\[lemma1\] $ET^*_e V_e = T_c^*EV_e$. Hence, $T^*_c$ has a unique fixed point $V_c^\ast = EV_e^\ast$.
$$\begin{aligned}
ET^eV_e(s) & = \E_{A\subseteq B} T^eV_e(s\circ A)\\
& = \E_{A\subseteq B}
\max_{k\in A} r^k_s +\gamma\sum_{s'\circ A'}
p^k_{s\circ A,s'\circ A'} V_e(s'\circ A') \\
& = \E_{A\subseteq B}
\max_{k\in A} r^k_s +\gamma\sum_{s'} p^k_{s,s'}
\E_{A'\subseteq B} V_e(s'\circ A') \\
& = \E_{A\subseteq B}
\max_{k\in A} r^k_s +\gamma\sum_{s'} p^k_{s,s'} EV^e(s') \\
& = T^cEV^e(s')
.\end{aligned}$$
\[lemma2\] Given the optimal value function $V^\ast_c$ for $\calM_c$, the optimal policy $\pi^\ast_e$ for $\calM_e$ can be constructed directly. Specifically, for any $s\circ A$, the optimal policy $\pi^\ast_e(s\circ A)$ and optimal value $V^\ast_e(s\circ A)$ at that embedded state can be computed in polynomial time.
Given $s\circ A$, the expected value of each action in $k\in A$ can be computed using a one-step backup of $V^\ast_c$. Then $\pi^\ast_e(s\circ A)$ is the action with maximum value, and $V^\ast_e(s\circ A)$ is its backed-up expected value.
Therefore, it suffices to work directly with the compressed MDP, which allows one to use value functions (and $Q$-functions) over the original state space. The price is that one needs to use state policies, since the best action at $s$ depends on the available set $A_s$. In other words, while the embedded MDP causes an exponential blow-up in state space, the compressed MDP causes an exponential blow-up in action space. We now turn to assumptions that allow us to effectively manage this action space blow-up.
Decision List Policies {#sec:lists}
----------------------
The embedded and compressed MDPs do not, *prima facie*, offer much computational or representational advantage, since they rely on an exponential increase in the size of the state space (embedded MDP) or decision space (compressed MDP). Fortunately, SAS-MDPs have optimal policies with a useful, concise form. We first focus on the policy representation itself, then describe the considerable computational leverage it provides.
A *decision list (DL) policy* $\mu$ is a type of policy for $\calM_e$ that can be expressed compactly using $O(nm \log m)$ space and executed efficiently. Let $\Sigma_B$ be the set of permutations over base action set $B$. A DL policy $\mu: S \rightarrow \Sigma_B$ associates a permutation $\mu(s) \in \Sigma_B$ with each state, and is executed at embedded state $s\circ A$ by executing $\min \{i \in \{1,\ldots,m\} : \mu(s)(i) \in A\}$. In other words, whenever base state $s$ is encountered and $A$ is the available set, the first action $k\in A$ in the order dictated by DL $\mu(s)$ is executed. Equivalently, we can view $\mu(s)$ as a state policy $\mu_s$ for $s$ in $\calM_c$. In our earlier example, one DL $\mu(s_2)$ is $[\Up,\Down]$, which requires taking (base) action $\Up$ if it is available, otherwise taking $\Down$.
For any SAS-MDP, we have optimal DL policies:
$\calM_e$ has an optimal policy that can be represented using a decision list. The same policy is optimal for the corresponding $\calM_c$.
Let $V^\ast$ be the (unique) optimal value function for $\calM_e$ and $Q^\ast$ its corresponding Q-function (see Sec. \[sec:valueiteralg\] for a definition). A simple inductive argument shows that no DL policy is optimal only if there is some state $s$, action sets $A \neq A'$, and (base) actions $j \neq k$, s.t. (i) $j,k \in A, A'$; (ii) for some optimal policy $\pi^\ast(s\circ A) = j$ and $\pi^\ast(s\circ A') = k$; and (iii) either $Q^\ast(s\circ A,j) > Q^\ast(s\circ A,k)$ or or $Q^\ast(s\circ A',k) > Q^\ast(s\circ A',j)$. However, the fact that the optimal Q-value of any action $k\in A$ at state $s\circ A$ is independent of the other actions in $A$ (i.e., it depends only on the base state) implies that these conditions are mutually contradictory.
The Product Distribution Assumption {#sec:pda}
-----------------------------------
The DL form ensures that optimal policies and value functions for SAS-MDPs can be expressed polynomially in the size of the base MDP $\calM$. However, their computation still requires the computation of expectations over action subsets, e.g., in Bellman or policy backups (Eqs. \[eq:compressedBellmanOp\], \[eq:compressedPolicyOp\]). This will generally be infeasible without some assumptions on the form the action availability distributions $P_s$.
One natural assumption is the *product distribution assumption (PDA)*. PDA holds when $P_s(A)$ is a product distribution where each action $k\in B$ is available with probability $\rho^k_s$, and subset $A \subseteq B$ has probability $\rho_s^A = \prod_{k\in A} \rho^k_s \prod_{k\in B\setminus A} (1-\rho^k_s)$. This assumption is a reasonable approximation in the settings discussed above, where state-independent exogenous processes determine the availability of actions (e.g., the probability that one advertiser’s campaign has budget remaining is roughly independent of another advertiser’s). For ease of notation, we assume that $\rho^k_s$ is identical for all states $s$ (allowing different availability probabilities across states has no impact on what follows). To ensure the MDP is well-founded, we assume some default action (e.g., no-op) is always available.[^3] Our earlier running example trivially satisifes PDA: at $s_2$, $\Up$’s availability probability ($p$) is independent of the availability of $\Down$ (1).
When the PDA holds, the DL form of policies allows the expectations in policy and Bellman backups to be computed efficiently without enumeration of subsets $A\subseteq B$. For example, given a fixed DL policy $\mu$, we have
$$\begin{aligned}
T^{\mu}_cV_c(s) &=
\sum_{i=1}^{m} \,\left[\prod_{j=1}^{i-1} (1-\rho^{\mu(s)(j)}_s)\right]\, \rho^{\mu(s)(i)}_s
\Bigg( r_s^{\mu(s)(i)} \nonumber \\
&~ + \gamma \sum_{s'} p^{\mu(s)(i)}_{s,s'} V_c(s') \Bigg).
\label{eq:pdabellman}\end{aligned}$$
The Bellman operator has a similar form. We exploit this below to develop tractable value iteration and policy iteration algorithms, as well as a practical LP formulation.
Arbitrary Distributions with Sampling (ADS) {#sec:ads}
-------------------------------------------
We can also handle the case where, at each state, the availability distribution is unknown, but is sampleable. In the longer version of the paper [@sasmdps_full:arxiv18], we show that samples can be used to approximate expectations w.r.t. available action subsets, and that the required sample size is polynomial in $|B|$, and not in the size of the *support* of the distribution.
Of course, when we discuss algorithms for policy computation, this approach does not allow us to compute the optimal policy exactly. However, it has important implications for sample complexity of learning algorithms like Q-learning. We note that the ability to sample available action subsets is quite natural in many domains. For instance, in ad domains, it may not be possible to model the process by which eligible ads are generated (e.g., involving specific and evolving advertiser targeting criteria, budgets, frequency capping, etc.). But the eligible subset of ads considered for each impression opportunity is an action-subset sampled from this process.
Under ADS, we compute approximate backup operators as follows. Let $\mathcal{A}_s = \{A_s^{(1)},\ldots,A_s^{(T)}\}$ be an i.i.d. sample of size $T$ of action subsets in state $s$. For a subset of actions $A$, an index $i$ and a decision list $\mu$, define $I_{[i,A,\mu]}$ to be 1 if $\mu(i) \in A$ and for each $j < i$ we have $\mu(j) \not\in A$, or 0 otherwise. Similar to \[eq:pdabellman\], we define: [ $$\begin{aligned}
T^{\mu}_cV_c(s) &\!=\! \frac{1}{T} \sum_{t=1}^T
\sum_{i=1}^{m} I_{\left[i,A_s^{(t)}, \mu(s)\right]}
\!
\Big( r_s^{\mu(s)(i)} \!+\! \gamma \!\sum_{s'} p^{\mu(s)(i)}_{s,s'} V_c(s') \Big).\end{aligned}$$ ]{}In the sequel, we focus largely on PDA; in most cases equivalent results can be derived in the ADS model.
Q-Learning with the Compressed MDP {#sec:qlearn}
==================================
Suppose we are faced with learning the optimal value function or policy for an SAS-MDP from a collection of trajectories. The (implicit) learning of the transition dynamics and rewards can proceed as usual; the novel aspect of the SAS model is that the action availability distribution must also be considered. Remarkably, Q-learning can be readily augmented to incorporate stochastic action sets: we require only that our training trajectories are augmented with the set of actions that were available at each state, $$\ldots s^{(t)}, A^{(t)}, k^{(t)}, r^{(t)},
s^{(t+1)}, A^{(t+1)}, k^{(t+1)}, r^{(t+1)}, \ldots,$$ where: $s^{(t)}$ is the realized state at time $t$ (drawn from distribution $P(\cdot|s^{(t-1)}, k^{(t-1)})$); $A^{(t)}$ is the realized available set at time $t$, drawn from $P_{s^{(t)}}$; $k^{(t)}\in A^{(t)}$ is the action taken; and $r^{(t)}$ is the realized reward. Such augmented trajectory data is typically available. In particular, the required sampling of available action sets is usually feasible (e.g., in ad serving as discussed above). *SAS-Q-learning* can be applied directly to the compressed MDP $\calM_c$, requiring only a minor modification of the standard Q-learning update for the base MDP. We simply require that each Q-update maximize over the *realized available actions* $A^{(t+1)}$:
$$\begin{aligned}
Q^{\new}(s^{(t)},k^{(t)}) &\leftarrow
(1-\alpha_t) Q^{\old}(s^{(t)},k^{(t)}) \\
&\quad +
\alpha_t [r^{(t)} + \gamma \max_{k\in A^{(t+1)}} Q^{\old}(s^{(t+1)},k)]~.\end{aligned}$$
Here $Q^{\old}$ is the previous $Q$-function estimate and $Q^{\new}$ is the updated estimate, thus it encompasses both online and batch Q-learning, experience replay, etc.; and $0\leq \alpha_t < 1$ is our (adaptive) learning rate.
It is straightforward to show that, under the usual exploration conditions, SAS-Q-learning will converge to the optimal Q-function for the compressed MDP, since the expected maximum over sampled action sets at any particular state will converge to the expected maximum at that state.
The SAS-Q-learning algorithm will converge w.p. 1 to the optimal Q-function for the (discounted, infinite-horizon) compressed MDP $\calM_c$ if the usual stochastic approximation requirements are satisfied. That is, if (a) rewards are bounded and (b) the subsequence of learning rates $\alpha_{t(s,k)}$ applied to $(s,k)$ satisfies $\sum \alpha_{t(s,k)} = \infty$ and $\sum \alpha^2_{t(s,k)} < \infty$ for all state-action pairs $(s,k)$ (see, e.g., [@watkins:mlj92]).
Moreover, function approximation techniques, such as DQN [@mnih2015], can be directly applied with the same action set-sample maximization. Implementing an optimal policy is also straightforward: given a state $s$ and the realization $A_s$ of the available actions, one simply executes $\arg\max_{k\in A_s} Q(s,k)$.
We note that extracting the optimal value function $V_c(s)$ for the compressed MDP from the learned Q-function is not viable without some information about the action availability distribution. Fortunately, one need not know the expected value at a state to implement the optimal policy.[^4]
Value Iteration in the Compressed MDP {#sec:valueiter}
=====================================
Computing a value function for $\calM_c$, with its “small” state space $S$, suffices to execute an optimal policy. We develop an efficient *value iteration (VI)* method to do this.
Value Iteration {#sec:valueiteralg}
---------------
Solving an SAS-MDP using VI is challenging in general due to the required expectations over action sets. However, under PDA, we can derive an efficient VI algorithm whose complexity depends only polynomially on the base set size $|B|$. Assume a current iterate $V^t$, where $
V^t(s) = \E_{A_s} [\max_{k\in A_s} Q^t(s,k) ]
$. We compute $V^{t+1}$ as follows:
- For each $s\in S, k\in B$, compute its $(t+1)$-stage-to-go Q-value: $Q^{t+1}(s,k) = r^k_s + \gamma \sum_{s'} p^k_{s,s'} V^t(s').$
- Sort these Q-values in descending order. For convenience, we re-index each action by its Q-value rank (i.e., $k_{(1)}$ is the action with largest Q-value, and $\rho_{(1)}$ is its probability, $k_{(2)}$ the second-largest, etc.).
- For each $s\in S$, compute its $(t+1)$-stage-to-go value:
$$\begin{aligned}
V^{t+1}(s) & = \E\nolimits_{A_s} \left[\max_{k\in A_s} Q^{t+1}(s,k)\right]\\
% & = p_1 Q^{t+1}(s,k_1) + (1 - p_1) p_2 Q^{t+1}(s,k_2) \\
% &\qquad + (1 - p_1)(1 - p_2) p_3 Q^{t+1}(s,k_3) + \ldots\\
& = \sum_{i=1}^{m-1} \left( \prod_{j=1}^{i-1} (1 - \rho_{(j)}) \right) \rho_{(i)} Q^{t+1}(s,k_{(i)})
.
\end{aligned}$$
Under ADS, we use the approximate Bellman operator:
$$\begin{aligned}
\widehat{V}^{t+1}(s) &= \E\nolimits_{A_s} \left[\max_{k\in A_s} \widehat{Q}^{t+1}(s,k)\right]
\\
&= \frac{1}{T} \sum_{t=1}^T
\sum_{i=1}^{m} I_{\left[i,A_s^{(t)}, \mu(s)\right]} \widehat{Q}^{t+1}(s,\mu(s)(i))~,
\end{aligned}$$
where $\mu(s)$ is the DL resulting from sorting $\widehat{Q}^{t+1}$-values.
The Bellman operator under PDA is tractable:
\[obs:VIperiteration\] The compressed Bellman operator $T^*_c$ can be computed in $O(n m\log m)$ time.
Therefore the per-iteration time complexity of VI for $\calM_c$ compares favorably to the $O(n m)$ time of VI in the base MDP. The added complexity arises from the need to sort Q-values.[^5] Conveniently, this sorting process immediately provides the desired DL state policy for $s$.
Using standard arguments, we obtain the following results, which immediately yield a polytime approximation method.
\[lemma3\] $T^*_c$ is a contraction with modulus $\gamma$ i.e., $||T^*_c v_c - T^*_c v'_c|| \leq \gamma ||v_c - v'_c||$.
For any precision $\veps < 1$, the compressed value iteration algorithm converges to an $\veps$-approximation of the optimal value function in $O(\log (L/\veps))$ iterations, where $L\leq [\max_{s,k} r^k_s ]/(1-\gamma)$ is an upper bound on $||V^\ast_e||$.
We provide an even stronger result next: VI, in fact, converges to an *optimal* solution in polynomial time.
The Complexity of Value Iteration {#sec:valueitercomplexity}
---------------------------------
Given its polytime per-iteration complexity, to ensure VI is polytime, we must show that it converges to a value function that induces an optimal policy in polynomially many iterations. To do so, we exploit the compressed representation and adapt the technique of [@tseng:ORLetters90].
Assume, w.r.t. the base MDP $\calM$, that the discount factor $\gamma$, rewards $r^k_s$, and transition probabilities $p^k_{s,s'}$, are rational numbers represented with a precision of $1/ \delta$ ($\delta$ is an integer). Tseng shows that VI for a standard MDP is strongly polynomial, assuming constant $\gamma$ and $\delta$, by proving that: (a) if the $t$’th value function produced by VI satisfies $$||V^t - V^\ast|| <
%\frac{1}{2 \delta^{2n+2} n^n}
1/(2 \delta^{2n+2} n^n)
,$$ then the policy induced by $V^t$ is optimal; and (b) VI achieves this bound in polynomially many iterations.
We derive a similar bound on the number of VI iterations needed for convergence in an SAS-MDP, using the same input parameters as in the base MDP, and applying the same precision $\delta$ to the action availability probabilities. We apply Tseng’s result by exploiting the fact that: (a) the optimal policy for the embedded MDP $\calM_e$ can be represented as a DL; (b) the transition function for any DL policy can be expressed using an $n\times n$ matrix (we simply take expectations, see above); and (c) the corresponding linear system can be expressed over the *compressed* rather than the embedded state space to determine $V_c^\ast$ (rather than $V_e^\ast$).
Tseng’s argument requires some adaptation to apply to the compressed VI algorithm. We extend his precision assumption to account for our action availability probabilities as well, ensuring $\rho^k_s$ is also represented up to precision of $1/\delta$.
Since $\calM_c$ is an MDP, Tseng’s result applies; but notice that each entry of the transition matrix for any state’s DL $\mu$, which serves as an action in $\calM_c$, is a product of $m+1$ probabilities, each with precision $1/\delta$. We have that $p^\mu_{s,s'}$ has precision of $1/\delta^{m+1}$. Thus the required precision parameter for our MDP is at most $\delta^{m+1}$. Plugging this into Tseng’s bound, VI applied to $\calM_c$ must induce an optimal policy at the $t$’th iteration if [$$||V^t - v^\ast|| < 1/(2({\delta^{(m+1)}})^{2n} n^n)
= 1/(2\delta^{(m+1)2n} n^n)~.$$ ]{}This in turn gives us a bound on the number of iterations of VI needed to reach an optimal policy:
\[thm:VIconvergence\] VI applied to $\calM_c$ converges to a value function whose greedy policy is optimal in $t^*$ iterations, where [$$t^* \leq \log(2\delta^{2n(m+1)} n^n M) / \log(1/\gamma)
% \quad\mbox{iterations}~.$$ ]{}
Combined with Obs. \[obs:VIperiteration\], we have:
\[cor:VIpolytime\] VI yields an optimal policy for the SAS-MDP corresponding to $\calM_c$ in polynomial time.
Under ADS, VI merely approximates the optimal policy. In fact, one cannot compute an exact optimal policy without observing the entire support of the availability distributions (requiring exponential sample size).
Policy Iteration in the Compressed MDP {#sec:policyiter}
======================================
We now outline a policy iteration (PI) algorithm.
Policy Iteration {#sec:policyiteralg}
----------------
The concise DL form of optimal policies can be exploited in PI as well. Indeed, *the greedy policy $\pi^V$ with respect to any value function $V$ in the compressed space* is representable as a DL. Thus the policy improvement step of PI can be executed using the same independent evaluation of action Q-values and sorting as used in VI above:
$$\begin{gathered}
Q^V(s,k) = r(s,k) + \gamma \sum_{s'} p^k_{s, s'} V(s')
,\\
Q^V(s,A_s) \!=\! \max_{k\in A_s} Q^V(s,k)
\;\textrm{, and } \;
\pi^V(s,A_s) \!=\! \arg\max_{k\in A_s} Q^V(s,k).
% Q^V(s,k) =&~ r(s,k) + \gamma \sum_{s'} p^k_{s, s'} V(s')
% ,\\
% Q^V(s,A_s) =&~ \max_{k\in A_s} Q^V(s,k)
% ,\\
% \pi^V(s,A_s) =&~ \arg\max_{k\in A_s} Q^V(s,k).
% \end{align*}\end{gathered}$$
The DL policy form can also be exploited in the policy evaluation phase of PI. The tractability of policy evaluation requires a tractable representation of the action availability probabilities, which PDA provides, leading to the following PI method that exploits PDA:
1. Initialize an arbitrary policy $\pi$ in decision list form.
2. Evaluate $\pi$ by solving the following linear system over variables $V^\pi(s), \forall s\in S$: (Note: We use $Q^{\pi}(s,k)$ to represent the relevant linear expression over $V^\pi$.)
$$\begin{aligned}
V^\pi(s)
% &= p_{(1)} Q^{\pi}(s,k_{(1)}) \\
% &\qquad + (1-p_{(1)})p_{(2)} Q^{V^\pi}(s,k_{(2)})+\ldots\\
&= \sum_{i=1}^{n} \,[\prod_{j=1}^{i-1} (1-\rho_{(j)})]\, \rho_{(i)} Q^{\pi}(s,k_{(i)})
\label{eq:pda_expect_val}
\end{aligned}$$
3. Let $\pi'$ denote the greedy policy w.r.t. $V^\pi$, which can be expressed in DL form for each $s$ by sorting Q-values $Q^{\pi}(s,k)$ as above (with standard tie-breaking rules). If $\pi'(s) = \pi(s)$, terminate; otherwise replace $\pi$ with $\pi'$ and repeat (Steps 2 and 3).
Under ADS, PI can use the approximate Bellman operator, giving an approximately optimal policy.
The Complexity of Policy Iteration {#sec:policyitercomplexity}
----------------------------------
The per-iteration complexity of PI in $\calM_c$ is polynomial: as in standard PI, policy evaluation solves an $n\times n$ linear system (naively, $O(n^3)$) plus the additional overhead (linear in $M$) to compute the compounded availability probabilities; and policy improvement requires $O(mn^2)$ computation of action Q-values, plus $O(nm\log m)$ overhead for sorting Q-values (to produce improving DLs for all states).
An optimal policy is reached in a number of iterations no greater than that required by VI, since: (a) the sequence of value functions for the policies generated by PI contracts at least as quickly as the value functions generated by VI (see, e.g., [@meister:ORSpektrum86; @hansen:jacm13]); (b) our precision argument for VI ensures that the greedy policy extracted at that point will be optimal; and (c) once PI finds an optimal policy, it will terminate (with one extra iteration). Hence, PI is polytime (assuming a fixed discount $\gamma<1$).
\[thm:PIpolytime\] PI yields an optimal policy for the SAS-MDP corresponding to $\calM_c$ in polynomial time.
In the longer version of the paper [@sasmdps_full:arxiv18], we adapt more direct proof techniques to derive polynomial-time convergence of PI for SAS-MDPs under additional assumptions. Concretely, for a policy $\mu$ and actions $k_1,k_2$, let $\eta_\mu(s,k_1,k_2)$ be the probability, over action sets, that at state $s$, the optimal $\mu^\star$ selects $k_1$ and $\mu$ selects $k_2$. Let $q > 0$ be such that $\eta_\mu(s,k_1,k_2) \ge q$ whenever $\eta_\mu(s,k_1,k_2) > 0$. We show:
The number of iterations it takes policy iteration to converge is no more than [ $$O\left( \frac{nm^2}{1-\gamma} \log \frac{m}{1-\gamma} \log \frac{e}{q} \right)~.$$ ]{}
Under PDA, the theorem implies *strongly-polynomial* convergence of PI if each action is available with constant probability. In this case, for any $\mu$, $k_i$, $k_j$, and $s$, we have $\eta_\mu(s,k_i,k_j) \ge \rho_s^{k_i} \cdot
\rho_s^{k_j} = \Omega(1)$, which in turn implies that we can take $q = \Omega(1)$ in the bound above.
Linear Programming in the Compressed MDP {#sec:lp}
========================================
An alternative model-based approach is linear programming (LP). The primal formulation for the embedded MDP $\calM_e$ is straightforward (since it is a standard MDP), but requires exponentially many variables (one per embedded state) and constraints (one per embedded state, base action pair). A (nonlinear) primal formulation for the compressed MDP $\calM_c$ reduces the number of variables to $|S|$:
$$\begin{aligned}
\min_{\mathbf{v}}\,\sum_{s\in S}\nolimits\alpha_s v_s,\quad\textrm{s.t. }\,v_s \geq \E\nolimits_{A_s}\max_{k\in A_s} Q(s,k)\quad \forall s.
\label{eq:compressedPrimalConstr}\end{aligned}$$
Here $\alpha$ is an arbitrary, positive state-weighting, over the embedded states corresponding to each base state and $$Q(s,k) = r_s^k + \sum_{s' \in S} p^k_{s,s'} v_{s'}$$ abbreviates the linear expression of the action-value backup at the state and action in question w.r.t. the value variables $v_s$. This program is valid given the definition of $\calM_c$ and the fact that a weighting over embedded states corresponds to a weighting over base states by taking expectations. Unfortunately, this formulation is non-linear, due to the max term in each constraint. And while it has only $|S|$ variables, it has factorially many constraints; moreover, the constraints themselves are not compact due to the presence of the expectation in each constraint.
PDA can be used to render this formulation tractable. Let $\sigma$ denote an arbitrary (inverse) permutation of the action set (so $\sigma(i)=j$ means that action $j$ is ranked in position $i$). As above, the optimal policy at base state $s$ w.r.t. a Q-function is expressible as a DL ( with actions sorted by Q-values) and its expected value given by the expression derived below. Specifically, if $\sigma$ reflects the relative ranking of the (optimal) Q-values of the actions at some fixed state $s$, then $V(s) = Q(s,\sigma(1))$ with probability $\rho_{\sigma(1)}$, i.e., the probability that $\sigma(1)$ occurs in $A_s$. Similarly, $V(s) = Q(s,\sigma(2))$ with probability $(1-\rho_{\sigma(1)})\rho_{\sigma(2)}$, and so on. We define the Q-value of a DL $\sigma$ as follows:
$$\begin{aligned}
Q^V_s(\sigma) = \sum_{i=1}^{n}
\,[\prod_{j=1}^{i-1} (1-\rho_{\sigma(j)})]\,
\rho_{\sigma(i)} Q^V(s,\sigma(i)).\end{aligned}$$
Thus, for any fixed action permutation $\sigma$, the constraint that $v_s$ at least matches the expectation of the maximum action’s Q-value is linear. Hence, the program can be recast as an LP by enumerating action permutations for each base state, replacing the constraints in Eq. (\[eq:compressedPrimalConstr\]) as follows:
$$\begin{aligned}
% \min_{\mathbf{v}}\,\sum_{s\in S}\nolimits\alpha_s v_s,\quad\textrm{s.t. }
v_s \geq Q^V_s(\sigma) \quad \forall s\in S, \forall \sigma \in \Sigma.
\label{eq:compressedPrimalConstr2}\end{aligned}$$
The constraints in this LP are now each compactly represented, but it still has factorially many constraints. Despite this, it can be solved in polynomial time. First, we observe that the LP is well-suited to constraint generation. Given a relaxed LP with a subset of constraints, a greedy algorithm that simply sorts actions by Q-value to form a permutation can be used to find the maximally violated constraint at any state. Thus we have a practical constraint generation algorithm for this LP since (maximally) violated constraints can be found in polynomial time.
More importantly from a theoretical standpoint, the constraint generation algorithm can be used as a separation oracle within an ellipsoid method for this LP. This directly yields an exact, (weakly) polynomial time algorithm for this LP [@GroetschelLovaszSchrijver1988].
Empirical Illustration {#sec:empirical}
======================
We now provide a somewhat more elaborate empirical demonstration of the effects of stochastic action availability. Consider an MDP that corresponds to a routing problem on a real-world road network (Fig. \[fig:routing\]) in the San Francisco Bay Area. The shortest path between the source and destination locations is sought. The dashed edge in Fig. \[fig:routing\] represents a bridge, available only with probability $p$, while all other edges correspond to action choices available with probability $0.5$. At each node, a no-op action (waiting) is available at constant cost; otherwise the edge costs are the geodesic lengths of the corresponding roads on the map. The optimal policies for different choices $p=0.1, 0.2$ and $0.4$ are depicted in Fig. \[fig:routing\], where line thickness and color indicate traversal probabilities under the corresponding optimal policies. It can be observed that lower values of $p$ lead to policies with more redundancy. Fig. \[fig:obliv\] investigates the effect of solving the routing problem obliviously to the stochastic action availability (assuming actions are fully available). The SAS-optimal policy allows graceful scaling of the expected travel time from source to destination as bridge availability decreases. Finalluy, the effects of violating the PDA assumption are investigated in the long version of this work [@TODO].
![Stochastic action MDPs applied to routing.\[fig:routing\]](bridgefigure.pdf){width="0.5\columnwidth"}
![ Expected trip time from source to destination under the SAS-optimal policy vs. under the oblivious optimal policy (the MDP solved as if actions are fully available) as a function of bridge availability.[]{data-label="fig:obliv"}](optimal_vs_obliv.pdf){width="\textwidth"}
Concluding Remarks {#sec:conclude}
==================
We have developed a new MDP model, *SAS-MDPs*, that extends the usual finite-action MDP model by allowing the set of available actions to vary stochastically. This captures an important use case that arises in many practical applications (e.g., online advertising, recommender systems). We have shown that embedding action sets in the state gives a standard MDP, supporting tractable analysis at the cost of an exponential blow-up in state space size. Despite this, we demonstrated that (optimal and greedy) policies have a useful decision list structure. We showed how this DL format can be exploited to construct tractable Q-learning, value and policy iteration, and linear programming algorithms. While our work offers firm foundations for stochastic action sets, most practical applications will not use the algorithms described here explicitly. For example, in RL, we generally use function approximators for generalization and scalability in large state/action problems. We have successfully applied Q-learning using DNN function approximators (i.e., DQN) using sampled/logged available actions in ads and recommendations domains as described in Sec. \[sec:qlearn\]. This has allowed us to apply SAS-Q-learning to problems of significant, commercially viable scale. Model-based methods such as VI, PI, and LP also require suitable (e.g., factored) representations of MDPs and structured implementations of our algorithms that exploit these representations. For instance, extensions of approximate linear programming or structured dynamic programming to incorporate stochastic action sets would be extremely valuable.
Other important questions include developing a polynomial-*sized* direct LP formulation; and deriving sample-complexity results for RL algorithms like Q-learning is also of particular interest, especially as it pertains to the sampling of the action distribution. Finally, we are quite interested in relaxing the strong assumptions embodied in the PDA model—of particular interest is the extension of our algorithms to less extreme forms of action availability independence, for example, as represented using consise graphical models (e.g., Bayes nets).
**Acknowledgments:** Thanks to the reviewers for their helpful suggestions.
[^1]: Models that trigger process termination when $A^{(t)}_s = \emptyset$ are well-defined, but we set aside this model variant here.
[^2]: Embedded states whose embedded action subsets have zero probability are unreachable and can be ignored.
[^3]: We omit the default action from analysis for ease of exposition.
[^4]: It is, of course, straightforward to learn an optimal value function if desired.
[^5]: The products of the action availability probabilities can be computed in linear time via caching.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'P.Ramírez-Moreta, L.Verdes-Montenegro, J.Blasco-Herrera, S.Leon, A.Venhola, M.Yun, V.Peris, R.Peletier, G.VerdoesKleijn, E.Unda-Sanzana, D.Espada, A.Bosma, E.Athanassoula, M.Argudo-Fernández, J.Sabater, J.C.Muñoz-Mateos, M.G.Jones, W.Huchtmeier, J.E.Ruiz, J.Iglesias-Páramo, M.Fernández-Lorenzo, J.Beckman, S.Sánchez-Expósito, J.Garrido'
bibliography:
- 'refs.bib'
date: 'Received Month Day, Year; accepted Month Day, Year'
title: Unveiling the environment and faint features of the isolated galaxy CIG96 with deep optical and HI observations
---
[Asymmetries in atomic hydrogen (HI) in galaxies are often caused by the interaction with close companions, making isolated galaxies an ideal framework to study secular evolution. The AMIGA project has demonstrated that isolated galaxies show the lowest level of asymmetry in their HI integrated profiles compared to even field galaxies, yet some present significant asymmetries. CIG96 (NGC 864) is a representative case reaching a 16% level.]{} [Our aim is to investigate the HI asymmetries of the spiral galaxy CIG96 and what processes have triggered the star-forming regions observed in the XUV pseudo-ring.]{} [We performed deep optical observations at CAHA1.23m, CAHA2.2m and VST (OmegaCAM wide-field camera) telescopes. We reach surface brightness (SB) limits of $\mu_{CAHA2.2m}$=27.5magarcsec$^{-2}$ (Cousins $R$) and $\mu_{VST}$=28.7magarcsec$^{-2}$ (SDSS$r$) that show the XUV pseudo-ring of the galaxy in detail. Additionally, a wavelet filtering of the HI data cube from our deep observations with VLA/EVLA telescope allowed us to reach a column density of $N_{HI}$ = 8.9$\times$10$^{18}$cm$^{-2}$(5$\sigma$) (28$\arcsec$ $\times$ 28$\arcsec$ beam), lower than in any isolated galaxy.]{} [We confirm that the HI of CIG96 extends farther than 4$\times$r$_{25}$ in all directions. Furthermore, we detect for the first time two gaseous structures ($\sim$10$^{6}$M$_{\odot}$) in the outskirts. The SDSS $g-r$ colour index image from CAHA1.23m shows extremely blue colours in certain regions of the pseudo-ring where $N_{HI} >8.5\times10^{20}$cm$^{-2}$, whereas the rest show red colours. Galactic cirrus contaminate the field, setting an unavoidable detection limit at 28.5magarcsec$^{-2}$ (SDSS$r$).]{} [At the current SB and $N_{HI}$ levels, we detect no stellar link within 1$\times$1or gaseous link within 40$\times$40 between CIG96 and any companion. The isolation criteria rule out interactions with other similar-sized galaxies for at least $\sim$2.7Gyr. Using existing stellar evolution models, the age of the pseudo-ring is estimated at 1Gyr or older. Undetected previously accreted companions and cold gas accretion remain as the main hypothesis to explain the optical pseudo-ring and HI features of CIG96.]{}
Introduction {#sec:intro}
============
Most galaxies in the nearby universe are either interacting with or gravitationally bound to nearby companions. Such events are directly responsible for a continuous change in their structural, dynamical and chemical properties [@toomre77]. A wide set of observed and, broadly, understood effects of such interactions (e.g. quenching or enhancement of the stellar formation, gaseous plumes and bridges, tidal streams, etc.) constitute some of the main drivers of the evolution of galaxies. Such interactions may prevail over the internal processes, hiding or even disrupting the key inner evolutionary mechanisms of each particular galaxy that, in the absence of large companions, might otherwise dominate its evolution. The bars present in nearly two$-$thirds of the spiral galaxies (e.g. @buta15), whether initially from external or internal origin, are among these inner elements that can crucially affect the evolution of the galaxy from their bulges or pseudo-bulges out to the outer Lindblad resonance in their external regions [e.g. @korken04; @buta05; @fernandezlorenzo14]. Additionally, the results from other cosmologically motivated studies point out that the interaction of the galaxies with dark matter halos might result in perturbations of the disc (e.g. @kazantzidis08 [@kazantzidis09]). Isolated galaxies, if selected with strict and robust criteria, constitute an ideal framework to study the secular evolution of galaxies since we can exclude the possibility of interactions with large companions. The Analysis of the interstellar Medium of Isolated GAlaxies (AMIGA) project[^1] [@verdes05] was designed to perform a multi-wavelength study of a large sample of galaxies selected with strict isolation criteria from the Catalog of Isolated Galaxies (CIG[^2], 1051 galaxies, @kar73).
With respect to the isolation level, a plethora of references to different definitions and selection criteria may be found throughout the literature of the last 40 years, (e.g. all references in @verdes05 or @muldrew12 among others). As part of the AMIGA project, this work makes use of its isolation criteria and parameters (local number density $\eta_{k}$ and tidal force estimation $Q$) in the version by [@verley07b] revised later by @argudo13 [@argudo14]. Both isolation parameters are defined in depth in the discussion of the environment (see Sect.\[sec:environment\]).
The results of the project are that variables expected to be enhanced by interactions are lower in isolated galaxies than in any other sample (e.g. L$_{FIR}$, @lisenfeld07, radio continuum emission, optical symmetry, @verdes10 and references therein, active galactic nucleus (AGN) rate, @sabater12). Among these, one specific result is especially significant in the context of the present work: the asymmetry level of the atomic gas (HI) integrated profiles of the CIG galaxies is also lower than any other sample, including field galaxies (@espada11b, see @jones18 for a full characterisation of the HI content of AMIGA sample). However, a number of galaxies show unusually high levels of asymmetry (up to 50%), the causes of which remain unknown. If asymmetries can only be generated by interactions, lopsidedness in an isolated galaxy such as CIG96 (NGC864) should not be observed. However, previous data from Green Bank as well as VLA observatories show a large HI envelope beyond 2$\times$ $r_{25}$ (q.v. Table\[table:cig96\]) that has an asymmetry level of 16% in its HI integrated profile [@espada05].
[@espada11a] report on a partial XUV ring (hereafter the pseudo-ring, see Sect.\[sec:sblimit\]) seen in near-UV (NUV) and far-UV (FUV) GALEX data, and located at 1.5$-$2$\times$ $r_{25}$. This pseudo-ring shows patchy regions with star formation (SF). It is not clear that such features can develop in galaxies free from interactions. In this paper we present additional data on this enigmatic object, in particular by obtaining further deep imaging at optical wavelengths.
[@erroz12] have studied the kinematics of the inner regions of CIG96 in H$\alpha$ but no previous study has provided convincing arguments that an external agent can explain both the HI and optical features of CIG96. As a consequence, this raises the question as to whether asymmetries might develop in galaxies free from interactions [@espada05; @espada11a], motivating the in-depth study of CIG96.
However, to support any internal agent as the main evolutionary process, it is necessary to first rule out any external influence. Neither tidal features nor gas$-$rich companions are found in HI maps even for the most asymmetric cases [e.g. @espada05; @espada11b; @portas11; @sengupta12] and current shallow optical images are surprisingly symmetric when dust patches are ignored. In the absence of interactions for the last $\sim$2.7Gyr (see Sect.\[sec:environment\]), any lopsided mode would have already dropped [@jog09]. Does this imply that secular evolution processes can lead to asymmetries? Since the early works of [@bosma78] and [@bosmafreeman93] we know that our understanding of a galaxy may change after performing and comparing deep observations that let us reach very low surface brightness (SB or $\mu$) levels of a galaxy and its surroundings. Therefore, this was the natural follow$-$up for CIG96. Additionally, as suggested by the N$-$body simulations of [@penarrubia05], the orbital properties of halo substructures are determined by the environment and can survive several gigayears, outliving HI tidal features. Within the last two decades, a number of works have unveiled many faint structures or companions that remained hidden in shallower observations [e.g. @martinezdelgado08; @martinezdelgado09; @martinezdelgado15; @duc15; @vandokkum15; @trujillo16; @trujillo17; @iodice17; @bosma17 among others].
[@espada05] also presented the discovery of a close and small companion situated at 15.2$\arcmin$ ($\sim$90 kpc, projected distance) to the east of CIG96. To account for the HI asymmetry, they rule out any encounter with a massive companion as well as any close or parabolic passage of another smaller galaxy. They leave the door open for a parallel passage through the equatorial plane of CIG96 at an intermediate distance, that is, outside the optical disc but within the extended HI disc. [@espada11a] studied the Kennicutt-Schmidt SF law and efficiency in the large atomic HI disc of CIG96, using the VLA observations mentioned in this work (see Sect.\[sec:HIobs\]), as well as NUV and FUV observations from GALEX. By comparing the VLA maps and UV images, they found a good spatial correlation between the HI and both NUV and FUV emission, especially outside the inner 1$\arcmin$. Also, the main star-forming regions lie on the enhanced HI emission of two spiral arm-like features that correspond to the HI pseudo-ring. They found that the (atomic) Kennicutt-Schmidt power-law index systematically decreases with the radius. Regarding the star formation efficiency (SFE), they saw that it decreases with radius where the HI component dominates and that there is a break in this correlation at *r* = 1.5$\times\ r_{25}$. However, mostly within the HI pseudo-ring structure, that is, between 1.5$\times\ r_{25}$ and 3.5$\times\ r_{25}$, SFE remains nearly constant. They concluded that this might be a common characteristic in extended UV disc galaxies and that a non-axisymmetric disc can drive the outer spiral arms, as the morphology of the galaxy allows.
In this work we present new and deep HI and optical data of CIG96 to study in detail its faint gaseous and stellar components as well as its surroundings in order to reveal any possible causes of its HI asymmetrical distribution and other effects on its evolution. Throughout this study, all mentions to distances between different parts of the galaxy and its surroundings are projected distances unless stated otherwise. Also, we have assumed a cosmology with $H_{0}$ = 75 kms$^{-1}$ Mpc$^{-1}$, $\Omega_{\Lambda 0}$ = 0.73 and $\Omega_{m 0}$ = 0.27.
--------------------- ------------------------------------
$\alpha_{(2000)}$ 02$^{h}$15$^{m}$27.6$^{s}$
$\delta_{(2000)}$ +6$^{\circ}$00$\arcmin$09$\arcsec$
Type SAB(rs)c
Distance 20.3 Mpc
$r_{25}$ 2.35$\arcmin$ / 13.9 kpc
Inclination 46.59$\degree$
$M_{dyn,\ CIG\ 96}$ 1.78 $\times$ 10$^{11}\ M_{\odot}$
Position angle 20.0$\degree$
A$_{int}$($r$) 0.185
A$_{k}$($r$) 0.006
A$_{int}$($g$) 0.255
A$_{k}$($g$) 0.008
--------------------- ------------------------------------
: Parameters of CIG96 (NGC 864)[]{data-label="table:cig96"}
Description of the observations and data processing {#sec:description}
===================================================
In this section we present all the HI and optical observations of CIG96 used in this work as well as the reduction and calibration processes we followed to obtain the final images. The most relevant data are summarised in Tables \[tableHI\] and \[tableOPT\].
HI observations {#sec:HIobs}
---------------
In two different epochs, 21 cm line observations of CIG96 were made using the NRAO Karl G. Jansky Very Large Array (hereafter VLA or EVLA) observatory. First, two VLA projects AV0276 and AV0282 were performed in July 2004 and July 2005, respectively. We obtained 3 hours in D-configuration (26 antennas used) and 7 hours in C-configuration (27 antennas used), respectively. Both observing projects had the same set up: 2 IF correlator mode, a bandwidth of 3.125 MHz per IF and a frequency resolution of 48.8 kHz that corresponds to a velocity resolution of 10.4kms$^{-1}$. Second, the Extended-VLA (EVLA) project 13A-341, fully dedicated to observing CIG96, was executed during 2013 as follows: 3 hours in March, in D-configuration; 3 hours in May in the hybrid DnC-configuration and 10 hours in July, in C-configuration. In all cases, 27 antennas were used. The set-up of these observations consisted of single IF correlator mode, a bandwidth of 2 MHz and a frequency resolution of 16 kHz, equivalent to a velocity resolution of 3.3kms$^{-1}$ that was smoothed to 10kms$^{-1}$ for the calculations. These data are summarised in Table\[tableHI\].
----------------- ----------------- ---- ------------------ ------ ------ ----- --
VLA 1249.5$-$1895.2 10 27.1$\times$23.6 0.31 2.68 1.5
EVLA 1330$-$1700 13 37.6$\times$20.0 0.84 6.19 4.1
VLA$+$EVLA 1330$-$1700 19 28.0$\times$28.0 0.25 1.78 1.4
VLA$+$EVLA$+$WF 1330$-$1700 19 28.0$\times$28.0 0.13 0.89 0.7
----------------- ----------------- ---- ------------------ ------ ------ ----- --
All VLA and EVLA data were fully calibrated and imaged using CASA software package [@casa] tasks. We used the CLEAN algorithm [@hogbom74] to produce the final datacube. Each data set or measurement set (MS) was scanned to remove bad data and RFI (radio-frequency interferences). They were separately calibrated in phase, amplitude and bandpass and imaged individually to check their suitability for our aims. We produced a set of two individual datacubes by combining all VLA data and all EVLA data, respectively. We discarded the hybrid DnC-configuration data due to the presence of a remarkable amount of RFI, making them too defective for our goals. The VLA data consisted of two individual MSs, one for D-configuration data and one for C-configuration data. EVLA data consisted of thirteen individual MSs: three MSs were obtained in D-configuration and ten MSs in C-configuration. All HI masses in this work have been computed as given by [@roberts62] and [@roberts75]:
$$\label{eq:mhi}
$$M_{HI}\ (M_{\odot}) = 2.356\times10^{5}\ D^{2}\ S\ \Delta V$$$$
where $D$ is distance ($Mpc$) and $S \Delta V$ is the velocity integrated HI flux ($Jy\ km s^{-1}$).
The column density $N_{HI}\ (cm^{-2})$ depends on the brightness temperature $T_{B} (K)$ integrated over the line width $dv$ ($km\ s^{-1}$). In turn, $T_{B}$ depends on the flux density $S$ $(Jy\ beam^{-1})$ and the product of the major and minor axes $Maj \times min$ ($arcsec^{-2}$). Respectively:
$$\label{eq:tb}
$$T_{B}\ (K) = 6.07\times10^{5}\ S \left(Maj\ \times \ min\right)^{-1}$$$$
$$\label{eq:nhi}
$$N_{HI}\ (cm^{-2}) = 1.823\times 10^{18} \int T_{B}\ dv$$$$
*VLA data cube.* All VLA data were used to produce a preliminary datacube via imaging using natural weighting. This led to a synthesized beam of 27.11$\arcsec$ $\times$ 23.60$\arcsec$ and a root mean square (rms) noise level of 0.31mJybeam$^{-1}$ (1$\sigma$), reaching a HI column density limit of $N_{HI} = 2.68\times$10$^{19}$ cm$^{-2}$ (5$\sigma$). Assuming a HI line width of 10 kms$^{-1}$, the achieved HI mass detection limit is $\sim 1.5 \times$10$^{6} M_{\odot}$ (5$\sigma$) and a HI column density of 2.7$\times$10$^{19}$ cm$^{-2}$ (5$\sigma$). *EVLA data cube.* All EVLA data in C and D configurations were combined and imaged with natural weighting in a preliminary datacube. This datacube had a median rms of 0.84 mJy beam$^{-1}$ (1$\sigma$) and a synthesized beam of 37.57$\times$ 19.97. Such beam elongation is due to the short right ascension range in which the observations were taken. With a velocity resolution of 10 kms$^{-1}$, the HI mass detection limit achieved was of $M_{HI}$ = 4.1$\times$10$^{6}$ M$_{\odot}$ (5$\sigma$).
*Combined EVLA and VLA data cubes: hereafter the HI cube.* After the rms-weighted[^3] concatenation of the VLA MS and EVLA MS we produced the final datacube of this work (hereafter referred as the HI cube). The corresponding weighting factors applied to the VLA and EVLA data were 10.40 and 1.42, respectively. The HI cube comprises a total of 19hours on target and has a synthesized beam of 28.16$\times$22.72 (2.77kpc$\times$2.24kpc at a distance of 20.3Mpc); it covers a velocity range from 1330kms$^{-1}$ to 1800 kms$^{-1}$ in 48 channels assuming spectral resolution of 10 kms$^{-1}$. We used the kinematical local standard of rest (LSRK) as the frame of reference for the radio velocities. Also, we worked with a smoothed beam of 28$\times$28 to simplify the physical interpretation of the results and avoid beam effects. The corresponding HI cube yielded a median rms of 0.25mJybeam$^{-1}$(1$\sigma$) that allowed us to reach a HI mass limit of $M_{HI}^{lim}\simeq$1.4$\times$10$^{6}$$M_{\odot}$(5$\sigma$) and a HI column density limit of $N_{HI}\simeq$1.78$\times$10$^{19}$cm$^{-2}$ (5$\sigma$). After performing a wavelet filtering (see Sect.\[sec:wavelets\]) over the HI cube, we improved these results by a factor of approximately two, reaching a final median rms of 0.126mJy beam$^{-1}$(1$\sigma$) per channel. The minimum HI mass detected is $M_{HI}^{lim} =$0.7$\times$10$^{6}$ $M_{\odot}$(5$\sigma$), the HI column density limit is $N_{HI} =$8.9$\times$10$^{18}$cm$^{-2}$(5$\sigma$) and the total HI mass is $M_{HI}^{total} =$9.77$\times$10$^{9} M_{\odot}$(5$\sigma$). The integrated intensity map, the velocity field and the channel maps are all presented in Sect.\[sec:HIres\].
Wavelet filtering of the HI cube {#sec:wavelets}
--------------------------------
A robust detection of faint HI features relies on reaching a column density ($N_{HI}$) that is as low as possible with the best signal$-$to$-$noise ratio (S/N). In order to further improve our $N_{HI}$ limit, we have applied a wavelet filtering to our HI cube which allows to achieve a higher S/N. An in-depth discussion of the wavelet transform is beyond the scope of this paper but we provide here an explanation of the method used in this work. As explained by [@leon16], the wavelet transform is a powerful signal-processing technique that provides a decomposition of the signal into elementary local contributions defined by a scale parameter [@grossmann]. The wavelets are the scalar products of shifted and dilated functions of constant shape. The data are unfolded in a space-scale representation that is invariant with respect to dilation of the signal. Such an analysis is particularly suited to studying signals that exhibit space-scale discontinuities and/or hierarchical features, as may be the case for the possible structures located in the outskirts of the HI envelope of CIG96.
Following the same procedure as [@leon16], we have used a $B_3$-spline scaling function defined by the following convolution matrix $M$:
$$M=
\left[ {\begin{array}{ccccc}
{{}^{1}\!/_{256}} & {{}^{1}\!/_{64}} & {{}^{3}\!/_{128}} & {{}^{1}\!/_{64}} & {{}^{1}\!/_{256}} \\
{{}^{1}\!/_{64}} & {{}^{1}\!/_{16}} & {{}^{3}\!/_{32}} & {{}^{1}\!/_{16}} & {{}^{1}\!/_{64}}\\
{{}^{3}\!/_{128}} & {{}^{3}\!/_{32}} & {{}^{9}\!/_{64}} & {{}^{3}\!/_{32}} & {{}^{3}\!/_{128}}\\
{{}^{1}\!/_{64}} & {{}^{1}\!/_{16}} & {{}^{3}\!/_{32}} & {{}^{1}\!/_{16}} & {{}^{1}\!/_{64}}\\
{{}^{1}\!/_{256}} & {{}^{1}\!/_{64}} & {{}^{3}\!/_{128}} & {{}^{1}\!/_{64}} & {{}^{1}\!/_{256}} \\
\end{array} } \right]$$
Similar to the Ricker function (mexican hat), it has a positive kernel surrounded by a negative annulus and the total integrated area is zero.
We have applied this wavelet over the HI calibrated data via the *A trous* algorithm (see @bij) as described by [@leon00]. This algorithm creates different filtered wavelet planes according to the scale parameters and a certain threshold level. The scale parameters have received values of 2$^{i}$ with $i$$\in$\[1,6\], each defining the $i-$th plane. Each $i-$th raw wavelet plane is defined as the subtraction of two components that, in turn, depend on the $i-$th scale parameter: the zeroth component corresponds to the image plane itself; the rest of the $i-$th components are defined as the result of convolving the $i-1-$th component with the previously defined kernel function. The last plane, namely, the last smoothed plane or LSP (in our case, scale parameter of 2$^{6}$) does not undergo any convolution; therefore, it is not a wavelet plane itself but the residuals of the last convolution. With the consequent exception of the LSP, each raw plane is filtered above a threshold to construct the $i-$th filtered wavelet plane. For this work, such a threshold was set at 5$\sigma_{i}$, where $\sigma_{i}$ is the rms noise for the $i-$th plane.
The combination of the filtered wavelet planes and the LSP is possible and may cause the rms to change. Since the original image is spread in different spatial scales, a limited combination of the planes implies the recovered flux will be a lower limit to the total emission contribution. Should all planes be combined, the recovery is complete and the total flux is conserved.
After filtering our HI cube, we combined all planes. The resulting rms and, accordingly, the HI column density limit, were improved by a factor of two, as specified in the last paragraph of Sect.\[sec:HIobs\] and summarised in Table\[tableHI\].
Blanking of the HI cube {#sec:lownoise}
-----------------------
We separated genuine emission from noise by blanking non-signal pixels using the following method.
First, we applied a spatial smoothing over the wavelet-filtered HI cube by convolving it with a Gaussian kernel four times the size of the adopted synthesized beam, that is, 56$\times$56. The resulting smoothed datacube was only used to create the masks, as described below, and its noise was $rms =$0.34mJybeam$^{-1}$ (1$\sigma$). Second, we created masks for each channel of the smoothed datacube. The shapes of these masks were defined by masking out the pixels with values below a 3.5$\times rms$ threshold ($\sim$1.2mJybeam$^{-1}$). Finally, the masks from the smoothed datacube were applied to the original datacube (non spatially smoothed) to create the moment maps[^4].
This method mainly has two advantageous consequences: one, the depth and spatial resolution of the original datacube remain unaffected by the masking and two, the threshold limit, for the integration, does not take into account the areas in each channel whose only contribution is noise. In other words, the blanking of the HI cube helps us to remove any remaining effect from the side lobes (either positive or negative) that might mimic nonexistent structures.
Optical observations {#sec:optobs}
--------------------
In order to obtain deep optical images of the outskirts and close environment of CIG96, we performed observations in three different observatories. Two datasets were observed with the 2.2m and 1.23m telescopes, respectively, at CAHA[^5] observatory in Spain. The first dataset is from CAHA2.2m, a deep image with good seeing in the Cousins$R$ band (see Sect.\[sec:22data\]). The second dataset consists of three images taken with photographic $B$, $G$, $R$ bands used to study colour index properties (see Sect.\[sec:123data\] and all 2.2m and 1.23m images combined in Fig.\[opt22\_123\]). The third dataset was obtained with the VLT Survey Telescope (ESO[^6]) in Chile (hereafter, VST) and provides a very deep and wide field image to study the surroundings of the galaxy (see image in Fig.\[optvst\] and Sect.\[sec:22data\]). The most relevant data are summarised in Table\[tableOPT\].
------------- -------------- ------------ ------ ----------------------------------------- ---------------- ------ ------
CAHA2.2m Cousins R 2$\times$2 1.04 3h 56m 12$\times$12 27.5 1.59
(CAFOS) (71 $\times$ 200 s) 71$\times$71
CAHA1.23m B, G, R 1$\times$1 1.04 3h 38m 15$\times$16 $-$ 1.56
(DLR-MKIII) photographic (30$_B$, 37$_G$, 42$_R$ $\times$ 120 s) 88$\times$94
VST SDSS $r$ 2$\times$2 0.21 5h 10m 60$\times$60 28.7 1.10
(OmegaCAM) (122 $\times$ 154 s) 350$\times$350
------------- -------------- ------------ ------ ----------------------------------------- ---------------- ------ ------
![A $12\arcmin \times12\arcmin$ combined image of the Cousins$R$ image from CAHA2.2m telescope and the three photographic $B$, $G$, $R$ images from CAHA1.23m telescope. This particular image is only used to show the outer faint structures of the galaxy (e.g. the pseudo-ring, the northern region in a magenta ellipse or the eastern diffuse emission pointed out by the yellow arrow), not for any physical measurement. The inner coloured area corresponds to an SDSS image of CIG96 down to $\sim$24magarcsec$^{-2}$ (SDSS$r$ band) and is used as reference.[]{data-label="opt22_123"}](cig96_allCAHASDSS_marks_invert){width="\hsize"}
![A $12\arcmin \times12\arcmin$ detail of the VST optical image of CIG96 with the SDSS colour image down to $\sim$24magarcsec$^{-2}$ (SDSS$r$). The red contour is set on 26.5magarcsec$^{-2}$ (SDSS$r$), to point out the faintest SB level of the pseudo-ring.[]{data-label="optvst"}](cig96_VST_linear26_29_265SBcont3sm){width="\hsize"}
### CAHA2.2m dataset {#sec:22data}
CIG96 was first observed in the second half of the night of September 11, 2012, with the CAFOS instrument at CAHA2.2m telescope. The CAFOS SITe1d detector has 2048$\times$2048 pixels with a pixel size of 24$\mu m$ (spatial scale of 0.53 pix$^{-1}$), providing an effective circular field of view of $\sim$12$\arcmin$ in diameter.
A total of 71 exposures of 200 s each build up a total time on source of 3h 56min. All images were taken in the Cousins$R$ filter, dithered by $\sim$20 and in 2$\times$2 binning mode, providing a pixel scale of 1.04/pixel. The night conditions were photometric during most of the night, with a median seeing of 1.59 (seeing ranging from 1.31 to 1.81). We used standard reduction and calibration techniques from *repipy* and *LEMON* packages [^7] and IRAF. No standard stars were measured in this campaign and so the extinction coefficient was computed by means of non-saturated stars present within the field of view of our observations. As a consequence, a larger uncertainty is introduced in the photometric calibration. In order to obtain the Zero Point of the night, we computed the Bouguer fit of eight non-saturated stars (visible in all images) and calibrated them with the corresponding data from SDSS [@ahn12]. Since this dataset was taken using Cousins$R$ filter, all fluxes were converted from SDSS magnitudes system to Cousins$R$ using the transformation by [@lupton], derived by matching photometry data from SDSS Data Release 4 (DR4) to Peter Stetson’s published photometry for stars: $$\label{eq:sdssRri}
\begin{gathered}
$$R_{ri} = r - 0.2936*(r-i) - 0.1439$$
\end{gathered}$$
in magnitudes, where $r$ and $i$ are the magnitudes in the SDSS$r$ and SDSS$i$ filters, respectively. The median Zero Point of the night (Cousins$R$ filter) is 24.28$\pm$0.12mag. We calculated the SB of the image by setting 40 square boxes of 20$\times$20size in the southern, western and northern areas of the image. The eastern side of the CAHA2.2m image is heavily contaminated by a star so we did not take into account any SB measurements of that side. There is a slightly uneven distribution of the light between the western side (median $\mu_{Cous\ R}=$27.5magarcsec$^{-2}$) and the northern and southern sides ($\mu_{Cous\ R} = $28magarcsec$^{-2}$). We cannot confirm whether the 0.5magarcsec$^{-2}$ difference comes from the residuals of the flat-fielding or from reflected light and the small field of view of the image prevents selecting a SB value over the rest so we set the SB limit of the image as the lowest value, $\mu_{Cous\ R}=$27.5magarcsec$^{-2}$ (approximately $\mu_{SDSS\ r} =$28.0magarcsec$^{-2}$).
### CAHA1.23m dataset {#sec:123data}
CIG96 was observed for a second time on the night of December 8, 2012 with the DLR-MKIII instrument at the CAHA 1.23m telescope. The camera is equipped with an e2v CCD231-84-NIMO-BI-DD sensor (4k$\times$4k pixels, 15$\mu m$pix$^{-1}$). The original field of view is 21.5$\times$21.5 but the observations were cropped down to the central 15$\times$16.
In this case, we used three different filters: photographic B, G and R (different from Johnson-Cousins) for which a total of 30, 37 and 42 exposures of 120s each were taken, respectively, in 1$\times$1 binning mode. The night conditions were stable for the most part of the night and all filters present a median seeing of 1.56 (seeing range from 1.48$\arcsec$ to 1.61$\arcsec$). The total integration time was 3h38min.
As with the previous dataset, standard reduction was applied to all the images in each filter separately. However, they were divided by a blank field. It was obtained from an adjacent galaxy-free field and corrected for bias and regular flat field too, so the remaining image would not show any residual gradient. Dividing the images by this blank field allows large-scale structures to be removed. We used the SDSS tabulated fluxes from several stars to calibrate the images via the following relation between SDSS and photographic filters: B(3900-5100 angstrom) would correspond to SDSS$g$ and R(5800-7000 angstrom) to SDSS$r$. However, G(4900-5800 angstrom) would lie right between SDSS$g$ and $r$ bands. For the conversion of G band to SDSS, we considered different scenarios in which the emission was split between SDSS$g$ and $r$ bands but it has not been used further in this work. Hereafter we focus on the empirical relations that we calculated for R and B bands with respect to SDSS$r$ and $g$. The initial relations between the corresponding magnitudes (not corrected from extinction) are: $$\label{eq:rRnoext}
$$m^{+ext}_{r_{\ SDSS}} = 1.01*m_{R_{phot}} - 9.83 \pm0.15$$$$ and $$\label{eq:gBnoext}
$$m^{+ext}_{g_{\ SDSS}} = 0.99*m_{B_{phot}} - 9.70 \pm0.33$$$$
Internal extinction and $k-$correction were applied to the fluxes in both $g$ and $r$ bands. We used the extinction laws by [@savage79] (in agreement with @fitzpatrick99) where $A(B) = 4.10 \times E_{B-V}$; the internal extinction and $k$-correction in the B band for CIG96 are $A_{int}(B) = 0.276$ and $A_k(B) = 0.009$, respectively [@fernandezlorenzo12]; the extinction-reddening relations for the SDSS bands are $A_x(g) = 3.793 \times E_{B-V}$ and $A_x(r) = 2.751 \times E_{B-V}$ [@stoughton02].
These relations yield the following internal and $k$-correction values for each band: $A_{int}\,(g)\,=\,0.255$, $A_{int}(r)\,=\,0.185$, $A_k(g)\,=\,0.008$ and $A_k(r)\,=\,0.006$.
Hence, the final empirical extinction$-$corrected equations that convert photographic B and R bands to SDSS$g$ and SDSS$r$ bands are: $$\label{eq:rR}
$$m_{r_{SDSS}} = 1.01*m_{R_{phot}} - 10.02 \pm0.15$$$$ and: $$\label{eq:gB}
$$m_{g_{SDSS}} = 0.99*m_{B_{phot}} - 9.96 \pm0.33$$\\$$
Finally, the images were average stacked applying an outlier-rejection algorithm.
With the two images from B and R bands already calibrated to SDSS$g$ and SDSS$r$ bands respectively, we built a $g-r$ image with the aim of studying the colour distribution in the most interesting regions of the galaxy (see Sect.\[sec:pseudocolor\]).
In Fig.\[opt22\_123\] we show the result of combining the reduced CAHA2.2m image (Cousins$R$ band) and the three reduced CAHA1.23m images (photographic filters). The lower resolution of these images (compared to the better resolution of VST, see Sect.\[sec:vstdata\]) provides a more clear visualization of the external structures of CIG96, especially the faint structure in the N and the very diffuse E side of the pseudo-ring, indicated in the image. However, we cannot calibrate them all to a common band, so this image must be taken only as an illustrative view of the galaxy.
### VST dataset {#sec:vstdata}
In order to study the larger-scale structure surrounding CIG96, we also observed the galaxy with OmegaCAM at the VST (runID: 098.B-0775(A)). This instrument has a field of view of 1 square degree sampled with a 32-CCD, 16k$\times$16k detector mosaic at 0.21/pix.
The 32 CCDs have intermediate spaces between the different chips in the vertical direction (5.64mm top and bottom gaps; 0.82mm central gap) and in the horizontal direction (1.5mm gap). Also, at the time these observations were designed, the user manual accounted for cross talk between CCDs 93 to 96 at <0.4% level (slightly above our aim of 0.35%). Further discussion with the telescope staff alerted to irregular gain variations in CCDs 82, 87 and 88. In order to avoid these CCDs as much as possible and guarantee a homogeneous coverage of the gaps, we initially designed a manual diagonal dithering pattern for the pilot observations exposures to sample the galaxy and its surroundings. With it, the 49 different offset positions of the galaxy (7 pointings with 7 offset positions each) were placed along a diagonal oriented from the southeast (SE) to the northwest (NW) of the chip, always leaving at least 1(both in RA and Dec) with the edge of the CCDs. After the pilot observations, we concluded that the previous diagonal dithering would not significantly differ from the already existing modes (JITTER and DITHER, since STARE was not useful for our aims) so we designed a new manual dithering pattern that would make a total of seven pointings, six of them to the corresponding apexes of a slightly irregular hexagon-shaped pattern plus one more central pointing.
A total of eight observing blocks (OBs) of 1 hour each were dedicated to observing CIG96. From these, 7OBs had 16 exposures and 1OB had 10 exposures, making a total of 122 exposures of 154s each. The 8OBs were carried out on the nights of October 6, 9 and 20, November 1 and 2 and December 2, 3 and 20, all in 2016. The total time spent on source was 5h 10 min. All observations were done under the following conditions: photometric sky transparency, maximum seeing of 2.0, airmasses below 2.0 (>63.4$\degree$), with an angular distance to the Moon of at least 60$\degree$ and its maximum illumination at 30%. We used a modified version [@venhola17] of Astro-WISE pipeline [@mcfarland13] to reduce and calibrate these data. The SB of the image was calculated as the median of the SB values computed in $\sim$60 square boxes of 20$\times$20spread in the central 40$\times$40of the image and avoiding stars. For this we used equation \[eq:SBvst\], in which the second term corresponds to the conversion from pix$^{-2}$ to arcsec$^{-2}$:
$$\begin{split}
\label{eq:SBvst}
\mu_{SDSS\ r} &= -2.5*log(F_{SDSS\ r}) - 2.5*log(0.21^{2}) \\ \\
&= -2.5*log(F_{SDSS\ r}) + 3.3889
\end{split}$$
Figure\[optvst\] shows an SDSS colour image of CIG96 on top of a subset of the VST image. Additionally, the faint SB reached with this image allowed us to detect Galactic cirrus around the galaxy (see Sect.\[sec:cirrus\]).
![*Top panel*: Integrated profile of CIG96, calculated from the EVLA and VLA combined HI cube (blue solid line), integrated spectrum of CIG96 (LSRK) obtained by [@haynes98] at Green Bank 43 m (heliocentric) (pink dashed line). Our integrated spectrum shows a central velocity that is lower than the Green Bank spectrum, therefore in order to match and facilitate the comparison between the two, we have shifted the latter by $-$17kms$^{-1}$. The green solid line is the integrated profile of the closest companion of CIG96: NGC864COM01. The horizontal green dashed line sets the width at 20% of the highest flux peak ($W20$) for the central radio velocity computation, shown as a blue dot ($V_{LSRK}\ (CIG\ 96) = 1544.15$ kms$^{-1}$). The vertical blue dotted line defines the two halves of the spectrum for the asymmetry parameter calculation. *Bottom panel*: Integrated HI profile of the companion NGC864COM01 with a rescaled flux density for an easier visualization. The green dot sets the central velocity of this galaxy ($V_{LSRK}\ (companion)$ = 1577.90 kms$^{-1}$).[]{data-label="lsrk"}](cig96-lsrkvelocity){width="\hsize"}
Planck and WISE images {#sec:planckwise}
----------------------
In order to inspect the cirrus around CIG96 (see Sect.\[sec:cirrus\]), we used images from the HFI camera of the Planck satellite at 857GHz / 350$\mu m$ band [@planck14]. Also, we have used a WISE band 3 image (12$\mu m$) since this band that traces hot dust and shows good correlation with the cirrus emission [@miville16]. Throughout this work, we will refer to these images as Planck857 and WISE3, respectively. Planck857 images were obtained from SkyView online tool [@mcglynn94] while the WISE3 image was obtained from the IRSA, NASA/IPAC archive and was reprocessed to improve the flat fielding and remove the stars.
HI results {#sec:HIres}
==========
Integrated emission and asymmetry level {#sec:velasym}
---------------------------------------
To calculate the total spectrum, we integrated the emission of each channel of the HI cube. Then, as discussed by [@fouque90], we computed the central velocity of the galaxy $V_{cen}$ as the average between the lowest and highest velocities measured at a width (or flux level) of the 20% of the highest flux peak in the integrated spectrum (abbreviated $W20$, name varies depending on the percentage used). The error can be estimated as: $\Delta V$=4$\frac{\sqrt{\delta\nu\,(W20 - W50)/2}}{S/N_{peak}}$ where $\delta\nu$ is the spectral resolution of the cube, $(W20 - W50) / 2$ represents the steepness of the edges of the HI profile at 20% and 50% of the maximum flux, and $S/N_{peak}$ is the S/N of the maximum flux peak. Taking these into account, the $W20$ central radio velocity of our HI cube is $V_{LSRK} (CIG\,96)$ = 1544.15 $\pm$ 0.23 kms$^{-1}$.
We find a difference of approximately 10 kms$^{-1}$ between our result for the central velocity of CIG96 and those calculated from single-dish data by [@espada05] (same method as in this work) and [@haynes98], 1561.6 and 1562 $\pm$ 1 kms$^{-1}$, respectively, both in heliocentric frame of reference, that is, approximately 1553 and 1554 kms$^{-1}$ when converted to LSRK, as is ours. [@kerr86] also provide a LSRK velocity of 1553 $\pm$ 1 kms$^{-1}$, showing the same shift with respect to our result. To identify the reason for this apparent inconsistency, we recalculated the central velocity of our HI cube and the one published by [@espada05] in different standards of rest and in the two optical and radio velocity conventions. In all cases, the differences remained within a few kms$^{-1}$, i.e., no change in the standard of rest or velocity convention would account for such a shift. The calibration process was also revised and the correct rest frequency for the HI line was confirmed, leaving us with the only hypothesis of an undetected error in the raw data or the calibration process. Taking this into account, we conclude this difference may be assumed, not to affect the interpretation of the data in any case since it is a small shift compared to the width of the profile.
CIG 96 has a close companion: NGC864COM01 (hereafter also referred to as the companion), detected in HI by [@espada05]. We determine a $W20$ central radio velocity of $V_{LSRK}$ (companion) = 1577.90 $\pm$ 2.62 kms$^{-1}$. The HI and optical properties of this galaxy, as well as its implication in the isolation of CIG96, are discussed further in Sect.\[sec:cigcomp\], \[sec:companion\] and \[sec:environment\].
In Fig.\[lsrk\] we compare the integrated emission spectra derived from our HI cube for CIG96 and its companion with the one obtained by [@haynes98] using data from Green Bank 43m single dish telescope, and still in the heliocentric system of reference. For a better comparison between the two spectra, we have shifted the latter by $-$17 kms$^{-1}$. The perfect match between them strongly suggests that our HI cube has a velocity shift of $-$10 kms$^{-1}$, after converting all velocities to LSRK.
In order to estimate the HI asymmetry level of a galaxy, quantified as A$_{flux\ ratio}$ [e.g. @haynes98; @kornreich01; @espada11b] we also use the HI integrated spectrum. A$_{flux\ ratio}$ is an areal asymmetry parameter defined as the emission ratio set between the two regions of the spectrum defined with respect to the central velocity and its lowest and highest velocity. While it provides a simple quantification of the gas distribution in the two halves of the galaxy, this global parameter does not give spatial information of any possible asymmetry.
We calculated the sources of the uncertainties of this parameter as described by [@espada11b], obtaining A$_{flux\ ratio}$=1.16$\pm$0.01, that is, 16 $\pm$ 1%, in full concordance with [@espada05].
Channel maps {#sec:chmap}
------------
The channel maps allow to inspect every channel of the HI data cube. Each one corresponds to a different velocity allowing us to trace any structures that might be connected to the gaseous envelope of the galaxy. In Fig.\[figvelchannelwz1\] we show a subset of the channel maps of the wavelet filtered HI cube on top of the CAHA1.23m optical image (band R) of CIG96. This image corresponds to the central $25\arcmin \times 25\arcmin$ of the primary beam and to the channels with emission, that is, from 1380 to 1690 kms$^{-1}$ (channels 6 to 37, respectively) where the channel width is 10 kms$^{-1}$. The systemic velocity of the galaxy ($V_{LSRK}\ (CIG\ 96) = 1544.15$ kms$^{-1}$, see Sect.\[sec:velasym\]) corresponds to channel 23 and the approaching and receding sides of the galaxy extend approximately 135 and 145 kms$^{-1}$, respectively. The synthesized beam is 28$\arcsec \times28\arcsec$, the rms is 0.126 $mJy\ beam^{-1}$ and the column density reached is $N_{HI} (5\sigma) = 8.9\times10^{18} cm^{-2}$.
The HI distribution is more symmetrical in the central channels ($\sim$1500 $-$ 1600 kms$^{-1}$) than in those with velocity differences of $\Delta V \geqslant$60 kms$^{-1}$ with respect to the central velocity. In the latter, the approaching side shows that the HI has a uniform distribution over a larger area in the southwest (SW) than in the receding side, where the distribution is more narrow and oriented towards the northeast (NE). The HI extension also differs, reaching $\sim$7.9($\sim$47 kpc) towards the SW and $\sim$9.3($\sim$55 kpc) towards the NE. Also, the receding NE side is less massive, as reflected in the asymmetrical shape of the integrated spectrum (Sec \[sec:velasym\]). In both the approaching and receding sides, the HI is extended beyond 4$\times r_{25}$ of the optical extension. From 1630 to 1670 kms$^{-1}$ (channels 31 to 35), there is a change in the orientation of the HI, especially visible in column densities below 1.0$\times10^{19}$ cm$^{-2}$ (outer contours of Fig.\[figvelchannelwz1\] and moment maps shown in Sect.\[sec:HImoments\]).
Focusing on the outermost regions, we note two previously undetected features:
- First, from 1480 to 1550 kms$^{-1}$ (channels 16 to 23), we notice a clumpy structure to the NW of the galaxy ($\alpha = 02^{h}15^{m}05.9^{s},\ \delta = 6\degree03\arcmin03\arcsec$), with an approximate size of $\sim$21 kpc ($\sim$3.5 arcmin, measured from channel 17 to 22), a column density of approximately $N_{HI}^{NW} \simeq$ 6.5$\times$10$^{19}$ cm$^{-2}$ and a total HI mass of $M_{HI}^{NW\ feat.} \simeq 3.1 \times 10^{6} M_{\odot}$. We refer to this as the NW HI feature and it is indicated with green marks in Fig.\[figvelchannelwz1\].
- Second, from 1600 to 1640 kms$^{-1}$ (channels 28 to 32), a structure shows up to the SE of the galaxy ($\alpha = 02^{h}15^{m}41.0^{s},\ \delta = 5\degree55\arcmin31\arcsec$), within a square region of approximately 8.8$\times$8.8 kpc ($\sim$90$\times$90) size, a column density of approximately $N_{HI}^{NW} \simeq$ 4.9$\times$10$^{19}$ cm$^{-2}$ and a total HI mass of $M_{HI}^{SE\ feat.} \simeq 1.6 \times 10^{6} M_{\odot}$. We refer to this as the SE HI feature and it is indicated with magenta marks in Fig.\[figvelchannelwz1\].
These structures are discussed further in Sect.\[sec:feats96\].
NGC864COM01, the HI rich companion of CIG96 {#sec:cigcomp}
-------------------------------------------
As described by [@espada05] and introduced in Sect.\[sec:velasym\], CIG96 has a small companion located at 15.2$\arcmin$ ($\sim$90 kpc) to the east with a *B* magnitude of $m_{B}=$ 16.38 mag. It shows emission throughout 11 channels (from 1540 to 1650kms$^{-1}$). Its central LSRK velocity is of $V_{LSRK}=$ 1577.90 kms$^{-1}$ and a total HI mass of $M_{HI} = 5.1\times10^{6}\ M_{\odot}$. The HI image of this galaxy is shown in Figs.\[mom0\] and \[mom1\]. Both CIG96 and its companion share a similar orientation of their minor axis. However, they show different kinematical orientation, that is, the companion is counter-rotating with respect to CIG96, and we do not find any signs of tidal features between them. The galaxy is studied further in Sect.\[sec:companion\].
Moment maps and position-velocity profiles {#sec:HImoments}
------------------------------------------
The integration of the flux density $S$ (or zeroth moment) is carried out from channel 6 (1380 kms$^{-1}$) to channel 38 (1700 kms$^{-1}$), i.e. one additional channel beyond the HI emission. The velocity field (or first moment) is the intensity-weighted velocity of the spectral line, i.e., a measure for the mean velocity of the gas. The zeroth moment is shown in Figs. \[mom0\] and \[mom0opt\]. The HI extends beyond 4$\times r_{25}$, that is, approximately up to 50 kpc (8.5$\arcmin$), reaching an integrated column density of $N_{HI}\ (5\sigma) = 1.2\times10^{20}$ cm$^{-2}$ with a beam size of 28$\times$ 28. As a comparison, in Fig.\[mom0\] we indicate with a black line the approximate $N_{HI}\ (5\sigma) = 8.7\times10^{20}$ cm$^{-2}$ column density reached by [@espada11a] with a beam size of 16.9$\times$ 15.6. Quantitatively, the current observations are roughly seven times deeper than the previous ones.
The first moment is shown in Fig.\[mom1\]. It allows the estimation of the position angle (from now on, PA) of the major and minor kinematical axes of the galaxy, indicated by the two black lines at PA=20$\degree$ and PA=110$\degree$, respectively. We have performed the position-velocity (P/V) profiles over the HI cube along the major and minor axes, as shown in Fig.\[pvcuts\]. The emission located at the largest radius in the SW region (indicated with a cyan arrow in the profile over the major axis, Fig.\[pvcuts\], top panel) was already detected by [@espada05]. It is visible in the channel maps at 1450 - 1470 kms$^{-1}$ (channels 13 to 15) and it shows a drop in velocity of about 30 $-$ 40 kms$^{-1}$ with respect to inner parts of the galaxy. Both its extension and velocity drop are in agreement with the previous work. The interruption in the emission to the NE is due to a $\sim$3$\times$3 kpc$^{2}$ region ($\sim$30$\times$30) with low HI emission. It is visible in the zeroth moment map ($RA = 2^{h}15^{m}34.935^{s}, DEC = 6\degree04\arcmin33.17\arcsec$) as well as in the channel maps at 1630 - 1640 kms$^{-1}$ (channels 31 and 32).
The P/V profile over the minor axis cuts through part of the NW HI feature (indicated with a red arrow, Fig.\[pvcuts\], bottom panel), the clumpy HI structure mentioned in Sect.\[sec:chmap\]. This feature shows a velocity gradient of $\sim$70 kms$^{-1}$ (approximately from 1480 to 1550 kms$^{-1}$) and it seems to connect with the galaxy in the channels around its central velocity (channels 23 to 25). Also, the central part of the galaxy shows emission in a wide range of velocities with respect to the central velocity. We discuss this effect further in Sect.\[sec:feats96\].
Optical data results {#sec:Optres}
====================
Surface brightness limit, dynamical masses and optical features {#sec:sblimit}
---------------------------------------------------------------
The images from CAHA2.2m and CAHA1.23m telescopes have a field of view of 12$\arcmin \times$12$\arcmin$, i.e. approximately 71$\times$71 kpc (see Fig.\[opt22\_123\]), while the VST covers 1$\degree \times$1$\degree$, that is, approximately a 350$\times$350 kpc field centred on the galaxy. The limiting SB reached is deeper than any other previously published, in particular with the VST image ($\mu_{r\ SDSS}$$(VST)$=28.7magarcsec$^{-2}$, see Fig.\[optvst\]) that reveals unprecedented detail of the extension, boundaries and structures of the external and faint pseudo-ring of CIG96 as well as its connection to the inner parts of the galaxy. The VST image also shows signs of Galactic cirrus (See Sect.\[sec:cirrus\]) so we set our reliable detection limit in $\mu_{r\ SDSS}$$(VST)$=28.4magarcsec$^{-2}$, just above the level where they start to become visible.
The total dynamical mass of CIG96 is $M_{dyn,\ CIG\ 96}$=1.78$\times$10$^{11}$M$_{\odot}$, following the calculation described by [@cou14]. It was estimated taking into account the inclination ($i$, in degrees, indicated in Table\[table:cig96\]), the radius of the galaxy along the major axis ($R$, in kpc) as well as the rotation velocity ($V$, in kms$^{-1}$). Both $R$ and $V$ are extracted from the HI data: $R$ of 6 (35.43kpc) from the rotation curve of the major axis (see Sect.\[sec:HImoments\]) and $V$ via measuring the velocity difference at such radius with respect the central velocity of the galaxy, resulting in 125kms$^{-1}$. The same calculation was made for the companion. We obtained a P/V cut of the galaxy along a PA of 35 to measure the peak $R$ and $V$, resulting in 35 (3.44kpc). However, with the current data we do not observe a turn over in the rotation curve so the mass calculation at this radius must be taken as a lower limit. We also assumed an inclination of 90 since it might be an edge-on galaxy (discussed further in Sect.\[sec:companion\]). The velocity extent measured at a 35 radius is of 60kms$^{-1}$. The dynamical mass of the companion is of $M_{dyn,\ comp}$=2.88$\times$10$^{9}$M$_{\odot}$. Hence, the dynamical mass relation between the host galaxy and its companion is approximately of $M_{dyn,\ CIG\ 96}/M_{dyn,\ comp}\simeq$62. The case of CIG96 can be considered similar to the one of the MW-mass galaxy M94 that, after a deep search performed as part of the recent work by [@smercina18], only shows two satellites.
The brightest stellar structures within the pseudo-ring ($\mu_{Cous\ R}\ (CAHA)$ = 25.5 $-$ 26.5magarcsec$^{-2}$) are located within a distance of *r*=1.5$-$2.0$\times$ $r_{25}$ from the galaxy centre (i.e. approximately 3.5$-$4.7or 15.0$-$20.5kpc). They are well defined and large to the north, thinner to the west and more diffuse to the south (see Fig.\[optvst\]). The east region shows very diffuse emission and no clear sign of the pseudo-ring structure, making the latter a partially closed pseudo-ring. Despite the SB limit reached, the numerous stars in the field and their PSFs may play a relevant role by overlapping with any fainter emission at such low SB, mimicking non-existent extragalactic stellar traces [@trujillo16]. In particular, this occurs in the eastern region where a few bright stars are located. However, the even deeper SB limit reached with the VST image has two immediate implications: one, the definition of certain regions of the pseudo-ring are greatly improved and two, the Galactic cirrus starts to become clearly visible at 28.5magarcsec$^{-2}$, hindering the detection of features beyond the pseudo-ring at SBs fainter than this level (see Sect.\[sec:cirrus\]).
Disc and pseudo-ring relative orientation {#sec:isophotes}
-----------------------------------------
A visual inspection of the CAHA2.2m optical image suggested an apparent misalignment between the pseudo-ring and the galactic disc. In order to quantify it, we performed elliptical fittings to the pseudo-ring structure as well as to the isophotes of the galaxy from 20.2 to 26.4magarcsec$^{-2}$ after removing the signatures of the close bright stars to avoid biased fittings.
The fittings of the innermost regions of the galaxy ($\mu_{Cous\ R}$=24.0magarcsec$^{-2}$ or brighter) were not reliable because of the strong influence of the spiral arms. Moreover, bright close stars contaminate the outer regions (fainter than $\mu_{Cous\ R}$=24.0magarcsec$^{-2}$). Even after removing them, too few points are left making reliable fittings difficult.
However, the optical images clearly showed the centre of the galaxy (error below 1). After fitting the pseudo-ring we found a shift of 12$\arcsec$ ($\sim$1.2kpc, the approximate length of the bar) between the centres of the pseudo-ring fitting and the disc and its orientation was PA$_{pseudo-ring\ fit}$=21.5, similar to the PA of the major axis of the galaxy (PA$_{maj}$=20). We also de-projected the image assuming a disc inclination of $i_{CIG\,96}$=46.59$\deg$ to confirm whether the pseudo-ring may be oval or in a different plane from the disc. We found the flattening or ellipticity of the pseudo-ring is of 0.04$-$0.05%, that is, practically circular, suggesting it to be slightly oval if seen at almost the same inclination as the inner disc of the galaxy.
![HI integrated intensity map of CIG96 and its companion after a 3.5$\sigma$ blanking (see Sect.\[sec:lownoise\]). We identify the NW and SE HI features mentioned in Sect.\[sec:chmap\] as well as the HI emission of the pseudo-ring. The black contour represents the column density of $N_{HI} = 8.7\times10^{20}$ cm$^{-2}$ (5$\sigma$) reached by [@espada11a]. The black circle at the bottom left represents the beam size of 28$\times$ 28.[]{data-label="mom0"}](cig96_mom0_overlay_nobar_feats){width="\hsize"}
![*Background*: VST optical image of CIG96 ranging from 26 to 28magarcsec$^{-2}$. *Foreground*: HI cube integrated profile contours showing column densities of 0.6, 7.1, 14.1, 28.2, 42.3, 56.5, 70.6, 80.4, 105.8, 127.0 and 141.1 $\times10^{20}$ cm$^{-2}$. The yellow circle at the bottom left represents the beam size of 28$\times$ 28.[]{data-label="mom0opt"}](cig96_mom0_overlay_optical){width="\hsize"}
![HI velocity field map of CIG96 and its companion after a 3.5$\sigma$ blanking (see Sect.\[sec:lownoise\]). The black lines indicate the orientation of the major and minor axis (PA$_{maj}$ = 20$\degree$ and PA$_{min}$ = 110$\degree$, respectively) along which the position-velocity cuts have been performed (see Fig.\[pvcuts\]). Grey contours represent the indicated velocities in kms$^{-1}$. The black circle at the bottom left represents the beam size of 28$\times$ 28.[]{data-label="mom1"}](cig96_mom1_nobar_vels_pvcuts){width="\hsize"}
![Position-velocity cuts along the major axis (*top panel*, PA=20$\degree$) and minor axis (*bottom panel*, PA=110$\degree$) of CIG96 HI cube. The column density is $N_{HI}\ (1\sigma) = 0.24\times10^{20}$ cm$^{-2}$ and the white and black contours correspond to 3.5$\sigma$ and 5$\sigma$, respectively. The cyan arrow points to the SW region where the velocity increases by approximately 30 $-$ 40 kms$^{-1}$ (see Sect.\[sec:HImoments\]). The red arrow points to the NW HI feature, the clumpy structure detected visible in channels 16 to 23 of the HI cube (see Sect.\[sec:chmap\]). As a reminder, the beam resolution is of 28$\times$ 28.[]{data-label="pvcuts"}](cig96_pvcuts){width="\hsize"}
CIG96 colour index image and optical features {#sec:pseudocolor}
---------------------------------------------
We analyse here the colour index image of CIG96 and the distribution along the pseudo-ring (further discussed in Sect.\[sec:feats96\]) via CAHA1.23m B and R images converted to SDSS$g$ and $r$ magnitudes, respectively (see Sect.\[sec:123data\]).
As a reference for the colour index values plotted in Fig.\[grpseudocuts\] left and central panels, we indicate the boundaries of the blue and red clouds from the SDSS $g-r$ optical colour$-$magnitude diagram. In particular we show the Green Valley interval of $\left(g - r\right)_{G.V.} = 0.60 - 0.75$ mag as defined by [@walker13] following the colour analysis by [@strateva01].
Qualitatively, we also note three striking features from the $g-r$ and optical images (see Fig.\[grpseudocuts\], left and central panels). The first feature is a diffuse arc in the east side of the pseudo-ring that almost closes it from north to south (Fig.\[grpseudocuts\], left panel, orange arrow); it is barely detectable (below $\sim$ $1.2\sigma$) in any individual image further than a diffuse emission due to the heavy contamination of nearby stars. The second structure is also barely detectable (below $1.2\sigma$) in any individual image despite there being no significant contamination by close stars in this region. It is located beyond the southern region of pseudo-ring, approximately 30kpc ($\sim$5) from the galaxy centre (Fig.\[grpseudocuts\], left and centre panels, cyan arrow). The third structure is indicated with yellow crosses in the central panel of Fig.\[grpseudocuts\]. This double structure has a SB of $\sim$26.0magarcsec$^{-2}$seems to connect the northern and southern inner parts of the galaxy with the western and eastern sides of the pseudo-ring, respectively.
Both in our VST image and in the DECaLS DR5 image, we detect a faint elongated (approximately $\sim$1 long) and diffuse structure to the NE of CIG96 (coordinates RA=$2^{h}15^{m}58.286^{s}$, DEC=+60439.15). It is located close to a bright star and barely a few kiloparsecs beyond the field of view covered by our $g-r$ image. As described further in Sect.\[sec:cirrus\], this structure lies on a region with a noticeable amount of background emission, mostly due to Galactic cirrus, and could therefore be part of it. However, we cannot rule out that this feature might be a tidal stream or a tidal disruption dwarf.
### Pseudo-ring colour index distribution {#sec:colordistribution}
We have studied the azimuthal variation of the colour index along the pseudo-ring by determining its median value in 33 circular non-overlapping apertures distributed in foreground star-free regions along its extent as shown in the central panel of Fig.\[grpseudocuts\], except for the NE region (PA in the range 38$\degree -$ 70$\degree$) due to the lack of reliable optical emission in this arc. We defined the apertures over a de-projected image of CIG96. For a better visualization, we have kept their spatial location and circular shape in the image presented in the previous figure, which is not de-projected. In order to discard any colour index changes in the pseudo-ring due to a gradient in the sky level, we determined the sky colour index of 62 regions set farther than the pseudo-ring, covering 360around CIG96 and free of bright stars. These apertures show $g-r$ values between $\sim$0.2 and $\sim$1.2. In Fig.\[pseudoapers\] we show the $g-r$ colour index distribution of all regions according to their PA and we find no colour index correlation between the apertures from the pseudo-ring and those from sky. However, as anticipated in the central panel of Fig.\[grpseudocuts\], we find a colour index change in between two PA ranges of the pseudo-ring. The 17 apertures of the SE arc (within PA=70$\degree -$258$\degree$) show a median colour of $g-r = 0.73$ mag ($st.dev. = 0.15$ mag), that is, a redder colour. Contrarily, the remaining 16 apertures of the NW arc (within PA=258$\degree -$38$\degree$) show a median value of $g-r = 0.31$ mag ($st.dev. = 0.11$ mag), that is, a bluer colour, making the difference between the two regions of $g-r \simeq$ 0.4 mag.
### Radial cuts {#sec:radcuts}
In order to compare the colour of the disc with the immediate pseudo-ring regions we computed radial profiles from individual $g$ and $r$ images. The right panel of Fig.\[grpseudocuts\] shows the de-projected $g$ image of CIG96 together with the lines along which those were calculated.
These profiles are shown in Fig.\[radprof\], where the bulge (the first 2.5 kpc, @espada11a), disc and pseudo-ring radii are marked as well. We selected the orientations due to the different structures crossed: disc, dust regions, arms, star-forming regions and thicker/thinner regions of the pseudo-ring. The profiles were then computed at PA of 6$\degree$, 16$\degree$, 30$\degree$ and 55$\degree$ and we will refer to them as PA6, PA16, PA30 and PA55, respectively.
To present the main results that these profiles yield, we have used a SB of 26.8magarcsec$^{-2}$ in the SDSS$r$ band. At this depth, the disc size varies in a range of $R_{disc}$=9.5$-$11kpc, depending on the PA.
The gap between the disc and pseudo-ring is not constant either: in the regions where the pseudo-ring and the disc are well resolved, the gap has an approximate width of $\simeq$1kpc. However, in regions where both the disc and pseudo-ring have a more diffuse emission, they prevent any reliable estimation of this separation.
The gap width, as well as its uncertainty, has a connection to the pseudo-ring dimensions: the more defined regions of the pseudo-ring have a width of $w_{pseudo-ring}\simeq$2kpc but it may rise up to $\sim$4kpc in some diffuse regions being hardly distinguishable from the disc.
Profile PA6 (green) shows red colours along the disc relative to the limit defined by the Green Valley strip. The peak at $\sim$7kpc corresponds to a foreground star ($m_{r\ SDSS}$=19.65mag). The pseudo-ring shows blue colours in most of its extent along this PA, matching the star-forming region ($\sim$0.7kpc size) present in this section of the cut, centred at a radius of approximately 12.5kpc. The colour difference of the disc and the pseudo-ring at this PA is $\sim$0.4mag.
Profile PA16 (pink) also shows the difference in colours between the disc and the pseudo-ring. The disc shows a stable red colour throughout its whole extension ($g-r \simeq$0.7mag). However, the pseudo-ring shows a colour gradient from $g-r \sim$0.7 to 0.1mag approximately, hence most of the pseudo-ring has blue colours. This profile was also aimed towards a large star forming region of $\sim$1.5kpc radius in the pseudo-ring and located at an approximate distance of 12kpc, so such a blue colour is expected. However, there is no apparent cause for the colour change.
Profile PA30 (yellow) shows a uniform disc colour within or right on the red edge of the Green Valley ($g-r \simeq$0.75mag) consistent with the rest of the profiles. There is an exceptionally red peak at 10.5kpc that, unlike in the case of PA6 (produced by a star), is the result of a region with large quantities of dust. The orientation of the previous profiles missed these dusty inner regions of the galaxy, easily visible in the left panel of Fig.\[grpseudocuts\], left. PA30 crosses the pseudo-ring through an area of diffuse emission and the redder colour is consistent throughout its extension. The orientation of this profile was chosen to obtain also the colour of the southern feature of $\sim$1kpc in width located at $\sim$18kpc indicated with a red vertical stripe (also marked with a cyan arrow in Fig.\[grpseudocuts\], left panel). Despite the fact that the feature is surrounded by the sky, its location and the surrounding 0.5kpc show a clearly blue colour. We have not considered this feature as part of the pseudo-ring so its width estimation remains between 1.5 and 3.5kpc and its radius $\sim$14kpc.
Profile PA55 (blue) shows a different behaviour along the disc. The mean colours are bluer than along the previous profiles; two regions that correspond to where the arms are crossed show very blue colours. This profile was selected to observe a much more diffuse and broad region of the pseudo-ring (width up to $\sim$3.5kpc). As in the disc, the pseudo-ring colour along this orientation is not homogeneous but it shows red colours ($g-r \sim$0.8mag) throughout most of its width. The farthest part of the pseudo-ring shows a steep change towards bluer colours, making it difficult to decipher whether it is an artifact of the sky or an existing structure with similar colour.
![Radial profiles obtained along four different orientations with PA of 6$\degree$, 16$\degree$, 30$\degree$ and 55$\degree$ (shown in Fig.\[grpseudocuts\], right panel) at the top, top-centre, bottom-centre and bottom panels, respectively. The horizontal green stripe represents the Green Valley in SDSS $g-r$ (see Sect.\[sec:pseudocolor\]. The bulge, disc and pseudo-ring limits are measured at 26.8magarcsec$^{-2}$. The bulge limit (2.5 kpc) and the disc limit are shown as the light blue dashed and black dotted lines, respectively. The pseudo-ring variable inner and outer limits are defined in each panel by the green dot-dashed lines. The light cyan band in the PA=30$\degree$ panel sets the location of the optical feature to the S marked as a cyan arrow in the left panel of Fig.\[grpseudocuts\] and a magenta ellipse in Fig.\[grHI\].[]{data-label="radprof"}](cig96_radprofiles){width="\hsize"}
Colour index and HI column density in the pseudo-ring {#sec:colorandHI}
-----------------------------------------------------
The black crosses of Fig.\[HIpseudomarks\] show the location of the apertures of the pseudo-ring on top of the HI 0$^{th}$ moment map. We find a remarkable spatial correlation between the optical pseudo-ring and the HI distribution, in agreement with [@espada11a]. In Fig.\[HIpseudomarks\] we indicate with a magenta ellipse the spatial location of the optical southern feature (shown in Fig.\[grpseudocuts\] left panel with a cyan arrow). It is too distant from the pseudo-ring ($\sim$4.1 kpc) as to confirm that both have a physical link and, unlike other star-forming regions of the pseudo-ring, we find no increase of the $N_{HI}$ in this region.
We have performed a detailed comparison between the pseudo-ring colour index $g-r$ and $N_{HI}$ for each selected aperture. With this aim, we have scaled each one of them by subtracting the mean value of the 33 apertures and dividing them by their sigma value (Fig.\[grHI\], top panel). We observe an anticorrelation between $g-r$ and $N_{HI}$ scaled values within PA=180$\degree -$ 40$\degree$, i.e. bluer colours correspond to larger column densities. It is only broken in the range PA=90$\degree -$ 180$\degree$, approximately, probably due to less reliable $g-r$ measurements in this side of the pseudo-ring, the most diffuse region. The anticorrelation is also confirmed in the bottom panel of Fig.\[grHI\] where we show $g-r$ as a function of $N_{HI}$: most of the bluer areas, mainly located in the NW side of the pseudo-ring (PA=260$\degree -$40$\degree$) show column densities of 8.5$-$13.5$\times$ 10$^{20}$ cm$^{-2}$, higher than most of the redder ones which show much lower levels instead. These results are discussed further in Sect.\[sec:feats96\].
![Central 53$\times$53 kpc (9$\times$9 arcmin) of the integrated HI emission map of CIG96. Column density is indicated with a colour gradient. Contours indicate 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 and 14 $\times$ 10$^{20}\ at\ cm^{-2}$. The black crosses indicate the central position of the 33 apertures used to measure the colours of the pseudo-ring (see Sect.\[sec:colordistribution\]). The magenta ellipse indicates the position of the southern feature indicated with a cyan arrow in Fig.\[grpseudocuts\], left panel, the $g-r$ colour index image. The yellow circle at the bottom left indicates the HI image synthesized beam of 28$\arcsec\times28\arcsec$.[]{data-label="HIpseudomarks"}](cig96_HI_pseudo_marks){width="\hsize"}
![*Top panel*: $g-r$ (brown circles) and $N_{HI}$ (green crosses) scaled (mean subtracted, sigma divided) values in each of the 33 apertures. A grey dashed line has been drawn at $g-r$ = 0 for reference. Redder colours and higher $N_{HI}$ are positive in this figure. *Bottom panel*: $g-r$ median colour index vs. $N_{HI}$ measured in the 33 apertures traced over the pseudo-ring. The green horizontal stripe represents the Green Valley in SDSS $g-r$.[]{data-label="grHI"}](cig96_HIvsColor){width="\hsize"}
Optical characteristics of the companion {#sec:companion}
----------------------------------------
We aimed to observe any possible optical structures connecting CIG96 and its companion (see Sect.\[sec:cigcomp\]). Given that the CAHA images have a field of view of 12$\arcmin \times$12$\arcmin$ (approximately 71kpc$\times$71kpc), it is only possible to studying any potential optical connection between the two galaxies with the VST 1$\degree\times$1$\degree$ image, as it provides continuous coverage across the $\sim$90kpc separation between them. Figure\[cig96companion\] shows a 10$\times$10 image centred between the two galaxies. At the current optical SB limit and in agreement with the HI map, we detect no sign of any stellar feature tracing any direct interaction between CIG96 and its companion.
Focusing on its companion, our VST image shows that it consists of an elongated structure oriented with a PA of 35plus a spheroidal component. Although the HI resolution of our data prevents us from separating both optical components, the HI emission is slightly elongated along the same PA, within the resolution of our HI data (28$\times$28, see Fig.\[mom0\]). Moreover, the HI kinematics shows the same orientation and is quite regular (within the limits given by the beam smearing), especially in the velocity range of 1560 to 1610kms$^{-1}$, where the main HI emission of the galaxy is found (see Fig.\[mom1\]).
In order to determine whether the optical image of NGC864COM01 is compatible with a disc$-$bulge system, we proceeded in two steps. We inspected the model and residuals provided by DECam Legacy Survey (DECaLS DR5, @schlegel15) and the model shows a good fit to the spheroidal component of the galaxy, with the residuals suggesting a blue and close to edge-on disc with a similar PA to the HI velocity field, as indicated above. Hence, from a morphological point of view it is compatible with a spiral galaxy with an Sa - Sb type.
As a next step we decided to perform a similar study to the analysis of the pseudo-ring of CIG96, aiming to determine the colour of each of the two components and whether they are consistent with values found for spheroidal components and discs in spiral galaxies. In particular, we used SDSS$g$ and $r$ images from the DECaLS survey whose exquisite seeing allowed to produce a $g-r$ colour index image and measure the colours in the sky (median value is 0.49$\pm$0.07mag) as well as in certain apertures, as shown in Fig.\[companionHIgr\], central and bottom panels. The measures in these apertures show that the elongated structure is bluer (0.33$-$0.45mag) than the spheroidal one, which is redder (0.55$-$0.70mag, right over the Green Valley range, see Fig.\[pseudoapers\]).
Lastly, we checked whether Sa-Sb galaxies exist with masses similar to that of the CIG 96 companion. We have used the value for the dynamical mass of NGC864COM01 as determined in Sect.\[sec:sblimit\] ($\sim$10$^{9}$M$_{\odot}$ which, as explained there, is a lower limit to the real one). According to [@nair10] there exist Sa and Sb galaxies within such a mass range. Hence all these characteristics lead us to favour the possibility of NGC 864 COM01 being an Sa or Sb galaxy.
![VST image in SDSS$r$ band of a 28$\arcmin\times28\arcmin$ region covering both CIG96 and its companion NGC864COM1 (indicated by the magenta contour to the east of the image). Optical contours are set in 24.0, 25.0, 26.0, 27.0, 27.5, 28.0 and 28.4magarcsec$^{-2}$ (SDSS$r$), smoothed with a Gaussian kernel of 11 pix radius. The purple contour marks the HI column density limit of the zeroth moment at 0.6$\times$10$^{20}\,cm^{-2}$, as in Fig.\[mom0opt\].[]{data-label="cig96companion"}](cig96_VST_comp28x28_11pix_sm_tags){width="\hsize"}
![*Top panel*: Velocity contours of the blanked HI datacube from 1560 to 1610kms$^{-1}$ (in steps of 10kms$^{-1}$) on top of the VST image of the companion of CIG96. The beam size of the HI data is indicated as a 28$\times$28grey disc in the bottom left corner of the image. The white frame represents the central 45$\times$45of the SDSS$g-r$ image of the bottom panel. *Central panel*: SDSS$g-r$ colour index image of the companion of CIG96 built from the corresponding SDSS$g$ and $r$ images from DECaLSDR5 survey. *Bottom panel*: Same image as in the central panel with the apertures used to measure the $g-r$ colour index. All apertures have a 1.75radius. The yellow-edge apertures trace the colours of the elongated feature while the black-edge apertures show the colours of the galaxy.[]{data-label="companionHIgr"}](cig96_companion_HI_gr2){width="\hsize"}
Cirrus {#sec:cirrus}
------
The possible large PSFs from close stars (whose radii reach several arcminutes, @trujillo16) and the necessity of a precise sky subtraction are key to discerning what may be artifacts from actual faint stellar components. However, there is another limiting factor at the current optical SB: the emission due to the presence of Galactic cold dust (@sandage76, confirmed with IRAS data by @neug84) known as cirrus in the Milky Way. This cirrus may cover large areas in the sky and, depending on the region, may start to be especially relevant from $\mu_{r\ SDSS} \simeq$ 25.5 $-$ 26.0magarcsec$^{-2}$ in the optical regime [@duc15].
As described in Sect.\[sec:planckwise\], we have used Planck857 and WISE3 images to identify the presence of cirrus in the field of CIG96. Their optical resolutions are 4.2$\arcmin$ and 6.5$\arcsec$, respectively, that is, approximately 242 and 6 times lower than the 1.04 arcsec pix$^{-1}$ resolution of our CAHA images or 1260 and 30 times lower than our VST image, respectively.
The Planck857 image is a good indicator of cirrus; however, we cannot extract reliable conclusions in the field of CIG96 due to its low spatial resolution. We used the central $40\arcmin\times40\arcmin$ of the VST image to inspect the cirrus (see Fig.\[cirrus\], top left panel). It shows that this area is populated with scattered emission visible from $\sim$28.5magarcsec$^{-2}$. While this image is not enough to conclude whether that emission is associated with cirrus or not, we can confirm it does not show up in any other observation considered in this work.
In the Planck857 image (Fig.\[cirrus\], bottom left panel), a $\sim$2$\sigma$ peak of emission (where $\sigma$ is the $rms$ measured on the clean SE region) is visible in the central pixels where the galaxy is located. In the surrounding area and close to the noise level, there are extended areas to the east, northwest and southwest of CIG96 which seem to match some of the emission observed with the VST image at 28.5magarcsec$^{-2}$.
The lack of cirrus structures brighter than 28.5magarcsec$^{-2}$ suggests they provide scarce (if any) contamination at brighter levels in our VST image, hence we set our detecion limit at 28.4magarcsec$^{-2}$ in SDSS$r$ band. From the opposite point of view, the low resolution of Planck857 makes it pointless to use such an image to find any cirrus structure in our optical image (Fig.\[cirrus\], bottom right panel). Should these exist, a positive matching between our VST and Planck 857 images would require extremely large and bright structures, easily detectable in both images; however, we do not find such large structures, preventing the use of the Planck857 image in this case.
WISE3 emission is shown in the top right panel of Fig.\[cirrus\] over the optical VST image. A quick glance at both Planck857 and WISE3 emission shows there is good correspondence between them, as expected. However, this is not the case for the optical$-$infrared images. Despite its higher resolution, the WISE3 image only matches some areas from the optical image, showing no significant emission nearby CIG96.
Following the same procedure as [@duc15], we aimed to trace the Galactic cirrus in the neighbourhood of CIG96 with our VST and the WISE3 images. After a careful revision of the reduction and calibration of the VST data, this image shows background emission from a SB level of 28.5magarcsec$^{-2}$ and fainter, that partially coincides with the IR emission from the WISE3 image (see Fig.\[cirrus\], top right panel). We cannot rule out an instrumental origin for some parts of the background emission of the field, yet the partial match between the VST and WISE3 images suggests that most of these structures are not undesired products of a deficient flat-field correction or scattered light but are actual Galactic cirrus, as expected in most latitudes from SB levels of 25magarcsec$^{-2}$ or higher [@guhathakurta89; @cortese10; @duc15]. Assuming that the emission belongs to the Galactic cirrus, reaching the very faint SB limit of 28.5magarcsec$^{-2}$ or fainter is crucial to assert their detection, implying that shallower images might be missing them. However, we cannot rule out an instrumental origin for the emission with the current data, and further optical and IR images of the same field with the same or fainter SB levels are desirable to confirm the nature of the background structures.
In summary, Planck857 and WISE3 images show no signs of large, diffuse and faint structures over CIG96 that might be interfering with our optical detection limits of the galaxy and its structures. However, we confirm a partial correspondence between the WISE3 image and the diffuse optical background emission in the surrounding field. The external (cirrus) or instrumental (flat-field correction) nature of the unmatched structures might be confirmed with further optical images of at least the same SB limit as our VST image. After a careful revision of the VST data processing, we must consider these background structures as a limiting factor to our images, setting the SB limit to 28.4magarcsec$^{-2}$.
Discussion of the optical and HI faint structures in the outskirts of CIG96 {#sec:discussion}
===========================================================================
In this section we discuss the implications of the above results with respect to the close environment of CIG96 as well as the origins of the different HI and optical features detected.
The environment of CIG96 {#sec:environment}
------------------------
![Representation in the isolation parameters of the subsample generated with spectroscopic data of CIG galaxies as calculated by [@argudo14] (q.v. Fig. 8 of that work). Lower values of local number density $\eta_k$ and tidal force estimation $Q$ represent higher isolation. The position of CIG96 is highlighted.[]{data-label="isol_argudo"}](cig96_isol_neighb_argudo){width="\hsize"}
---------------- ------- ------------ -------------------- ------
NGC864COM01 16.38 0.6 / 3.3 15.24 / 90 / E 1578
HIPASSJ0217+06 16.50 1.6 / 9.4 40.03 / 235 / NE 1549
UGC01677 18.00 0.9 / 5.3 76.70 / 450 / NW 1575
UGC01670 14.80 2.2 / 13.0 83.73 / 495 / NW 1593
UGC01803 14.70 2.8 / 16.5 89.26 / 527 / NE 1615
---------------- ------- ------------ -------------------- ------
The two parameters that quantify the isolation degree of a galaxy from CIG are the local number density, $\eta_{k}$, that accounts for the number of neighbour galaxies within a radius of 0.5 or 1 Mpc (depending on the criteria selected), and the tidal force estimation, $Q$, that quantifies how much the host is affected by its neighbourhood. According to the NASA/IPAC Extragalactic Database (NED), CIG96 has five identified galaxies within a radius of 1Mpc: NGC864COM01 (the aforementioned companion), HIPASSJ0217+06, UGC01677, UGC01670 and UGC01803. The properties of these neighbours are summarised in Table\[tablecomp\].
When taking the closest companion, NGC864COM01, the isolation time[^8] of CIG96 results in $\sim$2.7Gyr. The rest of the targets, located farther away, yield isolation times of $\sim$11 Gyr or higher than the age of the Universe, that is, they could not have interacted with CIG96 in the past.
We used the methods from two previous works by [@verley07b] (hereafter V07) and [@argudo13] (hereafter A13) to calculate the isolation parameters of CIG96 (see Sect.\[sec:intro\]). V07 consider the first $k-$th similar sized neighbours within a 500kpc radius. In the case of CIG96 only the closest two neighbours, i.e. NGC864COM01 and HIPASSJ0217+06 are selected and the isolation parameters are $\eta_{k}^{V07} = 0.877$ and $Q_{0.5}^{V07}\,=\,-1.981$. A13 extend the radius up to 1Mpc, that is, they consider the five neighbours mentioned above, and calculate the isolation parameters with photometric and spectroscopic data. The $Q$ parameter of CIG96 cannot be calculated with the photometric data because they contain no neighbour that violates the necessary criteria to measure the $Q$ between them and the host. However, the spectroscopic data allow to obtain the isolation parameters for CIG96 being $\eta_{k}^{A13}\,=\,0.68$ and $Q^{A13} =\,-\,3.41$. This method is also used by [@argudo14], where they focus on identifying the satellites around host galaxies from CIG, distinguish them from the Large Scale Structure (LSS) and quantify their effect over the host galaxy.
According to the spectroscopic data of A13, the position of CIG96 is highlighted in Fig.\[isol\_argudo\] where lower values of $\eta_k$ and $Q$ represent higher isolation. The remaining galaxies shown for comparison correspond to the subsample of galaxies used in [@argudo14] with 0, 1, 2 or 3 dynamically linked satellites (physically bound neighbours, q.v. Fig.5 in that work). The five closest companions of CIG96 within 1Mpc are responsible for raising the $\eta_k$ parameter, whereas the $Q$ value is similar to galaxies with one or two satellites. However, CIG96 has no satellite around it. This apparent contradiction reinforces a very relevant point: the isolation parameters are meant for statistical interpretations rather than to understand the environment of a particular galaxy. The physical interpretation of these parameters is that the populated environment of CIG96 does not contain massive galaxies and, among all the neighbours, the closest companion included, none are affecting the evolution of CIG96, as supported by all our observations. Despite the depth reached with our HI cube and optical images, we detect no signs of gaseous or stellar features between CIG96 and its gas$-$rich companion (see Sect.\[sec:cigcomp\] and \[sec:velasym\]) or any other more distant galaxies. The lack of detections supports the hypothesis of a long-lived isolation of CIG96 where its asymmetrical HI distribution (see Fig.\[mom0\]) may come from internal processes rather than an external agent with the exception of an absorbed small companion. HI asymmetries caused by major merger interactions are detectable within 0.4$-$0.7Gyr [@holwerda11c], that is, approximately within one or two full rotations of CIG96 ($T_{rot}^{CIG\,96}\simeq$ 0.4Gyr). However, the high isolation of the galaxy rules out encounters with major companions for at least the last $\sim$2.7Gyr.
The Galactic cirrus is observed from a SB level of $\mu_{r\ SDSS} =$28.5magarcsec$^{-2}$ and fainter, affecting a remarkable area of the field around CIG96. Infrared images may play a relevant role for the cirrus identification and future removal, therefore allowing for lower limits in SB to be reached. In this case, the two Planck857 and WISE3 images discussed do not manage to fully trace these structures at the sensitivity and spatial resolution required to remove them from the optical images. Since cirrus cannot be avoided by introducing any changes in the observational and data treatment strategy, they set an observational limit to our optical images, hampering the detection of fainter structures in polluted areas.
After reviewing the DECaLS DR5, a survey of similar characteristics with a SB limit of $\sim$28magarcsec$^{-2}$ ($\sim$0.7magarcsec$^{-2}$ brighter than our image from VST), we did not spot clear signs of the structures we detect in our VST image. While we cannot fully rule out an instrumental origin, most of the faint structures surrounding CIG96 only show up from $\sim$28.5magarcsec$^{-2}$ and fainter, suggesting that very faint SB levels are necessary to detect them, possibly being missed by slightly shallower images. We do confirm though the detection of the small stellar feature indicated in the left and central panels of Fig.\[grpseudocuts\] (cyan arrow), a candidate ultra-diffuse galaxy (UDG).
Origins of the features of CIG96 {#sec:feats96}
--------------------------------
*HI features*.
For decades it has been known that the HI in spiral galaxies frequently shows asymmetries and lopsidedness [@baldwin80; @richter94]. However, the origin of such features in the isolated galaxies from CIG is unknown. The asymmetries of CIG96 represent an excellent study case and have motivated all the observations and discussion of this work.
As described in Sect.\[sec:chmap\] and \[sec:HImoments\] and shown in Figs. \[figvelchannelwz1\], \[mom0\] and \[pvcuts\], we observe two external HI clouds, the NW and SE HI features, which do not seem to belong to the main HI body of CIG96 and a region with a remarkable receding velocity in the outermost SW region. The NW HI feature is a clumpy, almost continuous cloud spread throughout $\sim$70 kms$^{-1}$ and connected to the galaxy approximately along the direction of its minor axis. This connection may be traced down to channels 24 and 25 (1560 and 1570 kms$^{-1}$, respectively), where the HI cloud and the galaxy join showing no perturbation in the velocity field. The SE HI feature consists of a few gaseous clouds located in the same spatial location and spread along $\sim$40 kms$^{-1}$. However, the lack of any noticeable perturbation over the outermost HI of CIG96 suggest this feature has had little (if any) effect on the eastern side of the galaxy. It is also important to note that neither of these two features show any optical counterpart down to our detection limits.
The most distant HI region located at about $\sim$7$\arcmin$ ($\sim$41 kpc) from the centre along the direction of the major axis (indicated with a cyan arrow in Fig.\[pvcuts\]) shows a receding velocity that is $\sim$30 $-$ 40 kms$^{-1}$ above the expected velocity ($\sim$1430 kms$^{-1}$) of this region. [@espada05] proposed this region as a possible kinematically detached clump from the galaxy. The zeroth moment of our HI cube (Fig.\[mom0\]) and the P/V cuts over the major axis show such a kinematical detachment as well as small and scattered unresolved HI clouds. However, they do not provide evidence of a physical detachment in the SW region (indicated with the cyan arrow in Fig.\[pvcuts\], top panel), which is larger than any of the small HI clouds nearby. The inspection of the first-moment map suggests the farthest-south region might be warped but this would imply an external agent like a minor merger or the accretion of gas [@bournaud05; @jog09] that we do not find at our column density limit of $N_{HI}$=8.9$\times$10$^{18}$cm$^{-2}$. For this reason, we cannot rule out the cold gas accretion below the already low column density reached as a candidate to explain such features.
The P/V profile along the minor axis of CIG96 (Fig.\[pvcuts\]) shows emission in a wide gradient of velocities, going up to $+$85 kms$^{-1}$ in the receding side and down to $-$145 kms$^{-1}$ in the approaching side with respect to its central velocity. Beam effects may contribute to such dispersion [@bosma78] by introducing part of the surrounding emission at different velocities. However, with the current resolution of 28$\times$28, this effect can only explain dispersions up to approximately $\pm$60 $-$ 70 kms$^{-1}$. Counter-rotating gas clouds may also contribute to the high velocity dispersion but we do not detect any signatures of such features in the major axis, where they would likely be visible. An outflow or infall of extraplanar gas of different speeds may also explain the wide range of velocities but we do not detect any signs of either of these in any channel of the map.
*Accretion of cold HI clouds*.
High- and intermediate-velocity HI clouds around the Milky Way may reach masses of 10$^{7} M_{\odot}$. However, the clouds of the Local Group have smaller typical total masses of $M_{HI}^{cloud}\simeq $10$^{5}-$10$^{6} M_{\odot}$ [@wakker99] and they are expected to be several orders of magnitude below the total mass of their host galaxies.
With respect to CIG96, from our HI cube we compute a total HI mass of M$_{CIG\ 96}\eqsim$9.77$\times$10$^{9} M_{\odot}$. The two NW and SE HI features have total masses of $\sim3.1$ and 1.6$\times$10$^{6} M_{\odot}$, respectively, close to the detection limit achieved with our HI cube ($0.7 \times 10^{6} M_{\odot}$). The NW HI feature meets the HI mass, spatial distribution, and velocity range criteria to consider it a possible infalling cloud that overlaps with CIG96 in channels 23 $-$ 25 (velocities 1550 $-$ 1570 kms$^{-1}$).
As discussed by [@scott14], these clouds are not expected to fall on an extension of the rotation curve. This reinforces the idea that the previously discussed SW side of the galaxy is likely to be the warped edge of the HI disc instead of an accreted HI cloud.
*Pseudo-ring colour, column densities and minor mergers*.
HI is disrupted more easily than the faint optical halo substructures, which may live longer [@penarrubia05] than the $\sim$1.5$-$3 Gyr established by the quantified isolation criteria (see Sect.\[sec:environment\]). The external pseudo-ring of CIG96 is HI rich except for its southern and NE sides where the gas is scarce. The pseudo-ring colours in the southern side are clearly redder than in the rest of the ring, in particular in the bluer northern and NW sides where SF is taking place in a number of scattered regions according to their blue colours. This colour difference agrees with *GALEX* NUV and FUV results discussed by [@espada11a].
The star-forming regions with high UV emission match the bluer regions of the N-NW side of the pseudo-ring, and consistently show column densities above 8.5$\times$10$^{20}$ cm$^{-2}$. These are higher than the redder regions which coincide with areas where the HI seems disturbed and show much lower HI column densities, as shown in the bottom panel of Fig.\[grHI\]. The $N_{HI}$ is measured with a beam larger than the size of the optical regions. Once resolved, we would expect to achieve an even higher $N_{HI}$ level in the bluest regions, reaching the nominal SF value of $10^{21}$cm$^{-2}$. Also, the anticorrelation observed between the scaled $g-r$ and $N_{HI}$ seems to break in the 70$-$180arc. This is the eastern side of the pseudo-ring where the emission is very diffuse so any correlation or anticorrelation between $g-r$ and $N_{HI}$ in this area is uncertain.
Among the mechanisms that involve the SF enhancement or quenching as well as HI asymmetries are the minor mergers with low-mass dwarf galaxies. On the one hand, one or a few recent minor mergers with HI-rich small galaxies (wet mergers) are expected to leave clear HI footprints tracing such events, let alone orbiting stellar structures [@martinezdelgado09] or SF enhancement in the areas where the merging occurred: i.e., bluer colours would be expected in the southern and NE areas instead of the redder colours we find. Also, these mergers could have occurred in the more diffuse and incomplete regions of the pseudo-ring but, again, footprints of such events are missing. On the other hand, any minor mergers with one or a few HI-poor companions (dry mergers) might explain the stripped gas as well as the consequent SF quenching (even extinction) of the NE and southern regions. However, we find no optical or gaseous signature in any of these sections and the zeroth and first moments indicate no link between the possible warp in the southern arc of the pseudo-ring. One or various possible encounters with ultra$-$diffuse galaxies (UDG) of large mass$-$to$-$light ratios [@vandokkum16] might be responsible for the SF in the pseudo-ring. The only candidate to UDG lies to the south of the galaxy (see Fig.\[grpseudocuts\] and \[radprof\]) but we detect no signs of interaction between this feature and the galaxy at the SB level reached with our observations.
The tidal footprints from older minor mergers ($t$ > 0.7 Gyr, @holwerda11c) might have disappeared within one or two galaxy revolutions. Should they have left any optical counterparts that might still be visible, they would be expected to be weak and diffuse, such as those detected in the outskirts of other nearby galaxies (with no isolation classification) by [@morales18] at similar SB. Except for possible UDG mentioned before, we do not detect any other candidates or tidal streams brighter than our SB limit. However, the limiting SB level of $\mu_{r\ SDSS} =$ 28.5magarcsec$^{-2}$ prevents the detection of further potential candidates fainter than this limit, that is, based solely on these optical observations, we may not discard the hypothesis that one or more minor mergers may be responsible for the asymmetries of the galaxy and the colour index variation in the pseudo-ring, nor the possibility that faint optical counterparts to the HI features might exist. Consequently, old minor mergers remain as possible candidates to explain the stellar and gaseous features of CIG96.
*Age of the pseudo-ring of CIG96*
The blue colour of $g-r$ representations is a good tracer of the age of the stellar population in the optical regime. [@schawinski09] propose three different models to discuss age according to the $g-r$ colour observed: model 1 assumes an instantaneous burst of SF with an exponential decay of 100 Myr; model 2 considers an instantaneous burst of SF with instantaneous decay of 10 Myr; lastly, model 3 assumes a constant SFR.
The colour nature of the pseudo-ring of CIG96 does not fit with the age estimations of models 1 and 2 because so recent SF would either require nearby companions powerful enough to trigger it, or a transfer of gas from the inner parts towards the pseudo-ring in a lower period of time than the dynamical timescale of the galaxy. We do not detect signs of any of these requirements, leading us to consider model 3 as a more likely scenario. According to the latter, the continuous and slow ingestion of gas from the central parts of the galaxy into the pseudo-ring might explain a constant SF during at least 1 Gyr. Additionally, [@espada11a] found that the outer parts of CIG96 have ultraviolet colours of FUV$-$NUV = 0.1$-$0.2 mag$_{AB}$arcsec$^{-2}$. According to this interval FUV$-$NUV, the low-metallicity model proposed by [@smith10] fits with the expected lower metallicity in the outskirts of a galaxy. However, it yields ages lower than 100 Myr for the pseudo-ring. We discard the option of a companion since we would expect to detect it further than just its influence on the outskirts. *Origin of the pseudo-ring of CIG96*
The rings located 2$-$2.5 times the radius of the bar of the galaxy are labelled as outer rings (ORs) [@buta17]. Outer rings are typical in barred galaxies and their origin is gas accumulation in the outer Lindblad resonance (OLR) [@schommer76].
On the one hand, we find no signatures in either the optical or HI observations that suggest a collisional origin, in agreement with its high isolation level. On the other hand, the OLR is located at a distance slightly beyond twice the bar radius [@ath82]. In the case of CIG96, the bar and pseudo-ring optical radius are $\sim$2 kpc ($\sim$22$\arcsec$) and $\sim$14kpc ($\sim$145$\arcsec$), respectively, that is, the pseudo-ring is located over four times farther than the bar, making such a distance too large to be considered an OLR of the bar, as discussed by [@espada05].
An oval shape of the bright inner disc of the spiral galaxy might be a more reliable source to explain the origin of the pseudo-ring, as discussed briefly by [@espada11a]. These non-axisymmetric kinematical features can produce disturbances on the motions of gas clouds located in the outer HI disc, resulting in complete or partial resonance rings [@schwarz81; @verdes95; @buta96]. In agreement with what is expected to be found in these resonance rings, the pseudo-ring of CIG96 has a symmetric shape and it is partially defined, showing diffuse optical emission and low HI column densities in regions to the NE and SW.
An apparent misalignment between the galaxy centre and the pseudo-ring might be suggested by external isophotes of CIG96 (q.v. Figs. \[optvst\] and \[cig96companion\]). This is confirmed by the different elliptical fittings of the pseudo-ring that consistently show its centre lies $\sim$1.2kpc ($\sim$12$\arcsec$) to the south of the galaxy centre. However, the diffuse optical eastern side of the pseudo-ring and the high contamination of bright stars around the galaxy prevent a reliable global isophotal fitting analysis and further discussion on this topic.
The nearby stars also prevent us from clearly resolving the two elongated and faint arms or extensions that depart from the north and south of the disc. However, both the UV data analysed by [@espada11a] and our CAHA optical images (q.v., Figs. \[optvst\] and \[grpseudocuts\], central panel, yellow crosses) show that the northern and southern extensions connect the disc and the outermost regions of the galaxy where, respectively, they join the western and eastern sides of the pseudo-ring.
*Detection of fainter signatures of interactions*
Up to date, the studies performing deep optical observations of nearby galaxies mostly detect the brightest stellar features located in their outskirts. Standard $\Lambda$-CDM cosmological simulations of galactic halos [e.g. @bullock05; @johnston08; @cooper10] show that a large portion of the debris from old and minor mergers may be fainter than $\sim$30magarcsec$^{-2}$. For this reason, we can confirm that we do not detect signatures of these minor interactions in the outskirts of CIG96 down to the SB and $N_{HI}$ limits reached with our optical and HI observations, respectively, yet there may be unveiled features lying at fainter SB levels. Currently, for galaxies located close to the Milky Way, the only approach for detecting these very faint remnants of interactions consists of performing stellar density maps of evolved stellar components (e.g. RGB stars) in the halos of these galaxies (e.g. the PISCeS survey and study of Cen A and NGC253 @crnojevic16 [@crnojevic18]). These studies reach SB limits of 32$-$34magarcsec$^{-2}$, yet this technique is not feasible for more distant galaxies like CIG96 and farther with the current ground-based telescopes.
Conclusions {#sec:conclusions}
===========
The AMIGA project uses a sample that shows the most symmetric HI integrated profiles when compared to any other sample, even field galaxies. However, some of its members present very asymmetric profiles as well as other features whose origins remain unknown. If large asymmetries are mostly generated by interactions, the lopsidedness of an isolated galaxy such as CIG96 should not be observed. CIG96 is an isolated galaxy of the AMIGA sample that shows a 16% asymmetry in its HI profile as well as an actively star forming external pseudo-ring detected in the optical, UV and HI regimes. Our deep optical and HI observations have yielded unprecedented detail of the stellar and gas components of the galaxy and its outskirts. The wide field of view of 1$\times$1of VST telescope and the wavelet-filtered 21-cm data from VLA/EVLA telescope allowed us to reach a maximum SB and column density level of $\mu_{r\ SDSS}$ = 28.7magarcsec$^{-2}$ and $N_{HI}$ = 8.9 $\times$ 10$^{18}$ cm$^{-2}$ (5$\sigma$, beam size of 28$\times$28), respectively. The optical data reveal the detailed structure of the pseudo-ring as well as a gradient in its colour index. Moreover, the HI data show previously undetected features very close to the galaxy. Next, we present the main conclusions of this work:
Down to these limits, we do not find any gaseous or stellar connection between CIG96 and any galaxy in its close environment, including its closest, largest, and HI-rich companion NGC864COM01, located 15.2$\arcmin$ ($\sim$90 kpc) to the east (projected distance) that may be a close to edge-on Sa or Sb galaxy, as the optical and HI properties of the system suggest. Scattered Galactic cirrus shows up from 28.5magarcsec$^{-2}$ (SDSS$r$ band) in the surroundings of the galaxy and prevents any positive detection of further faint optical features beyond this depth.
We find two low-mass HI features ($\sim$10$^{6} M_{\odot}$) located to the NW and SE edges of the galaxy (the NW and SE HI features). The NW HI feature is visible along a number of immediate channels of the HI cube and depicts a low column density cloud ($N_{HI}^{NW} \simeq$ 6.5$\times$10$^{19}$ cm$^{-2}$) connected with CIG96 slightly to the N of its minor axis. We think that the SE HI feature, however, is a series of thin, small and spatially aligned clouds ($N_{HI}^{SE} \simeq$ 4.9$\times$10$^{19}$ cm$^{-2}$) that stand out in the zeroth moment. The individual channels of the HI cube show that the different clouds that compose this feature are not connected to one another and show no direct effect on the immediate gas of the disc edge despite its close proximity.
We find a colour index difference of $g-r \simeq$ 0.4 mag between two sides of the partially complete pseudo-ring (PA$_{redder}$ = 70$\degree- $258$\degree$ and PA$_{bluer}$ = 258$\degree -$38$\degree$) that cannot be assigned to any instrumental effect. No environmental cause (external gas accretion or minor merger) has been identified in our data as to explain such a change in the colour index. The outermost star-forming regions detected with NUV and FUV images from GALEX coherently match the blue regions of the pseudo-ring, which also show $N_{HI}$ values close to 10$^{21}$ cm$^{-2}$, the nominal SF value. The cause for the higher concentration of $N_{HI}$ in certain areas of the pseudo-ring is still to be found.
We have reviewed different SF models based on the FUV$-$NUV and $g-r$ colours to determine the approximate age of the pseudo-ring of CIG96. We may discard a short lived pseudo-ring ($\sim$100 Myr or younger) caused by a very recent encounter with either a similar-sized galaxy (the isolation criteria discard them) or a small galaxy (we would expect to see the merger remains); instead, they favour an older pseudo-ring ($\sim$1 Gyr).
Despite the fact that bars are usually relevant candidates to play a critical role in the secular evolution of the outskirts of a galaxy by leading the matter to concentrate in the OLR, such is not the case for CIG96. First, we do not find any significant matter concentration in the OLR of CIG96 and, second, the pseudo-ring is located at almost double the radius of the expected location of the OLR based on the bar size. For these reasons, we cannot consider a bar-driven accumulation of matter in the OLR as the pseudo-ring origin. Either an oval distortion or old, elongated arms expelled from the inner parts of the disc are more fitting explanations of the origin of the faint, distant (from the galaxy centre) and circular pseudo-ring of CIG96. Star-forming regions are expected in secular evolution and they may be triggered by external factors such as encounters with smaller and fainter infalling HI clouds. However, we do not find evidence of any external event that may explain the star-forming regions of the pseudo-ring of CIG96. Their origin remains unknown. The lack of any remarkable tidal features or other stellar components leads us to consider that, on the one hand, any major encounter with similar sized galaxies must have never occurred, as guaranteed by the isolation times and criteria (see Sect.\[sec:environment\]); on the other hand, any possible minor merger must have taken place before the last two revolutions of CIG96 ($t$>0.8Gyr, approximately), allowing the footprints of such encounters to disappear within such a time.
P.R.M. and all the coauthors thank the referee for the careful reading and valuable suggestions provided in the report, which have helped to improve this paper significantly. P.R.M. is funded by the project AMIGA4GAS project and the FPI Grant AYA2011-30491-C02-01 by the Ministry of Economy and Competitiveness of Spain. P.R.M., L.V.M., J.B.H., M.J., M.F.L., S.S.E., J.G.S. acknowledge support from the grant AYA2015-65973-C3-1-R (MINECO/FEDER, UE). P.R.M acknowdledges Víctor Terrón for his assistance during the calibration of the optical data. P.R.M. and the rest of the coauthors acknowledge Tom Jarrett for providing the WISE image used in this work; I.Trujillo for the discussion with respect to the deep optical observations and techniques; Monika Petr-Gotzens (ESO) for the help, follow$-$up and good advices provided with the VST observations and Enrique Pérez for his help with the dynamical mass estimations. L.V.M. acknowledges discussions with Alberto Fernández$-$Soto about the CAHA1.23m data. J.I.P. acknowledges financial support from the Spanish Ministerio de Economía y Competitividad under grant AYA2013-47742-C4-1-P, and from Junta de Andalucía Excellence Project PEX2011-FQM-7058. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. The National Radio Astronomy Observatory (NRAO) Karl G. Jansky Very Large Array (VLA) is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.. This research made use of Astropy, a community-developed core Python package for Astronomy [@astropy13; @astropy18] and of APLpy, an open-source plotting package for Python [@robi12]. We used the NASA/IPAC Extragalactic Database (NED), operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^1]: [http://amiga.iaa.es](http://amiga.iaa.es/)
[^2]: This catalog is referred to as K73 in SIMBAD and KIG in NED databases.
[^3]: Weighting computed as $w(i) = rms(i)$ $^{-2}$, where $rms(i)$ stands for the flux density rms of each cube in the same units.
[^4]: All tasks used to generate the described moment maps are part of the CASA Image Analysis toolkit.
[^5]: Based on observations collected at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucía (CSIC).
[^6]: Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme ID 093.B-0894 and 098.B-0775.
[^7]: *repipy* ([GitHub source](https://github.com/javierblasco/repipy)) reduction package by J. Blasco-Herrera, *LEMON* ([GitHub source](https://github.com/vterron/lemon)) calibration package by V. Terrón-Salas.
[^8]: As defined by [@verdes05] and in agreement with the isolation criteria, the isolation time estimates the minimum time a galaxy has been free of interactions with major companions. It is defined as follows: let $D$ be the diameter of CIG96; the time that a hypothetical companion of $d=4 \times D$ size needs to cover a distance of 20$\times d$ by travelling at a typical “field” velocity of 150 kms$^{-1}$ is 11.5 Gyr, that is, almost the age of the universe, implying no recent interaction. The closest companion NGC864COM01, travelling at a velocity of 33 kms$^{-1}$ with respect to CIG96, yields a minimum isolation time of 2.7Gyr while the other galaxies have never interacted with CIG96.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Roland Assaraf[^1], Benjamin Jourdain[^2], Tony Lelièvre[^3], Raphaël Roux[^4]'
bibliography:
- './computation\_sensitivities.bib'
title: Computation of sensitivities for the invariant measure of a parameter dependent diffusion
---
Introduction
============
We are interested in methods to compute the response of a Brownian dynamics to an infinitesimal change of a parameter $\lambda \in {\mathbb R}$. More precisely, we consider the dynamics in ${\mathbb R}^d$: $$\label{eq:EDS_lam}
\left\{
\begin{aligned}
{\mathrm d}X_t^\lambda&=F_\lambda(X_t^\lambda){\mathrm d}t+\sqrt2{\mathrm d}W_t,\\
X_0^\lambda&=X_0,
\end{aligned}
\right.$$ for $\lambda \in {\mathbb R}$ close to $0$, where $(W_t)_{t\geq0}$ is a standard $d$-dimensional Brownian motion independent of $X_0 \in {\mathbb R}^d$. Note that neither the initial condition $X_0$ nor the Brownian motion depend on $\lambda$. The family of vector fields $F_\lambda:{\mathbb R}^d\to{\mathbb R}^d$ is indexed by a real parameter $\lambda$. We assume that when $\lambda=0$, the vector field derives from some potential energy $V:{\mathbb R}^d\to{\mathbb R}$, namely $$F_0=-\nabla V,$$ where $\nabla$ denotes the gradient operator with respect to the space variables. For $\lambda$ close to zero, one can think of $(X_t^\lambda)_{t\geq0}$ as a physical system undergoing a potential energy $V$ to which one applies an external force $\lambda{\partial_\lambda^0}F_\lambda$. Here and in all the following, the notation ${\partial_\lambda^0}$ denotes the derivative with respect to $\lambda$ computed at $\lambda=0$.
Concerning the potential $V$, we assume that the following assumption holds.
The function $V$ satisfies the following assumptions:
(i) $V:{\mathbb R}^d\to{\mathbb R}$ is a $\mathcal C^2$ function such that $x \mapsto
\nabla^2 V(x)$ is locally Lipschitz.
(ii) $\displaystyle{\int_{{\mathbb R}^d} {e^{-V(x)}{\mathrm d}x}= 1}$ and $\displaystyle{\int_{{\mathbb R}^d} |\nabla V|^2(x) {e^{-V(x)}{\mathrm d}x}< \infty}$.
(iii) Pathwise existence and uniqueness hold for the process $(X_t^0)_{t\geq0}$.
Since $\nabla V$ is assumed to be locally Lipschitz, pathwise uniqueness is automatically ensured. Pathwise existence is ensured for instance as soon as there exists a finite constant $C$ such that for all $x \in {\mathbb R}^d,$ $\nabla V(x) \cdot x \le C (1 + |x|^2)$.
At $\lambda=0$, the dynamics is of the following gradient form $$\label{eq:EDS}
{\mathrm d}X_t^0=-\nabla V(X_t^0){\mathrm d}t+\sqrt2{\mathrm d}W_t.$$ Under the above assumptions, it can be checked that ${e^{-V(x)}{\mathrm d}x}$ is the unique invariant probability measure (see Lemma \[lem:erggen\] below) denoted in the following: $${\mathrm d}\pi_0 = {e^{-V(x)}{\mathrm d}x}.$$ [Then, from]{} classical results in ergodic theory, for any $f$ in ${\mathbb L}^1(\pi_0)$, almost surely, $$\label{eq:erg1}
\lim_{t \to \infty} \frac1t\int_0^tf(X^0_s){\mathrm d}s = \int_{{{\mathbb R}^d}}f {\mathrm d}\pi_0.$$
Let us now introduce the assumptions we need on the drift $(F_\lambda)_{\lambda\in{\mathbb R}}$.
There exists $\lambda_0 >0$ such that, for all $\lambda \in [0,\lambda_0]$,
(i) The function $F_\lambda-F_0$ is bounded by $C\lambda$ for some constant $C$ not depending on $x$. Moreover, as $\lambda
\to 0^+$, $\frac{F_\lambda(x)-F_0(x)}{\lambda}$ converges locally uniformly for $x\in{\mathbb R}^d$ to some limit ${\partial_\lambda^0}F_\lambda$. Note that ${\partial_\lambda^0}F_\lambda$ is bounded by $C$.
(ii) The function $x\mapsto F_\lambda(x)$ is locally Lipschitz. The function $x\mapsto{\partial_\lambda^0}F_\lambda(x)$ is differentiable on ${{{\mathbb R}^d}}$ and $\nabla \cdot{\partial_\lambda^0}F_\lambda$ is in ${\mathbb L}^2(\pi_0)$.
Under these assumptions, we will show (see Lemma \[lem:erggen\] below) that the dynamics is ergodic with respect to a probability measure $\pi_\lambda$: for any $f \in {\mathbb L}^1(\pi_\lambda)$, almost surely, $$\label{eq:erg2}
\lim_{t \to \infty} \frac1t\int_0^tf(X^\lambda_s){\mathrm d}s = \int_{{{\mathbb R}^d}}f {\mathrm d}\pi_\lambda.$$
The aim of this paper is to study the quantity: for a given observable $f$, $$\label{eq:but}
\lim_{\lambda\rightarrow0}
\frac{\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda-\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_0}\lambda
={\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right).$$ In particular, we will exhibit sufficient conditions for the existence of this derivative, derive various explicit formulae for this quantity, and discuss numerical techniques in order to approximate it.
The estimation of derivatives of the form is useful in various applications, in particular in molecular simulations (see for example the recent work [@warren-allen-12]): optimization procedure to fit a force field to some observations, study of phase transitions, estimate of forces in Variational Monte Carlo methods (see [@ACK-11]), or computation of transport coefficients. Transport coefficients are computed as the ratio of the magnitude of the response of the system submitted to a perturbation in its steady-state to the magnitude of the perturbation. These coefficients are related to macroscopic properties of the system through fluctuation dissipation theorems [@chandler-87; @evans-morriss-08]. Examples include the mobility or the thermal conductivity.
It is well-known that it is possible to approximate ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$ by considering a simulation at $\lambda=0$. For example, the celebrated Green-Kubo formula [@chandler-87; @evans-morriss-08] writes (see Theorem \[theo:GK\] in Section \[sect:GK\] for a proof in our specific context): $${\partial_\lambda^0}\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda
=\int_0^\infty{\mathbb E}_{\pi_0}\left[
f(X_0)\left(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot {\partial_\lambda^0}F_\lambda\right)(X_s^0)
\right]{\mathrm d}s$$ where the subscript $\pi_0$ indicates that the initial condition $X_0$ is distributed according to $\pi_0$. The derivative ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$ can thus be approximated by considering infinite-time integrals of auto-correlation functions for the stationary process at $\lambda=0$. This formula can be used to approximate ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$ numerically, which requires at least in some cases to be careful when choosing the truncation time in the integral, see for example [@chen-zhang-li-10]. Let us also mention another technique discussed in [@warren-allen-12], based on the use of Malliavin weights and the Bismut Elworthy Li formula (see [@bally-bavouzet-messaoud-07; @bismut-84; @elworthy-li-94]).
In this work, we are interested in so-called non-equilibrium molecular dynamics (NEMD) methods which consists in simulating two trajectories with $\lambda=0$ and $\lambda=\varepsilon$ small, and then considering the finite difference when $\varepsilon \to 0$ (see for example [@ciccotti-jacucci-75; @ciccotti-kapral-sergi-05]): $${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right) \simeq
\frac{\frac{1}{t} \int_0^t f(X^\varepsilon_s) \, ds - \frac{1}{t}
\int_0^t f(X^0_s) \, ds}{\varepsilon} \text{ when $\varepsilon \to
0$ and $t \to \infty$.}$$ Note that the consistency of this estimate is based on the ergodic properties and . To reduce the variance of the computation, it is natural to use the same driving Brownian motion for the two processes $(X^\varepsilon_s)_{s \ge
0}$ and $(X^0_s)_{s \ge 0}$ (see [@ciccotti-jacucci-75]) and we therefore end up with the natural following estimate: $$\label{eq:estim}
{\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right) \simeq
\frac1t\int_0^t{\partial_\lambda^0}(f(X_s^\lambda)){\mathrm d}s
\text{ when $t \to \infty$.}$$ As will be shown below (see Proposition \[proptt\]), it is easy to simulate ${\partial_\lambda^0}(f(X_t^\lambda))$ by using the formula $${\partial_\lambda^0}(f(X_t^\lambda))=T_t \cdot \nabla f(X^0_t)$$ where the so-called tangent vector $T_t \in {{{\mathbb R}^d}}$ is defined by $$T_t={\partial_\lambda^0}X^\lambda_t.$$ Indeed, the couple $(X^0_t, T_t)$ is a Markov process which satisfies the following extended version of the stochastic differential equation : $$\label{eq:EDS_extended}
\left\{
\begin{aligned}
{\mathrm d}X_t^0&=-\nabla V(X_t^0){\mathrm d}t+\sqrt2{\mathrm d}W_t \,,\\
{\mathrm d}T_t &= \left( {\partial_\lambda^0}F_\lambda(X_t^0)-\nabla^2V(X_t^0) T_t\right)
{\mathrm d}t\, ,
\end{aligned}
\right.$$ with initial conditions $X^0_0=X_0$ and $T_0=0$ (since, we recall, $X_0$ does not depend on $\lambda$). The estimate thus leads to a practical numerical method to evaluate the derivative ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right) $.
The main theoretical result of this paper consists in exhibiting sufficient conditions such that the following equalities hold true (see Theorem \[theo:interversion\]): $$\label{eq:CV_estim_1}
{\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)
=
\lim_{t\rightarrow\infty}\frac1t\int_0^t{\partial_\lambda^0}(f(X_s^\lambda)){\mathrm d}s
~~~\mbox{ a.s.}$$ and $$\label{eq:CV_estim_2}
{\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)=
\lim_{t\rightarrow\infty}{\mathbb E}\left[{\partial_\lambda^0}(f(X_t^\lambda))\right].$$ Therefore, two natural estimators of ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$ are $$\label{eq:estim_1}
\frac1t\int_0^t{\partial_\lambda^0}(f(X_s^\lambda)){\mathrm d}s$$ and $$\label{eq:estim_2}
{\mathbb E}\left[{\partial_\lambda^0}(f(X_t^\lambda))\right].$$ The second estimator is derived from the expected ergodic property on the time marginals: $\lim_{t \to \infty} {\mathbb E}(f(X^\lambda_t))=\int_{{{\mathbb R}^d}}f
{\mathrm d}\pi_\lambda$. For both estimators, Theorem \[theo:interversion\] can be seen as a rigorous justification of the interversion of the derivative ${\partial_\lambda^0}$ with the limit $\lim_{t \to
\infty}$ and an average in time for the first estimator and over the underlying probability space for the second one, since ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)={\partial_\lambda^0}\left(\lim_{t\rightarrow\infty}\frac1t\int_0^tf(X_s^\lambda){\mathrm d}s\right)$ and ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)={\partial_\lambda^0}\left(\lim_{t\to\infty}{\mathbb E}\left[f(X_t^\lambda)\right]\right)$. In addition, we also study the variance of the random variable ${\partial_\lambda^0}(f(X^0_t))=T_t\cdot \nabla f(X^0_t)$ which influences the statistical errors associated with the two estimators and .
The proof of and is based on two main ideas. First, the long-time limit (in law) of the couple $(X^0_t,T_t)$ is identified using a time-reversal argument (see Lemma \[lem:conv\_loi\]) in the spirit of the argument used in [@fontbona-jourdain-14] to study the long-time behavior of two interacting stochastic vortices. We are then able to identify the long-time limit of the estimators using the Green-Kubo formula which we prove in our setting in Section \[sect:GK\]. Second, the justification of the interversion of the derivative with the long-time limit and the integrals requires some integrability results, which are based on the study of the long-time behaviour of ${\mathbb E}\left[e^{-\int_0^t{\varphi}(Y_s^x){\mathrm d}s}\right]$ for ${\varphi}=\min
{\rm Spec}(\nabla^2 V)$, where $(Y^x_s)_{ s \ge 0}$ satisfies with $x$ as an initial condition: $$\label{eq:Ytx}
\left\{
\begin{aligned}
{\mathrm d}Y_t^x&= - \nabla V(Y_t^x) \, {\mathrm d}t + \sqrt{2} {\mathrm d}W_t,\\
Y_0^x&=x.\\
\end{aligned}
\right.$$ Let us emphasize that we prove all these results in a rather general setting: the state space is non compact (namely ${{{\mathbb R}^d}}$), the coefficients are only assumed to be locally Lipschitz and the potential $V$ is not necessarily strictly convex. The study of the long-time behaviour of the couple $(X^0_t,T_t)$ is very much related to the study of the long-time behaviour of the couple $(Y^x_t,DY^x_t)$, (see Lemma \[lemderci\]) which may be also useful to analyze other related numerical methods, see [@tailleur-kurchan-07].
The paper is organized as follows. In Section \[sect:poincare\], we give preliminary results on the stochastic differential equations and , in particular on their ergodic properties and the long-time behaviour of the associated Kolmogorov equations. In Section \[sect:vecteur\_tangent\], we then introduce the tangent vector $T_t$ and study its integrability. In Section \[sect:GK\], we derive and prove finite-time and infinite-time Green-Kubo formulae. We are then in position to prove the long-time convergence of the estimators and in Section \[sec:main\_result\]. Finally, the theoretical results are illustrated through various numerical experiments in Section \[sect:numeric\].
In all the following, we assume that Assumptions [**([[**Pot**]{}]{})**]{} and [**([[**Drift**]{}]{})**]{} hold, and we do not mention them explicitly in the statements of the mathematical results.
Preliminary results on and and the associated Kolmogorov equations {#sect:poincare}
====================================================================
In this section, we introduce partial differential equations related to the stochastic differential equations and , and study their long-time behaviors. These preliminary results will be crucial to analyze the numerical methods aimed at evaluating ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$ that we study.
Let us introduce a few notations. For any positive Borel measure $\mu$ on ${\mathbb R}^d$, we denote by ${\mathbb L}^2(\mu)$ the space of real valued measurable functions on ${\mathbb R}^d$ which are square integrable with respect to $\mu$. We denote by ${\mathbb L}_0^2({e^{-V(x)}{\mathrm d}x})$ the space of zero-mean square integrable functions: $$\label{eq:L20}
{\mathbb L}^2_0\left({e^{-V(x)}{\mathrm d}x}\right)
=\left\{f\in{\mathbb L}^2\left({e^{-V(x)}{\mathrm d}x}\right),~\int_{{{\mathbb R}^d}}f(x){e^{-V(x)}{\mathrm d}x}=0\right\},$$ and by ${\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$ the first order Sobolev space associated with the measure ${e^{-V(x)}{\mathrm d}x}$: $${\mathbb H}^1\left({e^{-V(x)}{\mathrm d}x}\right)
=\left\{f\in{\mathbb L}^2\left({e^{-V(x)}{\mathrm d}x}\right),~\nabla f\in{\mathbb L}^2\left({e^{-V(x)}{\mathrm d}x}\right)\right\},$$ where $\nabla f$ is to be understood in the distributional sense.
About the solution to and the regularity of the law of $X^\lambda_t$
---------------------------------------------------------------------
Let us first start by an existence and uniqueness result for the process $(X^\lambda_t)_{t \ge 0}$ solution to .
\[lem:EDS\_lam\] There exists a unique strong solution to .
The proof is rather standard. The existence of a weak solution to is obtained thanks to the Girsanov theorem and the existence assumption for the process $(X^0_t)_{t \ge
0}$ (see Assumption [**([[**Pot**]{}]{})**]{}-(iii)). Indeed, $$\begin{aligned}
{\mathrm d}X^0_t
&=-\nabla V(X^0_t) \, {\mathrm d}t + \sqrt{2} \, {\mathrm d}W_t\\
&=F_\lambda(X^0_t) \, {\mathrm d}t + \sqrt{2} \,{\mathrm d}\left( \int_0^t \frac{1}{\sqrt{2}}(F_0-F_\lambda)(X^0_s) \,
{\mathrm d}s + W_t \right).\end{aligned}$$ Indeed, under the probability ${\mathbb Q}$ such that, for all $t \ge 0$, ($({\mathcal F}_t)_{t \ge 0}$ being the natural filtration for $(W_t)_{t \ge 0}$), $$\frac{{\mathrm d}{\mathbb Q}}{{\mathrm d}{\mathbb P}}\Big|_{{\mathcal F}_t}=\exp\left(-\int_0^t
\frac{1}{\sqrt{2}}(F_0-F_\lambda)(X^0_s) {\mathrm d}W_s - \frac{1}{4} \int_0^t
|F_0-F_\lambda|^2(X^0_s) \, {\mathrm d}s\right)$$ the process $\widetilde{W}_t=\int_0^t \frac{1}{\sqrt{2}}(F_0-F_\lambda)(X^0_s) \,
{\mathrm d}s + W_t$ is a Brownian motion and therefore the triple $(X^0_t,\widetilde{W}_t,{\mathbb Q})$ is a weak solution to . Note that thanks to the Assumption [**([[**Drift**]{}]{})**]{}-(i), the Novikov conditions are satisfied which justifies the use of the Girsanov theorem [@karatzas-shreve-88 Section 3.5-D].
Moreover, it is standard to check that trajectorial uniqueness holds for the stochastic differential equation , since from Assumption [**([[**Drift**]{}]{})**]{}-(ii), $x \mapsto F_\lambda(x)$ is locally Lipschitz (see [@karatzas-shreve-88 Section 5.2-B, Theorem 2.5]). As a consequence, by the Yamada-Watanabe theorem (see for example [@karatzas-shreve-88 Section 5.3-D]), the stochastic differential equation admits a unique strong solution.
In the sequel, we will need some results about the Radon-Nikodym density of the distribution of the process with respect to the equilibrium measure. These properties are given in the two following Lemmas.
\[lem:p\] Whatever the choice of $X_0$, for each $\lambda\in[0,\lambda_0]$ and $t>0$, $X^\lambda_t$ admits a positive density with respect to the Lebesgue measure on ${\mathbb R}^d$.
Let us now consider $\lambda=0$ and $(Y_t^{x})_{t \ge 0}$ solution to . For all $t > 0$, the law of $Y_t^{x}$ admits a density $y \mapsto p(t,x,y)$ with respect to the Lebesgue measure which satisfies the reversibility property: $$\label{eq:rev}
\forall t > 0, e^{-V(x)}
p(t,x,y)=e^{-V(y)} p(t,y,x), \text{${\mathrm d}x \otimes {\mathrm d}y$-a.e.}.$$
A well-known corollary of is that if $X_0$ is distributed according to $\pi_0$, then the process $(X^0_t)_{t \ge 0}$ solution to is reversible: for any $t>0$, $$(X^0_s)_{s \in [0,t]} \text{ has the same law as } (X^0_{t-s})_{s \in [0,t]}.$$
Let $\psi:{\mathbb R}^d\to{\mathbb R}$ be a bounded measurable function. By the Girsanov theorem, $${\mathbb E}[\psi(X_t^\lambda)]
={\mathbb E}\left[
\psi(X_0+\sqrt2W_t)
e^{
\frac1{\sqrt2}\int_0^tF_\lambda(X_0+\sqrt2W_s){\mathrm d}W_s
-\frac14\int_0^t|F_\lambda(X_0+\sqrt2W_s)|^2{\mathrm d}s
}
\right].$$Here, the assumptions of the Girsanov theorem are satisfied. Indeed, according to [@rydberg-97 Theorem 2.1] these assumptions are satisfied if global-in-time existence and uniqueness in law hold for both Equation (see Lemma \[lem:EDS\_lam\]) and its driftless counterpart $${\mathrm d}Y_t=\sqrt2{\mathrm d}W_t,$$ which is a mere Brownian motion. The first assertion of Lemma \[lem:p\] is thus proved. For $\lambda=0$, we follow [@gihman-skorohod-72 page 91] to deduce that $$\begin{aligned}
{\mathbb E}[\psi(X_t^0)]
&={\mathbb E}\left[
\psi(X_0+\sqrt2W_t)
e^{
-\frac12V(X_0+\sqrt2W_t)
}
e^{ \frac12V(X_0)}
e^{
\frac14\int_0^t\left(2\Delta V-|\nabla V|^2\right)(X_0+\sqrt2W_s){\mathrm d}s
}
\right]. \label{eq:girsanov}\end{aligned}$$ Now, if one considers the Brownian bridge: $$\forall s \in [0,t], \, B_s^{x,y}=x+ \sqrt{2}W_s + \frac{s}{t} \left( y - x -
\sqrt{2}W_t\right)$$ one obtains by conditioning with respect to $W_t$: $${\mathbb E}(\psi(Y_t^{x}))= {\mathbb E}\left( \psi(x+\sqrt{2} W_t)
e^{- \frac12 V(x+\sqrt{2}W_t)} e^{\frac12 V(x)} g(x,x+\sqrt{2}
W_t) \right)$$ where $g(x,y)={\mathbb E}\left( e^{
\frac14\int_0^t\left(2\Delta V-|\nabla
V|^2\right)(B^{x,y}_s){\mathrm d}s
}\right)$. This shows that $Y_t^{x}$ admits a density with respect to the Lebesgue measure: $$\label{eq:pn}
p(t,x,y)=e^{- \frac12 V(y)} e^{\frac12 V(x)} g (x,y)
\frac{e^{-\frac{|x-y|^2}{4t}}}{(4 \pi t)^{d/2}}.$$ [F]{}rom the formula , it is straightforward to check that $$\label{eq:rev_approx}
e^{- V(x)} p(t,x,y) = e^{- V(y)} p(t,y,x)$$ by using the fact that $g(x,y)=g(y,x)$ which is a direct consequence of the fact that $(B^{x,y}_s)_{s \in [0,t]}$ has the same law as $(B^{y,x}_{t-s})_{s \in [0,t]}$. This concludes the proof of Lemma \[lem:p\].
Let us now state a few additional results on the dynamics when $\lambda=0$ and when $(X^0_t)_{t \geq 0}$ starts from a general random variable instead of a deterministic point.
\[lem:densite\_X\_t\] Let $X_0$ be distributed according to some probability measure $\mu_0$, and let $(X_t^0)_{t\geq0}$ evolve according to Equation . Denote by $\mu_t$ the distribution of the random variable $X_t^0$.
For all $t>0$, $\mu_t$ has a density $r(t,\cdot )$ with respect to $d\pi_0=e^{-V(x)} \, {\mathrm d}x$: $$\mu_t({\mathrm d}x) = r(t,x) e^{-V(x)} \, {\mathrm d}x.$$
Moreover, for $0 < s \le t$, for ${\mathrm d}x$-a.e. $x \in {\mathbb R}^d$, $$\label{eq:rtx}
r(t,x) = {\mathbb E}(r(s,Y_{t-s}^x))$$ where $Y_t^x$ satisfies . Equation holds for $s=0$ if $\mu_0$ has a density $r(0,\cdot)$ with respect to $d\pi_0=e^{-V(x)} \, {\mathrm d}x$.
If there exists $s \ge 0$ such that $\|r(s,\cdot)\|_{{\mathbb L}^\infty} <
\infty$, then, for all $t \ge s$, $$\label{eq:max_princ}
{{\rm ess\, inf}}_{x\in{{{\mathbb R}^d}}}r(s,x)\
\leq{{\rm ess\, inf}}_{x\in{{{\mathbb R}^d}}}r(t,x)
\leq{{\rm ess\, sup}}_{x\in{{{\mathbb R}^d}}}r(t,x)
\leq{{\rm ess\, sup}}_{x\in{{{\mathbb R}^d}}}r(s,x).$$ Finally, for any $q \in [1,\infty)$, if there exists $s \ge 0$ such that $r(s,\cdot) \in {\mathbb L}^q({e^{-V(x)}{\mathrm d}x})$, then for all $t \ge s$, $r(t,\cdot) \in {\mathbb L}^q({e^{-V(x)}{\mathrm d}x})$ and $$\label{eq:decroit_L2}
\forall t \ge s, \, \|r(t,\cdot)\|_{{\mathbb L}^q({e^{-V(x)}{\mathrm d}x})} \le \|r(s,\cdot)\|_{{\mathbb L}^q({e^{-V(x)}{\mathrm d}x})}.$$
Let $\psi:{\mathbb R}^d\to{\mathbb R}$ be a bounded measurable function. By conditioning with respect to $X_0$, and using the function $p(t,x,y)$ introduced in Lemma \[lem:p\] $$\begin{aligned}
{\mathbb E}(\psi(X^0_t))
&= \int \int \psi(y) p(t,x,y) \, {\mathrm d}y \, {{\mathrm d}}\mu_0(x) \label{eq:densite_X_t}\\
&= \int \psi(y) e^{- V(y)} e^{ V(y)} \int p(t,x,y) {{\mathrm d}}\mu_0(x) \,
{\mathrm d}y. \nonumber\end{aligned}$$ This shows that the law $\mu_t$ of $X^0_t$ is $r(t,y) e^{-V(y)} \, {\mathrm d}y$ with $$r(t,y)= e^{ V(y)} \int p(t,x,y) {{\mathrm d}}\mu_0(x).$$
Likewise, for any $s \in [0,t]$, by conditioning with respect to $X^0_s$, it is easy to check that ${\mathbb E}(\psi(X^0_t))=\int \psi(y) e^{- V(y)} e^{ V(y)} \int p(t-s,x,y)
{{\mathrm d}}\mu_s(x) \, {\mathrm d}y$. Now by taking $0 < s \le t$ and using the reversibility property , we get $$\begin{aligned}
{\mathbb E}(\psi(X^0_t))
&=\int \psi(y) e^{- V(y)} \int e^{ V(y)} p(t-s,x,y)
{{\mathrm d}}\mu_s(x) \, {\mathrm d}y\\
&=\int \psi(y) e^{- V(y)} \int e^{ V(x)} p(t-s,y,x)
r(s,x) {e^{-V(x)}{\mathrm d}x}\, {\mathrm d}y.\end{aligned}$$ Since the law of $X^0_t$ is $r(t,y) e^{-V(y)} \, {\mathrm d}y$, this shows that, $$\label{eq:rty}
{\mathrm d}y\text{-a.e.}, \, r(t,y)= \int p(t-s,y,x) r(s,x) \, {\mathrm d}x = {\mathbb E}(r(s,Y_{t-s}^y)).$$ This integral is well defined since $x \mapsto p(t-s,y,x)$ and $x
\mapsto r(s,x)$ are non negative measurable functions. This shows formula .
The maximum principle is then a direct consequence from this representation formula .
Finally, if $r(s,\cdot) \in {\mathbb L}^q({e^{-V(x)}{\mathrm d}x})$, then, for $t > s$, $r(t,\cdot) \in {\mathbb L}^q({e^{-V(x)}{\mathrm d}x})$ since (using the fact that $x \mapsto p(t-s,y,x)$ is a probability density function and the reversibility property ) $$\begin{aligned}
\int |r(t,y)|^q e^{-V(y)} \, {\mathrm d}y
&=
\int \left|\int p(t-s,y,x) r(s,x) \, {\mathrm d}x\right|^q e^{-V(y)} \, {\mathrm d}y\\
&\le
\int \int p(t-s,y,x) | r(s,x) |^q \, {\mathrm d}x \, e^{-V(y)} \, {\mathrm d}y\\
&=\int \int p(t-s,x,y) e^{-V(x)} | r(s,x) |^q \, {\mathrm d}x \, {\mathrm d}y\\
&=\int e^{-V(x)} | r(s,x) |^q \, {\mathrm d}x < \infty.\end{aligned}$$
In Appendix \[sec:hypdelta\], we discuss a stronger assumption on $V$ under which we are able to get more precise bounds on $p(t,x,y)$.
A Feynman-Kac formula and the Fokker-Planck equation
----------------------------------------------------
For two measurable functions $f:{\mathbb R}^d\to{\mathbb R}$ and ${\varphi}:{\mathbb R}^d\to{\mathbb R}$, with ${\varphi}$ locally integrable with respect to the Lebesgue measure, consider the Kolmogorov equation associated with the infinitesimal generator of the stochastic differential equation : $$\label{eq:EDP_phi}
\left\{
\begin{aligned}
\partial_tu(t,x)
&=\Delta u(t,x)
+F_\lambda(x)\cdot\nabla u(t,x)
-{\varphi}(x)u(t,x),\quad
t>0, \, x\in{{{\mathbb R}^d}},\\
u(0,x)&=f(x), \quad
x\in{{{\mathbb R}^d}}.
\end{aligned}
\right.$$ In all this section, $\lambda$ is a fixed parameter in the interval $[0,\lambda_0]$.
In the following, we will consider solutions to Equation in the following weak sense:
\[defi:solution\_EDP\] Let $u$ be a function in the space $${\mathbb L}^\infty\left([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})\right)
\cap{\mathbb L}^2\left([0,T],{\mathbb L}^2(|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x})\right)
\cap{\mathbb L}^2\left([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})\right)$$ for any $T>0$. For $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, we say that $u$ is a weak solution to if $u(0,\cdot)=f$ and for any function $v$ in ${\mathbb H}^1\left({e^{-V(x)}{\mathrm d}x}\right)\cap{\mathbb L}^2\left(|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x}\right)$, $$\label{eq:solution_EDP}
\begin{aligned}
\frac{{\mathrm d}}{{\mathrm d}t}\int_{{{\mathbb R}^d}}u(t,x)v(x){e^{-V(x)}{\mathrm d}x}&=-\int_{{{\mathbb R}^d}}\nabla u(t,x)\cdot\nabla v(x){e^{-V(x)}{\mathrm d}x}-\int_{{{\mathbb R}^d}}{\varphi}(x)u(t,x)v(x){e^{-V(x)}{\mathrm d}x}\\
&\quad + \int_{{{\mathbb R}^d}}(F_\lambda(x) -F_0(x))\cdot \nabla u(t,x) v(x){e^{-V(x)}{\mathrm d}x}\end{aligned}$$ in distributional sense.
Note that the last term in is well defined since for $\lambda \in [0,\lambda_0]$, from Assumption [**([[**Drift**]{}]{})**]{}-$(i)$, $\|F_\lambda-F_0\|_{{\mathbb L}^\infty({\mathbb R}^d)}\le C\lambda$. Moreover, note that the condition $u(0,\cdot)=f$ makes sense, since a function $u$ satisfying $$u\in{\mathbb L}^2([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}))
\mbox{ and }
\partial_tu\in{\mathbb L}^2([0,T],{\mathbb H}^{-1}({e^{-V(x)}{\mathrm d}x}))$$ actually lies in $\mathcal C([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}))$ (see for example [@temam-79 Lemma $1.2$ p. $261$]).
\[prop:EDP\_bien\_posee\] Assume $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ and that the function ${\varphi}$ is locally integrable with respect to the Lebesgue measure and bounded from below. Then, Equation admits a unique solution in the sense of Definition \[defi:solution\_EDP\]. Moreover, this solution is in ${\mathcal C}([0,+\infty),{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}))$.
In addition, if $ f \in
{\mathbb H}^1\left({e^{-V(x)}{\mathrm d}x}\right)\cap{\mathbb L}^2\left(|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x}\right)$, the solution $u$ is more regular: for any $T>0$, $$u \in
{\mathbb L}^\infty\left([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})
\cap{\mathbb L}^2(|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x})\right)
\cap{\mathbb H}^1\left([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})\right).$$
By Assumption [**([[**Drift**]{}]{})**]{}-$(i)$, there exists $C_0>0$ such that $\|F_\lambda-F_0\|_{{\mathbb L}^\infty({\mathbb R}^d)}\le C_0$. Let $C$ be a positive constant such that ${\varphi}+C-\frac{C_0^2}{2}$ is nonnegative. From [@temam-79 Lemma $1.2$ p. $261$], one can take $e^{-Ct}u(t,x)$ as a test function in and obtain the following estimate $$\begin{aligned}
&\frac{{\mathrm d}}{{\mathrm d}t}\left( \frac{e^{-Ct}}2\int_{{{\mathbb R}^d}}|u(t,x)|^2{e^{-V(x)}{\mathrm d}x}\right)\\
&= -e^{-Ct} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}- e^{-Ct}\int_{{{\mathbb R}^d}}|u(t,x)|^2({\varphi}(x)+C){e^{-V(x)}{\mathrm d}x}\\
&\quad + e^{-Ct} \int_{{{{\mathbb R}^d}}} (F_\lambda-F_0) \cdot \nabla u(t,x) u(t,x)
{e^{-V(x)}{\mathrm d}x}\\
&\le -e^{-Ct} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}- e^{-Ct}\int_{{{\mathbb R}^d}}|u(t,x)|^2({\varphi}(x)+C){e^{-V(x)}{\mathrm d}x}\\
&\quad + \frac{e^{-Ct}}{2} \int_{{{{\mathbb R}^d}}} |\nabla u(t,x)|^2
{e^{-V(x)}{\mathrm d}x}+ \frac{e^{-Ct}}{2} C_0^2 \int_{{{{\mathbb R}^d}}} |u(t,x)|^2
{e^{-V(x)}{\mathrm d}x}\\
&\le -\frac{e^{-Ct}}{2} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}- \int_{{{\mathbb R}^d}}e^{-Ct}|u(t,x)|^2\left({\varphi}(x)+C-\frac{C_0^2 }{2}\right){e^{-V(x)}{\mathrm d}x}.
\end{aligned}$$ Therefore, by integrating in time, one obtains the following estimate: $$\label{eq:estimation_a_priori}
\begin{aligned}
&\frac{e^{-Ct}}2\int_{{{\mathbb R}^d}}|u(t,x)|^2{e^{-V(x)}{\mathrm d}x}+\frac12\int_0^t\int_{{{\mathbb R}^d}}e^{-Cs}|\nabla u(s,x)|^2{e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s
\\
&+ \int_0^t\int_{{{\mathbb R}^d}}e^{-Cs}|u(s,x)|^2\left({\varphi}(x)+C-\frac{C_0^2
}{2}\right){e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s=\frac12\int_{{{\mathbb R}^d}}|f(x)|^2{e^{-V(x)}{\mathrm d}x}.
\end{aligned}$$ [From]{} this estimate, the uniqueness result follows from linearity by taking $f=0$ in . And thanks to this a priori estimate, existence can be proved by using a Galerkin method on a countable family of smooth functions dense in ${\mathbb H}^1({e^{-V(x)}{\mathrm d}x})\cap{\mathbb L}^2(|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x})$, which exists since the measure $|{\varphi}(x)|{e^{-V(x)}{\mathrm d}x}$ is finite on compact sets (see for example [@malliavin-95 chapter II, Theorem 3.5]). As explained above, the fact that $u \in {\mathcal
C}([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}))$ is then a standard result, see for example [@temam-79 Lemma $1.2$ p. $261$].
In order to obtain the additional regularity, let us take $\partial_t u(t,x)$ as a test function in : $$\begin{aligned}
\int_{{{\mathbb R}^d}}|\partial_t u(t,x)|^2{e^{-V(x)}{\mathrm d}x}&= -\frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}- \frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|u(t,x)|^2{\varphi}(x){e^{-V(x)}{\mathrm d}x}\\
&\quad + \int_{{{{\mathbb R}^d}}} (F_\lambda-F_0) \cdot \nabla
u(t,x) \partial_t u(t,x)
{e^{-V(x)}{\mathrm d}x}\\
&\le -\frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}- \frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|u(t,x)|^2{\varphi}(x){e^{-V(x)}{\mathrm d}x}\\
&\quad + C_0^2 \int_{{{{\mathbb R}^d}}} |\nabla u(t,x)|^2
{e^{-V(x)}{\mathrm d}x}+ \frac{1}{4} \int_{{{{\mathbb R}^d}}} |\partial_t u(t,x)|^2
{e^{-V(x)}{\mathrm d}x}.
\end{aligned}$$ Therefore, for a constant $C_1>0$ such that $\varphi + C_1$ is nonnegative, $$\begin{aligned}
&\frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}+ \frac{1}{2} \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|u(t,x)|^2 ({\varphi}(x) + C_1){e^{-V(x)}{\mathrm d}x}+ \frac{3}{4} \int_{{{\mathbb R}^d}}|\partial_t u(t,x)|^2{e^{-V(x)}{\mathrm d}x}\\
&\le C_0^2 \int_{{{{\mathbb R}^d}}} |\nabla u(t,x)|^2 + C_1 \int_{{{\mathbb R}^d}}u(t,x) \partial_t u(t,x) {e^{-V(x)}{\mathrm d}x}\\
& \le C_0^2 \int_{{{{\mathbb R}^d}}} |\nabla u(t,x)|^2 + C_1^2 \int_{{{\mathbb R}^d}}|u(t,x)|^2 {e^{-V(x)}{\mathrm d}x}+ \frac{1}{4} \int_{{{\mathbb R}^d}}|\partial_t u(t,x)|^2 {e^{-V(x)}{\mathrm d}x}.\end{aligned}$$ Using Grönwall’s Lemma, one obtains the estimate after integration in time: $$\begin{aligned}
& e^{-2C_0^2 t} \int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}+ e^{-2C_0^2 t} \int_{{{\mathbb R}^d}}|u(t,x)|^2 ({\varphi}(x) + C_1){e^{-V(x)}{\mathrm d}x}\\
&+ \int_0^t \int_{{{\mathbb R}^d}}e^{-2C_0^2 s} |\partial_t
u(s,x)|^2{e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s \le \int_{{{\mathbb R}^d}}|\nabla f(x)|^2{e^{-V(x)}{\mathrm d}x}\\
&+\int_{{{\mathbb R}^d}}|f(x)|^2 ({\varphi}(x) + C_1){e^{-V(x)}{\mathrm d}x}+ 2C_1^2 \int_0^t \int_{{{\mathbb R}^d}}e^{-2C_0^2 s}
|u(s,x)|^2 {e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s.\end{aligned}$$ The last term is bounded from above over finite time intervals by a constant times $\int_{{{\mathbb R}^d}}|f(x)|^2 {e^{-V(x)}{\mathrm d}x}$ thanks to . Again, this a priori estimate can be made rigorous through a Galerkin procedure, and yields the additional regularity stated in the Proposition (see for example [@quarteroni-valli-97 Proposition 11.1.1] for a similar reasoning).
\[prop:regularite\_EDP\_phi\] Let $u$ be a solution to the partial differential equation in the sense of Definition \[defi:solution\_EDP\], and assume that the initial condition $f$ is of class $\mathcal C^2$ with locally Lipschitz second order derivatives. If ${\varphi}$ is locally Lipschitz, then $u$, $\partial_tu$, $\nabla u$ and $\nabla^2u$ are continuous functions and $u$ is a classical solution to .
We use a bootstrap argument based on ${\mathbb L}^p_t{\mathbb L}^q_x$ regularity results for parabolic partial differential equations. In order to apply standard results which require $0$ as an initial condition, we consider $v=u-f$, which satisfies (in the weak sense, see Definition \[defi:solution\_EDP\]) the partial differential equation: $$\label{eq:EDP_v}
\left\{
\begin{aligned}
\partial_t v (t,x)
&=\Delta v (t,x)
+F_\lambda(x)\cdot\nabla v(t,x)
-{\varphi}(x)v(t,x) + g(x),\quad
t>0, \, x\in{{{\mathbb R}^d}},\\
v(0,x)&=0, \quad
x\in{{{\mathbb R}^d}},
\end{aligned}
\right.$$ where $$g(x) = \Delta f (x)
+ F_\lambda(x)\cdot\nabla f(x)
-{\varphi}(x)f(x)$$ is a locally Lipschitz function.
In this proof, we use the following notation $${\mathbb L}^p_t{\mathbb L}^q_x=\bigcap_{T>0}{\mathbb L}^p([0,T],{\mathbb L}^q({{{\mathbb R}^d}})),$$ where ${\mathbb L}^q({{{\mathbb R}^d}})$ is the ${\mathbb L}^q$ space associated with the Lebesgue measure. We will also use the notations ${\mathbb L}_t^p{\mathbb{W}}_x^{s,p}$ where ${\mathbb{W}}$ stands for the usual Sobolev space. We moreover introduce the notation $${\mathbb L}^{\infty-}=\bigcap_{2\leq q<\infty}{\mathbb L}^q.$$ Last, we set $${\mathbb{W}}^{1,p}_t{\mathbb L}_x^q
=\{
u\in{\mathbb L}^p_t{\mathbb L}_x^q,~
\partial_tu\in{\mathbb L}^p_t{\mathbb L}_x^q
\}.$$
Let $\chi$ be some function in the space $\mathcal C^\infty_0$ of smooth, compactly supported functions on ${{{\mathbb R}^d}}$. The function $\chi v$ satisfies, in the weak sense, the equation $$\left\{
\begin{aligned}
\partial_t(\chi v)
-\Delta(\chi v)
&=\Phi^\chi\mbox{ on }(0,+\infty)\times{\mathbb R}^d,\\
(\chi v)(0,x) &=0, \quad x\in{{{\mathbb R}^d}},
\end{aligned}
\right.$$ where $$\Phi^\chi= (\chi F_\lambda
-2\nabla\chi)\cdot\nabla v
-(\Delta \chi+\chi{\varphi})v + \chi g.$$ [From]{} parabolic regularity results, see for example [@von-wahl-82 Theorem III.1], one has the implication: $$\label{eq:implic1}
(\Phi^\chi\in{\mathbb L}^2_t{\mathbb L}^p_x)
\Rightarrow
(\chi v\in{\mathbb L}^2_t{\mathbb{W}}^{2,p}_x\cap{\mathbb{W}}^{1,2}_t{\mathbb L}^p_x).$$ In addition, [from]{} the definition of $\Phi^\chi$, one has $$\label{eq:implic2}
(\forall\chi\in\mathcal C^\infty_0,~\chi v\in{\mathbb L}^2_t{\mathbb{W}}^{2,p}_x)
\Rightarrow
(\forall\chi\in\mathcal C^\infty_0,~\Phi^\chi\in{\mathbb L}^2_t{\mathbb{W}}^{1,p}_x)$$ since ${\varphi}$ and $F_\lambda$ are locally Lipschitz functions (see Assumption [**([[**Drift**]{}]{})**]{}-(ii)).
Now, by Definition \[defi:solution\_EDP\], the function $v$ lies in ${\mathbb L}^\infty([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}))
\cap{\mathbb L}^2([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}))$, so that $\Phi^\chi$ lies in ${\mathbb L}^2_t{\mathbb L}^2_x$, for any function $\chi\in\mathcal C^\infty_0$. [First]{} assume $d>1$, and let $\bar p$ be the supremum of all those $p$ such that $\Phi^\chi$ lies in ${\mathbb L}^2_t{\mathbb L}^p_x$ for all $\chi$. Assume $\bar p<\infty$. Since $$\frac{d\bar p}{d+\bar p}<\min(\bar p,d),$$ one can find some $p\in\left(\frac{d\bar p}{d+\bar p},d\right)$, such that $\Phi^\chi$ belongs to ${\mathbb L}^2_t{\mathbb L}^p_x$ for any $\chi$ (note that $\frac{d\bar p}{d+\bar p}\geq1$, since $\frac1d+\frac1{\bar p}\leq\frac12+\frac12=1$). [From]{} $\chi v$ lies in ${\mathbb L}^2_t{\mathbb{W}}^{2,p}_x$, and hence from , $\Phi^\chi$ lies in ${\mathbb L}^2_t{\mathbb{W}}^{1,p}_x$. However, Sobolev embeddings yield $$\Phi^\chi\in{\mathbb L}^2_t{\mathbb{W}}^{1,p}_x\subset{\mathbb L}^2_t{\mathbb L}^{\frac{dp}{d-p}}_x,\mbox{ where
}\frac{dp}{d-p}>\bar p\mbox{ since }p>\frac{d\bar p}{d+\bar p},$$ which contradicts the definition of $\bar p$. As a conclusion, $\Phi^\chi$ lies in ${\mathbb L}^2_t{\mathbb L}^{\infty-}_x$ for any $\chi$. In the case $d=1$, one can directly deduce from $\Phi^\chi\in{\mathbb L}^2_t{\mathbb L}^2_x$ that $\chi v$ is in ${\mathbb L}^2_t{\mathbb{W}}^{2,2}_x\subset{\mathbb L}^2_t{\mathbb{W}}^{1,\infty-}_x$ for any $\chi$, and thus $\Phi^\chi\in{\mathbb L}^2_t{\mathbb L}^{\infty-}_x$ for any $\chi$. In any case, $\chi v$ lies in ${\mathbb L}^2_t{\mathbb{W}}^{2,\infty-}_x$ for any $\chi$.
Now consider the equation satisfied by $\chi\partial_i v$, for any coordinate $i$. One obtains $$\label{eq:chaleur_chi_di_u}
\left\{
\begin{aligned}
\partial_t(\chi\partial_i v)
-\Delta(\chi\partial_i v)
&=\Psi^\chi\mbox{ on }(0,+\infty)\times{\mathbb R}^d,\\
(\chi\partial_i v) (0,x) &= 0, \quad
x\in{{{\mathbb R}^d}},
\end{aligned}
\right.$$ where $$\begin{aligned}
\Psi^\chi
&=(\chi F_\lambda - 2\nabla\chi
)\cdot\nabla(\partial_iv)+\left(\chi \partial_i F_\lambda
-(\Delta\chi+\chi{\varphi})e_i\right)\cdot\nabla v-(\chi\partial_i{\varphi})v + \chi \partial_i g .\nonumber\end{aligned}$$ Since $\chi v$ lies in ${\mathbb L}^2_t{\mathbb{W}}^{2,\infty-}_x$ for any $\chi\in\mathcal C^\infty_0$, the function $\Psi^\chi$ is in ${\mathbb L}^2_t{\mathbb L}^{\infty-}_x$, from the boundedness of $\chi\partial_i{\varphi}$, $\chi\partial_i g$ and $\chi\partial_i F_\lambda$. Then, parabolic regularity for the heat equation implies $$\forall i,~\forall\chi\in\mathcal C^\infty_0,
~\chi\partial_iv\in{\mathbb L}^2_t{\mathbb{W}}^{2,\infty-}_x\cap{\mathbb{W}}^{1,2}_t{\mathbb L}^{\infty-}_x.$$ In particular, for any $\chi\in\mathcal C^\infty_0$, $\chi v$ is in ${\mathbb{W}}^{1,2}_t{\mathbb{W}}_x^{1,\infty-}$. [From]{} Sobolev embeddings, we deduce that $\chi v$ lies in $\mathcal C^{1/2}_t\mathcal C^{1-{\varepsilon}}_x$, for any ${\varepsilon}$ in $(0,1)$ ($\mathcal C^s$ stand for Hölder spaces). [From]{} the Hölder regularity of the initial condition, Hölder regularity theory for the heat equation now yields the desired regularity on $v$ (and thus on $u$), see for example [@krylov-96 Theorem 10.3.3].
\[prop:feynman-kac\] Assume $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ and that ${\varphi}$ is locally Lipschitz and bounded from below. Then the solution $u(t,x)$ to the partial differential equation given by Proposition \[prop:EDP\_bien\_posee\] admits the following probabilistic representation formula: for all $t \ge 0$, $$\label{eq:feynman-kac}
{\mathrm d}x\text{-a.e.}, \, u(t,x)
={\mathbb E}\left[f(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right]$$ where $(Y_t^{\lambda,x})_{t \ge 0}$ is defined by (notice that $(Y_t^{0,x})_{t \ge 0}=(Y_t^{x})_{t \ge 0}$ is defined by ) $$\label{eq:Ytlambdax}
\left\{
\begin{aligned}
{\mathrm d}Y_t^{\lambda,x}&= F_\lambda(Y_t^{\lambda,x}) \, {\mathrm d}t + \sqrt{2} {\mathrm d}W_t,\\
Y_0^{\lambda,x}&=x.\\
\end{aligned}
\right.$$
: Let us first prove a maximum principle for solutions to Equation in the sense of Definition \[defi:solution\_EDP\]. Assume that the initial condition $f$ of Equation is bounded from above by some nonnegative constant $M$. Let $C$ be a constant such that ${\varphi}+C$ is nonnegative. [From]{} [@temam-79 Lemma $1.2$ p. $261$], and [@gilbarg-trudinger-01 Lemma 7.6], one can take $e^{-Ct}(e^{-Ct}u(t,x)-M)^+$ as the test function in the weak formulation , and obtain (using the fact that from Assumption [**([[**Drift**]{}]{})**]{}-$(i)$, $\|F_\lambda-F_0\|_{{\mathbb L}^\infty({\mathbb R}^d)}\le C_0$ for some $C_0>0$) $$\begin{aligned}
&\frac12 \frac{{\mathrm d}}{{\mathrm d}t} \int_{{{\mathbb R}^d}}|(e^{-Ct}u(t,x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\\
&=-\int_{{{\mathbb R}^d}}|\nabla (e^{-Ct}u(t,x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\\
&\quad-\int_{{{\mathbb R}^d}}({\varphi}(x)+C)e^{-Ct}u(t,x)(e^{-Ct}u(t,x)-M)^+{e^{-V(x)}{\mathrm d}x}\\
&\quad+\int_{{{\mathbb R}^d}}(F_\lambda(x)-F_0(x))\cdot \nabla (e^{-Ct} u(t,x)
-M)^+ (e^{-Ct}u(t,x)-M)^+ {e^{-V(x)}{\mathrm d}x}\\
&\leq -\frac12 \int_{{{\mathbb R}^d}}|\nabla (e^{-Ct}u(t,x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\\
&\quad-\int_{{{\mathbb R}^d}}({\varphi}(x)+C)e^{-Ct}u(t,x)(e^{-Ct}u(t,x)-M)^+{e^{-V(x)}{\mathrm d}x}\\
&\quad+\frac{C_0^2}{2}\int_{{{\mathbb R}^d}}|(e^{-Ct}u(t,x)-M)^+|^2 {e^{-V(x)}{\mathrm d}x}.\end{aligned}$$ By using Grönwall’s Lemma, one therefore obtain after integration in time: $$\begin{aligned}
&\frac{e^{-C_0^2t}}{2}\int_{{{\mathbb R}^d}}|(e^{-Ct}u(t,x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\\
&\le -\frac12\int_0^t\int_{{{\mathbb R}^d}}e^{-C_0^2s}|\nabla
(e^{-Cs}u(s,x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s\\
&\quad-\int_0^t\int_{{{\mathbb R}^d}}e^{-C_0^2s} ({\varphi}(x)+C)e^{-Cs}u(s,x)(e^{-Cs}u(s,x)-M)^+{e^{-V(x)}{\mathrm d}x}\,
{\mathrm d}s\\
&\quad+\frac12 \int_{{{\mathbb R}^d}}|(f(x)-M)^+|^2{e^{-V(x)}{\mathrm d}x}\leq0\end{aligned}$$ so that the function $u(t,\cdot)$ is bounded from above by $Me^{Ct}$, for any positive $t$. By a similar argument, if $f$ is bounded from below by $-M$, with $M$ nonnegative, then $u(t,\cdot)$ is bounded from below by $-e^{-Ct}M$ for any positive $t$.
: Let us now prove the Feynman-Kac formula under the assumption $f \in
\mathcal C^\infty \cap {\mathbb L}^\infty({\mathbb R}^d)$. Let $x\in{{{\mathbb R}^d}}$, $t>0$ and $M>0$. Let $\tau_M$ be the first exit time from $B(x,M)$ (namely the ball centered at $x$ and of radius $M$) for the process $(Y_s^{\lambda,x})_{s\geq0}$. Since $s \mapsto Y_s^{\lambda,x}$ is continuous, $\tau_M$ goes to $\infty$ as $M$ goes to $\infty$. Let us consider the solution $(t,x) \mapsto u(t,x)$ to , which is ${\mathcal
C}^1$ with respect to $t$ and ${\mathcal
C}^2$ with respect to $x$ thanks to Proposition \[prop:regularite\_EDP\_phi\]. Applying Itō’s formula to $u(t-s,Y_s^{\lambda,x})$ in the time interval $[0,t\wedge\tau_M]$, one obtains $$u((t-\tau_M)^+,Y_{t\wedge\tau_M}^{\lambda,x})e^{-\int_0^{t\wedge\tau_M}{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}
=u(t,x)
+\sqrt2\int_0^{t\wedge\tau_M}\nabla u(t-s,Y_s^{\lambda,x})e^{-\int_0^s{\varphi}(Y_u^{\lambda,x}){\mathrm d}u}{\mathrm d}W_s.$$ On the interval $[0,t\wedge\tau_M]$, the integrand in the stochastic integral remains bounded, so that this integral has zero mean. Taking the expectation, one obtains $$u(t,x)
={\mathbb E}\left[u((t-\tau_M)^+,Y_{t\wedge\tau_M}^{\lambda,x})e^{-\int_0^{t\wedge\tau_M}{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right].$$ By the above maximum principle the function $u$ is bounded on $[0,t]\times{{{\mathbb R}^d}}$. With the lower bound on ${\varphi}$, the dominated convergence theorem yields, letting $M\to\infty$, $$u(t,x)
={\mathbb E}\left[f(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right].$$
: Let us now assume that $f$ is in ${\mathbb L}^\infty({\mathbb R}^d)$. Let $f_n$ be a sequence of $\mathcal C^\infty$ functions such that $\sup_{n \ge 1} \|f_n\|_{{\mathbb L}^\infty({\mathbb R}^d)} \le \|f\|_{{\mathbb L}^\infty({\mathbb R}^d)}$, and converging to $f$ almost everywhere. In particular, $f_n$ converges to $f$ in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ by Lebesgue’s theorem. Therefore, the solution $u_n$ to Equation starting from $f_n$ is such that $u_n(t,\cdot)$ converges to $u(t,\cdot)$ as $n\to\infty$ in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, from the a priori estimate . Moreover, one has $$\forall x\in{{{\mathbb R}^d}},~u_n(t,x)={\mathbb E}\left[f_n(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right]
={\mathbb E}[f_n(Y_t^{\lambda,x})\Gamma(Y_t^{\lambda,x})],$$ where $\Gamma$ is a bounded function satisfying $\Gamma(Y_t^{\lambda,x})={\mathbb E}\left[e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\Big|Y_t^{\lambda,x}\right]$. For the remaining of the proof, we assume that $t >0$ (the formula clearly holds for $t=0$). [F]{}rom Lemma \[lem:p\], the distribution of $Y_t^{\lambda,x}$ admits a density $p^\lambda(t,x,y)$ with respect to the Lebesgue measure. Therefore, by the Lebesgue theorem, ${\mathbb E}[f_n(Y_t^{\lambda,y})\Gamma(Y_t^{\lambda,y})]$ converges to ${\mathbb E}[f(Y_t^{\lambda,y})\Gamma(Y_t^{\lambda,y})]$ as $n\to\infty$. This shows the equality $u(t,x)={\mathbb E}\left[f(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right]$ for ${\mathrm d}x$-a.e. $x$.
: Let us now assume that $f$ is in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ and let us write $f=f^+-f^-$ where $f^+=\max(f,0)$ and $f^-= \max(-f,0)$. For $n \in
\mathbb{N}$, the functions $f^+_n=\min(f^+,n)$ (resp. $f^-_n=\min(f^-,n)$) are in ${\mathbb L}^\infty({\mathbb R}^d)$ and converge in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to $f^+$ (resp. $f^-$). Let us consider the solution $u^+_n$ (resp. $u ^-_n$) to Equation starting from $f^+_n$ (resp. $f^-_n$). Since $f^+_n -f^-_n$ converges as $n\to\infty$ to $f$ in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, for every $t \ge 0$, $u^+_n(t,\cdot) - u^-_n(t,\cdot)$ converges in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to $u(t,\cdot)$, where $u$ is the solution to Equation starting from $f$. Moreover, one has $${\mathrm d}x \text{-a.e.},~u^\pm_n(t,x)={\mathbb E}\left[f^\pm_n(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right]
={\mathbb E}[f^\pm_n(Y_t^{\lambda,x})\Gamma(Y_t^{\lambda,x})],$$ where $\Gamma$ is the bounded function defined above. By the monotone convergence theorem, ${\mathbb E}[f^\pm_n(Y_t^{\lambda,x})\Gamma(Y_t^{\lambda,x})]$ converges to ${\mathbb E}[f^\pm(Y_t^{\lambda,x})\Gamma(Y_t^{\lambda,x})]$. This shows the equality $u(t,x)={\mathbb E}\left[f(Y_t^{\lambda,x})e^{-\int_0^t{\varphi}(Y_s^{\lambda,x}){\mathrm d}s}\right]$ for ${\mathrm d}x$-a.e. $x$.
As a corollary of the previous result, we obtain that the law of $X^0_t$ satisfies a partial differential equation (the Fokker-Planck equation).
\[cor:FP\] Let $X_0$ be distributed according to some probability measure $\mu_0$, and let $(X_t^0)_{t\geq0}$ evolve according to Equation . Let us assume that $\mu_0$ has a density $r_0$ with respect to $d\pi_0=e^{-V(x)} \, {\mathrm d}x$ such that $r_0 \in
{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. Denote by $\mu_t$ the distribution of the random variable $X_t^0$, and by $x \mapsto r(t,x)$ the density of $\mu_t$ with respect to $d\pi_0=e^{-V(x)} \, {\mathrm d}x$ which exists by Lemma \[lem:densite\_X\_t\].
Then, $r(t,x)$ is the unique solution to the partial differential equation $$\label{eq:FP}
\left\{
\begin{aligned}
\partial_t r(t,x)
&=\Delta r(t,x)
-\nabla V(x)\cdot\nabla r(t,x), \quad
t>0, x\in{{{\mathbb R}^d}},\\
r(0,x)&=r_0(x), \quad
x\in{{{\mathbb R}^d}},
\end{aligned}
\right.$$ in the sense of Definition \[defi:solution\_EDP\].
[F]{}rom Lemma \[lem:densite\_X\_t\], we know that $$\forall t \ge 0, {\mathrm d}x\text{-a.e.}, \, r(t,x) = {\mathbb E}(r_0(Y_{t}^x)).$$ The conclusion is then a consequence of the Feynman-Kac representation formula .
Long-time behavior of the partial differential equation when $\lambda=0$ {#sec:longtime}
-------------------------------------------------------------------------
In this section, we are interested in the long-time behavior of the partial differential equation when $\lambda=0$, which is related to the stochastic differential equation through the Feynman-Kac formula .
### The case ${\varphi}=0$
To study the long-time behavior of the solution to with $\lambda={\varphi}= 0$, we introduce the following hypothesis (defined for any $\eta >0$).
The measure $e^{-V}$ satisfies a Poincaré inequality with constant $\eta>0$: for any function $v$ in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$, $$\label{eq:poincare}
\eta\int_{{{\mathbb R}^d}}|v|^2(x){e^{-V(x)}{\mathrm d}x}\leq\int_{{{\mathbb R}^d}}|\nabla v|^2(x){e^{-V(x)}{\mathrm d}x}.$$
Recall that ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$ denotes the functions in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ with zero mean with respect to $\pi_0$ (see ).
\[prop:tps\_long\] Let Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} be satisfied for some positive $\eta$, and let $u$ be a solution to in the sense of Definition \[defi:solution\_EDP\], in the case $\lambda={\varphi}= 0$, with an initial condition $f \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. Then $u$ converges exponentially fast to the constant function $\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}$ in the following sense: $$\forall t \ge 0, \, \left\|u(t,\cdot)- \int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}\right\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}
\leq e^{-\eta t} \left\|f - \int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}\right\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}.$$
Taking the constant function $\mathbf1$ as the test function in , one obtains that $\int_{{{\mathbb R}^d}}u(t,x){e^{-V(x)}{\mathrm d}x}=\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}$ for any $t\ge 0$. In particular, $u(t,\cdot)-\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}\in{\mathbb L}_0^2({e^{-V(x)}{\mathrm d}x})$. In addition, from [@temam-79 Lemma $1.2$ p. $261$], one can take $u(t,\cdot)-\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}$ as the test function in , which yields, using the Poincaré inequality, $$\begin{aligned}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\int_{{{\mathbb R}^d}}\left|u(t,x)-\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}\right|^2{e^{-V(x)}{\mathrm d}x}&=-\int_{{{\mathbb R}^d}}|\nabla u(t,x)|^2{e^{-V(x)}{\mathrm d}x}\\ & \leq-\eta\int_{{{\mathbb R}^d}}\left|u(t,x)-\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}\right|^2{e^{-V(x)}{\mathrm d}x}.
\end{aligned}$$ One concludes from Grönwall’s lemma.
This Proposition shows that, under the assumption of Corollary \[cor:FP\] (namely ${{\mathrm d}}\mu_0=r(0,x) {e^{-V(x)}{\mathrm d}x}$ with $r(0,\cdot) \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$), the density $r(t,x)$ of $X^0_t$ with respect to $\pi_0$ converges exponentially fast to $1$ if [**([[**Poinc($\eta$)**]{}]{})**]{} is satisfied for some positive $\eta$. Actually, the convergence of ${\mathrm d}\mu_t$ to ${e^{-V(x)}{\mathrm d}x}$ holds in total variation norm for any initial condition $\mu_0$.
\[cor:CV\_L1\] Let Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} be satisfied for some positive $\eta$, and let $(X_t^0)_{t\geq0}$ evolve according to Equation . Let us assume that $X_0$ is distributed according to some probability measure $\mu_0$. Denote by $\mu_t$ the distribution of the random variable $X_t^0$, and for all $t>0$, denote by $q(t,x)$ the density of $\mu_t$ with respect to the Lebesgue measure (which exists according to Lemma \[lem:p\]). Then $$\lim_{t \to \infty} \|q(t,\cdot ) - e^{-V}\|_{{\mathbb L}^1({\mathrm d}x)} = 0.$$
[F]{}rom Equation , for all $t>0$, one has $${\mathrm d}y\text{-a.e.,}\, q(t,y)= \int p(t,x,y) \mu_0({\mathrm d}x).$$
Let us fix a positive ${\varepsilon}$. Let us consider $t_0 >0$ (to be fixed later on) and $q^{\varepsilon}(t_0,x)$ a function in ${\mathbb L}^\infty({\mathbb R}^d)$ which is non-negative, with compact support, such that $\int
q^{\varepsilon}(t_0,x) \, {\mathrm d}x =1$ and $$\int_{{\mathbb R}^d} |q(t_0,x) - q^{\varepsilon}(t_0,x)| \, {\mathrm d}x \le {\varepsilon}.$$ To build such a function $q^{\varepsilon}(t_0,\cdot)$, one could for example consider for $n$ large enough $\frac{\min(q(t_0,x), n) 1_{|x| \le n}}{\int \min(q(t_0,x), n) 1_{|x| \le n}}$ which indeed converges to $q(t,x)$ in ${\mathbb L}^1({\mathrm d}x)$ when $n \to \infty$. Let us define the function $$\, q^{\varepsilon}(t,y) = \int p(t-t_0,x,y)
q^{\varepsilon}(t_0,x) {\mathrm d}x \qquad \forall t \ge t_0, \forall y \in {\mathbb R}^d.$$ For $t \ge t_0$, $q^{\varepsilon}(t,\cdot)$ is the density at time $t$ of the process $(X_t^{0,{\varepsilon}})_{t \ge t_0}$ solution to , with $X_{t_0}^{0,{\varepsilon}}$ distributed according to $q^{\varepsilon}(t_0,x)
{\mathrm d}x$. Let us now set $r^{\varepsilon}(t,x)= e^{V(x)} q^{\varepsilon}(t,x)$, the density of $X_t^{0,{\varepsilon}}$ with respect to ${e^{-V(x)}{\mathrm d}x}$. Since $r^{\varepsilon}(t_0,x) = e^{V(x)} q^{\varepsilon}(t_0,x) \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, from Corollary \[cor:FP\], $(s,x) \mapsto
r^{\varepsilon}(t_0+s,x)$ satisfies the following partial differential equation (with unknown $r$) $$\left\{
\begin{aligned}
\partial_t r(s,x)
&=\Delta r(s,x)
-\nabla V(x)\cdot\nabla r(s,x)
,\quad s\ge 0, x\in{{{\mathbb R}^d}},\\
r(0,x)&=r^{\varepsilon}(t_0,x),
\quad x\in{{{\mathbb R}^d}},
\end{aligned}
\right.$$ in the sense of Definition \[defi:solution\_EDP\]. In particular, from Proposition \[prop:tps\_long\], since $\int r^{\varepsilon}(t_0,x)
{e^{-V(x)}{\mathrm d}x}=1$, $$\forall t \ge t_0, \, \|r^{\varepsilon}(t,\cdot)- 1\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}
\leq e^{-\eta (t-t_0)}\|r^{\varepsilon}(t_0,\cdot) -1 \|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}$$ which is equivalent to $$\forall t \ge t_0, \, \|q^{\varepsilon}(t,\cdot)- e^{-V}
\|_{{\mathbb L}^2(e^{V(x)} \, {\mathrm d}x)}
\leq e^{-\eta (t-t_0)}\|q^{\varepsilon}(t_0,\cdot) -
e^{-V}\|_{{\mathbb L}^2(e^{V(x)} \, {\mathrm d}x)}.$$ By Cauchy-Schwarz inequality, we deduce that $\forall t \ge t_0$ $$\begin{aligned}
\|q^{\varepsilon}(t,\cdot)- e^{-V}
\|_{{\mathbb L}^1({\mathrm d}x)}&=
\int_{{\mathbb R}^d} \left|q^{\varepsilon}(t,x)- e^{-V(x)} \right| e^{V(x)/2}
\times e^{-V(x)/2} \, {\mathrm d}x
\\
&\le \|q^{\varepsilon}(t,\cdot)- e^{-V}
\|_{{\mathbb L}^2(e^{V(x)} \, {\mathrm d}x)}
\leq e^{-\eta (t-t_0)}\|q^{\varepsilon}(t_0,\cdot) -
e^{-V}\|_{{\mathbb L}^2(e^{V(x)} \, {\mathrm d}x)}.\end{aligned}$$
Moreover, we also have: $\forall t \ge t_0$ $$\begin{aligned}
\|q(t,\cdot)- q^\varepsilon(t,\cdot)\|_{{\mathbb L}^1({\mathrm d}x)}
&=
\int_{{\mathbb R}^d} \left| \int_{{\mathbb R}^d} (q(t_0,x)-
q^\varepsilon(t_0,x) ) p(t-t_0,x,y) \, {\mathrm d}x\right|\, {\mathrm d}y\\
& \le
\int_{{\mathbb R}^d} \int_{{\mathbb R}^d} \left| q(t_0,x)-
q^\varepsilon(t_0,x) \right| p(t-t_0,x,y) \, {\mathrm d}x\, {\mathrm d}y\\
&=\int_{{\mathbb R}^d} \left| q(t_0,x)-
q^\varepsilon(t_0,x) \right| \, {\mathrm d}x \le {\varepsilon}.\end{aligned}$$ We thus obtain: $\forall t \ge t_0$, $$\begin{aligned}
\|q(t,\cdot)- e^{-V}\|_{{\mathbb L}^1({\mathrm d}x)}
&\le \|q(t,\cdot) - q^\varepsilon(t,\cdot)\|_{{\mathbb L}^1({\mathrm d}x)}
+ \|q^\varepsilon(t,\cdot)- e^{-V} \|_{{\mathbb L}^1({\mathrm d}x)} \\
& \leq {\varepsilon}+ e^{-\eta (t-t_0)}\|q^{\varepsilon}(t_0,\cdot) - e^{-V}\|_{{\mathbb L}^2(e^{V(x)} \, {\mathrm d}x)}\end{aligned}$$ and the right-hand side is smaller than $2 {\varepsilon}$ for $t$ sufficiently large. This concludes the proof.
### The case ${\varphi}\neq 0$
In this section, we are going to investigate the long-time behavior of the function $u$ defined by $$\label{eq:a_majorer}
u(t,x)={\mathbb E}\left[e^{-\int_0^t{\varphi}(Y_s^x){\mathrm d}s}\right],$$ for a generic function ${\varphi}$ where, we recall, $(Y^x_s)_{s \ge0}$ satisfies . When ${\varphi}\geq\alpha$ for some positive constant $\alpha$, $u$ converges to $0$ exponentially fast as $t\to\infty$. We now look for hypotheses on ${\varphi}$ under which this convergence is preserved in the case $\inf {\varphi}\le 0$.
Notice that by ergodicity, almost surely, $\lim_{t \to \infty} \frac 1 t \int_0^t {\varphi}(Y^x_s) \, ds = \int {\varphi}{e^{-V(x)}{\mathrm d}x}$ and therefore, the almost sure exponential decay to zero is ensured if $\int_{{{\mathbb R}^d}}{\varphi}(x) {e^{-V(x)}{\mathrm d}x}> 0$, at any rate in $(0,\int_{{{\mathbb R}^d}}{\varphi}(x) {e^{-V(x)}{\mathrm d}x})$. The exponential decay to zero in $L^1$ is more complicated to establish. In Proposition \[prop:conv\_expo\] below, we prove this exponential decay under a sufficient condition which contains the assumption $\int_{{{\mathbb R}^d}}{\varphi}(x) {e^{-V(x)}{\mathrm d}x}> 0$.
When ${\varphi}$ is bounded from below and locally Lipschitz, from Proposition \[prop:feynman-kac\], the function $u$ defined by is solution (in the sense of Definition \[defi:solution\_EDP\]) to the partial differential equation with $f = 1$ and $\lambda=0$. As a consequence, the long-time behavior of is related to the spectrum of the operator $\Delta-\nabla V\cdot\nabla-{\varphi}$ which is self-adjoint in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$.
One way to study this spectrum is to perform the change of variable $v(t,x)=e^{-\frac12V(x)}u(t,x)$, making Equation become $$\partial_tv
=\Delta v
+\frac14\left(2\Delta V-|\nabla V|^2-4{\varphi}\right)v.$$ As a consequence, the long-time behavior of $v$ is characterized by the spectrum of the Schrödigner operator $\Delta+\frac14\left(2\Delta V-|\nabla V|^2-4{\varphi}\right)$, which can be controlled by the Cwikel-Lieb-Rozenblum bound (see for example [@cwikel-77; @lieb-76; @rozenbljum-72]). Indeed, for $d\geq3$, this bound states that the number $N$ of nonnegative eigenvalues of $\Delta+W$ satisfies $$N\leq L_d\int_{{{\mathbb R}^d}}\max(W(x),0)^{d/2}{\mathrm d}x,$$ where $L_d$ is some constant independent of $W$. In particular, if there exists ${\varepsilon}>0$ such that $$\label{eq:critere_lieb_thirring}
\int_{{{\mathbb R}^d}}\max\left(
{\varepsilon}+\frac12\Delta V(x)-\frac14|\nabla V(x)|^2-{\varphi}(x)
,0\right)^{d/2}
{\mathrm d}x
<\frac1{L_d},$$ then the spectrum of $\Delta+\frac14\left(2\Delta V-|\nabla V|^2-4{\varphi}\right)$ is included in $(-\infty,-{\varepsilon})$.
There are two main concerns with this approach. First, the constant $L_d$ is unknown, so that the criterion is not quantitative. Moreover, by Jensen’s inequality, the exponential convergence to $0$ of $${\mathbb E}\left[e^{-\delta\int_0^t{\varphi}(X_s^0){\mathrm d}s}\right]$$ for $\delta>1$ implies the exponential convergence of the function $u(t,x)$ in . However, in some cases, the criterion may apply to $\delta{\varphi}$ for some $\delta >1$ and not to ${\varphi}$. We are going to present another criterion which does not present these flaws.
\[prop:conv\_expo\] Assume that [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$, and that $$\begin{aligned}
&-\infty<\inf{\varphi}\leq 0,\;\;
\int_{{{\mathbb R}^d}}{\varphi}(x){e^{-V(x)}{\mathrm d}x}>0,\;\;
\int_{{{\mathbb R}^d}}{\varphi}(x)^2{e^{-V(x)}{\mathrm d}x}<\infty,\label{eq:crit_conv1}\\
\mbox{and }&-(\inf{\varphi})\frac{\int_{{{\mathbb R}^d}}{\varphi}^2(x){e^{-V(x)}{\mathrm d}x}}{\left(\int_{{{\mathbb R}^d}}{\varphi}(x){e^{-V(x)}{\mathrm d}x}\right)^2}
<\eta.\label{eq:crit_conv2}
\end{aligned}$$
Let $\mathbf E=\int_{{{\mathbb R}^d}}{\varphi}(x){e^{-V(x)}{\mathrm d}x}$, $\mathbf
V=\int_{{{\mathbb R}^d}}\left({\varphi}(x)-\mathbf E\right)^2{e^{-V(x)}{\mathrm d}x}$ and let $u$ be the weak solution to Equation in the sense of Definition \[defi:solution\_EDP\] for $\lambda=0$, with an initial condition $f \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. The quantity $\int_{{{\mathbb R}^d}}u^2e^{-V}$ converges exponentially fast to $0$ as $t\to\infty$: $$\exists C>0,~\forall t>0,~\int_{{{\mathbb R}^d}}u^2(t,x){e^{-V(x)}{\mathrm d}x}\leq Ce^{-\beta t},$$ with a rate $\beta$ given by $$\label{eq:taux}
\beta=\left(\eta+\inf{\varphi}+\frac{\mathbf E^2+\mathbf V}{2\mathbf
E}\right) -
\sqrt{\left(\eta+\inf{\varphi}-\frac{\mathbf E^2+\mathbf V}{2\mathbf E}\right)^2+2\frac{\eta\mathbf V}{\mathbf E}}
>0.$$
Note that the positivity of the rate is equivalent to the condition . Moreover, note that the left-hand side in condition is homogeneous of order $1$ in ${\varphi}$, unlike the criterion obtained using the Cwikel-Lieb-Rozenblum bound. As a consequence, if the criterion applies to $\delta{\varphi}$ for some real number $\delta>1$, then it applies to ${\varphi}$, as expected.
In this proof, for notational simplicity, we omit the time and space variables in the integrals, which are all considered with respect to the Lebesgue measure on ${{{\mathbb R}^d}}$.
[From]{} [@temam-79 Lemma $1.2$ p. $261$], one can take $u$ as the test function in , obtaining $$\begin{aligned}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\int_{{{\mathbb R}^d}}u^2e^{-V}
&=-\int_{{{\mathbb R}^d}}|\nabla u|^2e^{-V}-\int_{{{\mathbb R}^d}}{\varphi}u^2 e^{-V}\\
&\leq -\int_{{{\mathbb R}^d}}|\nabla u|^2e^{-V}-\inf{\varphi}\int_{{{\mathbb R}^d}}u^2 e^{-V}.\end{aligned}$$ Using , one deduces that $$\label{eq:e1}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\int_{{{\mathbb R}^d}}u^2e^{-V}
\leq-(\eta+\inf{\varphi})\int_{{{\mathbb R}^d}}u^2e^{-V}
+\eta\left(\int_{{{\mathbb R}^d}}ue^{-V}\right)^2.$$ On the other hand, taking the constant function $\mathbf1$ as the test function in , $$\begin{aligned}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\left(\int_{{{\mathbb R}^d}}u e^{-V}\right)^2
&=-\int_{{{\mathbb R}^d}}ue^{-V}\int_{{{\mathbb R}^d}}{\varphi}ue^{-V}\\
&=-\int_{{{\mathbb R}^d}}{\varphi}e^{-V}\left(\int_{{{\mathbb R}^d}}u e^{-V}\right)^2
+\int_{{{\mathbb R}^d}}ue^{-V}\left(\int_{{{\mathbb R}^d}}{\varphi}e^{-V}\int_{{{\mathbb R}^d}}u e^{-V}-\int_{{{\mathbb R}^d}}{\varphi}ue^{-V}\right).\end{aligned}$$ By Cauchy-Schwarz inequality, $$\begin{aligned}
\label{eq:e2}
\left|\int_{{{\mathbb R}^d}}{\varphi}e^{-V}\int_{{{\mathbb R}^d}}u e^{-V}-\int_{{{\mathbb R}^d}}{\varphi}ue^{-V}\right|
&=\left|\int_{{{\mathbb R}^d}}\left({\varphi}-\int_{{{\mathbb R}^d}}{\varphi}e^{-V}\right) \left(u-\int_{{{\mathbb R}^d}}u e^{-V}\right)e^{-V}\right|\nonumber\\
&\le\mathbf V^{1/2}\left(\int_{{{\mathbb R}^d}}\left(u-\int_{{{\mathbb R}^d}}ue^{-V}\right)^2e^{-V}\right)^{1/2}.\end{aligned}$$ Therefore, from the inequality $2ab\leq a^2+b^2$, $$\begin{aligned}
\label{eq:e3}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\left(\int_{{{\mathbb R}^d}}u e^{-V}\right)^2
&\leq-\frac{\mathbf E}2\left(\int_{{{\mathbb R}^d}}u e^{-V}\right)^2
+\frac{\mathbf V}{2\mathbf E}
\left(\int_{{{\mathbb R}^d}}\left(u-\int_{{{\mathbb R}^d}}ue^{-V}\right)^2e^{-V}\right)\nonumber\\
&=-\frac{\mathbf E^2+\mathbf V}{2\mathbf E}\left(\int_{{{\mathbb R}^d}}u e^{-V}\right)^2
+\frac{\mathbf V}{2\mathbf E}\int_{{{\mathbb R}^d}}u^2e^{-V}.\end{aligned}$$ By combining and , one obtains for $\delta\geq 0$, $$\begin{aligned}
\label{eq:majo_cle}
\frac12\frac{{\mathrm d}}{{\mathrm d}t}\left( \int_{{{\mathbb R}^d}}u^2e^{-V} + \delta \left(\int_{{{\mathbb R}^d}}u
e^{-V}\right)^2\right)
\leq-c_1(\delta)\int_{{{\mathbb R}^d}}u^2e^{-V}
-c_2(\delta)\left(\int_{{{\mathbb R}^d}}ue^{-V}\right)^2\end{aligned}$$ with $$\label{eq:expr_c1_c2}
c_1(\delta)
=\eta
+\inf{\varphi}-\frac{\delta\mathbf V}{2\mathbf E}
\mbox{ and }
c_2(\delta)
=-\eta
+\frac\delta2\frac{\mathbf V+\mathbf E^2}{\mathbf E}.$$ We want to find $\delta\geq0$ such that ensures exponential convergence to $0$ of $\int_{{{\mathbb R}^d}}u^2e^{-V}$ as $t\to+\infty$. If $\delta=0$, one has $c_2(0)=-\eta<0$, so we need $\delta>0$. We look for $\delta>0$ such that both $c_1(\delta)$ and $c_2(\delta)$ are positive.
[From]{} , $c_1(\delta)$ is positive if and only if $$\frac\delta{2\mathbf E}<\frac{\eta+\inf{\varphi}}{\mathbf V}$$ and $c_2(\delta)$ is positive if and only if $$\frac\delta{2\mathbf E}>\frac\eta{\mathbf V+\mathbf E^2}.$$ One concludes by checking that condition is necessary and sufficient for the interval $\left(\frac\eta{\mathbf V+\mathbf E^2},\frac{\eta+\inf{\varphi}}{\mathbf V}\right)$ to be nonempty.
[For]{} a given $\delta>0$, Equation gives a convergence rate of $2\min(c_1(\delta),\frac{c_2(\delta)}\delta)$. [From]{} the definition of $c_1$ and $c_2$, one can see that $c_1(\delta)$ is nonincreasing and, under , $\frac{c_2(\delta)}\delta$ is nondecreasing in $\delta$. As a consequence, $\min(c_1(\delta),\frac{c_2(\delta)}\delta)$ is maximized for $\delta c_1(\delta)=c_2(\delta)$. This last equation is quadratic, and one can check that its unique positive solution is $$\delta
=\frac{\mathbf E}{\mathbf V}
\left(
\eta+\inf{\varphi}-\frac{\mathbf V+\mathbf E^2}{2\mathbf E}+\sqrt{\left(\eta+\inf{\varphi}-\frac{\mathbf V+\mathbf E^2}{2\mathbf E}\right)^2+2\frac{\eta \mathbf V}{\mathbf E}}
\right),$$ giving the rate .
One can see that Equation is necessary and sufficient for the existence of $\delta>0$ such that $c_1(\delta)>0$ and $c_2(\delta)>0$. One can naturally wonder whether introducing more flexibility in the inequalities used in the proof of Proposition \[prop:conv\_expo\] could lead to a weaker condition. Actually, keeping track of the positive term $\int_{{{\mathbb R}^d}}|\nabla u|^2e^{-V}$ in , using inequality in , and using the inequality $2ab\leq\gamma a^2+\frac1\gamma b^2$ in leads to the exact same necessary and sufficient condition to ensure exponential convergence to $0$.
One can use the theory of large deviations to prove that the long-time behavior of quantities of the form is necessarily exponential, with a rate given by a variational formula.
Let $(X_t)_{t\geq0}$ evolve according to Equation . According to Donsker-Varadhan’s lemma, the random probability measure $\mu_t$ defined by the formula $$\mu_t(A)
=\frac1t\int_0^t\mathbf 1_{X_s\in A}{\mathrm d}s,$$ satisfies a large deviation principle with rate function $$I(\nu)
=\begin{cases}
\displaystyle\int_{{{\mathbb R}^d}}\left|\nabla\sqrt f\right|^2{e^{-V(x)}{\mathrm d}x}&\mbox{ if } \exists f:{\mathbb R}^d
\to {\mathbb R}, \, \nu=f(x){e^{-V(x)}{\mathrm d}x},\\
\infty&\mbox{ otherwise,}
\end{cases}$$ see for example [@den-hollander-00 Chapter IV.4]. As a consequence, in the long-time limit, $$-\frac1t\log\left({\mathbb E}\left[e^{-\int_0^t{\varphi}(X_s){\mathrm d}s}\right]\right)
=-\frac1t\log\left({\mathbb E}\left[e^{-t\left<{\varphi},\mu_t\right>}\right]\right)$$ converges to the constant $\alpha$ defined by $$\alpha=\inf_f
\int_{{{\mathbb R}^d}}\left|\nabla \sqrt f\right|^2(x){e^{-V(x)}{\mathrm d}x}+\int_{{{\mathbb R}^d}}{\varphi}(x)f(x){e^{-V(x)}{\mathrm d}x}$$ where the infimum is taken over all probability densities with respect to the measure ${e^{-V(x)}{\mathrm d}x}$, from Varadhan’s lemma in large deviations theory. By the change of variables $g^2=f$, $\alpha$ is equal to $$\inf_g
\frac{\int_{{{\mathbb R}^d}}\left|\nabla g\right|^2(x){e^{-V(x)}{\mathrm d}x}+\int_{{{\mathbb R}^d}}{\varphi}(x)g^2(x){e^{-V(x)}{\mathrm d}x}}
{\int_{{{\mathbb R}^d}}g^2(x){e^{-V(x)}{\mathrm d}x}},$$ which is the bottom of the spectrum of the operator $-\Delta+\nabla
V\cdot\nabla+{\varphi}$ which is self-adjoint in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, already discussed at the beginning of this section.
Let us give two examples where the result from Proposition \[prop:conv\_expo\] applies.
A first example is given by the double-well potential in dimension $1$, defined by $$V_\gamma(x)=x^4-\gamma x^2 + C_\gamma,$$ where $\gamma>0$ and $C_\gamma=\ln \left( \int_{{\mathbb R}} \exp(-x^4+\gamma x^2)\,
{\mathrm d}x \right)$. We want to apply Proposition \[prop:conv\_expo\] to the case where the function ${\varphi}$ is a multiple of $\min{\rm
Spec}\nabla^2V_\gamma(x)$ (see Section \[sec:expo\_cv\_R\] for a justification of this choice for ${\varphi}$). In the present case, this writes ${\varphi}_{\gamma,\delta}(x)=\delta(12x^2-2\gamma)$, where $\delta$ is the positive multiplicative factor.
Denote by $\eta_{\gamma}$ the optimal Poincaré constant associated to the potential $V_\gamma$. As $\gamma$ goes to $0$, the limit potential $x^4$ satisfies a Poincaré inequality with constant $\eta_0>0$, owing to its convexity. As a consequence, the Poincaré constants $\eta_{\gamma}$ converge to a positive limit. On the other hand, as $\gamma$ goes to $0$, the quantity $$-(\inf{\varphi}_{\gamma,\delta})
\frac
{\int_{\mathbb R}{\varphi}_{\gamma,\delta}^2(x)e^{-V_\gamma(x)}{\mathrm d}x}
{\left(\int_{\mathbb R}{\varphi}_{\gamma,\delta}(x)e^{-V_\gamma(x)}{\mathrm d}x\right)^2}$$ goes to $0$, since $\inf{\varphi}_{\gamma,\delta}$ goes to $0$ while $\frac
{\int_{\mathbb R}{\varphi}_{\gamma,\delta}^2(x)e^{-V_\gamma(x)}{\mathrm d}x}
{\left(\int_{\mathbb R}{\varphi}_{\gamma,\delta}(x)e^{-V_\gamma(x)}{\mathrm d}x\right)^2}$ converges to some positive constant. As a consequence, for any $\delta>0$ the inequality is satisfied for $\gamma$ smaller than some critical value depending on $\delta$. Notice that the inequalities are satisfied for any $\gamma
>0$ since, for any smooth potential $V:{\mathbb R}\mapsto {\mathbb R}$, $\int_{\mathbb R}V''(x){e^{-V(x)}{\mathrm d}x}>0$ holds from a mere integration by parts.
The second example is given by an identically vanishing potential $V(x)=0$, with the equation considered on the one-dimensional torus, identified with the segment $[0,2\pi]$ with periodic boundary conditions. The invariant measure is then the uniform measure on the torus. Consider the function ${\varphi}(x)=\sin(x)+\alpha$ with $\alpha\geq0$. In that cases, the mean value of ${\varphi}$ is given by $\alpha$, and ${\varphi}(x)$ is not nonnegative for all values of $x$ as soon as $\alpha<1$.
In that case, the Poincaré constant is given by $\eta=1$ and $\inf{\varphi}$ is given by $\alpha-1$. As a consequence, Equation writes $$(1-\alpha)\frac{1+2\alpha^2}{2\alpha^2}<1$$ The condition is thus satisfied if $\alpha>\alpha_0$ where $\alpha_0$ is the unique real root of the equation $$\alpha^3+\frac12\alpha-\frac12=0,$$ given by $\alpha_0\simeq0.590$. As a consequence, for $\alpha\in(\alpha_0,1)$, one has exponential convergence of to zero while the function ${\varphi}$ is not uniformly positive.
Existence and uniqueness of an invariant measure $\pi_\lambda$ for {#sec:inv_meas}
-------------------------------------------------------------------
In all this section, we assume that Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$. We would like to show that the stochastic differential equation admits a unique invariant probability measure that we denote in the following $\pi_\lambda$, and to give an explicit formula for this measure. Of course, for $\lambda=0$, we have $${\mathrm d}\pi_0={e^{-V(x)}{\mathrm d}x}$$ and one result of this section is that it is the unique invariant measure for . We will use $\pi_0$ as a reference measure to build functional spaces, and to construct the invariant measure $\pi_\lambda$ by perturbative arguments, using the crucial assumption on the boundedness of $F_\lambda+\nabla
V=F_\lambda-F_0$ (see Assumption[**([[**Drift**]{}]{})**]{}-$(i)$): for $\lambda
\in [0,\lambda_0],$ $$\|F_\lambda-F_0\|_{{\mathbb L}^\infty({{{\mathbb R}^d}})}\le C \lambda.$$
Let us begin with some notation. We denote by $\mathcal L_\lambda=F_\lambda\cdot\nabla+\Delta$ the generator of the process $(X_t^\lambda)_{t\geq0}$. In particular, $\mathcal L_0=-\nabla V\cdot\nabla+\Delta$. Also denote $$\label{eq:Tlambda}
\mathcal T_\lambda
=\mathcal L_\lambda-\mathcal L_0
=(F_\lambda+\nabla V)\cdot\nabla$$ The space ${\mathbb L}_0^2({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$ endowed with the symmetric bilinear form$$\label{eq:produit_scalaire}
(u,v)
\mapsto
\int_{{{\mathbb R}^d}}\nabla u(x)\cdot\nabla v(x){e^{-V(x)}{\mathrm d}x}$$ is a Hilbert space by Assumption [**([[**Poinc($\eta$)**]{}]{})**]{}. A consequence of the Riesz theorem is that for any $u\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, there exists a unique function $v$ in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$ such that $$\forall w\in{\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}),~
\int_{{{\mathbb R}^d}}\nabla v(x)\cdot\nabla w(x){e^{-V(x)}{\mathrm d}x}=\int_{{{\mathbb R}^d}}u(x)w(x){e^{-V(x)}{\mathrm d}x}.$$ This function is denoted $v=- \mathcal L_0^{-1}u$ since when $v$ is smooth, $\int_{{{\mathbb R}^d}}\nabla v(x)\cdot\nabla w(x){e^{-V(x)}{\mathrm d}x}= - \int_{{{\mathbb R}^d}}\mathcal L_0v(x)w(x){e^{-V(x)}{\mathrm d}x}$. We denote by $\mathcal D(\mathcal L_0)$ the domain of $\mathcal L_0$, defined by $$\mathcal D(\mathcal L_0)
=\left\{ v \in {\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}),\, \mathcal L_0 v \in {\mathbb L}_0^2({e^{-V(x)}{\mathrm d}x})\right\}.$$
For a function $u \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, from the Poincaré inequality, one has $$\begin{aligned}
\eta\|\mathcal L_0^{-1} u\|^2_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}
\leq\int_{{{\mathbb R}^d}}|\nabla (\mathcal L_0^{-1} u)(x)|^2{e^{-V(x)}{\mathrm d}x}&=-\int_{{{\mathbb R}^d}}(\mathcal L_0^{-1} u)(x) u(x){e^{-V(x)}{\mathrm d}x},\\
&\leq\|\mathcal L_0^{-1} u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}\| u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}\end{aligned}$$ which implies $$\label{eq:spectre_L0}
\eta\|\mathcal L_0^{-1} u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}\leq\|u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}.$$ In the following, we will use the orthogonal projection operator $\Pi_0$ from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ onto ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$ defined by: $$\label{eq:PI0}
\forall f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}),~
\Pi_0f
=f-\int_{{{\mathbb R}^d}}f(x){e^{-V(x)}{\mathrm d}x}.$$
Let us now explain formally how we obtain an explicit formula for the invariant measure $\pi_\lambda$ of . For any test function ${\varphi}$ and since ${\mathcal
L}_{\lambda} 1=0$, $\int_{{{{\mathbb R}^d}}}
{\mathcal L}_{\lambda} \Pi_0({\varphi}) \, {\mathrm d}\pi_\lambda =0 $. Thus, by considering $f={\mathcal L}_0 \Pi_0({\varphi})$, for any test function $f$, $\int_{{{{\mathbb R}^d}}}
{\mathcal L}_{\lambda} {\mathcal
L}_{0}^{-1} \Pi_0 f \, {\mathrm d}\pi_\lambda =0 $ which also writes $\int_{{{{\mathbb R}^d}}} (I + {\mathcal T}_\lambda
{\mathcal
L}_{0}^{-1} \Pi_0) \Pi_0 f \, {\mathrm d}\pi_\lambda =0 $ where $I$ denotes the identity operator. This is equivalent to: for any test function $f$, $\int_{{{{\mathbb R}^d}}} (I + {\mathcal T}_\lambda
{\mathcal
L}_{0}^{-1} \Pi_0) f \, \frac{{\mathrm d}\pi_\lambda}{{\mathrm d}\pi_0} {\mathrm d}\pi_0 =\int_{{{{\mathbb R}^d}}} f {\mathrm d}\pi_0 $ which yields $(I + {\mathcal T}_\lambda
{\mathcal L}_{0}^{-1} \Pi_0)^* \frac{{\mathrm d}\pi_\lambda}{{\mathrm d}\pi_0}= 1$, where $*$ denotes the dual operator on the Hilbert space ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. As a consequence, we are naturally led to study the operator $\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0$ defined from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. The aim of the next Lemma is to show rigorously that we can define an invariant measure $\pi_\lambda$ of by defining its Radon-Nikodym derivative with respect to $\pi_0$ as $(I + ({\mathcal T}_\lambda
{\mathcal L}_{0}^{-1} \Pi_0)^*)^{-1} 1$.
We can now state the result concerning the existence of an invariant measure for .
\[lem:pil\] Let us assume that Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$. Then there exists $\lambda_1 \in (0,\lambda_0]$ such that for $\lambda \in [0,\lambda_1]$, the dual operator $I +(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*$ on the Hilbert space ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ of the operator $I +\mathcal
T_\lambda\mathcal L_0^{-1}\Pi_0$ is invertible and has a bounded inverse.
Let us then introduce, for $\lambda \in [0,\lambda_1]$, the function $g_\lambda \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ and the associated measure $\pi_\lambda$ such that $$\label{eq:pi_lam}
{\mathrm d}\pi_\lambda= g_\lambda {{\mathrm d}\pi_0} \text{ where } g_\lambda=(I+(\mathcal T_\lambda\mathcal
L_0^{-1}\Pi_0)^*)^{-1}\mathbf1\,.$$ The measure $\pi_\lambda$ is a probability measure which is invariant for the process $(X_t^\lambda)_{t\geq0}$ solution to .
: Let us first study the operator $\mathcal T_\lambda
\mathcal L_0^{-1} \Pi_0$. [From]{} the boundedness assumption on $\nabla
V+F_\lambda=F_\lambda-F_0$ (see Assumption[**([[**Drift**]{}]{})**]{}-$(i)$), the definition of ${\mathcal L}_0^{-1}$ and , for any $u \in \mathcal {\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$, $$\begin{aligned}
\|\mathcal T_\lambda \mathcal L_0^{-1}u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}^2
&=\int_{{{\mathbb R}^d}}|(F_\lambda+\nabla V)(x)\cdot\nabla (\mathcal L_0^{-1}u)(x)|^2{e^{-V(x)}{\mathrm d}x}\\
&\leq C\lambda^2\int_{{{\mathbb R}^d}}|\nabla (\mathcal L_0^{-1}u)(x)|^2{e^{-V(x)}{\mathrm d}x}\\
&=-C\lambda^2\int_{{{\mathbb R}^d}}(\mathcal L_0^{-1}u)(x) u(x){e^{-V(x)}{\mathrm d}x}\\
&\leq C\lambda^2\|\mathcal L_0^{-1} u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}\|u\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}\leq C\frac{\lambda^2}\eta\|u\|^2_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}.\end{aligned}$$
As a consequence, the operator $\mathcal T_\lambda\mathcal L_0^{-1}$ is bounded from ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$ to ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, with: $$\label{eq:Ol}
\|\mathcal T_\lambda\mathcal L_0^{-1}\|_{\mathcal
L({\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x}),{\mathbb L}^2({e^{-V(x)}{\mathrm d}x}))} \le\sqrt{\frac{C}{\eta}} \lambda.$$ By composition, $\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0$ is thus a bounded operator from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to itself, with a norm of order $\mathcal O(\lambda)$, and so is $(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*$. As a consequence, for $\lambda$ small enough, the operator $I+(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*$ is invertible from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to itself.
: Let us now introduce the function $g_\lambda
\in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ defined by $$g_\lambda=(I+(\mathcal T_\lambda\mathcal
L_0^{-1}\Pi_0)^*)^{-1} \mathbf 1$$ and let us prove that ${\mathrm d}\pi_\lambda= g_\lambda {\mathrm d}\pi_0$ is invariant for the stochastic differential equation . Let $(Y^{\lambda,x}_t)_{t \ge}$ be the solution to with initial condition $Y^{\lambda,x}_0=x$ (see ). Using the Markov property, the aim is to prove that, for any ${\mathcal C}^\infty$ bounded test function $f:{{{\mathbb R}^d}}\to {\mathbb R}$, $$\label{eq:inv_egalite}
\int_{{{{\mathbb R}^d}}} {\mathbb E}(f(Y^{\lambda,x}_t)) g_\lambda(x) {e^{-V(x)}{\mathrm d}x}= \int_{{{{\mathbb R}^d}}} f(x)
g_\lambda(x) {e^{-V(x)}{\mathrm d}x}.$$ [From]{} Proposition \[prop:feynman-kac\], we know that $u(t,x)={\mathbb E}(f(Y^{\lambda,x}_t))$ is the solution to (with ${\varphi}=0$), and from Proposition \[prop:EDP\_bien\_posee\], we have for any $T>0$, $$u \in
{\mathbb L}^\infty\left([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})
\right)
\cap{\mathbb H}^1\left([0,T],{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})\right).$$ Moreover, from Proposition \[prop:regularite\_EDP\_phi\], $u$ is a classical solution to . Therefore, $$\begin{aligned}
\frac{d}{d t} \int_{{{{\mathbb R}^d}}} {\mathbb E}(f(Y^{\lambda,x}_t)) g_\lambda(x) {e^{-V(x)}{\mathrm d}x}&=
\frac{d}{d t} \int_{{{{\mathbb R}^d}}} u(t,x) g_\lambda(x) {e^{-V(x)}{\mathrm d}x}\\
&=
\int_{{{{\mathbb R}^d}}} \partial_t u(t,x) g_\lambda(x) {e^{-V(x)}{\mathrm d}x}\\
&=
\int_{{{{\mathbb R}^d}}} {\mathcal L}_\lambda u(t,x) g_\lambda(x) {e^{-V(x)}{\mathrm d}x},\end{aligned}$$ and ${\mathcal L}_\lambda u = \partial_t u \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. Now, notice that for any function $\psi$ which is the sum of a ${\mathcal C}^\infty$ function with compact support and a constant, $$\mathcal T_\lambda \psi=\mathcal T_\lambda\mathcal
L_0^{-1}\Pi_0\mathcal L_0 \psi$$ holds true since $\mathcal T_\lambda$ sends constant functions to $0$. Therefore, for any such function $\psi$, one has $$\begin{aligned}
\int_{{\mathbb R}^d} \mathcal L_\lambda \psi g_\lambda {\mathrm d}\pi_0=
\int_{{\mathbb R}^d} \left[ (\mathcal L_0+\mathcal T_\lambda)\psi \right] g_\lambda {\mathrm d}\pi_0
&=
\int_{{\mathbb R}^d} \left[ (I+\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)\mathcal
L_0\psi \right] g_\lambda {\mathrm d}\pi_0\\
&=
\int_{{\mathbb R}^d} \left[ \mathcal
L_0\psi \right] (I+(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*) g_\lambda {\mathrm d}\pi_0\\
&=
\int_{{\mathbb R}^d} \mathcal
L_0\psi \, {\mathrm d}\pi_0.\end{aligned}$$ Since $\pi_0$ is invariant for the dynamics with infinitesimal generator ${\mathcal L}_0$, the right-hand side is zero. By density, the equality $\int_{{\mathbb R}^d} \mathcal L_\lambda \psi
g_\lambda {\mathrm d}\pi_0=0$ holds for any function $\psi$ such that $\mathcal
L_\lambda \psi \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. Therefore, $\frac{d}{d t} \int_{{{{\mathbb R}^d}}}
{\mathbb E}(f(Y^{\lambda,x}_t)) g_\lambda(x) {e^{-V(x)}{\mathrm d}x}=0$ which yields after integration in time over $[0,t]$.
: Let us finally check that $\pi_\lambda$ is a probability measure. First, one has $$\begin{aligned}
\int_{{\mathbb R}^d}g_\lambda {\mathrm d}\pi_0=\int_{{\mathbb R}^d}(I+(\mathcal T_\lambda\mathcal
L_0^{-1}\Pi_0)^*)^{-1} \mathbf 1(I+(\mathcal T_\lambda\mathcal
L_0^{-1}\Pi_0))\mathbf 1{\mathrm d}\pi_0=\int_{{\mathbb R}^d}{\mathrm d}\pi_0=1.\end{aligned}$$ Second, one can prove that $g_\lambda \ge 0$. Indeed, from and the fact that $Y^{\lambda,x}_t$ admits a density $p^\lambda(t,x,y)$ with respect to the Lebesgue measure (see Lemma \[lem:p\]), we have $$\int_{{{{\mathbb R}^d}}} \int_{{{{\mathbb R}^d}}} f(y) p^\lambda(t,x,y) \, {\mathrm d}y g_\lambda(x) {e^{-V(x)}{\mathrm d}x}= \int_{{{{\mathbb R}^d}}} f(x)
g_\lambda(x) {e^{-V(x)}{\mathrm d}x}.$$ This equality holds for any smooth test function $f$ and, by a density argument, one can apply it to the bounded function $f(x)=\textrm{sgn}(g_\lambda(x))$, where $\textrm{sgn}(y)=1_{y \ge 0}
-1_{y <0}$ denotes the sign function. One thus obtains $$\int_{{{{\mathbb R}^d}}} \int_{{{{\mathbb R}^d}}} \big(\textrm{sgn}(g_\lambda(y)) \textrm{sgn}(g_\lambda(x)) -1\big) p^\lambda(t,x,y) \, {\mathrm d}y |g_\lambda|(x) {e^{-V(x)}{\mathrm d}x}= 0.$$ Thus, $\big(\textrm{sgn}(g_\lambda(y)) \textrm{sgn}(g_\lambda(x))
-1\big) p^\lambda(t,x,y) |g_\lambda|(x)=0$ ${\mathrm d}x \otimes {\mathrm d}y$-a.e.. Since $p^\lambda(t,x,y)>0$ ${\mathrm d}x \otimes {\mathrm d}y$-a.e. (see Lemma \[lem:p\]) and $\int_{{{{\mathbb R}^d}}} |g_\lambda|(x) \,
{\mathrm d}x >0$, this implies that ${\mathrm d}y$-a.e., $\textrm{sgn}(g_\lambda(y))=1$ or ${\mathrm d}y$-a.e $\textrm{sgn}(g_\lambda(y))=-1$. The conclusion then follows from the fact that $\int_{{\mathbb R}^d}g_\lambda {\mathrm d}\pi_0=1$.
Notice that in the case $\lambda=0$, we indeed have $g_0=1$. The next result states the uniqueness of the invariant measure for .
\[lem:erggen\] Let us assume that Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$. For $\lambda\in[0,\lambda_1]$ ($\lambda_1$ being the constant introduced in Lemma \[lem:pil\]), the unique invariant measure of the stochastic differential equation is the probability measure $\pi_\lambda$ defined by . This probability measure is equivalent to the Lebesgue measure on ${\mathbb R}^d$ and for any initial condition $X_0$, $$\label{eq:erggen}
\forall f\in{\mathbb L}^1(\pi_\lambda),\;{\mathbb P}\left(\lim_{t \to \infty} \frac1t\int_0^tf(X^{\lambda}_s){\mathrm d}s
= \int_{{{\mathbb R}^d}}f {\mathrm d}\pi_\lambda\right)=1.$$
Let $\lambda\in[0,\lambda_1]$. Lemma \[lem:pil\] ensures that $\pi_\lambda$ defined by is an invariant probability measure for ${\mathrm d}X_t^\lambda=F_\lambda(X_t^\lambda){\mathrm d}t+\sqrt2{\mathrm d}W_t$. For $X_0$ distributed according to any invariant probability measure, Lemma \[lem:p\] ensures that this measure is equivalent to the Lebesgue measure. As a consequence, all the invariant probability measures are equivalent and the dynamics admits exactly one invariant probability measure $\pi_\lambda$. Since $\pi_\lambda$ is the only invariant probability measure, it is ergodic (see for example [@rey-bellet-06 Theorem 3.8 and Equation (52)]) and denoting by $(Y^{\lambda,x}_t)_{t\geq 0}$ the solution to started from $Y_0=x\in{\mathbb R}^d$, $$\forall f\in{\mathbb L}^1(\pi_\lambda),\;{\mathrm d}x\mbox{ a.e.},\;{\mathbb P}\left(\lim_{t \to \infty} \frac1t\int_0^tf(Y^{\lambda,x}_s){\mathrm d}s
= \int_{{{\mathbb R}^d}}f {\mathrm d}\pi_\lambda\right)=1.$$ For any initial condition $X_0$, since the law of $X^{\lambda}_1$ is absolutely continuous with respect to the Lebesgue measure (see Lemma \[lem:p\]), follows by the Markov property.
Tangent vector of the diffusion {#sect:vecteur_tangent}
===============================
In all this Section, $(X^\lambda_t)_{t \ge 0}$ denotes the process solution to , with an initial condition $X_0$ which, we recall, does not depend on $\lambda$. We establish various results on the tangent vector $T_t$ defined by , which naturally appears in the estimators and to evaluate ${\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)$.
Definition and interpretations of the tangent vector
----------------------------------------------------
If the function $f$ is differentiable, one can write ${\partial_\lambda^0}\left(f(X_t^\lambda)\right)=T_t\cdot\nabla f(X_t^0)$, where the process $(T_t)_{t\geq0}$ is the so-called *tangent vector*, defined as $$\label{eq:definition_Tt}
T_t
={\partial_\lambda^0}X_t^\lambda,$$ and the existence of which is ensured by the following proposition.
\[proptt\]For any $t\geq0$, the function $\lambda\mapsto X_t^\lambda$ is almost surely differentiable, and the definition of the tangent vector makes sense. Moreover, $(T_t)_{t\geq0}$ almost surely satisfies the following ordinary differential equation whose coefficients depend on $(X_t^0)_{t\geq0}$: $$\label{eq:edo_Tt}
\left\{
\begin{aligned}
\frac{{\mathrm d}T_t}{{\mathrm d}t}
&={\partial_\lambda^0}F_\lambda(X_t^0)-\nabla^2V(X_t^0) T_t,\\
T_0&=0.
\end{aligned}
\right.$$
By [**([[**Drift**]{}]{})**]{}-$(i)$ and the continuity of $\nabla V$ and $(X^0_t)_{t\geq 0}$, $t\mapsto |{\partial_\lambda^0}F_\lambda(X_t^0)|+|\nabla^2V(X_t^0)|$ is locally bounded. Hence admits a unique solution $(T^0_t)_{t\geq 0}$ by the Cauchy-Lipschitz theorem. Let us prove that for $\bar t>0$, $\lambda\mapsto X^\lambda_{\bar t}$ is differentiable at $\lambda =0$ with derivative equal to $T^0_{\bar t}$. For $\lambda \in[0,\lambda_0]$, we set $\tau_\lambda=\inf\{t\geq 0:|X^\lambda_t|\geq \sup_{t\in[0,\bar t]}|X^0_t|+1\}$ with convention $\inf\emptyset=+\infty$. Let $L^{X^0}_{\bar t}=\sup_{x\in{\mathbb R}^d:|x|\leq \sup_{t\in[0,\bar t]}|X^0_t|+1}|\nabla^2V(x)|$. For $t\in[0,\bar{t}]$, one has $$\begin{aligned}
\sup_{s\in[0,t]}|X^\lambda_{s\wedge\tau_\lambda}-X^0_{s\wedge\tau_\lambda}|&\leq \int_0^{t\wedge\tau_\lambda}\left|F_\lambda(X^\lambda_s)+\nabla V(X^\lambda_s)\right|+\left|\nabla V(X^0_s)-\nabla V(X^\lambda_s)\right|{\mathrm d}s\\&\leq C\lambda t+L^{X^0}_{\bar t}\int_0^t|X^\lambda_{s\wedge\tau_\lambda}-X^0_{s\wedge\tau_\lambda}|{\mathrm d}s\end{aligned}$$ so that $\sup_{s\in[0,t]}|X^\lambda_{s\wedge\tau_\lambda}-X^0_{s\wedge\tau_\lambda}|\leq \frac{C\big(e^{L^{X^0}_{\bar t}t}-1\big)}{L^{X^0}_{\bar t}}\lambda$. For $\lambda\leq \frac{L^{X^0}_{\bar t}}{C\big(e^{L^{X^0}_{\bar t}t}-1\big)}$, one deduces that $\tau_\lambda\geq \bar{t}$ and $\sup_{s\in[0,t]}|X^\lambda_{s}-X^0_{s}|\leq \frac{C \big(e^{L^{X^0}_{\bar t}t}-1\big)}{L^{X^0}_{\bar t}}\lambda$. In particular, $X^\lambda_t$ converges to $X^0_t$ uniformly for $t\in[0,\bar t]$. Now, for $t\geq 0$, $$\begin{aligned}
X^\lambda_{t}-X^0_{t}=\int_0^{t}(F_\lambda(X^\lambda_s)-F_0(X^\lambda_s)){\mathrm d}s+\int_0^t\nabla^2V(\xi^\lambda_s)(X^0_{s}-X^\lambda_{s})ds,\end{aligned}$$ where, by a slight abuse of notations, $\nabla^2V(\xi^\lambda_s)$ stands for the matrix $(\partial_{ij}V(\xi^{\lambda,i}_s))_{1\leq
i,j\leq d}$ and $\forall i\in\{1,\hdots,d\}$, $\xi^{\lambda,i}_s\in[X^0_{s},X^\lambda_{s}]$. For $s\in[0,\bar t]$ and $\lambda\in \bigg (0,\lambda_0\wedge
\frac{L^{X^0}_{\bar t}}{C(e^{L^{X^0}_{\bar t}t}-1)}\bigg]$, $|\xi^{\lambda,i}_s|\leq\sup_{t\in[0,\bar t]}|X^0_t|+1$. Hence for $t\in[0,\bar{t}]$, $$\begin{aligned}
\sup_{s\in[0,t]} \left|\frac{X^\lambda_{s}-X^0_{s}}{\lambda}-T^0_s\right|\leq &\int_0^{t}\left|\frac{F_\lambda(X^\lambda_s)-F_0(X^\lambda_s)}{\lambda}-{\partial_\lambda^0}F_\lambda(X_s^0)\right|+|\nabla^2V(X^0_s)-\nabla^2V(\xi^\lambda_s)||T^0_s|{\mathrm d}s\\&+L^{X^0}_{\bar t}\int_0^t\left|\frac{X^\lambda_{s}-X^0_{s}}{\lambda}-T^0_s\right|{\mathrm d}s.\end{aligned}$$ By [**([[**Drift**]{}]{})**]{}-$(i)$-$(ii)$ and the uniform convergence of $X^\lambda_t$ to $X^0_t$ for $t\in[0,\bar t]$, $$\lim_{\lambda \to 0}
\int_0^{\bar
t}\left|\frac{F_\lambda(X^\lambda_s)-F_0(X^\lambda_s)}{\lambda}-{\partial_\lambda^0}F_\lambda(X_s^0)\right|+|\nabla^2V(X^0_s)-\nabla^2V(\xi^\lambda_s)||T^0_s|{\mathrm d}s = 0.$$With Grönwall’s lemma, one concludes that $\sup_{s\in[0,\bar{t}]}\left|\frac{X^\lambda_{s}-X^0_{s}}{\lambda}-T^0_s\right|$ converges to $0$ as $\lambda\to 0$.
We have the following expression of $(T_t)_{t\geq0}$ as an integral:
\[prop:tangent\] Define the *resolvent* $(R_{X^0}(s,t))_{s,t\geq0}$ associated with Equation as the solution, with values in ${\mathbb R}^{d\times d}$, to the following ordinary differential equation: $$\label{eq:resolvante}
\left\{
\begin{aligned}
\partial_tR_{X^0}(s,t)&=-\nabla^2V(X_t^0)R_{X^0}(s,t),~s,t\geq0,\\
R_{X^0}(s,s)&=I_d,~s\geq0,
\end{aligned}
\right.$$ where $I_d$ is the $d\times d$ identity matrix. The resolvent satisfies the following semigroup property $$\label{eq:semigroupe_R}
\forall r,s,t\in[0,\infty),~R_{X^0}(s,t)R_{X^0}(r,s)=R_{X^0}(r,t).$$ One can recover the tangent vector from the resolvent through the following formula: $$\label{eq:expression_Tt}
\forall t\geq0,
~T_t=\int_0^tR_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s.$$
The semigroup property is a consequence of uniqueness for Equation , satisfied by the two processes $(R_{X^0}(s,t))_{t\geq 0}$ and $(R_{X^0}(r,t)R_{X^0}(r,s)^{-1})_{t\geq0}$.
In view of the differential equations satisfied by $(T_t)_{t\geq0}$ and $(R_{X^0}(s,t))_{t\geq0}$, one has, from the equality $R_{X^0}(t,0)=R_{X^0}(0,t)^{-1}$, $$\begin{aligned}
\partial_t(R_{X^0}(t,0) T_t)
&=-R_{X^0}(t,0)\partial_t(R_{X^0}(0,t))R_{X^0}(t,0) T_t
+R_{X^0}(t,0)\partial_tT_t\\
&=R_{X^0}(t,0)\nabla^2V(X_t^0)R_{X^0}(0,t)R_{X^0}(t,0) T_t\\
&\quad+R_{X^0}(t,0){\partial_\lambda^0}F_\lambda(X_t^0)
-R_{X^0}(t,0)\nabla^2V(X_t^0) T_t\\
&=R_{X^0}(t,0){\partial_\lambda^0}F_\lambda(X_t^0).\end{aligned}$$ Integrating over $[0,t]$, one obtains $$R_{X^0}(t,0) T_t=\int_0^tR_{X^0}(s,0){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s,$$ and the result follows by using the semigroup property .
Notice that the resolvent is also the differential of the trajectory with respect to its initial condition.
\[lemderci\]Let $(Y_t^x)_{t \ge 0}$ solve . Then for any $t\geq 0$, $x\mapsto Y^x_t$ is ${\mathcal C}^1$ on ${\mathbb R}^d$ with Jacobian matrix $(D Y_t^x)_{i,j}=\partial_{x_j}Y_t^{i,x}$ given by $DY_t^x=R_{Y^x}(0,t)$.
By standard results on ordinary differential equations, $x\mapsto Y^x_t$ is ${\mathcal C}^1$ with Jacobian matrix $DY_t^x$ solving the equation $$\forall t\geq0,
~DY_t^x
=I_d
-\int_0^t\nabla^2V(Y_s^x)DY_s^x{\mathrm d}s,$$ obtained by spatial derivation of $Y_t^x
=x
-\int_0^t\nabla V(Y_s^x){\mathrm d}s
+\sqrt2W_t.$ By uniqueness for , one has $DY_t^x=R_{Y^x}(0,t)$.
In the following, we will need the following result about the link between the forward resolvent and its backward counterpart.
\[lem:retournement\_resolvante\] Let $(Y_s)_{0\leq s\leq t}$ satisfy Equation with $Y_0$ distributed according to $\pi_0$. [From]{} the reversibility of the dynamics , the process $(Z_s)_{0\leq s\leq t}$ defined by $Z_s=Y_{t-s}$ has the same law as $(Y_s)_{0\leq s\leq t}$, and one has the relation $$R_Y(0,s)
=R_Z^T(t-s,t),$$ where $R_Z^T$ is the transposed matrix of the resolvent associated with $Z$.
Uniqueness holds for the ordinary differential equation satisfied by $s \mapsto R_Y(0,s)$: $$\label{eq:EDO_semigroupe}
\left\{
\begin{aligned}
\frac{{\mathrm d}R}{{\mathrm d}s} (s)&=-\nabla^2V(Y_s)R(s),\\
R(0)&=I_d.
\end{aligned}
\right.$$ One can check that $s \mapsto R_Z^T(t-s,t)$ also solves . Indeed, since, by the semigroup property, $R_Z(t-s,t)=R_Z(t,t-s)^{-1}$, one has, for $s \in [0,t]$, $$\begin{aligned}
\partial_sR_Z(t-s,t)
&=-R_Z(t-s,t)\left(\partial_sR_Z(t,t-s)\right)R_Z(t-s,t)\\
&=-R_Z(t-s,t)\nabla^2V(Z_{t-s})R_Z(t,t-s)R_Z(t-s,t)\\
&=-R_Z(t-s,t)\nabla^2V(Z_{t-s})\\
&=-R_Z(t-s,t)\nabla^2V(Y_s).\end{aligned}$$ This concludes the proof.
Almost sure boundedness of $R_{X^0}(0,t)$ and $T_t$
---------------------------------------------------
The tangent vector can take large values, since the second term in the right-hand side of will provide exponential growth for $(T_t)_{t\geq0}$, typically when the trajectory $(X_t^0)_{t\geq0}$ is close to a local maximum of $V$, or when it crosses a saddle point of $V$. In the sequel, we need some assumptions on $V$ to control this behavior.
### Local-in-time boundedness of $R_{X^0}(s,t)$ and $T_t$
Let us first introduce an assumption which will be sufficient to get the local-in-time boundedness of $R_{X^0}(s,t)$ and $T_t$.
The matrix-valued function $\nabla^2V:{{{\mathbb R}^d}}\to{\mathbb R}^{d\times d}$ is bounded from below, in the sense that there exists $\alpha\in{\mathbb R}$ such that, for all $x,h\in{{{\mathbb R}^d}}$, $$h\cdot\nabla^2V(x)h\geq\alpha|h|^2.$$ Equivalently, the spectrum of $\nabla^2V(x)$ is bounded from below by $\alpha$, uniformly in $x$.
Under Assumption [**([[**min Spec**]{}]{})**]{}, the random variables $T_t$ and $R_{X^0}(s,t)$ are bounded:
\[lem:R\_T\_Linfty\] One has $$\label{eq:R_T_Linfty}
\forall 0 \le s<t,~
|R_{X^0}(s,t)|
\leq e^
{-\int_s^t\min\mathrm{Spec}\left(\nabla^2V(X_u^0)\right){\mathrm d}u},$$ ${\mathbb R}^{d\times d}$ being endowed with the matricial norm associated with the Euclidean norm on ${{{\mathbb R}^d}}$. In addition, if the Assumption [**([[**min Spec**]{}]{})**]{} is satisfied, for any $T>0$, the random variables $\sup_{0\leq s\leq t\leq T}|R_{X^0}(s,t)|$ and $\sup_{0\leq t\leq T}|T_t|$ lie in ${\mathbb L}^\infty(\Omega)$.
[For]{} any vector $x$, one has $$\begin{aligned}
\partial_t|R_{X^0}(s,t)x|^2
&=-2(R_{X^0}(s,t)x)^T\nabla^2V(X_t^0)(R_{X^0}(s,t)x)\\
&\leq-2\min\textrm{Spec}\left(\nabla^2 V(X_t^0)\right)|R_{X^0}(s,t)x|^2.\end{aligned}$$ As a consequence, one has the estimation $$|R_{X^0}(s,t)x|^2
\leq|x|^2e^
{-2\int_s^t\min\textrm{Spec}\left(\nabla^2V(X_u^0)\right){\mathrm d}u}$$ so that holds. If the Assumption [**([[**min Spec**]{}]{})**]{} is satisfied, this inequality proves that $R(s,t)$ is in ${\mathbb L}^\infty(\Omega)$ locally uniformly in time. [From]{} the expression of $T_t$ and the boundedness of ${\partial_\lambda^0}F_\lambda$, one also concludes that $T_t\in{\mathbb L}^\infty(\Omega)$, locally uniformly in time.
### Global-in-time boundedness of $R_{X^0}(0,t)$
We need some additional assumption on the convexity of the potential for $(R_{X^0}(0,t))_{t\geq0}$ to be bounded globally in time.
The potential $V$ is such that $$\label{eq:trou_moyen}
\int_{{{\mathbb R}^d}}\max\left(0,-\min\mathrm{Spec}\left(\nabla^2V(x)\right)\right){e^{-V(x)}{\mathrm d}x}<\infty
\mbox{ and }
\int_{{{\mathbb R}^d}}\min\mathrm{Spec}\left(\nabla^2V(x)\right){e^{-V(x)}{\mathrm d}x}>0.$$
In this assumption, the first inequality, always satisfied under [**([[**min Spec**]{}]{})**]{}, ensures that the integral $\int_{{{\mathbb R}^d}}\min\mathrm{Spec}\left(\nabla^2V(x)\right){e^{-V(x)}{\mathrm d}x}$ is well defined in $(-\infty,\infty]$. We refer to Appendix \[sec:annex\_conv\] for a discussion of this Assumption.
\[lem:convergence\_R\] Under Assumptions [**([[**min Spec**]{}]{})**]{} and [**([[**Conv**]{}]{})**]{}, the resolvent matrix $R_{X^0}(0,t)$ almost surely converges to $0$ as $t$ goes to infinity, with exponential rate. Namely, for any $\beta$ with $$0
<\beta
<\int_{{{\mathbb R}^d}}\min\mathrm{Spec}\left(\nabla^2V(x)\right){e^{-V(x)}{\mathrm d}x},$$ there exists an almost surely finite random variable $C>0$ such that $$\label{eq:convergence_R}
\forall t\geq0,~|R_{X^0}(0,t)|\leq Ce^{-\beta t}.$$
[From]{} Lemma \[lem:erggen\], by ergodicity (see ), one has $$\lim_{t\to\infty}\frac1t\int_0^t\min\textrm{Spec}\left(\nabla^2 V(X_s^0)\right){\mathrm d}s
=\int_{{{\mathbb R}^d}}\min\textrm{Spec}\left(\nabla^2V(x)\right)e^{-V(x)}{\mathrm d}x
~~\mbox{ a.s.}$$ We conclude by combining this limit with Equation .
\[rem:hypconv\] While Assumption [**([[**Conv**]{}]{})**]{} is automatically satisfied in dimension $1$ from a mere integration by parts, this is not the case in higher dimension. Indeed, if one applies the integration by parts formula in this case, one only obtains that $$\int_{{{\mathbb R}^d}}\nabla^2V(x){e^{-V(x)}{\mathrm d}x}=\int_{{{\mathbb R}^d}}\nabla V(x)\otimes\nabla V(x)e^{-V(x)}{\mathrm d}x$$ is a positive definite matrix (because of the integrability of $e^{-V}$, for any $y$ in ${{{\mathbb R}^d}}$, the function $x\mapsto\nabla V(x)\cdot y$ cannot be the zero function), so that the minimum of its spectrum is positive. A counterexample to Assumption [**([[**Conv**]{}]{})**]{} is given by a tensor potential $V(x)=U(x_1)+\hdots+U(x_d)$ with a well chosen function $U$. Indeed, in this case the left hand side of equation rewrites $$\int_{{{\mathbb R}^d}}\min_{i \in \{1, \ldots,d\}}(U''(x_i))e^{-\sum_i
U(x_i)}{\mathrm d}x_1\hdots{\mathrm d}x_d={\mathbb E}\left[\min_{i \in \{1, \ldots,d\}} U''(X_i)\right],$$ where $X_i$ are i.i.d random variables with distribution $e^{-U(x)}{\mathrm d}x$. If $U$ is chosen so that $U''$ is bounded and has a strictly negative minimum, then the sequence $\left(\displaystyle\min_{i \in \{1, \ldots,d\}}
U''(X_i)\right)_{d\geq1}$ converges almost surely as $d$ goes to infinity to the negative constant $\min U''$. Then, from the dominated convergence theorem, the quantity ${\mathbb E}\left[\displaystyle\min_{i \in \{1, \ldots,d\}} U''(X_i)\right]$ converges to $\min U''$, and is thus negative when $d$ is large enough.
Assumption [**([[**Conv**]{}]{})**]{} is not necessary for to hold. Indeed, if the matrices $\nabla^2V(x)$ commute, the matrix $R_{X^0}$ is given by $$R_{X^0}(0,t)=e^{- \int_0^t\nabla^2V(X_s^0){\mathrm d}s}.$$ and the convergence of $\frac1t\int_0^t\nabla^2V(X_s^0){\mathrm d}s$ to the positive definite matrix $\int_{{{\mathbb R}^d}}\nabla^2V(x){e^{-V(x)}{\mathrm d}x}$ implies that holds for $\beta < \min {\rm Spec} \left(\int_{{{\mathbb R}^d}}\nabla^2V(x){e^{-V(x)}{\mathrm d}x}\right)$, even in the cases when $V$ does not satisfy Assumption [**([[**Conv**]{}]{})**]{}. An example where the matrices $\nabla V^2(x)$ commute is the case of a tensor potential $V(x)=U(x_1)+\hdots+U(x_d)$. As seen before, $U$ and $d$ can be chosen such that $V$ does not satisfy Assumption [**([[**Conv**]{}]{})**]{}.
However, it is likely that Lemma \[lem:convergence\_R\] does not hold under the sole ergodicity property: $$\lim_{t\to\infty}\frac1t\int_0^t \left(\nabla^2 V(X_s^0)\right){\mathrm d}s
=\int_{{{\mathbb R}^d}}\nabla V(x) \otimes \nabla V(x) e^{-V(x)}{\mathrm d}x
~~\mbox{ a.s.}$$ Indeed, there exists some family of symmetric matrices $(A_t)_{t\geq0}$ converging in the Cesàro sense to a positive-definite matrix, for which the solution to $\frac{{\mathrm d}}{{\mathrm d}t} R_t=-A_tR_t$, $R_0=I_d$ does not converge to $0$ as $t$ goes to infinity. An example of this phenomenon is given by $$A_t
= \Omega_t\begin{pmatrix}-1&0\\0&3\end{pmatrix}\Omega_t^T,
\mbox{ where }
\Omega_t
=\begin{pmatrix}\cos t&-\sin t\\\sin t&\cos t\end{pmatrix}
.$$ Indeed, the family $(A_t)_{t\geq0}$ converges in the Cesàro sense to $I_d$ as $t$ goes to infinity, but the associated matrix $(R_t)_{t\geq0}$ diverges. To show this last point, consider the matrix $M_t=\Omega_t^TR_t$. Since $\partial_t\Omega_t=\Omega_t\begin{pmatrix}0&-1\\1&0\end{pmatrix}$, then $$\partial_tM_t
=\begin{pmatrix}1&1\\-1&-3\end{pmatrix}M_t$$ holds. As a consequence, $R_t=\Omega_t\exp\left(t\begin{pmatrix}1&1\\-1&-3\end{pmatrix}\right)$. The eigenvalues of the matrix $\begin{pmatrix}1&1\\-1&-3\end{pmatrix}$ are $-1-\sqrt3$ and $-1+\sqrt3$, the latter being positive, so that $R_t$ diverges as $t$ goes to infinity.
Boundedness of moments of $R_{X^0}(s,t)$ and $T_t$ {#sec:expo_cv_R}
--------------------------------------------------
In the sequel, we will need to control the moments of $T_t$. [From]{} Equation and the boundedness of ${\partial_\lambda^0}F_\lambda$ (see Assumption [**([[**Drift**]{}]{})**]{}-(i)), this boils down to estimating the moments of $R_{X^0}(s,t)$. For this purpose, from , it is enough control expectations of the form$~{\mathbb E}\left[e^{-\beta \int_0^t\min\mathrm{Spec}(\nabla^2V)(Y_s^x){\mathrm d}s}\right]$, where $\beta$ is a positive constant.
### Preliminary result when $X_0 \sim \pi_0$
One can deduce from Proposition \[prop:conv\_expo\] a criterion for exponential convergence of the moments of $R_{X^0}(0,t)$ to $0$ as $t\to\infty$. To state the result, we need to strengthen the assumptions [**([[**min Spec**]{}]{})**]{} and [**([[**Conv**]{}]{})**]{} which is the point of the following assumption. For any $\rho>0$, let us consider:
Assume that $$\begin{aligned}
&-\infty<\inf_{x \in {\mathbb R}^d} \min\mathrm{Spec}(\nabla^2V(x)) \leq 0,\\
& \int_{{{\mathbb R}^d}}\min\mathrm{Spec}(\nabla^2V(x)) {e^{-V(x)}{\mathrm d}x}>0 \text{ and } \int_{{{\mathbb R}^d}}\left(\min\mathrm{Spec}(\nabla^2V(x)) \right)^2{e^{-V(x)}{\mathrm d}x}<\infty,\\
&- (\inf
\min\mathrm{Spec}(\nabla^2V(x)))\frac{\int_{{{\mathbb R}^d}}\left(\min\mathrm{Spec}(\nabla^2V(x)) \right)^2{e^{-V(x)}{\mathrm d}x}}{\left(\int_{{{\mathbb R}^d}}\min\mathrm{Spec}(\nabla^2V(x)){e^{-V(x)}{\mathrm d}x}\right)^2}
<\rho.
\end{aligned}$$
Notice that for $\eta >0$ and $\beta >0$ under assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} and [**([[**Spec($\eta
/ \beta $)**]{}]{})**]{} then the assumptions and of Proposition \[prop:conv\_expo\] are satisfied with ${\varphi}(x)=\beta \min\mathrm{Spec}(\nabla^2V(x))$.
We are now in position to state a simple consequence of Proposition \[prop:conv\_expo\]:
\[cor:momr\] Let $(X^0_t)_{t\geq 0}$ solve starting from $X_0$ distributed according to $\pi_0$. Assume that [**([[**Poinc($\eta$)**]{}]{})**]{} and [**([[**Spec($\eta
/ \beta $)**]{}]{})**]{} hold for some $\eta >0$ and $\beta >0$. Then there is a constant $C\in(0,+\infty)$ such that $$\forall t\geq 0,\;{\mathbb E}_{\pi_0}[|R_{X^0}(0,t)|^\beta]\leq Ce^{-t/C}.$$
Here and in the following, the notation ${\mathbb E}_{\pi_0}$ means that the initial condition $X_0$ of the processes $(X^\lambda_t)_{t \ge 0}$ solution to is distributed according to $\pi_0$.
To apply Proposition \[prop:feynman-kac\] to the function ${\varphi}(x)=\beta \min\mathrm{Spec}(\nabla^2V(x))$, we need this function to be locally Lipschitz. Since $\nabla^2 V$ is locally Lipschitz (see Assumption [**([[**Pot**]{}]{})**]{}-$(i)$), this is a consequence of the Lemma \[lem:min\_spec\_lipschitz\] given below. By Proposition \[prop:feynman-kac\], the function $u(t,x)={\mathbb E}\left[
e^{-\int_0^t\beta\min\mathrm{Spec}\left(\nabla^2V(Y_s^x)\right){\mathrm d}s}
\right]$ is the solution to Equation in the sense of Definition \[defi:solution\_EDP\] for ${\varphi}(x)=\beta\min\mathrm{Spec}(\nabla^2V(x))$ and $f(x)=1$.
Since conditions and hold for this choice of ${\varphi}$, Equation and Proposition \[prop:conv\_expo\] give $${\mathbb E}_{\pi_0}\left[|R_{X^0}(0,t)|^\beta\right]\leq \int_{{\mathbb R}^d}u(t,x){e^{-V(x)}{\mathrm d}x}\leq \left(\int_{{\mathbb R}^d}u^2(t,x){e^{-V(x)}{\mathrm d}x}\right)^{1/2}
\leq Ce^{-t/C}$$ for some positive constant $C$.
\[lem:min\_spec\_lipschitz\] The function $A\mapsto\min\mathrm{Spec}(A)$ is a Lipschitz function on the space of symmetric $d\times d$ matrices.
Let $A$ be a symmetric matrix, and let $x$ be a vector in ${{{\mathbb R}^d}}$ such that $|x|=1$ and $\min\mathrm{Spec}(A)=x\cdot Ax$. Then, for any symmetric matrix $B$, one has $$\min\mathrm{Spec}(B)
\leq x\cdot Bx
=x\cdot (B-A)x+x\cdot Ax
\leq|A-B|+\min\mathrm{Spec}(A),$$ ${\mathbb R}^{d\times d}$ being endowed with the matricial norm associated with the Euclidean norm on ${{{\mathbb R}^d}}$. By exchanging $A$ and $B$ in the previous inequality, one obtains $$|\min\mathrm{Spec}(A)-\min\mathrm{Spec}(B)|
\leq|A-B|.$$
### Uniform-in-time boundedness of moments of $T_t$ {#sect:variance_bornee}
Numerically, the computation of through the long-time limit of a Monte Carlo approximation of the expression ${\mathbb E}[T_t\cdot\nabla f(X_t^0)]$ is only possible if $T_t$ has a bounded variance uniformly in time.
A case where this fact is easily proved is when the function $V$ is $\eta$-convex, where $\eta$ is a positive constant. We recall that this means that the spectrum of $\nabla^2V(x)$ is bounded from below by $\eta$, independently of $x$. More precisely, one has the following proposition.
\[prop:varalconv\] Assume that the $V$ is $\eta$-convex, for a positive constant $\eta$. Then, for any $\alpha
\ge 1$, $$\sup_{t \ge 0} {\mathbb E}|T_t|^\alpha < \infty.$$ In particular, $T_t$ has a bounded variance uniformly in time.
By and the boundedness of ${\partial_\lambda^0}F_\lambda$, one has $$\begin{aligned}
{\mathbb E}[|T_t|^\alpha]
={\mathbb E}\left[\left|\int_0^t R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right|^\alpha\right]\leq&{\mathbb E}\left[\left(C\int_0^te^{-\alpha(t-s)}{\mathrm d}s\right)^\alpha\right]<\infty,\end{aligned}$$ so that $T_t$ has a finite moment of order $\alpha$.
The convexity assumption on the potential can be loosened, as shown in the next Proposition.
\[prop:variance\] Let $\alpha\in[1,+\infty)$. Assume that [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$, that the initial condition $X_0$ is distributed according to a measure $\mu_0$ having a density with respect to the measure ${e^{-V(x)}{\mathrm d}x}$ which is in ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$ for some $p \in (1,
\infty]$, and Assumption [**([[**Spec($\eta (p-1)
/ (\alpha p) $)**]{}]{})**]{} holds (with the convention $\eta ( \infty-1)
/ (\alpha \infty)= \eta/\alpha$). Then, $$\sup_{t \ge 0} {\mathbb E}|T_t|^\alpha < \infty.$$ and, when $\alpha\geq 2$, $T_t$ has a bounded variance uniformly in time.
By and [**([[**Drift**]{}]{})**]{}-$(i)$, $${\mathbb E}^{1/\alpha}[|T_t|^\alpha]\leq \int_0^t{\mathbb E}^{1/\alpha}[|R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0)|^\alpha]{\mathrm d}s\leq C\int_0^t{\mathbb E}^{1/\alpha}[|R_{X^0}(s,t)|^\alpha]{\mathrm d}s$$
Let $\mu_s$ denote the law of $X^0_s$ for $s\geq 0$ and $(Y_t)_{t\geq
0}$ be a solution to with $Y_0$ distributed according to $\pi_0$. We notice that the Markov property gives: for $0\leq s\leq t$, $${\mathbb E}[|R_{X^0}(s,t)|^\alpha]
={\mathbb E}\left[|R_{Y}(s,t)|^\alpha\frac{{\mathrm d}\mu_s}{{e^{-V(x)}{\mathrm d}x}}(Y_s)\right].$$ Using Hölder inequality with $q=p/(p-1)$ ($q=1$ if $p=\infty$), Lemma \[lem:densite\_X\_t\] and Proposition \[cor:momr\], one deduces that for $t\geq s \ge 0$, $$\begin{aligned}
{\mathbb E}[|R_{X^0}(s,t)|^\alpha]
&\leq {\mathbb E}\left[|R_Y(s,t)|^{\alpha q}\right]^{\frac1q}
\left\|\frac{{\mathrm d}\mu_s}{{e^{-V(x)}{\mathrm d}x}}\right\|_{{\mathbb L}^p({e^{-V(x)}{\mathrm d}x})}\leq C e^{-\frac{t-s}{C}}.
\end{aligned}$$ This concludes the proof.
We are now in position to give sufficient conditions for the finiteness of the variance of the two estimators and .
Let $f: {{{\mathbb R}^d}}\to {\mathbb R}$ be a ${\mathcal C}^1$ function such that $\nabla
f$ is bounded. Let us assume that either $V$ is $\eta$-convex (for a positive constant $\eta$), or that there exists $\eta >0$ and $p\in
(1,\infty]$ such that [**([[**Poinc($\eta$)**]{}]{})**]{} holds, $X_0$ is distributed according to a measure $\mu_0$ having a density in ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$ with respect to ${e^{-V(x)}{\mathrm d}x}$ and Assumption [**([[**Spec($\eta (p-1)
/ (2 p) $)**]{}]{})**]{} holds. Then, $$\sup_{t \ge 0} {\rm Var}(T_t \cdot \nabla f (X^0_t) ) < \infty \text{ and
} \sup_{t \ge 0} {\rm Var}\left(\frac 1 t \int_0^t T_s \cdot \nabla f (X^0_s) \,
{\mathrm d}s\right) < \infty.$$
These results are simple consequences of the boundedness of $\nabla f$ and Proposition \[prop:variance\] for $\alpha=2$.
[From]{} the Central Limit Theorem for trajectorial averages (see for example [@duflo-97 Section 2.1.3, Theorem 6.3.20]), it is expected that the variance of $\frac 1 t \int_0^t T_s \cdot \nabla f (X^0_s) \,
{\mathrm d}s$ actually scales like $1/t$ in the limit $t \to
\infty$. This requires for example to prove the existence of a solution to the Poisson problem associated with the Markov process $(X^0_s,T_s)_{s \ge 0}$, which does not seem to be ensured under our set of assumptions. We leave the study of this issue to a future work.
Under the additional assumption [**([[**V**]{}]{})**]{} given in Appendix \[sec:hypdelta\], it is possible to extend the previous results to more general initial conditions. Assume that the initial condition $X_0$ is distributed according to a measure $\mu_0$ such that the measure $e^{\frac12V} {\mathrm d}\mu_0$ can be written as $$\label{eq:mu0V}
e^{\frac12V(x)} {\mathrm d}\mu_0 =f(x){\mathrm d}x+{\mathrm d}\nu,$$ where $f$ is some function in ${\mathbb L}^p({\mathrm d}x)$ with $p \in [1,2]$ and $\nu$ is some finite measure on ${{{\mathbb R}^d}}$. From , for any $t>0$, $\mu_t$ is absolutely continuous with respect to $e^{- V(x)}{\mathrm d}x$ with $\frac{{{\mathrm d}}\mu_t}{{e^{-V(x)}{\mathrm d}x}} \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$. Now, by the semi-group property satisfied by $R_{X^0}$, and the fact that $- \min{ \rm
Spec}(\nabla^2 V(x)) \le C < \infty$, one has for $\varepsilon>0$, $$|R_{X^0}(s,t)|
\leq|R_{X^0}(s\vee\varepsilon,t)R_{X^0}(s,s\vee\varepsilon)|
\leq e^{C(\varepsilon-s)^+}|R_{X^0}(s\vee\varepsilon,t)|.$$ For $\alpha>0$, using a similar change of measure as in the previous proof, the fact that $\frac{{{\mathrm d}}\mu_{\varepsilon}}{{e^{-V(x)}{\mathrm d}x}} \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ and Proposition \[cor:momr\], one deduces that under Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} and [**([[**Spec($\eta / (2\alpha)$)**]{}]{})**]{}, for $t\geq s\vee \varepsilon$, $$\begin{aligned}
{\mathbb E}[|R_{X^0}(s,t)|^\alpha]
&\leq e^{C(\varepsilon-s)^+}{\mathbb E}[|R_Y(s\vee\varepsilon,t)|^{2\alpha}]^{\frac12}
{\mathbb E}\left[\left(\frac{{\mathrm d}\mu_\varepsilon}{{e^{-V(x)}{\mathrm d}x}}(Y_\varepsilon)\right)^2\right]^{\frac12}\\
&\leq Ce^{C(\varepsilon-s)^+}e^{-\frac{t-s\vee\varepsilon}{C}}\leq Ce^{C\varepsilon}e^{-\frac{t-s}{C}}.
\end{aligned}$$ This estimation remains valid for $0\leq s\leq t\leq\varepsilon$ up to increasing $C$, since then, by and the fact that $- \min{ \rm
Spec}(\nabla^2 V(x)) \le C < \infty$, $|R_{X^0}(s,t)|\leq
e^{C\varepsilon}$. In conclusion, for $\alpha\geq 1$, under Assumptions [ **([[**V**]{}]{})**]{}, [**([[**Poinc($\eta$)**]{}]{})**]{} and [**([[**Spec($\eta / (2\alpha) $)**]{}]{})**]{}, $\sup_{t\geq 0}{\mathbb E}[|T_t|^\alpha]<\infty$ if $\mu_0$ satisfies .
The Green-Kubo formulae {#sect:GK}
=======================
A first way to compute the derivative is to use the Green-Kubo formula (see for example [@hairer-majda-10] for a mathematical approach and [@chandler-87; @evans-morriss-08] for physical motivations). This formula gives an expression of in terms of the time autocorrelations of $(X_t^0)_{t\geq0}$, where $(X_t^0)_{t\geq0}$ satisfies with an initial condition $X_0$ being distributed according to the equilibrium measure $\pi_0$.
Finite time Green-Kubo formula
------------------------------
We start with the Green-Kubo formula in finite time, which will not be used in the sequel of the paper, but motivates the infinite horizon Green-Kubo formula.
\[theo:GK\_tps\_fini\] Let $f \in{\mathbb L}^1({e^{-V(x)}{\mathrm d}x})$ be a Lipschitz function and let $\nabla f$ be its gradient in the sense of distributions which can be identified with its almost everywhere gradient. Suppose that the initial condition $X_0$ is distributed according to the equilibrium measure $\pi_0$ and that Assumption [**([[**min Spec**]{}]{})**]{} is satisfied. Then, for any $t\geq 0$, for any $\lambda\in[0,\lambda_0]$, $f(X_t^\lambda)
$ is integrable and $\lambda \mapsto {\mathbb E}_{\pi_0}[f(X_t^\lambda)]$ is differentiable at $0$ with derivative $$\label{eq:GK_tps_fini}
{\partial_\lambda^0}{\mathbb E}_{\pi_0}[f(X_t^\lambda)]
=\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right]{\mathrm d}s.$$
Since $X^0_t$ is distributed according to ${e^{-V(x)}{\mathrm d}x}$, Proposition \[proptt\] and the chain rule ensure that $\lambda\mapsto f(X^\lambda_t)$ is a.s. differentiable at $\lambda=0$ with derivative $\nabla f (X^0_t).T_t$.\
To justify the interchange between the derivation and the expectation, we need some integrability property. For $\lambda\in (0,\lambda_0]$ and $t\geq 0$, one has, using [**([[**min Spec**]{}]{})**]{} and [**([[**Drift**]{}]{})**]{}-$(i)$ for the inequality: $$\begin{aligned}
|X^\lambda _t-X^0_t|^2&=2\int_0^t(F_\lambda(X^\lambda_s)+\nabla V(X^\lambda_s)).(X^\lambda_s-X^0_s){\mathrm d}s+2\int_0^t(\nabla V(X^0_s)-\nabla V(X^\lambda_s)).(X^\lambda_s-X^0_s){\mathrm d}s\\
&\leq C\lambda^2t+\left(1-2\alpha\right)\int_0^t|X^\lambda _s-X^0_s|^2{\mathrm d}s.\end{aligned}$$ As a consequence, $$\forall \lambda\in (0,\lambda_0],\;\frac{|X^\lambda _t-X^0_t|^2}{\lambda^2}\leq C\frac{e^{(1-2\alpha)t}-1}{1-2\alpha}\label{majodif}$$ with the convention that the last ratio is equal to $t$ if $1-2\alpha=0$. With the Lipschitz continuity of $f$, one deduces that the random variable $\frac{f(X^\lambda_t)-f(X^0_t)}{\lambda}$ is bounded by a deterministic constant not depending on $\lambda$. The integrability of $f(X^\lambda_t)$ then follows from the integrability of $f(X^0_t)$ where $X^0_t$ is distributed according to ${e^{-V(x)}{\mathrm d}x}$ and $f\in {\mathbb L}^1({e^{-V(x)}{\mathrm d}x})$.
Moreover, by Lebesgue’s theorem, $\lambda \mapsto {\mathbb E}_{\pi_0}[f(X_t^\lambda)]$ is differentiable at $0$ with derivative $$\label{eq:green_kubo_preuve}
{\partial_\lambda^0}{\mathbb E}_{\pi_0}[f(X_t^\lambda)]
={\mathbb E}_{\pi_0}[\nabla f(X_t^0)\cdot T_t]
=\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(X_t^0)\cdot R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0)\right]{\mathrm d}s,$$ where we used for the second equality. All terms in Equation are well defined thanks to Lemma \[lem:R\_T\_Linfty\]. Let us now rewrite the right-hand side of . By introducing the process $(Y_s)_{0 \le s \le t}=(X^0_{t-s})_{0\leq s\leq t}$ (which has the same law as $(X^0_s)_{0 \le s \le t}$), using a change of variable $s\to t-s$ and Lemma \[lem:retournement\_resolvante\], we get $$\begin{aligned}
\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(X_t^0)\cdot R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0)\right]{\mathrm d}s
&=\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(X_{t}^0)\cdot R_{X^0}(t-s,t){\partial_\lambda^0}F_\lambda(X^0_{t-s})\right]{\mathrm d}s\\
&=\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(Y_{0})\cdot R_{Y}^T(0,s){\partial_\lambda^0}F_\lambda(Y_{s})\right]{\mathrm d}s\\
&=\int_0^t{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right] {\mathrm d}s.\end{aligned}$$ This completes the proof of .
The conclusion of Theorem \[theo:GK\_tps\_fini\] still holds if $f \in{\mathbb L}^1({e^{-V(x)}{\mathrm d}x})$ is a ${\mathcal C}^1$ function such that $\nabla f$ is uniformly continuous on ${\mathbb R}^d$ and $\nabla f\in{\mathbb L}^1({e^{-V(x)}{\mathrm d}x})$.
It is possible to give another expression of the right-hand side in .
\[prop:GK\_formule\_magique\] Let $f \in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ be a Lipschitz function. Assume [**([[**min Spec**]{}]{})**]{}and consider the process $(X_t^0)_{t\geq0}$ satisfying with an initial condition $X_0$ being distributed according to the equilibrium measure $\pi_0$. For almost every $s \ge 0$ $$\label{eq:GK_formule_magique}
{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right]
={\mathbb E}_{\pi_0}\left[
f(X_0)\left(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot {\partial_\lambda^0}F_\lambda\right)(X_s^0)
\right].$$
Since $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, by Proposition \[prop:EDP\_bien\_posee\], the partial differential equation $$\left\{
\begin{aligned}
\partial_tu(t,x)
&=\Delta u(t,x)
-\nabla V(x)\cdot \nabla u(t,x),
\,t>0,~x\in{\mathbb R},\\
u(0,x)
&=f(x),
\,x\in{\mathbb R}.
\end{aligned}
\right.$$ admits a unique solution $u$ in the sense of Definition \[defi:solution\_EDP\]. Moreover, according to Proposition \[prop:feynman-kac\], $$\forall s\geq 0,\;{\mathrm d}x\mbox{ a.e. },\;u(s,x)={\mathbb E}[f(Y_s^x)]
,$$ where $(Y_t^x)_{t\geq0}$ solves . When $s>0$, from Lemmas \[lem:p\], \[lemderci\] and \[lem:R\_T\_Linfty\] and Assumption [**([[**min Spec**]{}]{})**]{}, one can apply the dominated convergence theorem to differentiate ${\mathbb E}[f(Y_s^x)]$ with respect to $x$, obtaining $\nabla_x{\mathbb E}[f(Y_s^x)]={\mathbb E}\left[R_{Y^x}^T(0,s)\nabla f(Y_s^x)\right]$. Since $u\in\bigcap_{T>0}{\mathbb L}^2\left([0,T],{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})\right)$, ${\mathrm d}s$ a.e., $u(s,.)$ admits a distributional gradient denoted by $\nabla u(s,.)$ and $${\mathrm d}s\mbox{ a.e.},\;{\mathrm d}x\mbox{ a.e.},\;\nabla u(s,x)={\mathbb E}\left[R_{Y^x}^T(0,s)\nabla f(Y_s^x)\right].$$ When $X_0$ is distributed according to $\pi_0$, from reversibility of the dynamics and Lemma \[lem:retournement\_resolvante\], the random vectors $(X_0,X^0_s,R_{X^0}^T(0,s))$ and $(X^0_s,X_0,R_{X^0}(0,s))$ have the same law. Hence $$\label{eq:feynman-kac2}
{\mathrm d}s\mbox{ a.e.},\;\mbox{a.s.},\;{\mathbb E}_{\pi_0}\left[f(X_0)|X_s^0\right]
=u(s,X_s^0)
~\mbox{ and }~
{\mathbb E}_{\pi_0}\left[R_{X^0}(0,s)\nabla f(X_0)|X_s^0\right]
=\nabla u(s,X_s^0).$$ For $s$ such that Equation holds, one deduces that $$\begin{aligned}
{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right]
&={\mathbb E}_{\pi_0}\left[{\mathbb E}_{\pi_0}\left[R_{X^0}(0,s)\nabla f(X_0)|X_s^0\right]\cdot{\partial_\lambda^0}F_\lambda(X_s^0)\right]\\
&={\mathbb E}_{\pi_0}\left[\nabla u(s,X_s^0)\cdot{\partial_\lambda^0}F_\lambda(X_s^0)\right]\\
&=\int_{{{\mathbb R}^d}}\nabla u(s,x)\cdot{\partial_\lambda^0}F_\lambda(x){e^{-V(x)}{\mathrm d}x}\\
&=\int_{{{\mathbb R}^d}}u(s,x)\left(\nabla V(x)\cdot{\partial_\lambda^0}F_\lambda(x)-\nabla\cdot{\partial_\lambda^0}F_\lambda(x)\right){e^{-V(x)}{\mathrm d}x}\\
&={\mathbb E}_{\pi_0}\left[f(X_0)\left(\nabla V(X_s^0)\cdot{\partial_\lambda^0}F_\lambda(X_s^0)-\nabla\cdot{\partial_\lambda^0}F_\lambda(X_s^0)\right)\right],\end{aligned}$$ where we used Lemma \[lem:IPP\_eV\] below with $v(.)=u(s,.)$ which is in ${\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$ for the last but one equality.
\[lem:IPP\_eV\] Let $v$ be a function in ${\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$. Then $$\int_{{{\mathbb R}^d}}\nabla v(x)\cdot{\partial_\lambda^0}F_\lambda(x){e^{-V(x)}{\mathrm d}x}=\int_{{{\mathbb R}^d}}v(x)\left(\nabla V(x)\cdot{\partial_\lambda^0}F_\lambda(x)-\nabla\cdot{\partial_\lambda^0}F_\lambda(x)\right){e^{-V(x)}{\mathrm d}x}.$$
Let $\chi_n(x)=\chi(x/n)$ where $\chi$ is a smooth, $[0,1]$-valued, cutoff function such that $\chi(x)=1$ for $|x|<1$ and $\chi(x)=0$ for $|x|>2$. By integration by parts, one gets $$\begin{aligned}
\int_{{{\mathbb R}^d}}\chi_n(x)\nabla v(x)\cdot{\partial_\lambda^0}F_\lambda(x){e^{-V(x)}{\mathrm d}x}&=\int_{{{\mathbb R}^d}}\chi_n(x)v(x)\left(\nabla V(x)\cdot{\partial_\lambda^0}F_\lambda(x)-\nabla\cdot{\partial_\lambda^0}F_\lambda(x)\right){e^{-V(x)}{\mathrm d}x}\\
&\quad-\int_{{{\mathbb R}^d}}v(x)\nabla\chi_n(x)\cdot{\partial_\lambda^0}F_\lambda(x){e^{-V(x)}{\mathrm d}x}.\end{aligned}$$ The result then follows from Lebesgue’s theorem by taking the limit $n\to\infty$, using the fact that $\nabla V \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, ${\partial_\lambda^0}F_\lambda \in {\mathbb L}^\infty$ and $\nabla\cdot{\partial_\lambda^0}F_\lambda \in {\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ from Assumptions [**([[**Pot**]{}]{})**]{}-$(ii)$ and [**([[**Drift**]{}]{})**]{}-$(i)$-$(ii)$.
By combining and , one gets: for any $t \ge 0$, $${\partial_\lambda^0}{\mathbb E}_{\pi_0}[f(X_t^\lambda)]
=\int_0^t {\mathbb E}_{\pi_0}\left[
f(X_0)\left(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot {\partial_\lambda^0}F_\lambda\right)(X_s^0) \right] {\mathrm d}s,$$ where, we recall, the process $(X_t^0)_{t\geq0}$ satisfies with an initial condition $X_0$ being distributed according to the equilibrium measure $\pi_0$. Taking formally the limit $t\to\infty$, one obtains the classical Green-Kubo formula discussed in the next section.
Infinite time Green-Kubo formula
--------------------------------
\[theo:GK\] Consider the process $(X_t^0)_{t\geq0}$ satisfying with an initial condition $X_0$ being distributed according to the equilibrium measure $\pi_0$. Assume that Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$. Then, for any $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, $\lambda\mapsto \int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda$ is differentiable at $\lambda=0$ with derivative $$\label{eq:GK}
{\partial_\lambda^0}\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda
=\int_0^\infty{\mathbb E}_{\pi_0}\left[
f(X_0)\left(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot {\partial_\lambda^0}F_\lambda\right)(X_s^0)
\right]{\mathrm d}s.$$
Let us recall some results and notation from Section \[sec:inv\_meas\]. The generator of the process $(X_t^0)_{t\geq0}$ is $\mathcal L_0=-\nabla V\cdot\nabla+\Delta$. The generator of the process $(X_t^\lambda)_{t\geq0}$ is $\mathcal
L_\lambda=F_\lambda\cdot\nabla+\Delta=\mathcal L_0+\mathcal T_\lambda$ where $\mathcal T_\lambda
=(F_\lambda+\nabla V)\cdot\nabla$. The domain of the operator $\mathcal L_0$ is $$\mathcal D(\mathcal L_0)
=\left\{ v \in {\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}),\, \mathcal L_0 v \in {\mathbb L}_0^2({e^{-V(x)}{\mathrm d}x})\right\}.$$ For any $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, there exists a unique function $g=- \mathcal L_0^{-1}f$ in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$ such that $$\forall v\in{\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x}),~
\int_{{{\mathbb R}^d}}\nabla g(x)\cdot\nabla v(x){e^{-V(x)}{\mathrm d}x}=\int_{{{\mathbb R}^d}}f(x)v(x){e^{-V(x)}{\mathrm d}x}.$$
Let us start with a lemma which is a consequence of the results of Section \[sec:longtime\] on the long-time behaviour of ${\mathbb E}(f(Y_t^x))$.
\[lem:transformee\_laplace\_L0\] Let us assume that Assumption [**([[**Poinc($\eta$)**]{}]{})**]{} holds for some positive $\eta$. Let us introduce the semigroup $P_t$ associated to the Markovian evolution : for any $f \in {\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$, $$P_t f(x) = {\mathbb E}(f(Y_t^x))$$ where $(Y_t^x)_{t \ge 0}$ satisfies . We then have the following Laplace inversion formula for the operator $\mathcal L_0^{-1}$: for any $f$ in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$, $$\label{eq:poisson}
-\mathcal L_0^{-1}f
=\int_0^\infty P_tf \, {\mathrm d}t.$$
[F]{}rom Proposition \[prop:feynman-kac\], we know that $u(t,x)=P_t f (x) = {\mathbb E}[
f(Y_t^x) ]$ is well defined in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$ and satisfies the partial differential equation in the sense of Definition \[defi:solution\_EDP\]. From Proposition \[prop:tps\_long\], (since $\int_{{\mathbb R}^d} f(x) {e^{-V(x)}{\mathrm d}x}=0$) $$\label{eq:expo_decay}
\forall t \ge 0, \, \left\|P_t f \right\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}
\leq e^{-\eta t} \left\|f \right\|_{{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})}.$$ This shows that $\int_0^\infty P_t f \, {\mathrm d}t$ is well defined in ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$. Moreover, an adaptation with $\lambda={\varphi}=0$ (and thus $C_0=C=0$) of the first energy estimate in the proof of Proposition \[prop:EDP\_bien\_posee\] shows that $$\label{eq:PsfH1}
\int_0^\infty \int_{{{\mathbb R}^d}}|\nabla P_t f(x)|^2 {e^{-V(x)}{\mathrm d}x}\, {\mathrm d}t< \infty.$$ Therefore, $\int_0^\infty P_t f \, {\mathrm d}t \in {\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$.
[From]{} Definition \[defi:solution\_EDP\], for any test function $v \in
{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$, $$\int_{{{\mathbb R}^d}}f(x) v(x) {e^{-V(x)}{\mathrm d}x}= \int_{{{\mathbb R}^d}}P_tf(x) v(x) {e^{-V(x)}{\mathrm d}x}+ \int_0^t
\int_{{{\mathbb R}^d}}\nabla P_s f (x) \cdot \nabla v(x) {e^{-V(x)}{\mathrm d}x}\, {\mathrm d}s.$$ [From]{} Equation , $\lim_{t \to \infty} \int_{{{\mathbb R}^d}}P_t f(x) v(x) {e^{-V(x)}{\mathrm d}x}=0$ and thus, from , for any test function $v \in
{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})\cap{\mathbb H}^1({e^{-V(x)}{\mathrm d}x})$, $$\int_{{{\mathbb R}^d}}f(x) v(x) {e^{-V(x)}{\mathrm d}x}=
\int_{{{\mathbb R}^d}}\nabla \left( \int_0^\infty P_t f (x) {\mathrm d}t \right) \cdot \nabla v(x) {e^{-V(x)}{\mathrm d}x}.$$ This concludes the proof.
We recall that $\Pi_0$ the orthogonal projection from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ onto ${\mathbb L}^2_0({e^{-V(x)}{\mathrm d}x})$ (see ). We can now give a different expression for the right-hand side of . [From]{} Lemma \[lem:IPP\_eV\] applied to the constant $1$, one has $\int_{{{\mathbb R}^d}}(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(x){e^{-V(x)}{\mathrm d}x}=0$. Then, using successively this equality, the self-adjointness of $P_t$ in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ (which is a direct consequence of ), Equation and Lemma \[lem:IPP\_eV\], one has, $$\begin{aligned}
&\int_0^\infty {\mathbb E}_{\pi_0}\left[f(X_0)(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(X^0_t)\right]\, {\mathrm d}t\\
&=\int_0^\infty\int_{{{\mathbb R}^d}}{\mathbb E}\left[f(x)(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(Y_t^x)\right]{e^{-V(x)}{\mathrm d}x}\,{\mathrm d}t\\
&=\int_0^\infty\int_{{{\mathbb R}^d}}{\mathbb E}\left[\Pi_0f(x)(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(Y_t^x)\right]{e^{-V(x)}{\mathrm d}x}\, {\mathrm d}t\\
&=\int_0^\infty\int_{{{\mathbb R}^d}}\Pi_0f(x)P_t(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(x){e^{-V(x)}{\mathrm d}x}\, {\mathrm d}t\\
&=\int_0^\infty\int_{{{\mathbb R}^d}}(P_t\Pi_0f(x))(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(x){e^{-V(x)}{\mathrm d}x}\, {\mathrm d}t\\
&=-\int_{{{\mathbb R}^d}}(\mathcal L_0^{-1}\Pi_0f)(x)(\nabla V\cdot{\partial_\lambda^0}F_\lambda-\nabla\cdot{\partial_\lambda^0}F_\lambda)(x){e^{-V(x)}{\mathrm d}x}\\
&=-\int_{{{\mathbb R}^d}}{{\partial_\lambda^0}\mathcal T_\lambda}[(\mathcal L_0^{-1}\Pi_0f)](x){e^{-V(x)}{\mathrm d}x},\end{aligned}$$ where ${{\partial_\lambda^0}\mathcal T_\lambda}$ stands for the operator ${\partial_\lambda^0}F_\lambda\cdot\nabla$ (consistently with the definition of $\mathcal T_\lambda$). As a consequence, proving Equation boils down to proving: for any $f\in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$, $$\label{eq:GK_revisite}
{\partial_\lambda^0}\int_{{{\mathbb R}^d}}\Pi_0f(x){\mathrm d}\pi_\lambda(x)
=-\int_{{{\mathbb R}^d}}{{\partial_\lambda^0}\mathcal T_\lambda}[(\mathcal L_0^{-1}\Pi_0f)](x){e^{-V(x)}{\mathrm d}x}.$$
We are now in position to complete the proof of Theorem \[theo:GK\].
Recall that for $\lambda$ small enough, the operator $I+(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*$ is invertible from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to itself with bounded inverse (see Lemma \[lem:pil\]). For such a small $\lambda$, one has the equality $$\label{eq:(I+T)^-1}
(I+(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*)^{-1}
=I
-(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*
+\mathcal R_\lambda,$$ where (by ) the remainder $\mathcal R_\lambda
=(I+(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*)^{-1}
((\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*)^2$ has a norm from ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ to itself of order $\mathcal O(\lambda^2)$. Thus, by the analytical formula for $\pi_\lambda$ obtained in Lemma \[lem:pil\], $$\begin{aligned}
\int_{{{\mathbb R}^d}}f(x){\mathrm d}\pi_\lambda(x)
-\int_{{{\mathbb R}^d}}f(x){\mathrm d}\pi_0(x)
&=-\int_{{{\mathbb R}^d}}[ (\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)^*\mathbf1](x)f(x){e^{-V(x)}{\mathrm d}x}+\int_{{{\mathbb R}^d}}[\mathcal R_\lambda \mathbf1](x)f(x){\mathrm d}\pi_0\notag\\
&=-\int_{{{\mathbb R}^d}}[(\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0)f](x){e^{-V(x)}{\mathrm d}x}+\mathcal O(\lambda^2).\label{difespll0}\end{aligned}$$ Since, according to [**([[**Drift**]{}]{})**]{}-$(i)$, $\frac{F_\lambda+\nabla V}{\lambda}$ is bounded by $C$ for $\lambda\in(0,\lambda_0]$, one has, by Lebesgue’s theorem, $$\lim_{\lambda \to 0} \frac1\lambda\int_{{{\mathbb R}^d}}\mathcal T_\lambda\mathcal L_0^{-1}\Pi_0f(x){e^{-V(x)}{\mathrm d}x}=\int_{{{\mathbb R}^d}}({\partial_\lambda^0}\mathcal T_\lambda)\mathcal L_0^{-1}\Pi_0f(x){e^{-V(x)}{\mathrm d}x}.$$ Dividing by $\lambda$ and taking the limit $\lambda\to 0$, one concludes that $\lambda\mapsto \int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda$ is differentiable at $\lambda=0$ and holds.
Combining the previous result with , we obtain the following corollary.
\[cor:GK\] Let $f \in{\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$ be a Lipschitz function. Also assume that the Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} (for some positive $\eta$) and [**([[**min Spec**]{}]{})**]{} are satisfied. Then, one has $${\partial_\lambda^0}\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda
=\int_0^\infty{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right]{\mathrm d}s.$$
Long-time convergence of the estimators and {#sec:main_result}
=============================================
Statement of the main result
----------------------------
Let us study the long-time behavior of the two estimators and .
\[theo:interversion\] Let $f:{\mathbb R}^d\to{\mathbb R}$ be a ${\mathcal C}^1$ function such that $\nabla f$ is bounded.
- Assume the existence of $\eta>0$ such that either $V$ is $\eta$-convex or Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} and [ **([[**Spec($\eta$)**]{}]{})**]{} hold. Then $\lambda\mapsto\frac{1}{t}\int_0^t f(X^\lambda_s){\mathrm d}s$ is differentiable at $\lambda=0$ with derivative $\frac{1}{t}\int_0^t \nabla f(X^0_s).T_s {\mathrm d}s$ and $$\label{eq:cv_estim2}
\lim_{t\to\infty}{\partial_\lambda^0}\left(\frac1t\int_0^tf(X_s^\lambda){\mathrm d}s\right)
={\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right)
\mbox{ a.s.}.$$
- Assume either that $V$ is $\eta$-convex for a positive constant $\eta>0$ and ${\mathbb E}|X_0|<+\infty$, or that there exist $\eta>0$ and $p\in (1, \infty]$ such that [**([[**Poinc($\eta$)**]{}]{})**]{} holds, $X_0$ is distributed according to a measure having a density in ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$ with respect to ${e^{-V(x)}{\mathrm d}x}$ and [ **([[**Spec($\rho$)**]{}]{})**]{} holds for some $\rho<\eta(p-1)/p$ (with the convention $\eta(\infty-1)/\infty=\eta$). Then $\forall \lambda\in[0,\lambda_0]$, $\forall t\geq 0$, ${\mathbb E}|f(X^\lambda_t)|<+\infty$, $\lambda\mapsto{\mathbb E}[f(X^\lambda_t)]$ is differentiable at $\lambda=0$ with derivative ${\partial_\lambda^0}{\mathbb E}\left[f(X_t^\lambda)\right]={\mathbb E}[\nabla f(X^0_t).T_t]$ and $$\label{eq:cv_estim1}
\lim_{t\to\infty}{\partial_\lambda^0}{\mathbb E}\left[f(X_t^\lambda)\right]
={\partial_\lambda^0}\left(\int_{{{\mathbb R}^d}}f{\mathrm d}\pi_\lambda\right).$$
\[rem:mainres\] When $\pi_0$ is assumed to satisfy a logarithmic Sobolev inequality with constant $\eta$, which is stronger than the Poincaré inequality [**([[**Poinc($\eta$)**]{}]{})**]{}, then the second statement still holds as soon as $X_0$ is distributed according to a measure having a density in ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$ with respect to ${e^{-V(x)}{\mathrm d}x}$ for some $p>1$ and [ **([[**Spec($\rho$)**]{}]{})**]{} holds for some $\rho < \eta$, because of the hypercontractivity property of the semi-group associated with ensured by the Gross theorem.
Long-time behaviour of $(X^0_t,T_t)_{t\geq 0}$
----------------------------------------------
To prove Theorem \[theo:interversion\], one first needs to know the long-time limit of the trajectory and its tangent vector. We more generally consider $(X^0_t,T^0_t)_{t \ge0}$ solving $$\label{edsjointe}
\left\{
\begin{aligned}
{\mathrm d}X_t^0&=-\nabla V(X_t^0){\mathrm d}t+\sqrt2{\mathrm d}W_t\\
{\mathrm d}T^0_t
&=\left({\partial_\lambda^0}F_\lambda(X_t^0)-\nabla^2V(X_t^0) T^0_t\right){\mathrm d}t
\end{aligned}
\right.$$ with $(X^0_0,T^0_0)$ any initial condition independent from the Brownian motion $(W_t)_{t\geq 0}$. To write conveniently the long-time limit of $(X^0_t,T^0_t)$, we will run time backward and use Lemma \[lem:retournement\_resolvante\] about the link between the forward resolvent and its backward counterpart.
\[lem:conv\_loi\]
Under Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} (for some positive $\eta$), [**([[**min Spec**]{}]{})**]{} and [**([[**Conv**]{}]{})**]{}, the process $(X_t^0,T^0_t)_{t\geq0}$ converges in law as $t$ goes to infinity to the couple $$\left(Y_0,\int_0^\infty R_Y^T(0,t){\partial_\lambda^0}F_\lambda(Y_t){\mathrm d}t\right),$$ where $(Y_t)_{t\geq0}$ follows the dynamics with with $Y_0$ distributed according to ${e^{-V(x)}{\mathrm d}x}$. Moreover, the law ${\cal
V}$ of this couple is invariant by the dynamics and ergodic for this dynamics: for any test function $\varphi: {{{\mathbb R}^d}}\times{{{\mathbb R}^d}}\to {\mathbb R}$ in ${\mathbb L}^1({\cal V})$, for ${\cal V}$-a.e. deterministic initial condition $(X^0_0,T^0_0)$, $$\lim_{t \to \infty} \frac 1 t \int_0^t \varphi(X^0_s,T^0_s) \, {\mathrm d}s = \int_{{{{\mathbb R}^d}}\times
{{{\mathbb R}^d}}} \varphi(x,\tau) \, {\mathrm d}{\cal V}(x,\tau)\;a.s..$$
The integral $\int_0^\infty R_Y^T(0,t){\partial_\lambda^0}F_\lambda(Y_t){\mathrm d}t$ is almost surely well defined, from Lemma \[lem:convergence\_R\] and the boundedness of ${\partial_\lambda^0}F_\lambda$. To prove Lemma \[lem:conv\_loi\], we are going to use a time reversal argument.
For $t_0 >0$, we construct a coupling of the trajectory $(X^0_t)_{t\geq t_0}$ with another process $(\chi_t^{t_0})_{t\geq t_0}$ following the dynamics , but being at equilibrium. Denote by $q_{t_0}$ the density of the distribution of $X_{t_0}^0$ (which exists by Lemma \[lem:p\]), and define $\rho_{t_0}=\frac{q_{t_0}\wedge e^{-V}}{q_{t_0}}$. Let $U$ and $\zeta_{t_0}$ be mutually independent random variables which are independent of $X_0$ and of the Brownian motion $(W_t)_{t\geq0}$ driving $(X^0_t)_{t\geq0}$, such that $U$ is uniformly distributed over $[0,1]$ and, when $q_{t_0}\neq
e^{-V}$, $\zeta_{t_0}$ is distributed according to $C(e^{-V(x)}-q_{t_0}(x))^+{\mathrm d}x$, $C$ being a normalization constant ($\zeta_{t_0}$ does not need to be defined when $q_{t_0}=e^{-V}$). We define the position of the process $(\chi^{t_0}_t)_{t\geq{t_0}}$ at time ${t_0}$ by $\chi_{t_0}^{t_0}=X_{t_0}^0\mathbf1_{U\leq\rho_{t_0}(X_{t_0}^0)}+\zeta_{t_0}\mathbf1_{U>\rho_{t_0}(X_{t_0}^0)}$, which is distributed according to $\pi_0$. One has ${\mathbb P}(\chi_{t_0}^{t_0}\neq
X_{t_0}^0)=\frac12\|q_{t_0}(x)-e^{-V(x)}\|_{{\mathbb L}^1({\mathrm d}x)}$. For $t>{t_0}$, let $(\chi_t^{t_0})_{t\geq{t_0}}$ evolve according to the dynamics with Brownian motion $(W_t)_{t\geq0}$. Notice that $(\chi_{t+{t_0}}^{t_0})_{t \ge 0}$ has the same law as the process at equilibrium $(Y_t)_{t \ge
0}$ introduced in the statement of Lemma \[lem:conv\_loi\]. Moreover, $(\chi_t^{t_0})_{t \ge t_0}$ is such that $${\mathbb P}(\forall t\geq{t_0}, \chi_t^{t_0}=X_t^0)=1-\frac12\|q_{t_0}(x)-e^{-V(x)}\|_{{\mathbb L}^1({\mathrm d}x)}.$$
[From]{} an easy adaptation of Proposition \[prop:tangent\], one has on the one hand $$(X_t^0,T^0_t)=\left(X_t^0,R_{X^0}(0,t)T_0^0+\int_0^tR_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right).$$ On the other hand, for $0 < {t_0}\leq t$, we have the equalities, by using successively the time translation $s\to s-{t_0}$, the change of variable $u=t-{t_0}-s$, Lemma \[lem:retournement\_resolvante\] (using the notation, for $u
\in [0, t-{t_0}]$, $Z_u = Y_{t-{t_0}-u}$) and the time reversibility of the dynamics : $$\begin{aligned}
\left(
\chi_t^{t_0},
\int_{t_0}^tR_{\chi^{t_0}}(s,t){\partial_\lambda^0}F_\lambda(\chi_s^{t_0}){\mathrm d}s
\right)
&\stackrel{\mathcal D}=\left(
Y_{t-{t_0}},
\int_0^{t-{t_0}}R_Y(s,t-{t_0}){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s
\right)\nonumber\\
&=\left(
Y_{t-{t_0}},
\int_0^{t-{t_0}}R_Y(t-{t_0}-u,t-{t_0}){\partial_\lambda^0}F_\lambda(Y_{t-{t_0}-u}){\mathrm d}u
\right)\nonumber\\
&=
\left(Z_0,\int_0^{t-{t_0}}R_Z^T(0,u){\partial_\lambda^0}F_\lambda(Z_u){\mathrm d}u\right)\nonumber\\
&\stackrel{\mathcal D}=
\left(Y_0,\int_0^{t-{t_0}}R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right),\end{aligned}$$ where $\stackrel{\mathcal D}=$ stands for the equality in distribution. As a consequence, for any bounded Lipschitz function ${\varphi}:{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}$ (with Lipschitz constant ${\rm Lip}({\varphi})$), for $t \ge {t_0} > 0$ $$\begin{aligned}
&\left|{\mathbb E}\left[{\varphi}\left(Y_0,\int_0^{\infty}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)-{\varphi}(X_t^0,T^0_t)\right]\right|\nonumber\\
&\leq \left|{\mathbb E}\left[{\varphi}\left(Y_0,\int_0^{\infty}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)-{\varphi}\left(Y_0,\int_0^{t-{t_0}}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)\right]\right|\nonumber\\
&+\bigg|{\mathbb E}\bigg[{\varphi}\left(\chi_t^{t_0},\int_{t_0}^tR_{\chi^{t_0}}(s,t){\partial_\lambda^0}F_\lambda(\chi_s^{t_0}){\mathrm d}s\right)\nonumber\\
&\phantom{+\bigg|{\mathbb E}\bigg[}-{\varphi}\left(\chi_t^{t_0},R_{X^0}(0,t)T_0^0+\int_0^{t_0} R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s
+\int_{t_0}^tR_{\chi^{t_0}}(s,t){\partial_\lambda^0}F_\lambda(\chi_s^{t_0}){\mathrm d}s\right)\bigg]\bigg|\nonumber\\&+\left|{\mathbb E}\left[{\varphi}\left(\chi_t^{t_0},R_{X^0}(0,t)T_0^0+\int_0^{t_0} R_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s
+\int_{t_0}^tR_{\chi^{t_0}}(s,t){\partial_\lambda^0}F_\lambda(\chi_s^{t_0}){\mathrm d}s\right)-{\varphi}(X_t^0,T^0_t)\right]\right|\nonumber
\\
&\leq \left|{\mathbb E}\left[{\varphi}\left(Y_0,\int_0^{\infty}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)-{\varphi}\left(Y_0,\int_0^{t-{t_0}}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)\right]\right|\nonumber\\
&+{\mathbb E}\left[
\left(2\|{\varphi}\|_{{\mathbb L}^\infty(\Omega)}\right)\wedge\left( {\rm Lip}({\varphi})
\left|R_{X^0}({t_0},t)\left(R_{X^0}(0,{t_0})T_0^0+\int_0^{t_0} R_{X^0}(s,{t_0}){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right)\right|\right)
\right]\nonumber\\
&+2\|{\varphi}\|_{\infty}{\mathbb P}(X_{t_0}^0\neq \chi_{t_0}^{t_0}).
\label{eq:retourn_temps}\end{aligned}$$ Notice that we used the semi-group property of $R_{X^0}$ to obtain the last but one inequality. The first term in the right-hand side converges to $0$ as $t\to\infty$ by Lebesgue’s theorem. A direct adaptation of Lemma \[lem:convergence\_R\] shows that $R_{X^0}({t_0},t)$ goes to $0$ as $t$ goes to infinity, yielding from Lebesgue’s theorem that the second term in the right-hand side of goes to $0$ as $t$ goes to infinity. The third term in the right-hand side of can be rewritten as $2\|{\varphi}\|_{\infty}{\mathbb P}(X_{t_0}^0\neq \chi_{t_0}^{t_0})
=\|{\varphi}\|_{{\mathbb L}^\infty(\Omega)}\|e^{-V(x)}-p_{t_0}(x)\|_{{\mathbb L}^1({\mathrm d}x)}$ and thus goes to $0$ as ${t_0}$ goes to infinity, by Corollary \[cor:CV\_L1\]. Letting $t\to\infty$ and then $t_0\to\infty$ in , we conclude that the couple $(X_t^0,T^0_t)$ converges in law to $\left(Y_0,\int_0^{\infty}
R_Y^T(0,s){\partial_\lambda^0}F_\lambda(Y_s){\mathrm d}s\right)$ with law ${\cal V}$.
To check that ${\cal V}$ is invariant by the dynamics , we denote by $({\cal P}_{s})_{s\geq 0}$ the Markov semi-group associated with this dynamics. One has ${\mathbb E}[{\cal P}_s\varphi(X^0_t,T^0_t)]={\mathbb E}[\varphi(X^0_{t+s},T^0_{t+s})]$ where the right-hand side converges to $\int_{{\mathbb R}^d\times{\mathbb R}^d}\varphi(x,\tau){\mathrm d}{\cal V}(x,\tau)$ as $t\to\infty$ and ${\cal P}_s\varphi(x,\tau)={\mathbb E}[\varphi(Y^x_s,T^{x,\tau}_s)]$ with $$\left\{ \begin{aligned}
Y^x_t&=x-\int_0^t\nabla V(Y^x_s){\mathrm d}s+\sqrt2W_t \, ,\\
T^{x,\tau}_t
&=\tau+\int_0^t\left({\partial_\lambda^0}F_\lambda(Y^x_s)-\nabla^2V(Y^x_s)
T^{x,\tau}_s\right){\mathrm d}s \, .\end{aligned}
\right.$$ The continuity of $x\mapsto (Y^x_t)_{t\geq 0}$ for the topology of local uniform convergence on ${\mathcal C}({\mathbb R}_+,{\mathbb R}^d)$ together with the continuity of $\nabla^2V$ implies the continuity of $x\mapsto
(R_{Y^x}(s,t))_{s,t\geq 0}$ for the topology of local uniform convergence on ${\mathcal C}({\mathbb R}_+\times{\mathbb R}_+,
{\mathbb R}^{d\times d})$. With the continuity of ${\partial_\lambda^0}F_\lambda$, one deduces the continuity of $(x,\tau)\mapsto
T^{x,\tau}_s=R_{Y^x}(0,s)\tau+\int_0^sR_{Y^x}(r,s){\partial_\lambda^0}F_\lambda(Y^x_r){\mathrm d}r$. Hence, by Lebesgue’s theorem, ${\cal
P}_s\varphi(x,\tau)$ is continuous and bounded and ${\mathbb E}[{\cal
P}_s\varphi(X^0_t,T^0_t)]$ converges to $\int_{{\mathbb R}^d\times{\mathbb R}^d}{\cal
P}_s\varphi(x,\tau){\mathrm d}{\cal V}(x,\tau)$ as $t\to\infty$. Therefore $\int_{{\mathbb R}^d\times{\mathbb R}^d}{\cal P}_s\varphi(x,\tau){\mathrm d}{\cal
V}(x,\tau)=\int_{{\mathbb R}^d\times{\mathbb R}^d}\varphi(x,\tau){\mathrm d}{\cal
V}(x,\tau)$ and the probability measure ${\cal V}$ is invariant.
Since ${\cal V}$ is the unique invariant probability measure for the SDE , this measure is ergodic (see for example [@rey-bellet-06 Theorem 3.8 and Equation (52)]).
Let us deduce from the previous results the limit of $\frac{1}{t}\int_0^t\varphi(X^0_s) \cdot T^0_s{\mathrm d}s$ where $\varphi:{\mathbb R}^d\to{\mathbb R}^d$ is measurable and bounded.
\[lem:ergconv\]Assume the existence of $\eta>0$ such that either $V$ is $\eta$-convex or Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} and [ **([[**Spec($\eta$)**]{}]{})**]{} hold. Then $\int_{{\mathbb R}^d\times{\mathbb R}^d}|\tau|{\mathrm d}{\cal V}(x,\tau)<\infty$ (where the probability distribution ${\cal V}$ has been introduced in Lemma \[lem:conv\_loi\]) and for any function $\varphi:{\mathbb R}^d\to{\mathbb R}^d$ measurable and bounded, $\frac{1}{t}\int_0^t\varphi(X^0_s)\cdot T^0_s{\mathrm d}s$ converges a.s. to $\int_{{\mathbb R}^d\times{\mathbb R}^d}\varphi(x)\cdot\tau{\mathrm d}{\cal V}(x,\tau)$ as $t\to\infty$ whatever the choice of the initial condition $(X^0_0,T^0_0)$ independent of the Brownian motion $(W_t)_{t\geq 0}$.
Notice that if $V$ is $\eta$-convex, Assumptions [**([[**min Spec**]{}]{})**]{}, [**([[**Conv**]{}]{})**]{} and [**([[**Poinc($\eta$)**]{}]{})**]{} hold. In addition, Assumption [ **([[**Spec($\eta$)**]{}]{})**]{} implies Assumptions [**([[**min Spec**]{}]{})**]{} and [**([[**Conv**]{}]{})**]{}. Therefore, the conclusion of Lemma \[lem:conv\_loi\] holds under the two classes of hypotheses considered. Since ${\partial_\lambda^0}F_\lambda$ is bounded, one has $$\int_{{\mathbb R}^d\times{\mathbb R}^d}|\tau|{\mathrm d}{\cal V}(x,\tau)\leq \int_0^\infty {\mathbb E}\left[|R_Y^T(0,t)||{\partial_\lambda^0}F_\lambda(Y_t)|\right]{\mathrm d}t\leq C\int_0^\infty {\mathbb E}\left[|R_Y(0,t)|\right]{\mathrm d}t ,$$ where the right-hand side is finite by when $V$ is $\eta$-convex and by Proposition \[cor:momr\] otherwise.
In case the law of $(X_0^0,T^0_0)$ is absolutely continuous with respect to ${\cal V}$, the result of Lemma \[lem:ergconv\] is then a direct consequence of the ergodic property of the process $(X^0_t,T^0_t)_{t \ge 0}$ stated in Lemma \[lem:conv\_loi\].
To extend this result to more general initial conditions, we proceed as follows. By Lemma \[lem:densite\_X\_t\], the law of $X^0_1$ is absolutely continuous with respect to $\pi_0$ which is the marginal law of the $d$ first coordinates for the ergodic measure ${\cal V}$. Let $d{\cal V}_{T|X=x}(\tau)$ denote a regular conditional probability distribution of the $d$ last coordinates given the $d$ first ones under ${\cal V}$ and $\tilde{T}^0_1$ be a random vector independent of $(W_t-W_1)_{t\geq 1}$ with conditional law given $X^0_1$ equal to $d{\cal V}_{T|X=X^0_1}(\tau)$. Let for $t\geq 1$, $\tilde{T}^0_t=R_{X^0}(1,t)\tilde{T}^0_1+\int_1^tR_{X^0}(s,t){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s$. Then ${\mathrm d}\tilde{T}^0_t=\left({\partial_\lambda^0}F_\lambda(X_t^0)-\nabla^2V(X_t^0) \tilde{T}^0_t\right){\mathrm d}t$ so that $(X^0_t,\tilde{T}^0_t)_{t\geq 1}$ solves starting from $(X^0_1,\tilde{T}^0_1)$ the law of which is absolutely continuous with respect to the measure ${\cal V}$ ergodic for this stochastic differential equation (see Lemma \[lem:conv\_loi\]). As a consequence $\frac{1}{t}\int_1^t\varphi(X^0_s)\cdot\tilde{T}^0_s{\mathrm d}s$ converges a.s. to $\int_{{\mathbb R}^d\times{\mathbb R}^d}\varphi(x)\cdot\tau{\mathrm d}{\cal V}(x,\tau)$ as $t\to\infty$. Now, by an adaptation of Proposition \[prop:tangent\], one can check that for $t \ge 1$, $\tilde{T}^0_t-T^0_t=R_{X^0}(1,t)(\tilde{T}^0_1-T^0_1)=R_{X^0}(0,t)R_{X^0}(1,0)(\tilde{T}^0_1-T^0_1)$ so that $$\begin{aligned}
\frac{1}{t}\int_0^t\varphi(X^0_s)\cdot T^0_s{\mathrm d}s-\frac{1}{t}\int_1^t\varphi(X^0_s)\cdot\tilde{T}^0_s{\mathrm d}s&=\frac{1}{t}\int_0^1\varphi(X^0_s)\cdot T^0_s{\mathrm d}s\\
&+\frac{1}{t}\int_1^t\varphi(X^0_s)\cdot R_{X^0}(0,s){\mathrm d}sR_{X^0}(1,0)(T^0_1-\tilde{T}^0_1).\end{aligned}$$ The proof is completed by noticing that this quantity converges a.s. to $0$ (for the second term in the right-hand side, this is a consequence of the boundedness of $\varphi$ and of the almost sure estimate ).
Proof of Theorem \[theo:interversion\]
--------------------------------------
We are now in position to prove Theorem \[theo:interversion\].
Let us start by a preliminary result concerning the integrability of the random variable $\int_0^\infty \left|\nabla f(X_0) \right|
\left| R_{X^0}^T(0,t){\partial_\lambda^0}F_\lambda(X_t^0) \right| {\mathrm d}t$. By the boundedness of $\nabla f$ and ${\partial_\lambda^0}F_\lambda$ $$\begin{aligned}
{\mathbb E}_{\pi_0}\left[ \int_0^\infty \left|\nabla f(X_0) \right|
\left| R_{X^0}^T(0,t){\partial_\lambda^0}F_\lambda(X_t^0) \right| {\mathrm d}t\right]
\leq C\int_0^{+\infty}{\mathbb E}_{\pi_0}\left[|R_{X^0}(0,t)|\right] {\mathrm d}t,\end{aligned}$$ where, we recall, the subscript $\pi_0$ in ${\mathbb E}_{\pi_0}$ indicates that $X_0$ is distributed according to $\pi_0$. If $V$ is $\eta$-convex, using , almost surely, $$|R_{X^0}(0,t)| \le e^{-\eta t}.$$ If Assumptions [**([[**Poinc($\eta$)**]{}]{})**]{} and [ **([[**Spec($\eta$)**]{}]{})**]{} hold for some positive $\eta$ (notice that [**([[**Spec($\rho$)**]{}]{})**]{} for $\rho<\frac{\eta(p-1)}{p}$ with $p\in(1,+\infty]$ implies [**([[**Spec($\eta$)**]{}]{})**]{}), then, by Proposition \[cor:momr\] , $${\mathbb E}_{\pi_0}\left[|R_{X^0}(0,t)|\right]
\leq Ce^{-t/C}$$ for some positive constant $C$. Hence, in all cases, $$\label{eq:justif_fubini}
{\mathbb E}_{\pi_0}\left[ \int_0^\infty \left|\nabla f(X_0) \right|
\left| R_{X^0}^T(0,t){\partial_\lambda^0}F_\lambda(X_t^0) \right| {\mathrm d}t\right]<+\infty.$$
Let us now prove the first statement of Theorem \[theo:interversion\]. By , the boundedness of $\nabla f$ and Proposition \[proptt\], Lebesgue’s theorem implies that $\lambda\mapsto\frac{1}{t}\int_0^t f(X^\lambda_s){\mathrm d}s$ is differentiable at $\lambda=0$ with derivative $\frac{1}{t}\int_0^t
\nabla f(X^0_s)\cdot T_s {\mathrm d}s$. By Lemma \[lem:ergconv\], $\frac{1}{t}\int_0^t \nabla f(X^0_s)\cdot T_s {\mathrm d}s$ converges a.s. to $\int_{{\mathbb R}^d\times{\mathbb R}^d}\nabla f(x)\cdot \tau{\mathrm d}{\cal V}(x,\tau)={\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot\int_0^\infty R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right]$. The proof of is then completed by the following computations: $$\begin{aligned}
{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot\int_0^\infty R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right]
=&\int_0^\infty{\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0)\right]{\mathrm d}s\nonumber\\
=&{\partial_\lambda^0}\int_{{{\mathbb R}^d}}f(x){\mathrm d}\pi_\lambda(x)\label{eq:fub_GK}\end{aligned}$$ where we used Fubini’s theorem and for the first equality and Corollary \[cor:GK\] for the second one.
Let us finally deal with the second statement of Theorem \[theo:interversion\]. By , the boundedness of $\nabla f$ and Proposition \[proptt\], it is enough to check that ${\mathbb E}|f(X^0_t)|<+\infty$ to deduce that $\forall \lambda\in[0,\lambda_0]$, ${\mathbb E}|f(X^\lambda_t)|<+\infty$ and $\lambda\mapsto{\mathbb E}[f(X^\lambda_t)]$ is differentiable at $\lambda=0$ with derivative ${\partial_\lambda^0}{\mathbb E}\left[f(X_t^\lambda)\right]={\mathbb E}[\nabla
f(X^0_t)\cdot T_t]$. When ${e^{-V(x)}{\mathrm d}x}$ satisfies a Poincaré inequality, then according to [@bobkovledoux] and the references therein, since $f$ is a Lipschitz function, there exists a positive $\varepsilon$ such that $\int_{{\mathbb R}^d}e^{\varepsilon
|f|(x)}{e^{-V(x)}{\mathrm d}x}<+\infty$. Therefore, when the law $\mu_0$ of $X_0$ has a density with respect to ${e^{-V(x)}{\mathrm d}x}$ in ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$, by Lemma \[lem:densite\_X\_t\], $$\sup_{t\geq 0}{\mathbb E}|f(X^0_t)|\leq \|f\|_{{\mathbb L}^{\frac{p}{p-1}}({e^{-V(x)}{\mathrm d}x})}\left\|\frac{{\mathrm d}\mu_0}{{e^{-V(x)}{\mathrm d}x}}\right\|_{{\mathbb L}^p({e^{-V(x)}{\mathrm d}x})}<+\infty.$$
When $V$ is $\eta$-convex, then computing $|Y^x_t|^2$ by Itō’s formula, remarking that $$-2\nabla V(Y^x_t)\cdot Y^x_t=-2(\nabla
V(Y^x_t)-\nabla V(0))\cdot Y^x_t-2\nabla V(0)\cdot Y^x_t\leq
-\eta|Y^x_t|^2+\frac{|\nabla V(0)|^2}{\eta},$$ applying a localization procedure to get rid of the expectation of the stochastic integral, one obtains $\forall t\geq 0$, ${\mathbb E}[|Y^x_t|^2]\leq e^{-\eta
t}|x|^2+\frac{1-e^{-\eta t}}{\eta}\left( \frac{|\nabla
V(0)|^2}{\eta} + 2d\right)$. Hence, when the initial random variable $X_0$ with law $\mu_0$ is integrable, $$\begin{aligned}
{\mathbb E}|X^0_t|&\leq \int_{{\mathbb R}^d}{\mathbb E}|Y^x_t|{\mathrm d}\mu_0(x)\leq
\int_{{\mathbb R}^d}\sqrt{{\mathbb E}[|Y^x_t|^2]}{\mathrm d}\mu_0(x)\\
&\leq{\mathbb E}\left[\sqrt{e^{-\eta t}|X_0|^2+\frac{1-e^{-\eta t}}{\eta}\left( \frac{|\nabla
V(0)|^2}{\eta} + 2d\right) }\right] <+\infty,\end{aligned}$$ and ${\mathbb E}|f(X^0_t)|<+\infty$ by the Lipschitz continuity of $f$.
Notice that if $V$ is $\eta$-convex, Assumptions [**([[**min Spec**]{}]{})**]{}, [**([[**Conv**]{}]{})**]{} and [**([[**Poinc($\eta$)**]{}]{})**]{} hold. Moreover, Assumption [ **([[**Spec($\rho$)**]{}]{})**]{} implies Assumptions [**([[**min Spec**]{}]{})**]{} and [**([[**Conv**]{}]{})**]{}. Therefore, the conclusion of Lemma \[lem:conv\_loi\] holds under the two classes of hypotheses considered. The function $(x,\tau) \mapsto \nabla f(x)
\cdot \tau$ is continuous and the family $(\nabla f(X_t^0)\cdot
T_t)_{t \ge 0}$ is uniformly integrable by Proposition \[prop:varalconv\] when $V$ is $\eta$-convex and since $\sup_{t \ge 0} {\mathbb E}\left( \left|\nabla f(X_t^0)\cdot
T_t\right|^{\frac{\eta(p-1)}{\rho p}} \right) < \infty $, by Proposition \[prop:variance\], in the second framework. Therefore the convergence in distribution in Lemma \[lem:conv\_loi\] yields $$\lim_{t \to \infty} {\mathbb E}\left( \nabla f(X_t^0)\cdot T_t\right) = {\mathbb E}_{\pi_0}\left[\nabla f(X_0)\cdot\int_0^\infty R_{X^0}^T(0,s){\partial_\lambda^0}F_\lambda(X_s^0){\mathrm d}s\right],$$ where the right-hand side is equal to ${\partial_\lambda^0}\int_{{{\mathbb R}^d}}f(x){\mathrm d}\pi_\lambda(x)$ according to .
Numerical illustrations {#sect:numeric}
=======================
In this section, we illustrate through various numerical experiments the theoretical results obtained above. In Section \[sec:1d\], we study numerically on a one-dimensional toy model the integrability of the tangent vector $T_t$ and the sharpness of the integrability exponent obtained in Proposition \[prop:variance\]. In Section \[sec:WA\], we illustrate the interest of the estimator on a more realistic test case proposed in [@warren-allen-12]. Finally, we investigate in Section \[sec:roland\] on the one-dimensional toy model a variance reduction method for the estimator .
A one-dimensional toy model {#sec:1d}
---------------------------
In this section, we would like to study on a simple test case the integrability of the tangent vector $T_t$, and to compare the theoretical bounds obtained in Proposition \[prop:variance\], with a numerical estimation of the integrability exponent. Let us consider the potential $$\forall x \in {\mathbb R}, \, V_\lambda(x)
=x^4-\frac c2x^2+\lambda x,$$ where $c$ is some fixed constant, and $\lambda \in {\mathbb R}$ is the parameter. For $\lambda=0$, $V_0$ has curvature $-c$ at the origin, and for $c>0$, $V_0$ is a double-well potential, with wells located at $\pm{\sqrt c}/2$ and separated by a barrier with height $c^2/16$. In particular, as $c$ gets larger, the dynamics of $(X^0_t)_{t \ge 0}$ becomes more and more metastable.
Let us start with some explicit computation on $T_t$. When $\lambda\geq0$, the perturbative force pushes the system to the left. Therefore, one expects the tangent vector $T_t$ to be negative in the mean. In fact, one can prove that $T_t$ is in that case almost surely negative for $t>0$. Indeed, $(T_t)_{t\geq0}$ satisfies the equation $$\left\{
\begin{aligned}
\partial_t T_t
&= -1
+(c-12(X_t^0)^2)T_t,\\
T_0&=0,
\end{aligned}
\right.$$ which can be solved explicitly, since in dimension 1, the resolvent $R_{X^0}(s,t)$ is given by the exponential $R_{X^0}(s,t)=\exp\left(c(t-s)-12\int_s^t(X_u^0)^2{\mathrm d}u\right)$. Equation then becomes $$T_t
=-\int_0^t\exp\left(c(t-s)-12\int_s^t(X_u^0)^2{\mathrm d}u\right){\mathrm d}s
<0.$$
Concerning the upper bound on the integrability exponent obtained in Proposition \[prop:variance\], if the initial condition $X_0$ has a bounded density, then the tangent vector is bounded in ${\mathbb L}^\alpha$ uniformly in time for all $\alpha$ strictly smaller than $\eta/\rho$. Here, $\eta$ is the Poincaré constant of the potential $V_0$ and $\rho$ is the quantity $$\label{eq:definition_rho}
\rho
=-(\inf
\min\mathrm{Spec}(\nabla^2V(x)))\frac{\int_{{{\mathbb R}^d}}\left(\min\mathrm{Spec}(\nabla^2V(x)) \right)^2{e^{-V(x)}{\mathrm d}x}}{\left(\int_{{{\mathbb R}^d}}\min\mathrm{Spec}(\nabla^2V(x)){e^{-V(x)}{\mathrm d}x}\right)^2}
=c\frac
{\int_{\mathbb R}(12x^2-c)^2{e^{-V(x)}{\mathrm d}x}}
{\left(\int_{\mathbb R}(12x^2-c){e^{-V(x)}{\mathrm d}x}\right)^2}$$ appearing in Assumption [**([[**Spec($\eta/\alpha$)**]{}]{})**]{}. The real number $\rho$ can easily be approximated by one-dimensional numerical integration. Concerning the Poincaré constant $\eta$ of $V_0$, let us first notice that the potential $V_0$ can be written as the sum of a convex potential and a bounded perturbation and thus satisfies a Poincaré inequality thanks to the Holley-Stroock perturbation lemma (see [@ABC-00 Theorem 3.4.1]). The corresponding Poincaré constant can be computed numerically, since it is the second eigenvalue of the operator $L=\partial_x^2-V_0'(x)\partial_x=e^{V_0}\partial_x(e^{-V_0}\partial_x)$, whose first eigenvalue and eigenvector are $0$ and the constant function $\mathbf 1$. The numerical method we use to approximate $\eta$ is the following. First, notice that the spectrum of the operator $L$ is identical to the one of $\tilde
L=e^{-V_0/2} L e^{V_0/2} = e^{V_0/2} \partial_x(e^{-V_0}\partial_x(e^{V_0/2}\times\cdot))$ which is self-adjoint in the space ${\mathbb L}^2({\mathrm d}x)$. The operator $\tilde L$ is then discretized using a regular mesh with constant space step $\delta x$ by the infinite tridiagonal matrix $(M_{i,j})_{i,j\in\mathbb Z}$ defined by $$M_{i,i}=
-\frac1{\delta x^2}
\left(
e^{V(i\delta x)-V((i+1/2)\delta x)}
+e^{V(i\delta x)-V((i-1/2)\delta x)}
\right)$$ and $$M_{i,i+1}
=M_{i+1,i}
=\frac1{\delta x^2}
e^{\frac12V(i\delta x)
+\frac12V((i+1)\delta x)
-V((i+1/2)\delta x)},$$ (with $M_{i,j}=0$ whenever $|i-j|>1$). We consider the restriction to a finite set of indices $(M_{i,j})_{-N\leq i,j\leq N}$, which is equivalent to imposing homogeneous Dirichlet boundary conditions at $x=-N \delta x$ and $x=N \delta x$. These artificial boundary conditions are justified (in the limit $N \to \infty$) by the fact that the eigenvectors of $\tilde L$ go to $0$ at infinity. Since the matrix $(M_{i,j})_{-N\leq i,j\leq N}$ is a nonpositive symmetric matrix, one can successively compute its first eigenvalues by the inverse power method, using at each step a projection on the orthogonal of the eigenvector which have already been computed. We checked that the numerical approximation obtained for the second eigenvalue is converged when $\delta x \to 0$ and $N \to \infty$. The graphs of numerical approximations of both $\rho$ and the Poincaré constant are plotted on Figure \[fig:poincare\_rho\]. In particular, for a curvature constant $c$ located left to the intersection of the two curves (approximately $c\leq0.86$), Proposition \[prop:variance\] ensures that $T_t$ is bounded in ${\mathbb L}^1$, uniformly in time. Also, for curvature constants such that $\rho$ is less than half the Poincaré constant (corresponding approximately to $c\leq0.50$), $T_t$ is bounded in ${\mathbb L}^2$, and thus has a bounded variance uniformly in time. On Figure \[fig:integrabilite\_theorique\], we plot the critical exponent $\eta/\rho$ such that, according to Proposition \[prop:variance\], $T_t$ is in ${\mathbb L}^\alpha$ for $\alpha<\eta/\rho$.
Let us now explain how we estimate numerically the integrability exponent $\alpha$ such that $T_t$ actually is in ${\mathbb L}^\alpha$. This is done by computing the tail of the empirical cumulative distribution function of $T_t$. We simulate $10^6$ independent realizations of the process $(T_t,X_t)$, starting from $\sqrt c/2$ (that is, at the bottom of the right well), up to the time $t=40$, at which the systems seems to be at equilibrium. On Figure \[fig:integrabilite\_empirique\], we plot in logarithmic scale the tail of the empirical cumulative distribution function of those $N=10^6$ independent realizations $(T_t^i)_{i=1,\hdots, N}$, namely $$[0,\infty)\ni x
\mapsto\frac1N\sum_{i=1}^N\mathbf1_{|T_t^i|\geq x}
=\frac1N\sum_{i=1}^N\mathbf1_{T_t^i\leq -x}$$ with curvature $c$ being respectively 2, 3, 4 and 5, from bottom to top. Linear regression in those four cases gives the following slopes: $$\begin{array}{|c|c|c|c|c|}
\hline
c&2&3&4&5\\
\hline
{\rm slope}& -3.09 & -1.95 & -1.29 & -1.12 \\
\hline
\end{array}$$ We have checked that the results are the same for $t=40$ and for $t=80$. Note that an integrable random variable corresponds roughly to a slope less than $-1$, and a square integrable variable corresponds to a slope less than $-2$. We also plot on Figure \[fig:integrabilite\_theorique\] the empirical integrability exponent for different curvatures between $0$ and $3$. We observe that the results are in accordance with Proposition \[prop:variance\]: the theoretical lower bound is indeed smaller than the effective integrability exponent.
For a curvature larger than 3, the tangent vector $T_t$ at time $t=40$ does not seem to be of finite variance. This raises the question of appropriate variance reduction technique to be used in order to use the estimators or . We will investigate in Section \[sec:roland\] a first idea that could be used in this one-dimensional situation. Further studies related to this problem will be the subject of future works.
A many particle system {#sec:WA}
----------------------
In this section, we consider a more complex problem introduced in [@warren-allen-12], and motivated by experimental studies of colloidal particles in optical traps. Let us consider $X^\lambda_t=(Y^1_t,\cdots,Y^N_t)$ where $(Y^i_t)_{i=1\,\hdots,N}\in({\mathbb R}^2)^N$ are the positions of $N$ two-dimensional particles evolving according to $$\label{eq:warren-allen}
{\mathrm d}Y_t^i
= -\kappa Y_t^i{\mathrm d}t
+ \sum_{j=1}^N\nabla U(Y_t^i-Y_t^j){\mathrm d}t
+ \lambda e_1 {\mathrm d}t
+ {\mathrm d}W_t^i,\;1\leq i\leq N$$ with $\kappa>0$, $\lambda\in{\mathbb R}$, $e_1$ the normed vector directed along the first coordinate, and $U(x)=\Gamma e^{-|x|}/|x|$. The particles undergo a quadratic confining potential near the origin with strength $\kappa$, a repulsive interaction given by $U$, a shear in the $x$-direction with strength $\lambda$ and a thermal noise.
We study the case of $N=10$ particles with repulsion range $\Gamma=25$ and attraction intensity $\kappa=10$ corresponding to the parameters studied in [@warren-allen-12]. For those parameters, at equilibrium, particles are gathered around the origin. At $\lambda=0$, no particular direction appears in the dynamics, and the equilibrium measure is invariant with respect to rotations around the origin.
One wants to study the effect of shearing on the symmetry of the invariant measure. This symmetry can be measured by the empirical covariance $\Phi$ of the particle system, defined by $$\Phi(X^\lambda_t)=\frac1N\sum_{i=1}^N(Y_t^{i,1}-\bar Y_t^1)(Y_t^{i,2}-\bar Y_t^2),$$ where $Y_t^i=(Y_t^{i,1},Y_t^{i,2})$ and for $k\in\{1,2\}$, $\bar Y_t^k=\frac1N\sum_{i=1}^NY_t^{i,k}$. One is interested in computing the derivative ${\partial_\lambda^0}\int_{({\mathbb R}^2)^N}\Phi{\mathrm d}\pi_\lambda$.
On Figure \[fig:warren\_allen\], we plot the confidence interval obtained for the expectation ${\mathbb E}[{\partial_\lambda^0}\Phi(X_t^\lambda)]$, with $N=10^5$ independent simulations, as a function of the time $t$. The dynamics has been simulated using an explicit Euler-Maruyama scheme with time step $\delta t=10^{-5}$, and the expectation has been calculated through the Monte Carlo approximation $${\mathbb E}[{\partial_\lambda^0}\Phi(X_t^\lambda)]
={\mathbb E}[T_t\cdot\nabla\Phi(X_t^0)]
\simeq\frac1N\sum_{i=1}^NT_t^i\cdot\nabla\Phi(X_t^{0,i})$$ where the $(X_t^{0,i},T_t^i)_{1\leq i\leq N}$ are independent simulations of the Euler-Maruyama discretization of the dynamics ruling the evolution of $(X_t^0,T_t)$. As in [@warren-allen-12], we observe that the correlation function ${\mathbb E}[{\partial_\lambda^0}\Phi(X_t^\lambda)]$ increases as a function of time, before reaching a plateau. We have checked that similar results are obtained using a finite differenciation instead of the simulation of the couple $(X^0_t,T_t)$.
Particle merging {#sec:roland}
----------------
As mentioned in Section \[sec:1d\], in some situations, the variance of the tangent vector may become very large (or even infinite) which means that the estimators and become ineffective. Therefore, it is desirable to introduce variance reduction mechanisms. We explore in this section a first idea in the simple one-dimensional test case of Section \[sec:1d\]. Extensions and further variance reduction techniques will be the subject of forthcoming works.
A first simple idea to reduce the variance is to replace the tangent vector $T_s$ in the estimator of ${\mathbb E}[\nabla
f(X_s^0)\cdot T_s]$ by its conditional expectation given $X^0_s$. This corresponds in practice to replacing the tangent vector of particles which are at the same position at a given time $s$ by the average of their tangent vectors. Then, the particles evolve again following the dynamics . We refer to this procedure as “particle merging”. In practice, with this naive procedure the probability to observe two particles at the same position is zero, in dimension larger than one. A first simple practical way to implement this technique is to introduce small subsets of the configuration space, and to merge particles which are in the same subset, which of course reduces the variance but introduce some bias. The merging can be performed in a much efficient way and in larger dimensions, by correlating the particles, see e.g. [@kalos-pederiva-00]. This will be the scope of future work. Before studying the interest of particle merging on the simple case of Section \[sec:1d\], let us first state the theoretical result which justifies the use of this approach.
Assume [**([[**min Spec**]{}]{})**]{}. For $s\ge 0$, let $(\tilde T_t)_{t\geq s}$ be the solution to $$\left\{
\begin{aligned}
\frac{{\mathrm d}\tilde T_t}{{\mathrm d}t}&={\partial_\lambda^0}F_\lambda(X_t^0)-\nabla^2(X_t^0)\tilde T_t \text{ , for all }~t\geq s, \\
\tilde T_s&={\mathbb E}[T_s|X_s^0].
\end{aligned}
\right.$$ Then $\forall t\geq s,\;\tilde T_t={\mathbb E}[T_t|X^0_s,(W_r-W_s)_{r\in[s,t]}]$. Assume moreover that $f:{\mathbb R}^d\to{\mathbb R}^d$ is a Lipschitz function belonging to ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$ for some $p\in[1,\infty]$ and that the initial condition $X_0$ to admits a density with respect to $\pi_0$ belonging to ${\mathbb L}^{\frac{p}{p-1}}({e^{-V(x)}{\mathrm d}x})$ (where, by convention, $\frac{p}{p-1}=\infty$ if $p=1$). Then, for each $t\geq 0$, $f(X_t^\lambda)$ is integrable for $\lambda\in[0,\lambda_0]$, $\lambda\mapsto {\mathbb E}[f(X_t^\lambda)]$ is differentiable at $\lambda=0$ and $${\partial_\lambda^0}\left({\mathbb E}[f(X_t^\lambda)]\right)={\mathbb E}[\nabla f(X_t^0)\cdot T_t]={\mathbb E}[\nabla f(X_t^0)\cdot \tilde T_t],\mbox{ for each }t\ge s.$$
This Lemma shows that if, at a given time $s$, the particles at position $X^0_s$ replace their current tangent vectors by an average of these tangent vectors, and then follow the dynamics for $t \ge s$, the estimator is still consistent.
By Lemma \[lem:R\_T\_Linfty\], Assumption [**([[**min Spec**]{}]{})**]{} ensures that $T_t$ is integrable for each $t\geq 0$. In view of the equality and using the semigroup property of $R_{X^0}$, one gets that for $t\geq s\geq 0$, $$T_t=R_{X^0}(s,t)T_s+\int_s^tR_{X^0}(r,t){\partial_\lambda^0}F_\lambda(X_r^0)dr.$$ Since $(X^0_r)_{r\in[s,t]}$ and therefore $(R_{X^0}(r,t))_{r\in[s,t]}$ are measurable with respect to the sigma-field generated by $X^0_s$ and $(W_r-W_s)_{r\in[s,t]}$, one deduces that $${\mathbb E}[T_t|X^0_s,(W_r-W_s)_{r\in[s,t]}]=R_{X^0}(s,t){\mathbb E}[T_s|X^0_s,(W_r-W_s)_{r\in[s,t]}]+\int_s^tR_{X^0}(r,t){\partial_\lambda^0}F_\lambda(X_r^0)dr.$$ The independence of $(X^0_s,T_s)$ and $(W_r-W_s)_{r\in[s,t]}$ implies that ${\mathbb E}[T_s|X^0_s,(W_r-W_s)_{r\in[s,t]}]={\mathbb E}[T_s|X^0_s]$. Since, by an adaptation of Proposition \[prop:tangent\], $$\tilde{T}_t=R_{X^0}(s,t){\mathbb E}[T_s|X^0_s]+\int_s^tR_{X^0}(r,t){\partial_\lambda^0}F_\lambda(X_r^0)dr,$$ one concludes that $\tilde{T}_t={\mathbb E}[T_t|X^0_s,(W_r-W_s)_{r\in[s,t]}]$.
If the initial condition $X_0$ to admits a density with respect to $\pi_0$ belonging to ${\mathbb L}^{\frac{p}{p-1}}({e^{-V(x)}{\mathrm d}x})$, then so does $X^0_t$ for each $t\geq 0$ by Lemma \[lem:densite\_X\_t\]. When $f:{\mathbb R}^d\to{\mathbb R}^d$ is a Lipschitz function belonging to ${\mathbb L}^p({e^{-V(x)}{\mathrm d}x})$, the integrability of ${\mathbb E}[f(X_t^\lambda)]$, the differentiability of $\lambda\mapsto {\mathbb E}[f(X_t^\lambda)]$ at $\lambda=0$ and the equality ${\partial_\lambda^0}\left({\mathbb E}[f(X_t^\lambda)]\right)={\mathbb E}[\nabla f(X_t^0)\cdot T_t]$ are deduced from an adaptation of the beginning of the proof of Theorem \[theo:GK\_tps\_fini\]. Now, for $t \ge s$, $$\begin{aligned}
{\mathbb E}[\nabla f(X_t^0)\cdot T_t]&={\mathbb E}[{\mathbb E}[\nabla f(X_t^0)\cdot T_t|X^0_s,(W_r-W_s)_{r\in[s,t]}]]\\&={\mathbb E}[\nabla f(X_t^0)\cdot{\mathbb E}[T_t|X^0_s,(W_r-W_s)_{r\in[s,t]}]]={\mathbb E}[\nabla f(X_t^0)\cdot\tilde{T}_t].\end{aligned}$$
To test the interest of this approach, we consider again the setting of Section \[sec:1d\] with $c=2.9$ (which corresponds to case where the variance of tangent vector $T_t$, at $t=40$, is very large, see Figure \[fig:integrabilite\_theorique\]). The merging procedure is done as follows: a uniform mesh with step size $0.04$ is introduced, and, every ten timesteps, the tangent vectors of particles which are in the same bin are replaced by an average of these tangent vectors. On Figure \[fig:int\_conf\], we observe that this procedure divides approximately the variance by 4, while introducing a bias which is sufficiently small so that the confidence interval of the simulation with merging is included in the confidence interval of the simulation without merging. Figure \[fig:variance\_merging\] then gives more quantitative estimates of the variances of these two simulations (with and without merging), as a function of time. We have observed numerically that large values of $T_t$ become very unlikely with the merging procedure: using $10^3$ independant realizations of $10^3$ interacting particles over the time interval $(0,10)$, we did not observe any realization of $T_t$ with absolute value larger than $3$ (compare with what is reported on Figure \[fig:integrabilite\_empirique\]).
Alternative bounds on the density of $X^0_t$ {#sec:hypdelta}
============================================
In this section, we would like to present a few results which can be obtained under the assumption
The function $V$ is of class $\mathcal {\mathcal C}^2$ and satisfies $$C_V=\sup_{x\in{{{\mathbb R}^d}}}(2\Delta V(x)-|\nabla V(x)|^2)<+\infty.$$
Note that simple assumptions on the quantity $a_V(x)=2\Delta V(x)-|\nabla V(x)|^2$ can give strong results on the equilibrium measure ${e^{-V(x)}{\mathrm d}x}$. For instance, if $a_V(x)$ goes to $-\infty$ at infinity, then the equilibrium measures satisfies a Poincaré inequality (see for example the appendix in [@villani-09]).
Bounds on the density of $X^0_t$
--------------------------------
Consider the setting and the notation of Lemma \[lem:densite\_X\_t\] and let Assumption [**([[**V**]{}]{})**]{} hold. Assume that the measure $e^{\frac12V} {\mathrm d}\mu_0$ can be written as $$e^{\frac12V(x)} {\mathrm d}\mu_0 =f(x){\mathrm d}x+{\mathrm d}\nu,$$ where $f$ is some function in ${\mathbb L}^p({\mathrm d}x)$ with $p \in [1,2]$ and $\nu$ is some finite measure on ${{{\mathbb R}^d}}$. Then, for any $t>0$, $\mu_t$ is absolutely continuous with respect to $e^{-\frac12V(x)}{\mathrm d}x$ with $$\label{eq:majonorml2}
\left\|\frac{{\mathrm d}\mu_t}{e^{-\frac{1}{2}V(x)} {\mathrm d}x}\right\|_{{\mathbb L}^2({\mathrm d}x)}
\leq Ce^{Ct}\left(\frac{C_p}{ t^{(1/p-1/2)d/2}}+\frac{\nu({{{\mathbb R}^d}})}{t^{d/4}}\right).$$
Let $\psi:{\mathbb R}^d\to{\mathbb R}$ be a bounded measurable function and recall the formula $${\mathbb E}[\psi(X_t^0)]
={\mathbb E}\left[
\psi(X_0+\sqrt2W_t)
e^{
-\frac12V(X_0+\sqrt2W_t)
}
e^{ \frac12V(X_0)}
e^{
\frac14\int_0^t\left(2\Delta V-|\nabla V|^2\right)(X_0+\sqrt2W_s){\mathrm d}s
}
\right]$$ obtained by the Girsanov theorem, see Equation .
If $C$ is an upper bound for $\frac14(2\Delta V-|\nabla V|^2)$, and if one assumes $\psi\geq0$, one obtains $$\begin{aligned}
\int_{{{\mathbb R}^d}}\psi{\mathrm d}\mu_t
&\leq
e^{Ct}{\mathbb E}\left[
\psi(X_0+\sqrt2W_t)
e^{
-\frac12V(X_0+\sqrt2W_t)
}
e^{ \frac12V(X_0)}
\right]\\
&=\frac{e^{Ct}}{(4\pi t)^{d/2}}\int_{{{\mathbb R}^d}}\psi(y)
e^{-\frac12V(y)}
\left(\int_{{{\mathbb R}^d}}e^{\frac12V(x)}
e^{-\frac{|y-x|^2}{4t}}{\mathrm d}\mu_0(x)\right)
{\mathrm d}y.\end{aligned}$$ so that $\mu_t\ll e^{-\frac12V(x)}{\mathrm d}x$ with a Radon-Nikodym derivative $\frac{{\mathrm d}\mu_t}{e^{-\frac12V(x)}{\mathrm d}x}$ satisfying $$\begin{aligned}
\frac{{\mathrm d}\mu_t}{e^{-\frac12V(x)}{\mathrm d}x}
&\leq e^{Ct}\left(e^{\frac12V}\mu_0\right)*\gamma_t=e^{Ct}\left(f*\gamma_t+\nu*\gamma_t\right),\end{aligned}$$ where $*$ stands for the convolution product, and $\gamma_t(x)=\frac{1}{(4 \pi
t)^{d/2}} e^{-\frac{|x|^2}{4t}}$ denotes the centered Gaussian density with covariance matrix $2tI_d$. One concludes that holds by:
- the Young inequality $\|f * \gamma_t\|_{{\mathbb L}^2({\mathrm d}x)} \le \|f\|_{{\mathbb L}^p({\mathrm d}x)} \|\gamma_t\|_{{\mathbb L}^q({\mathrm d}x)}$ where $1/p+1/q =3/2$ ($p,q \in [1, \infty]$) and the heat kernel estimate $\|\gamma_t\|_{{\mathbb L}^q({\mathrm d}x)} \le C_q
t^{-\left(1-\frac{1}{q}\right) \frac{d}{2}}$;
- the estimate $\|\nu*\gamma_t\|^2_{{\mathbb L}^2({\mathrm d}x)}
\leq\|\nu*\gamma_t\|_{{\mathbb L}^1({\mathrm d}x)}\|\nu*\gamma_t\|_{{\mathbb L}^\infty({\mathrm d}x)}
\leq\frac{(\nu({{{\mathbb R}^d}}))^2}{(4\pi t)^{d/2}}$.
An additional result
--------------------
Assumption [**([[**V**]{}]{})**]{} can also be useful to prove the second point in Assumption [**([[**Pot**]{}]{})**]{}-$(ii)$ on the potential $V$.
\[lem:nablaV\_L2\] Under Assumption [**([[**V**]{}]{})**]{}, the function $\nabla V$ is in ${\mathbb L}^2({e^{-V(x)}{\mathrm d}x})$: $$\int_{{{\mathbb R}^d}}|\nabla V|^2(x){e^{-V(x)}{\mathrm d}x}\leq C_V.$$
Let $\chi_n(x)=\chi(x/n)$ where $\chi$ is a smooth, $[0,1]$-valued, cutoff function such that $\chi(x)=1$ for $|x|<1$ and $\chi(x)=0$ for $|x|>2$. $$\begin{aligned}
\int_{{{\mathbb R}^d}}|\nabla V|^2(x)\chi_n(x){e^{-V(x)}{\mathrm d}x}&=-\int_{{{\mathbb R}^d}}\left(\chi_n(x)\nabla V(x)\right)\cdot\nabla{e^{-V(x)}{\mathrm d}x}\\
&=\int_{{{\mathbb R}^d}}\nabla\chi_n(x)\cdot\nabla V(x){e^{-V(x)}{\mathrm d}x}+\int_{{{\mathbb R}^d}}\chi_n(x)\Delta V(x){e^{-V(x)}{\mathrm d}x}\\
&\leq\int_{{{\mathbb R}^d}}\Delta\chi_n(x){e^{-V(x)}{\mathrm d}x}+\frac12\int_{{{\mathbb R}^d}}\chi_n(x)|\nabla V|^2(x){e^{-V(x)}{\mathrm d}x}+\frac12C_V.\end{aligned}$$ As a consequence, $$\int_{{{\mathbb R}^d}}|\nabla V|^2(x)\chi_n(x){e^{-V(x)}{\mathrm d}x}\leq2\int_{{{\mathbb R}^d}}\Delta\chi_n(x){e^{-V(x)}{\mathrm d}x}+C_V$$ and the result follows from taking $n\to\infty$ by Fatou’s lemma for the left-hand side and Lebesgue’s theorem for the right-hand side.
About the Assumption [**([[**Conv**]{}]{})**]{} {#sec:annex_conv}
===============================================
In this section, we show that Assumption [**([[**Conv**]{}]{})**]{} is a natural one, since it appears as a sufficient condition in another related problem.
We recall that $(Y^x_t)_{t\geq 0}$ is defined in as the solution to $$\label{eq:Ytx_prime}
\forall t\geq0,~
Y^x_t
=x-\int_0^t\nabla V(Y^x_s){\mathrm d}s
+\sqrt2W_t.$$ Since $(R_{Y^x}(0,t))_{t\geq0}$ is the differential of the trajectory $(Y^x_t)_{t\geq0}$ with respect to $x$, we expect that a condition yielding long-time decay for $R_{Y^x}$ will imply that trajectories with same noise and close initial conditions will eventually converge toward each other. More precisely, we are interested in the joint long-time behavior of the so-called duplicated dynamics $(Y_t^x,Y_t^y)_{t\geq0}$, where $x$ and $y$ are two different initial conditions. Note here that the two processes $(Y^x_t)_{t\geq0}$ and $(Y^y_t)_{t\geq0}$ are driven by the same Brownian motion.
In [@lemaire-pages-panloup-13], the same problem is considered for a diffusion whose diffusion matrix may not be constant. In that case, an example is provided, where the process $Y_t^x-Y_t^y$ does not converge to $0$.
A similar problem is considered in [@burdzy-chen-jones-06]: the process is a Brownian motion reflected on the boundary of a domain $\Omega$. Such a dynamics can be formally seen as a singular case of the problem we consider, with $V=\infty\times\mathbf1_{\Omega^c}$. Equation then has to be written with a local time on the boundary in place of $\nabla V$. In that case, the difference $Y_t^x-Y_t^y$ will converge to $0$ if the domain $\Omega$ is smooth enough and has at most one hole. However, it is conjectured that the same result holds for much more general domains.
We will use the fact that $V$ is such that the dynamics is ergodic with respect to the invariant measure ${e^{-V(x)}{\mathrm d}x}$.
The one-dimensional case
------------------------
In the one-dimensional case, this question is especially simple, because of the order structure on the state space. In particular (see [@lemaire-pages-panloup-13]), it can be checked that if for any $x\in{\mathbb R}$, $Y_t^x$ converge weakly to $\pi_0$ as $t\to\infty$, then the only invariant distribution of the duplicated dynamics is the image of ${e^{-V(x)}{\mathrm d}x}$ by $x\mapsto(x,x)$. Actually, under additional assumption, one can show that $Y^x_t-Y^y_t$ converges in mean to $0$ in the long-time limit.
Assume that the dimension is $d=1$. If for any $x\in{\mathbb R}$ the time marginals of the process $(Y_t^x)_{t\geq0}$ converge weakly to $\pi_0$ as $t\to\infty$ and the random variables $(Y_t^x)_{t\geq0}$ are uniformly integrable, then, for any $x,y\in{\mathbb R}$, the process $(Y_t^x-Y_t^y)_{t\geq0}$ converges to $0$ in ${\mathbb L}^1(\Omega)$.
According to Corollary \[cor:CV\_L1\], the long-time convergence of the marginals holds for instance if the potential $V$ satisfies a Poincaré inequality (see Assumption [**([[**Poinc($\eta$)**]{}]{})**]{}).
First, from the uniform integrability of $(X_t^x)_{t\geq0}$ and the weak convergence of the time marginals, both ${\mathbb E}[Y_t^x]$ and ${\mathbb E}[Y_t^y]$ converge to $\int_{\mathbb R}x{e^{-V(x)}{\mathrm d}x}$ as $t\to\infty$.
Now assume, without loss of generality that $x\leq y$. Then, from a comparison theorem, $Y_t^x\leq Y_t^y$ holds for all positive times, and one obtains $${\mathbb E}[|Y_t^x-Y_t^y|]
={\mathbb E}[Y_t^x-Y_t^y]
={\mathbb E}[Y_t^x]-{\mathbb E}[Y_t^y]
\to0.$$
A general criterion
-------------------
The following facts hold true:
1. Assume that $$\label{eq:minoconv}
\forall x,y\in{{{\mathbb R}^d}},
\;(x-y)\cdot(\nabla V(x)-\nabla V(y))\geq \frac{v(x)+v(y)}{2}|x-y|^2$$ with $v:{{{\mathbb R}^d}}\to{\mathbb R}$ such that $\int_{{{\mathbb R}^d}}\max(0,-v(x)){e^{-V(x)}{\mathrm d}x}<\infty$ and $\int_{{{\mathbb R}^d}}v(x)e^{-V(x)}{\mathrm d}x>0$. Then for all $x,y\in{{{\mathbb R}^d}}$, $|Y^x_t-Y^y_t|$ converges [*a.s.*]{} to $0$, exponentially fast at any rate between $0$ and $\int_{{{\mathbb R}^d}}v(x){e^{-V(x)}{\mathrm d}x}$ as $t\to\infty$.
2. The exponential convergence to $0$ still holds if $V$ is convex and there exist $x_0\in{{{\mathbb R}^d}}$ and ${\varepsilon}>0$ such that the inequality $\inf_{x\in B(x_0,{\varepsilon})}\min{\rm Spec}(\nabla^2V(x))>0$ holds.
3. If $V$ is convex, then the only invariant measure of the duplicated dynamics is the image of ${e^{-V(x)}{\mathrm d}x}$ by $x\mapsto (x,x)$.
Let us start with a few remarks:
- The first point can be applied to the so-called Mexican hat potential $V(x)=\beta(|x|^4-\gamma |x|^2)$, with $\beta>0$ and $\gamma>0$, in dimension $d\geq 2$. For this potential, one has $$\begin{aligned}
(x-y)\cdot(\nabla V(x)-\nabla V(y))
&=2\beta|x-y|^2(|x|^2+|y|^2-\gamma)+2\beta(|x|^2-|y^2|)^2\\
&\geq\frac{v(x)+v(y)}{2}|x-y|^2,
\end{aligned}$$ for $v(x)=\beta(4|x|^2-2\gamma)$. In addition, one has $\int_{{{\mathbb R}^d}}v(x){e^{-V(x)}{\mathrm d}x}>0$ since $$\begin{aligned}
\frac{\int_{{\mathbb R}^d}|x|^2e^{\beta(\gamma
|x|^2-|x|^4)}{\mathrm d}x}{\int_{{\mathbb R}^d}e^{\beta(\gamma
|x|^2-|x|^4)}{\mathrm d}x}
&=\frac{\int_0^{+\infty}r^{\frac{d}{2}}e^{\beta(\gamma
r-r^2)}dr}{\int_0^{+\infty}r^{\frac{d}{2}-1}e^{\beta(\gamma
r-r^2)}dr}\\
&=\frac{\gamma}{2}+\frac{\int_0^{+\infty}r^{\frac{d}{2}-1}(r-\frac{\gamma}{2})e^{\beta(\gamma r-r^2)}dr}{\int_0^{+\infty}r^{\frac{d}{2}-1}e^{\beta(\gamma r-r^2)}dr}\\
&=\frac{\gamma}{2}+\frac{1_{\{d=2\}}}{2\beta
\int_0^{+\infty}e^{\beta(\gamma
r-r^2)}dr}+1_{\{d>2\}}\frac{(d-2)\int_0^{+\infty}r^{\frac{d}{2}-2}e^{\beta(\gamma
r-r^2)}dr}{4\beta\int_0^{+\infty}r^{\frac{d}{2}-1}e^{\beta(\gamma
r-r^2)}dr}\\
&>\frac{\gamma}{2}.
\end{aligned}$$
- Letting $y\to x$ in , one obtains that $\forall x\in{{{\mathbb R}^d}}$, $v(x)\leq \min{\rm Spec}(\nabla^2V(x))$. When $x\mapsto\min{\rm Spec}(\nabla^2V(x))$ is concave, $$\begin{aligned}
(x-y) \cdot (\nabla V(x)-\nabla V(y))
&=\int_0^1(x-y)\cdot\nabla^2V(\theta x+(1-\theta)y)(x-y){\mathrm d}\theta\\
&\geq |x-y|^2\int_0^1\min{\rm Spec}(\nabla^2V(\theta x+(1-\theta)y)){\mathrm d}\theta\\
&\geq |x-y|^2\int_0^1\theta\min{\rm Spec}(\nabla^2V(x))+(1-\theta)\min{\rm Spec}(\nabla^2V(y)){\mathrm d}\theta\\
&\geq \frac12\left(\min{\rm Spec}(\nabla^2V(x))+\min{\rm Spec}(\nabla^2V(y))\right)|x-y|^2
\end{aligned}$$ and one may choose $v(x)=\min{\rm Spec}(\nabla^2V(x))$ in .
- When $V=\bar{V}+\hat{V}$ with $\bar{V}$ such that $x\mapsto\min{\rm Spec}(\nabla^2\bar{V}(x))$ is concave and $\hat{V}$ such that $x\mapsto \nabla\hat{V}(x)$ is Lipschitz with constant $\delta$ and constant outside some Borel subset $A$ of ${{{\mathbb R}^d}}$, then one may choose $v(x)=\min{\rm Spec}(\nabla^2\bar{V}(x))-2\delta 1_A(x)$ in .
1. One has $$\begin{aligned}
\label{eq:decroissance}
{\mathrm d}|Y^x_t-Y^y_t|^2
&=-2(Y^x_t-Y^y_t)\cdot(\nabla V(Y^x_t)-\nabla V(Y^y_t)){\mathrm d}t\\
&\leq -(v(Y^x_t)+v(Y^y_t))|Y^x_t-Y^y_t|^2{\mathrm d}t,\nonumber
\end{aligned}$$ under . Hence $$|Y^x_t-Y^y_t|^2
\leq |x-y|^2e^{-\int_0^t(v(Y^x_s)+v(Y^y_s)){\mathrm d}s}.$$ Since, by , $\frac1t\int_0^t(v(Y^x_s)+v(Y^y_s)){\mathrm d}s$ converges [*a.s.*]{} to $2\int_{{{\mathbb R}^d}}v(x){e^{-V(x)}{\mathrm d}x}>0$, one easily deduces the first assertion.
2. When $V$ is convex, then $t\mapsto |Y^x_t-Y^y_t|$ is nonincreasing by . Now, for $z\in B(x_0,\frac{\varepsilon}2)$ and $w\in{{{\mathbb R}^d}}$, one has $$(z-w)\cdot(\nabla V(z)-\nabla V(w))
\geq|z-w|\inf_{B(x_0,{\varepsilon})}\min{\rm Spec}(\nabla^2V(\cdot))
\left(
\frac{\varepsilon}21_{B(x_0,{\varepsilon})^c}(w)+|z-w|1_{B(x_0,{\varepsilon})}(w)
\right).$$ As a consequence, $$(Y^x_t-Y^y_t)\cdot(\nabla V(Y^x_t)-\nabla V(Y^y_t))
\geq1_{B(x_0,\frac{{\varepsilon}}{2})}(Y^x_t)
\inf_{B(x_0,{\varepsilon})}\hspace{-3pt}\min{\rm Spec}(\nabla^2V(\cdot))
\left(\frac{\varepsilon}{2|x-y|}\wedge1\right)|Y^x_t-Y^y_t|^2.$$ One concludes by arguments similar to the ones used for the first assertion.
3. Let $V$ be convex and differentiable and let $x\neq y$ be such that $(x-y)\cdot(\nabla V(x)-\nabla V(y))=0$. Then $V$ is affine on the segment $[x,y]$ and $V(\frac{x+y}{2})=\frac{V(x)+V(y)}{2}$. For $z\in{{{\mathbb R}^d}}\setminus\{0\}$ and ${\varepsilon}\in {\mathbb R}$, $$\begin{aligned}
\frac{V(x)+V(y)}{2}
=V\left(\frac{x+y}2\right)
&\leq \frac{V(x+{\varepsilon}z)+V(y-{\varepsilon}z)}{2}\\
&=\frac{V(x)+V(y)}{2}+\frac{{\varepsilon}z}{2}\cdot(\nabla V(x)-\nabla V(y))+o({\varepsilon})
\end{aligned}$$ as $|{\varepsilon}|\to 0$. As a consequence $z\cdot(\nabla V(x)-\nabla V(y))=0$ and $\nabla V(x)=\nabla V(y)$.
Let $(X_t)_{t\geq0}$ and $(Y_t)_{t\geq0}$ be two solutions to the stochastic differential equation , such that $(X_0,Y_0)$ is distributed according to some invariant probability measure of the duplicated dynamics. Since $|X_t-Y_t|^2$ is [*a.s.*]{} non-increasing with $t$ and constant in distribution, [*a.s.*]{} $t\mapsto |X_t-Y_t|^2$ is constant and therefore ${\mathrm d}t$-[*a.e.*]{} $(X_t-Y_t)\cdot(\nabla V(X_t)-\nabla V(Y_t))=0$ which implies $\nabla V(X_t)=\nabla V(Y_t)$. One deduces that [ *a.s.*]{}, $t\mapsto X_t-Y_t$ is constant.\
Now, since $x\mapsto e^{-V(x)}$ is integrable, then $V$ cannot be affine in some direction and for any $z\in{{{\mathbb R}^d}}\setminus\{0\}$, $x\mapsto
z\cdot(\nabla V(x)-\nabla V(x-z))$ is not constant equal to zero. By continuity of $\nabla V$, one deduces the existence of $y\in{{{\mathbb R}^d}}$ and ${\varepsilon}>0$ such that $\forall x\in B(y,{\varepsilon})$, $z\cdot(\nabla
V(x)-\nabla V(x-z))>0$. With the ergodicity of $(X_t)_{t\geq 0}$ and the fact that ${\mathrm d}t$-[*a.e.*]{} $(X_0-Y_0)\cdot(\nabla V(X_t)-\nabla
V(X_t-X_0+Y_0))=0$, one concludes that [*a.s.*]{} $X_0=Y_0$.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement number 614492 and by the French National Research Agency under the grant ANR-12-BS01-0019 (STAB). The authors would like to thank fruitful discussions with G. Stoltz on nonequilibrium methods and Green-Kubo formulae.
[^1]: Laboratoire de Chimie Théorique, CNRS-UMR 7616 et Université Pierre et Marie Curie, 75252 Paris Cedex, France, `[email protected]`
[^2]: Université Paris-Est, CERMICS (ENPC), INRIA, F-77455 Marne-la-Vallée, France, `[email protected]`
[^3]: Université Paris-Est, CERMICS (ENPC), INRIA, F-77455 Marne-la-Vallée, France, `[email protected]`
[^4]: Laboratoire de Probabilités et Modèles Aléatoires, UMR 7599, UPMC, Case 188, 4 pl. Jussieu, F-75252 Paris Cedex 5, France, `[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Nowadays stereo cameras are more commonly adopted in emerging devices such as dual-lens smartphones and unmanned aerial vehicles. However, they also suffer from blurry images in dynamic scenes which leads to visual discomfort and hampers further image processing. Previous works have succeeded in monocular deblurring, yet there are few studies on deblurring for stereoscopic images. By exploiting the two-view nature of stereo images, we propose a novel stereo image deblurring network with **D**epth **A**wareness and **V**iew **A**ggregation, named **DAVANet**. In our proposed network, 3D scene cues from the depth and varying information from two views are incorporated, which help to remove complex spatially-varying blur in dynamic scenes. Specifically, with our proposed fusion network, we integrate the bidirectional disparities estimation and deblurring into a unified framework. Moreover, we present a large-scale multi-scene dataset for stereo deblurring, containing 20,637 blurry-sharp stereo image pairs from 135 diverse sequences and their corresponding bidirectional disparities. The experimental results on our dataset demonstrate that DAVANet outperforms state-of-the-art methods in terms of accuracy, speed, and model size.'
author:
- |
Shangchen Zhou$^1$ Jiawei Zhang$^1$ Wangmeng Zuo$^{2}$[^1] Haozhe Xie$^2$ Jinshan Pan$^3$ Jimmy Ren$^1$\
$^1$SenseTime Research $^2$Harbin Institute of Technology, Harbin, China\
$^3$Nanjing University of Science and Technology, Nanjing, China\
[<https://shangchenzhou.com/projects/davanet>]{}
title: 'DAVANet: Stereo Deblurring with View Aggregation'
---
Introduction
============
With the wide use of dual-lens smartphones, unmanned aerial vehicles and autonomous robots, stereoscopic vision has attracted increasing attention from researchers. Relevant studies not only covers traditional stereo tasks, such as stereo matching [@lecun2015stereo; @chang2018pyramid; @pang2018zoom] and scene flow estimation [@mayer2016large; @moritz2015kitti; @ilg2018occlusions], but also some novel tasks for improving visual effects of stereoscopic 3D contents, for example, stereo super-resolution [@jeon2018enhancing], stereo video retargeting [@li2018depth] and stereo neural style transfer [@chen2018stereoscopic; @gong2018neural]. However, stereo image deblurring has rarely been discussed. In fact, the images captured by handheld or on-board stereo cameras often contain blur due to camera shake and object motion. The blurry stereo images would cause visual discomfort to viewers and make it difficult for further image processing. Dynamic scene deblurring from a single blurry image is a highly ill-posed task. Due to depth variation and object/camera motion in dynamic scenes, it is difficult to estimate spatially variant blur with the limited information from single observation. Although the existing CNN based methods [@tao2018scale; @zhang2018dynamic; @kupyn2018deblurgan; @aittala2018burst; @nah2017deep; @su2017deep] have achieved encouraging results in monocular deblurring, they still fail when handling complicated non-uniform blur. To the best of our knowledge, there are few traditional methods [@xu2012depth; @sellent2016stereo; @pan2017simultaneous] proposed to exploit stereo information for deblurring, where a coarse depth or piecewise rigid 3D scene flow is utilized to estimate blur kernels in a hierarchical or iterative framework. However, they are time-consuming due to the complex optimization process. With a stereopsis configuration, our motivation is based on two observations: (1) Depth information can provide helpful prior information for estimating spatially-varying blur kernels. The near points are more blurry than the distant ones in a static scene which can be seen between the green and yellow boxes in Figure \[fig:varyingblur\]. Compared monocular-based algorithms, the proposed stereo-based method can obtain more accurate depth information by the disparity estimation. (2) The varying information in corresponding pixels cross two stereo views can help blur removal. In Section \[sec:motivation\], we demonstrate that the corresponding pixels in two different views have different blurs due to the motion perpendicular towards the camera and rotation, which is shown as the green boxes in Figure \[fig:varyingblur\]. The network can benefit from aggregated information, where the sharper pixel can be transferred and selected by using an adaptive fusion scheme. Two views can also share varying information, e.g., non-occlusion areas, caused by different viewpoints. Inspired by these two insights, we propose a novel depth-aware and view-aggregated stereo deblurring network, named *DAVANet*. It consists of *DeblurNet* and *DispBiNet*, for image deblurring and bidirectional disparities estimation respectively. The *DeblurNet* and the *DispBiNet* are integrated at feature domain by the proposed fusion network, named *FusionNet*. Specifically, the *DispBiNet* provides depth-integrated features and bidirectional disparities for the *FusionNet*. The *FusionNet* fully exploits these inputs and enriches the *DeblurNet* features with embedding depth and the other view information. With the perception of 3D scene information from stereo images, the proposed method is effective for dynamic scene deblurring. Finally, to obtain richer contextual information, a context module is designed to incorporate the multi-scale contextual information by applying several parallel atrous convolutions with different dilation rates. Currently, there is no particular dataset for stereo deblurring. As a result, we propose a large-scale multi-scene stereo blurry image dataset. It consists of 20,637 blurry-sharp stereo image pairs from 135 different sequences (98 for training and 37 for testing) and corresponding bidirectional disparities obtained from the ZED stereo camera [@stereolabs]. We adopt the blur generation method used in [@li2010generating; @nah2017deep; @su2017deep], that is, approximating a longer exposure by accumulating the frames in an image sequence. We first interpolate frame of captured videos to a very high frame rate (480 fps) using frame interpolation method proposed in [@niklaus2017iccv] and then average the sharp sequence to create a blurry image. The main contributions are summarized as follows:
- We propose a unified network for stereo deblurring. The *DispBiNet* predicts the bidirectional disparities for depth awareness as well as view information aggregation in the *FusionNet*, which helps the *DeblurNet* to remove dynamic scene blur from stereo images.
- We present a first large-scale multi-scene dataset for stereo deblurring, which consists of 20,637 stereo images from 135 diverse scenes. It is currently the largest dataset for deblurring.
- We both quantitatively and qualitatively evaluate our method on our dataset and show that it performs favorably against state-of-the-art algorithms in terms of accuracy, speed as well as model size.
Related Work
============
Our work is a new attempt for solving stereo image deblurring by integrating blur removal and disparity estimation into a unified network. The following is a review of relevant works on monocular single-image deblurring, monocular multi-image deblurring, as well as stereo image deblurring respectively. **Single-image Deblurring.** Many methods have been proposed for single-image deblurring. Some natural image priors are designed to help blur removal, such as $L_0$-regularized prior [@xu2013unnatural], dark channel prior [@pan2016blind], and discriminative prior [@li2019blind]. However, it is difficult for these methods to model spatially variant blur in dynamic scenes. To model the non-uniform blur, some depth-based methods [@lee2017joint; @park2017joint; @hu2014joint; @paramanand2013non] that utilize the predicted depth map to estimate different blur kernels. When the blur kernels are not be accurately estimated, they tend to generate visual artifacts in restored images. Moreover, they are computationally inefficient due to the complex optimization process. Recent years have witnessed significant advances in single image deblurring by CNN-based models. Several methods [@sun2015learning; @gong2017motion] use CNNs to estimate the non-uniform blur kernels. A conventional non-blind deblurring algorithm [@zoran2011learning] is used removing blur, which is time-consuming. More recently, many end-to-end CNN models for image deblurring have also been proposed [@nah2017deep; @noroozi2017motion; @zhang2017learning; @tao2018scale; @zhang2018dynamic; @kupyn2018deblurgan]. To obtain a large receptive field in the network for blur removal, [@tao2018scale] and [@tao2018scale] develop a very deep multi-scale networks in coarse-to-fine manner. Different from [@nah2017deep], Tao *et al.* [@tao2018scale] share the weights of the network at three different spatial scales and use the LSTM to propagate information across scales. To handle spatially variant blur in dynamic scenes, Zhang *et al.* [@zhang2018dynamic] adopt a VGG network to estimate the pixel-wise weights of the spatially variant RNNs [@liu2016learning] for blur removal in feature space. Noroozi *et al.* [@noroozi2017motion] build skip connections between the input and output, which reduces the difficulty of restoration and ensures color consistency. In addition, the adversarial loss is used in [@nah2017deep; @kupyn2018deblurgan] to restore more texture details. **Multi-image Deblurring.** Recently, several CNN-based methods [@su2017deep; @hyun2017online; @kim2018spatio; @aittala2018burst] have been proposed for monocular multi-image (video/burst) deblurring. [@su2017deep] and [@kim2018spatio] align the nearby frames with the reference frame to restore the sharp images, which can obtain more rich information cross different images. Kim *et al.* [@hyun2017online] propose a frame recurrent network to aggregate multi-frame features for video deblurring. By repeatedly exchanging the features across the burst images, Aittala *et al.* [@aittala2018burst] propose an end-to-end burst deblurring network in an order-independent manner. Based on the observations that the different images from video or burst are blurred differently, these multi-image fusion methods usually lead to good performance. **Stereo Deblurring.** So far, there are few traditional methods [@xu2012depth; @sellent2016stereo; @pan2017simultaneous] that leverage the scene information (i.e., disparity and flow) from stereo images for deblurring. Xu and Jia [@xu2012depth] partition the image into regions according to disparity (depth) estimated from stereo blurry images and estimate their blur kernels hierarchically. The methods [@sellent2016stereo; @pan2017simultaneous] propose a stereo video deblurring framework, where 3D scene flow estimation and blur removal are conducted jointly so that they can enhance each other with an iterative manner.
Proposed Method
===============
Motivation {#sec:motivation}
----------
The motivation that utilizing stereo camera for dynamic scene deblurring is inspired by two observations, which is exemplified in Figure \[fig:varyingblur\]. First, we find that nearby object points are more blurry than distant ones and stereo cameras can provide depth information (disparity). Second, the two views of the stereo camera may produce different sizes of the blur to the same object because of relative motion along the depth direction and camera rotation. The sharper view can help the other view to restore better by sharing its information. In this section, we analyze the above observations in details with the assumption that the stereo camera has already been rectified. **Depth-Varying Blur.** In [@xu2012depth], Xu and Jia have analyzed the relationship between blur size and depth. In Figure \[fig:depth\_view\_blur\](a), we simply restate it by only considering the relative translation parallel to the image plane $I$. According to the similar triangles theorem: $$\label{eq:depth_blur}
{\Delta X}/{\Delta P} = {f}/{z},$$ in which $\Delta X$, $\Delta P$, $f$ and $z$ denote the size of blur, the motion of object point, focal length, and depth of object point, respectively. Eq. \[eq:depth\_blur\] shows that blur size $\Delta X$ is inversely proportional to depth $z$ if motion $\Delta P$ is fixed, which means that the closer object will generate the larger blur. **View-Varying Blur.** For the stereo setups, the relative movements between the object point $P$ and two lens of stereo camera are different because the point $P$ is captured from different viewpoints. These differences make the object exhibit different blurs under the two views. Here, we consider two scenarios: relative translation along depth direction and rotation. For translation, we assume the object point $P$ moves from $P_t$ to $P_{t+1}$ along the depth direction in Figure \[fig:depth\_view\_blur\](b). According to the similar triangles theorem: $$\label{eq:view_blur_trans}
{\Delta X_L}/{\Delta X_R}={\overline{P_tM}}/{\overline{P_tN}}={h}/{(h+b)},$$ where $b$ is the baseline of the stereo camera and $h$ is the distance between left camera $C_L$ and line $\overline{P_tP_{t+1}}$. It demonstrates that the blur sizes for two views of a stereo camera are different due to relative translation in depth direction. As to relative rotation in Figure \[fig:depth\_view\_blur\](c), the velocities of two lens $v_{C_L}, v_{C_R}$ of the stereo camera are proportional to the corresponding radiuses of the rotation $\overline{C_LO}$, $\overline{C_RO}$: $$\label{eq:view_blur_rot}
{v_{C_L}}/{v_{C_R}}={\overline{C_LO}}/{\overline{C_RO}}.$$ In addition, the directions of the velocities are different due to relative rotation. As a result, both the size and direction of the blur vary between two views. The proposed network can utilize the information from the clearer view to help restore a better image for the more blurry one.
\[fig:pipeline\]
Network Architecture
--------------------
The overall pipeline of the proposed *DAVANet* is illustrated in Figure \[fig:pipeline\]. It consists of three sub-networks: *DeblurNet* for single-image deblurring, *DispBiNet* for bidirectional disparities estimation and *FusionNet* for fusing depth and two-view informations in an adaptive selection manner. Note that we adopt small convolution filters $(3\times3)$ to construct these three sub-networks and find that using the large filters does not significantly improve the performance. **DeblurNet.** The U-Net based structure of *DeblurNet* is shown in Figure \[fig:networks\](a). We use the basic residual block as the building block, which has been proved effectiveness in deblurring [@nah2017deep; @tao2018scale]. The encoder outputs features with $\frac{1}{4}\times\frac{1}{4}$ of the input size. Afterward, the following decoder reconstructs the sharp image with full resolution via two upsampled residual blocks. The skip-connections between corresponding feature maps are used between encoder and decoder. In addition, we also adopt a residual connection between the input and output. which makes it easy for the network to estimate the residual between blurry-sharp image pair and maintains color consistency. To enlarge the receptive field and obtain the multi-scale information, the scale-recurrent scheme is popularly adopted in [@nah2017deep; @tao2018scale]. Despite their performance improvement, they greatly increase the complexity of time and space. To solve this, we employ the two atrous residual blocks and a *Context Module* between encoder and decoder to obtain richer features. The *Context module* will be described in later a section. It should be noted that the *DeblurNet* uses shared weights for both views. **DispBiNet.** Inspired by DispNet [@mayer2016large] structure, we propose a small *DispBiNet* as shown in Figure \[fig:networks\](b). Different from DispNet, the proposed *DispBiNet* can predict bidirectional disparities in one forward process. The bidirectional prediction has been proved better than unidirectional prediction in scene flow estimation [@ilg2018occlusions]. The output is the full resolution with three times downsample and upsample in this network. In addtion, the residual block, atrous residual block, and context module are also used in *DispBiNet*.
\[fig:networks\]
**Context Module.** To embed the multi-scale features, we propose the *Context Module* (a slightly modified version of ASPP [@chen2018deeplab]) for *DeblurNet* and *DispBiNet*, which contains parallel dilated convolutions with different dilated rates, as show in Figure \[fig:networks\]. The four dilated rates are set to: $1, 2, 3, 4$. *Context Module* fuses richer hierarchical context information that benefit both blur removal and disparity estimation. **Fusion Network.** To exploit depth and two-view information for deblurring, we introduce the fusion network *FusionNet* to enrich the features with the disparities and the two views. For simplicity, we take left image as reference in this sections. As shown in Figure \[fig:fusion\], *FusionNet* takes the original stereo images $I^L, I^R$, the estimated disparity of left view $D^L$, features $F^D$ of the second last layer of *DispBiNet* and features $F^L, F^R$ from *DeblurNet* encoder as input in order to generate the fused features $F_{fuse}^L$. For two-view aggregation, the estimated left-view disparity $D^L$ is used to warp right-view features $F^R$ of *DeblurNet* to the left view, denoted as $W^L(F^R)$. Instead of directly concatenating $W^L(F^R)$ and $F^L$, the sub-network *GateNet* is employed to generate a soft gate map $G^L$ ranging from 0 to 1. The gate map can be utilized to fuse features $F^L$ and $W^L(F^R)$ in an adaptive selection scheme, that is, it selects helpful features and rejects incorrect ones from the other view. For example, at occlusion or false disparity regions, the values in the gate map tend to be 0, which suggest that only the features of reference view $F^L$ should be adopted. The *GateNet* consists of five convolutional layers as shown in Figure \[fig:fusion\]. Its input is absolute difference of input left image $I^L$ and the warped right image $W^L(I^R)$, namely $\left| I^L - W^L(I^R)\right|$, and the output is a single channel gate map. All feature channels share the same gate map to generate the aggregated features: $$F_{views}^L = F^L\odot(1-G^L)+W^L(F^R)\odot G^L,$$ where $\odot$ denotes element-wise multiplication.
\[fig:fusion\]
For depth awareness, a sub-network *DepthAwareNet* containing three convolutional layers is employed, and note that this sub-network is not shared by both views. Given the disparity $D^L$ and the second last layer features $F^D$ of *DispBiNet*, *DepthAwareNet-left* produces the depth-involved features $F_{depth}^L$. In fact, *DepthAwareNet* learns the depth-aware prior implicitly, which helps for dynamic scene blur removal. Finally, we concatenate the original left-view features $F^L$, view-aggregated features $F_{views}^L$, and depth-aware features $F_{depth}^L$ to generate the fused left-view features $F_{fuse}^L$. And then, we feed the $F_{fuse}^L$ to the decoder of *DeblurNet*. Note that the fusion processings of two views are the same.
Losses
------
**Deblurring Losses.** For Deblurring, we consider two loss functions to measure the difference between the restored image $\hat{I}$ and sharp image $I$ for both two views $L, R$. The first loss is MSE loss: $$\label{eq:mse_loss}
\mathcal{L}_{mse} = \frac{1}{2CHW}\sum _{k \in \{L, R\}} || \hat{I}^k- I^k ||^2,$$ where $C, H, W$ are dimensions of image. The other loss function is perceptual loss proposed in [@johnson2016perceptual], which is defined as the $l_2$-norm between the VGG-19 [@simonyan2015very] features of restored image $\hat{I}$ and sharp image $I$: $$\label{eq:percept_loss}
\mathcal{L}_{perceptual} = \frac{1}{2\mathcal{C}_j\mathcal{H}_j\mathcal{W}_j}\sum _{k \in \{L, R\}} ||\Phi_j(\hat{I}^k) - \Phi_j(I^k)||^2,$$ where $\mathcal{C}_j, \mathcal{H}_j, \mathcal{W}_j$ are dimensions of the features, and $\Phi_j(\cdot)$ denotes the features from the $j$-th convolution layer within the pretrained VGG-19 network. In our work we use the features from conv3-3 layer ($j$=15). The overall loss function for deblurring is: $$\label{eq:deblur_loss}
\mathcal{L}_{deblur} = \sum \limits_{k \in \{L, R\}} w_1 \mathcal{L}_{mse}^k + w_2 \mathcal{L}_{perceptual}^k,$$ where the weights $w_1, w_2$ of two losses are set to $1, 0.01$ in our experiments, respectively. **Disparity Estimation Loss.** For training *DispBiNet*, we consider MSE loss between estimated disparities $\hat{D}$ and ground truth $D$ at multiple scales and remove the invalid and occlusion regions with mask map $M$: $$\label{eq:disp_loss}
\mathcal{L}_{disp} = \sum \limits_{k \in \{L, R\}} \sum \limits_{i = 1} ^ m \frac{1}{H_iW_i} ||\hat{D}_i^k - D_i^k||^2\odot M_i^k,$$ where $m$ is the number of scales of the network and the loss at each scale $i$ is normalized.
Stereo Blur Dataset {#sec:data}
===================
Currently, there is no dataset specially designed for stereo image deblurring. Therefore, to train our network and verify its effectiveness, we propose a large-scale, multi-scene and depth-varying stereo blur dataset. It consists of a wide variety of scenarios, both indoor and outdoor. The indoor scenarios collect objects and persons, which usually with small depth. The outdoor scenarios include pedestrians, moving traffic and boats as well as natural landscapes. Moreover, we have diversified the dataset by considering various factors including illumination and weather. In the meantime, we have different photograph fashions including handheld shots, fixed shots, and onboard shots, to cover diverse motion patterns. Inspired by the dynamic scene blur image generation method in [@nah2017deep; @su2017deep; @hirsch2011fast], we average a sharp high frame rate sequence to generate a blurry image to approximate a long exposure. In practice, we use the ZED stereo camera [@stereolabs] to capture our data, which has the highest frame rate (60 fps) among the available stereo cameras. However, the frame rate is still not high enough to synthesize look-realistic blur, without generating undesired artifacts which exist in GOPRO dataset [@nah2017deep]. Therefore, we increase the video frame rate to 480 fps using a fast and high-quality frame interpolation method proposed in [@niklaus2017iccv]. Then, we average the varying number (17, 33, 49) of successive frames to generate different blur in size, which is temporally centered on a real-captured sharp frame (ground truth frame). For the synthesis, both two views of the stereo video have the same settings. In addition, to explore how the depth information helps with deblurring, our dataset also provides the corresponding bidirectional disparity of two views, acquired from a ZED camera. We also present the mask map for removing the invalid values in disparity ground truth and occlusion regions obtained by bidirectional consistency check [@sundaram2010dense]. In total, we collect 135 diverse real-world sequences of dynamic scenes. The dataset consists of 20,637 blurry-sharp stereo image pairs with their corresponding bidirectional disparities at $1280 \times 720$ resolution. We divide the dataset into 98 training sequences (17,319 samples) and 37 testing sequences (3,318 samples). The scenarios are totally different for training and testing sets, which avoids the over-fitting problem.
Experiments
===========
Implementation Details
----------------------
In our experiments, we train the proposed single and stereo image deblurring networks (i.e., *DeblurNet* and *DAVANet*) using our presented Stereo Blur Dataset. For more convincing comparison with single-image methods, we also train and evaluate *DeblurNet* on public GOPRO dataset [@nah2017deep], which contains 3,214 blurry-sharp image pairs (2,103 for training and 1,111 for evaluation). **Data Augmentation.** Despite our large dataset, we perform several data augmentation techniques to add diversity into the training data. We perform geometric transformations (randomly cropped to $256\times 256$ patches and randomly flipped vertically) and chromatic transformations (brightness, contrast and saturation are uniformly sampled within $\left[0.8, 1.2\right]$) using ColorJitter in PyTorch. To make our network robust, a Gaussian random noise from $\mathcal N(0, 0.01)$ is added to the input images. To keep the epipolar constraint of stereo images, we do not adopt any rotation and horizontal flip for data augmentation. **Training.** \[sec:training\] The overall proposed network *DAVANet* contains three sub-networks: *DeblurNet*, *DispBiNet* and *FusionNet*. We first pretrain our *DeblurNet* and *DispBiNet* on each task separately, then add *FusionNet* to the network and train them jointly as a whole. For all models, we set batch size to 2 and use the Adam [@kingma2015adam] optimizer with parameters $\beta_1 = 0.9$ and $\beta_2 = 0.999$. The initial learning rate in our experiments is set to $10^{-4}$ and decayed by 0.5 every 200k iterations. For the *DeblurNet*, we first train it on the presented dataset, where 2,000k iterations are sufficient for convergence. For the *DispBiNet*, we first train it using a subset (10,806 samples) of *FlyingThings3D* dataset. In this subset, the samples with large disparity ($>90$ $pixels$) are removed to ensure that the distribution of its disparity is the same as our dataset. Then we finetune the *DispBiNet* fully on our Stereo Blur Dataset until convergence. Finally, we jointly train the overall network on our dataset for 500k iterations.
Experimental Results
--------------------
\[tab:stereo\_psnr\_time\_size\]
-------------------------------------------------------------------------- ------------------------------------------------------------------------ --------------------------------------------------------------------------- --------------------------------------------------------------------------- ------------------------------------------------------------------------
{width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"}
\(a) Blurry image \(b) Hu *et al.* [@hu2014joint] \(c) Gong *et al.* [@gong2017motion] \(d) Nah *et al.* [@nah2017deep] \(e) Kupyn *et al.* [@kupyn2018deblurgan]
PSNR / SSIM 21.97 / 0.8196 28.18 / 0.9618 31.54 / 0.9678 28.17 / 0.9394
{width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"}
\(f) Zhang *et al.* [@zhang2018dynamic] \(g) Tao *et al.* [@tao2018scale] \(h) Ours-Single \(i) Ours-Stereo \(j) Ground Truth
32.61 / 0.9708 30.80 / 0.9732 31.08 / 0.9733 **34.97 / 0.9812** $+\infty$ / 1.0
{width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"}
\(a) Blurry image \(b) Hu *et al.* [@hu2014joint] \(c) Gong *et al.* [@gong2017motion] \(d) Nah *et al.* [@nah2017deep] \(e) Kupyn *et al.* [@kupyn2018deblurgan]
PSNR / SSIM 20.56 / 0.7664 25.00 / 0.8801 29.76 / 0.9119 27.26 / 0.8619
{width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"} {width="0.19\linewidth"}
\(f) Zhang *et al.* [@zhang2018dynamic] \(g) Tao *et al.* [@tao2018scale] \(h) Ours-Single \(i) Ours-Stereo \(j) Ground Truth
29.82 / 0.9149 30.72 / 0.9284 31.59 / 0.9364 **32.46 / 0.9445** $+\infty$ / 1.0
-------------------------------------------------------------------------- ------------------------------------------------------------------------ --------------------------------------------------------------------------- --------------------------------------------------------------------------- ------------------------------------------------------------------------
We quantitatively and qualitatively evaluate our single and stereo image deblurring networks (*DeblurNet* and *DAVANet*) on our dataset and compare them with the state-of-the-art deblurring algorithms, including conventional non-uniform deblurring algorithm [@whyte2012non], and CNN-based deblurring methods [@sun2015learning; @gong2017motion; @nah2017deep; @kupyn2018deblurgan; @zhang2018dynamic; @tao2018scale] in terms of PSNR and SSIM. To compare with other end-to-end CNN methods [@nah2017deep; @kupyn2018deblurgan; @zhang2018dynamic; @tao2018scale], we fully finetune their networks on our dataset until convergence with their released codes. For further comparison, we evaluate our single image deblurring network *DeblurNet* on GOPRO dataset [@nah2017deep] and compare it with aforementioned end-to-end CNN models. **Stereo blur dataset.** Although both [@nah2017deep] and [@tao2018scale] propose to use multi-scale recurrent scheme to improve the performance, it inevitably increases the computational cost. To solve this problem, we apply to use two atrous residual blocks and a *Context Module* to obtain the richer feature without a large network in the proposed *DeblurNet*. Table \[tab:stereo\_psnr\_time\_size\] shows that *DeblurNet* outperforms other state-of-the-art single-image deblurring algorithms under the proposed Stereo Blur Dataset. Although the proposed *DeblurNet* performs well with single view, we further evaluate the proposed stereo deblurring network *DAVANet* with other algorithms in Table \[tab:stereo\_psnr\_time\_size\]. It demonstrates that the proposed *DAVANet* performs better than the existing dynamic scene methods due to additional depth-aware and view-aggregated features. Figure \[fig:show1\] shows several examples from the our testing sets. The existing methods [@gong2017motion; @nah2017deep; @kupyn2018deblurgan; @zhang2018dynamic; @tao2018scale] cannot perfectly remove the large blur as depth information is not considered in their networks. Although depth information is used in [@hu2014joint], it is hard to estimate it accurately from a single image. In this way, their estimated blur kernels are ineffective and will introduce undesired artifacts into restored images. The proposed *DAVANet* estimates disparity considered as non-uniform prior information to handle spatially variant blur in dynamic scenes. Moreover, it also fuses two-view varying information, which provides more effective and additional information for deblurring. With depth awareness and view aggregation, Figure \[fig:show1\] shows our proposed *DAVANet* can restore sharp and artifact-free images.
------------------------------------------------------------------------- --------------------------------------------------------------------------- ------------------------------------------------------------------------------- --------------------------------------------------------------------------------- --------------------------------------------------------------------------- ----------------------------------------------------------------------- ------------------------------------------------------------------------- ----------------------------------------------------------------------------
{width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"} {width="0.118\linewidth"}
\(a) Blurry image \(b) Single \(c) w/o *DAVA* \(d) w/o *VA* \(e) Stereo (*DAVA*) \(f) Ground Truth \(g) Disparity (P) \(h) Disparity (T)
------------------------------------------------------------------------- --------------------------------------------------------------------------- ------------------------------------------------------------------------------- --------------------------------------------------------------------------------- --------------------------------------------------------------------------- ----------------------------------------------------------------------- ------------------------------------------------------------------------- ----------------------------------------------------------------------------
**GOPRO dataset.** Though our single image deblurring network *DeblurNet* performs well on our dataset, we further evaluate it on public GOPRO dataset [@nah2017deep] and compare it with the state-of-the-art CNN models. According to Table \[tab:psnr\_gopro\], the proposed *DeblurNet* with small size outperforms other algorithms in terms of PSNR and SSIM, which further demonstrates the effectiveness of *Context Module*.
\[tab:psnr\_gopro\]
**Running time and model size.** We implement our network using PyTorch platform [@pytorch]. To compare running time, we evaluate the proposed method and state-of-the-art image deblurring methods on the same server with an Intel Xeon E5 CPU and an NVIDIA Titan Xp GPU. As traditional blind or non-blind algorithms are used in [@whyte2012non; @sun2015learning; @gong2017motion], their methods are time-consuming. With GPU implementation, deep learning-based methods [@nah2017deep; @kupyn2018deblurgan; @zhang2018dynamic; @tao2018scale] are efficient. To enlarge the receptive field, multi-scale recurrent scheme and large CNN kernel size (e.g. $5\times5$) are used in [@nah2017deep; @tao2018scale]. For the same purpose, spatially variant RNNs are used in [@zhang2018dynamic]. They all lead to long computation time. We find that the proposed *Context Module*, which utilizes convolutions with different dilation rates, can embed multi-scale features and enlarge the receptive field at a low computational cost. In addition, only $3\times3$ convolutional layers are used in the proposed network which further reduces the size of network. According to Table \[tab:stereo\_psnr\_time\_size\], the proposed network is more efficiency with a small model, compared to the existing CNN-based methods.
Analysis and Discussions
------------------------
**Effectiveness of the disparity.** The proposed model *DAVANet* utilizes estimated disparities in two ways: Depth Awareness (*DA*) and View Aggregation (*VA*). To remove the effect of view aggregation, we do not warp features from the other view in the *FusionNet*, as shown in Figure \[fig:depth-analysis\](d). Furthermore, to remove the effect of both depth awareness and view aggregation, we feed two exactly the same images into the proposed network, where no depth information or disparity can be obtained, as shown in Figure \[fig:depth-analysis\](c). And we also compare the proposed *DAVANet* with the proposed single image network *DeblurNet*, as shown in Figure \[fig:depth-analysis\](b). The Figure \[fig:depth-analysis\] demonstrates that the proposed *DAVANet* with depth awareness and view aggregation performs better, using the accurate disparities provided by *DispBiNet*. **Ablation study.** The performance improvement of our proposed network should be attributed to three key components, including: *Context Module*, depth awareness, and view aggregation. To demonstrate the effectiveness of each component in the proposed networks, we evaluate the following three variant networks for controlled comparison: (a) To validate the effectiveness of the *Context Module*, we replace the *Context Module* of *DeblurNet* by the one-path convolution block with the same number of layers; (b) To remove the effect of depth information, we remove disparity loss of *DispBiNet* but keep the original input features to *DeblurNet*, where no depth information is involved. The whole network is updated by deblurring losses; (c) To remove the effect of view aggregation, we substitute the concatenation component, the view aggregated features $F_{views}^L$, with a copy of the reference view features $F^L$ in *FusionNet* (refer to Figure \[fig:fusion\] for clarification). We train these networks using the same strategy as aforementioned in Section \[sec:training\]. Table \[tab:ablation\] shows the proposed network is the best when all components are adopted.
\[tab:ablation\]
Conclusions
===========
In this paper, we present an efficient and effective end-to-end network, *DAVANet*, for stereo image deblurring. The proposed *DAVANet* benefits from depth awareness and view aggregation, where the depth and two-view information are effectively leveraged for spatially-varying blur removal in dynamic scenes. We also construct a large-scale, multi-scene and depth-varying dataset for stereo image deblurring, which consists of 20,637 blurry-sharp stereo image pairs from 135 diverse sequences. The experimental results show that our network outperforms the state-of-the-art methods in terms of accuracy, speed, and model size.
Acknowledgements
================
This work have been supported in part by the National Natural Science Foundation of China (No. 61671182 and 61872421) and Natural Science Foundation of Jiangsu Province (No. BK20180471).
[^1]: Corresponding author
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this note, we give a construction that provides a tight lower bound of $mn-1$ for the length of the shortest word in the intersection of two regular languages with state complexities $m$ and $n$.'
author:
- Thomas Ang
- Jeffrey Shallit
title: Length of the Shortest Word in the Intersection of Regular Languages
---
Introduction
============
Maslov observed that the state complexity of the intersection of two regular languages that have state complexities $m$ and $n$ has an upper bound of $mn$ [@Maslov]. One can easily verify this result using the usual cross-product construction . This means that the shortest word in such an intersection cannot be longer than $mn-1$. It is natural to wonder if this bound is the best possible, over a fixed alphabet size, for every choice of $m$ and $n$. Here we show that there is a matching lower bound.
First we define some notation. A deterministic finite automaton (DFA) is a quintuple $(Q, \Sigma, \delta, q_0, A)$ where $Q$ is the finite set of states, $\Sigma$ is the finite input alphabet, $\delta:Q\times\Sigma\rightarrow Q$ is the transition function, $q_0
\in Q$ is the initial state, and $A \subseteq Q$ is the set of accepting states. For a DFA $M$, $L(M)$ denotes the language accepted by $M$. For any $x \in \Sigma^*$, $|x|$ denotes the length of $x$, and $|x|_a$ for some $a \in \Sigma$ denotes the number of occurrences of $a$ in $x$. We also define two maps from nonempty languages to ${\mathbb N}$ as follows. For a nonempty language $L$, let ${\rm
lss}(L)$ denote the length of the shortest word in $L$. If $L$ is regular, then we let ${\rm sc}(L)$ denote the state complexity of $L$ (the minimal number of states in any DFA accepting $L$).
We previously stated that the upper bound on the state complexity of the intersection of two regular languages implies an upper bound the length of the shortest word in the intersection. More precisely, we have ${\rm lss}(L) < {\rm sc}(L)$, which follows directly from the pumping lemma for regular languages . So all that is left is to show that the upper bound of $mn-1$ can actually be attained for all $m$ and $n$. There is an obvious construction over a unary alphabet that works when $\gcd(m,n) = 1$: namely, set
- $L_1 = \{ x : |x| \equiv m-1 \pmod m\}$, and
- $L_2 = \{ x : |x| \equiv n-1 \pmod n\}$.
However, this construction fails when $\gcd(m,n) \neq 1$, so we provide a more general construction over a binary alphabet that works for all $m$ and $n$.
Our result
==========
\[prop:basic\] For all integers $m, n \geq 1$ there exist DFAs $M_1, M_2$ with $m$ and $n$ states, respectively, such that $L(M_1) \cap L(M_2) \not= \emptyset$, and ${\rm lss}(L(M_1) \cap L(M_2) ) = mn-1$.
The proof is constructive. Without loss of generality, assume $m \leq
n$, and set $\Sigma = \{0,1\}$. Let $M_1$ be the DFA given by $(Q_1, \Sigma,
\delta_1, p_0, A_1)$, where $Q_1 = \{p_0, p_1, p_2,\ldots, p_{m-1}\}$, $A_1 = p_0$, and for each $a$, $0 \leq a \leq m-1$, and $c \in \lbrace 0, 1 \rbrace $ we set $$\label{eq:delta1}
\delta_1(p_a, c) = p_{(a+c) \bmod m}.$$ Then $$L(M_1) = \{ x \in \Sigma^* : |x|_1 \equiv 0\!\! \pmod {m} \}.$$
Let $M_2$ be the DFA $(Q_2, \Sigma, \delta_2, q_0, A_2)$, illustrated in Figure \[fig:m2\], where $Q_2 = \{q_0, q_1, q_2,\ldots, q_{n-1}\}$, $A_2 = q_{n-1}$, and for each $a$, $0 \leq a \leq n-1$, $$\label{eq:delta2}
\delta_2(q_a, c) =
\begin{cases}
q_{a+c}, & \text{if } 0 \leq a < m-1; \\
q_{(a+1) \bmod n}, & \text{if }c = 0 \text{ and } m-1 \leq a \leq n-1; \\
q_0, & \text{if }c = 1 \text{ and } m-1 \leq a \leq n-1.
\end{cases} \nonumber$$
Focussing solely on the $1$’s that appear in some accepting path in $M_2$, we see that we can return to $q_0$
- via a simple path with $m$ $1$’s, or
- (if we go through $q_{n-1}$), via a simple path with $m-1$ $1$’s and ending in the transition $\delta(q_{n-1}, 0) = q_0$.
After some number of cycles through $q_0$, we eventually arrive at $q_{n-1}$. Letting $i$ denote the number of times a path of type (b) is chosen (including the last path that arrives at $q_{n-1}$) and $j$ denote the number of times a path of type (a) is chosen, we see that the number of $1$’s in any accepted word must be of the form $i(m-1) + jm$, with $i > 0$, $j \geq 0$. The number of $0$’s along such a path is then at least $i(n-m+1) - 1$, with the $-1$ in this expression arising from the fact that the last part of the path terminates at $q_{n-1}$ without taking an additional $0$ transition back to $q_0$.
Thus $$\begin{aligned}
L(M_2) \subseteq \{ x \in \Sigma^* : \exists i,j \in {\mathbb N}, \text{ such that } i > 0, j \geq 0, \text{ and } \\
|x|_1 = i(m-1) + jm, \ |x|_0 \geq i(n-m+1)-1 \}.\end{aligned}$$ Furthermore, for every $i,j \in {\mathbb N}, \text{ such that } i > 0, j \geq 0$, there exists an $x \in L(M_2)$ such that $|x|_1 = i(m-1) + jm$, and $|x|_0 = i(n-m+1)-1$. This is obtained, for example, by cycling $j$ times from $q_0$ to $q_{m-1}$ and then back to $q_0$ via a transition on $1$, then $j-1$ times from $q_0$ to $q_{n-1}$ and then back to $q_0$ via a transition on $0$, and finally one more time from $q_0$ to $q_{n-1}$.
It follows then that $$\begin{aligned}
L(M_1 \cap M_2) &\subseteq \{ x \in \Sigma^*: \exists i,j \in {\mathbb N}, \text{ such that } i > 0, j \geq 0, \text{ and } \\
& |x|_1 = i(m-1) + jm, \ |x|_0 \geq i(n-m+1)-1\\
& \text{and } i(m-1) + jm \equiv 0\!\! \pmod {m} \}.$$ Further, for every such $i$ and $j$, there exists a corresponding element in $L(M_1 \cap M_2)$. Since $m-1$ and $m$ are relatively prime, the shortest such word corresponds to $i = m$, $j = 0$, and satisfies $|x|_0 = m(n-m+1)-1$. In particular, a shortest accepted word is $(1^{m-1}0^{n-m+1})^{m-1}1^{m-1}0^{n-m}$, which is of length $mn-1$.
It is natural to try to extend the construction to an arbitrary number of DFAs. However, we have found empirically that, over a two-letter alphabet, the corresponding bound $mnp-1$ for three DFA’s does not always hold. For example, there are no DFA’s of $2, 2, $ and $3$ states for which the shortest word in the intersection is of length $2\cdot 2\cdot 3 - 1$.
[10]{}
Hopcroft, J.E., Ullman, J.D.: Introduction to Automata Theory, Languages, and Computation. Addison-Wesley (1979).
Maslov, A.N.: Estimates of the number of states of finite automata. Dokl. Akad. Navk. SSSR. **194** (1970) 1266–1268. In Russian. English translation in Soviet Math. Dokl. **11** (1970) 1373–1375.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Block coordinate descent (BCD) methods approach optimization problems by performing gradient steps along alternating subgroups of coordinates. This is in contrast to full gradient descent, where a gradient step updates all coordinates simultaneously. BCD has been demonstrated to accelerate the gradient method in many practical large-scale applications. Despite its success no convergence analysis for inverse problems is known so far. In this paper, we investigate the BCD method for solving linear inverse problems. As main theoretical result, we show that for operators having a particular tensor product form, the BCD method combined with an appropriate stopping criterion yields a convergent regularization method. To illustrate the theory, we perform numerical experiments comparing the BCD and the full gradient descent method for a system of integral equations. We also present numerical tests for a non-linear inverse problem not covered by our theory, namely one-step inversion in multi-spectral X-ray tomography.
**Keywords:** ill-posed problems, convergence analysis, regularization theory, coordinate descent, multi-spectral CT
**MSC2010:** 65J20, 44A12, 47J06.
author:
- Simon Rabanser
- Lukas Neumann
- Markus Haltmeier
title: 'Analysis of the Block Coordinate Descent Method for [Linear]{} Ill-Posed Problems'
---
Introduction {#sec:intro}
============
We consider the solution of inverse problems of the form $$\label{IP}
y^\delta = {\mathcal A}(x[1], \dots, x[{B}] ) + z$$ by block coordinate gradient descent (BCD) methods. Here ${\mathcal A}\colon {\mathcal X}\to {\mathcal Y}$ is a [linear]{} forward operator between Hilbert spaces ${\mathcal X}= {X}_1 \times \cdots \times {X}_{B}$ and ${\mathcal Y}$. Moreover, $x = (x[1], \dots, x[{B}])\in{\mathcal X}$ is the vector of blocks $x[{b}] \in {X}_{{b}}$ of unknown variables, ${ y^\delta}\in{\mathcal Y}$ are the given noisy data, and $z$ denotes the data perturbation that satisfies ${\left\Vertz\right\Vert} \leq \delta$ for some noise level $\delta\geq 0$.
For many inverse problems, the individual blocks $x[{b}]$ arise in a natural manner and might correspond to $x[{b}] = f[{b}]$, where $f[{b}] \colon \Omega_{b}\to {\mathbb R}$ are functions modeling unknown spatially varying parameter distributions. The blocks might also be formed by applying domain decomposition $ \Omega =\Omega_1 \cup \Omega_2 \cup \ldots \cup\Omega_{B}$ to a single function $f \colon \Omega \to {\mathbb R}$, and defining $x[{b}] = {{f}|{\Omega_{b}}}$ as the restriction of $f$ to $\Omega_{b}$.
Iterative regularization methods
--------------------------------
The characteristic feature of inverse problems is their ill-posedness which means that the solution of is unstable with respect to data perturbations. In such a situation, one has to apply regularization methods to obtain solutions in a stable way. There are at least two basic classes of regularization methods: iterative regularization and variational regularization [@Engl96; @scherzer2009variational]. In this paper we consider iterative regularization and introduce and analyze BCD as new member of this class of regularization methods.
The most established iterative regularization approaches for inverse problems are the Landweber iteration and its variants [@kaltenbacher08; @landweber1951iteration; @hanke1995convergence; @Neubauer17] $$\label{eq:iterL}
x_{k+1}^\delta
{:=}x_k^\delta - s^\delta_k {\mathcal A}^*
{\left( {\mathcal A}(x_k^\delta) - { y }^\delta \right)} \,,$$ where $x_0^\delta {:=}x_0 \in {\mathcal X}$ is an initial guess, $s^\delta_k$ is the step size and ${\mathcal A}^*$ denotes the adjoint of ${\mathcal A}$. If the step size is taken constant, then is the Landweber iteration [@landweber1951iteration; @hanke1995convergence]. Other step size rules yield the steepest descent and the minimal error method [@neubauer1995convergence] or a more recent variant analyzed in [@Neubauer17]. Kaczmarz type variants of for systems of ill-posed equations have been analyzed in [@decesaro2008steepest; @haltmeier2007kaczmarz1; @Hal09b; @kowar2002; @leitao16projective; @li2018averaged]. Kaczmarz methods make use of a product structure of the image space ${\mathcal Y}$, and are in this sense dual to BCD methods which exploit the product structure of the pre-image space ${\mathcal X}$.
We consider the product form ${\mathcal X}= {X}_1 \times \cdots \times {X}_{B}$, [where the forward operator can be written as ${\mathcal A}= [{\mathcal A}_1 , \dots, {\mathcal A}_B]$. ]{} As a consequence, the Landweber iteration takes the form $$\label{eq:lw}
\begin{pmatrix}
x_{k+1}^\delta[1] \\
x_{k+1}^\delta[2] \\
\vdots \\
x_{k+1}^\delta[{B}]
\end{pmatrix}
=
\begin{pmatrix}
x_k^\delta[1] \\
x_k^\delta[2] \\
\vdots \\
x_k^\delta[{B}]
\end{pmatrix}
- s^\delta_k
\begin{pmatrix}
{\mathcal A}_1^* \\
{\mathcal A}_2^* \\
\vdots \\
{\mathcal A}_B^*
\end{pmatrix}
{\left( {\mathcal A}(x_k^\delta) - { y }^\delta\right)}
\,.$$ We see that each iterative update requires computing ${B}$ separate updates, one for each of the blocks.
Block coordinate descent (BCD)
------------------------------
In order to [simplify]{} the iterative update in , a natural idea is to update only a single block in each iteration. This results in the BCD iteration $$\label{eq:cd}
x_{k+1}^\delta[{b}]
{:=}x_k^\delta[{b}]
-
s^\delta_k
\begin{cases}
{\mathcal A}_{b}^*
{\left( {\mathcal A}(x_k^\delta) - { y }^\delta\right)}
& \text{ if } {b}= {b}(k)
\\
0
& \text{ otherwise } \,,
\end{cases}$$ where the control ${b}(k) \in {\left\{1, \dots, {B}\right\}}$ selects the block that is updated in the $k$th iteration. If we apply the BCD iteration to [exact data where $\delta =0$,]{} we write $x_k$ instead of $x_k^\delta$. Rigorously studying the iteration in the context of ill-posed problems is the main aim of this paper. To guarantee convergence in the noisy case we will slightly modify the update rule of the BCD iteration by including a loping strategy which skips the $k$th iterative step if a certain residual term is sufficiently small (see Definition \[def:loping\]). [Under the reasonable assumption that the complexity of evaluating ${\mathcal A}$ is essentially $B$-times the complexity $M$ of evaluating ${\mathcal A}^*_{b}$, then one step of the Landweber Method has complexity $\mathcal{O} ( 2BM) $, whereas one step of the BCD method has complexity $\mathcal{O} ( (B+1)M)$. For the special form of ${\mathcal A}$ considered in the following section, the complexity of one step of the BCD method even reduces to $\mathcal{O} ( 2 M )$; see Remark \[rem:complexity\]. ]{}
Note that the iteration arises by applying the block gradient [descent]{} method, well known in optimization [@beck2013convergence; @nesterov2012efficiency; @saha2013nonasymptotic; @wright2015coordinate], to the residual functional $
\frac{1}{2} \| y^\delta - {\mathcal A}(x) \|^2$. In a finite dimensional setting, BCD and other coordinate descent type methods are well studied. However, existing convergence results mostly analyze convergence in the objective value. This only implies convergence in pre-image space, if the residual functional is strongly convex. Strong convexity does not hold for ill-posed problems. Therefore, existing convergence results and methods cannot be applied to ill-posed inverse problems. [Note that removing the strict convexity assumption can also also be achieved by coupling the BCD method with a proximal term; see [@bolte2014proximal] and the references therein.]{}
To the best of our knowledge, no convergence result for in the ill-posed setting is available. As the main contribution in this paper we will present a convergence analysis of BCD applicable to the ill-posed case. We show that under assumptions specified in Section \[sec:prelim\], for operators having a particular tensor product form, the BCD iteration yields a regularization method for solving ill-posed linear problems.
Outline
-------
This paper is organized as follows. In Section \[sec:prelim\] we present the main assumptions made in this paper, derive an auxiliary results and introduce the loping strategy. In Section \[sec:analysis\] we present the convergence analysis. In the exact data case, we show that the BCD iteration converges to a solution $x^*$ of the given equation as $k \to \infty$. In the noisy data case we show that the stopping index of the loping BCD iteration is finite and the corresponding iterates converge to $x^*$ as $\delta \to 0$. To illustrate the theory, in Section \[sec:int\] [we compare the BCD method with the gradient method for a system of integral equations. Additionally, in Section \[sec:xray\] we consider a non-linear example]{} not covered by our theory, namely one-step inversion in multi-spectral X-ray tomography [@rigie2015joint; @kazantsev2018joint; @atak2015dual; @barber2016algorithm]. The paper concludes with a short discussion presented in Section \[sec:conclusion\].
Preliminaries {#sec:prelim}
=============
In this section we formulate the main assumptions and derive basic results that we will use in the convergence analysis presented in Section \[sec:analysis\].
Note that for any Hilbert space ${X}$ we can write ${X}^{B}\simeq {\mathbb R}^{B}\otimes {X}$. For any ${b}\in {\left\{1, \dots, {B}\right\}}$ we define the projection operators $$\label{eq:pr}
{\mathcal P}_{b}= ( e_{b}e_{b}^{\mathsf{T}}) \otimes {\operatorname{Id}}_{X}\colon {X}^{B}\to {X}^{B}\colon
\begin{pmatrix}
x[1] \\
\vdots \\
x[{b}] \\
\vdots \\
x[{B}]
\end{pmatrix}
\mapsto
\begin{pmatrix}
0 \\
\vdots \\
x[{b}] \\
\vdots \\
0
\end{pmatrix}\,,$$ where $e_{b}$ denotes the ${b}$th standard basis vector in ${\mathbb R}^{{B}}$, defined by $e_{b}[{b}] =1$ and $e_{b}[{b}'] =0$ for ${b}' \neq {b}$. Using , the BCD method can be written in the compact form $$\label{eq:iter}
x_{k+1}^\delta
{:=}x_k^\delta - s^\delta_k {\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k^\delta)-y^\delta)\,.$$ Here ${b}(k) \in \{1,\dots,{B}\}$ is the selected block at the $k$th iteration, $s^\delta_k >0 $ is the step size, and $x_0^\delta:=x_0 \in {\mathcal X}$ is some initial guess. Recall that in the case of exact data we write $x_k$ instead of $x_k^\delta$.
Main assumptions
----------------
[We note that the main difficulty we encountered in the convergence analysis of the BCD method for ill-posed problems is that even for exact data ${ y }= {\mathcal A}(x^*)$, the error ${\left\Vertx_k - x^*\right\Vert}$ is not monotonically decreasing, except for some very special cases. This can be easily verified for linear operators in ${\mathbb R}^{B}$. On the other hand, the BCD is monotonically decreasing in the objective value, which is used in existing convergence theory for optimization problems [@beck2013convergence; @nesterov2012efficiency; @saha2013nonasymptotic; @wright2015coordinate]. However, this cannot be used directly for the convergence analysis in the ill-posed setting where the value of the residual functional gives no bounds for the error ${\left\Vertx_k - x^*\right\Vert}$. ]{}
[We present a complete convergence analysis under the following assumption that allows to separate the difficulties due to the ill-posedness and due to the non-monotonicity. ]{}
1. \[ass1\] ${\mathcal X}$, ${\mathcal Y}$ are Hilbert spaces of the form ${\mathcal X}= {X}^{B}$, [${\mathcal Y}= {Y}^{D}$ with ${D}, {B}\in {\mathbb N}$.]{}
2. \[ass2\] ${\mathcal A}\colon {\mathcal X}\to {\mathcal Y}$ has the form ${\mathcal A}= {V}\otimes {K}$, where
- [${K}\colon {X}\to {Y}$ is bounded linear]{};
- [${V}\in {\mathbb R}^{{D}\times {B}}$ has rank ${B}$ and non-vanishing columns $v_{b}\in {\mathbb R}^{D}$;]{}
3. \[ass4\] The control $b \colon {\mathbb N}\to {\left\{1, \dots, B\right\}}$ satisfies\
$\exists p \in {\mathbb N}\; \forall k \in {\mathbb N}\colon
{\left\{{b}(k), \dots, {b}(k + p-1) \right\}} = {\left\{1, \dots, {B}\right\}}$.
Let us introduce the operators [$$\begin{aligned}
&{\mathcal K}_{{B}} {:=}{\operatorname{Id}}_{{\mathbb R}^{B}} \otimes \, {K}\colon {\mathcal X}\to {\mathcal Y}\colon\begin{pmatrix}
x[1] \\ \vdots \\ x[{B}]
\end{pmatrix}
\mapsto
\begin{pmatrix}
{K}(x[1]) \\ \vdots \\ {K}(x[{B}])
\end{pmatrix}
\\
& {\mathcal V}_{Y}{:=}{V}\otimes {\operatorname{Id}}_{{Y}} \colon {\mathcal Y}\to {\mathcal Y}\colon
y
\mapsto
\sum_{{b}=1}^{B}v_{b}\, y[{b}] \,.$$ In a similar manner we denote ${\mathcal K}_{{D}} {:=}{\operatorname{Id}}_{{\mathbb R}^{D}} \otimes \, {K}$ and $ {\mathcal V}_{X}{:=}{V}\otimes {\operatorname{Id}}_{{X}}$. Then we have ${\mathcal A}= {\mathcal V}_{Y}\circ {\mathcal K}_{{B}} = {\mathcal K}_{{D}} \circ {\mathcal V}_{X}$.]{}
To overcome the above mentioned obstacles in the convergence analysis we will study the auxiliary sequence $({\mathcal V}_{X}x_k^\delta )_{k\in {\mathbb N}}$ which, by linearity, satisfies $$\begin{gathered}
\label{eq:iter1}
{\mathcal V}_{X}x_{k+1}^\delta
=
{\mathcal V}_{X}x_k^\delta - s^\delta_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k^\delta)-{ y^\delta})
\\
={\mathcal V}_{X}x_k^\delta - s^\delta_k {\Vertv_{{b}(k)}\Vert}^2 {\mathcal Q}^{{X}}_{{b}(k)} {\mathcal K}_{{D}}^*({\mathcal A}(x_k^\delta)-{ y^\delta})
\,.\end{gathered}$$ Here we have set [$${\mathcal Q}^{{X}}_{b}{:=}\frac{1}{{\left\Vertv_{b}\right\Vert}^2}(v_{b}v_{b}^{\mathsf{T}})
\otimes {\operatorname{Id}}_{X}\colon {\mathcal X}\to {\mathcal X}\,.$$ We will also use the notation ${\mathcal Q}^{{Y}}_{b}{:=}{\Vertv_{b}\Vert}^{-2} (v_{b}v_{b}^{\mathsf{T}})
\otimes {\operatorname{Id}}_{Y}$.]{} As an important auxiliary result we will show monotonicity for $({\mathcal V}_{X}x_k^\delta )_{k\in {\mathbb N}}$. This allows us to show that the BCD method combined with a loping strategy is a convergent regularization method. In fact, this is the reason for requiring the forward operator ${\mathcal A}$ to have the particular tensor product form specified in assumption \[ass2\]. The convergence analysis in the more general setting is still an open and challenging problem.
[Note that the assumption $\operatorname{rank} ({V}) = {B}$ is only necessary for the convergence of $({\mathcal V}_{X}x_k)_{k\in {\mathbb N}}$ implying convergence of $(x_k )_{k\in {\mathbb N}}$. In the case that ${V}$ has arbitrary rank, the main convergence results still hold true for the semi-norm ${\left\Vert{\mathcal V}_{X}({\,\cdot \,})\right\Vert}$ in place of the norm ${\left\Vert{\,\cdot \,}\right\Vert}$. ]{}
\[rem:complexity\] For the considered form ${\mathcal A}= {V}\otimes {K}$ and a cyclic control ${b}(k) = ((k-1) \operatorname{mod} B)+1$, one cycle of updates with the BCD method for $k \in {\left\{\ell B, \dots , (\ell+1) B -1\right\}}$ has essentially the same numerical complexity as one iteration with the Landweber iteration. To see this, we implement the BCD method in the following manner:
1. Initialization: $\forall b = 1, \dots, B$ do
- $x_{\rm BCD}[{b}] \gets x_0[{b}]$
- $h_{\rm BCD}[{b}] \gets {K}(x_{\rm BCD}[{b}])$.
2. Updates: $\forall i_0 = 1, \dots, N_{\rm cycle} \forall b = 1, \dots, B$ do
- $x_{\rm BCD}[b] \gets x_{\rm BCD}[b] - s_k {K}^*( ({\mathcal V}_{Y}^* ( {\mathcal V}_{Y}h_{\rm BCD} - y^\delta ))[b] )$
- $h_{\rm BCD}[b] \gets {K}( x_{\rm BCD}[b] )$.
Complexity of the above procedure is dominate by the evaluation of ${K}$, ${K}^*$ and the evaluation of ${\mathcal V}_{Y}$, ${\mathcal V}_{Y}^*$. Unless $B$ is very large (or evaluating ${K}$, ${K}^*$ is cheap), for typical inverse problems, the dominating parts are ${K}$, ${K}^*$. This shows that the complexity of one cycle of the BCD iteration in fact is similar to the complexity of one iteration of the Landweber iteration.
Monotonicity
------------
The following lemma is an important auxiliary result, which will be used at several places throughout this article.
\[lem:mon\] Let $x^* \in {\mathcal X}$ satisfy ${\mathcal A}(x^*) = y$ and set $$\label{eq:res}
{r}^\delta_k
{:=}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(k)} {\left( { y^\delta}-{\mathcal A}(x_k^\delta) \right)} \right\Vert}
\,.$$ Then, the following estimate holds: $$\begin{gathered}
\label{eq:mon}
\frac{1}{2}{\left\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\right\Vert}^2 - \frac{1}{2}{\left\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\right\Vert}^2
\leq
- s^\delta_k {r}^\delta_k {\Vertv_{{b}(k)}\Vert}^2 {\left( {r}^\delta_k - \delta_{{b}(k)} \right)}
\\
+\frac{(s^\delta_k)^2}{2} \, {\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*{\left( { y^\delta}- {\mathcal A}(x_k^\delta)\right)}\right\Vert}^2 \,.\end{gathered}$$ In particular, if ${\Vert{\mathcal Q}^{{Y}}_{{b}} (y-y^\delta)\Vert} \leq \delta_{b}$ and ${r}^\delta_k \geq \delta_{{b}(k)} $ and if the step size is chosen such that $$\label{eq:stepsize}
0 \leq s^\delta_k\leq \frac{2 {r}^\delta_k {\Vertv_{{b}(k)}\Vert}^2 \left({r}^\delta_k -\delta_{{b}(k)} \right)}{{\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*({ y^\delta}- {\mathcal A}(x_k^\delta))\right\Vert}^2}\,,$$ then ${\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\Vert}^2 \leq {\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}^2$.
Equation implies $$\begin{gathered}
\label{eq:monaux}
\frac{1}{2}{\left\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\right\Vert}^2 - \frac{1}{2}{\left\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\right\Vert}^2 \leq
{\left\langle {\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^* , {\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x_k^\delta\right\rangle} \\
+\frac{(s^\delta_k)^2}{2} \, {\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*{\left( { y^\delta}- {\mathcal A}(x_k^\delta)\right)}\right\Vert}^2 \,.\end{gathered}$$ We have $$\begin{aligned}
&{\left\langle{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^* ,{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x_k^\delta\right\rangle}
\\ & =
s^\delta_k {\Vertv_{{b}(k)}\Vert}^2
{\left\langle {\mathcal V}_{X}(x_k^\delta - x^* ) , {\mathcal Q}^{{Y}}_{{b}(k)} {\mathcal K}_{{D}}^*
{\left( { y }^\delta - {\mathcal A}(x_k^\delta) \right)} \right\rangle}
\\ & =
s^\delta_k {\Vertv_{{b}(k)}\Vert}^2
{\left\langle {\mathcal K}_{{B}} {\mathcal V}_{X}(x_k^\delta - x^* ) , {\mathcal Q}^{{Y}}_{{b}(k)}
{\left( { y }^\delta - {\mathcal A}(x_k^\delta) \right)} \right\rangle}
\\ & =
s^\delta_k {\Vertv_{{b}(k)}\Vert}^2
{\left\langle {\mathcal A}(x_k^\delta) - {\mathcal A}(x^*) , {\mathcal Q}^{{Y}}_{{b}(k)}
{\left( { y }^\delta - {\mathcal A}(x_k^\delta) \right)} \right\rangle}
\\ & =
s^\delta_k {\Vertv_{{b}(k)}\Vert}^2
{\left\langle {\mathcal A}(x_k^\delta) - { y }^\delta + { y }^\delta - {\mathcal A}(x^*) , {\mathcal Q}^{{Y}}_{{b}(k)}
{\left( { y }^\delta - {\mathcal A}(x_k^\delta) \right)} \right\rangle}
\\ & \leq
s^\delta_k {\Vertv_{{b}(k)}\Vert}^2
{\left(
- {\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({ y }^\delta - {\mathcal A}(x_k^\delta) ) \Vert}^2
+ \delta_{{b}(k)} {\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({ y }^\delta - {\mathcal A}(x_k^\delta) ) \Vert}
\right)}
\end{aligned}$$ By combining with the above estimate, we obtain $$\begin{gathered}
\frac{1}{2}{\left\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\right\Vert}^2 - \frac{1}{2}{\left\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\right\Vert}^2
\leq
s^\delta_k {r}^\delta_k {\Vertv_{{b}(k)}\Vert}^2 \Bigl( \delta_{{b}(k)}
- {r}^\delta_k \Bigr)
\\
+\frac{(s^\delta_k)^2}{2} \, {\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*{\left( { y^\delta}- {\mathcal A}(x_k^\delta)\right)}\right\Vert}^2 \,,\end{gathered}$$ which is the desired estimate . If $s^\delta_k$ is chosen according to , then the right hand side in inequality is less or equal to 0, which implies ${\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\Vert}^2 \leq {\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}^2$.
Loping BCD and discrepancy principle
------------------------------------
From Lemma \[lem:mon\] we see that if the residual term ${r}^\delta_k = {\Vert{\mathcal Q}^{{Y}}_{{b}(k)} {( { y^\delta}-{\mathcal A}(x_k^\delta) )} \Vert} $ satisfies , then the error ${\Vert {\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}$ is decreasing. In the case that does not hold, then an iterative update might increase the value of ${\Vert {\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}$. Following a similar strategy introduced in [@haltmeier2007kaczmarz1; @decesaro2008steepest] for Kaczmarz type iterative method we therefore modify by introducing a loping strategy as follows.
We\[def:loping\] define the loping BCD method by $$\begin{aligned}
\label{eq:iterS1}
x_{k+1}^\delta
&{:=}x_k^\delta - d^\delta_k s^\delta_k {\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k^\delta)-{ y^\delta})
\\ \label{eq:iterS2}
d^\delta_k
&{:=}\begin{cases}
1
& \text{ if }
{r}^\delta_k \geq \tau \delta_{{b}(k)}
\\
0 & \text{ otherwise }\,,
\end{cases}\end{aligned}$$ where ${r}^\delta_k= {\Vert{\mathcal Q}^{{Y}}_{{b}(k)} ({ y^\delta}-{\mathcal A}( x_{k}^\delta))\Vert} $ is as in Equation , and $$\label{eq:tau}
{{}\tau > 1 }\,.$$
In the case of exact data, we have $d^\delta_k =1$ and the loping BCD iteration reduces to the standard BCD. In the noisy data case the loping parameters $d^\delta_k$ ensure that no update is made if we cannot guarantee that an update would decrease ${\Vert {\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}$. Note that the choice of $\tau$ as in implies that condition is satisfied whenever we have $d^\delta_k =1$. For the loping BCD, Lemma \[lem:mon\] therefore implies that the error term ${\Vert {\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\Vert}$ is in fact monotonically decreasing. Moreover, we can show the following.
\[conv1\] Let $x^* \in {\mathcal X}$ satisfy ${\mathcal A}(x^*) = y$. Then the residuals ${r}^\delta_k {:=}{\Vert {\mathcal Q}^{{Y}}_{{b}(k)}({ y^\delta}- {\mathcal A}(x_k^\delta))\Vert}$ of the loping BCD iteration , satisfy $$\label{SumRes}
\sum_{k\in {\mathbb N}}
d^\delta_k s^\delta_k{\Vertv_{{b}(k)}\Vert}^2 ({r}^\delta_k)^2 \leq
\frac{{\left\Vert{\mathcal V}_{X}x_0 - {\mathcal V}_{X}x^*\right\Vert}^2}{\gamma_{\rm min} (2- \theta_{\rm max})}\,,$$ where, $s^\delta_k$, $\gamma_{\rm min}$, $\theta_{\rm max}$ are chosen such that
1. \[sum1\] $\forall k \in {\mathbb N}\colon d^\delta_k =1 \Rightarrow s^\delta_k \in (0, 2 A_k^\delta) $ with $ A_k^\delta {:=}\frac{ {\Vertv_{{b}(k)}\Vert}^2 {r}^\delta_k( {r}^\delta_k -\delta_{{b}(k)} )}{ {\left\Vert {\mathcal V}_{X}{\mathcal P}_{{b}(k)}
{\mathcal A}^*(
{ y^\delta}- {\mathcal A}(x_k^\delta) )\right\Vert}^2}$;
2. \[sum2\] $\forall k \in {\mathbb N}\colon d^\delta_k =1 \Rightarrow \theta_k {:=}s^\delta_k / A_k^\delta \leq \theta_{\rm max} < 2$;
3. \[sum3\] $1-1 / \tau \geq \gamma_{\rm min} >0 $.
We first show $$\begin{gathered}
\label{eq:saux}
{\left\Vert{\mathcal V}_{X}x_k^\delta - {\mathcal V}_{X}x^*\right\Vert}^2 - {\left\Vert{\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\right\Vert}^2
\\ \geq
(2- \theta_{\rm max}) d^\delta_k s^\delta_k
{\Vertv_{{b}(k)}\Vert}^2({r}^\delta_k)^2
{\left( 1 - 1/\tau \right)} \,.
\end{gathered}$$ If ${r}^\delta_k < \tau \delta$, then $d^\delta_k = 0$ and $x_{k+1}^\delta = x_k^\delta$ and therefore holds with equality. If ${r}^\delta_k \geq \tau \delta$, application of Lemma \[lem:mon\], \[sum2\] and \[sum1\] yield $$\begin{aligned}
\lVert {\mathcal V}_{X}x_k^\delta & - {\mathcal V}_{X}x^*\rVert^2
-
\lVert {\mathcal V}_{X}x_{k+1}^\delta - {\mathcal V}_{X}x^*\rVert^2 \\
&\geq
2s^\delta_k {\Vertv_{{b}(k)}\Vert}^2{r}^\delta_k (-\delta_{{b}(k)} +{r}^\delta_k) -(s^\delta_k)^2 {\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*({ y^\delta}- {\mathcal A}(x_k^\delta))\Vert}^2
\\& \geq
2s^\delta_k {\Vertv_{{b}(k)}\Vert}^2{r}^\delta_k (-\delta_{{b}(k)} + {r}^\delta_k ) -s^\delta_k \theta_{\rm max} A_k^\delta {\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(k)}{\mathcal A}^*({ y^\delta}- {\mathcal A}(x_k^\delta))\Vert}^2
\\ &
= 2s^\delta_k {\Vertv_{{b}(k)}\Vert}^2{r}^\delta_k (-\delta_{{b}(k)} + {r}^\delta_k ) -s^\delta_k \theta_{\rm max} {\Vertv_{{b}(k)}\Vert}^2{r}^\delta_k (-\delta_{{b}(k)} + {r}^\delta_k )
\\ &
=
(2- \theta_{\rm max})s^\delta_k{\Vertv_{{b}(k)}\Vert}^2{r}^\delta_k ( {r}^\delta_k
-\delta_{{b}(k)} )
\\ &
\geq
(2- \theta_{\rm max})s^\delta_k{\Vertv_{{b}(k)}\Vert}^2
({r}^\delta_k)^2
{\left( 1 -1/\tau \right)} \,.\end{aligned}$$ This shows with $d^\delta_k = 1$ in .
Summing over all $k \in {\mathbb N}$ and using \[sum3\] we obtain $${\left\Vert{\mathcal V}_{X}x_0 - {\mathcal V}_{X}x^*\right\Vert}^2 \geq (2-\theta_{\rm max})\gamma_{\rm min}
\sum_{k \in {\mathbb N}} d_k s^\delta_k{\Vertv_{{b}(k)}\Vert}^2 ({r}^\delta_k)^2 \,,$$ which shows after dividing by $(2-\theta_{\rm max})\gamma_{\rm min}$.
Note the conditions for the step sizes in Lemma \[conv1\] are inspired by [@Neubauer17], where a new step size rule for the gradient method for ill-posed problems has been introduced. From the definitions of ${r}^\delta_k, d^\delta_k$ we obtain ${r}^\delta_k - \delta_{{b}(k)} \geq = (1-1/\tau) {r}^\delta_k $. Moreover, recall that ${\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal V}_{X}^* = {\Vertv_{{b}(k)}\Vert}^2 {\mathcal Q}^{{X}}_{{b}(k)}$. Consequently, $$\begin{gathered}
A_k^\delta
=
\frac{ {\Vertv_{{b}(k)}\Vert}^2 {r}^\delta_k( {r}^\delta_k -\delta_{{b}(k)} ) }{{\Vert {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal V}_{X}^*
{\mathcal K}_{{D}}^*(
{ y^\delta}- {\mathcal A}(x_k^\delta) )\Vert}^2}
\geq
{\left( 1 - \frac{1}{\tau} \right)}
\frac{ {\Vert{\mathcal Q}^{{Y}}_{{b}(k)} ({ y^\delta}-{\mathcal A}( x_{k}^\delta))\Vert}^2 }
{{\Vert
{\mathcal Q}^{{X}}_{{b}(k)} {\mathcal K}_{{D}}^* (
{ y^\delta}- {\mathcal A}(x_k^\delta) )\Vert}^2} \\
\geq
\gamma_{\rm min} \frac{ {\Vert{\mathcal Q}^{{Y}}_{{b}(k)} ({ y^\delta}-{\mathcal A}( x_{k}^\delta))\Vert}^2 }
{{\Vert
{\mathcal K}_{{B}}^* {\mathcal Q}^{{Y}}_{{b}(k)} (
{ y^\delta}- {\mathcal A}(x_k^\delta) )\Vert}^2} \geq
\frac{ \gamma_{\rm min}}
{{\Vert{\mathcal K}_{{B}}^*\Vert}^2}
\,. \end{gathered}$$ This implies that we can choose the step sizes bounded from below. In particular, holds for any constant step size choice $s_k^\delta = s_\star \in (0, \gamma_{\rm min} / {{\Vert{\mathcal K}_{{B}}^*\Vert}^2}]$.
Convergence Analysis of the BCD method {#sec:analysis}
======================================
Throughout the following, let Assumption \[ass\] be satisfied. Moreover, we assume that the step sizes satisfy $s_{\rm min} \leq s_k^\delta \leq s_{\rm max}$ for some numbers $s_{\rm min} \leq s_{\rm max}$ independent of the iteration index $k \in {\mathbb N}$ and the noise level $\delta \geq 0$, and that \[sum1\]-\[sum3\] in Lemma \[conv1\] hold.
Convergence for exact data
--------------------------
In this subsection we show convergence of the BCD iteration in the noise free case. The proof closely follows [@decesaro2008steepest; @kowar2002].
\[thm:exact\] In the exact data case $\delta =0$, the BCD iteration $(x_k)_{k\in {\mathbb N}}$ defined by , satisfies $x_k \to x^\plus$, where $x^\plus$ is the solution of ${\mathcal A}(x) = y$ with minimal distance to $x_0$.
Let $x^* \in {\mathcal X}$ satisfy ${\mathcal A}(x^*) = y$ and define $\xi_k {:=}{\mathcal V}_{X}x_k- {\mathcal V}_{X}x^*$. We will show that $(\xi_k)_{k \in {\mathbb N}}$ is a Cauchy sequence. For $k =k_0 p+k_1$ and $l=l_0p+l_1$ with $k \leq l$ and $k_1,l_1 \in \{0,\dots,p-1\}$, let $n_0 \in \{k_0,\dots,l_0\}$ be such that $$\begin{aligned}
\label{eq:conv0}
\sum_{i_1=0}^{p-1}&{\left\Vert{\mathcal Q}^{{Y}}_{{b}(p n_0+i_1)}(y-{\mathcal A}(x_{p n_0+i_1}))\right\Vert}\\ \notag
&\leq \sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(p i_0+i_1)} (y-{\mathcal A}(x_{p i_0+i_1}))\right\Vert} \text{ for all } i_0 \in \{k_0,\dots,l_0\}\,.\end{aligned}$$ With $n {:=}p n_0 + p - 1$ we have $$\label{errnorm1}
{\left\Vert\xi_k-\xi_l\right\Vert} \leq {\left\Vert\xi_k-\xi_n\right\Vert}+{\left\Vert\xi_l-\xi_n\right\Vert}$$ and $$\begin{aligned}
\label{errnorm2}
{\left\Vert\xi_n-\xi_k\right\Vert}^2 &= 2{\left\langle\xi_n-\xi_k,\xi_n\right\rangle}+ {\left\Vert\xi_k\right\Vert}^2 - {\left\Vert\xi_n\right\Vert}^2\,, \\ \label{errnorm2a}
{\left\Vert\xi_n-\xi_l\right\Vert}^2 &= 2{\left\langle\xi_n-\xi_l,\xi_n\right\rangle} + {\left\Vert\xi_l\right\Vert}^2 - {\left\Vert\xi_n\right\Vert}^2\,.\end{aligned}$$ According to Lemma \[lem:mon\], the nonnegative sequence $({\left\Vert\xi_k\right\Vert})_{k \in {\mathbb N}}$ is monotonically decreasing and therefore converges to some ${\epsilon}\geq 0$. Consequently, the last two terms in equations and converge to $\varepsilon^2-\varepsilon^2 = 0$ for $k \to \infty$. In order to show that also ${\left\langle\xi_n-\xi_k,\xi_n\right\rangle}$ and ${\left\langle\xi_n-\xi_l,\xi_n\right\rangle}$ converge to zero, we set $i^* {:=}pn_0 + i_1$. Then using the definition of the BCD method in (\[eq:iter\]) for $i \in \{ 0, \dots, p-1 \}$ we obtain $$\begin{aligned}
\label{eq:conv}
& {\left\vert{\left\langle\xi_n-\xi_k,\xi_n\right\rangle}\right\vert} = {\left\vert{\left\langle{\mathcal V}_{X}x_n-{\mathcal V}_{X}x_k,{\mathcal V}_{X}x^*-{\mathcal V}_{X}x_n\right\rangle}\right\vert} \\ \notag
& = {\left\vert\sum_{i=k}^{n-1} s_i {\left\langle {\mathcal V}_{X}{\mathcal P}_{{b}(i)} {\mathcal A}^*({\mathcal A}(x_i)-y), {\mathcal V}_{Y}(x^*-x_n)\right\rangle} \right\vert} \\ \notag
& \leq v_{\rm max}^2 \sum_{i=k}^{n-1} s_i{\left\vert{\left\langle {\mathcal Q}^{{X}}_{{b}(i)} {\mathcal A}^*({\mathcal A}(x_i)-y),x^*-x_n\right\rangle}\right\vert} \\ \notag
& = v_{\rm max}^2 \sum_{i=k}^{n-1} s_i{\left\vert{\left\langle {\mathcal A}(x_i)-y,{\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal A}(x^*-x_n)\right\rangle}\right\vert} \\ \notag
& = v_{\rm max}^2 \sum_{i=k}^{n-1} s_i{\left\vert{\left\langle{\mathcal A}(x_i)-y,{\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal A}(x^*-x_{i^*}) + {\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal A}(x_{i^*}-x_n)\right\rangle}\right\vert}\\ \notag
&\leq v_{\rm max}^2 \sum_{i=k}^{n-1} s_i {\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)}({\mathcal A}(x_i)-y)\right\Vert}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal A}(x^*-x_{i^*})\right\Vert}\\ \notag
&\hspace{0.1\textwidth} + v_{\rm max}^2 \sum_{i=k}^{n-1}s_i {\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)}({\mathcal A}(x_i)-y)\right\Vert}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal A}(x_{i^*}-x_n)\right\Vert}\,,\end{aligned}$$ with $v_{\rm max} {:=}\max {\left\{{\Vertv_1\Vert}, \dots, {\Vertv_{{B}}\Vert}\right\}}$. Further we obtain $$\begin{aligned}
\label{eq:conv3}
&{\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)}{\mathcal A}(x_{i^*}-x_n)\right\Vert}
\\ \notag
& = {\left\Vert{\mathcal Q}^{{Y}}_{{b}(i)} {\mathcal K}_{{D}} {\mathcal V}_{X}(x_{i^*}-x_n)\right\Vert}
\\ \notag
& \leq {\Vert{\mathcal K}_{{D}}\Vert} {\left\Vert{\mathcal V}_{X}(x_{i^*}-x_n)\right\Vert}\\ \notag
&\leq {\Vert{\mathcal K}_{{D}}\Vert}\sum_{j=i_1}^{p-2}s_j{\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(p n_0 +j)} {\mathcal A}^*(y-{\mathcal A}(x_{p n_0+j}))\right\Vert}\\ \notag
&= {\Vert{\mathcal K}_{{D}}\Vert} \sum_{j=i_1}^{p-2}s_j{\left\Vert{\mathcal V}_{X}{\mathcal P}_{{b}(p n_0 +j)}{\mathcal V}_{X}^{*} {\mathcal K}_{{D}}^{*}(y-{\mathcal A}(x_{p n_0+j}))\right\Vert}\\ \notag
&= {\Vert{\mathcal K}_{{D}}\Vert} \sum_{j=i_1}^{p-2}s_j{\left\Vertv_{{b}(p n_0 +j)}\right\Vert}^2{\left\Vert {\mathcal Q}^{{X}}_{{b}(p n_0 +j)} {\mathcal K}_{{D}}^{*}(y-{\mathcal A}(x_{p n_0+j}))\right\Vert}\\ \notag
&\leq {\Vert{\mathcal K}_{{D}}\Vert} \sum_{j=i_1}^{p-2}s_j{\left\Vertv_{{b}(p n_0 +j)}\right\Vert}^2{\left\Vert {\mathcal K}_{{D}}^{*}({\mathcal Q}^{{Y}}_{{b}(p n_0 +j)}(y-{\mathcal A}(x_{p n_0+j})))\right\Vert}\\
&\leq {\Vert{\mathcal K}_{{D}}\Vert}^2 s_{\rm max} v_{\rm max}^2 \sum_{j=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(p n_0 +j)}(y-{\mathcal A}(x_{p n_0+j}))\right\Vert} \,.\end{aligned}$$ Substituting the estimate in (\[eq:conv\]), using the inequality $(\sum_{i=0}^{p-1}a_i)^2 \leq p \sum_{i=0}^{p-1}a_{i}^{2}$ and (\[eq:conv0\]) one concludes $$\begin{aligned}
\label{eq:conv4}
&{\left\vert{\left\langle\xi_n-\xi_k,\xi_n\right\rangle}\right\vert} \\ \notag
&\leq 2 p s_{\rm max} v_{\rm max}^2 \sum_{i_0=k_0}^{n_0-1}\sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(pi_0+i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert}^2\\ \notag
& +s_{\rm max}^2 {\Vert{\mathcal K}_{{B}}\Vert}^2 v_{\rm max}^4\sum_{i_0=k_0}^{n_0-1}\sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(pi_0+i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert} \sum_{j=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(p i_0 +j)}(y-{\mathcal A}(x_{p i_0 + j}))\right\Vert} \\ \notag
&\leq 2 p s_{\rm max} v_{\rm max}^2 \sum_{i_0=k_0}^{n_0-1}\sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(pi_0+i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert}^2\\ \notag
& \qquad + s_{\rm max}^2 {\Vert{\mathcal K}_{{B}}\Vert}^2 v_{\rm max}^4 \sum_{i_0=k_0}^{n_0-1}\Bigl(\sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(pi_0+i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert}\Bigr)^2 \\ \notag
&\leq C \sum_{i_0=k_0}^{n_0-1}\sum_{i_1=0}^{p-1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(pi_0+i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert}^2 \,,\end{aligned}$$ where we defined $ C{:=}s_{\rm max} v_{\rm max}^2 (2p + s_{\rm max} {\Vert{\mathcal K}_{{B}}\Vert}^2 v_{\rm max}^2 p)$. Finally, we have $${\left\vert{\left\langle\xi_n-\xi_k,\xi_n\right\rangle}\right\vert} \leq \frac{C}{s_{\rm min}}\sum_{i_0=k_0}^{n_0-1}\sum_{i_1 =0}^{p-1}s_{p i_0 +i_1}{\left\Vert{\mathcal Q}^{{Y}}_{{b}(i_1)}(y-{\mathcal A}(x_{p i_0 + i_1}))\right\Vert}^2\,.$$ Because of Lemma \[conv1\], the last sum converges to zero for $k=pk_0+k_1 \to \infty$ which implies ${\left\vert{\left\langle\xi_n-\xi_k,\xi_n\right\rangle}\right\vert}\to 0$. Similarly, one shows ${\left\vert{\left\langle\xi_n-\xi_l,\xi_n\right\rangle}\right\vert}\to 0$. Therefore, $\xi_k$ is Cauchy sequence and ${\mathcal V}_{X}x_k = {\mathcal V}_{X}x^{*} - \xi_k$ tends to an element [${\mathcal V}_{X}x^\plus$ with $x^\plus \in {\mathcal X}$. Because ${V}$ has rank ${B}$ and ${\Vert{\mathcal Q}^{{Y}}_{{b}(i)}(y-{\mathcal A}(x_i))\Vert} \to 0$, the element $x^\plus$ is a solution of ${\mathcal A}(x)={ y }$.]{} Further, $$x_{k+1}-x_k \in \operatorname{ran}({\mathcal A}^*) \subseteq \operatorname{ker}( {\mathcal A})^{\perp}
\qquad \text{ for all $k \in {\mathbb N}$}\,.$$ Because $\operatorname{ker}( {\mathcal A})^{\perp}$ is closed, its follows that $ x^* - x_0 \in \operatorname{ker}( {\mathcal A})^{\perp}$. Since $x^\plus$ is the only solution for which the latter holds true, we obtain $x_k \to x^\plus$.
Convergence for noisy data
--------------------------
In the noisy data case, we consider the loping version of the BCD. The iteration is terminated when for the first time all $x^\delta_k$ are equal within a cycle. That is, we stop the iteration at $$\label{eq:def-discrep}
k_*^\delta := \operatorname{arg\, min}{\left\{ k \in {\mathbb N}\mid
x_{k}^\delta = x_{k+1}^\delta = \cdots = x_{k+p-1}^\delta \right\}} \,.$$ [To simplify the notation, we assume that $\delta_{b}=\delta$ for all ${b}\in {\left\{1, \dots , {B}\right\}}$.]{} We first show that the stopping index is always finite.
\[prop:st-f\] If $\delta > 0$, then the stopping index $k_*^\delta$ defined in is finite, and we have $$\label{eq:sdk-monot-res}
\forall {b}= 1, \dots, {B}\colon \quad
{\left\Vert{\mathcal Q}^{{Y}}_{{b}} {\left( { y^\delta}-{\mathcal A}{\left( x_{k_*^\delta}^\delta \right)} \right)} \right\Vert} < \tau \delta
\, .$$
If for every $k \in {\mathbb N}$, there exists $\ell \in
{\left\{0,\dots, p-1\right\}}$ such that $x_{k + \ell} \neq x_{k}$, then from Lemma \[conv1\] we obtain $$\begin{gathered}
\label{eq:bound1}
\forall n \in {\mathbb N}\colon
\quad
{\left\Vert{\mathcal V}_{X}x_0 - {\mathcal V}_{X}x^*\right\Vert}^2 \geq
(2-\theta_{\rm max})
\gamma_{\rm min}
\sum_{k =0}^{ n p - 1} d^\delta_k s^\delta_k { {}{\Vertv_{{b}(k)}\Vert}^2 ({r}^\delta_k)^2}
\\
\geq (2-\theta_{\rm max})
\gamma_{\rm min}
C n p \tau \delta \,,
\end{gathered}$$ where $C>0$ is a lower bound of $s^\delta_k {\Vertv_{{b}(k)}\Vert}^2$. The right hand side of tends to infinity, which gives a contradiction. Consequently, the set $ \{ k \in {\mathbb N}\mid x_{k}^\delta = x_{k+1}^\delta = \cdots = x_{k+p-1}^\delta \} $ is nonempty and therefore contains a finite minimal element.
To prove (\[eq:sdk-monot-res\]) note that the finiteness of the stopping index and the definition of the loping BCD implies ${\Vert{\mathcal Q}^{{Y}}_{ {b}(k_*^\delta+ \ell) } {( { y^\delta}-{\mathcal A}{( x_{k_*^\delta}^\delta )} )} \Vert}
< \tau \delta $ for $\ell = 0, \dots , p - 1$. The assumption \[ass4\] on the control sequence ${b}(k)$ thus gives .
We call the step size selection $(s^\delta_k)_{k \in{\mathbb N}}$ continuous at $\delta =0$ if for all $k \in {\mathbb N}$ we have $$\label{eq:conts}
\lim_{\delta \to 0}
\sup \{ {\Vert s^\delta_k - s_k \Vert} \mid y^\delta \in {\mathcal Y}\wedge {\Verty - { y^\delta}\Vert} \leq \delta \}
= 0 \,.$$ An example for a continuous step size selection is the constant strep size $s_k^\delta = \gamma_{\rm min} / {\Vert{\mathcal K}_{{B}}\Vert}^2$. The next auxiliary result shows that the continuity of the step size selection implies continuity of $x_{k}^\delta$ at $\delta =0$.
\[lem:cont\] Suppose the step selection is continuous at $\delta =0$, and define $$\Delta_k(\delta, y, { y^\delta})
{:=}d^\delta_k s^\delta_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k^\delta)-{ y^\delta})
-
s_k {\mathcal V}_{Y}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k)-y) \,.$$ Then, for all $k \in {\mathbb N}$, we have $$\label{eq:cont}
\lim_{\delta \to 0}
\sup
\left\{ {\Vert \Delta_k(\delta, y, { y^\delta}) \Vert} \mid
y^\delta \in {\mathcal Y}\wedge
{\Verty - { y^\delta}\Vert} \leq \delta \right\}
= 0
\,.$$ Moreover, $x_{k+1}^\delta \to x_{k+1}$, as $\delta \to 0$.
We prove Lemma \[lem:cont\] by induction. Assume $k \geq 0$ and that (\[eq:cont\]) holds for all $k' < k$. First we note that implies $x_{k+1}^\delta \to x_{k+1}$, as $\delta \to 0$. For the proof of we consider two cases. In the first case, $d^\delta_k =1$, we have $$\begin{aligned}
{\left\Vert \Delta_k(\delta, y, y^\delta) \right\Vert}
&
= {\left\Vert s^\delta_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k^\delta)-{ y^\delta})
-
s_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k)-y))\right\Vert} \,.\end{aligned}$$ In the second case, $d^\delta_k =0$, we have ${\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({ y^\delta}-{\mathcal A}( x_{k}^\delta)) \right\Vert} < \tau \delta$. Consequently, $$\begin{aligned}
\|\Delta_k&(\delta, y, y^\delta)\| \\
&=
{\left\Vert s_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal A}^*({\mathcal A}(x_k)-y))\right\Vert} \\
&=
{\left\Vert s_k {\mathcal V}_{X}{\mathcal P}_{{b}(k)} {\mathcal V}_{X}^* {\mathcal K}_{D}^*({\mathcal A}(x_k)-y))\right\Vert} \\
&=
{\left\Vertv_{b}\right\Vert}^2 {\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} {\mathcal K}_{D}^*({\mathcal A}(x_k)-y))\right\Vert} \\
&\leq
{\left\Vertv_{b}\right\Vert}^2
{\left\Vert{\mathcal K}_{{D}}\right\Vert} \,
\Bigl(
{\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({\mathcal A}(x_k^\delta)-{\mathcal A}(x_k))\right\Vert}
+{\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({\mathcal A}(x_k^\delta)- { y^\delta})\right\Vert}
\\ & \hspace{0.2\textwidth} + {\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({ y^\delta}- y )\right\Vert}
\Bigr)
\\
&\leq
{\left\Vertv_{b}\right\Vert}^2 {\left\Vert{\mathcal K}_{{D}}\right\Vert} \,
\Bigl(
{\left\Vert {\mathcal Q}^{{Y}}_{{b}(k)} ({\mathcal A}(x_k^\delta)-{\mathcal A}(x_k))\right\Vert}
+({\Vert{\mathcal Q}^{{Y}}_{{b}(k)}\Vert} + \tau) \delta
\Bigr)
\,.
\end{aligned}$$ Now follows from the continuity of ${\mathcal A}$, and the induction hypothesis implying $x_k^\delta \to x_k$.
\[th:noisy\] Suppose the step selection $(s^\delta_k)_{k \in{\mathbb N}}$ is continuous at $\delta =0$. Let $(\delta_j)_{j\in {\mathbb N}} \in (0,\infty)^{\mathbb N}$ converge to $0$ and let $(y_j)\in {\mathcal Y}^{\mathbb N}$ be a sequence of noisy data with ${\Vert{\mathcal Q}_{{b}}^{Y}( y_j - y) \Vert} \leq \delta_j$. Let $(x_{j,k})_{k \in {\mathbb N}}$ be defined by the loping BCD iteration with data $y_j$ and stopped at $k_j := k_*(\delta_j, y_j) $ according to . Then $(x_{j,k_j})_{j \in {\mathbb N}} \to x^\plus$, where $x^\plus$ is the solution of ${\mathcal A}(x)=y$ with minimal distance to $x_0$.
From Lemma \[lem:cont\] and the continuity of ${\mathcal A}$ we have, for any $k \in {\mathbb N}$, that $x_{j,k} \to x_{k}$ and ${\mathcal A}( x_{j,k} ) \to {\mathcal A}(x_{k})$ as $j \to \infty $.
To show that $x_{j,k_j} \to x^\plus$, we first assume that $k_j$ has a finite accumulation point $k_*$. Without loss of generality we may assume that $k_j = k_*$ for all $j \in {\mathbb N}$. From Proposition \[prop:st-f\] we know that ${\Vert{\mathcal Q}^{{Y}}_{{b}} ( y_j -
{\mathcal A}(x_{j_,k_*}))\Vert} < \tau \delta_j$. By taking the limit $j \to \infty$, we obtain $ y = {\mathcal A}(x_{k_*}) $. Consequently, $x_{k_*} = x^\plus$ and $x_{j_,k_*} \to x^{*}$ as $j \to \infty$. It remains to consider the case where $k_j \to \infty$ as $j \to
\infty$. To that end let ${\epsilon}>0$. Without loss of generality we assume that $k_j$ is monotonically increasing. According to Theorem \[thm:exact\] we can choose $n \in {\mathbb N}$ such that ${\left\Vert{\mathcal V}_{X}x_{k_n} - {\mathcal V}_{X}x^\plus\right\Vert} < {\epsilon}/2$. Equation implies that there exists $j_0 > n$ such that ${\left\Vert{\mathcal V}_{X}x_{j,k_n} - {\mathcal V}_{X}x_{k_n}\right\Vert} < {\epsilon}/2$ for all $j \geq j_0$. Together with the monotonicity we obtain $$\begin{gathered}
{\left\Vert{\mathcal V}_{X}x_{j,k_j} - {\mathcal V}_{X}x^\plus\right\Vert}
\leq
{\left\Vert{\mathcal V}_{X}x_{j,k_n} - {\mathcal V}_{X}x^\plus\right\Vert}
\\
\leq {\left\Vert{\mathcal V}_{X}x_{j,k_n} - {\mathcal V}_{X}x_{k_n}\right\Vert}
+
{\left\Vert{\mathcal V}_{X}x_{k_n} - {\mathcal V}_{X}x^\plus\right\Vert}
< \frac{{\epsilon}}{2} + \frac{{\epsilon}}{2} = {\epsilon}\quad \text{ for } j \geq j_0 \,.\end{gathered}$$ Because ${\mathcal V}_{Y}$ is non-singular, this shows $x_{j,k_j} \to x^\plus$ as $j \to \infty$.
Example: System of linear integral equation {#sec:int}
===========================================
In this section we compare the BCD method to the standard Landweber method for an elementary system of linear integral equations.
![**Test phantoms and noisy data for a system of two integral equations.** Top: The two components $f^*[1]$ (left) and $f^*[2]$ (right) of the true unknown. Bottom: The two components $g^\delta[1]$ (left) and $g^\delta[2]$ (right) of the computed noisy data.[]{data-label="fig:I-data"}](X1-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Test phantoms and noisy data for a system of two integral equations.** Top: The two components $f^*[1]$ (left) and $f^*[2]$ (right) of the true unknown. Bottom: The two components $g^\delta[1]$ (left) and $g^\delta[2]$ (right) of the computed noisy data.[]{data-label="fig:I-data"}](X2-eps-converted-to.pdf "fig:"){width="49.00000%"}\
![**Test phantoms and noisy data for a system of two integral equations.** Top: The two components $f^*[1]$ (left) and $f^*[2]$ (right) of the true unknown. Bottom: The two components $g^\delta[1]$ (left) and $g^\delta[2]$ (right) of the computed noisy data.[]{data-label="fig:I-data"}](Y1-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Test phantoms and noisy data for a system of two integral equations.** Top: The two components $f^*[1]$ (left) and $f^*[2]$ (right) of the true unknown. Bottom: The two components $g^\delta[1]$ (left) and $g^\delta[2]$ (right) of the computed noisy data.[]{data-label="fig:I-data"}](Y2-eps-converted-to.pdf "fig:"){width="49.00000%"}
Forward problem
---------------
Consider the integration operator ${K}\colon L^2([0,1]) \to L^2([0,1])$ defined by $$\label{eq:intop}
{K}(f) \colon [0,1] \to {\mathbb R}\colon s \mapsto \int_0^s f(t) {\mathrm d}t \,.$$ According to the Cauchy-Schwarz inequality, we have $${\left\Vert{K}(f)\right\Vert}^2 = \int_0^1 {\left(\int_0^s f(t) {\mathrm d}t \right)}^2 {\mathrm d}s
\leq \int_0^1 s \int_0^1 f(t)^2 {\mathrm d}t {\mathrm d}s = \frac{1}{2} {\left\Vertf\right\Vert}^2$$ for all $ f \in L^2([0,1])$, which shows that ${K}$ is a well-defined linear bounded operator. Using the operator ${K}$ we consider the following forward model applied to a vector of functions $(f[{b}])_{{b}=1}^{B}\in {( L^2([0,1]) )}^{B}$.
![**Reconstruction from exact data using the Landweber (LW) and the BCD method.** Top: Reconstruction after 5000 cycles. Bottom: Reconstruction errors in the standard $2$-norm (left) and the ${V}$-norm (right) as a function of the iteration number. For both error measures the reconstruction error of BCD is smaller than the one of the Landweber method.[]{data-label="fig:I-exact"}](X1rec-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstruction from exact data using the Landweber (LW) and the BCD method.** Top: Reconstruction after 5000 cycles. Bottom: Reconstruction errors in the standard $2$-norm (left) and the ${V}$-norm (right) as a function of the iteration number. For both error measures the reconstruction error of BCD is smaller than the one of the Landweber method.[]{data-label="fig:I-exact"}](X2rec-eps-converted-to.pdf "fig:"){width="49.00000%"}\
![**Reconstruction from exact data using the Landweber (LW) and the BCD method.** Top: Reconstruction after 5000 cycles. Bottom: Reconstruction errors in the standard $2$-norm (left) and the ${V}$-norm (right) as a function of the iteration number. For both error measures the reconstruction error of BCD is smaller than the one of the Landweber method.[]{data-label="fig:I-exact"}](L2norm-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstruction from exact data using the Landweber (LW) and the BCD method.** Top: Reconstruction after 5000 cycles. Bottom: Reconstruction errors in the standard $2$-norm (left) and the ${V}$-norm (right) as a function of the iteration number. For both error measures the reconstruction error of BCD is smaller than the one of the Landweber method.[]{data-label="fig:I-exact"}](Vnorm-eps-converted-to.pdf "fig:"){width="49.00000%"}
For ${D}\geq {B}\geq 1$ and given matrix $V =(v_{d,b})_{d,b}\in {\mathbb R}^{D \times B}$ of rank ${B}$, we define the forward operator $${\mathcal A}\colon {( L^2([0,1]) )}^{B}\to {( L^2([0,1]))}^{D}\colon f \mapsto {\left( \sum_{b=1}^B v_{b,c} {K}(f[{b}])\right)}_{{d}=1}^{D}\,.$$
According to our general notion we have ${\mathcal A}= V \otimes {K}$ and the theory presented in the previous section can be applied for solving the inverse problem ${\mathcal A}(f) = g$. Note that this equation clearly is ill-posed because the range of ${K}$ is non-closed (and equal to the Sobolev space $H^1_{\diamond}([0,1]) {:=}{\left\{g \in L^2([0,1]) \mid g' \in L^2([0,1]) \wedge g(0)=0\right\}}$ of all weakly differentiable functions vanishing at $0$.)
More generally, one could replace the integration operator by any bounded (integral) operator ${K}\colon L^2([0,1]) \to L^2([0,1])$ with non-closed range.
![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](X1recN-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](X2recN-eps-converted-to.pdf "fig:"){width="49.00000%"}\
![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](Nnorm_noskip-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](Nnorm_skip-eps-converted-to.pdf "fig:"){width="49.00000%"}\
![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](L2norm_mon-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstructions from noisy data using the Landweber (LW) and the BCD method.** Top: Reconstruction using the BCD iteration (with the loping principle and the proposed stopping rule) and the Landweber method using the discrepancy principle as stopping rule. Middle: Reconstruction errors in the ${V}$-norm without (left) and with (right) loping. Bottom: reconstruction error for the first iterates in the $2$-norm (not monotonically decreasing) and in the ${V}$-norm (right).[]{data-label="fig:I-noisy"}](Vnorm_mon-eps-converted-to.pdf "fig:"){width="49.00000%"}
Reconstruction results
----------------------
For all presented numerical results we use ${B}={D}=2$ and take ${V}= \tilde {V}/{{\Vert\tilde {V}\Vert}}_{2,2}$ with $$\tilde {V}{:=}\begin{bmatrix} -3 & 1 \\ -1 & 0 \end{bmatrix} \,.$$ We discretize ${K}$ with the composite trapezoidal rule using $p = 100$ intervals such that the data and the unknowns are elements in $({\mathbb R}^p)^2 $. The true unknown $f^* =(f^*[1], f^*[2])$ and the noisy data $g^\delta =(g^\delta[1], g^\delta[2]) $ are shown in Figure \[fig:I-data\]. The exact data $g = {\mathcal A}f^*$ have been computed via numerical integration followed by application of ${V}$. Subsequently we computed noisy data by adding random white noise to ${ y }$ with a standard deviation of $0.001$. The resulting relative data errors are ${\left\Vertg-g^\delta\right\Vert} / {\left\Vert g \right\Vert} \simeq 0.015$, ${\left\Vert{\mathcal Q}_1(g-g^\delta)\right\Vert} / {\left\Vert{\mathcal Q}_1 g \right\Vert} \simeq 0.011$ and ${\left\Vert{\mathcal Q}_2(g-g^\delta)\right\Vert} / {\left\Vert{\mathcal Q}_2 g \right\Vert} \simeq 0.012$, respectively.
Reconstruction using the BCD and Landweber methods from simulated data are shown in Figure \[fig:I-exact\]. For each case we have used the maximum constant step-size, that lead to stable reconstruction. We evaluate the reconstruction error (norm of $f_k - f^*$) in terms of the standard 2-norm ${\left\Vert{\,\cdot \,}\right\Vert}_2$ and in the ${V}$-norm ${\left\Vert{\,\cdot \,}\right\Vert}_{{V}}$, $$\begin{aligned}
{\left\Vertf\right\Vert}_2^2
&{:=}{\left\Vertf[1]\right\Vert}^2 + {\left\Vertf[2]\right\Vert}^2\\
{\left\Vertf\right\Vert}_{{V}}^2
&{:=}{\left\Vertv_{1,1} f[1] + v_{1,2} f[2]\right\Vert}^2 +{\left\Vertv_{2,1} f[1] + v_{2,2} f[2]\right\Vert}^2 \,,\end{aligned}$$ respectively. As we can see from the bottom row in Figure \[fig:I-exact\], measured in both norms, the reconstruction error of the BCD is smaller than the error of Landweber iteration.
Figure \[fig:I-noisy\] shows reconstruction results for nosy data. Again, the error in the BCD method decreases faster than the one of the Landweber method. The BCD therefore requires less cycles than the Landweber method. Moreover, in the middle column of Figure \[fig:I-noisy\] we illustrate the need for the loping (or another regularization strategy). Without loping, the BCD iteration as well as the Landweber iteration start to diverge after around 2000 iterations. With loping (for the BCD method) and the with the discrepancy principle (for the Landweber method) both iterations stop. (Note that here we only show the error in the ${V}$-norm and that the Landweber method is monotonically decreasing in the $2$-norm when using the discrepancy principle.) Finally, the bottom row in Figure \[fig:I-noisy\] shows that the reconstruction error for the BCD iteration is not monotonically decreasing in the standard norm, whereas in the ${V}$-norm it is.
A nonlinear test: Multi-spectral X-ray tomography {#sec:xray}
=================================================
In this section we apply [a nonlinear generalization of the]{} BCD and the Landweber iteration to one-step inversion in multi-spectral X-ray tomography. [In particular, for nonlinear operators ${\mathcal A}$ in place of linear ones, we use the following generalization of the BCD iteration $$\label{eq:iterX}
x_{k+1}^\delta
{:=}x_k^\delta - s^\delta_k {\mathcal P}_{{b}(k)} {\mathcal A}'(x_k^\delta)^*({\mathcal A}(x_k^\delta)-y^\delta)\,.$$ Note that such problems are not covered be our theoretical analysis. We consider extending our theory to this class of examples a particularly interesting topic of future research.]{}
In the following we denote by $D_R \subseteq {\mathbb R}^2$ the disc with radius $R < 1$ centered at the origin. We define the fan beam Radon transform ${R}\mu \colon {\mathbb S}^1 \times {\mathbb S}^1 \to {\mathbb R}$ of a function $\mu \colon {\mathbb R}^2 \to {\mathbb R}$ supported in $D_R$ by $$\label{eq:radon}
({R}\mu)(\alpha,\beta) {:=}\int_{{\mathbb R}}\mu( \alpha + t\beta ){\mathrm d}t\,.$$ It can be easily verified that the fan beam Radon transform ${R}\colon L^2(D_R) \to L^2({\mathbb S}^1 \times {\mathbb S}^1)$ is linear and continuous [@natterer2001mathematics].
Mathematical modeling
---------------------
We assume that the tissue is composed of ${B}$ different materials each of them having a different energy dependent X-ray attenuation coefficient $\mu_{b}(E)$ with ${b}= 1, \dots, {B}$. The combined X-ray attenuation coefficient is then given by $$\mu( E, {\,\cdot \,}) = \sum_{{b}= 1}^{{B}}\mu_{b}(E) f[{b}] \,,$$ where $f[{b}] \colon {\mathbb R}^2 \to [0,1]$ is the fractional density map of the ${b}$th material. Our goal is to determine the fractional density maps $f[{b}] $ from multi-spectral X-ray transmission measurements.
![Normalized\[spec\] spectrum of a typical X-ray source [@atak2015dual; @barber2016algorithm]. This spectral energy distribution will be considered for our experiments.](spectrum-eps-converted-to.pdf){width="70.00000%"}
The energy sensitive X-ray transmission measurements result in the intensity [@barber2016algorithm] $$\label{xrayint}
I_W = \int_{W} s(E)\exp{\left( -{R}(\mu(E,{\,\cdot \,}))\right)} {\mathrm d}E\,.$$ Here $W \subseteq [0, \infty)$ denotes the energy window where the measurement is made and $s \colon [0, \infty) \to {\mathbb R}$ is the product of X-ray beam spectrum intensity and detector sensitivity. We assume the detector sensitivity to be constant and that the spectrum $s $ is known for energies ranging from $\SI{20}{\keV}$ to $\SI{120}{\keV}$ covering any energy window. The spectrum used for the numerical results is the same as in [@atak2015dual; @barber2016algorithm] and shown in Figure \[spec\].
In order to recover multiple material densities, we use multiple energy windows. We choose the same number ${B}$ of spectral windows as we have different materials. Moreover, to simplify the mathematical formulation we uniformly discretize the energy variable, $E_0 = \SI{20}{\keV} < E_1 < \cdots <
E_N =\SI{120}{\keV}$. The X-ray measurements corresponding to the ${b}$th energy window is given by $$\label{xraysum}
I[{b}] =
\sum_{i \in W_{b}}
s_i \exp(-{R}( \mu_i ))
\, \Delta E
=
\sum_{i \in W_{b}}
s_i \exp {\left( - {R}{\left( \sum_{{b}= 1}^{{B}} \mu_{i,{b}} f[{b}] \right)} \right)} \, \Delta E \,.$$ Here $W_{b}\subseteq {\left\{1, \dots , N\right\}}$ model discrete energy windows, $(s_i)_{i=1}^N$ is the discretized beam spectrum intensity, and $\Delta E {:=}(\SI{120}{\keV})/N$. Summarizing the above we define the following forward operator.
The measurement operator ${\mathcal A}$ with respect to the energy windows $W_1, \dots,W_{B}$ is given by $$\begin{aligned}
\nonumber
{\mathcal A}\colon (L^2(D_R))^{B}&\to L^2({\mathbb S}^1 \times {\mathbb S}^1)^{B}\\ \label{eq:ms}
f &\mapsto
{\left(
\sum_{i \in W_{b}}
s_i \exp {\left( - {R}{\left( \sum_{{b}= 1}^{{B}} \mu_{i,{b}} f[{b}] \right)} \right)}
\right)}_{{b}=1}^{B}\,.
\end{aligned}$$
We can decompose the operator ${\mathcal A}$ in the form $$\label{eq:bera}
{\mathcal A}(f) = {\left( {\mathcal V}_{Y}\circ {\mathcal E}\circ {\mathcal R}\circ {\mathcal U}\right)} (f)$$ where
- ${\mathcal U}\colon L^2(D_R)^{B}\to L^2(D_R)^N\colon f \mapsto ( \sum_{{b}= 1}^{{B}} \mu_{i,{b}} f[{b}])_{i=1}^{N} $
- ${\mathcal R}\colon L^2(D_R)^N \to L^2(D_R)^N \colon (\mu_i)_{i=1}^{N}
\mapsto ({R}\mu_i)_{i=1}^{N}$
- ${\mathcal E}\colon L^2(D_R)^N \to L^2(D_R)^N \colon (g_i)_{i=1}^{N}
\mapsto (\exp (- g_i))_{i=1}^{N}$
- ${\mathcal V}_{Y}\colon L^2(D_R)^N \to L^2(D_R)^{B}\colon (g_i)_{i=1}^{N}
\mapsto (\sum_{i \in W_{b}}
s_i g_i)_{{b}=1}^{B}$.
The operators ${\mathcal V}_{Y}, {\mathcal R}, {\mathcal U}$ are linear and bounded. To show the continuity and differentiability of ${\mathcal A}$ we have to verify that ${\mathcal E}$ is continuous and differentiable.
The operator ${\mathcal A}$ is continuous and Fréchet differentiable. For $f,h \in (L(D_R)^2)^{B}$ we have $$\label{eq:derexx}
{\mathcal A}'(f)(h) = {\left( {\mathcal V}_{Y}\circ {\mathcal E}'({\mathcal R}{\mathcal U}f) \circ {\mathcal R}\circ {\mathcal U}\right)} (h)$$ with $$\label{eq:derex}
{\mathcal E}'( g ) h = - ( \exp(-g_i) h_i)_{i=1}^N \,.$$
One only has to verify that $f \mapsto \exp(-f)$ is continuous and Fréchet differentiable on $L^2(D_R)$ with derivative given by ${\mathcal E}'( g ) h = \exp(-g) h$. For that purpose, let ${\left\Verth\right\Vert}_{2} \to 0$ which in particular implies its point wise convergence. Therefore $$\begin{aligned}
& \frac{{\left\Vert\exp( -g -h) - \exp(-g) + \exp( -g) h \right\Vert}_2}{{\left\Verth\right\Vert}_2}
\\&=
\frac{{\left\Vert\exp( -g) \exp(-h) - \exp(-g) + \exp( -g) h \right\Vert}_2}{{\left\Verth\right\Vert}_2} \\&
\leq
\frac{{\left\Vert\exp(-h) - 1 + h \right\Vert}_2}{{\left\Verth\right\Vert}_2}
\\&
\leq \frac{ {\left\Vert\mathcal O (h^2) \right\Vert}_2 }{{\left\Verth\right\Vert}_2}
\leq \frac{ \mathcal O {\left( {\left\Verth\right\Vert}_2^2\right)} }{{\left\Verth\right\Vert}_2}
= \mathcal O ( {\left\Verth\right\Vert}_2 ) \,.\end{aligned}$$ This shows , and follows by the chain rule.
In the context of the BCD method, the fractional density maps $f[{b}] $ play the roles of the blocks $x[{b}]$. The form of the forward operator ${\mathcal A}$ [has some similarity with the form that we used in the theoretical analysis of the BCD method, in the sense that the infinite dimensional smoothing operator is applied to several channels of a function. However, so far we have not been able to perform an analysis accounting for the non-linearity.]{} Additionally, we apply a preconditioning technique as outlined in the following subsection. Extending the convergence analysis of BCD such that it applies to multi-spectral CT is subject of future research.
Logarithmic scaling and preconditioning {#sec:scaling}
---------------------------------------
The energy dependence of the mass-attenuation coefficient of different materials can be quite similar. In order to enhance the dependence on the different materials we propose a logarithmic scaling and preconditioning technique (different from [@barber2016algorithm]). For simplicity we consider only the case ${B}= 2$, the general case can be treated in a similar manner.
The proposed preconditioned logarithmic data take the form $$\begin{gathered}
\label{precon}
{\mathcal H}(f) {:=}\begin{pmatrix}
{\mathcal H}_1(f)
\\
{\mathcal H}_2(f)
\end{pmatrix}=
\begin{pmatrix}
c_{1,1} & c_{1,2}
\\
c_{2,1} & c_{2,2}
\end{pmatrix}
\begin{pmatrix}
\log ( {\mathcal A}_1(f) )
\\
\log( {\mathcal A}_2(f) )
\end{pmatrix}
\\ =
\begin{pmatrix}
c_{1,1} \log( {\mathcal A}_1(f))+ c_{1,2} \log({\mathcal A}_2(f))
\\
c_{2,1} \log({\mathcal A}_1(f))+ c_{2,2} \log({\mathcal A}_2(f))
\end{pmatrix} \,,
\end{gathered}$$ where $f = (f[1],f[2])$ are the unknowns and $c_{1,1}$, $c_{1,2}$, $c_{2,1}$, $c_{2,2}$ are parameters. Moreover, recall that ${\mathcal A}_1(f)$ and ${\mathcal A}_2(f)$ are the X-ray intensities defined by corresponding to $W_1, W_2 \subseteq {\left\{1, \dots , N\right\}}$ modeling the discrete energy windows. The preconditioned inverse problem consists in solving the system $$\begin{aligned}
\label{eq:invms1}
v_1 &= {\mathcal H}_1(f[1],f[2])+z_1
\\ \label{eq:invms2}
v_2 &= {\mathcal H}_2(f[1],f[2]) +z_2\,,\end{aligned}$$ where $v_1,v_2$ are data perturbed by noise $(z_1,z_2)$.
In order to solve the equations in , with the BCD method we define the residual functionals $$\begin{aligned}
{\Phi}_1(f[1],f[2])&:= \frac{1}{2}{\left\Vert {\mathcal H}_1(f[1],f[2])-v_1\right\Vert}^2\,,\\
{\Phi}_2(f[1],f[2])&:= \frac{1}{2}{\left\Vert {\mathcal H}_2(f[1],f[2])-v_2\right\Vert}^2\,.\end{aligned}$$ Application of the BCD method requires the adjoint gradient of ${\mathcal A}_1$ and ${\mathcal A}_2$, that we compute next.
\[th:derivative\] Let $f, h \in L(D_R)^2$. The directional derivatives of ${\Phi}_1$ and ${\Phi}_2$ at $f$ in direction $h$ are given by $$\begin{gathered}
{\Phi}_{b}'(f)(h)
\\
=
- \sum_{m=1}^{2}
\sum_{k=1}^2
\sum_{i \in W_{b}}
\Big\langle {\mathcal H}_{b}(f)-v_{b},
\frac{c_{{b},k}}{{\mathcal A}_k(f)}
s_i
\exp(-{R}({\mathcal U}f)_i){R}(\mu_{i, m} h_m)
\Big\rangle_{L^2}
\,.
\end{gathered}$$
This follows from the chain rule.
From Proposition \[th:derivative\] we conclude that the partial gradients of the residual functionals ${\Phi}_{b}$ are given by $$\begin{gathered}
\partial_{m}
{\Phi}_{b}(f) = - \sum_{i \in W_{b}} \mu_{i,m}{R}^{*}\left[s_i \exp(-{R}({\mathcal U}f)_i)({\mathcal H}_{b}(f)-v_{b})\frac{c_{{b},1}}{{\mathcal A}_1(f)}\right]
\\
- \sum_{i \in W_{b}} \mu_{i,m} {R}^{*}\left[s_i \exp(-{R}({\mathcal U}f)_i)({\mathcal H}_{b}(f)-v_{b})\frac{c_{{b},2}}{{\mathcal A}_2(f)}\right] \,.\end{gathered}$$ These expressions will be used for the implementations of the BCD as well as the Landweber method applied to the preconditioned system .
![**Phantom $f = (f[1], f[2])$ used for the numerical results.** Left: brain density map $f[1]$. \[phantom\] Right: bone density map $f[2]$. Both are derived from the FORBILD head phantom, where a uniformly absorbing disc of value $1/2$ has been added to both channels.](brain-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Phantom $f = (f[1], f[2])$ used for the numerical results.** Left: brain density map $f[1]$. \[phantom\] Right: bone density map $f[2]$. Both are derived from the FORBILD head phantom, where a uniformly absorbing disc of value $1/2$ has been added to both channels.](bone-eps-converted-to.pdf "fig:"){width="49.00000%"}
![**Attenuation coefficients of brain and bone taken from NIST tables [@hubell04]**.\[massatt\] Left: Attenuation spectrum for brain. Right: Attenuation spectrum for bone.](massatt_brain-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Attenuation coefficients of brain and bone taken from NIST tables [@hubell04]**.\[massatt\] Left: Attenuation spectrum for brain. Right: Attenuation spectrum for bone.](massatt_bone-eps-converted-to.pdf "fig:"){width="49.00000%"}
![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](datenfenster1-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](datenfenster2-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](logdatenfenster1-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](logdatenfenster2-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](brain_data-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Simulated multi-energy X-ray data**. \[data\] Top: Data for energy window $[\SI{20}{keV},\SI{70}{keV}]$ (left) and $[\SI{70}{keV},\SI{120}{keV}]$ (right). Middle: Corresponding preconditioned logarithmic data. Third row: Simulated data for the full energy window $[\SI{20}{keV},\SI{120}{keV}]$, where the tissue consists only of the brain map (left) and the bone map (right). ](bone_data-eps-converted-to.pdf "fig:"){width="49.00000%"}
Numerical implementation
------------------------
For all our experiments we used fan beam geometry. Each channel of the discrete phantom has size $400 \times 400$. We discretized ${R}$ using 300 detector positions $\alpha_k$ equidistantly distributed on ${\mathbb S}^1$. For each detector position we compute $481$ line integrals for uniformly distributed angles $\beta_\ell$ in the interval $[-\pi/3,\pi/3]$. To actually compute ${R}f(\alpha_k, \beta_\ell)$ we used the trapezoidal rule and linear interpolation where we discretized the line integral using 400 equidistant sampling points in the interval $[0,2]$. The adjoint ${R}^* g$ is evaluated using the standard backprojection algorithm with linear interpolation. We used $N = 30$ equidistant discrete energy positions from $\SI{20}{keV}$ to $\SI{120}{keV}$.
For our numerical studies we apply one-step inversion in multi-spectral CT tomography to reconstruct a head phantom composed of two different material map derived from FORBID head. The phantom is shown in Figure \[phantom\] and consists of the pair $f = (f[1], f[2])$, where $f[1]$ corresponds to the fractional density of the brain and $f[2]$ to the fractional density of the bone material. We slightly modified the FORBID head phantom by inserting a disk with value $1/2$ in both components to demonstrate that the method can actually reconstruct mixed material distributions. The mass attenuation coefficients of the material maps (bone and brain) are taken from NIST tables [@hubell04] and are shown in Figure \[massatt\].
Figure \[data\] shows the data used for image reconstruction. In the first row original data ${\mathcal A}(f) = ({\mathcal A}_1(f)), {\mathcal A}_2(f)$ according to Definition \[def:ms\] are plotted, where the indices $1$ and $2$ corresponds to energy windows $[\SI{20}{keV},\SI{70}{keV}]$ and $[\SI{70}{keV},\SI{120}{keV}]$, respectively. One can observe, the data for both energy windows look quite similar. This is because of the similar energy dependence of the mass attenuation coefficients for $f[1]$ and $f[2]$; compare Figure \[massatt\]. For this reason, we make use of the proposed scaling and preconditioning outlined in Section \[sec:scaling\]. The second row shows the preconditioned data we use for the reconstruction. For comparison purpose, the last row in Figure \[data\] shows the negative logarithm of the X-ray intensities for the full energy window, with in each case containing only one of the material maps. We have chosen the constants $c_{1,1}=1$, $c_{1,2}=-1.35$, $c_{2,1}=-1$ and $c_{2,2}=2.3$ in such a way that each of the modified data blocks highlights different aspects of the material maps. Note that we have selected the constants for data of a very different phantom in order to avoid inverse crime.
![ **Reconstruction results for simulated data.**\[rec\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. For the Landweber method we have used $300$ iterations, for the BCD method $300$ cycles.](LandweberBrain-eps-converted-to.pdf "fig:"){width="49.00000%"} ![ **Reconstruction results for simulated data.**\[rec\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. For the Landweber method we have used $300$ iterations, for the BCD method $300$ cycles.](LandweberBone-eps-converted-to.pdf "fig:"){width="49.00000%"} ![ **Reconstruction results for simulated data.**\[rec\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. For the Landweber method we have used $300$ iterations, for the BCD method $300$ cycles.](CD_Brain-eps-converted-to.pdf "fig:"){width="49.00000%"} ![ **Reconstruction results for simulated data.**\[rec\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. For the Landweber method we have used $300$ iterations, for the BCD method $300$ cycles.](CD_Bone-eps-converted-to.pdf "fig:"){width="49.00000%"}
![**Relative reconstruction error for simulated data.**\[err\] Left: Reconstructed brain density. Right: Reconstructed bone density. The Landweber method is shown in dashed blue and the BCD method in solid red.](RelERR_bone-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Relative reconstruction error for simulated data.**\[err\] Left: Reconstructed brain density. Right: Reconstructed bone density. The Landweber method is shown in dashed blue and the BCD method in solid red.](RelERR_brain-eps-converted-to.pdf "fig:"){width="49.00000%"}
Numerical results
-----------------
For the following results we compare the performance of the BCD method with the standard gradient method as reference method. We use a cyclic control ${b}(k) = (k-1) \mod {B}$ and constant step sizes for both methods. Note that for the BCD as well as the Landweber method we included a positivity constraint. Figure \[rec\] shows reconstruction results for the bone and brain material map. Due to the applied logarithmic scaling and preconditioning, both methods are able to separate the materials after a reasonable number of iterations. One observes that even the mixed part can be reconstructed as well.
Figure \[err\] shows the relative squared reconstruction errors $$e[{b}] {:=}\frac{{\left\Vertf[{b}]-f_{\rm rec}[{b}]\right\Vert}^2}{{\left\Vertf_{\rm rec}[{b}]\right\Vert}^2}$$ of the bone and the brain map using the Landweber method and the BCD method. The horizontal axes show the number of iterations in the Landweber method and the number of cycles (number of iterations divided by the number of blocks) in the BCD method. A cycle for the BCD method has the same numerical complexity as one iteration for the Landweber method. The BCD method delivers a lower relative error for the brain map, the relative error of the reconstruction for the bone map is similar for both methods.
Reconstruction results for noisy data are shown in Figure \[rec:noise\]. To generate the noisy data, we added Gaussian white noise with standard deviation equal to $\SI{2}{\percent}$ of the maximal value of the exact data. In order to maintain stability of both iterations we stopped the Landweber iteration after $116$ iterations, accordingly the BCD-method is stopped after $116$ cycles. The relative squared reconstruction error is shown in Figure \[err:noise\]. Again, the BCD method is roughly a factor two faster than the Landweber method in recovering the brain map. For recovering the bone map, both methods are equally fast. We associate this different behavior to the particular form of preconditioning. As can be seen from the second line in Figure \[data\], both preconditioned data pairs contain significant parts of the data corresponding to the brain whereas the bone data is mainly contained in the second one. Investigating optimal weights for the preconditioning is an interesting aspect of future work.
![**Reconstruction results for noisy data.**\[rec:noise\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. The Landweber method we have used $116$ iterations and for the BCD method $116$ cycles.](LandweberBrain_noise-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstruction results for noisy data.**\[rec:noise\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. The Landweber method we have used $116$ iterations and for the BCD method $116$ cycles.](LandweberBone_noise-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstruction results for noisy data.**\[rec:noise\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. The Landweber method we have used $116$ iterations and for the BCD method $116$ cycles.](CD_Brain_noise-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Reconstruction results for noisy data.**\[rec:noise\] Top left: Reconstructed brain density with Landweber method. Top right: Reconstructed bone density with Landweber method. Bottom left: Reconstructed brain density with BCD method. Bottom right: Reconstructed bone density with BCD method. The Landweber method we have used $116$ iterations and for the BCD method $116$ cycles.](CD_Bone_noise-eps-converted-to.pdf "fig:"){width="49.00000%"}
![**Relative reconstruction error for noisy data.**\[err:noise\] Left: Reconstructed brain density. Right: Reconstructed bone density. The Landweber method is shown in blue and the BCD method in red. We observed the typical semi-convergence behaviour and therefore stopped the iterations at 116 cycles.](RelERR_brain_noise-eps-converted-to.pdf "fig:"){width="49.00000%"} ![**Relative reconstruction error for noisy data.**\[err:noise\] Left: Reconstructed brain density. Right: Reconstructed bone density. The Landweber method is shown in blue and the BCD method in red. We observed the typical semi-convergence behaviour and therefore stopped the iterations at 116 cycles.](RelERR_bone_noise-eps-converted-to.pdf "fig:"){width="49.00000%"}
Conclusion {#sec:conclusion}
==========
In this paper we analyzed the BCD (block coordinate descent) method for linear inverse problems. For a particular tensor product form we have shown that the BCD method combined with an appropriate loping and stopping strategy is a convergent regularization method for ill-posed inverse problems. The analysis in the present paper applies to operators having the tensor product form ${V}\otimes {K}(x) =
{V}( {K}(x[1]), \dots, {{}{K}(x[{B}]) } )$, where ${V}\in {\mathbb R}^{{D}\times {B}}$ and ${K}\colon {X}\to {Y}$ is linear. We presented two examples for numerically solving ill-posed problems with the BCD method. The first one is concerns a system of linear integral equations that is covered by our theory. As an outlook we applied the BCD method to an example not covered by our theory, namely one-step inversion in multi-spectral X-ray computed tomography.
Future work will be done to extend our analysis of the BCD method to more general forward [operators, in particular non-linear problems including examples like multi-spectral CT.]{} This is challenging as the BCD is not monotone in the reconstruction error ${\left\Vertx_k - x^*\right\Vert}$. However, we believe that the technique introduced in this paper of finding a suitable norm where monotonicity holds can be extended to more general situations.
Acknowledgments {#acknowledgments .unnumbered}
===============
The work Markus Haltmeier has been supported by the Austrian Science Fund (FWF), project P 30747-N32. Simon Rabanser acknowledges support of the Austrian Academy of Sciences (ÖAW) via the DOC Fellowship Programme. The authors thank the anonymous reviewers for valuable comments that helped to significantly improve the manuscript.
[10]{}
H. Atak and P. M. Shikhaliev. Dual energy ct with photon counting and dual source systems: comparative evaluation. , 60(23):8949, 2015.
R. F. Barber, E. Y. Sidky, T. G. Schmidt, and X. Pan. An algorithm for constrained one-step inversion of spectral [CT]{} data. , 61(10):3784, 2016.
A. Beck and L. Tetruashvili. On the convergence of block coordinate descent type methods. , 23(4):2037–2060, 2013.
J. Bolte, S. Sabach, and M. Teboulle. Proximal alternating linearized minimization for nonconvex and nonsmooth problems. , 146(1-2):459–494, 2014.
A. De Cezaro, M. Haltmeier, A. Leit[ã]{}o, and O. Scherzer. On steepest-descent-[K]{}aczmarz methods for regularizing systems of nonlinear ill-posed equations. , 202(2):596–607, 2008.
H. Engl, M. Hanke, and A. Neubauer. , volume 375. Springer Science & Business Media, 1996.
M Haltmeier. Convergence analysis of a block iterative version of the loping [Landweber]{}-[Kaczmarz]{} iteration. , 71(12):e2912–e2919, 2009.
M. Haltmeier, A. Leit[ã]{}o, and O. Scherzer. Kaczmarz methods for regularizing nonlinear ill-posed equations. [I]{}. [C]{}onvergence analysis. , 1(2):289–298, 2007.
M. Hanke, A. Neubauer, and O. Scherzer. A convergence analysis of the [L]{}andweber iteration for nonlinear ill-posed problems. , 72(1):21–37, 1995.
J. H. Hubell and S. M. Seltzer. Tables of x-ray mass attenuation coefficients and mass energy-absorption coefficients from 1 [keV]{} to 20 [MeV]{} for elements $z = 1$ to 92 and 48 additional substances of dosimetric interest. Technical report, 1995.
B. Kaltenbacher, A. Neubauer, and O. Scherzer. , volume 6. Walter de Gruyter, 2008.
D. Kazantsev, J. S. J[ø]{}rgensen, M. S. Andersen, W. R. B. Lionheart, P. D. Lee, and P. J. Withers. Joint image reconstruction method with correlative multi-channel prior for x-ray spectral computed tomography. , 34(6):064001, 2018.
R. Kowar and O. Scherzer. Convergence analysis of a [L]{}andweber-[K]{}aczmarz method for solving nonlinear ill-posed problems. In [*Ill-posed and inverse problems*]{}, pages 253–270. VSP, 2002.
L. Landweber. An iteration formula for fredholm integral equations of the first kind. , 73(3):615–624, 1951.
A. Leitão and B. F. Svaiter. On projective [L]{}andweber-[K]{}aczmarz methods for solving systems of nonlinear ill-posed equations. , 32(2):025004, 20, 2016.
H. Li and M. Haltmeier. The averaged [K]{}aczmarz iteration for solving inverse problems. , 11(1):618–642, 2018.
F. Natterer. , volume 32. SIAM, 2001.
Y. Nesterov. Efficiency of coordinate descent methods on huge-scale optimization problems. , 22(2):341–362, 2012.
A. Neubauer. A new gradient method for ill-posed problems. , 39, 12 2017.
A. Neubauer and O. Scherzer. A convergence rate result for a steepest descent method and a minimal error method for the solution of nonlinear ill-posed problems. , 14(2):369–377, 1995.
D. S. Rigie and P. J. La Rivi[è]{}re. Joint reconstruction of multi-channel, spectral ct data via constrained total nuclear variation minimization. , 60(5):1741, 2015.
A. Saha and A. Tewari. On the nonasymptotic convergence of cyclic coordinate descent methods. , 23(1):576–601, 2013.
O. Scherzer, M. Grasmair, H. Grossauer, M. Haltmeier, and F. Lenzen. , volume 167 of [*Applied Mathematical Sciences*]{}. Springer, New York, 2009.
S. J. Wright. Coordinate descent algorithms. , 151(1):3–34, 2015.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A triggered $140\ks$ observation of the narrow-line Seyfert 1 (NLS1) 335 in December 2015 caught the active galaxy at its lowest X-ray flux since 2007.Ê The NLS1 is relatively quiescent for the first $\sim120\ks$ of the observation before it flares in brightness by a factor of about five in the last $20\ks$.Ê Although only part of the flare is captured before the observation is terminated, the data reveal significant differences between the flare and quiescent phases.Ê During the low-flux state, 335 demonstrates a reflection-dominated spectrum that results from a compact corona around a Kerr black hole.Ê In addition to the rapid brightening, the flare is further described by spectral softening and a falling reflection fraction that are consistent with previous observations advocating at least part of the corona in 335 could be the base of an aborted jet. Ê The spectrum during the flaring interval reveals several residuals between the $2-3\sigma$ level that could be attributed to absorption lines from a highly ionised plasma that is moving outward at $v\sim0.12c$.Ê It could be that the increased luminosity during the flare enhances the radiation pressure sufficiently to launch a possible wind.Ê If the wind is indeed responding to the change in corona luminosity then it must be located within $\sim80\rg$.Ê The escape velocity at this distance is comparable to the estimated wind velocity.Ê If confirmed, this is the first example of a radio-quiet AGN exhibiting behaviour consistent with both diffuse and collimated outflow.'
date: 'Accepted. Received. '
title: ' Evidence for an emerging disc wind and collimated outflow during an X-ray flare in the narrow-line Seyfert 1 galaxy 335 '
---
\[firstpage\]
galaxies: active – galaxies: nuclei – galaxies: individual: 335 – X-ray: galaxies
Introduction {#sect:intro}
============
While the details are uncertain, it is true that the energy expelled from the central engine of active galactic nuclei (AGN) can play a significant role in galaxy evolution (e.g. Hopkins 2012; Fabian 2012). The ejection mechanism is clearly visible in many AGN, whether it be the collimated, relativistic jets in some radio-loud objects (e.g. Bridle & Perley 1984) or diffuse accretion disc winds seen in the UV spectral features of some radio-quiet AGN (e.g. Murray 1995). The more powerful relativistic jets may draw energy from a rapidly rotating black hole at the centre (e.g. Blandford & Znajek 1977), whereas the disc winds may be launched from a larger distance by radiation or magnetic pressure (e.g. Blandford & Payne 1982; Murray 1995; Proga & Kallman 2004). Although uncommon, there are examples of jets and winds co-existing in radio-loud AGN (e.g. Giroletti 2017; Tombesi 2010, 2012, 2014).
It has become clear that even low-luminosity, radio-quiet AGN are capable of launching jets (e.g. Giroletti & Panessa 2009; Foschini 2015; Lähteenmäki 2018) and evidence for winds are seen in the X-ray spectra of several Seyferts (e.g. Krongold 2003; Tombesi 2010; Longinotti 2015). Both processes are variable and may be working intermittently in some objects (e.g. Matzeu 2016; Parker 2017; Gonzalez 2017a).
In recent years, a number of AGN, specifically narrow-line Seyfert 1 galaxies (NLS1s; e.g. Gallo 2018; Komossa 2018), exhibit behaviour that could be reconciled if some significant part of the corona were attributed to the base of a jet (e.g. Gallo 2007, 2013; Wilkins 2015, 2017). Ghisellini (2003) propose that radio-quiet Seyferts may be hosting aborted jets that work intermittently, and whose ejecta do not reach the escape velocity of the system, but rather fall back colliding with other ejected blobs.
This description of the X-ray emitting region has been especially appropriate for the NLS1 335 ($z=0.025$). 335 has been extensively studied since falling to an X-ray dim state in 2007 (e.g. Grupe 2007, 2008). Over the past twelve years, 335 has exhibited persistent flickering and occasional high-amplitude flares (e.g. Grupe 2012; Wilkins 2015; Gallo 2018). While the low flux state appears consistent with a compact corona (e.g. Grupe 2007; Gallo 2013, 2015; Parker 2014; Wilkins & Gallo 2015), the flares have been attributed to a corona that may be attempting to launch material as in the aborted jet scenario (e.g. Gallo 2013, 2015; Wilkins 2015). During these flaring events, X-ray spectral modelling suggests the compact corona is beaming emission away from the disc.
In December 2015, 335 dropped to its lowest X-ray flux ever recorded triggering a and Target-of-Opportunity observation (PI: Longinotti) to study the UV and warm absorbers in the X-ray low state (Longinotti in prep). Here, the broadband X-ray data are examined with the blurred reflection model, which show evidence for the simultaneous onset of a collimated outflow and an accretion disc wind in the NLS1.
Observation and data reduction {#sect:data}
==============================
335 was observed for $\sim140\ks$ with (Jansen 2001) starting 30 December, 2015 as part of a coordinated campaign with the Hubble Space Telescope () to study the ionised absorbers and emitters in the UV and X-ray low-flux state. The observation was triggered by monitoring (Grupe 2015) and the - analysis is reported by Longinotti (in prep).
The data from the European Photon Imaging Camera (EPIC) are presented here. During the observations the EPIC detectors (Strüder 2001; Turner 2001) were operated in full-window mode and with the thin filter in place. The Observation Data Files (ODFs) were processed to produce calibrated event lists using the Science Analysis System ([SAS v16.0.0]{}). Light curves were extracted from these event lists to search for periods of high background flaring. Such periods were neglected resulting in a good-time exposure of $\sim117\ks$.
Spectra were extracted from a $35$ arcsec circular region centred on the source. The background photons were extracted from an off-source region on the same CCD. Pile-up was negligible during the observations. Single and double events were selected for the pn detector, and single-quadruple events were selected for the Metal Oxided Semi-conductor (MOS) detectors. EPIC response matrices were generated using the [SAS]{} tasks [ARFGEN]{} and [RMFGEN]{}. The MOS and pn data were compared for consistency and determined to be in agreement within known uncertainties.
The spectra were optimally binned following Kaastra & Bleeker (2016). Spectral fitting was performed using [XSPEC v12.9.1]{} (Arnaud 1996) and fit quality is tested using the $C$-statistic (Cash 1979). All parameters are reported in the rest frame of the source unless specified otherwise, but figures remain in the observed frame. The quoted errors on the model parameters correspond to a 90% confidence level for one interesting parameter. A value for the Galactic column density toward 335 of $3.56 \times 10^{20}\pscm$ (Kalberla 2005) is adopted in all of the spectral fits and abundances are from Wilms (2000).
Spectral variability {#sect:spec}
====================
Flux resolved spectroscopy
--------------------------
The $0.2-10\keV$ light curve of 335 shows the NLS1 exhibiting quiescent behaviour in the first $\sim120\ks$ of the observation before the onset of a large amplitude flare in the last $\sim20\ks$ (Fig. \[fig:lc\]). During the flaring period, the count rate increases by at least a factor of five before the observation ends. During the average low-flux interval, 335 is at its lowest flux ever observed with . For comparison, the spectrum during the first $120\ks$ is compared to the previous low-flux state captured in 2007 (Fig. \[fig:lowspec\]). The 2015 dim state shows overall lower flux across the band and marked hardening below $\sim2\keV$ compared to the 2007 low state.
The 2015 data are divided to create spectra in six flux intervals as marked in Fig. \[fig:lc\]. There is an average low state constituting data before $\sim120\ks$ in the light curve and a flare spectrum (average high) that includes data after $120\ks$. Data during the flare are divided into four, $5\ks$ intervals of increasing flux (A, B, C, and D). Although, spectrum D contains the highest flux data it is also of the shortest exposure (i.e. good-time interval, $\sim2.2\ks$). The six spectra are fitted simultaneously.
Only the blurred reflection model is considered in the examination of these data. The exercise of testing various models for 335 has been carried out in several earlier works (e.g. Gallo 2013, 2015; Grupe 2008, 2012). Gallo (2015) demonstrate the blurred reflection model is statistically preferred over partial-covering models. Partial covering is adopted to describe the continuum in these current data in the work of Longinotti (in prep). The testing of other models on this particular data set is the subject of future work.
In the blurred reflection model (e.g. Ross & Fabian 2005), the primary continuum source is a corona situated in some geometry above the disc and radiating isotropically in the rest-frame. Some of the primary emission (power law continuum) will illuminate the accretion disc producing back-scattered emission called the reflection spectrum, which in the energy band, is composed most prominently of fluorescent emission lines. Arising in the rapidly rotating accretion disc close to the supermassive black hole, the reflection spectrum will be blurred by Keplerian and relativistic (special and general) effects.
This scenario is modeled in [xspec]{} using the models listed in Table \[tab:fits\]. The model resembles that fitted to the spectra from the low-flux state analysis (Gallo 2015). Given the additional sensitivity below $\sim1\keV$ provided by , the model in this current work also includes a warm absorber and collisionally ionised emitter ([mekal]{} in [xspec]{}) that are phenomenologically consistent with the absorption and emission features found in the RGS analysis (Longinotti in prep).
Parameters that are not expected to vary over the decade-long time scales since 335 has been monitored (e.g. black hole spin, inclination, and iron abundance) have been left free to vary rather than constrained to previously measured values. The measurements in this analysis are consistent with earlier work, but a multi-epoch analysis (e.g. Wilkins & Gallo 2015; Keek & Ballantyne 2016) is left for the future.
The variability between the different flux intervals can be described very simply with changes in only three parameters: the power law normalisation, power law photon index ($\Gamma$), and disc ionisation parameter ($\xi=4\pi F/n$ where $n$ is the hydrogen number density and $F$ is the incident flux). The model fits the data well ($\Cdof=851/606$, Table \[tab:fits\]). The behaviour is very similar to that previously observed from 335 (Gallo 2013) and other NLS1s (e.g. Bonson 2018; Chiang 2015). The fact these parameters vary together is expected in the blurred reflection scenario as the power law is illuminating and ionising the inner accretion disc. Allowing only one component to vary independent of the other (e.g. either the power law parameters alone or the reflection parameters alone) increases the C-statistic significantly ($\Delta C > 169$ for 10 fewer free parameters).
The most prominent residuals that remain are seen between $7-8\keV$ in the average flare spectrum (lower panel, Fig. \[fig:meanfit\]). The residuals resulting from fitting the flare intervals ($A-D$) are shown in Fig. \[fig:highfit\]. The excess residuals between $7-8\keV$ are present in all the flare-intervals except for segment D, which has the highest flux, but shortest net exposure and fewest counts.
The same model simply scaled by a constant, fits the high- and low-flux MOS data equally well ($\Cdof=569/445$). The background becomes significant above $\sim5\keV$, but residuals at $E\sim7.5\keV$ are apparent in the high-flux spectra that are not present in the low-flux spectra (Fig. \[fig:meanfit\], lower panel). An inverted Gaussian profile at $E=7.58^{+0.11}_{-0.34}\keV$ and with unconstrained width improves the residuals in the MOS spectrum. An equivalent profile at $E=7.58^{+0.18}_{-0.34}\keV$ and $\sigma<300\eV$ improves the residuals in the pn spectrum. The residuals in the MOS instruments appear statistically consistent with those detected in the pn.
The model is self-consistent across all flux intervals and in agreement with previous observations of the source. Values of the distant reflector, warm absorber and ionised emitter are comparable to those measured in other works (e.g. Longinotti 2008, 2013, in prep; Grupe 2008; Gallo 2013, 2015) and do not appear to vary over the course of the observation. Some narrow features are missed when applying the pn model to the RGS data. The physical origin of the ionised emission features may be associated with some central starburst component, but it is difficult to determine if the emission arises in a collisional or photoionised plasma. On the other hand, the narrow emission features might also be originating from the distant narrow-line region.
The measurements of the blurring and accretion disc are in accord with the analyses of other data obtained at previous epochs and with different instruments (e.g. Grupe 2008, 2012; Gallo 2013, 2015; Parker 2014; Walton 2013). In general, the X-ray emission from 335 is consistent with originating from a compact corona, as noted by the steep inner emissivity profile ($q_{in}\approx6.4$) around a rapidly spinning black hole with a dimensionless spin parameter $a=cJ/GM^2 >0.986$, where $J$ is the angular momentum of a black hole of mass $M$. The high spin measurement is consistent with all previous spin measurements for 335 and robust to the cautions raised by Bonson & Gallo (2016). The accretion disc, which extends to the innermost stable circular orbit (ISCO), is modestly ionised ($\xi\approx 30-50\erg\cmps$) and overabundant in iron by a factor of $\sim3.7$. The reflection fraction ($\mathcal{R}$, the ratio between reflected and continuum flux between $0.1-100\keV$) is always greater than unity indicating the spectrum is dominated by the reflection component most likely because of light-bending effects (e.g. Miniutti & Fabian 2004).
Correlations between the different parameters in the low and high-flux intervals (black, circles) are shown in Fig. \[fig:corrs\]. A linear fit is applied to all the relations (solid, red curve) and in all cases the resulting correlation coefficient is high ($\approx \pm0.8$) indicating the trends are important. The behaviour is as expected from previous studies of AGN variability and what is anticipated in the blurred reflection scenario.
The power law photon index steepens with increasing power law flux (Fig. \[fig:corrs\], top left) as is expected in accreting sources (e.g. Wang 2004). The ionisation parameter (Fig. \[fig:corrs\], top right) and reflected flux (Fig. \[fig:corrs\], lower left) also increase as the power law flux rises during the flare. This is expected if there is some positive correlations with the number of photons striking the disc and number of photons reflected into the observers line-of-sight. The relation is corroborated in the lower right panel of Fig. \[fig:corrs\], demonstrating the expected trend between ionisation parameter and reflected flux. However, as the power law flux rises the reflection fraction ($\mathcal{R}$) diminishes suggesting that the relative contribution of the reflection spectrum becomes less important as the continuum brightens (Fig. \[fig:corrs\]). This might be explained, though not exclusively, if the source is emitting anisotropically.
Keek & Ballantyne (2016) investigated similar relations in multi-epoch data of 335 spanning $\sim14$ years. There are slight differences in the models used and parameters investigated that make direct comparison difficult, but it is clear that similar trends exist in the long-term data. Importantly, the Keek & Ballantyne (2016) results support the notion that the corona geometry is changing and that corona expansion might be in the vertical direction (see also Ballantyne 2017).
Model-independent analysis
--------------------------
The spectral variability is further examined by considering a flux-flux (e.g. Taylor 2003; Kammoun 2015) and principal component analysis (PCA) (e.g. Parker 2015).
The PCA is calculated for the 2015 observation using $5\ks$ intervals of the data. When comparing the variability of each principal component on the log-eigenvalue diagram (LEV) as described in Parker (2015), the first two eigenvectors appear significant above the noise level. The primary and secondary principal components (i.e. PC1 and PC2) account for approximately $93$ and $3$ per cent of the variability in the spectrum, respectively (Fig. \[fig:pca\]). The shape of PC1 and PC2 as a function of energy can reveal the physical component in the source spectrum that is driving the variability.
As PC2 crosses zero normalisation (i.e. the normalisation is positive at low energies and negative at high energies), this indicates the high- and low-energy variations are anti-correlated. Such behaviour is commonly associated with a pivoting power law (e.g. Parker 2015; Gallant 2018). In contrast, PC1 shows variability that is correlated in all energy bands (i.e. all data points are positive), however the degree of variability is energy dependent. Between $0.3-2\keV$, the variability is roughly comparable in each energy band, however as the energy increases above $\sim2\keV$, the variations diminish.
The spectral model above (Table \[tab:fits\]) is tested to determine if the scenario could account for the PCA. Spectra are created using the best-fit model above and allowing the power law normalisation, photon index, and ionisation parameter to vary in a correlated manner. The PCA is calculated from these simulated spectra that take into consideration the source brightness and background.
The PCA determined from the simulation are overplotted as a red band on the data in Fig. \[fig:pca\]. The simulation reproduces the general shape of PC1 and PC2 relatively well. Even sharp features in PC1 at energies of $\sim0.8-1\keV$ and $\sim6\keV$ that are likely associated with the warm absorber, emitter, or distant reflector are present and well-matched in the simulation.
Flux-flux plots are created comparing a soft ($0.4-0.8\keV$) and hard ($4-8\keV$) band to the $1-2\keV$ band (Fig. \[fig:ffp\]). Linear and power law models fitted to the flux-flux plots of 335 in the 2015 low state are of poor quality ($\redchi >2$). The same set of simulated data used for the PCA were manipulated to create flux-flux plots that represent the scenario derived above. These theoretical relations are overplotted on the 2015 data and show very good agreement (red curves in Fig. \[fig:ffp\]). The data from the 2007 low-flux state, which did not exhibit marked variability during the observation, are also included in the flux-flux plot and appear completely consistent with the 2015 model.
The blurred reflection scenario describes well the flux-resolved spectra as well as the variability behaviour resolved with the PCA and flux-flux plots. The variability can be completely described by brightening of the power law component (i.e. the corona) that likely drives the correlated changes in the photon index and the accretion disc ionisation parameter.
The low-state coronal geometry {#sect:corona}
==============================
To probe the geometry of the corona, the emissivity profile of the accretion disc (that is the pattern of illumination of the disc by the coronal X-ray source) is measured. The emissivity profile of the disc is encoded in the profile of the relativistically broadened line with the specific Doppler shift and redshift varying as a function of position on the disc, thus variations in reflected flux as a function of radius on the disc result in changes to the precise shape of the redshifted wing of the line.
We initially seek to measure the emissivity profile by fitting the profile of the line as the sum of contributions from successive annuli on the disc, following the method of Wilkins & Fabian (2011). However the signal to noise in the detection of the line above the continuum during this single orbit observation in the low flux state is insufficient to constrain the emissivity of the disc at each radius.
Instead, we are guided by the suggested broken-power law emissivity profile in combination with the emissivity profile of the accretion disc measured in previous observations of 335 with in 2006, 2007 and 2009, in 2006 and 2013, and in 2013 and 2014. The emissivity profile of the accretion disc was found to vary between a twice-broken power law, steeply falling in the inner regions, flattening to constant emissivity before breaking to the classically expected power law index of 3 over the outer disc, indicating a corona that is radially extended over the surface of the inner disc, during higher flux epochs to a once broken power law during the low flux epochs indicating illumination of the accretion disc by a compact X-ray source close to the black hole (Wilkins & Gallo 2015; Wilkins 2015). The outer break radius indicates the radial extent of the corona over the disc (Wilkins & Fabian 2012).
Instead of fitting a free function to the emissivity profile at each radius in the disc, we fit the profile of the line with a single profile with a twice-broken power law emissivity profile. We use this model to test for evidence of any extension of the corona over the inner regions of the accretion disc. The power law index of the middle region of the profile is frozen at zero (the flattened portion of the profile produced by the extended portion of the corona), while the inner and outer power law indices and the two break radii are fit as free parameters to the observed line in the spectrum from $1-10\keV$. Having run an initial fit by minimising the C-statistic with respect to the free model parameters, we run Markov Chain Monte Carlo (MCMC) calculations from this starting point in parameter space to obtain an estimate of the probability distribution of each of the parameters of the emissivity profile. MCMC was performed using the Goodman-Weare algorithm implemented through [xspec\_emcee]{},[^1] starting the walkers close to the best-fitting parameters found during the initial [xspec]{} fit, perturbed by random increments with standard deviation from the covariance matrix found during the fit. The MCMC used 80 walkers, running for 10000 steps, burning the first 1000 steps of each chain. We fit the model to the low flux state (before the flare) and the high flux state (during the flare) independently to understand the change in the corona as the flare is seen. Fig. \[fig:emis\] (centre and right panels) shows the probability distributions and contours between these parameters obtained from the MCMC analysis.
During both the low flux interval and the flare, the observed spectrum is consistent with there being no radial extension of the corona over the inner regions of the disc, hence the disc being illuminated by a central, compact source. This can be seen in the distribution in the difference between the break radii, showing the length of the flattened section (when $R_{b2} < R_{b1}$, the model produces simply a once broken power law breaking at $R_{b1}$ to the outer power law index). During the low flux interval, the difference in break radii is less than zero at the 94 per cent confidence level and less than 0.5 at the 98 per cent level. $R_{b2}$ is less than $5\rg$ at 96 per cent confidence. During the flare, the difference between the break radii is less than zero only at the 72 per cent confidence level, though less than 0.5 at the 90 per cent confidence level. The outer break radius is again less than $5\rg$ at 90 per cent confidence and less than $6\rg$ at 97 per cent confidence. It should be noted that from the twice-broken power law emissivity profile, it is not possible to measure radial extent of the corona below $5\rg$ since the inner steepening of the emissivity profile masks an outer break radius less than this, hence a measure of the break radius less than $5\rg$ is consistent with a compact corona that lies within this radius.
The most notable difference between the accretion disc emissivity profile between the low flux interval and the flare is the slope of the power law over the outer part of the accretion disc. During the low flux interval, the outer power law index is measured to be ($2.3\pm0.3$) revealing a flatter illumination profile over the outer disc than is expected typically from a compact X-ray source. Combined with the observation of a high reflection fraction, suggesting the bulk of the coronal emission is coming from close to the black hole such that the emission is focused onto the inner disc by strong light bending (e.g. Miniutti 2003; Suebswong 2006; Dauser 2013; Chainakun & Young 2015), This flattening of the profile may be indicative of a slight vertical extension of the corona above the black hole as was seen during the 2006 observations of 335 with (Wilkins & Gallo 2015), though it is difficult to infer this from these data alone.
On the other hand, during the flare, the illumination of the outer disc falls off very steeply with radius, with the outer index greater than 3 at the 98 per cent confidence level and greater than 4 at the 90 per cent confidence level. This finding implies that little illumination reaches the outer parts of the accretion disc which would be expected in the case of mildly relativistic vertical motion of a centrally-collimated corona. Beaming of emission away from the disc results in few X-rays reaching the outer parts of the disc, hence producing the steep profile, while the inner disc is still strongly irradiated by X-rays that are bent down towards the black hole and inner disc in the strong gravitational field close to the black hole. Beaming of emission away from the disc also results in a low reflection fraction as the majority of the flux emitted by the corona is emitted upwards, being detected directly as continuum emission rather than being reprocessed by the disc.
The onset of outflows {#sect:outflow}
=====================
Outflows have been previously reported for 335 (e.g. Longinotti 2008, 2013, in prep). These outflows have intermediate velocities that are faster than those associated with warm absorbers ($\ls2000\kmps$), but slower than those defined as ultrafast ($\sim10^4\kmps$) outflows. Highly ionised and ultrafast outflows have also been reported during a previous low X-ray flux state of 335 (e.g. see figures 3 and 4 of Parker 2014). These different velocity winds could have a common origin if the lower velocity outflows are manifestations of the ultrafast outflows that collide with the interstellar medium (ISM) (e.g. King & Pounds 2014).
The inset in the lower panel of Fig. \[fig:meanfit\], clearly shows negative residuals between $7-8\keV$ in the flaring spectrum. Various continuum models are used to test the importance of the residuals. The strength of the residuals do depend on the continuum model used, but they remain apparent in all cases. The residuals could be attributed to highly ionised iron that is significantly blueshifted.
Before further discussion of implications, the statistical importance of these features is examined in a robust manner. To determine the significance of features in the $1-10\keV$ band, a Monte Carlo simulation is carried out. One-thousand spectra of the low- and high-flux states are simulated adopting the best-fit model (Table \[tab:fits\]), signal-to-noise, and optimal binning used to analyse the true data. The high- and low-flux spectra are fit simultaneously over the entire $0.4-10\keV$ band and the residuals are examined between $1-10\keV$ at each flux interval. Constraining a physical continuum model over the entire broadband is more meaningful than using a power law over a limited energy range to examine the residuals. A Gaussian profile with a width of $10\eV$ is stepped by 0.1 keV steps through each spectrum separately and the improvement to the fit ($\Delta C$) is recorded, to determine the significance ($S$) of deviations at each energy. The same technique is then applied to the real spectra resulting in the significance (contour) plots in Fig. \[fig:mc\]. In the case of the real spectra, the model is linked between the flux states as in Section \[sect:spec\] (Table \[tab:fits\]), and a Gaussian profile is added and left free to vary in normalisation for the spectrum of interest while set to zero normalisation for the accompanying spectrum.
Interestingly, between $1-10\keV$, the low-flux spectrum does not show any significant deviations beyond what might be expected from random statistical fluctuations (top panel, Fig. \[fig:mc\]). However, the flare spectrum shows several deviations, both positive and negative, between the $2-3\sigma$ level (lower panel, Fig. \[fig:mc\]). Above $\sim9\keV$, residuals are regarded less important as the background level becomes noteworthy (Fig. \[fig:meanfit\]).
In Fig. \[fig:mc\], there are residuals between $2-3\sigma$ seen in absorption in the $7-8\keV$ band. Features in this energy band are often attributed to outflowing He- and H-like . The addition of two narrow ($\sigma=1\eV$) Gaussian absorption features at fixed energies of $6.7$ and $6.97\keV$ fit the residuals well with a common blueshift velocity of $v=0.12^{+0.08}_{-0.04}c$.
A highly ionised gas producing strong He- and H-like would also generate other spectral features from species at lower atomic numbers. An [xstar]{} grid with variable column density, ionisation parameter, and iron abundance is fitted to the high-flux spectrum to physically model the residuals. The inclusion of the grid improves the residuals, but is not well-constrained with the current data quality. The best fit parameters of the outflowing absorber are $\nh\approx10^{23}\pscm$, $\xi\approx10^4\erg\cmps$, $A_{Fe}=2$ solar, and $v=0.12\pm0.02c$. The stronger features produced by such a plasma, blueshifted by a velocity of $v\sim0.12c$ could reproduce some of the contours as indicated in Fig. \[fig:mc\].
Comparable features were reported in the 2013 observation by Parker (2014). Unfortunately, in the low-state observations that most closely resembles these current data in flux (Gallo 2015), the spectral region was background dominated by a Ni feature at $7.5\keV$.
Features of comparable $\sim2\sigma$ significance are seen in emission at lower energies of . Narrow Gaussian profiles can be fitted at $4.66^{+0.08}_{-0.09}$ and $5.38^{+0.11}_{-0.09}\keV$. The measured values correspond to rest-frame energies of Ti K$\alpha$ and Cr K$\alpha$, respectively. However, the features are much stronger than would be expected in a plasma of cosmic abundances and could just be resulting from an imperfect fit to the relativistic broad line.
Discussion
==========
The 2015 observation of 335 caught the NLS1 in its lowest X-ray flux state since the AGN became X-ray weak in 2007 (Grupe 2007, 2008). This extended X-ray weak state has been marked with continuous flickering in the X-ray and UV light curves, and occasional high amplitude X-ray flares in which the NLS1 can become $\sim50\times$ brighter (e.g. Grupe 2013; Wilkins 2015; Gallo 2018).
Since the onset of the monitoring in this X-ray weak state, there have been suggestions (Gallo 2013) that the X-ray flaring in 335 could be attributed to the base of an aborted jet (e.g. Ghisellini 2003). The jet-like corona may have even prevailed when 335 was in the stable, X-ray bright state prior to 2007. During that epoch the spectra were dominated by the power law component and exhibited low reflection fractions ($\mathcal{R}<1$) (e.g. Gallo 2015; Wilkins & Gallo 2015), which could be attributed to a corona moving at high velocity away from the disc.
The interpretation relies on the determination of the emissivity profile and simultaneous measurement of the reflection fraction when the source flares. In some NLS1s, the brightness is associated with the radial expansion of the corona over the disc. For instance, in h07, when the source was bright, the corona extended radially outwards, whereas when dim the source was confined to a small region close to the black hole (Wilkins 2014).
However, in previous observations of 335, the corona presumably extends vertically during flaring events. At the same time, the reflection fraction ($\mathcal{R}$) is seen to decrease even to values less than unity. This would indicate that the corona is illuminating anisotropically, preferentially away from the accretion disc. This might be expected from a corona that is moving at high velocities away from the disc and whose illumination is beamed (e.g. Beloborodov 1999; Gonzalez 2017b). This behaviour leaves signatures in the reverberation lag-energy spectrum (Wilkins 2016) that have been observed in 1 (Wilkins 2017), another NLS1 that shares a similar interpretation for its X-ray behaviour (e.g. Gallo 2007; Gallo 2018). Such timing signatures have not been reported in 335 since recent flaring observations were obtained with and , which lack the timing resolution to carry out such work effectively.
During the 2015 observation, 335 remained quiescent for the first $\sim120\ks$ before flaring in the last $\sim20\ks$. The end of the observation truncated the flare as it was still brightening (Fig. \[fig:lc\]). Though the complete flare was not recorded, the behaviour of 335 during this time was completely consistent with previous observed flares in the AGN. The source remained compact as it brightened, but the emissivity profile of the outer disc steepened indicating less illumination was reaching the disc at large distances (Section \[sect:corona\]). The power law spectrum softened (dropping in temperature or opacity) indicating the corona could be extending vertically since no radial extent was measured. Simultaneously, the reflection fraction dropped significantly from $\sim15$ to $\sim3$ in about $20\ks$ (Section \[sect:spec\]).
Values of $\mathcal{R}<1$ that would indicate beamed emission are never measured during the observation. However, in conjunction with previous analyses, the rapid decline in $\mathcal{R}$ is still supportive of the aborted jet scenario. During the flare the dynamic corona in 335 may be collimating and moving away from the disc. Alternatively, the corona could also have a blended configuration as suggested for 1 (e.g. Gallo 2007; Wilkins 2017), but it would still need to be rather compact.
While it is not certain if the reflection fraction continues to fall after the observation ends, the value of $\mathcal{R}\approx3$ can be used to estimate the velocity of the corona as a function of height ($h$) above the black hole (equation 15 of Gonzalez et al. 2017b) (Fig. \[fig:rzb\]). Under the assumptions of a maximally spinning Kerr black hole and a compact source geometry aligned with the spin axis of the black hole (i.e. “lamp-post” geometry), we may use the measured value of $\mathcal{R}$ at the peak of the flaring event to evaluate the source velocity (i.e. corona outflow velocity) at different source heights above the black hole. As the final measured value of $\mathcal{R}$ is still relatively high, we find the source remains close to the black hole ($h\ls4\rg$) for all possible velocities.
The escape velocity from the system has also been calculated using the black hole mass estimate for 335 of $2.5\pm0.3 \times 10^7 M_\odot$ (Grier 2012; similar mass as measured by Du 2014). For all combinations of velocity and height, the source never exceeds the escape velocity of the system (Fig. \[fig:rzb\]) so the material will likely fall back toward the accretion disc. Note, that despite the assumption of a compact source, Gonzalez (2017b) do not find the results for a collimated source to be significantly different.
The conjectured collimated, outflowing corona, may also be accompanied by a wind-like outflow (Section \[sect:outflow\]). The investigation of ultrafast outflows often relies on the presences of absorption features at energies above $\sim7\keV$. During the quiescent phase of the observation ($\ls120\ks$) no such features were present in the spectrum. However, during the flaring event, a number of features were detected between the $2-3\sigma$ level. Residuals were seen in emission and absorption, but an absorption feature at $\sim7.5\keV$ in the observed frame could be be attributed to highly ionised iron in outflow (Fig. \[fig:mc\]). Similar residuals at $\sim7.5\keV$ were also seen in the 2013 observation of 335 (e.g. see figures 3 and 4 of Parker 2014) . Comparison of the absorption features with a highly ionised plasma ($\xi=10^4\erg\cmps$) describe the data reasonably well with a single velocity shift of $v\sim0.12c$. The feature at $7.5\keV$ is physically described as arising from both He- and H-like iron. Other residuals in Fig. \[fig:mc\] might also be attributed with Ar <span style="font-variant:small-caps;">xviii</span>, S <span style="font-variant:small-caps;">vi</span>, and Ne <span style="font-variant:small-caps;">x</span>.
Alternatively, a disc origin for the absorption features seen in some AGN was proposed by Gallo & Fabian (2011, 2013). Similar spectral features could be reproduced if the ionising material were corotating with the disc rather than outflowing. However, in such a scenario, the absorption features would become stronger with increasing reflection fraction, which is in contrast to what is observed in 335.
The variability of the possible ultrafast wind in 335 is different than that described for the possible wind in another NLS1, 13 (Parker 2017; Pinto 2018; Jiang 2018). In 13, which is likely accreting near its Eddington limit, the wind features were seen when the source was dim and not when is was bright. Parameters other than radiation pressure might be driving the wind (e.g. King & Pounds 2003). The interpretation was that the wind is perhaps always present, but over-ionised when 13 was bright and therefore no wind features were observed. This is opposite of what is observed in 335, which is a sub-Eddington source (i.e. $L_{bol}/L_{Edd} < 1$) in the low-flux state (e.g. Keek & Ballantyne 2016). Here, like in X-ray binaries, the potential ultrafast outflow features are seen during the bright, flaring state and not when the source is dim. It could be that the wind in 335, is more massive and simply requires more energy to be moved and that the radiation pressure is not sufficient to launch the wind during the quiescent state and can only do so when the flare occurs.
On the other hand, the absorption features were not reported in the pre-2007 observations when 335 was $\sim10$-times more luminous (Crummy 2006; Longinotti 2007, 2013; Larsson 2007; Gallo 2015; Wilkins & Gallo 2015) despite evidence that the high-flux state maybe a more extreme example of the flare states. It could be that in the bright state (pre-2007 or prolonged flaring events since 2007), 335 behaves like 13 and over-ionises the purported wind.
If the flare originates in the compact corona as proposed, than the wind must be launched from nearby. The absorption features are reported in the average flare spectrum and they are even less significant in the short $\sim5\ks$ flare intervals. This would imply that the features appear on time scales of $<10\ks$ from the onset of the flare, which would correspond to a light-travel distance of $<80\rg$ from the compact corona, assuming a mass of $2.5\times10^7\Msun$ (Grier 2012) for the black hole in 335. The escape velocity at $80\rg$ is $\sim0.15c$ so such a wind could be ejecting energy out of the AGN environment.
In the scenario described by King & Pounds (2014), if this wind were to collide with the ISM, shocks could generate the fast warm absorber velocities reported by Longinotti (2013). It is worth noting that the star formation rate relative to AGN luminosity in 335 is the lowest seen in NLS1s (Sani 2010). This could be indication that the star formation in 335 has been suppressed by past activity that was similar to what we are witnessing currently.
Conclusions
============
A triggered $140\ks$ observation of the NLS1 335 finds the AGN in a relatively quiescent phase in the first $120\ks$ before flaring. The data are compared with the blurred reflection model. The spectrum is consistent with the blurred reflection scenario in which the corona is compact around a Kerr black hole and light bending is important. The variability analysis (flux-flux plots and PCA) as well as flux-resolved spectroscopy demonstrates that the variability can be attributed to brightening and softening of the power law component (i.e. the corona) and increased ionisation of the inner accretion disc. Of particular interest is that during the flare, the primary continuum is consistent with originating from a corona that might be extending vertically and beaming away from the disc, and that features appearing in the spectrum could arise in an accretion disc wind outflowing at $v\sim0.12c$. If confirmed, this is the first example of a radio-quiet AGN exhibiting behaviour consistent with diffuse and collimated outflow. Understanding such objects is necessary to determine the origin of the launching mechanism of winds and jets.
Acknowledgments {#acknowledgments .unnumbered}
===============
The project is an ESA Science Mission with instruments and contributions directly funded by ESA Member States and the USA (NASA). We are grateful to the observing team for preparing and activating the ToO. Many thanks to the referee for a thoughtful review and helpful comments.
Anders E., Grevesse N., 1989, GeCoA, 53, 197
Arnaud K., 1996, in: [*Astronomical Data Analysis Software and Systems*]{}, Jacoby G., Barnes J., eds, ASP Conf. Series Vol. 101, p17
Ballantyne D. R., 2017, MNRAS, 472, 60
Beloborodov A. M., 1999, MNRAS, 510, 123
Berton M. 2018, A&A, 614, 87
Blandford R. D., Znajek R. L., 1977, MNRAS, 179, 433
Blandford R. D., Payne D. G., 1982, MNRAS, 199, 883
Bonson K., Gallo L. C., 2016, MNRAS 458, 1927
Bonson K., Gallo L. C., Wilkins D. R., Fabian A. C., 2018, MNRAS, 477, 3247
Bridle A. H., Perley R. A., 1984, ARA&A, 22, 319
Cash W., 1979, ApJ, 228, 939
Chainakun P., Young A. J., 2015, MNRAS, 452, 333
Chiang Chia-Ying, Walton D. J., Fabian A. C., Wilkins D. R., Gallo L. C., 2015, MNRAS, 446, 759
Crummy J., Fabian A., Gallo L., Ross R., 2006, MNRAS, 365, 1067
Dauser T., García J., Wilms J., Böck M., Brenneman L. W., Falanga M., Fukumura K., Reynolds C. S., 2013, MNRAS, 430, 1694
Du P. 2014, ApJ, 782, 45
Fabian A. C., 2012, ARA&A, 50, 455
Foschini L. 2015, A&A, 575, 13
Gallant D., Gallo L. C., Parker M. L., 2018, MNRAS, 480, 1999
Gallo L. C., 2018 (arXiv 1807.09838)
Gallo L. C., Brandt W. N., Costantini E., Fabian A. C., 2007, MNRAS, 377, 1375
Gallo L. C., Fabian A. C., 2011, MNRAS, 418, 59
Gallo L. C., Fabian A. C., 2013, MNRAS, 434, 66
Gallo L. C. 2013, MNRAS, 428, 1191
Gallo L. C. 2015, MNRAS, 446, 633
Gallo L. C., Blue D. M., Grupe D., Komossa S., Wilkins D. R., 2018, MNRAS, 478, 2557
Giroletti M., Panessa F., 2009, ApJ, 706, 260
Giroletti M., Panessa F., Longinotti A. L., Krongold Y., Guainazzi M., Costantini E., Santos-Lleo M., 2017, A&A, 600, 87
Ghisellini G., Haardt F., Matt G., 2003, A&A, 413, 535
Gonzalez A. G., Waddell S. G. H., Gallo L. C., 2017a, MNRAS, 475, 128
Gonzalez A. G., Wilkins D. R., Gallo L. C., 2017b, MNRAS, 472, 1932
Grier C. J. 2012, ApJ, 744, 4
Grupe D., Komossa S., Gallo L. C., 2007, ApJ, 668, 111
Grupe D. 2008, ApJ, 681, 982
Grupe D., Komossa S., Gallo, L. C., Longinotti A. L., Fabian, A. C., Pradhan A. K., Gruberbauer M., Xu D, 2012, ApJS, 199, 28
Hopkins P. F., 2011, MNRAS, 420, L8
Jansen F. 2001, A&A, 365, L1
Jiang J. 2018, MNRAS, 477, 3711
Kaastra J. S., Bleeker J. A. M., 2016, A&A, 587, 151
Kalberla P. M. W., Burton W. B., Hartmann D., Arnal E. M., Bajaja E., Morras R., Pöppel W. G. L., 2005, A&A, 440, 775
Keek L., Ballantyne D. R., 2016, MNRAS, 456, 2722
King A. R., Pounds K. A., 2003, MNRAS, 345, 657
King A. R., Pounds K. A., 2014, MNRAS, 437, 81
Komossa S., 2018 (arXiv:1807.03666)
Krongold Y., Nicastro F., Brickhouse N. S., Elvis M., Liedahl D. A., Mathur S., 2003, ApJ, 597, 832
Lähteenmäki A., Järvelä E., Ramakrishnan V., Tornikoski M., Tammi J., Vera R. J. C., Chamani W., 2018, A&A, 614, 1
Larsson J., Miniutti G., Fabian A. C., Miller J. M., Reynolds C. S., Ponti G., 2008, MNRAS, 384, 1316
Longinotti A. L., Sim S., Nandra K., Cappi M., O’Neill P., MNRAS, 2007, 374, 237
Longinotti A. L., Nucita A., Santo-Lleo M., Guainazzi M., 2008, A&A, 484, 311
Longinotti A. L. 2013, ApJ, 766, 104
Longinotti A. L. 2015, ApJ, 813, 39
Matzeu G. A., Reeves J. N., Nardini E., Braito V., Costa M. T., Tombesi F., Gofford J., 2016, MNRAS, 458, 1311
Miniutti G., Fabian A.C., Goyder R., Lasenby A.N., 2003, MNRAS, 344, L22
Miniutti G., Fabian A.C., 2004, MNRAS, 349, 1699
Murray N., Chiang J., Grossman S. A., Voit G. M., 1995, ApJ, 451, 498
Parker M. L. 2014, MNRAS, 443, 1723
Parker M. L. et al. 2015, MNRAS 447, 72.
Parker M. L. 2017, Nature, 543, 83
Pinto C. 2018, MNRAS, 476, 1021
Proga D., Kallman T. R., 2004, ApJ, 616, 688
Ross R. R. & Fabian A. C. 2005, MNRAS 358, 211.
Sani E. 2010, MNRAS, 403, 1246
Strüder L. 2001, A&A, 365, L18
Suebsuwong T., Malzac J., Jourdain E., Marcowith A., 2006, A&A, 453, 773
Taylor R. D., Uttley P., McHardy I. M., 2003, MNRAS 342, L31.
Tombesi F., Cappi M., Reeves J. N., Palumbo G. G. C., Yaqoob T., Braito V., Dadina M. 2010, A&A 521, 35.
Tombesi F., Sambruna R. M., Marscher A. P., Jorstad S. G., Reynolds C. S., Markowitz A., 2012, MNRAS, 424, 754
Tombesi F., Tazaki F., Mushotzky R. F., Ueda Y., Cappi M., Gofford J., Reeves J. N., Guainazzi M., 2014, MNRAS, 443, 2154
Turner, M. J. L. 2001, A&A, 365, 27
Wang J., Watarai K., Mineshige S., 2004, ApJL, 607, L107
Walton D. J., Nardini E., Fabian A. C., Gallo L. C., Reis R. C., 2013, MNRAS, 428, 2901
Wilkins D.R., Fabian A.C., 2011, MNRAS, 414, 1269
Wilkins D.R., Fabian A.C., 2012, MNRAS, 424, 1284
Wilkins D. R., Kara, E.; Fabian, A. C., Gallo L. C., 2014, MNRAS 443, 2746
Wilkins D. R., Gallo L. C., 2015, MNRAS 449, 129
Wilkins D. R., Gallo L. C., Grupe D., Bonson K., Komossa S., Fabian A. C., 2015, MNRAS, 454, 4440
Wilkins D. R., Gallo L. C., Grupe D., Bonson K., Komossa S., Fabian A. C., 2015, MNRAS, 454, 4440
Wilkins D. R., Cackett E. M., Fabian A. C., Reynolds C. S., 2016, MNRAS, 458, 200
Wilkins D. R., Gallo L. C., Silva C. V., Costantini E., Brandt W. N., Kriss G. A., 2017, MNRAS, 471, 4436
Wilms J., Allen A., McCray R., 2000, ApJ, 542, 914
\[lastpage\]
[^1]: Made available by Jeremy Sanders (http://github.com/jeremysanders/xspec\_emcee)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
One of the outstanding issues in neutrino physics is the experimental determination of the neutrino mass hierarchy: Is the order of the neutrino masses “normal” — $m_1^2<m_2^2<m_3^2$ — or is it inverted — $m_3^2<m_1^2<m_2^2$, with $m_2^2-m_1^2\ll m_2^2,m_1^2$? While this issue can be resolved in next-generation long-baseline $\nu_{\mu}\leftrightarrow\nu_e$ neutrino oscillation studies if $|U_{e3}|^2\equiv\sin^2\theta_{13}$ is large enough, a clear strategy on how to resolve it if $|U_{e3}|^2$ is sufficiently small is still lacking.
Here, we study the capability of non-oscillation probes of neutrino masses to determine the neutrino mass ordering. We concentrate on studies of $m_{\nu_e}$, the kinematical neutrino mass to which precise studies of tritium $\beta$-decay are sensitive, $m_{ee}$, the effective mass to which the rate for neutrinoless double-beta decay is sensitive if the neutrinos are Majorana fermions, and $\Sigma$, the sum of the neutrino masses, to which cosmological probes of the energy budget of the Universe are sensitive. We find that combined measurements of $m_{ee}$, $\Sigma$, and $m_{\nu_e}$ are capable of establishing the neutrino mass hierarchy if these measurements are precise enough and if one “gets lucky.” We quantify the previous sentence in detail by performing a numerical analysis of a large number of theoretical data sets, for different measured values of $m_{\nu_e}$, $m_{ee}$, and $\Sigma$, keeping in mind the ultimate sensitivity that can be reached by next (and next-to-next) generation experiments.
author:
- André de Gouvêa
- James Jenkins
title: 'Non-Oscillation Probes of the Neutrino Mass Hierarchy and Vanishing $|U_{e3}|$'
---
Introduction {#sec:intro}
============
In order to explain all neutrino data [@TASI] (with the exception of those from LSND [@LSND], waiting to be confirmed by the ongoing MiniBooNE experiment), one is required to augment the standard model Lagrangian by adding operators that will render, after electroweak symmetry breaking, the neutrinos massive [@neutrino_theory].
Under the most conservative assumptions, the new standard model parameters that need to be determined from experiment are:
1. The three neutrino masses, $m_i$, $i=1,2,3$, chosen real and nonnegative. These are ordered as follows: $m_1^2<m_2^2$, and $0<\Delta m^2_{12}\equiv m^2_2-m^2_1<|\Delta m^2_{13}|$, where $\Delta m^2_{13}\equiv m^2_3-m^2_1$. A positive value of $\Delta m^2_{13}$ implies $m_3^2>m_2^2$ and a so-called normal mass hierarchy, while a negative value of $\Delta m^2_{13}$ implies $m_3^2<m_1^2$ and a so-called inverted mass hierarchy.
The three neutrino masses can be expressed in terms of $\Delta m^2_{12}$, $\Delta m^2_{13}$ (including its sign) and the lightest neutrino mass $m_l$. In the case of a normal mass hierarchy, $m_1=m_l$, $m_2=\sqrt{\Delta m^2_{12}+m^2_l}$, and $m_3=\sqrt{\Delta m^2_{13}+m^2_l}$, while if the mass hierarchy is inverted $m_1=\sqrt{-\Delta m^2_{13}+m^2_l}$, $m_2=\sqrt{-\Delta m^2_{13}+\Delta m^2_{12}+m^2_l}$, and $m_3=m_l$.
2. The six mixing parameters that characterize the leptonic mixing matrix. We adopt the particle data group parameterization of the leptonic mixing matrix (Eq. (13.32) in [@pdg]). Two of the six parameters are unphysical if the neutrinos are Dirac fermions.
Current experiments have been able to measure $\Delta m^2_{12}$ and $|\Delta m^2_{13}|$, together with the “solar angle” $\theta_{12}$ and the “atmospheric angle” $\theta_{23}$. The third mixing angle $\theta_{13}$ is constrained to be small ($\sin^2\theta_{13}<0.03$ at the 99% confidence level [@global_anal]), and we know nothing about the three CP-odd phases $\delta, \alpha_1, \alpha_2$. We also don’t know whether the neutrino mass hierarchy is normal or inverted, and the lightest neutrino mass is only modestly constrained: $m_l\in[0,2.0]$ eV, at the 99% confidence level [@global_anal].[^1] It is among the main goals of fundamental physics experiments to determine the currently unknown leptonic parameters.
In [@ours], together with Boris Kayser, we explored the issue of determining the mass hierarchy with neutrino oscillation experiments. In particular, we discussed the challenges associated with determining the neutrino mass hierarchy in the advent that $\sin^2\theta_{13}$ is vanishingly small. We concluded that, if $\theta_{13}$ happens to be small enough, the only way to determine the neutrino mass hierarchy with neutrino oscillation experiments would be to probe $\nu_{\mu}\to\nu_{\mu}$ and/or $\bar{\nu}_{\mu}\to\bar{\nu}_{\mu}$ oscillations at very long baselines ($L\gtrsim 5000$ km) and relatively small energies ($E_{\nu}\lesssim 500$ MeV). Given the enormous experimental challenges which one needs to overcome in order to achieve the conditions outlined in [@ours], it is of the utmost importance to explore other probes of the neutrino mass hierarchy.
Here, we study non-oscillation probes of neutrino masses and mixing, and determine their capabilities for extracting the neutrino mass hierarchy. We concentrate on three distinct observables: (i) the electron-type neutrino effective mass $m_{\nu_e}$, which can be probed in precise measurements of the end point of nuclear $\beta$ decay spectra, (ii) the effective neutrino mass $m_{ee}$ that characterizes the rate of neutrinoless double double beta decay (if the neutrinos are Majorana particles), and (iii) the sum of the active neutrino masses $\Sigma$, which is constrained, within the concordance cosmological model, by measurements of “cosmological” observables. All three are described in detail in Sec. \[sec:probes\]. We spell out the circumstances ([*e.g.*]{}, how precisely should $m_{\nu_e}$, $m_{ee}$, and $\Sigma$ be constrained) under which the observables listed above can add to our understanding of the neutrino mass spectrum in Sec. \[sec:fits\], emphasizing the importance of combining all three measurements.
We will concentrate on the limit $\sin^2\theta_{13}\to 0$, when “standard” matter affected $\nu_{\mu}\leftrightarrow\nu_e$ searches (see [@ours] and references therein) for the mass hierarchy are known to fail. Under these circumstances, as emphasized above and in [@ours], alternative probes of the character of the neutrino mass are not only welcome but absolutely necessary.
Non-Oscillation Probes of the Neutrino Mass {#sec:probes}
===========================================
The only evidence for neutrino masses comes from long-baseline, neutrino oscillation experiments. These take advantage of quantum mechanical interference in order to be sensitive to very small neutrino mass-squared differences.[^2] On the other hand, neutrino oscillations are impotent when it comes to determining the neutrino masses themselves — they are only sensitive to the differences among the various neutrino masses.
Of course, several other phenomena are also potentially sensitive to nonzero neutrino masses. We discuss three qualitatively distinct “non-oscillation probes” of nonzero neutrino masses, and explore their potential for determining the nature of the neutrino mass hierarchy under the assumption that it will not be revealed by future oscillation experiments.
$m_{\nu_e}$ — Energy Spectrum of Nuclear $\beta$-Decay
------------------------------------------------------
Conservation of energy and momentum dictates that any physical process involving neutrino absorption, emission, or scattering is sensitive to the values of the neutrino masses. Due to the smallness of neutrino masses, however, such effects are, in practice, almost always hopelessly unobservable.
Some of the most promising probes are decay processes in which the neutrino energy and momentum are well constrained and “easy” to measure with great precision and/or as small as possible. Among these are multi-body final state $\tau$ decays, $\pi^+$ decay at rest, and nuclear beta-decay, which are said to constrain, respectively, the tau-type, muon-type, and electron-type neutrino effective masses, which will be properly defined below.
Given the current constraints on neutrino masses-squared and neutrino mixing, however, the most precise probe, by far, of kinematical neutrino mass effects is the precise determination of the end point of the $\beta$-ray energy spectrum of tritium decay. The ‘end point’ is defined as the maximum energy that the daughter $\beta$-ray is kinematically allowed to have.
In a little more detail, the $\beta$-ray spectrum is a function of the three neutrino masses-squared $m_i^2$ and can be written, very schematically, as $$K=|U_{e1}|^2F(m_1^2/E_{\nu}^2,E_{\nu})+|U_{e2}|^2F(m_2^2/E^2_{\nu},E_{\nu})+|U_{e3}|^2F(m_3^2/E_{\nu}^2,E_{\nu}),
\label{eq:mb}$$ where $K$ is the $\beta$-ray spectrum as a function of the neutrino energy $E_{\nu}$,[^3] and $F$ is a function of the neutrino energy and the neutrino mass-squared in units of the neutrino energy. In order to obtain Eq. (\[eq:mb\]), we assume that tritium beta decay is described by $^3{\rm H}\to ^3{\rm He}+e^-+\bar{\nu}_i$, and that the probability that a specific mass eigenstate $\nu_i$ is emitted is $|U_{ei}|^2$.
Eq. (\[eq:mb\]) is independently sensitive to all three neutrino masses as long as $|U_{ei}|^2$ are known (which is, to a good approximation, the case) [@farzan_smirnov]. In the case of very small neutrino masses ($m_i^2/E_{\nu}^2\ll 1$), however, this is not useful in practice. We can write $K$, after expanding $F$ around $m_i^2/E_{\nu}^2=0$, as $$\begin{aligned}
K&=&\sum_i|U_{ei}|^2\left(F_0+\frac{m_i^2}{E_{\nu}^2}F'_0\right)+O\left(\frac{m_i^4}{E_{i}^4}\right), \\
&=&F_0+\frac{m_{\nu_e}^2}{E_{\nu}^2}F'_0+O\left(\frac{m_i^4}{E_{i}^4}\right),\end{aligned}$$ where $F_0=F$ evaluated at $m_i^2=0$, $F'_0=\partial F/\partial(m^2_i/E^2_{\nu})$ evaluated at $m_i^2=0$, and $$m_{\nu_e}^2\equiv\sum_i|U_{ei}|^2m_i^2
\label{eq:mb_def}$$ is defined as the electron neutrino effective mass-squared. Note that we made use of the fact that $\sum_i |U_{ei}|^2=1$. Henceforth, we will assume, given current bounds on $m_l$ and the neutrino oscillation parameters, that tritium beta-decay experiments are sensitive only to $m^2_{\nu_e}$.
Experimentally, $m_{\nu_e}^2$ is currently constrained to be less than $(2.0)^2$ eV$^2$ at the 99% confidence level [@pdg; @global_anal]. Given that this upper bound is much larger than $\Delta m^2_{12}$ and $|\Delta m^2_{13}|$, the current bound on $m_l^2$ is also much larger than the neutrino mass-squared differences (this is the so-called “quasi-degenerate neutrino masses” regime). Under these circumstances, $m_1^2\simeq m_2^2 \simeq m_3^2\simeq m_l^2$, independent of the mass hierarchy, and $m_{\nu_e}^2\simeq m_l^2\sum_i|U_{ei}|^2=m_l^2$. Hence, the bound $m_l\in[0,2.0]$ eV, quoted in Sec. \[sec:intro\]. Incidentally, the muon-type and tau-type neutrino effective masses-squared are given by $m^2_{\nu_{\alpha}}=\sum_im_i^2|U_{\alpha i}|^2$, $\alpha=\mu,\tau$. Given the current bound $m_l<2.0$ eV, both are constrained to be less than $(2.0)^2$ eV$^2$, orders of magnitude smaller than the current kinematical limits obtained from charged-current processes involving muons and taus [@pdg].
In the case $\sin^2\theta_{13}=0$, $m_{\nu_e}$ is simply given by $$\begin{aligned}
m_{\nu_e}^2&=&m_1^2\cos^2\theta_{12}+m_2^2\sin^2\theta_{12}, \\
&=&m_1^2+\Delta m^2_{12}\sin^2\theta_{12}, \\
&=&m_l^2+\Delta m^2_{12}\sin^2\theta_{12}~~~\rm (normal~hierarchy), \\
&=&m_l^2-\Delta m^2_{13}+\Delta m^2_{12}\sin^2\theta_{12}~~\rm (inverted~hierarchy).\end{aligned}$$ In the case of a normal mass hierarchy, the expressions above still hold for all values of $m_l$ and nonzero $|U_{e3}|$, as long as $|\Delta m^2_{13}|\sin^2\theta_{13}\ll\Delta m^2_{12}\sin^2\theta_{12}\to \sin^2\theta_{13}\ll 0.01$. In the case of an inverted hierarchy, the expressions above are valid if $\sin^2\theta_{13}\ll 1$, which is already experimentally guaranteed to be satisfied. Fig. \[fig\_ms\] depicts $m_{\nu_e}$ as a function of $m_l$ for an inverted and a normal mass hierarchy, and for $\Delta m^2_{12}=8.0\times 10^{-5}$ eV$^2$, $\Delta m^2_{13}=-2.44\times 10^{-3}$ eV$^2$, $\sin^2\theta_{12}=0.31$, and (as already mentioned) $\sin^2\theta_{13}=0$.
We are interested in extracting the neutrino mass hierarchy from experimental information on $m_{\nu_e}$, and address the question in the following way. We assume that all oscillation parameters are measured precisely enough that uncertainties on these are irrelevant (we worry about the precision with which oscillation parameters are measured in the next section). Furthermore, since we are working under the assumption that $|U_{e3}|=0$, we also assume that the neutrino mass hierarchy is not known, [*i.e.,*]{} there are two values of $\Delta m^2_{13}$ that perfectly fit all neutrino oscillation data. These will be referred to as $\Delta m^{2+}_{13}>0$ and $\Delta m^{2-}_{13}<0$. As discussed in detail in [@ours], it is quite likely that $\Delta m^{2+}_{13}-|\Delta m^{2-}_{13}|\equiv x\neq 0$. The value of $x$ will depend on the details of the measurement of $\Delta m^2_{13}$ [@ours].
Currently, the experimental upper bound on $m_{\nu_e}$ does not allow one to discriminate a normal from an inverted mass hierarchy. This remains true as long as $$m^2_{\nu_e}>-\Delta m^{2-}_{13}.
\label{eq:mb_bound}$$ Indeed, given our current knowledge of $\Delta m^2_{13}$, in order to determine the mass hierarchy from an upper bound (or measurement) of the electron-type effective neutrino mass, one should be able to constrain $m^2_{\nu_e}\lesssim 1.7\times 10^{-3}$ eV$^2$ (the current 99% upper bound on $\Delta m^{2-}_{13}$ [@global_anal]). This can only happen if the mass hierarchy is normal.
In other words, if the neutrino mass hierarchy is inverted, it is not possible to determine the character of the neutrino mass by measuring $m_{\nu_e}$, irrespective of how precise a measurement one is able to perform. Associated to the inverted mass hierarchy measurement of $m_l=m_l^-$, there is a different value of $m_l=m_l^+$ that provides an equally good fit to all neutrino data as long as $m_l^{2+}=m_l^{2-}-\Delta m^{2-}_{13}$ (we remind readers that $\Delta m^{2-}_{13}$ is negative-definite, such that $m_l^+>m_l^-$). If the mass hierarchy is normal and the true value of the lightest mass is $m_l^{2+}>-\Delta m^{2-}_{13}$, there is an equally good fit to all the neutrino data with an inverted hierarchy and a lightest mass $m_l^-$ given by $m_l^{2-}=m_l^{2+}+\Delta m^{2-}_{13}$. If $m_l^{2+}<-\Delta m^{2-}_{13}$, there is no inverted mass hierarchy “mirror solution.”
In the future, the KATRIN experiment, currently under construction, aims at being sensitive to values of $m_{\nu_e}>0.2$ eV at the 90% confidence level [@KATRIN]. Note that this still satisfies Eq. (\[eq:mb\_bound\]), such that KATRIN will not be able to establish the neutrino mass hiearchy, even if it observes a nonzero $m_{\nu_e}$ effect. We are not aware of experimental proposals to significantly improve the sensitivity to $m_{\nu_e}$ beyond the reach of KATRIN.
$m_{ee}$ — Neutrinoless Double Beta Decay
-----------------------------------------
If the neutrinos are Majorana fermions, lepton number is not a good quantum number.[^4] The most powerful probe of lepton number violation is the search for neutrinoless double beta decay, $0\nu\beta\beta$: $^AZ\to^A(Z+2)e^-e^-$. Assuming that the neutrino masses are the only relevant source of lepton number non-conservation, the rate for $0\nu\beta\beta$ is governed by the fundamental process $W^-W^-\to e^-e^-$, mediated by Majorana neutrino exchange.
Schematically, $$\Gamma_{0\nu\beta\beta}\propto \left|\sum_i U_{ei}^2 \frac{m_i}{Q^2+m_i^2}{\cal M}(m_i^2,Q^2)\right|^2,
\label{eq:mbb}$$ where $Q^2$ is a typical energy momentum transfer (squared) associated with the $0\nu\beta\beta$ process — $Q^2\sim (50)^2$ MeV$^2$ — and $\cal M$ contains the rest of the kinematics and the nuclear matrix element associated with the $Z\to (Z+2)$ transition. As is well known, the rate for $0\nu\beta\beta$ is proportional to a combination of the neutrino masses that vanishes in the limit $m_i\to 0$, $\forall i$. This is, of course, expected, since we are assuming that nonzero Majorana neutrino masses are the only source of lepton number violation.
All $m^2_i$ are much smaller than $Q^2$, so we can simplify Eq. (\[eq:mbb\]) by expanding the neutrino propagator $$\frac{m_i}{Q^2+m_i^2}=\frac{m_i}{Q^2}\left(1-\frac{m_i^2}{Q^2}+O\left(\frac{m_i^4}{Q^4}\right)\right),$$ and ${\cal M}(m_i^2,Q^2)={\cal M}(0,Q^2)(1+O(m_i^2/Q^2))$ such that, up to corrections proportional to $m^2_i/Q^2$, $$\Gamma_{0\nu\beta\beta}\propto |m_{ee}|^2,$$ where $$m_{ee}\equiv \sum_i U_{ei}^2m_i\equiv m_1|U_{e1}|^2e^{i\alpha_1}+m_2|U_{e2}|^2e^{i\alpha_2}+m_3|U_{e3}|^2e^{-2i\delta}
\label{eq:mbb_def}$$ is the effective mass for $0\nu\beta\beta$.[^5] We will assume henceforth that $0\nu\beta\beta$ decay experiments are capable of constraining (or measuring) $|m_{ee}|$.
In the case of interest here ($\sin^2\theta_{13}=0$), $$\begin{aligned}
|m_{ee}|&=&|m_1\cos^2\theta_{12}e^{i\alpha_1}+m_2\sin^2\theta_{12}e^{i\alpha_2}| \\
&=&\left(m_1^2\cos^4\theta_{12}+m_2^2\sin^4\theta_{12}+2m_1m_2\sin^2\theta_{12}\cos^2\theta_{12}
\cos\alpha\right)^{1/2}, \\
&=&\left(m_l^2\cos^4\theta_{12}+(m_l^2+\Delta m^2_{12})\sin^4\theta_{12}+m_l\sqrt{m_l^2+\Delta m^2_{12}}\frac{\sin^22\theta_{12}}{2}\cos\alpha\right)^{1/2}\rm (normal~hierarchy), \end{aligned}$$ and, $$\begin{aligned}
|m_{ee}|&=&\left((m_l^2-\Delta m^2_{13})\cos^4\theta_{12}+(m_l^2+\Delta m^2_{12}-\Delta m^2_{13})\sin^4\theta_{12} \right. \nonumber \\
&+&\left.\sqrt{(m^2_l-\Delta m^2_{13})(m_l^2+\Delta m^2_{12}-\Delta m^2_{13})}\frac{\sin^22\theta_{12}}{2}\cos\alpha\right)^{1/2}~~~\rm (inverted~hierarchy),\end{aligned}$$ where $\alpha\equiv \alpha_2-\alpha_1$ is the Majorana phase to which $m_{ee}$ is sensitive (in the limit $\sin^2\theta_{13}=0$). Given the current constraints on $\sin^2\theta_{13}$, the expressions above are an excellent approximation for $m_{ee}$ in the case of an inverted hierarchy, and also apply safely for a normal mass hierarchy as long as $\sqrt{|\Delta m^2_{13}|}\sin^2\theta_{13}\ll\sqrt{\Delta m^2_{12}}\sin^2\theta_{12}\to \sin^2\theta_{13}\ll 0.05$. Fig. \[fig\_ms\] depicts $|m_{ee}|$ as a function of $m_l$ for an inverted and a normal mass hierarchy, and for $\Delta m^2_{12}=8.0\times 10^{-5}$ eV$^2$, $\Delta m^2_{13}=-2.44\times 10^{-3}$ eV$^2$, $\sin^2\theta_{12}=0.31$, and $\sin^2\theta_{13}=0$. Associated to a fixed value of $m_l$ and a mass hierarchy there is a continuum of values of $|m_{ee}|$, one for each value of $\cos\alpha\in[-1,1]$.
The minimum (maximal) value of $m_{ee}$ for a fixed value of $m_l$ corresponds to $\cos\alpha=-1~(+1)$. $m_{ee}$ can vanish in the case of a normal mass hierarchy, when $m^2_l\cos^4\theta_{12}=(\Delta m^2_{12}+m_l^2)\sin^4\theta_{12}$. Hence, for $$m_l^2\sim \Delta m^2_{12}\frac{\sin^4\theta_{12}}{\cos^2\theta_{12}-\sin^2\theta_{12}}\sim (0.004)^2,$$ and an inverted mass hierarchy, $m_{ee}$ is especially suppressed.
As in the previous subsection, it is easy to show that if $$|m_{ee}|>\sqrt{\Delta m^{2-}_{13}}\cos^2\theta_{12}-\sqrt{\Delta m^{2-}_{13}+\Delta m^2_{12}}\sim 0.02~\rm eV,$$ a measurement of $|m_{ee}|$ has no discriminatory power when it comes to determining the nature of the neutrino mass hierarchy, irrespective of its precision. As before, such values can only be experimentally ruled out if the mass hierarchy is normal [@petcov_pascoli]. Hence, any “perfect” fit to all the neutrino data (including a potential measurement of $|m_{ee}|$) obtained under the assumption that the neutrino mass hierarchy is inverted can be matched by an as-perfect fit obtained under the assumption that the mass hierarchy is normal. A very recent and thorough analysis of the capabilities of a measurement of $|m_{ee}|$ to determine the mass hierarchy can be found in [@choubey_rodejohann].
In practice, the situation is significantly more involved, for three main reasons. One is that the extraction of $|m_{ee}|$ from the rate for $0\nu\beta\beta$ is severely clouded by uncertainties in computing the nuclear matrix elements. Different theoretical estimates that make use of different nuclear physics techniques can differ significantly, often by an order of magnitude [@nuclear_matrix]. We have nothing to add to the discussion, except for the fact that the situation should improve significantly in the next several years, especially if a positive signal for $0\nu\beta\beta$ is obtained. Another source of confusion is that $\Gamma_{0\nu\beta\beta}$ is only proportional to $|m_{ee}|$ if there are no other “beyond the standard model” sources of lepton number violation. We will come back to this issue briefly in the next section, and argue that the combined analysis of searches for $m_{\nu_e}$, $|m_{ee}|$ and $\Sigma$ (discussed in the next subsection) can shine light on this subject. Finally, the rate for $0\nu\beta\beta$ is only nonzero if lepton number conservation is not exact. As far as this discussion is concerned, this implies that while evidence for $0\nu\beta\beta$ can be translated into a nonzero $|m_{ee}|$, failed attempts to observe $0\nu\beta\beta$ cannot be translated into an upper bound on $|m_{ee}|$, and hence potentially vital information regarding the nature of the neutrino mass hierarchy. This should be contrasted with failed attempts to measure $m_{\nu_e}$ from tritium beta decay. These can be (much more model-independently) translated into information that may help resolve the nature of the neutrino mass hierarchy.
Currently, $m_{ee}$ is (conservatively) constrained to be $|m_{ee}|<0.91$ eV at the 99% confidence level (see [@global_anal] and references therein for details), and we will ignore the to-be-confirmed recent evidence for a nonzero rate for $0\nu\beta\beta$ [@klapdor]. As in the case for the current bound on $m_{\nu_e}$, no information regarding the neutrino mass hierarchy can be obtained from our current knowledge of $|m_{ee}|$. This would remain true even if we postulated that the neutrinos are Majorana fermions. Near-future experiments with sensitivity to $|m_{ee}|\gtrsim 0.1$ eV are currently being planned, and several of those aim at reaching an ultimate sensitivity to $|m_{ee}|\gtrsim 0.01$ eV [@0nubb_future]. It is fair to say that significant R&D efforts (plus time and resources) are necessary in order to improve the sensitivity to $|m_{ee}|$ beyond that.
$\Sigma$ — Cosmological Observables
-----------------------------------
There is a significant amount of indirect evidence that the Universe is filled with a very cold thermal population of relic neutrinos — the cosmic neutrino background. While we have no direct evidence that these really exist, their presence is required in order to match predictions of big bang nucleosynthesis with measurements of the primordial abundances of light nuclei [@BBN]. Indeed, before LEP experiments measured the invisible $Z$-boson width, bounds from big bang nucleosynthesis were known to provide the most stringent constraints on the number of light neutrino species [@pdg].
If neutrinos have mass, the cosmic neutrino background contributes to the energy budget of the Universe as hot dark matter. According to the very successful (if not very palatable) concordance cosmological model [@recent_cosmology], the fraction of hot dark matter in the Universe is constrained to be small. Given our current understanding of neutrino masses, $\Omega_{\rm hot}\simeq\sum m_i/(48~\rm eV)$, such that an upper bound on $\Omega_{\rm hot}$ can be translated directly into an upper bound on $\Sigma\equiv \sum_i m_i$. As was the case in previous subsections, cosmological observables are potentially sensitive to more than the sum of the neutrino masses, at least in principle. For example, hot dark matter has the property of preventing matter from “clumping” in the early Universe, leading to a suppression of power at small scales in the matter distribution power spectrum [@mc_bound_review]. Potential “kinks” in the spectrum are related to the mass of the individual hot dark matter species. We will ignore this possibility and will assume that cosmological observables are only sensitive to $\Sigma$.
$\Sigma$ depends only on the neutrino masses, and can be expressed in terms of $m_l$ and the mass-squared differences: $$\begin{aligned}
\Sigma&=&m_1+\sqrt{m_1^2+\Delta m^2_{12}}+\sqrt{m_1^2+\Delta m^2_{13}}, \label{eq:mc_def}\\
&=&m_l+\sqrt{m_l^2+\Delta m^2_{12}}+\sqrt{m_l^2+\Delta m^2_{13}}~~~(\rm normal~hierarchy), \\
&=&m_l+\sqrt{m_l^2-\Delta m^2_{13}}+\sqrt{m_l^2+\Delta m^2_{12}-\Delta m^2_{13}}~~~(\rm inverted~hierarchy).\end{aligned}$$ $\Sigma$ is the only observable discussed in this section that, in the limit $\theta_{13}\to 0$, is sensitive to $\Delta m^{2+}_{13}$, the value of the atmospheric mass hierarchy obtained assuming that the mass hierarchy is normal. Fig. \[fig\_ms\] depicts $\Sigma$ as a function of the lightest neutrino mass, $m_l$ for $\Delta m^2_{12}=8.0\times 10^{-5}$ eV$^2$, and $\Delta m^{2+}_{13}=2.50\times 10^{-3}$ eV$^2$, $\Delta m^{2-}_{13}=-2.44\times 10^{-3}$ eV$^2$, for both neutrino mass hierarchies.
As before, if $$\Sigma>\sqrt{-\Delta m^{2-}_{13}}+\sqrt{\Delta m^2_{12}-\Delta m^{2-}_{13}},
\label{mc_bound}$$ a measurement of $\Sigma$ is not sensitive to the nature of the neutrino mass hierarchy, independent of its precision. In light of the discussion in the two previous subsections, it should not be surprising that Eq. (\[mc\_bound\]) is automatically satisfied in the case of an inverted mass hierarchy, but could be violated in the case of a normal one, as long as $m_l$ is small enough ($m_l\lesssim 0.02$ eV). Quantitatively, for every candidate value of the lightest mass $m_l=m_l^-$ obtained by postulating that the mass hierarchy is inverted, there is another lightest mass value $m_l=m_l^+$ that fits the neutrino data just as well as long as one postulates that the mass hierarchy is normal and $m_l^+$ satisfies $$m_l^++\sqrt{m_l^{2+}+\Delta m^2_{12}}+\sqrt{m_l^{2+}+\Delta m^{2+}_{13}}=m^-_l+\sqrt{m_l^{2-}-\Delta m^{2-}_{13}}+\sqrt{m_l^{2-}+\Delta m^{2}_{12}-\Delta m^{2-}_{13}}.$$
As with measurements of $|m_{ee}|$, the map between cosmological measurements and $\Sigma$ is nontrivial and model dependent. The cosmological determination of $\Sigma$ depends on the history of the Universe and its particle content, and physics beyond the standard model and/or concordance cosmology can easily obscure, enhance or completely erase any relationship between cosmological observables and $\Sigma$ [@mc_bound_review; @avoid_mc].
According to [@global_anal], combined analysis of data from WMAP, SDSS and Lyman-alpha forest surveys constrain $\Sigma<0.94$ eV at the 99% confidence level. Different analyses of the same data or subsets thereof quote similar bounds (within 50%) [@mc_bound_review]. Upcoming missions have the potential to significantly improve on the current sensitivity to a nonzero value of $\Sigma$. It is expected that future data on the cosmic microwave background should be sensitive to $\Sigma\gtrsim 0.1$ eV [@sigma_shift], while current studies of the capabilities of weak lensing probes seem to indicate that these might be sensitive to $\Sigma$ values below the 0.1 eV level [@lensing]. For an overview see, for example, [@sigma_shift]. It has recently been speculated that one may indeed reach $\Sigma$ values as low as 0.03 eV [@wang_etal].
Combined Fits to the Mass Hierarchy {#sec:fits}
===================================
It has been widely recognized in the literature [@many; @petcov_wolfenstein; @lisi_silk; @new_petcov] that the combined analysis of searches for $m_{\nu_e}$, $m_{ee}$,[^6] and/or $\Sigma$ provide qualitatively more information than the separate analysis of each observable, including potential information regarding the mass hierarchy [@petcov_wolfenstein]. For example, a lower bound on $m_{\nu_e}$ combined with an upper bound for $m_{ee}$ may help determine, with the addition of some reasonable assumptions, that the neutrinos are Dirac fermions, while a large $m_{\nu_e}$ value, combined with an upper bound on $\Sigma$ would lead one to conclude that there is more to our understanding of particle physics or the history of the Universe than currently accepted. Here, we discuss the circumstances under which measurements of (or upper bounds for) $m_{\nu_e}$, $m_{ee}$, and/or $\Sigma$ help us determine the mass hierarchy.
It is easy to understand why a combined analysis should prove more powerful. The value of $m_l^+$ that renders a normal hierarchy fit to, say $\Sigma$, compatible with an inverted hierarchy one (associated to a value $m_l^-$ of the lightest mass) need not agree with the the equivalent quantity obtained from analyzing measurements of $m_{ee}$ or $m_{\nu_e}$. For example, if the true value of $m_l$ vanishes and the mass hierarchy is inverted, the mirror value $m_l=m_l^+$ that renders $\Sigma$ “the same” if one assumes that the mass hierarchy is normal is $m_l^+(\Sigma)\sim 0.5\sqrt{\Delta m^{2-}_{13}}\sim 0.02$ eV. On the other hand, the equivalent for $m_{\nu_e}$ is $m_l^+(m_{\nu_e})=\sqrt{\Delta m^{2-}_{13}}\sim0.05$ eV. Hence, for the wrong-hierarchy hypothesis, $m_l(m_{\nu_e})\neq m_l(\Sigma)\neq m_l(m_{ee})$ — even if the true neutrino mass hierarchy is inverted. This indicates that combined analyses of the three observables discussed in Sec. \[sec:probes\] can improve, sometimes significantly, our ability to determine the neutrino mass ordering.
Figure \[figure:2by2\] depicts the effective masses pairwise for both hierarchies, assuming $\Delta m^2_{12}=8.0\times 10^{-5}$ eV$^2$, $\Delta m^{2+}_{13}=2.50\times 10^{-3}$ eV$^2$, $\Delta m^{2-}_{13}=-2.44\times 10^{-3}$ eV$^2$, $\sin^2\theta_{12}=0.31$ and $\sin^2\theta_{13}=0$, allowing $m_l$ to vary between 0 and 0.5 eV, and the relevant Majorana phase to span its entire allowed physical range: $\cos\alpha\in[-1,1]$. One can easily identify sets of potential measurements that, with sufficiently small uncertainties, would imply a specific mass hierarchy.[^7] In particular, it is easy to verify, and improve on, the claims made in section \[sec:probes\] regarding the implication of the normal scheme for sufficiently small measurements of the effective masses. Evidence for the inverted hierarchy can also arise from such naive considerations. For instance, by examining the upper left panel it is clear that the inverted hierarchy could be established if accurate measurements of $m_{ee}$ and $\Sigma$ were made near 0.04 eV and 0.1 eV, respectively. The same conclusion is obtained if only an upper bound exists for $\Sigma$ near $0.2$ eV, while a measurement sets $m_{ee}$ near $0.05$ eV. It is easy to see that combined measurements of $m_{\nu_e}$ and $m_{ee}$ cannot distinguish an inverted mass hierarchy from a normal, quasi-degenerate scheme — knowledge of $\Sigma$ is extremely valuable when it comes to establishing an inverted mass hierarchy.
![Values of $m_{\nu_e}$, $|m_{ee}|$, and $\Sigma$ for both hierarchies, for $\Delta m^2_{12}=8.0\times 10^{-5}$ eV$^2$, $\Delta m^{2+}_{13}=2.50\times 10^{-3}$ eV$^2$, $\Delta m^{2-}_{13}=-2.44\times 10^{-3}$ eV$^2$, $\sin^2\theta_{12}=0.31$ and $\sin^2\theta_{13}=0$. The value of $m_l$ varies between 0 and 0.5 eV and the relevant Majorana phase spans its entire allowed physical range: $\cos\alpha\in[-1,1]$. See text for details.[]{data-label="figure:2by2"}](two_by_two.eps)
This simple picture is smeared out when realistic conditions are taken into account. The inclusion of parameter uncertainties will naturally wash out some of the interesting regions in this measurement space, but most of the main features remain robust. A complete statistical analysis, quantifying the very visual and intuitive line of reasoning outlined here, is presented below.
Parameters and Data Analysis
----------------------------
In order to quantitatively compare the different neutrino mass hierarchies we proceed as follows. We assume that all oscillation parameters $\Theta_i=\Delta m^2_{13},\Delta m^2_{12},\sin^2\theta_{12},\sin^2\theta_{13}$ are measured to be $\overline{\Theta_i}\pm\sigma_{\Theta_i}$, while the three non-oscillation parameters are constrained to be $m_{\nu_e}=\overline{m_{\nu_e}}\pm\sigma_{\nu_e}$, $m_{ee}=\overline{m_{ee}}\pm\sigma_{ee}$, and $\Sigma=\overline{\Sigma}\pm\sigma_{\Sigma}$. Note that we are assuming that the uncertainty related to the computation of the $0\nu\beta\beta$ nuclear matrix element is included in $\sigma_{ee}$. We assume no information concerning the three CP-odd phases $\alpha_1,\alpha_2,\delta$. We refer to all these measurements as the data. We further postulate that the neutrinos are Majorana fermions, and that there are no other sources of lepton number violation. Finally, we assume concordance cosmology and “standard” particle physics.
Armed with these results, we construct a $\chi^2$ function defined by $$\chi^2(m_l,\alpha_1,\alpha_2,\delta,\Theta) = \frac{[\overline{m_{\nu_e}} -
m_{\nu_e}(m_l,\Theta)]^2}{\sigma_{\nu_e}^2}
+\frac{[\overline{m_{ee}} -
m_{ee}(m_l,\Theta,\alpha_1,\alpha_2,\delta)]^2}{\sigma_{ee}^2} +
\frac{[\overline{\Sigma} - \Sigma(m_l,\Theta)]^2}{\sigma_{\Sigma}^2}
+ \sum_i{\frac{[\overline{\Theta_i} -
\Theta_i]^2}{\sigma_{\Theta_i}^2}}.$$ This function also depends implicitly on the choice of hierarchy from the structure of the effective masses $m_{\nu_e}$, $m_{ee}$, and $\Sigma$ given, respectively, by Eqs. (\[eq:mb\_def\],\[eq:mbb\_def\],\[eq:mc\_def\]).
For each hypothesis concerning the mass hierarchy, we minimize $\chi^2$ with respect to all neutrino parameters ($\Theta_i$, $m_l$, and the two relevant CP-odd phases \[say, $\alpha_2-\alpha_1$, and $-2\delta-\alpha_1$\]), and establish if a good fit to the data can be obtained. This is done by comparing the minimum value of $\chi^2=\chi^2_{\rm min}$ with expectations for three degrees of freedom. In particular, $\chi^2_{\rm min} < 7.8$ implies that the theory fits the data at the 95% confidence level. If only one of the hypotheses concerning the mass hierarchy fits the data at the 95% confidence level (or better), we conclude that it is established (at least) at the 95%, while if both hypotheses prove to fit the data at the 95% confidence, we conclude that the mass-hierarchy cannot be determined. Finally, if neither hypothesis fits the data, we are forced to conclude that there is a flaw in our theoretical model. Candidate flaws include the hypothesis that the neutrino is a Majorana fermion, that concordance cosmology is correct, etc.
The different data sets used in our numerical analysis are listed in Table \[table:parameters\]. The best fit values [@global_anal] from the global analysis of neutrino oscillations were used for $\overline{\Theta}$, except for $\overline{\Delta m^2_{13}}$. In the case of a normal mass hierarchy, we pick $\Delta m^{2+}_{13}=2.50\times 10^{-3}$ eV$^2$, while in the case of a normal mass hierarchy, we choose $\Delta m^{2-}_{13}=-\Delta m^{2+}_{13}+x$, where $x=2\Delta m^2_{12}\cos2\theta_{12}$, in order to take into account that fact that, in the case of very precise measurements of $\Delta m^2_{13}$, different hypothesis regarding the mass hierarchy yield different values for $|\Delta m^2_{13}|$ [@ours]. The value of $x$ above agrees with what one expects to obtain in accelerator $\nu_{\mu}\to\nu_{\mu}$ studies at relatively short baselines ($L\lesssim 3000$ km) or relatively large energies ($E_{\nu}\gtrsim 1$ GeV). We refer readers to [@ours] for details.
-------------------------------------------------- --------- ---------- --------- ---------- --------- ---------- ----------------------------------
Run 1 Run 2 Run 3
Parameter Central $\sigma$ Central $\sigma$ Central $\sigma$ comments
\[0.5ex\] $\Delta m^2_{12} (10^{-5}~{\rm eV}^2)$ 8.0 5% 8.0 5% 8.0 5% Reactor [@solar]
$\Delta m^{2+}_{13} (10^{-3}~{\rm eV}^2)$ 2.5 1% 2.5 1% 2.5 1% T2K, NO$\nu$A, MINOS [@Lindner]
$\Delta m^{2-}_{13} (10^{-3}~{\rm eV}^2)$ –2.44 1% –2.44 1% 2.5 1% T2K, NO$\nu$A, MINOS [@Lindner]
$\sin^2\theta_{12}$ 0.31 5% 0.31 5% 0.31 5% Reactor [@theta12][@solar]
$\sin^2\theta_{13}$ 0 0.01 0 0.01 0 0.01 T2K, NO$\nu$A,D-CHOOZ [@Lindner]
$m_{\nu_e} ~{\rm(eV)}$ 0 0.1 0 0.1 0 0.1 KATRIN [@KATRIN]
$m_{ee}~{\rm (eV)}$ y-axis 0.01 y-axis 0.05 y-axis 0.1
$\Sigma~{\rm (eV)}$ x-axis 0.01 x-axis 0.05 x-axis 0.1 [@mc_bound_review]
\[1ex\] \[table:parameters\]
-------------------------------------------------- --------- ---------- --------- ---------- --------- ---------- ----------------------------------
: Input values, including uncertaintites, for the simulated data sets. ‘x-axis’ and ‘y-axis’ refer to Fig. \[figure:run\]. The ‘comments’ column refers to the experiment (or class of experiments) expected to contribute most significantly to the uncertainty with which each neutrino observable is assumed to be measured.
In order to anticipate future developments, we assume that the uncertainties on the various oscillation parameters are smaller than the current ones. For instance, the uncertainty on the atmospheric mass-squared difference $|\Delta m^2_{13}|$ should be reduced to $\sim 3\%$ in the future, due to potential results from off-axis beam experiments, such as T2K and NO$\nu$A [@Lindner], and future endeavors can improve on that [@BNL]. Similarly, we expect that the the solar parameters, $\Delta m^2_{12}$ and $\sin^2\theta_{12}$ can be measured at the 5% level in, say, precise measurements of the low energy solar neutrino flux [@low_solar], improved reactor experiments boasting greater statistics and reduced systematic errors conducted at baselines at (or near) the first survival probability minimum [@theta12; @solar], or with broadband long-baseline $\nu_{\mu}\to\nu_e$ studies [@BNL].
Results
-------
Figure \[figure:run\] depicts regions of the $\overline{\Sigma}\times\overline{m_{ee}}$ plane consistent with both a normal and an inverted neutrino mass hierarchy (green \[gray\]), consistent only with a normal mass hierarchy (yellow \[lightest gray\]), consistent only with an inverted mass hierarchy (blue \[dark\]), or inconsistent with either hypothesis concerning the mass hierarchy (white). The uncertainties with which $\Sigma$ and $m_{ee}$ are expected to be measured, $\sigma_{m_{ee}}$ and $\sigma_{\Sigma}$, are held fixed for all values of $\overline{\Sigma}$, $\overline{m_{ee}}$, and are both equal to 0.01 eV in “Run 1” (left panel), 0.05 eV in “Run 2” (middle panel) and 0.1 eV in “Run 3” (right panel). Throughout, we assume that $m_{\nu_e}$ is experimentally constrained to be less than 0.1 eV (at the one sigma level).[^8] In the absence of new sources of lepton number violation, this upper bound on $m_{\nu_e}$ implies $|m_{ee}|\lesssim 0.1$ eV and, in the absence of nonstandard cosmology, $\Sigma\lesssim 0.3$ eV. We concentrate our exploration of the the parameter space to this region. If, on the other hand, the KATRIN experiment obtained statistically significant evidence that $m_{\nu_e}$ is nonzero (say, $m_{\nu_e}=0.3\pm 0.1$ eV) we would be constrained to the quasi-degenerate neutrino mass scenario, in which it is virtually impossible to determine the mass hierarchy by precisely measuring $m_{\nu_e}$, $m_{ee}$, and $\Sigma$ (see Fig. \[figure:2by2\]).
![Regions of the $\overline{\Sigma}\times\overline{m_{ee}}$ plane consistent with both a normal and an inverted neutrino mass hierarchy (green \[gray\]), consistent only with a normal mass hierarchy (yellow \[lightest gray\]), consistent only with an inverted mass hierarchy (blue \[dark\]), or inconsistent with either hypothesis concerning the mass hierarchy (white). Input parameters and expected uncertainties are tabulated in Table \[table:parameters\]. LEFT: $\sigma_{m_{ee}}=\sigma_{\Sigma}=0.01$ eV, MIDDLE: $\sigma_{m_{ee}}=\sigma_{\Sigma}=0.05$ eV, and RIGHT: $\sigma_{m_{ee}}=\sigma_{\Sigma}=0.1$ eV. See text for details.[]{data-label="figure:run"}](allRuns2.eps)
If both $\Sigma$ and $m_{ee}$ are only poorly measured ($\sigma_{m_{ee}}=\sigma_{\Sigma}=0.1$ eV, Fig. \[figure:run\](RIGHT)) it is, in general, not possible to determine the mass hierarchy, irrespective of $\overline{\Sigma}$ and $\overline{m_{ee}}$. The only exception corresponds to very large values of $\overline{m_{ee}}$, where, given an upper bound on $\Sigma$, only an inverted mass hierarchy provides a good fit. Very large values of $\overline{\Sigma}$ and $\overline{m_{ee}}$ are inconsistent with the upper bound on $m_{\nu_e}$, and, if observed, need to be interpreted as evidence for new cosmology/particle physics and sources of lepton number nonconservation other than the neutrino masses. If $\Sigma$ and $m_{ee}$ are measured somewhat more precisely ($\sigma_{m_{ee}}=\sigma_{\Sigma}=0.05$ eV, Fig. \[figure:run\](MIDDLE)), the situation is qualitatively the same: one will not be able to determine the mass hierarchy from non-oscillation probes of neutrino masses unless $m_{ee}$ is measured to be large ($\overline{m_{ee}}\gtrsim 0.1$ eV) and $\Sigma$ is “small” ($\overline{\Sigma}\lesssim 0.1$ eV), in which case only the inverted mass hierarchy provides a good fit to the data.
Only when both $\Sigma$ and $m_{ee}$ are measured very precisely ($\sigma_{m_{ee}}=\sigma_{\Sigma}=0.01$ eV, Fig. \[figure:run\](LEFT)) do we start to also positively discriminate a normal mass hierarchy. If $\overline{\Sigma}\lesssim 0.07$ eV, only the normal mass hierarchy hypothesis is capable of fitting the data, and only if $\overline{m}_{ee}\lesssim 0.04$ eV. As before, large $\overline{m_{ee}}$ values, appropriately correlated with large $\overline{\Sigma}$ values, point to an inverted mass hierarchy. Furthermore, when the sensitivity of both probes reaches the 0.01 eV level, we fail to fit the data if $\overline{m}_{ee}\lesssim 0.03$ eV and $\overline{\Sigma}\gtrsim0.3$ eV. If faced with such a scenario, we would be tempted to conclude that the neutrinos are Dirac fermions. Finally, note that, if $\sigma_{\Sigma}\simeq 0.01$ eV, we are guaranteed, in the absence of new cosmology/particle phisics, to obtain a nonzero value for $\overline{\Sigma}$ — hence the region in Fig. \[figure:run\](RIGHT) corresponding to $\overline{\Sigma}\lesssim 0.02$ eV, for all $\overline{m_{ee}}$, is white.
As alluded to above, if $m_{ee}\gtrsim 0.1$ eV [*and*]{} $\Sigma\gtrsim 0.3$, we enter the quasi-degenerate neutrino mass regime, where discrimination between the two mass orderings is virtually unachievable. This behavior can be clearly observed in all panels of Fig. \[figure:run\].
Concluding Remarks {#sec:end}
==================
While we have discovered that neutrinos have mass, there are still qualitative aspects of the neutrino mass spectrum that remain unknown. We have been able to measure, with good precision, the (absolute value of the) two mass-squared differences, but know very little about the magnitude of the individual neutrino masses. We are also ignorant when it comes to the neutrino mass hierarchy. We don’t know if the neutrino masses are “normal ordered” — $m_1^2<m_2^2<m_3^2$ — or whether the mass hierarchy is “inverted” — $m_3^2<m_1^2<m_2^2$ (such that $\Delta m^2_{12}\ll m_2^2,m_1^2$). The reason we have failed (so far) to discover the neutrino mass hierarchy via neutrino oscillations is that both $|U_{e3}|^2$ and $\Delta m^2_{12}/|\Delta m^2_{13}|$ happen to be small.
Next-generation, long-baseline neutrino oscillation experiments are poised to uncover the ordering of the neutrino masses, but they can only be successful if $|U_{e3}|^2$ is large enough. If, on the other hand, $|U_{e3}|^2$ is too small, it is quite challenging to determine the neutrino mass hierarchy via neutrino oscillations [@ours]. Other probes are necessary. Here we have studied in detail whether combined information from future studies of neutrinoless double beta decay ($m_{ee}$), tritium beta-decay ($m_{\nu_e}$) and cosmological probes of the energy composition of the Universe ($\Sigma$) can help establish the neutrino mass hierarchy.
Our results are summarized in Fig. \[figure:run\] and, as expected, the situation is quite challenging. Assuming we can rule out (at the one sigma level) $m_{\nu_e}>0.1$ eV and probe $m_{ee}$ and $\Sigma$ with $\sigma_{m_{ee}}\sim\sigma_{\Sigma}\sim 0.05$ eV, we can only hope to determine the neutrino mass hierarchy from nonoscillation experiments if it is inverted [*and*]{} if the Majorana phase to which $m_{ee}$ is sensitive is close to $0$ (“constructive interference,” $\cos\alpha\sim 1$).[^9] If $\sigma_{m_{ee}}\sim\sigma_{\Sigma}\sim 0.01$ eV, the situation is significantly improved, and there is also the possibility of establishing a normal mass hierarchy if $\Sigma\lesssim 0.07$ eV and $m_{ee}\lesssim 0.04$ eV. In the next several years ($\lesssim 10$), it is reasonable to expect that we will be somewhere between Fig. \[figure:run\](RIGHT) and Fig. \[figure:run\](MIDDLE), while it is not overly optimistic to assume we will eventually reach somewhere between Fig. \[figure:run\](MIDDLE) and Fig. \[figure:run\](LEFT).
It is curious to note that the “main goal” of non-oscillation neutrino experiments is, arguably, [*not*]{} to determine the mass hierarchy. These experiments are usually associated with establishing if the neutrinos are Majorana fermions, measuring the lightest neutrino mass $m_l$, and, ultimately, probing whether the Majorana phases are nontrivial ($\alpha_1,\alpha_2\neq 0,\pi$). While we have not concentrated on these issues here, it is clear that in the green \[gray\] regions of Fig. \[figure:run\], one is not able to properly measure $m_l$, even when $m_{ee}$ and $\Sigma$ are measured to be nonzero with good precision. In those regions, for every normal-hierarchy measurement of $m_l=m_l^+\pm\delta^+_{m_l}$, there is an inverted hierarchy $m_l=m_l^-\pm\delta^-_{m_l}$ that fits the data just as well.
We reemphasize that, in order to convert measurements of $\Gamma_{0\nu\beta\beta}$ and cosmological data into information regarding neutrino masses, we are required to rely on several untested hypothesis, including the fact that Majorana active neutrino masses are the only source of lepton number violation and the fact that we quantitatively understand the history of the Universe from times slightly before big-bang nucleosynthesis until the formation of the observed large-scale structure. It is reassuring that combined measurements of $m_{\nu_e}$, $m_{ee}$, and $\Sigma$ can help falsify these hypotheses in some regions of the parameter space, but there is a finite possibility that, after many years of data analysis, we will be left in a rather “confused” state. This would happen, for example, if the mass hierarchy were normal, $m_{l}\lesssim 0.01$ eV and $\sigma_{\Sigma}\gtrsim 0.03$ eV.
We conclude by pointing out that, as far as establishing the neutrino mass ordering is concerned, we would profit tremendously with more precise information on $m_{\nu_e}$. A measurement of (or upper bound for) $m_{\nu_e}$ at the several $\times 10^{-2}$ eV level, combined with $\sigma_{\Sigma}$ slightly below the 0.1 eV level would be sensitive to the mass hierarchy (see Fig. \[figure:2by2\]), and it is easy to appreciate that extracted values of $m_{\nu_e}$ are significantly less model dependent than those of $m_{ee}$ and $\Sigma$.
[**Note Added:**]{} While this work was being completed, [@new_petcov] was posted in the preprint ArXiv’s. It shares several of the results presented here (with a slightly different treatment of the “data”), and contains a very detailed discussion of the capabilities of $\Sigma$ and $m_{ee}$ (including a detailed treatment of the uncertainties related to nuclear matrix elements) to establish the neutrino mass hierarchy and the existence of nontrivial Majorana phases, and to measure $m_l$. Our contribution, on the other hand, contains a more detailed analysis (including a detailed qualitative discussion) of the capabilities of combined $m_{\nu_e}$, $m_{ee}$, and $\Sigma$ measurements to uncover the neutrino mass ordering, motivated by the fact that it will remain unknown if $|U_{e3}|$ is vanishingly small, and also discusses the importance of measuring (or further constraining) $m_{\nu_e}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are happy to thank Boris Kayser for enlightening conversations and comments on the manuscript, Thomas Schwetz for comments on the manuscript, and Silvia Pascoli for words of encouragement. This work is sponsored in part by the US Department of Energy Contract DE-FG02-91ER40684.
[99]{}
For a recent pedagogical discussion and many references, see A. de Gouvêa, hep-ph/0411274. A. Aguilar [*et al.*]{} \[LSND Collaboration\], Phys. Rev. D [**64**]{}, 112007 (2001). For recent general overviews of theoretical developments in neutrino physics see, for example, R.N. Mohapatra [*et al.*]{}, hep-ph/0412099; A. de Gouvêa, Mod. Phys. Lett. A [**19**]{} (2004) 2799. S. Eidelman [*et al.*]{} \[Particle Data Group Collaboration\], Phys. Lett. B [**592**]{}, 1 (2004). for an up-to-date analysis of all neutrino data, see A. Strumia and F. Vissani, hep-ph/0503246. See also G.L. Fogli, E. Lisi, A. Marrone and A. Palazzo, hep-ph/0506083.
A. de Gouvêa, J. Jenkins and B. Kayser, Phys. Rev. D [**71**]{}, 113009 (2005). For recent discussions, see, [*e.g.*]{}, Y. Farzan, O.L.G. Peres and A.Yu. Smirnov, Nucl. Phys. B [**612**]{}, 59 (2001) Y. Farzan and A.Yu. Smirnov, Phys. Lett. B [**557**]{}, 224 (2003); S.S. Masood, S. Nasri and J. Schechter, hep-ph/0505183. L. Bornschein \[KATRIN Collaboration\], Nucl. Phys. A [**752**]{}, 14 (2005). This point had, of course, been appreciated earlier. See, for example, S. Pascoli and S.T. Petcov, Phys. Lett. B [**544**]{}, 239 (2002); Phys. Lett. B [**580**]{}, 280 (2004).
S. Choubey and W. Rodejohann, hep-ph/0506102.
For an overview, see J. Suhonen, Nucl. Phys. A [**752**]{}, 53 (2005). For a recent, detailed, and critical analysis see V.A. Rodin, A. Faessler, F. Simkovic and P. Vogel, nucl-th/0503063. H.V. Klapdor-Kleingrothaus, I.V. Krivosheina, A. Dietz and O. Chkvorets, Phys. Lett. B [**586**]{}, 198 (2004) and references therein. C. Aalseth [*et al.*]{}, hep-ph/0412300. S. R. Elliott and P. Vogel, Ann. Rev. Nucl. Part. Sci. [**52**]{}, 115 (2002). For an overview of proposed experiments, see F.T. Avignone, Nucl. Phys. Proc. Suppl. [**143**]{} 233 (2005). For a recent review, see G. Steigman, astro-ph/0501591, and references therein. For recent pedagogical overviews, see, for example, M. Trodden and S.M. Carroll, astro-ph/0401547; K.A. Olive, astro-ph/0503065, and many references therein.
For recent overviews see S. Hannestad, New J. Phys. [**6**]{}, 108 (2004); Nucl. Phys. Proc. Suppl. [**145**]{}, 313 (2005); M. Tegmark, hep-ph/0503257; S. Pastor, hep-ph/0505148, and references therein.
S. Hannestad, astro-ph/0505551. J. Lesgourgues, S. Pastor and L. Perotto, Phys. Rev. D [**70**]{}, 045016 (2004). K. N. Abazajian and S. Dodelson, Phys. Rev. Lett. [**91**]{}, 041301 (2003); M. Kaplinghat, L. Knox and Y. S. Song, Phys. Rev. Lett. [**91**]{}, 241301 (2003). S. Wang, Z. Haiman, W. Hu, J. Khoury and M. May, astro-ph/0505390. For an incomplete list of references, see V. D. Barger and K. Whisnant, Phys. Lett. B [**456**]{}, 194 (1999); M. Czakon, J. Gluza, J. Studnik and M. Zralek, Phys. Rev. D [**65**]{}, 053008 (2002); S. Pascoli, S.T. Petcov and W. Rodejohann, Phys. Lett. B [**558**]{}, 141 (2003); H. Minakata and H. Sugiyama, Phys. Lett. B [**567**]{}, 305 (2003); J.N. Bahcall, H. Murayama and C. Peña-Garay, Phys. Rev. D [**70**]{}, 033012 (2004); C. Giunti, Nucl. Phys. Proc. Suppl. [**145**]{}, 231 (2005), and many references therein. An extensive list of references can be found in [@new_petcov].
See, for example, S. Pascoli, S.T. Petcov and L. Wolfenstein, Phys. Lett. B [**524**]{}, 319 (2002); S. Pascoli and S.T. Petcov, Phys. Atom. Nucl. [**66**]{}, 444 (2003) \[Yad. Fiz. [**66**]{}, 472 (2003)\]. G. L. Fogli [*et al*]{}, Phys. Rev. D [**70**]{}, 113003 (2004).
S. Pascoli, S.T. Petcov and T. Schwetz, hep-ph/0505226.
M. Lindner, hep-ph/0503101. M.V. Diwan [*et al.*]{}, Phys. Rev. D [**68**]{}, 012002 (2003). J.N. Bahcall and C. Peña-Garay, JHEP [**0311**]{}, 004 (2003).
H. Minakata, H. Nunokawa, W.J.C. Teves and R. Zukanovich Funchal, Phys. Rev. D [**71**]{}, 013005 (2005); A. Bandyopadhyay, S. Choubey, S. Goswami and S. T. Petcov, hep-ph/0410283. S. Choubey, hep-ph/0402288.
[^1]: This is the most conservative bound on $m_l$, provided by precision measurements of the end point spectrum of tritium beta-decay [@pdg], and ignores stronger constraints from searches for neutrinoless double-beta decay (which apply only if the neutrinos are Majorana fermions) and from cosmology (which rely on several nontrivial assumptions regarding the evolution of the Universe and its particle content). We will discuss these in more detail later.
[^2]: Hence the need for very long baselines — $L_{\rm osc}\propto E_{\nu}/\Delta m^2$ [@TASI]. It is always amusing to remember that solar neutrino experiments are sensitive to $\Delta m^2\gtrsim 10^{-11}$ eV$^2$.
[^3]: The neutrino energy is trivially related to the energy of the $\beta$-ray, which is experimentally accessible.
[^4]: To be more precise, $U(1)_{B-L}$ is no longer a global symmetry of the Lagrangian that describes standard model degrees of freedom.
[^5]: After electroweak symmetry breaking and in the weak basis where the charged current couplings and the charged lepton mass matrix are diagonal, $m_{ee}$ is the “$\nu_e\nu_e$” entry of the neutrino Majorana mass matrix.
[^6]: Henceforth, to simplify the notation, we define $m_{ee}\equiv|m_{ee}|$.
[^7]: The usefulness of looking at the non-oscillation observables as depicted in Fig. \[figure:2by2\] was pointed out in [@lisi_silk].
[^8]: This is in rough agreement with the sensitivity of the KATRIN experiment, which can exclude $m_{\nu_e}<0.2$ eV at the 90% confidence level if it does not observe any evidence for nonzero neutrino masses [@KATRIN].
[^9]: We remind readers that we are always assuming the value of $|U_{e3}|$ to be vanishingly small.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider networks of processes which interact with beeps. Various beeping models are used. The basic one, defined by Cornejo and Kuhn [@Cornejo10], assumes that a process can choose either to beep or to listen; if it listens it can distinguish between silence or the presence of at least one beep. The aim of this paper is the study of the resolution of paradigms such as collision detection, computation of the degree of a vertex, colouring, or $2$-hop-colouring in the framework of beeping models. For each of these problems we present Las Vegas or Monte Carlo algorithms and we analyse their complexities expressed in terms of the number of slots.
We present also efficient randomised emulations of more powerful beeping models on the basic one. We illustrate emulation procedures with an efficient degree computation algorithm in the basic beeping model; this algorithm was given initially in a more powerful model.
author:
- 'Y. Métivier, J.M. Robson and A. Zemmari'
bibliography:
- 'biblio.bib'
date:
title: On Distributed Computing with Beeps
---
[**keywords:**]{} Beeping model, Collision detection, Colouring, $2$-hop-colouring, Degree computation, Emulation.
Introduction
============
The problem
-----------
Distributed graph algorithms are studied according to standard criteria that are usually formulated: topological restriction (trees, rings, or triangulated networks ...), topological knowledge (size, diameter ...), and local knowledge to distinguish nodes (identities, port numbers). Another important parameter of these algorithms is the message size: no limit (local model), $O(\log n)$ (congest model, where $n$ is the size of the graph) or $O(1)$. For each of these criteria or parameters, we study in particular the number of steps (rounds) necessary to obtain the result. According to the hypotheses, solutions are deterministic or randomised.
Typically, if we consider the MIS[^1] problem, when no identifiers are available there are only randomised solutions. Since the major contribution due to Luby [@Luby], this problem has been extensively studied with parameters given above. More recently, Afek et al. [@Afek13], inspired by biological observations, study the MIS problem through the beeping model: at each step a vertex can either beep (emit a signal) or be silent, and if it is silent it can distinguish between silence or the presence of at least one beep in its neighbouring. This approach has been developed in several papers [@Cornejo10; @Schneider10; @Afek13; @HuangM13; @Scott13] for distributed problems such as MIS computation, (interval) colouring, conflict resolution, membership problem etc.
Let $G$ be a graph and let $v$ be a vertex of $G$; two kinds of collisions may happen from the point of view of $v$:
- $v$ beeps and simultaneously at least one neighbour of $v$ beeps, this collision is called an internal collision;
- at least two distinct neighbours of $v$ beep simultaneously, this collision is called a peripheral collision.
In this paper, we consider several variants of beeping models:
- if a process beeps, there are two cases:
1. it cannot know whether another process beeps simultaneously (see [@Cornejo10]), this case is denoted by $B$;
2. it can distinguish whether it beeped alone or if at least one neighbour beeped concurrently, it is an internal collision; this case is called sender side collision detection in [@Afek13] Section 6, and it is denoted in this paper $B_{cd}$;
- if a process listens, there are also two cases:
1. it can distinguish between silence or the presence of at least one beep (see [@Cornejo10]), this model is denoted $L$;
2. it can distinguish between silence or the presence of one beep or the presence of at least two beeps; in this case it is a peripheral collision, (see [@Schneider10],[@Afek13] Section 4), this model is denoted $L_{cd}$ in this paper.
Finally, a beeping model is defined by choosing between $B$ or $B_{cd}$ and between $L$ and $L_{cd}$. For example the basic beeping model introduced by Cornejo and Kuhn in [@Cornejo10] is $BL$; Afek et al. in [@Afek13] (Section 6) and Scott et al. in [@Scott13] study the MIS problem in the model $B_{cd}L$. In Section 4 of [@Afek13], Afek et al. study the MIS problem in $BL_{cd}$. In this paper we present algorithms in models $BL$, $B_{cd}L$ and $B_{cd}L_{cd}$.
Usually, the topology of a distributed system is modelled by a graph and paradigms of distributed systems are represented by classical problems in graph theory such as vertex degree, maximal independent set (MIS for short), $2$-MIS (we recall that a $2$-MIS of a graph $G$ is a MIS of the square of $G$, i.e., the graph with the set of vertices of $G$ in which there is an edge between any two different vertices $u$ and $v$ if the distance between $u$ and $v$ in $G$ is at most $2$), colouring (a colouring of a graph $G$ assigns colours to vertices such that two neighbours have different colours), $2$-hop-colouring (as for a $2$-MIS, a $2$-hop-colouring of a graph $G$ is a colouring of the square of $G$). Each solution to one of these problems is a building block for many distributed algorithms: symmetry breaking, topology control, routing, resource allocation or network synchronisation.
As explained in [@Peleg] (p. 79), a MIS or a colouring enables the construction of schedules such that two neighbouring vertices do not act concurrently. Furthermore, a MIS can help for the decomposition of a network into clusters. A $2$-MIS makes it possible to assign each vertex to exactly one leader. Channel assignment for a radio network with collision-freedom corresponds to a $2$-hop-colouring of the graph corresponding to the network since each colour corresponds to a channel [@KMR01]. The importance of the $2$-hop-colouring is also attested by Emek et al. [@EPSW14], they prove that in an anonymous network any randomised algorithm can be seen as the composition of a randomised $2$-hop-colouring and a deterministic algorithm. Finally, in an anonymous wireless network there are no port numbers, in this context a $2$-hop-colouring ensures that no node has two neighbours with the same colour, and colours act as port numberings.
The aim of this work is the study of the resolution of these problems in the framework of beeping models.
In this paper, results on graphs having $n$ vertices are expressed with high probability (w.h.p. for short), meaning with probability $1-o(n^{-1})$.
Let $G$ be a graph and let $v$ be a vertex of $G$. We denote by $\Delta$ the maximum degree of $G$. The neighbourhood of $v$, denoted $N(v)$, is the set of vertices adjacent to $v$ (at distance $1$ from $v$). We define $\overline{N}(v)$ by including $v$ itself in $N(v)$. We use also the set of vertices at distance at most $2$ from $v$ called the $2$-neighbourhood and denoted $N_2(v)$. We write $\log n$ for the natural logarithm of $n$ and $\log_2 n$ for the logarithm of $n$ to the base $2$.
The Network Model
-----------------
We consider a wireless network model and we follow definitions given in [@Cornejo10] and in [@Afek13]. The network is anonymous: unique identities are not available to distinguish the processes. The network communications are synchronous and encoded by a connected graph $G=(V,E)$ where the vertices $V$ represent processes and the edges $E$ represent pairs of processes that can hear each other. We assume that all processes wake up and start computation at the same step. Time is divided into discrete synchronised time intervals, and during each time interval all processors act in parallel and:
- beep or listen;
- perform local computations.
Usually, in the message passing point to point model each interval is called a round, and in the context of wireless network model each interval is called a slot.
\[listen\] In general, vertices are active or passive. When they are active they beep or listen; in the description of algorithms we say explicitely when a vertex beeps meaning that a non beeping active vertex listens.
The time complexity, also called the slot complexity, is the maximum number of slots needed until every vertex has completed its computation.
Algorithms are expressed with a for-loop or an until-loop; in this paper, we call a phase one execution of the body of the for-loop or of the until-loop.
An algorithm given in the beeping model induces an algorithm in the message passing model; thus any lower bound on the round complexity in the message passing model is a lower bound on the number of slots in the beeping model.
Distributed Probabilistic Algorithm
-----------------------------------
A probabilistic algorithm is an algorithm which makes some random choices based on some given probability distributions.
A distributed probabilistic algorithm is a collection of local probabilistic algorithms. The network is anonymous, and processes have no information on their degrees; thus their local probabilistic algorithms are identical and have the same probability distribution.
A Las Vegas algorithm is a probabilistic algorithm which terminates with a positive probability (in general $1$) and always produces a correct result.
A Monte Carlo algorithm is a probabilistic algorithm which always terminates; nevertheless the result may be incorrect with a certain probability.
Our Contribution
----------------
Classical considerations on symmetry breaking in anonymous beeping networks, see for example [@Afek13] (Lemma 4.1) , imply that:
There is no Las Vegas internal collision detection algorithm in the beeping models $BL$ and $BL_{cd}$. There is no Las Vegas peripheral collision detection algorithm in the beeping models $BL$ and $B_{cd}L$.
Finally, a first contribution may be summarised by the following table. 0.3cm
[|l\*[4]{}[|c]{}|]{} & $BL$ & $B_{cd}L$ & $BL_{cd}$ & $B_{cd}L_{cd}$\
Collision Detection & MC & MC&MC &LV\
Degree & MC &MC &MC & LV\
Colouring & MC & LV & MC & LV\
2-colouring & MC & MC & MC & LV\
MC means there exists a Monte Carlo algorithm and there exists no Las Vegas algorithm.
LV means there exists a Las Vegas algorithm.
### Collision Detection.
We present and analyse very simple Monte Carlo procedures which detect internal and peripheral collisions in the beeping model $BL$.
Let $G$ be a graph and let $v$ be a vertex of $G$. According to the initial knowledge (error probability $\epsilon$ and/or the size of the graph), we prove that, given $0<\epsilon <1$, any collision in $N(v)$ is detected in $O\left(\log(\frac 1 \epsilon)\right)$ slots with an error probability upper bounded by $\epsilon$ or in $O(\log n)$ slots with an error probability $1-o(\frac 1 {n^2})$. Any collision in $G$ is detected in $O(\log (\frac n \epsilon))$ slots with an error probability upper bounded by $\epsilon$ and in $O(\log n)$ slots with probability $1-o\left(\frac 1 {n}\right)$, i.e., w.h.p.
### Colouring and $2$-hop-colouring Algorithms.
Algorithms for colouring and $2$-hop-colouring are based on a repeat-loop whose body has three parts:
1. a vertex is candidate to a colour and beeps with a certain probability which can change after each iteration,
2. a candidate vertex tries to detect whether it is the only candidate or not in $\overline{N}(v)$ or $N_2(v)$,
3. according to the conclusion, it informs its neighbours (and possibly neighbours of its neighbours) and may adjust its probability to be once again candidate.
We present and analyse a Las Vegas colouring algorithm in the model $B_{cd}L$; its slot complexity is $76 \log_2 n + 112\Delta$. We present also a $2$-hop-colouring Las Vegas algorithm in the model $B_{cd}L_{cd}$; its slot complexity is $5\times(76 \log_2 n + 112\Delta^2)$. In both cases algorithms need no knowledge on $G$.
In the case where we know an upper bound $K$ on the maximum degree of the graph we provide a colouring algorithm with colours bounded by $K+1$ and with a slot complexity equals to $O\left( K(\log n +\log^2K)\right)$.
### Emulation.
Based on results of the section devoted to collision detection we propose emulation procedures of $B_{cd}$ and of $L_{cd}$ in $BL$. Let $G$ be a graph. Each beep or listen is emulated by $k=2\times\lceil \log_2\left(\frac n\varepsilon\right) \rceil$ slots, and the procedures are correct on $G$ with probability $1-\varepsilon$, or by $k=2\times\lceil \log_2\left(\frac 1 \varepsilon\right) \rceil$ slots, and, for any vertex $v$, the procedures are correct on $v$ with probability $1-\varepsilon$, or by $k = 2\times\lceil 2\log_2 n\rceil$ slots, and the procedures are correct on $G$ w.h.p. Finally, emulation procedures induce a logarithmic multiplicative factor for the slot complexity.
### Degree Computation.
First, we deduce from the $2$-hop-colouring a Las Vegas degree computation algorithm in $B_{cd}L_{cd}$; its slot complexity is $5\times(76 \log_2 n + 112\Delta^2)$.
We illustrate emulation procedures by applying them to the degree computation algorithm given in $B_{cd}L_{cd}$ and we obtain a Monte Carlo algorithm for the computation of the degrees of each vertex in $BL$. For any graph $G$ of size $n$, the new algorithm computes the degrees in $G$ in $O\left((\log n+\Delta^2)\log n\right)$, and the result is correct w.h.p.
For some problems, the design of some algorithms is more natural and easier in $B_{cd}L_{cd}$ than in $B_{cd}L$ or is more natural and easier in $B_{cd}L$ than in $BL$. In these cases emulation procedures enable safe and automatic translations of algorithms given in a strong model into a weaker model.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Problem Beeping model Time (number of slots) Information required at each node error probability
------------------------------- ---------------- ------------------------------------------------------------------ ----------------------------------------------------------- -------------------------------------------
Collision detection in $N(v)$ $BL$ $O\left(\log(\frac 1 \epsilon)\right)$ $\epsilon$ Monte Carlo at most $\epsilon$
Collision detection in $N(v)$ $BL$ $O(\log n)$ size of the graph Monte Carlo $o\left(\frac 1 {n^2}\right)$
Collision detection in $G$ $BL$ $O\left(\log(\frac size of the graph and $\epsilon$ Monte Carlo at most $\epsilon$
n \epsilon)\right)$
Collision detection in $G$ $BL$ $O\left(\log n\right)$ size of the graph Monte Carlo $o\left(\frac 1 {n}\right)$
MIS [@Scott13] $B_{cd}L$ $O(\log n)$ none Las Vegas
Colouring [@Cornejo10] $BL$ never stops stabilisation w.h.p. in $O(\Delta \log n)$ Each node knows its degree and an upper bound of $\Delta$ Monte Carlo
Colouring $B_{cd}L$ $O(\log n+\Delta)$ w.h.p. none Las Vegas
Colouring $B_{cd}L$ $O\left( K(\log n +\log^2K)\right)$ w.h.p. An upper bound $K$ on the maximum degree of $G$ Las Vegas
$2$-colouring $B_{cd}L_{cd}$ $O(\log n +\Delta^2)$ w.h.p. none Las Vegas
Degree computation $B_{cd}L_{cd}$ $O(\log n +\Delta^2)$ w.h.p. none Las Vegas
Degree computation $BL$ $O\left((\log n + \Delta^2)(\log(\frac{n}{\varepsilon}))\right)$ size of the graph and $\varepsilon$ Monte Carlo at most $\varepsilon$
Degree computation $BL$ $O\left((\log n+\Delta^2)\log n\right)$ size of the graph Monte Carlo $o(\left( \frac 1 n \right))$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Beeping algorithms on graphs with $n$ vertices.
Related Work
------------
As explained by Chlebus [@C01], in a radio network, a vertex can hear a message only if it was sent by a neighbour and this neighbour was the only neighbour that performed a send operation in that step. If no message has been sent to a vertex then it hears the background noise. If a vertex $v$ receives more than one message then we say that a collision occurred at the vertex $v$ and the vertex hears the interference noise. If vertices of a network can distinguish the background noise from the interference noise then the network is said to be with collision detection, otherwise it is without collision detection (see for example the Wake-up problem or the MIS problem for radio networks in [@GPP01; @Moswa; @CGK07; @JK15] where vertices do not make the difference between no neighbour sends a message and at least two neighbours send a message; see also the broadcasting problem in radio network in [@GHK13] where vertices make the difference between no neighbour sends a message, exactly one neighbour send a message and at least two neighbours send a message). In this context, an efficient randomised emulation of single-hop radio network with collision detection on multi-hop radio network without collision detection is presented and analysed in [@BGI91]. To summarise:
Detecting a collision in a radio network is to be able to distinguish between $0$ message and at least $2$ messages while detecting a collision in the beeping model is to be able to distinguish between $1$ message and at least $2$ messages.
Thus, from now on, we consider collisions as explained above for beeping models. Our collision detetection algorithm and the degree computation algorithm use similar ideas to those used for initialising a packet radio network [@HNO99] or for election in a complete graph with wireless communications [@BW12] (Algorithm 50, p. 132). The impact of collision detection is studied in [@Schneider10; @KP13], where it is proved that performances are improved, and in certain cases the improvement can be exponential. The complexity of the conflict resolution problem (the goal is to let every active vertex use the channel alone (without collision) at least once) is studied in [@HuangM13] (they assume that vertices are identified), and an efficient deterministic solution is presented and analysed.
General considerations and many examples of Las Vegas distributed algorithms related to MIS or colouring can be found in [@Peleg]. The computation of a MIS has been the object of extensive research on parallel and distributed complexity in the point to point message passing model [@Alonco; @Luby] [@Awerbuchco; @Linial]; Karp and Wigderson [@KarpW] proved that the MIS problem is in NC. Some links with distributed graph colouring and some recent results on this problem can be found in [@Kuhnco]. The complexity of some special classes of graphs such as growth-bounded graphs is studied in [@Kuhnmo]. Results have been obtained also for radio networks [@Moswa]. A major contribution is due to Luby [@Luby]. He gives a Las Vegas distributed algorithm. The main idea is to obtain for each vertex a [*local total order*]{} or a [local election]{} which breaks the local symmetry and then each vertex can decide locally whether it joins the MIS or not. Its time complexity is $O(\log n)$ and its bit complexity is $O(\log^2 n).$ Recently, a Las Vegas distributed algorithm has been presented in [@MRSZ11] which improved the bit complexity: its bit complexity is optimal and equal to $O(\log n)$ w.h.p. An experimental comparison between [@Luby] and [@MRSZ11] is presented in [@BK13]. If we remove the constraint on the size of messages or on the anonymity recent new results have been obtained for distributed symmetry breaking (MIS or colouring) in [@KothaP11; @BarenEPS12; @BarenE13; @BarenE14].
Afek et al. [@Afek13], from considerations concerning the development of certain cells, studied the MIS problem in the discrete beeping model $BL$ as presented in [@Cornejo10]. They consider, in particular, the wake-on-beep model (sleeping nodes wake up upon receiving a beep) and sender-side collision detection $B_{cd}L$: they give an $O((\log n)^2)$ rounds MIS algorithm. After this work, Scott et al. [@Scott13] presents in the model $B_{cd}L$ a randomised algorithm with feedback mechanism whose expected time to compute a MIS is $O(\log n)$. A vertex $v$ is candidate for joining the independent set (and beeps) with a certain probability (initially $1/2$); this value is decreased by some fixed factor if at least one neighbour whishes also to join the independent set. It is increased by the same factor (up to maximum $1/2$) if neither $v$ nor any neighbour of $v$ are candidates.
More generally, Navlakha and Bar-Joseph present in [@NB15] a general survey on similarities and differences between distributed computations in biological and computational systems and, in this framework, the importance of the beeping model.
In the model of point to point message passing, vertex colouring is mainly studied under two assumptions: - vertices have unique identifiers, and more generally, they have an initial colouring, - every vertex has the same initial state and initially only knows its own edges. If vertices have an initial colour, Kuhn and Wattenhofer [@Kuhnco] have obtained efficient time complexity algorithms to obtain $O(\Delta)$ colours in the case where every vertex can only send its own current colour to all its neighbours. In [@Johansson], Johansson analyses a simple randomised distributed vertex colouring algorithm for anonymous graphs. He proves that this algorithm runs in $O(\log n)$ rounds w.h.p. on graphs of size $n.$ The size of each message is $\log n,$ thus the bit complexity per channel of this algorithm is $O(\log^2 n).$ [@MRSZ10] presents an optimal bit and time complexity Las Vegas distributed algorithm for colouring any anonymous graph in $O(\log n)$ bit rounds w.h.p.
In [@Cornejo10], Cornejo and Kuhn study the interval colouring problem: an interval colouring assigns to each vertex an interval (contiguous fraction) of resources such that neighbouring vertices do not share resources (it is a variant of vertex colouring). They assume that each node knows its degree and an upper bound of the maximum degree $\Delta$ of the graph. They present in the beeping model $BL$ a probabilistic algorithm which never stops and stabilises with a correct $O(\Delta)$-interval coloring in $O(\log n)$ periods w.h.p., where: $n$ is the size of the graph, and a period is $Q$ time slots with $Q\geq \Delta$, thus it stabilises in $O(Q(\log n))$ slots.
Kothapalli et al. consider the family of anonymous rings and show in [@KOSS] that if only one bit can be sent along each edge in a round (point to point message passing model), then every Las Vegas distributed vertex colouring algorithm (in which every node has the same initial state and initially only knows its own edges) needs $\Omega(\log n)$ rounds w.h.p. to colour the ring of size $n$ with any finite number of colours. Kothapalli et al. consider also the family of oriented rings and they prove that the bit complexity in this family is $\Omega(\sqrt{\log
n})$ w.h.p.
[@FMRZ13] presents and analyses Las Vegas distributed algorithms which compute a MIS or a maximal matching for anonymous rings (in the point to point message passing model). Their bit complexity and time complexity are $O(\sqrt{\log n})$ w.h.p.
Emek and Wattenhofer introduce in [@EmekW13] a model for distributed computations which resembles the beeping model: networked finite state machines (nFSM for short). This model enables the sending of the same message to all neighbours of a vertex; however it is asynchronous, the states of vertices belong to a finite set, the degree of vertices is bounded and the set of messages is also finite. In the nFSM model they give a $2$-MIS algorithm for graphs of size $n$ using a set of messages of size $3$ with a time complexity equal to $O({\log n}^2).$
A Monte Carlo Collision Detection Algorithm in $BL$ {#collision}
===================================================
If we consider the beeping models presented in the Introduction, clearly the weakest is $BL$. This section presents simple and efficient probabilistic procedures for detecting collisions by using $BL$. Later (Section \[emuler\]) we will see how to emulate $B_{cd}$ or $L_{cd}$ in $BL$.
A phase $P$ is the sequence of the 3 following actions:
- vertices wishing to beep, randomly and uniformly select 0 or 1;
- slot $1$: vertices that have drawn 0 beep, the others listen;
- slot $2$: vertices that have drawn 1 beep, the others listen.
A vertex detects a collision if:
- it does not beep and it hears beeps at two slots in a phase,
- or if it beeps itself at a slot of a phase and hears a beep at the other slot of the same phase.
We address two questions:
Let $0<\epsilon<1$, how many phases must each vertex execute to decide whether there is a collision or not in its neighbourhood with an error probability bounded by $\epsilon$?
Let $0<\epsilon<1$, how many phases must each vertex execute to ensure that whether there is a collision or not over all the graph $G$ is detected with an error probability bounded by $\epsilon$?
**Var:**\
$k:$ **Global integer constant;**\
$collision:$ $boolean$ **Init** $false$;\
$i:$ $Integer;$\
$b:$ **in** $\{0,1\}$;\
We have:
\[lemma::collision\_local\] Let $G$ be a graph having $n$ vertices. Let $v$ be any vertex. Let $0<\epsilon<1$. Any collision in the neighbourhood of $v$ is detected in $O\left(\log_2(\frac 1 \epsilon)\right)$ phases (slots) with probability at least $1-\epsilon,$ and in $O\left(\log_2 n\right)$ phases (slots) with probability $1-o\left(\frac 1 {n^2}\right)$.
Let $v$ be any vertex having $d(v)\geq 1$ neighbours. If a collision happens between $u_1$, which is either $v$ or a neighbour of $v$ and $u_2$, a neighbour of $v$, then it will be detected if and only if $u_1$ chooses a slot different from $u_2$. This happens with probability $1/2$.
Thus, the probability that a collision happens and is not detected in the neighbourhood of $v$ within next $k$ phases is at most $\left(\frac 1 {2}\right)^k$. This probability is then less than $\epsilon$ (resp. less than $o\left(\frac 1 {n^2}\right)$) for any $k>\log_2(\frac 1 \epsilon)$ (resp. $k>2\log_2(n)$), which ends the proof.
Yielding:
Let $G$ be a graph having $n$ vertices. Any collision in $G$ is detected after at most $O\left(\log_2(\frac
n \epsilon)\right)$ phases (slots) with probability at least $1-\epsilon$, and after at most $O\left(\log_2 n\right)$ phases (slots) with probability $1-o\left(\frac 1 {n}\right)$.
Assume a collision occurs at time $t_0$ in $G$ and let $T$ denote the number of phases before it is detected in the whole graph. Clearly $T=\max\{
T_v\mid v\in V\}$, where $T_v$ denotes the time before a node $v$ detects a collision in its neighbourhood and then:
$$\begin{aligned}
{\mathbb P}r
\left(
T>\log_2\left(\frac n \epsilon\right)\right)
& \leq & n\times
{\mathbb P}r
\left(
T_v>\log_2\left(\frac n \epsilon\right)\right)\\
& = & n\times \frac 1 {2^{\log_2(\frac n \epsilon)}}= \epsilon.\end{aligned}$$
Which proves the first claim. The same argument, combined with the second claim of Lemma \[lemma::collision\_local\] proves the second claim of the corollary.
These results can be summarised by Algorithm 1 (Monte Carlo).
Colouring Algorithms
====================
A Las Vegas Colouring Algorithm in $B_{cd}L$ without any knowledge {#colouring}
------------------------------------------------------------------
This section presents and analyse a Las Vegas colouring algorithm in the model $B_{cd}L$ assuming that the vertices have no knowledge.
Initially each vertex is active. Each active vertex $v$ maintains a parameter $p$, its “beeping probability” initially equal to $1/2$. It maintains also a counter, denoted colour (initially equal to $0$), that is incremented at each iteration. In each phase each active vertex decides with probability $p$ to beep, indicating that it is a candidate to the current colour given by the counter. It succeeds and its colour is the value of the counter if and only if no neighbour has also beeped; in this case its state becomes coloured. Then after this slot, if $v$ is still active, it adjusts $p$, halving it if any neighbour beeped and doubling it if no neighbour beeped and it is not already $1/2$. If a neighbour has beeped we say that $v$ is “inhibited”.
\[Scott\]
**Var:**\
$state\in\{active,coloured\}$ **Init** $active$;\
$candidate:$ $Boolean$;\
$p:$ $real$ **Init** $1/2$;\
$colour:$ $Integer$ **Init** $0$;\
At the end of the body of the until-loop, we can add a slot which enables an uncoloured vertex to beep and finally a couloured vertex can detect the local termination of the colouring algorithm.
We first introduce some notation that we will use in this proof. For any vertex $v$, $p_v$ denote the parameter $p$ on the vertex $v$ and we define the following sum: $$q_v = \sum_{u\in N(v)}p_u.$$
We also note $q_v^* = \max\{q_v,1/5\}$ and finally $t_0 = 3 \log_2
(5q_v^*)- 2 \log_2 p_v$. We omit the subscript $v$ where there is no risk of ambiguity.
We finally write $l(q)$ for $\log_2 (5\max\{q,1/5\})$, that is $l(q)=\max\{\log_2(5q),0\}$.
Recall that $\overline{N}(v)$ is the set of vertices at distance less than or equal to $1$ from vertex $v$.
Then, we have the following theorem:
For any $t\ge 0$ and for any vertex $v$, its probability of remaining active after the next $t$ phases is at most $\alpha^{112d(v)+t_0-t}$ for the constant $\alpha=2^{1/36} \approx 1.01944$, where $d(v)$ is the degree of $v$ in the residual graph.
Note that $\alpha^{3\log_2q}=q^{3\log_2\alpha}=q^{1/12}$. The proof will be by induction on $t$. We have $t_0\ge 2$, so that if $t=0$, $\alpha^{t_0-t} >1$ and the claim is trivially true.
Let $t>0$. After one phase which does not colour $v$ we have by induction that the probability of remaining active for the following $t-1$ phases is at most $\alpha^{112d^\prime(v)+t^\prime_0-t+1}$ where $t^\prime_0$ is the new value of $t_0$, namely $ 3
l(q^{\prime})- 2\log_2 p^\prime$ and $d^\prime(v)$ is the new degree. So we conclude that the probability of survival is upper bounded by the mean of the random variable which is $\alpha^{112d^\prime(v)+t^\prime_0-t+1}$ if $v$ survives the first phase and $0$ otherwise. We refer to this mean as the [*bound*]{} and note that it is dependent on what happens outside the neighbourhood of $v$.
We will come back to the proof of the Theorem, but we first prove the following lemma:
The bound is maximised when what happens outside the neighbourhood of $v$ is that every neighbour $u$ of $v$ is inhibited from taking the current colour by an external neighbour beeping.
Proof\
Consider any external behaviour $E$ in which some $u$ is not inhibited; we will show that the bound is increased or unchanged if the behaviour is changed to $E^\prime$ in which $u$ is inhibited and there is no change for any other neighbours of $v$. (In a given graph there may be no such $E^\prime$ but we consider the maximum possible over any graph containing the neighbourhood $\overline{N}(v)$.) We consider fixed beeping decisions of all vertices in $\overline{N}(v)$ except $u$ and show that with these decisions $E^\prime$ gives a value of the bound greater than or equal to that of $E$. We consider two cases:
- Some neighbour of $u$ in $\overline{N}(v)$ beeps:\
$p_u$ will be halved whether or not $u$ is inhibited by $E^{\prime}$ and so $p^\prime$, $q^\prime$, $d^\prime(v)$ and the probability of survival are the same for $E$ and $E^\prime$. The bound is identical in the two cases.
- Otherwise:
Let the value of $p^\prime$ be $p_0$ if $u$ does not beep and $p_1$ if $u$ does beep. $p_1 \le p_0$.
Let the value of $q^\prime$ be $q_0$ if $u$ does not beep and is not inhibited, $q_1$ if it beeps and is inhibited and $q_2$ if it does not beep and is inhibited. Note that if $u$ beeps and is not inhibited, $u$ takes the current colour; we note the value of $q^\prime$ in this case as $q_3$ and note that $q_3<q_0$ since the effect of $u$ beeping is to remove $p_u$ from the sum for $q$ and possibly to halve the values of $p$ for some common neighbours of $u$ and $v$. We have $q_1\ge q_0/4$ since, at most, $u$’s beeping can result in a vertex $w$ halving $q_w$ when otherwise it would have doubled it. Similarly $q_2\ge q_0/4$ and $q_2 \ge q_0-3p_u/2$ since the inhibition results in $p_u$ being halved rather than potentially doubled.
Let $d_0$ be the new value of $d(v)$ if $u$ does not take the current colour; if it does, then the new value is $d_0-1$.
The bounds are thus $p_u\alpha ^ {112d_0+ 3 l(q_1)- 2 \log_2(p_1)-t+1}+
(1-p_u)\alpha ^ {112d_0+ 3 l(q_2)- 2 \log_2(p_0)-t+1}$ in the inhibited case and $(1-p_u)\alpha ^ {112d_0+ 3 l(q_0)- 2 \log_2(p_0)-t+1}+p_u\alpha^{112(d_0-1)+3l(q_3)-2\log_2(p_1)}$ in the uninhibited case. We claim that the ratio of the inhibited bound to the uninhibited is at least $1$. This ratio $\ge \frac{p_u\alpha^{ 3
l(q_1)}+(1-p_u)\alpha^{ 3 l(q_2)}} {(1-p_u)\alpha^{ 3 l(q_0)}+p_u\alpha^{3l(q_0)}/8}$ (since $p_1\le p_0$, $p_1\ge p_0/4$, $q_3<q_0$) and $\alpha^{108}=8$\
Remember that $p_u$ is a power of $1/2$. We consider four subcases:
- $q_0 \le 1/5$: $l(q_1)=l(q_2)=l(q_0)=0$ and the ratio $\ge (p_u+1-p_u)/(1-p_u+p_u/8) > 1$.
- $1/5 < q_0$ and $p_u\ge 1/8$: We use the bounds $q_1 \ge q_0/4$ and $q_2 \ge q_0/4$ giving that the ratio is at least $(p_u+1-p_u)\alpha^{-6}/(1-p_u+p_u/8)$ $= \alpha^{-6}/(1-p_u+p_u/8) \ge \alpha^{-6}/(7/8+1/64) \ge 1$.
- $1/5 <q_0 \le 4/5$ and $p_u \le 1/16$: We use the bounds $q_1
\ge q_0/4$ and $q_2 \ge q_0 -3p_u/2$ and the fact that for $0<x \le
15/32$, $(1-x)^{1/12} > 1-4/3(x/12)$ so that the ratio is at least $(p_u\alpha^{-6}/(1-p_u) + (1-3p_u/2q_0)^{3\log_2\alpha})\frac{1-p_u}{1-p_u(1-1/8)}$ $\ge (p_u\alpha^{-6} + (1-15p_u/2)^{1/12})\frac{1-1/16}{1-(1-1/8)/16}$ $\ge (p_u\alpha^{-6} + (1-(15p_u/2)/12\times (4/3)))\frac{120}{121}$ $\ge (1 + p_u(\alpha^{-6} - 5/6) )\frac{120}{121}> 1$.
- $q_0 > 4/5$ and $p_u\le 1/16$: Using the same bounds as in the previous subcase the ratio is greater than $(\frac{p_u}{1-p_u}\alpha^{ -6} + \alpha^{ 3 (l(q_0-3p_u/2)-l(q_0))})\frac{120}{121}$ $>(\frac{p_u}{1-p_u}\alpha^{-6 } + \alpha^{ 3
(l(4/5-3p_u/2)-l(4/5))})\frac{120}{121}$ and this is the bound already used for the case with $q_0=4/5$ and the same value of $p_u$ and so is greater than or equal to $1$.
This ends the proof that $E^\prime$ gives a value for the bound at least as great as that for $E$. The lemma is then proved by a simple induction on the number of uninhibited vertices.
We return to the inductive proof. Using the lemma we will always take $q^{\prime}=q/2$ giving probability of survival $\le \alpha^{ 112d^\prime(v)+3 l(q/2) - 2 \log_2 p^{\prime}-t+1}$ $\le \alpha^{ 112d(v)+3 l(q/2) - 2 \log_2 p^{\prime}-t+1}$.
We consider five cases.
- $q \ge 2/5$: We have $l(q/2)= l(q)-1$ and $p^\prime \ge p/2$ giving $${\mathbb P}r(survival) \le \alpha^{112d(v)+ 3 (l(q)-1)- 2 (\log_2 p-1)-t+1}
=\alpha^{112d(v)+ 3 l(q)- 2 (\log_2 p)-t}$$ as claimed.
- $1/5 \le q < 2/5$ and $p<1/2$: The probability that a neighbour of $v$ beeps is less than $q$ so that $p_v$ is doubled with probability at least $1-q$ and halved in the remaining cases. In all cases $l(q/2)=0$. Hence $P(survival) \le
\alpha^{112d(v)-2\log_2(p)-t+1}((1-q)\alpha^{- 2 } +q\alpha^{ 2 })$ and our claim is that it is at most $\alpha^{112d(v)+3\log_2(5q)-2\log_2(p)-t}$. That is the claim is valid since $(1-q)\alpha^{- 1 } +q\alpha^{ 3 }\le\alpha^{3\log_2(5q)}$ in the range $1/5 \le q < 2/5$. (It is valid at $q=1/5$ since $4\alpha^{-1}+\alpha^3<5$ and at $q=2/5$ since $3\alpha^{-1}+2\alpha^3<5\alpha^3$; between these two limits, the left hand side is linear and the right hand side ($(5q)^{3\log_2\alpha}$) has a negative second derivative so the inequality holds there also.)
- $1/5 \le q < 2/5$ and $p=1/2$: With probability greater than $1-q$ no neighbour of $v$ beeps and then $v$ has probability $1/2$ of taking the current colour; otherwise $p_v$ remains $1/2$. On the other hand, if a neighbour does beep, $p_v$ becomes $1/4$. In all cases $l(q/2)=0$. Thus the probability of survival $\le
\alpha^{112d(v)+2-t+1}((1-q)/2+q\alpha^2)$ and the claim is that it is at most $\alpha^{112d(v)+3\log_2(5q)+2-t}$. That is the claim is valid if $(1-q)\alpha/2+ q\alpha^{ 3 }\le\alpha^{3\log_2(5q)}$ a weaker condition than in the previous case.
- $q < 1/5$ and $p<1/2$: The probability that a neighbour of $v$ beeps is less than $1/5$ so that $p_v$ is doubled with probability at least $4/5$ and halved in the remaining cases. In all cases $l(q)$ decreases or is unchanged. Hence ${\mathbb P}r(survival) \le \alpha^{112d(v)+ 3
l(q) - 2 \log_2(p)-t+1} ((4/5)\alpha^{- 2 } +(1/5)\alpha^{ 2 })$ and this is less than $\alpha^{112d(v)+ 3 l(q) - 2 \log_2 p-t}$ as claimed, again since $4\alpha^{- 1 }+ \alpha^{ 3 }<5 $.
- $q < 1/5$ and $p=1/2$: With probability greater than $4/5$ no neighbour of $v$ beeps and then $v$ has probability $1/2$ of taking the current colour; otherwise $p_v$ remains $1/2$. On the other hand, if a neighbour does beep, $q$ decreases and $p_v$ becomes $1/4$. Hence ${\mathbb P}r(survival) \le (2\alpha^{112d(v)+ 3 l(q/2)- 2
\log_2(1/2)-t+1}+ \alpha^{ 3 l(q/2) - 2 \log_2(1/4)-t+1})/5$ $\le
\alpha^{ 3 l(q)- 2 \log_2(1/2)-t+1}(2+\alpha^{ 2 })/5$ which is at most $\alpha^{112d(v)+ 3 l(q) - 2 \log_2(1/2)-t}$ as claimed since $2+\alpha^{ 2 }<5\alpha^{-1} $.
This completes the proof of the theorem.
The complexity of Algorithm \[Scott\] is described by:
\[coloriagesans\] The number of phases (slots) taken by the colouring algorithm on any graph with $n$ nodes and maximum degree $\Delta$ is at most $76\log_2 n+112\Delta$ w.h.p.
Since initially $p_v=1/2$ and $q_v <n/2$ where the graph has $n$ vertices, we conclude that $t_0 < 3\log_2(5n/2)-2\log_2(1/2)<3\log_2
n+6$ so that after $t\ge 112\Delta +
76\log_2 n+6$ phases, any vertex has probability $\alpha^{3\log_2n+6-(76\log_2n+6)}=n^{-73/36}$ of survival and the probability that any vertex survives is at most $n^{-37/36}=o(n^{-1})$.
The number of colours used by the colouring algorithm is at most $76\log_2 n+112\Delta$ w.h.p.
A Las Vegas Colouring Algorithm with the Knowledge of an Upper Bound of the Maximum Degree in $B_{cd}L$ {#colwku}
-------------------------------------------------------------------------------------------------------
This section presents and analyses a Las Vegas colouring algorithm in the model $B_{cd}L$, assuming that the vertices are aware of an upper bound $K$ on the maximum degree $\Delta$ of the graph. So we aim to compute a $K+1$ colouring.
Each vertex has a counter (initially, its value is $0$) and a set of colours: $\{0,\cdots,K\}$. Each phase corresponds to three slots. In the first slot an uncoloured vertex tries to get a colour by beeping with a certain probability if the counter belongs to the set of colours. When a vertex beeps in the second slot, this means that it succeeds in choosing a colour (the current value of the counter), so there is no need to detect collision in this slot. Vertices which hear a beep at slot 2 withdraw the corresponding colour.
\[algorithm::coulouringBcdL\]
**Var:**\
$K:$ **Global integer constant upper bound on the maximum degree of $G$;**\
$state$ $\in \{Active, Inactive\}$ **Init** $Active;$\
$Colours$ $=\{0,\cdots,K\}$;\
$Colour$ $\in \{0,\cdots,K\}$ **Init** $0;$\
$counter$ $\in \{0,\cdots,K\}$ **Init** $0;$\
$slot:$ $Integer$;\
We can consider the [*modified*]{} colouring algorithm defined in the following way. By a [*cycle*]{} we mean $K$ rounds considering the $K$ colours. Now, every vertex uses the value of $|Colours|$ at the start of each cycle to decide the beeping probability it uses throughout this cycle.
### Analysis of the Algorithm.
We have the following theorem:
Let $G$ be a graph of size $n$, let $K$ be an upper bound on the maximum degree of $G$. The Colouring algorithm computes a $K+1$ colouring of $G$ in at most $O\left( K(\log n +\log^2K)\right)$ w.h.p.
We consider the Colouring algorithm in which every vertex has the same upper bound $K$ on the maximum degree. We consider both the [*basic*]{} algorithm in which $v$ uses the current value of $|Colours|$ to decide its beeping probability and also the [*modified*]{} algorithm in which it uses the value at the start of the current cycle. We recall that by a [*cycle*]{} we mean $K$ rounds considering the $|Colours|$ colours.
We consider $P_k$ the probability that vertex $v$ survives uncoloured over $k$ cycles.
In what follows
- $i$ ranges over $1..k$,\
- $c$ ranges over the $C_i$ colours possible for $v$ at the start of cycle $i$,\
- $u$ ranges over the neighbours of $v$ still uncoloured at the start of cycle $i$,\
- $p_u(i,c)$ is the probability that $u$ beeps at colour $c$ in cycle $i$.
First we consider the probability $p$ that $v$ survives uncoloured in a single round using a colour $c \in colours(v)$ . $$\begin{aligned}
p & = & {\mathbb P}r\left(v{\rm~does~not~beep~at~colour~}c{\rm~in~cycle~}i\right)\nonumber\\
& + & {\mathbb P}r\left(v{\rm~does~beep~and~some~neighbour~}u{\rm~also~beeps}\right)\nonumber\end{aligned}$$ but ${\mathbb P}r\left(v{\rm~does~beep}\right) \ge 1/2C_i$ and the beeping probabilities of $v$ and its neighbours are independent giving $$\begin{aligned}
p & \le & \left(1-1/2C_i\right) + {\mathbb P}r\left({\rm some~neighbour~beeps}\right)/2C_i\nonumber\\
& = & \left(1-1/2C_i\right)\left(1+{\mathbb P}r\left({\rm some~neighbour~beeps}\right)/(2C_i-1)\right)\nonumber\\
& \le & \left(1-1/2C_i\right)\left(1+\sum_u p_u(i,c)/(2C_i-1)\right).\nonumber\end{aligned}$$ After the first round, $p_u(i,c)$ and $C_i$ are random variables dependent on what has happened so far, and we consider the tree of all possible executions up to $k$ cycles, where each tree node has its own value of $p$. It is easily shown by induction that $P_k$ is upper bounded by the maximum over all paths in this tree of the product of the values of $p$ along the path. We fix a path which gives this maximum and bound the product for this path. We have the probability of surviving cycle $i$ $\le (exp(-1/2) * \prod_c (1+\sum_u p_u(i,c)/(2C_i-1)))$ $\le exp(-1/2 + \sum_c \sum_u p_u(i,c)/(2C_i-1))$ and so $P_k \le exp(-k/2 + \sum_i \sum_c \sum_u p_u(i,c)/(2C_i-1))$. We will give an upper bound on $\sum_i \sum_c \sum_u p_u(i,c)/(2C_i-1)$.
We number $v$’s neighbours in the initial graph from $1$ to $deg(v)$ in decreasing order of their [*lifetime*]{}, that is the number of rounds in which they remain uncoloured:\
Thus as long as $u_j$ is not coloured the degree of $v$ in the residual graph is at least $j$ and so $|colours(v)|>j$.
We write $p_u(i,c)$ as $base + \delta$ where $base = 1/2C_i$ and $\delta$ is what has been added as a result of $colours(u)$ being decreased before colour $c$ and we will bound $\sum_i \sum_u \sum_c base/(2C_i-1)$ and $\sum_u \sum_i \sum_c \delta/(2C_i-1)$ separately.
Firstly $base$: in cycle $i$, $v$ has $C_i$ colours available and so has less than $C_i$ neighbours; each neighbour $u$ has $\sum_c base \le 1/2$, giving, for this cycle, $\sum_u \sum_c base/(2C_i-1)\le 1/6$ so that $\sum_i \sum_u \sum_c base/(2C_i-1) \le k/6$.
Secondly $\delta$: For the modified algorithm $\delta=0$. In the basic algorithm, a vertex $u_j$ initially has $K$ colours available and when (if) this number decreases from $l$ to $l-1$, $p_u(i,c)$ increases from $1/2l$ to $1/2(l-1)$ and this increase of $1/2l(l-1)$ affects $\delta$ only for the, at most, $l-1$ colours still to be considered in this cycle so that $ \sum_c \delta$ for a cycle is at most $\sum_l 1/2l$, the sum being taken over those $l$ for which the number of colours is reduced from $l$. This gives an upper bound on $ \sum_i \sum_c \delta/(2C_i-1)$ of $\log K/2(2j+1)$ since $C_i>j$ and so $\sum_u \sum_i \sum_c \delta/(2C_i-1) < \sum_j \log K/2(2j+1) < \log^2 K/4$.
Hence, by standard arguments, after $k=O(\log n + \log^2 K)$ cycles for the basic algorithm or $O(\log n)$ cycles for the modified algorithm, $v$ has probability $o(1/n^2)$ of remaining uncoloured and the graph has probability $o(1/n)$ of having any uncoloured vertex.
A Las Vegas Algorithm for $2$-hop-colouring in $B_{cd}L_{cd}$ without any Knowledge {#2col}
===================================================================================
To calculate a $2$-hop-colouring of a graph $G$, we need to calculate a colouring of the “square” of $G$, that is the graph with the same vertices as $G$ and an edge between any pair $v$ and $w$ of vertices which either are neighbours in $G$ or have a common neighbour in $G$. Algorithm \[Scott\] (Section \[colouring\]) is modified to perform the computation of the colouring in the square of $G$, i.e., the $2$-hop-colouring of $G$ in $B_{cd}L_{cd}$.
At slot 1, an active vertex beeps with a certain probability. At slot 2, a vertex having two beeping neighbours beeps. Thus, a candidate vertex, which beeps without internal collision and which has no neighbours having detected a peripheral collision, has beeped alone among vertices at distance at most $2$ and it becomes coloured. At slot 3, an active vertex having heard at least one beep beeps. Finaly after slot 3, an active vertex knows whether at least one vertex beeped at distance at most $2$ to possibly change its probability (as in Algorithm \[Scott\]) to be candidate.
**Var:**\
$state\in\{active,coloured,turned$-$off\}$ **Init** $active$;\
$p:$ $real$ **Init** $1/2$;\
$slot:$ $Integer$;\
$colour:$ $Integer$ **Init** $0$ ;\
In this context, Theorem \[coloriagesans\] becomes:
\[2hopcolour\] The number of phases taken by the $2$-hop-colouring algorithm on any graph with $n$ nodes and maximum degree $\Delta$ is at most $76\log_2 n+112\Delta^2$ w.h.p. (the number of slots is at most $4\times(76\log_2 n+112\Delta^2)$).
The same transformation can be done the algorithm given in section \[colwku\] when we know an upper bound of the maximum degree.
A Las Vegas Degree Computation Algorithm in $B_{cd}L_{cd}$ {#degreecdcd}
==========================================================
The $2$-hop-colouring algorithm may be viewed as a degree computation algorithm. We present in this section a Las Vegas Degree Computation Algorithm in $B_{cd}L_{cd}$, Algorithm \[degree\], inspired by the $2$-hop-colouring algorithm given in Section \[2col\]. The idea is very simple: each vertex tries to be counted by its neighbours by beeping alone among vertices at distance at most two. When it is the case it informs its neighbours which increment their degree, it no longer tries to be counted and listens until the end of the algorithm for counting its neighbours. Slot 5 allows a vertex to detect the termination of the computation of its degree.
\[degree\]
**Var:**\
$state\in\{active,passive,turned$-$off\}$ **Init** $active$;\
$p:$ $real$ **Init** $1/2$;\
$slot:$ $Integer$;\
$deg:$ $Integer$ **Init** $0$ ;\
The degree algorithm allows each vertex to know its degree.
We deduce from Theorem \[2hopcolour\] that:
The number of phases taken by the degree algorithm on any graph with $n$ nodes and maximum degree $\Delta$ is at most $76\log_2 n+112\Delta^2$ w.h.p. (the number of slots is $5\times (76\log_2 n+112\Delta^2)$).
Emulating $B_{cd}$ or $L_{cd}$ in $BL$ {#emuler}
======================================
This section presents randomised emulation procedure (Algorithm \[algorithm::emulateBcd\] and Algorithm \[algorithm::procemulLcd\]) of $B_{cd}$ and $L_{cd}$ in $BL$.
The first procedure, $EmulateB_{cd}inBL()$, emulates a beeping slot in $B_{cd}$ in $BL$. The second, $EmulateL_{cd}inBL()$, emulates a listening slot in $L_{cd}$ in $BL$. Both procedures are Monte Carlo procedures and parametrized with an integer $k>1$ and an output boolean parameter $collision$ which indicates whether a collision has been detected. The parameter $k$ controls the probability of error for the collision detection.
Let $v$ be a vertex. Let $k$ be a vertex. We denote by $s$ the signature of the vertex $v$ which is the word formed by $k$ bits generated uniformly at random; it is denoted by $s:=gen(k)$.
Before any emulation each vertex generates its signature $s$ which depends on $k$: $s:=gen(k);$ then it uses the following procedures (Algorithm 6 and Algorithm 7).
\[algorithm::emulateBcd\]
**Procedure** $EmulateB_{cd}inBL$(**IN:**[$s:$ word of bits associated to the vertex]{}**;** **OUT:** [$collision: boolean$]{}) $collision := false;$
$ i := 0;$
**End Procedure**
\[algorithm::procemulLcd\]
**Procedure** $EmulateL_{cd}inBL$(**IN:**[$k$: Global integer constant]{}**;** **OUT:** [$collision: boolean$]{} ) $collision := false;$
$i := 0;$
**End Procedure**
The value of $k$ depends on the bound of the error probability we require, a straightforward adaptation of the analysis done in Section \[collision\] gives:
\[emule\] For any $\varepsilon>0$, and any $n>0$:
1. if $k=\lceil \log_2\left(\frac n\varepsilon\right) \rceil$, then, the procedures are correct on $G$ with probability $1-\varepsilon$,
2. if $k=\lceil \log_2\left(\frac 1 \varepsilon\right) \rceil$, then, for any vertex $v$, the procedures are correct on $v$ with probability $1-\varepsilon$,
3. if $k = \lceil 2\log_2(n)\rceil$, then, the procedures are correct on $G$ w.h.p.
In the emulation procedures, the for-loops are controlled by $k$ thus the first item of Lemma \[emule\] needs knowledge of $n$ and $\varepsilon$, the second item needs only knowledge of $\varepsilon$ and the last item needs knowledge of $n$.
Computing the Degree of each Vertex in $BL$
===========================================
This section illustrates emulation procedures by applying them to the Las Vegas degree algorithm presented in Section \[degreecdcd\] which computes the degree of each vertex in $B_{cd}L_{cd}$. We obtain a Monte Carlo Algorithm, denoted Algorithm 5’, which computes the degrees in the $BL$ model.
We will need two new boolean variables $collision_B$ and $collision_L$.
First each vertex generates its signature $s$. Then we modify Algorithm 5 as follows.
Collisions must be detected only in slot 1 of Algorithm 5. As is explained in Remark \[listen\], in this slot, active vertices which do not beep listen. Thus the instruction in slot 1:
[**if**]{} $candidate$ [**then**]{}\
beep\
becomes:
[**if**]{} $candidate$ [**then**]{} $EmulateB_{cd}inBL(s,collision_B)$\
$EmulateL_{cd}inBL(k,collision_L);$
and the first instructions in slot 2:
$ic:=$internal collision;\
$pc:=$ peripheral collision;\
become:
$ic := collision_B$;\
$pc := collision_L$.
The other instructions in the algorithm are not changed.
Finally, Algorithm 5’ (a degree computation algorithm in $BL$) is the concatenation of $gen(k)$ and Algorithm 5 modified as explained above. Then, we deduce from Lemma 6.1 the following results:
For any graph $G$ of size $n$ and any $0<\varepsilon<1$:
- if $k=\lceil \log_2\left(\frac n\varepsilon\right) \rceil$, Algorithm 5’ computes the degrees in $G$ in $O\left((\log n + \Delta^2)(\log(\frac{n}{\varepsilon}))\right)$, and the result is correct with probability at least $1-\varepsilon$.
- If $k=\lceil \log_2\left(\frac 1 \varepsilon\right) \rceil$, each vertex $v$ computes its degree in $O\left((\log n + \Delta^2
)(\log(\frac1{\varepsilon}))\right)$, and the result is correct with probability at least $1-\varepsilon$.
- If $k = \lceil 2\log_2(n) \rceil$, Algorithm 5’ computes the degrees in $G$ in $O\left((\log n+\Delta^2)\log n\right)$, and the result is correct with probability $1-o\left(\frac 1 n \right)$.
The same transformation can be done for the colouring or the $2$-hop-colouring algorithms.
Conclusion
==========
We present in this paper algorithms which detect collisions in the weakest beeping model with a logarithmic complexity. Then we consider more powerful beeping models which enable simple and efficient solutions to the colouring problem, to the 2-hop-colouring problem and to the degree computation. Finally, thanks to emulation procedures based on collision detection we give solutions to these problems in the weakest beeping model having a time complexity increased by a logarithmic factor.
[^1]: Let $G=(V,E)$ be a graph. An independent set of $G$ is a subset $I$ of $V$ such that no two members of $I$ are adjacent. An independent set $I$ is maximal, denoted MIS, if any vertex of $G$ is in $I$ or adjacent to a vertex of $I.$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
=0.6 cm
[**Abstract**]{}
We have investigated the strong gravitational lensing for the photons coupled to Weyl tensor in a Schwarzschild black hole spacetime. We find that in the four-dimensional black hole spacetime the equation of motion of the photons depends not only on the coupling between photon and Weyl tensor, but also on the polarization direction of the photons. It is quite different from that in the case of the usual photon without coupling to Weyl tensor in which the equation of motion is independent of the polarization of the photon. Moreover, we find that the coupling and the polarization direction modify the properties of the photon sphere, the deflection angle, the coefficients in strong field lensing, and the observational gravitational lensing variables. Combining with the supermassive central object in our Galaxy, we estimated three observables in the strong gravitational lensing for the photons coupled to Weyl tensor.
author:
- 'Songbai Chen[^1], Jiliang Jing [^2]'
title: '**Strong gravitational lensing for the photons coupled to Weyl tensor in a Schwarzschild black hole spacetime**'
---
=0.8 cm
Introduction
============
The interaction between the electromagnetic and gravitational fields should be important in physics because the electromagnetic force and gravity are two kinds of fundamental forces in nature. In the Lagrangian of the standard Einstein-Maxwell theory, only the quadratic term of Maxwell tensor is related directly to electromagnetic and gravitational fields, which can also be looked as the interaction between Maxwell field and the spacetime metric tensor. However, the interactions between electromagnetic field and curvature tensor are not included in this electromagnetic theory. Drummond *et al.*[@Drummond] found that such kind of the couplings could be appeared naturally in quantum electrodynamics with the photon effective action originating from one-loop vacuum polarization on a curved background spacetime. In this effective field theory, all of the coupling constants are very small and their values are of the order of the square of the Compton wave length of the electron $\lambda_e$ since the coupling between electromagnetic field and curvature tensor is just a quantum phenomenon in this case. However, the models with arbitrary coupling constant have been investigated widely in refs. [@Turner; @Ni; @Solanki; @Dereli1; @Balakin; @Hehl; @Bamba] for some physical motivation. In order to explain the power-law inflation in the early Universe and the large scale magnetic fields observed in clusters of galaxies, Turner *et al* [@Turner; @Bamba] reconsidered Drummond’s model [@Drummond] with the arbitrary coupling constant and found some interesting effects on the electromagnetic fluctuations. Ni [@Ni] proposed a classical generalized electromagnetic model in which electromagnetic field is interacted with curvature tensor through some special coupled terms. Considering that the coupling between electromagnetic field and curvature tensor should be emerged reasonably in the region near the classical supermassive compact objects at the center of galaxies due to their strong gravity and high mass density, Ni’s model has been investigated widely in astrophysics [@Solanki; @Dereli1] and black hole physics [@Balakin; @Hehl]. These investigation show that the coupling term modifies the equations of motion both for the electromagnetic and gravitational fields, which could yield time delays in the arrival of gravitational and electromagnetic waves.
Weyl tensor is an important tensor in general relativity, which describes a type of gravitational distortion in the spacetime. The couplings between Maxwell field and Weyl tensor can be treated actually as a special kind of interactions between electromagnetic field and curvature tensors since Weyl tensor is a function of Riemann tensor $R_{\mu\nu\rho\sigma}$, the Ricci tensor $R_{\mu\nu}$, and the Ricci scalar $R$. The electromagnetic theory with Weyl corrections have been investigated extensively in the literature [@Weyl1; @Wu2011; @Ma2011; @Momeni; @Roychowdhury; @zhao2013]. It is shown that the couplings with Weyl tensor change the universal relation with the $U(1)$ central charge in the holographic conductivity in the background of anti-de Sitter spacetime [@Weyl1] and modify the properties of the holographic superconductor including the critical temperature and the order of the phase transition [@Wu2011; @Ma2011; @Momeni; @Roychowdhury; @zhao2013]. Moreover, we find that with these couplings the dynamical evolution and Hawking radiation of electromagnetic field in the black hole spacetime depend on the coupling parameter and the parity of the field [@sb2013].
According to general theory of relativity, photons would be deviated from their straight path when they pass close to a compact and massive body, and the corresponding effects are called as gravitational lensing [@Einstein; @schneider; @Darwin]. The images of the source stars in the gravitational lensing carry the information about the source stars and gravitational lens itself, which could help us to identify the compact astrophysical objects in the Universe and examine further alternative theories of gravity in their strong field regime. Many investigations [@Einstein; @schneider; @Darwin; @Vir; @Vir1; @Vir11; @Fritt; @Bozza1; @Eirc1; @whisk; @Bozza2; @Bozza3; @Gyulchev; @sbnonk; @Bhad1; @TSa1; @AnAv; @gr1; @Kraniotis; @schen; @JH] have studied the propagation of the free photons in the background spacetimes and probed the effects of the spacetime parameters on the gravitational lensing. In general, gravitational lensing should be depended both on the properties of the background spacetime and the behavior of the photon itself. Eiroa [@Eirc2] has studied the behavior of photons in Born-Infeld electrodynamics and found that in this case photons did not follow geodesics of the metric but they followed geodesics of an effective metric depending on the Born-Infeld coupling, which modifies the properties of the gravitational lensing. Therefore, it is of interest to probe how the interaction between photon and spacetime curvature tensor affect the gravitational lensing. From the previous discussion, we know that the couplings between Maxwell tensor and curvature tensor will change the behavior of electromagnetic field in the background spacetime. It is well known that light is actually a kind of electromagnetic wave, which means that these couplings will modify the propagation of photons in the spacetime and lead to some particular phenomena of gravitational lensing. The deflection angle of the photons coupled to Riemann tensor in the weak field limit have been studied in [@Drummond]. Since the weak field limit takes only the first order deviation from Minkowski spacetime and it is valid only in the region far from the black hole, it is necessary to investigate further the gravity lensing in the strong field region near the black hole because that it starts from complete capture of the coupled photon and dominates the leading order in the divergence of the deflection angle. Moreover, in order to probe the universal features of the deflection angles of the photons coupled to spacetime curvature tensors, we here will study the strong gravitational lensing in the Schwarzschild black hole as the photons couple to Weyl tensor, and then explore the effect of these couplings on the deflection angle and the observables in the strong field limit.
The plan of our paper is organized as follows: In Sec.II, we derive the equations of motion for the photons coupled to Weyl tensor in the four-dimensional static and spherical symmetric spacetime, which can be obtained from the Maxwell equation with Weyl corrections by the geometric optics approximation [@Drummond; @Daniels; @Caip; @Cho1; @Lorenci]. In Sec.III, we will study the effects of these coupling on the photon sphere radius and the deflection angles of light ray in a Schwarzschild black hole spacetime. In Sec.IV, we will study the physical properties of the strong gravitational lensing for the coupled photons and then probe the effects of the coupling constant on the coefficients and the observables of the gravitational lensing in the strong field limit. We end the paper with a summary.
Equation of motion for the photons coupled to Weyl tensor
=========================================================
In this section, we will derive the equations of motion for the photons coupled to Weyl tensor in the four-dimensional static and spherical symmetric spacetime by the geometric optics approximation [@Drummond; @Daniels; @Caip; @Cho1; @Lorenci]. We begin with the action of the electromagnetic field coupled to Weyl tensor in the curved spacetime, which can be expressed as [@Weyl1] $$\begin{aligned}
S=\int d^4x \sqrt{-g}\bigg[\frac{R}{16\pi
G}-\frac{1}{4}\bigg(F_{\mu\nu}F^{\mu\nu}-4\alpha
C^{\mu\nu\rho\sigma}F_{\mu\nu}F_{\rho\sigma}\bigg)\bigg].\label{acts}\end{aligned}$$ Here $C_{\mu\nu\rho\sigma}$ is the Weyl tensor, which is defined as $$\begin{aligned}
C_{\mu\nu\rho\sigma}=R_{\mu\nu\rho\sigma}-\frac{2}{n-2}(
g_{\mu[\rho}R_{\sigma]\nu}-g_{\nu[\rho}R_{\sigma]\mu})+\frac{2}{(n-1)(n-2)}R
g_{\mu[\rho}g_{\sigma]\nu}.\end{aligned}$$ Here $n$ and $g_{\mu\nu}$ are the dimension and metric of the spacetime. The brackets around indices refers to the antisymmetric part. $F_{\mu\nu}$ is the usual electromagnetic tensor with a form $F_{\mu\nu}=A_{\nu;\mu}-A_{\mu;\nu}$. The coupling constant $\alpha$ has the dimension of length-squared. Varying the action (\[acts\]) with respect to $A_{\mu}$, one can obtain easily the corrected Maxwell equation $$\begin{aligned}
\nabla_{\mu}\bigg(F^{\mu\nu}-4\alpha
C^{\mu\nu\rho\sigma}F_{\rho\sigma}\bigg)=0.\label{WE}\end{aligned}$$ In order to get the equation of motion of the photons from the above corrected Maxwell equation (\[WE\]), one can adopt to the geometric optics approximation in which the photon wavelength $\lambda$ is much smaller than a typical curvature scale $L$, but is larger than the electron Compton wavelength $\lambda_e$, i.e., $\lambda_e<\lambda<L$. This ensures that the change of the background gravitational and electromagnetic fields with the typical curvature scale can be neglected in the process of photon propagation [@Drummond; @Daniels; @Caip; @Cho1; @Lorenci]. With this approximation, the electromagnetic field strength can be written as $$\begin{aligned}
F_{\mu\nu}=f_{\mu\nu}e^{i\theta},\label{ef1}\end{aligned}$$ where $f_{\mu\nu}$ is a slowly varying amplitude and $\theta$ is a rapidly varying phase. This means that the derivative term $f_{\mu\nu;\lambda}$ is not dominated and can be neglected in this approximation. The wave vector is $k_{\mu}=\partial_{\mu}\theta$, which can be treated as the usual photon momentum in the theory of quantum particle. The amplitude $f_{\mu\nu}$ is constrained by the Bianchi identity $$\begin{aligned}
D_{\lambda} F_{\mu\nu}+D_{\mu} F_{\nu\lambda}+D_{\nu} F_{\lambda\mu}=0,\end{aligned}$$ which leads to $$\begin{aligned}
k_{\lambda}f_{\mu\nu}+k_{\mu} f_{\nu\lambda}+k_{\nu} f_{\lambda\mu}=0.\end{aligned}$$ This means that the amplitude $f_{\mu\nu}$ can be written as $$\begin{aligned}
f_{\mu\nu}=k_{\mu}a_{\nu}-k_{\nu}a_{\mu},\label{ef2}\end{aligned}$$ where $a_{\mu}$ is the polarization vector satisfying the condition that $k_{\mu}a^{\mu}=0$. The amplitude $f_{\mu\nu}$ has just three independent components. Inserting Eqs.(\[ef1\]) and (\[ef2\]) into Eq.(\[WE\]), one can obtain the equation of motion of photon coupled to Weyl tensor $$\begin{aligned}
k_{\mu}k^{\mu}a^{\nu}+8\alpha
C^{\mu\nu\rho\sigma}k_{\sigma}k_{\mu}a_{\rho}=0.\label{WE2}\end{aligned}$$ Obviously, the coupling term with Weyl tensor affects the propagation of the coupled photon in the background spacetime.
Let us now to consider a four-dimensional static and spherical symmetric black hole spacetime, $$\begin{aligned}
ds^2&=&-fdt^2+\frac{1}{f}dr^2+r^2
d\theta^2+r^2\sin^2{\theta}d\phi^2,\label{m1}\end{aligned}$$ where the metric coefficient $f$ is a function of polar coordinate $r$. In order to introduce a local set of orthonormal frames, one can use the vierbein fields defined by $$\begin{aligned}
g_{\mu\nu}=\eta_{ab}e^a_{\mu}e^b_{\nu},\end{aligned}$$ where $\eta_{ab}$ is the Minkowski metric and the vierbeins $$\begin{aligned}
e^a_{\mu}=(\sqrt{f},\;\frac{1}{\sqrt{f}},\;r,\;r\sin\theta),\end{aligned}$$ with the inverse $$\begin{aligned}
e^{\mu}_a=(\frac{1}{\sqrt{f}},\;\sqrt{f},\;\frac{1}{r},\;\frac{1}{r\sin\theta}).\end{aligned}$$ Defining the notation for the antisymmetric combination of vierbeins [@Drummond; @Daniels] $$\begin{aligned}
U^{ab}_{\mu\nu}=e^a_{\mu}e^b_{\nu}-e^a_{\nu}e^b_{\mu},\end{aligned}$$ we find that the complete Weyl tensor can be rewritten in the following form $$\begin{aligned}
C_{\mu\nu\rho\sigma}=\mathcal{A}\bigg(2U^{01}_{\mu\nu}U^{01}_{\rho\sigma}-
U^{02}_{\mu\nu}U^{02}_{\rho\sigma}-U^{03}_{\mu\nu}U^{03}_{\rho\sigma}
+U^{12}_{\mu\nu}U^{12}_{\rho\sigma}+U^{13}_{\mu\nu}U^{13}_{\rho\sigma}-
2U^{23}_{\mu\nu}U^{23}_{\rho\sigma}\bigg),\end{aligned}$$ with $$\begin{aligned}
\mathcal{A}=-\frac{1}{12r^2}\bigg[r^2f''-2f'r+2f-2\bigg].\end{aligned}$$ In order to obtain the equation of motion for the coupled photon propagation, one can introduce three linear combinations of momentum components [@Drummond; @Daniels] $$\begin{aligned}
l_{\nu}=k^{\mu}U^{01}_{\mu\nu},\;\;\;\;\;\;\;\;\;\;
n_{\nu}=k^{\mu}U^{02}_{\mu\nu},\;\;\;\;\;\;\;\;\;\;
m_{\nu}=k^{\mu}U^{23}_{\mu\nu},\label{pvector}\end{aligned}$$ together with the dependent combinations $$\begin{aligned}
&&p_{\nu}=k^{\mu}U^{12}_{\mu\nu}=\frac{1}{k^0}\bigg(k^1n_{\nu}-k^2l_{\nu}\bigg),\nonumber\\
&&r_{\nu}=k^{\mu}U^{03}_{\mu\nu}=\frac{1}{k^2}\bigg(k^0m_{\nu}+k^3l_{\nu}\bigg),\nonumber\\
&&q_{\nu}=k^{\mu}U^{13}_{\mu\nu}=\frac{k^1}{k^0}m_{\nu}+
\frac{k^1k^3}{k^2k^0}n_{\nu}-\frac{k^3}{k^0}l_{\nu}.\label{vect3}\end{aligned}$$ The vectors $l_{\nu}$, $n_{\nu}$, $m_{\nu}$ are independent and orthogonal to the wave vector $k_{\nu}$. Using the relationship (\[vect3\]) and contracting the equation (\[WE2\]) with $l_{\nu}$, $n_{\nu}$, $m_{\nu}$, respectively, one can find that the equation of motion of the photon coupling with Weyl tensor can be simplified as a set of equations for three independent polarisation components $a\cdot l$, $a\cdot n$, and $a\cdot m$, $$\begin{aligned}
\bigg(\begin{array}{ccc}
K_{11}&0&0\\
K_{21}&K_{22}&
K_{23}\\
0&0&K_{33}
\end{array}\bigg)
\bigg(\begin{array}{c}
a \cdot l\\
a \cdot n
\\
a \cdot m
\end{array}\bigg)=0,\label{Kk}\end{aligned}$$ with the coefficients $$\begin{aligned}
K_{11}&=&(1+16\alpha \mathcal{A})(g_{00}k^0k^0+g_{11}k^1k^1)+(1-8\alpha \mathcal{A})(g_{22}k^2k^2+g_{33}k^3k^3),\nonumber\\
K_{22}&=&(1-8\alpha \mathcal{A})(g_{00}k^0k^0+g_{11}k^1k^1+g_{22}k^2k^2+g_{33}k^3k^3),
\nonumber\\
K_{21}&=&24\alpha \mathcal{A} \sqrt{g_{11}g_{22}}k^1k^2,\;\;\;\;\;\;\;\;\;\;
K_{23}=-24\alpha \mathcal{A}\sqrt{-g_{00}g_{33}}k^0k^3,\nonumber\\
K_{33}&=&(1-8\alpha \mathcal{A})(g_{00}k^0k^0+g_{11}k^1k^1)+(1+16\alpha \mathcal{A})(g_{22}k^2k^2+g_{33}k^3k^3).\end{aligned}$$ The condition of Eq.(\[Kk\]) with non-zero solution is $K_{11}K_{22}K_{33}=0$. The first root $K_{11}=0$ leads to the modified light cone $$\begin{aligned}
(1+16\alpha \mathcal{A})(g_{00}k^0k^0+g_{11}k^1k^1)+(1-8\alpha \mathcal{A})(g_{22}k^2k^2+g_{33}k^3k^3)=0, \label{Kk31}\end{aligned}$$ which corresponds to the case the polarisation vector $a_{\mu}$ is proportional to $l_{\mu}$ and the strength $f_{\mu\nu}\propto (k_{\mu}l_{\nu}-k_{\nu}l_{\mu})$. The second root $K_{22}=0$ means that $a\cdot l=0$ and $a\cdot m=0$ in Eq.(\[Kk\]), which implies $a_{\mu}=\lambda k_{\mu}$ and then $f_{\mu\nu}$ vanishes [@Drummond]. Thus, this root corresponds to an unphysical polarisation and it should be neglected for general directions of propagation of the coupled photon. The third root is $K_{33}=0$, i.e., $$\begin{aligned}
(1-8\alpha \mathcal{A})(g_{00}k^0k^0+g_{11}k^1k^1)+(1+16\alpha \mathcal{A})(g_{22}k^2k^2+g_{33}k^3k^3)=0,\label{Kk32}\end{aligned}$$ which means that the vector $a_{\mu}=\lambda m_{\mu}$ and the strength $f_{\mu\nu}=\lambda(k_{\mu}m_{\nu}-k_{\nu}m_{\mu})$. Therefore, the light-cone condition depends on not only the coupling between photon and Weyl tensor, but also on the polarizations. Moreover, the effects of Weyl tensor on the photon propagation are different for the coupled photons with different polarizations, which yields a phenomenon of birefringence in the spacetime [@Daniels; @Caip; @Cho1; @Lorenci]. When the coupling constant $\alpha=0$ the light-cone conditions (\[Kk31\]) and (\[Kk32\]) recover to the usual form without Weyl corrections.
Effects of Weyl corrections on the deflection angles for light ray in a Schwarzschild black hole spacetime
==========================================================================================================
In this section, we will study the deflection angles for light ray as photon couples to Weyl tensor in the background of a Schwarzschild black hole, and probe the effects of the coupling and the polarization types on the deflection angle.
For a Schwarzschild black hole spacetime, the metric function is $f=1-\frac{2M}{r}$ and then the light-cone conditions (\[Kk31\]) and (\[Kk32\]) can be expressed as $$\begin{aligned}
(1+\frac{16\alpha M}{r^3})(g_{00}k^0k^0+g_{11}k^1k^1)+(1-\frac{8\alpha M}{r^3} )(g_{22}k^2k^2+g_{33}k^3k^3)=0, \label{Kks31}\end{aligned}$$ for the photon with the polarization along $l_{\mu}$ (PPL) and $$\begin{aligned}
(1-\frac{8\alpha M}{r^3})(g_{00}k^0k^0+g_{11}k^1k^1)+(1+\frac{16\alpha M}{r^3} )(g_{22}k^2k^2+g_{33}k^3k^3)=0,\label{Kks32}\end{aligned}$$ for the photon with the polarization along $m_{\mu}$ (PPM), respectively. The light cone conditions (\[Kks31\]) and (\[Kks32\]) imply that the motion of the coupled photons is non-geodesic in the Schwarzschild metric. Actually, these photons follow null geodesics of the effective metric $\gamma_{\mu\nu}$, i.e., $\gamma^{\mu\nu}k_{\mu}k_{\nu}=0$ [@Breton]. The effective metric for the coupled photon can be expressed as $$\begin{aligned}
ds^2=-A(r)dt^2+B(r)dr^2+C(r)W(r)^{-1}(d\theta^2+\sin^2\theta d\phi^2),\label{l01}\end{aligned}$$ where $A(r)=B(r)^{-1}=1-\frac{2 M}{r}$ and $C(r)=r^2$. The quantity $W(r)$ is $$\begin{aligned}
W(r)=\frac{r^3-8\alpha M}{r^3+16\alpha M},\label{v11}\end{aligned}$$ for PPL and $$\begin{aligned}
W(r)=\frac{r^3+16\alpha M}{r^3-8\alpha M},\label{v12}\end{aligned}$$ for PPM, respectively. For simplicity, we here just consider that both the observer and the source lie in the equatorial plane in the Schwarzschild black hole spacetime and the whole trajectory of the photon is limited on the same plane. With this condition $\theta=\frac{\pi}{2}$, we can obtain the reduced effective metric in the form $$\begin{aligned}
ds^2=-A(r)dt^2+B(r)dr^2+C(r)W(r)^{-1}d\phi^2.\label{l1}\end{aligned}$$ For the photon moving in the equatorial plane ($\theta=\frac{\pi}{2}$), we have $k_2=0$ and $k_{\mu}=(k_0,k_1,0,k_3)$. And then the polarisation vectors $l_{\mu}$ and $m_{\mu}$ can be expressed further as $$\begin{aligned}
l_{\mu}=(-k^1,k^0,0,0),\;\;\;\;\;\; m_{\mu}=(0,0,-k^3,0).\end{aligned}$$ This means that $m_{\mu}$ is the polarization orthogonal to the plane of motion and $l_{\mu}$ lies on the plane of motion in this case.
In the four dimensional static spacetime (\[l01\]) with cyclic coordinates $t$ and $\phi$, it is easy to obtain two constants of motion of the geodesics $$\begin{aligned}
E=-g_{00}\dot{x}^{\mu}=A(r)\dot{t},\;\;\;\;\;\;\;\;\;
L=g_{33}\dot{x}^{\mu}=C(r)W(r)^{-1}\dot{\phi}.\end{aligned}$$ where a dot represents a derivative with respect to affine parameter $\lambda$ along the geodesics. $E$ and $L$ correspond to the energy and angular momentum of the coupled photon, respectively. Making use of these two constants and $k^{\mu}=\frac{dx^{\mu}}{d\lambda}$, one can find that the equations of motion of coupled photon can be expressed further as $$\begin{aligned}
\bigg(\frac{dr}{d\lambda}\bigg)^2
=\frac{1}{B(r)}\bigg[\frac{E^2}{A(r)}-W(r)\frac{L^2}{C(r)}\bigg].\label{v1}\end{aligned}$$ Comparing with Eqs.(8), (16) and (17) in ref.[@sb2013], we find that the equations of motion (\[v1\]) correspond actually to the radial equations of the electromagnetic perturbation with the even parity and the odd parity in the geometric optics limit, respectively. For a coupled photon coming from infinite, the form of the deflection angle in the Schwarzschild black hole spacetime is similar to that in the case without coupling [@Vir1] $$\begin{aligned}
\alpha(r_{0})=I(r_{0})-\pi,\end{aligned}$$ where $r_0$ is the closest approach distance. However, $I(r_{0})$ depends on the polarization directions of the photon in this case, i.e., $$\begin{aligned}
I(r_0)=2\int^{\infty}_{r_0}\frac{dr}{\sqrt{A(r) C(r)/ W(r)}
\sqrt{\frac{C(r)A(r_0)W(r_0)}{A(r)W(r)C(r_0)}-1}},\label{int1}\end{aligned}$$ which means that the deflection angle of PPL is different from that of PPM. Obviously, as $\alpha\rightarrow 0$, one can find that $W(r)\rightarrow 1$ and the deflection angle can be reduced to that of usual photon [@Vir1].
From Eqs.(\[v11\]) and (\[v12\]), we find that there is a singularity in Eq.(\[v1\]) at $r_{sin}=(-16\alpha M)^{1/3}$ for PPL and at $r_{sin}=(8\alpha M)^{1/3}$ for PPM because the quantity $\frac{dr}{d\lambda}$ in Eq.(\[v1\]) is divergent at this surface. The position of the singularity depends on the coupling parameter and the polarizations of the coupled photon. Considering that a photon should propagate continuously in the region outside the event horizon, the coupling constant $\alpha$ must satisfy $r^3_H+16\alpha M > 0$ (i.e., $\alpha>\alpha_{c1}=-M^2/2$) for PPL, and satisfy $r^3_H-8\alpha M > 0$ (i.e., $\alpha<\alpha_{c2}=M^2$) for PPM. With this constraint, the singularity $r_{sin}$ lies inside the event horizon $r_H$ and it does not affect the propagation of the photon. And then, we can use the usual methods [@Vir; @Bozza2] to study the deflection angles for the photon coupled to Weyl tensor in the background of a Schwarzschild black hole.
Using the photon sphere equation given in [@Vir], one can obtain that in a Schwarzschild spacetime the impact parameter and the equation of circular orbits of the coupled photon are $$\begin{aligned}
&&u=\sqrt{\frac{C(r)}{A(r)W(r)}},\label{u}\\
&&W(r)[A'(r)C(r)-A(r)C'(r)]+A(r)C(r)W'(r)=0.\label{sp}\end{aligned}$$ Here we set $E=1$. As the coupling parameter $\alpha\rightarrow 0$, we find that the function $W\rightarrow 1$, which results in that the impact parameter and the equation of circular photon orbits for PPL are the same as those for PPM. This means that gravitational lensing is independent of the polarization directions of the photon in the case without the coupling. Substituting Eqs.(\[v11\]) and (\[v12\]) into equation (\[sp\]), we can obtain the equation of circular photon orbits $$\begin{aligned}
2(r^3+16\alpha M)(r^3-8\alpha M)(r-3M)\pm72\alpha M r^3(r-2M)=0.\label{sppl}\end{aligned}$$ Here the signs “-" and “+" in the last term correspond to the cases of PPL and PPM, respectively. The biggest real roots of equations (\[sppl\]) outside the horizon can be defined as the photon sphere radius of the coupled photons.
{width="7cm"}{width="7cm"}
Obviously, the complex dependence of the equation (\[sppl\]) on the coupling parameter $\alpha$ yields that we can not get an analytical form for the photon sphere radius for the coupled photons. With the help of the numerical method, in Fig.1 we plot the dependence of the photon sphere radius $r_{ps}$ for the coupled photons on the coupling parameter $\alpha$. It tells us that with increase of $\alpha$, $r_{ps}$ increases for PPL and decreases for PPM in the Schwarzschild black hole spacetime. This implies that the properties of gravitational lensing for PPL is different from that for PPM. In other words, gravitational lensing for the photon coupled to Weyl tensor depends not only on the coupling constant $\alpha$, but also on the polarization directions of the photon. Moreover, we also find that as the couple constant $\alpha$ tends to the critical value $\alpha_{c1}$ or $\alpha_{c2}$, the photon sphere of the coupled photon is overlapped with the event horizon of the black hole.
Effects of Weyl Corrections on strong gravitational lensing in a Schwarzschild black hole spacetime
===================================================================================================
In this section, we will study the effects of the coupling with Weyl tensor on the coefficients and the observables of the gravitational lensing in the strong field limit. Following the evaluation method proposed by Bozza [@Bozza2], we can define a variable $$\begin{aligned}
z=1-\frac{r_0}{r},\end{aligned}$$ and rewrite the integral (\[int1\]) as $$\begin{aligned}
I(r_0)=\int^{1}_{0}R(z,r_0)F(z,r_0)dz,\label{in1}\end{aligned}$$ with $$\begin{aligned}
R(z,r_0)&=&2\frac{W(r)r^2\sqrt{
C(r_0)}}{r_0C(r)}=2W(z,r_0),\end{aligned}$$ $$\begin{aligned}
F(z,r_0)&=&\frac{1}{\sqrt{A(r_0)W(r_0)-\frac{A(z,r_0)W(z,r_0)C(r_0)}{C(z,r_0)}}}.\end{aligned}$$ The function $R(z, r_0)$ is regular for all values of $z
$ and $r_0$, but the function $F(z, \rho_s)$ diverges as $z$ tends to zero. This is similar to that in the Schwarzschild black hole spacetime without the coupling. Thus, one can split the integral (\[in1\]) into the divergent part $I_D(r_0)$ and the regular one $I_R(r_0)$ $$\begin{aligned}
I_D(r_0)&=&\int^{1}_{0}R(0,r_{ps})F_0(z,r_0)dz, \nonumber\\
I_R(r_0)&=&\int^{1}_{0}[R(z,r_0)F(z,r_0)-R(0,r_{ps})F_0(z,r_0)]dz
\label{intbr},\end{aligned}$$ where the function $F_0(z,r_{0})$ is obtained by expanding the argument of the square root in $F(z,r_{0})$ to the second order in $z$, $$\begin{aligned}
F_0(z,r_{0})=\frac{1}{\sqrt{p(r_{0})z+q(r_0)z^2}},\end{aligned}$$ with $$\begin{aligned}
p(r_0)&=&-\frac{r_0}{C(r_0)}\bigg\{W(r_0)[A'(r_0)C(r_0)-A(r_0)C'(r_0)]
+A(r_0)C(r_0)W'(r_0)\bigg\}, \nonumber\\
q(r_0)&=&\frac{r_0}{2C(r_0)}\bigg\{2\bigg[C(r_0)-r_0C'(r_0)\bigg]\bigg[A(r_0)W(r_0)C'(r_0)
-C(r_0)\bigg(A(r_0)W(r_0)\bigg)'\bigg] \nonumber\\
&+&r_0C(r_0)\bigg[A(r_0)W(r_0)C(r_0)''
-C(r_0)\bigg(A(r_0)W(r_0)\bigg)''\bigg]\bigg\}.\label{pq}\end{aligned}$$ If $r_0$ tends to the photon sphere radius $r_{ps}$, one can find that the coefficient $p(r_{0})$ approaches zero and then the integral (\[in1\]) diverges logarithmically since the leading term of the divergence in $F(z,r_{0})$ is $z^{-1}$. This means that when the photon is close to the photon sphere, the deflection angle can be approximated as [@Bozza2] $$\begin{aligned}
\alpha(\theta)=-\bar{a}\log{\bigg[\frac{\theta
D_{OL}}{u(r_{ps})}-1\bigg]}+\bar{b}+O[u-u(r_{ps})], \label{alf1}\end{aligned}$$ with $$\begin{aligned}
&\bar{a}&=\frac{R(0,r_{ps})}{2\sqrt{q(r_{ps})}},\;\;\;\;\;\;\;\;\; b_R=I_R(r_{ps}), \nonumber\\
&\bar{b}&=
-\pi+b_R+\bar{a}\log{\bigg[\frac{2r^2_{hs}u(r_{ps})''}{u(r_{ps})}\bigg]}.
\label{coa1}\end{aligned}$$ Here $D_{OL}$ is the distance between observer and gravitational lens. Substituting Eqs.(\[v11\]) and (\[v12\]) into equations (\[coa1\]), we can obtain the coefficients $\bar{a}$ and $\bar{b}$ in the strong gravitational lensing formula (\[alf1\]).
{width="7cm"}{width="7cm"}
{width="7cm"}{width="7cm"}
{width="7cm"}
In Figs.2-3, we plotted numerically the dependence of the coefficients ( $\bar{a}$ and $\bar{b}$ ) on the parameter $\alpha$. It is shown that with the increase of $\alpha$ the coefficient $\bar{a}$ decreases monotonously for PPL and increases for PPM. However, the change of $\bar{b}$ with $\alpha$ is more complicated. For PPL, the coefficient $\bar{b}$ first increases up to its maximum with $\alpha$ and then decreases down to its minimum with the further increase of $\alpha$ ; after that, it increases with $\alpha$ again. The variety of $\bar{b}$ with $\alpha$ for PPM is converse to that for PPL. The maximum and minimum of $\bar{b}$ and their corresponding values of $\alpha$ depend on the polarization of the coupled photons. Moreover, one can find that the coefficient $\bar{a}$ diverges as the couple constant $\alpha$ tends to the critical value $\alpha_{c1}$ or $\alpha_{c2}$, which implies that the deflection angle in the strong deflection limit (\[alf1\]) is not valid in the regime $\alpha<\alpha_{c1}$ for PPL and $\alpha>\alpha_{c2}$ for PPM, which is consistent with the previous discussion. Therefore, the presence of the coupling makes the change of the coefficients $\bar{a}$ and $\bar{b}$ more complicated because the effects of the coupling depend not only on the values of the parameter $\alpha$, but also on the direction of polarization of the coupled photon. Furthermore, we plotted in Fig.4 the change of the deflection angles evaluated at $u=u_{ps}+ 0.003$ with $\alpha$ for PPL and PPM, respectively. We find that the deflection angles in the strong field limit have similar behaviors of the coefficient $\bar{a}$, which means that the deflection angles of the light rays are dominated by the logarithmic term in this case.
We are now in the position to study the effect of the coupling constant $\alpha$ and the direction of polarization of the coupled photon on the observational gravitational lensing variables in the strong field limit. If the source and observer are far enough from the lens, one can find that the lens equation can be simplified further as [@Bozza3] $$\begin{aligned}
\gamma=\frac{D_{OL}+D_{LS}}{D_{LS}}\theta-\alpha(\theta) \; \text{mod}
\;2\pi,\end{aligned}$$ where $\gamma$ is the angle between the direction of the source and the optical axis. $D_{LS}$ and $D_{OL}$ are the lens-source distance and the observer-lens distance, respectively. The angle $\theta=u/D_{OL}$ is the angular separation between the lens and the image. As in ref.[@Bozza3], we here focus only on the simplest case in which the source, lens and observer are highly aligned so that the angular separation between the lens and the $n-$th relativistic image can be approximated as $$\begin{aligned}
\theta_n\simeq\theta^0_n\bigg(1-\frac{u_{ps}e_n(D_{OL}+D_{LS})}{\bar{a}D_{OL}D_{LS}}\bigg),\end{aligned}$$ with $$\begin{aligned}
\theta^0_n=\frac{u_{ps}}{D_{OL}}(1+e_n),\;\;\;\;\;\;e_{n}=e^{\frac{\bar{b}+|\gamma|-2\pi
n}{\bar{a}}},\label{st1}\end{aligned}$$ where $n$ is an integer and $\theta^0_n$ is the image position corresponding to $\alpha=2n\pi$. In the limit $n\rightarrow \infty$, we have $e_n\rightarrow 0$, which means that the relationship among the asymptotic position of a set of images $\theta_{\infty}$, the observer-lens distance $D_{OL}$ and the minimum impact parameter $u_{ps}$ can be rewritten as a simpler form $$\begin{aligned}
u_{ps}=D_{OL}\theta_{\infty}.\label{ups}\end{aligned}$$ In order to get the coefficients $\bar{a}$ and $\bar{b}$, we need at least another two observations. As in refs.[@Bozza2; @Bozza3], we consider a perfect situation in which only the outermost image $\theta_1$ is separated as a single image and all the remaining ones are packed together at $\theta_{\infty}$. In this situation the angular separation $s$ and the relative magnitudes $r_m$ between the first image and other ones can be simplified further as [@Bozza2; @Bozza3] $$\begin{aligned}
s&=&\theta_1-\theta_{\infty}=
\theta_{\infty}e^{\frac{\bar{b}-2\pi}{\bar{a}}},\nonumber\\
r_m&=&2.5\log{\mathcal{R}_0}=2.5\log{\bigg(\frac{\mu_1}{\sum^{\infty}_{n=2}\mu_n}
\bigg)}
=\frac{5\pi}{\bar{a}}\log{e},\label{sR}\end{aligned}$$ where $\mathcal{R}_0$ represents the ratio of the flux from the first image and those from the all other images. Through measuring these three simple observations $s$, $r_m$, and $\theta_{\infty}$, one can estimate the coefficients $\bar{a}$, $\bar{b}$ and the minimum impact parameter $u_{ps}$ in the strong deflection limit. Comparing their values with those predicted by the coupling theoretical model, we can extract the characteristics information stored in the strong gravitational lensing and examine whether this coupling exists in the Universe.
{width="6cm"} {width="6cm"}
{width="6cm"}{width="6cm"}
{width="6cm"}{width="6cm"}
The mass of the central object of our Galaxy is evaluated to be $4.4\times 10^6M_{\odot}$ and its distance from the earth is around $8.5kpc$ [@grf], which means the ratio $GM/D_{OL}
\approx2.4734\times10^{-11}$ . Combing with Eqs. (\[coa1\]), (\[ups\]) and (\[sR\]), we can estimate the values of the coefficients and observables in strong gravitational lensing as the photon couples to Weyl tensor in a Schwarzschild black hole spacetime. We present the dependence of these observables on the coupling constant $\alpha$ in Figs.(5)-(7). With the increase of the coupling constant $\alpha$, both the angular position of the relativistic images $\theta_\infty$ and the relative magnitudes $r_m$ increase for PPL and decrease for PPM. However, the variety of the angular separation $s$ with $\alpha$ is converse to the varieties of $\theta_\infty$ and $r_m$ with $\alpha$. Comparing with those for the photon without coupling to Weyl tensor, one can find that the behaviors of three observables in this case become more complicated. The main reason is that the coupling between the photon and Weyl tensor changes the equation of motion of the photon and makes the propagation of the light ray more complicated.
summary
=======
In this paper, we have investigated the equation of motion of the photon coupled to Weyl tensor and the corresponding strong gravitational lensing in a Schwarzschild black hole spacetime. We find that the coupling constant $\alpha$ and the polarization direction imprint in the propagation of the coupled photons and bring some new features for the physical quantities including the photon sphere radius, the deflection angle, the coefficients ($\bar{a}$ and $\bar{b}$) in the strong field lensing formulas, and the observational gravitational lensing variables. With increase of $\alpha$, the photon sphere radius $r_{ps}$ increases for PPL and decreases for the photon PPM in the Schwarzschild black hole spacetime. In the strong gravity limit, the coefficient $\bar{a}$ decreases with $\alpha$ for PPL and increases for PPM. The change of $\bar{b}$ with $\alpha$ is more complicated. Moreover, we find that the coefficient $\bar{a}$ diverges as the couple constant $\alpha$ tends to the critical value $\alpha_{c1}$ or $\alpha_{c2}$. Combining with the supermassive central object in our Galaxy, we estimated three observables in the strong gravitational lensing for the photons coupled to Weyl tensor. It is shown that with the increase of the coupling constant $\alpha$, both the angular position of the relativistic images $\theta_\infty$ and the relative magnitudes $r_m$ increase for PPL and decrease for PPM. However, the variety of the angular separation $s$ with $\alpha$ is converse to the varieties of $\theta_\infty$ and $r_m$ with $\alpha$ for two different kinds of coupled photons with different polarizations. These indicate that the gravitational lensing depends not only on the properties of background black hole spacetime, but also on the polarization of the coupled photon. It would be of interest to generalize our study to other black hole spacetimes, such as rotating black holes etc. Work in this direction will be reported in the future.
**Acknowledgments**
===================
This work was partially supported by the National Natural Science Foundation of China under Grant No.11275065, the construct program of key disciplines in Hunan Province and the Open Project Program of State Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy of Sciences, China (No.Y5KF161CJ1). J. Jing’s work was partially supported by the National Natural Science Foundation of China under Grant No.11475061.
[99]{} =0.5 cm
I. T. Drummond and S. J. Hathrell, *QED vacuum polarization in a background gravitational field and its effect on the velocity of photons*, Phys. Rev. D [**22**]{}, 343 (1980).
M. S. Turner and L. M. Widrow, *Inflation-produced, large-scale magnetic fields*,Phys. Rev. D [**37**]{} 2743 (1988);
F. D. Mazzitelli and F. M. Spedalieri, *Scalar electrodynamics and primordial magnetic fields*, Phys. Rev. D [**52**]{} 6694 (1995);
G. Lambiase and A. R. Prasanna, *Gauge invariant wave equations in curved space-times and primordial magnetic fields*, Phys. Rev. D [**70**]{}, 063502 (2004);
A. Raya, J. E. M. Aguilar and M. Bellini, *Gravitoelectromagnetic inflation from a 5D vacuum state: a new formalism*, Phys. Lett. B [**638**]{}, 314 (2006);
L. Campanelli, P. Cea, G. L. Fogli and L. Tedesco, *Inflation-produced magnetic fields in $R^nF^2$ and $IF^2$ models*, Phys. Rev. D [**77**]{}, 123002 (2008).
K. Bamba and S. D. Odintsov,*Inflation and late-time cosmic acceleration in non-minimal Maxwell-$F(R)$ gravity and the generation of large-scale magnetic fields*, JCAP [**0804**]{}, 024, (2008);
K. T. Kim, P. P. Kronberg, P. E. Dewdney and T. L. Landecker,*The halo and magnetic field of the Coma cluster of galaxies*, Astrophys. J. [**355**]{} 29 (1990);
K. T. Kim, P. C. Tribble and P. P. Kronberg, *Detection of excess rotation measure due to intracluster magnetic fields in clusters of galaxies*, Astrophys. J. [**379**]{} 80 (1991);
T. E. Clarke, P. P. Kronberg and H. Boehringer, *A New Radio - X-Ray Probe of Galaxy Cluster Magnetic Fields*, Astrophys. J. [**547**]{}, L111 ( 2001).
W. T. Ni, *Equivalence Principles and Electromagnetism*,Phys. Rev. Lett. [**38**]{}, 301 (1977);
W. T. Ni, *Equivalence principles and precision experiments. In Precision Measurements and Fundamental Constants II*, edited by B. N. Taylor and W. D. Phillips, U.S. National Bureau of Standards Publication 617 (U.S. GPO, Washington D.C., 1984), 647¨C651.
S. K. Solanki, O. Preuss, M. P. Haugan, A. Gandorfer, H. P. Povel,P. Steiner, K. Stucki, P. N. Bernasconi, and D. Soltau, *Solar constraints on new couplings between electromagnetism and gravity*, Phys. Rev. D [**69**]{}, 062001 (2004);
O. Preuss, M. P. Haugan, S. K. Solanki, and S. Jordan, *An astronomical search for evidence of new physics: Limits on gravity-induced birefringence from the magnetic white dwarf RE J0317-853*, Phys. Rev. D [**70**]{}, 067101 (2004);
Y. Itin and F. W. Hehl, *Maxwell¡¯s field coupled nonminimally to quadratic torsion: Axion and birefringence*, Phys. Rev. D [**68**]{}, 127701 (2003).
T. Dereli and O. Sert, *Non-minimal $\ln(R)F^2$ Couplings of Electromagnetic Fields to Gravity: Static, Spherically Symmetric Solutions*, Eur. Phys. J. C [**71**]{}, 1589 (2011).
A. B. Balakin and J. P. S. Lemos, *Non-minimal coupling for the gravitational and electromagnetic fields: A general system of equations*, Class. Quantum Grav. [**22**]{}, 1867 (2005);
A. B. Balakin, V. V. Bochkarev and J. P. S. Lemos, *Non-minimal coupling for the gravitational and electromagnetic fields: black hole solutions and solitons*, Phys. Rev. D [**77**]{}, 084013 (2008).
F. W. Hehl and Y. N. Obukhov, *How does the electromagnetic field couple to gravity, in particular to metric, nonmetricity, torsion, and curvature?*, Lect. Notes Phys. [**562**]{}, 479 (2001).
A. Ritz and J. Ward, *Weyl corrections to holographic conductivity*, Phys. Rev. D [**79**]{} 066003 (2009).
J. P. Wu, Y. Cao, X. M. Kuang, and W. J. Li, *The 3+1 holographic superconductor with Weyl corrections*, Phys. Lett. B [**697**]{}, 153 (2011). D. Z. Ma, Y. Cao, and J. P. Wu, *The Stückelberg Holographic Superconductors with Weyl corrections*, Phys. Lett. B [**704**]{}, 604 (2011).
D. Momeni, N. Majd, and R. Myrzakulov, *p-Wave holographic superconductors with Weyl corrections*, Europhys. Lett. 97, 61001 (2012). D. Roychowdhury, *Effect of external magnetic field on holographic superconductors in presence of nonlinear corrections*, Phys. Rev. D [**86**]{}, 106009 (2012);
D. Momeni, M. R. Setare, and R. Myrzakulov, *Condensation of the scalar field with Stuckelberg and Weyl Corrections in the background of a planar AdS-Schwarzschild black hole*, Int. J. Mod. Phys. A [**27**]{}, 1250128 (2012);
D. Momeni and M. R. Setare, *A note on holographic superconductors with Weyl Corrections*, Mod. Phys. Lett. A [**26**]{}, 2889 (2011).
Z. X. Zhao, Q. Y. Pan, J. L. Jing, *Holographic insulator/superconductor phase transition with Weyl corrections*, Phys. Lett. B [**719**]{}, 440 (2013).
S. Chen and J. Jing, *Dynamical evolution of the electromagnetic perturbation with Weyl corrections*, Phys. Rev. D [**88**]{}, 064058 (2013);
S. Chen and J. Jing, *Dynamical evolution of a vector field perturbation coupling to Einstein tensor*, Phys. Rev. D [**90**]{}, 124059 (2014);
H. Liao, S. Chen and J. Jing, *Absorption cross section and Hawking radiation of the electromagnetic field with Weyl corrections*, Phys. Lett. B [**728**]{} 457-461 (2014).
A. Einstein, *Lens-like action of a star by the deviation of light in the gravitational field*, *Science*, **84**, 506 (1936).
P. Schneider, J. Ehlers, and E. E. Falco, *Gravitational Lenses*, *Springer-Verlag, Berlin*, (1992).
C. Darwin, *The gravity field of a particle*, Proc. of the Royal Soc. of London [**249**]{} 180 (1959).
K. S. Virbhadra and G. F. R. Ellis, *Gravitational lensing by naked singularities*, Phys. Rev.D [**65**]{}, 103004(2002);
C. M. Claudel, K. S. Virbhadra, G. F. R. Ellis, *The geometry of photon surfaces*, J. Math. Phys. [**42**]{} 818 (2001).
K. S. Virbhadra, D. Narasimha and S. M. Chitre, *Role of the scalar field in gravitational lensing*, Astron. Astrophys. [**337**]{} 1 (1998).
K. S. Virbhadra, G. F. R. Ellis, *Schwarzschild black hole lensing*, Phys. Rev. D [**62**]{} 084003 (2000);
K. S. Virbhadra,*Relativistic images of Schwarzschild black hole lensing*, Phys. Rev. D [**79**]{}, 083004 (2009);
K. S. Virbhadra, *Time delay and magnification centroid due to gravitational lensing by black holes and naked singularities*, Phys. Rev. D [**77**]{}, 124014 (2008).
S. Frittelly, T. P. Kling and E. T. Newman, *Spacetime perspective of Schwarzschild lensing*, Phys. Rev. D [**61**]{}, 064021 (2000).
V. Bozza, S. Capozziello, G. lovane and G. Scarpetta, *Strong field limit of black hole gravitational lensing*, Gen. Rel. and Grav. [**33**]{}, 1535 (2001).
E. F. Eiroa, G. E. Romero and D. F. Torres, *Reissner-Nordstrom black hole lensing*, Phys. Rev. D [**66**]{}, 024010 (2002);
E. F. Eiroa, *A braneworld black hole gravitational lens: strong field limit analysis*, Phys. Rev. D [**71**]{}, 083010 (2005).
R. Whisker, *Strong gravitational lensing by braneworld black holes*, Phys. Rev. D [**71**]{}, 064004 (2005).
V. Bozza, *Gravitational lensing in the strong field limit*, Phys. Rev. D [**66**]{}, 103001 (2002).
V. Bozza, *Quasiequatorial gravitational lensing by spinning black holes in the strong field limit*, Phys. Rev. D [**67**]{}, 103006 (2003);
V. Bozza, F. De Luca, G. Scarpetta, and M. Sereno, *Analytic Kerr black hole lensing for equatorial observers in the strong deflection limit*, Phys. Rev. D [**72**]{}, 083003 (2005);
V. Bozza, F. De Luca, and G. Scarpetta, *Kerr black hole lensing for generic observers in the strong deflection limit*, Phys. Rev. D [**74**]{}, 063001 (2006).
G. N. Gyulchev and S. S. Yazadjiev, *Kerr-Sen dilaton-axion black hole lensing in the strong deflection limit*, Phys. Rev. D [**75**]{} 023006 (2007);
G. N. Gyulchev and S. S. Yazadjiev, *Gravitational lensing by rotating naked singularities*, Phys. Rev. D [**78**]{} 083004 (2008).
S. Chen and J. Jing, *Strong gravitational lensing by a rotating non-Kerr compact object*, Phys. Rev. D [**85**]{}, 124029 (2012). A. Bhadra, *Gravitational lensing by a charged black hole of string theory*, Phys. Rev. D [**67**]{}, 103009 (2003).
T. Ghosh and S. Sengupta, *Strong gravitational lensing across dilaton Anti-de Sitter black hole*, Phys. Rev. D [**81**]{}, 044013 (2010), arXiv: 1001.5129.
A. N. Aliev and P. Talazan, *Gravitational effects of rotating braneworld black holes*, Phys. Rev. D [**80**]{}, 044023 (2009), arXiv:0906.1465.
C. Ding, C. Liu, Y. Xiao, L. Jiang and R. Cai, *Strong gravitational lensing in a black hole spacetime dominated by dark energy*, Phys. Rev. D [**88**]{} 104007 (2013);
E. F. Eiroa and C. M. Sendra, *Regular phantom black hole gravitational lensing*, Phys. Rev. D [**88**]{}, 103007 (2013);
S. Wei, Y. Liu, C. Fu and K. Yang, *Strong field limit analysis of gravitational lensing in Kerr-Taub-NUT spacetime*, JCAP [**1210**]{}, 053 (2012);
S. Wei and Y. Liu, *Equatorial and quasi-equatorial gravitational lensing by Kerr black hole pierced by a cosmic string*, Phys. Rev. D [**85**]{}, 064044 (2012).
G. V. Kraniotis, *Precise analytic treatment of Kerr and Kerr-(Anti) de Sitter black holes as gravitational lenses*, Class. Quant. Grav. [**28**]{}, 085021 (2011).
Y. Liu, S. Chen and J. Jing, *Strong gravitational lensing in a squashed Kaluza-Klein black hole spacetime*, Phys. Rev. D [**81**]{},124017 (2010);
S. Chen, Y. Liu and J. Jing, *Strong gravitational lensing in a squashed Kaluza-Klein Godel black hole*, Phys. Rev. D [**83**]{}, 124019 (2011);
L. Ji, S. Chen, and J. Jing, *Strong gravitational lensing in a rotating Kaluza-Klein black hole with squashed horizons*, J. High Energy Phys. [**03**]{}, 089 (2014).
J. Sadeghi, H. Vaez, *Strong gravitational lensing in a charged squashed Kaluza-Klein Gödel black hole*, Phys. Lett. B [**728**]{}, 170-182 (2014), arXiv:1310.4486;
J. Sadeghi, A. Banijamali and H. Vaez, *Strong gravitational lensing in a charged squashed Kaluza-Klein black hole*, Astrophys. Space Sci. [**343**]{}, 559 (2013), arXiv:1205.0805.
E. F. Eiroa, *Gravitational lensing by Einstein-Born-Infeld black holes*, Phys. Rev. D [**73**]{}, 043002 (2006)
R. D. Daniels, and G. M. Shore, *“Faster than light" photons and charged black holes*, Nucl. Phys. B [**425**]{}, 634 (1994);
R. D. Daniels, and G. M. Shore, *‘Faster than light’ photons and rotating black holes* Phys. Lett. B [**367**]{}, 75 (1996).
G. M. Shore, *Faster than Light Photons in Gravitational Fields II: Dispersion and Vacuum Polarisation*, Nucl.Phys. B [**633**]{}, 271 (2002).
R. G. Cai, *Propagation of vacuum polarized photons in topological black hole spacetimes*, Nucl. Phys. B [**524**]{}, 639 (1998). H. T. Cho, *“Faster Than Light¡± Photons in Dilaton Black Hole Spacetimes*, Phys. Rev. D [**56**]{}, 6416-6424 (1997).
V. A. De Lorenci, R. Klippert, M. Novello, and J. M. Salim, *Light propagation in non linear electrodynamics*, Phys.Lett. B [**482**]{}, 134 (2000);
D. D. Dalvit, F. D. Mazzitelli, and C. M. Paris, *One-loop graviton corrections to Maxwell’s equations*, Phys. Rev. D [**63**]{}, 084023 (2001);
N. Ahmadi and M. N. Zonoz, *Quantum gravitational optics: the induced phase*, Class. Quant. Grav. [**25**]{}, 135008 (2008).
N. Breton, *Geodesic structure of the Born¨CInfeld black hole*, Class. Quantum Grav. [**19**]{}, 601 (2002).
R. Genzel, F. Eisenhauer and S. Gillessen, *The galactic center massive black hole and nuclear star cluster*, Rev. Mod. Phys. [**82**]{}, 3121 (2010); arXiv:1006.0064.
[^1]: [email protected]
[^2]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Turbulence in superfluids depends crucially on the dissipative damping in vortex motion. This is observed in the B phase of superfluid $^3$He where the dynamics of quantized vortices changes radically in character as a function of temperature. An abrupt transition to turbulence is the most peculiar consequence. As distinct from viscous hydrodynamics, this transition to turbulence is not governed by the velocity-dependent Reynolds number, but by a velocity-independent dimensionless parameter $1/q$ which depends only on the temperature-dependent mutual friction – the dissipation which sets in when vortices move with respect to the normal excitations of the liquid. At large friction and small values of $1/q \lesssim 1$ the dynamics is vortex number conserving, while at low friction and large $1/q \gtrsim 1$ vortices are easily destabilized and proliferate in number. A new measuring technique was employed to identify this hydrodynamic transition: the injection of a tight bundle of many small vortex loops in applied vortex-free flow at relatively high velocities. These vortices are ejected from a vortex sheet covering the AB interface when a two-phase sample of $^3$He-A and $^3$He-B is set in rotation and the interface becomes unstable at a critical rotation velocity, triggered by the superfluid Kelvin-Helmholtz instability.
PACS numbers: 47.37, 67.40, 67.57\
address: |
$^*$Low Temperature Laboratory, Helsinki University of Technology\
P.O.Box 2200, FIN-02015 HUT, Finland\
$^{\dag}$Kapitza Institute of Physical Problems, Kosygina 2, 119334 Moscow, Russia\
$^{\ddag}$Landau Institute for Theoretical Physics, Kosygina 2, 119334 Moscow, Russia
author:
- 'V.B. Eltsov$^{*\dag}$, M. Krusius$^*$, and G.E. Volovik$^{*\ddag}$'
title: Transition to Superfluid Turbulence
---
INTRODUCTION
============
Superfluid turbulence – the tangled motion of quantized vortex lines in superfluid $^4$He-II – has been known to exist for fifty years.[@Feynman] By increasing the applied flow velocity beyond some relatively low critical value, at which vortices become mobile, their turbulent motion is started and superfluid flow becomes dissipative.[@Vinen] In this context we mean with applied flow the [*counterflow velocity*]{} ${\bm v} = {\bm
v}_{\rm n} - {\bm v}_{\rm s}$, the difference between the velocities ${\bm v}_{\rm n}$ of the normal and ${\bm v}_{\rm s}$ of the superfluid components, which is created by external means.
Only recently it has been realized from measurements on superfluid $^3$He-B that turbulence is not necessarily a generic property of all superfluids, but one whose presence crucially depends on the damping in vortex motion.[@Nature] The damping, or more accurately the mutual friction dissipation, turns out to govern the onset of turbulence, dividing vortex motion in $^3$He-B to superconductor-like regular behavior at high temperatures, where the vortex number is conserved in dynamic processes, and to $^4$He-like disordered behavior at low temperatures, where turbulence easily sets in when perturbations are introduced in superfluid flow.
The transition to turbulence in superfluids can be compared to that in viscous flow, as discussed in Ref. \[\] on linear pipe flow with circular cross section. Here the transition is governed by the velocity-dependent Reynolds number $Re=UD/\nu$, where $U$ is the mean velocity, $D$ the characteristic length scale (the pipe diameter in Ref. \[\]), and $\nu$ the kinematic viscosity. In the 16m long circular pipe of Ref. \[\] laminar flow is stable at flow velocities $Re < 20\,000$. However, on injecting a controlled square pulse of perturbing flow azimuthally in the linear stream, the flow can be converted from laminar to turbulent over some length of the pipe. This turbulence travels downstream in the pipe, while upstream in the absence of the perturbation laminar flow again recovers. The critical amplitude of the perturbing mass flux $\Phi_{\rm inj}$, required to reach the transient turbulent state, was found to obey a scaling law of the form $\Phi_{\rm inj}/\Phi_{\rm pipe} = Re^{-\delta}$, where the exponent has the value $\delta = 1 \pm 0.01$.
To generate the transition to turbulence in superfluids, one or several vortices are injected in rotating vortex-free flow of $^3$He-B.[@Turbulence] It is then found that the transition is not governed by viscosity (which is absent for the superfluid fraction of the liquid), but by the mutual friction between vortices and the normal fraction of the liquid. As distinct from viscosity $\eta$, which enters the Reynolds number as the dimensional kinematic viscosity $\nu = \eta / \rho$, mutual friction is described by two dimensionless parameters which represent its dissipative and reactive components. In $^3$He-B in the range of the transition to turbulence both parameters are of comparable magnitude and have to be taken into account. In $^4$He-II the reactive mutual friction parameter is much smaller and is usually neglected (see review \[\]). In the superfluid, the viscosity dependent Reynolds number has to be replaced by a dimensionless characteristic number which is called $1/q$ and only depends on the two mutual friction parameters. In particular as opposed to Reynolds number, $1/q$ is velocity independent which is also observed in measurement at higher flow velocities as the limiting case. In this short review we focus on this limiting regime, which corresponds to the case when a sufficient number of closely spaced seed vortex loops is injected, so that they immediately start interacting and instantaneously produce turbulence.[@Nature]
If the number of seed vortices is reduced or their spacing is increased, then the flow perturbation is weakened and the transition to turbulence moves to higher values of $1/q$. In this case more new vortices need to be generated, before they can start interacting turbulently. Measurements with injection down to the limit of one single seed vortex are discussed in Refs. \[\]. Obviously such cases require some additional mechanism, which leads to an increase in the number of vortices in the low-density regime, when vortices do not yet interact. This is the single vortex instability in applied flow which via loop formation and reconnection generates new independent vortices. In this way the transition to turbulence becomes a complex process of series coupled mechanisms in the regime of small flow perturbation.[@Precursor] However, independently of the applied perturbation, in all these measurements the onset of turbulence is displayed as an abrupt transition, which takes place within a narrow distribution of $1/q$ values. Moreover, the average of this distribution of $1/q$ values proves to depend on the magnitude of the applied perturbation in a power-law manner.
Comparing measurements on the transition to turbulence in viscous and superfluid flow, we notice that they proceed in somewhat different manner. Nevertheless, there are similarities: In both cases (i) the initial state is perturbed externally by means of a quantitatively controlled disturbance which (ii) sets off turbulence for a short length of time if (iii) the perturbation is of sufficient amplitude, with power-law dependence on the relevant controlling parameter (which is $Re$ in viscous flow and the mutual friction dependent parameter $1/q$ in superfluids).
There are special reasons why the transition to turbulence as a function of mutual friction has not been observed in superfluid $^4$He-II and was only recently discovered in superfluid $^3$He-B. As distinct from $^4$He-II, $^3$He-B is a Fermi superfluid, where the superfluidity is caused by Cooper pairing. The mutual friction between vortices and the normal fraction of the liquid, which is composed of fermionic quasiparticles, is mediated by quasiparticles populating the vortex core states,[@KopninBook] the so-called fermion zero modes[@VolovikBook]. The scattering between the two types of quasiparticles leads to mutual friction and is described by a theory similar to the BCS theory of superconductivity. As a result the parameter $q(T)$ appears to be a dimensionless function of the dimensionless parameter $T/T_{\rm
c}$. In Fermi superfluids in the weak coupling approximation this parameter crosses unity at $T\sim 0.6 T_{\rm c}$, [*i.e.*]{} in the middle of the experimentally accessible temperature range of $^3$He-B. (In the cold superfluid fermionic gases discussed in Ref. \[\] $q$ can be adjusted with a magnetic field if the system is close to the Feshbach resonance.)
In contrast, in the boson superfluid $^4$He-II vortex dynamics is practically always in the turbulent regime (see review \[\]). Regular flow of vortices could be perhaps expected only within microkelvins from the superfluid transition temperature $T_\lambda$, but there is not yet enough information on vortex dynamics in this regime. Even there, the low viscosity of the normal component (in $^4$He-II the normal component is one of the least viscous fluids existing) causes its flow to become easily turbulent, which can in turn influence the flow of the superfluid component. In contrast, in $^3$He-B the normal component has $\sim 10^4$ times higher oil-like viscosity and is practically always in a state of laminar flow. The absence of turbulence in the flow of the normal component of $^3$He-B amounts to a considerable simplification and leads to new effects, which are absent in $^4$He-II. An example is a new scaling law for the Kolmogorov-Richardson cascade in developed homogeneous superfluid turbulence.[@LNV; @LNS]
The injection mechanism, which led to the discovery of the transition to turbulence as a function of $1/q$, is of particular interest. Here the injected seed vortices originate from the AB interface in a two-phase sample of $^3$He-A and $^3$He-B. The seed vortices are tossed as a tight bundle of some 10 loops across the AB interface from $^3$He-A into the vortex-free flow of $^3$He-B. This happens when the interface becomes unstable with respect to wave formation at a well-defined critical value for superfluid counterflow parallel to the interface. During the non-linear stage of this corrugation instability, the vortices in the deepest corrugation of the interface wave are ejected on the B-phase side of the interface. The instability itself, known as the superfluid Kelvin-Helmholtz shear flow instability, is reproducible and predictable, its measurements and theory match without fitting parameters. This is different from the ordinary Kelvin-Helmholtz instability at the interface between two viscous liquids or gases. In viscous fluids the initial state is not well described, since the shear-flow configuration is not an equilibrium situation and cannot be expressed as a solution of the Navier-Stokes equation. Interestingly, the superfluid Kelvin-Helmholtz instability shares some characteristics with the instability of quantum vacuum within the horizon or ergoregion of the black hole.[@BHandWH]
In this short review[@ROPreview] we shall first introduce the superfluid Kelvin-Helmholtz instability and the injection of vortex seed loops in Sec. 2. Then follows in Sec. 3 a description of the transition to turbulence in the case when turbulence is instantaneously started by the injected vortex loops.
SUPERFLUID KELVIN-HELMHOLTZ INSTABILITY {#experiment}
========================================
Kelvin-Helmholtz Instability in Viscous Liquids {#InterfaceInstability}
-----------------------------------------------
Kelvin–Helmholtz (KH) instability is one of the many interfacial instabilities in the hydrodynamics of liquids, gases, charged plasma, and even granular materials. It refers to the dynamic instability of an interface with discontinuous tangential flow velocities and can loosely be defined as the instability of a vortex sheet. Many natural phenomena have been attributed to this instability. The most familiar ones are the generation of capillary waves on the surface of water, first analyzed by Lord Kelvin,[@LordKelvin] and the flapping of sails and flags, first discussed by Lord Rayleigh.[@Rayleigh]
Many of the leading ideas in the theory of interfacial instabilities in hydrodynamics were originally inspired by considerations about ideal inviscid flow. A horizontal interface between two ideal liquids, stacked on top of each other by gravity because of their different mass densities $\rho_1$ and $\rho_2$, and flow parallel to the interface at velocities ${\bf v}_1$ and ${\bf v}_1$, leads to a corrugation instability at the critical differential flow velocity[@landau_fluid_dynamics] $$({\bf v}_1-{\bf v}_2)^4= 4\sigma g (\rho_1-\rho_2) \, \frac{
(\rho_1+\rho_2)^2}{ \rho_1^2\rho_2^2}~. \label{KHClassical}$$ Here $\sigma$ is the surface tension of the interface and $g$ gravitational acceleration. To separate the gravitational and inertial properties of the liquids, let us rewrite the threshold velocity in the following form $${\rho_1\rho_2\over \rho_1+\rho_2} \, ({\bf v}_1-{\bf v}_2)^2=2\sqrt{\sigma
F}~. \label{InstabilityCondition1}$$ We associate $F$ with the external field stabilizing the position of the interface, which in the gravitational field is the gravity force $$F=g(\rho_1-\rho_2)~, \label{GravityForce}$$ but which in the general case can originate from some other source. The surface mode of ripplons or capillary waves, which is first excited at the instability, has the wave number corresponding to the inverse ‘capillary length’, $$k_0=\sqrt{F/\sigma}~. \label{WaveVectorInstability}$$ However, ordinary fluids are not ideal and the correspondence between this theory and experiment is not good. One reason for this is that one cannot properly prepare the initial state – the shear-flow discontinuity is never in equilibrium in a viscous fluid. It is not a solution of the Navier-Stokes equation. That is why one cannot properly extend the ‘instability’ of the inviscid case to finite viscosities.
Kelvin-Helmholtz Instability in Superfluids {#kh_superfluid}
-------------------------------------------
In superfluids the criterion for the instability can be formulated in the absence of viscosity, since the tangential velocity discontinuity at the interface between $^3$He-A and $^3$He-B is a stable non-dissipative state. These two superfluid phases have different magnetic properties and their interface is stabilized by the gradient in the applied magnetic field $H(z)$ which provides the restoring force $F$ in Eq. (\[InstabilityCondition1\]): $$F= {1\over 2}\nabla \left((\chi_{\rm A} -\chi_{\rm B} ) H^2\right)
~. \label{InstabilityCondition2}$$ Here $\chi_{\rm A}>\chi_{\rm B}$ are the magnetic susceptibilities of the A and B phases, respectively. One might expect that by substituting this interfacial restoring force $F$ into Eq. (\[InstabilityCondition1\]) and using the superfluid densities of the A and B phases instead of the total density, one obtains the critical velocity for the KH instability of the A-B interface. However, it turns out that a proper extension of the KH instability to superfluids incorporates the criterion in Eq. (\[InstabilityCondition1\]) only as a particular limiting case.
The criterion for the KH instability of ideal fluids in Eq. (\[InstabilityCondition1\]) depends only on the relative velocity across the interface. In practice there always exists a preferred reference frame, imposed by the environment. In the superfluid case it is the frame of the normal component moving with velocity $\mathbf{v}_\mathrm{n}$ (in equilibrium $\mathbf{v}_\mathrm{n}=0$ in a frame frame fixed to the rotating container). Owing to this interaction of the AB interface with its environment, the instability occurs at a lower differential flow velocity than the classical criterion in Eq. (\[InstabilityCondition1\]) assumes (see Refs. \[\]): $${1\over 2} \rho_{\mathrm{s}B}({\bf v}_{\rm B} -{\bf v}_{\rm n})^2
+ {1\over 2} \rho_{\mathrm{s}A}({\bf v}_{\rm A} -{\bf v}_{\rm
n})^2 =\sqrt{\sigma F}~. \label{InstabilityConditionNewnon-zeroT}$$ Here ${\bf v}_{\rm A}$ and ${\bf v}_{\rm B}$ are the velocities of the superfluid components on the A- and B-phase sides of the interface; while $\rho_{\mathrm{s}A}$ and $\rho_{\mathrm{s}B}$ are the corresponding densities of these superfluid fractions.
Observation of Kelvin-Helmholtz Instability {#ObservSuperfluid}
-------------------------------------------
![Critical rotation velocity $\Omega_{\rm c} \approx
(v_{\rm n} -v_{\rm B})/R$ of the AB-interface instability versus temperature at different currents $I_{\rm b}$ in the barrier magnet which is employed to stabilize the A phase. If the barrier field $H(z)$ exceeds the critical value $H_{\rm AB}(T,P)$ needed for the A phase, then the AB interface resides at the two locations $z$ where $H(z) = H_{\rm AB}(T,P)$. These locations and the magnetic restoring force in Eq. (\[InstabilityCondition2\]) depend on temperature. The temperatures, where the AB interface disappears, are indicated by vertical arrows below the figure for different values of $I_{\rm b}$. The solid curves represent the instability criterion (\[InstabilityConditionNewnon-zeroT\]).[]{data-label="KHInstabilityCurvesFig"}](omega_cAB_29bar-c.eps){width="\linewidth"}
A comparison of Eq. (\[InstabilityConditionNewnon-zeroT\]) to the measured critical rotation velocity $\Omega_{\rm c}$ of the first KH instability event is shown in Fig. \[KHInstabilityCurvesFig\].[@Kelvin-HelmholtzInstabilitySuperfluids] Here we set $v_{\rm B}=\Omega R$ and $v_{\rm A}= v_{\rm n}=0$ ([*i.e.*]{} the velocities are given in the frame of the rotating container, and $R$ is the radius of the container). No fitting parameters are used. The curves have been drawn using generally accepted values for the different superfluid $^3$He parameters. Even the remaining residual differences between measurement and Eq. (\[InstabilityConditionNewnon-zeroT\]) have plausible explanations which are discussed in Ref. \[\]. Such remarkable agreement for a complicated phenomenon can only be achieved in superfluids, where shear flow is not dissipative until the instability threshold is reached. Although the instability in Eq. (\[InstabilityConditionNewnon-zeroT\]) depends on the reference frame fixed to the normal component, this does not mean that the renormalized instability criterion would depend on the magnitude of the interaction with the normal component – in fact, it is still determined by only thermodynamics. Waves are formed on the interface when the free energy of a corrugation becomes negative in the reference frame of the environment.
In terms of the jargon accepted in general relativity, the so-called ergoregion is then formed – the region in which the (free) energy of some excitations is negative. This connection between the interface instability, horizons, and ergoregions in black hole physics is worked out in Refs. \[\]. In the ergoregion the AB interface becomes thermodynamically unstable, since it becomes possible to reduce the energy via the normal component and its interaction with the solid sample boundary. The original classical instability condition Eq. (\[InstabilityCondition1\]) of the ideal inviscid fluid is restored if the interaction with the environment is not effective. This might occur in the superfluid, for example, during rapid rotational acceleration at very low temperatures when the instability caused by the interaction with the environment has not had enough time to develop.
However, measurements now attest that down to moderately low temperatures of $0.35\,T_{\rm c}$ the instability condition is not given by the classical KH expression (\[InstabilityCondition1\]) even in a perfectly inviscid superfluid, but by the renormalized criterion in Eq. (\[InstabilityConditionNewnon-zeroT\]). Its central property is that the instability condition does not depend on the relative velocity of the two superfluids, but on the velocity of each of the superfluids with respect to the environment. The instability will occur even if the two fluids have equal densities, $\rho_{\rm A} =\rho_{\rm B}$, and move with the same velocity, ${\bf v}_{\rm A}={\bf v}_{\rm B}$. The instability also occurs if there is only a single superfluid with a free surface. These new features arise from the two-fluid nature of the superfluid. The situation resembles that of a flag flapping in wind, which was originally discussed in terms of the KH instability of ideal inviscid fluids by Lord Rayleigh.[@Rayleigh] Newer explanations involve boundary interactions between the flag and the two gas streams which become turbulent within a narrow boundary layer. The superfluid analogue is an instability of a passive deformable membrane between two distinct parallel streams having the same density and velocity. The flag is the AB interface and the flagpole, which pins the flag, serves as the reference frame fixed to the environment so that Galilean invariance is violated.
![One scenario for the transfer of circulation across the AB interface in a KH instability. The A-phase vorticity is confined by the Magnus lift force to a vortex layer which covers the AB interface.[@Hanninen] The Magnus force arises from the tangential B-phase superflow below the interface. The instability creates a corrugation in the interface which becomes a potential well for A-phase vortices there. This pushes the vortices deeper in the potential well, deforming the corrugation to a trench with a droplet-like cross section. The trench, filled with A-phase vorticity, then propagates in the bulk B-phase where A phase is unstable and the multiply-quantized vortex relaxes to singly-quantized B-phase vortices. In this scenario the number of vortices ejected across the interface corresponds to the number of vortices in the A-phase vortex layer within one corrugation or one half wave-length of the interface ‘ripplon’, which is in accordance with measurements.[@Kelvin-HelmholtzInstabilitySuperfluids] []{data-label="KelvinInstabilityFig"}](KelvinInstability-a.eps){width="0.9\linewidth"}
Vortex Injection {#VortexInjection}
----------------
The non-linear stage of the interface instability, when the interface is distorted following the break-down of its ultimate equilibrium condition, leads to the injection of a bundle of vortices from the AB interface in the rapidly flowing B-phase, when viewed in the frame of the rotating sample container. The mechanism of injection has not been worked out in detail, but a simplified scenario is presented in Fig. \[KelvinInstabilityFig\]. In general, this mechanism can be associated with the non-linear development of the vortex sheet instability, as discussed [*e.g.*]{} in Ref. \[\].
In this context the detailed properties of the injection event are not all that important. Rather we are interested in the dynamic evolution of this bundle of vortices which initially (following the instability) protrudes out of the AB interface and then continues to the cylindrical sample boundary in the B-phase volume. The measurements display at first glance a rather unexpected result: The final B-phase state, [*i.e.*]{} the state which after some transient development will be stable in time (at constant rotation, temperature, and pressure), does not depend on the initial velocity of B-phase flow, but only on temperature.
SUPERFLUID TURBULENCE {#SuperfluidTurb}
=====================
Transition to Turbulence as a Function of Mutual Friction {#Exp}
---------------------------------------------------------
![Number of rectilinear B-phase vortex lines $N$ after KH instability, normalized to the equilibrium number $N_{\rm eq}$ and plotted versus temperature. At around $0.58\,T_\mathrm{c}$, a sudden change in the number of lines is observed. At temperatures above the transition the injection results in only a few lines, but below the transition close to the equilibrium number is counted. []{data-label="fillingratio"}](FillingRatio-c.eps){width="0.9\linewidth"}
At temperatures above $0.6\,T_{\rm c}$ the dynamics of vortex loops injected in vortex-free flow of superfluid $^3$He-B is regular and their number does not increase during their time-dependent evolution to rectilinear vortex lines of the rotating state. At lower temperatures it becomes possible for the vortices within a bundle of vortex loops, which is injected in a KH instability in the B phase, to start interacting turbulently. Surprisingly this happens within a narrow temperature regime of $0.06\,T_{\rm c}$ width centered at about $0.59\,T_{\rm c}$, as seen in Figs. \[fillingratio\] and \[PDFig\]. The consequence from this abrupt change in the dynamics, while the vortices evolve in the externally applied flow towards their rectilinear final state, is a radical change in the final number of vortex lines: While at temperatures above the transition the number of rectilinear vortices in the final state is a few and reproduces the distribution of vortices crossing the AB interface in one KH instability event, at temperatures below the transition the final state includes close to the equilibrium number of vortex lines.[@Nature] These properties can be examined in great detail, since the KH critical velocity stays well-behaved as a function of temperature and is a continuous smooth curve across this division line, continuing to follow the calculated dependence. All this indicates that the transition is not related to the properties of the KH instability itself, but arises from the change in the dynamics of vortices moving in rapidly flowing $^3$He-B.
![Temperature – velocity phase diagram of turbulence. Each data point represents the result from a KH injection measurement where $\Omega$ is increased from zero to $\Omega_{\rm c}$ at constant temperature and the number of rectilinear vortex lines $N$ is measured in the final state. The result is marked with an open symbol, when the number is close to that in the equilibrium state $N_{\rm eq}$ and when turbulence must have followed injection. Regular vortex expansion with no increase in $N$ is denoted with filled symbols. The vertical dashed line marks the boundary between the two final states. The lower plot shows the distribution of the final states in terms of gaussian fits. The horizontal top axis gives the mutual friction parameter $q(T)$. []{data-label="PDFig"}](KH-TurbulenceTransition-b.eps){width="0.61\linewidth"}
In Fig. \[PDFig\] the phase diagram of regular and turbulent vortex dynamics is shown as a function of flow velocity (vertical scale) and temperature (horizontal scale). The vertical axis is expressed in terms of the rotating B-phase counterflow velocity $|{\bf v}_{\rm B} - {\bf v}_{\rm n}|=\Omega R$. In this diagram each data point represents a KH injection measurement, obtained with different settings of the externally controlled parameters $T$ and $I_{\rm b}$, so that a variation as wide as possible for the critical velocity $\Omega_{\rm c}$ is obtained. What we are interested in here is the division in filled and open symbols: Injection events followed by a turbulent burst are marked with open symbols while events which lead to only a few rectilinear lines are marked with filled symbols.
The striking conclusion from Fig. \[PDFig\] is that the boundary between turbulence at low temperatures and regular dynamics at high temperatures is vertical, [*i.e.*]{} it is independent of the applied flow velocity above about 2.5mm/s. It will be next shown that this transition to turbulence is driven by the dimensionless velocity independent parameter $q(T)^{-1}$, which characterizes the friction force acting on vortices when they move with respect to the normal component of the liquid. It is this parameter which divides the vortex dynamics in superfluids into a low-$1/q$ regime with regular vortex number conserving motion and a high-$1/q$ regime where superfluid turbulence becomes possible.
Superfluid Equivalent of Reynolds Number {#Theory}
----------------------------------------
The velocity $\bm{v}_{\rm L}$ of an element of a vortex line is given by $$\bm{v}_{\rm L}=\bm{v}_{\rm s} +\alpha \hat{\bm{s}} \times
(\bm{v}_{\rm n}-\bm{v}_{\rm s}) -\alpha' \hat{\bm{s}} \times
[\hat{\bm{s}} \times (\bm{v}_{\rm n}-\bm{v}_{\rm s})]\,,
\label{vl}$$ where $\hat{\bm{s}}$ is a unit vector parallel to the vortex line element. This equation of motion depends on the dimensionless mutual friction parameters $\alpha(T)$ and $\alpha'(T)$, which originate from the dissipative and reactive forces acting on a vortex when it moves with respect to the normal component. For vortices in a fermionic system they were calculated by Kopnin.[@KopninBook] For $^3$He-B they were measured over a broad temperature range by Bevan [*et al.*]{}[@Bevan] (see also the monograph \[\], where these parameters are discussed in terms of the chiral anomaly). Evidently the nature of the solutions of Eq. (\[vl\]) for $\bm{v}_{\rm L}$ has to depend on the mutual friction parameters.
This property will become clearer if we form the coarse-grained hydrodynamic equation for superfluid vorticity $ {\bm
\omega}=\nabla\times {\bf v}_{\rm s}$, by averaging over vortex lines. This equation for the superfluid velocity can be obtained directly from the Euler equation for inviscid liquids where, instead of the viscous $\nabla^2{\bf v}$ term of the Navier-Stokes equation, one then has (see review in Ref. \[\]) $$\begin{aligned}
\frac{\partial {\bf v} }{ \partial t}+ \nabla\mu= {\bf v} \times
{\bm \omega}~-
\alpha'({\bf v} -{\bf v}_{\rm n})\times
{\bm \omega}+ \alpha~\hat {\bm \omega}
\times( {\bm \omega} \times({\bf v} -{\bf v}_{\rm n}) ) ~.
\label{SuperfluidHydrodynamics2}\end{aligned}$$ Here ${\bm v}\equiv {\bm v}_{\rm s}$ is the superfluid velocity, $
{\bm \omega}=\nabla\times {\bf v}$ the superfluid vorticity, and $\hat {\bm \omega}$ the unit vector $\hat {\bm \omega}= {\bm
\omega}/\omega$. In $^3$He-B the normal component has oil-like viscosity, its motion is practically always laminar, and we ignore its dynamics. In the frame where it is at rest, [*i.e.*]{} ${\bf
v}_{\rm n}=0$, the equation for superfluid hydrodynamics is simplified: $$\frac{\partial {\bf v} }{ \partial t}+ \nabla\mu= (1-\alpha'){\bf
v} \times
{\bm \omega}+ \alpha~\hat{\bm\omega} \times( {\bm \omega}
\times{\bf v} ) ~. \label{SuperfluidHydrodynamics}$$ After rescaling the time, $\tilde t=(1-\alpha')t$, one obtains an equation which depends on a single parameter $q=\alpha/(1-\alpha')$: $$\frac{\partial {\bf v} }{ \partial \tilde t}+ \nabla \tilde\mu-
{\bf v} \times
{\bm \omega}= q~\hat{\bm\omega} \times( {\bm \omega}
\times{\bf v} ) ~. \label{SuperfluidHydrodynamics3}$$ Let us compare this equation with the Navier-Stokes equation $$\frac{\partial {\bf v}}{ \partial t}+ \nabla\mu- {\bf v}\times
{\bm \omega}= \nu\nabla^2 {\bf v}
~. \label{NormalHydrodynamics}$$ The inertial terms on the lhs of Eq. (\[SuperfluidHydrodynamics3\]) for superfluid hydrodynamics are the same as those on the lhs of Eq. (\[NormalHydrodynamics\]) for viscous hydrodynamics. In contrast, the dissipative term on the rhs of Eq. (\[SuperfluidHydrodynamics3\]) is different from the corresponding viscous term on the rhs of Eq. (\[NormalHydrodynamics\]). This difference between the dissipative terms in the two equations of motion is of importance. Reynolds number, which characterizes the nature of the solutions, is formed as the dimensional equivalent of the ratio of the inertial and dissipative terms in the two hydrodynamic equations (\[SuperfluidHydrodynamics3\]) and (\[NormalHydrodynamics\]). For the Navier-Stokes case Eq. (\[NormalHydrodynamics\]) the ratio corresponds dimensionally to the conventional Reynolds number $Re=UD/\nu$, where $U$ is the characteristic velocity scale and $D$ the size of the large-scale flow. In superfluid hydrodynamics Eq. (\[SuperfluidHydrodynamics3\]) this ratio is simply given by the dimensionless intrinsic parameter $1/q$. Since $1/q$ does not depend on velocity or on the large-scale system size, turbulence becomes possible only if $1/q$ is large enough. In accordance to the usual characterization of the solutions of the Navier-Stokes equation (\[NormalHydrodynamics\]) we might thus expect the transition to turbulence to occur at $1/q \sim 1$. This is in agreement with the experimental phase diagram in Fig. \[PDFig\].
In the above considerations we used the coarse-grained description of superfluid vorticity, which does not take into account the discreteness of quantized vortices. This implies, that the flow velocity is large enough, [*i.e.*]{} it is essentially higher than the Feynman critical velocity, $U\gg \kappa/D$, where $\kappa$ is the circulation quantum. This condition is always fulfilled in $^3$He-B experiments, in which vortices are injected via the KH instability, since it occurs in the limit where the critical value of $1/q$ does not depend on the flow velocity, as seen in Fig. \[PDFig\]. However, at much lower velocities the critical value of $1/q$ is ultimately expected to increase.
Furthermore, the coarse-grained description of quantized vorticity for estimating the threshold value for $1/q$ assumes a sufficiently large number and density of injected vortices. The number of vortices injected in the KH instability is on average 10 (in a distribution which ranges from 3 to 30). They are packed close to each other since they originate from the same growing, but over-damped, ripplon corrugation (Fig. \[KelvinInstabilityFig\]). However, in other types of injection experiments the number and density of injected vortices can be less,[@Precursor; @Solntsev] and the critical value of $1/q$ increases. This is similar to what occurs in viscous pipe flow,[@Mullin] where the critical Reynolds number $Re$ for the onset of turbulence depends on the magnitude of the perturbation.
The transition from regular to turbulent dynamics as a function of the mutual friction parameter $1/q$ is a new phenomenon. It has not been observed in $^4$He-II, where $1/q$ is practically always large, so that a transition is expected only a few tens of microkelvins below $T_{\lambda}$ where $\rho_{\rm s}$ is vanishingly small, the coherence length $\xi(T)$ diverges, critical velocities approach zero, and vortex dynamics enters a regime where little is known. In $^3$He-B the transition at $1/q\sim 1$ is in the middle of the experimentally accessible temperature range and can be observed in one single experiment by scanning temperature from the superconductor-like dynamics (with no pinning, but high vortex damping) at high temperatures to superfluid $^4$He-like turbulence at low temperatures.
These features demonstrate that vortex dynamics in superfluids takes varied forms and that the traditional $^4$He-like behavior is just one extreme example. The opposite extreme is the strongly anisotropic superfluid $^3$He-A where $1/q$ is practically always in the range of regular dynamics and the unusually low temperatures, which would be required for conventional superfluid turbulence, are not experimentally realistic at this time. Instead, in $^3$He-A it is transitions in the structure of the vortex and in the global order parameter texture which provide a means to adjust to faster dynamics. An example of this is the dynamically driven transition from vortex lines to sheets in applied flow which is periodically alternating in direction[@LinesToSheets] and the dependence of the critical velocity on the global order parameter texture[@CritVel].
Turbulence in Uniform Rotation {#TurbulentBurst}
------------------------------
![Temporal evolution of vorticity in a rotating column, shown schematically (in the rotating frame) following the injection of seed vortices across the AB interface in a KH instability (at $T>0.40\,T_{\rm c}$). ([**a**]{}) Vortex-free initial state above AB interface. ([**b**]{}) A tight bundle of $\sim 10$ vortex loops is ejected across the AB interface in a KH instability. ([**c**]{}) A brief burst of turbulence spreads across the entire cross section next to the AB interface. Close to the equilibrium number of independent vortices is created and immediately polarized. ([**d**]{}) The vortices expand in spiral motion along the column with a sharp front towards the vortex-free section. ([**e**]{}) When the front reaches the top end plate, the twisted vortex state continues relaxing to rectilinear lines. If the twist is not too tight, then it is removed by the slow diffusive motion of the vortex ends on both end plates. ([**f**]{}) Final stable equilibrium vortex state with rectilinear lines.[]{data-label="TurbulenceInRotation"}](RotTurbulence-b.eps){width="\linewidth"}
So far measurements on the transition to turbulence have been performed in uniformly rotating flow of $^3$He-B. The long smooth-walled sample container was prepared from fused quartz with an aspect ratio (length to diameter) of almost 20. The sample can be examined in single phase configuration filled with only B phase, when the barrier magnet is not charged. In the opposite case the barrier field stabilizes over the central section of sample length a narrow layer of A phase which acts as a vortex barrier. The two end sections of equal length with B phase represent then two independent samples, whose lengths are roughly half of that of the single phase sample. In this way one can examine the influence of the length of the cylinder on the evolution and propagation of vortices in the rotating column. Another bonus is the possibility to study two B-phase samples in parallel when the A-phase barrier layer is present. This feature is of importance since the critical velocity of the sample container is an unknown predicament. The precondition for KH injection measurements is a high critical velocity. This restricts measurements to an open cylindrical volume with smooth clean walls, no internal measuring probes, and isolated from the rest of the liquid $^3$He volume by a small orifice in the center of one of the flat end plates. The number of rectilinear vortices in these two B phase sections is monitored with non-invasive NMR absorption measurements, by means of NMR detector coils mounted outside around both ends of the long quartz cylinder.
Rotation provides a strong polarization for the vorticity to be oriented along the rotation axis. It is known from studies of turbulence in rotating viscous fluids that turbulent disorder then tends to become limited to the transverse plane. A similar phenomenon occurs in thermal counterflow of rotating $^4$He-II, if the thermal current is imposed parallel to the rotation axis so that it counteracts the polarization from rotation.[@Tsubota] Thus turbulence in effect tends to become 2-dimensional in rotation. In the case of KH injection turbulence is limited to a brief burst which takes place in the cross section of the cylinder next to the AB interface where the injection occurs. In this short burst the equilibrium number of vortices is created and, because of rapid polarization, the free energy drops instantaneously within this cross section from the maximum close to the minimum value. In Fig. \[TurbulenceInRotation\] the ensuing evolution of the vorticity is shown schematically. It consists of the winding cork-screw-like propagation of the vorticity along the rotating column, and the final straightening of the helically twisted vortices towards rectilinear lines of the stable equilibrium vortex state.[@TwistedState]
The spiralling vortex motion in the rotating column can be seen to emerge from Eq. (\[vl\]). Let us write the expression for the velocity of the end point of a vortex, where it connects in perpendicular orientation to the cylindrical side wall of the sample container. In the vortex front this end point practically resides in vortex-free counterflow with ${\bm v}_{\rm n} = \Omega
R {\hat {\bm \phi}}$ and ${\bm v}_{\rm s} = 0$, so that its velocity is $${\bm v}_{\rm L} = -(1-\alpha^{\prime})\Omega R
{\hat {\bm \phi}} + \alpha \Omega R {\hat {\bm z}}~,$$ which includes two components. The first component enforces azimuthal motion which differs by the fraction $1-\alpha^{\prime}$ from the velocity $-\Omega R$ at which the superfluid component travels with respect to the cylinder wall. The second component moving with the velocity $\alpha \Omega R$ corresponds to the longitudinal expansion of the vortex end along the column.
With decreasing temperature the longitudinal motion slows down and ultimately becomes exceedingly slow. Since the KH instability occurs at a well-defined critical velocity, it can be started as a triggered event and the longitudinal expansion velocity can be measured by timing the time delay from the trigger to the moment when the vortex front reaches a detector coil.[@Turbulence] At higher temperatures above $0.40\,T_{\rm c}$ the longitudinal velocity of the front is found to be close to $\alpha \Omega
R$.[@TimeOfFlight] Thus the above scenario explains qualitatively the formation of the twisted vortex state behind the vortex front. The twisted state corresponds to dynamic equilibrium in a ‘force-free configuration’.[@TwistedState] This means that the residual superfluid velocity created by the twist of the vortices, which has both radial and longitudinal components, is strictly oriented along the vortex cores. With decreasing temperature the vortex front becomes sharper, the longitudinal expansion is slowed down, the twist in the trailing vortex bundle is wound ever tighter, and the residual superflow velocity along the vortices increases. Ultimately the flow along the vortices is expected to make them unstable with respect to the Glaberson-Donnelly Kelvin-wave instability and inter-vortex reconnections will become frequent behind the vortex front.[@KelvinWaveInstability] Such reconnections help to remove the twist and to speed up the relaxation of the twisted state towards the stable equilibrium vortex state with rectilinear lines (Fig. \[TurbulenceInRotation\]).
CONCLUSION {#ConclusionSec}
==========
Although the two helium superfluids, $^4$He-II and $^3$He-B, were expected to obey similar hydrodynamics, they differ somewhat in the actual values of their hydrodynamic properties. Rather surprisingly, these small differences have led to new insight in superfluid dynamics during recent years. An example is the transition between turbulent and regular vortex dynamics as a function of mutual friction dissipation $\alpha(T)$. In the superfluid literature a ‘transition to turbulence’ has so far been associated exclusively with the transition in $^4$He-II as a function of the applied flow velocity.[@Skrbek] These velocities, at which vortices typically become mobile in $^4$He-II, are very low compared to the intrinsic critical velocity of the bulk superfluid. In contrast, the mutual friction driven transition to turbulence at high flow velocities is velocity independent. Among the various coherent quantum systems $^3$He-B is unique in that $\alpha(T)$ spans the range where $1/q
= (1-\alpha^{\prime})/\alpha$ passes through the transition at $1/q \sim 1$. A similar situation might become available in atomic Bose-Einstein condensates, where mutual friction[@BEC-mf] has also been predicted to include the regime $1/q \sim 1$ and where turbulence might soon be experimentally realized[@BEC-turbulence].
A second advantage in using $^3$He-B for vortex dynamics studies is the high viscosity of its normal component. This guarantees that the normal component is practically always in a state of laminar flow and provides a well-behaved reference frame for superfluid dynamics. This simple situation is quite unlike that in $^4$He-II, where the coupled dynamics of the superfluid and normal components have to be included from the start. Finally we note that owing to the much larger superfluid coherence length $\xi(T)
> 10\,$nm, which determines the length scale for the radius of the vortex core in isotropic He superfluids, critical velocities can be controlled more efficiently in $^3$He-B. This allows new types of experimental studies. One example is the externally controlled introduction of seed vortices in vortex-free flow which made it possible to identify the mutual-friction-dependent transition to turbulence. The superfluid KH instability of the AB interface in a two-phase sample of $^3$He superfluids is one of the unique and most reproducible of such injection methods. At present time this is the method by which a direct transition to bulk turbulence is believed to start at the lowest possible value of $1/q$.
[**Acknowledgements:**]{} This work is supported in part by the EU Trans-national Access Programme FP6 ($\#$RITA-CT-2003-505313), by the Academy of Finland via its 2006 grant for visitors from Russia, by the Russian Ministry of Education and Science (Leading Scientific School grant $\#$1157.2006.2), and by the European Science Foundation COSLAB Program.
[9]{}
R.P. Feynman, Application of quantum mechanics to liquid helium, in [*Prog. Low Temp. Phys.*]{}, Vol. I, Ch. 2, editor C.J. Gorter (North-Holland Publ. Co., Amsterdam, 1955).
W.F. Vinen, Vortex lines in liquid helium II, in [*Prog. Low Temp. Phys.*]{}, Vol. III, Ch. 1, editor C.J. Gorter (North-Holland Publ. Co., Amsterdam, 1961).
A.P. Finne, T. Araki, R. Blaauwgeers, V.B. Eltsov, N.B. Kopnin, M. Krusius, L. Skrbek, M. Tsubota, and G.E. Volovik, [*Nature*]{}, [**424**]{}, 1022 (2003).
B. Hof, A. Juel, and T. Mullin, Scaling of the turbulence transition threshold in a pipe, [*Phys. Rev. Lett.*]{}, [**91**]{}, 24452 (2003); T. Mullin and J. Peixinho, Transition to turbulence in pipe flow, [*J. Low Temp. Phys.*]{}, this issue.
A.P. Finne, S. Boldarev, V.B. Eltsov, and M. Krusius, Measurement of turbulence in superfluid $^3$He-B, [*J. Low Temp. Phys.*]{}, [**136**]{}, 249 (2004).
W.F. Vinen, An introduction to quantum turbulence, [*J. Low Temp. Phys.*]{}, this issue.
A.P. Finne, S. Boldarev, V.B. Eltsov, and M. Krusius, Vortex formation in neutron-irradiated rotating superfluid $^3$He-B, [*J. Low Temp. Phys.*]{} [**135**]{}, 479 (2004).
R.E. Solntsev, R. de Graaf, V.B. Eltsov, R. Hänninen, and M. Krusius, Dynamic remnant vortices in superfluid $^3$He-B, Proc. Quantum Fluids and Solids Conf. 2006, [*J. Low Temp. Phys.*]{} (2007), cond-mat/0607323.
A.P. Finne, V.B. Eltsov, G. Eska, R. Hänninen, J. Kopu, M. Krusius, E.V. Thuneberg, and M. Tsubota, Vortex multiplication in applied flow: a precursor to superfluid turbulence, [*Phys. Rev. Lett.*]{} [**96**]{}, 085301 (2006); and to be published.
N.B. Kopnin, [*Theory of nonequilibrium superconductivity*]{} (Clarendon Press, Oxford, 2001).
G.E. Volovik, [*The universe in a helium droplet*]{} (Clarendon Press, Oxford, 2003).
R. Combescot, Ultracold Fermi gases: the BEC-BCS crossover, [*J. Low Temp. Phys.*]{}, this issue.
V.S. L’vov, S.V. Nazarenko, and G.E. Volovik, Energy spectra of developed superfluid turbulence, [*JETP Lett.*]{} [**80**]{}, 479 (2004); W.F. Vinen, Theory of quantum grid turbulence in superfluid $^3$He-B, [*Phys. Rev.*]{} B[**71**]{}, 24513 (2005).
V.S. L’vov, S.V. Nazarenko, and L. Skrbek, Energy spectra of developed turbulence in helium superfluids, [*J. Low Temp. Phys.*]{}, this issue, nlin.CD/0606002.
G.E. Volovik, Horizons and ergoregions in superfluids, [*J. Low Temp. Phys.*]{}, this issue, gr-qc/0603093.
See also a more extensive review in preparation: A.P. Finne, V.B. Eltsov, R. Hänninen, N.B. Kopnin, J. Kopu, M. Krusius, M. Tsubota and G.E. Volovik, Novel hydrodynamic phenomena in superfluid $^3$He, cond-mat/0606619.
Lord Kelvin (Sir William Thomson), [*Mathematical and physical papers*]{}, Vol. 4, [*Hydrodynamics and general dynamics*]{} (Cambridge University Press, 1910).
Lord Rayleigh (J.W. Strutt), [*Scientific papers*]{}, Vol. 1 (Cambridge University Press, 1899).
L.D. Landay and E.M. Lifshitz, [*Fluid Mechanics*]{} (Pergamon Press, Oxford, UK, 2nd edition, 1987).
G.E. Volovik, [*JETP Lett.*]{} [**75**]{}, 418Ê (2002) and [*JETP Lett.*]{} [**76**]{}, 240 (2002).
D. A. Abanin, [*JETP Lett.*]{} [**77**]{} 191 (2003).
R. Blaauwgeers, V.B. Eltsov, G. Eska, A.P. Finne, R.P. Haley, M. Krusius, J.J. Ruohio, L. Skrbek, and G.E. Volovik, Shear flow and Kelvin-Helmholtz instability in superfluids, [*Phys. Rev. Lett.*]{} [**89**]{}, 155301 (2002).
T. Ruokola and J. Kopu, Kelvin-Helmholtz instability in anisotropic superfluids, [*JETP Lett.*]{} [**81**]{}, 634 (2005); and to be published.
R. Schützhold and W.G. Unruh, [*Phys. Rev. D*]{} [**66**]{}, 044019 (2002).
M. Abid and A. Verga, Vortex sheet dynamics and turbulence, physics/0607108.
R. Hänninen, R. Blaauwgeers, V.B. Eltsov, A.P. Finne, M. Krusius, E.V. Thuneberg, and G.E. Volovik, Structure of the surface vortex sheet between two rotating $^3$He superfluids, Phys. Rev. Lett. [**90**]{}, 225301 (2003).
T.D.C. Bevan, A.J. Manninen, J.B. Cook, H. Alles, J.R. Hook and H.E. Hall, [*J. Low Temp. Phys.*]{} [**109**]{}, 423 (1997).
E.B. Sonin, Vortex oscillations and hydrodynamics of rotating superfluids, [*Rev. Mod. Phys.*]{} [**59**]{}, 87 (1987).
L. Skrbek, Flow phase diagram for the helium superfluids, [*JETP Lett.*]{} [**80**]{}, 484 (2004).
V.B. Eltsov, R. Blaauwgeers, N.B. Kopnin, M. Krusius, J.J. Ruohio, R. Schanen, and E.V. Thuneberg, Transitions from vortex lines to sheets: Interplay of topology and dynamics in an anisotropic superfluid, [*Phys. Rev. Lett.*]{} [**88**]{}, 065301 (2002).
V.M.H. Ruutu, J. Kopu, M. Krusius, Ü. Parts, B. Plaçais, E.V. Thuneberg, and W. Xu, Criitical velocity of vortex nucleation in rotating superfluid $^3$He-A, [*Phys. Rev. Lett.*]{} [**79**]{}, 5058 ((1997); J. Kopu and E.V. Thuneberg, One-dimensional textures and critical velocity in superfluid $^3$He-A, [*Phys. Rev.*]{} B[**62**]{}, 12374 (2000).
M.Tsubota, T. Araki, and C.F. Barenghi, Rotating superfluid turbulence, [*Phys. Rev. Lett.*]{} [**90**]{}, 205301 (2003). Based on measurements by: C.E. Swanson, C.F. Barenghi, and R.J. Donnelly, Rotation of a tangle of quantized vortex lines in He-II, [*Phys. Rev. Lett.*]{} [**50**]{}, 190 (1983).
V.B. Eltsov, A.P. Finne, R. Hänninen, J. Kopu, M. Krusius, M. Tsubota and E.V. Thuneberg, Twisted vortex state, [*Phys. Rev. Lett.*]{} [**96**]{}, 215302 (2006).
A.P. Finne, V.B. Eltsov, R. Blaauwgeers, M. Krusius, Z. Janu, and L. Skrbek, Time-of-flight measurements on quantized vortex lines in rotating $^3$He-B, [*J. Low Temp. Phys.*]{} [**134**]{}, 375 (2004).
R.M. Ostermeyer and W.I. Glaberson, Instability of vortex lines in the presence of axial normal flow, [*J. Low Temp. Phys.*]{} [**21**]{}, 191 (1975); [*Phys. Rev. Lett.*]{} [**33**]{}, 1197 (1974).
M. Kobayashi and M. Tsubota, Thermal dissipation in quantum turbulence, cond-mat/0607434.
N.G. Parker and C.S. Adams, Emergence and decay of turbulence in stirred atomic Bose-Einstein condensates, [*Phys. Rev. Lett.*]{} [**95**]{}, 145301 (2005); V. Schweikhard, I. Coddington, P. Engels, S. Tung, and E.A. Cornell, Vortex-lattice dynamics in rotating spinor Bose-Einstein condensates, [*Phys. Rev. Lett.*]{} [**93**]{}, 210403 (2004).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We explore the dynamics of cosmological models with two coupled dark components with energy densities $\rho_A$ and $\rho_B$ and constant equation of state (EoS) parameters $w_A$ and $w_B$. We assume that the coupling is of the form $Q=H\, q(\rho_A,\rho_B)$, so that the dynamics of the two components turns out to be scale independent, i.e. does not depend explicitly on the Hubble scalar $H$. With this assumption, we focus on the general linear coupling $q=q_o +q_A\,\rho_A+q_B\,\rho_B$, which may be seen as arising from any $ q(\rho_A,\rho_B)$ at late time and leads in general to an effective cosmological constant. In the second part of the paper we consider observational constraints on the form of the coupling from SN Ia data, assuming that one of the components is cold dark matter (CDM), i.e. $w_B=0$, while for the other the EoS parameter can either have a standard ($w_A>-1$) or phantom ($w_A<-1$) value. We find that the constant part of the coupling function is unconstrained by SN Ia data and, among typical linear coupling functions, the one proportional to the dark energy density $\rho_{A}$ is preferred in the strong coupling regime, $|q_{A}|>1$. Models with phantom $w_A$ favor a positive coupling function, increasing $\rho_A$. In models with standard $w_A$, not only a negative coupling function is allowed, transferring energy to CDM, but the uncoupled sub-case falls at the border of the likelihood.'
author:
- Claudia Quercellini
- Marco Bruni
- Amedeo Balbi
- Davide Pietrobon
bibliography:
- 'Bcoup.bib'
title: 'Late universe dynamics with scale-independent linear couplings in the dark sector'
---
Introduction {#intro}
============
The overall density of the observed universe, the growth of structures and their clustering properties cannot be explained by known forms of matter and energy [@2006Natur.440.1137S; @Khalil:2002]. In addition, cosmic microwave background (CMB) anisotropy observations show that the total density is close to critical, so that the gap between known and unknown cannot be accounted for by curvature [@Spergel:astro-ph/0603449; @Dunkley:2008]. Finally, observations of type Ia Supernovae (SNe) [@Riess:1998; @Perlmutter:1999; @Riess:astro-ph/0611572], baryon acoustic oscillations (BAO) [@Eisenstein:astro-ph/0501171; @Percival:2007p358], and integrated Sachs-Wolfe (ISW) effect[@Pietrobon:2006; @Giannantonio:2008] tell us that the universe expansion is currently accelerating. To explain these facts cosmologists need to assume the existence of a dark sector in the theory [@Copeland:2006], whose general properties have then to be tested against observations, i.e. with parameters that have to be deduced from indirect evidences. One possibility is that the dark sector is accounted for, partly or in full, by a modified gravity theory[^1], while a more conventional approach is to assume that gravity is well described by general relativity, with the dark sector made up of an unusual energy momentum tensor.
In the currently prevailing scenario, the dark sector consists of two distinct contributions. One component, [*cold dark matter*]{} (CDM), accounts for about one third of the critical density [@Percival:2007p1293] and is needed to explain the growth of inhomogeneities that we observe up to very large scales, as well as a host of other cosmological observations which goes from galactic scales, to clusters of galaxies, to redshift surveys. The other contribution, dubbed [*dark energy*]{}, accounts for the remaining two thirds of the critical density, and is needed to explain the observed late time acceleration of the universe expansion [@Perlmutter:1999; @Riess:astro-ph/0611572]. CDM can be modeled as a pressureless perfect fluid, representing unknown heavy particles, collisionless and cold, i.e. with negligible velocity dispersion. In its simplest form, dark energy consists of vacuum energy density, i.e. a cosmological constant $\Lambda$. Taken together, $\Lambda$ and CDM make up for the the so-called concordance $\Lambda$CDM model [@Spergel:2003; @Tegmark:2004]. This simple model fits observations reasonably well, but lacks a sound explanation in terms of fundamental physics, and a number of alternatives have been proposed. In general, dark energy can be modeled as a perfect fluid with an equation of state (EoS from now on) that violates the strong energy condition [@Visser:1997aa], such that it can dominate at late times and have sufficiently negative pressure to account for the observed accelerated expansion. Scalar fields can also be formally represented as perfect fluids (see e.g. [@Bruni:1992p718] and refs. therein). In a more exotic version, dubbed [*phantom energy*]{} [@Caldwell:2002p1819; @Caldwell:2003p1814], the EoS also violates the null energy condition [@Visser:1997aa], leading to the growth in time of the energy density with the cosmic expansion. Finally, another rather radical alternative to $\Lambda$CDM is to assume a single unified dark matter (UDM), able to mimic the essential features of $\Lambda$CDM that are necessary to build a viable cosmology. For example, in [@Balbi:astro-ph/0702423] we have considered observational constraints on a UDM model with an “affine" EoS, i.e. such that the pressure satisfies the affine relation $P=P_o+\alpha\rho$ with the energy density [@Ananda:astro-ph/0512224; @Ananda:2006]. This model is a one parameter ($\alpha$) generalization of $\Lambda$CDM, with the latter recovered for $\alpha=0$. There is no need to assume [*a-priori*]{} a $\Lambda$ term in Einstein equations, because the EoS $P=P_o+\alpha\rho$ leads to an effective cosmological constant with $\Omega_\Lambda =-8\pi G P_o/[3 H^2_o (1+\alpha)]$. The problem is thus shifted from justifying a $\Lambda$ term in Einstein equations to that of justifying the assumed EoS: a possible justification of this affine model can be given in terms of scalar fields, either of quintessence or k-essence type [@Quercellini:2007]. This type of model escapes typical constrains on many UDM models [@Sandvik:2004p316] (but cf. e.g. [@Gorini:2007]) because, for a given homogeneous isotropic background expansion, it allows multiple phenomenological choices for the speed of sound of the perturbations [@Pietrobon:2008].
In models of the dark sector consisting of two components, dark matter and dark energy are usually assumed to interact only through gravity, but they might exhibit other interactions without violating observational constraints [@Kunz:2007p2562]. Exploiting this degeneracy, here we depart from the standard scenario, and assume a cosmological model where the dark sector is made up of two coupled dark components, each described as a perfect fluid with its own constant EoS parameter $w$. This choice allows for the possibility that the observed evolution of the universe, although reasonably well explained by the $\Lambda$CDM model, is actually due to the dynamics of two rather general coupled components, possibly alleviating the so-called “coincidence problem", $\Omega_\Lambda \approx \Omega_{CDM}$, typical of the standard model [@Copeland:2006].
In this paper, our first aim is to characterise the dynamics of our cosmological model with the two general coupled components, taking into account general forms of interaction, parameterized in terms of a late time function $Q$ linear in the energy densities, Eqs. (\[Coup\]-\[coupling\]). To this end we will use standard dynamical system techniques [@KArrowsmith:1992; @Wainwright:1997], which are now rather common in the analysis of cosmological models, see e.g. [@Wands:1993; @Bruni:1993; @Amendola:1993; @Bruni:1994; @Bruni:1995a; @Bruni:1995b] and [@Copeland:1998p1287; @Ananda:astro-ph/0512224; @Ananda:2006; @Bohemer:2008]. To our knowledge, such an exhaustive analysis has not been carried out yet, although several sub-cases have been considered [@Majerotto:2004; @Olivares:2006; @Guo:2007; @Bohemer:2008; @Quartin:2008; @pettorino:2008; @Barrow:2006]. In our study, we restrict ourselves to the evolution of a homogeneous, isotropic cosmological background, leaving aside the question of what the effects of coupling could be in anisotropic models [@Ananda:2006], or when general perturbations are present [@Valiviita:2008p2415; @Dunsby:1992p713]. It is however worth noticing that, thanks to the particular form of coupling we choose, our analysis of the dynamics of the two components is valid in any theory of gravity, because is based only on the conservation equations, and not on specific field equations.
Secondly, as a way to gain some physical insight on the likelihood of some specific coupling models, we also explore the constraints on the predicted luminosity distance modulus derived from type Ia Supernovae observations, using a Monte Carlo Markov Chain (MCMC) approach. Needless to say, this is not intended as a full-fledged cosmological parameter estimation for these models, but only as a first exploration of the parameter space to rule out those models which are manifestly in contrast with observations. This analysis requires the use of the Friedmann equation, hence general relativity is assumed as the valid theory of gravity.
The paper is organized as follows. In Section II we study the general dynamics and solve the equations for the evolution of the coupled dark components; in Section III we focus on the cosmological effects of changing the parameters of the coupling term; in Section IV we derive constraints from the observation of type Ia SNe on some specific sub-classes of models; finally, in Section V, we present the main conclusions of our work.
dynamics of dark components {#coupledfluids}
===========================
Linear scale-free coupling
--------------------------
In general relativity, assuming a a flat Robertson-Walker universe, the dynamics is subject to the Friedmann constraint $$\label{einstein1}
H^2=\frac{8\pi G}{3}\,\rho_T\,,$$ where $\rho_T$ is the total energy density of the various components. Beside baryons and radiation, $\rho_T$ includes any other component contributing to the dark sector, i.e. that part of the total energy-momentum tensor that in the context of general relativity is needed to explain the observed universe, in particular the CMB [@Spergel:astro-ph/0603449; @Dunkley:2008], structure formation [@2006Natur.440.1137S; @Khalil:2002] and the late time acceleration of the expansion [@Riess:1998; @Perlmutter:1999; @Riess:astro-ph/0611572; @Eisenstein:astro-ph/0501171; @Percival:2007p358; @Pietrobon:2006; @Giannantonio:2008]. The dynamics itself is described by the evolution of the Hubble expansion scalar $H=\dot{a}/a$, given by the Raychaudhuri equation $$\label{ray}
\dot{H}=-H^2 -\frac{4\pi G}{3} (1+3\, w_T)\rho_T.$$ This is coupled to the evolution equations for the energy density of each of the matter components contributing to $\rho_T$. Since $\dot{H}+H^2=\ddot{a}/a$, with $a(t)$ the usual metric scale-factor (which we assume normalized to its present value), acceleration is achieved whenever $w_T=P_T/\rho_T<-1/3$, as it is well known.
The standard $\Lambda$CDM model assumes two dark components: the pressureless cold dark matter (CDM), with $w_{DM}=0$, and the cosmological constant $\Lambda$ with $w_\Lambda=-1$. CDM is needed to fill the gap between the baryon abundance and the amount of matter that is needed to explain the rotation curve of galaxies and structure formation in general, as well as to allow for a vanishing curvature model. In the context of general relativity, and under the Robertson-Walker homogeneus and isotropic assumption (see e.g. [@Celerier:2007] for alternatives), a cosmological constant $\Lambda$ is the simplest possible form of dark energy (DE) needed to generate the observed low redshift acceleration. While this simple scenario is preferred from the point of view of model comparison and selection [@Balbi:astro-ph/0702423], because of the low number of parameters, from a theoretical perspective is oversimplified, and it is worth exploring alternatives, even if purely phenomenological.
Here we shall consider two general coupled dark components with energy densities $\rho_{A}$ and $\rho_{B}$. Since we want to introduce a rather general type of coupling, focusing our analysis on its effects, we shall assume the simplest possible form for the EoS of these two dark components, i.e. we will assume that the EoS parameters $w_{A}$ and $w_{B}$ are constant. On the other hand, we shall not [*a priori*]{} restrict our study to the sub-class of models where one of the two components represents CDM with, for instance, $w_B=0$.
Due to the presence of the coupling, the two dark components satisfy the balance equations $$\begin{aligned}
\label{cons}
\dot{\rho_{A}}+3H(1+w_{A})\rho_{A}&=&Q\\
\label{cons2}
\dot{\rho_{B}}+3H(1+w_{B})\rho_{B}&=&-Q\,.\end{aligned}$$ Even assuming the linear form for the coupling $Q$ given in (\[Coup\]) and (\[coupling\]) below, this model allows us to explore a large number of alternatives. Here we will focus on models for the homogeneous and isotropic background expansion, assuming that for those models that will fit current observational data it might always be possible to construct an appropriate perturbative scheme allowing for structure formation, for instance by assuming a vanishing effective speed of sound in one component.
The coupled dark components $\rho_A$ and $\rho_B$ could be in principle be taken to represent DE only, i.e. they could be two extra dark components contributing to $\rho_T$ in (\[einstein1\]), [*in addition* ]{} to CDM. Leaving aside this possibility, and ignoring baryons and radiation as we will do in this Section, the sum of Eqs. (\[cons\]-\[cons2\]) gives the conservation equation for $\rho_T=\rho_{B}+\rho_{A}$. A positive coupling term $Q$ corresponds to a transfer of energy from $\rho_{B}$ to $\rho_{A}$, and vice versa, but in general $Q$ doesn’t need to have a definite sign.
An interaction term between two components has been considered several times in literature, starting from Wetterich [@Wetterich:1998; @Wetterich:1995] and Wands et al. [@Wands:1993; @Copeland:1998p1287] in scalar field models, and has been analysed by Amendola in dark energy models [@Amendola:2000; @Amendola:2001; @Amendola:2003; @Amendola:2004], and for example recently in [@Majerotto:2004; @Olivares:2006; @Guo:2007; @Bohemer:2008; @Quartin:2008; @pettorino:2008; @Delamacorra:2008; @Manera:2007].
The coupling term $Q$ can take any possible form $Q=Q(H,\rho_A,\rho_B,t)$. Here we shall consider the case of an autonomous ($t$ independent) coupling with a factorized $H$ dependence $$\label{Coup}
Q=\frac{3}{2} H q(\rho_A,\rho_B).$$ As we shall see below, with this assumption the effects of the coupling on the dynamics of $\rho_A$ and $\rho_B$ become effectively independent from the evolution of the Hubble scale $H$. For this reason, we may call this a “scale-independent" coupling. Furthermore, with the decoupling of the dynamics of the two dark components from that of $H$, the analysis of the next section is valid in any theory of gravity, because it is based on the conservation equations only: we don’t need to use (\[einstein1\])-(\[ray\]), i.e. the field equations of general relativity. Finally, we note that any coupling of this type can be approximated at late times by a linear expansion: $$\label{coupling}
q=q_{0}+q_{A}\rho_{A}+q_{B}\rho_{B}\;,$$ where $q_A, q_B$ are dimensionless coupling constants, and $q_0$ is a constant coupling term with dimensions of an energy density[^2]. In the following we shall analyse the dynamics arising from this general linear scale-independent coupling. Obvious sub-cases are: $q\propto \rho_T$ ($q_0=0, q_A=q_B$); $q\propto\rho_A$ ($q_0=0, q_B=0$); etc. We will come back to this in more detail in the next Section. Linear couplings have been frequently analysed in literature ([@Wetterich:1995; @Amendola:2000; @Majerotto:2004; @Guo:2007],[@Multa:2007; @Mainini:2007; @Bohemer:2008; @Quartin:2008]) both for mathematical simplicity, because they retain the linearity of system (\[cons\]-\[cons2\]) with no coupling, and because they can arise from string theory or Brans-Dicke-like Lagrangians after a conformal transformation of the metric.
Analysis of the scale-free linear dynamics {#stab}
------------------------------------------
### The linear dynamical system
In order to proceed with the analysis of the dynamics of the dark components, let us change variables, using the total density $\rho_{T}=\rho_{B}+\rho_{A}$ and the difference $\Delta=\rho_{B}-\rho_{A}$. We also set $$\begin{aligned}
w_{+}=(w_{B}+w_{A})/2\,, & ~~ & w_{-}=(w_{B}-w_{A})/2\,, \\
q_{+}=(q_{B}+q_{A})/2\,, & ~~ & q_{-}=(q_{B}-q_{A})/2\,.\end{aligned}$$ One reason for this choice is that ultimately the evolution of $\rho_T$ is the one that governs the general expansion law through (\[einstein1\]) and (\[ray\]). In addition, thanks to the particular form of the coupling (\[Coup\]) and assuming $H>0$, the dynamics can be made explicitly scale-independent, eliminating $H$ by adopting $N=\ln{(a)}$, the [*e*]{}-folding, as the independent variable. Then, denoting with a prime the derivative with respect to $N$, the system (\[cons\]-\[cons2\]) is transformed into $$\begin{aligned}
\label{constot}
\rho_{T}'+3\rho_{T}(1+w_{+})+3w_{-}\Delta&=&0 \qquad\qquad\qquad\qquad\qquad\quad\\
\label{cons2tot}
\Delta'+3\Delta(1+w_{+})+3w_{-}\rho_{T}&=&-3(q_{+}\rho_{T}+q_{-}\Delta+q_{0}).\end{aligned}$$ An effective EoS parameter $w_{eff}$ is implicitly defined from Eq. (\[constot\]): when $w_-=0$ the two EoS coincide giving rise to a constant $w_{eff}=w_+=w_{A}$ and $\rho_T$ scales accordingly, as a standard barotropic perfect fluid, but in general $$\label{weff}
w_{eff}=w_+ +w_-\frac{\Delta}{\rho_T}$$ changes with time. Notice that we can also define, using (\[Coup\]-\[coupling\]) in (\[cons\]), effective EoS parameters for the two components: $$\begin{aligned}
\label{wA}
w_{Aeff} & = & w_A -\frac{q_0+q_B\rho_B}{2\rho_A} -\frac{q_A}{2} \,,\\
w_{Beff} & = & w_B +\frac{q_0+q_A\rho_A}{2\rho_B} +\frac{q_B}{2} \,.
\label{wB}\end{aligned}$$ From now on we will characterize the cosmological evolution of any of the energy densities as standard/phantom behaviour. As mentioned in the introduction, standard/phantom respectively correspond to an energy density which is either a decreasing or an increasing function of time (the scale factor or the [*e*]{}-folding [*N*]{}). The phantom behaviour arises in the presence of coupling from an effective EoS parameter $<-1$, which corresponds to the violation of the null energy condition [@Visser:1997aa] for that given energy density. Thus, it follows from (\[weff\]) and (\[wA\]-\[wB\]) that we can have a phantom behaviour in the total energy density $\rho_T$ as well as in one or both of the single components $\rho_A$ and $\rho_B$, and that in principle the effective EoS parameter of each of these can pass through the $-1$ value, from phantom to standard or vice versa. On the other hand, we will also refer to constant parameters such as $w_A$ and $w_B$ as having a standard/phantom value, respectively $w_A>-1$ or $w_A<-1$, because the corresponding fluid would evolve in that way in the case of no coupling.
We will also refer to an “affine” evolution. As said in the introduction, for an uncoupled component with energy density $\rho$ this arises from an affine EoS of the form $P=P_o +\alpha\rho$. Inserted in the energy conservation equation this leads to $$\label{affine}
\rho=\rho_{\Lambda}+\rho_{0M}a^{-3(1+\alpha)}\, .$$ Therefore, starting from the Friedmann equations (\[einstein1\]-\[ray\]) with no cosmological constant term, the affine EoS and energy conservation lead to an effective cosmological constant $\rho_{\Lambda}$ plus an effective matter-like component with constant EoS parameter $\alpha$ (a barotropic perfect fluid) and today’s density $\rho_{0M}$ (cf. [@Ananda:astro-ph/0512224; @Ananda:2006; @Balbi:astro-ph/0702423; @Quercellini:2007] for a detailed analysis of the cosmological dynamics arising in this case). As we will see, it turns out that there are solutions of the system (\[constot\]-\[cons2tot\]) that evolve according to (\[affine\]).
In order to proceed with the analysis of Eqs. (\[constot\]-\[cons2tot\]) using standard dynamical system techniques [@KArrowsmith:1992], it is convenient to write it as $$\label{Xprime}
\bf{X}^\prime= { \bf J}\, {\bf X} +{\bf C}\,,$$ where the phase-space state vector $\bf X$ and the constant $\bf C$ are $$\label{ defs}
{\bf X}=\left(\begin{array}{c}
\rho_T \\
\Delta
\end{array}\right)\,, ~~~{\bf C}=\left(\begin{array}{c}
0 \\
-3q_0
\end{array}\right)\,,$$ and the matrix of coefficients ${\bf J}$ is given by $$\label{J}
{\bf J}=\left(\begin{array}{rr}
-3(1+w_+) & -3w_- \\
-3(w_-+q_+) & -3(1+w_++q_-)
\end{array}\right) \;.$$ Fixed points, if they exist, are solutions ${\bf X}_*$ of the equation $ { \bf J}\, {\bf X}_* +{\bf C}={\bf 0}$ and, given that the system (\[Xprime\]) is linear, $\bf J$ is also the Jacobian of the system at these fixed points. These fixed points correspond to constant values of $\rho_{T}$ and $\Delta$ and in turn of $\rho_{A}$ and $\rho_{B}$, that is to the emergence of an effective cosmological constant (when $\rho_T\not =0$, see below). Every constant form of energy is indeed alike the cosmological constant $\Lambda$, and plays exactly the same cosmological role: when dominates the evolution of the background, it drives an exponentially accelerated expansion, with an effective EoS parameter close to $-1$. Therefore, in this Section we will focus on the analysis of these fixed points.
Notice that - unlike the case with no coupling - there is no [*a priori*]{} guarantee from the equations above that $\rho_A$ and/or $\rho_B$, as well as $\rho_T$, will always be non-negative. However, one has to keep in mind that $\rho_T$ must be non-negative because of the Friedmann constraint (\[einstein1\]). This means that if $\rho_T$ is vanishing for some value of $N$ ($a$), then at that point the assumption $H>0$, on the basis of which Eq. (\[Xprime\]) is derived, is violated, and the solutions of (\[Xprime\]) no longer correspond to solutions of the original coupled system of Eqs. (\[ray\]) and (\[cons\]-\[cons2\]).
### Fixed points and stability analysis
Even if system (\[Xprime\]) is linear, many different possibilities arise from the fact that it depends on five parameters[^3]. There are two main cases, which we are now going to unfold.
[**Case 1:**]{} $\det({\bf J})\not =0$. In this case ${\bf J}^{-1}$ exists and there is a unique fixed point ${\bf X}_*=-{\bf J}^{-1} {\bf C}$. This can either be the origin in phase space, i.e. $\rho_{T*}=\Delta_*=0$, when $q_0=0$ (${\bf C}={\bf 0}$), or else this fixed point represents an effective cosmological constant, with $\rho_{T*}=\rho_\Lambda$, $\Delta_*=\Delta_\Lambda$ respectively given by $$\begin{aligned}
\label{effL}
\rho_\Lambda& = & 9\ \frac{w_-\ q_0}{\det({\bf J})}\,, \\
\Delta_\Lambda & = & -9\ \frac{(1+w_+)\ q_0}{\det({\bf J})}\,,
\label{effD}\end{aligned}$$ where $$\label{detJ}
\det({\bf J}) =9\left[(1+w_+) (1+w_+ +q_-) -w_-(w_- +q_+)\right],\qquad$$ with corresponding values for $\rho_{A\Lambda}$ and $\rho_{B\Lambda}$.
In order to analyse the stability properties of system (\[Xprime\]) at this fixed point, we now consider the eigenvalues of $\bf J$. Given that $\det({\bf J})\not =0$, there are two non-zero eigenvalues, given by $$\label{eigenvals}
\lambda_{\pm} = \frac{{\rm tr}( {\bf J}) }{2} \pm\sqrt{D}\,,$$ where $$\begin{aligned}
\label{trJ}
{\rm tr}( {\bf J}) & = & -3\left[2(1+w_+) +q_-\right]\,, \\
\label{D}
D & = &\left(\frac{{\rm tr}( {\bf J}) }{2}\right)^2 -\det({\bf J})= \\
&=&9\left[ \left( \frac{ q_{-} }{2} \right)^{2}+w_{-}(q_{+}+w_{-}) \right]\,,\end{aligned}$$ and the value of the discriminant $D$ determines the three possible Jordan canonical forms of $\bf J$. These three possible cases, to be further clarified in the next subsection, correspond to $D>0$, $D=0$, $D<0$, and are summarized below, with their sub-cases. [**Case 1a**]{}: real distinct eigenvalues ($D>0$). There are three sub-cases: [*i)*]{} $\lambda_- < \lambda_+ <0$, the fixed point is a stable node and both dark components have a standard behaviour, with decreasing energy densities; [*ii)*]{} $\lambda_+>\lambda_->0$, the fixed point is an unstable node and both dark components have a phantom behaviour, with increasing energy densities; [*iii)*]{} $\lambda_+>0>\lambda_-$, which requires $\det({\bf J})<0$; the fixed point is a saddle and both dark components have a first standard phase with decreasing energy density followed by a phantom phase. The connection between the phantom behaviour and sign of the eigenvalues will be rendered evident in the next Section, see Eq. (\[rhotot\]) and related comments.
[**Case1b**]{}: real equal eigenvalues ($D=0$). In this case $\lambda_+=\lambda_-=\lambda_0$ and the fixed point is an improper node, either stable (both dark components are standard) or unstable (both dark components are phantom). In both cases the two dark components, as well as $\rho_T$ and $\Delta$, follow a sort of affine evolution, with a modification term, see Eq. (\[improper\]).
[**Case 1c**]{}: complex eigenvalues ($D<0$). There are three sub-cases: [*i)*]{} if ${\rm tr}({\bf J})=0$ the fixed point is a centre; [*ii)*]{} if ${\rm tr}({\bf J})>0$ the fixed point is an unstable spiral; [*iii)*]{} if ${\rm tr}({\bf J})<0$ the fixed point is a stable spiral. This last case is the most interesting, with $\rho_T$ converging to an effective cosmological constant via a series of oscillations. Notice that this case has to be dealt with care, as in the past $\rho_T=0$ ($H=0$) at some point, and the time reversal of system (\[constot\]-\[cons2tot\]) should be considered prior to that.\
[**Case 2:**]{} $\det({\bf J}) =0$. In this case if $q_0\not =0$ (${\bf C} \not = {\bf 0}$) the system of linear equations $ { \bf J}\, {\bf X}_* +{\bf C}={\bf 0}$ is in general inconsistent and there are no fixed points. Alternatively, if $q_0 =0$ (${\bf C} = {\bf 0}$) there is an infinite number of fixed points, each of them representing a possible asymptotic state depending on the initial conditions. These infinite number of fixed points is represented by a straight line in phase space, corresponding to a conserved quantity for system (\[Xprime\]). It turns out that there are three possible combinations of the four parameters $w_+, w_-, q_+, q_-$ that give $\det({\bf J}) =0$ and they are shown in Table \[detJ0\], classified as [**Case 2a, 2b**]{} and [**2c**]{}. For [**Case 2b**]{}, $\rho_{T*}=0$ always, and either also $\Delta_{*}=0$ and then both $\rho_{A*}=\rho_{B*}=0$, or $\Delta_{*}\not =0$ and then $\rho_{A*}=-\rho_{B*}=\Delta_*/2$, so that one of the two is negative. Being the energy densities either null or negative, we can conclude that [**Case 2b**]{} corresponds to a non-physical situation. [**Case 2a**]{} and [**2c**]{} are more interesting, and are summarized in Table \[detJ0FP\]. In both cases, each fixed point on the line represents an effective cosmological constant. In particular, [**Case 2a**]{} corresponds to $w_{A}=w_{B}=-1$, i.e. two cosmological constant-like components whose energy densities scale in a different way because of the coupling. In addition, it turns out that for this very peculiar case $q_{0}$ can be non-zero.
Regarding the stability analysis, in [**Case 2**]{} $\det({\bf J})=0$ implies that one of the eigenvalues is null, namely $\lambda_-=0$ if ${\rm tr}( {\bf J})>0$ (and vice versa) (cfr. Eqs. (\[eigenvals\]) and (\[D\])). This implies that the total energy density, as well as the single dark components and $\Delta$, follow the affine evolution (\[affine\]); we will comment further on this below Eq. (\[rhotot\]). In this case, the non zero eingenvalue $\lambda_{-/+}$ for each sub-case is given in Table \[detJ0\]. Table \[detJ0FP\] gives instead the values the fixed points and the conditions for positive effective cosmological constants. Each of these fixed points is characterized by $\rho_{\Lambda}\propto \Delta_{\Lambda}$, which is equivalent to $\rho_{B\Lambda}\propto \rho_{A\Lambda}$. Explicitly, for [**Case 2a**]{} the energy densities are related by $\rho_{B\Lambda}=-\rho_{A\Lambda}q_{A}/q_{B}$, while for [**Case 2c**]{} $\rho_{B\Lambda}=-\rho_{A\Lambda}(1+w_{A})/(1+w_{B})$. Note that the same proportionality law holds for the fixed point of [**Case 1**]{} (cf. Eq. (\[effL\]-\[effD\])). This behaviour mimics that of scaling solutions [@Wands:1993; @Copeland:1998p1287], whose phase space typically admits fixed points where the contributions of the two fluids to the total energy density are constant (we will come back on this in Sec. \[critpoints\]). It is easy to verify that at all these fixed points the effective EoS parameter (\[weff\]) has the value $w_{eff}=-1$.
Case parameters $\lambda_{+/-}$
------ ---------------------------------------------------------------------------------------------- ----------------------
2a $w_-=0$, $w_+=-1$, $\forall q_+$, $\forall q_-$ $-3q_-$
2b $w_-=0$, $q_-=-(1+w_+)$, $\forall w_+$, $\forall q_+$ $-3(1+w_+)$
2c $q_+=\frac{(1+w_+)(1+w_+ + q_-)-w_-^2 }{w_-}$, $w_-\not =0$, $\forall w_+$, $\forall q_-$ $-3[2(1+w_+) + q_-]$
Case FPs $\rho_{A\Lambda}>0$ $\rho_{B\Lambda}>0$ $\rho_{A\Lambda}>0$, $\rho_{B\Lambda}>0$
------ ---------------------------------------------------- --------------------- --------------------- ----------------------------------------------------------------------------------
2a $\rho_{\Lambda}=-\frac{q_-}{q_+} \Delta_{\Lambda}$ $q_+>-q_-$ $q_+<q_-$ $\Delta_\Lambda >0$ and $-1<q_+/q_-<0$, or $\Delta_\Lambda <0$ and $0<q_+/q_-<1$
2c $\rho_{\Lambda}=-\frac{ \Delta_\Lambda}{R}$ $R>-1$ $ R<1$ $\Delta_\Lambda> 0$ and $-1 <R<0$, or $\Delta_\Lambda< 0$ and $0 <R<1$
An equivalent equation and its solutions
----------------------------------------
While the classification of the various possible phase portraits of system (\[constot\]-\[cons2tot\]) is best given by the representation (\[Xprime\]) used in the previous Section, given that this is a linear system it is useful to consider the equivalent second order linear equation with constant coefficients, and its solutions. This helps interpreting the formalism of Sec. \[stab\] from a cosmological point of view. For this reason, we shall focus on the cases of main interest.
First, notice that if $w_-=0$ then (\[constot\]) decouples and implies $$\label{w-0}
\rho_T = \rho_c a^{-3(1+w_+)}\ ,$$ with $\rho_c$ an integration constant, so that the total energy density can either be standard or phantom. Eq. (\[cons2tot\]) can then be integrated, giving $$\label{delta0}
\Delta=\Delta_0 a^{-3(1+w_+ +q_-)} +\Delta_{c} a^{-3(1+w_+)} + \Delta_* \,$$ where $\Delta_0$ is an integration constant, $\Delta_c\propto \rho_c$ and $\Delta_* \propto q_0$. This case is of limited interest, because in general the two components will have a negative energy density in the past or in the future.
Assuming now $w_-\not =0$, we obtain from (\[constot\]-\[cons2tot\]) a single equation for the total energy density $$\begin{aligned}
\label{rhodiff}
&\rho_{T}'' \, -{\rm tr}({\bf J})\, \rho_{T}'+\det({\bf J})\, \rho_{T}=9w_{-}q_{0},\end{aligned}$$ with ${\rm tr}({\bf J})$ and $\det({\bf J})$ respectively given by Eqs. (\[trJ\]) and (\[detJ\]). It is noticeable that - when $\det({\bf J})\not=0$ - this equation admits a constant particular solution corresponding to the constant source term, given by the fixed point (\[effL\]). Eq. (\[effL\]) with $\det({\bf J})\not=0$ is exactly this constant particular solution, which is the effective cosmological constant term and it vanishes if $q_{0}=0$.
When the eigenvalues (\[eigenvals\]) of the characteristic polynomial of Eq. (\[rhodiff\]) are distinct, the general solution can be written as $$\begin{aligned}
\label{rhotot}
\rho_{T}=\rho_{T+}a^{-3(1+\beta_{+})}+\rho_{T-}a^{-3(1+\beta_{-})}+\rho_{\Lambda}.\end{aligned}$$ When the eigenvalues are equal, $\lambda_+=\lambda_-=\lambda_0$, we have the special [**Case 1b**]{} of the previous Section and, denoting $T= {\rm tr}({\bf J})$, we obtain: $$\label{improper}
\rho_{T}=[\rho_{T1} +\rho_{T2} \ln(a)]\ a^{\frac{T}{2} } + \rho_\Lambda\,.$$ Given that this evolution law is similar to the affine one, Eq. (\[affine\]), except for the $\ln(a)$ correction term, and that the fixed point of this [**Case 1b**]{} is an improper node, we can refer to this as an improper affine evolution. Here $\rho_{T+}$, $\rho_{T-}$ and $\rho_{T1}$, $\rho_{T2}$ are integration constants related to the present values of the energy densities $\rho_A$ and $\rho_B$ of the two dark components; they are not independent in a flat universe, but related by imposing that the total energy density has the critical value at present. The effective cosmological constant $\rho_\Lambda$ is given by (\[effL\]). In both cases (\[rhotot\]) and (\[improper\]) the expansion of the universe is governed by this total energy density that enters the Friedmann equations (\[einstein1\]-\[ray\]). The parameters $\beta_{\pm}$ are related to the eigenvalues $\lambda_\pm$ by $\lambda_\pm= -3(1+\beta_\mp)$. That is, $\beta_\pm=\beta_{0}\pm\sqrt{D}/3$, with $\beta_{0}=w_{+}+q_{-}/2=-{\rm tr}({\bf J})/6-1$ and $D$ given by Eq. (\[D\]). The case (\[improper\]) arises from $D=0$, with $\lambda_0=-3(1+\beta_0)={\rm tr}({\bf J})/2$.
It is interesting to note the role of the parameters $\beta_{\pm}$ in (\[rhotot\]), assuming they are real. For a barotropic perfect fluid with constant EoS parameter $w$ one has $\rho\propto a^{-3(1+w)}$. Hence, $\beta_{\pm}$ simply represent two constant effective EoS parameters that drive the evolution of the total energy density (\[rhotot\]). In other words, the simple coupling (\[Coup\]-\[coupling\]) of the two dark components with constant $w_A$ and $w_B$ in general produces a total energy density (\[rhotot\]) equivalent to that of two uncoupled fluids with constant EoS parameters $\beta_{\pm}$, plus the effective cosmological constant $\rho_\Lambda$ [^4].
Apart from the asymptotic effective cosmological constant $\rho_\Lambda$ arising when $q_0\not=0$ and $w_-\not=0$ ([**Case 1**]{} of the previous Section), it is also possible to have asymptotic effective cosmological constants when $\det({\bf J})=0$ ([**Case 2**]{} of the previous Section) if $q_0=0$, corresponding to the vanishing of one of the eigenvalues (\[eigenvals\]), i.e. $\lambda_{-/+}=0$ and to $\beta_{+/-}=-1$. There are three cases, given in Table \[detJ0\], each with the corresponding non vanishing eigenvalue. In this cases $\rho_T$, $\Delta$, $\rho_A$ and $\rho_B$ have the affine behavior (\[affine\]). Hence the acceleration is guaranteed by the component of the total energy density for which an effective EoS parameter $\beta_{+/-}$ assumes value $-1$. These fixed points seem less interesting however, because even when the non-zero eigenvalue is negative they are not general attractor of the dynamics (since the other eigenvalue is null): there is a different asymptotic non zero value of $\rho_T$ for each possible initial condition.
We have already commented above on [**Case 1b**]{}, giving (\[improper\]). We now consider the two other sub-cases of [**Case 1**]{} of the previous Section from the point of view of the solution (\[rhotot\]). In both cases $\Delta$, $\rho_A$ and $\rho_B$ evolve with the scale factor as $\rho_T$ in (\[rhotot\]), i.e. they are all linear combinations of the two normal modes of the system, $a^{-3(1+\beta_{+})}$ and $a^{-3(1+\beta_{-})}$, plus a constant term.
[**Case 1a**]{}. When both eigenvalues are real and distinct $\rho_T$ scales as it would if we were considering two decoupled components with EoS parameters $\beta_+$ and $\beta_-$, plus a cosmological constant. If $\det({\bf J})<0$ the fixed point is a saddle, i.e. unstable, with $\lambda_+>0$ and correspondingly $\beta_-<-1$. Consequently $\rho_T$ scales as a mixture of a standard component and a phantom component. When $\det({\bf J})>0$ either both $\beta_{+/-}>-1$, or vice versa, with corresponding standard/phantom behaviour.
[**Case 1c**]{}. If $\lambda_\pm$ are complex conjugates, then $\beta_{\pm}=\beta_{0}\pm i\sqrt{|D|}/3$ and the total energy density can be written as $$\begin{aligned}
\rho_{T}& = & a^{-3(1+\beta_{0})}\left\{(\rho_{T-}-\rho_{T+})\sin{\left[\sqrt{|D|}\ln(a)\right]}\right. \nonumber \\
\label{complex}
& & \left. +(\rho_{T-}+\rho_{T+})\cos{\left[\sqrt{|D|}\ln(a)\right]}\right\} +\rho_\Lambda\end{aligned}$$ containing an oscillating function that modulates the power law scaling. We will now investigate the dynamics of the density parameters. In Sec. \[Markov\] we will present a MCMC analysis with SNe data.
Analysis of specific couplings {#critpoints}
==============================
Dynamics of density parameters {#dparam}
------------------------------
Introducing an interaction between two fluids can lead to interesting solutions for the energy densities, like attractor points in the phase space where the contributions of the two fluids to the total energy density are constants. In these points the value of the normalised energy densities depends only on the parameters of the model and, since they are attractors, they are reached from a wide range of initial conditions, thereby alleviating the coincidence problem. These are usually called “scaling solutions” [@Wands:1993; @Copeland:1998p1287] and are characterized by constant fractions of the energy density parameters, namely $\Omega_{A,B}=\rho_{A,B}/(3H^{2})$ (in units $8\pi G=1$, $c=1$).
In order to analyse the dynamics of the system, let us define the new variables: $$\begin{aligned}
\label{xy}
x=\frac{\rho_{A}}{3H^{2}};\qquad y=\frac{\rho_{B}}{3H^{2}}; \qquad z=\frac{\rho_{\Lambda}}{3H^{2}},\end{aligned}$$ where together with the coupled fluids we also include radiation to include the era when it’s the dominating component, when initial conditions are usually set. Note that $x=\Omega_{A}$, $y=\Omega_{B}$ and $\Omega_\gamma$ are constrained by $x+y+\Omega_{\gamma}=1$; $z$ is the energy density parameter of the total effective cosmological constant, and we neglect the baryons contribution, which is always subdominant. The system (\[cons\]-\[cons2\]) then becomes $$\begin{aligned}
\label{xycons}
x'&=&-x\Big[3(1+w_{+}-w_{-})+2\frac{H'}{H}\Big]\\
\nonumber&+& \frac{3}{2}\Big[(q_{+}-q_{-})x+(q_{+}+q_{-})y+\frac{\det({\bf J})}{9w_{-}}z\Big]\\
\label{xycons2}
y'&=&-y\Big[3(1+w_{+}+w_{-})+2\frac{H'}{H}\Big]\\
\nonumber&-& 3\Big[(q_{+}-q_{-})x+(q_{+}+q_{-})y+\frac{\det({\bf J})}{9w_{-}}z\Big]\\
\label{xycons3}
z'&=&-2z\frac{H'}{H},\end{aligned}$$ where $$\begin{aligned}
\label{eq:Raychaudhuri}\nonumber
\frac{H'}{H} & = & -1-\frac{1}{2} \left[x(1+3(w_{+}-w_{-})) \right. \\
& & \left. +y(1+3(w_{+}+w_{-}))+2(1-x-y) \right] \end{aligned}$$ is a rewriting of the Raychaudhuri equation (\[ray\]) for the Hubble expansion scalar.
The fixed points, namely the points satisfying $x'=y'=z'=0$, are presented in Table \[table1\], labeled by capital letters, together with the corresponding eigenvalues. To the best of our knowlegde, this is the first complete analysis of the dynamics of a three components cosmological system where two of the barotropic fluids are coupled via a general linear coupling function of the form (\[coupling\]). The effective EoS parameters at each of the fixed points $w_{eff}=p_{tot}/\rho_{tot}$ is also listed, where $\rho_{tot}=\rho_{A}+\rho_{B}+\rho_{\gamma}$ and therefore $w_{eff}=(w_{+}-w_{-})x+(w_{+}+w_{-})y+\Omega_{\gamma}/3$.
Points $x$ $y$ $z$ $w_{eff}$ $\lambda_{x}$ $\lambda_{y}$ $\lambda_{z}$
-------- -------------------------------------------- ------------------------------------------------------------------------------------ ----- --------------- ------------------ ----------------------------------------------------------- -----------------------------------------------------------
A $0$ $0$ $0$ $\frac{1}{3}$ $4$ $1-3\beta_{+}$ $1-3\beta_{-}$
B $-\frac{q_{-}-2w_{-}+2\sqrt{D}/3}{4w_{-}}$ $\frac{(q_{-}-2w_{-}+2\sqrt{D}/3)(q_{+}+2w_{-}+2\sqrt{D}/3)}{4w_{-}(q_{+}+q_{-})}$ $0$ $\beta_{+}$ $3(1+\beta_{+})$ $-\frac{1-3(\beta_{+}-2\beta_{-}+q_{-})+\sqrt{F_{+}}}{2}$ $-\frac{1-3(\beta_{+}-2\beta_{-}+q_{-})-\sqrt{F_{+}}}{2}$
C $-\frac{q_{-}-2w_{-}-2\sqrt{D}/3}{4w_{-}}$ $\frac{(q_{-}-2w_{-}-2\sqrt{D}/3)(q_{+}+2w_{-}-2\sqrt{D}/3)}{4w_{-}(q_{+}+q_{-})}$ $0$ $\beta_{-}$ $3(1+\beta_{-})$ $-\frac{1-3(\beta_{-}-2\beta_{+}+q_{-})+\sqrt{F_{-}}}{2}$ $-\frac{1-3(\beta_{-}-2\beta_{+}+q_{-})-\sqrt{F_{-}}}{2}$
D $\frac{1+w_{-}+w_{+}}{2w_{-}}$ $-\frac{1-w_{-}+w_{+}}{2w_{-}}$ $1$ $-1$ $-4$ $-3(1+\beta_{+})$ $-3(1+\beta_{-})$
All the fixed points shown in Table \[table1\] exist for $w_{-}\neq0$, when the EoS parameters of the two fluids are the different. As aforementioned, the only physically reasonable fixed point for system (\[constot\]-\[cons2tot\]) corresponding to $w_{-}=0$ is [**Case 2a**]{}, where $\det({\bf J})=0$.
The fixed points A corresponds to the radiation dominated era, while B, C and D represent epochs that are dominated by the two fluids. In particular, at the fixed point D the constant energy densities of $x$ and $y$ ($\rho_{A\Lambda}$ and $\rho_{B\Lambda}$) cause the accelerated expansion with $w_{eff}=-1$. From the expression for $x$ and $y$ at this latter fixed point it is easy to see that $y=-x(1+w_{A})/(1+w_{B})$, which is exactly the proportionality that holds for [**Case 1**]{} and [**Case 2c**]{}, as discussed at the end of Sec. \[stab\]: this point is characterized by the final domination of an effective cosmological constant, either driven by $q_{0}$ ([**Case 1**]{}, see Fig. \[fig:omegasbis\]) or not ([**Case 2c**]{}, see Fig. \[fig:omegastris\]). In the first case whenever $|\beta_{\pm}|<1$ D is always an attractor, while in the second case it is not because one of the eigenvalues is null. Notice that its existence is completely independent on $q_{+}$ and $q_{-}$. Whenever the system settles into the fixed points B or C the role of $\beta_{+}$ and $\beta_{-}$ is exactly that of effective EoS parameters (see Table \[table1\]) which allows for phantom line crossing at late time (see Fig. \[fig:omegas\]), i.e. line for which the effective total EoS parameter is $w_{eff}=-1$.
In the following we will examine in more detail three special classes of the coupling function and in Sec. \[Markov\] we will make a first comparison of the models to the data using MCMC applied to type Ia SNe distance modulus.
I. $q_{+}=q_{-}$ {#single .unnumbered}
----------------
Imposing $q_{+}=q_{-}$ is equivalent to choosing $q_{A}=0$ and $q_{B}=q$; therefore among the range of possible couplings represented by (\[coupling\]) we are restricting to the class of models where $Q/H$ is proportional solely to the energy density of one fluid (in our case e.g. $\rho_{B}$), and it reads $$\label{coupl1}
\frac{Q}{H}=\frac{3}{2}(q\rho_{B}+q_{0}).$$ This assumption also includes models with $q_{+}=-q_{-}$ since the coefficients $q_{A}$ and $q_{B}$ can be either negative of positive. In this case the dynamics is the same as for $q_{+}=q_{-}$, the roles of $x$ and $y$ being simply interchanged. We will refer to this subclass of models as model I.
In this model $\sqrt{D}$ is automatically real, since $D=9(q/2+w_{B}-w_{A})^{2}/4$; as a consequence the scaling function (\[rhotot\]) always drives a power law expansion, with $\beta_{+}=q/2+w_{B}$ and $\beta_{-}=w_{A}$ if $\beta_{0}>0$ (i.e. $(w_{B}+w_{A}+q/2)>0$), vice versa if $\beta_{0}<0$. Hence the total fluid ends up as if it was made up of: [*i)*]{} a component scaling as the original fluid $\rho_{A}$ with no coupling, [*ii)*]{} a second component characterized by a new EoS parameter and [*iii)*]{} an effective cosmological constant term $\rho_{\Lambda}$. Moreover a pure affine behaviour (\[affine\]), or its improper modification (\[improper\]), is obtained in three cases: [*i*]{}) $q=-2(w_{B}-w_{A})$, which gives (\[improper\]); [*ii*]{}) $q=-2(1+w_{B})$, that corresponds to $\beta_{+/-}=-1$ (even for $\rho_{\Lambda}=0$, i.e. $q_{0}=0$, an effective cosmological constant is generated); [*iii*]{}) $w_{A}=-1$, where one of the two fluids is [*ab initio*]{} a constant term. Notice however that generally, because $\beta_{+}=\beta_{-}+2\sqrt{D}/3$, models with $\beta_{+}=-1$ and $\beta_{-}>-1$ are not feasible. In particular, the $\Lambda$CDM evolution is exactly recovered in case [*ii*]{}) for $w_{A}=0$, that is if one of the fluids [*is* ]{} dust; in case [*iii*]{}) for $q_{B}=-2w_{B}$. The fixed point D is characterized by the domination of the constant part of the total energy density $\rho_{\Lambda}$; along it, the values of $x$ and $y$ are both positive only if either $w_{A}$ or $w_{B}$ have phantom values, i.e. $w_{A}<-1$ or $w_{B}<-1$. This statement holds true also for models II and III. However, if $w_{A}<-1$ D is no longer an attractor, as $\lambda_{y}=-3(1+w_{A})$ is greater than zero. On the other hand $w_{B}<-1$ requires $q>-2(1+w_{B})$ to let the fixed point be an attractor: in this case $q$ is positive. A strong and positive $q$ corresponds to a transfer of energy from $\rho_{A}$ to the other fluid with $w_B<-1$. Therefore in order to fall at late time into the cosmological constant dominated era a fluid with a phantom EoS parameter $w_B$ must absorb energy from the other non-phantom fluid. It is worth stressing that the effective cosmological constant, i.e. $q_{0}$, is somewhat redundant whenever the fixed point D is not an attractor (see Fig. \[fig:omegas\]). In Fig. (\[fig:omegas\]) an example of this dynamics of the background is shown; the effective cosmological constant is not noticeable, since, after the evolution on the saddle point B, the system is trapped in the attractor point C.
![ Upper panel: evolutions of the energy density parameters $\Omega_{A}$ (thin solid line), $\Omega_{B}$ (dotted line) and $\Omega_{\gamma}$ (thick solid line) for a model with $q_{+}=q_{-}=0.25$; for comparison, the dashed lines are the values of $x$ and $y$ at the fixed points B (thin short-dashed lines) and C (thick long-dashed lines). For this model the parameters are: $\Omega_{0A}=\Omega_{\Lambda}=0.5$, $w_{A}=0$, $w_{B}=-1.5$, $\beta_{+}=0$ and $\beta_{-}=-1.25$. Lower panel: the total effective EoS parameter for the same model : $w_{eff}$ evolves from the value $1/3$ in the radiation dominated era, approaches the value $~0$ in the matter dominated era and then asymptotically evolves toward a constant phantom value, in this case $\beta_{-}=-1.25$. []{data-label="fig:omegas"}](tot_omw1.eps){width="7.truecm"}
![ Upper panel: evolutions for the energy density parameters for a model with $q_{-}=0$ and $q_{+}=-0.5$; for comparison, the dashed lines are the values of $x$ and $y$ at the fixed points B (thin short-dashed lines) and D (thick long-dashed lines). For this model the parameters are: $\Omega_{0A}=\Omega_{\Lambda}=0.5$, $w_{A}=-1.1$, $w_{B}=0.2$. Lower panel: effective EoS for the same model; for comparison, we plot the EoS parameter of the fixed point B, $\beta_{+}=-0.14$.[]{data-label="fig:omegasbis"}](tot_omw2.eps){width="7.truecm"}
![ Upper panel: evolutions for the energy density parameters for a model with $q_{+}=0$ and $q_{-}=-0.18$; for comparison, the dashed lines are the values of $x$ and $y$ at the fixed point B (thin short-dashed lines) and D (thick long-dashed lines). For this model the parameters are: $\Omega_{0A}=0.5$, $\Omega_{\Lambda}=0$, $w_{A}=-0.9$, $w_{B}=0.$ (dust). The EoS parameters at B are $\beta_{+}=-0.08$ and $\beta_{-}=-1$. Lower panel: effective EoS for the same model. []{data-label="fig:omegastris"}](tot_omw3.eps){width="7.truecm"}
II. $q_{-}=0$ {#total .unnumbered}
-------------
If $q_{-}=0$ the resulting coupling function $Q/H$ is linearly dependent on the sum of the energy densities of the two fluids, approximately equivalent to the total energy density (these models have been examined for example in [@Olivares:2006] and [@Abdalla:2007]) and is as follows $$\label{coupl2}
\frac{Q}{H}=\frac{3}{2}(q\rho_{T}+q_{0}).$$ With this assumption $q_{A}=q_{B}=q_{+}=q$ and $\beta_{\pm}=w_{+}\pm\sqrt{D}/3$ where $D=9(w_{B}-w_{A})(2q+w_{B}-w_{A})/4$. If $q$ is positive these effective EoS are real for $w_{B}> w_{A}$ or $w_{B}\le w_{A} -2q$, while if $q$ is negative the same relations hold but with opposite inequality signs. We will label this model II.
In this model the affine evolution is recovered for $w_{B}=(qw_{A}+2w_{A}+2)/(q-2w_{A}-2)$, corresponding to $\beta_{-}=-1$. In this case, which is indeed [**Case 2c**]{} of Section \[coupledfluids\], an effective cosmological constant arises even for $\rho_{\Lambda}=0$. Again, because $\beta_{+}=\beta_{-}+2\sqrt{D}$, models with $\beta_{+}=-1$ and $\beta_{-}>-1$ are not feasible. From a cosmological point of view this means that a matter-like evolution cannot be generated together with a cosmological constant. The $\Lambda$CDM limit is achieved if$w_{A}=(-1+q\pm\sqrt{q^{2}+1})/2$ and $w_{B}=-1-w_{A}$. The evolution of the energy densities for a special choice of the parameters is illustrated in Fig. \[fig:omegasbis\]: the effective cosmological constant (\[effL\]) arises at late time, driving the acceleration, and $\rho_{\Lambda}$ is caused by a non-zero $q_{0}$ ($\Omega_{\Lambda}\neq 0$).
III. $q_{+}=0$ {#difference .unnumbered}
--------------
The subgroup of models with $q_{+}=0$ ( from now on model III) includes the couplings that are proportional to the difference of the energy densities $\Delta$ (for example recently analysed in [@Chimento:2007bis]). With this assumption $q_{-}=q_{B}=-q_{A}=q$ and the discriminant $D=9(q^{2}+(w_{B}-w_{A})^{2})/4$ is always positive, so that oscillating solutions (\[complex\]) are never permitted. The coupling function reads $$\label{coupl3}
\frac{Q}{H}=\frac{3}{2}(q\Delta+q_{0}).$$ As before, the affine expansion (\[affine\]) may only be generated if one of the two effective EoS parameters assumes the value of the EoS of a cosmological constant, that is either $\beta_{+}=-1$ or $\beta_{-}=-1$. In particular if $\beta_{+}=-1$, $\beta_{-}=-1-2\sqrt{D}/3$ is always phantom. In this case none of the terms in Eq. (\[rhotot\]) can play the role of matter. On the other hand if $\beta_{-}=-1$ (corresponding to $w_{A}=(-1-q\pm\sqrt{1-q^{2}})/2$), $\beta_{+}=-1+2\sqrt{D}/3$ is always grater than $-1$, i.e. always standard. An example of this dynamics is shown in Fig. \[fig:omegastris\], where the effective cosmological constant (\[effL\]) arises at late time with no need of $q_{0}$, driving the acceleration ([**Case 2c**]{}). Then typically for $w_{B}=-q-w_{A}-1$ we have that $\beta_{+}=0$ and the $\Lambda$CDM model is recovered.
Markov chains with supernovae {#Markov}
=============================
Methods
-------
Given the large number of parameters, the task of finding the minimum $\chi^{2}$ and mapping its distribution in the entire parameter space can be computationally expensive. To this end we adopt a MCMC. In this work we only want to test our models as a description of the homogenous isotropic background expansion (regardless of perturbations), hence supernovæ are ideal for this purpose. We use the 192 type Ia SNe distance modulus data set provided in [@Davis:2007]. In particular we want to see wether supernovae can qualitatively distinguish different kind of couplings, included what we called model I, II and III.
Type Ia SNe light curves allow a determination of an extinction-corrected distance moduli, $$\mu_0=m-M=5\log \left(d_L/{\rm Mpc}\right)+25
\label{distance_modulus}$$ where $d_L=(L /4\pi F)^{1/2}=(1+z)\int_0^z dz' / H(z')$ is the luminosity distance. We compare our theoretical predictions to the values of $\mu_0$ with $H^{2}=8\pi G/3(\rho_{A}+\rho_{B}+\rho_{\gamma}+\rho_{b})$, where we account also for the baryon energy density $\rho_{b}$. We fix the value of the dimensionless Hubble constant to be $h=0.72$ [@HSTKeyP01] and the baryon energy density at present $\Omega_{b}h^{2}=0.02229$ according to [@Spergel:astro-ph/0603449]. The smaller is the EoS parameter of a single fluid the later can be the domination era for this fluid. Hence, counting the role of $\beta_{\pm}$ as effective EoS parameters, whenever $\beta_{\pm}>0$ a baryonic era might emerge at recent time. The absolute distance modulus $M$ is intrinsically affected by uncertainty; therefore we treat it as a nuisance parameter and marginalize over it.
The parameters that are representative of the models are $\{\Omega_{0A},\Omega_{\Lambda},q_{A},q_{B},w_{A},w_{B}\}$, or otherwise $\{\Omega_{0A},\Omega_{\Lambda},q_{+},q_{-},w_{+},w_{-}\}$ and, as functions of these, the two effective EoS introduced in Eq. (\[rhotot\]): $\beta_{+}$ and $\beta_{-}$. For the ensuing analysis it is worth reminding our classification of models: I) model with a coupling function proportional to only one of the two energy densities; II) model with a coupling function proportional to the sum of the energy densities; III) model with a coupling function proportional to the difference of the energy densities.
We shall now focus our analysis on the case $w_{B}=0$, i.e. $\rho_B$ would represent standard CDM if it wasn’t for the coupling with the DE component.
Results: CDM - DE coupled models {#results}
--------------------------------
The first result we obtain is that $\Omega_{\Lambda}$ is completely unconstrained, independently on which model we consider. This means that SNe are not sensitive to the constant term of the coupling. As we have seen in Sec. \[coupledfluids\] the dynamics of the system can easily generate the acceleration settling on fixed points D, where $w_{eff}=-1$, even for $\Omega_{\Lambda}=0$ (see Fig. \[fig:omegastris\]), or B and C, where the total energy density can also exhibit phantom evolutions.
In Fig. \[fig:res1\] and \[fig:res2\] we present MCMC chains in a two-dimensional diagram $[q_{+},q_{-}]$ ($[q_{A},q_{B}]$ on the right hand side). As said above, we consider a model where one of the two fluid represents a CDM component, i.e. $w_{B}=0$, a reasonable assumption considered all the other cosmological probes pointing towards the existence of a form of cold dark matter (see e.g. [@Khalil:2002]), and we let $w_{A}$ assume three different values that characterize $\rho_{A}$ as a DE component (phantom-like behaviour is shown in the top panels, cosmological constant-like in the second row panels and non-phantom model in the bottom panels).
Note that the case where $w_{A}=0$ and $\rho_{B}$ is DE can be easily derived from the previous one, corresponding in the diagram to a reflection with respect to the line $q_{B}=-q_{A}$. In fact, interchanging the two EoS and swapping the roles of the two energy densities, and applying the transformation ($q_{A}\rightarrow -q_{B}$, $q_{B}\rightarrow -q_{A}$, i.e. $q_{+}\rightarrow -q_{+}$, $q_{-}\rightarrow q_{-}$), one recovers the aforementioned model.
In addition, the straight lines corresponding to models I, II and III are drawn, and diagrams of Fig. \[fig:res2\] are derived from the same choice of parameters as in Fig. \[fig:res1\] except for $\Omega_{\Lambda}\neq 0$ (it is by eye easily verifiable that there is no dependence on $\Omega_{\Lambda}$). Finally, the short-dashed curves represent the improper affine evolution (\[improper\]), while the short-dashed straight line represents affine models (\[affine\]) with $\beta_{+/-}=-1$ ([**Case 2c**]{}).
As a first step we derived the unidimensional likelihood for $\{\Omega_{0A},q_{+},q_{-}\}$. The best fit of the energy density parameter for the three class of models presented in Fig. \[fig:res1\] and \[fig:res2\] is respectively $\Omega_{0A}=0.63,0.65,0.76$ with an error of $2\sigma=0.1$; this best fit does not change including $\Omega_{\Lambda}$. In the diagrams $\Omega_{0A}$ is therefore fixed to these best fit values. It is worth stressing that here we are not just analysing the typical models considered in literature (namely I, II and III) but the results incorporate [*all*]{} the possible linear couplings, and, we might say, all the possible expansions at recent times of a generic coupling function $Q$ (Eq. \[Coup\]). Hence we are not interested in deriving constraints on single parameters, a route that might be hard to follow with SN Ia in view the high number of parameters and their degeneracies. We instead want to see what kind of linear couplings are preferred by the data and provide a qualitative way to distinguish the type and the direction of the interaction.
The first noticeable thing in the $[q_{A},q_{B}]$ diagram is that the points lie almost on a horizontal branch of the diagram, close to the line representing model I, in particular with $Q\propto \rho_{A}$. So if we allow the interaction term to be strong and move out of the weak coupling regime (i.e. [$|q_{A,B}|>1$]{}), the most “frequent” linear coupling function emerging from the chains is the one proportional to the DE density ($\rho_{A}$). In addition, strong couplings are favoured for positive value of $q_{A}$ (see Fig. \[fig:res1\] and \[fig:res2\]): the energy is transferred from dark matter to DE. Increasing the value of $w_{A}$, that is moving from phantom-like values towards quintessence-like ones (going downwards in the right hand side column of fig. (\[fig:res1\])), this horizontal branch tends to negative values of $q_{B}$. Models with a phantom $w_A$ show an increasing energy density with the scale factor, $a$, while for DE model characterized by $w_A>-1$ the energy density is diluted with the universe expansion: this second kind of models requires a lower transfer of energy from CDM to DE. Apart from a small spot in the origin of the axis (weak couplings), the coupling or type II does not seem to be favored by SN data, the effect increasing with higher values of $w_{A}$, i.e. for non-phantom values. Another evidence that arises from diagrams Fig. \[fig:res1\] and \[fig:res2\] is that for non-phantom values of $w_A$ the uncoupled case (namely $[q_{A},q_{B}]=[0,0]$) falls almost outside the border of the likelihood.
Since today $\rho_{0B}\simeq \rho_{0A}$ we can say that the sign of the coupling function ($Q \simeq q_{A}\rho_{0A}+q_{B}\rho_{0B}$) changes along the straight line $q_{B}=-q_{A}$ (long dashed line): above this line the exchange term reverses the energy transfer from CDM to DE (i.e. positive $Q$), while below it is the opposite (negative $Q$). Again, the higher is $w_{A}$ the bigger is the number of points that we can find below this line. Therefore for DE components with $w_A<-1$ an exchange of energy from DE to CDM is less probable, independently on the type of linear coupling. This reflects the fact that an increasing energy density (characteristic of phantom behaviour) favors more and more absorbing and positive DE couplings at present, while non-phantom values of $w_A$ seem to need a negative exchange term, most of all for weak couplings, to explain supernovae data. It is worth stressing that eventually the likelihood seems to exclude the uncoupled case.
{width="17.truecm"}
The connection between where the points lie in the diagrams, i.e. the region favoured by the likelihood, and where the cosmological background evolution is affine is an interesting issue; this directly connects coupled DE models to an effective evolution of the total energy density that is completely equivalent to a cosmological constant plus a component with constant EoS parameter $\alpha$, Eq. (\[affine\]). If one looks at the left side diagrams of Fig. \[fig:res2\], a short-dashed curve and a short-dashed straight line are drawn on it. The former corresponds to the improper affine evolution (\[improper\]), obtained for $q_{-}=\pm\sqrt{-4w_{-}(q_{+}+w_{-})}$. Hence the only affine models are those that correspond to the straight line for witch $\beta_{+/-}=-1$ (for the model with $w_{A}=-1$ this coincides with the line representing model II, the only possibility to recover the affine evolution with no coupling). For a DE model with a phantom $w_A$ the affine evolution coexisting with a non-zero $\Omega_{\Lambda}$ is somewhat ruled out and, among the models indicated in the last Section, is more compatible with a coupling function proportional to $\rho_{A}$ (DE, model I) and possibly to $\Delta$ (model III). For DE models with $w_A>-1$ the situation is different: the data seem to favor an affine evolution generated in models with a coupling function proportional to $\rho_{B}$ (matter, model I) and again model III. In addition, for DE models with standard $w_A$ an improper affine evolution together with a non-vanishing $q_{0}$ ($\Omega_{\Lambda}$) is allowed, in a region where the coupling function shifts towards negative sign, thus representing a transfer of energy from DE to CDM.
{width="17.truecm"}
Conclusions {#conclusion}
===========
We have analysed the dynamics of two coupled dark components represented by two barotropic perfect fluids characterized by constant EoS parameters $w_A$ and $w_B$. We have assumed a flat, homogeneous and isotropic cosmology and a general linear coupling between the two barotropic perfect fluids. This scale-independent coupling takes a linear form proportional to the single energy densities plus a constant term: any coupling of this type can approximate at late time a more general coupling function. We have studied the stability of the system and shown that an effective cosmological constant can arise both from the constant part $q_{0}$ of the function $Q$ and from an effective cosmological constant-like EoS. We have also examined the dynamics of the energy density parameters, and evaluated the fixed points and the corresponding eigenvalues, for the most general form of linear coupling. We have then restricted the analysis to some specific linear couplings previously considered in the literature (model I, II, III). Since we are restricting to the background expansion and we have modeled the coupling function as a late time first order Taylor expansion, a comparison with distance modulus from SN Ia data appeared as our natural step further. We have presented a MCMC analysis for a model with dark matter plus DE using the data set provided in [@Davis:2007]. Considering two representative specific values of the DE parameter $w_A$, one standard ($w_A>-1$) and the other phantom ($w_A<-1$), we have condensed our results in coupling diagrams, where the points arising from the MCMC chains are drawn together with lines for model I, II and III and for the improper affine (\[improper\]) affine (\[affine\]) evolutions, the latter including the $\Lambda$CDM model as a subcase. Couplings proportional to the DE density seem favored, mostly for strong couplings $|q_A|>1$. The total sign of the exchange term sets the direction of the interaction: models with phantom $w_A$ definitely prefer positive coupling, i.e. an energy transfer from dark matter to DE. On the other hand, models with non-phantom $w_A$ not only allow for negative $Q$, but forces the uncoupled model to fall at the border of the likelihood. For further and stronger constraints more complementary data are required, like CMB spectra or matter power spectra. These observables necessitate an accurate relativistic perturbation analysis which is nor obvious neither uniquely defined in phenomenological coupled models as those considered here. Moreover, simplified observable that make no use of perturbation analysis, like the CMB shift parameter, can be strongly model-dependent and, although straightforward, should not be used in models where the evolution, even just that of the unperturbed background, detaches significantly from that of the $\Lambda$CDM model. These extended investigations can only be settled with future work.
Acknowledgements
----------------
We are gratuful to Roy Maartens, Betta Majerotto, Jussi Valiviita and other members of ICG (Portsmouth) for useful discussions. MB work was partly funded by STFC.
[^1]: See e.g. [@Durrer:2008] and other articles in the same special issue on dark energy.
[^2]: Strictly speaking, an expansion about today would lead to $q=\hat{q}_{0}+\hat{q}_{A}(\rho_{A}- \rho_{A0})+\hat{q}_{B}(\rho_{B} -\rho_{B0})$, but constants can always be re-defined in order to put the coupling $q$ in the form (\[coupling\]).
[^3]: Only four of them are independent: one can always rewrite Eqs. (\[constot\]-\[cons2tot\]) renormalising the parameters to any of them, but we don’t want to assume that any of $w_+, w_-, q_+, q_-, q_0$ is non null. In particular, the value of $q_0$ is irrelevant to the existence of the fixed points according to the eigenvalues of $\bf J$, but physically its value is relevant, because it determines the effective cosmological constant (\[effL\]).
[^4]: Mathematically, this is easily understood in terms of the dynamical system formalism of the previous Section: because the system (\[Xprime\]) is linear, it can be globally (in phase space) written in Jordan normal form [@KArrowsmith:1992]. When the eigenvalues (\[eigenvals\]) $\lambda_\pm$ are real and distinct this Jordan form is diagonal, i.e. the dynamical system separates into two normal modes, physically corresponding to two new effective uncoupled fluids with constant EoS parameters $\beta_{\pm}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The contagion threshold for diffusion of innovations is defined and calculated in finite graphs (two-dimensional regular lattices, regular random networks (RRNs), and two kinds of scale-free networks (SFNs)) with and without the bilingual option. Without the bilingual option, degree inhomogeneity and clustering enhance the contagion threshold in non-regular networks except for those with an unrealistically small average degree. It is explained by the friendship paradox and detour effect. We found the general boundary of the cost that makes the bilingual option effective. With a low-cost bilingual option, among regular lattices, SFNs, and RRNs, the contagion threshold is largest in regular lattices and smallest in RRNs. The contagion threshold of regular random networks is almost the same as that of the regular trees, which is the minimum among regular networks. We show that the contagion threshold increases by clustering with a low-cost bilingual option.'
address:
- 'Division of Liberal Arts and Sciences, Gwangju Institute of Science and Technology, Gwangju 61005, South Korea'
- 'Department of Physics and Photon Science, Gwangju Institute of Science and Technology, Gwangju 61005, South Korea'
author:
- 'Jeong-Ok Choi'
- Unjong Yu
bibliography:
- 'diff\_inno.bib'
title: 'Diffusion of innovations in finite networks: effects of heterogeneity, clustering, and bilingual option on the threshold in the contagion game model'
---
Contagion game ,Diffusion of innovations ,Complex network , Scale-free network ,Regular random network ,Clustering coefficient
Introduction
============
*Diffusion of innovations* describes the diffusion process of a new idea, technology, or product through a social network [@Rogers03; @Kleinberg07]. The rapid developments of communication, transportation technologies, and social networking services are making connections between people denser and more complex [@Lazer09], and so the mechanism of diffusion of innovations is becoming more and more complex.
In the diffusion of innovations, people tend to adopt the innovation when the exposure, which is the proportion of innovation adopters in their neighborhood, is larger than a certain threshold [@Valent96; @Centola10; @Pei13]. It is because people usually judge the benefit of adopting new innovation based on the exposure. As a special case of diffusion of innovations, the epidemic begins from one or a few instigators (or infected people) and diffuses through a network making all the agents infected when the process is finished. In this paper, our main interest is the epidemics.
This problem was modeled mathematically as a contagion game by Morris [@Morris]. In this contagion game model, every agent takes one option from two strategies: $A$ and $B$. During the game an agent can get (positive) payoff only when it has the same option as its neighbor; they get the payoff $(1-q)$ and $q$ when they have the strategy $A$ and $B$, respectively. To begin with, Morris considered an infinite graph where all the agents have strategy $B$ (status quo option) except for only finite number of instigators that adopt strategy $A$ (innovative option) and keep their strategies $A$. The innovative option $A$ is assumed to be better than the status quo option ($q<0.5$) because there is no incentive to transform to the innovative option otherwise. All the agents but instigators change their strategies to maximize their payoffs in their local environments (best-response update) infinite times. It is called that the strategy $A$ become *epidemic* on the graph in the game when the whole graph adopts the innovative option in the end by this dynamics. The contagion threshold is defined to be the largest $q$ value that can make the game on the graph epidemic from all the possible finite-size instigators [@Morris].
Another important aspect of a contagion game is the possibility of the bilingual option, which is compatible with both options by paying an extra cost. Immorlica [*et al.*]{} extended the contagion game of Morris by including the bilingual option, which is represented by the payoff matrix of Table \[Table\_payoff\] [@Immorlica07]. This bilingual option makes the dynamics in a contagion game more complex. Since it costs $c$ for an agent to choose the bilingual option (i.e., the cost is $c/\Delta \times \Delta = c$ if it has $\Delta$ neighbors), choosing the bilingual option is likely to give the agent larger total payoffs than choosing $A$ or $B$ when $c$ is low enough. A bilingual option with low cost may help a smooth transition to the innovative option $A$ for an agent with more neighbors. As a result, the contagion threshold increases. [@Immorlica07] obtained contagion threshold $q_0^c$ as a function of cost $c$ in some infinite regular graphs. Recently, it was proved that the infinite regular tree has the smallest contagion threshold among infinite regular graphs of the same degree for every positive $c$ [@Yu_Choi17].
In this paper, we extend the study of [@Morris; @Immorlica07; @Yu_Choi17] to finite and irregular graphs. We define the contagion threshold for finite graphs and calculate it for regular lattices, regular random networks (RRNs), and two kinds of scale-free networks (SFNs) with and without the bilingual option. To see the effect of each network parameter more clearly, we restrict this work to model networks. Effects of degree heterogeneity, clustering, and bilingual option on the contagion threshold are discussed.
----------------------- ------------- ---------------- -------------- ----------------
option $A$ option $B$ option $AB$
option $A$ $1-q$ $0$ $1-q$
Strategy of the agent option $B$ $0$ $q$ $q$
option $AB$ $1-q-c/\Delta$ $q-c/\Delta$ $1-q-c/\Delta$
----------------------- ------------- ---------------- -------------- ----------------
Model and methods
=================
In a contagion game on an infinite graph with the payoff matrix in Table \[Table\_payoff\], the [*contagion threshold*]{} is defined the maximum value of $q$ to be epidemic. Initially in the game, some agents are instigators and take $A$ while all the rest of agents take $B$. Note that with a bilingual option, the contagion threshold is a function of $c$. In the definition of the contagion threshold for infinite graphs, there is no restriction in the set of instigators as long as its size is finite [@Morris; @Immorlica07; @Yu_Choi17]. If the same definition is used for finite graphs, then the contagion threshold will be always 1 with a trivial set of instigators–the entire vertex set. Another problem is that it is practically impossible to find the optimal set of instigators that maximize diffusion of innovations in complex networks [@Kempe03]. Therefore, we define the contagion threshold for finite irregular graphs as follows: in a [*finite*]{} graph, the contagion threshold $q_0$ is the expectation value of maximum $q$ for the innovative option $A$ to spread the whole graph from a state in which all nodes adopt the option $B$ except [*one node*]{} (an instigator) that adopts and keeps the option $A$, through the contagion game that all nodes except the instigator perform best-response update as many times as possible. This definition is justified when the density of instigators is low and each instigator acts independently of the other instigators. We denote $q_0^c$ the contagion threshold for finite graphs with the bilingual option of cost $c$. This definition gives the same value as that by the original definition for the infinite lattice in the square lattice, but they may give different results in other networks.
![Regular lattice network, regular random network, and Barab[á]{}si-Albert scale-free network. Red circles and lines represent nodes and edges, respectively. In (a), black solid lines represent edges of lattice of $\Delta \geq 6$; blue dashed lines, $\Delta \geq 8$; and green curves, $\Delta = 10$. $\Delta$ is the degree of each agent.[]{data-label="Fig_lattice"}](figure1.eps){width="10.0cm"}
Three kinds of two-dimensional regular lattices with a degree ($\Delta$) of 6, 8, and 10 are studied in this work. (See Fig. \[Fig\_lattice\](a).) The lattices of $\Delta=6$ and $\Delta=8$ are the same as the triangular lattice and the square lattice with Moore neighborhood, respectively. Two third-nearest neighbors are added to the lattice of $\Delta=8$ to make a lattice with $\Delta=10$. The periodic boundary condition is used. The [*clustering coefficient*]{} ($\xi$) of a graph is the ratio of the number of triangles over the number of connected triples of nodes. The clustering coefficients are $\xi=2/5$, $3/7$, and $7/15$ for $\Delta=6$, $8$, and $10$, respectively.
RRNs with $N$ nodes of degree $\Delta$ were generated by the Steger-Wormald algorithm [@Steger99]. It is faster than the original pairing model [@Bollobas80] and generates uniform RRNs asymptotically when $N$ is large [@Kim03]. The algorithm begins by making a set of unpaired arms of nodes $\{ (1,1),$ $(1,2), \cdots, (1,\Delta),$ $(2,1), \cdots, (N,\Delta-1),$ $(N,\Delta)\}$, whose element $(a,b)$ represents the $b$-th unpaired arm of node $a$. Two elements $(i_1,i_2)$ and $(j_1,j_2)$ are chosen at random from the set. If $i_1 \neq j_1$ and the two nodes have not been connected before, then nodes $i_1$ and $j_1$ are connected. After connecting $i_1$ and $j_1$, the two elements $(i_1,i_2)$ and $(j_1,j_2)$ are eliminated from the set of unpaired arms. This pairing procedure is repeated until no unpaired arm is left. The probability that this algorithm fails to make a $\Delta$-regular network without multiple edges is very small. The clustering coefficient of RRNs is $(\Delta-1)/N$ asymptotically, which approaches zero for large $N$ (see the appendix).
SFNs were constructed using the growing method [@Barabasi99]. Starting from a network of $m$ isolated nodes, a new node is introduced and $m$ edges connect the new node and existing nodes avoiding multiple edges. This growth process (i.e., an addition of one node and $m$ edges to the network) is repeated until the network has $N$ nodes. The node to connect to a new node is chosen by two algorithms: preferential attachment (PA) and triad formation (TF) [@Holme02]. As for the PA, the probability to be selected is proportional to its degree, while the node is chosen randomly among neighbors of the node chosen immediately before by the PA, in the TF. For the first edge of a new node, the PA is used, and from the second connection, TF is chosen with probability $P_{\mathrm{TF}}$ if it is possible. The Barab[á]{}si-Albert SFN, whose clustering coefficient vanishes for large network size [@bollobas03], is obtained with $P_{\mathrm{TF}}=0$. The clustering coefficient can be enhanced by adjusting $P_{\mathrm{TF}}>0$; it is called the Holme-Kim SFN. The two SFNs have almost the same degree distribution, which is $P(\Delta) \sim \Delta^{-3}$ [@Holme02]. The average degree of each SFN is $\langle \Delta \rangle = 2m$.
As for regular lattices, where network structure is unique and all nodes are symmetric, the contagion threshold $q_0^c$ can be calculated analytically. For other networks, it was obtained numerically. The location of the instigator can be crucial, especially, in SFNs. We confirmed that high-degree instigators tend to give higher $q_0^c$, but there are exceptions; it is known to be NP-hard to determine an optimal set of instigators that makes the epidemic process most effective [@Kempe03]. The contagion threshold for each instigator was calculated and averaged to determine $q_0^c$ of the network. Ten kinds of networks were generated for each case, and the results of them were averaged. Differences in the results by different implementations of networks are very small.
Results and discussion
======================
{width="13.0cm"}
Contagion threshold without bilingual option
--------------------------------------------
The contagion threshold $q_0$ without bilingual option ($c=\infty$) is shown in Fig. \[Fig\_AB\]. Size dependence on $q_0$ is very small if it exists. We consider the effects of degree heterogeneity and clustering, which is not conclusive yet in spite of various studies [@Watts2002; @Montanari10; @Young11; @Acemoglu11; @Coupechoux2011; @Coupechoux2014].
A node $v_i$ having only one neighbor that had adopted innovative option $A$ changes its strategy into $A$ if and only if $q<1/d(v_i)$, where $d(v_i)$ is its degree. Therefore, in regular graphs such as regular lattices and RRNs, the contagion threshold $q_0$ is $1/\Delta$. In SFNs, $q_0$ also decreases with average degree $\langle\Delta\rangle$, but the decrement is smaller than regular graphs, and $q_0$ is higher than $1/\langle\Delta\rangle$ for $\langle\Delta\rangle \geq 8$. We confirmed that SFNs have higher $q_0$ than regular graphs at least up to $\langle\Delta\rangle=50$. Thus, the contagion threshold is positively correlated with the degree heterogeneity in the two families of graphs except for networks of extremely small average degree ($\langle\Delta\rangle < 6$). According to [@Dunbar93], humans have about 150 relations on average at any given time, among which about 12-20 relations are very intimate (sympathy group) [@Zhou05]. Therefore, we conclude that degree heterogeneity promotes the epidemics in contagion games for realistic cases. In [@Watts2002], however, it was shown that classical random networks, which have the Poisson degree distribution, have larger contagion threshold than SFNs for average degree $1 < \langle \Delta \rangle \le 30$. This discrepancy of the results in [@Watts2002] and our result is not necessarily contradictory, and we claim that it is mainly due to the structural differences in SFNs.
Heterogeneity of networks affects the epidemic process in two ways. The first one is the [*friendship paradox effect*]{}. The instigator is chosen at random and its average degree is $\langle \Delta \rangle$, but the average degree of its neighbor $\langle \Delta' \rangle$ is larger than $\langle \Delta \rangle$ in non-regular networks: $$\begin{aligned}
\langle \Delta' \rangle &=& \frac{\sum_{v \in V(G)} \left[d(v)\right]^2}{\sum_{v \in V(G)} d(v)}
= \frac{\sum_{v \in V(G)} \left[d(v)\right]^2}{2e}
= \langle \Delta \rangle + \frac{\sigma^2}{ \langle \Delta \rangle} ,\end{aligned}$$ where $e$ is the number of links and $\sigma$ is the standard deviation of the degree distribution of the graph [@Feld91]. Therefore, it is harder to persuade its neighbors in a highly-heterogeneous network, whose standard deviation of the degree distribution is large. On the other hand, heterogeneity may promote the contagion process through the mechanism that we call the [*detour effect*]{}. On average, it is harder for the instigator to persuade its neighbors by the friendship paradox effect, but some neighbors of the instigator may have a degree smaller than $\langle \Delta \rangle$ and are easy to persuade; the contagion process can continue through these channels. Neighbors with a large degree are hard to persuade, but it is possible to persuade them later with the help of other paths from the instigator to the nodes. Therefore, heterogeneity may enhance the contagion process when there are cycles. The detour effect becomes dominant only for high degree nodes, and so the enhancement of the contagion threshold by heterogeneity is effective only in networks with a high average degree as shown in Fig. \[Fig\_AB\](a). The most probable degree is $\langle \Delta \rangle/2$ in the SFNs made by the growing method in this work. To the contrary, the SFNs used in [@Watts2002] are made by the configuration model and the most frequent value of the degree is $\Delta=1$; therefore, the detour effect is ineffective and heterogeneity suppresses the contagion threshold through the friendship paradox effect.
Figure \[Fig\_AB\] shows a positive correlation between the clustering coefficient and contagion threshold on SFNs: when the average degree is fixed, $q_0$ is higher in Holme-Kim SFNs than in Barabási-Albert SFNs. Interestingly, this effect vanishes in networks with an extremely small average degree, consistently with [@Coupechoux2014]. It can be understood also by the detour effect in heterogeneous networks: the detour effect is reinforced by short cycles, and so clustering increases the contagion threshold in heterogeneous networks except when the average degree is extremely small.
Contagion threshold with bilingual option
-----------------------------------------
In the model with the bilingual option $AB$, Fig. \[Fig\_ABAB\] presents the contagion threshold $q_0^c$ for lattices, RRNs, and SFNs. In spite of the cost $c$, the bilingual option ($AB$) can be more advantageous since it is compatible with both options $A$ and $B$ resulting in higher payoffs. Thus, the contagion threshold of a contagion game with the bilingual option is a function of the cost $c$.
In the $\Delta$-regular finite graph, to be epidemic, at least one of the neighbors of the instigator must choose either $A$ or $AB$ at the first step. If $q < {1}/{\Delta}$ and $c > (\Delta - 1)q$, then it chooses $A$ (see Fig. \[Fig\_ABAB\_schematic\](a)). In this case, the contagion game continues to be epidemic without the bilingual option. If $c < (\Delta - 1)q$ and $c < 1-q$, then a neighbor of the instigator chooses $AB$ (see Fig. \[Fig\_ABAB\_schematic\](b)), and the final state can be epidemic or a mixture of the three strategies. Therefore, the contagion threshold $q_0^c$ is at most $1/\Delta$ if $c > ({\Delta-1})/{\Delta}$, and $q_0^c < 1-c$ if $c < ({\Delta-1})/{\Delta}$ for a $\Delta$-regular finite graph. Interestingly, for various $c$ and $q$ on $\Delta$-regular graphs, there appears qualitative change across lines $c = (\Delta - 1)q$ and $c=1-q$; $AB$ becomes effective only if $c$ is low below these lines. We call this region ($c < (\Delta - 1)q$ and $c < 1-q$) as the [*low-cost region*]{}.
{width="13.0cm"}
Figure \[Fig\_ABAB\] shows that qualitative change occurs across the boundary of the low-cost region also in non-regular networks as well as in regular graphs. Therefore, the boundary is supposed to be general, though it was derived in regular graphs. In the high-cost region ($c>(\langle\Delta\rangle - 1)q$ or $c > 1-q$), $q_0^c$ is the same as $q_0$ without bilingual option or approaches $q_0$ fast. Hence, the bilingual option reflects its effect in the low-cost region ($c < (\langle\Delta\rangle - 1)q$ and $c<1-q$). In the low-cost region, $q_0^c$ increases as the bilingual cost is reduced except RRNs, which have somewhat smaller $q_0^c$ than $q_0$ near the boundary of $c=1-q$. The decrease of $q_0^c$ means that the bilingual option can hinder the contagion process by protecting the status quo option from extinction. Contagion threshold is largest in lattices, intermediate in SFNs, and smallest in RRNs. Interestingly, the contagion threshold of RRNs is almost the same as that of the infinite regular tree, which is proved to have the smallest value among all infinite regular graphs [@Yu_Choi17]. It is easy to see that $q_0^c$ for a finite regular lattice is as small as the contagion threshold defined for an infinite regular graph with a bilingual option as in [@Immorlica07; @Yu_Choi17]. Therefore, it is highly probable that RRN has relatively small contagion threshold among all finite graphs with $N$ nodes. It is understandable because RRN has a tree-like structure and the (average) number of cycles with a fixed length is reasonably small [@Bollobas]. In other words, both networks have small clustering and no short cycles. In contrast to the high-cost region, $q_0^c$ increases with the average degree in all networks considered in this paper. Clustering always enhances the contagion threshold. Size-dependence is negligible except for the Barabási-Albert SFNs, whose $q_0^c$ decreases a little bit as the number of nodes in the network.
![ Regions $(q,c)$ for the neighbor of the instigator to change its strategy into $A$ in (a) and into $AB$ in (b) for $\Delta$-regular graphs. The red, blue, and purple straight lines represent $c=(\Delta-1)q$, $q=1/\Delta$, and $c=1-q$. Yellow and white regions in (b) show the low-cost and high-cost regions, respectively.[]{data-label="Fig_ABAB_schematic"}](figure4.eps){width="10.0cm"}
Summary
=======
We defined and calculated the contagion threshold in finite graphs: regular lattices, RRNs, and SFNs. Without the bilingual option, the contagion threshold $q_0$ in regular graphs is $q_0 = 1/\Delta$ and independent of the network structure. The contagion threshold in SFNs is larger than that of regular graphs for average degree $\langle\Delta\rangle\geq 8$. This can be understood by the friendship paradox effect and detour effect in SFNs. Clustering in SFNs enhances the detour effect to increase the contagion threshold. The bilingual option makes a qualitative change of contagion threshold when the bilingual cost $c$ is lower than $(1-q)$ and $(\langle\Delta\rangle-1)q$. With the bilingual option of low cost, $q_0^c$ is largest in lattices, intermediate in SFNs, and smallest in RRNs. The contagion threshold of RRNs of degree $\Delta$ is almost the same as the $\Delta$-regular tree, which has the smallest $q_0^c$ among $\Delta$-regular networks. In both cases of regular networks and SFNs, the innovative option is easier to spread as the clustering coefficient of the network increases. We found little network size dependence on the contagion threshold except in Barabási-Albert SFNs with a low-cost bilingual option.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by GIST Research Institute (GRI) grant funded by the GIST in 2019.
References {#references .unnumbered}
==========
Clustering coefficient in regular random networks
=================================================
We consider a $\Delta$-regular random network of $N$ vertices, which is assumed to be large. Suppose that a vertex $x$ is connected to vertices $y$ and $z$. The clustering coefficient $\xi$ is the probability that vertices $y$ and $z$ are connected. Among $\Delta$ arms of $y$, one edge connects $x$ and $y$, and $y$ has $(\Delta-1)$ arms to be connected. And there are $(N-2)$ vertices that can be connected to $y$ excluding $x$ and $y$. Therefore, The probability that vertices $y$ and $z$ are not connected $(1-\xi)$ can be calculated as $$\begin{aligned}
&&1-\xi = \left( \frac{N-3}{N-2} \right)
\left( \frac{N-4}{N-3} \right)
\left( \frac{N-5}{N-4} \right)
\cdots
\left( \frac{N-(\Delta+1)}{N-\Delta} \right) \\
&&~~~~~~~ = \left( 1 - \frac{1}{N-2} \right)
\left( 1 - \frac{1}{N-3} \right)
\left( 1 - \frac{1}{N-4} \right)
\cdots
\left( 1 - \frac{1}{N-\Delta} \right)\\
&&~~~~~~~ \approx 1 - \left(\frac{1}{N-2}\right) - \left(\frac{1}{N-3}\right)
- \left(\frac{1}{N-4}\right)
- \cdots - \left(\frac{1}{N-\Delta}\right)\end{aligned}$$ in the limit of $N \gg \Delta$. Therefore, the clustering coefficient or the probability for $y$ and $z$ to be connected to each other is $$\begin{aligned}
\xi \approx \left(\frac{1}{N-2}\right) + \left(\frac{1}{N-3}\right)
+ \left(\frac{1}{N-4}\right)
+ \cdots + \left(\frac{1}{N-\Delta}\right) \approx \frac{\Delta-1}{N} .\end{aligned}$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have measured the neutron structure function g$_{2}^{n}$ and the virtual photon-nucleon asymmetry A$_{2}^{n}$ over the kinematic range $0.014\leq x \leq 0.7$ and $1.0 \leq Q^{2} \leq 17.0$ by scattering 48.3 GeV longitudinally polarized electrons from polarized $^{3}$He. Results for A$_{2}^{n}$ are significantly smaller than the $\sqrt{R}$ positivity limit over most of the measured range and data for g$_2^{n}$ are generally consistent with the twist-2 Wandzura-Wilczek prediction. Using our measured g$_{2}^{n}$ we obtain results for the twist-3 reduced matrix element $d_{2}^{n}$, and the integral $\int$g$_{2}^{n}(x)dx$ in the range $0.014\leq x \leq 1.0$. Data from this experiment are combined with existing data for g$_{2}^{n}$ to obtain an average for $d_{2}^{n}$ and the integral $\int$g$_{2}^{n}(x)dx$.'
address:
- 'The American University, Washington D.C. 20016'
- 'Université Blaise Pascal, LPC IN2P3/CNRS, F-63170 Aubière Cedex, France'
- 'University of California, Berkeley, California 94720-7300'
- 'University of California, Los Angeles, California 90024'
- 'California Institute of Technology, Pasadena, California 91125'
- 'Centre d’Etudes de Saclay, DAPNIA/SPhN, F-91191 Gif-sur-Yvette, France'
- 'Kent State University, Kent, Ohio 44242'
- 'University of Massachusetts, Amherst, Massachusetts 01003'
- 'University of Michigan, Ann Arbor, Michigan 48109'
- 'National Institute of Standards and Technology, Gaithersburg, Maryland 20899'
- 'Northwestern University, Evanston, Illinois 60201'
- 'Old Dominion University, Norfolk, Virginia 23529'
- 'University of Pennsylvania, Philadelphia, Pennsylvania 19104'
- 'Princeton University, Princeton, New Jersey 08544'
- 'Smith College, Northampton, Massachusetts 01063'
- 'Southern Oregon State College, Ashland, Oregon 97520'
- 'Stanford Linear Accelerator Center, Stanford, California 94309'
- 'Syracuse University, Syracuse, New York 13210'
- 'Temple University, Philadelphia, Pennsylvania 19122'
- 'Tohoku University, Aramaki Aza Aoba, Sendai, Miyagi, Japan'
- 'College of William and Mary, Williamsburg, Virginia 23187'
- 'University of Wisconsin, Madison, Wisconsin 53706'
author:
- 'K. Abe'
- 'T. Akagi'
- 'B. D. Anderson'
- 'P. L. Anthony'
- 'R. G. Arnold'
- 'T. Averett'
- 'H. R. Band'
- 'C. M. Berisso'
- 'P. Bogorad'
- 'H. Borel'
- 'P. E. Bosted'
- 'V. Breton'
- 'M. J. Buenerd'
- 'G. D. Cates'
- 'T. E. Chupp'
- 'S. Churchwell'
- 'K. P. Coulter'
- 'M. Daoudi'
- 'P. Decowski'
- 'R. Erickson'
- 'J. N. Fellbaum'
- 'H. Fonvieille'
- 'R. Gearhart'
- 'V. Ghazikhanian'
- 'K. A. Griffioen'
- 'R. S. Hicks'
- 'R. Holmes'
- 'E. W. Hughes'
- 'G. Igo'
- 'S. Incerti'
- 'J. R. Johnson'
- 'W. Kahl'
- 'M. Khayat'
- 'Yu. G. Kolomensky'
- 'S. E. Kuhn'
- 'K. Kumar'
- 'M. Kuriki'
- 'R. Lombard-Nelsen'
- 'D. M. Manley'
- 'J. Marroncle'
- 'T. Maruyama'
- 'T. Marvin'
- 'W. Meyer'
- 'Z.-E. Meziani'
- 'D. Miller'
- 'G. Mitchell'
- 'M. Olson'
- 'G. A. Peterson'
- 'G. G. Petratos'
- 'R. Pitthan'
- 'R. Prepost'
- 'P. Raines'
- 'B. A. Raue'
- 'D. Reyna'
- 'L. S. Rochester'
- 'S. E. Rock'
- 'M. V. Romalis'
- 'F. Sabatie'
- 'G. Shapiro'
- 'J. Shaw'
- 'T. B. Smith'
- 'L. Sorrell'
- 'P. A. Souder'
- 'F. Staley'
- 'S. St. Lorant'
- 'L. M. Stuart'
- 'F. Suekane'
- 'Z. M. Szalata'
- 'Y. Terrien'
- 'A. K. Thompson'
- 'T. Toole'
- 'X. Wang'
- 'J. W. Watson'
- 'R. C. Welsh'
- 'F. R. Wesselmann'
- 'T. Wright'
- 'C. C. Young'
- 'B. Youngman'
- 'H. Yuta'
- 'W.-M. Zhang'
- 'P. Zyla'
title: 'Measurement of the Neutron Spin Structure Function g$_{2}^{n}$ and Asymmetry A$_{2}^{n}$'
---
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
and
The deep inelastic spin structure functions g$_{1}(x,Q^{2})$ and g$_{2}(x,Q^{2})$, which depend on the Bjorken scaling variable $x$ and the virtual photon four-momentum squared $-Q^{2}$, provide insight into the internal spin structure of the nucleon. A large set of data for g$_{1}$ now exists for the proton, deuteron [@SMC; @E143] and neutron [@E142_2; @E154]. These data have been used to test the fundamental Bjorken sum rule, and within the framework of the quark-parton model (QPM), to measure the quark contribution to the nucleon’s spin. The g$_{2}$ structure function contains contributions from both the longitudinal and transverse polarization distributions within the nucleon. It is sensitive to higher twist effects such as quark-gluon correlations and quark mass contributions, and is not easily interpreted in the QPM where such effects are not included. However, by interpreting g$_{2}$ using the operator product expansion (OPE) within QCD [@Vain; @Jaffe], it is possible to study contributions to the nucleon spin structure beyond the simple QPM.
The OPE allows us to write the hadronic matrix element in deep inelastic scattering (DIS) in terms of a series of renormalized operators of increasing twist [@Vain; @Jaffe]. The leading contribution is twist-2, with higher twist terms suppressed by powers of $1/Q$. Keeping only terms up to twist-3, the moments of g$_{1}$ and g$_{2}$ at fixed $Q^{2}$ can be related to the twist-2 and twist-3 reduced matrix elements, $a_{j}$ and $d_{j}$ [@Jaffe], $$\begin{aligned}
\label{eq:moments}
\int_{0}^{1}x^{j}{\rm{g}}_{1}(x,Q^{2})dx=\frac{a_{j}}{2},
\;\;j=0,2,4,... \nonumber \\
\int_{0}^{1}x^{j}{\rm{g}}_{2}(x,Q^{2})dx=\frac{1}{2}\frac{j}{j+1}(d_{j}-a_{j}),
\;\;j=2,4,...\end{aligned}$$ In the expressions above, $d_{j}$ directly appears in the equation for g$_{2}$ allowing us to study the higher twist structure of the nucleon at leading order. An expression for the twist-2 part of g$_{2}$ was derived by Wandzura and Wilczek [@g2ww] from these sum rules assuming that the twist-3 contributions $d_{j}$, are negligible, $${\rm{g}}_{2}^{WW}(x,Q^{2})=-{\rm{g}}_{1}(x,Q^{2}) + \int_{x}^{1}
\frac{{\rm{g}}_{1}(x',Q^{2})}{x'} dx'.
\label{eq:g2ww}$$ Comparing measured values of g$_{2}$ with this prediction enables us to extract information about higher twist contributions to g$_2$. There is an additional twist-2 contribution to g$_{2}$ [@Song; @PCR] beyond the g$_{2}^{WW}$ term which arises from the transverse polarization density in the nucleon, $h_{T}(x,Q^{2})$. However, this term is suppressed by the ratio of the quark to nucleon mass $m/M$ in DIS [@Song] and will be neglected in this analysis.
The structure function g$_{2}(x,Q^{2})$ may be expressed in terms of two measurable asymmetries, $A_{\parallel}(x,Q^{2})$ and $A_{\perp}(x,Q^{2})$, corresponding to longitudinal and transverse target polarization with respect to the incoming electron beam helicity, $${\rm{g}}_{2}(x,Q^{2})=\frac{F_{2}(x,Q^{2})(1+\gamma ^{2})}
{2x\left[1+R(x,Q^{2})\right]}
\frac{y}{2d\,{\rm{sin}}\, \theta}
\biggl[A_{\perp}\frac{E+E'{\rm{cos}}\, \theta}{E'}
-A_{\parallel}\: {\rm{sin}}\, \theta \biggr],$$ where $E$ and $E'$ are the incident and scattered electron energies, $\theta$ is the scattering angle, $\gamma=2Mx/\sqrt{Q^{2}}$, $y=(E-E')/E$, $d=(1-\epsilon)(2-y)/y[1+\epsilon R(x,Q^{2})]$, and $\epsilon^{-1}=
1+2\left[1+\gamma ^{-2}\right]{\rm{tan}}^{2}(\theta /2)$. Fits to existing data were used for the unpolarized structure function $F_{2}(x,Q^{2})$ [@NMC] and for $R(x,Q^{2})$ [@R1990], the ratio of longitudinal to transverse virtual photon absorption cross sections. At small scattering angles, the term $A_{\parallel}\:{\rm{sin}}\,\theta$ is small, and consequently the dominant contribution to g$_2$ comes from $A_{\perp}$.
Spin dependent DIS can also be described in terms of the spin asymmetries A$_{1}(x,Q^{2})$ and A$_{2}(x,Q^{2})$ for virtual photon absorption. The asymmetry A$_2(x,Q^{2})$ is bounded by the positivity limit $|{\rm{A}}_{2}(x,Q^{2})|\leq \sqrt{R(x,Q^{2})}$, and like g$_{2}$, it is dominated by $A_{\perp}$, $${\rm{A}}_{2}(x,Q^{2})=\frac{\gamma (2-y)}{2d\,
{\rm{sin}}\,\theta }\left[A_{\perp}
\frac{y(1+xM/E)}{(1-y)} + A_{\parallel}\: {\rm{sin}}\,\theta \right].$$
Measurements of g$_2$ and A$_{2}$ exist for the proton [@E143g2; @SMCg2] and deuteron [@E143g2], and in the case of the neutron, a measurement was made at the Stanford Linear Accelerator Center (SLAC) by scattering 26 GeV polarized electrons from polarized $^3$He [@E142_2]. In this Letter, we report a new measurement of g$_{2}$ and A$_{2}$ for the neutron made during experiment E154 at SLAC. For this experiment, the beam energy was increased to 48.3 GeV and two new large-acceptance spectrometers were constructed to provide broader kinematic coverage than previously measured. Results from this experiment for g$_{1}^{n}$ and A$_{1}^{n}$ have been reported elsewhere [@E154], and we focus here on the measurement of $A_{\perp}$ and the subsequent determination of g$_{2}^{n}$ and A$_{2}^{n}$.
The target was a 30 cm long, thin-walled glass cell containing approximately 10.5 atmospheres (as measured at $20^{\circ}$C) of $^{3}$He gas. The helium nuclei were polarized by spin-exchange collisions with rubidium atoms that were polarized by optical pumping [@target]. The system was designed to allow continuous pumping of the target polarization in the longitudinal direction only. Therefore, to obtain transverse polarization, the $^3$He spins were first pumped to a longitudinal polarization of 48% and then rotated to the transverse direction using two orthogonal sets of Helmholtz coils. In the transverse orientation, the polarization decreased to 33% over a period of 24 hours, at which time the target was re-polarized. Approximately $7\times 10^{6}$ electron events were recorded during two cycles of transverse running.
The electron beam was produced in 250 ns pulses at a rate of 120 Hz, each containing approximately $3\times 10^{10}$ electrons. The average beam polarization was measured to be $0.826 \pm 0.023$ using a M[ø]{}ller polarimeter [@moller], and the helicity of each pulse was chosen randomly to reduce helicity-dependent systematic errors. Scattered electrons were measured using two independent spectrometers at scattering angles centered around $2.75 ^{\circ}$ and $5.5^{\circ}$, and the asymmetry $A_{\perp}$ was calculated as $$A_{\perp}=\left(\frac{N^{-}-N^{+}}{N^{-}+
N^{+}}\right)\frac{1}{P_{b}P_{t}f},$$ where $N^{+}$ and $N^{-}$ are the measured electron rates for positive and negative beam helicities corrected for detection efficiency and normalized to the incident charge, $P_{b}$ and $P_{t}$ are the beam and target polarizations, and $f$ is the dilution factor which corrects for electrons that scattered from materials in the target system other than $^3$He.
Our measured asymmetry included not only DIS events, but also pions mis-identified as electrons, and electrons produced in charge symmetric hadron decays. The rates and asymmetries for these backgrounds were measured and used to correct $A_{\perp}$. The asymmetry was also corrected for internal [@Kukhto] and external [@Tsai] radiative effects. Uncertainties in the radiative corrections were estimated by varying the input models over a range consistent with the measured data. Finally, a neutron result was extracted from $A_{\perp}$ by applying a correction for the $^{3}$He nuclear wave function [@he3_1] and using g$_{2}^{WW}$ obtained from a fit to existing proton data [@SMC; @E143] for g$_{1}^{p}$.
Results for A$_{2}^{n}$ and g$_{2}^{n}$ from both spectrometers are given in Table \[tab:data\] with statistical and systematic errors. The data cover the kinematic range $0.014\leq x \leq 0.7$ and $1.0\leq Q^{2} \leq 17.0$ (GeV/c)$^{2}$ with an average $Q^2$ of $3.6$ (GeV/c)$^2$. Systematic errors are dominated by the uncertainty in the radiative corrections, but are significantly smaller than the statistical error over the entire data range. No evidence of $Q^{2}$ dependence was seen for A$_{2}^{n}$ or g$_{2}^{n}$ within the experimental errors and the data from both spectrometers were averaged. The results for A$_{2}^{n}$ are shown in Fig. \[a2\] along with the $\sqrt{R}$ positivity limit and previous data from SLAC experiment E142 [@E142_2]. The data are in good agreement with the E142 measurement and are significantly smaller than the positivity limit over most of the measured range. Results for $x{\rm{g}}_{2}^{n}$ are shown in Fig. \[xg2\] along with the twist-2 prediction, $x{\rm{g}}_{2}^{WW}$. To calculate g$_{2}^{WW}$, we assume that g$_{1}/F_{1}$ is independent of Q$^2$, and use a fit to our measured g$_1^n$ data [@E154]. A comparision of our data with g$_{2}^{WW}$ over the measured range gives a $\chi^{2}/$(dof) of $1.02$ indicating good overall agreement. However, the data clearly do not rule out the possibility of large twist-3 contributions and show marginal agreement with the twist-2 prediction in the region $0.03<x<0.1$.
[rrcrr]{} & & & &\
& & & &\
\
0.014 - 0.02 & 0.017 & 1.2 & $0.03\pm 0.07\pm 0.01$ & $7.36\pm
15.74\pm 2.24$\
0.02 - 0.03 & 0.025 & 1.6 & $0.00\pm 0.06\pm 0.01$ & $0.15\pm
7.19\pm 0.98$\
0.03 - 0.04 & 0.035 & 2.1 & $-0.11\pm 0.06\pm 0.01$ & $-7.90\pm
4.91\pm 0.96$\
0.04 - 0.06 & 0.049 & 2.6 & $0.10\pm 0.06\pm 0.01$ & $4.60\pm
2.50\pm 0.54$\
0.06 - 0.10 & 0.077 & 3.4 & $0.06\pm 0.06\pm 0.01$ & $1.32\pm
1.34\pm 0.25$\
0.10 - 0.15 & 0.122 & 4.1 & $0.13\pm 0.11\pm 0.03$ & $1.22\pm
0.95\pm 0.24$\
0.15 - 0.20 & 0.173 & 4.7 & $-0.03\pm 0.18\pm 0.03$ & $-0.08\pm
0.81\pm 0.14$\
0.20 - 0.30 & 0.242 & 5.1 & $-0.25\pm 0.24\pm 0.05$ & $-0.48\pm
0.51\pm 0.11$\
0.30 - 0.40 & 0.341 & 5.6 & $0.63\pm 0.55\pm 0.13$ & $0.54\pm
0.46\pm 0.15$\
0.40 - 0.50 & 0.425 & 5.9 & $0.16\pm 1.40\pm 0.04$ & $0.04\pm
0.57\pm 0.02$\
\
0.06 - 0.10 & 0.084 & 5.5 & $0.16\pm 0.10\pm 0.02$ & $4.08\pm
2.40\pm 0.43$\
0.10 - 0.15 & 0.123 & 7.2 & $0.01\pm 0.08\pm 0.02$ & $0.23\pm
1.00\pm 0.20$\
0.15 - 0.20 & 0.173 & 8.9 & $0.05\pm 0.11\pm 0.02$ & $0.40\pm
0.72\pm 0.15$\
0.20 - 0.30 & 0.242 & 10.7 & $0.15 \pm 0.14\pm 0.03$ & $0.48\pm
0.41\pm 0.10$\
0.30 - 0.40 & 0.342 & 12.5 & $-0.21\pm 0.27\pm 0.03$ & $-0.22\pm 0.31\pm 0.04$\
0.40 - 0.50 & 0.442 & 13.8 & $-0.36\pm 0.53\pm 0.05$ & $-0.16\pm 0.24\pm 0.03$\
0.50 - 0.70 & 0.564 & 15.0 & $-0.04\pm 0.96\pm 0.06$ & $-0.01\pm 0.13\pm 0.01$\
\[tab:data\]
In order to quantify the possible contribution of higher twist effects to g$_{2}^{n}$, Eq. \[eq:moments\] can be solved for the twist-3 reduced matrix elements $d_{j}^{n}$ at fixed $Q^2$, $$d_{j}^{n}(Q^{2})=2\int_{0}^{1}x^{j}\left[{\rm{g}}_{1}^{n}(x,Q^{2})
+\left(
\frac{j+1}{j}\right){\rm{g}}_{2}^{n}(x,Q^{2})\right]dx,\;\; j=2,4,...
\label{eq:d2}$$ The combination of g$_{1}$ and g$_{2}$ in the above expression effectively cancels any twist-2 components allowing us to look for a net twist-3 contribution to g$_{2}$. The matrix element was calculated using our g$_{2}(x,Q^{2})$ data and a fit to our measured g$_{1}(x,Q^{2})$. Because the integrand in Eq. \[eq:d2\] is purely twist-3, we assumed the unmeasured region $0.7\leq x \leq 1$ behaves like $(1-x)^{2}$ as suggested by Brodsky [*[et al.]{}*]{} [@Brodsky], and fit our last data point to this form to extrapolate to $x=1$. We neglected any contribution from the region $0\leq x < 0.014$ because it is suppressed by the $x^{j}$ term. For the $j\!=\!2$ moment, we obtain a value of $d_{2}^{n}=-0.004\pm 0.038 \pm 0.005$ with an average $Q^{2}$ of $3.6$ (GeV/c)$^{2}$. The contribution from the high-$x$ extrapolation is much smaller than the experimental errors and does not significantly change the result for the matrix element.
Data from SLAC experiments E142 [@E142_2] and E143 [@E143g2] were combined with this experiment to yield a average neutron result for g$_{2}$ with an average $Q^{2}$ of $3.0$ (GeV/c)$^{2}$. Neutron results were extracted from the E143 proton and deuteron data assuming a 5% D-state in the deuteron. The results are shown in Fig. \[fig3\] along with the g$_{2}^{WW}$ prediction. Comparing the combined data with g$_{2}^{WW}$ gives a $\chi^{2}/$(dof) of $1.01$, again indicating good agreement with g$_{2}^{WW}$. Using the combined data, we obtain the result $d_{2}^{n}=-0.010\pm 0.015$ at an average $Q^{2}$ of $3.0$ (GeV/c)$^{2}$. The measured $d_{2}^{n}$ along with various model predictions are summarized in Table \[tab:d2\], and while the data are consistent with zero, the precision is insufficient to rule out models which contain significant twist-3 contributions.
$d_{2}^{n}\times 10^{2}$ $Q^{2}$ (GeV/c)$^{2}$
--------------------------------- -------------------------- -----------------------
E154 result $-0.4\pm 3.8$ $3.6$
SLAC Average $-1.0\pm 1.5$ $3.0$
Bag model [@Song] $-0.253$ $5.0$
Bag model [@E143g2; @Stratmann] $0.03$ $5.0$
Bag model [@Ji] $0$ $5.0$
QCD sum rule [@Stein] $-3\pm 1$ $1.0$
QCD sum rule [@BBK] $-2.7\pm 1.2$ $1.0$
Lattice QCD [@LQCD] $-0.39\pm 0.27$ $4.0$
: Comparison of experimental and theoretical results for the reduced twist-3 matrix element $d_{2}^{n}$.
\[tab:d2\]
From the OPE it is not possible to obtain an expression for the $j\!=\!0$ moment of g$_{2}$. However, Burkhardt and Cottingham [@BCsum] have derived the sum rule $\int_{0}^{1}{\rm{g}}_{2}(x)dx=0$, which is valid to first order in pQCD [@altarelli], using dispersion relations for virtual Compton scattering. To evaluate the integral, the g$_{2}^{WW}$ expression in Eq. \[eq:g2ww\] was used to evolve the twist-2 part of our measured g$_{2}$ to a $Q^{2}$ of $3.6$ (GeV/c)$^{2}$ assuming g$_{1}/F_{1}$ is independent of $Q^{2}$ and fitting our g$_{1}^{n}$ data as before. At large $x$, we see from Eq. \[eq:g2ww\] that g$_{2}^{WW} \approx -$g$_1$ and we therefore assume that g$_{2}\propto (1-x)^{3}$ for the extrapolation to $x=1$. The result is $\int_{0.014}^{1}{\rm{g}}_{2}(x)dx=0.19\pm 0.17 \pm 0.02$ with an average $Q^{2}$ of $3.6$ (GeV/c)$^{2}$. The $Q^{2}$ evolution and high-$x$ extrapolation do not contribute significantly to the integral and the uncertainties in these quantites are included in the error. Combining this result with data from SLAC experiments E142 [@E142_2] and E143 [@E143g2] yields a result of $\int_{0.014}^{1}{\rm{g}}_{2}(x)dx=0.06\pm 0.15$ at an average $Q^{2}$ of $3.0$ (GeV/c)$^{2}$, which is consistent with the Burkhardt-Cottingham sum rule. However, this does not represent a conclusive test of the sum rule because the behavior of g$_{2}^{n}$ as $x\rightarrow\!0$ is not known.
In summary, we have presented a new measurement of A$_{2}^{n}$ and g$_{2}^{n}$ in the kinematic range $0.014 \leq x \leq 0.7$ and $1.0 \leq Q^{2}
\leq 17.0$ (GeV/c)$^{2}$. Our results for A$_{2}^{n}$ are significantly smaller than the $\sqrt{R}$ positivity limit over most of the measured range and data for g$_{2}^{n}$ are generally consistent with the twist-2 g$_{2}^{WW}$ prediction. The values obtained for the twist-3 matrix element $d_{2}^{n}$ from this measurement and the SLAC average are also consistent with zero. However, further measurements are needed to make a conclusive statement about the higher twist content of the nucleon.
We wish to thank the personnel of the SLAC accelerator department for their efforts which resulted in the successful completion of the E154 experiment. We would also like to thank J. Ralston for useful discussions and guidance. This work was supported by the Department of Energy; by the National Science Foundation; by the Kent State University Research Council (GGP); by the Jeffress Memorial Trust (KAG); by the Centre National de la Recherche Scientifique and the Commissariat a l’Energie Atomique (French groups); and by the Japanese Ministry of Education, Science and Culture (Tohoku).\
[99]{} B. Adeva [*[et al.]{}*]{}, Phys. Lett. B [**[302]{}**]{} (1993) 533; D. Adams [*[et al.]{}*]{}, Phys. Lett. B [**[329]{}**]{} (1994) 399 ; B. Adeva [*[et al.]{}*]{}, Phys. Lett. B [**[357]{}**]{} (1995) 248. K. Abe [*[et al.]{}*]{}, Phys. Rev. Lett. [**[74]{}**]{} (1995) 346; K. Abe [*[et al.]{}*]{}, Phys. Rev. Lett. [**[75]{}**]{} (1995) 25. P. Anthony [*[et al.]{}*]{}, Phys. Rev. D [**[54]{}**]{} (1996) 6620. K. Abe [*[et al.]{}*]{}, SLAC–PUB–7459 (1997), submitted to Phys. Rev. Lett. E. Shuryak and A. Vainshtein, Nuc. Phys. (1982) 141. R. Jaffe and X. Ji, Phys. Rev. D [**[43]{}**]{} (1991) 724. S. Wandzura and F. Wilczek, Phys. Lett. B [**[72]{}**]{} (1977) 195. X. Song, Phys. Rev. D [**[54]{}**]{} (1996) 1955. J. Cortes, B. Pire and J. Ralston, Z. Phys. C [**[55]{}**]{} (1992) 409. M. Arneodo [*[et al.]{}*]{}, Phys. Lett. B [**[364]{}**]{} (1995) 107. L. Whitlow [*[et al.]{}*]{}, Phys. Lett. B [**[250]{}**]{} (1990) 193. K. Abe [*[et al.]{}*]{}, Phys. Rev. Lett. [**[76]{}**]{} (1996) 587. D. Adams [*[et al.]{}*]{}, Phys. Lett. B [**[336]{}**]{} (1994) 125. J. Johnson [*[et al.]{}*]{}, Nuc. Instrum. Meth. A [**[356]{}**]{} (1995) 148. H. R. Band [*[et al.]{}*]{}, SLAC-PUB-7370 (1997), submitted to Nucl. Instrum. Meth. T. Kuchto and N. Shumeiko, Nuc. Phys. [**[B219]{}**]{} (1983) 412; I. Akusevich and M. Shumeiko, J. Phys. G [**[20]{}**]{} (1994) 513. Y. Tsai, SLAC-PUB-848 (1971); Rev.Mod. Phys. [**[46]{}**]{} (1974) 815. C. Ciofi degli Atti [*[et al.]{}*]{}, Phys. Rev. C [**[48]{}**]{} (1993) 968. S. Brodsky [*[et al.]{}*]{}, Nucl. Phys. (1992) 519. M. Stratmann, Z. Phys. C [**[60]{}**]{} (1993) 763. X. Ji and P. Unrau, Phys. Lett. B [**[333]{}**]{} (1994) 228. E. Stein [*[et al.]{}*]{}, Phys. Lett. B [**[343]{}**]{} (1995) 369. I. Balitsky, V. Braun and A. Kolesnichenko, Phys.Lett. B [**[242]{}**]{} (1990) 245; B [**[318]{}**]{} (1993) 648 (Erratum). M. Göckeler [*[et al.]{}*]{}, Phys. Rev. D [**[53]{}**]{} (1996) 2317. H. Burkhardt and W. Cottingham, Ann. Phys. (1970) 453. G. Altarelli [*[et al.]{}*]{}, Phys. Lett. B [**[334]{}**]{} (1994) 187.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate galactic winds driven by supernova (SN) explosions in an isolated dwarf galaxy using high-resolution (particle mass $m_{\rm gas} = 1{\rm M_\odot}$, number of neighbor $N_{\rm ngb} = 100$) smoothed-particle hydrodynamics simulations that include non-equilibrium cooling and chemistry, individual star formation, stellar feedback and metal enrichment. Clustered SNe lead to the formation of superbubbles which break out of the disk and vent out hot gas, launching the winds. We find much weaker winds than what cosmological simulations typically adopt at this mass scale. At the virial radius, the time-averaged loading factors of mass, momentum and energy are 3, 1 and 0.05, respectively, and the metal enrichment factor is 1.5. Winds that escape the halo consist of two populations that differ in their launching temperatures. Hot gas acquires enough kinetic energy to escape when launched while warm gas does not. However, warm gas can be further accelerated by the ram pressure of the subsequently launched hot gas and eventually escape. The strong interactions between different temperature phases highlight the caveat of extrapolating properties of warm gas to large distances based on its local conditions (e.g. the Bernoulli parameter). Our convergence study finds that wind properties converge when the cooling masses of individual SNe are resolved, which corresponds to $m_{\rm gas}=5 {\rm M_\odot}$ with an injection mass of $500 {\rm M_\odot}$. The winds weaken dramatically once the SNe become unresolved. We demonstrate that injecting the terminal momentum of SNe, a popular sub-grid model in the literature, fails to capture SN winds irrespective of the inclusion of residual thermal energy.'
author:
- |
Chia-Yu Hu$^{1}$[^1]\
$^{1}$Center for Computational Astrophysics, Flatiron Institute, 162 5th Ave NY NY\
bibliography:
- 'literatur.bib'
title: 'Supernova-driven winds in simulated dwarf galaxies'
---
galaxies: dwarf - galaxies: evolution - galaxies: ISM
Introduction
============
Galactic winds, galactic scale outflows of gas originated from the interstellar medium (ISM), play a critical role in the formation and evolution of galaxies (see for a review). Understanding the origin of galactic winds is still an active field of research. Broadly speaking, the driving forces fall into two categories: active galactic nuclei (AGNs) and stellar activities (mostly due to massive stars). AGN-driven winds are thought to be important for galaxies embedded in massive dark matter halos (halo mass M$_h > 10^{12} {\rm M_\odot}$), while winds driven by stellar feedback are more important in lower-mass halos (M$_h < 10^{12} {\rm M_\odot}$). Among several mechanisms of stellar feedback, supernova (SN) explosions are of major importance in driving winds due to its dominant energy budget and its bursty nature.
The evolution of a supernova remnant (SNR) can be divided into three phases [@1988ApJ...334..252C; @1998ApJ...500..342B; @1998ApJ...500...95T]. The first phase is the “free-expansion” phase, where the fast-moving stellar ejecta expand unimpededly and gradually sweep up the interstellar medium (ISM). Once the mass of the swept-up ISM becomes comparable to the ejecta mass, it enters the energy-conserving Sedov-Taylor phase [@1959sdmm.book.....S; @1950RSPSA.201..159T]. During this phase, the radial momentum increases as the pressurized bubble expands. When the thermal energy starts to escape through radiative cooling, the radial momentum saturates and the SNR enters the momentum-conserving phase. Recent hydrodynamical simulations have reproduced the analytical results and it has been shown that the key to faithfully simulate the dynamical impact of a SN on its ambient ISM is to resolve its energy-conserving phase [@2015MNRAS.450..504M; @2015ApJ...802...99K; @2015ApJ...809...69S].
Modeling galactic winds originating from a multi-phase ISM with numerical simulations is challenging because of the high resolution it requires. This is currently not feasible for large-scale cosmological simulations, and is expected to remain so in near future. Therefore, phenomenological approaches have to be taken, making these galactic winds an input of the model rather than a prediction (see for a review).
While simulating AGN-driven winds that resolve the entire dynamic range of relevance seem to remain out of reach, it has become possible recently to simulate SN-driven winds in a representative volume of a galaxy with the required resolution and relevant ISM physics. The common approach is to simulate a patch of the galaxy (typically a Milky Way-like galaxy) with a gravitationally stratified box, assuming periodic boundary conditions along the disk plane and open boundaries in perpendicular directions [@2012ApJ...750..104H; @2016ApJ...816L..19G; @2016MNRAS.456.3432G; @2017MNRAS.466.1903G; @2018ApJ...853..173K]. While significant processes have been made, the assumption of periodicity in this type of simulations renders them insufficient to predict the large-scale properties of winds (see Appendix C in [@2016MNRAS.459.2311M] for a discussion), which can only be inferred indirectly [@2017ApJ...841..101L; @2018ApJ...853..173K].
Star-forming dwarf galaxies, besides being interesting objects on their own, provide a unique opportunity for galactic scale simulations (i.e., simulating the entire galaxy rather than just a patch of it) with the required resolution thanks to their small sizes [@2016MNRAS.458.3528H; @2017MNRAS.471.2151H]. They serve as an ideal laboratory where numerical experiments can be conducted so as to strengthen our understanding of not only the physics of the ISM but also the numerical model we adopt. High-resolution galactic scale simulations can therefore provide valuable guidelines for developing sub-grid models for cosmological simulations.
In this work, we present high-resolution ($1 {\rm M_\odot}$) hydrodynamical simulations of SN-driven winds originated from the multi-phase ISM of an isolated dwarf galaxy from pc-scales where individual SNe are resolved all the way to the virial radius of the halo. We quantify the amount of mass, momentum, energy and metals carried by the winds and investigate how winds escape the halo, taking advantage of the Lagrangian nature of our code. We study the convergence properties of winds by gradually coarsening the resolution, and we show that the resolution requirement to obtain converged wind properties in our simulations is $5 {\rm M_\odot}$, which is the scale that starts to resolve individual SNe. Finally, we explore different schemes for SN feedback and its impact on galactic winds. We demonstrate that the wind properties are insensitive to the SN schemes as long as the SNe are resolved. In addition, we find that injecting the terminal momentum of SNe, a widely adopted sub-grid model, is unable to improve the convergence due to its assumption that the thermal energy has been radiated away right after the injection, and the fact that winds are driven by thermal pressure.
This paper is structured as follows. In section \[sec:method\], we present the adopted numerical methods and the simulation setup. In section \[sec:results\], we discuss the wind properties in our simulations, present the convergence study and demonstrate the effect of different SN schemes. In section \[sec:discussion\], we discuss our results comparing to other studies in the literature. In section \[sec:summary\], we summarize our work.
Numerical Method {#sec:method}
================
The numerical methods we adopt in this paper are largely based on @2016MNRAS.458.3528H and @2017MNRAS.471.2151H, with a number of revisions. Here we briefly summarize the methods and give a detailed description of the revised aspects.
For gravity and hydrodynamics, we use [Gadget-3]{} [@2005MNRAS.364.1105S] with a modified implementation of smoothed particle hydrodynamics (SPH) [@2014MNRAS.443.1173H]. We follow non-equilibrium cooling and heating processes and a chemistry network for molecular hydrogen based on @1997ApJ...482..796N, @2007ApJS..169..239G and @2012MNRAS.421..116G.
We adopt a stochastic star formation recipe: for each gas particle that is eligible for star formation on a timestep $\Delta t$, it has a probability of $\epsilon_{\rm sf} \Delta t / t_{\rm ff}$ ($t_{\rm ff}$ is the gas free-fall time and $\epsilon_{\rm sf}$ is an efficiency parameter) to be converted into a star particle, with its position, velocity, mass and metallicity unchanged but acts like a collisionless particle. We assume $\epsilon_{\rm sf} = 0.5$ in this work. This high value of efficiency is justified by our high resolution which is able to resolve the structure of dense clouds. Star formation occurs only when (*i*) local velocity divergence becomes negative and (*ii*) the local Jeans mass of the gas particle drops below the total mass within an SPH smoothing kernel (the “kernel mass”). At our fiducial resolution, each star particle represents an actual star whose mass is drawn from a Kroupa stellar initial mass function (IMF) [@2001MNRAS.322..231K] rather than a stellar population. We implement a revised sampling method that will be described in details in Section \[sec:imf\_sample\].
We consider three types of stellar feedback: photoelectric heating, photoionization and SN explosions. Photoelectric heating is self-consistently calculated using the FUV fluxes originated from star particles. The FUV radiation is assumed to be optically thin between the stellar sources and gas particles, albeit with a local shielding treatment using the smoothing length of the gas particles. This is a fair approximation in dwarf galaxies that have low dust-to-gas ratios like ours. Photoionization is based on a Strömgren type approximation with an iterative method that can cope with overlapping HII regions.
SN feedback includes both supernova type II (SNII) and type Ia (SNIa). All massive stars ($\geqq 8 {\rm M_\odot}$) will explode as SNII at the end of their lifetimes. A small fraction of low-mass stars ($< 8 {\rm M_\odot}$) will explode as SNIa following a delay-time distribution (see Section \[sec:dtd\]). We adopt a revised implementation of SN feedback which will be described in Section \[sec:SNfeedback\].
We track eleven individual elements including hydrogen and helium which evolve as passive scalars and contribute to radiative cooling. We include metal enrichment from asymptotic giant branch (AGB) stars, SNII and SNIa based on the metal yields from @2010MNRAS.403.1413K, @2004ApJ...608..405C and @1999ApJS..125..439I, respectively [^2]. We also include metal diffusion following @2010MNRAS.407.1581S and @2013MNRAS.434.3142A to account for the sub-grid mixing of metals.
IMF Sampling {#sec:imf_sample}
------------
The newly formed star particles inherit their progenitor SPH particle mass $m_{\rm gas}$, which is 1${\rm M_\odot}$ in our fiducial simulation. The technical challenge is how to convert a group of star particles with a uniform mass distribution ($m_{\rm gas} = 1 {\rm M_\odot}$) into particles that follow an IMF while keeping the total mass unchanged. @2017MNRAS.471.2151H introduced a method where star particles represent individual stars that collectively follow an assumed IMF, while when resolution is poor ($m_{\rm gas} \gtrsim 100 {\rm M_\odot}$) they represent stellar populations with nonuniformly sampled IMFs. In this work, we revise the method by adding a distance constraint of mass transfer between star particles which we describe as follows.
The strategy of our IMF sampling is similar to @2017MNRAS.471.2151H: we draw an array of stellar masses $m^{\rm IMF}_j$ from a given IMF and assign them to a star particle $a$ with mass $m^*_a$ until $M_{\rm IMF} \equiv \Sigma_j m^{\rm IMF}_j \geqq m^*_a$. In practice, $M_{\rm IMF}$ will almost always exceeds $m^*_a$. In @2017MNRAS.471.2151H, this last sample is still kept while the residual mass $\Delta m \equiv M_{\rm IMF} - m^*_a$ will be “borrowed” from the next newly formed star particle $b$, i.e., $\Delta m $ is subtracted from the mass of the star particle $b$. After the sampling, particle masses will be adjusted such that $m^*_a \gets M_{\rm IMF}$ and $m^*_b \gets (m^*_b - \Delta m)$. While this method ensures that the total mass of all star particles is conserved after the sampling, it implies a mass transfer between star particles that could in principle be spatially well-separated.
In this work, we add a distance constraint such that the mass transfer can only happen within a predefined searching radius $r_{\rm sea}$. We sample a Kroupa IMF with a mass range of \[0.08M$_\odot$, 50M$_\odot$\] until $M_{\rm IMF} \geqq m^*_a$. If $M_{\rm IMF} > m^*_a$, we search for another star particle $b$ that has not yet sampled its stellar masses within $r_{\rm sea}$. If $M_{\rm IMF} \leqq m^*_a + m^*_b$ (meaning that there is enough mass in $b$ that can be transfered), we transfer mass from $b$ to $a$ such that $m^*_a \gets M_{\rm IMF}$ and $m^*_b \gets (m^*_b - \Delta m)$. If $M_{\rm IMF} > m^*_a + m^*_b$, we search for yet another star particle $c$ within $r_{\rm sea}$ for the mass transfer, and if $M_{\rm IMF} < m^*_a + m^*_b + m^*_c$, we do the mass transfer such that $m^*_a \gets M_{\rm IMF}$, $m^*_c \gets (m^*_c + m^*_b - \Delta m)$ and $m^*_b \gets 0$[^3]. We do so iteratively until either (*i*) enough neighbors are found and we can keep the last sample, or (*ii*) there are not enough neighbors to do the mass transfer and so we need to discard this sample. We set $r_{\rm sea} = 4$pc and do the sampling every 2Myr. Each star particle will end up having an array of stellar masses[^4] which provide information of stellar lifetimes, UV luminosities and metal yields.
As discussed in @2017MNRAS.471.2151H, this method has the advantage of being very flexible. When the resolution is poor (e.g., $m^* \sim 10^5 {\rm M_\odot}$), a star particle represents a star cluster following a fairly-sampled IMF and this method reduces to the stellar population model commonly adopted in cosmological simulations. When $m^* \sim 100 - 1000 {\rm M_\odot}$, star particles represent low-mass star clusters whose IMFs vary strongly from particle to particle due to the small samples, which is more physical than assuming the IMF is still well-sampled at this scale. When $m^* \sim 1 {\rm M_\odot}$, star particles represent individual stars which collectively follow the assumed IMF. It also has the advantage that the stellar dynamics will be better resolved as $m^* \approx m_{\rm gas}$, as opposed to the sink-particle approach where typically $m^* \gg m_{\rm gas}$.
SNIa distribution {#sec:dtd}
-----------------
For SNIa, we adopt a power-law delay-time distribution ${\rm DTD}(t) = 0.15 t^{-1.12}$ where $t$ is the delay time in units of Gyr. The normalization is chosen such that $\int_{t_{min}}^{t_{max}} {\rm DTD}(t) dt = 1$. The minimal and maximal delay times are $t_{min}$ = 0.04Gyr and $t_{max}$ = 21Gyr, chosen as the main sequence lifetime of the most massive and least massive SNIa progenitor stars, respectively. For each low-mass star ($< 8 {\rm M_\odot}$) that has become a white dwarf (i.e., when its age is larger than its stellar lifetime), its probability of exploding as an SNIa during the time \[$t, t + \Delta t$\] is $N_{Ia} m^{\rm IMF} f_{\rm WD}(t)^{-1}{\rm DTD}(t) \Delta t$ where $\Delta t$ is the timestep, $N_{Ia} = 1.5\times 10^{-3} {\rm M_\odot^{-1}}$ is the average number of SNIa progenitor stars per total mass of a single stellar population and $f_{\rm WD} (t)$ is the mass fraction of white dwarfs. We construct a table to obtain $f_{\rm WD} (t)$ as a function of $t$ for a Kroupa IMF.
SN feedback {#sec:SNfeedback}
-----------
{width="0.75\linewidth"}
In contrast to @2017MNRAS.471.2151H where the energy of each SN $E_{\rm SN}$ is injected into the ISM all in the form of thermal energy, we implement a revised energy injection scheme which we describe as follows.
### Exact energy partition for resolved SNe
When an SN is resolved (see section \[sec:subgridSN\] for definition), we inject $0.28 E_{\rm SN}$ in terms of kinetic energy while the rest $0.72 E_{\rm SN}$ is injected as thermal energy. This energy partition is based on the exact solution of a SNR that is still in the energy conserving Sedov-Taylor (ST) phase before radiative cooling starts to kick in. As such, the energy partition is guaranteed to be correct right after the energy injection. We take the canonical value $E_{\rm SN} = 10^{51}$ erg for each SN, and inject this energy into $N_{\rm inj} = 96$ gas particles [^5] with a pixelwise scheme (see section \[sec:healpix\]).
The definition of “injecting” kinetic energy can be ambiguous when there are strong relative motions between gas and SNe. For example, gas flows converging rapidly to an SN may even be decelerated by the SN blastwave, meaning that the gas kinetic energy actually decreases. In practice, however, the SN kinetic energy tends to be much higher than the kinetic energy of the gas to be injected in typical ISM conditions. The kinetic energy injected to each particle is $e_k = 0.28 E_{\rm SN} N_{\rm inj}^{-1} $ and the corresponding velocity kick is $\Delta v = \sqrt{2e_k / m_{\rm gas}} \approx 541$ km s$^{-1} (m_{\rm gas} / {\rm M_\odot})^{-0.5}$ along the outward radial direction. This velocity kick is much higher than the typical velocity dispersion in the ISM ($\sim 10 {\rm km s^{-1}}$) for the resolution we explore in this paper, making it possible to have a well-defined kinetic energy injection [^6].
### HealPix-based pixelwise injection {#sec:healpix}
In @2017MNRAS.471.2151H, the SN energy is injected into the nearest $N_{\rm inj}$ gas particles. While this works fine for thermal energy injection, it has potential issues for kinetic energy/momentum injection when the nearest gas particles are highly clumpy as it implies an anisotropic injection, i.e., the kinetic energy will be preferentially injected in high-density regions where most particles are located. To address this issue, we implement a pixel-by-pixel injection scheme using the [HealPix]{} tessellation library [@2011ascl.soft07018G] and divide the $4\pi$ solid angle into $N_{\rm pix} = 12$ pixels. Within each pixel we look for the nearest $N_{\rm inj} / N_{\rm pix} = 8$ gas particles for the energy injection. By doing so, we guarantee that the kinetic energy injection is isotropic at the pixel level. Although there could still be an anisotropic particle distribution within each pixel, this should be sufficiently small as particles are not expected to be highly clumpy at scales $N_{\rm pix}$ times smaller than the SPH kernel mass.
### Sub-grid model for unresolved SNe {#sec:subgridSN}
Both the pure thermal feedback and the one based on the exact ST energy partition become problematic when the ST phase is unresolved. This is defined as the injection radius (typical a few resolution elements) becomes larger than the cooling radius $R_c$ where radiative cooling kicks in and the system enters a pressure-driven “snowplow” phase followed by a momentum-conserving phase with a terminal momentum $p_{\rm term}$. In Lagrangian codes like ours, the equivalent definition would be when the injection mass $M_{\rm inj} = N_{\rm inj} m_{\rm gas}$ becomes larger than the cooling mass $M_c = (4\pi/3) \rho_0 R_c^3$ where $\rho_0$ is the ambient density. We adopt $R_c = 22.6$ pc $E_{51}^{0.29} n_{\rm H}^{-0.42}$ from @2015ApJ...802...99K [^7] where $E_{51}$ is the SN energy in units of $10^{51}$ erg and $n_{\rm H}$ is the hydrogen number density of the ambient medium in units of cm$^{-3}$. This leads to the following requirement of resolving the ST phase: $$\label{eq:coolmass}
m_{\rm gas} < 1680 {\rm M_\odot} E_{51}^{0.87} n_{\rm H}^{-0.26} N_{\rm inj}^{-1} .$$ When Eq. \[eq:coolmass\] is not satisfied, physically, a significant amount of energy has already been radiated away right after the injection, and therefore injecting $10^{51}$ erg into $M_{\rm inj}$ leads to an incorrect evolution of the SNR. Injecting thermal energy results in the well-known “over-cooling” problem which underestimates the terminal momentum as most energy would radiate away almost immediately without having much dynamical impact on the ISM. On the other hand, ST injection (or pure kinetic injection) overestimates the terminal momentum [@2017ApJ...846..133K; @2018MNRAS.477.1578H] as it wrongly assumes that the SNR still has 0.28 $E_{51}$ of kinetic energy at that scale while physically the kinetic energy has already decayed due to momentum conservation.
We therefore adopt a sub-grid model which has been widely adopted in the literature (see e.g. [@2013ApJ...776....1K; @2014MNRAS.445..581H; @2015MNRAS.449.1057G; @2015MNRAS.450..504M; @2015ApJ...809...69S; @2015MNRAS.451.2900K; @2015MNRAS.454..238W; @2017MNRAS.466...11R; @2017arXiv170206148H; @2018ApJ...853..173K]) when the SNe are unresolved. Instead of injecting energy, we inject the terminal momentum $$\label{eq:pterm}
p_{\rm term} = 3.3 \times 10^5 {\rm M_\odot ~ km ~ s^{-1} } E_{51}^{0.93} n_{\rm H}^{-0.13},$$ where the scalings are again taken from @2015ApJ...802...99K while the normalization is recalibrated by our numerical experiments (cf. Fig. \[fig:snrconv\]). This momentum is injected into $N_{\rm inj}$ gas particles using the same pixel-by-pixel scheme as described above. Each particle will receive a momentum of $N_{\rm inj}^{-1} p_{\rm term}$ and the corresponding velocity kick is $\Delta v = 3.44 \times 10^3$ km s$^{-1} n_{\rm H}^{-0.13} (m_{\rm gas} / {\rm M_\odot})^{-1}$. Following the scaling relation reported in @1998ApJ...500...95T, We also include the residual thermal energy of the SNR: $$\label{eq:resEth}
E^{\rm res}_{\rm th} = 0.72 E_{\rm SN} \Big(\frac{M_{\rm inj} }{M_c}\Big)^{-2.17}.$$ As shown in Appendix \[app:resEth\], this residual thermal energy has little effect on the wind properties and is included mainly for consistency.
When the injection region is highly inhomogeneous, a single SN may be resolved along certain pixels while unresolved along the other pixels. We therefore calculate Eq. \[eq:coolmass\], \[eq:pterm\] and \[eq:resEth\] pixel by pixel using the hydrogen number density of each pixel.
Our default SN scheme can be summarized as follows:
1. When an SN is resolved (Eq. \[eq:coolmass\] is true), inject $0.28 E_{\rm SN}$ in terms of kinetic energy and $0.72 E_{\rm SN}$ in terms of thermal energy into the neighboring particles found by the [HealPix]{}-based algorithm.
2. When an SN is unresolved (Eq. \[eq:coolmass\] is not true), inject the terminal momentum $p_{\rm term}$ and the residual thermal energy $E^{\rm res}_{\rm th}$ given by Eq. \[eq:pterm\] and \[eq:resEth\], respectively.
### Numerical test
Fig. \[fig:snrconv\] shows the time evolution of the kinetic energy (top panels) and momentum (bottom panels) of an SNR with $E_{\rm SN} = 10^{51}$ erg in a uniform medium of $n_{\rm H} = 1$ cm$^{-3}$ at four different resolutions ($m_{\rm gas}$ = 0.1, 1, 10 and 100${\rm M_\odot}$, and the corresponding injection mass $M_{\rm inj} = 10, 100, 10^3$ and $10^4 {\rm M_\odot}$, respectively). The SNR with pure thermal injection is shown in the left panels while the SNR with our default injection scheme is shown in the right panels. The dashed horizontal lines indicate the expected kinetic energy in the ST phase ($0.28 E_{\rm SN}$).
In the case of thermal injection (left panels), the time it takes to convert the thermal energy into kinetic energy to the expected amount of $0.28 E_{\rm SN}$ in the ST phase lengthens as the resolution becomes worse, which reflect the capability of the system to converge to the exact solution and generate the right amount of momentum in the ST phase. Once the radiative cooling kicks in, the momentum of the SNR rapidly plateaus and eventually drops when the SN bubble falls back and closes. According to Eq. \[eq:coolmass\], the SN is resolved when $m_{\rm gas} \lesssim 20 {\rm M_\odot}$. As expected, the terminal momentum is converged ($p_{\rm term} \approx 3.3 \times 10^5 {\rm M_\odot ~ km ~ s^{-1} }$) in cases of $m_{\rm gas}$ = 0.1, 1 and 10 ${\rm M_\odot}$ while it becomes significantly underestimated in the $m_{\rm gas} = 100 {\rm M_\odot}$ case, which is a demonstration of the well-known “over-cooling” problem.
With the default injection scheme (right panels), in the resolved cases ($m_{\rm gas}$ = 0.1, 1 and 10${\rm M_\odot}$), the SNR acquires $0.28 E_{\rm SN}$ of kinetic energy by construction right from the beginning of the injection, and it remains so for the entire ST phase. The terminal momentum is converged and agrees well with that in the thermal injection case, which is consistent with @2012MNRAS.419..465D and @2015ApJ...809...69S. In the unresolved case ($m_{\rm gas} = 100 {\rm M_\odot}$), since we put in the expected terminal momentum (Eq. \[eq:pterm\]) by hand and this momentum has to be conserved afterwards, we obtain the same terminal momentum as the resolved ones, which is around a factor of 3 higher than that in the $m_{\rm gas} = 100 {\rm M_\odot}$ case with pure thermal injection.
We emphasize that putting in the expected $p_{\rm term}$ by hand for the unresolved SNe is ultimately still a sub-grid model. There are other aspects that this model cannot capture (for example, the low-density bubble and the dense shell), which can play an important role in simulations. This motivates us to test whether this sub-grid model provides better convergence compared to the pure thermal injection when applied to more realistic simulations (see section \[sec:SNinjectResult\]).
Initial conditions
------------------
The initial conditions consist of an exponential gaseous disk subject to an external gravitational potential of a dark matter halo. The halo follows the Navarro-Frenk-White profile [@1997ApJ...490..493N] with virial mass $M_{\rm vir} = 10^{10} {\rm M_\odot}$, virial radius $R_{\rm vir} = 44.4$kpc and concentration $c = 17$. The choice of $c$ is motivated by the scaling relation from cosmological N-body simulations in @2014MNRAS.441.3359D. The total mass of the disk is $M_{\rm gas} = 10^7 {\rm M_\odot}$. The initial temperature of gas is $T = 10^4$ K, while the density profile of the disk (in a cylindrical coordinate) is $$\rho_{\rm gas} (R,z) = \frac{M_{\rm gas}}{4\pi R^2_0 z_0} \exp(-\frac{R}{R_0}) \exp(-\frac{z}{z_0})$$ where $R_0$ and $z_0$ are the scale-length and scale-height of the disk, respectively. We set $r_0$ = 0.4 kpc such that the gas surface density at $r$ = 0 is 10 ${\rm M_\odot pc^{-2}}$. We set $z_0$ = 0.5 kpc, which is a somewhat arbitrary choice as the scale-height of the disk changes rapidly once the simulation begins (since the gas will cool and collapse onto the midplane). The initial gas metallicity is $Z = 0.1 Z_{\rm \odot}$ and the dust-to-gas ratio is 0.1% in mass. The rotation velocity of the disk follows the circular velocity of the halo including the correction of pressure gradient. The SPH particle mass is $m_{\rm gas} = 1 {\rm M_\odot}$. The gravitational softening length for gas is set to be the same as the SPH smoothing length. Once a gas particle is converted into a star particle, we adopt a fixed softening length of 0.3 pc [^8].
Results {#sec:results}
=======
Definitions {#sec:def}
-----------
Before discussing our results, we first need to give clear definitions of wind quantities and how we measure them in the simulations. We quantify the winds by the fluxes of mass ($F_{\rm m}$), momentum ($F_{\rm p}$) and energy ($F_{\rm e}$) as in the fluid equations, viz. $$F_{\rm m} = \rho {\bf v} ,~~
F_{\rm p} = \rho {\bf v}{\bf v} + P ,~~
F_{\rm e} = (\rho e_{\rm tot} + P) {\bf v},$$ where $\rho$, ${\bf v}$ and $P$ are the density, velocity and pressure of gas, respectively. $e_{\rm tot} = 0.5 ||{\bf v}||^2 + u$ is the total energy (kinetic + thermal) per mass and $u$ is the specific thermal energy. Integrating the fluxes over a measuring surface $S$, we obtain the flow rate of mass ($\dot{M}$), momentum ($\dot{p}$)[^9] and energy ($\dot{E}$) as $$\dot{M} = \int_{S} F_{\rm m} \cdot \hat{n} dA, ~~\dot{p} = \int_{S} F_{\rm p} \cdot \hat{n} dA, ~~\dot{E} = \int_{S} F_{\rm e} \cdot \hat{n} dA$$ where $\hat{n}$ is the outward (away from the galaxy) unit normal vector of the area $dA$. Far away from the disk, we define the measuring surface as a spherical shell of radius $r_s$ less the region where $|z| / r_s < 1 / \sqrt{2}$, which encloses the majority of the outflowing gas and excludes the flaring disk component at large galactocentric radii. On the other hand, as the winds are mostly vertical rather than radial in the vicinity of the disk, a spherical shell is no longer an appropriate measuring surface. Therefore, when $r_s \leqq 1.5$kpc, we measure the wind properties with two planes parallel to the disk’s midplane at $z = \pm r_s$[^10].
The flow rate of mass and energy can be further decomposed into the outflow (where the radial velocity $v_r \equiv {\bf v}\cdot \hat{n} > 0$) and inflow ($v_r < 0$) components such that $$\dot{M} = \dot{M}_{out} - \dot{M}_{in}, ~~ \dot{E} = \dot{E}_{out} - \dot{E}_{in}.$$ Note that we define the inflow rates to be positive by absorbing the minus sign in ${\bf v}\cdot \hat{n}$. The momentum flow rate cannot be decomposed into outflow and inflow as it is always positive.
In SPH simulations, the discretized flow rates can be expressed as $$\begin{aligned}
\dot{M}_{out} &=& \sum_{i, v_{r,i}>0} \dfrac{m_i v_{r,i}}{dr}, \\
\dot{M}_{in} &=& - \sum_{i, v_{r,i}<0} \dfrac{m_i v_{r,i}}{dr}, \\
\dot{p} &=& \sum_{i} \dfrac{m_i (v^2_{r,i} + (\gamma - 1) u_i)}{dr} \label{eq:eta_p}, \\
\dot{E}_{out} &=& \sum_{i, v_{r,i}>0} \dfrac{m_i (v^2_i + \gamma u_i) v_{r,i}}{dr} \label{eq:eta_eout}, \\
\dot{E}_{in} &=& - \sum_{i, v_{r,i}<0} \dfrac{m_i (v^2_i + \gamma u_i) v_{r,i}}{dr} \label{eq:eta_ein}, \end{aligned}$$ using the relation $P = (\gamma - 1) \rho u$. The summations are over particles that lie within the interval $r_s \pm 0.5 dr$. We adopt $dr = 0.1 r_s$.
It is useful to express the flow rates normalized to the star formation activities that drive the winds. We define the following quantities:
- outflow mass loading factor $\eta_m^{out} \equiv \dot{M}_{out} / {\rm \overline{SFR}}$
- inflow mass loading factor $\eta_m^{in} \equiv \dot{M}_{in} / {\rm \overline{SFR}}$
- momentum loading factor $\eta_p \equiv \dot{p} / (p_{\rm ej} R_{\rm SN}) $
- outflow energy loading factor $\eta_e^{out} \equiv \dot{E}_{out} / (E_{\rm SN} R_{\rm SN})$
- inflow energy loading factor $\eta_e^{in} \equiv \dot{E}_{in} / (E_{\rm SN} R_{\rm SN})$
where $R_{\rm SN} = {\rm \overline{SFR}} / (100{\rm M_\odot}) $ is the SN rate for a Kroupa IMF, $p_{\rm ej} = 3\times 10^4 {\rm M_\odot~km~s^{-1}}$ is the momentum carried by the ejecta of a typical SN, and ${\rm \overline{SFR}}$ is the time-averaged star formation rate of the galaxy. We take the time-averaged SFR over the entire simulation time excluding the initial transient phase $t \in $ \[0.4 Gyr, 1.6 Gyr\][^11] because our system reaches a quasi-steady state where the instantaneous SFR fluctuates a lot but the mean SFR remains roughly constant.
{width="0.95\linewidth"}
Likewise, the discretized mass flow rates of metals can be defined as $$\begin{aligned}
\dot{M}_{Z, out} &=& \sum_{i, v_{r,i}>0} \dfrac{Z_i m_i v_{r,i}}{dr}, \\
\dot{M}_{Z, in} &=& - \sum_{i, v_{r,i}<0} \dfrac{Z_i m_i v_{r,i}}{dr},\end{aligned}$$ where $Z_i$ is the metallicity of particle $i$ (mass fraction of all heavy elements). Correspondingly, we define the following quantities:
- outflow metal loading factor $\eta_Z^{out} \equiv \dot{M}_{Z, out} / (m_Z R_{\rm SN})$
- inflow metal loading factor $\eta_Z^{in} \equiv \dot{M}_{Z, in} / (m_Z R_{\rm SN})$
where $m_Z = 2.5 {\rm M_\odot}$ is the IMF-averaged metal mass ejected by an SN.
The metal loading factor only measures how much metals is passing through the measuring surface normalized to how much metals ejected by SNe. A high $\dot{M}_{Z, out}$ means that a large fraction of the newly produced metals is being blown out, but it has no information about whether metals are “preferentially” blown out. To quantify the metal enrichment of the outflowing gas, we need to define the following quantities:
- outflow enrichment factor $y_Z^{out} \equiv \dot{M}_{Z, out} / (\dot{M}_{out} Z_{\rm ISM})$
- inflow enrichment factor $y_Z^{in} \equiv \dot{M}_{Z, in} / (\dot{M}_{in} Z_{\rm ISM})$
where $Z_{\rm ISM}$ is the time-averaged metallicity of the disk, defined as the region where $R < 1$ kpc and $|z| < 0.5$ kpc. Note that the relation between $y_Z^{out}$ and $\eta_Z^{out}$ can be expressed as $y_Z^{out}= 0.025 ~\eta_Z^{out} / (\eta_m^{out} Z_{\rm ISM})$.
Morphology
----------
In Fig. \[fig:ismfaceedgestars0399\], we show the maps of column density (top panels) and temperature (slices through the origin, bottom panels) of the simulated galaxy at simulation time $t$ = 454 Myr. The left and middle panels show the face-on and edge-on views, respectively, with massive stars ($> 8 {\rm M_\odot}$) over-plotted as white circles. The right panels show the zoom-out edge-on views. The fluid velocity is over-plotted as arrows (not shown in all panels).
The ISM shows a multi-phase structure which is shaped by gravity, thermal instability and SN feedback. SNe collectively create superbubbles with sizes of a few hundred pc, within which the gas is hot and diffuse. Several massive stars are located inside the bubbles. These massive stars, when later on explode as SNe, will have a stronger dynamical impact on the ISM due to the low-density environments they reside where radiative energy losses are inefficient. Energy input from SN feedback keeps the gaseous disk thick. When the hot bubbles break out of the disk (i.e., when they expand well above the scale-height of the disk), the over-pressurized hot gas ($\sim 10^6$K) vents out with velocities much higher than the warm gas ($\sim 10^4$K), and this hot gas preferentially escapes the disk through low-density channels. In contrast, before the bubbles break out, the velocities of the bubbles are much lower. In steady state, the winds form a gaseous halo and exhibit a multi-phase structure. In the zoom-out edge-on view, temperature discontinuities can be seen where the fast moving hot winds are pushing out the warm halo gas.
Time evolution
--------------
The top panel of Fig. \[fig:ofvstime\] shows the time evolution of the total star formation rate (SFR) of the simulated galaxy, measured by the total stellar mass that have formed within the last 20 Myr divided by 20 Myr. The middle panel shows the time evolution of $\dot{M}_{out}$, $\dot{M}_{in}$ and the (mass-weighted) mean temperature of gas $T$ within $r_s \pm 0.5 dr$ where $r_s$ = 1kpc. The bottom panel is the same as the middle panel but measured at $r_s$ = 10kpc.
The SFR experiences an initial transient phase of starburst and then settles to a quasi-steady state after $t =$ 0.2 Gyr with a mean SFR $\approx 10^{-4} {\rm M_\odot yr^{-1}}$ and strong temporal fluctuations (more than an order of magnitude). Although there is no cosmological inflows to replenish the gas reservoir, the SFR can still be sustained at the same level throughout the simulation as the galactic depletion time is much longer than the simulation time. At $r_s$ = 1kpc, $\dot{M}_{out}$ exhibits strong fluctuations and it anticorrelates with $\dot{M}_{in}$ but with $\dot{M}_{out} > \dot{M}_{in}$ for most of the time, meaning that there is a net mass outflow. The mean temperature is only around $2\times 10^4$ K, not much higher than the typical warm phase of the ISM. At $r_s$ = 10kpc, all the gas is flowing outward ($\dot{M}_{in} \sim 0$), and the typical timescale of variation becomes longer. The mean temperature is higher than that at $r_s$ = 1kpc, and there is a strong temporal correlation with $\dot{M}_{out}$: whenever there is a rapid rise in temperature, the mass outflow rate also increases rapidly. Note that the dynamical timescale becomes longer at large $r_s$. Therefore, it is necessary to run the simulations for long enough time ($\gtrsim$ 1Gyr) to ensure that the entire system within the virial radius has settled to a quasi-steady state.
Time-averaged wind properties
-----------------------------
In Fig. \[fig:mpeloadingz\], we show the time-averaged mass loading factor $\eta_m$ (upper left), metal enrichment factor $y_Z$ (upper right), momentum loading factor $\eta_p$ (lower left) and energy loading factor $\eta_e$ (lower right) as a function of measuring radius $r_s$. $\eta_p$ and $\eta_e$ are further decomposed into kinetic (subscript “$k$”) and thermal (subscript “$th$”) components which correspond to the first and the second terms in Eq. \[eq:eta\_p\], \[eq:eta\_eout\] and \[eq:eta\_ein\], respectively. The lines represent the time-averaged values over the interval $t \in $ \[0.4 Gyr, 1.6 Gyr\] while the shaded regions show the temporal fluctuations between 16 percentile and 84 percentile.
At $r_s < 2$kpc, inflow roughly balances outflow ($\eta_m^{in} \approx \eta_m^{out}$), indicating that the majority of the outflowing gas in this regime will eventually fall back onto the disk. This regime is therefore dominated by the so-called “fountain flows”. The momentum flux at small $r_s$ is dominated by the thermal component $\eta_{p,th}$, which indicates that the outflowing gas is mainly accelerated by the thermal pressure when it is within or close to the disk. Similar trend can be seen in the energy fluxes, though the difference between the thermal and kinetic components at small $r_s$ is smaller than that of the momentum flux, mainly because the kinetic energy includes contribution from the rotation of the disk. Unlike the mass flux where $\eta_m^{out} \approx \eta_m^{in}$, the energy flux has a clear net outflow with both $\eta_{e,k}^{out} > \eta_{e,k}^{in}$ and $\eta_{e,th}^{out} > \eta_{e,th}^{in}$. As will be shown later, this is because most of the energy is carried by the hot gas, which has a net outflow. This is contrary to the mass flux which has almost no net outflow and is dominated by the warm gas. There is almost no metal enrichment ($y_Z^{out} \approx 1$) in this regime where the fountain flows dominate, as a significant amount of the unenriched gas is entrained by the pressure-driven hot gas, diluting the enriched gas.
As $r_s$ increases, $\eta_m^{in}$ drops rapidly towards zero and the actual winds start to emerge. Once a gas parcel manages to reach this regime ($r_s > 2$kpc), it will most likely escape the halo eventually, as $\eta_m^{out}$ flattens out and remains almost constant all the way to $R_{\rm vir}$ where $\eta_m^{out} \approx$ 3. The thermal component of the momentum loading factor$\eta_{p,th}$ drops rapidly and at large $r_s$ while the kinetic component $\eta_{p,k}$ overtakes and remains almost constant up to $R_{\rm vir}$ with $\eta_{p,k}^{out} \approx$ 1. This suggests that ram pressure (rather than thermal pressure) becomes the main source of acceleration in this regime. Similar trend can be seen in the energy fluxes, where $\eta_{e,th}^{out}$ drops rapidly due to the expansion of gas while $\eta_{e,k}^{out}$ remains almost constant up to $R_{\rm vir}$ with $\eta_{e,k}^{out} \approx 0.05$. The majority of energy injected by the SNe is lost and only 5% is carried with the escaping winds mainly in the form of kinetic energy. As for metals, $y_Z^{out}$ increases only slightly up to 1.5 at $R_{\rm vir}$, i.e., the winds are mildly metal enriched (by around 50%) relative to the ISM.
The fact that $\eta_m^{out}$, $\eta_p^{out}$ and $\eta_e^{out}$ all flatten out at large $r_s$ indicates that the winds are indeed in a quasi-steady state in a time-averaged sense: for a given spherical shell, the mass, momentum and energy fluxes that flow in balance those that flow out, which forms a steady (but not static) gaseous halo.
### Multi-phase winds
In Fig. \[fig:mpeloadingzphases\], we show the same quantities as Fig. \[fig:mpeloadingz\] but separate the gas into three different temperature phases: hot gas ($T > 3\times 10^5$K, in red), ionized gas ($3\times 10^4 {\rm K} < T < 3\times 10^5$K, in green) and warm gas ($T < 3\times 10^4$K, in blue)[^12]. At $r_s < 2$kpc, the warm gas dominates both inflow and outflow which corresponds to the fountain flows. The hot gas occupies only a small mass fraction of the flows and has a net outflow of $\eta_m^{out} \sim 1$ at $r_s = 2$kpc. The ionized gas shows similar characteristics to the hot gas in this regime. These two gas populations provide the necessary thermal pressure to launch the winds. While most of the mass and momentum is carried by the warm gas, most of the energy is carried by the hot gas.
Note that the three different phases do not evolve independently. As $r_s$ increases, the fraction of hot gas in the winds gradually decreases while the fraction of ionized gas increases. This is not because the hot gas is slowing down as it moves outwards. Instead, it is because the temperature of the hot gas decreases drops below $3\times 10^5$K. Cooling is mainly due to the expansion of gas rather than radiative cooling, as the latter is very inefficient at large $r_s$ where the density is low. Similar phase transition occurs between the ionized gas and warm gas at larger $r_s$. On the other hand, some of the warm gas is accelerated by the ram pressure of the fast-moving hot gas. When it gets shock-heated, it evolves from the warm phase to the ionized phase. As a result, in the “halo” regime ($r_s > 2$kpc), the ionized gas turns out to be the dominant phase.
At small $r_s$, the warm gas is unenriched while the hot and ionized gas are metal enriched by a factor of around 3.6 and 1.6, respectively. Though significantly metal enriched, as the hot gas only occupies a small mass fraction of the flows, the overall fountain flows are close to unenriched due to the dilution by the warm gas (cf. Fig. \[fig:mpeloadingz\]). As $r_s$ increases, the metallicity of different phases gradually mix with each other. When the winds emerge ($r_s \gtrsim 2$kpc), the hot gas has around three times higher metallicity than the warm phase. As the winds reach $R_{\rm vir}$, all three phases are well-mixed with $y_Z \approx 1.5$.
### Radial profile of winds
As the system reaches a steady state, the winds form a steady but not static gaseous halo. In Fig. \[fig:windprofiletimeave\], we show the radial velocity $v_{\rm r}$ and the sound speed $c_s$ (upper panel) and the hydrogen number density $n_{\rm H}$ (lower panel) of gas as a function of $r_s$. All quantities are (mass-weighted) spatially- and temporally-averaged. The shaded areas show the temporal fluctuations between 16 percentile and 84 percentile. The escape velocity, defined as $$v_{\rm esc}(r) = \sqrt{2 (\Phi(R_{\rm vir}) - \Phi(r))}$$ where $\Phi(r)$ is the gravitational potential of the NFW halo, is overplotted in the upper panel.
The mean radial velocity increases monotonically with $r_s$ from 3 km s$^{-1}$ at $r_s$ = 0.1kpc to 100 km s$^{-1}$ at $R_{\rm vir}$. Winds are gradually accelerated as $r_s$ increases, and only at $r_s > 7$kpc do they acquire enough kinetic energy to escape the halo ($v_{\rm r} > v_{\rm esc}$). At $r_s < 2$kpc, the low level of $v_{\rm r}$ is mainly due to the fountain flows which move much slower than the actual winds. At $r_s > 2$kpc where the winds emerge, the fact that $v_{\rm r}$ increases with $r_s$ may seem counterintuitive. As will be shown shortly, the acceleration is due to the hot gas which acquires very high velocities ($> 100$ km s$^{-1}$) already at small $r_s$. On the other hand, $c_s$ shows a much flatter profile compared to the monotonically increasing $v_{\rm r}$, which leads to a transition of winds from subsonic ($r_s < 2$kpc) to supersonic ($r_s > 2$kpc) regime. $n_{\rm H}$ decreases monotonically and scales roughly as $r^{-3}$ due to the expansion of gas, Comparing the profiles of both $c_s$ and $n_{\rm H}$, it is clear that the winds are not expanding adiabatically. A significant amount of energy is added to the winds as they travel towards $R_{\rm vir}$.
Interaction between warm gas and hot gas
----------------------------------------
In this section, we investigate how winds actually escape the dark matter halo by backtracking the history of the escaping gas particles, taking advantages of the Lagrangian nature of our code. In Fig. \[fig:historywarm\], we show the history of 500 randomly selected gas particles that lie within the interval $r_s \in$ \[$0.95 R_{\rm vir}$, $1.05 R_{\rm vir}$\] at $t = 1.6$ Gyr. Four different particle quantities are shown as a function of $r_s$: $v_{\rm r}$ (top left), $T$ (top right), $A$ (bottom left) and $B$ (bottom right), Each thin gray trajectory represents the history of each gas particle back-traced from $t = 1.6$Gyr. $A$ is the entropy of gas defined as $A \equiv P / \rho^{\gamma}$, and $B$ is the Bernoulli parameter defined as[^13] $$B = 0.5 v^2 + \gamma u + \Phi(r) - \Phi(0).$$ The escape velocity is overplotted (dotted line) in the top left panel. Five particles which have their maximum temperatures lower than $3\times 10^4$K within the interval $r_s \in$ \[1,2\] kpc (the “launching radius”) are highlighted in color with thicker trajectories. These highlighted particles are launched as warm gas, and their velocities are typically lower than the escape velocity when they are launched and then slowed down by the gravitational pull of the halo. Without further acceleration, they will not be able to escape the halo. Because these particles are already supersonic when launched, there is not much thermal energy is available for the acceleration from adiabatic expansion. However, many of them experience multiple episodes of abrupt acceleration subsequently, which eventually give them enough kinetic energy to escape the halo. The acceleration episodes typically correspond to a sudden rise of both temperature and entropy, which is indicative of shocks. This explains why the mean sound speed in Fig. \[fig:windprofiletimeave\] does not drop (or even increase) as what may be expected from adiabatic expansion. In between the shock episodes, gas is decelerated by gravity and cooled nearly adiabatically (with $A$ being constant) due to expansion. The Bernoulli parameter, instead of staying constant, increases with $r_s$ suggesting that these particles keep acquiring energy as they travel towards $R_{\rm vir}$. Interestingly, one of the highlighted particle (in purple) reaches $10^6$K at $r_s = 0.2$kpc but then rapidly cooled back to the warm phase. This is an example of an SN bubble that failed to break out of the disk before radiating away its thermal energy.
We now turn to the gas particles that were hot when launched. Fig. \[fig:historyhot\] shows the same quantities as Fig. \[fig:historywarm\] but highlights five particles which have their minimum temperatures higher than $3\times 10^5$K within the interval r $\in$ \[1,2\] kpc. These particles correspond to the hot gas vented out from the superbubbles as they break out of the disk. Their velocities are already higher than $v_{\rm esc}$ right after they were launched, and remain constant or slightly decrease afterwards. They expand and cool almost adiabatically with their entropies remain constant. Their Bernoulli parameters are mainly dominated by the kinetic component. These particles could easily escape the halo without any further interaction. Nevertheless, they sweep up the preexisting warm gas particles in the halo (those that do not have enough energy to escape on their own), feeding energy and momentum to the latter, and both manage to escape eventually. The winds are therefore composed of two populations of gas with different launching temperatures. Note that as the winds travel to large $r_s$, because the hot gas cools adiabatically and the warm gets shock-heated, the temperature difference between these two phases decreases and eventually the winds become single-phase.
Convergence and SN injection schemes {#sec:conv}
------------------------------------
In this section, we investigate the convergence properties by running the simulations at four different resolutions where $m_{\rm gas} / {\rm M_\odot}$ = 1, 5, 25 and 125 (with the corresponding injection mass $M_{\rm inj} / {\rm M_\odot}$ = 100, 500, 2500 and $1.25\times 10^4 {\rm M_\odot}$). The gravitational softening of stars is set to be equal to the SPH smoothing length at the typical densities where star formation occurs (recall that our star formation threshold is resolution dependent), which turn out to be 0.2, 0.8, 4 and 10 pc, from the highest resolution run to the lowest, respectively.
In Fig. \[fig:sfrconv\], we show the time evolution of the total SFR at four different resolutions. The time-averaged SFR in units of ${\rm M_\odot yr^{-1}}$ are, from the highest resolution run to the lowest, $1.3\times 10^{-4}$, $1.5\times 10^{-4}$, $1.5\times 10^{-4}$ and $3.6\times 10^{-4}$, respectively. The SFR converges at $m_{\rm gas} = 25 {\rm M_\odot}$, while in the 125${\rm M_\odot}$-run the time-averaged SFR is higher by a factor of 2.5 and is also more bursty than the other three runs.
![Time evolution of the total SFR at four different resolutions. The SFR converges at $m_{\rm gas} = 25 {\rm M_\odot}$ while it becomes slightly higher and more bursty at $m_{\rm gas} = 125 {\rm M_\odot}$.[]{data-label="fig:sfrconv"}](SFR_conv.pdf){width="1\linewidth"}
In Fig. \[fig:mloadconv\], we show $\eta^{out}_m$ (top panel) and $\eta^{out}_e$ (bottom panel) as a function of $r_s$ at four different resolutions. The vertical error bars indicate the temporal fluctuations between 16 and 84 percentiles. Both $\eta^{out}_m$ and $\eta^{out}_e$ converge at $m_{\rm gas} = 5 {\rm M_\odot}$ in the entire range of $r_s \in [0.1$kpc, $R_{\rm vir}]$. The $25 {\rm M_\odot}$-run is marginally converged, where $\eta^{out}_m$ and $\eta^{out}_e$ are lower than the converged values at $R_{\rm vir}$ by a factor of 2 and 5, respectively. In the $125 {\rm M_\odot}$-run, however, both $\eta^{out}_m$ and $\eta^{out}_e$ vanish rapidly at $r_s > 2$kpc, indicating the failure of launching winds despite its slightly higher and more bursty SFR. Winds are more sensitive to resolutions and are more difficult to converge compared to the SFR.
![$\eta^{out}_m$ (top panel) and $\eta^{out}_e$ (bottom panel) as a function of $r_s$ at four different resolutions. The vertical error bars indicate the temporal fluctuations between 16 and 84 percentiles. Wind properties converge at $m_{\rm gas} = 5 {\rm M_\odot}$ ($M_{\rm inj} = 500 {\rm M_\odot}$) where individual SNe are properly resolved. Our 125M$_\odot$-run fails to launch winds.[]{data-label="fig:mloadconv"}](Mload_conv_resEth){width="0.9\linewidth"}
In Fig. \[fig:pd2by2converge\], we show the time-averaged phase diagram ($n_{\rm H}$ vs. $T$) in runs with $m_{\rm gas}$ = 1M$_\odot$ (upper left), 5M$_\odot$ (lower left), 25M$_\odot$ (upper middle) and 125M$_\odot$ (lower middle), respectively. The solid blue line on the lower-left corner of each phase diagram indicates the (resolution-dependent) star formation threshold. The dashed black line indicates a constant pressure of $10^3$Kcm$^{-3}$. The bottom panels show the time-averaged probability density function (PDF) of $T$ (upper right) and $n_{\rm H}$ (lower right) at four different resolutions.
In our highest-resolution run ($m_{\rm gas} = 1{\rm M_\odot}$), the ISM has a multi-phase structure and different phases are roughly in pressure equilibrium with $P \sim 10^3 {\rm K cm^{-3}}$. Heating from SN feedback keeps the diffuse gas ($n_{\rm H} \sim 0.1-1$cm$^{-3}$) at $T = 10^4$K, consistent with previous studies [@2016MNRAS.458.3528H; @2017MNRAS.471.2151H]. The thin horizontal line at $T = 10^4$K is due to photoionization. At $n_{\rm H} > 100 {\rm cm^{-3}}$, cold gas enters the isothermal regime: it decouples with the ISM, reaches a much higher pressure, and undergoes gravitationally collapse. Star formation occurs in the densest gas where $n_{\rm H} \sim 10^3 - 10^4$ cm$^{-3}$. This is possible thanks to the high resolution we adopt which allows us to properly follow the gravitationally collapse up such high densities, and it also provides a physical justification to our choice of a high star formation efficiency as we resolve the Jeans mass up to the high densities where star formation is expected to be very efficient. Due to the resolution-dependent star formation threshold, the highest density the gas can reach decreases as $m_{\rm gas}$ increases. In the 125M$_\odot$-run, the densest gas only extend to $n_{\rm H} \sim 100$ cm$^{-3}$. While this may lead to less clustered star formation, it is partly compensated by its highly bursty star formation history (cf. Fig. \[fig:sfrconv\]).
The density and temperature PDFs are well-converged at $m_{\rm gas} = 5 {\rm M_\odot}$, and start to show deviations in the 25M$_\odot$-run, especially at the high-temperature part. The most striking feature in the 125M$_\odot$-run (which fails to launch winds), compared to the other higher-resolution runs, is its complete lack of hot phase. Since winds are driven by the thermal pressure which is mostly carried by the hot gas, the failure of wind launching is a direct consequence of having no hot gas.
{width="0.99\linewidth"}
### SN environment {#sec:SNenv}
In Fig. \[fig:snrhot\], we show the cumulative distribution functions of hydrogen number density ($n_{\rm SN}$, top panel) and temperature ($T_{\rm SN}$, bottom panel) of gas where the SNe occur at four different resolutions. The vertical dashed lines in the top panel indicate the (resolution-dependent) critical density for which the SNe are resolved (cf. Eq. \[eq:coolmass\]). All SNe are properly resolved in the 1M$_\odot$- and 5M$_\odot$-runs. Around 20% SNe in the 25M$_\odot$-run are unresolved, and almost all SNe in the 125M$_\odot$-run are unresolved.
In the highest resolution run ($m_{\rm gas} = 1 {\rm M_\odot}$), around 60% SNe occur in places hotter than the typical warm ISM ($10^4$K). This indicates that these SNe occur in preexisting hot bubbles generated by previous SNe. The hot bubbles provide low-density environments (lower than the typical ISM $n_{\rm H} \sim 0.1 ~{\rm cm}^{-3}$) where radiative losses becom inefficient, magnifying the dynamical impact of SNe. In fact, this is already hinted qualitatively in Fig. \[fig:ismfaceedgestars0399\] where several massive stars can be found within the hot and diffuse superbubbles. There is almost no SN that occurs in dense star-forming gas. This is due to the effect of photoionization which evacuates the gas prior to the SN events, consistent with previous studies [@2017MNRAS.466.3293P; @2017MNRAS.471.2151H]. The cumulative distribution functions are converged at $m_{\rm gas} = 5 {\rm M_\odot}$. In the 25M$_\odot$-run, SNe occur in colder and denser environments, though the majority of them are still resolved. In the 125M$_\odot$-run, no SNe occur in environments where $T_{\rm SN} > 10^4$K due to the lack of hot phase (cf. Fig. \[fig:pd2by2converge\]), and most SNe occur in typical ISM density $n_{\rm H} \sim 0.01 - 1 ~{\rm cm}^{-3}$. The fact that $n_{\rm SN}$ increases with $m_{\rm gas}$ makes SNe even more poorly-resolved. As almost all SNe are unresolved, SN feedback is mostly done by $p_{\rm term}$-injection.
### SN injection scheme {#sec:SNinjectResult}
In this section, we investigate how different SN injection schemes affect our results. We compare the following three different schemes:
- default scheme as described in section \[sec:SNfeedback\]
- default scheme without the [HealPix]{}-based injection
- pure thermal injection
In Fig. \[fig:mloadinject\], we show the outflow mass loading factor as a function of radius with three different injection schemes of SN feedback. The top, middle and bottom panels are runs with $m_{\rm gas} = $5, 25 and 125 M$_\odot$, respectively. The vertical error bars show the temporal fluctuations between 16 and 84 percentiles.
In the 5M$_\odot$-runs, three different injection schemes lead to almost indistinguishable radial profile of $\eta_m^{out}$ in the entire range of $r_s \in$ \[0.1kpc, 44.4kpc\]. It has been shown that for an SNR in a uniform medium, its evolution is insensitive to the form of injected energy (kinetic vs. thermal) as long as the SN is resolved [@2012MNRAS.419..465D; @2015ApJ...809...69S]. It is reassuring that we can further extend this statement in a complicated system like our simulated galaxy: the wind properties are insensitive to the form of injected energy as long as individual SNe are resolved, because there is enough time for each SNR to naturally evolve towards the ST solution. The effect of anisotropic injection is also negligible here, presumably because the ambient density structure of SNe is properly resolved such that even without the [HealPix]{} approach, the injection is still sufficiently isotropic.
As the resolution coarsens, the injection scheme starts to make a difference. In the 25M$_\odot$-runs, the default injection with and without [HealPix]{} show a factor of two difference in $\eta_m^{out}$ in the range of $r_s \in$ \[2kpc, $R_{\rm vir}$\]. Without [HealPix]{}, kinetic energy is preferentially (and unphysically) injected into overdensities of gas, which weakens the winds in an inhomogeneous medium as the overdensities are more difficult to be accelerated. The pure thermal injection also leads to a slightly lower $\eta_m^{out}$ at $R_{\rm vir}$ (which coincidentally agrees with the default injection without [HealPix]{}) compared to the default injection. This suggests that the pure thermal injection may already suffer from the overcooling problem at $m_{\rm gas} = $25 M$_\odot$, as the SNe are only marginally resolved (cf. Fig \[fig:snrhot\]). We note that the differences between these runs are only a factor of two which are comparable to or even less than the temporal fluctuations.
In the 125M$_\odot$-runs, the injection scheme has a significant effect on $\eta_m^{out}$. Since most SNe are unresolved at this resolution, the pure thermal injection suffers from the overcooling problem as expected and thus underestimates $\eta_m^{out}$ by around one order of magnitude. Surprisingly, the default injection, which is mainly $p_{\rm term}$-injection at this resolution (cf. Fig \[fig:snrhot\]), fails even more in terms of wind launching, with $\eta_m^{out}$ three orders of magnitude lower than the converged values already at $r_s \sim 7$kpc and almost zero mass flux at $R_{\rm vir}$. Momentum injection without [HealPix]{} fails even more due to its mass bias, which preferentially inject momentum into the directions parallel to the disk plane.
The reason that $p_{\rm term}$-injection leads to even weaker winds than pure thermal injection does is due to its assumption that the thermal energy has already been radiated away right from the injection. As such, winds can only be momentum-driven in this case. Apparently, it is insufficient to launch winds solely by momentum (at least in our simulations). In Fig. \[fig:coldbubble\], we show the maps of column density (top panels) and temperature slices (bottom panels) in the edge-on view with $m_{\rm gas} = 125 {\rm M_\odot}$. The left panels are with pure thermal injection while the right panels are with our default scheme which is mostly $p_{\rm term}$-injection at this resolution, shown at the time where the most significant SN-bubbles are developed. Massive stars are overplotted as white circles. The fluid velocity is over-plotted as arrows. The SN-bubbles in the left and right panels are similar in size, but they have strikingly different gas temperatures. Thermal injection could generate hot bubbles in situations where SNe are highly clustered, while $p_{\rm term}$-injection generates warm bubbles by construction even with highly clustered SNe. By the time the bubbles break out, only the hot bubble is able to vent out hot gas with high velocities and launch the pressure-driven winds. We emphasize that pure thermal injection does still suffer from the overcooling problem: most of the time (when SNe are not clustered enough) the hot gas cools numerically and winds can not be launched as efficiently.
Discussion {#sec:discussion}
==========
Comparison with previous studies
--------------------------------
@2017ApJ...841..101L use stratified-box simulations to study galactic winds across a range of gas surface density. They find $\eta_m^{out} \sim 6$, $\eta_e^{out} \sim 0.2$ and $\eta_Z^{out} \sim 0.7$ in their low-surface-density model (their Fig. 10). This agrees surprisingly well with our model which finds $\eta_m^{out} \sim 7$, $\eta_e^{out} \sim 0.1$ and $\eta_Z^{out} \sim 0.55$ at the launching radius, which is remarkable as their numerical model is very different from ours: they use the grid-based code [Enzo]{} [@2014ApJS..211...19B] without self-gravity and star formation and inject SNe at random locations with a predefined SN rate.
@2017MNRAS.470L..39F conduct isolated-disk simulations with the grid-based code [Athena]{} [@2008ApJS..178..137S]. They adopt a similar initial conditions to ours but their numerical model do not include self-gravity, cooling below $10^4$K and star formation, and the SN rate is determined by the local gas free-fall time and an efficiency parameter mimicking a Schmidt-law star formation model. Their model with lowest gas surface density ($\Sigma_{\rm gas} = 10 {\rm M_\odot}$) and individually injected SNe shows $\eta_m^{out} \sim 1$ and $\eta_e^{out} \sim 0.01$, which are significantly smaller than ours, especially for the energy loading factor.
The fact that our results agrees better with @2017ApJ...841..101L than with @2017MNRAS.470L..39F is likely due to the SN environment. @2017ApJ...841..101L put SNe at random locations and so many of them will occur in diffuse environments. In our case, as shown in Fig. \[fig:snrhot\], most of SNe also occur in densities lower than the peak ISM density because of SN clustering. In contrast, in @2017MNRAS.470L..39F, SNe preferentially occur in high-density environments, which would lead to much weaker winds as shown in @2016MNRAS.456.3432G. Indeed, when they group several SNe together to mimic the effect of SN clustering, they find loading factors much similar to ours ($\eta_m^{out} \sim 3$ and $\eta_e^{out} \sim 0.1$ in their $f_{\rm cl}=10$ model).
Interestingly, the winds predicted by our model are much weaker than what cosmological simulations typically adopt (or effectively adopt) to reproduce realistic galaxy population (see e.g. [@2011MNRAS.415...11D; @2014MNRAS.444.1518V; @2015MNRAS.446..521S]). This suggest that either other important feedback processes are still missing in our model, or that the wind loading factors in cosmological simulations need to be revised down. More observational constraints on galactic winds of dwarf galaxies are highly desirable (e.g. [@2018MNRAS.477.3164M]).
@2018ApJ...853..173K investigate galactic winds in a solar-neighborhood condition with stratified-box simulations and a self-consistent model that follows self-gravity, cooling, star formation and SN feedback. They find that the warm gas forms fountain flows while the hot gas has nearly constant $\eta_m^{out}$ and $\eta_e^{out}$ above 1 kpc. Our dwarf galaxy shows similar wind properties in the sense that the fountain flows are mostly warm and that the loading factors of winds remain nearly constant after launching ($r_s \gtrsim 2$kpc). In addition, the typical velocity of warm gas at launching is also comparable (50 $-$ 100km s$^{-1}$). However, they find much lower loading factors compared to ours ($\eta_m^{out} \sim 0.1$ and $\eta_e^{out} \sim 0.02$). In addition, beyond the launching radius, they do not find energy transfer from hot gas to warm gas to be significant, while in our case a fraction of warm gas is efficiently accelerated by the hot gas and eventually escapes the halo. The discrepancy should not be too surprising as these two set of simulations model very different galaxies. Our dwarf galaxy has a much shallower gravitational potential which makes warm gas easier to be accelerated to large $r_s$. As it moves outward and expands, its cross section increases, making it even easier to be accelerated by ram pressure.
Isotropic momentum injection: when does it matter?
--------------------------------------------------
@2018MNRAS.477.1578H have introduced a method of injecting SN momentum isotropically by carefully distributing momentum based on the effective areas of neighboring particles. Their method has the advantage that isotropy is imposed on a particle level, as opposed to our method where there can still be anisotropy within each pixel (though the number of pixels can easily be increased if desired). However, since their method still relies on a single neighbor search to find the $N_{\rm inj}$ nearest neighbors, in extreme situations where all nearest neighbors are locates within a small solid angle of an SN (e.g. a dense clump), the injection can still be anisotropic. Our [HealPix]{}-based method has the advantage that isotropy is guaranteed on a pixel level (i.e. there will be no empty pixels) and thus can properly cope with the above-mentioned situation, at the expense of having to do $N_{\rm pix}$ neighbor searches per SN (which is never the bottleneck in simulations).
@2018MNRAS.477.1578H demonstrated that the properties of their simulated galaxies can differ dramatically by switching from the simple inject scheme (that suffers from the anisotropic problem) to their improved scheme. They reported that the difference is more pronounced at high resolution ($m_{\rm gas} = 7\times 10^3 {\rm M_\odot}$) than at low resolution ($m_{\rm gas} = 5.6\times 10^4 {\rm M_\odot}$) as the anisotropic error in velocity increases for the same momentum budget when the injection region becomes smaller. In contrast, our improved scheme makes little difference to the galactic winds, especially at high resolution. This is not necessarily a conflict with @2018MNRAS.477.1578H as we are exploring a much higher resolution range. In our high-resolution run ($m_{\rm gas} = 5{\rm M_\odot}$), all SNe are properly resolved and 72% of energy is injected as thermal energy which does not suffer from the anisotropic problem. Although 28% of energy is still injected as kinetic energy, the difference between the two schemes turns out to be negligible. In our lowest-resolution run ($m_{\rm gas} = 125{\rm M_\odot}$), the difference becomes the largest, but in this case both schemes result in almost no winds anyway, and thus the anisotropic problem is hardly a concern (compared to the lack of hot gas).
Injecting terminal momentum as a sub-grid model for SN feedback?
----------------------------------------------------------------
In large-scale cosmological simulations, SN feedback will be necessarily unresolved and one always needs to adopt a sub-grid model. Our numerical experiments suggest that while injecting $p_{\rm term}$ as a sub-grid model can be a viable way to drive turbulence in the ISM and regulate star formation [@2013ApJ...776....1K; @2015MNRAS.450..504M; @2015MNRAS.451.2900K], it is not a suitable sub-grid model when it comes to launching galactic winds, as its underlying assumption is that thermal energy has already been radiated away right after the injection. Such an assumption leads to a huge underproduction of hot gas, which has been shown to be the dominant population that can be launched as winds in many SNe-resolved stratified-box simulations [@2016MNRAS.456.3432G; @2017MNRAS.466.1903G; @2017ApJ...841..101L; @2017MNRAS.470L..39F; @2018ApJ...853..173K] as well as in our global-disk simulations. In other words, this sub-grid model breaks down because SN winds are pressure-driven rather than momentum-driven.
In principle, it is possible to generate some hot gas with $p_{\rm term}$-injection if the cumulative terminal momentum from a group of SNe becomes strong enough to shock-heat the ISM. In practice, however, this mechanism seems to be very inefficient in dwarf galaxies like ours. In fact, in our simulations, $p_{\rm term}$-injection generates even less hot gas (and hence weaker winds) than pure thermal injection. This should not be too surprising. After all, this sub-grid model is designed to recover the correct terminal momentum rather than generating the right amount of hot gas. On the other hand, in a solar-neighborhood setup, @2017ApJ...846..133K showed that the amount of hot gas in the ISM actually increases (and so does the outflow rate) as they degrade the resolution to cell size $\Delta x > 16$pc and stop seeing convergence (their Fig. 17). This overproduced hot gas is not due to shock-heating via $p_{\rm term}$-injection. Instead, it is due to their sink particle approach which suffers from artificially enhanced SN clustering at low resolution. This leads to a blown-out of the entire disk, and the subsequent SNe then become “resolved” (i.e. thermal energy is injected) which give rise to the hot gas-dominated ISM due to the so-called “thermal runaway” [@2015MNRAS.449.1057G].
We stress that we are not advocating using pure thermal injection as a sub-grid model to launch winds, as it suffers from the well-known over-cooling problem. Our results simply suggest that both schemes fail to launch winds if SNe are unresolved. It remains unclear how to devise a sub-grid model that can not only inject the right amount of momentum into the ISM but also properly generate hot gas that can be launched as winds.
Resolution requirement for convergence {#sec:resreq}
--------------------------------------
In our simulations, the wind properties converge at $m_{\rm gas} = 5 {\rm M_\odot}$. This exact number may depend on the number of neighboring particles affected by an SN ($N_{\rm inj}$). We adopt $N_{\rm inj} = 96$ in order to ensure that the injected region is properly resolved by at least one resolution element. In SPH, this corresponds to the particle number in a smoothing kernel $N_{\rm ngb}$ (=100 in this work). Such a large $N_{\rm ngb}$ (and hence $N_{\rm inj}$) is required in order to suppress the numerical noise and improve convergence in SPH [@2012MNRAS.425.1068D]. More aggressive choice (smaller) of $N_{\rm inj}$ can be adopted, though at the expense of not properly resolving the injected region, which may trigger numerical fluid instabilities. We explore the choice of $N_{\rm inj}$ and its effect on wind properties in Appendix \[app:Ninj\]. In other particle codes such as the meshless-finite-mass method [@2015MNRAS.450...53H], it is reasonable to use a lower $N_{\rm inj}$ as the method has been shown to perform well with fewer neighbors (the standard value is $N_{\rm ngb} = 32$). Another subtlety is whether energy/momentum is distributed to neighboring particles uniformly or weighted with the kernel function. If a kernel weighting is adopted, the effective $N_{\rm ngb}$ would be even smaller as most energy/momentum will be distributed to the nearest neighbors that have higher weightings. This means that the resolution requirement (Eq. \[eq:coolmass\]) can actually be less stringent. We chose to be conservative by using a large $N_{\rm inj}$ in this work, and thus our convergence requirement ($m_{\rm gas} = 5 {\rm M_\odot}$) can be viewed as a lower limit. However, our claim that one must resolve individual SNe to generate hot gas and launch winds should remain qualitatively robust.
In practice, it is not enough to only marginally satisfy the resolution requirement in Eq. \[eq:coolmass\]. In our 25M$_\odot$-run, despite resolving 70% of SNe, the SN environments failed to converge (see Fig. \[fig:snrhot\]). This is because a marginally resolved SN defined as Eq. \[eq:coolmass\] is still unable to resolve the diffuse bubble and dense shell of an SNR, leading to a smoothed-out density structure. When subsequent SNe occur within this no-so-diffuse bubble, they will not be as efficient as they physically should be, which in turn weakens the winds. Our results suggest that the cooling mass needs to be resolved by around 5 injection mass, which is consistent with other grid-based simulations [@2015ApJ...809...69S; @2015ApJ...802...99K] who show that the cooling radius needs to be resolved by at least a few cells. **We expect a similar resolution requirement (modulo the slightly larger $N_{\rm inj}$ required by SPH) would apply for other Lagrangian methods such as MFM and the moving mesh codes, though this still needs to be explicitly tested. Indeed, @2018MNRAS.478..302S who use the moving mesh code [Arepo]{} [@2010MNRAS.401..791S] show that in their galactic scale simulations even with $m_{\rm gas} = 20 {\rm M_\odot}$ their wind properties are still not converged.**
**Intriguingly, the recent FIRE-2 cosmological “zoom-in” simulations presented in @2018MNRAS.477.1578H, show that their stellar mass evolution of similar dwarf galaxies is well converged at $m_{\rm gas} = 2000 {\rm M_\odot}$, which appears to be in conflicted with the results in both @2018MNRAS.478..302S and in this work.** The reason for the discrepancy requires further investigation, but a few points are worth noting: (i) The convergence of stellar mass does not necessarily imply the convergence of wind properties, though these two are expected to be closely related. (ii) At $m_{\rm gas} = 2000 {\rm M_\odot}$, most SNe in the FIRE-2 simulations are far from resolved (even with $N_{\rm inj} = 1$). Therefore, the fact that the convergence can still be achieved suggests that the galactic winds in the FIRE-2 simulations may not be driven by SNe but by other included feedback processes such as photoionization, stellar winds and radiative pressure[^14].
Summary {#sec:summary}
=======
We have conducted high-resolution smoothed-particle hydrodynamics simulations ($m_{\rm gas} = 1{\rm M_\odot}$) of an isolated dwarf galaxy with a self-consistent ISM model which includes self-gravity, non-equilibrium cooling and heating, an H$_2$ chemistry network, individual star formation, stellar feedback and metal enrichment. We study the properties of SN-driven winds of the galaxy and investigate the convergence of our model. Our main results are summarized as follows.
1. Our self-consistent ISM model naturally leads to clustered SNe. Around 60% of SNe occur within the preexisting hot bubbles (Fig. \[fig:snrhot\]), which greatly enhances their dynamical impact on the ISM and leads to the formation of superbubbles with sizes of a few hundred pc (Fig. \[fig:ismfaceedgestars0399\]). These superbubbles are able to break out of the disk, venting out hot and over-pressurized gas, which becomes the driving force of winds.
2. In steady state, the gas flows can be characterized by the fountain flows and the winds. The fountain flows dominate (in mass) at small $r_s$ but they rapidly decline with $r_s$. At $r_s > 2$kpc, the fountain flows vanish and the actual winds emerge with nearly constant loading factors all the way to $R_{\rm vir}$. We find $\eta_m^{out} \sim 3$, $\eta_p^{out} \sim 1$ and $\eta_e^{out} \sim 0.05$ at $R_{\rm vir}$ (Fig. \[fig:mpeloadingz\]). As a significant amount of gas in the ISM is blown out of the disk, the winds are only slightly metal-enriched ($y_Z^{out} \sim 1.5$ at $R_{\rm vir}$) due to the dilution by the unenriched ISM. The corresponding metal loading factor is $\eta_Z^{out} \sim 0.2$. The predicted winds are much weaker than what is typically adopted in cosmological simulations, which implies either our model is missing important feedback mechanisms or the winds in cosmological simulations need to be revised down.
3. At $r_s < 2$kpc, most mass and momentum are carried by the warm fountain flows, while most energy is carried by the outflowing hot gas (Fig. \[fig:mpeloadingzphases\]). After the winds are launched ($r_s > 2$kpc), they can be separated into two populations depending on their launching temperatures (Fig. \[fig:historywarm\] and \[fig:historyhot\]). The hot gas acquires very high velocity (larger than $v_{\rm esc}$) after the breakout of SN bubbles and is able to escape the halo on its own, with a nearly constant Bernoulli parameter (or slightly decreasing as it feeds energy to the warm gas). On the other hand, the warm gas acquires velocity lower than $v_{\rm esc}$ at launching, but is later on accelerated and shock-heated by the subsequently launched hot winds, and eventually obtains enough energy to escape the halo. Therefore, its Bernoulli parameter keeps increasing with $r_s$. This highlights the caveat of extrapolating properties of warm gas from the launching radius to $R_{\rm vir}$ in small-box simulations.
4. The wind properties converge at $m_{\rm gas} = 5 {\rm M_\odot}$ where the injection mass is $500 {\rm M_\odot}$ (Fig. \[fig:mloadconv\]). This corresponds to the resolution where the cooling mass (Eq. \[eq:coolmass\]) of individual SNe can be resolved by at least 5 resolution elements, or 5 kernel masses (Fig. \[fig:snrhot\]). When SNe are unresolved, no hot gas can be generated and therefore no winds will be launched (Fig. \[fig:pd2by2converge\]).
5. When SNe are properly resolved, winds are insensitive to the injection scheme of SN feedback (Fig. \[fig:mloadinject\]). Injecting 100% thermal energy leads to the same wind properties as injecting a mixture of thermal and kinetic energy based on the exact Sedov-Taylor solution. In addition, our new pixel-by-pixel momentum injection scheme with improved isotropy based on the [HealPix]{} method has little effect on winds when SNe are resolved.
6. When SNe are unresolved, injecting thermal energy suffers from the overcooling problem as expected. Injecting the terminal momentum $p_{\rm term}$, a popular sub-grid model in the literature, also fails to capture winds (injecting the residual thermal energy does not help, as shown in Appendix \[app:resEth\]). This is because winds are driven by thermal pressure carried by the hot gas during the breakout of SN bubbles, rather than driven by the momentum carried mostly by the warm gas. Since $p_{\rm term}$-injection assumes that thermal energy has already been radiated away, it is unable to generate hot gas to launch winds (Fig. \[fig:coldbubble\]). Further investigations are required to devise a sub-grid model that can not only generate turbulence in the ISM and regulate star formation but also produce enough hot gas to launch winds.
Acknowledgments {#acknowledgments .unnumbered}
===============
I thank the referee for the useful comments on the paper. I would also like to thank Greg Bryan, Chang-Goo Kim, Miao Li, Mordecai Mac Low, Thorsten Naab and Phil Hopkins for enlightening discussions, and Bernhard Röttgers for his help on [pygad]{}[^15], which I use for visualization in this work. The Center for Computational Astrophysics is supported by the Simons Foundation.
Effect of the residual thermal energy {#app:resEth}
=====================================
As discussed in Section \[sec:subgridSN\], injecting the terminal momentum of the SNR has become a widely-used sub-grid model in the literature. However, the residual thermal energy is not always included (see, e.g., [@2015MNRAS.449.1057G; @2015MNRAS.454..238W; @2017ApJ...846..133K]). In this appendix, we investigate the effect of the residual thermal energy by repeating the 25${\rm M_\odot}$- and 125${\rm M_\odot}$-runs without injecting $E^{\rm res}_{\rm th}$. In Fig. \[fig:comparereseth\], we show $\eta^{out}_m$ (top panel) and $\eta^{out}_e$ (bottom panel) as a function of $r_s$ for these two runs (dashed lines) along with the runs that include $E^{\rm res}_{\rm th}$ (solid lines). The vertical error bars indicate the temporal fluctuations between 16 and 84 percentiles.
The residual thermal energy has a negligible effect on the wind loading factors. This is not surprising as thermal energy drops rapidly after cooling kicks in. In addition, since the unresolved SN is not able to clear out the gas and form a low-density bubble, any residual thermal energy is forced to be injected in a dense region instead of in the low-density bubble as it physically should be, and will therefore be radiated away even more rapidly for numerical reasons. As a result, it has little impact on the wind properties.
![$\eta^{out}_m$ (top panel) and $\eta^{out}_e$ (bottom panel) as a function of $r_s$ for the 25${\rm M_\odot}$- and 125${\rm M_\odot}$-runs, with (solid lines) and without (dashed lines) injecting the residual thermal energy. The vertical error bars indicate the temporal fluctuations between 16 and 84 percentiles. Wind loading factors are insensitive to whether or not the residual thermal energy is included.[]{data-label="fig:comparereseth"}](compareResEth){width="0.99\linewidth"}
Effect of the injection mass {#app:Ninj}
============================
As discussed in Section \[sec:resreq\], we adopt a conservative choice of $N_{\rm inj} = 96$ to make sure that the injection region comparable to the SPH resolution and hence is properly resolved. In this appendix, we explore a more aggressive (around 5 times smaller) choice of $N_{\rm ing} = 20$ for the non-converged runs and test whether the convergence can be improved. In Fig. \[fig:meloadingninj\], we show the same plot as Fig. \[fig:mloadconv\] but overplotting with three new runs with 5 times smaller injection mass (dashed lines) and $m_{\rm gas} = 25, 125$ and $625 {\rm M_\odot}$, respectively.
The 25${\rm M_\odot}$-run is insensitive to the choice of $N_{\rm inj}$ and underpredicts $\eta^{out}_m$ and $\eta^{out}_e$ by a factor of 3 and 5 at $R_{\rm vir}$, respectively. Note that the 25${\rm M_\odot}$-run with $N_{\rm inj} = 20$ has nearly the same $M_{\rm inj}$ as the 5${\rm M_\odot}$-run with $N_{\rm inj} = 96$, but the former is not converged while the latter is. This is because, despite having the same $M_{\rm inj}$, the one that has poorer SPH resolution will not be able to resolve the gas structure (e.g. diffuse bubbles, dense shells and low density channels). As such, it is dangerous to presume that convergence can be achieved at $x$-times higher $m_{\rm gas}$ just by simply adopting a $x$-times smaller $N_{\rm inj}$.
The 125${\rm M_\odot}$-run shows a significant stronger winds when a smaller $N_{\rm inj}$ is adopted, though both $\eta^{out}_m$ and $\eta^{out}_e$ are still a factor of 5 and 10 lower than the converged values, respectively. This is because with this injection mass most of the SNe are marginally resolved (cf. Fig. \[fig:snrhot\]). Meanwhile, the 625${\rm M_\odot}$-run has almost no winds. This is in line with our conclusion that winds weaken dramatically once the SNe become unresolved, and by adopting a smaller $N_{\rm inj}$, the transition scale is shifted towards higher mass scale.
To sum up, lowering $N_{\rm inj}$ does increase the transition particle mass where winds weaken dramatically (from $m_{\rm gas} = 125{\rm M_\odot}$ to $625 {\rm M_\odot}$), which is defined by Eq. \[eq:coolmass\]. However, the particle mass required for converged results ($m_{\rm gas} = 5{\rm M_\odot}$) does not increase with a smaller $N_{\rm inj}$.
![Same as Fig. \[fig:mloadconv\] but overplotting with three new runs with 5 times smaller injection mass (dashed lines) and $m_{\rm gas} = 25, 125$ and $625 {\rm M_\odot}$, respectively. Lowering $N_{\rm inj}$ does increase the transition particle mass where winds weaken dramatically (from $m_{\rm gas} = 125{\rm M_\odot}$ to $625 {\rm M_\odot}$), which is defined by Eq. \[eq:coolmass\]. However, the particle mass required for converged results ($m_{\rm gas} = 5{\rm M_\odot}$) does not increase with a smaller $N_{\rm inj}$. []{data-label="fig:meloadingninj"}](MEloading_Ninj){width="0.99\linewidth"}
[^1]: [email protected]
[^2]: Note that since we are tracking individual stars, the metal yields we adopt are mass-dependent rather than IMF-averaged.
[^3]: The particles that are assigned with zero mass during the process will be removed.
[^4]: At our fiducial resolution (1M$_\odot$), each particle will only have a single stellar mass.
[^5]: This value is chosen to be close to $N_{\rm ngb} = 100$, the number of particles within an SPH kernel, such that the injected region is hydrodynamically resolved.
[^6]: This is not the case when $m_{\rm gas} \gtrsim 100{\rm M_\odot}$. However, at this resolution, most SNe become unresolved anyway and we would switch to the sub-grid model described in \[sec:subgridSN\].
[^7]: We note that the metallicity ($Z$) in our initial setup is 10 times lower than that in @2015ApJ...802...99K. However, $R_c$ only has a very weak dependence on metallicity ($R_c \propto Z^{-0.14}$, see e.g. [@1988ApJ...334..252C]).
[^8]: With our resolution ($m_{\rm gas} = 1 {\rm M_\odot}$), star particles actually represent individual stars and therefore the two-body interaction is physical rather than numerical. We still adopt a small but finite softening length just to prevent them from having too small timesteps.
[^9]: Note that the integral of $F_{\rm p}$ is a vector. We define $\dot{p}$ by taking only the vector component normal to $S$.
[^10]: This leads to a small discontinuity of the wind properties at $r_s \approx 1.5$kpc as the transition is not perfectly smooth (cf. Fig. \[fig:mpeloadingz\], \[fig:mpeloadingzphases\] and \[fig:windprofiletimeave\]).
[^11]: While the disk reaches a quasi-steady state after $t = 0.2$Gyr, it takes a bit longer (0.4Gyr) for the winds at $R_{\rm vir}$ to do so.
[^12]: The warm gas here actually also includes what is conventionally called the cold ($T < 300$K) gas. However, the fraction of cold gas is negligible here. The majority of the warm gas has $T \approx 10^4$K.
[^13]: Following @2018ApJ...853..173K, we subtract out $\Phi(0)$ in the definition of $B$ such that the gravitational term ($\Phi(r) - \Phi(0)$) is always positive.
[^14]: We note that these feedback processes are all treated by simplified recipes (the same applies to the photoionization treatment in our model) and further investigations are required to understand their effects.
[^15]: https://bitbucket.org/broett/pygad
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the energy and angular distributions of the ions reaching the electrodes in low-pressure, capacitively coupled oxygen radio-frequency discharges. These distributions, as well as the possibilities of the independent control of the ion flux and the ion energy are analysed for different types of excitation: single- and classical dual-frequency, as well as valleys- and sawtooth-type waveforms. The studies are based on kinetic, particle-based simulations that reveal the physics of these discharges in great details. The conditions cover weakly collisional to highly collisional domains of ion transport via the electrode sheaths. Analytical models are also applied to understand the features of the energy and angular distribution functions.'
address: |
$^1$Institute for Solid State Physics and Optics, Wigner Research Centre for Physics, Hungarian Academy of Sciences, 1121 Budapest, Konkoly Thege Miklós str. 29-33, Hungary\
$^2$Center for Atomic and Molecular Technologies, Graduate School of Engineering, Osaka University, 2-1 Yamadaoka, Suita, Osaka 565-0871, Japan\
$^3$Department of Physics, West Virginia University, Morgantown, USA\
$^4$Institute for Electrical Engineering, Ruhr-University-Bochum, Bochum, Germany\
$^5$Evatec AG, 9477 Truebbach, Switzerland\
author:
- 'Zoltán Donkó$^{1,2}$, Aranka Derzsi$^{1,3}$, Máté Vass$^1$, Julian Schulze$^{3,4}$ Edmund Schuengel$^5$, Satoshi Hamaguchi$^2$'
title: 'Ion energy and angular distributions in low-pressure capacitive oxygen RF discharges driven by tailored voltage waveforms'
---
Introduction
============
The active species created in discharge plasmas provide the basis of various surface modification techniques, such as etching, deposition, surface microstructuring and functionalisation [@G1; @G2; @G3; @G4]. The species utilised in these applications are usually ions and radicals, which participate in different physical and chemical interactions with the surface layers. Controlling the flux, the energy and the angle of incidence of these active species at the surfaces is of primary importance. In order to be able to regulate the processes taking place at the plasma-surface interface the connection between the operating conditions of the plasma and the resulting ion properties, such as the ion flux, the mean ion energy, the Ion Flux-Energy Distribution Function (IFEDF), and the Ion Angular Distribution Function (IADF), has to be understood. Both the formation and the characteristics of the IFEDF and the IADF [@Wild; @Kawamura; @Eddi15; @Chen18; @WangS; @Hamaguchi; @Kushner; @Kratzer; @Woodworth; @Rakhimova; @Huang; @NEW2] received significant attention during the past decades. While the IFEDF is important in all applications, the IADF is highly relevant, e.g. in high aspect ratio etching [@Wu; @Donnelly].
Capacitively Coupled Plasmas (CCPs) represent one of the most important sources used in the applications mentioned above. In these systems the ions flying towards the electrodes have to traverse the sheaths (over which a significant voltage drop is present) and the ion energy and angular distributions are established within these regions [@Wild; @Kawamura; @Eddi15; @Chen18], the collisionality of which is an important factor in forming these distributions. Low pressures and narrow sheath widths result in collisionless or nearly collisionless transfer and a narrow angular distribution. At the other extreme, when the sheaths are much longer than the free path of the ions, the transport is highly collisional. This results in low-energy ions arriving with a broad angular distribution at the surfaces. At low pressures, the other important parameter is the ratio of the ion transit time to the period of the applied radio-frequency (RF) waveform. Under conditions when this ratio is small, the ions acquire an energy that corresponds closely to the instantaneous sheath voltage, while at high values of this ratio, the ions fly through the sheath during several RF periods, and their energy is determined by the time-averaged sheath voltage.
By changing the pressure, the driving frequency, and the voltage amplitude, a variety of IFEDFs can be realised [@DonkoEPS]. Nonetheless, various ways for an additional control of the ion properties have been searched for during the past decades. Solutions were found in the extra degrees of freedom provided by the driving voltage waveform, which can be more complex than the “default” single harmonic signal.
An independent control of the ion properties (the mean ion energy and the ion flux) was first made possible by introducing [*Dual-Frequency (DF) excitation*]{} to drive capacitively coupled plasmas in 1992 [@Goto]. When significantly different driving frequencies are used, the plasma production and charged particle densities are primarily controlled by the amplitude of the high-frequency voltage, while the transport of the ions across the sheaths is primarily determined by the low-frequency voltage amplitude. Properties of plasma sources, operated in various gases and under different conditions, driven by DF waveforms have thoroughly been studied [@DF1; @DF2; @DF3; @DF4]. These investigations have also revealed that the independent control of ion flux and energy is limited by “frequency coupling” effects [@coupling1; @coupling2] and secondary electron emission from the electrodes [@secondaries1; @secondaries2].
A further major step in the control of ion properties has been the discovery of the [*Electrical Asymmetry Effect*]{} (EAE) [@EAE; @Heil2] in 2008, by using a base frequency and its second harmonic to excite the plasma. Such a driving voltage was shown to lead to the development of a DC self-bias even in geometrically symmetrical systems. This self-bias can be controlled by the phase between the two harmonics and has a direct effect on the energy of the ions at the electrodes, while the ion flux remains approximately constant [@DZ-EAE]. The performance of the EAE in various electropositive and electronegative gases has been investigated in details, including electron power absorption mechanisms and transitions between them, the effects of the secondary electron emission and of the driving frequencies, etc. [@EAE-neg; @EAE-freq; @EAE-freq2; @ZYR; @EAE-general].
The possibility of using a higher number of harmonics, leading to [*peaks- and valleys-type waveforms*]{}, and the optimization of the harmonic voltage amplitudes have also been explored [@Eddi15; @optim1; @optim2; @Wendt1; @Wendt2]. These types of waveforms belong to the set of [*Tailored Voltage Waveforms*]{} (TVW, discussed in details in a comprehensive review [@Trevor]), which also cover [*sawtooth-type waveforms*]{} (introduced in 2014) [@sawtooth; @sawtooth2]. Through the past years TVWs were utilised as well in practical applications, e.g. in silicon thin film deposition [@EAE-depos; @EAE-depos2] and etching [@J_1; @J_2].
In this work we investigate oxygen plasmas. The choice of this gas is motivated by the complex physics of oxygen plasmas, as well as by the practical importance of this gas: oxygen plasmas have been used, e.g., in high-tech applications based on etching by reactive plasma species [@Coburn; @Childres], in patterning of highly oriented pyrolytic graphite [@Lu], resist stripping for multilayer lithography [@Hartney], in the modification of various coatings and films [@Cvelbar; @Nakamura; @Chou], in the creation of micropatterns of chemisorbed cell adhesion-repellent films [@Tourovskaia; @H1; @H2] and in the production of highly porous SnO$_2$ fibers [@Zhang].
Oxygen discharges have been the subject of numerous fundamental discharge studies. The EAE in oxygen CCPs was studied experimentally and via particle simulations by Schuengel [*et al.*]{}[@Eddi-O2] and Zhang [*et al.*]{} [@Zhang-O2]. Emission patterns this is the last candidate. next esc will revert to uncompleted text. hat reveal information about the spatio-temporal distribution of the excitation rates caused by different species have been analysed by Dittmann [*et al.*]{} [@Dittmann]. The surface recombination of the singlet delta oxygen metastable molecules has been addressed in [@Greb; @Greb2; @NEW]. A comparison between experiments and simulations for peaks- and valleys-type waveforms, focusing on the DC self-bias, the ion flux, the discharge power, and the ion flux-energy distributions has been presented in [@Derzsi2016]. These investigations have been extended by Phase Resolved Optical Emission Spectroscopy (PROES) studies. Derzsi [*et al.*]{} [@Derzsi2017] presented a detailed comparison between experimental and simulation data for excitation maps for peaks- and valleys-type, as well as sawtooth-type waveforms. Transitions between the electron power absorption modes associated with the sheath dynamics (“$\alpha$-mode”) and with the bulk and ambipolar electric fields (“Drift-Ambipolar” or “DA”-mode [@DAmode]) have been identified as a function of the operating conditions [@Derzsi2017; @Gud2017]. A (limited) sensitivity analysis of some of the model parameters has been presented by Donkó [*et al*]{} [@DonkoEPS2018], while the effect of the driving frequency has recently been analysed via Particle-in-Cell simulations incorporating Monte Carlo collisions (PIC/MCC) by Gudmundsson [*et al.*]{} [@G2018]. The effect of creation of negative ions at the electrode surfaces has been addressed in [@Matt]. These (and several other) studies have uncovered much of the effects taking place in low-pressure oxygen CCPs, however, the analysis of the formation of the ion angular and energy distribution functions in discharges driven by different voltage waveforms is not yet understood and warrants further investigations.
Therefore, in this paper, we investigate the ion properties in low-pressure oxygen CCPs driven by various waveform types: single-frequency, classical dual-frequency, valleys and sawtooth. In particular, we focus on the flux-energy distributions, the angular distributions, as well as the joint energy and angular distributions of the impinging positive ions at the electrodes, as a function of (i) the driving voltage waveform, (ii) the gas pressure, and (iii) the ion-induced secondary electron emission coefficient (SEEC, $\gamma$). We present, as well, a simple model for the formation of the joint energy and angular distribution at conditions when the sheaths are weakly collisional and use a more elaborated model [@Eddi] to derive the IFEDF for any conditions. We also discuss general characteristics of oxygen plasmas such as electron power absorption modes, electronegativity, and behaviour of the DC self-bias under the conditions of excitation by TVWs. The discharge model (including the set of elementary processes and the definition of the driving voltage waveforms) is presented in section \[sec:model\]. The results of our investigations are presented in section 3, while section 4 gives a brief summary of our studies.
Discharge model, computational implementation, and discharge conditions {#sec:model}
=======================================================================
Below, we present the main features of the discharge model and its computational implementation. Subsequently, we specify the discharge conditions, including the various driving voltage waveforms that are used in our studies of oxygen CCPs.
Model of oxygen CCPs and computational implementation
-----------------------------------------------------
Our model of the oxygen CCPs is largely based on the “xpdp1” set of elementary processes [@Vahedi] and its recent revision [@Gudmundsson], and is the same as described in our previous works [@Derzsi2016; @Derzsi2017; @DonkoEPS2018]. Therefore, only a brief summary of the features of the model is given below and the reader is referred to [@Derzsi2016] for details.
The charged species considered in the model are O$_2^+$ and O$^-$ ions and electrons. The set of elementary collision processes between the electrons and O$_2$ neutral molecules includes elastic scattering, excitation to rotational, vibrational and electronic levels, ionisation, dissociative excitation, dissociative attachment, impact detachment, as well as dissociative recombination. For O$_2^+$ ions elastic collisions with O$_2$ are taken into account; we include the symmetric charge exchange process and an additional channel with isotropic scattering in the center-of-mass frame (the cross section for the isotropic channel is set to be 50% of the charge exchange cross section as suggested in [@Gudmundsson]). For O$^-$ ions the model includes elastic scattering with O$_2$ neutrals, detachment in collisions with electrons and O$_2$ molecules, mutual neutralization with O$_2^+$ ions, as well as collisions with metastable singlet delta oxygen molecules, O$_{2}(a^{1}\Delta_{\rm g})$. This latter species is known to play an important role in oxygen CCPs [@Greb; @Greb2; @NEW], especially in establishing the negative ion balance. In our model we assume that this species has a spatially uniform density that is computed from the balance between their creation rate in the gas phase by e$^-$+O$_2$ collisions and their loss rate at the electrode surfaces [@Derzsi2016].
Compared to the original xpdp1 set, we replace the elastic collision cross-section with the elastic momentum transfer cross-section of [@Biagi] and assume isotropic electron scattering, replace the xpdp1 ionisation cross-section with that recommended in [@Gudmundsson], and adopt as well all the ion-molecule and ion-ion collision cross-sections from [@Gudmundsson].
The model is implemented into a 1d3v Particle-in-Cell simulation code incorporating Monte Carlo treatment of collision processes (PIC/MCC) [@PIC1; @PIC2; @PIC3]. We assume plane and parallel electrodes with a gap of $L$ = 2.5 cm. One of the electrodes is driven by a voltage waveform (specified below), while the other electrode is at ground potential. Electrons are assumed to be elastically reflected from the electrodes with a probability of 0.2. The code allows including the emission of secondary electrons from the electrodes; we use SEEC values of $\gamma=0$ (disregard electron emission), $\gamma=0.06$ (a value that is characteristic for metal surfaces), and $\gamma=0.4$ (a value typical for dielectric surfaces). The gas temperature is fixed at $T_{\rm g}=350$ K. For the surface quenching probability of O$_{2}(a^{1}\Delta_{\rm g})$ singlet delta molecules we use the value of $\alpha = 6 \times 10^{-3}$, which resulted in our previous studies in a good overall agreement between the experimental and simulation data for the ion fluxes and the ion flux-energy distribution functions at the electrodes [@Derzsi2016; @Derzsi2017; @DonkoEPS2018].
In the cases when a DC self-bias voltage develops, its value is determined in an iterative manner to ensure equal losses of positive and negative charges at each electrode over one period of the fundamental driving frequency [@DZ-EAE]. This, self-consistently computed value is added to the excitation waveform applied at the powered electrode.
The computations are carried out using a spatial grid with $N_x$ = 100-1600 points and $N_t$ = 2000-85000 time steps within the fundamental RF period. These parameters have been set to fulfil the stability criteria of the computational scheme. We note that despite the relatively low number of different plasma species and elementary processes included in our model, our previous studies have concluded that this model is able to predict experimentally observable discharge characteristics with a reasonable accuracy, for a wide domain of conditions [@Derzsi2016; @Derzsi2017; @DonkoEPS2018].
Throughout this paper the distribution functions, viz. the ion flux energy distribution function, $F(\varepsilon)$, the ion angular distribution function, $F(\Theta)$, as well as the joint ion energy-angular distribution function, $F(\varepsilon,\Theta)$, represent the number of ions reaching the electrodes as a function of their energy ($\varepsilon$) and/or incidence angle ($\Theta$) [@Shihab]. This way, these distribution functions are “measured” in units of m$^{-2}$ s$^{-1}$ eV$^{-1}$, m$^{-2}$ s$^{-1}$ deg$^{-1}$, and m$^{-2}$ s$^{-1}$ eV$^{-1}$ deg$^{-1}$, respectively. It is noted that another definition of the angular distribution function that corresponds to the flux over solid angle is also commonly used; for a discussion see [@Shihab].
Excitation waveforms
--------------------
Simulations are conducted with various excitation waveforms, specified below. Table 1 lists these waveforms and summarises the parameters used. Subsequently, the mathematical forms of the excitation voltage waveforms are given.
[@lllll]{} Driving waveform&Frequency / Frequencies & Pressure \[Pa\] & SEEC $\gamma$ & Waveform\
Single harmonic & $f_1$ = 27.12 MHz & 5, 10, 20 & 0, 0.06, 0.4 & eq. (\[eq:exc1\])\
Classical dual frequency & $f_1$ = 27.12 MHz & $f_2 = f_1 / 14$ & 5, 10, 20 & 0, 0.06, 0.4 & eq. (\[eq:exc2\])\
Peaks and valleys & $f_1$ = 15 MHz, $f_k = k f_1$ & 5, 10, 20 & 0, 0.4 & eq. (\[eq:exc3\])\
Sawtooth & $f_1$ = 15 MHz, $f_k = k f_1$ & 5, 10, 20 & 0, 0.4 & eq. (\[eq:exc4\])\
The excitation waveforms for the different cases are:
- [single-frequency excitation: $$\phi(t) = \phi_1 \cos(2 \pi f_1 t),
\label{eq:exc1}$$]{}
- [classical dual-frequency excitation: $$\phi(t) = \phi_1 \cos(2 \pi f_1 t) + \phi_2 \cos(2 \pi f_2 t),
\label{eq:exc2}$$ ]{}
- [peaks- and valleys-type waveforms: $$\phi(t) = \sum_{k=1}^N \phi_k \cos(2 \pi k f_1 t + \theta_k),
\label{eq:exc3}$$ where $\theta_k$ are the phase angles and $N$ is the number of harmonics (for which we use a maximum number of 4). The $\phi_k $ amplitudes of the individual harmonics are set according to $$\phi_k = \frac{2(N-k+1)}{(N+1)^2} \phi_{\rm pp},
\label{eq:exc3a}$$ where $\phi_{\rm pp}$ is the peak-to-peak voltage. The peaks-type voltage waveforms can be realised by setting all $\theta_k$ angles to zero, while for the valleys-type waveforms the phase angles of all even harmonics have to be set to $\pi$. We note that by “reversing” the waveform (peaks $\leftrightarrow$ valleys), in the geometrically symmetrical configuration considered here the plasma is mirrored with respect to the centre of the electrode gap. (Therefore, it is sufficient to study discharges with one of the waveform types; we chose the valleys-type waveform in the present study.)]{}
- [sawtooth-type waveforms: $$\phi(t) = \pm \phi_{\rm a} {\sum_{k=1}^N \frac{1}{k}\sin(2 \pi k f_1 t)},
\label{eq:exc4}$$ where the minus and plus signs define, respectively, the sawtooth-up and sawtooth-down waveforms. The value of the $\phi_{\rm a}$ prefactor is set (for each value of the number of harmonics, $N$) in a way that the waveform has the specified peak-to-peak voltage, $\phi_{\rm pp}$.]{}
Examples of driving voltage waveforms specified above are displayed in figure \[fig:waves\].
![Driving voltage waveforms covered in this work: (a) single- ($\phi_2$ = 0 V) and classical dual-frequency ($\phi_2 >$ 0 V) waveforms with $\phi_1$ = 200 V, (b) valleys-type and (c) sawtooth-type waveforms with different number of harmonics ($N$), for $\phi_{\rm pp}$ = 400 V. $T_1$ is the period of the fundamental frequency $f_1$.[]{data-label="fig:waves"}](waveforms.pdf){width="\textwidth"}
Results
=======
Single-frequency excitation – the general behaviour of the discharge
--------------------------------------------------------------------
First, we address the discharge behaviour under the simplest form of excitation, driving the plasma with a single-frequency harmonic voltage source given by eq.(\[eq:exc1\]), with $f_1=$ 27.12 MHz. Figure \[fig:singlefreq1\] presents the simulation results for pressure values of 5 Pa and 20 Pa, for $\phi_{\rm pp}$ = 400 V peak-to-peak driving voltage, $L$ = 2.5 cm electrode gap, and an ion-induced secondary electron emission coefficient $\gamma=0$. Panels (a) and (d) show the time-averaged density distributions of the charged species. The discharge is highly electronegative at both pressures, the electronegativity decreases with increasing pressure, but the peak negative ion density remains approximately an order of magnitude higher than the peak electron density even at 20 Pa. (We define the electronegativity ($\beta$) as the ratio of the spatially averaged densities of negative ions and electrons.) A similar dependence of the electronegativity of the plasma on the oxygen pressures was also observed in [@Eddi-O2] and [@LVL]. Panels (b) and (e) of figure \[fig:singlefreq1\] depict the electron power absorption rate, space- and time-resolved (for five radiofrequency periods). The power absorption peaks in both cases in the vicinity of the expanding sheath edges. The position of the sheath edges is found by the criterium defined by Brinkmann [@Brinkmann]. Taking as an example the sheath adjacent to the powered electrode situated at $x=0$, the position of the sheath edge, $s$, is found from $$\int_0^s n_{\rm e}(x) {\rm d} x = \int_s^{x^\ast} [n_{\rm i}(x)- n_{\rm e}(x)] {\rm d} x,$$ where $n_{\rm i}$ and $n_{\rm e}$ are, respectively, the total positive and negative charged particle densities, and $x^\ast$ is a position towards the centre of the discharge where quasineutrality holds.
Power absorption in the bulk plasma domain and at the edges of the collapsing sheaths, which is a sign of the Drift-Ambipolar power absorption mode [@DAmode], is revealed at the lower pressure of 5 Pa only. In this latter case the mean energy of the electrons in the centre of the plasma is 6.3 eV, whereas in the 20 Pa case it is only 2.1 eV. The high value of the mean electron energy at the lower pressure is required because of the need to conduct the current by the electrons having a depleted density. At higher pressure, where the electron density in the bulk is about an order of magnitude higher with respect to the 5 Pa case (the electronegativity decreases), such a high electron energy is not required and the mean electron energy decays significantly.
![Characteristics of single-frequency ($f_1=$ 27.12 MHz) oxygen discharges at 5 Pa (left) and 20 Pa (right) pressures. Panels (a, d) show the time-averaged density distributions of the charged species, while (b, e) display the spatio-temporal electron power absorption rate for five RF periods ($T$). In panels (c, f) the energy and angular distribution of the O$_2^+$ ions, $F(\varepsilon,\Theta)$, is given (in arbitrary units) at the electrodes. The insets illustrate the energy and angular distributions (respective integrals of $F(\varepsilon,\Theta)$ according to incidence angle and energy). All distributions have been normalised to a maximum value of 1, for easier comparison. $\phi_{\rm pp}$ = 400 V, $\gamma=0$. The black lines in panels (b) and (e) mark the sheath edges.[]{data-label="fig:singlefreq1"}](singlefreq1.pdf){width="85.00000%"}
We note that oxygen plasmas are highly electronegative at low pressures, while discharges operated in, e.g., CF$_4$, are highly electronegative at high pressures and much less electronegative at low pressures. In O$_2$, detachment due to collisions of negative ions with metastable molecules causes the decrease of the electronegativity as a function of pressure (as confirmed by our simulation results, not discussed here in details). Such different behaviour of electronegative plasmas can lead to different modes of discharge operation with drastic consequences for process control.
The $F(\varepsilon,\Theta)$ flux energy and angular distribution of the O$_2^+$ ions upon their impingement at the electrodes is displayed in figures \[fig:singlefreq1\](c) and (f), for the two pressures studied. Insets of these panels show the flux energy distribution $F(\varepsilon)$ (integral of $F(\varepsilon,\Theta)$ over all angles) and the flux angular distribution $F(\Theta)$ (integral of $F(\varepsilon,\Theta)$ over all energies). At low pressure (5 Pa) most of the ions arrive at angles smaller than a few degrees and the energy distribution peaks at $\varepsilon \approx$ 100 eV, that corresponds to the time average of the sheath voltage. This indicates a weakly collisional sheath, which the ions traverse, however, over several RF cycles. At the higher pressure of 20 Pa the $F(\varepsilon,\Theta)$ distribution reveals several discrete spots that originate from the charge exchange collisions, which, in the presence of a periodically varying sheath electric field, synchronise the motion of groups of ions. (Ions that have undergone a charge exchange collision at times of low electric field “gather” until the field grows again and moves these ions – for details see [@Wild; @Eddi].) In general, the angular distribution of the incoming ions broadens and the energy distribution, due to the higher collisionality of the electrode sheaths, becomes a decreasing function of the ion energy (except of its peaks originating from charge exchange collisions).
![The colour map shows the $F(\overline{\varepsilon}_{\rm a},\Theta)$ distribution of the O$_2^+$ ions arriving at the electrodes, as obtained from the model. The distribution has been normalised to a maximum value of 1. The lines show the loci of the possible $(\overline{\varepsilon}_{\rm a},\Theta)$ pairs after an ion collision at different normalised positions, $\overline{\xi}_{\rm c}$, (see text) within the sheath. The arrow shows the direction of an increasing $\chi_{\rm lab}$ from 0$^\circ$ to 90$^\circ$ along these lines.[]{data-label="fig:mushroom"}](mushroom.pdf){width="50.00000%"}
The “mushroom” shape of the $F(\varepsilon,\Theta)$ distribution at low pressure (e.g. 5 Pa, figure \[fig:singlefreq1\](c)) can be understood based on an analytical model of the motion of O$_2^+$ ions. In this model we assume that each ion undergoes [*exactly one collision*]{} within the sheath, at a [*random position*]{}. The energy of the ions entering the sheath is assumed to be zero. As the ion transit time through the sheath is significantly longer than the RF period, ions can be considered to sense the time-averaged electric field. The latter is taken to be linear with respect to the position within the sheath. We consider the grounded sheath and fix the zero of our coordinate system ($\xi=0$) at the position of the maximum sheath width. At this point we also set the electric field and the potential to zero, $E = 0$ V/m, $\Phi=0$ V. Under these conditions the electric field and the potential within the sheath vary as: $$\begin{aligned}
E(\xi) = -\frac{2 \phi_{\rm s}}{s^2} \, \xi= -\frac{2 \phi_{\rm s}}{s} \,\overline{\xi} \\
\Phi(\xi) = \Phi_{\rm s} \biggl( \frac{\xi}{s} \biggr)^2 = \Phi_{\rm s} \, \overline{\xi}^2,\end{aligned}$$ where $\Phi_{\rm s}$ is the time-averaged sheath voltage, $s$ is the maximum width of the sheath, and $\overline{\xi}= \xi / s$ is the normalised spatial coordinate within the sheath. If the ion undergoes a collision at $\overline{\xi}_{\rm c}$, the pre- and post-collision energies are, respectively, $$\begin{aligned}
\varepsilon_{\rm c} = Q \, \Phi_{\rm s} \, \overline{\xi}^2 \\
\varepsilon_{\rm s} = \varepsilon_{\rm c} - \Delta \varepsilon = \varepsilon_{\rm c} \cos^2 \chi_{\rm lab} = Q \, \Phi_{\rm s} \,\overline{\xi}^2 \cos^2 \chi_{\rm lab},\end{aligned}$$ as the relative change of the energy is $\Delta \varepsilon / \varepsilon_{\rm c} = (1- \cos \chi_{\rm com}) /2$ and $\chi_{\rm lab} = \chi_{\rm com}/2$, because the masses of the collision partners are the same. Here $\chi_{\rm com}$ and $\chi_{\rm lab}$ are the scattering angles in the center-of-mass (COM) and laboratory (LAB) frames, respectively, and $Q$ is the charge of the ion. The energy of the ion upon arrival at the electrode is: $$\varepsilon_{\rm a} = \varepsilon_{\rm s} + Q \, \bigl[ \Phi_{\rm s} - \Phi(\overline{\xi}_{\rm c}) \bigr] = Q \, \Phi_{\rm s} \bigl[ 1 - \overline{\xi}_{\rm c}^2 (1 - \cos^2 \chi_{\rm lab}) \bigr].$$ Introducing $\overline{\varepsilon}_{\rm a}$ as the arrival energy normalised by its maximum possible value $\varepsilon_{{\rm a,max}} = Q \, \Phi_{\rm s}$, we obtain $$\overline{\varepsilon}_{\rm a} = 1 -\overline{\xi}_{\rm c}^2 (1 - \cos^2 \chi_{\rm lab}).
\label{eq:model_energy}$$ The angle of incidence can be computed from the perpendicular component of the post-collision velocity (that does not change along the post-collision trajectory), $v_{{\rm s},\perp}$, and the magnitude of the velocity upon arrival, $v_{\rm a}$, as $$\sin \Theta = \frac{v_{{\rm s},\perp}}{v_{\rm a}} = \frac{\overline{\xi}_{\rm c} \sin (2 \chi_{\rm lab})}{2 \sqrt{\overline{\varepsilon_{\rm a}}}}.
\label{eq:model_angle}$$
![(a) The effect of the SEEC $\gamma$ on the flux-energy distribution of O$_2^+$ ions, at $p$ = 10 Pa ($f_1=$ 27.12 MHz, $\phi_{\rm pp}$ = 400 V). The inset shows the time dependence of the width of the powered sheath, $s_{\rm p}$. (b) Time dependence of the sheath voltages ($\phi_{\rm sp}$, $\phi_{\rm sg}$) and their sum for $\gamma=0$, the applied voltage waveform ($\phi$), the bulk voltage drop ($\phi_{\rm bulk}$) for the $\gamma=0$ and $\gamma=0.4$ cases, and the difference between the voltage drops over the powered sheath between the two cases of $\gamma=0$ and $\gamma=0.4$ ($\Delta \phi_{\rm sp}$). Note that the $\phi_{\rm bulk}$ and $\Delta \phi_{\rm sp}$ values are multiplied by a factor of 10. $T_1$ is the period of the driving frequency.[]{data-label="fig:single_idf_pot"}](single_idf_pot.pdf){width="45.00000%"}
Equations (\[eq:model\_energy\]) and (\[eq:model\_angle\]) give the normalised arrival energy and the incidence angle at the electrode for any position of collision, $\overline{\xi}_{\rm c}$, and laboratory scattering angle, $\chi_{\rm lab}$. The $F(\overline{\varepsilon}_{\rm a},\Theta)$ distribution can be simulated by generating a high number of random samples of these two variables. The result of such a simulation is shown in figure \[fig:mushroom\]. We note that in this simulation the collisions were executed by the same approach as in the PIC/MCC simulations: charge exchange (backward scattering in the COM system) was supposed to occur with a probability of 2/3 and isotropic scattering (in the COM system) was supposed to occur with a probability of 1/3. The results of the computation resemble closely the PIC/MCC simulation result shown in figure \[fig:singlefreq1\], despite the very simple nature of the model. The formation of the main feature seen in the $F(\overline{\varepsilon}_{\rm a},\Theta)$ colour map of figure \[fig:mushroom\] can be understood by considering scattering events at specific $\overline{\xi}_{\rm c}$ positions. Figure \[fig:mushroom\] shows sets of $(\overline{\varepsilon},\Theta)$ pairs (forming lines) at fixed values of $\overline{\xi}_{\rm c}$. These lines were generated by scanning $\chi_{\rm lab}$ between 0$^\circ$ and 90$^\circ$. The $F(\overline{\varepsilon}_{\rm a},\Theta)$ distribution is nothing else but a superposition of an infinite number of such sets having infinitesimal widths. The model, on the other hand, cannot reproduce the low-energy part of the distribution obtained in the PIC/MCC simulation, which is a result of multiple collision (isotropic scattering or charge exchange) events.
The effect of the SEEC $\gamma$ on the flux-energy distribution of the O$_2^+$ ions is illustrated in figure \[fig:single\_idf\_pot\](a) by PIC/MCC results. One can observe several differences between the two curves, corresponding to $\gamma$ = 0 and $\gamma=0.4$. First, the integral of $F(\varepsilon)$ increases as $\gamma$ is increased. This is attributed to an enhanced plasma density, that, in turn, results in shorter sheaths and lower collisionality, in the presence of appreciable secondary electron emission. (Note that while figure \[fig:single\_idf\_pot\](a) presents the $F(\varepsilon)$ data in arbitrary units, the data are proportional to the real fluxes in the two cases). Second, $F(\varepsilon)$ exhibits fewer peaks at higher $\gamma$, which is the consequence of the fewer number of charge exchange collisions under such conditions, caused by the decrease of the sheath widths; see the time dependence of the width of the powered sheath, $s_{\rm p}$ in the inset of figure \[fig:single\_idf\_pot\](a).
The third difference between the $F(\varepsilon)$ curves computed at $\gamma$ = 0 and 0.4 is that the maximum ion energy decreases and the dominant peak of $F(\varepsilon)$ shifts towards lower energies by $\approx$ 5 eV at $\gamma=0.4$. The origin of this shift is analysed in figure \[fig:single\_idf\_pot\](b) that shows several voltage components: the voltage drops over the powered and the grounded sheaths ($\phi_{\rm sp}$ and $\phi_{\rm sg}$, respectively) and the sum of them for $\gamma=0$, the applied voltage waveform ($\phi$), as well as the bulk voltage drop ($\phi_{\rm bulk}$) for $\gamma=0$ and $\gamma=0.4$. (These voltages can be determined in a straightforward way when the sheath edge positions, $s_{\rm p}$ and $s_{\rm g}$ are known.) The figure also shows the difference of the voltage drops over the powered sheath in the two cases with $\gamma=0$ and $\gamma=0.4$, $\Delta \phi_{\rm sp} = \phi_{\rm sp,\gamma=0} - \phi_{\rm sp,\gamma=0.4}$. Note that by definition $\phi_{\rm sp} < 0$ [@optim1].
The data reveal that $\Delta \phi_{\rm sp}(t)$ is negative at all times, i.e. the magnitude of the sheath voltage gets lower when $\gamma$ is increased. The temporal average of $| \Delta \phi_{\rm sp} |$ is $\approx$ 5 V, which explains the shift of the high-energy part of the $F(\varepsilon)$ distribution. Note that while we discussed the case of the powered sheath, the same arguments also hold for the grounded sheath. It is also to be mentioned that the sheath voltages do not add up to the applied voltage waveform, as a significant voltage drop over the bulk plasma $\phi_{\rm bulk} = \phi - (\phi_{\rm sp} +\phi_{\rm sg})$ forms, characteristic for electronegative plasmas. This voltage drop is actually, found to be responsible for the change of $\phi_{\rm sp}$, as the peak value of $| \phi_{\rm bulk} (t) |$ of $\approx$ 10 V at $\gamma=0$ decreases to $\approx$ 5 V at $\gamma=0.4$. This change is in turn, caused by the decreasing electronegativity with increasing $\gamma$: in the presence of an additional (surface) supply of electrons the conductivity of the bulk plasma increases and this requires a lower $\phi_{\rm bulk}$. As a secondary reason the decrease of the floating potential (by about 1.5 V) with increasing $\gamma$ can be identified that also leads to a decrease of the sheath voltage drop that is available to accelerate the positive ions. The effect of the SEEC on the “cutoff ion energy” is similar at other pressures, too. For both 5 Pa and 20 Pa we find lowering of the energy limit by a similar magnitude as was found and discussed for the case of 10 Pa.
Classical dual-frequency excitation {#sec:dual}
-----------------------------------
Next, we address the properties of oxygen CCPs under classical DF excitation as specified by eq.(\[eq:exc2\]). Results will be presented for the high-frequency voltage amplitude $\phi_1$ set to 200 V and using a low-frequency voltage amplitude between $\phi_2$ = 0 V and 300 V. For some conditions the range of $\phi_2$ was limited by the adverse effect of $\phi_2$ on the charged particle densities, especially at low values of the SEEC.
![The effect of the low-frequency voltage amplitude ($\phi_2$) on the density of O$_2^+$ ions in classical dual-frequency discharges at 5 Pa (left column) and 20 Pa (right column), at the $\gamma$ values specified. $\phi_1$ = 200 V.[]{data-label="fig:DF_densities"}](DF_densities.pdf){width="50.00000%"}
![The spatio-temporal variation of the mean electron energy in the vicinity of the powered electrode for $p$ = 20 Pa, $\phi_1$ = 200 V, and $\gamma=0.4$, at (a) $\phi_2$ = 0 V and (b) $\phi_2$ = 200 V. $T_2$ is the period of the low-frequency excitation.[]{data-label="fig:meane-xt"}](meane-xt.pdf){width="50.00000%"}
![The electronegativity, $\beta$, of the plasma (a) and the mean electron energy, $\overline{\varepsilon}_{\rm e}$, in the centre of the bulk (b), for classical dual-frequency excitation, as a function of $\phi_2$ for different pressures and SEEC values. $\phi_1$ = 200 V for all cases. The dashed green horizontal line in (b) indicates the energy (6.5 eV) where the attachment cross section peaks.[]{data-label="fig:DF_electronegativity"}](DF_electronegativity.pdf){width="40.00000%"}
The spatial distribution of the density of O$_2^+$ ions is shown in figure \[fig:DF\_densities\], for different values of the gas pressure and the SEEC ($\gamma$). Each panel compares the distributions obtained at $\phi_2$ = 0 V and 200 V. At low pressure (5 Pa) the presence of the low-frequency component in the driving waveform results in a significant decrease of the ion density at zero / low $\gamma$ values (figures \[fig:DF\_densities\](a),(b)). This is a consequence of the widening of the sheath due to the higher voltage, which, on the other hand, does not contribute significantly to higher ionisation because of its low frequency, i.e. the frequency coupling mechanism [@coupling1; @coupling2]. At the highest value of the SEEC, $\gamma=0.4$, the enhanced sheath voltage contributes to the ionisation as it accelerates secondary electrons to high energies, which can then ionise as well. This effect almost exactly compensates the effect described above, leaving the ion density at the same value as in the case of $\phi_2$ = 0 V (see figure \[fig:DF\_densities\](c)). At 20 Pa pressure, as shown in figures \[fig:DF\_densities\](d) and (e) for the lower $\gamma$ values, the sheath widths also slightly decrease and the ion density profiles become narrower and the peak ion density somewhat increases when $\phi_2$ = 200 V is applied.
Comparing the cases shown in figures \[fig:DF\_densities\](a) and (d) we observe an opposite effect of $\phi_2$ on the ion density at 5 Pa and 20 Pa. The application of the low-frequency voltage increases the sheath width in both cases, and this results in the reduction of the space available for the electrons to dissipate their energy acquired at sheath expansion. At 20 Pa the electrons dissipate all their energy in the bulk, whereas this is not the case at 5 Pa, where the electron mean free path is significantly longer.
At high $\gamma$ (figure \[fig:DF\_densities\](f)) a significant enhancement of the ion density is observed due to the effect of an efficient multiplication and an enhanced mean energy of the secondary electrons. The behaviour of the latter quantity is shown in figure \[fig:meane-xt\], for 20 Pa pressure, $\gamma$ = 0.4, for $\phi_2$ = 0 V (panel (a)) and $\phi_2$ = 200 V (panel (b)). These results indicate for $\phi_2$ = 200 V a significant ($\sim$ factor of two) increase of the maximum value of the mean electron energy in space and time, with respect to the $\phi_2$ = 0 V case, when both $\phi_1$ and $\phi_2$ take their maximum values simultaneously in time.
The specific driving voltage waveform has a significant effect as well on the electronegativity ($\beta$) of the plasma. The results are shown in figure \[fig:DF\_electronegativity\](a), for different pressures and SEEC values. In the case of $\gamma=0$, at low pressure (5 Pa) there is a remarkable decrease of the electronegativity with increasing $\phi_2$, whereas, for the higher pressures (10 Pa and 20 Pa) the opposite trend is observed: the increasing $\phi_2$ results in higher electronegativity. This behaviour of $\beta$ is discussed together with that of the mean electron energy in the centre of the plasma, $\overline{\varepsilon}_{\rm e}$, shown in figure \[fig:DF\_electronegativity\](b). At $\phi_2$ = 0 V, there is a strong correlation between a higher electronegativity and a higher mean electron energy, due to the reasons explained in the case of single-frequency discharges. When $\phi_2$ is increased, the same correlation is maintained at $p$ = 10 Pa, both the electronegativity and the mean electron energy increase with $\phi_2$ at zero $\gamma$ and decrease at high $\gamma$. At 20 Pa pressure the increase of the mean energy is less significant than the increase of the electronegativity, and at 5 Pa an opposite trend is observed at $\gamma=0$. The dependence of these observed trends on the pressure can be understood by noting that the electron attachment cross sections peaks at $\cong$ 6.5 eV, indicated by a horizontal line in the plot showing the mean electron energy. (This cross section has a sharp maximum around this value and has another broad peak above 20eV, with a smaller magnitude [@Vahedi]. This second feature of the cross section is not expected to influence our discussion.)
The application of the low-frequency voltage results in an increase of $\overline{\varepsilon}_{\rm e}$ at $\gamma=0$ (see figure \[fig:DF\_electronegativity\](b)). When the mean energy is lower than the energy corresponding to the peak of the electron attachment cross section, the attachment rate and, thus, the electronegativity increases with $\phi_2$. This is the scenario at 10 Pa and 20 Pa pressures. Here, the increased attachment rate results in an increased electric field within the bulk plasma due to the depleted electron density, that in turn, increases $\overline{\varepsilon}_{\rm e}$, i.e. there is a positive feedback mechanism leading to the establishment of the observed operating conditions. At 5 Pa, however, the mean energy $\overline{\varepsilon}_{\rm e}$ is near 6.5 eV at $\phi_2 =0$ V and its increase at $\phi_2 > 0$ V decreases the efficiency of the attachment and results in a decrease of the electronegativity, as seen in figure \[fig:DF\_electronegativity\](a).
The non-monotonic behaviour of the electronegativity at $p$ = 10 Pa and 20 Pa with increasing $\phi_2$ at $\gamma=0$ is a result of two competing mechanisms. (i) As the mean electron energy approaches the value corresponding to the optimum of negative ion formation, the negative ion density increases in the bulk plasma. (ii) At the same time, also caused by the increase of the low-frequency voltage, the width of the bulk decreases, this way limiting the domain where negative ions can accumulate. As the electronegativity is defined “globally”, i.e. as the ratio of the [*spatially averaged*]{} negative ion density and the [*spatially averaged*]{} electron density, the interplay of these two effects causes a transition, in our case at around e.g., $\phi_2 \approx 200$ V at 20 Pa.
At the high SEEC value of $\gamma=0.4$ both $\beta$ and $\overline{\varepsilon}_{\rm e}$ exhibit only a weak dependence on $\phi_2$, because of the significant additional supply of electrons from the electrodes, which are accelerated to high energies and are collisionally multiplied within the sheaths.
![Flux-energy distributions of O$_2^+$ ions at the electrodes of oxygen discharges driven by classical dual-frequency waveforms, at different pressures: 5 Pa (a), 10 Pa (b), and 20 Pa (c). Each curve is normalised to a maximum value of 1.0. $\phi_1$ = 200 V and $\gamma=0$ for all cases.[]{data-label="fig:DF_ifedf"}](DF_ifedf.pdf){width="40.00000%"}
The flux-energy distributions of O$_2^+$ ions at the electrodes, as a function of $\phi_2$ are displayed in figure \[fig:DF\_ifedf\] for the pressures of 5 Pa (a), 10 Pa (b), and 20 Pa (c), for $\gamma=0$. (As the discharge is symmetrical, these distributions are the same at both electrodes.) The results indicate that for each pressure the range of ion energies is expanded as the low-frequency voltage amplitude grows. The influence of $\phi_2$ on the sheath widths (cf. figure \[fig:DF\_densities\]) and consequently, on the collisionality, also influences the shapes of the IFEDFs at low pressures. At 5 Pa pressure, e.g., while the IFEDF is peaked near the maximum energy when $\phi_2$ = 0 V, the energy spectrum exhibits from $\sim$100 eV onwards a fall with increasing energy, as the ions have little probability to cross the enlarged sheath without charge exchange collisions. At 20 Pa pressure the energy range is expanded without a major change of the shape of the spectrum. A notable difference is, however, the disappearance of the distinct peaks of the IFEDF, which are only present at $\phi_2$ = 0 V.
The peaks in the IFEDFs can form as a consequence of charge exchange collisions (forming slow ions) in the region between the electrode and the maximum sheath width. When the electric field is zero for considerable periods of time within this region, these “cold” ions can accumulate and are accelerated together during the next high-field periods. While in a single-frequency discharge (i.e. $\phi_2$ = 0 V) the electric field exhibits zero values within this region for considerable time periods (several tens of nanoseconds), whenever $\phi_2 \neq$ 0 V, the sheath electric field lacks these intervals, making it impossible for the ions to aggregate and be accelerated in a synchronised manner.
This phenomenon is analysed in more details based on the model of the ion motion in the sheaths, described Schuengel [*et al.*]{} [@Eddi]. This model is based on approximations of the ion density profile in the sheath as given in [@Wild] and of the temporal evolution of the sheath voltage from an equivalent circuit model [@EAE-control3]. It allows the calculation of the ion motion in the sheaths and the IFEDFs from the set of input data that consists of the applied voltage waveform (specified for the given case), the maximum sheath width (which is taken as a result of the PIC/MCC simulations and is slightly corrected in the model, for details see [@Eddi]), and the ion mean free path (that can be computed from the cross sections). By computing ion trajectories, the energy of the ions upon the arrival at the electrodes as well as the ion transit time can be obtained for ions starting from arbitrary positions. Based on these, the structure of the IFEDFs was successfully explained in [@Eddi] for various discharge conditions. Here, we make use of the same model and carry out calculations to explain the characteristic changes of the IFEDFs as an effect of varying $\phi_2$ in classical DF discharges operated in oxygen.
![(a) Energy of those ions arriving at the electrode, which experienced a charge exchange collision within the sheath at a normalised position $x/s_{\rm max}$ and at a normalised time $t/T_2$ and arrive at the electrode without further collisions and (b) the time that these ions need to arrive at the electrode, $\overline{\tau} = t / T_1$, for $\phi_2$ = 0 V. (c) and (d) are zoomed regions of (a) and (b). (e) and (f) are the same as (a) and (b), but for $\phi_2$ = 100 V. (g) and (h) show the IFEDFs obtained from the model in comparison with the PIC/MCC simulation results for $\phi_2$ = 0 V and $\phi_2$ = 100 V, respectively. $s_{\rm max}$ is the maximum sheath width, $T_1$ is the period of the high-frequency component and $T_2$ is the period of the low-frequency component of the excitation waveform. $\phi_1$ = 200 V and $p$ = 20 Pa for all cases.[]{data-label="fig:DF_EDDI"}](DF_EDDI.pdf){width="62.00000%"}
![Mean energy, $\langle \varepsilon \rangle$, (a) and flux, $\Gamma$, (b) of O$_2^+$ ions at the electrodes, as a function of the low-frequency voltage amplitude, at different pressure and $\gamma$ values specified. $\phi_1$ = 200 V for all cases.[]{data-label="fig:DF_results"}](DF_results.pdf){width="40.00000%"}
Figure \[fig:DF\_EDDI\](a) shows the energy of those ions arriving at the electrode, which experienced a charge exchange collision within the sheath at a normalised position $x/s_{\rm max}$ (where $s_{\rm max}$ is the maximum sheath width) and at a normalised time $t/T_2$ (where $T_2$ is the period of the low-frequency component of the excitation waveform) and arrive at the electrode without further collisions, in the case $\phi_2$ = 0 V. Panel (b) shows the time that these ions need to arrive at the electrode, $\overline{\tau}$, that is normalised by $T_1$ (where $T_1$ is the period of the high-frequency component of the excitation waveform). These plots, as well as panels (c) and (d), which are zoomed parts of panels (a) and (b), confined to narrower domains of space and time, indicate that there are certain zones in space and time, where ions, that have undergone a charge exchange collision, can accumulate. For these domains the same ion energy at the electrode is achieved. As an example, the peaks in the IFEDF shown in figure \[fig:DF\_EDDI\](g) at about 21 eV and 29 eV (marked with arrows in the plot) originate from ions that were “born” in $x/s_{\rm max}$ of about 0.14 and 0.22, respectively (blue and yellow regions in panel (c)). These ions need 4 and 5 periods of the 27.12 MHz cycle, respectively, to reach the electrode (as inferred from figure \[fig:DF\_EDDI\](d). The IFEDF constructed from the model this way (see figure \[fig:DF\_EDDI\](g)) matches extremely well the distribution obtained from the PIC/MCC simulation.
The energy of the arriving ions and the time needed to reach the electrode following a charge exchange collision is displayed, respectively, in figures \[fig:DF\_EDDI\](e) and (f), for $\phi_2$ = 100 V, while the resulting IFEDF is depicted in figure \[fig:DF\_EDDI\](h). At these conditions there are no clear zones in space and time, where ions that have undergone a charge exchange collision may accumulate. Ions that are born at the same spatial position but at different times are accelerated differently due to the presence of the low-frequency field. Therefore, they reach the electrode with different energies. The model again predicts an IFEDF that is in very good agreement with the one obtained from the PIC/MCC simulation, and explains the changes of the IFEDF due to a change of $\phi_2$ in an elementary way.
According to figure \[fig:DF\_ifedf\] the effect of $\phi_2$ on the IFEDFs is appreciable at all pressures. This shows the possibility of the control of the mean ion energy by the $\phi_2$ amplitude. This voltage amplitude should, in principle, have a negligible effect on the flux of the ions. Previous studies of the independent control of ion properties have, however, shown that this is violated due to (i) frequency coupling effects and (ii) the influence of secondary electron emission from the electrodes. The first effect results from the fact that the sheath dynamics is determined by both voltage amplitudes and as an increasing low-frequency voltage expands the sheaths, in general, the high-frequency sheath width oscillations take place in the domain of higher ion densities, thereby decreasing the modulation of the sheath width and the speed of expansion [@coupling1; @coupling2]. The second effect (the influence of secondary electrons) results from the higher total accelerating voltage which may give rise to significantly higher multiplication of secondary electrons creating a higher plasma density and higher ion flux [@secondaries1; @secondaries2]. The results of our investigation of the performance of the separate control of the mean ion energy and ion flux are presented in figure \[fig:DF\_results\]. Panel (a) shows the mean ion energy, $\langle \varepsilon \rangle$, while panel (b) shows the O$_2^+$ ion flux, $\Gamma$, as a function of the low-frequency voltage amplitude. As the discharge is symmetrical, these quantities are the same at both electrodes. The data reveal that the value of $\phi_2$ controls well the mean ion energy in all cases, however, for some cases this is accompanied by a change of the ion flux. Figure \[fig:DF\_results\](b) shows that this effect is most critical at the highest value of the SEEC, $\gamma=0.4$. In this case the independent control is impossible at the higher pressures (10 Pa and 20 Pa). In the other parameter combinations the flux remains reasonably constant, providing a way of controlling the mean ion energy (closely) independently of the ion flux $\Gamma$.
Valleys-waveforms {#sec:valleys}
-----------------
In the following we present simulation results for multi-frequency waveforms, defined by eq.(\[eq:exc3\]), with a base frequency of $f_1$ = 15 MHz and including harmonics up to $N$ = 4. This type of waveform was found in previous studies to be efficient for an independent control of the ion properties [@Derzsi2016; @EAE-control1; @EAE-control2; @EAE-control3]. The basis of this control is the self-bias voltage that develops (even in geometrically symmetrical reactors) due to the amplitude asymmetry of this waveform. The control parameters, which determine the value of the DC self-bias (at fixed $\phi_k$ harmonic amplitudes) are the $\theta_k$ phase angles in eq.(\[eq:exc3\]). As mentioned in section 2.1, the value of the DC self-bias voltage is determined in an iterative manner in the simulations, to ensure equal losses of positive and negative charges at each electrode over one period of the fundamental driving frequency [@DZ-EAE], i.e. the self-bias voltage is not pre-defined for the given waveform, but it is self-consistently calculated.
Below, examples will be given for valleys-type waveforms that give rise to a positive self-bias. Therefore, the ion energy range is expected to be extended at the grounded electrode due to the increase of the time-averaged sheath voltage drop at that side of the plasma. In contrast, a decrease of the mean ion energy at the powered electrode is contemplated with increasing self-bias.
We do not carry out a variation of the phase angles, the presentation of the results is restricted to phases $\theta_k =0$ for each odd harmonic and $\theta_k =\pi$ for each even harmonic, specific for [*valleys-type*]{} waveforms. The possibility of the control of the mean ion energy is, however, still revealed at this choice of the phase angles, as the self-bias voltage takes its extremum value near $\theta_k =0$ [@DZ-EAE]. Thus, the mean ion energies at the two electrodes are (very nearly) maximum and minimum values, i.e. the range between them represents the control range for the mean ion energy.
![Self-bias voltage, $\eta$, (a), electronegativity, $\beta$, (b), and (temporally averaged) mean electron energy, $\overline{\varepsilon}_{\rm e}$, in the discharge centre (c) in oxygen discharges driven by valleys-type waveforms (specified by eq.(\[eq:exc3\])) consisting of $N$ harmonics, at different pressures and $\gamma$ values. The dashed horizontal line in (c) indicates the energy (6.5 eV) where the attachment cross section peaks. $\phi_{\rm pp}$ = 400 V for all cases.[]{data-label="fig:valleys_bias"}](valleys_bias.pdf){width="40.00000%"}
![Spatio-temporal distribution of the ionisation rate in oxygen discharges driven by valleys-type waveforms (specified by eq.(\[eq:exc3\])) consisting of $N=1$ (a) and $N=4$ (b) harmonics. $\phi_{\rm pp}$ = 400 V, $p$ = 20 Pa and $\gamma=0$ for both cases. $T_1$ is the period of the fundamental frequency ($f_1$).[]{data-label="fig:valeys_ion_rate"}](valeys_ion_rate.pdf){width="40.00000%"}
First, we present the PIC/MCC simulation results for the computed DC self-bias voltage ($\eta$) and the electronegativity ($\beta$) of the plasma, in figure \[fig:valleys\_bias\](a) and (b), respectively. The data shown in figure \[fig:valleys\_bias\](a) confirm that a significant DC self-bias voltage is established when the waveform includes an increasing number of harmonics. Accompanied by this change is a decrease of the electronegativity, as seen in figure \[fig:valleys\_bias\](b). The latter effect is most pronounced at 20 Pa. The change of $\beta$ strongly correlates with the change of the (temporally averaged) mean electron energy at the discharge centre, shown in figure \[fig:valleys\_bias\](c). In the case of valleys waveforms the observed decrease of the mean electron energy ($\overline{\varepsilon}_{\rm e}$) as a function of $N$ is caused by an electron power absorption mode transition, as shown in figure \[fig:valeys\_ion\_rate\] that compares the spatio-temporal distribution of the ionisation rates obtained for $N=1$ and $N=4$ (at $\phi_{\rm pp}$ = 400 V, $p$ = 20 Pa and $\gamma=0$). At $N=1$ a hybrid $\alpha$ + DA power absorption mode is revealed, with significant ionisation within the plasma bulk region. This is the consequence of the high electric field within that domain under the conditions of high electronegativity (see figure \[fig:valleys\_bias\](b)) i.e., strongly depleted electron density. When $N$ is changed to 4, a pure $\alpha$-mode electron power absorption establishes, ionisation is concentrated near the expanding sheath edges, with a strong dominance of the region near the grounded electrode, due to the strong positive self-bias at these conditions (see figure \[fig:valleys\_bias\](a)). The data indicate a $\approx$ 15 times higher peak ionisation rate in the case of $N$ = 4, compared to the case of $N$ = 1, while the spatio-temporal average of these also shows an increase by a factor of $\approx$ 2.5. Increasing $N$ enhances the sheath expansion heating of electrons. This corresponds to an enhanced source of energetic electrons, which explains the decrease of the mean electron energy observed in figure \[fig:valleys\_bias\](c). Similar to the classical dual-frequency scenario, a change of the mean electron energy is coupled to a change of the electronegativity ($\beta$) via the energy dependence of the attachment cross section (see section \[sec:dual\]).
![Mean energy of O$_2^+$ ions at the powered electroed (PE) and at the grounded electrode (GE) of oxygen discharges driven by valleys-type waveforms (specified by eq.(\[eq:exc3\])) consisting of $N$ harmonics, at different pressures. $\phi_{\rm pp}$ = 400 V and $\gamma=0$ for all cases. The arrows in the case of 5 Pa show the control range of the mean ion energy that can be covered by changing the phase angles in the driving voltage waveform. []{data-label="fig:valleys_flux_energy"}](valleys_energy.pdf){width="40.00000%"}
The change of the DC self-bias with increasing number of harmonics ($N$) is expected to change the mean energy of ions arriving at the electrodes. This effect is confirmed by the results shown for the mean ion energy in figure \[fig:valleys\_flux\_energy\]. We note that at 2 Pa pressure it was not possible to ignite the plasma with a single harmonic, $N=1$. Taking the $p$ = 5 Pa case as an example, the data show that using the highest number of harmonics, the maximum and minimum mean ion energies are 130 eV and 30 eV, respectively, i.e., $\langle \varepsilon \rangle$ can be changed by more than a factor of 4 at the same conditions by changing the phase angles of the driving voltage waveform. The mean ion energy is insensitive to the value of $\gamma$.
![Flux-energy distributions of O$_2^+$ ions at the grounded electrode of oxygen discharges driven by valleys-type waveforms (specified by eq.(\[eq:exc3\])) consisting of $N$ harmonics, obtained at different pressures: 2 Pa (a), 5 Pa (b), and 20 Pa (c). Each curve is normalised to a maximum value of 1.0. $\phi_{\rm pp}$ = 400 V and $\gamma=0$.[]{data-label="fig:valleys_iedfs"}](valleys_iedfs.pdf){width="40.00000%"}
The flux-energy distribution function of the O$_2^+$ ions at the grounded electrode of the CCP are shown in figure \[fig:valleys\_iedfs\] for different pressures, as a function of the number of harmonics in the valleys-type excitation waveform. The range of ion energy is significantly extended at all pressures when the number of applied harmonics is increased. We observe, however, a significant difference of the shape of the IFEDF compared to that obtained for the classical dual-frequency excitation, compare, e.g., the 5 Pa cases (figure \[fig:DF\_ifedf\](a) vs. figure \[fig:valleys\_iedfs\](b)). This difference originates from the fact that while in the case of the classical DF excitation the low-frequency voltage strongly increases the sheath width, this is not the case when the valleys-type waveform is applied. Therefore, these two types of excitation result in noteworthily different IFEDFs. Enhancing the ion energies without increasing the collisionality of the sheaths is definitely advantageous when a well-directed beam of ions is necessary for surface processing. The valleys-type (or peaks-type) waveforms provide this possibility, as figure \[fig:valleys\_iedfs\] confirms. For the lowest pressure case a nearly mono-energetic beam of ions arrives at the electrode for each $N$. More details about these cases (at 2 Pa pressure) are given in figure \[fig:valleys\_angular\], which shows the $F(\varepsilon,\Theta)$ combined, energy and angularly-resolved distribution function.
![Energy and angular distribution of the O$_2^+$ ion flux, $F(\varepsilon,\Theta)$ (in arbitrary units) at the grounded electrode in discharges driven by valleys-type waveforms with $N$ = 2 (a) and $N$ = 4 (b) harmonics, at 2 Pa pressure. The insets illustrate the energy and angular distributions (respective integrals of $F(\varepsilon,\Theta)$ according to incidence angle and energy). All distributions have been normalised to a maximum value of 1, for easier comparison. $\phi_{\rm pp}$ = 400 V, $\gamma=0$.[]{data-label="fig:valleys_angular"}](valleys_angular.pdf){width="40.00000%"}
![Comparison of flux energy (a) and angular (b) distributions of the O$_2^+$ ions (in arbitrary units) at the grounded electrode in discharges driven by a DF waveform ($\phi_1 = \phi_2$ = 200 V) at 5 Pa, a valleys-waveform with $N$ = 3 at 5 Pa, and a valleys-waveform with $N$ = 4 at 2 Pa. These distributions have been normalised to peak values of 1, for an easier comparison of their [*shapes*]{}. $\gamma=0$.[]{data-label="fig:comp_angular"}](comp_angular.pdf){width="40.00000%"}
Figure \[fig:comp\_angular\] compares the IFEDFs and IADFs obtained with DF and valleys waveforms. The DF excitation at 5 Pa and $\phi_1 = \phi_2$ = 200 V and the valleys-type excitation with 3 harmonics, at 5 Pa and $\phi_{\rm pp}$ = 400 V, form a pair of cases in the sense that the maximum ion energy is very nearly the same. Significantly different flux-energy distributions are obtained for these cases as already mentioned above. The difference between the sheath widths in these two cases, that causes the different shapes of the IFEDFs, also results in different angular distributions of the arriving ions; in the case of valleys-type excitation the distribution shifts to lower values of $\Theta$. A further narrowing of the $F(\Theta)$ distribution can be achieved by operating the plasma at a lower pressure, with a higher number of harmonics, as it is illustrated in figure \[fig:comp\_angular\] for valleys-type excitation with $N$ = 4 harmonics, at 2 Pa and $\phi_{\rm pp}$ = 400 V.
Sawtooth-waveforms
------------------
![Self-bias voltage, $\eta$, (a) and electronegativity, $\beta$, (b) in oxygen discharges driven by sawtooth-type waveforms (specified by eq.(\[eq:exc4\])) consisting of $N$ harmonics, at different pressures. $\phi_{\rm pp}$ = 400 V for all cases.[]{data-label="fig:sawtooth_bias"}](sawtooth_bias.pdf){width="38.00000%"}
![Spatio-temporal distribution of the ionisation rate in oxygen discharges driven by sawtooth-down waveforms (specified by eq.(\[eq:exc4\])) consisting of $N$ = 4 harmonics, for $p$ = 2 Pa (a) and $p$ = 20 Pa (b) pressures. $T_1$ is the period of the fundamental frequency ($f_1$). $\phi_{\rm pp}$ = 400 V and $\gamma=0$ for both cases.[]{data-label="fig:sawtooth_ionisation"}](sawtooth_ionisation.pdf){width="42.00000%"}
![Mean energy of the O$_2^+$ ions at the powered electrode (PE, filled symbols) and at the grounded electrode (GE, open symbols) of oxygen discharges driven by sawtooth-down waveforms (specified by eq.(\[eq:exc4\])) consisting of $N$ harmonics, at different pressures. $\phi_{\rm pp}$ = 400 V and $\gamma=0$ for all cases.[]{data-label="fig:sawtooth_energy_flux"}](sawtooth_energy.pdf){width="40.00000%"}
Finally, we address the properties of CCPs driven by sawtooth-waveforms. We consider [*sawtooth-down*]{} waveforms resulting from a plus sign in eq.(\[eq:exc4\]). Unlike in the case of valleys-type (or peaks-type) waveforms that exhibit different positive and negative extrema (“amplitude asymmetry effect”), the sawtooth-type waveforms have equal positive and negative extrema. The difference of the rise and fall times of the waveform, however, still establishes an asymmetry of the discharge dynamics (“slope asymmetry”) that can give rise to a self-bias voltage, different sheath properties and IFEDFs at the two electrodes, spatial distributions of the excitation rates, etc. (The asymmetry of the discharge depends strongly on the type of the buffer gas used; discharges in Ar, CF$_4$, and H$_2$ gases have been studied and completely different excitation dynamics were found at the same driving voltage waveforms [@Bastien; @Bastien2].) In the present study, the number of harmonics ranges between $N=1$ and $N$ = 4. We find that the behaviour of the self-bias for oxygen gas is rather specific, as it can be seen in figure \[fig:sawtooth\_bias\](a). At low pressures (2 Pa and 5 Pa) a small positive self-bias voltage is generated when the number of harmonics is increased. In contrast with this, a negative self-bias appears at the high pressure (20 Pa) case, with a significantly higher magnitude compared to that in the low-pressure cases. The electronegativity of the plasma decreases with increasing number of harmonics, similarly to the behaviour observed in the case of the valleys-type waveforms. The strongest dependence is also observed here at the highest pressure. These changes can be understood by analysing the spatio-temporal distributions of the ionisation rate, which are shown in figure \[fig:sawtooth\_ionisation\] for $p$ = 2 Pa and $p$ = 20 Pa pressures, at $N$ = 4 harmonics, $\phi_{\rm pp}$ = 400 V and $\gamma=0$. The change of the gas pressure induces a change of the electron power absorption mode. At the lower pressure the highest rate of ionisation appears in the bulk plasma, while at the higher pressure it is confined within a narrow $x-t$ region, near the edge of the expanding sheath. Similar to the observations of Bruneau [*et al.*]{} [@Bastien2] this changes the symmetry of the discharge and therefore the sign of the self-bias. We note that the present pressure dependence is opposite to that observed in [@Bastien2] for CF$_4$ gas, because O$_2$ is electronegative at low pressure and CF$_4$ is electronegative at high pressure.
![(a) Angular distribution of the O$_2^+$ ions at the grounded electrode in discharges driven by sawtooth-down waveforms with different number of harmonics ($N$), at pressures of $p$ = 2 Pa (thin lines) and 5 Pa (thick lines). The distributions have been normalised to a maximum value of 1, for easier comparison. (b) Time dependence of the sheath width at the grounded electrode for the same cases as shown in (a). $\phi_{\rm pp}$ = 400 V and $\gamma$ = 0.[]{data-label="fig:sawtooth_angular0"}](sawtooth_angular0.pdf){width="40.00000%"}
![Energy and angular distribution of the O$_2^+$ ion flux, $F(\varepsilon,\Theta)$ (in arbitrary units) at the grounded electrode in discharges driven by sawtooth-down waveforms with $N$ = 1 (a) and $N$ = 4 (b) harmonics. The insets illustrate the energy and angular distributions (respective integrals of $F(\varepsilon,\Theta)$ according to incidence angle and energy, respectively.) All distributions have been normalised to a maximum value of 1, for easier comparison. $\phi_{\rm pp}$ = 400 V and $p$ = 20 Pa.[]{data-label="fig:sawtooth_angular"}](sawtooth_angular.pdf){width="40.00000%"}
Figure \[fig:sawtooth\_energy\_flux\] shows the mean ion energy at different pressures, for different number of harmonics. A comparison of the results at different pressures confirms the expected behaviour that the mean ion energy decreases with increasing pressure (due to the varying collisionality of the sheaths). The data, on the other hand, indicate a marginal effect of $N$ on the mean energy of the ions, at any fixed pressure. As explained in section \[sec:valleys\] a “reversed” waveform (in the present case sawtooth down/up) mirrors the plasma on the symmetry mid-plane, this way the difference between the mean energies at the powered and grounded electrodes represents the control range of $\langle \varepsilon \rangle$ by phase control. Such a control, i.e. achieving different ion energies at the two electrodes, as inferred from figure \[fig:sawtooth\_energy\_flux\], is not feasible with sawtooth-type waveforms. Furthermore, as figure \[fig:sawtooth\_angular0\](a) reveals, the angular distribution function (IADF) of the ions is also rather insensitive to the value of $N$. At 2 Pa pressure a slight influence of $N$ is found at small angles, whereas at 5 Pa the IADFs are practically identical at various numbers of harmonics, except for $N$ = 1. This behaviour originates from the fact that the maximum sheath widths ($s$) and the voltage drops over the sheaths are weakly influenced by the operating conditions. As confirmed in figure \[fig:sawtooth\_angular0\](b), e.g., the maximum width of the sheath at the grounded electrode, $s_{\rm g}$, is very nearly the same for all conditions. The energy- and angle-resolved distributions of the ions at the grounded electrode, $F(\varepsilon,\Theta)$, show small differences when the number of harmonics is changed, as illustrated in figure \[fig:sawtooth\_angular\] for the case of 20 Pa pressure. The peaks in the IFEDF become less pronounced when the number of excitation harmonics is increased. This is caused by the reduction of the size of those domains in space and time, where ions that have undergone a charge exchange collision, can accumulate, similarly to the case of the dual-frequency waveform analysed in figure \[fig:DF\_EDDI\].
Summary
=======
In this work, we have investigated the properties of low-pressure oxygen CCPs, with focus on the ion properties: possibilities of controlling the mean energy of ions bombarding the electrode surfaces, the flux-energy, the angular- as well as the joint energy-angular distribution of these ions. These characteristics have been studied for various excitation waveforms, including single-frequency, classical dual-frequency, valleys-type and sawtooth-type waveforms.
In the case of single-harmonic excitation the shape of the energy-angular distribution at low pressures was understood by a simple analytical model. A more elaborated model was applied to account for the shape of the IFEDF at dual-frequency excitation.
A remarkable difference has been found between the distributions emerging under classical DF and valleys-type excitation: in the case of the classical DF excitation the application of the low-frequency voltage component was found to increase the width of the electrode sheaths significantly, thereby increasing their collisionality. In the case of the valleys-type waveforms the enhancement of the mean ion energy was found to occur without this effect, and as a result a more confined beam of ions (in terms of angular spread) was found to reach the electrode surfaces. Hence, applying valley-type waveforms allows for a control of the ion energy without strongly affecting the narrow ion angular distribution function. In the case of sawtooth-type waveforms a weak effect of the waveform shape (defined by the number of harmonics, $N$) on the angular distributions was found.
The electronegativity in the presence of different driving voltage waveforms was studied as a function of the voltage amplitudes and the number of harmonics. Its changes were understood based on the electron power absorption dynamics and changes of the mean electron energy that strongly affects the electronegativity via the energy dependence of the attachment cross section.
The complex waveforms, especially the classical dual-frequency waveform were found to “smoothen” the IFEDFs by making disappear the wide spatio-temporal domains within the sheath region, where ions that have undergone a charge exchange collision accumulate. The application of such waveforms, is therefore advantageous whenever a smooth IDEDF is preferred. When, on the other hand, a distinct peak in the IFEDF would be required, the method presented in [@Eddi15] may be followed.
Although our discharge model and its numerical implementation have been benchmarked with various experimental data for a wide range of operating conditions and excitation waveforms [@Derzsi2016; @Derzsi2017; @DonkoEPS2018], additional experimental data for multi-frequency discharges (e.g. on electron density, electronegativity, singlet delta molecule density, etc.) would be desired to allow further refining and verification of the discharge model. Such data, at present, exist mostly for single-frequency discharges [@Kullig], data for dual-frequency discharges are limited [@newref1; @newref2], to our best knowledge.
This work has been supported by National Research, Development and Innovation Office of Hungary (NKFIH, K119357, PD-121033), by the US National Science foundation (PHY 1601080), the DFG via SFB TR87, the International Joint Research Promotion Program (Type A) of Osaka University, the JSPS Grants-in-Aid for Scientific Research (S)15H05736 and the J. Bolyai Scholarship of the Hungarian Academy of Sciences (AD).
References {#references .unnumbered}
==========
[99]{}
Lieberman M A and Lichtenberg A J 2005 [*Principles of Plasma Discharges and Materials Processing*]{}
Makabe T and Petrović Z Lj 2006 [*Plasma Electronics: Applications in Microelectronic Device Fabrication*]{}
Chabert P and Braithwaite N 2011 [*Physics of Radio-Frequency Plasmas*]{}
Oehrlein G S and Hamaguchi S 2018 [*Plasma Sources Sci. Technol.*]{} [**27**]{} 023001
Kawamura E, Vahedi V, Lieberman M A and Birdsall C K 1999 [*Plasma Sources Sci. Technol.*]{} [**8**]{} R45
Schuengel E, Donkó Z, Hartmann P, Derzsi A, Korolov I and Schulze J 2015 [*Plasma Sources Sci. Technol.*]{} [**24**]{} 045013
Chen W, Zhang X and Diao D 2018 [*Applied Physics Express*]{} [**11**]{} 056201
Wang S, Xu X and Wang Y-N 2007 [*Phys. Plasmas*]{} [**14**]{} 113501 Hamaguchi S, Farouki R T and Dalvie M 1991 [*Phys. Rev. A*]{} [**44**]{} 3804
Kushner M J 1985 [*J. Appl. Phys.*]{} [**58**]{} 4024 Kratzer M, Brinkmann R P, Sabisch W and Schmidt H 2001 [*J. Appl. Phys.*]{} [**90**]{} 2169 Woodworth J R, Riley M E, Meister D C, Aragon B P, Le M S and Sawin H H 1996 [*J. Appl. Phys.*]{} [**80**]{} 1304 Rakhimova T V, Braginsky O V, Ivanov V V, Kovalev A S, Lopaev D V, Mankelevich Y A, Olevanov M A, Proshina O V, Rakhimov A T, Vasilieva A N and Voloshin D G 2007 [*IEEE Trans. Plasma Science*]{} [**35**]{} 1229
Huang S and Gudmundsson J T 2014 [*IEEE Trans. Plasma Science*]{} [**42**]{} 2854
Fischer G, Ouaras K, Drahi E, Bruneau B and Johnson E V 2018 [*Plasma Sources Sci. Technol.*]{} accepted manuscript DOI: https://doi.org/10.1088/1361-6595/aaca05
Wu B, Kumar A and Pamarthy S 2010 [*J. Appl. Phys.*]{} [**108**]{} 051101 Donnelly V M and Kornblit A 2013 [*J. Vacuum Science & Technol. A*]{} [**31**]{} 050825 Donkó Z, Schulze J, Czarnetzki U, Derzsi A, Hartmann P, Korolov I and Schuengel E 2012 [*Plasma Phys. Control. Fusion*]{} [**54**]{} 124003
Goto H H, Lowe H D, Ohmi T 1992 [*J. Vac. Sci. Technol. A* ]{} [**10**]{} 3048
Boyle P C, Ellingboe A R and Turner M M 2004 [*Plasma Sources Sci. Technol.*]{} [**13**]{} 493
Kitajima T, Takeo Y, Petrović Z Lj and Makabe T 2000 [*Appl. Phys. Lett.*]{} [**77**]{} 489
Lee J K, Manuilenko O V, Babaeva N Yu, Kim H C and Shon J W 2005 [*Plasma Sources Sci. Technol.*]{} [**14**]{} 89
Kawamura E, Lieberman M A and Lichtenberg A J 2006 [*Phys. Plasmas*]{} [**13**]{} 053506
Gibson A R, Greb A, Graham, W G and Gans T 2015 [*Appl. Phys. Lett.*]{} [**106**]{} 054102
Derzsi A, Schuengel E, Donkó Z and Schulze J 2015 [*Open Chem.*]{} [**13**]{} 346
Donkó Z, Schulze J, Hartmann P, Korolov I, Czarnetzki U and Schuengel E 2010 [*Appl. Phys. Lett.*]{} [**97**]{} 081501
Schulze J, Donkó Z, Schuengel E and Czarnetzki U 2011 [*Plasma Sources Sci. Technol.*]{} [**20**]{} 045007
Heil B G, Schulze J, Mussenbrock T, Brinkmann R P and Czarnetzki U 2008 [*IEEE Trans. Plasma Sci.*]{} [**36**]{} 1404
Heil B G, Czarnetzki U, Brinkmann R P and Mussenbrock T 2008 [*J. Phys. D: Appl. Phys.*]{} [**41**]{} 165202
Donkó Z, Schulze J, Heil B G and Czarnetzki U 2009 [*J. Phys. D: Appl. Phys.*]{} [**42**]{} 025205
Korolov I, Donkó Z, Czarnetzki U and Schulze J 2012 [*J. Phys. D Appl. Phys.*]{} [**45**]{} 465205
Lafleur T and Booth J P 2013 [*Appl. Phys. Lett.*]{} [**102**]{} 154104
Zhang Y-R, Hu Y-T, Gao F, Song Y-H and Wang Y-N 2018 [*Plasma Sources Sci. Technol.*]{} in press
Zhang Q-Z, Jiang W, Hou L-J and Wang Y-N 2011 [*J. Appl. Phys.*]{} [**109**]{} 013308
Bora B, Bhuyan H, Favre M, Wyndham E and Wong C S 2013 [*J. Appl. Phys.*]{} [**113**]{} 153301
Schulze J, Schuengel E, Donkó Z and Czarnetzki U 2011 [*Plasma Sources Sci. Technol.*]{} [**20**]{} 015017
Schulze J, Schuengel E, Czarnetzki U and Donkó Z 2009 [*J. Appl. Phys.*]{} [**106**]{} 063307
Wang S-B and Wendt A E 2000 [*J. Appl. Phys.*]{} [**88**]{} 643
Patterson M M, Chu H Y and Wendt A E 2007 [*Plasma Sources Sci. Technol.*]{} [**16**]{} 257
Lafleur T 2015 [*Plasma Sources Sci. Technol.*]{} [**25**]{} 013001
Bruneau B, Novikova T, Lafleur T, Booth J P and Johnson E V 2014 [*Plasma Sources Sci. Technol.*]{} [**23**]{} 065010
Bruneau B, Gans T, O’Connell D, Greb A, Johnson E V and Booth J P 2015 [*Phys. Rev. Lett.*]{} [**114**]{} 125002
Hrunski D, Mootz F, Zeuner A, Janssen A, Rost H, Beckmann R, Binder S, Schuengel E, Mohr S, Luggenhölscher D, Czarnetzki U and Grabosch G 2013 [*Vacuum*]{} [**87**]{} 114
Schuengel E, Hofmann R, Mohr S, Schulze J, Röpcke J and Czarnetzki U 2015 [*Thin Solid Films*]{} [**574**]{} 60
Zhang Y, Kushner M J, Sriraman S, Marakhtanov A, Holland A and Paterson A 2015 [*J. Vacuum Science & Technology A*]{} [**33**]{} 031302
Wang J K and Johnson E V 2017 [*Plasma Sources Sci. Technol.*]{} [**26**]{} 01LT01
Coburn J W and Winters H F 1979 [*Journal of Applied Physics*]{} [**50**]{} 3189
Childres I, Jauregui L A, Tian J and Chen Y P 2011 [*New J. Phys.*]{} [**13**]{} 025008
Lu X, Huang H, Nemchuk N and Ruoff R S 1999 [*Appl. Phys. Lett.*]{} [**75**]{} 193
Hartney M A, Hess D W and Soane D S 1989 [*J. Vacuum Sci. Technol. B*]{} [**7**]{} 1
Cvelbar U, Mozetic M and Klanjsek-Gunde M 2005 [*IEEE Trans. Plasma Sci.*]{} [**33**]{} 236
Nakamura Y, Suzuki Y and Watanabe Y 1996 [*Thin Solid Films*]{} [**290-291**]{} 367
Chou N J, Tang C H, Paraszczak J and Babich E 1985 [*Appl. Phys. Lett.*]{} [**46**]{} 31
Tourovskaia A, Barber T, Wickes B T, Hirdes D, Grin B, Castner D G, Healy K E and Floch A 2003 [*Langmuir*]{} [**19**]{} 4754
Andoa A, Uno H, Urisu T and Hamaguchi S 2013 [*Applied Surface Science*]{} [**276**]{} 1
Ando A, Asano T, Sayed Md. A, Tero R, Kitano K, Urisu T and Hamaguchi S 2012 [*Japanese Journal of Applied Physics*]{} [**51**]{} 036201
Zhang Y, Li J, An G and He X 2010 [*Sensors and Actuators B: Chemical*]{} [**144**]{} 43
Schuengel E, Zhang Q-Z, Iwashita S, Schulze J, Hou L-J , Wang Y-N and Czarnetzki U 2011 [*J. Phys. D: Appl. Phys.*]{} [**44**]{} 285205
Zhang Q-Z, Jiang W, Hou L-J and Wang Y N 2011 [*J. Appl. Phys.*]{} [**109**]{} 013308
Dittmann K, Drozdov D, Krames B and Meichsner J 2007 [*J. Phys. D: Appl. Phys.*]{} [**40**]{} 6593
Greb A, Niemi K, O[’]{}Connell D and Gans T 2013 [*Appl. Phys. Lett.*]{} [**103**]{} 244101
Greb A, Gibson A R, Niemi K, O’Connell D and Gans T 2015 [*Plasma Sources Sci. Technol.*]{} [**24**]{} 044003
Proto A and Gudmundsson J T 2018 [*Plasma Sources Sci. Technol.*]{} accepted manuscript DOI: https://doi.org/10.1088/1361-6595/aaca06
Derzsi A, Lafleur T, Booth J-P, Korolov I and Donkó Z 2016 [*Plasma Sources Sci. Technol.*]{} [**25**]{} 15004
Derzsi A, Bruneau B, Gibson A R, Johnson E, O’Connell D, Gans T, Booth J-P and Donkó Z 2017 [*Plasma Sources Sci. Technol.*]{} [**26**]{} 034002
Schulze J, Derzsi A, Dittmann K, Hemke T, Meichsner J and Donkó Z 2011 [*Phys. Rev. Lett.*]{} [**107**]{} 275001
Gudmundsson J T and Snorrason D I 2017 [*J. Appl. Phys.*]{} [**122**]{} 193302
Donkó Z, Derzsi A, Korolov I, Hartmann P, Brandt S, Schulze J, Berger B, Koepke M, Bruneau B, Johnson E, Lafleur T, Booth J-P, Gibson A R, O’Connell D and Gans T 2018 [*Plasma Phys. Control. Fusion*]{} [**60**]{} 014010
Gudmundsson J T, Snorrassn D I and Hannesdottir H 2018 [*Plasma Sources Sci. Technol.*]{} [**27**]{} 025009
Matthias P, Bandelow G, Matyash K, Duras J, Hacker P, Kahnfeld D, Kemnitz S, Lewerentz L, L[ü]{}skow K F, Meichsner J and Schneider R 2018 [*Eur. Phys. J. D*]{} [**72**]{} 86
Schuengel E, Donkó Z and Schulze J. 2017 [*Plasma Process Polym.*]{} [**14**]{} 1600117
Vahedi V and Surendra M 1995 [*Computer Phys. Commun.*]{} [**87**]{} 179
Gudmundsson J T, Kawamura E and Lieberman M A 2013 Plasma Sources Sci. Technol. [**22**]{} 035011
Biagi-v8.9 database (Cross sections extracted from PROGRAM MAGBOLTZ, VERSION 8.9 March 2010), www.lxcat.net, retrieved on November 25, 2014.
Birdsall C K 1991 [*IEEE Trans. Plasma Science*]{} [**19**]{} 65
Verboncoeur J P 2005 [*Plasma Phys. Control. Fusion*]{} [**47**]{} A231
Matyash K, Schneider R, Taccogna F, Hatayama A, Longo S, Capitelli M, Tskhakaya D and Bronold F X 2007 [*Contrib. Plasma Phys.*]{} [**47**]{} 595
Shihab M and Mussenbrock T 2017 [*Phys. Plasmas*]{} [**24**]{} 113510
Lichtenberg A J, Vahedi V and Lieberman M A 1994 [*J. Appl. Phys.*]{} [**75**]{} 2339
Brinkmann R P 2007 [*J. Appl. Phys.*]{} [**102**]{} 093303
Czarnetzki U, Schulze J, Schuengel E and Donkó Z 2011 [*Plasma Sources Sci. Technol.*]{} [**20**]{} 024010
Schulze J, Schuengel E, Donkó Z and Czarnetzki U 2010 [*Plasma Sources Sci. Technol.*]{} [**19**]{} 045028
Schulze J, Derzsi A and Donkó Z 2011 [*Plasma Sources Sci. Technol.*]{} [**20**]{} 045008
Bruneau B, Korolov I, Lafleur T, Gans T, O’Connell D, Greb A, Derzsi A, Donkó Z, Brandt S, Schuengel E, Schulze J, Johnson E and Booth J-P 2016 [*J. Appl. Phys.*]{} [**119**]{} 163301
Bruneau B, Lafleur T, Gans T, O’Connell D, Greb A, Korolov I, Derzsi A, Donkó Z, Brandt S, Schuengel E, Schulze J, Diomede P, Economou D J, Longo S, Johnson E and Booth J-P 2016 [*Plasma Sources Sci. Technol.*]{} [**25**]{} 01LT02
Küllig C, Dittmann K and Meichsner J, 2010 [*Plasma Sources Sci. Technol.*]{} [**19**]{} 065011
Liu Y X, Zhang Q Z, Liu J, Song Y H, Bogaerts A and Wang Y N 2013 [*Plasma Sources Sci. Technol.*]{} [**22**]{} 025012
Liu Y X, Zhang Q Z, Liu J, Song Y H, Bogaerts A and Wang Y N 2012 [*Appl. Phys. Lett.*]{} [**101**]{} 114101
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Arianto$^{(1,2)}$, Freddy P. Zen$^{(1)}$, Bobby E. Gunara$^{(1)}$, Triyanta$^{(1)}$ and Supardi$^{(1,3)}$\
$^{(1)}$Theoretical Physics Laboratory, THEPI,\
Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung\
Jl. Ganesha 10 Bandung 40132, Indonesia.\
$^{(2)}$Department of Physics, Udayana University\
Jl. Kampus Bukit Jimbaran Kuta-Bali 80361, Indonesia.\
$^{(3)}$Department of Physics, Sriwijaya University\
Jl. Raya Palembang-Prabumulih, Inderalaya, Indonesia.\
title: Some Impacts of Lorentz Violation on Cosmology
---
Introduction
============
Scalar field theory has become the generic playground for building cosmological models related to particle physics, in particular for obtaining inflationary cosmologies which is one of the most reliable concepts to describe the early stage of the universe. The key property of the laws of physics that makes inflation possible is the existence of states of matter that have a high energy density which cannot be rapidly lowered. The Inflationary scenario [@Lindeetc] relies on the potential energy of a scalar (inflaton) field to drive a period of early universe acceleration. It has been thought that the early universe could be well characterized by a series of phase transitions, in which topological defects could be formed [@Vilenkin]. In the context of the string theory, the natural values of the gauge and gravitational couplings in our 4d universe are explained by the dynamics of ’moduli’ scalar fields [@4; @5]. Moreover, originating from the work of Sen [@Sen] (see also [@Gibbons]), the possibility of the tachyon field being a candidate for the inflaton has been extensively studied. The tachyon action is of the Dirac Born Infeld form [@Ashoke; @Sen] which leads to an equation of state interpolating between $-1$ at early times and $0$ at late-times. This suggests the possibility that the tachyon can play the role of the inflaton at early times and the dark matter at late-times. Serious difficulties, however, plagues the tachyonic inflation [@Kofman]. These include large density perturbations, problem with reheating and formation of caustics.
Recent observational evidences especially from the Type Ia Supernovae [@Riess; @Jassal] and WMAP satellite missions [@Bennett:2003bz], indicate that we live in a favored spatially flat universe consisting approximately of 30% dark matter and 70% dark energy. In the framework of the General Relativity this means that about two thirds of the total energy density of the Universe consists of dark energy, the still unknown component with a relativistic negative pressure $p < -\rho/3$. The simplest candidate for dark energy is the cosmological $\Lambda$-term. During the cosmological evolution the $\Lambda$-term component has the constant (Lorentz invariant) energy density $\rho$ and pressure $p = -\rho$. However, it has got the famous and serious fine-tuning problem, while the also elusive dark matter candidate might be a lightest and neutral supersymmetry particle with only gravity interaction. For this reason the different forms of dynamically changing dark energy with an effective equation of state $w < -1/3$ were proposed, instead of the constant vacuum energy density. As a particular example of dark energy, the scalar field with a slow rolling potential (quintessence) [@Wetterich] is often considered. The possible generalization of quintessence is a k-essence [@Picon], the scalar field with a non-canonical Lagrangian. Any such behavior would have far-reaching implications for particle physics. However, recent theory of gravity with the Lorentz violation [@Sato:2000vu; @Coleman:1998ti] are proposed.
More recently, authors in Ref. [@KS] explored the Lorentz violating scenario in the context of the scalar-vector-tensor theory. They showed that the Lorentz violating vector affects the dynamics of the inflationary model. One of the interesting features of this scenario, is that the exact Lorentz violating inflationary solutions are related to the absence of the inflaton potential. In this case, the inflation is completely associated with the Lorentz violation. Depending on the value of the coupling parameter, the three kind of exact solutions are found: the power law inflation, de Sitter inflation, and the super-inflation.
The purpose of this paper is to study the dynamics of a scalar field in the framework of Lorentz violating scalar-vector-tensor model, taking into account the effect of the dynamically coupling vector. In this framework, we explore a class of exact solutions such as evolution of a scalar field and equation of state parameters. We discuss an inflationary scenario with a power-law coupling vector model with two typical potentials: an inverse power-law potential and a power-law potential. Then, we show that it is possible to find attractor solutions in the Lorentz violating scalar-vector-tensor model in which both the coupling function and the potential function are specified.
The organization of this paper is the following: in Section \[secII\], we set down the general formalism for the scalar-vector-tensor theory where the Lorentz symmetry is spontaneously broken due to the unit-norm vector field. In Section \[secIII\], we use our formalism to find an exact solution of the equation of state. In Section \[secIV\], we study Lorentz violating (inverse) power law inflation. In Section \[secV\], the critical points of the global system and their stability are presented. The final Section is devoted to the conclusions.
General Formalism {#secII}
==================
In the present section, we develop the general reconstruction scheme for the scalar-vector-tensor theory. We will consider the properties of general four-dimensional universe, i.e. the universe where the four-dimensional space-time is allowed to contain any non-gravitational degree of freedom in the framework of Lorentz violating scalar-tensor-vector theory of gravity. Let us assume that the expectation values of a vector field $u^\mu$ is $<0| u^\mu u_\mu |0> = -1$. The action can be written as the sum of three distinct parts: $$\begin{aligned}
S&=& S_g + S_u + S_{\phi} \ ,
\label{eq:action}\end{aligned}$$ where the actions for the tensor field $S_g$, the vector field $S_u$, and the scalar field $S_{\phi}$ are, $$\begin{aligned}
S_g &=& \int d^4 x \sqrt{-g}~ {1\over 16\pi G}R \ ,
\label{eq:act-grav} \\
S_u &=& \int d^4 x \sqrt{-g} \left[
- \beta_1 \nabla^\mu u^\nu \nabla_\mu u_\nu
-\beta_2 \nabla^\mu u^\nu \nabla_\nu u_\mu -\beta_3 \left( \nabla_\mu u^\mu \right)^2 \right. \nonumber\\
&& \left.
-\beta_4 u^\mu u^\nu \nabla_\mu u^\alpha \nabla_\nu u_\alpha
+ \lambda \left( u^\mu u_\mu +1 \right) \right] \ ,
\label{eq:act-VT} \\
S_{\phi} &=& \int d^4 x \sqrt{-g}~ {\cal{L}}_{\phi} \ .
\label{eq:act-matter}\end{aligned}$$ In the above $\beta_i(\phi)$ ($i=1,2,3,4$) are arbitrary parameters which has the dimension of mass squared. It means that $\sqrt{\beta_i}$ gives the mass scale of symmetry breakdown. ${\cal{L}}_{\phi}$ is the Lagrangian density for scalar field, expressed as a function of the metric $g_{\mu\nu}$ and the scalar field $\phi$. Then, the action (\[eq:action\]) describes the scalar-tensor-vector theory of gravity. The dimensionless vector field, $u^\mu$, satisfies the constraint $$\begin{aligned}
u^\mu u_\mu = -1 .\end{aligned}$$ For the background solutions, we use the homogeneity and isotropy of the universe spacetime $$\begin{aligned}
ds^2 = - {\mathcal{N}}^2 (t) dt^2 + e^{2\alpha(t)} \delta_{ij} dx^i dx^j \ ,\end{aligned}$$ where ${\mathcal{N}}$ is a lapse function and the scale of the universe is determined by $\alpha$. We take the constraint $$\begin{aligned}
u^\mu = \left( {1\over {\mathcal{N}}} , 0 ,0 ,0 \right) \ ,\end{aligned}$$ where ${\mathcal{N}} =1$ is taken into account after the variation. Varying the action (\[eq:action\]) with respect to $g^{\mu\nu}$, we have field equations $$\begin{aligned}
R_{\mu\nu}-{1\over 2}g_{\mu\nu}R = 8\pi G T_{\mu\nu} \ ,
\label{eq:einstein-eq}\end{aligned}$$ where $T_{\mu\nu} =T_{\mu\nu}^{(u)} + T_{\mu\nu}^{(\phi)}$ is the total energy-momentum tensor, $T_{\mu\nu}^{(u)}$ and $T_{\mu\nu}^{(\phi)}$ are the energy-momentum tensors of vector and scalar fields, respectively, defined by the usual formulae $$\begin{aligned}
T_{\mu\nu}^{(k)} = -2\frac{\partial {\cal{L}}^{(k)} }{\partial g^{\mu\nu}
}+ g_{\mu\nu} {\cal{L}}^{(k)} , \qquad k=u, {\phi} \ .\end{aligned}$$ The time and space components of the total energy-momentum tensor are given by $$\begin{aligned}
T^{0}_{0} = - \rho_u -\rho_{\phi} \ , \qquad T^{i}_i = p_u+ p_{\phi} \ ,
\label{eq:00-ii-compot}\end{aligned}$$ where the energy density and pressure of the vector field are given by $$\begin{aligned}
&& \rho_u = -3\beta H^2 \ ,
\label{eq:rho-vf} \\
&&p_u = \left(3 + 2{H^{\prime}\over H} + 2{\beta^{\prime}\over \beta} \right)\beta H^2 \ ,
\label{eq:pres-vf} \\
&& \beta \equiv \beta_1 +3 \beta_2 + \beta_3 \ .
\label{eq:def-beta}\end{aligned}$$ From the above equations, one can see that $\beta_4$ does not contribute to the background dynamics. A prime denotes the derivative of any quantities $X$ with respect to $\alpha$. $X^{\prime}$ is then related to its derivative with respect to $t$ by $X^{\prime}=(dX/dt)H^{-1}=\dot{X} H^{-1}$ where $H=d\alpha/dt=\dot{\alpha}$ is the Hubble parameter. From Eqs. (\[eq:rho-vf\]) and (\[eq:pres-vf\]), one obtains the energy equation for the vector field $u$ $$\begin{aligned}
{\rho}^{\prime}_u + 3({\rho}_u + p_u)=+3H^2 \beta^{\prime} \ ,
\label{eq:eos-vec}\end{aligned}$$ and for the scalar field $$\begin{aligned}
{\rho}^{\prime}_{\phi} + 3({\rho}_{\phi} + p_{\phi})=-3H^2 \beta^{\prime} \ .
\label{eq:eos-mat}\end{aligned}$$ The total energy equation in the presence of both the vector and the scalar fields is, accordingly, $${\rho}^{\prime} + 3({\rho} + p)=0 \ , \quad (\rho = \rho_u + \rho_{\phi}) \ .
\label{eq:eos-total}$$ This energy conservation equation can also be obtained by equating the covariant divergence of the total energy-momentum tensor to zero, since the covariant divergence of the Einstein tensor is zero by its geometric construction. It follows from contraction of the geometric Bianchi identity.
Substituting Eq. (\[eq:00-ii-compot\]) into the Einstein equations (\[eq:einstein-eq\]), we obtain two independent equations, called the Friedmann equations, as follows: $$\begin{aligned}
&& -3H^2 = 8\pi G \left(3\beta H^2 -\rho_{\phi} \right) \ ,
\label{eq:Friedmann1} \\
&& -2HH' - 3H^2 = 8\pi G \left[\left(3 + 2{H^{\prime}\over H} + 2{\beta^{\prime}\over \beta} \right)\beta H^2 +p_{\phi} \right] \ .
\label{eq:Friedmann2}\end{aligned}$$ These Friedmann equations can be rewritten as $$\begin{aligned}
&& \left( 1 + \frac{1}{8\pi G \beta} \right) H^2={1\over 3\beta} \rho_{\phi} \ ,
\label{eq:Friedmann} \\
&& \left( 1 + \frac{1}{8\pi G \beta} \right) \left( HH'+H^2\right)=-{1\over 6} \left( {\rho_{\phi}\over \beta} + {3p_{\phi}\over \beta} \right) - H^2 {\beta'\over \beta} \ .
\label{eq:Friedmannsec}\end{aligned}$$ The second term on RHS of Eq. (\[eq:Friedmannsec\]) is a consequence of the coupling vector field as a function of scalar field. If $\beta_i =0$, thus without the vector field, the above equations reduce to the conventional ones. And in the case $\beta=const$., the above equations are lead to the Friedmann equations given in Ref. [@Carroll:2004ai].
Using Eqs. (\[eq:Friedmann\]) and (\[eq:eos-mat\]), we obtain a set of equations as follows: $$\begin{aligned}
&& {H^{\prime}\over H} + {\bar{\beta}^{\prime}\over \bar{\beta}}
+ {3\over 2}(1+\omega_{\phi}) =0 \ ,
\label{eq:hb} \\
&& {H^{\prime}\over H} - {\rho^{\prime}_{\phi}\over \rho_{\phi}}
- {3\over 2}(1+\omega_{\phi})=0 \ ,
\label{eq:hr} \\
&& {\rho^{\prime}_{\phi}\over \rho_{\phi}} +
{\bar{\beta}^{\prime}\over \bar{\beta}} + 3(1+\omega_{\phi})=0
\label{eq:rb} \ ,\end{aligned}$$ where $$\begin{aligned}
\bar{\beta} &=& \beta + {1\over 8\pi G} \ ,
\label{eq:twopart}\end{aligned}$$ and $\omega_{\phi} ={p_{\phi} /\rho_{\phi}}$ is the equation of state of the scalar field. It is easy to check that the equations (\[eq:hb\])–(\[eq:rb\]) satisfy the following constraint $$\begin{aligned}
&& 2{H^{\prime}\over H} + {\bar{\beta}^{\prime}\over \bar{\beta}}
- {\rho^{\prime}_{\phi}\over \rho_{\phi}}=0 \ .
\label{eq:cls}\end{aligned}$$ In order to solve the Eqs. (\[eq:hb\])–(\[eq:rb\]) and (\[eq:cls\]), we have to specify the model and the matter content of the universe. The general solution of these equations can be written as $$\begin{aligned}
&& H\bar{\beta} \propto \exp \left[-\int {3\over
2}(1+\omega_{\phi}(\alpha)) d\alpha \right] \ ,
\label{eq:hb-sol} \\
&& {H\over \rho_{\phi}} \propto \exp \left[\int {3\over
2}(1+\omega_{\phi}(\alpha)) d\alpha \right] \ ,
\label{eq:hr-sol} \\
&& \rho_{\phi} \bar{\beta} \propto \exp \left[-\int
3(1+\omega_{\phi}(\alpha)) d\alpha \right] \label{eq:rb-sol} \ .\end{aligned}$$ If the functions $\omega_{\phi}$ and $\bar{\beta}$ are given, then we can find the evolution of the Hubble parameter under the Lorentz violation. For example, the cosmological constant corresponds to a fluid with a constant equation of state $\omega_{\phi}=-1$. Thus the above equations reduce to: $
H\bar{\beta} \propto 1$, $H \propto \rho_{\phi}$ and $\rho_{\phi}\bar{\beta} \propto 1$ where $H$, $\rho_{\phi}$ and $\beta$ are functions of $\alpha$. If $\omega_{\phi}$ is a constant parameter of a simple one component fluid, and for a given $\alpha(t)$, Eqs. (\[eq:hb\])–(\[eq:rb\]) can be used to determine $\beta(\alpha)$ and $\rho_{\phi}(\alpha)$. We, then, are able to determine the potential of the Lorentz violation model.
Dynamical Equations for Scalar Fields {#secIII}
=====================================
For a given scalar field Lagrangian with the FRW background, we can obtain the equations of motion for a scalar field by using Eq. (\[eq:eos-mat\]) and Eqs. (\[eq:hb\])–(\[eq:rb\]). Let us consider the Lagrangian density of a scalar field $\phi$ with a potential $V(\phi)$ in Eq. (\[eq:action\]): $$\begin{aligned}
{\cal{L}}_{\phi}= -{\eta\over 2}(\nabla \phi)^2 - V(\phi) \ ,\end{aligned}$$ where $(\nabla \phi)^2=g^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi$. Ordinary scalar fields correspond to $\eta = 1$ while $\eta = -1$ is for phantoms. For the homogeneous field the density $\rho_{\phi}$ and pressure $p_{\phi}$ of the scalar field, may be found as follows $$\begin{aligned}
&&\rho_{\phi} = {\eta\over 2} H^2 \phi^{\prime 2} + V(\phi) \ ,
\label{eq:infrho} \\
&&p_{\phi}= {\eta\over 2} H^2 \phi^{\prime 2} - V(\phi) \ .
\label{eq:infp}\end{aligned}$$ The corresponding equation of state parameter is, accordingly $$\begin{aligned}
\omega_{\phi}={p_{\phi}\over \rho_{\phi}} = - \frac{1- \eta H^2
\phi^{\prime 2}/2V}{1 + \eta H^2 \phi^{\prime 2}/2V} \ .
\label{phant-eos}\end{aligned}$$ Substituting Eq. (\[eq:infrho\]) into Eq. (\[eq:Friedmann\]), the Friedmann equation leads to $$\begin{aligned}
H^2 = \frac{1}{3\bar{\beta}} \left[
\frac{\eta}{2} H^2 \phi^{\prime 2} + V(\phi) \right] \ .
\label{eq:01-1}\end{aligned}$$ Now, differentiating Eq. (\[eq:infrho\]) with respect to $\alpha$ and using Eq. (\[eq:eos-mat\]), and also differentiating Eq. (\[eq:01-1\]) with respect to $\alpha$ and using Eq. (\[eq:03-1\]) give, respectively, $$\begin{aligned}
\phi'' &=&- \left( \frac{H'}{H} + 3\right)\phi' - \eta \frac{V_{,\phi}}{H^2}
- 3\eta \bar{\beta}_{,\phi} \ ,
\label{eq:03-1}\\
\phi^{\prime} &=&-2\eta \bar{\beta}\left(\frac{H_{,\phi}}{H} +
\frac{\bar{\beta}_{,\phi}}{\bar{\beta}} \right) \ .
\label{eq:02-1}\end{aligned}$$ Substituting Eq. (\[eq:02-1\]) into the Friedmann equation the potential of the scalar field can be written as $$\begin{aligned}
V = 3\bar{\beta} H^2 \left[ 1-{2\over 3}\eta\bar{\beta}\left({\bar{\beta}_{,\phi}\over \bar{\beta}} + {H_{,\phi}\over H}\right)^2 \right] \ .\end{aligned}$$ Note that in the above equations the Hubble parameter $H$ has been expressed as a function of $\phi$, $H=H(\phi(t))$. From Eq. (\[eq:hb\]), the equation of state can be written as $$\begin{aligned}
\omega_\phi &=& -1 + {4\over 3}\eta\bar{\beta}\left(\frac{H_{,\phi}}{H} + \frac{\bar{\beta}_{,\phi}}{\bar{\beta}} \right)^2 \nonumber\\
&=&-1 + {1\over 3}\eta \frac{\phi^{\prime 2}}{\bar{\beta}} \ .
\label{eos1-1}\end{aligned}$$ Equations (\[eq:02-1\]) and (\[eos1-1\]) are two equations that we need to solve for the scalar field $\phi$ and the equation of state $\omega_\phi$. This is achieved only if the Hubble parameter $H(\phi)$ and the coupling vector ${\bar{\beta}}(\phi)$ are known. For different choice of the Hubble parameter $H(\phi)$ and the coupling vector ${\bar{\beta}}(\phi)$, it is possible to extract a class of exact solutions of Eqs. (\[eq:02-1\]) and (\[eos1-1\]). We shall solve Eqs. (\[eq:02-1\]) and (\[eos1-1\]) to obtain the following physical quantities ($V$ and $K$ are the potential and kinetic energies, respectively): $$\begin{aligned}
&&V={3\over 2}(1-\omega_\phi){\bar{\beta}}H^2, \quad K={3\over 2}(1+\omega_\phi){\bar{\beta}}H^2, \nonumber\\
&&\rho_\phi=3{\bar{\beta}}H^2 \ ,~~\quad\quad \qquad p_\phi=3\omega_\phi {\bar{\beta}}H^2 \ .
\label{quant-phys}\end{aligned}$$ In the following two subsections we will explore a class of exact solutions.
Exact solutions and the behavior of scalar fields
-------------------------------------------------
We shall have to solve equations (\[eq:02-1\]) and (\[eos1-1\]) for $H$, $\omega_\phi$, $\bar{\beta}$, and $V$, which is not possible unless two are known. In the present subsection, we consider an example to find an exact solution of the equation of state of the scalar field in the quadratic coupling vector. The equation of state for the scalar field has been intensively studied in [@Zlatev] for the so called tracking cosmological solutions introduced in [@Steinhardt99], and some classes of potentials allowing for the field equation of state were described.
Let us consider a simple model $$H=H_0 \ , \qquad \bar{\beta}(\phi) = m\phi^2 \ ,
\label{model-hconst1}$$ where $H_0$ and $m$ are positive constant parameters. The equation (\[eq:02-1\]) can now be integrated to yield the evolution of the scalar field $$\phi(t)=\phi_0 \exp \left[-4\eta mH_0(t-t_0) \right] \ ,
\label{sol-n2}$$ where $\phi(t=t_0)\equiv \phi_0$ is a constant. Then, it is easy to find the equation of state of the scalar field by using Eq. (\[eos1-1\]). We obtain $$\begin{aligned}
\omega_\phi &=& -1 + {16\over 3} m \ , \quad \textit{for ordinary scalar fields} \ ,
\label{const-eos-ord} \\
\omega_\phi &=& -1 - {16\over 3} m \ , \quad \textit{for phantom fields} \ .
\label{const-eos-phant}\end{aligned}$$ Then, the potential and the kinetic energies, the energy density and the pressure of the scalar field evolve according to $$\begin{aligned}
V(t) &=& mH_0^2\phi_0^2(3 - 8m)\exp \left[-8\eta mH_0(t-t_0) \right] \ , \\
K(t) &=& 8\eta\left(mH_0\phi_0\right)^2 \exp \left[-8\eta mH_0(t-t_0) \right]\ , \\
\rho(t) &=& 3mH_0^2\phi_0^2 \exp \left[-8\eta mH_0(t-t_0) \right] \ , \\
p(t) &=& \eta m H_0^2\phi_0^2 \left( 16m -3\eta \right)\exp \left[-8\eta mH_0(t-t_0) \right] \ .\end{aligned}$$ The above solutions are completely associated with the Lorentz violation. The model (\[model-hconst1\]) depicts that the cosmic evolution starts from a constant value of the scale factor and grows exponentially, $a(t)=a_0 e^{H_0(t-t_0)}$. The coupling vector decreases exponentially for the ordinary scalar field and increases exponentially for the phantom field from a constant value of $m\phi_0^2$. Hence the potential energy, the kinetic energy, the energy density and the pressure decrease exponentially for the ordinary scalar field. For the phantom field, on the other hand, the potential energy and energy density as well as the absolute values of the kinetic energy and the pressure increase exponentially. Note that the kinetic energy and the pressure begin with the negative values. The Eqs. (\[const-eos-ord\])–(\[const-eos-phant\]) show that the equation of state $\omega_\phi$ is non-dynamical because it only depends on the value of the coupling vector parameter $m$, both for the ordinary scalar and the phantom fields. Since an accelerated expansion occurs for $\omega_\phi < -1/3$ then we have $m<1/8$ for the ordinary scalar field. However, the present data of the Universe seems to tell that $\omega_\phi$ might be less than $-1$. Thus, the value $m$ may be chosen in order to fit the present observable constraint on the equation of state parameter.
In other case, for instance, $H(\phi)=H_0\phi^\xi$ and $\bar{\beta}(\phi)=m\phi^2$, we also find the constant equation of state, $$\begin{aligned}
\omega_\phi = -1 + {4\over 3} \eta m (\xi + 2)^2 \ .
\label{const-eos-gen}\end{aligned}$$ The condition for the accelerating Universe $\ddot{a}$ or $H'/H >
-1$ yields $$\begin{aligned}
\eta m < {1\over 2\xi (\xi + 2)} \ .
\label{cond-acc}\end{aligned}$$ This model gives a power law expansion $$\begin{aligned}
{a(t)\over a_o}= \left[1 + {H_0\phi_0^\xi \over p}(t-t_0) \right]^p \ , \qquad p>1 \ ,\end{aligned}$$ where $$\begin{aligned}
p={1\over 2\eta m\xi(\xi+2)} \ .\end{aligned}$$ The scalar field evolve as $$\phi(t)=\phi_0\left(1+{H_0\phi_0^\xi\over p}(t-t_0)\right)^{-1/\xi} \ .
\label{phi-lain}$$ Hence, the complete set of solutions is found by substituting Eqs. (\[const-eos-gen\]) and (\[phi-lain\]) into Eqs. (\[quant-phys\]).
In the following subsection, we will see that the equation of state may be dynamics. For this purpose we generalize the coupling vector to $\bar{\beta}(\phi)=m\phi^n$, $n>2$.
Variable equation of states
---------------------------
Let us consider a model where the coupling vector is a power law of the scalar field, $$H=H_0 \ , \quad \bar{\beta}(\phi) = m\phi^n \ ,\quad n >2 \ ,
\label{model-hconst}$$ where $H_0$, $m$ and $n$ are constant positive parameters. Following the same above procedure, the scalar $\phi$ can be evaluated as, $$\phi(t)=\frac{\phi_0}{ \left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{1\over{n-2}}} \ ,$$ the coupling vector is given by $$\bar{\beta}(t)=\frac{m\phi_0^n}{ \left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{n\over{n-2}}} \ ,$$ and the dynamical equation of state (\[eos1-1\]) is $$\begin{aligned}
\omega(t) = -1 + \frac{4\eta mn^2\phi_0^{n-2}/3}{1+2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0)} \ .
\label{eos-narbri}\end{aligned}$$ Then, the potential and kinetic energies, the energy density and the pressure of the scalar field are given by $$\begin{aligned}
V(t) &=& 3mH_0^2\phi_0^n\left[1- \frac{2\eta mn^2\phi_0^{n-2}/3}{1+2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0)}\right] \times \nonumber\\
&&\times {1\over\left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{{n\over{n-2}}}} \ ,\\
K(t) &=& \frac{2\eta\left(mnH_0\phi_0^{n-1}\right)^2}{ \left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{{2(n-1)\over{n-2}}}} \ , \\
\rho(t) &=& \frac{3mH_0^2\phi_0^n}{ \left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{{n\over{n-2}}}} \ , \\
p(t) &=& 3mH_0^2\phi_0^n \left[ -1 + \frac{4\eta mn^2\phi_0^{n-2}/3}{1+2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0)} \right] \nonumber\\
&&\times {1\over\left[1 + 2\eta mnH_0(n-2)\phi_0^{n-2}(t-t_0) \right]^{{n\over{n-2}}}} \ .\end{aligned}$$ Thus, the model (\[model-hconst\]) describes that the cosmic evolution grows exponentially from a constant value of the scale factor, $a(t)=a_0 e^{H_0(t-t_0)}$, while the coupling vector $\bar{\beta}$ started from a constant value of the scalar field, $m\phi_0^n$. The equation of state $\omega_\phi$ is dynamical both for the ordinary scalar and phantom fields. Then the potential energy, kinetic energy, the energy density and the pressure decrease for the ordinary scalar field. For the phantom field, on the other hand, the potential and energy density increase while the kinetic energy and pressure begin with the negative values.
Lorentz Violating Inflation Scenario {#secIV}
=====================================
As it has been studied by authors in Ref. [@KS], the Lorentz violation on the inflationary scenario can be divided into two parts: the Lorentz violations stage $8\pi G \beta \gg 1$ and the standard slow roll stage $8\pi G \beta \ll 1$. The first stage corresponds to $\bar{\beta} = \beta$ in Eq. (\[eq:twopart\]) and the second stage corresponds to $\bar{\beta} = 1/ 8\pi G $, then we have the usual dynamical equations.
In this section we will consider the inflationary scenario for the scalar field (inflaton). In particular, we consider a power-law coupling vector, $\beta(\phi) = m \phi^n$, with two types of the potential: $V(\phi) = \mu^{4+\nu}\phi^{-\nu}$ and $V(\phi) = {1\over 2}M^2\phi^2$. Here $\mu$, $\nu$ and $M$ are parameters. Thus, the dynamics of each particular inflationary model are determined by the Friedmann equation and the scalar field equation of motion once the functional form of the inflaton potential and the coupling parameter have been specified. Let us collect the dynamics-related equations for the inflaton the Friedmann equation (\[eq:01-1\]) in inflationary models $$\begin{aligned}
H^2 = \frac{1}{3\bar{\beta}} \left[
\frac{1}{2} H^2 \phi^{\prime 2} + V(\phi) \right] \ ,
\label{eq:01}\end{aligned}$$ the constraint equation (obtained from Eqs. (\[phant-eos\]) and (\[eq:hb\])) $$\begin{aligned}
\frac{H'}{H} + \frac{1}{2} \frac{\phi^{\prime 2}}{\bar{\beta}} + \frac{\bar{\beta}'}{\bar{\beta}} =0 \ ,
\label{eq:02}\end{aligned}$$ and the equation of motion Eq. (\[eq:03-1\]) $$\begin{aligned}
\phi'' + \frac{H'}{H} \phi' + 3\phi' + \frac{V_{,\phi}}{H^2}+ 3 \bar{\beta}_{,\phi} =0 \ .
\label{eq:03}\end{aligned}$$ $\bar{\beta}$ is given by Eq. (\[eq:twopart\]). Then, at the critical value of $\phi$, the effective coupling vector becomes $$\begin{aligned}
8\pi G \beta (\phi_c ) =1 \ .\end{aligned}$$ For example, a coupling parameter of the form $\beta=m\phi^2$ gives the critical value $$\begin{aligned}
\phi_c = \frac{M_{pl}}{\sqrt{8m \pi }} \ , M_{pl} = G^{-1} \ .\end{aligned}$$ Let $\phi_i$ be the corresponding initial value of the scalar field. Putting $\phi_i \sim 3 M_{pl}$, the Lorentz violation implies the criterion $m > 1/(72\pi) \sim 1/226$.
The set of Eqs. (\[eq:01\])–(\[eq:03\]) constitutes the equations we have to solve for the problem specified by the coupling parameter $\beta(\phi)$ and the potential $V (\phi)$. In following subsection, we consider with a model with the coupling parameter $\beta(\phi)$ is given by $$\begin{aligned}
\beta(\phi) = m \phi^n \ ,
\label{eq:model-1}\end{aligned}$$ where $n$ and $m$ are parameters. For the model (\[eq:model-1\]), we obtain the critical value of the scalar field and the criterion for Lorentz violation $$\begin{aligned}
\phi_c = \left( \frac{M_{pl}^2}{8m \pi }\right)^{1/n} \quad \textit{and}\quad m > \frac{M_{pl}^2}{8\pi(3M_{pl})^{n} } \ .\end{aligned}$$ Now, we consider two typical potentials appear in many cosmological implications: an inverse power-law potential and a power-law potential. We discuss those solutions and analyze the two regimes separately.
Inverse power law potential: $V(\phi) = \mu^{4+\nu}\phi^{-\nu}$
---------------------------------------------------------------
In this subsection, we consider the class of power law potential $$\begin{aligned}
V(\phi) = \mu^{4+\nu}\phi^{-\nu} \ ,\end{aligned}$$ where $\mu$ and $\nu$ are constants. Inverse power law models are interesting for a number of reasons. In conventional cosmology, they drive ‘intermediate’ inflation [@Barrow] and typically produce significant tensor perturbations for almost scale-invariant scalar fluctuations. They arise in supersymmetric condensate models of QCD [@Binetruy] and can in principle act as a source of quintessence [@Wetterich; @Balbi].
### Lorentz violating stage
Let us first consider the Lorentz violating stage, $8\pi G \beta \gg 1$ ($\bar{\beta}=\beta$), we have the equations (\[eq:01\])–(\[eq:03\]). In this stage both the coupling function and the potential function are relevant.
An inflationary epoch, in which the scale factors $a$ are accelerating, requires the scalar field $\phi$ to evolve slowly compared to the expansion of the universe. Thus, the following conditions of slow-rolling are required: $$\begin{aligned}
H^2 \phi^{\prime 2} \ll V \ , \quad \phi'' \ll \phi' \ , \quad \phi^{\prime 2} \ll \beta , \quad \textit{and} \quad \beta' \ll \beta \ .
\label{eq:slowroll-inflaton}\end{aligned}$$ The formalism which gives these slow roll conditions are discussed in Ref. [@KS]. This is sufficient to guarantee inflation. Under the slow-roll conditions Eq. (\[eq:slowroll-inflaton\]), the Eqs. (\[eq:01\])–(\[eq:03\]) can be simplified. We obtain the slow roll equations $$\begin{aligned}
H^2 \simeq \frac{V}{3\beta} \ , \quad \textit{and} \quad \phi' \simeq - \beta\left({\beta_{,\phi} \over \beta}+ \frac{V_{,\phi}}{V} \right) \ .
\label{eq:sl-inflaton-1}\end{aligned}$$ Inserting Eq. (\[eq:model-1\]) and the potential of the form $V(\phi) = \mu^{4+\nu}\phi^{-\nu}$ into Eq. (\[eq:sl-inflaton-1\]), we have $$\begin{aligned}
H^2 &=& \frac{\mu^{4+\nu}}{3m} \phi^{-(\nu+n)} \ ,
\label{eq:eolut-sol-Frid-inflaton}\\
\phi' &=& -m(n-\nu)\phi^{(n-1)} \ .
\label{eq:eolut-sol-scl-inflaton}\end{aligned}$$ One can then solve for $\phi$ from Eq. (\[eq:eolut-sol-scl-inflaton\]), $$\begin{aligned}
\phi(\alpha) = \left[\phi_i^{2-n}+m(n-2)(n-\nu)(\alpha-\alpha_i)\right]^{-{1\over n-2}} \ , \quad \textit{for} \quad n \neq 2, n \neq \nu \ ,
\label{eq:eolut-sol1-scl-inflaton}\end{aligned}$$ where $\phi(\alpha =\alpha_i)\equiv \phi_i$ is a constant. The Friedmann equation gives $$\begin{aligned}
H^2(\alpha) = \frac{\mu^{4+\nu}}{3m} \left[\phi_i^{2-n}+m(n-2)(n-\nu)(\alpha-\alpha_i)\right]^{{n+\nu \over n-2}} \ .
\label{eq:eolut-sol1-Frid-inflaton}\end{aligned}$$ The solution (\[eq:eolut-sol1-scl-inflaton\]) and the slow roll conditions (\[eq:slowroll-inflaton\]) during the Lorentz violating stage give $n>2$ because $$\begin{aligned}
&& \phi^{\prime 2}\left(\sim \alpha^{-2(1-n)/(2-n)}\right) \ll \beta\left(\sim \alpha^{n/(2-n)}\right) \ , \\
&&\beta' \left(\sim \alpha^{-2(1-n)/(2-n)}\right) \ll \beta
\left(\sim \alpha^{n/(2-n)}\right) \ .\end{aligned}$$ From Eq. (\[eq:eolut-sol1-Frid-inflaton\]), the universe expands during the Lorentz violating stage as $$\begin{aligned}
{a(t) \over a_i} = \exp \left\{ -{B\over C}+ {1\over C} \left( {1\over B^{D-1}}-AC(D-1)(t-t_i) \right)^{-{1\over D-1}} \right\} \ , \quad C \neq 0 \ ,
\label{eq:scale-univ-inflaton}\end{aligned}$$ where the constants $A, B, C$ and $D$ are $$\begin{aligned}
A=\sqrt{{\mu^{4+\nu}\over 3m}} \ , \quad B=\phi_i^{2-n} \ , \quad C=m(n-2)(n-\nu) \ , \quad D=\frac{n+\nu}{2(n-2)} \ .\end{aligned}$$ Combining Eqs. (\[eq:scale-univ-inflaton\]), (\[eq:eolut-sol1-scl-inflaton\]) and (\[eq:eolut-sol1-Frid-inflaton\]), we obtain the physical quantities $$\begin{aligned}
&& \phi(t) = \left( {1\over B^{D-1}}-AC(D-1)(t-t_i) \right)^{{1\over (D-1)(n-2)}} \ ,
\label{eq:eolut-sol1-scl-time} \\
&& H(t) = A \left( {1\over B^{D-1}}- AC(D-1)(t-t_i) \right)^{-{D\over (D-1)}} \ ,
\label{eq:eolut-sol1-Frid-time}\\
&& \beta(t)=n\left( {1\over B^{D-1}}- AC(D-1)(t-t_i) \right)^{{n\over (D-1)(n-2)}} \ .
\label{eq:eolut-sol1-beta-time}\end{aligned}$$ The scalar field energy density, on the other hand, evolves according to $$\begin{aligned}
\rho (t)\simeq V =3mA^2\left( {1\over B^{D-1}}-AC(D-1)(t-t_i) \right)^{{n-2D(n-2)\over (D-1)(n-2)}} \ .\end{aligned}$$ One can see that the Hubble parameter $H$ decreases during the Lorentz violation stage.
For $n=2, \nu\neq 2$, Eq. (\[eq:model-1\]) and the second part of Eq. (\[eq:sl-inflaton-1\]) gives $$\begin{aligned}
\phi (\alpha)= \phi_i e^{-n(2-\nu)(\alpha -\alpha_i)} \ ,
\label{eq:sol1-sl-inflaton}\end{aligned}$$ where $\phi(\alpha=\alpha_i)\equiv \phi_i$. For this solution to satisfy slow roll conditions (\[eq:slowroll-inflaton\]), we need $ m <1/(2-\nu)^2$. Thus, we have the range $1/226 <m <1/(2-\nu)^2$ of the parameter for which the Lorentz violating inflation is relevant. The Hubble parameter as a function of the scale factor, $\alpha$, is given by $$\begin{aligned}
H^2(\alpha) = H^2_i e^{-m(\nu^2-4)(\alpha - \alpha_i)} \ ,
\label{eq:sol2-sl-inflaton}\end{aligned}$$ where $$\begin{aligned}
H^2_i =\frac{\mu^{4+\nu}}{3 m \phi_i^{(2+\nu)}} \ ,\end{aligned}$$ while the scale factor $a(t) = e^{\alpha}$ is of the form: $$\begin{aligned}
{a(t)\over a_i} = \left[{\phi (t) \over \phi_i}\right]^{{1\over m(\nu-2)}} \ .\end{aligned}$$ Now we obtain the evolution of some physical quantities as follows $$\begin{aligned}
&& {a(t)\over a_i} =\left[m(\nu^2-4)H^2_i(t-t_i)\right]^{1\over m(\nu^2-4)} \ ,
\label{eq:eolut-m2-a} \\
&& {\phi(t)\over\phi_i} = \left[m(\nu^2-4)H^2_i(t-t_i)\right]^{1\over \nu + 2} \ ,
\label{eq:eolut-m2-scl-time} \\
&& {H(t)\over H^2_i} = \sqrt{m(4-\nu^2)(t-t_i)} \ ,
\label{eq:eolut-m2-Frid-time}\\
&& {\beta(t)\over m\phi^2_i}= \left[m(\nu^2-4)H^2_i(t-t_i)\right]^{2\over \nu + 2} \ ,
\label{eq:eolut-m2-beta-time}\\
&&\rho (t)={\mu^{4+\nu}\over \phi_i}\left[m(\nu^2-4)H^2_i(t-t_i)\right]^{-{\nu \over \nu + 2}} \ .\end{aligned}$$
### Standard slow roll stage
The governing equations (\[eq:01\])–(\[eq:03\]) in the standard slow roll stage $8\pi G\beta \ll 1$ ($\bar{\beta}=(8\pi G)^{-1}$), are, accordingly, $$\begin{aligned}
&& H^2 = \frac{8\pi G}{3} \left[ \frac{1}{2} H^2 \phi^{\prime
2} + V \right] \ ,
\label{eq:ssrs1}\\
&& \frac{H'}{H} + 4\pi G \phi^{\prime 2} =0 \ ,
\label{eq:ssrs2}\\
&& \phi'' + \frac{H'}{H} \phi' + 3\phi' + \frac{V_{,\phi}}{H^2}
=0 \ . \label{eq:ssrs3}\end{aligned}$$ In this case the slow roll equations are given by $$\begin{aligned}
H^2 \simeq \frac{8\pi G}{3} V \ , \quad
\phi'\simeq - \frac{1}{8\pi G} \frac{V_{,\phi}}{V} \ .\end{aligned}$$ For the potential model $V(\phi) = \mu^{4+\nu}\phi^{-\nu}$, the evolution of the inflaton and the Hubble parameter can be solved as $$\begin{aligned}
\phi^2 (\alpha) &=& \phi_c^2 + \frac{\nu}{4\pi G }(\alpha-\alpha_c) \ , \\
H^2(\alpha) &=& \frac{8\pi G}{3} \mu^{4+\nu} \left[ \phi_c^2 + \frac{\nu}{4\pi G }(\alpha-\alpha_c)\right]^{-\nu/2} \ ,\end{aligned}$$ and the scale factor is given by $$\begin{aligned}
{a (t)\over a_c} = \exp \left[ {4\pi G \over \nu}( \phi^2 (t) -\phi_c^2) \right] \ .\end{aligned}$$ The evolution equations are given by $$\begin{aligned}
&& {a(t)\over a_c} = \exp\left\{-{B_s\over C_s}+{1\over C_s} \left[B_s^{D_s+1} +A_s C_s(D_s+1)(t-t_c) \right]^{{1\over D_s+1}}\right\} \ ,
\label{eq:eolut-sdr-a} \\
&& \phi(t)= \left[B_s^{D_s+1} +A_s C_s(D_s+1)(t-t_c) \right]^{{1\over 2(D_s+1)}} \ ,
\label{eq:eolut-sdr-scl-time} \\
&& H(t) = A_s\left[B_s^{D_s+1} +A_s C_s(D_s+1)(t-t_c) \right]^{{D_s\over (D_s+1)}} \ ,
\label{eq:eolut-sdr-Frid-time}\end{aligned}$$ and the scalar field energy density evolves as $$\begin{aligned}
\rho(t)= {3\over 8}\left(\frac{A_s^2 C_s}{D_s} \right)\left[B_s^{D_s+1} +A_s C_s(D_s+1)(t-t_c) \right]^{{2D_s\over (D_s+1)}} \ .
\label{eq:eolut-sdr-beta-time}\end{aligned}$$ where $A_s$, $B_s$, $C_s$ and $D_s$ are the constants, $$\begin{aligned}
A_s = \sqrt{\frac{8\pi G}{3} \mu^{4+\nu}} \ , \quad B_s =\phi_c^2 \ , \quad C_s =\frac{\nu}{4\pi G } \ , \quad D_s =\frac{\nu}{4} \ .\end{aligned}$$ Note that, in the standard slow roll stage, the Hubble parameter $H$ increases.
Another interesting quantity is the number of e-folding during the inflationary phase. The total e-folding number reads $$\begin{aligned}
N &=& -{B\over C}+ {1\over C} \left( {1\over B^{D-1}}- AC(D-1)(t_c-t_i) \right)^{-{1\over D-1}} \nonumber\\
&&-{B_s\over C_s}+{1\over C_s} \left[B_s^{D_s+1} +A_s C_s(D_s+1)(t_e-t_c) \right]^{{1\over D_s+1}} \nonumber\\
&=&{1\over C} \left( \phi^{2-n}_c - \phi_i^{2-n} \right)+ \frac{4\pi G }{\nu} \left( \phi^2_e - \phi_c^2 \right) \ ,\end{aligned}$$ for $m > 2, \nu \neq m$ and $$\begin{aligned}
N &=&{1\over n(\nu^2-4)}\log \left[n(\nu^2-4)H^2_i(t_c-t_i)\right] \nonumber\\
&& -{B_s\over C_s}+{1\over C_s} \left[B_s^{D_s+1} +A_s C_s(D_s+1)(t_e-t_c) \right]^{{1\over D_s+1}} \nonumber\\
&=&{1\over m(2-\nu)} \log \left( {\phi_i \over \phi_c} \right)+ \frac{4\pi G }{\nu} \left( \phi^2_e - \phi_c^2 \right) \ ,\end{aligned}$$ for $n = 2, \nu \neq 2$. Note that the first terms of the above equations arise from the Lorentz violating stage. As an example, let us take the values: $N=70$, $m = 10^{-2}$, $n=2$ and $\nu=1$. If $\phi_e \sim 0.3 M_{pl}$ is the value of scalar field at the end of inflation, then, $\phi_c \sim 2 M_{pl}$. The contribution from the inflation end is still relevant. Therefore, we get $\phi_i \sim 2.5 M_{pl}$.
Power law potential: $V(\phi) = {1\over 2}M^2\phi^2$
----------------------------------------------------
### Lorentz violating stage
The most realistic inflationary universe scenarios are chaotic models. For the model $V(\phi) = {1\over 2}M^2\phi^2$, assuming the slow roll conditions, we find the slow roll equations during the Lorentz violating regime as follows $$\begin{aligned}
H^2 &=& \frac{M^2}{6n} \phi^{-(n-2)} \ ,
\label{eq:eolut-sol-Frid-chao}\\
\phi' &=& -m(n+2)\phi^{(n-1)} \ .
\label{eq:eolut-sol-scl-chao}\end{aligned}$$ Then we find the solution (\[eq:eolut-sol-scl-chao\]) as $$\begin{aligned}
\phi(\alpha) = \left[\phi_i^{2-n}+m(n^2-4)(\alpha-\alpha_i)\right]^{{1\over 2-n}} \ ,
\label{eq:eolut-neq2-scl-inflaton}\end{aligned}$$ for $m\neq 2$ and $$\begin{aligned}
\phi(\alpha) = \phi_i e^{-4m(\alpha-\alpha_i)} \ ,
\label{eq:eolut-e2-scl-inflaton}\end{aligned}$$ for $n= 2$. The inflationary scenario of this model was already obtained in Ref. [@KS] where the Hubble parameter becomes constant during the Lorentz violating regime and $1/226<m<1/16$ is the range of parameter $m$. We concern here the solution for $n\neq 2$. The solution for the Hubble parameter is given by $$\begin{aligned}
H^2(\alpha) = \frac{M^2}{6m} \left[\phi_i^{2-n}+m(n^2-4)(\alpha-\alpha_i)\right] \ ,
\label{eq:eolut-sol1-Frid-chao}\end{aligned}$$ and $$\begin{aligned}
{a(t)\over a_i}= \exp \left[ \frac{1}{m(n^2-4)} \left( {1\over \phi^{n-2}(t)} -{1\over \phi_i^{n-2}}\right) \right] \ ,
\label{eq:eolut-scale1-chao}\end{aligned}$$ which is the solution for the scale factor. As in the previous subsection, we also obtain $n>2$ which the effect of Lorentz violation occurs in this regime. The time evolution of the above equations can be obtained by integrating Eq. (\[eq:eolut-sol1-Frid-chao\]), we get $$\begin{aligned}
\alpha(t) =\alpha_i -{b\over c}+ {1\over c} \left[ b^{1/2}+{1\over 2}dc(t-t_i) \right]^2 \ ,
\label{eq:eolut-scale1-time}\end{aligned}$$ where $$\begin{aligned}
b=\phi_i^{2-n} \ , \quad c =m(n^2-4) \ , \quad d = \sqrt{\frac{M^2}{6m}} \ , \quad n > 2 \ .\end{aligned}$$ Then the evolution equations are given by $$\begin{aligned}
&& {a(t)\over a_i}= \exp \left\{ -{b\over c}+{1\over c} \left[ b^{1/2} +{1\over2}d c(t-t_i) \right]^{2}\right\} \ ,
\label{eq:1} \\
&& \phi(t) = \left[ b^{1/2} +{1\over2}d c(t-t_i) \right]^{-{2\over n-2}} \ ,
\label{eq:2} \\
&& H(t)= d \left[ b^{1/2} +{1\over2}d c(t-t_i) \right] \ ,
\label{eq:3}\\
&& \beta(t) = n\left[ b^{1/2}+{1\over2}d c(t-t_i) \right]^{-{2n\over n-2}} \ ,
\label{eq:4}\\
&& \rho(t) = 3nd^2 \left[ b^{1/2} +{1\over2}d c(t-t_i) \right]^{-{4\over n-2}} \ .
\label{eq:5}\end{aligned}$$ Since $b$, $c$ and $d$ are positive constants, one can see that the Hubble parameter $H$ and the scale factor $a$ increase during the Lorentz violating stage for $n>2$. In the case $n=2$, the Hubble parameter is constant. In the following subsection, we will see that the Hubble parameter decreases in the standard slow roll stage.
### Standard slow roll stage
Now, let us consider the chaotic inflationary scenario in the standard slow roll stage. A set of the dynamical equations of the scalar field are given by Eqs. (\[eq:ssrs1\])–(\[eq:ssrs3\]). Assuming the standard slow roll conditions, we find the slow roll equations $$\begin{aligned}
&& H^2 \simeq \frac{4\pi G}{3} M^2 \phi^2 \ , \\
&& \phi' \simeq - \frac{1}{4\pi G} \phi^{-1} \ .\end{aligned}$$ The evolution of the inflaton can be solved as $$\begin{aligned}
\phi^2 (\alpha) = \phi_c^2 - \frac{1}{2\pi G }(\alpha-\alpha_c) \ .\end{aligned}$$ The Hubble parameter and the scale factor $a(t) = e^{\alpha}$ can be also obtained as $$\begin{aligned}
&& H^2 = \frac{4\pi GM^2 }{3} \left( \phi_c^2 - \frac{1}{2\pi G }(\alpha-\alpha_c) \right) \ ,
\label{eq:h1} \\
&& a (t) =a_c \exp \left[ 2\pi G ( \phi_c^2 -\phi^2 (t) ) \right] \ .
\label{eq:a1}\end{aligned}$$ From Eq. (\[eq:h1\]), we obtain $$\begin{aligned}
\alpha(t) -\alpha_c= {b_s\over c_s}-{1\over c_s} \left[b_s^{1/2} -{1\over 2}d_s c_s(t-t_c) \right]^{2} \ ,\end{aligned}$$ and the dynamical evolutions are given by $$\begin{aligned}
&& {a(t)\over a_c}= \exp \left[ {b_s\over c_s}-{1\over c_s} \left(b_s^{1/2} -{1\over2}d_s c_s(t-t_c) \right)^{2}\right] \ ,
\label{eq:1st} \\
&& \phi (t) = b_s^{1/2} -{1\over2}d_s c_s(t-t_c) \ ,
\label{eq:2nd} \\
&& H(t) = d_s \left[b_s^{1/2} -{1\over2}d_s c_s(t-t_c) \right] \ ,
\label{eq:3td}\\
&& \rho(t) = {3c_s d_s\over 4}\left[b_s^{1/2} -{1\over2}d_s c_s(t-t_c) \right]^{2} \ ,
\label{eq:5fi}\end{aligned}$$ with $$\begin{aligned}
b_s= \phi_c^2 \ , \quad c_s = \frac{1}{2\pi G } \ , \quad d_s = \frac{4\pi GM^2 }{3} \ .\end{aligned}$$ Note that the Hubble parameter decreases in the standard slow roll stage.
In the case of chaotic potential, the total e-folding number reads $$\begin{aligned}
N &=& -{b\over c}+{1\over c} \left[ b^{1/2} +{1\over2}d c(t_c-t_i) \right]^{2} +{b_s\over c_s}-{1\over c_s} \left[b_s^{1/2} -{1\over2}d_s c_s(t_e-t_c) \right]^{2} \nonumber\\
&=&\frac{1}{m(n^2-4)}\left( \phi_c^{2-n} - \phi_i^{2-n}\right) +2\pi G \left( \phi_c^2 -\phi_e^2 \right) \ ,\end{aligned}$$ where $\phi_e$ is the value of scalar field at the end of inflation. Notice that the first term arises from the Lorentz violating stage.
Phase-space analysis {#secV}
====================
In this section, we investigate the global structure of the dynamical system via phase plane analysis and compute the cosmological evolution by numerical analysis. Introducing the following variables: $$\begin{aligned}
&&x\equiv{\phi'\over \sqrt{6\bar{\beta}}} \ , \qquad y\equiv\sqrt{{V\over 3H^2\bar{\beta}}} \ ,
\label{def-xy}\\
&& \lambda_1 \equiv -{\bar{\beta}_{,\phi}\over \sqrt{\bar{\beta}}} \ , \qquad \lambda_2 \equiv - \sqrt{\bar{\beta}}{V_{,\phi}\over V} \ ,
\label{def-lambda}\\
&& \Gamma_1 \equiv \frac{\bar{\beta} \bar{\beta}_{,\phi\phi}}{\bar{\beta}_{,\phi}^2} \ , \qquad \Gamma_2 \equiv \frac{V V_{,\phi\phi}}{V_{,\phi}^2} \ ,
\label{def-gamma}\end{aligned}$$ the Eqs. (\[eq:02\]) and (\[eq:03\]) can be written as a plane-autonomous system $$\begin{aligned}
&&x'=-3x(1-x^2)+\sqrt{{3\over 2}} (\lambda_1+\lambda_2)y^2 \ ,
\label{auto-x}\\
&& y' = \left[3x-\sqrt{{3\over 2}} (\lambda_1+\lambda_2)\right]xy \ ,
\label{auto-y}\\
&& \lambda_1^{\prime} =-\sqrt{6}\lambda_1^2 \left(\Gamma_1 -{1\over 2}\right)x \ ,
\label{auto-L1}\\
&&\lambda_2^{\prime} =-\sqrt{6}\lambda_2^2 \left[ \Gamma_2 -\left(1-{\lambda_1\over 2\lambda_2}\right) \right] x \ ,
\label{auto-L2}\end{aligned}$$ where the prime denotes a derivative with respect to the logarithm of the scale factor, $\alpha=\ln a$. The functions $\lambda_1(\phi)$ and $\lambda_2(\phi)$ determine a type of the coupling vector and the potential, respectively. The Friedmann constraint, Eq. (\[eq:01\]), takes the simple form $$\begin{aligned}
x^2 +y^2 =1 \ .
\label{constrps}\end{aligned}$$ The equation of state for the scalar field could be expressed in terms of the new variables as $$\begin{aligned}
\omega_\phi =\frac{p_\phi}{\rho_\phi}=\frac{x^2-y^2}{x^2+y^2} \ .\end{aligned}$$ Notice that $x^2$ measures the contribution to the expansion due to the scalar field kinetic energy and the coupling function, while $y^2$ measures the contribution to the expansion due to the potential energy and the the coupling function.
Equations (\[auto-x\])–(\[auto-L2\]) are written as an autonomous phase system of the form ${\bf x}'={\bf f}({\bf x})$ where ${\bf x}=(x,y,\lambda_1,\lambda_2)$. The use of this form for the dynamical equations allows the fixed points of the system to be readily identified, and the so-called critical points ${\bf
x}_0$ are solutions of the system of equations ${\bf f}({\bf
x}_0)=0$. To determine their stability we need to perform linear perturbations around the critical points in the form ${\bf x} =
{\bf x}_0+{\bf u}$, which results in the following equations of motion ${\bf u}' =M{\bf u}$, where $$\begin{aligned}
M_{ij} =\frac{\partial f_i}{\partial x_j}\Big |_{{\bf x}_0} \ .\end{aligned}$$ In the case of the dynamical equations (\[auto-x\])–(\[auto-L2\]), [**u**]{} is a 4-column vector consisting of the perturbations of $x$, $y$, $\lambda_1$ and $\lambda_2$. Thus, $M_{ij}$ is a $4 \times 4$ matrix. The stability of the critical points is determined by the eigenvalues $\mu_i$ of the matrix $M$ at the critical points. A non-trivial critical point is called stable (unstable) whenever the eigenvalues of $M$ are such that $Re(\mu_i) < 0$ ($Re(\mu_i) >
0$). If neither of the aforementioned cases are accomplished, the critical point is called a saddle point.
![The phase plane of Lorentz violating kinetic dominated solution for $m>3/8$. []{data-label="figure1"}](figure1.eps){height="10cm" width="12cm"}
In the following, we will study the simplest model, $$\begin{aligned}
\bar{\beta}(\phi)= m \phi^2 \ , \qquad V(\phi)= {1\over 2}M^2 \phi^2 \ ,
\label{model-ps}\end{aligned}$$ where $m$ and $M$ are parameters. Substituting Eqs. (\[model-ps\]) into Eqs. (\[def-lambda\]) and (\[def-gamma\]), respectively, we obtain $$\begin{aligned}
\lambda_1 = \lambda_2 = -2\sqrt{m} \ , \qquad \Gamma_1=\Gamma_2={1\over 2} \ ,\end{aligned}$$ and Eqs. (\[auto-L1\]) and (\[auto-L2\]) are trivially satisfied. In the former, Eqs. (\[auto-x\]) and (\[auto-y\]) can be fused into the single equation, $$\begin{aligned}
x'&=&-\left[3x-\sqrt{{3\over 2}} (\lambda_1+\lambda_2)\right](1-x^2) \nonumber\\
&=& -\left(3x+2\sqrt{6m}\right)(1-x^2) \ ,
\label{auto-single}\end{aligned}$$ which is one dimensional phase-space corresponding to the unit circle. Critical points correspond to fixed points where ${\bf
x}'=0$, and there are Lorentz violation self-similar solutions with $$\begin{aligned}
{H'\over H} &=& -3x^2 +\sqrt{6}\lambda_1 x \ .
\label{self-sim}\end{aligned}$$ Note that the second term arises from Lorentz violation. Applying the above procedure, setting ${\bf x}'=0$, the critical points $(x_0,y_0)$ of the system are $(1,0)$, $(-1,0)$, and $(-\sqrt{8m/3},\sqrt{1-8m/3})$. For any form of the potential in the Lorentz violation stage, the critical points $(1,0)$ or $(-1,0)$ correspond to two Lorentz violation kinetic-dominated solutions. Then, the critical point $(-\sqrt{8m/3},\sqrt{1-8m/3})$ corresponds to a Lorentz violation potential-kinetic solution. Integration of Eq. (\[self-sim\]) with respect to $\alpha$ will show that all critical points, ${\bf x}_0$, correspond to the Hubble parameter $$\begin{aligned}
H \propto \exp \left(-{\alpha\over p}\right) \ .\end{aligned}$$
![The phase plane of Lorentz violating kinetic-potential solution for $m<3/8$. []{data-label="figure2"}](figure2.eps){height="10cm" width="12cm"}
This relates to an expanding universe with a scale factor $a(t)$ given by $a(t)\sim t^p$, where $$\begin{aligned}
p \equiv {1\over 3x^2_0 -\sqrt{6}\lambda_1 x_0} = {1\over 3x^2_0 +2\sqrt{6m}x_0} \ .
\label{slope:p}\end{aligned}$$ The linear perturbation about the points $x_{0+}=+1$ and $x_{0-}=-1$ give the eigenvalues $\mu_{+}=6+4\sqrt{6m}$ and $\mu_{-}=6-4\sqrt{6m}$, respectively. Thus for positive $m$, $x_{0+}=+1$ is always unstable and $x_{0-}=-1$ is stable for $m>3/8$ but unstable for $m<3/8$. Moreover, in the linear perturbation about the Lorentz violation potential-kinetic solution, we obtain the eigenvalue $\mu=8m-3$. The solution is stable for $m<3/8$. In Figs. \[figure1\] and \[figure2\], we show the phase plane plot for $m>3/8$ and $m<3/8$. We note that the trajectories are confined inside the circle given by $x^2+y^2=1$.
Another remarkable feature of the above model is that the equation of state is given by $$\begin{aligned}
\omega_\phi = -1 + {16\over 3} m \ ,\end{aligned}$$ completely determined by the parameter $m$ of the coupling vector. Thus, we always have $\omega_\phi >-1$ for ordinary scalar field.
Conclusions
===========
In this paper, we have studied the dynamics of a scalar field in the Lorentz violating scalar-vector-tensor theory of gravity, taking into account the effect of the power-law effective coupling vector. Since the effective coupling vector be dynamics variable, the equation of state is dependent on the coupling parameter. For the model with the power law Hubble parameter and coupling vector, we find an exact solution of the equation of state. A constant equation of state corresponds to $n=2$ while for $n >2$ leads to dynamics equation of states. In this case, the scalar fields are completely associated with the Lorentz violation.
Also, the different form in the coupling vector and the potential models lead to the different qualitative evolution in two regimes of inflation. The results show that, for the inverse power-law potential, the Hubble parameter decreases during the Lorentz violation stage and increases in the standard slow roll stage. For the power-law potential, the Hubble parameter increases during the Lorentz violation stage but it decreases in the standard slow roll stage.
From the qualitative study of the dynamical system, we have demonstrated the attractor behavior of inflation driven by a scalar field in the context of scalar-vector-tensor theory of gravity. We have found that there exists the Lorentz violating kinetic dominated solution and the Lorentz violating potential-kinetic dominated solution, depending on the region of the coupling parameter in the simplest Lorentz violating chaotic inflation model. The quadratic coupling vector and the chaotic potential correspond to the constants $\lambda_1=\lambda_2=-2\sqrt{m}$ and $\Gamma_1=\Gamma_2=1/2$. There are two important results of this study, which are different from the scalar-tensor theory of gravity: the condition for the accelerating universe, Eq. ([\[self-sim\]]{}) and the slope, $p$, Eq. ([\[slope:p\]]{}). The first one yields $\lambda_1{\bf x}_0 >\sqrt{3/8}(\omega_\phi+1/3)$. The analysis of the critical points show that we may obtain an accelerated expansion provided that the solutions are approaching the Lorentz violation kinetic dominated solution with $m>1/6$ and approaching the Lorentz violation potential-kinetic dominated solution with $m<3/8$. When the accelerating condition is satisfied, the slope $p$ characterizes the properties of the inflating universe: power-law inflation ($p>0$), de Sitter inflation ($p=0$) and superinflation ($p\equiv -|p|<0$). In other cases, if $\lambda_1$ and $\lambda_2$ are constants, one finds that the coupling vector is still quadratic in scalar field, $\bar{\beta}\sim \phi^2$, while the potential as a function of scalar field $\phi$ is given by a power-law potential, $V(\phi) \sim \phi^{2\gamma}$ and $\Gamma_1=1/2$, $\Gamma_2=1-1/2\gamma$, where $\gamma = \lambda_2/\lambda_1$. Moreover, in order to obtain dynamical evolution of the system, we need to solve Eqs. (\[auto-L1\]) and (\[auto-L2\]) together with Eqs. (\[auto-x\]) and (\[auto-y\]). For a realistic model, the effect of an additional component (matter field) would be interesting [@ari].
Finally, we would like to emphasize that there exists an attractor solution in the Lorentz violating scalar-vector-tensor theory of gravity.
[99]{}
A. Linde, *Particle Physics and Inflationary Cosmology*, (Harwood academic publishers, 1980); E. W. Kolb, and M. S. Turner, *The Early Universe*, (Perseus Publishing, 1990); A. R. Liddle, and D. H. Lyth, *Cosmological Inflation and Large-Scale Structure*, (Cambridge University Press, Cambridge, 2000).
A Vilenkin, *Cosmic Strings and Domain Walls*, Phys.Rep. [**121**]{}, 263 (1985); T. W. B. Kibble, *Topology of Cosmic Domains and Strings*, J. Phys. [**A9**]{}, 1387 (1976); A. Vilenkin and E. P. S. Shellard, *Cosmic strings and other topological defects*, Cambridge Univ. Press (Cambridge 1994); M. B. Hindmarsh and T. W. Kibble, *Cosmic strings*, Rept. Prog. Phys. [**58**]{}, 477 (1995) \[arXiv:hep-ph/9411342\].
M. B. Green, J. H. Schwarz and E. Witten, *Superstring Theory*, Vol. 1 and 2, Cambridge Univ. Press (Cambridge 1987).
T. Barreiro, B. de Carlos and E. J. Copeland, *Stabilizing the Dilaton in Superstring Cosmology*, Phys. Rev. D [**58**]{}, 083513 (1998) \[arXiv:hep-th/9805005\]; G. Huey, P. J. Steinhardt, B. A. Ovrut and D. Waldram, *A Cosmological Mechanism for Stabilizing Moduli*, Phys. Lett. B [**476**]{}, 379 (2000) \[arXiv:hep-th/0001112\]; T. Barreiro, B. de Carlos and N. J. Nunes, *Moduli Evolution in Heterotic Scenarios*, Phys. Lett. B [**497**]{}, 136 (2001) \[arXiv:hep-ph/0010102\].
A. Sen, *Rolling Tachyon*, JHEP [**0204**]{}, 048 (2002) \[arXiv:hep-th/0203211\]; A. Sen, *Tachyon Matter*, JHEP [**0207**]{}, 065 (2002) \[arXiv:hep-th/0203265\].
G. W. Gibbons, *Cosmological Evolution of the Rolling Tachyon*, Phys. Lett. B [**537**]{}, 1 (2002).
A. Sen, *Supersymmetric World-volume Action for Non-BPS D-branes*, JHEP [**9910**]{}, 008 (1999) \[arXiv:hep-th/9909062\]; M. Garousi, *Tachyon couplings on non-BPS D-branes and Dirac-Born-Infeld action*, Nucl. Phys. B [**584**]{}, 284 (2000) \[arXiv:hep-th/0003122\]; E. Bergshoeff, M. de Roo, T. de Wit, E. Eyras and S. Panda, *T-duality and Actions for Non-BPS D-branes*, JHEP [**0005**]{}, 009 (2000) \[arXiv:hep-th/0003221\]; J. Kluson, *Proposal for non-BPS D-brane action*, Phys. Rev. D [**62**]{}, 126003 (2000) \[arXiv:hep-th/0004106\].
L. Kofman and A. Linde, *Problems with Tachyon Inflation*, JHEP [**07**]{}, 004 (2002). \[arXiv:hep-th/0205121\]; A. Frolov, L. Kofman and A. Starobinsky, *Prospects and Problems of Tachyon Matter Cosmology*, Phys.Lett. B [**545**]{} 8 (2002) \[arXiv:hep-th/0204187\].
A.G. Riess *et al*., *Type Ia Supernova Discoveries at $z>1$ From the Hubble Space Telescope: Evidence for Past Deceleration and Constraints on Dark Energy Evolution*, Astrophys. J. [**607**]{}, 665 (2004) \[arXiv:astro-ph/0402512\].
H. Jassal, J. Bagla and T. Padmanabhan, *The vanishing phantom menace*, \[arXiv:astro-ph/0601389\].
C. L. Bennett *et al*., *First Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Preliminary Maps and Basic Results*, Astrophys. J. Suppl. [**148**]{}, 1 (2003) \[arXiv:astro-ph/0302207\].
C. Wetterich, *Cosmology and the fate of dilatation symmetry*, Nucl. Phys. B [**302**]{}, 668 (1988); P. J. E. Peebles and B. Ratra, *Cosmology with a time-variable cosmological constant*, Astrophys. J. [**325**]{}, L17 (1988); B. Ratra and P. J. E. Peebles, *Cosmological Consequences of a Rolling Homogeneous Scalar Field*, Phys. Rev. D [**37**]{}, 3406 (1988); J. A. Frieman, C. T. Hill, A. Stebbins and I. Waga, *Cosmology with Ultra-light Pseudo-Nambu-Goldstone Bosons*, Phys. Rev. Lett. [**75**]{}, 2077 (1995) \[arXiv:astro-ph/9505060\]; R. R. Caldwell, R. Dave and P. J. Steinhardt, *Cosmological Imprint of an Energy Component with General Equation of State*, Phys. Rev. Lett. [**80**]{}, 1582 (1998) \[arXiv:astro-ph/9708069\]; I. Zlatev, L. Wang and P. J. Steinhardt, *Quintessence, Cosmic Coincidence, and the Cosmological Constant*, Phys. Rev. Lett. [**82**]{}, 896 (1999) \[arXiv:astro-ph/9807002\].
C. Armendariz-Picon, T. Damour and V. Mukhanov, *k-Inflation*, Phys. Lett. B [**458**]{}, 209 (1999) \[arXiv:hep-th/9904075 \]; C. Armendariz-Picon V. Mukhanov and P.J. Steinhardt, *A Dynamical Solution to the Problem of a Small Cosmological Constant and Late-time Cosmic Acceleration*, Phys. Rev. Lett. [**85**]{}, 4438 (2000) \[arXiv:astro-ph/0004134\]; T. Chiba, T. Okabe and M. Yamaguchi, *Kinetically Driven Quintessence*, Phys. Rev. D [**62**]{}, 023511 (2000) \[arXiv:astro-ph/9912463\].
H. Sato, *Extremely high energy and violation of Lorentz invariance*, \[arXiv:astro-ph/0005218\].
S. R. Coleman and S. L. Glashow, *High-energy tests of Lorentz invariance*, Phys. Rev. D [**59**]{}, 116008 (1999) \[arXiv:hep-ph/9812418\].
S. Kanno and J. Soda, *Lorentz violating inflation*, Phys.Rev. D [**74**]{}, 063505 (2006) \[arXiv:hep-th/0604192\].
S. M. Carroll and E. A. Lim, *Lorentz-violating vector fields slow the universe down*, Phys. Rev. D [**70**]{}, 123525 (2004) \[arXiv:hep-th/0407149\].
P. J. Steinhardt L. Wang, and I. Zlatev, *Cosmological Tracking Solutions*, Phys. Rev. D [**59**]{}, 123504 (1999) \[arXiv:astro-ph/9812313\].
I. Zlatev and P. J. Steinhardt, *A tracker solution to the cold dark matter cosmic coincidence problem*, Phys. Lett. B [**459**]{}, 570 (1999) \[arXiv:astro-ph/9906481\].
J. D. Barrow, *Graduated inflationary universe*, Phys. Lett. B [**235**]{}, 40 (1990).
P. Binetruy, *Models of dynamical supersymmetry breaking and quintessence*, Phys. Rev. D [**60**]{}, 063502 (1999) \[arXiv:hep-ph/9810553\]; P. Brax and J. Martin, *The robustness of quintessence*, Phys. Rev. D [**61**]{}, 103502 (2000) \[arXiv:astro-ph/9912046\].
A. Balbi, C. Baccigalupi, S. Matarrese, F. Perrotta, and N. Vittorio, *Implications for quintessence models from MAXIMA-1 and BOOMERANG-98*, Astrophys. J. [**547**]{}, L89 (2001) \[arXiv:astro-ph/0009432\].
Arianto and Freddy P. Zen, in preparation.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Catharina Graafland$^{1}$, José M. Gutiérrez$^1$, Juan M. López$^1$,\
Diego Pazó$^1$ & Miguel A. Rodr[í]{}guez$^1$
date: ' $^1$Instituto de F[í]{}sica de Cantabria, CSIC–Universidad de Cantabria, Avenida de Los Castros, E-39005 Santander, Spain\'
title: |
The Probabilistic Backbone of Data-Driven Complex Networks: An example in Climate:\
Supplementary Figures
---
![Maps for Correlation Networks (CNs) of size (**a**) 1,901, (**b**) 3,118, (**c**) 5,086 and (**d**) 8,284 and (**e**) Bayesian Network of size 1,796 in which gridboxes are coloured in function of their betweenness values. Raw betweenness values $BC_i$ are transformed to $\ln(1+BC_i)$ and every gridbox presents the mean of the betweenness values of its direct neighbors for visualization purposes. CN maps (**a**)-(**d**) show alternation in the assignation of regions with high and low betweenness. The alternation between the CN maps of size (**c**) 5,086 and (**d**) 8,284 is small and one deducts a more ore less established pattern. The white boxes (zero betweenness) in CNs of size (**a**) 1,901 and (**b**) 3,118 mostly indicate variables that are unconnected to the network (see largest connected component size in Supplementary Figures 2 and 3). The white boxes in other maps indicate variables that do not belong to any geodesic. CN maps (**c**) and (**d**) share high betweenness regions in the mid-east Pacific Ocean, the Northern part of South America together with the Caribbeans, the South-West part of the Indian Ocean, the Philippines and the Chinese Sea and part of the North Atlantic Ocean against Greenland. They also share low betweenness regions for the mid-west Pacific Ocean, the west Pacific Ocean against Mexico, the Eastern part of the Indian Ocean against Australia and South-East part of the Pacific Ocean on the height of Chili. The BN map in (**e**) coincides on the above mentioned high betweenness regions. The pattern of the BN is little more flattened due to a lower value of the clustering coefficient as displayed in Figure 3.[]{data-label="fig:betweeness"}](sup_fig1){width="\textwidth"}
![The diameter of a network is the length of the longest geodesic (shortest path) in the network. For a graph that is not fully connected we define the diameter as the length of the longest geodesic of the Largest Connected Component (LCC). The figure displays the diameter (left y-axis, dots) and the size of the largest connected component (right y-axis, dashed lines) versus the number of edges in Correlation Networks (CNs; magenta), Bayesian Networks (BNs; green), Erdös-Renyi random graphs (ERs; yellow) and Regular lattices (RLs; blue). ERs are random graphs in which every arc is included with probability $2|E|/(N(N-1))$ with $N$ the total number of variables ($N = 648$) and $|E|$ the number of edges in the graph. RLs are deterministic graphs and augment locally. The smallest RL corresponds to a network in which all variables are connected with their direct neighbours in a $36 \times 18$ rectangular grid (order 1 connection). The second RL corresponds to a network in which nodes are connected with direct neighbours and neighbours of direct neighbours (order 2 connection). And so on.\
A network is fully connected if the size of the LCC equals $N = 648$. All RLs are fully connected by construction. BNs, ERs and CNs are fully connected at sizes $|E| =$ 1,000, 1,900 and 5,750, respectively.\
The maximum diameter value for ERs is 33 at size 500, is 65 for BNs at size 500 and is 70 for CNs at size 2,200. All maximum diameter values are found when the network is almost fully connected. BNs and CNs have similar maxima, but as BNs connect earlier, the size of the BN that yields the maximum diameter is four times smaller than that of the CN. BN diameter values first tend towards those of a local lattice (at size 500), but then tend to efficient ER values (around size 2,000). CN diameter values increase much slower during the network construction process. Values are similar to local lattice structure up to 3,500 edges and do not approach the efficient ER structure up to networks of 8,000 edges and above. []{data-label="fig:diameter"}](sup_fig2){width="\textwidth"}
![The global triangular clustering of a network is the ratio between the number of triangles and the total number of possible triangles that could exist: $\text{clustering} = \text{number of closed paths of length two} \div \text{number of paths of length two}.$ The figure displays the clustering (left y-axis, dots) and size of the Largest Connected Component (LCC, right y-axis, dashed lines) versus number of edges in Correlation Networks (CNs; magenta), Bayesian Networks (BNs; green), Erdös-Renyi random graphs (ERs; yellow) and Regular lattices (RLs; blue). ERs are random graphs in which every arc is included with probability $2|E|/(N(N-1))$ with $N$ the total number of variables ($N = 648$) and $|E|$ the number of edges in the graph. RLs are deterministic graphs and augment locally. The smallest RL corresponds to a network in which all variables are connected with their direct neighbors (rectangular grid; order 1 connection). The second RL corresponds to a network in which nodes are connected with direct neighbors and neighbors of direct neighbors (order 2 connection). And so on.\
ERs do not possess intrinsic clustering structure, and the observed minimal grow of clustering values is only due to network saturation. RLs do have a local clustering structure. This contribution to the clustering coefficients is totally local by definition. Clustering coefficients grow in function of size. CNs of all sizes posses high clustering values. A peak value of almost 1 is obtained around 1,700 edges, even in the correlation range with relatively more large distance links clustering values remain high; clustering in CNs occurs at local and global scale (see Figure 1(**b**-**c**)). BNs posses relative low clustering coefficients when compared with CNs of similar size, however the values are significantly higher then the coefficients of ERs that do not posses clustering structure at all. The peak value 0.35 ($\approx$ one out three connected triples is triangular) belongs to a graph that contains around 500 edges. This graph is almost fully connected, but the structure is still local. BNs with more edges have lower clustering coefficients. The BN of 1,796 edges (see corresponding network in Figure 1(**c**) has a clustering coefficient of 0.2; this BN does exhibit clustering, but the long range edges do not contribute positively to the value of the clustering coefficient as the connection between two locally clustered regions is captured with a minimal amount of long distant edges, instead of an redundant edge bundel as is the case for CNs.[]{data-label="fig:clustering"}](sup_fig3){width="\textwidth"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $k$ be an algebraically closed field of characteristic $p>0$ and $C$ a connected nonsingular projective curve over $k$ with genus $g \geq 2$. This paper continues the work begun in [@LM], namely the study of “big actions”, i.e. the pairs $(C,G)$ where $G$ is a $p$-subgroup of the $k$-automorphism group of $C$ such that$\frac{|G|}{g} >\frac{2\,p}{p-1}$. If $G_2$ denotes the second ramification group of $G$ at the unique ramification point of the cover $C \rightarrow C/G$, we display necessary conditions on $G_2$ for $(C,G)$ to be a big action, which allows us to pursue the classification of big actions. Our main source of examples comes from the construction of curves with many rational points using ray class field theory for global function fields, as initiated by J-P. Serre and followed by [@Lau] and [@Au]. In particular, we obtain explicit examples of big actions with $G_2$ abelian of large exponent.'
author:
- 'Michel Matignon and Magali Rocher.'
title: 'On smooth curves endowed with a large automorphism $p$-group in characteristic $p>0$.'
---
Introduction.
=============
*Setting.* Let $k$ be an algebraically closed field and $C$ a connected nonsingular projective curve over $k$, with genus $g \geq 2$. If $char(k)=0$, Hurwitz exhibits a linear bound for the $k$-automorphism group of the curve $C$, namely: $|Aut_k(C)| \leq 84\, (g-1)$. If $char(k)=p>0$, the Hurwitz bound is no longer true when $g$ grows large, but the finiteness result still holds (cf. [@Sch]) and one gets polynomial bounds on $|Aut_k(C)|$ (cf. [@St] and [@Sin]). In this situation, the full automorphism group may be very large as compared with the case $char(k)=0$. This is due to the appearance of wild ramification, which leads us to concentrate on large automorphism p-groups in $char(k)=p>0$. In this spirit, Nakajima (cf. [@Na]) studies the size of Sylow $p$-subgroups of $Aut_k(C)$ and emphasizes the influence of another important invariant of the curve: the $p$-rank, denoted by $\gamma$. Indeed, if $G$ is a Sylow $p$-subgroup of $Aut_k(C)$, we deduce from [@Na] that $|G| \leq \frac{2\,p}{p-1} \, g$, except for $\gamma=0$. On the contrary, when $\gamma=0$, the upper bound on $|G|$ is no more linear in $g$, namely $|G| \leq \max \{g, \frac{4\,p}{(p-1)^2} \, g^2\}$. As shown in [@St], the quadratic upper bound $\frac{4\,p}{(p-1)^2} \, g^2$ can really be attained, which demonstrates that, in this case, $Aut_k(C)$ may be especially large. Following Nakajima’s work, Lehr and Matignon explore the “big actions”, that is to say the pairs $(C,G)$ where $G$ is a $p$-subgroup of $Aut_k(C)$ such that $\frac{|G|}{g}> \frac{2\,p}{p-1}$ (see [@LM]). In particular, they exhibit a classification of the big actions that satisfy $\frac{4}{(p-1)^2} \leq \frac{|G|}{g^2}$.
*Motivation and outline of the paper.* Let $(C,G)$ be a big action. As shown in [@LM], there is a point of $C$, say $\infty$, such that $G$ is equal to the wild inertia subgroup $G_1$ of $G$ at $\infty$. Let $G_2$ be the second ramification group of $G$ at $\infty$ in lower notation. Then, the quotient curve $C/G_2$ is isomorphic to the projective line ${\mathbb P}_k^1$ and the quotient group $G/G_2$ acts as a group of translations of ${\mathbb P}_k^1$ fixing $\infty$, through $X \rightarrow X+y$, where $y$ runs over a subgroup $V$ of $k$. In this way, the group $G$ appears as an extension of $G_2$ by the $p$-elementary abelian group $V$ via the exact sequence: $$0 \longrightarrow G_2 \longrightarrow G=G_1 \longrightarrow V \simeq
({\mathbb Z}/\,p\, {\mathbb Z})^v \longrightarrow 0$$ The aim of this paper is, on the one hand, to give necessary conditions on $G_2$ for $(C,G)$ to be a big action and, on the other hand, to display realizations of big actions with $G_2$ abelian of large exponent. In section 2, we first prove that $G_2$ must be equal to $D(G)$, the commutator subgroup of $G$. In section 3, given a big action $(C,G)$ and an additive polynomial map: ${\mathbb P}_k^1 \rightarrow C/G_2 \simeq {\mathbb P}_k^1$, we display a new big action $(\tilde{C}, \tilde{G})$ such that $\tilde{G}_2\simeq G_2$. In section 4, we demonstrate that $G_2$ cannot be cyclic except when $G_2$ has order $p$. Some of these results on $G_2$ are necessary to pursue the classification of big actions initiated by Lehr and Matignon, more precisely to explore the case: $\frac{4}{(p^2-1)^2} \leq \frac{|G|}{g^2}$. Indeed, we prove in section 5 that such an inequality requires $G_2$ to be an elementary abelian $p$-group whose order divides $p^3$. In sequel papers, M. Rocher goes further: she studies big actions with a $p$-elementary abelian $G_2$ (see [@MR2]), which enables her to display the classification of big actions satisfying $\frac{4}{(p^2-1)^2} \leq \frac{|G|}{g^2}$ (see [@MR3]). In section 6, following [@Lau] and [@Au], we consider the maximal abelian extension of $K:={\mathbb F}_q(X)$ ($q=p^e$) denoted by $K_S^m$, which is unramified outside $X=\infty$, completely split over the set $S$ of the finite rational places and whose conductor is smaller than $m \, \infty$, with $m \in {\mathbb N}$. Class field theory gives a description of the Galois group $G_S(m)$ of this extension, but also precises its upper ramification groups, which allows us to compute the genus of the extension. Moreover, it follows from the unicity and the maximality of $K_S^m$ that the group of translations $\{X \rightarrow X+y, \, y \in {\mathbb F}_q \}$ extends to a $p$-group of ${\mathbb F}_q$-automorphisms of $K_S^m$, say $G(m)$, with the exact sequence: $$0 \longrightarrow G_S(m) \longrightarrow G(m) \longrightarrow \, {\mathbb F}_q \longrightarrow 0$$ This provides examples of big actions with $G_2$ abelian of exponent as large as we want, but also relates the problem of big actions to the search of algebraic curves with many rational points compared with their genera. In particular, we conclude section 6 by exhibiting specific $K$-subextensions of $K_S^m$, for a well-chosen conductor $m\,\infty$, giving examples of big actions such that $G_2 \simeq {\mathbb Z}/p^2{\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t$ with a small $p$-rank, namely $t=O(log_p \,g)$. In the final section, we use Katz-Gabber theorem to highlight the link between big actions on curves and an analogous ramification condition for finite $p$-groups acting on $k((z))$.
*Notation and preliminary remarks.* Let $k$ be an algebraically closed field of characteristic $p>0$. We denote by $F$ the Frobenius endomorphism for a $k$-algebra. Then, $\wp$ means the Frobenius operator minus identity. We denote by $k\{F\}$ the $k$-subspace of $k[X]$ generated by the polynomials $F^i(X)$, with $i \in {\mathbb N}$. It is a ring under the composition. Furthermore, for all $\alpha$ in $k$, $F\, \alpha=\alpha^p\,F$. The elements of $k\{F\}$ are the additive polynomials, i.e. the polynomials $P(X)$ of $k[X]$ such that for all $ \alpha$ and $\beta$ in $k$, $P(\alpha+ \beta) = P(\alpha)+ P(\beta)$. Moreover, a separable polynomial is additive if and only if the set of its roots is a subgroup of $k$ (see [@Go] chap. 1).\
Let $f(X)$ be a polynomial of $k[X]$. Then, there is a unique polynomial $red (f)(X)$ in $k[X]$, called the reduced representative of $f$, which is $p$-power free, i.e. $red(f)(X) \in \bigoplus_{(i,p)=1} k\, X^i$, and such that $red(f)(X)=f(X)$ mod $\wp (k[X]).$ We say that the polynomial $f$ is reduced mod $\wp(k[X])$ if and only if it coincides with its reduced representative $red(f)$. The equation $W^p-W=f(X)$ defines a $p$-cyclic étale cover of the affine line that we denote by $C_f$. Conversely, any $p$-cyclic étale cover of the affine line $Spec \, k[X]$ corresponds to a curve $C_f$ where $f$ is a polynomial of $k[X]$ (see [@Mi] III.4.12, p. 127). By Artin-Schreier theory, the covers $C_f$ and $C_{red(f)}$ define the same $p$-cyclic covers of the affine line. The curve $C_f$ is irreducible if and only if $red(f) \neq 0$.\
Throughout the text, $C$ always denotes a nonsingular smooth projective curve with genus $g$ and $Aut_k(C)$ means its $k$-automorphism group. Our main references for ramification theory are [@Se] and [@Au].
First results on “big actions”.
===============================
To precise the background of our work, we begin by collecting and completing the first results on big actions already obtained in [@LM]. The expression “big actions” stands for curves endowed with a big automorphism $p$-group. The first task is to recall what we mean by “big”.
Let $G$ be a subgroup of $Aut_k (C).$ We say that the pair $(C,G)$ is a big action if $G$ is a finite $p$-group, if $g \neq 0$ and if $$\label{N}
\frac{ |G|}{g} > \frac{2\,p}{p-1}$$
[@LM] Assume that $(C,G)$ is a big action with $g \geq 2$. Then, there is a point of $C$ (say $\infty$) such that $G$ is the wild inertia subgroup of $G$ at $\infty$: $G_1$. Moreover, the quotient $C / G$ is isomorphic to the projective line ${\mathbb P}^1_k$ and the ramification locus (respectively branch locus) of the cover $\pi: \, C \rightarrow C/G$ is the point $\infty$ (respectively $\pi(\infty))$. For all $i \geq 0$, we denote by $G_i$ the $i$-th lower ramification group of $G$ at $\infty$. Then,
1. $G_2$ is non trivial and it is strictly included in $G_1$.
2. The Hurwitz genus formula applied to $C \rightarrow C/G$ reads: $$\label{genus}
2 \, g = \, \sum_{i \geq 2} (|G_i|-1)$$
3. The quotient curve $C/G_2$ is isomorphic to the projective line ${\mathbb P}_k^1.$ Moreover, the quotient group $G /G_2$ acts as a group of translations of the affine line $C/G_2 - \{\infty \}=Spec \,k[X]$, through $X \rightarrow X+y$, where $y$ runs over a subgroup $V$ of $k$. Then, $V$ is an ${\mathbb F}_p$-subvector space of $k$. We denote by $v$ its dimension. Thus, we obtain the exact sequence: $$0 \longrightarrow G_2 \longrightarrow G=G_1 \stackrel{\pi}{\longrightarrow} V \simeq
({\mathbb Z}/\,p\, {\mathbb Z})^v \longrightarrow 0$$ where $$\pi:
\left\{
\begin{array}{ll}
G \rightarrow V\\
g \rightarrow g(X)-X
\end{array}
\right.$$
4. Let $H$ be a normal subgroup of $G$ such that $g_{C/H} >0$. Then, $(C/H, G/H)$ is also a big action. Moreover, the group $G/H$ fixes the image of $\infty$ in the cover $C \rightarrow C/H$. In particular, if $g_{C/H}=1$, then $p=2$, $C/H$ is birational to the curve $W^2+W=X^3$ and $G/H$ is isomorphic to $Q_8$, the quarternion group of order $8$ (see [@Si], Appendix A, Prop. 1.2).
Note that, for $g=1$, one can find big actions $(C,G)$ such that $G$ is not included in a decomposition group of $Aut_k(C)$ as in Proposition 2.2.
The following lemma generalizes and completes the last point of Proposition 2.2.
Let $G$ a finite $p$-subgroup of $Aut_k(C)$. We assume that the quotient curve $C/G$ is isomorphic to ${\mathbb P}_k^1$ and that there is a point of $C$ (say $\infty$) such that $G$ is the wild inertia subgroup of $G$ at $\infty$: $G_1$. We also assume that the ramification locus (respectively branch locus) of the cover $\pi: \, C \rightarrow C/G$ is the point $\infty$ (respectively $\pi(\infty))$. Let $G_2$ be the second ramification group of $G$ at $\infty$ and $H$ a subgroup of $G$.
1. Then, $C/H$ is isomorphic to ${\mathbb P}_k^1$ if and only if $H \supset G_2$.
2. In particular, if $(C,G)$ is a big action with $g \geq 2$ and if $H$ is a normal subgroup of $G$ such that $H \subsetneq G_2$, then $g_{C/H}>0$ and $(C/H,G/H)$ is also a big action.
**Proof**: When applied to the cover $C \rightarrow C/G \simeq {\mathbb P}_k^1$, the Hurwitz genus formula (see e.g. [@St93]) reads: $2(g-1)=2|G|\, (g_{C/G}-1) + \sum_{i\geq 0}\, (|G_i|-1)$. When applied to the cover $C \rightarrow C/H$, it yields: $2(g-1)=2|H|\, (g_{C/H}-1) + \sum_{i\geq 0} \,(|H \cap G_i|-1)$. Since $H \subset G=G_0=G_1$, it follows that: $$2 |H| g_{C/H}= -\, 2(|G|-|H|) + \sum_{i\geq 0} \, (|G_i|-|H \cap G_i|)= \sum_{i\geq 2} \,\, (|G_i|-|H \cap G_i|)$$ Therefore, $g_{C/H}=0$ if and only if for all $i \geq 2$, $G_i=H \cap G_i$, i.e. $G_i \subset H$, which is equivalent to $G_2 \subset H$. The second point then derives from Proposition 2.2.4. $\square$
The very first step to study big actions is to precise their description when $G_2 \simeq {\mathbb Z}/p {\mathbb Z}$. The following proposition aims at gathering and reformulating the results already obtained for this case in [@LM] (cf. Prop. 5.5, 8.1 and 8.3).
[@LM]. Let $(C,G)$ be a big action, with $g \geq 2$, such that $G_2 \simeq {\mathbb Z}/p {\mathbb Z}$.
1. Then, $C$ is birational to the curve $C_f: \, W^p-W=f(X)=X\,S(X)+c\, X \in k[X]$, where $S$ in $k\{F \}$ is an additive polynomial with degree $s\geq 1$ in $F$. If we denote by $m$ the degree of $f$, then $m=1+p^s=i_0$, where $i_0 \geq 2$ is the integer such that: $$G=G_0=G_1 \supsetneq G_2 = G_3=\cdots=G_{i_0} \supsetneq G_{i_0+1} =\cdots=$$
2. Write $ S(F) = \sum _{j=0}^{s} a_j F^j$, with $a_{s} \neq 0$. Then, following [@El] (section 4), we can define an additive polynomial related to $f$, called the “palindromic polynomial” of $f$: $$Ad_{f}:= \frac{1}{a_{s}} \, F^{s} \, (\sum _{j=0}^ {s} \,a_j\, F^j + F^{-j}\, a_j)$$ The set of roots of $Ad_f$, denoted by $Z(Ad_f)$, is an ${\mathbb F}_p$-subvector space of $k$, isomorphic to $({\mathbb Z}/p{\mathbb Z})^{2s}$. Besides, $Z(Ad_f)= \{y \in k,\, f(X+y)-f(X)=0 \, \mod \, \wp(k[X] \}$.
3. Let $G_{\infty,1}$ be the wild inertia subgroup of $Aut_k(C)$ at $\infty$. Then, $G_{\infty,1}$ is a central extension of $ {\mathbb Z}/p {\mathbb Z}$ by the elementary abelian $p$-group $Z(Ad_f)$ which can be identified with a subgroup of translations $ \{ X \rightarrow X+y ,\, \, y \in k \}$ of the affine line. Furthermore, if we denote by $Z(G_{\infty,1})$ the center of $G_{\infty,1}$ and by $D(G_{\infty,1})$ its commutator subgroup, $Z(G_{\infty,1})=D(G_{\infty,1})=<\sigma>$, where $\sigma(X)=X$ and $\sigma(W)=W+1$. Thus, we get the following exact sequence: $$0 \longrightarrow Z(G_{\infty,1})=D(G_{\infty,1})\simeq {\mathbb Z}/p {\mathbb Z}\longrightarrow G_{\infty,1} \stackrel{\pi}{\longrightarrow} Z(Ad_f) \simeq ({\mathbb Z}/ p {\mathbb Z}) ^{2s} \longrightarrow 0$$ where $$\pi:
\left\{
\begin{array}{lc}
G_{\infty,1} \rightarrow Z(Ad_f) \simeq ({\mathbb Z}/ p {\mathbb Z}) ^{2s}\\
g \rightarrow g(X)-X
\end{array}
\right.$$ For $p>2$, $G_{\infty,1}$ is the unique extraspecial group with exponent $p$ and order $p^{2s+1}$. The case $p=2$ is more complicated (see [@LM] 4.1).
4. There exists an ${\mathbb F}_p$-vector space $V \subset Z(Ad_f) \simeq ({\mathbb Z}/ p {\mathbb Z}) ^{2s}$ such that $G=\pi^{-1}(V) \subset G_{\infty,1}$ and such that we get the exact sequence: $$0 \longrightarrow G_2 \simeq {\mathbb Z}/p {\mathbb Z}\longrightarrow G \stackrel{\pi}{\longrightarrow} V \longrightarrow 0.$$
Therefore, the key idea to study big actions is to use Proposition 2.2.4 and Lemma 2.4.2 to go back to the well-known situation described above. This motivates the following
Let $(C,G)$ be a big action with $g \geq 2$. Let $\mathcal{G}$ be a normal subgroup in $G$ such that $\mathcal{G}$ is strictly included in $G_2$. Then, there exists a group $H$, normal in $G$, such that $\mathcal{G} \subset H \subsetneq G_2$ and $[G_2:H]=p$. In this case, $(C/H,G/H)$ enjoys the following properties.
1. The pair $(C/H,G/H)$ is a big action and the exact sequence of Proposition 2.2: $$0 \longrightarrow G_2 \longrightarrow G \stackrel{\pi}{\longrightarrow} V \longrightarrow 0$$ induces the following one: $$0 \longrightarrow G_2/H=(G/H)_2 \simeq {\mathbb Z}/p{\mathbb Z}\longrightarrow G/H \stackrel{\pi}{\longrightarrow} V \longrightarrow 0$$
2. The curve $C/H$ is birational to $C_f$: $W^p-W=f(X)=X \,S(X)+c\,X \in k[X]$, where $S$ is an additive polynomial of degree $s \geq 1$ in $F$. Let $Ad_f$ be the palindromic polynomial related to $f$ as defined in Proposition 2.5. Then, $V \subset Z(Ad_f)\simeq ({\mathbb Z}/p{\mathbb Z})^{2s}$.
3. Let $E$ be the wild inertia subgroup of $Aut_k(C/H)$ at $\infty$. We denote by $D(E)$ its commutator subgroup of $E$ and by $Z(E)$ its center. Then, $E$ is an extraspecial group of order $p^{2s+1}$ and $$0 \longrightarrow D(E)=Z(E) \simeq {\mathbb Z}/p {\mathbb Z}\longrightarrow E \stackrel{\pi}{\longrightarrow} Z(Ad_f) \simeq ({\mathbb Z}/p {\mathbb Z}) ^{2s} \longrightarrow 0$$
4. Moreover, $G/H$ is a normal subgroup in $E$. It follows that $G_2$ is equal to $D(G)$, the commutator subgroup of $G$, which is also equal to $D(G)G^p$, the Frattini subgroup of $G$.
**Proof:** First of all, the existence of the group $H$ comes from [@Su1] (Chap. 2, Thm. 1.12). We deduce from Lemma 2.4.2 that $(C/H, G/H)$ is still a big action. Then, $G=G_1 \supsetneq G_2$ (resp. $G/H=(G/H)_1 \supsetneq (G/H)_2$). As the first jump always coincides in lower and upper ramification, it follows that $G_2=G^2$ (resp.$(G/H)_2=(G/H)^2$). By [@Se] (Second Part, Chap. IV, Prop. 14), $(G/H)_2=(G/H)^2= G^2H/H=G_2H/H=G_2/H$. The first assertion follows. The second and the third point directly derive from Proposition 2.5.\
We now prove the last statement. By Proposition 2.5, $Z(E)=(G/H)_2=G_2/H \subset G/H$. So, $G/H$ is a subgroup of $E$ containing $Z(E)$. Moreover, since $({\mathbb Z}/p{\mathbb Z})^{2s}$ is abelian, $\pi(G/H)$ is normal in $E/Z(E)$. It follows that $G/H$ is normal in $E$. We eventually show that $G_2=D(G)$. On the one hand, since $G/G_2$ is abelian, $D(G)$ is included in $G_2.$ On the other hand, assume that $D(G)$ is strictly included in $G_2$. Then, the first point applied to $\mathcal{G}=D(G)$ ensures the existence of a group $H$, normal in $G$, with $D(G) \subset H \subset G_2$, $[G_2:H]=p$ and such that $(C/H, G/H)$ is a big action. Since $D(G) \subset H$, $G/H$ is an abelian subgroup of $E$. As $G/H$ is also a normal group in $E$, [@Hu] (Satz 13.7) implies $|G/H| \leq p^{s+1}$. Hence $\frac{|G/H|}{g_{C/H}} \leq \frac{2\,p}{p-1}$, which contradicts condition for the big action $(C/H, G/H)$. It follows that $D(G)=G_2$. In addition, as $G/G_2$ is an elementary abelian $p$-group, then $G^p=G_1^p \subset G_2=D(G)$. As a consequence, $G_2=D(G)G^p$ which is equal to the Frattini subgroup of $G$, since $G$ is a $p$-group. $\square$
When applying Theorem 2.6 to $\mathcal{G}=G_{i_0+1}$, where $i_0$ is defined as in Proposition 2.5, one obtains Theorem 8.6(i) of [@LM]. In particular, for all big actions $(C,G)$ with $g \geq 2$, there exists an index $p$-subgroup $H$ of $G_2$, normal in $G$, such that $(C/H,G/H)$ is a big action with $C/H$ birational to $W^p-W=f(X)=X\,S(X)+c\,X \in k[X]$, where $S$ is an additive polynomial of degree $s\geq 1$ in $F$. Note that, in this case, $i_0=1+p^s$.
As $G_2$ cannot be trivial for a big action, we gather from the last point of Theorem 2.6 the following result.
Let $(C,G)$ be a big action with $g \geq 2$. Then $G$ cannot be abelian.
It is natural to wonder whether $G_2$ can be non abelian. Although we do not know yet the answer to this question, we can mention a special case in which $G_2$ is always abelian, namely:
Let $(C,G)$ be a big action with $g \geq 2$. If the order of $G_2$ divides $ p^3$, then $G_2$ is abelian.
**Proof:** There is actually only one case to study, namely: $|G_2|=p^3$. We denote by $Z(G_2)$ the center of $G_2$. The case $|Z(G_2)| = 1$ is impossible since $G_2$ is a $p$-group. If $|Z(G_2)| = p$, then $Z(G_2)$ is cyclic. But, as $G_2$ is a $p$-group, normal in $G$ and included in $D(G)$ (see Theorem 2.6), [@Su2] (Prop. 4.21, p. 75) implies that $G_2$ is also cyclic, which contradicts the strict inclusion of $Z(G_2)$ in $G_2$. If $|Z(G_2)| = p^2$, then $G_2/Z(G_2)$ is cyclic and $G_2$ is abelian, which leads to the same contradiction as above. This leaves only one possibility: $|Z(G_2)|= p^3$, which means that $G_2=Z(G_2)$. $\square$
Let $(C,G)$ be a big action with $g \geq 2$. We keep the notation of Proposition 2.2. Let $G_{\infty,1} $ be the wild inertia subgroup of $Aut_k(C)$ at $\infty$. Then, $(C,G_{\infty,1} )$ is a big action whose second lower ramification group is equal to $D(G_{\infty,1})=D(G)$. In particular, $G$ is equal to $G_{\infty,1}$ if and only if $|G/D(G)| =|G_{\infty,1} /D(G_{\infty,1})|$.
**Proof:** As $G$ is included in $G_{\infty,1}$, then $D(G) \subset D(G_{\infty,1})$. If the inclusion is strict, one can find a subgroup $\mathcal{G}$ such that $G \subsetneq \mathcal{G} \subset G_{\infty,1}$ with $[\mathcal{G}:G]=p$ (see [@Su1], Chap. 2, Thm. 19). Then, $G$ is a normal subgroup of $\mathcal{G}$. It follows that $D(G)$ is also a normal subgroup of $\mathcal{G}$. As $|G| \leq |\mathcal{G}|$, the pair $(C,\mathcal{G})$ is a big action. So, by Theorem 2.6, $\mathcal{G}_2=D(\mathcal{G})$. Since $D(G)$ is normal in $\mathcal{G}$ and $g(C/D(G))=0$, we gather from Lemma 2.4.1 that $D(G)=\mathcal{G}_2=D(\mathcal{G})$. The claim follows by reiterating the process. $\square$
Let $(C,G_{\infty,1})$ be a big action as in Corollary 2.10. Then, $G_{\infty,1}$ is a $p$-Sylow subgroup of $Aut_k(C)$. Moreover, we deduce from [@Gi] (Thm. 1.3) that $G_{\infty,1}$ is the unique $p$-Sylow subgroup of $Aut_k(C)$ except in four special cases: the hyperelliptic curves: $W^{p^n}-W=X^2$ with $p>2$, the Hermitian curves and the Deligne-Lusztig curves arising from the Suzuki groups and the Ree groups (see equations in [@Gi], Thm. 1.1).
Base change and big actions.
============================
Starting from a given big action $(C,G)$, we now display a way to produce a new one: $(\tilde{C}, \tilde{G})$, with $\tilde{G}_2\simeq G_2$ and $g_{\tilde{C}}=p^{s} \, g_C$. The main tool is a base change associated with an additive polynomial map: ${\mathbb P}_k^1 \stackrel{S}{\longrightarrow} C/G_2\simeq {\mathbb P}_k^1 $.
Let $(C,G)$ be a big action with $g \geq 2$. We denote by $L:=k(C)$ the function field of the curve $C$, by $k(X):=L^{G_2}$ the subfield of $L$ fixed by $G_2$ and by $k(T):=L^{G_1}$, with $T=\prod_{v \in V} (X-v)$. Write $X=S(Z)$, where $S(Z)$ is a separable additive polynomial of $k[Z]$ with degree $p^{s}$, $s \in {\mathbb N}$.
1. Then, $L$ and $k(Z)$ are linearly disjoined over $k(X)$.
2. Let $\tilde{C}$ be the smooth projective curve over $k$ with function field $k(\tilde{C}):=L[Z]$. Then, $k(\tilde{C})/k(T)$ is a Galois extension with group $\tilde{G} \simeq G \times ({\mathbb Z}/p{\mathbb Z})^{s}$. Furthermore, $g_{\tilde{C}}=p^{s} \, g_C$. It follows that $\frac{|\tilde{G}|}{g_{\tilde{C}}}= \frac{|G|}{g}$. So, $(\tilde{C},\tilde{G})$ is still a big action with second ramification group $\tilde{G_2}\simeq G_2\times \{0\} \subset G \times ({\mathbb Z}/p{\mathbb Z})^{s}$. This can be illustrated by the following diagram:
$$\begin{array}{clc}
C & \longleftarrow & \tilde{C}\\
\downarrow & & \downarrow \\
C/G_2\simeq {\mathbb P}_k^1& \stackrel{S}{\longleftarrow}& {\mathbb P}_k^1\\
\end{array}$$
The proof of this proposition requires two preliminary lemmas.
Let $K:=k((z))$ be a formal power series field over $k$. Let $K_1/K$ be a Galois extension whose group $\mathcal{G}$ is a $p$-group. Let $K_0/K$ be a $p$-cyclic extension. Assume that $K_0$ and $K_1$ are linearly disjoined over $K$. Put $L:=K_0K_1$. $$\begin{array}{clc}
K_1 & - & L=K_0 K_1\\
\quad \\
\mathcal{G} \, \vert& & \vert \quad \\
\quad \\
K & -& K_0
\end{array}$$ We suppose that the conductor of $K_0/K$ (see e.g. [@Se] Chap. 15, Cor. 2) is $2$. Then, $L/K_1$ also has conductor $2$.
**Proof:** Consider a chief series of $\mathcal{G}$ (cf. [@Su1], Chap. 2, Thm. 1.12), i.e. a series such that: $$\mathcal{G} =\mathcal{G}_0 \supsetneq \mathcal{G}_1 \cdots \supsetneq \mathcal{G}_n=\{0\}$$ with $\mathcal{G}_i$ normal in $\mathcal{G}$ and $[\mathcal{G}_{i-1}: \mathcal{G}_i]=p$. One shows, by induction on $i$, that the conductor of the extension $K_0K_1^{ \mathcal{G}_i}/ K_1^{ \mathcal{G}_i}$ is $2$. Therefore, one can assume $\mathcal{G} \simeq {\mathbb Z}/p{\mathbb Z}$. In this case, $L/k((z))$ is a Galois extension with group $G \simeq ({\mathbb Z}/p{\mathbb Z})^2$. Write the ramification filtration of $G$ in lower notation: $$G=G_0=\cdots=G_{i_0} \supsetneq G_{i_0+1} =\cdots$$
1. First, assume that $G_{i_0+1}=\{0\}$. Then, an exercise shows that, for any index $p$-subgroup $H$ of $G$, the extensions $L/L^H$ (case $(\alpha)$) and $L^H/K$ (case $(\beta)$) are $p$-cyclic with conductor $i_0+1$. When applied to $H=Gal(L/K_0)$, case $(\beta)$ gives $i_0=1$. Therefore, one concludes by applying case $(\alpha)$ to $H=Gal(L/K_1)$.
2. Now, assume that $G_{i_0+1}\neq \{0\}$. As above, let $H$ be an index $p$-subgroup of $G$. An exercise using the classical properties of ramification theory shows that:
1. If $H=G_{i_0+1}$, then $L/L^{H}$ (resp. $L^{H}/K$) is a $p$-cyclic extension with conductor $i_0+i_1+1$ (resp. $i_0+1$).
2. If $H\neq G_{i_0+1}$, then $L/L^{H}$ (resp. $L^{H}/K$) is a $p$-cyclic extension with conductor $i_0+1$ (resp. $i_0+\frac{i_1}{p}+1$).
Apply this result to $H:=Gal(L/K_0)$. Since $K_0/K$ has conductor $2$, it follows that $i_0+1=2$, so $i_0=1$ and $Gal(L/K_0)=G_{i_0+1}$. Therefore, $Gal(L/K_1) \neq G_{i_0+1}$ and we infer from case (b) that $L/K_1$ has conductor $i_0+1=2$. $\square$
Let $W$ be a finite ${\mathbb F}_p$-subvector space of $k$. Let $W_1$ and $W_2$ be two ${\mathbb F}_p$-subvectors spaces of $W$ such that $W =W_1 \bigoplus W_2$. Define $T:=\prod_{w \in W} (Z-w)$ and $T_i:=\prod_{w \in W_i} (Z-w)$, for $i$ in $\{1,2\}$. Then, $k(T) \subset k(T_i) \subset k(Z)$. Moreover,
1. The extensions $k(T_1)/k(T)$ and $k(T_2)/k(T)$ are linearly disjoined over $k(T)$.
2. For all $i$ in $\{1,2\}$, $k(Z)/k(T)$ (resp. $k(Z)/k(T_i)$) is a Galois extension with group isomorphic to $W$ (resp. $W_i$).
3. For all $i$ in $\{1,2\}$, $k(T_i)/k(T)$ is a Galois extension with group isomorphic to $\frac{W}{W_i}$.
This induces the following diagram: $$\begin{array}{lll}
k(T_1) & \stackrel{W_1}{-} & k(Z)\\
& &\quad \\
\vert \, \frac{W}{W_1} & & \vert \, W_2 \\
& &\quad \\
k(T) & \stackrel{\frac{W}{W_2}}{-} & k(T_2)
\end{array}$$
**Proof:** Use for example [@Go] (1.8). $\square$
**Proof of Proposition 3.1:**
1. The first point derives from Lemma 2.4.1.
2. Put $W:=S^{-1}(V)$, with $V$ defined as in Proposition 2.2.3, $W_1:=S^{-1}(\{0\})\simeq ({\mathbb Z}/p{\mathbb Z})^{s}$, since $S$ is an additive separable polynomial of $k[Z]$ with degree $p^{s}$ (see e.g. [@Go] chap. 1). Call $W_2$ any ${\mathbb F}_p$-subvector space of $W$ such that $W=W_1 \bigoplus W_2$. Then, Lemma 3.3 applied to the extension $k(Z)/k(T)$ induces the following diagram: $$\begin{array}{clc}
L=k(C) & - & k(\tilde{C})\\
\quad \\
G_2 \,\vert & & \vert \\
\quad \\
L^{G_2}=k(X)=k(Z)^{W_1} & \stackrel{W_1}{-} & k(Z)\\
\quad \\
\frac{W}{W_1} \, \vert & & \vert \, W_2 \\
\quad \\
L^{G_1}=k(T)=k(Z)^{W} & \stackrel{\frac{W}{W_2}}{-} & k(Z)^{W_2}
\end{array}$$ In particular, Lemma 3.3 implies that $k(Z)^{W_1} \cap k(Z)^{W_2}=k(T)$. Since $k(C)\cap k(Z)=k(X)$ (cf. first point of the proposition), we deduce that $k(C)$ and $k(Z)^{W_2}$ are linearly disjoined over $k(T)$. As $k(Z)^{W_2}/k(T)$ is a Galois extension with group $\frac{W}{W_2} \simeq W_1 \simeq ({\mathbb Z}/p {\mathbb Z})^{s}$, it follows that $k(\tilde{C})/k(T)$ is a Galois extension with group $\tilde{G} \simeq Gal(k(C)/k(T)) \times Gal(k(Z)^{W_2} /k(T)) \simeq G \times ({\mathbb Z}/p{\mathbb Z})^{s}$.\
Now, consider a flag of ${\mathbb F}_p$-subvector spaces of $W_1$: $$W_1=W_1^{(1)} \supsetneq W_1^{(2)} \supsetneq \cdots \supsetneq W_1^{(s+1)}=\{0\}$$ such that $[W_1^{(i-1)}:W_1^{(i)}]=p$. It induces the following inclusions: $$k(Z)=k(Z)^{W_1^{(s+1)}} \supsetneq k(Z)^{W_1^{(s)}} \supsetneq \cdots \supsetneq k(Z)^{W_1^{(1)}} =k(X)$$ Then, apply Lemma 3.2 to $K_1/K$: the completion at $\infty$ of the extension $k(C)/k(X)$, whose group $G_2$ is a $p$-group, and to $K_0/K$: the completion at $\infty$ of the $p$-cyclic extension $k(Z)^{W_1^{(i)}}/k(Z)^{W_1^{(i-1)}}$ whose conductor is $2$. By induction, we thus prove that the extension $k(\tilde{C})/k(T)$ also has conductor $2$. It follows from the Hurwitz genus formula that $g_{\tilde{C}}=p^{s} \, g_C$. Finally, the last statement on $\tilde{G}_2$ is a consequence of Lemma 2.4.1. $\square$
Under the conditions of Proposition 3.1, it can happen that $G$ is a $p$-Sylow subgroup of $Aut_k(C)$ without $\tilde{G}$ being a $p$-Sylow subgroup of $Aut_k(\tilde{C})$.\
Indeed, take $C$ : $W^p-W=X^{1+p}$ and $S(Z)=Z^p-Z$. Then, $\tilde{C}$ is parametrized by $\tilde{W}^p-\tilde{W}=(Z^p-Z)\, (Z^{p^2}-Z^p)=-Z^2+2\, Z^{1+p} -Z^{1+p^2}$ mod $\wp(k[Z])$. We denote by $G_{\infty,1}(C)$ (resp. $G_{\infty,1}(\tilde{C})$) the wild inertia subgroup of $Aut_k(C)$ (resp. $Aut_k(\tilde{C})$) at $X=\infty$ (resp. $Z=\infty$). Note that $G_{\infty,1}(C)$ (resp. $G_{\infty,1}(\tilde{C})$) is a $p$-Sylow subgroup of $Aut_k(C)$ (resp. $Aut_k(\tilde{C})$). Take $G:=G_{\infty,1}(C)$. From Proposition 2.5, we deduce that $|\tilde{G}|=p\,|G|=p\,|G_{\infty,1}(C)|=p^4$, whereas $|G_{\infty,1}(\tilde{C})|=p^5$.
A new step towards a classification of big actions.
===================================================
If big actions are defined through the value taken by the quotient $\frac{|G|}{g}$, it occurs that the key criterion to classify them is the value of another quotient: $\frac{|G|}{g^2}$. Indeed, the quotient $\frac{|G|}{g^2}$ has, to some extent, a “sieve” effect among big actions. In what follows, we pursue the work of Lehr and Matignon who describe big actions for the two highest possible values of this quotient, namely $\frac{|G|}{g^2}=\frac{4\, p}{(p-1)^2}$ and $\frac{|G|}{g^2}=\frac{4}{(p-1)^2}$ (cf. [@LM] Thm. 8.6). More precisely, we investigate the big actions $(C,G)$ that satisfy: $$\label{eq*}
M:= \frac{4}{(p^2-1)^2} \leq \frac{|G|}{g^2}$$ The choice of the lower bound $M$ can be explained as follows: as shown in the proof of ([@LM], Thm. 8.6), a lower bound $M$ on the quotient $\frac{|G|}{g^2}$ involves an upper bound on the order of the second ramification group, namely: $$\label{eq**}
|G_2| \leq \frac{4}{M} \frac{|G_2/ G_{i_0+1}| ^2}{( |G_2/ G_{i_0+1}| -1)^2}$$ where $i_0$ is defined as in Proposition 2.5. Therefore, we have to choose $M$ small enough to obtain a wide range of possibilities for the quotient, but meanwhile large enough to get serious restrictions on the order of $G_2$. The optimal bound seems to be $M:=\frac{4}{(p^2-1)^2}$, insofar as, for such a choice of M, the upper bound on $G_2$ implies that its order divides $p^3$, and then that $G_2$ is abelian (cf. Corollary 2.9).
Let $(C,G)$ be a big action with $g \geq 2$ satisfying condition . Then, the order of $G_2$ divides $p^3$. It follows that $G_2$ is abelian.
**Proof:** Put $p^m:=|G_2/G_{i_0+1}|$, with $m \geq 1$, and $A_m := \frac{4}{M} \frac{|G_2/ G_{i_0+1}|}{(|G_2/ G_{i_0+1}| -1)^2}=\frac{4}{M}
\frac{p^{m}}{(p^{m}-1)^2}$. Then, inequality reads: $1< |G_2| =p^{m} |G_{i_0+1}|\leq p^{m} A_m$, which gives: $1 \leq |G_{i_0+1}| \leq A_m$. Since $(A_m)_{m \geq 1}$ is a decreasing sequence with $A_4 <1$, we conclude that $m \in \{1,2,3 \}$.\
If $m=3$, then $1 \leq |G_{i_0+1}| \leq A_3 <p$. So $|G_{i_0+1}|=1$ and $|G_2|=p^3$. If $m=2$, then $1 \leq |G_{i_0+1}| \leq A_{2}=p^2$. So $|G_2| = p^2 \, |G_{i_0+1}|$, with $ |G_{i_0+1}| \in \{1,p,p^2 \}$. This leaves only one case to exclude, namely $|G_{i_0+1}|=p^2$. In this case, $|G_2|=p^4$ and formula yields a lower bound on the genus, namely: $ 2\, g \geq \, (i_0-1) (p^4-1).$ Let $s$ be the integer defined in Remark 2.7. Then, $i_0=1+p^s$. Besides, by Theorem 2.6, $V \subset ({\mathbb Z}/p{\mathbb Z})^{2s}$. Consequently, $|G|=|G_2||V| \leq p^{4+2s}$ and $$\frac{|G|}{g^2} \leq \frac{4\, p^{4+2s}}{p^{2s}(p^4-1)^2}=\frac{4}{(p^2-1)^2} \frac{p^4}{(p^2+1)^2} <\frac{4}{(p^2-1)^2}$$ which contradicts equality .\
If $m=1$, then $1 \leq |G_{i_0+1}| \leq A_{1}$ with $ A_{1}:=p \, (p+1)^2 <
\left\{
\begin{array}{ll}
p^4 \,, \quad if \, p \geq 3 \\
p^5 \,, \quad if \, p=2
\end{array}
\right.
$.\
Since $G_{i_0+1}$ is a $p$-group, we get: $
\left\{
\begin{array}{ll}
1 \leq |G_{i_0+1}| \leq p^3 \,, \quad if \, p \geq 3 \\
1 \leq |G_{i_0+1}| \leq p^4 \,, \quad if \, p=2
\end{array}
\right.
$. As $|G_2|=p\,|G_{i_0+1}|$, there are two cases to exclude: $|G_{i_0+1}|=p^{3+\epsilon}$, with $\epsilon =0$ if $p \geq 3$ and $\epsilon \in \{0,1\}$ if $p=2$. Then $|G_2|=p^{4+\epsilon}$. If $\epsilon =0$, we are in the same situation as in the previous case. If $\epsilon=1$, yields $2\, g \geq (i_0-1) (p^{5}-1).$ Since this case only occurs for $p=2$, we eventually get an inequality: $$\frac{|G|}{g^2} \leq \frac{4\, p^{5+2s}}{p^{2s}\,(p^5-1)^2}= \frac{128}{961} <\frac{4}{9}= \frac{4}{(p^2-1)^2}$$ which contradicts condition . Therefore, the order of $G_2$ divides $p^3$. Then, we gather from Corollary 2.9 that $G_2$ is abelian. $\square$
But we can even prove better: under these conditions, $G_2$ has exponent $p$.
Let $(C,G)$ be a big action with $g \geq 2$ satisfying condition .Then $G_2$ is abelian with exponent $p$.
**Proof:** By Proposition 4.1, $G_2$ is abelian, with order dividing $p^3$. As a consequence, if $G_2$ has exponent strictly greater than $p$, either $G_2$ is cyclic with order $p^2$ or $p^3$, or $G_2$ is isomorphic to ${\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p {\mathbb Z}.$ We begin with a lemma excluding the second case. Note that one can find big actions $(C,G)$ with $G_2$ abelian of exponent $p^2$. Nevertheless, it requires the $p$-rank of $G_2$ to be large enough (see section 6).
Let $(C,G)$ be a big action with $g \geq 2$ satisfying condition . Then $G_2$ cannot be isomorphic to ${\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p {\mathbb Z}$.
**Proof:** Assume $G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p {\mathbb Z}$. Then, the lower ramification filtration of $G$ reads as in one of the four following cases:
i\) $G=G_1 \supsetneq G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+1} \simeq {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+i_1+1} = \{0 \}.$
ii\) $G=G_1 \supsetneq G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+1}\simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset G_{i_0+i_1+1} = \{0 \}.$
iii\) $G=G_1 \supsetneq G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+1} \simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset G_{i_0+i_1+1} \simeq {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+i_1+i_2} = \{0 \}.$
iv\) $G=G_1 \supsetneq G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p {\mathbb Z}\supset G_{i_0+1} \simeq {\mathbb Z}/p^2 {\mathbb Z}\supset G_{i_0+i_1+1} \simeq {\mathbb Z}/p{\mathbb Z}\supset G_{i_0+i_1+i_2} = \{0 \}.$
We now focus on the ramification filtration of $G_2$, temporary denoted by $H$ for convenience. Then, for all $i \geq 0$, the lower ramification groups of $H$ are: $H_i=H \cap G_i$.\
In case i), the lower ramification of $H$ reads: $$H=H_0=\cdots=H_{i_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+1}=\cdots=H_{i_0+i_1} \simeq {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+i_1+1} = \{0 \}.$$ Consider the upper ramification groups: $H^{\nu_0}=H^{\varphi(i_0)}=H_{i_0}$ and $H^{\nu_1}=H^{\varphi(i_0+i_1)}=H_{i_0+i_1}$, where $\varphi$ denotes the Herbrand function. Then, the ramification filtration in upper notation reads: $$H^0=\cdots=H^{\nu_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_0+1}=\cdots
=H^{\nu_1} \simeq {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_1+1} =\{0 \}.$$ Since $H$ is abelian, it follows from Hasse-Arf theorem that $\nu_0$ and $\nu_1$ are integers. Consequently, the formula: $$\forall \, m \in {\mathbb N}, \quad
\varphi(m)+1= \frac{1}{|H_0|} \, \sum_{i=0}^m |H_i|$$ gives $\nu_0=i_0$ and $\nu_1 = i_0+\frac{i_1}{p^2}$. Besides, [@Ma] (Thm. 6) implies $H^{\nu_0} \supsetneq H^{p\,\nu_0} \supset (H^{\nu_0})^p$ with $ (H^{\nu_0})^p=H^p=G_2^p \simeq {\mathbb Z}/p {\mathbb Z}$. Thus, $ H^{p\nu_0} \supset H^{\nu_1}$, which involves: $ p \nu_0 \leq \nu_1$ and $i_1 \geq p^2(p-1)i_0$. Then, the Hurwitz genus formula applied to $C \rightarrow C/H \simeq {\mathbb P}^1_k$ yields a lower bound for the genus: $$2\, g =(i_0-1) (|H|-1)+i_1(|H_{i_0+1}|-1) \geq (p-1) (i_0+1) (p^3+p+1).$$ Let $s$ be the integer defined in Remark 2.7. Then, $i_0=1+p^s$. Moreover, by Theorem 2.6, $|G|=|G_2||V| \leq p^{3+2s}$. It follows that $ \frac{|G|}{g^2} \leq \frac{4}{(p^2-1)^2} \frac{p^3(p+1)^2}{(p^3+p+1)^2} $. Since $\frac{p^3(p+1)^2}{(p^3+p+1)^2}<1$ for $p \geq 2$, this contradicts condition .
In case ii), the lower ramification filtration of $H$ reads: $$H=H_0=\cdots=H_{i_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+1}=\cdots H_{i_0+i_1} \simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset H_{i_0+i_1+1} =\{0 \}.$$ Keeping the same notation as in case i), the upper ramification filtration reads: $$H=H^0=\cdots=H^{\nu_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_0+1}=\cdots
=H^{\nu_1}\simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset H^{\nu_1+1} =\{0 \}.$$ with $\nu_0=\varphi(i_0)=i_0$ and $\nu_1=\varphi(i_0+i_1)=i_0+\frac{i_1}{p}$. Once again, $H^{p\nu_0} \supset (H^{\nu_0})^p \simeq {\mathbb Z}/p {\mathbb Z}$ implies $ H^{p \,\nu_0} \supset H^{\nu_1}$, which involves $p \, \nu_0 \leq \nu_1$ and $i_1 \geq i_0 \, p \, (p-1).$ Then, the Hurwitz genus formula yields: $$2\, g=(i_0-1) (|H|-1)+i_1(|H_{i_0+1}|-1) \geq (p-1) \, p^s \, (p^3+p^2+1) \geq (p-1)p^s (p^3+p+1).$$ Thus, we get the same lower bound on the genus as in the preceding case, hence the same contradiction.
In case iii), the lower ramification filtration of $H$ reads: $$H_{i_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+1}=\cdots=H_{i_0+i_1} \simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset H_{i_0+i_1+1} =\cdots=H_{i_0+i_1+i_2} \simeq {\mathbb Z}/ p {\mathbb Z}\supset \{ 0 \}.$$ Keeping the same notation as above and introducing $H^{\nu_2}=H^{\varphi(i_0+i_1+i_2)}=H_{i_0+i_1+i_2}$, the upper ramification filtration reads: $$H^{\nu_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_0+1}=\cdots
=H^{\nu_1} \simeq ({\mathbb Z}/p{\mathbb Z})^2 \supset H^{\nu_1+1} =\cdots=H^{\nu_2} \simeq {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_2+1} = \{0\}$$ with $\nu_0=\varphi(i_0)=i_0$, $\nu_1=\varphi(i_0+i_1)=i_0+\frac{i_1}{p}$ and $\nu_2= \varphi(i_0+i_1+i_2)= i_0+ \frac{i_1}{p} + \frac{i_2}{p^2}$. Since $H^{p\nu_0} \supset (H^{\nu_0})^p \simeq {\mathbb Z}/ p {\mathbb Z}$, we obtain: $H^{p\,\nu_0} \supset H^{\nu_2}$. Then, $p \, \nu_0 \leq \nu_2$, which involves $p^2\, (p-1) \, i_0 \leq i_1 \, p +i_2$. With such inequalities, the Hurwitz genus formula gives a new lower bound for the genus, namely: $$2 \,g=(i_0-1) (|H|-1)+i_1(|H_{i_0+1}|-1)+i_2
(|H_{i_0+i_1+1}|-1) \geq (p-1) \, (p^s \, (p^2+p+1) +(p^s+1) \, (p-1) \, p^2)$$ From $ 2\, g \geq (p-1) \, (p^{3+s}+p^{1+s}+p^s+p^3-p^2 ) \geq (p-1) \, p^s(p^3+p)$, we infer the inequality: $$\frac{|G|}{g^2} \leq
\frac{4}{(p^2-1)^2} \, \frac{p^{2s+3}(p+1)^2}{p^{2s}\, (p^3+p )^2} =
\frac{4}{(p^2-1)^2} \, \frac{p\, (p+1)^2}{(p^2+1 )^2}$$ Since $\frac{p\, (p+1)^2}{(p^2+1 )^2}<1$ for $p \geq 2$, this contradicts condition .
In case iv), the lower ramification filtration of $H$ : $$H_{i_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+1}=
\cdots=H_{i_0+i_1} \simeq ({\mathbb Z}/p^2{\mathbb Z}) \supset H_{i_0+i_1+1} =\cdots=H_{i_0+i_1+i_2} \simeq {\mathbb Z}/ p {\mathbb Z}\supset \{ 0 \}.$$ induces the following upper ramification filtration: $$H^{\nu_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\times {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_0+1}=\cdots
=H^{\nu_1} \simeq ({\mathbb Z}/p^2{\mathbb Z}) \supset H^{\nu_1+1} =\cdots=H^{\nu_2} \simeq {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_2+1} = \{0\}.$$ This is almost the same situation as in case iii), except that $H_{i_0+1}$ is isomorphic to ${\mathbb Z}/p^2 {\mathbb Z}$ instead of $({\mathbb Z}/p {\mathbb Z})^2$. But, since the only thing that plays a part in the proof is the order of $H_{i_0+1}$ , which is the same in both cases, namely $p^2$, we conclude with the same arguments as in case iii). $\square$
The previous method based on the analysis of the ramification filtration of $G_2$ fails to exclude the case $G_2 \simeq {\mathbb Z}/ p^2 {\mathbb Z}$ for a big action satisfying . Indeed, if $H:=G_2\simeq {\mathbb Z}/ p^2 {\mathbb Z}$, the lower ramification filtration of $H$: $$H_0=\cdots=H_{i_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\supset H_{i_0+1}=\cdots H_{i_0+i_1} \simeq {\mathbb Z}/p{\mathbb Z}\supset H_{i_0+i_1+1} = \{0 \}.$$ induces the upper ramification filtration: $$H^0=\cdots=H^{\nu_0} \simeq {\mathbb Z}/p^2 {\mathbb Z}\supset H^{\nu_0+1}=\cdots
=H^{\nu_1} \simeq {\mathbb Z}/p{\mathbb Z}\supset H^{\nu_1+1} =\{0 \}.$$ with $\nu_0=\varphi(i_0)=i_0$ and $\nu_1=\varphi(i_0+i_1)=i_0+\frac{i_1}{p}$. Since $H^{p\nu_0} \supset (H^{\nu_0})^p \simeq {\mathbb Z}/ p {\mathbb Z}$, we obtain: $p \, \nu_0 \leq \nu_1,$ hence $i_1 \geq (p-1) \, p \, i_0$. Let $s$ be the integer defined in Remark 2.7. Then, the Hurwitz genus formula yields: $$2\, g=(i_0-1) (|H|-1)+i_1(|H_{i_0+1}|-1) \geq (p-1) \, (p^s \, (p^2+1) +p^2-p) \geq (p-1) \, p^s \, (p^2+1).$$ If we denote by $v$ the dimension of the ${\mathbb F}_p$-vector space $V$, we eventually get: $$\frac{|G|}{g^2} \leq
\frac{4}{(p^2-1)^2} \, \frac{p^{2+v}(p+1)^2}{p^{2s} \, (p^2+1)^2} .$$ In this case, condition requires $p^{1+\frac{v}{2}-s} (p+1) > p^2$. Since $\frac{v}{2} \leq s$, this implies $p+1 > p^{1+s-\frac{v}{2}} \geq p$, hence $\frac{v}{2} = s$. This means that $V= Z(Ad_f)$, where $f$ is the function defined in Remark 2.7 and $Ad_f$ its palindromic polynomial as defined in Proposition 2.5. Therefore, one does not obtain yet any contradiction.
Accordingly, to exclude the cyclic cases $G_2
\simeq {\mathbb Z}/ p^2 {\mathbb Z}$ and $G_2 \simeq {\mathbb Z}/ p^3 {\mathbb Z}$ and thus complete the proof of Proposition 4.2, we need to shift from a ramification point of view on $G_2$ to the embedding problem: $G_2
\subsetneq G_1$. This enables us to prove the more general result on big actions formulated in the next part.
Big actions with a cyclic second ramification group $G_2$.
==========================================================
The aim of this section is to prove that there does not exist any big action whose second ramification group $G_2$ is cyclic, except for the trivial case $G_2\simeq {\mathbb Z}/p{\mathbb Z}$.
Let $(C,G)$ be a big action. If $G_2 \simeq ({\mathbb Z}/ p^n {\mathbb Z})$, then $n=1$.
**Proof:**\
Let $(C,G)$ be a big action with $ G_2 \simeq {\mathbb Z}/ p^n {\mathbb Z}$.
1. *First of all, we prove that we can assume $n=2$.*\
Indeed, for $n >2$, $\mathcal{H}:=G_2^{p^{n-2}}$ is a normal subgroup in $G$, strictly included in $G_2$. So Lemma 2.4.2 asserts that the pair $(C/\mathcal{H},G/\mathcal{H})$ is a big action. Besides, the second lower ramification group of $G/\mathcal{H}$ is isomorphic to ${\mathbb Z}/p^2{\mathbb Z}.$\
2. *Notation and preliminary remarks.*\
We denote by $L:=k(C)$ the function field of $C$ and by $k(X):=L^{G_2}$ the subfield of $L$ fixed by $G_2$. Following Artin-Schreier-Witt theory (see [@Bo] Chap. IX, ex. 19), we define the $W_2({\mathbb F}_p)$-module $$\tilde{A}:= \frac{ \wp (W_2(L)) \cap W_2(k(X))}{ \wp (W_2(k(X)))}$$ where $W_2(L)$ denotes the ring of Witt vectors of length 2 with coordinates in $L$. The inclusion $k[X] \subset k(X)$ induces an injection $$A:=\frac{ \wp (W_2(L)) \cap W_2(k[X])}{ \wp (W_2(k[X]))}\hookrightarrow \tilde {A}$$ Since $L/L^{G_2}$ is étale outside $X=\infty$, it follows from [@Mi] (III, 4.12) that we can identify $A$ with $\tilde{A}$. Consider the Artin-Schreier-Witt pairing: $$\left\{
\begin{array}{ll}
G_2 \times A \longrightarrow W_2({\mathbb F}_p)\\
(g, \overline{\wp \, x}) \longrightarrow [g, \overline{\wp \, x} > := gx-x
\end{array}
\right.$$ where $g \in G_2 \subset Aut_k(L)$, $x \in L$ such that $\wp x \in k[X]$ and $\overline{\wp x}$ denotes the class of $\wp x$ mod $\wp (k[X])$. This pairing is non degenerate, which proves that, as a group, $A$ is dual to $G_2$.\
As a ${\mathbb Z}$-module, $A$ is generated by $(f_0(X),g_0(X))$ in $W_2(k[X])$ and then, $L=k(X,W_0,V_0)$ with $\wp(W_0,V_0)= (f_0(X), g_0(X))$. An exercise left to the reader shows that one can choose $f_0(X)$ and $g_0(X)$ reduced mod $\wp (k[X])$ (see definition of a reduced polynomial in section 1). We denote by $m_0$ (resp. $n_0$) the degree of $f_0$ (resp. $g_0$). Note that they are prime to $p$. The $p$-cyclic cover $L^{G_2^p}/L^{G_2}$ is parametrized by: $W_0^p-W_0=f_0(X)$. We deduce from Proposition 2.5 that $f_0(X)=XS(X)+c\,X$, where $S$ is an additive polynomial with degree $s \geq 1$ in $F$. After an homothety on $X$, we can assume $S$ to be monic. Furthermore, note that $s \geq 2$. Indeed, if $s=1$, the two inequalities established in Remark 4.4: $|G| \leq p^{2+2s} \leq p^4$ and $2\, g \geq (p-1) \, (p^s \, (p^2+1) +p^2-p) = (p-1) \, (p^3+p^2)$ imply $\frac{|G|}{g} \leq \frac{2\,p}{p-1} \, \frac{p^3}{p^3+p^2} <\frac{2\,p}{p-1}$, which contradicts .\
3. *The embedding problem.*\
For any $y \in V$, the class of $ (f_0(X+y),g_0(X+y))$ in $A$ induces a new generating system of $A$, which means that : $$\label{e6}
{\mathbb Z}(f_0(X),g_0(X)) \,= \, {\mathbb Z}(f_0(X+y),g_0(X+y)) \, \mod \wp(W_2(k[X])).$$ As $A$ is isomorphic to ${\mathbb Z}/p^2 {\mathbb Z}$, ensures the existence of an integer $n(y)$ such that $$\label{e7}
(f_0(X+y),g_0(X+y))=n(y)\, (f_0(X),g_0(X)) \qquad \mod \wp(W_2(k[X]))$$ where $n(y):=a_0(y)+b_0(y)\, p$, with $a_0(y) \in {\mathbb N}$, $0<a_0(y) <p$, and $b_0(y) \in {\mathbb N}$, $0 \leq b_0(y) <p$. We calculate $n(y)\, (f_0(X),g_0(X))=a_0(y) \, (f_0(X),g_0(X))+ b_0(y) p \, (f_0(X),g_0(X))$. On the one hand, $a_0(y) \, (f_0(X),g_0(X))=(a_0(y) f_0(X), a_0(y)g_0(X)+ c(a_0(y)) f_0(X))$, where $c(a_0(y))$ is given by the recursive formula: $$\forall \, i \in {\mathbb N}, \quad c(i+1)=c(i)+ \frac{1}{p} \,(1+i^p-(1+i)^p) \quad \mod \, p$$ On the other hand, $$b_0(y) \,p \, (f_0(X),g_0(X))=b_0(y) \,(0,f_0(X)^p)= (0, b_0(y) f_0(X)) \, \mod \wp(W_2(k[X]))$$ As a conclusion, reads: $$\label{e9}
(f_0(X+y),g_0(X+y))= (a_0(y) f_0(X), a_0(y) g_0(X)+ \ell_0(y) f_0(X)) \, \mod \wp(W_2(k[X]))$$ where $\ell_0(y):=c(a_0(y)) +b_0(y) $. We notice that, for all $y$ in $V$, $a_0(y)=1$ mod $p$. Indeed, the equality of the first coordinate of Witt vectors in implies $f_0(X+y)=a_0(y) \, f_0(X) \mod \, \wp(k[X])$. Thus, by induction, $f_0(X+py)=a_0(y)^p \, f_0(X) \mod \, \wp(k[X])$. Since $V$ is an elementary abelian $p$-group, $f_0(X+py)=f_0(X)$, which involves: $a_0(y)^p=1$ mod $p$ and $a_0(y)=1$ mod $p$. So, becomes: $$\label{e10}
(f_0(X+y),g_0(X+y))= (f_0(X),g_0(X)+ \ell_0(y) f_0(X)) +(P^p(X),Q^p(X))-(P(X),Q(X))$$ with $P(X)$ and $Q(X)$ polynomials of $k[X]$. In order to circumvent the problem related to the special formula giving the opposite of Witt vectors for $p=2$, we would rather write as follows: $$\label{e11}
(f_0(X+y),g_0(X+y))+(P(X),Q(X))= (f_0(X),g_0(X)+ \ell_0(y) \,f_0(X)) +(P(X)^p,Q(X)^p)$$ The first coordinate of reads: $$\label{e12}
f_0(X+y)+P(X)=f_0(X)+P(X)^p$$ On the second coordinate of , the addition law in the ring of Witt vectors gives the following equality in $k[X]$: $$\label{e13}
g_0(X+y)+Q(X)+\psi(f_0(X+y),P(X))=g_0(X)+\ell_0(y)\, f_0(X)+Q(X)^p+ \psi(f_0(X),P(X)^p)$$ where $\psi$ is defined as follows: $$\psi(a,b):= \frac{1}{p}\, (a^p+b^p-(a+b)^p)
= \frac{-1}{p}\, \sum_{i=1}^{p-1}\, \binom pi \, a^i\, b^{p-i}= \sum_{i=1}^{p-1}\, \frac{(-1)^i}{i} \, a^i \, b^{p-i}\quad \mod \, p$$ As a consequence, gives: $$\label{e14}
\Delta_y (g_0):= g_0(X+y)-g_0(X)=\ell_0(y) \, f_0(X)+ \delta \qquad \mod \, \wp (k[X])$$ with $$\begin{array}{ll}
\delta&:=\psi(f_0(X),P(X)^p)-\psi(f_0(X+y),P(X))\\
&= \sum_{i=1}^{p-1} \,\frac{(-1)^i}{i}\, \{ f_0(X)^i \, P(X)^{p(p-i)}-f_0(X+y)^i \,P(X)^{p-i} \}
\end{array}$$
With the notation defined above, $\delta$ is equal to: $$\label{e15}
\delta =\sum_{i=1}^{p-1} \frac{(-1)^i}{i} \, y^{p-i} X^{i+p^{s+1}} +\mbox {lower degree terms in X }$$
**Proof:** We search for the monomials in $\delta$ that have degree in $X$ greater or equal to $p^{s+1}+1$. We first focus on $f_0(X)^i \, P(X)^{p(p-i)}$. We can infer from equality that $P(X)$ has degree $p^{s-1}$ and that its leading coefficient is $y^{1/p}$. Furthermore, [@LM] (proof of Prop. 8-1) shows that $P(X)-P(0)$ is an additive polynomial. So, we can write: $P(X)=y^{1/p} \, X ^{p^{s-1}} +P_1(X)$, where $P_1(X)$ is a polynomial of $k[X]$ with degree at most $p^{s-2}$. Then, for all $i$ in $\{1,\cdots,p-1\}$, $f_0(X)^i\, P(X)^{p\, (p-i)}= f_0(X)^i \, (y\, X^{p^s}+P_1(X)^p)^{p-i}= f_0(X)^i \, (\sum_{j=0}^{p-i} \binom {p-i}j \, y^j \, X^{jp^s} \, P_1(X)^{p(p-i-j)})$. Since $f_0(X)$ has degree: $1+p^s$, this gives in $\delta$ a monomial of degree at most: $i\, (1+p^s)+j\, p^s +p\, (p-i-j)\, p^{s-2}=p^s+(i+j)\, (p-1) \,p^{s-1}+i.$ If $j \leq p-i-1$, this degree is at most: $p^s+(p-1)^2\,p^{s-1}+i=(p-1)\,p^s+p^{s-1}+i$, which is strictly lower than $p^{s+1}+1$, for $s \geq 2$ and $1 \leq i \leq p-1$ . As a consequence, the monomials of degree greater or equal to $p^{s+1}+1$ can only occur when the index $j$ is equal to $p-i$, namely in $f_0(X)^i \, y^{p-i} \, X^{p^s(p-i)}$. As $f_0(X)=X \,S(X)+c \, X$, where $S$ is a monic additive polynomial of degree $s$ in $F$, $f_0$ reads: $f_0(X)=X^{1+p^s} +P_2(X)$ where $P_2(X)$ is a polynomial in $k[X]$ with degree at most $1+p^{s-1}$. Then, for all $i$ in $\{1,\cdots,p-1\}$, $f_0(X)^i \, y^{p-i} \, X^{p^s(p-i)}= y^{p-i} \, X^{p^s(p-i)} \, (\sum_{k=0}^i \binom ik X^{(1+p^s)j}\, P_2(X)^{i-k})$. Accordingly, we get a monomial of degree at most: $ p^s\, (p-i) +k\, (1+p^s)+(i-k) \,(1+p^{s-1})=p^s \, (p-i) +i \, (1+p^{s-1} ) +k \, (p^s-p^{s-1})$. When $0 \leq k \leq i-1$, the maximal degree obtained in this way is $i + p^{s-1}-p^s+p^{s+1}$ which is stricly lower than $p^{s+1}+1$. Therefore, for all $i$ in $\{1,\cdots,p-1\}$, the only contibution to take into account is $k=i$, which produces in $\delta$ the sum: $ \sum_{i=1}^{p-1} \frac{(-1)^i}{i}\, y^{p-i} X^{i+p^{s+1}}$.\
We now search for monomials with degree greater or equal to $p^{s+1}+1$ in the second part of $\delta$, namely: $f_0(X+y)^i \,P(X)^{p-i}$. This has degree at most: $i\, (1+p^s) +(p-i)\,p^{s-1}=i\,p^s+(p-i)\, p^{s-1}+i$, which is strictly lower than $p^{s+1}+1$, for $s\geq 2$ and $1 \leq i \leq p-1$. Therefore, $f_0(X+y)^i \,P(X)^{p-i}$ does not give any monomial in $\delta$ with degree greater or equal to $p^{s+1}+1$. Thus, we get the expected formula. $\square$ $\square$
4. *We notice that $g_0(X)$ cannot be of the form $X\, \Sigma(X)+\gamma\,X$, with $\Sigma \in k\{F\}$ and $\gamma \in k$.*\
Otherwise, the left-hand side of reads: $\Delta_y(g_0):=
g_0(X+y)-g_0(X)=X\,\Sigma(y)+y\,\Sigma(X)+y\,\Sigma(y)+\gamma \,y$, which only gives a linear contribution in $X$ after reduction mod $\wp(k[X])$. By Lemma 5.2, $deg \, f_0=1+p^s < deg\, \delta= p^{s+1}+p-1$, which involves that the degree of the right-hand side of is $p-1+p^{s+1} >1$, hence a contradiction.\
Therefore, we can define an integer $a \leq n_0=deg \, g_0$ such that $X^{a}$ is the monomial of $g_0(X)$ with highest degree which is not of the form $1+p^n$, with $n\in {\mathbb N}$. Note that since $g_0$ is reduced mod $\wp(k[X])$, $a \not \equiv 0$ mod $p$. We also notice that the monomials in $g_0(X)$ with degree strictly greater than $a$ are of the form $X^{1+p^n}$, and so, as explained above, they only give linear monomials in $\Delta_y(g_0)$ mod $\wp(k[X])$. Therefore, after reduction mod $\wp(k[X])$, the degree of the left-hand side of is at most $a-1$. Since the degree of the right-hand side is $p^{s+1}+p-1$, it follows that: $$\label{a}
a-1 \geq p^{s+1}+p-1 \qquad$$
5. *We show that $p$ divides $a-1$.*\
Assume that $p$ does not divide $a-1$. In this case, the monomial $X^{a-1}$ is reduced mod $\wp(k[X])$. Since the monomials of $g_0(X)$ with degree strictly greater than $a$ only give a linear contribution in $\Delta_y(g_0)$ mod $\wp \, (k[X])$, reads as follows, for all $y$ in $V$: $$c_a(g_0) \, a\, y X^{a-1}+ \mbox{lower degree terms } = -y \, X^{p^{s+1}+p-1} +
\mbox{ lower degree terms} \mod \, \wp\, (k[X])$$ where $c_a(g_0)\neq 0$ denotes the coefficient of $X^a$ in $g_0$. If $a-1 > p^{s+1}+p-1$, the coefficient $c_a(g_0)\, a\, y=0$, for all $y$ in V. Since $a
\neq 0$ mod $p$, it leads to $V=\{0\}$, so $G_1=G_2$, which is impossible for a big action (see Proposition 2.2.1). We gather from that $a-1= p^{s+1}+p-1$, which contradicts: $a \neq 0$ mod $p$.\
Thus, $p$ divides $a-1$. So, we can write $a=1+\lambda\, p^t$, with $t>0$, $\lambda$ prime to $p$ and $\lambda \geq 2$ because of the definition of $a$. We also define $j_0:=a-p^t=1+(\lambda-1)\, p^t$. Note that $pj_0 > a$. Indeed, $$pj_0 \leq a \Leftrightarrow p(1+(\lambda-1)p^t) \leq 1+ \lambda\,p^t \Leftrightarrow \lambda \leq \frac{1-p+p^{t+1}}{p^t(p-1)}=\frac{-1}{p^t}+ \frac{p}{p-1} < \frac{p}{p-1} \leq 2$$ which is impossible since $\lambda \geq 2$.
6. *We determine the coefficient of $X^{j_0}$ in the left hand-side of .*\
Since $p$ does not divide $j_0$, the monomial $X^{j_0}$ is reduced mod $\wp(k[X])$. In the left-hand side of , namely $\Delta_y(g_0)$ mod $\wp(k[X])$, the monomial $X^{j_0}$ comes from monomials of $g_0(X)$ of the form: $X^b$, with $b$ in $\{j_0+1,\cdots,a\}$. As a matter of fact, the monomials of $g_0(X)$ with degree strictly greater than $a$ only give a linear contribution mod $\wp(k[X])$, whereas $j_0=1+(\lambda-1)\, p^t>1$. For all $b \in \{j_0+1,\cdots, a\}$, the monomial $X^b$ of $g_0(X)$ generates $\binom b{j_0}\, y^{b-j_0} \, X^{j_0}$ in $\Delta_y(g_0)$. Since $p\, j_0 > a \geq b$ (see above), these monomials $X^b$ do not produce any $X^{j_0\, p^n}$, with $n \geq 1$, which would also give $X^{j_0}$ after reduction mod $\wp(k[X])$. It follows that the coefficient of $X^{j_0}$ in the left-hand side of is $T(y)$ with $T(Y):= \sum _{b=j_0+1}^a \, c_b(g_0)\, \binom b{j_0}\, Y^{b-j_0}$, where $c_b(g_0)$ denotes the coefficient of $X^b$ in $g_0(X)$. As the coefficient of $Y^{a-j_0}$ in $T(Y)$ is $c_a(g_0) \, \binom a{j_0}=c_a(g_0)\, \binom {1+\lambda p^t}{1+(\lambda-1)p^t} \equiv c_a(g_0) \, \lambda \not \equiv 0 \, \mod \, p$, the polynomial $T(Y)$ has degree $a-j_0=p^t$.\
7. *We identify with the coefficient of $X^{j_0}$ in the right-hand side of and gather a contradiction.*\
We first assume that the monomial $X^{j_0}$ does not occur in the right-hand side of . Then, $T(y)=0$ for all $y$ in $V$, which means that $V$ is included in the set of roots of $T$. Thus, $|V| \leq p^t$. To compute the genus $g$, put $M_0:=m_0$ and $M_1:=\max \{ p \,m_0, \,n _0 \}$. Then, by [@Ga], the Hurwitz genus formula applied to $C \rightarrow C/G_2 \simeq {\mathbb P}_k^1$ yields: $2\, (g-1)=2\, |G_2|\, (g_{C/G_2}-1) +d=-2\,p^2+d$, with $d:=(p-1)\,(M_0+1)+p\,(p-1) \,(M_1+1)$. From $p \, m_0= p\, (p^s+1)=p^{s+1}+p$ and $p^{s+1}+p-1 < n_0$, we infer $M_1=n_0$. Moreover, since $n_0 \geq a =1+\lambda \, p^t \geq 1+2\,p^t >2\,p^t$, we obtain a lower bound for the genus: $2\, g=(p-1)\,p\, (n_0-1+p^{s-1}) \geq 2\,p^{t+1} \, (p-1)$. As $|G|=|G_2||V|\leq p^{2+t}$, it entails an inequality $\frac{|G|}{g} \leq \frac{2\:p}{p-1} \, \frac{p^{1+t}}{2\,p^{1+t}}= \frac{1}{2} \, \frac{2\,p}{p-1}$ which contradicts .\
As a consequence, the monomial $X^{j_0}$ appears in the right-hand side of , which implies that $j_0 \leq p^{s+1}+p-1$. Using , we get: $ j_0=1+(\lambda-1)\,p^t \leq p^{s+1}+p-1 < a=1+\lambda \,p^t $. This yields: $$\label{t2}
\lambda-1 \leq p^{s+1-t} + \frac{p-2}{p^t} < \lambda$$ If $ s+1-t \leq -1$, since $t \geq 1$, gives: $\lambda-1 \leq \frac{1}{p} +\frac{p-2}{p} <1$, which contradicts $\lambda \geq 2$. It follows that $s+1-t \geq 0$. Then, combined with: $0 \leq \frac{p-2}{p^t} <1$ leads to $\lambda-1=p^{s+1-t}$. We gather that $j_0=1+(\lambda-1)\, p^t =1+p^{s+1} > deg \, f_0=1+p^s$. Therefore, in the right-hand side of , the monomial $X^{j_0}=X^{1+p^{s+1}}$ only occurs in $\delta$. By Lemma 5.2, the coefficient of $X^{j_0}=X^{1+p^{s+1}}$ in $\delta$ is $-y^{p-1}$. By equating the coefficient of $X^{j_0}$ in each side of , we get $T(y)=-y^{p-1}$, for all $y$ in $V$. Put $\tilde{T}(Y):=
T(Y)+Y^{p-1}$. Since $deg \, T =p^t > p-1$, the polynomial $\tilde{T}$ has still degree $p^t$ and satisfies: $\tilde{T}(y)=0$ for all $y$ in $V$. Once again, it leads to $|V| \leq p^t$, which contradicts as above. $\square$
Therefore, when $(C,G)$ is a big action, $G_2 \simeq ({\mathbb Z}/p^n {\mathbb Z})$ implies $n=1$. More generally, if $G_2$ is abelian of exponent $p^n$, with $n \geq 2$, there exists an index $p$-subgroup of $G_2^p$, say $H$, normal in $G$ such that the pair $(C/H, G/H)$ is a big action with $(G/H)_2=G_2/H \simeq {\mathbb Z}/p^2 {\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t$, with $t \in {\mathbb N}^*$. A natural question is to search for a lower bound on the $p$-rank: $t$ depending on the genus $g$ of the curve. As seen in the proof of Theorem 5.1, the difficulty lies in the embedding problem, i.e. in finding an extension which is stable under the translations by $V$. In the next section, we exhibit big actions with $G_2$ abelian of exponent at least $p^2$. In particular, we construct big actions $(C,G)$ with $G_2 \simeq {\mathbb Z}/p^2{\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t$ where $t= O(log_p \, g)$.
Examples of big actions with $G_2$ abelian of exponent strictly greater than $p$.
=================================================================================
In characteristic $0$, an anologue of big actions is given by the actions of a finite group $G$ on a compact Riemann surface $C$ with genus $g_C \geq 2$ such that $|G| = 84(g_C-1)$. Such a curve $C$ is called a *Hurwitz curve* and such a group $G$ a *Hurwitz group* (cf. [@Co]). In particular, the lowest genus Hurwitz curves are the Klein’s quartic with $G \simeq PSL_2({\mathbb F}_7)$ (cf. [@El2]) and the Fricke-Macbeath curve with genus $7$ and $G \simeq PSL_2({\mathbb F}_8)$ (cf. [@Mc65]).\
Let $C$ be a Hurwitz curve with genus $g_c$. Let $n \geq 2$ be an integer and let $C_n$ be the maximal unramified Galois cover whose group is abelian, with exponent $n$. The Galois group of the cover $C_n/ C$ is isomorphic to $({\mathbb Z}/ n {\mathbb Z})^{2g_C}.$ We infer from the unicity of $C_n$ that the ${\mathbb C}$-automorphims of $C$ have $n^{2g_c}$ prolongations to $C_n$. Therefore, $g_{C_n}-1
= n^{2g}(g_{C}-1)$. Consequently, $C_n$ is still a Hurwitz curve (see [@Mc]).\
Now, let $(C,G)$ be a big action. Then $C \rightarrow C/G$ is an étale cover of the affine line whose group is a $p$-group. From the Deuring-Shafarevich formula (see e.g. [@Bou]), it follows that the Hasse-Witt invariant of $C$ is zero. This means that there are no nontrivial connected étale Galois covers of $C$ with group a $p$-group. Therefore, if we want to generalize the method mentionned above to produce Galois covers of $C$ corresponding to big actions, it is necessary to introduce ramification. A means to do so is to consider ray class fields of function fields, as studied by K. Lauter [@Lau] and R. Auer [@Au]. Since the cover $C \rightarrow C/G_2$ is an étale cover of the affine line $Spec \, k[X]$ totally ramified at $\infty$, we focus on the special case of ray class fields of the rational function field ${\mathbb F}_q(X)$, where $q=p^e$ (see [@Au], III.8). Such ray class fields allows us to produce families of big actions $(C,G)$ (where $C$ is defined over $k={\mathbb F}_p^{alg}$) with specific conditions imposed on ramification and endowed with an abelian $G_2$ of exponent as large as we want.
([@Au], Part II) Let $K$ be the rational function field: ${\mathbb F}_q(X)$, with $q=p^e$ and $e \in {\mathbb N}^*$. Let $S$ be the set of all finite rational places, namely $\{ (X-y),\, y \in {\mathbb F}_q \}$. Let $m \geq 0$ be an integer. Fix $K^{alg}$ an algebraic closure of $K$ in which all extensions of $K$ are assumed to lie. We define $K_S^m \subset K^{alg}$ as the largest abelian extension $L / K$ with conductor $\leq m \,\infty$, such that every place in $S$ splits completely in $L$.
1. We define the splitting set of any finite Galois extension $L/K$, denoted by $S(L)$, as the set consisting of the places of $K$ that split completely in $L$. If $K_S^m/K$ is the extension defined in Definition 6.1, then $S\subset S(K_S^m)$.
2. In what follows, we only consider finite Galois extensions $L/K$ that are unramified outside $X=\infty$ and (totally) ramified at $X=\infty$. Therefore, the support of the conductor of $L/K$ is reduced to the place $\infty$. So, we systematically confuse the conductor $m\,\infty$ with its degree $m$.
3. We could more generally define $K_S^m$ for $S$ a non-empty subset of the finite rational places, i.e. $S:= \{ (X-y), \, y \in V \subset {\mathbb F}_q\}$. However, to get big actions, it is necessary to consider the case where $V$ is a subgroup of ${\mathbb F}_q$. In what follows, we focus on the case $V={\mathbb F}_q$, as announced in Definition 6.1.
We keep the notation of Definition 6.1.
1. The existence of the extension $K_S^m / K$ is based on global class field theory (see [@Au], Part II).
2. $K_S^m/K$ is a finite abelian extension whose full constant field is ${\mathbb F}_q$.
3. The reason for Lauter and Auer’s interest in such ray class fields is that they provide for examples of global function fields with many rational places, or what amounts to the same, of algebraic curves with many rational points. Indeed, let $C(m) / {\mathbb F}_q$ be the nonsingular projective curve with function field $K_S^m$. If we denote by $N_m:= |C(m)({\mathbb F}_q)|$ the number of ${\mathbb F}_q$-rational points on the curve $C(m)$, then: $ N_m=1+q \,[K_S^m:K]$. The main difficulty lies in computing $[K_S^m:K]$. We first wonder when $K_S^m$ coincide with $K$. Here are partial answers.
4. Let $q=p^e$, with $ e\in {\mathbb N}$. If $e$ is even, put $r:=\sqrt{q}$ and if $e$ is odd, put $r:=\sqrt{qp}$. Then, for all $i$ in $\{0,\cdots, r+1\}$, $K_S^i=K={\mathbb F}_q(X)$. (see [@Au], III, Lemma 8.7 and formula (13)). Note that the previous estimate $ N_m=1+q \,[K_S^m:K]$, combined with the Hasse-Weil bound (see e.g. [@St93] V.2.3), furnishes another proof of $K_S^i=K$ when $i <1+r$.
5. More generally, Lauter displays a method to compute the degree of the extension $K_S^m/K$ via a formula giving the order of its Galois group: $G_S(m)$ (see [@Lau], Thm. 1). Her proof consists in starting from the following presentation of $G_S(m)$: $$G_S(m) \simeq \frac{1+Z\, {\mathbb F}_q [[Z]]} {< 1+Z^m\, {\mathbb F}_q[[Z]], 1-yZ,\, y \in {\mathbb F}_q >}$$ where $Z=X^{-1}$, which indicates that $G_S(m)$ is an abelian finite $p$-group. Then, she transforms the multiplicative structure of the group into an additive group of generalized Witt vectors. In particular, she deduces from this theorem the smallest conductor $m$ such that $G_S(m)$ has exponent stricly greater than $p$ (see next proposition).
([@Lau], Prop. 4) We keep the notation defined above. If $q=p^e$, the smallest conductor $m$ for which the group $G_S(m)$ is not of exponent $p$ is $m_2:= p^{\lceil e/2\rceil +1} +p+1$, where $\lceil e/2 \rceil$ denotes the upper integer part of $e/2$.
We now emphasize the link with big actions. Let $F$ be a function field with full constant field ${\mathbb F}_q$. Let $C/{\mathbb F}_q$ be the smooth projective curve whose function field is $F$ and $C^{alg}:= C\times_{{\mathbb F}_q} k$ with $k={\mathbb F}_p^{alg}$. If $G$ is a finite $p$-subgroup of $Aut_{{\mathbb F}_q} C$, then $G$ can be identified with a subgroup of $Aut_{k} C^{alg}$. In this case, $(C^{alg}, G)$ is a big action if and only if $g_{C^{alg}}=g_C >0$ and $\frac{|G|}{g_C} >\frac{2\,p}{p-1}$. For convenience, in the sequel, we shall say that $(C,G)$ is a big action if $(C^{alg}, G)$ is a big action.\
In what follows, we consider the curve $C(m)/{\mathbb F}_q$ whose function field is $K_S^m$ and, starting from this, we construct a $p$-group $G(m)$ acting on $C(m)$ by extending the translations $X \rightarrow X+y$, with $y \in {\mathbb F}_q$. In particular, we obtain an upper bound for the genus of $C(m)$, which allows us to circumvent the problem related to the computation of the degree $[K_S^m:K]$ when checking whether $(C(m), G(m))$ is a big action.
We keep the notation defined above.
1. Let $C(m) / {\mathbb F}_q$ be the nonsingular projective curve with function field $K_S^m$. Then, the group of translations: $X \rightarrow X+y$, $y\in {\mathbb F}_q$, extends to a $p$-group of ${\mathbb F}_q$-automorphisms of $C(m)$, say $G(m)$, with the following exact sequence: $$0 \longrightarrow G_S(m) \longrightarrow G(m) \longrightarrow {\mathbb F}_q \longrightarrow 0$$
2. Let $L$ be an intermediate field of $K_S^m/K$. Assume $L=(K_S^m)^H$, i.e. the extension $L/K$ is Galois with group: $G_S(m)/H$. For all $i \geq 0$, we define $L^i$ as the $i$-th upper ramification field of $L$, i.e. the subfield of $L$ fixed by the $i$-th upper ramification group of $G_S(m)/H$ at $\infty$: $G_S^i(m)H/H$, where $G_S^i(m)$ denotes the $i$-th upper ramification group of $G_S(m)$ at $\infty$. Then, $$\forall \, i \geq 0, \quad L^i= L \cap K_S^{i}$$ In particular, when $L=K_S^m$ and $ i \leq m$, $L^i=K_S^i$, i.e. $G_S^i(m)=Gal(K_S^m/K_S^i)$.
3. Let $L$ be an intermediate field of $K_S^m/K$. Define $n:=\min \{i \in {\mathbb N}, L \subset K_S^i\}$. Then, the genus of the extension $L/K$ is given by the formula: $$g_{L} = 1+ [L:K] \,(-1 + \frac{n}{2}) -\frac{1}{2}\, \sum_{j=0}^{n-1} \,[L \cap K_S^j:K]$$ where the sum is empty for $n=0$.\
In particular, $g_L=0$ if and only if $n:=\min \{i \in {\mathbb N}, L \subset K_S^i\}=0$.\
Note that if $n >0$, then $g_L < [L:K] \, (-1+\frac{n}{2})$.
4. If $m \geq r+2$, $\frac{|G(m)|}{g_{K_S^{m}}} >\frac{q}{-1+\frac{m}{2}}$. It follows that if $\frac{q}{-1+\frac{m}{2}} \geq \frac{2\,p}{p-1}$, the pair $(C(m), G(m))$ is a big action. In this case, the second lower ramification group of $G(m)$: $G_2(m)$, is equal to $G_S(m)$. In particular, for $p >2$, (resp. $p=2$), if $e\geq 4$ (resp. $e \geq 6$) and if $m_2$ is the integer defined in Proposition 6.4, the pair $(C(m_2), G(m_2))$ is a big action whose second ramification group: $G_S(m_2)$, is abelian of exponent $p^2$.
**Proof:**
1. The set $S$ is globally invariant under the translations: $X \rightarrow X+y$, $y\in {\mathbb F}_q$. That is the same for $\infty$, so the translations by ${\mathbb F}_q$ do not change the conditions imposed on ramification. As a consequence, owing to the maximality and the unicity of $K_S^m$, they can be extended to ${\mathbb F}_q$-automorphisms of $K_S^m$. This proves the first assertion.
2. The second point directly derives from [@Au] (II, Thm. 5.8).
3. The genus formula is obtained by combining the preceding results, the Hurwitz genus formula and the Discriminant formula (see [@Au], I, 3.7). Now assume that $n=0$. Then, $L\subset K_S^0={\mathbb F}_q(X)$ and $g_L=0$. Conversely, assume $g_L=0$. If $n \neq 0$, Remark 6.3.4 implies that $n \geq r+2 \geq 3$. Using the preceding formula and Remark 6.3.4, $g_L=0$ reads: $$2+ (n-2)\, [L:K] = \sum_{j=0}^{n-1} \, [K_S^j \cap L:K]= 2+ \sum_{j=2}^{n-1} \, [K_S^j \cap L:K] \leq 2+ (n-2)\, [L:K]$$ It follows that, for all $j$ in $\{2,\cdots,n-1\}$, $K_S^j \cap L=L$. In particular, $L \subset K_S^2=K_S^0$, hence a contradiction. Finally, since $n>0$ implies $n \geq 3$ and since $K=K_S^0=K_S^1$, one notices that $$g_L=[L:K] \,(-1 + \frac{n}{2}) -\frac{1}{2}\, \sum_{j=2}^{n-1} \,[L \cap K_S^j:K] <[L:K]\, (-1+\frac{n}{2})$$
4. Assume that $m \geq r+2$. We gather from Remark 6.3.4 that $n:=\min \{i \in {\mathbb N}, K_S^m \subset K_S^i\} \geq r+2 \geq 3$. Then, it follows from the previous point that $$g_{K_S^m} <[K_S^m:K]\, (-1+\frac{n}{2}) \leq [K_S^m:K] \,(-1 + \frac{m}{2})$$ As $|G(m)|=q [K_S^m:K]$, we deduce the expected inequality. In particular, when $\frac{q}{-1+\frac{m}{2}} > \frac{2\,p}{p-1}$, the pair $(C(m), G(m))$ is a big action. It remains to show that, in this case, $G_2(m)$ is equal to $G_S(m)$. Lemma 2.4.2 first proves that $G_S(m) \supset G_2(m)$. Let $L:=(K_S^m)^{G_2(m)}$ be the subfield of $L$ fixed by $G_2(m)$ and define $n:=\min \{i \in {\mathbb N}, L \subset K_S^i\}$. Assume $G_S(m) \supsetneq G_2(m)$. Then $L \supsetneq (K_S^m)^{G_S(m)}=K$. We infer from Remark 6.3.4 that $n \geq r+2$, which proves, using the previous point, that $g_L>0$. But, since $(C(m), G(m))$ is a big action, $C/G_2(m) \simeq {\mathbb P}_k^1$, so $g_L=0$, hence a contradiction. We eventually explain the last statement. By Proposition 6.5.2, $G_S^{m_2-1}(m_2)=Gal(K_S^{m_2}/K_S^{m_2-1})$, which induces the following exact sequence: $$0 \longrightarrow G_S^{m_2-1}(m_2) \longrightarrow G_S(m_2) \longrightarrow G_S(m_2-1) \longrightarrow 0$$ We infer from Proposition 6.4 that $G_S(m_2-1)$ has exponent $p$ whereas the exponent of $G_S(m_2)$ is at least $p^2$. It follows that $G_S^{m_2-1}(m_2)$ cannot be trivial. Since $G_S^{m_2}(m_2)=\{0\}$ (use Proposition 6.5.2), we deduce from the elementary properties of the ramification groups that $G_S^{m_2-1}(m_2)$ is $p$-elementary abelian. Therefore, $G_S(m_2)$ has exponent smaller than $p^2$ and the claim follows. $\square$
Let $N_m$ be the number of ${\mathbb F}_q$-rational points on the curve $C(m)$ as defined in Remark 6.3.3. Then, $N_m=1+q \, |G_S(m)|=1+ |G(m)|$. This highlights the equivalence of the two ratios: $\frac{|G(m)|}{g_{C(m)}}$ and $\frac{N_m}{g_{C(m)}}$. In particular, this equivalence emphasizes the link between the problem of big actions and the search of algebraic curves with many rational points.
As seen in Remark 6.3.4, $K_S^i=K$ for all $i$ in $\{0,\cdots,r+1\}$, where $r=\sqrt{q}$ or $\sqrt{qp}$ according to whether $q$ is a square or not. The following extensions $K_S^m$, for $m \geq r+2$, are partially parametrized, at least for the first ones, in [@Au] (Prop. 8.9). In the table below, we exhibit a complete description of the extensions $K_S^m$ for $m$ varying from $0$ to $m_2=p^{\lceil e/2\rceil +1}+p+1$, in the special case $p=5$ and $e=4$. This involves $q=p^e=625$, $s=e/2=2$, $r=p^s=25$ and $m_2=131$. The table below should suggest the general method to parametrize such extensions.
conductor $m$ $[K_S^m:K]$ New equations
--------------------------------- ------------- --------------------------------------------------
$0 \leq m \leq r+1=26$ $1$
$r+2=27 \leq m \leq 2r+1=51$ $5^2$ $W_{0}^r+W_{0}=X^{1+r}$
$m=2r+2=52$ $5^6$ $W_{1}^q-W_{1}=X^{2r}\, (X^q-X)$
$2r+3=53 \leq m \leq 3r+1=76$ $5^8$ $W_{2}^r+W_{2}=X^{2(1+r)}$
$m=3r+2=77$ $5^{12}$ $W_{3}^q-W_{3}=X^{3r}\, (X^q-X)$
$m=3r+3=78$ $5^{16}$ $W_{4}^q-W_{4}=X^{3r}\, (X^{2q}-X^2)$
$3r+4=79 \leq m \leq 4r+1=101$ $5^{18}$ $W_{5}^r+W_{5}=X^{3(1+r)}$
$m=4r+2=102$ $5^{22}$ $W_{6}^q-W_{6}=X^{4r}\, (X^q-X)$
$m=4r+3=103$ $5^{26}$ $W_{7}^q-W_{7}=X^{4r}\, (X^{2q}-X^2)$
$m=4r+4=104$ $5^{30}$ $W_{8}^q-W_{8}=X^{4r}\, (X^{3q}-X^3)$
$4r+5=105 \leq m \leq 5r+1=126$ $5^{32}$ $W_{9}^r+W_{9}=X^{4(1+r)}$
$m=5r+2=127$ $5^{36}$ $W_{10}^q-W_{10}=X^{5r}\, (X^{q}-X)$
$m=5r+3=128$ $5^{40}$ $W_{11}^q-W_{11}=X^{5r}\, (X^{2q}-X^2)$
$m=5r+4=129$ $5^{44}$ $W_{12}^q-W_{12}=X^{5r}\, (X^{3q}-X^3)$
$m=5r+5=130$ $5^{48}$ $W_{13}^q-W_{13}=X^{5r}\, (X^{4q}-X^4)$
$m=m_2=131$ $5^{50}$ $[W_{0}, W_{14}]^r+[W_{0}, W_{14}]=[X^{1+r},0] $
In this case, $$\label{quotient}
\frac{|G(m_2)|}{g_{K_S^{m_2}}} \simeq 9,6929 \cdots$$
**Comments on the construction of the table:** For all $i$ in $\{0,\cdots,14\}$, put $L_i:=K(W_0, \cdots, W_i)$.
1. We first prove that the splitting set of each extension $K(W_i)/K$ (see Remark 6.2.1) contains $S$. Indeed, fix $y$ in ${\mathbb F}_q$ and call $P_y$ the corresponding place in $S$: $(X-y)$. We have to distinguish three cases. By [@St93] (Prop. VI. 4.1), $P_y$ completely splits in the extension $K(W)/K$, where $W^r+W=X^{u\,(1+r)}$, with $1 \leq u \leq 4$, if the polynomial $T^r+T-y^{u\,(1+r)}$ has a root in $K$, which is true since $y^{u(1+r)}=(F^s+I) \, (\frac{1}{2}\, y^{u(1+r)})$. Likewise, $P_y$ completely splits in the extension $K(W)/K$, where $W^q-W=X^{u\,r} \, (X^{v\,q}-X^v)$, with $1 \leq v < u \leq 5$, since $y^{vq}-y^v=0$. Finally, $P_y$ completely splits in the extension $K(W,\tilde{W})/K$, where $[W,\tilde{W}]^r+[W,\tilde{W}]=[X^{1+r},0]$, since $[y^{1+r},0]= (F^s+I)\, [\frac{1}{2} \,y^{1+r},-\frac{2^p-2}{4p}\, y^{(1+r)\,p}]$. To conclude, we remark that $L_i=L_{i-1} \, K(W_i)$ for all $i$ in $\{1,\cdots,14\}$. Then, $S(L_i)=S(L_{i-1}) \cap S(K(W_i))$ (cf. [@Au], Cor. 3.2.b), which allows us to gather, by induction on $i$, that the splitting set of each $L_i$ contains $S$.
2. We now compute the conductor $m(K(W_i))$ of each extension $K(W_i)/K$. As above, we have to distinguish three kinds of extensions. First, the extension $K(W)/K$, where $W^r+W=X^{u\,(1+r)}$, with $1 \leq u \leq 4$, has conductor $ur+u+1$ (see [@Au], Prop. 8.9.a). Besides, the extension $K(W)/K$, where $W^q-W=X^{u\,r} \, (X^{v\,q}-X^v)$, with $1 \leq v < u \leq 5$, has conductor $ur+v+1$ (see [@Au], Prop. 8.9.b). Finally, the conductor of the extension $K(W,\tilde{W})/K$, where $[W,\tilde{W}]^r+[W,\tilde{W}]=[X^{1+r},0]$ is given by the formula: $1+ \max \{p(1+r),0\} = 1+p+p^{s+1}=m_2$ (see [@Ga], Thm. 1.1). As a conclusion, since $m(L_i)= \max \{m(L_{i-1}), m(K(W_i)) \}$ (cf. [@Au], Cor. 3.2.b), an induction on $i$ allows us to obtain the expected conductor for $L_i$.
3. We gather from the two first points the inclusions: $K(W_0) \subset K_S^{27}$, $K(W_0,W_1) \subset K_S^{52}, \cdots$\
$ K(W_0,\cdots, W_{14}) \subset K_S^{m_2}$. Equality is eventually obtained by calculating the degree of each extension $K_S^m/K$ via [@Lau] (Thm. 1) or [@Au] (p. 54-55, formula (13)). $\square$
We deduce from what preceeds an example of big actions with $G_2$ abelian of exponent $p^2$, with a small $p$-rank. More precisely, we construct a subextension of $K_S^{m_2}$ with the commutative diagram: $$\begin{matrix}
0&\longrightarrow &G_S(m_2)&\longrightarrow& G(m_2)&\longrightarrow&{\mathbb F}_q&\longrightarrow &0\\
& & \varphi \, \downarrow& &\downarrow& &\vert \vert& & \\
0&\longrightarrow &H&\longrightarrow& G&\longrightarrow&{\mathbb F}_q&\longrightarrow &0\\
& & \downarrow& &\downarrow& & & & \\
& & 0& &0& & & &
\end{matrix}$$ such that the pair $(C(m_2)/Ker(\varphi), G)$ is a big action where $G_2 \simeq {\mathbb Z}/p^2 {\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t$ with $t=O(log_p \,g)$, $g$ being the genus of the curve $C(m_2)/Ker(\varphi)$. Contrary to the previous case where the stability under the translations by ${\mathbb F}_q$ was ensured by the maximality of $K_S^{m_2}$, the difficulty now lies in producing a system of equations defining a subextension of $K_S^{m_2}$ which remains globally invariant through the action of the group of translations $X \rightarrow X+y$, $y \in {\mathbb F}_q$. Write $q=p^e$. We have to distinguish the case $e$ even and $e$ odd.
We keep the notation defined above. In particular, $K={\mathbb F}_q(X)$ with $q=p^e$. Assume $e=2\,s$, with $s \geq 1$, and put $r:=p^s$. We define $$f_0(X):= a\, X^{1+r} \; \; \mbox{with} \; \; a \neq 0, \, \; a \in \Gamma:=\{ \gamma \in {\mathbb F}_q, \gamma^r+\gamma=0\}$$ and $$\forall \, i \in \{1,\cdots,p-1\}, \, \; f_i(X)=X^{ir/p} \, (X^q-X)=X^{ip^{s-1}}\, (X^q-X)$$ Let $L:=K(W_i)_{1 \leq i \leq p}$ be the extension of $K$ parametrized by the Artin-Schreier-(Witt) equations: $$W_0^p-W_0=f_0(X) \quad \forall \, i\, \in \{1,\cdots,p-1\}, \; W_i^q-W_i=f_i(X) \quad \mbox{and} \quad [W_0,W_p]^p-[W_0,W_p]= [f_0(X),0]$$ For all $i$ in $\{0,1,\cdots, p-1\}$, put $L_i:=K(W_0, \cdots,W_i)$.
1. $L$ is an abelian extension of $K$ such that every place in $S$ completely splits in $L$. Moreover, $$L_0 \subset K_S^{r+2} \quad, \forall \, i \in \{1,\cdots,p-1\}, \, \, L_i \subset K_S^{p^{s+1}+i+1} \,\, \mbox{with} \, L \subset K_S^{m_2}$$ where $m_2=p^{s+1}+p+1$ is the integer defined in Proposition 6.4. (see table below).
2. The extension $L/K$ has degree $[L:K]=p^{2+(p-1)e}$. Let $G_L$ be its Galois group. Then $$G_L \simeq {\mathbb Z}/p^2{\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t \quad \mbox{with} \, \, t=(p-1)\,e$$
3. The extension $L/K$ is stable under the translations: $X\rightarrow X+y$, with $y \in {\mathbb F}_q$. Therefore, the translations by ${\mathbb F}_q$ extend to form a $p$-group of ${\mathbb F}_q$-automorphisms of $L$, say $G$, with the following exact sequence: $$0\longrightarrow G_L\longrightarrow G \longrightarrow{\mathbb F}_q\longrightarrow 0$$
4. Let $g_L$ be the genus of the extension $L/K$. Then, $$g_L= \frac{1}{2} \, \{ \, p^{2+2\,s\,(p-1)} \, (p^{s+1}+p-1)- p^{s} \, (p^2-p+1)- p^{2\,s+1}\, (\sum_{i=0}^{p-2}\, q^i)\, \}$$ In particular, when $e$ grows large, $g_L\sim \frac{1}{2}\, p^{(2p-1)\, \frac{e}{2}+3}$ and $t=O(log_p \,g_L)$.\
Note that, for $p=5$ and $e=4$, one gets $\frac{|G|}{g_L} \simeq 9,7049 \cdots$, which is slightly bigger than the quotient obtained for the whole extension $K_S^{m_2}$ (see ).
**Proof:**
1. Fix $y$ in ${\mathbb F}_q$ and call $P_y:=(X-y)$, the corresponding place in $S$. As $f_i(y)=0$ for all $i$ in $\{1,\cdots, p-1\}$, the place $P_y$ completely splits in each extension $K(W_i)$ with $W_i^q-W_i=f_i(X)$. Therefore, to prove that $P_y$ completely splits in $L$, it is sufficient to show that $[f_0(y), 0] \in \wp(W_2({\mathbb F}_q))$. By [@Bo] (Chap. IX, ex. 18), this is equivalent to show that $Tr([f_0(y),0])=0$, where $Tr$ means the trace map from $W_2({\mathbb F}_q)$ to $W_2({\mathbb F}_p)$. We first notice that, when $y$ is in ${\mathbb F}_q$, $\gamma:=f_0(y)=a\,y^{1+r}$ lies in $\Gamma.$ It follows that: $$Tr([\gamma,0])=\sum_{i=0}^{2s-1} \, F^i \, [\gamma,0]= \sum_{i=0}^{s-1} \, [\gamma^{p^i},0]+\sum_{i=0}^{s-1} \,[\gamma^{r\,p^i},0]=\sum_{i=0}^{s-1} \, [\gamma^{p^i},0]+\sum_{i=0}^{s-1} \,[-\gamma^{p^i},0]$$ For $p>2$, $[-\gamma^{p^i},0]=-[\gamma^{p^i},0]$ and $Tr([\gamma,0])=0$. For $p=2$, since $p\,[\gamma,0]=[0,\gamma^p]$, one gets: $$Tr([\gamma,0])=[0,\gamma^p +\gamma^{p^2} +\cdots+\gamma^{p^s}]=[0,(\gamma +\gamma^{p} +\cdots+\gamma^{p^{s-1}})^p]= [0,Tr_{{\mathbb F}_r/{\mathbb F}_p} (\gamma)^p]$$ As $\Gamma$ coincides with $\{\beta^{r}-\beta, \, \beta \in {\mathbb F}_q\}$ (see e.g. [@Au] p. 58), $Tr_{{\mathbb F}_r/{\mathbb F}_p} (\gamma)=0$ and $Tr([\gamma,0])=0$. To establish the expected inclusions, it remains to compute the conductor of each extension $L_i$. First of all, [@Au] (I, ex. 3.3) together with [@St93] (Prop III,7.10) shows that the conductor of $L_0$ is $r+2$. Thus, $L_0 \subset K_S^{r+2}$. Moreover, as $f_i(X)=X^{i+p^{s+1}}-X^{1+ip^{s-1}}$ mod $\wp({\mathbb F}_q[X])$, we infer from [@Au] (I, ex. 3.3) and [@Au] (I, Cor. 3.2) that the conductor of $L_i$ is $1+i+p^{s+1}$. So, $L_i \subset K_S^{1+i+p^{s+1}}$. To complete the proof, it remains to show that $L$ has conductor $m_2$, which derives from [@Ga] (see comments above).
The equations, conductor and degree of each extension $L_i$ are finally gathered in the table below.
$L_i$ conductor $m$ $[L_i:K]$ New equations
----------------- ----------------------------------- ----------------- -------------------------------------
$K$ $0 \leq m \leq r+1=p^s+1$ $1$
$L_0$ $r+2 \leq m \leq p^{s+1}+1=m_2-p$ $p$ $W_{0}^p-W_{0}=f_0(X)$
$L_1$ $m=p^{s+1}+2=m_2-(p-1)$ $p^{1+e}$ $W_1^q-W_1=f_1(X)$
$L_2$ $m=p^{s+1}+3=m_2-(p-2)$ $p^{1+2e}$ $W_2^q-W_2=f_2(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_i$ $m=p^{s+1}+i+1=m_2-(p-i)$ $p^{1+ie}$ $W_i^q-W_i=f_i(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_{p-1}$ $m=p^{s+1}+p=m_2-1$ $p^{1+(p-1)e}$ $W_{p-1}^q-W_{p-1}=f_{p-1}(X)$
$L$ $m=p^{s+1}+p+1=m_2$ $p^{2+(p-1)e}$ $[W_0,W_p]^p-[W_0,W_p]=[f_0(X),0] $
2. See table above.
3. Fix $y$ in ${\mathbb F}_q$. Consider $\sigma$ in $G(m_2)$ (defined as in Proposition 6.5) such that $\sigma(X)=X+y$.
1. We first prove that $\sigma(W_0) \in L_0$. Indeed, as $y \in {\mathbb F}_q$ and $a \in \Gamma=\{ \gamma \in {\mathbb F}_q, \gamma^r+\gamma=0\}$, $$\begin{array}{ll}
\wp(\sigma(W_0)-W_0)&= \sigma(\wp(W_0))-\wp(W_0)\\
& =f_0(X+y)-f_0(X)\\
&= a\,y\,X^r+a\,y^r\,X+f_0(y)\\
&=-a^r \,y^{r^2} \, X^r+a\,y^r\,X+f_0(y)\\
&=\wp(P_y(X))+f_0(y)
\end{array}$$ where $P_y(X):=(I+F+F^2+\cdots +F^{s-1}) \, (-a\, y^r \, X)$. Since $f_0(y) \in \wp({\mathbb F}_q)$ (see proof of the first point), it follows that $\wp(P_y(X))+f_0(y)$ belongs to $\wp({\mathbb F}_q[X])$. Therefore, $\sigma(W_0) \in L_0={\mathbb F}_q(X,W_0)$.
2. We now prove that, for all $i$ in $\{1,\cdots,p-1\}$, $\sigma(W_i) \in L_i$. Indeed, $$\begin{array}{ll}
(F^e-{id})\, (\sigma(W_i)-W_i)&= \sigma (W_i^q-W_i)-(W_i^q-W_i)\\
\quad \\
&=f_i(X+y)-f_i(X)\\
\quad \\
&=(X+y)^{i \,p^{s-1}} \, (X^q-X)-X^{i\, p^{s-1}}\, (X^q-X)\\
\quad \\
&=(X^{p^{s-1}}+y^{p^{s-1}})^{i} \, (X^q-X)-X^{i\, p^{s-1}}\, (X^q-X)\\
\quad \\
&=\sum_{j=1}^{i-1} \, \binom ij \, y^{(i-j)p^{s-i}}\, f_j(X) \, \mod (F^e-{id}) \, ({\mathbb F}_q[X])\\
\quad \\
&=(F^e-{id}) \, (\sum_{j=1}^{i-1} \, \binom ij \, y^{(i-j)p^{s-i}}\, W_j) \, \mod (F^e-{id}) \, ({\mathbb F}_q[X])\\
\end{array}$$ where the sum is empty for $i=1$. It follows that $\sigma(W_i) \in L_i={\mathbb F}_q(X,W_0, W_1, \cdots, W_i)$.
3. To conclude, we show that $\sigma(W_p) \in L$, which requires the use of Remark 6.3.4. Indeed, compute: $$\begin{array}{ll}
\Delta&:=\wp(\sigma \, [W_0,W_p]-[W_0,W_p])\\
&= \sigma (\wp([W_0,W_p])-\wp([W_0,W_p])\\
&=[f_0(X+y),0]-[f_0(X),0]
\end{array}$$ As shown in the proof of the first point, $[f_0(y),0]$ lies in $\wp(W_2({\mathbb F}_q))$. Then, $$\Delta=[f_0(X+y),0]-[f_0(X),0]-[f_0(y),0]-[P_y(X),0]+[P_y(X),0]^p \mod \,\wp(W_2({\mathbb F}_q[X]))$$ with $y$ in ${\mathbb F}_q$ and $P_y$ defined as above. Let $W({\mathbb F}_q)$ be the ring of Witt vectors with coefficients in ${\mathbb F}_q$. Then, for any $y \in {\mathbb F}_q$, we denote by $\tilde{y}$ the Witt vector $\tilde{y}:=(y,0,0,\cdots) \in W(k)$. For any $P(X):=\sum_{i=0}^{s} \, a_i \, X^i \in {\mathbb F}_q[X]$, we denote by $\tilde{P}(X):=\sum_{i=0}^s \, \tilde{a_i} \, X^i \in W({\mathbb F}_q)[X]$. The addition in the ring of Witt vectors yields: $$\begin{array}{ll}
\Delta&= [0,A] \quad \mod \,\wp(W_2({\mathbb F}_q[X]))
\end{array}$$ where $A$ is the reduction modulo $p\, W_2({\mathbb F}_q)[X]$ of: $$\frac{1}{p} \{\tilde{f_0}(X+\tilde{y})^p-\tilde{f_0}(X)^p-\tilde{f_0}(\tilde{y})^p+\tilde{P}_y(X)^p-\tilde{P}_y(X)^{p^2} -(\tilde{f_0}(X+\tilde{y})-\tilde{f_0}(X)-\tilde{f_0}(\tilde{y})-\tilde{P}_y(X)+\tilde{P}_y(X)^p)^p\}$$ Since $\tilde{f_0}(X+\tilde{y})-\tilde{f_0}(X)-\tilde{f_0}(\tilde{y})+\tilde{P}_y(X)-\tilde{P}_y(X)^p=0$ mod $p\, W({\mathbb F}_q)[X]$, $A$ becomes: $$A=\frac{1}{p} \, \{\tilde{f_0}(X+\tilde{y})^p-\tilde{f_0}(X)^p-\tilde{f_0}(\tilde{y})^p+\tilde{P}_y(X)^p-\tilde{P}_y(X)^{p^2}\} \quad \mod \, p\,W({\mathbb F}_q)[X]$$ We observe that: $$\begin{array}{lll}
\tilde{f_0}(X+\tilde{y})^p&=\tilde{a}^p \, (X+\tilde{y})^p \, (X+\tilde{y})^{p^{s+1}}& \mod \, p^2\,W({\mathbb F}_q)[X]\\
\quad \\
&= \tilde{a}^p \, (X+\tilde{y})^p \, (X^{p^s}+\tilde{y}^{p^s})^{p} \,& \mod \, p^2\,W({\mathbb F}_q)[X]\\
\quad \\
&=\tilde{a}^p \, \sum_{i=0}^p \,\sum_{j=0}^p \, \binom pi \, \binom pj \, X^{j+ip^s} \, \tilde{y}^{p-j+p^s\,(p-i)} \, &\mod \, p^2\,W({\mathbb F}_q)[X]\\
\end{array}$$ As $\binom pi \, \binom pj=0$ mod $p^2$ when $0<i<p$ and $0<j<p$, one obtains: $$\tilde{f_0}(X+\tilde{y})^p-\tilde{f_0}(X)^p-\tilde{f_0}(\tilde{y})^p=\tilde{a}^p \, \sum_{(i,j) \in I}
\, \binom pi \, \binom pj \, X^{j+ip^s} \, \tilde{y}^{p-j+p^s\,(p-i)} \quad \mod \, p^2\,W({\mathbb F}_q)[X]\\$$ with $$I:=\{(i,j) \in {\mathbb N}^2, \, 0 \leq i \leq p, 0\leq j \leq p,\, ij=0 \mod p, (i,j)\neq(0,0) \,, (i,j)\neq(p,p)\}$$ Besides, $$\begin{array}{lll}
\tilde{P}_y(X)^p-\tilde{P}_y(X)^{p^2}&= (\sum_{i=0}^{s-1} \, (-\tilde{a} \, \tilde{y}^r \, X)^{p^i}) ^{p} -
(\sum_{i=0}^{s-1} \, (-\tilde{a} \, \tilde{y}^r \, X)^{p^i}) ^{p^2} &\, \mod \, p^2\,W({\mathbb F}_q)[X] \\
\quad \\
&= (\sum_{i=0}^{s-1} \, (-\tilde{a} \, \tilde{y}^r \, X)^{p^i})^{p} -
(\sum_{i=0}^{s-1} \, (-\tilde{a} \, \tilde{y}^r \, X)^{p^{i+1}})^{p} &\, \mod \, p^2\,W({\mathbb F}_q)[X] \\
\quad \\
&= -\tilde{a}^p\, \tilde{y}^{rp} \, X^p+\tilde{a}^{rp} \, \tilde{y}^{r^2p} \, X^{pr} +p \, \tilde{T}_y(X) & \mod \, p^2\,W({\mathbb F}_q)[X] \\
\end{array}$$ with $\tilde{T}_y(X) \in W({\mathbb F}_q)[X]$. As $y \in {\mathbb F}_q$ and $a\in \Gamma$, we get: $$\begin{array}{lll}
\tilde{P}_y(X)^p-\tilde{P}_y(X)^{p^2}&= -\tilde{a}^p\, \tilde{y}^{rp} \, X^p-\tilde{a}^{p} \, \tilde{y}^{p} \, X^{pr} +p \, \tilde{T}_y(X) & \mod \, p^2 \,W({\mathbb F}_q)[X]\\
\end{array}$$ As a consequence, $$A=\tilde{a}^p \, \sum_{(i,j)\in I_1}
\, \frac{1}{p} \, \binom pi \, \binom pj \, X^{j+ip^s} \, \tilde{y}^{p-j+p^s\,(p-i)} +\tilde{T}_y(X) \quad \mod \, p\, \wp({\mathbb F}_q[X])$$ with $$I_1:=\{(i,j) \in I,\, (i,j)\neq(0,p) \,, (i,j)\neq(p,0)\}$$ So, $A$ reads: $$A=a^p \, \sum_{(i,j)\in I_1}
\, \frac{1}{p} \, \binom pi \, \binom pj \, X^{j+ip^s} \, y^{p-j+p^s\,(p-i)} +T_y(X)$$ with $T_y \in {\mathbb F}_q[X]$. We first consider the sum. Since, for $i=0$, $i=p$ and $j=p$, one gets monomials whose degree (after eventual reduction mod $\wp({\mathbb F}_q[X])$) is strictly lower than $1+p^{s}$, one can write: $$A=a^p \,\sum_{j=1}^{p-1} \, \frac{1}{p} \, \binom pj \, X^{j+ip^s} \, y^{p-j} +R_y(X)+T_y(X) \quad\, \mod \, \wp({\mathbb F}_q[X])$$ where $R_y(X)$ is a polynomial of ${\mathbb F}_q[X]$ with degree strictly lower than $1+p^s=1+r$. We now focus on the polynomial $T_y(X) \in {\mathbb F}_q[X]$. It is made of monomials which read either $X^{i_0+i_1\,p+\cdots +i_{s-1}\,p^{s-1}}$ with $i_0+i_1+\cdots+i_{s-1}=p$ or $X^{i_1\,p+\cdots+i_{s} \, p^s}$, with $i_1+i_2+\cdots+i_{s}=p$. As $X^{i_1\,p+\cdots+i_{s} \, p^s}=X^{i_1+\cdots+i_{s} \, p^{s-1}}$ mod $\wp({\mathbb F}_q[X])$, it follows that $T_y$ does not have any monomial with degree higher than $1+p^s$ after reduction mod $\wp({\mathbb F}_q[X])$. So, $$A=a^p \,\sum_{j=1}^{p-1} \,\frac{1}{p} \, \binom pj \, X^{j+ip^s} \, y^{p-j} +R^{[1]}_y(X) \quad\, \mod \, \wp({\mathbb F}_q[X])$$ where $R^{[1]}_y(X)$ is a polynomial of ${\mathbb F}_q[X]$ with degree strictly lower than $1+r$. Since, for all $j$ in $\{1,\cdots,p-1\}$, $f_j(X)= X^{j+p^{s+1}}-X^{1+jp^{s-1}}$ mod $\wp({\mathbb F}_q[X])$, we gather: $$A=a^p \,\sum_{j=1}^{p-1} \,\frac{1}{p} \, \binom pj \, y^{p-j}\, f_j(X) +R^{[2]}_y(X) \quad\, \mod \, \wp({\mathbb F}_q[X])$$ where $R^{[2]}_y(X)$ is a polynomial of ${\mathbb F}_q[X]$ with degree strictly lower than $1+r$. Then, $$\begin{array}{lll}
A &= \sum_{j=1}^{p-1} \, c_j(y) \, f_j(X)+R^{[2]}_y(X) \, & \mod \, \wp({\mathbb F}_q[X])\,
\end{array}$$ with $c_j(y):=a^p\, \frac{1}{p} \, \binom pj \, y^{p-j} \in {\mathbb F}_q$. It follows that: $$\begin{array}{lll}
A&= \sum_{j=1}^{p-1} \, (F^e-{id}) \, (c_j(y) \, W_j) +R^{[2]}_y(X) \, &\mod \, \wp({\mathbb F}_q[X]) \\
&\\
&= (F-{id})\, \sum_{j=1}^{p-1} \, \, P_j( W_j) +R^{[2]}_y(X) \, & \mod \, \wp({\mathbb F}_q[X])
\end{array}$$ where $P_j(W_j)= ({id}+F+\cdots +F^{e-1})\, (c_j(y)\, W_j) \in {\mathbb F}_q[W_j]$. We gather that: $$\wp(\sigma \, [W_0,W_p]-[W_0,W_p])=\wp \, ([0,\sum_{j=1}^{p-1} \, \, P_j( W_j)])+ [0, R^{[2]}_y(X)] \, \mod \,\wp(W_2({\mathbb F}_q[X]))$$ As a consequence, $[0, R^{[2]}_y(X)]$ lies in $\wp(W_2(K_S^{m_2}))$ and so, there exists $V \in K_S^{m_2}$ such that $V^p-V=R^{[2]}_y(X)$ Accordingly, $K(V)$ is a $K$-subextension of $K_S^{m_2}$ with conductor $1+deg(R^{[2]}_y(X)) \leq 1+r$. In particular, $K(V) \subset K_S^{r+1}=K={\mathbb F}_q(X)$, which implies that $R^{[2]}_y(X)\in \wp(K)$. Therefore, $$\wp(\sigma \, [W_0,W_p]-[W_0,W_p])=\wp \, ([0,\sum_{j=1}^{p-1} \, P_j( W_j)]) \, \mod \,\wp(W_2(K))$$ which allows to conclude that $\sigma \, (W_p)$ is in $L=K(W_0,W_1,\cdots,W_p)$.
4. As $L \subset K_S^{m_2}$ and $L \not \subset K_S^{m_2-1}$, the formula established in Proposition 6.5.3. yields: $$\begin{array}{ll}
g_L&=1+[L:K] \,(-1+\frac{m_2}{2} )-\frac{1}{2} \sum_{j=0}^{m_2-1} [K_S^{j}:K] \\
&\\
&=1+p^{2+(p-1)e}\, (-1+\frac{p^{s+1}+p+1}{2})-\frac{1}{2}(r+2+(m_2-p-(r+2)+1)\, p+
\sum_{i=1}^{p-1} \, p^{1+i\,e} )\\
&\\
&= \frac{1}{2} \, p^{2+(p-1)e} \, (p^{s+1}+p-1) -\frac{1}{2} \, (p^s+p^{s+2}-p^{s+1}+
\sum_{i=1}^{p-1} \, p^{1+i\,2\,s}) \\
&\\
&=\frac{1}{2} \, p^{2+(p-1)e} \, (p^{s+1}+p-1) -\frac{1}{2} \,p^s (p^2-p+1) -\frac{1}{2} \, p^{2s+1} (1+q+q^2+\cdots +q^{p-2}) \qquad \square
\end{array}$$
The preceding proposition can be generalized to construct a big action endowed with a second ramification group $G_2$ abelian of exponent as large as we want.
We keep the notation defined above. In particular, $q=p^e$, with $e=2s$ and $s\geq 1$. Let $n \geq 2$. Put $m_n:=1+p^{n-1}\, (1+p^s)$. If $\frac{q}{-1+m_n/2} >\frac{2\,p}{p-1}$, the pair $(C(m_n), G(m_n))$, as defined in Proposition 6.5, is a big action with a second ramification group $G_S(m_n)$ abelian of exponent at least $p^n$.
**Proof:** Proposition 6.5.4 first ensures that $(C(m_n), G(m_n)$ is a big action. Consider the $p^n$-cyclic extension $K(W_1,\cdots, W_n)/K$ parametrized as follows, with Witt vectors of length $n$: $$[W_1,\cdots,W_n]^p-[W_1,\cdots,W_n]=[f_0(X), 0 ,\cdots, 0]$$ where $f_0(X)=a\,X^{1+r}$ is defined as in Proposition 6.7, i.e. $r=p^s$, $a^r+a=0$ , $a\neq 0$. The same proof as in Proposition 6.7.1 shows that all places of $S$ completely split in $K(W_1,\cdots,W_n)$. Moreover, by [@Ga] (Thm. 1.1) the conductor of the extension $K(W_1,\cdots,W_n)$ is $1+max \{p^{n-1}\, (1+p^s),0\}=m_n$. It follows that $K(W_1,\cdots, W_n)$ is included in $K_S^{m_n}$. Therefore, $G_S(m_n)$ has a quotient of exponent $p^n$ and the claim follows. $\square$
Tne next proposition is an analogue of Proposition 6.7 in the case where $e$ is odd. We does not mention the proof which is mainly similar to the proof of Proposition 6.7.
We keep the notation defined above. In particular, $K={\mathbb F}_q(X)$ with $q=p^e$. Assume $e=2\,s-1$, with $s \geq 2$, and put $r:=\sqrt{qp}=p^{s}$. We define $$\forall \, i \in \{1,\cdots,p-1\}, \; f_i(X)=X^{ir/p} \, (X^q-X)=X^{ip^{s-1}}\, (X^q-X)$$ $$\forall \, i \in \{1,\cdots,p-1\}, \; g_i(X)=X^{ir/{p^2}} \, (X^q-X)=X^{ip^{s-2}}\, (X^q-X)$$ Let $L:=K(W_i,V_j)_{1\leq i \leq p, 1 \leq j \leq p-1 }$ be the extension of $K$ parametrized by the Artin-Schreier-(Witt) equations: $$\forall \, i \in \{1,\cdots,p-1\}, \; W_i^q-W_i=f_i(X) \quad \mbox{and} \quad
\forall \, j\, \in \{1,\cdots,p-1\}, \; V_j^q-V_j=g_j(X)$$ $$[W_1,W_p]^p-[W_1,W_p]= [X^{1+p^s},0]-[X^{1+p^{s-1}},0]$$ For all $i$ and $j$ in $\{1,\cdots, p-1\}$, put $L_{i,0}:=K(W_k)_{1\leq k \leq i}$ and $L_{p-1,j}:=K(W_i,V_k)_{1\leq i \leq p-1, 1 \leq k \leq j}$.
1. $L$ is an abelian extension of $K$ such that every place in $S$ completely splits in $L$. Then, $$\forall \, i, \, j \in \{1,\cdots,p-1\}, \, L_{i,0} \subset K_S^{p^s+i+1} \quad, \, \, L_{p-1,j} \subset K_S^{p^{s+1}+j+1} \quad \mbox{and} \quad L \subset K_S^{m_2}$$ where $m_2=p^{s+1}+p+1$ is the integer defined in Proposition 6.4. (see table below.)
2. The extension $L/K$ has degree $[L:K]=p^{2(p-1)e+1}$. Let $G_L$ be its Galois group. Then $$G_L \simeq {\mathbb Z}/p^2{\mathbb Z}\times ({\mathbb Z}/p{\mathbb Z})^t \quad \mbox{with} \,\; t=2\,(p-1)\,e-1$$
3. The extension $L/K$ is stable under the translations: $X\rightarrow X+y$, with $y \in {\mathbb F}_q$. Therefore, the translations by ${\mathbb F}_q$ extend to form a $p$-group of ${\mathbb F}_q$-automorphisms of $L$, say $G$, with the following exact sequence: $$0\longrightarrow G_L\longrightarrow G \longrightarrow{\mathbb F}_q\longrightarrow 0$$
4. Let $g_L$ be the genus of the extension $L/K$. Then, $$g_L= \frac{1}{2} \, \{ \, p^{1+(2p-1)e} \, (p^{s+1}+p-1)- p^{(p-1)e} \, (p^{s+1}-p^s-p+1)- p^s+p^e\, (\sum_{i=0}^{2p-3} q^i) \}$$ In particular, when $e$ grows large, $g_L\sim \frac{1}{2}\, p^{2+4s(p-1)+s}$ and $t=O(log_p \, g_L)$.
We gather in the table below the conductors, degrees and equations of each extension.
$L_{i,j}$ conductor $m$ $[L_{i,j}:K]$ New equations
----------------- --------------------------------- ------------------- ------------------------------------
$K$ $0 \leq m \leq r+1=p^s+1$ $1$
$L_{1,0}$ $m=r+2=p^s+2 $ $p^e$ $W_{1}^q-W_{1}=f_1(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_{i,0} $ $m=p^{s}+i+1$ $p^{ie}$ $W_i^q-W_i=f_i(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_{p-1,0}$ $p^{s}+p \leq m \leq p^{s+1}+1$ $p^{(p-1)e}$ $W_{p-1}^q-W_{p-1}=f_{p-1}(X)$
$L_{p-1,1}$ $m=p^{s+1}+2=m_2-(p-1) $ $p^{pe}$ $V_{1}^q-V_{1}=g_1(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_{p-1,j}$ $m=p^{s+1}+j+1=m_2-(p-j)$ $p^{(p+j-1)e}$ $V_j^q-V_j=g_j(X)$
$\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$ $\cdots \cdots$
$L_{p-1,p-1}$ $ m=p^{s+1}+p=m_2-1$ $p^{2(p-1)e}$ $V_{p-1}^q-V_{p-1}=g_{p-1}(X)$
$L$ $m=p^{s+1}+p+1=m_2$ $p^{1+2\,(p-1)e}$ $[W_1,W_p]^p-[W_1,W_p]=$
$[X^{1+p^s},0]-[X^{1+p^{s-1}},0] $
A local approach to big actions.
================================
Let $(C,G)$ be a big action. We recall that there exists a point $\infty \in C$ such that $G$ is equal to $G_1(\infty)$ the wild inertia subgroup of $G$ at $\infty$, which means that the cover $\pi: C \rightarrow C/G$ is totally ramified at $\infty$. Moreover, the quotient curve $C/G$ is isomorphic to the projective line: ${\mathbb P}_k^1$ and $\pi$ is étale above the affine line: ${\mathbb A}_k^1={\mathbb P}_k^1-\pi(\infty)=Spec\, k[T]$. The inclusion $k[T] \subset k((T^{-1}))$ induces a Galois extension $k(C) \otimes_{k(T)} k((T^{-1}))=:k((Z))$ over $k((T^{-1}))$ with group equal to $G$ and ramification groups in lower notation equal to $G_i:=G_i({\infty})$. Then, the genus of $C$ is given by the formula: $g=\frac{1}{2}\, (\sum_{i\geq 2} (|G_i|-1)) >0$ (see ). It follows that: $$\frac{|G|}{\sum_{i\geq 2} (|G_i|-1)}=\frac{|G|}{2\, g} > \frac{p}{p-1}.$$ This leads to the definition below.
We call “local big action” any pair $(k((Z)),G)$ where $G$ is a finite $p$-subgroup of $Aut_k(k((Z))$ whose ramification groups in lower notation at $\infty$ satisfy the two inequalities: $$g(G):=\frac{1}{2} (\sum_{i\geq 2} (|G_i|-1))>0 \qquad \mbox{and} \qquad \frac{|G|}{g(G)} > \frac{2\: p}{p-1}.$$
It follows from the Katz-Gabber Theorem (see [@Ka] Thm. 1.4.1 or [@PGi] cor. 1.9) that big actions $(C,G)$ and local big actions $(k((Z)),G)$ are in 1-to-1 correspondance via the following functor induced by the inclusion $k[T] \subset k((T^{-1}))$:
$$\left\{
\aligned \mbox{finite \'etale Galois covers of Spec k[T]} \\
\mbox{ with Galois group a p-group}
\endaligned
\right\} \quad \longrightarrow \quad \left\{
\aligned \mbox{finite \'etale Galois covers of Spec } k((T^{-1}))\\
\mbox{with Galois group a p-group} \endaligned
\right\}$$
Therefore, we can infer from the global point of view properties related to local extensions that would be difficult to prove directly. For instance, if $(k((Z)), G)$ is a local big action, we can deduce that $G_2$ is stricly included in $G_1$. Furthermore, we obtain: $$\frac{|G|}{g(G)^2} \leq \frac{4\, p}{(p-1)^2}.$$
[Co-Ka-Ko]{} R. Auer, *Ray Class Fields of Global Function Fields with Many Rational Places*. Dissertation at the University of Oldenburg, www.bis.uni-oldenburg.de/dissertation/ediss.html, (1999). R. Auer, *Ray class fields of global function fields with many rational places.* Acta Arith. **95** (2000), no. 2, 97-122. N. Bourbaki, *Algèbre commutative.* Eléments de Mathématiques, Masson, Paris, (1983) I. Bouw, *The $p$-rank of curves and covers of curves.* Courbes semi-stables et groupe fondamental en géométrie algébrique (Luminy, 1998), 267–277, Progr. Math., 187, Birkhäuser, Basel, (2000). M. Conder, *Hurwitz groups: a brief survey*. Bull. Amer. Math. Soc. (N.S.) **23** (1990), no. 2, 359-370. N. Elkies *Linearized algebra anf finite groups of Lie type. I. Linear and symplectic groups* in *Applications of curves over finite fields*, (Seattle, WA, 1997), Contemporary Mathematics, vol 245, American Mathematical Society, (Providence, RI, 1999) N. Elkies, *The Klein quartic in number theory. The eightfold way,* 51-101, Math. Sci. Res. Inst. Publ., 35, Cambridge Univ. Press, Cambridge, (1999). M. Garuti, *Linear systems attached to cyclic inertia.* Arithmetic fundamental groups and noncommutative algebra (Berkeley, CA, 1999), 377-386, Proc. Sympos. Pure Math., **70**, Amer. Math. Soc., Providence, RI, 2002. P. Gille, *Le groupe fondamental sauvage d’une courbe affine en caractéristique $p>0$.* Courbes semi-stables et groupe fondamental en géométrie algébrique (Luminy, 1998), 217–231, Progr. Math., 187, Birkhäuser, Basel, 2000. M. Giulietti, G. Korchmáros, *On large automorphism groups of algebraic curves in positive characteristic* arXiv:0706.2320 , 15 Jun 2007 D. Goss, *Basic structures of function field arithmetic*, Ergebnisse der Mathematik und ihrer Grenzgebiete (3), vol.35, (Springer, Berlin, 1996) B. Huppert, *Endliche Gruppen. I.*, Die Grundlehren der Mathematischen Wissenschaften, Band 134 Springer-Verlag, Berlin-New York (1967) N. Katz, *Local-to-global extensions of representations of fundamental groups.* Ann. Inst. Fourier (Grenoble) **36** (1986), no. 4, 69-106. K. Lauter, *A Formula for Constructing Curves over Finite Fields with Many Rational Points*. Journal of Number Theory **74** (1999), no. 1, 56-72. C. Lehr, M. Matignon, *Automorphism groups for $p$-cyclic covers of the affine line*. Compositio Math. **141** (2005). M. Marshall, *Ramification groups of abelian local field extensions.* Canad. J. Math. **23** (1971) 271–281 A.M. Macbeath, *On a theorem of Hurwitz*. Proc. Glasgow Math. Assoc. **5** 90-96 (1961). A.M. Macbeath, *On a curve of genus $7$.* Proc. London Math. Soc. (3) **15** 527-542 (1965). J. S. Milne, *Etale cohomology*, Princeton Mathematical Series, 33, Princeton University Press, (Princeton, N.J., 1980) S. Nakajima, *$p$-ranks and automorphism groups of algebraic curves*. Trans. Amer. Math. Soc. **303** (1987). M. Rocher, *Large $p$-groups actions with a $p$-elementary abelian second ramification group.* (2008), available on arXiv. M. Rocher, *Large $p$-groups actions with $\frac{|G|}{g^2} \geq
\frac{4}{(p^2-1)^2}$* (2008). In preparation. H. L. Schmid, *Über die Automorphismen eines algebraischen Funktionenkörpers von Primzahlcharacteristik*, J.Reine Angew. Math. **179**, p5-15 (1938) J-P. Serre, *Corps locaux.* Deuxième édition. Hermann, Paris, (1968). J. Silverman, *The arithmetic of elliptic curves.* Graduate Texts in Mathematics, 106. Springer-Verlag, New York, 1986. B. Singh, *On the group of automorphisms of function field of genus at least two* J. Pure Appl. Algebra **4** (1974), 205-229. H. Stichtenoth, *Über die Automorphismengruppe eines algebraischen Funktionkorpers von Primzahlcharakteristik I, II*, Arch. Math. (Basel) **24** (1973). H. Stichtenoth, *Algebraic function fields and codes.* Universitext. Springer-Verlag, Berlin, (1993). M.Suzuki, *Group Theory I*. Grundlehren der Mathematischen Wissenschaften , 247. (Springer-Verlag, Berlin-New York, 1982.) M. Suzuki, *Group theory. II*. Grundlehren der Mathematischen Wissenschaften, 248. Springer-Verlag, New York, (1986)
Michel MATIGNON\
Institut de Mathématiques de Bordeaux, Université de Bordeaux I, 351 cours de la Libération, 33405 Talence Cedex, France\
e-mail : [[email protected]]{}
Magali ROCHER\
Institut de Mathématiques de Bordeaux, Université de Bordeaux I, 351 cours de la Libération, 33405 Talence Cedex, France\
e-mail : [[email protected]]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present optical observations of supernova SN2014C, which underwent an unprecedented slow metamorphosis from H-poor type Ib to H-rich type IIn over the course of one year. The observed spectroscopic evolution is consistent with the supernova having exploded in a cavity before encountering a massive shell of the progenitor star’s stripped hydrogen envelope. Possible origins for the circumstellar shell include a brief Wolf-Rayet fast wind phase that overtook a slower red supergiant wind, eruptive ejection, or confinement of circumstellar material by external influences of neighboring stars. An extended high velocity H$\alpha$ absorption feature seen in near-maximum light spectra implies that the progenitor star was not completely stripped of hydrogen at the time of core collapse. Archival pre-explosion Subaru Telescope Suprime-Cam and Hubble Space Telescope Wide Field Planetary Camera 2 images of the region obtained in 2009 show a coincident source that is most likely a compact massive star cluster in NGC7331 that hosted the progenitor system. By comparing the emission properties of the source with stellar population models that incorporate interacting binary stars we estimate the age of the host cluster to be $30 - 300$ Myr, and favor ages closer to $30$ Myr in light of relatively strong H$\alpha$ emission. SN2014C is the best-observed member of a class of core-collapse supernovae that fill the gap between events that interact strongly with dense, nearby environments immediately after explosion and those that never show signs of interaction. Better understanding of the frequency and nature of this intermediate population can contribute valuable information about the poorly understood final stages of stellar evolution.'
author:
- |
D. Milisavljevic, R. Margutti, A. Kamble, D. J. Patnaude, J. C. Raymond, J. J. Eldridge, W. Fong, M. Bietenholz, P. Challis, R. Chornock, M. R. Drout, C. Fransson, R. A. Fesen,\
J. E. Grindlay, R. P. Kirshner, R. Lunnan, J. Mackey, G. F. Miller,\
J. T. Parrent, N. E. Sanders, A. M. Soderberg, B. A. Zauderer
title: 'Metamorphosis of SN2014C: Delayed interaction between a hydrogen poor core-collapse supernova and a nearby circumstellar shell'
---
Introduction {#sec:Intro}
============
Recent observations have convincingly demonstrated that massive stars can experience an eruptive mass loss episode $\sim 1$ yr before core collapse. Such episodes have been confirmed in H-rich type IIn [@Ofek13; @Smith14; @Ofek14] and H-poor, He-rich type Ibn [@Pastorello07; @Foley07] supernovae. Although the brief timescales between eruption and supernova explosion have been anticipated in special cases of very massive stars [@Woosley07; @Quataert12], a growing number of systems are being discovered that fall outside most theoretical regimes and that challenge many long held notions of stellar evolution [@Stritzinger12; @Pastorello13; @Mauerhan13; @Margutti14; @SmithArnett14; @Moriya15].
The most favored candidate progenitors of type IIn and Ibn supernovae are luminous blue variable (LBV) stars [@Kotak06; @Pastorello07; @Gal-Yam09-05gl], and stars leaving the LBV phase and entering a Wolf-Rayet (W-R) phase [@Smith12; @Pastorello15] that may be exceptionally brief ($\la
10^3$ yr; @Dwarkadas11). LBV stars are prime suspects because they undergo recurrent mass-loss episodes that eject large portions of their hydrogen envelope ($\ga 1\;M_{\odot}$; @Smith06). However, traditional stellar evolutionary theory has predicted LBV stars to be a transitional phase lasting $10^4$ to $10^5$ yr before evolving to a compact and hydrogen-poor W-R star lasting a few $\times 10^5$ yr [@Meynet94; @Langer94; @Maeder00]. This predicted timescale is discrepant with the $\sim 1$ yr timescale between eruption and core collapse observed in at least some type IIn and Ibn systems [@Pastorello07; @Pastorello13; @Smith14].
{width="\linewidth"}
The extent to which eruptive and/or accelerated mass loss may directly precede core collapse in a wider range of supernovae is poorly constrained, particularly in type IIb, Ib, and Ic supernovae (SN Ibc) where the progenitor star has been significantly stripped of its hydrogen envelope. Some insight into this issue was provided by optical [@Maeda15] and radio+X-ray observations [@Kamble15] of the type IIb SN2013df. Both studies concluded that the progenitor star experienced enhanced mass loss of $(3-8) \times 10^{-5}\;
M_{\odot}\;\rm yr^{-1}$ (for wind velocity of 10 ) in the final centuries leading up to the supernova explosion. Intriguingly, not all type IIb supernovae exhibit the same mass loss enhancement (e.g., SN2011dh; @Krauss12 [@Maeda14; @deWitt15]). The underlying physical reasons for this dichotomy are presently unclear, but may be related to the progenitor star size and various channels of binary interaction [@Maeda15].
W-R stars have long been suggested as an obvious progenitor of SN Ibc because they are deficient in hydrogen [@Gaskell86]. However, only progenitor stars with much cooler atmospheres than those of W-R stars have been unambiguously detected at the explosion sites of type IIb supernovae [@Aldering94; @Maund11; @VanDyk14], and no secure direct identification has yet been made of a type Ib or Ic progenitor system [@vanDyk03; @Smartt09; @Eldridge13]. A possible exception is the type Ib iPTF13bvn [@Cao13; @Groh13; @Bersten14; @Eldridge15; @Fremling14].
The observed number of W-R stars is probably insufficient to account for all SN Ibc, and low mass He stars in binary systems are more likely to be the dominant progenitor channel [@Podsiadlowski92; @Wellstein99; @Smartt09; @Smith11; @Claeys11; @Langer12; @Eldridge13; @Dessart15; @Eldridge15]. However, because W-R stars are difficult to detect in broadband images and large populations may presently be unaccounted for [@Shara13; @Massey14; @Massey15], W-R stars may still represent a significant fraction of SN Ibc.
{width="0.95\linewidth"}
In this paper we present and analyze optical observations of a supernova that underwent a remarkable metamorphosis consistent with a delayed interaction between the core-collapse explosion and a massive circumstellar shell formed from the progenitor star’s partially stripped hydrogen envelope. In Section \[sec:Observations\], we present multi-epoch spectra of SN2014C that follow its slow spectroscopic transformation from normal type Ib to strongly interacting type IIn, as well as archival pre-explosion images that show a luminous source coincident with the supernova. In Section \[sec:Discussion\], we discuss properties of the partially stripped progenitor star and its host environment, as well as possible physical mechanisms behind the formation of the circumstellar shell. Finally, in Section \[sec:Conclusions\], we summarize our conclusions and outline how future work on late-interacting SN Ibc systems like SN2014C can contribute unique information about poorly understood stages of stellar evolution and mass loss that immediately precede core collapse.
Observations and Results {#sec:Observations}
========================
SN 2014C was discovered at coordinates $\alpha = 22^{\rm h}37^{\rm
m}05\fs60$ and $\delta = +34\degr24'31\farcs9$ (J2000.0) in the nearby spiral galaxy NGC 7331 on 2014 January 5.1 UT by the Lick Observatory Supernova Search [@Kim14]. NGC7331 is known to have hosted two previous supernovae: SN1959D [@Humason59] and SN2013bu [@Itagaki13]. In Figure \[fig:images\], a finding chart for SN 2014C and its immediate environment made from an $r'$-band image obtained with the 6.5m MMT telescope and the MMTCam instrument[^1] is shown.
Shortly after the supernova’s discovery, optical spectra obtained by J. Zhang & X. Wang with the 2.4m telescope of Yunnan Observatories and L. Tartaglia et al. with the Asiago 1.82m Copernico Telescope led to the classification of the supernova as a hydrogen-poor type Ib near maximum light. It was at this point that our group initiated a multi-wavelength (radio-to-X-ray) observing campaign. Optical/ultraviolet light curves and X-ray data are presented in Margutti et al. (2015), and radio data are presented in Kamble et al. (2015). Results from a search for pre-explosion images and optical spectroscopy during the first year of monitoring are presented here. Throughout this paper we adopt the Cepheid distance of $14.7 \pm 0.6$ Mpc to the host galaxy NGC7331 [@Freedman01], and uncertainties are quoted at the $1\sigma$ confidence level (unless otherwise noted).
Pre-Explosion Images {#sec:pre-explosion}
--------------------
NGC7331 is an early type spiral galaxy that has been studied closely from X-ray to radio wavelengths (see, e.g., @Thilker07). Consequently, extensive archival data exist that cover the region of SN2014C prior to explosion. We retrieved optical images originally obtained on 2009 August 21 with the 8.3m Subaru Telescope and the Suprime-Cam instrument [@Miyazaki02], via the SMOKA Science Archive[^2]. $B$, $V$, and $R$ filters were used with exposure lengths of 200, 300, and 560s, respectively. Images were bias-corrected and flat-fielded following standard procedures with the IRAF software[^3], and the absolute positions were obtained using the `IMWCS` software[^4] and the US Naval Observatory B-1.0 catalog [@Monet03].
We also retrieved archival Hubble Space Telescope (HST) images covering the location of SN2014C from the Mikulski Archive for Space Telescopes. The images were obtained with the Wide Field Planetary Camera 2 (WFPC2) and the F658N filter ($\lambda_{\rm C} = 6591$ Å; $\delta\lambda$ = 29 Å) on 2009 January 1 with a total integration time of $3\times600$s under program 11966 (PI: Regan). We used the AstroDrizzle package of the DrizzlePac 2.0 software[^5] to remove geometric distortion, correct for sky background variations, flag cosmic-rays, and drizzle the individual frames together.
[lccccccc]{} 2014 Jan 09 & 56666.08 & -4 & FLWO&FAST & 300 & $3500-7400$ &7.0\
2014 May 06 & 56783.43 & 113 & MMT&Blue Channel & 300 & $3300-8500$ &7.0\
2014 Oct 22 & 56953.28 & 282 & LBT&MODS & Dual &$5000-10000$&3.0\
2015 Jan 21 & 57043.08 & 373 & MMT&Blue Channel & 300 & $3500-8500$&7.0\
2015 Apr 25 & 57137.47 & 467 & MMT&Blue Channel &1200&$4270-5300$&1.5
\[tab:speclog\]
A visible source is seen in close proximity with SN2014C in all three filters of the Subaru images (Figure \[fig:images\]), as well as in the HST F658N image. The source in the Subaru images is unresolved, with a point spread function (PSF) that we fit with a Gaussian having a full width at half maximum (FWHM) comparable to that observed in nearby stars ($\approx 0\farcs 8$; average of five stars). The source seen in the HST image is resolved, non-uniform and extended in distribution. It is slightly elongated in the north-south direction and is contained within an approximate effective radius of $\approx 0 \farcs 24$.
We used the `GEOMAP` task of IRAF to determine a spatial transformation function between the MMT, Subaru, and HST images, and the `GEOTRAN` task to apply the transformation. Figure \[fig:enlarge\] shows the location of the pre-explosion sources with respect to SN2014C as observed in the MMT image. A small offset of $0\farcs 24 \pm 0.05$ exists between the centroid of SN 2014C observed in the MMT image and the emission peak of the source observed in the Subaru image. A small offset is also seen between the location of SN 2014C and the approximate center of the extended region observed in the HST F658N image.
We performed PSF photometry on the Subaru images using the `DAOPHOT` package in IRAF in order to constrain emission properties of the potential progenitor system. The images were calibrated via relative photometry using 10 stars in Sloan Digital Sky Survey (SDSS) images that cover the field of view. Photometric transformations were made from @Jordi06 to put SDSS photometry into the $BVRI$ system. The apparent magnitudes of the coincident source are $m_B = 22.18 \pm 0.13$, $m_V = 21.13 \pm 0.09$, and $m_R =
20.28 \pm 0.06$. The region surrounding the source has considerable galaxy light and the reported apparent magnitudes may overestimate the brightness. The uncertainties reflect the error in PSF fitting and do not include possible error due to contamination from galaxy light.
We measured the flux contained within a $0\farcs 5$ aperture centered on the coincident emission source observed in the HST/WFPC2 F658N image. The sum count rate $1.00 \pm 0.01$ counts s$^{-1}$ was multiplied by the modified PHOTFLAM parameter ($9.87 \times 10^{-17}\;
\rm erg\,s^{-1}\,cm^{-2}$Å$^{-1}$) and the effective bandpass of the filter given by the RECTW parameter (39.232 Å). The integrated flux, which is a sum of narrow H$\alpha$, \[\] $\lambda\lambda$6548, 6583, and continuum emission, is $(3.87 \pm
0.04) \times 10^{-15}\;\rm erg\,s^{-1}\,cm^{-2}$. By estimating the contribution from continuum emission from the spectral energy distribution (SED) of the Subaru photometry, and assuming the contribution from the \[\] lines to be $\approx 0.13$ per cent of the continuum subtracted flux, we derive an observed H$\alpha$ flux of $2.9 \times 10^{-15}\;\rm erg\,s^{-1}\,cm^{-2}$, and an unabsorbed luminosity of $4.3\times 10^{38}\;\rm erg\,s^{-1}$ after correcting for $E(B-V)_{\rm total}$ (cf. Section \[sec:spectra\]).
{width="0.80\linewidth"}
Optical Spectroscopy {#sec:spectra}
--------------------
Low-resolution optical spectra of SN2014C were obtained from three telescopes: The F. L. Whipple Observatory (FLWO) 1.5m Tillinghast telescope mounted with the FAST instrument [@Fabricant98], the 6.5m MMT Telescope mounted with the Blue Channel instrument [@Schmidt89], and the $2\times8.4$m Large Binocular Telescope mounted with the MODS instrument [@Pogge10]. Details of the observations are provided in Table \[tab:speclog\].
![The spectrum of SN 2014C compared to those of other type Ib supernovae near maximum light. The spectra of SN2014C and SN2008D have been corrected for extinction using $E(B-V)$ values of 0.75 mag (this paper) and 0.5 mag [@Soderberg08; @Modjaz09], respectively, whereas the spectrum of iPTF13bvn has not been corrected [@Cao13; @Srivastav14]. The data were originally published in @Modjaz09 and @Srivastav14 and were digitally retrieved from WISEREP [@Yaron12]. Also shown is a synthetic spectrum of SN2014C created with `SYN++`. Some absorption features dominated by single ions with projected Doppler expansion velocities of $(1.3 \pm 0.1) \times 10^{4}$ are identified, as well as high velocity (HV) hydrogen that spans $(1.3-2.1) \times 10^{4}$ .[]{data-label="fig:syn++"}](f4.eps){width="\linewidth"}
Our five epochs of optical spectra of SN2014C are plotted in Figure \[fig:spectra\]. The explosion date is not tightly constrained by the light curve, so we use the peak in the V-band on 2014 January 13 as the reference from which phase in days is measured (Margutti et al. 2015).
Standard procedures to bias-correct, flat-field, and flux calibrate the data were followed using the IRAF/PYRAF software[^6] and our own IDL routines. A recession velocity of 990 , determined from many narrow emission lines including \[\] $\lambda\lambda$4959, 5007, H$\alpha$, and \[\] $\lambda\lambda$6716, 6731, was removed from all spectra. Line identifications and estimates of expansion velocities of the photospheric spectra were made with the supernova spectrum synthesis code `SYN++` [@Thomas11]. Narrow emission line identifications in late-time spectra were made using lists provided in @Fesen96.
The foreground extinction due to the Milky Way is $E(B-V)_{mw} = 0.08$ mag [@Schlafly11]. The host internal extinction was estimated by measuring the equivalent width (EW) of D absorption in our optical spectra and following the prescriptions of @Turatto03. The EW(NaID) from the day $-4$ spectrum is $4.25
\pm 0.07$ Å. Using the lower branch of the @Turatto03 relation (see their Fig. 3), this measurement of EW() implies $E(B-V)_{\rm host} = 0.67 \pm 0.01$ mag. We adopted a total extinction of $E(B-V)_{\rm total}=0.75 \pm 0.08$ mag, which combines the Galactic extinction with the inferred host extinction. We confirmed that this extinction estimate provides an appropriate $(B-V)$ color correction to the color indices of SN2014C to match those of other type Ib supernovae (Margutti et al. 2015). All extinction corrections made in this paper use $E(B-V)_{\rm total}$, in combination with the standard reddening law of @Cardelli89 assuming $R_{\rm V} = 3.1$.
Spectroscopic Metamorphosis {#sec:morph}
---------------------------
Our optical spectrum obtained 2014 January 09 (day $-4$) shows features clearly associated with , , , exhibiting velocities of $(1.3 \pm 0.1) \times
10^4$ (Figure \[fig:syn++\]). These are features regularly seen in type Ib supernovae [@Filippenko97], including well-known examples SN2008D [@Modjaz09] and iPTF13bvn [@Srivastav14]. An absorption centered around 6150 Å is typically associated with , however we find that the spectrum is best fit with a combination of high velocity (HV) H$\alpha$ features spanning $(1.3-2.1) \times 10^4$ . This choice of line identification is discussed in greater detail in Section \[sec:hydrogen\].
Follow-up spectroscopic observations did not resume until day 113 when the supernova returned from behind the Sun. Interestingly, the day 113 spectrum of SN2014C exhibits a mix of standard and non-standard emissions. The broad \[\] $\lambda\lambda$ 6300, 6364 and \[\] $\lambda\lambda$7291, 7324 emission observed in the spectrum are a normal feature of type Ib supernovae several months after explosion and are associated with inner metal-rich ejecta that are radioactively heated by $^{56}$Co. Most unusual, however, was conspicuous emission centered around the H$\alpha$ line with an overall FWHM of 1400 that had emerged. This emission continued to grow in strength relative to other emission lines over the next several months (Figure \[fig:spectra\]).
[*An extraordinary event must have occurred while SN2014C was hidden behind the Sun.*]{} An intermediate-width H$\alpha$ feature is normally seen only in type IIn supernovae, where it is associated with radiative shocks in dense clouds [@Chugai94]. The interaction decelerates the blast wave and a dense shell traveling approximately at the shock velocity is formed. Accordingly, we interpret the conspicuous change in H$\alpha$ emission from SN2014C to be the result of the supernova having encountered dense H-rich circumstellar material (CSM) between Februrary and May 2014. Consistent with this scenario (see, e.g., @Chevalier06), strong radio and X-ray emission accompanied the sudden increase in H$\alpha$ emission in SN2014C as the shock continued to strongly interact with the density spike in CSM (Margutti et al. 2015; Kamble et al. 2015).
![Evolution of the emission line profile around H$\alpha$. On day $-4$ only unresolved (FWHM $\approx 250$ ) H$\alpha$ and \[\] $\lambda\lambda$6548, 6583 lines are observed. Subsequently, beginning on day 113, an intermediate component (FWHM $\approx 1200$ ) is observed. A model fitting the day 373 profile is shown. Scaling parameters are given to the left of each spectrum.[]{data-label="fig:profile"}](f5.eps){width="0.8\linewidth"}
We attempted to decompose the complex line profile centered on H$\alpha$ in the spectra into multiple Gaussian features. After removing the linear continuum and running a least-squares fitting routine, we found that the profile could be reasonably reproduced with narrow components of H$\alpha$ and \[\] 6548, 6583 lines having instrumentally unresolved FWHM velocities of $\approx 250$ and an intermediate-width component with a FWHM width of $1200
\pm 100$ (Figure \[fig:profile\]). The H$\alpha$ profile extends to 6520 Å and 6610 Å, which sets limits on the velocity of the emitting shocked CSM to $-2000$ and $+2200$ . The narrow components are presumably associated with wind material that is being photoionized by X-rays of the forward shock, and the intermediate component is associated with the shock and/or ejecta running into CSM. The day $-4$ spectrum, which has only narrow components, is most likely a combination of emission local to the supernova and from the entire host massive star cluster (see Section \[sec:pre-explosion\]). We interpret emission at later epochs to be be dominated by emission from supernova-CSM interaction.
Beginning with the day 282 spectrum and continuing with the day 373 spectrum, the emissions are increasingly complex and originate from several distinct regions. Figure \[fig:lineID\] shows an enlargement of the day 373 spectrum corrected for extinction and a complete list of identified emission features. Several narrow, unresolved emission lines are observed including \[\] $\lambda$4363 and $\lambda\lambda$4959, 5007, \[\] $\lambda$3869, $\lambda$4686, and \[\] $\lambda$5755. Also seen are several narrow, unresolved coronal lines including \[\], \[\], \[\], \[\], and \[\]. We attribute this emission to ionization of the pre-shock circumstellar gas by X-rays emitted by the shocked gas. The strongest constraint on the wind velocity comes from the day 474 spectrum that has the highest resolution of all our data. We measure a FWHM width of 1.5 Å (which is unresolved) in the \[\] $\lambda$4363 emission line. This sets an upper limit of $< 100$ for the unshocked wind velocity.
{width="\linewidth"}
Broad emission centered around the \[\] $\lambda\lambda$4959, 5007 lines is seen in our spectra beginning on day 282 and continues to be visible through our last spectrum obtained on day 474. The width of the emission is difficult to measure since it blends with H$\beta$ blueward of 4959 Å and another source of emission redward of 5070 Å. We estimate that the velocity width must be $\ga 3500$ , meaning that the emission originates from a region different than the shocked CSM. Presumably it is emission from oxygen-rich stellar ejecta being excited by the reverse shock. Broad \[\] emission is normally only seen in supernovae many years to decades after core collapse [@Milisavljevic12]. But in the case of SN2014C, the supernova-CSM interaction may have accelerated its dynamical evolution.
By comparing the overall emission profile around 7300 Å to that observed around \[\] $\lambda\lambda$6300, 6364, we find that the velocity distribution is best matched when measured with respect to the \[\] $\lambda\lambda$7291, 7324 lines ($\lambda_C =
$7307 Å) as opposed to the \[\] $\lambda\lambda$7319, 7330 lines ($\lambda_C = $7325 Å). Detecting emission from \[\] and \[\], but not from \[\], implies that the ejecta associated with this emission have densities that are $\ga 10^{6}$ cm$^{-3}$ [@Fesen99].
The dereddened day 373 optical spectrum exhibits a blue pseudo-continuum, which is sometimes seen in type IIn and Ibn supernovae and has been attributed to fluorescence from a number of blended emission lines (@Fransson02 [@Foley07; @Smith09]; Figure \[fig:IInIbn\]). One of the strongest expected fluorescence-pumped lines is at 8451 Å[@Fransson02]. Near this wavelength region we detect a minor emission peak centered at 8446 Å in our day 282 spectrum that is blended with the near-infrared triplet. This feature is most likely the recombination line at 8446 Å and not . We also do not observe Fe II lines at 9071, 9128, or 9177 Å, which would be expected to accompany the Fe II $\lambda$8451 line [@Fransson02].
Properties of the unshocked CSM
-------------------------------
Properties of the surrounding unshocked CSM shed from the progenitor star of SN2014C can be estimated using the relative line strengths of narrow features observed in our optical spectra. The forbidden oxygen lines provide a lower limit to the density. We do not detect narrow \[\] $\lambda$3727, but we do detect \[\] $\lambda\lambda$4959, 5007, indicating that electron densities are well above $10^4$ cm$^{-3}$ in this emitting region [@Osterbrock06]. An upper limit to the density can be estimated from the relative line strengths of \[\]. Comparing these line strengths with the CHIANTI database [@Landi13], we find that the density is less than $10^7$ cm$^{-3}$. An estimate of the temperature can be derived using the \[\] line diagnostic $R =
\lambda$(4959+5007)/$\lambda$4363 for densities between $10^{5} -
10^{6}\;\rm cm^{-3}$ using the `TEMDEN` task in IRAF [@Shaw94]. We measure $R = 9.8 \pm 0.2$, which is associated with temperatures between $(2-8) \times 10^4$K.
The ratio of \[\] $\lambda$6583 / H$\alpha \approx 0.3$ in the day 373 spectrum is higher than the ratio observed in the day $-4$ spectrum (i.e., before interaction commenced) where it is $\approx
0.15$. An increase in temperature, which would be anticipated from the hard ionizing spectrum and high density that suppresses some of the forbidden line cooling, may explain the high ratio. The increase could also be indicative of nitrogen-enriched CSM from CNO processing in the progenitor star.
Knowledge of the X-ray luminosity $L_x$ around day 373 can further constrain properties of the unshocked CSM. The ionization parameter is defined as $$\xi = L_x/(n r^2),$$ where $n$ is the electron density number, and $r$ is the radius of the emitting region. $L_x$ at this time is $\sim 5 \times 10^{40}$ erg s$^{-1}$ and consistent with a temperature of 18 keV (Margutti et al.2015). @Kallman82 find that the ion fraction peaks at about 30% for $\xi \sim 10$ in models with photoionizing 10 keV bremsstrahlung spectra that are expected to be at least approximately applicable to SN2014C. Thus, $$n r^2 = L_x/\xi \sim 5 \times 10^{39}\;\rm cm^{-1}
\label{eq:nr2}$$ Although there must be a range of ionization parameters present, the ratio of \[\] $\lambda$5303/\[\] $\lambda$6374 is only $0.18 \pm 0.03$, which indicates that $\xi$ only goes up to $\approx 25$.
The emission measure $EM$ is defined as $$\begin{split}
EM & = (4 \pi/3) n^2 r^3 f \\
& = \frac{L_{\rm [Fe~VII]} f}{ j_{\epsilon} A_{(Fe)} \eta}
\label{eq:EM1}
\end{split}$$ where $j_{\epsilon}$ is the emissivity, $A_{(Fe)}$ is the Fe abundance ($2.75\times 10^{-5}$), $\eta$ is the ion fraction, and $f$ is the filling factor. We measure the line flux of \[\] $\lambda$6087 to be $5.2 \times 10^{-15}$ erg cm$^{-2}$ s$^{-1}$, which translates to a luminosity of $1.4 \times 10^{38}$ erg s$^{-1}$. The emissivity for the \[\] $\lambda$6087 line from CHIANTI is $1.3 \times 10^{-20}$ ergs$^{-1}$cm$^{-3}$sr$^{-1}$.
{width="0.8\linewidth"}
Solving equation (\[eq:EM1\]) for $r$ in terms of equation (\[eq:nr2\]), $$r = \frac{(n r^2)^2}{EM(4 \pi/3)f},$$ yields a radius of $ 4.9\times 10^{15}\;f^{-1}\;\rm cm$. Recasting equation (\[eq:nr2\]) in terms of $n$, and substituting this value of $r$ we find that $$n = L_x/(\xi r^2) \sim 2 \times 10^8\;f^2\;\rm cm^{-3}.
\label{eq:n}$$ With the constraint that $n \la 10^{6}\;\rm cm^{-3}$ from the forbidden oxygen and iron line diagnostics, equation (\[eq:n\]) implies that $f \la 0.1$, which is reasonable since the gas that contributes to the highest-ionization Fe lines may fill only some of the volume. The temperatures, densities, and filling factor we derive suggest that the emission is not uniformly distributed and originates from clumped material. A full treatment of emission line modeling is left for future work.
Discussion {#sec:Discussion}
==========
SN2014C represents the first time an ordinary type Ib supernova has been seen to slowly evolve into a strongly interacting type IIn. This Ib-IIn metamorphosis is consistent with a delayed interaction between an H-poor star’s supernova explosion and a local H-rich shell presumably formed from material stripped from the progenitor star. The strong supernova-CSM interaction must have started between February and May of 2014 while the supernova was behind the Sun. Type Ib supernovae have typical rise times of $\la 20$ days [@Drout11; @Bianco14], thus the strong interaction occurred somewhere between $\approx$ 20 and 130 days after explosion. Assuming a minimum forward shock velocity of $0.1\,c$, we can loosely constrain the radius of the shell to be $\ga 3 \times 10^{16}$ cm from the explosion center.
In Figure \[fig:IInIbn\], SN2014C is compared to two representative objects: the type IIn SN2005ip [@Stritzinger12] and the type Ibn SN2006jc [@Modjaz14]. Spectroscopic differences and similarities between these supernovae reflect the distance of the CSM shell with respect to the explosion and the shell’s chemical composition. In type IIn and Ibn supernovae, massive and dense CSM is located nearby and interaction commences shortly ($\la 1$ day) after explosion. Consequently, an opaque CSM hides the interior and O-rich material is often never observed. This contrasts with SN2014C, where the longer delay between explosion and interaction is associated with a larger, less dense and less opaque shell. This latter situation is advantageous as it permits observations of interior processes that are normally hidden.
Previous examples of delayed interaction
----------------------------------------
The closest analog to SN2014C in overall properties is SN2001em. That supernova was spectroscopically classified as a hydrogen deficient type Ib/c [@Filippenko01][^7], two years before it was re-detected as a highly luminous radio source [@Stockdale04], and showed intermediate-width H$\alpha$ emission with FWHM $\sim 1800$ in its optical spectra [@Soderberg04]. Unlike SN2014C, the intermediate stages between H-deficient to H-dominated emission in SN2001em was completely missed.
Although it was originally thought that SN2001em may harbor a bipolar relativistic off-axis jet that had decelerated to mildly relativistic velocities [@Granot04], later observations with very long baseline interferometry ruled this model out by demonstrating that the expansion velocity of the supernova was below 6000 and that the supernova was therefore not driven by a relativistic jet [@Bietenholz05; @Bietenholz07; @Schinzel09]. @Chugai06 found that the multi-wavelength properties of SN 2001em could be modeled in a scenario in which non-relativistic ejecta from the supernova explosion collided with a dense and massive ($\sim 3\;M_{\odot}$) circumstellar shell at a distance of $\sim 7 \times 10^{16}$ cm from the star. The circumstellar shell was presumably formed by a vigorous mass loss episode with a mass loss rate of $\sim 2 \times 10^{-3}$ M$_{\odot}$ yr$^{-1}$ approximately $1000$ to $2000$ yr prior to the supernova explosion. The hydrogen envelope was completely lost and subsequently swept up by the fast wind of the pre-supernova star and accelerated to a velocity of 30-50 . Such a high-rate mass-loss event could be explained by binary interaction, but could also be explained by a powerful eruption from an LBV star. A similar scenario is possible for SN2014C, but in this case, less of the progenitor’s H-rich envelope had been stripped at the time of explosion (see additional discussion in section \[sec:shell\]).
Late-time interaction with dense CSM shells has been observed in a handful of supernovae with H-rich ejecta. For example, the progenitor of SN1996cr evacuated a large cavity just prior to exploding, and the forward blast wave likely spent 1-2 yr in relatively uninhibited expansion before eventually striking a dense shell of CSM [@Bauer08]. SN1996cr may be a “wild cousin” of SN1987A, which has slowly interacted with a dense ring at a radius of $6 \times
10^{17}$ cm (approximately an order of magnitude larger than the shell radius inferred for SN2014C) beginning as early as 1995 [@Sonneborn98; @Lawrence00].
Partially stripped progenitor star {#sec:hydrogen}
----------------------------------
Our modeling of the near-maximum light optical spectrum of SN2014C associates absorption around 6150 Å with high velocity hydrogen (Figure \[fig:spectra\]). Attempts to model the 6150 Å absorption with were unsuccessful because the velocity as inferred from the minimum of the feature ($v \approx 8000$ ) was inconsistent with those inferred from all other ions ($v \approx
13000$ , cf. Section \[sec:morph\]). Alternative identifications for this absorption feature that have been investigated elsewhere for other SN Ibc, including and [@Harkness87; @Deng00], were ruled out as strong contributors in SN2014C. Our model is consistent with recent work by @Parrent15, who find that is unlikely to be the dominant contributor to the absorption feature near 6150 Å seen in many SNIbc.
The detection of hydrogen absorption around the time of maximum light implies that the progenitor star of SN2014C was only partially stripped of hydrogen at the time of explosion. This is an important observation because it informs about the evolutionary status of the star at the time of core collapse. It is also relevant for debates on the extent of H and He in SNIbc [@Matheson01; @Branch02; @Branch06; @Hachinger12; @Milisavljevic13; @Modjaz14; @Milisavljevic15; @Parrent15]. The range of projected Doppler velocities used in fitting the optical spectrum with ($1.3-2.1 \times 10^4$ ) is consistent with a detached envelope consisting of an extended layer of hydrogen beyond the otherwise fairly sharp photosphere. It is beyond the scope of this paper to accurately estimate the mass of hydrogen associated with the detached envelope, but the models of @Hachinger12 provide useful upper limits. Hachinger et al. conclude that $\approx 0.03\;
M_{\odot}$ of hydrogen is sufficient for H$\alpha$ absorption to dominate over absorption due to $\lambda$6355, as it does in SN2014C.
Mass loss in massive stars is correlated with metallicity [@Vink01; @Vink05]. The oxygen abundance of the explosion site of SN2014C measured from our day $-4$ spectrum using the $N2$ scale of @Pettini04 is $\log({\rm O/H})+12 = 8.6 \pm 0.1$, which is near the solar value ($\log({\rm O/H})_{\odot} + 12 = 8.69$; @Asplund05). Our measurement is consistent with an independent measurement of the oxygen abundance of the center of NGC7331 ($\log({\rm O/H}) + 12 = 8.75 \pm 0.18$; @Gusev12). The explosion site metallicity of SN2014C is above the median metallicity of type Ib supernovae with secure classifications, which is $\log({\rm O/H})_{\odot} + 12 = 8.43 \pm 0.14$ [@Sanders12], and in line with the mean metallicity of type IIn, which is $\log({\rm
O/H})_{\odot} + 12 = 8.63 \pm 0.03$ [@Taddia15] (both measured with the same $N2$ scale). However, the explosion site metallicity is consistent with the metallicity distribution observed for each class, which are largely overlapping. SN 2014C was discovered in a targeted survey, which is known to favor the discovery of supernovae in brighter and more metal-rich galaxies [@Sanders12].
Pre-explosion detection of progenitor system
--------------------------------------------
Valuable information about the progenitor system comes from the Subaru and HST pre-explosion imaging (Figures \[fig:images\] and \[fig:enlarge\]). Correcting for extinction using $E(B-V)_{\rm
total}$, the absolute magnitude of the coincident source is $M_V =
-12.0 \pm 0.17$ mag. The luminosity rivals that of the most luminous stars known, suggesting that the source is not a single star and is instead a massive star cluster. Supporting this conclusion are the facts that 1) the source is extended and composite in the HST F658N image with an effective radius of $\approx 17$ pc (cf. Section \[sec:pre-explosion\]) that is consistent with observed sizes of massive star clusters (see, e.g., @Bastian13), and 2) SN2014C has a noticeable offset from the center of the source (Fig. \[fig:images\]).
We compared the photometry and estimate of H$\alpha$ luminosity of the pre-explosion source to the Binary Population and Spectral Synthesis (BPASS) stellar population models (@Eldridge09; Eldridge et al., in preparation; http://bpass.auckland.ac.nz). Assuming the total extinction $E(B-V)_{\rm total}$, we find that the best fitting age of the stellar population is between 30 to 300 Myr. The detection of H$\alpha$ emission from the pre-explosion source favors younger populations closer to 30 Myr, and binary star models match estimates derived from the H$\alpha$ luminosity more closely than those derived from single star models. The turn off mass of stellar populations in the favored age range is between 3.5 to $9.5\;M_{\odot}$, and the mass of the cluster is estimated to be between $3 \times 10^5$ and $10^6
M_{\odot}$. These age constraints are the same for three metallicities close to that of the environment of $Z=0.008$, $0.014$ and $0.020$, where solar metallicity is between $0.014$ and $0.020$.
A source of uncertainty in our analysis of the host massive star cluster is in our estimate of the foreground extinction. It is unknown whether the extinction estimated from the day $-4$ optical spectrum of SN2014C is local to the supernova, or along the line of sight to the entire cluster. An additional complication is that if the extinction is local, then it is possible that the extinction may have changed with time. Thus, within the stated uncertainties, the uncorrected absolute magnitudes represent lower limits to the luminosity of the cluster.
Origin of the CSM shell {#sec:shell}
-----------------------
Massive stars are known to become stripped of their outer layers in a variety of ways including line-driven winds and binary interactions [@Podsiadlowski92; @Woosley95; @Wellstein01; @Vink01; @Puls08; @Yoon10], but the details behind the precise physical mechanisms that are involved and the stages that this mass loss takes place are not well constrained (see @Eldridge08 [@Langer12; @SmithARAA14] and references therein). Any plausible explanation of the origin of the dense CSM shell with which SN2014C interacted must be able to account for two key properties of the system. One property is the distance of the shell from the progenitor star of SN2014C, which is much more distant than the shells in the majority of type IIn and Ibn supernovae. Another property is that the progenitor star exploded as a type Ib supernova, meaning that it had been largely stripped of its hydrogen envelope at the time of core collapse.
Below, we discuss three plausible scenarios for the physical origin of the massive shell that surrounded SN2014C at the time of explosion.
### An unusually short W-R phase
W-R stars have significant stellar winds, with mass loss rates of a few times $10^{-5} M_{\odot}$ yr$^{-1}$ and terminal wind velocities of $1000-3000$ [@Crowther07]. These winds can run into and overtake the much slower winds ($< 100$ ) of a previous red supergiant (RSG) phase and create massive CSM shells, or “bubbles,” (see, e.g., @Garcia96). Approximately half of the Galactic population of W-R stars are associated with a CSM shell, some of which have anomalous abundances associated with ejected CNO-processed stellar material [@Miller93]. However, the majority of these CSM shells have diameters of many parsecs, which is much larger than the radius of the shell inferred for SN 2014C ($\sim 0.01$ pc). Thus, SN2014C did not interact with an environment like those typically observed around W-R stars.
However, it is possible that the progenitor star of SN2014C evolved through a brief W-R phase lasting much less than $10^4$ yr. Although traditionally it has been believed that massive stars should spend $0.5-1$ Myr in the core-He-burning W-R phase before finally exploding as normal SN Ibc [@Heger03], advances in stellar evolution modeling show many pathways that may potentially lead to short-lived W-R stars. Examples include low mass binary stars [@Eldridge08], and RSGs with strong pulsation-driven superwinds [@Yoon10-RSG]. Analyses of circumstellar environments around supernova remnants also support the notion that progenitors of stripped-envelope supernovae may pass through brief W-R phases [@Schure08; @Hwang09].
However, a short W-R phase associated with a typical mass loss rate of a few $\times 10^{-5}\;M_{\odot}\,\rm yr^{-1}$ in itself cannot explain the massive shell SN2014C interacted with. Though a short W-R phase will sweep up a dense shell close to the star, the shell will not have much mass because the RSG wind was expanding freely at this small radius. A possible solution is if the short W-R phase was accompanied with an enhanced mass loss rate (e.g., $\sim
10^{-3}\;M_{\odot}\;\rm yr^{-1}$), which is a scenario proposed by @Chugai06 for SN2001em.
### Shell ejection in a single eruption
Alternatively, in light of the established connection between type IIn supernovae and LBV stars, it is plausible that the shell surrounding the SN2014C progenitor system was ejected in a single eruptive event rather than a brief period of enhanced W-R stellar wind. Shell ejections have typical velocities of $\ga 100$ , which would mean that the major mass loss event would have taken place in the last $\la
100$ years prior to core collapse.
Providing the pre-explosion source was a massive star cluster, the most probable turn-off mass range estimated from the BPASS models ($3.5 - 9.5\; M_{\odot}$) is too low to be compatible with typical LBV progenitors ($M_{ZAMS} \ga 20 M_{\odot}$). However, because uncertainties in our fitting allow room for ages as low as 10 Myr, an LBV-like progenitor is not ruled out completely. It is worth noting that the type IIn-LBV connection is thus far only robust for stars with H-rich envelopes and to date there is no direct detection of the progenitor star of an interacting H-poor supernova. Indirect arguments that the progenitors of type Ibn supernovae exploded in the transition from LBV to W-R phases [@Foley07; @Pastorello07; @Smith12], potentially due to instabilities initiated in the final stages of He core burning [@Pastorello15], have been advanced.
Ejection of an H-rich common envelope in a binary system is also a plausible scenario. Stellar evolution calculations and population models that incorporate stellar duplicity show that stars less than 20 M$_{\odot}$ in interacting binaries can be stripped of their hydrogen envelopes via mass transfer to a companion and/or the loss of the common envelope, and end their lives as low mass helium stars that explode as SN Ibc [@Yoon10; @Eldridge13]. This type of progenitor system was proposed for the type Ib iPTF13bvn [@Eldridge15], which has an optical spectrum at maximum light very similar to that of SN2014C (see Figure \[fig:syn++\]).
### CSM confinement
The location of SN2014C’s progenitor star within a compact massive star cluster makes it plausible that hot gas from stellar winds and prior supernovae could have provided a large external pressure by which to confine the RSG wind into a shell near the star. Simulations have shown that it is difficult to confine RSG wind to $<0.3$ pc with external thermal or ram pressure [@vanMarle06; @Eldridge06], thus this scenario is not plausible unless the wind had considerable asymmetry originating from stellar rotation and/or duplicity [@Eldridge07].
A viable alternative scenario is through the process of photoionization-confinement. @Mackey14 demonstrate how an external radiation field generated by neighboring stars can form a standing shock in the neutral part of an outflowing wind and create an almost static, photoionization-confined shell that traps up to 35 per cent of all mass lost during the RSG phase close to the star until it explodes. Their model was specific to Betelgeuse and its $\sim
0.1\;M_{\odot}$ circumstellar shell, but the model is applicable to other stars that might have much more massive shells.
Complicating this interpretation is the fact that the progenitor star of SN2014C was H-stripped at the time of explosion and thus no longer in the RSG phase. If the star evolved through a W-R phase with a typical timescale of $\sim 10^5$ yr, the momentum of the winds traveling $\ga 1000$ make it unlikely that a shell $\la
1\;M_{\odot}$ in mass with radius $\sim 3 \times 10^{16}$cm would have survived. The photoionization-confined shell scenario remains plausible if the shell was initially much more massive at the end of the RSG phase (a few $\times 1\;M_{\odot}$) and/or if the W-R phase was short.
Conclusion {#sec:Conclusions}
==========
We have presented spectroscopic observations of SN2014C that follow its evolution from an ordinary type Ib supernova to an interacting type IIn. Our observations are consistent with the supernova having exploded in a cavity before encountering a dense, nearby ($\ga 3
\times 10^{16}$ cm) H-rich shell formed from mass lost from the progenitor star. We considered three possible origins to the shell: 1) a W-R fast wind phase that overtook a slower RSG wind, 2) an eruptive ejection, or 3) various forms of CSM confinement. We find that all explanations require that the progenitor star experienced a brief $\la
1000$yr W-R phase. The brief W-R phase may have been associated with larger-than-normal mass loss rates ($> 10^{-4}\;M_{\odot}\;\rm
yr^{-1}$). Alternatively, the prior RSG wind may have been distributed asymmetrically and/or confined via a photoionization layer. Our observations disfavor a sudden eruption in an LBV-like event, but do not rule out such a scenario. Ejection of an H-rich common envelope in a binary system is also possible.
We also presented archival Subaru and HST pre-explosion images covering the field of SN2014C that show a luminous coincident source. The SED and source size as measured by HST are both consistent with a compact massive star cluster. We estimated the age of the cluster to be $30 - 300$ Myr, and favor models incorporating interacting binary systems with ages closer to $30$ Myr in light of relatively strong H$\alpha$ emission.
Because extensive archival data of the host galaxy exist, it is possible that a historical light curve could be constructed. Such a light curve could be used to look for changes in apparent brightness in the system that may be associated with precursor activity from an unstable massive star (see, e.g., @Pastorello13 and @Ofek14). Notably, SN2013bu was discovered in NGC7331 on 2013 April 21 [@Itagaki13], which was only eight months before the discovery of SN2014C. Thus, multi-epoch observations of SN2013bu that cover the region of SN2014C would be appropriate for this type of archival analysis.
We searched for previous HST observations, but unfortunately SN2014C fell slightly outside the footprint except for the F658N images presented here. Future observations can pinpoint the location within the host cluster and investigate the immediate stellar environment when the supernova has faded. Our results strongly motivate observations at ultraviolet wavelengths from which line diagnostics can be performed to investigate potentially enhanced abundances consistent with CNO processed material (e.g., @Fransson02 [@Fransson14]). It is anticipated that additional emission lines will develop as the CSM cools over time (see, e.g., SN2005ip; @Smith09). Line diagnostics at these wavelengths would reveal important information about the evolutionary transitions a massive star may undergo in its final stages approaching core collapse.
SN 2014C is a well-observed example of a class of core-collapse supernovae that fill a gap between events that interact strongly with nearby environments immediately after explosion (type IIn and Ibn) and events that are never observed to interact at all (the majority of SNIbc). Previous surveys of late-time radio emission from SNIbc suggest that events like SN2014C are infrequent [@Bietenholz14]. However, considering that these follow-up observations are not densely sampled and cover many SNIbc that are $> 50$ Mpc, the frequency of other late interaction events where the mass lost is less extreme and/or concentrated at greater distances is unknown.
The shell of mass loss material surrounding the progenitor star of SN2014C was close enough to be detected via subsequent interaction with the blast wave, and yet fortuitously distant enough to permit a clear view of the underlying supernova. Improved constraints on the frequency with which SNIbc interact with their stripped shells can contribute important information about the final stages of mass loss and stellar evolution. Given that the hydrodynamic instabilities that lead to enhanced pre-supernova mass loss may be related to deviations from spherical symmetry in the progenitor star structure [@SmithArnett14], understanding how the underlying physical mechanisms may permit a range of pre-supernova mass loss rates and time lags may also help with our understanding of the core collapse process itself.
We thank an anonymous referee who provided many helpful comments and suggestions that improved the quality and presentation of this paper. Observations reported here were obtained at the MMT Observatory, a joint facility of the Smithsonian Institution and the University of Arizona, as well as the 6.5 m Magellan Telescopes located at Las Campanas Observatory, Chile. This paper uses data taken with the MODS spectrographs built with funding from NSF grant AST-9987045 and the NSF Telescope System Instrumentation Program (TSIP), with additional funds from the Ohio Board of Regents and the Ohio State University Office of Research. Based in part on data collected at Subaru Telescope and obtained from the SMOKA, which is operated by the Astronomy Data Center, National Astronomical Observatory of Japan. Some of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX09AF08G and by other grants and contracts. This paper made use of the Weizmann interactive supernova data repository (WISEREP) - http://wiserep.weizmann.ac.il. J. M. acknowledges support from the Deutsche Forschungsgemeinschaft priority program 1573, Physics of the Interstellar Medium. D. M. thanks M. Shara and A. Pagnotta for helpful discussions.
natexlab\#1[\#1]{}
, G., [Humphreys]{}, R. M., & [Richmond]{}, M. 1994, , 107, 662
, M., [Grevesse]{}, N., & [Sauval]{}, A. J. 2005, in Astronomical Society of the Pacific Conference Series, Vol. 336, Cosmic Abundances as Records of Stellar Evolution and Nucleosynthesis, ed. [T. G. Barnes III & F. N. Bash]{}, 25–+
, N., [Schweizer]{}, F., [Goudfrooij]{}, P., [Larsen]{}, S. S., & [Kissler-Patig]{}, M. 2013, , 431, 1252
, F. E., [Dwarkadas]{}, V. V., [Brandt]{}, W. N., [et al.]{} 2008, , 688, 1210
, M. C., [Benvenuto]{}, O. G., [Folatelli]{}, G., [et al.]{} 2014, , 148, 68
, F. B., [Modjaz]{}, M., [Hicken]{}, M., [et al.]{} 2014, , 213, 19
, M. F., & [Bartel]{}, N. 2005, , 625, L99
—. 2007, , 665, L47
, M. F., [De Colle]{}, F., [Granot]{}, J., [Bartel]{}, N., & [Soderberg]{}, A. M. 2014, , 440, 821
, D., [Jeffery]{}, D. J., [Young]{}, T. R., & [Baron]{}, E. 2006, , 118, 791
, D., [Benetti]{}, S., [Kasen]{}, D., [et al.]{} 2002, , 566, 1005
, Y., [Kasliwal]{}, M. M., [Arcavi]{}, I., [et al.]{} 2013, , 775, L7
, J. A., [Clayton]{}, G. C., & [Mathis]{}, J. S. 1989, , 345, 245
, R. A., & [Fransson]{}, C. 2006, , 651, 381
, N. N., & [Chevalier]{}, R. A. 2006, , 641, 1051
, N. N., & [Danziger]{}, I. J. 1994, , 268, 173
, J. S. W., [de Mink]{}, S. E., [Pols]{}, O. R., [Eldridge]{}, J. J., & [Baes]{}, M. 2011, , 528, A131
, P. A. 2007, , 45, 177
, A., [Bietenholz]{}, M. F., [Kamble]{}, A., [et al.]{} 2015, ArXiv e-prints, arXiv:1503.00837
, J. S., [Qiu]{}, Y. L., [Hu]{}, J. Y., [Hatano]{}, K., & [Branch]{}, D. 2000, , 540, 452
, L., [Hillier]{}, D. J., [Woosley]{}, S., [et al.]{} 2015, , 453, 2189
, M. R., [Soderberg]{}, A. M., [Gal-Yam]{}, A., [et al.]{} 2011, , 741, 97
, V. V. 2011, , 412, 1639
, J. J. 2007, , 377, L29
, J. J., [Fraser]{}, M., [Maund]{}, J. R., & [Smartt]{}, S. J. 2015, , 446, 2689
, J. J., [Fraser]{}, M., [Smartt]{}, S. J., [Maund]{}, J. R., & [Crockett]{}, R. M. 2013, , 436, 774
, J. J., [Genet]{}, F., [Daigne]{}, F., & [Mochkovitch]{}, R. 2006, , 367, 186
, J. J., [Izzard]{}, R. G., & [Tout]{}, C. A. 2008, , 384, 1109
, J. J., & [Stanway]{}, E. R. 2009, , 400, 1019
, D., [Cheimets]{}, P., [Caldwell]{}, N., & [Geary]{}, J. 1998, , 110, 79
, R. A., & [Hurford]{}, A. P. 1996, , 106, 563
, R. A., [Gerardy]{}, C. L., [Filippenko]{}, A. V., [et al.]{} 1999, , 117, 725
, A. V. 1997, , 35, 309
, A. V., & [Chornock]{}, R. 2001, , 7737, 3
, R. J., [Smith]{}, N., [Ganeshalingam]{}, M., [et al.]{} 2007, , 657, L105
, C., [Chevalier]{}, R. A., [Filippenko]{}, A. V., [et al.]{} 2002, , 572, 350
, C., [Ergon]{}, M., [Challis]{}, P. J., [et al.]{} 2014, , 797, 118
, W. L., [Madore]{}, B. F., [Gibson]{}, B. K., [et al.]{} 2001, , 553, 47
, C., [Sollerman]{}, J., [Taddia]{}, F., [et al.]{} 2014, , 565, A114
, A., & [Leonard]{}, D. C. 2009, , 458, 865
, G., [Langer]{}, N., & [Mac Low]{}, M.-M. 1996, , 316, 133
, C. M., [Cappellaro]{}, E., [Dinerstein]{}, H. L., [et al.]{} 1986, , 306, L77
, J., & [Ramirez-Ruiz]{}, E. 2004, , 609, L9
, J. H., [Georgy]{}, C., & [Ekstr[ö]{}m]{}, S. 2013, , 558, L1
, A. S., [Pilyugin]{}, L. S., [Sakhibov]{}, F., [et al.]{} 2012, , 424, 1930
, S., [Mazzali]{}, P. A., [Taubenberger]{}, S., [et al.]{} 2012, , 422, 70
, R. P., [Wheeler]{}, J. C., [Margon]{}, B., [et al.]{} 1987, , 317, 355
, A., [Fryer]{}, C. L., [Woosley]{}, S. E., [Langer]{}, N., & [Hartmann]{}, D. H. 2003, , 591, 288
, M. L. 1959, IAUC, 1682, 1
, U., & [Laming]{}, J. M. 2009, , 703, 883
, K., [Noguchi]{}, T., [Nakano]{}, S., [et al.]{} 2013, Central Bureau Electronic Telegrams, 3498, 1
, K., [Grebel]{}, E. K., & [Ammon]{}, K. 2006, , 460, 339
, T. R., & [McCray]{}, R. 1982, , 50, 263
, A., [Margutti]{}, R., [Soderberg]{}, A. M., [et al.]{} 2015, ArXiv e-prints, arXiv:1504.07988
, M., [Zheng]{}, W., [Li]{}, W., [et al.]{} 2014, Central Bureau Electronic Telegrams, 3777, 1
, R., & [Vink]{}, J. S. 2006, , 460, L5
, M. I., [Soderberg]{}, A. M., [Chomiuk]{}, L., [et al.]{} 2012, ArXiv e-prints, arXiv:1201.0770
, E., [Young]{}, P. R., [Dere]{}, K. P., [Del Zanna]{}, G., & [Mason]{}, H. E. 2013, , 763, 86
, N. 2012, , 50, 107
, N., [Hamann]{}, W.-R., [Lennon]{}, M., [et al.]{} 1994, , 290, 819
, S. S., [Sugerman]{}, B. E., [Bouchet]{}, P., [et al.]{} 2000, , 537, L123
, J., [Mohamed]{}, S., [Gvaramadze]{}, V. V., [et al.]{} 2014, , 512, 282
, K., [Katsuda]{}, S., [Bamba]{}, A., [Terada]{}, Y., & [Fukazawa]{}, Y. 2014, , 785, 95
, K., [Hattori]{}, T., [Milisavljevic]{}, D., [et al.]{} 2015, , 807, 35
, A., & [Meynet]{}, G. 2000, , 38, 143
, R., [Milisavljevic]{}, D., [Soderberg]{}, A. M., [et al.]{} 2014, , 780, 21
, P., [Neugent]{}, K. F., & [Morrell]{}, N. 2015, , 807, 81
, P., [Neugent]{}, K. F., [Morrell]{}, N., & [Hillier]{}, D. J. 2014, , 788, 83
, T., [Filippenko]{}, A. V., [Li]{}, W., [Leonard]{}, D. C., & [Shields]{}, J. C. 2001, , 121, 1648
, J. C., [Smith]{}, N., [Filippenko]{}, A. V., [et al.]{} 2013, , 430, 1801
, J. R., [Fraser]{}, M., [Ergon]{}, M., [et al.]{} 2011, , 739, L37
, G., [Maeder]{}, A., [Schaller]{}, G., [Schaerer]{}, D., & [Charbonnel]{}, C. 1994, , 103, 97
, D., [Fesen]{}, R. A., [Chevalier]{}, R. A., [et al.]{} 2012, , 751, 25
, D., [Margutti]{}, R., [Soderberg]{}, A. M., [et al.]{} 2013, , 767, 71
, D., [Margutti]{}, R., [Parrent]{}, J. T., [et al.]{} 2015, , 799, 51
, G. J., & [Chu]{}, Y.-H. 1993, , 85, 137
, S., [Komiyama]{}, Y., [Sekiguchi]{}, M., [et al.]{} 2002, , 54, 833
, M., [Li]{}, W., [Butler]{}, N., [et al.]{} 2009, , 702, 226
, M., [Blondin]{}, S., [Kirshner]{}, R. P., [et al.]{} 2014, , 147, 99
, D. G., [Levine]{}, S. E., [Canzian]{}, B., [et al.]{} 2003, , 125, 984
, T. J. 2015, , 803, L26
, E. O., [Sullivan]{}, M., [Cenko]{}, S. B., [et al.]{} 2013, , 494, 65
, E. O., [Sullivan]{}, M., [Shaviv]{}, N. J., [et al.]{} 2014, , 789, 104
, D. E., & [Ferland]{}, G. J. 2006, [Astrophysics of gaseous nebulae and active galactic nuclei]{}, 2nd edn. (Sausalito, CA: University Science Books)
, J. T., [Milisavljevic]{}, D., [Soderberg]{}, A. M., & [Parthasarathy]{}, M. 2015, ArXiv e-prints, arXiv:1505.06645
, A., [Smartt]{}, S. J., [Mattila]{}, S., [et al.]{} 2007, , 447, 829
, A., [Cappellaro]{}, E., [Inserra]{}, C., [et al.]{} 2013, , 767, 1
, A., [Benetti]{}, S., [Brown]{}, P. J., [et al.]{} 2015, , 449, 1921
, M., & [Pagel]{}, B. E. J. 2004, , 348, L59
, P., [Joss]{}, P. C., & [Hsu]{}, J. J. L. 1992, , 391, 246
, R. W., [Atwood]{}, B., [Brewer]{}, D. F., [et al.]{} 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7735, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 0
, J., [Vink]{}, J. S., & [Najarro]{}, F. 2008, , 16, 209
, E., & [Shiode]{}, J. 2012, , 423, L92
, N. E., [Soderberg]{}, A. M., [Levesque]{}, E. M., [et al.]{} 2012, , 758, 132
, F. K., [Taylor]{}, G. B., [Stockdale]{}, C. J., [Granot]{}, J., & [Ramirez-Ruiz]{}, E. 2009, , 691, 1380
, E. F., & [Finkbeiner]{}, D. P. 2011, , 737, 103
, G. D., [Weymann]{}, R. J., & [Foltz]{}, C. B. 1989, , 101, 713
, K. M., [Vink]{}, J., [Garc[í]{}a-Segura]{}, G., & [Achterberg]{}, A. 2008, , 686, 399
, M. M., [Bibby]{}, J. L., [Zurek]{}, D., [et al.]{} 2013, , 146, 162
, R. A., & [Dufour]{}, R. J. 1994, in Astronomical Society of the Pacific Conference Series, Vol. 61, Astronomical Data Analysis Software and Systems III, ed. D. R. [Crabtree]{}, R. J. [Hanisch]{}, & J. [Barnes]{}, 327
, S. J. 2009, , 47, 63
, N. 2014, , 52, 487
, N., & [Arnett]{}, W. D. 2014, , 785, 82
, N., [Li]{}, W., [Filippenko]{}, A. V., & [Chornock]{}, R. 2011, , 412, 1522
, N., [Mauerhan]{}, J. C., & [Prieto]{}, J. L. 2014, , 438, 1191
, N., [Mauerhan]{}, J. C., [Silverman]{}, J. M., [et al.]{} 2012, , 426, 1905
, N., & [Owocki]{}, S. P. 2006, , 645, L45
, N., [Silverman]{}, J. M., [Chornock]{}, R., [et al.]{} 2009, , 695, 1334
, A. M., [Gal-Yam]{}, A., & [Kulkarni]{}, S. R. 2004, GRB Coordinates Network, 2586, 1
, A. M., [Berger]{}, E., [Page]{}, K. L., [et al.]{} 2008, , 453, 469
, G., [Pun]{}, C. S. J., [Kimble]{}, R. A., [et al.]{} 1998, , 492, L139
, S., [Anupama]{}, G. C., & [Sahu]{}, D. K. 2014, , 445, 1932
, C. J., [Van Dyk]{}, S. D., [Sramek]{}, R. A., [et al.]{} 2004, , 8282, 2
, M., [Taddia]{}, F., [Fransson]{}, C., [et al.]{} 2012, , 756, 173
, F., [Sollerman]{}, J., [Fremling]{}, C., [et al.]{} 2015, ArXiv e-prints, arXiv:1505.04719
, M., [Yamanaka]{}, M., [Maeda]{}, K., [et al.]{} 2009, , 700, 1680
, D. A., [Boissier]{}, S., [Bianchi]{}, L., [et al.]{} 2007, , 173, 572
, R. C., [Nugent]{}, P. E., & [Meza]{}, J. C. 2011, , 123, 237
, M., [Benetti]{}, S., & [Cappellaro]{}, E. 2003, in From Twilight to Highlight: The Physics of Supernovae, ed. [W. Hillebrandt & B. Leibundgut]{}, 200
, S. D. 2010, in Astronomical Society of the Pacific Conference Series, Vol. 425, Hot and Cool: Bridging Gaps in Massive Star Evolution, ed. C. [Leitherer]{}, P. D. [Bennett]{}, P. W. [Morris]{}, & J. T. [Van Loon]{}, 73
, S. D., [Li]{}, W., & [Filippenko]{}, A. V. 2003, , 115, 1
, S. D., [Zheng]{}, W., [Fox]{}, O. D., [et al.]{} 2014, , 147, 37
, A. J., [Langer]{}, N., [Achterberg]{}, A., & [Garc[í]{}a-Segura]{}, G. 2006, , 460, 105
, J. S., & [de Koter]{}, A. 2005, , 442, 587
, J. S., [de Koter]{}, A., & [Lamers]{}, H. J. G. L. M. 2001, , 369, 574
, S., & [Langer]{}, N. 1999, , 350, 148
, S., [Langer]{}, N., & [Braun]{}, H. 2001, , 369, 939
, S. E., [Blinnikov]{}, S., & [Heger]{}, A. 2007, , 450, 390
, S. E., [Langer]{}, N., & [Weaver]{}, T. A. 1995, , 448, 315
, O., & [Gal-Yam]{}, A. 2012, , 124, 668
, S.-C., & [Cantiello]{}, M. 2010, , 717, L62
, S.-C., [Woosley]{}, S. E., & [Langer]{}, N. 2010, , 725, 940
[^1]: http://www.cfa.harvard.edu/mmti/wfs.html
[^2]: http://smoka.nao.ac.jp/
[^3]: The Image Reduction and Analysis Facility is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^4]: http://tdc-www.harvard.edu/wcstools/
[^5]: http://drizzlepac.stsci.edu/
[^6]: PYRAF is a product of the Space Telescope Science Institute, which is operated by AURA for NASA.
[^7]: The text of @Filippenko01 reads “type Ib or Ic (most likely Ic), perhaps a month after maximum brightness.” We visually inspected these data published in @vanDyk10 and agree with their classification.
|
{
"pile_set_name": "ArXiv"
}
|
[Available at: `http://publications.ictp.it`]{}IC/2008/...\
United Nations Educational, Scientific and Cultural Organization\
and\
International Atomic Energy Agency\
THE ABDUS SALAM INTERNATIONAL CENTRE FOR THEORETICAL PHYSICS\
THE VACUUM STRUCTURE, SPECIAL RELATIVITY THEORY
AND QUANTUM MECHANICS REVISITED: A FIELD THEORY-NO-GEOMETRY APPROACH
\
Nikolai N. Bogolubov (Jr.)[^1]\
*The V.A. Steklov Mathematical Institute of RAN, Moscow, Russian Federation\
and\
The Abdus Salam International Centre for Theoretical Physics, Trieste, Italy,*\
Anatoliy K. Prykarpatsky[^2]\
*The AGH University of Science and Technology, Kraków 30-059, Poland,\
and\
The Ivan Franko State Pedagogical University, Drogobych, Lviv region, Ukraine\
* and\
Ufuk Taneri[^3]\
*The Department of Applied Mathematics and Computer Science, Eastern Mediterranean University EMU, Famagusta, North Cyprus\
and\
Kyrenia American University GAU, Institute of Graduate Studies, Kyrenia, North Cyprus.*\
MIRAMARE – TRIESTE\
September 2008\
**Abstract**
The main fundamental principles characterizing the vacuum field structure are formulated and the modeling of the related vacuum medium and charged point particle dynamics by means of devised field theoretic tools are analyzed. The Maxwell electrodynamic theory is revisited and newly derived from the suggested vacuum field structure principles and the classical special relativity theory relationship between the energy and the corresponding point particle mass is revisited and newly obtained. The Lorentz force expression with respect to arbitrary non-inertial reference frames is revisited and discussed in detail, and some new interpretations of relations between the special relativity theory and quantum mechanics are presented. The famous quantum-mechanical Schrödinger type equations for a relativistic point particle in the external potential and magnetic fields within the quasiclassical approximation as the Planck constant $\hbar
\rightarrow 0$ and the light velocity $c\rightarrow \infty $ are obtained.
Introduction
============
It is a generally accepted statement that no one physical theory can present the absolutely true picture of Nature, assuming that there always exist boundaries of its application, which are approved by experiments and new experience data. This statement concerns, evidently, the relativity theory which, as it was pondered by A. Einstein, should logically arise from both the light velocity constance principle with respect to the inertial reference frames and the generalized equivalence principle with respect to gravity and inertial masses. This theory, namely the special relativity theory, proved to be very thoroughly confirmed by many nuclear physics experiments, which simultaneously became the nuclear energetics backgrounds, widely used today worldwide.
Nevertheless, the nature of space-time and surrounding matter objects was and persists to be one of the most intriguing and challenging problems facing mankind and, in particular, natural scientists. As we know, one of the most brilliant inventions in physics of 19-th century was the combining of electricity and magnetism within the Faraday-Maxwell electromagnetism theory. This theory explained the main physical laws of light propagation in space-time and posed new questions concerning the nature of vacuum. Nonetheless, almost all the attempts aiming to unveil the real state of art of the vacuum problem appeared to be unsuccessful in spite of new ideas suggested by Mach, Lorentz, Poincaré, Einstein and some others physicists. Moreover, the non-usual way of treating the space-time devised by Einstein, in reality, favored eclipsing both its nature and the related physical vacuum origin problems [Fe,Ge,Ma,TW,Ba,PB,Ca]{}, reducing them to some physically unmotivated formal mathematical principles and recipes, combined in the well-known special relativity theory (SRT). The SRT appeared to be adapted only to the inertial reference systems and met with hard problems of the electromagnetic Lorentz forces explanation and relationships between inertial and gravity forces. The latter was artificially “dissolved” by means of the well-known “equivalence principle” owing to which the “inertial” mass of a material object was postulated to coincide with its “gravity” mass. In contrast, E. Mach suggested that any motion of a material point in the space-time, both straight-linear and curvi-linear, can also be only relative. In particular, as the inertion law is also relative, the measure of material point inertion, that is its mass, should also be relative and depend on mutual interactions between all material bodies in the Universe. These E. Mach’s ideas influenced so strongly A. Einstein that he dreamed of them his whole life, trying to reconcile Mach’s relativity principle with his own approach to the general relativity theory. Nonetheless, despite the titanic efforts of A. Einstein, he failed to include the Mach relativity principle into his general relativity theory, since they appeared to be mutually excluding each other.
Simultaneously, the vacuum origin as a problem almost completely disappeared from the Einstein theory being replaced by the geometrization of space-time nature and all related physical phenomena. Meanwhile, the impressive success of 20-th century quantum physics, especially quantum electrodynamics, have demonstrated clearly enough [Fe1,So,BS,BjD,AB,Sch,B-B,Ba]{} that the vacuum polarization and electron-positron annihilation phenomena make it possible to pose new questions about space-time and vacuum structures, and further to revisit [@Br; @Ca; @Ma; @Fa; @BD; @Lo; @Fo1; @Re; @PB] the existing points of view on them.
As well known, the classical mechanics uses the notion of “potential” energy, being a scalar function of spatial variables, very important for formulating dynamical equations, in spite of the fact that it is determined up to arbitrary constant. One can also observe a similar situation in the classical electrodynamic theory, which effectively uses the notions of scalar and potentials related to each other via the well-known Lorentz compatibility gauge constraint [@BjD; @AB; @BS; @Pa] and defined up to suitable gauge transformations. These, in some sense, “different” potential functions were later deeply reanalyzed within the classical Einstein relativity theory by many physicists [So,BjD,Fa,Br,Lo,Pa,Re,Fo1,BD,Ma,We,LK,L-L,Me1,Se]{} that gave rise to the understanding of their fundamental role in combining two great theories - the electromagnetism and the gravity.
Moreover, new important problems arose owing to the famous Einstein relationship between the internal energy and the velocity dependent mass, belonging to a material particle. But, as it was mentioned by L. Brillouin [@Br], the relationship between the particle mass and its internal energy takes into account only the kinetic energy, describing no mass pertaining to the potential energy, which makes the classical relativity theory a not completely closed and, physically, not compatible theoretical construction. And as written by P.W. Bridgeman in [@Bri], “... at construction of his general relativity theory A. Einstein did not make use of those lessons, which he had us taught himself, and of his deep penetration, which he demonstrated us in his special relativity theory”.
Below we try to unveil some nontrivial aspects of the real space-time and vacuum origin problems, deeply related with the relativity theory and electrodynamics, to derive, from the natural field theory principles, all of the well-known Maxwell electromagnetism and special relativity theories results, to show their relative or only visible coincidence with real physical phenomena and to feature new perspectives facing the modern fundamental physics.
Moreover, having further developed the field approach to the microscopic vacuum structure, previously suggested in [@Re] and accounted for in [@PB], we obtained, within the quasi-classical approximation, a new derivation of main quantum mechanical relationships describing evolution of microscopic particle systems, coinciding as $\hbar \rightarrow 0$ with those devised at the beginning of the 20-th century by the great physicists Schrödinger, Heisenberg and Dirac.
The Maxwell electromagnetism theory: new look and interpretation
================================================================
We start from the following field theoretical model [@PB] of the microscopic vacuum structure, considered as some physical reality imbedded into the standard three-dimensional Euclidean space reference system marked with three spatial coordinates $r\in \mathbb{R}^{3},$ endowed with the standard scalar product $<\cdot ,\cdot >,$ and parameterized by means of the scalar temporal parameter $t\in \mathbb{R}.$ We will describe the physical vacuum matter endowing it with an everywhere enough smooth four-vector potential function $(W,A):\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}\times \mathbb{R}^{3},$ naturally related to light propagation properties. The material objects, imbedded into the vacuum, we will model (classically here) by means of the scalar charge density function $\ \rho :\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ and the vector current density $\ J:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3},$ being also everywhere enough smooth functions.
1. The *first* field theory principle regarding the vacuum we accept is formulated as follows: the four-vector function $(W,A):\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}\times \mathbb{R}^{3}$ satisfies the standard Lorentz type continuity relationship$$\frac{1}{c}\frac{\partial W}{\partial t}+<\nabla ,A>=0, \label{M1.1}$$where, by definition, $\nabla :=\partial /\partial r$ is the usual gradient operator.
2. The *second* field theory principle we accept is a dynamical relationship on the scalar potential component $W:\mathbb{R}^{3}\times
\mathbb{R}\rightarrow \mathbb{R}:$$$\frac{1}{c^{2}}\frac{\partial ^{2}W}{\partial t^{2}}-\nabla ^{2}W=\rho ,
\label{M1.2}$$assuming the linear law of the small vacuum uniform and isotropic perturbation propagations in the space-time, understood here, evidently, as a first (linear) approximation in the case of weak enough fields.
3. The *third* principle is similar to the first one and means simply the continuity condition for the density and current density functions:$$\frac{\partial \rho }{\partial t}+<\nabla ,J>=0. \label{M1.3}$$
We need to note here that the vacuum field perturbations velocity parameter $c>0,$ used above, coincides with the vacuum light velocity, as we are trying to derive successfully from these first principles the well-known Maxwell electromagnetism field equations, to analyze the related Lorentz forces and special relativity relationships. To do this, we first combine equations (\[M1.1\]) and (\[M1.2\]):$$\frac{1}{c^{2}}\frac{\partial ^{2}W}{\partial t^{2}}=-<\nabla ,\frac{1}{c}\frac{\partial A}{\partial t}>=<\nabla ,\nabla W>+\rho ,$$whence $$<\nabla ,-\frac{1}{c}\frac{\partial A}{\partial t}-\nabla W>=\rho .
\label{M1.4}$$Having put, by definition,$$E:=-\frac{1}{c}\frac{\partial A}{\partial t}-\nabla W, \label{M1.5}$$we obtain the first material Maxwell equation$$<\nabla ,E>=\rho \label{M1.6}$$for the electric field $E:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3}.$ Having now applied the rotor-operation $\ \ \nabla \times $ to expression (\[M1.5\]) we obtain the first Maxwell field equation$$\frac{1}{c}\frac{\partial B}{\partial t}-\nabla \times E=0 \label{M1.7}$$on the magnetic field vector function $B:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3},$ defined as $$B:=\nabla \times A. \label{M1.8}$$
To derive the second Maxwell field equation we will make use of (\[M1.8\]), (\[M1.1\]) and (\[M1.5\]):$$\begin{aligned}
\nabla \times B &=&\nabla \times (\nabla \times A)=\nabla <\nabla ,A>-\nabla
^{2}A= \notag \\
&=&\nabla (-\frac{1}{c}\frac{\partial W}{\partial t})-\nabla ^{2}A=\frac{1}{c}\frac{\partial }{\partial t}(-\nabla W-\frac{1}{c}\frac{\partial A}{\partial t}+\frac{1}{c}\frac{\partial A}{\partial t})-\nabla ^{2}A= \notag
\\
&=&\frac{1}{c}\frac{\partial E}{\partial t}+(\frac{1}{c^{2}}\frac{\partial
^{2}A}{\partial t^{2}}-\nabla ^{2}A). \label{M1.9}\end{aligned}$$We have from (\[M1.5\]), (\[M1.6\]) and (\[M1.3\]) that $$<\nabla ,\frac{1}{c}\frac{\partial E}{\partial t}>=\frac{1}{c}\frac{\partial
\rho }{\partial t}=-\frac{1}{c}<\nabla ,J>,$$or $$<\nabla ,-\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}-\nabla (\frac{1}{c}\frac{\partial W}{\partial t})+\frac{1}{c}J>=0. \label{M1.10}$$Now making use of (\[M1.1\]), from (\[M1.10\]) we obtain that$$\begin{aligned}
&<&\nabla ,-\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}-\nabla (\frac{1}{c}\frac{\partial W}{\partial t})+\frac{1}{c}J>=<\nabla ,-\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}+\nabla <\nabla ,A>+\frac{1}{c}J=
\notag \\
&=&<\nabla ,-\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}+\nabla
^{2}A+\nabla \times (\nabla \times A)+\frac{1}{c}J>= \notag \\
&=&<\nabla ,-\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}+\nabla
^{2}A+\frac{1}{c}J>=0. \label{M1.11}\end{aligned}$$Thereby, equation (\[M1.11\]) yields$$\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}-\nabla ^{2}A=\frac{1}{c}(J+\nabla \times S) \label{M1.12}$$for some smooth vector function $S:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3}.$ Here we need to note that continuity equation (\[M1.3\]) is defined, concerning the current density vector $J:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3},$ up to a vorticity expression, that is $J\simeq J+\nabla \times S$ and equation (\[M1.12\]) can finally be rewritten down as $$\frac{1}{c^{2}}\frac{\partial ^{2}A}{\partial t^{2}}-\nabla ^{2}A=\frac{1}{c}J. \label{M1.13}$$Having substituted (\[M1.13\]) into (\[M1.9\]) we obtain the second Maxwell field equation
$$\nabla \times B-\frac{1}{c}\frac{\partial E}{\partial t}=\frac{1}{c}J.
\label{M1.14}$$
In addition, from (\[M1.8\]) one also finds the magnetic no-charge relationship$$<\nabla ,B>=0. \label{M1.15}$$
Thus, we have derived all the Maxwell electromagnetic field equations from our three main principles (\[M1.1\]), (\[M1.2\]) and (\[M1.3\]). The success of our undertaking will be more impressive if we adapt our results to those following from the well known relativity theory in the case of point charges or masses. Below we will try to demonstrate the corresponding derivations based on some completely new physical conceptions of the vacuum medium first discussed in [@Re; @PB].
It is interesting to analyze a partial case of the first field theory vacuum principle (\[M1.1\]) when the following local conservation law for the scalar potential field function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ holds:$$\frac{d}{dt}\int_{\Omega _{t}}Wd^{3}r=0, \label{M1.16}$$where $\Omega _{t}\subset \mathbb{R}^{3}$ is any open domain in space $\mathbb{R}^{3}$ with the smooth boundary $\partial \Omega _{t}$ for all $t\in \mathbb{R}$ and $d^{3}r$ is the standard volume measure in $\mathbb{R}^{3}$ in a vicinity of the point $r\in \Omega _{t}.$
Having calculated expression (\[M1.16\]) we obtain the following equivalent continuity equation $$\frac{1}{c}\frac{\partial W}{\partial t}+<\nabla ,\frac{v}{c}W>=0,
\label{M1.17}$$where $\nabla :=\nabla _{r}$ is, as above, the gradient operator and $v:=dr/dt$ is the velocity vector of a vacuum medium perturbation at point $r\in \mathbb{R}^{3}$ carrying the field potential quantity $W.$ Comparing now equations (\[M1.1\]), (\[M1.17\]) and using equation (\[M1.3\]) we can make the suitable identifications: $$A=\frac{v}{c}W,\text{ \ \ \ \ }J=\rho v, \label{M1.18}$$well known from the classical electrodynamics and superconductivity theory [@Fe1; @Ge]. Thus, we are faced with a new physical interpretation of the conservative electromagnetic field theory when the vector potential $A:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3}$ is completely determined via expression (\[M1.18\]) by the scalar field potential function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}.$ It is also evident that all the Maxwell electromagnetism filed equations derived above hold as well in the case (\[M1.18\]), as it was first demonstrated in [@Re] (but with some mathematical inaccuracies) and in [@PB].
Consider now the conservation equation (\[M1.16\]) jointly with the related integral “vacuum momentum” conservation relationship $$\frac{d}{dt}\int_{\Omega _{t}}(\frac{vW}{c^{2}})d^{3}r=0,~~~\Omega
_{t}|_{t=0}=\Omega _{0}, \label{M1.19}$$where, as above, $\Omega _{t}\subset \mathbb{R}^{3}$ is for any time $t\in
\mathbb{R}$ an open domain with the smooth boundary $\partial {\Omega _{t}},$ whose evolution is governed by the equation $$dr/dt=v(r,t) \label{M1.20}$$for all $x\in \Omega _{t}$ and $t\in \mathbb{R},$ as well as by the initial state of the boundary $\partial {\Omega }_{0}.$ As a result of relation (\[M1.19\]) one obtains the new continuity equation $$\frac{d(vW)}{dt}+vW<\nabla ,v>=0. \label{M1.21}$$Now making use of (\[M1.17\]) in the equivalent form $$\frac{dW}{dt}+W<\nabla ,v>=0,$$we finally obtain a very interesting local conservation relationship $$dv/dt=0\ \label{M1.22}$$on the vacuum matter perturbations velocity $v=dr/dt,$ which holds for all values of the time parameter $t\in \mathbb{R}.$ As it is easy to observe, the obtained relationship completely coincides with the well-known hydrodynamic equation [@MC] of ideal compressible liquid without any external exertion, that is, any external forces and field “pressure” are equally identical to zero. We received a natural enough result where the propagation velocity of the vacuum field matter is constant and equals exactly $v=c,$ that is the light velocity in the vacuum, if to recall the starting wave equation (\[M1.2\]) owing to which the small vacuum field matter perturbations propagate in the space with the light velocity.
Special relativity theory and dynamical field equations
=======================================================
From classical electrodynamics we know that the main dynamical relationship relates the particle mass acceleration to the Lorentz force which strongly depends on the absolute charge velocity. For the electrodynamics to be independent on the reference system physicists were forced to reject the Galilean transformations and replace them with the artificially postulated Lorentz transformations. This resulted later in the Einstein relativity theory which has partly reconciled the problems concerned with deriving true dynamical equations for a charged point particle.
We will now start from the scalar field vacuum medium function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ in the conservation condition case (\[M1.16\]) discussed above. This means, obviously, that the vacuum medium field vector potential $A:\mathbb{R}^{3}\times \mathbb{R}\rightarrow
\mathbb{R}^{3},$ charge and current densities $(\rho ,J):\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ $\times \mathbb{R}^{3}$ are related owing to expressions (\[M1.18\]).
Consider now vacuum field medium conservation equations (\[M1.17\]) and (\[M1.2\]) at the density $\rho =0:$$$\begin{aligned}
-\frac{1}{c^{2}}\frac{\partial ^{2}W}{\partial t^{2}} &=&\frac{1}{c^{2}}\frac{\partial }{\partial t}(-\frac{\partial W}{\partial t})=\frac{1}{c^{2}}\frac{\partial }{\partial t}(<\nabla ,vW>)= \notag \\
&=&<\nabla ,\frac{\partial }{\partial t}(\frac{Wv}{c^{2}})>=-<\nabla ,\nabla
W>. \label{M2.1}\end{aligned}$$From relation (\[M2.1\]) it follows that$$\frac{\partial }{\partial t}(\frac{Wv}{c^{2}})+\nabla W=\nabla \times F,
\label{M2.2}$$where $F:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}^{3}$ is some smooth function, which we put, by definition, to be zero owing to the *a priori* assumed vortexless vacuum medium dynamics. So, our dynamical equation on the vacuum medium scalar field function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ looks like $$\frac{\partial }{\partial t}(\frac{Wv}{c^{2}})+\nabla W=0. \label{M2.3}$$
Consider now a charged point particle $q$ in the space point $r=R(t):=R_{0}+\dint\limits_{0}^{t}u(t)dt\in \mathbb{R}^{3},$ depending on time parameter $t\in \mathbb{R}$ and initial point $R_{0}\in \mathbb{R}^{3}$ at** **$t=0.$ Since the vacuum medium field is described by means of the potential field function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow
\mathbb{R},$ which is naturally disturbed by the charged particle $q,$ we will model this fact approximately as the following resulting functional relationship:$$W(r,t)=\tilde{W}(r,R(t)) \label{M2.4}$$for some scalar function $\tilde{W}:\mathbb{R}^{3}\times \mathbb{R}^{3}\rightarrow \mathbb{R}.$ This function must satisfy equation (\[M1.17\]), that is $$<\frac{\partial \tilde{W}}{\partial R},u>+<\nabla ,\tilde{W}v>=0.
\label{M2.5}$$As we are interested in differential properties of the function $\tilde{W}:\mathbb{R}^{3}\times \mathbb{R}^{3}\rightarrow \mathbb{R}$ as $r\rightarrow
R(t)\in \mathbb{R}^{3},$ where the charged point particle is located, we obtain from (\[M2.5\]) that $$<\frac{\partial \tilde{W}}{\partial R}+\frac{\partial \tilde{W}}{\partial r},u>|_{r\rightarrow R(t)}+\tilde{W}<\nabla ,v>|_{r\rightarrow R(t)}=0,$$giving rise to the relationship $$\frac{\partial \tilde{W}}{\partial R}=-\frac{\partial \tilde{W}}{\partial r}
\label{M2.6}$$as $r\rightarrow R(t),$ since $v|_{r\rightarrow R(t)}\rightarrow
dR(t)/dt:=u(t)$ and $<\nabla ,v>|_{r\rightarrow R(t)}\rightarrow <\nabla
,u(t)>=0$ for all $t\in \mathbb{R}.$
Returning now to equation (\[M2.3\]) we can write, owing to (\[M2.6\]), that $$\begin{array}{c}
\frac{1}{c^{2}}\left. \left( \frac{\partial \tilde{W}}{\partial t}v+\tilde{W}\frac{\partial v}{\partial t}\right) \right\vert _{r\text{ }\rightarrow
R(t)}=\frac{1}{c^{2}}\left. \left( -<\frac{\partial \tilde{W}}{\partial r},v>v+\tilde{W}\frac{\partial v}{\partial t}\right) \right\vert
_{r\rightarrow R(t)}= \\
=\frac{1}{c^{2}}\left. \left( <\frac{\partial \tilde{W}}{\partial R},u>u+\tilde{W}\frac{du}{dt}\right) \right\vert _{r\rightarrow R(t)}\Rightarrow
\frac{1}{c^{2}}\frac{d}{dt}(\bar{W}u)=-\left. \frac{\partial \tilde{W}}{\partial r}\right\vert _{r\rightarrow R(t)}=\frac{\partial \bar{W}}{\partial
R},\end{array}
\label{M2.7}$$where we put, by definition, $\bar{W}:=\tilde{W}(r,R(t))|_{r\rightarrow
R_{0}}.$ Thus, we obtained from (\[M2.7\]) that the function $\bar{W}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ satisfies the determining dynamical equation $$\frac{d}{dt}(-\frac{\bar{W}}{c^{2}}u)=-\frac{\partial \bar{W}}{\partial R}
\label{M2.8}$$at point $R(t)\in \mathbb{R}^{3},$ $t\in \mathbb{R},$ of the point charge $q$ location.
Now we need to proceed with our calculations and would like to make the following very important **assumption:** we will interpret the quantity $-\frac{\bar{W}}{c^{2}}$ as the real “dynamical” mass of our point charge $q$ at point $R(t)\in \mathbb{R}^{3},$ that is $$m:=-\frac{\bar{W}}{c^{2}}. \label{M2.9}$$This, in part, means that the whole observed particle mass $m$ depends only on the vacuum field potential $\bar{W}$ owing to both its gravitational and electromagnetic interactions with long distant and closely ambient it material particles! This statement, evidently, in many points coincides with the well-known Mach principle [@Br; @Re; @PB] and formalizes it concerning the real field structure of vacuum. We need here to mention that this idea was also earlier claimed, but not realized practically, by L. Brillouin in [@Br]. We press here that no assumption about the equivalence of the inertial mass and gravitational mass is made and, moreover, such a kind of statement is completely alien within the theory devised here and in [Re,PB]{}.
Using further (\[M2.9\]) we can rewrite equation (\[M2.8\]) as $$\frac{dp}{dt}=-\frac{\partial \bar{W}}{\partial R}, \label{M2.10}$$where the quantity $p:=mu$ has the natural momentum interpretation.
The obtained equation (\[M2.10\]) is very interesting from the dynamical point of view. Really, from equation (\[M2.10\]) we obtain that $$<u,\frac{d}{dt}(mu)>=c^{2}<\frac{\partial m}{\partial R},u>=c^{2}\frac{dm}{dt}. \label{M2.11}$$As a result of (\[M2.11\]) we easily derive, following [@Re; @BP], the conservative relationship$$\frac{d}{dt}\left( m\sqrt{1-\frac{u^{2}}{c^{2}}}\right) =0 \label{M2.12}$$for all $t\in \mathbb{R}.$ Really, based on (\[M2.11\]), we have that$$m<u,\frac{du}{dt}>+<u,u>\frac{dm}{dt}=c^{2}\frac{dm}{dt}, \label{M2.12a}$$ or, equivalently, $$\frac{1}{2}m\frac{du^{2}}{dt}-(c^{2}-u^{2})\frac{dm}{dt}=0. \label{M2.12b}$$As a result of (\[M2.12b\]), we obtain $$-\frac{1}{2}\frac{1}{(1-\frac{u^{2}}{c^{2}})}\frac{d}{dt}(\frac{u^{2}}{c^{2}})+\frac{1}{m}\frac{dm}{dt}=0, \label{M2.12c}$$giving rise, by means of simple integrating, to the following differential expression:$$\frac{d}{dt}\ln (\sqrt{1-\frac{u^{2}}{c^{2}}})+\frac{d\ln m}{dt}=\frac{d}{dt}\ln \left( m\sqrt{1-\frac{u^{2}}{c^{2}}}\right) =0. \label{M2.12d}$$The latter is, evidently, equivalent to result (\[M2.12\]), that is the quantity $$m\sqrt{1-\frac{u^{2}}{c^{2}}}=m_{0} \label{M2.13}$$is constant for all $t\in \mathbb{R},$ giving rise to the well known relativistic expression for the mass of a point particle:$$m=\frac{m_{0}}{\sqrt{1-\frac{u^{2}}{c^{2}}}}. \label{M2.14}$$As we can see, the point particle mass $m$ depends, in reality, not on the coordinate $R(t)\in \mathbb{R}^{3}$ of the point particle $q,$ but on its velocity $u:=dR(t)/dt.$ Since the field potential $\bar{W}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ consists of two parts $$\bar{W}=\bar{W}_{0}+\Delta \bar{w}, \label{M2.15}$$where $\bar{W}_{0}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ is constant and responsible for the external influence of all long distant objects in the Universe upon the point particle $q$ and $\Delta \bar{w}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ is responsible for the local field potential perturbation by the point charge $q$ and its closest ambient neighborhood. Then, obviously, $$\Delta m:=m-m_{0}=-\Delta \bar{w}/c^{2} \label{M2.16}$$is the strictly dynamical mass component belonging to the point particle $q.$ Moreover, since the full momentum $p=mu$ satisfies equation (\[M2.10\]), one can easily obtain that the quantity $$\bar{W}^{2}-c^{2}p^{2}=E_{0}^{2} \label{M2.17}$$is not depending on time $t\in \mathbb{R},$ that is, $dE_{0}/dt=0,$ where $E_{0}:=m_{0}c^{2}$ is the so-called [@Fe1; @Di; @Fo1; @TW] “*internal energy"* of a point particle $q.$ The result (\[M2.17\]) demonstrates the important property of the energy essence: the point particle $q$ is, in reality, endowed with the only dynamical energy $\Delta E:=\Delta mc^{2}.$ Concerning this ”internal“ particle energy $E_{0}=m_{0}c^{2}$ we see that it has nothing to do with the real particle energy, since its origin is determined completely owing to the long distant objects of the Universe and can not be used for any physical processes, contrary to the known Einstein theory statements about a ”huge“ internal energy stored inside the particle mass. Equivalently, the Einstein theory statement about the ”equivalence“ of the mass and the ”internal“ energy of particle appears to be senseless, since the main part of the field potential function $\bar{W}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ at the location point of the point particle $q$ is constant and results owing to the long distant objects in the Universe, which obviously can not be used for so called ”practical applications".
Nonetheless, we have observed above, as a by-product, the well known “relativistic” effect of the particle mass growth depending on the particle velocity in the form (\[M2.14\]). As it was already mentioned in [Re,PB]{} this “mass growth” is, in reality, completely of dynamical nature and is not a consequence of the Lorentz transformations, as it was stated within the Einstein SRT. Moreover, we can state that all of the so-called “relativistic” effects have nothing to do with the mentioned above Lorentz transformations and with such artificial “effects” as length “shortening” and time “slowing”. There is also no reasonable cause to identify the particle mass with its real energy and vice versa. Concerning the interesting physical effect called particles “annihilation” we need to stress here that it has also nothing to do with the transformation of particles masses into energy. The field theoretical explanation of this phenomenon consists in creating their very special bonding state, whose interaction with ambient objects is vanishing. As a result the visible inertial or dynamical mass of this bound state is also zero, exactly what the experiment shows, and nothing else. Inversely, if an intensive enough photon meets such a bound state of two particles, it can break them back into two separate particles, what the experiment shows to happen. Here we recall a similar analogy borrowed from the modern quantum physics of infinite particle systems described by means of the second quantization scenario [@Fo; @Be; @BP; @BPT] suggested in 1932 by V. Fock. Within this scenario there are also realized creating-annihilation effects which are present owing to the inter-particle interaction forces. Moreover, as we know from the modern superconductivity theory within this picture one can describe special bound states of particles, so-called “Cooper pairs”, whose interaction with each other completely vanishes and whose combined mass strongly differs from the sum of the suitable components and equals the so-called “effective” compound mass, depending strongly on the potential field intensity inside the superconductor matter.
We now proceed to discuss the relation (\[M2.17\]) derived above, where the conserved quantity $E_{0}$ is naturally interpreted as the total energy of a particle moving with velocity $u:=$ $dR(t)/dt$ in vacuum endowed with the field potential $\bar{W}$ in a vicinity of the particle $q$ located at point $R(t)\in \mathbb{R}^{3}.$ We see that the total particle energy $E_{0}$ strictly depends on both the field potential $\bar{W}$ and its velocity $u,
$ as the particle momentum $p=m_{0}u/\sqrt{1-\frac{u^{2}}{c^{2}}}$ depends strictly relativistically on its velocity $u.$
As it was mentioned by L. Brillouin in [@Br] the Einstein SRT postulates that the total energy $E_{0}$ of a moving particle in a potential field $U$ equals $E_{0}=m_{0}c^{2}/\sqrt{1-\frac{u^{2}}{c^{2}}}+U.$ He writes: *“This means that any possibility of existing the particle mass related with the external potential energy is completely excluded... Thereby, the strange situation appears: the internal (particle) energy is endowed with the mass but the external - is not”.* (The citation is taken from* *[@Br]).
Contrary to this inference from the Einstein SRT, the relation (\[M2.17\]) obtained above naturally takes into account both the kinetic energy of the particle motion with velocity $u$ and the field potential energy $\bar{W}
$ in a vicinity of the particle $q$ located at point $R(t)\in \mathbb{R}^{3}.
$
Moreover, L. Brillouin in [@Br] writes: “*...Einstein tends by any way to reduce gravity to geometry by means of changing the Newtonian gravitational potential by a tensor potential of second order, realizing the joint description of gravity and geometry; this is achieved owing to the appearance of a huge gap between gravity and electromagnetism.... The (Einstein) article - a genuine mathematical work but its application to the physical reality remains to be open."*** ** Later V. Fock in his famous book [@Fo1] tried to rescue the situation that arisen with gravity and electromagnetism but his approach was also based on the Einstein geometrization ideas and no real success was achieved. Evaluating the gravity and electromagnetism theories state of art L. Brillouin in [@Br] states that ”*In general, the necessity of considering a curved space-time Universe is still not proved; the physical meaning of the general relativity is still very vague".*
Having analyzed, from this point of view, the results formulated above we see that all of the Einstein SRT statements were obtained within the classical Euclidean space-time scenario and no four-dimensional space-time geometry, like the invariance of the four-dimensional metric interval with respect to the Lorentz transformations, was involved. We will show below that within the approaches devised above and in [@Re; @PB] we will derive the next very important result of the SRT and Maxwell electromagnetism theory, namely that related to the nature of the Lorentz force acting on a moving in space charged particle. As is well known [@Br; @Fo1; @BD; @Re], wishing to make the Lorentz force expression compatible with the postulated relativity principles appeared to be decisive in Einstein’s endeavors to construct his SRT and later the general relativity theory.
As a very interesting aspect of the vacuum field theory devised above and in [@PB; @Re] we need here to mention a very close relationship of electromagnetic and gravity fields. Namely, the vacuum field potential, in general, consists of two components: $W:=W_{g}+W_{em},$ where $W_{g}$ corresponds to the gravity interaction between material particles and $W_{em}
$ corresponds to the electromagnetic interaction between their charges, where we have accepted that these two physical vacuum realities are different and independent. Thecorresponding full mass of a particle is given then by expression (\[M2.9\]) in the form$$m:=-\bar{W}/c^{2}=-(\bar{W}_{g}+\bar{W}_{em})/c^{2}, \label{M2.18}$$including both the gravity and electromagnetic interactions between particles. Then, following the derivation of electromagnetic Maxwell equations above, based on the electromagnetic vacuum field potential $W:=W_{em},$ we can derive, in the same way, the corresponding Maxwell type gravimagnetic equations on the gravity potential $W_{g},$ as follows$$\begin{array}{c}
\frac{1}{c^{2}}\frac{\partial ^{2}W_{g}}{\partial t^{2}}-\nabla
^{2}W_{g}=\rho _{g},\frac{1}{c^{2}}\frac{\partial ^{2}A_{g}}{\partial t^{2}}-\nabla ^{2}A_{g}=J_{g}, \\
\text{ } \\
\nabla \times B_{g}-\frac{1}{c}\frac{\partial E_{g}}{\partial t}=\frac{1}{c}J_{g},\frac{1}{c}\frac{\partial B_{g}}{\partial t}-\nabla \times E_{g}=0, \\
\\
<\nabla ,E_{g}>=\rho _{g},\text{ \ }<\nabla ,B_{g}>=0,\end{array}
\label{M2.19}$$where, by definition, $$E_{g}:=-\nabla W_{g}-\frac{1}{c}\frac{\partial A_{g}}{\partial t},\text{ }B_{g}:=\nabla \times A_{g}. \label{M2.20}$$Here $\rho _{g}$ and $J_{g}$ denote the gravity particle mass density and mass current density, respectively. The form of governing equations ([M2.19]{}) is, evidently, strongly motivated by the well known similarity between the Coulomb electrical and Newtonian gravitational forces expressions, and was very deeply previously discussed in [@Br; @Ca; @Re; @Pe].
As already shown above, in the conservative case, the following representations $$J_{g}=\rho _{g}v,\text{ \ }A_{g}=\frac{v}{c}W_{g}, \label{M2.21}$$similar to (\[M1.18\]), hold. As a consequence, we can state that the gravitational waves exist, propagating in vacuum with the same light velocity $c$ as electromagnetic waves, and are described by means of the Maxwell type gravimagnetic equations (\[M2.19\]). Here we note that similar inferences have been done many years ago in classical oeuvres of famous physicists of the XIX century J.C. Maxwell [@Max] and O. Heaviside [@He], as well as in works of other physicists of the past and present centuries [@BL; @Pe; @Bu; @Re].
The Lorentz force and the relativity theory principles revisited
================================================================
It is a well known fact that the Einstein special relativity theory is applicable only for physical processes related to each other by means of the inertial reference systems, moving with constant velocities. In this case one can make use of the Lorentz transformations and calculate the components of suitable four-vectors and the resulting mass growth of particles owing to formula (\[M2.11\]). A nontrivial problem arises when we wish to analyze these quantities with respect to non-inertial reference systems moving with some nonzero acceleration. Below we will revisit this problem from the vacuum field theory scenario devised above and show that the whole “special” relativity theory emerges as its partial case or by-product and is free of the artificial “inertial reference systems” problems mentioned above.
Really, our vacuum field theory structure is described by the dynamical equation (\[M2.3\]), which we would like to investigate in a neighborhood of two interacting to each other point particles $q_{f}$ at point $R_{f}(t)\in \mathbb{R}^{3}$ and $\ q$ at point $R(t):=R_{0}+\dint\limits_{0}^{t}u(t)dt\in \mathbb{R}^{3},$ respectively, depending on time parameter $t\in \mathbb{R}$ and initial point $R_{0}\in \mathbb{R}^{3}$ at** **$t=0.$ As was already done in Section 2 we assume that thevacuum potential field function $W:\mathbb{R}^{3}\times \mathbb{R}\rightarrow \mathbb{R}$ can be represented as $W=\tilde{W}(r;R_{f}(t),R(t))$ for some function $\tilde{W}:\mathbb{R}^{3}\times \mathbb{R}^{3}\times
\mathbb{R}^{3}\rightarrow \mathbb{R}$ and all $t\in \mathbb{R}.$ Then, based on the continuity equation (\[M1.17\]) we obtain$$<\frac{\partial \tilde{W}}{\partial R_{f}},u_{f}>+<\frac{\partial \tilde{W}}{\partial R},u>+<\frac{\partial \tilde{W}}{\partial r},v>+\tilde{W}<\nabla
,v>=0. \label{M3.1}$$We will now be interested in the potential field function $\tilde{W}:\mathbb{R}^{3}\times \mathbb{R}^{3}\times \mathbb{R}^{3}\rightarrow \mathbb{R}$ in a vicinity of the relative distance vector $\tilde{R}(t):=R(t)-R_{f}(t)\in
\mathbb{R}^{3},$ keeping in mind that the interaction between particles $q_{f}$ and $q_{\text{ }}$depends on this relative interparticle distance $\tilde{R}(t)\in \mathbb{R}^{3}.$ From (\[M3.1\]), as vector $r\rightarrow \tilde{R}(t),$ we easily derive that $$\left. \frac{\partial \tilde{W}}{\partial R_{f}}+\frac{\partial \tilde{W}}{\partial R}\right\vert _{r\rightarrow \tilde{R}(t)}=0,\text{ \ \ \ \ }\left.
\frac{\partial \tilde{W}}{\partial R}+\frac{\partial \tilde{W}}{\partial r}\right\vert _{r\rightarrow \tilde{R}(t)}=0. \label{M3.2}$$Combining relations (\[M3.2\]) with the dynamical field equations ([M2.3]{}) we obtain that $$\begin{array}{c}
\frac{1}{c^{2}}\frac{\partial }{\partial t}\left. \left( \tilde{W}v\right)
\right\vert _{r\text{ }\rightarrow \tilde{R}(t)}=\frac{1}{c^{2}}\left.
\left( <\frac{\partial \tilde{W}}{\partial R_{f}},u_{f}>v+<\frac{\partial
\tilde{W}}{\partial R},u>v+\tilde{W}\frac{\partial v}{\partial t}\right)
\right\vert _{r\rightarrow \tilde{R}(t)}= \\
=\frac{1}{c^{2}}\left. \left( <\frac{\partial \tilde{W}}{\partial \tilde{R}},u-u_{f}>(u-u_{f})\right) \right\vert _{r\rightarrow \tilde{R}(t)}=-\left.
\frac{\partial \tilde{W}}{\partial r}\right\vert _{r\rightarrow \tilde{R}(t)}=\left. \frac{\partial \tilde{W}}{\partial R}\right\vert _{r\rightarrow
\tilde{R}(t)},\end{array}$$whence one derives the new dynamical equation $$\frac{d}{dt}(-\frac{\bar{W}}{c^{2}}(u-u_{f}))=-\frac{\partial \bar{W}}{\partial R} \label{M3.3}$$on the resulting function $\bar{W}:=\tilde{W}|_{r\rightarrow \tilde{R}(t)}$
Equation (\[M3.3\]) possesses a very important feature of depending on the only relative quantities not depending on the reference system. Moreover, we have not, on the whole, met the necessity to use other transformations of coordinates different from the Galilean transformations. We mention here that the dynamical equation (\[M3.3\]) was also derived in [@Re] making use of some not completely true relationships and mathematical manipulations. But the main corollary of [@Re] and our derivation [PB]{}, saying that equation (\[M3.3\]) fits for all reference systems, both inertial and accelerated, appears to be fundamental and gives rise to new unexpected results in the modern electrodynamics and gravity theory. Below we will proceed with one of very important relativity physics aspects, concerned with the well-known Lorentz force expression measuring the action exerted by external electromagnetic field on a charged point particle $q$ at space point $R(t)\in \mathbb{R}^{3}$ for any time moment $t\in \mathbb{R}.$
To do this we accept, owing to the vacuum field theory, that the resulting potential field function $\bar{W}:\mathbb{R}^{3}\rightarrow \mathbb{R}$ can be represented in the vicinity of the charged point particle $q$ as $$\bar{W}=\bar{W}_{0}+q\varphi , \label{M3.3a}$$where $\varphi :\mathbb{R}^{3}\rightarrow \mathbb{R}$ is a suitable local electromagnetic field potential and $\bar{W}_{0}:\mathbb{R}^{3}\rightarrow
\mathbb{R}$ is a constant vacuum field potential owing to the particle interaction with the external distant Universe objects. Then, having substituted (\[M3.3a\]) into the main dynamical field equation ([M3.3]{}) we obtain that $$\begin{aligned}
\frac{d}{dt}(-\frac{\bar{W}}{c^{2}}u) &=&\frac{d}{dt}(-\frac{\bar{W}}{c^{2}}u_{f})-\nabla \bar{W}=-\nabla \bar{W}+\frac{\partial }{\partial t}(-\frac{\bar{W}}{c^{2}}u_{f})+<u,\nabla >(-\frac{\bar{W}}{c^{2}}u_{f})= \notag \\
&=&-\nabla \bar{W}+\frac{1}{c}\frac{\partial }{\partial t}(-\frac{\bar{W}}{c}u_{f})-u\times (u_{f}\times \nabla \frac{\bar{W}}{c^{2}})-<u,u_{f}>\nabla
\bar{W}= \notag \\
&=&-\nabla \bar{W}(1+\frac{<u,u_{f}>}{c^{2}})+\frac{1}{c}\frac{\partial }{\partial t}(-\frac{\bar{W}}{c}u_{f})-\frac{1}{c^{2}}u\times (u_{f}\times
\nabla \bar{W})= \notag \\
&=&-q\nabla \varphi (1+\frac{<u,u_{f}>}{c^{2}})-\frac{q}{c}\frac{\partial }{\partial t}(\frac{\varphi }{c}u_{f})+\frac{q}{c}u\times (\nabla \times \frac{\varphi u_{f}}{c})= \notag \\
&=&-q\nabla \varphi (1+\frac{<u,u_{f}>}{c^{2}})-\frac{q}{c}\frac{\partial A}{\partial t}+\frac{q}{c}u\times (\nabla \times A), \label{M3.3b}\end{aligned}$$where we denoted $u:=dR(t)/dt,$ $u_{f}:=dR_{f}(t)/dt,$ $\nabla :=\partial
/\partial R=-\partial /\partial R_{f}$ and $A:=\varphi u_{f}/c$ to be the related magnetic potential. Since we have already shown that the Lorentz force $$F:=\frac{d}{dt}(-\frac{\bar{W}}{c^{2}}u)=\frac{d}{dt}\left( \frac{m_{0}u}{\sqrt{1-\frac{u^{2}}{c^{2}}}}\right)$$is given by expression (\[M3.3b\]), it can be rewritten in the form $$\begin{aligned}
F &=&\frac{d}{dt}\left( \frac{m_{0}u}{\sqrt{1-\frac{u^{2}}{c^{2}}}}\right)
=qE+\frac{q}{c}u\times B-\frac{q}{c^{2}}\nabla \varphi <u,u_{f}>=
\label{M3.4} \\
&=&qE+\frac{q}{c}u\times B-\frac{q}{c}\nabla <u,A>, \notag\end{aligned}$$which was derived also in [@Re] and where we put, by definition, $E:=-\nabla \varphi -\frac{1}{c}\frac{\partial A}{\partial t},$ $B:=\nabla
\times A,$ being respectively the suitable electric and magnetic vector fields.
The resulting expression (\[M3.4\]) is almost completely equivalent to the well-known [@AB] classical Lorentz force expression $F$ up to the additional “inertial” term $$F_{c}:=-\frac{q}{c}\nabla <u,A>, \label{M3.5}$$which is absent in the Einstein special relativistic theory. Namely, owing to the absence of the term (\[M3.5\]) the classical relativistic Lorentz force expression in the four-vector form was not invariant with respect to any reference frame transformations, except inertial ones. And, as was noticed in [@Re; @Br], owing only to this crucial fact A. Einstein introduced the Lorentz transformations and related with them visible length shortening and time slowing effects! Moreover, they gave rise to such strange enough and non-adequate notions as non-Euclidean time-spaces [Ho,Me,Ba1,Br]{}, black holes [@Da; @Gr; @Ho; @Ba] and some other singular and nonphysical objects.
Now based on the Lorentz force expression (\[M3.4\]) we can easily obtain the corresponding energy conservation law for a moving charged point particle $q$ and the ambient magnetic field:$$\frac{dE_{0}^{2}}{dt}=-2c<\nabla _{R_{f}}\bar{W},qA>=-\frac{dE_{f}^{2}}{dt_{f}},\text{ } \label{M3.6}$$where we put, by definition, that $$E_{0}^{2}:=\bar{W}^{2}-c(p^{^{\prime }})^{2},\text{ \ \ \ }E_{f}:=\bar{W}|_{t=t_{f}} \label{M3.7}$$where $p^{\prime }:=p-\frac{q}{c}A$ is the *“shifted” *momentum of the charged point particle $q$ in the external magnetic field $B=\nabla
\times A,$ and took into account that the parameter $\ R_{f}:=\
R_{f}(t)|_{t=t_{f}}\in \mathbb{R}^{3}$ and the evolution parameter $t_{f}=t\in \mathbb{R}$ corresponds to the change of the only magnetic field potential energy $E_{f}.$ The obtained expressions (\[M3.6\]) and (\[M3.7\]) take, evidently, into account the natural balance of energies belonging to both the moving charged point particle $q$ and the ambient magnetic field.
Quantum mechanics backgrounds revisited
=======================================
We will start from relation (\[M2.17\]) rewritten in the following form:$$E_{0}^{2}=\bar{W}-p^{2}c^{2},\text{ \ }dE_{0}/dt=0, \label{M4.1}$$and make the following canonical quantization replacements: $$p\rightarrow \hat{p}:=\frac{\hbar }{i}\nabla ,\text{ \ \ }E_{0}\rightarrow
\hat{E}_{0}:=-\frac{\hbar }{i}\frac{\partial }{\partial t},\text{ \ \ }\bar{W}\rightarrow \bar{W}\circ , \label{M4.2}$$where $\hat{p}=\frac{\hbar }{i}\nabla $ is the standard spatial translation generator in $\mathbb{R}^{3},$ $\hat{E}_{0}=-\frac{\hbar }{i}\frac{\partial
}{\partial t}$ is the standard time translation operator along the real time axis $\mathbb{R}$ and $\bar{W}\circ $ is a usual scalar multiplication operator on the function $\bar{W}:\mathbb{R}^{3}\rightarrow \mathbb{R},$ all acting in the Hilbert space $\mathcal{H}:=$ $L_{2}(\mathbb{R}^{3};\mathbb{C})$ with the standard scalar product $(\cdot ,\cdot ).$ As an elementary result of this replacement we can easily write, following the quantization recipes similar to those from [@Di; @AB; @BS; @BjD], that the observable squared energy $E_{0}^{2}$ is the average value $$E_{0}^{2}:=(\hat{E}_{0}\psi ,\hat{E}_{0}\psi )=(\psi ,(\bar{W}^{2}+\hbar
^{2}c^{2}\nabla ^{2})\psi ), \label{M4.3}$$being realized on the one-particle quantum mechanical state vector $\psi \in
\mathcal{H}.$ Taking into account representation (\[M4.3\]) and assuming that the Planck constant $\hbar \rightarrow 0,$ we now will try to factorize the linear differential operator $\bar{W}^{2}+\hbar
^{2}c^{2}\nabla ^{2},$ well defined on a suitable dense linear subset $D(\hat{E}_{0})\subset \mathcal{H},$ in the following *a priori* nonnegative canonical form:$$\bar{W}^{2}+\hbar ^{2}c^{2}\nabla ^{2}=\hat{P}^{+}\hat{P}, \label{M4.4}$$where the sign $"+"$ means the standard conjugation operation in the Hilbert space $\mathcal{H}.$ It is easy to find from (\[M4.4\]) that $$\hat{P}:=\hat{S}(1+\hbar ^{2}c^{2}\bar{W}^{-1}\circ \nabla ^{2}\circ \bar{W}^{-1})^{1/2}\bar{W}\circ \label{M4.5}$$with an arbitrary unitary operator $\hat{S}:\mathcal{H\rightarrow H},$ satisfying the usual condition $\hat{S}^{+}\hat{S}=1.$ As a result, we obtain from (\[M4.3\]) and (\[M4.4\]) that $$(\hat{E}_{0}\psi ,\hat{E}_{0}\psi )=(\psi ,\hat{P}^{+}\hat{P}\psi )=(\hat{P}\psi ,\hat{P}\psi ). \label{M4.6}$$Thereby, based on expression (\[M4.6\]), we can derive the following Schrödinger type linear evolution equation:$$\hat{E}_{0}\psi :=i\hbar \frac{\partial \psi }{\partial t}=\hat{U}\hat{P}\psi \label{M4.7}$$for all $t\in \mathbb{R},$ where $\hat{U}:\mathcal{H\rightarrow H}$ is some unitary operator to be determined later.
Let us now symbolically calculate the operator expression (\[M4.5\]) up to the symbolic operator accuracy $O(\hbar ^{4}):$$$\hat{P}=\hat{S}(1+\hbar ^{2}c^{2}\bar{W}^{-1}\circ \nabla ^{2}\circ \bar{W}^{-1})^{1/2}\bar{W}=\hat{S}(1+\frac{\hbar ^{2}c^{2}}{2\bar{W}}\nabla
^{2}\circ \bar{W}^{-1})\bar{W}+O(\hbar ^{4}). \label{M4.8}$$Having substituted result (\[M4.8\]) into (\[M4.7\]) one obtains (up to the operator accuracy $O(\hbar ^{4})$) the next evolution equation:$$i\hbar \frac{\partial \psi }{\partial t}=\hat{U}\hat{S}(-\frac{\hbar ^{2}}{2(-\bar{W}/c^{2})}\nabla ^{2}+\bar{W})\psi . \label{M4.9}$$Taking into account that the energy operator $\hat{E}_{0}=-\frac{\hbar }{i}\frac{\partial }{\partial t}$ is formally self-adjoint, from a physical point of view we need to choose in (\[M4.9\]) that $\hat{U}:=\hat{S}^{-1},
$ as the operator $\ \hat{H}:=-\frac{\hbar ^{2}}{2(-\bar{W}/c^{2})}\nabla
^{2}+\bar{W}$ is formally self-adjoint.
Recalling now that owing to (\[M2.9\]) and (\[M2.14\]) the expression $-\bar{W}/c^{2}:=m(u)=m_{0}/\sqrt[2]{1-\frac{u^{2}}{c^{2}}}$ is the dynamical relativistic mass of a particle, whose quantum motion is under study, equation (\[M4.9\]) can be finally rewritten in the following classical self-adjoint Schrödinger type form:$$i\hbar \frac{\partial \psi }{\partial t}=\hat{H}\text{ }\psi :=-\frac{\hbar
^{2}}{2m(u)}\nabla ^{2}\psi +\bar{W}\psi , \label{M4.11}$$where $\hat{H}:\mathcal{H\rightarrow H}$ is the corresponding self-adjoint Hamiltonian operator for a relativistic point particle under an external potential field. It is easy to observe that equation (\[M4.11\]) in the classical non-relativistic case when the velocity $c\rightarrow \infty $ reduces to the well-known classical Schrödinger equation $$i\hbar \frac{\partial \psi }{\partial t}=\hat{H}_{0}\text{ }\psi :=-\frac{\hbar ^{2}}{2m_{0}}\nabla ^{2}\psi +\bar{W}\psi , \label{M4.12}$$derived at the beginning of the past century from completely different points of view by the great physicists Schrödinger, Heisenberg and Dirac.
Now we proceed to the similar quantization procedure of expression ([M3.7]{}) for the case when the magnetic potential is not equal to zero. Making use of the standard quantization scheme (\[M4.2\]) we obtain, by definition, that the average squared energy value $E_{0}^{2}$ on a suitable quantum state vector $\psi \in \mathcal{H}$ equals $$E_{0}^{2}:=(\hat{E}_{0}\psi ,\hat{E}_{0}\psi )=(\psi ,[\hat{W}^{2}-c^{2}(\frac{\hbar }{i}\nabla +\frac{q}{c}\hat{A})^{2}]\psi ), \label{M4.13}$$where, as before, $\hat{W}:=\bar{W}\circ $ and $\hat{A}:=\bar{A}\circ $ are the scalar multiplication operators. The operator expression in the squared brackets on the right-hand side of (\[M4.13\]) can easily be represented in the following strongly nonnegative form:$$\hat{W}^{2}-c^{2}(\frac{\hbar }{i}\nabla +\frac{q}{c}\hat{A})^{2}:=\hat{P}^{+}\hat{P}, \label{M4.14}$$where, as above, one can take$$\hat{P}:=[1-\hat{W}^{-1}(\frac{\hbar }{i}\nabla +\frac{q}{c}\hat{A})^{2}\hat{W}^{-1}]^{1/2}\hat{W}. \label{M4.15}$$Then from equation (\[M4.13\]) and expression (\[M4.14\]) we can write down the following “magnetic” Schrödinger type evolution equation:$$i\hbar \frac{\partial \psi }{\partial t}:=\hat{P}\psi , \label{M4.16}$$which gives rise, under simultaneous conditions $\hbar \rightarrow 0$ and $c\rightarrow \infty ,$ if the product $c\hbar =const,$ to the following result:$$i\hbar \frac{\partial \psi }{\partial t}:=\hat{H}\psi =[\frac{1}{2m(u)}(\frac{\hbar }{i}\nabla +\frac{q}{c}\hat{A})^{2}+W_{0}+q\varphi ]\psi
\label{M4.17}$$with accuracy $O(\hbar ^{4}).$ Equation (\[M4.17\]) reduces to the well-known [@Di] classical “magnetic” Schrödinger type evolution equation with accuracy $O(\hbar ^{4})$: $$i\hbar \frac{\partial \psi }{\partial t}:=\hat{H}_{0}\psi =[\frac{1}{2m_{0}}(\frac{\hbar }{i}\nabla +\frac{q}{c}\hat{A})^{2}+W_{0}+q\varphi ]\psi ,
\label{M4.18}$$where, as before, we put $m_{0}=-\bar{W}_{0}/c^{2},$ since the potential energy relationship $|q\varphi /\bar{W}_{0}|<<1.$
The results obtained above convey very eloquently that our quasiclassical approach to the description of the vacuum field structure, as the Planck constant $\hbar \rightarrow 0$ and the light velocity $c\rightarrow \infty ,$ give rise to the related classical quantum Schrödinger dynamics very naturally, being simultaneously deeply physically motivated. It is also worthy to mention that in all of our derivations we have used no Lorentz invariance in spite of the fact that the dynamical mass $m=m_{0}/\sqrt[2]{1-\frac{u^{2}}{c^{2}}}$ is expressed by means of a suitable Lorentz factor. It is, evidently, a by-product result of the *second* field theory principle equation (\[M1.2\]), which is *a priori* Lorentz invariant.
To finish the discussion we make here a plausible claim that the suitable completely relativistic quantum field theoretic analysis of our dynamical vacuum field model via the well-known Dirac type factorization approach [Di]{} will necessary shed new light on the so complicated and sophisticated nature of the quantum world of elementary particles.
Concerning the results described above we could state that the vacuum field theory approach of [@Re; @PB; @Br; @Ca; @BD; @Me1; @B-B] to fundamental physical phenomena appears to be really a powerful tool in the hands of researchers, who wish to penetrate into the hidden properties of the surrounding Universe. As the microscopical quantum level of describing the vacuum field matter structure is, with no doubt, very important, we see the next challenging steps, both in understanding the backgrounds of quantum processes within the approaches devised in [@Br; @Re; @PB] and in this work, and in deriving new physical relationships, which will help us to explain the Nature more deeply and adequately.
Acknowledgments
===============
The Authors are cordially thankful to the Abdus Salam International Centre for Theoretical Physics in Trieste, Italy, for the hospitality during their research 2007-2008 scholarships. A.P. is, especially, grateful to Profs. P.I. Holod (Kyiv, UKMA), J.M. Stakhira (Lviv, NUL), B.M. Barbashov (Dubna, JINR), Z. K[á]{}kol (Krak[ó]{}w, AGH), J. S[ł]{}awianowski (Warsaw, IPPT), Z. Peradzy[ń]{}ski (Warsaw, UW) and M. B[ł]{}aszak (Poznań, UP) for fruitful discussions, useful comments and remarks. Last but not least thanks go to Prof. O. N. Repchenko for the discussion of some controversial vacuum field theory aspects and to Mrs. Dilys Grilli (Trieste, Publications office, ICTP) for professional help in preparing the manuscript for publication.
[99]{} Feynman R. P M. . Lectures on gravitation. Notes of California Inst. of Technology, 1971
Feynman R., Leighton R. and Sands M. The Feynman lectures on physics. Electrodynamics, v. 2, Addison-Wesley, Publ. Co., Massachusetts, 1964
Dirac P.A.M. The principles of quantum mechanics. Second edition. Oxford, Clarendon Press, 1935
Bogolubov N. and Shirkov D. Introduction to the theory of quantized fields. Interscience, New York, 1959
Bjorken J.D. and Drell S.D. Relativistic quantum fields. Mc Graw-Hill Book Co., NY, 1965
Akhiezer A.I. and Berestecky V.B. Quantum electrodynamics. Nauka Publisher, Moscow, 1969
Schwinger J. Quantum electrodynamics. Dover Publ. Inc. New York, 1958
de Gennes P.-G. Superconductivity of metals and alloys, Benjamin, USA, 1964
Brillouin L. Relativity reexamined. Academic Press Publ., New York and London, 1970
Brillouin L. and Lucas R. Journ. Phys. Radium, 27, 1966, p. 229
Heaviside O. Electromagnetic Theory. New York, 1893, p. 115-118
Carstoiu J. Compt. Rend., 268, 1969, p. 201
Markov M.A. The Mach principle and physical vacuum in general relativity. Problems of Theoretical Physics. Essays dedicated to Nikolai N. Bogolubov on the occasion of his sixtith birthday. Nauka Publisher, Moscow, 1969, p. 26-27
Faddeev L.D. Hamiltonian approach to the gravity. Russian Physical Surveys, Nauka Publisher, Moscow, 1986
Maxwell J.C. Oeuvres on Electromagnetic Theory. Cambrridge, 1950
Petrov A.Z. Proceedings of the USSR Academy of Sciences, 190, 1970, p. 305
Burghardt R. Acta Phys. Austr., 32, 1970, p. 272-281
Repchenko O. Field physics. Moscow, “Galeria” Publ., 2005
Brans C.H. and Dicke R.H. Mach’s principle and a relativistic theory of gravitation. Phys. Rev., 124, 1961, p. 925
Bialynicky-Birula I. Phys Rev., 155, 1967, p. 1414; 166,1968, p. 1505
Logunov A.A. Lectures on relativity theory and gravitation. Nauka Publisher, Moscow, 1987
Fock V.A. Theory of space, time and gravity. Nauka Publisher, Moscow, 1955
Fock V.A. Konfigurationraum und zweite Quantelung. Zeischrift Phys., Bd. 75, 1932, p. 622-647
Sommerfeld A. Mechanics. v.1, New York, 1952
t’Hooft G. Introduction to general relativity. Institute for Theoretical Physics Utrecht University, Princetonplein 5, 3584 CC Utrecht, the Netherlands, 2002 (www.phys.uu.nl/ thooft/lectures/genrel.pdf)
Taylor E.F. and Wheeler J. A. Space-time physics. Freeman and Company, San Francisco and London, 1966
Pauli W. Theory of relativity. Oxford Publ., 1958
Bridgeman P.W. Reflections of a Scientist. Philosophical Library, New York, 1955
Weinstock R. New approach to special relativity. Am. J. Phys., 33, 1965, p. 640-645.
Lee A.R., Kalotas T.M. Lorentz transformations from the first postulate. Am. J. Phys., 43, 1975, p. 434-437.
Levy-Leblond J.M. One more derivation of the Lorentz transformation. Am. J. Phys., 44, 1976, 271-277.
Mermin N.D. Relativity without light. Am. J. Phys., 52, 1984, p. 119-124.
Sen A. How Galileo could have derived the special theory of relativity. Am. J. Phys., 62, 1994, p. 157-162.
Mermin D.N. It’s About Time: Understanding Einstein’s Relativity, Princeton, NJ., Princeton University Press, 2005
Barbashov B. M. and Nesterenko V. V. Introduction to the Relativistic String Theory, World Scientific, Singapore, 1990
Collins H. Gravity’s Shadow: The Search for Gravitational Waves, Chicago: University of Chicago Press, 2004
Marsden J. and Chorin A. Mathematical foundations of the mechanics of liquid. Springer, New York, 1993
Barbashov B.M., Efimov G.V. and others. (Editors) Selected Problems of Modern Physics. Proceedings of the XII-th International Conference on Selected Problems of Modern Physics. Section 1, Dubna, 2003
Prykarpatsky A.K. and Bogolubov N.N. (Jr.) The field structure of vacuum, Maxwell equations and relativity theory aspects. Preprint ICTP, Trieste, IC/2008/051 (http://publications.ictp.it; arXiv lanl: 0807.3691v.8 \[gr-gc\] 24.08.2008 )
Bogolubov N.N. (Jr.), Prykarpatsky A.K. , Golenia J. and Taneri U. Introductive backgrounds of modern quantum mathematics and application to nonlinear dynamical systems. Intern. Journal of Theor. Physics, v.47, N2, 2008; Preprint ICTP, Trieste, IC/2007/108 (http://publications.ictp.it)
Bogolubov N.N. (Jr.), Prykarpatsky A.K. and Taneri U. Quantum field theory with application to quantum nonlinear optics. World Scientific Publishing Co., 2003
Berezin F.A. The method of second quantization. Nauka Publisher, Moscow, 1965
Damour T. General Relativity and Experiment. Proceedings of the XI International Congress on Mathematical Physics, Intern. Press, 1995. Proceedings of the International Conference
Green B. The elegant Universe. Vintage Books Inc., New York, 1999
[^1]: nikolai\[email protected]
[^2]: [email protected], [email protected]
[^3]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study a stochastic extended Korteweg - de Vries equation driven by a multiplicative noise. We prove the existence of a martingale solution to the equation studied. The proof of the solution is based on two approximations of the problem considered and the compactness method.'
address:
- 'Faculty of Mathematics, Computer Science and Econometrics, University of Zielona Góra, Szafrana 4a, 65-516 Zielona Góra, Poland'
- ' Faculty of Mathematics, Computer Science and Econometrics, University of Zielona Góra, Szafrana 4a, 65-516 Zielona Góra, Poland'
author:
- Anna Karczewska
- Maciej Szczeciński
title: 'Martingale solution to stochastic extended Korteweg - de Vries equation'
---
[^1]
Introduction {#s1}
============
The celebrated Korteweg - de Vries equation (*KdV* for short) [@kdv], derived from the set of Eulerian shallow water and long wavelength equations, become a paradigm in the field of nonlinear partial differential equations. KdV appears as the lowest approximations of wave motion in several fields of physics, see, e.g., monographs [@DrJ; @EIGR; @Newell; @Rem; @Whit] and references therein.
KdV is, however, the result of an approximation of the set of the Euler equations within perturbation approach limited to the first order in expansion with respect to parameters assumed to be small. Several authors have extended KdV to the second order (the *extended KdV* or *KdV2*), e.g. [@MS90; @MS96; @BS; @KRI; @KRR; @KRI2; @KRIR; @Yang], which is a more exact approximation of the Euler equations but far more difficult since it contains higher nonlinear term and higher derivatives.
Despite its non-integrability, KdV2 has exact analytic solutions both solitonic [@KRI] and periodic [@RK; @RKI]. These solutions have the same form as corresponding solutions to KdV but with slightly different coefficients.
A natural continuation of the study of the extended KdV equation seems to be considering stochastic versions of such equation. KdV2 equation driven by random noise can be a model of several kinds of waves (e.g., surface water waves, waves in plasma) influenced by random factors. Two cases of the stochastic KdV2 equation are possible - the case with additive noise and the case with the multiplicative noise. The additive case we studied in [@KaSz1], where a mild solution to KdV2 has been established.
In this paper, we consider the stochastic extended Korteweg - de Vries equation with multiplicative random noise. We prove the existence of martingale solution to stochastic KdV2 equation driven by cylindrical Wiener process. In the proof, we generalize the methods used in papers [@Deb] and [@Gat]. We have to emphasize that the method used in [@Deb] for estimations was not suitable in our case. We adapted for our purposes (proof of Lemma 2.4) the approach used by Flandoli and Gątarek in [@Gat].
Existence of martingale solution
================================
We consider initial value problem for Korteweg - de Vries type equation $$\label{Apr}
\begin{cases}
{\mathop{}\mathrm{d}}u(t,x) + \big[ u_{3x}(t,x) + u(t,x)u_{x}(t,x) + u(t,x)u_{3x}(t,x) \\
\hspace{11ex} +\, 3 u_{x}(t,x)u_{2x}(t,x) \big] {\mathop{}\mathrm{d}}t = \Phi\left(u(t,x)\right) {\mathop{}\mathrm{d}}W(t), \\
u(0,x) = u_{0}(x).
\end{cases}$$ In (\[Apr\]), $W(t)$, $t\geq 0$, is a cylindrical Wiener process adapted to the filtration $\left\{\mathscr{F}_{t}\right\}_{t\geq 0}$, $u_{0} \in L^{2}(\mathbb{R}) $ is a deterministic function, $u(\omega,\cdot,\cdot):\mathbb{R}_{+}\times\mathbb{R} \rightarrow \mathbb{R}$ for all $\omega\in\Omega$. Moreover, we assume that $|u(t,x)|+|u|_{L^{2}(\mathbb{R})}<\lambda<\infty, ~\lambda>0, \mbox{~for~all~} t\in\mathbb{R}_{+} \mbox{~and~} x\in\mathbb{R}$, what reflects finitness of solutions to deterministic version of the equation (\[Apr\]) (see, e.g., [@KRI; @RK; @RKI]). The operator $\Phi$ is a continuous mapping from $H^{2}(\mathbb{R})$ to $L_{2}^{0}(L^{2}(\mathbb{R}))$, the space of Hilbert-Schmidt operators from $L^{2}(\mathbb{R})$ to itself. The operator $\Phi$ is such that for any $u\in H^{2}(\mathbb{R})$ the following conditions hold: $$\label{W1}
\displaystyle\mathop{\exists}_{\kappa_{1}, \kappa_{2}>0} \quad \left\|\Phi(u(x))\right\|_{L_{0}^{2}(L^{2}(\mathbb{R}))} \leq \kappa_{1} \max \left\{\left|u(x)\right|^{2}_{L^{2}(\mathbb{R})}, \left|u(x)\right|_{L^{2}(\mathbb{R})}\right\} + \kappa_{2};$$ $$\begin{aligned}
\label{W3}
&\mbox{there~exist~such~functions}~ a,b\in L^{2}(\mathbb{R})~ \mbox{with~compact~support,~that~the~mapping} \\
&u \mapsto \left( \Phi(u)a , \Phi(u)b \right)_{L^{2}(\mathbb{R})}~ \mbox{is~continuous~in~topology}~ L^{2}_{loc}(\mathbb{R}).\nonumber\end{aligned}$$
We say that the problem (\[Apr\]) has a [martingale solution]{} on the interval $[0,T]$, $T>0$, if there exists a stochastic basis $(\Omega,\mathscr{F},\left\{\mathscr{F}_{t}\right\}_{t\geq 0}, \mathbb{P}, \left\{W_{t}\right\}_{t\geq 0} )$, where $\left\{W_{t}\right\}_{t\geq 0}$ is a cylindrical Wiener process, and there exists the process $\left\{u(t,x)\right\}_{t\geq 0}$ adapted to the filtration $\left\{\mathscr{F}_{t}\right\}_{t\geq 0}$ with trajectories belonging to the space $$\nonumber
L^{\infty}(0,T;L^{2}(\mathbb{R}))\cap L^{2}(0,T;L^{2}_{loc}(\mathbb{R})\cap \mathscr{C}(0,T;H^{s}_{loc}(\mathbb{R}), \quad s<0, \quad \mathbb{P} - \text{a.s.}$$ such that $$\nonumber
\begin{aligned}
& \left\langle u(t,x), v(x) \right\rangle + \int_{0}^{t} \left\langle u_{3x}(s,x) + u(s,x)u_{x}(s,x) + u(s,x)u_{3x}(s,x) \right. \\
& \left. + 3 u_{x}(s,x)u_{2x}(s,x), v(x) \right\rangle {\mathop{}\mathrm{d}}s = \left\langle u_{0}(x), v(x)\right\rangle + \left\langle \int_{0}^{t} \Phi(u(s,x)) {\mathop{}\mathrm{d}}W(s), v(x) \right\rangle
\end{aligned}$$ for any $t\in[0,T]$ and $v\in H^{1}_{loc}(\mathbb{R})$.
Now, we can to formulate the main result of the paper.
\[P4.1\] For all $u_{0}\in L^{2}(\mathbb{R})$ and $T>0$ there exists a martingale solution to (\[Apr\]) with conditions (\[W1\]) and (\[W3\]).
Let $\varepsilon>0$. Consider
$$\label{par}
\begin{cases}
{\mathop{}\mathrm{d}}u^{\varepsilon}(t,x) + \big[ \varepsilon u^{\varepsilon}_{4x}(t,x) + u^{\varepsilon}_{3x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{x}(t,x) + 3u_{x}^{\varepsilon}(t,x)u^{\varepsilon}_{2x}(t,x) \\
\hspace{12ex} + u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x) \big] {\mathop{}\mathrm{d}}t
= \Phi\left(u^{\varepsilon}(t,x)\right){\mathop{}\mathrm{d}}W(t) \\
u^{\varepsilon}_{0}(x) = u^{\varepsilon}(0,x) .
\end{cases}$$
\[parMart\] For any $\varepsilon>0$ there exists a martingale solution to the problem (\[par\]) with conditions (\[W1\]) and (\[W3\]).
\[szac4.1\] There exists $\varepsilon_{0} > 0$, such that $$\begin{aligned}
\label{4.1a}
\exists_{C_{1}>0}\forall_{0 < \varepsilon < \varepsilon_{0}} \varepsilon\mathbb{E}\left( \left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{2}(\mathbb{R}))} \right) & \leq \tilde{C}_{1}, \\
\label{4.1c}
\forall_{k\in X_{k}}\exists_{C_{2}(k)>0}\forall_{0 < \varepsilon < \varepsilon_{0}} \mathbb{E}\left( \left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{1}(-k,k))} \right) & \leq \tilde{C}_{2}(k), \end{aligned}$$ where $X_{k} = \big\{k>0: \left|k\right| \leq \min\left\{-x_{1}, x_{2}\right\}\big\}$.
\[PropCias\] The family of distributions $\mathscr{L}(u^{\varepsilon})$ is tight in $L^{2}(0,T;L^{2}_{loc})\cap\mathscr{C}(0,T;H^{-3}_{loc}(\mathbb{R}))$.
Proofs of Lemmas \[parMart\], \[szac4.1\] and \[PropCias\] are given in sections \[sec3\] and \[sec4\].
Substitute in Prohorov’s theorem (e.g., see Theorem 5.1 in [@Bil]) $S:=L^{2}(0,T;L^{2}_{loc}(\mathbb{R}))\cap\mathscr{C}(0,T;H^{-3}_{loc}(\mathbb{R}))$ and $\mathscr{K}:=\left\{\mathscr{L}(u^{\varepsilon})\right\}_{\varepsilon>0}$. Since $\mathscr{K}\subset\mathscr{P}(S)$ is tight in $S$, then it is sequentially compact, so there exists a subsequence of $\left\{\mathscr{L}(u^{\varepsilon})\right\}_{\varepsilon>0}$ converging to some measure $\mu$ in $\bar{\mathscr{K}}$.
Because $\left\{\mathscr{L}(u^{\varepsilon})\right\}_{\varepsilon>0}$ is convergent, then it is also weakly convergent. Therefore in Skorohod’s theorem (e.g., see Theorem 6.7 in [@Bil]) one can substitute $\mu_{\varepsilon}:=\left\{\mathscr{L}(u^{\varepsilon})\right\}_{\varepsilon>0}$, $\mu:=\lim_{\varepsilon\to 0} \mu_{\varepsilon}$. Then there exists a space $(\bar{\Omega}, \bar{\mathscr{F}}, \left\{\bar{\mathscr{F}}_{t}\right\}_{t\geq 0}, \bar{\mathbb{P}})$ and random variables $\bar{u}^{\varepsilon}$, $\bar{u}$ with values in $L^{2}(0,T;L^{2}_{loc})\cap \mathscr{C}(0,T;H^{-3}_{loc}(\mathbb{R}))$ such that $\bar{u}^{\varepsilon}\rightarrow\bar{u}$ in $L^{2}(0,T;L^{2}_{loc})$ and in $\mathscr{C}(0,T;H^{-3}_{loc}(\mathbb{R}))$. Moreover $\mathscr{L}(\bar{u}^{\varepsilon}) = \mathscr{L}(u^{\varepsilon})$.
Then due to Lemma \[szac4.1\], for any $p\in\mathbb{N}$ there exist constants $\tilde{C}_{1}(p)$, $\tilde{C}_{2}$ such that $$\nonumber
\begin{aligned}
\mathbb{E}(\sup_{t\in[0,T]} \left|\bar{u}^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{2p}) \leq \tilde{C}_{1}(p) \quad \mbox{and} \quad
\mathbb{E}(\left|\bar{u}^{\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{2}(\mathbb{R}))}) \leq \tilde{C}_{2}.
\end{aligned}$$ Additionally, $$\nonumber
\bar{u}^{\varepsilon}(t,x)\in L^{2}(0,T; H^{1}(-k,k))\cap L^{\infty}(0,T; L^{2}(\mathbb{R})), \quad \mathbb{P} - \text{a.s.}$$ Then one can conclude that $\bar{u}^{\varepsilon}\rightarrow \bar{u}$ weakly in $L^{2}(\bar{\Omega},L^{2}(0,T;H^{1}(-k,k)))$.
Let $x\in\mathbb{R}$ be fixed and denote $$\nonumber
\begin{aligned}
M^{ \varepsilon}(t) := & u^{ \varepsilon}(t,x) - u_{0}^{ \varepsilon}(x) + \int_{0}^{t}\bigg[ \varepsilon u^{\varepsilon}(t,x)_{4x}(t,x) + u^{ \varepsilon}(t,x)u^{ \varepsilon}_{x}(t,x)\\
& + u^{ \varepsilon}_{3x}(t,x) + 3u^{ \varepsilon}_{x}(t,x) u^{ \varepsilon}_{2x}(t,x) + u^{ \varepsilon}(t,x) u^{ \varepsilon}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s, \\
\bar{M}^{ \varepsilon}(t) := & \bar{u}^{ \varepsilon}(t,x) - \bar{u}_{0}^{ \varepsilon}(x) + \int_{0}^{t}\bigg[ \varepsilon \bar{u}^{\varepsilon}(t,x)_{4x}(t,x) + \bar{u}^{ \varepsilon}(t,x)\bar{u}^{ \varepsilon}_{x}(t,x)\\
& + \bar{u}^{ \varepsilon}_{3x}(t,x) + 3\bar{u}^{ \varepsilon}_{x}(t,x) \bar{u}^{ \varepsilon}_{2x}(t,x) + \bar{u}^{ \varepsilon}(t,x) \bar{u}^{ \varepsilon}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s.
\end{aligned}$$
Note, that $$\nonumber
\begin{aligned}
M^{ \varepsilon}(t) = & u_{0}^{ \varepsilon}(x) - \int_{0}^{t}\bigg[ \varepsilon u^{\varepsilon}(t,x)_{4x}(t,x) + u^{ \varepsilon}(t,x)u^{ \varepsilon}_{x}(t,x) + u^{ \varepsilon}_{3x}(t,x) + 3u^{ \varepsilon}_{x}(t,x) u^{ \varepsilon}_{2x}(t,x)\\
& + u^{ \varepsilon}(t,x) u^{ \varepsilon}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s + \int_{0}^{t} \left( \Phi \left( u^{ \varepsilon} (s,x) \right) \right) {\mathop{}\mathrm{d}}W^{ \varepsilon} (s) - u_{0}^{ \varepsilon}(x) + \int_{0}^{t}\bigg[ \varepsilon u^{\varepsilon}(t,x)_{4x}(t,x) \\
& + u^{ \varepsilon}(t,x)u^{ \varepsilon}_{x}(t,x) + u^{ \varepsilon}_{3x}(t,x) + 3u^{ \varepsilon}_{x}(t,x) u^{ \varepsilon}_{2x}(t,x) + u^{ \varepsilon}(t,x) u^{ \varepsilon}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s \\
= & \int_{0}^{t} \left( \Phi \left( u^{ \varepsilon} (s,x) \right) \right) {\mathop{}\mathrm{d}}W (s), \\
\end{aligned}$$ so, $M^{ \varepsilon}(t)$, $t\geq 0$, is a square integrable martingale with values in $L^{2}(\mathbb{R})$, adapted to filtration $\sigma\left\{u^{ \varepsilon}(s,x), 0\leq s \leq t\right\}$ with quadratic variation $$\nonumber
\left[M^{ \varepsilon}\right](t) = \int_{0}^{t}\Phi(u^{ \varepsilon}(s,x))\left[\Phi(u^{ \varepsilon}(s,x))\right]^{*} {\mathop{}\mathrm{d}}s .$$
Substitute in the Doob inequality (e.g., see Theorem 2.2 in [@Gaw]) $M_{t}:=M^{ \varepsilon}(t)$ and $p:=2p$. Then $$\label{szacDoob}
\mathbb{E}\left[\left( \sup_{t\in[0,T]} \left|M^{ \varepsilon}(t)\right|_{L^{2}(\mathbb{R})}^{p} \right)\right] \leq \left( \frac{p}{p-1} \right)^{p} \mathbb{E} \left(\left|M^{ \varepsilon}(T)\right|_{L^{2}(\mathbb{R})} \right).$$
Assume $0\leq s \leq t \leq T$ and let $\varphi$ be a bounded continuous function on $L^{2}(0,s;L^{2}_{loc}(\mathbb{R}))$ or $C(0,s;H^{-3}_{loc}(\mathbb{R}))$. Let $a,b\in H^{3}_{0}(-k,k)$, $k\in\mathbb{N}$, be arbitrary and fixed. Since $M^{ \varepsilon}(t)$ is a martingale and $\mathscr{L}(\bar{u}^{ \varepsilon}) = \mathscr{L}(u^{ \varepsilon})$, then (see [@Gat], p. 377-378) $$\nonumber
\begin{aligned}
\mathbb{E} & \Big( \left\langle M^{ \varepsilon}(t) - M^{ \varepsilon}(s); a \right\rangle \varphi\left(u^{ \varepsilon}(t,x)\right)\Big) = 0 , \\
\mathbb{E} & \Big( \left\langle \bar{M}^{ \varepsilon}(t) - \bar{M}^{ \varepsilon}(s); a \right\rangle\varphi\left(\bar{u}^{ \varepsilon}(t,x)\right)\Big) = 0 \\
\end{aligned}$$ and $$\nonumber
\begin{aligned}
\mathbb{E} & \bigg\{\bigg[\left\langle M^{ \varepsilon}(t);a \right\rangle \left\langle M^{ \varepsilon}(t);b \right\rangle - \left\langle M^{ \varepsilon}(s);a \right\rangle \left\langle M^{ \varepsilon}(s);b \right\rangle \\
& - \int_{s}^{t} \left\langle \left[\Phi\left(u^{ \varepsilon}(\xi,x)\right)\right]^{*}a ; \left[\Phi\left(u^{ \varepsilon}(\xi,x)\right)\right]^{*}b \right\rangle {\mathop{}\mathrm{d}}\xi\bigg]\varphi(u^{ \varepsilon}(t,x))\bigg\} = 0, \\
\mathbb{E} & \bigg\{\bigg[\left\langle \bar{M}^{ \varepsilon}(t);a \right\rangle \left\langle \bar{M}^{ \varepsilon}(t);b \right\rangle - \left\langle \bar{M}^{ \varepsilon}(s);a \right\rangle \left\langle \bar{M}^{ \varepsilon}(s);b \right\rangle \\
& - \int_{s}^{t} \left\langle \left[\Phi\left(\bar{u}^{ \varepsilon}(\xi,x)\right)\right]^{*}a ; \left[\Phi\left(\bar{u}^{ \varepsilon}(\xi,x)\right)\right]^{*}b \right\rangle {\mathop{}\mathrm{d}}\xi\bigg]\varphi(\bar{u}^{ \varepsilon}(t,x))\bigg\} = 0 .
\end{aligned}$$
Denote $$\bar{M}(t) := \bar{u}(t,x) - \bar{u}_{0}^{ \varepsilon}(x) + \int_{0}^{t}\bigg[ \bar{u}(t,x)\bar{u}_{x}(t,x)
+ \bar{u}_{3x}(t,x) + 3\bar{u}_{x}(t,x) \bar{u}_{2x}(t,x) + \bar{u}(t,x) \bar{u}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s.$$ If $\varepsilon \rightarrow 0 $, to $ \bar{M}^{ \varepsilon}(t) \rightarrow \bar{M}(t)$ and $\bar{M}^{ \varepsilon}(s) \rightarrow \bar{M}(s)$, $\bar{\mathbb{P}} - \text{a.s.}$ in $H^{-3}_{loc}(\mathbb{R})$. Moreover, since $\varphi$ is continuous, then $\varphi(\bar{u}^{ \varepsilon}(s,x)) \rightarrow \varphi(\bar{u}(s,x))$, $\bar{\mathbb{P}} - \text{a.s.}$. Therfeore, if $\varepsilon \rightarrow 0$, then $$\nonumber
\begin{aligned}
\mathbb{E}& \Big( \left\langle \bar{M}^{ \varepsilon}(t) - \bar{M}^{ \varepsilon}(s); a \right\rangle \varphi(\bar{u}^{ \varepsilon}(t,x))\Big) \rightarrow \mathbb{E} \Big( \left\langle \bar{M}(t) - \bar{M}(s); a \right\rangle \varphi(\bar{u}(t,x))\Big) .
\end{aligned}$$
Additionaly, because $\Phi$ is a continuous operator in topology $L_{loc}^{2}(\mathbb{R})$ and (\[szacDoob\]) holds, therefore if $\varepsilon \rightarrow 0$, then $$\nonumber
\begin{aligned}
\left\langle \left(\Phi(\bar{u}^{ \varepsilon}(s,x))\right)^{*}a; \left(\Phi(\bar{u}^{ \varepsilon}(s,x))\right)^{*}b\right\rangle \rightarrow \left\langle \left(\Phi(\bar{u}(s,x))\right)^{*}a; \left(\Phi(\bar{u}(s,x))\right)^{*}b \right\rangle
\end{aligned}$$ and $$\nonumber
\begin{aligned}
\mathbb{E} & \bigg\{\bigg[\left\langle \bar{M}^{ \varepsilon}(t);a \right\rangle \left\langle \bar{M}^{ \varepsilon}(t);b \right\rangle - \left\langle \bar{M}^{ \varepsilon}(s);a \right\rangle \left\langle \bar{M}^{ \varepsilon}(s);b \right\rangle \\
& - \int_{s}^{t} \left\langle \left[\Phi\left(\bar{u}^{ \varepsilon}(s,\xi)\right)\right]^{*}a ; \left[\Phi\left(\bar{u}^{ \varepsilon}(s,\xi)\right)\right]^{*}b \right\rangle {\mathop{}\mathrm{d}}\xi\bigg]\varphi(\bar{u}^{ \varepsilon}(t,x))\bigg\} \\
\rightarrow \mathbb{E} & \bigg\{\bigg[\left\langle \bar{M}(t);a \right\rangle \left\langle \bar{M}(t);b \right\rangle - \left\langle \bar{M}(s);a \right\rangle \left\langle \bar{M}(s);b \right\rangle \\
& - \int_{s}^{t} \left\langle \left[\Phi\left(\bar{u}(s,\xi)\right)\right]^{*}a ; \left[\Phi\left(\bar{u}(s,\xi)\right)\right]^{*}b \right\rangle {\mathop{}\mathrm{d}}\xi\bigg]\varphi(\bar{u}(t,x))\bigg\}.
\end{aligned}$$
Then $\bar{M}(t)$ is also a square integrable martingale adapted to the filtration $\sigma\left\{\bar{u}(s), 0\leq s \leq t\right\}$ with quadratic variation equal $\int_{0}^{t} \Phi(\bar{u}(s,x))\left(\Phi(\bar{u}(s,x))\right)^{*} d s$.
Substitute in the representation theorem (e.g., see Theorem 8.2 in [@dPZ]), $M_{t}:=\bar{M}(t)$, $[M_{t}]:=\int_{0}^{t} \Phi(\bar{u}(s,x))\times\left(\Phi(\bar{u}(s,x))\right)^{*} d s$ and $\Phi(s):=\Phi(\bar{u}(s,x))$.
Then there exists a process $\tilde{M}(t) = \int_{0}^{t}\Phi(\bar{u}(s,x))d W(s)$, such that $\tilde{M}(t)=\bar{M}(t)$, $\mathbb{\bar{P}}-\text{a.s.}$, and $$\nonumber
\begin{aligned}
& \bar{u}(t,x) - u_{0}(x) + \int_{0}^{t}\bigg[\bar{u}(t,x)\bar{u}_{x}(t,x) + \bar{u}_{3x}(t,x) + 3\bar{u}_{x}(t,x) \bar{u}_{2x}(t,x) + \bar{u}(t,x) \bar{u}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s \\
&= \int_{0}^{t}\Phi(\bar{u}(s,x)){\mathop{}\mathrm{d}}W(s) .
\end{aligned}$$ This implies $$\nonumber
\begin{aligned}
& \bar{u}(t,x) = u_{0}(x) - \int_{0}^{t}\bigg[\bar{u}(t,x)\bar{u}_{x}(t,x) + \bar{u}_{3x}(t,x) + 3\bar{u}_{x}(t,x) \bar{u}_{2x}(t,x) + \bar{u}(t,x) \bar{u}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s \\
& + \int_{0}^{t}\Phi(\bar{u}(s,x)){\mathop{}\mathrm{d}}W(s) , \\
\end{aligned}$$ so $\bar{u}(t,x)$ is a solution to (\[Apr\]), what finishes the proof of Theorem \[P4.1\] .
Proofs of Lemmas \[szac4.1\] and \[PropCias\] {#sec3}
=============================================
Let $p: \mathbb{R} \rightarrow \mathbb{R}$, be a smooth function fulfilling conditions
- $p$ is increasing in $\mathbb{R}$;
- $\forall_{x\in{\mathbb{R}}}$ $~p > \delta_{0} > 0$;
- $\forall_{n\in\mathbb{N}}~ \left|\frac{\partial ^{n}}{\partial x^{n}} p(x)\right|<\delta_{n}$;
- $\left(\lambda - 2\right) \delta_{2} \geq \delta_{3}$ .
Let $F(u^{\varepsilon}) := \int_{X}p(x)(u^{\varepsilon}(x))^{2}{\mathop{}\mathrm{d}}x$. Applying the Itô formula for $F(u^{\varepsilon})$, we obtain $$\begin{aligned}
{\mathop{}\mathrm{d}}F(u^{\varepsilon}(t,x)) = & \left\langle F'(u^{\varepsilon}(t,x));\Phi(u^{\varepsilon}(t,x)) \right\rangle {\mathop{}\mathrm{d}}W(t) - \left\langle F'(u^{\varepsilon}(t,x));\varepsilon u^{\varepsilon}_{4x}(t,x) + u^{\varepsilon}_{3x}(t,x) \right.\\
&\left. + u^{\varepsilon}(t,x)u^{\varepsilon}_{x}(t,x) + 3u_{x}^{\varepsilon}(t,x)u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x)\right\rangle {\mathop{}\mathrm{d}}t \\
&+ \frac{1}{2}tr\left\{F''(u^{\varepsilon}(t,x))\Phi(u^{\varepsilon}(t,x))\left[\Phi(u^{\varepsilon}(t,x))\right]^{*}\right\}{\mathop{}\mathrm{d}}t ,
\end{aligned}$$ where $$\left\langle F'(u^{\varepsilon}(t,x));v(t,x) \right\rangle \!= \! 2\!\!\int_{X} \!p(x)u^{\varepsilon}(t,x)v(t,x){\mathop{}\mathrm{d}}x \quad \mbox{and} \quad
F''(u^{\varepsilon}(t,x))v(t,x) \!= \!2p(x)v(t,x).$$
We use the following estimates from [@Deb p.242]. There exist $C_{1}, C_{2}, C_{3}$, such that $$\nonumber
\begin{aligned}
\int_{\mathbb{R}} p(x)u^{\varepsilon}(t,x)u^{\varepsilon}_{4x}(t,x) {\mathop{}\mathrm{d}}x \geq & \frac{1}{2}\int_{\mathbb{R}}p(x)\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - C_{1}\left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(\mathbb{R})} \\
&- C_{2}\int_{\mathbb{R}}p'(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x ; \\
\int_{\mathbb{R}} p(x)u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x) {\mathop{}\mathrm{d}}x \geq & \frac{3}{2}\int_{\mathbb{R}}p'(x)\left[u^{\varepsilon}_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - \frac{1}{2}\int_{\mathbb{R}}p'''(x)\left[u(t,x)\right]^{2}{\mathop{}\mathrm{d}}x ; \\
\int_{\mathbb{R}} p(x)\left[u^{\varepsilon}(t,x)\right]^{2}u^{\varepsilon}_{x}(t,x) {\mathop{}\mathrm{d}}x = & -\frac{1}{3} \int_{\mathbb{R}}p'(x)\left[u^{\varepsilon}(t,x)\right]^{3} {\mathop{}\mathrm{d}}x \\
\geq & -C_{3}\left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) - \frac{1}{2}\int_{\mathbb{R}}p'(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x .
\end{aligned}$$ Similarly as above, one has $$\begin{aligned}
\int_{\mathbb{R}} & p(x)\left[3u_{x}^{\varepsilon}(t,x)u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x)\right]
= \int_{\mathbb{R}}p''(x)u_{x}^{\varepsilon}(t,x)\left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \int_{\mathbb{R}}p'(x)\left[u_{x}^{\varepsilon}(t,x)\right]^{2}u^{\varepsilon}(t,x) {\mathop{}\mathrm{d}}x + \int_{\mathbb{R}}p(x)u^{\varepsilon}(t,x)u_{x}^{\varepsilon}(t,x)u_{2x}^{\varepsilon}(t,x) {\mathop{}\mathrm{d}}x \\
\geq & - \frac{1}{3} \int_{\mathbb{R}}p'''(x)\left[u^{\varepsilon}(t,x)\right]^{3} {\mathop{}\mathrm{d}}x - \int_{\mathbb{R}}p'(x)\left|u^{\varepsilon}(t,x)\right|\left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& - \int_{\mathbb{R}}p(x)\left|u^{\varepsilon}(t,x)\right|u_{x}^{\varepsilon}(t,x)u_{2x}^{\varepsilon}(t,x) {\mathop{}\mathrm{d}}x \\
$$$$\begin{aligned}
\geq & -C_{4}\left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) - \frac{1}{2}\int_{\mathbb{R}}p'''(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - \lambda \int_{\mathbb{R}}p'(x)\left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& - \lambda \int_{\mathbb{R}}p(x)u_{x}^{\varepsilon}(t,x)u_{2x}^{\varepsilon}(t,x) {\mathop{}\mathrm{d}}x \\
= & -C_{4}\left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) - \frac{1}{2}\int_{\mathbb{R}}p'''(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - \lambda \int_{\mathbb{R}}p'(x)\left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \frac{1}{2} \lambda \int_{\mathbb{R}}p'(x)\left(u_{x}^{\varepsilon}(t,x)\right)^{2} {\mathop{}\mathrm{d}}x.
\end{aligned}$$
In consequence we have $$\label{DebL2}
\begin{aligned}
&\left\langle F'(u^{\varepsilon}(t,x));\varepsilon u^{\varepsilon}_{4x}(t,x) + u^{\varepsilon}_{3x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{x}(t,x) + 3u_{x}^{\varepsilon}(t,x)u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x)\right\rangle \\
\geq & \varepsilon\int_{\mathbb{R}}p(x)\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - 2 \varepsilon C_{1}\int_{\mathbb{R}}\left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - 2 \varepsilon C_{2}\int_{\mathbb{R}}p'(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x \\
& + 3\int_{\mathbb{R}}p'(x)\left[u^{\varepsilon}_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - \int_{\mathbb{R}}p'''(x)\left[u(t,x)\right]^{2}{\mathop{}\mathrm{d}}x -2 C_{3}\left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) \\
& - \int_{\mathbb{R}}p'(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x -2 C_{4}\left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) - \int_{\mathbb{R}}p'''(x)\left[u_{x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x \\
& - 2 \lambda \int_{\mathbb{R}}p'(x)\left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x + \lambda \int_{\mathbb{R}}p'(x)\left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \end{aligned}$$ $$\nonumber
\begin{aligned}
= & \varepsilon\int_{\mathbb{R}}p(x)\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x + \int_{\mathbb{R}} \left[-2\varepsilon C_{2} p'(x) + 3p'(x) - p'(x) - p'''(x) -\lambda p'(x)\right] \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \int_{\mathbb{R}} \left[-2 \varepsilon C_{1} - p'''(x) \right] \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - C_{5} \left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) \\
= & \varepsilon\int_{\mathbb{R}}p(x)\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x + \int_{\mathbb{R}} \left[\left(-2\varepsilon C_{2} - \lambda + 2\right) p'(x) - p'''(x) \right] \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \int_{\mathbb{R}} \left[-2 \varepsilon C_{1} - p'''(x) \right] \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - C_{5} \left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) \\
\geq & \varepsilon\int_{\mathbb{R}}p(x)\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x + \int_{\mathbb{R}} \left[2\delta_{1}\varepsilon C_{2} +\delta_{1}\left( \lambda - 2\right) - \delta_{2} \right] \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \int_{\mathbb{R}} \left[-2 \varepsilon C_{1} - p'''(x) \right] \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - C_{5} \left(1+\left|u^{\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}^{6}\right) \\
\geq & \varepsilon\delta\int_{\mathbb{R}}\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x + 2\delta_{1}\varepsilon C_{2} \int_{\mathbb{R}} \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - \left[2 \varepsilon C_{1} + \delta_{3} \right] \int_{\mathbb{R}} \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x - C_{5} \left(1+\lambda^{6}\right).
\end{aligned}$$
Let $\left\{e_{1}\right\}_{i\in\mathbb{N}}$ be an orthonormal basis in $L^{2}(\mathbb{R})$. Then there exists a constant $C_{4}>0$, such that $$\label{DebL3}
\begin{aligned}
\text{Tr}\left(F''(u)\Phi(u)\left[\Phi(u)\right]^{*}\right) =& 2\sum_{i\in\mathbb{N}} \int_{\mathbb{R}}p(x)\left|\Phi\left(u^{\varepsilon}(t,x)\right)e_{i}(x)\right|^{2} {\mathop{}\mathrm{d}}x \leq C_{4}\left|\Phi\left(u^{\varepsilon}(t,x)\right)\right|^{2}_{L_{0}^{2}\left(L^{2}(X)\right)} \\
\leq & C_{6}\left(\kappa_{1}\left|u^{\varepsilon}(t,x)\right|_{L^{2}(X)}^{2}+\kappa_{2}\right)^{2}.
\end{aligned}$$ Due to (\[DebL2\]) and (\[DebL3\]) we have $$\nonumber
\begin{aligned}
\mathbb{E}F(u^{\varepsilon}(t,x)) \leq & F\left(u^{\varepsilon}_{0}\right) - \varepsilon\delta\mathbb{E}\int_{0}^{t}\int_{\mathbb{R}}\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x - 2\delta_{1}\varepsilon C_{2} \mathbb{E}\int_{0}^{t}\int_{\mathbb{R}} \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x \\
& + \left[2 \varepsilon C_{1} + \delta_{3} \right] \mathbb{E}\int_{0}^{t}\int_{\mathbb{R}} \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x + tC_{5} \left(1+\lambda^{6}\right) \\
& + C_{6}\mathbb{E}\int_{0}^{t}\left(\kappa_{1}\left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(\mathbb{R})}+\kappa_{2}\right)^{2}{\mathop{}\mathrm{d}}t, \\
\end{aligned}$$ so, $$\nonumber
\begin{aligned}
&\mathbb{E}F(u^{\varepsilon}(t,x)) + \varepsilon\delta\mathbb{E}\int_{0}^{t}\int_{\mathbb{R}}\left[u^{\varepsilon}_{2x}(t,x)\right]^{2}{\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t + 2\delta_{1}\varepsilon C_{2} \mathbb{E}\int_{0}^{t}\int_{\mathbb{R}} \left[u_{x}^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t \\
\leq & F\left(u^{\varepsilon}_{0}\right) + \left[2 \varepsilon C_{1} + \delta_{3} \right] \mathbb{E}\int_{0}^{t}\int_{\mathbb{R}} \left[u^{\varepsilon}(t,x)\right]^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t + C_{5} t\left(1+\lambda^{6}\right) \\
& + C_{6}\mathbb{E}\int_{0}^{t}\left(\kappa_{1}\left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(\mathbb{R})}+\kappa_{2}\right)^{2}{\mathop{}\mathrm{d}}t \\
= & F\left(u^{\varepsilon}_{0}\right) + \left[2 \varepsilon C_{1} + \delta_{3} \right] t\lambda^{2} + C_{5} t\left(1+\lambda^{6}\right) + C_{6}t\left(\kappa_{1}\lambda^{2}+\kappa_{2}\right)^{2} \\
\leq & F\left(u^{\varepsilon}_{0}\right) + \left[2 \varepsilon C_{1} + \delta_{3} \right] T\lambda^{2} + C_{5} T\left(1+\lambda^{6}\right) + C_{6}T\left(\kappa_{1}\lambda^{2}+\kappa_{2}\right)^{2} \leq \varepsilon C_{7} + C_{8}.
\end{aligned}$$
Let $\varepsilon_{0}>0$ be fixed. Then for all $0<\varepsilon<\varepsilon_{0}$ one has $$\nonumber
\begin{aligned}
\varepsilon\, \mathbb{E} &\left( \left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{2}(X))} \right) = \varepsilon \,\mathbb{E} \! \int_{0}^{T} \!\! \int_{\mathbb{R}} \! \left[u^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t + \varepsilon \mathbb{E} \! \int_{0}^{T} \!\! \int_{\mathbb{R}} \! \left[u_{2x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t \\ &
\leq \varepsilon\, T \lambda^{2} + \varepsilon\, \mathbb{E} \! \int_{0}^{T} \!\!\int_{\mathbb{R}} \!\left[u_{2x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t
= \varepsilon\, T \lambda^{2} + \varepsilon\, \mathbb{E} \! \int_{0}^{T} \!\! \int_{\mathbb{R}} \!\frac{1}{\delta} \delta \left[u_{2x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t \\ &
\leq \varepsilon\, T \lambda^{2} + \frac{1}{\delta} \varepsilon\, \delta\, \mathbb{E} \! \int_{0}^{T} \!\! \int_{\mathbb{R}} \! \left[u_{2x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x
\leq \varepsilon \,T \lambda^{2} + \frac{1}{\delta}\, \varepsilon\left(\varepsilon C_{7}(T) + C_{8}(T)\right)\\ &
\leq \varepsilon_{0} T \lambda^{2} + \frac{\varepsilon_{0}^{2} C_{7}(T) + \varepsilon_{0}C_{8}(T)}{\delta},
\end{aligned}$$ what proves (\[4.1a\]). Moreover one has $$\nonumber
\begin{aligned}
\mathbb{E} & \left( \left|u^{\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{1}(-k,k))} \right) = \mathbb{E} \int_{0}^{T} \int_{-k}^{k} \left[u^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t + \mathbb{E} \int_{0}^{T} \int_{-k}^{k} \left[u_{x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x {\mathop{}\mathrm{d}}t \\ &
\leq \varepsilon\, T \lambda^{2} + \mathbb{E} \int_{0}^{T} \int_{-k}^{k} \left[u_{x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x
\leq \varepsilon\, T \lambda^{2} + \mathbb{E} \int_{0}^{T} \int_{\mathbb{R}} \left[u_{x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x \\ &
\leq \varepsilon T \lambda^{2} + \frac{1}{2\delta_{1}\varepsilon C_{2}} 2\delta_{1}\varepsilon C_{2} \mathbb{E} \int_{0}^{T} \int_{\mathbb{R}} \left[u_{x}^{\varepsilon}(t,x)\right] ^{2} {\mathop{}\mathrm{d}}x \\ &
\leq \varepsilon T \lambda^{2} + \frac{1}{2\delta_{1}\varepsilon C_{2}} \left(\varepsilon C_{7}(T) + C_{8}(T)\right) \leq \varepsilon_{0} T \lambda^{2} + \frac{\varepsilon_{0} C_{7}(T) + C_{8}(T)}{2\delta_{1}\varepsilon_{0} C_{2}},
\end{aligned}$$ what proves inequality (\[4.1c\]).
Let $k\in \mathbb{N}$ be arbitrary fixed and let $0<\varepsilon<\varepsilon_{0}$. Then $$\label{dec}
\begin{aligned}
u^{\varepsilon}(t,x) = & u_{0}^{\varepsilon}(x) - \int_{0}^{t} \bigg[\varepsilon u^{\varepsilon}_{4x}(t,x) + u^{\varepsilon}_{3x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{x}(t,x) \\
& + 3u^{\varepsilon}_{x}(t,x)u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x)u^{\varepsilon}_{3x}(t,x) \bigg] {\mathop{}\mathrm{d}}s + \int_{0}^{t} \left( \Phi(u^{\varepsilon}(s,x)) \right) {\mathop{}\mathrm{d}}W(s) .
\end{aligned}$$
Denote $$\nonumber
\begin{aligned}
J_{1}:= & u_{0}^{\varepsilon}(x) ; \quad
J_{2}:= - \varepsilon\int_{0}^{t} u^{\varepsilon}_{4x}(t,x) {\mathop{}\mathrm{d}}s ; \quad
J_{3}:= - \int_{0}^{t} u^{\varepsilon}(s,x) u^{\varepsilon}_{x}(s,x) {\mathop{}\mathrm{d}}s ; \\
J_{4}:= & - \int_{0}^{t} u^{\varepsilon}_{3x}(t,x) {\mathop{}\mathrm{d}}s ; \quad
J_{5}:= - \left(3\int_{0}^{t}\!\!u^{\varepsilon}_{x}(s,x) u^{\varepsilon}_{2x}(t,x) {\mathop{}\mathrm{d}}s + \int_{0}^{t}\!\! u^{\varepsilon}(t,x) u^{\varepsilon}_{3x}(t,x) {\mathop{}\mathrm{d}}s\right); \\
J_{6}:= & \int_{0}^{t} \left( \Phi(u^{\varepsilon}(s,x)) \right) {\mathop{}\mathrm{d}}W(s) .
\end{aligned}$$ There exists a constant $C_{1}>0$, that $\mathbb{E} \left|J_{1}\right|^{2}_{W^{1,2}(0,T,H^{-2}(-k,k))} = C_{1}$.\
There exists a constant $C_{2} > 0$, such that $$\nonumber
\left| - \varepsilon u^{\varepsilon}_{4x}(t,x) \right|_{H^{-2}(-k,k)} = \varepsilon \left| u^{\varepsilon}_{4x}(t,x) \right|_{H^{-2}(-k,k)}
\leq C_{2}\varepsilon\left| u^{\varepsilon}(s,x) \right|_{H^{2}(-k,k)}.$$ Therefore, due to Lemma \[szac4.1\], we can write $$\nonumber
\begin{aligned}
\mathbb{E}& \left| - \varepsilon u^{\varepsilon}_{4x}(t,x) \right|^{2}_{L^{2}(0,T;H^{-2}(-k,k))} = \mathbb{E} \int_{0}^{T} \left| - \varepsilon u^{\varepsilon}_{4x}(t,x) \right|^{2}_{H^{-2}(-k,k)} {\mathop{}\mathrm{d}}s \\ &
\leq C_{2}^{2}\varepsilon^{2} \mathbb{E} \int_{0}^{T} \left| u^{\varepsilon}(s,x) \right|^{2}_{H^{2}(-k,k)} {\mathop{}\mathrm{d}}s \leq C_{3}(k), \mbox{~where~} C_{3}(k)>0.
\end{aligned}$$ So, there exists a constant $C_{4}(k)>0$, such that $$\nonumber
\begin{aligned}
\mathbb{E} \left|J_{2}\right|^{2}_{W^{1,2}(0,T,H^{-2}(-k,k))} \leq C_{4}(k) .
\end{aligned}$$
Now, we use the result from [@Deb p.243]. There exists a constant $C_{5}(k)>0$, that the following inequality holds $$\left|u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x)\right|_{H^{-1}(-k,k)} \leq C_{5}(k)\left|u^{\varepsilon}(s,x)\right|^{\frac{3}{2}}_{L^{2}(-k,k)}\left|u^{\varepsilon}(s,x)\right|^{\frac{1}{2}}_{H^{1}(-k,k)}.$$ This estimate implies the existence of a constant $C_{8}(k)>0$, such that $$\nonumber
\begin{aligned}
& \left| - u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|_{H^{-2}(-k,k)} = \left| u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|_{H^{-2}(-k,k)} \\ &
\leq C_{6} \left| u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|_{H^{-1}(-k,k)}
\leq C_{7}(k) \left|u^{\varepsilon}(s,x)\right|^{\frac{3}{2}}_{L^{2}(-k,k)}\left|u^{\varepsilon}(s,x)\right|^{\frac{1}{2}}_{H^{1}(-k,k)} \\ &
\leq C_{7}(k) \left|u^{\varepsilon}(s,x)\right|_{L^{2}(-k,k)}\left|u^{\varepsilon}(s,x)\right|^{\frac{1}{2}}_{L^{2}(-k,k)}\left|u^{\varepsilon}(s,x)\right|^{\frac{1}{2}}_{H^{1}(-k,k)} \\ &
\leq C_{7}(k) \left[\left(2k\lambda^{2}\right)^{\frac{1}{2}}\right]\left|u^{\varepsilon}(s,x)\right|^{\frac{1}{2}}_{H^{1}(-k,k)}
\leq C_{8}(k) \lambda\left|u^{\varepsilon}(s,x)\right|_{H^{1}(-k,k)} .
\end{aligned}$$
Due to Lemma \[szac4.1\] there exists a constant $C_{9}(k)>0$, that we can write $$\nonumber
\begin{aligned}
\mathbb{E}& \left|- u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|^{2}_{L^{2}(0,T;H^{-2}(-k,k))} = \mathbb{E} \int_{0}^{T} \left| - u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|^{2}_{H^{-2}(-k,k)} {\mathop{}\mathrm{d}}s \\ &
\leq C_{8}^{2}(k) \lambda^{2} \mathbb{E} \int_{0}^{T} \left|u^{\varepsilon}(s,x)\right|^{2}_{H^{1}(-k,k)} {\mathop{}\mathrm{d}}s
= C_{8}^{2}(k) \lambda^{2} \mathbb{E} \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T;H^{1}(-k,k))} \leq C_{9}(k) \lambda^{2} .
\end{aligned}$$
Then, there exists a constant $C_{10}(k)>0$, such that $$\mathbb{E} \left|J_{3}\right|^{2}_{W^{1,2}(0,T,H^{-2}(-k,k))} \leq C_{10}(k).$$ We have $$\nonumber
\begin{aligned}
\left| - u^{\varepsilon}_{3x}(t,x)\right|_{H^{-2}(-k,k)} = & \left| u^{\varepsilon}_{3x}(t,x) \right|_{H^{-2}(-k,k)}
\leq C_{11}\left| u^{\varepsilon}(s,x) \right|_{H^{1}(-k,k)} \\
\leq & C_{12}\left| u^{\varepsilon}(s,x) \right|_{H^{2}(-k,k)}, \mbox{~where~} C_{12}>0.
\end{aligned}$$ Lemma \[szac4.1\] implies the existence of a constant $C_{13}>0$, such that $$\nonumber
\begin{aligned}
\mathbb{E} & \left|- u^{\varepsilon}_{3x}(t,x) \right|^{2}_{L^{2}(0,T;H^{-2}(-k,k))} = \mathbb{E} \! \int_{0}^{T}\!\!\! \left| - u^{\varepsilon}_{3x}(t,x) \right|^{2}_{H^{-2}(-k,k)} {\mathop{}\mathrm{d}}s \leq
C^{2}_{12} \,\mathbb{E} \!\int_{0}^{T}\!\!\! \left|u^{\varepsilon}(s,x)\right|^{2}_{H^{2}(-k,k)} {\mathop{}\mathrm{d}}s \\ &
= C^{2}_{12} \,\mathbb{E} \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T;H^{2}(-k,k))} \leq C^{2}_{12} \,\mathbb{E} \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T;H^{2}(\mathbb{R}))} \leq C_{13} .
\end{aligned}$$ So, there exists a constant $C_{14}>0$, such that $~\mathbb{E} \left|J_{4}\right|^{2}_{W^{1,2}(0,T,H^{-2}(-k,k))} \leq C_{14}$. There exist constants $C_{15}, C_{16}(k) > 0$, such that $$\nonumber
\begin{aligned}
\left| \right. - & \left.\left(3 u^{\varepsilon}_{x}(s,x) u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x) u^{\varepsilon}_{3x}(t,x) \right)\right|_{H^{-2}(-k,k)}\\\leq & C_{15}\left| u^{\varepsilon}(s,x)u^{\varepsilon}_{x}(s,x) \right|_{L^{2}(-k,k)}
\leq C_{16}(k) \lambda^{2}\left|u^{\varepsilon}(s,x)\right|_{H^{1}(-k,k)}.
\end{aligned}$$ Due to Lemma \[szac4.1\] there exists a constant $C_{17}(k)>0$, such that $$\nonumber
\begin{aligned}
\mathbb{E} & \left| - \left(3 u^{\varepsilon}_{x}(s,x) u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x) u^{\varepsilon}_{3x}(t,x) \right) \right|^{2}_{L^{2}(0,T;H^{-3}(-k,k))} \\ &
= \mathbb{E} \int_{0}^{T} \left| - \left(3 u^{\varepsilon}_{x}(s,x) u^{\varepsilon}_{2x}(t,x) + u^{\varepsilon}(t,x) u^{\varepsilon}_{3x}(t,x) \right) \right|^{2}_{H^{-3}(-k,k)} {\mathop{}\mathrm{d}}s \\ &
\leq C_{16}^{2}(k) \lambda^{4} \mathbb{E} \! \int_{0}^{T}\!\!\! \left|u^{\varepsilon}(s,x)\right|^{2}_{H^{1}(-k,k)} {\mathop{}\mathrm{d}}s
= C_{16}^{2}(k) \lambda^{4} \mathbb{E} \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T;H^{1}(-k,k))} \leq C_{17}(k) \lambda^{4} .
\end{aligned}$$ So, there exists a constant $C_{18}(k)>0$, such that $~\mathbb{E} \left|J_{5}\right|^{2}_{W^{1,2}(0,T,H^{-3}(-k,k))} \leq C_{18}(k)$. Substitute in [@Gat Lemma 2.1] $f(s) := \Phi(u(s,x))$, $K=H=L^{2}(\mathbb{R})$. Then $\mathscr{I}(f)(t) = \int_{0}^{t}\Phi(u(s,x)) {\mathop{}\mathrm{d}}W(s)$ and for all $p\geq 1$ and $\alpha<\frac{1}{2}$ there exists a constant $C_{22}(p,\alpha)>0$, such that $$\nonumber
\begin{aligned}
\mathbb{E}\left|\int_{0}^{t}\Phi(u^{m}(s,x)) {\mathop{}\mathrm{d}}W(s)\right|^{2p}_{W^{\alpha ( p),2p}(0,T;L^{2}(\mathbb{R}))} \leq & C_{22}(2p,\alpha) \mathbb{E} \left( \int_{0}^{T} \left|\Phi(u^{m}(s,x))\right|^{2p}_{L_{2}^{0}(L^{2}(\mathbb{R}))} {\mathop{}\mathrm{d}}s \right).
\end{aligned}$$ Then, due to condition (\[W1\]), there exists a constant $C_{23}>0$, that $$\nonumber
\begin{aligned}
\mathbb{E}\left|\int_{0}^{t}\Phi(u^{m}(s,x)) {\mathop{}\mathrm{d}}W(s)\right|^{2p}_{W^{\alpha,2p}(0,T;L^{2}(\mathbb{R}))} \leq & C_{23} (p,\alpha) .
\end{aligned}$$ Substitution in the above inequality $p:=1$ yields $$\label{It\^o3}
\mathbb{E}\left|J_{6}\right|^{2}_{W^{\alpha,2}(0,T;L^{2}(\mathbb{R}))} = \mathbb{E}\left|\int_{0}^{t}\Phi(u(s,x)) {\mathop{}\mathrm{d}}W(s)\right|^{2}_{W^{\alpha,2}(0,T;L^{2}(\mathbb{R}))} \leq C_{23}(2,\alpha) = C_{24}(\alpha).$$ Let $\beta\in\left(0,\frac{1}{2}\right)$ and $\alpha\in\left(\beta + \frac{1}{2}, \infty\right)$ be arbitrary fixed. Note, that the following inclusion relations hold $$\hspace{2.5ex} W^{\alpha,2}(0,T;L^{2}(\mathbb{R})) \subset W^{\alpha,2}(0,T;H^{-2}([-k,k));$$ $$\mbox{and} \hspace{3ex} W^{1,2}(0,T,H^{-2}(-k,k)) \subset W^{\alpha,2}(0,T,H^{-2}(-k,k)).$$ Then, there exists a constant $C_{25}(\alpha) > 0$, such that $$\nonumber
\begin{aligned}
\mathbb{E} & \left|u^{m}(s,x)\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))}^{2} = \mathbb{E}\left|\sum_{i=1}^{6} J_{i}\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))}^{2} \leq \mathbb{E} \left( \sum_{i=1}^{6} \left|J_{i}\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} \right)^{2} \\
= & \mathbb{E} \left[ \sum_{i=1}^{6} \left|J_{i}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} + 2\sum_{i=1}^{6} \sum_{j=i+1}^{6} \left|J_{i}\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))}\left|J_{j}\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} \right]\\
\leq & \mathbb{E} \left[ \sum_{i=1}^{6} \left|J_{i}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} + 2\sum_{i=1}^{6} \sum_{j=i+1}^{6} \left(\left|J_{i}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} + \left|J_{j}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))}\right) \right]\\
= & \mathbb{E} \left[ 8 \sum_{i=1}^{6} \left|J_{i}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))}\right]
= 8 \sum_{i=1}^{6} \left[ \mathbb{E} \left|J_{i}\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} \right]
\leq C_{25}(\alpha) .
\end{aligned}$$ Moreover $$\hspace{7ex} W^{\alpha,2}(0,T,H^{-2}(-k,k)) \subset C^{\beta}(0,T;H^{-3}_{loc}(-k,k);$$ $$\mbox{and} \hspace{3ex} W^{\alpha,2}(0,T,H^{-2}(\mathbb{R})) \subset W^{\alpha,2}(0,T,H^{-2}(-k,k)).$$
So, there exist constants $C_{27}(k), C_{28}(k, \alpha) >0$, such that $$\label{Cszac}
\begin{aligned}
& \mathbb{E}\left|u^{\varepsilon}(s,x)\right|_{C^{\beta}(0,T;H^{-3}(-k,k)}^{2} \leq C_{26} \mathbb{E}\left|u^{\varepsilon}(s,x)\right|_{W^{\alpha,2}(0,T,H^{-3}(-k,k))}^{2} \leq C_{27}(k,\alpha) \\
& \mathbb{E}\left|u^{\varepsilon}(s,x)\right|_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} \leq C_{28}(k, \alpha).
\end{aligned}$$
Let $\eta>0$ be arbitrary fixed. Due to Lemma \[szac4.1\] there exists a constant $C_{30}(k)>0$, that $$\label{rwnszac}
\begin{aligned}
\mathbb{E}\left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T,H^{-1}(-k,k))} \leq & C_{29}(k)\mathbb{E}\left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T,H^{-1}(\mathbb{R}))} \tilde{C}_{2} = C_{30}(k).
\end{aligned}$$
Substituting in [@Deb Lemma 2.1] $\alpha_{k}:=\eta^{-1}2^{k} \left( C_{30}(k) + C_{27}(k,\alpha) + C_{28}(k,\alpha) \right)$ and using Markov inequality [@Pap p. 114] for $$X := \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T,H^{-1}(-k,k))} + \left|u^{\varepsilon}(s,x)\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} + \left|u^{\varepsilon}(s,x)\right|_{C^{\beta}(0,T;H^{-3}_{loc}(-k,k)}^{2}$$ and $ \varepsilon := \eta^{-1}2^{k} \left( C_{30}(k) + C_{27}(k,\alpha) + C_{28}(k,\alpha) \right)$, we obtain $$\nonumber
\begin{aligned}
&\mathbb{P} \Big(u^{\varepsilon} \in A\left(\left\{\alpha _{k} \right\} \right) \Big) =
1 - \mathbb{P} \Big( \left|u^{\varepsilon}(s,x)\right|^{2}_{L^{2}(0,T,H^{-1}(-k,k))} + \left|u^{\varepsilon}(s,x)\right|^{2}_{W^{\alpha,2}(0,T,H^{-2}(-k,k))} \\
& ~+ \left|u^{\varepsilon}(s,x)\right|_{C^{\beta}(0,T;H^{-3}_{loc}(-k,k))}^{2} \geq \eta^{-1}2^{k} \left( C_{30}(k) + C_{27}(k,\alpha) + C_{28}(k,\alpha) \right) \Big) \\
= &\hspace{2ex} 1 - \frac{C_{30}(k) + C_{27}(k,\alpha) + C_{28}(k,\alpha)}{\eta^{-1}2^{k} \left( C_{30}(k) + C_{27}(k,\alpha) + C_{28}(k,\alpha)\right)}
= 1 - \frac{\eta}{2^{k}} > 1 - \eta .
\end{aligned}$$
Let $K$ be the following mapping for $\eta>0$: $K\left( \eta \right) = A\left(\left\{a_{k}^{(\eta)}\right\}\right)$, where $\left\{a_{k}^{(\eta)}\right\}$ is an increasing sequence of positive numbers, which can, but does not have to, depend on $\eta$. Note, that due to [@Deb Lemma 2.1], the set $K(\eta)$ is compact for all $\eta>0$. Moreover, $\mathbb{P}\left\{K\left( \eta \right)\right\} > 1-\eta$, then the family $\mathscr{L}\left(u^{\varepsilon}\right)$ is tight.
Proof of Lemma \[parMart\] {#sec4}
==========================
Let $\left\{e_{i}\right\}_{i\in\mathbb{N}}$ be an orthonormal basis in space $L^{2}(\mathbb{R})$. Denote by $P_{m}$, for all $m\in\mathbb{N}$, the orthogonal projection on $Sp(e_{0},...,e_{m})$. Consider finite dimensional approximation of the problem (\[par\]) in the space $P_{m}L^{2}(\mathbb{R})$ of the form $$\label{Galerkin}
\begin{cases}
{\mathop{}\mathrm{d}}u^{m,\varepsilon}(t,x) + \left[\varepsilon \theta\left(\frac{\left|u^{m,\varepsilon}_{4x}(t,x)\right|^{2}}{m}\right) u^{m,\varepsilon}_{4x}(t,x) + \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)\right|^{2}}{m} \right) u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x)\right. \\ \hspace{14ex}
+ \theta\left(\frac{\left|u^{m,\varepsilon}_{3x}(t,x)\right|^{2}}{m} \right)u^{m,\varepsilon}_{3x}(t,x) + 3 \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|^{2}}{m} \right)u^{m,\varepsilon}_{2x}(t,x) \\ \left. \hspace{14ex}
+ \theta\left(\frac{\left|u^{m,\varepsilon}_{3x}(t,x)\right|^{2}}{m} \right)u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{3x}(t,x) \right] {\mathop{}\mathrm{d}}t
= P_{m}\Phi\left(u^{m,\varepsilon}(t,x)\right){\mathop{}\mathrm{d}}W^{m}(t) \\
u^{m,\varepsilon}_{0}(x) = P_{m}u^{\varepsilon}(0,x) ,
\end{cases}$$ where $\theta\in C^{\infty}(\mathbb{R})$ fulfils conditions $$\begin{cases}
\theta(\xi) = 1, \quad &\textrm{when} \quad \xi\in [0,1] \\
\theta(\xi) \in [0,1], \quad &\textrm{when} \quad \xi\in (1,2) \\
\theta(\xi) = 0, \quad &\textrm{when} \quad \xi\in \left.[2,\infty)\right. .
\end{cases}$$
Let $m\in\mathbb{N}$ be arbitrary fixed and $$\nonumber
\begin{aligned}
b(u(t,x)) := & \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)\right|^{2}}{m} \right) u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x) + \theta\left(\frac{\left|u^{m,\varepsilon}_{3x}(t,x)\right|^{2}}{m} \right) u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{3x}(t,x) \\
& + 3 \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|^{2}}{m} \right)u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x) , \\
\sigma (u(t,x)) := & \Phi(u^{m,\varepsilon}(t,x)).
\end{aligned}$$ Then $$\nonumber
\begin{aligned}
\left| b(u(t,x)) \right|_{L^{2}(\mathbb{R})} \leq & \left| \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)\right|^{2}}{m} \right) u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x) \right|_{L^{2}(\mathbb{R})} \\
& + \left| \theta\left(\frac{\left|u^{m,\varepsilon}_{3x}(t,x)\right|^{2}}{m} \right) u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{3x}(t,x) \right|_{L^{2}(\mathbb{R})} \\
& + 3 \left| \theta\left(\frac{\left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|_{L^{2}(\mathbb{R})}^{2}}{m} \right)u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x) \right|_{L^{2}(\mathbb{R})} \\
= : & J_{1} + J_{2} + 3 J_{3}.
\end{aligned}$$ Note, that $$\nonumber
J_{1} =
\begin{cases}
0, \quad \mbox{when} \quad \left|u^{m,\varepsilon}_{x}(t,x)\right| \geq \sqrt{2m} \\
\lambda \left| u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x) \right|_{L^{2}(\mathbb{R})}, \quad \mbox{when} \quad \left|u^{m,\varepsilon}_{x}(t,x)\right| \leq \sqrt{2m}
\end{cases}$$ where $\lambda \in [0,1]$, therefore $$\nonumber
\begin{aligned}
J_{1} \leq \left| u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x) \right|_{L^{2}(\mathbb{R})} \leq \sqrt{2m} \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} .
\end{aligned}$$ Analogously, $$\nonumber
\begin{aligned}
J_{2} \leq \left| u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{3x}(t,x) \right|_{L^{2}(\mathbb{R})} \leq \sqrt{2m} \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} .
\end{aligned}$$ Moreover $$\nonumber
J_{3} =
\begin{cases}
0, \quad \mbox{when} \quad \left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|_{L^{2}(\mathbb{R})}^{2} \geq \sqrt{2m} \\
\lambda \left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|^{2}, \quad \mbox{when} \quad \left|u^{m,\varepsilon}_{x}(t,x)u^{m,\varepsilon}_{2x}(t,x)\right|_{L^{2}(\mathbb{R})}^{2} \leq \sqrt{2m}
\end{cases}$$ where $\lambda \in [0,1]$, so $$\nonumber
\begin{aligned}
3J_{3} \leq 3 \left| u^{m,\varepsilon}(t,x)u^{m,\varepsilon}_{x}(t,x) \right|_{L^{2}(\mathbb{R})} \leq 3 \sqrt{2m} .
\end{aligned}$$ Finally, $$\nonumber
\begin{aligned}
& \left| b( u^{m,\varepsilon}(t,x)) \right|_{L^{2}(\mathbb{R})} \leq 2\sqrt{2m} \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} + 3 \sqrt{2m} .
\end{aligned}$$ Additionally, due to the condition (\[W1\]), there exist constants $\kappa_{1}, \kappa_{2} > 0$, such that $$\left\|\Phi(u^{m,\varepsilon}(t,x))\right\|_{L_{0}^{2}(L^{2}(\mathbb{R}))} \leq \kappa_{1} \left|u^{m,\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})} + \kappa_{2} ,$$ then $$\nonumber
\begin{aligned}
& \left| b(u^{m,\varepsilon}(t,x)) \right|_{L^{2}(\mathbb{R})} + \left\|\sigma(u^{m,\varepsilon}(t,x))\right\|_{L_{0}^{2}(L^{2}(\mathbb{R}))} \\
\leq & 2\sqrt{2m} \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} + 3 \sqrt{2m} + \kappa_{1} \left|u^{m,\varepsilon}(t,x)\right|_{L^{2}(\mathbb{R})}+ \kappa_{2} \\
= & \left( 2\sqrt{2m} + \kappa_{1} \right) \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} + 3 \sqrt{2m} + \kappa_{2} \\
\leq & \left( 3\sqrt{2m} + \max \left\{ \kappa_{1}, \kappa_{2} \right\} \right) \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} + 3\sqrt{2m} + \max \left\{ \kappa_{1}, \kappa_{2} \right\} \\
= & \left( 3\sqrt{2m} + \max \left\{ \kappa_{1}, \kappa_{2} \right\} \right) \left( \left| u^{m,\varepsilon}(t,x) \right|_{L^{2}(\mathbb{R})} + 1\right).
\end{aligned}$$
Therefore, from [@Kar Prop. 3.6 and 4.6], when $b(u(t,x))$ and $\sigma(u(t,x))$ are as above, for all $m\in\mathbb{N}$, there exists a martingale solution to (\[Galerkin\]). Moreover, applying the same methods as in section \[sec3\] one can show that for all $m$ the following inequalities hold $$\begin{aligned}
\exists_{C_{1}(\varepsilon)>0}\mathbb{E}\left( \left|u^{m,\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{2}(\mathbb{R}))} \right) & \leq \tilde{C}_{1}(\varepsilon), \\
\forall_{k\in X_{k}}\exists_{C_{2}(k,\varepsilon)>0} \mathbb{E}\left( \left|u^{m,\varepsilon}(t,x)\right|^{2}_{L^{2}(0,T;H^{1}(-k,k))} \right) & \leq \tilde{C}_{2}(k,\varepsilon);\end{aligned}$$ and the family of distributions $\mathscr{L}(u^{m,\varepsilon})$ is tight in $L^{2}(0,T;L^{2}_{loc})\cap C(0,T;H^{-3}_{loc}(\mathbb{R}))$. Then application of the same methods, as used already on pages –, leads to the proof of the existence of martingale solution to (\[par\]).
[99]{}
Billingsley, P.: *Convergence of Probability Measures*, John Wiley & Sons, New York 1999.
Burde, G.I. and Sergyeyev, A.: Ordering of two small parameters in the shallow water wave problem. J. Phys. A: Math. Theor., **46**, 075501 (2013).
Da Prato, G. and Zabczyk, J.: *Stochastic equations in infinite dimensions* “Encyclopedia of Mathematics and its Application,” Cambridge University Press, Cambridge 1992.
de Bouard, A. and Debussche A.: On the Stochastic Korteweg - de Vries Equation, J. Funct. Anal., **154**, 215-251, (1998).
Drazin, P. G. and Johnson, R.S.: *Solitons: An Introduction*. Cambridge University Press, Cambridge 1989.
Flandoli, F. and Gątarek, D.: Martingale and stationary solutions for stochastic Navier-Stokes equations, Probability Theory and Related Fields **102**, 367-391, (1995).
Gawarecki, L. and Mandrekar, V.: *Stochastic differential equations in infinite dimensions*, Springer, New York 2011.
Infeld, E. and Rowlands, G.: *Nonlinear Waves, Solitons and Chaos. 2nd edition*. Cambridge University Press, Cambridge (2000).
Karatzas, I. and Shreve, S.E.: *Brownian Motion and Stochastic Calculus, 2nd ed*. Springer, New York, Berlin (1991).
Karczewska, A., Rozmej, P. and Infeld, E.: Shallow-water soliton dynamics beyond the Korteweg - de Vries equation. Phys. Rev. E, **90**, 012907 (2014).
Karczewska, A., Rozmej, P. and Infeld, E.: Energy invariant for shallow-water waves and the Korteweg - de Vries equation: Doubts about the invariance of energy. Phys. Rev. E, **92**, 053202 (2015).
Karczewska A., Rozmej, P., Infeld, E. and Rowlands, G.: Adiabatic invariants of the extended KdV equation. Phys. Lett. A, **381**, 270–275 (2017).
Karczewska, A., Rozmej, P. and Rutkowski, L .: A new nonlinear equation in the shallow water wave problem. Physica Scripta **89**, 054026 (2014).
Karczewska, A. and Szczeciński, M.: The existence of mild solution to stochastic extended KdV equation, (2018), submitted.
Korteweg, D. J. and de Vries, F.: On the change of form of long waves advancing in a rectangular canal, and on a new type of long stationary waves. Philosophical Magazine **39**, 422–443 (1895).
Marchant, T.R. and Smyth, N.F.: The extended Korteweg–de Vries equation and the resonant flow of a fluid over topography. J. Fluid Mech., **221**, 263-288 (1990).
Marchant, T.R. and Smyth, N.F.: Soliton interaction for the extended Korteweg-de Vries equation. IMA J. Appl. Math., **56**, 157-176 (1996).
Newell, A. C.: *Solitons in Mathematics and Physics*. Society for Industrial and Applied Mathematics, Philadelphia, PAS 1985.
Papoulis, A.: *Probability, Random Variables, and Stochastic Processes, 3rd ed.*, McGraw-Hill, New York 1991.
Remoissenet, M.: *Waves Called Solitons: Concepts and Experiments*. Springer, New York, Berlin 1999.
Rozmej, P. and Karczewska, A.: New exact superposition solutions to KdV2 equation. Adv. Math. Phys. **2018**, ID 5095482 (2018).
Rozmej, P., Karczewska, A. and Infeld, E.: Superposition solutions to the extended KdV equation for water surface waves. Nonlinear Dyn. **91** (2), 1085–1093 (2018).
G.B. Whitham G. B.: *Linear and Nonlinear Waves*. Wiley, New York 1974.
Yang, J.: Dynamics of embedded solitons in the extended Korteweg-de Vries equations. Studies in Applied Mathematics **106**, 337–365 (2001).
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The triplet superconductivity in $UGe_2$ and $URhGe$ coexists with itinerant ferromagnetism such that in the pressure-temperature phase diagram the whole region occupied by the superconducting state is situated inside a more vast ferromagnetic region. In the same family metal $UCoGe$ the pressure dependent critical lines $T_{Curie}(P)$ and $T_{sc}(P)$ of the ferromagnet and the superconducting phase transitions intersect each other. The two-band multidomain superconducting ferromagnet state arises at temperatures below both of these lines. Here I describe the symmetry and the order parameters of the paramagnet as well of the multidomain ferromagnet superconducting states. The Josephson coupling between two adjacent ferromagnet superconducting domains is discussed.'
author:
- 'V.P.Mineev'
bibliography:
- 'sample.bib'
title: Coexistence of triplet superconductivity and itinerant ferromagnetism
---
[address=[Commissariat a l’Energie Atomique, INAC/SPSMS, 38054 Grenoble, France]{} ]{}
Introduction
============
A phase transition of the second order breaks some symmetry such that below the critical temperature the ordered phase of lower symmetry in comparison with the initial state is formed. As it was first pointed out by L.D.Landau [@Lan] an intersection of critical lines on the phase diagram leads to formation of an ordered phase with symmetry lower than the symmetries of both initial ordered states existing below of each critical lines separately. Here we study the symmetries of ordered phases arising at intersection of critical lines of ferromagnet and superconducting phase transitions.
The co-existence of superconductivity and ferromagnetism in several uranium compounds $UGe_2$, [@Saxena] $URhGe$, [@Aoki] and the recently revealed $UCoGe$. [@Huy07] is found to arise as a co-operative phenomena rather than as the overlap of two-mutually competing orders. In all these compounds the substantial reduction of the ordered moment as compared with the Curie-Weiss moment provides clear evidence of $5f$ itineracy. In the first two compounds the Curie temperatures $T_{Curie}$ is more than the order of magnitude higher than their critical temperatures for superconductivity. In $UCoGe$ the ratio $T_{Curie}/T_{sc}$ at ambient pressure is about four. The large exchange field and also high upper critical field at low temperatures strongly exceeding the paramagnetic limiting field [@Huxley01; @Hardy051; @Huy08] indicate that here we deal with Cooper pairing in the triplet state.
{height=".5\textheight"}
The singlet superconductivity coexists with ferromagnetism in a form known as the Anderson-Suhl or cryptoferromagnetic superconducting state (for review see [@Buzdin]) characterized by the formation of a transverse domain-like magnetic structure. The structure period or domain size is larger than interatomic distance and smaller than the superconducting coherence length that weakens the depairing effect of the exchange field. The latter is irrelevant in the case of triplet superconductivity. Hence, there is no reason for the formation of a cryptomagnetic state. Indeed, no traces of space modulation of magnetic moments directions on the scale smaller than the coherence length has been revealed.[@Aoki; @Aso5; @Kotegawa; @Ohta] On the other hand the neutron depolarization measurements on $UGe_2$ down to 4.2 K (that is in the ferromagnet but not superconducting region) establish, that the magnetic moment strictly aligned along a-axis, with a typical domain size in the bc-plane of the order $4.4\times10^{-4}$ cm [@Sakarya] that is about two orders of magnitude larger than the largest superconducting coherence length in b-direction $\xi_b\approx 7\times10^{-6}$ cm.
Arising at temperatures far below the corresponding Curie temperature the superconductivity in $UGe_2$ and $URhGe$ coexists with ferromagnetism in some pressure interval such that in the $(P,T)$ phase diagram the whole region occupied by the superconducting state is situated inside a more vast ferromagnetic region.[@Huxley03; @Hardy052] In $URhGe$ the Curie temperature increases up to the highest pressure achieved $(130~ kbar)$. The superconducting critical temperature decreases slowly up to 20 kbar. It is more peculiar the behavior of $UGe_2$, where at low temperatures the ferromagnetism and the superconductivity abruptly ( by means the first-order-type transition) disappears at the same critical pressure $P_c\approx 15~ kbar$.
The observation that the superconductivity in $UGe_2$ is confined to the ferromagnet state can be trivially explained by an assumption that the ferromagnetism in this compound is formed by f-electrons with half-metallic bands filling. Namely, the band with the spin-down electrons is completely filled, whereas the band with spin-up electrons filled up to the Fermi level. The triplet spin-up superconducting state formed in this band persists so long the Fermi level intersects this band. The pressure induced the Fermi level lifting above the band upper boundary kills both the itinerant ferromagnetism and the superconductivity. The particular one-band superconducting state was chosen [@Hardy051] for successful explanation of temperature dependences of the upper critical field in $URhGe$ in different crystallographic directions. This state is also appropriate for the description [@Mineev2006] of the transition driven by the change of orientation of the ordered magnetic moment in this compound by the application of magnetic field in perpendicular direction accompanying by the arising a reentrant superconducting state [@Levy]. The one band superconductivity, of course, does not exclude the existence of the other conducting but not superconducting bands, or, more exactly, the bands with negligibly small superconducting gaps. The latter is in correspondence with reduced specific heat jump in comparison with BCS value, and the finite residual zero temperature ratio $(C(T)/T)_{T\to0}$ comparable with its magnitude in the normal state found in all uranium superconducting ferromagnets.
The phase diagram of the new ferromagnetic superconductor $UCoGe$ is qualitatively different (see Figure 1).[@Hassinger] At ambient pressure, the ferromagnetism ($T_{Curie}\approx 3 K$) coexists with superconducting state ($T_{sc}\approx 0.7 K$). Then at applied pressure, the Curie temperature decreases such that no indication of ferromagnetic order is observed above $P^*\approx 10~kbar$. The resistive superconducting transition is, however, quite stable with changes in temperature and persists up to the highest measured pressure of about $24~ kbar$. Thus, the pressure dependent transition lines $T_{Curie}(P)$ and $T_{sc}(P)$ apparently intersect each other and the superconductivity exists both in the paramagnet and in the ferromagnet state.
The ferromagnet superconducting state in an orthorhombic metal is similar to the superfluid $^3He-A$ in an external magnetic field known as $A_2$ state. The superfluid $^3He-A$ is the spin nonpolarized state formed by the spin-up and the spin-down Cooper pairs in equal amounts. There is also the spin-polarized $A_1$ state where the pairing only spin-up particles occurs.[@Leggett] The $^3He-A_1$ arises from the normal Fermi liquid in an external magnetic field. Then, at lower temperature, the liquid passes to the $A_2$ state where the paired spin-up and spin-down states are almost equally populated. The presence of spin-orbital coupling admixes some amount of the spin-down Cooper pairs to pure $A_1$ state [@Kojima], such that the $A_1$ and $A_2$ states are in fact qualitatively indistinguishable. The phase transition between these two states is a crossover, looking as a phase transition due to the smallness of the spin-orbital coupling in superfluid $^3He$. The nonunitary two-band superconducting state arising in the ferromagnet state of $UCoGe$ can be considered as an analog of superfluid $A_2$ phase arising from the normal liquid $^3He$ under magnetic field.
The increasing pressure causes the decrease of the exchange field that suppresses the spin-up and spin-down band difference. The restoration of the time reversal symmetry occurs at recreation of spin-up and spin-down band degeneracy by the phase transition from the ferromagnet [*axiplanar*]{} superconducting state to the paramagnet superconducting state similar to the [*planar*]{} state of the superfluid $^3He$ (for the superfluid $^3He$ phase definitions see for instance [@MinSam]). Thus, the ferromagnet superconducting state is separated from the normal state by the more symmetric, paramagnet planar-like state. We see, that the $(P,T)$ phase diagram in $UCoGe$ is quite naturally explained in terms of two band superconducting state in this material. The observation of the upward curvature in the temperature dependence of the upper critical field in $UCoGe$ [@Huy08] adds the additional argument in support of this point.
The symmetries and the order parameters of unconventional superconducting states arising from the normal state with a ferromagnetic order in orthorhombic crystals with strong spin-orbital coupling have been found in the paper. [@Mineev] Then it was pointed out that superconducting states in triplet ferromagnet superconductors represent a special type of two band superconducting states. [@Cham; @Min04]. There were obtained several results based on phenomenlogical (Ginzburg-Landau) and microscopic descriptions of two-band superconductivity. It was proved, however, that the superconducting ferromagnet classes pointed there have been found improperly. Although, it leaves untouched the main results of [@Mineev; @Cham; @Min04], the based on these papers description of possible $(P,T)$ phase diagrams for two band superconducting states in an orthorhombic itinerant ferromagnet is incorrect.[@Min08] To make it correctly we return to the definition of the superconducting ferromagnet classes. It will be proven that $T_{Curie}(P)$ and $T_{sc}(P)$ can intersect each other as the critical lines of the phase transition of the second order. The symmetry and the order parameters of the multidomain ferromagnet as well of the paramagnet superconducting states are established. The Josephson coupling between neighboring superconducting domains is also discussed.
Two-band superconducting ferromagnet phase diagram
==================================================
All uranium ferromagnetic superconductors are orthorhombic metals. The symmetry of its normal paramagnetic state is determined by the elements of the group $$G_N=D_2\times U(1)\times R,
\label{e-1}$$ where $D_{2}=(E, C_{2}^{z},
C_{2}^{x}, C_{2}^{y})$ is the point symmetry group including the operations $C_{2}^{x}, C_{2}^{y},C_{2}^{z}$ of rotation on the angle $\pi$ about the $x,y,z$- axes correspondingly, $U(1)$ is the group of gauge transformations, and $R$ is the time reversal operation.
If the pressure dependent transition lines $T_{Curie}(P)$ and $T_{sc}(P)$ intersect each other at $P=P^*$ then the region of the coexistence of superconductivity and ferromagnetism is separated from the normal state by the region of ferromagnet normal state at $P<P^*$ and by the region of the superconducting state at $P>P^*$ (see Figure 1).
In the transition from the normal paramagnet state to the normal ferromagnet state the magnetic moment directed along one crystallographic axis appears. We chose this direction as the $\hat z$ axis. Hence, in the ferromagnet state the symmetry reduces to the $$G_F=D_2(C_2^z)\times U(1),
\label{e0}$$ where $$D_2(C_2^z)=(E,C_2^z,RC_2^x,RC_2^y)
\label{e00}$$ is the so called [*magnetic class*]{} [@Landau] or the point symmetry group of the ferromagnet. The rotations on the angle $\pi$ about the $x$- and $y$- directions are accompanied by the time inversion $R$ that changes the direction of magnetization to the opposite one.
In the transition from the normal paramagnet state to the superconducting paramagnet state the gauge symmetry is broken, such that the symmetry of this, so called [*conventional*]{} superconducting state is $$G_S=(E,C_2^z, C_2^x, C_2^y)\times R.
\label{es}$$ There is another possibility related to the formation of [*nonconventional*]{} superconducting state where, in the addition to the gauge symmetry, the point symmetry is also broken. We shall not discuss it here.
Now, we shall consequently describe the phase transitions from the normal ferromagnet state (F) to the superconducting ferromagnet state (FS) taking place at $P<P^*$ and from the paramagnet superconducting state (S) to the superconducting ferromagnet state (FS) at $P>P^*$.
### F to FS phase transition
As it was remarked in [@Cham] superconducting state in an itinerant ferromagnet represents the special type of two band superconducting state consisting of pairing states formed by spin-up electrons from one band and by spin-down electrons from another band. Hence, a superconducting state characterizes by two component order parameter $${\bf d}_1({\bf k})=\Delta_{\uparrow}({\bf k})(\hat{x}+i\hat{y}),~~~
{\bf d}_2({\bf k})=
\Delta_{\downarrow}({\bf k})(\hat{x}-i\hat{y}).
\label{e1}$$ Here, $\hat{x}$ and $\hat{y}$ are the unit vectors of the spin coordinate system pinned to the crystal axes.
The unconventional superconducting states arising from the normal state with a ferromagnetic order in orthorhombic crystals with strong spin-orbital coupling belong to the two different corepresentations $A$ and $B$.[@Mineev] All the states relating to the given corepresentation obey the same critical temperature. The order parameter amplitudes for $A$ and $B$ states correspondingly are given by $$\begin{aligned}
\Delta_{\uparrow}^A({\bf k})=\eta_1(k_xu_1+ik_yu_2),\nonumber\\
\Delta_{\downarrow}^A({\bf k})=
\eta_2(k_xu_3+ik_yu_4),
\label{e2}\end{aligned}$$ $$\begin{aligned}
\Delta_{\uparrow}^B({\bf k})=\eta_1(k_zv_1+ik_xk_yk_zv_2),\nonumber\\
\Delta_{\downarrow}^B({\bf k})=
\eta_2(k_zv_3+ik_xk_yk_zv_4).
\label{e2'}\end{aligned}$$ They are odd functions of the momentum directions of pairing particles on the Fermi surface. The functions $u_i=u_i(k_x^2, k_y^2, k_z^2)$ and $v_i=v_i(k_x^2, k_y^2, k_z^2)$ are invariant in respect of all transformations of orthorhombic group. For the brevity, in that follows, we shall discuss only the $A$ state. This state is related to the family of nonunitary axiplanar states.
The complex order parameter amplitudes $\eta_1=|\eta_1|e^{i\varphi_1}$ and $\eta_2=|\eta_2|e^{i\varphi_2}$ are not completely independent. The relative phase difference $\varphi_1-\varphi_2$ is chosen such that the quadratic in the order parameter part of the Ginzburg-Landau free energy density should be minimal. In an ordinary two-band superconductor it is $$F=\alpha_1|\eta_1|^2+\alpha_2|\eta_2|^2+\gamma(\eta_1^*\eta_2+\eta_1\eta_2^*),
\label{44}$$ and $\varphi_1-\varphi_2=\pi$ for $\gamma>0$ and $\varphi_1-\varphi_2=0$ for $\gamma<0$. In the case of ferromagnetic normal state the time reversal symmetry is broken and the quadratic in the order parameter components free energy density has the form $$\begin{aligned}
F=\alpha_1|\eta_1|^2+\alpha_2|\eta_2|^2+\gamma(\eta_1^*\eta_2+\eta_1\eta_2^*)
\nonumber\\+
i\delta(\eta_1^*\eta_2-\eta_1\eta_2^*).
\label{45}\end{aligned}$$ Here, all the coefficients are the functions of the exchange field $h$. The last term breaks the time reversal symmetry. In the absence of exchange field $\delta=0$. Minimization of free energy fixes the order parameter components phase difference $\tan(\varphi_1-\varphi_2)=\delta/\gamma$. After substitution of this value back to we come to the expression $$F=\alpha_1|\eta_1|^2+\alpha_2|\eta_2|^2
-\sqrt{\gamma^2+\delta^2}( \eta_1^*\eta_2+\eta_1\eta_2^*).
\label{freeen}$$ Here $\alpha_i=\alpha_{i0}(T-T_{ci})$, $i=1,2$ are the band indices, $T_{ci}$ are the critical temperatures in each band in the absence of band mixing. Unlike eqn. the complex amplitudes $\eta_1=|\eta_1|e^{i\theta}$, $\eta_2=|\eta_2|e^{i\theta}$ in the eqn. have common phase factors with $\theta=(\varphi_1+\varphi_2)/2$. This form of free energy valid near the phase transition from the ferromagnet state to the ferromagnet superconducting state has been used in the papers. [@Cham; @Min04] The common for the each band superconductivity critical temperature is given by $$T_{sc}=\frac{T_{c1}+T_{c2}}{2}+\sqrt{\left (\frac{T_{c1}-T_{c2}}{2}\right )^2+\frac{\gamma^2+\delta^2}{\alpha_{10}\alpha_{20}}}$$
In the superconducting $A$-state the gauge symmetry is broken. Acting on the order parameters (5), (6) by the elements $g$ of $D_2(C_2^z)=(E,C_2^z,RC_2^x,RC_2^y)$ group we obtain the following coefficients of transformation, or matrices of corepresentation $$\Gamma_1=(1,1, e^{-2i\varphi_1}, e^{-2i\varphi_1}),~~~
\Gamma_2=(1,1, e^{-2i\varphi_2}, e^{-2i\varphi_2}),
\label{Gamma}$$ correspondingly. Corepresentations $\Gamma_1$ and $\Gamma_2$ are equivalent or they are transformed each other by an unitary matrix $U$ as $\Gamma_1(g) =U^{-1}\Gamma_2(g)U$ if the element $g$ does not include the time inversion, and as $\Gamma_1(g) =U^{-1}\Gamma_2(g)U^*$ if the element $g$ includes the time inversion. It is easy to check that here the matrix of transformation is $U=e^{i(\varphi_2-\varphi_1)}$.
The order parameter component ${\bf d}_1({\bf k})$ relating to the spin-up band is invariant in respect to the following group of transformations $$G_{FS}=(E,C_2^z, RC_2^x, RC_2^y)=D_2(C_2^z).
\label{46}$$ Action of the time reversal operation $R$ on superconducting order parameter implies also the multiplication of it by the square of its phase factor: $R\to e^{2i\varphi_1}R$. The second component ${\bf d}_2({\bf k})$ possess the same symmetry. So, the group of symmetry of superconducting ferromagnet state $A$ called also by [*the superconductiing magnetic class*]{} is $D_2(C_2^z)$. This group is the subgroup of the group of symmetry of the ferromagnet state .
### Superconducting ferromagnet domains
The Cooper pairing changes the magnitude of spontaneous magnetization in respect to its value in normal ferromagnet state. Namely, the superconducting spin magnetic moment density is $${\bf M}_s=\mu_B\left [N'_{0\uparrow} \langle|\Delta_{\uparrow}¥({\bf k})|^2\rangle-N'_{0\downarrow}\langle|\Delta_{\downarrow}¥({\bf k})|^2\rangle \right].$$ Here, in the first term, $N'_{0\uparrow}$ is the derivative of the density of states at the Fermi surface of the spin-up band , and the angular brackets means the averaging over it. The second term presents the corresponding input of the spin-down band. One can write also the orbital magnetic moment density. [@MinSam]
Along with the introduced state $A$, there is its time reversed state $A^*$ characterized by the complex conjugate order parameter components $$\begin{aligned}
{\bf d}_1^*({\bf k})=\zeta_1(\hat{x}-i\hat{y})(k_xu_1-ik_yu_2),\nonumber\\
{\bf d}_2^*({\bf k})=\zeta_2(\hat{x} +i\hat{y})(k_xu_3-ik_yu_4).\end{aligned}$$ The states $A$ and $A^*$ occupy neighboring domains with the opposite direction of magnetization. The state $A^*$ order parameter amplitudes are $\zeta_1=|\zeta_1| e^{i\phi_1}$ and $\zeta_2=|\zeta_2| e^{i\phi_2}$. The phase difference is fixed by $\tan(\phi_1-\phi_2)=\delta(-h)/\gamma$.
The matrices of corepresentations for the state $A^*$ are obtained from by the substitution $\varphi_{1,2} \to \phi_{1,2}$. So, they transformed each other by means the matrices $U_i=e^{i(\varphi_i-\phi_i)}$. It means, that the corepresentations for the state $A^*$ are equivalent to the corepresentations for the state $A$. Hence, the superconducting states in the neighboring domains obey the same critical temperature.
The symmetry of the time reversed states $A^*$ belong to the same [*superconducting ferromagnet class*]{} $D_2(C_2^z)$ as the $A$-states.
### S to FS phase transition
In $P>P^*$ region at temperature decrease $UCoGe$ pass to the nonmagnetic superconducting state. Let us assume the simplest and quite natural situation that it is the superconducting state with the order parameter $${\bf d}({\bf k})=2\eta(
k_xw_1\hat {x}+k_yw_2\hat y),
\label{d}$$ transforming according to the unit representation of the normal state point symmetry group $D_2$. Here $\eta=|\eta|e^{i\varphi}$ and the functions $w_{1,2}=w_{1,2}(k_x^2, k_y^2, k_z^2)$ are invariant in respect of all transformations of orthorhombic group. This state reminds planar phase of superfluid $^3He$. The paramagnet superconducting state is invariant in respect to the group which can be rewritten as $$G_S=D_2(C_2^z)+R\times D_2(C_2^z).
\label{G_S}$$
By further decrease the temperature we approach to $T_{Curie}(P)$. At this temperature the exchange field appears, and the Kramers degeneracy between spin-up and spin-down electron states is lifted accompanied by arising of deviation from of the order parameter $$\begin{aligned}
&{\bf d}({\bf k})=2\eta(
k_xw_1\hat {x}+k_yw_2\hat y)~~~~~~~~~~~~\nonumber\\ &=\eta(k_xw_1-ik_yw_2)(\hat x+i\hat y)+
\eta(k_xw_1+ik_yw_2)(\hat x-i\hat y)\nonumber
\\
&\to \tilde{\bf d}({\bf k})=\eta_1(k_xw_1-ik_yw_2)(\hat x+i\hat y)\nonumber\\ &+
\eta_2(k_xw_1+ik_yw_2)(\hat x-i\hat y)
\label{d'}\end{aligned}$$ The order parameter $\tilde{\bf d}({\bf k})$ transforms according to corepresentation of the symmetry group of the paramagnet superconducting state.
Along with increase of the band splitting the two component of the order parameter $\tilde{\bf d}({\bf k})$ are transformed to the order parameters of spin-up and spin-down bands given by eqns. , . The ferromagnet superconducting state determined by eqn. as well by the eqns. and is invariant in respect to the group $$G_{FS}=D_2(C_2^z).$$ The latter is the subgroup of the group of symmetry of ferromagnet state $G_F$ as well as of the symmetry group of paramagnet superconducting state $G_S$ . So, the lines of the ferromagnet and the superconducting phase transitions can intersect each other as the critical lines of the phase transitions of the second order.
Interdomain Josephson coupling
==============================
Let us consider a flat domain wall dividing magnetic moment-up and -down domains in single band ferromagnet. This case, the localized at $x=0$ domain wall contribution to the superconducting free energy density is given by [@Sam] $$\begin{aligned}
F_{DW}=\left [\gamma_1(|\eta|^2+|\zeta|^2)+\gamma_2(\eta^*\zeta+\eta\zeta^*)
\right.\nonumber\\+
\left.i\gamma_3(\eta^*\zeta-\eta\zeta^*)\right ]\delta(x).
\label{Fs}\end{aligned}$$ Here $\eta=|\eta|e^{i\varphi}$ and $\zeta=|\zeta|e^{i\phi}$ are the superconducting order parameters in the left (magnetic moment-up) domain and in the right (magnetic moment-down) domain, correspondingly. The boundary conditions at $x=0$ are derived by the minimization of the sum of domain wall and the gradient free energies.[@MinSam] $$\begin{aligned}
K\frac{\partial\zeta}{\partial x}=\gamma_1\zeta+(\gamma_2+i\gamma_3)\eta\nonumber\\
-K\frac{\partial\eta}{\partial x}=\gamma_1\eta+(\gamma_2-i\gamma_3)\zeta.
\label{bc}\end{aligned}$$ Here, the rigidity coefficients $K\sim\hbar^2/m$. The solutions of left and right domain nonlinear Ginzburg-Landau equations supplemented by these boundary conditions determine the order parameter distribution of two domain superconducting structure. The solution of corresponding linear problem is physically relevant only in the case of stimulation of superconductivity by the domain wall when the localized near domain wall superconducting state arises at temperatures higher than the temperature of superconducting phase transition in single domain geometry.
The situation for two band superconductivity is much more complicated. This case the two-band domain wall free energy density is obtained by the addition to the eqn. the corresponding terms for the second band order parameters $\eta_2$ and $\zeta_2$ and also the interband terms symmetric in respect to the substitutions $\eta_i$ by $\zeta_i$ and vice versa.
Substituting the boundary conditions in the sum of the left domain and the right domain current through the domain wall (see for instance [@MinSam]) we obtain the density of the interdomain Josephson current: $${\bf j}=\frac{8eK}{\hbar}|\zeta||\eta|\left [ \gamma_2\sin(\phi-\varphi)-\gamma_3\cos(\phi-\varphi)\right ]$$ Thus, due to the time reversal breaking ($\gamma_3\ne0$) the expression for the Josephson current between the adjacent superconducting domains with spin-up and spin-down magnetization differs from the usual weak link Josephson current formula. In the equilibrium, the phase difference between domains is fixed: $\tan(\phi-\varphi)=\gamma_3/\gamma_2$, and a spontaneous interdomain current is absent.
In conclusion of this section it is worth to be noted that the existence of the interdomain Josephson coupling bilinear in respect of $|\eta|$ and $|\zeta|$ is typical for the $A$ superconducting states. The order paramer for the $B$ states is vanishing in the equatorial plane $k_z=0$. This case, there is only the higher order Josephson coupling between the domains divided by a flat domain wall parallel to the magnetization direction.
Conclusion
==========
The superconducting state in the itinerant ferromagnet uranium compound $UCoGe$ manifests the properties naturally explained in terms of two band superconductivity with triplet pairing. We discussed the symmetry and the order parameters of such a state. It was proven that the pressure dependent critical lines of the ferromagnet $T_{Curie}(P)$ and the superconducting phase transition $T_{sc}(P)$ can intersect each other in the correspondence with the experimental observations.[@Hassinger] The Josephson coupling in between the adjacent superconducting ferromagnet domains is found to be different in comparison with the usual Josephson coupling between two superconductors.
[99]{}
L. D. Landau, *ZhETF* **7**, 19 (1937) \[*Phys. Zs. Sowjet.* **11**, 26 (1937)\].
S. S. Saxena, P. Agarval, K. Ahilan, F. M. Grosche, R. K. W. Hasselwimmer, M. J. Steiner, E. Pugh, I. R. Walker, S. R. Julian, P. Monthoux, G. G. Lonzarich, A. Huxley, I. Sheikin, D. Braithwaite, and J. Flouquet, *Nature* **406**, 587 (2000).
D. Aoki, A. Huxley, E. Ressouche, D. Braithwaite, J. Flouquet, J.-P. Brison, E. Lhotel and C. Paulsen, *Nature* **413**, 613 (2001).
N. T. Huy, A. Gasparini, D. E. de Nijs, Y. Huang, J. C. P. Klaasse, T. Gortenmulder, A. de Visser, A. Hamann, T. Gorlach, and H. v. Lohneysen, *Phys. Rev. Lett.* **99**, 067006, (2007).
A. Huxley, I. Sheikin, E. Ressouche, N. Kernavanois, D. Braithwaite, R. Calemzuk, and J. Flouquet, *Phys. Rev. B* **63**, 144519 (2001).
F. Hardy and A. D. Huxley, *Phys. Rev. Lett.* **94**, 247006 (2005).
N. T. Huy, D. E. de Nijs, Y. Huang, and A. de Visser, *Phys.Rev.Lett.* **100**, 007002 (2008).
L. N. Bulaevskii, A. I. Buzdin, M. L. Kulic, and S. V. Panjukov, *Adv. Phys.* **34**, 175 (1985).
N. Aso, H. Nakane, G. Motoyama, N. K. Sato, Y. Uwatoko, T. Takeuchi, Y. Homma, Y.Shiokawa, K. Hirota, *Physica B* **359-361**, 1051 (2005).
H. Kotegawa, A. Harada, S. Kawasaki, Y. Kawasaki, Y. Kitaoka, Y. Haga, E. Yamamoto, Y. Onuki, K. M. Itoh, E. E. Haller, and H. Harima, *J. Phys. Soc. Jpn.* **74**, 705 (2005).
T. Ohta, Y. Nakai, Y. Ihara, K. Ishida, K. Deguchi, N. K. Sato, and I. Satoh, *J. Phys. Soc. Jpn.* **77**, 023707 (2008).
S. Sakarya and N. H. van Dijk, E. Bruck, *Phys. Rev.B* **71**, 174417 (2005).
A. Huxley, E. Ressouche, B. Grenier, D. Aoki, J. Flouquet, and C. Pfleiderer, *J. Phys.: Condens. Matter* **15**, 1945 (2003).
F. Hardy, A. Huxley, J. Flouquet, B. Salce, G. Knebel, D. Braithwaite, D. Aoki, M. Uhlarz, C. Pfleiderer, *Physica B* **359-361**, 1111 (2005).
V. P. Mineev, *C. R. Physique* **7**, 35 (2006).
F. Levy, F. Hardi, I. Sheikin, B. Grenier, A. D. Huxley, *Science* **309**, 1343 (2005).
E. Hassinger, D. Aoki, G. Knebel, and J. Flouquet, *J. Phys. Soc. Jap.* **77**, 073703 (2008).
A. Leggett, *Rev. Mod. Phys* **47**, 331 (1975).
A Yamaguchi, S. Kobayashi, H. Ishimoto and H. Kojima, *Nature* **444**, 909 (2006).
V. P. Mineev, K. V. Samokhin, *Introduction to Unconventional Superconductivity*, Gordon and Breach Science Publishers, Amsterdam, 1999.
V. P. Mineev, *Phys. Rev. B* **66**, 134504 (2002).
V. P. Mineev and T. Champel, *Phys. Rev. B* **69**, 144521 (2004).
V. P. Mineev, *Int. J. Mod. Phys.* **18**, 2963 (2004).
V. P. Mineev, *J. Phys. Soc. Jpn.* **77**, 103702 (2008).
L. D. Landau and E. M. Lifshitz, *Electrodynamics of Continuous Media*, Pergamon Press, Oxford, 1984.
K. V. Samokhin and D. Shirokoff, *Phys. Rev. B* **71**, 104527 (2005).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Lower bounds are derived on the amplitude $B$ of intergalactic magnetic fields (IGMFs) in the region between Galaxy and the blazar Mrk 421, from constraints on the delayed GeV pair-echo flux that are emitted by secondary $e^-e^+$ produced in $\gamma\gamma$ interactions between primary TeV gamma-rays and the cosmic infrared background. The distribution of galaxies mapped by the Sloan Digital Sky Survey shows that this region is dominated by a large intergalactic void. We utilize data from long-term, simultaneous GeV-TeV observations by the [*Fermi*]{} Large Area Telescope and the ARGO-YBJ experiment extending over 850 days. For an assumed value of $B$, we evaluate the daily GeV pair-echo flux expected from the TeV data, select the dates where this exceeds the [*Fermi*]{} 2-$\sigma$ sensitivity, compute the probability that this flux is excluded by the [*Fermi*]{} data for each date, and then combine the probabilities using the inverse normal method. Consequently, we exclude $B < 10^{-20.5}~{\rm G}$ for a field coherence length of 1 kpc at $\sim$ 4-$\sigma$ level, as long as plasma instabilities are unimportant for cooling of the pair beam. This is much more significant than the 2-$\sigma$ bounds we obtained previously from observations of Mrk 501, by virtue of more extensive data from the ARGO-YBJ, as well as improved statistical analysis. Compared with most other studies of IGMF bounds, the evidence we present here for a non-zero IGMF is more robust as it does not rely on unproven assumptions on the primary TeV emission during unobserved periods.'
author:
- 'Keitaro Takahashi, Masaki Mori, Kiyotomo Ichiki, Susumu Inoue, Hajime Takami,'
title: 'Lower Bounds on Magnetic Fields in Intergalactic Voids from Long-Term GeV-TeV Light Curves of the Blazar Mrk 421'
---
Introduction
============
Intergalactic magnetic fields (IGMF), particularly those inside intergalactic void regions, have attracted much interest as possible remnants of primordial magnetic fields that were generated in the early Universe [e.g. @Gnedin; @Langer; @Takahashi1; @Ichiki2]. While such fields can be amplified later within galaxies and galaxy clusters by dynamo processes, they may remain unaffected by subsequent astrophysical effects deep inside voids. Thus, IGMFs are expected to be a window onto the early Universe. For comprehensive reviews on primordial and intergalactic magnetic fields, see Widrow (2002), Widrow et al. (2012) and Ryu et al. (2012).
However, the predicted amplitudes for IGMFs of primordial origin are generally very small, $B = 10^{-25} - 10^{-15}~{\rm G}$, and difficult to probe through Faraday rotation measurements in distant radio sources or their effects on the anisotropy of the cosmic microwave background (CMB). In this context, a method that is sensitive to weak IGMFs utilizing delayed secondary emission from high-energy gamma-ray sources was proposed by Plaga (1995) and subsequently developed by many authors [@Dai; @Razzaque; @Murase1; @Murase2; @Ichiki1; @Takahashi4; @Elyiv; @Neronov1; @Takahashi2]. Such emission that we refer to as “pair echos” is expected to occur typically at GeV energies, for which the [*Fermi*]{} Large Area Telescope (LAT) is currently the most sensitive instrument. Since the echo flux is predicted to be larger for smaller $B$, a GeV upper limit on such components translates into a lower bound on $B$.
In our previous study [@Takahashi3], we focused on a specific TeV flare of Mrk 501 observed in 2009 by the VERITAS (Very Energetic Radiation Imaging Telescope Array System) and MAGIC (Major Atmospheric Gamma-ray Imaging Cherenkov) telescopes. Comparing the expected light curves of the pair echo from the flare and the concurrent quiescent emission with simultaneous [*Fermi*]{} observations, we obtained a lower bound on the IGMF amplitude of $B > 10^{-20}~{\rm G}$ at $90\%$ confidence level assuming a field coherence length of 1 kpc. This was obtained with minimal assumptions about the primary TeV emission during unobserved periods or spectral bands, and can be considered more robust in comparison with previous studies [@Neronov3; @AndoKusenko; @Tavecchio1; @Tavecchio2; @Dolag; @Dermer; @Neronov2; @Taylor; @Arlen].
Here we focus on the TeV blazar Mrk 421 located at $z = 0.031$. As seen in Fig. \[fig:Mrk421\], maps of the local galaxy distribution from the Sloan Digital Sky Survey reveal that a large void lies between our galaxy and the supercluster containing Mrk 421 [@Abazajian; @Blanton]. This is also seen to be the case for Mrk 501. Thus, Mrk 421 is a desirable target for probing IGMFs. Mrk 421 has been monitored continuously at TeV energies by the ARGO-YBJ experiment over the period from 2007 November to 2010 February [@Bartoli] (hereafter B11), during which many flares were observed so that more statistically significant bounds on IGMFs can be expected. Note that compared with Cherenkov telescopes, such air shower detectors have a much higher duty cycle and allow uninterrupted long-term observations, albeit at lower sensitivity.
TeV and GeV Emission from Mrk 421
=================================
First we discuss the TeV spectrum and light curve of Mrk 421 with which we evaluate the pair echo. In B11, the daily fluxes at energies above $0.3~{\rm TeV}$ are presented for approximately 850 days. For some days, negative numbers are reported that are presumably caused by systematic errors, and we simply set them to zero. Although the spectra are not available separately for each day, average spectra were derived for four different flux states based on the X-ray count rate. Since the TeV flux was shown to be tightly correlated with that in X-rays, here we choose to define three flux states according to the daily TeV counts, “high" (count $> 80$), “medium" ($40 <$ count $< 80$) and “low" (count $< 40$), which correspond respectively to the X-ray flux levels 4, 3 and 1+2 of B11. Note that levels 1 and 2 can be treated together for our purposes as their TeV spectra are very similar. According to the daily flux, we assume that the TeV spectral index for each day takes the average value of the corresponding flux state. We also impose a maximum spectral cutoff at 5 TeV as the highest energy photons detected by ARGO-YBJ, as well as a minimum cutoff at 0.1 TeV. In Fig. \[fig:spectrum\], exemplary spectra for the three states are shown, with and without the effects of intergalactic $\gamma\gamma$ absorption using the cosmic infrared background (CIB) model of Franceschini et al. (2008), also adopted in B11. It turns out that the resulting constraints on the IGMF is largely determined by the high state emission, and the low state is of very little relevance.
For GeV gamma-rays, we utilize the data from [*Fermi*]{} LAT that has been performing continuous observations of Mrk 421 in the survey mode from MJD 54683. We obtained the data through the [*Fermi*]{} Science Support Center (FSSC) and adopt the standard analysis tools provided by the FSSC. In our analysis, we divide the energy band in three, that is, 100 MeV-1 GeV, 1-10 GeV, and $>$10 GeV, and derive flux probability distribution functions for each day assuming Poisson statistics. Because the statistics is small for this short interval (1-day bins), we adopt the aperture photometry method where we count events located within 1 degree from the source. Note that we can neglect the background events above 1 GeV for this timescale at the high Galactic latitude of Mrk 421. Below we use data during MJD 54683 - 55255, when both TeV and GeV observations were performed, focusing on the energy range of $1-10~{\rm GeV}$ where [*Fermi*]{} LAT is most sensitive and the strongest constraints on the pair echo can be obtained.
Pair Echo
=========
We summarize briefly the basic physics of pair echos (for details, see e.g. Ichiki et al. 2008 and Takahashi et al. 2012). The mean free path of primary gamma-rays with energy $E_{\gamma} \gtrsim 1 ~ {\rm TeV}$ for $\gamma\gamma$ interactions with the CIB is $
\lambda_{\gamma \gamma}
= 1/(0.26 \sigma_T n_{\rm IR})
= 190~{\rm Mpc}~( n_{\rm IR}/0.01 ~ {\rm cm}^{-3} )^{-1} ,
$ where $\sigma_T$ is the Thomson cross section and $n_{\rm IR}$ is the number density of relevant CIB photons. The interaction results in an $e^-e^+$ pair with energy $E_e \approx E_{\gamma}/2$, which can then inverse-Compton (IC) upscatter ambient CMB photons to produce the pair echo, that is, secondary gamma rays with energy $
\langle E_{\rm echo} \rangle
= 2.7 T_{\rm CMB} \gamma_e^2
= 2.5~{\rm GeV}~
( E_{\gamma}/2 ~ {\rm TeV} )^2,
$ where $T_{\rm CMB} = 2.7~{\rm K}$ is the CMB temperature and $\gamma_e = E_e/m_e c^2$. For primary gamma rays with $E_\gamma \simeq 1 - 5~{\rm TeV}$, $E_{\rm echo} \simeq 1 - 10~{\rm GeV}$. As long as plasma instabilities are unimportant (see below), the pairs continue successive IC scattering until they lose a large fraction of their energy over a length scale $
\lambda_{\rm IC, cool}
= 3 m_e^2/(4 E_e \sigma_T U_{\rm CMB})
= 350 ~ {\rm kpc}~( E_e/1 ~ {\rm TeV} )^{-1},
$ where $U_{\rm CMB}$ is the CMB energy density. Comparing typical values for $\lambda_{\gamma \gamma}$ and $\lambda_{\rm IC, cool}$, we see that the pairs are generated mostly far away from the source, and then cool over a much smaller scale. Thus, for Mrk 421, the pairs are likely to be produced deep inside and propagate only within the large, intervening void (Fig. \[fig:Mrk421\]).
It has been suggested recently that rather than IC cooling, the beam of the $\gamma\gamma$-produced pairs may lose much of their energy by heating the intergalactic gas through two-stream-like plasma instabilities [@Broderick; @Schlickeiser]. If true, it may considerably reduce the pair echo signal, while causing some non-trivial consequences for the evolution of galaxies and the intergalactic medium [@Chang; @Pfrommer]. However, the actual efficiency and eventual fate of such instabilities has been debated [@Miniati] and is highly uncertain at the moment. Below, we proceed on the assumption that such instabilities are insignificant.
A crucial attribute of the pair echo is the time delay compared with the primary gamma-rays, caused by two effects. One is the angular spreading inherent in the pair production and IC scattering processes, for which the typical delay time $
\Delta t_{\rm ang}
= (\lambda_{\gamma\gamma} + \lambda_{\rm IC, cool})/2 \gamma_e^2
\approx 3 \times 10^3 ~ {\rm sec} ~ (E_{\rm echo}/1 ~ {\rm GeV})^{-1}
(n_{\rm IR}/0.01 ~ {\rm cm}^{-3})^{-1}
$ (Ichiki et al. 2008). The second is deflections of the pairs in the IGMF with typical delay time $
\Delta t_{\rm B}
= (\lambda_{\gamma \gamma} + \lambda_{\rm IC, cool})
\langle \theta_{\rm B}^2 \rangle /2
$, where $
\langle \theta_{\rm B}^2 \rangle^{1/2}
= \max [\lambda_{\rm IC, cool}/r_{\rm L}, (\lambda_{\rm IC, cool} r_{\rm coh}/6)^{1/2} / r_{\rm L}]
$ is the typical deflection angle, $r_{\rm L}$ the Larmor radius and $r_{\rm coh}$ the coherence length of the IGMF. If $r_{\rm coh} \ll \lambda_{\rm IC, cool}$, that is, the IGMF is sufficiently tangled on the IC cooling scale, $$\begin{aligned}
\Delta t_{\rm B}
&\approx& 2 \times 10^4 ~ {\rm sec} ~
(E_{\rm echo}/1 ~ {\rm GeV})^{-3/2} (B / 10^{-19} ~ {\rm G})^2
\nonumber \\
&& \times (r_{\rm coh}/1 ~ {\rm kpc}) (n_{\rm IR}/0.01 ~ {\rm cm}^{-3})^{-1},\end{aligned}$$ where $B$ is the field amplitude. Hereafter we take a fiducial value $r_{\rm coh} = 1~{\rm kpc}$ (see e.g. Langer et al. 2005), although the results can be trivially scaled for other values as it always occurs in the combination $B^2 r_{\rm coh}$ if $r_{\rm coh} \lesssim \lambda_{\rm IC, cool}$. The total delay time is approximately $\Delta t = \Delta t_{\rm ang} + \Delta t_{\rm B}$, and the magnetic field properties are reflected in the delay as long as $\Delta t_{\rm ang} \lesssim \Delta t_{\rm B}$.
To calculate the pair echo spectra and light curves, we follow Ichiki et al. (2008). First, the time-integrated flux of secondary pairs is $$\frac{dN_{e,{\rm 0}}}{d\gamma_e} (\gamma_e)
= 4 m_e
\frac{dN_{\gamma}}{dE_{\rm \gamma}}(E_{\gamma} = 2 m_e \gamma_e)
\left[1-e^{-\tau_{\gamma \gamma}(E_{\gamma} = 2 \gamma_e m_e)}\right],
\label{eq:dN0dgamma}$$ where $dN_{\gamma}/dE_{\gamma}$ is the primary gamma-ray fluence and $\tau_{\gamma \gamma}(E_\gamma)$ is the $\gamma\gamma$ optical depth in the CIB. The time-dependent pair-echo spectrum is $$\frac{d^2 N_{\rm echo}}{dt dE_\gamma}
= \int d\gamma_e \frac{dN_e}{d{\gamma_e}}
\frac{d^2 N_{\rm IC}}{dt dE_\gamma},$$ where $d^2 N_{\rm IC}/dt dE_\gamma$ is the IC spectrum from a single electron/positron, and $dN_e/d{\gamma_e}$ is the total flux of pairs relevant for the echo gamma-rays observed at time $t$. This formalism was extended to account for the finite probability of pair production near the observer in Takahashi et al. (2012) (see also Dai et al. 2002).
Weaker IGMFs generally lead to higher echo fluxes, as long as the time delay is dominated by $\Delta t_{\rm B}$ rather than $\Delta t_{\rm ang}$. For $r_{\rm coh} = 1~{\rm kpc}$, $\Delta t_{\rm B}$ is of the same order as $\Delta t_{\rm ang}$ if $B \sim 10^{-20}~{\rm G}$. Fig. \[fig:lc\_TeV-pa2\] compares the daily TeV counts with the $1-10~{\rm GeV}$ light curves of the pair echo during a period of 150 days for two values of $B$. While the two are generally correlated, for weaker IGMF, the peak flux of the echo is larger and its response to the primary emission is quicker. Although the magnetic deflection implies that the pair echo emission should also be spatially extended around the primary source, the extension is much smaller than the [*Fermi*]{} angular resolution and can be neglected for the field strengths of $B \sim 10^{-20}~{\rm G}$ considered here.
Statistical Analysis
====================
We now compare the expected pair echo with the [*Fermi*]{}-LAT data and derive constraints on the IGMF. Compared with our previous paper [@Takahashi3], we have a much greater number of independent flux bins (each representing the daily count), so a more sophisticated method of deriving the constraints is necessary. First, we compute the probability $P_i$ that a specific value of the the IGMF amplitude is excluded by the $i$-th flux bin, using the probability distribution function of the true flux obtained from the [*Fermi*]{}-LAT observation. Then, we combine the probabilities to derive the total probability $P_{\rm tot}$ using meta-analysis.
Note that it would not be appropriate to simply combine such probabilities for all bins. If the TeV flux for the $i$-th bin is low enough for the expected echo flux to be below the [*Fermi*]{} sensitivity for that bin, the probability $P_i$ would be small, irrespective of $B$. If we combine all such probabilities, the total probability $P_{\rm tot}$ can become so small that no constraints on $B$ can be obtained, even if some values of $P_i$ are sufficiently large for bins during TeV flares. Thus, we must select data bins for which the expected echo flux would be detectable by [*Fermi*]{}, depending on the assumed value of $B$. As explained above, larger $B$ results in a weaker echo that can only be detected for bins with higher TeV flux, so the number of such bins will be smaller. Here we set this selection threshold such that the echo flux exceeds the 2-$\sigma$ sensitivity of [*Fermi*]{}-LAT. In Fig. \[fig:lc\_sensitivity-UL\], this is compared with the echo light curves for $B = 10^{-20.5}~{\rm G}$ and $10^{-20}~{\rm G}$ at $1-10~{\rm GeV}$ during a particular 50-day period (only a small fraction of the entire data set). Here 4 and 3 bins exceed the [*Fermi*]{}-LAT sensitivity for $B = 10^{-20.5}~{\rm G}$ and $10^{-20}~{\rm G}$, respectively, which correspond to large TeV flares as seen in Fig. \[fig:lc\_TeV-pa2\].
Fig. \[fig:lc\_sensitivity-UL\] also plots the $50\%$ confidence [*Fermi*]{}-LAT upper limits on the daily flux. For the first flare (MJD 55147), the expected pair-echo flux for $B = 10^{-20.5}~{\rm G}$ greatly exceeds the upper limit, and the probability that this value of $B$ is excluded is very large. Although that for $B = 10^{-20}~{\rm G}$ also exceeds the limit, it does not reach the 2-$\sigma$ sensitivity, so the bin is not counted to compute $P_{\rm tot}$ for this $B$ value. For the second (MJD 55152) and third (MJD 55166) flares, the echo fluxes surpass the upper limits as well as the sensitivity for both $B = 10^{-20.5}~{\rm G}$ and $10^{-20}~{\rm G}$. For the fourth flare (MJD 55182), the echo flux for $B = 10^{-20.5}~{\rm G}$ is comparable to the $50\%$ confidence upper limit, neither favoring nor excluding this $B$ value, whereas that for $B = 10^{-20}~{\rm G}$ is not constrained by the limit and this $B$ value remains allowed.
We now consider the probability distribution function of the true flux and calculate the probability $P_i$ that it is less than the expected pair-echo flux for the $i$-th bin. To combine $P_i$, we use the inverse normal method, a type of meta-analysis. First, we derive the Z value of the normal distribution for the $i$-th bin, $Z_i$, which is the percentile (point) of the one-sided P value $P_i$. Note that $Z_i$ is negative if $P_i < 0.5$. Next, we compute the total Z value $Z_{\rm tot}$ as $$Z_{\rm tot} = \frac{1}{\sqrt{N}} \sum_{i=1}^{N} Z_i,
\label{eq:meta}$$ where $N$ is the number of the selected bins. Finally, we derive the one-sided P value $P_{\rm tot}$ of the normal distribution that corresponds to the above $Z_{\rm tot}$. We can interpret $P_{\rm tot}$ such that the assumed value of $B$ is excluded at a confidence level of $P_{\rm tot}$.
Fig. \[fig:constraint\] shows $Z_{\rm tot}$ as a function of $B$. For $B \leq 10^{-20.5}~{\rm G}$, the delay time of the pair echo is determined by angular spreading and becomes independent of $B$. Such weak IGMFs including $B=0$ is excluded by about 4-$\sigma$ significance. The significance decreases for larger $B$, and no constraints are obtained for $B \gtrsim 10^{-19.7}~{\rm G}$. This is a consequence of the lack of any time bins for which the pair-echo flux exceeds the 2-$\sigma$ [*Fermi*]{}-LAT sensitivity when $B \geq 10^{-19.5}~{\rm G}$.
Here we have not considered emission components other than the pair echo in the GeV band. In reality, there is likely to be primary GeV emission from the blazar, and possibly also other types of secondary GeV emission (e.g. Essey et al. 2011). If such components can be reliably accounted for, stronger upper limits on the pair echo and hence stronger lower bounds on the IGMF would be obtainable from the same [*Fermi*]{} data.
Discussion and Summary
======================
Using data from long-term, simultaneous GeV-TeV observations of Mrk 421 by [*Fermi*]{}-LAT and ARGO-YBJ, we have constrained the flux of secondary pair echos and derived lower bounds on the IGMF strength in the large void region lying between our Galaxy and Mrk 421. This was done by: 1) calculating the daily pair-echo flux from the TeV data over 600 days, 2) selecting the dates where the expected pair-echo flux exceeds the [*Fermi*]{}-LAT 2-$\sigma$ sensitivity, 3) computing the probability that an assumed value of the IGMF is excluded by the [*Fermi*]{}-LAT data for each date, and 4) combining these probabilities to derive the total probability using the inverse normal method. Consequently, as long as plasma instabilities are inconsequential, IGMFs weaker than $10^{-20.5}~{\rm G}$ are excluded by about 4-$\sigma$ for a field coherence length of 1 kpc. For general values of $r_{\rm coh}$, the derived constraint is $B \gtrsim 10^{-22}~\max[(r_{\rm coh}/350~{\rm kpc})^{-1/2},1]~{\rm G}$, where the latter case corresponds to IGMFs that are coherent over the IC cooling length.
Improving on our previous analysis using Mrk 501 [@Takahashi3], no assumptions are made here concerning the TeV emission during unobserved periods. The obtained constraints are thus more robust than from other studies, particularly those based on limits to the spatially-extended halo emission from secondary pairs that inevitably involves very long time delays, often longer than the typical lifetimes of blazars. Although the value of the lower limit obtained here is similar to our previous work, the statistical significance has increased remarkably, from less than 2-$\sigma$ to about 4-$\sigma$, thanks to the much larger data and improved statistical analysis.
In our study, the errors in the TeV flux, which propagate to the errors in the expected pair-echo flux, have not been considered. However, assuming a Gaussian distribution for the errors, the probability that the true echo flux is larger (or smaller) than the central value is $50\%$, so the errors in the expected echo flux should cancel out among the data bins to some extent and is unlikely to affect the total $P$ value significantly.
Here we have used the [*Fermi*]{}-LAT data only as daily upper limits to the GeV fluxes. Because the pair-echo flux is strongly dependent on the TeV flux, we can obtain potentially tighter constraints on the IGMF by investigating statistical correlations between the [*Fermi*]{}-LAT data and the ARGO-YBJ data. This will be presented elsewhere in the near future.
We thank Songzhan Chen who kindly provided us with the ARGO-YBJ data. This work is supported in part by the Grant-in-Aid from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan, No. 23740179, No. 24111710 and No. 24340048 (KT), No. 22540315 (MM), No. 24340048 (KI) and No. 22540278 (SI), and by the Grant-in-Aid for the global COE program “Quest for Fundamental Principles in the Universe: from Particles to the Solar System and the Cosmos” at Nagoya University from MEXT of Japan. The work of HT is supported by JSPS fellowship.
Abazajian, K. N., et al. 2009, ApJS 182, 543 Ando, S., & Kusenko, A. 2010, ApJ, 722, L39 Arlen, T. C., Vassiliev, V. V., Weisgarber, T., Wakely, S. P., & Yusef Shafi, S., ArXiv eprints, 1210.2802 Blanton, M., et al. 2005 AJ, 129, 2562 Bartoli, B., et al. 2011, ApJ, 734, 110 (B11) Broderick, A. E., Chang, P., & Pfrommer, C. 2012, ApJ, 752, 22 Chang, P., Broderick, A. E., & Pfrommer, C. 2012, ApJ, 752, 23 Chen, S. for the ARGO-YBJ Collaboration, private communication. Dai, Z. G., Zhang, B., Gou, L. J., M[é]{}sz[á]{}ros, P., & Waxmann, E. 2002, ApJL, 580, L7 Dermer, C. D., Cavadini, M., Razzaque, S., Finke, J. D., & Lott, B. 2011, ApJL, 733, L21 Dolag, K., Kachelrie[ß]{}, M., Ostapchenko, S., & Tom[á]{}s, R. 2011, ApJL, 727, L4 Elyiv, A., Neronov, A., & Semikoz, D. V. 2009, Phys. Rev. D, 80 023010 Essey, W., Ando, S. & Kusenko, A. 2011, Astropart. Phys., 35, 135 Franceschini, A., Rodighiero, G., & Vaccari, M. 2008, A[&]{}A, 487, 837 Gnedin, N. Y., Ferrara, A. & Zweibel, E. 2000, ApJ, 539, 505 Ichiki, K., Inoue, S., & Takahashi, K. 2008, ApJ, 682, 127 Ichiki, K., Takahashi, K., Ohno, H., Hanayama, H., & Sugiyama, N. 2006, Science, 311, 827 Langer, M., Aghanim, N., & Puget, J.-L. 2005, A[&]{}A, 443, 367 Miniati, F. & Elyiv, A. 2012, arXiv:1208.1761 Murase, K., Asano, K., & Nagataki, S. 2007, ApJ, 671, 1886 Murase, K., Takahashi, K., Inoue, S., Ichiki, K., & Nagataki, S. 2008, ApJ, 686, L67 Neronov, A., & Semikoz, D. V. 2009, Phys. Rev. D, 80, 123012 Neronov, A., Semikoz, D. V., Tinyakov, P. G., & Tkachev, I. I. 2011, A[&]{}A, 526, A90 Neronov, A., & Vovk, I. 2010, Science, 328, 73 Pfrommer, C., Chang, P., & Broderick, A. E., 2012, ApJ, 752, 24 Plaga, R. 1995, Nature, 374, 430 Razzaque, S., M[é]{}sz[á]{}ros, P., & Zhang, B. 2004, ApJ, 613, 1072 Ryu, D., Schleicher, D. R. G., Treumann, R. A., Tsagas, C. G., & Widrow, L. M. 2012, Space Science Reviews, 166, 1 Schlickeiser, R., Ibscher, D., & Supsar, M. 2012, ApJ, 758, 102 Takahashi, K., Ichiki, K., Ohno, H., & Hanayama, H. 2005, Phys. Rev. Lett., 95, 121301 Takahashi, K., Inoue, S., Ichiki, K., & Nakamura, T. 2011, MNRAS, 410, 2741 Takahashi, K., Mori, M., Ichiki, K., & Inoue, S. 2012, ApJL, 744, L7 Takahashi, K., Murase, K., Ichiki, K., Inoue, S., & Nagataki, S., 2008, ApJ, 687, L5 Tavecchio, F., Ghisellini, G., Bonnoli, G., & Foschini, L. 2010a, ArXiv e-prints, 2011, MNRAS, 414, 3566 Tavecchio, F., Ghisellini, G., Foschini, L., Bonnoli, G., Ghirlanda, G., & Coppi, P. 2010, MNRAS, 406, L70 Taylor, A. M., Vovk, I., & Neronov, A. 2011, A[&]{}A, 529, 9 Widrow, L. M. 2002, Reviews of Modern Physics, 74, 775 Widrow, L. M., Ryu, D., Schleicher, D. R. G., Subramanian, K., Tsagas, C. G., & Treumann, R. A. 2012, Space Science Reviews, 166, 37
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Given a semisimple stable autonomous tensor category over a field $K$, to any group presentation with finite number of generators we associate an element $Q(P)\in K$ invariant under the Andrews-Curtis moves. We show that in fact, this is the same invariant as the one produced by the algorithm introduced by Frank Quinn in [@Q:lectures]. The new definition allows us to present a relatively simple proof of the invariance and to evaluate $Q(P) $ for some presentations. On the basis of some numerical calculations over different Gelfand-Kazhdan categories, we make a conjecture which allows us to relate the value of $Q(P)$ for two different classes of presentations.'
address: 'Virginia Polytechnic Institute and State University Blacksburg, VA, 24061'
author:
- Ivelina Bobtcheva
date: 'April 1998.'
title: 'On Quinn’s Invariants of 2-dimensional CW-complexes'
---
[^1]
Introduction {#acmoves}
============
The original formulation of the Andrews-Curtis conjecture [@AC] has the following algebraic form. Suppose that $<x_1,x_2,\dots ,x_n|\; R_1,R_2,\dots ,R_n>$ is a presentation of the trivial group. Then this can be reduced to the empty presentation by a finite sequence of operations of the following types:
- The places of $R_1$ and $R_s$ are interchanged.
- $R_1$ is replaced with $gR_1g^{-1}$, where $g$ is any element in the group.
- $R_1$ is replaced with $R_1^{-1}$.
- $R_1$ is replaced with $R_1R_2$.
- Adding of an additional generator $y$ and an additional relator $y$.
- The reverse of (v).
We will refer to these six operations as AC-moves. The Andrews-Curtis conjecture is a particular case of the more general deformation conjecture stating that a simple homotopy equivalence of 2-complexes can be obtained by a deformation through 2-complexes (2-deformation [^2]). It is known that if two complexes $X$ and $Y$ are simple homotopy equivalent, then for some $k$, there exists a 2-deformation from the one point union of $X$ with $k$ copies of $S^{2}$ to the one point union of $Y$ with $k$ copies of $S^{2}$.
An algebraic formulation of this deformation conjecture comes from the following result [@Q:handle]: two 2-complexes are simple homotopy equivalent if and only if the corresponding group presentations have the same difference \# relation $-$\# generators, and they can be reduced to the same group presentation via the moves (i)-(vi) above plus the additional move
- $<x_1,x_2,\dots ,x_n|\; R_1,R_2,\dots ,R_m>$ is replaced with $<x_1,x_2,\dots ,x_n|\; Q_1,Q_2,\dots ,Q_m>$ if there are words $
w_{i,j}$, $v_{i,j}$ for $1\leq i\leq m$, $1\leq j\leq k(i)$, and indices $r(i,j)$, $s(i,j)$ such that for all $i$, $$R_{i}(Q_{i})^{-1}=\sqcap
_{j=1}^{k(i)}R_{i,j}Q_{i,j}R_{i,j}^{-1}Q_{i,j}^{-1},$$
where $R_{i,j}=w_{i,j}R_{r(i,j)}^{\pm 1}w_{i,j}^{-1}$ and $Q_{i,j}= v_{i,j}Q_{s(i,j)}^{\pm 1}v_{i,j}^{-1}$. Then, in its general form, the Andrews-Curtis conjecture states that two presentations which have the same difference \# relation $-$ \# generators and are equivalent through the moves (i)-(vii), are actually equivalent through the moves (i)-(vi).
In [@Q:lectures] an algorithm is described for computing topological quantum field theories on $1+1$-complexes associated to stable subcategories of the autonomous tensor categories of Gelfand and Kazhdan over the finite field $Z_{p}$. These two dimensional topological theories have the property that the invariant corresponding to $S^2$ is 0, and therefore give the opportunity of looking for counterexamples of the Andrews-Curtis conjecture. The invariants have been approached numerically [@Q:lectures; @thesis; @Q:num.pres.] where every CW-complex is being sliced in standard elementary slices. Then the invariant of the complex is obtained by composing the morphisms corresponding to each slice. Up to now no counterexamples have been detected, and the analytical value of the invariant even for simple complexes wasn’t known.
The goal of this work is to present a “global” definition of the invariant, in sense that, given a group presentation $P$ with a finite number of generators, we associate to it an element $Q(P)\in K$, which is invariant under the AC-moves. There is an explicit formula for $Q(P)$ in terms of some kind of trace (partial trace ) of a given morphism in the underlying semisimple stable autonomous tensor category .
The paper is structured in the following way. In section 2 some basic definitions and properties of semisimple stable autonomous tensor categories are presented. In section 3 the diagrammatic notation , which will be the main tool of manipulating the expressions is described . Section 4 is concerned with the definition and properties of partial traces in the category. The invariant $Q(P)$ is defined in section 5, and the proof of the invariance under the AC-moves is presented in section 6. Section 7 reviews the algorithm introduced in [@Q:lectures] and shows that it produces the same invariant. In the last section we list some conjectures on the dimension functions in the Gelfand-Kazhdan categories and on the basis of one of them show that if $P'$ is a presentation obtained from $P$ by adding a generator $y$ and a relator $xyx^{-1}y^{-1}$ (where $x$ is a generator of $P$), and if $P''$ is obtained from $P$ by simply adding a relator $x$, then $Q(P')=NQ(P'')$, where $N$ is the number of the simple objects in the category.
I would like to thank Frank Quinn for introducing me into the subject, and patiently explaining to me many basic concepts and ideas. His help and encouragement have been decisive for the completion of this work.
General category statements
===========================
is a tensor category if it is supplied with a bifunctor $\diamond :{\mbox {${\mathcal A}$}}\times {\mbox {${\mathcal A}$}}\rightarrow {\mbox {${\mathcal A}$}}$ and an identity object ${\mbox {${\bf 1}$}}$, such that for any $A,B,C\in {\mbox {${\mathcal A}$}}$ there are natural isomorphisms $$\begin{aligned}
&& \alpha _{A,B,C}:(A\diamond B)\diamond
C\rightarrow A\diamond (B\diamond C)\quad \mbox{(associativity morphisms)}, \\
&&\etaa{A} :{\mbox {${\bf 1}$}}\diamond A\rightarrow A, \quad\aeta{A} :A\diamond{\mbox {${\bf 1}$}}\rightarrow A,\\
&&\gamma =\gamma _{A,B} :A\diamond B\rightarrow B\diamond A
\quad \mbox{(commutativity morphisms)}.\end{aligned}$$ Moreover, $\gamma _{AB}\gamma_{BA} =id_{AB}$. These isomorphisms satisfy a number of axioms as listed in [@MacL:nat].
In order to simplify the notation, in the future we will often write $A B$ instead of $A\diamond B$.
\[bracketing\] Let be the free $(\otimes ,I)$-algebra on a single symbol $O$. The number of $O$’s in an element $T$ in will be called its length. Let now be a tensor category,. Following [@KL], we define a category ${\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}$, with objects $T[\underline{B}]\equiv T[B_{1}, B_{2},\dots ,B_{n}]$, where $T$ is an object of length $n$ in , and $\underline{ B}=(B_{1}, B_{2},\dots ,B_{n})$ is a sequence of objects in . The morphisms have the form $\pi [f_1, f_2, \dots , f_n]:T[B_{1}, B_{2},\dots ,B_{n}]\rightarrow
S[A_{1}, A_{2},\dots ,A_{n}]$, where $\pi $ is a bijection from the $O$’s in $T$ to the $O$’s in $S$, i.e. an element in the symmetric group ${\bf S}_n$, and $f_i:B_{\pi ^{-1}(i)}\rightarrow A_i$ are maps in . We use the following convention for the special maps: $$\begin{aligned}
&&T[f_1, f_2, \dots , f_n]\equiv \iota (n)[f_1, f_2, \dots , f_n]:
T[B_{1}, B_{2},\dots ,B_{n}]\rightarrow
T[A_{1}, A_{2},\dots ,A_{n}],\;\\
&&\pi\equiv \pi [id_{A_1},id_{A_2},\dots ,id_{A_n}]:
T[A_{\pi(1)}, A_{\pi(2)},\dots ,A_{\pi(n)}]\rightarrow
S[A_{1}, A_{2},\dots ,A_{n}],\end{aligned}$$ where $f_i:B_i\rightarrow A_i$, and $\iota (n)$ denotes the identity element ${\bf S_n}$. We will refer to the objects in ${\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}$ as bracketings in , and to the maps of type $\pi=\pi [id_{A_1},id_{A_2},\dots ,id_{A_n}]$ as permutations.
\[tenstrPB\] ${\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}$ possesses a structure of a tensor category with identity object $I[\; ]$, and product $\otimes $ defined in the following way: $$S[B_{1}, B_{2},\dots ,B_{n}] \otimes T[B_{n+1}, B_{2},\dots ,B_{m}]=
(S\otimes T)(B_{1}, B_{2},\dots ,B_{n+m}).$$ The associativity and commutativity morphisms are defined to be the corresponding permutations. Then the coherence theorem of MacLane [@MacL:nat] is equivalent to the assertion that there is a strict tensor functor ${\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}\rightarrow {\mbox {${\mathcal A}$}}$. We write $T(B_{1}, B_{2},\dots ,B_{n})$ and $\pi (f_1, f_2, \dots , f_n)$ for the images of $T[B_{1}, B_{2},\dots ,B_{n}]$ and $\pi [f_1, f_2, \dots , f_n]$ under this functor; for example $((O\otimes O)\otimes I)(a,b)=(a\diamond b)\diamond {\mbox {${\bf 1}$}}$. The images of the permutations will be called again permutations, and we will denote them with the corresponding cyclic decomposition. For example, $(1,2)(3,5,4): T(A,B,C,D,G)\rightarrow S(B,A,D,G,C)$.
In the case when $B_{i}=B$ for any $i$, we will use the convention $T(B^k)=T(B_{1}, B_{2},\dots ,B_{k})$. Then the permutations define an action of ${\bf S}_{k}$ on $T(B^k)$, and we denote the corresponding representation of ${\bf S}_{k}$ as $\rho [T(B^k)]$.
Let $P{\mbox {${\mathcal A}$}}$ be the category with the same objects as ${\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}$, but has different set of morphisms: for any morphism $T(\underline{B})
\rightarrow R(\underline{ B})$ in there is a morphism $T[\underline{ B}]
\rightarrow R[\underline{ C}]$, which will be denoted in the same way. The composition is obvious. Then there are functors $ {\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}\rightarrow P{\mbox {${\mathcal A}$}}$ and $P{\mbox {${\mathcal A}$}}\rightarrow {\mbox {${\mathcal A}$}}$, whose compositions is exactly the functor $ {\mbox {${\mathcal P}$}}\circ {\mbox {${\mathcal A}$}}\rightarrow {\mbox {${\mathcal A}$}}$ given by the coherence result of MacLane.
\[defn.satc\] An [*autonomous tensor*]{} category in the terminology of [@Shum] is a tensor category in which to every object $A$ is assigned an object $A^*\in {\mbox {${\mathcal A}$}}$ and morphisms $\Lambda _{A}:{\mbox {${\bf 1}$}}\rightarrow A^{*}A$ (coform) and $\lambda _{A}:AA^{*}\rightarrow {\mbox {${\bf 1}$}}$ (form) such that the compositions $$\begin{aligned}
&&A^* {\smash{
\mathop{\longrightarrow}\limits^{\etaa{A^*} ^{-1}}}} {\mbox {${\bf 1}$}}\, A^* {\smash{
\mathop{\longrightarrow}\limits^{\Lambda _{A}\diamond
id_{A^*}}}}
(A^{*}A)A^* {\smash{
\mathop{\longrightarrow}\limits^{\alpha }}} A^*(AA^*)
{\smash{
\mathop{\longrightarrow}\limits^{id_{A^*}\diamond \lambda _{A}}}} A^*\,{\mbox {${\bf 1}$}}{\smash{
\mathop{\longrightarrow}\limits^{\etaa{A^*}}}}
A^*, \\
&&A {\smash{
\mathop{\longrightarrow}\limits^{\aeta{A} ^{-1}}}} A\, {\mbox {${\bf 1}$}}{\smash{
\mathop{\longrightarrow}\limits^{id_{A}\diamond
\Lambda _{A}}}}
A(A^{*}A) {\smash{
\mathop{\longrightarrow}\limits^{\alpha ^{-1}}}} (AA^{*})A
{\smash{
\mathop{\longrightarrow}\limits^{ \lambda _{A}\diamond id_{A}}}} {\mbox {${\bf 1}$}}\,A
{\smash{
\mathop{\longrightarrow}\limits^{\etaa{A}}}} A \end{aligned}$$ act as identities. In any such category there is a canonical isomorphism $w_{A}:A^{**}\rightarrow A$, for which $
\gamma_{A^{*}A}\circ \Lambda _{A}=(w_{A}\diamond id)\circ\Lambda _{A^{*}}
$ We call an autonomous tensor category [*stable*]{} if $(A^{*})^{*}=A$ and $w_{A}=id_{A}$.
\[defn.semis\] A category is called [*semisimple*]{} if it is abelian over a field $K$, and if there is a finite subset of objects $\Sigma $ in , such that every other objects is isomorphic to a direct sum of objects in $\Sigma $, and for any $a,b\in\Sigma$, $hom _{{\mbox {${\mathcal A}$}}}(a,b)\simeq \left\{ \begin{array}{c} 0,\mbox{ if } a\neq b,\\
K,\mbox{ if } a= b.
\end{array}
\right. $. In general, we will use small latin letters $a,b,c,\dots $ to denote objects in $\Sigma $, and capital letters to indicate an arbitrary object in the category. We also use the notation $F(A,B)=hom_{{\mbox {${\mathcal A}$}}}(A,B)$.
The fact that an object $A$ is isomorphic to a direct sum of objects in $\Sigma $ means that for any $a \in \Sigma$ there exists a basis $\{\epsilon _i(a,A):a\rightarrow A\}_{i}$ of $F(a,A)$ and a basis $\{\epsilon _i(a,A)^*:A\rightarrow a\}_{i}$ of $F(A,a)$, which are dual in sense that $\epsilon _i(a,A)^*\circ \epsilon _j(a,A)=\delta _{i,j}id_a$. Moreover, $\sum _{a,j}\epsilon _j (a,A) \circ \epsilon _j(a,A)^*= id _{A}$. Then, given the objects $A,B$ and $C$, the map $$\begin{aligned}
\nabla : &\oplus _{b\in\Sigma} F(a,A\diamond (b\diamond C))\otimes
F(b,B)\rightarrow &F(a,A\diamond (B\diamond C)),\\
&\psi \otimes \phi\rightarrow &(id_A\diamond (\phi\diamond id_C))
\circ\psi\end{aligned}$$ is an isomorphism. In fact, the inverse of $\nabla$ is given by the map $$\begin{aligned}
&\nabla^{-1}:F(a,A\diamond (B\diamond C))\rightarrow
&\oplus _{b\in\Sigma} F(a,A\diamond (b\diamond C))\otimes F(b,B),\\
&\phi \rightarrow &\oplus _{b\in\Sigma ,i}
[(id_A\diamond (\epsilon _i(b,B)^*\diamond id_C))\circ \phi]\otimes
\epsilon _i(b,B).\end{aligned}$$
In a semisimple tensor category for any simple object $b$, the representation $\rho [T(a^k)]$ from \[tenstrPB\], induces representations $\rho _{b}[T(a^k)]$ of ${\bf S_{k}}$ on the space $F(b, T(a^k))$.
\[perp.cat.\] The concrete categories, used in the numerical computations of quantum invariants of 2-dimensional CW-complexes, are the ones studied in [@GK]. These are semisimple tensor categories, defined over the finite field $K=Z_p$, and are constructed as a quotient of the category of some modular finite-dimensional modules of a semisimple algebraic group $G$, in the case when $ p$ is greater then the Coxeter number of the corresponding Lie-algebra. If is such Gelfand-Kazhdan category, its simple objects correspond to the simple modules with highest weights inside the fundamental alcove. We remind that [*involution*]{} on a set $S$ is a bijection of $S$ into itself whose square is equal to the identity. Then on the set of simple objects $\Sigma $ there is an involution denoted by $a\rightarrow a^{*}$ , such that $a^{*}$ is isomorphic to the dual module of $a$. A form and coform can be introduced, and this makes an autonomous tensor category. But this category is not always stable, i.e. the condition $
\gamma _{a,a^{*}}\circ \Lambda _{a^{*}} =\Lambda _{a}:{\mbox {${\bf 1}$}}\rightarrow a^{*}\diamond a.
$ is not always satisfied. The simplest examples come from the categories of modular representations of $sl(2)$. In this case if $a$ is an odd highest weight module in $\Sigma $, $
\gamma _{a,a^{*}}\circ \Lambda _{a^{*}}=-\Lambda _{a}.
$ But the subcategory generated by the even highest weights modules is stable.
The general situation is not very different from the one of $sl(2)$, and is described, for example, in [@thesis]. It is shown that there is always a subcategory of which is stable, for example the one generated by the modules with highest weights in the root lattice. In general, there are number of sublattices of the weight lattice which produce subcategories with the desired property.
Diagrammatic notation
=====================
The diagrammatic notation used here is based on the coherence results in [@KL; @Shum; @FY]. The particular case of autonomous tensor categories is studied in [@KL], and the results there are extended to tortile categories in [@Shum; @FY]. The version we use here is the one described in [@Y], specified to the case of stable autonomous tensor category.
The idea behind the diagrammatic notation is that to each composition of ${\mbox {${\mathcal A}$}}$-morphisms there corresponds a labeled link diagram with coupons. Then there are rules for transforming these diagrams, and the statement is that if by such transformations the diagrams corresponding to $\psi
_{1}:A\rightarrow B$ and $\psi
_{2}:A\rightarrow B$ can be reduced to the same diagram, then $\psi _{1}=\psi _{2}$. In this sense, a morphism in is determined by its diagram.
\[invprop\] Suppose that the group $Z_2\times Z_2=<\theta _{1}, \theta _{-1}|\;
\theta _{1}^{2}, \theta _{-1}^{2}>$ acts on a set $S$. Let $S_{i}$, $i=\pm 1$, be the set of fixed points of the subgroup generated by $\theta _{i}$, and suppose that $S_{1}\cap S_{-1}=\emptyset $. Then if $\tilde{S}=S\setminus (S_{1}\cup S_{-1})$ and ${\mathcal O}$ is any orbit, we have that either ${\mathcal O}\setminus \tilde{S}$ is a two elements set, or ${\mathcal O}\subset \tilde{S} $.
The proof is straightforward.
The [*coupon category*]{} associated to a stable autonomous tensor category $({\mbox {${\mathcal A}$}}, \alpha ,\gamma ,\eta ,\lambda ,\Lambda ,*)$ is defined as a quotient of another category $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$. The objects of $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$ are finite sequences of objects in . Below, we will use the notation $\underline{A}$ both for the sequence $(A_{1},A_{2}, \ldots , A_{l})$ of length $|\underline{A}|=l$, and for the set of objects $\{A_i\}_{i=1}^{l}$ in . In particular, $\underline{A}\sqcup \underline{B}$ denotes the disjoint union of the two sets, and $\underline{A}\setminus \underline{B}$ denotes the complement of $\underline{B}$ in $\underline{A}$.
We introduce a $*$ operation on the objects in $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$: $
\underline{A}=(A_{1},A_{2}, \ldots ,
A_{l})\rightarrow \underline{A}^{*}=(A_{l}^{*},A_{l-1}^{*}, \ldots , A_{1}^{*})$. Then a morphism $(F,\theta ):\underline{X}\rightarrow \underline{Y}$ in $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$ is given by a finite set $F=\{f^i:T^{i}[\underline{A}^{i}]\rightarrow
R^{i}[\underline{B}^{i}]\}_{i}$ of morphisms in $P{\mbox {${\mathcal A}$}}$ , and an involution $\theta $ on the set $S_{F}=\underline{X}\sqcup \underline{Y}^{*}\sqcup
(\sqcup _{i}(\underline{A}^{i}{}^{*}\sqcup \underline{B}^{i}))$ such that $\theta (C)^{*}=C$ and $\theta$ is free of fixed points. The elements of $\underline{A}^{i}$ and $\underline{B}^{i}$ are called correspondingly the lower and the upper ends of $f^{i}$. The identity morphisms in $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$, $id :S\rightarrow S$ are given by $(\{\}, *)$.
Let now $(F,\theta ):\underline{X}\rightarrow \underline{Y}$ and $(G,\eta ):\underline{Y}\rightarrow
\underline{Z}$ be two morphisms in $\tilde {{\mbox {${\bf C}{\mathcal A}$}}}$. To define their composition, we would like to take $F\sqcup G$ with the involution map obtained by identifying the two copies of $\underline{Y}$, but in this way one obtains an involution, which may contain fixed points. To avoid this, identity morphisms are introduced as follows.
Let $S'= S_{F}\setminus \underline{Y}^{*}$ and $S''=S_{G}^{*}\setminus \underline{Y}^{*}$. We extend $\theta$ to an involution on $S=
S'\sqcup \underline{Y}^{*}\sqcup S''$ by defining $\theta$ to be the identity on the elements in $S''$. In a similar way we define an involution $\eta '$ on $S$ as $\eta' (x)=\left\{
\begin{array}{c} \eta (x^*)^{*} \mbox{ if } x\in \underline{Y}^{*}\sqcup S''\\
x \mbox{ otherwise }
\end{array}\right. $. Then $\theta$ and $\eta '$ define an action of $Z_{2}\times Z_{2}$ on $S$, which satisfy the requirements of proposition \[invprop\], where $\tilde{S}=\underline{Y}^{*}$, $S_1=S'$, and $S_{-1}=S''$. Then, according to the proposition, the orbits under this action are divided in two - the ones which are entirely contained in $\underline{Y}^{*}$, and ones which have exactly two elements outside this set. Let ${\bf E}$ be the subset of orbits which are entirely contained in $\underline{Y}^{*}$. If ${\mathcal O}\in {\bf E}$, let $A_{\mathcal O}$ be the smallest element in the orbit ( with respect to the ordering which comes with $\underline{Y}^{*}$), and let ${\bf T}=\{id_{A_{\mathcal O}}\}_{{\mathcal O}\in {\bf E}}$. Then composition of $(F,\theta )$ and $(G,\eta )$ is given by $(F\sqcup G\sqcup T, \xi )$, where $\xi$ is an involution on the set $S'\sqcup (S'')^{*}\sqcup
(\sqcup _{{\mathcal O}\in {\bf E}}\{A_{\mathcal O}^{*},A_{\mathcal O}
\})$ defined in the following way. Let $\imath :S'\sqcup (S'')^{*}
\stackrel{id\sqcup *}{\longrightarrow}
S'\sqcup S''\subset S$. Then if $A$ is an element in $S'\sqcup (S'')^{*}$, we define $\xi (A)=\imath ^{-1}(B)$, where $B$ is the unique element in $S'\sqcup S''$ which belongs to the same $Z_{2}\times Z_{2}$-orbit as $\imath (A)$. Moreover, $\xi (A_{\mathcal O})=A_{\mathcal O}^{*}$.
\[carel\] The coupon category is defined to be the quotient of $\tilde{{\mbox {${\bf C}{\mathcal A}$}}}$ under the following relations. Let $(F,\theta )$ be a morphism in $\tilde{{\mbox {${\bf C}{\mathcal A}$}}}$. Then the following replacements between morphisms in $F$ can be made:
- $id _{A}:O[A]\rightarrow O[A] $, whose ends are not mapped into each other, can be removed. Then $\theta $ changes to $\theta '$ where $\theta '(X)$ is equal to $\theta (X)$ if $X$ is not mapped into some of the ends of $id _{A}$, otherwise $\theta '(\theta (A))=\theta (A^{*})$.
- If $f:T[\underline{A}]\rightarrow Q[\underline{B}]$ and $g:Q[\underline{C}]\rightarrow R[\underline{D}]$ are such that $|\underline{B}|=|\underline{C}|$ and the upper ends of $f$ are mapped into the lower ends of $g$, i.e. $\theta (B_{i})=C_{i}^{*}$, then $f,g$ can be replaced with $f\circ g$.
- $f:T[\underline{A}]\rightarrow Q[\underline{B}
]\;\longleftrightarrow \;f'=\iota (l)^{-1}\circ f\circ \iota (k):
T'[\underline{A}]\rightarrow Q'[\underline{B}]$, where $k=|\underline{A}|$, $l=|\underline{B}|$, and $\iota (k):
T'(\underline{A})\rightarrow T(\underline{A})$ and $\iota (l):
Q'(\underline{B})\rightarrow Q(\underline{B})$ are the identity permutations as defined in \[bracketing\].
- $\gamma _{A,B}:(O\otimes O)[A,B]\rightarrow (O\otimes
O)[B,A]\;\longleftrightarrow \;
id :(O\otimes O)[A,B]\rightarrow (O\otimes O)[A,B]$.
- $\{f:T[\underline{A}]\rightarrow T'[\underline{B}],\,
g:Q[\underline{C}]\rightarrow Q'[\underline{D}]\}
\;\longleftrightarrow \;
g\diamond f: T[\underline{A}]\otimes Q[\underline{B}]\rightarrow
T'[\underline{C}]\otimes Q'[\underline{D}] $.
- $\eta _{A}:(O\otimes I)[A] \rightarrow O[A]
\;\longleftrightarrow \;
id_{A}: O[A]\rightarrow O[A]$, and the analogous statement for ${}_{A}\eta $.
- $\lambda _{A}:(O\otimes O)[A,A^{*}] \rightarrow
I[\{ \}]\;\longleftrightarrow \;
id_{A^{*}}:O[A^{*}]\rightarrow O[A^{*}]$.
- $\Lambda _{A}:I[\{ \}]\rightarrow(O\otimes O)[A^{*},A]
\;\longleftrightarrow \;
id_{A}: O[A]\rightarrow O[A]$.
The category is a stable autonomous tensor category. The product $\underline{A}\otimes \underline{B}$ is simply the sequence obtained by putting the elements of $\underline{B}$ after the elements of $\underline{A}$, the identity object is the empty set, and the associativities, $\eta _{\underline{A}}$ and ${}_{\underline{A}}\eta $ are the identity morphisms. $\gamma $, $\lambda $ and $\Lambda $ are given by $(\{ \}, *)$. Moreover, there is a strict tensor functor $P{\mbox {${\mathcal A}$}}\rightarrow {\mbox {${\bf C}{\mathcal A}$}}$ which maps $T[\underline{ B}]$ into $
\underline{ B}$ and $f:T[\underline{ B}]\rightarrow R[\underline{ C} ]$ into $(\{f\},*):\underline{ B}\rightarrow \underline{ C}$. Then the coherence result which underlines the use of the diagrams is that if $\psi ,\phi:
T(\underline{ B})\rightarrow R(\underline{ C})$ are two morphisms in which have the same image in , then $\psi =\phi$. In this way, given $T,R\in {\mbox {${\mathcal P}$}}$, every morphism $(F,\theta ):\underline{ B}\rightarrow
\underline{ C}$ determines a unique morphism $T(\underline{ B})\rightarrow R(\underline{ C})$ in , which also will be denoted as $(F,\theta )$.
A morphism $(F,\theta ):\underline{X}\rightarrow \underline{Y}$ in can be described by a diagram in the following way. First we draw a set of $|\underline{X}|+ |\underline{Y}|$ points labeled by the corresponding elements in $\underline{X}$ and $\underline{Y}$, and for each morphism $f:T[\underline{A}]\rightarrow
R[\underline{B}]$ we draw a rectangle (coupon) labeled by $f$ with $|\underline{A}|$ lines labeled by $A_i$’s, attached to its lower side, and $|\underline{B}|$ lines labeled by $B_j$’s, attached to its upper side. Then we connect with lines every two ends of coupons and labeled points which are mapped into each other by the involution $\theta $. In this way one looses any information about the elements in associated with the morphism of a coupon except their length, but note that according to \[carel\] (c) this is the only essential one. An example for the type of diagram obtained in this way is presented in figure \[gendiag\]. Then the equivalence relations in \[carel\] imply that one can perform the local diagram moves represented on figure \[relations\]. Note that if in the diagram there is a line which starts and finishes in a coupon and is labeled by the identity object , this line can be removed.
(5,3) (0,-0.3)
(15,12) (0,0)
We introduce a special notation for the basic elements of the homomorphism spaces. If $\{\epsilon _{i}\}_{i}$ is a basis of $F(b,
T(A_{1},A_{2},\dots , A_{l}))$ and $\{\epsilon _{i}^{*}\}_{i}$ is its dual basis of $F(T(A_{1},A_{2},\dots , A_{l}),b)$ (see \[defn.semis\]), we represent these basic elements in the diagrams as it is shown in figure \[basis\] (a) and (b). In the special and most often used case of a basis for $F(a,bc)$, where $a,b,c\in
\Sigma$, the notation will be simplified by replacing the coupon with a vertex labeled by $i$. Figure \[id.basis\] represents the defining identities of the dual basis. From the discussion in \[defn.semis\] it follows that if $\{\zeta _k\}_{k}$ is a basis for $F(a,bC)$ and $\{\epsilon _i\}_{i}$ is a basis for $F(b,AB)$, then as one varies $b,i,k$, the morphisms $(\epsilon _{i}\diamond id_{C})\circ \zeta _{k}$ form a basis of $F(a,(AB)C)$ (figure \[basis\] (c) ).
(12,2) (0,-0.3)
(12,3) (0,-0.3)
\[conj.basis\] Given a morphism $\varphi$ of $F(A,B)$, we define $$\hat{\varphi}= \eta _{A^{*}}(\lambda _{B^{*}}\diamond id_{A^{*}})
\alpha _{B^{*},B,A^{*}}^{-1}
(id_{B^{*}}\diamond (\varphi \diamond id_{A^{*}}))
(id_{B^{*}}\diamond\Lambda _{A^{*}})({}_{B^{*}}\eta )^{-1}\in F(B^*,A^*).$$ The corresponding diagram is presented in figure \[hatbasis\]. It is easy to see that if $\{\epsilon _i\}_i$ is a basis for $F(A,B)$, then $\{\hat{\epsilon }_{i}\}_i$ is a basis for $F(B^*,A^*)$ with dual $\{\hat{\epsilon ^{*}_{i}}\}_i$.
(4,2) (0,-0.3)
\[rank\] If $\varphi :A\rightarrow A$ is a morphism in , define the rank of $\varphi$ to be the morphism $r(\varphi)=\lambda _{A^*}\circ (id_{A^{*}}\diamond \varphi )\circ
\Lambda _{A}$, and $r(A)\equiv r(id_{A})$. We think of $r(\varphi)$ as an element in the field via the standard isomorphism $F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})\sim
K$. The diagrammatic notation for the rank of a morphism is presented in figure \[fig.rank\] (a). As an easy application of the diagrams one can see that
- $r(A)=r(A^{*})$ ;
- $r(T(A_{1},A_{2}, \ldots ,
A_{l}))=r(A_{1})r(A_{2}) \ldots r(A_{l})$;
\(a) allows as to use the simplified notation on figure \[fig.rank\] (b) for a rank of an object. The identity in figure \[fig.rank\] (c) follows from the fact that for any simple object $b$, $F(b ,b)\sim K$.
(9,1.5) (0,-0.3)
Let $\{\epsilon _{i}(a,bc)\}_{i}$ be a basis for $F(a,bc)$, and $\{\epsilon _{i}(a,bc)^*\}_{i}$ be the corresponding dual basis. Then the morphisms $\bar{\zeta
}_{i}(c,b^{*}a)=
\etaa{c}(\lambda _{b^*}\diamond id_{c})(id_{b^*}\diamond \epsilon _{i})$ and $\zeta _{i}(c,b^{*}a)=(id _{b^*}\diamond \epsilon _{i}^*)
(\Lambda _{b}\diamond
id_{c})\etaa{c}^{-1}$ represent sets of elements in $F(b^{*}a,c)$ and $F(c,b^{*}a)$ correspondingly.
\[dbasis\] $\{\zeta_{i}(c,b^{*}a)\}_i$ forms a basis for $F(c,b^*a)$ with a dual basis given by $\zeta _{i}(c,b^{*}a)^*=\frac{r(c)}{r(a)}\bar{\zeta}_{i}(c,b^{*}a)$. In particular, $\sum _{c, i} \frac{r(c)}{r(a)}(\zeta _{i}(c,b^{*}a)
\circ\bar{\zeta}_{i}(c,b^{*}a))=id_{b^*a}$.
The proof is presented in figure \[fig.dbasis\].
(9,2.2) (0,-0.3)
To make it easier in the future to construct diagrams and to study parts of them, we introduce the notion of an [*open morphism*]{}, correspondingly [*open diagrams*]{}. An open morphism in is a morphism $(F,\theta ):\underline{A}=(A_{1},A_{2}, \ldots ,A_{l})\rightarrow
\underline{B}=(B_{1},B_{2}, \ldots ,B_{k})$ where some subset $In(F)$ of the lower ends is labeled as $in$-ends, and some subset $Out(F)$ of the upper ends is labeled as $out$-ends. We refer to the set of $in$- and $out$- ends as [*free*]{} ends.
If $(F,\theta):\underline{A}\rightarrow \underline{B}$ and $(G,\xi ):\underline{C}\rightarrow \underline{D}$ are two open morphisms, and $S\subset Out(F)\cap In(G)$, we define the product of the two morphisms over $S$ to be the open morphism obtained by mapping the $out$-ends of $(F,\theta)$ labeled by $S$ into the $in$-ends of $(G,\xi )$ labeled by $S$ as shown on figure \[fig.prodopen\] (a). To be more precise, let $\underline{B}'=
\underline{B}\setminus S$ and $\underline{C}'=\underline{C}\setminus
S$. Then $(F,\theta)\circ _{S}(G,\xi ):\underline{A}\otimes
\underline{C}'\rightarrow \underline{B}'\otimes
\underline{D}$ with $In( G\circ _{S}F)=In(F)\sqcup
(In(G)\setminus S)$ and $Out( G\circ _{S}F)=(Out(F)\setminus S)\sqcup Out(G)$ is the following composition in : $$(F,\theta)\circ _{S}(G,\xi )=(id_{\underline{B}'}\otimes (G,\xi ))
\circ (\{\},\epsilon )\circ
((F,\theta)\otimes id_{\underline{C}'}),$$ where $\epsilon :\underline{B}\sqcup \underline{C}'\rightarrow
(\underline{B}'\sqcup \underline{C})^{*}$ is the involution map sending $S\subset \underline{B}\rightarrow
(S\subset \underline{C})^{*}$.
(16,6.5) (0,-0.3)
In the diagrams we indicate the $in$-ends with a circle and an arrow which comes out of it, and the $out$-ends with a circle and an arrow which goes into it. We will think of an open diagram as a sub-diagram of a bigger one, obtained by cutting out in the places of the free ends. For this reason, often the free ends will be left hanging without putting them in a specific place among the lower or the upper ends of the diagram, indicating in this way that the only important thing is how these will glue with the free ends of other open diagrams in the final “closed” one.
Given an open morphism $(F,\theta):\underline{A}\rightarrow \underline{B}$ we define the closure and the reverse of $(F,\theta)$. Given $S\subset In(F)\cap Out(F)$, the closure of $(F,\theta)$ over $S$ is defined to be the open morphisms $\circ _{S}(F,\theta )=(F,\theta ')$, where $\theta '$ is the extension of $\theta$ obtained by mapping into each other $S\subset \underline{A}$ and $S^*\subset \underline{B}^*$ as shown in figure \[fig.prodopen\] (b). By definition $In(\circ _{S}(F))=In(F)\setminus S$ and $Out(\circ _{S}(F))=Out(F)\setminus S$.
The reverse of $(F,\theta)$ is defined to be a morphism $(F,\theta)^{\hat{}}=(\hat{F},\hat{\theta})
:Out(F)^{*}\otimes (\underline{A}\setminus
In(F))\rightarrow (\underline{B}\setminus
Out(F))\otimes In(F)^{*}$ with $In(\hat{F})=Out(F)^{*}$ and $Out(\hat{F})=In(F)^{*}$. $(F,\theta)^{\hat{}}$ is obtained from $(F,\theta )$ by exchanging the places of the $in$- and $out$- ends through composition with the corresponding form and coform in as it is shown in figure \[fig.prodopen\] (c).
Partial traces
==============
Given a homomorphism $\phi :T(A_{1},A_{2},\dots ,A_{l})
\rightarrow T(A_{1},A_{2},\dots ,A_{l})$, $b\in \Sigma $ and a basis $\{\epsilon _i\}_{i}$ of $F(b,A_{k})$, the partial trace of $\phi $ as $A_{k}$ goes to $b$ is defined to be the morphism $$\begin{aligned}
&&{\tt Tr}_{A_{k}\rightarrow b}^{\epsilon _i}(\phi )
=\sum _{i} (id\diamond \epsilon ^{*}_i \diamond id
)\circ\phi\circ
(id\diamond \epsilon _i \diamond id):\\
&&\qquad\quad T(A_{1},\dots , A_{k-1}, b,A_{k+1}, \dots ,A_{l})\rightarrow
T(A_{1},\dots , A_{k-1}, b,A_{k+1}, \dots ,A_{l}),\end{aligned}$$ where $\{\epsilon ^{*}_i\}_{i}$ is the dual basis of $F(A_{k},b)$, and $(id\diamond \psi\diamond id)$ stands for $T(id_{A_{1}},\dots,id_{A_{k-1}},\psi ,id_{A_{k+1}},\dots
,id_{A_{l}})$. The corresponding diagram is presented in figure \[diagr.patr\] (a).
(14,3.5) (0,-0.3)
\[ptrprop\] Let $\phi $ be as above.
- If $\psi:A_{k}\rightarrow A'_{k}$ is an isomorphism and $\epsilon _i,\epsilon _i'$ are bases of $F(b,A_{k})$ and $F(b,A_{k}')$ correspondingly, then $${\tt Tr}_{A_{k}\rightarrow b}^{\epsilon _i}(\phi )={\tt Tr}_{A_{k}'\rightarrow
b}^{\epsilon _i'}((id\diamond \psi\diamond id)\circ\phi \circ
(id\diamond \psi ^{-1}\diamond id));$$
- If $k\neq m$, and $a,b\in \Sigma$, then $${\tt Tr}_{A_{m}\rightarrow a}({\tt Tr}_{A_{k}\rightarrow
b}(\phi )) ={\tt Tr}_{A_{k}\rightarrow b}({\tt Tr}_{A_{m}\rightarrow
a}(\phi ))\equiv {\tt Tr}_{\{A_{k}\rightarrow b,A_{m}\rightarrow
a\}}(\phi );$$
- If $A_{k}=R(B_{1},B_{2}, \ldots , B_{m})$, i.e. for some $T'$, $$T(A_{1},A_{2},\dots ,A_{l})=
T'(A_{1},A_{2},\dots , A_{k-1}, B_{1}, \ldots ,B_{m},A_{k+1}, \dots
,A_{l}),$$ then $${\tt Tr}_{A_{k}\rightarrow b}(\phi ) =
\sum _{a\in \Sigma }{\tt Tr}_
{R(B_{1}, \ldots ,B_{s-1},a, B_{s+1},\ldots , B_{m})\rightarrow b}(
{\tt Tr}_{B_{s}\rightarrow a}(\phi ));$$
- If $\pi$ is a permutation in ${\bf S}_l$, then $
\pi\circ {\tt Tr}_{A_{k}\rightarrow b}(\phi )\circ \pi ^{-1}=
{\tt Tr}_{A_{\pi (k)}\rightarrow b}
(\pi\circ \phi \circ \pi ^{-1}) .
$
\(a) follows from the semisimplicity of the category in the following way: $$\begin{aligned}
&&\sum _i(id\diamond (\epsilon '_i{}^* \circ \psi)\diamond id)\circ \phi\circ
(id\diamond (\psi ^{-1} \circ \epsilon '_i )\diamond id)=\\
&&\qquad =\sum _{i,j}(id\diamond (\epsilon '_i{}^* \circ \psi\circ \epsilon _j)
\diamond id)\circ (id\diamond \epsilon ^*_j \diamond id)\circ \phi\circ
(id\diamond (\psi ^{-1} \circ \epsilon '_i )\diamond id)=\\
&&\qquad =\sum _{i,j}(id\diamond \epsilon ^*_j \diamond id)\circ \phi\circ
(id\diamond (\psi ^{-1} \circ \epsilon '_i )\diamond id)\circ
(id\diamond (\epsilon '_i {}^* \circ \psi\circ \epsilon _j)
\diamond id)=\\
&&\qquad =\sum _j(id\diamond \epsilon ^*_j \diamond id)\circ \phi\circ
(id\diamond \epsilon _j \diamond id).\end{aligned}$$
\(b) follows directly from the definition. To prove (c), we observe that from \[defn.semis\] follows that if $\{\tau _{i}\}_{i}$ is a basis for $F(b,R(B_{1}, \ldots ,B_{s-1},a, B_{s+1},\ldots , B_{m}))$, and $\{\epsilon _{j}\}_{j}$ is a basis for $F(a,B_{s})$, then $\{R(id_{B_{1}},\ldots ,id_{B_{s-1}},\epsilon _{j},id_{B_{s+1}} \ldots,
id_{B_{m}})\circ \tau
_{i}\}_{a, i,j}$ is a basis for $F(b,A_{k})$. The use of this basis in the evaluation of the partial trace leads to the expression in (c). The statement in (d) is straightforward: every permutation is a product of ones of the form $(k-1,k)$, and for those is shown in figure \[diagr.patr\] (b).
As a consequence of (a) above, the indication of the basis in the notation of the partial trace will be omitted.
Definition of the invariant
===========================
To simplify the notation, we make a standard choice of bracketing. When $T$ is of the form $(\dots ((O\otimes O)\otimes O)\dots \otimes O)\otimes O$, we will write $T(B^k)$ as simply $B^k$, and $T(B_{1}, B_{2},\dots ,B_{k})$ as $B_{1}\diamond B_{2}\diamond\dots \diamond B_{k}$. In particular, $A(BC)B^{3}C=(((A,(BC))((B,B),B))C$.
\[fbl.prop\] We start by introducing some definitions. Given any $b\in\Sigma$ define $$b^l\equiv \left\{ \begin{array}{cc} b^l\mbox{ if } l> 0, \\
(b^{*})^{|l|}\mbox{ if } l< 0
\end{array}\right.$$ Then if $l>0$, let $f(b,l)=id_{b^{l+1}}:(b,b, \ldots,b)\rightarrow
(b,b, \ldots,b)$ be the open morphism in , where $In(f(b,l))$ is the first copy of $b$ in the lower ends and $Out(f(b,l))$ is the last copy of $b$ among the upper ends. For $l<0$, we define $f(b,l)=f(b^{*},|l|)^{\hat{}}:(b,b^*, \ldots,b^*)\rightarrow (b^*,
\ldots,b^*,b)\ $. The diagrams corresponding to $f(b,l)$ in both cases are presented in figures \[fbl\] (a) and (b). Note that with this definition, for any integer $l$, $f(b,l)^{\hat{}}=f(b^*,-l)$.
(11,3) (0,-0.3)
\[fblfbl\] Suppose we are given a nonzero integer $l$ and $b_{1},b_{2}, a_{i}\in \Sigma$, $1\leq i\leq |l|$. Let $
\chi :b_{1}b_{1}^lb_{2}b_{2}^l\rightarrow
(b_{1}b_{2})(\diamond _{i}(b_{1}b_{2})_{i})$ and $\chi ':b_{1}^lb_{1}b_{2}^lb_{2}\rightarrow
(b_{1}b_{2})(\diamond _{i}(b_{1}b_{2})_{i})$ be the corresponding permutations, where $$(b_{1}b_{2})_{i}=\left\{ \begin{array}{cc} b_{1}b_{2}\mbox{ if } l> 0, \\
b_{1}^{*}b_{2}^{*}\mbox{ if } l< 0
\end{array}\right. .$$ Then ${\tt Tr}_{\{(b_{1}b_{2})_{i}\rightarrow a_{i}\}_{i}}
(\chi '\circ(f(b_{1},l)\otimes f(b_{2},l))\circ\chi ^{-1}):
b_{1}b_{2}(\diamond _{i}a_{i})\rightarrow
b_{1}b_{2}(\diamond _{i}a_{i})$ is zero unless $a_{i}$ are not all the same, in which case is closely related to $f(a_{1},l)$. To be able to give the precise statement, let $\{\epsilon _{k}(a_{i},b_{1}b_{2}):a_{i}\rightarrow b_{1}b_{2}\}_{k}$ be a basis for $F(a_{i},b_{1}b_{2})$ and $\{\epsilon
_{k}(a_{i},b_{1}b_{2}))^*\}_{k}$ be its dual basis. We will denote with the same symbols the open morphism in : $(a_{i})\rightarrow (b_{1},b_{2})$ ($(b_{1},b_{2})\rightarrow (a_{i})$) which have a unique coupon labeled by $\epsilon _{k}(a_{i},b_{1}b_{2})$ (correspondingly $(\epsilon _{k}(a_{i},b_{1}b_{2}))^*$ ), and all lower ends are $in$, and all upper ends are $out$. Let $
f^{\epsilon}(a_{i},l)=\sum_{k}\epsilon _{k}(a_{i},b_{1}b_{2})^{*} \circ _{a_{i}}
f(a_{i},l)\circ _{a_{i}}\epsilon _{k}(a_{i},b_{1}b_{2}):
b_{1}b_{2}a_{1}^l\rightarrow a_{1}^lb_{1}b_{2}$. Then if $l>0$ we have $$\begin{aligned}
&&{\tt Tr}_{\{(b_{1}b_{2})_{i}\rightarrow a_{i}\}_{i}}
(\chi ' \circ(f(b_{1},l)\otimes f(b_{2},l))\circ\chi ^{-1})=
\sqcap _{i=1}^{l-1}
\delta _{a_{i},a_{i+1}}\sum _{k} \pi\circ f^{\epsilon}(a_{1},l).\end{aligned}$$ Here $\pi :a_1^lb_1b_2\rightarrow b_1b_2a_1^l$ is the corresponding permutation. The proof is shown in figure \[fptrA\].
(13,4.5) (0,0)
When $l<0$, the morphisms $\{
\zeta _{k}(a_{i}^{*},b_{1}^{*}b_{2}^{*})=\gamma _{b_{2}^{*},b_{1}^{*}} \circ
\hat{\epsilon} _{k} (a_{i},b_{1}b_{2})
\}_{k}$ form a basis for $F(a_{i}^{*},b_{1}^{*}b_{2}^{*})$. Then taking the partial trace with respect to this basis, also in this case, we obtain that $$\begin{aligned}
&&{\tt Tr}_{\{(b_{1}b_{2})_{i}\rightarrow a_{i}^{*}\}_{i}}
(\chi '\circ(f(b_{1},l)\otimes f(b_{2},l))\circ\chi ^{-1})=
\sqcap _{i=1}^{|l|-1}
\delta _{a_{i},a_{i+1}}\sum _{k}\pi\circ f^{\epsilon}(a_{1},l).\end{aligned}$$ The proof is shown in figure \[fptrB\].
(13,4.5) (0,0)
First we will define the invariant for the case of a presentation with one relation $P=<x_1,x_2,\dots ,x_n|\; R>$, where $$R=x_{i_{1}}^{l_{1}}x_{i_{2}}^{l_{2}} \ldots x_{i_{k}}^{l_{k}}.$$ To the relation $R$ and to any $b\in\Sigma$ we associate an element $R(b)=b^{l_{1}}b^{l_{2}} \ldots b^{l_{k}}\in{\mbox {${\mathcal A}$}}$ and a morphism $[R,b]:R(b)\rightarrow R(b)$ defined in the following way: $$[R,b]=\circ _{b}(r(b)\,f(b,l_{1})\circ _{b}f(b,l_{2})\circ _{b}\ldots
\circ _{b} f(b,l_{k})).$$ In other words, the $out$-end of $f(b,l_{i})$ is connected with the $in$-end of $f(b,l_{i+1})$ and at the end the morphism is closed by connecting the $out$-end of $f(b,l_{k})$ with the $in$-end of $f(b,l_{1})$. The result is multiplied by the rank of $b$. The open morphism without the final closure will be denoted with $[R,b]^{o}$. We will refer to the factors $b^l$ in $R(l)$ as $b$-factors. As examples, the diagrams corresponding to $[x^2y^{-2}x^{-1}y,b]^o$ and $[x^2y^{-2}x^{-1}y,b]$ are presented in figure \[exampl.diag\] (a) and (b).
(15,9) (0,-0.3)
\[onerel\] Now to the relation $R$, to $b\in\Sigma $, and to any generator $x_{k}$, we associate an element $g_{k}(R,b)\in{\mbox {${\mathcal A}$}}$, defined as the product of all b-factors in $R(b)$ corresponding to the generator $x_{k}:
g_{k}(R,b)=\diamond _{r|x_{i_{r}}=x_{k}}b^{l_{r}}$. Let $\kappa (R):R(b)\rightarrow g_{1}(R,b)g_{2}(R,b) \ldots
g_{n}(R,b)$ be the permutation of factors. Then the invariant of the presentation $P$ is defined to be: $$Q(P)=\sum _{b\in\Sigma }{\tt
Tr}_{\{g_{k}(R,b)\rightarrow {\mbox {${\bf 1}$}}\}_{k}}(\kappa (R)\circ[R, b]
\circ \kappa (R)^{-1}).$$ The diagram corresponding to $Q(<x,y|x^2y^{-2}x^{-1}y>)$ is presented in figure \[exampl.diag\] (c). The evaluation of the invariant for some simple examples is presented on figures \[examples\] and the results are collected below.
- $Q(<x|x^k>)=\sum _{b\in\Sigma}r(b){\tt Trace}
[\rho _{{\mbox {${\bf 1}$}}}[b^k](1,2,3,\ldots ,k)]$. In particular, $
Q(<x|x^2>)=\sum _{b\in\Sigma |b=b^{*}}r(b);
$
- The invariant for the standard presentation of the fundamental group of a surface of genus $n$ is $Q(<x_1,x_2,\dots ,x_{2n} |\;\sqcap
_{j=1}^{n-1}x_{2j}x_{2j+1}x_{2j}^{-1}x_{2j+1}^{-1}>)= \sum _{b\in\Sigma
}r(b)^{-2(n-1)}$;
- $Q(<x,y| xyxy^{-1}x^{-1}y^{-1}>)=\sum _{b\in\Sigma
|b=b^{*}}{\tt Trace}[\rho _{{\mbox {${\bf 1}$}}}[b^3](1,2,3)]$;
- $Q(<x,y| xyx^{-1}y>)=|\{b\in\Sigma
|b=b^{*}\}|$.
Here ${\tt Trace}$ denotes the usual trace of the corresponding element in the representation of the symmetric group.
(15,13) (0,-0.3)
\[morerel\] Now we extend the definition to the case of presentations with more then one relation. Let $P=<x_1,x_2,\dots ,x_n|\; R_1,R_2,\dots ,R_m>$ be such a presentation, and let $\underline{b}=(b_{1},b_{2} \ldots b_{m})\in \Sigma ^{\times m}$. Then define $$\begin{aligned}
&&Rel(\underline{b})=\diamond _{j=1}^m R(b_{j}),Ê\\
&&[P,\underline{b}]=[R_{1},b_{1}]\diamond [R_{2},b_{2}]\diamond \ldots
\diamond [R_{m},b_{m}]:Rel(\underline{b})\rightarrow Rel(\underline{b}),\\
&&G_{k}(\underline{b})=\diamond _{j=1}^m
g_{k}(R_{1},b_{1})g_{k}(R_{2},b_{2}) \ldots
g_{k}(R_{m},b_{m}),\\
&&\xi (P): Rel(\underline{b})\rightarrow \diamond _{k=1}^nG_{k}(\underline{b})
\mbox{ to be the corresponding permutation of $b$-factors}.\end{aligned}$$ Note that the permutation $\xi (P)$ groups the $b$-factors in $Rel(\underline{b})$ according to the generator to which they correspond. Then the invariant of the presentation $P$ is defined to be $$Q(P)=\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\xi(P)\circ[P, \underline{b}]
\circ \xi (P)^{-1}).$$ As example the diagram corresponding to $Q(<x,y|x^2y^2,
xyx^{-1}y^{-1}>)$ is shown in figure \[tworel\].
(12,4) (0,0)
$Q(P)$ is invariant under the AC-moves (i)-(vi).
The proof is contained in the next section.
Invariance under the AC - moves
===============================
We consider each move separately. The presentation obtained after performing the move will be denoted with $P'$ and the corresponding products of $b$-factors will be denoted with $Rel'(\underline{b})$ and $G_{k}'(\underline{b})$.
The first move consists in exchanging the places of $R_{1}$ and $R_{s}$. Let $$\underline{b}'=(b_{s}, b_{2},b_{3}, \ldots ,
b_{s-1},b_{1},b_{s+1}, \ldots, b_{m})$$ and let $\pi _{k}:G_{k}(\underline{b})\rightarrow
G_{k}'(\underline{b}')$ be the permutation which exchanges the places of $g_{k}(R_{1},b_{1})$ and $g_{k}(R_{s},b_{s})$. Then $\pi = \pi
_{1}\diamond \pi _{2}\diamond \ldots \diamond \pi _{n}:
\diamond _{k=1}^nG_{k}(\underline{b})\rightarrow
\diamond _{k=1}^nG_{k}'(\underline{b}')$. As morphisms in we have that $$\pi\circ \xi (P)\circ [P,\underline{b}]\circ \xi (P)^{-1}\circ \pi
^{-1}= \xi (P')\circ [P',\underline{b}']\circ \xi (P')^{-1}.$$
(9,6) (0,0)
This fact is illustrated in figure \[moveone\]. Then according to \[ptrprop\] (a) we have $$\begin{aligned}
&Q(P')&=\sum _{\underline{b}'\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}'(\underline{b}')\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\pi\circ\xi(P)\circ [P, \underline{b}]
\circ \xi (P)^{-1}\circ\pi ^{-1}) = \\
&&=\sum _{\underline{b}'\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\xi(P)\circ[P, \underline{b}]
\circ \xi (P)^{-1}) = Q(P).\end{aligned}$$ This shows the invariance under the first move.
The second move conjugates $R_{1}$ with an arbitrary element in the group. Obviously, it is enough to show the invariance under conjugation with one of the generators $R_{1}\rightarrow R_{1} '=x_{l}R_{1}x_{l}^{-1}$. Then $G_{k}$ changes only when $k=l$, and $G_{l}'(\underline{b})= b_{1}g_{l}(R_{1},b_{1})b_{1}^*g_{l}(R_{2},b_{2})\ldots
g_{l}(R_{m},b_{m})$. Let $$\pi :(\diamond _{k<l}G_{k}(\underline{b}))
G_{l}'(\underline{b})
(\diamond _{k>l}G_{k}(\underline{b}))
\rightarrow (\diamond _{k<l}G_{k}(\underline{b}))
b_{1}b_{1}^* G_{l}(\underline{b})
(\diamond _{k>l}G_{k}(\underline{b}))$$ be the corresponding permutation of factors. The goal now is to show that $$E={\tt Tr}_{G_{l}'(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}
(\xi(P')\circ[P', \underline{b}]\circ\xi (P')^{-1})=
{\tt Tr}_{G_{l}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}
(\xi(P)\circ[P, \underline{b}]\circ\xi (P)^{-1}),$$ which according to \[ptrprop\] (b) would imply the statement. By using properties \[ptrprop\] (a) and (c) of the partial trace we obtain $$\begin{aligned}
&E&={\tt Tr}_{b_{1}b_{1}^*G_{l}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}
(\pi \circ\xi(P')\circ[P', \underline{b}]\circ\xi (P')^{-1}\circ\pi ^{-1})=\\
&&=\sum _{a\in\Sigma}{\tt Tr}_{aG_{l}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}[
{\tt Tr}_{b_{1}b_{1}^*\rightarrow a}
(\pi \circ\xi(P')\circ[P', \underline{b}]\circ\xi (P')^{-1}\circ\pi ^{-1})].\end{aligned}$$ We remind that $[P',\underline{b}]=\diamond _{i=1}^{m}[R'_{i},b_{i}]$, but the partial trace ${\tt Tr}_{b_{1}b_{1}^*\rightarrow a}$ actually involves only the morphism $[R'_{1},b_{1}]$, in sense that $[R'_{1},b_{1}]$ maps $b_1R_1(b_1)b_1^*$ into itself, as shown on figure \[movetwo\] (a), and the effect of the permutation $\pi$ is to move the $b$-factors $b_1$ and $b_1^*$ together, so that the partial trace can be taken. Then figure \[movetwo\] (b) proves that $${\tt Tr}_{b_{1}b_{1}^*\rightarrow a}
(\pi \circ\xi(P')\circ[P', \underline{b}]\circ\xi (P')^{-1}\circ\pi ^{-1})=
\delta _{a,{\bf 1}}\,\nu \circ(id_{{\bf 1}}\diamond (\xi(P)
\circ [P, \underline{b}]
\circ\xi (P)^{-1}))\circ\nu ^{-1}.$$ Here $\nu: {\mbox {${\bf 1}$}}\, G_{1}(\underline{b})G_{2}(\underline{b}) \ldots
G_{2}(\underline{b})\rightarrow G_{1}(\underline{b})G_{2}(\underline{b}) \ldots
G_{l-1}(\underline{b}) \,{\mbox {${\bf 1}$}}\, G_{l}(\underline{b}) \ldots
G_{n}(\underline{b})$ is the corresponding permutation. The factor can be removed from ${\tt Tr}_{{\mbox {${\bf 1}$}}G_{l}(\underline{b})\rightarrow
{\mbox {${\bf 1}$}}}$ by conjugating the argument with $\etaa{G_{l}(\underline{b})}$ and using \[ptrprop\] (a). This completes the proof of the invariance under (ii).
(10,7.3) (0,0)
The third move consists of replacing $R_{1}$ with $R_{1}^{-1}$. First we observe that if $R=x_{i_{1}}^{l_{1}}x_{i_{2}}^{l_{2}} \ldots
x_{i_{s}}^{l_{s}}$ is a relation, then $$\begin{aligned}
&[R^{-1},b]&=r(b)\circ _{b}(f(b,-l_{s})\circ _{b}f(b,-l_{s-1})\circ _{b}
\ldots \circ _{b}f(b,-l_{1}))=\\
&&=r(b^{*})\circ _{b}(f(b^{*},l_{s})^{\hat{}}\circ _{b}
f(b^*,l_{s-1})^{\hat{}}\circ _{b}
\ldots \circ _{b}f(b^*,l_{1})^{\hat{}}\,)=\\
&&=\pi (R)\circ [R,b^*]\circ \pi (R)^{-1},\end{aligned}$$ where $\pi (R):b^{-l_{1}}b^{-l_{2}} \ldots b^{-l_{s}}\rightarrow
b^{-l_{s}}b^{-l_{s-1}} \ldots b^{-l_{1}}$ is the permutation of $b$-factors. The last equality is illustrated in figure \[movethree\].
(12,4) (0,-0.3)
In the case we want to study, $\pi (R_{1})$ induces a permutation $\pi = \pi (R_{1})\diamond id :R_{1}(b_{1}^{*})\diamond (\diamond
_{i=2}^{m}R_{i}(b_{i}))\rightarrow R_{1}^{-1}(b_{1})\diamond (\diamond
_{i=2}^{m}R_{i}(b_{i}))$. Then from the discussion above it follows that $$[P',\underline{b}]=\pi \circ [P,\hat{\underline{b}}]\circ \pi ^{-1},$$ where $\hat{\underline{b}}=(b_{1}^{*}, b_{2}, \ldots, b_{m})$. Moreover, $G_{k}'(\underline{b})$ can be obtained from $G_{k}(\hat{\underline{b}})$ by reversing the order of the $b$-factors of the form $b^{l_{j}}_{1}$. Let the corresponding permutation be $\chi _{k}: G_{k}'(\underline{b})\rightarrow
G_{k}(\hat{\underline{b}})$, and let $\chi =\chi _{1}\diamond \chi
_{2}\diamond \ldots \diamond \chi _{m}: \diamond _{k=1}^n
G_{k}'(\underline{b})\rightarrow \diamond _{k=1}^n G_{k}(\hat{\underline{b}})$. Then from property \[ptrprop\] we obtain $$\begin{aligned}
&Q(P')&=\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}'(\underline{b})\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\xi(P')\circ[P', \underline{b}]
\circ \xi (P')^{-1})=\\
&&=\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}(\hat{\underline{b}})\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\chi\circ \xi(P')\circ
\pi \circ [P,\hat{\underline{b}}]\circ \pi ^{-1}
\circ \xi (P')^{-1}\circ \chi ^{-1})=\\
&&=\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{\{G_{k}(\hat{\underline{b}})\rightarrow {\mbox {${\bf 1}$}}\}_{k}}
(\xi(P)\circ[P, \hat{\underline{b}}]
\circ \xi (P)^{-1})=Q(P).\end{aligned}$$ This completes the proof of the invariance under move (iii).
The forth move consists of replacing $R_{1}$ with $R_{1}R_{2}$. Figure \[movefourA\] presents the corresponding change in $[P,\underline{b}]$.
(12,2) (0,0)
Let $
R_{2}=x_{i_{1}}^{l_{1}}x_{i_{2}}^{l_{2}} \ldots
x_{i_{s}}^{l_{s}}
$, and $\chi : R_{1}(b_{1})R_{2}(b_{1})R_{2}(b_{2})\rightarrow
R_{1}(b_{1})\diamond _{i=1}^s \diamond _{j=1}^{|l_{i}|}(b_{1}b_{2})_{i,j}$ be the corresponding permutation, where $$(b_{1}b_{2})_{i,j}=\left\{\begin{array}{c} b_{1}b_{2} \mbox{ if }
l_{i}>0\\
b_{1}^{*}b_{2}^{*} \mbox{ if }
l_{i}<0
\end{array}\right. .$$ Then $\chi \circ([R_{1}R_{2},b_{1}]\diamond [R_{2},b_{2}])\circ \chi ^{-1}$ maps $R_{1}(b_{1})\diamond _{i=1}^s \diamond _{j=1}^{|l_{i}|}
(b_{1}b_{2})_{i,j}$ into itself. As a first step we prove that $$\begin{aligned}
&W&=\sum _{b_{2}\in\Sigma} {\tt Tr}_{\{(b_{1}b_{2})_{i,j}
\rightarrow a_{i,j}\}_{i,j}}
(\chi \circ([R_{1}R_{2},b_{1}]\diamond [R_{2},b_{2}])\circ\chi ^{-1})=\\
&&\qquad \qquad \sqcap _{i=1}^{s-1}(
\delta _{a_{i,1},a_{i+1,1}}
\sqcap _{j=1}^{|l_{i}|-1}
\delta _{a_{i,j},a_{i,j+1}})
([R_{1},b_{1}]\diamond [R_{2},a_{1,1}])\end{aligned}$$ Let $a_{i}\equiv a_{i,1}$, $1\leq i\leq s$. Recalling \[fbl.prop\], where the partial trace involving $f(b_1,l)\otimes f(b_2,l)$ was evaluated, and we have that $W=\circ
_{b_{1}}([R_{1},b_{1}]^{o}\circ _{b_{1}}G)$, where $$G=(\sqcap _{i=1}^{s-1}\sqcap _{j=1}^{|l_{i}|-1}
\delta _{a_{i,j},a_{i,j+1}})\sum _{b_{2}\in\Sigma} \circ _{b_{2}}
(f^{\epsilon}(a_{1},l_{1}) \circ _{\{b_{1},b_{2}\}}f^{\epsilon}(a_{2},l_{2})
\circ _{\{b_{1},b_{2}\}}
\ldots \circ _{\{b_{1},b_{2}\}}f^{\epsilon}(a_{s},l_{s})).$$ Then the statement for $W$ follows from the fact that $\epsilon _{k}(a_{i},b_{1}b_{2})^{*}\circ
\epsilon _{l}(a_{i+1},b_{1}b_{2})=\delta _{a_{i},a_{i+1}}\delta
_{k,l}\,id_{a_{i}}$ and from proposition \[dbasis\] as it is shown in figure \[movefourB\].
(9,6) (0,-0.3)
Now the invariance under move (iv) is easy to show. Fix $k$ between 1 and $n$. Then after the move, $G_{k}'(\underline{b})=g_{k}(R_{1},b_{1})g_{k}(R_{2},b_{1})
g_{k}(R_{2},b_{2})\ldots g_{k}(R_{m},b_{m})$. Let $G^{>2}_{k}(\underline{b})=\diamond _{i=3}^s g_{k}(R_{i},b_{i})$ and $g_{k}(R_{2},b)=\diamond _{r=1}^{s_{k}}b^{l_{i_{r}}}$. Then define $\tau _{k}: G_{k}'(\underline{b})\rightarrow
g_{k}(R_{1},b_{1})\diamond _{r=1}^{s_{k}}
\diamond _{j=1}^{|l_{i_{r}}|}
(b_{1}b_{2})_{i_{r},j}G^{>2}_{k}(\underline{b})$ to be the corresponding permutation, and let $\tau =\tau _{1}\diamond \tau _{2}\diamond \ldots \diamond \tau _{m}:
\diamond _{k=1}^{n}G_{k}'(\underline{b})\rightarrow
\diamond _{k=1}^{n} g_{k}(R_{1},b_{1})\diamond _{r=1}^{s_{k}}
\diamond _{j=1}^{|l_{i_{r}}|}
(b_{1}b_{2})_{i_{r},j}G^{>2}_{k}(\underline{b})$. Then from \[ptrprop\] (a) and (c) we obtain, $$\begin{aligned}
&&\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{G_{k}'(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}
(\xi(P')\circ[P', \underline{b}]
\circ \xi (P')^{-1})=\\
&&\quad =\sum _{\underline{b}\in\Sigma ^{\times m}}{\tt
Tr}_{g_{k}(R_{1},b_{1})\diamond _{r=1}^{s_{k}}
\diamond _{j=1}^{|l_{i_{r}}|}
(b_{1}b_{2})_{i_{r},j}G^{>2}_{k}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}
(\tau \circ \xi (P')\circ[P', \underline{b}]
\circ \xi (P')^{-1}\circ \tau ^{-1})=\\
&&\quad =\sum _{\underline{b}\in\Sigma ^{\times m}}
\sum _{\underline{a}_{r}\in\Sigma ^{\times s_{k}}}{\tt
Tr}_{g_{k}(R_{1},b_{1})\diamond _{r=1}^{s_{k}}
\diamond _{j=1}^{|l_{i_{r}}|}
a_{r,j}G^{>2}_{k}(\underline{b})\rightarrow {\mbox {${\bf 1}$}}}{\tt
Tr}_{\{(b_{1}b_{2})_{i_{r},j}\rightarrow a_{r,j}\}_{r,j}}\\
&&\qquad\qquad\qquad (\tau \circ \xi(P')\circ[P', \underline{b}]
\circ \xi (P')^{-1}\circ \tau ^{-1}).\end{aligned}$$ Now we observe that $$\tau \circ \xi(P')\circ[P', \underline{b}]
\circ \xi (P')^{-1}\circ \tau ^{-1}=
\xi (P)\circ (
(\chi \circ([R_{1}R_{2},b_{1}]\diamond [R_{2},b_{2}])\circ \chi ^{-1})
\diamond _{i=3}^m [R_i,b_i])\circ \xi (P) ^{-1}.$$ Here the use of $\xi (P)$ is somewhat abusive. By definition $\xi (P)$ is the permutation of $b$-factors $\diamond _{j=1}^m R(b_j)\rightarrow
\diamond _{k=1}^{n}G_k(\underline{b})$, while on the r.h.s. above the same notation is used to indicate the same permutation between expressions where each factors $b_2^{l_i}$ has been replaced with $\diamond _{j=1}^{l_i}(b_1b_2)_{i,j}$. Then the statement follows from \[ptrprop\] (d).
The invariance under the moves (v) and (vi) is straightforward: according to the definition, these moves change the value of the invariant by multiplication or division by $$\sum _{b\in\Sigma}r(b){\tt Tr}_{b\rightarrow {\mbox {${\bf 1}$}}}id _b =1.$$
The definition of Quinn’s invariant
===================================
In this section we show that the invariant defined above is actually the one produced by the algorithm described in [@Q:lectures]. First we introduce some morphisms which are being used in the algorithm.
- Given any $a,b,c\in\Sigma $ define $cycl(a,b,c) :F(a,b c)\rightarrow F(b^*,c a^*)$ as $$\begin{aligned}
&cycl(a,b,c)(\phi ):&b^{*}\rightarrow b^{*}{\mbox {${\bf 1}$}}{\smash{
\mathop{\longrightarrow}\limits^{\Lambda _{a^{*}}}}}
b^{*}(aa^{*}){\smash{
\mathop{\longrightarrow}\limits^{id_{b^{*}}\diamond (\phi\diamond id_{a^*})}}}\\
&&b^{*}((bc)a^{*})\rightarrow b^{*}b(ca^{*})
{\smash{
\mathop{\longrightarrow}\limits^{\lambda _{b^{*}}}}}{\mbox {${\bf 1}$}}(ca^{*})\rightarrow ca^{*}.\end{aligned}$$ $cycl(a,b,c)$ is actually an isomorphisms with inverse $
cycl(c^*,a^*,b)\circ cycl(b^*,c,a^*)$.
- Given any $a,b\in\Sigma $ and $A\in\Sigma$ define $$\begin{aligned}
&m_{b}:F(a,A)\longrightarrow F({\mbox {${\bf 1}$}},(b b^*) A),
& \psi\rightarrow (\Lambda _{b^*}\diamond \psi )\circ\etaa{a}^{-1} \\
&d_{b}:F({\mbox {${\bf 1}$}},(b b^*) A) \longrightarrow F(a,A),
&\phi \rightarrow \etaa{A}\circ (\lambda _b\diamond id_A)\circ
\phi.\end{aligned}$$
The corresponding diagrams are presented in figures \[diagr.md\].
(12,3) (0,0)
\[algorithm\] The invariant of a presentation $P=(x_1,x_2,\dots ,x_n|\, R_1,R_2,\dots, R_m)$ as defined in [@Q:lectures] is a map $Q(P):F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})\rightarrow F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})$, i.e. an element in $F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})\simeq
K$. This map is obtained as a composition of the morphisms which are listed below. Here we will refer to the space $V(k)=\oplus _{y_j\in\Sigma} F({\mbox {${\bf 1}$}},y_1y^*_1y_2y^*_2\dots y_ky^*_k)$ as the [*state space of $k$ generators*]{}, and to the summand $F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}}\dots {\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}})$ of $V(k)$ as the trivial summand of this state space.
- [**Beginning presentation.**]{} This morphism maps $F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})$ into $V(n)$ by embedding $F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})$ into the trivial summand. $$\begin{aligned}
K=&F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}})\rightarrow &\oplus _{y_j\in\Sigma}
F({\mbox {${\bf 1}$}},y_1y^*_1 y_2y^*_2\dots y_ny^*_n),\\
&1 \rightarrow &1\in F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}}\dots {\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}}).\end{aligned}$$
- [**Beginning relation.**]{} Starting a relation leads to the appearance of an additional generator, and therefore maps $V(n)$ into $V(n+1)$ by using the map $m_b$: $$\oplus _{y_j\in\Sigma} F({\mbox {${\bf 1}$}},y_1y^*_1y_2y^*_2\dots y_ny^*_n)
\stackrel{\oplus _br(b)^2m_{b^*}}{\longrightarrow}
\oplus _{b,y_j\in\Sigma}
F({\mbox {${\bf 1}$}},b^*by_1y^*_1y_2y^*_2\dots y_ny^*_n).$$
- [**The morphism corresponding to a factor $(y_i)^s$ in a relation.**]{} The main ingredient in it is the [*circulator*]{} which is the following map of $\oplus _{x\in\Sigma }F(a,bxx^*)$ in itself. $$\begin{aligned}
&CR(a,b): &\oplus _{x\in\Sigma }F(a,bxx^*)
\stackrel{(1,3,2)}{\longrightarrow}
\oplus _{x\in\Sigma }F(a,x(x^*b))
\stackrel{\nabla ^{-1}}{\rightarrow}\\
&&\oplus _{x,z\in\Sigma } F(a,xz^*)\otimes F(z^*,x^*b)
\stackrel{id\diamond cycl}{\longrightarrow}
\oplus _{x,z\in\Sigma } F(a,xz^*)\otimes F(x,bz)
\stackrel{\nabla }{\rightarrow}\\
&& \oplus _{z\in\Sigma }F(a,bzz^*). \end{aligned}$$ Here $\nabla $ is as in \[defn.semis\]. The diagram describing $CR(a,b)$ is presented in figure \[def.circ\]. We note that the circulator is actually an isomorphism and a specific expression for its inverse will be provided later.
(5.5,3.7) (0,-0.3)
Then a factor $(y_i)^s$ in a relation corresponds to the following composition of morphisms. First, in the state space of the $n+1$ generators, $y_iy^*_i$ is moved next to $b$. Then the part corresponding to $by_iy^*_i$ is separated from the rest by using the isomorphism $\nabla ^{-1}$, and the circulator $CR$ is applied $s$ times on it. Then following the inverse steps one goes back to the original form of the state space. $$\begin{aligned}
&&\oplus _{b,y_j\in\Sigma} F({\mbox {${\bf 1}$}},b^*by_1y^*_1
y_2y^*_2\dots y_ny^*_n)
\rightarrow \\
&&\oplus _{b,y_j\in\Sigma} F({\mbox {${\bf 1}$}},b^*(by_iy^*_i)y_1y^*_1\dots
\hat{y}_i\hat{y^*}_i\dots y_ny^*_n)
\stackrel{\nabla ^{-1}}{\rightarrow } \\
&&\oplus _{y_j\in\Sigma ,j\neq i}\oplus _{a,b\in\Sigma} F({\mbox {${\bf 1}$}},b^*ay_1y^*_1\dots
\hat{y}_i\hat{y^*}_i\dots y_ny^*_n) \otimes
(\oplus _{y_i\in\Sigma }F(a,by_iy^*_i))
\stackrel{id\otimes (CR(b,a))^s}{\longrightarrow}\\
&&\oplus _{y_j\in\Sigma ,j\neq i}\oplus _{a,b\in\Sigma} F({\mbox {${\bf 1}$}},b^*ay_1y^*_1\dots
\hat{y}_i\hat{y^*}_i\dots y_ny^*_n) \otimes
(\oplus _{y_i\in\Sigma }F(a,by_iy^*_i))
\stackrel{\nabla}{\rightarrow } \\
&&\oplus _{b,y_j\in\Sigma} F({\mbox {${\bf 1}$}},b^*(by_iy^*_i)y_1y^*_1\dots
\hat{y}_i\hat{y^*}_i\dots y_ny^*_n)
\rightarrow
\oplus _{b,y_j\in\Sigma} F({\mbox {${\bf 1}$}},b^*by_1y^*_1
y_2y^*_2\dots y_ny^*_n),\end{aligned}$$ where $\hat{y}$ indicates that the corresponding term is missing.
- [**Ending relation.**]{} This is a map from $V(n+1)$ generators back into $V(n)$ proportional to the one induced by $d_b$. $$\oplus _{b,y_j\in\Sigma}
F({\mbox {${\bf 1}$}},b^*by_1y^*_1y_2y^*_2\dots y_ny^*_n)
\stackrel{\oplus _br(b)^{-1}d_b}{\longrightarrow}
\oplus _{y_j\in\Sigma} F({\mbox {${\bf 1}$}},y_1y^*_1y_2y^*_2\dots y_ny^*_n).$$
- [**Ending presentation.**]{} This is a map from $V(n)$ generators back to the ring $V(0)$, which is injective on the trivial summand on the state space and sends all other summands of $V(n)$ to 0. $$\oplus _{y_j\in\Sigma}
F({\mbox {${\bf 1}$}},y_1y^*_1 y_2y^*_2\dots y_ny^*_n)
\rightarrow F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}}) ,\;
1\in F({\mbox {${\bf 1}$}},{\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}}\dots {\mbox {${\bf 1}$}}\,{\mbox {${\bf 1}$}})\rightarrow 1.$$
\[qequiv\] Given a group presentation $P$, the invariant of $P$, produced by the above algorithm, is exactly $Q(P)$.
The key is to understand the powers of the circulator. Given $b\in \Sigma$ and a nonzero integer $l$, we use the definitions of $b^l$ and $f(b,l)$ made in \[fbl.prop\]. Let $x,w\in \Sigma$, $\{\epsilon _i\}_{i}$ be a basis for $F(w^*,x^*b^l)$ and let $\{\epsilon _i^{*}\}_{i}$ be its dual. We define morphisms $\phi (l,b,x,w,i): bxx^*\rightarrow x^*(bb^l)w$ and $\psi (l,b,x,w,i): x^*(b^lb)w\rightarrow bww^*$ in the following way: $$\begin{aligned}
&\phi (l,b,x,w,i): &bxx^*\rightarrow bxx^*{\mbox {${\bf 1}$}}\stackrel{(1,2,3)\diamond \Lambda _w}
{\longrightarrow }(x^*bx)(w^*w)
\stackrel{id\diamond (\epsilon _i\diamond id_w)}{\longrightarrow }\\
&&\rightarrow (x^*bx)(x^*b^lw)\rightarrow x^*(b(xx^*)b^l)w
\stackrel{\lambda _x}{\longrightarrow } x^*(bb^l)w, \\
&\psi (l,b,x,w,i): &x^*(b^lb)w\rightarrow (x^*b^l)bw
\stackrel{\epsilon _i^{*}}{\longrightarrow } w^*bw
\stackrel{(1,3,2)}{\longrightarrow }bww^*.\end{aligned}$$
\[powcirc\] Given $f\in F(a,bxx^*)$, $
CR(b,a)^l(f)=\oplus _{w\in\Sigma} cr(l,b,x,w)\circ f,
$ where $cr(l,b,x,w)=\sum _i \psi(l,b,x,w,i)\circ (id_{x^*}\diamond
f(b,l)\diamond id_w)\circ \phi (l,b,x,w,i)$.
The corresponding diagram is shown in figure \[circ.pow\]. Note that $cr(l,b,x,w)$ can actually be written as the composition of a partial trace and another morphism, and therefore it is independent on the particular choice of the basis $\{\epsilon _i\}_{i}$.
(9,6) (0,0)
The fact that the circulator $CR(a,b):F(a,bxx^*)\rightarrow F(a,bww^*)$ acts as $cr(1,b,x,w)$ follows from the comparison of the diagrams in figure \[def.circ\] and \[circ.pow\] (a). Then in figure \[inv.circ\] we prove that $cr(-1,b,x,w)$ is a right inverse of the circulator. In a similar fashion one can see that it is a left inverse as well. Then, by assuming that for $l>1$, the $(l-1)$-th power of the circulator acts as $cr(l-1,b,x,w)$, figure \[ind.circ\] proves that the statement is also true for the $l$-th power. The inductive proof for the negative powers of the circulator goes in a similar fashion.
(9,11) (0,-0.3)
(15,8) (0,-0.3)
Now theorem \[qequiv\] follows from the following statement. Let $P$ be a presentation with $n$ generators, $R=x_{i_{1}}^{l_{1}}x_{i_{2}}^{l_{2}}\ldots x_{i_{s}}^{l_{s}}$ be a relation in $P$, and given $\underline{y},\underline{z}\in\Sigma ^{\times n}$ and $b\in\Sigma$, let $\{\epsilon _{j}^{i}(z^*_{i},y^*_{i}g_i(R,b))\}_{j}$, $i=1\div n$, be a set of bases for the homomorphisms spaces $F(z^*_{i},y^*_{i}g_i(R,b))$. Then the map $V(n+1)\rightarrow V(n+1)$, which the relation $R$ induces according to the algorithm in \[algorithm\], is given by $
\varphi \rightarrow \oplus _{\underline{y},\underline{z},j}
rl(\underline{y},\underline{z},j)\circ \varphi,
$ where $rl(\underline{y},\underline{z},j)$ is the following composition of maps: $$\begin{aligned}
&&b^{*}by_{1}y_{1}^{*}y_{2}y_{2}^{*}\dots y_{n}y_{n}^{*}
\rightarrow
b^{*}(\diamond _{i}y_{i}^{*})b(\diamond _{i}y_{i})
\stackrel{id\diamond (\diamond _{i}\Lambda
_{z_{i}})}{\longrightarrow}\\
&&b^{*}(\diamond _{i}y_{i}^{*})b(\diamond _{i}y_{i})
(\diamond _{i}z_{i}^{*}z_{i})
\stackrel{id\diamond (\diamond _{i}(\epsilon ^{i}_{j}\diamond
id_{z_{i}}))}
{\longrightarrow}\\
&& b^{*}(\diamond _{i}y_{i}^{*})b(\diamond _{i}y_{i})
(\diamond _{i}y_{i}^{*}g_i(R,b)z_{i})
\rightarrow
b^{*}(\diamond _{i}y_{i}^{*})b(\diamond _{i}y_{i}
y_{i}^{*})(\diamond _{i}g_i(R,b))(\diamond _{i}z_{i})
\stackrel{id\diamond (\diamond _{i}\lambda _{y_{i}})\diamond id}
{\longrightarrow}\\
&&b^{*}(\diamond _{i}y_{i}^{*})b(\diamond _{i}g_i(R,b))(\diamond _{i}z_{i})
\stackrel{id\diamond \phi(R)\diamond id}{\longrightarrow}
b^{*}(\diamond _{i}y_{i}^{*})(\diamond _{i}g_i(R,b))b(\diamond _{i}z_{i})
\rightarrow\\
&&b^{*}(\diamond _{i}y_{i}^{*}g_i(R,b))b(\diamond _{i}z_{i})
\stackrel{id\diamond (\diamond _{i}\epsilon ^{i}_{j}{}^*)\diamond id}
{\longrightarrow}
b^{*}(\diamond _{i}z^{*}_{i})b(\diamond _{i}z_{i})
\rightarrow
b^{*}bz_{1}z_{1}^{*}z_{2}z_{2}^{*}\dots z_{n}z_{n}^{*},\end{aligned}$$ where $\phi(R)=(id_{b}\diamond \kappa (R))[R,b]^{o}(\kappa
(R)^{-1}\diamond id_{b}):b(\diamond _{i}g_i(R,b))\rightarrow
(\diamond _{i}g_i(R,b))b$ (see \[onerel\]). The corresponding diagram is presented in figure \[endstep\], where for simplicity we have restricted ourselves to the case of two generators. The statement is proved by induction over $s$. For $s=1$ it is reduced to proposition \[powcirc\]. Assume it is true for $R$, and let $R'=Rx_{k}^{l}$ for some $k$ and $l$. Then $$\begin{aligned}
&g_{i}(R',b)=\left\{\begin{array}{c}g_{k}(R,b)b^{l} \mbox{ if } i=k,\\
g_i(R,b) \mbox{ otherwise }
\end{array}\right. ,
&\mbox{ and }\quad [R',b]=\circ _{b}([R,b]^{o}\circ _{b}f(b,l)).\end{aligned}$$ Then the statement follows from figure \[endind\].
(12,9) (0,0)
(15,8) (0,-0.3)
Conjectures and consequences
============================
As we said, this work was inspired by an extensive numerical study of the invariants defined in [@Q:lectures]. The numerical project has been carried out by Frank Quinn, the author, and Luoqi Zhang, and the autonomous tensor categories used in it are subcategories of the Gelfand-Kazhdan categories, as described in \[perp.cat.\]. We remind that such a subcategory is determined by a type and rank of simple Lie algebra ($A,B,C,D,E$ or $F$), a prime number $p$, and an invariant sublattice of the weight lattice (containing the root lattice). But actually many of the categories corresponding to different algebras are equivalent. Conjecturally, any two autonomous tensor categories with sets of simple objects $\Sigma _1$ and $\Sigma _2$, such that there is a bijection $\mu :\Sigma _1\rightarrow \Sigma _2$, and $dim(F(a,bc))=
dim(F(\mu(a),\mu(b)\mu(c)))$, are equivalent. In other words, the dimensions of the homomorphism spaces determine the category. Let $dim(a,bc)$ denote the dimension of $F(a,bc)$ and we will refer to those as dimension functions. On web page http://www.math.vt.edu/quantum\_topology has been collected a list with the values of the dimension functions for some categories corresponding to algebras type $A,B,C,D$, small primes, and either the full weight lattice or the root lattice. Let $L$ be the order of the abelian group obtained as a quotient of the lattice over which the category is defined, modulo the root lattice. Then all numerically generated examples satisfy the following conjecture:
\[conj1\]In the case of stable Gelfand-Kazhdan categories, for any $b\in\Sigma$
- $dim(a,b^{Lp})= 0 \quad (\mbox{mod $ p$}),\mbox{ if } a\neq {\mbox {${\bf 1}$}}$;
- $r(b)^{-1}\sum _{c\in\Sigma}dim(b,cc^*) =\left\{
\begin{array}{c} |\Sigma | \mbox{ if $b$ belongs to the root lattice,}\\
0 \mbox{ otherwise}
\end{array} \quad (\mbox{mod $ p$})\right.$.
We make few comments on this. From \[rank\] and (a) above we obtain $$dim({\mbox {${\bf 1}$}},b^{Lp})=\sum _{a\in\Sigma}dim(a,b^{Lp})r(a)=r(b^{Lp})=
r(b)^{Lp}=r(b)^L, \quad (mod\; p).$$ In connection with (b) we comment that the statement can be proved easily for algebra type $A_1$. Moreover, the 0 part can be derived from a conjecture of Frank Quinn, according to which the category over an arbitrary lattice is a product of the category over the root lattice and a finite group category. If this is true, it would imply that the only interesting categories are the ones coming from the root lattice, since a finite group category brings to a classical invariant. For categories over the root lattice, the conjecture \[conj1\] brings to the following result.
\[corr\]Let $P=<x_1,x_2, \dots ,x_n|R_1,R_2,\dots ,R_m>$ be a group presentation. If we are working with a Gelfand-Kazhdan category over the root lattice, and \[conj1\] (b) is true, then $
Q(P')=|\Sigma|\, Q(P''),
$ where $$\begin{aligned}
&&P'=<x_1,x_2, \dots ,x_n,y|R_1,R_2,\dots ,R_m, x_kyx_k^{-1}y^{-1}>\mbox{ and}\\
&&P''=<x_1,x_2, \dots ,x_n|R_1,R_2,\dots ,R_m, x_k>.\end{aligned}$$
Let $\underline{b}\in\Sigma ^{\times m}$, and $a,c\in\Sigma$. If $(\underline{b},c)$ denotes the sequence $(b_{1},b_{2}, \ldots
,b_{m},c)$ we have that $$\begin{aligned}
&&[P',(\underline{b},c)]=
[P,\underline{b}]\diamond
[x_kyx_k^{-1}y^{-1},c]:Rel(\underline{b})ccc^{*}c^{*}\rightarrow
Rel(\underline{b})ccc^{*}c^{*},\\
&&G_l'(\underline{b},c)=\left\{\begin{array}{l}
G_k(\underline{b}) (cc^*)\mbox{ if }l=k,\\
G_l(\underline{b})\mbox{ otherwise,
}\end{array}\right.\quad\mbox{ and }\quad
G_y'(\underline{b},c)=cc^*, \\
&&\xi (P'): Rel(\underline{b})ccc^*c^*
\stackrel{\xi (P)\diamond (2,3)}{\longrightarrow}
(\diamond _{l}G_{l}(\underline{b}))cc^*cc^*
\rightarrow \\
&&\qquad \qquad\qquad\rightarrow
(\diamond _{l<k}G_{l}(\underline{b}))G_k(\underline{b}) (cc^*)
(\diamond _{l>k}G_{l}(\underline{b}))cc^*.\end{aligned}$$ Moreover, $$\begin{aligned}
&&[P'',(\underline{b},a)]= r(a)
[P,\underline{b}]\diamond
id_{a}:Rel(\underline{b})a\rightarrow
Rel(\underline{b})a,\\
&&G_l''(\underline{b},a)=\left\{\begin{array}{l}
G_k(\underline{b}) a\mbox{ if }l=k,\\
G_l(\underline{b})\mbox{ otherwise, }\end{array}\right.\\
&&\xi (P'')= Rel(\underline{b})a
\stackrel{\xi (P)\diamond id_{a}}{\longrightarrow}
(\diamond _{l}G_{l}(\underline{b}))a
\rightarrow
(\diamond _{l<k}G_{l}(\underline{b}))G_k(\underline{b}) a
(\diamond _{l>k}G_{l}(\underline{b})),\end{aligned}$$ where we have used the notations in \[morerel\]. Then by using \[ptrprop\] (c) and (d) we obtain the following for the part of the partial trace involving $G_k'(\underline{b},c)$ and $G_y'(\underline{b},c)$: $$\begin{aligned}
&W&=\sum _{c\in\Sigma}{\tt Tr}_{\{G_k'(\underline{b},c)\rightarrow {\mbox {${\bf 1}$}},
\;G_y'(\underline{b},c)\rightarrow {\mbox {${\bf 1}$}}\}}
(\xi (P')\circ ([P,\underline{b}]\diamond [x_kyx_k^{-1}y^{-1},c])\circ
\xi (P')^{-1})=\\
&&=\sum _{c,a\in\Sigma}{\tt Tr}_{G_k(\underline{b})a\rightarrow {\mbox {${\bf 1}$}}}(
\xi (P'')\circ ([P,\underline{b}]\diamond K)\circ\xi (P'')^{-1}),\end{aligned}$$ where $K={\tt Tr}_{\{cc^*\rightarrow a, cc^*\rightarrow {\mbox {${\bf 1}$}}\}}(
(2,3)\circ[x_kyx_k^{-1}y^{-1},c]\circ (2,3))$. Figure \[fig.corr\] shows that in fact, $K=dim(a,cc^*)id_a\diamond id_{{\mbox {${\bf 1}$}}}$. Hence we obtain : $$\begin{aligned}
&W&=\sum _{c,a\in\Sigma}dim(a,cc^*){\tt Tr}_{G_k(\underline{b})a\rightarrow {\mbox {${\bf 1}$}}}(
\xi (P'')\circ ([P,\underline{b}]\diamond id_{a})\circ\xi (P'')^{-1})=\\
&&=|\Sigma |\sum _{a\in\Sigma}{\tt Tr}_{G_k(\underline{b})a\rightarrow {\mbox {${\bf 1}$}}}(
\xi (P'')\circ [P'',(\underline{b},a)]\circ\xi (P'')^{-1}),\end{aligned}$$ which is exactly the part of the partial trace in the definition of $Q(P'')$ involving $G_k''(\underline{b},a)$.
(10,3) (0,-0.3)
Another conjecture we would like to discuss concerns the order of the circulator. It is observed in [@Q:lectures], that for all numerically generated categories, the order of the circulator is in fact $Lp$. On the basis of this and the expression for the powers of the circulator \[powcirc\] we may state the following
\[conj2\] Let $\sigma^{Lp+1}(b)=(1,2,3,\dots ,Lp+1):bb^{Lp}\rightarrow
bb^{Lp}$. Then in the case of stable Gelfand-Kazhdan categories, $${\tt Tr}_{b^{Lp}\rightarrow w}\sigma^{Lp+1}(b)=\left\{
\begin{array}{ll}
0 &\mbox{if }\quad w\neq {\mbox {${\bf 1}$}},\\
id_b \diamond id _{{\mbox {${\bf 1}$}}} &\mbox{otherwise.}
\end{array} \right.$$
The corresponding diagram is presented in figure \[fig.conj2\].
(4,3) (0,0)
The conjecture about the finite order of the circulator implies that given a presentation, the exponents of the generators in the relations matter only mod $Lp$, in particular it is enough to look at presentations $P$ with only positive exponents. In this last case, the morphism $[P,\underline{b}]$ has a particularly simple form - it is a product of cyclic permutations: $$[P,\underline{b}]=(1,2,3,\dots ,e_1)\diamond (1,2,3,\dots ,e_2)
\diamond \dots \diamond (1,2,3,\dots ,e_m),$$ where $e_i$ is the total exponent of the relation $R_i$. This fact doesn’t seem to help though with evaluating the invariant, at least until some more information is available for the symmetry group representations $\rho [b^k]$.
In conclusion we point out that clearly the main question is how the invariant behaves under the move (vii) in section 1, i.e. under simple homotopy equivalence. The numerical examples haven’t exhibited any nontrivial behavior, but using the formalism in this paper to write down the change in the value of the invariant under (vii), brings to expressions which look quite complicated and we don’t know how to analyze. We suspect that if the behavior of the computed invariant under this move is actually trivial, this is a consequence of some specific properties of the Gelfand - Kazhdan categories being used ( like the ones listed in conjecture \[conj1\]), rather then a consequence of the general setup of the theory.
[99]{}
J.Andrews, M.Curtis, [*Free groups and handlebodies*]{}, Proc.AMS [**16**]{} (1965),192-195.
I.Bobtcheva, [*Numerical generation of semisimple tortile categories coming from quantum groups*]{}, PhD thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA, August 1996.
P.Freyd and D.Yetter, [*Brided compact closed categories with applications in law-dimensional topology*]{}, Adv. in Math. [**77**]{} (1989),156-182.
S.Gelfand and D.Kazhdan, [*Examples of tensor categories*]{}, Invent.Math. [**109**]{} (1992),595-617.
C.Hog-Angeloni, W.Metzler and A.Sieradski, [*Two-dimensional homotopy and combinatorial group theory*]{}, Cambridge University Press, London Mathematical Society Lecture Notes Series, vol.[**197**]{}, 1993.
G.Kelly and M.Laplaza, [*Coherence for compact closed categories*]{}, J.Pure and Applied Math. [**19**]{} (1980),193-213.
S.MacLane, [*Natural associativities and commutativities*]{}, Rice Univ. Studies [**49**]{} (1963),28-46.
F.Quinn, [*Lectures on Axiomatic Topological Quantum Field Theory*]{}, LAS/Park City Mathematical Series, vol.[**1**]{}, 1995.
I.Bobtcheva and F.Quinn, [*Numerical presentations of tortile categories*]{}, internal report.
F.Quinn, [*Handlebodies and 2-complexes*]{}, Lecture Notes in Mathematics [**1167**]{}, Geometry and Topology, Proceedings, University of Maryland, 1983-1984, ,245-259.
Mei Chee Shum, [*Tortile tensor categories*]{}, Journal of Pure and Applied Algebra [**93**]{} Berlin Heidelberg New York (1994), 57-110.
D.Yetter, [*State-sum invariants of 3-manifolds associated to artinian semisimple tortile categories*]{}, Topology and its Applications [**58**]{} (1994),47-80.
[^1]: Partially supported by grant from the US National Science Fondation.
[^2]: In part of the literature [@CM:book] this is actually called a 3-deformation, since it can be achieved through expansions and collapses of disks of dimension at most three.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Improvements of in-orbit calibration of GSO scintillators in the Hard X-ray Detector on board Suzaku are reported. To resolve an apparent change of the energy scale of GSO which appeared across the launch for unknown reasons, consistent and thorough re-analyses of both pre-launch and in-orbit data have been performed. With laboratory experiments using spare hardware, the pulse height offset, corresponding to zero energy input, was found to change by $\sim 0.5$% of the full analog voltage scale, depending on the power supply. Furthermore, by carefully calculating all the light outputs of secondaries from activation lines used in the in-orbit gain determination, their energy deposits in GSO were found to be effectively lower, by several percent, than their nominal energies. Taking both these effects into account, the in-orbit data agrees with the on-ground measurements within $\sim $5%, without employing the artificial correction introduced in the previous work (Kokubun et al. 2007). With this knowledge, we updated the data processing, the response, and the auxiliary files of GSO, and reproduced the HXD-PIN and HXD-GSO spectra of the Crab Nebula over 12-300 keV by a broken powerlaw with a break energy of $\sim 110$ keV.'
author:
- |
Shin’ya <span style="font-variant:small-caps;">Yamada</span>, Kazuo <span style="font-variant:small-caps;">Makishima</span>, Kazuhiro <span style="font-variant:small-caps;">Nakazawa</span>, Motohide <span style="font-variant:small-caps;">Kokubun</span>,\
Madoka <span style="font-variant:small-caps;">Kawaharada</span>, Takao <span style="font-variant:small-caps;">Kitaguchi</span>, Shin <span style="font-variant:small-caps;">Watanabe</span>,\
Hiromitsu <span style="font-variant:small-caps;">Takahashi</span>, Hirofumi <span style="font-variant:small-caps;">Noda</span>, Hiroyuki <span style="font-variant:small-caps;">Nishioka</span>, Kazuyoshi <span style="font-variant:small-caps;">Hiragi</span>,\
Katsuhiro <span style="font-variant:small-caps;">Hayashi</span>, Kenta <span style="font-variant:small-caps;">Nakajima</span>, Makoto <span style="font-variant:small-caps;">Tashiro</span>, Makoto <span style="font-variant:small-caps;">Sasano</span>,\
Sho <span style="font-variant:small-caps;">Nishino</span>, Shunsuke <span style="font-variant:small-caps;">Torii</span>, Soki <span style="font-variant:small-caps;">Sakurai</span>, Tadayuki <span style="font-variant:small-caps;">Takahashi</span>,\
Tsunefumi <span style="font-variant:small-caps;">Mizuno</span>, Teruaki <span style="font-variant:small-caps;">Enoto</span>, Takayuki <span style="font-variant:small-caps;">Yuasa</span>, Takaaki <span style="font-variant:small-caps;">Tanaka</span>,\
Tomomi <span style="font-variant:small-caps;">Kouzu</span>, Toshio <span style="font-variant:small-caps;">Nakano</span>, Yasushi <span style="font-variant:small-caps;">Fukazawa</span>, Yukikatsu <span style="font-variant:small-caps;">Terada</span>,\
Yasunobu <span style="font-variant:small-caps;">Uchiyama</span>, Wataru <span style="font-variant:small-caps;">Iwakiri</span>, and the HXD team
title: |
Improvements in calibration of GSO scintillators\
in the *Suzaku* Hard X-ray Detector
---
Introduction {#section:1}
============
The fifth Japanese X-ray satellite Suzaku ([@Mitsuda2007]), carries the X-ray Imaging Spectrometer (XIS; [@Koyama2007]) located at the foci of the X-ray Telescope (XRT [@Serlemitsos2006]) and a non-imaging hard X-ray instrument, the Hard X-ray Detector (HXD). The HXD covers a hard X-ray energy range of 10–600 keV, utilizing Si-PIN diodes (hereafter HXD-PIN), and gadolinium silicate scintillators (Gd$_2$SiO$_5$:Ce, hereafter HXD-GSO) which are placed behind HXD-PIN. The detailed design of the HXD is summarized in Takahashi et al. (2007) (hereafter Paper I), followed by a report on its in-orbit performance ([@Kokubun2007]; hereafter Paper II) and that of the timing property [@Terada2008].
Prior to the launch on 2005 July 10, the performance of the HXD was tested and calibrated at every stage of its integration, and also after the instrument was mounted on the spacecraft. In addition to these pre-launch experiments, several Monte-Carlo simulations were performed, to better understand the HXD performance (including its background) expected in near-Earth radiation environments ([@Terada2005]; Paper I). We have also calibrated the HXD performance extensively after the launch (Paper II), and compared all propertied that can be obtained in-orbit with the results of the ground measurements. As a result, we have successfully verified that the HXD performs in-orbit generally as expected (Paper II), and reconfirmed most of the results of the pre-launch calibration.
In spite of these successful in-orbit calibration efforts, we are still left with a few unsolved issues of the HXD performance. One of them is an apparent change in the energy scale (relation between the incident X-ray energy and detector pulse height) of HXD-GSO. We found this issue when we observed the X-ray pulsar A0535+26, which exhibits an absorption line at $\sim 50$ keV [@Terada2006]; this spectral feature appeared at discrepant energies in the HXD-PIN and HXD-GSO spectra that were produced based on the nominal pre-launch calibration (see subsection 2.3). Since the energy scale of HXD-PIN was kept unchanged across the launch (figure 5 of Paper II), we tentatively concluded that the energy scale of HXD-GSO changed for some unspecified reasons, and hence added an artificial non-linearity to the GSO energy scale (figure 12 of Paper II) as a temporary solution.
The aim of the present paper is to more fundamentally solve this issue, by revisiting the pre-launch and in-orbit calibrations, and conduct some laboratory experiments using spare hardware. Section 2 is used to review the way of conversion from energy to pulse height, and to present the investigation strategy. In section 3, we study the spare hardware performance, followed in section 4 by a detailed comparison between the on-ground and in-orbit data, and construction of a revised energy scale. We present in section 5 how well our new solution worked on actual observation data, and summarized these works in section 6.
The HXD Experiment
==================
We start with a brief summary of the HXD in section 2.1, and then the procedure of the energy scale in section 2.2. In section 2.3, the in-orbit calibration that had been performed until Paper II are summarized. Other unsolved issues are introduced in section 2.4. We present our methods to tackle the problem in section 2.5.
\[section:2\]
(140mm,140mm)[figure1.eps]{}
The hardware configuration
--------------------------
As detailed in Paper I, the HXD consists of three major components: a sensor part (HXD-S), an analog electronics part (HXD-AE), and a digital electronics system (HXD-DE). HXD-S has a compound-eye configuration, made of 4$\times$4 well-type phoswich detectors (“Well-counter units”), surrounded by 20 thick active shields (“Anti-counter units”). A Well-counter unit consists of 4 Si-PIN diodes and 4 GSO scintillators, all placed inside a deep well of active shields formed by BGO scintillators. Each unit outputs 4 Si-PIN signals and one GSO signal, after scintillation light from the 4 GSOs is detected by a single photomultiplier tube (PMT) which also detects light from the BGO shield. All signals of 36 PMTs (16 Well-counter and 20 Anti-counter units) and 64 PIN diodes are fed in parallel into HXD-AE, where they are amplified, pulse-shape discriminated, and analog-to-digital converted. The digitized signals are further processed by HXD-DE using an on board CPU.
The present paper pertains to the GSO scintillators (HXD-GSO) of the Well-counter units, as well as to HXD-AE, but not to HXD-DE. It does not deal with the Anti-counter units in HXD-S, either.
Energy to pulse-height conversion
---------------------------------
Let us briefly review the flow of converting the pulse-height into energy as described in Paper II. In any radiation detector that detects individual particles (including photons) as electronic pulses, the energy scale is critically important to its performance. Its determination is significantly more difficult in hard X-rays, than in energies below $\sim 25$ keV where we can utilize various celestial/instrumental atomic lines and built-in calibration radio-isotopes. In the particular case of HXD-GSO, the energy $E$ of a detected photon is converted into pulse height, or more properly, “pulse invariant”, $P$, through a process shown in figure 1. A photon with an energy of $E$, totally absorbed in a GSO scintillator, creates a photoelectron which immediately produces many secondary electrons. These electrons deposit their energies in the scintillator, creating optical photons of a total number $L$ (which is called light yeild). The photons are (partially) detected by a PMT (Hamamatsu R6231), to get multiplied into an electronic pulse with charge $Q$. This charge is converted into a voltage pulse $V$ in a charge sensitive amplifier (CSA). The process up to this stage takes place in HXD-S. The pulse sent to HXD-AE is further amplified and shaped therein by a shaper, then peak-held in an application-specific integration circuit, and finally digitalized by a 12-bit analog-to-digital converter. Thus, a photon energy up to $E \sim 600$ keV is converted into a 12-bit pulse height, $H$.
Although the pulse height $H$ is approximately proportional to $E$ with a constant coefficient of proportionality, there are at least following effects that can cause deviations and/or time variations in the $E$ vs. $H$ proportionality.
1. Non-linearity in the GSO light yeild [@Uchiyama2001].
2. The overall coefficient of proportionality, or “gain”, donated $G$.
3. Differential and integrated non-linearity in the amplification chain in HXD-AE [@Kitaguchi2006].
4. The offset signal in analog electronics, or the “pedestal” defined in Paper II, denoted $H_{\rm{o}}$ in this peper. This means a value of $H$ when no signal is present in GSO ($E=L=Q=0$).
The effect (2) is further decomposed into the GSO light yield, the PMT gain, the CSA capacitance, and electronics settings. As detailed in Paper II, the GSO light yield depends on the temperature, while the PMT gain varies by $\sim 10$% in a complex way, depending on the temperature, time since the launch, and time after passing through South Atlantic Anomaly, wherein the PMT high voltage is turned off for $\sim 20$ minutes or less. Therefore, this coefficient, $G$, must be determined from moment to moment. In contrast, the effects (1), (3), and (4) are considered approximately constant with time, and were determined in the ground measurements; therefore, we can apply these results to the in-orbit data unless something in hardware did change.
Given the knowledge on the effects (1)-(4), we can convert the pulse height $H$ of each event into a quantity called pulse invariant ($P$) through the reverse procedure of figure 1. If (1)–(4) are correct, $P$ should coincide with $E$ within the detection energy resolution (statistical effect) and calibration uncertainties in the energy scale (systematic effects).
In-orbit calibration until Paper II
-----------------------------------
During the initial in-orbit calibration phase of Suzaku, we applied the ground-determined (1), (3), and (4) to the actual in-orbit HXD-GSO data. As explained in section 2.2, $G$ must be determined every $\sim 50$ minutes, so we used for that purpose GSO background lines at 511 keV, $\sim$350 keV, and 153 keV (table. 7 of Paper II for details). At that time, the energies of these lines were considered to be well defined, because they are based on nuclear physics, and the effect of non-linearity of (1) and (3) are negligible at least in $> 100 $ keV. Thus, by fitting these lines on the $E$ vs. $H$ plane with a linear function \[equation (1) in Paper II\], we determined the slope $G$ and the intercept of the line with the $H$ axis, hereafter $H_{\rm{i}}$, allowing both to vary. As a typical case, the employed data points (on the $E$ vs. $H$ plane) and the linear function (green) for one Well Unit (W31) are shown in figure \[ehplot\], which is almost the same as figure 12 in Paper II. The data points include background gamma-ray lines at 70 keV, 150 keV, 193 keV, $\sim$ 350 keV, and 511 keV. Here $H$ is expressed in units of ADC channels, with its maximum, 4096, corresponding to $\sim$ 700 keV. Then, $H_{\rm{i}}$ has appeared at $\sim -50$ channels (Paper II), in contrast to $H_{\rm{i}} \sim -10$ channels measured in the ground calibration (Paper I). This means apparent changes by $\Delta H_{\rm{i}} = -40$ between the post- and pre-launch calibrations.
(100mm,100mm)[figure2.eps]{}
Although the non-zero value of $\Delta H_{i}$ suggests some changes in the GSO energy scale in energies of $\lesssim 70 $ keV, there are no appropriate spectral features in the $\sim$ 10 to $\sim$ 70 keV range to be used for calibration. Accordingly, a celestial object named A0535+26 was utilized to calibrate the energy scale down to $\sim$ 50 keV [@Terada2006], because it is known to exhibit a cyclotron resonance absorption line as shown in black in figure \[a0535\]. The feature was clearly detected by HXD-PIN at an energy of $\sim 45$ keV ([@Terada2008]). However, as shown in green in figure \[a0535\] and plotted in cyan in figure \[ehplot\], the same feature was seen at about 60 keV ($H \sim 205$) in the GSO spectrum, when the detected $H$ was converted back to $P$ by applying (1)-(4) and extrapolating the energy scale determined over 70–511 keV (doted green line in figure 2) to energies below 70 keV. More specifically, this procedure appears to assign too high a value of $E$ for a given $H$ of HXD-GSO, or conversely, to underestimate $H$ for GSO events with $E \sim 45$ keV. This could be regarded as being caused by the change of $\Delta H_{\rm{i}} \simeq -40$ mentioned above.
The observation of A0525+26 thus revealed a $\sim$ 20% discrepancy between the energy scales of HXD-PIN and HXD-GSO at lower energies. However, the energy calibration of HXD-PIN is more reliable than that of GSO, because its energy scale is mostly determined by the preamplifier capacitance. Furthermore, an instrumental peak due to escaping Gd-K line photons, expected at 43 keV, correctly appeared at $43\pm1$ keV in the in-orbit background spectra of HXD-PIN. Therefore, we had to conclude that the problem is in HXD-GSO, rather than in HXD-PIN.
In order to reconfirm the inference derived from the A0535+26 data, Paper II utilized pedestal information $H_{\rm{o}}$ on the PMT signal, i.e., the value of $H$ when the energy deposit $E$ approaches 0. This $H_{\rm{o}}$ can be obtained not only during the pre-launch tests, but also from the in-orbit data, by referring to values of $H$ of those events which were triggered by the PIN diodes in the same Well-counter unit: even for such events, the PMT pulse heights are latched, and sent out to telemetry, for later use in the background rejection. In fact, this gives $H_{\rm{o}} + \delta $, where $\delta$ is contribution from noise in the PMT signal and can be estimated to be less than $\sim $1 keV ( $\sim$ 4 ch). The value of $H_{\rm{o}}$ derived in this way is given in figure 2b as a data point at $E \sim 0 $ and $H_{\rm{o}} \sim 80$. Thus, the problem uncovered by the A0535+26 data was reconfirmed. This effect exceeds what can be explained by (1) and (3).
To temporarily solve this apparent discrepancy in the energy scale of GSO below $\sim 70$ keV, an additional non-linearity (the third term of equation 2 in Paper II) has been artificially introduced. As a result, the GSO spectrum of A0535+26 moved to lower energies (blue spectrum in figure \[a0535\]), and the cyclotron absorption-line energy measured with HXD-GSO became consistent with that of HXD-PIN. Although the blue spectrum in figure \[a0535\] still shows a somewhat higher absorption-line energy than the black one, the difference can be explained away by a poorer energy resolution of HXD-GSO than HXD-PIN, coupled with a steeply declining continuum. In figure \[a0535\], the entire spectra are divided by the Crab Nebula spectrum which behaves as $\propto E^{-2.1}$.
(100mm,100mm)[figure3.eps]{}
Other differences between on ground and in-orbit data
-----------------------------------------------------
In addition to the above issue of the GSO energy scale, the in-orbit data exhibited some other anomalies which cannot be fully explained by the pre-launch calibration. When Paper II was complied, it was not clear whether either or both of these two unexpected effects comes from the same origin as the GSO energy scale problem. These issues are discussed in section 5.
One is the performance of pulse-shape discrimination, which is applied to the PMT signal in HXD-AE to distinguish GSO and BGO events (see figure 15 in Paper II). As described later in subsection 5.1, width of the GSO branch on so-called slow-fast diagram, appears to have charged across the launch. The other is that the current response matrix for GSO is overpredicting the spectrum by 10% in energies below 100 keV, as evidenced by a Crab spectrum of GSO (figure 18 in Paper II). For a temporary settlement, an artificial ancillary response file has been provided.
Working hypotheses and verification methods
-------------------------------------------
Although the unexpected change in the GSO energy scale below $\sim 70$ keV is at present absorbed by the artificial correction factor, this is a temporary and empirical solution. More fundamentally, we need to understand the cause of the change, and if necessary, update the GSO energy scale.
In the following, we conduct two approaches towards the above objectives. One of them is to examine the HXD hardware performance for any launch-related changes. This is carried out in section 3, partially incorporating laboratory experiments using scintillators, phototubes, preamplifiers, and analog electronics that are all (nearly) equivalent to the actual flight-model HXD.
The other approach to our goal is to examine the process of energy calibration, searching for any inaccuracy or inconsistency between the pre- and post-launch calibrations. In section 4, we therefore re-analyze the in-orbit GSO data, as well as the pre-launch calibration data which were taken by irradiating various calibration isotopes to the HXD. Through a consistent data analysis, we try to understand more accurately how the energy scale had changed across the launch.
Examination of the hardware performance {#section:3}
=======================================
Possible changes in the HXD hardware
------------------------------------
The first of the two approaches mentioned in section 2.5 is based on a concern that the present issue can be a result of some changes in the HXD hardware performance, including its scintillators, PMTs, CSAs, bleeders, and electronics; in other words, anywhere in the diagram of figure 1. This is further subdivided into the following three general possibilities. One is that the launch stress caused some irreversible changes (including damages) to the HXD hardware. Another is that frequent large-amplitude scintillation pulses, produced by charged particles in the space environment, brought about some differences in the low-energy response. The other is that the spacecraft power supply, which is known to be less stable than those used in laboratory, may have caused some changes in the electronics response to low-pulse-height signals.
We found that the former two possibilities do not affect the energy scale, so that we summarized them in Appendix 1 and 2, respectively. In the next subsection, we examine the third possibilities that actually affected the energy scale of the GSO.
Effects of the power supply {#subsection:3-}
---------------------------
A clear difference in the hardware environment between the ground and in-orbit measurements is the power supply. Our ground calibration was performed using general-purpose power supplies (+12 V, -12 V, +5 V analog, and +5 digital), which are more stable than the flight one. This could affect the HXD-AE performance, particularly at low pulse heights. Although HXD-AE does not have its own calibrators, the raw pulse height accumulated during no signal inputs in GSO, i.e., pedestal ($H_{\rm{o}}$ in figure \[ehplot\]), can be used for this purpose. The values of the pedestal have been compared among three data sets: (1) the on-ground calibration data acquired in June 2004, using a laboratory power supply, where the HXD was not yet mounted on the spacecraft; (2) the on-ground data taken in 2004 August with the HXD mounted on the spacecraft, and driven by the flight power supply; and (3) in-orbit data acquired from a black sky in June 2007 during a period when the PMT high voltage was temporary turned off to avoid over/under shoot caused by cosmic ray signals. As explained in section 2.3, the pedestal signal, $H_{\rm{o}}$, was obtained by selecting cosmic rays or background events triggered uniquely by HXD-PIN.
Figure \[ped\] shows histograms of $H_{\rm{o}}$ for the 16 Well-counter units. Thus, they show $H_{\rm{o}} \sim 81$ channels when HXD-AE was driven by laboratory power supplies. In contrast, they exhibited $H_{\rm{o}} \sim 73$ channels when the flight power supply was used, both before or after the launch. The change by $\Delta H_{\rm{o}} \sim -8$ can be attributed to the difference in the power supplies, because everything else was the same. However, as illustrated in figure \[ehplot\], the effect, at most $\sim$4% for the 50 keV signals (equivalent to $H \sim 200$), is not large enough to account for the $\sim 20$% change of energy scale at $\sim$ 50 keV (section 2.3); some other effects are needed.
(70mm,70mm)[figure4.eps]{}
A Unified Analysis of Ground and In-Orbit data sets {#section:3}
===================================================
We start with examination of ground calibration data in section 4.1, and then, employing newly introduced “$Q$ vs. $H$" plot in section 4.2, compare the ground data to in-orbit data in section 4.3. The examination on light yield and reanalysis of in-orbit data are summarized in section 4.4 and 4.5.
Summary of ground calibrations
------------------------------
Since we have found that the change in hardware cannot explain the change of the $E$ vs. $H$ relation, we then proceed to the second approach introduced in section 2.5, and try to review our calibration processes, starting with that on ground data.
Although an inorganic scintillator has a roughly constant $L$ (subsection 2.2) of its own, the $E$ vs. $L$ proportionality is in many cases only approximate, and needs detailed calibration. such measurements of GSO used in the HXD were performed by Kitaguchi et al. (2005), irradiating gamma-rays from various radio-isotopes.
$Q$ vs. $H$ calibration on ground in June 2004 were performed by Kawaharada et al. (2004), by inputting electronic test pulses of various pulse heights simultaneously to HXD-AE and a laboratory MCA (the same as used in section 3.2). When the latter was used as an ideal reference, HXD-AE was found to have a good $Q$ vs. $H$ proportionality at $H \gtrsim 200$ ($\gtrsim$50 keV), with slight non-linearity below $H \sim 200$.
In addition to these basic measurements, end-to-end pre-launch calibrations were performed by irradiating gamma-ray isotopes to GSO from outside and processing the signals with HXD-AE. The acquired data are of high importance, because they can be compared directly with the in-orbit data.
Here, we have analyzed such pre-launch GSO calibration data, taken by Kawaharada et al. (2004) in June 2004 under the same condition as (1) in section 3.2 (i.e., using a laboratory power supply). Of the 16 Well-counter units, we use the W31 unit for simplicity, because the HXD-AE performance is almost the same among the 16 channels (Paper I). Thus, we can use the results to compare them with the in-orbit data.
Introduction of “Q vs H” plots
------------------------------
When comparing different sets of calibration data, we usually create from each data set an “$E$ vs. $H$ plot", wherein the measured pulse heights $H$ are plotted against the known photon energies $E$. A comparison between a pair of $E$ vs. $H$ plots requires their appropriate corrections for the difference in the overall gain $G$, which varies depending mostly on the PMT gain. This is usually done by renormalizing the $H$ values from one of the plots to a certain constant factor, so that these plots become close to each other. This renormalization procedure is correct if, for example, the $E$ vs. $H$ relation deviates from an exact proportionality mainly due to the GSO non-linearity between $E$ and $L$; however, it is not if the deviation is dominated by electronics effects and hence is dependent on $H$, not $E$ or $L$. More generally, an $E$ vs. $H$ relation may be approximated as $H= GE+H_{\rm{o}}$ at small values of $E$, and it is not trivial how to deal with $H_{\rm{o}}$ when re-normalizing; should we calculate as $H \rightarrow \beta(GE+H_{\rm{o}})$, or $\beta GE+H_{\rm{o}}$, or in other ways (where $\beta$ is the renormalization factor)?
To avoid the above problem, we have introduced a new concept of “$Q$ vs. $H$" plot (Yamada 2008), where $Q$ means the charge (figure 1) coming out of the PMT. The new plot can be obtained from an ordinary $E$ vs. $H$ plot in the following manner. First, the energy $E$ (assumed to be known) of each data point is converted into the light yield $L$ through the knowledge on the GSO light yeild (Kitaguchi 2006), or $E$ vs. $L$ relation, which is implicitly assumed to be rather stable across the launch. Then, the charge $Q$ is defined as a quantity proportional to $L$, without any zero-point offsets; this is justified by the general property of PMTs. The constant of proportionality between $Q$ and $L$ is determined so that the value of $Q$ becomes identical to that of $E$ (in units of keV) at the most reliable high-energy data point. Thus, $Q$ is defined to take roughly the same values as $E$, and is considered to be free from the GSO non-linearity and the PMT gain changes. The most time-variable factor, namely $G$, is re-normalized at this stage. In other words, we compare different calibration plots by adjusting their energy axes, rather than by stretching/compressing their pulse-height axes. This procedure implicitly assume little changes in the HXD-AE performance.
Comparison between the ground and in-orbit calibration results
--------------------------------------------------------------
Figure \[newplot\] is the “$Q$ vs. $H$” plot constructed via the above procedure from the ground calibration data (green) and in-orbit data (red). The gain of the ground data were re-adjusted (along the horizontal axis) so that they match the in-orbit ones. To guide the eyes, a proportionality line pointing from the origin to the data corresponding to 511 keV is superposed, and ratios of the data to this line are shown in figure \[newplot\]b. In the lower-ADC channel ($\lesssim$ 200 ch) region, the ground calibration data exhibit $\sim10$% deviation above the linear function. As described in subsection 3.3, this offset is attributable to the non-ideal $Q$ vs. $H$ properties of HXD-AE (Kawaharada et al. 2004). Actually, when the on-ground data points are corrected for this effect, the deviation from the proportionality line diminishes to less than $4$%; this means that the energy scale of GSO for the on-ground data consists of all known factors.
(70mm,70mm)[figure5.eps]{}
Figure \[newplot\]b reveals two systematic differences between the pre-launch and in-orbit calibration data sets. One is that the in-orbit data points fall systematically below the line of proportionality, in energies from 60 to 400 keV: this effect is not observed in the ground calibration data. The other is that the in-orbit data points in the pulse-height range below $H \sim 250$ ch exhibit much milder excess above the proportionality line, than the ground data points. Thus, the energy scale appears to have really changed across the launch. However, it is apparently inconsistent with the result that in-orbit data exhibit [*larger*]{} excess toward lower pulse heights than the pre-launch data (cf. figure 2; figure 12 in Paper II).
As explained in section 2.2 and figure \[ehplot\], the first approximation of energy scale was derived by connecting the 150 keV, 350 keV, and 511 keV data points with a linear (not proportional) function, in which the slope and offset are both allowed to change. This linear function, with an intercept of $H_{\rm{i}} = -50$ (figure \[ehplot\]), has a steeper slope than the line of proportionality (with $H_{\rm{i}} = 0$) employed in figure \[ehplot\]. If this steeper line is instead used in figure \[newplot\], the in-orbit data in panel (b) would exhibit much larger positive residuals in $H \lesssim 200$, while the intermediate-energy data points would fall right on the line, thus reproducing the behavior seen in figure \[ehplot\]. Considering the result in section 3 that the hardware performance has not changed as much, it is therefore more reasonable to consider that the significant low-enegy excess observed in figure \[ehplot\] and Paper II is caused artificially because we started from an inappropriate linear function with a steeper slope (and hence with a large negative $H_{\rm{i}}$). If this is true, the next question is what has caused the line, connecting 150, 350, and 511 keV points, to have a wrong slope. Let us consider how the energies of these activation/annihilation lines seen in the in-orbit background spectra are measured with HXD-GSO.
Re-examination of Light yield from activation lines
---------------------------------------------------
Among the instrumental background lines used in the in-orbit energy calibration of HXD-GSO, the 511 keV electron-positron annihilation line is based on a simple elementary process, with its photo-peak created by absorption of a single mono-energetic gamma-ray line. In contrast, as described in Table 7 in Paper II, the activation lines at 70 keV, 150 keV, and 196 keV are statistical mixtures of various radiation components, including nuclear transition gamma-rays, fluorescent X-rays, Auger electrons, and internal conversion electrons. This raise a suspect that the values of $L$ for the 150 keV and 350 keV activation lines would differ from their nominal values.
(70mm,70mm)[figure6.eps]{}
The values of light yield, $L$, from activation lines have been so far calculated by assuming that their energy is deposited in GSO by a single gamma-ray photon. However, this approximation might be over-simplified, because electrons and photons give different $L$ in GSO and their energy deposited in GSO are not perfectly proportional to $L$. Figure \[relaL\] shows the light yield against energy for electrons (Uchiyama et al. 2001) and photons (Kitaguchi et al. 2006), which are both experimentally measured and normalized at 344 keV. Thus, the two curves are both considerably non-linear, and are different from each other. Therefore, even if the overall energy released by a nuclear decay is the same, the light output depends on how the total energy is split into individual components, and whether they are photons or electrons. Taking into account figure \[relaL\], we evaluate the light output for the three background lines (70 keV, 148 keV, and 196 keV).
(70mm,70mm)[figure7.eps]{}
For each activation line, we analyzed the decay processes involved therein, and calculated their energy releases and branching ratios. This was done using such information as nuclear activation energies, the energies of fluorescent X-rays, the fluorescent yields, and the internal conversion coefficients. Then, we calculated light outputs for all the electrons and photons using the corresponding relative light yields in figure \[relaL\], and summed up the derived light outputs weighted by their branching ratios.
Figure \[decay\] shows an example of the decay scheme of the 70 keV line. It begins with electron capture of $^{151}$Gd into $^{151}$Eu. This releases an energy of $\sim$ 50 keV, via emission of a fluorescent X-ray photon or an Auger electron, with a branching ratio of $\sim$ 90% and $\sim$ 10%, respectively. This leaves $^{151}$Eu in an excited state with a remaining energy of 20 keV, which is released by either a photon or an electron through internal conversion. A simplified example of evaluation is as follows. When a single photon emission with an energy of 70 keV is considered, the light yield is estimated as $70 \times 0.92 = 64.4$, where 0.92 can be read from figure \[relaL\]; in contrast, where a photon of 20 keV, plus a 9:1 contribution from a photon of 50 keV and an electron of 50 keV are considered, the light yield can be calculated as $20 \times 0.78 + 0.9 \times 50 \times 0.9 + 0.1 \times 50 \times 0.96 = 60.9$, which is 5.4 % lower than the single-photon value of 64.4 keV. We have calculated more precisely (Yamada 2008), and described them more in detail in Appendix 3. As a result, we have found that the light outputs of 70 keV, 148 keV, and 196 keV lines decrease by 5.7%, 5.2% and 3.0%, respectively. In other words, for example, the activation line which totally deposits 70 keV in GSO should have been regarded as a light output from a single photon with an effective energy of 65.8 keV.
Re-analysis of the In-orbit Data
--------------------------------
When the revised output (subsection 4.4) is used, the in-orbit data points for the 70, 150, 196 keV lines move to the left in figure \[newplot\]a. Then, their ratio to the proportional line have become as shown in figure \[newplot\]c. Thus, the in-orbit data agrees better with the ground data. However, the ratio of the in-orbit data still shows slightly lower values over 50–400 keV than the on-ground ones.
Finally, the change of the pedestal, $\Delta H_{\rm{o}} \sim -8$, as found in section 3.2, was considered. As a result, the low-enegy in-orbit data points moved in figure \[newplot\]a slightly upward along the $H - \Delta H$ axis, and their ratios, figure \[newplot\]d, became in full agreement with those of the on-ground data. We hence conclude that the in-orbit energy scale of GSO is consistent with the on-ground one within an accuracy of $\sim$ 5%, and the apparent change of the in-orbit data is caused partially by the differences of HXD-AE power supplies, and more fundamentally, by the overestimation of light outputs from the activation lines employed in the in-orbit calibration. The artificial nonlinearity introduced in Paper II and mentioned in section 1 is no longer needed.
As a further confirmation, we overlaid the GSO spectra of A0535+26 in red in figure 3. The red spectrum agrees with the blue one at $P \sim 50$, which means that the energy scale of GSO agrees with that of PIN at $\sim 50$ keV; while the red one around $P \sim$ 70–200 is slightly shifted to the left side by $\sim P$ channels $\sim$ keV, which is caused by the recalculation of the light yield preformed in section 4.4. Thus, we have succeeded in obtaining more correct energy scale of the GSO than the previous one, without relying on the artificial correction introduced in Paper II.
Application of the revised energy scale of HXD-GSO
==================================================
Since we have confirmed that the hardware did not change except for the slight pedestal shift, we have created a new energy scale of GSO and implemented it in a ftool [hxdpi]{} in HEADAS 6.9 or later. In this section, we described some after effects of this new energy scale, and apply it to the spectra of the Crab Nebula.
GSO branch width
----------------
Let us examine the issue of GSO branch width (for definition see figure 15 in Paper II). Figure \[psd\] shows four GSO branch widths from the unit W00 (see caption for details). As shown in Paper II and mentioned in section 2.4, the in-orbit GSO branch (yellow) became wider below $100$ keV than those of the ground calibration (green). This change can be explained as an effect of the artificial nonlinearity introduced previously, because the effect of nonlinearity works differently for fast and slow shaped pulse heights (unless they happen to have the same value). This causes a wider data distribution in a fast-slow diagram (cf. figure 2 in Paper II).
As shown in red in figure 8, this problem has nearly disappeared when we re-analysed the same in-orbit data using the new GSO energy scale established in section 4.5. Thus, we have confirmed that the in-orbit GSO branch width is consistent with the on-ground one and that the apparent change as explained in section 2.4 was caused by the incorrect energy scale of GSO. The revised GSO branch width is released as [ae\_hxd\_gsopsd\_20090812.fits]{}, which must be used in [hxdgrade]{}.
(70mm,70mm)[figure8.eps]{}
Update of GSO response
======================
The final task of our re-calibtation is to update the response matrix of GSO. The calculation of the HXD energy response employes a Monte-Carlo simulator, called “simHXD" (Ozaki et al. 2006; Terada et al. 2005), based on a Geant4 framework ([@Sulkimo2003] and [@Allison2006]). which has been constructed using the measured detector performance and the actual detector geometry. This simHXD assumes a bunch of input photons obeying a flat spectrum, and produces the corresponding fake HXD event data, by simulating how each input photon interacts with the HXD-PIN and/or HXD-GSO and what kind of signals it produces. Then, the simulated data are processed in the same way, using the same event selection criteria, as the actual data are analyzed. This gives an acceptance probability (and hence an effective area) for an input photon to be judged as a valid HXD-PIN or HXD-GSO event.
In the present paper, we have improved simHXD in the following three points. (1) The HXD-GSO energy scale was revised (subsection 5.1), while calculation on energy deposits were not changed. (2) The GSO branch width was updated referring to the red data in figure 8. (2) The simulator was further optimized, with respect to its usage of the GSO brach width information, and its emulation of the event selections in HXD-DE (Paper I).
In figure \[psd\], we overlaid a result of a improved simHXD (Terada et al. 2005; Ozaki et al. 2006). The GSO branch width calculated with simHXD should be consistent with the others. In fact, they became consistent as shown in figure \[psd\]. Thus, the on-ground, in-orbit and simulated data became all consistent. The new response matrices of HXD-GSO, [ae\_hxd\_gsoxi(hx)nom\_20100524.rsp]{}, has been released in May of 2010 [^1].
The HXD spectra of the Crab Nebula
----------------------------------
(140mm,140mm)[figure9.eps]{}
---------------------------- ------------------- ------------------- ---------------- ------------------ -----------------------
Target position Photon index1 Normalization Photon index2 $E_{\rm{break}}$ $\chi_{\nu}^2$(d.o.f)
\[1.05ex\] Target position Photon index1 Normalization Photon index2 $E_{\rm{break}}$ $\chi_{\nu}^2$(d.o.f)
XIS nominal 2.109 $\pm$ 0.003 11.65 $\pm$ 0.011 - - 1.23 (132)
HXD nominal 2.096 $\pm$ 0.003 11.14 $\pm$ 0.011 - - 1.52 (132)
\[1.3ex\] XIS nominal 2.104 $\pm$ 0.004 11.47 $\pm$ 0.12 2.21+0.06-0.04 105 $\pm$ 20 1.04 (130)
HXD nominal 2.089 $\pm$ 0.004 10.91 $\pm$ 0.12 2.28+0.09-0.10 117 $\pm$ 19 1.09 (130)
\[1mm\]
---------------------------- ------------------- ------------------- ---------------- ------------------ -----------------------
: Fitted parameters with a single and broken powerlaw for the PIN and GSO spectra of the Crab nebula.[]{data-label="tbl:pin_crabfit"}
The Crab Nebula has been used as a standard candle in the X-ray range, because it is bright and approximately stable. Thus, it is essential to examine whether the new GSO response, created above, can explain the Crab data, without invoking the artificial correction factor which had been introduced to supplement the old response.
We analyzed two data sets of the Crab, acquired consecutively on 2005 September 15 each for 6 hours: one was obtained at the HXD nominal position from 14:00, and the other at the XIS nominal position from 19:50. The GSO data were reprocessed by [hxdpi]{} and [hxdgrade]{} with the new GSO branch width, [ae\_hxd\_gsopsd\_20090812.fits]{} (section 5.1). Non X-ray backgrounds of GSO created from the same procedure, released as version 2.4, were used. We extracted HXD-PIN spectra from cleaned events, and subtracted non X-ray background (Fukazawa et al 2009) from it and used standard responses, [ae\_hxd\_pinxi(hx)nome1\_20080129.rsp]{}.
The discrepancy in normalization between PIN and GSO became $\sim 20$%, which is probably due to differences between actually observed data and simHXD on the anti-coincidence efficiency mainly triggered from Gd-K fluorescence emission lines or the depth of the depletion layers in PIN. The GSO responses around 50-60 keV is hard to calibrate due to its poorer energy resolution at these energy ranges and the photo-absorption at Gd-L shell that is not negligible at the lower energy ranges and makes the absorption process more complex than in higher energy ranges. Since these effects do not significantly affect the spectral shape over $\sim$ 70 keV in GSO, we created the GSO auxiliary files, [ae\_hxd\_gsoxi(hx)nom\_crab\_20100526.arf]{} [^2], which trims (by $\sim$ 20%) the PIN vs. GSO cross normalization, and adjust the HXD-GSO response around Gd-K edge energy (50–60 keV). The cross-normalization between HXD-PIN and HXD-GSO was fixed at 1:1, and the fitting range of HXD-PIN and HXD-GSO were 12-70 keV and 50-300 keV, respectively. The column density for photoelectric absorption was fixed at $3 \times 10^{21}$ cm$^2$. The error refers to 90% significance.
We started to fit the HXD-PIN and the HXD-GSO spectra with a powerlaw. The obtained parameters are summarized in table 1, while the spectra with the best-fitted model (blue) and the ratio to the model are shown in figure \[crab\]a and b, respectively. The XIS- and HXD-nominal data were both reproduced approximately by a powerlaw, with reduced $\chi^2$ (d.o.f) of 1.23 (132) and 1.52 (132), respectively. In both cases, the derived photon index of $\sim 2.1$ and the normalization of $\sim$11 photons s$^{-1}$ cm$^{-2}$ keV$^{-1}$ are close to those reported previously. However, the ratio to the model, as shown in the bottom panel in figure \[crab\]a and b, falls $\sim$ 10% below unity in energies above $\sim$100 keV. This suggests that there is a mild spectral cutoff at $\gtrsim 100$ keV in the Crab spectrum, as reported in some previous studies including in particular that with INTEGRAL [@crabintegral].
To express the high-energy turn over, we fitted the Crab spectra with a broken powerlaw, and obtained the results as summarized in table 1 (bottom two rows) and figure 9 (panels c and d). The fits became significantly improved to reduced $\chi^2$ (d.o.f) of 1.04 (130) and 1.09 (130), for the XIS and HXD nominal data, respectively, and the model vs. data discrepancy has been reduced to $<$10%. The results imply that the photon index steepens from $\sim$ 2.1 to 2.2–2.3 across $\sim$ 110 keV. These are consistent with those obtained by INTEGRAL [@crabintegral]. Thus, we confirmed that the new response of GSO can successfully reproduce the Crab spectrum in terms of a broken powerlaw, without using the artificial correction factor needed previously.
Conclusion
==========
We have performed re-calibration of the energy scale of the GSO scintillators which covers the harder energy ranges of the Suzaku HXD. Through laboratory experiments using flight-spare hardware, we found that the pedestal of HXD-AE decreased by 8 ch when the HXD is driven by the spacecraft power supply, than the case where a more stable laboratory power supply is used. Calculating all light outputs of secondaries from activation lines revealed that the energy of calibrators, used to determine the in-flight gain of GSO, was previously over-estimated by several percent. Taking both effects into account, the in-orbit data have been confirmed to be consistent with the pre-launch calibration.
When the GSO energy scale was thus revised, the in-orbit GSO branch width became also consistent with the ground one. Taking into account the energy scale thus updated, as well as the GSO branch width, the response matrixes have been created. Then, the Crab Nebula spectra of HXD-PIN and HXD-GSO, over 12–300 keV, has been expressed by a broken powerlaw with a break energy of $\sim 110$ keV and a photon index of $\sim$ 2.1 below this energy.
As a result of these works, major issues with the in-orbit calibration of HXD-GSO have been successfully solved.
The authors would like to express their thanks to all who have contributed to the design, development, operation, and the calibration of the HXD experiment, and all of the Suzaku supporters. The present work was supported by Grant-in-Aid for JSPS Fellows.
Irreversible Changes in the HXD Hardware
========================================
The heavy mechanical stress during the launch can produce cracks in the GSO crystals, or degrade their optical contact to the PMT entrance window. However, the energy resolution of GSO, measured after the launch using activation lines, agreed very well with that measured on ground (Paper II). Therefore, the response of the GSO crystals is unlikely to have changed significantly.
The PMTs are another critical item subject to the launch vibration. Since the launch, the PMTs have been working normally, with gain changes by $\sim$0.5% and $\sim$20% on short ($\sim$hours) and long ($\sim$month) time scales, respectively. Then, these gain changes may have slightly affected the PMT linearity, because the PMT bleeders use nonlinear electronic parts, such as clump diodes and Zener diodes (Paper I). However, the effects of these non-linear components would appear in the highest energy range under large signals caused by cosmic rays, than in the lower energy range as we observed. Therefore, the PMT performance is considered to have remained essentially unchanged.
Electronics parts in the CSA and HXD-AE could in principle be affected by the launch stress. However, it is highly unlikely that the 16 CSAs and 16 amplification chains, which individually process signals from the 16 Well-counter units, were damaged in the same manner. Considering these, we conclude that any conceivable launch-related irreversible hardware change cannot explain away the present issue of the GSO energy-scale change.
Effects of Large Signals in the Space Environment
=================================================
The HXD is exposed in-orbit to a high flux of cosmic rays, so that the PMTs receive frequent large-amplitude signals. This could affect detailed PMT responses, because their bleeders employ non-linear components as mentioned in the previous subsection. Furthermore, the response of HXD-AE may change when frequent large signals are present. The effects of large signals on PMTs and HXD-AE were studied on ground, by irradiating several isotopes to GSO, and optical pulses to the PMT at the same time (Tanihata 1998; Kawaharada 2002). According to the results, any non-linear effect did not appear in energy ranges above $\sim$ 80 keV. However, the lower energy range has remained yet to be examined.
(140mm,140mm)[figure10.eps]{}
To study effects of large signals on low-energy events, we performed laboratory experiments using a GSO scintillator, a PMT with bleeder, a CSA, and HXD-AE, which are all the flight spares or flight equivalent components. The flight-spare GSO is of the same size, made by the same process, and its energy resolution is several percent worse than those used in the flight hardware. in-orbit, a Well-counter unit experiences a typical upper-discriminator (UD) hit rate of $\sim$ 100 Hz (Paper II), which are mostly caused by primary cosmic rays passing through its BGO shields, since the UD threshold is set at $\sim$ 1 MeV. Note that a typical minimum-ionizing proton with an energy of $\sim$ 1 GeV deposits an energy of $\sim$ 80 MeV, when it passes through the BGO bottom part (with a density of 7 g cm$^{-3}$) for a path length of $\sim 6$ cm. Besides the GSO scintillator attached to the PMT, we placed a light-emiting diode (LED) therein, and made it flash by a pulse generator to simulate the cosmic-ray signals. The LED pulses were adjusted to have a decay time of 500 ns, and an energy deposit per pulse of 200 MeV BGO equivalent ( or 50 MeV GSO equivalent) with a repetition rate of 100 Hz ($\approx$ the UD hit rate). This value is about twice larger than an estimated energy deposit, $\sim$80 keV, just to exaggerate possible effects. Then, the PMT was operated at 900 V, and the output (last dynode) signal was amplified by the CSA at a time constant of 0.5 $\mu$s. The experimental setup is shown in figure \[block\].
To confirm the performance from GSO to CSA, a commercial multi-chainnel analyzer (Amptek MCA 8000A) was used as an ideal ADC, which is known to be linear to within 1 %. To test the CSA itself, we input test pulses, created by a random pulse generator at a voltage of 5 V (equivalent to $\sim$ 662 keV signals), into the CSA and measure its output using the MCA. To avoid the effects of overshoot or undershoot when a large signal (LED signal) comes into the PMT, the PMT anode signal was used to veto data acquisition in the MCA via a gate generator. As a result, the ADC channels of the test pulses were not affected by the LED signals by more than 1%. The result means that the CSA itself does not change under large signals.
We then proceed to the test from GSO to PMT, using the same setup, by irradiating $^{152}$Eu and $^{109}$Cd simultaneously to the GSO scintillator. We placed Sn and Cu plates of 5 mm thickness in front of $^{152}$Eu, as shown in gray boxes in figure \[block\], to stop lower energy ($\lesssim$ 100 keV) events. This allowed us to utilize 122 keV and 344 keV lines from $^{152}$Eu, and 22 keV line from $^{109}$Cd to determine the gain and non-linearity of GSO. Consequently, the GSO spectrum remained unchanged under the LED irradiation within $<$ 1%. We have thus confirmed that the performance from GSO to CSA, or before HXD-AE, does not change under large signals To test the HXD-AE performance under large signals, a fight-equivalent HXD-AE system (hereafter HXD-AE’) was used, under the setup shown in figure \[block\]b. We put simultaneously into the CSA both the LED signals created by the PMT and test pulses, and measured pulse heights of the test pulse using HXD-AE’. Each test pulse from the random pulse generator was fed simultaneously to the CSA (with a 1 $\mu$s delay) and “PIN-trigger” input to HXD-AE’, where the latter was used to acquire the former signal. This scheme was necessary to avoid large dead times if the CSA signals were used as self-trigger source. Under the presence and absence of the large LED signals, the pulse heights of the random test pulses were measured with HXD-AE’, for various test-pulse voltages over 0–6 V. As a result, the pulse heights measured under the two conditions agreed with $1$ % without any unexpected effects. Consequently, the effect of large signals cannot explain the change of the GSO energy scale.
Calculation of Light Outputs from Activation Lines
==================================================
Preparation for the Calculation
-------------------------------
Here we show how to calculate light outputs for the 70, 150, 196 keV lines. We denote the light outputs of GSO to electron as $L_e(E)$, and the one to photon as $L_{\gamma}$. To calculate the former, we adopted the approximation expression based on the experimental measurement of the light output to electron (Uchiyama et al. 2001), $$L_e (E) = \left[ 1 - \frac{1.5}{E} \right] \times E.$$ Based on the ground measurement (Kitaguchi et al. 2005), the latter is described as $$E = E_{\mathrm{LO}}(E) ~ \left[ 1 - \exp ( - 0.058 \times E_{\mathrm{LO}}(E) - 0.53 ) \right].$$ $$L_{\gamma}(E) = \frac{\left( E_{\mathrm{LO}} (E) - 8.1 ~ \mathrm{keV} \right)}{ \left( E_{\mathrm{LO} } (344 ~ \mathrm{keV}) - 8.1 ~ \mathrm{keV} \right) } \times \mathrm{344},$$ where $E_{\mathrm{LO}}$ is an intervening variable. The two functions, $L_e (E)/E$ and $L_{\gamma}(E)/E$ are plotted in figure 6. The binding energy, the fluorescence energies, and the fluorescence yields for several shells (K, L, M, ..) are denoted as $E_{K, L, M, ..}$, $E_{K\alpha, ..}$ and $f_{K, L, ..}$, respectively. These parameters of $^{151}$Eu are taken from the database of ENSDF[^3].
The light output caused by releasing the energy of the K(L, M, ..)-shell vacancy, after the electron bounded in K(L, M..) shell is ejected, should be calculated and defined as $LK(LL, LM, ..)$. There are two processes; the fluorescence X-ray emitted with the probability of $f_{K(L, M, ..)}$, and the Auger electron emitted with the probability of 1 - $f_{K(L, M, ..)}$. Considering both effects, we can approximately express $LK$ as $$LK = f_K \left[ L_{\gamma} (E_{K\alpha}) + L_e (E_K - E_{K\alpha}) \right] + ( 1 - f_K ) L_e (E_K),$$ and $LL$ as $$LL = f_L \left[ L_{\gamma} (E_{L\alpha}) + L_e (E_L - E_{L\alpha}) \right] + ( 1 - f_L ) L_e (E_L).$$ In the case of M shell, we assume that all energies are released via the Auger electron, expressing $LM$ as $$LM = L_e (E_M).$$
Then, we define the light output from ejection of the electron from the K(L, M, ..) shell via the internal conversion, and denoted them as $L_{eK(eL, eM, ..)}(E)$. It is written as $$L_{eK, eL, eM, ..} (E) = L_e ( E - E_{K(L, M, ..)}) + LK(LM, LL, ..),$$ where $E - E_{K(L, M, ..)}$ represents the kinetic energy of the electron escaping from K(L, M, ..)-shell.
Finally, we define the light output via the gamma decay with the released energy of $E$ as $L_O(E)$. $L_O(E)$ becomes a linear combination of $L_{\gamma}(E)$ and $L_{eK(eL, eM, ..)} (E)$. Each weight of the sum is derived from the internal conversion coefficient for each shell, $\alpha$(K, L, M, ..) (ref. ENSDF). Therefore, the expression of $L_O(E)$ is $$L_O (E) = \frac{ \left[ L_{\gamma}(E) + \alpha (K) L_{eK} (E) + \alpha (L) L_{eL} (E) + \alpha (M) L_{eM} (E) ... \right] }{ N },$$ where the $N$ is the normalization factor and defined as $$N \equiv 1 + \alpha (K) + \alpha (L) + \alpha (M) + ...$$ In the following calculation, we considered the contribution from the electron and gamma emission inside the N-shell.
Results of the Light Outputs
----------------------------
We utilized the internal conversion coefficient of $^{151}$Eu, because all the three lines are originated from the beta-decay of Gd. The light output of the 196 keV line is a sum of 174 .70 keV and 21.53 keV, as described in Paper II. By using the notation defined in section A.3.1, this can be written as $$L_{\gamma} (196.2 ~{\rm keV}) \longrightarrow L_O (174.7 ~{\rm keV}) + L_O (21.5 ~{\rm keV})$$ According to the equation (A7), we obtained $L_O (174.7 {\rm keV})=168.7$ and $L_O (21.5 {\rm keV})=18.3$, resulting in 186.9 by summing them, while $L_{\gamma} (196.2 {\rm keV})= 192.7$. Therefore, the derived light output becomes lowered by 3.0%.
Likewise, the 150 keV line is the sum of the K-shell binding energy and the average energy of 43% from 103.2 keV and 39.2% from 97.4 keV, which are expressed as $$L_{\gamma} (148.9 ~{\rm keV}) \longrightarrow \frac{[0.43 L_O (103.2 ~{\rm keV}) + 0.40 L_O (97.4 ~{\rm keV})]}{0.43 + 0.40} + LK.$$ We obtained $L_O$ (103.2 keV) = 97.3, $L_O$ (97.4 keV) = 91.4, and $LK$ = 42.3, giving 136.8 in total, while $L_{\gamma}$ (148.9 keV) = 144.3. Thus, the light output from the 150 keV line becomes lowered by 5.2 %. The 70 keV line is the sum of the K shell energy and 21.5 keV, expressed by $$L_{\gamma} (70 ~{\rm keV}) \longrightarrow L_O (21.5 ~{\rm keV}) + LK.$$ Considering $L_{\gamma}$ (70 keV) = 64.2, the light output from the 70 keV line becomes 60.5, lowered by 5.7 %.
Fukazawa, Y., [*et al.*]{} 2006, SPIE
Fukazawa, Y. et al. 2009, PASJ, 61, S17
Jourdain, E., Roques, J. P., 2009, ApJ, 704, 17
Kawaharada, M. 2002, Master thesis, Dept. of Physics, University of Tokyo
Kawaharada, M., [*et al.*]{} 2004, Proc. SPIE, 5501, 286
Kitaguchi, T., [*et al.*]{} 2006, SPIE, Vol. 6319
Kokubun, M. [*et al.*]{} 1999, [*IEEE Trans. Nucl. Sci.*]{}, 46, 371
Kokubun, M. [*et al.*]{} 1999, 59, 53-76
Koyama, K. [*et al.*]{} 2007, , 59, 23-33
Mitsuda, K. [*et al.*]{} 2007, , 59, 1-7
Ozaki, M. [*et al.*]{} 2005, IEEE, Trans. Nucl. Sci, 53, 3, 1310–1316
Serlemitsos, T. [*et al.*]{} 2006, , 59, 9-21
Sulkimo, J. et al. 2003, Nuclear Instruments and Methods in Physics Research Section A, 506, 3, 1, 250-303 Alison, J. et al. 2006, IEEE Transactions on Nuclear Science, 53, 1, 270-278
Takahashi, T. [*et al.*]{} 2007,
Tanihata, S. 1998, Master thesis, Dept. of Physics, University of Tokyo
Terada, Y. [*et al.*]{} 2005, IEEE, Trans. Nucl. Sci, 52, 4, 902
Terada, Y. [*et al.*]{} 2006, , 648, 139
Terada, Y. [*et al.*]{} 2008, , 60, SP-1, 2008
Terada, Y. [*et al.*]{} 2005, [*IEEE Trans. Nucl. Sci.*]{}, 52, 902
Uchiyama, Y., [*et al.*]{} 2001, [*IEEE Trans. Nucl. Sci.*]{}, 48, 379
Yamada, S. 2008, Master thesis, Dept. of Physics, University of Tokyo
[^1]: http://heasarc.gsfc.nasa.gov/docs/suzaku/analysis/gso\_newgain.html
[^2]: http://www.astro.isas.ac.jp/suzaku/analysis/hxd/gsoarf2/
[^3]: http://www.nndc.bnl.gov/ensdf/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This survey summarizes and illustrates the main qualitative properties of hydrodynamics models for collective behavior. These models include a velocity consensus term together with attractive-repulsive potentials leading to non-trivial flock profiles. The connection between the underlying particle systems to the swarming hydrodynamic equations is performed through kinetic theory modelling arguments. We focus on Lagrangian schemes for the hydrodynamic systems showing the different qualitative behavior of the systems and its capability of keeping properties of the original particle models. We illustrate known results concerning large time profiles and blow-up in finite time of the hydrodynamic systems to validate the numerical scheme. We finally explore unknown situations making use of the numerical scheme showcasing a number of conjectures based on the numerical results.'
address:
- 'Department of Mathematics, Imperial College London, SW7 2AZ, UK'
- 'Fakultät für Mathematik, Technische Universität München, Boltzmannstra[ß]{}e 3, 85748, Garching bei München, Germany'
- 'ETSIAE, Technical University of Madrid, Pza. de Cardenal Cisneros, 3, 28040, Madrid, Spain'
author:
- 'José A. Carrillo'
- 'Young-Pil Choi'
- Sergio Pérez
title: 'A review on attractive-repulsive hydrodynamics for consensus in collective behavior'
---
[^1]
[^2]
[^3]
Introduction {#sec:1}
============
Modelling the collective behavior of a large number of interacting individuals is a very challenging problem in animal behavior, pedestrian flow, cell adhesion and chemotaxis problems, and many other biological applications, see for instance [@CK; @PEK; @CDFSTB; @BD; @BMP; @PBSG] and the literature therein. Most of the literature is based on Individual Based Models (IBMs) which are particle descriptions from a kinetic modelling perspective. These particle systems typically include three basic effects: attraction, repulsion and alignment or re-orientation of the individuals, called the first principles of swarming. The way in which these three effects are taken into account has given rise to a large number of different and interesting models for collective behavior. These basic 3-zone models were introduced by theoretical biologists [@Aoki; @HW] for fisheries control as well as computer scientists [@Reyn] in order to mimick animal behavior in animation movies. These models have evolved toward more complete descriptions involving particular species interactions and adapted to particular animals such as birds [@HCH], fish [@HH; @KTIHC; @BTTYB], ducks [@LLE; @LLE2], and insects [@BT] for instance.
These basic particle descriptions can be coarsened to macroscopic descriptions when the number of individuals is large leading to nonlocal macroscopic models both at the level of the mass density [@MEK; @MEBS] or hydrodynamic descriptions [@CDMBC; @CDP]. This connection to continuum models is better done by passing to the intermediate description provided by kinetic modelling. The kinetic theory approach via mean-field limits of interacting particle systems has offered a mathematical underpinning to derive kinetic equations in a rigorous manner from particle descriptions. The connection towards macroscopic equations is done either via closure assumptions or moment approximations [@CDMBC; @CDP] and large friction limits [@LT]. One of the most famous particle models was introduced by Vicsek and his collaborators [@VCBCS] showing a phase transition behavior that has also been studied through kinetic modelling and self-organized hydrodynamics [@dm2008; @dfl2013; @dfl2015]. We refer to [@CFTV; @KCBFL] and the references therein for a good account of the different levels of description and the state of the art of these models in the applied mathematical community.
We will focus in this review on two velocity consensus models [@CS0; @CS; @MT; @MT2] that lead to asymptotic convergence for the large time toward a fixed velocity under certain conditions, a phenomenon that is called asymptotic flocking. These models have been studied extensively in the last years due to their apparent simplicity in formulating the possibility of consensus in velocity. These models are connected to the Vicsek model in which all particles travel to a fixed speed by large friction limits [@BC]. They also present a phase transition in terms of noise as the original Vicsek model [@BCCD]. In this survey, we concentrate in the basic properties of the hydrodynamic models incoporating also the effects of attraction and repulsion through an interaction potential.
In Section 2, we give a brief account of the particle descriptions making particular emphasis to the consensus in velocity models with interaction potentials and their flock solutions. Section 3 is first devoted to explain the link between these particle models and hydrodynamic descriptions via kinetic modelling. We propose a Lagrangian approach to solve the one dimensional hydrodynamic descriptions. We numerically explore different qualitative aspects of the hydrodynamic models such as critical thresholds [@CCTT] and their sharpness for consensus models with and without interaction potentials. We also analyse the effect of the singularity of the potential in the long time asymptotics of global solutions.
Microscopic descriptions: Discrete models
=========================================
In this section, we review some of the basic individual based attractive-repulsive models containing an additional velocity alignment force. The social interaction between individuals of the swarm is modelled by an effective interaction potential encapsulating the short-range repulsion and the long-range attraction forces at the particle level as discussed in the introduction. On top, we will also consider cases in which there is a tendency of behaving similarly to other individuals of the group, this mimicking behavior can be modelled in many different ways. One of the simplest manners of incorporating this gregarious behavior is to assume that each individual averages its relative velocity vector with nearby individuals according to some weights that we call the communication function. All the modelling in these simple descriptions are reduced to find biologically reasonable potentials and communication functions for the particular application or adapted to a particular species. Many authors have studied what are the most probable interaction regions for different animals, see [@HH; @LLE2; @KTIHC] and the references therein. We will showcase some of the different behaviors in these models by choosing toy-example for potentials and communication functions. Although not too biologically reasonable, these choices give us generic behaviors for these models.
For the velocity alignment force, we use two different types of forces proposed by Cucker and Smale [@CS0; @CS] and Motsch and Tadmor [@MT]. More precisely, let $(x_i,v_i)$ be the position and velocity of $i$-th individual. Then our main system reads as $$\label{eq_par}
\begin{cases}
{\displaystyle \frac{dx_{i}}{dt}=v_{i},\quad i = 1,\cdots,N,\quad t>0,}\\[2mm]
{\displaystyle \frac{dv_{i}}{dt}=\frac{1}{S_i(x)}\sum_{j=1}^{N}\psi\left(x_{i}-x_{j}\right)\left(v_{j}-v_{i}\right) - \frac1N\sum_{j \neq i} \nabla K(x_i - x_j).}
\end{cases}$$ The first term on the right hand side of $\eqref{eq_par}_2$ represents a nonlocal velocity alignment force, where $\psi$ is the communication function. The second term on the right hand side of $\eqref{eq_par}_2$ serves as attractive/repulsive forces through the interaction potential $K(x)$. Typical assumptions on $K(x)$ are that is radially symmetric and smooth outside the origin possibly decaying to zero for large distances, one particular example widely used in the literature is the Morse potential, see [@DCBC; @CMP; @ABCV; @CHM1] for more details. Here the scaling function $S_i(t)$ and the communication function $\psi$ are given by $$\label{def_s}
S_i(x) := \left\{ \begin{array}{ll}
N & \mbox{ for the Cucker-Smale model},\\[1mm]
\displaystyle \sum_{k=1}^N \psi(x_i - x_k) & \mbox{ for the Motsch-Tadmor model},
\end{array}\right.$$ and $$\psi(x)=\frac{1}{(1+|x|^{2})^{\beta/2}},\quad\beta\geq 0,$$ respectively. These scalings are related to the mean-field limit for the system of the N interacting particles. Assuming that the effect of each individual on another one via the social force decays as $1/N$ is intuitive, if we want to obtain some non trivial limit as $N\to \infty$, since we should keep the total kinetic and potential energy and velocity of each individual to be of order 1 in that limit. More discussions about the mean-field limit can be found in [@BH; @Dob; @Spo; @Szn; @Gol; @Hau; @CCR; @CCR2; @BCC; @CFTV; @AIR; @BV; @FHM; @CCH; @CCH2].
In [@CS], Cucker and Smale introduced the Newton-type particle system for flocking phenomena. The local averaging of relative velocities is weighted by the communication function $\psi$ in such a way that closer individuals have stronger influence than further ones. Note that the velocity alignment force of the Cucker-Smale (in short CS) model is scaled with the total mass. Later, Motsch and Tadmor proposed in [@MT] a new model for self-organized dynamics. They pointed out that the CS model is inadequate for far-from-equilibrium scenarios since the communication function is normalized by the total number of agents $N$. By taking into account the velocity-alignment force normalized with a local average density, the Motsch-Tadmor (in short MT) model takes into account not only the relative distance between agents but also their relative weights compared to the CS model. Note that the MT model does not have the symmetry property due to the normalization.
Velocity-alignment models without interaction forces {#sec_p1}
----------------------------------------------------
We begin our discussion with the case in which the individuals are only interacting through the velocity alignment force as $$\begin{cases}
{\displaystyle \frac{dx_{i}}{dt}=v_{i},\quad i = 1,\cdots,N,\quad t>0,}\\[2mm]
{\displaystyle \frac{dv_{i}}{dt}=\frac{1}{S_i(x)}\sum_{j=1}^{N}\psi\left(x_{i}-x_{j}\right)\left(v_{j}-v_{i}\right),\quad\psi(x)=\frac{1}{(1+|x|^{2})^{\beta/2}},\quad\beta\geq0,}
\end{cases}\label{eq_cs}$$ with the initial data $$(x_{i}(0),v_{i}(0))=:(x_{i0},v_{i0}),\quad i = 1,\cdots,N.\label{ini_eq_cs}$$ Here the scaling function $S_i(x)$ is given in . We notice that the standard Cauchy-Lipschitz theory yields the existence and uniqueness of global in time smooth solutions to the system with $S_i(x) \equiv N$ since the communication function $\psi$ is bounded and globally Lipschitz. For the MT model, we can also show that the communication function $\psi$ is bounded from below for any time $T < \infty$, and this again enables us to apply the Cauchy-Lipschitz theory to the MT model to have the existence and uniqueness of solutions. Let us first remind the main analytical results concerning the flocking behavior for the system . Then, we present several numerical results to illustrate the analytical ones. We also compare the time behavior of solutions to the CS and MT systems, i.e., with $S_i(x) \equiv N$ and $S_i(x) = \sum_{k=1}^N \psi(x_i - x_k)$.
For the large-time behavior of solutions, we first introduce the definition of universal asymptotic flocking for the system .
Let $(x_i,v_i)_{i=1}^N$ be a given solution of the particle system -. Then the $(x_i,v_i)_{i=1}^N$ leads to asymptotic flocking if and only if it satisfies the following two conditions: $$\lim_{t \to \infty}\max_{1 \leq i,j \leq N}|v_i(t) - v_j(t)| = 0 \quad \mbox{and} \quad \sup_{0 \leq t < \infty}|x_i(t) - x_j(t)| < \infty.$$
We then define diameters in position and velocity phase spaces as follows: $$R^{x}(t):= \max_{1\leq i,j\leq N}\left|x_{i}(t)-x_{j}(t)\right|,\quad R^{v}(t):=\max_{1\leq i,j\leq N}\left|v_{i}(t)-v_{j}(t)\right|.\label{def:xv}$$
For the system , rigorous estimates showing the emergence of flocking depending on the decay rate of the communication function are provided in [@CS0; @CS]. Later, the flocking estimates are refined in [@CFRT; @HL; @HT; @MT; @Tan]. Flocking models with vision cones or topological interactions are studied in [@AP; @CCHS; @Hasko] and with noise [@Choi; @DFT].
In the theorem below, we sumarize the flocking estimates for the system . The proof follows the blueprint of [@CFRT; @ACHL; @HL; @MT; @Tan], so we omit it here.
\[thm\_fl\]Let $(x,v)$ be any global smooth solution to the CS system -.
- If $0\leq\beta\leq1$, then we have unconditional asymptotic flocking, that is $$R^{v}(0)e^{-t}\leq R^{v}(t)\leq R^{v}(0)e^{-\psi(\tilde{R})t}\quad t\geq0,
\label{new}$$ where $\tilde{R}$ is implicitly given by $$R^{v}(0)=\int_{R^{x}(0)}^{\tilde{R}}\varphi(s)\,ds.$$
- If $\beta>1$ and the initial diameters $R^{x}(0)$ and $R^{v}(0)$ satisfy $$R^{v}(0)<\int_{R^{x}(0)}^{\infty}\psi(s)\,ds,\label{a:gene-1}$$ then estimate also holds.
In Figs. \[fig:sim-beta08\] and \[fig:beta>1\], we observe typical particle simulations of the CS model in 2D. As stated in Theorem \[thm\_fl\], the unconditional asymptotic flocking occurs for any initial data in the case simulated in Fig. \[fig:sim-beta08\] while Fig. \[fig:beta>1\] shows a comparison between flocking and non flocking cases.
In Fig. \[fig:sim-beta08\], the initial positions and velocities $\{(x_{i0},v_{i0})\}_{i=1}^N$ with $N=50$ are generated randomly from the uniform distribution $[-10, 10]^2 \times \big\{ [-5,5]\times [-4.3,5.7] \big\}$ with the aim of having the mean velocity equal to $\frac{1}{N} \sum_{i=1}^N v_i(0) \approx (0,0.7)$.
In Fig. \[fig:beta>1\], the flocking behavior happens depending on the initial data. The initial positions and velocities $\{(x_{i0},v_{i0})\}_{i=1}^N$ with $N=50$ are generated randomly from the uniform distribution $[-10, 10]^2 \times \big\{ [-5,5]\times [-4.3,5.7] \big\}$ with the aim of having the mean velocity equal to $\frac{1}{N} \sum_{i=1}^N v_i(0) \approx (0,0.7)$. With this initial configuration, it results that $R^x(0)=26.23$ and $R^v(0)=12.25$. Then, the initial data for (A) satisfies since $R^{v}(0)<\int_{R^{x}(0)}^{\infty}\psi(s)\,ds=16.43$. On the other hand, the initial data for (B) do not satisfy the condition because $R^{v}(0)>\int_{R^{x}(0)}^{\infty}\psi(s)\,ds=2.60$.
For the CS model, i.e., $S_i(x) \equiv N$ in , if we set an averaged quantity $v_c(t):= \frac1N \sum_{i=1}^N v_i(t)$, then $v_c(t)$ satisfies $v_c^\prime(t) = 0$, i.e., $v_c(t) = v_c(0)$ due to the symmetry of the communication function $\psi$. Thus, if the global flocking occurs, then we have that for all $i \in \{1, \cdots, N\}$ $$v_i(t) \to v_c(0) = \frac1N\sum_{i=1}^Nv_i(0) \quad \mbox{as} \quad t \to \infty.$$ On the other hand, in the MT model, i.e., $S_i(x) = \sum_{k=1}^N \psi(x_i - x_k)$, the momentum is not conserved. Thus identifying the asymptotic flocking state in terms of the initial data is a very intriguing question. Partial answers to asymptotic flocking have been provided in [@MT].
In Fig. \[fig:Initial-conditions-mt\], we show the different behavior of the CS and MT velocity averaging. We choose the initial positions $\{x_{i0}\}_{i=1}^{55}$ divided into two groups, $G_1 := \{x_{i0}\}_{i=1}^{50}$ and $G_2 := \{x_{i0}\}_{i=1}^5$, they are generated randomly from the uniform distribution $[-10,10]^2$ and $[60,63] \times [-1.5,1.5]$, respectively. The initial velocities $\{v_{i0}\}_{i=1}^{55}$ are generated randomly from the uniform distribution $[-5,5]\times [-4.3,5.7] $ with the aim of having the mean velocity equal to $\frac{1}{N} \sum_{i=1}^N v_i(0) \approx (0,0.7)$. We observe the much faster decay of the velocity radius of the support $R^v(t)$ defined in in the MT model compared to the CS model, and thus the asymptotic flocking is achieved faster in the MT model than in the CS model.
Finally, we show in Fig. \[fig:Evolution-mt-5s\] a comparison between the time evolution of the system with $S_i(x) \equiv N$ and $S_i(x) = \sum_{k=1}^N \psi(x_i - x_k)$. Subplots (a) and (b) show a snapshot of the solutions at $t=5$, while subplots (c) and (d) show a snapshot of the solutions at $t=50$, for the CS and the MT models respectively. From Fig. \[fig:Evolution-mt-5s\] (a), we find that the CS flocking particles in the small group $G_2$ are not interacting with others in the large group $G_1$ in the beginning. It seems that the particles tend to move with their own initial velocities. On the other hand, the particles in the MT model are trying to be aligned with their neighbors from the beginning. Even though both models exhibit flocking behavior and the analytical results require the same conditions for flocking, numerical simulations demonstrate again that the decay rate of convergence of the MT model to the flocking state is faster than the one of the CS model. We are not aware of results comparing the rate of decay to flocking for both models.
Attractive-Repulsive models {#sec_p2}
---------------------------
Now, we turn to the case in which we incorporate attractive-repulsive forces to the system with the CS alignment force: $$\begin{cases}
\dot{x}_{i}=v_{i},\quad i = 1,\cdots,N, \quad t > 0,\\[2mm]
\displaystyle \dot{v}_{i}=\frac{1}{N}\,\sum_{j=1}^{N}\psi\left(x_{i}-x_{j}\right)\left(v_{j}-v_{i}\right)-\frac{1}{N}\,\sum_{j=1}^{N}\nabla K\left(x_{i}-x_{j}\right).
\end{cases}\label{cs-simul}$$ For the interaction potential $K$, we will choose in most of our simulations a repulsive Newtonian potential confined by a quadratic one of the form $$\label{int_pot}
K(x) = \alpha\frac{\left|x\right|^{2}}{2} + k \phi(x) \quad \mbox{where} \quad -\Delta_x \phi(x) = \delta_0(x).$$ By choosing $\alpha >0$, we confine our particles in a bounded region and they are repelled by Newtonian interaction choosing $k<0$. In general, $K$ is typically chosen repulsive at the origin and attractive at infinity in such a way that there is a typical length of stable interaction between two particles. Other popular choices as mentioned above are Morse-like and power-law like potentials as in [@DCBC; @CMP; @CHM1; @CH] and the references therein.
The existence of particular solutions, called flock solutions, for the system has recently received lots of attention due to the ubiquitous appearance of this kind of solutions in several swarming models.
A flock solution of the particle model is a spatial configuration $\hat{x}$ with zero net interaction force on every particle, that is: $$\sum_{j\neq i}\nabla K(\hat x_i-\hat x_j)=0, \quad i=1,\cdots,N\,,$$ that translates at a uniform velocity $m_0\in{\mathbb{R}}^d$, hence $(x_i(t),v_i(t))=(\hat{x_i}-tm_0,m_0)$. \[def-flock\]
The richness of the qualitative properties of the spatial configurations for the flock solution, also called flock profile, depending on the potential $K$ is quite impressive, see [@KSUB]. Other stable patterns were observed for related systems, for instance single or double rotating mills [@LRC; @DCBC; @CDP; @CKMT; @CKR]. However, these milling patterns are typically eliminated due to the presence of the CS alignment term. The stability of flock patterns for the particle system has recently been established in [@ABCV; @CHM2].
As the total number of individuals gets large, the system of differential equations is difficult to analyse and usually a continuum description based on mean-field limits is adopted, either at the kinetic level for the particle distribution function [@CDP; @CFTV] or at the hydrodynamic level for the macroscopic density and velocities [@CDMBC; @CDP] as we will discuss in the next section. At the continuum level, the flock profiles are characterized by searching for continuous probability densities or probability measures $\rho$ of particle locations such that the total force acting on each individual balances out. This is equivalent to finding probability densities or measures $\rho$ such that $$\label{eq:basic}
\nabla K\ast \rho = 0 \quad\mbox{ on supp}(\rho) \,.$$ Being the problem posed on the support of the unknown density $\rho$ implies that the equation is highly nonlinear. In fact, characterizing the interaction potentials $K$ such that these profiles are continuous or regular in their support is a very challenging question. Explicit formulas for solutions to for particular potentials such as Morse-like and power-law like potentials are possible due to the particular properties of associated differential operators [@LBT; @BT; @FHK; @CMP; @CHM1; @CH]. In particular, it is known from classical potential theory that the solution to in the case of the confined repulsive Newtonian potential is given by a characteristic of a ball whose radius is determined to have the right total mass of the system.
In Fig. \[fig:cs\_simul\], we show a typical simulation for the system with interaction potential given in . The initial positions and velocities $\{(x_{i0},v_{i0})\}_{i=1}^N$ with $N=50$ are generated randomly from the uniform distribution $[-10, 10]^2 \times \big\{ [-5,5]\times [-4.3,5.7] \big\}$ with the aim of having the mean velocity equal to $\frac{1}{N} \sum_{i=1}^N v_i(0) \approx (0,0.7)$. This simulation shows the generic flock formation after some time in which particles distribute more or less uniformly in a certain ball. Of course, we know that as $N\to \infty$ this distribution of particles will be getting closer and closer to the characteristic of the ball, see [@BLL; @KSUB; @BKSUV] for instance.
Let us finally mention that the continuum spatial profile of the flock solutions can also be found by steepest descent methods from first-order models of swarming. These models can be found formally from the previous second-order particle models by substituting the CS alignment term by simple linear friction force and assuming that inertia is negligible with respect to other terms, see [@MEK; @MEBS]. In this limit, they lead to $$\label{1storder}
\frac{d}{dt}x_i=-\frac{1}{N} \sum_{j\neq i}\nabla K(|x_i-x_j|), \quad i=1,\cdots,N\,.$$ By taking the mean-field limit in , as $N\to \infty$, one derives the so-called aggregation equation $$\label{aggreg}
\rho_t = \nabla\cdot(\rho\nabla K*\rho)\,,$$ for the evolution of the mass density of particles. The aggregation equation with repulsive-attractive potentials has attracted lots of attention in the last years in the mathematical analysis community due to its rich regularity structure for steady states and solutions depending on the singularity of the potential at the origin, see [@CMV; @CMV2; @Lau; @TB; @TBL; @BL; @BCL; @BLR; @BLL; @BV; @BC; @BCLR; @BCLR2; @CDFLS; @CDFLS2; @CCH] and the references therein.
Macroscopic descriptions: Flocking behavior and finite-time blow-up phenomena
=============================================================================
In this section, we study some of the features and properties of continuum models for collective behavior being capable of describing flocking behavior. By using BBGKY hierarchies or mean-field limits [@CFTV], we can derive a Vlasov-type equation from the system . More precisely, when the number of individuals goes to infinity, i.e., $N \to \infty$, the mesoscopic observables for the system can be calculated from the velocity moments of the density function $f = f(x,v,t)$ which is a solution to the following Vlasov-type equation: $$\label{ki_cs}
\begin{cases}
\partial_{t}f+v\cdot\nabla_{x}f+\nabla_{v}\cdot\left(F(f)f - (\nabla_x K \star \rho) f\right)=0,\quad x,v\in{\mathbb{R}}^{d}, t>0,\\[2mm]
{\displaystyle \mathbf{\mathrm{\mathit{\mathit{F}(f)= \int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}\psi(x-y)(w-v)f(y,w,t)\,dydw}}}},\\[2mm]
{\displaystyle \rho(x,t) = \int_{{\mathbb{R}}^d} f(x,v,t)\,dv }.
\end{cases}$$ The derivation of the kinetic equation is well studied only for regular potentials [@BH; @Dob; @Neun]. If $K$ vanishes, rigorous mean-field limit, existence of weak solutions, and large-time behavior of measure-valued solutions are studied in [@CFRT; @HL; @CCH]. We also refer to [@BCHK; @BCHK2; @BCHK3; @BCHK4; @CCK; @Choi2] for a dynamics of flocking particles interacting with homogeneous/inhomogenous fluids. For the equation under certain conditions for $K$, quite general frameworks are proposed in [@BCC; @CCR2; @CCH2]. For not too singular interaction potentials $K$, the rigorous derivation of is studied in [@HJ].
The kinetic description has to be taken as usual as an intermediate mesoscopic description of the system leading to macroscopic models for collective behavior via asymptotic limits or closure assumptions. By taking moments on the kinetic equation together with a zero temperature closure or monokinetic assumption for the local hydrodynamics solution, one can obtain hydrodynamic descriptions of the system . This procedure has been perforemed in different ways by different authors, see for instance [@CDP; @CDMBC; @HT]. Associated to the kinetic distribution function $f(x,v,t)$, one can define the mean velocity as $$\rho(x,t) u(x,t) = \int_{{\mathbb{R}}^d} v f(x,v,t)\,dv\,.$$ By taking the first two moments with respect to $v$ on the kinetic equation , one formaly obtains $$\begin{cases}
\partial_{t}\rho+\nabla_x\cdot\left(\rho u\right)=0,\quad x\in{\mathbb{R}}^{d},\quad t>0,\\[2mm]
{\displaystyle {\partial}_{t}(\rho u)+\nabla_x\cdot(\rho u\otimes u)+ \nabla_x\cdot\left(
\int_{{\mathbb{R}}^d} (v-u)\otimes (v-u) f(x,v,t)\,dv \right)=}\\[2mm]
{\displaystyle \qquad\qquad\qquad\qquad\,\,\,\rho\int_{{\mathbb{R}}^{d}}\psi(x-y)(u(y)-u(x))\rho(y)\,dy - \rho {\left}(\nabla_x K \star \rho{\right}).}\\
\end{cases}\label{h2_CS}$$ Assuming that the distribution function is not far from monokinetic, that is $$f(x,v,t) \simeq \rho (x,t) \, \delta (v - u(x, t)),$$ the hydrodynamic system is reduced to the following pressureless Euler-type equations given by $$\begin{cases}
\partial_{t}\rho+\nabla_x\cdot\left(\rho u\right)=0,\quad x\in{\mathbb{R}}^{d},\quad t>0,\\[2mm]
{\displaystyle {\partial}_{t}(\rho u)\!+\!\nabla_x\!\cdot\!(\rho u\otimes u)\!=\!\rho\!\!\left(\!\int_{{\mathbb{R}}^{d}}\!\!\psi(x-y)(u(y)-u(x))\rho(y)\,dy \!- \!\nabla_x K \star \rho\!\!{\right}),}
\end{cases}\label{h_CS}$$ where $\rho = \rho(x,t)$ and $u = u(x,t)$ represent the particle density and their corresponding mean velocity, respectively. For now on, $\Omega(t)$ denotes the interior of the support of the density $\rho$, i.e., $\Omega(t) := \{x \in {\mathbb{R}}: \rho(x,t) >0\}$. We assume that $\Omega(0)=:\Omega_0$ is a bounded open set. The hydrodynamic system has to be complemented with initial conditions $$\label{ini_h_CS}
\left.{\left}(\rho(\cdot, t),u(\cdot, t){\right})\right|_{t=0}=\left(\rho_{0},u_{0}\right)\,.$$
As usual with hydrodynamic equations, we observe that they are mathematically challenging due to the nonlinearity introduced by the material derivative of the velocity field that can lead to blow-up of the velocity profile. On the other hand, flock profiles are also solutions of the hydrodynamic equations . Actually, if the density $\rho$ satisfies and the velocity is constantly given by $u(x,t)=u_0\in {\mathbb{R}}^d$, then they form a particular solution of the hydrodynamic equations . This is the flocking solution at the hydrodynamical level of description for collective behavior.
In the next subsections, we will numerically explore the derived hydrodynamic equations in one dimension providing numerical evidence showing the flocking behavior and the finite-time blow-up of solutions giving some insight in the stability of flock solutions and the conditions for blow-up of the solutions. Analytical results concerning these hydrodynamic equations are very few in the literature, we will compare to existing analytical results in the relevant sections below.
Numerical scheme {#subsec:numerics}
----------------
For the numerical simulation of the hydrodynamic system , we use a Lagrangian numerical scheme. With this purpose, we consider the characteristic flow $\eta(x,t)$ associated to the fluid velocity $u$ defined by $$\label{eq:char}
\frac{d \eta(x,t)}{dt} = u(\eta(x,t),t) =: v(x,t) \quad \mbox{with} \quad \eta(x,0) = x.$$ Set $h(x,t) := \rho(\eta(x,t),t)$, then using the characteristic flow , we can rewrite the system in one dimension as $$\begin{aligned}
\begin{aligned}
\begin{cases}
\displaystyle h(x,t) = \displaystyle \rho_0(x){\left}(\frac{{\partial}\eta}{{\partial}x}(x,t){\right})^{-1},\\[2mm]
\displaystyle \frac{dv}{dt}(x,t) = \displaystyle\int_{{\mathbb{R}}}\psi(\eta(x,t)-\eta(y,t))\left(v(y,t)-v(x,t)\right)\rho_0(y)\,dy\\[2mm]
\qquad\qquad \displaystyle \,\,\,-\int_{{\mathbb{R}}} \frac{\partial K}{\partial x}(\eta(x,t)-\eta(y,t))\rho_0(y)\,dy,
\end{cases}\label{eq:eqs-lagrange_notation}
\end{aligned}\end{aligned}$$ with the initial data $$\left.{\left}(h(\cdot, t),v(\cdot, t){\right})\right|_{t=0}=\left(\rho_{0},u_{0}\right).$$ Note that the continuity equation for the density $\eqref{eq:eqs-lagrange_notation}_1$ is decoupled from the rest. Thus it can easily be solved from the information obtained from the characteristic and momentum equations.
We do a spatial discretization of the equation $\eqref{eq:eqs-lagrange_notation}_2$ choosing a uniform mesh of length $\Delta x$ of the initial positions of the particles with a number of points given by $n$. For each node $i$, the position, density and velocity through the characteristics of the ith-particle will be denoted as $\eta_i(t)$, $h_i(t)$ and $v_i(t)$. At each node the term $\frac{{\partial}\eta_i(t)}{{\partial}x}$ is computed from $\eta_j(t)$, that is with the information about how the position of the nodes change through the characteristics in time, by standard finite differences of fourth order, and afterwards that value is inverted. The end values use one-sided finite differences to avoid unknown values of the Lagrangian density. The spatial derivative of the interaction potential appearing in $\eqref{eq:eqs-lagrange_notation}_2$ is obtained analytically and not approximated numerically. For the integral terms in $\eqref{eq:eqs-lagrange_notation}_2$, we approximate them by direct numerical quadrature formulas as $$\begin{aligned}
\int_{\Omega_0}\psi(\eta(x,t)-\eta(y,t)) v(y,t) \rho_0(y)\,dy &\sim\, \Delta x\sum_{j=1,j\neq i}^{n} \psi(\eta_i(t)-\eta_j(t))\,v_j(t)\,\rho_j(0), \cr
\int_{\Omega_0}\psi(\eta(x,t)-\eta(y,t))\rho_0(y)\,dy &\sim\,\Delta x\sum_{j=1,j\neq i}^{n} \psi(\eta_i(t)-\eta_j(t))\,\rho_j(0),\cr
\int_{\Omega_0} \frac{\partial K}{\partial x}(\eta(x,t)-\eta(y,t))\rho_0(y)\,dy &\sim \,\Delta x\sum_{j=1,j\neq i}^{n} \frac{\partial K}{\partial x}(\eta_i(t)-\eta_j(t))\,\rho_j(0).
\end{aligned}$$ Taking everything into consideration, the system of equations , for each node $i$, takes the form $$\begin{cases}
\displaystyle \frac{d\eta_i(t)}{dt}=v_i(t),\\[2mm]
\displaystyle h_i(t)=\rho_i(0){\left}(\frac{{\partial}\eta_i(t)}{{\partial}x}{\right})^{-1},\\[2mm]
\displaystyle \frac{dv_i(t)}{dt}\!\!= \Delta x\!\!\!\!\!\!\!\sum_{j=1,j\neq i}^{n} \!\!\!\!\!\!\!\rho_j(0)\!\left[\psi(\eta_i(t)\!-\!\eta_j(t))\!\!\left(v_j(t)\!-\!v_i(t)\right)\!-\!\!\frac{\partial K}{\partial x}\!(\!\eta_i(t)\!-\!\eta_j(t)\!)\right]
\end{cases}\label{eq:spatial-discret}$$ A temporal discretization is carried out in the system composed by the equations $\eqref{eq:spatial-discret}_1$ and $\eqref{eq:spatial-discret}_3$, in order to obtain the evolution of position and velocity through the caracteristics. The employed numerical scheme is a classical fourth-order Runge-Kutta explicit scheme using the built-in $ode45$ Matlab command. Subsequently, the density is obtained from $\eqref{eq:spatial-discret}_2$. The numerical experiments have been performed with the version R2014a of Matlab, using a computer ACER Aspire V3-572G. Similar Lagrangian approaches have been taken by other authors in related problems [@CKMT; @KT].
In all the test cases below, we choose the same initial conditions for the particle positions, the initial position of the nodes has been set uniformly distributed inside the interval $[-0.75,\,0.75]$. Thus, the initial position of each node $i$ is given by $$\eta_i(0)=-0.75+\frac{1.5}{n-1}\left(i-1\right)\quad \mbox{for} \quad i=1,\cdots,n.$$ In most of our simulations $n=200$ if it is not specified otherwise. The initial conditions of density and velocity will be specified for each case that will be treated subsequently.
Euler-alignment system {#subsec:euler-alig}
----------------------
The aim of this subsection is to analyze numerically some of the open problems related to the theoretical results studied in [@CCTT] on critical threshold phenomena for the system without the interaction forces. We will take the communication function of CS model given by $$\psi(x)=\frac{1}{(1+|x|^{2})^{\beta/2}}, \quad \beta > 0.$$ In this case, the global regularity or the finite time blow up of the solution can be determined by the initial configurations with sharp conditions. We use this case as validation to our scheme being capable of showing either the global consensus in velocity of the blow-up in the velocity field and density as shown by the theory [@CCTT]. We also numerically compare the large-time behaviors of solutions to the CS and the MT models at the hydrodynamic level. The simulations in this subsection are done with initial density $$\rho_{i}(0)=\frac{1}{\gamma}\cos\left(\pi\,\frac{x_{i}(0)}{1.5}\right)\quad \mbox{for each node } i=1,\cdots, n,$$ where the constant $\gamma$ is fixed by the mass normalization, i.e., $\sum_{i=1}^n \rho_{i}(0) = 1$. Concerning the initial velocity, we choose $$u_{i}(0)=-c\text{\,}\sin\left(\pi\,\frac{x_{i}(0)}{1.5}\right)\quad \mbox{for each node } i=1,\cdots, n,$$ where the constant $c >0$ will be varied to study different initial conditions in the simulations except for the comparison between the MT and CS models.
### Hydrodynamic Cucker-Smale model
In this case, a critical threshold leading to sharp dichotomy between global-in-time existence or finite-time blowup of solutions is provided in [@CCTT]. The region where the solutions blow up in a finite time is called “[*Supercritical region*]{}”, otherwise the region in which the solutions globally exist in time it called “[*Subcritical region*]{}”.
\[thm\_cri\] Let $(\rho,u)$ be classical solutions to the system - with $K=0$.
- (Subcritical region) If $\partial_{x}u_{0}(x)\geq-\psi\star \rho_{0}(x)$ for all $x\in\mathbb{R}$, then the system has a global classical solution.
- (Supercritical region) If there exists an $x$ such that $\partial_{x}u_{0}(x)<-\psi\star \rho_{0}(x)$, then there is a finite time blow up of the solution. Moreover, this blow-up happens as an infinite negative slope in the velocity and divergence value of the density at the same location.
For the numerical simulations, three different cases corresponding with the values of the constant $c =0.2, 0.4, 0.5$ are treated. The first two cases lie in subcritical region, and the third one lies in the supercritical region, see Fig. \[fig:IC-eu-alig\].
![\[fig:IC-eu-alig\]Values of $\partial_{x}u_{0}(x)+\psi\star \rho_{0}(x)$ for different values of $c$.](CheckIC)
Since the initial configurations of the first two cases lie in the subcritical region, we have global regularity of solutions. Both initial configurations are symmetric, thus the initial mean velocity and center of mass are zero, which are kept through their evolution. The numerical simulations in Fig. \[fig:eu-alig-c02\] demonstrate that, and they are consistent with Theorem \[thm\_cri\] leading to global in time solutions. Even though both cases, $c = 0.2, 0.4$ are inside the subcritical region, the steady density for the case $c = 0.4$ shows more concentrated behavior at the middle of the domain(Fig. \[fig:eu-alig-c02\]. (c)) than the case $c = 0.2$(Fig. \[fig:eu-alig-c02\]. (a)). For both cases, the velocity converges to zero as time goes on which gives the global consensus or flocking behavior in this case. We see that the profile of density depends in a complicated way on the initial density configuration as it happens in the particle system.
When $c=0.5$, the initial data lies in the supercritical region, and thus a blow up should be expected from Theorem \[thm\_cri\]. Indeed, the numerical scheme is capable to show this blow-up phenomenon. We can observe that the derivative of the velocity becomes sharper and sharper at the origin, see the inlet in Fig. \[fig:eu-alig-c05\](b), while the density value at the origin gets larger and larger, Fig. \[fig:eu-alig-c05\](a). Actually, the simulation can not be continued after $t = 2.7311$ due to the high value of the density and the large negative derivated of the velocity at the origin. Fig. \[fig:eu-alig-c05\] depicts the time-behavior of density until $t=2$.
### Numerical comparison of Cucker-Smale and Motsch-Tadmor equations
The aim of this subsection is to compare the solutions of the hydrodynamic CS and MT systems. Taking a similar strategy for the CS model, the hydrodynamic MT system can be formally derived from the kinetic and the particle levels of the MT model, leading to $$\label{h_MT}
\begin{cases}
\partial_{t}\rho+ {\partial}_x \left(\rho u\right)=0,\quad x\in{\mathbb{R}},\quad t>0,\\[2mm]
\displaystyle {\partial}_{t}(\rho u)+{\partial}_x (\rho u^2) = \frac{1}{\psi\ast \rho}\int_{{\mathbb{R}}}\psi(x-y)(u(y)-u(x))\rho(x)\rho(y)\,dy.
\end{cases}$$ For the system , the large-time behavior of solutions showing the velocity alignment is studied in [@MT] and the critical thresholds phenomena is provided in [@TT].
For the numerical simulations, we take the parameter $\beta = 0.5$ in the communication function $\psi$. The initial conditions have been chosen following the simulations carried out at the particle level, see Fig. \[fig:Initial-conditions-mt\] and \[fig:Evolution-mt-5s\]. The objective has been to establish two zones in the initial domain with an important difference in mass. In addition, those two regions start with opposite velocity, so that the direction of the velocity corresponding to the largest zone will prevail.
The initial distances of Fig. \[fig:Initial-conditions-mt\] have been preserved in the simulation, although here they have been divided by $10$ for visualization. Then, the initial positions of the nodes are $$\eta_i(0)=-1+\frac{7.5}{n-1}\left(i-1\right)\quad \mbox{for} \quad i=1,\cdots,n.$$ Concerning the initial density, the sum of the mass of both regions is unit, and the relation between the individual masses is the same as in the particle level, $50/5$. Then the initial density is a piecewise function that satisfies $$\rho_{i}(0)=
\begin{cases}
\frac{1}{\gamma_1}\cos\left(\pi\,\dfrac{x_{i}(0)}{2}\right)\quad \mbox{if } \eta_i(0) \in [-1,1],\\[2mm]
0\quad \mbox{if } \eta_i(0) \in (1,5.5),\\[2mm]
\frac{1}{\gamma_2}\cos\left(\pi\,\dfrac{x_{i}(0)-6}{1}\right)\quad \mbox{if } \eta_i(0) \in [5.5,6.5]\\[2mm]
\end{cases}\quad \mbox{for} \quad i=1,\cdots,n,$$ where $\gamma_1$ and $\gamma_2$ are chosen to satisfy the conditions on the masses. With respect to the initial velocities, both groups start with opposite velocities given by $$u_{i}(0)=
\begin{cases}
0.1\cos\left(\pi\,\dfrac{x_{i}(0)}{2}\right)\quad \mbox{if } \eta_i(0) \in [-1,1],\\[2mm]
0\quad \mbox{if } \eta_i(0) \in (1,5.5),\\[2mm]
-0.1\cos\left(\pi\,\dfrac{x_{i}(0)-6}{1}\right)\quad \mbox{if } \eta_i(0) \in [5.5,6.5]\\[2mm]
\end{cases}\quad \mbox{for} \quad i=1,\cdots,n.$$
In Fig. \[fig:mt-cs-hydro\]. (a) and (b) the evolution of the density and velocity are showed, at $t=20$. As it happened in the microscopic case, for MT system the small group tends to keep the initial configuration, while for the CS case it varies more.
In order to compare the convergence rate to steady states of these systems, we consider the following quantity which measures the difference between velocities on the support of density: $$R_{\rho}^{v}(t):=\sup_{x,y\in\,\Omega(t)}|u(x,t)-u(y,t)|.$$ As depicted in Fig. \[fig:mt-cs-hydro\]. (c), the MT model shows faster decay rate than the CS model, which is already observed at the particle level, see Fig. \[fig:Initial-conditions-mt\]. (b).
Attractive-Repulsive models {#attractive-repulsive-models}
---------------------------
In this subsection, we consider the system with the following power-law potential in one dimension: $$K(x) = \frac{|x|^a}{a} - \frac{|x|^b}{b},$$ with the convention that $|x|^0/0 = \log |x|$. In one dimension, $K(x) = k|x|$ corresponds to the attractive ($k>0$) or repulsive ($k<0$) Poisson force. In Section \[sec\_EPA\], we deal with this case and study the critical thresholds analytically and numerically. In particular, there is a gap between the known subcritical and supercritical regions for the repulsive Poisson force (see Theorem \[thm\_cri2\]), we numerically analyse initial data in this region. In Section \[sec\_EPAP\], we study the case $a= 2$ and $b = 1$. In this case, it is well-known, as discussed in Section 2.2 and at the beginning of Section 3, that there is a flocking solution given by a density profile satisfying whose solution is the characteristic function of an interval. In particular, there is an steady state with $u=0$ with this density profile. We numerically show the convergence in time of density to this steady state whenever solutions are globally defined. We also study the blow-up phenomena in this case which is not present in first-order models of swarming such as the aggregation equation . Finally, the case $a = 2$ and $b = 0$, i.e., when the repulsive force is more repulsive than Newtonian is also numerically explored in Section \[sec\_EPAL\]. In this case, the flocking profile/stationary state for the density satisfying is given by the semi-circle law as shown in [@CFP] and the references therein.
### Euler-Poisson-alignment system {#sec_EPA}
In this part, we consider the system with Poisson forcing term in one dimension: $$\label{h_CS_P}
\begin{cases}
\partial_{t}\rho+{\partial}_x \left(\rho u\right)=0,\quad x\in{\mathbb{R}},\quad t>0,\\[2mm]
{\displaystyle {\partial}_{t}(\rho u)+ {\partial}_x(\rho u^2)=\int_{{\mathbb{R}}}\psi(x-y)(u(y)-u(x))\rho(x)\rho(y)\,dy - \rho {\left}({\partial}_x K \star \rho{\right}),}
\end{cases}$$ with $K\left(x\right)=k|x|$ attractive when $k>0$ and repulsive when $k<0$. Note that $k=\pm 0.5$ corresponds to the Newtonian interaction. For the system , the critical thresholds are studied in [@CCTT] depending on the sign of $k$. The result in [@CCTT] is as follows.
\[thm\_cri2\] Let $(\rho,u)$ be solutions to the Euler-Poisson-alignment model .
1. Attractive potential $\left(k>0\right)$: A unconditional finite-time blow up of the solution will appear no matter the initial conditions.
2. Repulsive potential $\left(k<0\right)$: As in the Euler-alignment system, there are two different zones:
- (Subcritical region) If $\partial_{x}u_{0}(x)<-\psi\star \rho_{0}(x)+\sigma_{+}(x)$ for all $x\in\mathbb{R}$, then the system has a global classical solution. Here, $\sigma_{+}(x)=0$ whenever $\rho_{0}(x)=0$ and elsewhere $\sigma_{+}(x)$ is the unique negative root of the equation $$\rho_{0}^{-1}(x)-\frac{1}{\psi_{M}^{2}}\left(2k+\frac{\psi_{M}\sigma_{+}(x)}{\rho_{0}(x)}-2ke^{\frac{\psi_{M}\sigma_{+}(x)}{2k \rho_{0}(x)}}\right)=0,\text{\ }\rho_{0}(x)>0.\label{eq:4.13}$$
- (Supercritical region) If there exists an x such that $$\partial_{x}u_{0}(x)<-\psi\star \rho_{0}(x)+\sigma_{-}(x),\quad\sigma_{-}:=-\sqrt{-4k \rho_{0}(x)}.$$ then the solution blows up in a finite time.
Note that there is a gap between the thresholds in the case of repulsive potential due to the non-locality of the velocity alignment force. If we choose the constant communication function $\psi \equiv 1$, we can close this gap, see Corollary \[cor\_cri\].
For the initial density and velocity for the numerical simulations, we take them as in Section \[subsec:euler-alig\]: for each node $i=1,\cdots,n$, $$\rho_{i}(0)=\frac{1}{\gamma}\cos\left(\pi\,\frac{x_{i}(0)}{1.5}\right) \quad \mbox{and} \quad u_{i}(0)=-c\text{\,}\sin\left(\pi\,\frac{x_{i}(0)}{1.5}\right).$$ Similarly as before, we change the values of $c$ to consider the subcritical and supercritical regions stated in Theorem \[thm\_cri2\].
Fig. \[fig:eu-po-ali-k1\] shows the evolution of density and velocity for the system with $c=0.4$ and $k=0.5$. As stated in Theorem \[thm\_cri2\], the density is blowing up at a finite time in this case. It occurs when $t=1.0788$.
For the repulsive potential case, we fix the value $k=-0.5$ and change the parameter $c = 0.95, 1.08, 1.2$ to consider both subcritical and supercritical regions described in Theorem \[thm\_cri2\], see Fig. \[fig:eu-po.ali-IC\]. To be more precise, when $c = 0.95$, $\partial_{x}u_{0}(x)+\psi\star\rho_{0}(x)$ lies in the subcritical region, for $c = 1.08$ it is between subcritical and supercritical regions. Thus it is not clear to have whether global regularity or finite-time blow-up of solutions for that case. When $c = 1.2$, it is inside the supercritical region and finite-time blowing up solution is expected from Theorem \[thm\_cri2\]. Since the initial density is independent of the parameter $c$, $\sigma_-$ and $\sigma_+$ are same for all three cases.
A Trust-region with the Dogleg method is used in to obtain the value of $\sigma_+$. Basically it consists in a Newton-Raphson’s method that applies a trust-region technique to improve robustness when starting far from the solution. Additionally, it introduces a procedure denoted as Powell dogleg. For more information about this method, see [@Pow]. In Matlab it could be easily implemented by the subroutine *fsolve*.
![\[fig:eu-po.ali-IC\]Value of $\partial_{x}u_{0}(x)+\psi\star\rho_{0}(x)$ depending on $c$.](check-ic)
The numerical simulations of the density and the velocity for the case $c=0.95$ at different times are shown in Fig. \[fig:eu-po-al-k-1-c095\]. In the beginning, the density tends to be concentrated near the origin, but after some time the repulsive force changes the sign of slope of the velocity. Consequently, the density spreads and the size of support of the density is increasing. Thus there is no finite-time blow-up of solutions in this case, which is consistent with Theorem \[thm\_cri2\].
When $c=1.08$, some values of $\partial_{x}u_{0}(x)+\psi\star\rho_{0}(x)$ are between $\sigma_+(x)$ and $\sigma_-(x)$. We do not know analytically what to expect in this case. We numerically observe finite-time blow up in Fig. \[fig:eu-po-al-k-1-c108\] after $t = 0.8107$. Our numerical exploration did not find initial data leading to global regularity when some values of $\partial_{x}u_{0}(x)+\psi\star\rho_{0}(x)$ are between $\sigma_+(x)$ and $\sigma_-(x)$. It is an open problem to decide if dichotomy of coexistence of finite time blow-up and global existence can happen in this region.
Finally, when $c=1.2$, the initial conditions are inside the supercritical region, and it is expected from Theorem \[thm\_cri2\] the blow-up of solutions in finite-time. The numerical simulations in this case show that indeed, see Fig. \[fig:eu-po-al-k-1-c12\]. The repulsive force is not enough to prevent the blow up phenomena of the system with these initial data, and the blow-up is produced after $t = 0.6204$.
As mentioned before, the gap between thresholds occurs due to the non-locality of the velocity alignment force. In fact, if we choose the constant communication function, i.e., $\psi \equiv 1$ or $\beta = 0$, then we have a sharp critical threshold for the repulsive potential case.
\[cor\_cri\] Let $(\rho,u)$ be solutions to the system with $\psi \equiv 1$ and $k < 0$.
- (Subcritical region) If $\partial_{x}u_{0}(x)>- \|\rho_{0}\|_{L^1} + \sigma(x)$ for all $x\in\mathbb{R}$, then the system has a global classical solution. Here, $\sigma(x)=0$ whenever $\rho_{0}(x)=0$ and elsewhere $\sigma(x)$ is the unique negative root of the equation $$\rho_{0}^{-1}(x)-2k-\frac{\sigma(x)}{\rho_{0}(x)}+2ke^{\frac{\sigma(x)}{2k \rho_{0}(x)}}=0,\text{\ }\rho_{0}(x)>0.$$
- (Supercritical region) If there exists an x such that $\partial_{x}u_{0}(x)<- \|\rho_{0}\|_{L^1}+\sigma(x)$, where the value of $\sigma(x)$ is the one given in the $\mathit{subcritical\,region}$, then the solution blows up in a finite time.
This particular case has also been checked to validate our code. In fact, our code captures the dichotomy in this case quite nicely leading to similar simulations as in the case of non constant communication function.
### Euler-Poisson-alignment system with power-law potential {#sec_EPAP}
In this part, we consider the repulsive Newtonian potential confined by a quadratic attractive potential: $$\label{pot_np}
K(x) = -\frac{|x|}{2} + \frac{x^2}{2}.$$ For simplicity, we also consider a linear damping in the momentum equation instead of the nonlocal velocity alignment. In this situation, our main system reads as $$\label{h_np}
\begin{cases}
\partial_{t}\rho+{\partial}_x \left(\rho u\right)=0,\quad x\in{\mathbb{R}},\quad t>0,\\[2mm]
{\displaystyle {\partial}_{t}(\rho u)+ {\partial}_x(\rho u^2)= -\rho u - \rho {\left}({\partial}_x K \star \rho{\right}),}
\end{cases}$$
If we take the constant communication function $\psi \equiv 1$ and assume the initial momentum is zero, i.e., $\int_{\mathbb{R}}\rho_0(x) u_0(x)\,dx = 0$, the system can be derived from the system with the potential $K$ given in .
Concerning the initial density for the numerical simulations, we again define it, for each node $i = 1,\cdots, n$, as $$\rho_{i}(0)=\frac{1}{\gamma}\cos\left(\pi\,\frac{x_{i}(0)}{1.7}\right),$$ where the constant $\gamma$ is fixed so that the total mass $M_0 := \int_{\mathbb{R}}\rho_0\,dx$ has the required value. For the initial velocity, we choose $$u_{i}(0)=-c\text{\,}x_{i}(0)\quad\mbox{for each node } i=1,\cdots, n,$$ where we again choose different constants $c > 0$ to deal with the subcritical and supercritical regions.
For the system with the potential $K$ given in , the sharp critical thresholds are classified in [@CCZ] according to the size of initial mass $M_0$. Time-asymptotic behaviors of density and velocity are also studied. The results in [@CCZ] for the case $M_0 < 1/4$ are as follows.
\[thm\_crinew\] Let $(\rho,u)$ be a classical solution to the system with the potential . Suppose that the initial density is compactly supported with the initial mass satisfying $M_0 <1/4$. Then the solution blows up in finite time if and only if there exists a $x^* \in \Omega_0 :=$ supp$(\rho_0)$ such that $$\partial_x u_0(x^*)<0,\quad M_0 - \rho_0(x^*)<\lambda_1\partial_x u_0 (x^*)$$ and $$\rho_0(x^*)\leq \left(\lambda_1 \partial_x u_0 (x^*) + \rho_0(x^*)-M_0\right)^{-\lambda_2/\sqrt{\Xi}} \left(\lambda_2 \partial_x u_0 (x^*) + \rho_0(x^*)-M_0\right)^{\lambda_1\sqrt{\Xi}}.$$ Here the constants $\lambda_i<0, i=1,2$ and $\Xi > 0$ are given as $$\lambda_1 \coloneqq \dfrac{-1+\sqrt{1-4M_0}}{2},\quad\lambda_2 \coloneqq \dfrac{-1-\sqrt{1-4M_0}}{2},\quad \textrm{and}\quad \Xi := 1-4M_0.$$ Furthermore, if there is no finite-time blow-up, we have $$\rho(x,t) \to M_0 \mathbf{1}_{[a,b]} \quad \mbox{in } L^1 \quad \mbox{and} \quad u(x,t) \to 0, \quad \mbox{in } L^\infty \quad \mbox{as} \quad t \to \infty,$$ exponentially fast, where $a,b$ are constants given by $$\begin{aligned}
b &= \frac{1}{M_0}{\left}(\int_{\mathbb{R}}x \rho_0(x)\,dx + \int_{\mathbb{R}}(\rho_0 u_0)(x)\,dx {\right}) + \frac12,\cr
a &= \frac{1}{M_0}{\left}(\int_{\mathbb{R}}x \rho_0(x)\,dx + \int_{\mathbb{R}}(\rho_0 u_0)(x)\,dx {\right}) - \frac12.
\end{aligned}$$
We refer to [@CCZ] for the sharp critical thresholds and the large-time behavior for global-in-time solutions for $M_0 \geq 1/4$.
In order to check numerically the critical thresholds stated in Theorem \[thm\_crinew\], two numerical simulations are carried out in Fig. \[fig:new\_critthres\]. First, the mass is set to be $M_0=0.2$, and then two cases $c = 0.9, 1.1$ are considered. Note that our initial conditions for $\rho_0$ and $u_0$ imply $$\int_{\mathbb{R}}x \rho_0(x)\,dx = \int_{\mathbb{R}}(\rho_0 u_0)(x)\,dx = 0.$$ When $c=0.9$, the initial data lie in the subs critical region, i.e., the initial data do not satisfy the conditions in Theorem \[thm\_crinew\], and subsequently, the density and the velocity converge to $M_0 \mathbf{1}_{[-1/2,1/2]}$ and $0$ as time goes on, respectively. On the other hand, for the case $c=1.1$, there is a finite time blow up caused by an infinite slope of the velocity on the boundary. No numerical simulation has been provided for the case of $M_0 \geq 1/4$ because the critical threshold established in [@CCZ] for that case involves a more involved requirement on the initial conditions. However, the dichotomy of behaviors obtained is similar.
By employing the argument proposed in [@CCTT], we provide an estimate on the blow-up time for the system with the potential . It is worth mentioning that this blow-up analysis does not depend on the size of mass. Differentiating the momentum equation of system with respect to $x$, we can rewrite it as $$\label{h_np_blowup}
\begin{cases}
\rho' = - \rho d\\[2mm]
d' = -d^2 - d + 2\rho - M_0
\end{cases}$$ where $\{ \}'$ denotes the time derivative along the characteristic flow $\eta$ defined in and $d := {\partial}_x u$.
\[thm\_blowupnew\] Let $(\rho,u)$ be a classical solution to the system with the potential on the time interval $[0,T]$. Suppose that there exists an x such that $$d_0(x) < 0 \quad \mbox{and} \quad \frac{1 + d_0(x)}{\rho_0(x)} +2\log\left(1 - \frac{d_0(x)}{2\rho_0(x)}\right) \leq 0.$$ Then the life span $T$ of the solution should be finite time and moreover $$T \leq 2\inf_{x \in \mathcal{S}}\log\left(1 - \frac{d_0(x)}{2\rho_0(x)}\right),$$ where $\mathcal{S}$ is defined as $$\mathcal{S}:= {\left}\{x \in {\mathbb{R}}\,:\, d_0(x) < 0 \quad \mbox{and} \quad \frac{1 + d_0(x)}{\rho_0(x)} +2\log\left(1 - \frac{d_0(x)}{2\rho_0(x)}\right) \leq 0{\right}\}.$$
Set $\beta = d/\rho$. Then it follows from that $\beta$ satisfies $$\beta' = \frac{1}{\rho^2}{\left}(d' \rho - d\rho' {\right}) = \frac{1}{\rho^2}{\left}( -d\rho + 2\rho^2 - \rho M_0{\right}) = -\beta + 2 - \frac{M_0}{\rho} \leq -\beta + 2.$$ This yields $$\beta \leq 2 + (\beta_0 - 2)e^{-t} \quad \mbox{for} \quad t \geq 0.$$ On the other hand, it follows from the continuity equation that $$\rho' = -\rho^2 \beta, \quad \mbox{i.e.,} \quad \rho^{-1} = \rho_0^{-1} + \int_0^t \beta(s)\,ds \leq \rho_0^{-1} + 2t + (\beta_0 - 2)(1 - e^{-t}).$$ Set $f(t) := \rho_0^{-1} + 2t + (\beta_0 - 2)(1 - e^{-t})$, then $$f_0 = \rho_0^{-1}>0 \quad \mbox{and} \quad \lim_{t \to +\infty}f(t) = \infty.$$ On the other hand, $f'(t) = 2 + (\beta_0 - 2)e^{-t}$, thus if there exists a $t_* > 0$ such that $f'(t_*) = 0$ and $f(t_*) \leq 0$, then the density $\rho$ is blowing up until this time $t_* > 0$. Note that $f'(t_*) = 0$ implies $e^{-t_*} = 2/(2-\beta_0)$. This yields $\beta_0 < 0$, i.e., $d_0 < 0$ due to $e^{-t_*} \in (0,1)$. Then for $d_0 < 0$ we get $$f(t_*) = \rho_0^{-1} + \beta_0 + 2t_* = \rho_0^{-1} +\beta_0 +2\log\left(\frac{2-\beta_0}{2}\right).$$ Hence if there exists a $x$ such that $$d_0(x) < 0 \quad \mbox{and} \quad \rho_0^{-1}(x) +\beta_0(x) +2\log\left(\frac{2-\beta_0(x)}{2}\right) \leq 0,$$ then the life-span $T$ of the solution $(\rho,u)$ should be finite. Furthermore, the time $T$ satisfies $$T \leq 2\inf_{x \in \mathcal{S}}\log {\left}(\frac{2-\beta_0(x)}{2}{\right}).$$
A study concerning the qualitative properties of the dynamics of the system with the potential are also conducted. Depending on the initial conditions, the density and position may converge to the steady state with or without oscillations around that state. In Fig. \[fig:new\_comparison\], it is depicted how the density and position of the boundary nodes evolves depending on the initial mass or the initial velocity. Generally, it shows in Fig. \[fig:new\_comparison\] (a) and (c) that there is a limit mass below in which there are no oscillations. However, this limit mass may change with the initial velocity. With respect to the influence of the initial velocity for a fixed mass according to Fig. \[fig:new\_comparison\] (b) and (d), it is possible to deduce that the more negative the initial slope of the velocity is, the larger the tendency to the oscilations is.
Finally, some numerical simulations with the CS nonlocal velocity alignment instead of linear damping are conducted. Although no theoretical result is known for that case, our numerical simulations demonstrate that the total mass of the system and the initial conditions affect the global behavior of the solution in similar way as in the linear damping case.
In Fig. \[fig:new\_cs\], two different simulations are depicted, and they differ in the initial velocity. The first one, Fig. \[fig:new\_cs\] (a), corresponds to $M_0=1$ and $c=0.2$, and we observe that there is no blow up in finite time reaching eventually the steady state. In contrast, the other simulation, increasing the value of the parameter c up to $0.5$, shows finite-time blow-up due to the infinite slope of the velocity on the boundary. The numerical time of blow up is $t=2.22$.
### Stronger repulsive potential than Newtonian {#sec_EPAL}
In this part, we deal with a potential more repulsive than Newtonian at the origin confined by the quadratic attractive potential given by $$\label{pot_nnp}
K(x) = -\log|x| + \frac{x^2}{2}.$$ For simplicity, we again consider the linear damping in the momentum equation. Steady solutions for this problem corresponds again to find density profiles $\rho$ satisfying together with $u=0$ on the support of the density $\rho$. They can be characterized as the global minimizer of certain energy functional and they can be computed explicitly, see [@CFP] and [@ST]. In fact, they are given by the semicircle law, that is, their graph is a semicircle.
We choose as initial density for the numerical simulations the positive part of a cosine function. For each node $i \in \{1,\cdots, N\}$, it is given by $$\rho_{i}(0)=\frac{1}{\gamma}\cos\left(\pi\,\frac{x_{i}(0)}{1.5}\right),$$ where the constant $\gamma$ is computed so that the total mass $M_0 := \int_{\mathbb{R}}\rho_0\,dx$ has the required value. With respect to the initial velocity, it is chose to be $$u_{i}(0)=-c\text{\,}x_{i}(0)\quad \mbox{for } \:i=1,...,n,$$ where the constant $c\in\mathbb{R}^{+}$ will be varied to study different initial conditions in the simulations.
We first find a blow up estimate for the system with the potential . By differentiating the momentum equation of system , it is obtained that $$\begin{aligned}
\label{gd_est}
\begin{aligned}
&\rho' = -\rho d,\cr
&d' = - \left(d^2 + d + \int_{\Omega(t)} \frac{\rho(y)}{|x-y|^2}\,dy+ M_0\right) \leq - (d^2 + d + M_0) ,
\end{aligned}\end{aligned}$$ and for the case of $1 - 4M_0 \geq 0 $ it is found that $$\begin{aligned}
\label{gd_est}
\begin{aligned}
&\rho' = -\rho d,\cr
&d' \leq -(d - d_-)(d - d_+),
\end{aligned}\end{aligned}$$ where $$d_\pm := \frac{-1 \pm \sqrt{1 - 4M_0}}{2}.$$
\[thm\_blowupnonnew\]Let $(\rho,u)$ be a classical solution to the system with the potential on the time interval $[0,T]$. Suppose that $1 - 4M_0 \geq 0$. Then the life span $T$ of the solution $(\rho,u)$ should be finite if there exists a x such that $\partial_x u_0(x)<\frac{-1 - \sqrt{1 - 4M_0}}{2}$. Moreover, $$T \leq \frac{1}{d_- - d_0}.$$
We divide the proof into two steps.
[**Step 1.-**]{} If $d_0 < d_-$, then $d(t) < d_-$ for $t \in [0,T]$. Set $$\mathcal{T}:= {\left}\{t \in [0,T] : d(s) < d_- \mbox{ for } s \in [0,t] {\right}\}.$$ Then $\mathcal{T}$ is not empty since $0 \in \mathcal{T}$. Furthermore if we set $\mathcal{T}^* = \sup \mathcal{T}$, then $\mathcal{T}^* > 0$ since $d(t)$ is continuous. Suppose that $\mathcal{T}^* < T$, then we get $\lim_{t \to \mathcal{T^*}-} d(t) = d_-$. On the other hand, it follows from that $$d' \leq -(d - d_-)(d - d_+) \quad \mbox{for} \quad t \in [0,\mathcal{T}^*),$$ and this yields $d'(t) < 0$ for $t \in [0,\mathcal{T}^*)$ since $d(t) < d_- < d_+$ for $t \in [0,\mathcal{T}^*)$. This is a contradiction to $$d_- = \lim_{t \to \mathcal{T^*}-} d(t) \leq d_0 < d_-.$$ Thus $\mathcal{T}^* \geq T$ and $d(t) < d_-$ for $t \in [0,T]$.
[**Step 2.-**]{} If $d_0 < d_-$, then the life-span of smooth solutions $T$ should be finite. Since $d(t) < d_- < d_+$ for $t \in [0,T]$, we get $$d' \leq -(d - d_-)(d - d_+) \leq -(d - d_-)^2, \quad \mbox{i.e.,} \quad (d - d_-)'\leq -(d - d_-)^2.$$ This implies $$d(t) \leq \frac{1}{(d_0 - d_-)^{-1} + t} + d_-.$$ Since $d_0 < d_-$, the right hand side of the above equality diverges to $- \infty$ when $t \rightarrow (d_- - d_0)^{-1}$. This concludes that the life-span $T$ should be less than $(d_- - d_0)^{-1}$.
In Fig. \[fig:nonnew\_blowup\], two numerical simulations are depicted with the objective of supporting Theorem \[thm\_blowupnonnew\]. The first case, with $M_0=0.2$ and $c=0.3$, the initial conditions of the blow up estimate of Theorem \[thm\_blowupnonnew\] are not satisfied. The numerical experiment seems to indicate that there exists a global-in-time solution converging toward the steady solution given by the semicircle law. We do not have sharp critical thresholds for this system so its behavior could not be predicted beforehand. On the other hand, in the second case it is set $M_0=0.2$ and $c=1$, and those initial conditions satisfy the blow up estimate of Theorem \[thm\_blowupnonnew\]. It can be observed in Fig. \[fig:nonnew\_blowup\] (f) that the finite-time blow-up is again produced by the infinite slope of the velocity at the boundary. Finding critical thresholds for this case is an open problem.
To conclude, a numerical study of the CS system with has been carried out. In fact, as above the total mass of the system and the initial conditions affect the global behavior of the solution leading to global existence of solutions converging to the semicircle law or finite time blow-up. The qualitative behavior is similar to the case of linear damping in case of initial data with zero mean velocity. Otherwise, the system can lead to travelling wave solutions with the same density profile.
### Euler-aligment system with attractive power-law potential and pressure
In this part, we consider the pressure term $p$ in the Euler-alignment system with the attractive power-law potential $K(x) = x^2/2$ leading to the system $$\label{press}
\begin{cases}
\partial_{t}\rho+{\partial}_x \left(\rho u\right)=0,\quad x\in{\mathbb{R}},\quad t>0,\\[2mm]
{\displaystyle {\partial}_{t}(\rho u)+ {\partial}_x(\rho u^2)+{\partial}_x p(\rho)= \rho{\left}[\psi*(\rho u)-(\psi*\rho)u - {\partial}_x K \star \rho{\right}],}
\end{cases}$$ where the pressure-law is given by $p(\rho) = \rho^2$.
For the numerical approximation, we will use a variation of the Lagragian scheme as in Section \[subsec:numerics\]. Notice that we can not directly apply that scheme due to the presence of pressure $p$. In order to overcome this new difficulty, we take into account the approximated pressure term ${\partial}_x p_\epsilon(\rho) := 2\rho({\partial}_x \delta_\epsilon \star \rho)$, where $\delta_\epsilon$ is a mollification of the Dirac delta function $\delta_0$ given by a Gaussian $$\delta_\epsilon(x) = \frac{1}{\sqrt{2\pi\epsilon}}e^{\frac{-x^2}{2\epsilon}}.$$ Note that $\delta_\epsilon$ converges weakly as mesures to $\delta_0$ as $\epsilon \to 0$, and subsequently, this implies ${\partial}_x p_\epsilon(\rho) = 2\rho(\delta_\epsilon \star {\partial}_x \rho) \to 2\rho{\partial}_x \rho = {\partial}_x p(\rho)$ as $\epsilon \to 0$, for smooth enough mass densities $\rho$. Using this approximation, we can rewrite the system $\eqref{press}_2$ as $${\partial}_{t}(\rho u)+ {\partial}_x(\rho u^2) = \int_{{\mathbb{R}}}\psi(x-y)(u(y)-u(x))\rho(x)\rho(y)\,dy - \rho {\left}({\partial}_x \tilde K \star \rho{\right}),$$ where the interaction potential $\tilde K$ is given by $$\tilde K(x) = -\frac{1}{\sqrt{2\pi \epsilon}}e^{\frac{-x^2}{2\epsilon}} + \frac{x^2}{2}.$$ This enables us to use the previous Lagrangian scheme for the numerical simulations. It is worth mentioning that the parameter $\epsilon$ can not be too small for the numerical simulation and it should be chosen according to both the number of nodes and the distance between them. In our setting, we take the parameter $\epsilon = 10^{-4.1}$ for the numerical simulations.
Similarly as before, we consider the CS nonlocal velocity alignment force and the linear damping for the numerical simulations. We again remind the reader that if we choose the constant communication function $\psi \equiv 1$, then CS velocity alignment force becomes the linear damping by assuming that the initial momentum is zero, i.e., $\int_{\mathbb{R}}(\rho_0 u_0)(x)\,dx = 0$.
Let us next investigate the steady solution for the system . Let us first look for steady solutions of the form $\rho = \rho_\infty(x)$ and $u = u_\infty \equiv 0$. Since the initial momentum is zero, it is straightforward to check that the center of mass of the density $\rho$ is preserved on time. Let us assume without loss of generality that the center of mass is zero. Plugging $\rho_\infty$ and $u_\infty$ into , we easily find $$2{\partial}_x \rho_\infty(x) = -(x \star \rho_\infty)(x) = -x M_0\quad \mbox{on} \quad \mbox{supp}(\rho_\infty).$$ This yields, by solving the ODE and fixing the mass to be $M_0$ with zero center of mass, that $$\rho_\infty (x) = \left\{ \begin{array}{ll}
\displaystyle -\frac{M_0}{4}\left(x+\sqrt[3]{3}\right)\left(x-\sqrt[3]{3}\right) \quad & \textrm{for} \quad x\in\left[-\sqrt[3]{3},\sqrt[3]{3}\right],\\[3mm]
0 & \textrm{otherwise}.
\end{array} \right.$$
The initial density and velocity for the numerical simulations are defined as $$\rho_{i}(0)=\frac{1}{\gamma}\cos\left(\pi\,\frac{x_{i}(0)}{1.5}\right) \quad \mbox{and} \quad u_{i}(0)=-c\text{\,}\sin\left(\pi\,\frac{x_{i}(0)}{1.5}\right),$$ for each node $i=1,\cdots,n$, where $\gamma$ is chosen so that the mass of the system is unit, and $c=0.2$. Actually, in our numerical experiments we add to the initial data the positive constant $\alpha=0.05$. The presence of vacuum areas lead to numerical artifacts at the boundary if we strictly impose zero value of the density. This is mainly due to the pressure term since the density is expected to become non-differentiable at the tip of the support, as for the steady state $\rho_\infty$. Therefore, we opt by adding this small constant to our initial data over the whole simulation interval. Furthermore, for the CS nonlocal velocity alignment force we have added a constant value to the initial velocity of $0.1$, in order to have a nonzero initial momentum.
Fig. \[fig:press\] shows the time evolutions of the density and the velocity for the approximated system with the linear damping (a), (b) and the CS nonlocal velocity alignment force (c), (d). We observe a convergence to the steady state for linear damping, while for the CS model a convergence towards a travelling wave profile due to the nonzero initial momentum. In both cases, we observe very fast convergences toward the steady state/travelling wave with time modulated decaying oscillations. Furthermore, it shows that the shape of the asymptotic density profiles are consistent with the theoretical results.
Acknowledgements {#acknowledgements .unnumbered}
================
J. A. C. was partially supported by the Royal Society via a Wolfson Research Merit Award. J. A. C. and Y.-P. C. were partially supported by EPSRC grant EP/K008404/1. Y.-P. C. was supported by the ERC-Stating grant HDSPCONTR “High-Dimensional Sparse Optimal Control”. S.P. was partially supported by a Erasmus+ scholarship.
[^1]:
[^2]:
[^3]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we develop a new parameterized algorithm for the [Independent Feedback Vertex Set]{} (IFVS) problem. Given a graph $G=(V,E)$, the goal of the problem is to determine whether there exists a vertex subset $F\subseteq V$ such that $V-F$ induces a forest in $G$ and $F$ is an independent set. We show that there exists a parameterized algorithm that can determine whether a graph contains an IFVS of size $k$ or not in time $O(4^kn^{2})$. To our best knowledge, this result improves the known upper bound for this problem, which is $O^{*}(5^{k}n^{O(1)})$.'
author:
- 'Yinglei Song[^1]'
title: An Improved Parameterized Algorithm for the Independent Feedback Vertex Set Problem
---
\[section\] \[section\] \[section\] \[section\] \[section\]
\#1
[**Keywords:**]{} independent feedback vertex set, parameterized algorithm, dynamic programming
Introduction
============
A [*feedback vertex set*]{} in a graph $G=(V,E)$ is a vertex subset $F \subseteq V$ such that vertices in $V-F$ induce a forest in $G$. Feedback vertex sets have important applications in deadlock recovery in the development of operating systems [@silberschatz] and database management systems [@molina]. Moreover, $F$ is an [*independent feedback vertex set*]{} if it induces an independent set in $G$.
In the past two decades, the [Feedback Vertex Set]{} (FVS) problem has been intensively studied. The goal of the problem is to determine whether a graph contains a FVS of size $k$ or not. In [@karp], it is shown that the problem is NP-complete. So far, a few exact algorithms have been developed to compute a minimum FVS in a graph. For example, in [@razgon], an algorithm that needs $O(1.9053^{n})$ time was developed to find a minimum FVS in a graph, this algorithm is also the first one that breaks the trivial $O(2^{n}n)$ bound. In [@fomin], an elegant algorithm that can enumerate all induced forests in a graph is used to find the minimum FVS. This algorithm needs $O(1.7548^{n})$ computation time. Recently, the upper bound of this problem is improved to $O(1.7347^n)$ [@fomin2].
Parameterized computation provides another potentially practical solution for problems that are computationally intractable. Specifically, one or a few parameters in some intractable problems can be identified and parameterized computation studies whether efficient algorithms exist for these problems while some or all of the parameters are small. A parameterized problem may contain a few parameters $k_1, k_2, \cdots, k_l$ and the problem is [*fixed parameter tractable*]{} if it can be solved in time $O(f(k_1, k_2, \cdots, k_l)n^{c})$,where $f$ is a function of $k_1, k_2, \cdots, k_l$, $n$ is the size of the problem and $c$ is a constant independent of all parameters. For example, the [Vertex Cover]{} problem is to determine whether a graph $G=(V,E)$ contains a vertex cover of size at most $k$ or not. The problem is NP-complete. However, a simple parameterized algorithm can solve the problem in time $O(2^{k}|V|)$ [@downey]. In practice, this algorithm can be used to efficiently solve the [Vertex Cover]{} problem when the parameter $k$ is fixed and small. On the other hand, some problems do not have known efficient parameterized solutions and are therefore parameterized intractable. Similar to the conventional complexity theory, a hierarchy of complexity classes has been constructed to describe the parameterized complexity of these problems [@downey]. For example, the [Independent Set]{} problem is to decide whether a graph contains an independent set of size $k$ or not and has been shown to be W\[1\]-complete [@downey2]. It cannot be solved with an efficient parameterized algorithm unless all problems in W\[1\] are fixed parameter tractable. A thorough investigation on these parameterized complexity classes are provided in [@downey].
The parameterized FVS problem uses the size of an FVS as the parameter $k$. In [@bodlaender; @dom; @downey; @downey2; @dehne; @guo; @guo2; @kanj; @raman; @raman2], it is shown that the problem is fixed parameter tractable and a few algorithms that can solve the problem in time $O(2^{O(k)}n^{O(1)})$ are developed. In [@chen], it is shown that the problem can be solved in time $O(5^{k}kn^{2})$. Recently, the upper bound of this problem is further improved to $O(3.83^{k}kn^{2})$ [@cao].
Although the FVS problem has attracted a lot of attention from researchers in computer science, little effort has been made toward the IFVS problem. Recently, it is shown that the FVS problem can be reduced to the IFVS problem in polynomial time while preserving the parameter $k$ [@misra]. The reduction is simply subdividing each graph edge once. The IFVS problem is thus more general than the FVS problem. In [@misra], it is shown that the IFVS problem can be solved in time $O^{*}(5^kn^{O(1)})$.
In this paper, we develop a new parameterized algorithm that can determine whether a graph contains an IFVS of size $k$ or not in time $O(4^{k}n^2)$. Our algorithm follows the [*iterative compression*]{} approach developed in [@reed]. This technique has been used to develop parameterized algorithms for many problems [@dehne; @dom; @guo2]. Specifically, given a graph $G$, the process starts with an empty graph and adds vertices one by one to reconstruct $G$. When a vertex is added, edges are also added to join the vertex with its neighbors that have been included in the graph. A series of graph $G_0,G_1,\cdots,G_n$ can thus be constructed during the process, where $G_0$ is an empty graph and $G_n=G$. Starting with $G_{2}$, the approach processes each $G_i$ ($2 \leq i \leq n$) to determine whether it contains an IFVS of size $k$ or not. If the answer is “yes”, the approach finds such a set $F_i$ in $G_i$ and continues to process $G_{i+1}$, otherwise the approach outputs “no” since $G$ does not contain an IFVS of size $k$ if one of its subgraph does not contain such an IFVS. The approach outputs “yes” if eventually $G$ is found to contain an IFVS of size $k$. A detailed description of the approach can be found in [@reed] and some other related work such as [@dehne; @dom; @guo2].
It is not difficult to see that, the key part of the above approach is to develop an algorithm that can determine whether $G_{i+1}$ contains an IFVS of size $k$ or not, given an FVS of size $k+1$. The FVS can be obtained by including all vertices in $F_i$ and the additional vertex that is added to $G_i$ to create $G_{i+1}$. The major contribution of this paper is the development of such an algorithm. Other than using a branching approach, we develop a dynamic programming based algorithm that can accomplish the task in time $O(4^{k}n)$. Since the graph contains $n$ vertices in total, the total amount of computation time needed by the approach is thus at most $O(4^{k}n^2)$. Based on the simple parameter preserving polynomial time reduction that reduces the FVS problem to IFVS problem, we also immediately obtain a parameterized algorithm that solves the FVS problem in time $O(4^{k}n^2)$.
Notations and Preliminaries
===========================
The graphs in this paper are undirected graphs without loops. For a given graph $G=(V,E)$ and a vertex subset $V' \subseteq V$, $G[V']$ is the subgraph induced on the vertices in $V'$. A vertex $u$ is a [*neighbor*]{} of $V'$ if it is joined to one of the vertices in $V'$ by an edge in $G'$. We use $N_{G}[V']$ to denote the set of neighbors of $V'$ in $G$. We may omit the subscript $G$ when the underlying graph is clear in the context. Given two graphs $G_1=(V_1, E_1)$ and $G_2=(V_2, E_2)$, we use $G_1 \cup G_2$ to denote the graph $(V_1 \cup V_2, E_1 \cup E_2)$. Furthermore, to simplify the notation, we use $G-V'$ to denote subgraph $G[V-V']$. A vertex subset $V'$ is a [*feedback vertex cover*]{} in $G$ if $G-V'$ is a forest. The objective of our algorithm is to find a feedback vertex cover that contains minimum number of vertices.
The Algorithm
=============
Our algorithm is based on the following Lemma.
\[lm1\] Given a graph $G=(V,E)$ and an FVS $F$ in $G$, a minimum IFVS in $G$ can be computed in time $O(4^{|F|}|V|)$.
Since $F$ is an FVS, $G[V-F]$ is a forest. As the first step of the algorithm, we arbitrarily choose a vertex in each tree in $G[V-F]$ as the root of the tree. We color each vertex in $V-F$ with black. We then check the number of children of each vertex in the tree and find those that have more than two children. For each such vertex $u$, we remove the edges that join it with its children $c_1,c_2,\cdots, c_d$, where $d$ is the number of children in the tree. We then create a path of white vertices $w_1,w_2,\cdots,w_{d-1}$. $v_{d-1}$ is joined to $u$ with an edge. To construct a binary tree, $w_1$ is joined to $c_1$ and $c_2$ with two edges and each white vertex $w_t$ ($2 \leq t \leq d-1$) in the path is joined to $c_{t+1}$ with an edge. Each white vertex $w_t$ is said to be [*equal*]{} to $u$. We use $H$ to denote the resulting forest. The number of white vertices in $H$ is at most $$\sum_{i=1}^{|V|-|F|}{d_i}\leq 2(|V|-|F|)$$ where $d_i$’s are the degrees of all vertices in $G[V-F]$. The number of vertices in $H$ is thus at most $3(|V|-|F|)$. Figure \[fig1\] illustrates the approach the algorithm uses to convert $T_i$ into a binary tree.
![The conversion the algorithm uses to convert a tree in $G[V-F]$ into a binary tree. (a)An internal node $u$ that has $d$ children $c_1,c_2,\cdots,c_d$; (b) A path of $d-1$ white vertices $w_1,w_2,\cdots,w_{d-1}$ are added to convert the tree into a binary tree.[]{data-label="fig1"}](ffigure1){width="10.0cm" height="5.8cm"}
In order to find a minimum IFVS $F'$, we consider the vertices in $F' \cap F$ first. We enumerate all subsets in $F$. For each such subset $S$, we check whether $S$ is an independent set and $G[F-S]$ is acyclic, if it is not the case, $S$ cannot be $F' \cap F$ and we can continue to the next subset. Otherwise, we need to find a minimum subset $S' \subseteq G-F$ such that $S' \cup S$ form an IFVS in $G$.
Without loss of generality, we assume $G[F-S]$ contains $l$ connected components $C_1, C_2, \cdots, C_l$ and let $C=\{C_1,C_2,\cdots,C_l\}$. Since $H$ is a forest, for each tree $T_i \in H$, we use a dynamic programming algorithm to find a minimum subset $S'_{i}$ in $T_i$ such that $(T_i -S'_{i}) \cup (F-F')$ is acyclic. We then obtain $S'$ by taking the union of $S'_i$’s for all $T_i$’s in $G-F$.
We use $r_i$ to denote the root of $T_i$. The algorithm starts with the leaves of the $T_i$ and follows a bottom-up order to scan through all vertices in $T_i$. For a tree node $u$ in $T_i$ and a set $S_u \subseteq C$, we consider the following two cases.
1. [$u$ is not in the IFVS, there exists a path that connects $u$ to each component in $S_u$ in the subtree rooted at $u$ in $T_i$. Moreover, none of the vertices in the path are in the IFVS,]{}
2. [$u$ is in the IFVS.]{}
The algorithm maintains dynamic programming table $M[u, S_u]$ for case 1 and table $P[u]$ for case 2. The two tables store the minimum number of vertices that can be included in the IFVS in the subtree rooted at $T_u$.
For a leaf node $v$ of $T_i$ and a subset $S_v$ of connected components in $G[V-F]$, the values of $M[v, S_v]$ are initialized to be $0$ if $S_v$ is the set of components that are directly connected to $v$ with an edge. Otherwise, $M[v,S_v]$ is set to be $|V|+1$. If $v \in N_{G}(S)$, $v$ is not in the IFVS, we thus set $P[v]$ to be $|V|+1$, otherwise $P[v]$ is set to be $1$.
For an internal node $u$ in $T_i$, since $T_i$ is a binary tree, we use $c_1,c_2$ to denote its children in $T_i$. In the case where $u \in N_{G}(S)$, $u$ cannot be in the IFVS and we thus set $P[u]$ to be $|V|+1$. Otherwise, both $M[u, P[u]]$ and $P[u]$ need to be determined with the recursion relations of the dynamic programming.
Next, we determine the recursion relations of the dynamic programming. For a black vertex $u$, we use $W_u$ to denote the set of connected components in $G[F-S]$ that are directly connected to $u$ with an edge. If $W_u$ contains a component that is not in $S_u$, $M[u, S_u]$ is set to be $|V|+1$, since $u$ is connected to any component in $W_u$ regardless of its child nodes in $T_i$. For a white vertex $u$, $W_u=W_b$, where $b$ is the black vertex that is equal to $u$ in $T_i$. The algorithm thus only needs to compute the values of $M[u, S_u]$ for these $S_u$’s that are supersets of $W_u$.
We now assume $W_u \subset S_u$ and $u$ has two children. Based on the color of $u$, we use the following recursion relations to compute $M[u, W_u]$ and $M[u, S_u]$, if $u$ is a black vertex, the recursion relation to compute $M[u,S_u]$ are $$M[u, W_u] =\min{\{M[c_1,\epsilon]+M[c_2,\epsilon],M[c_1,\epsilon]+P[c_2], P[c_1]+M[c_2,\epsilon],P[c_1]+P[c_2]\}}$$ $$M[u, S_u] = \min{\{\min_{S_{c_1},S_{c_2}}{\{M[c_1, S_{c_1}]+M[c_2,S_{c_2}]\}},
M[c_1,S_{u}-W_{u}]+P[c_2], P[c_1]+M[c_2,S_{u}-W_{u}]\}}$$ where $S_{c_1}$ and $S_{c_2}$ are disjoint subsets of $S_u$ and $S_u-W_u=S_{c_1}\cup S_{c_2}$. The algorithm exhaustively enumerates all such subsets and computes the minimum value of $M[c_1,S_{c_1}]+M[c_2,S_{c_2}]$ over all possible subsets $S_{c_1}$ and $S_{c_2}$.
If $u$ is a white vertex and both $c_1$ and $c_2$ are black vertices, the recursion relation to compute $M[u,S_u]$ is the same as above. If $u$ is white and one of $c_1,c_2$ is white, without loss of generality, we assume $c_1$ is white. Since $u$ and $c_1$ represents the same vertex in $G$, if $u$ is not in the IFVS, $c_1$ cannot be in the IFVS. The recursion relation is thus $$\begin{aligned}
M[u, W_u] &=& \min{\{M[c_1,W_u]+M[c_2,\epsilon],M[c_1,W_u]+P[c_2]\}} \\
M[u, S_u]&=& \min{\{\min_{S_{c_1},S_{c_2}}{\{M[c_1, S_{c_1}\cup W_u]+M[c_2,S_{c_2}]\}},M[c_1,S_{u}]+P[c_2]\}}\end{aligned}$$ where $S_{c_1}$ and $S_{c_2}$ are disjoint subsets of $S_u$ and $S_{c_1} \cup S_{c_2}=S_u-W_u$.
If $u$ is black and $u\not\in N_G(S)$, since none of $c_1$ and $c_2$ are included in the IFVS, the recursion relation to compute $P[u]$ is as follows. $$P[u]=\min_{S_u}{\{M[c_1, S_u]\}}+\min_{S_u}{\{M[c_2,S_u]\}}+1$$ where the minimum is taken over all possible subsets $S_u$ of $C$.
If $u$ is white, $u\not\in N_G(S)$, and both $c_1$ and $c_2$ are black, the recursion for computing $P[u]$ is the same as above. If $u$ is white and one of $c_1,c_2$ is white, without loss of generality, we assume $c_1$ is white. Since $u$ and $c_1$ represents the same vertex and $c_2$ is not included in the IFVS, the recursion relation is $$P[u]=P[c_1]+\min_{S_u}{\{M[c_2,S_u]\}}$$ where the minimum is taken over all possible subsets $S_u$ of $C$.
If $u$ has only one child, $u$ must be black and we use $c$ to denote its child. If $c$ is black, we use the following recursion relations to compute $M[u, W_u]$, $M[u,S_u]$, and $P[u]$. $$\begin{aligned}
M[u,W_u] &=& \min{\{M[c,\epsilon],P[c]\}} \\
M[u,S_u] &=& M[c, S_u-W_u] \\
P[u] &=& \min_{S_u}{\{M[c,S_u]\}}+1\end{aligned}$$ where the minimum is taken over all possible subsets $S_u$ of $C$.
If $c$ is white, $c$ is equal to $u$, we use the following recursion relations to compute $M[u, S_u]$ and $P[u]$. $$\begin{aligned}
M[u,S_u] &=& M[c,S_u] \\
P[u] & = & P[c]\end{aligned}$$
After the dynamic programming tables are filled by the algorithm, for each $T_i$($1 \leq k \leq l$), we can find the minimum of $P[r_i]$, $M[r_i, S_{r_i}]$ for all subsets $S_{r_i}$’s of $C$.We then follow a top-down trace back procedure to find $S'_i$. Note that the above recursions split $S_u$ into two disjoint subsets $S_{c_1}$ and $S_{c_2}$, which can guarantee that no cycles are formed if $u$ is not included in the IFVS. The correctness of this algorithm can thus be easily proved with the principle of induction and its recursion relations.
Before the dynamic programming starts, the algorithm needs to construct $H$ from the trees in $G[V-F]$, it needs at most $O(|V|)$ time. In addition, the algorithm also needs to compute $W_u$ for each internal node in the trees in $H$ and $N_G(S)$. However, since $S$ contains at most $|F|$ vertices, this needs at most $O(|F||V|)$ time.
During the dynamic programming process, for a given node $u$ and $S_u$, the algorithm needs to consider all possible disjoint pairs of subsets $S_{c_1}$ and $S_{c_2}$, such that $S_{c_1}\cup S_{c_2} =S_u-W_u$. We denote $t=|S_u|$, the number of such pairs is at most $2^{t}$. The total amount of computation time needed by the dynamic programming on one single vertex in $T_i$ is thus at most $$O(\sum_{t=0}^{l}{2^{t}\binom{l}{t}})=O(3^{l})$$ where the right hand side of the equation can be directly obtained with the binomial theorem.
Since $H$ contains at most $3(|V|-|F|)$ vertices, the dynamic programming algorithm can find $S'$ in time $O(3^{l}|V|)$.For each $S \subseteq F $ such that $G[F-S]$ is acyclic, we can find a feedback vertex cover $S \cup S'$. The one that contains the minimum number of vertices is a minimum IFVS. The total amount of the computation time needed by the algorithm is thus at most $$O(\sum_{l=1}^{|F|}{3^{l}\binom{|F|}{l}|V|})=O(4^{|F|}|V|)$$ The right hand side of the equation again is the direct result of applying the binomial theorem to the left hand side. The theorem has been proved.
Based on the algorithm we have developed in Lemma \[lm1\], we can use the iterative compression technique developed in [@reed] to develop a parameterized algorithm that can determine whether a graph contains an IFVS of size $k$ in time $O(4^kn^2)$. We thus have the following theorem.
\[the1\] Given a graph $G=(V,E)$ and a positive integer parameter $k$, there exists a parameterized algorithm that can determine whether $G$ contains an IFVS of size $k$ or not in time $O(4^k|V|^2)$.
We use the iterative compression technique to develop the algorithm. Let $n=|V|$, the algorithm can be sketched as follows.
1. [Starting with an empty graph $G_0$, add vertices one by one to reconstruct graph $G$, when a vertex $u_i$ is added to $G_i$ ($0 \leq i <n $), edges are added to $G_i$ to join $u_i$ with its neighbors in $G$ that have been included in $G_i$ to construct $G_{i+1}$. We thus can obtain a series of graphs $G_0,G_1,\cdots, G_{n}$, where $G_0$ is an empty graph and $G_{n}=G$;]{}
2. [Arbitrarily choose one vertex from $G_{2}$ and include it in set $F_2$;]{}
3. [For each $i=2$ to $n-1$, given $F_i \cup \{u_i\}$, use the algorithm presented in the proof of Lemma \[lm1\] to compute a minimum IFVS $F_{i+1}$ in $G_{i+1}$, if $F_{i+1}$ contains more than $k$ vertices, return “no”;]{}
4. [return “yes”.]{}
For computation time, step 1 can be accomplished in time $O(n^2)$. From Lemma \[lm1\], the computation time needed by step 3 is at most $$\sum_{i=2}^{n}{O(4^{k}i)}=O(4^{k}n^2)$$ The theorem has been proved.
In [@misra], it is shown that the FVS problem can be reduced to the IFVS problem through a simple polynomial time reduction that also preserves the parameter $k$. Theorem \[the1\] thus implies there exists a dynamic programming based algorithm that solves the FVS problem in time $O(4^kn^2)$. In fact, it is not difficult to see that the algorithm presented in the proof of Lemma \[lm1\] can be slightly changed to solve the FVS problem.
Conclusions
===========
In this paper, we develop a new parameterized algorithm that can determine whether a graph contains an IFVS of size $k$ or not in time $O(4^{k}n^2)$. To our best knowledge, this result improves the previously known upper bound for this problem, which is $O^*(5^kn^{O(1)})$.Our technique combines a dynamic programming approach with the recently developed iterative compression technique, which has been proved to be useful for developing parameterized algorithms for many NP-complete problems.
Our algorithm does not employ the method of branching, which is an often used approach for developing exact and parameterized algorithms. Indeed, the IFVS is a subset with certain global structural properties. In general, a branching algorithm employs the local structure of a graph and can hardly be used to further improve the upper bound of the problem. However, we are still looking for the possibility to combine our method and the branching approach to further improve the upper bound of this problem. In fact, recent work [@fomin2] has shown that this combination can lead to improved upper bound results for many NP-hard graph optimization problems.
[21]{}
Bodlaender H.,“On Disjoint Cycles”, [*International Journal on Foundations of Computer Science*]{}, 5(1):59-68, 1994. Y. Cao, J. Chen, and Y. Liu, “On Feedback Vertex Set New Measure and New Structures”,[*SWAT 2010*]{}, 93-104,2010. J. Chen, F. V. Fomin, Y. Liu, S. Lu,and Y. Villanger, “Improved algorithms for feedback vertex set problems”, [*Journal of Computer System Science*]{}, 74(7): 1188-1198, 2008. F. K. H. Dehne, M. R. Fellows, M. A. Langston, F. A. Rosamond, and K. Stevens, “An $O(2^{O(k)}n^3)$ Algorithm for The Undirected Feedback Vertex Cover Problem”, [*Proceedings of the 11th Annual International Conference on Computing and Combinatorics*]{}, pp. 859-869, 2005 (COCOON 2005). M. Dom, J. Guo, F. Huffner, R. Niedermeier, and A. Trub,“Fixed-parameter tractability results for feedback set problems in tournaments”, [*Algorithms and complexity*]{}, volume 3998 of Lecture Notes in Computer Science, pages 320–331. Springer, Berlin, 2006. Downey, R., Fellows, M., “Fixed parameter tractability and completeness, [*Complexity Theory: Current Research*]{}, pp. 191–225. Cambridge University Press, 1992. Downey, R., Fellows, M., [*Parameterized Complexity*]{}, Springer-Verlag, New York, 1999. F. V. Fomin, S. Gaspers, and A. V. Pyatkin, “Finding A Minimum Feedback Vertex Cover in Time $O(1.7548^n)$”, [*Proceedings of 2006 International Workshop on Parameterized and Exact Computation*]{}, 2006 (IWPEC 2006). F. V. Fomin, S. Gaspers, S. Saurabh, and A. A. Stepanov, “On Two Techniques of Combining Branching and Treewidth”, [*Proceedings of the 17th International Symposium on Algorithm and Computation*]{}, 2006 (ISAAC 2006). J. Guo, R. Niedermeier, and S. Wernicke, “Parameterized Complexity of Generalized Vertex Cover Problems”, [*Proceedings of the 9th Workshop on Algorithms and Data Structures*]{}, pp. 36-48, 2005 (WADS 2005). J. Guo, J. Gramm, F. H ?uffner, R. Niedermeier, and S. Wernicke, “Compression-based fixed-parameter algorithms for feedback vertex set an d edge bipartization”, [*Journal of Computer and System Science*]{}, 72(8):1386–1396, 2006. I. Kanj, M. Pelsmajer, and M. Schaefer, “Parameterized Algorithms for Feedback Vertex Set”, [*Proceedings of 2004 International Workshop on Parameterized and Exact Computation*]{}, pp. 234-247, 2004 (IWPEC 2004). R. M. Karp, “Reducibility Among Combinatorial Problems”, [*Complexity of Computer Computations*]{}, Plenum Press, New York, pp. 85-103, 1972. N. Misra, G. Philip, V. Raman, and, S. Saurabah, “On Parameterized Independent Feedback Vertex Set”, [*Proceedings of COCOON 2011*]{}, pp.98-109, 2011. H. G. Molina, J. D. Ullman, and J. Widom, [*Database Systems: The Complete Book (2nd Edition)*]{}, Pearson Prentice Hall, 2009. V. Raman, S. Saurabh, and C. R. Subramanian, “Faster Fixed Parameter Tractible Algorithms for Undirected Feedback Vertex Set”,[*ACM Transactions on Algorithms*]{}, 2(3): 403-415, 2006. Raman, V., Saurabh, S., Subramanian, C., “Faster fixed parameter tractable algorithms for undirected feedback vertex set”, [*ISAAC 2002, LNCS vol. 2518*]{}, pp. 241–248. Springer, 2002. I. Razgon, “Exact Computation of Maximum Induced Forest”, [*Proceedings of the 10th Scandinavian Workshop on Algorithm Theory*]{},pp. 241-248, 2002. B. Reed, K. Smith, and A. Vetta, “Finding odd cycle transversals”, [*Operations Research Letters*]{}, 32(4):299–301, 2004. A. Silberschatz, P. B. Galvin,and G. Gagne, [*Operating System Concepts, 9th Edition*]{}, John Wiley and Sons, Inc.,2012.
[^1]: School of Computer Science and Engineering, Jiangsu University of Science and Technology, Zhenjiang, Jiangsu 212003, China. Email: [[email protected].]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The past few years have witnessed the flourishing of crowdsourced medical question answering (Q&A) websites. Patients who have medical information demands tend to post questions about their health conditions on these crowdsourced Q&A websites and get answers from other users. However, we observe that a large portion of new medical questions cannot be answered in time or receive only few answers from these websites. On the other hand, we notice that solved questions have great potential to solve this challenge. Motivated by these, we propose an end-to-end system that can automatically find similar questions for unsolved medical questions. By learning the vector presentation of unsolved questions and their candidate similar questions, the proposed system outputs similar questions according to the similarity between vector representations. Through the vector representation, the similar questions are found at the question level, and the diversity of medical questions expression issue can be addressed. Further, we handle two more important issues, i.e., training data generation issue and efficiency issue, associated with the LSTM training procedure and the retrieval of candidate similar questions. The effectiveness of the proposed system is validated on a large-scale real-world dataset collected from a crowdsourced maternal-infant Q&A website.'
author:
- |
Yaliang Li$^1$, Liuyi Yao$^2$, Nan Du$^1$, Jing Gao$^2$,\
Qi Li$^3$, Chuishi Meng$^4$, Chenwei Zhang$^5$, and Wei Fan$^1$
title: |
Finding Similar Medical Questions from\
Question Answering Websites
---
[28]{}
\#1 \#1[\#1]{}\#1 \#1 \#1 \#1 \#1[\#1]{} \#1[\#1]{}
[bahdanau2014neural]{} . . ().
[berger2000bridging]{} . . In . .
[blei2003latent]{} . . (), .
[cai2011learning]{} . . In . .
[cao2010generalized]{} . . In . .
[cao2009use]{} . . In . .
[das2016mirror]{} . . In . .
[das2016together]{} . . In . .
[graves2013speech]{} . . In . .
[hochreiter1997long]{} . . , (), .
[jeon2005finding]{} . . In . .
[ji2012question]{} . . In . .
[kiros2015skip]{} . . In . .
[le2014distributed]{} . . In , Vol. . .
[li2010routing]{} . . In . .
[mikolov2013distributed]{} . . In . .
[palangi2016deep]{} . . , (), .
[ponte1998language]{} . . In . .
[qiu2015convolutional]{} . . In . .
[sutskever2014sequence]{} . . In . .
[wan2015deep]{} . . ().
[xue2008retrieval]{} . . In . .
[zhai2001study]{} . . In . .
[zhang2016mining]{} . . In . .
[zhang2016learning]{} . . In . .
[zhang2014question]{} . . In . .
[zhou2011phrase]{} . . In . .
[zhou2016learning]{} . . (), .
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Completely positive and completely bounded mutlipliers on rigid $C^{\ast}$-tensor categories were introduced by Popa and Vaes. Using these notions, we define and study the Fourier-Stieltjes algebra, the Fourier algebra and the algebra of completely bounded multipliers of a rigid $C^{\ast}$-tensor category. The rich structure that these algebras have in the setting of locally compact groups is still present in the setting of rigid $C^{\ast}$-tensor categories. We also prove that Leptin’s characterization of amenability still holds in this setting, and we collect some natural observations on property (T).'
address:
- 'Yuki Arano Department of Mathematics, Kyoto university Kitashirakawa-Oiwakecho, Sakyo-ku, Kyoto, 606-8502, Japan'
- 'Tim de Laat Mathematisches Institut, Westfälische Wilhelms-Universität Münster Einsteinstrasse 62, 48149, Germany'
- 'Jonas Wahl KU Leuven, Department of Mathematics Celestijnenlaan 200B – Box 2400, 3001 Leuven, Belgium'
author:
- Yuki Arano
- Tim de Laat
- Jonas Wahl
title: 'The Fourier algebra of a rigid $C^{\ast}$-tensor category'
---
Introduction
============
Let $G$ be a locally compact group. The Fourier algebra $A(G)$ of $G$ is the Banach algebra consisting of the matrix coefficients of the left-regular representation of $G$, and the Fourier-Stieltjes algebra $B(G)$ of $G$ is the Banach algebra consisting of the matrix coefficients of all unitary representations of the group. Both these algebras were introduced by Eymard [@eymard]. Nowadays, they play an important role in analytic group theory, in particular in the study of approximation and rigidity properties for groups (see e.g. [@brownozawa]). In this respect, also their relation to operator algebras is fundamental: the dual of $A(G)$ can be identified with the group von Neumann algebra $L(G)$ of $G$, and $B(G)$ can be identified with the dual of the universal group $C^{\ast}$-algebra $C^{\ast}(G)$ of $G$. In the study of approximation and rigidity properties for groups, also the Banach algebra of completely bounded Fourier multipliers plays a fundamental role. Analogues of the Fourier-Stieltjes algebra, the Fourier algebra and the algebra of completely bounded Fourier multipliers have been studied quite extensively in the setting of quantum groups, starting with the work of Daws [@daws].
A couple of years ago, fundamentally new results on approximation and rigidity properties for quantum groups were proven [@decommerfreslonyamashita], [@arano1]. Partly relying on these results, Popa and Vaes formulated the theory of unitary representations for “subfactor related group-like objects” (e.g. quantum groups, subfactors and $\lambda$-lattices), in the setting of rigid $C^{\ast}$-tensor categories [@popavaes] (see also [@neshveyevyamashita] and [@ghoshjones]). The representation category of a compact quantum group and the standard invariant of a subfactor are important and motivating examples of rigid $C^{\ast}$-tensor categories. Intimately related to the unitary representation theory are the notions of completely positive and completely bounded multipliers, which are particularly important in the study of approximation and rigidity properties. Popa and Vaes studied such properties for subfactor related group-like objects in [@popavaes]. Recently, more new results on approximation and rigidity properties for subfactor related group-like objects were proven [@cjones], [@aranovaes], [@aranodelaatwahl], [@tarragowahl].
The aim of this article is to study the structure of the collections of completely positive and completely bounded multipliers. In particular, we define and study analogues of the Fourier-Stieltjes algebra, the Fourier algebra and the algebra of completely bounded Fourier multipliers in the setting of rigid $C^{\ast}$-tensor categories. It turns out that in this setting they also form Banach algebras, and that the operator algebraic structure of these algebras is still present.
It turns out that in the setting of rigid $C^{\ast}$-tensor categories, we still have Leptin’s characterization of amenability (see Theorem \[thm:leptin\]). Moreover, we collect some observations on property (T) for quantum groups and rigid $C^{\ast}$-tensor categories (see Section \[sec:propertyt\]).
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank Stefaan Vaes for valuable discussions, suggestions and remarks, and for permitting them to include his proof of Theorem \[thm:fima\].
Preliminaries
=============
Rigid $C^{\ast}$-tensor categories {#subsec:cstc}
----------------------------------
A $C^{\ast}$-tensor category is a category that behaves similar to the category of Hilbert spaces. For the basic theory of $C^{\ast}$-tensor categories and the facts mentioned in this subsection, we refer to [@neshveyevtuset Chapter 2].
In what follows, all tensor categories will be assumed to be strict, unless explicitly mentioned otherwise. This is not a fundamental restriction, since every tensor category can be strictified.
Let $\mathcal{C}$ be a $C^{\ast}$-tensor category. An object $\bar{u}$ in $\mathcal{C}$ is conjugate to an object $u$ in $\mathcal{C}$ if there are $R \in \operatorname{Mor}({\mathbbm{1}}, \bar{u} {\otimes}u)$ and $\bar{R} \in \operatorname{Mor}({\mathbbm{1}}, u {\otimes}\bar{u})$ such that $$u \xrightarrow{1 {\otimes}R} u {\otimes}\bar{u} {\otimes}u \xrightarrow{\bar{R}^* {\otimes}1 } u \ \ \text{and} \ \ \bar{u} \xrightarrow{1 {\otimes}\bar{R}} \bar{u} {\otimes}u {\otimes}\bar{u} \xrightarrow{R^* {\otimes}1} \bar{u}$$ are the identity morphisms. Conjugate objects are uniquely determined up to isomorphism. If every object has a conjugate object, then the category $\mathcal{C}$ is called a rigid $C^{\ast}$-tensor category.
Let $\mathrm{Irr}(\mathcal{C})$ denote the set of equivalence classes of irreducible objects in $\mathcal{C}$. Using the same notation as above, if $u$ is an irreducible object with a conjugate, then $d(u) = \|R \| \| \bar{R} \|$ is independent of the choice of the morphisms $R$ and $\bar{R}$. An arbitrary object $u$ in a rigid $C^{\ast}$-tensor category is unitarily equivalent to a direct sum $u \cong \bigoplus_{k} u_k$ of irreducible objects, and we put $d(u) = \sum_{k} d(u_k)$. The function $d : \mathcal{C} \to [0, \infty)$ defined in this way is called the intrinsic dimension of $\mathcal{C}$.
Multipliers on rigid $C^{\ast}$-tensor categories
-------------------------------------------------
Multipliers on rigid $C^{\ast}$-tensor categories were introduced by Popa and Vaes [@popavaes].
\[mult\] A multiplier on a rigid $C^{\ast}$-tensor category ${\mathcal{C}}$ is a family of linear maps $$\theta_{\alpha,\beta} : \operatorname{End}(\alpha {\otimes}\beta) \to \operatorname{End}(\alpha {\otimes}\beta)$$ indexed by $\alpha, \beta \in {\mathcal{C}}$ such that $$\begin{aligned}
\theta_{\alpha_2,\beta_2}(UXV^{\ast}) &= U\theta_{\alpha_1,\beta_1}(X)V^{\ast}, \nonumber \\
\theta_{\alpha_1 {\otimes}\alpha_2,\beta_1 {\otimes}\beta_2} (1 {\otimes}X {\otimes}1) &= 1 {\otimes}\theta_{\alpha_2, \beta_1}(X) {\otimes}1 \label{eq:equation1}\end{aligned}$$ for all $\alpha_i, \beta_i \in {\mathcal{C}}, X \in \operatorname{End}(\alpha_2 {\otimes}\beta_1)$ and $U,V \in \mathrm{Mor}(\alpha_1,\alpha_2) {\otimes}\mathrm{Mor}(\beta_2,\beta_1)$.
A multiplier $(\theta_{\alpha,\beta})$ is said to be completely positive (or a cp-multiplier) if all maps $\theta_{\alpha,\beta}$ are completely positive. A multiplier $(\theta_{\alpha,\beta})$ is said to be completely bounded (or a cb-multiplier) if all maps $\theta_{\alpha,\beta}$ are completely bounded and $\| \theta \|_{\mathrm{cb}} = \sup_{\alpha, \beta \in {\mathcal{C}}} \| \theta_{\alpha,\beta}\|_{\mathrm{cb}} < \infty$. By [@popavaes Proposition 3.6], every multiplier $(\theta_{\alpha, \beta})$ is uniquely determined by a family of linear maps $\operatorname{Mor}(\alpha {\otimes}\bar{\alpha},{\mathbbm{1}}) \to \operatorname{Mor}(\alpha {\otimes}\bar{\alpha}, {\mathbbm{1}}), \ \alpha \in \operatorname{Irr}({\mathcal{C}})$. Since $\operatorname{Mor}(\alpha {\otimes}\bar{\alpha}, {\mathbbm{1}})$ is one-dimensional whenever $\alpha$ is irreducible, each of these linear maps is given by multiplication with a scalar $\varphi(\alpha) \in {\ensuremath{\mathbb{C}}}, \ \alpha \in \operatorname{Irr}({\mathcal{C}}),$ and hence every multiplier corresponds uniquely to a function $\varphi: \operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}$. Therefore, when we speak of a multiplier we will often mean the underlying function $\varphi: \operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}$.
The fusion algebra and admissible $\ast$-representations {#subsec:admissiblerepresentations}
--------------------------------------------------------
Recall that the fusion algebra ${\ensuremath{\mathbb{C}}}[{\mathcal{C}}]$ of a rigid $C^{\ast}$-tensor category ${\mathcal{C}}$ is defined as the free vector space with basis $\operatorname{Irr}({\mathcal{C}})$ and multiplication given by $$\alpha \beta = \sum_{\gamma \in \operatorname{Irr}({\mathcal{C}})} \operatorname{mult}(\alpha {\otimes}\beta, \gamma) \gamma, \ \ \ \alpha, \beta \in \operatorname{Irr}({\mathcal{C}}).$$ In fact, the fusion algebra is a $\ast$-algebra when equipped with the involution $\alpha^{\sharp}=\bar{\alpha}$.
In [@popavaes], Popa and Vaes defined the notion of admissible $\ast$-representation of $\mathbb{C}[\mathcal{C}]$ as a unital $\ast$-representation $\Theta: {\ensuremath{\mathbb{C}}}[{\mathcal{C}}] \to B({\mathcal{H}})$ such that for all $\xi \in {\mathcal{H}}$ the map $$\operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}, \; \alpha \to d(\alpha)^{-1} \langle \Theta(\alpha) \xi, \xi \rangle$$ is a cp-multiplier. Moreover, they proved the existence of a universal admissible $\ast$-representation and denoted the corresponding enveloping $C^{\ast}$-algebra of ${\ensuremath{\mathbb{C}}}[{\mathcal{C}}]$ by ${C_u(\mathcal C)}$.
The tube algebra
----------------
In [@ghoshjones], Ghosh and Jones related the representation theory of rigid $C^{\ast}$-tensor categories to Ocneanu’s tube algebra, which was introduced in [@ocneanu]. More precisely, Ghosh and Jones proved that a representation of ${\ensuremath{\mathbb{C}}}[{\mathcal{C}}]$ is admissible in the sense of Popa and Vaes if and only if it is unitarily equivalent to the restriction of a $*$-representation of the tube algebra to ${\ensuremath{\mathbb{C}}}[{\mathcal{C}}]$. While we do not elaborate on this in detail, the tube algebra picture is convenient when studying completely bounded multipliers (see in particular Proposition \[ArVaes\]).
Let us recall the definition of the tube algebra. Let ${\mathcal{C}}$ be a rigid $C^{\ast}$-tensor category. For each equivalence class $\alpha \in \mathrm{Irr}(\mathcal{C})$, choose a representative $X_{\alpha} \in \alpha$, and let $X_0$ denote the representative of the tensor unit. Moreover, let $\Lambda$ be a countable family of equivalence classes of objects in ${\mathcal{C}}$ with distinct representatives $Y_{\beta} \in \beta$ for every $\beta \in \Lambda$. The annular algebra with weight set $\Lambda$ is defined as $${\mathcal{A}}\Lambda = \bigoplus_{{\alpha}, {\beta}\in \Lambda, \ {\gamma}\in \operatorname{Irr}({\mathcal{C}})} \operatorname{Mor}(X_{{\gamma}} {\otimes}Y_{{\alpha}}, Y_{{\beta}} {\otimes}X_{{\gamma}}).$$ The algebra ${\mathcal{A}}\Lambda$ comes equipped with the structure of an associative $*$-algebra. We will always assume the weight set $\Lambda$ to be full, i.e. every irreducible object is equivalent to a subobject of some element in $\Lambda$. The annular algebra with weight set $\Lambda = \operatorname{Irr}({\mathcal{C}})$ is called the tube algebra of Ocneanu, and we write ${\mathcal{A}}\Lambda = {\mathcal{AC}}$.
Unitary half braidings {#sec:unitaryhalfb}
----------------------
Another approach to the representation theory of a rigid $C^*$-tensor category was developed in [@neshveyevyamashita] in terms of unitary half braidings on ind-objects. This approach is particularly well behaved when one is interested in taking tensor products of representations, a fact we will make use of in the proof of Theorem \[FourierStieltjesBanach\]. Let us recall that intuitively, an ind-object $X \in \operatorname{ind}{\mathcal{C}}$ is a possibly infinite direct sum of objects in the rigid $C^*$-tensor category ${\mathcal{C}}$ and that $\operatorname{ind}{\mathcal{C}}$ is a $C^*$-tensor category containing ${\mathcal{C}}$, albeit generically not a rigid one. For a rigorous definition and additional details, see [@neshveyevyamashita]. A unitary half braiding $\sigma$ on an ind-object $X \in \operatorname{ind}{\mathcal{C}}$ was defined in [@neshveyevyamashita] as a family of unitary morphisms $\sigma_{\alpha} \in \operatorname{Mor}( \alpha {\otimes}X, X {\otimes}\alpha ), \ \alpha \in {\mathcal{C}}$ satisfying
- $\sigma_{{\mathbbm{1}}} = {\mathord{\text{\rm id}}}$;
- $(1 {\otimes}V) \sigma_{\alpha} = \sigma_{\beta} (V {\otimes}1)$ for all $V \in \operatorname{Mor}(\alpha,\beta)$;
- $\sigma_{\alpha {\otimes}\beta} = (\sigma_{\alpha} {\otimes}1) (1 {\otimes}\sigma_{\beta})$.
Every pair $(X, \sigma)$ consisting of an ind-object $X$ and a unitary half braiding $\sigma$ on $X$, defines a $*$-representation of ${\ensuremath{\mathbb{C}}}[{\mathcal{C}}]$ on the Hilbert space ${\mathcal{H}}_{(X, \sigma)} = \operatorname{Mor}_{\operatorname{ind}{\mathcal{C}}}({\mathbbm{1}}, X)$ with inner product $\langle \xi, \eta \rangle 1 = \eta^* \xi$. More concretely, if we choose a set of representatives $Y_{\alpha}$ for ${\alpha}\in \operatorname{Irr}({\mathcal{C}})$ with standard solution of the conjugate equations $(R_{Y_{\alpha}}, \bar{R}_{Y_{\alpha}} )$, then $$\pi_{(X, \sigma)}: {\ensuremath{\mathbb{C}}}[{\mathcal{C}}] \to B({\mathcal{H}}_{(X, \sigma)}), \quad \pi(\alpha) \xi = (1 {\otimes}\bar{R}_{Y_{\alpha}}^*)(\sigma_{Y_{\alpha}} {\otimes}1)(1 {\otimes}\xi {\otimes}1) \bar{R}_{Y_{\alpha}}$$ defines a $*$-representation. Note that a different choice of representatives yields a unitarily equivalent $*$-representation. It was shown in [@neshveyevyamashita] that any admissible representation is unitarily equivalent to a representation of the above form. More generally, for an explicit bijection between unitary half braidings on ind-objects and (non-degenerate) $*$-representations of the tube algebra, see [@popashlyakhtenkovaes Proposition 3.14].
The Fourier-Stieltjes algebra {#sec:fourieralgebra}
=============================
Let $\mathcal{C}$ be a rigid $C^{\ast}$-tensor category, and let $\mathbb{C}[\mathcal{C}]$ denote its fusion algebra. The notion of admissible $\ast$-representation and the universal admissible $\ast$-representation, as introduced by Popa and Vaes in [@popavaes], were recalled in Section \[subsec:admissiblerepresentations\]. Admissible $\ast$-representations can be used to define the Fourier-Stieltjes algebra of a $C^{\ast}$-tensor category.
The Fourier-Stieljes algebra $B({\mathcal{C}})$ of a rigid $C^{\ast}$-tensor category ${\mathcal{C}}$ is the algebra of functions $\varphi: \operatorname{Irr}(\mathcal{C}) \to {\ensuremath{\mathbb{C}}}$ of the form $$\varphi(\alpha) = d(\alpha)^{-1} \langle \Theta(\alpha) \xi, \eta \rangle \ \ \ (\alpha \in \operatorname{Irr}(\mathcal{C})),$$ where $\Theta : \mathbb{C}[{\mathcal{C}}] \to B({\mathcal{K}})$ is an admissible $\ast$-representation of the fusion algebra and $\xi, \eta \in {\mathcal{K}}$. We call such a function $\varphi$ a (matrix) coefficient of $\Theta$. The algebra structure is given by pointwise multiplication.
To see that the Fourier-Stieltjes algebra is a unital algebra, note that it can also be defined as the span of the cp-multipliers $CP(\mathcal{C})$ on ${\mathcal{C}}$, i.e. $${\mathcal{B}}({\mathcal{C}}) = \left\{ \sum_{i=1}^n \lambda_i \varphi_i \;\Bigg\vert\; n \in {\ensuremath{\mathbb{N}}}, \ \lambda_i \in \mathbb{C}, \ \varphi_i \in CP({\mathcal{C}}), \ i=1,\dots,n \right\}.$$ Indeed, it follows from the definition of cp-multiplier that the product of two such multipliers is a cp-multiplier again.
We will now equip ${\mathcal{B}}({\mathcal{C}})$ with a norm that turns it into a Banach algebra.
\[FSa\] The map $\Phi_0 : CP({\mathcal{C}}) \to {C_u(\mathcal C)}^{\ast}_+, \ \Phi_0(\varphi)(\alpha) = \omega_{\varphi}(\alpha) = d(\alpha) \varphi(\alpha)$ extends linearly to an isomorphism of vector spaces $\Phi: {\mathcal{B}}({\mathcal{C}}) \to {C_u(\mathcal C)}^{\ast} $. Moreover, for an element $\varphi \in {\mathcal{B}}({\mathcal{C}})$, we have the following equality of norms: $$\| \varphi \|_{{\mathcal{B}}({\mathcal{C}})} := \| \Phi(\varphi) \| = \min \{\| \xi \| \| \eta \| \mid \ \varphi(\cdot) = d(\cdot)^{-1} \langle \Theta(\cdot) \xi, \eta \rangle, \ \Theta \text{ admissible} \}.$$
By definition of ${C_u(\mathcal C)}$ and [@popavaes Proposition 4.2], the map $\Phi_0$ is well-defined, and so is $\Phi$. It is clear that $\Phi$ defines a bijection. The second part follows directly from the following lemma.
Let $A$ be a unital $C^*$-algebra. For all $\omega \in A^*$, we have the following equality of norms: $$\| \omega \| = \min \{\| \xi \| \| \eta \| \mid \ \omega(\cdot) = \langle \Theta(\cdot) \xi, \eta \rangle, \ \Theta \ *-\text{representation of } A \}.$$
Although this is a well-known result, for the sake of completeness, we include a proof.
Since we can view $A^{\ast}$ as the predual of the von Neumann algebra $A^{**}$, we can consider the polar decomposition of $\omega$, i.e. there exists a positive normal functional $| \omega | \in A^{\ast}_+ $ and a partial isometry $V \in A^{**}$ such that $\omega = V |\omega|$ and $\| \omega \| = \| \ | \omega | \ \| $. Consider the GNS-representation $\Theta: A \to B({\mathcal{K}})$ of $|\omega|$, which has a cyclic vector, say $\eta$, i.e. $|\omega|(x) = \langle \Theta(x) \eta, \eta \rangle$ for all $x \in A$. As a consequence, we obtain that $$\omega(x) = (V |\omega|)(x) = |\omega|(xV) = \langle \Theta(x) \Theta'(V) \eta, \eta \rangle \ \ \forall x \in A,$$ where $\Theta'$ is the unique extension of $\Theta$ to $A^{**}$. Defining $\xi = \Theta'(V) \eta $, we have $\| \xi \| \leq \| \eta \|$, since $V$ is a partial isometry. Altogether, the set on the right side of the equality which is to be proven is nonempty and we have $\| \omega \| = \| \ | \omega | \ \| = \| \eta \|^2 \geq \| \xi \| \| \eta \|$.\
On the other hand, for every $*$-representation $\Theta:A \to B({\mathcal{H}})$ and $\xi, \eta \in {\mathcal{H}}$ such that $\omega(\cdot) = \langle \Theta(\cdot) \xi, \eta \rangle$ we have $$| \omega(x) | = | \langle \Theta(x) \xi, \eta \rangle | \leq \| x \| \| \xi \| \| \eta \| \quad \forall x \in A.$$
\[FourierStieltjesBanach\] Let ${\mathcal{C}}$ be a rigid $C^{\ast}$-tensor category. Then ${\mathcal{B}}({\mathcal{C}})$ is a Banach algebra with respect to the norm defined in the previous proposition.
The definition of $\| \cdot \|_{{\mathcal{B}}({\mathcal{C}})}$ directly implies that $({\mathcal{B}}({\mathcal{C}}), \| \cdot \|_{{\mathcal{B}}({\mathcal{C}})})$ is a Banach space and hence we are only left with showing that $\| \varphi_1 \varphi_2 \|_{{\mathcal{B}}({\mathcal{C}})} \leq \| \varphi_1 \|_{{\mathcal{B}}({\mathcal{C}})} \| \varphi_2 \|_{{\mathcal{B}}({\mathcal{C}})}$ for $\varphi_1, \varphi_2 \in {\mathcal{B}}({\mathcal{C}})$. Now, by Proposition \[FSa\] and the discussion in Section \[sec:unitaryhalfb\], for $i=1,2$ we can find pairs $(X_i, \sigma_i)$ of ind-objects $X_i \in \operatorname{ind}{\mathcal{C}}$ and unitary half braidings $\sigma_i$ on $X_i$ as well as $\xi_i, \eta_i \in {\mathcal{H}}_{(X_i,\sigma_i)}$ such that $$\varphi_i(\alpha) = d(\alpha)^{-1} \langle \pi_{(X_i,\sigma_i)}(\alpha) \xi_i, \eta_i \rangle \quad \text{and} \quad \| \varphi_i \|_{{\mathcal{B}}({\mathcal{C}})} = \| \xi_i \| \| \eta_i \|.$$ Following [@neshveyevyamashita], $\sigma = (1 {\otimes}\sigma_2)(\sigma_1 {\otimes}1)$ defines a unitary half braiding on $X = X_1 {\otimes}X_2 \in \operatorname{ind}{\mathcal{C}}$. Recall from [@neshveyevyamashita] that, in the same way as unitary half braidings are generalizations of group representations, this new half braiding is the proper analogue of the tensor product of the unitary half braidings $(X_1, \sigma_1)$ and $(X_2, \sigma_2)$. We have $\xi = (\xi_1 {\otimes}1) \xi_2, \eta = (\eta_1 {\otimes}1) \eta_2 \in \operatorname{Mor}_{\operatorname{ind}{\mathcal{C}}}({\mathbbm{1}}, X_1 {\otimes}X_2) = {\mathcal{H}}_{(X,\sigma)}$ with $\| \xi \| = \| \xi_1 \| \| \xi_2 \|, \| \eta \| = \| \eta_1 \| \| \eta_2 \|. $ Choosing representatives $Y_{\alpha}$ for irreducible objects ${\alpha}\in \operatorname{Irr}({\mathcal{C}})$ as in Section \[sec:unitaryhalfb\] and using the fact that $\bar{R}_{Y_{\alpha}}^* \bar{R}_{Y_{\alpha}} = d(\alpha) \in \operatorname{Mor}({\mathbbm{1}}, {\mathbbm{1}})$, we compute $$\pi_{(X,\sigma)}(\alpha) \xi = d(\alpha)^{-1} (\pi_{(X_1,\sigma_1)}(\alpha) \xi_1 {\otimes}1) (\pi_{(X_2,\sigma_2)}(\alpha) \xi_2),$$ and hence $$\varphi_1(\alpha) \varphi_2(\alpha) = d(\alpha)^{-1} \langle \pi_{(X,\sigma)}(\alpha) \xi, \eta \rangle,$$ which finishes the proof.
The Fourier algebra
===================
Recall that by [@popavaes Corollary 4.4], the left regular representation of $\mathbb{C}[{\mathcal{C}}]$ given by $$\lambda: \mathbb{C}[{\mathcal{C}}] \to B(\ell^2(\operatorname{Irr}({\mathcal{C}}))), \ \lambda(\alpha) \delta_{\beta} = \sum_{\gamma \in \operatorname{Irr}({\mathcal{C}})} \operatorname{mult}(\alpha {\otimes}\beta, \gamma) \delta_{\gamma}$$ is admissible and corresponds to the cp-multiplier defined by $\varphi_{\lambda}(\alpha) = \delta_{\alpha, {\mathbbm{1}}} \ (\alpha \in \operatorname{Irr}({\mathcal{C}}))$.
The Fourier algebra $A({\mathcal{C}})$ of a rigid $C^{\ast}$-tensor category ${\mathcal{C}}$ is defined as the predual of the von Neumann algebra $\lambda(\mathbb{C}[{\mathcal{C}}])''$.
Recall that there is a one-to-one correspondence between functions on $\operatorname{Irr}({\mathcal{C}})$ and functionals $\omega : \mathbb{C}[{\mathcal{C}}] \to \mathbb{C}$ given by $\varphi \mapsto \omega_{\varphi}$, where $\omega_{\varphi}({\alpha}) = d({\alpha}) \varphi({\alpha})$. By this correspondence, $A({\mathcal{C}})$ can also be interpreted as an algebra of functions on $\operatorname{Irr}({\mathcal{C}})$.
\[Fa\] For every $\omega \in A({\mathcal{C}})$, there exist $\xi, \eta \in \ell^2(\operatorname{Irr}({\mathcal{C}}))$ such that $\omega(x) = \langle \lambda(x) \xi, \eta \rangle$. In addition, $$\| \omega \|_{A({\mathcal{C}})} = \min \{\| \xi \| \| \eta \| \mid \omega(\cdot) = \langle \lambda(\cdot) \xi, \eta \rangle, \ \xi,\eta \in \ell^2(\operatorname{Irr}({\mathcal{C}})) \}.$$
Since $M = \lambda(\mathbb{C}[{\mathcal{C}}])''$ is nothing but the GNS-representation with respect to $\omega_{\varphi}$, where $\varphi_{\lambda}(\alpha) = \delta_{\alpha, {\mathbbm{1}}} \ (\alpha \in \operatorname{Irr}({\mathcal{C}}))$, we can represent every positive normal functional on $M$ as a vector state on $M$ by [@takesaki Chapter IX, Lemma 1.6]. The result for a general normal functional follows as in Proposition \[FSa\] by polar decomposition.
It is an immediate consequence of Proposition \[Fa\] that we have $$\| \varphi \|_{B({\mathcal{C}})} \leq \| \varphi \|_{A({\mathcal{C}})}$$ for $ \varphi \in A({\mathcal{C}})$, and it is not hard to see that the norms are actually equal. Indeed, the dual $C_r({\mathcal{C}})^*$ of the reduced $C^*$-algebra $C_r({\mathcal{C}})=\overline{\lambda(\mathbb{C}[\mathcal{C}])}$ identifies isometrically with the dual of a quotient of $C_u(\mathcal{C})$ and hence with the annihilator of a closed ideal in ${C_u(\mathcal C)}$. Consequently, $$\| \varphi \|_{A({\mathcal{C}})} = \| \varphi \|_{C_r({\mathcal{C}})^*} =\| \varphi \|_{{C_u(\mathcal C)}^*}$$ for $ \varphi \in A({\mathcal{C}})$. This means that we could also have defined $A({\mathcal{C}})$ as the closure of the coefficients of the left regular representation in ${\mathcal{B}}({\mathcal{C}})$. Moreover, we will see in Corollary \[CorCB\] that $A({\mathcal{C}})$ is a closed ideal in ${\mathcal{B}}({\mathcal{C}})$ and in particular a Banach algebra itself.
Completely bounded multipliers {#sec:cbm}
==============================
In this section, we study the algebra of completely bounded multipliers $$M_0A({\mathcal{C}}) = \{ \varphi : \operatorname{Irr}({\mathcal{C}}) \to \mathbb{C} \mid \varphi \ \text{cb-multiplier} \}.$$
While the Fourier algebra $A({\mathcal{C}})$ is only defined in terms of the fusion algebra $\mathbb{C}[{\mathcal{C}}]$, the Fourier-Stieltjes algebra $B(\mathcal{C})$ and the algebra $M_0A(\mathcal{C})$ of completely bounded multipliers use considerably more information on the category $\mathcal{C}$. Therefore, there is no apparent reason why completely bounded multipliers should correspond to completely bounded maps on the von Neumann algebra $\lambda(\mathbb{C}[{\mathcal{C}})])''$. However, in the tube algebra setting, the situation is more convenient. Indeed, whenever $\varphi: \operatorname{Irr}(\mathcal{C}) \to \mathbb{C}$ is a function on the irreducibles of $\mathcal{C}$ and $\Lambda$ is a full family of objects, then there is a canonical linear map $M_{\varphi}: {\mathcal{A}}\Lambda \to {\mathcal{A}}\Lambda$ given by $$M_{\varphi}(x) = \varphi(\gamma) x \ \ \ \text{whenever} \ \ \ x \in \operatorname{Mor}(X_{\gamma} {\otimes}Y_{{\alpha}}, Y_{{\beta}} {\otimes}X_{\gamma}).$$ Let us recall here that a multiplier $\varphi: \operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}$ is called completely bounded if $\| \varphi \|_{cb} = \sup_{{\alpha}, \beta \in {\mathcal{C}}} \| \theta^{\varphi}_{{\alpha}, \beta} \|_{cb} < \infty$, where $(\theta^{\varphi}_{{\alpha}, \beta})_{{\alpha}, \beta \in {\mathcal{C}}}$ denotes the family of linear maps associated to $\varphi$ as in Definition \[mult\] and the discussion thereafter. In terms of the maps $M_{\varphi}$ on the level of the tube algebra, the characterization of completely bounded multipliers is analogous to the group case. This leads to the following proposition which was proven by Vaes and the first-named author [@aranovaes Proposition 5.1].
\[ArVaes\] Let ${\mathcal{C}}$ be a rigid $C^{\ast}$-tensor category, let $\Lambda$ be a full family of objects of ${\mathcal{C}}$, and let $\varphi: \operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}$ be a function. Moreover, let $M_{\varphi}: {\mathcal{A}}\Lambda \to {\mathcal{A}}\Lambda$ be defined as above. Then $\|M_{\varphi}\|_{\mathrm{cb}} = \| \varphi \|_{\mathrm{cb}}$. If this cb-norm is finite, then $M_{\varphi}$ extends uniquely to a normal completely bounded map on $ {\mathcal{A}}\Lambda'' \subset B(L^2({\mathcal{A}}\Lambda))$.
\[CorCB\] Let $\varphi$ be a completely bounded multiplier. Then, the multiplication operator $$T_{\varphi} : A({\mathcal{C}}) \to A({\mathcal{C}}), \qquad \theta \mapsto \varphi \theta \quad (\theta \in A({\mathcal{C}}))$$ is well defined and completely bounded with $\| T_{\varphi} \|_{\mathrm{cb}} \leq \| \varphi \|_{\mathrm{cb}}$.
The dual map of the multiplication operator $T_{\varphi}$ is given by restricting the map $M_{\varphi}$ to $A({\mathcal{C}})^*$. By the previous proposition and standard results in operator space theory, the map $T_{\varphi}$ is completely bounded with $ \| T_{\varphi} \| = \| T^*_{\varphi} \| \leq \| \varphi \|_{\mathrm{cb}}$.
\[dual\] Let ${\mathcal{C}}$ be a rigid $C^{\ast}$-tensor category. Then $M_0A({\mathcal{C}})$ carries the structure of a dual Banach algebra if we endow it with pointwise addition and multiplication and the cb-norm $\| \cdot \|_{\mathrm{cb}}$.
Pick a full family of objects $\Lambda$, say $\Lambda = \operatorname{Irr}({\mathcal{C}})$, and denote the reduced $C^{\ast}$-algebra of ${\mathcal{A}}\Lambda$ by $A$ and its enveloping von Neumann algebra by $M = {\mathcal{A}}\Lambda''$. It follows from a well-known result in operator theory due to Effros and Ruan [@effrosruan] and, independently, due to Blecher and Paulsen [@blecherpaulsen], that the space of completely bounded maps $CB(A,M)$ is a dual operator space with predual $A \hat{{\otimes}} M_*$. Here, $\hat{{\otimes}}$ denotes the projective tensor product of operator spaces (see [@pisier Chapter 4] for details). Let us show that the image of the isometric embedding $ M_0A({\mathcal{C}}) \to CB(A,M), \ \varphi \mapsto \widetilde{M}_{\varphi}$ is w$^*$-closed in $CB(A,M)$, where $\widetilde{M}_{\varphi}$ denotes the unique extension of $M_{\varphi}$ to $A$. This will then imply that $ M_0A({\mathcal{C}})$ is isomorphic as a Banach space to the dual of a quotient of $A \hat{{\otimes}} M_*$ and in particular to a dual Banach algebra. So, let $(\varphi_{i})$ be a net in $M_0A({\mathcal{C}})$ such that $(\widetilde{M}_{\varphi_{i}})$ converges to a completely bounded map $\Psi \in CB(A,M)$. In particular, this means that $$\omega(\widetilde{M}_{\varphi_{i}}(x)) \to \omega(\Psi(x)) \quad \text{as} \quad i \to \infty$$ for all $x \in A, \ \omega \in M_*$. By choosing $x \in \operatorname{Mor}(X_{\gamma} {\otimes}{\mathbbm{1}}, {\mathbbm{1}}{\otimes}X_{\gamma})$ and $\omega \in M_*$ such that $\omega(x) \neq 0$ and by applying the definition of $\widetilde{M}_{\varphi_{i}}(x)$, we find that $\varphi_{i}$ converges pointwise to a bounded function $\varphi$. It follows from a short computation that the restriction of $\Psi$ to ${\mathcal{A}}\Lambda$ is equal to $M_{\varphi}$. As a consequence, $\varphi$ is completely bounded by the previous proposition with $\Psi = \widetilde{M}_{\varphi}$. Lastly, it is easy to see that pointwise multiplication of completely bounded maps in $M_0A({\mathcal{C}})$ is separately w$^*$-continuous, so $M_0A({\mathcal{C}})$ is a dual Banach algebra.
Leptin’s characterization of amenability
========================================
As defined by Popa and Vaes [@popavaes Definition 5.1], a rigid $C^{\ast}$-tensor category $\mathcal{C}$ is said to be amenable if there exists a net of finitely supported cp-multipliers $\varphi_i:\operatorname{Irr}(\mathcal{C}) \to \mathbb{C}$ that converges to $1$ pointwise.
In [@leptin], Leptin proved that a locally compact group is amenable if and only if the Fourier algebra of the group admits a bounded approximate unit. We finish this section by proving a version of Leptin’s theorem for rigid $C^*$-tensor categories. Before doing so, we note that, using the dimension function $d:\operatorname{Irr}({\mathcal{C}}) \to {\ensuremath{\mathbb{C}}}$, one can turn $\operatorname{Irr}({\mathcal{C}})$ into a discrete hypergroup (see [@muruganandam] for the definition of a hypergroup and its Fourier algebra). In the setting of discrete hypergroups, the existence of a bounded approximate unit on the Fourier algebra implies amenability, but the converse implication does not hold (see [@alaghmandan]).
We now state our version of Leptin’s theorem in the setting of rigid $C^{\ast}$-tensor categories.
\[thm:leptin\] A rigid $C^*$-tensor category $\mathcal{C}$ is amenable if and only if $A({\mathcal{C}})$ admits a bounded approximate unit, i.e. a net $(\varphi_{i})$ in $A({\mathcal{C}})$ such that $\sup_{i} \| \varphi_{i} \|_{A({\mathcal{C}})} < \infty$ and for all $f \in A({\mathcal{C}})$, $$\| \varphi_{i} f - f\|_{A({\mathcal{C}})} \to 0 \quad \text{as} \quad i \to \infty.$$
In order to prove this theorem, we first prove the following lemma.
\[lem.approx\] The space of finitely supported functions in the unit ball $A({\mathcal{C}})_1$ is norm dense in $A({\mathcal{C}})_1$, i.e. $A({\mathcal{C}})_1 = \overline{c_c(\operatorname{Irr}({\mathcal{C}})) \cap A({\mathcal{C}})_1}^{A({\mathcal{C}})}$.
Note first that if $\xi, \eta \in c_c(\operatorname{Irr}({\mathcal{C}}))$ are finitely supported functions, the same holds for the matrix coefficient $\varphi_{\xi,\eta}({\alpha}) = d({\alpha})^{-1} \langle \lambda({\alpha}) \xi, \eta \rangle, \ {\alpha}\in \operatorname{Irr}({\mathcal{C}})$. Since we can approximate any function in $\ell^2(\operatorname{Irr}({\mathcal{C}}))$ by finitely supported ones of smaller norm, every $\varphi \in A({\mathcal{C}})_1$ can be approximated in norm by functions of the form $\varphi_{\xi,\eta}$ with $\xi, \eta \in c_c(\operatorname{Irr}({\mathcal{C}}))$ and $\| \xi \|, \| \eta \| \leq 1$. More precisely, this follows from the inequality $$\| \varphi_{\xi_1,\eta_1} - \varphi_{\xi_2,\eta_2} \| \leq \| \xi_1 - \xi_2 \| \| \eta_1 \| + \| \eta_1 - \eta_2 \| \| \xi_2 \|$$ for all $\xi_i, \eta_i \in A({\mathcal{C}}), \ i=1,2$, which is easily established.
Assume first that ${\mathcal{C}}$ is amenable. By [@popavaes Proposition 5.3], this means that the trivial representation $\epsilon$ given by $\epsilon({\alpha}) = d({\alpha}), \ {\alpha}\in \operatorname{Irr}({\mathcal{C}})$ extends to a character on $C_r({\mathcal{C}})$ which we can extend to a (not necessarily normal) state on $C_r({\mathcal{C}})''$. Since the unit ball of every Banach space is w$^*$-dense in the unit ball of its double dual, there exists a net of normal states $(\omega_{i})$ on $C_r({\mathcal{C}})''$ such that for all $x \in C_r({\mathcal{C}})''$, $$\omega_{i}(x) \to \epsilon(x) \quad \text{as} \quad i \to \infty.$$ Let $\varphi_{i} \in A({\mathcal{C}})$ such that $\omega_{i} = \omega_{\varphi_{i}}$. By the previous lemma, it suffices to show that for all $f \in c_c(\operatorname{Irr}({\mathcal{C}})) \cap A(\mathcal{C})_1$, $$\| \omega_{\varphi_{i} f - f}\| \to 0 \quad \text{as} \quad i \to \infty.$$ Let $f \in c_c(\operatorname{Irr}({\mathcal{C}}))$. The operator given by $$T_{f}: {\ensuremath{\mathbb{C}}}[{\mathcal{C}}] \to {\ensuremath{\mathbb{C}}}[{\mathcal{C}}], \ T_{f}( {\alpha}) = f({\alpha}) {\alpha}\quad {\alpha}\in \operatorname{Irr}({\mathcal{C}})$$ extends to a (completely) bounded finite rank operator on $C_r({\mathcal{C}})''$ (see Proposition \[ArVaes\]) with norm $\| T_f \| = K$ for some $K > 0$. We have $$\| \omega_{\varphi_{i} f - f}\| = \sup_{\| x \| \leq 1} | \omega_{\varphi_{i} -1} (T_f(x)) | \leq \sup_{ y \in {\operatorname{Ran}}T_f, \ \|y \| \leq K } | \omega_{\varphi_{i} -1} (y)|.$$ But since ${\operatorname{Ran}}T_f$ is finite-dimensional and $\omega_{\varphi_{i} -1} \to 0$ as $i \to \infty$ in the w$^*$-topology, the result follows.\
Let us now assume that $A({\mathcal{C}})$ admits a bounded approximate unit $(\varphi_{i})$ with $\| \varphi_{i} \|_{A({\mathcal{C}})} \leq 1$ for all $i$. Let ${\alpha}\in \operatorname{Irr}({\mathcal{C}})$. By putting $f = \delta_{{\alpha}} \in A({\mathcal{C}})$, the characteristic function of ${\alpha}$, and $x = d({\alpha})^{-1} \lambda({\alpha})$, we obtain $$|\varphi_i({\alpha}) -1| \| \lambda({\alpha}) \| = |\omega_{\varphi_{i}f -f} (x) | \to 0 \quad \text{as } i \to \infty$$ and hence $(\varphi_{i})$ converges to $1$ pointwise. Now, using Lemma \[lem.approx\], we can approximate every $\varphi_{i}$ by a net $(\phi^j_{i})$ in $A({\mathcal{C}})_1 \cap c_c(\operatorname{Irr}({\mathcal{C}}))$ and since $\varphi_{i}$ is a positive element of $A({\mathcal{C}})$, the function $\phi^j_{i}$ can also be chosen to be positive for all $i$ and $j$. The net $(\phi_{i}^j)_{(i,j)}$ in $A({\mathcal{C}})_1 \cap c_c(\operatorname{Irr}({\mathcal{C}}))$ converges to $1$ pointwise, which proves the amenability of the category by [@popavaes Proposition 5.3].
Remarks on property (T) {#sec:propertyt}
=======================
The material of Section \[sec:fourieralgebra\] gives rise to some observations on property (T) in the setting of $C^{\ast}$-tensor categories that are motivated by Kazhdan’s property (T) in the setting of groups. Kazhdan’s property (T) is a rigidity property for locally compact groups that has numerous consequences and applications in mathematics. It was introduced in [@kazhdan], in which it was also shown that countable discrete groups with property (T) are finitely generated. Property (T) can be generalized to other settings, such as quantum groups and rigid $C^{\ast}$-tensor categories, and usually the natural analogue of finite generation is still an important consequence of property (T). In particular, Popa and Vaes showed that this is indeed the case in the setting of rigid $C^{\ast}$-tensor categories [@popavaes Proposition 5.4].
The definition of property (T) in the context of rigid $C^{\ast}$-tensor categories by Popa and Vaes and two characterizations of this property obtained in [@popavaes], are given in the following definition.
A rigid $C^{\ast}$-tensor category ${\mathcal{C}}$ has property (T) if one (and hence all) of the following equivalent conditions is satisfied.
(i) Every net $(\varphi_{\lambda})$ of cp-multipliers $\varphi_{\lambda} : \operatorname{Irr}{\mathcal{C}}\to \mathbb{C}$ converging to $\varphi_{{\varepsilon}}$ pointwise converges uniformly, i.e. $\sup_{x \in \operatorname{Irr}{\mathcal{C}}} |\varphi_{\lambda}(x) -1| \to 0$.
(ii) If $(\omega_{\lambda})$ is a net of states on ${C_u(\mathcal C)}$ converging to ${\varepsilon}$ in the weak\*-topology, it must already converge in norm.
(iii) There exists a unique nonzero projection $p \in {C_u(\mathcal C)}$ such that $\alpha p = d(\alpha) p$ for all $\alpha \in {\mathcal{C}}$. Such a projection is the analogue of a Kazhdan projection in the setting of groups.
Let ${\mathcal{C}}$ be a rigid $C^{\ast}$-tensor category, and let $W({\mathcal{C}}) = B({\mathcal{C}})^* = {C_u(\mathcal C)}^{**}$ be the enveloping von Neumann algebra of the full $C^{\ast}$-algebra of $\mathcal{C}$. Since the multiplier $\varphi_{{\varepsilon}}: \operatorname{Irr}({\mathcal{C}}) \to \mathbb{C}$ given by $\varphi_{{\varepsilon}}({\alpha}) = 1$ is completely positive by [@popavaes Corollary 4.4], the counit ${\varepsilon}: {\ensuremath{\mathbb{C}}}[{\mathcal{C}}] \to {\ensuremath{\mathbb{C}}}$ extends to a normal $*$-homomorphism on $W({\mathcal{C}})$.
It is known that for every locally compact group $G$, the Fourier–Stieltjes algebra $B(G)$ has a unique invariant mean. This goes back to [@godement Chapitre III]. This result was generalized to the setting of locally compact quantum groups in [@dawsskalskiviselter]. The next proposition asserts the existence of an invariant mean on the Fourier-Stieltjes algebra of ${\mathcal{C}}$, but we formulate it in terms of the existence of a central projection on $W(\mathcal{C})$.
Let $A$ be a unital $C^*$-algebra and let $\chi: A \to {\ensuremath{\mathbb{C}}}$ be a character on $A$. There exists a unique projection $p$ in the von Neumann algebra $A^{**}$ such that $$x p = p x = \chi(x) p \quad \text{ for all $x \in A^{**}$.}$$ In particular, setting $A = C_u({\mathcal{C}}), \ \chi= {\varepsilon}$, we find a unique projection $p \in B({\mathcal{C}})^* = W({\mathcal{C}})$ such that ${\varepsilon}(p) = 1$ and $\langle \omega, {\alpha}p \rangle = \langle \omega, p {\alpha}\rangle = d({\alpha}) \langle \omega, p \rangle$ for all ${\alpha}\in \operatorname{Irr}({\mathcal{C}})$ and $\omega \in B({\mathcal{C}})$.
Uniqueness of $p$ is immediate. To prove the existence, note that, since $\chi$ is a normal $\ast$-homomorphism, its kernel ${\operatorname{Ker}}(\chi)$ is weakly closed and therefore a von Neumann algebra itself. Denote its unit by $e_{\chi}$. Then the central cover $p = 1 - e_{\chi}$ of $\chi$ is a projection in $A^{**}$ satisfying $qp = pq = p$ for all $q$ with $\chi(q) = 1$. On the other hand, if $q$ is a projection in ${\operatorname{Ker}}(\chi)$, we have $pq =0$. Since every von Neumann algebra is the norm closure of the span of its projections, and $\chi$ is in particular norm continuous, the result follows.
In the group case, it was shown in [@akemannwalter Lemma 1] (see also [@valette Lemma 3.1] and [@haagerupknudbydelaat Proposition 4.1]) that a locally compact group $G$ has Kazhdan’s property (T) if and only if the unique invariant mean on $B(G)$ is weak$^{\ast}$-continuous, i.e. the mean is an element of $C^{\ast}(G)$ rather than just $C^{\ast}(G)^{\ast\ast}$. In fact, under the natural map from $B(G)^{\ast}$ to $C^{\ast}(G)^{\ast\ast}$, the mean is mapped to the Kazhdan projection, which is by weak$^{\ast}$-continuity actually an element of $C^{\ast}(G)$.
By characterization (iii) above, we see that the same thing happens for $C^{\ast}$-tensor categories: a rigid $C^{\ast}$-tensor category $\mathcal{C}$ has property (T) if and only if the mean on $B(\mathcal{C})$ is weak$^{\ast}$-continuous.
In the group case, the unique invariant mean on $B(G)$ is the restriction to $B(G)$ of the unique invariant mean on the space $\mathrm{WAP}(G)$ of weakly almost periodic functions on $G$, which is well-known to have a unique invariant mean. Indeed, note that $B(G) \subset \mathrm{WAP}(G)$. Hence, the only thing one needs to show is that this restriction is the unique invariant mean on $B(G)$. In a similar fashion, it is shown (see [@haagerupknudbydelaat Theorem A]) that the space $M_0A(G)$ of completely bounded Fourier multipliers on $G$ admits a unique invariant mean, using that $B(G) \subset M_0A(G) \subset \mathrm{WAP}(G)$. It is not known whether the space $M_0A(\mathcal{C})$ of a rigid $C^{\ast}$-tensor category admits a unique invariant mean, in particular because it is not known what the natural analogue of $\mathrm{WAP}(G)$ for rigid $C^{\ast}$-tensor categories should be. For locally compact quantum groups, WAP algebras were studied more thoroughly in [@dasdaws]. However, to the authors’ knowledge, the existence of an invariant mean on the WAP algebra of a locally compact quantum group ${\mathbb{G}}$ is only known in the case where ${\mathbb{G}}$ is amenable [@runde].
The unique invariant mean on $M_0A(G)$ leads in [@haagerupknudbydelaat] to the notion of property (T$^{\ast}$), defined in terms of the mean on $M_0A(G)$ being weak$^{\ast}$-continuous, which obstructs the Approximation Property of Haagerup and Kraus (see [@haagerupkraus]). The first examples of groups without the latter property were provided only recently (see [@lafforguedelasalle], [@haagerupdelaat1], [@haagerupdelaat2], [@haagerupknudbydelaat] and [@liao]). It is still an open problem to find an example of a quantum group without the analogue of the Approximation Property.
We will now compare property (T) for rigid $C^{\ast}$-tensor categories with other versions of property (T). In the case of discrete quantum groups, we have the following definition of property (T) (see [@kyed]), which is equivalent to the one introduced by Fima in [@fima].
\[yuki1\] Let ${\mathbb{G}}$ be a compact quantum group. The discrete dual ${\hat{\mathbb G}}$ has property (T) if one (and hence all) of the following equivalent conditions is satisfied.
(i) If a net of states $(\omega_\lambda)$ in ${C_u(\mathbb G)}^*$ converges to ${\varepsilon}$ pointwise, then it converges in norm.
(ii) There exists a projection $p \in {C_u(\mathbb G)}$ such that $x p = \varepsilon(x) p$ for all $x \in {C_u(\mathbb G)}$.
It was shown in [@dawsskalskiviselter] that, also in the more general framework of locally compact quantum groups, condition (i) of the previous definition is equivalent to the conventional notion of property (T) in terms of (almost) invariant vectors. The first part of the following theorem is [@popavaes Proposition 6.3], and the second part was proven in [@arano1].
\[propTtheorem\] Let ${\mathbb{G}}$ be a compact quantum group. The following conditions are equivalent:
(i) the category $\operatorname{Rep}{\mathbb{G}}$ has property (T) for rigid $C^{\ast}$-tensor categories,
(ii) the discrete dual ${\hat{\mathbb G}}$ has central property (T), i.e. if a net $(\omega_\lambda)$ of central states on $C_u(\mathbb{G})^{\ast}$ converges in the weak$^{\ast}$-topology, then it converges in norm.
Moreover, if we assume the Haar state on ${\mathbb{G}}$ to be tracial, this is equivalent to the discrete dual ${\hat{\mathbb G}}$ having (non-central) property (T).
We will now complete the picture by involving property (T) for von Neumann algebras. We use the following two characterizations of this property (see [@brownozawa Chapter 12] for the equivalence).
A finite von Neumann algebra $(M,\tau)$ has property (T) if one (and hence all) of the following equivalent conditions is satisfied.
(i) If $(\Phi_{\lambda}: M \to M)$ is a net of unital completely positive $\tau$-preserving maps converging to the identity pointwise on $L^2(M)$, i.e. $ \| \Phi_{\lambda}(x) - x \|_2 \to 0, \ \lambda \to \infty $ for all $x \in M$, then it already converges in norm, i.e. $$\sup_{x \in M_1} \| \Phi_{\lambda}(x) - x \|_2 \to 0 \quad \textrm{as} \quad \lambda \to \infty.$$
(ii) For any $M$-bimodule $\mathcal H$ and any net $(\xi_{\lambda})$ of unit vectors satisfying $$\langle x \xi_{\lambda} y, \xi_{\lambda} \rangle_{\mathcal H} \to \tau(xy) \quad \text{as} \quad \lambda \to \infty$$ for all $x,y \in M$ and $\tau(x) = \langle x \xi_{\lambda}, \xi_{\lambda} \rangle = \langle \xi_{\lambda} x, \xi_{\lambda} \rangle$ for all $\lambda$, there exists a net of $M$-central vectors $(\mu_{\lambda})$ with $$\| \xi_{\lambda} - \mu_{\lambda} \| \to 0 \quad \textrm{as} \quad \lambda \to \infty.$$
The following theorem is a generalization of [@fima Theorem 3.1]. However, as Stefaan Vaes pointed out to us, the proof in [@fima] contains a mistake. Indeed, at a critical point in the proof of [@fima Theorem 3.1], it is stated that for two irreducible objects $x,y \in \operatorname{Irr}({\mathbb{G}})$, one has $x \subset x {\otimes}y$ if and only if $y = {\mathbbm{1}}$. This is false, whenever $\hat{{\mathbb{G}}}$ is not a group. We thank Stefaan Vaes for providing us with a new proof, which we include here with his kind permission.
\[thm:fima\] Let ${\mathbb{G}}$ be a compact quantum group with a tracial Haar state. Then ${\hat{\mathbb G}}$ has (central) property (T) if and only if $L^{\infty}(\mathbb{G})$ has property (T).
Suppose that ${\hat{\mathbb G}}$ has property (T), and let $\mathcal H$ be a ${L^{\infty}(\mathbb G)}$-bimodule and $(\xi_{\lambda})$ a net of unit vectors in $\mathcal H$ such that $\langle x \xi_{\lambda} y , \xi_{\lambda} \rangle \to h(xy) \ \forall x,y \in {L^{\infty}(\mathbb G)}$ and $ h(x) = \langle x \xi_{\lambda}, \xi_{\lambda} \rangle = \langle \xi_{\lambda} x, \xi_{\lambda} \rangle$ for all $\lambda$ and $x\in M$. We have to find a net $(\mu_{\lambda})$ of ${L^{\infty}(\mathbb G)}$-central vectors such that $ \| \xi_{\lambda} - \mu_{\lambda} \| \to 0$. For every $\pi \in \operatorname{Irr}({\mathbb{G}})$, choose a unitary matrix $u^{\pi} = (u_{ij}^{\pi})$ representing $\pi$. Since the Haar state is tracial, we can assume that $u^{\bar{\pi}} = \overline{u^{\pi}}$. Define the linear map $$\Theta: {\mathcal{O}}({\mathbb{G}}) \to B({\mathcal{H}}) \ ; \ \Theta(u_{ij}^{\pi}) \xi = \sum_{k=1}^{d(\pi)} u_{ik}^{\pi} \xi ( u_{jk}^{\pi})^{*}, \ \ \ (\pi \in \operatorname{Irr}({\mathbb{G}}))$$ and, denoting the the coinverse of ${\mathbb{G}}$ by $S$, observe that $\Theta = \vartheta \circ \Delta$ where $\vartheta: {\mathcal{O}}({\mathbb{G}}) {\otimes}{\mathcal{O}}({\mathbb{G}}) \to B({\mathcal{H}})$ is the $*$-homomorphism defined by $\vartheta(a {\otimes}b) \xi = a \xi S(b), \ \xi \in {\mathcal{H}}$. Hence $\Theta$ is a $*$-homomorphism as well and therefore extends to ${C_u(\mathbb G)}$. Moreover, the conditions on $(\xi_{\lambda})$ imply $$\| \Theta(x)\xi_{\lambda} - {\varepsilon}(x) \xi_{\lambda} \| \to 0 \ \ \ \forall x \in {C_u(\mathbb G)}.$$ Indeed, it suffices to show this for $x$ being a coefficient of a irreducible corepresentation $\pi \in \operatorname{Irr}({\mathbb{G}})$ and in that case one computes $$\begin{aligned}
\| \Theta(u_{ij}^{\pi})\xi_{\lambda} - \delta_{ij} \xi_{\lambda} \|^2 \xrightarrow{\lambda} \sum_{k,l=1}^{d(\pi)} h((u_{il}^{\pi})^*u_{ik}^{\pi}(u_{jk}^{\pi})^*u_{jl}^{\pi}) - 2 \sum_{k=1}^{d(\pi)} h(u_{ik}^{\pi}(u_{jk}^{\pi})^*) + \delta_{ij} = 0.\end{aligned}$$ Since ${\hat{\mathbb G}}$ has property (T), by Definition \[yuki1\], we can find a projection $q \in {C_u(\mathbb G)}$ such that $xq = \varepsilon(x) q$ for all $x \in {C_u(\mathbb G)}$ and in particular we have $\varepsilon(q) =1$. Defining $\mu_{\lambda} = \Theta(q) \xi_{\lambda}$, it follows that $ \| \xi_{\lambda} - \mu_{\lambda} \| \to 0$. It only remains to prove that the vector $\mu_{\lambda}$ is ${L^{\infty}(\mathbb G)}$-central for every $\lambda$. To see this, observe first that for $\pi \in \operatorname{Irr}({\mathbb{G}})$, we have $$\sum_{k=1}^{d(\pi)} u_{ik}^{\pi} \mu_{\lambda}(u_{jk}^{\pi})^* = \Theta(u_{ij}^{\pi}) \mu_{\lambda} = \Theta(u_{ij}^{\pi}q) \xi_{\lambda} = \delta_{ij} \mu_{\lambda}.$$ Therefore, the computation $$\begin{aligned}
\mu_{\lambda}u_{il}^{\pi} = \sum_{j=1}^{d(\pi)} \delta_{ij} \mu_{\lambda}u_{jl}^{\pi}
= \sum_{j,k=1}^{d(\pi)} u_{ik}^{\pi} \mu_{\lambda}(u_{jk}^{\pi})^* u_{jl}^{\pi}
= u_{il}^{\pi} \mu_{\lambda}\end{aligned}$$ for $\pi \in \operatorname{Irr}({\mathbb{G}}), \ i,l= 1, \dots, d(\pi),$ concludes the argument.\
Let us now assume that ${L^{\infty}(\mathbb G)}$ has property (T). We prove that $\operatorname{Rep}{\mathbb{G}}$ has property (T), which is equivalent to central property (T) by Theorem \[propTtheorem\]. Let $(\varphi_{\lambda})_{\lambda}$ be a net of cp-multipliers converging to $\varepsilon$ pointwise. Without loss of generality, we can assume that $\varphi_{\lambda}(1) = 1$ for all $\lambda$. By Proposition 6.1 in [@popavaes], we obtain a net of $h$-preserving unital completely positive maps $\Psi_{\lambda}: {L^{\infty}(\mathbb G)}\to {L^{\infty}(\mathbb G)}$ such that $\Psi_{\lambda}(u_{ij}^{\pi}) = \varphi_{\lambda}(\pi) u_{ij}^{\pi} $ for all $\pi \in \operatorname{Irr}({\mathbb{G}}), \ i,j = 1, \dots, \dim \pi$. The pointwise convergence of the net $(\varphi_{\lambda})_{\lambda}$ then implies that the unital completely positive maps $\Psi_{\lambda}: {L^{\infty}(\mathbb G)}\to {L^{\infty}(\mathbb G)}$ converge pointwise to the identity, i.e. $$\| \Psi_{\lambda}(x) - x \|_2 \to 0, \ \ \ \forall x \in {L^{\infty}(\mathbb G)}\ \ \ \text{as} \ \ \ \lambda \to \infty.$$ It follows from the assumption that ${L^{\infty}(\mathbb G)}$ has property (T) that $$\sup_{x \in {L^{\infty}(\mathbb G)}_1}\| \Psi_{\lambda}(x) - x \|_2 \to 0 \ \ \ \text{as} \ \ \ \lambda \to \infty.$$ Now, for all $\pi \in \operatorname{Irr}({\mathbb{G}})$ and all $\lambda$, consider the unital completely positive map $${\mathord{\text{\rm id}}}_{\pi} {\otimes}\Psi_{\lambda}: B(H_{\pi}) {\otimes}{L^{\infty}(\mathbb G)}\to B(H_{\pi}) {\otimes}{L^{\infty}(\mathbb G)}$$ and note that $({\mathord{\text{\rm id}}}_{\pi} {\otimes}\Psi_{\lambda})(u^{\pi}) = \varphi_{\lambda}(\pi) u^{\pi} $. Hence, $$\sup_{\pi \in \operatorname{Irr}({\mathbb{G}})} |\varphi_{\lambda}(\pi) -1| = \sup_{\pi \in \operatorname{Irr}({\mathbb{G}})} \| (\varphi_{\lambda}(\pi) -1) u^{\pi} \|_2 = \sup_{\pi \in \operatorname{Irr}({\mathbb{G}})} \| ({\mathord{\text{\rm id}}}_{\pi} {\otimes}\Psi_{\lambda})(u^{\pi}) - u^{\pi} \|_2 \to 0,$$ which establishes property (T) in the categorial sense.
[99]{} C.A. Akemann and M.E. Walter, *Unbounded negative definite functions*, Canad. J. Math. **33** (1981), 862–871.
M. Alaghmandan, *Amenability notions of hypergroups and some applications to locally compact groups*, Math. Nachr. **290** (2017), 2088–2099.
Y. Arano, *Unitary spherical representations of Drinfeld doubles*, J. Reine Angew. Math., published online (2016).
Y. Arano, T. de Laat and J. Wahl, *Howe-Moore type theorems for quantum groups and rigid $C^{\ast}$-tensor categories*, Compos. Math. **154** (2018), 328–341.
Y. Arano and S. Vaes, *$C^{\ast}$-tensor categories and subfactors for totally disconnected groups*, Operator Algebras and Applications, The Abel Symposium 2015, Springer, 2016, pp. 1–43.
D. Blecher and V. Paulsen, *Tensor products of operator spaces*, J. Funct. Anal. **99** (1991), 262–292.
N.P. Brown and N. Ozawa, *$C^{\ast}$-Algebras and Finite-Dimensional Approximation Properties*, Amer. Math. Soc., Providence, 2008.
B. Das and M. Daws, *Quantum Eberlein compactifications and invariant means*, Indiana Univ. Math. J. **65** (2016), 307–352.
M. Daws, *Multipliers, self-induced and dual Banach algebras*, Dissertationes Math. (Rozprawy Mat.) **470** (2010), 62 pp.
M. Daws, A. Skalski and A. Viselter, *Around property (T) for quantum groups*, Comm. Math. Phys. **353** (2017), 69–118.
K. De Commer, A. Freslon and M. Yamashita, *CCAP for universal discrete quantum groups*, Comm. Math. Phys. **331** (2014), 677–701.
E. Effros and Z.J. Ruan, *A new approach to operator spaces*, Canadian Math. Bull. **34** (1991), 329–337.
P. Eymard, *L’algèbre de Fourier d’un groupe localement compact*, Bull. Soc. Math. France **92** (1964), 181–236.
P. Fima, *Property T for discrete quantum groups*, Internat. J. Math. **21** (2010), 47–65.
S.K. Ghosh and C. Jones, *Annular representation theory for rigid $C^{\ast}$-tensor categories*, J. Funct. Anal. **270** (2016), 1537–1584.
R. Godement, *Les fonctions de type positif et la théorie des groupes*, Trans. Amer. Math. Soc. **63** (1948), 1–84.
U. Haagerup, S. Knudby and T. de Laat *A complete characterization of connected Lie groups with the Approximation Property*, Ann. Sci. Éc. Norm. Supér. (4) **49** (2016), 927–946.
U. Haagerup and J. Kraus, *Approximation properties for group [$\operatorname{C}^*$-algebras]{} and group von [Neumann]{} algebras*, Trans. Amer. Math. Soc. **344** (1994), 667–699.
U. Haagerup and T. de Laat, *Simple Lie groups without the Approximation Property*, Duke Math. J. **162** (2013), 925–964.
U. Haagerup and T. de Laat, *Simple Lie groups without the Approximation Property II*, Trans. Amer. Math. Soc. **368** (2016), 3777–3809.
C. Jones, *Quantum $G_2$ categories have property (T)*, Internat. J. Math. **27** (2016), 1650015.
D.A. Kazhdan, *Connection of the dual space of a group with the structure of its closed subgroups*, Funk. Anal. Appl. **1** (1967), 63–65.
D. Kyed, *A cohomological description of property (T) for quantum groups*, J. Funct. Anal. **261** (2011), 1469–1493.
V. Lafforgue and M. de la Salle, *Noncommutative [$L^p$]{}-spaces without the completely bounded approximation property*, Duke. Math. J. **160** (2011), 71–116.
H. Leptin, *Sur l’algèbre de Fourier d’un groupe localement compact*, C. R. Acad. Sci. Paris Sér. A **266** (1968), 1180–1182.
B. Liao, *Approximation properties for $p$-adic symplectic groups and lattices*, preprint (2015), arXiv:1509.04814.
V. Muruganandam, *Fourier algebra of a hypergroup. I*, J. Aus. Math. Soc. **82(1)** (2007), 59–83.
S. Neshveyev and L. Tuset, *Compact Quantum Groups and their Representation Categories*, Société Mathématique de France, Paris, 2013.
S. Neshveyev and M. Yamashita, *Drinfeld center and representation theory for monoidal categories*, Comm. Math. Phys. **345** (2016), 385–434.
A. Ocneanu, *Chirality for operator algebras*, Subfactors (Kyuzeso, 1993), World Sci. Publ., River Edge, 1994, pp. 39–63.
G. Pisier, *Introduction to Operator Spaces*, Cambridge University Press, Cambridge, 2003.
S. Popa and S. Vaes, *Representation theory for subfactors, $\lambda$-lattices and $C^{\ast}$-tensor categories*, Comm. Math. Phys. **340** (2015), 1239–1280.
S. Popa, D. Shlyakhtenko and S. Vaes, *Cohomology and $L^2$-Betti numbers for subfactors and quasi-regular inclusions*, to appear in Int. Math. Res. Not.
V. Runde, *Uniform continuity over locally compact quantum groups*, J. London Math. Soc. **80** (2009), 55–71.
M. Takesaki, *Theory of Operator Algebras II*, Springer-Verlag, Berlin, 2003.
P. Tarrago and J. Wahl, *Free wreath product quantum groups and standard invariants of subfactors*, preprint (2016), arXiv:1609.01931.
A. Valette, *Minimal projections, integrable representations and property (T)*, Arch. Math. (Basel) **43** (1984), 397–406.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The zero temperature quenching dynamics of the ferromagnetic Ising model on a densely connected small world network is studied where long range bonds are added randomly with a finite probability $p$. We find that in contrast to the sparsely connected networks and random graph, there is no freezing and an initial random configuration of the spins reaches the equilibrium configuration within a very few Monte Carlo time steps in the thermodynamic limit for any $p \ne 0$. The residual energy and the number of spins flipped at any time shows an exponential relaxation to equilibrium. The persistence probability is also studied and it shows a saturation within a few time steps, the saturation value being 0.5 in the thermodynamic limit. These results are explained in the light of the topological properties of the network which is highly clustered and has a novel small world behaviour.'
address: ' Department of Physics, University of Calcutta, 92 Acharya Prafulla Chandra Road, Kolkata 700009, India. '
author:
- Pratap Kumar Das and Parongama Sen
title: Zero temperature dynamics of Ising model on a densely connected small world network
---
=-.3in
PACS no: 89.75.-k
Reprint no: CU-physics-04/2005
[2]{}
Introduction
============
It is well known that the one dimensional Ising model with nearest neighbour interaction does not have any non-trivial phase transition. However, a drastic change is observed in its critical behavior when even a few long range interactions are added randomly. Such a one dimensional lattice with extra random connections is known to behave as a small-world network (SWN)[@Watts; @newman; @BA], which means that the average shortest distance between any two sites in this lattice scales with the logarithm of the number of sites. The Ising model on small world networks not only has a finite temperature phase transition [@Barrat; @Gitter], but the critical behavior is also mean-field like [@kim; @herrero; @hong; @lopes; @doro; @leone]. A network in which the distribution of the number of links follows a power-law is known as a scale free network (SFN). Here also a finite temperature phase transition of the Ising model with a diverging critical temperature [@scale] has been observed.
Recently, quite a few studies on the dynamical properties of the Ising model on random graphs as well as networks have been reported both at finite and zero temperatures [@sven1; @hagg; @boyer; @sven2; @zhu; @jeong; @castell]. Dynamics of Ising models is a much studied phenomenon and has emerged as a rich field of present-day research. Models having same static critical behavior may display different behavior when dynamic critical phenomena are considered [@Ho_Ha]. An important dynamical feature commonly studied is the quenching phenomenon below the critical temperature. In a quenching process, the system has a disordered initial configuration corresponding to a high temperature and its temperature is suddenly dropped. This results in quite a few interesting phenomena like domain growth [@gunton; @bray], persistence [@satya1; @derrida; @stauffer; @krap1; @Krap_Redner] etc.
In one dimension, a zero temperature quench of the Ising model ultimately leads to the equilibrium configuration, i.e., all spins point up (or down). The average domain size $D$ increases in time $t$ as $D(t)\sim t^{1/z}$, where $z$ is the dynamical exponent. As the system coarsens, the magnetisation also grows in time as $m(t)\sim t^{1/2z}$. In two or higher dimensions, however, the system does not always reach equilibrium [@Krap_Redner] although these scaling relations still hold good.
Apart from the domain growth phenomenon, another important dynamical behavior commonly studied is persistence. In Ising model, persistence is simply the probability that a spin has not flipped till time $t$ and is given by $P(t)\sim t^{-\theta}$. $\theta$ is called the persistence exponent and is unrelated to any other static or dynamic exponents. Persistence can also be studied at finite temperatures and the exponent may change at the critical temperature [@stauffer; @derrida2].
We have studied the zero temperature quenching dynamics of the ferromagnetic Ising model on a SWN. It is important to carefully describe the type of network under consideration. There are two well-known methods of generating a small world network starting from a chain of nodes having connections with nearest neighbours only. These are (a) the addition type, where new long range (LR) bonds are added randomly keeping the nearest neighbour connections intact, and (b) the rewiring type where the existing nearest neighbour bonds are rewired to distant neighbours randomly. In the first case, when bonds are added with probability $p$, the total number of LR bonds is $pN^2$ for large $N$ ($N$ is the number of nodes). Even with $p \sim 1/N$, i.e., with a finite number of LR bonds, a phase transition has been observed in the Ising model [@Gitter]. While considering the dynamics of Ising models on addition type network, again $p = \gamma/N$ has been the usual choice, where $\gamma $ is a finite quantity. The dynamical behaviour of such a network for any value of $\gamma>1$ is comparable to that of the random graph in the sense that the system fails to reach its global minimum energy with zero temperature Glauber dynamics even in the thermodynamic limit [@sven1; @hagg]. Quenching dynamics of the Ising model on the scale-free network has also been considered recently [@castell] with an average connectivity $k=pN$ for each node. Here $k$ was fixed such that $p \sim 1/N$. The results again show a freezing at zero temperature.
In our study, we have considered an addition type network generated from a one-dimensional chain of Ising spins with nearest neigbours. Here each node has $pN$ number of LR bonds with $p$ fixed (i.e., finite in the limit $N \to \infty$) such that the network is a densely connected network. In this network, we have shown that the freezing or blocking effect is removed for any $p > 0$ in the thermodynamic limit. This conclusion is reached by studying various quantities like the domain sizes, magnetisation, residual energy, number of flipped spins etc. as functions of time. We also find that the relaxation to the equilibrium state is exponential.
The study of persistence of the Ising model on this network shows that there is no algebraic decay as it reaches a constant finite value similar to what happens in lattices of dimension greater than three [@stauffer]. The behaviour of $P(t)$ with $p$ is described in detail later in the paper.
In section II we describe the dynamical model under consideration and the physical quantities calculated. The topological properties of the network are described in section III where we find that there is a novel behaviour of the network as far as small world property is concerened. The results are discussed in section IV. Summary and some conclusive comments have been given in section V.
The Models and Dynamic properties studied
=========================================
We have considered a one-dimensional ferromagnetic Ising model on a network, in which, apart from nearest neighbour links, there exist some random long range connections with probability $p$. The Hamiltonian for this system is $$H=-\sum_{i}J_{ij}s_{i}s_{j},$$ where $s_i=\pm1$ is the state of the spin at the $i$th site, $J_{ij}=1 $ for nearest neighbours and for other neighbours equal to 1 with probability $p$. The Hamiltonian should be divided by a factor of $pN$, however, at $T=0$, only the sign of the energy differences are required for the Glauber dynamics and therefore this factor has not been included. We have simulated this system with periodic boundary condition. The initial configuration is random and single spin flip Glauber dynamics has been used for subsequent updating, i.e., a spin is picked up at random and flipped if the resulting configuration has lower energy, never flipped if the energy is raised and flipped with probability 1/2 if there is no change in energy on flipping.\
We have estimated the following quantities in the system.\
$(1)$ Average domain size $D(t)$ at time $t$.\
$(2)$ Magnetization $m(t)$ as a function of time (the average magnetisation being zero from symmetry, here $m(t)$ has been calculated by taking the average of the absolute values of the magnetisation).\
$(3)$ Residual energy $E_r(t) = E(t) -E_g$ where $E_g$ is the energy of the ground state and $E(t)$ the energy at time $t$.\
$(4)$ $N_{flip}(t)$, the number of spin flips at time $t$ per unit time.\
$(5)$ Persistence probability $P(t)$ defined as in section $I$.\
When considering the domains, we have in mind the original one-dimensional lattice and measure the domain size along it.
In finite dimensional nearest neighbour Ising models, the dynamics of these quantities is governed by the exponents $\theta$ and $z$, i.e.,\
$D(t)\sim t^{1/z}$\
$m(t)\sim t^{1/{2z}}$\
$E_r(t) \sim t^{-1/z}$\
$N_{flip}(t)\sim t^{-1/z}$\
$P(t)\sim t^{-\theta}$\
with $\theta=0.375$ and $z=2$ in one dimension.\
In principle, it is not essential to study all these quantities to determine the characteristics of the dynamical system. However, in a numerical study, it is better to check that the behavior of these different quantities is consistent with a single $z$ and $\theta$ .
Since freezing is a key question here, we have also calculated the freezing probability $F(p)$ as a function of $p$ for various system sizes. Several other quantities and exponents related to domain dynamics in the quenching phenomena can be defined [@krap2], but here we have restricted our study to those mentioned above.
In the simulations, we have restricted the system sizes to $N\leq 1000 $ as a large number of configurations is required to have accurate data. The results have been averaged out over $1000$ initial configurations and network configurations (typically). As the network is densely connected, the updatings consume a lot of CPU time forcing us to restrict our study to rather small system sizes.
Topological properties of the network
=====================================
The essential difference between a simple one dimensional lattice with nearest neighbour links and the present network lies in the topology of the networks. The topological properties of a network like the average shortest distances $\langle S \rangle$ (here distance means number of steps required to reach another node) and the clustering coefficient $\cal C$ may help in understanding the static and dynamical phenomena of Ising model on such network. In the nearest neighbour lattice, $\langle S \rangle$ is proportional to $N$ and $\cal C$ is zero (as there are no loops).
[*Super small world effect*]{}: The average shortest distance behaves in the expected manner; $\langle S\rangle$ is very small ($O(10)$) for very small values of $p$ and decreases to the exact value 1 in the $p\to 1$ limit. In a small world network, the shortest distance is supposed to scale as $\ln(N)$. Here, if $p$ is kept fixed, $\langle S\rangle$ actually decreases with $N$ showing that it approaches the behaviour corresponding to $p=1$ in the thermodynamic limit. This is a novel behaviour which one can call a “super small world” effect as noted earlier in [@zhou]. One can also compare data for different $N's$ by keeping the number of edges per site, $pN$, a constant rather than $p$. Therefore, in two networks of size $N_1$ and $N_2$, the values of $p$ are kept $p_1$ and $p_2$ respectively such that $p_1N_1 = p_2N_2$. We then observe that $\langle S_{N_1,p_1}\rangle /\langle S_{N_2,p_2}\rangle = \ln(N_1)/\ln(N_2)$ which is true for the conventional small world networks. For example, for $N_1=100$ and $p_1=0.05$, $\langle S_{100,p_1}\rangle=2.58 $ and for $N_2=500$ and $p_2=0.01, ~~\langle S_{500,0.01}\rangle=3.43$, and the ratio of these two quantities is very close to that of $\ln(100)$ and $\ln(500)$.
The clustering coefficient is estimated in the standard way: the probability that two neighbours of a particular node are also connected to each other is a measure of the clustering tendency. $\cal C$ is also calculated for different $p$ and $N$ values. Obviously it increases with $p$ and is equal to one at $p=1$ when all the nodes are connected to each other. Since the nearest neighbours are already connected, clustering coefficient would increase if the long range bonds happen to be next nearest neighbours [@PS]. For small values of $p$, this is more probable in smaller lattices and therefore clustering coefficient shows a marked decrease with $N$. Fig. $1$ shows the data for both $\langle S \rangle$ and $\cal C$.
These results will be helpful to interpret the observations which we have made and will be referred to in the later sections.
RESULTS and DISCUSSIONS
=======================
We first discuss the results for the growth of the domains sizes and magnetisation. The domain sizes have been scaled by the system sizes $N$ such that $D(t) \leq 1$. We have verified that for $p=0$, all the physical quantities under consideration follow the known behaviour summarised in section II (with $z$ and $\theta$ assuming the values corresponding to one dimension). As soon as a non-zero $p$ value is introduced, both $m(t)$ and $D(t)$ quickly reach a saturation value such that a variation with time occurs only over a short initial period of time (Fig. 2).
For small system sizes and at low $p$, the saturation values of $m(t)$ and $D(t)$ are far from those of the equilibrium values (both unity at $T$=0) similar to the results on random graphs and sparsely connected small world networks. This apparently suggests that the system gets “frozen” in one of the metastable states. However, in the present model, blocking seems to be effective only for finite sizes, as the saturation values of both magnetisation and average domain size approach unity (i.e., the equilibrium value) when the system size is increased (Fig. 3). This is true for any finite value of $p \neq 0$. The apparent blocking effect is more prominent for small values of $p$.
The comparison of $m(t)$ and $D(t)$ for different system sizes also shows that the initial growth becomes rapidly sharper with the system size, so that any time dependence in the initial period loses its significance. The period over which this growth takes place also shrinks in size in larger system sizes (Fig. 3) signifying a very fast growth in the magnetization.
The study of the distribution of magnetisation is also consistent with the fact that blocking occurs for finite sizes only. The distribution has finite values for all values of $m$ ($ -1 < m < 1$) for small $N$, but for larger sizes has non-zero values very close to $m =\pm 1$ only.
The fast growth in $m(t)$ and $D(t)$ is supported by the behaviour of $E_r(t)$ and $N_{flip}(t)$ as both show an exponential decay ($E_r(t), ~N_{flip}(t) \sim
\exp(-\alpha t)$) with $\alpha =1$ for all $p$ (Fig. 4).
[*[Freezing probability]{}*]{}
The above observations suggest that for finite sizes, the probability that the system goes to a frozen state is finite for small values of $p$. Since for the limiting cases $p=1$ and $p=0$, there is no freezing, it is expected that the freezing probability $F(p)$ will have a peak for a non-zero value of $p$. We study the freezing probability for fixed values of $N$ and find out some interesting features. There is indeed a peak occurring at $p=p_m$ with $p_m \sim 1/N $. Except for very small values of $p$, $F(p)$ decreases with $N$ signifying the disappearance of freezing in the thermodynamic limit. For small values of $p$, the behaviour may be different. Specifically, if $p = 1/N$, we find that the freezing probability $increases$ with $N$, which is in consistency with the observation of [@castell]. But at this value of $p$, the network is a sparsely connected one and therefore it has a freezing tendency indicated by this increase. In fact, keeping $p$ fixed at a certain value such that $p < p_m$ (for the system sizes concerned) we find that the freezing tendency gets enhanced with $N$. Also, we find that $F(p)$ shows an exponential decay beyond $p_m$: $F(p) \sim \exp(-\gamma p)$ for large values of $p$ with $\gamma \sim N$.
These results (shown in Fig. 5) indicate again that there is a sharp discontinuity in the dynamical behaviour at $p=0$ and that freezing disappears for any non-zero $p$ for large values of $N$ as long as $p$ is a finite quantity.
In reference [@boyer], it had been shown how the domain walls get pinned when the number of extra bonds is $ O(N) $, i.e., $p\sim 1/N$ and the system freezes. The domain walls become mobile as soon as the number of long range interactions increase to $O(N^2)$ and ultimately they disappear by annihilating each other. Thus, in a densely connected infinite network, freezing disappears for any $p\neq 0$.
We next study the behaviour of the persistent probability $P(t)$ with time for different values of $p$ (Fig. 6). $P(t)$ follows the well known power law decay for $p=0$, but quickly falls to a finite saturation value $P_{sat}$ for any non-zero value of $p$. The decay is sharper for higher values of $p$. The saturation behaviour is similar to that of $D(t)$ and $m(t)$.
The behaviour of $P_{sat}$ with different system sizes however, shows an interesting feature. In a finite $d$-dimensional lattice of size $L$, when $P(t)$ decays algebraically in time with the exponent $\theta$, the persistent probability as a function of $t$ and $L$ is given by
$$P(L, t) \sim t^{-\theta}f(t/L^z),$$
where $f(x) =$ constant for small $x$ and $f(x) =x^\theta$ for large $x$ such that $P(L, t\to \infty) \sim L^{-z\theta}$ which indicates that the time independent persistence probability decreases with $N$, where $N=L^d$, the total number of lattice sites. For non-zero values of $p$, when there is no algebraic decay of the persistent probability with time, we find that (Fig. 7) there exists a value of $p=p^*$ below which the persistence probability actually increases with $N$, the increase with $N$ being slower than a power law. Beyond $p^*$, $ P_{sat}$ decreases with $N$ but the decrease is fairly weak. However for both $p < p^*$ and $p > p^*$, $P(N,t,p)$ approaches a constant close to 0.5 from below and above respectively. That the saturation value of the persistence probability is close to 0.5 can be justified: initially fifty persent spins are up/down, and spins of only one type are flipped only within the short time the system reaches the equilibrium state. The value of $p^* \simeq 0.25$.
Summary and Conclusions
=======================
We have studied the dynamics of the ferromagnetic Ising model on a small-world network at zero temperature. The network is densely connected in the sense that there is a finite number of extra bonds for each node. It is observed that addition of long range bonds with probability $p$ brings the initial random configuration of Ising spins to the equilibrium ferromagnetic configuration within a very few time steps even if $p$ is very small. Magnetisation and average domain size quickly reach the saturation values without showing any scaling behaviour with time. Consistent with this observation, the residual energy and the number of flippings at any time shows an exponential decay. The study of the freezing probability shows a peak occurring in the distribution at $p_m \sim 1/N$ and an exponential decay which becomes faster with the system size. Persistence probability also reaches a saturation value within a few time steps. These results obtained for the quenching dynamics indicate that the dynamical behaviour of this densely connected network is much different compared to that of random graphs and sparsely connected SWN where the same dynamics leads to a frozen state not equivalent to the equilibrium ground state. That there is no power law behaviour but exponential relaxation is consistent with the mean field behaviour of the network [@silva].
In order to explain our observations, we notice that the present model is highly clustered as the density of connections is large (see section III). In contrast, the addition type small world network (with two nearest neighbour only), generated with a $p \sim 1/N$ has a vanishing clustering coefficient even though it has small world property [@PS]. This is in fact the reason for its behaviour as a random graph which also has a small world effect but vanishing clustering coefficient. The large clustering in our model is effective in making the domains entirely non-local and therefore the system can reach the global equilibrium state very fast. The super small world effect, by which we mean that the average shortest distance decreases with $N$ (tending to unity), also helps in understanding the results. For any $p \neq 0$, the network flows towards the $p=1$ fixed point (for large sized networks) for which one does not expect any freezing.
Our results are consistent with some very recent observations [@castell] where the dynamics of Ising model on a scale free network with increasing number of connectivity $k=pN$ has been studied. From the data shown, one can obtain the freezing probability as a function of a fixed $p$ (e.g., $p=0.1$) as well, which shows that freezing will $disappear$ for large system sizes.
Usually persistence probability has a unique behaviour and is governed by an independent exponent $\theta$ not related to the dynamic exponent $z$ which dictates the behaviour of $D(t), ~m(t), E_r(t)$ and $N_{flip}(t)$. Here one cannot make any statement about the independence of persistence and the domain growth phenomena. The only distinctive behaviour of persistence seems to be a difference in behaviour with finite sizes occurring for $p< p^*$ where $p^* \simeq 0.25$. At present our understanding of the network and the dynamics is not enough to explain the significance of $p^*$, although the limiting value of the persistence probability being 0.5 is easily explained from the mean field nature of the network.
From the results reported in the present paper, we conclude that for quenching dynamics, for any $p \le 1/N$ the time evolution leads to a frozen state far from equilibrium whereas with a finite $p$ freezing is overcome. The finite density of connections thus acts as a driving force, like a finite temperature, which drives the system out of the frozen state. This may appear as a dynamical phase transition in finite systems, however, in the thermodynamic limit ($N \to \infty$) of course, there is no such transition.
We finally remark that as far as statics is concerned, a finite value of $p$ is not required to get a phase transition but the finiteness of $p$ is essential to remove the dynamic frustration when dynamics is considered.
.5cm
Acknowledgments: We thank S. Dasgupta and S. N. Majumdar for helpful discussions. P. K. Das acknowledges support from CSIR grants no. 9/28(608)/2003-EMR-I. PS acknowledges DST grant number SP/S2/M-11/99.
.8cm Email: [pra$\_$tapdas]{}@rediffmail.com, [email protected]
[references:]{} D. J. Watts and S. H. Strogatz, Nature, [**[393]{}**]{} 440 (1998). M. E. J. Newman, D. J. Watts, Phys. Rev. E [**60**]{} 7332 (1999). R. Albert and A.-L. Barabási, Rev. Mod. Phys. [**[74]{}**]{} 47 (2002). A. Barrat and M. Weigt, Eurphys. J. B [**[13]{}**]{} 547 (2000). M. Gitterman, J. Phys.A [**[33]{}**]{} 8373 (2000). B. J. Kim etal, [Phys. Rev. E **[64]{}**]{} 056135 (2001). C. P. Herrero, Phys. Rev. E [**[65]{}**]{} 066110 (2002). H. Hong, B. J. Kim and M. Y Choi, Phys. Rev. E [**[66]{}**]{} 011107 (2002). J. Viana Lopes, Yu. G. Pogorelov, J. M. B. Lopes dos Santos and R. Toral, [Phys. Rev. E [**[70]{}**]{} 026112 (2004)]{}.
S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, Phys. Rev. E [**[66]{}**]{} 016104 (2002). M. Leone, A. Vazquez, A. Vespignani and R. Zecchina, Eur. Phys. J B [**[28]{}**]{} 191 (2002). A. Aleksiejuk, J. A. Holyst and D. Stauffer, Physica A [**310**]{} 260 (2002); G. Bianconi, Phys. Lett. A [**303**]{} 166 (2002) ; J. Viana Lopes, Yu. G. Pogorelov, J. M. B. Lopes dos Santos and R. Toral, [Phys. Rev. E [**[70]{}**]{} 026112 (2004)]{}. P. Svenson, Phys. Rev. E [**[64]{}**]{} 036122 (2001). O. Haggstrom, [Physica A [**[310]{}**]{} 275 (2002)]{}. D. Boyer and O. Miramontes, [Phys. Rev. E [**[67]{}**]{} R035102 (2003)]{}. P. Svenson and D. A. Johnson, [Phys. Rev. E [**[65]{}**]{} 036105 (2002)]{}. J. Y. Zhu and H. Zhu, [Phys. Rev. E [**[67]{}**]{} 026125 (2003)]{}. D. Jeong, M. Y. Choi and H. Park, [Phys. Rev. E [**[71]{}**]{} 036103 (2005)]{}. C. Castellano, V. Loreto, A. Barrat, F. Cecconi and D. Parisi, [Phys. Rev. E [**[71]{}**]{} 066107 (2005)]{} . P. C. Hohenberg and B. I. Halperin, [Rev. Mod. Phys. [**[49]{}**]{} 435 (1977)]{}. J. D. Gunton, M. San Miguel and P. S. Sahni, [*[Phase Transitions and critical phenomena]{}*]{}, Vol 8, eds. C. Domb and J. L. Lebowitz (Academic, NY 1983). A. J. Bray, [Adv. Phys. [**[43]{}**]{} 357 (1994) and the references therein]{}. For a review, see S. N. Majumdar, [Curr. Sci. [**[77]{}**]{} 370 (1999)]{}. B. Derrida, A. J.Bray and C. Godreche, [J.Phys. A [**[27]{}**]{} L357 (1994)]{} D. Stauffer, [J. Phys. A [**[27]{}**]{} 5029 (1994)]{}. P. L. Krapivsky, E.Ben-Naim and S. Redner, [Phys. Rev.E [**[50]{}**]{} 2474 (1994)]{}. V. Spirin, P. L. Krapivsky and S. Redner, [Phys. Rev.E [**[63]{}**]{} 036118 (2001)]{}. B. Derrida, [Phys. Rev. E [**[55]{}**]{} 3705 (1997)]{}. P. L. Krapivsky and E. Ben-Naim, Phys. Rev. E [**[56]{}**]{} 3788 (1997). N. R. da Silva and J. M. Silva, [Phys. Lett. A [**[135]{}**]{} 373 (1989)]{}. Tao Zhou, Bing-Hong Wang, P. M. Hui and K. P. Chan, cond-mat/0405258; in the first version of this paper the authors used the term “super small world”. P. Sen, K. Banerjee and T. Biswas, [Phys. Rev. E [**[66]{}**]{} 037102 (2002)]{}.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a transfer learning-based solution for the problem of multiple class novelty detection. In particular, we propose an end-to-end deep-learning based approach in which we investigate how the knowledge contained in an external, out-of-distributional dataset can be used to improve the performance of a deep network for visual novelty detection. Our solution differs from the standard deep classification networks on two accounts. First, we use a novel loss function, *membership loss*, in addition to the classical cross-entropy loss for training networks. Secondly, we use the knowledge from the external dataset more effectively to learn *globally negative filters*, filters that respond to generic objects outside the known class set. We show that thresholding the maximal activation of the proposed network can be used to identify novel objects effectively. Extensive experiments on four publicly available novelty detection datasets show that the proposed method achieves significant improvements over the state-of-the-art methods.'
author:
- |
Pramuditha Perera and Vishal M. Patel\
Department of Electrical and Computer Engineering,\
Johns Hopkins University, Baltimore, MD 21218, USA\
[[email protected], [email protected]]{} [^1]
bibliography:
- 'egbib.bib'
title: Deep Transfer Learning for Multiple Class Novelty Detection
---
Introduction
============
In recent years, intelligent systems powered by artificial intelligence and computer vision that perform visual recognition have gained much attention [@he15deepresidual],[@NIPS2012_ALEX],[@VGG]. These systems observe instances and labels of known object classes during training and learn association patterns that can be used during inference. A practical visual recognition system should first determine whether an observed instance is from a known class. If it is from a known class, then the identity of the instance is queried through classification.
The former process is commonly known as novelty detection (or novel class detection) [@Markou03noveltydetection] in the literature. Given a set of image instances from known classes, the goal of novelty detection is to determine whether an observed image during inference belongs to one of the known classes. Novelty detection is generally a more challenging task than out-of-distribution detection [@OOD] since novel object samples are expected to be from a similar distribution to that of known samples.
![Novelty detection in dog-breed classification. **Left:** Sample images. **Right:** Feature representation. Both known (first row) and novel (second row) images are images of dogs. Given known images, the goal of novelty detection is to reject novel images. In order to do so, the knowledge of out-of-distributional images (final row), in this case non-dog images, are used to learn a suitable representation. []{data-label="fig:negative"}](negative3){width="\linewidth"}
In practice, the knowledge on unknown classes is not entirely absent. Given a set of known classes from a certain problem domain, generally unknown class data from the same problem domain is unavailable. However, in some cases it is possible to obtain data outside the known class from different problem domains, which we refer to as *out-of-distributional* samples. For example, for a dog-breed recognition application, ImageNet dataset [@ILSVRC] that contains images of objects may be considered as *out-of-distributional* data as shown as in Figure \[fig:negative\]. However, since the *out-of-distributional* data are from a different problem domain, they do not approximate the distribution of the *novel* samples well.
Nevertheless, since the deep-models produce generalizable features, the knowledge of *out-of-distributional* samples can be transferred to the original problem to aid novelty detection. When the problem considered is a $c$ class problem, and when the *out-of-distributional* data of $\mathcal{C}$ classes are available, the following three strategies are used to transfer knowledge for novelty detection in the literature:
1\. **Fine-tuning**: Network is first pre-trained on the *out-of-distributional* data and later fine-tuned on the training data of the given domain. Novelty is queried by thresholding the final activation score [@BendaleB16].
2\. **Feature Extraction**: Conventional novelty detection techniques [@Bodesheim_2013_CVPR],[@Liu],[@kextremes] are used based on the fine-tuned features.
3\. **Fine-tune $(c+\mathcal{C})$**: Network is first pre-trained on the *out-of-distributional* data. Both the training data and the *out-of-distributional* data are used to perform fine-tuning in $(c+\mathcal{C})$ classes together. Novelty is determined in the same way as in approach 1.
We note that in all these baselines, the *out-of-distributional* data is employed in the training process. In fact, any novelty detection method operating on the pre-trained/finetuned deep features are implicitly making use of the *out-of-distributional* data. In this work, we introduce a new framework to perform novelty detection based on transfer learning. First, we show that using cross-entropy loss alone for training is not optimal for the novelty detection task. Secondly, we empirically show that the *out-of-distributional* data can be used more effectively in training to produce better novelty detection performance with respect to considered baseline solutions. Specifically, we make following primary contributions in this paper.
1\. We propose an end-to-end novelty detection framework based on deep learning. To the best of our knowledge, this is one of the first end-to-end deep learning solutions targeting visual novelty detection.
2\. We introduce a new loss function, *membership loss* which has a similar functionality to that of the cross-entropy loss but encourages an embedding that produces high activation for known object classes consistently.
3\. We propose to take advantage of large-scale external datasets to learn the *globally negative filters* to reduce high activations caused by the novel images.
4\. We show empirically, that the proposed method outperforms the baseline novelty detection methods across four publicly available datasets.
Related Work
============
Object classification schemes are often equipped with a suitable mechanism to detect novel objects. For example, Eigenfaces [@Turk1991] was accompanied by a reconstruction error-based novel object detection method. In sparse representation-based classification (SRC) algorithm [@Wright:2009:RFR:1495801.1496037], Sparsity Concentration Index (SCI) was proposed for the same purpose. In contrast, there is no formal novelty detection mechanism proposed for deep-learning based classification. In its absence, thresholding the highest class activation score of the deep model has been used as a baseline in the literature [@BendaleB16]. As an alternative, several recent works have proposed novelty detection schemes based on deep features [@BendaleB16],[@kextremes]. In the same spirit, it is also a possibility to use classical novelty detection tools such as Kernel PCA [@HOFFMANN2007863], Kernel null space-based novelty detection (KNFST) [@Bodesheim_2013_CVPR] and its variants [@localnovelty],[@Liu] on deep features. KNFST operating on deep-features produces the current state of the art performance in visual novelty detection [@Liu]. However, advantages of deep-learning are not properly exploited in all of these approaches due to the absence of an end-to-end learning framework.
On the other hand, novelty detection problem has a close resemblance to both anomaly detection [@oza2019one], [@2018arXiv180105365P], [@Chandola:2009:ADS:1541880.1541882],[@oza2019active] and open-set recognition problems [@Scheirer_2013_TPAMI],[@BendaleB16]. Therefore, it is possible to solve anomaly detection using tools proposed in these alternative domains. In anomaly detection, given a single *normal* class, the objective is to detect out-of-class instances. One-class SVM [@Scholkopf:2001:ESH:1119748.1119749] and SVDD [@Tax:2004:SVD:960091.960109] are two of the most widely used tools in anomaly detection. Novelty detection can be viewed as an anomaly detection problem if all known classes are considered as a single augmented class. On the other hand, objective in open-set recognition (OSR) is similar to that of novelty detection. But in addition, OSR requires correct classification of samples detected as known samples. Therefore, it is also possible to use open-set recognition tools to perform novelty detection. However, we note that due to subtle differences in objectives, OSR algorithms are not optimal for novelty detection.
In the proposed framework, maximal activation of the final layer of a deep network is considered as a statistic to perform novelty detection. We design the network and choose loss functions appropriately so that this statistic is low for novel objects compared to the known object classes.
[Background]{} {#mechanics}
==============
In this section, we briefly review how deep networks produce activations in response to input stimuli. Based on this foundation, we introduce the notion of *positive filters* and *negative filters*. Consider a $c$ class fully-supervised object classification problem with a training image set $\mathbf{x} = {{x_1},{x_2},\dots,{x_n}}$ and the corresponding labels $\mathbf{y} = {{y_1}, {y_2}, \dots ,{y_n}}$ where $y_i \in \{1,2, \dots c\}$. Deep convolutional neural networks (CNNs) seek to learn a hierarchical, convolutional filter bank with filters that respond to visual stimuli of different levels. In $c$ class classification, the top most convolutional filter activation $\mathbf{g}$ is subjected to a non-linear transformation to generate the final activation vector $\mathbf{f} \in \mathbb{R}^{c}$ (for example, $\mathbf{g}$ is the conv5-3 layer in VGG16 and conv5c in Resnet50. $\mathbf{f}$ is the fc8 and fc1000 layers in the respective networks). In a supervised setting, network parameters are learned such that $ \arg\max \mathbf{f} = {y_i}$ for $ \forall i \in \{1,2,\dots ,n \}$. This is conventionally done by optimizing the network parameters based on the cross-entropy loss.
If there exist $k$ filters in the top most convolution filter bank, its output $\mathbf{g}$ is a set of $k$ number of activation maps. The final activation vector of the network $\textbf{f}$ is a function of $\mathbf{g}$. For a given class $i$, there exists some $k_i$ filters in the filter bank ($1 \leq k_i \leq k$) that generally generates positive activation values. These activations provide supporting (positive) evidence that an observed image is from class $i$. Conversely, all remaining filters provide evidence against this hypothesis. Activation score of each class in $\textbf{f}$ is determined by taking into account the evidence for and against each class. For the remainder of the paper, we call filters that provide evidence for and against a particular class as *positive filters* and *negative filters* of the class, respectively.
![Positive and negative filters of the *sand snake* class in the Resnet50 trained on ILSVRC12 dataset. Top: weights of the fully connected layer corresponding to the *sand snake* class. We call filters associated with positive weights as positive filters of *sand snake* class. All other filters are named as negative filters. Bottom: Visualization of top negative and positive filters. These patterns are likely to produce high activation in these filters. We note top positive filters are activated by snake-like structures. []{data-label="fig:snake"}](snake){width="1\linewidth"}
This concept can be easily explained by taking the Resnet architecture [@he15deepresidual] as an example. In Resnet, final convolution output $\mathbf{g}$ is subjected to global average pooling followed by a fully connected layer. Therefore, the $i^{th}$ component of the final activation vector $\mathbf{f}$ can be written as $\mathbf{f}_i = W_i \times GAP(g)$, where $GAP$ is global average pooling operation (mean of filter map) and $W$ is the weight matrix of the fully connected layer. Here, activation of the $i^{th}$ class is a weighted summation of mean feature maps found in $\mathbf{g}$. From the above definition, filters associated with positive weights for a given class in $W$ can be identified as *positive filters* for that particular class. Conversely, filters associated with the negative weights become *negative filters* of the class.
For example consider the *Sand Snake* class appearing in the ILSVRC12 dataset [@ILSVRC]. Shown in Figure \[fig:snake\] (top) are the weights associated with the *Sand Snake* class in the final fully connected layer of the Resnet50 network trained on the ILSVRC12 dataset. We recognize filters associated with positive and negative weights as positive and negative filters, respectively for the given class. In Figure \[fig:snake\] (bottom) we visualize per-unit visualization of top positive and top negative filters for the considered class using the DeepVis toolbox [@yosinski-2015-ICML-DL-understanding-neural-networks] (these are the images that are most likely to activate the corresponding filters). By observation, we notice that the top *positive filters* are activated when the network observes structures similar to snakes. On the other hand, the top *negative filters* are unrelated to the appearance of *sand snakes*.
Deep Novelty Detection {#obj}
======================
Based on the above background, we propose to learn the distributions of known object classes using a CNN framework with the objective of performing joint classification and novelty detection. In our formulation, assuming each known class has a unique single label, we force the final activation vector $\mathbf{f}$ to model the probability distribution vector of known classes. Formally, for a given data-label pair $(x_i,y_i)$, we expect $\mathbf{f}_i=1$ and $\mathbf{f}_j=0,~ \forall j\neq i$. Once such a representation is learned, $\arg \max \mathbf{f}$ returns the most-likely class of an observed sample. Then, $\max \mathbf{f}$ yields the likeliness of the sample belonging to the most likely class. Similar to binary classification, identity $I$ of a test instance can be queried using hard thresholding. In order to learn a representation suitable for the stated objective, we use conventional classification networks as the foundation of our work and propose the following two alternations.
**1. Membership loss.** Assuming each known class has a unique single label, if the observed image is from a known class, only a single positive activation should appear in $\mathbf{f}$. We observe that when cross-entropy loss is used, this is not the case. To alleviate this, we introduce a new loss called *[membership loss]{}* in addition to the cross-entropy loss.
**2. Globally negative filters.** In a classification setting, a negative filter of a certain class is also a positive filter of another class. In other words, there exist no explicit negative filters. In our formulation, we propose to generate *globally negative filters* (filters that generate negative evidence for all known classes) to reduce the possibility of a novel sample registering high activation scores.
{width="1\linewidth"}
[Limitations of Cross-Entropy Loss]{}
-------------------------------------
When a classification network is trained, each element $f_i$ of the activation vector $\mathbf{f}$ is first normalized using the softmax function to arrive at a normalized activation vector $\mathbf{\tilde{f}}$ as in, $\tilde{f_j} = e^{{f}_j} / {\sum\limits_{j=1}^{c}e^{{f}_j}}.$ When it is assumed that all image classes appearing during inference are known ahead of time, $j^{th}$ element of vector $\tilde{\mathbf{f}}$ is interpreted as the likelihood of the input image $x_i$ belonging to the $j^{th}$ class. Neural network-based classification systems are learned by minimizing the cross-entropy loss which is the negative log likelihood of the correct class $\tilde{\mathbf{f}}$. However, since this is a relative measure, the learned representation deviates from our objective due to the following reasons.
Firstly, even a low activation of the ground truth class could yield a low cross-entropy provided that the activations of all other (non-matching) classes are very low. As a result, lower score values may not get heavily penalized during training. Therefore, a model trained using the cross-entropy loss may end up producing low activation scores for known classes during inference. In closed set classification, this behavior will not cause complications as long as the correct class records the highest score. However, in threshold-based novelty detection, this poses a problem as having low scores for the positive class will result in false negatives. Secondly, the cross-entropy loss does not penalize activations of unrelated classes as long as the correct class produces the highest activation. As a result, inaccurate cross-class relationships are encouraged during training.
In order to illustrate this point, we trained a VGG16 [@DBLP:journals/corr/SimonyanZ14a] based CNN classification network using the first 128 classes of the Caltech256 dataset. For the considered example, the Calculator class (indexed at 27) is a known class and the Playing Cards class (indexed at 163) is a novel class. Shown in Figure \[fig:acti2\] are the activations of conv5-3 and fc8 layers of the network for two inputs of the two classes. As can be seen from this figure, when the network observes a calculator object (known object), it correctly associates the highest score in **f** to the correct class (class 27). However, there is also a significant miss-association between the calculator class and coin (class 43), keyboard (class 45), dice (class 55) and joystick classes (class 120).
Membership Loss
---------------
In our approach, we first independently translate each activation score value $f_i$ into the range $0 - 1$ using the sigmoid$(\sigma)$ function. We interpret each transformed activation score as the probability of the input image belonging to each individual class. If the ground truth label of a given observation $x$ is $y$, we aim at learning a function that produces absolute probabilities for the membership of each class as follows $$\mathbb{P}(y=i) = \sigma(f(x)_i) ~~\forall i \in \{1,2,\dots c\}.$$ Ideally, the learned transformation will produce $f(x)_i = 1 $ for $i=y$ and $f(x)_i = 0,$ otherwise. We denote the risk of associating a higher score with a wrong class ($f(x)_i = 1 $ for $i \neq y$ ) as $R_{W1}$ and risk of associating a low score with the correct class ($f(x)_i = 0 $ for $i=y$) as $R_{C0}$. We define the *membership loss $L_M$* as the risk of classification as $$\nonumber L_M(x,y) = R_{C0}(x,y) + \lambda R_{W1}(x,y),$$ where $\lambda$ is a positive scalar. With our formulation, we define $R_{W1}(x,y) = [1-\mathbb{P}(y=1)]^2 = [1-\sigma(f(x)_y)]^2$. Here, the quadratic term is introduced to impose a heavy penalty on very high deviations. Similarly, $R_{C0}(x,y)$ becomes, $$\begin{aligned}
\nonumber R_{C0}(x,y) &= \frac{1}{c-1}\sum\limits_{i=1, i\neq y}^{c}[\mathbb{P}(i=1)]^2\\ &= \frac{1}{c-1}\sum\limits_{i=1, i\neq y}^{c} [\sigma(f(x)_i)]^2.\end{aligned}$$ By substitution, we get $$\nonumber L_M(x,y) = [1-\sigma(f(x)_y) ]^2+ \lambda \frac{1}{c-1}\sum\limits_{i=1, i\neq y}^{c} [\sigma(f(x)_i)]^2.$$ Here, the parameter $\lambda$ controls relative weight given to each risk source. In our experiments, we set $\lambda = 5$. Taking the partial derivative of the membership loss yields the following back-propagation formula $$\nonumber \frac{\partial L_M(x,y) }{\partial f(x)_i}=
\begin{cases}
-2[1-\sigma(f(x)_i)] \times \sigma(f(x)_i)'& \text{for } i=y\\
\nonumber \frac{2\lambda}{c-1} \sigma(f(x)_i) \times \sigma(f(x)_i)'& \text{for } i\neq y,
\end{cases}$$ where, $\sigma(f(x)_i)' = \sigma(f(x)_i) (1-\sigma(f(x)_i))$.
The proposed *membership loss* does not operate on the closed-set assumption. It takes individual score values into account in an absolute sense. Therefore, when the membership loss is used, known samples that produce small activations will be penalized regardless of the score values of the other classes. When the membership loss is used together with the cross-entropy loss, the network learns a representation that produces relatively higher activation scores for the correct class. For example, consider the $fc8$ activation map of the proposed method for the Calculator object input shown in Figure \[fig:acti2\]. There, we observe that the correct class (indexed at 27) produces a large positive score whereas all other classes produce negative scores.
Globally Negative Filters
-------------------------
When a conventional classification network is used, novel images are often able to produce very high activation scores there by leading to false positive detections. Such an example is shown in Figure \[fig:acti2\](bottom) where a Playing Cards instance has produced a very high activation score in the index corresponding to the Calculator class (indexed at 27). Final activation score of a class is generated based on the responses of the positive and negative filters as discussed in Section \[mechanics\]. Once the network is trained, given an input of a particular known class, the input stimulates some *positive filters* and *negative filters* associated with the class. If the model is well trained, the response of the positive filters exceeds the response of the negative filters to produce a high positive activation score. Given this background, it is interesting to investigate how a novel sample is able to produce a high activation score. Let us revisit activations of Playing Cards image (novel image) shown in Figure \[fig:acti2\] (bottom). In this example, Playing Cards image has stimulated some positive filters of the Calculator class despite the differences in content. At the same time, by chance, it has not produced sufficient stimulation in negative filters of the Calculator class, thereby producing a large positive activation in $\mathbf{f}$. This can be clearly observed in Figure \[fig:acti2\] where both the Calculator and the Playing Cards images have activated similar filters in the conv5-3 layer.
To this end, we make the following proposal. We wish to learn a set of filters that are stimulated generally by natural images and produce evidence against all known classes. In other words, these filters are *negative filters* with respect to all known classes - hence we call them *globally negative filters*. If any of such filters are stimulated during inference, it would prove greater evidence that the observed image is novel. However, this proposal will succeed only if the *globally negative filters* are stimulated by arbitrary images outside the known class set. In order to learn the *globally negative filters*, we propose a joint-learning network structure. In addition to the known object dataset, we use the *out-of-distributional* data samples in training. For the remainder of the paper we refer the *out-of-distributional* dataset as the *reference dataset*. We learn features that can perform classification in both the known dataset and the reference dataset. If the *reference dataset* has $\mathcal{C}$ classes, once trained, the filter bank will contain positive filters of all $c+\mathcal{C}$ classes. Filters associated with the reference dataset will likely act as *negative filters* for all classes in the known dataset, thereby be globally negative. In this framework, the *globally negative filters* are likely to respond to arbitrary natural images provided that the reference dataset is a large-scale diverse dataset.
In Figure \[fig:acti2\], we show the impact of using the *globally negative filters*. Visualization of top activated filters for the Calculator class are shown at the top in Figure \[fig:acti2\](b). As can be seen from this figure, these filters are positively co-related with the Calculator class. With the new formulation, we observe that playing cards object activates some extra filters which are not in common with the calculator class (highlighted in red). At the bottom of Figure \[fig:acti2\](b) we visualize filters with the highest activation for the Playing Cards object. By inspection, these two visualizations look arbitrary and do not have an obvious association with any of the Caltech256 classes. We interpret these filters as instances of the *globally negative filters*. Due to the availability of more negative evidence, the overall activation value of the playing cards object has been drastically reduced.
Training Procedure
------------------
We propose a network architecture and a training mechanism to ensure that the network learns the *globally negative filters*. For this purpose, we use an external multi-class labeled dataset which we refer to as the *reference dataset*.
We first select a CNN backbone of choice (this could be a simple network such as Alexnet [@NIPS2012_ALEX] or a very deep/complex structure such as DenseNet [@huang2017densely]). Two parallel CNN networks of the selected backbone are used for training as shown in Figure \[fig:nw\](a). The only difference between the two parallel networks is the final fully-connected layer where the number of outputs is equal to the number of classes present in either dataset. For the purpose of our discussion, we refer the sub-network up to the penultimate layer of the CNN as the feature extraction sub-network.
Initially, weights of the two feature extraction sub-networks are initialized with identical weights and they are kept identical during training. Final layer of both parallel networks are initialized independently. Weights of these two layers are learned during training without having any dependency between each other. During training, two mini batches from two datasets (reference dataset (R) and known classes (T)) are considered and they are fed into the two branches independently. We calculate the cross-entropy loss $(L_{ce})$ with respect to the samples of the reference dataset and both the membership loss $(L_{m})$ and the cross-entropy loss with respect to the samples of known classes. The cumulative loss of the network then becomes a linear combination of the two losses as follows, $$\nonumber
Cumulative Loss = L_{ce}(R)+\alpha_1~L_{ce}(T)+\alpha_2~L_{m}(T).$$
In our experiments, we keep $\alpha_1, \alpha_2 = 1$. The cumulative loss is back-propagated to learn the weights of the two CNN branches. Reducing membership loss and cross-entropy loss with respect to the known-class dataset increases the potential of performing novelty detection in addition to classification as discussed in the preceding sub-section. On the other hand, having good performance (low cross-entropy loss) in the reference dataset suggests the existence of filters that are responsive to generic objects provided that the reference dataset is sufficiently diverse. When classes appearing in the reference dataset do not intersect with known classes, these filters serve as the *globally negative filters*.
![Proposed architecture for novelty detection. We use an external multi-class dataset (reference dataset (R)) in addition to the known object dataset (T). Two parallel CNN networks with identical structure and weights are used to extract features from both datasets. We train separate classifier networks operating on the same feature to perform classification in either dataset. During inference, novelty detection is performed by thresholding the maximal activation of the bottom branch of the network.[]{data-label="fig:nw"}](noveltydual){width="1\linewidth"}
Testing (Novelty Detection)
---------------------------
During inference, we propose to use the setup shown in Figure \[fig:nw\](b) where we only consider the bottom CNN branch of the training network. Given a test image $x$, we perform a forward pass using the learned CNN network to obtain the final feature $\textbf{f}(x)$. The largest element of $\textbf{f}(x)$, $\max \textbf{f}(x)$ is thresholded using a predetermined threshold $\gamma$ to arrive at the identity of the test image. If the yielded score is below the threshold $\gamma$, we identify the test sample to be novel. In a practical system, threshold $\gamma$ is chosen considering the percentile of the matched score distribution (for example threshold can be chosen to be 95th percentile if the accepted false negative rate is 5%) . In addition to the novelty detection procedure, the same network structure can be used to perform classification as well. Here, $\arg \max \textbf{f}(x)$ yields the identity of the predicted class for the test sample $x$. We note that this step is identical to the classification procedure used in the standard CNN-based classification.
Experimental Setup and Results
==============================
In this section, we present experimental results for the novelty detection task. We first describe the baseline methods used for comparison. Then, we introduce the four datasets used for evaluation. Finally, we discuss the obtained results followed by the analysis of the proposed method.
Baseline Methods
----------------
We evaluate the proposed method on four novelty detection databases and we compare its performance with the standard novelty detection schemes. We use the following baseline comparisons based on the AlexNet [@NIPS2012_ALEX] and the VGG16 [@DBLP:journals/corr/SimonyanZ14a] features fine-tuned on the given dataset.
**1. Finetune [@DBLP:journals/corr/SimonyanZ14a]:** $fc8$ feature scores of the trained deep model are thresholded to detect novel samples.\
**2. One-class SVM [@Scholkopf:2001:ESH:1119748.1119749]:** A one-class SVM classifier is trained for all known classes. The maximum SVM score is considered during the inference.\
**3. KNFST [@Bodesheim_2013_CVPR], [@Liu]:** Deep features are normalized and histogram intersection kernel method is used to generate inner products between the samples.\
**4. Local KNFST [@localnovelty]:** Deep features with histogram intersection kernel is considered with 600 local regions.\
**5. OpenMax [@BendaleB16]:** Activations of penultimate layer of a deep model are used to construct a single channel class-wise mean activation vectors (MAV) and the corresponding Weibull distributions.\
**6. K-extremes [@kextremes]:** Mean activations of the VGG16 $fc7$ features are considered for each class and top 0.1 activation indexes are binarized to arrive at the Extreme Value Signatures.\
**7. Finetune$(c+\mathcal{C})$:** A $(c+\mathcal{C})$ class CNN is trained by treating classes of the reference dataset as the additional class.\
In addition, we evaluate the performance based on the pretrained deep features (trained on the ILSVRC12 database) for KNFST and local KNFST methods. Whenever pre-trained features are use they are denoted by the suffix *pre*.
Datasets
--------
We use four publicly available multi-class datasets to evaluate the novelty detection performance of the proposed method.
![Sample images from the evaluation datasets. Each column contains images taken from a single class of each dataset.[]{data-label="fig:samples"}](samples "fig:"){width="1\linewidth"} -5pt
**Caltech256 Dataset.** The Caltech256 dataset is a fully annotated dataset which consists of 30607 images from 256 object classes. Following the protocol presented in [@Liu], we first sorted the class names alphabetically and picked the first 128 classes as the known classes and considered the images from the remaining 128 classes as the novel images.
**Caltech-UCSD Birds 200 (CUB 200) Dataset.** The CUB-200 dataset includes 6033 images belonging to 200 distinct bird categories. Ground truth labels for each image are provided. In our experiment, we sorted names of the bird categories alphabetically and used the first 100 classes as the known classes. The remaining classes were used to represent novel images.
**Stanford Dogs Dataset.** This dataset is a subset of the ImageNet dataset and was originally intended for fine-grain classification. There are 20580 images belonging to 120 different dog breeds in this dataset. We considered the first 60 classes as the known classes and treated the remaining classes as the novel classes during performance evaluation.
**FounderType-200 Dataset.** This dataset is a collection of Chinese character images in different font types. The dataset is organized based on the font-type. In total there are 200 different font-types with 6763 images from each class in this dataset. Following the same convention as before, we picked the first 100 classes to represent the enrolled classes. The remaining 100 classes were used to simulate the novel images.
In all datasets, following the protocol in [@Liu], images of the enrolled classes were randomly split into two even sets to form training and testing datasets of the enrolled classes. Images of the novel classes were used only during testing. When finetuning/extracting features from the caltech256 dataset following [@imagenet_cvpr09], we used the pretrained model trained on the Places365 dataset [@zhou2017places]. For all other tasks, we used the pretrained model trained on the ILSVRC12 dataset. Accordingly, the validation sets of Places365 was used as the reference dataset for Caltech256. For all other tasks the validation set of ILSVRC12 was considered.
Results {#sec:results}
-------
We evaluated all methods based on the VGG16 and the AlexNet features. We used the training codes provided by the authors when evaluating the KNFST [@Bodesheim_2013_CVPR] and the local KNFST [@localnovelty] methods. Performance of each method is evaluated using the area under the receiver operating characteristics (AUC) curve. Obtained AUC values for each method are tabulated in Table \[table:nov\] for all datasets[^2].
When baseline methods are considered, a variance in performance can be observed across datasets. In general, K-extremes has reported below-par performances compared to the other methods. When the number of enrolled classes are very high, the mean activation signature of a class looses its uniqueness. This is why K-extremes method fails when very large number of classes are enrolled as suggested in [@kextremes]. In the Caltech-256 and CUB-200 datasets, thresholding deep activations and OpenMax has yielded better results among the baseline methods. In Caltech256, this has improved marginally when the reference dataset (ILSVRC12) is incorporated. This method has performed reasonably well in the FounderType-200 dataset but it’s performance in the Standford Dogs dataset is not convincing. In general, KNFST has out-performed local KNFST except for in the Standford Dogs dataset. KNFST (and local KNFST) operating on the finetuned deep features have performed better in general compared to the pre-trained deep features. This trend has changed only in the Standford Dogs dataset. Here we note that none of the baseline methods have yielded consistent performance across datasets.
In comparison, the proposed method is able to produce the best performance across all datasets. When AlexNet is used as the back-bone network, there is an improvement of about 3.0% over the baselines in the CUB-200 and Standford Dogs datasets. In the other two datasets this margin is 2.0%. In the Caltech256, CUB-200 and FounderType-200 datasets, the improvements in AUC are in excess of 2.0% for the VGG16 model. In the Standford Dogs dataset, the proposed method is able to introduce a significant advancement of more than 7.0% in AUC compared with the baseline methods. In general, we note that in datasets where the baseline performance is already very good, as in the CUB-200 and FounderType 200 datasets, the improvement of the proposed method is relatively small. On the other hand, when the baseline performance is poor, the proposed method is able to generate a significant improvement in the performance.
Ablation Study
--------------
In this subsection, we investigate the impact of each individual component of the proposed framework. For the purpose of the ablation study, we use the validation dataset of the ILSVRC12 dataset as the reference dataset. It should be noted that figures reported in this subsection are different from Table \[table:nov\] due to this reason. Starting from the traditional CNN architecture, we added one component of the proposed framework at a time and evaluated the novelty detection performance on the Caltech-256 dataset as a case study. Testing protocol presented in the preceding subsection was followed in all cases. Considered cases are as follows.
**a) Single CNN with the cross-entropy loss (AUC 0.854).** This is the CNN baseline where a CNN is trained using the enrolled classes conventionally.
**b) Single CNN with the cross-entropy loss+membership loss (AUC 0.865).** The network architecture is the same as in case (a). In addition to the cross-entropy loss, the membership loss is calculated with respect to the enrolled dataset.
**c) Two Parallel CNNs with cross-entropy loss (AUC 0.864).** The network structure proposed in Figure \[fig:nw\](a) is used. In contrast, only the cross-entropy loss is used in the bottom sub-network.
**d) Proposed method (AUC 0.906).** Proposed structure Figure \[fig:nw\](a) is used for training.
In the proposed method, we introduced membership loss and a parallel network structure as contributions. From the case study conducted, it appears that the novelty detection performance improves compared to the baseline even when one of the contributions are used. Moreover, we observe that the two contributions compliment each other and generate even better results when combined together.
Impact of the Reference Dataset
-------------------------------
In the proposed method, we assumed the availability of a reference dataset with large number of classes. In this subsection, we investigate the impact of the reference dataset by varying the reference dataset of choice. In particular, we use the ILSVRC12, Caltech-256 and Standford Dogs datasets as the reference datasets to perform novelty detection using the proposed method in the CUB-200 dataset. Results obtained are tabulated in Table \[table:ref\]. Here we have included the performance of the best baseline method for the CUB-200 dataset (Finetune) from Table \[table:nov\] as a baseline.
Compared to ILSVRC12, when Caltech-256 is used as the reference dataset, AUC drops by 0.005%. This further drops by 0.008% when the Standford Dogs dataset is used. The ILSVRC12 dataset contains 1000 image classes and has significant variance in images within each class. Caltech-256 is a similar multi-class dataset but with fewer classes. Both of these datasets contain natural images. However since ILSVRC12 has more classes and more intra-class variance, we expect it to generate *globally negative filters* better. Therefore, the performance drop of Caltech-256 compared to ILSVRC12 is expected. On the other hand, the Standford Dogs dataset only contains images of dogs. Therefore, filters learned using this dataset may not be generic to get stimulated by the arbitrary inputs. Therefore, the drop in the performance is justified. In conclusion, we note that the proposed method is able to out-perform baseline novelty detection methods even when the reference dataset is varied. However, better results are obtained when a larger dataset with high degree of intra-class variation is used as the reference dataset.
Impact on Classification Accuracy
---------------------------------
When a test image is present, the proposed method produces a set of class activation scores. It is still possible to perform classification using the same system by associating the test image with the class containing the highest activation. In what follows, we consider test samples of the known classes and perform closed-set classification in the same experimental setup described in Section \[sec:results\]. In other words, we do not consider novel samples for the purpose of this study. Obtained classification accuracies for the four datasets are tabulated in Table \[table:class\]. Although the proposed method is designed for the purpose of novelty detection, we note that the proposed changes have contributed towards increasing the classification accuracy of the system as well. This is because the *membership loss* explicitly enforces correct class to have a high score and all other classes to have scores closer to zero.
Conclusion
==========
We presented an end-to-end deep learning-based solution for image novelty detection. We build up on the conventional classification networks and introduce two novel contributions; namely, *membership loss* and a training procedure that produces *globally negative filters*. In the proposed method, novelty is quarried simply by thresholding the highest activation of the output vector. We demonstrate the effectiveness of the proposed method on four publicly available multi-class image datasets and obtain state-of-the-art results.
[^1]: This work was supported by the NSF grant 1801435.
[^2]: Source code of the proposed method is made available at https://github.com/PramuPerera/TransferLearningNovelty
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'As Computer Vision algorithms move from passive analysis of pixels to active reasoning over semantics, the breadth of information algorithms need to reason over has expanded significantly. One of the key challenges in this vein is the ability to identify the information required to make a decision, and select an action that will recover this information. We propose an reinforcement-learning approach that maintains an distribution over its internal information, thus explicitly representing the ambiguity in what it knows, and needs to know, towards achieving its goal. Potential actions are then generated according to particles sampled from this distribution. For each potential action a distribution of the expected answers is calculated, and the value of the information gained is obtained, as compared to the existing internal information. We demonstrate this approach applied to two vision-language problems that have attracted significant recent interest, visual dialog and visual query generation. In both cases the method actively selects actions that will best reduce its internal uncertainty, and outperforms its competitors in achieving the goal of the challenge.'
author:
- |
Ehsan Abbasnejad$^1$, Qi Wu$^1$, Iman Abbasnejad$^2$, Javen Shi$^1$, Anton van den Hengel$^1$\
$^1$`{ehsan.abbasnejad,qi.wu01,javen.shi,anton.vandenhengel}@adelaide.edu.au`\
$^2$``\
$^1$Australian Institute of Machine Learning & The University of Adelaide, Australia\
$^2$Queensland University of Technology, Australia
bibliography:
- 'lib.bib'
title: |
An Active Information Seeking Model\
for Goal-oriented Vision-and-Language Tasks
---
Introduction
============
In most problems in Computer Vision it is assumed that all of the information required is available a-priori, and suitable to be embodied in the code or the weights of the solution. This assumption is so pervasive that it typically goes unsaid. In fact this assumption is satisfied by a small subset of problems of practical interest. Problems in this set must be self-contained, tightly specified, relate to a very prescribed form of data drawn from a static distribution, and be completely predictable. Many important problems do not satisfy these criteria, even though researchers have found many that do.
The majority of problems that computer vision might be applied to are solvable only by an agent that is capable of actively seeking the information it needs. This might be because the information required is not available at training time, or because it is too broad to be embodied in the code or weights of an algorithm. The ability to seek the information required to complete a task enables a degree of flexibility and robustness that cannot be achieved through other means, but also enables behaviors that lie towards the Artificial Intelligence end of the spectrum.
![ Two goal-oriented vision-and-language tasks, broken up into four constituent parts: a context encoder, an information seeker, a answerer and a goal executor. The given examples are chosen from a goal-oriented visual dialog dataset GuessWhat [@guesswhat_game] (upper from the red dash-line) and, a compositional VQA dataset CLEVR [@clevr] (lower). ](img/fig1_new.pdf){width="1\columnwidth"}
\[fig:intro\]
Applications that lie at the intersection of vision and language are examples of such problems. They are more challenging than conventional computer vision problems because they often require an agent (model) to acquire information on the fly to help to make decisions, such as visual dialog [@das2016visual; @visdial_rl] and visual question answering [@answer_questioner_mind; @guesswhat_game; @strub2017end]. More recently, a range of tasks have been proposed that use ‘language generation’ as a mechanism to gather more information in order to achieve another specific goal. These tasks offer a particular challenge because the information involved is inevitably very broad, which makes concrete representations difficult to employ practically.
In visual dialog, and particularly goal-oriented visual question generation, an agent needs to understand the user request and complete a task via asking a limited number of questions. Similarly, compositional VQA (e.g. [@clevr]) is a visual query generation problem that requires a model first to convert a natural language question to a sequence of ‘programs’ and then obtain the answer by running the programs on an engine. The question-to-program model represents an information ‘seeker’, while the broader goal is to generate an answer based on the information acquired.
Agents applicable to these tasks normally consist of three parts: a *context encoder*, an *information seeker* and a *goal executor*, as shown in Fig.\[fig:intro\]. The context encoder is responsible for encoding information such as images, questions, or dialog history to a feature vector. The information seeker is a model that is able to generate new queries (such as natural language questions and programs) based on the goal of the given task and its seeking strategy . The information returned will then join the context and internal information to be sent to the goal executor model to achieve the goal. The seeker model plays a crucial role in goal-oriented vision-and-language tasks as the better seeking strategies that recovers more information, the more likely it is that the goal can be achieved. Moreover, the seeker’s knowledge of the value of additional information is essential in directing the seeker towards querying what is needed to achieve the goal. In this paper, we focus on exploring the seeker model.
The conventional ‘seeker’ models in these tasks follow a sequence-to-sequence generation architecture, that is, they translate an image to a question or translate a question to a program sequence via supervised learning. This requires large numbers of ground-truth training pairs. Reinforcement learning (RL) is thus employed in such goal-oriented vision-language to mediate this problem due to its ability to focus on achieving a goal through directed trial and error [@guesswhat_game; @zhang2018goal]. A policy in RL models specifies how the seeker asks for additional information. However, these methods generally suffer from two major drawbacks: (1) they maintain a single policy that translates the input sequence to the output while disregarding the strategic diversity needed. Intuitively a single policy is not enough in querying diverse information content for various goals–we need multiple strategies. In addition, (2) the RL employed in these approaches can be prohibitively inefficient since the question generation process does not consider its effect in directing the agent towards the goal. In fact, the agent does not have a notion of what information it needs and how it benefits in achieving its goal.
To this end, in contrast to conventional methods that use a single policy to model a vision-and-language task, we instead maintain a ***distribution of policies***. By employing a Bayesian reinforcement learning framework for learning this distribution of the seeker’s policy, our model incorporates the expected gain from a query towards achieving its goal. Our framework uses recently proposed Stein Variational Gradient Descent [@svgd] to perform an efficient update of the posterior policies. Having a distribution over seeking policies, our agent is ***capable of considering various strategies for obtaining further information***, analogous to human contemplation of various ways to ask a question. Each sample from the seeker’s policy posterior represents a policy of its own, and seeks a different piece of information. This allows the agent to further contemplate the outcome of the various strategies before seeking additional information and considers the consequence towards the goal. We then formalize an approach for the agent to ***evaluate the consequence of receiving additional information*** towards achieving its goal.
We apply the proposed approach to two complex vision-and-language tasks, namely GuessWhat [@guesswhat_game] and CLEVR [@clevr], and show that it outperforms the baselines and achieves the state-of-art results.
Conclusion
==========
The ability to identify the information needed to support a conclusion, and the actions required to obtain it, is a critical capability if agents are to move beyond carrying out low-level prescribed tasks towards achieving flexible high semantic level goals. The method we describe is capable of reasoning about the information it holds, and the information it will need to achieve its goal, in order to identify the action that will best enable it to fill the gap between the two. Our approach thus actively seeks the information it needs to achieve its goal on the basis of a model of the uncertainty in its own understanding. If we are to enable agents that actively work towards a high-level goal the capability our approach demonstrates will be critical.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Low-rank quaternion tensor completion method, a novel approach to recovery color videos and images is proposed in this paper. We respectively reconstruct a color image and a color video as a quaternion matrix (second-order tensor) and a third-order quaternion tensor by encoding the red, green, and blue channel pixel values on the three imaginary parts of a quaternion. Different from some traditional models which treat color pixel as a scalar and represent color channels separately, whereas, during the quaternion-based reconstruction, it is significant that the inherent color structures of color images and color videos can be completely preserved. Under the definition of Tucker rank, the global low-rank prior to quaternion tensor is encoded as the nuclear norm of unfolding quaternion matrices. Then, by applying the ADMM framework, we provide the tensor completion algorithm for any order $(\geq 2)$ quaternion tensors, which theoretically can be well used to recover missing entries of any multidimensional data with color structures. Simulation results for color videos and color images recovery show the superior performance and efficiency of the proposed method over some state-of-the-art existing ones.'
address:
- 'Department of Mathematics, Faculty of Science and Technology, University of Macau, Macau 999078, China'
- 'School of Mathematics and Quantitative Economics, Shandong University of Finance and Economics, Shandong, China'
author:
- Jifei Miao
- Kit Ian Kou
- Wankai Liu
bibliography:
- 'refs.bib'
title: 'Low-rank quaternion tensor completion for recovering color videos and images'
---
[1.95]{}
Quaternion ,color videos ,color images ,tensor completion ,low-rank.
Introduction {#section_1}
============
A color video or color image contains red, blue, and green channels. In most cases, some data of the acquired videos or images are missed during acquisition and transmission, which poses great challenges to further processing of them. Hence, a well-performed recovery technology is important and necessary to recover complete videos or images from their incomplete observations. The core of the missing value estimation lies on how to exactly build a proper low-rank regularizer to measure the global structure of the underlying video or image data according to the fact that the data inherently possess a low-rank structure [@DBLP:journals/pr/LiWLT19].
In the past few decades, the low-rank matrix completion problem has been widely studied and proven very useful in the application of images and even videos recovery [@DBLP:journals/pr/FanC17a; @DBLP:journals/pr/FanC18]. Commonly, the method is to stack all the image or video pixels as column vectors of a matrix, and recovery theories and algorithms are adopted to the resulting matrix which is low-rank or approximately low-rank. However, these image and video recovery models are usually developed for grey-level pixels. For color videos and color images processing, traditional matrix-based methods usually ignore the mutual connection among channels, because these recovery methods are applied to red, green, and blue channels separately, which is likely to result in color distortion during the recovery process [@DBLP:journals/tip/ZouKW16].
On the other hand, with the success of low-rank matrix completion, low-rank tensor completion is an extension to process the multidimensional data [@DBLP:journals/pr/QinJHLLZDZF19; @DBLP:journals/tip/ZhouLLZ18; @DBLP:journals/sigpro/LongLCZ19]. A color image with red, blue, and green channels can be naturally regarded as a third-order tensor. Each frontal slice of this third-order tensor corresponds to a channel of the color image. Analogously, a video comprised of color images is a fourth-order tensor with an additional index for a temporal variable [@DBLP:journals/sigpro/LongLCZ19]. Nevertheless, there are still some underlying restrictions on tensor-based completion algorithms, especially for color video recovery problems. For example there are plenty of completion algorithms using tensor singular value decomposition (t-SVD) and the tubal rank [@6909886TNN; @DBLP:journals/tip/ZhouLLZ18], *etc.*, can not be well applied to color video recovery problem, since the t-SVD and the tubal rank theories [@DBLP:journals/siammax/KilmerBHH13] they are based on are defined for third-order tensors. Therefore, for grey scalar videos (third-order tensors), they can obtain well performance, however, for color videos (fourth-order tensors), they may ignore the inherent color structures, and then can not offer a satisfying recovery result. In addition, there are some factorization based approaches for tensor completion, such as CANDECOMP/PARAFAC (CP) and Tucker factorizations based approaches [@DBLP:journals/pami/ZhaoZC15; @DBLP:journals/tsp/YokotaZC16], tensor unfolding based approaches [@DBLP:journals/pami/LiuMWY13; @1930-8337_2015_2_601], *etc.*. These methods can deal with color videos directly, however, the factorization or matricization operation may destroy color pixel structure and lead to color distortion during the recovery process [@DBLP:journals/pr/YuWGGXP19].
Different from conventional matrix and tensor based models, in this paper, we make use of quaternion tensors to represent color videos and color images[^1], and study the problem of quaternion tensors completion to estimate missing data of them. Actually, the quaternion has achieved excellent results in color image processing problems including histopathological image analysis [@DBLP:journals/pr/ShiZWGZY19], color image denoising and representation [@DBLP:journals/mssp/GaiYW015; @DBLP:journals/pr/HosnyD19], color object recognition [@DBLP:journals/pr/LiLHS15; @DBLP:journals/pr/ShaoSWCC14], and so on. The quaternion based method is to encode the red, green, and blue channel pixel values on the three imaginary parts of a quaternion [@DBLP:journals/pr/LiLHS15]. That is $$\label{equ1}
\dot{t}=0+t_{r}i+t_{g}j+t_{b}k,$$ where $\dot{t}$ denotes a color pixel, $t_{r}$, $t_{g}$ and $t_{b}$ are, respectively, the red, green and blue channel pixel values, and $i$, $j$ and $k$ are the three imaginary units[^2]. The graphical of a pure quaternion representing a color pixel can be seen in Figure.\[quaternion\_fig\].
![The graphical of a pure quaternion representing a color pixel. $i$, $j$ and $k$ correspond to three channels ($R$, $G$ and $B$) of the color pixel.[]{data-label="quaternion_fig"}](quaternion){width="3cm" height="2.2cm"}
By using (\[equ1\]), an $M\times N\times3$ color image, and an $M\times N\times3\times T$ (image row $\times$ image column $\times$ RGB $\times$ frame) color video are respectively described by a second-order quaternion tensor (quaternion matrix) with size $M\times N$ and a third-order quaternion tensor with size $M\times N \times T$ whose entries are pure quaternions. The main advantage of this quaternion representation is that it processes a color pixel holistically as a vector field and handles the coupling between the color channels naturally [@DBLP:journals/pr/HosnyD19], and color information of source video and image is fully utilized.
However, the existing quaternion based methods mainly consider the color image issues but not consider the higher dimensional data structures, for example, color videos. In this paper, we reconstruct an $M\times N\times3\times T$ color video (fourth-order tensor) as an $M\times N \times T$ third-order quaternion tensor with each frontal slice being a quaternion matrix. It is important to highlight that different from traditional tensor model, the three color channels of each pixel in quaternion tensor can be fully connected by the model (\[equ1\]), and it is clear that even if the matricization operations (*e.g.*, tensor unfolding) can not destroy the color pixel structure, *i.e.,* the relative positions of the three color channel pixels of one pixel will remain unchanged just as Figure.\[quaternion\_fig\]. Hence, Under the definition of Tucker rank, the global low-rank prior to quaternion tensor is encoded as the nuclear norm of unfolding quaternion matrices. Then, we provide the completion algorithm for any order $(\geq 2)$ quaternion tensors by applying the alternating direction method of multipliers (ADMM) [@DBLP:journals/ftml/BoydPCPE11] framework. Simulation results for color videos and color images recovery show that the performance of the proposed method is better than that of the testing methods.
The rest of this paper is organized as follows. Section \[section\_2\] introduces some notations and preliminaries for quaternion algebra and quaternion tensor. Section \[sec3\] reviews the tensor completion theory and proposes our quaternion-based tensor completion model. The detailed overview of the quaternion tensor completion method is presented in Section \[sec4\]. Section \[sec5\] provides some simulations to illustrate the performance of our approach, and compare it with some state-of-the-art methods. Finally, some conclusions are drawn in Section \[sec6\].
Notations and preliminaries {#section_2}
===========================
In this section, we first summarize some main notations and then introduce some basic knowledge of quaternion algebra and quaternion tensor.
Notations and definitions
-------------------------
In this paper, $\mathbb{R}$, $\mathbb{C}$, and $\mathbb{H}$ respectively denote the set of real numbers, the set of complex numbers and the set of quaternions. A scalar, a vector, a matrix, and a tensor are written as $a$, $\mathbf{a}$, $\mathbf{A}$, and $\mathcal{A}$, respectively. For a tensor $\mathcal{A}$, we use the Matlab notation $\mathcal{A}(:,:,k)$ to denote its $k$-th frontal slice and the $\mathbf{A}_{[k]}$ to denote its mode-k unfolding. A dot (above the variable) is used to denote a quaternion variable (*e.g.,* [@DBLP:journals/tip/ZouKW16; @DBLP:journals/tip/XuYXZN15]), $\dot{a}$, $\dot{\mathbf{a}}$, $\dot{\mathbf{A}}$ and $\dot{\mathcal{A}}$ respectively represent a quaternion scalar, a quaternion vector, a quaternion matrix and a quaternion tensor. $(\cdot)^{\ast}$ and $(\cdot)^{H}$ denote the conjugation and conjugate transpose. $|\cdot|$, $\|\cdot\|_{F}$ and $\|\cdot\|_{\ast}$ are respectively the moduli, the Frobenius norm, and the nuclear norm. ${\rm{tr}}\{\cdot\}$, ${\rm{rank}}(\cdot)$ and $\nabla_{sub}$ denote the trace, rank and subgradient operators respectively. ${\rm{Unfold}}_{k}$ is the Mode-k unfolding operator of tensors, and we use ${\rm{Fold}}_{k}$ to denote the inverse operator of ${\rm{Unfold}}_{k}$.
Basic knowledge of quaternion algebra
-------------------------------------
Quaternions were discovered in 1843 by W.R. Hamilton [@doi:10.1080/14786444408644923][^3]. A quaternion $\dot{q}\in\mathbb{H}$ is a four-dimensional (4D) hypercomplex number and has a Cartesian form given by: $$\label{equ2}
\dot{q}=q_{0}+q_{1}i+q_{2}j+q_{3}k,$$ where $q_{l}\in\mathbb{R}\: (l=0,1,2,3)$ are called its components, and $i, j, k$ are square roots of -1 and are related through the famous relations: $$\begin{aligned}
\left\{
\begin{array}{lc}
i^{2}=j^{2}=k^{2}=ijk=-1,\\
ij=-ji=k,
jk=-kj=i,
ki=-ik=j.
\end{array}
\right.\end{aligned}$$ A quaternion $\dot{q}\in\mathbb{H}$ can be decomposed into a real part R$(\dot{q})$ and an imaginary part I$(\dot{q})$: $$\label{equ3}
\dot{q}={\rm{R}}(\dot{q})+{\rm{I}}(\dot{q}),$$ where ${\rm{R}}(\dot{q})=q_{0}$, ${\rm{I}}(\dot{q})=q_{1}i+q_{2}j+q_{3}k$. Then, $\dot{q}\in\mathbb{H}$ will be called a pure quaternion if its real part is null, *i.e.*, if ${\rm{R}}(\dot{q})=0$. Given two quaternions $\dot{p}$ and $\dot{q}\in\mathbb{H}$, the sum and multiplication of them are respectively: $$\label{equ4}
\dot{p}+\dot{q}=(p_{0}+q_{0})+(p_{1}+q_{1})i+(p_{2}+q_{2})j+(p_{3}+q_{3})k$$ and $$\begin{aligned}
\label{equ5}
\dot{p}\dot{q}=&(p_{0}q_{0}-p_{1}q_{1}-p_{2}q_{2}-p_{3}q_{3})
+(p_{0}q_{1}+p_{1}q_{0}+p_{2}q_{3}-p_{3}q_{2})i\\
&+(p_{0}q_{2}-p_{1}q_{3}+p_{2}q_{0}+p_{3}q_{1})j
+(p_{0}q_{3}+p_{1}q_{2}-p_{2}q_{1}+p_{3}q_{0})k.\end{aligned}$$ It is noticeable that the multiplication of two quaternions is not commutative so that in general $\dot{p}\dot{q}\neq\dot{q}\dot{p}$.
The conjugate and the modulus of a quaternion $\dot{q}\in\mathbb{H}$ are, respectively, defined as follows: $$\begin{aligned}
\dot{q}^{\ast}=q_{0}-q_{1}i-q_{2}j-q_{3}k,\\
|\dot{q}|=\sqrt{q_{0}^{2}+q_{1}^{2}+q_{2}^{2}+q_{3}^{2}}.\end{aligned}$$ Every quaternion $\dot{q}\in\mathbb{H}$ can be uniquely represented as the Cayley–Dickson (CD) form: $$\label{equ6}
\dot{q}=z_{1}+z_{2}j,$$ where $z_{1}=q_{0}+q_{1}i$ and $z_{2}=q_{2}+q_{3}i$ are complex numbers.
Quaternion matrix and tensor
----------------------------
The quaternion matrix is denoted as $\dot{\mathbf{Q}}=(\dot{q}_{n_{1},n_{2}})\in\mathbb{H}^{N_{1}\times N_{2}}$, $1\leq n_{1}\leq N_{1}$, $1\leq n_{2}\leq N_{2}$, where each entry is a quaternion [@ZHANG199721]. We often rewritten it as $$\dot{\mathbf{Q}}=\mathbf{Q}_{0}+\mathbf{Q}_{1}i+\mathbf{Q}_{2}j+\mathbf{Q}_{3}k,$$ where $\mathbf{Q}_{l}\in\mathbb{R}^{N_{1}\times N_{2}}\: (l=0,1,2,3)$, $\dot{\mathbf{Q}}$ is named a pure quaternion matrix when $\mathbf{Q}_{0}=\mathbf{0}$. In addition to scalar representations, based on the CD form (\[equ6\]), there exists their isomorphic complex representation denoted as $f(\dot{\mathbf{Q}})\in\mathbb{C}^{2N_{1}\times 2N_{2}}$, is of the form: $$\label{definef}
f(\dot{\mathbf{Q}})=\left(\begin{array}{cc}
\mathbf{Z}_{1} & \mathbf{Z}_{2}\\
-\mathbf{Z}^{\ast}_{2} & \mathbf{Z}^{\ast}_{1}
\end{array}\right),$$ where $\mathbf{Z}_{1}=\mathbf{Q}_{0}+\mathbf{Q}_{1}i$, $\mathbf{Z}_{2}=\mathbf{Q}_{2}+\mathbf{Q}_{3}i\in\mathbb{C}^{N_{1}\times N_{2}}$.
Note that the multiplication between quaternion matrices can be defined similar to classical multiplication between real or complex matrices, except that the multiplication between two quaternion numbers is employed.
(The rank of quaternion matrix [@ZHANG199721]) The maximum number of right (left) linearly independent columns (rows) of a quaternion matrix $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$ is called the rank of $\dot{\mathbf{Q}}$.
(Quaternion singular value decomposition (QSVD) [@ZHANG199721]) Let $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$ be of rank r. Then, there exist two unitary quaternion matrices[^4] $\dot{\mathbf{U}}\in\mathbb{H}^{N_{1}\times N_{1}}$ and $\dot{\mathbf{V}}\in\mathbb{H}^{N_{2}\times N_{2}}$ such that $$\dot{\mathbf{U}}^{H}\dot{\mathbf{Q}}\dot{\mathbf{V}}=\mathbf{\Lambda}=\left(\begin{array}{cc}
\mathbf{\Sigma}_{r} & \mathbf{0}\\
\mathbf{0} & \mathbf{0}
\end{array}\right),$$ where $\mathbf{\Sigma}_{r}={\rm{diag}}(\sigma_{1},\ldots,\sigma_{r})$ is a real diagonal matrix and has $r$ positive entries $\sigma_{k}, \,(k=1,\ldots,r)$ on its diagonal (*i.e.*, positive singular values of $\dot{\mathbf{Q}}$).
The relation between the QSVD of quaternion matrix $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$ and the SVD of its isomorphic complex matrix $f(\dot{\mathbf{Q}})\in\mathbb{C}^{2N_{1}\times 2N_{2}}$ ($f(\dot{\mathbf{Q}})=\mathbf{U}\check{\mathbf{\Lambda}}\mathbf{V}^{H}$) is defined as [@DBLP:journals/tip/XuYXZN15]: $$\begin{aligned}
\label{qsvd}
\left\{
\begin{array}{lc}
\mathbf{\Lambda}={\rm{row}}_{odd}({\rm{col}}_{odd}(\check{\mathbf{\Lambda}})),\\
\dot{\mathbf{U}}={\rm{col}}_{odd}(\mathbf{U}_{1})+{\rm{col}}_{odd}(-(\mathbf{U}_{2})^{\ast})j,\\
\dot{\mathbf{V}}={\rm{col}}_{odd}(\mathbf{V}_{1})+{\rm{col}}_{odd}(-(\mathbf{V}_{2})^{\ast})j,
\end{array}
\right.\end{aligned}$$ such that $\dot{\mathbf{Q}}=\dot{\mathbf{U}}\mathbf{\Lambda}\dot{\mathbf{V}}^{H}$, where $$\begin{aligned}
\mathbf{U}=\left(\begin{array}{c}
(\mathbf{U}_{1})_{N_{1}\times 2N_{1}} \\
(\mathbf{U}_{2})_{N_{1}\times 2N_{1}}
\end{array} \right),\quad
\mathbf{V}=\left(\begin{array}{c}
(\mathbf{V}_{1})_{N_{2}\times 2N_{2}} \\
(\mathbf{V}_{2})_{N_{2}\times 2N_{2}}
\end{array} \right),\end{aligned}$$ and ${\rm{row}}_{odd}(\mathbf{M})$, ${\rm{col}}_{odd}(\mathbf{M})$ respectively extract the odd rows and odd columns of matrix $\mathbf{M}$. Based on the QSVD, we define the quaternion matrix nuclear norm (QMNN) below.
(QMNN) Given a quaternion matrix $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$, the QMNN of it is defined as $$\label{qmnn}
\|\dot{\mathbf{Q}}\|_{\ast}=\sum_{k=1}^{{\rm{min}}\{N_{1},N_{2}\}}\sigma_{k},$$ where $\sigma_{k}$ is the singular value of $\dot{\mathbf{Q}}$, which can be obtained by the QSVD of $\dot{\mathbf{Q}}$.
In addition, the Frobenius norm of the quaternion matrix $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$ is defined as [@ZHANG199721]: $\|\dot{\mathbf{Q}}\|_{F}=\sqrt{\sum_{n_{1}=1}^{N_{1}}\sum_{n_{2}=1}^{N_{2}}|\dot{q}_{n_{1},n_{2}}|^{2}}=\sqrt{{\rm{tr}}\{(\dot{\mathbf{Q}})^{H}\dot{\mathbf{Q}}\}}$.
Analogously, in this paper, we generalize the definition of quaternion matrix to higher dimensional quaternion array, *i.e.*, quaternion tensor.
(Quaternion tensor) A multidimensional array or an Nth-order tensor is called a quaternion tensor if its entries are quaternion numbers, *i.e.*, $$\begin{aligned}
\label{qtensor1}
\dot{\mathcal{T}}&=(\dot{t}_{n_{1},n_{2},\ldots,n_{N}})\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}} \nonumber\\
&=\mathcal{T}_{0}+\mathcal{T}_{1}i+\mathcal{T}_{2}j+\mathcal{T}_{3}k,\end{aligned}$$ where $\mathcal{T}_{l}\in\mathbb{R}^{N_{1}\times N_{2} \times\ldots \times N_{N}}\: (l=0,1,2,3)$, $\dot{\mathcal{T}}$ is named a pure quaternion tensor when $\mathcal{T}_{0}$ is a zero tensor.
(Mode-k unfolding) For an Nth-order quaternion tensor $\dot{\mathcal{T}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$, its mode-k unfolding is defined as a quaternion matrix $${\rm{Unfold}}_{k}(\dot{\mathcal{T}})=\dot{\mathbf{T}}_{[k]}\in\mathbb{H}^{N_{k}\times N_{1}\ldots N_{k-1}N_{k+1}\ldots N_{N}}\quad \text{with entries}$$ $$\dot{\mathbf{T}}_{[k]}(n_{k}, n_{1}\ldots n_{k-1}n_{k+1}\ldots n_{N})=\dot{t}_{n_{1},n_{2},\ldots,n_{N}},$$ where $\dot{t}_{n_{1},n_{2},\ldots,n_{N}}$ is the $(n_{1},n_{2},\ldots,n_{N})$th entry of $\dot{\mathcal{T}}$.
(Tucker rank [@DBLP:journals/siamrev/KoldaB09]) Given a quaternion tensor $\dot{\mathcal{T}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$, the Tucker rank of it is defined as $$\label{Tucker_rank}
{\rm{rank_{tucker}}}(\dot{\mathcal{T}})=({\rm{rank}}(\dot{\mathbf{T}}_{[1]}),{\rm{rank}}(\dot{\mathbf{T}}_{[2]}),\ldots,{\rm{rank}}(\dot{\mathbf{T}}_{[N]})),$$ where ${\rm{rank}}(\dot{\mathbf{T}}_{[k]})$ denotes the rank of the mode-k unfolding quaternion matrix $\dot{\mathbf{T}}_{[k]}$.
Problem formulation {#sec3}
===================
In this section, we first review the tensor completion theory and then propose our quaternion based tensor completion model.
Tensor completion theory
------------------------
The tensor completion problem consists of recovering a tensor from a subset of its entries. The key is to build up the relationship between the available and the missing entries [@DBLP:journals/pami/LiuMWY13]. The usual structural assumption on a tensor that makes the problem well-posed is that the tensor is low-rank or approximate low-rank. Mathematically, the optimization model for tensor completion problem can be formulated as: $$\label{equ7}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\mathcal{T}}\quad {\rm{rank}}(\mathcal{T})\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\mathcal{T})=\mathcal{Y},
\end{split}$$ where $\mathcal{Y}$ is the underlying complete tensor, $\mathcal{T}$ is the observed tensor, and $\mathcal{P}_{\Omega}$ denotes the random sampling operator which is defined by:
$$\mathcal{P}_{\Omega}(\mathcal{T})=\left\{
\begin{array}{lc}
t_{n_{1},n_{2},\ldots,n_{N}},\qquad &(n_{1},n_{2},\ldots,n_{N})\in \Omega, \\
0, &\text{otherwise}.
\end{array}
\right.$$
However, there is no unique definition for the rank of tensors, such as CP rank [@doi:10.1137/0614071], Tucker rank [@DBLP:journals/pami/LiuMWY13], tubal rank [@DBLP:journals/tip/ZhouLLZ18], tensor train rank [@DBLP:journals/tip/BenguaPTD17], *etc.*. With different definitions of tensor rank, there are many methods optimization models for tensor completion problem. Among all definitions for the rank of tensors, the Tucker rank is widely used to depict the low-rankness of the underlying tensor. based on minimizing Tucker rank, (\[equ7\]) can be formulated as: $$\label{equ8}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\mathcal{T}}\quad {\rm{rank_{tucker}}}(\mathcal{T})\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\mathcal{T})=\mathcal{Y},
\end{split}$$ According to the definition of Tucker rank, (\[equ8\]) can be written as [@DBLP:journals/pami/LiuMWY13; @DBLP:journals/ijon/TanCWZR14]: $$\label{equ9}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\mathbf{T}_{[n]}}\quad \sum_{n=1}^{N}\alpha_{n}{\rm{rank}}(\mathbf{T}_{[n]})\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\mathcal{T})=\mathcal{Y},
\end{split}$$ where $\alpha_{n}$ are nonnegative constants. However, directly optimizing the problem (\[equ9\]) is NP-hard [@DBLP:journals/siammax/GillisG11]. Inspired by matrix nuclear norm, the tightest convex surrogate of the matrix rank, Liu et al. [@DBLP:journals/pami/LiuMWY13] established the following definition of the nuclear norm for tensors: $$\label{equ10}
\|\mathcal{T}\|_{\ast}=\sum_{n=1}^{N}\alpha_{n}\|\mathbf{T}_{[n]}\|_{\ast}.$$ Then, based on (\[equ10\]), the problem (\[equ9\]) can be finally rewritten as: $$\label{equ11}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\mathbf{T}_{[n]}}\quad \sum_{n=1}^{N}\alpha_{n}\|\mathbf{T}_{[n]}\|_{\ast}\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\mathcal{T})=\mathcal{Y}.
\end{split}$$
Proposed formulation of quaternion tensor completion
----------------------------------------------------
Quaternion tensor completion can be regarded as the generalization of the traditional tensor completion problem in the quaternion number field, which is to estimate the missing values of a quaternion tensor $\dot{\mathcal{T}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$ under a given subset $\Omega$ of its entries $\{\dot{\mathcal{T}}_{n_{1},n_{2},\ldots,n_{N}}|(n_{1},n_{2},\ldots,n_{N})\in\Omega \}$. That is $$\label{equ12}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\dot{\mathcal{T}}}\quad {\rm{rank}}(\dot{\mathcal{T}})\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\dot{\mathcal{T}})=\dot{\mathcal{Y}},
\end{split}$$ where $\dot{\mathcal{Y}}$ is the underlying complete quaternion tensor, $\dot{\mathcal{T}}$ is the observed quaternion tensor.
Based on the previous **Definition 3**, **Definition 4** and **Definition 5**, and followed by traditional tensor case, we finally translate the problem (\[equ12\]) into the following low-rank quaternion tensor completion formulation: $$\label{equ13}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\dot{\mathbf{T}}_{[n]}}\quad \sum_{n=1}^{N}\alpha_{n}\|\dot{\mathbf{T}}_{[n]}\|_{\ast}\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\dot{\mathcal{T}})=\dot{\mathcal{Y}}.
\end{split}$$ The formulation (\[equ13\]) can be well used to recover missing entries of any multidimensional data with color structures. For special cases, *i.e.*, $N=3$ and $N=2$, we can deal with color videos and color images recovery problems. It is important to notice that in these kinds of applications the formulation (\[equ13\]) outperforms that of (\[equ11\]), because for traditional tensor the Mode-k unfolding operation may destroy the color pixel structure, but for quaternion tensor, the inherent color structures can be completely preserved during this process.
Proposed algorithm {#sec4}
==================
In this section, we show how to solve the optimization problem (\[equ13\]), then we provide complexity analyses of the proposed method.
Optimization Procedure
----------------------
The optimization problem (\[equ13\]) can be solved by various methods. For efficiency, we adopt the ADMM framework in this paper, which can support the convergence of the algorithm [@DBLP:journals/ftml/BoydPCPE11]. By using additional quaternion tensor $\dot{\mathcal{X}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$ with $N$ Mode-k unfolding quaternion matrices $\dot{\mathbf{X}}_{[1]},\dot{\mathbf{X}}_{[2]},\ldots,\dot{\mathbf{X}}_{[N]}$, we first convert (\[equ13\]) to the following equivalent problem: $$\label{equ14}
\begin{split}
\mathop{{\rm{minimize}}}\limits_{\dot{\mathcal{T}},\dot{\mathbf{X}}_{[n]}}\quad &\sum_{n=1}^{N}\alpha_{n}\|\dot{\mathbf{X}}_{[n]}\|_{\ast}\\
\text{subject to} \quad &\dot{\mathbf{T}}_{[n]}=\dot{\mathbf{X}}_{[n]}, \ \text{for all} \ n=1,2,\ldots,N\\
&\mathcal{P}_{\Omega}(\dot{\mathcal{T}})=\dot{\mathcal{Y}}.
\end{split}$$ This problem can be solved by the ADMM framework, which minimizes the following augmented Lagrangian function: $$\begin{aligned}
\label{equ15}
&\mathcal{L}(\dot{\mathcal{T}}, \{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N},\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N},\{\beta_{n}\}_{n=1}^{N})\nonumber \\
&=\sum_{n=1}^{N}\left(\alpha_{n}\|\dot{\mathbf{X}}_{[n]}\|_{\ast}+\langle\dot{\mathbf{F}}_{[n]},\dot{\mathbf{T}}_{[n]}-\dot{\mathbf{X}}_{[n]}\rangle+\frac{\beta_{n}}{2}\|\dot{\mathbf{T}}_{[n]}-\dot{\mathbf{X}}_{[n]}\|_{F}^{2}\right),\end{aligned}$$ where $\{\beta_{n}\}_{n=1}^{N}$ are the penalty parameters, $\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N}$ are the Lagrange multipliers, which are $N$ Mode-k unfolding quaternion matrices of quaternion tensor $\dot{\mathcal{F}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$.
It is clear that although the objective function of (\[equ15\]) is not jointly convex for all variables, it is convex concerning each variable independently. Hence, A natural way to solve the problem is to iteratively optimize the augmented Lagrangian function (\[equ15\]) over one variable, while fixing the others. To update each variable, in the $\tau+1$th iteration, perform the following steps:
- **Step 1:** $\dot{\mathcal{T}}^{(\tau+1)}=\mathop{{\rm{arg\, min}}}\limits_{\dot{\mathcal{T}}}\:\mathcal{L}\left(\dot{\mathcal{T}}, (\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N})^{(\tau)},(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N})^{(\tau)},(\{\beta_{n}\}_{n=1}^{N})^{(\tau)}\right)$,
- **Step 2:** $(\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N})^{(\tau+1)}=\mathop{{\rm{arg\, min}}}\limits_{\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N}}\:\mathcal{L}\left(\dot{\mathcal{T}}^{(\tau+1)}, \{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N},(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N})^{(\tau)},(\{\beta_{n}\}_{n=1}^{N})^{(\tau)}\right)$,
- **Step 3:** $(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N})^{(\tau+1)}=\mathop{{\rm{arg\, min}}}\limits_{\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N}}\:\mathcal{L}\left(\dot{\mathcal{T}}^{(\tau+1)}, (\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N})^{(\tau+1)},\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N},(\{\beta_{n}\}_{n=1}^{N})^{(\tau)}\right)$,
- **Step 4:** Updating $(\{\beta_{n}\}_{n=1}^{N})^{(\tau+1)}$.
For **Step 1**, it is easy to find that the optimal solution of $\dot{\mathcal{T}}^{(\tau+1)}$ is $$\label{equ16}
\dot{\mathcal{T}}^{(\tau+1)}=\mathcal{P}_{\Omega^{c}}
\left(\frac{1}{N}\sum_{n}^{N}\bigg({\rm{Fold}}_{n}(\dot{\mathbf{X}}_{[n]}^{(\tau)})-\frac{1}{\beta_{n}^{(\tau)}}{\rm{Fold}}_{n}(\dot{\mathbf{F}}_{[n]}^{(\tau)})\bigg)\right)+\dot{\mathcal{Y}},$$ where $\Omega^{c}$ is the complement of $\Omega$, and we have used the fact that $\mathcal{P}_{\Omega^{c}}(\dot{\mathcal{Y}})=\mathbf{0}$ in (\[equ16\]).
For **Step 2**, it can be decomposed into $N$ independent optimization problems, which can be solved in paralleled. For each $n$, $\dot{\mathbf{X}}_{[n]}^{(\tau+1)}$ is the optimal solution of the following problem: $$\begin{aligned}
\label{equ17}
\dot{\mathbf{X}}_{[n]}^{(\tau+1)}&=\mathop{{\rm{arg\, min}}}\limits_{\dot{\mathbf{X}}_{[n]}}\ \alpha_{n}\|\dot{\mathbf{X}}_{[n]}\|_{\ast}+\langle\dot{\mathbf{F}}_{[n]}^{(\tau)},\dot{\mathbf{T}}_{[n]}^{(\tau+1)}-\dot{\mathbf{X}}_{[n]}\rangle+\frac{\beta_{n}^{(\tau)}}{2}\|\dot{\mathbf{T}}_{[n]}^{(\tau+1)}-\dot{\mathbf{X}}_{[n]}\|_{F}^{2} \nonumber\\
&=\mathop{{\rm{arg\, min}}}\limits_{\dot{\mathbf{X}}_{[n]}}\ \alpha_{n}\|\dot{\mathbf{X}}_{[n]}\|_{\ast}+\frac{\beta_{n}^{(\tau)}}{2}\|\dot{\mathbf{X}}_{[n]}-(\dot{\mathbf{T}}_{[n]}^{(\tau+1)}+\frac{1}{\beta_{n}^{(\tau)}}\dot{\mathbf{F}}_{n}^{(\tau)})\|_{F}^{2} \nonumber\\
&=\mathop{{\rm{arg\, min}}}\limits_{\dot{\mathbf{X}}_{[n]}}\ \frac{\alpha_{n}}{\beta_{n}^{(\tau)}}\|\dot{\mathbf{X}}_{[n]}\|_{\ast}+\frac{1}{2}\|\dot{\mathbf{X}}_{[n]}-(\dot{\mathbf{T}}_{[n]}^{(\tau+1)}+\frac{1}{\beta_{n}^{(\tau)}}\dot{\mathbf{F}}_{n}^{(\tau)})\|_{F}^{2}.\end{aligned}$$ Then, the optimal solution of (\[equ17\]) can be obtained by the following theorem.
[^5] Let $\dot{\mathbf{Q}}\in\mathbb{H}^{N_{1}\times N_{2}}$ be a given quaternion matrix, then the QSVD (defined on **theorem 1**) of $\dot{\mathbf{Q}}$ with rank $r$ is $$\label{ssvd}
\dot{\mathbf{Q}}=\dot{\mathbf{U}}\mathbf{\Lambda}\dot{\mathbf{V}}^{H}=\dot{\mathbf{U}}_{r}\mathbf{\Sigma}_{r}\dot{\mathbf{V}}_{r}^{H},$$ where $\dot{\mathbf{U}}_{r}=[\dot{\mathbf{u}}_{1},\dot{\mathbf{u}}_{2},\ldots,\dot{\mathbf{u}}_{r}]\in\mathbb{H}^{N_{1}\times r}$ and $\dot{\mathbf{V}}_{r}=[\dot{\mathbf{v}}_{1},\dot{\mathbf{v}}_{2},\ldots,\dot{\mathbf{v}}_{r}]\in\mathbb{H}^{N_{2}\times r}$, $\mathbf{\Sigma}_{r}={\rm{diag}}(\sigma_{1},\ldots,\sigma_{r})$. Define the quaternion matrix singular value thresholding operator $\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})=\dot{\mathbf{U}}_{r}\breve{\mathbf{\Sigma}}_{r}\dot{\mathbf{V}}_{r}^{H}$, where $\breve{\mathbf{\Sigma}}_{r}={\rm{diag}}\{{\rm{max}}(\sigma_{n}-\xi,0) \} (n=1,2,\ldots,r)$. Then the operator $\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})$ obeys $$\label{equ18}
\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})=\mathop{{\rm{arg\, min}}}\limits_{\dot{\mathbf{X}}}\ \xi\|\dot{\mathbf{X}}\|_{\ast}+\frac{1}{2}\|\dot{\mathbf{X}}-\dot{\mathbf{Q}}\|_{F}^{2}.$$
$\mathit{Proof.}$ It is obvious that the function $\mathfrak{N}(\dot{\mathbf{X}})=\xi\|\dot{\mathbf{X}}\|_{\ast}+\frac{1}{2}\|\dot{\mathbf{X}}-\dot{\mathbf{Q}}\|_{F}^{2}$ is strictly convex, hence there indeed exists a unique minimizer. We say that $\dot{\mathbf{X}}_{\ast}$ minimizes $\mathfrak{N}(\dot{\mathbf{X}})$ if and only if $\mathbf{0}$ is a subgradient of the function $\mathfrak{N}(\dot{\mathbf{X}})$ at the point $\dot{\mathbf{X}}_{\ast}$, *i.e.*, $$\label{cftj}
\mathbf{0}\in\xi\nabla_{sub}\|\dot{\mathbf{X}}_{\ast}\|_{\ast}+\dot{\mathbf{X}}_{\ast}-\dot{\mathbf{Q}},$$ where $\nabla_{sub}\|\dot{\mathbf{X}}_{\ast}\|_{\ast}$ denotes the subgradient set of $\|\dot{\mathbf{X}}_{\ast}\|_{\ast}$, which can be obtained by the following Lemma.
[@DBLP:journals/nla/JiaNS19] Suppose that $\dot{\mathbf{X}}\in\mathbb{H}^{N_{1}\times N_{2}}$ with rank $r$ has the QSVD as $\dot{\mathbf{A}}_{r}\mathbf{D}_{r}\dot{\mathbf{B}}_{r}^{H}=\sum_{n=1}^{r}d_{n}\dot{\mathbf{a}}_{n}\dot{\mathbf{b}}_{n}^{H}$, then $$\label{subd}
\nabla_{sub}\|\dot{\mathbf{X}}\|_{\ast}=\{\sum_{n=1}^{r}\dot{\mathbf{a}}_{n}\dot{\mathbf{b}}_{n}^{H}+\dot{\mathbf{E}}\mid\dot{\mathbf{E}}\in\mathbb{H}^{N_{1}\times N_{2}},\ \dot{\mathbf{A}}_{r}^{H}\dot{\mathbf{E}}=\mathbf{0},\ \dot{\mathbf{E}}\dot{\mathbf{B}}_{r}=\mathbf{0}, \ \|\dot{\mathbf{E}}\|\leq 1\}.$$
We rewritten the QSVD of $\dot{\mathbf{Q}}$ (*see* (\[ssvd\])) as $$\dot{\mathbf{Q}}=\dot{\mathbf{U}}_{r}\mathbf{\Sigma}_{r}\dot{\mathbf{V}}_{r}^{H}=\sum_{n=1}^{r}\sigma_{n}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}=\sum_{n=1}^{r_{0}}\sigma_{n}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}+\sum_{n=r_{0}+1}^{r}\sigma_{n}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H},$$ where $\sigma_{n}\leq \xi$ when $r_{0}+1 \leq n\leq r$. Then, setting $\dot{\mathbf{X}}_{\ast}=\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})$, we have $$\dot{\mathbf{X}}_{\ast}=\sum_{n=1}^{r_{0}}(\sigma_{n}-\xi)\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}+\sum_{n=r_{0}+1}^{r}0\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H},$$ and as a result $$\dot{\mathbf{X}}_{\ast}-\dot{\mathbf{Q}}=-\xi\left(\sum_{n=1}^{r_{0}}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}+\sum_{n=r_{0}+1}^{r}\frac{\sigma_{n}}{\xi}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}\right)
=-\xi\left(\sum_{n=1}^{r_{0}}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}+\dot{\mathbf{E}}\right),$$ where $\dot{\mathbf{E}}=\sum_{n=r_{0}+1}^{r}-\frac{\sigma_{n}}{\xi}\dot{\mathbf{u}}_{n}\dot{\mathbf{v}}_{n}^{H}$. According to **Lemma 1**, it is obvious that $\dot{\mathbf{X}}_{\ast}-\dot{\mathbf{Q}}\in-\xi\nabla_{sub}\|\dot{\mathbf{X}}\|_{\ast}$, *i.e.*, when $\dot{\mathbf{X}}_{\ast}=\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})$, (\[cftj\]) holds. Consequently, the $\mathfrak{S}_{\xi}(\dot{\mathbf{Q}})$ obeys the optimization problem (\[equ18\]).
Therefore, based on **Theorem 2**, we can easily obtain the following optimization result of (\[equ17\]) $$\label{equ19}
\dot{\mathbf{X}}_{[n]}^{(\tau+1)}=\mathfrak{S}_{\frac{\alpha_{n}}{\beta_{n}^{(\tau)}}}\left(\dot{\mathbf{T}}_{[n]}^{(\tau+1)}+\frac{1}{\beta_{n}^{(\tau)}}\dot{\mathbf{F}}_{n}^{(\tau)}\right).$$
For **Step 3**, it can also be decomposed into $N$ independent optimization problems and be solved in paralled. For each $n$, $\dot{\mathbf{F}}_{[n]}^{(\tau+1)}$ is updated by the following equation: $$\label{equ20}
\dot{\mathbf{F}}_{[n]}^{(\tau+1)}=\dot{\mathbf{F}}_{[n]}^{(\tau)}-\beta_{n}^{(\tau)}(\dot{\mathbf{X}}_{[n]}^{(\tau+1)}-\dot{\mathbf{T}}_{[n]}^{(\tau+1)}).$$
For **Step 4**, since the dynamical $(\{\beta_{n}\}_{n=1}^{N})$ are usually preferred to speed up the convergence of the algorithm [@DBLP:journals/pami/HuZYLH13], we use the following way to adaptive update $\beta_{n}$, for $n=1,2,\ldots,N$. $$\label{equ21}
\beta_{n}^{(\tau+1)}={\rm{min}}(\beta_{n}^{max},\eta\beta_{n}^{(\tau)}),$$ where $\beta_{n}^{max}$ is the default maximum of $\beta_{n}$, and $\eta\geq 1$ is a constant parameter. Generally, we set $\eta=\eta_{0}> 1$ if $\|\dot{\mathcal{T}}^{(\tau+1)}-\dot{\mathcal{T}}^{(\tau)}\|_{F}$ is small enough (*e.g.*, 0.01), $\eta=1$ otherwise.
Finally, the proposed **L**ow-**R**ank **C**ompletion for **Q**uaternion **T**ensor (**LRC-QT**) method can be summarized in Table \[tab\_algorithm\].
------------------------------------------------------------------------
\[tab\_algorithm\]
The observed quaternion tensor data $\dot{\mathcal{T}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$, the observed index set $\Omega$. **Initialize** $(\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N})^{(0)}$, $(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N})^{(0)}$, $\{\alpha_{n}\}_{n=1}^{N}$, $(\{\beta_{n}\}_{n=1}^{N})^{(0)}$, $\{\beta_{n}^{max}\}_{n=1}^{N}$, $\eta_{0}$, $\epsilon$. **Repeat** Update $\dot{\mathcal{T}}^{(\tau+1)}$ using equation (\[equ16\]). % Lines 5 and 6 all can be performed in parallel. Update $(\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{N})^{(\tau+1)}$ using equation (\[equ19\]). Update $(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{N})^{(\tau+1)}$ using equation (\[equ20\]). Update $\beta_{n}^{(\tau+1)}$ using equation (\[equ21\]). $\tau\longleftarrow \tau+1$. **Until** $\|\dot{\mathcal{T}}^{(\tau+1)}-\dot{\mathcal{T}}^{(\tau)}\|_{F}\leq\epsilon$. $\dot{\mathcal{T}}$.
------------------------------------------------------------------------
In the rest of this section, to facilitate direct processing of color image recovery issues, as a special case of the aforementioned method, we consider the low-rank quaternion matrix (second-order tensor) completion problem, *i.e.*, $$\label{equ22}
\begin{split}
&\mathop{{\rm{minimize}}}\limits_{\dot{\mathbf{T}}}\quad \alpha\|\dot{\mathbf{T}}\|_{\ast}\\
&\text{subject to} \quad \mathcal{P}_{\Omega}(\dot{\mathbf{T}})=\dot{\mathbf{Y}},
\end{split}$$ where $\dot{\mathbf{Y}}\in\mathbb{H}^{M\times N}$ is the underlying complete quaternion matrix, $\dot{\mathbf{T}}\in\mathbb{H}^{M\times N}$ is the observed quaternion matrix. For color image recovery, model (\[equ22\]) is different from traditional matrix based and third-order tensor based models and is more advantageous than them. Since the traditional matrix based models are inherently developed for gray-level images. Although third-order tensor based models can deal with this problem, for this type of algorithms, the rank of a tensor is generally pretty hard to determine [@DBLP:journals/tip/ZhouLLZ18], so they usually cannot offer the best low-rank approximation to a tensor. Besides, the tensor factorization or matricization based methods (*see, e.g.*, [@DBLP:journals/pami/LiuMWY13]) are likely to destroy color pixel structure. In brief, the recovery theory for low-rank tensor completion problem is not well established compared with that of matrix based completion problems [@DBLP:journals/nla/JiaNS19].
For problem (\[equ22\]), adding an additional variable quaternion matrix $\dot{\mathbf{X}}\in\mathbb{H}^{M\times N}$, we can obtain the following equivalent formulation: $$\label{equ23}
\begin{split}
\mathop{{\rm{minimize}}}\limits_{\dot{\mathbf{T}}}\quad &\alpha\|\dot{\mathbf{X}}\|_{\ast}\\
\text{subject to} \quad &\dot{\mathbf{T}}=\dot{\mathbf{X}},\quad \mathcal{P}_{\Omega}(\dot{\mathbf{T}})=\dot{\mathbf{Y}}.
\end{split}$$ Then we define the following augment Lagrangian function: $$\label{equ24}
\mathcal{L}(\dot{\mathbf{T}}, \dot{\mathbf{X}}, \dot{\mathbf{F}},\beta)
=\alpha\|\dot{\mathbf{X}}\|_{\ast}+\langle\dot{\mathbf{F}},\dot{\mathbf{T}}-\dot{\mathbf{X}}\rangle+\frac{\beta}{2}\|\dot{\mathbf{T}}-\dot{\mathbf{X}}\|_{F}^{2},$$ where $\beta$ is the penalty parameter, $\dot{\mathbf{F}}$ is the Lagrange multiplier. According to the ADMM framework, we independently update $\dot{\mathbf{T}}$, $\dot{\mathbf{X}}$, $\dot{\mathbf{F}}$, $\beta$, and we summarize the proposed **L**ow-**R**ank **C**ompletion for **Q**uaternion **M**atrix (**LRC-QM**) method in Table \[tab\_algorithm2\].
------------------------------------------------------------------------
\[tab\_algorithm2\]
The observed quaternion matrix data $\dot{\mathbf{T}}\in\mathbb{H}^{M\times N}$, the observed index set $\Omega$. **Initialize** $\dot{\mathbf{X}}^{(0)}$, $\dot{\mathbf{F}}^{(0)}$, $\alpha$, $\beta^{(0)}$, $\beta^{max}$, $\eta_{0}$, $\epsilon$. **Repeat** $\dot{\mathbf{T}}^{(\tau+1)} \longleftarrow \mathcal{P}_{\Omega^{c}}
\left(\dot{\mathbf{X}}^{(\tau)}-\frac{1}{\beta}\dot{\mathbf{F}}^{(\tau)}\right)+\dot{\mathbf{Y}}$ (where $\mathcal{P}_{\Omega^{c}}(\dot{\mathbf{Y}})=\mathbf{0}$). $\dot{\mathbf{X}}^{(\tau+1)} \longleftarrow\mathfrak{S}_{\frac{\alpha}{\beta^{(\tau)}}}\left(\dot{\mathbf{T}}^{(\tau+1)}+\frac{1}{\beta^{(\tau)}}\dot{\mathbf{F}}^{(\tau)}\right)$. $\dot{\mathbf{F}}^{(\tau+1)}\longleftarrow \dot{\mathbf{F}}^{(\tau)}-\beta^{(\tau)}(\dot{\mathbf{X}}^{(\tau+1)}-\dot{\mathbf{T}}^{(\tau+1)})$. $\beta^{(\tau+1)}\longleftarrow{\rm{min}}(\beta^{max},\eta\beta^{(\tau)})$ (we set $\eta=\eta_{0}> 1$ if $\|\dot{\mathbf{T}}^{(\tau+1)}-\dot{\mathbf{T}}^{(\tau)}\|_{F}$ is small enough (*e.g.*, 0.01), $\eta=1$ otherwise). $\tau\longleftarrow \tau+1$. **Until** $\|\dot{\mathbf{T}}^{(\tau+1)}-\dot{\mathbf{T}}^{(\tau)}\|_{F}\leq\epsilon$. $\dot{\mathbf{T}}$.
------------------------------------------------------------------------
The computational complexity analysis
-------------------------------------
For LRC-QT, the observed quaternion tensor data $\dot{\mathcal{T}}\in\mathbb{H}^{N_{1}\times N_{2} \times\ldots \times N_{N}}$. We assume, for simplicity, that $N_{1}= N_{2} =\ldots = N_{N}=I$. It is easy to see that the main per-iteration computational complexity lies in the update of $\dot{\mathbf{X}}_{[n]}, n=1,2,\ldots,N$, which requires computing $N$ QSVD of $I\times I^{N-1}$ quaternion matrices. There has been some quaternionic algorithms (*e.g.*, [@Bihan2007Jacobi]) were proposed for computing the QSVD, nevertheless, they are too time-consuming. We propose to compute the QSVD using the isomorphism between $\mathbb{H}^{N_{1}\times N_{2}}$ and $\mathbb{C}^{2N_{1}\times 2N_{2}}$ (*see* (\[definef\])). According to (\[qsvd\]), the computation of the QSVD for $N$ $I\times I^{N-1}$ quaternion matrices is equivalent to the computation of the SVD of $N$ $2I\times 2I^{N-1}$ complex matrices, which can be performed using well-established classical algorithms of SVD over $\mathbb{C}^{2N_{1}\times 2N_{2}}$ (the built-in function ‘svd’ in MATLAB 2014b is used by us). We still assume that the computational complexity of SVD for a $K\times K$ complex matrix is about $\mathcal{O}(K^3)$. Therefore, the whole computational complexity of LRC-QT for one iteration is about $\mathcal{O}(8NI^{2N-1})$. Analogously, For LRC-QM, the observed quaternion matrix data $\dot{\mathbf{T}}\in\mathbb{H}^{M\times N}$. Assuming $M=N=I$, the main per-iteration computational complexity lies in the update of $\dot{\mathbf{X}}$, which is about $\mathcal{O}(8I^{3})$.
Simulation results {#sec5}
==================
In this section, simulations on some natural color videos and images are conducted to evaluate the performance of the proposed LRC-QT and LRC-QM methods. And we compare them with several existing state-of-the-art approaches, including SiLRTC [@DBLP:journals/pami/LiuMWY13], SPC [@DBLP:journals/tsp/YokotaZC16], TMac (involving TMac-inc and TMac-dec) [@1930-8337_2015_2_601] and TCTF (which is not applicable in color video simulation) [@DBLP:journals/tip/ZhouLLZ18]. All the simulations are run in MATLAB $2014b$ under Windows $7$ on a personal computer with $2.20$GHz CPU and $8.00$GB memory.
All color videos and images, in our simulations, are initially represented by fourth-order tensors $\mathcal{T}_{V}\in \mathbb{R}^{M\times N\times 3 \times T}$ and third-order tensors $\mathcal{T}_{I}\in \mathbb{R}^{M\times N\times 3}$ respectively. For LRC-QT, each color video is reshaped as a pure quaternion tensor $\dot{\mathcal{T}}\in\mathbb{H}^{M\times N \times T}$ by using the following way: $$\dot{\mathcal{T}}=\mathcal{T}_{V}(:,:,1,:)i+\mathcal{T}_{V}(:,:,2,:)j+\mathcal{T}_{V}(:,:,3,:)k.$$ For LRC-QM, each color image is reshaped as a pure quaternion matrix $\dot{\mathbf{T}}\in\mathbb{H}^{M\times N}$ by using the following way: $$\dot{\mathbf{T}}= \mathcal{T}_{I}(:,:,1)i+\mathcal{T}_{I}(:,:,2)j+\mathcal{T}_{I}(:,:,3)k.$$ In addition, we uniformly generate the index set $\Omega$ at Gaussian random distribution, and define the sampling ratio (SR) as: ${\rm{SR}}=\frac{{\rm{numel}}(\Omega)}{M\times N\times 3 \times T}100\% \ (\text{or}\ {\rm{SR}}=\frac{{\rm{numel}}(\Omega)}{M\times N\times 3}100\%\ \text{for color images})$, where ${\rm{numel}}(\Omega)$ represents the number of observation entries in the index set $\Omega$.
**Quantitative assessment:** To evaluate the performance of proposed methods, except visual quality, we employ three quantitative quality indexes, including the peak signal-to-noise ratio (PSNR), the structure similarity (SSIM) and the feature similarity (FSIM), which are respectively defined as: ${\rm{PSNR}}=10{\rm{log}}10\left( \frac{{\rm{Peakval}}^{2}}{{\rm{MSE}}}\right)$, where ${\rm{Peakval}}$ is taken from the range of the pixel value datatype (*e.g.*, for uint8 pixel value, it is 255), ${\rm{MSE}}$ is the mean square error, *i.e.*, ${\rm{MSE}}=\|\mathcal{X}-\mathcal{T}\|_{F}^{2}/{\rm{numel}}(\mathcal{X})$, where $\mathcal{X}$ and $\mathcal{T}$ are the recovered and truth data, respectively; ${\rm{SSIM}}=\frac{(2\mu_{\mathcal{T}}\mu_{\mathcal{X}}+C_{1})(2\sigma_{\mathcal{T}\mathcal{X}}+C_{2})}{(\mu_{\mathcal{T}}^{2}+\mu_{\mathcal{X}}^{2}+C_{1})(\sigma_{\mathcal{T}}^{2}+\sigma_{\mathcal{X}}^{2}+C_{2})}$, where $\mu_{\mathcal{T}}$, $\mu_{\mathcal{X}}$, $\sigma_{\mathcal{T}}$, $\sigma_{\mathcal{X}}$ and $\sigma_{\mathcal{T}\mathcal{X}}$ are the local means, standard deviations, and cross-covariance for images $\mathcal{T}$ and $\mathcal{X}$, $C_{1}=(0.01L)^{2}$, $C_{2}=(0.03L)^{2}$, $C_{3}=C_{2}/2$, $L$ is the specified dynamic range of the pixel values (average structure similarity index over frames (ASSIM) is chosen for color videos); ${\rm{FSIM}}=\frac{\sum_{z\in \Delta}S_{L}(z)PC_{m}(z)}{\sum_{z\in \Delta}PC_{m}(z)}$, where $\Delta$ demotes the whole image spatial domain. The phase congruency for position $z$ of image $\mathcal{T}$ is denoted as $PC_{x}(\mathcal{T})$, then $PC_{m}(z)={\rm{max}}\{PC_{\mathcal{T}(z)},PC_{\mathcal{X}(z)}\}$, $S_{L}(z)$ is the gradient magnitude for position $z$ (average feature similarity index over frames (AFSIM) is chosen for color videos).
For LRC-QT, we let $\mathcal{X}(:,:,1,:)={\rm{Imag}}_{1}(\dot{\mathcal{T}})$, $\mathcal{X}(:,:,2,:)={\rm{Imag}}_{2}(\dot{\mathcal{T}})$ and $\mathcal{X}(:,:,3,:)={\rm{Imag}}_{3}(\dot{\mathcal{T}})$, where ${\rm{Imag}}_{n}(\dot{\mathcal{T}})\:(n=1,2,3)$ denotes $n$-th imaginary part of $\dot{\mathcal{T}}$. Analogously, for LRC-QM, $\mathcal{X}(:,:,1)={\rm{Imag}}_{1}(\dot{\mathbf{T}})$, $\mathcal{X}(:,:,2)={\rm{Imag}}_{2}(\dot{\mathbf{T}})$ and $\mathcal{X}(:,:,3)={\rm{Imag}}_{3}(\dot{\mathbf{T}})$.
**Datasets:** In the simulations, we use the color video dataset: YUV Video Sequence[^6] where each sequence contains at least $150$ frames, and the color image dataset: Berkeley Segmentation Dataset (BSD)[^7] which includes $300$ clean color images of size $481\times 321\times 3$.
We first show that these color videos and color images can be well approximated and reconstructed by the low-rank quaternion tensors[^8] and low-rank quaternion matrices, respectively. As mentioned in [@4797640; @DBLP:journals/tip/ZhouLLZ18], *etc.*, when the videos or images data are arranged into tensors or matrices, they lie on a union of low-rank subspaces approximately, which indicate the low-rank structure of the video or image data. This is also true for quaternion tensor and quaternion matrix data. For instance, in Figure.\[fig\_rank\] we display the singular values of one color video with size $288\times 352\times3\times20$ (reconstructed as third-order pure quaternion tensor with size $288\times 352\times20$) and one color image with size $481\times 321\times 3$ (reconstructed as pure quaternion matrix with size $481\times 321$), which are selected from the two datasets randomly. One can obviously see that most of the singular values are very close to $0$ and much smaller than the first several larger singular values. So we could say that the color videos and color images can be well approximated by the low-rank quaternion tensors and low-rank quaternion matrices, respectively, as we desired.
\
**Parameter and initialization settings:** For LRC-QT in Table \[tab\_algorithm\], $(\{\dot{\mathbf{X}}_{[n]}\}_{n=1}^{3})^{(0)}=(\{\dot{\mathbf{F}}_{[n]}\}_{n=1}^{3})^{(0)}=(\mathbf{0}, \mathbf{0}, \mathbf{0})$, $\{\alpha_{n}\}_{n=1}^{3}=(2, 2, 10^{-3})$, $(\{\beta_{n}\}_{n=1}^{3})^{(0)}=(0.08, 0.08, 1)$, $\{\beta_{n}^{max}\}_{n=1}^{3}=(10^{3}, 10^{3}, 10^{3})$, $\eta_{0}=1.05$, $\epsilon=10^{-3}$. For LRC-QM in Table \[tab\_algorithm2\], $\dot{\mathbf{X}}^{(0)}=\dot{\mathbf{F}}^{(0)}=\mathbf{0}$, $\alpha=2$, $\beta^{(0)}=0.08$, $\beta^{max}=10^{3}$, $\eta_{0}=1.05$, $\epsilon=10^{-3}$. For LRC-QT and LRC-QM, the parameters and initialization settings are just based on our experience and simulation results, and there may be better settings. For SiLRTC[^9], SPC[^10], TMac[^11] and TCTF[^12], the codes of them are provided by their corresponding authors. The parameter settings and initialization methods of these algorithms are all based on the suggestions of their corresponding papers. Besides, for LRC-QT, as equations (\[equ19\]) and (\[equ20\]) show, we can update all $\dot{\mathbf{X}}_{[n]}$ and $\dot{\mathbf{F}}_{[n]}$ $(n=1,2,\ldots,N)$ parallelly, but for a fair comparison of the algorithm running time, we still employ the serial updating scheme in our code.
**Simulation 1:** In this simulation, we use four color videos (Bus, News, Salesman and Suzie) reconstructed by third-order pure quaternion tensors with size $288\times 253\times 20$, $144\times 176\times 20$, $144\times 176\times 20$ and $144\times 176\times 20$ respectively, and shown as Figure.\[fig\_zs\], to evaluate the proposed LRC-QT for color video recovery. Figures.\[busvideo\]-\[suzievideo\] show the recovery results by different methods for the 1st, 8th, 15th and 20th frames of Bus video with ${\rm{SR}}=50\%$, News video with ${\rm{SR}}=30\%$, Salesman video with ${\rm{SR}}=20\%$ and Suzie video with ${\rm{SR}}=10\%$, respectively. We see from Figures.\[busvideo\]-\[suzievideo\] that the color video frames recovered by LRC-QT are visually better than those recovered by the other compared approaches. At very low SR (*e.g.*, SR$=\%10$, *see* Figure.\[suzievideo\]), the advantage of LRC-QT seems to be more obvious. Table \[videotable\] summaries the PSNR, ASSIM, AFSIM, iterations and running time of different methods on the four color videos with various SRs. From the results, one can find that for PSNR, ASSIM and AFSIM, LRC-QT reaches the highest values in most cases. We have reason to believe that this is mainly due to the advantage of quaternion representation of color pixel values. For the comparison of the number of iterations and running time, our approach is to compare them required by different compared methods when the PSNR obtained by all methods reaches a certain common pre-set value. We can see that although our method is not the fastest one, it can also achieve an acceptable PSNR value faster, and if we consider the fact that it can perform parallelly, the running time will be greatly reduced (Specifically, we can see that to achieve an acceptable PSNR value, our method requires only a few iterations, but the total time is not the shortest, which is closely related to the per-iteration computational complexity and how optimized the code is.).
{width="13cm" height="5.1cm"}
{width="13cm" height="5.1cm"}
{width="13cm" height="5.1cm"}
{width="13cm" height="5.1cm"}
\[videotable\]
**Simulation 2:** In this simulation, we use BSD dataset to evaluate the proposed LRC-QM for color image recovery. We randomly select 50 color images from this dataset reconstructed by pure quaternion matrices with size $481\times 321$.
{width="14cm" height="4cm"}
{width="14cm" height="4cm"}
{width="14cm" height="4cm"}
In Figures.\[imag1\]-\[imag3\] we display the comparison of PSNR, SSIM and FSIM results of different methods for color image recovery on $50$ color images with various SRs. From the results, one can find that our LRC-QM approach performs better than the other methods in the vast majority of images. Besides the superiority on PSNR, the good performance of our method on SSIM and FSIM also demonstrates the advantages of the quaternion-based model.
Conclusions {#sec6}
===========
Focusing on color videos and images recovery problems, this paper utilizing quaternions to represent the color pixels with RGB channels proposed a low-rank quaternion tensor completion method. Quaternion representation processes a color video or image holistically as a vector field and handles the coupling between the color channels naturally, and color information of source video or image is fully preserved. Although color video and images can also be represented as higher-order real tensors, the color structure will be destroyed during the process of matricization (*e.g., mode-k unfolding*). Due to the special structure of the three imaginary parts of the quaternion, the relative positions of the three color channel pixels of one pixel are insensitive to the deformation of the quaternion tensor, which means that the color structure of the video can be completely maintained in the process of matricization. We adopted the ADMM framework to optimize the proposed model, which can guarantee the convergence of the algorithm. In addition, to facilitate direct processing of color image recovery issues, as a special case, we displayed the low-rank quaternion matrix completion model and optimization procedure separately. Theoretically, the proposed method can be well used to recover missing entries of any multidimensional data with color structures. In the simulation section, we mainly considered the color videos and images recovery problems. The results demonstrate the competitive performance (*w.r.t.*, PSNR, SSIM and FSIM) of the proposed methods compared with several state-of-the-art approaches.
Note that although the proposed method can well recover color videos and images, it needs to compute QSVD in each iteration, which is time-consuming and storage-intensive for large matrices. While the characteristic of less iteration of the algorithm can alleviate this shortcoming to some extent, in the future, we still aim to further explore better QSVD method to improve the efficiency of the algorithm, or to use some low-rank decomposition approaches to replace the nuclear norm minimization model.
[2]{}
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by The Science and Technology Development Fund, Macau SAR (File no. FDCT/085/2018/A2).
[^1]: Color images are represented by second-order quaternion tensors, we also call them quaternion matrices in the paper.
[^2]: A detailed introduction to the quaternion can be found in Section \[section\_2\].
[^3]: Here we just give some fundamental algebraic operations used in our work briefly, which follow the definition in [@DBLP:journals/pr/ShaoSWCC14; @Girard2007Quaternions]. Readers can find more details on quaternion algebra in the references.
[^4]: A unitary quaternion matrix $\dot{\mathbf{U}}\in\mathbb{H}^{N\times N}$ has the following property: $\dot{\mathbf{U}}\dot{\mathbf{U}}^{H}=\dot{\mathbf{U}}^{H}\dot{\mathbf{U}}=\mathbf{I}_{N}$, with $\mathbf{I}_{N}\in\mathbb{R}^{N\times N}$ being the quaternion identity matrix which is the same as the classical identity matrix.
[^5]: Similar theorems for traditional real matrix case can be found in [@Furnival2017Denoising; @DBLP:journals/tgrs/HeZZS16; @DBLP:journals/tip/BenguaPTD17], *etc.*, we will demonstrate that the theorem still holds true for the quaternion matrix case in this paper.
[^6]: <http://trace.eas.asu.edu/yuv/>
[^7]: [ https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/]( https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/)
[^8]: The rank of quaternion tensor here refers to the Tucker rank defined by **Definition 5**.
[^9]: [ http://www.cs.rochester.edu/ jliu/publications.html]( http://www.cs.rochester.edu/ jliu/publications.html)
[^10]: [ http://ieeexplore.ieee.org/document/7502115/media]( http://ieeexplore.ieee.org/document/7502115/media)
[^11]: [http://www.caam.rice.edu/ yx9/TMac/](http://www.caam.rice.edu/ yx9/TMac/)
[^12]: <https://panzhous.github.io/>
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
and Christian Hagen\
Institute for Theoretical Physics, University of Regensburg, D-93040 Regensburg, Germany
- |
Erek Bilgici and Christof Gattringer\
Institut für Physik, FB Theoretische Physik, Universität Graz, A-8010 Graz, Austria\
\
E-mails:
title: Dressed Polyakov loops and center symmetry from Dirac spectra
---
The main phenomena in QCD at the finite temperature transition are deconfinement and chiral symmetry restoration. There has always been the question whether there is a mechanism connecting the two, in particular since in the quenched theory the corresponding phase transitions occur at the same critical temperatures [@kogut:83]. Here we define an observable that indeed links the two and discuss numerical findings from quenched lattice configurations [@bilgici:08].
The Polyakov loop, on the one hand, is the order parameter of confinement, being traceless in the confined phase and moving towards the center of the gauge group at temperatures above $T_c$. The spectral density of the Dirac operator at the origin, $\rho(0)$, on the other hand, is the order parameter of chiral symmetry. It is related to the condensate by the famous Banks-Casher relation [@banks:80] $\langle\bar{\psi}\psi\rangle=-\pi\rho(0)$ and vanishes above $T_c$. How does confinement leave a trace in the Dirac spectrum? The answer lies in the dependence on temporal boundary conditions, as we will show now.
We use the lattice as a regulator. The (untraced) Polyakov loop is the product of temporal links. For the lattice Dirac operator we use the staggered one [@kogut:74a], which can be viewed as hopping by one link. It is well-known that the $k$-th power of the Dirac operator at the same argument, $D^k(x,x)$, contains all products of links along closed loops of length $k$ starting at $x$. To distinguish the Polyakov loop from ‘trivially closed’ loops (like the plaquette), one needs phase boundary conditions for the fermions [@gattringer:06b], $\psi(x_0+\beta,\vec{x})=e^{i\varphi} \psi(x_0,\vec{x})$. These boundary conditions amount to an imaginary chemical potential. They can be easily implemented by replacing the temporal links $U_0$ in some time slice by $e^{i\varphi}U_0$. As a consequence, all loops get a factor $e^{i\varphi q}$ according to their winding number $q$ around the temporal direction, trivially closed loops remain unchanged.
In this way, the Polyakov loop can be reconstructed from the Dirac spectrum by using at least three boundary conditions, see [@bruckmann:06b]. This ‘thin’ Polyakov loop, however, has poor renormalization and scaling properties and it turned out that in this approach it is UV dominated [@bruckmann:06b].
Influenced by the Jena group [@synatschke:07a], we instead consider the propagator $1/(m+D)$ with some probe mass $m$ and use a geometric series obtaining all powers of the Dirac operator. Denoting the Dirac operator at a particular boundary condition $\varphi$ by $D_\varphi$, i.e., including the factors of $e^{i\varphi}$, the propagator is given as a product of links along all closed loops $l$, $$\tr\,\frac{1}{m+D_{\varphi}}=
\frac{1}{m}\sum_{{\rm loops}\:\: l}
\frac{{\rm sign}(l)}{(2am)^{|l|}}\:
e^{\,i\varphi q(l)}\:\,
\tr_{\!c}\!\!\!\!\prod_{(x,\mu)\in l} U_\mu(x)\,,
\label{eq_pre_formula}$$ where $|l|$ is the length of the loop and ${\rm sign}(l)$ comes from the staggered factor. Of importance in (\[eq\_pre\_formula\]) is the phase factor, since one can project onto a particular winding $q$ by a Fourier transform w.r.t. the boundary angle, $\int_0^{2\pi} d\varphi \, e^{-i\varphi q}\ldots$. Specifying to a single winding, $q=1$, like for the Polyakov loop, we arrive at [@bilgici:08] $$\ts\equiv
\int_0^{2\pi} \frac{d\varphi}{2\pi} \, e^{-i\varphi}
\frac{1}{V}\Big\langle \tr \frac{1}{m+D_\varphi}
\Big\rangle =
\frac{1}{mV}
\sum_{q(l)=1}
\frac{{\rm sign}(l)}{(2am)^{|l|}}\:
\Big\langle\tr_{\!c}\!\!\!\!\prod_{(x,\mu)\in l} U_\mu(x)\Big\rangle\,.
\label{eq_the_formula}$$ We refer to the new observable $\ts$ as the ‘dual condensate’, because it is obtained through a Fourier transform from the trace of the propagator. Indeed, in the massless limit (after the infinite volume limit as usual) we obtain the chiral condensate integrated with a phase factor over the boundary conditions and, using the Banks-Casher relation at every individual angle $\varphi$, the corresponding representation in terms of the eigenvalue densities $\rho(0)_\varphi$. The right hand side of (\[eq\_the\_formula\]) represents the ‘dressed Polyakov loop’, that is the set of all loops which wind once around the temporal direction. In the infinite mass limit, detours become suppressed and only the straight Polyakov loop survives as it is the shortest possible loop in this set.
![Left: The dressed Polyakov loop $\ts$ as a function of $T$ at various lattices (for $m =$ 100 MeV). Right: The corresponding plot for the conventional Polyakov loop.[]{data-label="fig_order_parameter"}](sigma_100_tilde "fig:"){width="0.46\linewidth"} ![Left: The dressed Polyakov loop $\ts$ as a function of $T$ at various lattices (for $m =$ 100 MeV). Right: The corresponding plot for the conventional Polyakov loop.[]{data-label="fig_order_parameter"}](thinloop "fig:"){width="0.46\linewidth"}
Fig. \[fig\_order\_parameter\] shows that $\ts$ is indeed an order parameter. Keeping the mass $m$ fixed, $\ts$ vanishes below the critical temperature (which is about 280 MeV in the quenched case) and develops an expectation value for higher temperatures. One finds that the results (when expressed in physical units) obtained for different volumes and with different resolution essentially fall on a universal curve. This illustrates the good renormalization properties of our observable, which are inherited from $\langle\bar{\psi}\psi\rangle$ and may be viewed as an effect of the dressing rendering such loops less UV dominated.
![Individual and accumulated contributions to the spectral sum (\[eq\_spectral\_rep\]) at two different masses.[]{data-label="fig_spectral_contributions"}](contrib_accum){width="0.6\linewidth"}
In a spectral representation, the Dirac operator in $\ts$ is replaced by an eigenvalue sum, $$\ts=\int_0^{2\pi} \frac{d\phi}{2\pi} \, e^{-i\phi}
\frac{1}{V}\Big\langle \sum_i
\frac{1}{m+\lambda_\phi^{(i)}}\Big\rangle\,.
\label{eq_spectral_rep}$$ As the eigenvalues appear in the denominator, we expect the sum to be dominated by the IR modes. As Fig. \[fig\_spectral\_contributions\] shows, this is confirmed by our lattice data.
How is a finite resp. vanishing order parameter $\ts$ built up by the eigenvalues? Fig. \[fig\_response\_integrand\] shows that they respond differently to the boundary conditions in the confined vs. deconfined phase. The eigenvalues are independent of the boundary condition in the confined phase, which leads to a vanishing order parameter $\ts$. In the deconfined phase, on the other hand, the eigenvalues show a typical cosine-type of modulation. Together with the Fourier factor this yields a nonvanishing $\ts$.
![The integrand $\langle \sum_i (m+\lambda_\varphi^{(i)})^{-1}\rangle/V$ as a function of the boundary condition $\varphi$ for two values of $am$ and two temperatures, for configurations with real Polyakov loop. []{data-label="fig_response_integrand"}](integrand_2pi){width="0.45\linewidth"}
The chiral condensate has to behave essentially in the same way, as it is the integrand in the massless limit. Although the chiral condensate is finite in the confined phase, it is independent of the boundary condition $\varphi$ and hence results in a vanishing dual condensate. The situation in the deconfined phase might be more confusing at first glance as the spectrum has a gap there. For boundary conditions in line with the original Polyakov loop, however, the chiral condensate persists above the critical temperature $T_c$, which has been seen in several studies [@gattringer:02b]. This mechanism is needed to ensure a finite $\ts$ and should actually be at work for all $T>T_c$.
In the quenched case the conventional Polyakov loop is the order parameter for center symmetry. Under center transformations the dressed Polyakov loop behaves in the same way. Hence, $\ts$ is an order parameter for center symmetry, which is underneath our numerical findings and which we therefore expect to be independent of the choice of the lattice Dirac operator.
[99]{}
J. Kogut [*et. al.*]{}, [*Phys. Rev. Lett.*]{} [ **50**]{} (1983) 393. E. Bilgici, F. Bruckmann, C. Gattringer and C. Hagen, [*Phys. Rev.*]{} [**D77**]{} (2008) 094007. T. Banks and A. Casher, [*Nucl. Phys.*]{} [**B169**]{} (1980) 103.
J. B. Kogut and L. Susskind, [*Phys. Rev.*]{} [**D11**]{} (1975) 395. C. Gattringer, [*Phys. Rev. Lett.*]{} [**97**]{} (2006) 032003. F. Bruckmann, C. Gattringer and C. Hagen, [*Phys. Lett.*]{} [**B647**]{} (2007) 56. F. Synatschke, A. Wipf and C. Wozar, [*Phys. Rev.*]{} [ **D75**]{} (2007) 114003. C. Gattringer and S. Schaefer, [*Nucl. Phys.*]{} [**B654**]{} (2003) 30; V. G. Bornyakov [*et. al.*]{}, arxiv: [[0807.1980]{}](http://arXiv.org/abs/0807.1980) \[hep-lat\]; T. G. Kovacs, [*PoS*]{} [**LAT2008**]{} (2008) 198; M. A. Stephanov, [*Phys. Lett.*]{} [**B375**]{} (1996) 249.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A general method is obtained for finding recurrences involving the number of spanning trees of grid graphs, obtained by taking the graph product of an arbitrary graph and path or cycle. The results in this paper extend the work by Desjarlais and Molina and give concrete methods for finding the recurrences. Many new recurrences are found, yielding conjectures on the order of the linear recurrences of grid graphs and graphs obtained by taking the product of a complete graph and a path.'
author:
- Paul Raff
bibliography:
- 'refs.bib'
date: 'July 25, 2008'
title: Spanning Trees in Grid Graphs
---
Introduction
============
The Matrix Tree Theorem of Kirchhoff, a generalization of Cayley’s Theorem from complete graphs to arbitrary graphs [@Stanley], gives the number of spanning trees on a labeled graph as a determinant of a specific matrix. If $A = (a_{ij})$ is the adjacency matrix of a graph $G$, then the number of spanning trees can be found by computing any cofactor of the Laplacian matrix of $G$, or specific to the $(n,n)$-cofactor:
$$\mbox{Number of spanning trees of $G$ } =
\begin{vmatrix}
a_{12} + \ldots + a_{1n} & -a_{12} & \cdots & -a_{1,n-1} \\
-a_{21} & a_{21} + \cdots + a_{2n} & & -a_{2,n-1} \\
\vdots & & \ddots & \vdots \\
-a_{n-1,1} & -a_{n-1,2} & \cdots & a_{n-1,1} + \cdots + a_{n-1,n}
\end{vmatrix}$$
Since determinants are easy to compute, then the Matrix Tree Theorem allows for the computation for the first few numbers in the sequence of spanning trees for families of graphs dependent on one or more parameters. However, the downside of the Matrix Tree Theorem is that it can only produce a sequence of numbers, and cannot *a priori* assist in finding out the recurrence involved with said sequence. In this paper, the motivation is the following families of graphs:
1. $k \times n$ grid graphs, with $n \to \infty$.
2. $k \times n$ cylinder graphs, with $n \to \infty$.
3. $k \times n$ torus graphs, with $n \to \infty$.
All of the families of graphs mentioned above can be placed into a more general class of graphs of the form $G \times P_n$ or $G \times C_n$, where $P_n$ and $C_n$ denote the path and cylinder graph on $n$ vertices, respectively. For each of these classes, a general method is obtained for finding recurrences for all of the above families of graphs, and explicit recurrences are found for many cases. The only drawback, as it stands, is the amount of computational power needed to obtain these recurrences, as the recurrences are obtained through characteristic polynomials of large matrices. The result is at least 15 new sequences of numbers, plus improvements on the best-known recurrences known for other sequences.
History and Outline
===================
The main source of the historical results is a paper [@FaasePaper] and website [@FaaseSite] by Faase, where the main motivation is to count the number of hamiltonian cycles in certain classes of graphs. Later on, in 2000, Desjarlais and Molina [@Desjarlais] discuss the number of spanning trees in $2 \times n$ and $3 \times n$ grid graphs. In 2004, Golin and Leung [@Golin] discuss a technique called *unhooking* which will be used in this paper to reduce the problem of counting spanning trees in cylinder graphs to the problem of counting spanning trees in grid graphs.
In the first two papers, and this one, the general idea is the same: our goal is to count the number of spanning trees, but the method we use requires us to count other related objects, also. The paper by Faase appeals to the Transfer-Matrix Method, used widely in statistical mechanics (for more about the Transfer-Matrix Method, see \[[@Stanley]\]). The main distinction of this paper from [@Desjarlais] is the direct application of the Cayley-Hamilton Theorem to achieve recurrences for the sequences we are investigating. Overall, the results from this paper yield sequences for the number of spanning trees of the graphs $G \times P_n$ and $G \times C_n$ for any graph $G$. Along with these sequences, our methods find the minimal recurrence, generating function, and closed-form formulae for all of these sequences. As a consequence, we also find the sequences and recurrences for many, many other types of subgraphs.
The bulk of the paper focuses on the steps involved in finding the transition matrix for a given graph. In doing so, we will have to count other, related spanning forests with special properties.
Notation.
=========
All of the graphs we will be dealing with depend on two parameters, which we will call $k$ and $n$. In all cases, we will think of $k$ as fixed and $n \to \infty$.
The $k \times n$ *grid graph* $G_k(n)$ is the simple graph with vertex and edge sets as follows:
$$\begin{aligned}
V\left(G_k(n)\right) &= {\{v_{ij} \mid 1 \leq i \leq k, 1 \leq j \leq n\}} \\
E\left(G_k(n)\right) &= {\{v_{i,j}v_{i',j'} \mid |i - i'| + |j - j'| = 1\}}\end{aligned}$$
In order to keep the diagrams clean, Figure \[labeling\] shows the vertex naming conventions we will use.
(0,0)(6,3) (0,0)[0.1]{}[A]{} (0,1)[0.1]{}[B]{} (0,2)[0.1]{}[C]{} (0,3)[0.1]{}[D]{} (2,0)[0.1]{}[E]{} (2,1)[0.1]{}[F]{} (2,2)[0.1]{}[G]{} (2,3)[0.1]{}[H]{} (6,0)[0.1]{}[I]{} (6,1)[0.1]{}[J]{} (6,2)[0.1]{}[K]{} (6,3)[0.1]{}[L]{} (0,1.3)[0.03]{}[a]{} (0,1.5)[0.03]{}[b]{} (0,1.7)[0.03]{}[c]{} (2,1.3)[0.03]{}[d]{} (2,1.5)[0.03]{}[e]{} (2,1.7)[0.03]{}[f]{} (6,1.3)[0.03]{}[g]{} (6,1.5)[0.03]{}[h]{} (6,1.7)[0.03]{}[i]{} (3.2,0)[0.03]{}[j]{} (4,0)[0.03]{}[k]{} (4.8,0)[0.03]{}[l]{} (3.2,1)[0.03]{}[m]{} (4,1)[0.03]{}[n]{} (4.8,1)[0.03]{}[o]{} (3.2,2)[0.03]{}[p]{} (4,2)[0.03]{}[q]{} (4.8,2)[0.03]{}[r]{} (3.2,3)[0.03]{}[s]{} (4,3)[0.03]{}[t]{} (4.8,3)[0.03]{}[u]{}
When showing examples, usually of spanning trees or spanning forests, we will always show the underlying graph in one form or another. A concrete example is given in figure \[edge-ex\]: we will use black edges for edges in the subgraph exemplified; all unused edges will show up in light grey.
(0,0)(2,2) (0,0)[0.1]{}[A]{} (0,1)[0.1]{}[B]{} (0,2)[0.1]{}[C]{} (1,0)[0.1]{}[D]{} (1,1)[0.1]{}[E]{} (1,2)[0.1]{}[F]{} (2,0)[0.1]{}[G]{} (2,1)[0.1]{}[H]{} (2,2)[0.1]{}[I]{}
When dealing with grids of arbitrary size, we will mainly be interested in the very right-most end of the grid, so we will represent the rest of the graph we do not care about by a gray box, as shown in figure \[arb-ex\].
(0,0)(3,1) (0,0)(0,1)(2,1)(2,0) (2,1)[.1]{}[A]{} (3,1)[.1]{}[B]{} (3,0)[.1]{}[C]{} (2,0)[.1]{}[D]{}
The $k \times n$ *cylinder graph* $C_k(n)$ can be obtained by “wrapping” the grid graph around, specifically by adding the following edges: $$E\left(C_k(n)\right) = E\left(G_k(n)\right) \bigcup {\{{\{v_{1,i},v_{n,i}\}} \mid 1 \leq i \leq k\}}.$$ Note that $C_k(n) = P_k \times C_n$.
The $k \times n$ *torus graph* $T_k(n)$ can be obtained by “wrapping” the cylinder graph around the other way, specifically by adding the following edges: $$E\left(T_k(n)\right) = E\left(C_k(n)\right) \bigcup {\{{\{v_{i,1},v_{i,k}\}} \mid 1 \leq i \leq n\}}$$ Note that $T_k(n) = C_k \times C_n$.
Throughout this paper, we will be dealing with partitions of the set $[k] = {\{1, 2, \ldots, k\}}$. We denote by $\mathcal{B}_k$ the set of all such partitions, and $B_k = |\mathcal{B}_k|$ are the Bell numbers. We will impose an ordering on $\mathcal{B}_k$, which we will call the lexicographic ordering on $\mathcal{B}_k$:
Given two partitions $P_1$ and $P_2$ of $[k]$, for $i \in [k]$, let $X_i$ be the block of $P_1$ containing $i$, and likewise $Y_i$ the block of $P_2$ containing $i$. Let $j$ be the minimum value of $i$ such that $X_i \neq Y_i$. Then $P_1 < P_2$ iff
1. $|P_1| < |P_2|$ or
2. $|P_1| = |P_2|$ and $X_j \prec Y_j$, where $\prec$ denotes normal lexicographic ordering.
For example, $\mathcal{B}_3$ in order is $$\mathcal{B}_3 = {\{{\{{\{1,2,3\}}\}}, {\{{\{1\}},{\{2,3\}}\}}, {\{{\{1,2\}},{\{3\}}\}}, {\{{\{1,3\}},{\{2\}}\}}, {\{{\{1\}},{\{2\}},{\{3\}}\}}\}}$$ However, we will use shorthand notation for set partitions as follows: $$\mathcal{B}_3 = {\{123,1/23,12/3,13/2,1/2/3\}}.$$ Since our examples will only deal with $k<10$, we will not have to worry about double-digit numbers on our shorthand notation.
We will find many recurrences in this paper, all pertaining to the number of spanning trees of the graphs mentioned above. Since we will be dealing with each type of graph separately, we will always denote by $T_n$ the number of spanning trees of whatever graph we are dealing with at the moment, which will be unambiguous.
Grid Graphs: The Example For $k = 2$.
=====================================
What follows is mainly from [@Desjarlais] and is the inspiration for the other results on grid graphs. We would like to find a recurrence for $T_n$, which for now will represent the number of spanning trees in $G_2(n)$. If we started out with a spanning tree on $G_2(n-1)$, then there are three different ways to add the additional two vertices to still make a spanning tree on $G_2(n)$:
(0,0)(11,1) (0,0)(0,1)(2,1)(2,0) (2,1)[.1]{}[A]{} (3,1)[.1]{}[B]{} (3,0)[.1]{}[C]{} (2,0)[.1]{}[D]{} (4,0)(4,1)(6,1)(6,0) (6,1)[.1]{}[A]{} (7,1)[.1]{}[B]{} (7,0)[.1]{}[C]{} (6,0)[.1]{}[D]{} (8,0)(8,1)(10,1)(10,0) (10,1)[.1]{}[A]{} (11,1)[.1]{}[B]{} (11,0)[.1]{}[C]{} (10,0)[.1]{}[D]{}
However, there is also a way to create a spanning tree on the $2 \times n$ grid from something that isn’t a spanning tree on $G_k(n-1)$. Let $x = v_{1,n-1}$ and $y = v_{2,n-2}$ be the end vertices on $G_k(n-1)$. If we have a spanning forest on $G_k(n-1)$ with the property that there are two trees in the forest and $x$ and $y$ are in distinct trees, then we can append the following edges to create a spanning tree in $G_k(n)$:
(0,0)(3,1) (0,0)(0,1)(2,1)(2,0) (2,1)[.1]{}[A]{} (3,1)[.1]{}[B]{} (3,0)[.1]{}[C]{} (2,0)[.1]{}[D]{}
Therefore, in counting $T_n$ it is useful to also count $F_n$, which we define as the number of spanning forests in $G_k(n)$ consisting of two trees with the additional property that the end vertices $v_{1,n}$ and $v_{2,n}$ are in distinct trees. From the preceding two paragraphs we can now obtain the recurrence $$T_n = 3T_{n-1} + F_{n-1}$$ and through similar reasoning we can also find the recurrence $$F_n = 2T_{n-1} + F_{n-1}$$ At this point, let us note that we have enough information to find $T_n$ (or $F_n$) in time linear in $n$. However, our goal is to provide explicit recurrences for $T_n$ alone. If we let $v_n$ denote the column vector $$v_n = \left[ \begin{array}{c}
T_n \\
F_n
\end{array} \right]$$ And if we define the matrix $A$ by $$A = \left[\begin{array}{cc}
3 & 1 \\
2 & 1
\end{array}\right]$$ Then we satisfy $$Av_{n-1} = v_n.$$ With the starting conditions $$v_1 = \left[
\begin{array}{c}
1 \\
1
\end{array} \right].$$ The characteristic polynomial of $A$ is $$\chi_\lambda(A) = \lambda^2 - 4\lambda + 1$$ so by the Cayley-Hamilton Theorem, we satisfy $$A^2 - 4A + 1 = 0.$$ This can be re-written as $$A^2 = 4A - 1$$ and if we multiply by the vector $v_n$ on the right we obtain $$\left[ \begin{array}{c} T_{n+2} \\ F_{n+2} \end{array} \right] = 4\left[ \begin{array}{c} T_{n+1} \\ F_{n+1} \end{array} \right] - \left[ \begin{array}{c} T_{n} \\ F_{n} \end{array} \right].$$ Hence, we now see that $T_n$ and $F_n$ satisfy the same recurrence: $$\begin{aligned}
T_{n+2} &= 4T_{n+1} - T_n \\
F_{n+2} &= 4F_{n+1} - F_n\end{aligned}$$ with starting conditions $$\begin{array}{cc}
T_0 = 1 & T_1 = 4 \\
F_0 = 1 & F_1 = 3
\end{array}.$$ We now have all the information we need to obtain more information, such as the generating function and, finally, a closed-form formula for $T_n$. All of these items can be found in [@Desjarlais].
The General Case For Grid Graphs.
=================================
We want to use the same ideas for general $k$, but it requires a bit more bookkeeping. To extend the idea of $F_n$ in the previous section, we need to consider partitions of $[k] = {\{1,2, \ldots, k\}}$ and the forests that come from these partitions.
Given a spanning forest $\mathcal{F}$ of $G_k(n)$, the *partition induced by $\mathcal{F}$* is obtained from the equivalence relation $$i \sim j \iff v_{n,i}, v_{n,j} \mbox{ are in the same tree of } \mathcal{F}.$$
For example, the partition induced by a spanning tree of $G_k(n)$ is $123\cdots n$ and the partition induced by the forest with no edges is $1/2/3/\cdots/n-1/n$.
Given a spanning forest $\mathcal{F}$ of $G_k(n)$ and a partition $P$ of $[k]$, we say that $\mathcal{F}$ *is consistent with* $P$ if:
1. The number of trees in $\mathcal{F}$ is precisely $|P|$.
2. $P$ is the partition induced by $\mathcal{F}$.
Given a graph $G$ on $k$ vertices and a partition $P$ of $[k]$, let $T_G(P,n)$ be the number of spanning trees of the graph $G \times P_n$. We will often omit $G$ when it is clear from the context, or irrelevant. Recall that we have an ordering of partitions, so we will define $T_G(i,n) = T_G(P_i,n)$.
In the previous section, since $B_2 = 2$, we were counting two things: $T_n$ , which corresponds to $T(12,n)$, and $F_n$, which corresponds to $T(1/2,n)$. Therefore, for arbitrary $k$ we are now tasked with counting $B_k$ different objects at once, so we are to find the $B_k \times B_k$ matrix that represents the $B_k$ simultaneous recurrences between these objects.
Define by $E_{n}$ the set of edges $$E_{n} = E(G_k(n)) \setminus E(G_k(n-1))$$ Note that $\left|E_{n}\right| = 2k-1$ edges.
Given some forest $\mathcal{F}$ of $G_k(n-1)$ and some subset $X \subseteq E_{n}$, we can combine the two to make a forest of $G_k(n)$. If we are only interested in the number of trees in the new forest and its induced partition, then we only need to know the same information from $\mathcal{F}$, and this is all independent of $n$. Therefore, we have the following definition:
Given two partitions $P_1$ and $P_2$ in $\mathcal{B}_k$, a subset $X \subseteq E_{n}$ *transfers from $P_1$ to $P_2$* if a forest consistent with $P_1$ becomes a forest consistent with $P_2$ after the addition of $X$.
Figure \[4x4-ex\] shows a spanning forest of $G_4(4)$ where, from left to right, the edges transfer from $1/23/4$ to $1234$, from $1234$ to $12/34$, and from $12/34$ to $1/2/34$.
(0,0)(2,3) (0,0)[0.1]{}[A]{} (0,1)[0.1]{}[B]{} (0,2)[0.1]{}[C]{} (0,3)[0.1]{}[D]{} (1,0)[0.1]{}[E]{} (1,1)[0.1]{}[F]{} (1,2)[0.1]{}[G]{} (1,3)[0.1]{}[H]{} (2,0)[0.1]{}[I]{} (2,1)[0.1]{}[J]{} (2,2)[0.1]{}[K]{} (2,3)[0.1]{}[L]{} (3,0)[0.1]{}[M]{} (3,1)[0.1]{}[N]{} (3,2)[0.1]{}[O]{} (3,3)[0.1]{}[P]{}
Therefore, we can define the $B_k \times B_k$ matrix $A_k$ by the following: $$A_k(i,j) = |{\{A \subseteq E_{n+1} \mid A \mbox{ is compatible from } P_j \mbox{ to } P_i\}}|.$$ The $2 \times 2$ matrix in the previous section is $A_2$. Brute-force search with straightforward Mathematica code [@RaffSite] can produce more matrices:
$$A_3 = \begin{bmatrix}
8 & 3 & 3 & 4 & 1 \\
4 & 3 & 2 & 2 & 1 \\
4 & 2 & 3 & 2 & 1 \\
1 & 0 & 0 & 1 & 0 \\
3 & 2 & 2 & 2 & 1
\end{bmatrix}$$
$$A_4 =\left[ \begin{array}{p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}}
21 & 8 & 9 & 11 & 8 & 14 & 11 & 15 & 3 & 3 & 4 & 3 & 4 & 5 & 1 \\
9 & 8 & 6 & 4 & 4 & 6 & 5 & 8 & 3 & 3 & 4 & 2 & 2 & 2 & 1 \\
6 & 4 & 9 & 4 & 4 & 4 & 4 & 4 & 3 & 2 & 2 & 3 & 2 & 2 & 1 \\
3 & 0 & 0 & 3 & 1 & 2 & 1 & 2 & 0 & 0 & 0 & 0 & 1 & 1 & 0 \\
9 & 4 & 6 & 5 & 8 & 6 & 4 & 8 & 2 & 3 & 2 & 3 & 4 & 2 & 1 \\
1 & 0 & 0 & 1 & 0 & 3 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
3 & 1 & 0 & 1 & 0 & 2 & 3 & 2 & 0 & 0 & 1 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
5 & 4 & 6 & 4 & 3 & 4 & 3 & 4 & 3 & 2 & 2 & 2 & 2 & 2 & 1 \\
5 & 4 & 4 & 3 & 4 & 6 & 3 & 4 & 2 & 3 & 2 & 2 & 2 & 2 & 1 \\
1 & 1 & 0 & 0 & 0 & 0 & 1 & 2 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\
5 & 3 & 6 & 3 & 4 & 4 & 4 & 4 & 2 & 2 & 2 & 3 & 2 & 2 & 1 \\
1 & 0 & 0 & 1 & 1 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
1 & 0 & 0 & 1 & 0 & 2 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
4 & 3 & 4 & 3 & 3 & 4 & 3 & 4 & 2 & 2 & 2 & 2 & 2 & 2 & 1
\end{array} \right]$$
$A_5,A_6,$ and $A_7$ have also been found; they are shown in [@RaffSite]. Once these matrices are known, then everything about the sequence of spanning trees can be found. The following table shows some results obtained for grid graphs; results obtained for arbitrary graphs of the form $G \times P_n$ for all graphs $G$ with at most five vertices are in [@RaffSite].
-----------------------------------------------------------------------------------------------------------------------------------------------------
$G_2(n):$ ([@Desjarlais])
$T_n = 4T_{n-1}-T_{n-2}$
Sequence: ${\{1,4,15,56,209, \ldots\}}$ (OEIS A001353)
Generating Function: $\frac{x}{1 - 4x + x^2}$
$G_3(n):$ ([@FaaseSite])
$T_n = 15T_{n-1} - 32T_{n-2} + 15T_{n-3} - T_{n-4}$
Sequence: ${\{1, 15, 192, 2415, 30305, \ldots\}}$ (OEIS A006238)
Generating Function: $\frac{3x(1 + 49x + 1152x^2)}{1 + 24x - 24x^2 + x^3}$
$G_4(n):$ ([@FaaseSite])
$T_n = 56T_{n-1} - 672T_{n-2} + 2632T_{n-3} - 4094T_{n-4} + 2632T_{n-5} - 672T_{n-6} + 56T_{n-7} - T_{n-8}$
Sequence: ${\{1, 56, 2415, 100352, 4140081, \ldots\}}$ (OEIS A003696)
Generating Function: $\frac{16x(1 + 12x + x^2)}{1 - 204x + 1190x^2 - 204x^3 + x^4}$
$G_5(n):$ ([@FaaseSite], with improvements from this paper)
$T_n = 209T_{n-1} - 11936T_{n-2} + 274208T_{n-3} - 3112032T_{n-4} + 19456019T_{n-5}$
$- 70651107T_{n-6} + 152325888T_{n-7} - 196664896T_{n-8} + 152325888T_{n-9}$
$- 70651107T_{n-10} + 19456019T_{n-11} - 3112032T_{n-12} + 274208T_{n-13}$
$- 11936T_{n-14} + 209T_{n-15} - T_{n-16}$
Sequence: ${\{1, 209, 30305, 4140081, 557568000, \ldots\}}$ (OEIS A003779)
Generating Function: $\frac{125x(1 + 4656x + 10616686x^2 + 23432228161x^3 + 51714958501250x^4)}{1 + 2255x - 105985x^2 + 105985x^3 - 2255x^4 + x^5}$
$G_6(n):$ (new)
$T_n = 780 T_{n-1}-194881 T_{n-2}+22377420 T_{n-3}-1419219792 T_{n-4}$
$+ 55284715980T_{n-5}-1410775106597 T_{n-6}+24574215822780 T_{n-7} $
$-300429297446885 T_{n-8}+2629946465331120 T_{n-9}-16741727755133760 T_{n-10} $
$+78475174345180080 T_{n-11}-273689714665707178 T_{n-12}+716370537293731320 T_{n-13} $
$-1417056251105102122 T_{n-14}+2129255507292156360 T_{n-15}-2437932520099475424 T_{n-16} $
$+2129255507292156360 T_{n-17}-1417056251105102122 T_{n-18}+716370537293731320 T_{n-19} $
$-273689714665707178 T_{n-20}+78475174345180080 T_{n-21}-16741727755133760 T_{n-22} $
$+2629946465331120 T_{n-23}-300429297446885 T_{n-24}+24574215822780 T_{n-25} $
$-1410775106597 T_{n-26} +55284715980 T_{n-27}-1419219792 T_{n-28}+22377420 T_{n-29} $
$-194881 T_{n-30}+780 T_{n-31}-T_{n-32} $
Sequence: ${\{1, 780, 380160, 170537640, 74795194705, \ldots\}}$ (OEIS A139400)
Generating Function: See [@RaffSite]
-----------------------------------------------------------------------------------------------------------------------------------------------------
Extending to Generalized Graphs of the Form $G \times P_n$
==========================================================
For the results above, it was not necessary that the graph we were dealing with was a grid. We could have repeated the same process as above for any sequences of graphs $G_n$ defined by $$G_n = G \times P_n$$ for some predefined graph $G$. In fact, the Mathematica code in the appendix handles any such general case. Therefore, it leads to the following theorem:
Let a graph $G$ be given with $k$ vertices, and define the sequence of graphs ${\{G_n\}}$ by $G_n = G \times P_n$. Then there is a $B_k \times B_k$ matrix $M$ and a vector $v$, both taking on integer values, such that $$T_n = M^nv[1]$$ where $T_n$ is the number of spanning trees in $G_n$. Furthermore, $M^nv[i]$ lists the number of spanning forests consistent with $P_i$ in $G_n$.
Let a graph $G$ be given with $k$ vertices, and consider the sequence ${\{T_n\}}$. Then $T_n$ satisfies a linear recurrence of order $B_k$.
From investigations, we have a few conjectures:
For the matrix $M$ given in the theorem above, the characteristic polynomial $\chi_\lambda(M)$ factors over the integers into monomials whose degree is always a power of 2.
For any graph $G$, the recurrence ${\{T_n\}}$ satisfies a linear recurrence whose coefficients alternate in sign.
The recurrence for the grid graph $G_k(n)$ has order $2^{k-1}$.
The recurrence for the graph $K_k \times P_n$ has order $k$.
For the time being, we will only prove the special case of Conjecture 3 for the grid graphs $G_2(n)$. We will give a combinatorial proof that we hope can be adjusted accordingly to the higher cases. To aid in the proof, we will introduce the concept of *grid addition*, which is simply a shorthand way of creating the union of two grids.
If $G_1$ is a $k \times n_1$ grid and $G_2$ is a $k \times n_2$ grid, then $G_1 + G_2$ is the $k \times (n_1 + n_2 - 1)$ grid defined as the graph obtained by identifying the right-most vertices of $G_1$ with the left-most vertices of $G_2$. Any overlapping edges remain.
Figure \[grid-add\] shows the addition of a tree on $G_2(3)$ with a tree on $G_2(2)$ to obtain a subgraph of $G_2(4)$.
(0,0)(10,1) (0,0)[.1]{}[A]{} (0,1)[.1]{}[B]{} (1,0)[.1]{}[C]{} (1,1)[.1]{}[D]{} (2,0)[.1]{}[E]{} (2,1)[.1]{}[F]{} (3,0.5)[$+$]{} (4,0)[.1]{}[G]{} (4,1)[.1]{}[H]{} (5,0)[.1]{}[I]{} (5,1)[.1]{}[J]{} (6,0.5)[$=$]{} (7,0)[.1]{}[A]{} (7,1)[.1]{}[B]{} (8,0)[.1]{}[C]{} (8,1)[.1]{}[D]{} (9,0)[.1]{}[E]{} (9,1)[.1]{}[F]{} (9,0)[.1]{}[G]{} (9,1)[.1]{}[H]{} (10,0)[.1]{}[I]{} (10,1)[.1]{}[J]{}
The number of spanning trees of the graphs $G_2(n)$ satisfies the linear recurrence $T_n = 4T_{n-1} - T_{n-2}$ with the initial conditions $T_1 = 1$, $T_2 = 4$.
Showing the initial conditions is a minor exercise. We will prove this recurrence in the equivalent form $T_n + T_{n-2} = 4T_{n-1}$. Let $\mathcal{T}_k$ denote the set of spanning trees of the graph $G_2(k)$. We will associate $T_{n-2}$ with the set $\mathcal{T}_{n-2}$ with an addition at the end, as shown by Figure \[int-T-n-2\].
(6,1) (0,0)(0,1)(4,1)(4,0) (4,0)[.1]{}[A]{} (4,1)[.1]{}[B]{} (5,0)[.1]{}[C]{} (5,1)[.1]{}[D]{} (6,0)[.1]{}[E]{} (6,1)[.1]{}[F]{}
In this way, we can think of $\mathcal{T}_{n-2}$ as being trees of $G_2(n)$. Similarly, as Figure \[int-4T-n-1\] shows, we will associate $4T_{n-1}$ with the set of trees from $\mathcal{T}_{n-1}$ with each of the four trees of $G_2(2)$ added at the end.
(9,3) (0,0)(0,1)(2,1)(2,0) (2.5,0.5)[$+$]{} (3,0)[.1]{}[A]{} (3,1)[.1]{}[B]{} (4,0)[.1]{}[C]{} (4,1)[.1]{}[D]{} (5,0)(5,1)(7,1)(7,0) (7.5,0.5)[$+$]{} (8,0)[.1]{}[A]{} (8,1)[.1]{}[B]{} (9,0)[.1]{}[C]{} (9,1)[.1]{}[D]{} (0,2)(0,3)(2,3)(2,2) (2.5,2.5)[$+$]{} (3,2)[.1]{}[A]{} (3,3)[.1]{}[B]{} (4,2)[.1]{}[C]{} (4,3)[.1]{}[D]{} (5,2)(5,3)(7,3)(7,2) (7.5,2.5)[$+$]{} (8,2)[.1]{}[A]{} (8,3)[.1]{}[B]{} (9,2)[.1]{}[C]{} (9,3)[.1]{}[D]{}
If we have a tree from $\mathcal{T}_n$, then we can decompose it depending on what the ending of the tree looks like. Figure \[decomp-Tn\] shows all of the possibilities, along with their decompositions. Note that the decompositions are of the same form as we dictated for $4T_{n-1}$.
(10,13) (0,0)(0,1)(2,1)(2,0) (2,0)[.1]{}[A]{} (2,1)[.1]{}[B]{} (3,0)[.1]{}[C]{} (3,1)[.1]{}[D]{} (4.5,0.5)[$\to$]{} (6,0)(6,1)(8,1)(8,0) (8.5,0.5)[$+$]{} (8,0)[.1]{}[E1]{} (8,1)[.1]{}[E2]{} (9,0)[.1]{}[A]{} (9,1)[.1]{}[B]{} (10,0)[.1]{}[C]{} (10,1)[.1]{}[D]{} (0,2)(0,3)(2,3)(2,2) (2,2)[.1]{}[A]{} (2,3)[.1]{}[B]{} (3,2)[.1]{}[C]{} (3,3)[.1]{}[D]{} (4.5,2.5)[$\to$]{} (6,2)(6,3)(8,3)(8,2) (8.5,2.5)[$+$]{} (8,2)[.1]{}[E1]{} (8,3)[.1]{}[E2]{} (9,2)[.1]{}[A]{} (9,3)[.1]{}[B]{} (10,2)[.1]{}[C]{} (10,3)[.1]{}[D]{} (0,4)(0,5)(2,5)(2,4) (2,4)[.1]{}[A]{} (2,5)[.1]{}[B]{} (3,4)[.1]{}[C]{} (3,5)[.1]{}[D]{} (4.5,4.5)[$\to$]{} (6,4)(6,5)(8,5)(8,4) (8.5,4.5)[$+$]{} (8,4)[.1]{}[E1]{} (8,5)[.1]{}[E2]{} (9,4)[.1]{}[A]{} (9,5)[.1]{}[B]{} (10,4)[.1]{}[C]{} (10,5)[.1]{}[D]{} (0,6)(0,7)(2,7)(2,6) (2,6)[.1]{}[A]{} (2,7)[.1]{}[B]{} (3,6)[.1]{}[C]{} (3,7)[.1]{}[D]{} (4.5,6.5)[$\to$]{} (6,6)(6,7)(8,7)(8,6) (8.5,6.5)[$+$]{} (8,6)[.1]{}[E1]{} (8,7)[.1]{}[E2]{} (9,6)[.1]{}[A]{} (9,7)[.1]{}[B]{} (10,6)[.1]{}[C]{} (10,7)[.1]{}[D]{} (0,8)(0,9)(2,9)(2,8) (2,8)[.1]{}[A]{} (2,9)[.1]{}[B]{} (3,8)[.1]{}[C]{} (3,9)[.1]{}[D]{} (4.5,8.5)[$\to$]{} (6,8)(6,9)(8,9)(8,8) (8.5,8.5)[$+$]{} (8,8)[.1]{}[E1]{} (8,9)[.1]{}[E2]{} (9,8)[.1]{}[A]{} (9,9)[.1]{}[B]{} (10,8)[.1]{}[C]{} (10,9)[.1]{}[D]{} (0,10)(0,11)(2,11)(2,10) (2,10)[.1]{}[A]{} (2,11)[.1]{}[B]{} (3,10)[.1]{}[C]{} (3,11)[.1]{}[D]{} (4.5,10.5)[$\to$]{} (6,10)(6,11)(8,11)(8,10) (8.5,10.5)[$+$]{} (8,10)[.1]{}[E1]{} (8,11)[.1]{}[E2]{} (9,10)[.1]{}[A]{} (9,11)[.1]{}[B]{} (10,10)[.1]{}[C]{} (10,11)[.1]{}[D]{} (0,12)(0,13)(2,13)(2,12) (2,12)[.1]{}[A]{} (2,13)[.1]{}[B]{} (3,12)[.1]{}[C]{} (3,13)[.1]{}[D]{} (4.5,12.5)[$\to$]{} (6,12)(6,13)(8,13)(8,12) (8.5,12.5)[$+$]{} (8,12)[.1]{}[E1]{} (8,13)[.1]{}[E2]{} (9,12)[.1]{}[A]{} (9,13)[.1]{}[B]{} (10,12)[.1]{}[C]{} (10,13)[.1]{}[D]{}
Similarly, if we have a tree from $\mathcal{T}_{n-2}$ modified as explained above, then Figure \[decomp-Tn-2\] shows the decomposition. Again, note that the decompositions are of the same form as we dictated for $4T_{n-1}$.
(12,1) (0,0)(0,1)(2,1)(2,0) (2,0)[.1]{}[A]{} (2,1)[.1]{}[B]{} (3,0)[.1]{}[C]{} (3,1)[.1]{}[D]{} (4,0)[.1]{}[E]{} (4,1)[.1]{}[F]{} (5.5,0.5)[$\to$]{} (7,0)(7,1)(9,1)(9,0) (10.5,0.5)[$+$]{} (9,0)[.1]{}[A]{} (9,1)[.1]{}[B]{} (10,0)[.1]{}[E1]{} (10,1)[.1]{}[E2]{} (11,0)[.1]{}[C]{} (11,1)[.1]{}[D]{} (12,0)[.1]{}[E]{} (12,1)[.1]{}[F]{}
The reader can verify that the map described is invertible, yielding the desired bijection.
Extending to Cylinder Graphs
============================
In this section we will discuss the changes necessary to extend the above arguments to find recurrences for cylinder graphs and generalized cylinder graphs. We shall take advantage of the “unhooking” technique covered in [@Golin]. The technique is a reduction from a cylinder graph to a grid graph. Recall that the vertex sets of $C_k(n)$ and $G_k(n)$ are the same.
For a given $k$, we define $\mathcal{E}_k$ by $$\mathcal{E}_k = E(C_k(n))\setminus E(G_k(n))$$
If we unhook (i.e. remove) the edges in $\mathcal{E}_k$ then what we have left is precisely $G_k(n)$. Now we have to consider what structures in $G_k(n)$ yield a spanning tree in $C_k(n)$ by the addition of some subset of edges from $\mathcal{E}_k$. Since we are going to add edges that go from one end of the grid to another, we must look at both ends of the grid now, as opposed to only looking at one end. For example, Figure \[cyl-ex\] shows a spanning forest of $G_3(3)$ will never yield a spanning tree of $G_3(n)$ for any $n > 3$ through the method described in the previous sections, but this spanning forest would create two different spanning trees of $C_3(3)$ through the addition of either edge $v_{1,1}v_{3,1}$ or $v_{1,2}v_{3,2}$.
(0,0)(2,2) (2,2)(2.25,2.25)(-.25,2.25)(0,2) (2,1)(2.25,1.25)(-.25,1.25)(0,1) (2,0)(2.25,0.25)(-.25,0.25)(0,0) (0,0)[0.1]{}[A]{} (0,1)[0.1]{}[B]{} (0,2)[0.1]{}[C]{} (1,0)[0.1]{}[D]{} (1,1)[0.1]{}[E]{} (1,2)[0.1]{}[F]{} (2,0)[0.1]{}[G]{} (2,1)[0.1]{}[H]{} (2,2)[0.1]{}[I]{}
Therefore, we can keep the same basic idea used with grid graphs, with some modifications. We must now keep track of how our spanning forests affects the vertices *at each end*.
Given a spanning forest $\mathcal{F}$ of $G_k(n)$, the partition $P$ of $[2k]$ *induced by* $\mathcal{F}$ is obtained from the equivalence relation $$i \sim j \iff v_i, v_j \mbox{ are in the same tree of } \mathcal{F}$$ where we identify the vertices $v_1, v_2, \ldots , v_k$ with $v_{1,1}, v_{1,2}, \ldots , v_{1,k}$, respectively, and the vertices $v_{k+1},v_{k+2},\ldots,v_{2k}$ with $v_{n,1}, v_{n,2}, \ldots, v_{n,k}$, respectively.
Given a spanning forest $\mathcal{F}$ of $G_k(n)$ and a partition $P$ of $[2k]$, we say that $\mathcal{F}$ *is cylindrically consistent with* $P$ if:
1. The number of trees in $\mathcal{F}$ is precisely $|P|$.
2. $P$ is the partition induced by $\mathcal{F}$.
For example, the forest shown in Figure \[cyl-ex\] is consistent with the partition $12/3456$. It’s important to know what partition a certain forest of $G_k(n)$ is cylindrically consistent with, as that determines how many different ways edges can be added to achieve a spanning tree of $C_k(n)$. Since each spanning tree of $C_k(n)$ is uniquely determined by the underlying spanning forest of $G_k(n)$ and the extra edges from $\mathcal{E}_k$, we have all the information we need to count the number of spanning trees of $C_k(n)$.
For a given $k$, the *tree-counting vector* $d_k$ is the vector, indexed by the partitions of $[2k]$, such that $d_k(i)$ is the number of ways that edges from $E(C_k(n)) \setminus E(G_k(n))$ can be added to get from a forest cylindrically consistent with partition $i$ to a spanning tree of $C_k(n)$. Notice that this is independent of $n$.
It can be verified that the following information produces $d_2$: $$\begin{array}{|c|c|}
\hline
1234 & 1 \\
\hline
1/234 & 1 \\
\hline
12/34 & 2 \\
\hline
134/2 & 1 \\
\hline
123/4 & 1 \\
\hline
14/23 & 2 \\
\hline
124/3 & 1 \\
\hline
13/24 & 0 \\
\hline
1/2/34 & 1 \\
\hline
1/23/4 & 1 \\
\hline
1/24/3 & 0 \\
\hline
12/3/4 & 1 \\
\hline
13/2/4 & 0 \\
\hline
14/2/3 & 1 \\
\hline
1/2/3/4 & 0 \\
\hline
\end{array}$$ $$d_2 = (1,1,2,1,1,2,1,0,1,1,0,1,0,1,0)$$ To count the number of spanning trees for $C_k(n)$ we can produce the $B_{2k} \times B_{2k}$ matrix in the same way as we did for the grid graphs, and using this matrix we can find the number of spanning forests of $G_k(n)$ consistent with each of the partitions of $\mathcal{B}_{2k}$, which can be expressed as a vector of length $B_{2k}$. Then, when we take the dot product of this vector with $d_k$, we obtain the number of spanning trees of $C_k(n)$. For example, it can be verified that the following is the matrix related to $C_2(n)$: $$A =
\left[\begin{array}{p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}p{6pt}}
3 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 3 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\
1 & 0 & 3 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 3 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 1 & 0 \\
1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 3 & 1 & 1 & 0 & 1 & 1 & 1 \\
0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\
1 & 0 & 2 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 2 & 1 & 1 & 0 & 1 & 1 & 1
\end{array} \right]$$ The initial vector is as follows: $$v = (1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0)$$ We then obtain $$\begin{aligned}
(Av)\cdot d_2 &= 12 \\
(A^2v)\cdot d_2 &= 75 \\
(A^3v) \cdot d_2 &= 384 \\
\vdots\end{aligned}$$ which yields the sequence of the number of spanning trees on $C_2(n)$.
Similar to the process with grids, there is nothing specific here to the simple cylinder graph - these methods can be used to obtain sequences for graph families of the form $G \times C_n$ for arbitrary $G$. However, due to the rapid growth of $B_{2k}$, the ability to find the appropriate matrices becomes somewhat impossible starting at graphs with five vertices. Nevertheless, we still have the following:
For a given graph $G$ on $k$ vertices, there is a $B_{2k} \times B_{2k}$ matrix $M$ and a vector $v$ of length $B_{2k}$ such that $$(M^nv)\cdot d_k$$ is the number of spanning trees of the graph $G \times C_n$.
For a given graph $G$ on $k$ vertices, the number of spanning trees ${\{T_n\}}$ of $G \times C_n$ satisfies a linear recurrence of order at most $B_{2k}$.
Although the sequence for $C_2(n)$ is already known, these methods used were able to obtain sequences for $C_3(n)$ and $K_3 \times C_n$, which we now state:
----------------------------------------------------------------------------------------------------------------
$C_2(n):$ ([@FaaseSite], with improvements)
$T_n = 10T_{n-1} - 35T_{n-2} + 52T_{n-3} - 35T_{n-4} + 10T_{n-5} - T_{n-6}$
Sequence: ${\{1, 12, 75, 384, 1805, \ldots\}}$ (OEIS A006235)
Generating Function: $\frac{x (1 + 2 x - 10 x^2 + 2 x^3 + x^4)}{(-1 + 5 x - 5 x^2 + x^3)^2}$
$C_3(n):$ (new)
$T_n = 48T_{n-1} - 960T_{n-2} + 10622T_{n-3} - 73248T_{n-4} + 335952T_{n-5} - 1065855T_{n-6} + 2396928T_{n-7}$
$-3877536T_{n-8} + 4548100T_{n-9} - 3877536T_{n-10} + 2396928T_{n-11} - 1065855T_{n-12} + 335952T_{n-13}$
$-73248T_{n-14} + 10622T_{n-15} - 960T_{n-16} + 48T_{n-17} - T_{n-18}$
Sequence: ${\{1, 70, 1728, 31500, 508805, \ldots\}}$ (OEIS to be submitted)
Generating Function: See [@RaffSite]
$K_3 \times P_n:$ (new)
$T_n = 58T_{n-1} - 1131T_{n-2} + 8700T_{n-3} - 29493T_{n-4} + 43734T_{n-5} - 29493T_{n-6} + 8700T_{n-7} $
$- 1131T_{n-8} + 58T_{n-9} - T_{n-10}$
Sequence: ${\{3, 318, 12960, 410700, 11870715, \ldots\}}$ (OEIS to be submitted)
Generating Function: $\frac{3 x (1 + 48 x - 697 x^2 - 2474 x^3 + 9918 x^4 + 62 x^5 - 2045 x^6 +
96 x^7 + 5 x^8)}{(-1 + 29 x - 145 x^2 + 145 x^3 - 29 x^4 + x^5)^2}$
----------------------------------------------------------------------------------------------------------------
Acknowledgements
================
Special thanks to Andrew Baxter for thoroughly reviewing the paper and suggesting many helpful additions. Thanks also to Prof. Doron Zeilberger for taking me on as his seventh concurrent student, even though his self-proclaimed limit is four.
|
{
"pile_set_name": "ArXiv"
}
|
[**The Optical Gravitational Lensing Experiment.\
OGLE-III Photometric Maps of the Large Magellanic Cloud[^1]**]{} 1.5cm [**A. U d a l s k i$^1$, I. S o s z y ñ s k i$^1$, M.K. S z y m a ñ s k i$^1$, M. K u b i a k$^1$,\
G. P i e t r z y ñ s k i$^{1,2}$, £. W y r z y k o w s k i$^{3,1}$, O. S z e w c z y k$^{2,1}$,\
K. U l a c z y k$^1$ and R. P o l e s k i$^1$**]{} $^1$Warsaw University Observatory, Al. Ujazdowskie 4, 00-478 Warszawa, Poland\
e-mail: (udalski,soszynsk,msz,mk,pietrzyn,wyrzykow,szewczyk,kulaczyk,rpoleski) @astrouw.edu.pl\
$^2$ Universidad de Concepci[ó]{}n, Departamento de Fisica, Casilla 160–C, Concepci[ó]{}n, Chile\
$^3$ Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK
ABSTRACT
We present the OGLE-III Photometric Maps of the Large Magellanic Cloud. They cover about 40 square degrees of the LMC and contain mean, calibrated [*VI*]{} photometry and astrometry of about 35 million stars observed during seven observing seasons of the third phase of the Optical Gravitational Lensing Experiment – OGLE-III.
We discuss the quality of data and present color–magnitude diagrams of selected fields. The OGLE-III Photometric Maps of the LMC are available to the astronomical community from the OGLE Internet archive.
[ ]{}[*Magellanic Clouds – Surveys – Catalogs – Techniques: photometric*]{}
Introduction
============
One of the important results of the second phase of the Optical Gravitational Lensing Experiment was publication of the OGLE Photometric Maps of Dense Stellar Regions (Udalski [[*et al. *]{}]{}1998, 2000, 2002) containing precisely calibrated [*BVI*]{} photometry and astrometry of the SMC, LMC and Galactic bulge fields observed during OGLE-II. Because these regions of the sky are extremely interesting from the astrophysical point of view, OGLE maps have been widely used by astronomers worldwide for many astrophysical applications.
OGLE-III phase of the OGLE project, started on June 12, 2001 and still in operation, was a significant extension of the OGLE survey. Much larger observing capabilities made it possible to cover practically entire area of the LMC and SMC and large fraction of the Galactic bulge. After seven years of continuous observations the huge collection of OGLE images was re-reduced to obtain the final precise photometry, calibrated to the standard system (Udalski [[*et al. *]{}]{}2008). Thus it became possible to extend the OGLE-II maps to new regions, not observed by OGLE before.
In this paper we present the OGLE-III photometric maps of the Large Magellanic Cloud. They are available to the astronomical community from the OGLE Internet archive.
Observations
============
OGLE-III images of the LMC used for construction of the OGLE-III Photometric Maps of this galaxy were collected between July 2001 and March 2008 and cover seven observing seasons of the LMC. Observations were carried out at Las Campanas Observatory, operated by the Carnegie Institution of Washington, with 1.3 m Warsaw telescope equipped with the eight chip mosaic camera (Udalski 2003). One full mosaic image covers approximately $35{\ifmmode{'}\else$'$\fi}\times35{\ifmmode{'}\else$'$\fi}$ on the sky with the scale of 0[$''$]{}26/pixel.
As the main goal of the OGLE survey is the search for variability of observed objects, the majority of observations were obtained through the single, namely [*I*]{}-band filter. Although this filter well approximates the standard one for $V-I<3$ mag colors, one has to be aware that for very red objects some deviation may be present. Several hundred [*I*]{}-band images were collected for each of the observed fields. From time to time the LMC fields were also observed in the [*V*]{}-band – typically about 40–50 times in the considered period. The exposure time was set to 180 and 240 seconds in the [*I*]{} and [*V*]{}-band, respectively.
Because the majority of observed fields have high stellar density, observations were conducted only in good seeing and transparency conditions. When the seeing exceeded 1[$''$]{}8 observations were stopped. The median seeing of the [*V*]{} and [*I*]{}-band OGLE-III LMC datasets is equal to 1[$''$]{}2.
Table 1 lists all LMC fields observed during OGLE-III phase. It also provides equatorial coordinates of their centers and number of stars detected in the [*I*]{}-band. The total observed area reaches 40 square degrees. Fig. 1 presents a combined image of the LMC taken by the ASAS survey program (Pojmañski, 1997) with contours of the OGLE-II and OGLE-III fields.
Data Reductions
===============
Initial pre-reductions of the collected images – de-biasing, flatfielding – are done at the telescope, immediately after the last pixel of the image is written to the main data acquisition computer (Udalski 2003). Although the first, provisional photometry is done also in almost real time at the telescope, photometry used in this paper comes from the final off-line re-reductions of the entire LMC dataset collected so far. Photometry is based on the Difference Image Analysis method (DIA – Alard and Lupton 1998, Wo¼niak 2000) and all details on the OGLE-III implementation and calibration to the standard system can be found in Udalski [[*et al. *]{}]{}(2008). Comparison with the OGLE-II photometric maps indicates that the mean difference of the magnitudes between the calibrated OGLE-III and OGLE-II photometry for about 800000 stars brighter than $I<18$ mag and $V<19$ mag in overlaping fields is negligible ($-0.004\pm0.013,
-0.003\pm0.013$ in the [*I*]{} and [*V*]{}, respectively when calibrating directly with standard stars from Landolt’s (1992) fields or $0.001\pm0.010, 0.000\pm0.007$ for final calibrations of OGLE-III photometry with OGLE-II maps) implying that OGLE-II and OGLE-III maps are photometrically fully consistent.
Astrometric transformation of the pixel grid to equatorial system was done in similar way as in the OGLE-II maps. Details can also be found in Udalski [[*et al. *]{}]{}(2008).
[cccrccccr]{} Field & RA & DEC & $N_{\rm Stars}$ &$\phantom{xxxxx}$& Field & RA & DEC & $N_{\rm Stars}$\
& (2000) & (2000) & && & (2000) & (2000) &\
LMC100 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}2 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}07{\ifmmode{''}\else$''$\fi}$ & 1075711 && LMC158 & 4[[\^[h]{}$^{\rm h}$]{}]{}30[[\^[m]{}$^{\rm m}$]{}]{}59[[\^[s]{}$^{\rm s}$]{}]{}9 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}01{\ifmmode{''}\else$''$\fi}$ & 34264\
LMC101 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}03[[\^[s]{}$^{\rm s}$]{}]{}1 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}19{\ifmmode{''}\else$''$\fi}$ & 548707 && LMC159 & 5[[\^[h]{}$^{\rm h}$]{}]{}25[[\^[m]{}$^{\rm m}$]{}]{}11[[\^[s]{}$^{\rm s}$]{}]{}4 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}58{\ifmmode{''}\else$''$\fi}$ & 234254\
LMC102 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}03[[\^[s]{}$^{\rm s}$]{}]{}4 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}48{\ifmmode{''}\else$''$\fi}$ & 266963 && LMC160 & 5[[\^[h]{}$^{\rm h}$]{}]{}25[[\^[m]{}$^{\rm m}$]{}]{}20[[\^[s]{}$^{\rm s}$]{}]{}9 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}24{\ifmmode{''}\else$''$\fi}$ & 437897\
LMC103 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}9 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}26{\ifmmode{''}\else$''$\fi}$ & 833034 && LMC161 & 5[[\^[h]{}$^{\rm h}$]{}]{}25[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}5 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}59{\ifmmode{''}\else$''$\fi}$ & 787467\
LMC104 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}4 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}03{\ifmmode{''}\else$''$\fi}$ & 472959 && LMC162 & 5[[\^[h]{}$^{\rm h}$]{}]{}25[[\^[m]{}$^{\rm m}$]{}]{}43[[\^[s]{}$^{\rm s}$]{}]{}3 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}24{\ifmmode{''}\else$''$\fi}$ & 1171172\
LMC105 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}6 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}31{\ifmmode{''}\else$''$\fi}$ & 353469 && LMC163 & 5[[\^[h]{}$^{\rm h}$]{}]{}25[[\^[m]{}$^{\rm m}$]{}]{}52[[\^[s]{}$^{\rm s}$]{}]{}2 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}55{\ifmmode{''}\else$''$\fi}$ & 806912\
LMC106 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}0 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}36{\ifmmode{'}\else$'$\fi}57{\ifmmode{''}\else$''$\fi}$ & 236384 && LMC164 & 5[[\^[h]{}$^{\rm h}$]{}]{}26[[\^[m]{}$^{\rm m}$]{}]{}08[[\^[s]{}$^{\rm s}$]{}]{}4 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}23{\ifmmode{''}\else$''$\fi}$ & 381352\
LMC107 & 5[[\^[h]{}$^{\rm h}$]{}]{}13[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}5 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}52{\ifmmode{'}\else$'$\fi}57{\ifmmode{''}\else$''$\fi}$ & 226287 && LMC165 & 5[[\^[h]{}$^{\rm h}$]{}]{}26[[\^[m]{}$^{\rm m}$]{}]{}20[[\^[s]{}$^{\rm s}$]{}]{}9 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}37{\ifmmode{'}\else$'$\fi}01{\ifmmode{''}\else$''$\fi}$ & 329028\
LMC108 & 5[[\^[h]{}$^{\rm h}$]{}]{}13[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}9 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}40{\ifmmode{''}\else$''$\fi}$ & 257801 && LMC166 & 5[[\^[h]{}$^{\rm h}$]{}]{}31[[\^[m]{}$^{\rm m}$]{}]{}20[[\^[s]{}$^{\rm s}$]{}]{}1 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}51{\ifmmode{''}\else$''$\fi}$ & 314715\
LMC109 & 5[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}53[[\^[s]{}$^{\rm s}$]{}]{}3 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}04{\ifmmode{'}\else$'$\fi}06{\ifmmode{''}\else$''$\fi}$ & 294145 && LMC167 & 5[[\^[h]{}$^{\rm h}$]{}]{}31[[\^[m]{}$^{\rm m}$]{}]{}39[[\^[s]{}$^{\rm s}$]{}]{}6 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}32{\ifmmode{''}\else$''$\fi}$ & 394585\
LMC110 & 5[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}43[[\^[s]{}$^{\rm s}$]{}]{}6 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}42{\ifmmode{''}\else$''$\fi}$ & 529605 && LMC168 & 5[[\^[h]{}$^{\rm h}$]{}]{}32[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}4 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}00{\ifmmode{''}\else$''$\fi}$ & 636559\
LMC111 & 5[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}7 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}02{\ifmmode{''}\else$''$\fi}$ & 696121 && LMC169 & 5[[\^[h]{}$^{\rm h}$]{}]{}32[[\^[m]{}$^{\rm m}$]{}]{}22[[\^[s]{}$^{\rm s}$]{}]{}8 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}26{\ifmmode{''}\else$''$\fi}$ & 1092986\
LMC112 & 5[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}21[[\^[s]{}$^{\rm s}$]{}]{}5 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}21{\ifmmode{''}\else$''$\fi}$ & 752907 && LMC170 & 5[[\^[h]{}$^{\rm h}$]{}]{}32[[\^[m]{}$^{\rm m}$]{}]{}48[[\^[s]{}$^{\rm s}$]{}]{}1 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}53{\ifmmode{''}\else$''$\fi}$ & 878425\
LMC113 & 5[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}10[[\^[s]{}$^{\rm s}$]{}]{}9 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}48{\ifmmode{''}\else$''$\fi}$ & 561679 && LMC171 & 5[[\^[h]{}$^{\rm h}$]{}]{}33[[\^[m]{}$^{\rm m}$]{}]{}10[[\^[s]{}$^{\rm s}$]{}]{}6 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}30{\ifmmode{''}\else$''$\fi}$ & 513519\
LMC114 & 5[[\^[h]{}$^{\rm h}$]{}]{}11[[\^[m]{}$^{\rm m}$]{}]{}58[[\^[s]{}$^{\rm s}$]{}]{}9 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}22{\ifmmode{''}\else$''$\fi}$ & 221912 && LMC172 & 5[[\^[h]{}$^{\rm h}$]{}]{}33[[\^[m]{}$^{\rm m}$]{}]{}34[[\^[s]{}$^{\rm s}$]{}]{}4 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}36{\ifmmode{'}\else$'$\fi}54{\ifmmode{''}\else$''$\fi}$ & 423099\
LMC115 & 5[[\^[h]{}$^{\rm h}$]{}]{}07[[\^[m]{}$^{\rm m}$]{}]{}09[[\^[s]{}$^{\rm s}$]{}]{}7 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}52{\ifmmode{'}\else$'$\fi}59{\ifmmode{''}\else$''$\fi}$ & 234177 && LMC173 & 5[[\^[h]{}$^{\rm h}$]{}]{}37[[\^[m]{}$^{\rm m}$]{}]{}29[[\^[s]{}$^{\rm s}$]{}]{}3 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}50{\ifmmode{''}\else$''$\fi}$ & 221971\
LMC116 & 5[[\^[h]{}$^{\rm h}$]{}]{}07[[\^[m]{}$^{\rm m}$]{}]{}00[[\^[s]{}$^{\rm s}$]{}]{}9 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}29{\ifmmode{''}\else$''$\fi}$ & 205536 && LMC174 & 5[[\^[h]{}$^{\rm h}$]{}]{}37[[\^[m]{}$^{\rm m}$]{}]{}59[[\^[s]{}$^{\rm s}$]{}]{}8 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}26{\ifmmode{''}\else$''$\fi}$ & 340208\
LMC117 & 5[[\^[h]{}$^{\rm h}$]{}]{}06[[\^[m]{}$^{\rm m}$]{}]{}55[[\^[s]{}$^{\rm s}$]{}]{}3 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}58{\ifmmode{''}\else$''$\fi}$ & 527346 && LMC175 & 5[[\^[h]{}$^{\rm h}$]{}]{}38[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}3 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}01{\ifmmode{''}\else$''$\fi}$ & 497977\
LMC118 & 5[[\^[h]{}$^{\rm h}$]{}]{}06[[\^[m]{}$^{\rm m}$]{}]{}25[[\^[s]{}$^{\rm s}$]{}]{}4 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}25{\ifmmode{''}\else$''$\fi}$ & 694697 && LMC176 & 5[[\^[h]{}$^{\rm h}$]{}]{}39[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}6 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}30{\ifmmode{''}\else$''$\fi}$ & 576984\
LMC119 & 5[[\^[h]{}$^{\rm h}$]{}]{}06[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}5 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}02{\ifmmode{''}\else$''$\fi}$ & 817851 && LMC177 & 5[[\^[h]{}$^{\rm h}$]{}]{}39[[\^[m]{}$^{\rm m}$]{}]{}38[[\^[s]{}$^{\rm s}$]{}]{}0 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}49{\ifmmode{''}\else$''$\fi}$ & 817966\
LMC120 & 5[[\^[h]{}$^{\rm h}$]{}]{}05[[\^[m]{}$^{\rm m}$]{}]{}39[[\^[s]{}$^{\rm s}$]{}]{}8 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}28{\ifmmode{''}\else$''$\fi}$ & 617701 && LMC178 & 5[[\^[h]{}$^{\rm h}$]{}]{}40[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}1 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 471477\
LMC121 & 5[[\^[h]{}$^{\rm h}$]{}]{}05[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}4 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}59{\ifmmode{''}\else$''$\fi}$ & 442352 && LMC179 & 5[[\^[h]{}$^{\rm h}$]{}]{}40[[\^[m]{}$^{\rm m}$]{}]{}52[[\^[s]{}$^{\rm s}$]{}]{}3 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}36{\ifmmode{'}\else$'$\fi}58{\ifmmode{''}\else$''$\fi}$ & 296759\
LMC122 & 5[[\^[h]{}$^{\rm h}$]{}]{}04[[\^[m]{}$^{\rm m}$]{}]{}52[[\^[s]{}$^{\rm s}$]{}]{}9 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}25{\ifmmode{''}\else$''$\fi}$ & 288282 && LMC180 & 5[[\^[h]{}$^{\rm h}$]{}]{}40[[\^[m]{}$^{\rm m}$]{}]{}51[[\^[s]{}$^{\rm s}$]{}]{}5 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}12{\ifmmode{'}\else$'$\fi}28{\ifmmode{''}\else$''$\fi}$ & 258973\
LMC123 & 5[[\^[h]{}$^{\rm h}$]{}]{}01[[\^[m]{}$^{\rm m}$]{}]{}18[[\^[s]{}$^{\rm s}$]{}]{}0 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}53{\ifmmode{'}\else$'$\fi}00{\ifmmode{''}\else$''$\fi}$ & 242827 && LMC181 & 5[[\^[h]{}$^{\rm h}$]{}]{}43[[\^[m]{}$^{\rm m}$]{}]{}35[[\^[s]{}$^{\rm s}$]{}]{}7 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}58{\ifmmode{''}\else$''$\fi}$ & 201579\
LMC124 & 5[[\^[h]{}$^{\rm h}$]{}]{}01[[\^[m]{}$^{\rm m}$]{}]{}00[[\^[s]{}$^{\rm s}$]{}]{}3 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 290411 && LMC182 & 5[[\^[h]{}$^{\rm h}$]{}]{}44[[\^[m]{}$^{\rm m}$]{}]{}16[[\^[s]{}$^{\rm s}$]{}]{}0 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}32{\ifmmode{''}\else$''$\fi}$ & 311362\
LMC125 & 5[[\^[h]{}$^{\rm h}$]{}]{}00[[\^[m]{}$^{\rm m}$]{}]{}36[[\^[s]{}$^{\rm s}$]{}]{}1 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}54{\ifmmode{''}\else$''$\fi}$ & 357288 && LMC183 & 5[[\^[h]{}$^{\rm h}$]{}]{}45[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}8 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}59{\ifmmode{''}\else$''$\fi}$ & 361838\
LMC126 & 5[[\^[h]{}$^{\rm h}$]{}]{}00[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}4 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}31{\ifmmode{''}\else$''$\fi}$ & 530735 && LMC184 & 5[[\^[h]{}$^{\rm h}$]{}]{}45[[\^[m]{}$^{\rm m}$]{}]{}43[[\^[s]{}$^{\rm s}$]{}]{}2 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}33{\ifmmode{''}\else$''$\fi}$ & 486666\
LMC127 & 4[[\^[h]{}$^{\rm h}$]{}]{}59[[\^[m]{}$^{\rm m}$]{}]{}33[[\^[s]{}$^{\rm s}$]{}]{}6 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}54{\ifmmode{''}\else$''$\fi}$ & 547901 && LMC185 & 5[[\^[h]{}$^{\rm h}$]{}]{}46[[\^[m]{}$^{\rm m}$]{}]{}30[[\^[s]{}$^{\rm s}$]{}]{}8 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}51{\ifmmode{''}\else$''$\fi}$ & 630366\
LMC128 & 4[[\^[h]{}$^{\rm h}$]{}]{}59[[\^[m]{}$^{\rm m}$]{}]{}03[[\^[s]{}$^{\rm s}$]{}]{}6 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}24{\ifmmode{''}\else$''$\fi}$ & 406243 && LMC186 & 5[[\^[h]{}$^{\rm h}$]{}]{}47[[\^[m]{}$^{\rm m}$]{}]{}21[[\^[s]{}$^{\rm s}$]{}]{}2 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}24{\ifmmode{''}\else$''$\fi}$ & 376966\
LMC129 & 4[[\^[h]{}$^{\rm h}$]{}]{}58[[\^[m]{}$^{\rm m}$]{}]{}24[[\^[s]{}$^{\rm s}$]{}]{}6 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}07{\ifmmode{''}\else$''$\fi}$ & 304616 && LMC187 & 5[[\^[h]{}$^{\rm h}$]{}]{}48[[\^[m]{}$^{\rm m}$]{}]{}12[[\^[s]{}$^{\rm s}$]{}]{}6 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}36{\ifmmode{'}\else$'$\fi}52{\ifmmode{''}\else$''$\fi}$ & 298070\
LMC130 & 4[[\^[h]{}$^{\rm h}$]{}]{}57[[\^[m]{}$^{\rm m}$]{}]{}50[[\^[s]{}$^{\rm s}$]{}]{}8 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}20{\ifmmode{''}\else$''$\fi}$ & 231039 && LMC188 & 5[[\^[h]{}$^{\rm h}$]{}]{}48[[\^[m]{}$^{\rm m}$]{}]{}26[[\^[s]{}$^{\rm s}$]{}]{}6 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}12{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 152482\
LMC131 & 4[[\^[h]{}$^{\rm h}$]{}]{}55[[\^[m]{}$^{\rm m}$]{}]{}28[[\^[s]{}$^{\rm s}$]{}]{}6 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}52{\ifmmode{'}\else$'$\fi}46{\ifmmode{''}\else$''$\fi}$ & 248421 && LMC189 & 5[[\^[h]{}$^{\rm h}$]{}]{}50[[\^[m]{}$^{\rm m}$]{}]{}37[[\^[s]{}$^{\rm s}$]{}]{}9 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}26{\ifmmode{''}\else$''$\fi}$ & 153656\
LMC132 & 4[[\^[h]{}$^{\rm h}$]{}]{}55[[\^[m]{}$^{\rm m}$]{}]{}00[[\^[s]{}$^{\rm s}$]{}]{}6 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}36{\ifmmode{''}\else$''$\fi}$ & 217827 && LMC190 & 5[[\^[h]{}$^{\rm h}$]{}]{}51[[\^[m]{}$^{\rm m}$]{}]{}33[[\^[s]{}$^{\rm s}$]{}]{}2 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}55{\ifmmode{''}\else$''$\fi}$ & 200564\
LMC133 & 4[[\^[h]{}$^{\rm h}$]{}]{}54[[\^[m]{}$^{\rm m}$]{}]{}29[[\^[s]{}$^{\rm s}$]{}]{}2 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}47{\ifmmode{''}\else$''$\fi}$ & 346558 && LMC191 & 5[[\^[h]{}$^{\rm h}$]{}]{}52[[\^[m]{}$^{\rm m}$]{}]{}20[[\^[s]{}$^{\rm s}$]{}]{}1 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}24{\ifmmode{''}\else$''$\fi}$ & 243558\
LMC134 & 4[[\^[h]{}$^{\rm h}$]{}]{}53[[\^[m]{}$^{\rm m}$]{}]{}49[[\^[s]{}$^{\rm s}$]{}]{}2 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}18{\ifmmode{''}\else$''$\fi}$ & 304271 && LMC192 & 5[[\^[h]{}$^{\rm h}$]{}]{}53[[\^[m]{}$^{\rm m}$]{}]{}24[[\^[s]{}$^{\rm s}$]{}]{}1 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}51{\ifmmode{''}\else$''$\fi}$ & 228516\
LMC135 & 4[[\^[h]{}$^{\rm h}$]{}]{}53[[\^[m]{}$^{\rm m}$]{}]{}05[[\^[s]{}$^{\rm s}$]{}]{}2 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}51{\ifmmode{''}\else$''$\fi}$ & 284846 && LMC193 & 5[[\^[h]{}$^{\rm h}$]{}]{}54[[\^[m]{}$^{\rm m}$]{}]{}21[[\^[s]{}$^{\rm s}$]{}]{}7 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}34{\ifmmode{''}\else$''$\fi}$ & 138223\
LMC136 & 4[[\^[h]{}$^{\rm h}$]{}]{}52[[\^[m]{}$^{\rm m}$]{}]{}23[[\^[s]{}$^{\rm s}$]{}]{}7 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}25{\ifmmode{''}\else$''$\fi}$ & 244490 && LMC194 & 5[[\^[h]{}$^{\rm h}$]{}]{}55[[\^[m]{}$^{\rm m}$]{}]{}29[[\^[s]{}$^{\rm s}$]{}]{}7 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}36{\ifmmode{'}\else$'$\fi}59{\ifmmode{''}\else$''$\fi}$ & 77816\
LMC137 & 4[[\^[h]{}$^{\rm h}$]{}]{}51[[\^[m]{}$^{\rm m}$]{}]{}30[[\^[s]{}$^{\rm s}$]{}]{}2 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}01{\ifmmode{''}\else$''$\fi}$ & 200058 && LMC195 & 5[[\^[h]{}$^{\rm h}$]{}]{}56[[\^[m]{}$^{\rm m}$]{}]{}00[[\^[s]{}$^{\rm s}$]{}]{}0 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}12{\ifmmode{'}\else$'$\fi}25{\ifmmode{''}\else$''$\fi}$ & 43002\
LMC138 & 4[[\^[h]{}$^{\rm h}$]{}]{}49[[\^[m]{}$^{\rm m}$]{}]{}34[[\^[s]{}$^{\rm s}$]{}]{}7 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}53{\ifmmode{'}\else$'$\fi}07{\ifmmode{''}\else$''$\fi}$ & 117664 && LMC196 & 5[[\^[h]{}$^{\rm h}$]{}]{}56[[\^[m]{}$^{\rm m}$]{}]{}54[[\^[s]{}$^{\rm s}$]{}]{}7 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}29{\ifmmode{''}\else$''$\fi}$ & 116258\
LMC139 & 4[[\^[h]{}$^{\rm h}$]{}]{}49[[\^[m]{}$^{\rm m}$]{}]{}05[[\^[s]{}$^{\rm s}$]{}]{}2 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}30{\ifmmode{''}\else$''$\fi}$ & 128858 && LMC197 & 5[[\^[h]{}$^{\rm h}$]{}]{}58[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}7 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}06{\ifmmode{''}\else$''$\fi}$ & 92295\
LMC140 & 4[[\^[h]{}$^{\rm h}$]{}]{}48[[\^[m]{}$^{\rm m}$]{}]{}18[[\^[s]{}$^{\rm s}$]{}]{}2 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}04{\ifmmode{'}\else$'$\fi}05{\ifmmode{''}\else$''$\fi}$ & 188164 && LMC198 & 5[[\^[h]{}$^{\rm h}$]{}]{}59[[\^[m]{}$^{\rm m}$]{}]{}02[[\^[s]{}$^{\rm s}$]{}]{}5 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}35{\ifmmode{''}\else$''$\fi}$ & 71992\
LMC141 & 4[[\^[h]{}$^{\rm h}$]{}]{}47[[\^[m]{}$^{\rm m}$]{}]{}26[[\^[s]{}$^{\rm s}$]{}]{}7 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}36{\ifmmode{''}\else$''$\fi}$ & 191197 && LMC199 & 6[[\^[h]{}$^{\rm h}$]{}]{}00[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}7 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}00{\ifmmode{''}\else$''$\fi}$ & 63841\
LMC142 & 4[[\^[h]{}$^{\rm h}$]{}]{}46[[\^[m]{}$^{\rm m}$]{}]{}31[[\^[s]{}$^{\rm s}$]{}]{}9 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}08{\ifmmode{''}\else$''$\fi}$ & 225807 && LMC200 & 6[[\^[h]{}$^{\rm h}$]{}]{}01[[\^[m]{}$^{\rm m}$]{}]{}27[[\^[s]{}$^{\rm s}$]{}]{}5 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}36{\ifmmode{''}\else$''$\fi}$ & 58765\
LMC143 & 4[[\^[h]{}$^{\rm h}$]{}]{}45[[\^[m]{}$^{\rm m}$]{}]{}43[[\^[s]{}$^{\rm s}$]{}]{}1 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}19{\ifmmode{''}\else$''$\fi}$ & 155672 && LMC201 & 6[[\^[h]{}$^{\rm h}$]{}]{}02[[\^[m]{}$^{\rm m}$]{}]{}45[[\^[s]{}$^{\rm s}$]{}]{}9 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}37{\ifmmode{'}\else$'$\fi}04{\ifmmode{''}\else$''$\fi}$ & 91051\
LMC144 & 4[[\^[h]{}$^{\rm h}$]{}]{}44[[\^[m]{}$^{\rm m}$]{}]{}40[[\^[s]{}$^{\rm s}$]{}]{}2 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}01{\ifmmode{''}\else$''$\fi}$ & 116363 && LMC202 & 6[[\^[h]{}$^{\rm h}$]{}]{}03[[\^[m]{}$^{\rm m}$]{}]{}28[[\^[s]{}$^{\rm s}$]{}]{}3 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}12{\ifmmode{'}\else$'$\fi}34{\ifmmode{''}\else$''$\fi}$ & 79469\
LMC145 & 4[[\^[h]{}$^{\rm h}$]{}]{}43[[\^[m]{}$^{\rm m}$]{}]{}47[[\^[s]{}$^{\rm s}$]{}]{}5 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}52{\ifmmode{'}\else$'$\fi}43{\ifmmode{''}\else$''$\fi}$ & 64628 && LMC203 & 6[[\^[h]{}$^{\rm h}$]{}]{}03[[\^[m]{}$^{\rm m}$]{}]{}29[[\^[s]{}$^{\rm s}$]{}]{}9 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}48{\ifmmode{'}\else$'$\fi}04{\ifmmode{''}\else$''$\fi}$ & 71396\
LMC146 & 4[[\^[h]{}$^{\rm h}$]{}]{}43[[\^[m]{}$^{\rm m}$]{}]{}03[[\^[s]{}$^{\rm s}$]{}]{}0 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}17{\ifmmode{''}\else$''$\fi}$ & 86656 && LMC204 & 6[[\^[h]{}$^{\rm h}$]{}]{}03[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}6 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}25{\ifmmode{''}\else$''$\fi}$ & 108173\
LMC147 & 4[[\^[h]{}$^{\rm h}$]{}]{}42[[\^[m]{}$^{\rm m}$]{}]{}07[[\^[s]{}$^{\rm s}$]{}]{}8 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}03{\ifmmode{'}\else$'$\fi}55{\ifmmode{''}\else$''$\fi}$ & 103604 && LMC205 & 6[[\^[h]{}$^{\rm h}$]{}]{}04[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}9 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}04{\ifmmode{''}\else$''$\fi}$ & 82072\
LMC148 & 4[[\^[h]{}$^{\rm h}$]{}]{}41[[\^[m]{}$^{\rm m}$]{}]{}06[[\^[s]{}$^{\rm s}$]{}]{}8 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 110885 && LMC206 & 6[[\^[h]{}$^{\rm h}$]{}]{}05[[\^[m]{}$^{\rm m}$]{}]{}40[[\^[s]{}$^{\rm s}$]{}]{}3 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 78179\
LMC149 & 4[[\^[h]{}$^{\rm h}$]{}]{}40[[\^[m]{}$^{\rm m}$]{}]{}05[[\^[s]{}$^{\rm s}$]{}]{}1 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}57{\ifmmode{''}\else$''$\fi}$ & 115117 && LMC207 & 6[[\^[h]{}$^{\rm h}$]{}]{}07[[\^[m]{}$^{\rm m}$]{}]{}04[[\^[s]{}$^{\rm s}$]{}]{}2 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}55{\ifmmode{''}\else$''$\fi}$ & 70889\
LMC150 & 4[[\^[h]{}$^{\rm h}$]{}]{}39[[\^[m]{}$^{\rm m}$]{}]{}05[[\^[s]{}$^{\rm s}$]{}]{}3 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}16{\ifmmode{''}\else$''$\fi}$ & 96039 && LMC208 & 6[[\^[h]{}$^{\rm h}$]{}]{}08[[\^[m]{}$^{\rm m}$]{}]{}30[[\^[s]{}$^{\rm s}$]{}]{}4 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}27{\ifmmode{''}\else$''$\fi}$ & 87619\
LMC151 & 4[[\^[h]{}$^{\rm h}$]{}]{}37[[\^[m]{}$^{\rm m}$]{}]{}51[[\^[s]{}$^{\rm s}$]{}]{}6 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}25{\ifmmode{'}\else$'$\fi}45{\ifmmode{''}\else$''$\fi}$ & 89935 && LMC209 & 6[[\^[h]{}$^{\rm h}$]{}]{}10[[\^[m]{}$^{\rm m}$]{}]{}07[[\^[s]{}$^{\rm s}$]{}]{}0 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}37{\ifmmode{'}\else$'$\fi}00{\ifmmode{''}\else$''$\fi}$ & 64885\
LMC152 & 4[[\^[h]{}$^{\rm h}$]{}]{}37[[\^[m]{}$^{\rm m}$]{}]{}54[[\^[s]{}$^{\rm s}$]{}]{}1 & $-66{\ifmmode^{\circ}\else$^{\circ}$\fi}52{\ifmmode{'}\else$'$\fi}52{\ifmmode{''}\else$''$\fi}$ & 46107 && LMC210 & 6[[\^[h]{}$^{\rm h}$]{}]{}10[[\^[m]{}$^{\rm m}$]{}]{}55[[\^[s]{}$^{\rm s}$]{}]{}7 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}12{\ifmmode{'}\else$'$\fi}37{\ifmmode{''}\else$''$\fi}$ & 70282\
LMC153 & 4[[\^[h]{}$^{\rm h}$]{}]{}37[[\^[m]{}$^{\rm m}$]{}]{}01[[\^[s]{}$^{\rm s}$]{}]{}7 & $-67{\ifmmode^{\circ}\else$^{\circ}$\fi}28{\ifmmode{'}\else$'$\fi}30{\ifmmode{''}\else$''$\fi}$ & 56016 && LMC211 & 6[[\^[h]{}$^{\rm h}$]{}]{}11[[\^[m]{}$^{\rm m}$]{}]{}22[[\^[s]{}$^{\rm s}$]{}]{}0 & $-72{\ifmmode^{\circ}\else$^{\circ}$\fi}48{\ifmmode{'}\else$'$\fi}04{\ifmmode{''}\else$''$\fi}$ & 61205\
LMC154 & 4[[\^[h]{}$^{\rm h}$]{}]{}35[[\^[m]{}$^{\rm m}$]{}]{}59[[\^[s]{}$^{\rm s}$]{}]{}1 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}04{\ifmmode{'}\else$'$\fi}02{\ifmmode{''}\else$''$\fi}$ & 65095 && LMC212 & 6[[\^[h]{}$^{\rm h}$]{}]{}11[[\^[m]{}$^{\rm m}$]{}]{}04[[\^[s]{}$^{\rm s}$]{}]{}0 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}14{\ifmmode{'}\else$'$\fi}50{\ifmmode{''}\else$''$\fi}$ & 81878\
LMC155 & 4[[\^[h]{}$^{\rm h}$]{}]{}34[[\^[m]{}$^{\rm m}$]{}]{}49[[\^[s]{}$^{\rm s}$]{}]{}4 & $-68{\ifmmode^{\circ}\else$^{\circ}$\fi}39{\ifmmode{'}\else$'$\fi}32{\ifmmode{''}\else$''$\fi}$ & 71173 && LMC213 & 6[[\^[h]{}$^{\rm h}$]{}]{}12[[\^[m]{}$^{\rm m}$]{}]{}17[[\^[s]{}$^{\rm s}$]{}]{}9 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}37{\ifmmode{''}\else$''$\fi}$ & 52207\
LMC156 & 4[[\^[h]{}$^{\rm h}$]{}]{}33[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}7 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}15{\ifmmode{'}\else$'$\fi}00{\ifmmode{''}\else$''$\fi}$ & 72805 && LMC214 & 6[[\^[h]{}$^{\rm h}$]{}]{}13[[\^[m]{}$^{\rm m}$]{}]{}58[[\^[s]{}$^{\rm s}$]{}]{}2 & $-70{\ifmmode^{\circ}\else$^{\circ}$\fi}26{\ifmmode{'}\else$'$\fi}08{\ifmmode{''}\else$''$\fi}$ & 60410\
LMC157 & 4[[\^[h]{}$^{\rm h}$]{}]{}32[[\^[m]{}$^{\rm m}$]{}]{}23[[\^[s]{}$^{\rm s}$]{}]{}8 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}50{\ifmmode{'}\else$'$\fi}26{\ifmmode{''}\else$''$\fi}$ & 44165 && LMC215 & 6[[\^[h]{}$^{\rm h}$]{}]{}15[[\^[m]{}$^{\rm m}$]{}]{}36[[\^[s]{}$^{\rm s}$]{}]{}4 & $-71{\ifmmode^{\circ}\else$^{\circ}$\fi}01{\ifmmode{'}\else$'$\fi}28{\ifmmode{''}\else$''$\fi}$ & 60801\
Construction of Photometric Maps
================================
Because the OGLE-III photometric databases are constructed separately for [*I*]{} and [*V*]{}-band data, the first step of map preparation was the cross-identification of [*V*]{}-band counterparts to each [*I*]{}-band object. Due to some shifts between respective [*I*]{} and [*V*]{}-band reference images of a given subfield and larger than single detector dimensions size of the reference images to mask the gaps between chips of the OGLE-III mosaic camera, the [*V*]{}-band counterparts of objects located close to the border in the [*I*]{}-band reference image may be present in more than one field. Therefore in the first step third order transformation between pixel grids of a given [*I*]{}-band subfield reference image and all [*V*]{}-band reference images of subfields that even partially overlap were derived. Then for all [*I*]{}-band database objects of a given subfield their corresponding [*V*]{}-band counterparts in all fields were found.
The mean photometry was derived for all objects with minimum of 4 and 6 observations in the [*V*]{} and [*I*]{}-band, respectively, by averaging all observations after removing $5\sigma$ deviating points. In the case of multiple [*V*]{}-band cross-identifications all [*V*]{}-band observations formed a single dataset and then were averaged because they are independent measurements. Finally, the color term correction was applied for each object to its database average magnitude according to the transformation equations and color term coefficients presented in Udalski [[*et al. *]{}]{}(2008). After the transformation of the $V-I$ color to the standard system, [*I*]{} and [*V*]{} magnitudes were appropriately adjusted. For objects that do not have color information the average color of the LMC population $V-I=0.7$ mag was used for color correction of respective [*I*]{} or [*V*]{} magnitude.
Table 2 presents the first 25 entries from the map of the LMC100.1 subfield. The columns contain: (1) ID number; (2,3) equatorial coordinates J2000.0; (4,5) $X,Y$ pixel coordinates in the [*I*]{}-band reference image; (6,7,8) photometry: [*V*]{}, $V-I$, [*I*]{}; (9,10,11) number of points for average magnitude in [*V*]{}, number of $5\sigma$ removed points in [*V*]{}, $\sigma$ of magnitude for [*V*]{}-band; (12,13,14) same as (9,10,11) for the [*I*]{}-band. 9.999 or 99.999 markers mean “no data”. $-1$ in column (10) indicates multiple [*V*]{}-band cross-identification (the average magnitude and standard deviation are calculated for merged photometry).
The full set of the OGLE-III Photometric Maps is available from the OGLE Internet archive (see below).
Discussion
==========
OGLE-III Photometric Maps form a significant extension of the OGLE-II maps that covered only central regions of the LMC ([[*cf. *]{}]{}Fig. 1). The new maps include practically entire area of the LMC and can be used for many projects studying the global structure of this galaxy. Precise, well calibrated [*VI*]{} photometry and astrometry makes this dataset a unique tool for many astrophysical applications.
To show the accuracy of the OGLE-III Photometric Maps Figs. 2 and 3 present standard deviation of magnitudes as a function of magnitude in the [*V*]{} and [*I*]{}-band for two LMC fields: LMC100.1 – very dense field located in the central bar and LMC209.1 – sparsely populated field form the outer parts of this galaxy. As one can expect the accuracy of photometry depends on the stellar density and, for example, $\sigma=0.1$ mag photometry scatter is reached for stars by about 0.3 magnitude brighter in the densest bar fields than in the uncrowded fields.
{width="9cm"}
Fig. . Histogram of magnitudes for the central bar subfield LMC100.1.
{width="9cm"}
Fig. . Same as in Fig. 5 for the outer subfield LMC209.1.
Figs. 4 and 5 show the histograms of objects in the [*V*]{} and [*I*]{}-band for the same two fields. One can notice that the completeness of the maps is high and reaches $V\approx20.5{-}21$ mag and $I\approx20{-}20.5$ mag. Again, as one can expected, completeness is a function of stellar crowding.
[ r@ c@ c@ r@ c@ c@ r@ c@ c@ r@ c@ r@ c@ c]{} ID & RA & DEC & $X$ & $Y$ & $V$ & $V-I$ & $I$ & $N_V$ & $N^{\rm bad}_V$ & $\sigma_V$ & $N_I$ & $N^{\rm bad}_I$ & $\sigma_I$\
&(2000) & (2000)&&&&&&&&&&&\
1 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}05[[\^[s]{}$^{\rm s}$]{}]{}97 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}33{\ifmmode{'}\else$'$\fi}17{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}4$ & 57.08 & 56.26 & 16.647 & $ 2.483$ & 14.164 & 35 & $ 0$ & 0.127 & 6 & 0 & 0.058\
2 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}08[[\^[s]{}$^{\rm s}$]{}]{}62 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}32{\ifmmode{'}\else$'$\fi}15{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}3$ & 296.00 & 110.54 & 16.331 & $ 1.922$ & 14.409 & 67 & $-1$ & 0.208 & 399 & 0 & 0.077\
3 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}08[[\^[s]{}$^{\rm s}$]{}]{}67 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}56{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}1$ & 600.20 & 112.84 & 17.715 & $ 3.062$ & 14.652 & 47 & $ 0$ & 0.157 & 399 & 0 & 0.065\
4 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}09[[\^[s]{}$^{\rm s}$]{}]{}13 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}39{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 663.16 & 122.23 & 16.318 & $ 1.847$ & 14.471 & 48 & $ 0$ & 0.018 & 402 & 0 & 0.010\
5 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}10[[\^[s]{}$^{\rm s}$]{}]{}04 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}29{\ifmmode{'}\else$'$\fi}46{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 866.65 & 141.47 & 14.495 & $ 0.354$ & 14.141 & 48 & $ 0$ & 0.005 & 402 & 0 & 0.007\
6 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}11[[\^[s]{}$^{\rm s}$]{}]{}95 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}32{\ifmmode{'}\else$'$\fi}00{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}9$ & 351.08 & 177.73 & 17.470 & $ 3.182$ & 14.288 & 44 & $ 0$ & 0.680 & 404 & 0 & 0.214\
7 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}13[[\^[s]{}$^{\rm s}$]{}]{}76 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}00{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}3$ & 814.39 & 216.32 & 16.530 & $ 2.092$ & 14.438 & 48 & $ 0$ & 0.029 & 404 & 0 & 0.015\
8 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}12 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}29{\ifmmode{'}\else$'$\fi}03{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}2$ &1033.50 & 224.43 & 14.141 & $ 0.188$ & 13.953 & 48 & $ 0$ & 0.004 & 404 & 0 & 0.006\
9 & 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}14[[\^[s]{}$^{\rm s}$]{}]{}53 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}32{\ifmmode{'}\else$'$\fi}48{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}2$ & 168.82 & 228.84 & 16.917 & $ 2.326$ & 14.591 & 87 & $-1$ & 0.072 & 385 & 0 & 0.033\
10& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}16[[\^[s]{}$^{\rm s}$]{}]{}07 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}03{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}1$ & 803.47 & 262.79 & 14.739 & $ 0.780$ & 13.959 & 48 & $ 0$ & 0.005 & 404 & 0 & 0.005\
11& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}17[[\^[s]{}$^{\rm s}$]{}]{}31 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}31{\ifmmode{'}\else$'$\fi}47{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}3$ & 402.93 & 285.85 & 15.366 & $ 1.367$ & 13.998 & 48 & $ 0$ & 0.006 & 404 & 0 & 0.005\
12& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}17[[\^[s]{}$^{\rm s}$]{}]{}46 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}29{\ifmmode{'}\else$'$\fi}50{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}9$ & 850.10 & 290.92 & 16.813 & $ 2.172$ & 14.641 & 48 & $ 0$ & 0.072 & 404 & 0 & 0.031\
13& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}19[[\^[s]{}$^{\rm s}$]{}]{}14 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}12{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 766.32 & 324.45 & 17.416 & $ 3.007$ & 14.408 & 47 & $ 0$ & 0.286 & 404 & 0 & 0.189\
14& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}19[[\^[s]{}$^{\rm s}$]{}]{}61 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}32{\ifmmode{'}\else$'$\fi}51{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 155.22 & 330.93 & 16.432 & $ 2.245$ & 14.187 & 84 & $-1$ & 0.048 & 366 & 0 & 0.021\
15& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}19[[\^[s]{}$^{\rm s}$]{}]{}54 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}31{\ifmmode{'}\else$'$\fi}44{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}6$ & 413.08 & 330.83 & 16.979 & $ 2.371$ & 14.608 & 48 & $ 0$ & 0.131 & 404 & 0 & 0.062\
16& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}23[[\^[s]{}$^{\rm s}$]{}]{}42 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}33{\ifmmode{'}\else$'$\fi}20{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}3$ & 44.83 & 407.14 & 16.310 & $ 9.999$ & 99.999 & 40 & $ 0$ & 0.052 & 0 & 0 & 9.999\
17& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}25[[\^[s]{}$^{\rm s}$]{}]{}72 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}29{\ifmmode{'}\else$'$\fi}44{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}3$ & 874.65 & 457.67 & 15.044 & $ 1.200$ & 13.844 & 48 & $ 0$ & 0.006 & 404 & 0 & 0.004\
18& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}27[[\^[s]{}$^{\rm s}$]{}]{}42 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}31{\ifmmode{'}\else$'$\fi}18{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 511.94 & 489.95 & 15.476 & $ 1.214$ & 14.262 & 48 & $ 0$ & 0.008 & 404 & 0 & 0.005\
19& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}27[[\^[s]{}$^{\rm s}$]{}]{}78 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}30{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}2$ & 698.01 & 498.27 & 14.010 & $ 9.999$ & 99.999 & 48 & $ 0$ & 0.153 & 0 & 0 & 9.999\
20& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}29[[\^[s]{}$^{\rm s}$]{}]{}18 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}57{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}9$ & 591.71 & 525.78 & 13.920 & $-0.135$ & 14.055 & 48 & $ 0$ & 0.006 & 404 & 0 & 0.008\
21& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}47 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}34{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}9$ & 679.44 & 592.49 & 15.383 & $ 9.999$ & 99.999 & 48 & $ 0$ & 0.061 & 0 & 0 & 9.999\
22& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}32[[\^[s]{}$^{\rm s}$]{}]{}64 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}32{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}9$ & 687.26 & 596.09 & 16.676 & $ 9.999$ & 99.999 & 48 & $ 0$ & 0.026 & 0 & 0 & 9.999\
23& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}33[[\^[s]{}$^{\rm s}$]{}]{}45 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}18{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}0$ & 744.46 & 612.71 & 14.765 & $ 1.573$ & 13.192 & 48 & $ 0$ & 0.008 & 404 & 0 & 0.006\
24& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}33[[\^[s]{}$^{\rm s}$]{}]{}84 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}30{\ifmmode{'}\else$'$\fi}18{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}0$ & 744.28 & 620.53 & 14.906 & $ 1.322$ & 13.584 & 48 & $ 0$ & 0.008 & 404 & 0 & 0.007\
25& 5[[\^[h]{}$^{\rm h}$]{}]{}19[[\^[m]{}$^{\rm m}$]{}]{}37[[\^[s]{}$^{\rm s}$]{}]{}93 & $-69{\ifmmode^{\circ}\else$^{\circ}$\fi}31{\ifmmode{'}\else$'$\fi}46{\makebox[0pt][l]{.}}{\ifmmode{''}\else$''$\fi}7$ & 402.99 & 700.88 & 15.501 & $ 2.142$ & 13.359 & 48 & $ 0$ & 0.046 & 404 & 0 & 0.017\
Figs. 6–9 show several color–magnitude diagrams (CMDs) constructed directly from the OGLE-III maps. They include CMD of one of the densest central fields of the LMC: LMC100.1, high extinction region near the Tarantula nebula: LMC175.6, as well as regions in the outskirts of this galaxy (LMC121.1 in the western wing and LMC186.1 in the eastern wing). Figs. 6–9 clearly illustrate the quality of data and reveal in details the characteristic features of the LMC stellar populations, as well as show how interstellar extinction affects the CMDs.
Data Availability
=================
The OGLE-III Photometric Maps of the LMC are available to the astronomical community from the OGLE Internet Archive:
[*http://ogle.astrouw.edu.pl ftp://ftp.astrouw.edu.pl/ogle3/maps/lmc/*]{}
Beside tables with photometric data and astrometry for each of the subfields the [*I*]{}-band reference images are also included. Usage of the data is fully allowed, requiring only the proper acknowledgment to the OGLE project.
This paper was partially supported by the Polish MNiSW grant N20303032/4275 to AU and NN203293533 to IS and by the Foundation for Polish Science through the Homing Program.
[Alard, C., and Lupton, R.]{} 1998, [**]{}, [**503**]{}, [325]{}.
[Landolt, A.U.]{} 1992, [**]{}, [**104**]{}, [372]{}.
[Pojma[ń]{}ski, G.]{} 1997, [**]{}, [**47**]{}, [467]{}.
[Udalski, A.]{} 2003, [**]{}, [**53**]{}, [291]{}.
[Udalski, A., Szymański, M., Kubiak, M., Pietrzyński, G., Woźniak, P., and [Ż]{}ebruń, K.]{} 1998, [**]{}, [**48**]{}, [147]{}.
[Udalski, A., Szymański, M., Kubiak, M., Pietrzyński, G., Soszy[ń]{}ski, I., Woźniak, P., and [Ż]{}ebruń, K.]{} 2000, [**]{}, [**50**]{}, [307]{}.
[Udalski, A., Szymański, M., Kubiak, M., Pietrzyński, G., Soszy[ń]{}ski, I., Woźniak, P., [Ż]{}ebruń, K., Szewczyk, O., and Wyrzykowski, [Ł]{}.]{} 2000, [**]{}, [**52**]{}, [217]{}.
[Udalski, A., Szymański, M.K., Soszy[ń]{}ski, I., and Poleski, R.]{} 2008, [**]{}, [**58**]{}, [69]{}.
[Wo[ź]{}niak, P.R.]{} 2000, [**]{}, [**50**]{}, [421]{}.
Captions of JPEG figures.
Fig. 1. OGLE-III fields in the LMC (black squares: 100–215) overplotted on the image obtained by the ASAS all sky survey. Red strips (1–21) mark OGLE-II fields. Fig. 2. Standard deviation of magnitudes as a function of magnitude for the central bar subfield LMC100.1. Fig. 3. Same as in Fig. 2 for the outer subfield LMC209.1 Fig. 6. Color–magnitude diagram for the central bar subfield LMC100.1. Fig. 7. Color–magnitude diagram for the variable extinction subfield LMC175.6. Fig. 8. Color–magnitude diagram for the western wing subfield LMC121.1 Fig. 9. Color–magnitude diagram for the eastern wing subfield LMC186.1
[^1]: Based on observations obtained with the 1.3 m Warsaw telescope at the Las Campanas Observatory of the Carnegie Institution of Washington.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we consider a cognitive multi-hop relay secondary user (SU) system sharing the spectrum with some primary users (PU). The transmit power as well as the hop selection of the cognitive relays can be dynamically adapted according to the local (and causal) knowledge of the instantaneous channel state information (CSI) in the multi-hop SU system. We shall determine a low complexity, decentralized algorithm to maximize the average end-to-end throughput of the SU system with dynamic spatial reuse. The problem is challenging due to the decentralized requirement as well as the causality constraint on the knowledge of CSI. Furthermore, the problem belongs to the class of stochastic [*Network Utility Maximization*]{} (NUM) problems which is quite challenging [[@NUM]]{}. We exploit the time-scale difference between the PU activity and the CSI fluctuations and decompose the problem into a master problem and subproblems. We derive an asymptotically optimal low complexity solution using divide-and-conquer and [ illustrate]{} that significant performance gain can be obtained through dynamic hop selection and power control. The worst case complexity and memory requirement of the proposed algorithm is ${\cal O}(M^2)$ and ${\cal
O}(M^3)$ respectively, where $M$ is the number of SUs.
author:
-
title: 'Decentralized Dynamic Hop Selection and Power Control in Cognitive Multi-hop Relay Systems'
---
\[section\] \[section\] \[section\] \[section\] \[section\]
Introduction
============
Cooperative Communication and Dynamic Spectrum Access (DSA) are two important technologies that drive the evolution of the next generation wireless systems. For instance, cooperative communication [@broadcast; @cooperative; @commnumcation] exploits the broadcast nature of the wireless channel and enhances the reliability of the packet against channel fading and hence, increases the coverage of wireless systems. There are a lot of works studying multi-hop relay network. In [@dual; @hop; @perf1], the authors analyzed the performance of a dual-hop relaying communications over fading channels. Performance bounds of multi-hop relay system is analyzed in [@multi; @hop; @bound]. However, these works did not consider dynamic resource adaptation in the relay system. In [@ARF2], the authors investigated the minimum energy per bit treating both capacity and power consumption as optimization parameters in the wireless ad-hoc network. The minimization of the transmit power under the assumption of orthogonal transmissions was studied in [@ARF; @ARF4], in which the optimal parallel-relay channel power allocation for [ *Amplify and Forward* (AF) and *Decode and Forward* (DF)]{} were derived. However, in all these works, the power control solution adapts on the path loss only and failed to exploit the dynamic fluctuations of microscopic fading. In [@Kim:2008], the authors considered dynamic power control for multi-hop relay but the solution is centralized and requires knowledge of the global channel state information about the entire adhoc network, which is very difficult to realize in practice. Furthermore, a fixed number of hops to deliver a packet to the destination is always assumed in the above works. Due to the [*store-and-forward*]{} penalty in the end-to-end throughput of multi-hop relaying, it is not always optimal to involve a fixed number of hops in the multi-hop network. To tackle this issue, various [*opportunistic*]{} multi-hop relaying protocols were proposed in [@CU1; @CU2; @CU3]. In these designs, the number of hops to deliver a packet to the destination node changes dynamically according to the channel conditions. However, in these works, the opportunistic multi-hop protocols are heuristic in nature and the performance is studied by simulation and empirical measurements. In [@one; @relay; @perf], performance analysis on one-hop relay protocol is [studied.]{} In [@CU4; @CU5], performance analysis on some simple opportunistic multi-hop relaying protocols [is studied]{}. Furthermore, they all assume constant transmit power and deterministic channels where the effects of random fading is ignored.
On the other hand, DSA is an important new paradigm of spectrum access in which a secondary system dynamically shares [medium]{} with a higher priority primary system. Using [*cognitive radios*]{} [(CRs)]{} [@CR1; @CR2], the nodes in the secondary user (SU) systems sense the activity of the primary users (PUs) and access the spectrum only if the primary system is idle. In other words, the SU system dynamically share the spectrum with the PU systems by exploiting the burstiness of the PU traffic in the temporal, frequency and spatial domains. One key issue of DSA or [CR]{} is the efficiency of spectrum sharing between the SU and PU systems. In [@CR3; @CR4], the authors considered a [CR]{} system based on the [*interference avoidance approach*]{} in which the SU could transmit only if there are no active PUs within the coverage of the SU system. While such approach exploits the burstiness of the PU activity without requiring the knowledge of PU signal structure, the access opportunity of the SU system will be quite low for SU separated by a large distance as such access opportunity exists only if all the PU along the SU coverage are idle simultaneously. As a result, cognitive multi-hop relay for the SU systems is a promising solution to resolve this issue of low probability of access for distant secondary users. While intuitively, cognitive multi-hop relay could significantly enhance the spectrum sharing efficiency between the SU and PU systems, there are still a number of technical challenges to overcome as listed below.
- [**Jointly Optimal Opportunistic Hop Selection and Power Control for Cognitive Multi-hop relays:**]{} Most of the existing works only considered either the power control [@ARF; @ARF4] or the opportunistic multi-hop relaying protocols. It is very important to jointly optimize both the forward hopping strategy and the power control policy to exploit the instantaneous fluctuations of PU activities and the microscopic fading in order to improve the performance of the cognitive multi-hop relays.
- [**Dynamic Spatial Reuse in Cognitive Multihop Relaying:**]{} In most of the existing works studying power control or forward hopping in multihop relay [@CU1; @CU2; @CU3], they focus entirely on the multihop aspects of the problem and assume that the multi-hop network does not have to share spectrum with any PU systems. This simplifies the problem significantly. While this is a reasonable assumption in the regular multihop network without PU, such symmetric spatial reuse is not always possible in cognitive multihop relay network due to the random PU activities on any hops.
- [**Decentralized Solution with Local Knowledge of Channel State Information (CSI):**]{} An additional level of difficulty in solving the forward hopping and power control problem is the requirement of decentralized solution. In practice, it is very difficult to obtain and keep track of an up-to-date knowledge of the instantaneous channel state information for the entire multi-hop network. As a result, it is desirable to have a decentralized solution which requires knowledge of local (rather than global) channel state information only. In [@distributedchain], the authors considered a distributed resource management scheme for multi-hop CR networks but no power control is considered and the solution is based on heuristic design.
- [**Causal Knowledge of Channel States in the Multi-hop Relay Network:**]{} In most of the existing works [@Kim:2008], not only global knowledge but also non-causal[^1] knowledge of channel states in the multi-hop network is assumed. Specifically, at $t=0$, the centralized controller is assumed to have knowledge of all the channel states in all the hops of the entire multi-hop relay network. However, by the time the packets are delivered in the $n$-th hop, the actual channel state may have changed and the constraint of having causal knowledge of channel states have not been taken into account in the previous works of power optimization in multihop relay network.
In this paper, we shall try to address the above technical challenges. We consider a cognitive multi-hop SU system with a source, a destination and $M$ half-duplex cognitive relays scattered between the source and the destination. The SU system dynamically shares the spectrum with a PU system (with many PU nodes). The transmit power of the SU nodes as well as the hopping sequence of the cognitive relays are adaptive according to the local (and causal) knowledge of channel states in the multi-hop SU system to optimize the average end-to-end throughput. The solution also accommodates dynamic spatial reuse across the cognitive multi-hop system. The problem belongs to the class of stochastic NUM[^2] problems, which is well-known to be challenging. To obtain a decentralized solution for the throughput optimization problem we exploit the time-scale difference between the PU activity and the CSI fluctuations and decompose the problem into a master problem and subproblems (operating at different time scales). To deal with the causality requirement[^3], we express the subproblems into recursive forms and solve them using [*divide-and-conquer*]{}. We show that significant performance gains on the throughput of the SU system can be obtained using joint forward hopping and power control over a wide range of PU activity. Furthermore, we show that the decentralized solution has worst case complexity of ${\cal O}(M^2)$ and is asymptotically optimal for large $M$.
System Model, Control Policy and End-to-End Throughput {#sect2}
======================================================
Fig.\[fig:system\] illustrates the system model of the cognitive multi-hop relay system. The SU system consists of a cognitive source, a destination and several randomly distributed relays.
The system adopts certain Layer 3 ([network layer]{}) protocol to determine a route from the SU source node to the SU destination node, where the [route]{} is defined as a sequence of ordered nodes $\mathbb{R}=<R_0,R_1,...R_M>$, where $R_0, R_M$ are source node and destination node respectively. This route is assumed to be fixed throughout the communication session.
Denote the source as $R_0$, destination $R_M$ and $M-1$ cognitive relays, $\{R_1,...,R_{M-1}\}$, which are distributed between $R_0$ and $R_M$. The PU system consists of short-range wireless systems where the PU nodes are assumed to distribute uniformly (with a density of $\rho_p$) over the SU coverage area. Each of the PU node is assumed to have bursty activity with an active probability of $P_a$. The PU and the SU systems share common frequency spectrum and the SU system [can]{} access the channel only when all the involved PU nodes are idle. In the following, we shall elaborate on the channel model, control policy and the end-to-end throughput of the SU cognitive multi-hop relay system.
Channel Model
-------------
Figure \[fig:obtainCSI\] illustrates the signaling flow in multi-hop relay system. For the sensing of PU activity, we adopt the distributed sensing and centralized data fusion model as in IEEE 802.22. For instance, there are periodic quiet periods in the SU system that enable the sensing of PU activity. During the quiet periods, the [SUs]{} sense the PU activity locally and sends the sensing results to [the other]{} SU nodes. The SUs exchange the sensing results and update *continuous segment* (to be defined in the next subsection) information for data fusion. Define $A_m\in\{0,1\}$ as the sensing result which represents the availability of the shared spectrum to the SU system ($A_m=1$ denotes that the shared spectrum is available to SU node $R_m$) and $\mathbf{A}=(A_1,...,A_m)$ be the vector of PU activity states for the $M$ SU nodes. We assume an SU node $R_m$, [$m=\{0,1...,M\}$]{} has access ($A_m=1$) to the shared spectrum [if and only if]{} the nearest active PU node is at least $D_0$ meters away from the SU node[^4]. Furthermore, assume that $\mathbf{A}$ remains quasi-static between [two consecutive]{} sensing periods. This is a reasonable assumption as the burstiness of the PU nodes are of a longer time scale compared to the packet frame duration.
The received signal at the $j$-th SU node from the $i$-th SU node at the $k$-th frame is given by: $$\begin{aligned}
Y_{ij}(k)={\color{deep-blue} H_{ij}(k)\sqrt{D_{ij}}} X_{ij}(k) + Z_{ij}\end{aligned}$$ [where]{} $X_{ij}(k)$ is the transmitted data symbol from node $i$ to node $j$, $Z_{ij}$ is the zero-mean complex Gaussian channel noise (with normalized variance 1) and $G_{ij}(k)=|H_{ij}(k)|^2 D_{ij}$ is the combined channel loss (including both the large-scale path loss $D_{ij}$ and the microscopic fading $H_{ij}$) between node $i$ and $j$. The microscopic fading $H_{ij}$ is modeled as zero-mean, unit-variance complex Gaussian i.i.d (independent for different users) random variables. Let $\mathbf{G}(k)=\{G_{ij}(k):i\neq j,
i,j\in\{0,1,...,M\}\}$ be the global channel state (GCS) information. We assume $\mathbf{G}$ is quasi-static within a frame. For practical considerations, we have the following restrictions on the knowledge of the channel states.
- [**Local Knowledge of Channel States:**]{} We assume each of the SU node only has knowledge of the local channel state (LCS, to be defined below) and global PU activity state $\mathbf{A}$ (which remains quasi-static between two consecutive sensing periods).
- [**Causal Knowledge of Channel States:**]{} We assume that each SU node only has causal knowledge of the LCS and cannot predict into the future.
Specifically, we assume at the $k$-th frame, SU node m only has knowledge about the current LCS: $\mathbf{G}_m(k)$.
System State, Hopping and Power Control Policy, System State Transition Kernel.
-------------------------------------------------------------------------------
In this section, we shall formally define the control policy in the cognitive multi-hop relaying system. The multi-hop relay network operates in a DF manner with half-duplex constraint. At each frame, the upstream SU node transmits a packet of $B$ bits to its down-stream nodes using a transmit power which could be dynamically adjusted based on the current LCS knowledge. The down-stream SU node(s) attempt to decode the $B$-bits packet before it [can]{} forward to the next hop.
In this paper we consider dynamic spatial reuse in the cognitive multi-hop relay system as illustrated in Figure \[fig:relay scheme\]. For any given PU states $\mathbf{A}$, the multi-hop relay chain will be partitioned into several segments, which is defined as:
\[def:segment\] A continuous segment $L_{ij}$ in the cognitive multi-hop relay chain is defined as a sequence of nodes $<R_i,...,R_j> \subseteq \mathbb{R}$ such that: $$\begin{aligned}
S_{i-1} =0, S_{i} = \ldots = S_{j} =1, S_{j+1} = 0,
\;\; i,j\in\{1,2...,M\} \label{eqn:conseg}\end{aligned}$$
(Define $S_{-1} = S_{M+1} = 0$). The nodes $R_i$ and $R_j$ are called the [*head-node*]{} and the [*end-node*]{} of the continuous segment respectively. Define the probability of $\{R_i,...,R_j\}$ forms a continuous segment as $\Pr(i,j)=\Pr(S_{i-1} =0, S_{i} =
\ldots = S_{j} =1, S_{j+1} = 0)$.
Spatial reuse is allowed only for relays in different segments of the partition. Hence, relays in different segments can transmit different information simultaneously without interfering each other. Packets are stored at the end-node of each continuous segment and the end-node are not allowed to transmit except when the down-stream PU activity becomes idle. However, for relays in one continuous segment, they have to obey the TDMA constraint and cannot transmit different information simultaneously at any given time.
Within a continuous segment $L_{ij}$ induced by the PU activity $\mathbf{A}$, we shall define the hopping and power control policy as follows:
Suppose $R_i\sim R_j$ induced by a continuous segment $L_{ij}$ under a PU activity state $\mathbf A$. System state of $L_{ij}$ at frame index[^5] $k\in\{1,2...j-i\}$ is given by: $\eta_{ij}(k)=\{s_{ij}(k),\mathbf{G}_{s_{ij}(k)}\}$, where $s_{ij}(k)\in\{i,i+1,...j\}$ denotes the index of the source node at frame $k$, $s_{ij}(1)=i$; $\mathbf{G}_{s_{ij}(k)}$ is the LCS at node $s_{ij}(k)$.
A stationary policy $\Omega_{ij}$ is a mapping from the current system state $\eta_{ij}(k)$ to the corresponding hopping and power control actions. The policy $\Omega_{ij} = \{\mathcal{L}_{ij}, \mathcal{P}_{ij}\} $, where:
- [*Forward hopping policy $\mathcal{L}_{ij} $*]{}: $l_{ij}(k)=\mathcal{L}_{ij}(\eta_{ij}(k)), k\in \{1,2...j-i\} $, where the hopping control action (destination node index at frame $k$) has to satisfy the constraint: $s_{ij}(k)\le l_{ij}(k)\le j $, with the left inequality strictly holds when $s_{ij}(k)<j$.
- [*Dynamic power control policy $\mathcal{P}_{ij} $*]{}: $P_{ij}(k)=\mathcal{P}_{ij}(\eta_{ij}(k)),k\in \{1,2...j-i\} $, where the power control action (transmitting power at frame $k$) shall satisfy $P_{ij}(k)>0$.
The source node of at the $k+1$-th frame $s_{ij}(k+1)$, is determined by the hopping control action in the previous frame $l_{ij}(k)$. Furthermore, the distribution of the channel state $\mathbf{G}_{s_{ij}(k+1)}$ is independent of the previous system states $\eta_{ij}(k)$ due to the casual knowledge assumption. Hence, the [*state transition kernel*]{} of the system state $\{\eta_{ij}(k)\}$ is given by: $$\begin{aligned}
\Pr(\eta_{ij}(k+1)|\eta_{ij}(k),\Omega_{ij})=\mathbf{1}\left(s_{ij}(k+1)=l_{ij}(k)\right)\cdot\Pr(\mathbf{G}_{s_{ij}(k+1)})\label{eqn:kernel}\end{aligned}$$
Strictly speaking, the forward hopping policy $\mathcal L$ does not contain all possible hopping sequences w.r.t. a given route $\mathbb{R}$. For example, potential loops (e.g. $R_i\rightarrow
R_j\rightarrow R_i$) are excluded. Note that it is an intractable problem to optimize w.r.t. general hopping policies (including loops) due to the enormous possible policies involved. Instead, we shall restrict to forward hopping policy only and from which, we could exploit the structure in the policy space to derive much simpler solutions.
End-to-End Throughput with Dynamic Spatial Reuse and Forward Hopping Control
----------------------------------------------------------------------------
In order for a SU node to forward a packet, in any [*continuous segment*]{}, the node itself must be able to decode the packet first (DF). Suppose a node is able to decode if and only if the total mutual information received is no less than B bits. Hence, we have: $$\begin{aligned}
T_{ij}(k)\cdot\log(1+ G_{s_{ij}(k)l_{ij}(k)}(k)
P_{ij}(k)) \geq B, \;\; k \in
\{1,2...j-i\}, s_{ij}<j\label{eqn:info}\end{aligned}$$
where $i,j$ satisfy and $T_{ij}(k)$ is the transmitting time of the $k$-th frame in *continuous segment* $L_{ij}$.
We first formally define the [*per-hop reward and cost*]{} below.
Define the [*reward*]{} at the $k$-th frame as the time taken to transmit 1 bit at the $k$-th frame: $$\begin{aligned}
T(\eta_{ij}(k),\Omega_{ij})=\left\{\begin{array}{l}\frac{1}{\log(1+
G_{s_{ij}(k)l_{ij}(k)}(k) P_{ij}(k))}\mbox{ when: } s_{ij}(k)<j
\\ 0 \mbox{ otherwise.}
\end{array}\right. \end{aligned}$$ Define the cost at the $k$-th frame as the power consumed to transmit 1 bit at the $k$-th frame: $$\begin{aligned}
P(\eta_{ij}(k),\Omega_{ij})=\left\{\begin{array}{l}\frac{P_{ij}(k) }{\log(1+
G_{s_{ij}(k)l_{ij}(k)}(k) P_{ij}(k))}\mbox{ when:
}s_{ij}(k)<j\\0\mbox{ otherwise.}\end{array}\right. \end{aligned}$$
Note that $T_{ij}(k)=B\cdot T(\eta_{ij}(k),\Omega_{ij})$ and hence the average data rate in the [*continuous segment*]{} $L_{ij}$ can be expressed as: $$\begin{aligned}
U_{ij}= E^{\Omega_{ij}}\left(\frac{B}{\sum_{k\in\{1,2...j-i\},s_{ij}(k)<j}T_{ij\_k}}\right)
=\label{eqn:av-thp-sub1}E^{\Omega_{ij}}
\left(\frac{1}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right)\end{aligned}$$
where the expectation $E^{\Omega_{ij}}$ is taken w.r.t. the probability measure [*induced*]{} by the control policy $\Omega_{ij}$ and the transition kernel in . Similarly, average power consumption $\overline{P}_{ij}$ in $L_{ij}$ can be expressed as: $$\overline{P}_{ij}=E^{\Omega_{ij}}\left(\frac{\sum_{k=1}^{j-1}P(\eta_{ij}(k),\Omega_{ij})}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right)\label{eqn:av-subpower}$$
The end-to-end average throughput of the cognitive multi-hop system can be written as the weighted sum of average data rate of all [*continuous segments*]{} with end-node $R_M$: $$\label{eqn:av-thp1}\overline{U}(\Omega) =
\sum_{i=0}^{M-1} \Pr(i,M) U_{iM}$$
The average sum-power constraint is given by: $$\label{eqn:tx-pwr1} \sum_{i=0}^{M-1}\sum_{j=i+1}^{M}\Pr(i,j)
\overline{P}_{ij}\leq P_0$$
Moreover, the conventional flow-balance constraint[^6] is given by: $$\begin{aligned}
\label{eqn:flow-balance1} \sum_{i=0}^{m-1}\Pr(i,m) U_{im}&\ge&
\sum_{j=m+1}^{M}\Pr(m,j) U_{mj} \;\forall m \in \{1 ,..., M-1\}\end{aligned}$$
Problem Formulation
===================
[ Note that the conventional flow-balance constraint in may not be convex[^7]. ]{} To solve this issue, we introduce [a]{} new balance criteria, namely the [*section flow-balance criteria*]{}. For instance, we consider the sum of average data rate passing through each section (rather than each node). Specifically, the sum-average data rate passing through the $m$-th section ($m \in
\{1,2,...M\}$ as illustrated in Fig.\[fig:sectionbalance\]). Define: $ \bar{U}_m = \sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j)
U_{ij}$. The [*section flow-balance criteria*]{} is given by: $$\label{eqn:flow-balance2} \overline{U}_m \ge \overline{U}_{m+1},\;
\forall m \in \{1 ,..., M-1\}$$
In the following lemma, we shall illustrate that the section flow-balance criteria is in fact equivalent to the conventional per-node flow-balance:
\[Lem:flow-balance\] \[[*Equivalence of the flow balance criteria*]{}\] The conventional [*per-node*]{} flow balance constraint in (\[eqn:flow-balance1\]) is equivalent to the [*per-section*]{} flow balance criteria in (\[eqn:flow-balance2\]).
please refer to Appendix \[PLem:flow-balance\] for the proof.
Lemma \[Lem:flow-balance\] gives an equivalent form for traditional flow-balance criteria. Moreover, note that the objective $\overline{U}((\Omega))$ in (\[eqn:av-thp1\]) is equal to: $$\begin{aligned}
\nonumber \overline{U}(\Omega) &=& \sum_{i=0}^{M-1}
\Pr(i,M) U_{iM} \;= \sum_{i=0}^{M-1} \sum_{j=M}^{M} \Pr(i,j)
U_{ij}\;= \bar{U}_{M}
\\ \label{eqn:av-thp2} &=& \min(\{\bar{U}_{1}, \bar{U}_{2},...,\bar{U}_{M}
\}) \mbox{ (Due to section flow balance criteria
(\ref{eqn:flow-balance2}))}\end{aligned}$$ where $\Omega$ is the overall control policy: $\Omega=\{\Omega_{ij}, \forall i,j$ that satisfies under a PU actively state $\mathbf{A}\}$.
From (\[eqn:av-thp2\]), the optimization problem can be formulated as:
\[prob:org\] $$\begin{aligned}
&&\label{eqn:obj_2}\bar{U}=\max_{\Omega}\left[\min_{m\in\{1,..., M\}}
\sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j){U}_{ij}\right]
\\&&\nonumber \mbox{Subject to:}
\\\label{eqn:power_2} && \sum_{i=0}^{M-1}\sum_{j=i+1}^{M}\Pr(i,j)
\overline{P}_{ij}\leq P_0\end{aligned}$$
where: ${U}_{ij}$, $\overline{P}_{ij}$ is given by and respectively.
Decomposition of Main Problem
-----------------------------
The optimization problem in is too complex to solve directly. Furthermore, due to the causality constraint in the control policies ${\cal P}$ and ${\cal L}$, the solution is not trivial and brute-force solution will not lead to viable solutions. However, it is worthy noting that for a given PU activity state $\mathbf{A}$, operations on different *continuous segment* are naturally separated from each other. (e.g. as in Fig \[fig:relay scheme\], when $S_4=0$, hopping and power control policy in segment $R_0\sim R_3$ has no direct [ influence]{} on that in $R_5\sim R_6$). Making use of this insight, we shall first decompose the problem into a [*master problem*]{} and a [*sub-problem*]{}. Define: ${\cal{P}}_{main}=\{\overline{P}_{ij}\}, \;i,j \in \{0,1,...,M\},
\;i<j$.
We have the following decomposition theory:
Optimization problem consisting of a master problem ( Problem \[prob:mas\], with ${\cal{P}}_{main}$ as the optimization policy) and $\frac{M(M-1)}{2}$ subproblems (Problem \[prob:sub\], with ${\cal L}_{ij},{\cal P}_{ij}$ as the optimization policies) is equivalent to Problem \[prob:org\].
\[prob:mas\] $$\begin{aligned}
&&\label{eqn:obj_3}\overline{U}=\max_{{\cal
P}_{main}}\left[\min_{m\in\{1,..., M\}}
\sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j){U}^*_{ij}(\overline{P}_{ij})\right]
\\&&\nonumber \mbox{Subject to:}
\\\label{eqn:power_3b} && \sum_{i=0}^{M-1}\sum_{j=i+1}^{M}\Pr(i,j) \overline{P}_{ij}\leq P_0\end{aligned}$$
\[prob:sub\] $$\begin{aligned}
\label{eqn:obj_sub_1} &&{U}^*_{ij}(\overline{P}_{ij}) =\max_{\mathcal{L}_{ij},\mathcal{P}_{ij}}E^{\Omega_{ij}}
\left(\frac{1}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right)
\\&&\nonumber \mbox{Subject to:}
\\\label{eqn:power_sub_1} &&E^{\Omega_{ij}}\left(\frac{\sum_{k=1}^{j-1}P(\eta_{ij}(k),\Omega_{ij})}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right)\le\overline{P}_{ij}\end{aligned}$$
\[thm:decompose\]
Please refer to Appendix \[Pthm:decompose\] for the proof.
Decentralized Hop Selection and Power Control Algorithm
=======================================================
Solving the Sub Problem
-----------------------
To satisfy the causality constraint of the control policy on the local CSI, we have to model the subproblem in a recursive form so as to apply dynamic programming (DP) [@DPbook]. However, problem (\[eqn:obj\_sub\_1\]) cannot be expressed in a recursive form and hence, could not be divide-and-conquered. To tackle the challenges, we shall solve a lower bound version of the problem. We shall show that the lower bound solution is indeed asymptotically tight for large number of nodes.
### **Asymptotically Optimal Solution**
We first elaborate a suboptimal solution for the subproblem (Problem \[prob:sub\]). Let $$\begin{aligned}
\Omega_{ij}^{LB}= \arg\min_{\Omega_{ij}}E^{\Omega_{ij}} \left[\sum_{k=1}^{j-1}
T(\eta_{ij}(k),\Omega_{ij}) +
\lambda_{ij}\left(P(\eta_{ij}(k),\Omega_{ij})-\overline{P}_{ij}T(\eta_{ij}(k),\Omega_{ij})\right)
\right] \label{eqn:LB3}\end{aligned}$$ where the parameter $\lambda_{ij}$ in the suboptimal solution $\Omega_{ij}^{LB}$ is given by the roots of the equation[^8]: $$\label{eqn:power_sub_3}
E^{\Omega_{ij}}\left(\frac{\sum_{k=1}^{K_{ij}}P(\eta_{ij}(k),\Omega_{ij})
}{\sum_{k=1}^{K_{ij}}T(\eta_{ij}(k),\Omega_{ij})}\right)
=\overline{P}_{ij}$$
Note that the solution $\Omega_{ij}^{LB}$ is a feasible but suboptimal solution of the subproblem (Problem \[prob:sub\]). We have the following lemma about the property of the suboptimal solution $\Omega_{ij}^{LB}$.
\[Lem:lagrange\] If the following conditions are satisfied: 1) For any $\epsilon>0$, there exists a finite $C>0$ such that when $|s-t|\geq C$, $G_{st} < \epsilon$; and 2) $G_{st}\ge G_{st'}$, $G_{st} \ge G_{s't}$ when $t'\ge t>s\ge s'$; then we have: $U_{ij}^{LB}(\overline{P}_{ij}) \rightarrow
U^*_{ij}(\overline{P}_{ij})$, as $|j-i|\rightarrow \infty$. where $U_{ij}^{LB}(\overline{P}_{ij})$ is the average throughput of the segment $R_i\sim R_j$ using the suboptimal control $\Omega_{ij}^{LB}$.
Please refer to Appendix \[pLem:langrange\] for the proof.
\[lab:sub\] The condition 1) in Lemma \[Lem:lagrange\] means that the nodes are not *“over concentrated”* on one spot. This is is a mild requirement, which only excludes the special topologies where there are infinite number of nodes over a finite coverage area. The condition 2) refers to the path loss dominated situations, which applies for medium-range (over 2-5 km) multi-hop networks.
[As a result, the suboptimal solution $\Omega^{LB}_{ij}$ has reasonable performance in general cases (as will be illustrated in Section V) and it is asymptotically optimal for large number of nodes. ]{}In order to derive $\Omega^{LB}_{ij}$, we shall first express into a recursive form and solve the problem by divide-and-conquer using DP. Define $$\label{eqn:g_1}g(\eta_{ij}(k);P_{ij}(k),l_{ij}(k))=\frac{1+\lambda_{ij}(P_{ij}(k)-\overline{P}_{ij})}{\log(1+P_{ij}(k)G_{s_{ij}(k)l_{ij}(k)}(k))}$$ then the problem can be expressed recursively as: $$\label{eqn:obj_sub_DP1}
J(s_{ij}(k))=E_{\mathbf{G}_{s_{ij}(k)}}
[\min_{P_{ij}(k),l_{ij}(k)}(g(\eta_{ij}(k);P_{ij}(k),l_{ij}(k))+J(l_{ij}(k)))]$$ where $J(m)$ is called the [*expected cost*]{} from node $R_m$ to $R_j$. Note that $J(j)=0$ and $J(s_{ij}(1))=J(i)$ gives the value of (\[eqn:LB3\]). As a result of the recursive form in , the [*backward recursion algorithm*]{} to solve problem is summarized in the following.
$$\label{eqn:opt_p_sub}\frac{G_{s_{ij}(k)l_{ij}(k)}(k)}{(1+P_{ij}(k)G_{s_{ij}(k)l_{ij}(k)}(k))
\log(1+P_{ij}(k)G_{s_{ij}(k)l_{ij}(k)}(k))+(\overline{P}_{ij}-P_{ij}(k))G_{s_{ij}(k)l_{ij}(k)}(k)}
= \lambda_{ij}$$
\[Remark:sub\] Note that the memory size of the table in the offline recursion is $j-i$. The computational complexity for the online algorithm in each step $k$ is only of the order $j-i$. Hence, the online algorithm has worst case complexity ${\cal O}(M^2)$ and worst case memory requirement ${\cal O}(M)$ for each continuous segment $i,j$.
Solving the Main Problem
------------------------
After solving for the subproblem, we shall focus on solving the main problem based on $U^{LB}_{ij}(\overline{P}_{ij})$ (which is of a longer time scale) in this section. We first establish the following Theorem regarding the concavity of $U^{LB}_{ij}(\overline{P}_{ij})$ w.r.t. $\overline{P}_{ij}$.
\[Lem:subconcave\] The lower bound ($U^{LB}_{ij}(\overline{P}_{ij})$) of $U^*_{ij}(\overline{P}_{ij})$ is a concave function of $\overline{P}_{ij}$.
Please refer to Appendix \[pLem:subconcave\] for the proof.
From Lemma \[Lem:subconcave\], it is easy to deduce that the lower-bound version of the master problem in (\[eqn:obj\_3\]) \[with $U_{ij}^*(\overline{P}_{ij}$) replaced by $U_{ij}^{LB}(\overline{P}_{ij})$\] is a convex optimization problem. As a result, the standard gradient search could be applied to solve the master problem. Please refer to Figure \[fig:alg\] for the detailed algorithm description.
Note that the offline recursion needs to be updated only when there are changes [ in]{} the PU statistics or the SU path loss and in practice, the above offline algorithm is computed over a long time scale. Combining the master problem and the subproblems the total memory requirement of the offline table in algorithm 1 is $\mathcal {O}(M^3)$.
Simulation Results
==================
In this section, we shall illustrate the performance of the proposed scheme by simulation. We consider a multi-hop cognitive relay system with $6$ nodes ($\{R_0,R_1,...R_5\}$) and 6 PUs (one PU in the neighborhood of each SU node). The distance between $R_0$ and $R_5$ is 5, and the other 4 nodes randomly scatter between them. Path loss between two nodes $R_i, R_j$ is given by the “flat-earth model” [@flatearth]: $\log_{10}D_{ij}= -\alpha\log_{10}d_{ij} \;(dB)$ where $d_{ij}$ is the distance between the two nodes and $\alpha$ is the path loss exponent. The proposed scheme is compared with four schemes below:
- [**Direct transmission only (Baseline 1):**]{} $R_0$ transmit directly to $R_5$ when all PU remain silent ($S_i = 1, \forall i\{0,1,...5\}$). This is equivalent to the case without relay.
- [**Per-node transmission only (Baseline 2):** ]{} if $R_m$ ($\forall m\{0,1,...4\}$) received a packet in previous frames, it transmits this packet to $R_{m+1}$ when the PU activity permits ($A_m = S_{m+1} =1$). This corresponds to the traditional DF multi-hop relay scheme.
- [**Direct (per-node) transmission with dynamic spatial reuse (Baseline 3/4)**]{}: These two schemes adopt the same dynamic spatial reuse method as the proposed scheme. Yet, within each [*continuous segment*]{}, they adopt direct and per-node transmission respectively.
Figure \[fig:Main\_SNR\] and Figure \[fig:Main\_Pr\] illustrate the average end-to-end throughput ($\overline{U}$) versus the average SNR ($P_0$) and PU activity level ($\Pr(A_m=0)$) respectively. The proposed scheme achieves significant throughput gains over a wide range of SNR and PU activities. This gain is contributed by both the dynamic hop selection as well as dynamic power control. Comparison with baseline 1 illustrates how cognitive relay could help to increase the probability of access and efficiency of spectrum sharing in general. Comparison with baseline 2 and 3 illustrates the importance of joint dynamic power and opportunistic hop selection in cognitive multihop systems. The gain contributed by the dynamic hop selection is most significant under moderate SNR. At very high SNR, the dynamic hopping performance approaches that of the baseline 3, illustrating the system always perform one-hop direct transmission to avoid the half-duplex penalty. At very low SNR, the performance of the proposed scheme approaches that of the baseline 4, illustrating that the system prefer hop-by-hop transmission for SNR gain.
Figure \[fig:Main\_Pr\] illustrate that the dynamic hopping gain is more prominent under low PU activity. This is because at low PU activity, there is a higher chance of forming a longer [*continuous segment*]{} and hence, more flexible choices for the dynamic hop selection. Figure \[fig:Main\_converge\] illustrate the convergence rate of the off-line recursion for the *Main problem* (Algorithm 2). The proposed algorithm can achieve $90\%$ of the converged performance within 10 iterations and converges after about 30 iterations. This iteration efficiency is good enough for off-line algorithms.
Figure \[fig:Main\_node\] illustrates the normalized throughput $\frac{\bar{U}}{U_{max}}$ versus the average transmit SNR ($P_0$) for various number of cognitive relay nodes where $\frac{\bar{U}}{U_{max}}$ is obtained from brute-force numerical optimization of Problem 1. With $N=6$, we have over $95\%$ of the optimal performance. This illustrates that the proposed scheme is not only order-optimal but achieves close-to-optimal performance even in small to moderate number of cognitive relay nodes.
Summary
=======
In this paper, we have derived a low complexity hop selection and dynamic power control policies to maximize the average end-to-end throughput of the cognitive multi-hop SU system with dynamic spatial reuse. By exploiting the time-scale difference between the PU activity and the CSI dynamics, we decompose the problem into a *Master problem* and several *Sub Problems*. The solution obtained is decentralized in the sense that each node determines its next hop and transmit power based on the local and causal CSI only. The solution consists of an offline recursion and an online algorithm with worst case complexity ${\cal O}(M^2)$ and worst case memory requirement ${\cal O}(M^3)$. Furthermore, the solution is asymptotically optimal for large number of nodes. Significant throughput performance has been demonstrated.
Proof of Lemma \[Lem:flow-balance\] {#PLem:flow-balance}
===================================
$$\begin{aligned}
\nonumber\bar{U}_{m} - \bar{U}_{m+1}&=&
\sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j) U_{ij} -
\sum_{i=0}^{m}\sum_{j=m+1}^{M}\Pr(i,j) U_{ij}
\\ \nonumber &=& \sum_{i=0}^{m-1}\Pr(i,m) U_{i,m} -
\sum_{j=m+1}^{M}\Pr(m,j)U_{m,j} \;\;\;\;\; \forall m \in \{1 ,...,
M-1\}\end{aligned}$$
Hence: $$\begin{aligned}
\nonumber U_m \ge U_{m+1}&\Leftrightarrow&
\nonumber \sum_{i=0}^{m-1}\Pr(i,m) U_{i,m} -
\sum_{j=m+1}^{M}\Pr(m,j)U_{m,j} \ge 0 \\&\Leftrightarrow &
\sum_{i=0}^{m-1}\Pr(i,m) U_{i,m}\ge \sum_{j=m+1}^{M}\Pr(m,j)
U_{m,j}\end{aligned}$$
Proof of Lemma \[thm:decompose\] {#Pthm:decompose}
================================
To prove Problem \[prob:mas\] and Problem \[prob:sub\] are equivalent to Problem \[prob:org\], we first prove the following Lemma:
\[Lem:exchange\] Define: $$\begin{aligned}
V&=&\max_{\mathbf{X}}\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f_i(x_i)\right)\label{eqn:v_1}
\\V'&=&\min_{m\in\{1,2...M\}}\max_{\mathbf{X}}\left(\sum_{i=1}^{L}A_{mi}f_i(x_i)\right)\label{eqn:V_2}\end{aligned}$$ where $\mathbf{X}=\{x_i\in \mathbb{C}_i,i\in\{1,2...L\}\}$ are a set of independent variables. If $f_i(x_i)$ is finite and $\forall
m\in\{1,2...M\}, i\in\{1,2...L\}$, then: $$\begin{aligned}
V=V'=\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f^*_i\right)\label{eqn:exchange_result}\end{aligned}$$ where $f^*_i=\max_{x_i\in\mathbb{C}_i}f_i(x_i)$.
In general, switching of “$\max$” and “$\min$” is not allowed but there are two specific structures in Lemma \[Lem:exchange\]. that we are exploiting.
- [**Independency Property:**]{} $f_i(x_i), \forall i$ are mutually independent (i.e. they are not coupled by any common variables), as $\mathbf{X}=\{x_i,i\in\{1,2...L\}\}$ is a set of independent variables.
- [**Monotone Property:**]{} Since for every $m$ and $i$, $A_{mi}\ge 0$: $\forall m,i$, $\sum_{i=1}^{L}A_{mi}f_i(x_i)$ is an non-decreasing function of $f_i(x_i)$. As a result, $V$ is an non-decreasing function of $f_i(x_i)$, $\forall i\in\{1,2...L\}$.
[ Since $V$ is an non-decreasing function of $f_i(x_i)$ (*Monotone Property*), $f_i(x_i)\le f^*_i$, $\forall i\in\{1,2...M\}$:]{} $$\begin{aligned}
V\le
\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f^*_i\right)\label{eqn:v_le}\end{aligned}$$
[ Moreover, denote $x^*_i=\arg\max_{x_i\in\mathbb{C}_i}f_i(x_i)$, from the *Independency Property*, $\{x_i=x^*_i, i\in\{1,2...M\}\}$ is a feasible point for $V$. Hence:]{} $$\begin{aligned}
V\ge
\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f^*_i\right)\label{eqn:v_ge}\end{aligned}$$
Combining , $V=\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f^*_i\right)$.
On the other hand, since $\forall m$, $\max_{\mathbf{X}}\left(\sum_{i=1}^{L}A_{mi}f_i(x_i)\right)=\sum_{i=1}^{L}A_{mi}f^*_i$: $$\begin{aligned}
V'=\min_{m\in\{1,2...M\}}\left(\sum_{i=1}^{L}A_{mi}f^*_i\right)\end{aligned}$$
In Problem \[prob:org\], for a fixed $\mathcal{P}_{main}$, denote: $$\begin{aligned}
\nonumber \overline{U}(\mathcal{P}_{main})&=&\max_{{\mathcal
L},{\mathcal P}|\mathcal{P}_{main}}\min_{m\in\{1,..., M\}}
\left(\sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j){U}_{ij}\right)
\\&=&\max_{{\mathcal
L},{\mathcal P}|\mathcal{P}_{main}}\min_{m\in\{1,..., M\}}
\left(\sum_{i=0}^{M-1}\sum_{j=1}^{M}A(i,j,m){U}_{ij}\right)
\\\nonumber \mbox{where:}&& A(i,j,m)=\left\{
\begin{array}{l}\Pr(i,j)\;\;\mbox{if: } i<m\le j;
\\0\;\;\;\;\mbox{else} \end{array}\right.\end{aligned}$$
Note that: a) From (\[eqn:av-thp-sub1\]), , $U_{ij}$ and $\overline{P}_{ij}$ depends on different set of variables ${\cal L}_{ij}$ and ${\cal P}_{ij}$. Hence, for a given ${\cal{P}}_{main}=\{\overline{P}_{ij}\}$, constraint is decoupled and $\{\mathcal{L}_{ij},
\mathcal{P}_{ij}\}$ become independent variables for different $\{i,j\}$.
b\) From , for all $i,j$, $\Pr(i,j)\ge 0$, $A(i,j,m)\ge0$, $\forall i,j,m$.
Combining a) and b), we can apply Lemma \[Lem:exchange\] and obtain: $$\begin{aligned}
\label{eqn:exchange1}\overline{U}({\mathcal P}_{main})&=&\min_{m\in\{1,...,
M\}}\left( \sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j)
U^*_{ij}(\overline{P}_{ij})\right)\end{aligned}$$ where $U^*_{ij}(\overline{P}_{ij})$ is given by the solution of Problem \[prob:sub\]. Hence, we can rewrite the objective function as: $$\begin{aligned}
\overline{U}=\max_{{\mathcal P}_{main}}\overline{U}({\mathcal P}_{main})=\max_{{\mathcal P}_{main}}\min_{m\in\{1,...,
M\}}\left( \sum_{i=0}^{m-1}\sum_{j=m}^{M}\Pr(i,j)
U^*_{ij}(\overline{P}_{ij})\right)\end{aligned}$$ which is exactly the objective function in Problem \[prob:mas\]. Therefore, the optimal solution given by Problem \[prob:mas\] and Problem \[prob:sub\] shall be the same as Problem \[prob:org\].
Proof of Lemma \[Lem:lagrange\] {#pLem:langrange}
===============================
We shall prove that the suboptimal solution $\Omega^{LB}_{ij}$ is asymptotically optimal under the two conditions in Lemma \[Lem:lagrange\]. We shall first prove the following Lemma:
We partition the continuous segment $R_i\sim R_j$ into $R=\lceil\frac{j-i}{C}\rceil$ clusters: $\mathbb{V}_r=\{i+rC,i+rC+1,...\min(i+r(C+1)-1,j)\}$, $r\in\{0,1...R-1\}$. As for any $\epsilon>0$, there exists a finite $C>0$ such that when $|s-t|\geq C$, $G_{st} < \epsilon$, let $\epsilon\ll\frac{1}{\overline{P}_{ij}}$, we have:[^9] $$\begin{aligned}
l_{ij}(k)-s_{ij}(k)< C,
\forall k\in\{1,2...j-i\} \label{eqn:hop}\end{aligned}$$
Denote $T_r=\sum_{l_{ij}(k)\in\mathbb{V}_r}T(\eta_{ij}(k),\Omega_{ij})
=\sum_{s_{ij}(k)\neq j,l_{ij}(k)\in\mathbb{V}_r}\frac{1}{\log(1+
G_{s_{ij}(k)l_{ij}(k)}(k) P_{ij}(k))}$, then: $\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})=\sum_{r=0}^{R-1}T_r$. Moreover, from , we have: $1\le |s_{ij}(k)\neq
j,l_{ij}(k)\in\mathbb{V}_r|\le C$. Moreover, as in practice, the time duration to transmit one bit should be positive and finite, there should exist $T_{\min},T_{\max}\in\mathbb{R}^+$ such that $T_{\min}\le T(\eta_{ij}(k),\Omega_{ij})\le T_{\max}$, $\forall
\eta_{ij}(k)$. Hence we have: $$\begin{aligned}
T_{\min}\le T_r \le C T_{\max},\forall
r\in\{0,1,...R-1\}\label{eqn:range}\end{aligned}$$
As we shall proof in Appendix \[pLem:cov\], we have the following results concerning the covariance between $\{T_r\}$:
With Lemma \[lem:cov\] and , we have: $$\begin{aligned}
\nonumber \mbox{Var}(\frac{\sum_{r=0}^{R-1}T_r}{E^{\Omega_{ij}}\sum_{r=0}^{R-1}T_r})
&=&\frac{\sum_{r=0}^{R-1}\mbox{Var}(T_r)+2\sum_{r=1}^{R-1}\mbox{Cov}(T_r,\sum_{s=0}^{r-1}T_s)}{\left(\sum_{r=0}^{R-1}E^{\Omega_{ij}}T_r\right)^2}
\le\frac{\sum_{r=0}^{R-1}\mbox{Var}(T_r)}{R^2T^2_{\min}}
\\&\le&\frac{RC^2T^2_{\max}}{R^2T^2_{\min}}\rightarrow
0 \;\; \mbox{as: } R=\lceil\frac{j-i}{C}\rceil\rightarrow
\infty\label{eqn:variance}\end{aligned}$$
Substitute into Chebyshev inequality, is proved; can also be proved through similar process. We shall omit the details due to page limit.
From , : $$\begin{aligned}
\nonumber \label{eqn:asyn2}&&E^{\Omega_{ij}}\left[\frac{1}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right]
\rightarrow
\frac{1}{\sum_{k=1}^{j-i}E^{\Omega_{ij}}T(\eta_{ij}(k),\Omega_{ij})}\mbox{ and}
\\ \nonumber \label{eqn:asyn2}&&E^{\Omega_{ij}}\left[\frac{\sum_{k=1}^{j-i}P(\eta_{ij}(k),\Omega_{ij})}{\sum_{k=1}^{j-i}T(\eta_{ij}(k),\Omega_{ij})}\right]
\rightarrow
\frac{\sum_{k=1}^{j-i}E^{\Omega_{ij}}P(\eta_{ij}(k),\Omega_{ij})}{\sum_{k=1}^{j-i}E^{\Omega_{ij}}T(\eta_{ij}(k),\Omega_{ij})}\mbox{ in probability when } j-i\rightarrow
\infty\end{aligned}$$
Hence, for sufficiently large $j-i$, Problem \[prob:sub\] can be equivalently rewritten as: $$\begin{aligned}
&&\min_{\Omega_{ij}}\sum_{k=1}^{j-i}E^{\Omega_{ij}}T(\eta_{ij}(k),\Omega_{ij})
\\&\mbox{S.t.:}&\sum_{k=1}^{j-i}E^{\Omega_{ij}}\left(P(\eta_{ij}(k),\Omega_{ij})-\overline{P}_{ij}T(\eta_{ij}(k),\Omega_{ij})\right)\le0\end{aligned}$$
Observe that the Lagrangian dual function of the above problem is exactly (\[eqn:LB3\]). Hence, $U^{LB}(\overline{P}_{ij})
\rightarrow U^*_{ij}(\overline{P}_{ij})$ for sufficiently large $j-i$.
Proof of Lemma \[lem:cov\] {#pLem:cov}
==========================
We shall first prove the following Lemma:
Denote $N^-_{a}=|\{n:a_n<0\}|$, $N^+_{b}=|\{n:b_n>0\}|$, where $|\mathbb{A}|$ means the cardinality or set $\mathbb{A}$. If $N^-_a=N^+_b$, then obviously $\sum_{n=0}^{N}a_nb_n\le0$; Otherwise, without loss of generality, assume $N^-_a>N^+_b$ and then: $$\begin{aligned}
\nonumber
\sum_{n=1}^{C}p_na_nb_n=\sum_{n=0}^{N^+_b-1}p_na_nb_n+\sum_{n=N^+_b}^{N^-_a-1}p_na_nb_n+\sum_{n=N^-_a}^{N}p_na_nb_n\le\sum_{n=0}^{N^+_b-1}p_na_nb_n+\sum_{n=N^+_b}^{N^-_a-1}p_na_nb_n
\\\le a_{(N^+_b-1)}\sum_{n=0}^{N^+_b-1}p_nb_n+a_{(N^+_b)}\sum_{n=N^+_b}^{N^-_a-1}p_nb_n \le
(a_{(N^+_b-1)}-a_{(N^+_b)})\sum_{n=0}^{N^+_b-1}p_nb_n\le0
\label{eqn:loser}\end{aligned}$$
Recall the system state transition kernel: $$\begin{aligned}
\Pr(\eta_{ij}(k)|\eta_{ij}(k-1),\Omega_{ij})=\mathbf{1}\left(s_{ij}(k)=l_{ij}(k-1))\right)\Pr(\mathbf{G}_{s_{ij}(k)})\end{aligned}$$
It can be observed that conditioned on the source node at the frame $k$: $s_{ij}(k)$, $\{\eta_{ij}(k),\eta_{ij}(k+1),...\}$ are independent of $\{\eta_{ij}(1),\eta_{ij}(2),...\eta_{ij}(k-1)\}$. Correspondingly, $\{T(\eta_{ij}(s),\Omega^{ij}),s\in\{k,k+1...\}\}$ are conditionally independent of $\{T(\eta_{ij}(s),\Omega^{ij}),s\in\{1,2...,k-1\}\}$. Denote $l_{r\_\min}=\min(l_{ij}(k):l_{ij}(k)\in\mathbb{V}_r)$. Then: conditional on $l_{r\_\min}$, $T_r$ is independent of $\{T_s,s\in\{0,1,...r-1\}\}$. Hence: $$\begin{aligned}
E^{\Omega_{ij}}\left.\left(T_r\cdot\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}=x\right)=E^{\Omega_{ij}}(T_r|l_{r\_\min}=x)E^{\Omega_{ij}}\left.\left(\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}=x\right)\label{eqn:ind}\end{aligned}$$ where $x\in\{i+rC,i+rC+1,...i+rC+|\mathbb{V}_r|-1\}$. Denote $k_{r\_\min}=\min(k:l_{ij}(k)\in\mathbb{V}_r)$. Since $G_{st}\ge
G_{st'}$ when $s< t\le t'$, $T(\eta_{ij}(k_{r\_\min})-1)$ is an non-decreasing function of $l_{r\_\min}$. Correspondingly, $E^{\Omega_{ij}}\left.\left(\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}\right)$ is a non-decreasing function of $l_{r\_\min}$. Similarly, as $G_{st}\ge G_{s't}$ when $s'\le s< t$, $E^{\Omega_{ij}}(T_r|l_{r\_\min})$ is a non-increasing function of $l_{r\_\min}$. Let $E^{\Omega_{ij}}\left.\left(\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}=x\right)-E^{\Omega_{ij}}\left(\sum_{s=1}^{r-1}T_s\right)=a_x$, $E^{\Omega_{ij}}(T_r|l_{r\_\min}=x)-E^{\Omega_{ij}}(T_r)=b_n$, $\Pr(l_{r\_\min}=x)=p_n$ and substitute to Lemma \[Lem:ab\]: $
\sum_{x=i+rC}^{i+rc+|\mathbb{V}_r|-1}\Pr(l_{r\_\min}=x)\left(E^{\Omega_{ij}}(T_r|l_{r\_\min}=x)-E^{\Omega_{ij}}(T_r)\right)
\left(E^{\Omega_{ij}}\left.\left(\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}=x\right)-E^{\Omega_{ij}}\left(\sum_{s=1}^{r-1}T_s\right)\right)\le0
$. From this result and : $$\begin{aligned}
\nonumber
&&\mbox{Cov}(T_r,\sum_{s=0}^{r-1}T_s)=E^{\Omega_{ij}}\left(T_r\sum_{s=1}^{r-1}T_s\right)-E^{\Omega_{ij}}\left(T_r\right)E^{\Omega_{ij}}\left(\sum_{s=1}^{r-1}T_s\right)
\\\nonumber &&=\sum_{x=i+rC}^{i+rc+|\mathbb{V}_r|-1}\Pr(l_{r\_\min}=x)\left(E^{\Omega_{ij}}(T_r|l_{r\_\min}=x)-E^{\Omega_{ij}}(T_r)\right)
\\&&\;\;\;\;\cdot\left(E^{\Omega_{ij}}\left.\left(\sum_{s=1}^{r-1}T_s\right|l_{r\_\min}=x\right)-E^{\Omega_{ij}}\left(\sum_{s=1}^{r-1}T_s\right)\right)\le0\end{aligned}$$
Proof of Lemma \[Lem:subconcave\] {#pLem:subconcave}
=================================
Due to the [*Theorem of Lagrangian*]{} ([@FirstOpt], section 5.2.3), we have $$\label{eqn:sub_grad} \frac{\partial U^{LB}_{ij}}{\partial
\overline{P}_{ij}}=\lambda^*_{ij}(\overline{P}_{ij})$$ where $\lambda^*_{ij}(\overline{P}_{ij})$ is the Lagrange [ multiplier]{} obtained in the subproblem via Algorithm \[Alg:sol\_sub\]. Hence, Lemma \[Lem:subconcave\] holds if and only if $\lambda^*_{ij}(\overline{P}_{ij})$ is a non-increasing function of $\overline{P}_{ij}$. Note that in (\[eqn:opt\_p\_sub\]), $\forall
k,l_k,G_{s_{ij}(k)l_{ij}(k)}(k)
>0$: $P_{ij}(k)$ decreases as $\lambda^*_{ij}$ increases. Substitute this result to (\[eqn:power\_sub\_3\]) and it is obvious that $\lambda^*_{ij}(\overline{P}_{ij})$ decreases as $\overline{P}_{ij}$ increases.
[1]{}
Valenti, M.C.; “Practical relay networks: a generalization of hybrid-ARQ Bin Zhao”; *IEEE Journal on Selected Areas in Communications*, Volume 23, Issue 1, Jan. 2005 Page(s):7 - 18
Zhong Zhou; Shengli Zhou; Jun-Hong Cui; Shuguang Cui; “[Energy-Efficient Cooperative Communication Based on Power Control and Selective Single-Relay in Wireless Sensor Networks]{}”, 1em plus 0.5em minus 0.4em*IEEE Transactions on Wireless Communications*, Volume 7, Issue 8, August 2008 pp.:3066 - 3078.
Channels Ikki, S.; Ahmed, M.H.; “[Performance Analysis of Dual-Hop Relaying Communications over Generalized Gamma Fading Channels]{}”, 1em plus 0.5em minus 0.4em*IEEE GLOBECOM ’07*. 26-30 Nov. 2007 pp.:3888 - 3893
G.K.; Tsiftsis, T.A.; Mallik, R.K.; “[Bounds for multihop relayed communications in nakagami-m fading Karagiannidis]{}”,1em plus 0.5em minus 0.4em*IEEE Transactions on Communications*, Volume 54, Issue 1, Jan. 2006 pp.:18 - 22
Hanbyul Seo; Inyoung Mok; Byeong Gi Lee; “[Determination of optimal transmission power in wireless relay networks with generalized error model]{}” 1em plus 0.5em minus 0.4em *IEEE Transactions on Wireless Communications*, Volume 6, Issue 12, December 2007 Page(s):4233 - 4237
Farhadi, G.; Beaulieu, N.C.; Hanbyul Seo; Inyoung Mok; Byeong Gi Lee; “[On the performance of amplify-and-forward cooperative systems with fixed gain relays]{}”, 1em plus 0.5em minus 0.4em*IEEE Transactions on Wireless Communications*, Volume 7, Issue 5, Part 2, May 2008 Page(s):1851 - 1856
Ribeiro, A.; Xiaodong Cai; Giannakis, G.B.; “[Symbol error probabilities for general Cooperative links]{}” 1em plus 0.5em minus 0.4em*IEEE Transactions on Wireless Communications*, Volume 4, Issue 3, May 2005 Page(s):1264 - 1273
Seung-Jun Kim; Xiaodong Wang; Madihian, M.;*Optimal resource allocation in multi-hop OFDMA wireless networks with cooperative relay* 1em plus 0.5em minus 0.4emIEEE Transactions Wireless Communications, on Volume 7, Issue 5, Part 2, May 2008 Page(s):1833 - 1838
S. Biswas and R. Morris; “[Opportunistic routing in multi- hop wireless networks]{}”, 1em plus 0.5em minus 0.4em*ACM SIGCOMM Computer Com- munication Review*, vol. 34, number 1, pp. 69¨C74, 2004.
E. Rozner, J. Seshadri, and Y. Mebta, “[Simple opportunis- tic routing protocol for wireless mesh networks]{}”, 1em plus 0.5em minus 0.4em *WiMesh 2006*. 2nd IEEE Workshop on, pp. 48¨C54, 2006.
N. Xiao, L. Ding, M. Li, and M. Wu, “[Geographic Proba- bilistic Routing Protocol forWirelessMesh Network]{}”,1em plus 0.5em minus 0.4em*LECTURE NOTES IN COMPUTER SCIENCE*, vol. 4847, p. 477, 2007.
Ibrahim, A.S.; Sadek, A.K.; Weifeng Su; Liu, K.J.R; “[Cooperative communications with relay-selection: when to cooperate and whom to cooperate with?]{}” *IEEE Transactions Wireless Communications*, on Volume 7, Issue 7, July 2008 Page(s):2814 - 2827
R. Shah, S. Wietholter, A. Wolisz, and J. Rabaey, “[Mod- eling and analysis of opportunistic routing in low traffic scenarios]{}”,1em plus 0.5em minus 0.4em*IEEE WiOpt* 2005.
R. Shah, S. Wietholter, A. Wolisz, and J. Rabaey, “[When does opportunistic routing make sense?]{}”,1em plus 0.5em minus 0.4em*IEEE Third Int. Conf. on Pervasive Computing and Com- munications Workshops (PerCom)*, 2005.
Joseph Mitola, “[Cognitive Radio]{}”,1em plus 0.5em minus 0.4em*Licentiate proposal*, KTH, Stockholm, Sweden, December 1998.
S. Haykin, “[Cognitive radio: brain-empowered wireless communications]{}”,1em plus 0.5em minus 0.4em*IEEE Journal on Selected Areas in Communications*, Vol.23, NUMBER2, 2005 pp. 201-220.
Jafar, S. A., Srinivasa, S., “[Capacity limits of cognitive radio with distributed and dynamic spectral activity]{}”,1em plus 0.5em minus 0.4em*IEEE Journal on Selected Areas in Communications*, Volume 25, Issue 3, April 2007 pp. 529-537.
Xiaofei Zhou, Honggang Zhang, Chlamtac, I., “[Space-Frequency Coded Cooperative Scheme Among Distributed Nodes In Cognitive UWB Radio]{}” 1em plus 0.5em minus 0.4em*IEEE 16th International Symposium on Publication* Sept. 2005, Volume 1, Page(s): 461-465.
Lai, K.K.; Cheng, R.S.; Wong, A.K.; “[Channel Reuse for Accumulative Repetition Message-Forwarding on Wireless Ad-Hoc Networks]{}” 1em plus 0.5em minus 0.4em*IEEE Wireless Communications and Networking Conference, 2008*. March 31 2008-April 3 2008 Page(s): 2301 - 2306.
Shiang, H.; van der Schaar, M.; “[Distributed Resource Management in Multi-Hop Cognitive Radio Networks for Delay Sensitive Transmission]{}”1em plus 0.5em minus 0.4em*IEEE Transactions on Vehicular Technology*, Accepted for future publication.
Palomar, D.P.; Mung Chiang; “[A tutorial on decomposition methods for network utility maximization]{}” *IEEE JSAC*, Volume 24, Issue 8, Aug. 2006 Page(s):1439 - 1451
Dimitri P. Bertsekas, “[Dynamic Programming and optimal control]{}”,1em plus 0.5em minus 0.4emISBN: 1-886529-08-6.
Rangaralan K. Sundaram, “A First Course in Optimization Theory”, Cambridge University press, ISBN 0-521-49719-1.
URL: http://en.wikipedia.org/wiki/Path\_loss.
[^1]: Causality here refers to whether the source knows about the future channel states along the entire multihop transmission event from the source to the destination. In existing works, one way to justify the “non-causal knowledge” is to assume the channel state remains quasi-static across the sum of frame durations in the multihop transmission from the source to the destination.
[^2]: Stochastic NUM refers to a Network Utility Maximization problem where the objective function involves expectation w.r.t. the stochastic system state and the optimization variables involve not just actions at a given system state realization but rather a collection of actions for all system state realizations. This is a challenging problem because of the huge dimension of variables involved as well as the lack of explicit closed form expression for the objective function in terms of the control policy.
[^3]: In our paper, we allow the CSI to be time varying across different hops in the multi-hop transmission and the control policy is adaptive to the current information (but not the future CSI knowledge) only.
[^4]: $D_0$ is determined by the mean interference constraint to PU. For instance, denote $P_{int}$ as the interference constraint from SU to PU, $P_0$ is the mean transmitting power of SU, then $D_{0}\ge\left(\frac{P_0}{P_{int}}\right)^{\frac{1}{\alpha}}$, where $\alpha$ is the path loss factor.
[^5]: The frame index $k$ is equal to the number of hops already experienced by the packet currently transmitting in a continuous segment and will be reset to 1 when this packet is successfully delivered to the end node. Hence, $k$ might be different from segment to segment.
[^6]: The conventional flow balance constraint ensures that the output flow does not exceed the input flow at any SU node.
[^7]: A convex (concave) function subtracting another convex (concave) function is neither convex nor concave in general.
[^8]: For any given $\lambda_{ij}$, $\Omega^{LB}_{ij}$ is determined by . Substitute both policy to the , the LHS become a function of $\lambda_{ij}$
[^9]: Otherwise, $T(\eta_{ij}(k),\Omega_{ij})=\frac{1}{\log(1+
G_{s_{ij}(k)l_{ij}(k)}(k)
P_{ij}(k))}\sim\mathcal{O}\frac{1}{\epsilon
\overline{P}_{ij}(k)}\rightarrow \infty$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'High altitude platform station (HAPS) systems have recently attracted renewed attention. While terrestrial and satellite technologies are well-established for providing connectivity services, they face certain shortcomings and challenges, which could be addressed by complementing them with HAPS systems. In this paper, we envision a HAPS as a super macro base station, which we refer to as HAPS-SMBS, to provide connectivity in a plethora of applications. Unlike a conventional HAPS, which targets remote areas or disaster recovery, we envision next-generation HAPS-SMBS to have the necessary capabilities to address the high capacity, broad coverage, low latency, and computing requirements especially for highly populated metropolitan areas. This article focuses mainly on the potential opportunities, target use cases, and challenges that we expect to be associated with the design and implementation of the HAPS-SMBS based future wireless access architecture.'
author:
- 'Md Sahabul Alam, Gunes Karabulut Kurt, Halim Yanikomeroglu, Peiying Zhu, and Ngc Dũng Đào'
bibliography:
- 'maker\_IEEE.bib'
title: 'High Altitude Platform Station based Super Macro Base Station (HAPS-SMBS) Constellations'
---
Introduction {#S:Intro .unnumbered}
============
Parameters UAV HAPS VLEO LEO
------------------------------ ----------------------- ------------------------------------------- ---------------- -----------------
Operational altitude $100-400$ m $20-50$ km $250-500$ km $400-2000$ km
Cost Low Medium Medium High
Round-trip propagation delay $0.66-2.66$ $\micro$s $0.13-0.33$ ms $1.66-3.33$ ms $2.66-13.33$ ms
Communication endurance Short Long Long Long
Resource limitation High Low (empowered by solar battery charging) High High
Mobility Varying speeds Quasi-stationary Fast Fast
Coverage area Small Wider Wider Wider
Path loss Low Low Medium High
It is widely acknowledged that flexible and agile solutions for wireless connectivity will play a key role in future wireless communication systems. Currently, the connectivity requirements in terrestrial networks are addressed mainly by the densification of network infrastructures [@bhushan2014network]. However, densification solutions do not appear to be sufficient to address the ever-increasing user demands which are getting more and more unpredictable in space and time. In other words, no matter how dense most parts of the network is (with small base stations (BSs) in addition to macro BSs), a demanding application (such as immersed reality) can temporary arise at a locality in which the network infrastructure may be relatively sparse. In light of this, the seamless integration of terrestrial and aerial networks, known as vertical heterogeneous networks (VHetNets), has emerged as a promising architecture [@alzenad2019coverage]. In the current state-of-the-art, the emergence of low earth orbit (LEO) satellite constellations have been identified as a promising solution for enhancing network coverage [@zhou2019coverage]. However, we note that LEO constellations have two major shortcomings: a) direct LEO to user equipment (UE) connection is difficult with the current technology due to high path loss, and b) frequent handoffs will be encountered due to the high mobility of LEO nodes. Hence, a potential complementary solution to the wireless capacity and coverage enhancement lies in aerial platforms. The utilization of aerial platforms for 5G wireless communication systems have already been considered in 3GPP Release 17 [@3gpp17].
The envisioned aerial network is composed of two interacting sub-layers which offer agile network functionalities. The first sub-layer includes moving unmanned aerial vehicle (UAV) nodes, whereas the second sub-layer is composed of high altitude platform station (HAPS) systems, which are the quasi-stationary network elements. The International Telecommunication Union (ITU) has defined a HAPS in Article 1.66A as “A station on an object at an altitude of 20 to 50 km and at a specified, nominal, fixed point relative to the Earth". Most current deployment plans target an altitude range of 18 to 21 km. We believe that this HAPS sub-layer will provide important functions both in terms of capacity and coverage improvements by enabling the best features of both terrestrial and satellite communications. Motivated by these advantages, this paper discusses the significant role that HAPS systems can play in the future wireless access networks.
![Promises and target use cases of HAPS-SMBS networks.[]{data-label="haps_promises"}](HAPS_promises_5.pdf "fig:"){width="\columnwidth"}\
HAPS research dates back to the late 1990s [@tozer2001high]. Despite many advantages that HAPS deployment promised, their implementation was very limited at that time. In recent years, there has been a substantial increase in research efforts and commercial application plans for different HAPS technologies [@arum2020energy]. These developments have made HAPS systems more viable network element thanks also to the evolution of communication networks and advances in solar panel efficiency, battery energy density, lightweight composite materials, autonomous avionics, and antennas. With these advancements, in practice, the potential applications of a HAPS can be substantially broader than the conventional scenarios targeting remote and disaster applications.
In this article, we envision HAPS as a super macro BS, which we refer to as HAPS-SMBS, to cover a large metropolitan area in line with the smart city paradigm. The urgency of increasing traffic volume in complex urban scenarios as well as the problems of deploying terrestrial BSs and LEOs motivate us to consider the deployment of HAPS-SMBS. For example, in order to provide coverage for such a metropolitan area, a large number of ground BSs, as well as a backhaul network, may be needed. This high cost of infrastructure would be a major concern compared to a HAPS-SMBS. By contrast, a HAPS-SMBS is an excellent interface to mask both high path loss and the high mobility effects of LEO constellations. To solve the first problem, the UE can connect to a HAPS-SMBS with radio and a HAPS-SMBS to an LEO with free space optics (FSO). Since HAPS are almost geostationary, there are no mobility management related problems. The envisioned HAPS-SMBS can provide wireless services and assist the terrestrial network with the provision of distinct features, such as data acquisition, computing, caching, and processing. Fig. \[haps\_promises\] summarizes the promises and novel target use cases of a HAPS-SMBS as a main component of wireless access architecture. This is detailed in the subsequent sections.
Aerial Networks {#S:architecture .unnumbered}
=================
Nowadays, aerial networks have received growing interest for their potential to improve network design both in terms of capacity and coverage. Aerial networks consist of two network components: HAPS and UAV nodes.
High Altitude Platform Station (HAPS) Systems {#high-altitude-platform-station-haps-systems .unnumbered}
---------------------------------------------
HAPS was a popular research topic in the late 1990s and early 2000s with many distinct areas of investigation [@tozer2001high]. However, all these earlier visions as well as the current Google Loon project aim to bring remote parts of the globe online and for disaster applications. In contrast, we envision a HAPS-SMBS as another type of BS in a multi-tier VHetNet architecture to be deployed particularly in dense urban areas. Table \[Table:summary\] summarizes the features of a HAPS compared to a UAV and a very low earth orbit (VLEO) & LEO satellite.
Unmanned Aerial Vehicles (UAVs) {#unmanned-aerial-vehicles-uavs .unnumbered}
-------------------------------
Unlike HAPS with their quasi-stationary positions, UAVs are networking elements with relatively high mobility. The UAVs can be of two functions:
**UAV Base Stations:** The coverage and capacity improvements offered through the use of UAV mounted aerial BSs is a well-studied topic in the literature [@alzenad20173; @IREM2]. This concept is being actively investigated by 3GPP [@3gpp17]. The 5G system should be able to support UxNB (the 3GPP term for a UAV-BS) to provide enhanced and more flexible radio coverage. **UAV as User Equipment:** The use of UAVs as UE, such as drones is already supported through existing terrestrial networks. In particular, the use of UAVs as UE is currently being promoted by mega-retailers who would like to use the drones to carry courier packages.
To this end, an overview of how to make use of UAVs in wireless networks is provided. In the following sections, we investigate the latent opportunities and challenges of HAPS in future wireless access networks.
HAPS Advantages {#haps_advantages .unnumbered}
===============
The promise of HAPS as a main component of wireless network architecture can be listed as follows.
**Favorable channel conditions:** HAPS are expected to be only 20 km away with line-of-sight (LoS) links. This distance, combined with the high probability of LoS channels provides a relatively low channel attenuation. Hence, a direct link with ground UE is possible. At a low altitude, when compared to LEO satellites located 400 km to 2000 km away, this provides a much more favorable link budget. In terms of downlink, the corresponding favorable channel conditions provide a high signal-to-noise ratio (SNR) for the downlink and a coverage advantage, including for highly populated areas. **(Almost) Geostationary positions:** The position of a HAPS is relatively stationary. This means capacity is not wasted by orbiting over unpopulated areas (e.g., oceans), at all times connectivity from the same location can be enabled. The stationary status of a HAPS avoids the introduction of a significant Doppler shift. Furthermore, no tracking of the devices are needed. The stationarity of a HAPS also provides a basis for a main mobility management node, which can contribute towards the handoff management.
**Smaller footprint compared to satellite nodes:** The smaller footprint due to lower altitudes, when compared to satellite nodes, provides a higher area throughput, and improved resource utilization capability.
**Large platform:** A HAPS can be larger than a big building, and according to the recommendations of the ITU standard, its position should be maintained in a cylinder with a radius of $400$ m and height of $\pm 700$ m [@lTURF1500]. Hence it is suitable for multiple input multiple output (MIMO) and massive-MIMO (M-MIMO). Compounded by millimeter-wave (mmWave), very narrow beams are possible that improve the signal-to-interference-and-noise ratio (SINR) for all users. For example, using a HAPS to cover even a temporary hot spot at the ground is possible. In addition, due to the large size of a HAPS, it can be equipped with wide solar panels and energy storage systems to sustain it with the energy it requires.
**Even lower latency:** The relatively low altitude of a HAPS also provide a 40 km to 100 km round-trip distance, which corresponds to a round-trip delay of 0.13 ms to 0.33 ms. Hence, HAPS based connectivity does not suffer from the high-latency problems of satellite networks, which makes a HAPS suitable for low-latency applications.
**Hybrid connectivity:** ITU has already dedicated 600 MHz of spectrum for HAPS [@cianca2005integrated]. In addition to the dedicated band and terrestrial cellular bands, FSO is a promising alternative for providing multi-connectivity to robust and/or high data rate communication systems. One leading solution is to generalize the multi-band radio frequency (RF) links with hybrid RF-FSO connections. This approach will trigger a change of the classical radio access architecture to a more generalized *wireless access architecture* paradigm.
HAPS Super Macro Base Stations {#haps-super-macro-base-stations .unnumbered}
==============================
{width="\textwidth"}
A macro BS is a fundamental element in any HetNet wireless infrastructure for providing coverage and support capacity. Due to the inherent characteristics of quasi-stationarity, the larger footprint compared to UAVs, and the LoS channels, the envisioned HAPS mounted super macro BS can serve as a powerful platform to enhance coverage and capacity, as shown in Fig. \[HAPS\_macro\_BS\]. HAPS-SMBS improve the flexibility of the network design. The presence of HAPS-SMBS reduces the need for communication network over-engineering, which is done to match the requirements of peak demands. Therefore, the terrestrial network can be designed to satisfy the average user demands, and the rapidly changing (and often unpredictable) high demands can be simply addressed through a HAPS-SMBS. It should be emphasized that we refer to this BS as a *super macro BS* because of its large coverage area with M-MIMO and the provision of supporting distinct features, such as data acquisition, computing, caching, and processing.
However, with increasing the interest in HAPS-SMBS, it is imperative to access the feasibility of its deployment mainly considering the energy consumption constraint. In this vein, there has been a successful deployment of aircraft-based solar-powered HAPS [@arum2020review]. Energy management of HAPS-SMBS requires the investigation of how much energy will be consumed and how much solar energy can be harvested. For example, the authors in [@arum2020energy] estimated a HAPS BS power consumption for a service area radius of 60 km. It was shown that the total energy required for 24 hours of continuous HAPS mounted macro BS operation at full capacity was approximately 70 kWh. By contrast, the available solar energy provisioning using a 35 m wingspan HAPS platform is approximately 80 kWh. Hence, a solar power based HAPS-SMBS with extended coverage is potentially feasible in the near future. Additionally, significant developments in aeronautics, wingspan, and other areas of design will certainly create a HAPS-SMBS platform with higher energy efficiency.
Future HAPS-SMBS wireless network architecture can support data acquisition, computing, caching, and processing in a plethora of application domains. Some of them are shown in Fig. \[HAPS\_macro\_BS\], as detailed below.
**HAPS-SMBS for IoT applications:** It is expected that in future 5G/B5G networks, HAPS-SMBS will play a key role in different applications including the internet of things (IoT). In the past, there have been several research projects on HAPS; however, they are limited to civil applications, such as disaster monitoring or earth observation. IoT networks are characterized by an enormous number of devices each with low-rate links which are ideal for a single base station with a wide footprint. Due to its larger coverage, HAPS-SMBS in future 5G/B5G can support improved coverage for the realization of diverse outdoor IoT applications in a seamless, efficient, and cost-effective manner.
**HAPS-SMBS for backhauling outdoor small cell BSs:** Although the concept of a small cell base station has been widely acknowledged and studied for extremely high data rate coverage in 4G LTE wireless framework and is still perceived as a 5G key enabler, this concept cannot be realized in a straightforward manner mainly due to the difficulty and cost of backhauling a high number of small cell base stations. Motivated by the recent advances in HAPS and FSO research, backhauling outdoor small cell BSs can be realized through FSO and HAPS-SMBS [@alzenad2018fso], i.e., by placing the outdoor small cell BSs wherever appropriate without much concern about backhaul, and then focusing the laser on the HAPS-SMBS for the backhaul connectivity.
![Data rate vs transmit power of a vertical FSO link for different weather conditions. It is assumed that a HAPS-SMBS is placed at a distance of 18 km.[]{data-label="haps_performance"}](HAPS_FSO_rate.pdf "fig:"){width="\columnwidth"}\
For illustration, Fig. \[haps\_performance\] shows the achievable data rate of an FSO link where it is assumed that the terrestrial small cell BSs are connected to a HAPS-SMBS through this link. The achievable data rate of a given FSO link is calculated according to [@alzenad2018fso Eqn. (3)] where the parameters listed in [@alzenad2018fso Table 2] are used to obtain the numerical results. To take into account the impact of different weather conditions on the performance, we adopt the approaches developed in [@alzenad2018fso] and [@muhammad2005channel] for fog, rain, and cloud attenuation. From Fig. \[haps\_performance\], it can be observed that the data rates in the range of multi Gb/s can be achieved in clear weather conditions. It can also be observed that the achievable data rate is mostly affected by the rain. So the system may use FSO when there are clear skies or foggy conditions, and it can switch to RF during rainy conditions. **HAPS-SMBS to cover temporary unpredictable events:** The proposed HAPS-SMBS based wireless network architecture can provide additional coverage in case of temporary events which are hard to predict. For example, a HAPS-SMBS can provide additional beams to support the instantaneous capacity requirements in densely populated areas. Such flash events normally happen in cities, particularly when there are large gatherings, which leads to network congestion.
**HAPS-SMBS to support agile computational off-loading:** The main idea of computational off-loading is to do the computations at the network edge near the end user in order to reduce response time and enable real-time applications. In the future, as many applications (e.g., augmented reality) will require high computational capabilities, it is expected that enabling efficient computational offloading will be a necessity. HAPS-SMBS will play a significant role in providing computational services as part of the integrated network. HAPS-SMBS have more computational power than UE (e.g., aerial UE) and can provide better coverage with LoS links due to their high position which avoids the possibility of disconnection while offloading data. **HAPS-SMBS as a flying data center:** HAPS-SMBS will also enable the possibility of flying data centers. These data centers can provide a back-up computational facility, that can also be functional in case of emergency scenarios where the ground infrastructure fails to function.
**HAPS-SMBS for coverage holes:** HAPS-SMBS can assist existing terrestrial networks by providing coverage holes through a cost-effective manner. This problem happens when the terrestrial UE received signal strength in an area falls below a predetermined level that is required for robust radio performance due to physical obstructions. For this, a HAPS-SMBS needs to steer a beam in a specific direction. Through their physical advantages, HAPS-SMBS can perform 3D beamforming that enables the creation of separate beams in the three-dimensional space at the same time for different users as shown in Fig. \[HAPS\_macro\_BS\]. At present, it is widely seen that existing terrestrial networks fail to overcome the coverage holes even in a modern metropolitan area without having a viable access point to which to connect.
**HAPS-SMBS for ubiquitous coverage for intelligent transportation systems:** HAPS-SMBS can play a key role for the ubiquitous coverage of intelligent transportation systems (ITS)/connected and autonomous vehicle (CAV) paradigms shown in Fig. \[haps\_ITS\]. Recent advances in sensors and the introduction of in-car wireless communication capabilities have paved the way for CAVs that enables unprecedented scenarios for road transportation. Nevertheless, huge data fusion and processing are necessary for many ITS applications. As vehicles are limited in processing capabilities, they may offload the data to cloud or fog computing nodes for delay-tolerant applications require large computation power. However, due to the high mobility of vehicles, data offloading will be interrupted by frequent handovers. In addition, the data processing outcome needs to be delivered through the BS, which is accessible by the vehicle. Fortunately, a HAPS-SMBS can provide both the large area coverage and computational capabilities with low communication delays. Thus, a HAPS-SMBS can eliminate the effect of frequent handovers in vehicular networks. In addition, HAPS-SMBS can be used as a sidelink for vehicle-to-vehicle communications where the terrestrial BS fails to provide enough coverage. Moreover, a HAPS-SMBS can have a wide view of a vehicular network which is essential for coordinating vehicle-to-everything communications, especially for areas with limited infrastructure.
![HAPS-SMBS constellation to support ITS applications.[]{data-label="haps_ITS"}](Figure4_revised.pdf "fig:"){width="\columnwidth"}\
**HAPS-SMBS to cover a massive amount of aerial UE:** Using HAPS-SMBS for coverage can also provide an essential tool for cargo drones that will likely disrupt the retail industry in the near future. As the retail industry is evolving with the possibility of using cargo drones, 3D highways can be expected to serve the cargo package distributions of platooning autonomous drones. Considering 1 delivery/home/day for 1 million homes will require 12 drone launches per second which is the equivalent of approximately 10,000 drones in the air at any given time. To enable reliable connectivity for such a high amount of aerial UEs, a single HAPS-SMBS can be used to cover a city as shown in Fig. \[haps\_UEs\].
![A HAPS-SMBS for providing coverage to a massive number of aerial UE’s.[]{data-label="haps_UEs"}](Figure5_revised.pdf "fig:"){width="\columnwidth"}\
**HAPS-SMBS as an intelligent aerial network enabler:** HAPS-SMBS can be equipped with powerful processors that can provide computational support for limited-resources aerial network elements. Besides, due to their quasi-stationary positions and large coverage areas, HAPS-SMBS can collect data from large portions of the aerial network and use such data as a real-time input for machine learning (ML) algorithms. In this way, a HAPS-SMBS can dynamically learn about the network status, resources and topology, and with a minimum dependence on terrestrial-based control, a HAPS-SMBS can control and manage the aerial network intelligently.
**HAPS-SMBS as an interface to provide seamless communication to LEO satellites:** LEO satellites move at a very high speed resulting in frequent disconnections and handovers at the terrestrial gateways. Compared to terrestrial gateways, a HAPS-SMBS has a wide upper footprint that can cover many LEO satellites simultaneously. Therefore, it is envisioned that a HAPS-SMBS can act as an interface of the satellite network and provide seamless satellite communication to the aerial and terrestrial networks. In this scenario, HAPS-SMBS will handle the frequent handover of LEO satellites, and if user devices can communicate with a HAPS-SMBS directly then users do not have to use special devices or ground stations to communicate with satellites. In this regard, supervised ML can be utilized by the HAPS-SMBS to learn the mobility patterns of the satellites in order to predict their handover then establish a connection to an approaching satellite before losing the current connection.
Open Challenges and Future Research {#S:Challenges .unnumbered}
===================================
Distributed RRM {#distributed-rrm .unnumbered}
---------------
The HAPS-SMBS based wireless access architecture we envision has many distinct and even critical characteristics compared to terrestrial networks. The distinct characteristics may make it inefficient to apply the standards, protocols, and design methodologies that are optimized for radio resource management (RRM) in terrestrial wireless networks in the design of HAPS-SMBS aided VHetNets directly. The RRM algorithms can be operated either in a centralized or distributed fashion. In conventional terrestrial networks, the RRM problems are typically addressed through a centralized approach. However, this may not be a feasible choice for HAPS-SMBS based VHetNets due to issues related to network heterogeneity, computational complexity, cost, spectrum overhead for channel state information (CSI) transmission, and scalability. Although each technology has distinct advantages and drawbacks, distributed RRM technology may help provide HAPS-SMBS aided VHetNets with improved agility and resilience.
In addition, a distributed radio access network (RAN) should be coupled with distributed RRM technology to maximize its full potential. In fact, the concepts of advanced RAN and advanced RRM are inseparable. Besides, distributed RRM should have enough cognition (cognitive radio) to decide when to transmit and which subcarriers to transmit with. Furthermore, the potential of ML should be explored in developing distributed RRM algorithms.
Capacity Improvement {#capacity-improvement .unnumbered}
--------------------
There are several ways to improve the capacity of communication networks. Some of them are as follows:
**Spectral efficiency improvement:** MIMO is one of the most promising techniques for improving spectrum efficiency in HAPS networks. However, many challenges have to be addressed for implementing MIMO in HAPS-SMBS. Despite the challenges, there are a considerable number of studies that have investigated the use of MIMO techniques in HAPS communications [@michailidis2010three]. In addition, research needs to be undertaken in aerial distributed massive MIMO, where antennas are coordinated from geographically distributed HAPS for improving spectrum efficiency. In particular, 3D MIMO, also known as full dimension MIMO, can yield higher overall system throughput.
Beamforming is also believed to have an important role in addressing the capacity demands of aerial networks at a reduced power level. Beamforming at HAPS-SMBS is more challenging than beamforming at ground BSs, where both the location and target coverage are generally fixed. Some possible directions for research on beamforming at HAPS-SMBS are the following:
- 3D beamforming at HAPS-SMBS for coverage holes and unpredictable hot spots on the ground.
- 3D beamforming at HAPS-SMBS for aerial-UEs (such as cargo drones).
Accurate beam-steering/alignment can be a challenge in moving networks; nevertheless, the quasi-stationarity of a HAPS-SMBS will help in this regard. **NOMA:** Non-orthogonal multiple access (NOMA) has recently been introduced as an effective approach that can potentially provide spectral efficiency, presenting a promising candidate solution for future radio systems. NOMA can also be exploited to improve spectral efficiency at HAPS-SMBS to cover a massive number of aerial UE; however, the successful operation of NOMA in HAPS-SMBS requires numerous associated challenges to be addressed, including the power coefficient determination in regard to the channel uncertainty of HAPS-SMBS to UE.
**Extension to mmWave bands:** Extending the spectrum to extremely high frequencies, such as mmWave bands, can be regarded as the most efficient proposal for improving transfer rates in HAPS-SMBS. In addition to the bands already dedicated for HAPS usage, for example, 47.2 – 47.5 GHz and 47.9 – 48.2 GHz, ITU during the World Radiocommunication Conference 2019 (WRC-19) congress discussed that the frequency bands 21.4 - 22 GHz and 24.25 - 27.5 GHz can be used by HAPS [@WRC-19]. The application of mmWave techniques may offer many advantages for HAPS-SMBS, such as higher bandwidth, higher Tx/Rx antenna gain, beamforming and spatial multiplexing gain, placement of a large number of antennas in small dimensions, etc. However, many challenges have to be addressed for HAPS-SMBS mmWave communication networks, including the large coverage with mmWave.
Network Management/Control {#network-managementcontrol .unnumbered}
--------------------------
The need for the joint communication, control, computing, and caching in a HAPS-SMBS to meet the intrinsic requirements of envisioned applications raises unprecedented challenges in the network management.
There have been gradual developments to make communication networks more autonomous, self-organizing, self-configuring, and self-sustaining. To support these developments, potential solutions have been introduced in the literature; network slicing (NS), software-defined network (SDN), network function virtualization (NFV), are among them. The VHetNets architecture is highly dynamic and heterogeneous. The exploitation of NS, SDN, and NFV in the presence of a HAPS-SMBS should be explored to facilitate network reconfiguration and improve network agility and resilience. For example, the HAPS NS should consider dynamic spectrum slicing to avoid underutilization or overutilization. Furthermore, the application of machine learning techniques to derive an in-network solution in HAPS-SMBS systems is a promising research topic. The ability of having in-network solution will eliminate the need for direct human intervention on many operation levels and allow HAPS systems to make intelligent decisions in a collaborative manner.
Interference Management/Control {#interference-managementcontrol .unnumbered}
-------------------------------
HAPS were previously deployed in isolation, so there was no or few problems of interference. In metropolitan areas, one of the key challenges of deploying HAPS is interference management. In this case, owing to the simultaneous data transmission from HAPS-SMBS with other segments of the integrated network, more interference will be generated which may result in a higher link outage probability. To access the impact of the aggregated interference produced by HAPS-SMBS on the integrated VHetNet architecture, proper interference analysis and management of interference is required. Furthermore, intelligent interference management is necessary, which can be achieved through implementing ML algorithms that can learn and adapt to changes in network environments. Thus, interference management through beamforming and frequency reuse can be done in an intelligent way.
HAPS Constellations and Inter-HAPS Networking: {#haps-constellations-and-inter-haps-networking .unnumbered}
----------------------------------------------
We also envision deploying a network (constellation) of interconnected HAPS-SMBSs whenever necessary. For instance, the 7,800 km long trans-Canada highway from the Pacific Ocean to the Atlantic can be served by a linear HAPS constellation towards a coast-to-coast intelligent transportation system. Coordinating a HAPS network through ground stations would not be a feasible choice due to response delays, and a ground station with its limited footprint cannot have communication coverage to all the HAPS network. Therefore, we envision that HAPS networks will be self-organized with either centralised or distributed control and management system. In the centralized approach, a HAPS is elected to be the manager while the others are followers. In the distributed approach, the available HAPS in a network need to negotiate and coordinate in distributing the communication tasks in order to avoid interference, wasting resources, or overlapping footprints. In this regard, a comprehensive study of the HAPS constellation design methodology and inter-HAPS networking with proper interference management becomes critical.
Conclusion {#S:con .unnumbered}
==========
In this article, we shed light on the potential opportunities and target use cases of HAPS-SMBS aided wireless access architecture. We pointed out that, while research on HAPS goes back to the late 1990s, the concept has attracted new attention in recent years, both in academia and industry, as a promising solution in future wireless networks. This momentum is fueled by the ever-increasing demand from wireless networks and also by advances in solar panel efficiency, battery energy density, lightweight composite materials, autonomous avionics, and antennas. We illustrated that the proposed VHetNets architecture empowered by HAPS-SMBS nodes will enable the network to increase the overall throughput, improve coverage, and also to provide a platform to perform the near-user computation to significantly reduce the end-to-end delays. Furthermore, HAPS-SMBS nodes will also enable the network to address unpredictable congestion instances as well as coverage holes in populated areas.
[MD SAHABUL ALAM]{} ([email protected]) received the Ph.D. degree in electrical engineering from ETS, Montreal, QC, Canada. Currently, he is working as a Postdoctoral fellow in systems and communications engineering department of Carleton University with prestigious FRQNT PDF fellowship. In Ph.D., Dr. Alam awarded the Governor General of Canada Gold Medal. His research interests include non-terrestrial communications and smart grid communications.
[GUNES KARABULUT KURT]{} ([email protected]) received the Ph.D. degree in electrical engineering from the University of Ottawa, Ottawa, ON, Canada, in 2006. Since 2010, she has been with ITU. She is also an Adjunct Research Professor at Carleton University. She is serving as an Associate Technical Editor of IEEE Communications Magazine.
[HALIM YANIKOMEROGLU]{} \[F\] ([email protected]) is a full professor in the Department of Systems and Computer Engineering at Carleton University, Ottawa, Canada. His research interests cover many aspects of 5G/5G+ wireless networks. His collaborative research with industry has resulted in 38 granted patents. He is a Fellow of the Engineering Institute of Canada and he is a Distinguished Speaker for IEEE Communications Society and IEEE Vehicular Technology Society.
[PEIYING ZHU]{} \[F\] ([email protected]) is a Huawei Fellow. She is currently leading 5G wireless system research in Huawei. The focus of her research is advanced wireless access technologies with more than 150 granted patents. In recent years she organized and chaired various 5G workshops. She is a frequent keynote speaker in major IEEE conferences.
[NGC DŨNG ĐÁO]{} ([email protected]) is a principle engineer of Huawei Technologies Canada. His research interest covers several aspects of 5G and beyond 5G mobile networks, including architecture design, data analytics, and vertical applications. He is a technical editor of IEEE Communications Magazine, associate editor of IEEE Communications Surveys & Tutorials, and editor of IEEE Transactions on Vehicular Technologies.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Creating and labelling datasets of videos for use in training Human Activity Recognition models is an arduous task. In this paper, we approach this by using 3D rendering tools to generate a synthetic dataset of videos, and show that a classifier trained on these videos can generalise to real videos. We use five different augmentation techniques to generate the videos, leading to a wide variety of accurately labelled unique videos. We fine tune a pre-trained I3D model on our videos, and find that the model is able to achieve a high accuracy of $73\%$ on the HMDB51 dataset over three classes. We also find that augmenting the HMDB training set with our dataset provides a $2\%$ improvement in the performance of the classifier. Finally, we discuss possible extensions to the dataset, including virtual try on and modeling motion of the people.'
author:
- |
Ollie Matthews\
Princeton University\
[[email protected]]{}
- |
Koki Ryu\
Princeton University\
[[email protected]]{}
- |
Tarun Srivastava\
Princeton University\
[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'Creating a Large-scale Synthetic Dataset for Human Activity Recognition'
---
Introduction
============
A significant challenge in Human Activity Recognition (HAR) is obtaining large amounts of video data on which deep networks can be trained. Labelling and storing video data is significantly more labour and memory intensive than images, and the largest version of the Kinetics dataset [@carreira2019short] which contains $650,000$ videos of crowd-sourced data over $700$ classes is still two orders of magnitude smaller than ImageNet [@deng2009imagenet].
####
In this paper we attempt to see if this problem can be approached with synthetic data. The recently published MoVi dataset [@ghorbani2020movi] provides joint position data for $90$ subjects doing $21$ different actions. Combining this data with human body rendering packages, we are able to generate videos of people doing the actions from any angle, and superimpose them onto any background.
####
The aim is to use this to create a large dataset with a diverse set of angles and camera motions, and to see if models trained on this dataset are able to generalise to real videos. For evaluation we use the I3D model [@carreira2017quo], which is pre-trained on Kinetics, and test it on the HMDB51 dataset [@6126543] on a set of classes which do not appear in Kinetics.
####
In particular, our contributions are:
- We *introduce* a potentially infinite source of data for Human Activity Recognition. We cover three common actions - walking, waving, and sitting down. These actions were chosen for testing due to their appearance in the HMDB51 dataset, but our framework could easily be extended to more actions provided in the MoVi dataset.
- We *validate* our approach of using synthetic data by fine-tuning the I3D model, pretrained on Kinetics, with a combination of our synthetic data and real videos from the MoVi. We show that our fine-tuned model outperforms the same I3D model fine-tuned with only real videos taken from the HMDB training set.
- Lastly, we *provide* the ability to extend our framework to any real indoor scene through the use of RGB-D depth map fusion. We reconstruct a variety of indoor scenes using a pipeline based on the work done by Choi et al.[@choi2015robust] and synthesise videos of human subjects performing the chosen actions. Rendering our subjects in such 3D environments was shown to further improve the performance of our framework.
Related Work
============
![Examples images taken from the SURREAL dataset.[]{data-label="fig:surrealsucks"}](./images/surrealsucks.png){width="\linewidth"}
####
The use of a synthetic dataset is actively being researched by the deep learning community. Peng et al [@peng14] crowdsource CAD models online to augment datsets for object detection algorithms. Gaidon et al. [@Gaidon16] have created a Virtual KITTI dataset that generates synthetic videos of cars to study object detection and tracking. There has not, however, been much research into the use of synthetic data for Human Activity Recognition. The only widely available dataset for HAR is SURREAL by Varol et al. [@varol17].
####
Figure \[fig:surrealsucks\] shows example images from their synthetic dataset, found in their paper. These images are far from realistic, and do not make geometrical sense. People are often in impossible positions, as seen in the first image, or out of proportion like in the second image. By generating a 3D model of our subjects within the background scene, we can ensure the results are not geometrically impossible, and the videos produced from these 3D renderings can include realistic effects like shadows and occlusion.
Datasets
========
We make use of a number of datasets for generation and evaluation of our videos. The videos of the generated people are based on MoVi, an opensource dataset of Inertial Measurement Unit (IMU), motion capture and video data for actors performing a range of activities. The $60$ female and $30$ male actors are of a range of ages, and perform $20$ prescribed actions as well as one individually chosen action.
####
We generate realistic 3D human body meshes based on the captured body poses from MoVi with: SMPL+H [@MANO:SIGGRAPHASIA:2017], which is a parameterized human body model with hand motions; and Dyna [@Dyna:SIGGRAPH:2015], which generates soft-tissue deformations on the body model. The 3D points of MoVi converted into the body parameters of SMPL+H and Dyna are stored in the AMASS[@AMASS:2019] dataset. We use these parameters to generate human action videos.
####
As a part of this project, we also reconstruct indoor 3D scenes to use as backgrounds for our synthetic dataset. The underlying idea being that any user can extend this synthetic dataset to their specific purpose by reconstructing their own environment and synthesising more data. As a proof of concept, in this project we reconstructed open-source dataset with readily available RGB-D streams. We reconstructed scenes from the Redwoord Indoor LiDAR dataset [@Park2017], the ICL-NUIM dataset[@handa:etal:ICRA2014], the BundleFusion dataset [@dai2017bundlefusion], and lastly the TUM RGB-D SLAM dataset [@sturm12iros].
####
We evaluate our results on the HMDB51 dataset, which contains $6,849$ clips of $51$ different actions mostly extracted from films.
Challenges in the HMDB Dataset
------------------------------
The HMDB dataset is notoriously challenging, with only $70$ videos for training. The videos also come from films, and the person doing the action is often not central in the video.
####
Figure \[fig:hmdbbad\] shows an example of a problem that can occur because of this. The image is a frame taken from one of the ’hand waving’ videos in HMDB. The frame has a wide aspect ratio, with the man waving only appearing at the right of the screen. The classifier we use requires inputs to have equal aspect ratios, so at training, we have to crop out sections of the frame to be fed into the classifier. For many of the crops, the label will be ’hand waving’ even though the man waving is not in shot. We deal with this at test time by convolving our classifier across the image and averaging the results, but these videos can harm the performance of the network if they come up during training.
![An frame from a challenging video in the HMDB dataset where the subject is not in the center.[]{data-label="fig:hmdbbad"}](./images/hmdbbad.png){width="\linewidth"}
####
The HMDB dataset also includes videos where the camera is moving during the shot, and others where the subject is occluded. These problems make the dataset hard to train on, and we aim to deal with them in generating our synthetic dataset. The problem of the subject not being in the shot, for example, does not occur when we create the videos ourselves and choose where the subject is placed.
Approach
========
Rendering
---------
We render realistic human action videos based on the body parameters from AMASS with Trimesh [@trimesh] and Pyrender [@pyrender]. We create Trimesh objects corresponding to the human body and the background and add them to a 3D scene model in Pyrender. Pyrender exports the scene as MP4 video.
####
We convert the body parameters to human body mesh objects with VPoser [@SMPL-X:2019]. We create the background in two different ways. In the first, we create a flat wall as a Trimesh object and add a background image as the texture. In the second, we convert the OBJ file of the living room in ICL-NUIM to a GLB file with a free converter [@anyconv]. The GLB format is readable by Trimesh, and can then be imported as a 3D object.
####
We use a Pyrender scene set up by VPoser to render videos. We modify the positions of cameras and lights in the scene and add the background and human model as objects to it. To make the rendered video realistic, we fine-tune the size and positions of the objects manually.
Data Augmentation {#sec:aug}
-----------------
We render videos based on the actions of 15 subjects in MoVi dataset. For each subject, we choose three actions, ’walking’, ’sitting down’ and ’hand waving’, and generate 10 videos per action. We try different methods to add variation to the multiple videos made from a single pose.
### Background Images + Rotation (BG+R) {#bg_rot}
For this and the next method, we use six images from Google Images to provide the backgrounds. These show three inside and three outside scenes, and are used as texture for the back wall. The human objects are generated with a random rotation from -90 to 90 . We then add the human model in the center of the scene and the wall behind the human, and the scene is rendered as a video. Examples are shown in Figures \[fig:bgr1\] and \[fig:bgr2\].
![BG+R (2).[]{data-label="fig:bgr2"}](./images/bgr1.png){width="\linewidth"}
![BG+R (2).[]{data-label="fig:bgr2"}](./images/bgr2.png){width="\linewidth"}
### Background Images + Rotation, Resizing and Rotation (BG+R^2^T) {#bg_rot_scale}
In \[bg\_rot\], the human model is centered in the generated videos. Since the subject is not necessarily in the center of the shot in a real video, we apply scaling and translation to the human poses to generate another set of videos. For each action, we choose random variables $s\in[0.7,1.3]$, $x\in[-0.5,0.5]$, and $y\in[-0.1,0.1]$. If we denote by $h$ the height of the human body in the video, for each scene we change the coordinates of the body by $(h \cdot x, h \cdot y)$, then scale the body by $s$. Examples are shown in Figures \[fig:bgrrt1\] and \[fig:bgrrt2\].
![BG+R$^2$T (2).[]{data-label="fig:bgrrt2"}](./images/bgrrt1.png){width="\linewidth"}
![BG+R$^2$T (2).[]{data-label="fig:bgrrt2"}](./images/bgrrt2.png){width="\linewidth"}
### 3D Model + Rotation (3D+R)
In all of the following methods, we use a 3D object as the background of the generated videos. We create a human pose 3D mesh with the same method as the one used in \[bg\_rot\], then place this human body in a fixed point of the living room scene from the ICL-NUIM. To add some variation to the background, we pick a random color for the background from 13 colors: ’pink’, ’purple’, ’cyan’, ’red’, ’green’, ’yellow’, ’brown’, ’blue’, ’offwhite’, ’white’, ’orange’, ’grey’, and ’yellow’.
####
As described in \[bg\_rot\], we also pick a random angle from the range of -90 to 90 . Instead of rotating the human body by this angle, we rotate the camera and the light around the body by the angle so that each rendered video shows the different part of the living room.
####
Finally, we change the position of the light and the angle of view of the camera to make the human pose look clear in the closed object. Figures \[fig:3dr1\] and \[fig:3dr2\] show some frames from these.
![3D+R (2).[]{data-label="fig:3dr2"}](./images/3dr1.png){width="\linewidth"}
![3D+R (2).[]{data-label="fig:3dr2"}](./images/3dr2.png){width="\linewidth"}
### 3D Model + Camera Motion (3D+M) {#sec:3DM}
Having a full 3D model allows us to add camera motion to the video. We choose random variables $x_1$ and $x_2$ from the range of -0.5 to 0.5, and choose $y_1$ and $y_2$ from the range of -0.1 to 0.1. We move the camera and the light in the video so that the coordinates of the body is $(x_1 \cdot h, y_1 \cdot h)$ at the beginning of the video and $(x_2 \cdot h, y_2 \cdot h)$ at the end.
####
We also choose two angles, $\theta_1$ and $\theta_2$, from the range of -90and 90 and rotate the camera and the light from $\theta_1$ to $\theta_2$. Figure \[fig:3dm\] shows a sequence of frames from one of these videos, where the camera motion can be seen.
![A sequence of frames from a video in 3D+M.[]{data-label="fig:3dm"}](./images/3dm_seq.png){width="\linewidth"}
### Reconstructed 3D Model + Rotation (R3D+R)
In this final method, we use 3D reconstructions of scenes to lead to more realistic images, and give users the ability to cater the synthetic dataset to their specific purpose. Given that there are several commercially available structured light sensors available, users could reconstruct their own scenes for use in the synthetic dataset. A good example would be human activity recognition in sports - if a user wished to identify human actions in a school basketball court, they could reconstruct the scene and generate videos for that purpose.
####
We use the Python implementation of the Open3D [@open3d] library to reconstruct scenes from a sequence of RGB-D frames. In real life, users will not have precise camera pose data, so our 3D scene reconstructions were performed without ground truth pose data. We begin by aligning the RGBD frames by using the in-built ’compute RGB-D odometry’ function that estimates 6D rigid body motion between two frames. The input to this function is a rough estimate of the alignment provided by a 5 point RANSAC algorithm. With these estimates, we perform pose optimisation which results in fragments of the scene. A global posegraph is then computed after performing multiple registration on the scene fragments. Multiple registration is performed again to refine the posegraph, and lastly these fragments are integrated, resulting in a .ply file.
####
The human can then be placed into the .ply file in the same way as in Section \[sec:3DM\]. Figure \[fig:3drecons\] shows two of the generated scenes, and Figures \[fig:r3dr1\] and \[fig:r3dr2\] show frames from the final videos.
![Examples of reconstructed scenes used for the images in Figures \[fig:r3dr1\] and \[fig:r3dr2\].[]{data-label="fig:3drecons"}](./images/3dreconstructions.png){width="\linewidth"}
![R3D+R (2).[]{data-label="fig:r3dr2"}](./images/3d+real.png){width="\linewidth"}
![R3D+R (2).[]{data-label="fig:r3dr2"}](./images/3d+real_1.png){width="\linewidth"}
Optical Flow
------------
To calculate optical flow, we use the TV-L1 algorithm. This is also used in the I3D model’s original paper as it provides a dense and accurate output. We use the OpenCV implementation [@opencv_library], and Figure \[fig:optflow\] shows an example of an output for a 3D+M video.
![A visualized optical flow generated from a video in 3D+M.[]{data-label="fig:optflow"}](./images/optflow.png){width="\linewidth"}
Dataset Evaluation
==================
To evaluate our dataset, we fine-tune a pre-trained video classifier on our videos and test it on the HMDB dataset. The classifier used is the I3D classifier, pre-trained on the original version of Kinetics [@kay2017kinetics] which contains at least $400$ videos per class for $400$ classes.
Classes
-------
We evaluate over classes that fit two criteria: they appeared in both MoVi and HMDB, and did not appear in Kinetics. The second condition ensures that our pre-trained model had no advantage in predicting certain activities, and restricts us to three classes - ’walking’, ’sitting down’ and ’hand waving’.
Model
-----
The I3D model utilises a two-stream 3D Convolutional Neural Network (CNN) approach based on Inception-v1 [@ioffe2015batch]. Before being trained on Kinetics, the classifier is bootstrapped from 2D CNNs trained on ImageNet.
####
We use a publicly available version of the I3D model [@I3D] with the final layer removed to extract $1024$-dimensional features from RGB and optical flow videos. The model is fully convolutional, and provides around 4 feature vectors per second of video.
####
For both the RGB and optical flow features, we connect a fully connected layer to the features to a $3$-dimensional output with a softmax applied. A dropout rate of $0.5$ is applied to the features. The class probabilities from both networks are then summed to provide combined probabilities. For testing, we also average the probabilities over the temporal dimension of the video.
Data Pre-Processing {#sec:preprocess}
-------------------
Before being fed into the I3D model, the videos need to be converted to RGB and optical flow arrays. The I3D model takes videos at $25$ frames per second and $224\times224$ pixels. While our videos can be generated under these constraints, the HMDB videos need to be resampled to $25$ fps and reshaped. The HMDB videos do not have equal aspect ratios, so we resize them - keeping the aspect ratio constant - so that the vertical dimension is $224$ pixels, and then extract an equally spaced set of frames spanning the width of the videos.
####
The RGB and flow arrays are also centered and normalised. As suggested in the original I3D paper, the optical flows are truncated at $\pm 20$ prior to this.
Training
--------
To train the fully connected layer, we use an Adam optimiser [@kingma2014adam] with a learning rate of $10^{-4}$ for $300$ epochs. Each video has multiple feature vectors over the time dimension, and we train the fully-connected layer individually on each of these with a categorical cross-entropy loss function.
Results
=======
We use the first suggested split of the HMDB dataset for evaluation, which gives $70$ training videos and $30$ testing videos for each class. We then evaluate the test accuracies for the RGB network, the optical flow network, and the two networks combined.
Ablation Study 1: Comparison of Augmentation Methods {#sec:augmethods}
----------------------------------------------------
Table \[tab:res1\] show the accuracy achieved when training the classifier on the subset of videos from each augmentation technique listed in Section \[sec:aug\].
RGB Optical Flow Combined
---------- ---------- -------------- ----------
BG+R 0.58 0.60 0.67
BG+R^2^T 0.53 0.68 0.68
3D+R 0.58 **0.70** 0.66
3D+M **0.59** 0.56 0.67
R3D+R **0.59** **0.70** **0.69**
: Test accuracies for model trained on individual augmentation methods.[]{data-label="tab:res1"}
The simplest background with rotation method performs well in RGB and combined, but underperforms in optical flow, indicating that it is important to have different sizes of models for the optical flow network to learn to be size-invariant. Varying the size and location of the subjects helps to deal with this, but the RGB accuracy decreases.
####
Using the living room model as a 3D scene leads to better accuracy in the individual networks. This can be explained by the fact that in this augmentation method the humans are not necessarily centered, which helps enforce size and location invariance in the network, but the range of sizes is not as large as in BG+R^2^T which could explain the better accuracy in RGB and optical flow. Since the videos are rendered from a full 3D model, they are also more realistic geometrically. Adding camera motion improves RGB performance, but dramatically decreases the performance of the optical flow network. This could be due to the camera motion being too fast, which causes non linear behavious if the optical flow values truncation at the limits described in Section \[sec:preprocess\]. It could also be because it trains the classifier to always expect camera motion while in reality this is only the case in some videos.
####
The best performance overall is from using the reproduced 3D scenes. These scenes have realistic backgrounds, and the proportions and shadows are geometrically correct. As seen in the feet of the figure in Figure \[fig:3dr2\], these videos also can have occlusion, which is not present in any of the other videos and often occurs in HMDB51.
Full Dataset Evaluation
-----------------------
RGB Weight Flow Weight
---------- ------------ -------------
BG+R 1 0
BG+R^2^T 0 1
3D+R 1 1
3D+M 1 1
R3D+R 1 1
HMDB51 8 3
: The weights applied to each set of videos in COMBW (note HMDB51 is only included in COMBW+HMDB).[]{data-label="tab:weights"}
Table \[tab:res2\] shows the results from evaluating the model on the combined augmentation methods. In COMB, all of the augmentation methods are given equal weight. In COMBW, they are weighted differently, with weights shown in Table \[tab:weights\] applied to each set of videos. These weights were informed by the results of Section \[sec:augmethods\], with augmentation methods which lead to poor classifiers being left out. Note that the camera motion is included in training the optical flow network even though it leads to poor performance in our ablation study. It turns out there was a $1\%$ reduction in accuracy when it was omitted, which implies that while not all of the training videos should have camera motion, it is an important factor to include.
####
HMDB shows the results from training on HMDB alone, and COMBW+HMDB shows the results of training on HMDB combined with COMBW. We weight HMDB more heavily than the other sets, in particular on the RGB training, since it contains less samples and the classifier performance is so much higher when trained on it. However we do not pretend the weights chosen are optimal - they were found through trial and error, and could potentially be improved upon in future work.
RGB Optical Flow Combined
------------ ------ -------------- ----------
COMB 0.54 0.74 0.70
COMBW 0.59 0.75 0.73
HMDB 0.78 0.77 0.81
COMBW+HMDB 0.78 0.82 0.83
: Test accuracies for model trained on combined datasets.[]{data-label="tab:res2"}
Using the entire dataset leads to significant improvements in performance compared to only using certain parts. This shows that the different augmentation methods work well in conjunction, for example while using an entire dataset with camera motion is not effective, including this with the other augmentation techniques makes for a more robust classifier.
####
The RGB performance on COMB is poor, however, and it seems that the BG+R^2^T set has a significant detrimental effect on the performance of the RGB network as removing it in COMBW dramatically improves performance. The flow network also sees moderate improvements when the BG+R set is removed and more emphasis is put on the R3D+R scenes. The final accuracy of the COMBW classifier of $0.73$ is very impressive seeing as the classifier has never seen any of the videos in HMDB before.
####
Training on HMDB leads to significant improvements in particular in the RGB network, which is not entirely surprising. While we try to make the RGB videos as realistic as possible, they are still based on reconstructions and animations. The optical flow, on the other hand, can largely be reproduced in the synthetic videos regardless of how realistic the actual videos are.
####
Using a combination of our dataset and HMDB improves the optical flow network performance by $5\%$, and the combined performance of the classifier by $2\%$. This shows that our data is realistic enough, at least in terms of optical flow, to bolster smaller datasets and improve performance by increasing the training set size.
####
Figure \[fig:bar\] shows the combined performance of the I3D model trained on each of the datasets, for each class.
![Per-class accuracy of the model trained on each dataset.[]{data-label="fig:bar"}](./images/barchart.pdf)
####
The main source of confusion in the classifiers is between ’walking’ and ’hand waving’. Since these both mostly happen from an upright position, they look more similar than ’sitting down’, on which the accuracies are close to $1.0$. The generated videos seem to struggle in particular with walking, which we attribute to two problems. The first is that walking shots often have close ups on the upper body of the person, and our generated videos do not have this. The second is that all of our models are moving on a constant point in their scenes, since there was no readily available position information in the dataset. This means that images like that shown in Figure \[fig:badwalk\], where there is a close up and all the motion is in the background of the scene are difficult for the model trained on our dataset. In contrast, the model performs well on walking videos where the entire body is in the frame, as seen in Figure \[fig:goodwalk\].
![Model predicts ’walking’.[]{data-label="fig:goodwalk"}](./images/walkgood.png){width="\linewidth"}
![Model predicts ’walking’.[]{data-label="fig:goodwalk"}](./images/goodwalk.png){width="\linewidth"}
####
The I3D network trained on the generated videos has higher accuracy on hand-waving than that trained on HMDB. This could be in part because the classifier is more biased towards hand waving rather than the classifier being better, but using the generated videos to augment the HMDB training set does improve performance on hand waving, with only a slight fall in walking accuracy.
####
Figures \[fig:goodsit\] and \[fig:badwave\] show two more examples of predictions from the classifier trained on our dataset. The classifier is able to accurately recognise the person sitting in Figure \[fig:goodsit\] even though they are occluded by the desk, but struggles with the person who is waving in Figure \[fig:badwave\] where only the upper part of their body is shown, and they look like they could be sitting down.
![Model predicts ’sitting down’.[]{data-label="fig:badwave"}](./images/sitgood.png){width="\linewidth"}
![Model predicts ’sitting down’.[]{data-label="fig:badwave"}](./images/badwave.png){width="\linewidth"}
Ablation Study 2
----------------
To justify the complexity of generating the synthetic, we try training the I3D model on the human body models without any background (HBM-B), and on the raw videos of the subjects performing the activities from the MoVi dataset (RV). The results are shown in Table \[tab:ablation2\].
RGB Optical Flow Combined
------- ------ -------------- ----------
HBM-B 0.51 0.67 0.65
RV 0.58 0.66 0.68
: Test accuracies for model trained on ablation datasets.[]{data-label="tab:ablation2"}
The results indicate that including a background to the scene is important for RGB classification, but has less of an effect on the optical flow network. This is unsurprising as the stationary backgrounds should not have an effect on the optical flow output. The optical flow on our synthetic dataset is still higher, which shows that the occlusion and camera motion effects in the video generation are important for performance.
####
Using the training videos achieves high RGB performance and optical flow performance, but not as high as our generated videos. While the videos are more realistic, they do not offer the same range of viewing angles as the generated videos, and do not include other augmentation techniques like occlusion and camera motion.
Conclusion
==========
In this paper, we have shown that it is possible to train a high performing classifier with a purely synthetic dataset. Moreover, we have shown that synthetic data can be used alongside real data to improve on performance, with particular improvements in optical flow processing.
####
These results serve as a proof of concept, implying that synthetic data is a valuable tool in human activity recognition. By being able to generate infinite permutations of humans in environments, we are able to provide a rich dataset for training even if the videos are not completely realistic. We believe that with more sophisticated video augmentation techniques, the improvements could be even more dramatic. In particular, more realistic videos could allow for improvements in the RGB stream as well as the optical flow stream.
Future Work
===========
We identify some key areas in which our video generation could be improved.
####
*More variation in the videos.* We found that many of the videos in the test set had close-ups on the upper body of people. Including these into our data would be easy, and could improve performance. We could also include more occlusion by actively placing objects in the 3D scenes between the camera and the human.
####
*Adding clothing to the subjects.* An obvious technique to make our subjects more realistic would be to generate them with clothes. The work done by Lassner et al.[@lassner] achieves exactly this aim. They have built a deep learning generative model for people in clothing. They train models on a large image base build on top of the Chictopia 10K dataset. During our initial attempts at building our dataset, we ran into issues with their current codebase and had to opt for SMPL based models instead. Using their generative model could be a promising step towards more realistic videos.
####
*Implementing movement.* The humans we generated are all fixed on one spot in their scenes, not moving during the video. This was done because the movement information is not directly provided in the MoVi dataset, but if movement were infered from the IMU sensors and videos, our generated people could be made to walk across the scenes instead of walking on the spot. This would make for more realistic optical flow as well as RGB frames.
####
*Automating human placement.* Placing the humans in the 3D scenes was a very time consuming job as the rotation and translation parameters needed to be tuned one by one. This limited the number of scenes we could recreate. The process could be made semi-autonomous by implementing a GUI where a user could rotate the scene and place the human in acceptable locations. The possible human placement locations could also be automatically inferred from surface normal and geometric information in the scenes. A better improvement still would be to be able to automatically locate ’sittable’ objects in the environment so that the people do not sit in mid air.
####
*Generating videos at train time.* Since the videos were time consuming to generate, we generated them all before training and trained on each video multiple times. One potential advantage of a synthetic dataset is that the videos could be generated before each training epoch, to discourage the model from overfitting by never showing it the same video twice.
####
As well as improving the dataset, we could see how well our dataset works for other classifiers, and try to improve on state-of-the-art performance by augmenting with our videos.
.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a generalized structure of Bell inequalities for arbitrary $d$-dimensional bipartite systems, which includes the existing two types of Bell inequalities introduced by Collins-Gisin-Linden-Massar-Popescu \[, 040404 (2002)\] and Son-Lee-Kim \[, 060406 (2006)\]. We analyze Bell inequalities in terms of correlation functions and joint probabilities, and show that the coefficients of correlation functions and those of joint probabilities are in Fourier transform relations. We finally show that the coefficients in the generalized structure determine the characteristics of quantum violation and tightness.'
author:
- 'Seung-Woo Lee'
- Yong Wook Cheong
- Jinhyoung Lee
title: 'A generalized structure of Bell inequalities for bipartite arbitrary-dimensional systems'
---
= 10000
Introduction
============
Local-realistic theories impose constraints on any correlations obtained from measurement between two separated systems [@Bell64; @CHSH69; @tBell]. It was shown that these constraints, known as Bell inequalities, are incompatible with the quantitative predictions by quantum mechanics in case of entangled states. For example, the original Bell inequality is violated by a singlet state of two spin-1/2 particles [@Bell64]. The Clauser-Horne-Shimony-Holt (CHSH) inequality is another common form of Bell inequality, allowing more flexibility in local measurement configurations [@CHSH69]. These constraints are of great importance for understanding the conceptual features of quantum mechanics and draw the boundary between local-realistic and quantum correlations. One may doubt if there is any well-defined constraint for many high-dimensional subsystems which would eventually simulate a classical system as increasing its dimensionality to infinity [@tBell]. Therefore, constraints for more complex systems such as multi-partite or high-dimensional systems have been proposed and investigated intensively [@sWerner; @Collins02; @Kaszli00; @Masanes03; @Cerf02; @Laskowski04; @Acin02; @JLee04; @Zohren06; @WSon05; @WSon06; @Brukner02; @WSon04; @Werner01; @Peres99].
For bipartite high-dimensional systems, Collins [*et al*]{}. suggested a local-realistic constraint, called CGLMP inequality [@Collins02]. It is violated by quantum mechanics and its characteristics of violation are consistent with the numerical results provided by Kaszlikowski [*et al.*]{} [@Kaszli00]. Further, Masanes showed that the CGLMP inequality is tight [@Masanes03], which implies that the inequality has no interior bias as a local-realistic constraint. However, Acin *et al.* found that the CGLMP inequality shows maximal violation by non-maximally entangled state [@Acin02]. Zohren and Gill found the similar results when they applied CGLMP inequality to infinite dimensional systems [@Zohren06]. Recently, Son [*et al.*]{} [@WSon05] suggested a generic Bell inequality and its variant for arbitrary high-dimensional systems. The variant will be called SLK inequality throughout this paper. They showed that the SLK inequality is maximally violated by maximally entangled state. Very recently, the CGLMP inequality was recasted in the structure of the SLK inequality by choosing appropriate coefficients [@WSon06].
In this paper, we propose a generalized structure of Bell inequalities for bipartite arbitrary $d$-dimensional systems, which includes various types of Bell inequalities proposed previously. A Bell inequality in the given generalized structure can be represented either in the correlation function space or joint probability space. We show that a Bell inequality in one space can be mapped into the other space by Fourier transformation. The two types of high-dimensional Bell inequalities, CGLMP and SLK, are represented in terms of the generalized structure with appropriate coefficients in both spaces (Sec. \[section:GBI\]). We investigate the violation of Bell inequalities by quantum mechanics. The expectations of local-realistic theories and quantum mechanics are determined by the coefficients of correlation functions or joint probabilities. The CGLMP inequality is maximally violated by non-maximally entangled state while the SLK is by maximally entangled state (Sec. \[section:QV\]). We also investigate the tightness of Bell inequalities which represents whether they contain an interior bias or not at the boundary between local-realistic and quantum correlations. Then we show that the SLK is a non-tight Bell inequality while the CGLMP is tight (Sec. \[section:TI\]).
Generalized arbitrary dimensional Bell inequality {#section:GBI}
=================================================
We generalize a Bell inequality for bipartite arbitrary $d$-dimensional systems. Suppose that each observer independently choose one of two observables denoted by $A_1$ or $A_2$ for Alice, and $B_1$ or $B_2$ for Bob. Here we associate a hermitian observables $H$ to a unitary operator $U$ by the simple correspondence, $U=\exp(iH)$, and call $U$ a unitary observable [@Cerf02; @Brukner02; @JLee04; @WSon05]. We note that unitary observable representation induces mathematical simplifications without altering physical results [@note]. Each outcome takes the value of an element in the set of order $d$, $V=\{1,\omega,...,\omega^{d-1}\}$, where $\omega=\exp(2\pi i/d)$. The assumption of local-realistic theories implies that the outcomes of observables are predetermined before measurements and the role of the measurements is just to reveal the values. The values are determined only by local hidden variables $\lambda$, i.e., $A_a(\lambda)$ and $B_b(\lambda)$ for $a,b =1,2$.
We denote a correlation between specific measurements taken by two observers, as $A_a(\lambda)B^*_b(\lambda)$. Based on the local hidden-variable description, the correlation function is the average over many trials of the experiment as $$\begin{aligned}
\label{eq:CorrFn1}
C_{ab} = \int d\lambda ~\rho(\lambda)A_a(\lambda)B^*_b(\lambda),\end{aligned}$$ where $\rho(\lambda)$ is the statistical distribution of the hidden variables $\lambda$ with the properties of $\rho(\lambda)
\ge 0$ and $\int d\lambda \rho(\lambda) = 1$. The correlation function can be expanded in terms of joint probability functions over all possible outcome pairs $(k,l)$ with complex-valued weight as $$\begin{aligned}
\label{eq:CorrFn2}
C_{ab} = \sum_{k,l=0}^{d-1}\omega^{k-l}P(A_{a}=k, B_{b}=l),\end{aligned}$$ where $\omega^{k-l}$ is called a correlation weight and $P(A_{a}=k, B_{b}=l)$ is a joint probability of Alice and Bob obtaining outcomes $\omega^k$ and $\omega^l$ respectively. Here we use the powers $k$ and $l$ of the outcomes $\omega^k$ and $\omega^l$ for the arguments of the joint probability as there is one-to-one correspondence.
We assume in general a correlation weight $\mu_{k,l}$ to satisfy certain conditions [@WSon04]. \[\] The correlation expectation vanishes for a bipartite system with a locally unpolarized subsystem: $\sum_{k}\mu_{k,l}=0,\forall l$ and $\sum_{l}\mu_{k,l}=0,\forall k$ \[\] The correlation weight is unbiased over possible outcomes of each subsystem (translational symmetry within modulo $d$): $
\mu_{k,l}=\mu_{k+\gamma, l+\gamma},\forall \gamma$. \[\] The correlation weight is uniformly distributed modulo $d$: $|\mu_{k+1,l}-\mu_{k,l}|=|\mu_{k,l+1}-\mu_{k,l}|,\forall k,l$. The correlation weight in Eq. (\[eq:CorrFn2\]) $\omega^{k-l}$ satisfies all the conditions, can be written as $\omega^{\alpha}$ where $\alpha\equiv k-l \in\{0,1,...,d-1\}$ and it obeys $\sum_{\alpha}\omega^{\alpha}=0$.
Let us now consider higher-order($n$) correlations following also the local hidden-variable description. The $n$-th order correlation function averaged over many trials of the experiment corresponds to the $n$-th power of $1$-st order correlation as $$\begin{aligned}
\label{eq:CorrFn3}
C^{(n)}_{ab} &=& \int d\lambda
~\rho(\lambda) \left(A_a(\lambda) B^*_b(\lambda)\right)^n \nonumber \\
&=&\sum_{k,l=0}^{d-1}\omega^{n(k-l)}P(A_{a}=k, B_{b}=l) \nonumber \\
&=&\sum_{\alpha=0}^{d-1}\omega^{n\alpha}P(A_{a}\doteq
B_{b}+\alpha),\end{aligned}$$ where the $n$-th order correlation weight $\omega^{n\alpha}$ also satisfies the above conditions, , and , and $P(A_{a}\doteq B_{b}+\alpha)$ is the joint probability of local measurement outcomes differing by a positive residue $\alpha$ modulo $d$. Here we note that the higher-order correlations Eq. (\[eq:CorrFn3\]) show the periodicity of $C^{(d+n)}_{ab}=C^{(n)}_{ab}$ and they have the Fourier relation with the joint probabilities as given in Eq. (\[eq:CorrFn3\]).
We present a generalized Bell function for arbitrary $d$-dimensional system using higher-order correlation functions as $$\begin{aligned}
\label{eq:GeneralBF}
{\cal B}=\sum_{a,b}\sum_{n=0}^{d-1}f_{ab}(n)C^{(n)}_{ab},\end{aligned}$$ where coefficients $f_{ab}(n)$ are functions of the correlation order $n$ and the measurement configurations $a$, $b$. They determine the constraint of local-realistic theories with a certain upper bound and its violation by quantum mechanics will be investigated in Sec. \[section:QV\]. The zero-th order correlation has no meaning as it simply shift the value of ${\cal
B}$ by a constant and is chosen to vanish, i.e., $\sum_{a,b}f_{ab}(0)=0$. The Bell function in Eq. (\[eq:GeneralBF\]) is rewritten in terms of the joint probabilities given in Eq. (\[eq:CorrFn3\]), as $$\begin{aligned}
\label{eq:GeneralBF2}
{\cal B}=\sum_{a,b}\sum_{\alpha=0}^{d-1}\epsilon_{ab}(\alpha)P(A_{a}\doteq B_{b}+\alpha),\end{aligned}$$ where $\epsilon_{ab}(\alpha)$ are coefficients of the joint probabilities $P(A_{a}\doteq B_{b}+\alpha)$.
We note that the coefficients $\epsilon_{ab}(\alpha)$ are obtained by the Fourier transformation of $f_{ab}(n)$ based on the kernel of a given correlation weight as $$\begin{aligned}
\label{eq:CoeffRelation1}
\epsilon_{ab}(\alpha)&=&\sum_{n=0}^{d-1}f_{ab}(n)\omega^{n\alpha},\\
\label{eq:CoeffRelation2}
f_{ab}(n)&=&\frac{1}{d}\sum_{\alpha=0}^{d-1}\epsilon_{ab}(\alpha)\omega^{-n\alpha}.\end{aligned}$$ It is remarkable that one can represent a given Bell function either in the correlation function space or joint probability space by using the Fourier transformation of the coefficients between them. This is the generalization of the Fourier transformation in 2-dimensional Bell inequalities provided by Werner *et al.* [@Werner01].
Different Bell inequalities can be represented by altering coefficients of the generalized structure, including previously proposed Bell inequalities in bipartite systems. In the case of $d=2$, CHSH-type inequalities can be obtained with coefficients as $f(1)=(1,1,-1,1)$ and $\epsilon_{ab}(\alpha)=f_{ab}(1)(-1)^{\alpha}$ where $\alpha\in\{0,1\}$. For arbitrary $d$-dimensional systems, the two types of Bell inequalities, CGLMP and SLK, are represented in terms of the generalized structure with appropriate coefficients obtained as follows.
[*CGLMP inequality*]{} - As it was originally proposed in terms of joint probabilities [@Collins02], the Bell function of the CGLMP inequality is in the form of (\[eq:GeneralBF2\]) and its coefficients are given as $$\begin{aligned}
\label{eq:CoeffCGLMP}
\nonumber
\epsilon_{11}(\alpha)&=&1-\frac{2\alpha}{d-1},~~~
\epsilon_{12}(\alpha)=-1+\frac{2\dot{(\alpha-1)}}{d-1},\\
\epsilon_{21}(\alpha)&=&-1+\frac{2\alpha}{d-1},~~~
\epsilon_{22}(\alpha)=1-\frac{2\alpha}{d-1},\end{aligned}$$ where the dot implies the positive residue modulo $d$. By using the inverse Fourier transformation in Eq. (\[eq:CoeffRelation2\]) the coefficients for the correlation function representation are obtained as $$\begin{aligned}
\label{eq:CoeffCGLMP2}
\nonumber
f_{11}(n\neq0)&=&\frac{2}{d-1}\left(\frac{1}{1-\omega^{-n}}\right),\\
\nonumber f_{12}(n\neq0)&=&\frac{2}{d-1}\left(\frac{1}{1-\omega^{n}}\right),\\
\nonumber f_{21}(n\neq0)&=&\frac{2}{d-1}\left(\frac{1}{\omega^{-n}-1}\right),\\
\nonumber f_{22}(n\neq0)&=&\frac{2}{d-1}\left(\frac{1}{1-\omega^{-n}}\right),\\
f_{ab}(n=0)&=&0 ~~~~~\forall a,b,\end{aligned}$$ where the sum of the 0-th order coefficients vanishes and does not affect the characteristics of the Bell inequality.
[*SLK inequality*]{} - It was introduced in terms of correlation functions [@WSon05], and the coefficients are given by $$\begin{aligned}
\label{eq:CoeffSLK}
\nonumber
f_{11}(n\neq0)&=&(\omega^{n\delta}+\omega^{(n-d)\delta})/4,\\
\nonumber f_{12}(n\neq0)&=&(\omega^{n(\delta+\eta_1)}+\omega^{(n-d)(\delta+\eta_1)})/4,\\
\nonumber f_{21}(n\neq0)&=&(\omega^{n(\delta+\eta_2)}+\omega^{(n-d)(\delta+\eta_2)})/4,\\
\nonumber f_{22}(n\neq0)&=&(\omega^{n(\delta+\eta_1+\eta_2)}+\omega^{(n-d)(\delta+\eta_1+\eta_2)})/4,\\
f_{ab}(n=0)&=&0 ~~~~~\forall a,b,\end{aligned}$$ where $\delta$ is a real number, called a variant factor, and $\eta_{1,2}\in\{+1/2,-1/2\}$. By varying $\delta$ and $\eta_{1,2}$, one can have many variants of SLK inequalities. For all the variants the coefficients in the joint probability picture are obtained as $$\begin{aligned}
\label{eq:CoeffSLK2}
\nonumber
\epsilon_{11}(\alpha)&=&S(\delta+\alpha),\\
\nonumber \epsilon_{12}(\alpha)&=&S(\delta+\eta_1+\alpha),\\
\nonumber \epsilon_{21}(\alpha)&=&S(\delta+\eta_2+\alpha),\\
\epsilon_{22}(\alpha)&=&S(\delta+\eta_1+\eta_2+\alpha),\end{aligned}$$ where $$\begin{aligned}
\nonumber
S(x\neq0)&=&\frac{1}{4}(\cot{\frac{\pi}{d}x}\sin{2\pi x}-\cos{2\pi
x}-1),\\
S(x=0)&=&\frac{1}{2}(d-1).\end{aligned}$$ We have shown that those two types of high-dimensional inequalities have different coefficients but the same generalized structure. In the frame work of the generalized structure we will now study how the coefficients determine the characteristics of Bell inequalities such as the degree of violation and tightness.
Violation by Quantum Mechanics {#section:QV}
==============================
In order to see the violation of Bell inequalities by quantum mechanics we need to know the upper bound by local hidden variable theories. We note that a probabilistic expectation of a Bell function is given by the convex combination of all possible deterministic values of the Bell function and the local-realistic upper bound is decided by the maximal deterministic value. Let $\alpha_{ab} = \alpha$ such that $P(A_a \doteq B_b + \alpha) = 1$. The assumption of local-realistic theories implies that the values $\alpha_{ab}$ are predetermined. For a Bell function in the form of Eq. (\[eq:GeneralBF\]) or (\[eq:GeneralBF2\]), they obey the constraint, $\alpha_{11}+\alpha_{22}\doteq
\alpha_{12}+\alpha_{21}$, because of the identity, $A_1 - B_1 +
A_2 -B_2 = A_1 - B_2 + A_2 - B_1$. The local-realistic upper bound of the Bell function is therefore given by
$$\begin{aligned}
\label{eq:LRmax}
{\cal
B}^{\mathrm{max}}_{\mathrm{LR}}=\max_{\alpha_{ab}}[\sum_{a,b}\epsilon_{ab}(\alpha_{ab})|\alpha_{11}+\alpha_{22}\doteq
\alpha_{12}+\alpha_{21}].\end{aligned}$$
The quantum expectation value for arbitrary quantum state $\hat{\rho}$ is written by $$\begin{aligned}
\label{eq:QBvalue}
\nonumber
{\cal B}_{\mathrm{QM}}(\hat{\rho})&=&\mathrm{Tr}(\hat{{\cal B}}\hat{\rho})\\
&=&\sum_{a,b}\sum_{n=0}^{d-1}f_{ab}(n)\mathrm{Tr}(\hat{C}_{ab}^{(n)}\hat{\rho}),\end{aligned}$$ where $\hat{{\cal B}}$ is the Bell operator defined by replacing the correlation function in Eq. (\[eq:GeneralBF\]) with correlation operator, $\hat{C}_{ab}^{(n)}=\sum_{k,l}\omega^{n(k-l)}\hat{P}_a\otimes\hat{P}_b$ where $\hat{P}_a$,$\hat{P}_b$ are projectors onto the measurement basis denoted by $a,b$. If an expectation value of any quantum state exceeds the local realistic bound ${\cal
B}_{\mathrm{LR}}^{\mathrm{max}}$, i.e., the Bell inequality is violated by quantum mechanics, the composite system is entangled and shows nonlocal quantum correlations. The maximal quantum expectation is called quantum maximum ${\cal
B}_{\mathrm{QM}}^{\mathrm{max}}$ and corresponds to the maximal eigenvalue of the Bell operator. In the case of $d=2$, with the coefficients $f(1)=(1,1,-1,1)$ we can obtain the quantum maximum, ${\cal B}_{\mathrm{QM}}^{\mathrm{max}}=2\sqrt{2}$, which is in agreement with the Cirel’son bound [@Cirel80].
In the presence of white noise, a maximally entangled state $|\psi_m \rangle$ becomes $\hat{\rho}=p|\psi_{m} \rangle \langle
\psi_{m}|+(1-p)\openone/d^2$ where $p$ is the probability that the state is unaffected by noise. Then, the minimal probability for the violation is $p^{\mathrm{min}}={\cal
B}^{\mathrm{max}}_{\mathrm{LR}}/{\cal B}_{\mathrm{QM}}(|\psi_{m}
\rangle)$. We now investigate the violation of two types of Bell inequalities, CGLMP and SLK, and compare them as follows.
[*CGLMP inequality*]{} - The local-realistic upper bound, ${\cal
B}^{\mathrm{max}}_{\mathrm{LR}}=2$, can be obtained as Eq. (\[eq:LRmax\]). The quantum expectation can also be obtained as Eq. (\[eq:QBvalue\]) and it is consistent with the result in Ref. [@Collins02]. Acin *et al.* found, however, that the CGLMP inequality shows maximal violation for non-maximally entangled states [@Acin02]. For $3$-dimensional system, the quantum maximum is ${\cal B}_{\mathrm{QM}}^{\mathrm{max}}\simeq2.9149$ for the non-maximally entangled state, $(1/\sqrt{n})(|00\rangle+\gamma|11\rangle+|22\rangle)$ where $\gamma\simeq0.7923$ and $n=2+\gamma^2$. It is higher than the expectation by maximally entangled state, ${\cal B}(|\psi_{m}
\rangle)\simeq2.8729$. The expectation of the CGLMP is shown in Fig. \[fig1\] against the entanglement degree $\gamma$, once the local measurements are chosen such that they maximize the Bell function for the maximally entangled state. Further, we also note that the minimal violation probability($p^{\mathrm{min}}$) of the CGLMP decreases as the dimension $d$ increases.
![The expectation value of (a) the CGLMP and (b) the optimal SLK for $d=3$ as varying the value $\gamma$ for the quantum state, $(1/\sqrt{n})(|00\rangle+\gamma|11\rangle+|22\rangle)$ where $n=2+\gamma^2$. The SLK takes the maximum $2$ when the state is maximally entangled ($\gamma=1$), whereas the CGLMP takes the maximum $2.9149$ for a partially entangled state ($\gamma\simeq0.7923$). The dashed lines indicate the local-realistic upper bounds.[]{data-label="fig1"}](Figure1){width="45.00000%"}
[*SLK inequality*]{} - Many variants of the SLK Bell inequality are obtained by varying $\delta$ and $\eta_{1,2}$. All the variants of the SLK have the same quantum maximum $d-1$ for a maximally entangled state $|\psi_m\rangle$, ${\cal B}^{\mathrm{max}}_{\mathrm{QM}}={\cal
B}_{\mathrm{QM}}(|\psi_{m} \rangle)$, as we prove in Appendix \[app:QUB\]. On the other hand, the local-realistic upper bounds depend on the variants. The local-realistic upper bound ${\cal
B}^{\mathrm{max}}_{\mathrm{LR}}$ is a function of the variant factor $\delta$. It shows a periodicity, ${\cal
B}^{\mathrm{max}}_{\mathrm{LR}}(\delta)={\cal
B}^{\mathrm{max}}_{\mathrm{LR}}(\delta+1/2)$, and without loss of generality it suffices to consider $0\leq\delta < 1/2$. If $\delta=0$, the local-realistic upper bound is the same as the quantum maximum $d-1$, and thus the corresponding Bell inequality is not violated by quantum mechanics. When $\delta=1/4$, we have the lowest local-realistic upper bound as $$\begin{aligned}
\label{eq:lowestLRUB} \min_{\delta}[{\cal
B}^{\mathrm{max}}_{\mathrm{LR}}(\delta)]=
\frac{1}{4}(3\cot{\frac{\pi}{4d}}-\cot{\frac{3\pi}{4d}})-1,\end{aligned}$$ and for other cases the bound values are symmetric at $\delta=1/4$, i.e., ${\cal B}^{\mathrm{max}}_{\mathrm{LR}}(1/4+\epsilon)={\cal
B}^{\mathrm{max}}_{\mathrm{LR}}(1/4-\epsilon)$ for $0<\epsilon\leq1/4$. Therefore, we will call the variant of $\delta=1/4$, which gives the maximal difference between quantum maximum and local-realistic upper bound, as the [*optimal*]{} SLK inequality and use it for comparing to the CGLMP. In Fig. \[fig1\], we present the quantum expectation of the SLK for 3-dimensional systems against the degree $\gamma$, where the local measurements are chosen such that they maximize the Bell function for the maximally entangled state. Note that the SLK inequality shows the maximal violation by maximally entangled states and the minimal violation probability $p^{\mathrm{min}}$ increases as the dimension $d$ increases.
By investigating the violation of two inequalities, CGLMP and SLK, based on the generalized structure of Bell inequalities, we showed that those two types have very different characteristics. The SLK inequality is maximally violated by maximally entangled states as being consistent with our intuition whereas the CGLMP is maximally violated by non-maximally entangled states. We remark that the coefficients of the given generalized structure determine the characteristics of quantum violations.
Tightness of Bell inequalities {#section:TI}
==============================
The set of possible outcomes for a given measurement setting forms a convex polytope in the joint probability space or alternatively in the correlation function space [@Peres99; @Werner01; @Masanes03; @Laskowski04]. Each generator of the polytope represents the predetermined measurement outcome called local-realistic configuration. All interior points of the polytope are given by the convex combination of generators and they represent the accessible region of local-realistic theories associated with the probabilistic expectations of measurement outcomes. Therefore, every facet of the polytope is a boundary of halfspace characterized by a linear inequality, which we call *tight Bell inequality*. There are non-tight Bell inequalities which contain the polytope in its halfspace. As the non-tight Bell inequality has interior bias at the boundary between local-realistic and quantum correlations, one might say it to be the worse detector of the nonlocal test [@Werner01; @Masanes03; @Laskowski04].
The Bell polytope is lying in the joint probability space of dimension $h$, the degrees of freedom for the measurement raw data. For a bipartite system, two observables per party and $d$-dimensional outcomes, the joint probability, $P(A_a=k,B_b=l)$ where $k,l=0,1,...,d-1$ and $a,b=1,2$, can be arranged in a $4d^2$-dimensional vector space. However, the joint probabilities have two constraints, i.e., normalization and no-signaling constraints, which reduce dimension by $4d$ [@Masanes03]. The generators in the $h$-dimensional space can be written, following the notations in Ref. [@Masanes03], as $$\label{eq:generator}
\mathbf{G}=|A_1,B_1\rangle\oplus|A_1,B_2\rangle\
\oplus|A_2,B_1\rangle\ \oplus|A_2,B_2\rangle\,$$ where $|n\rangle$ stands for $|n \mod d\rangle$ and is the $d$-dimensional vector with a 1 in the $n$-th component and $0$s in the rest.
In order to examine the tightness of a given generalized Bell inequality, in general one considers the following conditions that every tight Bell inequality fulfills [@Masanes03]. \[\] All the generators must belong to the half space of a given facet. \[\] Among the generators on the facet, there must be $h$ which are linearly independent. First, it is straightforward that all generators fulfill the inequality as the Bell inequality derived to do. As the local-realistic upper bound is the maximum among expectation values of local-realistic configurations, all generators are located below the local-realistic upper bound, ${\cal
B}^{\mathrm{max}}_{\mathrm{LR}}$. Thus the first condition is fulfilled. Second, we examine whether there are $h$ linearly independent generators which give the value of the local-realistic bound, ${\cal B}^{\mathrm{max}}_{\mathrm{LR}}$. By the predetermined local-realistic values $\alpha_{ab}$, the generators (\[eq:generator\]) become $$\begin{aligned}
\nonumber &|A,A-\alpha_{11}\rangle\oplus|A,A-\alpha_{12}\rangle
\oplus|A-\alpha_{12}+\alpha_{22},A-\alpha_{11}\rangle\ \\
&\oplus|A-\alpha_{11}+\alpha_{21},A-\alpha_{12}\rangle,\end{aligned}$$ where $A\in\{0,1,...,d-1\}$ and the number of linearly independent generators is determined by the number of sets $\{\alpha_{ab}\}$ that give the local-realistic upper bound. If the number of linear independent generators is not smaller than $h=4d(d-1)$, the corresponding Bell inequality is tight.
[*CGLMP inequality*]{} - For the CGLMP inequality the local-realistic upper bound is achieved when $\alpha_{11}+\alpha_{22}-\dot{(\alpha_{12}-1)}-\alpha_{21}+d-1=0$. The condition allows the sufficient number of linearly independent generators and the CGLMP inequality is tight [@Masanes03].
[*SLK inequality*]{} - For the optimal SLK inequality, the upper bound is obtained in the case that $\{\alpha_{11},\alpha_{12},\alpha_{21},\alpha_{22}\}$ is equal to one of four sets; $\{0,0,d-1,d-1\}$, $\{0,0,0,0\}$,$\{0,1,d-1,0\}$,$\{d-1,0,d-1,0\}$. Thus there are four types of generators as $$\begin{aligned}
\nonumber &&|A,A\rangle\oplus|A,A\rangle\ \oplus|A-1,A\rangle\
\oplus|A-1,A\rangle\,\\
\nonumber &&|A,A\rangle\oplus|A,A\rangle\ \oplus|A,A\rangle\
\oplus|A,A\rangle\,\\
\nonumber &&|A,A\rangle\oplus|A,A-1\rangle\ \oplus|A-1,A\rangle\
\oplus|A-1,A-1\rangle\,\\
&&|A,A+1\rangle\oplus|A,A\rangle\ \oplus|A,A+1\rangle\
\oplus|A,A\rangle\,\end{aligned}$$ which are linearly independent with $A\in\{0,1,...,d-1\}$. There are only $4d$ linearly independent generators which are smaller than $h=4d(d-1)$, the tightness condition T.2. Thus the optimal SLK inequality is non-tight. On the other hand, the SLK inequality for $\delta=0$ is tight but it is not violated by quantum mechanics.
Remarks
=======
In summary, we presented a generalized structure of the Bell inequalities for arbitrary $d$-dimensional bipartite systems by considering the correlation function specified by a well-defined complex-valued correlation weight. The coefficients of a given Bell inequality in the correlation function space and the joint probability space were shown to be in the Fourier relation. Two known types of high-dimensional Bell inequalities, CGLMP and SLK, were shown to have the generalized structure in common and we found their coefficients in both spaces.
Based on the generalized structure, we investigated characteristics of the Bell inequalities such as quantum violation and tightness. We found that the CGLMP and SLK inequalities show different characteristics. For instance, the SLK inequality is maximally violated by maximally entangled states, which is consistent with the intuition “the larger entanglement, the stronger violation against local-realistic theories,” whereas the CGLMP inequality is maximally violated by the non-maximally entangled state as previously shown by Acin [*et al.*]{} [@Acin02]. On the other hand, in analyzing the tightness of the inequalities, the CGLMP is tight but the SLK inequality is found to be non-tight for $\delta\neq 0$, implying that the SLK inequality has interior bias at the boundary between local-realistic and quantum correlations.
The correlation coefficients of Bell inequalities play a crucial role in determining their characteristics of quantum violation and tightness. This implies that by altering the coefficients in the generalized structure one can construct other Bell inequalities. The present work opens a possibility of finding a new Bell inequality that fulfills both conditions of the maximal violation by maximal entanglement and the tightness.
The authors thanks D. Jaksch for useful discussions. This work was supported by MOST/KOSEF through the Quantum Photonic Science Research Center and the Korean Research Foundation Grant funded by the Korean Government (MOEHRD) (KRF-2005-041-C00197). S.-W.Lee was supported by the EU through the STREP project OLAQUI.
Quantum Maximum of all variant SLK inequalities {#app:QUB}
===============================================
We shall prove that all the variant SLK inequalities take $d-1$ as the quantum maximum. The Bell operator of the SLK variants can be written as $$\begin{aligned}
\hat{{\cal
B}}_S&=&\frac{1}{2}\sum^{d-1}_{n=1}\boldsymbol{\alpha}\cdot\boldsymbol{\tilde\beta},\end{aligned}$$ where $\boldsymbol{\alpha}=(\hat{A}^{\dagger n}
_1,\hat{A}^{\dagger
n}_2)^{T}$ and $\boldsymbol{\tilde\beta}=\mathbf{U}\boldsymbol{\beta}$ with $\boldsymbol{\beta}=(\hat{B}^n_1,\hat{B}^n_2)^{T}$ and $\mathbf{U}$ is a $2 \times 2$ unitary matrix with elements, $$\begin{aligned}
\nonumber U_{11}&=&(\omega^{n\delta}+\omega^{(n-d)\delta})/2,\\
\nonumber U_{12}&=&(\omega^{n(\delta+\eta_1)}+\omega^{(n-d)(\delta+\eta_1)})/2,\\
\nonumber U_{21}&=&(\omega^{n(\delta+\eta_2)}+\omega^{(n-d)(\delta+\eta_2)})/2,\\
U_{22}&=&(\omega^{n(\delta+\eta_1+\eta_2)}+\omega^{(n-d)(\delta+\eta_1+\eta_2)})/2,\end{aligned}$$ where $\eta_{1,2}\in\{1/2,-1/2\}$.
The expectation of the Bell operator is given by $$\begin{aligned}
\label{eq:qefvbo}
&&\frac{1}{2}\left|\sum_{n=1}^{d-1} \bra{\psi} \boldsymbol{\alpha} \cdot
\boldsymbol{\tilde\beta} \ket{\psi} \right| \le
\frac{1}{2}\sum_{n=1}^{d-1} \left| \bra{\psi} \boldsymbol{\alpha} \cdot
\boldsymbol{\tilde\beta} \ket{\psi} \right| \nonumber \\
&\le&
\frac{1}{2}\sum_{n=1}^{d-1} \left(\left| \bra{\psi} \alpha_1 \otimes {\tilde\beta}_1
\ket{\psi} \right| + \left| \bra{\psi} \alpha_2 \otimes {\tilde\beta}_2
\ket{\psi}\right| \right) \nonumber \\
&\le& \frac{1}{\sqrt{2}}
\sum_{n=1}^{d-1} \sqrt{\left| \bra{\psi} \alpha_1 \otimes {\tilde\beta}_1
\ket{\psi} \right|^2 + \left| \bra{\psi} \alpha_2 \otimes {\tilde\beta}_2
\ket{\psi}\right|^2} \nonumber \\
&=&\frac{1}{\sqrt{2}}
\sum_{n=1}^{d-1} \sqrt{\sum_{i=1}^2 \left| \bra{\psi} \alpha_i \otimes
{\tilde\beta}_i \ket{\psi} \right|^2},\end{aligned}$$ where we consecutively used the triangle inequality and the arithmetic-geometric means inequality, $ 2 |a||b| \le
|a|^2+|b|^2$. Note that $$\begin{aligned}
\nonumber
\sum_{i=1}^2 \left| \bra{\psi} \alpha_i \otimes {\tilde\beta}_i
\ket{\psi} \right|^2 &\le& \sum_{i=1}^2 \bra{\psi} \left(\alpha_i^\dag
\otimes {\tilde\beta}_i^\dag \right) \left(\alpha_i \otimes {\tilde\beta}_i
\right) \ket{\psi}\\
&=&\sum_{i=1}^2 \bra{\psi} \openone \otimes {
\tilde\beta}_i^\dag {\tilde\beta}_i \ket{\psi},\end{aligned}$$ where we used that $\alpha_i$ is unitary. Here the above inequality is obtained by reasoning that $\hat{Q} \equiv \openone
- \ket{\psi}\bra{\psi}$ is a positive operator as $\bra{\phi}
\hat{Q} \ket{\phi} = 1 - |\bracket{\phi}{\psi}|^2 \ge 0$ for any $\ket{\phi}$, and $|\bra{\psi} \hat{C} \ket{\psi}|^2=\bra{\psi}
\hat{C}^\dag \ket{\psi}\bra{\psi} \hat{C} \ket{\psi} = \bra{\psi}
\hat{C}^\dag (\openone - \hat{Q}) \hat{C} \ket{\psi} =
\bra{\psi}\hat{C}^\dag\hat{C} \ket{\psi} -
\bra{\psi}\hat{Q}\ket{\psi} \le \bra{\psi}\hat{C}^\dag\hat{C}
\ket{\psi}$, where $\hat{C}\equiv \alpha_i \otimes
{\tilde\beta}_i$. Since $\sum_i \tilde{\beta}_i^\dag
\tilde{\beta}_i = \sum_{jk} \sum_{i} U_{ij}^* U_{ik} \beta_j^\dag
\beta_k = \sum_{jk} \delta_{jk} \beta_j^\dag \beta_k = \sum_{i}
\beta_i^\dag \beta_i = 2 \openone$, it is clear that $$\begin{aligned}
\sum_{i=1}^2 \left| \bra{\psi} \alpha_i \otimes \tilde{\beta}_i
\ket{\psi} \right|^2 \le \bra{\psi} \openone \otimes \sum_{i=1}^2 {
\beta}_i^\dag {\beta}_i \ket{\psi}= 2.\end{aligned}$$ Hence the upper bound for all variants of the SLK is $$\begin{aligned}
\left|\bra{\psi} \hat{{\cal B}}_S \ket{\psi} \right| \le d-1.\end{aligned}$$ Since all SLK Bell operators have the eigenvalue $d-1$ for maximally entangled states, the upper bound is reachable. Therefore, $d-1$ is the quantum maximum for all variants of the SLK inequality.
[99]{} J. S. Bell, Physics [**1**]{}, 195 (1964). J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt, , 880 (1969). J. S. Bell, [*Speakable and unspeakable in quantum mechanics*]{} (Cambridge Univ., Cambridge 1987), p198. Open Problems, http://www.imaph.tu-bs.de/qi/problems/1.html. D. Collins, N. Gisin, N. Linden, S. Massar, and S. Popescu, , 040404 (2002). D. Kaszlikowski, P. Gnacinski, M. Zukowski, W. Miklaszewski, and A. Zeilinger, , 4418 (2000); T. Durt, D. Kaszlikowski, and M. Zukowski, , 024101 (2001). L. Masanes, Quant. Inf. Comp. [**3**]{}, 345 (2003). A. Acín, T. Durt, N. Gisin, and J. I. Latorre, , 052325 (2002). S. Zohren, R. D. Gill, e-print arXiv:quant-ph/0612020 (2006). W. Son, J. Lee, and M. S. Kim, , 060406 (2006). W. Son, Č. Brukner, and M. S. Kim, , 110401 (2006). N. J. Cerf, S. Massar, and S. Pironio, , 080402 (2002). J. Lee, S.-W. Lee, and M. S. Kim, , 032316 (2006). Č. Brukner, M. Zukowski, and A. Zeilinger, , 197901 (2002). W. Son, J. Lee, and M. S. Kim, J. Phys. A [**37**]{}, 11897 (2004). R. F. Werner and M. M. Wolf, , 032112 (2001). W. Laskowski, T. Paterek, M. Zukowski, and Č. Brukner, , 200401 (2004). A. Peres, [*All the Bell Inequalities*]{}, Foundations of Physics [**29**]{}, 589-614 (1999). Instead of the complex eigenvalues, one may consider real eigenvalues but then he/she has to employ rather complicated form of correlation weights to obtain Bell inequalities equivalent to ones derived in the present paper. B. S. Cirel’son, Lett. Math. Phys. [**4**]{}, 93 (1980).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Telescope Array (TA) is the largest ultrahigh energy (UHE) cosmic ray observatory in the northern hemisphere TA is a hybrid experiment with a unique combination of fluorescence detectors and a stand-alone surface array of scintillation counters. We will present the spectrum measured by the surface array alone, along with those measured by the fluorescence detectors in monocular, hybrid, and stereo mode. The composition results from stereo TA data will be discussed. Our report will also include results from the search for correlations between the pointing directions of cosmic rays, seen by the TA surface array, with active galactic nuclei.'
author:
- 'C. C. H. Jui$^{\dagger}$, for the Telescope Array Collaboration'
title: 'Cosmic Ray in the Northern Hemisphere: Results from the Telescope Array Experiment'
---
Introduction
============
The Telescope Array (TA) experiment [@dpf2011_cchjui:TA] is a collaboration of 26 universities and research institutions in Japan, U.S., South Korea, Russia, and Belgium. The core of the collaboration consists of key members from the Akeno Giant Air Shower Array (AGASA) [@dpf2011_cchjui:agasa] in Japan, and the High Resolution Fly’s Eye (HiRes) experiment [@dpf2011_cchjui:hires]. TA combines the large area scintillation ground array technique developed by AGASA with that of the air fluorescence method pioneered by the Fly’s Eye (FE) [@dpf2011_cchjui:fe] at the University of Utah, and later improved by the HiRes group.
Telescope Array is located in the central western desert of Utah, near the city of Delta, about 250 km south west of Salt Lake City. The arrangement of the experiment is shown in figure \[dpf2011\_cchjui:fig001\]. The new experiment consists of three fluorescence detector (FD) stations, marked in the figure by the green squares, located at the periphery of a ground array of 507 surface detectors (SD). Each SD unit, shown in figure \[dpf2011\_cchjui:fig001\] by the black squares, consists of a scintillation counter mounted on a raised steel frame. They are deployed in a square grid of 1.2 km nearest-neighbor spacing, and the full array covers a total of about 730 km$^{2}$.
![The layout of the Telescope Array experiment. The black squares show the location of the 507 SD’s. The green squares mark the FD stations at the periphery of the ground array. The three communications towers near the FD stations are indicated by orange circles. The central laser facility, at the center of the array, is shown by the blue cross.[]{data-label="dpf2011_cchjui:fig001"}](dpf2011_cchjui_fig001.eps){width="80mm"}
Each of the three FD stations has a field of view (FOV) of about 30$^{\circ}$ in elevation and about 110$^{\circ}$ in azimuth. The central laser facility (CLF), which initially operated one vertical pulsed YAG laser (355 nm), is marked by the blue cross in figure \[dpf2011\_cchjui:fig001\]. The location of the CLF is equidistant from all three FD stations so that the vertical pulses can be used for cross-calibration of the three stations independent of the aerosol concentration in the air. The three FD stations are also oriented such that the CLF lies at the center of view of each.
TA Surface Detectors
====================
Each TA scintillation counter contains two slabs of double-layered plastic scintillators with an overall collection area of 3.0 m$^{2}$. The scintillation photons are collected by wavelength-shifting optical fibers laid in extruded grooves on the surface of the scintillators. All of the light collected from the top and bottom layers are each separately collected into a single photomultiplier tube (PMT). Each SD unit is powered entirely by its own solar panel-battery power supply, and communicates over a 24 GHz wireless point-to-point link with one of three communication towers. The towers are located near each of the three FD stations.
The SD counters are self-calibrated using minimum-ionizing cosmic muons. these muons provide a convenient unit of “vertical equivalent muon” (VEM). The output of the PMT from each scintillator is monitored continuously at 40 million samples per second (40 MSPS). Pulse data are recorded into a storage buffer when a cluster of at least 1/3 VEM is observed. An event trigger is formed when a minimum of three adjacent counters each detects a cluster of at least 3 VEM each. A typical event is shown in figure \[dpf2011\_cchjui:fig002\].
![(a) Left: Display of a typical air shower event captured by the TA surface detector array, where each circle represents a triggered SD unit. (b) Top Right: Fit of onset times of the triggered SD units to determine the shower arrival direction. (c) Bottom Right: Fit of the measured particle density vs. distance from the shower axis and the interpolation to obtain the density at 800 m (S800). []{data-label="dpf2011_cchjui:fig002"}](dpf2011_cchjui_fig002.eps){width="120mm"}
In figure \[dpf2011\_cchjui:fig002\](a), the hit counters and the their recorded densities are shown by the circles and their area. The locations of the counters are given by row and column number (at 1.2 km spacing). The color scheme shows the arrival time in terms of equivalent distance traveled by light, divided by the detector spacing. The location, signal size, and onset time detected by each counter is used to fit for the core (centroid) location and arrival direction of the shower. This geometry fit uses a modified Linsley time delay function [@dpf2011_cchjui:linsley] to describe the curvature of shower front, and the AGASA lateral density function (LDF) [@dpf2011_cchjui:agasa-ldf] to predict the fall of density from the core of the shower. The result of this fit is illustrated in figure \[dpf2011\_cchjui:fig002\](b), which plots the onset time vs. distance along the direction indicated by the arrow in figure \[dpf2011\_cchjui:fig002\](a). The particle densities from the hit SD units are then plotted as a function of its perpendicular distance to the shower core, and the density at 800 meters (S800) is interpolated from the fit to the AGASA LDF function, as shown in figure \[dpf2011\_cchjui:fig002\](c). The S800 value is then compared to the average from simulated events (shown in figure \[dpf2011\_cchjui:fig003\]) and the measured energy is interpolated according to the measured zenith angle.
![ The variation of the average particle density at 800 meters from core (S800) with the energy and zenith angle of simulated air showers. Energies of real air showers are determined by comparing the measured S800 to this plot. []{data-label="dpf2011_cchjui:fig003"}](dpf2011_cchjui_fig003.eps){width="70mm"}
TA Fluorescence Detectors
=========================
A total of 38 fluorescence telescopes are divided into three stations. The first of these was constructed on Black Rock (BR) Mesa, at the southeastern corner of the surface array. A second station is located at Long Ridge (LR) on the southwestern flank of the SD array. The BR and LR detectors were built in Japan based on essentially the same specifications as the telescopes used by the HiRes experiment, but with larger mirrors of 6.8 m$^{2}$ area (compared to 5.2 m$^{2}$ for HiRes). Each site consists of 12 telescopes with 256-pixel (16x16 in a triangular lattice) cameras. Each pixel, instrumented by a hexagonal PMT, covers a cone of 1.1$^{\circ}$ in the sky. A typical event captured by the FD station at Black Rock is shown in figure \[dpf2011\_cchjui:fig004\]. For each triggered event, the 10 MHz FADC data from each channel are scanned for pulses. Those channels containing a three sigma excess over background (primarily night sky) fluctuations are displayed as circles, with the area of the circle being proportional to the integrated pulse area.
![ Display of a typical downward air shower event captured by the FD station at Black Rock. The circles correspond to channels with pulses in excess of three sigma over the (night sky) background, and the area of the circles represent the integrated pulse area. Those pixels associated with the air shower are easily identified by their size and correlation in direction and time. The shower-detector plane (SDP) is shown by the fitted curve. The colors indicate arrival time of the signal light, with blue indicating the earliest and red the latest pulses. []{data-label="dpf2011_cchjui:fig004"}](dpf2011_cchjui_fig004.eps){width="120mm"}
As can be seen in figure \[dpf2011\_cchjui:fig004\], the pixels corresponding to the actual air shower are easily identified by their size as well as spatial and temporal correlation, and are marked in color. The pointing directions of these channels are then used to fit for a shhower-detector plane (SDP). Because of the distortion inherent in the Miller cylindrical projection used for the event display, the fitted SDP appears as a curve in figure \[dpf2011\_cchjui:fig004\]. The colors of event pixels indicate time progression: blue represents the earliest arrival times at the top of the event, and red represents the latest at the bottom. The event depicted was clearly a downward going air shower. Once the SDP is obtained, the trajectory of the air shower can be completely determined in one of two ways. For monocular observation, where only the measurement from a single fluorescence station is used, the shower axis can be determined by fitting the arrival time at each pixel to equation \[dpf2011\_cchjui:eqn\_timefit\]. $$t_{i}=t_{0}+\frac{R_{P}}{c}\tan{\left(\frac{\pi-\psi-\chi_{i}}{2}\right)}
\label{dpf2011_cchjui:eqn_timefit}$$ As illustrated in figure \[dpf2011\_cchjui:fig005\](a), $R_{P}$ is the impact parameter of the shower (nearest distance of approach of the shower axis to the FD), and $\psi$ is the angle made by the shower axis to the line of intersection between the SDP and the ground. The value $t_{0}$ physically corresponds to the time at which the shower passes the point of nearest approach. The output parameters from the timing fit are $t_{0}$, $R_{P}$, and $\psi$. The inputs are the measured times $t_{i}$ and the angles $\chi_{i}$ of the pixels involved in the event. As seen in figure \[dpf2011\_cchjui:fig005\](a), $\chi$ is the angle made between the PMT pointing direction (projected onto the SDP), and the ground, measured within the shower detector plane. Alternately, with two FD stations viewing the same event in stereoscopic mode, the shower trajectory can be determined from the intersection of the two SDPs. This stereo reconstruction method is illustrated in figure \[dpf2011\_cchjui:fig005\](b). Typically at energies in the ultra high energy (UHE) regime, monocular reconstruction gives $R_{P}$ and $\psi$ resolutions of about 10% and $5^{\circ}$, respectively, whereas the stereo reconstruction improves these to about 5% and $1^{\circ}$.
![ (a) Illustration of the monocular FD time fit to determine the shower trajectory. (b) Illustration of the intersecting plane method of finding the shower axis for stereo FD observations. []{data-label="dpf2011_cchjui:fig005"}](dpf2011_cchjui_fig005.eps){width="120mm"}
Figure \[dpf2011\_cchjui:fig006\](a) shows the timing fit described above for the event shown in figure \[dpf2011\_cchjui:fig004\]. The amount of curvature in the data determines the in-plane angle, $\psi$. For a given $\psi$, the overall slope of the data then determines the impact parameter, $R_{P}$. Having determined the shower trajectory, the pointing directions of the PMTs are then converted to slant depth. The signal is then fitted to a parametric function, usually the Gaisser-Hillas form [@dpf2011_cchjui:gh] for the shower size vs. depth, and includes scattered and direct Ĉerenkov light in addition to the fluorescence signal. The profile fit for this same event is shown in figure \[dpf2011\_cchjui:fig006\](b). The Energy is extracted from the overall area of the curve, and the depth of the shower maximum, $X_{max}$, is extracted from the fit. Over many showers, the $X_{max}$ values give a statistical measure of the composition of the primary particles.
![(a) Timing fit to determine $R_{P}$ and $\psi$. (b) Profile fit to determine the energy and $X_{max}$ for the shower shown originally in figure \[dpf2011\_cchjui:fig004\]. []{data-label="dpf2011_cchjui:fig006"}](dpf2011_cchjui_fig006.eps){width="120mm"}
The High Resolution Fly’s Eye experiment used two alternative monocular reconstruction techniques. Between 1992-1996, the HiRes prototype, in the tower configuration (14 telescopes viewing up to $70^{\circ}$ in elevation), operated in coincidence with the CASA/MIA arrays. The HiRes/MIA monocular reconstruction included the timing information from the MIA array. This combination became known as the hybrid reconstruction method, and yields $R_{P}$ and $\psi$ resolutions comparable to that of stereo reconstruction. The Telescope Array experiment is primarily designed for hybrid FD reconstruction. An example of the hybrid timing fit for a TA event is shown in figure \[dpf2011\_cchjui:fig007\].
![Timing fit for a hybrid event that includes times of the surface detectors. []{data-label="dpf2011_cchjui:fig007"}](dpf2011_cchjui_fig007.eps){width="50mm"}
Previously for the HiRes-1 site, where the telescopes only view up to 17$^{\circ}$, the observed tracks from air showers are too short for the timing fit alone to give reliable results. Instead, another variant of the monocular reconstruction was used that combined the timing and profile fits of figure \[dpf2011\_cchjui:fig006\]. This technique uses the form of the shower profile to constrain the range of geometries. The profile-constrained fit (PCF) gives resolutions that are comparable to monocular fit at the highest energies for the one-layer HiRes-1 detector but quickly degrades and becomes unusable below about $3\times{10}^{18}$ eV. The third fluorescence detector station on Middle Drum (MD) Mountain, located at the northern end of TA, was built with 14 refurbished telescopes from HiRes-1. This commonality between HiRes and TA allows us to compare the results of the two experiments directly. For this purpose, the initial analysis of the MD monocular FD data used exactly the same simulation and reconstruction codes as was used for HiRes-1, changing only the pointing geometry of the detectors, and lowering the trigger threshold in the simulation to reflect the reduced ambient background light.
TA Energy Spectrum Measurement
==============================
One of the early objectives for building the Telescope Array experiment was to resolve the discrepancy between the observation of the Greisen-Zatsepin-K’uzmin [@dpf2011_cchjui:gzk] cut-off in the UHE cosmic ray spectrum. Using the fluorescence technique alone, HiRes reported the first observation of the GZK cut-off in 2008 [@dpf2011_cchjui:hires_prl2008], whereas earlier measurements by AGASA, using a scintillation ground array alone reported a continuing spectrum [@dpf2011_cchjui:agasa_spectrum]. Figure \[dpf2011\_cchjui:fig008\] shows the monocular spectrum from the Middle Drum FD station from its first three years of observation. The HiRes monocular spectrum [@dpf2011_cchjui:hires_prl2008] is also shown in the figure. The MD station uses 14 refurbished telescopes from HiRes-1, the latter having provided most of the statistical significance for the GZK cut-off. The two sets of spectra are in excellent agreement both in the shape and overall normalization. The new TA result is also consistent with a flux suppression at the expected GZK threshold.
![ TA monocular spectrum from the Middle Drum FD station from its first three years of observation, overlaid with the monocular spectra from HiRes. The TA and HiRes spectra are in excellent agreement. The TA spectrum is also consistent with the presence of the GZK cut-off. []{data-label="dpf2011_cchjui:fig008"}](dpf2011_cchjui_fig008.eps){width="80mm"}
From a compilation of TA hybrid events seen by both the SD and the FD, the SD was seen to give a consistently higher energy. After the first year of observation, the SD energies was determined to be consistently 1.27 times higher than the FD energies. Figure \[dpf2011\_cchjui:fig009\](a) shows a histogram of the difference between the FD and SD energies, with the SD energies scaled down by 1.27. A scatter-plot of log FD energy vs log SD energy for the same events is shown in figure \[dpf2011\_cchjui:fig009\](b). The latter shows a linear relationship between the two energy measurements over the 1.5 decades of energy above $3\times{10}^{18}$ eV.
![ (a) A histogram of the difference between FD and SD energies for hybrid events above $3\times{10}^{18}$ eV. (b) A scatter-plot of the log FD energy vs. log SD energy for the same hybrid events. []{data-label="dpf2011_cchjui:fig009"}](dpf2011_cchjui_fig009.eps){width="100mm"}
Figure \[dpf2011\_cchjui:fig010\](a) shows the energy spectrum of UHE cosmic rays compiled from the first three years of TA surface array data. The energy of each event was rescaled by the factor of 1.27. This spectrum is overlaid with the monocular FD spectrum (previously shown in \[dpf2011\_cchjui:fig008\]). With the rescaling of energies alone, the SD spectrum obtained is in excellent agreement with the monocular FD spectrum, and in turn, with the HiRes spectrum both in normalization and in shape. Figure \[dpf2011\_cchjui:fig010\](b) shows a preliminary hybrid spectrum from the Middle Drum FD data overlaid with the SD spectrum. Again the two are in excellent agreement. We have divided the three TA spectra shown (MD FD monocular, SD, and MD FD hybrid) and the HiRes spectrum into three plots in order to avoid clutter. The conclusion we draw here is that with a 1.27 energy scaling factor for the SD, the TA SD and MD FD spectra are completely consistent with the HiRes results. Monocular and hybrid spectra from BR and LR, not shown here, are also consistent with those of the SD, and MD.
![(a) The TA surface detector spectrum with event energies scaled down by a factor of 1.27, overlaid with the monocular FD spectrum from Middle Drum. (b) The hybrid FD spectrum from Middle Drum overlaid with the energy-rescaled SD spectrum. []{data-label="dpf2011_cchjui:fig010"}](dpf2011_cchjui_fig010.eps){width="140mm"}
Composition and Anisotropy
==========================
Since 2009, there has been a discrepancy in the $X_{max}$-based composition results between the AUGER and HiRes collaborations. AUGER claims to see a trend toward heavier composition at above ${10}^{19}$ eV [@dpf2011_cchjui:auger_comp], whereas HiRes results are consistent with a predominantly proton composition [@dpf2011_cchjui:hires_comp]. Figure \[dpf2011\_cchjui:fig011\] shows the first TA composition result based on $X_{max}$ from stereo events. In Figure \[dpf2011\_cchjui:fig011\](a), the distribution of $X_{max}$ for TA stereo events is compared to those of iron and proton events simulated with CORSIKA using the QGSJET-II hadronic model. It is clear that in mean value and in width of the distribution, the TA results are consistent with a predominantly protonic composition.
![(a) Distribution of shower maximum depth ($X_{max}$) of TA stereo data compared to CORSIKA simulation for proton and iron, based on the QGSJET-II hadronic interaction model. (b) Plot of mean $X_{max}$ vs log energy for the same TA stereo data set. The accompanying curves show CORSIKA simulation results, including detector response and trigger selection effects for proton and iron with three different hadronic interaction models. []{data-label="dpf2011_cchjui:fig011"}](dpf2011_cchjui_fig011.eps){width="140mm"}
Figure \[dpf2011\_cchjui:fig011\](b) shows the plot of mean $X_{max}$ vs. log energy for the same stereo data set. The various curves show the predictions (folding in detector response and trigger selection) of CORSIKA simulations with different hadronic interaction models. The TA data, like that for HiRes, is again consistent with a predominately protonic composition, especially when compared to QGSJET models. Composition studies based on the width of the $X_{max}$ distributions, and on the width of the shower profiles as well as those using hybrid events are nearing completion.
The anisotropy searches in TA are based primarily on the SD data. After the first three years of observations, the data is entirely consistent with isotropy. We did check the TA data against the claim made by the AUGER collaboration in the 2007 Science article [@dpf2011_cchjui:auger_2007science], where 8 of 13 AUGER events above $5.7\times{10}^{19}$ eV were seen to be within 3.1$^{\circ}$ of Active galactic nuclei in the Veron-Cetty catalog [@dpf2011_cchjui:veron-cetty] with $z<0.018$. For the northern sky, the corresponding prediction for TA would have been 15 correlations out of 20 TA events seen above $5.7\times{10}^{19}$ eV, whereas an isotropic distribution predicts five accidental correlations. Of these 20 events, eight were seen to be in coincidence with AGNs. This result is not a particularly significant departure ($p=0.13$) from the null hypothesis. More TA data is needed for further anisotropy searches.
The Telescope Array experiment is supported by the Japan Society for the Promotion of Science through Grants-in-Aid for Scientific Research on Specially Promoted Research (21000002) “Extreme Phenomena in the Universe Explored by Highest Energy Cosmic Rays”, and the Inter-University Research Program of the Institute for Cosmic Ray Research; by the U.S. National Science Foundation awards PHY-0307098, PHY-0601915, PHY-0703893, PHY-0758342, and PHY-0848320 (Utah) and PHY-0649681 (Rutgers); by the National Research Foundation of Korea (2006-0050031, 2007-0056005, 2007-0093860, 2010-0011378, 2010-0028071, R32-10130); by the Russian Academy of Sciences, RFBR grants 10-02-01406a and 11-02-01528a (INR), IISN project No. 4.4509.10 and Belgian Science Policy under IUAP VI/11 (ULB). The foundations of Dr. Ezekiel R. and Edna Wattis Dumke, Willard L. Eccles and the George S. and Dolores Dore Eccles all helped with generous donations. The State of Utah supported the project through its Economic Development Board, and the University of Utah through the Office of the Vice President for Research. The experimental site became available through the cooperation of the Utah School and Institutional Trust Lands Administration (SITLA), U.S. Bureau of Land Management and the U.S. Air Force. We also wish to thank the people and the officials of Millard County, Utah, for their steadfast and warm support. We gratefully acknowledge the contributions from the technical staffs of our home institutions and the University of Utah Center for High Performance Computing (CHPC).
[99]{} Fukushima, M., Institute for Cosmic Ray Research Mid-term (2004 - 2009) Maintenance Plan Proposal Book “Cosmic Ray Telescope Project”. 2002, Tokyo University. http://www.telescopearray.org
N. Chiba [*et al.*]{}, Nucl. Instrum. Meth. [**A311**]{} (1992) 338; H. Ohoka, [*et al.*]{}, Nucl. Instrum. Meth. [**A385**]{} (1997) 268.
T. Abu-Zayyad [*et al.*]{}, Proc. 26th ICRC, 4, 349 (1999); J. H. Boyer [*et al.*]{}, Nucl. Inst. Meth. [**A482**]{}, 457 (2002).
R. M. Baltrusaitis [*et al.*]{}, Nucl Instrum. Meth. [**A20**]{}, 410 (1985).
J. Linsely, J. Phys. G: Nucl. Phys., [**12**]{}, 51-57, 1986; D. Ikeda [*et al.*]{} Astrophys. Space Sci. Trans., [**7**]{}, 257–263, (2011).
S. Yoshida, N. Hayashida, K. Honda [*et al.*]{} J. Phys. G: Nucl. Phys., [**20**]{}, 651–664, (1994).
T. Gaisser and A. M. Hillas, Proc. 15th ICRC [**8**]{}, p. 353, (1977).
K. Greisen, Phys. Rev. Lett. [**16**]{}, 48 (1968); T. Zatsepin and V. A Kuzmin, Pis’ma Zh. Eksp. Teor. Fiz., [**4**]{}, 114 (1966).
R. U. Abbasi [*et al.*]{} Phys. Rev. Lett. [**92**]{}, 151101 (2008).
M. Takeda [*et al.*]{}, Astropart. Phys. [**19**]{}, 447 (2003).
J. Abraham [*et al.*]{}, Phys. Rev. Lett. [**104**]{} 091101 (2010).
R. U. Abbasi [*et al.*]{}, Phys. Rev. Lett. [**104**]{} 161101 (2010).
The Auger Collaboration, Science, 318 p. 938, (2007).
M.-P. Cetty and P. Veron, Astronomy and Astrophysics, [**455**]{}(2), p. 773, (2006).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present MHD simulations exploring the launching, acceleration and collimation of jets and disk winds. The evolution of the disk structure is consistently taken into account. Extending our earlier studies, we now consider the self-generation of the magnetic field by an $\alpha^2\Omega$ mean-field dynamo. The disk magnetization remains on a rather low level, that helps to evolve the simulations for $T > 10,000$ dynamical time steps on a domain extending 1500 inner disk radii. We find a magnetic field of the inner disk similar to the commonly found open field structure, favoring magneto-centrifugal launching. The outer disk field is highly inclined and predominantly radial. Here, differential rotation induces a strong toroidal component that plays a key role in outflow launching. These outflows from the outer disk are slower, denser, and less collimated. If the dynamo action is not quenched, magnetic flux is continuously generated, diffuses outward the disk, and fills the entire disk. We have invented a toy model triggering a time-dependent mean-field dynamo. The duty cycles of this dynamo lead to episodic ejections on similar timescales. When the dynamo is suppressed as the magnetization falls below a critical value, the generation of the outflows and also accretion is inhibited. The general result is that we can steer episodic ejection and [*large-scale jet knots*]{} by a [*disk-intrinsic dynamo*]{} that is time-dependent and regenerates the jet-launching magnetic field.'
author:
- 'Deniss Stepanovs, Christian Fendt, Somayeh Sheikhnezami'
title: 'Modelling MHD accretion-ejection - episodic ejections of jets triggered by a mean-field disk dynamo'
---
Introduction
============
Astrophysical jets as highly collimated beams of high- velocity material and outflows of a small degree of collimation and lower speed are a ubiquitous phenomenon in astrophysical sources of rather different ranges in energy output and a physical extension. Previous calculations have shown that jets and winds could be produced by the interplay of large scale magnetic fields with the accretion disk .
Two main mechanisms compete in the acceleration of the material that is lifted from the disk into the outflow. In addition to the classical magneto-centrifugal acceleration mechanism proposed by @1982MNRAS.199..883B that is usually compared to a sling-shot mechanism of material along poloidally dominated magnetic field lines, acceleration may also be driven by a pressure gradient of the toroidal magnetic field, comparable to a mechanical spring mechanism. This mechanism has been studied extensively both analytically [@1994MNRAS.267..146L; @1995MNRAS.275..244L; @1996MNRAS.279..389L] and numerically [@1995ApJ...439L..39U]. If the toroidal magnetic field is generated contineously, a inflation of the poloidal field structure results and the material enclosed by the poloidal magnetic loops is accelerated in vertical direction. This kind of jet structure is known as a [*tower jet*]{}, or Poynting-dominated jet.
The jet launching and collimation problem[^1] is usually being addressed numerically applying a large scale initial poloidal magnetic field. This holds in particular for simulations considering the acceleration and collimation process only and assuming the underlying disk as a boundary condition .
Also simulations treating the launching mechanism, i.e. simulations of the accretion-ejection structure that include the time evolution of the disk dynamics, so far have assumed a global large-scale magnetic field as initial condition .
These studies have provided a deep insight in the launching mechanism, i.e. the connection between the outflow and the underlying disk. It is clear today that the magnetic field plays [*the*]{} crucial role in lifting the matter out of the disk and accelerating it to high velocity. By knowing the disk magnetization one can refer many details of the launched outflow, namely its energetics and the ejection efficiency (see , hereafter Paper I; @paper-2).
It is still an open issue what the exact structure and the strength of the magnetic field in the disk is, and where its origin is. Besides a central stellar magnetic field or advection of magnetic field from the ambient medium, a turbulent dynamo can be a major source of the disk magnetic field [@1981MNRAS.195..881P; @1981MNRAS.195..897P; @1995ApJ...446..741B]. In order to study the disk dynamo in the context of outflow launching, only a few numerical experiments were performed in which the magnetic field was generated [*ab initio*]{} . These authors were first to show how accretion disks start producing the outflow if the magnetic field is amplified by the dynamo to about its equipartition value.
A further motivation for considering a disk dynamo for jet-launching is seemingly the time-dependent ejection of the jet material. For protostellar jets the typical timescales for ejection derived from the observed knot separation and jet velocity are in the range of 10-100 years. The typical timescale of the jet-launching area is, however, about 10-20 days, that is the Keplerian period close to the inner disk radius. A time-variable dynamo may be responsible for changing the jet-launching conditions on longer timescales. We may here refer to the dynamo cycle of the Solar magnetic field that is longer than the Sun’s rotation period[^2].
Our main concern in this paper is structure and time evolution of the dynamo-generated magnetic field, the launching of outflows by such a disk self-generated magnetic field, and the interrelation between the dynamo and the episodic ejection of jets, possibly leading to the co-called jet knots. Such a study has not yet been presented in the literature.
Disk dynamos were discussed in the literature concerning episodic accretion and ejection events in dwarf novae, and also as a possible physical process to generate MHD instabilities and turbulence, allowing for angular momentum transfer and accretion. @1996ApJ...457..332A discussed a disk dynamo mechanism in accretion disks as a cause for dwarf novae eruption, similar to what could probably happen in the jet-launching disks. @1992MNRAS.259..604T discuss a disk dynamo action in order to physically produce the magnetic disk viscosity. However, @1998ApJ...492L..75G showed that for low Reynolds numbers the MHD disk turbulence and angular momentum transport dies out, possibly leading to episodic accretion in dwarf novae.
@1995MNRAS.276.1179R discuss a model of a disk dynamo driven by magnetic buoyancy which does not directly involve a disk turbulence. This model was re-visited by finding that accretion could in fact be established by a Parker instability-driven dynamo. According to accretion could be based on the interaction of Parker and magneto-rotational instabilities. In this scenario the vertical component of the magnetic field is generated by Parker instability (PI) and serves as a source for magnetorotational instability (MRI).
Applying a mean-field $\alpha^2\Omega$ dynamo [@1980opp..bookR....K] we present the step by step evolution of the magnetic field. In our approach turbulence is being addressed in the mean field approach, and is not self-consistently generated (e.g. by the MRI).
We also study episodic jet-launching scenarios by means of a simple toy model in which we artificially switch on/off the dynamo action. We discuss whether similar processes in which the dynamo does change its strength may lead to episodic jet ejection and the jet knots. We also discuss in detail how the magnetic field can be regenerated by re-establishing the dynamo action.
Our paper is organized as follows. In Section 2 we briefly describe our numerical setup. For a more complete discussion we refer to Paper I. In particular, we discuss the implementation of the mean-field dynamo equations and the model approach for the magnetic diffusivity and the dynamo-$\alpha$. In Section 3 we present our reference dynamo simulation where the jet-launching magnetic field structure is dynamo-generated from a weak seed field. We discuss the difference between dynamo and non-dynamo simulations. In Section 4 we present simulations during which the disk dynamo is switched on and off repeatedly, leading to episodic ejection of the disk material into the collimated outflow. We summarize our paper in Section 5.
Model Approach {#sec:model}
==============
For our numerical simulations, we apply the MHD code PLUTO[^3] [@2007ApJS..170..228M], solving the time-dependent, resistive MHD equations on a spherical grid. Our simulations have been performed in 2D axisymmetry, applying spherical coordinates $(R,\theta)$. We refer to $(r,z)$ as cylindrical coordinates.
We have specified the equations considered in detail in Paper I. Here we stress in particular the induction equation that we have modified in the code according to the mean-field dynamo formalism [@1980opp..bookR....K], $$\label{eq:induction}
\frac{\partial {{ \textbf{\emph{B}} }}}{\partial t} =
\nabla\times ({{ \textbf{\emph{V}} }}\times {{ \textbf{\emph{B}} }}+ { {\overline{\overline{\alpha}}}_{\rm dyn} }{{ \textbf{\emph{B}} }}- {\overline{\overline{\eta}}}{{ \textbf{\emph{J}} }}),$$ where the tensor ${ {\overline{\overline{\alpha}}}_{\rm dyn} }$ describes the $\alpha$-effect of the mean-field dynamo, and the tensor ${\overline{\overline{\eta}}}$ the magnetic diffusivity (see below).
As no physical scales are introduced in the equations we solve, the results of simulations are presented in non-dimensional units. Lengths are given in units of $R_0$, corresponding to inner disk radius. Velocities are given in units of $V_{\rm K,0}$, corresponding to the Keplerian speed at $R_0$. Thus times are given in $T_0 \equiv R_0 / V_{\rm K,0}$ units. Note, that ${ 2 \pi T_0}$ corresponds to one rotation at the innermost orbit. Densities are given in units of $\rho_0$, corresponding to $R_0$. Pressure is measured in $P_{0} = \epsilon^2 \rho_{0} V_{0}^2$, where $\epsilon$ is the ratio of the initial isothermal sound speed to Keplerian speed taken at the disk midplane. All our simulations were performed with $\epsilon=0.1$.
We normalize all variables, namely $R, \rho, V, B$, to their values at the inner disk radius ${R_{\rm 0}}$. We thus may apply our scale-free simulations to a variety of jet sources. For the typical astrophysical scaling of the code units we refer to Table 1 of Paper I.
We apply a numerical grid with equidistant spacing in $\theta$-direction, but stretched cell sizes in radial direction, considering $\Delta R = R \Delta\theta$. Our computational domain of a size $R=[1, 1500 R_0],\theta=[0,\pi/2]$ is discretized with $(N_R \times N_\theta)$ grid cells. We use a general resolution of $N_\theta = 128$. In order to cover a factor 1500 in radius, we apply $N_R = 600$. This gives a resolution of 16 cells per disk height ($2 \epsilon$) in the general case. We have also performed a resolution study with 1.5 times higher (lower) resolution, thus using $900 \times 192$ ($450 \times 64$) cells for the domain, or 26 (11) cells per disk height.
Initial Conditions
------------------
As a measure for the magnetic field strength, we use the magnetization defined as the ratio between magnetic and thermal pressure, $$\mu = \frac{B^2}{2P}$$ We have used different prescriptions for the magnetization, however, in all cases the local magnetic field pressure $B^2/2$ is related to the gas pressure $P$ at the midplane.
All dynamo simulations we perform start from a very weak initial magnetization ${\mu_{\rm init}}= 10^{-5}$. Therefore the initial structure of the accretion disk can be obtained as the solution to the steady-state force equilibrium equation, neglecting the contribution by the Lorentz force, $$\nabla P +\rho \nabla {\Phi_{\rm g}}- \frac{1}{R} \rho {V_{\rm \phi}}^2
({{ \textbf{\emph{e}} }}_R \sin\theta + {{ \textbf{\emph{e}} }}_\theta \cos\theta) = 0.$$ Assuming a self-similar disk structure this equation can be solved analytically.
All our simulations are initialized with a purely radial magnetic field, confined within the disk and defined via the vector potential $\vec B = \nabla \times A \vec{e_\phi}$, and $$A = B_{\rm p,0} r^{-1} e^{-8{z/H}^2}.$$ The parameter $B_{\rm p,0} = \epsilon \sqrt{2 \mu_{\rm init} } $ denotes the strength of the initial magnetic field, while $\epsilon = 0.1$ is ratio of isothermal sound[^4] to Keplerian speed. Although this magnetic field distribution may be considered as somewhat artificial, we found that it provides a smooth evolution during the initial phase. We have also performed simulations starting from a purely toroidal magnetic field as the initial condition, leading to very similar results.
In contrast, purely vertical magnetic field would generate strong currents at the disk surface region because of the strong initial shear between the rotating disk and the non-rotating corona. This would greatly impact the initial evolution of the accretion-ejection structure. As long as the [*initial*]{} magnetization ${\mu_{\rm init}}$ is low, it does not play a substantial role for the initial disk evolution. This is the result of the exponential evolution of the magnetic field amplification by the dynamo.
Boundary Conditions
-------------------
[lllllllll]{} & $\rho$ & P & ${V_{R}}$ & ${V_{\rm \theta}}$ & ${V_{\rm \phi}}$ & ${B_{\rm R}}$ &${B_{ \theta}}$ & ${B_{ \phi}}$\
Inner disk &$\sim r^{-3/2}$&$\sim r^{-5/2}$&$\sim r^{-1/2}, \leq 0$&0&$\sim r^{-1/2}$ &Slope&Slope&$\sim r^{-1}$\
Inner corona &$\sim r^{-3/2}$&$\sim r^{-5/2}$&$0.2 cos(\varphi)$&$0.2 sin(\varphi)$&$\sim r^{-1/2}$ &0&div B =0&$0$\
Outer disk &$\sim r^{-3/2}$&$\sim r^{-5/2}$ &Outflow, $\le 0$& Outflow& Outflow&div B =0&Outflow&$\sim r^{-1}$\
Outer corona &$\sim r^{-3/2}$&$\sim r^{-5/2}$ &Outflow, $\geq 0$& Outflow& Outflow&div B =0&Outflow&$\sim r^{-1}$\
Axis & Sym & Sym & Sym & Anti-sym & Anti-sym & Sym & Anti-sym & Anti-sym\
Equator& Sym & Sym & Sym & Anti-sym & Sym & Anti-sym & Sym & Anti-sym\
\[tbl:bc\]
The boundary conditions are adapted from Paper I. The only change was made for the coronal region of the inner boundary. Here, we do not allow any magnetic flux to penetrate the inner coronal region and not only set ${B_{ \phi}}= 0$, but also $B_R = 0$. Since the magnetic field vanishes in that area, we therefore prescribe a purely radial profile of the inflow into the corona (in contrast to an inflow aligned to the magnetic field considered previously), keeping the same inflow velocity ${V_{R}}= 0.2$. We summarize all boundary conditions in Table \[tbl:bc\]
Since the magnetic field is suppressed in the inner coronal region, the shear in the area between the coronal region and the disk boundary can develop strong electric currents. This makes the region between the axis and the jet subject to small-scale perturbations, especially in the runs with high resolution. On the other hand, the jet-launching area of the inner disk always shows a smooth, stable, and non-fluctuating evolution.
Magnetic Diffusivity
--------------------
In Paper I we studied in detail models applying both a [*standard diffusivity*]{} and a so-called [*strong diffusivity*]{}. We have shown numerically that the standard diffusivity model is prone to the accretion instability. In the present paper, studying the dynamo action, we also performed simulations using both models. These models qualitatively share many similar features. Therefore, we will present simulations applying the strong diffusivity model, however, commenting on differences between diffusivity model.
The main mediator in the magnetic diffusivity models is the magnetization of the underlying disk. In case of simulations with a substantially strong initial magnetic field (see Paper I), the disk magnetization is set by the magnetic field at the disk midplane.
Since in the dynamo simulations presented here the initial magnetic field does not intersect with the midplane[^5], and may also remain low for quite some time, the parametrization of the diffusivity model with the magnetization had to be revised. We keep the same notation as in Paper I for the strong diffusivity model, $$\label{eq:strongdiff}
{{\alpha}_{\rm ssm}}= {{\alpha}_{\rm m}}\sqrt{2\mu_0} \left(\frac{{ \mu_{\rm D} }}{\mu_0}\right)^2,$$ where the [*disk magnetization*]{}, $$\label{eq:mudisk}
{ \mu_{\rm D} }= \frac{<B_{\rm D}>^2}{2P},$$ is defined by means of the [*average*]{} total magnetic field $<B_{\rm D}>$ for a certain radius within the disk (up to $H$), normalized to the midplane pressure. A non-zero magnetic diffusivity allows for reconnection and diffusion of the magnetic field across the midplane. An assumption that the magnetic diffusivity is dependent on the total magnetic field strength is consistent with the fact that the MRI is excited by both toroidal and poloidal magnetic field components [@2013EAS....62...95F].
The Dynamo Model
----------------
We apply a standard mean-field $\alpha^2\Omega$ dynamo formalism [@1980opp..bookR....K], where $\alpha$ represents the dynamo effect by turbulence, and $\Omega$ stands for the differential rotation of the plasma. According to the mean-field dynamo theory, an extra electromotive force term ${ {\overline{\overline{\alpha}}}_{\rm dyn} }{{ \textbf{\emph{B}} }}$ enters the induction equation (Equation \[eq:induction\]) and is responsible for the generation of the magnetic field. In general, ${ {\overline{\overline{\alpha}}}_{\rm dyn} }$ is a tensor, however non-diagonal components are less relevant for the dynamo process , in particular when a moderately strong magnetic field is present [@2012SSRv..169..123B]. Therefore we neglect all non-diagonal components and set the diagonal values equal to one parameter, ${{\alpha}_{\rm dyn}}$. The sign of ${{\alpha}_{\rm dyn}}$ as well as its number value in real disks has been widely debated . It has been shown that in order to get a dipolar structure of the mean magnetic field (as opposed to a quadrupolar structure) a negative sign of alpha should be chosen .
Following a dimensional analysis, we may scale ${{\alpha}_{\rm dyn}}$ as the Keplerian velocity, thus applying $${{\alpha}_{\rm dyn}}= { \alpha_{\rm D} }\,r^{-1/2} { F_{\rm \alpha} }(z),$$ where the vertical profile of $\alpha$-effect is defined by $${ F_{\rm \alpha} }(z) = \left\{
\begin{array}{ll}
\sin( \pi \frac{z}{H} ) & z\leq H \\
0 & z > H \\
\end{array}
\right.$$ Here, $H$ denotes the disk scale height and is approximated as constant in time. The profile ${ F_{\rm \alpha} }(z)$ restricts the $\alpha$-effect to the disk area.
It is generally believed that in case of the strong magnetic field the dynamo is quenched [@2005PhR...417....1B]. The main reason is that a strong global magnetic field suppresses the turbulence, and thus the turblulent dynamo. The quenching is commonly applied by multiplying the ${{\alpha}_{\rm dyn}}$-term by a quenching function, $$\label{eq:quen}
Q = \frac{1}{1 + 2\mu_{\rm x}},$$ where $\mu_{\rm x}$, in contrast to the ${ \mu_{\rm D} }$ is the [*local*]{} magnetization. In order to be consistent with and directly affect the resulting magnetic field, we parameterize $\mu_{\rm x} = q_{\mu}{ \mu_{\rm D} }$. By setting different $q_{\mu}$ we can limit the magnetic field growth to a certain value. Typically, we choose a rather high $q_{\mu}$ in order to quench the dynamo already for low magnetizations.
However, there is another possibility of limiting the magnetic field strength. We find that ${ \mu_{\rm 0} }$ in the strong diffusivity model (Equation \[eq:strongdiff\]) is in fact a good measure for the resulting actual disk magnetization. This comes from the functional form of the diffusivity profile - any further growth of the disk magnetization has a strong feedback of the diffusivity (see Paper I).
Both direct dynamo quenching and indirect limiting the magnetization by applying the strong diffusivity model lead to the saturation of the magnetic field. The difference between this approaches is that in case of leaving the dynamo working in the disk, the magnetic flux is being continuously generated and the disk is being filled with the magnetic field. If the standard diffusivity model is applied, then the dynamo quenching is the only mechanism to stop further magnetic field amplification. Therefore, since we apply the strong diffusivity model in the simulations we present, these simulations are run without dynamo quenching.
The expected dynamo number for accretion disks is given by $$|{D}| = |{C}_{\alpha} {C}_{\omega}|\lesssim \frac{3}{2} {{{\alpha}_{\rm ssm}}}^{-2},$$ where ${C}_{\alpha} = ({{\alpha}_0}H / \eta_0)$ and ${C}_{\omega} = (|\Delta\Omega|H^2 / \eta_0)$ represent the strength of $\alpha$-effect and shear $d\Omega/dr$, respectively. Since our main concern is the resulting jet-launching magnetic field, we choose the maximum dynamo number in order to generate the magnetic field structure as rapid as possible. The maximum dynamo number is provided by ${ \alpha_{\rm D} }= - 0.1$. Note, that the dynamo number $D$ is strongly dependent on ${{\alpha}_{\rm ssm}}$.
A Reference Dynamo Simulation
=============================
In this section we discuss simulations applying the dynamo model and the resulting configuration of disk-jet system.
We will refer to our reference dynamo simulation as to the simulation with the parameters ${ \alpha_{\rm D} }= -0.1$, ${{\alpha}_{\rm m}}= 1.65$, ${ \mu_{\rm 0} }= 0.01$. Figure \[fig:dyn\_tevol\] shows the time evolution of our reference dynamo simulation, that can be seen as typical for our model setup. The simulation starts from a weak (${ \mu_{\rm seed} }= 10^{-5}$), purely radial magnetic field, confined within the disk. Once the simulation is started, the toroidal component of the magnetic field is being continuously generated from the radial magnetic field simply by stretching. For the poloidal magnetic field component, the only generation mechanism is the dynamo effect that induces the poloidal component from the toroidal one.
{width="18cm"}
Since the toroidal magnetic field is antisymmetric to equator, the poloidal magnetic field loops that are generated by the dynamo first, do [*not*]{} cross the equator. When they evolve in time magnetic reconnection is enforced by the equatorial plane boundary condition and the magnetic diffusivity in the disk. As a consequence the magnetic loops (in the upper and lower hemisphere) merge and do traverse the equatorial plane. Since our diffusivity model depends on the average magnetic field in the disk there is always a substantial diffusivity present in the disk.
As described by , the toroidal component of the magnetic field is continuously amplified until it reaches the buoyancy limit and starts moving upward, away from the disk. The upward motion changes the structure of the magnetic field lines from a predominantly radial into a vertical direction. When the magnetic field is sufficiently inclined, magneto-centrifugal launching [@1982MNRAS.199..883B] can strongly accelerated the plasma on these field lines. The outflowing gas carries with it the toroidal magnetic field generated in the disk, thus setting a limit for the toroidal magnetic field strength in the disk (see Paper I).
All dynamo simulations evolve into three distinct domains in the accretion-ejection structure. Starting from the innermost disk, in the first region the magnetic field has the typical structure of field lines inclined with respect to the disk surface. Although magnetic field generation by the dynamo can still take place (if the $\alpha$-effect is not quenched), the magnetization in that region has become sufficiently high in order to operate the standard magneto-centrifugal jet driving. The second region is where the poloidal magnetic field is mostly radial. Here the velocity shear in the disk creates a strong toroidal component of the magnetic field. In this area, the outflow is launched mainly by the buoyancy of the toroidal magnetic field.
The third region is the outer disk, where the magnetic field is rather weak, with somewhat irregular structure. Here, the magnetic flux has originated from the same magnetic loops as the innermost field, but because of longer distance from the inner disk the magnetic field strength has become much lower. Note, that the magnetic field in the outer disk has the opposite polarity. In summary, the overall structure of the magnetic field has a strong gradient that leads to strong outward diffusion of magnetic flux. Due to this diffusion, the whole disk is filled by a non-zero net magnetic field (${B_{\rm P}}|_{\rm midplane}\neq 0$).
Dynamo Effect Versus Magnetic Diffusion
---------------------------------------
As discussed in detail in Paper I, the evolution of the disk-jet system is mainly set by the two opposite processes, the diffusion and advection of the magnetic field. It is more complicated to reach this balance. In case of dynamo simulations a third process contributes to the induction equation, the dynamo, and it is more complex to reach to an equilibrium situation. Nevertheless, the main effects of these processes can be disentangled. The dynamo term in the induction equation manifests itself by generating loops of the poloidal magnetic field. Because of the magnetic flux conservation along the magnetic loops, the magnetization in the inner disk (inner footpoint of the loop) is always higher than in the outer disk (outer footpoint of the loop). Therefore a strong gradient of the magnetic field develops, that evolves primarily following the magnetic diffusivity model.
By smoothing out the gradient of the magnetic field, the diffusivity plays a key role in the overall evolution of the magnetic field - first, it diffuses the magnetic field outward, thus filling the outer disk with the magnetic flux, second, at the same time the diffusivity destroys some flux within the magnetic loop by reconnection.
In general, if the dynamo is not sufficiently strong (in case of low dynamo numbers), the generated magnetic field will quickly decay (being diffused) and the magnetization necessary for jet-launching will not be reached.
Structure of the Tower Jet
--------------------------
![Snapshot at $T=1,000$ of physically different regions of the disk-jet structure. Shown is the mass density (in logarithmic scale) and streamlines of the poloidal velocity (black lines with arrows). The red line marks the magnetic field line that is rooted at the innermost disk area along the midplane. The upper white dashed line separates the area where $V_p || B_p$ from the rest of the disk. The accretion and ejection areas are separated by a white line indicating $V_r = 0$, and a black line indicating $F_\phi = 0$, respectively. The lower white dashed line separates the actual accretion area where $V_r >> V_\theta$ from the rest of the disk.[]{data-label="fig:dyn_struct"}](\figurepath/dyn_struct-eps-converted-to.pdf){width="9cm"}
{width="18cm"} {width="18cm"}
Figure \[fig:dyn\_struct\] shows the snapshot of the initial evolution at $T=1,000$.
What can be immediately seen is that the disk structure, namely the structure of the velocity field and the magnetic field (see Figure \[fig:dyn\_tevol\]) is completely different from the non-dynamo simulations [@paper-1; @paper-2]. We find two distinct regions in which the poloidal component of the magnetic field is inclined slightly (for $r<10$) or strongly (for $r>10$) with respect to the disk surface. While the magnetic field of the inner disk favors a standard magneto-centrifugal launching, in this section, we primarily concentrate mainly on the outer disk.
In this region, the strong toroidal magnetic field is induced by the differential rotation of both the inclined poloidal magnetic field and by the magnetic loops that are rooted at radially different footpoints along the disk. The mechanism we observe is similar to the well-known tower jet [@1994MNRAS.267..146L; @1995MNRAS.275..244L; @1996MNRAS.279..389L; @1995ApJ...439L..39U]. The increasing toroidal magnetic field pressure leads to an inflation of the poloidal magnetic loops and the material enclosed by that poloidal loops is accelerated in vertical direction. This structure – typical for a tower jet– is clearly seen in the extended loops in Figure \[fig:jet\_cyl\]. In Figure \[fig:jet\_cyl\], we further see that the fast jet, the one that is launched from the inner region of the disk, becomes collimated already close to the disk. On the contrary, for the tower jet - the expanding loop structures - it takes a while to collimate.
Almost everywhere in the jet region, the toroidal magnetic field dominates the poloidal magnetic field. Numerical simulations have shown that such structures, for example, naturally result from the interaction between a stellar dipole magnetic field penetrating the accretion disk . Note, however, that the magnetic loops presented in the disk are generated by the disk dynamo. The tower jet origins from the magnetic loop structure, and as the simulation evolves, that magnetic loop structure, and thus the base of the ”tower”, constantly moves outward. Around the magnetic loop structure ($r \approx 10$), we find that it is the the buoyancy force of the toroidal magnetic field that is the main force responsible for the lifting of the disk material into the outflow. Starting from the disk surface, defined as a surface of zero radial velocity (${V_{R}}= 0$), the matter is further accelerated by the pressure gradient of the toroidal magnetic field. The latter is, in fact, consistent with the simulations by @1995ApJ...439L..39U. The early evolution of the disk-jet system (Figure \[fig:jet\_cyl\] and Figure \[fig:dyn\_tevol\] at $T=200$) clearly show the similarity to the magnetic towers.
We typically find that the launching region, that was defined in Paper I as the region where the velocity of the plasma changes from being perpendicular to the magnetic field to almost parallel to the magnetic field, is broader in dynamo simulations than in the non-dynamo simulations. Also, the disk surface, where the radial velocity changes sign by definition, is located at higher altitudes, although the thermal disk scale height is still about constant in time and about its initial value.
Outflow Launching: Accretion-ejection {#sec:olac}
-------------------------------------
The magnetic field of the inner disk that is established in the simulations is similar to the usual structure favoring the magneto-centrifugal launching of the outflow. This type of jet formation has been previously found and discussed by many authors as well in our Paper I.
Therefore, in this section we concentrate on the (outer) disk region where the magnetic field has evolved into a structure completely different from the previous simulations. Namely into the structure with the poloidal magnetic field being predominantly radial, and very strong toroidal component. In this part of the disk, it is the toroidal magnetic field that plays the key role in the launching (see Fig.\[fig:dyn\_btbp\]).
![Distribution of the magnetic field components at $T=1000$. Shown is the ratio of the toroidal to the poloidal magnetic field (colors, [*logarithmic scale*]{}), the poloidal magnetic field lines (thin black lines); the sonic surface (red line); and the locations where the Lorentz force components change sign, $F_\phi = 0$ (white line) and $F_\theta = 0$ (blue line). Further denoted are the locations where i) the gas pressure force is equal to the Lorentz force both projected parallel to the magnetic field (black dashed line), and ii) where the gas pressure force is equal to the Lorentz force both projected parallel to the velocity field (green line). []{data-label="fig:dyn_btbp"}](\figurepath/dyn_btbp-eps-converted-to.pdf){width="9cm"}
As the accretion-ejection process is governed by the magnetic torques, these torques need to be discussed in detail. The white line in Figure \[fig:dyn\_btbp\] marks the region where the magnetic torque changes sign. The torque is negative in the inner disk (inside the white line), where the angular momentum extraction from disk to outflow takes place, and positive in the disk corona, leading to the acceleration of the outflow material. In the region that is dominated by the magnetic loops at radii of $R \simeq 15$ we find that the torque is purely positive, thus playing a major role in the acceleration of the plasma. The blue line in Figure \[fig:dyn\_btbp\] separates two regions, where i) the magnetic forces accelerate the matter in the direction of the outflow ($F_\theta >0$, above the line), and ii) where the magnetic forces pinch the disk (below the line). In the disk area below this line the main force lifting the matter into the outflow is the thermal pressure. The lines that mark the area where the pressure force is equal to the Lorentz force projected parallel to the magnetic field (black dashed line) and parallel to the velocity field (green line) are also shifted closer toward the disk midplane.
In the area above the loop-like structure the toroidal magnetic field dominates the poloidal field. In this region acceleration is mainly governed by the toroidal magnetic field pressure gradient.
It is worth to note that such a configuration does not reach a steady state. This is already indicated by the misalignment between the magnetic field lines and velocity field. Furthermore, the blue lines in Figure \[fig:dyn\_struct\] denote the launching area (see Paper I) where the velocity field changes from a direction perpendicular to parallel to magnetic field lines. The longer the simulation evolves, the larger the area that reaches a steady state. In other words, the non-steady loop structure is moving outward along the disk.
Dynamical Profiles of a Dynamo-disk driving jets {#sub:profiles}
------------------------------------------------
Here we discuss the overall disk structure of the reference dynamo simulation. Figure \[fig:dyn\_profiles\] presents the radial profiles of a number of MHD variables measured at the disk midplane, and the fits to them by power-laws. The slight deviation between these lines shows how the disk structure changes after a long time evolution. At time $T=10,000$ we find distinct power-laws for the profiles along the disk for radii up to $R \leq 40$. This is the radius that marks a steady state area from the rest of the disk, where the magnetic field is continuously generated. This is most easy to infer from the profile of the poloidal magnetic field profile, that starts deviating from the approximated power-law at $R=40$.
![Physical quantities along the disk midplane for the reference dynamo simulation at $T=10,000$. Colored lines correspond to different physical quantities, density $\rho$, sound speed $C_{\rm s}$, rotational velocity $V_{\phi}$, radial velocity $V_{R}$, and the magnetic field component $B_{\theta}$. The thick dashed lines show the corresponding approximation by a power-law. The thin dashed lines show the initial power-law distribution, slightly offset from the actual distribution at $T=10,000$. []{data-label="fig:dyn_profiles"}](\figurepath/res_def_profiles-eps-converted-to.pdf){width="8.5cm"}
In order to better compare the analytical fits to the radial profiles resulting from the dynamo simulation to those without the dynamo (Paper I), we plotted the fits with the same power-law indices as for Paper I. We see that the actual profiles (for example, the density) show only a tiny deviation.
At time $T= 10,000$, we find the following numerical values for the power-law coefficients $\beta_X$ for different variables at the midplane $X(r, \theta = \pi/2) \sim r^{\beta_X}$. The disk rotation remains Keplerian over the whole time evolution, therefore $\beta_{{V_{\rm \phi}}} = -1/2$. The radial profiles for density and thermal pressure slightly change from their initial distributions. The power-law index of the density changes from $\beta_\rho = -3/2$ to about $\beta_\rho = -4/3$, while for the pressure it changes from $\beta_{\rm P}= -5/2$ to about $\beta_{\rm P}= -20/9$. We find $\beta_{{V_{R}}} = -2/5$ for the accretion velocity, and $\beta_{{B_{ \theta}}} = -5/4$ for the poloidal magnetic field. The accretion velocity remains subsonic everywhere in the disk with an accretion Mach number of ${M_{R}}\equiv V_R / {C_{\rm s}}\simeq 0.08$. As expected, we find strong fluctuations in the area where the dynamo is active and field magnetic generation is ongoing.
Following and considering the mass accretion rate ${\dot{M}_{\rm acc}}\sim R^2 \rho {V_{R}}$, it is easy to derive the ejection index [ $\xi = 2+ \beta_\rho +\beta_{{V_{R}}}$[^6]]{}, that is a measure of the efficiency of the outflow. We find $\xi = 0.25$, about the same value as for the non-dynamo simulation (see Paper I).
In this section we have demonstrated that the radial profiles for disk dynamics along the midplane are very similar for the simulations including dynamo action and those without dynamo (in Paper I). We find several reasons for explanation. First, in the case of moderately weak magnetic field, the disk dynamics is primarily governed by the hydrodynamical quantities, but not so much the magnetic field. The power-law nature of the Keplerian rotation dominates the dynamics and forces the other hydrodynamical profiles into a power-law distributions as well. Second, the magnetic field strength resulting from the disk evolution is of the same order for both approaches, $\mu \approx 0.01$, that also can lead to profiles with to the same distribution.
Dynamo Versus Non-dynamo Simulations
------------------------------------
Here we discuss the major differences between the simulations with and without the mean-field $\alpha^2\Omega$ dynamo. As pointed out in the previous sections, the major difference between dynamo and non-dynamo simulations is the structure of the magnetic field and not so much the accretion disk hydrodynamics. The dynamo generates the magnetic field that is continuously spreading over the whole disk. In the early stage of the disk evolution, this makes a substantial difference. Later, when the inner part of the disk has reached a steady state, this part of the disk looks very much the same except a few details. One difference can be found concerning the disk wind close to the disk surface. In dynamo simulation the sonic surface and the Alfvén surface are located 20%-30% further up into the outflow. However, the magnetic lever arm (radius of the Alfvén point) is about the same. This is the result of the vertical component of the magnetic field being stronger, respectively the lower inclination of the magnetic field with respect to the disk surface. (more inclined toward the disk surface). The launching area, namely the area where the velocity field changes from being almost perpendicular to the magnetic field into a direction parallel to the field (see Paper I), is now wider, while the disk surface stays at about the same level. Note that because of the evolving loop-like structure of the magnetic field, the field inclination with respect to the disk or the sonic surface does change in time and space - except the inner disk where a steady-state has established.
In order to study the jet properties with respect to the actual disk magnetization, we have performed several simulations, varying the ${ \mu_{\rm 0} }$ parameter in the definition of the diffusivity (Equation \[eq:strongdiff\]). This parameter indirectly governs the resulting disk magnetization, as discussed above and also in Paper I. By running simulations with different $\mu_0$, we were able to probe the resulting actual magnetization of the poloidal component in the inner disk over a range ${ \mu_{\rm act} }= 0.01-0.05$. As shown in Paper I, it is in this range of magnetization where a change in the dominant launching mechanism takes place.
In comparison, we find that in our dynamo simulations the disk and jet properties do not really differ for different actual magnetization, and we cannot disentangle different launching mechanisms in the dynamo simulations. In contrast, the disk quantities as well as the jet integrals (see Paper I) behave rather similar in this range of magnetization for both simulations.
In order to disentangle the causes for this similar evolution, we may recall two points. The main reason why we can distinguish two different mechanisms in the non-dynamo simulation is the ability to generate a strong magnetic shear with sufficiently weak poloidal magnetic field. This is possible because the diffusivity in the standard model depends only on the poloidal magnetic field. Thus, for a weak poloidal field, the diffusivity is also small, that helps to sustain the strong magnetic shear. In contrast, in the current study, the magnetic diffusivity depends on the average [*total*]{} magnetic field in the disk. Therefore, a strong magnetic shear (a stronger toroidal magnetic field) directly increases the diffusivity, and, as a consequence, limits the magnetic shear.
This more subtle interrelation between the simulation parameters and physical processes again emphasizes the impact of the magnetic diffusivity model applied.
Episodic Jet Ejection Triggered by a Time-variable Disk Dynamo
==============================================================
In this section we present simulation results of a toy model applying a time-dependent dynamo action. Our motivation is the following. The dynamo is intrinsically a stochastic phenomenon that in real accretion disks can be subject to strong fluctuations. Some accretion disks may exist in which the dynamo action is suppressed, while in other disks if certain conditions are met the dynamo can start to operate. Also, dynamo quenching mechanism can stop an already working dynamo, and thus lead to a reconfiguration of the disk-jet system. Here we refer to the well-known solar cycle as an example. It is believed that a solar dynamo is responsible for the reconfiguration of the magnetic field of the Sun. The strong toroidal field component reveals itself as sunspots with a cyclic appearance. This solar periodicity can be understood as triggered by a constantly operating dynamo (or maybe different types of dynamo). Another interesting feature of the solar activity in this respect is that it exhibits long-term minima [@1976Sci...192.1189E], during which there were no sunspots observed at all. It is believed that during these minima the dynamo action is either strongly suppressed or completely switched off.
For our accretion-ejection simulations we follow a preliminary approach and apply a simple toy model to explore the impact of such an effect for jet-launching. We multiply the spacial $\alpha$ profile with a time-dependent function. Here we apply a periodic step function (Figure \[fig:steps\]), by which we continuously switch on and off the dynamo in the disk. The periodic step function is characterized by its period $T_0$ and time length of a step function $\Delta T$. In other words, $T_0$ is the period of the dynamo cycle and $\Delta T$ the activity cycle of the dynamo. Below we present the simulation with $T_0 = 1000$ and $\Delta T = 400$.
![Periodic step function applied for the toy model of a time-variable disk dynamo. Here the dynamo alpha is simply modulated by the periodic step function. Thus, the dynamo is switched on after periods of $n T_0$ and switched off at $n T_0 + \Delta T$. []{data-label="fig:steps"}](\figurepath/steps-eps-converted-to.pdf){width="9cm"}
![The time evolution of the actual disk magnetization ${ \mu_{\rm act} }$ of the reference dynamo simulation (blue line) and time-dependent dynamo simulation (green line).[]{data-label="fig:dyn_tmu"}](\figurepath/dyn_tmu-eps-converted-to.pdf){width="9cm"}
![The time evolution of the accretion Mach number ${M_{R}}$ of the reference dynamo simulation (blue line) and time-dependent dynamo simulation (green line).[]{data-label="fig:dyn_tmr"}](\figurepath/dyn_tmr-eps-converted-to.pdf){width="9cm"}
{width="18cm"}
Essentially, the modulation of the dynamo-$\alpha $ leads to the variation of the disk magnetization (Figure \[fig:dyn\_tmu\]) with the same periodicity. The strength of diffusivity is chosen such that without a dynamo working in the disk, the advection of the magnetic flux with the accreted material cannot balance the outward diffusion of the magnetic field. This eventually results in a decrease of the disk magnetization. As previously shown (see Paper I for details), there exists a limit of the strength of the magnetization below which the disk cannot sustain a jet. When the disk magnetization decays below the level of $\mu \approx 10^{-3}$ the strong jet disappears. On the other hand, if the dynamo action is re-established, the disk magnetization grows again and the outflow is re-launched.
We performed a series of parameter runs, varying both $T_0$ and $\Delta T$. In principle, different scenarios are possible, depending not only on the period of step function, but also on the magnetic diffusivity and the dynamo parameter (and various combinations of those).
In order to generate episodic ejection events, thus in order to re-establish a jet-driving magnetic field structure, several conditions have to be met by the dynamo process. First, the dynamo must be strong enough in order to generate the magnetic field sufficiently fast. Second, in order to suppress the jet ejection during consecutive switch-off periods, $(n T_0 - \Delta T)$, these periods should be sufficiently long, and/or the magnetic diffusivity should be sufficiently high. Only when the inner disk magnetization decreases below $\mu \approx 10^{-3}$, the jet-launching can no be sustained and the strong outflow disappears. The overall evolution of these processes depends both on the periods $T_0$ and $\Delta T$ of the step function.
The interplay between dynamo action, accretion and diffusion may lead to different scenarios for the episodic ejection events. If the switch-off period of the the dynamo is shorter than the timescale for the magnetic field to diffuse out, the jet ejection will be constantly sustained, and the jet mass and energy fluxes will be just modulated. If the disk magnetization decays below the value that is necessary to drive a jet, and if the dynamo is weak or works only for a short time, the magnetic field will not be re-established adequately, and, therefore no new jet will be ejected.
Applying such a toy model we are able to produce [*episodic jet events*]{}, during which the jet as well as the disk variables undergo substantial changes. The change of the disk magnetization directly affects both accretion and ejection processes. As the magnetization varies in time, also the other physical quantities vary. As discussed in Paper I, the accretion Mach number ${M_{R}}= {V_{R}}/{C_{\rm s}}$ is tightly related to the disk magnetization. Figure \[fig:dyn\_tmr\] clearly shows that variations in the disk magnetization triggered by the time-dependent dynamo directly affect the disk accretion.
Figure \[fig:epi\_tevol\] shows the time evolution of a time-dependent dynamo simulation. As for the case when we discuss the evolution of the reference simulation (Figure \[fig:dyn\_tevol\]), we show only a small cylindrical part of the whole spherical domain. We kept all parameters the same as in the reference dynamo simulation, just folding the dynamo term with the periodic step function. Obviously, compared to the reference dynamo simulation, the overall structure of the disk-jet system changes in time - substantially, and not smoothly as for the case of a constant dynamo effect. The dynamo is working until $T=400$ when it is switched off. Therefore before $T=400$ the evolution of the disk and outflow was identical to previously discussed (see Figure \[fig:dyn\_tevol\]). Between $T=400$ and $T=1,000$ the generation of the magnetic field by the dynamo was switched off. As a consequence, the magnetic field substantially diffuses and the jet velocity decreases. Although the disk magnetization is continuously decreasing, a weak outflow is still present. It is the period when the dynamo is switched off ($T_0 - \Delta T$), that indirectly limits the disk magnetization. The more time is given to diffuse away the magnetic field, the smaller the resulting disk magnetization will be in the period when the dynamo is switched off.
At $T=1,000$, the dynamo is switched on again, and the generation of the magnetic field is re-established. Because the dynamo-$\alpha$ is rather high and a substantial magnetic flux has remained from the previous cycle, it takes a rather little time to reach again sufficient magnetization for strong outflow launching. Once the substantial magnetization of the disk is reached, $\mu\approx 10^{-3}$, the outflow is re-launched. from the inner part of the disk, the outflow re-establishes to the outward direction. Advection of the magnetic flux, together with the accretion material leads to amplification of the magnetization. At $T=1,400$ the typical magneto-centrifugal structure of the magnetic field is re-established and a quasi-steady outflow re-appears, thus closing the activity cycle. Essentially, these magnetic cycles and subsequent jet ejection are repetitive. However, they are not necessarily identical, due to the magnetic field structure remaining from the previous cycle.
We note that the dynamo mechanism discussed above is able to regenerate the magnetic field in the disk completely. This is indeed different from the case of a simple modulation by the change in the magnetic diffusivity parameter ${{\alpha}_{\rm ssm}}$. In Paper I we have shown that the diffusivity parameter ${{\alpha}_{\rm ssm}}$ is very crucial for determining the [*actual*]{} disk magnetization. Without a disk dynamo, we were able to modulate the outflow simply by varying ${{\alpha}_{\rm ssm}}$ parameter, however, it was impossible to drastically affect the structure of the magnetic field, as we can now do by the dynamo.
Structure and Evolution of the Episodic Jets
--------------------------------------------
Figure \[fig:knot\_struct\] presents a time series of snapshots of the high-speed outflow propagating close to the axis. Again, we show only a small cylindrical part of the whole spherical domain, choosing time and space scales in order to display on the main outflow features. In order to show typical stages of the episodic ejection generation and propagation we have chosen the three dynamical time steps times $T= 450, 1,450, 3,450$ of our simulation lasting in fact 10,000 dynamical time steps.
First, we see how the outflow is generated initially (thus from the first cycle of dynamo activity) and then propagates throughout the hydrostatic corona. After switching off the dynamo at $T=400$, the jet weakens until it almost completely disappears. The dynamo becomes active again at $T=1,000$. At $T=1,450$ that first ejection has already reached $R \approx 1,100$, just when another “jet” has been launched. At even later time, $T=3,450$, this second knot has established an outflow, and a third knot is launched.
The timescales we have chosen are such that we can follow multiple ejection events on our grid. The time $T=1,000$ would correspond to about $T/2\pi$ inner disk rotations, thus about a year if we apply the inner disk radius $R_0 = 0.1\,$AU. For comparison, jet observations of young stars suggest the timescales between the knots $\tau_{\rm knot} \simeq \Delta L / v_{\rm jet} \sim$ of about 10 years. If the knot generation mechanism is indeed triggered by a disk dynamo, the timescale for the field reversals must be longer.
In the two middle panels of Figure \[fig:knot\_struct\] we clearly see two fast rapidly moving gas ejections. These parcels of ejected material are separated from each other by the typical period of the dynamo $T_0$, corresponding to about $1000 {r_{\rm 0}}$ distance between them. The ejected material rams into the gas which is left from the previous parcel and which moves with lower velocity. Shocks are generated that can be clearly seen in the density map[^7]. We may interpret the repeated ejections as jets knots, however, a more detailed (future) investigation would be necessary to confirm this picture.
We may clearly identify the signatures of the inflow from the inner coronal boundary into the domain along the outflow axis for low $z<200$. As discussed above, this axial inflow is essential to provide the gas pressure that balances the collimation forces of the outflow in the vicinity of the disk. In our simulations it is injected artificially by the boundary condition, however, an astrophysical interpretation could be that of a wind driven by the central object.
{width="15cm"} {width="15cm"} {width="15cm"} {width="15cm"} {width="15cm"} {width="15cm"}
Self-induced Magnetic Field Regeneration
----------------------------------------
![Example simulation resulting in two opposite magnetic loops generated by dynamo. Shown is the mass density (colors, logarithmic scale), the poloidal magnetic field lines (black lines, dashed lines show opposite magnetic flux), the sonic surface (red line), and the Alfvén surface (white line) at time $T=1000$. Arrows show normalized velocity vectors. []{data-label="fig:dyn_snapshot"}](\figurepath/dyn_snapshot-eps-converted-to.pdf){width="9cm"}
As discussed in the beginning of this section, the dynamo action is a stochastic, highly non-linear process. We can expect that under certain conditions the non-linear evolution of the dynamo is more pronounced than under other conditions. In this section we show preliminary results of simulations evolving in a more stochastic way and, by that, may be considered as a more natural ["]{}switch["]{} for the dynamo mechanism. No artificial switch on/off has been applied. These simulations consider a self-induced regeneration of the magnetic field without applying any additional constraints such as a periodic step function in the time-dependent dynamo profile.
In our simulations we have observed very similar self-induced regeneration processes of the magnetic field under different conditions. Thus, there seem to be several ways how a self-induced switch of the dynamo regeneration can take place. Some of them require the presence of a quenching mechanism. Without that, the magnetic flux will be contineously generated in the inner disk and will eventually fill the entire disk with magnetic field of one dominant polarity.
One possibility to establish a self-induced switch for the magnetic field, is to initiate the simulation with the disk filled by the toroidal magnetic field of different polarity. Then the poloidal magnetic field that is generated by the dynamo will have different polarities as well. Constant field amplification will lead to the accretion/advection of this magnetic structure. As these structures move toward the center, the generation of the magnetic field in the innermost structure will be quenched, while the magnetic flux in the outer disk will continue to grow. As a consequence these structures merge (by diffusion) and decay, and quiescent period of outflow launching follows.
Another way is to link the dynamo term to the toroidal magnetic field. As discussed above, the poloidal magnetic field in the outer disk has an opposite polarity with respect to the inner disk magnetic field. Thus, in this case an additional feedback channel is provided that - under certain circumstances - can lead to a more fluctuating evolution of the disk-outflow system. The last example that surprisingly showed such self-induced regeneration of the magnetic field is our reference simulation, but with lower dynamo-term ${ \alpha_{\rm D} }=-0.03$. Figure \[fig:dyn\_snapshot\] shows the magnetic field structure in the disk of one of such simulation. We have observed that sometimes the dynamo generates several magnetic field loops in the disk. While the magnetic field in the inner disk is able to quench the dynamo, in the outer disk the magnetic field is continuously amplified. The magnetic flux generated in the outer disk is of opposite polarity. If advected inward, it will eventually reconnect with the magnetic flux in the inner disk. During this cancellation (reconnection) process, the disk magnetization in the jet launching region will decrease below a critical level, jet-launching will decay, and the outflow will disappear. At later stages, when the magnetic field remaining from the reconnection process becomes sufficiently amplified by the dynamo, the outflow will be launched again.
The details of the process of magnetic field regeneration in fact depends on many model parameters, in particular the magnetic diffusivity model. Although this might be an interesting mechanism triggering episodic events, we do not present details here, since we were not be able yet to get it work robustly, and run the simulation longer than for just a few regenerations.
Conclusions
===========
We have presented results of MHD simulations investigating the generation of the magnetic field by the accretion disk dynamo in the context of jet and outflows launching. The time evolution of the disk structure is self-consistently taken into account. The simulations are performed in axisymmetry treating all three field components. We apply the MHD code PLUTO-4.0, that we have modified for the mean-field $\alpha^2\Omega$ dynamo problem in the induction equation.
In the present work we explored the generation of a large scale, global magnetic field. Our simulations were initiated by the purely radial magnetic field with magnetization ${\mu_{\rm init}}= 10^{-5}$. We showed in detail how the magnetic field is being generated and through which consecutive stages it evolves, acquiring in the end the ability to launch jet and outflows. In this respect our simulations can be seen as a continuation of early work by and . In our paper we are concentrating more on the jet and outflow generation and propagation.
One advantage of our simulations is that our model keeps the disk magnetization at a rather low level. Therefore, the magnetic field does not substantially affect the disk hydrodynamics, and allows to evolve our simulations for very long time. Each simulation has been evolved at least up to $T=10,000$ on a spherical domain with $R=[1, 1500]$.
In the following, we summarize our main results.
\(1) In our simulations treating a mean-field disk dynamo, we may distinguish two main features in the magnetic field structures. The magnetic field of the inner disk that is similar to the commonly found open field structure, favoring a magneto-centrifugal launching of the outflow. The poloidal magnetic field of the outer part of the disk is highly inclined, and predominantly radial. Differential rotation induces a very strong toroidal component from it. Such a structure is similar to what is known as tower jet or Poynting jet in literature. In this part of the disk, it is toroidal magnetic field that plays a key role in outflow launching. First, below the disk surface (${V_{R}}= 0$) the matter is lifted by the buoyant force of the magnetic field, thus, by the gradient of the thermal pressure. Starting from the disk surface (${V_{R}}= 0$), the matter is further accelerated by the pressure gradient of the toroidal magnetic field. The outflows from the outer part of the disk are typically slower, denser, and less collimated, thus corresponding to a higher mass loading.
\(2) In principle, the dynamo can fill the entire accretion disk with magnetic flux. Thus, if the dynamo action is not quenched, magnetic flux is continuously generated, diffuses outward along the disk until it fills the entire disk. This loop-like structure of the magnetic field that is typical for a dynamo, propagates further outward.
\(3) As soon as the disk magnetization reaches a critical limit, $\mu > 10^{-3}$, disk winds are launched and can be accelerated to super-magnetosonic speed. This result is complementary to our earlier simulations that do not consider dynamo [@paper-1], and where the critical magnetization was obtained just from advection of magnetic flux by accretion. Thus, again we can confirm the long-standing belief that the disk magnetization plays the key role in the outflow launching. In the inner disk, the rate of generation of the magnetic field by the dynamo is higher, leading to a strong gradient of the disk magnetization.
\(4) We have further invented a toy model triggering a time-dependent efficiency of the mean-field dynamo. In that model approach, we periodically switch on and off the dynamo. This strongly affects the magnetic field structure. The decay of magnetic flux by diffusion can be completely balanced by the dynamo that regenerates the magnetic field. As a consequence, the duty cycles of the dynamo action can lead to episodic jet ejection, just depending on the disk magnetization obtained during dynamo activity. When the dynamo is suppressed and the disk magnetization falls below a critical value, $\mu \approx 10^{-3}$, the generation of outflows as well as the accretion is substantially inhibited. We had chosen the timescale of the dynamo cycle and the corresponding timescale for the episodic ejections somehow shorter compared to the observed values - just because we wanted to follow several events in the same simulation box. However, the main - and general - result is that we can steer episodic ejection and [*large-scale jet knots*]{} by the [*disk intrinsic dynamo*]{} that is time-dependent and regenerates the jet-launching magnetic field.
\(5) Concerning the disk hydrodynamics, we find that the accretion velocity follows the same power-law $\beta_{{V_{R}}} \approx -2/5$ for the simulations with and without dynamo. This interesting also, because we have applied slightly different diffusivity model leading to different magnetic field structure. Nevertheless, the accretion profiles are approximately the same. As a consequence, we also find approximately the same ejection index $\xi \approx 0.25$.
\(6) Although the dynamo and non-dynamo simulations are significantly different, launching mechanism of the fast jet is primarily the same. Thus, from a pure observational point of view, one would not yet be able to distinguish whether the jets are launched from a dynamo-generated magnetic field or from a magnetic field advected from the interstellar medium. In summary, we have shown the accretion-ejection evolution considering a magnetic field self-generated by a mean-field disk dynamo. Repetitive ejection could be obtained by a time-dependent dynamo-$\alpha$. A future step could be to consider the dynamo-action for the strong-field case. That might be realized by implementing both a MRI dynamo and a Parker dynamo by means of different dynamo-$\alpha$. Another step would be to have a more direct link between the actual magnetic field and the dynamo.
We thank Andrea Mignone and the PLUTO team for the possibility to use their code. The simulations were performed on the THEO cluster of Max Planck Institute for Astronomy. This work was partly financed by the SFB 881 of the German science foundation DFG.
Resolution Study
================
We shortly discuss the results of our resolution study. We have performed simulations with a grid resolution of 0.75, and 1.5 times our standard resolution of $N_\theta=128$ cells per quadrant, corresponding to $N_\theta=96$ and $N_\theta=192$ cells per quadrant, or approximately 12 or 24 cells per disk height $2\epsilon$ compared to 16 cells per disk height in our reference run[^8].
Figure \[fig:res\_all\] shows the dynamical profiles the three simulations of our resolution study. The radial profiles are plotted along the midplane for various dynamical variables at time $T=10,000$. As discussed in Section \[sub:profiles\] these profiles can be nicely fitted by power-laws. The same power-law index also provides the same ejection to accretion index. Therefore, we conclude that our results are not resolution dependent.
However, several differences between these curves can be noticed. The inner part of the disk for the simulation with lower resolution indicates a substantial deviation from the corresponding power-law. We also find that for a lower resolution the accretion speed increases, and, as a consequence, the overall density in the disk decreases. This highlights the effect of the numerical viscosity that enhances the angular momentum transfer in the disk. On the other hand, the magnetic field is diffused out faster and to a larger distance. This indicates a higher numerical resistivity for the case of lower resolution.
![Resolution study. Physical quantities along the midplane for the simulations with different resolution at $T=10,000$. From top to bottom the resolution is (12, 16, 24) cells per disk height ($2\epsilon$). Colors show different variable profile, thick dashed lines correspond to certain power-law, the mismatched thin dashed lines correspond to initial distributions of variables. []{data-label="fig:res_all"}](\figurepath/res_u_profiles-eps-converted-to.pdf "fig:"){width="5.9cm"} ![Resolution study. Physical quantities along the midplane for the simulations with different resolution at $T=10,000$. From top to bottom the resolution is (12, 16, 24) cells per disk height ($2\epsilon$). Colors show different variable profile, thick dashed lines correspond to certain power-law, the mismatched thin dashed lines correspond to initial distributions of variables. []{data-label="fig:res_all"}](\figurepath/res_def_profiles-eps-converted-to.pdf "fig:"){width="5.9cm"} ![Resolution study. Physical quantities along the midplane for the simulations with different resolution at $T=10,000$. From top to bottom the resolution is (12, 16, 24) cells per disk height ($2\epsilon$). Colors show different variable profile, thick dashed lines correspond to certain power-law, the mismatched thin dashed lines correspond to initial distributions of variables. []{data-label="fig:res_all"}](\figurepath/res_r_profiles-eps-converted-to.pdf "fig:"){width="5.9cm"}
[55]{} natexlab\#1[\#1]{}
, R. & [R[ü]{}diger]{}, G. 2001, , 374, 1035
, P. J., [Livio]{}, M., & [Pringle]{}, J. E. 1996, , 457, 332
, A., [von Rekowski]{}, B., [Dobler]{}, W., [Brandenburg]{}, A., & [Shukurov]{}, A. 2001, , 370, 635
, R. D. & [Payne]{}, D. G. 1982, , 199, 883
, A. & [Campbell]{}, C. 1997, in Lecture Notes in Physics, Berlin Springer Verlag, Vol. 487, Accretion Disks - New Aspects, ed. E. [Meyer-Hofmeister]{} & H. [Spruit]{}, 109
, A. & [Donner]{}, K. J. 1997, , 288, L29
, A., [Nordlund]{}, A., [Stein]{}, R. F., & [Torkelsson]{}, U. 1995, , 446, 741
, A., [Sokoloff]{}, D., & [Subramanian]{}, K. 2012, , 169, 123
, A. & [Subramanian]{}, K. 2005, , 417, 1
, A. & [von Rekowski]{}, B. 2007, , 78, 374
, F. & [Keppens]{}, R. 2002, , 581, 988
—. 2004, , 601, 90
, J. A. 1976, Science, 192, 1189
, C. 2006, , 651, 272
—. 2009, , 692, 346
, C. & [Elstner]{}, D. 1999, , 349, L61
—. 2000, , 363, 208
, C. & [Sheikhnezami]{}, S. 2013, , 774, 12
, C. & [[Č]{}emelji[ć]{}]{}, M. 2002, , 395, 1045
, J. 1997, , 319, 340
, J. & [Pelletier]{}, G. 1995, , 295, 807
, S. 2013, in EAS Publications Series, Vol. 62, EAS Publications Series, 95–142
, C. F. & [Menou]{}, K. 1998, , 492, L75
, M. R., [Shibata]{}, K., & [Matsumoto]{}, R. 1996, , 468, L37
, A. & [Levin]{}, Y. 2008, , 490, 501
, Y., [Hayashi]{}, M. R., & [Matsumoto]{}, R. 2004, , 600, 338
, R., [Li]{}, Z., & [Blandford]{}, R. 1999, , 526, 631
, F. & [Rädler]{}, K.-H. 1980, [Mean-field magnetohydrodynamics and dynamo theory]{}
, R. V. E., [Romanova]{}, M. M., & [Bisnovatyi-Kogan]{}, G. S. 1995, , 275, 244
, D. 1996, , 279, 389
, D. & [Boily]{}, C. 1994, , 267, 146
, Z., [Casse]{}, F., & [Sauty]{}, C. 2006, , 460, 1
, A., [Bodo]{}, G., [Massaglia]{}, S., [Matsakos]{}, T., [Tesileanu]{}, O., [Zanni]{}, C., & [Ferrari]{}, A. 2007, , 170, 228
, G. C., [Ferreira]{}, J., & [Zanni]{}, C. 2010, , 512, A82+
, R. & [Pudritz]{}, R. E. 1997, , 482, 712
, G. & [Pudritz]{}, R. E. 1992, , 394, 117
, O. & [Fendt]{}, C. 2010, , 709, 1100
, R. E. 1981, , 195, 897
—. 1981, , 195, 881
, R. E., [Ouyed]{}, R., [Fendt]{}, C., & [Brandenburg]{}, A. 2007, Protostars and Planets V, 277
, R. E., [Rogers]{}, C. S., & [Ouyed]{}, R. 2006, , 365, 1131
, M. v., [R[ü]{}diger]{}, G., & [Elstner]{}, D. 2000, , 353, 813
, M., [Joung Turner]{}, N., & [Bodenheimer]{}, P. 1995, , 276, 1179
, S., [Fendt]{}, C., [Porth]{}, O., [Vaidya]{}, B., & [Ghanbari]{}, J. 2012, , 757, 65
, K. & [Uchida]{}, Y. 1985, , 37, 31
, D. & [Fendt]{}, C. 2014,
—. 2014,
, C. A. & [Pringle]{}, J. E. 1992, , 259, 604
, P., [Ferrari]{}, A., [Mignone]{}, A., [Zanni]{}, C., [Bodo]{}, G., & [Massaglia]{}, S. 2009, , 400, 820
, G. V., [Koldoba]{}, A. V., [Romanova]{}, M. M., [Chechetkin]{}, V. M., & [Lovelace]{}, R. V. E. 1995, , 439, L39
, G. V., [Lovelace]{}, R. V. E., [Romanova]{}, M. M., [Li]{}, H., & [Colgate]{}, S. A. 2000, , 541, L21
, B., [Fendt]{}, C., & [Beuther]{}, H. 2009, , 702, 567
, B. & [Brandenburg]{}, A. 2004, , 420, 17
, B., [Brandenburg]{}, A., [Dobler]{}, W., [Dobler]{}, W., & [Shukurov]{}, A. 2003, , 398, 825
, C., [Ferrari]{}, A., [Rosner]{}, R., [Bodo]{}, G., & [Massaglia]{}, S. 2007, , 469, 811
[^1]: In this paper we apply the following notation. With jet [*launching*]{} we denote the process that lifts accreting material out of the disk and couples it to a disk wind - the accretion-ejection structure. With jet [*formation*]{} we denote the acceleration and collimation of that slow disk wind from the disk surface to a high velocity of super-escape speed, super-Alfénic speed, and possibly super-magnetosonic speed
[^2]: By coincidence, the difference in the respective time scales - a magnetic cycle of 22 years and a rotational period of 35 days - is comparable to the protostellar jets.
[^3]: Version 4.0, released 2013
[^4]: Note however, that we use adiabatic equation of state
[^5]: The initial field is purely radial or toroidal
[^6]: Steady state and a power-law nature of an accretion rate is implicitly assumed
[^7]: The shock structure is also visible in the pressure distribution and also in the jumps in the velocity profile along the jet (not shown as figure).
[^8]: Note that once the resolution in $\theta$-direction and the radial extent of the disk is chosen, the resolution in $R$-direction is uniquely determined (see Section \[sec:model\], or Paper I)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we generalize the Givental theory for Frobenius manifolds and cohomological field theories to flat F-manifolds and F-cohomological field theories. In particular, we define the notion of Givental cone for flat F-manifolds, and we provide a generalization of the Givental group as a matrix loop group acting on them. We show that this action is transitive on semisimple flat F-manifolds. We then extend this action to F-cohomological field theories in all genera. We show that, given a semisimple flat F-manifold and a Givental group element connecting it to the constant flat F-manifold at its origin, one can construct a family of F-CohFTs in all genera, parameterized by a vector in the associative algebra at the origin, whose genus $0$ part is the given flat F-manifold. If the flat F-manifold is homogeneous, then the associated family of F-CohFTs contains a subfamily of homogeneous F-CohFTs. However, unlike in the case of Frobenius manifolds and CohFTs, these homogeneous F-CohFTs can have different conformal dimensions, which are determined by the properties of a certain metric associated to the flat F-manifold.'
address:
- 'A. Arsie:Department of Mathematics and Statistics, The University of Toledo, 2801W. Bancroft St., 43606 Toledo, OH, USA'
- 'A. Buryak:School of Mathematics, University of Leeds, Leeds, LS2 9JT, United Kingdom'
- 'P. Lorenzoni:Dipartimento di Matematica e Applicazioni, Università di Milano-Bicocca, Via Roberto Cozzi 53, I-20125 Milano, Italy'
- 'P. Rossi:Dipartimento di Matematica “Tullio Levi-Civita”, Università degli Studi di Padova,Via Trieste 63, 35121 Padova, Italy'
author:
- Alessandro Arsie
- Alexandr Buryak
- Paolo Lorenzoni
- Paolo Rossi
title: 'Semisimple flat F-manifolds in higher genus'
---
Introduction {#introduction .unnumbered}
============
In a series of influential papers [@Giv01a; @Giv01b; @Giv04] A. Givental, inspired by localization formulas in the Gromov–Witten theory of projective spaces, described a technique to express the all genera descendant Gromov–Witten potential of a target variety with semisimple quantum cohomology in terms of the action of a certain operator on $N$ copies (where $N$ is the dimension of the target variety’s cohomology) of the descendant potential of a point (also known as the Witten–Kontsevich tau-function).\
Restricting the attention to the genus $0$ descendant potential, the situation is described as the action of a certain loop group of matrices on the space of descendant potentials of calibrated Frobenius manifolds (i.e solutions to certain differential equations inspired by Gromov–Witten theory and well-known in the theory of Frobenius manifolds, see [@DZ05])). This action is in fact transitive when further restricted to semisimple Frobenius manifolds. In this sense, taking as a starting point the $N$-fold product of the trivial $1$-dimensional Frobenius manifold, the descendant potential of any other fixed calibrated semisimple Frobenius manifold can be recovered by the action of an operator representing an element of the Givental group connecting these two semisimple Frobenius manifolds.\
The geometric set up for this result is interpreting the genus $0$ descendant potential as the generating function for a Lagrangian cone in an infinite dimensional symplectic vector space. The Givental group is then a group of symplectic transformations acting on the set of all such cones.\
In [@Giv01b], Givental conjectured how to extend such action to the potential at all genera as a canonical quantization of the above symplectic action. In particular, Givental’s formula can be seen as a way to reconstruct higher genus descendant Gromov–Witten invariants of a target variety from its (genus $0$) quantum cohomology, as long as this is semisimple.\
In [@Tel12], Teleman proved Givental’s reconstruction formula by reformulating the problem in the language of cohomological field theories (CohFTs), families of cohomology classes on the moduli space of stable algebraic curves introduced in [@KM94] to axiomatize the properties of Gromov–Witten invariants. In this context, the Givental symplectic loop group is seen as acting directly on the space of all CohFTs, and this action restricts to a transitive action on the space of semisimple CohFTs. The explicit form of this action had been known to experts for a while before being first accurately described in [@PPZ15].\
In this language, the Givental–Teleman reconstruction theorem relies on two results: transitivity of the action of the Givental group in genus $0$ (i.e. on descendant potentials of semisimple calibrated Frobenius manifolds) and the fact that a semisimple CohFT is essentially uniquely reconstructable from its genus $0$ part (in fact up to insertions of Hodge classes in the general case, and uniquely for homogeneous CohFTs). This shows in particular that the Givental group acts transitively on semisimple CohFTs.\
In this paper we generalize Givental’s theory to flat F-manifolds and F-CohFTs. F-manifolds were introduced by Hertling and Manin in [@HM99] (see also the book [@Man99]). They are generalizations of Frobenius manifolds, for which one drops the potentiality condition and the presence of a metric. Flat F-manifolds were first studied by Getzler [@Get04] and Manin [@Man05] (in Getzler’s paper they are called Dubrovin manifolds) and they are often useful to capture interesting structures in singularity theory and algebraic geometry (for instance they appear in genus $0$ open Gromov–Witten theory, as remarked in [@BB19]).\
In Section \[section:flat F-manifolds\], we develop the theory of semisimple flat F-manifolds in canonical coordinates, reformulating some known facts and introducing stronger or more precise technical results needed for Section \[section:Givental theory for flat F-manifolds\], where we introduce a suitably generalized version of Givental theory for flat F-manifolds.\
In particular, we prove that to a calibrated flat F-manifold one can univocally associate a sequence of descendant vector potentials describing a cone in an infinite dimensional vector space. When the flat F-manifold is Frobenius, such cone is Lagrangian with respect to the symplectic structure constructed via the flat metric, as proved in [@Giv04]. We then introduce a larger Givental-type loop group (which is not symplectic anymore) and a corresponding action, which is defined on the space of such descendant cones. We prove that the action is transitive for semisimple flat F-manifolds, thereby completely generalizing the genus $0$ Givental theory. We also recall the definition of homogeneous flat F-manifold given in [@BB19] and the related notions of Saito structure without metric (introduced in [@Sab98]) and of bi-flat F-manifold (introduced in [@AL13]). For homogeneous flat F-manifolds we show that the $R$-matrix defining an element of the generalized Givental group is uniquely determined.\
In Section \[section:F-CohFT\], we study the corresponding generalization of CohFTs, which we call F-CohFTs. They were introduced in [@BR18], where the first, all genera, explicit example, relevant for open Gromov–Witten theory, was constructed. They are in fact generalizations of partial CohFTs: the gluing axiom at a nonseparating node is dropped and moreover the complete equivariance of the classes with respect to permutation of marked points is broken, as one of them carries a co-vector, instead of a vector. This removes the necessity of a metric, which is then also dropped. Indeed, partial CohFTs in genus $0$ still reduce to Frobenius manifolds, while F-CohFTs give flat F-manifolds.\
Finally, in Section \[section:group action of F-CohFTs\], we extend our generalized Givental group action to F-CohFTs in all genera and we show our main result: given any semisimple flat F-manifold we can construct an F-CohFT with that F-manifold as its genus $0$ part. In fact, because of the absence of the gluing axiom at nonseparating nodes, some genus $1$ information is needed to fix the nonzero genus part. This is done by specifying the degree $0$ part of the F-CohFT on ${\overline{\mathcal M}}_{1,1}$, which amounts to a vector $G_0$ in the F-CohFT phase space $V$. For each choice of such a vector, we construct a different F-CohFT with the given flat F-manifold as its genus $0$ part. If the flat F-manifold is homogeneous, then we construct a decomposition $V=\oplus_{i\in I} V_i$ and prove that for $G_0\in V_i$ the resulting F-CohFT is homogeneous of conformal dimension $\gamma_i$. The collection of numbers $\gamma_i$, $i\in I$ is determined by the properties of a certain metric that is associated to the flat F-manifold.\
We remark that, beside the aforementioned applications to singularity theory and moduli spaces of curves, another motivation for studying F-CohFTs and developing a corresponding Givental-type theory comes from integrable systems. Indeed, as shown in [@BR18], to any F-CohFT one can associate, via a suitable generalization of the double ramification hierarchy construction of [@Bur15; @BR16], an infinite hierarchy of compatible evolutionary PDEs (in particular, systems of conservation laws). The dispersionless limit of this hierarchy is the principal hierarchy associated with the corresponding flat F-manifold, see Section \[section:Givental theory for flat F-manifolds\].\
The results of the present paper then allow to construct a family of dispersive deformations of the principal hierarchy of any semisimple flat F-manifold, parametrized by a vector $G_0$ at its origin. In the homogeneous case, choices of $G_0$ exist for which the deformation is homogeneous. We will study in detail such dispersive deformations and the properties of the double ramification hierarchy of an F-CohFT in our next paper.\
Acknowledgments {#acknowledgments .unnumbered}
---------------
A. B. was supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 797635. P. L. is supported by MIUR - FFABR funds 2017 and by funds of H2020-MSCA-RISE-2017 Project No. 778010 IPaDEGAN.\
Flat F-manifolds around a semisimple point {#section:flat F-manifolds}
==========================================
After recalling the definition of flat F-manifold as a generalization of the notion of Frobenius manifold, in this section we show that a flat F-manifold around a semisimple point possesses a metric and we construct rotation coefficients and a sequence of $R$-matrices. These objects will play an important role in our later construction of an F-cohomological field theory in all genera associated to a flat F-manifold.
Flat F-manifolds and Frobenius manifolds
----------------------------------------
We recall here the following facts and definitions from [@Get04; @Man05], see also [@AL18] and [@Dub96].
A flat F-manifold $(M,\nabla,\circ,e)$ is the datum of an analytic manifold $M$, an analytic connection $\nabla$ in the tangent bundle $T M$, an algebra structure $(T_p M,\circ)$ with unit $e$ on each tangent space, analytically depending on the point $p\in M$, such that the one-parameter family of connections $\nabla_z=\nabla+z\circ$ is flat and torsionless for any $z\in{\mathbb C}$, and $\nabla e=0$.
From the flatness and the torsionlessness of $\nabla_z$ one can deduce the commutativity and the associativity of the algebras $(T_pM,\circ)$. Moreover, if one choses flat coordinates $t^\alpha$, $1\le\alpha\le N$, $N=\dim M$, for the connection $\nabla$, then it is easy to see that locally there exist analytic functions $F^\alpha(t^1,\ldots,t^N)$, $1\leq\alpha\leq N$, such that the second derivatives $$\begin{gathered}
\label{eq:structure constants of flat F-man}
c^\alpha_{\beta\gamma}=\frac{\d^2 F^\alpha}{\d t^\beta \d t^\gamma}\end{gathered}$$ are the structure constants of the algebras $(T_p M,\circ)$, $$\begin{gathered}
\frac{\d}{\d t^\beta}\circ\frac{\d}{\d t^\gamma}=c^\alpha_{\beta\gamma}\frac{\d}{\d t^\alpha}.\end{gathered}$$ Also, in the coordinates $t^\alpha$ the unit $e$ has the form $e=A^\alpha\frac{\d}{\d t^\alpha}$ for some constants $A^\alpha\in{\mathbb C}$. Note that we use Einstein’s convention of sum over repeated Greek indices. From the associativity of the algebras $(T_p M,\circ)$ and the fact that the vector field $A^\alpha\frac{\d}{\d t^\alpha}$ is the unit it follows that $$\begin{aligned}
A^\mu\frac{\d^2 F^\alpha}{\d t^\mu\d t^\beta} &= \delta^\alpha_\beta, && 1\leq \alpha,\beta\leq N,\label{eq:axiom1 of flat F-man}\\
\frac{\d^2 F^\alpha}{\d t^\beta \d t^\mu} \frac{\d^2 F^\mu}{\d t^\gamma \d t^\delta} &= \frac{\d^2 F^\alpha}{\d t^\gamma \d t^\mu} \frac{\d^2 F^\mu}{\d t^\beta \d t^\delta}, && 1\leq \alpha,\beta,\gamma,\delta\leq N.\label{eq:axiom2 of flat F-man}\end{aligned}$$ The $N$-tuple of functions ${\overline{F}}=(F^1,\ldots,F^N)$ is called the [*vector potential*]{} of the flat F-manifold.\
Conversely, if $M$ is an open subset of ${\mathbb C}^N$ and $F^1,\ldots,F^N\in{\mathcal{O}}(M)$ are functions satisfying equations and , then these functions define a flat F-manifold $(M,\nabla,\circ,A^\alpha\frac{\d}{\d t^\alpha})$ with the connection $\nabla$ given by $\nabla_{\frac{\d}{\d t^\alpha}}\frac{\d}{\d t^\beta}=0$, and the multiplication $\circ$ given by the structure constants .
Consider a flat F-manifold $(M,\nabla,\circ,e)$ and a symmetric nondegenerate bilinear form $g$ (often called a metric) on the tangent spaces $T_pM$ analytically depending on the point $p\in M$. We say that $g$ is compatible with the product $\circ$ if $$\begin{gathered}
g(X\circ Y, Z)=g(X,Y\circ Z)\end{gathered}$$ for any local vector fields $X,Y,Z $ on $M$.
A point $p\in M$ of an $N$-dimensional flat F-manifold $(M,\nabla,\circ,e)$ is called *semisimple* if $T_pM$ has a basis of idempotents $\pi_1,\dots,\pi_N$ satisfying $\pi_k \circ \pi_l = \delta_{k,l} \pi_k$. Moreover, locally around such a point one can choose coordinates $u^i$ such that $\frac{\d}{\d u^k}\circ\frac{\d}{\d u^l}=\delta_{k,l}\frac{\d}{\d u^k}$. These coordinates are called [*canonical coordinates*]{}. In particular, this means that semisimplicity at a point is an open property on $M$. In canonical coordinates we have $e=\sum_i\frac{\d}{\d u^i}$.\
A flat F-manifold given by a vector potential $(F^1,\ldots,F^N)$ is called [*homogeneous*]{} if there exists a vector field $E$ of the form $$\label{Euler}
E=\sum_{\alpha=1}^N(\underbrace{(1-q_\alpha)t^\alpha+r^\alpha}_{=:E^\alpha})\frac{\d}{\d t^\alpha},\quad q_\alpha,r^\alpha\in{\mathbb C},$$ satisfying $[e,E]=E$ and such that $$\begin{gathered}
E^\mu\frac{\d F^\alpha}{\d t^\mu}=(2-q_\alpha)F^\alpha+A^\alpha_\beta t^\beta+B^\alpha,\quad\text{for some $A^\alpha_\beta,B^\alpha\in{\mathbb C}$}.\end{gathered}$$ Note that this equation can be written more invariantly as ${\mathrm{Lie}}_E(\circ)=\circ$, where ${\mathrm{Lie}}_E$ denotes the Lie derivative. The vector field $E$ is called the [*Euler vector field*]{}.
A flat F-manifold $(M,\nabla,\circ,e)$ is called a Frobenius manifold if it is equipped with a metric $\eta$ compatible with the product $\circ$ and such that $\nabla \eta = 0$. The connection $\nabla$ is then the Levi-Civita connection associated to $\eta$. A Frobenius manifold will be denoted by a tuple $(M,\eta,\circ,e)$.
Homogeneous Frobenius manifolds are sometimes called conformal Frobenius manifolds.\
In case a flat F-manifold is actually Frobenius, the vector potential ${\overline{F}}$ can be shown to descend locally from a Frobenius potential $F(t^*)$ as $F^\alpha(t^*) = \eta^{\alpha\mu} \frac{\d F(t^*)}{\d t^\mu}$ and the Frobenius potential $F(t^*)$ satisfies $$\begin{aligned}
A^\mu\frac{\d^3 F}{\d t^\mu\d t^\alpha \d t^\beta} &= \eta_{\alpha\beta}, && 1\leq \alpha,\beta\leq N,\label{eq:axiom1 for Frobenius manifolds}\\
\frac{\d^3 F}{\d t^\alpha \d t^\beta \d t^\mu} \eta^{\mu \nu}\frac{\d^3 F}{\d t^\nu \d t^\gamma \d t^\delta} &= \frac{\d^3 F}{\d t^\alpha \d t^\gamma \d t^\mu} \eta^{\mu \nu}\frac{\d^3 F}{\d t^\nu \d t^\beta \d t^\delta}, && 1\leq \alpha,\beta,\gamma,\delta\leq N.\label{eq:axiom2 for Frobenius manifolds}\end{aligned}$$ In particular, the structure functions $c^\alpha_{\beta\gamma}$ of the algebras $(T_pM,\circ)$ can be written as $c^\alpha_{\beta\gamma} =\eta^{\alpha\mu} \frac{\d^3 F}{\d t^\mu \d t^\beta \d t^\gamma}$, $1\leq \alpha,\beta,\gamma\leq N$.
Metric, rotation coefficients and $R$-matrices {#subsection:metric for a flat F-manifold}
----------------------------------------------
Consider a flat F-manifold $(M,\nabla,\circ,e)$ around a semisimple point. Let $u^1,\ldots,u^N$ be the canonical coordinates. By $t^1,\ldots,t^N$ we denote the flat coordinates.\
In general, our flat F-manifold is not Frobenius and so it doesn’t possess a metric which is covariantly constant with respect to $\nabla$ and compatible with the product $\circ$. However, there is a natural metric compatible with the product $\circ$, which was first constructed in [@AL13]. Introduce a matrix ${\widetilde{\Psi}}$ by $${\widetilde{\Psi}}:=\left(\frac{\d u^i}{\d t^\alpha}\right).$$ Note that, in canonical coordinates, the connection $\nabla_z=\nabla+z\circ$ is given by $$\nabla+z\circ=d-d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}+z dU,$$ where $U:={\mathrm{diag}}(u^1,\ldots,u^N)$.
\[proposition:properties of tGamma\] 1. The matrix $d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}$ has the form $$d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}={\widetilde{D}}+[{\widetilde{\Gamma}},dU],$$ where ${\widetilde{D}}$ is a diagonal matrix consisting of one-forms and ${\widetilde{\Gamma}}$ is a matrix with vanishing diagonal entries.\
2. We have $d{\widetilde{D}}=0$ and $$\begin{gathered}
\label{eq:system for tGamma}
d[{\widetilde{\Gamma}},dU]={\widetilde{D}}\wedge [{\widetilde{\Gamma}},dU]+[{\widetilde{\Gamma}},dU]\wedge{\widetilde{D}}+[{\widetilde{\Gamma}},dU]\wedge [{\widetilde{\Gamma}},dU].\end{gathered}$$
Denote $M:=d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}$. The flatness of $\nabla_z$ is equivalent to the equation $$\begin{gathered}
\label{eq:equations for M}
-d M+(-M+zdU)\wedge(-M+zdU)=0\Leftrightarrow
\left\{
\begin{aligned}
&M\wedge dU+dU\wedge M=0,\\
&dM=M\wedge M.
\end{aligned}
\right.\end{gathered}$$ Part 1 of the proposition follows from the equation $M\wedge dU+dU\wedge M=0$. For Part 2 we write $$\begin{gathered}
d\left({\widetilde{D}}+[{\widetilde{\Gamma}},dU]\right)=\left({\widetilde{D}}+[{\widetilde{\Gamma}},dU]\right)\wedge\left({\widetilde{D}}+[{\widetilde{\Gamma}},dU]\right)\Leftrightarrow\\
\Leftrightarrow d{\widetilde{D}}+d[{\widetilde{\Gamma}},dU]={\widetilde{D}}\wedge [{\widetilde{\Gamma}},dU]+[{\widetilde{\Gamma}},dU]\wedge{\widetilde{D}}+[{\widetilde{\Gamma}},dU]\wedge [{\widetilde{\Gamma}},dU],\end{gathered}$$ and it remains to note that the diagonal parts of the matrices $[{\widetilde{\Gamma}},dU]$ and $[{\widetilde{\Gamma}},dU]\wedge[{\widetilde{\Gamma}},dU]$ are equal to zero. The proposition is proved.
Let $\Gamma^i_{jk}$ be the Christoffel symbols of the connection $\nabla$ in canonical coordinates, ${\widetilde{\Gamma}}=({\widetilde{\gamma}}^i_j)$ and ${\widetilde{D}}={\mathrm{diag}}({\widetilde{D}}_1,\ldots,{\widetilde{D}}_N)$, where ${\widetilde{D}}_i$ are one-forms ${\widetilde{D}}_i=\sum_j{\widetilde{D}}_{ij}du^j$. Note that ${\widetilde{D}}_{ij}=-\Gamma^i_{ji}$. Proposition \[proposition:properties of tGamma\] together with the fact $\sum_k\frac{\d{\widetilde{\Psi}}}{\d u^k}=0$ implies that $$\begin{aligned}
&\Gamma^i_{jk}=0,&&k\ne i\ne j\ne k,\label{eq:Gamma in terms of tgamma1}\\
&\Gamma^i_{ij}=\Gamma^i_{ji}=-\Gamma^i_{jj}={\widetilde{\gamma}}^i_j,&& i\ne j,\label{eq:Gamma in terms of tgamma2}\\
&\Gamma^i_{ii}=-\sum_{k\ne i}{\widetilde{\gamma}}^i_k,\label{eq:Gamma in terms of tgamma3}\end{aligned}$$ and that the functions ${\widetilde{\gamma}}^i_j$ satisfy the following system: $$\begin{aligned}
\frac{\d{\widetilde{\gamma}}^i_j}{\d u^k}=&-{\widetilde{\gamma}}^i_j{\widetilde{\gamma}}^i_k+{\widetilde{\gamma}}^i_j{\widetilde{\gamma}}^j_k+{\widetilde{\gamma}}^i_k{\widetilde{\gamma}}^k_j, && i\ne k\ne j\ne i,\label{eq:Darboux-Egoroff for tgamma,1}\\
\sum_k\frac{\d{\widetilde{\gamma}}^i_j}{\d u^k}=&0,&& i\ne j.\label{eq:Darboux-Egoroff for tgamma,2}\end{aligned}$$
Since $d{\widetilde{D}}=0$, there exists a nondegenerate diagonal matrix $H={\mathrm{diag}}(H_1,\ldots,H_N)$ satisfying $$dH\cdot H^{-1}=-{\widetilde{D}}.$$ The functions $H_i$ are defined by this property uniquely up to rescalings $H_i\mapsto\lambda_i H_i$, $\lambda_i\in{\mathbb C}^*$. Define a metric $g=\sum_i g_i(d u^i)^2$ on our flat F-manifold by $g_i:=H_i^2$. It is clearly compatible with the product $\circ$. If our flat F-manifold is Frobenius, then there exist numbers $\lambda_i\in{\mathbb C}^*$ such that the metric $\sum_i \lambda_i g_i(d u^i)^2$ coincides with the metric $\eta$.\
Define matrices $\Psi$ and $\Gamma$ by $$\Psi:=H{\widetilde{\Psi}},\qquad\Gamma=(\gamma^i_j):=H{\widetilde{\Gamma}}H^{-1}.$$ Let us call the coefficients $\gamma^i_j$ the [*rotation coefficients*]{}.
We have $$\begin{gathered}
\label{eq:formulas for dPsi and dGammadU}
d\Psi=[\Gamma,dU]\Psi,\qquad d[\Gamma,dU]=[\Gamma,dU]\wedge[\Gamma,dU].\end{gathered}$$
We compute $$\begin{aligned}
d\Psi\cdot\Psi^{-1}=&d\left(H{\widetilde{\Psi}}\right){\widetilde{\Psi}}^{-1}H^{-1}=dH\cdot H^{-1}+{\widetilde{D}}+H[{\widetilde{\Gamma}},dU]H^{-1}=[\Gamma,dU],\\
d[\Gamma,dU]=&d\left(H[{\widetilde{\Gamma}},dU]H^{-1}\right)=\\
=&-{\widetilde{D}}\wedge[\Gamma,dU]+\left({\widetilde{D}}\wedge[\Gamma,dU]+[\Gamma,dU]\wedge{\widetilde{D}}+[\Gamma,dU]\wedge[\Gamma,dU]\right)-[\Gamma,dU]\wedge{\widetilde{D}}=\\
=&[\Gamma,dU]\wedge[\Gamma,dU].\end{aligned}$$
Note that the matrix equation $d[\Gamma,dU]=[\Gamma,dU]\wedge[\Gamma,dU]$ is equivalent to the system $$\begin{gathered}
\left\{\begin{aligned}
&\frac{\d\gamma^i_j}{\d u^k}=\gamma^i_k\gamma^k_j,\quad k\ne i\ne j\ne k,\\
&\sum_{k=1}^N\frac{\d \gamma^i_j}{\d u^k}=0,
\end{aligned}\right.\end{gathered}$$ which is the classical Darboux–Egorov system. In the case of a Frobenius manifold the coefficients $\gamma^i_j$ are the rotation coefficients of the metric and satisfy the additional symmetry property $\gamma^i_j=\gamma^j_i$.\
Note also that we have the system $$\frac{\d H_i}{\d u^j}=
\begin{cases}
\gamma^i_j H_j,&\text{if $i\ne j$},\\
-\sum_{k\ne i}\gamma^i_k H_k,&\text{if $i=j$}.
\end{cases}$$ Introducing a column-vector ${\overline{H}}:=(H_1,\ldots,H_N)$, this system can be equivalently written as $$\begin{gathered}
\label{eq:equation for oH}
d{\overline{H}}=[\Gamma,dU]{\overline{H}}.\end{gathered}$$
\[proposition:matrices R\_k\] 1. There exists a sequence of matrices $R_0={\mathrm{Id}},R_1,R_2,\ldots$ satisfying the equations $$\begin{gathered}
\label{eq:equations for R_k}
d R_{k-1}+R_{k-1}[\Gamma,dU]=[R_k,dU],\quad k\ge 1.\end{gathered}$$ 2. The matrices $R_i$ are determined uniquely up to a transformation $$\begin{gathered}
\label{eq:ambiguity of R_i}
{\mathrm{Id}}+\sum_{i\ge 1}R_i z^i\mapsto \left({\mathrm{Id}}+\sum_{i\ge 1}D_iz^i\right)\left({\mathrm{Id}}+\sum_{i\ge 1}R_i z^i\right),\end{gathered}$$ where $D_i$, $i\ge 1$, are arbitrary diagonal matrices with constant entries.
1\. The matrices $R_i$ can be recursively constructed in the following way. Suppose that the matrices $R_0={\mathrm{Id}},R_1,\ldots,R_m$, $m\ge 0$, are already constructed. We define the nondiagonal entries of $R_{m+1}$ by $$\begin{gathered}
\label{eq:non-diagonal part of R_m}
(R_{m+1})^i_j:=(R_m)^i_i\gamma^i_j-\frac{\d(R_m)^i_j}{\d u^i},\quad i\ne j,\end{gathered}$$ and then determine the diagonal entries by the equation $$\begin{gathered}
\label{eq:diagonal part of R_m}
d(R_{m+1})^i_i=-\sum_{j\ne i}(R_{m+1})^i_j\gamma^j_i(du^i-du^j),\end{gathered}$$ where the integration constants can be arbitrary. Let us check that this procedure is well defined and gives a solution of equations .\
Suppose that $n\ge 0$ steps of our procedure are well defined and produce matrices $R_0={\mathrm{Id}},R_1,\ldots,R_n$ satisfying equations with $k\le n$. Let us first check that $$d[R_n,dU]=[R_n,dU]\wedge[\Gamma,dU].$$ For $n=0$ this is trivial and for $n\ge 1$ we compute $$\begin{aligned}
d[R_n,dU]=&d\left(dR_{n-1}+R_{n-1}[\Gamma,dU]\right)=dR_{n-1}\wedge[\Gamma,dU]+R_{n-1}d[\Gamma,dU]=\label{eq:computation with R_n}\\
=&dR_{n-1}\wedge[\Gamma,dU]+R_{n-1}[\Gamma,dU]\wedge[\Gamma,dU]=[R_n,dU]\wedge[\Gamma,dU].\notag\end{aligned}$$
Then note that equation with $m=n-1$ implies that the diagonal part of the matrix $dR_n+R_n[\Gamma,dU]$ is equal to zero. Moreover, we have $$\begin{gathered}
(dR_n+R_n[\Gamma,dU])\wedge dU+dU\wedge(dR_n+R_n[\Gamma,dU])=d[R_n,dU]-[R_n,dU]\wedge[\Gamma,dU]=0,\end{gathered}$$ and, therefore, $dR_n+R_n[\Gamma,dU]=[R,dU]$ for some matrix $R$ whose nondiagonal entries are given exactly by formula , $R^i_j:=(R_n)^i_i\gamma^i_j-\frac{\d(R_n)^i_j}{\d u^i}$, $i\ne j$. In order to check that the diagonal part of $R_{n+1}$ can be defined by equation with $m=n$, we have to check that $$\begin{gathered}
d(R[\Gamma,dU])^{\mathrm{diag}}=0,\end{gathered}$$ where $(\cdot)^{\mathrm{diag}}$ denotes the diagonal part of a matrix. We compute $$d(R[\Gamma,dU])^{\mathrm{diag}}=\left((dR+R[\Gamma,dU])\wedge[\Gamma,dU]\right)^{\mathrm{diag}}$$ and it remains to check that the expression $$(dR+R[\Gamma,dU])\wedge dU+dU\wedge(dR+R[\Gamma,dU])=d[R,dU]-[R,dU]\wedge[\Gamma,dU]$$ is equal to zero, which is true by the same computation as in . This completes the proof of Part 1 of the proposition.\
2. Clearly, transformations preserve the space of solutions of equations . Suppose that a sequence of matrices $R_0={\mathrm{Id}},R_1,\ldots$ satisfies equations . Equation for a fixed $k$ determines the nondiagonal entries of the matrix $R_k$ in terms of the matrix $R_{k-1}$ and this gives formula . Since $[R_k,dU]^{\mathrm{diag}}=0$, equation determines the differential of the diagonal part of $R_{k-1}$ in terms of the nondiagonal part of $R_{k-1}$. This gives formula . So all solutions of equations are given by the procedure described in the proof of the first part of the proposition. At each step of this procedure the integration constants for the diagonal entries of $R_i$ are totally ambiguous. It is easy to check by induction that, fixing some choice of integration constants, any other choice can be obtained by a transformation of the form . This completes the proof of the proposition.
Note that equation for $k=1$ implies that $R_1-\Gamma$ is a diagonal matrix.\
Note also that, introducing the generating series $R(z):={\mathrm{Id}}+\sum_{i\ge i}R_i z^i$, the system equations can be equivalently written as $$\begin{gathered}
\label{eq:equation for R(z)}
z(dR(z)+R(z)[\Gamma,dU])=[R(z),dU].\end{gathered}$$
In order to explain the meaning of relations , let us consider the system $\nabla_{z^{-1}}{\widetilde{\xi}}=0$ for $1$-forms ${\widetilde{\xi}}=\sum_{i=1}^N{\widetilde{\xi}}_i(u^*,z) du^i$ depending on $z$ that are covariantly constant with respect to the connection $\nabla_{z^{-1}}=\nabla+\frac{1}{z}\circ$. In canonical coordinates this system reads $$\begin{aligned}
\frac{\d{\widetilde{\xi}}_i}{\d u^j}=&\Gamma^i_{ji}{\widetilde{\xi}}_i+\Gamma^j_{ji}{\widetilde{\xi}}_j,\quad j\ne i,\\
\frac{\d{\widetilde{\xi}}_i}{\d u^i}=&\sum_{l\ne i}\left(-\Gamma^i_{il}{\widetilde{\xi}}_i+\Gamma^l_{ii}{\widetilde{\xi}}_l\right)+\frac{1}{z}{\widetilde{\xi}}_i.\end{aligned}$$ Let us rewrite the above system for the unknown functions $\xi_i$ defined as ${\widetilde{\xi}}_i=H_i\xi_i$. We obtain the system $$\begin{aligned}
\frac{\d\xi_i}{\d u^j}=&\gamma^j_i\xi_j,\quad j\ne i\\
\frac{\d\xi_i}{\d u^i}=&-\sum_{l\ne i}\gamma^l_i\xi_l+\frac{1}{z}{\widetilde{\xi}}_i,\end{aligned}$$ which, introducing a row-vector ${\overline{\xi}}:=(\xi_1,\ldots,\xi_N)$, is equivalent to the equation $$\begin{gathered}
\label{eq:equation for xi}
d{\overline{\xi}}={\overline{\xi}}[dU,\Gamma]+\frac{1}{z}{\overline{\xi}}dU.\end{gathered}$$ For $\gamma^i_j=0$ (trivial flat F-manifold) a fundamental matrix of solutions of this equation is $\Xi^0=e^{U/z}$. It is straightforward to check that, looking for a fundamental matrix $\Xi$ of solutions of equation in the form $$\Xi=e^{U/z}\left({\rm Id}+\sum_{k\ge 1}R_k z^k\right),$$ one obtains system . We will see that the formal series in the brackets can be interpreted as an element of a group acting on the space of flat F-manifolds.
Flat F-manifolds and Riemannian F-manifolds
-------------------------------------------
In the previous section we saw that, around a semisimple point, a flat F-manifold possesses a metric. In this section we show that there is a correspondence between semisimple (at any point) flat F-manifolds and Riemannian F-manifolds.\
Let us first recall the local description of flat F-manifolds around a semisimple point in canonical coordinates.
\[theorem:local description of semisimple flat F-manifolds\] Let ${\widetilde{\gamma}}^i_j$, $i\ne j$, be a solution of equations , . Then the connection $\nabla$, given by equations –, the structure constants, given by $c^i_{jk}:=\delta^i_j\delta^i_k$, and the vector field $e:=\sum_i\frac{\d}{\d u^i}$ define a semisimple flat F-manifold structure, where the coordinates $u^i$ are canonical. Moreover, any flat F-manifold around a semisimple point can be obtained in this way.
The fact that any flat F-manifold around a semisimple point can be obtained in this way was proved in the previous section. The converse statement is a direct computation (see [@AL15] for details).
In the previous section we associated to any flat F-manifold around a semisimple point a metric $g=\sum_i g_i(d u^i)^2$ satisfying the condition $$\begin{gathered}
\label{eq:defining property for gi}
\frac{1}{g_i}\frac{\d g_i}{\d u^j}=2\Gamma^i_{ij},\quad 1\le i,j\le N. \end{gathered}$$ This condition determines a metric uniquely up to rescalings $g_i\mapsto\lambda_ig_i$, $\lambda_i\in{\mathbb C}^*$. Thus, we actually get a family of metrics parameterized by a vector $(\lambda_1,\ldots,\lambda_N)\in({\mathbb C}^*)^N$. This family can be described in a more invariant way, without going to the canonical coordinates.
\[proposition:two descriptions of a metric\] The family of metrics $g$ on a flat F-manifold $M$ around a semisimple point given, in canonical coordinates, by coincides with the family of metrics $g$ compatible with the product $\circ$ and satisfying the condition $$\begin{gathered}
\label{eq:invariant definition of a metric}
\nabla_kg_{ij}=\frac{1}{2}\sum_l\left(c^l_{ik}(d\theta)_{lj}+\frac{1}{2}c^l_{jk}(d\theta)_{li}\right),\quad 1\le i,j,k\le N,\end{gathered}$$ where $X,Y,Z$ are local vector fields on $M$ and $\theta$ is the counit, $\theta(\cdot)=g(e,\cdot)$.
Suppose that a metric $g=\sum_i g_i(du^i)^2$ satisfies condition . Note that $\theta=\sum_i g_i du^i$. Then the proof of property becomes a simple direct computation based on the expression of the Christoffel symbols $\Gamma^i_{jk}$ in terms of the functions $g_i$.\
Suppose now that a metric $g=\frac{1}{2}\sum_{i,j} g_{ij}du^idu^j$ is compatible with the product $\circ$ and satisfies condition . The compatibility with the product $\circ$ immediately implies that $g_{ij}=0$ for $i\ne j$. For $i=j$ the right-hand side of is zero, while the left-hand side is $\frac{\d g_{ii}}{\d u^k}-2\Gamma^i_{ik}g_{ii}$, which gives equation for $g_{ii}=g_i$.
[@HM99] An F-manifold $(M,\circ,e)$ is the datum of an analytic manifold $M$, a commutative associative algebra structure $(T_pM,\circ)$ on each tangent space analytically depending on the point $p\in M$ and a unit vector field $e$ such that the condition $$\begin{gathered}
\label{eq:HM condition,invariant}
{\mathrm{Lie}}_{X\circ Y}(\circ)=X\circ{\mathrm{Lie}}_Y(\circ)+Y\circ{\mathrm{Lie}}_X(\circ)\end{gathered}$$ is satisfied for any local vector fields $X,Y$ on $M$. The above condition is called the Hertling–Manin condition.
If $t^1,\ldots,t^N$ are some coordinates on $M$, then condition is equivalent to the following condition for the structure constants $c^i_{jk}$ of the multiplication $\circ$: $$\begin{gathered}
\sum_{s=1}^N\left(\frac{\d c^k_{jl}}{\d t^s}c^s_{im}+\frac{\d c^s_{im}}{\d t^j}c^k_{sl}-\frac{\d c^k_{im}}{\d t^s}c^s_{jl}-\frac{\d c^s_{jl}}{\d t^i}c^k_{sm}-\frac{\d c^s_{jl}}{\d t^m}c^k_{si}+\frac{\d c^s_{mi}}{\d t^l}c^k_{js}\right)=0,\quad 1\le i,j,k,l,m\le N.\end{gathered}$$
In the remaining part of this section we will focus on the semisimple case describing the relation between flat F-manifolds and a special class of F-manifolds called Riemannian F-manifolds.
A semisimple (pseudo-)Riemannian F-manifold $(M,g,\circ,e)$ is the datum of a semisimple F-manifold $(M,\circ,e)$ equipped with a metric $g$ compatible with the product $\circ$ and such that $$\begin{gathered}
\label{eq:condition for Riemannian F-manifold}
Z\circ R(W,Y)(X)+W\circ R(Y,Z)(X)+Y\circ R(Z,W)(X)=0,\end{gathered}$$ where $R$ is the curvature operator for the Levi-Civita connection ${\widetilde{\nabla}}$ associated to $g$ and $X,Y,W,Z$ are local vector fields on $M$. If also the condition $$\begin{gathered}
{\mathrm{Lie}}_e g=0\end{gathered}$$ is satisfied, then the manifold is called a Riemannian F-manifold with Killing unit vector field.
The notion of a Riemannian F-manifold appears also in [@LPR09] and [@DS11]. In both cases the definition involves some extra conditions. In [@LPR09] the connection and the product are required to satisfy the condition ${\widetilde{\nabla}}_l c^i_{jk}=\nabla_j c^i_{lk}$, while in [@DS11] the counit $\theta$ is required to be closed.
We have the following theorem.
1\. Consider a flat F-manifold $(M,\nabla,\circ,e)$ around a semisimple point and a metric $g$ compatible with the product $\circ$ and satisfying . Then the tuple $(M,g,\circ,e)$ is a Riemannian F-manifold with Killing unit vector field.\
2. Consider a Riemannian F-manifold $(M,g,\circ,e)$ with Killing unit vector field around a semisimple point. Then there exists a unique torsionless connection satisfying . With this connection the tuple $(M,\nabla,\circ,e)$ is a flat F-manifold.
Consider a Riemannian F-manifold $(M,g,\circ,e)$ around a semisimple point. Let $u^1,\ldots,u^N$ be the canonical coordinates. The compatibility with the product $\circ$ implies that the metric $g$ has the form $g=\sum_i g_i(du^i)^2$ and, therefore, the Christoffel symbols ${\widetilde{\Gamma}}^i_{jk}$ of the connection ${\widetilde{\nabla}}$ are given by $$\begin{aligned}
{\widetilde{\Gamma}}^i_{jk}&=0,&& i\ne j\ne k\ne i,\\
{\widetilde{\Gamma}}^i_{ij}&=\frac{1}{2g_i}\frac{\d g_i}{\d u^j},&& 1\le i,j\le N,\\
{\widetilde{\Gamma}}^j_{ii}&=-\frac{1}{2g_j}\frac{\d g_i}{\d u^j},&& i\ne j.\end{aligned}$$
Equation reads $$\begin{gathered}
\label{eq:Riemann tensor condition,1}
R^m_{lij}\delta^m_{k}+R^m_{lki}\delta^m_{j}+R^m_{ljk}\delta^m_{i}=0,\quad 1\le i,j,k,l,m\le N.\end{gathered}$$ Note that because of the skewsymmetry of the Riemann tensor with respect to the second and the third lower indices the condition above is satisfied if some of the indices $i,j,k$ coincide. Note also that condition is trivially satisfied when $m$ is distinct from $i,j,k$. Therefore, condition is nontrivial only if the indices $i,j,k$ are distinct and $m$ coincides with one of them, which gives the system $$R^k_{lij}=0,\quad k\ne i\ne j\ne k.$$
The vanishing of the Christoffel symbols ${\widetilde{\Gamma}}^i_{jk}$ with distinct indices $i,j,k$ implies that $R^k_{lij}$ vanishes if all the indices $i,j,k,l$ are distinct. Therefore, we come to the system $$\begin{gathered}
R^k_{kij}=0,\qquad R^k_{iij}=0,\qquad k\ne i\ne j\ne k.\end{gathered}$$ The skewsymmetry of the tensor $R_{ijkl}=g_i R^i_{jkl}$ with the respect to the first two lower indices implies that $R^k_{kij}=0$. Thus, condition is equivalent to the system $$R^k_{iij}=0,\qquad k\ne i\ne j\ne k.$$ Since $$\begin{gathered}
R^k_{iij}=-\frac{\d{\widetilde{\Gamma}}^k_{ii}}{\d u^j}+{\widetilde{\Gamma}}^k_{ii}{\widetilde{\Gamma}}^i_{ij}-{\widetilde{\Gamma}}^k_{jj}{\widetilde{\Gamma}}^j_{ii}-{\widetilde{\Gamma}}^k_{kj}{\widetilde{\Gamma}}^k_{ii},\quad i\ne j\ne k\ne i,\end{gathered}$$ we conclude that the datum of a Riemannian F-manifold in canonical coordinates is equivalent to a diagonal metric $g=\sum_i g_i(du^i)^2$ satisfying the Darboux-Tsarev system $$\begin{gathered}
\label{eq:Darboux-Tsarev system}
-\frac{\d{\widetilde{\Gamma}}^k_{ii}}{\d u^j}+{\widetilde{\Gamma}}^k_{ii}{\widetilde{\Gamma}}^i_{ij}-{\widetilde{\Gamma}}^k_{jj}{\widetilde{\Gamma}}^j_{ii}-{\widetilde{\Gamma}}^k_{kj}{\widetilde{\Gamma}}^k_{ii}=0,\quad i\ne j\ne k\ne i.\end{gathered}$$
Let us prove Part 1 of the theorem. Consider a flat F-manifold $(M,\nabla,\circ,e)$ and a metric $g$ compatible with the product $\circ$ and satisfying . By Proposition \[proposition:two descriptions of a metric\], in canonical coordinates we have $g=\sum_i g_i(du^i)^2$, where the $g_i$’s satisfy . Note that $$\begin{aligned}
&{\widetilde{\Gamma}}^i_{ij}=\Gamma^i_{ij}={\widetilde{\gamma}}^i_j,\qquad{\widetilde{\Gamma}}^j_{ii}=-\frac{g_i}{g_j}{\widetilde{\Gamma}}^i_{ij}=-\frac{g_i}{g_j}{\widetilde{\gamma}}^i_j,&& i\ne j,\\
&{\widetilde{\Gamma}}^i_{ii}=\Gamma^i_{ii}=-\sum_{j\ne i}{\widetilde{\gamma}}^i_j.\end{aligned}$$ Using these formulas it is now easy to check that system follows from system . Since $\sum_j\frac{1}{g_i}\frac{\d g_i}{\d u^j}=\sum_j\Gamma^i_{ij}=0$, the unit vector field is Killing.\
Let us prove Part 2 of the theorem. Consider a Riemannian F-manifold $(M,g,\circ,e)$ with Killing unit vector field. Let ${\widetilde{\gamma}}^i_j:={\widetilde{\Gamma}}^i_{ij}$, $i\ne j$. It is easy to check that the Darboux–Tsarev system implies that the functions ${\widetilde{\gamma}}^i_j$ satisfy system . The fact that the unit vector field $e$ is Killing implies that $$\sum_k\frac{\d g_i}{\d u^k}=0\Rightarrow\sum_k\frac{\d{\widetilde{\Gamma}}^i_{ij}}{\d u^k}=0.$$ Therefore, equation is also satisfied and, thus, by Theorem \[theorem:local description of semisimple flat F-manifolds\], the connection $\nabla$ given by equations – defines a flat F-manifold $(M,\nabla,\circ,e)$. The fact that $\nabla$ satisfies condition or, equivalently, condition is obvious.\
It remains to check that condition determines a connection $\nabla$ uniquely. Denote the tensor on the right-hand side of this equation by $\Delta_{kij}$. Let us write three equations corresponding to the cyclic permutations of the indices $i,j,k$: $$\begin{aligned}
&\frac{\d g_{ij}}{\d u^k}-\Gamma_{kij}-\Gamma_{kji}=\Delta_{kij},\\
&\frac{\d g_{ki}}{\d u^j}-\Gamma_{jki}-\Gamma_{jik}=\Delta_{jki},\\
&\frac{\d g_{jk}}{\d u^i}-\Gamma_{ijk}-\Gamma_{ikj}=\Delta_{ijk},\end{aligned}$$ where $\Gamma_{kij}:=\sum_s\Gamma^s_{ki}g_{sj}$. Summing the first and the third equations and subtracting the second one, we get $$\Gamma_{kij}=\frac{1}{2}\left(\frac{\d g_{ij}}{\d u^k}-\frac{\d g_{ik}}{\d u^j}+\frac{\d g_{jk}}{\d u^i}-\Delta_{kij}+\Delta_{jik}-\Delta_{ijk}\right),$$ which completes the proof of the theorem.
Homogeneous flat F-manifolds
----------------------------
In this section we will prove that the metric, constructed in Section \[subsection:metric for a flat F-manifold\], in the case of homogeneous flat F-manifolds satisfies an additional homogeneity property. We will also show that an additional homogeneity property allows to fix uniquely the choice of $R$-matrices from Proposition \[proposition:matrices R\_k\].\
Consider a flat F-manifold $(M,\nabla,\circ,e)$ around a semisimple point. Let $t^1,\ldots,t^N$ be flat coordinates. Suppose that our flat F-manifold is homogeneous with an Euler vector field $$E=\sum_{\alpha=1}^N((1-q_\alpha) t^\alpha+r^\alpha)\frac{\d}{\d t^\alpha},\quad q_\alpha,r^\alpha\in{\mathbb C}.$$ In canonical coordinates we have $E=\sum_i(u^i+a^i)\frac{\d}{\d u^i}$ for some $a^i\in{\mathbb C}$. After an appropriate shift of the coordinates we can assume that $$E=\sum_i u^i\frac{\d}{\d u^i}.$$ Consider the matrices ${\widetilde{\Psi}},{\widetilde{\Gamma}}=({\widetilde{\gamma}}^i_j),\Gamma=(\gamma^i_j)$, the diagonal matrix ${\widetilde{D}}$ (consisting of one-forms) and the metric $g=\sum_i g_i(du^i)^2$, $g_i=H_i^2$, constructed in Section \[subsection:metric for a flat F-manifold\].
\[proposition:homogeneity of H and gamma\] 1. The diagonal matrix $i_E{\widetilde{D}}$ is constant, $i_E{\widetilde{D}}=-{\mathrm{diag}}(\delta_1,\ldots,\delta_N)$, $\delta_i\in{\mathbb C}$.\
2. We have $$\begin{gathered}
\label{eq:homogeneity for H and Gamma}
\sum_j u^j\frac{\d H_i}{\d u^j}=\delta_i H_i,\qquad \sum_k u^k\frac{\d\gamma^i_j}{\d u^k}=(\delta_i-\delta_j-1)\gamma^i_j.\end{gathered}$$
Define a diagonal matrix $Q$ by $Q:={\mathrm{diag}}(q_1,\ldots,q_N)$. In [@BB19] the authors introduced a family of connections ${\widetilde{\nabla}}^\lambda$ on $M\times{\mathbb C}^*$, depending on a complex parameter $\lambda$, by $$\begin{aligned}
{\widetilde{\nabla}}^\lambda_X Y:=&\nabla_X Y+z X\circ Y,\notag\\
{\widetilde{\nabla}}^\lambda_{\frac{\d}{\d z}}Y:=&\frac{\d Y}{\d z}+E\circ Y+\frac{\lambda-Q}{z}Y,\label{eq:connection tnabla,2}\\
{\widetilde{\nabla}}^\lambda_X\frac{\d}{\d z}={\widetilde{\nabla}}^\lambda_{\frac{\d}{\d z}}\frac{\d}{\d z}:=&0,\notag\end{aligned}$$ where $z$ is the coordinate on ${\mathbb C}^*$ and $X,Y$ are local vector fields on $M\times{\mathbb C}^*$ having zero component along ${\mathbb C}^*$. The authors proved that the connection ${\widetilde{\nabla}}^\lambda$ is flat for any value of $\lambda$. Let us show how to derive the proposition from the flatness of ${\widetilde{\nabla}}^\lambda$.\
Note that equation can be rewritten as $${\widetilde{\nabla}}^\lambda_{\frac{\d}{\d z}}Y=\frac{\d Y}{\d z}+E\circ Y+\frac{1}{z}\nabla_Y E+\frac{\lambda-1}{z}Y.$$ In canonical coordinates we have $$\nabla_Y E=\sum_i Y^i\frac{\d}{\d u^i}+\left({\widetilde{\Psi}}\sum_iY^i\frac{\d({\widetilde{\Psi}}^{-1})}{\d u^i}\right)E,$$ and it is easy to check that $$\left({\widetilde{\Psi}}\sum_iY^i\frac{\d({\widetilde{\Psi}}^{-1})}{\d u^i}\right)E=\left({\widetilde{\Psi}}\sum_iu^i\frac{\d({\widetilde{\Psi}}^{-1})}{\d u^i}\right)Y.$$ For an $N\times N$ matrix $A=(A^i_j)$ denote by ${\mathrm{Ext}}(A)$ the $(N+1)\times(N+1)$ matrix defined by $${\mathrm{Ext}}(A)^i_j:=
\begin{cases}
A^i_j,&\text{if $1\le i,j\le N$},\\
0,&\text{otherwise}.
\end{cases}$$ We see that in canonical coordinates the connection ${\widetilde{\nabla}}^\lambda$ is given by $${\widetilde{\nabla}}^\lambda=d+{\mathrm{Ext}}\left(-d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}+zdU+\left(U+\frac{\lambda}{z}-\frac{1}{z}\left(\sum_iu^i\frac{\d{\widetilde{\Psi}}}{\d u^i}\right){\widetilde{\Psi}}^{-1}\right)dz\right).$$
The flatness of the connection ${\widetilde{\nabla}}^\lambda$ for any value of $\lambda$ is equivalent to equation together with the equations $$\begin{aligned}
dB+[B,M]=&0,\label{eq:B and M}\\
[U,M]+[B,dU]=&0,\notag\end{aligned}$$ where $M:=d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}$ and $B:=i_E M=\left(\sum_iu^i\frac{\d{\widetilde{\Psi}}}{\d u^i}\right){\widetilde{\Psi}}^{-1}$. Equation implies that $$dB+i_E(M\wedge M)=0\stackrel{\text{by eq.~\eqref{eq:equations for M}}}{\Rightarrow}dB+i_E(dM)=0.$$ The diagonal part of $dM$ is $d{\widetilde{D}}=0$. Therefore, the diagonal part of $dB$, which is $d(i_E{\widetilde{D}})$, vanishes. This proves the first part of the proposition.\
Denote $\Delta:=-i_E{\widetilde{D}}$. Since $H$ is defined by $dH\cdot H^{-1}=-{\widetilde{D}}$, we get $\sum_i u^i\frac{\d H}{\d u^i}=\Delta H$, which is exactly the first equation in . We see that $B=-\Delta+[{\widetilde{\Gamma}},U]$. Therefore, equation implies that $$d[{\widetilde{\Gamma}},U]+[-\Delta+[{\widetilde{\Gamma}},U],{\widetilde{D}}+[{\widetilde{\Gamma}},dU]]=0.$$ Substituting ${\widetilde{\Gamma}}=H^{-1}\Gamma H$, we get $$d[\Gamma,U]=[[\Delta,\Gamma],dU]+[[\Gamma,dU],[\Gamma,U]].$$ Applying the contraction $i_E$ to both sides of this equation we get $$\left[\sum_iu^i\frac{\d\Gamma}{\d u^i},U\right]+[\Gamma,U]=[[\Delta,\Gamma],U]\Rightarrow\sum_iu^i\frac{\d\Gamma}{\d u^i}=[\Delta,\Gamma]-\Gamma,$$ which is exactly the second equation in . This completes the proof of the proposition.
If our homogeneous flat F-manifold is a conformal Frobenius manifold of conformal dimension $\delta$, meaning that ${\mathrm{Lie}}_E\eta=(2-\delta)\eta$, then $\delta_i=-\frac{\delta}{2}$, $1\le i\le N$.
\[proposition:unique R-matrix in the homogeneous case\] There exists a unique sequence of matrices $R_0={\mathrm{Id}},R_1,R_2,\ldots$ satisfying the differential equation and the homogeneity condition $$\begin{gathered}
\label{eq:homogeneity for Rk}
\sum_s u^s\frac{\d(R_k)^i_j}{\d u^s}=(\delta_i-\delta_j-k)(R_k)^i_j,\quad k\ge 0.\end{gathered}$$
As we know from the proof of Proposition \[proposition:matrices R\_k\], without requiring property a sequence of matrices $R_k$ is recursively determined in the following way. Suppose that matrices $R_0,R_1,\ldots,R_m$, $m\ge 0$, are already constructed. Then the nondiagonal entries of $R_{m+1}$ are given by $$\begin{gathered}
\label{eq:nondiagonal entries of R}
(R_{m+1})^i_j=(R_m)^i_i\gamma^i_j-\frac{\d(R_m)^i_j}{\d u^i},\quad i\ne j.\end{gathered}$$ The diagonal entries of $R_{m+1}$ are determined by the equation $$\begin{gathered}
\label{eq:diagonal entries of R}
\frac{\d(R_{m+1})^i_i}{\d u^j}=
\begin{cases}
(R_{m+1})^i_j\gamma^j_i,&\text{if $j\ne i$},\\
-\sum_{k\ne i}(R_{m+1})^i_k\gamma^k_i,&\text{if $j=i$},
\end{cases}\quad 1\le i,j\le N,\end{gathered}$$ uniquely up to constants, which can be arbitrary. Suppose that the matrix $R_m$ satisfies the homogeneity condition . From formula and Proposition \[proposition:homogeneity of H and gamma\] it follows that the nondiagonal entries of $R_{m+1}$ also satisfy the homogeneity condition . Formula implies that $\sum_s u^s\frac{\d}{\d u^s}\frac{\d(R_{m+1})^i_i}{\d u^j}=-(m+2)\frac{\d(R_{m+1})^i_i}{\d u^j}$ for any $i$ and $j$. Thus, unique functions $(R_{m+1})^i_i$ satisfying equations and are given by $$(R_{m+1})^i_i=-\frac{1}{m+1}\sum_{j\ne i}(u^j-u^i)(R_{m+1})^i_j\gamma^j_i.$$ The proposition is proved.
In [@AL13] the authors introduced the notion of a bi-flat F-manifold, which is the datum of two different flat F-manifold structures $(\nabla,\circ,e)$ and $(\nabla^{*},*,E)$ on the same manifold $M$ intertwined by the following conditions:
1. $[e,E]=e$, ${\mathrm{Lie}}_E(\circ)=\circ$;
2. $X*Y=(E\circ)^{-1}\,X\circ Y$ (or $X\circ Y=(e*)^{-1}X*Y$) for all local vector fields $X,Y$ on $M$;
3. $(d_{\nabla}-d_{\nabla^{*}})(X\,\circ)=0$ for all local vector fields $X$ on $M$, where $d_{\nabla}$ is the exterior covariant derivative.
Using the fact that the connection $\nabla^*$ can be expressed in terms of the other data, in [@AL17] the authors proved that in the semisimple case the structure of a bi-flat F-manifold is equivalent to the datum of a flat F-manifold, equipped with an invertible vector field $E$ satisfying condition (1) from the list above and also the property $\nabla\nabla E=0$ (see [@KMS18] for the discussion of the regular case). In other words on the complement of the discriminant (i.e. at the points where $E$ is invertible) any regular homogeneous flat F-manifold is equipped with a bi-flat structure.
The structure $(\nabla,\circ,e,E)$ where $E$ is a linear Euler vector field can be also characterized in terms of a flat meromorphic connection on the bundle $\pi^*TM$ on $\mathbb{P}\times M$ called Saito structure without metric (see [@Sab98] for details). In particular, assuming the existence of flat coordinates diagonalizing the matrix $\nabla E$, it turns out that Saito structures without metric are equivalent to homogeneous flat structures.
### Example: extended $2$-spin theory {#subsubsection:example1}
An example of a flat F-manifold is given by the extended $r$-spin theory, constructed in [@JKV01] and then studied in details in [@BCT19; @Bur18; @BR18]. Let us consider the case $r=2$ and compute the metric, the rotation coefficients and the $R$-matrices.\
The vector potential of the flat F-manifold of the extended $2$-spin theory is given by (see e.g. [@BR18 Section 4.3]) $$F^1(t^1,t^2)=\frac{(t^1)^2}{2},\qquad F^2(t^1,t^2)=t^1t^2-\frac{(t^2)^3}{12}.$$ The unit is $\frac{\d}{\d t^1}$. The flat F-manifold is homogeneous with the Euler vector field $$E=t^1\frac{\d}{\d t^1}+\frac{1}{2}t^2\frac{\d}{\d t^2}.$$ The set of nonsemisimple points coincides with the $t^1$-axis. Consider our flat F-manifold around a point $(0,\tau)$, $\tau\ne 0$. The canonical coordinates satisfy the system of differential equations $$\frac{\d u^i}{\d t^\alpha}\frac{\d u^i}{\d t^\beta}=c_{\alpha\beta}^\gamma\frac{\d u^i}{\d t^\gamma},\quad 1\le\alpha,\beta,i\le 2,$$ from which we find $$\left\{
\begin{aligned}
u^1=&t^1,\\
u^2=&t^1-\frac{(t^2)^2}{2},
\end{aligned}
\right.\qquad
\left\{
\begin{aligned}
t^1=&u^1,\\
t^2=&\sqrt{2(u^1-u^2)},
\end{aligned}
\right.\qquad
{\widetilde{\Psi}}=\begin{pmatrix}
1 & 0\\
1 & -t^2
\end{pmatrix}.$$ We then compute $$d{\widetilde{\Psi}}\cdot{\widetilde{\Psi}}^{-1}=\begin{pmatrix}
0 & 0\\
-\frac{dt^2}{t^2} & \frac{dt^2}{t^2}
\end{pmatrix},
\qquad
dU=\begin{pmatrix}
dt^1 & 0\\
0 & dt^1-t^2dt^2
\end{pmatrix},$$ and find $${\widetilde{\Gamma}}=\begin{pmatrix}
0 & 0\\
-\frac{1}{(t^2)^2} & 0
\end{pmatrix},
\qquad
H=\begin{pmatrix}
\lambda_1 & 0\\
0 & \frac{\lambda_2}{t^2}
\end{pmatrix},\quad\lambda_1,\lambda_2\in{\mathbb C}^*.$$ Let us choose the parameters $\lambda_1,\lambda_2$ such that $H|_{(t^1,t^2)=(0,\tau)}={\mathrm{Id}}$. We obtain $$H=
\begin{pmatrix}
1 & 0\\
0 & \frac{\tau}{t^2}
\end{pmatrix},\qquad
\Gamma=
\begin{pmatrix}
0 & 0\\
-\frac{\tau}{(t^2)^3} & 0
\end{pmatrix},\qquad
\Psi=
\begin{pmatrix}
1 & 0\\
\frac{\tau}{t^2} & -\tau
\end{pmatrix}.$$ We see that $$\delta_1=0,\qquad\delta_2=-\frac{1}{2}.$$ It is easy to check that a unique sequence of $R$-matrices given by Proposition \[proposition:unique R-matrix in the homogeneous case\] is the following: $$R_m=\begin{pmatrix}
0 & 0\\
(-1)^m\tau\frac{(2m-1)!!}{(t^2)^{2m+1}} & 0
\end{pmatrix},\quad m\ge 1.$$ In particular, we have $$\Psi\big|_{(t^1,t^2)=(0,\tau)}=
\begin{pmatrix}
1 & 0\\
1 & -\tau
\end{pmatrix},\qquad
R_m\big|_{(t^1,t^2)=(0,\tau)}=\begin{pmatrix}
0 & 0\\
(-1)^m\frac{(2m-1)!!}{\tau^{2m}} & 0
\end{pmatrix},\quad m\ge 1.$$\
Givental-type theory for flat F-manifolds {#section:Givental theory for flat F-manifolds}
=========================================
In this section we recall the notion of a calibrated flat F-manifold, introduced in [@BB19], and interpret such an object as a certain infinite dimensional submanifold in the space ${\mathbb C}^N[[z,z^{-1}]]$. This generalizes a similar result of A. Givental [@Giv04] about Frobenius manifolds and allows to introduce a group action on the space of calibrated flat F-manifolds. We then prove that this group action, combined with linear changes of coordinates, is transitive on the space of semisimple calibrated flat F-manifolds.\
Let us fix a point ${\overline{t}}_{\mathrm{orig}}=(t^1_{\mathrm{orig}},\ldots,t^N_{\mathrm{orig}})\in{\mathbb C}^N$ and denote by ${\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}$ the ring of formal power series in the shifted variables $t^\alpha-t^\alpha_{\mathrm{orig}}$. In this section we consider flat F-manifolds defined on a formal neighbourhood of ${\overline{t}}_{\mathrm{orig}}$, which means that the functions describing the structure of our flat F-manifolds belong to the ring ${\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}$.
Calibrated flat F-manifolds and descendant vector potentials {#subsection:calibrated flat F-manifold}
------------------------------------------------------------
Consider a flat F-manifold structure given by a vector potential ${\overline{F}}=(F^1,\ldots,F^N)$, $F^\alpha\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}$, and a unit $A^\alpha\frac{\d}{\d t^\alpha}$, $A^\alpha\in{\mathbb C}$. We will often denote the unit by $\frac{\d}{\d t^\un}$. A [*calibration*]{} of our flat F-manifold is a collection of functions $\Omega^{\alpha,d}_{\beta,0}\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}$, $1\le\alpha,\beta\le N$, $d\ge -1$, satisfying $\Omega^{\alpha,-1}_{\beta,0}=\delta^\alpha_\beta$ and the property $$\begin{gathered}
\label{eq:property of a calibration}
\frac{\d\Omega^{\alpha,d}_{\beta,0}}{\d t^\gamma}=c^\mu_{\gamma\beta}\Omega^{\alpha,d-1}_{\mu,0}, \quad d\ge 0,\end{gathered}$$ where $c^\alpha_{\beta\gamma}=\frac{\d^2 F^\alpha}{\d t^\beta\d t^\gamma}$. Introduce $N\times N$ matrices $\Omega^d_0$, $d\ge -1$, by $(\Omega^d_0)^\alpha_\beta:=\Omega^{\alpha,d}_{\beta,0}$. Equation implies that $c^\alpha_{\beta\gamma}=\frac{\d\Omega^{\alpha,0}_{\beta,0}}{\d t^\gamma}$ and, thus, equation can be written in the matrix form as $$d\Omega^p_0=\Omega^{p-1}_0\cdot d\Omega^0_0,\quad p\ge 0,$$ where $d(\cdot)$ denotes the full differential. A calibration is determined uniquely up to a transformation $$\begin{gathered}
\label{eq:transformation of calibration}
\left({\mathrm{Id}}+\sum_{d\ge 1}\Omega^{d-1}_0 z^d\right)\mapsto G(z)\left({\mathrm{Id}}+\sum_{d\ge 1}\Omega^{d-1}_0 z^d\right),\quad G(z)\in{\mathrm{End}}({\mathbb C}^N)[[z]],\quad G(0)={\mathrm{Id}}.\end{gathered}$$ A flat F-manifold together with a calibration is called a [*calibrated flat F-manifold*]{}.\
To our calibrated flat F-manifold one can associate a two-parameter family of matrices $\Omega^p_q$, $p,q\ge 0$, in the following way. Let us introduce matrices $\Omega^0_d=(\Omega^{\alpha,0}_{\beta,d})$, $d\ge 0$, by the equation $$\begin{gathered}
\label{eq:upper-lower relation}
\left({\mathrm{Id}}+\sum_{d\ge 1}(-1)^d\Omega^0_{d-1}z^d\right)\left({\mathrm{Id}}+\sum_{d\ge 1}\Omega^{d-1}_0 z^d\right)={\mathrm{Id}}.\end{gathered}$$ By definition, we put $\Omega^0_{-1}:={\mathrm{Id}}$. The matrices $\Omega^0_p$ satisfy the equation $$\begin{gathered}
\label{eq:TRR with differential}
d\Omega^0_p=d\Omega^0_0\cdot\Omega^0_{p-1},\quad p\ge 0.\end{gathered}$$ We define matrices $\Omega^p_q=(\Omega^{\alpha,p}_{\beta,q})$, $p,q\ge 0$, by $$\begin{gathered}
\label{eq:definition of Omegapq}
\Omega^p_q:=\sum_{i=0}^q(-1)^{q-i}\Omega^{p+q-i}_0\Omega^0_{i-1}\stackrel{\scriptsize{\text{eq. \eqref{eq:upper-lower relation}}}}{=}\sum_{i=0}^p(-1)^{p-i}\Omega^{i-1}_0\Omega^0_{p+q-i},\end{gathered}$$
Let us present the construction of the descendant vector potentials associated to our calibrated flat F-manifold. We will use the notation $\Omega^{\alpha,p}_{\un,q}:=A^\beta\Omega^{\alpha,p}_{\beta,q}$. Equation implies that $\Omega^{\alpha,0}_{\un,0}-t^\alpha$ is a constant, $\Omega^{\alpha,0}_{\un,0}=t^\alpha+c^\alpha$, ${{\overline{c}}}=(c^1,\ldots,c^N)\in{\mathbb C}^N$. Let $$(\Omega^{{\overline{c}}})^{\alpha,p}_{\beta,q}:=\left.\Omega^{\alpha,p}_{\beta,q}\right|_{t^\gamma\mapsto t^\gamma-c^\gamma}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}}.$$ Consider the [*principal hierarchy*]{} associated to our calibrated flat F-manifold: $$\begin{gathered}
\label{eq:principal hierarchy for F-man}
\frac{\d v^\alpha}{\d t^\beta_d}=\d_x\left(\left.(\Omega^{{\overline{c}}})^{\alpha,0}_{\beta,d}\right|_{t^\gamma=v^\gamma}\right),\quad 1\le\alpha,\beta\le N,\quad d\ge 0.\end{gathered}$$ The flows of the principal hierarchy pairwise commute. Since $(\Omega^{{\overline{c}}})^{\alpha,0}_{\un,0}=t^\alpha$, we can identify the flows $\frac{\d}{\d x}$ and $\frac{\d}{\d t^\un_0}=A^\alpha\frac{\d}{\d t^\alpha_0}$. Clearly, the functions $v^\alpha=t^\alpha_0$ satisfy the subsystem of system given by the flows $\frac{\d}{\d t^\beta_0}$. Denote by $(v^{\mathrm{top}})^\alpha\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}+{{\overline{c}}}}[[t^*_{\ge 1}]]$ the solution of the principal hierarchy specified by the initial condition $$\left.(v^{\mathrm{top}})^\alpha\right|_{t^*_{\ge 1}=0}=t^\alpha_0.$$ It is often called the [*topological solution*]{}.\
Let $$(\Omega^{\mathrm{top}})^{\alpha,p}_{\beta,q}:=\left.(\Omega^{{\overline{c}}})^{\alpha,p}_{\beta,q}\right|_{t^\gamma\mapsto(v^{\mathrm{top}})^\gamma}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}}[[t^*_{\ge 1}]].$$ The [*descendant vector potentials*]{} ${\overline{\mathcal{F}}}^a=({\mathcal{F}}^{1,a},\ldots,{\mathcal{F}}^{N,a})$, $a\ge 0$, of our calibrated flat F-manifold are defined by $${\mathcal{F}}^{\alpha,a}:=\sum_{b\ge 0}(\Omega^{\mathrm{top}})^{\alpha,a}_{\beta,b}q^\beta_b\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}}[[t^*_{\ge 1}]],\quad a\ge 0,$$ where $q^\beta_b:=t^\beta_b-A^\beta\delta_{b,1}$. We have the property $$\begin{gathered}
\label{eq:derivative of descendant vector potential}
\frac{\d{\mathcal{F}}^{\alpha,a}}{\d t^\beta_b}=(\Omega^{\mathrm{top}})^{\alpha,a}_{\beta,b}.\end{gathered}$$ The function $F^\alpha$ coincides with the function $\left.{\mathcal{F}}^{\alpha,0}\right|_{\substack{t^\beta\mapsto t^\beta+c^\beta\\t^*_{\ge 1}=0}}$ up to an affine function in the variables $t^\gamma$.
\[remark:about shifts\] We see that in the construction of the descendant vector potentials he have to go from the functions $\Omega^{\alpha,p}_{\beta,q}$ to the shifted functions $(\Omega^{{\overline{c}}})^{\alpha,p}_{\beta,q}$. Actually, if we start from the shifted vector potential $((F^{{\overline{c}}})^1,\ldots,(F^{{\overline{c}}})^N)$ given by $(F^{{\overline{c}}})^\alpha:=F^\alpha|_{t^\beta\mapsto t^\beta-c^\beta}$, then the functions $(\Omega^{{\overline{c}}})^{\alpha,p}_{\beta,0}$ define a calibration and in the further construction of the descendant vector potentials we don’t have to do any shifts and get the same functions ${\mathcal{F}}^{\alpha,a}$. That is why in [@BB19] the authors don’t write explicitly the shifts needed in the construction of the descendant vector potetentials. We do it because we want to study the action of transformations on descendant vector potentials and such a transformation shifts a point around which formal functions ${\mathcal{F}}^{\alpha,a}$ are defined.
Let us adopt the convention $${\mathcal{F}}^{\alpha,a}:=(-1)^{a+1}q^\alpha_{-a-1},\quad\text{if $a<0$}.$$
\[proposition:descendant vector potentials,equivalent approach\] A sequence of $N$-tuples of functions $({\mathcal{F}}^{1,a},\ldots,{\mathcal{F}}^{N,a})$, ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}[[t^*_{\ge 1}]]$, $a\ge 0$, is a sequence of descendant vector potentials of a flat F-manifold if and only if the following equations are satisfied: $$\begin{aligned}
\sum_{b\ge 0}q^\beta_{b+1}\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\beta_b}=&-{\mathcal{F}}^{\alpha,a-1},&& a\in{\mathbb Z},\label{eq:vector string equation}\\
\sum_{b\ge 0}q^\beta_b\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\beta_b}=&{\mathcal{F}}^{\alpha,a},&& a\in{\mathbb Z},\label{eq:vector dilaton equation}\\
\frac{\d^2{\mathcal{F}}^{\alpha,0}}{\d q^\beta_{b+1}\d q^\gamma_c}=&\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\beta_b}\frac{\d^2{\mathcal{F}}^{\alpha,0}}{\d q^\mu_0\d q^\gamma_c},&& b,c\ge 0,\label{eq:basic vector TRR}\\
\frac{\d{\mathcal{F}}^{\alpha,a+1}}{\d q^\beta_b}+\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\beta_{b+1}}=&\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\mu_0}\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\beta_b},&& a,b\ge 0.\label{eq:important vector relation}\end{aligned}$$
It is not hard to check that equations and imply the following generalizations of equation : $$\begin{aligned}
\frac{\d^2{\mathcal{F}}^{\alpha,a}}{\d q^\beta_{b+1}\d q^\gamma_c}=&\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\beta_b}\frac{\d^2{\mathcal{F}}^{\alpha,a}}{\d q^\mu_0\d q^\gamma_c},&& a,b,c\ge 0,\label{eq:vector TRR1}\\
\frac{\d^2{\mathcal{F}}^{\alpha,a+1}}{\d q^\beta_b\d q^\gamma_c}=&\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\mu_0}\frac{\d^2{\mathcal{F}}^{\mu,0}}{\d q^\beta_b\d q^\gamma_c},&& a,b,c\ge 0.\label{eq:vector TRR2}\end{aligned}$$
Suppose that $N$-tuples $({\mathcal{F}}^{1,a},\ldots,{\mathcal{F}}^{N,a})$, $a\ge 0$, are descendant vector potentials of a flat F-manifold. Equation was proved in [@BB19]. Also we have $\sum_{b\ge 0}q^\beta_b\frac{\d (v^{{\mathrm{top}}})^\alpha}{\d q^\beta_b}=0$ [@BB19], which implies that $\sum_{b\ge 0}q^\beta_b\frac{\d(\Omega^{\mathrm{top}})^{\alpha,a}_{\gamma,c}}{\d q^\beta_b}=0$. Therefore, equation is true. Equation follows from equations and . The last equation follows from equation .\
Suppose now that functions ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}[[t^*_{\ge 1}]]$ satisfy equations –. Then it is straightforward to check that the functions $F^\alpha$ and $\Omega^{\alpha,d}_{\beta,0}$ given by $$F^\alpha:=\left.{\mathcal{F}}^{\alpha,0}\right|_{t^*_{\ge 1}=0},\qquad \Omega^{\alpha,d}_{\beta,0}:=\left.\frac{\d{\mathcal{F}}^{\alpha,d}}{\d t^\beta_0}\right|_{t^*_{\ge 1}=0},$$ define a calibrated flat F-manifold such that the $N$-tuples $({\mathcal{F}}^{1,a},\ldots,{\mathcal{F}}^{N,a})$ are its descendant vector potentials.
Let ${\overline{\mathcal{F}}}^a(t^*_*)$, $a\ge 0$, be a collection of descendant vector potentials and consider a linear change of variables $$\begin{gathered}
t^\alpha_a\mapsto{\widetilde{t}}^\alpha_a(t^*_*)=M^\alpha_\mu t^\mu_a,\end{gathered}$$ where $M=(M^\alpha_\beta)\in{\mathrm{GL}}({\mathbb C}^N)$. Then it is easy to see that the collection of functions $(M.{\mathcal{F}})^{\alpha,a}({\widetilde{t}}^*_*)$ defined by $$\begin{gathered}
\label{eq:transformation of descendant potentials from linear changes}
(M.{\mathcal{F}})^{\alpha,a}:=\left.M^\alpha_\mu{\mathcal{F}}^{\mu,a}\right|_{t^\beta_b=(M^{-1})^\beta_\gamma{\widetilde{t}}^\gamma_b},\quad a\ge 0,\end{gathered}$$ satisfies all the equations – and, thus, gives a collection of descendant vector potentials. The unit vector field ${\widetilde{A}}^\alpha\frac{\d}{\d{\widetilde{t}}^\alpha}$ of the associated flat F-manifold is given by ${\widetilde{A}}^\alpha=M^\alpha_\mu A^\mu$. Clearly, this defines a ${\mathrm{GL}}({\mathbb C}^N)$-action on collections of descendant vector potentials. We will use the notation $$\overline{M.{\mathcal{F}}}^a:=((M.{\mathcal{F}})^{1,a},\ldots,(M.{\mathcal{F}})^{N,a}).$$
Ancestor vector potentials
--------------------------
Consider a collection of descendant vector potentials ${\overline{\mathcal{F}}}^a$, ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}}[[t^*_{\ge 1}]]$. They are called [*ancestor*]{} if ${\overline{t}}_{{\mathrm{orig}}}=0$ and the following property is satisfied: $$\label{eq:ancestor vector potential}
\left.\frac{\d{\mathcal{F}}^{\alpha,a}}{\d t^\beta_b}\right|_{t^*_*=0}=0,\quad a,b\ge 0.$$
Consider a flat F-manifold structure given by a vector potential ${\overline{F}}$ with $F^\alpha\in{\mathbb C}[[t^1,\ldots,t^N]]$. Then there exists a unique calibration giving a collection of ancestor potentials.
It is easy to check that such a calibration is uniquely determined by equation and the properties $\Omega^{\alpha,0}_{\beta,0}=\frac{\d F^\alpha}{\d t^\beta}-\left.\frac{\d F^\alpha}{\d t^\beta}\right|_{t^*=0}$ and $\left.\Omega^{\alpha,a}_{\beta,0}\right|_{t^*=0}=0$, $a\ge 0$.
A unique calibration described by this lemma will be called the [*ancestor calibration*]{}.\
Let us assign to the variable $t^\beta_b$ degree $b-1$, $\deg t^\beta_b:=b-1$.
\[lemma:technical property of ancestor potentials\] Consider a collection of ancestor vector potentials ${\overline{\mathcal{F}}}^a$, $a\ge 0$. Then $$\begin{gathered}
\label{eq:degree of Falphaa}
\deg{\mathcal{F}}^{\alpha,a}\le -a-2,\quad a\ge 0,\end{gathered}$$ which means that all the monomials that form the power series ${\mathcal{F}}^{\alpha,a}$ have degree less or equal to $-a-2$.
By equation , we have $\left.{\mathcal{F}}^{\alpha,a}\right|_{t^*_*=0}=0$. Therefore, property follows from the property $$\begin{gathered}
\deg\left(\frac{\d{\mathcal{F}}^{\alpha,a}}{\d t^\beta_b}\right)\le -a-b-1,\quad a,b\ge 0,\end{gathered}$$ which can be easily checked by induction using relations and . The lemma is proved.
\[remark:polynomiality of ancestor potentials\] Lemma \[lemma:technical property of ancestor potentials\] implies that for a collection of ancestor vector potentials ${\overline{\mathcal{F}}}^a$ we have $${\mathcal{F}}^{\alpha,a}\in{\mathbb C}[t^*_{\ge 2}][[t^*_0,t^*_1]].$$ Therefore, for any collection of constants $c^\beta_b\in{\mathbb C}$ such that $c^\beta_0=c^\beta_1=0$ the formal power series ${\mathcal{F}}^{\alpha,a}$ can be expressed as a formal power series in the shifted variables $(t^\beta_b+c^\beta_b)$, for which property still holds.
Constant flat F-manifolds
-------------------------
A flat F-manifold given by a vector potential ${\overline{F}}=(F^1,\ldots,F^N)$ is called [*constant*]{} if its structure constants $c^\alpha_{\beta\gamma}=\frac{\d^2 F^\alpha}{\d t^\beta\d t^\gamma}$ are constants.\
In any rank $N\ge 1$ the simplest constant flat F-manifold has the vector potential ${\overline{F}}=\left(\frac{(t^1)^2}{2},\ldots,\frac{(t^N)^2}{2}\right)$ and the unit $\sum_{\alpha=1}^N\frac{\d}{\d t^\alpha}$. Its ancestor vector potentials are given by $${\mathcal{F}}^{\alpha,a}=\sum_{n\ge a+2}\sum_{\substack{d_1,\ldots,d_n\ge 0\\\sum d_i=n-2-a}}\frac{1}{n(n-1)}\frac{\prod t^\alpha_{d_i}}{a!\prod d_i!},\quad a\ge 0.$$ This flat F-manifold is called the [*trivial flat F-manifold of dimension $N$*]{}. Applying to it the ${\mathrm{GL}}({\mathbb C}^N)$-action we can get the ancestor vector potentials of any semisimple constant flat F-manifold. Clearly, in canonical coordinates the Christoffel symbols of the connection $\nabla$ and the rotation coefficients of any semisimple constant flat F-manifold vanish.
Flat F-manifolds and the geometry of ${\mathbb C}^N[[z,z^{-1}]]$
----------------------------------------------------------------
Let ${\mathcal{H}}:={\mathbb C}^N[[z,z^{-1}]]$. Denote by $\phi_1,\ldots,\phi_N$ the standard basis in ${\mathbb C}^N$. Any element $f(z)\in{\mathcal{H}}$ can be expressed as $$f(z)=\sum_{d\ge 0}\frac{(-1)^{d+1}p^\alpha_d\phi_\alpha}{z^{d+1}}+\sum_{d\ge 0}q^\alpha_d\phi_\alpha z^d,\quad p^\alpha_d,q^\alpha_d\in{\mathbb C},$$ in a unique way. We view the coefficients $p^\alpha_d$ and $q^\alpha_d$ as coordinates on the space ${\mathcal{H}}$. We have the decomposition ${\mathcal{H}}={\mathcal{H}}_+\oplus{\mathcal{H}}_-$, where ${\mathcal{H}}_+={\mathbb C}^N[[z]]$ and ${\mathcal{H}}_-=z^{-1}{\mathbb C}^N[[z^{-1}]]$.\
Consider a collection of functions ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}}[[t^*_{\ge 1}]]$, $1\le\alpha\le N$, $a\ge 0$. Recall that we relate the variables $t^\beta_b$ and $q^\beta_b$ by $q^\beta_b=t^\beta_b-A^\beta\delta_{b,1}$. We view ${\mathcal{F}}^{\alpha,a}$ as formal functions on the space ${\mathcal{H}}_+$ near the point ${\overline{t}}_{\mathrm{orig}}-\phi_\un z$, where $\phi_\un:=A^\alpha\phi_\alpha$. Consider the graph of the collection of functions ${\mathcal{F}}^{\alpha,a}$ in the space ${\mathcal{H}}$: $${\mathcal{C}}:=\left\{(p^*_*,q^*_*)\in{\mathcal{H}}|p^\alpha_a={\mathcal{F}}^{\alpha,a}\right\}\subset{\mathcal{H}}.$$
\[theorem:flat F-manifold as a cone\] The functions ${\mathcal{F}}^{\alpha,a}$ satisfy equations – if and only if the following three conditions are satisfied:
- ${\mathcal{C}}$ is a cone with the vertex at the origin;
- for any point $f\in{\mathcal{C}}$ we have $f\in zT_f{\mathcal{C}}$, where $T_f{\mathcal{C}}\subset{\mathcal{H}}$ denotes the tangent space to ${\mathcal{C}}$ at the point $f$;
- for any point $f\in{\mathcal{C}}$ the tangent space $T_f{\mathcal{C}}$ is tangent to ${\mathcal{C}}$ along $zT_f{\mathcal{C}}$.
Note that, comparing to a similar result about Frobenius manifolds [@Giv04 Theorem 1], we just drop the condition that the cone is Lagrangian and the condition that any tangent space $L$ is tangent to ${\mathcal{C}}$ [*exactly*]{} along $zL$. Actually, the second of these two conditions follows automatically from the conditions of Theorem \[theorem:flat F-manifold as a cone\]. We will show it in Lemma \[lemma:last condition\] after the proof of the theorem.
Clearly, ${\mathcal{C}}$ is a cone with the vertex at the origin if and only if the vector field $\sum_{a\ge 0}q^\alpha_a\frac{\d}{\d q^\alpha_a}+\sum_{a\ge 0}p^\alpha_a\frac{\d}{\d p^\alpha_a}$ is tangent to ${\mathcal{C}}$, which is equivalent to equation .\
Assume now that equation is satisfied. Therefore, for any $f\in{\mathcal{C}}$ the tangent space $T_f{\mathcal{C}}$ passes through the origin. Since ${\mathcal{C}}$ is the graph of the collection of functions ${\mathcal{F}}^{\alpha,a}$, we can lift the coordinate vector fields $\frac{\d}{\d q^\alpha_a}$ on ${\mathcal{H}}_+$ to vector fields $e_{\alpha,a}$ on ${\mathcal{C}}$, which give a natural basis in the tangent spaces to ${\mathcal{C}}$. We have $$e_{\alpha,a}=\frac{\d}{\d q^\alpha_a}+\sum_{b\ge 0}\frac{\d{\mathcal{F}}^{\beta,b}}{\d q^\alpha_a}\frac{\d}{\d p^\beta_b},\quad1\le\alpha\le N,\quad a\ge 0.$$ Consider now a point $f=(p^\beta_b,q^\alpha_a)\in{\mathcal{C}}$. We have $z^{-1}f=({\widetilde{p}}^\beta_b,{\widetilde{q}}^\alpha_a)$, where $${\widetilde{q}}^\alpha_a=q^\alpha_{a+1},\qquad
{\widetilde{p}}^\beta_b=
\begin{cases}
-q^\beta_0,&\text{if $b=0$},\\
-{\mathcal{F}}^{\beta,b-1},&\text{if $b\ge 1$}.
\end{cases}$$ The condition $f\in zT_f{\mathcal{C}}\Leftrightarrow z^{-1}f\in T_f{\mathcal{C}}$ is satisfied if and only if the vector $$\sum_{a\ge 0}q^\alpha_{a+1}\frac{\d}{\d q^\alpha_a}-q^\alpha_0\frac{\d}{\d p^\alpha_0}-\sum_{a\ge 0}{\mathcal{F}}^{\alpha,a}\frac{\d}{\d p^\alpha_{a+1}}\in T_f{\mathcal{H}}$$ belongs to $T_f{\mathcal{C}}$. The last property is equivalent to the equation $$\sum_{a\ge 0}q^\alpha_{a+1}\frac{\d{\mathcal{F}}^{\beta,b}}{\d q^\alpha_a}=
\begin{cases}
-{\mathcal{F}}^{\beta,b-1},&\text{if $b\ge 1$},\\
-q^\beta_0,&\text{if $b=0$},
\end{cases}$$ which is exactly equation .\
Let us assume that equations and are satisfied. For a tangent space $T_f{\mathcal{C}}$ a basis in $z T_f{\mathcal{C}}$ is given by the vectors $$z e_{\alpha,a}=\frac{\d}{\d q^\alpha_{a+1}}-\frac{\d{\mathcal{F}}^{\beta,0}}{\d q^\alpha_a}\frac{\d}{\d q^\beta_0}-\sum_{b\ge 1}\frac{\d{\mathcal{F}}^{\beta,b}}{\d q^\alpha_a}\frac{\d}{\d p^\beta_{b-1}},\quad 1\le\alpha\le N,\quad a\ge 0.$$ Thus, $$\begin{gathered}
zT_f{\mathcal{C}}\subset T_f{\mathcal{C}}\Leftrightarrow ze_{\alpha,a}=e_{\alpha,a+1}-\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\alpha_a}e_{\mu,0}\Leftrightarrow \frac{\d{\mathcal{F}}^{\beta,b+1}}{\d q^\alpha_a}+\frac{\d{\mathcal{F}}^{\beta,b}}{\d q^\alpha_{a+1}}=\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\alpha_a}\frac{\d{\mathcal{F}}^{\beta,b}}{\d q^\mu_0},\end{gathered}$$ where the last equation coincides with .\
Let us assume now that equations , and are satisfied. We already know that for any point $f\in{\mathcal{C}}$ we have $f\in zT_f{\mathcal{C}}\subset T_f{\mathcal{C}}$. Therefore, the space $T_f{\mathcal{C}}$ is tangent to ${\mathcal{C}}$ along $zT_f{\mathcal{C}}$ if and only if the Lie derivative of the coefficient of $\frac{\d}{\d p^\alpha_a}$ in $e_{\gamma,c}$ along the vector field $z e_{\beta,b}$ is zero. The last property is equivalent to the equation $$\left(\frac{\d}{\d q^\beta_{b+1}}-\frac{\d{\mathcal{F}}^{\mu,0}}{\d q^\beta_b}\frac{\d}{\d q^\mu_0}\right)\frac{\d{\mathcal{F}}^{\alpha,a}}{\d q^\gamma_c}=0.$$ Since we have assumed that equation is satisfied, the last equation is equivalent to equation . The theorem is proved.
The cone ${\mathcal{C}}\subset{\mathcal{H}}$ obtained from a calibrated flat F-manifold by the construction described above will be called a [*generalized Givental cone*]{}. The generalized Givental cone corresponding to an ancesor calibration will be called an [*ancestor cone*]{}. The ancestor cone corresponding to the trivial flat F-manifold of dimension $N$ will be denoted by ${\mathcal{C}}^{\mathrm{triv}}_N$.
\[lemma:last condition\] Consider functions ${\mathcal{F}}^{\alpha,a}$ satisfying equations – and the associated cone ${\mathcal{C}}\subset{\mathcal{H}}$. Then each tangent space $L$ to the cone ${\mathcal{C}}$ is tangent to it exactly along $zL$.
The subspace $zL\subset L$ has codimension $N$ and a basis in $L/zL$ is given by the vectors $e_{\alpha,0}$. The coefficients of $\frac{\d}{\d p^\beta_b}$ in the vector fields $e_{\alpha,a}$ give natural functions on the space of tangent spaces to ${\mathcal{C}}$. Consider the coefficient ${\mathrm{Coef}}_{\d/\d p^\beta_0}e_{\un,0}=\frac{\d{\mathcal{F}}^{\beta,0}}{\d q^\un_0}$ and the Lie derivative $L_{e_{\alpha,0}}$ of it along the vector field $e_{\alpha,0}$. Since $$\left.\left(L_{e_{\alpha,0}}\frac{\d{\mathcal{F}}^{\beta,0}}{\d q^\un_0}\right)\right|_{q^\gamma_c=\delta_{c,0}t^\gamma_{\mathrm{orig}}-\delta_{c,1}A^\gamma}=\delta^\beta_\alpha,$$ we see that any tangent space $L$ to ${\mathcal{C}}$ is tangent to it exactly along $zL$.
Consider a map $M\in{\mathrm{GL}}({\mathbb C}^N)$, $M\phi_{\alpha}=M^\beta_\alpha\phi_\beta$. It induces a linear map ${\mathcal{H}}\mapsto{\mathcal{H}}$ and, obviously, the conditions for a cone ${\mathcal{C}}$ formulated in Theorem \[theorem:flat F-manifold as a cone\] are preserved by this map. It is easy to see that the transformation of the corresponding descendant vector potentials is described by formula .
$J$-function
------------
Consider a generalized Givental cone ${\mathcal{C}}\subset{\mathcal{H}}$. In the same way as for the case of Frobenius manifolds [@Giv04] one can introduce the $J$-function associated to ${\mathcal{C}}$. For this we consider the intersection of the cone ${\mathcal{C}}$ with the affine space $-z\phi_\un+z{\mathcal{H}}_-$. Via the projection to $-z\phi_\un+{\mathbb C}^N$ along ${\mathcal{H}}_-$ the intersection becomes the graph of a function from ${\mathbb C}^N$ to ${\mathcal{H}}$ called the [*$J$-function*]{}: $${\mathbb C}^N\ni t^\alpha\phi_\alpha\mapsto J(-z,t^*)=-z\phi_\un+t^\alpha\phi_\alpha+\sum_{j\ge 1}J_k(t^*)(-z)^{-k}.$$ If ${\overline{\mathcal{F}}}^a$ are the descendant vector potentials, corresponding to the cone ${\mathcal{C}}$, then, clearly, $$J_k(t^*)=\left.{\mathcal{F}}^{\alpha,k-1}\phi_\alpha\right|_{t^*_{\ge 1}=0},\quad k\ge 1.$$ The cone ${\mathcal{C}}$ is uniquely determined by its $J$-function. This fact can be derived, for example, from equation .\
The $J$-function of the cone ${\mathcal{C}}^{\mathrm{triv}}_N$ is $$J(z,t^*)=\sum_{\alpha=1}^N ze^{t^\alpha/z}\phi_\alpha.$$
Generalized Givental group action
---------------------------------
Consider two groups $$\begin{aligned}
G_+:=&\left\{R(z)={\mathrm{Id}}+\sum_{i\ge 1}R_i z^i\in{\mathrm{End}}({\mathbb C}^N)[[z]]\right\},\\
G_-:=&\left\{S(z)={\mathrm{Id}}+\sum_{i\ge 1}S_i z^{-i}\in{\mathrm{End}}({\mathbb C}^N)[[z^{-1}]]\right\}.\end{aligned}$$ Let us call the groups $G_+$ and $G_-$ the [*upper triangular*]{} and the [*lower triangular group*]{}, respectively. If we ignore for a moment potential problems caused by infinite summation, we can say that the groups $G_+$ and $G_-$ act on the space ${\mathcal{H}}$ by the left multiplication. Moreover, the conditions for a cone ${\mathcal{C}}$ formulated in Theorem \[theorem:flat F-manifold as a cone\] are preserved by these actions. Thus, we get $G_{\pm}$-actions on the space of calibrated flat F-manifolds. Let us analyze these actions more carefully.
### Upper triangular group {#subsubsection:upper triangular group}
\[proposition:R-action is well-defined\] The $G_+$-action on the space of ancestor cones is well defined.
Let ${\mathcal{C}}\subset{\mathcal{H}}$ be the cone corresponding to a collection ${\overline{\mathcal{F}}}^a$ of ancestor vector potentials. Consider a point $$f(z)=\sum_{a\ge 0}\frac{(-1)^{a+1}{\mathcal{F}}^{\alpha,a}\phi_\alpha}{z^{a+1}}+\sum_{a\ge 0}q^\alpha_a\phi_\alpha z^a\in{\mathcal{C}}.$$ For $R(z)\in G_+$ we have $$R(z)f(z)=\sum_{a\ge 0}\frac{(-1)^{a+1}{\widetilde{\mathcal{F}}}^{\alpha,a}\phi_\alpha}{z^{a+1}}+\sum_{a\ge 0}{\widetilde{q}}^\alpha_a\phi_\alpha z^a,$$ where $$\begin{aligned}
{\widetilde{q}}^\alpha_a=&q^\alpha_a+\sum_{i=1}^a(R_i)^\alpha_\mu q^\mu_{a-i}+\sum_{i\ge a+1}(R_i)^\alpha_\mu(-1)^{i-a}{\mathcal{F}}^{\mu,i-a-1},&& a\ge 0,\label{eq:R-transformation of q}\\
{\widetilde{\mathcal{F}}}^{\alpha,a}=&{\mathcal{F}}^{\alpha,a}+\sum_{i\ge 1}(-1)^i(R_i)^\alpha_\mu{\mathcal{F}}^{\mu,a+i},&& a\ge 0.\label{eq:R-transformation of mcF}\end{aligned}$$ Lemma \[lemma:technical property of ancestor potentials\] implies that the infinite sums on the right-hand sides of these two equations are well defined. Lemma \[lemma:technical property of ancestor potentials\] also implies that the transformation $q^\alpha_a\mapsto{\widetilde{q}}^\alpha_a$ can be considered as a change of variables. Therefore, we can express the functions ${\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)$ as functions of the variables ${\widetilde{q}}^\beta_b$, ${\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)={\widehat{\mathcal{F}}}^{\alpha,a}({\widetilde{q}}^*_*)$. Note that the function ${\widehat{\mathcal{F}}}^{\alpha,a}$ is a formal function around the point $$R(z)(-\phi_\un z)=-\phi_\un z-\sum_{i\ge 1}(R_i)^\mu_\un\phi_\mu z^{i+1}\in{\mathcal{H}}_+,$$ but, by Remark \[remark:polynomiality of ancestor potentials\], it can be expressed as a formal function around the point $-\phi_\un z$.\
It remains to check that the descendant vector potentials $({\widehat{\mathcal{F}}}^{1,a},\ldots,{\widehat{\mathcal{F}}}^{N,a})$ are ancestor. Let us express the functions ${\widehat{\mathcal{F}}}^{\alpha,a}$ as formal power series in the variables ${\widetilde{t}}^\beta_b={\widetilde{q}}^\beta_b+A^\beta\delta_{b,1}$. Formulas and imply that property holds for the functions ${\widehat{\mathcal{F}}}^{\alpha,a}$. Therefore, $\left.\frac{\d{\widehat{\mathcal{F}}}^{\alpha,a}}{\d{\widetilde{t}}^\beta_b}\right|_{{\widetilde{t}}^*_*=0}=0$ and the descendant vector potentials $({\widehat{\mathcal{F}}}^{1,a},\ldots,{\widehat{\mathcal{F}}}^{N,a})$ are ancestor.
Consider a collection of ancestor vector potentials ${\overline{\mathcal{F}}}^a$ and the corresponding cone ${\mathcal{C}}\subset{\mathcal{H}}$. For $R(z)\in G_+$ denote by $\overline{R(z).{\mathcal{F}}}^a=\left((R(z).{\mathcal{F}})^{1,a},\ldots,(R(z).{\mathcal{F}})^{N,a}\right)$ the ancestor vector potentials, corresponding to the cone $R(z){\mathcal{C}}\subset{\mathcal{H}}$. Recall that we use the convention $${\mathcal{F}}^{\alpha,a}=(-1)^{a+1}q^\alpha_{-a-1},\quad\text{if $a<0$}.$$
\[proposition:infinitesimal R-action\] The infinitesimal action of the group $G_+$ on the space of ancestor vector potentials is given by $$\begin{gathered}
\label{eq:infinitesimal R-action}
\left.\frac{d}{d{\varepsilon}}(e^{{\varepsilon}r(z)}.{\mathcal{F}})^{\alpha,a}\right|_{{\varepsilon}=0}=\sum_{i\ge 1}(-1)^i(r_i)^\alpha_\mu{\mathcal{F}}^{\mu,a+i}+\sum_{i\ge 1,\,j\ge 0}(-1)^{i-j-1}(r_i)^\mu_\nu\frac{\d{\mathcal{F}}^{\alpha,a}}{\d t^\mu_j}{\mathcal{F}}^{\nu,i-j-1},\quad a\ge 0,\end{gathered}$$ where $r(z)=\sum_{i\ge 1}r_i z^i$, $r_i\in{\mathrm{End}}({\mathbb C}^N)$.
We have $$\left.e^{{\varepsilon}r(z)}[{\mathcal{F}}]^{\alpha,a}\right|_{q^\beta_b\mapsto{\widetilde{q}}^\beta_b}={\widetilde{\mathcal{F}}}^{\alpha,a},$$ where ${\widetilde{q}}^\beta_b$ and ${\widetilde{\mathcal{F}}}^{\alpha,a}$ are given by formulas and with $R(z)=e^{{\varepsilon}r(z)}$. Differentiating both sides of this equation with respect to ${\varepsilon}$ and setting ${\varepsilon}=0$, we get formula .
### Lower triangular group {#subsubsection:lower triangular group}
Consider a flat F-manifold with a vector potential ${\overline{F}}$, $F^\alpha\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}}$, its calibration given by matrices $\Omega^p_0$, and the associated descendant vector potentials ${\overline{\mathcal{F}}}^a$, ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}}[[t^*_{\ge 1}]]$, where $c^\alpha=\Omega^{\alpha,0}_{\un,0}-t^\alpha$. Let ${\mathcal{C}}\subset{\mathcal{H}}$ be the associated cone.
\[proposition:lower triangular group action\] Consider an arbitrary element $S(z)={\mathrm{Id}}+\sum_{i\ge 1}S_i z^{-i}\in G_-$ and let $\log S(z)=s(z)=\sum_{i\ge 1}s_i z^{-i}$.\
1. The cone $S(z){\mathcal{C}}$ is well defined. It corresponds to a collection of descendant vector potentials, which we denote by $\overline{S(z).{\mathcal{F}}}^a=\left((S(z).{\mathcal{F}})^{1,a},\ldots,(S(z).{\mathcal{F}})^{N,a}\right)$, with $(S(z).{\mathcal{F}})^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}-S_1\phi_\un}[[t^*_{\ge 1}]]$ given by $$\begin{gathered}
\label{eq:S-action on mcF}
(S(z).{\mathcal{F}})^{\alpha,a}=e^{-\widehat{s(z)}}\left({\mathcal{F}}^{\alpha,a}+\sum_{i=1}^a(-1)^i(S_i)^\alpha_\mu{\mathcal{F}}^{\mu,a-i}+(-1)^{a+1}\sum_{i\ge a+1}(S_i)^\alpha_\mu q^\mu_{i-a-1}\right),\end{gathered}$$ where $\widehat{s(z)}=\sum_{\substack{i\ge 1\\j\ge 0}}(s_i)^\alpha_\beta q^\beta_{i+j}\frac{\d}{\d q^\alpha_j}$.\
2. The descendant vector potentials $\overline{S(z).{\mathcal{F}}}^a$ correspond to the same flat F-manifold, but with the different calibration given by the matrices ${\widetilde\Omega}^d_0$ defined by $${\mathrm{Id}}+\sum_{d\ge 1}{\widetilde\Omega}^{d-1}_0 z^d=S(-z^{-1})\left({\mathrm{Id}}+\sum_{d\ge 1}\Omega^{d-1}_0 z^d\right).$$
1\. For a point $$f(z)=\sum_{a\ge 0}\frac{(-1)^{a+1}{\mathcal{F}}^{\alpha,a}\phi_\alpha}{z^{a+1}}+\sum_{a\ge 0}q^\alpha_a\phi_\alpha z^a\in{\mathcal{C}}$$ we have $$S(z)f(z)=\sum_{a\ge 0}\frac{(-1)^{a+1}({\mathcal{F}}')^{\alpha,a}\phi_\alpha}{z^{a+1}}+\sum_{a\ge 0}(q')^\alpha_a\phi_\alpha z^a,$$ where $$\begin{aligned}
(q')^\alpha_a=&q^\alpha_a+\sum_{i\ge 1}(S_i)^\alpha_\mu q^\mu_{a+i}=e^{\widehat{s(z)}} q^\alpha_a,&& a\ge 0,\\
({\mathcal{F}}')^{\alpha,a}=&{\mathcal{F}}^{\alpha,a}+\sum_{i=1}^a(-1)^i(S_i)^\alpha_\mu{\mathcal{F}}^{\mu,a-i}+(-1)^{a+1}\sum_{i\ge a+1}(S_i)^\alpha_\mu q^\mu_{i-a-1},&& a\ge 0.\end{aligned}$$ Expressing the function $({\mathcal{F}}')^{\alpha,a}$ as a function of the variables $(q')^{\beta}_b$, we get formula . This proves Part 1 of the proposition.\
2. Denote $S(z)^{-1}={\mathrm{Id}}+\sum_{i\ge 1}{\widetilde{S}}_i z^{-i}$. Then we have ${\widetilde\Omega}^0_p=\sum_{i=0}^{p+1}\Omega^0_{p-i}{\widetilde{S}}_i$, $p\ge -1$. From this it is easy to see that the topological solution $({\widetilde v}^{\mathrm{top}})^\alpha$ and the matrices $({\widetilde\Omega}^{\mathrm{top}})^0_b$ corresponding to the new calibration are given by $$({\widetilde v}^{\mathrm{top}})^\alpha=e^{-\widehat{s(z)}}(v^{\mathrm{top}})^\alpha,\qquad ({\widetilde\Omega}^{\mathrm{top}})^0_b=e^{-\widehat{s(z)}}\left(\sum_{i=0}^{b+1}(\Omega^{\mathrm{top}})^0_{b-i}{\widetilde{S}}_i\right).$$ Then for the descendant vector potentials $({\widetilde{\mathcal{F}}}^{1,a},\ldots,{\widetilde{\mathcal{F}}}^{N,a})$ corresponding to the new calibration we get $$\begin{aligned}
{\widetilde{\mathcal{F}}}^{\alpha,0}=&\sum_{b\ge 0}q^\beta_b({\widetilde\Omega}^{\mathrm{top}})^{\alpha,0}_{\beta,b}=\sum_{b\ge 0}q^\beta_b e^{-\widehat{s(z)}}\left(\sum_{i=0}^{b+1}(\Omega^{\mathrm{top}})^{\alpha,0}_{\mu,b-i}({\widetilde{S}}_i)^\mu_\beta\right)=\\
=&e^{-\widehat{s(z)}}\left(\sum_{b,j\ge 0}(S_j)^\beta_\nu q^\nu_{b+j}\sum_{i=0}^{b+1}(\Omega^{\mathrm{top}})^{\alpha,0}_{\mu,b-i}({\widetilde{S}}_i)^\mu_\beta\right)=\end{aligned}$$ $$\begin{aligned}
=&e^{-\widehat{s(z)}}\left(\sum_{i,j,l\ge 0}(\Omega^{\mathrm{top}})^{\alpha,0}_{\mu,l}({\widetilde{S}}_i)^\mu_\beta(S_j)^\beta_\nu q^\nu_{i+j+l}+\sum_{b,j\ge 0}({\widetilde{S}}_{b+1})^\alpha_\beta(S_j)^\beta_\nu q^\nu_{b+j}\right)=\\
=&e^{-\widehat{s(z)}}\left({\mathcal{F}}^{\alpha,0}-\sum_{l\ge 0}(S_{l+1})^\alpha_\nu q^\nu_l\right)=(S(z).{\mathcal{F}})^{\alpha,0}.\end{aligned}$$ Knowing that ${\widetilde{\mathcal{F}}}^{\alpha,0}=(S(z).{\mathcal{F}})^{\alpha,0}$ is enough to conclude that ${\widetilde{\mathcal{F}}}^{\alpha,a}=(S(z).{\mathcal{F}})^{\alpha,a}$ for all $a\ge 0$. This completes the proof of part 2 of the proposition.
\[proposition:all from ancestor\] Any generalized Givental cone can be obtained from some ancestor cone by the action of an element from the group $G_-$.
Choosing a matrix $S_1$ such that $S_1\phi_\un={\overline{t}}_{\mathrm{orig}}+{{\overline{c}}}$ and using Proposition \[proposition:lower triangular group action\], we get $$(({\mathrm{Id}}+S_1z).{\mathcal{F}})^{\alpha,a}\in{\mathcal{R}}^0[[t^*_{\ge 1}]].$$ So without loss of generality we can assume that ${\overline{t}}_{\mathrm{orig}}={{\overline{c}}}=0$. Using again Proposition \[proposition:lower triangular group action\] we see that the cone $$\left({\mathrm{Id}}+\sum_{j\ge 1}(-1)^jz^{-j}\Omega^{j-1}_0(0)\right)^{-1}{\mathcal{C}}$$ is ancestor. The proposition is proved.
Let ${\overline{F}}=(F^1,\ldots,F^N)$, $F^\alpha\in{\mathbb C}[[t^1,\ldots,t^N]]$, be a vector potential of a flat F-manifold. Denote by ${\mathcal{C}}$ the associated ancestor cone and by $\Omega^j_{0}$ the matrices defining the ancestor calibration. Consider a family of vector potentials ${\overline{F}}_{{{\overline{\tau}}}}$, depending on formal parameters $\tau^1,\ldots,\tau^N$, ${{\overline{\tau}}}=(\tau^1,\ldots,\tau^N)$, defined by $$F^\alpha_{{{\overline{\tau}}}}:=F^\alpha|_{t^\beta\mapsto t^\beta+\tau^\beta}\in{\mathbb C}[[\tau^*]][[t^*]].$$ Denote by ${\mathcal{C}}_{{{\overline{\tau}}}}$ the associated family of ancestor cones.
\[lemma:ancestor cones at shifted points\] We have $$\begin{gathered}
\label{eq:ancestor cones at shifted points}
{\mathcal{C}}_{{{\overline{\tau}}}}=\left({\mathrm{Id}}+\sum_{j\ge 1}(-1)^j\Omega^{j-1}_0(\tau^*)z^{-j}\right)^{-1}{\mathcal{C}}.\end{gathered}$$
According to Proposition \[proposition:lower triangular group action\], the object on the right-hand side of equation is a germ of a graph over the point $-\tau^\alpha\phi_\alpha-\phi_\un z\in{\mathcal{H}}_+$, while ${\mathcal{C}}_{{{\overline{\tau}}}}$ is a germ of a graph over the point $-\phi_\un z\in{\mathcal{H}}_+$. However, since the parameters $\tau^\alpha$ are formal, the cone on the right-hand side can be considered as a germ of a graph over the point $-\phi_\un z\in{\mathcal{H}}_+$.
By Proposition \[proposition:lower triangular group action\], the cone on the right-hand side of equation corresponds to the same flat F-manifold given by the vector potential ${\overline{F}}$ together with the calibration $${\mathrm{Id}}+\sum_{j\ge 1}{\widetilde\Omega}^{j-1}_0(t^*)z^j=\left({\mathrm{Id}}+\sum_{j\ge 1}\Omega^{j-1}_0(\tau^*)z^j\right)^{-1}\left({\mathrm{Id}}+\sum_{j\ge 1}\Omega^{j-1}_0(t^*)z^j\right).$$ By Remark \[remark:about shifts\], this cone also corresponds to the flat F-manifold given by the vector potential ${\overline{F}}_{{{\overline{\tau}}}}$ together with the calibration ${\mathrm{Id}}+\sum_{j\ge 1}({\widetilde\Omega}^{-{{\overline{\tau}}}})^{j-1}_0z^j$. Since $$\left.\left({\mathrm{Id}}+\sum_{j\ge 1}({\widetilde\Omega}^{-{{\overline{\tau}}}})^{j-1}_0z^j\right)\right|_{t^\alpha=0}=\left.\left({\mathrm{Id}}+\sum_{j\ge 1}{\widetilde\Omega}^{j-1}_0z^j\right)\right|_{t^\alpha=\tau^\alpha}={\mathrm{Id}},$$ the cone on the right-hand side of is indeed the ancestor cone corresponding to the flat F-manifold with the vector potential ${\overline{F}}_{{{\overline{\tau}}}}$.
Reconstruction of semisimple flat F-manifolds {#section:reconstruction}
---------------------------------------------
Consider a flat F-manifold given by a vector potential ${\overline{F}}=(F^1,\ldots,F^N)$, $F^\alpha\in{\mathbb C}[[t^1,\ldots,t^N]]$. Consider the constant flat F-manifold given by the structure constants of our flat F-manifold at the origin. We call this constant flat F-manifold the [*constant part of our flat F-manifold*]{}.
\[theorem:reconstruction\] Suppose that our flat F-manifold is semisimple at the origin. Then the corresponding ancestor cone ${\mathcal{C}}$ can be obtained from the ancestor cone ${\mathcal{C}}^{\mathrm{const}}$ of the constant part by some element $R(z)$ of the group $G_+$.
Before proving the theorem, let us present an important technical result. Consider a calibrated flat F-manifold given by a vector potential ${\overline{F}}$ and matrices $\Omega^p_0$, $F^\alpha,\Omega^{\alpha,p}_{\beta,0}\in{\mathcal{R}}^{{\overline{t}}_{\mathrm{orig}}}$, and the associated descendant vector potentials ${\overline{\mathcal{F}}}^a$, ${\mathcal{F}}^{\alpha,a}\in{\mathcal{R}}^{{\overline{t}}_{{\mathrm{orig}}}+{{\overline{c}}}}[[t^*_{\ge 1}]]$. Recall that the constants $c^\alpha$ are given by $\Omega^{\alpha,0}_{\un,0}=t^\alpha+c^\alpha$. Consider the associated cone ${\mathcal{C}}\subset{\mathcal{H}}$.
\[proposition:another parameterization\] The cone ${\mathcal{C}}$ has the following parameterization: $${\mathcal{C}}=\left\{\left(\sum_{j\ge 0}(-1)^j\Omega^{j-1}_0 z^{-j}\right)\sum_{i\ge 1}q^\alpha_i\phi_\alpha z^i\right\}.$$
During the proof of the proposition we will denote the coordinates on ${\mathcal{H}}$ by ${\widetilde{p}}^\beta_b$ and ${\widetilde{q}}^\alpha_a$. We have $$\left(\sum_{j\ge 0}(-1)^j\Omega^{j-1}_0 z^{-j}\right)\sum_{i\ge 1}q^\alpha_i\phi_\alpha z^i=\sum_{a\ge 0}\frac{(-1)^{a+1}{\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)\phi_\alpha}{z^{a+1}}+\sum_{a\ge 0}{\widetilde{q}}^\alpha_a(q^*_*)\phi_\alpha z^a,$$ where $$\begin{aligned}
{\widetilde{q}}^\alpha_a(q^*_*)=&
\begin{cases}
t^\alpha_0+c^\alpha+\sum_{i\ge 1}(-1)^i\Omega^{\alpha,i-1}_{\mu,0}t^\mu_i,&\text{if $a=0$},\\
q^\alpha_a+\sum_{i\ge 1}(-1)^i\Omega^{\alpha,i-1}_{\mu,0} q^\mu_{i+a},&\text{if $a\ge 1$},
\end{cases}\\
{\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)=&\sum_{i\ge 1}(-1)^i\Omega^{\alpha,i+a}_{\mu,0}q^\mu_i,\hspace{3.1cm} a\ge 0.\end{aligned}$$ We see that the transformation $q^\alpha_a\mapsto{\widetilde{q}}^\alpha_a(q^*_*)$ can be considered as a change of variables and, therefore, we can express the functions ${\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)$ as functions of the variables ${\widetilde{q}}^\beta_b$, ${\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)=\left.{\widehat{\mathcal{F}}}^{\alpha,a}({\widetilde{q}}^*_*)\right|_{{\widetilde{q}}^\beta_b={\widetilde{q}}^\beta_b(q^*_*)}$, where ${\widehat{\mathcal{F}}}^{\alpha,a}\in{\mathbb C}[[{\widetilde{q}}^\beta_0-(t^\beta_{\mathrm{orig}}+c^\beta),{\widetilde{q}}^\beta_1+A^\beta,{\widetilde{q}}^\beta_{\ge 2}]]$.\
Let us first check that the submanifold ${\mathcal{C}}'\subset{\mathcal{H}}$ given by $${\mathcal{C}}':=\left\{\left(\sum_{j\ge 0}(-1)^j\Omega^{j-1}_0 z^{-j}\right)\sum_{i\ge 1}q^\alpha_i\phi_\alpha z^i\right\}=\left\{\sum_{i\ge 1,\,j\ge 0}(-1)^j\Omega^{\mu,j-1}_{\nu,0}q^\nu_i\phi_\mu z^{i-j}\right\}$$ satisfies the conditions from Theorem \[theorem:flat F-manifold as a cone\]. Regarding the first condition, we see that a vector $$f=\sum_{i\ge 1,\,j\ge 0}(-1)^j\Omega^{\mu,j-1}_{\nu,0}q^\nu_i\phi_\mu z^{i-j}\in{\mathcal{C}}'\subset{\mathcal{H}}$$ depends linearly on the variables $q^\nu_i$, with $i\ge 1$. Therefore, ${\mathcal{C}}'$ is a cone with the vertex at the origin.\
For $f\in{\mathcal{H}}$ denote by $f^\alpha_i$ the coefficient of $\phi_\alpha z^i$ in $f$ and let us use these coefficients as coordinates on ${\mathcal{H}}$. For $f\in{\mathcal{C}}'$ a basis in the tangent space $T_f{\mathcal{C}}'$ is given by the vectors $$e_{\alpha,a}:=
\begin{cases}
\sum_{i\ge 1,\,j\ge 0}(-1)^j\frac{\d\Omega^{\mu,j-1}_{\nu,0}}{\d q^\alpha_0}q^\nu_i\frac{\d}{\d f^\mu_{i-j}}=-\sum_{i\ge 1,\,j\ge 0}(-1)^j q^\nu_i c^\theta_{\nu\alpha}\Omega^{\mu,j-1}_{\theta,0}\frac{\d}{\d f^\mu_{i-j-1}},&\text{if $a=0$},\\
\sum_{j\ge 0}(-1)^j\Omega^{\mu,j-1}_{\alpha,0}\frac{\d}{\d f^\mu_{a-j}},&\text{if $a\ge 1$}.
\end{cases}$$ One can immediately see that after the obvious identification of the vector spaces ${\mathcal{H}}$ and $T_f{\mathcal{H}}$ the vectors $z^{-1}f$ and $-e_{\un,0}$ become equal. Thus, the second condition from Theorem \[theorem:flat F-manifold as a cone\] is satisfied.\
Let us check the third condition. We compute $$z e_{\alpha,a}=
\begin{cases}
e_{\alpha,a+1},&\text{if $a\ge 1$},\\
-\sum_{i\ge 1}q^\nu_i c^\theta_{\nu\alpha}e_{\theta,i},&\text{if $a=0$},
\end{cases}$$ which implies that $z T_f{\mathcal{C}}'\subset T_f{\mathcal{C}}'$. Define a matrix $M=(M^\alpha_\beta)$ by $M^\alpha_\beta:=-q^\theta_1 c^\alpha_{\theta\beta}$. Note that, since $M^\alpha_\beta=\delta^\alpha_\beta-t^\theta_1 c^\alpha_{\theta\beta}$, the matrix $M$ is invertible. Therefore, the vectors $e_{\alpha,a}$ with $a\ge 1$ give a basis in $z T_f{\mathcal{C}}'$. We see that in order to check the third condition from Theorem \[theorem:flat F-manifold as a cone\] it is sufficient to check that $\frac{\d e_{\alpha,a}}{\d q^\beta_b}\in T_f{\mathcal{C}}'$ for $a\ge 0$ and $b\ge 1$. Clearly, $\frac{\d e_{\alpha,a}}{\d q^\beta_b}=0$ if $a,b\ge 1$. For $b\ge 1$ we have $$\frac{\d e_{\alpha,0}}{\d q^\beta_b}=-\sum_{j\ge 0}(-1)^jc^\theta_{\alpha\beta}\Omega^{\mu,j-1}_{\theta,0}\frac{\d}{\d f^\mu_{b-1-j}}=
\begin{cases}
-c^\theta_{\alpha\beta}e_{\theta,b-1},&\text{if $b\ge 2$},\\
-\sum_{j\ge 0}(-1)^jc^\theta_{\alpha\beta}\Omega^{\mu,j-1}_{\theta,0}\frac{\d}{\d f^\mu_{-j}},&\text{if $b=1$},
\end{cases}$$ and it remains to check that $$h_\alpha:=\sum_{j\ge 0}(-1)^j\Omega^{\mu,j-1}_{\alpha,0}\frac{\d}{\d f^\mu_{-j}}\in T_f{\mathcal{C}}'.$$ We can easily see that $$M^\nu_\alpha h_\nu=e_{\alpha,0}+\sum_{i\ge 2}q^\nu_i c^\theta_{\nu\alpha}e_{\theta,i-1},$$ and, since the matrix $M$ is invertible, we get $h_\alpha\in T_f{\mathcal{C}}'$.\
Let us finally prove that ${\mathcal{C}}={\mathcal{C}}'$. Note that $\left.{\widetilde{q}}^\beta_b(q^*_*)\right|_{t^*_{\ge 1}=0}=\delta_{b,0}(t^\beta_0+c^\beta)-\delta_{b,1}A^\beta$ and $$\begin{gathered}
\left.{\widetilde{\mathcal{F}}}^{\alpha,a}(q^*_*)\right|_{t^*_{\ge 1}=0}=\Omega^{\alpha,a+1}_{\un,0}={\mathcal{F}}^{\alpha,a}\Big|_{\substack{t^\gamma_0\mapsto t^\gamma_0+c^\gamma\\t^*_{\ge 1}=0}},\quad a\ge 0.\end{gathered}$$ Thus, the $J$-functions of the cones ${\mathcal{C}}$ and ${\mathcal{C}}'$ coincide, which implies that ${\mathcal{C}}={\mathcal{C}}'$. The proposition is proved.
Consider the canonical coordinates $u^i(t^*)$ of our flat F-manifold. Making an appropriate shift we can assume that $u^i(0)=0$. Consider the matrices $H$ and $\Psi$ constructed in Section \[subsection:metric for a flat F-manifold\], and the matrices $R_i$ given by Proposition \[proposition:matrices R\_k\]. We consider these matrices as functions of the variables $t^\alpha$, $H=H(t^*)$, $\Psi=\Psi(t^*)$, $R_i=R_i(t^*)$. Denote $$R(z,t^*):=1+\sum_{i\ge 1}R_i(t^*)z^i.$$ Let us prove that $${\mathcal{C}}=\Psi^{-1}(0)R^{-1}(-z,0)\Psi(0){\mathcal{C}}^{\mathrm{const}}.$$ Since ${\mathcal{C}}_N^{\mathrm{triv}}=H^{-1}(0)\Psi(0){\mathcal{C}}^{\mathrm{const}}$, it is sufficient to prove that $${\mathcal{C}}^{\mathrm{triv}}_N=H^{-1}(0)R(-z,0)\Psi(0){\mathcal{C}}.$$
We say that a Laurent series $\sum_{i\in{\mathbb Z}}f_i(t^*)z^i$, $f_i\in{\mathbb C}[[t^1,\ldots,t^N]]$, is [*admissible*]{} if for each $i\le 0$ the formal power series $f_i(t^*)$ consists of monomials $t^{\alpha_1}\cdots t^{\alpha_k}$ with $k\ge -i$. Note that the product of any two admissible Laurent series is well defined and is also an admissible Laurent series. Any Laurent series $f$ with matrix coefficients can be considered as a matrix, whose entries are Laurent series, and we say that $f$ is admissible, if all the entries are admissible.\
Consider the following Laurent series with matrix coefficients: $$\begin{gathered}
S_1(z,t^*):=\sum_{j\ge 0}\Omega^{j-1}_0(t^*)z^{-j},\qquad S_2(z,t^*):=e^{U(t^*)/z}R(z,t^*)\Psi(t^*),\end{gathered}$$ where $U(t^*)={\mathrm{diag}}(u^1(t^*),\ldots,u^N(t^*))$ and the matrices $\Omega^{j-1}_0(t^*)$ define the ancestor calibration for our flat F-manifold. Since $U(0)=0$, the Laurent series $S_2(z,t^*)$ is well defined and is admissible. By Lemma \[lemma:technical property of ancestor potentials\], the Laurent series $S_1(z,t^*)$ is also admissible and we can consider the product $$S_2(-z,t^*)S_1^{-1}(-z,t^*).$$ Clearly, $R(-z,0)\Psi(0)=S_2(-z,0)S_1^{-1}(-z,0)$. Define an ancestor cone ${\mathcal{C}}_0$ by $${\mathcal{C}}_0:=H^{-1}(0)S_2(-z,0)S_1^{-1}(-z,0){\mathcal{C}}.$$ We have to prove that ${\mathcal{C}}_0={\mathcal{C}}^{\mathrm{triv}}_N$.\
Both Laurent series $S_1$ and $S_2$ satisfy the same differential equation $$d S_{1,2}=z^{-1}S_{1,2}d\Omega^0_0.$$ The equation $d S_1=z^{-1}S_1 d\Omega^0_0$ is a part of the definition of a calibration. In order to prove the equation $d S_2=z^{-1}S_2 d\Omega^0_0$, one should use the formula $d\Omega^0_0=\Psi^{-1}dU\Psi$ together with formulas and . This implies that the product $S_2(z,t^*)S_1^{-1}(z,t^*)$ does not depend on the variables $t^\alpha$. By Proposition \[proposition:another parameterization\], the cone ${\mathcal{C}}_0$ can be parameterized as follows: $${\mathcal{C}}_0=\left\{H^{-1}(0)S_2(-z,0)S_1^{-1}(-z,0)S_1(-z,t^*)\left(-zA^\alpha\phi_\alpha+\sum_{i\ge 1}t^\alpha_i\phi_\alpha z^i\right)\right\}.$$ Since $S_2(-z,0)S_1^{-1}(-z,0)=S_2(-z,t^*)S_1^{-1}(-z,t^*)$, we get the following parameterization of the cone ${\mathcal{C}}_0$: $${\mathcal{C}}_0=\left\{e^{-U(t^*)/z}H^{-1}(0)R(-z,t^*)\Psi(t^*)\left(-z A^\alpha\phi_\alpha+\sum_{i\ge 1}t^\alpha_i\phi_\alpha z^i\right)\right\}.$$
Consider a family of vectors $f(z,t^*)\in z{\mathcal{H}}_+$, depending on $t^1,\ldots,t^N$, defined by $$f(z,t^*):=\Psi^{-1}(t^*)R^{-1}(-z,t^*)H(0)\left(-z\sum_{\alpha=1}^N\phi_\alpha\right).$$ This family has the form $$f(z,t^*)=-zA^\alpha\phi_\alpha+\sum_{i\ge 1}f^\alpha_i(t^*)\phi_\alpha z^i,\quad f^\alpha_i\in{\mathbb C}[[t^*]],\quad f^\alpha_i(0)=0,$$ and we clearly have $$H^{-1}(0)R(-z,t^*)\Psi(t^*)f(z,t^*)=-z\sum_{\alpha=1}^N\phi_\alpha.$$ Therefore, $$e^{-U(t^*)/z}\left(-z\sum_{\alpha=1}^N\phi_\alpha\right)\in{\mathcal{C}}_0.$$
We see that the cone ${\mathcal{C}}_0$ contains the family of vectors $e^{-U(t^*)/z}\left(-z\sum_{\alpha=1}^N\phi_\alpha\right)$ parameterized by $t^1,\ldots,t^N$. This implies that the $J$-function of the cone ${\mathcal{C}}_0$ is $\sum_{\alpha=1}^N ze^{t^\alpha/z}\phi_\alpha$. Therefore, ${\mathcal{C}}_0={\mathcal{C}}^{{\mathrm{triv}}}_N$. This completes the proof of the theorem.
Combining this theorem with Proposition \[proposition:all from ancestor\], we get the following result.
Any generalized Givental cone ${\mathcal{C}}$ such that the algebra structure of the flat F-manifold at ${\overline{t}}_{{\mathrm{orig}}}$ is semisimple can be obtained from the cone ${\mathcal{C}}^{\mathrm{triv}}_N$ by the following composition of operators: $${\mathcal{C}}=S(z)R(z)M{\mathcal{C}}^{\mathrm{triv}}_N,$$ for some $S(z)\in G_-$, $R(z)\in G_+$ and $M\in{\mathrm{GL}}({\mathbb C}^N)$.\
F-cohomological field theories {#section:F-CohFT}
==============================
F-cohomological field theories (F-CohFTs for short) were introduced in [@BR18] as a generalization of the notion of cohomological field theory (or CohFT) [@KM94] and of partial cohomological field theory [@LRZ15]. We recall here their definition and their relation with Frobenius and flat F-manifolds. In what follows we denote by ${\overline{\mathcal M}}_{g,n}$ the Deligne–Mumford moduli space of genus $g$ stable curves with $n$ marked points, where $g,n\geq 0$ and $2g-2+n>0$.
F-CohFTs, partial CohFTs and CohFTs
-----------------------------------
We will denote by $H^*(X)$ the cohomology ring with coefficients in ${\mathbb C}$ of a topological space $X$. When considering the moduli space of stable curves, $X={\overline{\mathcal M}}_{g,n}$, the cohomology ring $H^2k({\overline{\mathcal M}}_{g,n})$ can optionally be replaced by the Chow ring $A^k({\overline{\mathcal M}}_{g,n})$, $k\geq 0$.
\[definition:F-CohFT\] An F-cohomological field theory (or F-CohFT) is a system of linear maps $$c_{g,n+1}\colon V^*\otimes V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1}),\quad 2g-1+n>0,$$ where $V$ is an arbitrary finite dimensional vector space, together with a special element $e\in V$, called the unit, such that, chosen any basis $e_1,\ldots,e_{\dim V}$ of $V$ and the dual basis $e^1,\ldots,e^{\dim V}$ of $V^*$, the following axioms are satisfied:
- the maps $c_{g,n+1}$ are equivariant with respect to the $S_n$-action permuting the $n$ copies of $V$ in $V^*\otimes V^{\otimes n}$ and the last $n$ marked points in ${\overline{\mathcal M}}_{g,n+1}$, respectively.
- $\pi^* c_{g,n+1}(e^{\alpha_0}\otimes \otimes_{i=1}^n e_{\alpha_i}) = c_{g,n+2}(e^{\alpha_0}\otimes \otimes_{i=1}^n e_{\alpha_i}\otimes e)$ for $1 \leq\alpha_0,\alpha_1,\ldots,\alpha_n\leq \dim V$, where $\pi\colon{\overline{\mathcal M}}_{g,n+2}\to{\overline{\mathcal M}}_{g,n+1}$ is the map that forgets the last marked point.\
Moreover, $c_{0,3}(e^{\alpha}\otimes e_\beta \otimes e) = \delta^\alpha_\beta$ for $1\leq \alpha,\beta\leq \dim V$.
- ${\mathrm{gl}}^* c_{g_1+g_2,n_1+n_2+1}(e^{\alpha_0}\otimes \otimes_{i=1}^{n_1+n_2} e_{\alpha_i}) = c_{g_1,n_1+2}(e^{\alpha_0}\otimes \otimes_{i\in I} e_{\alpha_i} \otimes e_\mu)\otimes c_{g_2,n_2+1}(e^{\mu}\otimes \otimes_{j\in J} e_{\alpha_j})$ for $1 \leq\alpha_0,\alpha_1,\ldots,\alpha_{n_1+n_2}\leq \dim V$, where $I \sqcup J = \{2,\ldots,n_1+n_2+1\}$, $|I|=n_1$, $|J|=n_2$, and ${\mathrm{gl}}\colon{\overline{\mathcal M}}_{g_1,n_1+2}\times{\overline{\mathcal M}}_{g_2,n_2+1}\to {\overline{\mathcal M}}_{g_1+g_2,n_1+n_2+1}$ is the corresponding gluing map.
There is an obvious generalization of the notion of F-CohFT, where the maps $c_{g,n+1}$ take value in $H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})\otimes K$, where $K$ is a ${\mathbb C}$-algebra. We will call such objects [*F-cohomological field theories with coefficients in $K$*]{}.\
Given an F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$, $\dim V=N$, and a basis $e_1,\ldots,e_N\in V$, an $N$-tuple of functions $(F^1,\ldots,F^N)$ satisfying equations and can be constructed as the following generating functions: $$\label{eq:genus 0 vector potential of an F-CohFT}
F^\alpha(t^1,\ldots,t^N):=\sum_{n\geq 2}\frac{1}{n!}\sum_{1\leq\alpha_1,\ldots,\alpha_n\leq N}\left(\int_{{\overline{\mathcal M}}_{0,n+1}}c_{0,n+1}(e^\alpha\otimes\otimes_{i=1}^n e_{\alpha_i})\right)\prod_{i=1}^n t^{\alpha_i},$$ thus yielding an associated flat F-manifold structure on a formal neighbourhood of $0$ in $V$. Note that the unit vector field of the flat F-manifold is $\frac{\d}{\d t^\un}=A^\alpha\frac{\d}{\d t^\alpha}$, where $A^\alpha e_\alpha=e$. More in general, we have the following result involving genus $0$ intersection numbers of the F-CohFT with psi classes, where $\psi_i \in H^2({\overline{\mathcal M}}_{g,n})$, $1\leq i\leq n$, is the first Chern class of the $i$-th tautological bundle on ${\overline{\mathcal M}}_{g,n}$ whose fiber at a point representing the class of a marked stable curve $(C,x_1,\ldots,x_n)$ is the cotangent line to $C$ at $x_i\in C$.
For $1\leq \alpha\leq N$ and $a\geq 0$, the formal power series $$\label{eq:genus 0 ancestor potential of an F-CohFT}
{\mathcal{F}}^{\alpha,a}(t^*_*):=\sum_{n\geq 2}\frac{1}{n!}\sum_{\substack{1\leq\alpha_1,\ldots,\alpha_n\leq N\\ a_1,\ldots,a_n\geq 0}}\left(\int_{{\overline{\mathcal M}}_{0,n+1}}c_{0,n+1}(e^\alpha\otimes\otimes_{i=1}^n e_{\alpha_i})\psi_1^{a}\prod_{i=1}^n\psi_{i+1}^{a_i}\right)\prod_{i=1}^n t^{\alpha_i}_{a_i}$$ form a sequence of ancestor vector potentials of a flat F-manifold.
We are going to use the characterization of ancestor vector potentials given by Proposition \[proposition:descendant vector potentials,equivalent approach\] and equation . Therefore, we need to show that equations , , , and are valid for the power series .\
Equation follows from the fact that the formal power series ${\mathcal{F}}^{\alpha,a}$ are always at least quadratic in the variables $t^*_*$, by definition .\
Equations and can be proved by computing $\frac{\d {\mathcal{F}}^{\alpha,a}}{\d t^\un_0}$ and $\frac{\d {\mathcal{F}}^{\alpha,a}}{\d t^\un_1}$ using Axiom (ii) of Definition \[definition:F-CohFT\] together with the equations $$\begin{gathered}
\label{eq:monomial in psi's and forgetful map}
\psi_1^{a_1}\ldots\psi_n^{a_n} =\pi^* (\psi_1^{a_1}\ldots\psi_n^{a_n})+ \sum_{i=1}^n \pi^*(\psi_1^{a_1}\ldots\psi_i^{a_i-1}\ldots\psi_n^{a_n}) \delta^{\{i,n+1\}}_0\end{gathered}$$ and $$\pi_*\psi_{n+1} = 2g-2+n,$$ respectively, where $\pi\colon{\overline{\mathcal M}}_{g,n+1}\to{\overline{\mathcal M}}_{g,n}$ is the morphism forgetting the last marked point and $\delta^{\{i,n+1\}}_0$ is the class of the irreducible boundary divisor in ${\overline{\mathcal M}}_{g,n+1}$ that is the closure of the locus of marked stable curves $(C,x_1,\ldots,x_{n+1})$ with two irreducible components, one of genus $0$ carrying the marked points $x_i$ and $x_{n+1}$ and the other of genus $g$ carrying all the other markings.\
Proving equation requires Axiom (iii) of Definition \[definition:F-CohFT\] together with the following formula for the psi class $\psi_i$, valid in genus $0$ for fixed $i,j,k \in \{1,\ldots,n+1\}$ with $i\neq j\neq k\neq i$: $$\begin{gathered}
\label{eq:psi class in genus 0}
\psi_i = \sum_{\substack{I\subset \{1,\ldots,n+1\}\\|I|\geq 2,\, i\in I,\, j,k \in I^c}} \delta^I_0 \in H^2({\overline{\mathcal M}}_{0,n+1}),\end{gathered}$$ where $\delta^I_0$ is the closure of the locus of genus $0$ marked stable curves $(C,x_1,\ldots,x_{n+1})$ with two irreducible components, one carrying the marked points with labels in $I$ and the other carrying the marked points with labels in the complement $I^c=\{1,\ldots,n+1\}\backslash I$. Applying the formula to the case $i,j>0$, $k=1$, one obtains equation .\
Summing two equations , where $i$ and $j$ are swapped, we obtain $$\psi_i+\psi_j = \sum_{\substack{I\subset \{1,\ldots,n+1\}\\|I|,|I^c|\geq 2,\, i\in I,\, j\in I^c}} \delta^I_0 \in H^2({\overline{\mathcal M}}_{0,n+1}),$$ which implies, taking $i=1$, equation .
[@LRZ15]\[definition:pCohFT\] A partial CohFT is a system of linear maps $$c_{g,n}\colon V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n}),\quad 2g-2+n>0,$$ where $V$ is an arbitrary finite dimensional vector space, together with a special element $e\in V$, called the unit, and a symmetric nondegenerate bilinear form $\eta\in (V^*)^{\otimes 2}$, called a metric, such that, chosen any basis $e_1,\ldots,e_{\dim V}$ of $V$, the following axioms are satisfied:
- the maps $c_{g,n}$ are equivariant with respect to the $S_n$-action permuting the $n$ copies of $V$ in $V^{\otimes n}$ and the $n$ marked points in ${\overline{\mathcal M}}_{g,n}$, respectively.
- $\pi^* c_{g,n}( \otimes_{i=1}^n e_{\alpha_i}) = c_{g,n+1}(\otimes_{i=1}^n e_{\alpha_i}\otimes e)$ for $1 \leq\alpha_1,\ldots,\alpha_n\leq \dim V$, where $\pi\colon{\overline{\mathcal M}}_{g,n+1}\to{\overline{\mathcal M}}_{g,n}$ is the map that forgets the last marked point.\
Moreover $c_{0,3}(e_{\alpha}\otimes e_\beta \otimes e) =\eta(e_\alpha\otimes e_\beta) =:\eta_{\alpha\beta}$ for $1\leq \alpha,\beta\leq \dim V$.
- ${\mathrm{gl}}^* c_{g_1+g_2,n_1+n_2}( \otimes_{i=1}^{n_1+n_2} e_{\alpha_i}) = \eta^{\mu \nu}c_{g_1,n_1+1}(\otimes_{i\in I} e_{\alpha_i} \otimes e_\mu)\otimes c_{g_2,n_2+1}(\otimes_{j\in J} e_{\alpha_j}\otimes e_\nu)$ for $1\leq\alpha_1,\ldots,\alpha_{n_1+n_2}\leq \dim V$, where $I \sqcup J = \{1,\ldots,n_1+n_2\}$, $|I|=n_1$, $|J|=n_2$, and ${\mathrm{gl}}\colon{\overline{\mathcal M}}_{g_1,n_1+1}\times{\overline{\mathcal M}}_{g_2,n_2+1}\to {\overline{\mathcal M}}_{g_1+g_2,n_1+n_2}$ is the corresponding gluing map and where $\eta^{\alpha\beta}$ is defined by $\eta^{\alpha \mu}\eta_{\mu \beta} = \delta^\alpha_\beta$ for $1\leq \alpha,\beta\leq \dim V$.
Clearly, given a partial CohFT $c_{g,n}\colon V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$, the system of linear maps $c^\bullet_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ defined as $c^\bullet_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^n e_{\alpha_i}):=\eta^{\alpha_0\mu}c_{g,n+1}(e_\mu\otimes\otimes_{i=1}^n e_{\alpha_i})$ forms an F-CohFT, called the associated F-CohFT.
Given a partial CohFT $c_{g,n}\colon V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$, $\dim V=N$, and a basis $e_1,\ldots,e_N\in V$, a function $F(t^1,\ldots,t^N)$ satisfying equations and can be constructed as the following generating function: $$F(t^1,\ldots,t^N):=\sum_{n\geq 3}\frac{1}{n!}\sum_{1\leq\alpha_1,\ldots,\alpha_n\leq N}\left(\int_{{\overline{\mathcal M}}_{0,n}}c_{0,n}(\otimes_{i=1}^n e_{\alpha_i})\right)\prod_{i=1}^n t^{\alpha_i},$$ thus yielding an associated Frobenius manifold structure on a formal neighbourhood of $0$ in $V$.
\[definition:CohFT\] A CohFT is a partial CohFT $c_{g,n}\colon V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$ such that the following extra axiom is satisfied:
- ${\mathrm{gl}}^* c_{g+1,n}(\otimes_{i=1}^n e_{\alpha_i}) = c_{g,n+2}(\otimes_{i=1}^n e_{\alpha_i}\otimes e_{\mu}\otimes e_\nu) \eta^{\mu \nu}$ for $1 \leq\alpha_1,\ldots,\alpha_n\leq \dim V$, where ${\mathrm{gl}}\colon{\overline{\mathcal M}}_{g,n+2}\to {\overline{\mathcal M}}_{g+1,n}$ is the gluing map, which increases the genus by identifying the last two marked points.
Formal shift of an F-CohFT
--------------------------
Let $\tau^1,\ldots,\tau^N$ be formal variables. For an arbitrary F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ with a fixed basis $e_1,\ldots,e_N\in V$ consider a system of maps $$c_{g,n+1}^{{\overline{\tau}}}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})\otimes{\mathbb C}[[\tau^1,\ldots,\tau^N]]$$ defined by $$\begin{gathered}
\label{eq:formal shift of an F-CohFT}
c^{{\overline{\tau}}}_{g,n+1}(\omega\otimes \otimes_{i=1}^n v_i):=\sum_{m\geq 0} \frac{1}{m!}\pi_{m*} c_{g,n+m+1}(\omega\otimes \otimes_{i=1}^n v_i \otimes (\tau^\alpha e_\alpha)^{\otimes m}),\end{gathered}$$ where $\omega\in V^*$, $v_i\in V$, ${{\overline{\tau}}}=(\tau^1,\ldots,\tau^N)$ and $\pi_m\colon{\overline{\mathcal M}}_{g,n+m+1}\to{\overline{\mathcal M}}_{g,n+1}$ is the map that forgets the last $m$ marked points. It is straightforward to check that the maps $c^{{\overline{\tau}}}_{g,n+1}$ form an F-CohFT with coefficients in ${\mathbb C}[[\tau^1,\ldots,\tau^N]]$. Moreover, if ${\overline{F}}$ is the vector potential given by formula , then the flat F-manifold of the F-CohFT $c^{{\overline{\tau}}}$ is described by the vector potential ${\overline{F}}_{{\overline{\tau}}}$ from Section \[subsubsection:lower triangular group\]. We will call the F-CohFT $c^{{\overline{\tau}}}_{g,n+1}$ the [*formal shift*]{} of the F-CohFT $c_{g,n+1}$.
Partial and F-topological field theories
----------------------------------------
A CohFT $c_{g,n}\colon V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$ is called a [*topological field theory*]{} (TFT) when $c_{g,n}(V^{\otimes n})\subset H^0({\overline{\mathcal M}}_{g,n})$. In this case each class $c_{g,n}(\otimes_{i=1}^n e_{\alpha_i})\in H^0({\overline{\mathcal M}}_{g,n})$ is represented by a constant complex valued function on ${\overline{\mathcal M}}_{g,n}$. Evaluating such functions at a point of ${\overline{\mathcal M}}_{g,n}$ representing a maximally degenerate stable curve, i.e. a nodal curve whose irreducible components are all ${\mathbb P}^1$’s with three special (nodal or marked) points, using Axioms (iii) and (iv) of Definition \[definition:CohFT\], we see that the entire CohFT can be reconstructed from the linear map $c_{0,3}\colon V^{\otimes 3}\to H^0({\overline{\mathcal M}}_{0,3})$ only. This amounts to the datum of a Frobenius algebra with unit $(V,\bullet,\eta,e)$, where the metric is given by $\eta(e_\alpha,e_\beta)= c_{0,3}(e\otimes e_\alpha\otimes e_\beta)$ and the product is given by $e_{\alpha}\bullet e_{\beta} = c_{0,3}(e_\alpha\otimes e_\beta\otimes e_{\mu})\eta^{\mu \nu}e_{\nu}$.\
Let us generalize these notions and observations to partial and F-CohFTs.
A partial CohFT $c_{g,n}\colon V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$ is called a partial topological field theory (partial TFT) when $c_{g,n}(V^{\otimes n})\subset H^0({\overline{\mathcal M}}_{g,n})$.
An F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ is called an F-topological field theory (F-TFT) when $c_{g,n+1}(V^*\otimes V^{\otimes n})\subset H^0({\overline{\mathcal M}}_{g,n+1})$.
\[proposition:reconstruction of F-TFTs\] A partial TFT $c_{g,n}\colon V^{\otimes n}\to H^0({\overline{\mathcal M}}_{g,n})$ can be uniquely reconstructed from the linear maps $c_{0,3}\colon V^{\otimes 3}\to H^0({\overline{\mathcal M}}_{0,3})$ and $c_{1,1}\colon V\to H^0({\overline{\mathcal M}}_{1,1})$ only. This amounts to the datum of a Frobenius algebra with unit $(V,\bullet,\eta,e)$, where the metric is given by $\eta(e_\alpha,e_\beta)= c_{0,3}(e\otimes e_\alpha\otimes e_\beta)$ and the product is given by $e_{\alpha}\bullet e_{\beta} = c_{0,3}(e_\alpha\otimes e_\beta\otimes e_{\mu})\eta^{\mu \nu}e_{\nu}$, together with a special covector $\omega\in V^*$ given by $\langle\omega,e_{\alpha}\rangle = c_{1,1}(e_\alpha)$.
An F-TFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^0({\overline{\mathcal M}}_{g,n+1})$ can be uniquely reconstructed from the linear maps $c_{0,3}\colon V^*\otimes V^{\otimes 2}\to H^0({\overline{\mathcal M}}_{0,3})$ and $c_{1,1}\colon V^*\to H^0({\overline{\mathcal M}}_{1,1})$ only. This amounts to the datum of a commutative associative algebra with unit $(V,\bullet,e)$, where the product is given by $e_{\alpha}\bullet e_{\beta} = c_{0,3}(e^\mu \otimes e_\alpha\otimes e_\beta)e_{\mu}$, together with a special vector $w\in V$ given by $w=c_{1,1}(e^\alpha)e_\alpha$.
As for TFTs, the classes $c_{g,n}(\otimes_{i=1}^n e_{\alpha_i})\in H^0({\overline{\mathcal M}}_{g,n})$ for partial TFTs and $c_{g,n+1}(e^{\alpha}\otimes \otimes_{i=1}^n e_{\alpha_i})\in H^0({\overline{\mathcal M}}_{g,n+1})$ for F-TFTs are represented by constant complex valued functions on ${\overline{\mathcal M}}_{g,n}$ and ${\overline{\mathcal M}}_{g,n+1}$, respectively. Evaluating such functions at points of the moduli spaces representing stable curves with separating nodes only and whose irreducible components are either ${\mathbb P}^1$’s with three special (marked or nodal) points or elliptic curves with one special point, and using Axiom (iii) of Definition \[definition:pCohFT\] or \[definition:F-CohFT\] we obtain the desired result.
Given a CohFT, a partial CohFT or an F-CohFT, their degree zero parts are naturally a TFT, a partial TFT or an F-TFT, respectively.
Homogeneous F-CohFTs
--------------------
Since $H^*({\overline{\mathcal M}}_{g,n})$ is a graded ${\mathbb C}$-vector space, it is natural to consider the special case of F-CohFTs for which:
- the vector spaces $V$ and $V^*$ are also graded, $\deg e = 0$, and the pairing between $V$ and $V^*$ has degree $0$, i.e. $\deg e^\alpha =-\deg e_\alpha$ for a homogeneous basis $e_1,\ldots,e_{\dim V}$ of $V$,
- the maps $c_{g,n+1}\colon V^*\otimes V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$ are homogeneous of degree $\deg c_{g,n+1}$.
Then, because of Axiom (ii) in Definition \[definition:F-CohFT\], $\deg c_{g,n+1}$ does not depend on $n$ and $\deg c_{0,n+1}=0$ for any $n\geq 2$. Moreover, because of Axiom (iii), $\deg c_{g,n+1}$ is a linear function of $g$, which implies that the general form of a grading compatible with the axioms of F-CohFT is $$\deg c_{g,n+1} = \gamma g , \quad \gamma \in {\mathbb C}.$$ Thus, setting $q_\alpha:=\deg e_\alpha$, we get the following condition for the classes $c_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})$: $$\begin{gathered}
\label{eq:homogeneous F-CohFT without correction}
\deg c_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})=\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}+\gamma g,\end{gathered}$$ where $\deg$ denotes half of the cohomological degree. Now, in order to get a generalization of the usual notion of homogeneous CohFT, let us correct the left-hand side of by adding a term $\pi_{1*}c_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)$, where $r^\gamma\in{\mathbb C}$ and $\pi_1\colon{\overline{\mathcal M}}_{g,n+2}\to{\overline{\mathcal M}}_{g,n+1}$ is the map that forgets the last marked point. We finally arrive to the following definition.
An F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ is called homogeneous if $V$ is a graded vector space with a homogeneous basis $e_1,\ldots,e_{\dim V}$, $\deg e=0$, and complex constants $r^\alpha$, $1\leq \alpha\leq \dim V$, and $\gamma$ exist, such that the following condition is satisfied: $$\begin{gathered}
\operatorname{Deg}c_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})+\pi_{1*}c_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)=\\
=\left(\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}+\gamma g\right)c_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}),\end{gathered}$$ where $q_\alpha:=\deg e_\alpha$ and by $\operatorname{Deg}\colon H^*({\overline{\mathcal M}}_{g,n})\to H^*({\overline{\mathcal M}}_{g,n})$ we denote the operator that acts on $H^i({\overline{\mathcal M}}_{g,n})$ by the multiplication by $\frac{i}{2}$. The constant $\gamma$ is called the conformal dimension of our F-CohFT.
The flat F-manifold associated to a homogeneous F-CohFT is homogeneous with the Euler vector field given by $$E=\sum_{\alpha=1}^{\dim V}\left((1-q_\alpha)t^\alpha+r^\alpha\right)\frac{\d}{\d t^\alpha}.$$
Suppose that a homogeneous F-CohFT comes from a partial CohFT $c_{g,n}\colon V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n})$, with the metric $\eta$ on $V$, seen as the map $\eta\colon V^{\otimes 2}\to {\mathbb C}$, having degree $\deg \eta = -\delta$. Then our partial CohFT satisfies the condition $$\begin{gathered}
\operatorname{Deg}c_{g,n}(\otimes_{i=1}^ne_{\alpha_i})+\pi_{1*}c_{g,n+1}(\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)=\left(\sum_{i=1}^n q_{\alpha_i}+\gamma g-\delta\right)c_{g,n}(\otimes_{i=1}^ne_{\alpha_i}).\end{gathered}$$
Finally, if our partial CohFT is a CohFT, then the last property has to be compatible with the extra gluing axiom at nonseparating nodes and this imposes the further condition $\gamma=\delta$, i.e. $$\begin{gathered}
\operatorname{Deg}c_{g,n}(\otimes_{i=1}^ne_{\alpha_i})+\pi_{1*}c_{g,n+1}(\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)=\left(\sum_{i=1}^n q_{\alpha_i}+\delta(g-1)\right)c_{g,n}(\otimes_{i=1}^ne_{\alpha_i}).\end{gathered}$$ This is exactly the homogeneity condition in the definition of homogeneous CohFTs (see e.g. [@PPZ15 Definition 1.7]). Note that the constant $\delta$ is the conformal dimension of the corresponding Frobenius manifold.\
Group action on F-CohFTs {#section:group action of F-CohFTs}
========================
In this section we define a generalization of the notion of Givental group acting on the space of CohFTs (see [@PPZ15]) to a corresponding generalized Givental group acting on the space of F-CohFTs. Using this action we then present a construction of a family of F-CohFTs associated to any given flat F-manifold that is semisimple at the origin. The family is parameterized by a vector $G_0\in{\mathbb C}^N$.
$R$-matrices
------------
Given a vector space $V$, a system of linear maps $$c_{g,n+1}\colon V^*\otimes V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1}),\quad 2g-1+n>0,$$ satisfying Axioms (i) and (iii) of Definition \[definition:F-CohFT\] is called an F-CohFT without unit.
Consider now the group $G_+$ of ${\mathrm{End}}(V)$-valued power series of the form $R(z)={\mathrm{Id}}+\sum_{i\geq 1} R_iz^i$, and let us denote by $R^{-1}(z)$ the inverse element to $R(z)$ and by $R(z)^t$ the transposed ${\mathrm{End}}(V^*)$-valued power series. We refer to such an element of $G_+$ as an [*$R$-matrix*]{}.\
Let $\Gamma$ be a stable graph of genus $g$ with $n$ marked legs (see [@PPZ15 Section 0.2] for the definition) and $V(\Gamma)$, $E(\Gamma)$ be its sets of vertices and edges, each vertex $v\in V(\Gamma)$ marked with a genus $g(v)$ and with valence $n(v)$. Let $\xi_{\Gamma}\colon\prod_{v\in V(\Gamma)} {\overline{\mathcal M}}_{g(v),n(v)} \to {\overline{\mathcal M}}_{g,n}$ be the natural map whose image is the closure of the locus of stable curves whose dual graph is $\Gamma$. The degree of $\xi_\Gamma$ is $|\operatorname{Aut}\Gamma|$, the number of automorphisms of the graph $\Gamma$.\
Let $T_{g,n+1}$ be the set of stable trees of genus $g$ with $n+1$ marked legs. Then $\Gamma \in T_{g,n+1}$ can be seen as a stable rooted tree where the root is the vertex to which leg $1$ is attached and each edge $e\in E(\Gamma)$ is splitted into two half edges $e'$ and $e''$, where $e'$ is closer to the root and $e''$ is farther from the root.\
The action of $R\in G_+$ on an F-CohFT without unit $c_{g,n+1}\colon V^*\otimes V^{\otimes n} \to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ is the system of maps $$\label{eq:R-action}
(Rc)_{g,n+1} := \sum_{\Gamma \in T_{g,n+1}}\xi_{\Gamma*}\left[\prod_{v\in V(\Gamma)}c_{g(v),n(v)}R(-\psi_1)^t \prod_{k=2}^{n+1} R^{-1}(\psi_k)\prod_{e\in E(\Gamma)} \frac{{\mathrm{Id}}-R^{-1}(\psi_{e'})R(-\psi_{e''})}{\psi_{e'}+\psi_{e''}}\right].$$ Understanding how this formula gives a linear map from $V^*\otimes V^{\otimes n}$ to $H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ requires some explanation. First, the covector and the $n$ vectors are fed to the external leg terms $R(-\psi_1)^t$ and $R^{-1}(\psi_k)$, $2\leq k \leq n+1$, which are elements of $H^*({\overline{\mathcal M}}_{g,n+1})\otimes {\mathrm{End}}(V^*)$ and $H^*({\overline{\mathcal M}}_{g,n+1})\otimes {\mathrm{End}}(V)$, respectively. The result is an element in $H^*({\overline{\mathcal M}}_{g,n+1})\otimes V^*$ and $n$ elements in $H^*({\overline{\mathcal M}}_{g,n+1})\otimes V$, the first factor of which acts by multiplication in cohomology, while the second factor is fed to the $c_{g(v),n(v)}$ sitting at the vertex $v$ to which the corresponding leg is attached.\
Second, the edge term $$\frac{{\mathrm{Id}}-R^{-1}(\psi_{e'})R(-\psi_{e''}) }{\psi_{e'}+\psi_{e''}}$$ is an element of $H^*({\overline{\mathcal M}}_{g(v'),n(v')})\otimes H^*({\overline{\mathcal M}}_{g(v''),n(v'')})\otimes V\otimes V^*$, where $v'$ is the vertex to which $e'$ is attached and $v''$ is the vertex to which $e''$ is attached (indeed, the ${\mathrm{End}}(V)$-valued power series in the psi classes at the numerator is in the ideal generated by the denominator, since $R(z)R^{-1}(z)={\mathrm{Id}}$). The first two factors act by multiplication in cohomology, the third factor is fed to one of the vector entries of the $c_{g(v'),n(v')}$ sitting at the vertex $v'$ and the third factor is fed to the covector entry of the $c_{g(v''),n(v'')}$ sitting at the vertex $v''$.\
This way all entries of the vertex terms $c_{g(v),n(v)}$ are exhausted by either a leg or an edge term, and all that is left is a product of (even) cohomology classes.
Note that, unlike in the analogous formula for the $R$-action on CohFTs without unit [@PPZ15 Section 2.1], we don’t have the factor $\frac{1}{|\operatorname{Aut}(\Gamma)|}$ in formula . This is because stable trees don’t have nontrivial automorphisms.
If $c_{g,n+1}$ is an F-CohFT without unit, then $(Rc)_{g,n+1}$ is an F-CohFT without unit. The resulting action is a left group action.
The proof strictly follows the ideas of the analogous proofs, found in [@PPZ15 Section 2], for the $R$-matrix action on a CohFT. The $S_n$-equivariance of $(RC)_{g,n+1}$ follows from the $S_n$-equivariance of $c_{g,n+1}$ and the definition of the $R$-matrix action. For the pullback of $(Rc)_{g,n+1}$ to a boundary divisor of curves with dual graph $\Phi$ with two vertices and one separating edge, one follows the argument in the proof of [@PPZ15 Proposition 2.3] to show that such pullback is an expression similar to , the only differences being that the sum runs over all stable rooted trees $\Gamma$ which are degenerations of $\Phi$ and that the edge term assigned to the distinguished separating edge $e$ is simply $R^{-1}(\psi_{e'})R(-\psi_{e''})$. This shows that $(Rc)_{g,n+1}$ satisfies Axiom (iii) in Definition \[definition:F-CohFT\].\
To show that the resulting action is a left group action we follow again the argument in [@PPZ15 Proposition 2.4].
$R$-matrix action on F-CohFTs
-----------------------------
Consider the abelian group of $V$-valued power series of the form $T(z)=\sum_{i\geq 2} T_i z^i$. We refer to such power series as a [*translation*]{}. Its action on an F-CohFT without unit $c_{g,n+1}$ is given by the formula $$\label{eq:T-action}
(Tc)_{g,n+1}(\omega\otimes \otimes_{i=1}^n v_i): = \sum_{m\geq 0} \frac{1}{m!} \pi_{m*} c_{g,n+m+1}(\omega\otimes \otimes_{i=1}^n v_i \otimes \otimes_{k=n+2}^{n+m+1} T(\psi_k) ),$$ where $\omega\in V^*$ and $v_i\in V$, $1\leq i \leq n$.
If $c_{g,n+1}$ is an F-CohFT without unit, then $(Tc)_{g,n+1}$ is an F-CohFT without unit. The resulting action is an abelian group action. If two translations $T'$ and $T''$ are related by an $R$-matrix $R$ via the equation $T'(z)=R(z)T''(z)$, then $(T'Rc)_{g,n+1} = (RT''c)_{g,n+1}$.
The proof follows closely the proofs of [@PPZ15 Proposition 2.7, 2.8, 2.9] with the obvious transpositions.
In order to obtain a well-defined $R$-matrix action on F-CohFTs with unit we need to combine it with an appropriate translation. Given an F-CohFT with unit $e$ and an $R$-matrix $R$, let us define the translations $T_R'(z):=z[R(z)e-e]$ and $T_R''(z):=z[e-R^{-1}(z)e]$.
\[theorem:R-action on F-CohFTs\] If $c_{g,n+1}$ is an F-CohFT with unit $e$, then $(T_R'Rc)_{g,n+1}=(RT_R''c)_{g,n+1}$ is an F-CohFT with the same unit $e$. The resulting action is a left group action of the group $G_+$ on F-CohFTs with unit.
Again, the proof follows closely the proof of [@PPZ15 Proposition 2.12].
Following [@PPZ15] let us use the notation $$R.c:=T_R'Rc$$ for the constructed action of the group $G_+$ on F-CohFTs with unit.
\[proposition:R-invariance of F-TFTs\] The action of $G_+$ on F-CohFTs with unit leaves their degree $0$ part unchanged.
From equations and it is easy to see that both $(Rc)_{g,n+1} - c_{g,n+1}$ and $(Tc)_{g,n+1}-c_{g,n+1}$, seen as elements of $H^*({\overline{\mathcal M}}_{g,n+1})\otimes V\otimes (V^*)^{\otimes n}$, have no cohomological degree $0$ term.
The constructed action of the group $G_+$ on F-CohFTs (with unit) induces a $G_+$-action on the corresponding ancestor vector potentials. Let us prove that the latter coincides with the $G_+$-action on ancestor vector potentials constructed in Section \[subsubsection:upper triangular group\].
\[theorem:consistency of R-actions\] Consider an F-CohFT (with unit) $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ and choose a basis $e_1,\ldots,e_N\in V$. Consider the corresponding sequence of ancestor vector potentials $\overline{{\mathcal{F}}}^a$, $a\ge 0$, and an $R$-matrix $R\in G_+$. Then the sequence of ancestor vector potentials corresponding to the F-CohFT $(R.c)_{g,n+1}$ coincides with the sequence $\overline{R.{\mathcal{F}}}^a$, $a\ge 0$.
It is sufficient to check the statement of the theorem infinitesimally, i.e. to prove that for any $r=\sum_{i\ge 1}r_i z^i$, $r_i\in{\mathrm{End}}(V)$, we have $$\begin{aligned}
&\sum_{n\geq 2}\frac{1}{n!}\sum_{\substack{1\leq\alpha_1,\ldots,\alpha_n\leq N\\ a_1,\ldots,a_n\geq 0}}\left(\int_{{\overline{\mathcal M}}_{0,n+1}}\left.\frac{d}{d{\varepsilon}}\left[(e^{{\varepsilon}r}.c)_{0,n+1}(e^\alpha\otimes\otimes_{i=1}^n e_{\alpha_i})\right]\right|_{{\varepsilon}=0}\psi_1^{a}\prod_{i=1}^n\psi_{i+1}^{a_i}\right)\prod_{i=1}^n t^{\alpha_i}_{a_i}=\\
=&\left.\frac{d}{d{\varepsilon}}(e^{{\varepsilon}r}.{\mathcal{F}})^{\alpha,a}\right|_{{\varepsilon}=0},\quad 1\le\alpha\le N,\quad a\ge 0.\end{aligned}$$ Directly from the definition of the $R$-action on F-CohFTs, it is easy to see that $$\begin{aligned}
&\left.\frac{d}{d{\varepsilon}}\left[(e^{{\varepsilon}r}.c)_{0,n+1}(e^\alpha\otimes\otimes_{i=1}^n e_{\alpha_i})\right]\right|_{{\varepsilon}=0}=\\
=&\sum_{k\ge 1}(-1)^k(r_k)^\alpha_\mu c_{0,n+1}(e^\mu\otimes\otimes_{i=1}^n e_{\alpha_i})\\
&-\sum_{i=1}^n\sum_{k\ge 1}\psi_{i+1}^k(r_k)^\mu_{\alpha_i}c_{0,n+1}\left(e^\alpha\otimes \otimes_{j=1}^{i-1}e_{\alpha_j}\otimes e_\mu\otimes\otimes_{j=i+1}^n e_{\alpha_j}\right)\\
&+\sum_{\substack{I\sqcup J=\{1,\ldots,n\}\\|I|\ge 1,\,|J|\ge 2}}\sum_{p,q\ge 0}(-1)^q(r_{p+q+1})^\mu_\nu{\mathrm{gl}}_*\left(\psi_{|I|+2}^p c_{0,|I|+2}(e^{\alpha}\otimes \otimes_{i\in I} e_{\alpha_i} \otimes e_\mu)\otimes\psi_1^q c_{0,|J|+1}(e^{\nu}\otimes \otimes_{j\in J} e_{\alpha_j})\right)\\
&+\sum_{k\ge 1}(r_k)^\mu_\un\pi_{1*}\left(\psi_{n+2}^{k+1}c_{0,n+2}(e^\alpha\otimes\otimes_{i=1}^n e_{\alpha_i}\otimes e_\mu)\right),\end{aligned}$$ where ${\mathrm{gl}}\colon{\overline{\mathcal M}}_{0,|I|+2}\times{\overline{\mathcal M}}_{0,|J|+1}$ is the gluing map. Multiplying the right-hand side by $\psi_1^{a}\prod_{i=1}^n\psi_{i+1}^{a_i}$, integrating over ${\overline{\mathcal M}}_{0,n+1}$ and taking the generating series, we obtain exactly the expression on the right-hand side of formula . This completes the proof of the theorem.
Define a ${\mathrm{GL}}(V)$-action on an F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ by $$(M.c)_{g,n+1}(\omega\otimes\otimes_{i=1}^n v_i):=c_{g,n+1}(M^t\omega\otimes\otimes_{i=1}^n M^{-1}v_i),\quad M\in{\mathrm{GL}}(V),$$ where $\omega\in V^*$ and $v_i\in V$. Clearly, if $e\in V$ is the unit of the F-CohFT $c_{g,n+1}$, then $Me\in V$ is the unit of the F-CohFT $(M.c)_{g,n+1}$. The following proposition says that this ${\mathrm{GL}}(V)$-action on F-CohFTs is consistent with the ${\mathrm{GL}}({\mathbb C}^N)$-action on descendant vector potentials defined in Section \[subsection:calibrated flat F-manifold\].
Consider an F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$ and choose a basis $e_1,\ldots,e_N\in V$. Consider the corresponding sequence of ancestor vector potentials $\overline{{\mathcal{F}}}^a$, $a\ge 0$, and an element $M\in{\mathrm{GL}}(V)={\mathrm{GL}}({\mathbb C}^N)$. Then the sequence of ancestor vector potentials corresponding to the F-CohFT $(M.c)_{g,n+1}$ coincides with the sequence $\overline{M.{\mathcal{F}}}^a$, $a\ge 0$.
Direct computation.
For an F-CohFT $c_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^{\mathrm{even}}({\overline{\mathcal M}}_{g,n+1})$, an element $M\in{\mathrm{GL}}(V)$ and an $R$-matrix $R$ we will use the notation $$MR.c:=M.(R.c).$$
Construction of F-CohFTs from semisimple F-manifolds
----------------------------------------------------
Recall from Section \[section:F-CohFT\] that to an F-CohFT we associated a flat F-manifold whose vector potential is given by equation and only involves genus $0$ intersection numbers. Recall moreover from Proposition \[proposition:reconstruction of F-TFTs\] that the degree $0$ part of an F-CohFT is an F-TFT and, as such, it can be uniquely reconstructed from the datum of an associative unital algebra together with an element of such algebra. Such associative algebra coincides, by definition, with the one on the tangent space at the origin of the associated genus $0$ flat F-manifold, while its special element is genus $1$ information.
\[theorem:F-CohFT associated to a flat F-manifold\] Consider a flat F-manifold given by a vector potential $\overline{F}=(F^1,\ldots,F^N)$, where $F^\alpha\in {\mathbb C}[[t^1,\ldots,t^N]]$, that is semisimple at the origin $t^*=0$. Let $G_0=G_0^\alpha \frac{\d}{\d t^\alpha}$, where $G_0^\alpha \in {\mathbb C}$, be an element of its tangent space at the origin. There exists an F-CohFT whose associated flat F-manifold is the one considered and whose degree $0$ part is the F-TFT defined by the associative unital algebra on the tangent space at the origin of this F-manifold together with the element $G_0$ of this algebra.
Let us start with the F-TFT reconstructed uniquely from the unital associative algebra at the origin $t^*=0$ of the F-manifold and its element $G_0$, according to Proposition \[proposition:reconstruction of F-TFTs\]. It is defined on the $N$ dimensional ${\mathbb C}$-vector space generated by $e_\alpha=\frac{\d}{\d t^\alpha}$, $1\leq\alpha\leq N$, with unit $e$ and structure constants $\left.\frac{\d^2 F^\alpha}{\d t^\beta \d t^\gamma}\right|_{t^*=0}$, $1\leq \alpha,\beta,\gamma\leq N$, with respect to the basis $e_1,\ldots,e_n$. Since an F-TFT is an F-CohFT, we can associate to it a flat F-manifold and a sequence of ancestor vector potentials. By definition, it coincides with the constant part of the starting flat F-manifold, and its ancestor cone will be denoted by ${\mathcal{C}}^{\mathrm{const}}\subset {\mathcal{H}}$, according to the terminology introduced in Section \[section:reconstruction\].\
Thanks to semisimplicity at the origin, according to Theorem \[theorem:reconstruction\], there exists an $R$-matrix $R(z)\in G_+$ acting on ${\mathcal{C}}^{\mathrm{const}}$ to produce the ancestor cone ${\mathcal{C}}$ of the starting flat F-manifold.\
We then make this $R$-matrix $R(z)$ act on the F-TFT via the action described in Theorem \[theorem:R-action on F-CohFTs\], obtaining an F-CohFT. By Theorem \[theorem:consistency of R-actions\], the flat F-manifold associated to this F-CohFT coincides with the starting F-manifold. Proposition \[proposition:R-invariance of F-TFTs\] ensures that the degree $0$ part is the starting F-TFT.
Homogeneous F-CohFTs corresponding to homogeneous flat F-manifolds
------------------------------------------------------------------
Here we present a construction of a family of homogeneous F-CohFTs associated to any given homogeneous flat F-manifold that is semisimple at the origin.\
Before considering the homogeneous case, let us discuss the construction from the proof of Theorem \[theorem:F-CohFT associated to a flat F-manifold\] in more details. So let us consider a flat F-manifold given by a vector potential $\overline{F}=(F^1,\ldots,F^N)$, where $F^\alpha\in {\mathbb C}[[t^1,\ldots,t^N]]$, that is semisimple at the origin $t^*=0$, and the associated ancestor cone ${\mathcal{C}}$. As in the proof of Theorem \[theorem:reconstruction\], we consider the canonical coordinates $u^i(t^*)$ and the matrix $H={\mathrm{diag}}(H_1,\ldots,H_N)$ constructed in Section \[subsection:metric for a flat F-manifold\]. Recall that the functions $H_i$ were defined uniquely up to the rescalings $H_i\mapsto\lambda_iH_i$, where the constants $\lambda_i\in{\mathbb C}^*$ can be chosen arbitrarily. Let us make a unique choice such that $H_i|_{t^*=0}=1$ for any $i$. Consider then the matrix $\Psi$ and the matrices $R_i$ given by Proposition \[proposition:matrices R\_k\]. We consider these matrices as functions of the variables $t^\alpha$, $H=H(t^*)$, $\Psi=\Psi(t^*)$, $R_i=R_i(t^*)$. From the proof of Theorem \[theorem:reconstruction\] we know that ${\mathcal{C}}=\Psi^{-1}(0)R^{-1}(-z,0)\Psi(0){\mathcal{C}}^{\mathrm{const}}$ or, equivalently, $${\mathcal{C}}=\Psi^{-1}(0)R^{-1}(-z,0){\mathcal{C}}^{\mathrm{triv}}_N.$$
Let $V={\mathbb C}^N$ and $e_1,\ldots,e_N\in{\mathbb C}^N$ be the standard basis in ${\mathbb C}^N$. A family of F-TFTs $c^{{\mathrm{triv}},G_0}_{g,n+1}\colon V^*\otimes V^{\otimes n}\to H^0({\overline{\mathcal M}}_{g,n+1})$, parameterized by a vector $G_0=(G_0^1,\ldots,G_0^N)\in{\mathbb C}^N$, corresponding to the trivial flat F-manifold of dimension $N$ is given by $$c^{{\mathrm{triv}},G_0}_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j}):=
\begin{cases}
(G_0^{i_0})^g,&\text{if $i_0=i_1=\ldots=i_n$},\\
0,&\text{otherwise}.
\end{cases}$$ Note that $c^{{\mathrm{triv}},G_0}_{1,1}(e^i)=G_0^i$. By the proof of Theorem \[theorem:F-CohFT associated to a flat F-manifold\], a family of F-CohFTs associated to our flat F-manifold is given by $$\begin{gathered}
\label{eq:oF-G_0 F-CohFT}
c^{{\overline{F}},G_0}:=\Psi^{-1}(0)R^{-1}(-z,0).c^{{\mathrm{triv}},G_0}.\end{gathered}$$ Note that the degree zero part of $c^{{\overline{F}},G_0}_{1,1}(e^\alpha)$ is equal to $\sum_j(\Psi^{-1}(0))^\alpha_j G_0^j$.\
Suppose now that our F-manifold is homogeneous with an Euler vector field $$E=E^\alpha\frac{\d}{\d t^\alpha}=((1-q_\alpha)t^\alpha+r^\alpha)\frac{\d}{\d t^\alpha}.$$ Then, by Proposition \[proposition:homogeneity of H and gamma\], we have $E^\alpha\frac{\d H_i}{\d t^\alpha}=\delta_i H_i$ for some $\delta_i\in{\mathbb C}$, $1\le i\le N$. Consider then a unique sequence of matrices $R_i$, $i\ge 1$, given by Proposition \[proposition:unique R-matrix in the homogeneous case\]. For any $\delta\in{\mathbb C}$ define a subspace $V_\delta\subset{\mathbb C}^N$ by $$V_\delta:=\{w=(w^1,\ldots,w^N)\in{\mathbb C}^N|w^i=0\text{ if } \delta_i\ne\delta\}.$$ Let $\mathcal{D}:=\{\delta_i\}_{1\le i\le N}$. We get the decomposition ${\mathbb C}^N=\oplus_{\delta\in\mathcal{D}}V_{\delta}$.
\[theorem:homogeneous F-CohFT\] For any $1\le l\le N$ and a vector $G_0\in V_{\delta_l}$ the F-CohFT $$c^{{\overline{F}},G_0}=\Psi^{-1}(0)R^{-1}(-z,0).c^{{\mathrm{triv}},G_0}$$ is homogeneous of conformal dimension $-2\delta_l$.
The proof of the theorem is based on the following crucial result, which is true without the homogeneity assumption. For arbitrary vectors $w=(w^1,\ldots,w^N)\in({\mathbb C}^*)^N$ and $G_0=(G_0^1,\ldots,G_0^N)\in{\mathbb C}^N$ define an F-TFT $c^{w,G_0}$ with the phase space $V={\mathbb C}^N$ by $$\begin{gathered}
\label{eq:definition of cwG0}
c^{w,G_0}_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j}):=
\begin{cases}
\frac{(G_0^{i_0})^g}{(w^{i_0})^{g+n-1}},&\text{if $i_0=i_1=\ldots=i_n$},\\
0,&\text{otherwise}.
\end{cases}\end{gathered}$$ This F-TFT corresponds to the constant flat F-manifold with the vector potential $\left(\frac{(t^1)^2}{2w^1},\ldots,\frac{(t^N)^2}{2w^N}\right)$ and the unit $\sum_{i=1}^N w^i\frac{\d}{\d t^i}$.\
As at the beginning of this section, consider a flat F-manifold given by a vector potential ${\overline{F}}$, $F^\alpha\in {\mathbb C}[[t^1,\ldots,t^N]]$, that is semisimple at the origin $t^*=0$, and the associated matrices $H(t^*)$, $\Psi(t^*)$ and $R_i(t^*)$ such that $H_i(0)=1$.
\[proposition:crucial proposition\] For an arbitrary $G_0\in{\mathbb C}^N$ we have $$\begin{gathered}
\label{eq:crucial proposition}
c^{{\overline{F}},G_0,{\overline{t}}}=\Psi^{-1}(t^*)R^{-1}(-z,t^*).c^{{\overline{H}}(t^*),H^{-1}(t^*)G_0},\end{gathered}$$ where $c^{{\overline{F}},G_0,{\overline{t}}}$ is the formal shift of the F-CohFT $c^{{\overline{F}},G_0}$, ${\overline{t}}=(t^1,\ldots,t^N)$.
Obviously, both sides of equation are equal if we set $t^\alpha=0$. From equation it is clear that the left-hand side of satisfies the differential equation $$\frac{\d}{\d t^\beta}\left(c^{{\overline{F}},G_0,{\overline{t}}}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})\right)=\pi_{1*}\left(c^{{\overline{F}},G_0,{\overline{t}}}_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes e_\beta)\right).$$ Therefore, it is sufficient to check that the right-hand side of satisfies the same differential equation, $$\begin{gathered}
\frac{\d}{\d t^\beta}\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})\right)=\\
=\pi_{1*}\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes e_\beta)\right),\end{gathered}$$ or, equivalently, $$\begin{gathered}
\label{eq:differential equation for RHS}
d\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})\right)=\\
=\pi_{1*}\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes e_\beta dt^\beta)\right).\end{gathered}$$
Recall that the unit of the F-CohFT $c^{{\overline{H}},H^{-1}G_0}$ is ${\overline{H}}=\sum_{i=1}^n H_i e_i$. We have $$\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}=\Psi^{-1}R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0},$$ where $T''_{R^{-1}(-z)}=z\left({\overline{H}}-R(-z){\overline{H}}\right)$. Let us introduce the notation $$R_{\ge k}(z):=\sum_{i\ge k}R_i z^i,\quad k\ge 1.$$ We see that $$\left(T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j})=
\begin{cases}
\Omega^k_{g,n+1},&\text{if $i_0=i_1=\ldots=i_n=k$},\\
0,&\text{otherwise},
\end{cases}$$ where $$\Omega^k_{g,n}=\sum_{m\ge 0}\frac{1}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-2}}\pi_{m*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\right)\in H^*({\overline{\mathcal M}}_{g,n})\otimes{\mathbb C}[[t^1,\ldots,t^N]],$$ and $\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k$ denotes the $k$-th component of the vector $R_{\ge 1}(-\psi_i){\overline{H}}$.\
Let $$T^N_{g,n+1}:=\left\{(\Gamma,f)\left|\begin{smallmatrix}\Gamma\in T_{g,n+1}\\f\colon V(\Gamma)\to\{1,\ldots,N\}\end{smallmatrix}\right.\right\}.$$ We denote by $H(\Gamma)$ the set of half-edges of $\Gamma\in T_{g,n+1}$. A function $f\colon V(\Gamma)\to\{1,\ldots,N\}$ induces a function $H(\Gamma)\to\{1,\ldots,N\}$, denoted by the same letter $f$, by $f(h):=f(v(h))$, where $h\in H(\Gamma)$ and $v(h)$ is the vertex of $\Gamma$ incident to $h$. We denote by $l_i(\Gamma)$ the leg of $\Gamma$ marked by $i$, $1\le i\le n+1$. Let us also introduce the notations $${\widetilde{R}}(z):=R(z)\Psi,\qquad {\mathrm{ET}}(x,y):=\frac{{\mathrm{Id}}-R(-x)R^{-1}(y)}{x+y}.$$ Then the F-CohFT $\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}$ can be described in the following way: $$\begin{gathered}
\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})=\\
=\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[\prod_{v\in V(\Gamma)}\Omega_{g(v),n(v)}^{f(v)}{\widetilde{R}}^{-1}(\psi_1)^{\alpha_0}_{f(l_1)}\prod_{k=2}^{n+1}{\widetilde{R}}(-\psi_k)^{f(l_k)}_{\alpha_{k-1}}\prod_{e\in E(\Gamma)}{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')}\right].\end{gathered}$$
For a pair $(\Gamma,f)\in T^N_{g,n+1}$ and $v\in V(\Gamma)$, $e\in E(\Gamma)$ and $2\le k\le n+1$ let us introduce the following classes in $H^*\left(\prod_{v\in V(\Gamma)}{\overline{\mathcal M}}_{g(v),n(v)}\right)\otimes{\mathbb C}[[t^1,\ldots,t^N]]$: $$\begin{aligned}
{\mathrm{Cont}}_{\Gamma,f}^v:=&\prod_{\substack{{\widetilde v}\in V(\Gamma)\\{\widetilde v}\ne v}}\Omega_{g({\widetilde v}),n({\widetilde v})}^{f({\widetilde v})}{\widetilde{R}}^{-1}(\psi_1)^{\alpha_0}_{f(l_1)}\prod_{k=2}^{n+1}{\widetilde{R}}(-\psi_k)^{f(l_k)}_{\alpha_{k-1}}\prod_{e\in E(\Gamma)}{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')},\\
{\mathrm{Cont}}_{\Gamma,f}^{l_1}:=&\prod_{v\in V(\Gamma)}\Omega_{g(v),n(v)}^{f(v)}\prod_{k=2}^{n+1}{\widetilde{R}}(-\psi_k)^{f(l_k)}_{\alpha_{k-1}}\prod_{e\in E(\Gamma)}{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')},\\
{\mathrm{Cont}}_{\Gamma,f}^{l_k}:=&\prod_{v\in V(\Gamma)}\Omega_{g(v),n(v)}^{f(v)}{\widetilde{R}}^{-1}(\psi_1)^{\alpha_0}_{f(l_1)}\prod_{\substack{2\le{\widetilde{k}}\le n+1\\{\widetilde{k}}\ne k}}{\widetilde{R}}(-\psi_{{\widetilde{k}}})^{f(l_{{\widetilde{k}}})}_{\alpha_{{\widetilde{k}}-1}}\prod_{e\in E(\Gamma)}{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')},\\
{\mathrm{Cont}}_{\Gamma,f}^e:=&\prod_{v\in V(\Gamma)}\Omega_{g(v),n(v)}^{f(v)}{\widetilde{R}}^{-1}(\psi_1)^{\alpha_0}_{f(l_1)}\prod_{k=2}^{n+1}{\widetilde{R}}(-\psi_k)^{f(l_k)}_{\alpha_{k-1}}\prod_{\substack{{\widetilde{e}}\in E(\Gamma)\\{\widetilde{e}}\ne e}}{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')}.\end{aligned}$$ Then we can write $$\begin{aligned}
&d\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})\right)=\\
=&\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[\sum_{v\in V(\Gamma)}d\Omega_{g(v),n(v)}^{f(v)}{\mathrm{Cont}}_{\Gamma,f}^v+d{\widetilde{R}}^{-1}(\psi_1)^{\alpha_0}_{f(l_1)}{\mathrm{Cont}}_{\Gamma,f}^{l_1}+\sum_{k=2}^{n+1}d{\widetilde{R}}(-\psi_k)^{f(l_k)}_{\alpha_{k-1}}{\mathrm{Cont}}_{\Gamma,f}^{l_k}\right.+\\
&\left.\hspace{2.7cm}+\sum_{e\in E(\Gamma)}d{\mathrm{ET}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')}{\mathrm{Cont}}_{\Gamma,f}^e\right].\end{aligned}$$
Equations , and imply that $$\begin{aligned}
&d\left(zR_{\ge 1}(z){\overline{H}}\right)=[R_{\ge 2}(z),dU]{\overline{H}}, && z\left(dR^{-1}(z)-[\Gamma,dU]R^{-1}(z)\right)=[R^{-1}(z),dU],\\
&d{\widetilde{R}}(z)=z^{-1}[R(z),dU]\Psi, && d{\widetilde{R}}^{-1}(z)=z^{-1}\Psi^{-1}[R^{-1}(z),dU],\end{aligned}$$ which gives the following equations: $$\begin{aligned}
d\Omega^k_{g,n}=&\sum_{m\ge 0}\frac{2g+n+m-2}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\left([dU,\Gamma]{\overline{H}}\right)^k\pi_{m*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\right)+\label{eq:formula for dOmega}\\
&+\sum_{m\ge 0}\frac{1}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\pi_{(m+1)*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\cdot\left([R_{\ge 2}(-\psi_{n+m+1}),dU]{\overline{H}}\right)^k\right),\notag\\
d{\widetilde{R}}^{-1}(z)=&z^{-1}\Psi^{-1}[R^{-1}(z),dU],\label{eq:formula for dtRinv}\\
d{\widetilde{R}}(-z)=&z^{-1}[dU,R(-z)]\Psi,\label{eq:formula for dtR}\\
d{\mathrm{ET}}(x,y)=&\frac{y[R(-x),dU]R^{-1}(y)+xR(-x)[dU,R^{-1}(y)]}{x y(x+y)}.\label{eq:formula for dET}\end{aligned}$$
For the right-hand side of we compute $$\begin{gathered}
\pi_{1*}\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes e_\beta dt^\beta)\right)=\label{eq:sum on RHS of differential equation}\\
=\pi_{1*}\left(\sum_{(\Gamma,f)\in T^N_{g,n+2}}\xi_{\Gamma*}\left[{\widetilde{R}}(-\psi_{n+2})^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right).\end{gathered}$$ Define a subset ${\widetilde{T}}^N_{g,n+2}\subset T^N_{g,n+2}$ by $${\widetilde{T}}^N_{g,n+2}:=\left\{(\Gamma,f)\in T^N_{g,n+2}\left|\begin{smallmatrix}g(v(l_{n+2}(\Gamma)))=0\\n(v(l_{n+2}(\Gamma)))=3\end{smallmatrix}\right.\right\}.$$ Let us first compute the part of the sum on the right-hand side of where $(\Gamma,f)\in {\widetilde{T}}^N_{g,n+2}$: $$\begin{aligned}
&\pi_{1*}\left(\sum_{(\Gamma,f)\in{\widetilde{T}}^N_{g,n+2}}\xi_{\Gamma*}\left[{\widetilde{R}}(-\psi_{n+2})^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right)=\label{eq:sum for contracted components,1}\\
&\hspace{2cm}=\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[\left(\Psi^{-1}dU\frac{{\mathrm{Id}}-R^{-1}(\psi_1)}{\psi_1}\right)^{\alpha_0}_{f(l_1)}{\mathrm{Cont}}_{\Gamma,f}^{l_1}+\right.\label{eq:sum for contracted components,2}\\
&\left.\hspace{5.1cm}+\sum_{k=2}^{n+1}\left(\frac{{\mathrm{Id}}-R(-\psi_k)}{\psi_k}dU\Psi\right)_{\alpha_{k-1}}^{f(l_k)}{\mathrm{Cont}}_{\Gamma,f}^{l_k}+\right.\label{eq:sum for contracted components,3}\\
&\left.\hspace{5.1cm}+\sum_{e\in E(\Gamma)}\left(\frac{{\mathrm{Id}}-R(-\psi_{e'})}{\psi_{e'}}dU\frac{{\mathrm{Id}}-R^{-1}(\psi_{e''})}{\psi_{e''}}\right)_{f(e'')}^{f(e')}{\mathrm{Cont}}_{\Gamma,f}^e\right].\label{eq:sum for contracted components,4}\end{aligned}$$ Here, the sum in line corresponds to the part of the sum in line where $v(l_1(\Gamma))=v(l_{n+2}(\Gamma))$, the sum in line corresponds to the part of the sum in line where $v(l_k(\Gamma))=v(l_{n+2}(\Gamma))$ for $2\le k\le n+1$, and the sum in line corresponds to the part of the sum in line where the leg $l_{n+2}(\Gamma)$ is a unique leg incident to the vertex $v(l_{n+2}(\Gamma))$.\
Let us now compute the part of the sum on the right-hand side of where $(\Gamma,f)\in T^N_{g,n+2}\backslash{\widetilde{T}}^N_{g,n+2}$. Consider the following diagram of forgetful maps: $$\begin{gathered}
\xymatrix{
{\overline{\mathcal M}}_{g,n+m+1}\ar[d]_{{\widetilde{\pi}}_1}\ar[r]^{{\widetilde{\pi}}_m}\ar[dr]^{\pi_{m+1}} & {\overline{\mathcal M}}_{g,n+1}\ar[d]^{\pi_1}\\
{\overline{\mathcal M}}_{g,n+m}\ar[r]^{\pi_m} & {\overline{\mathcal M}}_{g,n}
}\end{gathered}$$
\[lemma:forgetful map\] Consider integers $a_1,\ldots,a_{n+1}\ge 0$ and $b_1,\ldots,b_m\ge 2$. Then we have $$\begin{gathered}
\pi_{1*}\left[\prod_{j=1}^{n+1}\psi_j^{a_j}{\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)\right]=\\
=\begin{cases}
\sum\limits_{i=1}^n\psi_i^{a_i-1}\prod\limits_{\substack{1\le j\le n\\j\ne i}}\psi_j^{a_j}\pi_{m*}\Big(\prod\limits_{j=1}^m\psi_{n+j}^{b_j}\Big)+\prod\limits_{j=1}^n\psi_j^{a_j}\pi_{m*}\Big(\sum\limits_{i=1}^m\psi_{n+i}^{b_i-1}\prod\limits_{j\ne i}\psi_{n+j}^{b_j}\Big),&\text{if $a_{n+1}=0$},\\
(2g+n+m-2)\prod\limits_{j=1}^n\psi_j^{a_j}\pi_{m*}\Big(\prod\limits_{j=1}^m\psi_{n+j}^{b_j}\Big),&\text{if $a_{n+1}=1$},\\
\prod\limits_{j=1}^n\psi_j^{a_j}\pi_{(m+1)*}\Big(\psi_{n+1}^{a_{n+1}}\prod\limits_{j=1}^m\psi_{n+1+j}^{b_j}\Big),&\text{if $a_{n+1}\ge 2$}.
\end{cases}\end{gathered}$$
Suppose $a_{n+1}=0$, then using we compute $$\begin{aligned}
&\pi_{1*}\left[\prod_{j=1}^n\psi_j^{a_j}{\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)\right]=\\
=&\pi_{1*}\left[\pi_1^*\left(\prod_{j=1}^n\psi_j^{a_j}\right){\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)+\sum_{i=1}^n\pi_1^*\left(\psi_i^{a_i-1}\prod_{j\ne i}\psi_j^{a_j}\right)\delta_0^{\{i,n+1\}}{\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)\right]=\\
=&\pi_{(m+1)*}\left[\pi_{m+1}^*\left(\prod_{j=1}^n\psi_j^{a_j}\right)\prod_{j=1}^m\psi_{n+1+j}^{b_j}+\sum_{i=1}^n\pi_{m+1}^*\left(\psi_i^{a_i-1}\prod_{j\ne i}\psi_j^{a_j}\right){\widetilde{\pi}}_m^*(\delta_0^{\{i,n+1\}})\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right].\end{aligned}$$ Since $b_1,\ldots,b_m\ge 2$, we have ${\widetilde{\pi}}_m^*\left(\delta_0^{\{i,n+1\}}\right)\prod_{j=1}^m\psi_{n+1+j}^{b_j}=\delta_0^{\{i,n+1\}}\prod_{j=1}^m\psi_{n+1+j}^{b_j}$. Therefore, we can continue the last chain of equations as follows: $$\begin{aligned}
&\pi_{(m+1)*}\left[\pi_{m+1}^*\left(\prod_{j=1}^n\psi_j^{a_j}\right)\prod_{j=1}^m\psi_{n+1+j}^{b_j}+\sum_{i=1}^n\pi_{m+1}^*\left(\psi_i^{a_i-1}\prod_{j\ne i}\psi_j^{a_j}\right)\delta_0^{\{i,n+1\}}\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right]=\\
=&\pi_{m*}\left[\pi_m^*\left(\prod_{j=1}^n\psi_j^{a_j}\right)\sum_{i=1}^m\psi_{n+i}^{b_i-1}\prod_{j\ne i}\psi_{n+j}^{b_j}+\sum_{i=1}^n\pi_m^*\left(\psi_i^{a_i-1}\prod_{j\ne i}\psi_j^{a_j}\right)\prod_{j=1}^m\psi_{n+j}^{b_j}\right]=\\
=&\prod_{j=1}^n\psi_j^{a_j}\pi_{m*}\left(\sum_{i=1}^m\psi_{n+i}^{b_i-1}\prod_{j\ne i}\psi_{n+j}^{b_j}\right)+\sum_{i=1}^n\psi_i^{a_i-1}\prod_{j\ne i}\psi_j^{a_j}\pi_{m*}\left(\prod_{j=1}^m\psi_{n+j}^{b_j}\right).\end{aligned}$$
If $a_{n+1}\ge 1$, then using again we obtain $$\begin{aligned}
\pi_{1*}\left[\prod_{j=1}^{n+1}\psi_j^{a_j}{\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)\right]=&\pi_{1*}\left[\pi_1^*\left(\prod_{j=1}^n\psi_j^{a_j}\right)\psi_{n+1}^{a_{n+1}}{\widetilde{\pi}}_{m*}\left(\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)\right]=\\
=&\pi_{(m+1)*}\left[\pi_{m+1}^*\left(\prod_{j=1}^n\psi_j^{a_j}\right){\widetilde{\pi}}_m^*(\psi_{n+1}^{a_{n+1}})\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right].\end{aligned}$$ Noticing that ${\widetilde{\pi}}_m^*(\psi_{n+1}^{a_{n+1}})\prod_{j=1}^m\psi_{n+1+j}^{b_j}=\psi_{n+1}^{a_{n+1}}\prod_{j=1}^m\psi_{n+1+j}^{b_j}$ we get $$\pi_{(m+1)*}\left[\pi_{m+1}^*\left(\prod_{j=1}^n\psi_j^{a_j}\right)\psi_{n+1}^{a_{n+1}}\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right]=\prod_{j=1}^n\psi_j^{a_j}\pi_{(m+1)*}\left(\psi_{n+1}^{a_{n+1}}\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right).$$ If $a_{n+1}\ge 2$, then this proves the lemma. If $a_{n+1}=1$, then we just note that $$\pi_{(m+1)*}\left(\psi_{n+1}\prod_{j=1}^m\psi_{n+1+j}^{b_j}\right)=(2g+n+m-2)\pi_{m*}\left(\prod_{j=1}^m\psi_{n+j}^{b_j}\right).$$
We write $$\begin{aligned}
&\pi_{1*}\left(\sum_{(\Gamma,f)\in T^N_{g,n+2}\backslash{\widetilde{T}}^N_{g,n+2}}\xi_{\Gamma*}\left[{\widetilde{R}}(-\psi_{n+2})^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right)=\notag\\
&\hspace{3cm}=\pi_{1*}\left(\sum_{(\Gamma,f)\in T^N_{g,n+2}\backslash{\widetilde{T}}^N_{g,n+2}}\xi_{\Gamma*}\left[\Psi^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right)+\label{eq:crucial proposition,sum1}\\
&\hspace{3.5cm}+\pi_{1*}\left(\sum_{(\Gamma,f)\in T^N_{g,n+2}\backslash{\widetilde{T}}^N_{g,n+2}}\xi_{\Gamma*}\left[(-\psi_{n+2})(R_1\Psi)^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right)+\label{eq:crucial proposition,sum2}\\
&\hspace{3.5cm}+\pi_{1*}\left(\sum_{(\Gamma,f)\in T^N_{g,n+2}\backslash{\widetilde{T}}^N_{g,n+2}}\xi_{\Gamma*}\left[{\widetilde{R}}_{\ge 2}(-\psi_{n+2})^{f(l_{n+2})}_\beta dt^\beta{\mathrm{Cont}}_{\Gamma,f}^{l_{n+2}}\right]\right).\label{eq:crucial proposition,sum3}\end{aligned}$$ By Lemma \[lemma:forgetful map\], the expression in line is equal to $$\begin{aligned}
&\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[\sum_{v\in V(\Gamma)}A^{f(v)}_{g(v),n(v)}{\mathrm{Cont}}_{\Gamma,f}^v+\left(\Psi^{-1}\frac{R^{-1}(\psi_1)-{\mathrm{Id}}}{\psi_1}dU\right)^{\alpha_0}_{f(l_1)}{\mathrm{Cont}}_{\Gamma,f}^{l_1}+\right.\\
&\hspace{1cm}+\sum_{k=2}^{n+1}\left(dU\frac{R(-\psi_k)-{\mathrm{Id}}}{\psi_k}\Psi\right)^{f(l_k)}_{\alpha_{k-1}}{\mathrm{Cont}}_{\Gamma,f}^{l_k}+\\
&\hspace{1cm}\left.+\sum_{e\in E(\Gamma)}\left(\frac{{\mathrm{ET}}(\psi_{e'},\psi_{e''})-{\mathrm{ET}}(\psi_{e'},0)}{\psi_{e''}}dU+dU\frac{{\mathrm{ET}}(\psi_{e'},\psi_{e''})-{\mathrm{ET}}(0,\psi_{e''})}{\psi_{e'}}\right)^{f(e')}_{f(e'')}{\mathrm{Cont}}_{\Gamma,f}^e\right],\end{aligned}$$ where $$\begin{aligned}
A^k_{g,n}=&\sum_{m\ge 0}\frac{1}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\pi_{(m+1)*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\cdot(-1)\left(dU R_{\ge 1}(-\psi_{n+m+1}){\overline{H}}\right)^k\right)=\\
=&\sum_{m\ge 0}\frac{1}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\pi_{(m+1)*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\cdot(-1)\left(dU R_{\ge 2}(-\psi_{n+m+1}){\overline{H}}\right)^k\right)+\\
&+\sum_{m\ge 0}\frac{2g+n+m-2}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\left(dU R_1{\overline{H}}\right)^k\pi_{m*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\right).\end{aligned}$$ Lemma \[lemma:forgetful map\] implies that the expressions in lines and are equal to $$\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[B^{f(v)}_{g(v),n(v)}{\mathrm{Cont}}^v_{\Gamma,f}\right]\quad\text{and}\quad\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[C^{f(v)}_{g(v),n(v)}{\mathrm{Cont}}^v_{\Gamma,f}\right],$$ respectively, where $$\begin{aligned}
B^k_{g,n}=&-\sum_{m\ge 0}\frac{2g+n+m-2}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\left(R_1dU{\overline{H}}\right)^k\pi_{m*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\right),\\
C^k_{g,n}=&\sum_{m\ge 0}\frac{1}{m!}\frac{(G_0^k)^g}{H_k^{2g+n+m-1}}\pi_{(m+1)*}\left(\prod_{i=n+1}^{n+m}(-\psi_i)\left(R_{\ge 1}(-\psi_i){\overline{H}}\right)^k\cdot\left(R_{\ge 2}(-\psi_{n+m+1})dU{\overline{H}}\right)^k\right).\end{aligned}$$
Summarizing the above computations we get $$\begin{aligned}
&\pi_{1*}\left(\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes e_\beta dt^\beta)\right)=\\
=&\sum_{(\Gamma,f)\in T^N_{g,n+1}}\xi_{\Gamma*}\left[\sum_{v\in V(\Gamma)}{\mathcal{V}}_{g(v),n(v)}^{f(v)}{\mathrm{Cont}}_{\Gamma,f}^v+{\widetilde{\mathcal{L}}}(\psi_1)^{\alpha_0}_{f(l_1)}{\mathrm{Cont}}_{\Gamma,f}^{l_1}+\sum_{k=2}^{n+1}{\mathcal{L}}(\psi_k)^{f(l_k)}_{\alpha_{k-1}}{\mathrm{Cont}}_{\Gamma,f}^{l_k}\right.+\\
&\left.\hspace{2.7cm}+\sum_{e\in E(\Gamma)}{\mathcal{E}}(\psi_{e'},\psi_{e''})^{f(e')}_{f(e'')}{\mathrm{Cont}}_{\Gamma,f}^e\right],\end{aligned}$$ where $$\begin{aligned}
{\mathcal{V}}^k_{g,n}=&A^k_{g,n}+B^k_{g,n}+C^k_{g,n}\stackrel{\text{eq.~\eqref{eq:formula for dOmega}}}{=}d\Omega^k_{g,n},\\
{\widetilde{\mathcal{L}}}(z)=&\Psi^{-1}dU\frac{{\mathrm{Id}}-R^{-1}(z)}{z}+\Psi^{-1}\frac{R^{-1}(z)-{\mathrm{Id}}}{z}dU=\Psi^{-1}z^{-1}[R^{-1}(z),dU]\stackrel{\text{eq.~\eqref{eq:formula for dtRinv}}}{=}d{\widetilde{R}}^{-1}(z),\\
{\mathcal{L}}(z)=&\frac{{\mathrm{Id}}-R(-z)}{z}dU\Psi+dU\frac{R(-z)-{\mathrm{Id}}}{z}\Psi=z^{-1}[dU,R(-z)]\Psi\stackrel{\text{eq.~\eqref{eq:formula for dtR}}}{=}d{\widetilde{R}}(-z),\\
{\mathcal{E}}(x,y)=&\frac{{\mathrm{Id}}-R(-x)}{x}dU\frac{{\mathrm{Id}}-R^{-1}(y)}{y}+\frac{{\mathrm{ET}}(x,y)-{\mathrm{ET}}(x,0)}{y}dU+dU\frac{{\mathrm{ET}}(x,y)-{\mathrm{ET}}(0,y)}{x}=\\
=&\frac{y[R(-x),dU]R^{-1}(y)+xR(-x)[dU,R^{-1}(y)]}{x y(x+y)}\stackrel{\text{eq.~\eqref{eq:formula for dET}}}{=}d{\mathrm{ET}}(x,y).\end{aligned}$$ We conclude that equation is true.
We have to check that $$\begin{gathered}
\label{eq:homogeneity property of F-CohFT}
\operatorname{Deg}c^{{\overline{F}},G_0}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})+\pi_{1*}c^{{\overline{F}},G_0}_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)=\\
=\left(\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}-2\delta_l g\right)c^{{\overline{F}},G_0}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}).\end{gathered}$$ Since $$\pi_{1*}c^{{\overline{F}},G_0}_{g,n+2}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}\otimes r^\gamma e_\gamma)=\left.E^\alpha\frac{\d}{\d t^\alpha}c^{{\overline{F}},G_0,{\overline{t}}}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})\right|_{t^*=0},$$ equation follows from the equation $$\begin{gathered}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)c^{{\overline{F}},G_0,{\overline{t}}}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})=\left(\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}-2\delta_l g\right)c^{{\overline{F}},G_0,{\overline{t}}}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}),\end{gathered}$$ which, by Proposition \[proposition:crucial proposition\], is equivalent to the equation $$\begin{gathered}
\label{eq:main homogeneity}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})=\\
=\left(\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}-2\delta_l g\right)\left(\Psi^{-1}R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i}).\end{gathered}$$ Let us prove it.\
From definition and the assumption $G_0\in V_{\delta_l}$ it follows that $$\begin{gathered}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)c^{{\overline{H}},H^{-1}G_0}_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j})=\left(\delta_{i_0}-\sum_{j=1}^n \delta_{i_j}-2\delta_l g\right)c^{{\overline{H}},H^{-1}G_0}_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j}).\end{gathered}$$ Recall that the unit of the F-CohFT $c^{{\overline{H}},H^{-1}G_0}$ is ${\overline{H}}=\sum_{i=1}^n H_i e_i$. We express $$R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}=R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0},$$ where $T''_{R^{-1}(-z)}=z({\overline{H}}-R(-z){\overline{H}})$. Consider the definition for the action of $T''_{R^{-1}(-z)}$ on the F-CohFT $c^{{\overline{H}},H^{-1}G_0}$. Let $\Delta:={\mathrm{diag}}(\delta_1,\ldots,\delta_N)$. Since $$\left(z\frac{\d}{\d z}+E^\alpha\frac{\d}{\d t^\alpha}\right)R(z)=[\Delta,R(z)],\qquad E^\alpha\frac{\d}{\d t^\alpha}{\overline{H}}=\Delta{\overline{H}},$$ we have $$\left(z\frac{\d}{\d z}+E^\alpha\frac{\d}{\d t^\alpha}\right)\left(R(z){\overline{H}}\right)=\Delta\left(R(z){\overline{H}}\right),$$ which implies that $$\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)T''_{R^{-1}(-\psi_k)}=({\mathrm{Id}}+\Delta)T''_{R^{-1}(-\psi_k)}\in V\otimes H^*({\overline{\mathcal M}}_{g,n+m+1})\otimes{\mathbb C}[[t^1,\ldots,t^N]].$$ Since the map $\pi_{m*}\colon H^*({\overline{\mathcal M}}_{g,n+m+1})\to H^*({\overline{\mathcal M}}_{g,n+1})$ decreases the cohomological degree by $2m$, we obtain $$\begin{gathered}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)\left(T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j})=\\
=\left(\delta_{i_0}-\sum_{j=1}^n \delta_{i_j}-2\delta_l g\right)\left(T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j}).\end{gathered}$$
Consider now definition of the action of $R^{-1}(-z)$ on the F-CohFT without unit (and with coefficients in ${\mathbb C}[[t^1,\ldots,t^N]]$) $T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}$. For the leg terms $R^{-1}(\psi_1)^t\in{\mathrm{End}}(V^*)\otimes H^*({\overline{\mathcal M}}_{g,n+1})\otimes{\mathbb C}[[t^1,\ldots,t^N]]$ and $R(-\psi_k)\in {\mathrm{End}}(V)\otimes H^*({\overline{\mathcal M}}_{g,n+1})\otimes{\mathbb C}[[t^1,\ldots,t^N]]$ we have $$\begin{gathered}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)R^{-1}(\psi_1)^t=[R^{-1}(\psi_1)^t,\Delta],\qquad
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)R(-\psi_k)=[\Delta,R(-\psi_k)].\end{gathered}$$ For the edge term $\frac{{\mathrm{Id}}-R(-\psi_{e'})R^{-1}(\psi_{e''})}{\psi_{e'}+\psi_{e''}}$ we compute $$\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)\frac{{\mathrm{Id}}-R(-\psi_{e'})R^{-1}(\psi_{e''})}{\psi_{e'}+\psi_{e''}}=\left[\Delta,\frac{{\mathrm{Id}}-R(-\psi_{e'})R^{-1}(\psi_{e''})}{\psi_{e'}+\psi_{e''}}\right]-\frac{{\mathrm{Id}}-R(-\psi_{e'})R^{-1}(\psi_{e''})}{\psi_{e'}+\psi_{e''}}.$$ Note that the map $$\xi_{\Gamma*}\colon H^*\left(\prod_{v\in V(\Gamma)}{\overline{\mathcal M}}_{g(v),n(v)}\right)\to H^*({\overline{\mathcal M}}_{g,n+1})$$ increases the cohomological degree by $2|E(\Gamma)|$. Summarizing the above computations for the action the operator $\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}$ on the vertex, the leg and the edge terms, we see that the contribution of each stable tree $\Gamma\in T_{g,n+1}$ in formula to a class $$\left(R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j})$$ is an eigenvector of the operator $\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}$ with the eigenvalue $\delta_{i_0}-\sum_{j=1}^n \delta_{i_j}-2\delta_l g$. Therefore, $$\begin{gathered}
\label{eq:last step for homogeneity}
\left(\operatorname{Deg}+E^\alpha\frac{\d}{\d t^\alpha}\right)\left(R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j})=\\
=\left(\delta_{i_0}-\sum_{j=1}^n \delta_{i_j}-2\delta_l g\right)\left(R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}\right)_{g,n+1}(e^{i_0}\otimes\otimes_{j=1}^n e_{i_j}).\end{gathered}$$
It remains to act by $\Psi^{-1}$ on the F-CohFT (with coefficients in ${\mathbb C}[[t^1,\ldots,t^N]]$) $$R^{-1}(-z)T''_{R^{-1}(-z)}c^{{\overline{H}},H^{-1}G_0}=R^{-1}(-z).c^{{\overline{H}},H^{-1}G_0}.$$ We have $$E^\alpha\frac{\d}{\d t^\alpha}\frac{\d u^i}{\d t^\beta}=\frac{\d}{\d t^\beta}\left(E^\alpha\frac{\d u^i}{\d t^\alpha}\right)-\frac{\d E^\alpha}{\d t^\beta}\frac{\d u^i}{\d t^\alpha}=q_\beta\frac{\d u^i}{\d t^\beta}\quad\Rightarrow\quad E^\alpha\frac{\d}{\d t^\alpha}\Psi^i_\beta=(\delta_i+q_\beta)\Psi^i_\beta.$$ Together with equation this immediately implies equation .
### Example: extended $2$-spin theory in all genera
Let us apply the above construction to the flat F-CohFT of the extended $2$-spin theory (see Section \[subsubsection:example1\]): $$F^1(t^1,t^2)=\frac{(t^1)^2}{2},\qquad F^2(t^1,t^2)=t^1t^2-\frac{(t^2)^3}{12},\qquad E=t^1\frac{\d}{\d t^1}+\frac{1}{2}t^2\frac{\d}{\d t^2}.$$ The unit is $\frac{\d}{\d t^1}$. The flat F-manifold is not semisimple at the origin, so we consider it around a semisimple point $(0,\tau)$, $\tau\in{\mathbb C}^*$. We have $\delta_1=0$ and $\delta_2=-\frac{1}{2}$. By Theorem \[theorem:homogeneous F-CohFT\], there exist two families of homogeneous F-CohFTs with the associated flat F-manifolds given by the vector potential ${\overline{F}}_{(0,\tau)}$.\
First, for any $\lambda\in{\mathbb C}^*$ the F-CohFT $$c^{{\overline{F}}_{(0,\tau)},(\lambda,0)}=\left.\left(\Psi^{-1}R^{-1}(-z)\right)\right|_{(t^1,t^2)=(0,\tau)}.c^{{\mathrm{triv}},(\lambda,0)}$$ is homogeneous of conformal dimension $0$. Here the matrices $\Psi$ and $R(z)={\mathrm{Id}}+\sum_{k\ge 1}R_k z^k$ were computed in Section \[subsubsection:example1\]. It is easy to see that $$c^{{\overline{F}}_{(0,\tau)},(\lambda,0)}_{g,k+l+1}(e^1\otimes e_1^{\otimes k}\otimes e_2^{\otimes l})=\begin{cases}
0,&\text{if $l\ge 1$},\\
\lambda^g\in H^0({\overline{\mathcal M}}_{g,k+1}),&\text{if $l=0$}.
\end{cases}$$ An argument from [@BR18 Section 6] shows that the F-CohFT $c^{{\overline{F}}_{(0,\tau)},(\lambda,0)}$ doesn’t have a limit when $\tau\to 0$.\
Second, for any $\lambda\in{\mathbb C}^*$ the F-CohFT $$c^{{\overline{F}}_{(0,\tau)},(0,\lambda)}=\left.\left(\Psi^{-1}R^{-1}(-z)\right)\right|_{(t^1,t^2)=(0,\tau)}.c^{{\mathrm{triv}},(0,\lambda)}$$ is homogeneous of conformal dimension $1$. On the other hand, in [@BR18 Theorem 3.9] the authors constructed a homogeneous F-CohFT $c^{2,{\mathrm{ext}}}$, also of conformal dimension $1$, with the associated flat F-manifold given by the vector potential ${\overline{F}}$. Consider its formal shift $c^{2,{\mathrm{ext}},(0,\tau)}$. The property (see [@BR18 Theorem 3.9]) $$\begin{gathered}
\deg c^{2,{\mathrm{ext}}}_{g,n+1}(e^{\alpha_0}\otimes\otimes_{i=1}^ne_{\alpha_i})=\left(\sum_{i=1}^n q_{\alpha_i}-q_{\alpha_0}+g\right),\quad q_1=0,\quad q_2=\frac{1}{2},\end{gathered}$$ together with the fact that $q_\alpha<1$ implies that the sum on the right-hand side of is finite. Therefore, the F-CohFT $c^{2,{\mathrm{ext}},(0,\tau)}$ is well defined for any value of $\tau$. It is interesting to compare the F-CohFTs $c^{{\overline{F}}_{(0,\tau)},(0,\lambda)}$ and $c^{2,{\mathrm{ext}},(0,\tau)}$.\
Using the properties of the F-CohFT $c^{2,{\mathrm{ext}}}$ [@BR18 Theorem 3.9], it is easy to compute that $$c^{2,{\mathrm{ext}},(0,\tau)}_{1,1}(e^\alpha)=
\begin{cases}
0,&\text{if $\alpha=1$},\\
\tau\in H^0({\overline{\mathcal M}}_{1,1}),&\text{if $\alpha=2$}.
\end{cases}$$ On the other hand, the degree zero part of $c^{{\overline{F}}_{(0,\tau)},(0,\lambda)}_{1,1}(e^\alpha)$ is equal to $$\lambda\left.(\Psi^{-1})^\alpha_2\right|_{(t^1,t^2)=(0,\tau)}=\begin{cases}
0,&\text{if $\alpha=1$},\\
-\frac{\lambda}{\tau},&\text{if $\alpha=2$}.
\end{cases}$$ We see that the degree zero parts of the F-CohFTs $c^{{\overline{F}}_{(0,\tau)},(0,-\tau^2)}$ and $c^{2,{\mathrm{ext}},(0,\tau)}$ coincide. The whole F-CohFTs can not coincide because, for example, $c^{{\overline{F}}_{(0,\tau)},(0,-\tau^2)}_{g,n+1}(e^1\otimes e_1^{\otimes n})=0$ for $g\ge 1$, and $c^{2,{\mathrm{ext}},(0,\tau)}_{g,n+1}(e^1\otimes e_1^{\otimes n})=\lambda_g$. However, we expect that the F-CohFTs $c^{{\overline{F}}_{(0,\tau)},(0,-\tau^2)}$ and $c^{2,{\mathrm{ext}},(0,\tau)}$ coincide after the restriction to the moduli space of curves of compact type $\cM^{\mathrm{ct}}_{g,n+1}$ (those whose dual graph is a tree). If this is true, then it would be interesting to study whether the existence of the limit $\lim_{\tau\to 0}c^{{\overline{F}}_{(0,\tau)},(0,-\tau^2)}$ gives new relations in the cohomology or Chow ring of $\cM^{\mathrm{ct}}_{g,n+1}$.\
We finally remark that the partial failure, just observed for the extended $2$-spin F-CohFT, of the Givental-type theory to reconstruct an F-CohFT from its F-TFT and genus $0$ restriction is not unexpected. Indeed, a crucial difference between the $R$-matrix action on F-CohFTs and the corresponding $R$-matrix action on CohFTs [@PPZ15 Section 2.1] is that in the sum runs over stable trees only, instead of all stable graphs. This restriction seems natural, as the vertex contributions, i.e. the maps $c_{g,n+1}$, need one input and $n$ outputs, but it is clear that, in general, some parts of the full F-CohFT, supported in particular on ${\overline{\mathcal M}}_{g,n+1}\setminus \cM_{g,n+1}^{\mathrm{ct}}$, can be lost. One can hence try to enlarge the Givental group (for instance introducing non-separating-edge contributions different from the $R$-matrix) and modify the action adding back stable graphs that are not trees, in the effort to recover the lost transitivity.\
Alternatively, one can look at F-CohFTs $c_{g,n+1}$ as maps to the cohomology groups $H^*(\cM_{g,n+1}^{\mathrm{ct}})$ instead of $H^*({\overline{\mathcal M}}_{g,n+1})$ (simply by restriction) and, since the restriction of any cohomology class on ${\overline{\mathcal M}}_{g,n+1}$ supported on ${\overline{\mathcal M}}_{g,n+1} \setminus \cM_{g,n+1}^{\mathrm{ct}}$ to $\cM_{g,n+1}^{\mathrm{ct}}$ is zero, any contribution to the action of stable graphs that are not trees becomes irrelevant.\
Moreover, given an F-CohFT $c_{g,n+1}$ on $\cM_{g,n+1}^{\mathrm{ct}}$, it is possible to produce a canonical F-CohFT on ${\overline{\mathcal M}}_{g,n+1}$ by multiplying it by the top Chern class $\lambda_g$ of the Hodge bundle on ${\overline{\mathcal M}}_{g,n+1}$. The result is a well defined F-CohFT on ${\overline{\mathcal M}}_{g,n+1}$, because the restriction of the class $\lambda_g$ to ${\overline{\mathcal M}}_{g,n+1} \setminus \cM_{g,n+1}^{\mathrm{ct}}$ is zero. Notice also that multiplying by $\lambda_g$ commutes with the $R$-matrix action on F-CohFTs. This is particularly relevant in view of future applications to the double ramification hierarchy, see [@Bur15; @BR16; @BR18], which only depends on $\lambda_g \cdot c_{g,n+1}$.\
[KMS18]{}
A. Arsie, P. Lorenzoni. [*Flat F-manifolds, Miura invariants, and integrable systems of conservation laws*]{}. Journal of Integrable Systems [**3**]{} (2018), no. 1, xyy004.
A. Arsie, P. Lorenzoni. [*Complex reflection groups, logarithmic connections and bi-flat F-manifolds*]{}. Letters in Mathematical Physics [**107**]{} (2017), no. 10, 1919–1961.
A. Arsie and P. Lorenzoni. [*F-manifolds, multi-flat structures and Painlevé transcendents*]{}. arXiv:1501.06435v5 (2015), accepted for publication in the Asian Journal of Mathematics.
A. Arsie, P. Lorenzoni. [*From the Darboux–Egorov system to bi-flat F-manifolds*]{}. Journal of Geometry and Physics [**70**]{} (2013), 98–116.
A. Basalaev, A. Buryak. [*Open WDVV equations and Virasoro constraints*]{}. Arnold Mathematical Journal [**5**]{} (2019), no. 2–3, 145–186.
A. Buryak. [*Double ramification cycles and integrable hierarchies*]{}. Communications in Mathematical Physics 336 (2015), no. 3, 1085–1107.
A. Buryak. [*Extended $r$-spin theory and the mirror symmetry for the $A_{r-1}$-singularity*]{}. arXiv:1802.07075. To appear in the Moscow Mathematical Journal.
A. Buryak, E. Clader, R. J. Tessler. [*Closed extended $r$-spin theory and the Gelfand–Dickey wave function*]{}. Journal of Geometry and Physics 137 (2019), 132–153.
A. Buryak, P. Rossi. [*Recursion relations for double ramification hierarchies*]{}. Communications in Mathematical Physics 342 (2016), no. 2, 533–568.
A. Buryak, P. Rossi. [*Extended $r$-spin theory in all genera and the discrete KdV hierarchy*]{}. arXiv:1806.09825v2.
L. David, I. A. B. Strachan. [*Dubrovin’s duality for F-manifolds with eventual identities*]{}. Advances in Mathematics [**226**]{} (2011), no. 5, 4031–4060.
B. Dubrovin. [*Geometry of 2D topological field theories*]{}. Integrable systems and quantum groups (Montecatini Terme, 1993), 120–348, Lecture Notes in Math., 1620, Fond. CIME/CIME Found. Subser., Springer, Berlin, 1996.
B. Dubrovin, Y. Zhang. [*Normal forms of hierarchies of integrable PDEs, Frobenius manifolds and Gromov–Witten invariants*]{}. A new 2005 version of arXiv:math/0108160.
E. Getzler. [*The jet-space of a Frobenius manifold and higher-genus Gromov–Witten invariants*]{}. Frobenius manifolds, 45–89, Aspects Math., E36, Friedr. Vieweg, Wiesbaden, 2004.
A. Givental. [*Semisimple Frobenius structures at higher genus*]{}. International Mathematics Research Notices 2001, no. 23, 1265–1286.
A. Givental. [*Gromov–Witten invariants and quantization of quadratic Hamiltonians*]{}. Moscow Mathematical Journal [**1**]{} (2001), no. 4, 551–568.
A. Givental. [*Symplectic geometry of Frobenius structures*]{}. Frobenius manifolds, 91–112, Aspects Math., E36, Friedr. Vieweg, Wiesbaden, 2004.
C. Hertling, Yu. Manin. [*Weak Frobenius manifolds*]{}. International Mathematics Research Notices 1999, no. 6, 277–286.
T. J. Jarvis, T. Kimura, A. Vaintrob. [*Gravitational descendants and the moduli space of higher spin curves*]{}. Advances in algebraic geometry motivated by physics (Lowell, MA, 2000), 167–177, Contemp. Math., 276, Amer. Math. Soc., Providence, RI, 2001.
Y. Konishi, S. Minabe, Y. Shiraishi. [*Almost duality for Saito structure and complex reflection groups*]{}. Journal of Integrable Systems [**3**]{} (2018), no. 1, xyy003.
M. Kontsevich, Yu. Manin. [*Gromov–Witten classes, quantum cohomology, and enumerative geometry*]{}. Communications in Mathematical Physics [**164**]{} (1994), no. 3, 525–562.
S.-Q. Liu, Y. Ruan, Y. Zhang. [*BCFG Drinfeld–Sokolov hierarchies and FJRW-Theory*]{}. Inventiones Mathematicae [**201**]{} (2015), no. 2, 711–772.
P. Lorenzoni. [*Darboux–Egorov system, bi-flat F-manifolds and Painlevé VI*]{}. International Mathematics Research Notices 2014, no. 12, 3279–3302.
P. Lorenzoni, M. Pedroni, A. Raimondo. [*F-manifolds and integrable systems of hydrodynamic type*]{}. Archivum Mathematicum [**47**]{} (2011), no. 3, 163–180.
Y. Manin. [*Frobenius manifolds, quantum cohomology, and moduli spaces*]{}. American Mathematical Society Colloquium Publications, vol. 47, American Mathematical Society, Providence, RI, 1999.
Y. Manin. [*F-manifolds with flat structure and Dubrovin’s duality*]{}. Advances in Mathematics [**198**]{} (2005), no. 1, 5–26.
R. Pandharipande, A. Pixton, D. Zvonkine. [*Relations on ${\overline{\mathcal M}}_{g,n}$ via $3$-spin structures*]{}. Journal of the American Mathematical Society [**28**]{} (2015), no. 1, 279–309.
C. Sabbah. [*Frobenius manifolds: isomonodromic deformations and infinitesimal period mappings*]{}. Expositiones Mathematicae [**16**]{} (1998), no. 1, 1–58.
C. Teleman. [*The structure of 2D semi-simple field theories*]{}. Inventiones Mathematicae [**188**]{} (2012), no. 3, 525–588.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using a Scanning Tunneling Microscope (STM), we investigate the Local Density of States (LDOS) of artificially fabricated normal metal nano-structures in contact with a superconductor. Very low temperature local spectroscopic measurements ($100\ mK$) reveal the presence of well defined subgap peaks at energy $|E| < \Delta$ in the LDOS at various positions of the STM tip. Although no clear correlations between the LDOS and the shape of the samples have emerged, some of the peak features suggest they originate from quasi-particle bound states within the normal metal structures (De Gennes S$^t$James states). Refocusing of electronic trajectories induced by the granular structure of the samples can explain the observation of spatially uncorrelated interference effects in a non-ballistic medium.'
author:
- 'W. Escoffier'
- 'C. Chapelier'
- 'F. Lefloch'
title: Ballistic effects in a proximity induced superconducting diffusive metal
---
[^1]
When a normal metal (N) is in electrical contact with a superconductor (S), it acquires superconducting properties close to its interface: a phenomenon known as the proximity effect. Although this effect was unveiled and well studied in the 1960s within the framework of the Ginsburg-Landau theory [@Bogoliubov; @DeGennes64], a revival of interest arose in the 1990s as several groups applied the techniques of mesoscopic physics to fabricate and study hybrid samples on a mesoscopic scale [@Gueron; @Esteve; @Moussy; @Vinet]. One of the key concepts relies on Andreev reflection, which provides a way to transfer Cooper pairs from the superconductor into the normal metal or reciprocally electrons from the normal metal into Cooper pairs inside the superconductor [@Andreev]. Proximity induced superconducting correlations in N in the vicinity of the S-N interface are of a different quantum nature from those of conventional superconductors. They can be regarded as correlated electron-hole pairs (Andreev pairs) propagating without any attraction potential. In diffusive metals, a microscopic description of the proximity effect emerged, based on the Usadel equation itself derived from the quasi-classical theory [@Usadel]. In particular, the Local Density of States (LDOS) of the normal part of the SN junction has been predicted to exhibit a radically different behaviour depending on the characteristic size of N, say, its width $d_n$ in a 1-D model. When the normal metal is semi-infinite ($d_n >> \xi_n$ where $\xi_n=\sqrt{\hbar D_n/\Delta}$ stands for the typical Andreev pairs coherence length, $D_n$ is the electron diffusion constant in N and $\Delta$ the superconducting gap) the LDOS displays a dip near the Fermi energy whose amplitude progressively decreases as the probing distance from the S-N contact increases. However when the normal metal is finite ($d_n < \xi_n$), a position independent mini-gap appears, whose width is of the order of the Thouless energy defined as $E_{Th}=\hbar D_n / d_n^2$ [@Belzig; @Gupta]. On the other hand, the properties of S-N bilayers are predicted to change radically from the above description if the normal metal is a ballistic medium. Andreev’s pairs propagating straight and bouncing back and forth in the normal part of the junction give rise to quantum bound states (coined De-Gennes S$^t$James states [@DeGennes63]), which yield peaks at energies $E_k < | \Delta |$ in the density of states of the system. Clearly, such peaks result from interference effects which, to our knowledge, have never been studied with a thermal resolution much lower than $\Delta$. The first observation of De-Gennes S$^t$James resonances have been reported by Tessmer [*et. al.*]{} [@Tessmer] who measured small sub-gap oscillations in the local density of states of NbSe$_2$/Au contacts with the help of a Scanning Tunneling Microscope (STM) at $1.6\ K$. Two years later, Levi [*et. al.*]{} [@Levi] observed similar effects in diffusive NbTi/Cu junctions at $4.2\ K$. Even if the temperature-limited resolution of their experiment restricted its extent, it appears that ballistic effects still can be seen in diffusive metals.
In this paper, we investigate the LDOS characteristics of artificially fabricated diffusive SNS systems of mesoscopic dimensions using a very low temperature STM. We observed sub-gap peaks of a similar nature to those reported in Levi [*et al*]{}, but energetically better resolved since the temperature was only $100\ mK$. These peaks are believed to be related to quantum bound states involving the quasi-ballistic motion of electrons in N.\
Samples consist of mesoscopic structures of normal metal (Au) embedded by a superconducting matrix (Nb). They were designed as dots, crosses, squares, triangles or hexagons with a side length of about $1\ \mu m$. Their width and thickness are $200\ nm$ and $20\ nm$ respectively (see Fig. \[figure1\] or Fig. \[figure4\] for a schematic representation of a square-shaped structure). These structures are placed in the center of a $0.5 \times 0.5\ mm^2$ lithographed test card to allow their quick location using the STM X-Y displacement stage, prior to cooling down the system. Samples were fabricated using electron beam lithography (with negative photo-resist) on an initial $20\ nm$ sputtered gold layer on $Si$. Argon etching, through the lithographed photo-resist mask, has been used to create the structures, before sputter-depositing a $20\ nm$ thick niobium film. Mechanically assisted lift-off of niobium on top of the gold structures and cleaning are the final steps of the process. This fabrication method has been developed in order to fulfill several requirements:\
(i) The interface between the superconductor and the normal metal is transparent enough with no significant barrier in order to yield an Andreev reflection coefficient close to 1 and a fully developed superconducting proximity effect [@Blonder].\
(ii) Niobium superconducting properties at the free surface of the samples are not damaged by the fabrication process and a gap in the LDOS close to its bulk value is measured on top of the film [@Hoss].\
(iii) Sample roughness and cleanliness is compatible with the STM imaging capabilities.
Although much care is taken during cleaning, some photo-resist residues remain difficult to remove. We systematically tried to avoid the dirtiest regions of the sample when positioning the STM tip and since no peculiar effects have ever been observed when tunneling close to the residues, we believe they do not affect our conclusions. The electronic mean free path of gold, extracted from resistivity measurements of films made in the same conditions as mentioned earlier, has been measured to be approximately $22\ nm$. Assuming a Fermi velocity $v_F=1.39\times10^6\ m.s^{-1}$, the diffusion constant $D_n$ is about $100\ cm^2.s^{-1}$. These films show a granular structure with a mean grain diameter of $50\ nm$.\
![\[figure1\] Room temperature images of parts of the sample: a) tilted SEM image and b) STM image. Here the niobium layer has been slightly over-deposited and the overall sample surface is not perfectly flat. Some electro-resist bits could not be removed and appear as white “bubbles” in STM images. They sometimes can alter the quality of the STM images at low temperature, preventing an accurate location of the SN interface.](fig1.ps){width="7.0cm"}
We used a home-made STM mounted in a dilution fridge, operating at a base temperature of $100\ mK$, to measure the local density of states of the normal metal structures at various positions. A small AC modulation of $20\ \mu V$ rms was added to the sample-tip DC bias voltage $V$ and the differential conductance $\frac{dI}{dV}(V)$ was obtained with a lock-in amplifier technique. Depending on experimental parameters, a typical spectrum acquisition lasts between 5 and 10 s. Noisy or unphysical spectra are automatically removed from analysis. All spectra are renormalised so that the flat density of states at energies far from $\pm \Delta$ equals 1 . When tunneling above niobium, spectra can be reproduced well using the BCS model for the density of states (see Fig. \[figure2\](b)), with $\Delta=1.17\ meV$ and an effective temperature of $T_{eff}=300\ mK$ (suggesting an electron heating effect by unfiltered electromagnetic radiation).
Now we would like to focus on spectra acquired when tunneling above the normal metal structures. A STM topographic image of one of the vertices of a triangle-like structure is shown in Fig. \[figure2\](a). A series of spectra has been acquired on 256 points spaced by $2.3\ nm$ along the horizontal dashed line. Four consecutive spectra have been automatically acquired on each spot in order to check their reproducibility. Near the middle of the structure, most of the spectra show weakened BCS-like LDOS similar to the typical niobium spectra reported above, indicating a strong proximity effect. A few of them however show small symmetric peaks with respect to the Fermi energy $E_F=0$ inside the gap (e.g. spectra of Fig. \[figure2\](d) and (e)). Although these peaks remain almost identical in the four spectra acquired consecutively at the same position, their number, energetic position and amplitude can change radically when the STM tip is slightly moved (e.g. spectra of Fig. \[figure2\](d) and (e) have been recorded only $17\ nm$ apart but show very different peak configurations). As discussed later, we suggest these peaks originate from interfering Andreev’s pairs trajectories in the normal metal. Isotropic diffusion or inelastic scattering effects affect the Andreev’s pairs coherence and explain the small amplitude of the peaks. Next to the SN boundaries, we observed non-BCS spectra, similar to spectra of Fig. \[figure2\](c) and (f), having a non-zero density of states at energies inferior, but close to $\pm \Delta$. We have no definitive interpretation for the general shape of these spectra yet. However, we suggest that such states in the LDOS involve numerous, short and poorly defined different electronic trajectories connected to the SN interface which yield a broad distribution of the aforementioned peaks below the gap edges.
![\[figure2\] Spectra have been recorded automatically along the horizontal dashed line in the center of the image, crossing one end of a triangle-like structure whose boundaries have been marked by the dotted lines. Only the most significant spectra are displayed, at the positions indicated by the arrows.](fig2a.ps "fig:"){height="3.5cm"} ![\[figure2\] Spectra have been recorded automatically along the horizontal dashed line in the center of the image, crossing one end of a triangle-like structure whose boundaries have been marked by the dotted lines. Only the most significant spectra are displayed, at the positions indicated by the arrows.](fig2b.ps "fig:"){width="8cm"}
\
In a separate experiment, we investigated the LDOS of square-shaped structures. Although STM images were rather clear at the beginning of the experiment, their quality kept decreasing on cooling down the system, preventing us from accurately correlating spectroscopic and topographic information. Nevertheless, it did not seem to affect the quality of spectroscopic measurements, which at different locations of the STM tip show alternately BCS-like density of states and spectra with very well-defined sub-gap peaks, similar to, but sharper and higher than those reported earlier in the triangle structure. On the other hand, peaks turned out to be very stealthy, as several consecutive spectral acquisitions at the same nominal position failed to display the same peak configuration (energetic position, amplitude and number). We interpret this effect as an extreme sensitivity of the peaks to the tunneling position, as discussed later. The relatively large number of spectra showing such peaks (75 selected spectra in total) allows us to summarize their characteristic features:\
(i) Only even numbers of peaks (up to six) have been observed.\
(ii) Energetic peak positions are always symmetric with respect to the Fermi energy and to a lesser extent in amplitude.\
(iii) Apart from a couple of spectra which show peaks at energies $\pm E_k$ slightly above $\Delta$, the majority of them lie within the energy gap.
To our knowledge such strong peaks have never been observed previously and we never saw them in our samples when we performed similar experiments above 1K. It is the reason why we believe they result from electronic interference effects, where long coherence lengths are required.\
Although peaks appear to be position-dependent, the observation of quasi-similar peaked spectra on both the triangle-shaped and square-shaped structures indicates that the macroscopic geometry has very little influence on the LDOS. Nevertheless, the above mentioned characteristics of the peaks support the hypothesis they result from resonant De Gennes S$^t$ James states due to the confinement of Andreev pairs inside the normal metal structure.
![\[figure3\] Peak energetic positions reported for two typical spectra (shown in inserts) acquired on one side of the square-shaped normal metal structures. For $E > 0$, three peaks are clearly visible in both spectra and have been labeled with index $n=0,1,2$. Solutions of the De Gennes S$^t$James equation have been displayed with adjusted trajectory lengths $L^{tot}_n=7525\ nm$ for spectrum 1 and $L^{tot}_n=7450\ nm$ for spectrum 2.](fig3a.ps "fig:"){width="8cm"} ![\[figure3\] Peak energetic positions reported for two typical spectra (shown in inserts) acquired on one side of the square-shaped normal metal structures. For $E > 0$, three peaks are clearly visible in both spectra and have been labeled with index $n=0,1,2$. Solutions of the De Gennes S$^t$James equation have been displayed with adjusted trajectory lengths $L^{tot}_n=7525\ nm$ for spectrum 1 and $L^{tot}_n=7450\ nm$ for spectrum 2.](fig3b.ps "fig:"){width="8cm"}
Spectra containing a large number of peaks offer the opportunity to study their relative energetic positions. Fig. \[figure3\] shows two spectra with six peaks each. Restricting our analysis to positive energies (peaks are symmetric), their labeled energetic positions $E^{exp}_k$ (n=0,1,2) have been reported in table \[table1\].
On the other hand, the De Gennes $S^t$ James quantification equation in the case of a SNS system states: $$\frac{E_k L^{tot}_n}{\hbar v_F}=k \pi + arcos\left(\frac{E_k}{\Delta}\right)$$ where $L^{tot}_n$ corresponds to the length of the interfering Andreev pairs trajectory. Using $\Delta=1.17\ meV$ and $v_F=1.39\times 10^6\ m.s^{-1}$, $L^{tot}_n$ is the only adjustable parameter that has to be tuned in order to match the set of $E_k$ with the experimental ones. A good agreement has been achieved using $L_n^{tot}=7525\ nm$ for spectrum 1 and $L_n^{tot}=7450\ nm$ for spectrum 2 (see table \[table1\]).
[|c||c||c|]{} & Spectrum 1 & Spectrum 2\
-------
Index
0
1
2
-------
: \[table1\] Theoretical and experimental energies (meV) of the peaks for spectrum 1 and 2 of Fig. \[figure3\].
&
$E_k$ $E_k^{exp}$
-------- -------------
$0.17$ $0.17$
$0.52$ $0.51$
$0.85$ $0.87$
: \[table1\] Theoretical and experimental energies (meV) of the peaks for spectrum 1 and 2 of Fig. \[figure3\].
&
$E_k$ $E_k^{exp}$
-------- -------------
$0.17$ $0.16$
$0.52$ $0.53$
$0.86$ $0.84$
: \[table1\] Theoretical and experimental energies (meV) of the peaks for spectrum 1 and 2 of Fig. \[figure3\].
\
This striking agreement raises the question of the nature of the electron-hole trajectories in the diffusive gold structure. Since the mean free path and the grain diameter are of the same order of magnitude, it is very likely that most of the scattering events occur predominantly at grain boundaries. Because these extended defects are much bigger than the Fermi wavelength, we expect specular reflection or very anisotropic diffusion to dominate at least for specific trajectories. Therefore, the absence of isotropic quantum diffraction on point-like impurities allows us to consider quasi-ballistic paths in a semi-classical picture as represented in Fig. \[figure4\]. The very existence of such a trajectory bouncing back and forth between SN boundaries with an unfolded total length $L^{tot}_n$ can explain why the observed experimental peaks are well described by the ballistic De Gennes S$^t$ James model. Moreover, the corresponding diffusion length $d^{diff}_n=\sqrt{D_nL^{tot}_n/v_F}$ in N is of the order of $230\ nm$, which is close to the nominal width $d_n=200\ nm$ of the normal metal structures.
![\[figure4\] Schematic representation of the system under study. $d_n=200\ nm$ is the nominal width of the sides of the square-shaped normal metal structure while $L^{tot}_n$ corresponds to the total length of the Andreev pair trajectory connecting the two SN interfaces.](fig4.ps){width="7cm"}
The above analysis considers a unique Andreev pair and its associated trajectory, satisfying the conditions for an interference effect to occur. According to reference [@DeGennes63], the presence of many trajectories of different length results in a peak broadening effect, associated with a non-zero sub-gap density of states. This theoretical work however considers the spatially integrated density of states of a SN slab. In our case, the STM tip constitutes a local probe and implies a selection among all the possible Andreev pairs’ trajectories. This explains why we observed very well defined peaks, instead of the sawtooth structure of the density of states initially proposed in reference [@DeGennes63]. Furthermore, according to reference [@Shytov], the discreteness of the peaks can also result from a localization effect of the Andreev pair induced by the granular structure of the samples: grains could act as lenses which re-focus diverging hole and electron trajectories of the Andreev pair and enhance interference effects. If the STM tip is in the vicinity of a localized state, it will give rise to a peak in the LDOS. As grains are randomly distributed, authors of ref. [@Shytov] predict that peaks measured by STM should be spatially uncorrelated, as indeed we observed. To a further extent, it could also explain the reported high sensitivity of the peaks’ configuration between two successive spectroscopic acquisitions at the same nominal position. Actually, between these two acquisitions, the feedback loop of the STM is activated for a short time: this can result in a very small uncontrolled shift of the STM tip position and drive the system out of resonance.\
To conclude, we report the observation of sharp sub-gap peaks in the local density of states of S-N systems at $100\ mK$ using the scanning tunneling spectroscopy technique. Their characteristics can accurately be described by De Gennes S$^t$James resonant states in the normal metal islands, although the short electronic mean free path of the samples should normally prevent their existence. If these peaks are indeed due to electronic interference effects, this would bring about new fundamental questions concerning the limit between ballistic, quasi-ballistic and diffusive systems.
[16]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
[^1]: On leave from CEA-DSM-DRFMC-SPSMS, CEA Grenoble, France
|
{
"pile_set_name": "ArXiv"
}
|
[ ]{}
Hiroshi Kawai$^*$, Osamu Narikiyo
[ ]{}
(Received )
[ **Abstract**]{}
The vibration of the dimer excited by STM current on Si(001) surface is investigated. We describe this system by the Hamiltonian which has the electron-vibration coupling term as the key ingredient. In order to characterize the transition rates induced by STM current between vibrational states we have introduced the effective temperature of the vibration which differs from the temperature of the substrate. The behavior of the effective temperature depends on the substrate temperature and STM current in highly nonlinear manner and qualitatively changes around 50K of the substrate temperature. At lower temperatures, the effective temperature strongly deviates from the substrate temperature and reaches a few hundred Kelvin for the typical values of STM current. At higher substrate temperatures, the effective temperature reduces to the substrate temperature. On the basis of these behaviors of the effective temperature, we solve the puzzle of the symmetric-asymmetric crossover in dimer images of STM observation in the ordered state of c(4 $\times$ 2).\
: Si(001), symmetric image, STM, dimer vibration
$^*$ E-mail: [email protected]
[ **1. Introduction**]{}
Intensive studies have been made on the atomic structure of the reconstructed Si(001) surface both experimentally and theoretically. On Si(001) surface, neighboring atoms form buckled dimers which have two stable tilting angles. Along the $\langle 110 \rangle$ direction, the dimer rows is formed by the dimers. The principal features of the surface reconstruction on Ge(001) surface are similar to Si(001) surface. The ground state of the structure of the dimer arrangement of both surfaces is c(4 $\times$ 2) and the ordered c(4 $\times$ 2) phase turns into a disordered state at room temperature through the order-disorder phase transition.$^{1-14)}$ In the ordered state, the dimers are observed as the asymmetric images by scanning tunneling microscopy (STM). In the disordered state, the symmetric appearing images are observed by STM, and these symmetric appearing images have been attributed to the rapid repeat of the orientational change (flip-flop motion).
In the STM observation, semiconductor substrates are doped into n-type ones or p-type ones for the electric conductivity. On Si(001) surface, serious differences in the property of the dimer system between the n-type substrates and p-type ones have not been reported at temperatures higher than 60K.
Recently, p(2 $\times$ 2) phase on Si(001) surface$^{15-17)}$ is observed on highly doped n-type substrates below 40K by STM. A phase transition is also observed by low-energy electron diffraction$^{18)}$ (LEED) at 40K on highly doped n-type substrates. On the p-type substrates, c(4 $\times$ 2) phase on Si(001) surface is observed down to 9K.$^{17)}$
At much low temperatures, the symmetric STM images of p(2 $\times$ 1) structure$^{17,19,20)}$ are observed at some conditions of the tip bias voltages $V_{\rm t}$ and the tip currents $I_{\rm t}$ in STM observations. Yokoyama$^{19)}$ reported that STM images of the dimers at 5K on boron doped (B-doped) p-type Si(001) substrate at the condition of $V_{\rm t}= \pm 1$V and $I_{\rm t}=50$pA appear to be symmetric, although most of the dimer images on the substrate are well asymmetric to make the ordered c(4 $\times$ 2) structure at 63K, and that the symmetric STM images at 5K are observed also on Antimony doped n-type Si(001) substrate. Kondo$^{19)}$ reported subsequently that the STM images at 20K on phosphorus doped n-type Si(001) substrate at the condition of $V_{\rm t}= 2.0$V and -1.6V, and $I_{\rm t}=300$pA appear to be symmetric. These observations of the symmetric images in the ordered phase are puzzling and have not been explained yet.
The first-principles calculations (FPC) with high accuracy$^{6,7,21)}$ show that the ground state on Si(001) surface is c(4 $\times$ 2). The results of FPC show that the energy of the electronic system decreases, and the energy of the lattice system increases, as the tilting angle of the dimer increases. The equilibrium angle is determined by the competition between the energies of two systems. Thermal equilibrium properties on the surface at temperatures higher than 100K have been successfully explained by Monte Carlo simulations (MCS) with Metropolis algorithm based on the results of FPC$^{6,7)}$ by mapping the two stable orientations of the dimer onto the Ising system.$^{6,7,22-25)}$ The time-resolving dynamical Monte Carlo simulations (TDMCS) performed for the dimer systems of Si(001)$^{26,27)}$ surface and Ge(001)$^{28)}$ surface show that the most of the long range fluctuations of the dimer arrangement on Si(001) and Ge(001) surfaces are carried by the phase boundary of the c(4 $\times$ 2) domain in the dimer row at low temperatures. In order to perform TDMCS on Si(001) surface and on Ge(001) surface, the model potentials for the continuous values of the tilting angle are obtained based on the FPC results$^{27,28)}$. The TDMCS results with the model potentials reproduce well the results of the time-resolving STM observations$^{12,29)}$. Since at temperatures higher than 60K, as mentioned above, serious dopant effect are not reported on Si(001) surface, the reported MCS results based on the results of FPC remain valid.
The surface localized states of $\pi$-band and $\pi^*$-band$^{6,7,17,19,21,30-32)}$ are formed mainly through the hybridization of the dangling bonds on the dimers. FPC$^{6,7,21)}$ show that the energies of the top of the $\pi$-band and the bottom of $\pi^*$-band stay a few ten meV above from the top of the valence band of Si and about 500meV below from the bottom of the conduction band of Si, respectively. The band widths are obtained to be about 1eV in both of $\pi$-band and $\pi^*$-band. These features are observed by angle-resolved UPS$^{30)}$ (ARUPS) and scanning tunneling spectroscopy $^{19,31,32)}$ (STS). On the n-type substrates, the Fermi level of the system stays near the bottom of the conduction band, which is much higher than the bottom of $\pi^*$-band. The large band bending appears near the surface region of the n-type substrate at much low temperatures. On the p-type substrates, the Fermi level stays near the top of the valence band. Therefore, the top of the $\pi$-band is near the Fermi level. The large band bending does not appear near the surface region on the p-type substrates. Because the dopant effect on the substrate is not taken into account in the FPC, the results of FPC reproduce well the surface properties of p-type substrate, but some modifications are necessary for n-type substrate at much low temperatures.
In the present study, we take the scope to the vibration excited by STM current on the p-type substrate. We will introduce the Hamiltonian which has the electron-vibration coupling term. The excitation rates by the STM currents for the vibration will be obtained. The rates will be characterized by introducing the effective temperature of the vibration. The mechanism for the symmetric-asymmetric crossover in dimer images of STM observation on Si(001) surface will be presented.\
[ **2. Model**]{}
We introduce the Hamiltonian $H$ representing the coupling between the electronic system and the localized vibrational system,
$$\begin{aligned}
H &=& H_{\rm e} + H_{\rm e-v}, \nonumber \\
H_{\rm e} &=& \sum_k \varepsilon_k c_k^\dagger c_k
+ \sum_p \varepsilon_p c_p^\dagger c_p
+ \Gamma \left\{(\sum_k \gamma_k c_k^\dagger)(\sum_p \gamma_p^* c_p) + {\rm H.c.}
\right\} \nonumber \\
&=& \sum_\alpha \varepsilon_\alpha c_\alpha^\dagger c_\alpha
+ \varepsilon_a a^\dagger a
+ (\sum_\alpha \Gamma_\alpha a^\dagger c_\alpha + {\rm H.c.})
+ \sum_p \varepsilon_p c_p^\dagger c_p
+ (\sum_p \Gamma_p a^\dagger c_p + {\rm H.c.}) \nonumber, \\
H_{\rm e-v}&=& \hbar \omega (b^\dagger b + \frac{1}{2})
+ \delta \varepsilon (b^\dagger +b) a^\dagger a, \nonumber \\
\varepsilon_a &=& \sum_k |\gamma_k|^2 \varepsilon_k, \nonumber \\
a^\dagger &=& \sum_k \gamma_k c_k^\dagger, \nonumber \\
\Gamma_p &=& \Gamma \gamma_p,\end{aligned}$$
where $c_k$, $c_p$ and $b$ are the annihilation operators for the electronic states in the $\pi$-band, for the states in the conduction band of the STM tip, and for the vibrational state in the rocking mode of the dimer on Si(001) surface beneath the STM tip. We assume that the vibrational states are the harmonic oscillator, and the energy of the state with the vibrational number $n$ is $\hbar \omega (n+1/2)$. The energy of the vibrational state $\hbar \omega$ is measure by the electron energy loss spectroscopy$^{33)}$ as 20meV. $a^\dagger = \sum_k \gamma_k c_k^\dagger$ and $\sum_p \gamma_p c_p^\dagger$ are the creation operator for the spatially localized states $|a \rangle$ in the surface dimer and in the prominence of the tip, respectively. The tunneling current is through the the spatial localized states. The coefficients of the spatial localized states are normalized as $\sum_k |\gamma_k|^2=1$ and $\sum_p |\gamma_p|^2=1$. The diagonalized term of $\sum_k \varepsilon_k c_k^\dagger c_k$ is rewritten by $a$ and $c_\alpha$ as the first three terms in the third line of eq. (1), where $c_\alpha$ is the annihilation operators for the electronic states orthogonalized with $|a \rangle$. The amplitude of the vibration modifies the one electron energy of $|a \rangle$, and $H$ has the electron-vibration coupling term, because the energy of the electronic system, as mentioned above, decreases as the tilting angle of the dimer increases. The term of $\delta \varepsilon$ represents the electron-vibration coupling. $\delta \varepsilon$ is the coupling constant between the amplitude of the vibration and the energy of the one electron state of $|a \rangle$. The vibration is excited or deexcited through the electron-vibration coupling by the STM current. The similar Hamiltonian of eq. (1) has been used in the precedent studies of vibrations of molecules adsorbed on metal surfaces$^{34-41)}$. In the present study, however, the Hamiltonian is applied to totally different system, the surface localized semiconductor state on Si(001) surface.
In the present study, we treat the case that the tip bias voltages $V_{\rm t}$ is positive and so large that the Fermi level in the conduction band of the STM tip is below the bottom of the $\pi$-band of the p-type Si(001) substrate; $V_{\rm t}$ is assumed to be slightly larger than 1V, because the band width of the $\pi$-band is about 1eV. The elastic inter-band transition of electrons from the $\pi$-band to the conduction band of the tip is induced by the presence of the tip. The elastic inter-band transition rate $\sigma_{\rm elas}$ of electrons from the surface to the tip without coupling to the vibration$^{34-41)}$ is given by the T-matrix element and the Fermi distribution functions:
$$\begin{aligned}
\sigma_{\rm elas} &=& 2\frac{2 \pi}{\hbar}\sum_{\alpha, p}
|\Gamma_p^* G_a(\varepsilon) \Gamma_\alpha|^2 \nonumber \\
&& \quad {} \times
(\frac{1}{\exp(\beta(\varepsilon_\alpha-\varepsilon_{\rm F}))+1})
(1-\frac{1}{\exp(\beta(\varepsilon_p+eV_{\rm t}-\varepsilon_{\rm F}))+1})
\delta(\varepsilon_\alpha-\varepsilon_p) \nonumber \\
&=& \frac{4}{\pi \hbar}
\int_{-D_\pi+E_\pi}^{E_\pi} {\rm d} \varepsilon \>
\Delta_{\rm t} (\varepsilon)
\Delta_{\rm s} (\varepsilon)
|G_a(\varepsilon)|^2 \nonumber \\
&& \quad {} \times
(\frac{1}{\exp(\beta(\varepsilon_\alpha-\varepsilon_{\rm F}))+1})
(\frac{1}{1+\exp(-\beta(\varepsilon_p+eV_{\rm t}-\varepsilon_{\rm F}))})
\nonumber \\
&\approx& \frac{4}{\pi \hbar} \Delta_{\rm t} (\varepsilon_a) \Delta_{\rm s} (\varepsilon_a)
|G_a (\varepsilon_a)|^2 D_\pi, \nonumber \\
G_a (\varepsilon)&=& \langle a | (\varepsilon + i 0^+ - H_{\rm e})^{-1} | a \rangle
\nonumber \\
&=& \frac{1}{\varepsilon -\varepsilon_a - \Lambda(\varepsilon) + i \Delta (\varepsilon)},
\nonumber \\
\Delta (\varepsilon) &=& \Delta_{\rm s} (\varepsilon) + \Delta_{\rm t} (\varepsilon),
\nonumber \\
\Delta_{\rm s} (\varepsilon) &=& \pi \sum_\alpha |\Gamma_\alpha|^2
\delta(\varepsilon-\varepsilon_\alpha),
\nonumber \\
\Delta_{\rm t} (\varepsilon) &=& \pi \sum_p |\Gamma_p|^2
\delta(\varepsilon-\varepsilon_p),
\nonumber \\
\Lambda (\varepsilon) &=& \frac{1}{\pi} P
\int_{-\infty}^\infty {\rm d} \varepsilon'
\frac{\Delta (\varepsilon')}{\varepsilon-\varepsilon'},\end{aligned}$$
where $\varepsilon_{\rm F}$ is the Fermi level of the electronic system, $G_a (\varepsilon)$ is the Green function for the electron system, $\Delta (\varepsilon)$ is the width of the projected density of states for $|a \rangle$, $\Delta_{\rm s} (\varepsilon)$ and $\Delta_{\rm t} (\varepsilon)$ are the components of $\Delta (\varepsilon)$ in the surface and in the tip, respectively, $P$ denotes the Cauchy principal value, and $E_\pi$ and $D_\pi$ are the energy level of the top of the $\pi$-band and the width of the $\pi$-band, respectively. In the present study, $-(E_\pi - \varepsilon_{\rm F})$ is assumed to be of the order of a few ten meV and $D_\pi$ is assumed to be large enough $\beta D_\pi \gg 1$. The temperature of the system is assumed to be so low that the broadness of the Fermi distribution around $\varepsilon_{\rm F}$ in the surface state and around $\varepsilon_{\rm F}-eV_{\rm t}$ in the tip state are not contributed to the evaluation of the integration in eq.(2). We assume that $\Delta_{\rm t} (\varepsilon_a) / \Delta_{\rm s} (\varepsilon_a)$ is small in the STM observation in the typical conditions enough to be approximated as $\Delta (\varepsilon_a) \approx \Delta_{\rm s} (\varepsilon_a)$. Because in the STS experiment at 5.5K, the tunneling spectra of the $\pi$-band are observed to be almost symmetric band, we assume that $\varepsilon_a$ stay at the center of the $\pi$-band in energy, and $\Lambda (\varepsilon_a)=0$. Therefore, $\sigma_{\rm elas}$ is approximately obtained as
$$\begin{aligned}
\sigma_{\rm elas} &\approx&
\frac{4}{\pi \hbar}
\Delta_{\rm t} (\varepsilon_a) \Delta_{\rm s}^{-1} (\varepsilon_a) D_\pi .\end{aligned}$$
The inter-band transition of electrons from the $\pi$-band to the conduction band of the tip induced by the presence of the tip can excite the vibrational state in the dimer through the electron-vibration coupling. The inelastic inter-band transition rate $\sigma_{0 \to 1}^{\rm int}$ of electrons$^{34-41)}$ from the surface to the tip coupling with the excitation from the ground state of the vibrational number $n=0$ to the excited state of $n=1$ is obtained in the lowest order in $|\delta \varepsilon|$ as
$$\begin{aligned}
\sigma_{0 \to 1}^{\rm int} &=& 2 \frac{2\pi}{\hbar} \sum_{\alpha', p'}
| \langle p', 1 |H_{\rm e-v}|\alpha', 0 \rangle |^2
\delta(E(p') - \hbar \omega - E(\alpha')) \nonumber \\
&=& \frac{4\pi |\delta \varepsilon|^2}{\hbar}
\int_{-D\pi+E_\pi}^{E_\pi} {\rm d} \varepsilon \>
\rho_{\rm s}(\varepsilon) \rho_{\rm t}(\varepsilon) \nonumber \\
&& \quad {} \times
(\frac{1}{\exp(\beta(\varepsilon_\alpha-\varepsilon_{\rm F}))+1})
(1-\frac{1}{\exp(\beta(\varepsilon_p - \hbar \omega +eV_{\rm t}-\varepsilon_{\rm F}))+1}) \nonumber \\
&\approx& \frac{4\pi |\delta \varepsilon|^2}{\hbar}
\rho_{\rm s}(\varepsilon_a) \rho_{\rm t}(\varepsilon_a)D_\pi, \nonumber \\
\rho_{\rm s}(\varepsilon) &=& \sum_{\alpha'} |\langle \alpha' | a \rangle|^2
\delta(\varepsilon - \varepsilon_{\alpha'}) \nonumber \\
&=& \frac{1}{\pi}
(\frac{\Delta_{\rm s}(\varepsilon)}{(\varepsilon-\varepsilon_a-\Lambda(\varepsilon))^2
+\Delta^2(\varepsilon))}), \nonumber \\
\rho_{\rm t}(\varepsilon) &=& \sum_{p'} |\langle p' | a \rangle|^2
\delta(\varepsilon - \varepsilon_{p'}) \nonumber \\
&=& \frac{1}{\pi}
(\frac{\Delta_{\rm t}(\varepsilon)}{(\varepsilon-\varepsilon_a-\Lambda(\varepsilon))^2
+\Delta^2(\varepsilon))}), \end{aligned}$$
where $E(\alpha')$ and $E(p')$ are the one electron energies of $| \alpha' \rangle$ and $| p' \rangle$, respectively, and $\rho_{\rm s}$ and $\rho_{\rm t}$ are the projected density of states for $| a \rangle$. $| \alpha' \rangle$ and $| p' \rangle$ are the stationary states of $H_{\rm e}$ connected to $| \alpha \rangle$ and $| p \rangle$, respectively. Namely, the stationary states are given as $|\alpha' \rangle = | \alpha \rangle
+
(\varepsilon_\alpha +i 0^+ - H_{\rm e})^{-1} \Gamma_\alpha | a \rangle $ and $| p' \rangle = | p \rangle
+ (\varepsilon_p +i 0^+ - H_{\rm e})^{-1} \Gamma_p | a \rangle$. The projected density of states are approximated in the same way as eq. (3): $\rho_{\rm s}(\varepsilon_a) \approx
(\pi \Delta_{\rm s}(\varepsilon_a))^{-1}$ and $\rho_{\rm t}(\varepsilon_a) \approx \Delta_{\rm t}(\varepsilon_a)
(\pi \Delta_{\rm s}^2 (\varepsilon_a))^{-1}$. The inelastic transition rate $\sigma_{0 \to 1}^{\rm int}$ is given by use of the approximation for the projected density of states as
$$\begin{aligned}
\sigma_{0 \to 1}^{\rm int} &\approx&
\frac{4 |\delta \varepsilon|^2}{\pi \hbar}
\Delta_{\rm t} (\varepsilon_a) \Delta_{\rm s}^{-3} (\varepsilon_a) D_\pi .\end{aligned}$$
The elastic transition rate $\sigma_{\rm elas}$ and the inelastic transition rate $\sigma_{0 \to 1}^{\rm int}$ depend critically on $I_{\rm t}$ and do not on $T$, because in eqs. (3) and (5), $\Delta_{\rm t}$ depends critically on the spatial tunneling gap of STM. The fraction of the inelastic transition rate to the elastic transition rate$^{34-41)}$ $W=\sigma_{0 \to 1}^{\rm int}/\sigma_{\rm elas}$, however, does not depend on $\Delta_{\rm t}$, and is obtained from eqs. (3) and (5) as
$$\begin{aligned}
W &=&
(\frac{|\delta \varepsilon |}{\Delta_{\rm s}(\varepsilon_a)})^2 .\end{aligned}$$
Eq. (6) shows that the fraction of $\sigma_{0 \to 1}^{\rm int}/\sigma_{\rm elas}$ is given dominantly from the intrinsic property of Si(001) surface, and independent of the STM current. In the present study, as mentioned already, $V_{\rm t}$ is assumed to be so large that the Fermi level in the conduction band of the tip is below the bottom of the $\pi$-band, and $\hbar \omega \ll eV_{\rm t}$. From these assumption, the inelastic inter-band transition rate $\sigma_{1 \to 0}^{\rm int}$ of electrons from the surface to the tip coupling with the deexcitation from the vibrational number $n=1$ to $n=0$ is well approximated to take the same value as $\sigma_{0 \to 1}^{\rm int}$. Thus the inelastic inter-band transition rate $\sigma_{n \to n \pm 1}^{\rm int}$ of electron from the surface to the tip coupling with the vibrational transition from $n$ to $n \pm 1$ is given as $\sigma_{n \to n \pm 1}^{\rm int} = {\rm max}(n,n \pm 1)
\sigma_{0 \to 1}^{\rm int}$.
At finite temperatures $T$, the deexcitation of the vibrational number $n=1$ to $n=0$ through the inner-band excitation of the electron-hole pair creation in the $\pi$-band can occur. In the inner-band excitation, both the initial state and the final state of the electron are within the states of $| \alpha' \rangle$, which have dominant component in the $\pi$-band on the surface. The deexcitation of the vibration with the inner-band excitation does not occur at 0K. The rate of the deexcitaion $\sigma_{1 \to 0}^{\rm inn}$ of vibration through the inner excitation is independent of $I_{\rm t}$ and depends on $T$:
$$\begin{aligned}
\sigma_{1 \to 0}^{\rm inn} &=& 2\frac{2\pi}{\hbar}
\sum_{\alpha'_1, \alpha'_2} |\langle \alpha'_2,0 |H_{\rm e-v} | \alpha'_1,1 \rangle|^2,
\delta(E(\alpha'_2)-\hbar \omega -E(\alpha'_1)) \nonumber \\
&=& \frac{4 |\delta \varepsilon|^2}{\pi \hbar}
\int_{-D_\pi + E_\pi}^{E_\pi-\hbar \omega} {\rm d} \varepsilon \>
\rho_{\rm s}(\varepsilon)\rho_{\rm s}(\varepsilon + \hbar \omega) \nonumber \\
&& {} \quad \times
(\frac{1}{\exp(\beta(\varepsilon-\varepsilon_{\rm F}))+1})
(1-\frac{1}{\exp(\beta(\varepsilon + \hbar \omega -\varepsilon_{\rm F}))+1}) \nonumber \\
&\approx& 4\pi |\delta \varepsilon|^2 \rho_{\rm s}^2 (\varepsilon_a)
(\frac{F(T)}{1-\exp(-\beta \hbar \omega)}) \nonumber \\
&\approx& \frac{4}{\pi} W
(\frac{F(T)}{1-\exp(-\beta \hbar \omega)}), \nonumber \\
F(T) &=& \frac{1}{\beta \hbar}
\log(\frac{1+\exp(\beta (E_\pi-\varepsilon_{\rm F}))}
{1+\exp(\beta (E_\pi-\hbar \omega -\varepsilon_{\rm F}))}).\end{aligned}$$
The transition processes of $\sigma_{\rm elas}$, $\sigma_{1 \to 0}^{\rm int}$, $\sigma_{0 \to 1}^{\rm int}$, $\sigma_{1 \to 0}^{\rm inn}$, and $\sigma_{0 \to 1}^{\rm inn}$ are schematically shown in Fig. 1.
The essential framework of the formalism for $\sigma_{\rm elas}$, $\sigma_{0 \to 1}^{\rm int}$, and $\sigma_{1 \to 0}^{\rm int}$ is similar to that of theoretical studies$^{34-41)}$ on the STM induced vibration of molecules adsorbed on metal surfaces. In the problems of molecules adsorbed on metal surfaces, transitions corresponding to the inner-band transitions in the present study have also been treated. The transitions corresponding to the inner-band transitions, however, do not exhibit any temperature dependence in the case of metal surfaces. Eq. (7) shows that $\sigma_{1 \to 0}^{\rm inn}$ depend strongly on $T$; $\pi \beta \hbar \sigma_{1 \to 0}^{\rm inn} \approx
4W\exp(\beta(E_\pi - \varepsilon_{\rm F}))$ for $-\beta(E_\pi - \varepsilon_{\rm F}) \gg 1$ and $\pi \beta \hbar \sigma_{1 \to 0}^{\rm inn} \approx 2W$ for $-\beta(E_\pi -\hbar \omega - \varepsilon_{\rm F}) \ll 1$. The rate of the excitation $\sigma_{0 \to 1}^{\rm inn}$ of vibration through the inner-band deexcitation is easily obtained by the essentially same way as $\sigma_{1 \to 0}^{\rm inn}$: $\sigma_{0 \to 1}^{\rm inn} \approx (4/\pi) W [F(T)/(\exp(\beta \hbar \omega)-1)]$.
The total transition rate of the excitation $R_{0 \to 1}$ of the vibrational number $n=0$ to 1 and the deexcitation $R_{1 \to 0}$ of the vibrational number $n=1$ to 0 are given by the sum of rates as
$$\begin{aligned}
R_{0 \to 1} &=& (\frac{\tau_0^{-1}}{\exp(\beta \hbar \omega)-1})
+\sigma_{0 \to 1}^{\rm inn}+\sigma_{0 \to 1}^{\rm int} \nonumber \\
&=& (\frac{1}{\exp(\beta \hbar \omega)-1})
(\tau_0^{-1}+\frac{4}{\pi} W F(T)) + W \sigma_{\rm elas}, \nonumber \\
R_{1 \to 0} &=& (\frac{\tau_0^{-1}}{1-\exp(-\beta \hbar \omega)})
+\sigma_{1 \to 0}^{\rm inn} +\sigma_{1 \to 0}^{\rm int} \nonumber \\
&=& (\frac{1}{1-\exp(-\beta \hbar \omega)})
(\tau_0^{-1}+\frac{4}{\pi} W F(T)) + W \sigma_{\rm elas},\end{aligned}$$
where $\tau_0$ is the life time of the vibration through the background phonon system at 0K.
The probability $P_n$ to find the vibrational state at $n$ satisfies the rate equation represented by $R_{0 \to 1}$ and $R_{1 \to 0}$:
$$\begin{aligned}
\frac{{\rm d} P_n}{{\rm d}t} &=&
(n+1)R_{1 \to 0}P_{n+1}+nR_{0 \to 1}P_{n-1}
-(nR_{1 \to 0}+(n+1)R_{0 \to 1})P_n, \quad 1 \le n, \nonumber \\
\frac{{\rm d} P_0}{{\rm d}t} &=& R_{1 \to 0}P_1-R_{0 \to 1}P_0.\end{aligned}$$
In the stationary distribution of $P_n$, $P_n$ is derived as $P_n=r^n (1-r)$ from eq. (9), where $r=R_{0 \to 1}/R_{1 \to 0}$.
The tip current $I_{\rm t}$ are given by the sum of the elastic inter-band transitions and the inelastic inter-band transitions:
$$\begin{aligned}
\frac{I_{\rm t}}{e}&=&
\sum_{n=0}^\infty P_n \sigma_{\rm elas}
+\sum_{n=0}^\infty P_n \sigma_{n \to n+1}^{\rm int}
+\sum_{n=1}^\infty P_n \sigma_{n \to n-1}^{\rm int} \nonumber \\
&=& [1+W \frac{1+r}{1-r}] \sigma_{\rm elas}.\end{aligned}$$
When $T$, $W$, and $I_{\rm t}$ are given, $R_{0 \to 1}$ and $R_{1 \to 0}$ are obtained selfconsistently with eqs. (8) and (10). The distribution of the dimer vibration in the STM observation is characterized by the effective temperature $T_{\rm ef}$ of the dimer vibration defined as
$$\begin{aligned}
r=\exp(-\hbar \omega / (k_{\rm B} T_{\rm ef})).\end{aligned}$$
Throughout this paper, $T$ denotes the temperature of the substrate. $T_{\rm ef}$ differs from $T$ in general and depends on $T$ and $I_{\rm t}$ nonlinearly. The key quantity for $T_{\rm ef}$ is the fraction $W=\sigma_{0 \to 1}^{\rm int}/\sigma_{\rm elas}$, which is derived from $|\delta \varepsilon|$ and $\Delta_{\rm s}$ as shown in eq. (6). The coupling constant $\delta \varepsilon$ between the amplitude of the vibration and the energy of $| a \rangle$, is approximately estimated as
$$\begin{aligned}
|\delta \varepsilon| &=& \Delta \theta \frac{\partial \varepsilon_a}{\partial \theta}_{\theta=\theta_0},\end{aligned}$$
where $\theta_0$ (= $19.1^\circ = 0.333$rad ) is the equilibrium tilting angle of dimer in the c(4 $\times$ 2) structure, and $\Delta \theta$ is the mean value of the amplitude of the vibrational number $n=1$. $\Delta \theta$ is obtained by $\omega$ and the effective value of the moment of inertia $I$ as $\Delta \theta = (3\hbar/(2 I \omega))^{1/2}$. As mentioned already, the model potentials for the continuous values of the tilting angles of the dimers on Si(001) surface are obtained based on the FPC results. We can get approximate values of $I$ and $\partial \varepsilon_a / \partial \theta$ at $\theta=\theta_0$ by the model potentials:
$$\begin{aligned}
\frac{\partial \varepsilon_a}{\partial \theta}_{\theta=\theta_0} &=& - k \theta_0,
\nonumber \\
I &=& \frac{k}{\omega^2}, \end{aligned}$$
where $k$ is the value of the second partial derivative of the model potentials for the tilting angle of the dimer beneath the tip at $\theta=\theta_0$. We derive the value of $k$ from the model potentials, as $k=1.1 \times 10^4{\rm meV}/{\rm rad}^2$. Using eq. (13), we get approximate values as $I=1.2 \times 10^{-23}{\rm meVs}^2/{\rm rad}^2$, $\partial \varepsilon_a / \partial \theta =
-3.7 \times 10^3{\rm meV}/{\rm rad}$, and $\Delta \theta = 5.2 \times 10^{-2}{\rm rad}$. From eq. (12), $|\delta \varepsilon|$ is approximately estimated to be $1.9 \times 10^2{\rm meV}$. From the half width of half maximum of STS for $\pi$-band of Si(001) at 5.5K, $\Delta_{\rm s}$ is roughly estimated to be $2.5 \times 10^2$meV. From these values, the key quantity $W$ is obtained to be about 0.6.\
[ **3. Results and Discussions**]{}
In the present study, we take the scope on the p-type substrate. The Fermi level of the p-type substrate is well approximated to be at the center of the top of the valence band of the substrate and the acceptor level of dopant at low temperatures. At temperatures low but higher than a few Kelvin, $E_\pi$ is practically fixed at the top of the valence band. In the present study, $E_\pi - \varepsilon_{\rm F}$ is approximated to be $-22.5$meV for B-doped substrate and $-32.5$meV for Ga-doped substrate, respectively. The life time of the vibration $\tau_0$ through the background phonon system at 0K is expected to be of the order of $10^{-9}$s for semiconductors. We assumed for $\tau_0$ to be 8ns.
The excitation rate $R_{0 \to 1}$ and the deexcitation rate $R_{1 \to 0}$ for B-doped substrate are shown in Fig. 2.
{width="16cm"}
Both rates are monotonically increasing function of $T$ and $I_{\rm t}$ in highly nonlinear manner. The rates scarcely depend on $T$ for $T<20$K, because the inter-band transitions are dominant in them and the inner-band transitions scarcely contribute to them. When $T$ increases around 20K, the excitation and deexcitaion of the vibration through the inner-band transitions of electrons become effective and the rates increase steeply for $T$. The deexcitation rate $R_{1 \to 0}$ increases much steeper for $T$ than $R_{0 \to 1}$ around 20K $\sim$ 40K. The rates scarcely depend on $I_{\rm t}$ for $T>70$K, because the inner-band transitions are dominant in them and the inter-band transitions scarcely contribute to them.
Since the magnitudes of the rates are smaller enough than the frequency of the vibration 4.8THz derived from $\hbar \omega =
20$meV, the effective temperature of the dimer vibration $T_{\rm ef}=-\hbar \omega / (k_{\rm B} \log(R_{0 \to 1}/R_{1 \to 0}))$ well characterizes the distribution of the vibration. $T_{\rm ef}$ for B-doped substrate are shown in Fig. 3.
{width="8.5cm"}
At low temperatures of $T<20$K, $T_{\rm ef}$ scarcely depending on $T$, increases steeply with $I_{\rm t}$, becomes about 250K at $I_{\rm t}=50$pA, and reaches about 500K at $I_{\rm t}=200$pA. The rapid flip-flop is expected in such high $T_{\rm ef}$. These high $T_{\rm ef}$ appearing at low temperatures of $T<20$K, essentially explains why the symmetric STM images are observed at low temperatures. When $T$ increases around 20K, $T_{\rm ef}$ starts to decrease steeply, because of steep increase of $R_{1 \to 0}$. $T_{\rm ef}$ becomes 70K around $T \approx 40$K and $I_{\rm t} \approx 150$pA. At higher temperatures $T>60$K, $T_{\rm ef}$ increases gradually with $T$, scarcely depends on $I_{\rm t}$ within the typical values of $I_{\rm t}$, and takes the nearly same value as $T$. These low $T_{\rm ef}$ explains why the asymmetric STM images recover in temperature range higher than 50K.
$T_{\rm ef}$ for Ga-doped substrate are shown in Fig. 4. The essential feature of $T_{\rm ef}$ shown in Fig. 4 is same as in Fig. 3. The steep decrease of $T_{\rm ef}$ in Fig. 4, however, is starting around $T \approx 30$K which is higher than for B-doped substrate shown in Fig. 3. The temperature where the inner-band transitions exceed the inter-band transitions in magnitude on Ga-doped substrate, is higher than that on B-doped substrate and the ratio of these two temperatures is almost the same value as the ratio of $\varepsilon_{\rm F}-E_\pi$, 32.5/22.5=1.44, since $\sigma_{1 \to 0}^{\rm inn}$, as mentioned already, is approximated as $\sigma_{1 \to 0}^{\rm inn} \approx
4W \exp(\beta(E_\pi-\varepsilon_{\rm F}))/(\pi \beta \hbar)$ for $-\beta(E_\pi-\varepsilon_{\rm F}) \gg 1$.
{width="8.5cm"}
At low temperature $T<20$K, the atom at the up-position beneath the tip is rapidly transferred to the down-position through the dimer flip-flop motion by quasi-thermal excitation of the vibration of high $T_{\rm ef}$. After the flip-flop motion to the down-position, $T_{\rm ef}$ becomes much lower than at the up-position, because tip current is expected to be much smaller at down-position than at up-position. Because of the low effective temperature $T_{\rm ef}$, the transition rate of the atom from the down-position to the up-position is much smaller than that of the opposite transition; the fraction of the staying time at the down-position is much larger than that at the up-position. The spatial STM gap between the prominence of the tip and the surface dimer is tuned for $I_{\rm t}$ to be the STM condition. The tip current $I_{\rm t}$ is time averaged one, and is tuned almost at the down-positon at low temperatures. In the scanning of STM tip at much low temperatures, the atoms beneath the tip are almost always tied to the down-positions, and the atoms at the up-position scarcely contribute to the STM images. As the result of the large fraction at the down-position, the symmetric dimer images of the p(2 $\times$ 1) structure is observed even in the ordered state of c(4 $\times$ 2). These feature that the tip ties the atom to the down-positions in STM observation on Si(001) surface corresponds to the experimental results reported by Mitsui and Takayanagi$^{42)}$.
In the present study, we demonstrate that the inner-band transitions play the important role to decrease $T_{\rm ef}$ at the temperatures of a few ten Kelvin where deexcitation of the vibration through the inner-band excitations of the electron-hole creation occur. This deexcitation through the inner-band excitation can occur also when enough amount of holes in the $\pi$-band or enough amount of electrons in the $\pi^*$-band is induced. For example, when the electrons and holes are induced by infrared photons or by other methods, $T_{\rm ef}$ will be expected to decrease in much low temperature in STM observation. Here we propose the notion of “the excitation-induced cooling”, namely, the decrease of $T_{\rm ef}$ essentially induced by the electronic excitation in $\pi$-band or $\pi^*$-band.\
[ **4. Conclusion**]{}
In order to explain the symmetric-asymmetric crossover in dimer image of STM observation on Si(001) surface, we calculate the effective temperature $T_{\rm ef}$ of dimer vibration. The crossover behavior is semiquantitatively understood by the nonlinear dependence of $T_{\rm ef}$ on the substrate temperature and STM current. The symmetric image in the ordered state of c(4 $\times$ 2) results from the large fraction of the atom at the down-position induced by STM current. We have proposed the notion of “the excitation-induced cooling” under more general situations than that induced by the substrate temperature.\
[ **Acknowledgment**]{}
The author would like to thank K. Makoshi, H. Ueba, and T. Mii for valuable discussions. This work was supported in part by Grant-in-Aid for Scientific Research (C) from Japan Society for the Promotion of Science.
[ **References**]{}\
------- ----------------------------------------------------------------------------------------------
$1)$ R. E. Schlier and H. E. Farnsworth: J. Chem. Phys. [30]{} (1959) 917.
$2)$ J. J. Lander and J. Morrison: J. Chem. Phys. [37]{} (1959) 729.
$3)$ T. Tabata, T. Aruga, and Y. Murata: Surface Science [**179**]{} (1987) L63.
$4)$ R. J. Hamers, R.M. Tromp, and J. E. Demuth: Phys. Rev. B [**34**]{} (1986) 5343.
$5)$ R. A. Wolkow: Phys. Rev. Lett. [**68**]{} (1992) 2636.
$6)$ K. Inoue, Y. Morikawa, K. Terakura, and M. Nakayama:
Phys. Rev. B [**49**]{} (1994) 14774.
$7)$ K. Terakura, T. Yamasaki, and Y. Morikawa: Phase Transition [**53**]{} (1995) 143.
$8)$ S. D. Kevan and N. G. Stoffel: Phys. Rev. Lett. [**53**]{} (1984) 702.
$9)$ S. D. Kevan: Phys. Rev. B [**32**]{} (1985) 2344.
$10)$ S. Ferrer, X. Torrelles, V. H. Etgens, H. A. van der Vegt, and P. Fajardo:
Phys. Rev. Lett. [**75**]{} (1995) 1771.
$11)$ C. A. Lucas, C. S. Dower, D. F. McMorrow, G. C. L. Wong, F. J. Lamelas, and
P. H. Fuoss: Physi. Rev. B [**47**]{} (1993) 10375.
$12)$ T. Sato, M. Iwatsuki, and H. Tochihara: J. Electron Microsc. [**48**]{} (1999) 1.
$13)$ Y. Yoshimoto, Y. Nakamura, H. Kawai, M. Tsukada, and M. Nakayama:
Phys. Rev. B [**61**]{} (2000) 1965.
$14)$ Y. Yoshimoto, Y. Nakamura, H. Kawai, M. Tsukada, and M. Nakayama:
Surf. Rev. and Lett. [**6**]{} (1999) 1045.
$15)$ H. Shigekawa, K. Miyake, M. Ishida, K. Hata, H. Oigawa, Y. Nannichi,
R. Yoshizaki, A. Kawazu, T. Abe, T. Ozawa, and T. Nagamura:
Jpn. J. Appl. Phys. [**35**]{} (1996) L1081.
$16)$ S. Yoshida, O. Takeuchi, K. Hata, R. Morita, M. Yamashita, and H. Shigekawa:
Jpn. J. Appl. Phys. [**41**]{} (2002) 5017.
$17)$ K. Hata, S. Yoshida, and H. Shigekawa: Phys. Rev. Lett. [**89**]{} (2002) 286104.
$18)$ M. Matsumoto, K. Fukutani, and T. Okano: Phys. Rev. Lett. [**90**]{} (2003) 106103.
$19)$ T. Yokoyama and K. Takayanagi: Phys. Rev. B [**61**]{} (2000) R5078.
$20)$ K. Kondo, T. Amakusa, M. Iwatsuki, and H. Tokumoto: Surface Science [**453**]{} (2001) L318.
$21)$ A. Ramstad, G. Brocks, and P. J. Kelly: Phys. Rev. B [**51**]{} (1994) 14504.
$22)$ H. Tochihara, Y. Nakamura, H. Kawai, M. Nakayama, T. Sato, T. Sueyoshi,
T. Amakusa, and M. Iwatsuki: J. Phys. Soc. Jpn. [**67**]{} (1998) 2330.
$23)$ Y. Nakamura, H. Kawai, and M. Nakayama: Phys. Rev. B [**52**]{}52 (1995) 8231.
------- ----------------------------------------------------------------------------------------------
------- -------------------------------------------------------------------------------------------------
$24)$ Y. Nakamura, H. Kawai, and M. Nakayama: Surface Science [**357-358**]{} (1996) 500.
$25)$ Y. Nakamura, H. Kawai, and M. Nakayama: Phys. Rev. B [**55**]{} (1997) 10549.
$26)$ H. Kawai, Y. Nakamura, and M. Nakayama: J. Phys. Soc. Jpn. [**68**]{} (1999) 3936.
$27)$ H. Kawai, R. Miyata, Y. Yoshimoto, and M. Tsukada: submitted to J. Phy. Soci. Japn.
$28)$ H. Kawai, Y. Yoshimoto, H. Shima, Y. Nakamura, and Masaru Tsukada:
J. Phys. Soc. Jpn [**71**]{} (2002) 2192.
$29)$ K. Hata, Y. Saino, and H. Shigekawa: Phys. Rev. Lett. [**86**]{} (2001) 3084.
$30)$ Y. Enta, S. Suzuki, and S. Kono: Phys. Rev. Lett. [**65**]{} (1990) 2704.
$31)$ T.Yokoyama, M. Okamoto, and K. Takayanagi: Phys. Rev. Lett. [**81**]{} (1998) 3423.
$32)$ K. Hata, Y. Shibata, and H. Shigekawa: Phys. Rev. B [**64**]{} (2001) 235310.
$33)$ N. Takagi, S. Shimonaka, T. Aruga, and N. Nishijima: Phys. Rev. B [**60**]{} (1999)
10919.
$34)$ B. N. J. Persson and M. Persson: Solid State Comm. [**36**]{} (1980) 175.
$35)$ B. N. J. Persson and A. Baratoff: Phys. Rev B [**59**]{} (1987) 339.
$36)$ S. Gao, M. Persson, and B. I. Lundqvist: Phys. Rev. B [**55**]{} (1997) 4825.
$37)$ N. Mingo and K. Makoshi: Surface Science [**438**]{} (1999) 261.
$38)$ N. Mingo and K. Makoshi: Phys. Rev. Lett. [**84**]{} (2000) 3694.
$39)$ K. Makoshi, N. Mingo, T. Mii, H. Ueba, and S. Tikhodeev: Surface Science [**493**]{} (2001) 71.
$40)$ N. Mingo, K. Makoshi, T. Mii, and H. Ueba: Surface Science [**482-485**]{} (2001) 96.
$41)$ B. N. J. Persson, H. Ueba: Surface Science [**502-503**]{} (2002) 18.
$42)$ T. Mitsui and K. Takayanagi: Phys. Rev. B [**62**]{} (2000) R16251.
------- -------------------------------------------------------------------------------------------------
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A countable discrete group $G$ is called Choquet-Deny if for every non-degenerate probability measure $\mu$ on $G$ it holds that all bounded $\mu$-harmonic functions are constant. We show that a finitely generated group $G$ is Choquet-Deny if and only if it is virtually nilpotent. For general countable discrete groups, we show that $G$ is Choquet-Deny if and only if none of its quotients has the infinite conjugacy class property. Moreover, when $G$ is not Choquet-Deny, then this is witnessed by a symmetric, finite entropy, non-degenerate measure.'
address:
- California Institute of Technology
- Northwestern University
author:
- Joshua Frisch
- Yair Hartman
- Omer Tamuz
- Pooya Vahidi Ferdowsi
bibliography:
- 'main.bib'
title: 'Choquet-Deny groups and the infinite conjugacy class property'
---
Introduction
============
Let $G$ be a countable discrete group. A probability measure $\mu$ on $G$ is [*non-degenerate*]{} if its support generates $G$ as a semigroup.[^1] A function $f \colon G \to {\mathbb{R}}$ is [*$\mu$-harmonic*]{} if $f(k) = \sum_{g \in G}\mu(g) f(k g)$ for all $k \in G$. We say that the [*measured group*]{} $(G,\mu)$ is [*Liouville*]{} if all the bounded $\mu$-harmonic functions are constant; this is equivalent to the triviality of the Poisson boundary $\Pi(G,\mu)$ [@furstenberg1963noncommuting; @furstenberg1971random; @furstenberg1973boundary] (also called the Furstenberg-Poisson boundary; for formal definitions see also, e.g., Furstenberg and Glasner [@furstenberg2009stationary], Bader and Shalom [@bader2006factor], or a survey by Furman [@furman2002random]).
When $G$ is non-amenable, $(G,\mu)$ is not Liouville for every non-degenerate $\mu$ [@furstenberg1973boundary]. Conversely, when $G$ is amenable, then there exists some non-degenerate $\mu$ such that $(G,\mu)$ is Liouville, as shown by Kaimanovich and Vershik [@kaimanovich1983random] and Rosenblatt [@rosenblatt1981ergodic]. It is natural to ask for which groups $G$ it holds that $(G,\mu)$ is Liouville for [*every*]{} non-degenerate $\mu$. We call such groups [*Choquet-Deny*]{} groups; as we discuss in §\[sec:CD-defs\], there are a few variants of this definition (see, e.g., [@guivarc1973croissance; @glasner1976proximal; @glasner1976choquet], or [@jaworski2007choquet]), which, however, we show to be equivalent.
The classical Choquet-Deny Theorem (which was first proved for ${\mathbb{Z}}^d$ by Blackwell [@blackwell1955transient]) states that abelian groups are Choquet-Deny [@cd1960]; the same holds for virtually nilpotent groups [@dynkin1961random]. There are many examples of amenable groups that are not Choquet-Deny: first examples of such groups[^2] are due to Kaimanovich [@kaimanovich1983examples] and Kaimanovich and Vershik [@kaimanovich1983random], they include locally finite groups; Erschler shows that finitely generated solvable groups that are not virtually nilpotent are not Choquet-Deny [@erschler2004liouville], and that even some groups of intermediate growth are not Choquet-Deny [@erschler2004boundary]. Kaimanovich and Vershik [@kaimanovich1983random p. 466] conjecture that “Given an exponential group G, there exists a symmetric (nonfinitary, in general) measure with non-trivial boundary.” See Bartholdi and Erschler [@bartholdi2017poisson] for additional related results and further references and discussion.
Our main result is a characterization of Choquet-Deny groups. We say that $G$ has the [*infinite conjugacy class*]{} property (ICC) if it is non-trivial, and if each of its non-trivial elements has an infinite conjugacy class. We say that $\mu$ is [*fully supported*]{} if $\operatorname{supp}\mu=G$; obviously this implies that $\mu$ is non-degenerate.
\[thm:main\] A countable discrete group $G$ is Choquet-Deny if and only if it has no ICC quotients. Moreover, when $G$ does have an ICC quotient, then there exists a fully supported, symmetric, finite entropy probability measure $\mu$ on $G$ such that $(G,\mu)$ is not Liouville. In particular, if $G$ is finitely generated, then it is Choquet-Deny if and only if it is virtually nilpotent.
That a group with no ICC quotients is Choquet-Deny was shown by Jaworski [@jaworski2004countable Theorem 4.8].[^3] Our contribution is therefore in the proof of the converse, which appears in §\[sec:hard\].
Groups with no ICC quotients are known as FC-hypercentral (see, e.g., [@mclain1956remarks; @duguid1956fc], or [@robinson1972finiteness §4.3]). This class is closed under forming subgroups, quotients, direct products and finite index extensions, and includes all virtually nilpotent groups. Among finitely generated groups, virtually nilpotent groups are precisely those with no ICC quotients (see [@mclain1956remarks Theorem 2] and [@duguid1956fc Theorem 2]); this implies the result in Theorem \[thm:main\] for finitely generated groups. Since finitely generated groups of exponential growth are not virtually nilpotent, Theorem \[thm:main\] implies that the above mentioned conjecture of Kaimanovich and Vershik [@kaimanovich1983random] is correct.
A very recent result by three of the authors of this paper shows that a countable discrete group is strongly amenable if and only if it has no ICC quotients [@frisch2018strong]. This implies that $G$ is strongly amenable if and only if $(G,\mu)$ is Liouville for every non-degenerate $\mu$, paralleling the above mentioned characterization of amenability as equivalent to the existence of a non-degenerate $\mu$ such that $(G,\mu)$ is Liouville. While the proofs of these two similar results are different, it is natural to ask whether there is some deeper connection between strong amenability and the Choquet-Deny property.
Different possible definitions of Choquet-Deny groups {#sec:CD-defs}
-----------------------------------------------------
Our definition of Choquet-Deny groups is not the usual one, which states that a group is Choquet-Deny if $(G,\mu)$ is Liouville for every [*adapted*]{} measure $\mu$, where $\mu$ is called adapted if its support generates $G$ as a [*group*]{} (rather than as a semigroup, as in the non-degenerate case) [@guivarc1973croissance; @glasner1976proximal; @glasner1976choquet]. Yet another definition used in the literature requires that for [*every*]{} $\mu$, every bounded $\mu$-harmonic function is constant on the left cosets of $G_{\mu}$, where $G_\mu$ is the subgroup of $G$ generated by the support of $\mu$ [@jaworski2007choquet].
While a priori these are different definitions, they are equivalent, as demonstrated by our result and by Jaworski’s Theorem 4.8 in [@jaworski2004countable]. Jaworski’s result shows that groups with no ICC quotients are Choquet-Deny according to any of these definitions. Since our construction of $\mu$ with a non-trivial boundary yields measures that are supported on all of $G$ (hence non-degenerate, hence adapted), it shows that groups with ICC quotients are not Choquet-Deny according to any of these definitions. Moreover, our result shows that the class of Choquet-Deny groups (whether defined with adapted or with non-degenerate measures) is closed under taking subgroups, which, to the best of our knowledge, was also not previously known.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We would like to thank Anna Erschler and Vadim Kaimanovich for many useful comments on the first draft of this paper. We thank Wojciech Jaworski for bringing a number of errors to our attention and suggesting many improvements. We likewise thank an anonymous referee for many helpful suggestions.
Proofs {#sec:hard}
======
In this section we prove the main result of our paper, Theorem \[thm:main\]. Unless stated otherwise, we will assume that all groups are countable and discrete.
Recall that a probability measure $\mu$ on $G$ is symmetric if $\mu(g)=\mu(g^{-1})$ for all $g \in G$. Its Shannon entropy (or just entropy) is $H(\mu)=-\sum_{g \in G}\mu(g)\log\mu(g)$.
Our Theorem \[thm:main\] is a direct consequence of [@jaworski2004countable Theorem 4.8], which proves it for the case of groups with no ICC quotients, and of the following proposition, which handles the case of groups with ICC quotients.
\[prop:hard-direction\] Let $G$ be a group with an ICC quotient. Then there exists a fully-supported, symmetric, finite entropy probability measure $\mu$ on $G$ such that $\Pi(G,\mu)$ is non-trivial.
The main technical effort in the proof of Proposition \[prop:hard-direction\] is in the proof of the following proposition.
\[prop:non-invariant\] Let $G$ be an amenable ICC group. For every $h \in G \setminus \{e\}$ there exists a fully supported, symmetric, finite entropy probability measure $\mu$ such that $$\begin{aligned}
\label{eq:mu}
\lim_{m\to\infty}{\left\lVerth\mu^{*m}-\mu^{*m}\right\rVert} > 0.
\end{aligned}$$
Here $\mu^{*m}$ is the $m$-fold convolution $\mu * \cdots * \mu$. We will prove this Proposition later, and now turn to the proof of Proposition \[prop:hard-direction\].
The case of non-amenable $G$ is known, so assume that $G$ is amenable and has an ICC quotient $Q$. Let $h$ be a non-identity element of $Q$. Applying Proposition \[prop:non-invariant\] to $Q$ and $h$ yields a finite entropy, symmetric measure $\bar{\mu}$ on $Q$ that is fully supported, and satisfies .
Since $\bar{\mu}$ has full support and satisfies , it follows from [@glasner1976choquet Theorem 2] that $(Q,\bar{\mu})$ has a non-trivial Poisson boundary. Let $\mu$ be any symmetric, finite entropy non-degenerate probability measure on $G$ that is projected to $\bar{\mu}$; the existence of such a $\mu$ is straightforward. Then $(G,\mu)$ has a non-trivial Poisson boundary.
Switching Elements {#sec:switching}
------------------
Here we introduce two notions: switching elements and super-switching elements. We will use these notions in the proof of Proposition \[prop:non-invariant\].
Let $X$ be a finite symmetric subset of a group $G$.
- We call $g\in G$ a [*switching element for $X$*]{} if $$X \cap g X g^{-1} \subseteq \{e\}.$$
- We call $g\in G$ a [*super-switching element for $X$*]{} if $$X \cap \big(g Xg\cup g Xg^{-1} \cup g^{-1} Xg \cup g^{-1} Xg^{-1}\big) \subseteq \{e\}.$$
Note that since $X$ is symmetric, $g\in G$ is a switching element for $X$ if and only if $g^{-1}$ is a switching element for $X$.
\[claim:super-switching\] Let $X$ be a finite symmetric subset of a group $G$ and let $g\in G$ be a super-switching element for $X$. If $g^{w_1} x g^{w_2} = y$ for $x,y\in X$ and $w_1, w_2\in \{-1, +1\}$, then $x=y=e$.
Let $g^{w_1} x g^{w_2} = y$ for $x,y\in X$ and $w_1, w_2\in \{-1, +1\}$. Since $$y=g^{w_1} x g^{w_2} \in \big(g Xg\cup g Xg^{-1} \cup g^{-1} Xg \cup g^{-1} Xg^{-1}\big)$$ and $y\in X$, it follows from the definition of a super-switching element for $X$ that $y=e$.
From $g^{w_1} x g^{w_2} = y$, we get $g^{-w_1} y g^{-w_2} = x$. So, by symmetry, the same argument shows $x=e$.
\[prop:switching-element\] Let $G$ be a discrete (not necessarily countable) amenable ICC group, and let $X$ be a finite symmetric subset of $G$. The set of super-switching elements for $X$ is infinite.
Fix an invariant finitely additive probability measure $d$ on $G$. For $A \subseteq G$, we call $d(A)$ the density of $A$. We will need the fact that infinite index subgroups have zero density, and that $d(A) = 0$ for every finite subset $A \subset G$.
Let ${C_G(x)}$ be the centralizer of a non-identity $x\in X$. Then, since $X$ is finite, there is a finite set of cosets of ${C_G(x)}$ that includes all $g \in G$ such that $g^{-1}xg\in X$. So, non-switching elements for $X$ are in the union of finitely many cosets of subgroups with infinite index, since $G$ is ICC. This means that the set of non-switching elements for $X$ has zero density, and so the set $S$ of switching elements for $X$ has density one.
Let $T$ be the set of all super-switching elements for $X$. Let $A \subseteq G$ be the set of involutions $\{g\in G \ | \ g^2=e\}$.
If $d(A) > 0$, then $d(A\cap S)>0$. On the other hand, for any $g\in A\cap S$, since $g$ is switching for $X$ and $g^{-1}=g$, $g$ is super-switching for $X$. Hence $A\cap S \subseteq T$. This shows that if $d(A)>0$, then $d(T) \geq d(A\cap S) > 0$, and so we are done.
So, we can assume that $d(A)=0$. For any $x, y\in X$, let $S_{x,y} = \{ g\in S\ | \ gxg = y\}$. Note that $$T = S \setminus \bigcup_{\substack{x, y\in X\\ (x,y)\neq(e,e)}} S_{x,y}.$$ It is thus enough to be shown that each $S_{x,y}$ has zero density when $(x,y)\neq(e,e)$. So assume for the sake of contradiction that $d(S_{x,y})>0$. Fix $g\in S_{x,y}$. We have the following for all $h\in g^{-1}S_{x,y}$. $$\begin{aligned}
gxg=y=gh x gh\implies&\ (xg)= h(xg)h\\
\implies&\ (xg)^{-1} h^{-1} (xg) = h\\
\implies&\ h = (xg)^{-1} h^{-1} (xg)\\
&\quad = (xg)^{-1} [(xg)^{-1} h^{-1} (xg)]^{-1} (xg)\\
&\quad = (xg)^{-2} h (xg)^2\\
\implies&\ \text{$h$ is in the centralizer of $(xg)^2$.}
\end{aligned}$$ So, the centralizer of $(xg)^2$ includes $g^{-1} S_{x,y}$, which has a positive density. So, the centralizer of $(xg)^2$ has finite index. This implies that $(xg)^2 = e$, because in an ICC group only the identity can have a finite index centralizer. Hence $x g \in A$ for all $g \in S_{x,y}$. So $x S_{x,y} \subseteq A$. Hence $S_{x,y}$ also has zero density, which is a contradiction.
A Heavy-Tailed Probability Distribution on ${\mathbb{N}}$. {#sec:heavy-tailed}
----------------------------------------------------------
Here we state and prove a lemma about the existence of a probability distribution on ${\mathbb{N}}=\{1,2,\ldots\}$ such that infinite i.i.d. samples from this measure have certain properties. We will use this distribution in the proof of Proposition \[prop:non-invariant\].
\[lemma:random-string\] Let $p$ be the following probability measure on ${\mathbb{N}}$: $p(n) = cn^{-5/4}$, where $1/c=\sum_{n=1}^\infty n^{-5/4}$. Then $p$ has finite entropy and the following property: for any ${\varepsilon}>0$ there exist constants $K_{\varepsilon},N_{\varepsilon}\in{\mathbb{N}}$ such that for any natural number $m\geq K_{\varepsilon}$ there exists an $E_{{\varepsilon},m}\subseteq {\mathbb{N}}^m$ such that:
1. $p^{\times m}(E_{{\varepsilon},m}) \geq 1-{\varepsilon}$, where $p^{\times m}$ is the $m$-fold product measure $p \times \cdots \times p$.
2. For any $s=(s_1,\ldots,s_m)\in E_{{\varepsilon},m}$, the maximum of $\{s_1,\ldots,s_{K_{\varepsilon}}\}$ is at most $N_{\varepsilon}$.
3. For any $s=(s_1,\ldots,s_m)\in E_{{\varepsilon},m}$ and for any $K_{\varepsilon}\leq k\leq m$, the maximum of $\{s_1,\ldots,s_k\}$ is at least $k^2$.
4. For any $s=(s_1,\ldots,s_m)\in E_{{\varepsilon},m}$ and for any $K_{\varepsilon}\leq k\leq m$, the maximum of $\{s_1,\ldots,s_k\}$ appears in $(s_1,\ldots,s_k)$ only once.
It is straightforward to see that $p$ has finite entropy.
Let $s=(s_1,s_2,\ldots)\in {\mathbb{N}}^\infty$ have distribution $p^{\times \infty}$; i.e., $s$ is a sequence of i.d.d. random variables with distribution $p$. Since each $s_i$ has distribution $p$, for each $n\in{\mathbb{N}}$ we have: $$\begin{aligned}
\label{eq:p}
\Pr{s_i \geq n} = \sum_{m=n}^\infty p(m) = c \sum_{m=n}^\infty m^{-5/4} \geq c \int_{n}^\infty x^{-5/4} \dd x = 4c n^{-1/4}.
\end{aligned}$$
For $k\geq 1$, let $$\begin{aligned}
M_k \coloneqq & \max\{s_1,\ldots,s_k\},
\end{aligned}$$ and let $$\begin{aligned}
\text{next}(k) \coloneqq & \min\{i > k \ | \ s_i \geq M_k \}.
\end{aligned}$$ In words, $\text{next}(k)$ is the first index $i > k$ for which $s_i$ matches or exceeds $M_k$.
We first show that with probability one, $M_k \geq k^2$ for all $k$ large enough. To this end, let $A_k$ be the event that $M_k < k^2$. We have: $$\begin{aligned}
\Pr{A_k} = & \ \Pr{s_i < k^2 \ \forall i\in\{1,\ldots,k\}}\\
=& \ (1-\Pr{s_1 < k^2})^k\\
\leq & \ (1-4c(k^2)^{-1/4})^k\\
\leq& \ e^{-4ck^{1/2}}.
\end{aligned}$$ Since the sum of these probabilities is finite, by Borel-Cantelli we get that $$\begin{aligned}
\Pr{A_k\ \text{infinitely often}} = 0.
\end{aligned}$$ Hence $M_k \geq k^2$ for all $k$ large enough, almost surely. Furthermore, the expectation of $1/M_k$ is small: $$\begin{aligned}
\label{eq:emk}
\E{\frac{1}{M_k}}
=
\E{\frac{1}{M_k}}{A_k}\Pr{A_k} +
\E{\frac{1}{M_k}}{\neg A_k}\Pr{\neg A_k}
\leq e^{-4ck^{1/2}} + \frac{1}{k^2}\cdot
\end{aligned}$$
Next, we show that, with probability one, $s_{\text{next}(k)} >
M_k$ for all $k$ large enough. That is, for large enough $k$, the first time that $M_k$ is matched or exceeded after index $k$, it is in fact exceeded.
Let $B_k$ be the event that $s_{\text{next}(k)} = M_k$. We would like to show that this occurs only finitely often. Note that $$\begin{aligned}
\Pr{B_k}{M_k}
&= \Pr{s_{\text{next}(k)}=M_k}{M_k}\\
&= \sum_{i=k+1}^\infty\Pr{s_i=M_k, \text{next}(k)=i}{M_k}.
\end{aligned}$$ Applying the definition of $\text{next}(k)$ yields $$\begin{aligned}
\Pr{B_k}{M_k}
&= \sum_{i=k+1}^\infty\Pr{s_i=M_k, s_{k+1},\ldots,s_{i-1}<M_k}{M_k}.
\end{aligned}$$ By the independence of the $s_i$’s we can write this as $$\begin{aligned}
\Pr{B_k}{M_k}
&= \sum_{i=k+1}^\infty\Pr{s_i=M_k}{M_k}\prod_{n=1}^{i-(k+1)}\Pr{s_{k+n}<M_k}{M_k}\\
&= \sum_{i=k+1}^\infty\frac{c}{M_k^{5/4}}\Pr{s_{k+1}<M_k}{M_k}^{i-(k+1)}.
\end{aligned}$$ By , $\Pr{s_{k+1}<M_k}{M_k} \leq
1-4 c M_k^{-1/4}$. Hence $$\begin{aligned}
\Pr{B_k}{M_k}
\leq \frac{c}{M_k^{5/4}} \cdot \frac{1}{4 c M_k^{-1/4}}
=\frac{1}{4 M_k}\cdot
\end{aligned}$$ Using it follows that $$\begin{aligned}
\Pr{B_k} = \E{\Pr{B_k}{M_k}} \leq \E{\frac{1}{4M_k}} \leq \frac{1}{4}e^{-4ck^{1/2}} + \frac{1}{4k^2}.
\end{aligned}$$ Hence $\sum_k \Pr{B_k} < \infty$, and so by Borel-Cantelli $B_k$ occurs only finitely often.
Since $A_k$ and $B_k$ both occur for only finitely many $k$, the (random) index $\text{ind}'$ at which they stop occurring is almost surely finite, and is given by $$\text{ind}' = \min\{\ell \in {\mathbb{N}}\,:\, s \not \in A_k \cup B_k \text{ for all } k \geq \ell\}.$$ Let $$\text{ind} = \text{next}(\text{ind}').$$ Hence for $k \geq \text{ind}$, $M_k \geq k^2$ and $M_k$ appears in $(s_1,\ldots,s_k)$ only once.
Fix ${\varepsilon}> 0$. Since $\text{ind}$ is almost surely finite, then for large enough constants $K_{\varepsilon}\in{\mathbb{N}}$ and $N_{\varepsilon}\in {\mathbb{N}}$ the event $$E_{\varepsilon}= \{\text{ind} \leq K_{\varepsilon}\text{ and } M_{K_{\varepsilon}}\leq N_{\varepsilon}\}$$ has probability at least $1-{\varepsilon}$, and additionally, conditioned on $E_{\varepsilon}$ it holds that $k \geq \text{ind}$ for all $k \geq K_{\varepsilon}$, and hence $M_k \geq k^2$ and $M_k$ appears in $(s_1,\ldots,s_k)$ only once. Therefore, if for $m \geq K_{\varepsilon}$ we let $E_{{\varepsilon},m}$ be the projection of $E_{\varepsilon}$ to the first $m$ coordinates, then $E_{{\varepsilon},m}$ satisfies the desired properties.
Proof of Proposition \[prop:non-invariant\]
-------------------------------------------
Let $\frac{1}{8}>{\varepsilon}>0$. Let $p$, $K_{\varepsilon}\in {\mathbb{N}},N_{\varepsilon}\in {\mathbb{N}}$, and $E_{{\varepsilon},m}\subseteq {\mathbb{N}}^m$ be the probability measure, the constants, and the events from Lemma \[lemma:random-string\]. To simplify notation let $N=N_{\varepsilon}$ and $K=K_{\varepsilon}$.
Let $G = \{a_1,a_2,\ldots\}$, where $a_1=a_2=\cdots=a_N=e$. We define $(g_n)_{n}$, $(A_n)_{n}$, $(B_n)_n$ and $(C_n)_n$ recursively. Given $g_1,\ldots,g_{n}$, let $A_n = \{g_n, g_n^{-1},a_n, a_n^{-1}\}$ and $B_n=\cup_{i\leq n} A_i$. Denote $C_n = B_n\cup\{h^{-1},h\}$. Note that $A_n$, $B_n$, and $C_n$ are finite and symmetric for any $n\in {\mathbb{N}}$. Let $g_1=g_2=\ldots=g_N=e$. For $n +1 > N$, given $C_n$, let $g_{n+1} \in G$ be a super-switching element for $(C_n)^{2n+1}$ which is not in $(C_n)^{8n+1}$. The existence of such a super-switching element is guaranteed by Proposition \[prop:switching-element\] and the facts that $(C_n)^{2n+1}$ is a finite symmetric subset of $G$ and that $(C_n)^{8n+1}$ is finite.
For $n \in {\mathbb{N}}$, define a symmetric probability measure $\mu_n$ on $A_n$ by $$\mu_n = {\varepsilon}2^{-n} (\frac{1}{2}\delta_{a_n} + \frac{1}{2} \delta_{a_n^{-1}}) + (1-{\varepsilon}2^{-n}) (\frac{1}{2}\delta_{g_n} + \frac{1}{2}\delta_{g_n^{-1}}).$$ Here $\delta_g$ is the point mass on $g \in G$. Finally, let $$\mu = \sum_{n=1}^{\infty} p(n) \mu_n.$$ Obviously $\mu$ is symmetric and $\operatorname{supp}\mu=G$. Since $p$ has finite entropy and each $\mu_n$ has support of size at most 4, it follows easily that $\mu$ has finite entropy.
We want to show that $$\lim_{m\to\infty}{\left\lVerth\mu^{*m}- \mu^{*m}\right\rVert} > 0.$$
Fix $m\in {\mathbb{N}}$ larger than $K$ and $N$. For each $n \in {\mathbb{N}}$ define $f_n:\{1,2,3,4\}\to A_n$ by $$\begin{aligned}
f_n(1) = a_n,\
f_n(2) = a_n^{-1},\
f_n(3) = g_n,\
f_n(4) = g_n^{-1},
\end{aligned}$$ and define $\nu_n:\{1,2,3,4\}\to [0,1]$ by $$\begin{aligned}
\nu_n(1) = \nu_n(2) = \frac{1}{2} {\varepsilon}2^{-n},\
\nu_n(3) = \nu_n(4) = \frac{1}{2} (1-{\varepsilon}2^{-n}).
\end{aligned}$$ Let $$\Omega = \{(s,w)\ | \ s\in{\mathbb{N}}^m, \ w\in\{1,2,3,4\}^m \}.$$
We define the measure $\eta$ on the countable set $\Omega$ by specifying its values on the singletons: $$\eta(\{(s,w)\}) = p^{\times m}(s)\ \nu_{s_1}(w_1)\ \nu_{s_2}(w_2)\ \ldots\ \nu_{s_m}(w_m).$$ It follows immediately from this definition that $\eta$ is a probability measure.
Define $r:\Omega \to G$ by $$r(s,w) = f_{s_1}(w_1) f_{s_2}(w_2) \ldots f_{s_m}(w_m).$$ It is not difficult to see that $r_{*}\eta = \mu^{*m}$, and so we need to show that ${\left\lVerth r_*\eta - r_*\eta\right\rVert}$ is uniformly bounded away from zero for $m$ larger than $K$ and $N$.
Recall that $E_{{\varepsilon},m}\subseteq {\mathbb{N}}^m$ is the event given by Lemma \[lemma:random-string\]. Fix $s\in E_{{\varepsilon},m}$. Define $$\begin{aligned}
i_{s,1} =&\ \min\{j\in \{1\ldots,m\} \ | \ s_j > N\},\\
i_{s,2} =&\ \min\{j > i_{s,1} \ | \ s_j \geq s_{i_{s,1}}\},\\
& \vdots \\
i_{s,l(s)} =&\ \min\{j > i_{s,l(s)-1} \ | \ s_j \geq s_{i_{s,l(s)-1}} \}.
\end{aligned}$$ Note that by the second property of $E_{{\varepsilon},m}$ in Lemma \[lemma:random-string\], we know that $$\begin{aligned}
K < i_{s,1} < i_{s,2} < \cdots < i_{s,l(s)},
\end{aligned}$$ and by the fourth property, $$\begin{aligned}
N< s_{i_{s,1}} < s_{i_{s,2}} < \cdots < s_{i_{s,l(s)}} = \max\{ s_1,\ldots, s_m\}.
\end{aligned}$$ Let $$W_{\varepsilon}^{s}=\{w\in\{1,2,3,4\}^m \ | \ \forall k\leq l(s) \ w_{i_{s,k}}=3,4 \}.$$
For $s \in {\mathbb{N}}^m$ let $\eta_s$ be the measure $\eta$, conditioned on the first coordinate equalling $s$. I.e., let $$\eta_s(A) = \frac{\eta(A \cap \Omega^{s})}{\eta(\Omega^{s})},$$ where $\Omega^{s} = \{s\}\times\{1,2,3,4\}^m \subseteq \Omega$.
Then $$\begin{aligned}
\eta_s(\{s\}\times W_{\varepsilon}^{s})
=&\ 1 - \eta_s(\{w_{i_{s,1}}=1,2; \text{ or } w_{i_{s,2}}=1,2;\ \ldots; \text{ or } w_{i_{s,l(s)}}=1,2\ \})\\
\geq&\ 1 - \sum_{k=1}^{l(s)} \eta_s(\{w_{i_{s,k}}=1,2\})\\
=&\ 1 - \sum_{k=1}^{l(s)} {\varepsilon}2^{-s_{i_{s,k}}}\\
\geq&\ 1 - \sum_{j=1}^\infty {\varepsilon}2^{-j}\\
=&\ 1-{\varepsilon},
\end{aligned}$$ where the first inequality follows from the union bound, and the last inequality holds since $s_{i_{s,1}} < s_{i_{s,2}} < \cdots < s_{i_{s, l(s)}}$.
Finally, let $$\Omega_{\varepsilon}= \{(s,w)\in\Omega \ | \ s\in E_{{\varepsilon},m}, \ w\in W_{\varepsilon}^{s} \}.$$ By the above, and since $\eta(E_{{\varepsilon},m}\times\{1,2,3,4\}^m) \geq 1-{\varepsilon}$ by Lemma \[lemma:random-string\], we have shown that $$\begin{aligned}
\eta(\Omega_{\varepsilon})\geq\ (1-{\varepsilon})(1-{\varepsilon})>\ 1-2{\varepsilon}.
\end{aligned}$$
\[claim:E\_eps\] For any $\alpha, \beta\in \Omega_{\varepsilon}$, we have $h r(\alpha) \neq r(\beta)$.
We prove this claim after we finish the proof of the Proposition.
Let $\eta_1$ be equal to $\eta$ conditioned on $\Omega_{\varepsilon}$, and $\eta_2$ be equal to $\eta$ conditioned on the complement of $\Omega_{\varepsilon}$. We have $\eta = \eta(\Omega_{\varepsilon}) \eta_1 + (1-\eta(\Omega_{\varepsilon})) \eta_2$, and by the above claim we know ${\left\lVerthr_*\eta_1 - r_*\eta_1\right\rVert} = 2$. So for $m$ larger than $K$ and $N$ $$\begin{aligned}
{\left\lVerth\mu^{*m}- \mu^{*m}\right\rVert} =&\ {\left\lVerthr_*\eta - r_*\eta\right\rVert}\\
=&\ {\left\lVert\eta(\Omega_{\varepsilon})(h r_*\eta_1 - r_*\eta_1) + (1-\eta(\Omega_{\varepsilon}))(h r_*\eta_2 - r_*\eta_2)\right\rVert}\\
\geq&\ \eta(\Omega_{\varepsilon}) {\left\lVerth r_*\eta_1 - r_*\eta_1\right\rVert} - 2 (1-\eta(\Omega_{\varepsilon}))\\
\geq&\ 2(1-2{\varepsilon}) -2(2{\varepsilon})= 2-8{\varepsilon},
\end{aligned}$$ which is uniformly bounded away from zero since ${\varepsilon}< \frac{1}{8}$. Since ${\left\lVerth\mu^{*m}- \mu^{*m}\right\rVert}$ is a decreasing sequence, this completes the proof of Proposition \[prop:non-invariant\].
Let $\alpha=(s,w),\ \beta=(t,v)\in \Omega_{\varepsilon}$. Hence $\max\{K,N\}< m$, $s\in E_{{\varepsilon},m}$, $t\in E_{{\varepsilon},m}$, $w\in W_{\varepsilon}^{s}$, and $v\in W_{\varepsilon}^{t}$. Assume that $hr(\alpha) = r(\beta)$. So, we have $$hf_{s_1}(w_1)\cdots f_{s_m}(w_m) = f_{t_1}(v_1)\cdots f_{t_m}(v_m).$$ Let $K<i_1<i_2<\cdots<i_{l(s)}$ and $K<j_1<j_2<\cdots<j_{l(t)}$ be the indices we defined for $s$ and $t$ in the proof of Proposition \[prop:non-invariant\]. We remind the reader that the unique maximum of $(s_1,\ldots,s_m)$ is attained at $i_{l(s)}$, with a corresponding statement for $(t_1,\ldots,t_m)$ and $j_{l(t)}$. So we have $$\begin{aligned}
h &\overbrace{f_{s_1}(w_1)\cdots f_{s_{i_{l(s)}-1}}(w_{i_{l(s)}-1})}^{b_1} f_{s_{i_{l(s)}}}(w_{i_{l(s)}}) \overbrace{f_{s_{i_{l(s)}+1}}(w_{i_{l(s)}+1})\cdots f_{s_m}(w_m)}^{b_2}\\
=&\ \underbrace{f_{t_1}(v_1)\cdots f_{t_{j_{l(t)}-1}}(v_{j_{l(t)}-1})}_{c_1} f_{t_{j_{l(t)}}}(v_{j_{l(t)}}) \underbrace{f_{t_{j_{l(t)}+1}}(v_{j_{l(t)}+1})\cdots f_{t_m}(v_m)}_{c_2}.
\end{aligned}$$ Let $p = s_{i_{l(s)}} = \max\{s_1,\ldots,s_m\}$ and $q = t_{j_{l(t)}} = \max\{t_1,\ldots,t_m\}$. Since $w\in W_{\varepsilon}^{s}$ and $v\in W_{\varepsilon}^{t}$, we know $f_{s_{i_{l(s)}}}(w_{i_{l(s)}}) = g_p^{\pm 1}$ and $f_{t_{j_{l(t)}}}(v_{j_{l(t)}}) = g_q^{\pm 1}$, so $$\begin{aligned}
\label{eq:hrr}
h b_1 g_p^{\pm 1} b_2 = c_1 g_q^{\pm 1} c_2.
\end{aligned}$$ Since $p = \max\{s_1,\ldots,s_m\}$, and since $m\geq K$, we know that $m\leq m^2 \leq p$. So $b_1,b_2\in (B_{p-1})^{p-1}\subseteq (C_{p-1})^{p-1}$. Similarly $c_1,c_2\in (C_{q-1})^{q-1}$.
Consider the case that $p>q$. Then $c_1,c_2,g_q^{\pm 1}\in (C_q)^q\subseteq (C_{p-1})^{p-1}$. Hence $g_p^{\pm 1} = [b_1^{-1}] h^{-1} [c_1 g_q^{\pm 1} c_2 b_2^{-1}]$ by , and so $$g_p \in (C_{p-1})^{4(p-1)} \{h,h^{-1}\} (C_{p-1})^{4(p-1)}\subseteq (C_{p-1})^{8(p-1)+1},$$ which is a contradiction with our choice of $g_p$, since $p>N$. Similarly, if $p<q$, we get a contradiction. So we can assume that $p=q$.
If $p = q$, then by we have $$h b_1 g_p^{\pm 1} b_2 = c_1 g_p^{\pm 1} c_2,$$ and $c_1, c_2, b_1, b_2 \in (C_{p-1})^{p-1}$. So, for $x=c_1^{-1} h b_1\in (C_{p-1})^{2(p-1)+1}$ we have $g_p^{\pm 1} x g_p^{\pm 1} = c_2 b_2^{-1} \in (C_{p-1})^{2(p-1)}\subseteq (C_{p-1})^{2(p-1)+1}$. By the fact that $g_p$ is a super-switching element for $(C_{p-1})^{2(p-1)+1}$ and from Claim \[claim:super-switching\], we get that $x$ is the identity.
So $h b_1 = c_1$, i.e. $$hf_{s_1}(w_1)\cdots f_{s_{i_{l(s)}-1}}(w_{i_{l(s)}-1}) = f_{t_1}(v_1)\cdots f_{t_{j_{l(t)}-1}}(v_{j_{l(t)}-1}).$$ By the exact same argument, we can see this leads to a contradiction unless $$hf_{s_1}(w_1)\cdots f_{s_{i_{l(s)-1}-1}}(w_{i_{l(s)-1}-1}) = f_{t_1}(v_1)\cdots f_{t_{j_{l(t)-1}-1}}(v_{j_{l(t)-1}-1}).$$ And again, this leads to a contradiction unless $$hf_{s_1}(w_1)\cdots f_{s_{i_{l(s)-2}-1}}(w_{i_{l(s)-2}-1}) = f_{t_1}(v_1)\cdots f_{t_{j_{l(t)-2}-1}}(v_{j_{l(t)-2}-1}).$$ Note that if $l(s)\neq l(t)$, at some point in this process we get that either all the $s_i$’s or all the $t_i$’s are at most $N$ while the other string has characters strictly greater than $N$. This leads to a contradiction similar to the case $p\neq q$, which we explained before. So, by continuing this process, we get a contradiction unless $$\begin{aligned}
\label{eq:last step}
h f_{s_1}(w_1)\cdots f_{s_{i_1-1}}(w_{i_1-1}) = f_{t_1}(v_1)\cdots f_{t_{j_1-1}}(v_{j_1-1}).
\end{aligned}$$ Note that $s_1,\ldots,s_{i_1-1}\leq N$, which implies $$f_{s_1}(w_1)=\cdots=f_{s_{i_1-1}}(w_{i_1-1})=e.$$ Similarly, $t_1,\ldots,t_{j_1-1}\leq N$ implies that $$f_{t_1}(v_1)=\cdots=f_{t_{j_1-1}}(v_{j_1-1})=e.$$ So, from we get $h=e$, which is a contradiction.
[^1]: In the context of Markov chains such measures are called [*irreducible*]{}.
[^2]: In the Lie group setting, an example of an amenable group that is not Choquet-Deny was already known to Furstenberg [@furstenberg1963noncommuting].
[^3]: In fact, Jaworski proves there a stronger statement; see the discussion in §\[sec:CD-defs\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper formulates and presents a solution to the new problem of budgeted semantic video segmentation. Given a video, the goal is to accurately assign a semantic class label to every pixel in the video within a specified time budget. Typical approaches to such labeling problems, such as Conditional Random Fields (CRFs), focus on maximizing accuracy but do not provide a principled method for satisfying a time budget. For video data, the time required by CRF and related methods is often dominated by the time to compute low-level descriptors of supervoxels across the video. Our key contribution is the new budgeted inference framework for CRF models that intelligently selects the most useful subsets of descriptors to run on subsets of supervoxels within the time budget. The objective is to maintain an accuracy as close as possible to the CRF model with no time bound, while remaining within the time budget. Our second contribution is the algorithm for learning a policy for the sparse selection of supervoxels and their descriptors for budgeted CRF inference. This learning algorithm is derived by casting our problem in the framework of Markov Decision Processes, and then instantiating a state-of-the-art policy learning algorithm known as Classification-Based Approximate Policy Iteration. Our experiments on multiple video datasets show that our learning approach and framework is able to significantly reduce computation time, and maintain competitive accuracy under varying budgets.'
author:
- |
Behrooz Mahasseni, Sinisa Todorovic, and Alan Fern\
Oregon State University\
Corvallis, OR 97331, USA\
[[email protected], [email protected], [email protected]]{}
bibliography:
- 'library.bib'
title: Approximate Policy Iteration for Budgeted Semantic Video Segmentation
---
Introduction {#sec:introduction}
============
[**Motivation:**]{} The design of vision approaches is typically informed by a trade-off between efficiency and accuracy. Good computational efficiency is usually achieved by taking a number of heuristic pre- and post-processing steps and integrating them with the main approach. For example, vision practitioners heuristically limit the types of features to be extracted (e.g., low-cost ones), as well as locations and scales in images and video from which they are extracted. While these steps have been satisfactory for small-scale problems, their heuristic nature makes an adaptation of existing systems to settings with stringent runtime requirements very difficult.
[**Our goal**]{} is to formulate a principled framework for optimally adapting a vision system to varying time budgets imposed by particular application settings, so as to maximize overall performance for any budget and maintain an accuracy as close as possible to the system’s performance with no time bound. In this paper, we focus our presentation of theory and experiments in the context of semantic video segmentation. Nevertheless, it is worth noting that our framework is based on fairly non-restrictive assumptions, and thus is suitable for many other computer vision problems, beyond the scope of this paper.
[**The Focus Vision Problem:**]{} Given a video, our goal is to assign a class label to every pixel from the set of semantic classes seen in training, under any time budget. We call this new problem *budgeted semantic video segmentation*. For example, in a video of a street, we want to efficiently segment spatiotemporal subvolumes occupied by cars, pedestrians, and buildings in less time than the user-specified bound. This is an important problem with a wide range of applications (e.g., driverless cars, sports video analytics) which require highly accurate and timely estimates of space-time extents of objects in the scene.
[**The key idea:**]{} We assume that a given approach to semantic video segmentation specifies an inference procedure (e.g., loopy belief propagation, graph-cut) which takes input features and outputs an inference result. Rather than pixels, most existing approaches, first, label supervoxels, obtained from an unsupervised (low-level) video partitioning, and then transfer these labels to the corresponding pixels. Thus, a typical input consists of supervoxels and their [*descriptors*]{}, where the descriptors may be computed locally at every supervoxel, between pairs of neighboring supervoxels, and globally across the entire video. Computing all descriptors for all supevoxels is costly. Thus, a sparse section of supervoxels and choosing their most useful and least costly descriptors is our key problem. Note that we do not modify the inference procedure. Rather, we adapt the descriptor computation step to suit budgeted settings of a broad family of approaches. Consequently, we do not seek to improve prediction accuracy, but rather maintain the existing level of accuracy while reducing runtime.
![We formalize a sequential inference policy aimed at adopting a fairly general family of approaches to the problem of budgeted semantic video segmentation. Our focus domain is a holistic CRF-based inference, but other approaches to semantic video segmentation could be considered. Given an unsupervised video partitioning into supervoxels, a set of feature extractors, and a user-specified time budget, our policy sequentially selects the best pair (supevoxel, feature) toward maximizing performance of the CRF inference until the time expires.[]{data-label="fig:BlockDiagram"}](BlockDiagram){width="\columnwidth"}
[**Our framework**]{} is illustrated in Fig. \[fig:BlockDiagram\]. We assume access to a Conditional Random Field (CRF) and associated inference procedure—one of the most popular formulations of semantic video segmentation—and design a [*sequential*]{} inference policy that interacts with the given approach. Given an unsupervised video partitioning into supervoxels and a user-specified time budget, our policy is presented with a sequence of candidate supervoxels until time expires and must decide for each one which new descriptor to run for it, if any, with the goal of maximizing performance of the CRF inference. The sequential selection should take into account both the immediate and long-term value of the decisions toward overall inference accuracy.
We formalize such sequential decision making in the framework of Markov Decision Processes (MDPs), where, for a given state, the policy selects the highest value action among possible actions. The policy state is defined by the previous selections of descriptors for supervoxels, which is conveniently represented via special policy features. The policy actions include running a descriptor for the currently presented supervoxel or [Finished]{}, which specifies that no further descriptors should be computed for the current supervoxel. The policy is defined as a [*linear*]{} ranking function for balancing efficiency and expressiveness, such that its execution consumes resources negligibly, and that it captures sufficient information about the current policy state to support good decisions. For training the policy, we use the state-of-the-art policy learning approach of Classification-Based Approximate Policy Iteration (CAPI), which is able to leverage state-of-the-art classification learning techniques for policy optimization.
Related Work and Our Contributions {#sec:priorwork}
==================================
Semantic video segmentation is mostly formulated as a graphical-model based labeling of supervoxels in the video [@chen2010propagating; @budvytis2010label; @brostow2008segmentation; @miksik2013efficient; @chen2011temporally; @liumulti; @taylor2013semantic; @wojek2008dynamic]. For example, graphical models were used for: i) Propagating manual annotations of supervoxels of the first few frames to other supervoxels in the video [@miksik2013efficient; @chen2010propagating; @budvytis2010label], or ii) Supervoxel labeling based on week supervision in training [@Liu_2014_CVPR]. The accuracy of such labeling can be improved by CRF-based reasoning about 3D relations [@kundu2014joint] or context [@liumulti] among object classes in the scene. Therefore, it seems reasonable that we develop our framework for an existing CRF-based labeling of supervoxels. None of these methods explicitly studied their runtime efficiency, except for a few empirical results of sensitivity to the total number of supervoxels used [@jain2013coarse].
Prior work has considered the issue of how to reduce inference runtime by specifying efficient approximations to the original inference [@zhang2012efficient; @krahenbuhl2012efficient; @a22001fast]. Their work is not related to ours, since we keep the given inference procedure intact. A few approaches have addressed cost-sensitive inference for activity recognition [@Mohamed_ECCV12; @Amer_ICCV13], image classification [@Efficiency_CVPR13; @AnytimeRecognition_CVPR14], and object detection [@Wu_ICCV2013]. Our fundamental difference is that these methods precompute all features and then adapt the very inference procedure such that it uses only an optimal subset of the features to meet the time budget. In contrast, we do not compute any features before our inference policy makes the decision to do so.
Closely related work improves runtime efficiency by reducing the costs of feature extraction [@Weiss_2013_ICCV; @NIPS2013_5142; @roig_active_2013; @liu2003multiclass; @grubb2013speedmachines]. For example, CRF-based semantic scene labeling in images is made more efficient by computing only a small subset of unary potentials for the CRF, and efficiently predicting the missing potentials from neighbors [@roig_active_2013]. This approach can be viewed as a special case of our framework, since they only select superpixels and then use [*all*]{} features for computing the unary potentials, whereas we would select both superpixels and feature types in their domain. In [@NIPS2013_5142], a budget constrained reinforcement learning is used to select optimal features for tracking human poses in relatively simple videos. Since they extract a chosen feature over the entire video, this approach can be viewed as another special case of our framework, because we would additionally make the decision about where to extract the feature in the video. This is a crucial difference for large videos, where computing even a single feature/descriptor across an entire video may exceed the budget constraint. Additional differences from these two approaches are explained in Sec. \[sec:problem-statement\]. We efficiently compute low-level descriptors for a selected subset of supervoxels in addition to the object level potentials in [@liu2003multiclass]. Unlike [@grubb2013speedmachines] we do not change the inference module, which makes it possible to augment any state of the art with our approach. It is not clear how to extend [@grubb2013speedmachines] to use higher order potentials. More importantly our approach is able to exploit the high correlation among the features of neighboring supervoxels to avoid the extra feature extraction cost if it does not help the final inference.
[**Contributions:**]{} 1) First formalization of the budgeted semantic video segmentation problem. 2) Design of the first learning algorithm that can tune inference policies for varying time budgets. 3) Specification and evaluation of the inference policy representation as a linear function that supports learning.
Budgeted Semantic Video Segmentation {#sec:problem-statement}
====================================
Our framework takes as input a time budget $B$ and video that has been partitioned into a set of supervoxels $V$. The goal is to accurately assign a label to each supervoxel in $V$ in time less than $B$. Below, we first present a common CRF formulation for semantic video segmentation, which our framework is based on. Next, we formulate our framework for bounding the CRF inference time.
A Common CRF Formulation {#sec:CRF}
------------------------
We consider a standard pairwise CRF model, which specifies the following score for any labeling $\{y_i\}$ of supervoxels $i \in V$ and their spatiotemporal neighborhood relationships $(i,j)\in E \subset V\times V$: $$\begin{aligned}
\label{eq:CRF}
\sum_{i \in V} \w_u\cdot \bpsi_u(\x_i,y_i) {+}
\sum_{(i, j) \in E} \w_p\cdot \psi_p(\x_i,\x_j,y_i,y_j),\end{aligned}$$ where $\w_u\cdot \bpsi_u(\x_i,y_i)$ denotes the unary potential specified in terms of an $n_u$-dimensional *unary feature vector* $\bpsi_u(\x_i,y_i)$ estimated for observation $\x_i$ at supervoxel $i$ when $i$ is assigned label $y_i$ from the set of labels $\mathcal{L}$, and the corresponding weight vector $\w_u$. Also, we have that $\w_p\cdot \psi_p(\x_i,\x_j,y_i,y_j)$ assigns pairwise “compatibility" scores of assigning labels $y_i$ to $i$ and $y_j$ to neighboring supervoxel $j$, where $\bpsi_p(\x_i,y_i,\x_j,y_j)$ is an $n_p$-dimensional *pairwise feature vector* and $\w_p$ as the associated weight vector. We specify $n_u = L = |\mathcal{L}|$ and $n_p=L^2$, but more general specifications are also possible.
We consider a common form of unary features defined in terms of a probabilistic multi-class classifier. In particular, we will learn a probabilistic classifier $H$ that returns an $L$-dimensional probability vector, such that $H(i)$ is a predicted distribution over all labels $y_i\in\mathcal{L}$ for supervoxel $i$. The input to $H$, used as a basis for prediction, is the concatenation of multiple descriptors of supervoxel $i$. In this paper, we use logistic regression for $H$. Given $H$, the unary feature vector $\bpsi_u(\x_i,y_i)$ is constructed in the standard way by setting all elements of $\bpsi_u(\x_i,y_i)$ to zero, except for the element corresponding to label $y_i$ which is set to the probability $H(i)$ of predicting $y_i$ for $i$. Thus, the cost of $\bpsi_u(\x_i,y_i)$ is dominated by the cost of computing $H(i)$.
The pairwise features is specified in the standard way as the difference between descriptors of the pair of neighboring supervoxels. Thus, the pairwise “compatibility" score is defined to increase as this difference becomes smaller for the supervoxels with the same label, and to decrease for small descriptor differences when the pair of supervoxels have different labels. Non-overlapping features are replaced by the expected features conditioned on the label.
[**CRF inference**]{} involves two main steps. First, unary and pairwise potentials are computed for all combinations of supervoxels and labels. Second, a standard approximate CRF inference procedure is applied, such as belief propagation or $\alpha$-expansion (used in our experiments), which returns a high scoring (ideally optimal) label assignment. The overall computational cost is, thus, the total time for computing the potentials and running the inference procedure.
[**Learning the CRF and $H$.**]{} Given labeled training data, we learn our CRF model, $\mathcal{M}$, using the unary and pairwise feature vectors computed over all training supervoxels. This is done by computing all descriptors for all supervoxels in the training data, and then using a standard CRF library to obtain $\mathcal{M}$ and the associated logistic regression $H$.
Recall, that the main idea of our framework is to reduce time by only computing a subset of descriptors for some supervoxels. Thus, in order to produce the unary feature vector, $H$ must be able to make label predictions for a supervoxel using any subset of its descriptors. One approach to obtaining such predictions would be to train $H$ on all descriptors, and then use one of a number of common strategies for making predictions with “missing information" (e.g., replacing the missing descriptor values with zero or expected values).
Rather, we take a more brute force approach, with the advantage of not requiring any method for handling missing descriptors. Instead of just training a single classifier based on all descriptors, we train a [*collection of classifiers*]{}, one classifier for each possible subset of descriptors. $H$ is then represented by this collection of classifiers. That is, each logistic regression is trained by removing all descriptor information from the training set other than its assigned descriptor subset. When $H$ is asked to make a prediction for a supervoxel during budgeted inference, it uses the classifier corresponding to the set of descriptors that have been run for that supervoxel. Given that classifier training is very fast, this will often be practical even with thousands of possible subsets. This is indeed the case in our experiments as described in Sec. \[sec:results\].
Budgeted Unary Feature Computation {#sec:budgeted}
----------------------------------
Computing the unary features $\bpsi_u$ typically dominates overall computation time, since this involves computing a number of low-level descriptors over all supervoxels. The pairwise features $\bpsi_p$ are much cheaper, in comparison, since they are generally based on comparing descriptors already computed for $\bpsi_u$. For this reason, our framework focuses on bounding the time of computing $\bpsi_u$, and we will let $B$ denote this time bound for the remainder of this paper. The overall CRF inference will typically be a small constant larger than $B$, when $B$ is non-trivial.
Our hypothesis is that similar accuracies in inference can be achieved with less cost by intelligently selecting for each supervoxel a sparse set of descriptors to compute, including the empty set, which are then used by $H$ to generate $\bpsi_u$. Below, we describe how to make these decisions.
Alg. \[alg:inference\] presents our iterative approach to selecting descriptor subsets for time bounded inference. Throughout the iterations, we maintain a set of *candidate supervoxels*, $\mathcal{C}$, which are currently being considered for descriptor computation. $\mathcal{C}$ is initialized to a small random subset of $V$. We also maintain a set of *finished supervoxels*, $\mathcal{F}$, which will no longer be considered for descriptor computation. Each iteration consists of two steps. The first step calls the function $\mbox{Select}(\mathcal{C})$, which returns a supervoxel $i \in \mathcal{C}$ to be considered next. As described in Sec. \[sec:results\], we consider two versions of $\mbox{Select}(\mathcal{C})$: a) Random selection, and b) Priority-based selection. The second step applies *policy* $\pi$ for selecting either a new descriptor for $i$ to compute, or mark $i$ as being finished. Specifically, $\pi(i)$ returns an *action* for $i$ that is either a descriptor index or [Finished]{}. In the latter case, $i$ is moved from $\mathcal{C}$ to $\mathcal{F}$, and all neighbors of $i$ that are not already in $\mathcal{F}$ are added to $\mathcal{C}$. The iterations continue until the runtime reaches $B$. When $B$ is reached, the CRF unary features $\bpsi_u(\x_i,y_i)$ are computed for all supervoxels $i$ where at least one descriptor has been computed. For a supervoxel $i$, where no descriptors are computed, $\bpsi_u(\x_i,y_i)$ is estimated based on the available unary features $\{\bpsi_u(\x_j,y_j)\}$ of $i$’s neighbors, i.e. $\bpsi_u(\x_i,y_i)$ is set to a weighted sum of $\{\bpsi_u(\x_j,y_j)\}$, where the weights are the inverse Euclidean distances between the centroids of $i$ and the neighbors.
$\mathcal{C} \leftarrow \mbox{small random subset of } V$; $\mathcal{F} \leftarrow \emptyset$ Interpolate unary features for supervoxels $i$ with no computed descriptors (see text)\
Apply CRF inference
The key element in the above framework is the policy $\pi$. Given a supervoxel $i$, $\pi$ must weigh the cost of computing a new descriptor of $i$ versus the potential improvement in accuracy of the CRF inference. It is critical that $\pi$ makes these decisions efficiently, otherwise the cost of evaluating $\pi$ would negate any potential reduction in descriptor computation that it provides. As our key contributions, in the following Sec. \[sec:learn\], we specify a suitable representation and learning algorithm for such a $\pi$.
Policy Representation and Learning {#sec:learn}
==================================
This section first describes the linear representation we use for policies, and then formulates our policy learning algorithm within the MDP framework.
Policy Representation {#sec:representation}
---------------------
At each iteration of our budgeted inference, $\pi$ is shown a supervoxel $i$, and asked to decide whether or not to run a new descriptor for $i$ and if so which one. We call these choices actions. $\pi$ selects an action $a$ based on the information available at the time, which we call the *inference state* $s$, $a=\pi(s)$. An inference state is a tuple $s=(i,b,\mathcal{C},\mathcal{F},\mathcal{D})$, where $i$ is the supervoxel currently being considered, $b$ is the remaining budget, and $\mathcal{C}$ and $\mathcal{F}$ are the sets of candidate and finished supervoxels as described in Sec. \[sec:budgeted\]. Finally, $\mathcal{D}$ is the set of descriptor outputs that have been produced so far for supervoxels in $\mathcal{C}$ and $\mathcal{F}$.
Since $\pi$ is called many times during inference, it is critical that the time required to select an action be significantly smaller than the time required to run descriptors. To support efficiency we represent $\pi$ as a linear function that ranks the possible actions at an inference state $s$ based on an easy to compute vector of $\pi$-features $\bphi(s)$.
[**Policy Features.**]{} The $\pi$-features have three subvectors $\bphi(s)=[\bphi_1(s),\bphi_2(s),\bphi_3(s)]$ that capture different aspects of the inference state $s$. $\bphi_1(s)$ is a binary vector that indicates which descriptors have already been run for $i$. $\bphi_1(s)$ allows the policy to learn the value of taking certain actions, given various combinations of computed descriptors characterizing $s$. $\bphi_2(s)$ is a vector that is equal to a weighted average of the unary potential features $\{\bpsi_u\}$ of “finished" neighbors of $i$ that are in $\mathcal{F}$. Recall that $\bpsi_u$ corresponds to probability distributions over class labels. Thus, $\bphi_2(s)$ allows the policy to base its decisions on the confidence of neighboring supervoxels about the various semantic labels. For example, the policy can learn that if all neighbors are very confident about a particular label then it is not worth computing further descriptors for $i$. Finally, $\bphi_3(s)$ is the standard shape-context descriptor capturing the spatiotemporal layout of finished supervoxels in $\mathcal{F}$ around $i$. $\bphi_3(s)$ is computed by binning the space-time neighborhood of $i$ (all supervoxels that touch $i$) into 8 bins ($\{up, down, left, right\} \times \{before, after\}$), and counting the finished supervoxels that fall in each bin. $\bphi_3(s)$ allows the policy to base its decisions in part on the density of surrounding finished supervoxels.
Given the $\pi$ features for an inference state $s$, $\bphi(s)$, our policy is a linear ranking function over policy actions $a$, represented in terms of a weight vector $\w_a$ for each action. $$\pi(s)=\arg\max_a ~\w_a\cdot \bphi(s)$$ Thus, training the policy entails training the weights $\w_a$ so as to make the best possible decisions.
Policy Learning {#sec:pi-learning}
---------------
It is important to note that ground-truth annotations of our training videos do not directly provide ground-truth decisions that should be made by $\pi$. The supervised training data for $\pi$ would need to label inference states by best policy actions. Since this information is not available in training, $\pi$ cannot be learned via pure supervised learning. The training data does, however, provide the means for evaluating the quality of any policy. In particular, given any $\pi$ and budget $B$, we can run time-bounded inference on each training video using $\pi$, as summarized in Alg. \[alg:inference\], and then measure the prediction accuracy of the CRF relative to the available ground truth. In practice, large numbers of such policy evaluations can be run quickly by precomputing all descriptors for all supervoxels across the training videos. This allows for the budgeted inference process to be “simulated" on the training data without requiring descriptors to be recomputed for each policy evaluation. The question then is how to use this fast policy evaluation on the training data in order to learn an effective policy?
Policy learning is complicated by the fact that the policy is inherently solving a sequential decision making problem, where each decision may have long-term impacts on the overall solution accuracy. Optimizing the long-term value of policies is challenging due to the fact that each inference process will involve many policy decisions and assigning relative credit to those decisions toward the overall accuracy is non-trivial. Such sequential decision making problems are naturally formalized in the MDP framework [@MDP_1994].
[**MDP Formulation:**]{} An MDP specifies a set of states $S$, a set of actions $A$ that can be taken by a policy, and a transition function $T$, which describes how the state of the system changes when actions are taken. In addition, an MDP specifies a reward function, which evaluates the relative goodness of various system states. In our time-bounded inference application, the states correspond to inference states $(i,b,\mathcal{C},\mathcal{F},\mathcal{D})$ as described in Sec. \[sec:representation\]. The actions $A$ correspond to policy actions of either selecting to run a descriptor for the current supervoxel $i$, or returning [Finished]{}, which indicates that we are finished computing descriptors for $i$.
The transition function describes how an action $a$ changes a state $s=(i,b,\mathcal{C},\mathcal{F},\mathcal{D})$. If the action $a$ is to run a descriptor, then the new state is equal to $s'=(i',b',\mathcal{C},\mathcal{F},\mathcal{D}')$, where $\mathcal{D}'$ updates $\mathcal{D}$ with the newly computed descriptor information for $i$. In addition, the new budget $b'$ will be equal to $b$ minus the cost of $a$, and $i'\in \mathcal{C}$ will be the supervoxel selected next for processing. When the action is [Finished]{}, the new state is $s'=(i',b,\mathcal{C}',\mathcal{F}',\mathcal{D})$, where $\mathcal{C}'$ and $\mathcal{F}'$ are updated as described in Sec. \[sec:budgeted\], and $i'$ is the newly selected supervoxel. Note that when $b=0$, no further actions are allowed. Importantly, the reward function is zero for all states, except for final states with $b=0$, where the reward is equal to the accuracy achieved by the CRF inference using the selected descriptors run by the $\pi$.
Given the above MDP formulation, the problem of optimizing $\pi$ to maximize expected long-term reward in the MDP is identical to finding a policy that maximizes inference accuracy within our budgeted inference framework. Thus, in principle, any policy learning algorithm from the MDP literature could be employed for our problems. Prior work [@NIPS2013_5142] used reinforcement learning (RL) for learning an approximate Q-function. In that work, the Q-function $Q(s,a)$ of an MDP gives the expected future reward of being in state $s$ and taking action $a$. Given $Q(s,a)$ the policy is defined to select the action with largest Q-value. Unfortunately, the Q-function can be extremely complicated to represent for problems that involve long sequences of decisions. In that prior work, the number of decisions was bounded by the number of descriptors, which is quite small. Rather, in this paper, the number of decisions is related to the number of supervoxels, which is substantially larger. Our early experiments showed that standard approaches for learning Q-functions, represented using our $\pi$-features, were ineffective for the problem scales we address here.
[**Classication-Based Approximate Policy Iteration (CAPI):**]{} Our approach is motivated by the fact that we do not actually need to learn the Q-function, but only learn to rank good actions above bad actions, e.g., using the linear policy representation described in Sec. \[sec:representation\]. This suggests considering approaches that directly learn the decision boundary between good and bad actions. Classification-Based Approximate Policy Iteration (CAPI) is one such state-of-the-art technique that we follow here.
CAPI was originally proposed by [@Fern03; @Lagoudakis03]. It has demonstrated a number of empirical successes, and has been the subject of theoretical analysis providing various performance guarantees [@Fern06; @Dimitrakakis08; @Lazaric10]. A key distinction of CAPI is that it is able to leverage state-of-the-art learning algorithms for classification and ranking (e.g. SVMs).
CAPI is conceptually simple. Given an initial policy $\pi_0$, which is random in our experiments, CAPI iteratively applies an *approximate policy improvement operator* $\mathcal{PI}$, which takes an input policy $\pi$ and returns an (approximately) improved policy $\mathcal{PI}[\pi]$. Thus, the CAPI algorithm produces a sequence of improving policies $\pi_{t+1} = \mathcal{PI}(\pi_t)$ and terminates when no further improvement is observed or a training time bound is reached. Recall that for our linear policy representation this will correspond to a sequence of weight vectors. It remains to describe the approximate policy improvement operator $\mathcal{PI}$.
Given a current policy $\pi$, CAPI computes the improved policy $\mathcal{PI}(\pi)$ using a two step process: [**Step 1 – Training Set Generation.**]{} Create a training set of state-action pairs $\mbox{Trn}=\{(s_i,a_i)\}$ such that $a_i$ is an “improved" action for $s_i$, i.e. better or at least as good as the current action $\pi(s_i)$. [**Step 2 – Classifier Learning.**]{} Apply a classifier learner to $\mbox{Trn}$ to obtain a policy $\pi'$ that achieves high accuracy in selecting the improved actions. In Step 2 we use a multi-class SVM classifier to learn the weights of $\pi$ based on the training set generated by Step 1.
Step 1 requires selecting a set of states for the training set and then computing the improved actions. In our experiments, we have found that an effective and simple approach is to run the current policy $\pi$ on the training videos, and to let the training states correspond to all inference states encountered during inference using $\pi$ across all training videos. For each such inference state $s_i$ we must now compute a label $a_i$ that corresponds to an improved action relative to the action selected by $\pi$.
In order to compute an improved action, CAPI uses the Monte Carlo simulation technique of *policy rollout* [@Tesauro96]. Fig. \[fig:policy\_rollout\] illustrates the main components of policy rollout. Simply stated, policy rollout computes a score for each action $a$ at a state $s_i$ that is equal to the accuracy achieved by the final CRF inference after taking action $a$ in $s_i$ and then taking actions according to $\pi$ until the budget is zero. The action leading to the highest accuracy is then selected. That is, rollout considers all one-step action departures from the current policy $\pi$ at $s_i$ and selects the action that resulted in the highest final CRF accuracy. Note that the transition function for actions may be stochastic, e.g., when the selection of the next supervoxel is implemented by random selection. In these cases, policy rollout runs multiple simulations for each action and the average accuracies across simulations are used to score actions.
For deterministic transition functions, policy rollout is guaranteed to return an improved action if the current policy is not optimal in $s_i$. For stochastic transitions, in the limit of infinite simulations, the action returned by rollout is guaranteed to be an improved action. A polynomial sample complexity bound exists on the quality of the action returned by rollout compared to that of $\pi$ [@Fern06]. Since our classifier is linear the learning time is linear in sample size. Note that while the training process may require significant time, the end result is a single policy $\pi$ from the last iteration of CAPI that can be used efficiently at test time for time-bounded inference.
![Monto Carlo simulation of policy rollout. All possible actions of a current state are evaluated by running the current policy $\pi$ starting from the next states corresponding to the actions being taken, until the time budget is reached. The CRF inference is then applied to compute the Hamming loss. This is used to improve the current policy.[]{data-label="fig:policy_rollout"}](policy_rollout){width="\columnwidth"}
Results {#sec:results}
=======
Our goal is to empirically support the claim that we can optimally adapt a given method for semantic video segmentation to varying time budgets, such that it yields satisfactory performance for any budget, and maintains an accuracy as close as possible to its performance for no time bound. As the given method was originally designed to perform best without time constraints, it is important to note that our performance is inherently upper-bounded by the method’s accuracy for an infinite time budget. Therefore, our evaluation differs from much work in computer vision, where the focus is on demonstrating improvements in accuracy.
**Datasets.** For evaluation, we use three benchmark datasets: 1) CamVid [@brostow2008segmentation], 2) MPIScene [@wojek2010monocular], and 3) SUNY Buffalo-Xiph.org 24-class [@chen2010propagating]. CamVid consists of 5 videos with an average length of 5000 frames. The videos are captured with a moving camera recording road scenes. Following prior work, we focus on the 11 most common object class labels and use the standard split of training and test frames as in [@brostow2008segmentation; @liumulti]. MPIScene consists of 4 videos with an average length of 150 frames. The videos show driving scenes recorded from a car. Almost 25% of the frames are labeled with 5 object classes. For MPIScene, we use the split of 1/2 training and 1/2 test frames as in [@wojek2008dynamic]. The SUNY Buffalo dataset consists of 10 videos of diverse scenes with an average length of 80 frames. They are fully annotated with 24 class labels. As in [@jain2013coarse], we use one-half of the frames for training and the other half for testing.
**Supervoxels and Descriptors.** All videos are partitioned into supervoxels using the hierarchical graph-based approach of [@grundmann2010efficient], and its software implementation presented in [@xu2013flattening]. For CamVid, the video partitioning is based on the 9th level of the generated supervoxel tree, producing on average 3500 supervoxels per video. For, MPIScene, we used the 9th level of the supervoxel hierarchy, producing on average 1000 supervoxels per video. In this paper, we do not consider the computation time of extracting supervoxels.
Each supervoxel has access to algorithms for computing appearance and motion descriptors, as controlled by our inference policy. These descriptors include the following. We use dense trajectories [@wang:2011] by tracking a set of densely sampled points in two different scales on a grid. The descriptor extraction finds the tightest cube of a given supervoxel, and uses the dense trajectories to generate HOG, HOF, and MBH (motion boundary histogram) for each track. We also use the color histogram in CIE-Lab color space for each supervoxel. In addition, we also use object detectors as mid-level descriptors of supervoxels to identify the probability of observing a corresponding object class in a supervoxel. Specifically, given a supervoxel, we run a Deep Convolutional Neural Net (DCNN) on all pixels of the supervoxel’s first, middle and last frame. Then, the output of the logistic regression layer of DCNN is used as a descriptor of the superpixel. For CamVid, computing HOG, HOF and MBH for the entire video takes 6-8, 15-17, 9-11 minutes, respectively; computing the color histogram for all supervoxels in a video takes 3-5 minutes; and running DCNN takes 0.1 seconds per supervoxel. On average the full descriptor extraction for an entire video in CamVid requires 43-53 minutes. For MPIScene, computing HOG, HOF and MBH for an entire video takes 10-12, 18-20, 11-12 seconds, respectively; computing the color histogram for all supervoxels in a video takes 2-4 seconds; and running DCNN takes 0.1 seconds per supervoxel. On average the full feature extraction for an entire video in MPIScene requires 42-50 seconds. For SUNY Buffalo-Xiph.org 24-class computing HOG, HOF and MBH for the entire video takes 5-7, 8-10, 5-6 seconds respectively. Computing the color histogram over all supervoxels in a video takes 0.6-1 second. On average, full feature extraction for the entire video from the SUNY Buffalo dataset requires 19-22 seconds.
**Variations of Our Framework:** We evaluate our framework using different supervoxel selection strategies (Sec. \[sec:budgeted\]) and different sets of available descriptors. We consider three variations: 1) *-RndRnk* randomly selects a supervoxel from $\mathcal{C}$, and uses only HOG, HOF, MBH and color histogram; 2) *-NhbRnk* ranks the supervoxels based on the confidence of $H$ classifiers for neighbors in $\mathcal{F}$ (Sec. \[sec:budgeted\]), and uses only HOG, HOF, MBH and color histogram; 3) *-Full* is similar to -NhbRnk, but additionally uses DCNN-based descriptor.
**Upper-Bound Performance:** For evaluating our upper-bound performance, we compare our time-bounded accuracies to two unbounded CRF models. The first model, referred to as CRF, uses only HOG, HOF, MBH and color histogram descriptors. The second model, referred to as CRF-Full, additionally uses DCNN-based descriptor.
**Baselines:** We specify a number of reasonable baseline approaches. This comparison serves to evaluate how our proposed learning of the inference policy affects performance relative to alternative strategies. 1) *Baseline1* randomly selects a sequence of descriptors to be computed for all supervoxels until time runs out; 2) *Baseline2* randomly selects a subset of supervoxels, such that there is time for computing all descriptors for each supervoxel in the subset, and 3) *Baseline3* is aimed to recreate the related approach of [@NIPS2013_5142] in our domain, i.e., *Baseline3* learns an inference policy using Q-learning for selecting a sparse set of descriptors that are computed for all supervoxels in the entire video.
**Implementation:** is done in C++. Darwin library(http://drwn.anu.edu.au/) is used for training the CRF and $H$. The $\alpha$-expansion algorithm [@a22001fast] is used for CRF inference. We perform our experiments on an Intel quad core-i7 CPU and 16GB RAM PC. We use Caffe deep learning framework [@jia2014caffe] for our DCNN implementation. Fine tuning of the DCNN parameters based on the Alexnet model is done for each dataset. The training set for the objects is obtained as in [@liumulti].
Table \[tab:results\] compares per-class and average video labeling accuracy of the aforementioned variations of our framework with those of time-unconstrained CRF models and baselines, for three different time budgets, on CamVid. As can be seen, for the smallest budget, the accuracy of all baselines is much worse than that of all our framework variations. We observed that Baseline1 and Baseline2 were not able to use HOF and MBH, for this budget, because the cost of computing motion descriptors for the entire video was above the budget. This demonstrates the importance of our intelligent descriptor selection.
In Table \[tab:results\], we also see that all variations of our framework are able to continually improve accuracy as the time budget increases. At the largest time budget of 45 minutes, -Full achieves nearly the same accuracy as the unbounded CRF-Full which in turn requires 50-55min for computation of all descriptors in the entire video. This demonstrates that we are able to maintain a similar level of accuracy of the original method under reduced runtimes, namely for a 5 min time reduction.
Interestingly, the results in Table \[tab:results\] show that for lower budgets the variations of our framework give superior accuracy for the dominant class labels relative to unbounded inference. This is likely due to our explicit capturing of the contextual information from neighboring supervoxels. The results suggest that our policy successfully learned to avoid computing redundant descriptors when the neighbors can provide strong evidence of the label. Thus, the main impact of increasing the budget is to improve accuracy on the non-dominant labels.
Tables \[tab:sunny-mpi\] show the per-class and average video labeling accuracy on MPI-Scene and SUNY Buffalo-Xiph.org 24-class dataset.
\[\]\[c\][ ]{}
\[\]\[c\][ ]{}
Fig. \[fig:plot\](top) how decisions of our policy, learned in *-Full*, differ across various budgets for CamVid. Specifically, the figure shows the histogram of certain types of descriptors selected for different budgets. We see that HOF and MBH are seldom used for small budgets, as expected, since they incur higher costs. Also, the color histogram descriptor is more frequently selected when we are given small budgets, as expected, since they incur small costs.
Fig. \[fig:plot\](bottom) shows how the distribution of descriptor selection changes in time during our budgeted inference by *-NhbRnk* for CamVid. Specifically, the figure shows the histogram of descriptor selections made by the policy at various moments in time of the inference process. We see that the distribution changes as the inference time increases until the budget is reached. For example, initially, the distribution is highly skewed toward selections of the cheap color descriptor but then becomes more uniform. This suggests that the policy successfully learned to select low-cost descriptors initially for facilitating later policy decisions.
Conclusion
==========
We formulated the new problem of budgeted semantic video segmentation, where the pixels of a video must be semantically labeled under a time budget. We presented a budgeted inference framework for this problem that intelligently selects supervoxel descriptors to run, which are then used for CRF inference. Since descriptor computation often dominates the cost of CRF inference, our framework can provide substantial time savings in a principled manner. We formulated the inference policy for selecting among descriptors to run for each supervoxel in a video. We introduced a principled algorithm for learning such policies based on labeled training videos by formulating budgeted inference in the framework of an MDP. Our experiments show that we are able to learn policies for budgeted inference that significantly improve on the accuracies of several baselines. The results also demonstrate that we can optimally adapt a method, design to operate with no time bound, to varying time budgets, such that it yields satisfactory performance for any budget, and maintains an accuracy as close as possible to its performance for no time bound.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we propose an autonomous UAV path planning framework using deep reinforcement learning approach. The objective is to employ a self-trained UAV as a flying mobile unit to reach spatially distributed moving or static targets in a given three dimensional urban area. In this approach, a Deep Deterministic Policy Gradient (DDPG) with continuous action space is designed to train the UAV to navigate through or over the obstacles to reach its assigned target. A customized reward function is developed to minimize the distance separating the UAV and its destination while penalizing collisions. Numerical simulations investigate the behavior of the UAV in learning the environment and autonomously determining trajectories for different selected scenarios.'
author:
- '\'
title: 'Autonomous UAV Navigation: A DDPG-based Deep Reinforcement Learning Approach'
---
Autonomous navigation, deep reinforcement learning, obstacle avoidance, unmanned aerial vehicle.
Introduction {#Introduction}
============
Smart cities are witnessing a rapid development to provide satisfactory quality of life to its citizens [@7539244]. The establishment of such cities requires the integration and use of novel and emerging technologies. In this context, unmanned areal vehicles (UAV), *aka* drones, are continuously proving their efficiency in leveraging multiple services in several fields, such as good delivery and traffic monitoring (e.g. Amazon is starting to use UAVs to deliver packages to customers). UAVs are easy to deploy with a three dimensional (3D) mobility as well as a flexibility in performing difficult and remotely located tasks while providing bird-eye view [@8613833; @8731861]. Path planning remains one of key challenges that need to be solved to improve UAV navigation especially in urban areas. The core idea is to devise optimal or near-optimal collision-free path planning solutions to guide UAVs to reach a given target, while taking into consideration the environment and obstacle constraints in the area of interest.
In recent studies, such as [@8359869], the authors adopted the ant colony optimization algorithm to determine routes for UAVs while considering obstacle avoidance for modern air defence syste. In [@8833455], a combination of grey wolf optimization and fruit fly optimization algorithms is proposed for the path planning of UAV in oilfield environment. In [@7978637; @8747351; @xian], the UAV path planning problems were modeled as mixed integer linear programs (MILP) problem. Also, in [@7829099], a 3D path planning method for multi-UAVs system or single UAV is proposed to find a safe and collision-free trajectory in an environment containing obstacles. However, most of the solutions are based on MILP which are computationally complex or evolutionary algorithms, which do not necessarily reach near-optimal solutions. Moreover, the existing approaches remain centralized where a central node, e.g. a control center runs the algorithm and provides to the UAV its path plan. Centralized approaches restrain the system and limit its capabilities to deal with real-time problems. They impose a certain level of dependency and cost additional communication overhead between the central node and the flying unit. Hence, artificial intelligence (AI), precisely, reinforcement learning (RL) come out as a new research tendency that can grant the flying units sufficient intelligence to make local decisions to accomplish necessary tasks. In [@8443226] and [@leme], the authors presented a Q-learning algorithm to solve the autonomous navigation problem of UAVs. Q-learning was also employed to establish paths while avoiding obstacles in [@8742915]. However, the authors used discrete actions (i.e. the environment is modeled as a grid world with limited UAV action space, degree of freedom). which may reduce the UAV efficiency while dealing with real-world environment, where the flying units operate according to a continuous action space.
![Illustration of the autonomous obstacle-aware UAV navigation in an urban environment.[]{data-label="system"}](model.PNG){width="7cm"}
Unlike existing RL-based solutions which are usually operating on a discretized environment, the proposed framework aims to provide UAV autonomous navigation with continuous action space to reach fixed or moving targets dispersed within a 3D space area while considering the UAV safety. A deep deterministic gradient decent (DDPG)-based approach is modeled with the objective to allow an UAV determine the best course to accomplish its missions safely, i.e. obstacle avoidance. A reward function is designed to guide the UAV toward its destination while penalizing any crash. During the training phase, we adopt a transfer learning approach to train the UAV how to reach its destination in a free-space environment (i.e., source task). Then, the learned model is fed to other models (i.e., new task) dedicated to different environments with specific obstacles’ locations so that the UAV can learn how to avoid obstacles to navigate to the destination. During the prediction phase, it determines the path within the training environment by figuring out which route to take to reach any randomly generated static or dynamic destination from any arbitrary starting position. In the simulations, we investigate the behavior of the autonomous UAVs for different scenarios including obstacle-free and urban environments. It is shown that the UAV smartly selects paths to reach its target while avoiding obstacles either by crossing over or deviating them.
![Architecture of the actor-critic learning model.[]{data-label="alg_fig"}](acric.PNG){width="8cm"}
Problem Description {#SystemModel}
===================
System Model
------------
Autonomous navigation for UAVs in real environment is complex. Hence, Without loss of generality, we create a virtual 3D environment with high matching degree to the real-world urban areas. Unlike most of the existing virtual environments, which are studied in literature and usually modeled as a grid world, in this paper, we focus on a free space environment containing 3D obstacles that may have diverse shapes as illustrated in Fig. \[system\]. Consequently, the UAV has the freedom to take any direction and speed to reach its target unlike grid world, which restricts the freedom of UAV into a finite set of actions. The goal is to train the UAV to fly safely from any arbitrary starting position to reach any destination in the considered area with continuous action space. The UAV, defined as $u$, is characterized by its 3D Cartesian geographical location $loc_u = [x, y, z]$ and initially situated at $loc_u(0) = [x_0, y_0, z_0]$. The destination $d$ is defined by its 3D location $loc_{d} = [x_{d}, y_{d}, z_{d}]$. In this paper, the investigated system assumes the following assumptions:
- The environment obstacles have different heights. Each one of them is represented by a 3D polygon characterized by its the starting point \[$x_{obs}$, $y_{obs}$\], the set containing the edges of the base $edg_{obs}$, and its height $h_{obs}$. Hence, if having an altitude higher than the obstacle’s height, the UAV can cross over the obstacles. Otherwise, the UAV can avoid it by flying around.
- The destination location is known to the UAV and it can be either static or dynamic (i.e., the target location can evolve over time). If the destination location is dynamic then it follows a random pre-defined trajectory, that is unknown by the UAV.
UAV Actions
-----------
For each taken action, we assume that the UAV chooses a distance to cross according to a certain direction in the 3D space during $\Delta t$ units of time. The action is modeled using the spherical coordinates $(\rho, \phi, \psi)$ as follows: $$\begin{aligned}
&x_u = x_u + \rho \sin{\phi} \cos{\psi}, \quad y_u = y_u + \rho \sin{\phi},\notag\\
&\text{and }z_u = z_u + \rho \cos{\phi},\end{aligned}$$ where $\rho$ is the traveled radial distance by the UAV in each step $\left(\rho \in [\rho_{min}, \rho_{max}]\right)$, where $\rho_{max}$ is the maximum distance that the UAV can cross during the step length $\Delta t$. Its value depends on the maximum speed of the UAV denoted by $v_{max}$. The parameter $\psi$ denotes the inclination angle $\left(\psi \in [0, 2\pi]\right)$, and $\phi$ represents the elevation angle $\left(\phi \in [0,\pi]\right)$. For instance:\
$\bullet$ if $\rho = \rho_{max}$, $\phi = \pi$, and any value of $\psi$, the UAV moves by $\rho_{max}$ along the Z axis.\
$\bullet$ if $\rho = \rho_{max}$, $\phi = \pi/2$, and $\psi=0$, the UAV moves along the x axis.\
The distance between the UAV and its target is defined as $D(u,d)$.
![Example representing a collision scenario. The UAV’s altitude is less than the obstacle’s height (obs1). The action is chosen such that the UAV crosses the obstacle. This results in an obstacle penalty reflecting the underwent depth.[]{data-label="depth"}](crashdepth.PNG){width="8cm"}
DDPG Learning for Autonomous Scheduling {#ProblemFormulation}
=======================================
DDPG was developed as an extension of deep Q-network (DQN) algorithms introduced by Mnih et al. [@mnih], which was the first approach combining deep and reinforcement learning but only by handling low-dimensional action spaces. DDPG is also a deep RL algorithm, that has the capability to deal with large-dimensional/infinite action spaces. It tries to find an efficient behavior strategy for the agent to obtain maximal rewards in order to accomplish its assigned tasks [@Lillicrap]. This DPG algorithm has the capability to operate over continuous action spaces which is a major hurdle for classic RL methods like Q-learning.
DDPG is based on the actor-critic algorithm. It is essentially a hybrid method that combines the policy gradient and the value function together. The policy function $\mu$ is known as the actor, while the value function $Q$ is referred to as the critic. Essentially, the actor output is an action chosen from a continuous action space, given the current state of the environment $a = \mu(s|\theta^{\mu})$, which, in our case, has the form of a tuple $a=[\rho,\phi,\psi]$. As for the critic, its output $Q(s, a|\theta^{\mu})$ is a signal having form of a Temporal Difference (TD) error to criticize the actions made by the actor knowing the current state of the environment. A diagram summarizing the actor-critic architecture is given in Fig. \[alg\_fig\].
Note that the training phase of the DDPG model is executed for $M$ episodes where each one of them accounts for $T$ steps. We use the index $t$ to denote an iteration within a single episode where $t=1,\dots, T$. The actor and critic are designed with neural networks. The value network is updated based on Bellman equation [@Murat] by minimizing the mean-squared loss between the updated Q value and the origin value, which can be formulated as shown in Algorithm 1 (line 11). As for the policy network’s update (line 13), it is based on the deterministic policy gradient theorem [@ddpg].
There are also some practical tricks that are used to enhance the performance of the framework. A trade off between exploration and exploitation is made by the use of $\epsilon$-greedy algorithm, where a random action $a_t$ is selecting with $\epsilon$ probability, otherwise a precise action $a_t=\mu(s_t|\theta^{\mu})$ is selected according to the current policy with a $1-\epsilon$ probability. Furthermore, an experience replay buffer $b$, with size $B$, is used during the training phase to break the temporal correlations. Each interaction with the environment is stored as tuples in the form of $[s_{t}, a, r, s_{t+1}]$, which are the current state, the action to take, the reward of performing action $a$ at state $s_t$, and the next state, respectively (Algorithm 1 (line 9)) and, during the learning phase, a randomly extracted set of data from the buffer is used (Algorithm 1 (line 10)). Also, target networks are exploited to avoid the divergence of the learning algorithm caused by the direct updates of the networks weights with the gradients obtained from the TD error signal.
Randomly initialize critic $Q(s, a|\theta^{\mu})$ and actor $\mu(s|\theta^{\mu})$ neural networks with weights $\theta^Q$ and $\theta^\mu$. Initialize target networks $Q'$ and $\mu'$ with weights $\theta^{Q'} \gets \theta^Q$, $\theta^{\mu'} \gets \theta^\mu $. Initialize replay buffer $b$. Receive first observation $s_1$. Select $a_t$ based on $\epsilon$-greedy algorithm: select random action $a_t$ with $\epsilon$ probability, otherwise $a_t=\mu(s_t|\theta^{\mu})$ according to the current policy. Execute action $a_t$ and observe reward $r_t$ and new state $s_{t+1}$. Store transition $[s_t, a_t, r_t, s_{t+1}]$ in $b$. Sample a random batch of $N$ transitions $[s_j, a_j, r_j, s_{j+1}]$. Set $\hspace{0.1cm}y_j = r_j + \gamma Q'(s_{j+1}, \mu'(s_{j+1}|\theta^{\mu'} )|\theta^{Q'})$. Update critic by minimizing the loss:
$\hspace{1cm}L = \frac{1}{N}\sum_{j}(y_j - Q(s_j, a_j|\theta^Q))^2$ Update the actor policy using policy gradient:
$\nabla_{\theta^\mu}\mu|_{s_j}\approx \frac{1}{N}\sum_j\nabla_a Q(s, a|\theta^Q)|_{s=s_j,a=\mu(s_j)}\nabla_{\theta^{\mu}}\mu(s|\theta^\mu)|_{s_j} $ Update the target networks:
$\hspace{1cm}\theta^{Q'} \gets \nu\theta^Q + (1-\nu)\theta^{Q'}$
$\hspace{1cm}\theta^{\mu'} \gets \nu\theta^\mu + (1-\nu)\theta^{\mu'}$
Reward Function
---------------
In an obstacle-constrained environment, the UAV must avoid obstacles and autonomously navigate to reach its destination in real-time. Therefore, the reward function, denoted by $f_r$, is modeled such that it encourages the UAV to reach its destination and, at the same time, penalizes it when crashing. Thus, the reward function is composed of two terms: target guidance reward and obstacle penalty. The target guidance reward, denoted by $f_{gui}$, is used to motivate the flying unit to reach its target as fast as possible, while the obstacle penalty, denoted by $f_{obp}\textbf{}$ is responsible for alerting the UAV to keep a certain safety distance off the obstacles. The reward function is formulated as follows:
$$\begin{aligned}
&\hspace{-0.2cm}f_{r}(D(u,d), \sigma) = (1-\beta) f_{gui}(D(u,d))+\beta f_{obp}(\sigma),\\
&\hspace{-0.2cm}f_{gui}(D(u,d)) = \exp(-5D(u,d)^2),\\
&\hspace{-0.2cm}f_{obp}(\sigma) = \exp(-100\sigma)-1, \end{aligned}$$
![Illustration of the transfer-learning technique.[]{data-label="tramsfer"}](transfer.PNG){width="8cm"}
![Path followed by UAV based on the dynamic target location. Yellow dot: initial UAV position, blue dashed line: trajectory of the UAV, blue dot: target initial location, and dashed black line: the target trajectory.[]{data-label="dyn"}](dynn.PNG){height="5cm" width="8cm"}
where $\sigma$ is the crash depth explained in Fig. \[depth\] and $\beta$ is a variable that regulates the balance between $f_{obp}$ and $f_{gui}$. The obstacle penalty is modeled as a function of the crash depth $\sigma$ to conserve the continuous nature of the reward function instead of using discrete penalty, which proved to be more efficient to help the model to converge. In fact, when the crash depth is high, the UAV receives a higher penalty, whereas a small crash depth results in a lower penalty. The use of this approach helps the UAV learn efficiently over the training episodes how to adjust its trajectory to avoid obstacles.
Training Phase and Transfer Learning
------------------------------------
Transfer learning is a machine learning technique used to transfer the knowledge to speed up training and improve the performance of deep learning models. The proposed approach to train the UAV consists in two steps. Initially, we train the model in an obstacle-free environment. Training in such environment, grants the UAV the capability to reach any target in the covered 3D area with continuous space action. Then, the trained model on the obstacle-free environment will serve as a base for future models trained on other environments with obstacles. Afterwards, we transfer the acquired knowledge (i.e. source task) and use it to improve the UAV learning of new tasks where it updates its path based on the obstacle locations while flying toward its target. The adopted transfer learning technique applied to DDPG for autonomous UAV navigation is illustrated in Fig. \[tramsfer\]. Once the training phase is completed offline, the UAV is capable to make instant decisions, while interacting with the environment, to manage real-time missions.
Simulation Results
==================
In this section, we study the behavior of the system for selected scenarios. We also visualize the efficiency of the framework in terms of crash rate and tasks accomplishment. To do so, we assume that the UAV starting location $loc_u$, its target location $loc_{d}$, and the obstacles’ parameters are randomly generated within a cube-shaped area with 100 $m$ edge length. We make sure that the locations of both the targets and the UAV are outside the obstacles. The rest of the simulation parameters are set as follows:
----------- ------- ------------------------------------- --------- --------------- ------- --- -----
$\beta$ 4 $\nu$ 0.99 $\rho_{\max}$ 0.2 m T 100
\[1ex\] M 40000 $\epsilon_{end}$,$\epsilon_{start}$ 0.1,0.9 B 10000 N 256
----------- ------- ------------------------------------- --------- --------------- ------- --- -----
![The reward received by the UAV during the training phase: (a) the source task (obstacle-free) and (b) the environment with obstacles []{data-label="last"}](last.PNG){width="9cm"}
The simulations are executed using Python. For the sake of clarity, the figures concerning the UAV path planning are presented in only 2D dimension area (i.e., plan(x,y)) and we provide beside each dot, the altitude of either the target or the UAV. In the first scenario, we consider an obstacle-free environment. The destination location is assumed to be dynamic, that it keeps moving in a randomly generated way. As shown in Fig. \[dyn\], the UAV is successfully adapting its trajectory based on the location of its target until it reaches it. This shows that the UAV succeeded in learning how to update each direction in order to “catch” its assigned destination.
In the next scenarios, the obstacles are added in a random disposition with different heights as shown in Fig. \[sys\]. In these cases, we assume that the target destinations are static. In Fig. \[sys\](a), the UAV successfully reached its destination location while avoiding the obstacles. In Fig. \[sys\](b), on its way to the destination, the UAV crossed over $obs2$ ($z_u=0.63 > h_{obs2}= 0.5$) in order to reach faster its target location unlike the case in Fig. \[sys\](a), where the UAV could not cross over $obs2$ to reach its destination as soon as possible because of the obstacle height (maximum height). In Fig. \[sys\](c), having a higher altitude than $obs6$, the UAV crossed over $obs6$ to reach its target. In all cases, scenarios show some lacking in precision to reach the target location due to the fact of using infinite action space which makes it hard to get pinpoint accuracy. These scenarios showed that the UAV successfully learned how to avoid obstacles to reach its destination.
In Fig. \[last\], we present the reward received by the UAV during its training phase, in Fig. \[last\](a) shows that the UAV learns to obtain the maximum reward value in an obstacle-free environment. We successfully obtained a trained model capable of reaching targets in 3D environment with continuous action space. Then, using the knowledge gathered by the first training, we trained the model to be able to avoid obstacles. Fig. \[last\](b) shows that the UAV model has converged and reached the maximum possible reward value.
During the testing phase and as shown in Table I, for the obstacle-free environment, the UAV successfully reached its target for the tested cases, 100% success rate for 1000 test case. As for the environment with obstacles, in the case of env1, the UAV successfully reached its target safely for 84% of the 1000 tested scenarios and in the case of env2, the reached its target safely for 82% of the 1000 tested scenarios.
[|c | c||c | c|]{}
Scenarios & obstacle free & obstacles env1 & obstacles env2\
\[0.5ex\] Completion rate & 100% & 84% & 82%\
\[1ex\]
\[cc\]
Conclusion
==========
In this paper, we have developed an efficient framework for autonomous obstacle-aware UAV navigation in urban areas. Using a DDPG-based deep reinforcement learning approach, the UAV determines its trajectory to reach its assigned static or dynamic destination within a continuous action space. A transfer learning approach is devised in order to maximize a reward function balancing between target guidance and obstacle penalty. The simulation results exhibit the capability of UAVs in learning from the surrounding environment to determine their trajectories in real-time.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
S. P. [Mohanty]{}, U. [Choppali]{}, and E. [Kougianos]{}, “Everything you wanted to know about smart cities: The internet of things is the backbone,” *IEEE Consumer Electronics Magazine*, vol. 5, no. 3, pp. 60–70, July. 2016.
M. B. [Ghorbel]{}, D. [Rodríguez-Duarte]{}, H. [Ghazzai]{}, M. J. [Hossain]{}, and H. [Menouar]{}, “Joint position and travel path optimization for energy efficient wireless data gathering using unmanned aerial vehicles,” *IEEE Transactions on Vehicular Technology*, vol. 68, no. 3, pp. 2165–2175, Mar. 2019.
H. [Ghazzai]{}, H. [Menouar]{}, A. [Kadri]{}, and Y. [Massoud]{}, “Future uav-based its: A comprehensive scheduling framework,” *IEEE Access*, vol. 7, pp. 75678–75695, June 2019.
J. [Chen]{}, F. [Ye]{}, and T. [Jiang]{}, “Path planning under obstacle-avoidance constraints based on ant colony optimization algorithm,” in *IEEE 17th International Conference on Communication Technology (ICCT’17)*, China, Oct. 2017.
F. [Ge]{}, K. [Li]{}, W. [Xu]{}, and Y. [Wang]{}, “Path planning of uav for oilfield inspection based on improved grey wolf optimization algorithm,” in *Chinese Control And Decision Conference (CCDC’19)*, China, June. 2019.
Z. [Zhang]{}, J. [Wang]{}, J. [Li]{}, and X. [Wang]{}, “Uav path planning based on receding horizon control with adaptive strategy,” in *29th Chinese Control And Decision Conference (CCDC’17), Chongqing, China*, May 2017.
A. [Bahabry]{}, X. [Wan]{}, H. [Ghazzai]{}, H. [Menouar]{}, G. [Vesonder]{}, and Y. [Massoud]{}, “Low-altitude navigation for multi-rotor drones in urban areas,” *IEEE Access*, vol. 7, pp. 87716–87731, June 2019.
A. [Bahabry]{}, X. [Wan]{}, H. [Ghazzai]{}, G. [Vesonder]{}, and Y. [Massoud]{}, “Collision-free navigation and efficient scheduling for fleet of multi-rotor drones in smart city,” *IEEE International Midwest Symposium on Circuits and Systems (MWSCAS’19), Dallas, TX, USA*, Aug. 2019.
, [S.]{}, [L. Shuandao]{}, and [W. Jiang]{}, “Path planning for uavs based on improved artificial potential field method through changing the repulsive potential function,” in *IEEE Chinese Guidance, Navigation and Control Conference (CGNCC’16)*, China, Aug. 2016.
C. [Yan]{} and X. [Xiang]{}, “A path planning algorithm for uav based on improved q-learning,” in *2nd International Conference on Robotics and Automation Sciences (ICRAS’18), China*, June 2018.
O. [Bouhamed]{}, H. [Ghazzai]{}, H. [Besbes]{}, and Y. [Massoud]{}, “Q-learning based routing scheduling for a multi-task autonomous agent,” *IEEE International Midwest Symposium on Circuits and Systems (MWSCAS’19), Dallas, TX, USA*, Aug. 2019.
V. N. [Sichkar]{}, “Reinforcement learning algorithms in global path planning for mobile robot,” in *International Conference on Industrial Engineering, Applications and Manufacturing (ICIEAM’19), Sochi, Russia.*, Mar 2019, pp. 1–5.
V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep reinforcement learning,” *Nature*, vol. 518, Feb. 2015.
T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning.” in *ICLR*, 2016.
R. Bellman, *Dynamic Programming*, ser. Dover Books on Computer Science.1em plus 0.5em minus 0.4emDover Publications, 2013.
T. Lillicrap, J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,” *CoRR*, 09 2015.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a family of self-consistent, spherical, lowered isothermal models, consisting of one or more mass components, with parameterized prescriptions for the energy truncation and for the amount of radially biased pressure anisotropy. The models are particularly suited to describe the phase-space density of stars in tidally limited, mass-segregated star clusters in all stages of their life-cycle. The models extend a family of isotropic, single-mass models by Gomez-Leyton and Velazquez, of which the well-known Woolley, King and Wilson (in the non-rotating and isotropic limit) models are members. We derive analytic expressions for the density and velocity dispersion components in terms of potential and radius, and introduce a fast model solver in (), that can be used for data fitting or for generating discrete samples.'
date: 'Accepted 2015 August 7. Received 2015 August 6; in original form 2015 June 26'
title: A family of lowered isothermal models
---
methods: analytical – methods: numerical – stars: kinematics and dynamics – globular clusters: general – open clusters and associations: general – galaxies: star clusters: general
Introduction {#Sect:Intro}
============
The evolution of globular clusters (GCs) is the result of an interplay between stellar astrophysics (stellar and binary evolution, stellar mergers, etc.), dynamical two-body relaxation and the interaction with the tidal field of their host galaxy [@2003gmbp.book.....H]. Despite this plethora of physical processes at work on their respective time-scales, the [*instantaneous*]{} surface brightness profiles and kinematics of GCs are well described by relatively simple distribution function (DF) based models that need very few assumptions .
The relative simplicity of GC properties is owing to the absence of gas and non-baryonic dark matter and the collisional nature of their evolution, which drives them to tractable properties, such as spherical symmetry, isotropy and (quasi-)equipartition between different mass species [e.g. @1987degc.book.....S]. Because the relaxation time-scale of GCs is much longer than their dynamical time, their instantaneous properties can be described by models that satisfy the collisionless Boltzmann equation [see e.g. Chapter 8 in @2014dyga.book.....B].
Two-body interactions in GCs evolve the velocity distribution of stars towards a Maxwell-Boltzmann distribution, at least in the core, where the relaxation time-scale is short. Models with isothermal cores are therefore a good choice for fitting properties of GCs. An obvious starting point for a discussion on model choice is, therefore, the isothermal model. This model has an infinite spatial extent and infinite mass [@1939isss.book.....C] and to make the model applicable to real star clusters, the assumption of the idealized Maxwell-Boltzmann distribution of velocities needs to be relaxed. This can be done by changing the model such that stars have a finite escape velocity. @1954MNRAS.114..191W developed such a model by simply ‘lowering’ the (specific) energy $E$ by a constant. The DF, which describes the density in six-dimensional phase-space as a function of $E$, is then simply $f(E) = A\exp[-(E-\phi(\rt))/s^2]$, for $E\le\phi(\rt)$, and $f(E)=0$ for $E>\phi(\rt)$. Here $s$ is a velocity scale, which in the isothermal model equals the one-dimensional velocity dispersion and $E$ is reduced by the specific potential at the truncation radius $\rt$, $\phi(\rt)$. This truncation in energy mimics the role of tides due to the host galaxy, which makes it easier for stars to escape by reducing the escape velocity. The resulting models are nearly isothermal in the core, and have a finite mass and extent.
The DF of these models is discontinuous at $E=\phi(\rt)$. @1963MNRAS.125..127M and @1966AJ.....71...64K avoided this by subtracting a constant from the DF introduced by Woolley, which makes the models continuous at $E=\phi(\rt)$. Compared to the Woolley models, the density of stars near the escape energy is reduced in these models (hereafter referred to as King models), and they display a more gentle truncation of their density profile. The resulting, more extended, low-density envelopes make these models resemble real GCs more [for an in depth discussion on the effect of the truncation on the density profiles see @1977AJ.....82..271H]. The spherical, non-rotating limit of the models introduced by @1975AJ.....80..175W, hereafter called Wilson models, are models that are continuous both in the DF and its derivative. This is achieved by subtracting an additional term linear in $E$ from the DF. These models are yet more spatially extended than King models. For some GCs in Local Group galaxies, the Wilson models provide a better description of the observed surface brightness profiles compared to the King models (@2005ApJS..161..304M; @2012MNRAS.419...14C also show that models that are more extended than King models better describe the surface brightness profiles of some GCs).
An additional outcome of the two-body relaxation process is that it drives the velocity distribution of the stars towards isotropy. Isotropic models, defined by a DF that only depends on $E$, are therefore a natural choice for clusters that are in late stages of their evolution, near dissolution. At early phases, however, the velocity distribution in the outer parts is expected to be radially anisotropic. This is, first, because the (incomplete) violent relaxation process that takes place during their formation results in a halo of radial orbits [@1967MNRAS.136..101L]. Secondly, two-body ejections from the dense core populate the halo with radial orbits on a two-body relaxation time-scale [@1972ApJ...173..529S]. @1963MNRAS.125..127M proposed a separable DF, dependent on $E$ and on the (specific) angular momentum $J$ to introduce radial anisotropy (hereafter referred to as Michie-King models). The DF of the Michie-King models is the product of the isotropic DF with an exponential term with a $J^2$ dependent argument. This is similar to Eddington’s method of including radial anisotropy in the isothermal model [@1915MNRAS..75..366E]. As a result, the inner parts of the models remain approximately isothermal and isotropic, which is appropriate to GCs because there the relaxation time is short, and anisotropy becomes important at larger distances from the centre. Near the truncation radius the models become isotropic again as a result of the energy truncation. The latter property has a somewhat coincidental resemblance to GCs, because near the Jacobi radius the orbits of stars gain angular momentum due to the interaction with the (tri-axial) tidal potential [@1992ApJ...386..519O], therewith suppressing the amount of radial anisotropy near the truncation energy. A review of the effect of anisotropy on model properties can be found in .
In real GCs, which contain multiple mass components, the relaxation process drives the systems towards equipartition, resulting in the heavier components being more centrally concentrated, a state which is often referred to as mass segregated. King models with different mass species were first introduced by @1976ApJ...206..128D and have since been applied to take into account the effects of mass segregation in mass-modelling efforts of Galactic GCs (e.g. M3: @1979AJ.....84..752G, Omega Cen: and larger samples of GCs: @1993ASPC...50..357P [@2012ApJ...755..156S]). Mass segregation is important for almost all of the Galactic GCs, given their short relaxation time-scales, relative to their ages [@H61; @2011MNRAS.413.2509G]. Approximating multimass systems by single-mass models can lead to severe biases in the inferred properties of GCs [@2015MNRAS.448L..94S; @sollima15] and it is, therefore, desirable to have the ability to include multiple mass components in a dynamical model of a GC.
It is our aim to develop a family of models that capture the general behaviour of collisional systems discussed above, and whose properties can be varied by parameters that can be constrained by observational data. showed that the expressions for the DF of the isotropic Woolley, King and Wilson models can be generalized by a DF in which the exponential function of $E$ is reduced by the leading orders of its series expansion. This approach was further generalized by @2014JSMTE..04..006G [hereafter GV14], who showed that solutions [*in between*]{} these models can be obtained (these models are briefly reviewed in Section \[ssec:iso\]). In this paper we extend the models of GV14 to allow for the presence of (radially biased) pressure anisotropy and multiple mass components. We present an efficient Poisson solver in to facilitate the use of these models in fitting observational data, and in drawing samples from the models, which can be used as initial conditions for numerical simulations.
The paper is organized as follows: in Section \[sec:model\], we define the models and in Section \[sec:properties\], we illustrate their main properties. In Section \[sec:limepycode\], we present the code [^1] and our conclusions and a discussion are presented in Section \[sec:conclusion\]. Supporting material can be found in the appendices.
Model definition and scaling {#sec:model}
============================
Single-mass models
------------------
### Distribution function (DF) {#ssec:df}
The DF of the single-mass family of models is $$f(E,J^2) =
\displaystyle A\exp\left(-\frac{J^2}{2\ra^2 s^2}\right)\Eg\left(g, -\frac{E-\phi(\rt)}{s^2}\right)
\label{eq:dfani}$$ for $E\le\phi(\rt)$, and 0 for $E>\phi(\rt)$. The DF depends on two integrals of motion: the specific energy $E = v^2/2 + \phi(r)$, with $v$ the velocity and $\phi(r)$ the specific potential at distance $r$ from the centre, and the norm of the specific angular momentum vector $J = |{\textbf{\emph{r}}}\times {\textbf{\emph{v}}}|=rv\sin\vartheta$, where ${\textbf{\emph{r}}}$ and ${\textbf{\emph{v}}}$ are the position vector and velocity vector, respectively, and $\vartheta$ is the angle between them. The energy $E$ is lowered by the potential at the truncation radius $\phi(\rt)$.
In equation (\[eq:dfani\]) we introduced the function $$\begin{aligned}
\Eg(a, x) =
\begin{cases}
\exp(x) & a=0 \\
\displaystyle\exp(x) P(a, x) & a>0,
\end{cases}
\label{eq:eg}\end{aligned}$$ where $P(a, x) \equiv \gamma(a, x)/\Gamma(a)$ is the regularized lower incomplete gamma function (see Appendix \[app:gamma\] for the definition of this function and its properties). Combining the exponential and the incomplete gamma function into a single function $\Eg(a, x)$ has advantages in deriving the model properties (see GV14 and Appendix \[AppD:Eg\] for details on the behaviour of this function).
A model is specified by three parameters: the central potential, which is a required boundary condition for solving Poisson’s equation and defines how concentrated the model is; the anisotropy radius $\ra$, which determines the amount of anisotropy present in the system (for increasing $\ra$ the models are more isotropic); the truncation parameter $g$, which controls the sharpness of the truncation of the model (this parameter is called $\gamma$ in GV14). The physical units of a model are defined by two scales: the velocity scale $s$, and the normalization constant $A$, which sets the phase-space density and therewith the total mass $M$. For more information regarding scales and parameters of the models we refer the reader to Section \[ssec:scaling\].
The isotropic models ($\ra\rightarrow\infty$) and their properties are discussed in detail in GV14. For these models, and integer values of $g$, three well-known families of models are recovered: when $g=0$ we retrieve the @1954MNRAS.114..191W models, for $g=1$ we recover the King models [@1963MNRAS.125..127M; @1966AJ.....71...64K], and for $g=2$ we find the (isotropic, non-rotating) Wilson models [@1975AJ.....80..175W][^2]. In practice, the models defined by equation (\[eq:dfani\]) are radially anisotropic for $\ra\lesssim\rt$, because of the $J^2$ dependence in the first exponential. When $g=1$, the DF is the Michie-King model [@1963MNRAS.125..127M], which is often used to fit GC data .
The potential $\phi(r)$ is found by solving Poisson’s equation. For the self-consistent problem we consider here, the potential is completely determined by the density $\rho$ associated with the DF. This problem is non-linear, because the DF depends on the potential. Since the models defined by equation (\[eq:dfani\]) are spherically symmetric, Poisson’s equation is $$\begin{aligned}
\frac{1}{r^2}\frac{\dr}{\dr r}\left(r^2\frac{\dr\phi}{\dr r}\right)& = 4\pi G\rho,\end{aligned}$$ where the density is obtained by means of an integration of the DF over all velocities $$\rho = \int \dr^3 v \, f(E, J^2).
\label{eq:rho}$$ In Sections \[ssec:iso\] and \[ssec:ani\], we derive analytic expressions for $\rho$ as a function of $\phi$ and $r$. Note that only in the anisotropic case the dependence on the radial coordinate $r$ is both implicit (through $\phi$, as in the isotropic case), and explicit, i.e. $\rho(\phi, r)$. Having analytic expressions for $\rho(\phi, r)$, avoids the need of solving a double integral at each radial step, making it significantly faster to obtain the solution to Poisson’s equation. In the next section we introduce a convenient set of units to solve the model.
### Scaling and units {#ssec:scaling}
To solve Poisson’s equation, we use a dimensionless (positive) energy $\ehat = \phihat - \khat$, with dimensionless potential $\phihat =
(\phi(\rt) - \phi)/s^2$, and $\khat \equiv v^2/(2 s^2)$. As in @1966AJ.....71...64K, we consider the dimensionless density by normalizing $\rho$ to its central value, i.e. $\rhohat
= \rho/\rho_0$. In this way, Poisson’s equation in dimensionless form reads $$\frac{1}{\rhat^2}\frac{\dr}{\dr\rhat}\left(\rhat^2\frac{\dr\phihat}{\dr\rhat}\right) = -9\rhohat.$$ The dimensionless radius is now defined by the other scales: $\rhat =
r/\rs$, with $\rs^2 = 9 s^2/(4\pi G\rho_0)$. This radial scale was introduced in @1966AJ.....71...64K and is often referred to as the King radius. The factor of 9 was introduced to give $\rs$ the meaning of a core radius, because for models with moderately high central concentration, the projected density at $\rs$ is about one half of its central value.
The Poisson equation is solved by assuming the boundary conditions at $\rhat = 0$: $\phihat = \phihat_0$ and $\dr\phihat/\dr r = 0$. As mentioned in Section \[ssec:df\], the central potential $\phihat_0$ is one of the parameters that define the model[^3].
### Isotropic models {#ssec:iso}
We first briefly review the isotropic version of these models, as introduced by GV14. Many quantities can be derived from the DF. The density $\rho$ is found by integrating the DF over all velocities (equation \[eq:rho\]) and the pressure is found by taking the second velocity moment of the DF [^4] $$\begin{aligned}
\rho &= (2\pi s^2)^{3/2}A\rhoint, \label{eq:rhodef} \\
\rho\mvsq &= (2\pi s^2)^{3/2}s^2A \vsqint.\end{aligned}$$ Here $\sigma^2=3\sigmaoned^2$ is the mean-square velocity, $\sigmaoned$ is the one-dimensional velocity dispersion and we introduce a dimensionless density integral ($\rhoint$) and a dimensionless pressure integral ($\vsqint$) $$\begin{aligned}
\rhoint &= \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr \khat \, \khat^{1/2}\Eg(g, \phihat - \khat) = \Eg(\gthree, \phihat), \label{eq:rhointiso}\\
\vsqint &= \frac{4}{\sqrt{\pi}}\int_0^{\phihat} \dr \khat \, \khat^{3/2}\Eg(g, \phihat - \khat) = 3\Eg(\gfive, \phihat). \label{eq:vsqintiso}\end{aligned}$$ The results of these integrations follow straightforwardly from the convolution formula of the $\Eg(a,x)$ function (equation \[EG\_convolution\]). An alternative derivation by means of fractional calculus is presented in Appendix \[app:fractional\]. The dimensionless density that appears in Poisson’s equation is therefore $\rhohat=\rhoint/\rhointnull$, where $\rhointnull$ is the result of equation (\[eq:rhointiso\]) evaluated at $\phihat=\phihat_0$. The dimensionless mean-square velocity is found from $\sigmahat^2 = \sigma^2/s^2 = \vsqint/\rhoint$.
### Anisotropic models {#ssec:ani}
Here we present the relevant quantities for the anisotropic case. The details of the derivations can be found in Appendix \[app:series\], and the derivations by means of fractional calculus can be found in Appendix \[app:fractional\]. To solve the anisotropic models, we introduce $t=\cos\theta$, such that we can write the integral over the angles as $4\pi\int_0^1 \dr t$. We further introduce $\rrahat= \rhat/\rahat$ such that the density integral becomes
$$\begin{aligned}
\rhoint&=\frac{2}{\sqrt{\pi}}\!\int_0^{\phihat}\!\dr\khat\int_{0}^{1}\! \dr t \, \exp\left[\khat \phat^2 (t^2-1)\right] \khat^{1/2}\Eg(g, \phihat - \khat) \nonumber \\
&=\frac{2}{\sqrt{\pi}} \int_0^{\phihat}\dr\khat\, \frac{F(\phat\khat^{1/2})}{\phat}\Eg(g, \phihat - \khat).
\label{eq:rhointani}\end{aligned}$$
Here $F(x)$ is Dawson’s integral and we refer to Appendix \[App:Dawson\] for some properties of this function. To first order, $F(x) \propto x$, and we thus find that for large $\rahat$, i.e. small $\phat$, equation (\[eq:rhointani\]) converges to the integral of the isotropic model (equation \[eq:rhointiso\]). The solution of the integration gives $\rhoint$ as a function of $\phihat$ and $\rrahat$ $$\rhoint \!=\! \frac{\Eg( \gthree, \phihat)}{1+\phat^2}+\frac{\phat^2}{1+\phat^2} \frac{\phihat^{g+{\tfrac{3}{2}}}{_1}F_1(1, \gfive, -\phihat\phat^2)}{\Gamma(g+{\tfrac{5}{2}})}.
\label{eq:rhointanires}$$ Here $\hyp(a,b, x)$ is the confluent hypergeometric function whose properties are given in Appendix \[1F1\]. For small $\phat$, the second term on the right-hand-side goes to zero and the solution converges to the isotropic result of equation (\[eq:rhointiso\]). This expression for the density integral allows for fast computations of the right-hand-side of Poisson’s equation and facilitates efficient solving of the anisotropic models.
For the anisotropic models, we need to calculate both the radial and the tangential[^5] components of the pressure tensor, as well as the total pressure. The radial and tangential component of the velocity vector are defined as $\vr =
v\cos\theta$ and $\vt = v\sin\theta$ and for the corresponding integrals we find $$\begin{aligned}
\displaystyle\vrsqint&\hspace{-0.cm}=\hspace{-0.cm}\frac{4}{\sqrt{\pi}}\!\int_0^{\phihat}\!\!\!\dr\khat\!\int_{0}^{1}\!\!\!\dr t\,\exp\!\left[\rrahat^2\khat(t^2\!\!-\!1)\right]t^2\khat^{3/2}\Eg(g, \phihat-\khat),\\
\displaystyle\vtsqint&\hspace{-0.cm}=\hspace{-0.cm}\frac{4}{\sqrt{\pi}}\!\int_0^{\phihat}\!\!\!\dr\khat\!\int_{0}^{1}\!\!\!\dr t\,\exp\!\left[\rrahat^2\khat(t^2\!-\!1)\right](1\!-t^2)\khat^{3/2}\Eg(g, \phihat-\khat),\\
\displaystyle\vsqint&\hspace{-0.cm}=\hspace{-0.cm}\frac{4}{\sqrt{\pi}}\!\int_0^{\phihat}\!\!\!\dr\khat\!\int_{0}^{1}\!\!\!\dr t\,\exp\!\left[\rrahat^2\khat(t^2\!-\!1)\right]\khat^{3/2}\Eg(g, \phihat-\khat).\end{aligned}$$ By carrying out these integrals as described in Appendices \[sapp:sigma\_deriv\] and \[sapp:sigma\_deriv\_frac\], we obtain $$\begin{aligned}
\hspace{-0.25cm}\vrsqint &\!=\!\frac{\Eg(\gfive, \phihat)}{1+\phat^2}\!+\!\frac{\phat^2}{1+\phat^2} \frac{\phihat^{g+{\tfrac{5}{2}}}\hyp(1, g\!+\!{\tfrac{7}{2}}, \!-\phihat\phat^2)}{\Gamma(g+{\tfrac{7}{2}})}, \label{eq:vrsqintres} \\
\hspace{-0.25cm}\vtsqint &= \frac{\Eg(\gfive, \phihat)}{1+\phat^2}\frac{2}{(1+\phat^2)} + \frac{2\phat^2}{1+\phat^2}\frac{\phihat^{g+{\tfrac{5}{2}}}}{\Gamma(g+{\tfrac{7}{2}})} \nonumber\\
&\times\left[ \frac{\hyp(1, g+{\tfrac{7}{2}}, -\phihat\phat^2)}{1+\phat^2} + \hyp(2, g+{\tfrac{7}{2}}, -\phihat\phat^2) \right], \label{eq:vtsqintres} \\
\hspace{-0.25cm}\vsqint&\!=\!\frac{\Eg(\gfive, \phihat)}{1+\phat^2}\frac{(3+\phat^2)}{(1+\phat^2)} +\frac{\phat^2}{1+\phat^2} \frac{\phihat^{g+{\tfrac{5}{2}}}}{\Gamma(g+{\tfrac{7}{2}})} \nonumber\\
\times&\left[\!\frac{3+\phat^2}{1+\phat^2}{_1}F_1(1, g\!+\!{\tfrac{7}{2}},\!-\phihat\phat^2)\!+\!2{_1}F_1(2, g\!+\!{\tfrac{7}{2}},\!-\phihat\phat^2)\right]\!. \label{eq:vsqintres} \end{aligned}$$ Note that the expression for $\vrsqint$ resembles the expression for $\rhoint$ of equation (\[eq:rhointanires\]), in the sense that the functional form is the same, but all arguments and the power index that include $g$ are increased by 1. We already saw a similar resemblance between $\rhoint$ and $\vsqint$ in the isotropic case (equations \[eq:rhointiso\] and \[eq:vsqintiso\], respectively).
With these expressions for the density and pressure integrals, we defined most of the properties of these models that are of direct relevance for comparison to data. In Section \[ssec:project\], we discuss how the projected quantities can be derived.
### Limits {#sssec:limits}
In this section we consider some limits of the models. In the core, where $\phat$ is small ($\rhat\ll\rahat$), the model is isotropic. This is because the second terms in equations (\[eq:rhointanires\]), (\[eq:vrsqintres\]), (\[eq:vtsqintres\]) and (\[eq:vsqintres\]) vanish due to the multiplication by $\phat^2$.
Near the truncation radius the models behave like polytropes and are, therefore, also isotropic, because $$\begin{aligned}
\lim_{\phihat\rightarrow0} \hyp(1, a\!+\!1,-\phat^2\phihat) &= 1,\\
\lim_{\phihat\rightarrow0} \Eg(a, \phihat) &= \frac{\phihat^a}{\Gamma(a+1)},\end{aligned}$$ and the $\rrahat$ dependence disappears. In this regime, we find $$\begin{aligned}
\lim_{\phihat\rightarrow0} \rhohat & = \frac{\phihat^{g+3/2}}{\Gamma(g+5/2)}, \label{eq:poly}\\
\lim_{\phihat\rightarrow0} \rhohat \sigmahat^2 & = 3\frac{\phihat^{g+5/2}}{\Gamma(g+7/2)},\\
\lim_{\phihat\rightarrow0} \rhohat \sigmarhat^2 & = \frac{1}{3}\lim_{\phihat\rightarrow0} \rhohat \sigmahat^2 = \frac{1}{2} \lim_{\phihat\rightarrow0} \rhohat \sigmathat^2.\end{aligned}$$ This suppression of the velocity anisotropy near the truncation radius results naturally from the mathematical definition of the truncation, and is appropriate for tidally truncated systems [@1992ApJ...386..519O]. In $N$-body models a tangentially biased anisotropy is observed near $\rt$ [@sollima15], which cannot be reproduced by the models presented here. However, it is likely that most of the stars with tangentially biased velocities are above the escape energy, so-called potential escapers and these are not considered by these models, nor any other model we are aware off.
Models with $\phihat_0\rightarrow0$ are close to pure polytropes over their entire radial range. In this regime, and for $g=7/2$ (i.e. a polytropic index $n=5$, equation \[eq:poly\]), we recover the @1911MNRAS..71..460P model, which is infinite in extent ($\rho \propto r^{-5}$ at large radii), but finite in mass. Polytropes with $n\ge 5$ (i.e. $g\ge7/2$) are infinite in extent and will not be considered here. For $g<7/2$ models can have a finite $\rt$ depending on both $\phihat_0$ and $\ra$ (see GV14 and Section \[sec:properties\]).
In the cores of models with $\phihat_0\gg0$ the DF approaches the isothermal sphere, because $$\begin{aligned}
\lim_{\phihat\rightarrow\infty} \Eg(a,\phihat) & =\exp(\phihat).\end{aligned}$$ Models with $g\rightarrow\infty$ also approach the isothermal sphere. To conclude, these models approach the isothermal sphere in the limit of $\phihat_0\rightarrow\infty$, independent of $g$, but also in the limit of $g\rightarrow\infty$, independent of $\phihat_0$.
Multimass models {#ssec:multimassmodel}
----------------
It is possible to consider models with multiple mass components, by considering the DF as the sum of DFs of the form of equation (\[eq:dfani\]), each of which describes a different mass component with a mass-dependent velocity scale parameter. The first to do this were @1976ApJ...206..128D, who calculated multimass King models. For a multimass model with $\Ncomp$ mass components, $2\Ncomp+2$ parameters are required in addition to the ones introduced in Section \[ssec:df\] for single-mass models. These additional parameters are the values for the component masses $m_j$, the amount of mass in each component $M_j$, $\delta$ and $\eta$. The latter two parameters set the mass dependence of the velocity scale $s_j$ and the anisotropy radius of each component $\rajhat$, for which we adopt power-law relations $$\begin{aligned}
s_j &= s\mu_j^{-\delta}, \label{eq:delta} \\
\rajhat &= \rahat\mu_j^{\eta}. \label{eq:eta}\end{aligned}$$ Here $\mu_j= m_j/\bar{m}$ is the dimensionless mass of component $j$ and $\bar{m}$ is the central density weighted mean-mass $$\bar{m} = \frac{\sum_j m_j\rho_{0j}}{\sum_j \rho_{0j}}.$$ Note that in the multimass models, the values of $s$ and $\rahat$ are the velocity scale and anisotropy radius corresponding to $\bar{m}$. The definitions of $\delta$ and $\eta$ are such that the anisotropy profiles are approximately mass independent when $\delta=\eta$ (see equation \[eq:dfani\]). The typical values considered for these parameters are $\delta = 1/2$ and $\eta = 0$.
We notice that in the limit of infinite $\phihat_0$ the velocity scale $s_j$ approaches the one-dimensional velocity dispersion of mass component $j$, $\sigma_{{\rm 1d},j}$, hence the traditional assumption for $\delta=1/2$ implies equipartition ($m_j s_j^2 =\bar{m}\sigma_{{\rm 1d},j}^2=$ constant). However, it is important to keep in mind that for multimass models with typical and realistic values of $\phihat_0$, the velocity dispersion of each component in the centre is smaller than $s_j$ and, therefore, there is no equipartition (see Section \[ssec:delta\] and @1981AJ.....86..318M [@2006MNRAS.366..227M]).
To solve a multimass model self-consistently, we compute the density for each mass component as in equation (\[eq:rho\]) and add all components on the right-hand-side of Poisson’s equation. The detailed procedure is described in @1979AJ.....84..752G, and here we only briefly summarize the required steps. The dimensionless Poisson equation to solve is $$\hat{\nabla}^2\phihat = -9\sum_j\alpha_j\rhohat_j,$$ where $\alpha_j$ is the ratio of the central density of the $j$-th mass component to the total central density, such that $$\sum_j \alpha_j = \sum_j \frac{\rho_{0j}}{\rho_0} = 1$$ and $$\rhohat_j = \frac{\rho_j}{\rho_{0j}} = \frac{\rhoint(\mu^{2\delta}\phihat,\hat{r})}{{\rhoint(\mu^{2\delta}\phihat_0, 0)}}.
\label{eq:rhointratio}$$ By considering multiple mass components, we introduce an eigenvalue problem in the solution of Poisson’s equation, because the values of $\rho_{0j}$ that yield the desired ${M}_j$ values are not known a priori. Therefore, as a first step to solve the model, we assume that $\alpha_j = {M}_j/\sum_j M_j$, and we obtain the solution by iteration (see Section \[sec:limepycode\] for details).
Normalization and potential energy
----------------------------------
In solving the models we have chosen to define the dimensionless quantities in terms of the density scale $\rho_0$ and the velocity scale $s$ (Section \[ssec:scaling\]). In some cases it is useful to have an expression for the normalization constant $A$ in the DF (equation \[eq:dfani\]), for example, when fitting models to discrete data. From equation (\[eq:rhodef\]) we find that $A$ relates to the other scales as
$$\begin{aligned}
A & = \frac{\rho_0}{(2\pi s^2)^{3/2}\rhointnull}.\end{aligned}$$
For the multimass models there is a normalisation for each component, $A_j$. The relation with the mass scale $\ms=M/\mhat$ is $\ms=\rs^3\rho_0=\rs^3(2\pi s^2)^{3/2} A\rhointnull$, where we introduced $\mhat = \int\rhohat\dr^3\rhat$.
The total dimensionless (positive) gravitational energy $\Uhat$ of the model is calculated from integrating the potential [@1966AJ.....71...64K] $$\begin{aligned}
\Uhat &= \frac{1}{2} \int_{0}^{\mhat} \dr \hat{m} \, \phihat + \frac{\hat{G}\mhat^2}{2\rthat}.
\label{eq:Uhat}\end{aligned}$$ The second term has to be added because $\phihat$ is a lowered potential. Note that this integration of $\phihat$ over mass is readily obtained from solving Poisson’s equation.
Model properties {#sec:properties}
================
Single-mass models
------------------
### Density and velocity dispersion profiles
In Fig. \[fig:rhovel\], we show the density profiles, the velocity dispersion profiles, and the anisotropy profiles for isotropic and anisotropic models with different values of the truncation parameter $g$. The anisotropy profile is computed from $\sigmat^2$ and $\sigmar^2$ as $$\beta = 1-\frac{\sigmat^2}{2 \sigmar^2}.
\label{eq:beta}$$ In the case of isotropy $\beta = 0$, $0<\beta\le1$ indicates radially biased anisotropy (with $\beta=1$ implying fully radial orbits) and for tangentially biased anisotropy $\beta<0$. Because $\beta$ is a measure of anisotropy locally, we also quantify the total amount of anisotropy with $$\kappa = \frac{2K_{\rm r}}{K_{\rm t}},
\label{eq:kappa}$$ introduced by @1981SvA....25..533P. Here $K_{\rm r}$ and $K_{\rm t}$ are the radial and tangential components of the kinetic energy, respectively. For isotropic models $\kappa =1$, and for radially biased anisotropic models $\kappa>1$. @1981SvA....25..533P found that for $\kappa>1.7\pm0.25$ radial orbit instability occurs. We use this criterion to check the stability of the anisotropic models we calculate.
![Dimensionless density profile (top), velocity dispersion profile (middle) and anisotropy profile (bottom) for models with different truncation parameters $g$ (different colours). Isotropic models are shown with solid lines, anisotropic models with $\rahat/\rhhat=1.5$ with dashed lines.[]{data-label="fig:rhovel"}](rho){width="\columnwidth"}
In Fig. \[fig:rhovel\] we show anisotropic models characterized by $\rahat/\rhhat=1.5$. Because the (dimensionless) half-mass radius $\rhhat$ is not known before solving the model, we find the value of $\rahat$ that gives the correct ratio $\rahat/\rhhat$ iteratively. We see that all models are approximately isothermal in the centre. When increasing $g$, the models become more extended. Including radial anisotropy also results in a larger truncation radius.
Note that, with this choice of $\rahat/\rhhat$, the maximum value assumed by the anisotropy function for $g = 0$ (Woolley model) is about $0.4$, while for $g = 2$ (Wilson model) it is possible to achieve $\beta\simeq1$ in the outer parts of the model. This dependence of the maximum value of $\beta$ on $g$ does not imply that there are differences in the total amount of anisotropy: for all the anisotropic models shown in Fig. \[fig:rhovel\], indeed, we find $\kappa\simeq1.2$. The ability to calculate models with more radial orbits (larger $\beta$) without increasing the radial component of the total kinetic energy is important to keep in mind when considering other physical effects that can enhance or suppress the amount of radial orbits, such as the presence of a dark matter halo [@2013MNRAS.428.3648I] and the galactic tides [@1992ApJ...386..519O]. In a forthcoming study, we quantify the presence of radial orbits in direct $N$-body models of tidally limited clusters [@2016MNRAS.462..696Z].
### DF, density of states and differential energy distribution
In the top panels of Fig. \[fig:dmde\_comp\], we show the DF as a function of $\ehat$, for isotropic models, with different values of $g$ and $\phihat_0$. In the middle panels we show the density of states $\gdos(\ehat)$, which is the phase-space volume per unit of energy (see equation \[densityofstates\] for a definition). The bottom panels display the differential energy distribution $\dr \mhat/\dr \ehat$, which is the amount of mass per unit energy. For the isotropic models it is simply the product of $f(\ehat)$ and $\gdos(\ehat)$ (equation \[eq:dmde\]). Details on how this was derived for the models presented here, and on the procedure for anisotropic models, are given in Appendix \[app:dmde\]. A general discussion on the differential energy distribution can be found in chapter 4 of @BT1987.
{width="\textwidth"}
In the first and third columns (linear $x$-scale), we recognize the exponential behaviour of $f(\ehat)$ for the $g=0$ model, and the exponential behaviour at high $\ehat$ for $g>0$ models. From the second and fourth column, we see that at low $\ehat$, the DF scales as $f(\ehat) \propto \ehat^g$, which corresponds to the regime where the models behave as polytropes. From Fig. \[fig:dmde\_comp\] it is also evident that when $\ehat \simeq \phihat_0$, the model behaviour is independent of $g$.
From the differential energy distribution, we see that only for $g=0$ there is a non-zero mass at $\ehat=0$. For models with $g>0$, the truncation is such that $f(\ehat=0)=\dr \hat{M}/\dr \ehat\left|_{\ehat=0}\right.=0$. These models give rise to more realistic looking density profiles, but in real GCs the number of particles with the escape energy is not zero [@2001MNRAS.325.1323B], because of the gradual scattering of particles over the critical energy for escape by two-body relaxation, and because of the finite time for stars to escape from the Jacobi surface imposed by the galactic tidal field [@2000MNRAS.318..753F].
### Finite and infinite models
As discussed in Section \[sssec:limits\], there are no models with finite extent if $g\ge3.5$. GV14 showed that the maximum value $g_{\rm
max}$ to get models with a finite extent depends on $\phihat_0$, and $g_{\rm max}=3.5$ holds in the limit of $\phihat_0\rightarrow0$. GV14 show that all their isotropic models are finite for $g \lesssim 2.1$.
We note that there is a class of isotropic models that are finite in extent, but are not relevant to star clusters, and that are not discussed in GV14. This is illustrated in Fig. \[fig:rho\_multi\], where we show density profiles for models with different $\phihat_0$ and $g=2.75$. The model with $\phihat_0=3$ converges to a finite $\rthat$ and has a density profile comparable in shape to the ones shown in Fig. \[fig:rhovel\]. The model with $\phihat_0=9$ is infinite in extent, and only plotted up to $\log\,\rhat=10$. The models with $\phihat_0=5$ and $\phihat_0=7$ are finite, but show a sharp upturn in the density profile at large radii, which causes them to have a lot of mass in the envelope, but little energy, which makes these models inapplicable to real stellar systems. Their extreme density contrast between the core and the extended halo makes these models perhaps applicable to red giant stars [see the density profiles for red giants in @2012ApJ...744...52P]. To quantify the boundary between models with, and without the core-halo structure, we compute the ratio of the dimensionless virial radius $\rvhat=-G\mhat^2/(2\Uhat)$ over $\rhhat$ for a grid of models with $0\leq\phihat_0\leq20$ and $0\leq g\leq3.5$, and we show the result as contours in Fig. \[fig:rvrh\_phi0\]. We find that for a given $g(\phihat_0)$, when increasing $\phihat_0(g)$, the change in $\rvhat/\rhhat$ is large and abrupt once the models develop the core-halo structure. We identify the value of $\rvhat/\rhhat\simeq0.64$ as the one separating the two classes of models. In the remaining discussion, we only consider models with $\rvhat>0.64\rhhat$.
When considering anisotropic models, we find that for each $\phihat$ and $g$, there is a minimum value of $\rahat$ that can be used to obtain a model that has a finite extent. We note that models with infinite extent can have a finite total mass, but because we envision an application of these models to tidally limited systems we do not consider them here. In Fig. \[fig:ramin\], we show the minimum $\rahat$ for which models are finite in extent. The lines show, as a function of $\phihat_0$, and for different $g$, the values of $\rahat$ that are needed to get $\rthat = 10^7$. Note that this minimum for $\rahat$ goes up approximately exponentially with $\phihat_0$, and also increases with $g$.
![Density profiles for isotropic models with truncation parameter $g=2.75$. Models with $\phihat_0 = 3$, $5$, and $7$ (blue, green, and red line, respectively) have a finite truncation radius, but only the model with $\phihat_0=3$ is relevant when describing GCs; the model with $\phihat_0=9$ (light blue line) is infinite in extent.[]{data-label="fig:rho_multi"}](rho_multi){width="\columnwidth"}
![Ratio of dimensionless virial radius to half-mass radius, $\rvhat/\rhhat$, for models with different $\phihat_0$ and $g$. We consider models with $\rvhat/\rhhat\ge0.64$ as relevant to describe star clusters. Models that have an infinite $\rthat$ are plotted as $\rvhat/\rhhat=0$ (i.e. they correspond to the white region in the plot).[]{data-label="fig:rvrh_phi0"}](rvrh_phi0){width="\columnwidth"}
![Minimum $\rahat$ for finite sized models, for different $\phihat_0$ and $g$.[]{data-label="fig:ramin"}](ramin){width="8cm"}
### Entropy
@1966AJ.....71...64K suggested that in the process of core collapse, clusters evolve along a sequence of models with increasing central concentration. He also noted that his models are probably not able to describe the late stages of core collapse, because for large central concentration the variation in energy due to a change in the central concentration occurs in the envelope, and not in the core. Further support for this idea comes from @1968MNRAS.138..495L, who showed that a maximum in entropy occurs at $\phihat_0 \simeq 9$ for both Woolley and King models at constant mass and energy. The entropy of a self-gravitating system is obtained from the DF as
$$S = -\int \dr^3 r \, \dr^3v \, f \ln f \ .
\label{eq:entropy}$$
Because two-body encounters continuously increase the total entropy of the system, we do not expect King models to be able to describe a system in the late stages of core collapse (i.e. $\phihat_0\gtrsim9$). This was confirmed by Fokker-Planck models of isolated star clusters going into core collapse [@1980ApJ...242..765C], for which the entropy increase follows that of King models with increasing central concentration, up to a value of $\phihat_0 \simeq 9$, but then it continues to rise during the gravothermal catastrophe. Cohn concluded that in this regime, the isotropic King models are not able to describe the entropy evolution in his simulations.
In Fig. \[fig:entropy\], we show the entropy $S$, computed as in equation (\[eq:entropy\]), for the isotropic King models (black solid line), which shows a maximum at $\phihat_0 \simeq 9$. We also show the entropy curves for different values of $g$, and for selected anisotropic models. All models are scaled to the same $M$ and total energy $\Etot$, in the conventional $N$-body units: $G=M=-4\Etot=1$ . For $0\lesssim \phihat_0 \lesssim 1$, the anisotropic models are similar to their corresponding isotropic models, and therefore they have similar entropy. From this plot it is apparent that evolution at constant mass and energy, and with increasing entropy is possible beyond $\phihat\gtrsim9$ if $g$ is increased, and/or $\ra$ is decreased (i.e. including more anisotropy). A local maximum in entropy is seen near $\phihat_0\simeq17$. Similar oscillating behaviour of the entropy was found for isothermal models in a non-conducting sphere and we refer to @1968MNRAS.138..495L and @1989ApJS...71..651P for detailed discussions. A study of equilibria in lowered isothermal models of the Woolley and King-type can be found in @1978ApJ...223..299K; for a discussion on the evolutionary sequence of quasi-equilibrium states in $N$-body systems we refer to @2005MNRAS.364..990T. It would be of interest to compare the models discussed here to the phase-space density of particles in an $N$-body system undergoing core collapse.
![Entropy curves for isotropic and anisotropic models with different truncation prescriptions (i.e. different values of $g$). All models are scaled to the same mass and energy. The anisotropic models are shown as dashed lines and for these models we used $\ra=\rv$. For $g\ge1.5$ the anisotropic models are not finite for all $\phihat_0$, and the corresponding curves are therefore not plotted. This figure shows that the entropy can be increased by increasing $g$, and/or by decreasing $\ra$.[]{data-label="fig:entropy"}](entropy){width="\columnwidth"}
In Fig. \[fig:gphi0\], we illustrate the dependence of the entropy on $g$ and $\phihat_0$ for isotropic models. For a model with $g=1$ and a low concentration, the entropy can be increased by moving to the right in this diagram, and near $\phihat_0\simeq9$ the entropy can be increased by increasing $g$.
![Entropy contours for isotropic models, all scaled to $G=M=-4\Etot=1$, with different $\phihat_0$ and $g$. Contours of constant $\rthat$ are shown as black lines. Moving to the right at constant $g$ leads to an increase of entropy up to $\phihat_0\simeq9$ [@1968MNRAS.138..495L; @1980ApJ...242..765C]. The entropy can grow further by increasing $g$ at constant $\phihat_0\simeq9$. The maximum entropy is found near $\phihat_0 \simeq9$ and $g\simeq2.2$.[]{data-label="fig:gphi0"}](gphi0){width="\columnwidth"}
In Fig. \[fig:rarh\_phi0\], we show the dependence of entropy on anisotropy, expressed here in terms of $\ra/\rh$, for models with $g=0$. We see that for constant $\ra/\rh\gtrsim1$, the entropy can increase by increasing $\phihat_0$, up to about $\phihat_0\simeq9$ [this was also found by @1998MNRAS.301...25M in a study of anisotropic Woolley, King and Wilson models]. The entropy can be increased further by decreasing the anisotropy radius. A maximum is found near $\phihat_0\simeq9$ and $\ra\simeq\rh$.
![Entropy for models, all scaled to $G=M=-4\Etot=1$, with $g=0$ and different concentrations and different amounts of anisotropy, quantified here in terms of $\ra/\rh$. Contours of constant $\rthat$ are shown as black lines. A maximum in entropy is found at $\phihat_0\simeq9$ and $\ra\simeq\rh$.[]{data-label="fig:rarh_phi0"}](rarh_phi0){width="\columnwidth"}
multimass models {#ssec:multimass}
----------------
Multimass models with $\Ncomp$ mass bins require, in addition to the parameters of the single-mass models, $2\Ncomp+2$ parameters (Section \[ssec:multimassmodel\]). There is, therefore, a large variety of models that can be considered, and many properties that we can chose from to illustrate the behaviour of these models. We decide to focus on two properties that highlight important features of these multimass models in relation to mass segregation. In a follow-up study (Peuten et al., in preparation) we present a detailed comparison between the multimass models and $N$-body simulations of clusters with different mass functions.
### The role of $\delta$ {#ssec:delta}
In Fig. \[fig:vrms0\] we show the dimensionless central velocity dispersion of each mass component, $\hat{\sigma}_{{\rm 1d},j0}$, as a function of its mass $m_j$ for isotropic, 20-component models with different $\phihat_0$ and $g$. The mass bins are logarithmically spaced between $0.1\,\msun$ and $1\,\msun$ (note that the units are not important, because the model behaviour depends only on the dimensionless values $m_j/\bar{m}$), and $M_j\propto m_j^{0.7}$, which corresponds to a power-law mass function $\dr N/\dr m_j \propto m_j^{-1.3}$ (i.e. a GC-like mass function). The mass segregation parameter was set to $\delta=1/2$ (for the definition of $\delta$, see equation \[eq:delta\]).
Despite the fact that $m_j s_j^2$ is constant for all mass bins, there is no equipartition between the different mass species, i.e. $\sigmaonedj$ does not scale as $m_j^{-1/2}$ for the different mass components. This is because only in the limit of infinite central concentration $\phihat_0\rightarrow\infty$, $s_j = \sigmaonedj$, but for realistic values of $\phihat_0$, the ratio $\sigmaonedj/s_j<1$. Because the central potential for the lower mass components is smaller than the global $\phihat_0$ that defines the model, the truncation in energies reduces $\sigmaonedj$ more for low-mass components [@1981AJ.....86..318M; @2006MNRAS.366..227M]. This is illustrated by the $\phihat_0=16$ model in Fig. \[fig:vrms0\], for which a constant $m_j\sigmaonedj^2$ only holds for the most massive bins.
@2013MNRAS.435.3272T recently observed very similar trends between $\sigmaonedj$ and $m_j$ in $N$-body models of GCs (see their fig. 1) as those shown in Fig. \[fig:vrms0\]. They concluded that modelling techniques that assume equipartition, such as multimass Michie-King models, are ‘approximate at best’. We stress that multimass models that are widely used in literature, i.e. those with $\delta=1/2$ [@1976ApJ...206..128D; @1979AJ.....84..752G], are [*not*]{} in a state of equipartition, as is illustrated in Fig. \[fig:vrms0\] and has been stated previously [@1981AJ.....86..318M; @2006MNRAS.366..227M]. In fact, from a comparison of the model behaviour in Fig. \[fig:vrms0\] and the $N$-body models of Trenti and van der Marel we conclude that the most commonly chosen flavour of multimass models (i.e. King models with $\delta = 1/2$) do a good job in reproducing the degree of mass segregation in evolved stellar system [see also @sollima15].
![Dimensionless central velocity dispersion for each mass component of multimass models with a power-law mass function, and different $\phihat_0$ and $g$. The value of the mass segregation parameter is $\delta=1/2$. Equipartition in energy is only reached for large values of $m_j$ for the model with $\phihat_0=16$. The dash-dotted lines show the velocity dispersion each of the models would have in the case of equipartition.[]{data-label="fig:vrms0"}](vrms0){width="\columnwidth"}
### The role of $\eta$
In Fig. \[fig:beta\], we illustrate the effect of the parameter $\eta$ that sets the anisotropy radius of the different mass components (for the definition of $\eta$ see equation \[eq:eta\]). We show the anisotropy profiles for three-component models with $m_j = [0.2, 0.4,
0.8]$, and the same mass function as before (i.e. $M_j \propto
m_j^{0.7}$), and for different values of $\eta$. All models have $\phihat_0=9$, $g=1.5$, $\delta=1/2$ and $\rahat = 20$.
In the multimass models used in the literature $\eta$ is implicitly assumed to be $0$. From Fig. \[fig:beta\] we can see that this implies that the $\beta$ profile of the high-mass stars rises to larger values. It is tempting to interpret this as that massive stars are on more radial orbits. However, the more massive stars are also more centrally concentrated, where the velocity distribution is more isotropic. To quantify the importance of this effect, we show in each panel the values of the parameter $\kappa_j$ for each mass component (equation \[eq:kappa\]). From this we can see that in fact for the $\eta=0$ models the intermediate mass component is the most anisotropic. The relation between $\beta_j$ and $\kappa_j$ depends on the mass function, $\phihat_0$, and $g$ and this is therefore not a general property of $\eta=0$ models.
We note that for $\eta=\delta=1/2$ the $\beta_j$ profiles are nearly mass independent. Again, this does not mean that the kinetic energy in radial orbits relative to that in tangential orbits is constant, as can be seen from the values of $\kappa_j$. When considering a value of $\eta > 1/2$ we observe that the component for which $\beta_j$ assumes the largest values is the least massive one.
![Anisotropy profiles for three-component models ($m_j = [0.2,
0.4, 0.8]$) with different values for the anisotropy parameter $\eta$, that sets the anisotropy radius of the individual components as a function of their mass. All models have $\phihat_0=9$, $g=1.5$, $\delta=1/2$ and $\rahat=20$. The values of $\kappa_j$ are shown for each component in the individual panels.[]{data-label="fig:beta"}](beta){width="\columnwidth"}
The code {#sec:limepycode}
=========
General implementation {#ssec:generalimplementation}
----------------------
We introduce a -based code that solves the models and allows the user to compute some useful quantities from the DF. The code is called [ Lowered Isothermal Model Explorer in PYthon]{} (), and is available from: <https://github.com/mgieles/limepy>.
One of the main features of the code is its flexibility: the user can easily solve isotropic or anisotropic models, and include one or more mass components. The type of model to calculate is determined by the input parameters:
1. the dimensionless central potential $\phihat_0$;
2. the truncation parameter $g$;
3. the anisotropy radius $\rahat$ (for anisotropic models);
4. two arrays $m_j$, $M_j$ and $\delta$ and $\eta$ (for multimass models).
By default, the model is solved in the dimensionless units described in Section \[ssec:scaling\]. There we pointed out that the scales of the models are set by $A$ and $s$, which correspond to a mass density (in six-dimensional phase space) and a velocity scale. These two scales, combined with the gravitational constant $G$ then define the radial scale. To allow a user to scale a model to physical units, we decided to use the mass and radial scale as input, and from this the velocity scale is computed internally. The reason for this is that we foresee that an important application of the code is to recover the GC mass and radius from a comparison of the models to data. It is possible to scale the model to physical units by specifying $M$ in $\msun$ and a radial scale (either $\rv$ or $\rh$) in $\pc$. The resulting unit of velocity is then $\kms$, with $G=0.004302\,\pc\,(\kms)^2/\msun$. Alternative units, such as the units $G=\rv=M=1$ , can be considered by redefining the scales. After solving the model, the values of all typical radii are available: the King radius $\rs$, the half-mass radius $\rh$, the truncation radius $\rt$, the anisotropy radius $\ra$, and the virial radius $\rv$.
The code solves Poisson’s equation with the ‘dopri5’ integrator [@hairer1993solving], which is a Runge-Kutta integrator with adaptive step-size to calculate fourth and fifth order accurate solutions. It is supplied by the sub-package . The relative and absolute accuracy parameters are chosen as a compromise between speed and accuracy and can be adjusted by the user. The basic version of the code allows us to obtain, as a result of this integration, only the potential as a function of radius. The full model solution contains, in addition to the potential, the density, the radial and tangential components of the velocity dispersion, the global velocity dispersion profile, and the anisotropy profile (equation \[eq:beta\]). It is possible to use the potential calculated in this way to compute the value of the DF as a function of input $E$ (isotropic models), or $E$ and $J$ (anisotropic models), or positions and velocities.
After solving a model, the code carries out a simple test to see whether it is in virial equilibrium: $2\hat{K} - \Uhat = 0$, where $\hat{K}$ is the dimensionless total kinetic energy (recall that $\Uhat$ is defined to be positive, equation \[eq:Uhat\]). For models that are infinite in extent, the solver stops at a large radius, the virial equilibrium assertion fails and the lack of convergence is flagged.
For multimass models the central densities of the components need to be found by iteration (Section \[ssec:multimass\]). @1979AJ.....84..752G proposed a recipe in which the ratios of central densities over the total density, $\alpha_j$, are set equal to $M_j/\sum_j M_j$ in the first iteration. Because for $\delta>0$ the more massive components are more centrally concentrated, the amount of mass in these components is underestimated in the first iteration, while the mass in low-mass stars is overestimated. After each iteration, $\alpha_j$ is multiplied by the ratio of $M_j/M_j^\prime$, where $M_j^\prime$ is the array of masses obtained in the previous step, and then normalized again. This is repeated until convergence. However, we found that for models with low $\phihat_0$ and a wide mass spectrum the mass function does not always converge with this method. We found that multiplying $\alpha_j$ by $\sqrt{M_j/M_j^\prime}$, instead, is more reliable and results in a similar number of iterations for models that do converge with the method proposed by Gunn and Griffin.
Solutions are not numerically stable when considering large values of the arguments of the hypergeometric functions. To stabilize the calculations, we adopt the asymptotic behaviour of the hypergeometric function $\hyp(1, b, -x)$ and $\hyp(2, b, -x)$ for $x\ge700$ (see equations \[eq:1f1asym1\] and \[eq:1f1asym2\]). For multimass models with a wide mass spectrum (e.g. when stellar mass black holes are considered in addition to the stellar mass function), the central potential of the massive component can be too large for the computation of $\rhoint(\mu^{2\delta}\phihat, \rhat)$ (see equation \[eq:rhointratio\]) in the first iteration. We therefore use the approximation $\rhohat_j
= \exp[\mu^{2\delta}(\phihat-\phihat_0)]$ if $\mu^{2\delta}\phihat
>700$.
Model properties in projection {#ssec:project}
------------------------------
In order to compare the models to observations of GCs, it is necessary to compute the model properties in projection. For a spherically symmetric system, it is straightforward to compute the projected properties as a function of the projected radial coordinate $R$ [for a more detailed discussion, see for example @BT1987].
The projected surface mass density is found from the intrinsic mass density as $$\Sigma(R) = 2 \int_0^{\rt} \dr z \, \rho (r),
\label{Proj_Sigma}$$ where $r^2 = R^2+z^2$, and $z$ is along the direction of the line-of-sight. The velocity dispersion along the line of sight is given by the following integral $$\sigma_{\rm LOS}^2(R) = \frac{2}{\Sigma(R)} \int_0^{\rt} \dr z \, \rho (r) \sigma_z^2(r),
\label{Proj_Disp1}$$ where $\sigma_z^2$ is the contribution of the velocity dispersion tensor to the $z$-direction. For isotropic models, $\sigma_z^2
= \sigma^2/3$. For anisotropic models, it is possible to calculate $$\sigma_z^2(r) = \sigma_{\rm r}^2 \cos ^2 \xi + \sigma_{\theta}^2 \sin ^2 \xi,
\label{eq:sigma2zr}$$ where $\sin \xi = R/r$. We recall that, for the anisotropic models considered here, $\sigma_{\theta}^2 = \sigma_{\varphi}^2
= \sigmat^2/2$. The component $\sigma_{\varphi}^2$ does not contribute to $\sigma_z^2$, because it is always perpendicular to the line of sight. We can use the anisotropy profile $\beta$ (see equation \[eq:beta\]) to rewrite equation (\[eq:sigma2zr\]) as $$\sigma_z^2(r) = \sigma_{\rm r}^2 \left[ 1 - \beta(r) \frac{R^2}{r^2} \right].$$
The quantity $\sigma_{\rm LOS}(R)$ is useful when comparing the models to the velocity dispersion profiles that are calculated from radial (i.e. line-of-sight) velocities. Now that proper motions data are becoming available for an increasing number of GCs [@2014ApJ...797..115B], it is also interesting to compare the velocity dispersion components that can be measured on the plane of the sky with those calculated from the models. This comparison is particularly important because it is a direct way to detect the presence of anisotropy in the systems. We calculate, therefore, the radial and tangential projected components of the velocity dispersion as $$\begin{aligned}
\sigma_{\rm R}^2(R) &= \frac{2}{\Sigma(R)}\int_0^{\rt} \dr z \, \rho (r) \sigma_{\rm S}^2(r), \\
\sigma_{\rm T}^2(R) &= \frac{2}{\Sigma(R)}\int_0^{\rt} \dr z \, \rho (r) \sigma_{\rm \varphi}^2(r),\end{aligned}$$ where $\sigma_{\rm S}^2$ is given by $$\sigma_{\rm S}^2(r) = \sigma_{\rm r}^2 \left[ 1 - \beta(r) \left(\frac{1 - R^2}{r^2}\right) \right].$$
In the case of multimass models, the projected quantities introduced above are calculated separately for each mass component, by replacing every quantity in the equations above with the respective $j$th profile.
Generating discrete samples from the DF
---------------------------------------
A separate sampling routine [<span style="font-variant:small-caps;">limepy.sample</span>]{} is provided that generates discrete samples from the models. The routine takes a object containing a model as input and the number of points $N$ that need to be sampled. In the case of a multimass models the input $N$ is ignored and computed from the total mass $M$ and the pair $m_j, M_j$. Radial positions are sampled by generating numbers between 0 and 1 and interpolating the corresponding $r$ values from the (normalized) cumulative mass profile(s).
To obtain velocities, we first sample values of $x$, where $x=\khat^{3/2}=(\hat{v}^2/2)^{3/2}$. The probability density function (PDF) for $x$ can be written as $$P(x) = \frac{F(\phat x^{1/3})}{\phat
x^{1/3}}\Eg(g, \phihat(\rhat)-x^{2/3}),$$ where $\phat=\rhat/\rahat$. The function $P(x)$ has a maximum at $x=0$, and declines monotonically to 0 at $x=\phihat(\rhat)^{3/2}$. These properties make it easier to efficiently sample values for $x$ from $P(x)$, than sampling values of $v$ from $v^2 f(r,v)$. To make the rejection sampling more efficient, we adapt a supremum function $F(x)$, which consists of 10 segments between 11 values $x_i$ which are linearly spaced between $0$ and $\phihat(\rhat)^{3/2}$, and for each segment $x_i< x<x_{i+1}$, $F_i(x) = P(x_i)$. We then sample values from the function $F(x)$, reject the points that are above $P(x)$ and resample the rejected points until all points are accepted. Typically, a handful of iterations are needed.
For anisotropic models we also need to sample angles $\theta$. We do this by sampling values for $t=\cos\theta$. From the DF it follows that the PDF for $t$ is $$P(t) = \exp\left[\phat^2\khat(t^2-1)\right].
\label{eq:Pt}$$ By integrating equation (\[eq:Pt\]) we find that the cumulative DF is the imaginary error function. This function cannot be inverted analytically, hence the values for $t$ need to be found by numerically inverting this function, which can be done accurately with built-in routines.
When values for $r, \vt$ and $\vr$ are obtained, these are converted to Cartesian coordinates by generating three additional random angles.
Conclusions and discussion {#sec:conclusion}
==========================
In this study we present a family of lowered isothermal models, with the ability to consider multiple mass components and a variable amount of radially biased pressure anisotropy. The models extend the single-mass family of isotropic models recently developed by GV14. The new additions we propose here make the models ideally suited to be compared to data of resolved GCs.
The models are characterized by an isothermal and isotropic core, and a polytropic halo. The shape of the halo is set by the truncation parameter $g$, that controls the sharpness of the energy truncation, i.e. the prescription of lowering the isothermal model. For integer values of $g$, several well-known isotropic models are recovered: for $g=0$ we recover the @1954MNRAS.114..191W models, for $g=1$ the @1963MNRAS.125..127M, or @1966AJ.....71...64K models and $g=2$ corresponds to the non-rotating, isotropic @1975AJ.....80..175W models. The DF proposed by GV14, with the introduction of the continuous parameter $g$ to determine the truncation, allows us to consider models [*in between*]{} these models. The advantage of this prescription for the truncation is that it is now possible to control the sharpness of the truncation by means of a parameter.
We present [ Lowered Isothermal Model Explorer in PYthon]{} ($\limepy$), a -based code that solves the models, and computes observable quantities such as the density and velocity dispersion profile in projection. In addition, the code can be used to draw random positions and velocities from the DF, which can be used to generate initial conditions for numerical simulations.
It is interesting to discuss possible extensions of, and improvements to the models. One obvious pitfall is that the tidal field is not included in a self-consistent way. To quantify the effect of the omission of the tidal field, we can consider the specific energy $E$ at $\rt$. In our models $E(\rt)=\phi(\rt)=-GM/\rt$, whilst inclusion of the tidal terms would give (for a cluster on a circular orbit, in a reference frame corotating with the galactic orbit) a specific Jacobi energy of $E_{\rm J}(\rt) = -(3/2)GM/\rt$. Therefore, the properties of stars near the critical energy for escape are described only approximately by these models, because in this energy range the galactic tidal potential is of comparable importance as the cluster potential. Another simplification of the models is that the galactic tidal potential is triaxial, whilst our models are spherical. Both of these points could be improved upon by including a galactic tidal potential in the solution of Poisson’s equation, following the methods described by @1995MNRAS.272..317H or @2008ApJ...689.1005B, @2009ApJ...703.1911V.
The models do not include a prescription for rotation, which can be an important factor to take into account when describing real GCs . Self-consistent models with realistic rotation curves exist and have been successful in describing the rotational properties of several Galactic GCs [@2013ApJ...772...67B]. It is feasible to include rotation in the models presented in this paper, for example, in the way it is done in the @1975AJ.....80..175W model, by multiplying the DF in equation (\[eq:dfani\]) by a $J_{\rm z}$ dependent exponential term. Including the rotation, and a description of the galactic tidal field, would make the models more realistic and, therefore, a worthwhile exercise for future studies.
Lastly, we note that our models could be useful in modelling nuclear star clusters. Despite the fact that these systems are not tidally truncated in the same way as clusters on an orbit around the galaxy centre, their profiles are well described by lowered isothermal models [e.g. @2014MNRAS.441.3570G]. For a general application to nuclear star clusters, it is desirable to include the effect of the presence of a black hole in the centre, which generates a point-mass potential. provided a method to self-consistently solve King models with an external point-mass potential: this recipe could be used to include the effect of a massive black hole in the models presented here, to make them more versatile in describing nuclear star clusters.
The aim of this project was to introduce models that can be used to describe the phase-space density of stars in tidally limited, mass-segregated star clusters, in any stage of their life-cycle. At early stage, GCs are dense with respect to their tidal density [e.g. @2013MNRAS.432L...1A] and at the present day about half of the GCs is still much denser than their tidal density [@2010MNRAS.401.1832B; @2011MNRAS.413.2509G]. These GCs ought to have a population of stars with radial orbits in their envelopes, either as a left-over of the violent relaxation process during their formation [@1967MNRAS.136..101L], and/or because of two-body ejections from the core [@1972ApJ...173..529S]. In this phase we expect models with high values of $g$, and small $\ra$ to describe GCs well. These models can thus describe GCs with large Jacobi radii, relative to $\rh$. This applies to a large fraction of the Milky Way GC population, and these objects are beyond the reach of King models [@2010MNRAS.401.1832B]. In later stages of evolution, GCs will be more tidally limited, and isotropic, hence we expect $g$ to reduce and $\ra$ to increase during the evolution (up to a value that practically corresponds to having isotropic models). Capturing these variations in GCs properties with continuous parameters has the advantage that these parameters can be inferred from data. This avoids the need of a comparison of goodness-of-fit parameters of different models.
When only surface brightness data are available, it is challenging to distinguish between models with different truncation flavours and pressure anisotropy, because their role has an impact mostly on the low-density outer parts, far from the centre of the cluster, where foreground stars and background stars are dominating. The addition of kinematical data of stars in the outer region of GCs greatly aids in discriminating between models, but this is challenging at present. Precise proper motions ($\lesssim1\,\kms$) can be obtained with the [*Hubble Space Telescope*]{} [[*HST*]{}; e.g. @2006ApJS..166..249M; @2015arXiv150200005W], but the field of view of [*HST*]{} limits observations to the central parts of Milky Way GCs. Radial velocity measurements of stars in the outer parts of GCs are expensive because of the contamination of non-member stars [@2012ApJ...751....6D]. The upcoming data of the ESA-[*Gaia*]{} mission will improve this situation: the availability of all-sky proper motions and photometry measurements will facilitate membership selection, and for several nearby GC the proper motions will be of sufficient quality that they can be used for dynamical modelling and to unveil the properties of the hidden low-energy stars [@2012MNRAS.420.2562A; @2013MmSAI..84...83P; @sollima15]. The models presented in this paper allow for higher level of inference of physical properties of GCs from these upcoming data.
In two forthcoming studies we will compare the family of models to a series of direct $N$-body simulations of the long term evolution of single-mass star clusters [@2016MNRAS.462..696Z] and multimass clusters (Peuten et al., in preparation) evolving in a tidal field.
Acknowledgements {#acknowledgements .unnumbered}
================
MG acknowledges financial support from the European Research Council (ERC-StG-335936, CLUSTERS) and the Royal Society (University Research Fellowship) and AZ acknowledges financial support from the Royal Society (Newton International Fellowship). This project was initiated during the [*Gaia*]{} Challenge (<http://astrowiki.ph.surrey.ac.uk/dokuwiki>) meeting in 2013 (University of Surrey) and further developed in the follow-up meeting in 2014 (MPIA in Heidelberg). The authors are grateful for interesting discussions with the [*Gaia*]{} Challenge participants, in particular Antonio Sollima, Anna Lisa Varri, Vincent Hénault-Brunet, and Adriano Agnello. Miklos Peuten and Eduardo Balbinot are thanked for doing some of the testing of and Maxime Delorme for suggestions that helped to improve the code. We thank Giuseppe Bertin for comments on an earlier version of the manuscript and the anonymous referee for constructive feedback. Our model is written in the programming language and the following open source modules are used for the code and for the analyses done for this paper: [^6], [^7], [^8].
P. E. R., [Gieles]{} M., 2013, , 432, L1
J., [Evans]{} N. W., [Deason]{} A. J., 2012, , 420, 2562
H., 2001, , 325, 1323
H., [Parmentier]{} G., [Gieles]{} M., [Vesperini]{} E., 2010, , 401, 1832
M., [Bragaglia]{} A., [Carretta]{} E., [Gratton]{} R. G., [Lucatello]{} S., [Catanzaro]{} G., [Leone]{} F., 2012, , 538, A18
A., [Anderson]{} J., [van der Marel]{} R. P., [Watkins]{} L. L., [King]{} I. R., [Bianchini]{} P., [Chanam[é]{}]{} J., [Chandar]{} R., [Cool]{} A. M., [Ferraro]{} F. R., [Ford]{} H., [Massari]{} D., 2014, , 797, 115
G., 2014, [Dynamics of Galaxies]{}. Cambridge University Press, Cambridge
G., [Varri]{} A. L., 2008, , 689, 1005
P., [Varri]{} A. L., [Bertin]{} G., [Zocchi]{} A., 2013, , 772, 67
J., 1982, , 20, 399
J., [Tremaine]{} S., 1987, [Galactic Dynamics]{}. Princeton University Press, Princeton, NJ
L., [Idczak]{} D., 2014, preprint (arXiv:1402.0319)
J. A., [Gieles]{} M., [Sollima]{} A., [Koposov]{} S., [Mart[í]{}nez-Delgado]{} D., [Pe[ñ]{}arrubia]{} J., 2012, , 419, 14
S., 1939, [An Introduction to the Study of Stellar Structure]{}. [Chicago Univ. Press, Chicago, IL]{}
H., 1980, , 242, 765
G. S., 2012, , 751, 6
G. S., [Freeman]{} K. C., 1976, , 206, 128
E., 1977, , 61, 391
A. S., 1915, , 75, 366
T., [Heggie]{} D. C., 2000, , 318, 753
I. Y., [B[ö]{}ker]{} T., 2014, , 441, 3570
M., [Heggie]{} D. C., [Zhao]{} H., 2011, , 413, 2509
Y. J., [Velazquez]{} L., 2014, J. Stat. Mech.: Theory Exp., 4, 6 (GV14)
J. E., [Griffin]{} R. F., 1979, , 84, 752
Hairer E., N[ø]{}rsett S., Wanner G., 1993, Solving Ordinary Differential Equations I: Nonstiff Problems. Springer-Verlag, Berlin
D., [Hut]{} P., 2003, [The Gravitational Million-Body Problem: A Multidisciplinary Approach to Star Cluster Dynamics]{}. Cambridge Univ. Press, Cambridge
D. C., [Ramamani]{} N., 1995, , 272, 317
M., 1961, Ann. Astrophys., 24, 369; English translation: preprint (arXiv:1103.3499)
M. H., 1971, , 14, 151
C., 1977, , 82, 271
R., [Nipoti]{} C., [Sollima]{} A., [Bellazzini]{} M., [Chapman]{} S. C., [Dalessandro]{} E., 2013, , 428, 3648
J., [Horwitz]{} G., [Dekel]{} A., 1978, , 223, 299
I. R., 1966, , 71, 64
D., 1967, , 136, 101
D., [Wood]{} R., 1968, , 138, 495
D. E., [Anderson]{} J., [Meylan]{} G., [Gebhardt]{} K., [Pryor]{} C., [Minniti]{} D., [Phinney]{} S., 2006, , 166, 249
D. E., [van der Marel]{} R. P., 2005, , 161, 304
M., [Pucacco]{} G., [Vesperini]{} E., 1998, , 301, 25
D., 1981, , 86, 318
G., 1987, , 184, 144
G., [Heggie]{} D. C., 1997, A&AR, 8, 1
R. W., 1963, , 125, 127
P., 2006, , 366, 227
P., 2010, , 514, A52
K. S., [Lin]{} D. N. C., 1992, , 386, 519
T., 1989, , 71, 651
E., [Bellazzini]{} M., [Marinoni]{} S., 2013, , 84, 83
J.-C., [De Marco]{} O., [Fryer]{} C. L., [Herwig]{} F., [Diehl]{} S., [Oishi]{} J. S., [Mac Low]{} M.-M., [Bryan]{} G. L., [Rockefeller]{} G., 2012, , 744, 52
H. C., 1911, , 71, 460
V. L., [Shukhman]{} I. G., 1981, , 25, 533
C., [Meylan]{} G., 1993, in [Djorgovski]{} S. G., [Meylan]{} G., eds, ASP Conf. Ser. Vol. 50, [Structure and Dynamics of Globular Clusters]{}. Astron. Soc. Pac. San Francico, p. 357
Samko S. G., Kilbas A. A., Marichev O. I., 1993, Fractional integrals and derivatives: theory and applications. Gordon & Breach, Philadelphia, PA
R. L., [Gieles]{} M., 2015, , 448, L94
A., [Baumgardt]{} H., [Zocchi]{} A., [Balbinot]{} E., [Gieles]{} M., [Henault-Brunet]{} V., [Varri]{} A. L., 2015, MNRAS, 451, 2185
A., [Bellazzini]{} M., [Lee]{} J.-W., 2012, , 755, 156
L., 1987, [Dynamical Evolution of Globular Clusters]{}. Princeton, NJ, Princeton University Press, 1987
Jr. L., [Shapiro]{} S. L., 1972, , 173, 529
A., [Sakagami]{} M.-a., 2005, , 364, 990
M., [van der Marel]{} R., 2013, , 435, 3272
A. L., [Bertin]{} G., 2009, , 703, 1911
A. L., [Bertin]{} G., 2012, , 540, A94
L. L., [van der Marel]{} R. P., [Bellini]{} A., [Anderson]{} J., 2015, , 803, 29
C. P., 1975, , 80, 175
R. V. D. R., 1954, , 114, 191
A., [Bertin]{} G., [Varri]{} A. L., 2012, , 539, A65
Zocchi A., Gieles M., H[é]{}nault-Brunet V., Varri A. L. 2016, , 462, 696
Derivations {#app:series}
===========
The DF introduced in equation (\[eq:dfani\]) can be expressed as a function of the dimensionless quantities $\phihat$, $\khat$ and $\phat$ defined in Sections \[ssec:scaling\] and \[ssec:ani\] as $$f = A \exp\left(-\khat \phat^2 \sin^2\theta\right)\Eg\left(g, \phihat - \khat\right).$$ We want to calculate, for these models, the density and velocity dispersion components. We recall that these quantities can be obtained from the DF in the following way:[^9] $$\begin{aligned}
\rho &= \int \dr^3 v \, f, \\
\sigma^2_{i} &= \frac{1}{\rho} \int \dr^3 v \, f v^2_i, \label{DF_vel_disp_i}\end{aligned}$$ where the subscript $i$ denotes the $i$-th component of the velocity vector. To carry out these integrals of the DF in the three-dimensional velocity volume we can use the dimensionless variable $\khat$ and the variable $t = \cos \theta$ $$\dr^3 v = \dr v \dr\theta \dr\varphi \, v^2 \sin\theta = - \dr \khat \dr t \dr\varphi \, \sqrt{\khat} (2 s^2)^{3/2}.$$ In calculating the relevant quantities mentioned above, we encounter the following integrals with respect to the variable $t$ $$\begin{aligned}
&\int_0^{1} \dr t \, \exp\left[-\khat \phat^2 (1-t^2)\right]
= \frac{F\left(\sqrt{\khat}\phat\right)}{\sqrt{\khat}\phat} , \label{Itheta1} \\
&\int_0^{1} \dr t \, t^2 \exp\left[-\khat \phat^2 (1-t^2)\right]
= \frac{1}{2 \khat\phat^2} - \frac{F\left(\sqrt{\khat}\phat\right)}{2 (\sqrt{\khat}\phat)^3}, \label{Itheta2} \end{aligned}$$ where $F(x)$ is the Dawson integral, whose properties are presented in Section \[App:Dawson\]. We use the above results to proceed and derive the density and velocity components.
Density profile
---------------
The density is calculated as $$\begin{aligned}
\rho &= \int \dr^3v \, f \nonumber \\
&= \frac{\tilde{A}}{\sqrt{\pi}} \int_0^{\phihat} \dr\khat \, \int_0^{1} \dr t \, \khat^{1/2}\exp\left[\khat \phat^2 (t^2-1)\right] \Eg\left(g, \phihat-\khat\right) \nonumber \\
&= \tilde{A} \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \frac{F\left(\sqrt{\khat} \phat\right)}{\phat }\Eg\left(g, \phihat-\khat\right) \nonumber\\
&= \tilde{A} \rhoint,\end{aligned}$$ where we replaced $\Gamma(3/2)$ by $\sqrt{\pi}/2$, and we introduced $\tilde{A} = A \left(2 \pi s^2 \right)^{3/2}$ and we solved the integral over $t$ as shown in equation (\[Itheta1\]). The integral $\rhoint$ can be solved by first doing an integration by parts (by using the results in equations \[Egintegral\] and \[DawsDeriv\]) and by then using the convolution formula of equation (\[EG\_convolution\]) and the recurrence relation of equation (\[RecurrencyEg\]) in the following way: $$\begin{aligned}
\rhoint &= \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \Eg\left(g, \phihat-\khat\right) \frac{F\left(\sqrt{\khat} \phat\right)}{\phat} \label{Irho} \\
&= \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \Eg\left(g+1, \phihat-\khat\right)\times\left[ \frac{1}{2\sqrt{\khat}} - \phat F\left(\sqrt{\khat} \phat\right) \right]\nonumber\\
&= \Eg\left(g+{\tfrac{3}{2}}, \phihat\right)
- \phat^2 \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \Eg\left(g, \phihat-\khat \right) \frac{F\left(\sqrt{\khat} \phat\right)}{\phat} \nonumber \\
& \hspace{1cm} +\phat^2 \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \frac{\left(\phihat-\khat\right)^g}{\Gamma(g+1)} \frac{F\left(\sqrt{\khat} \phat\right)}{\phat} \nonumber \\
&= \Eg\left(g+{\tfrac{3}{2}}, \phihat \right) - \phat^2 \rhoint + \phat^2 \fintone.\end{aligned}$$ The integral $\fintone$ can be calculated by substituting the Dawson function for its series representation (see equation \[DawsonSeries\]), by changing variable to $y
= \khat/\phihat$, by using the Beta function of equation (\[Beta\]), and by recognizing the expression in equation (\[1F1DefSer\]) $$\begin{aligned}
\fintone &= \frac{2}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \frac{\left(\phihat-\khat\right)^g}{\Gamma(g+1)} \frac{F\left(\sqrt{\khat} \phat\right)}{\phat } \nonumber \\
&= \frac{\phihat^{g+\frac{3}{2}}\hyp\left(1,g+{\tfrac{5}{2}},-\phat^2\phihat\right)}{\Gamma\left(g+\frac{5}{2}\right)},
\label{IntegralDawson1}\end{aligned}$$ where $ \hyp\left(a,b,x\right)$ is the confluent hypergeometric function (see Section \[1F1\]). Therefore, we can finally write the density integral as $$\rhoint \!= \!\frac{\Eg(g+{\tfrac{3}{2}}, \phihat )}{1+\rrahat^2} + \frac{\phat^2}{1+\phat^2} \frac{\phihat^{g+\frac{3}{2}}\hyp(1,g+\frac{5}{2},-\phat^2\phihat)}{\Gamma\left(g+\frac{5}{2}\right)}.
\label{DENSITY_Anis}$$
Velocity dispersion profiles {#sapp:sigma_deriv}
----------------------------
The velocity dispersion profile can be computed in a similar way as the density, by using again the result in equation (\[Itheta1\]) $$\begin{aligned}
\sigma^2\rho &= \int \dr^3v \, v^2 f \nonumber \\
&= \frac{2\tilde{A} s^2}{\sqrt{\pi}} \int_0^{\phihat} \dr\khat\!\int_0^{1} \dr t \, \exp\left[{-\khat \phat^2 (1-t^2)}\right] \khat^{3/2} \Eg(g, \phihat-\khat ) \nonumber \\
&= \frac{4 \tilde{A} s^2}{\sqrt{\pi} } \int_0^{\phihat} \dr\khat \,\frac{F(\sqrt{\khat} \phat)}{\phat} \khat \Eg(g, \phihat-\khat ) \nonumber \\
&= \tilde{A} s^2 \vsqint.\end{aligned}$$ The integral $\vsqint$ can be solved with an integration by parts, then by using equation (\[EG\_convolution\]), and finally, after having used the recurrence relation of equation (\[RecurrencyEg\]), by recognizing the presence of the integral $\rhoint$ found when calculating the density $$\begin{aligned}
\vsqint &= \frac{4}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \khat \Eg\left(g, \phihat-\khat \right) \frac{F\left(\sqrt{\khat} \phat\right)}{\phat }\label{DefVSQint}\\
&= \frac{4}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \Eg\left(g+1, \phihat-\khat \right) \nonumber \\
& \hspace{1cm}\times \left[\frac{\sqrt{\khat}}{2} + \frac{F\left(\sqrt{\khat} \phat\right)}{\phat } -\khat\phat^2 \frac{F\left(\sqrt{\khat} \phat\right)}{\phat} \right] \nonumber \\
&= \Eg\left(g+{\tfrac{5}{2}}, \phihat \right) + 2 \rhoint_{g+1} + \phat^2 \left(\finttwo - \vsqint \right),\end{aligned}$$ where $\rhoint_{g+1} = \rhoint(g+1, \rrahat, \phihat)$. We then solve the integral $\finttwo$ in a similar way as we did for $\fintone$, to get $$\begin{aligned}
\finttwo &= \frac{4}{\sqrt{\pi}}\int_0^{\phihat} \dr\khat \, \khat \frac{F\left(\sqrt{\khat} \phat\right)}{\phat } \frac{\left(\phihat-\khat\right)^g}{\Gamma(g+1)} \nonumber \\
&= \frac{\phihat^{g+\frac{5}{2}}}{\Gamma\left(g+{\tfrac{7}{2}}\right)} \left[\hyp\left(1,g+{\tfrac{7}{2}},-\phat^2\phihat\right) \right. \nonumber \\
& \left. + 2 \hyp\left(2,g+{\tfrac{7}{2}},-\phat^2\phihat\right) \right].
\label{IntegralDawson2}\end{aligned}$$ Therefore, we finally have $$\begin{aligned}
\sigma^2\rho &= \frac{\tilde{A} s^2}{(\phat^2+1)} \left\lbrace \Eg\left(g+{\tfrac{5}{2}}, \phihat \right)\left(\frac{3+\phat^2}{1+\phat^2}\right) \right. \nonumber \\
& \left. + \frac{\phat^2 \phihat^{g+{\tfrac{5}{2}}}}{\Gamma\left(g+\frac{7}{2}\right)} \left[ 2 \, \hyp\left(2,g+{\tfrac{7}{2}},-\phat^2\phihat\right) \right.\right. \nonumber \\
& \left.\left. + \hyp\left(1,g+{\tfrac{7}{2}},-\phat^2\phihat\right) \left(\frac{3+\phat^2}{1+\phat^2} \right) \right] \right\rbrace.
\label{DISP_Anis}\end{aligned}$$ The radial component of the velocity dispersion is given by $$\begin{aligned}
\sigmar^2\rho &= \int \dr^3 v \, (v \cos\theta)^2 f \nonumber \\
&= \frac{2\tilde{A} s^2}{\sqrt{\pi}} \int_0^{\phihat} \dr\khat \, \khat^{3/2} \Eg\left(g, \phihat-\khat \right) \left[ \frac{1}{\khat\phat^2} - \frac{F\left(\sqrt{\khat}\phat\right)}{(\sqrt{\khat}\phat)^3} \right] \nonumber \\
&= \frac{\tilde{A} s^2}{\phat^2} \left[ \Eg\left(g+{\tfrac{3}{2}}, \phihat \right) - \rhoint \right],
\label{DISP_r_Anis_int}\end{aligned}$$ where we solved the integral by using equations (\[Itheta2\]) and (\[EG\_convolution\]). We point out that we can express this quantity by means of the density integrals of the isotropic and the anisotropic case (see also equations \[eq:rhointiso\] and \[eq:rhointani\] in Section \[sec:model\]). We finally obtain $$\sigmar^2\rho = \tilde{A} s^2 \left[ \frac{\Eg\left(g+{\tfrac{3}{2}}, \hat{\phi} \right)}{(1+\hat{p}^2)} - \frac{\hat{\phi}^{g+{\tfrac{3}{2}}} \hyp\left(1,g+{\tfrac{5}{2}},-\hat{p}^2\hat{\phi}\right)}{(1+\hat{p}^2) \Gamma\left(g+\frac{5}{2}\right)} \right],
\label{DISP_r_Anis}$$ which, by using equations (\[RecurrencyEg\]) and (\[RecurrencyHyp2\]), can be rewritten as $$\thickmuskip=2mu
\medmuskip=2mu
\sigmar^2\rho = \tilde{A} s^2 \left[ \frac{\Eg\left(g+{\tfrac{5}{2}}, \hat{\phi} \right)}{(1+\hat{p}^2)} + \frac{\phat^2 \hat{\phi}^{g+{\tfrac{5}{2}}} \hyp\left(1,g+\frac{7}{2},-\hat{p}^2\hat{\phi}\right)}{(1+\hat{p}^2) \Gamma\left(g+\frac{7}{2}\right)} \right].
\label{DISP_r_Anis_REW}$$ To calculate the tangential component of the velocity dispersion we solve the integral over $t$ by expressing it as the difference between equation (\[Itheta1\]) and equation (\[Itheta2\]), and we carry out an integration by parts $$\begin{aligned}
\sigmat^2\rho &= \int \dr^3 v \, (v \sin\theta)^2 f \nonumber \\
&= \frac{2 \tilde{A} s^2}{\sqrt{\pi}} \int_0^{\phihat} \dr\khat \, \khat^{3/2} \Eg\left(g, \phihat-\khat \right) \nonumber\\
&\hspace{0.25cm}\times\left[\frac{2 F\left(\sqrt{\khat}\phat\right)}{\sqrt{\khat}\phat} - \frac{1}{\khat\phat^2} + \frac{F\left(\sqrt{\khat}\phat\right)}{(\sqrt{\khat}\phat)^3} \right] \nonumber \\
&= \frac{\tilde{A} s^2}{\phat^2} \left[\phat^2 \vsqint - \Eg\left(g+{\tfrac{3}{2}}, \phihat \right) + \rhoint \right] .
\label{DISP_T_Anis_int}\end{aligned}$$ After recognizing the integrals we solved above, we can finally write $$\begin{aligned}
\sigmat^2\rho &= \frac{\tilde{A} s^2}{(1+\phat^2)} \left\lbrace \Eg\left(\gfive, \phihat \right)\frac{2}{(1+\phat^2)} \right. \nonumber \\
& \left. + \, \frac{2\phat^2 \phihat^{g+{\tfrac{5}{2}}}}{\Gamma(g+{\tfrac{7}{2}})} \left[ \frac{\hyp\left(1, g+{\tfrac{7}{2}}, -\phihat\phat^2\right)}{(1+\phat^2)} \right.\right. \nonumber \\
& \left.\left. + \, \hyp\left(2, g+{\tfrac{7}{2}}, -\phihat\phat^2\right) \right] \right\rbrace \ .
\label{DISP_T_Anis_REW}\end{aligned}$$
Derivations using fractional calculus {#app:fractional}
=====================================
Fractional calculus is a branch of mathematics that considers real numbers for the orders of derivatives and integration. Because the integrals that need to be solved contain terms like $\khat^{1/2}$ and $\khat^{3/2}$, we can use semi-derivatives and semi-integrals and integration by parts to solve them. By following @2014arXiv1402.0319B, we define the left- and right-sided Riemann-Liouville fractional integrals of order $\alpha >
0$ of a function $q \in L^1$ as $$\begin{aligned}
I^{\alpha}_{a+} q(t) &= \frac{1}{\Gamma(\alpha)} \int_{a}^{t} \dr x \, q(x) (t-x)^{\alpha-1}, \label{leftFracInt} \\
I^{\alpha}_{b-} q(t) &= \frac{1}{\Gamma(\alpha)} \int_{t}^{b} \dr x \, q(x) (x-t)^{\alpha-1}. \label{rightFracInt}\end{aligned}$$ In the remainder of this section, we will use the result illustrated by @2014arXiv1402.0319B in their Proposition 2 [for a proof see @samko1993] $$\begin{aligned}
\int_{a}^{b} \dr t \, \left(I^{\alpha}_{a+} q_1\right)(t) q_2(t) = \int_{a}^{b} \dr t \, q_1(t) \left(I^{\alpha}_{b-} q_2\right)(t).
\label{Prop2}\end{aligned}$$
Density
-------
When considering the isotropic limit of the DF, the integral to be solved to calculate the density is: $$\rhoint = \frac{2}{\sqrt{\pi}} \int_0^{\phihat}\dr\khat\, \khat^{1/2} \Eg\left(g, \phihat-\khat \right).
\label{densISOFrac1}$$ We can use fractional calculus to solve it, by changing variable of integration (using $x = \phihat-\khat$) and by considering that $$\begin{aligned}
q_1(x) &= \Eg\left(g, x \right), \\
q_2(x) &= 1, \\
I^{1/2}_{0+} q_1(x) &= \Eg\left(g + {\tfrac{1}{2}}, x \right), \\
I^{1/2}_{\phihat-} q_2(x) &= \frac{2(\phihat-x)^{1/2}}{\sqrt{\pi}},\end{aligned}$$ thus obtaining $$\begin{aligned}
\rhoint &= \int_0^{\phihat}\dr x\, \Eg(g +{\tfrac{1}{2}}, x) = \Eg(g +{\tfrac{3}{2}}, \phihat),\end{aligned}$$ which was solved by using equation (\[Egintegral\]).
The density of the anisotropic models is calculated by solving the integral $\rhoint$ introduced in equation (\[Irho\]). To do this, we can use the result shown in equation (\[Prop2\]) by noticing (see equations \[EG\_convolution\] and \[DawsonDef\]) that $$\begin{aligned}
q_1(\khat) &= \exp\left(-\khat\phat^2\right), \label{q11}\\
q_2(\khat) &= \Eg\left(g, \phihat-\khat\right), \\
I^{1/2}_{0+} q_1(\khat) &= \frac{2F(\sqrt{\khat}\phat)}{\sqrt{\pi} \phat}, \label{q1i1} \\
I^{1/2}_{\phihat-} q_2(\khat) &= \Eg\left(g + {\tfrac{1}{2}}, \phihat-\khat \right),\end{aligned}$$ and by rewriting the integral of equation (\[Irho\]) as $$\rhoint = \int_0^{\phihat} \dr \khat \, \exp\left(-\khat\phat^2\right) \Eg\left(g + {\tfrac{1}{2}}, \phihat-\khat \right).
\label{densFrac1}$$ This integral can be solved by parts, by using the expression of the derivative of the lower incomplete gamma function (equation \[GammaDeriv\]) and by using equation (\[IntDefHyp\]) to obtain $$\rhoint = \frac{\Eg\left(g+{\tfrac{1}{2}}, \phihat \right)}{1+\phat^2} - \frac{\phihat^{g+{\tfrac{1}{2}}}\hyp(1,g+{\tfrac{3}{2}}, -\phihat\phat^2)}{(1+\phat^2)\Gamma(g+{\tfrac{3}{2}})}.$$ The last step can be rewritten as equation (\[DENSITY\_Anis\]) by using the recurrence relations shown in equations (\[RecurrencyEg\]) and (\[RecurrencyHyp2\]).
Velocity dispersion {#sapp:sigma_deriv_frac}
-------------------
In the isotropic limit of the DF, the velocity dispersion is calculated by means of an integral with the same structure as the one found in equation (\[densISOFrac1\]). The velocity dispersion of the anisotropic models is calculated by solving the integral $\vsqint$ introduced in equation (\[DefVSQint\]). We can use the result shown in equation (\[Prop2\]) also in this case, by considering the function $q_1$ introduced in equation (\[q11\]), its fractional integral (equation \[q1i1\]), and $$\begin{aligned}
\thickmuskip=2mu
\medmuskip=2mu
q_2(\khat) &= \khat \, \Eg\left(g, \phihat-\khat \right), \\
I^{1/2}_{\phihat-} q_2(\khat) &= \frac{1}{2} \Eg\left(g+{\tfrac{3}{2}}, \phihat-\khat \right) + \khat \Eg\left(g+{\tfrac{1}{2}}, \phihat-\khat \right),\end{aligned}$$ and by rewriting the integral $\vsqint$ as $$\begin{aligned}
\vsqint &= \int_0^{\phihat} \dr \khat \exp\left(-\khat\phat^2\right) \Eg\left(g +{\tfrac{3}{2}}, \phihat-\khat \right) \nonumber\\
& +2 \int_0^{\phihat} \dr \khat \exp\left(-\khat\phat^2\right) \khat \Eg\left(g +{\tfrac{1}{2}}, \phihat-\khat \right).\end{aligned}$$ The first integral is in the same form as the one we found for the density, and the second one can be reduced to something similar with an integration by parts. By solving the integrals, we obtain: $$\begin{aligned}
\vsqint &= \frac{1}{1+\phat^2 } \left[ \Eg\left(g+{\tfrac{3}{2}}, \phihat\right)\left(\frac{3+\phat^2}{1+\phat^2}\right) - \frac{2 \phihat^{g+{\tfrac{3}{2}}}}{\Gamma\left(g+{\tfrac{3}{2}}\right)} \right. \nonumber\\
& \left. +\frac{2 \phihat^{g+{\tfrac{3}{2}}}\hyp(1,g+\frac{5}{2}, -\phihat\phat^2)}{\Gamma\left(g+\frac{5}{2}\right)} \left( g + \phat^2\phihat + \frac{\phat^2}{1+\phat^2} \right) \right],\end{aligned}$$ which can be rewritten as equation (\[DISP\_Anis\]) by using the recurrence relations shown in equations (\[RecurrencyEg\]), (\[RecurrencyHyp1\]), and (\[RecurrencyHyp2\]).
By inspecting equations (\[DISP\_r\_Anis\_int\]) and (\[DISP\_T\_Anis\_int\]), it is immediate to notice that the radial and tangential components of the velocity dispersion are calculated with integrals that can be written as a combination of those solved in this section by means of fractional calculus.
Differential energy distribution {#app:dmde}
================================
The differential energy distribution gives the amount of mass per units of energy [@BT1987]. Here we briefly recall how to calculate it for isotropic and anisotropic systems.
For a DF that only depends on $E$, the differential energy distribution can be calculated as: $$\frac{\dr M}{\dr E} \equiv f(E)\gdos(E),
\label{eq:dmde}$$ where $f(E)$ is the DF, and $\gdos(E)$ is the density of states, which is the volume of phase space per unit energy and is defined as $${\gdos}(E) = \int \dr^3r \, \dr^3v \, \delta(E-H),
\label{densityofstates}$$ where $\delta(x)$ is the Dirac delta function. For a spherically symmetric systems, this integral can be expressed as $$\thickmuskip=2mu
\medmuskip=2mu
\gdos(E) = 16 \pi^2 \, \int_0^{\rmm(E)} \dr r \, r^2 \, \int \dr v \, v^2 \, \delta\left(\frac{1}{2} v^2 + \phi - E\right),
\label{densityofstates_1}$$ where $\rmm(E)$ is the radius at which $\phi = E$. By changing variable (using $y = v^2/2$), we finally obtain $$\gdos(E) = 16 \pi^2 \, \int_0^{\rmm(E)} \dr r \, r^2 \, \sqrt{2(E - \phi)},
\label{densityofstates_2}$$ and the differential energy distribution is therefore $$\frac{\dr M}{\dr E} = 16 \pi^2 \, f(E) \, \int_0^{\rmm(E)} \dr r\, r^2 \, \sqrt{2(E-\phi)}.
\label{N_E_iso}$$
When considering anisotropic systems, for which the DF depends also on the angular momentum $J$, the differential energy distribution is obtained as $$\frac{\dr M}{\dr E} = \int \dr^3r \, \dr^3v \, \delta(E-H) f(H,J).
\label{differentialenergydistribution}$$ This integral can be expressed as $$\frac{\dr M}{\dr E}= 8 \pi^2 \int \dr r \, r^2 \int \dr \vr \, \dr \vt \, \vt \, \delta(E-H) f(H,J),$$ and it can then be rearranged by changing variable and introducing $J = r \, \vt$ in the following way $$\frac{\dr M}{\dr E} = 8 \pi^2 \int \dr r \int \dr \vr \, \dr J \, J \, \delta(E-H) f(H,J).$$ The integral over $\vr$ is solved by using the fact that $$\vr^2 = 2(E-\Phi) - \frac{J^2}{r^2},$$ to obtain $$\begin{aligned}
\frac{\dr M}{\dr E} = 16 \pi^2 \, \int \dr r \int \dr J \, \frac{J f(E,J)}{\sqrt{2(E-\Phi) - J^2/r^2}}.\end{aligned}$$ By using the expression for the DF of the models presented in this paper (equation \[eq:dfani\]) we can perform the integration over $J$ in this last equation and write the differential energy distribution as a function of the part of the DF that depends on energy only, $f(E)$: $$\begin{aligned}
\frac{\dr M}{\dr E}= 16 \pi^2 \, f(E) \, \int_0^{\rmm(E)}\dr r\, \sqrt{2} \, \ra s \, r \, F\left(\frac{r \sqrt{E-\phi}}{\ra s}\right),\end{aligned}$$ where $F(x)$ is the Dawson integral (see Appendix \[App:Dawson\]). In the limit of $\ra\rightarrow\infty$ this reduces to the result for the isotropic case shown in equation (\[N\_E\_iso\]), which follows from substituting the leading term of equation (D14) in equation (C11).
Useful properties of mathematical functions {#app:functions}
===========================================
Useful properties of the gamma functions {#app:gamma}
----------------------------------------
The gamma function of a positive integer $n$ is defined as $$\Gamma(n) = (n-1)!,$$ while for non-integer arguments $a$, it can be written as an integral $$\Gamma(a) = \int_0^{\infty} \dr t \, t^{a-1}\exp(-t).$$ The lower incomplete gamma function is given by $$\gamma(a, x) = \int_0^x \dr t \, t^{a-1}\exp(-t),$$ and its derivative is $$\frac{\dr \gamma(a,x)}{\dr x} = x^{a-1}\exp(-x).
\label{GammaDeriv}$$
Useful properties of the $\Eg(a,x)$ function {#AppD:Eg}
--------------------------------------------
The exponential function $\Eg(a,x)$ is defined as $$\Eg(a,x) = \frac{1}{\Gamma(a)} \int_0^x \dr t \, t^{a-1} \exp(x-t),$$ and an alternative expression is given by means of the lower incomplete gamma function $$\Eg(a,x) = \frac{\exp(x) \gamma(a,x)}{\Gamma(a)}.
\label{E_gamma_exp}$$ The series representation of this function is $$\Eg(a,x) = \sum_{i=0}^{\infty} \frac{x^{i+a}}{\Gamma(i+a+1)}.
\label{Eg_series_repr}$$ The following recurrence relation holds $$\Eg(a,x) = \Eg(a+1, x) + \frac{x^a}{\Gamma(a+1)}.
\label{RecurrencyEg}$$ The derivative and the integral of $\Eg(a,x)$ are given by $$\begin{aligned}
\frac{\dr \Eg(a, x)}{\dr x} &= \Eg(a-1,x), \\
\int \dr x \, \Eg(a, x) &= \Eg(a+1,x) + \mathrm{constant}. \label{Egintegral}\end{aligned}$$ A proof of these equations can be easily obtained by writing $\Eg(a,x)$ as in equation (\[E\_gamma\_exp\]), and by considering equation (\[GammaDeriv\]) and the recurrence relation (equation \[RecurrencyEg\]). The convolution formula $$\frac{1}{\Gamma(b)} \int_0^x \dr y \, \Eg(a,x-y) y^{b-1} = \Eg(a+b,x)
\label{EG_convolution}$$ can be obtained by using the series representation of $\Eg(a,x)$ (equation \[Eg\_series\_repr\]) and by changing variable, to express the integral with a form that allows us to recognize the Beta function: $$B(m,n) = \int_0^1 \dr y \, (1-y)^{m-1}y^{n-1} = \frac{\Gamma(m)\Gamma(n)}{\Gamma(m+n)}.
\label{Beta}$$ The identity of equation (\[EG\_convolution\]) accounts for the fractional integration of $\Eg(a,x)$ (see equation \[leftFracInt\]).
Useful properties of the Dawson integral {#App:Dawson}
----------------------------------------
The Dawson integral (sometimes called the Dawson function) is defined as $$F(x) = \exp(-x^2) \int_0^x \dr y \exp(y^2).
\label{DawsonDef}$$ It is also possible to express $F(x)$ as a sum as $$F(x) = \sum_{i=0}^{\infty} \frac{(-1)^i x^{2i+1} \Gamma\left({\tfrac{3}{2}}\right)}{\Gamma\left(i+{\tfrac{3}{2}}\right)} \ .
\label{DawsonSeries}$$ The Dawson integral is an odd function, and its derivative is $$\frac{\dr F(x)}{\dr x} = 1 - 2 x F(x).
\label{DawsDeriv}$$
Useful properties of the confluent hypergeometric function {#1F1}
----------------------------------------------------------
The confluent hypergeometric function is defined as $$\begin{aligned}
\hyp(a,b,x) &= \displaystyle\sum_{i=0}^{\infty} \frac{\Gamma(a+i)}{\Gamma(a)}\frac{\Gamma(b)}{\Gamma(b+i)}\frac{x^i}{\Gamma(i+1)}.
\label{1F1DefSer}\end{aligned}$$ It can also be defined by means of an integral, as $$\hyp(a,b,x) = \frac{\Gamma(b)}{\Gamma(a) \Gamma(b-a)} \int_{0}^{1} \dr y \, \exp(xy) y^{a-1} (1-y)^{b-a-1},
\label{IntDefHyp}$$ which holds for ${\rm Re}(b)>{\rm Re}(a)>0$. The following recurrence relations hold $$\begin{aligned}
\hyp(2,b,x) &= (2-b+x) \, \hyp(1,b,x) + b - 1, \label{RecurrencyHyp1} \\
x \, \hyp(1,b+1,x) &= b \, \hyp(1,b,x) - b. \label{RecurrencyHyp2}\end{aligned}$$ We also note that this function is related to the exponential function, and for $b=a$ we have $\hyp(a,a,x)=\exp(x)$. Another useful property of this function is that $$\hyp(a, b, 0) = 1.
\label{Hyp1}$$
The density (equation \[eq:rhointanires\]) and the velocity moments (equations \[eq:vrsqintres\] - \[eq:vsqintres\]) of anisotropic models are expressed by means of the function $\hyp(a,b,x)$ with $a=1$. When considering $a = 1$ in equation (\[IntDefHyp\]), we obtain $$\begin{aligned}
\hyp(1,b,x) &= \frac{(b-1)}{x^{b-1}} \, \exp(x) \, \gamma(b-1,x) .
\label{1F1_alt}\end{aligned}$$ We point out that when $x<0$ two of the quantities appearing in equation (\[1F1\_alt\]) are imaginary. This is the reason why in general this expression cannot be used to speed up the code by expressing the equations mentioned above in a more compact way. When considering integer values of $b$, however, we can simplify the hypergeometric function, and express it by means of exponentials and polynomials. The smallest value of $b$ we use is $g +{\tfrac{5}{2}}$, therefore $b$ assumes the smallest integer value when $g = {\tfrac{1}{2}}$ for $b=3$ we calculate $$\begin{aligned}
\hyp(1,3,x) &= \frac{2}{x^{2}} \left[ \exp(x) -1 -x\right].\end{aligned}$$ Combined with the recurrence relation mentioned earlier, the results of the anisotropic models and half-integer values of $g$ can be expressed by means of these elementary functions.
The asymptotic series expansion for the confluent hypergeometric function when $|x| \rightarrow \infty$ is given by $$\begin{aligned}
\hyp(a,b,x) & \propto \frac{\Gamma(b)}{\Gamma(b-a)} (-x)^{-a} \left[ 1 + \mathcal{O}\left( \frac{1}{z} \right) \right] \nonumber \\
& +\frac{\Gamma(b)}{\Gamma(a)} \exp(x) x^{a-b} \left[ 1 + \mathcal{O}\left( \frac{1}{z} \right) \right].\end{aligned}$$ This is useful to compute the density and velocity moments, because for very large $|x|$ the evaluation of this function becomes inaccurate in (see section \[ssec:generalimplementation\]). In particular, the functions that are needed to compute these quantities, $\hyp(1,b,-x)$ and $\hyp(2,b,-x)$, have the following behaviour for large $|x|$: $$\begin{aligned}
\lim_{x\rightarrow \infty}\hyp(1,b,-x) &= \frac{b-1}{x},\label{eq:1f1asym1}\\
\lim_{x\rightarrow \infty}\hyp(2,b,-x) &= \frac{(b-1)(b-2)}{x^2}.\label{eq:1f1asym2}\end{aligned}$$
[^1]: is available from <https://github.com/mgieles/limepy.>
[^2]: The Woolley, King and Wilson DFs follow straightforwardly from equations (\[eq:dfani\]) and (\[eq:eg\]), because $\Eg(0,x) = \exp(x)$, $P(1,x) =
1-\exp(-x)$, such that $\Eg(1,x) = \exp(x)-1$ and $P(2,x) =
1-\exp(-x)-x\exp(-x)$, such that $\Eg(2,x) = \exp(x) - 1 - x$.
[^3]: This parameter is called $W_0$ in @1966AJ.....71...64K.
[^4]: By considering the first velocity moment of the DF we find the mean velocity: for these models, this quantity vanishes everywhere. We also note that expressions for higher order moments of the velocity distribution can be derived, but these are beyond the scope of this paper.
[^5]: The tangential velocity comprises the two components $\vt^2 = \vtheta^2+\vphi^2$, where $\vtheta =
v\sin\theta\sin\varphi$ and $\vphi = v\sin\theta\cos\varphi$. The corresponding components of the velocity dispersion tensor are equal to each other, and each of them accounts for half of the tangential component: $\sigmat^2 = 2 \sigmatheta^2 = 2\sigmaphi^2$.
[^6]: <http://www.numpy.org>
[^7]: <http://www.scipy.org>
[^8]: <http://matplotlib.sourceforge.net>
[^9]: As noticed already in Section \[ssec:iso\], equation (\[DF\_vel\_disp\_i\]) holds because the mean velocity for the systems described by these models is zero everywhere.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The polarized Raman scattering spectra from freshly cleaved $ab$, $ac$, and $bc$ surfaces of high quality twin free YBa$_2$Cu$_3$O$_{6.5}$ (Ortho-II) single crystals ($T_c$=57.5 K and $\Delta T = 0.6$ K) were studied between 80 and 300 K. All eleven $A_g$ Raman modes expected for the Ortho-II structure as well some modes of $B_{2g}$ and $B_{3g}$ symmetry were identified in close comparison with predictions of lattice dynamical calculations. The electronic scattering from the $ab$ planes is strongly anisotropic and decreases between 200 and 100 K within the temperature range of previously reported pseudogap opening. The coupling of phonons to Raman active electronic excitations manifested by asymmetric (Fano) profiles of several modes also decreases in the same range. Among the new findings that distinguish the Raman scattering of Ortho-II from that of Ortho-I phase is the unusual relationship ($\alpha_{xx} \approx
-\alpha_{yy}$) between the elements of the Raman tensor of the apex oxygen $A_g$ mode.
author:
- 'M. N. Iliev, V. G. Hadjiev'
- 'S. Jandl, D. Le Boeuf'
- 'V. N. Popov'
- 'D. Bonn, R. Liang, W. N. Hardy'
title: 'Raman study of Twin Free YBa$_2$Cu$_3$O$_{6.5}$ (Ortho-II) Single Crystals '
---
Introduction
============
The properties of underdoped YBa$_2$Cu$_3$O$_x$ ($6 < x < 7$) have been studied intensively within the efforts to unravel the mechanism of high temperature superconductivity. The Ortho II phase, corresponding to oxygen content $x = 6.5$ and characterized by alternating full and empty chains, has attracted particular attention as it is both underdoped and free of disorder. Although the existence of this phase has been documented experimentally in considerable number of reports, it has also been established that as a rule the Ortho II domains coexists with domains of different oxygen ordering or/and strong disorder even in the case $x\simeq
6.5$. An improved procedure developed by Liang at al.[@liang2000] has made possible preparation of twin free, highly ordered Ortho II single crystals. Such crystals have recently been used in several studies of specific properties of the Ortho II phase by means of neutron scattering[@stock2002; @stock2004], x-ray diffraction[@zimmermann2003], time-resolved spectroscopy[@gedik2004; @gedik2004a], infrared spectroscopy[@hwang2006], nuclear magnetic resonance[@yamani2006], microwave spectroscopy[@harris2006], and resistance measurements in high magnetic fields.[@leyraud2007].
There have been several attempts to identify the Raman modes of the Ortho-II phase by measuring the spectra of oxygen deficient YBa$_2$Cu$_3$O$_x$ ($x\approx 6.5$) single crystals.[@iliev1993; @misochko1994; @iliev1996] Compared to the well known Raman spectra of Ortho-I ($x=7$) and T ($x=6$) phases, in the case of Ortho-II one expects shift of the corresponding Raman modes and activation of additional modes due to the doubling of the unit cell. A significant number of additional modes have been observed, but their identification has met definite difficulties due to ambiguities in the local structure of YBa$_2$Cu$_3$O$_x$ even in the case $x=6.5$. Indeed, in the idealized Ortho-I, Ortho-II and T structures the Cu1 and O1 atoms in the basal Cu-O planes are at centrosymmetrical sites and their vibrations are not Raman active. The oxygen arrangement in the basal planes of a real YBa$_2$Cu$_3$O$_x$ material, however, is characterized by chain fragments instead of infinite chains and part of oxygen atoms are outside the chains in otherwise vacant O5 sites. This creates a number of local non-centrosymmetrical surroundings for Cu and O atom, in particular those at the end of chain fragments, and activates their vibrations in the Raman spectrum.[@iliev1996; @falques1997] Some of these defects can also be produced by local laser annealing[@iliev1997; @iliev1997a] or photoactivation.[@osada2005] Another issue, as a rule neglected but of particular importance in the case of Ortho-II phase, is the possible lack of correspondence between the oxygen content and arrangement in the volume of the crystal and its surface layer(s), where the Raman scattering occurs. The cause for such discrepancy is the in- and out-diffusion of oxygen, which depends strongly on the type of the surface ($ab$, $bc$ or $ac$), starting oxygen content, ambient oxygen pressure, temperature and exposure time.[@conder2000]
In this paper we present results of a temperature-dependent Raman study on freshly-cleaved $ab$, $ac$, and $bc$ surfaces of a twin free YBa$_2$Cu$_3$O$_{6.5}$ (YBCO6.5) single crystal with high degree of Ortho-II type ordering. This allowed us with great certainty to identify the proper Raman modes of the Ortho-II phase and assign them to definite atomic motions via close comparison with predictions of lattice dynamical calculations, as well as to measure the symmetry and strength of electronic scattering.
Samples and Experimental
========================
We used a high quality mechanically de-twinned Ortho-II YBCO6.5 single crystal with $T_c$=57.5 K and $\Delta T =0.6~K$, grown by a flux method in BaZrO3 crucible.[@liang2000; @leyraud2007] Immediately before mounting the sample on the cold finger of a Microstat$\circledR$He (Oxford Instruments) optical cryostat, a small area of the surface ($ab$, $ac$, or $bc$) to be used for Raman measurements was cleaved out to ensure that the scattering volume has an Ortho-II oxygen arrangement. The Raman spectra were measured under microscope ($\times 50$ magnification) using a triple T64000 (Horiba Jobin Yvon) spectrometer. In most experiments we used 633 nm excitation with less than 1 mW incident laser power focused at a spot of 2-3 $\mu$m diameter. Comparative measurements with 515, 488, and 458 nm excitations were also done. All spectra were corrected for the Bose factor. For description of the scattering configurations we use the Porto’s notation $a(BC)d$, where the first and fourth letters denote, respectively, the directions of incident and scattered light in a Cartesian $xyz$ system with axes along the crystallographic directions. The polarization of the incident and scattered light is given by the second and third letters, respectively.
Results and Discussion
======================
Raman phonons in the Ortho-II phase
-----------------------------------
In Figure 1 are compared the Raman spectra obtained with 633 nm excitation at room temperature from freshly cleaved and aged areas on the $ab$ and $bc$ surfaces of the same Ortho-II crystal. The spectral shapes and Raman line frequencies from aged surfaces are consistent with those from earlier reports on the $XX/YY$ and $ZZ$ spectra of twinned YBCO6.5[@misochko1994] and YBCO$x$ ($x\approx
6.5$ samples.[@palles1996; @hong2007] The corresponding spectra from freshly cleaved and aged surfaces, however, exhibit definite differences in the positions and appearance of some of the peaks.
{width="8cm"}
Figure 2 shows the Raman spectra from freshly cleaved surfaces at 90 K in all available exact scattering configurations. From symmetry considerations one expects for the Ortho-II structure an increased number of Raman active modes, $11A_g + 4B_{1g} + 11B_{2g} +
8B_{3g}$, compared to the $5A_g + 5B_{2g} + 5B_{3g}$ modes of the Ortho-I structure. The atomic displacements of the eleven $A_g$ modes, as predicted by lattice dynamical calculations, (LDC)[@iliev1996] are shown in Figure 3. For most modes, the experimentally observed frequencies (at 90 K) are in good agreement with those predicted by LDC, shown in parenthesis. In addition to the five $A_g$ modes at 126, 147, 342, 447, and 487 cm$^{-1}$ corresponding to displacements along the $c$-axis of Ba, Cu2, O2-O3, O2+O3, O4 in the Ortho-I structure, there are six more $A_g$ modes in the extended Ortho-II cell. Four of them can be positively identified at 107 cm$^{-1}$ \[Cu2($z$)-Cu2’($-z$) out-of-phase\], 171 cm$^{-1}$ \[Y($x$)\], 352 cm$^{-1}$ \[O4($z$)-O4’($-z$) out-of-phase\], and 579 cm$^{-1}$ \[O2($x$)\]. The mixed Ba/Cu mode, predicted at 146 cm$^{-1}$, may be very weak and not observable. The weak Raman line at 381 cm$^{-1}$ can be tentatively assigned to the mode involving mainly O4’ displacements along $c$, with a predicted frequency of 414 cm$^{-1}$. In the $ZZ$ spectra one observes additional lines of $A_g$ character at 224 and 601 cm$^{-1}$. These positions are close to those of defect modes related to displacements of Cu1 and O1 at the end of broken chains,[@iliev1996; @iliev1997; @hong2007] which could preexist or be created by photoactivation in our Ortho-II sample.
{width="8cm"}
{width="8cm"}
In the $XZ$($B_{2g}$) and $YZ$($B_{3g}$) spectra one observes relatively strong Raman peaks at 199 and 327 cm$^{-1}$, respectively. The closest $B_{2g}$ (200 and 213 cm$^{-1}$) and $B_{3g}$ (300 cm$^{-1}$) frequencies predicted by LDC, correspond to modes involving mainly displacements of O4 and O4’ along $a$ and $b$ directions, respectively.
A more careful look at the appearance of the apex oxygen mode near 487 cm$^{-1}$ in the $X'X'$ and $X'Y'$ spectra of Fig. 2 reveals a significantly different behavior compared to that in the corresponding spectra of the Ortho-I phase. Indeed, with these scattering configurations, the intensity of an $A_g$ Raman line is proportional to $(\alpha_{xx} + \alpha_{yy})^2$ and $(\alpha_{xx} -
\alpha_{yy})^2$, respectively, where $\alpha_{xx}$ and $\alpha_{yy}$ are diagonal elements of the corresponding Raman tensors. Negligible intensity in the $X'X'$ spectra may be expected if $\alpha_{xx}
\approx -\alpha_{yy}$, which in the Ortho-I phase is satisfied only for the out-of-phase O2-O3 mode at $\approx 336$ cm$^{-1}$, but not for the apex oxygen mode near 500 cm$^{-1}$ (seen in the $X'X'$ but not in the $X'Y'$ spectrum). Here, for the Ortho-II phase the apex oxygen mode has an appearance similar to that of the out-of-phase mode: it is practically not seen in the $X'X'$ spectrum, but well pronounced in the $X'Y'$ spectrum with an intensity comparable to that in the $XX$ and $YY$ spectra. This allows us to conclude that for the apex oxygen mode of the Ortho-II phase, the relation $\alpha_{xx} \approx -\alpha_{yy}$ is satisfied.
Electronic scattering and electron-phonon interaction
-----------------------------------------------------
In addition to discrete phonon lines, the Raman spectra of Ortho-II phase contain structureless background with intensity stronger with $YY$ than with $XX$, and negligible intensity with $ZZ$ polarization (Fig.4). Such background scattering, observed in the normal and superconducting states of high $T_c$ superconductors has been attributed to electronic scattering and has been intensively studied both experimentally and theoretically.[@strohm1997; @bock1999; @opel2000; @devereaux2007] It has been shown that for optimally doped YBCO the electronic scattering is practically independent of temperature for $T > T_c$. With the opening of a superconducting gap $\Delta$ at $T < T_c$, there is a redistribution of electronic scattering intensity from lower to higher energies and a maximum associated with pair-breaking is formed at $\omega_{max} =
2\Delta$. For underdoped YBCO a slight decrease of electronic scattering intensity at $\omega < 600$ cm$^{-1}$ has been observed below a characteristic temperature $T^*$ well above $T_c$ and this has been considered a manifestation of a pseudogap opening.[@opel2000] There are to our knowledge no reports on the variation of electronic scattering near $T^*$.
With few exceptions,[@krantz1995; @strohm1997] in experimental studies of the electronic scattering in YBCO, twinned samples have been used and the theoretical models are based on the tetragonal approximation for the crystal structure.
{width="8cm"}
Figs. 4 and 5 illustrate that the electronic scattering from the $ab$ surface of the Ortho-II crystal is strongly anisotropic and temperature dependent below 600 cm$^{-1}$. Its intensity in the $z(YY)\bar{z}$ spectra is stronger by a factor 2 than in the $z(XX)\bar{z}$ and much stronger than in the $z(XY)\bar{z}$, $z(X'Y')\bar{z}$, $x(ZZ)\bar{x}$, and $x(YZ)\bar{x}$ spectra. We attribute the stronger electronic Raman intensity for $YY$ polarization to the additional scattering channels involving two CuO$_2$ planar and a CuO chain bands crossing the Fermi surface along the $\Gamma\,(0,0,0)$ - $Y\,(0,\pi,0)$ direction[@bandstr] in the Brillouin zone .
{width="8cm"}
The coupling of the phonons to the electronic excitations contributing to the Raman background is manifested in our spectra by the asymmetric Fano profile of some of the $A_g$ phonon lines, most clearly pronounced for these near 118 cm$^{-1}$(Ba), 168 cm$^{-1}$ (Y) and 486 cm$^{-1}$ (apex oxygen), which interestingly include mainly motions of atoms not belonging to the Cu-O planes. For a phonon coupled to an electronic background the Fano profile $$I(\omega)=I_0\frac{(\epsilon + q)^2}{(1+\epsilon^2)}+B(\omega)$$ is generally used to describe the line shape, where $\epsilon =
(\omega - \omega_p)/\Gamma$, $\omega_p$ is the renormalized phonon frequency that includes all contributions resulting from the interaction of the phonon with elementary excitations, $\Gamma$ is the linewidth, $q$ is the asymmetry parameter, and $B(\omega)$ is the non-interacting with the phonon part of the electronic excitations continuum. In the case of real phonon and electronic scattering amplitudes $t_p$ and $t_e$, and a flat scattering background around the phonon frequency $$\frac{1}{q} = \frac{t_e}{t_p} \pi \rho V,$$ where $\rho(\omega)={\rm const}$ is the density of electronic excitations coupled to the particular phonon, $V$ is the related electron-phonon coupling constant, and the intensity of the interacting with the phonon part of the electronic continuum in Eq.(1) can be expressed as $I_0=\pi \rho t_e^2$. Under the reasonable assumption that $t_e$, $t_p$, and $V$ are only weakly dependent on $T$, the variations with temperature of the absolute value of $1/q$ and $I_0$ will be governed mainly by $\rho(T)$. Fig. 6 shows in more detail the variations between 260 and 80 K of profile of the Raman line near 485 cm$^{-1}$, which corresponds to the apex oxygen $A_g$ mode. In contrast to the known results for twin free Ortho-I crystals, the slopes of Fano shaped profiles in the $XX$ and $YY$ spectra of Ortho-II are of opposite sign, which is not surprising if we take into account that $t_p^{xx} \propto
\alpha_{xx}$ and $t_p^{yy} \propto \alpha_{yy}$ are of opposite sign for this mode. It is also evident from Fig. 6 that the Raman lines become more symmetric at lower temperature. The Fano fit of the experimental profiles showed that the $1/q$ factor for the $XX$ and $YY$ decreases, respectively, from -0.6 and +0.8 at 300 K to -0.2 and +0.1 at 80 K. On the other hand, the electronic background $B(\omega,T)$ remains nearly constant between 300 and 200 K but with further cooling decreases significantly between 200 and 100 K (Fig.5). On the basis of considerations above concerning $1/q(T)$ and $I_0(T)$ being governed by $\rho(T)$, one therefore concludes that with decreasing temperature the density of electronic excitations coupled to this phonon also decreases, concurrently to the decrease of the electronic background $B(\omega,T)$ already noticed. Such a temperature behavior of the electronic Raman scattering is most consistent with an opening of a gap in the electronic excitations as the pseudogap in the underdoped high temperature superconductors.[@timusk1999; @letacon2006; @hinkov2007] It is worth noting here that the characteristic temperature T$^*
\approx 150$ K, below which Opel at al.[@opel2000] have observed weak spectral weight loss in the $z(XY)\bar{z}$ spectra of twinned films of YBa$_2$Cu$_3$O$_{6.5}$ (T$c \approx 60$ K), is in the same temperature range.
{width="6cm"}
Conclusions
===========
The experimental results reported here strongly suggest that some of the previous Raman scattering data[@misochko1994; @palles1996; @opel2000] obtained from aged twinned surfaces of YBCO$_x$ ($x \approx 6.5$) samples may not be representative of the ordered Ortho-II structure. This study provides more reliable data for identification of the Ortho-II Raman modes, as well as information on the variation of electronic scattering and electron-phonon interactions in the temperature range where opening of a pseudogap has been claimed.
This work is supported in part by the State of Texas through the Texas Center for Superconductivity at the University of Houston (TcSUH) and by the U.S. Air Force Office of Scientific Research (SPRING award ID FA9550-06-1-0401). The work at the University of Sherbrooke has been supported by the National Science and Engineering Research Council of Canada, and the Fonds Québécois de la Recherche sur la Nature et les Technologies.
[99]{} R. Liang, D, A. Bonn, W. N. Hardy, Physica C [**336**]{}, 57 (2000).
C. Stock, W. J. L. Buyers, Z Tun, R. Liang, D. Peets, D. Bonn, W. N. Hardy, L. Taillefer, Phys. Rev. B [**66**]{}, 024505 (2002).
C. Stock, W. J. L. Buyers, R. Liang, D. Peets, Z. Tun, D. Bonn, W. N. Hardy, R. J. Birgeneau Phys. Rev. B [**69**]{}, 014502 (2004).
M. v. Zimmermann, J. R. Schneider, T. Frello, N. H. Andersen, J. Madsen, M. Kall, H. F. Poulsen, R. Liang, P. Dosanjh, W. N. Hardy, Phys. Rev. B [**68**]{}, 104515 (2003).
N. Gedik, J. Orenstein, R. X. Liang, D. A. Bonn, W. Hardy, Physica C [**408**]{}, 690 (2004).
N. Gedik, P. Blake, R. C. Spitzer, J. Orenstein, Ruixing Liang, D. A. Bonn, W. N. Hardy, Phys. Rev. B [**70**]{}, 014504 (2004).
J. Hwang, J. Yang, T. Timusk, S. G. Sharapov, J. P. Carbotte, D. A. Bonn, R. Liang, W. N. Hardy. Phys. Rev. B [**73**]{}, 014508 (2006).
Z. Yamani, B. W. Statt, W. A. MacFarlane, R. X. Liang, D. A. Bonn, W. N. Hardy, Phys. Rev. B [**73**]{}, 212506 (2006).
R. Harris, P. J. Turner, S. Kamal, A. R. Hosseini, P. Dosanjh, G. K Mullins, J. S. Bobowski, C. P. Bidinosti, D. M. Broun, R. Liang, W. N. Hardy, D. A. Bonn, Phys. Rev. B [**74**]{} 104508 (2006).
N. Doiron-Leyraud, C. Proust, D. LeBoeuf, J. Levallois, J. B. Bonnemaison, R. X. Liang, D. A. Bonn, W. N. Hardy, L. Taillefer, Nature [**447**]{}, 565 (2007).
M. Iliev, C. Thomsen, V. Hadjiev, M. Cardona, Phys. Rev. B [**47**]{}, 12341 (1993).
O. V. Misochko, S. Tajima, S. Miyamoto, N. Koshizuka, Solid State Commun. [**92**]{}, 877 (1994).
M. N. Iliev, V. G. Hadjiev, V. G. Ivanov, J. Raman Spectroscopy [**27**]{}, 333 (1996).
E. Faulques, V. G. Ivanov, Phys. Rev. B [**55**]{}, 3974 (1997).
M. Iliev, H.-U. Habermeier, M. Cardona, V. G. Hadjiev, R. Gajic, Physica C [**279**]{}, 63 (1997).
M. N. Iliev, P. X. Zhang, H.-U. Habermeier, M. Cardona, Journal of Alloys and Compounds [**251**]{}, 99 (1997).
M. Osada, M. Käll, J. Bäckström, M. Kakihana, N. H. Andersen, L. Börjesson, Phys. Rev. B [**71**]{}, 214503 (2005)
S. I. Bredikhin, G. A. Emelchenko, V. S. Shechtman, A. A. Zhokhov, S. Carter, R. J. Chater, J. A. Kilner, B. C. H. Steele, Physica C [**179**]{}, 286 (1991).
D. Palles, N. Poulakis, E. Liarokapis, K. Conder, E. Kaldis, K. A. Müller, Phys. Rev. B [**54**]{}, 6721 (1996).
S. Hong, K. Kim, H. Cheong, G. Park, Physica C [**454**]{}, 82 (2007).
Y. B. Li, L. F. Cohen, A. D. Caplin, R. A. Stradling, W. Kula, R. Sobolewski, J. L. MacManus-Driscoll, J. Appl. Phys. [**80**]{}, 2929 (1996).
K. Conder, Materials Science & Engineering, R32, 41 (2000).
T. Strohm and M. Cardona, Phys. Rev. B [**55**]{}, 12725 (1997).
A. Bock, Ann. Phys. (Leipzig) [**6**]{}, 441 (1999).
M. Opel, R. Nemetschek, C. Hoffmann, R. Philipp, P. F. Müller, R. Hackl, I. Tüttő, A. Erb, B. Revaz, E. Walker, H. Berger, and L. Forró, Phys. Rev. B [**61**]{}, 9752 (2000).
T. Devereaux and R. Hackl, Rev. Mod. Phys. [**79**]{}, 175 (2007).
M. Krantz and M. Cardona, J. Low Temp. Phys. [**99**]{}, 205 (1995).
E. Bascones,T. M. Rice, A. O. Shorikov, A. V. Lukoyanov, and V. I. Anisimov, Phys. Rev. B [**71**]{}, 012505 (2005); I. S. Elfimov, G. A. Sawatzky, A. Damascelli, arXiv:0706.4276v2 \[cond-mat.str-el\].
T. Timusk and B. W. Statt, Rep. Prog. Phys., 61 (1999).
M. Le Tacon, A. Sacuto, A. Georges, G. Kotliar, Y. Gallais, D. Colson, and A. Forget, Nature Physics [**2**]{}, 537 (2006).
V. Hinkov, P. Bourges, S. Pailhes, Y. Sidis, A. Ivanov, C. D. Frost, T. G. Perring, C. T. Lin, D. P. Chen, and B. Keimer B, Nature Physics [**3**]{}, 780 (2007).
.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We find generators for the full rational loop group of ${{{\operatorname{GL}}}}(n,{\mathbb C})$ as well as for the subgroups consisting of loops that satisfy a reality condition with respect to one of the noncompact real forms ${{{\operatorname{GL}}}}(n,{\mathbb R})$ and ${{{\operatorname{U}}}}(p,q)$. We calculate the dressing action of some of those generators on the positive loop group, and apply this to the ZS–AKNS flows and the $n$–dimensional system associated to ${{{\operatorname{GL}}}}(n,{\mathbb R})/{{{\operatorname{O}}}}(n)$.'
address: 'Mathematisches Institut der Universität zu Köln, Weyertal 86–90, 50931 Köln, Germany'
author:
- Oliver Goertsches
title: Generating Rational Loop Groups With Noncompact Reality Conditions
---
Introduction
============
The interest in finding generators for rational loop groups, i.e. groups of meromorphic maps from ${{{\mathbb C}\mathbb{P}^{1}}}$ into a complex Lie group, originated from dressing actions [@Terng2000] and their various geometric applications; cf. the survey [@Terng2008] and the references therein. Terng and Uhlenbeck introduced the idea of simple elements, i.e. rational loops with as few poles as possible that generate the loop group, in order to obtain explicit formulae for the dressing action.
Uhlenbeck [@Uhlenbeck1989] found simple elements for the group of ${{{\operatorname{GL}}}}(n,{\mathbb C})$-valued rational loops satisfying the ${{{\operatorname{U}}}}(n)$-reality condition, and Terng and Wang [@Terng2006] extended this to the twisted loop group associated to ${{{\operatorname{U}}}}(n)/{{{\operatorname{O}}}}(n)$. Motivated by this work, Donaldson, Fox and the author [@DFG2008] found generators for the rational loop groups of all classical groups and ${{{\operatorname{G}}}}_2$ with reality condition given by the respective compact real form, and most of their twisted loop groups.
Looking at the above results, it suggests itself to ask for generators of rational loop groups, where the reality condition is given by a *noncompact* real form. In this paper, we solve this question for the easiest case, namely the noncompact real forms ${{{\operatorname{GL}}}}(n,{\mathbb R})$ and ${{{\operatorname{U}}}}(p,q)$ of ${{{\operatorname{GL}}}}(n,{\mathbb C})$. It turns out that the task of finding generators is actually easier if we do not impose any reality condition at all: in Section \[sec:GLnC\], we show that any ${{{\operatorname{GL}}}}(n,{\mathbb C})$-valued loop can be written as a product of loops of the form $$p_{{\alpha},\beta,V,W}({\lambda})=\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)\pi_V +\pi_W,$$ where the projections $\pi_V$ and $\pi_W$ are defined via a decomposition ${\mathbb C}^n=V\oplus W$ into complex subspaces, and $$m_{{\alpha},k,N}({\lambda})=\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)^k N,$$ where $k$ is a positive integer and $N$ is a two-step nilpotent map, i.e. $N^2=0$.
Whereas the first type of simple elements is the obvious generalization of those used in [@Uhlenbeck1989], the loops of the form $m_{{\alpha},k,N}$ are of a different nature, mainly because they have only one singularity. This also reflects itself in the proof, which is split into two parts. Using only the first type and with the same arguments as in the proofs of the theorems mentioned above, we first reduce to the case of a loop with only one singularity; afterwards, a different argument shows that this loop is a product of loops of the second type.
In Sections \[sec:GLnR\] and \[sec:Upq\], we give refinements of this proof to generate the subgroup of loops satisfying the reality condition given by ${{{\operatorname{GL}}}}(n,{\mathbb R})$ and ${{{\operatorname{U}}}}(p,q)$, respectively. In each case, we need those of the loops above that satisfy the reality condition, as well as products of two simple elements of the type $p_{{\alpha},\beta,V,W}$ that do not satisfy the reality condition by themselves. For ${{{\operatorname{U}}}}(p,q)$ with $p\neq q$ we also need a generalization of the $m_{{\alpha},k,N}$ in which we allow $N$ to be three-step nilpotent.
We would like to remark that as previously done in the literature, we formulate the theorems for groups of negative loops, i.e. loops that are normalized at $\infty$. All of them are true without this assumption, if we allow more general linear fractional transformations in the definition of the simple factors than those that send $\infty$ to $\Id$.
Sections \[sec:dressing1\], \[sec:dressing2\] and \[sec:systems\] are independent of the generating theorems in Sections \[sec:GLnC\], \[sec:GLnR\], and \[sec:Upq\]. In Section \[sec:dressing1\], we consider the dressing action of simple elements of the form $m_{{\alpha},k,N}$ with $k=1$, and apply this to the ZS-AKNS flows. To apply dressing to the twisted flows in the ${{{\operatorname{SL}}}}(n)/{{{\operatorname{SO}}}}(n)$–hierarchy, we also prove a permutability formula that enables us to find certain products $s_{{\alpha},N}$ of simple elements $m_{{\alpha},1,N}$ that satisfy the twisting condition, see . In Section \[sec:dressing2\] we briefly consider the case $k=2$.
Finally, in Section \[sec:systems\], we make the observation that the $n$–dimensional system associated to a symmetric space $U/K$ is equivalent to the system associated to its dual symmetric space $U^*/K$. The space of solutions of the ${{{\operatorname{U}}}}(n)/{{{\operatorname{O}}}}(n)$–system, which by the work of Terng and Wang [@Terng2006] can be identified with the space of $\partial$–invariant flat Egoroff metrics, is therefore acted on by the group of negative loops in ${{{\operatorname{GL}}}}(n,{\mathbb C})$ satisfying the ${{{\operatorname{GL}}}}(n)$–reality and the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–twisting condition, in particular by the $s_{{\alpha},N}$. We calculate the action of the $s_{{\alpha},N}$ on those Egoroff metrics and their associated families of flat Lagrangian immersions in ${\mathbb C}^n$.
*Acknowledgements.* The work on this paper was started when the author was visiting Chuu-Lian Terng in Irvine, supported by a DAAD postdoctoral scholarship. He wishes to thank UC Irvine, and especially Chuu-Lian Terng, for their hospitality. Furthermore, he thanks Chuu-Lian Terng, Daniel Fox, and Neil Donaldson for valuable comments and useful discussions.
Preliminaries
=============
For any complex reductive Lie group $G$ and representation $\rho:G\to{{{\operatorname{GL}}}}(V)$, the *rational loop group* associated to $\rho$ is given by $${{{\mathcal L}}}(G,V)=\{g:{{{\mathbb C}\mathbb{P}^{1}}}\to G\mid \rho\circ g:{{{\mathbb C}\mathbb{P}^{1}}}\to {{{\operatorname{End}}}}(V) \text{ is meromorphic}\};$$ see [@DFG2008] for some basic examples on how the rational loop group of $G$ depends on the chosen representation. If $\tau$ is an antiholomorphic involution of $G$, we say that a loop $g\in {{{\mathcal L}}}({{{\operatorname{GL}}}}(n,{\mathbb C}))$ satisfies the *reality condition* with respect to $\tau$ if $$\tau(g({\lambda}))=g({\overline{{\lambda}}}).$$ If $\sigma$ additionally is an holomorphic involution on $G$ commuting with $\tau$, then we say that $g$ is [*twisted*]{} with respect to $\sigma$ if $$\sigma(g(-\lambda))=g(\lambda).$$ A loop $g$ is called $\emph{negative}$ if it is normalized at $\infty$, i.e. $g(\infty)=\Id$. We use superscripts to denote the reality and twisting conditions, and subscripts to denote negativity; for example, the group of negative rational loops satisfying the $\tau$-reality and the $\sigma$-twisting condition will be denoted by ${{{\mathcal L}}}_-^{\tau,\sigma}(G,V)$.
If $g\in {{{\mathcal L}}}(G,V)$ is given, we say that $\alpha\in {{{\mathbb C}\mathbb{P}^{1}}}$ is a *pole* of $g$ if ${\alpha}$ is a pole of $\rho\circ g:{{{\mathbb C}\mathbb{P}^{1}}}\to {{{\operatorname{End}}}}(V)$. If ${\alpha}$ is not a pole of $g$, we say that ${\alpha}$ is a *zero* of $g$ if $\rho(g({\alpha}))\in{{{\operatorname{End}}}}(V)$ is singular. Finally, ${\alpha}$ is a *singularity* of $g$ if it is a pole or a zero.
If ${\alpha}\in {{{\mathbb C}\mathbb{P}^{1}}}$ is a pole of $g$, there is a unique number $k\ge 1$ such that the map $({\lambda}-{\alpha})^{k-1}g$ has a pole at ${\alpha}$, but $({\lambda}-{\alpha})^k g$ has none. If we denote the evaluation of this map at ${\alpha}$ by $A\in {{{\operatorname{End}}}}(V)$, we call the pair $(k,\rm{rk }\, A)$ the *pole data* of $g$ at ${\alpha}$. There is a natural ordering on the possible pole data: $(k_1,n_1)<(k_2,n_2)$ if and only if $k_1<k_2$ or ($k_1=k_2$ and $n_1<n_2$). It thus makes sense to compare degrees of poles.
The full rational loop group {#sec:GLnC}
============================
In this section, we prove a Generating Theorem for the full rational loop group of ${{{\operatorname{GL}}}}(n,{\mathbb C})$ associated to the standard representation on ${\mathbb C}^n$. The simple elements needed for that are given in Table \[tab:GLnC\].
Name Definition Conditions
-------------------------- ----------------------------------------------------------------------- ---------------------------------------------
$p_{{\alpha},\beta,V,W}$ $\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)\pi_V +\pi_W$ ${\mathbb C}^n=V\oplus W$
$m_{{\alpha},k,N}$ $\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)^k N$ $N:{\mathbb C}^n\to {\mathbb C}^n,\, N^2=0$
: Simple elements for ${{{\mathcal L}}}_-({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$
\[tab:GLnC\]
Here, $\alpha,\beta$ are distinct complex numbers. and the maps $\pi$ are projections along the decomposition in the column ’Conditions’ onto the subspace in the subscript. Note that the $p_{{\alpha},\beta,V,W}$ have two singularities, whereas the $m_{{\alpha},k,N}$ have only one; furthermore, the determinant of $m_{{\alpha},k,N}$ is $1$ at each value ${\lambda}\neq{\alpha}$.
\[thm:gennoreality\] The rational loop group ${{{\mathcal L}}}_-({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$ is generated by the simple elements given in Table \[tab:GLnC\].
In the case $n=1$, no simple factors of the form $m_{{\alpha},k,N}$ exist. The theorem becomes the well-known statement that any meromorphic map $f:{{{\mathbb C}\mathbb{P}^{1}}}\to {{{\mathbb C}\mathbb{P}^{1}}}$ with $f(\infty)=1$ is of the form $f({\lambda})=\frac{p({\lambda})}{q({\lambda})}$, where $p$ and $q$ are monic polynomials of equal degree.
Let $g\in {{{\mathcal L}}}_-({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$. The first step in the proof is to multiply simple elements $p_{{\alpha},\beta,V,W}$ to the left of $g$ to remove all but at most one singularity. This works similarly to the proofs of existing generating theorems:
Assume first that $g$ has at least two singularities. Let $\alpha\in {\mathbb C}$ be a pole of $g$ – which exists since otherwise $g$ had to be constant – and $\beta\in {\mathbb C}$ another singularity. If we define $\varphi({\lambda})=\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$, the map $g\circ \varphi^{-1}$ has a pole at $0$, so we can write its Laurent expansion around $0$ as $g\circ\varphi^{-1}({\lambda})=\sum_{j=-k}^\infty {\lambda}^j g_j$ with $g_{-k}\neq 0$. Composing with $\varphi$, we obtain the Laurent expansion of $g$ in $\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$ around ${\alpha}$: $$g({\lambda})=\sum_{j=-k}^\infty \left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)^j g_j.$$ Let $V={\operatorname{im}}g_{-k}$, choose an arbitrary complement $W$ of $V$, and regard $$\begin{aligned}
p_{{\alpha},\beta,V,W}({\lambda})g({\lambda})&=\left( \left(\frac{{\lambda}-\alpha}{{\lambda}-\beta}\right) \pi_V +\pi_W\right)\left(\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^kg_{-k}+\ldots\right),\end{aligned}$$ which obviously has a pole at ${\alpha}$ of lower degree. Inductively, we can remove the pole at ${\alpha}$ by multiplying simple elements of the first type and are left with a loop (which we again call $g$) whose Laurent expansion around ${\alpha}$ in $\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$ is of the form $$g({\lambda})=g_0+\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)g_1+\ldots$$ If $g_0$ is invertible, we have completely removed the singularity ${\alpha}$. If $g_0$ is not invertible, we continue as follows: The map ${\lambda}\mapsto \det g({\lambda})$ has a zero at ${\alpha}$ of a certain order, say $k$. If we set $W={\operatorname{im}}g_0$ and let $V$ be an arbitrary complement, the loop $
\tilde{g}=p_{\beta,{\alpha},V,W}g$ has no pole at ${\alpha}$, and the order of the zero of ${\lambda}\mapsto \det \tilde{g}({\lambda})$ is lower than $k$. Using induction, we arrive at a loop whose evaluation at ${\alpha}$ is invertible, i.e. in ${{{\operatorname{GL}}}}(n,{\mathbb C})$. This loop has strictly less singularities than the one we started with.
For this procedure, it was essential to be able to choose two distinct singularities. Therefore, we can only repeat this process until we are left with a loop $g$ that has exactly one pole, say ${\alpha}\in {\mathbb C}$, and no other singularity, i.e. $g({\lambda})\in {{{\operatorname{GL}}}}(n,{\mathbb C})$ for all ${\lambda}\in {{{\mathbb C}\mathbb{P}^{1}}},{\lambda}\neq {\alpha}$. We can therefore write $g$ explicitly as $$g({\lambda})=({\lambda}-{\alpha})^{-r} A_r+\ldots+({\lambda}-{\alpha})^{-1} A_1+A_0$$ with $A_r\neq 0$. The normalization condition says $A_0=\Id$. Since $\det g({\lambda})$ is a polynomial in $({\lambda}-{\alpha})^{-1}$, and complex polynomials always have at least one pole and one zero on ${{{\mathbb C}\mathbb{P}^{1}}}$, it follows that $\det g({\lambda})=1$ for all ${\lambda}\neq {\alpha}$.
For the second part of the proof, we need some notation. For any $i\ge 0$, we define $K_i=\bigcap_{j\ge i} \ker A_j$ and $$V_i:=\sum_{j\ge i} A_j(K_{j+1}).$$ We have filtrations $$\label{eqn:filtration}
0=K_0\subset K_1\subset \ldots \subset K_r\subset K_{r+1}={\mathbb C}^n$$ and $$0=V_{r+1}\subset V_r\subset \ldots \subset V_1\subset V_0\subset {\mathbb C}^n.$$ Let ${\mathcal K}$ be the set of tuples of nonnegative integers $(a_i)_{i\ge 0}$ satisfying $\sum_i a_i=n$. We introduce a total ordering on ${\mathcal K}$ by setting $$(a_i)_i< (b_i)_i\Longleftrightarrow \text{There exists } j\ge 0 \text{ such that } a_i= b_i \text{ for } i>j \text{ and } a_j < b_j.$$ Note that the unique minimum with respect to this ordering of $\mathcal K$ is the tuple $(n,0,0,\ldots)$. For a loop $g$ as above, we define an associated tuple ${\epsilon}(g)=(a_i)_i\in {\mathcal K}$ by $a_i:=\dim K_{i+1}-\dim K_i=\dim A_i(K_{i+1})$; the tuple ${\epsilon}(g)$ really is an element of $\mathcal K$ since $\sum_{i\ge 0} (\dim K_{i+1}-\dim K_i)=\dim K_{r+1}-\dim K_0=n$. The only loop $g$ whose associated tuple ${\epsilon}(g)$ is the minimum $(n,0,0,\ldots)$, is the constant loop $g({\lambda})=\Id$. We show by induction on ${\epsilon}(g)$ that $g$ can be written as a product of simple elements of the form $m_{{\alpha},k,N}$, the induction basis being trivial.
Let $s\ge 0$ be the smallest number such that ${\operatorname{im}}A_i\subset V_i$ for all $i\ge s$. Since ${\operatorname{im}}A_r=V_r$ by definition, we have $s\le r$.
Let us first regard the case that $s>0$; the case $s=0$ will be treated later. By definition of $V_{s-1}$, the space $A_{s-1}(K_s)$ is a subset of $V_{s-1}$, but by definition of $s$, the space $A_{s-1}(K_{r+1})={\operatorname{im}}A_{s-1}$ is not, so the smallest number $l$ such that ${\operatorname{im}}A_{s-1}(K_l)\not\subset V_{s-1}$ satisfies $s<l\le r+1$. Let $v\in K_l$ be such that $A_{s-1}(v)\notin V_{s-1}$, and note that $A_{l-1}(v)\neq 0$.
Let $N$ be a two-step nilpotent map satisfying $N(V_{s-1})=0$ and $N(A_{s-1}(v))=-A_{l-1}(v)\in V_{l-1}\subset V_{s-1}$. It follows that $NA_i=0$ for all $i\ge s$ since ${\operatorname{im}}A_i\subset V_i\subset V_{s-1}$ for such $i$. Therefore, the product $$\begin{aligned}
\tilde{g}({\lambda})&=m_{{\alpha},l-s,N}({\lambda})g({\lambda})= (\Id+({\lambda}-{\alpha})^{s-l}N)\sum_{i=0}^r ({\lambda}-{\alpha})^{-i}A_i\\
&=\Id+\ldots + ({\lambda}-{\alpha})^{-l+1} (NA_{s-1}+A_{l-1}) +\sum_{i\ge l} ({\lambda}-{\alpha})^{-i} A_i\end{aligned}$$ coincides with $g$ starting with the $({\lambda}-{\alpha})^{-l}$-coefficient. The $({\lambda}-{\alpha})^{-l+1}$-coefficient satisfies $$(NA_{s-1}+A_{l-1})(K_{l-1})=NA_{s-1}(K_{l-1})\subset N(V_{s-1})=0$$ and $$(NA_{s-1}+A_{l-1})(v)=-A_{l-1}(v)+A_{l-1}(v)=0,$$ so ${\epsilon}(\tilde{g})<{\epsilon}(g)$, and induction may be applied.
It remains to regard the case $s=0$, i.e. ${\operatorname{im}}A_i\subset V_i$ for all $i\ge 0$. In particular, $V_0={\mathbb C}^n$. For dimensional reasons, we have a direct decomposition $$\label{eqn:basicproof_s=0}
{\mathbb C}^n=\bigoplus_{i\ge 0} A_i(K_{i+1}).$$ Let $\mathcal B$ be a basis of ${\mathbb C}^n$ compatible with the filtration . More precisely, let $W_i$ be a complement of $K_{i}$ in $K_{i+1}$, i.e. $${\mathbb C}^n=K_r\oplus W_r=K_{r-1}\oplus W_{r-1}\oplus W_r=\ldots=\bigoplus_{i\ge 0} W_i,$$ choose bases ${\mathcal B}_i$ of $W_i$, and let ${\mathcal B}=\bigcup_i {\mathcal B}_i$. Note that $A_i(K_{i+1})=A_i(W_i)$, so by , we get a second basis $\mathcal B'$ of ${\mathbb C}^n$ by defining ${\mathcal B}'=\bigcup_i A_i({\mathcal B}_i)$. Since ${\operatorname{im}}A_i \subset V_i$, the matrix representation of $A_i$ with respect to these bases (${\mathcal B}$ as basis of the domain of definition, and ${\mathcal B'}$ as basis of the target) is of the form $$[A_i]_{\mathcal B\mathcal B'}=\left(\begin{array}{c|c|c} * & {\bf{0}} & {\bf{0}} \\ \hline
* & {\bf{1}} & {\bf{0}}\\ \hline
\smash{\underbrace{\bf{0}}_{n-\dim K_{i+1}}} & \smash{\underbrace{\bf{0}}_{\dim K_{i+1}-\dim K_i}} & \smash{\underbrace{\bf{0}}_{\dim K_i}} \end{array} \right),$$ where $*$ signifies unknown entries and $\bf 1$ represents a diagonal matrix of the appropriate dimension. From this, we can calculate the leading term of $\det g({\lambda})$ as a polynomial in $({\lambda}-{\alpha})^{-1}$: $$\det g({\lambda})=({\lambda}-{\alpha})^{-\sum_{i\ge 0} i \dim W_i} +\ldots$$ On the other hand, we know that $\det g({\lambda})=1$ for all ${\lambda}\neq {\alpha}$, which is therefore only possible if $\dim W_i=0$ for all $i\ge 1$, i.e. $g({\lambda})=\Id$ for all ${\lambda}$.
The ${{{\operatorname{GL}}}}(n,{\mathbb R})$–reality condition {#sec:GLnR}
==============================================================
In this section, we prove a generating theorem for the group of ${{{\operatorname{GL}}}}(n,{\mathbb C})$-valued loops satisfying the reality condition with respect to the noncompact real form ${{{\operatorname{GL}}}}(n,{\mathbb R})$. Denote by $\tau:{{{\operatorname{GL}}}}(n,{\mathbb C})\to {{{\operatorname{GL}}}}(n,{\mathbb C})$ the antiholomorphic involution $\tau(A)={\overline{A}}$; we are interested in the loop group ${{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$, i.e. the group of rational loops $g:{{{\mathbb C}\mathbb{P}^{1}}}\to {{{\operatorname{GL}}}}(n,{\mathbb C})$ satisfying ${\overline{g({\overline{{\lambda}}})}}=g({\lambda})$ and the normalization condition $g(\infty)=\Id$.
To generate this group, we need several types of simple elements, see Table \[tab:glnr\].
[|c|c|c|]{} Name & Definition & Conditions\
$p_{{\alpha},\beta,V,W}$ & $\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)\pi_V +\pi_W$ &
----------------------------------------------------------
$\alpha,\beta\in {\mathbb R},\, {\mathbb C}^n=V\oplus W$
${\overline{V}}=V,\, {\overline{W}}=W$
----------------------------------------------------------
: Simple elements for the ${{{\operatorname{GL}}}}(n,{\mathbb R})$-reality condition
\
$q_{\alpha,\beta,V,W}$ & $\frac{({\lambda}-\alpha)({\lambda}-{\overline{{\alpha}}})}{({\lambda}-\beta)({\lambda}-{\overline{\beta}})}\pi_V+\pi_W$ &
--------------------------------------------------------------------
${\alpha}$ or $\beta\notin {\mathbb R},\, {\mathbb C}^n=V\oplus W$
${\overline{V}}=V,\, {\overline{W}}=W$
--------------------------------------------------------------------
: Simple elements for the ${{{\operatorname{GL}}}}(n,{\mathbb R})$-reality condition
\
$r_{\alpha,\beta,V,W}$ & $\left(\frac{{\lambda}-\alpha}{{\lambda}-\beta}\right)\pi_V + \pi_W + \left(\frac{{\lambda}-{\overline{\alpha}}}{{\lambda}-{\overline{\beta}}}\right) \pi_{{\overline{V}}}$ &
------------------------------------------------
${\mathbb C}^n=V\oplus W\oplus {\overline{V}}$
$V\cap {\overline{V}}=0,\, {\overline{W}}=W$
------------------------------------------------
: Simple elements for the ${{{\operatorname{GL}}}}(n,{\mathbb R})$-reality condition
\
$m_{{\alpha},k,N}$ & $\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)^k N$ & ${\alpha}\in{\mathbb R},\, N^2=0,\, {\overline{N}}=N$\
\[tab:glnr\]
Note that all of these simple elements are either ${{{\operatorname{GL}}}}(n,{\mathbb C})$-simple factors or products of two ${{{\operatorname{GL}}}}(n,{\mathbb C})$-simple factors that do not satisfy the reality condition by themselves: $q_{\alpha,\beta,V,W}=p_{{\alpha},\beta,V,W}p_{{\overline{{\alpha}}},{\overline{\beta}},V,W}$ and $r_{\alpha,\beta,V,W}=p_{{\alpha},\beta,V,W}p_{{\overline{{\alpha}}},{\overline{\beta}},{\overline{V}},W}$.
\[thm:glnr\] The rational loop group ${{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$ is generated by the simple elements given in Table \[tab:glnr\].
Let $g\in {{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$. Observe that if $\alpha\in {\mathbb C}$ is a singularity of $g$, then so is ${\overline{{\alpha}}}$. We first regard the case that $g$ has at least two singularities, not all of which are real. Let ${\alpha}\in {\mathbb C}\setminus {\mathbb R}$ be a singularity of $g$. If $\alpha$ and ${\overline{\alpha}}$ are the only singularities of $g$, let $\beta$ be a random real number; otherwise let $\beta$ be a (real or complex) singularity of $g$ different from $\alpha$ and ${\overline{{\alpha}}}$. We will remove the singularity at $\alpha$ (and simultaneously at ${\overline{{\alpha}}}$) by multiplying with simple elements of the type $q$ and $r$, so although in the first case we might introduce a new singularity at the real value $\beta$, we will have reduced the total number of singularities in any case.
If $g$ has a pole at ${\alpha}$, write the Laurent expansion of $g$ in $\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$ around ${\alpha}$ as $$g({\lambda})=\sum_{j=-k}^\infty \left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)^jg_j$$ with $g_{-k}\neq 0$; otherwise, continue with below. If there exists a nonzero space $V\subset {\operatorname{im}}g_{-k}$ with $V={\overline{V}}$, let $W$ be an arbitrary complement of $V$ with ${\overline{W}}=W$, and regard $$\begin{aligned}
q_{{\alpha},\beta,V,W}({\lambda})g({\lambda})&= \left(\frac{({\lambda}-\alpha)({\lambda}-{\overline{{\alpha}}})}{({\lambda}-\beta)({\lambda}-{\overline{\beta}})}\pi_V + \pi_W \right)\left(\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^kg_{-k}+\ldots\right)\\
&=\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^k \pi_W\circ g_{-k}+\ldots.\end{aligned}$$ This loop has a pole of lower degree at ${\alpha}$, since the kernel of $\pi_W\circ g_{-k}$ contains not only the kernel of $g_{-k}$, but also the preimage of $V$ under $g_{-k}$.
If such a space does not exist, let $V={\operatorname{im}}g_{-k}$ be the full image of $g_{-k}$. We have $V\cap {\overline{V}}=0$ and can therefore choose an arbitrary complement $W$ of $V\oplus {\overline{V}}$ with ${\overline{W}}=W$. Regard $$r_{{\alpha},\beta,V,W}({\lambda})g({\lambda})=\left(\left(\frac{{\lambda}-\alpha}{{\lambda}-\beta}\right)\pi_V + \pi_W + \left(\frac{{\lambda}-{\overline{\alpha}}}{{\lambda}-{\overline{\beta}}}\right) \pi_{{\overline{V}}}\right)\left(\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^kg_{-k}+\ldots\right),$$ which has a pole of lower degree; in fact, its $\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^k$-coefficient vanishes completely.
Continuing this, we obtain a loop (again denoted by $g$) without pole at ${\alpha}$, whose Laurent expansion in $\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$ around ${\alpha}$ we write as $$\label{eq:gonlyzero}
g({\lambda})=g_0+\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)g_1+\ldots$$ If $g_0$ is invertible, ${\alpha}$ is no singularity, so assume that $g_0$ is singular. Denote by $k$ the order of the zero ${\alpha}$ of the map ${\lambda}\mapsto \det g({\lambda})$. Let $W_0\subset {\operatorname{im}}g_{0}$ be a maximal subspace with $W_0={\overline{W_0}}$, and write ${\operatorname{im}}g_{-k}=W_0\oplus W_1$, where $W_1$ is an arbitrary complement of $W_0$ in ${\operatorname{im}}g_{-k}$. We have necessarily $W_1\cap {\overline{W_1}}=0$.
If $W_1$ is not empty, let $V={\overline{W_1}}$ and $W=W_0\oplus W_2$, where $W_2$ is an arbitrary complement of $W_0\oplus W_1\oplus {\overline{W_1}}$ in ${\mathbb C}^n$ with $W_2={\overline{W_2}}$. We have constructed a decomposition $${\mathbb C}^n=V\oplus W\oplus {\overline{V}}$$ with ${\operatorname{im}}g_{-k}\subset W\oplus {\overline{V}}$. Then, the loop $\tilde{g}=r_{\beta,{\alpha},V,W}g$ has no pole at ${\alpha}$ since ${\alpha}\neq {\overline{{\alpha}}}$; furthermore, the map ${\lambda}\mapsto \det \tilde{g}({\lambda})$ has a zero at ${\alpha}$ of lower order than $k$.
If $W_1$ is empty, we have ${\operatorname{im}}g_{-k}=W_0$, i.e. ${\operatorname{im}}g_{-k}={\overline{{\operatorname{im}}g_{-k}}}$. In this case, let $W={\operatorname{im}}g_{-k}$ and $V$ be an arbitrary complement with $V={\overline{V}}$. Then we reduce the order of the zero by regarding $\tilde{g}=q_{\beta,{\alpha},V,W}g$.
By induction, we have removed the singularity $\alpha$ (and simultaneously ${\overline{\alpha}}$). Repeating this step removes all nonreal singularities.
After having removed all nonreal singularities, we have to deal with the case of several real singularities. If ${\alpha}\neq \beta$ are two real singularities of $g$, we can continue as in the first step, the difference being that the reality condition implies that the image of $g_{-k}$ (and the image of $g_0$, after having removed the pole) is invariant under conjugation. This simplifies matters insofar as we only need to make use of the simple factors $p$; in the notation of the previous step, there always exists a nonzero $V\subset {\operatorname{im}}g_{-k}$ with ${\overline{V}}=V$ (in fact, we may choose $V={\operatorname{im}}g_{-k}$), and the space $W_1$ is always empty.
Finally, we are left with a loop $g\in {{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$ with exactly one singularity $\alpha\in {\mathbb R}$. We can therefore write $g$ explicitly as $$g({\lambda})=({\lambda}-{\alpha})^{-r} A_r+\ldots+({\lambda}-{\alpha})^{-1} A_1+A_0$$ with $A_r\neq 0$. The reality condition implies immediately that ${\overline{A_i}}=A_i$ for all $i$. We may continue the proof exactly as in Theorem \[thm:gennoreality\], because due to the reality of the $A_i$, the nilpotent endomorphisms $N$ constructed there may all be chosen to be real.
The ${{{\operatorname{U}}}}(p,q)$–reality condition {#sec:Upq}
===================================================
In this section, we generate the group of ${{{\operatorname{GL}}}}(n,{\mathbb C})$-valued loops satisfying the ${{{\operatorname{U}}}}(p,q)$-reality condition. Let us assume that $p\le q$. The case $p=0$ was proved by Uhlenbeck in [@Uhlenbeck1989]. The group ${{{\operatorname{U}}}}(p,q)$ is the fixed point set of the antiholomorphic involution $\tau(A)=(A^*)^{-1}$, where $A^*$ is the adjoint of $A$ with respect to the inner product $\langle v,w\rangle=-\sum_{i=1}^p {\overline}{v_i}{w_i}+\sum_{i=p+1}^{p+q} {\overline}{v_i}{w_i}$. Also denote $$s=\left(\begin{matrix}-I_p & 0 \\ 0 & I_q \end{matrix}\right).$$ The simple elements needed for the proof are given in Table \[tab:upq\]; the loops $n_{{\alpha},k,N}$ are needed only for the case $p\neq q$.
[|c|c|c|]{} Name & Definition & Conditions\
$p_{{\alpha},V}$ & $\left(\frac{{\lambda}-{\alpha}}{{\lambda}-{\overline{{\alpha}}}}\right)\pi_V+\pi_{V^\perp}$ & ${\mathbb C}^n=V\oplus V^\perp,\, V\cap V^\perp=0$\
$q_{\alpha,\beta,V}$ & $\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)\pi_V+\pi_{(V\oplus sV)^\perp} +\left(\frac{{\lambda}-{\overline}{\beta}}{{\lambda}-{\overline}{{\alpha}}}\right) \pi_{sV}
$ &
------------------------------------------------------
${\mathbb C}^n=V\oplus (V\oplus sV)^\perp \oplus sV$
$V$ isotropic
------------------------------------------------------
: Simple elements for the ${{{\operatorname{U}}}}(p,q)$-reality condition
\
$m_{{\alpha},k,N}$ & $\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)^k N$ & ${\alpha}\in{\mathbb R},\, N^2=0,\, N^*=-N$\
$n_{{\alpha},k,N}$ & $\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)^k N+\frac{1}{2}\left(\frac{1}{{\lambda}-{\alpha}}\right)^{2k} N^2$ &
---------------------------------------------
${\alpha}\in {\mathbb R}$, $N=M-M^*$, where
$M:(V\oplus sV)^\perp\to V$,
$V$ max. isotropic
---------------------------------------------
: Simple elements for the ${{{\operatorname{U}}}}(p,q)$-reality condition
\
\[tab:upq\]
One easily sees that the first two types of simple elements satisfy the reality condition. Also, the loops $p_{{\alpha},V}$ already appear in [@Terng2000], Section 11. Note that the $q_{{\alpha},\beta,V}$ are products of two ${{{\operatorname{GL}}}}(n,{\mathbb C})$-simple elements: $q_{{\alpha},\beta,V}:=p_{{\alpha},\beta,V,(sV)^\perp}p_{{\overline{\beta}},{\overline{{\alpha}}},sV,V^\perp}$. Furthermore, there is an overlap between the first two types: $ q_{{\alpha},{\overline}{{\alpha}},V}=p_{{\alpha},V\oplus sV}$. To show that the last two types of simple elements satisfy the reality condition observe that they fit into the following framework:
Let $N^*=-N$ and $N^r=0$ for some $r\ge 1$. Then, for any ${\alpha}\in {\mathbb R}$ and $k\in {\mathbb N}$, $$g({\lambda})=\exp(({\lambda}-{\alpha})^{-k} N)=\sum_{j=0}^{r-1} \frac{1}{j!} ({\lambda}-{\alpha})^{-jk}N^{j}$$ is a rational loop satisfying the ${{{\operatorname{U}}}}(p,q)$-reality condition.
Rationality is clear because of the nilpotency of $N$. For the reality condition we calculate $$\tau(g({\overline{{\lambda}}}))=\exp(d\tau(({\overline{{\lambda}}}-{\alpha})^{-k}N))=\exp(-({\lambda}-{\alpha})^{-k}N^*)=g({\lambda}),$$ where we used $d\tau(X)=-X^*$ for all $X\in {{{\mathfrak {gl}}}}(n,{\mathbb C})$.
The following lemmas about the existence of simple factors of type $m_{{\alpha},k,N}$ and $n_{{\alpha},k,N}$ will be crucial for the second part of the proof.
\[lem:Nexist\] Let $N^*=-N$ and $N^2=0$. Then, ${\operatorname{im}}N$ is isotropic, and $\ker N=({\operatorname{im}}N)^\perp$. Conversely, for any $V\subset {\mathbb C}^n$ isotropic, and vectors $v\in V$ and $w\notin V^\perp$ such that $\langle v,w\rangle\in i\cdot {\mathbb R}$, there exists a skewsymmetric endomorphism $N$ of ${\mathbb C}^n$ with $N^2=0$ satisfying $N(V^\perp)=0$, ${\operatorname{im}}N\subset V$ and $N(w)=v$.
Let $N:{\mathbb C}^n\to {\mathbb C}^n$ be a two-step nilpotent skewsymmetric endomorphism. Since $$0=\langle N^2v,w\rangle =-\langle Nv,Nw\rangle,$$ for all $v$ and $w$, the image of $N$ is isotropic. It follows $\ker N=({\operatorname{im}}N^*)^\perp=({\operatorname{im}}N)^\perp$.
The map $N$ is defined by its restriction $N:sV\to V$, and the map $N\circ s:V\to V$ is skewsymmetric with respect to the standard inner product $(v,w)=\langle sv,w\rangle$: $$(Ns(v),w)=\langle sNsv,w\rangle=-\langle sv,Ns(w)\rangle =-(v,Ns(w)).$$
So for the converse direction, if $v\in V$ and $w\notin V^\perp$ with $\langle v,w\rangle\in i\cdot {\mathbb R}$ are given, write $w=s\tilde{w}_0 + \tilde{w}$ with $w\in V$ and $\tilde{w}\in V^\perp$ and note that $(\tilde{w}_0,v)=\langle w,v\rangle \in i\cdot {\mathbb R}$. We can therefore choose a map $\tilde{N}:V\to V$, skewsymmetric with respect to $(\cdot,\cdot)$, that satisfies $\tilde{N}(\tilde{w}_0)=v$. Then $N:{\mathbb C}^n\to {\mathbb C}^n$, defined by $N(V^\perp)=0$ and $\left.N\right|_{sV}=\tilde{N}s$, is skewsymmetric with respect to $\langle \cdot,\cdot\rangle$ and $N(w)=N(s\tilde{w}_0)=\tilde{N}(\tilde{w}_0)=v$.
Let $V$ be maximal isotropic, and $M:(V\oplus sV)^\perp\to V$ be an arbitrary linear map. We denote also by $M$ the extension of $M$ by zero on $V\oplus sV$. Then $M^*$ satisfies $V^\perp\subset \ker M^*$ and $M^*(sV)\subset (V\oplus sV)^\perp$. It follows that $N=M-M^*$ sends $sV$ to $(V\oplus sV)^\perp$, which in turn is sent to $V$. In particular, $N^2$ sends $sV$ to $V$ and $N^3=0$.
\[lem:N2exist\] Assume that $p\neq q$ and let $V$ be maximal isotropic. If $u,v,w$ are vectors such that $w\in V$, $v\in V^\perp\setminus V$ and $u\notin V^\perp$, satisfying $\langle u,w\rangle\in {\mathbb R}$ and $$2\cdot \langle u,w\rangle + \langle v,v\rangle=0,$$ then there exists a linear map $M:(V\oplus sV)^\perp \to V$ such that the skewsymmetric endomorphism $N=M-M^*$ satisfies $$\frac{1}{2}N^2(u)+N(v)+w=0.$$
Let $M:(V\oplus sV)^\perp\to V$ be defined by $M(v)=-2w$ and $M(v^\perp\cap (V\oplus sV)^\perp)=0$, and define $N:=M-M^*$. We have $$\langle N(u),v\rangle =-\langle u,N(v)\rangle=2\cdot \langle u,w\rangle$$ and consequently, $$N(u)\in 2 \frac{\langle u,w\rangle}{\langle v,v\rangle}\cdot v + v^\perp\cap (V\oplus sV)^\perp+V=-v+v^\perp\cap (V\oplus sV)^\perp+V.$$ It follows $$\frac{1}{2}N^2(u)+N(v)+w=w-2w+w=0$$ as desired.
We are now in position to prove the Generating Theorem.
The rational loop group ${{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$ is generated by the simple elements given in Table \[tab:upq\].
Let $g\in {{{\mathcal L}}}_-^\tau({{{\operatorname{GL}}}}(n,{\mathbb C}),{\mathbb C}^n)$. Observe that $g$ is holomorphic (in particular ${{{\operatorname{GL}}}}(n,{\mathbb C})$-valued) at a point ${\alpha}$ if and only if it is holomorphic at ${\overline}{{\alpha}}$. We proceed in three steps as in the proof of Theorem \[thm:glnr\].
Let us first consider the case of $g$ having more than one singularity, and not all of them real. Let ${\alpha}\in {\mathbb C}\setminus {\mathbb R}$ be one of them, and choose $\beta\neq {\alpha},{\overline{{\alpha}}}$ to be another (real or complex) singularity; if there is none, let $\beta$ be an arbitrary real number. If ${\alpha}$ is a pole, write down the Laurent expansion of $g$ in $\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}$ around ${\alpha}$ explicitly as $$g({\lambda})=\sum_{j=-k}^\infty \left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)^j g_j$$ with $g_{-k}\neq 0$; otherwise continue with . If there exists an isotropic subspace $V\subset {\operatorname{im}}g_{-k}$, the loop $$\begin{aligned}
q_{{\alpha},\beta,V}({\lambda}) g({\lambda})&=\left(\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)\pi_V+\pi_{(V\oplus sV)^\perp} + \left(\frac{{\lambda}-{\overline}{\beta}}{{\lambda}-{\overline}{{\alpha}}} \right)\pi_{sV}\right)g({\lambda})\\
&=\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^k \left(\pi_{(V\oplus sV)^\perp}+\left(\frac{{\alpha}-{\overline{\beta}}}{{\alpha}-{\overline{{\alpha}}}}\right)\pi_{sV}\right)g_{-k}+\ldots\end{aligned}$$ has a pole of lower degree at ${\alpha}$; note that we used here that ${\alpha}$ is nonreal.
If there does not exist an isotropic subspace of ${\operatorname{im}}g_{-k}$, let $V={\operatorname{im}}g_{-k}$ and note that $V\cap V^\perp=0$. Then, $$\begin{aligned}
p_{{\alpha},V}({\lambda})g({\lambda})&=\left(\left(\frac{{\lambda}-{\alpha}}{{\lambda}-{\overline{{\alpha}}}}\right)\pi_V+\pi_{V^\perp}\right)\left(\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^kg_{-k}+\ldots\right)\\
&=\left(\frac{{\lambda}-\beta}{{\lambda}-{\alpha}}\right)^k\pi_{V^\perp}\circ g_{-k}+\ldots,\end{aligned}$$ has a pole of lower degree at ${\alpha}$.
Repeating this, we obtain a loop $g$ without pole at ${\alpha}$, whose Laurent expansion we write as $$\label{eq:nopoleupq}
g({\lambda})=g_0+\left(\frac{{\lambda}-{\alpha}}{{\lambda}-\beta}\right)g_1+\ldots.$$ If $g_0$ is invertible, we have removed the singularity at $\alpha$, so assume that $g_0$ is singular. As in the previous proofs, we try to reduce the order of the zero of the map ${\lambda}\mapsto \det g({\lambda})$ by multiplying simple factors.
If ${\operatorname{im}}g_0\cap ({\operatorname{im}}g_0)^\perp=0$, we can reduce the order of the zero by regarding $p_{{\overline{{\alpha}}},V}g$ for $V=({\operatorname{im}}g_0)^\perp$.
If ${\operatorname{im}}g_0\cap ({\operatorname{im}}g_0)^\perp\neq 0$, let $V={\operatorname{im}}g_0\cap ({\operatorname{im}}g_0)^\perp\neq 0$ equal this intersection, and note that $V$ is isotropic and ${\operatorname{im}}g_0\subset V^\perp$. The order of the zero is now reduced by regarding $q_{\beta,{\alpha},sV}g$.
Repeating this, we obtain a loop that is holomorphic at ${\alpha}$.
If $g$ has several singularities, but all of them are real, let ${\alpha}$ and $\beta$ be two of those. If $\alpha$ is a pole and we expand $g$ in a Laurent series as before, it follows from the reality condition that $V:={\operatorname{im}}g_{-k}$ is isotropic. Then, the product $q_{\beta,{\alpha},sV}g$ has a pole of lower degree than $g$. Repeating this, we obtain a loop $g$ that can be evaluated at ${\alpha}$. Because ${\alpha}\in {\mathbb R}$, the reality condition says $g({\alpha})^*g({\alpha})=\Id$; in particular, $g({\alpha})$ is invertible and we have removed the singularity ${\alpha}$.
We are left with the case of $g$ having a single singularity ${\alpha}\in {\mathbb R}$; write $$g({\lambda})=({\lambda}-{\alpha})^{-r}A_r+\ldots+({\lambda}-{\alpha})^{-1}A_1+\Id$$ with $r\ge 1$ and $A_r\neq 0$. The reality condition written out explicitly is $$\label{eq:orthcond}
\sum_{i+j=k} \langle A_iv,A_jw\rangle=0$$ for all $k\ge 1$.
The type of induction we will use is the same as in the second part of the proof of Theorem \[thm:gennoreality\]; recall the following notation: $K_i=\bigcap_{j\ge i} \ker A_j$ for $i\ge 0$, so that $${\mathbb C}^n=K_{r+1}\supset K_r\supset\ldots \supset K_1\supset K_0=0,$$ and $V_i:=\sum_{j\ge i} A_j(K_{j+1})$, so that $$0=V_{r+1}\subset V_r\subset\ldots\subset V_1\subset V_0.$$ For $i\ge 1$, the spaces $V_i$ are isotropic, and $V_0$ is perpendicular to $V_1$, so since $g$ is supposed to be nonconstant, $V_0\neq {\mathbb C}^n$. Thus, no analogue of the last part of the proof of Theorem \[thm:gennoreality\] is needed here.
Let ${\mathcal K}=\{ (a_i)_{i\ge 0}\mid a_i\in {\mathbb N},\, \sum_i a_i=n\}$, equipped with the total ordering $$(a_i)_i< (b_i)_i\Longleftrightarrow \text{There exists } j\ge 0 \text{ such that } a_i=b_i \text{ for } i>j \text{ and } a_j < b_j.$$ For a loop $g$ as above, we define an associated tuple ${\epsilon}(g)=(a_i)_i\in {\mathcal K}$ by $a_i:=\dim K_{i+1}-\dim K_i=\dim A_i(K_{i+1})$.
Since the remainder of the proof is significantly different in the case $p=q$, we treat it in two separate propositions.
With the notation as above, in the case $p=q$ the loop $g$ is a product of simple factors of the form $m_{{\alpha},k,N}$.
We will prove the claim by induction on ${\epsilon}(g)$, the induction basis being trivial since the unique minimum is attained only for $g({\lambda})=\Id$.
Let $k\ge 0$ be the smallest integer such that ${\operatorname{im}}A_k\subset V_k$. Since $A_0=\Id$ and as noted above, $V_0\neq {\mathbb C}^n$ unless $g$ is the constant identity loop, we can assume $k\ge 1$. By definition of $V_{k-1}$, we have $A_{k-1}(K_k)\subset V_{k-1}$, and by definition of $k$, we have $A_{k-1}(K_{r+1})={\operatorname{im}}A_{k-1}\not\subset V_{k-1}$. Let $l$ be the smallest integer satisfying $r+1\ge l >k$ such that $A_{k-1}(K_l)\not\subset V_{k-1}$.
Choose a vector $v\in K_l$ such that $A_{k-1}(v)\notin V_{k-1}$; then automatically $A_{l-1}(v)\neq 0$, since otherwise $v\in K_{l-1}$. Since $l>k$, we therefore have $0\neq A_{l-1}(v)\in V_{l-1}\subset V_{k-1}$. Equation gives $$\begin{aligned}
0&=\sum_{i+j=k+l-2} \langle A_i v,A_j v\rangle\\
&= \langle A_{k-1}v,A_{l-1} v\rangle + \langle A_{l-1} v,A_{k-1}v\rangle=2\cdot {\operatorname{Re}}\,\langle A_{k-1}v,A_{l-1} v\rangle,\end{aligned}$$ since all the other summands vanish either because $v\in K_{l}$, or because $V_k$ is isotropic and for all indices $i,j\ge k$, the images ${\operatorname{im}}A_i$ and ${\operatorname{im}}A_j$ are in $V_k$. Therefore with the help of Lemma \[lem:Nexist\] we may choose a two-step nilpotent, skewsymmetric map $N$ with $N(W)=0$ and $N(A_{k-1}v)=-A_{l-1} v$, where $W\supset V_{k-1}$ is a maximal isotropic subspace with $A_{k-1}v\not\perp W$. Note that we use here that maximal isotropic subspaces are their own orthogonal complement, which is true since we assumed $p=q$. We define $$\begin{aligned}
\tilde{g}({\lambda})&=m_{{\alpha},l-k,N}({\lambda})g({\lambda})\\
&=\Id+\ldots + ({\lambda}-{\alpha})^{-l+1} (NA_{k-1}+A_{l-1}) + ({\lambda}-{\alpha})^{-l} A_{l} +\ldots + ({\lambda}-{\alpha})^{-r} A_r\\
&=:\sum_i ({\lambda}-{\alpha})^{-i}\tilde{A}_i,\end{aligned}$$ and wish to show that ${\epsilon}(\tilde{g})<{\epsilon}(g)$ in order to use induction. To show this we have to investigate how $\tilde{K}_i=\bigcap_{j\ge i} \ker \tilde{A}_j$ has changed compared to $K_i$. Obviously, $K_i=\tilde{K}_i$ for $i\ge l$. If $w\in K_{l-1}$, we have $$(NA_{k-1}+A_{l-1})(w)=NA_{k-1}(w)\in NA_{k-1}(K_{l-1})\subset N(V_{r-1})=0,$$ so $K_{l-1}\subset \tilde{K}_{l-1}$, and furthermore, $(NA_{k-1}+A_{l-1})(v)=0$, so ${{{\operatorname{K}}}}_{l-1}\subsetneq\tilde{K}_{l-1}$, which means ${\epsilon}(\tilde{g})<{\epsilon}(g)$; by induction, the statement follows.
With the same notation as above, in the case $p\neq q$ the loop $g$ is a product of simple factors of the form $m_{{\alpha},k,N}$ and $n_{{\alpha},k,N}$.
We will prove the claim by induction on ${\epsilon}(g)$.
The same argument as in the previous proposition shows that we can reduce to the following situation: $V:=V_{k}$ is maximal isotropic, and ${\operatorname{im}}A_{k}\subset V^\perp$ but ${\operatorname{im}}A_k\not\subset V$. Since $p\neq q$, we have a decomposition $${\mathbb C}^n=V\oplus (V\oplus sV)^\perp \oplus sV,$$ and the inner product is definite on $W:=(V\oplus sV)^\perp$.
Since $A_0=\Id$, we may define $s\ge 1$ to be the integer such that $$\label{eq:imagesperp}
{\operatorname{im}}A_{k-1},\ldots,{\operatorname{im}}A_{k-s+1}\perp V, \text{ but }{\operatorname{im}}A_{k-s}\not\perp V.$$ We claim that for all $-s+1\le i\le 0$, $$\label{eq:strangeAKinV}
A_{k+i}(K_{k+s+2i})\subset V.$$ For $i=0$, this follows from : for any $v\in K_{k+s}$, $$\langle A_{k}(v),A_{k}(v)\rangle=-\sum_{j=-s+1}^{-1} \langle \underbrace{A_{k-j}(v)}_{\in V},\underbrace{A_{k+j}(v)}_{\in V^\perp}\rangle-\sum_{j=1}^{s-1} \langle \underbrace{A_{k-j}(v)}_{\in V^\perp},\underbrace{A_{k+j}(v)}_{\in V}\rangle=0.$$ The vector $A_k(v)$ is therefore isotropic and perpendicular to the maximal isotropic subspace $V$; it follows $A_k(K_{k+s})\subset V$.
If we assume that we have shown for $0\ge i\ge i_0+1$, we can show it for $i=i_0\ge -s+1$ again via : for any $v\in K_{k+s+2i_0}$, $$\begin{aligned}
\langle &A_{k+i_0}(v),A_{k+i_0}(v)\rangle\\
&=-\sum_{j=-s-i_0+1}^{-1} \langle A_{k+i_0-j}(v),\underbrace{A_{k+i_0+j}(v)}_{\in V^\perp}\rangle - \sum_{j=1}^{s+i_0-1}\langle \underbrace{A_{k+i_0-j}(v)}_{\in V^\perp},A_{k+i_0+j}(v)\rangle=0,\end{aligned}$$ since e.g. for $j>0$, we have $$A_{k+i_0+j}(v)\in A_{k+i_0+j}(K_{k+s+2i_0})\subset A_{k+i_0+j}(K_{k+s+2(i_0+j)})\subset V$$ by assumption.
Let $l\ge -s+1$ be such that $$\label{eq:whichKsperp}
A_{k-s}(K_{k+l})\perp V \text{ but } A_{k-s}(K_{k+l+1})\not\perp V$$ and choose a vector $v\in K_{k+l+1}$ with $A_{k-s}(v)\not\perp V$. It follows $A_{k+l}(v)\neq 0$. Look at the reality condition $$\begin{aligned}
0&=\sum_{i+j=-s+l}\langle A_{k+i}(v),A_{k+j}(v)\rangle \nonumber\\
&=\langle A_{k-s}(v),A_{k+l}(v)\rangle+\ldots + \langle A_{k+l}(v),A_{k-s}(v)\rangle. \label{eq:realitydifficult}\end{aligned}$$ Let us look at the summands of each at a time. If neither of $i$ and $j$ is $-s$ and at least one, say $i$, is positive, the respective summand vanishes since then, ${\operatorname{im}}A_{k+i}\subset V$ and ${\operatorname{im}}A_{k+j}\subset V^\perp$. What happens if $i$ and $j$ are both nonpositive, and neither of them is equal to $-s$?
First of all, this is only possible if $-s+l\le 0$, i.e. $l\le s$. We have $A_{k+i}(v)\in A_{k+i}(K_{k+l+1})$ and $A_{k+j}(v)\in A_{k+j}(K_{k+l+1})$, so we see from , that if $k+l+1\le k+s+2i$ or $k+l+1\le k+s+2j$, the respective summand vanishes. If neither of these inequalities is valid, it follows from $i+j=-s+l$ that $$k+l+1> k+s+2i=k+(l-i-j)+2i=k+l+i-j \Longrightarrow 1>i-j$$ and $$k+l+1> k+s+2j=k+(l-i-j)+2j=k+l+j-i \Longrightarrow 1>j-i,$$ so $i=j$. We therefore see: If $-s+l\le 0$ is odd or $-s+l>0$, becomes $$\label{eq:realitydifficultodd}
\langle A_{k-s}(v),A_{k+l}(v)\rangle+\langle A_{k+l}(v),A_{k-s}(v)\rangle=0,$$ and if $-s+l\le 0$ is even, becomes $$\label{eq:realitydifficulteven}
\langle A_{k-s}(v),A_{k+l}(v)\rangle + \langle A_{k+\frac{-s+l}{2}}(v),A_{k+\frac{-s+l}{2}}(v)\rangle + \langle A_{k+l}(v),A_{k-s}(v)\rangle=0.$$
If we are dealing with , we have $$\langle A_{k-s}(v),A_{k+l}(v)\rangle\in i\cdot {\mathbb R},$$ so by Lemma \[lem:Nexist\] there exists a skewsymmetric two-step nilpotent map $N$ with $N(V^\perp)=0$ and $N(A_{k-s}(v))=-A_{k+l}(v)$. We claim that the loop $$\tilde{g}({\lambda})=m_{{\alpha},s+l,N}({\lambda})g({\lambda})=:\sum_i ({\lambda}-{\alpha})^{-i}\tilde{A}_i$$ satisfies ${\epsilon}(\tilde{g})<{\epsilon}(g)$. Since $N(V^\perp)=0$, we have $\tilde{A}_i=NA_{i-s-l}+A_i=A_i$ for all $i>k+l$ by . For $i=k+l$, $$\tilde{A}_{k+l}(K_{k+l})=(NA_{k-s}+A_{k+l})(K_{k+l})=NA_{k-s}(K_{k+l})\subset N(V^\perp)=0$$ by and $\tilde{A}_{k+l}(v)=NA_{k-s}(v)+A_{k+l}(v)=0$ although $A_{k+l}(v)\neq 0$, so ${\epsilon}(\tilde{g})<{\epsilon}(g)$ and we may use induction.
If we are dealing with , there are three subcases. If $\langle A_{k-s}(v),A_{k+l}(v)\rangle\in i\cdot {\mathbb R}$, the middle summand vanishes, so the same argument as before applies with a skewsymmetric map $N$ sending $A_{k-s}(v)$ to $-A_{k+l}(v)$.
If $\langle A_{k-s}(v),A_{k+l}(v)\rangle\in {\mathbb R}$, we can apply Lemma \[lem:N2exist\] to find a map $M:(V\oplus sV)^\perp\to V$ such that $N=M-M^*$ satisfies $$\label{eq:newk+l}
\frac{1}{2}N^2(A_{k-s}(v))+N(A_{k+\frac{-s+l}{2}}(v))+A_{k+l}(v)=0;$$ we want to show that the product $$\begin{aligned}
\tilde{g}({\lambda})&=n_{{\alpha},\frac{s+l}{2},N}({\lambda})g({\lambda})=\left(\Id+({\lambda}-{\alpha})^{-\frac{s+l}{2}} N+\frac{1}{2}({\lambda}-{\alpha})^{-s-l} N^2\right)g({\lambda})\\
&=:\sum_i ({\lambda}-{\alpha})^{-i} \tilde{A}_i \end{aligned}$$ satisfies ${\epsilon}(\tilde{g})<{\epsilon}(g)$. For that, we claim that $$\tilde{A}_{k+l+j}(K_{k+l+j})=\left(\frac{1}{2}N^2A_{k-s+j}+NA_{k+\frac{-s+l}{2}+j} +A_{k+l+j}\right)(K_{k+l+j})=0$$ for all $j\ge 0$. The first summand vanishes for $j= 0$ since $A_{k-s}(K_{k+l})\subset V^\perp$ by , and for $j> 0$ since then, ${\operatorname{im}}A_{k-s+j}\subset V^\perp$ by . The third summand vanishes trivially, so it remains to regard the second. By , $$V\supset A_{k+\frac{-s+l}{2}+j}(K_{k+s+2\frac{-s+l}{2}+2j})=A_{k+\frac{-s+l}{2}+j}(K_{k+l+2j})\supset A_{k+\frac{-s+l}{2}+j}(K_{k+l+j}),$$ so the second summand vanishes as well. Then, shows ${\epsilon}(\tilde{g})<{\epsilon}(g)$.
The third case is that $\langle A_{k-s}(v),A_{k+l}(v)\rangle$ is neither real nor purely imaginary. The idea is to multiply with a simple factor of the type $m$ to make this inner product purely real. Let $N$ be a skewsymmetric map with $N(V^\perp)=0$ and $N^2=0$ such that $$N(A_{k-s}(v))=-i\cdot \frac{{\operatorname{Im}}\langle A_{k-s}(v),A_{k+l}(v)\rangle}{\langle A_{k-s}(v),A_{k+l}(v)\rangle}\cdot A_{k+l}(v);$$ Lemma \[lem:Nexist\] allows us to do so since $$\begin{aligned}
\langle A_{k-s}(v), -i\cdot \frac{{\operatorname{Im}}\langle A_{k-s}(v),A_{k+l}(v)\rangle}{\langle A_{k-s}(v),A_{k+l}(v)\rangle}\cdot A_{k+l}(v)\rangle=-i\cdot {\operatorname{Im}}\langle A_{k-s}(v),A_{k+l}(v)\rangle\in i\cdot {\mathbb R}.\end{aligned}$$ Consider the product $$\begin{aligned}
\tilde{g}({\lambda})=m_{{\alpha},s+l,N}({\lambda})g({\lambda})=\sum_i ({\lambda}-{\alpha})^{-i} \tilde{A}_i,\end{aligned}$$ where $\tilde{A}_i=NA_{i-s-l}+A_i$. Since the kernel of $N$ contains $V^\perp$, we have $\tilde{A}_i=A_i$ for all $i> k+l$, and for $i\le k+l$, they differ only by endomorphisms with values in $V$. We see that $$\begin{aligned}
\langle\tilde{A}_{k-s}(v),\tilde{A}_{k+l}(v)\rangle &=\langle \tilde{A}_{k-s}(v),NA_{k-s}(v)+A_{k+l}(v)\rangle \\
&=-i\cdot {\operatorname{Im}}\langle A_{k-s}(v),A_{k+l}(v)\rangle+\langle A_{k-s}(v),A_{k+l}(v)\rangle\in {\mathbb R}, \end{aligned}$$ and $$\begin{aligned}
2\cdot\langle &\tilde{A}_{k-s}(v),\tilde{A}_{k+l}(v)\rangle + \langle \tilde{A}_{k+\frac{-s+l}{2}}(v),\tilde{A}_{k+\frac{-s+l}{2}}(v)\rangle\\
&=\langle A_{k-s}(v),A_{k+l}(v)\rangle + \langle A_{k+\frac{-s+l}{2}}(v),A_{k+\frac{-s+l}{2}}(v)\rangle + \langle A_{k+l}(v),A_{k-s}(v)\rangle\\
&=0,\end{aligned}$$ so we have reduced to the assumptions of case two: we can now multiply $\tilde{g}$ with a simple factor of the type $n_{{\alpha},\frac{s+l}{2},N}$ for an appropriate $N$ as explained in the previous case.
Nilpotent dressing: Simple poles {#sec:dressing1}
================================
Recall how the Birkhoff factorization theorem yields the *dressing action* [@Terng2000] of the negative loop group ${{{\mathcal L}}}_-({{{\operatorname{GL}}}}(n,{\mathbb C}))$ on the positive loop group ${{{\mathcal L}}}_+({{{\operatorname{GL}}}}(n,{\mathbb C}))$: Given generic $g_\pm\in {{{\mathcal L}}}_\pm({{{\operatorname{GL}}}}(n,{\mathbb C}))$, there exist $\hat{g}_\pm\in {{{\mathcal L}}}_\pm({{{\operatorname{GL}}}}(n,{\mathbb C}))$ such that $g_-g_+=\hat{g}_+ \hat{g}_-$; the dressing action of $g_-$ on $g_+$ is then defined by $g_- * g_+:= \hat{g}_+$. Under presence of a $\tau$–reality and/or a $\sigma$–twisting condition, the dressing action restricts correspondingly (e.g. we obtain an action of ${{{\mathcal L}}}_-^{\tau,\sigma}({{{\operatorname{GL}}}}(n,{\mathbb C}))$ on ${{{\mathcal L}}}_+^{\tau,\sigma}({{{\operatorname{GL}}}}(n,{\mathbb C}))$). Let us consider the dressing action of a nilpotent simple element $$m_{\alpha,1,N}=\Id + \left(\frac{1}{{\lambda}-{\alpha}}\right) N,$$ where $N^2=0$.
\[prop:nilpotentdressing\] Let $f\in {{{\mathcal L}}}_+(GL(n,{\mathbb C}))$, i.e. $f:{\mathbb C}\to GL(n,{\mathbb C})$ is holomorphic on all of ${\mathbb C}$. Let $f_1:=\left.\frac{d}{d{\lambda}}\right|_{{\lambda}={\alpha}} f({\lambda})f({\alpha})^{-1}\in {{{\mathfrak {gl}}}}(n,{\mathbb C})$, and assume that $\Id+Nf_1$ is invertible. If we define $\tilde{N}:=f({\alpha})^{-1}(\Id+Nf_1)^{-1}Nf({\alpha})$, then $\tilde{N}^2=0$ and $$m_{{\alpha},1,N} * f = m_{{\alpha},1,N}f m_{{\alpha},1,\tilde{N}}^{-1}\in {{{\mathcal L}}}_+(GL(n,{\mathbb C})).$$
To prove that $\tilde{N}$ is two-step nilpotent, multiply its defining equation $$\label{eq:defineNtilde}
(\Id+Nf_1)f(\alpha)\tilde{N}f(\alpha)^{-1}=N$$ from the left with $N$ to obtain $$\label{eq:NNtilde}
Nf(\alpha)\tilde{N}f(\alpha)^{-1}=0.$$ Then, multiplying from the right with $f(\alpha)\tilde{N}f(\alpha)^{-1}$, we get $$(\Id+Nf_1)f(\alpha)\tilde{N}^2f(\alpha)^{-1}=0,$$ which is only possible if $\tilde{N}^2=0$.
To show holomorphicity, we only need to show that the loop is holomorphic at $\alpha$, i.e. that the negative terms in its Laurent series expansion at ${\alpha}$ vanish. But the $({\lambda}-{\alpha})^{-2}$–coefficient is $$-Nf(\alpha)\tilde{N}=0$$ using , and the $({\lambda}-{\alpha})^{-1}$–coefficient is $$Nf({\alpha})-f({\alpha})\tilde{N}-Nf_1f({\alpha})\tilde{N} = Nf({\alpha}) -(\Id+Nf_1)(\Id+Nf_1)^{-1}Nf({\alpha})=0.$$ Thus, the new loop is holomorphic.
To give some first application of this proposition, let us quickly review the construction of the ZS-AKNS flows, developed by Zakharov and Shabat [@ZS] and Ablowitz, Kaup, Newell and Segur [@AKNS]. See e.g. Section 2 of [@Terng2000] for a detailed exposition. For a non–zero diagonal matrix $a\in {{{{\mathfrak {sl}}}}}(n,{\mathbb C})$, define $${{{{{\mathfrak {sl}}}}}}(n,{\mathbb C})_a=\{y\in {{{{{\mathfrak {sl}}}}}(n,{\mathbb C})}\mid [a,y]=0\},\quad {{{{{\mathfrak {sl}}}}}}(n,{\mathbb C})_a^\perp=\{y\in {{{{{\mathfrak {sl}}}}}(n,{\mathbb C})}\mid {{\rm tr}}(ay)=0\},$$ and denote by $S({\mathbb R},{{{{\mathfrak {sl}}}}}(n,{\mathbb C})_a^\perp)$ the space of rapidly decaying maps. For $b\in {{{{\mathfrak {sl}}}}}(n,{\mathbb C})$ such that $[a,b]=0$ and any positive integer $j$, there is a unique family of ${{{{\mathfrak {sl}}}}}(n,{\mathbb C})$–valued maps $Q_{b,j}$ such that $$(Q_{b,j}(u))_x+[u,Q_{b,j}(u)]=[Q_{b,j+1}(u),a]$$ and the asymptotic expansion $\sum_{j=0}^\infty Q_{b,j}(u){\lambda}^{-j}$ is conjugate to $b$. Then, the $(b,j)$*–flow* on $S({\mathbb R},{{{{\mathfrak {sl}}}}}(n,{\mathbb C})_a^\perp)$, also called the *$j$–th flow in the ${{{{\mathfrak {sl}}}}}(n,{\mathbb C})$–hierarchy defined by $b$*, is given by $$u_t=(Q_{b,j}(u))_x+[u,Q_{b,j}(u)].$$ If $u$ is a solution of the $j$–th flow defined by $b$, then there exists a unique *trivialization* of $u$, i.e. a solution $E(x,t,{\lambda})$ of $$\begin{aligned}
E^{-1}E_x&=a{\lambda}+ u \\
E^{-1}E_t&= b{\lambda}^j+Q_{b,1}(u){\lambda}^{j-1} + \ldots + Q_{b,j}(u) \\
E(0,0,{\lambda})&=\Id.\end{aligned}$$ Assume that $u$ is a solution admiting a *local reduced wave function* $\omega(x,t,{\lambda})$, as in Definition 2.4 of [@Terng2000]. In particular, $$E(x,t,{\lambda})=\omega(0,0,{\lambda})^{-1}e^{a{\lambda}x + b{\lambda}^j t} \omega(x,t,{\lambda}).$$ Then we can adapt Theorem 4.3 of [@Terng2000] to our situation:
\[prop:dressing\] Let $u$ be a local solution of the $j$–th flow defined by $b$ with trivialization $E$ that admits a local reduced wave function $\omega$. Choose $\alpha\in {\mathbb C}$ and a two-step nilpotent map $N:{\mathbb C}^n\to {\mathbb C}^n$. Let $E_1(x,t)=\left.\frac{d}{d{\lambda}} \right|_{{\lambda}={\alpha}} E(x,t,{\lambda})E(x,t,{\alpha})^{-1}$, and define $\tilde{N}$ as in Proposition \[prop:nilpotentdressing\]: $$\tilde{N}(x,t)=E(x,t,{\alpha})^{-1}(\Id+NE_1(x,t))^{-1}NE(x,t,{\alpha}),$$ wherever this is well-defined. Then, $\tilde{u}(x,t)=u(x,t)-[a,\tilde{N}(x,t)]$ is another solution of the $j$–th flow. Its trivialization is $$\tilde{E}(x,t)=m_{{\alpha},1,N} E(x,t) m_{{\alpha},1,\tilde{N}(x,t)}^{-1}$$ and it has the local reduced wave function $$\tilde{\omega}(x,t,{\lambda})=\omega(x,t,{\lambda})m_{{\alpha},1,\tilde{N}(x,t)}({\lambda})^{-1}.$$
The proof is as in [@Terng2000]. Since $m$ is a local reduced wave function of $u$, we have $$E(x,t,{\lambda})=\omega(0,0,{\lambda})^{-1}e^{a{\lambda}x + b{\lambda}^j t} \omega(x,t,{\lambda}).$$ Thus, if we define $\tilde{E}$ and $\tilde{m}$ as in the proposition, we have $$\begin{aligned}
\tilde{E}(x,t,{\lambda})&=m_{{\alpha},1,N} ({\lambda}) E(x,t,{\lambda}) m_{{\alpha},1,\tilde{N}(x,t)}({\lambda})^{-1}\\
&=m_{{\alpha},1,N} ({\lambda}) \omega(0,0,{\lambda})^{-1}e^{a{\lambda}x + b{\lambda}^j t} \omega(x,t,{\lambda}) m_{{\alpha},1,\tilde{N}(x,t)}({\lambda})^{-1}\\
&=\tilde{\omega}(0,0,{\lambda})^{-1} e^{a{\lambda}x + b{\lambda}^j t} \tilde{\omega}(x,t,{\lambda}).\end{aligned}$$ Therefore, Proposition 2.11 of [@Terng2000] shows that if $$\tilde{\omega}(x,t,{\lambda})=\Id+\tilde{\omega}_1(x,t){\lambda}^{-1}+\tilde{\omega}_2(x,t){\lambda}^{-2}+\ldots$$ is the expansion of $\tilde{\omega}$ at $\infty$, then $\tilde{u}=[a,\tilde{\omega}_1]$ is a solution of the $j$–th flow with trivialization $\tilde{E}$ and local reduced wave function $\tilde{\omega}$. We have $$m_{{\alpha},1,\tilde{N}(x,t)}({\lambda})^{-1}=\Id-\tilde{N}(x,t)({\lambda}-{\alpha})^{-1} = \Id - \tilde{N}(x,t){\lambda}^{-1}+\ldots,$$ and hence $\tilde{\omega}_1=\omega_1-\tilde{N}$. Thus, $\tilde{u}=u-[a,\tilde{N}]$.
\[exam:jthflowsimplepole\] Let us apply Proposition \[prop:dressing\] to the vacuum solution $u=0$ of the $j$–th flow in the ${{{{\mathfrak {sl}}}}}(2,{\mathbb C})$–hierarchy defined by $a=\left(\begin{matrix} 1 & 0 \\ 0 & -1 \end{matrix}\right)$. Its trivialization $E$ is given by $E(x,t,{\lambda})=e^{a({\lambda}x + {\lambda}^j t)}$, and its local reduced wave function is $\omega(x,t,{\lambda})=\Id$. If we denote $$\xi(x,t)=x+j{\alpha}^{j-1}t,$$ then the power series expansion of $E(x,t,{\lambda})E(x,t,{\alpha})^{-1}$ in ${\lambda}={\alpha}$ reads $$E(x,t,{\lambda})E(x,t,{\alpha})^{-1}= \Id+a\xi(x,t)({\lambda}-{\alpha})+\ldots,$$ hence $$\begin{aligned}
\tilde{N}(x,t)&=\left(\begin{matrix}e^{-{\alpha}x-{\alpha}^j t} & 0 \\ 0 & e^{{\alpha}x + {\alpha}^j t}\end{matrix}\right)(\Id+Na\xi(x,t))^{-1}N\left(\begin{matrix}e^{{\alpha}x+{\alpha}^j t} & 0 \\ 0 & e^{-{\alpha}x - {\alpha}^j t}\end{matrix}\right).\end{aligned}$$ We write the nilpotent matrix $N$ in the form $N=\left(\begin{matrix}n_1 & n_2 \\Ên_3 & -n_1 \end{matrix}\right)$, with $\det N=-n_1^2-n_2n_3=0$. A direct calculation shows that $$\begin{aligned}
\tilde{N}(x,t)&= \frac{1}{1+2n_1\xi(x,t)}\left(\begin{matrix}n_1 & n_2 e^{-2{\alpha}x -2{\alpha}^j t} \\ n_3 e^{2{\alpha}x + 2{\alpha}^j t} & -n_1 \end{matrix}\right),\end{aligned}$$ and hence $$\begin{aligned}
\tilde{u}(x,t)&=-[a,\tilde{N}(x,t)]= \frac{2}{1+2n_1\xi(x,t)} \left(\begin{matrix}0 & -n_2 e^{-2{\alpha}x -2{\alpha}^j t} \\ n_3 e^{2{\alpha}x + 2{\alpha}^j t} & 0 \end{matrix}\right).\end{aligned}$$ We see that the new solution $\tilde{u}$ is smooth on all of ${\mathbb R}^2$ if and only if $n_1=0$, i.e. $N=\left(\begin{matrix} 0 & n_2 \\ 0 & 0 \end{matrix}\right)$ or $N=\left(\begin{matrix} 0 & 0 \\ n_3 & 0 \end{matrix}\right)$. If $n_1\neq 0$, then $\tilde{u}$ is singular along the line $x+j{\alpha}^{j-1}t=-\frac{1}{2n_1}$.
Consider the involutions $\sigma$ and $\tau$ on ${{{{\mathfrak {sl}}}}}(n,{\mathbb C})$, given by $\tau(A)=\overline{A}$ and $\sigma(A)=-A^t$. The Cartan decomposition of the symmetric space ${{{\operatorname{SL}}}}(n)/{{{\operatorname{SO}}}}(n)$ is the eigenspace decomposition of $\sigma$, restricted to ${{{{\mathfrak {sl}}}}}(n,{\mathbb R})$: ${{{{\mathfrak {sl}}}}}(n,{\mathbb R})={{{\mathfrak {so}}}}(n)\oplus {{{\mathfrak {p}}}}$. For odd positive integer $j$, the *$j$–th flow in the ${{{\operatorname{SL}}}}(n)/{{{\operatorname{SO}}}}(n)$–hierarchy defined by $b$* is given by the restriction of the $j$–th flow in the ${{{\operatorname{SL}}}}(n,{\mathbb C})$–hierarchy to $S({\mathbb R},{{{{\mathfrak {sl}}}}}(n,{\mathbb R})_{a,\sigma}^\perp)$, where ${{{{\mathfrak {sl}}}}}(n,{\mathbb R})_{a,\sigma}^\perp={{{\mathfrak {so}}}}(n)\cap {{{{\mathfrak {sl}}}}}(n,{\mathbb C})_a^\perp$.
To apply dressing to twisted hierarchies, we need to find products of simple elements that satisfy the twisting condition. For that, a permutability formula is essential:
Let $\alpha \neq \beta$ and $N,M$ satisfy $N^2=M^2=0$. If $$\hat{N}=\left(\Id+\left(\frac{1}{{\alpha}-\beta}\right)M\right)\left(\Id+\left(\frac{1}{{\alpha}-\beta}\right)^2NM\right)^{-1}N\left(\Id-\left(\frac{1}{{\alpha}-\beta}\right)M\right)$$ and $$\hat{M}=\left(\Id+\left(\frac{1}{\beta-{\alpha}}\right)N\right)\left(\Id+\left(\frac{1}{\beta-{\alpha}}\right)^2MN\right)^{-1}M\left(\Id-\left(\frac{1}{\beta-{\alpha}}\right)N\right).$$ are well-defined, then we have $$m_{\beta,1,\hat{M}}m_{{\alpha},1,N}=m_{{\alpha},1, \hat{N}}m_{\beta,1,M}.$$
This follows from Proposition \[prop:dressing\] as usual.
For ${\alpha}\in {\mathbb C}$ and a two-step nilpotent map $N$ such that $$\label{eq:twistingN'}
N'=\left(\Id-\frac{1}{2{\alpha}}N\right)\left(\Id+\frac{1}{4{\alpha}^2} N^tN\right)^{-1}N^t\left(\Id+\frac{1}{2{\alpha}}N\right)$$ is well-defined, let $$\label{eq:salphaN}
s_{{\alpha},N}:=m_{-{\alpha},1,N'}m_{{\alpha},1,N}.$$
\[cor:simpletwistedelements\] For ${\alpha}\in {\mathbb R}$ and $N$ two-step nilpotent such that is well-defined, $s_{{\alpha},N}\in {{{\mathcal L}}}_-^{\tau,\sigma}({{{\operatorname{GL}}}}(n,{\mathbb C}))$.
The third flow in the ${{{\operatorname{SL}}}}(2,{\mathbb R})/{{{\operatorname{SO}}}}(2)$–hierarchy defined by $a=\left(\begin{matrix}1 & 0 \\ 0 & -1 \end{matrix}\right)$ is the modified KdV equation $$q_t=\frac14 (q_{xxx}+6q^2q_x),$$ where $u=\left(\begin{matrix}0 & q \\ -q & 0 \end{matrix}\right)$, see [@Terng2000], Example 3.12. Let ${\alpha}\in {\mathbb R}$ and $N=\left(\begin{matrix}n_1 & n_2 \\ n_3 & -n_1 \end{matrix}\right)$ with $\det N=0$. To perform dressing with $s_{{\alpha},N}$ on the vacuum solution $u=0$, we need to apply Proposition \[prop:dressing\] twice. Using notation and the calculations of Example \[exam:jthflowsimplepole\], one finds the new solution $\hat{q}$ as the upper right entry of $\hat{u}=\tilde{u}-[a,\tilde{N'}]$, where $\tilde{N'}$ is constructed as follows: $$\tilde{N'}(x,t)=\tilde{E}(x,t,-{\alpha})^{-1}(\Id+N'\tilde{E}_1(x,t))^{-1}N'\tilde{E}(x,t,-{\alpha}),$$ where $$\tilde{E}(x,t)=m_{{\alpha},1,N}E(x,t)m_{{\alpha},1,\tilde{N}(x,t)}^{-1}$$ and $\tilde{E}_1(x,t)=\left.\frac{d}{d{\lambda}}\right|_{{\lambda}=-{\alpha}} \tilde{E}(x,t,{\lambda})\tilde{E}(x,t,-{\alpha})^{-1}$. With the help of a computer one finds $$\hat{q}=-{\alpha}e^{2{\alpha}x + 2 {\alpha}^3 t}\frac{(A(x,t)-8n_1) n_3e^{4{\alpha}x+4{\alpha}^3t}+(A(x,t)+8n_1) n_2}{n_3^2 e^{8ax+8a^3t}+B(x,t) e^{4ax+4a^3t}+n_2^2},$$ where $$A(x,t)=16n_1{\alpha}x+48n_1{\alpha}^3t+8{\alpha}$$ and $$B(x,t)=16{\alpha}^2 n_1^2 x^2+96 {\alpha}^4 n_1^2 x t+16 {\alpha}^2 n_1 x+48 {\alpha}^4 n_1 t+144 {\alpha}^6 n_1^2 t^2+4{\alpha}^2+2 n_1^2.$$
Nilpotent dressing: Higher pole order {#sec:dressing2}
=====================================
Dressing with simple elements $m_{{\alpha},k,N}$ with $k\geq 2$ is still possible, but the formulas become more and more complicated as $k$ grows. We only give an idea for $k=2$.
\[prop:dressingpoleorder2\] Let $f\in {{{\mathcal L}}}_+(GL(n,{\mathbb C}))$, choose ${\alpha}\in {\mathbb C}$ and a two-step nilpotent map $N$, and write the power series expansion of $f$ in ${\alpha}$ as $f({\lambda})=\sum_{i=0}^\infty f_i ({\lambda}-{\alpha})^i$. If $X=(Nf_3+f_1)(Nf_2+f_0)^{-1}$, $$M_1=(Nf_2+f_0-XNf_1)^{-1}(XNf_0-Nf_1)$$ and $$M_2=-(Nf_2+f_0)^{-1}(Nf_1M_1+Nf_0)$$ are well-defined, then the loop $$m_{{\alpha},2,N} f \left(\Id+\left(\frac{1}{{\lambda}-{\alpha}}\right)M_1+\left(\frac{1}{{\lambda}-{\alpha}}\right)^2M_2\right)$$ is holomorphic at ${\alpha}$.
The principal part of the Laurent series in ${\alpha}$ of the new loop reads $$\begin{aligned}
&\left(\frac{1}{{\lambda}-{\alpha}}\right)^4 Nf_0M_2 \label{eq:dressingterm4} \\
+&\left(\frac{1}{{\lambda}-{\alpha}}\right)^3[Nf_1M_2+Nf_0M_1] \label{eq:dressingterm3}\\
+&\left(\frac{1}{{\lambda}-{\alpha}}\right)^2[(Nf_2+f_0)M_2 +Nf_1M_1+Nf_0] \label{eq:dressingterm2}\\
+&\left(\frac{1}{{\lambda}-{\alpha}}\right)^1[(Nf_3+f_1)M_2+(Nf_2+f_0)M_1+Nf_1]. \label{eq:dressingterm1}\end{aligned}$$ If the terms and vanish, then also and , as one can see by multiplying them from the left with the two-step nilpotent map $N$. But and vanish if $M_1$ and $M_2$ are chosen as in the statement of the proposition.
Already the formulas in Example \[exam:jthflowsimplepole\] become very complicated if one replaces the simple element $m_{{\alpha},1,N}$ by $m_{{\alpha},2,N}$. Therefore, we restrict ourselves to the case of the third flow, and the simple element having its pole at $0$. For $u=\left(\begin{matrix} 0 & q \\ r & 0 \end{matrix} \right)$, the third flow in the ${{{{\mathfrak {sl}}}}}(2,{\mathbb C})$–hierarchy defined by $a=\left(\begin{matrix}1 & 0 \\ 0 & -1 \end{matrix}\right)$ is given by $$q_t=\frac14 (q_{xxx}-6qrq_x),\quad r_t=\frac14 (r_{xxx}-6qrr_x),$$ see [@Terng2000], Example 2.8. Applying Proposition \[prop:dressingpoleorder2\] to the vacuum solution $u=0$ and the simple element $m_{0,2,N}$, with $N=\left(\begin{matrix}n_1 & n_2 \\ n_3 & -n_1\end{matrix}\right)$ satisfying $\det N=0$, a direct calculation provides the solution $$\tilde{u}=\frac{4}{4n_1^2x^4-12n_1^2xt+3} \left(\begin{matrix} 0 & n_2 (2n_1x^3+3x+3n_1t) \\ -n_3(2n_1x^3-3x+3n_1t) & 0 \end{matrix}\right).$$
The $n$–dimensional systems {#sec:systems}
===========================
Let $U/K$ be a rank $n$ symmetric space with Cartan decomposition ${{{\mathfrak {u}}}}={{{\mathfrak {k}}}}\oplus {{{\mathfrak {p}}}}$, and choose a maximal abelian subalgebra ${{{\mathfrak {a}}}}\subset {{{\mathfrak {p}}}}$ with basis $a_1,\ldots,a_n$. Recall that the $n$–dimensional system associated to $U/K$ is the following system of first order partial differential equations for $v:{\mathbb R}^n\to {{{\mathfrak {a}}}}^\perp\cap {{{\mathfrak {p}}}}$: $$[a_i,v_{x_j}]-[a_j,v_{x_i}]=[[a_i,v],[a_j,v]],$$ which is independent of the choice of basis.
Associated to any symmetric space $U/K$ is its dual symmetric space $U^*/K$, which has the Cartan decomposition ${{{\mathfrak {u}}}}^*={{{\mathfrak {k}}}}\oplus i{{{\mathfrak {p}}}}$. Choosing the maximal abelian subspace $i{{{\mathfrak {a}}}}\subset i{{{\mathfrak {p}}}}$ with basis $ia_1,\ldots,ia_n$, we see
$v:{\mathbb R}^n\to {{{\mathfrak {a}}}}^\perp \cap {{{\mathfrak {p}}}}$ is a solution of the $U/K$–system if and only if $-iv:{\mathbb R}^n\to (i{{{\mathfrak {a}}}})^\perp \cap i{{{\mathfrak {p}}}}$ is a solution of the $U^*/K$–system.
Therefore, the $U/K$–system and the $U^*/K$–system are the same, and we do not only have a dressing action of the rational loop group ${{{\mathcal L}}}_-^{\tau,\sigma}(U)$ on the space of solutions of the $U/K$–system, but also one of ${{{\mathcal L}}}_-^{\tau,\sigma}(U^*)$. Furthermore, whatever geometric interpretation of the solutions of the particular $U/K$–system has been found, also applies to the $U^*/K$–system.
Let us apply this observation to the system associated to the symmetric space ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$, which we now have seen to be the same as the system associated to ${{{\operatorname{U}}}}(n)/{{{\operatorname{O}}}}(n)$. The Cartan decomposition of ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$ is ${{{\mathfrak {gl}}}}(n)={{{\mathfrak {so}}}}(n)\oplus {{{\mathfrak {p}}}}$, where ${{{\mathfrak {p}}}}$ is the space of symmetric matrices. Let $a_i=e_{ii}$ be the standard basis of the Cartan subalgebra ${{{\mathfrak {a}}}}\subset {{{\mathfrak {p}}}}$ of diagonal matrices, i.e. $a_i$ is the matrix with zeros everywhere except a $1$ at the $ii$–entry. Then, $\beta:{\mathbb R}^n\to {{{\mathfrak {p}}}}$ is a solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system if and only if $$\label{eq:GLnOnsystem}
\begin{cases} (\beta_{ij})_{x_k}=\beta_{ik}\beta_{kj} & i,j,k \text{ distinct} \\
(\beta_{ij})_{x_i}+(\beta_{ij})_{x_j}+\sum_k \beta_{ik}\beta_{kj}=0 & i\neq j,
\end{cases}$$ see [@Terng2006]. On the other hand, $\beta$ is a solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system if and only if $\omega_{\lambda}= \sum_i ({\lambda}a_i + [a_i,\beta])dx_i$ is flat for all ${\lambda}$. In this case, there is a unique frame $E(x,{\lambda})$ satisfying $$E^{-1}dE=\sum_i ({\lambda}a_i + [a_i,\beta])dx_i,\quad E(0,{\lambda})=\Id.$$ This frame satisfies the ${{{\operatorname{GL}}}}(n,{\mathbb R})$–reality and the ${{{\operatorname{O}}}}(n)$–twisting condition: $$E(x,{\overline{{\lambda}}})=\overline{E(x,{\lambda})},\qquad E(x,-{\lambda})^t E(x,{\lambda})=\Id.$$
\[rem:GLnUn\] Observe that $F(x,{\lambda})=E(x,i{\lambda})$ satisfies the ${{{\operatorname{U}}}}(n)$–reality condition: $$F(x,\overline{{\lambda}})^*F(x,{\lambda})=E(x,i\overline{{\lambda}})^*E(x,i{\lambda}) = E(x,-i{\lambda})^tE(x,i{\lambda})=\Id.$$ This is not surprising as $F^{-1}dF=\sum_i (i{\lambda}a_i + [a_i,\beta])dx_i$, i.e. $F$ is the frame of the solution $-i\beta$ of the ${{{\operatorname{U}}}}(n)/{{{\operatorname{O}}}}(n)$–system.
Let ${\alpha}\in {\mathbb R}$ and $N$ be two-step nilpotent such that both $$\tilde{N}(x)=E(x,{\alpha})^{-1}(\Id+NE_1(x))^{-1}NE(x,{\alpha})$$ and $$\tilde{N'}(x)=\tilde{E}(x,-{\alpha})^{-1}(\Id+N'\tilde{E}_1(x))^{-1}N\tilde{E}(x,-{\alpha})$$ are well-defined. Here, $E_1$ and $\tilde{E}_1$ are given by $E_1(x):=\left.\frac{d}{d{\lambda}}\right|_{{\lambda}={\alpha}} E(x,{\lambda})E(x,{\alpha})^{-1}$ and $\tilde{E}_1(x):=\left.\frac{d}{d{\lambda}}\right|_{{\lambda}=-{\alpha}} \tilde{E}(x,{\lambda}) \tilde{E}(x,-{\alpha})^{-1}$, and $N'$ is given by . Now we may consider the dressing action of a simple element $s_{{\alpha},N}$ on $E$ and obtain the new frame $$\hat{E}=s_{{\alpha},N}*E=m_{-{\alpha},1,N'}m_{{\alpha},1,N} E m_{{\alpha},1,\tilde{N}}^{-1} m_{-{\alpha},1,\tilde{N'}}^{-1}.$$ The calculation of $\hat{E}^{-1}d\hat{E}$ is then implicit in Proposition \[prop:dressing\]: $$\hat{E}^{-1}d\hat{E}=\sum_i ({\lambda}a_i + [a_i,\beta-\tilde{N}-\tilde{N'}])dx_i,$$ We have proved:
\[prop:dressingGLnOn\] Let $\beta$ be a solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system, and $E(x,{\lambda})$ its frame. Then $$s_{{\alpha},N}* \beta= \beta-(\tilde{N}+ \tilde{N'})_*$$ is the solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system obtained by dressing with $s_{{\alpha},N}$. Here, we denote by $(\tilde{N}+ \tilde{N'})_*$ the trace–free part of $\tilde{N}+ \tilde{N'}$.
Let us quickly review parts of the connection between solutions of the ${{{\operatorname{U}}}}(n)/{{{\operatorname{O}}}}(n)$–system (resp. the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system) and Egoroff metrics, as found by Terng and Wang [@Terng2006]. A local orthogonal system $(x_i)$ of ${\mathbb R}^n$ is called Egoroff if there exists a function $\phi(x)$ such that the Euclidean metric $ds^2$ written in this coordinate system is of the form $ds^2=\sum_i h_i^2(x) dx_i^2$, where $h_i^2(x)=\frac{\partial \phi}{\partial x_i}$. The rotation coefficient matrix $\beta$ of the Egoroff metric $\sum h_i^2dx_i^2$ is defined by $\beta_{ij}=\frac{(h_i)_{x_j}}{h_j}$ for $i\neq j$, and $\beta_{ii}=0$. If $\beta$ is the rotation coefficient matrix of a flat Egoroff metric, then $\beta$ solves , i.e. is a solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system. Conversely, if $\beta$ is a solution of the ${{{\operatorname{GL}}}}(n)/{{{\operatorname{O}}}}(n)$–system, then $\beta$ is the rotation coefficient matrix of a flat Egoroff metric.
A flat Egoroff metric is called $\partial$–invariant or spherical, if $\partial h_i=0$, where $\partial=\sum_j \frac{\partial}{\partial x_i}$ — see Proposition 2.4 of [@Terng2006], where four equivalent conditions for being $\partial$–invariant are listed. Recall also statements (1) and (3) of Theorem 2.5 of [@Terng2006]: If $\sum_i h_i^2 dx_i^2$ is a $\partial$–invariant flat Egoroff metric, and $E$ the frame of $\sum_i ({\lambda}a_i + [a_i,\beta])dx_i$, then $h$ can be reconstructed via the formula $E(x,0)h(x)=h(0)$. Furthermore, there is an associated family of flat Lagrangian immersions into ${\mathbb C}^n$ given by $$X(x,{\lambda})=-i{\lambda}^{-1}(E(x,i{\lambda})h(x)-h(0)).$$ Note that the additional factor $i$ in front of ${\lambda}$ is explained by Remark \[rem:GLnUn\]. Then, we have the following analogue of Theorem 4.2 of [@Terng2006]:
Let $\sum_i h_i^2dx_i^2$ be a $\partial$–invariant flat Egoroff metric with coefficient matrix $\beta$ and frame $E(x,{\lambda})$. Let $c=h(0)$. If $\hat{E}=s_{{\alpha},N} * E$ and $\hat{c}$ is a constant, then we have a new $\partial$–invariant flat Egoroff metric $$\hat{h}(x)=\hat{E}(x,0)c$$ with associated family of flat Lagrangian submanifolds $$\hat{X}(x,{\lambda})=-i{\lambda}^{-1} (\hat{E}(x,i{\lambda})\hat{E}(x,0)^{-1}\hat{c}-\hat{c}).$$
[10]{} \[1\][`#1`]{} \[2\]\[\][[\#2](#2)]{}
Mark J. Ablowitz, David J. Kaup, Alan C. Newell, and Harvey Segur: *The inverse scattering transform – Fourier analysis for nonlinear problems*, Studies in Appl. Math. **53** (1974), 249–315.
Neil Donaldson, Daniel Fox, and Oliver Goertsches: *Generators for rational loop groups*, [arXiv:0803.0029]{}. to appear in Trans. Amer. Mat. Soc.
Chuu-Lian Terng, Karen Uhlenbeck: *Bäcklund transformations and loop group actions*, Comm. Pure Appl. Math. **53**(1): (2000) 1–75.
Chuu-Lian Terng: *Geometries and symmetries of soliton equations and integrable elliptic equations*, Adv. Stud. Pure Math. **51** (2008), Surveys on geometry and integrable systems, 401–488.
Chuu-Lian Terng, Erxiao Wang: *Transformations of flat Lagrangian immersions and Egoroff nets*, Asian J. Math. **12**(1): (2008), 99–119.
Karen Uhlenbeck: *Harmonic maps into [L]{}ie groups: classical solutions of the chiral model*, J. Differential Geom. **30**(1): (1989) 1–50.
Vladimir E. Zakharov and Alexey B. Šabat: *Integration of nonlinear equations of mathematical physics by the method of the inverse scattering problem, II*, Functional Anal. Appl. **13** (1979), no. 3, 166–173.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In recent work, we constructed a new near octagon $\mathcal{G}$ from certain involutions of the finite simple group $G_2(4)$ and showed a correspondence between the Suzuki tower of finite simple groups, $L_3(2) < U_3(3) < J_2 < G_2(4) < Suz$, and the tower of near polygons, $\mathrm{H}(2,1) \subset \mathrm{H}(2)^D \subset \mathsf{HJ} \subset \mathcal{G}$. Here we characterize each of these near polygons (except for the first one) as the unique near polygon of the given order and diameter containing an isometrically embedded copy of the previous near polygon of the tower. In particular, our characterization of the Hall-Janko near octagon ${\mathsf{HJ}}$ is similar to an earlier characterization due to Cohen and Tits who proved that it is the unique regular near octagon with parameters $(2, 4; 0, 3)$, but instead of regularity we assume existence of an isometrically embedded dual split Cayley hexagon, $\mathrm{H}(2)^D$. We also give a complete classification of near hexagons of order $(2, 2)$ and use it to prove the uniqueness result for $\mathrm{H}(2)^D$.'
author:
- Anurag Bishnoi and Bart De Bruyn
title: Characterizations of the Suzuki tower near polygons
---
**Keywords:** near polygon, generalized polygon, Suzuki tower\
**MSC2010:** 05E18, 51E12, 51E25
Introduction {#sec1}
============
Near polygons, introduced by Shult and Yanushka [@Sh-Ya], form an important class of point-line geometries. They are related to polar spaces [@Cameron], distance-regular graphs [@BCN], finite simple groups [@Wilson], and they have a rich theory of their own [@bdb-book]. While dual polar spaces and generalized polygons (which can be seen as a special class of near polygons) are related to classical groups and exceptional groups of Lie type, there are also some near polygons corresponding to sporadic simple groups. One of the most famous examples is the Hall-Janko near octagon ${\mathsf{HJ}}$ (also known as the Cohen-Tits near octagon [@Br]) constructed by Cohen [@Co] using the conjugacy class of $315$ central involutions of the Hall-Janko sporadic simple group $J_2$. The collinearity graph of ${\mathsf{HJ}}$ is a distance-regular graph with intersection array $\{10, 8, 8, 2; 1, 1, 4, 5\}$ which is uniquely determined by these intersection numbers, or equivalently, ${\mathsf{HJ}}$ is the unique regular near octagon with parameters $(2,4; 0, 3)$ (cf. [@Co-Ti Thm. 3]). It is well known that ${\mathsf{HJ}}$ contains subgeometries isomorphic to the dual split Cayley hexagon of order $(2,2)$, henceforth denoted by $\mathrm{H}(2)^D$. One of our main results in this paper is an alternative characterization of ${\mathsf{HJ}}$ where instead of assuming regularity with parameters $(2, 4; 0, 3)$, we only assume the order $(2,4)$ and the existence of an isometrically embedded $\mathrm{H}(2)^D$ (see Theorem \[thm:HJ\]).
In [@ab-bdb:2] we constructed a new near octagon ${\mathcal{G}}$ of order $(2, 10)$ which has its full automorphism group isomorphic to $G_2(4){:}2$ and proved that ${\mathcal{G}}$ contains ${\mathsf{HJ}}$ as an isometrically embedded subgeometry. Our main construction in [@ab-bdb:2] was inspired by the construction of ${\mathsf{HJ}}$ given by Cohen [@Co] where the points of ${\mathsf{HJ}}$ are the $315$ central involutions of the group $J_2$, and all the lines are the triples $\{x, y, xy\}$ formed by taking two distinct commuting central involutions $x$ and $y$. We recall the construction of ${\mathcal{G}}$ here:
*Let $C$ be the conjugacy class $2A$ [^1] consisting of the $4095$ central involutions of the group $G = G_2(4){:}2$. Then ${\mathcal{G}}$ is the point-line geometry with point set $C$ whose lines are all $15015$ three element subsets $\{x, y, xy\}$ of $C$ where $x$ and $y$ are distinct commuting involutions in $C$ satisfying $[G : N_G(\langle x, y \rangle)] \in \{1365, 13650\}$.*
The Suzuki tower (a name coined by Tits, as it has been mentioned in [@Leemans05]) is the sequence of five finite simple groups $L_3(2) < U_3(3) < J_2 < G_2(4) < Suz$ where each group (except the last one) is a maximal subgroup of the next group in the sequence. The five groups in the Suzuki tower correspond to five vertex-transitive graphs $\Gamma_0$, $\Gamma_1$, $\Gamma_2$, $\Gamma_3$, $\Gamma_4$ where the automorphism groups of the $\Gamma_i$’s are $L_3(2){:}2$, $U_3(3){:}2$, $J_2{:}2$, $G_2(4){:}2$ and $Suz{:}2$, respectively [@Suzuki]. Using the properties of ${\mathcal{G}}$ that we have described in [@ab-bdb:2], we gave a correspondence between the Suzuki tower and the tower of four near polygons $\mathrm{H}(2,1) \subset \mathrm{H}(2)^D \subset {\mathsf{HJ}}\subset {\mathcal{G}}$, where $\mathrm{H}(2,1)$ is the unique generalized hexagon of order $(2,1)$ (which is the point-line dual of the incidence graph of the Fano plane). The first four simple groups in the Suzuki tower act as automorphism groups of the corresponding near polygons. Moreover, we were able to construct all of the graphs $\Gamma_i$’s using these near polygons. We note that there are various other geometries associated with the Suzuki tower that have been studied before, see for example [@Ne82], [@S92] and [@Leemans05].
For all $i \in \{1, 2, 3, 4\}$ the graph $\Gamma_i$ is strongly regular and contains $\Gamma_{i-1}$ as a local graph. In fact, it has been proved in [@Pa] that for $i \in \{ 2, 3 \}$ the graph $\Gamma_i$ is the unique connected graph which is locally $\Gamma_{i-1}$, while there exist two connected graphs which are locally $\Gamma_3$, the graph $\Gamma_4$ and the graph $3\Gamma_4$ constructed by Soicher [@S93] which is a $3$-fold antipodal cover of the Suzuki graph $\Gamma_4$. In [@BFS93], it was proved that there are up to isomorphism three connected graphs that are locally $\Gamma_0$, one of which is the $U_3(3)$-graph $\Gamma_1$. In the present paper, we give some characterizations of the geometries in the “Suzuki tower of near polygons” by proving the following three theorems.
\[thm:H2\] The dual split Cayley hexagon of order $(2,2)$ is the unique near hexagon of order $(2,2)$ that contains the generalized hexagon $\mathrm{H}(2,1)$ as an isometrically embedded subgeometry.
\[thm:HJ\] The Hall-Janko near octagon is the unique near octagon of order $(2,4)$ that contains the dual split Cayley hexagon of order $(2,2)$ as an isometrically embedded subgeometry.
\[thm:G24\] The $G_2(4)$ near octagon is the unique near octagon of order $(2,10)$ that contains the Hall-Janko near octagon as an isometrically embedded subgeometry.
It is known that the second constituent of the ${\mathsf{HJ}}$-graph $\Gamma_2$ is isomorphic to the distance-$2$ graph of the generalized hexagon $\mathrm{H}(2)^D$. Similarly, the second constituent of the $G_2(4)$-graph $\Gamma_3$ is isomorphic to the distance-2 graph of the Hall-Janko near octagon ${\mathsf{HJ}}$. It has been communicated to us by one of the referees that a similar result holds for the near octagon ${\mathcal{G}}$. Let $\Gamma = 3\Gamma_4$ be the triple cover of the Suzuki graph $\Gamma_4$ as constructed by Soicher [@S93]. Then $G = \mathrm{Aut}(\Gamma) \cong 3 \cdot Suz{:}2$ acts distance-transitively on the vertices and the vertices split into orbits of length $1, 416, 4095, 832, 2$ under the action of the stabilizer $G_x \cong G_2(4){:}2$ of a fixed vertex $x$ of $\Gamma$. Let $V$ be the set of $4095$ vertices at distance $2$ from $x$ in $\Gamma$. There exists a bijective correspondence between $V$ and the involution class $2A$ of $G_x$ (i.e., the points of the near octagon ${\mathcal{G}}$), see Pasechnik [@Pa Proposition 3]. The collinearity graph of ${\mathcal{G}}$, which uniquely determines the geometry ${\mathcal{G}}$, can be constructed from $V$ by taking the edges as pairs $\{u, v\} \subset V$ whose stabilizer has index $4095$ or $40950$ in $G_x$. The subgraph of $\Gamma$ induced on $V$ is isomorphic to the graph defined on the point set of $\mathcal{G}$, by calling two vertices adjacent whenever they lie at distance 2 and have a unique common neighbour. These claims can be verified with the aid of the permutation representation of $3 \cdot Suz{:}2$ on the $5346$ vertices of $\Gamma$ given in the ATLAS [^2].
We do not know if there is a near polygon which corresponds to $Suz$, but certainly the involution geometry of $Suz$ studied in the literature [@Yoshiara88; @Bardoe96] is not a near polygon; we can directly see from the suborbit diagram [@Bardoe96 Fig. 1] that there are point-line pairs $(p, L)$ in this geometry where every point of the line $L$ is at the same distance $4$ from $p$. However, this involution geometry is a near $9$-gon in the sense of [@BCN Sec. 6.4]. Indeed, it is clear from the suborbit diagram [@Bardoe96 Fig. 1] that for every point-line pair $(p, L)$ with ${\mathrm{d}}(p, L) < 4$, there is a unique point on $L$ nearest to $p$. Similarly, one can verify that the involution geometry of the Conway group $Co_1$ [@Bardoe99 Fig. 1] which contains $Suz$ is a near $11$-gon. The techniques involving valuations of near polygons used in this paper do not work for near $(2d+1)$-gons, and thus we are unable to give similar characterizations of these involution geometries. We do not know if there are more near polygons hiding in larger groups that can extend the Suzuki tower of near polygons (see [@S92] for a possible extension of the Suzuki tower of groups).
This paper is organised as follows. In Section \[sec:Prelim\] we give the basic properties of near polygons and their valuations that we will use in our proofs. In Section \[sec:H2\] we first prove that up to isomorphism there are only three near hexagons of order $(2, 2)$, the two generalized hexagons of order $(2,2)$ (split Cayley hexagon and its dual) and the product near polygon $\mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$. Then we show that out of these three near hexagons only the dual split Cayley hexagon contains $\mathrm{H}(2,1)$ as a subgeometry, thus proving Theorem \[thm:H2\].
Sections \[sec:HJ\] and \[sec:G24\] are then devoted to proving Theorems \[thm:HJ\] and \[thm:G24\], respectively. The main tool that we use in these proofs is the theory of valuations of near polygons introduced by De Bruyn and Vandencasteele in [@bdb-pvdc]. This is a purely combinatorial tool which has since been developed and used to obtain several classification results for near polygons (see [@bdb-val] for a survey). In [@ab-bdb:1] we used the so-called valuation geometry of $\mathrm{H}(2)^D$, obtained with the help of a computer, to prove that it is not contained in any semi-finite near hexagon as a full isometrically embedded subgeometry, thus giving a partial answer to the famous open problem about existence of semi-finite generalized polygons in this particular case. Also, the $G_2(4)$ near octagon ${\mathcal{G}}$ was first constructed using the valuation geometry of ${\mathsf{HJ}}$ (see [@ab-bdb:2 Appendix]). The algorithm provided in [@ab-bdb:1] to compute the valuation geometry of $\mathrm{H}(2)^D$ can also be used to compute the valuation geometry of ${\mathsf{HJ}}$ and both these valuation geometries will be crucial to our proofs. We have made the computer code, written in GAP [@Gap], required for these computations available online [@ab-bdb:code]. The valuation geometry of $\mathrm{H}(2)^D$ is constructed in the file `Suz1.g` while the valuation geometry of ${\mathsf{HJ}}$ is constructed in `Suz2.g`. This code is also used to prove Lemmas \[lemHJ:connected\], \[lem:ValBC\], \[lem:ValB\], \[lem:ValC\] and \[lemG:connected\].
Near polygons and their valuations {#sec:Prelim}
==================================
In a partial linear space we can identify each line with the set of points incident with it, and then the incidence relation becomes set inclusion. We would do so whenever it is convenient. A partial linear space is said to have [*order*]{} $(s,t)$ if every line is incident with precisely $s+1$ points and if every point is incident with precisely $t+1$ lines. All distances in a partial linear space ${\mathcal{S}}=({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ will be measured in its collinearity graph, and denoted by ${\mathrm{d}}_{{\mathcal{S}}}(\cdot, \cdot)$ or by ${\mathrm{d}}(\cdot,\cdot)$ when no confusion could arise. If $x$ is a point of ${\mathcal{S}}$, then $\Gamma_i(x)$, for $i \in \mathbb{N}$, will denote the set of points of ${\mathcal{S}}$ at distance $i$ from $x$. Similarly, for a nonempty subset $X \subseteq {\mathcal{P}}$, we define $$\Gamma_i(X) = \{x \in {\mathcal{P}} \, | \, i = {\mathrm{d}}(x, X) := \min_{y \in X} {\mathrm{d}}(x, y)\}.$$ Let $X$ be a set of points of ${\mathcal{S}}$. Then $X$ is called a *subspace* if for every pair of distinct collinear points in $X$, the line joining those points is contained in $X$. A subspace $X$ is called *convex*, if for every pair of points $x, y \in X$ all points on a shortest path between $x$ and $y$ are contained in $X$.
A near $2d$-gon with $d \in \mathbb{N}$ is a partial linear space ${\mathcal{N}} = ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ defined by the following axioms:
1. the collinearity graph of ${\mathcal{N}}$ is a connected graph of diameter $d$;
2. for every $x \in {\mathcal{P}}$ and every line $L \in {\mathcal{L}}$ there exists a unique point $\pi_L(x)$ incident with $L$ that is nearest to $x$.
It follows that every near $4$-gon is a possibly degenerate generalized quadrangle (see [@Pa-Th]). In fact, generalized $2d$-gons with $d \geq 2$ are near $2d$-gons that satisfy the following extra property:
1. For every pair of points $x, y$ at distance $i \in \{1, 2, \dots, d-1\}$ from each other, the set $\{z \in {\mathcal{P}} \, | \, {\mathrm{d}}(x, z) = i - 1, {\mathrm{d}}(z, y) = 1\}$ is a singleton.
Two lines $L_1$ and $L_2$ of a near polygon are called *parallel* at distance $i$ if ${\mathrm{d}}(L_1, L_2) = i$ and for each point $x_1$ on $L_1$, there is a unique point $x_2$ on $L_2$ such that ${\mathrm{d}}(x_1, x_2) = i$, or equivalently, if ${\mathrm{d}}(L_1, L_2) = i$ and for each point $x_2$ on $L_2$, there is a unique point $x_1$ on $L_1$ such that ${\mathrm{d}}(x_1, x_2) = i$.
Let ${\mathcal{N}} = ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ be a subgeometry of another near polygon ${\mathcal{N}}' = ({\mathcal{P}}', {\mathcal{L}}', {\mathrm{I}}')$, i.e., ${\mathcal{P}} \subseteq {\mathcal{P}}'$, ${\mathcal{L}} \subseteq {\mathcal{L}}'$ and ${\mathrm{I}}= {\mathrm{I}}' \cap ({\mathcal{P}} \times {\mathcal{L}})$. Then ${\mathcal{N}}$ is called a *full* subgeometry of ${\mathcal{N}}'$ if for every line $L \in {\mathcal{L}}$ we have $\{x \in {\mathcal{P}} : x ~{\mathrm{I}}~ L\} = \{x \in {\mathcal{P}}' : x ~{\mathrm{I}}'~ L \}$ and it is called *isometrically embedded* if for all $x, y \in {\mathcal{P}}$ we have ${\mathrm{d}}_{{\mathcal{N}}}(x, y) = {\mathrm{d}}_{{\mathcal{N}}'}(x, y)$.
For every nonempty convex subspace $X$ of a near polygon ${\mathcal{N}} = ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$, we can define a full isometrically embedded subgeometry of ${\mathcal{N}}$ by taking the elements of $X$ as the points of the subgeometry and the lines $L \in {\mathcal{L}}$ satisfying $\{x \in X : x ~{\mathrm{I}}~ L\} = \{x \in {\mathcal{P}}: x ~{\mathrm{I}}~ L\}$ as the lines of the subgeometry, with the incidence relation being the one induced by ${\mathrm{I}}$.
The most important class of subgeometries of near polygons are the *quads*. They were introduced by Shult and Yanushka in [@Sh-Ya], and the theory of near polygons with quads was further developed by Brouwer and Wilbrink in [@BW]. A quad $Q$ of a near polygon ${\mathcal{N}}$ is a set of points that satisfies the following properties.
1. The maximum distance between two points of $Q$ is $2$.
2. If $x, y \in Q$ are distinct and collinear, then every point incident with the line $xy$ lies in $Q$.
3. If $x$ and $y$ are two non-collinear points in $Q$, then every common neighbor of $x$ and $y$ is in $Q$.
4. The subgeometry of ${\mathcal{N}}$ determined by those points and lines that are contained in $Q$ is a non-degenerate generalized quadrangle.
Or succinctly, a quad $Q$ is a convex subspace of ${\mathcal{N}}$ that induces a subgeometry isomorphic to a non-degenerate generalized quadrangle. Sufficient conditions for existence of quads were given by Shult and Yanushka in [@Sh-Ya Proposition 2.5], where they proved that if $a$ and $b$ are two points of a near polygon at distance $2$ from each other, and if $c$ and $d$ are two common neighbors of $a$ and $b$ such that at least one of the lines $ac, ad, bc, bd$ contains at least three points, then $a$ and $b$ are contained in a unique quad. We will implicitly use this result in our proofs.
In a near polygon with three points per line, each quad induces a generalized quadrangle of order $(2,t)$, and each such generalized quadrangle is isomorphic to either the $(3 \times 3)$-grid, $W(2)$ or $Q(5, 2)$ (see eg. Section 1.10 in [@bdb-book]). We will call the quad a [*grid-quad*]{}, a [*$W(2)$-quad*]{} or a [*$Q(5,2)$-quad*]{} depending on which case occurs. The $(3 \times 3)$-grid is an example of a so-called product near polygon (see Section 1.6 in [@bdb-book] for the precise definition) as it can be obtained by taking the direct product $\mathbb{L}_3 \times \mathbb{L}_3$, where $\mathbb{L}_3$ is a line with three points. We will see another example of a product near polygon, $\mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$, in Section \[sec:H2\], which is a near hexagon of order $(2, 2)$.
Let ${\mathcal{N}}= ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ be a near $2d$-gon. A function $f: {\mathcal{P}}\rightarrow \mathbb{Z}$ is called a *semi-valuation* of ${\mathcal{N}}$ if every line $L$ contains unique point $x_L$ such that $f(x)=f(x_L)+1$ for every point $x$ of $L$ distinct from $x_L$. A [*valuation*]{} of ${\mathcal{N}}$ is a semi-valuation $f$ for which $\min_{x \in {\mathcal{P}}}f(x) = 0$.
Let $f$ be a valuation of ${\mathcal{N}}$. Then $M_f$ denotes the maximum value attained by $f$ and $\mathcal{O}_f$ denotes the set of points with $f$-value $0$. It can be easily checked that the set of points of ${\mathcal{N}}$ that have $f$-value strictly less than $M_f$ is a [*hyperplane*]{} $H_f$ of ${\mathcal{N}}$, i.e., a proper subset of ${\mathcal{P}}$ having the property that each line has either one or all its points in it. From the near polygon axioms it also follows that the function $f : {\mathcal{P}}\rightarrow \mathbb{Z}$ defined by $f(y) = {\mathrm{d}}(x, y)$, where $x$ is a fixed point of ${\mathcal{N}}$, is a valuation of ${\mathcal{N}}$. This is known as the *classical valuation* of ${\mathcal{N}}$ with center $x$. Two valuations $f_1$ and $f_2$ of ${\mathcal{N}}$ are called [*isomorphic*]{} if there exists an automorphism $\theta$ of ${\mathcal{N}}$ such that $f_2 = f_1 \circ \theta$. Thus, all classical valuations of ${\mathcal{N}}$ are isomorphic if $\mathrm{Aut}({\mathcal{N}})$ acts transitively on the points of ${\mathcal{N}}$, which is true for all the Suzuki tower near polygons.
Two valuations $f_1$ and $f_2$ of ${\mathcal{N}}$ are called *neighboring valuations* if there exists an $\epsilon \in \mathbb{Z}$ such that $|f_1(x) - f_2(x) + \epsilon| \leq 1$ for every point $x$ of $\mathcal{N}$. The number $\epsilon$ (necessarily belonging to $\{ -1,0,1 \}$) is uniquely determined, except when $f_1=f_2$, in which case there are three possible values for $\epsilon$, namely $-1$, $0$ and $1$. Suppose ${\mathcal{N}}$ is a near polygon in which every line has precisely three points. Let $f_1, f_2$ be two neighboring valuations and let $\epsilon \in \{-1, 0, 1\}$ be such that $|f_1(x) - f_2(x) + \epsilon| \leq 1$ for all points $x$. For points $x$ satisfying $f_1(x) = f_2(x) - \epsilon$ we define $f_3'(x) = f_1(x) - 1 = f_2(x) - \epsilon - 1$ and for other points $x$ we define $f_3'(x) = \max \{f_1(x), f_2(x) - \epsilon\}$. Let $m$ be the minimal value attained by $f_3'$. Then it is easily seen that $f_3'$ is a semi-valuation and thus the function defined by $f_3(x) = f_3'(x) - m$ is a valuation of ${\mathcal{N}}$, which we denote by $ f_1 \ast f_2$. If $f_1=f_2$, then we have $f_3=f_1=f_2$, regardless of the value of $\epsilon$. For neighboring valuations $f_1, f_2$ of ${\mathcal{N}}$ and $f_3 = f_1 \ast f_2$ we have (i) $f_2 \ast f_1 = f_1 \ast f_2 = f_3$; (ii) $f_1$ and $f_3$ are neighboring valuations and $f_1 \ast f_3 = f_2$; (iii) $f_2$ and $f_3$ are neighboring valuations and $f_2 \ast f_3 = f_1$. For more on the basic theory of valuations we refer to [@ab-bdb:1 Section 2] and [@bdb:Gn]. The following result establishes the main connection between valuations and isometric embeddings of near polygons.
\[lem:linetypes\] \[lem:embeddings\] Let ${\mathcal{N}}= ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ be a near polygon which is an isometrically embedded full subgeometry of a near polygon ${\mathcal{N}}' = ({\mathcal{P}}', {\mathcal{L}}', {\mathrm{I}}')$. Then the following holds:
1. For every point $x$ in ${\mathcal{P}}'$ the function $f_x: {\mathcal{P}}\rightarrow \mathbb{N}$ defined by $f_x(y) := d(x,y) - d(x, {\mathcal{P}})$ is a valuation of ${\mathcal{N}}$.
2. For every pair of distinct collinear points $x_1$ and $x_2$ in ${\mathcal{N}}'$, the valuations $f_{x_1}$ and $f_{x_2}$ are neighboring.
3. Say every line of ${\mathcal{N}}'$ is incident with three points and let $\{ x_1,x_2,x_3 \}$ be a line of ${\mathcal{N}}'$. Then $f_{x_1} \ast f_{x_2} = f_{x_3}$. In particular, if two of $f_{x_1},f_{x_2},f_{x_3}$ coincide then they are all equal.
See [@ab-bdb:1 Lemma 2.2].
\[new\] Let ${\mathcal{N}}= ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$ be a near polygon which is an isometrically embedded full subgeometry of a near $2d$-gon ${\mathcal{N}}' = ({\mathcal{P}}', {\mathcal{L}}', {\mathrm{I}}')$, and for every point $x$ of ${\mathcal{N}}'$, let $f_x$ be the valuation of ${\mathcal{N}}$ as defined in Lemma \[lem:linetypes\]. Then:
1. If $x$ is a point of ${\mathcal{N}}'$ such that ${\mathrm{d}}(x,{\mathcal{P}})=i$, then $M_{f_x} \leq d - i$.
2. If $x$ is a point at distance $1$ from ${\mathcal{N}}$ such that $|{\mathcal{O}}_{f_x}| = 1$, then there is a unique point $\pi_{{\mathcal{N}}}(x)$ in ${\mathcal{N}}$ collinear with $x$.
This immediately follows from the definition of the map $f_x$.
Let ${\mathcal{N}}$ be a near polygon that has three points on each line and let $V$ be the set of valuations of ${\mathcal{N}}$. The *valuation geometry* of ${\mathcal{N}}$ is the partial linear space ${\mathcal{V}}$ defined by taking the set $V$ as points and the triples $\{f_1, f_2, f_3\}$ of pairwise distinct and neighboring valuations that satisfy $f_1 \ast f_2 = f_3$ as lines. We observe that $\mathrm{Aut}({\mathcal{N}})$ acts on the valuation geometry ${\mathcal{V}}$ by the map $(f, \theta) \in V \times \mathrm{Aut}({\mathcal{N}}) \mapsto f \circ \theta^{-1}$ (and thus $f(x) = (f \circ \theta^{-1}) (\theta(x))$, $\forall x \in {\mathcal{P}} ~ \forall f \in V ~ \forall \theta \in \mathrm{Aut}({\mathcal{N}})$). When computing valuations we will only record the information about different orbits under this action by giving each orbit a different label (see eg. Table \[tab1:HD2\]) and noting the essential properties of the valuations in that orbit. Similarly, the lines of ${\mathcal{V}}$ will be given a type, which is just a sorted tuple of the type of points on that line, and we record the information about the number of ${\mathcal{V}}$-lines of a given type incident to a fixed valuation of a given type in a separate table (see eg. Table \[tab2:HD2\]). Say ${\mathcal{N}}$ is a full isometrically embedded subgeometry of another near polygon ${\mathcal{N}}'$. Then by Lemma \[lem:embeddings\] the points and lines of ${\mathcal{N}}'$ induce points and lines of the valuation geometry ${\mathcal{V}}$ of ${\mathcal{N}}$. We define the type of a point or line in ${\mathcal{N}}'$ to be the type of the corresponding point or line of ${\mathcal{V}}$. Note that the points/lines of ${\mathcal{N}}'$ of the same type are not necessarily isomorphic under the action of $\mathrm{Aut}({\mathcal{N}}')$. Lines of the valuation geometry ${\mathcal{V}}$ of a fixed near polygon ${\mathcal{N}}$ will be referred to as ${\mathcal{V}}$-lines and the valuation of ${\mathcal{N}}$ induced by a point $x$ of ${\mathcal{N}}'$ (see Lemma \[lem:embeddings\]) will be denoted by $f_x$.
Near hexagons of order $(2,2)$ {#sec:H2}
==============================
In this section we classify all near hexagons ${\mathcal{N}}$ of order $(2, 2)$. We know that if every pair of points in ${\mathcal{N}}$ at distance $2$ from each other have a unique common neighbor, then ${\mathcal{N}}$ is a generalized hexagon. In that case we can use the result of Cohen and Tits [@Co-Ti] to say that ${\mathcal{N}}$ is either isomorphic to the split Cayley hexagon $\mathrm{H}(2)$ or its dual $\mathrm{H}(2)^D$. Similarly, if every pair of points in ${\mathcal{N}}$ at distance $2$ have more than one common neighbor then we can use [@BCHW Theorem 1.1] to conclude that ${\mathcal{N}}$ must be isomorphic to $\mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$ (number (xi) in their classification). We will prove that these are the only possible cases. First we need a basic result on near polygons with an order.
\[lem:count\] Let ${\mathcal{N}} = ({\mathcal{P}}, {\mathcal{L}}, I)$ be a finite near $2d$-gon, $d \geq 1$, of order $(s,t)$ and let $x$ be a point of ${\mathcal{N}}$. Then $$\sum_{y \in {\mathcal{P}}} \left( \frac{-1}{s}\right)^{\mathrm{d}(x,y)} = 0.$$
By (NP2), for every line $L$ the sum $\sum_{y {\mathrm{I}}L} (\frac{-1}{s})^{{\mathrm{d}}(x,y)}$ is $0$. Therefore, we have $$0 = \sum_{L \in {\mathcal{L}}} \sum_{y {\mathrm{I}}L} \left(\frac{-1}{s}\right)^{{\mathrm{d}}(x,y)} = \sum_{y \in {\mathcal{P}}} \sum_{L {\mathrm{I}}y} \left(\frac{-1}{s}\right)^{{\mathrm{d}}(x,y)} = (t+1) \sum_{y \in {\mathcal{P}}} \left(\frac{-1}{s}\right)^{{\mathrm{d}}(x,y)} .$$
\[lem:quad\_order\] Let ${\mathcal{N}}$ be a finite near hexagon of order $(s, t)$ and $Q$ a quad of ${\mathcal{N}}$ that has order $(s, t')$. Then $t' < t$.
We know that $t' \leq t$. For the sake of contradiction, assume that $t' = t$. Let $x$ be a point of $Q$. Since all lines of ${\mathcal{N}}$ through $x$ are already contained in $Q$, $x$ cannot be collinear with any point that is not contained in $Q$. But then, there cannot be any points of ${\mathcal{N}}$ that lie outside $Q$, as the collinearity graph of ${\mathcal{N}}$ is connected. Thus ${\mathcal{N}} = Q$, which is a contradiction.
Let ${\mathcal{N}}$ be a near hexagon of order $(2, 2)$. Then the number of common neighbors of a pair of points at distance $2$ from each other is a constant $c \in \{1, 2\}$.
Let $v$ denote the total number of points of ${\mathcal{N}} = ({\mathcal{P}}, {\mathcal{L}}, {\mathrm{I}})$. For a fixed point $x$ let $n_i(x)$ denote the number of points at distance $i \in \{ 0,1,2,3 \}$ from $x$. For all $x \in {\mathcal{P}}$, we have $n_0(x) = 1$, $n_1(x) = 6$, and thus $$\label{eq:1}
n_2(x) + n_3(x) = v - 7.$$\
By Lemma \[lem:count\] we have $$\label{eq:2}
n_0(x) - \frac{n_1(x)}{2} + \frac{n_2(x)}{4} - \frac{n_3(x)}{8} = 0.$$\
Solving equations (\[eq:1\]) and (\[eq:2\]) we get that $n_2(x) = (v + 9)/3$ and $n_3(x) = (2v - 30)/3$ for all $x \in {\mathcal{P}}$. Therefore these numbers only depend on $v$, and we can define constants $$n_0 = 1, n_1 = 6, n_2 = (v + 9)/3, n_3 = (2v - 30)/3.$$\
By Lemma \[lem:quad\_order\] all quads of ${\mathcal{N}}$ are grid-quads ($\cong \mathbb{L}_3 \times \mathbb{L}_3$). For a point $x$, let $N(x)$ be the number of grid-quads that contain $x$. Then the number of points at distance $2$ from $x$ that are contained in a grid along with $x$ is equal to $4N(x)$ since there is a unique quad through a pair of points at distance $2$ which have more than one common neighbor. Double counting edges between $\Gamma_1(x)$ and $\Gamma_2(x)$ we get that $2 \cdot 4N(x) + 1 \cdot (n_2 - 4N(x)) = n_1 \cdot 4$, and hence $N(x) = (63 - v)/12$. So, the total number of grid-quads through a point is a constant given by $N := (63 - v)/12$.
It is known that there cannot be more than one quad through a pair of intersecting lines (see eg. [@bdb-book Thm. 1.4]). Since the number of lines through each point is $3$, we must have $N \in \{0, 1, 2, 3\}$. Since $v = 63 - 12N$, using double counting we get that the total number of grid-quads in ${\mathcal{N}}$ is $$N(63 - 12N)/9.$$ This number is not an integer if $N \in \{1, 2\}$. Therefore, $N$ must be $0$ or $3$. If $N$ is $0$, then there are no quads, and hence every two points at distance $2$ from each other have a unique common neighbor. Say $N$ is equal to $3$ and let $x, y$ be a pair of points at distance $2$ from each other. Every line through $x$ is contained in precisely two of the three grid-quads through $x$. Thus for every neighbor $z$ of $x$, the two lines through $z$ that contain a point of $\Gamma_2(x)$ are contained in grids through $x$. This implies that there is a grid through $x$ and $y$, i.e., the number of common neighbors between $x$ and $y$ is $2$.
Every near hexagon of order $(2,2)$ is isomorphic to one of the following: $\mathrm{H}(2)$, $\mathrm{H}(2)^D$, $\mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$.
It is known that the generalized hexagon $\mathrm{H}(2)^D$ has subgeometries isomorphic to $\mathrm{H}(2,1)$, while $\mathrm{H}(2)$ does not have such subgeometries. We can thus finish the proof of Theorem \[thm:H2\] by showing that also $\mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$ does not contain any subgeometry isomorphic to $\mathrm{H}(2, 1)$. So, let ${\mathcal{H}}_1 \cong \mathrm{H}(2,1)$ be a subgeometry of ${\mathcal{H}}_2 \cong \mathbb{L}_3 \times \mathbb{L}_3 \times \mathbb{L}_3$. We know that ${\mathcal{H}}_1$ has $21$ points and ${\mathcal{H}}_2$ has $27$ points. Each point of ${\mathcal{H}}_1$ is collinear with exactly two points of ${\mathcal{H}}_2 \setminus {\mathcal{H}}_1$. Since $|{\mathcal{H}}_2 \setminus {\mathcal{H}}_1| = 6$, there must be a point in ${\mathcal{H}}_2 \setminus {\mathcal{H}}_1$ collinear with at least $(21 \times 2)/6 = 7$ points of ${\mathcal{H}}_1$. This contradicts the fact that ${\mathcal{H}}_2$ has order $(2,2)$.
Characterization of the Hall-Janko near octagon {#sec:HJ}
===============================================
For this section let ${\mathcal{N}}$ be a near octagon of order $(2,4)$ with a generalized hexagon ${\mathcal{H}}$ isomorphic to $\mathrm{H}(2)^D$ isometrically embedded in it. The valuation geometry ${\mathcal{V}}$ of ${\mathcal{H}}$ is described in Tables \[tab1:HD2\] and \[tab2:HD2\]. In Table \[tab1:HD2\], the column Value Distribution denotes the distribution of points of ${\mathcal{H}}$ as per the value they have. In Table \[tab2:HD2\], the entries denote the number of lines of a given type through a point of a given type in ${\mathcal{V}}$. From Table \[tab2:HD2\] it can be seen that the set of valuations of type $A$ and $B$ form a subspace of ${\mathcal{V}}$, and hence we can define a full subgeometry ${\mathcal{V}}_{A,B}$ of ${\mathcal{V}}$ induced by these valuations. In this section we will show that ${\mathcal{N}}$ is isomorphic to ${\mathcal{V}}_{A, B}$. Since ${\mathsf{HJ}}$ is a near octagon of order $(2,4)$ with $\mathrm{H}(2)^D$ isometrically embedded in it, this will prove Theorem \[thm:HJ\].
Type $\#$ $M_f$ $|{\mathcal{O}}_f|$ $|H_f|$ Value distribution
------ -------- ------- --------------------- --------- --------------------
$A$ $63$ $3$ $1$ $31$ $[1, 6, 24, 32]$
$B$ $252$ $3$ $1$ $47$ $[1, 14, 32, 16]$
$C$ $252$ $2$ $1$ $23$ $[ 1, 22, 40, 0 ]$
$D$ $1008$ $2$ $5$ $31$ $[5, 26, 32, 0]$
: The valuations of $\mathrm{H}(2)^D$[]{data-label="tab1:HD2"}
Type $A$ $B$ $C$ $D$
------- ------ ----- ------ ------
$AAA$ $3$ – – –
$ABB$ $2$ $1$ – –
$ACC$ $2$ – 1 –
$ADD$ $24$ – – $3$
$BBB$ – $4$ – –
$BCC$ – $1$ $2$ –
$BDD$ – $4$ – $2$
$CCC$ – – 8 –
$CCD$ – – $40$ $5$
$CDD$ – – $4$ $2$
$DDD$ – – – $10$
: The lines of the valuation geometry $\mathcal{V}$ of $\mathrm{H}(2)^D$[]{data-label="tab2:HD2"}
\[lemHJ:basic\]
1. Each point of ${\mathcal{N}}$ is at distance at most $2$ from ${\mathcal{H}}$.
2. Points at distance $1$ from ${\mathcal{H}}$ must be of type $A$, $B$ or $C$.
3. Points at distance $2$ from ${\mathcal{H}}$ must be of type $C$ or $D$.
Since the maximum value of a valuation of ${\mathcal{H}}$ is at least $2$ (see Table \[tab1:HD2\]) and the diameter of ${\mathcal{N}}$ is $4$, by Lemma \[new\](1) the distance of any point of ${\mathcal{N}}$ to ${\mathcal{H}}$ is at most $2$. This proves ($a$). Again by Lemma \[new\](1), if a point $x$ of ${\mathcal{N}}$ lies at distance 2 from ${\mathcal{H}}$, then $f_x$ has maximum value at most $2$, implying that $x$ can only be of type $C$ or $D$.
Now, let $x$ be a point of type $D$ at distance $1$ from ${\mathcal{H}}$. The five points with $f_x$-value $0$ in ${\mathcal{H}}$ must be collinear with $x$ and necessarily be of type $A$ (all points in ${\mathcal{H}}$ are of type $A$). By Lemma \[lem:linetypes\] this gives rise to five distinct ${\mathcal{V}}$-lines of type $ADD$ through a valuation of type $D$ in the valuation geometry ${\mathcal{V}}$ of ${\mathcal{H}}$, which contradicts the corresponding entry in Table \[tab2:HD2\].
\[lemHJ:dist1\] Each point of ${\mathcal{N}}$ at distance $1$ from ${\mathcal{H}}$ must be of type $A$ or $B$.
Let $x$ be a point of type $C$ at distance $1$ from ${\mathcal{H}}$. By Lemma \[new\](2) and Table \[tab1:HD2\], there is a unique point $x'$ in ${\mathcal{H}}$ collinear with $x$. Again from Table \[tab1:HD2\] we see that there are $22$ points with $f_x$-value $1$ in ${\mathcal{H}}$, which must necessarily be at distance $2$ from $x$. Six of these points are neighbors of $x'$ in ${\mathcal{H}}$ and these are the only ones that have a common neighbor with $x$ that lies inside ${\mathcal{H}}$ (namely $x'$). The remaining $16$ points give rise to neighbors of $x$ that lie outside ${\mathcal{H}}$. Since the order of ${\mathcal{N}}$ is $(2, 4)$ and $x'$ lies in ${\mathcal{H}}$ there are only $9$ neighbors of $x$ that lie outside ${\mathcal{H}}$. Therefore, at least one such neighbor $y$ must be collinear with more than one point in ${\mathcal{H}}$. By Lemma \[lemHJ:basic\] the point $y$ must be of type $A$, $B$ or $C$ and for each of these possibilities we have $|{\mathcal{O}}_{f_y}| = 1$. This contradicts Lemma \[new\](2).
\[lemHJ:AB\] If $x$, $y$ are two points of ${\mathcal{N}}$, not contained in ${\mathcal{H}}$, of type $A$ and $B$ respectively, then $x$ and $y$ cannot be collinear.
Let $x$, $y$ be such points and suppose they are collinear. By Lemma \[lemHJ:basic\] they must be at distance $1$ from ${\mathcal{H}}$. The three valuations induced by the three points on the line $xy$ must be distinct (see Lemma \[lem:linetypes\]) and therefore, the line $xy$ gives rise to a ${\mathcal{V}}$-line of ${\mathcal{H}}\cong \mathrm{H}(2)^D$. From Table \[tab2:HD2\] it follows that the line $xy$ is of type $ABB$. Let $y'$ be the unique neighbor of $y$ in ${\mathcal{H}}$. Then by a similar reasoning the line $yy'$ is also of type $ABB$. But, in the valuation geometry there is a unique line of type $ABB$ through a valuation of type $B$. Therefore, the points $x$ and $y'$ induce the same type $A$ valuation, which shows that ${\mathcal{O}}_{f_x} = \{y'\}$ and hence, $x$ and $y'$ are collinear. This contradicts (NP2).
\[lemHJ:B\] If $x$ is a point of type $B$ in ${\mathcal{N}}$, then it has a unique neighbor in ${\mathcal{H}}$ and all the other neighbors of $x$ must induce distinct type $B$ valuations of ${\mathcal{H}}$.
Let $x$ be such a point, necessarily at distance $1$ from ${\mathcal{H}}$ by Lemma \[lemHJ:basic\]. By Lemma \[new\](2) and Table \[tab1:HD2\], it has a unique neighbor, say $x'$, in ${\mathcal{H}}$. There are $14$ points with $f_x$-value $1$ in ${\mathcal{H}}$ and $6$ of them are the neighbors of $x'$. The remaining $8$ must give rise to neighbors of $x$ lying outside ${\mathcal{H}}$. Let $y$ be such a neighbor. By Lemmas \[lemHJ:dist1\] and \[lemHJ:AB\], $y$ must be of type $B$ and then by Lemma \[new\](2) it cannot lie on the line $xx'$. Therefore, we get $8$ type $B$ neighbors of $x$ each corresponding to a distinct valuation of ${\mathcal{H}}$ (since the set ${\mathcal{O}}_f$ is distinct for each such valuation). The third point on the line $xx'$ must also be of type $B$ and induce a valuation distinct from all other type $B$ neighbors of $x$. Since the order of ${\mathcal{N}}$ is $(2,4)$, we have accounted for all neighbors of $x$.
The following is an immediate consequence of Lemma \[lemHJ:B\].
\[corHJ:B\] There are no lines in ${\mathcal{N}}$ of type $BCC$ or $BDD$.
\[lemHJ:noCD\] There is no point in ${\mathcal{N}}$ of type $C$ or $D$.
Let $x$ be such a point, necessarily at distance $2$ from ${\mathcal{H}}$ (see Lemmas \[lemHJ:basic\] and \[lemHJ:dist1\]). We treat the two cases separately.\
*Case 1*: Let $x$ be of type $D$. By Table \[tab1:HD2\], $|{\mathcal{H}}\cap \Gamma_2(x)| = 5$. Every line through $x$ that contains a point in $\Gamma_1({\mathcal{H}})$ must be of type $ADD$ by Table \[tab2:HD2\], Lemma \[lemHJ:dist1\] and Corollary \[corHJ:B\]. Since each of these lines has exactly one point which lies in $\Gamma_1({\mathcal{H}})$, and since that point (of type $A$) has a unique neighbor in ${\mathcal{H}}$, it must be the case that all five lines through $x$ are of type $ADD$. In fact, this also shows that these five lines correspond to five distinct lines in the valuation geometry. But we know from Table \[tab2:HD2\] that there are only three lines of type $ADD$ through a point of type $D$ in the valuation geometry, a contradiction.\
*Case 2*: Let $x$ be of type $C$. Since $|{\mathcal{O}}_{f_x}| = 1$, we have ${\mathcal{O}}_{f_x} = \Gamma_2(x) \cap {\mathcal{H}}= \{x'\}$ for some $x' \in {\mathcal{H}}$. Since there are no points of type $D$ (by *Case 1*) and no lines of type $BCC$ (by Corollary \[corHJ:B\]), all lines through $x$ must be of type $ACC$ or $CCC$ by Table \[tab2:HD2\]. Each of the type $A$ neighbors of $x$ induces the valuation $f_{x'}$ of ${\mathcal{H}}$. Therefore, besides the $6$ neighbors of $x'$ in ${\mathcal{H}}$, every point of ${\mathcal{H}}$ that has $f_x$-value $1$ must be at distance $2$ from a type $C$ neighbor of $x$. There are $16 = 22 - 6$ points with $f_x$-value $1$ in ${\mathcal{H}}$ that are not neighbor of $x'$. Since there are at most $9$ type $C$ neighbors of $x$, there must be a type $C$ neighbor $y$ of $x$ which is at distance $2$ from two distinct points of ${\mathcal{H}}$. This contradicts the fact that $|{\mathcal{O}}_{f_y}| = 1$ ($y$ is a type $C$ point).
By Lemmas \[new\](2), \[lemHJ:basic\] and \[lemHJ:noCD\], we have:
\[new2\] Every point $x$ of ${\mathcal{N}}$ not contained in ${\mathcal{H}}$ has type A or B, and lies at distance 1 from a unique point of ${\mathcal{H}}$ (the projection of $x$ in ${\mathcal{H}}$).
\[lemHJ:QuadIntersection\] Let $Q$ be a quad of ${\mathcal{N}}$ that intersects ${\mathcal{H}}$ nontrivially. Then $Q \cap {\mathcal{H}}$ is either a singleton or a line.
Say $Q \cap {\mathcal{H}}$ is not a singleton. Since $Q \cap {\mathcal{H}}$ is a subspace, it suffices to show that there are no two non-collinear points in $Q \cap {\mathcal{H}}$. Let $x$, $y$ be two non-collinear points in $Q \cap {\mathcal{H}}$. Since $Q$ is a non-degenerate generalized quadrangle, there are at least two common neighbors of $x$ and $y$ in $Q$. Since points at distance $2$ in ${\mathcal{H}}\cong \mathrm{H}(2)^D$ have a unique common neighbor and $Q \cap {\mathcal{H}}$ is a convex subspace, at least one of these common neighbors must lie outside ${\mathcal{H}}$. This gives rise to a point at distance $1$ from ${\mathcal{H}}$ with two neighbors ($x$ and $y$) in ${\mathcal{H}}$, which contradicts Corollary \[new2\].
\[lemHJ:W2\] If $x$ is a point of type $A$ in ${\mathcal{N}}$ which is not contained in ${\mathcal{H}}$, then there exists a unique $W(2)$-quad $Q$ containing $x$ and its projection $x'$ in ${\mathcal{H}}$. For this quad $Q$ we have:
1. $Q$ intersects ${\mathcal{H}}$ in a line $M$.
2. Let $M_1$ and $M_2$ be the two lines through $x'$ in ${\mathcal{H}}$ other than $M$. Then the two lines through $x$ that are not contained in $Q$ can be labeled $L_1$ and $L_2$ such that $L_1, M_1$ are parallel and at distance $1$ from each other, and $L_2, M_2$ are parallel and at distance $1$ from each other.
Let $x$ be a point of type $A$ outside ${\mathcal{H}}$ and let $x'$ be the unique point in ${\mathcal{H}}$ collinear with $x$. We have $f_x = f_{x'}$. Each of the four lines through $x$ which lie outside ${\mathcal{H}}$ are of type $AAA$ by Lemmas \[lemHJ:AB\] and \[lemHJ:noCD\]. From Table \[tab2:HD2\] we see that there are only three distinct lines of type $AAA$ through a valuation of type $A$ in the valuation geometry. Therefore, there exists two lines $K_1$, $K_2$ through $x$ which lie in $\Gamma_1({\mathcal{H}})$ and induce the same set of valuations on ${\mathcal{H}}$. Let $\{f_1, f_2, f_3\}$ be this set with $f_1 = f_x = f_{x'}$. Since all three lines through $x'$ which lie inside ${\mathcal{H}}$ correspond to distinct lines of type $AAA$ in the valuation geometry, at least one of them, say $M$, must induce the set $\{f_1, f_2, f_3\}$ of valuations.
Let $y \neq x'$ be a point on $M$. Then $y$ is collinear with a point on $K_1$ and a point on $K_2$ both of which induce the valuation equal to $f_y$. Therefore, $x$, $y$ are two points at distance $2$ in a near polygon with at least three common neighbors. From the existence result of quads, it follows that $x$ and $y$ lie in a unique quad $Q$. By the classification of quads of order $(2,t)$ and Lemma \[lem:quad\_order\], $Q$ must be isomorphic to $W(2)$, the unique generalized quadrangle of order $(2,2)$. From Lemma \[lemHJ:QuadIntersection\] it follows that $Q \cap {\mathcal{H}}= M$ and hence, for each point on $M$, the two lines through it going out of ${\mathcal{H}}$ are contained in $Q$. Hence none of the points on $M$ can be contained in a $W(2)$-quad other than $Q$.
Now let $L$ be a line through $x$ not contained in $Q$. $L$ necessarily induces a set of valuations other then $\{f_1, f_2, f_3\}$. There are only two other possibilities and both of them are induced by lines through $x'$ contained in ${\mathcal{H}}$, but not in $Q$. Therefore there must be a line $L'$ through $x'$ inducing the same set of valuations as $L$. The correspondence $L \mapsto L'$ between the set of lines through $x$ not contained in $Q$ and the set of lines through $x'$ in ${\mathcal{H}}$ distinct from $M$ is a bijection as otherwise there would exist another $W(2)$-quad through the line $xx'$ but we have already proved that there is a unique such quad.
There is no point in ${\mathcal{N}}$ of type $A$ outside ${\mathcal{H}}$.
Let $x$ be a point of type $A$ outside ${\mathcal{H}}$. By Lemma \[lemHJ:W2\] it lies in a unique $W(2)$-quad $Q$ which intersects ${\mathcal{H}}$ in a line $L$. By Lemma \[lemHJ:AB\] and Corollary \[new2\], all points of $Q \setminus L$ have type A. Let $x'$ be the projection of $x$ in ${\mathcal{H}}$ and $y'$ a neighbor of $x'$ in ${\mathcal{H}}$ lying on a line through $x'$ other than $L$. By Lemma \[lemHJ:W2\], there exists a unique neighbor $y$ of $y'$ outside ${\mathcal{H}}$ and collinear with $x$. Again by Lemma \[lemHJ:AB\] and Corollary \[new2\], the point $y$ has type A. So, by Lemma \[lemHJ:W2\], there exists a unique $W(2)$-quad $S$ containing $y$ and $y'$. The $W(2)$-quads $Q$ and $S$ are disjoint.
Suppose $p$ is a neighbor of $x'$ contained in $Q \setminus L$. As $p$ has type A, there exists by Lemma \[lemHJ:W2\] a unique line through $p$ disjoint from ${\mathcal{H}}$ that is parallel and at distance 1 from the line $x'y'$, implying that there is a common neighbor of $p$ and $y'$ in $S \setminus {\mathcal{H}}$. This implies that we can label the two lines of $Q$ through $x'$ distinct from $L = Q \cap {\mathcal{H}}$ by $T_1$ and $T_2$ and the two lines of $S$ through $y'$ distinct from $S \cap {\mathcal{H}}$ by $U_1$ and $U_2$ such that $T_1,U_1$ are parallel and at distance 1 from each other, and $T_2,U_2$ are parallel and at distance 1 from each other.
Now, consider a point $z$ in $S$ which is not collinear with $y'$. If $z$ is at distance $1$ from $Q$ (necessarily from a point of type A of $Q \setminus L$), then by Lemma \[lemHJ:W2\] its projection $z' \in S \cap {\mathcal{H}}$ is collinear with a point on the line $L$ in ${\mathcal{H}}$, contradicting the fact that ${\mathcal{H}}$ is a generalized hexagon. So $z$ (as well as every point of $S$ non-collinear with $y'$) must be at distance at least $2$ from $Q$. The two lines of $S$ through $y'$ distinct from $S \cap {\mathcal{H}}$ are parallel and at distance 1 from a line of $Q$. So, taking the projection of $z$ on these two lines, we see that there are two points in $Q$ at distance $2$ from $z$. It is known that the point $z$ induces a classical or an ovoidal valuation of $Q$ (see e.g. [@bdb-book Thm. 1.22]). Since there are two points in $Q$ at distance $2$ from $z$, the point $z$ must induce an ovoidal valuation of $Q$. Since there are five points in an ovoid in $Q$ ($\cong W(2)$), each of the five lines through $z$ must contain a (necessarily unique) point at distance $1$ from $Q$. Thus the projection of $z$ on ${\mathcal{H}}$, $z'$, must be collinear with a point in $Q$. Now, all the lines through $z'$ are either in ${\mathcal{H}}$ or in $S$. Since $S$ and $Q$ are disjoint and ${\mathcal{H}}$ is a generalized hexagon, we have a contradiction.
Therefore, ${\mathcal{N}}$ has the following description:
- each point of ${\mathcal{N}}$ is at distance at most $1$ from ${\mathcal{H}}$;
- each point of ${\mathcal{N}}$ that lies in ${\mathcal{H}}$ induces a valuation of type $A$, and
- each point of ${\mathcal{N}}$ that does not lie in ${\mathcal{H}}$ induces a valuation of type $B$.
Obviously, distinct points of ${\mathcal{N}}$ induce distinct type A valuations. To prove that ${\mathcal{N}}$ is isomorphic to ${\mathcal{V}}_{A, B}$, we first show that no two points in ${\mathcal{N}}$ can induce the same type $B$ valuation of ${\mathcal{H}}$, which will give us a bijection between the point sets of these geometries.
\[lemHJ:B1\] If $g_1$ and $g_2$ are two distinct valuations of type $B$ collinear to each other in the valuation geometry, then we have $|\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_1\}| = |\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_2 \}|$
Let $g_1$, $g_2$ be two such valuations of type $B$. Say a point $y$ in ${\mathcal{N}}$ induces the valuation $g_1$ of ${\mathcal{H}}$. If $g_1$ and $g_2$ lie on ${\mathcal{V}}$-line of type $ABB$, then the third point on the line joining $y$ and the unique neighbor of $y$ that lies in ${\mathcal{H}}$ induces the valuation $g_2$, giving us a bijection between $\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_1\}$ and $\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_2 \}$. So, say $g_1$ and $g_2$ lie on ${\mathcal{V}}$-line of type $BBB$. Then by Lemma \[lemHJ:B\] we know that each of the four lines through $y$ which do not intersect ${\mathcal{H}}$ must induce distinct ${\mathcal{V}}$-lines of type $BBB$. But we know from Table \[tab2:HD2\] that there are exactly four such ${\mathcal{V}}$-lines containing $g_1$, and hence $g_2$ is contained in exactly one of them. Therefore, there must be precisely one neighbor of $y$ in ${\mathcal{N}}\setminus {\mathcal{H}}$ which induces the valuation $g_2$. This gives a bijection between $\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_1\}$ and $\{x \in {\mathcal{N}}\setminus {\mathcal{H}}: f_x = g_2\}$.
\[lemHJ:B2\] \[lemHJ:connected\] The subgeometry of ${\mathcal{V}}$ defined on the type $B$ valuations by the lines of type $ABB$ and $BBB$ is connected.
This is checked by computer computation, see [@ab-bdb:code].
\[corHJ:BOnce\] For each type $B$ valuation $f$ of ${\mathcal{H}}$, there exists exactly one point $x \in {\mathcal{N}}$ with $f_x = f$.
Since each point of ${\mathcal{H}}$ is collinear with precisely four points of ${\mathcal{N}}\setminus {\mathcal{H}}$, and each point of ${\mathcal{N}}\setminus {\mathcal{H}}$ has a unique neighbor in ${\mathcal{H}}$, we have $|{\mathcal{N}}\setminus {\mathcal{H}}| = 4 \times |{\mathcal{H}}| = 252$. By Lemmas \[lemHJ:B1\] and \[lemHJ:B2\] we know that for every pair of type B valuations there exist equally many points in ${\mathcal{N}}$ which induce those valuations. But from Table \[tab1:HD2\] we can see that there are exactly $252$ valuations of type $B$. Therefore each type $B$ valuation is induced exactly once.
Now we can prove that ${\mathcal{N}}$ is isomorphic to ${\mathcal{V}}_{A,B}$ as follows. Map every point $x$ of ${\mathcal{N}}$ to the valuation of type $T \in \{A, B\}$ that it induces. Since no two points of ${\mathcal{N}}$ induce the same valuation (see Corollary \[corHJ:BOnce\]), for every line $L = \{x, y, z\}$ of ${\mathcal{N}}$ the triple $\{f_x, f_y, f_z\}$ is a ${\mathcal{V}}$-line. Map every line of ${\mathcal{N}}$ to this corresponding line of ${\mathcal{V}}_{A, B}$. Since ${\mathcal{N}}$ and ${\mathcal{V}}_{A,B}$ have the same number of points and the same order $(2,4)$ the above maps between the point and line sets of ${\mathcal{N}}$ and ${\mathcal{V}}_{A,B}$ are bijections and define an isomorphism between the two geometries. Thus, every near polygon of order $(2,4)$ that contains an isometrically embedded generalized hexagon ${\mathcal{H}}$ isomorphic to $\mathrm{H}(2)^D$ must be isomorphic to ${\mathcal{V}}_{A, B}$, which proves Theorem \[thm:HJ\].
Characterization of the $G_2(4)$ near octagon {#sec:G24}
=============================================
For this section let ${\mathcal{N}}$ be a near octagon with three points on each line containing a suboctagon ${\mathcal{H}}$ isomorphic to ${\mathsf{HJ}}$ isometrically embedded in it. The valuation geometry ${\mathcal{V}}$ of ${\mathcal{H}}\cong {\mathsf{HJ}}$ is given in Tables \[tab1:HJ\] and \[tab2:HJ\]. The main purpose of this section is to show that if ${\mathcal{N}}$ has order $(2,10)$, then ${\mathcal{N}}$ is isomorphic to the $G_2(4)$ near octagon. In [@ab-bdb:2 Appendix] it was shown that the $G_2(4)$ near octagon can be constructed by taking the valuations of type $A$, $B$ and $C$ as points and the ${\mathcal{V}}$-lines of type $AAA$, $ABB$, $ACC$, $BBC$ and $CCC$ as lines. Therefore, we will show that if ${\mathcal{N}}$ has order $(2,10)$, then it consists of points of type $A$, $B$ or $C$ and lines of type $AAA, ABB, ACC, BBC$ or $CCC$, with each type occurring exactly once. First we derive some general results that are true for any near octagon ${\mathcal{N}}$ with three points on each line that contains ${\mathcal{H}}$ as a full isometrically embedded subgeometry and later restrict ourselves to the case when ${\mathcal{N}}$ has order $(2,10)$. Lemmas \[lem:ValBC\] to \[lemG:connected\] are proved using the computer model of the valuation geometry that we have constructed, see [@ab-bdb:code].
Type \# $M_f$ $|\mathcal{O}_f|$ value distribution
------ -------- ------- ------------------- ------------------------ --
$A$ $315$ $4$ $1$ $[1,10, 80, 160, 64]$
$B$ $630$ $3$ $1$ $[1, 10, 112, 192, 0]$
$C$ $3150$ $3$ $1$ $[1, 26, 128, 160, 0]$
$D$ $1008$ $2$ $5$ $[5, 110, 200, 0, 0]$
$E$ $2016$ $2$ $25$ $[25, 130, 160, 0, 0]$
: The valuations of Hall-Janko near octagon ${\mathsf{HJ}}$[]{data-label="tab1:HJ"}
Type $A$ $B$ $C$ $D$ $E$
------- ----- ------ ----- ------ -----
$AAA$ $5$ – – – –
$ABB$ $1$ $1$ – – –
$ACC$ $5$ – $1$ – –
$BBB$ – $5$ – – –
$BBC$ – $10$ $1$ – –
$CCC$ – – $9$ – –
$CDD$ – – $4$ $25$ –
$DDD$ – – – $6$ –
$DEE$ – – – $1$ $1$
$EEE$ – – – – $6$
: The lines of the valuation geometry $\mathcal{V}$ of ${\mathsf{HJ}}$[]{data-label="tab2:HJ"}
\[lem:ValBC\] Let $f$ be a valuation of type $C$ and let $g \not= f$ and $h \not= f$ be valuations of type $B$ or $C$ lying on distinct ${\mathcal{V}}$-lines through $f$. Then $g$ and $h$ are non-collinear.
\[lem:ValuationGeometryHJ\] \[lem:ValB\] Let $f$ be a valuation of type $B$ and let $x \in {\mathcal{H}}$ be the unique point in ${\mathcal{O}}_f$. Then the map $\{f, g, h\} \mapsto {\mathcal{O}}_f \cup {\mathcal{O}}_g \cup {\mathcal{O}}_h$ is a bijection between the set of five ${\mathcal{V}}$-lines of type $BBB$ through $f$ and the set of five lines of ${\mathcal{H}}$ through $x$.
\[lem:ValC\] Let $f$ be a valuation of type $C$. Then there is a unique ${\mathcal{V}}$-line $\{f, g, h\}$ of type $CCC$ through $f$ such that ${\mathcal{O}}_f \cup {\mathcal{O}}_g \cup {\mathcal{O}}_h$ is a line of ${\mathsf{HJ}}$. For every other ${\mathcal{V}}$-line $\{ f,g',h' \}$ of type $CCC$ through $f$, the set ${\mathcal{O}}_f \cup {\mathcal{O}}_{g'} \cup {\mathcal{O}}_{h'}$ is a set of three pairwise non-collinear points.
A ${\mathcal{V}}$-line $\{ f,g,h \}$ of type CCC will be called [*special*]{} if ${\mathcal{O}}_f \cup {\mathcal{O}}_g \cup {\mathcal{O}}_h$ is a line of ${\mathsf{HJ}}$. If that is not the case, then $\{ f,g,h \}$ will be called an [*ordinary*]{} ${\mathcal{V}}$-line. This concept of *special* and *ordinary* is then extended to the lines of ${\mathcal{N}}$ that induce ${\mathcal{V}}$-lines of type $CCC$.
\[lemG:connected\] The subgeometry of ${\mathcal{V}}$ defined on the type $C$ valuations by the lines of type $ACC$ and the ordinary lines of type $CCC$ is connected.
***Remark***: Lemmas \[lem:ValBC\] and \[lem:ValC\] could alternatively be verified by a geometric reasoning inside the $G_2(4)$ near octagon, keeping in mind its above mentioned construction using the valuations of type $A$, $B$ and $C$ of ${\mathsf{HJ}}$.
\[lemG:basic\] Every point of ${\mathcal{N}}$ is at distance at most $2$ from ${\mathcal{H}}$. Points of ${\mathcal{H}}$ are of type $A$, points at distance $1$ from ${\mathcal{H}}$ are of type $B$ or $C$ and those at distance $2$ are of type $D$ or $E$.
Since ${\mathcal{H}}$ is isometrically embedded in ${\mathcal{N}}$, all points of ${\mathcal{H}}$ induce type $A$ valuations. From Lemma \[new\](1) and the column $M_f$ of Table \[tab1:HJ\] we see that the points in $\Gamma_1({\mathcal{H}})$ cannot be of type $A$, but the points in $\Gamma_2({\mathcal{H}})$ must be of type $D$ or $E$. If $x \in \Gamma_1({\mathcal{H}})$, then there exists a line through $x$ that intersects ${\mathcal{H}}$, which must necessarily be of type $ABB$ or $ACC$ by Table \[tab2:HJ\], implying that $x$ has type $B$ or $C$.
\[corG:dist1\] Every point of ${\mathcal{N}}$ at distance $1$ from ${\mathcal{H}}$ is collinear with a unique point of ${\mathcal{H}}$.
Such points are of type $B$ or $C$ and valuations of type $B$ and $C$ have exactly one point of value $0$ (see column $|{\mathcal{O}}_f|$ in Table \[tab1:HJ\]).
\[lemG:noE\] There are no points of type $E$ in ${\mathcal{N}}$.
Let $x$ be a type $E$ point of ${\mathcal{N}}$. By Lemma \[lemG:basic\], $x$ must be at distance $2$ from ${\mathcal{H}}$. Let $y$ be a neighbor of $x$ which lies at distance $1$ from ${\mathcal{H}}$. Then $x$ has type $B$ or $C$ by Lemma \[lemG:basic\]. Since the valuations $f_x$ and $f_y$ are not equal, the line $xy$ gives rise to a ${\mathcal{V}}$-line in the valuation geometry of ${\mathsf{HJ}}$ (see Lemma \[lem:linetypes\]). But, by Table \[tab2:HJ\] there are no ${\mathcal{V}}$-lines with both type $E$ and type $T$ points on it, for $T \in \{B, C\}$.
Let $x$ be a point of ${\mathcal{N}}$ at distance $1$ from ${\mathcal{H}}$ which by Lemma \[lemG:basic\] is of type $B$ or $C$. We will call the unique point of ${\mathcal{H}}$ collinear with $x$ (see Corollary \[corG:dist1\]) the *projection* of $x$, and denote it by $\pi(x)$. From now onward we implicitly use the fact that points at distance $1$ from ${\mathcal{H}}$ are of type $B$ or $C$. For a line $L = \{x, y, z\}$ contained in $\Gamma_1({\mathcal{H}})$ we define the projection $\pi(L)$ of $L$ to be the set $\{\pi(x), \pi(y), \pi(x)\}$ of points of ${\mathcal{H}}$. Since ${\mathcal{N}}$ is a near polygon, $\pi(L)$ and $L$ have the same size for every line $L$ in $\Gamma_1({\mathcal{H}})$. But, this projection may or may not be a line of ${\mathcal{H}}$.
\[lemG:B\] Let $x$ be a type $B$ point of ${\mathcal{N}}$ and let $y$ be a point on a line through $x$ which does not intersect ${\mathcal{H}}$. Then $y$ is at distance $1$ from ${\mathcal{H}}$ and the projections $\pi(x)$ and $\pi(y)$ are collinear.
The point $y$ must be of type $B$ or $C$ since there are no ${\mathcal{V}}$-lines containing both type $B$ and type $T$ points for $T \in \{D, E\}$ (see Table \[tab2:HJ\]) and hence at distance $1$ from ${\mathcal{H}}$ by Lemma \[lemG:basic\]. The projections $\pi(x)$ and $\pi(y)$ have $f_x$-values $0$ and $1$, respectively. Since $f_x$ is of type $B$, there are exactly ten points of ${\mathcal{H}}$ that have $f_x$-value $1$ (see Table \[tab1:HJ\]). Clearly, every point in ${\mathcal{H}}$ at distance 1 from $\pi(x)$ has $f_x$-value $1$. Since ${\mathcal{H}}$ has order $(2,4)$, there are precisely ten such points and hence $\pi(y)$ must be one of them.
\[corG:B\] If $x$ is a point of type $B$ in ${\mathcal{N}}$, then every line through $x$ that does not intersect ${\mathcal{H}}$ is parallel to and at distance $1$ from a unique line of ${\mathcal{H}}$.
\[lemG:BBC\] Every type $B$ point $x$ of ${\mathcal{N}}$ is incident with a line of type $BBC$.
Let $x$ be a point of type $B$. Then every point of ${\mathcal{H}}$ at distance 1 from $\pi(x)$ has $f_x$-value 1. Since ${\mathcal{O}}_{f_x} = \{ \pi(x) \}$ every point of ${\mathcal{H}}$ at distance 2 from $\pi(x)$ should have $f_x$-value 2, and since ${\mathcal{H}}$ is a regular near octagon with parameters $(2, 4; 0, 3)$, there are $80$ such points. By Table \[tab1:HJ\] there are $112$ points of ${\mathcal{H}}$ with $f_x$-value $2$. Let $y$ be one of the other $112-80=32$ points with $f_x$-value 2 at distance at least 3 from $\pi(x)$. Since $f_x(y) = 2$, we have ${\mathrm{d}}(x, y) = 3$. Let $x, u, v, y$ be a path of length $3$ connecting $x$ and $y$. By Lemmas \[lemG:basic\] and \[lemG:B\], the point $u$ has type $A$, $B$ or $C$. We will show that $u$ is of type $C$, hence proving that the line $xu$ is of type $BBC$.
If $u$ is of type $A$, then $u = \pi(x)$, which would be in contradiction with ${\mathrm{d}}(\pi(x), y) > 2$. Suppose $u$ is of type $B$, and hence at distance $1$ from ${\mathcal{H}}$. From Corollary \[corG:B\] we see that $\pi(u)$ and $\pi(x)$ are collinear (or equal). We cannot have $v = \pi(u)$ as that would imply that ${\mathrm{d}}(\pi(x), y) \leq 2$. Therefore, $v$ lies outside ${\mathcal{H}}$ and $y$ must be equal to $\pi(v)$. Again by Corollary \[corG:B\] $y=\pi(v)$ and $\pi(u)$ must be collinear (or equal), which contradicts the fact that ${\mathrm{d}}(\pi(x), y) > 2$. So, $u$ is of type $C$.
\[corG:Cexists\] There exist type $C$ points in ${\mathcal{N}}$.
As there exist points at distance $1$ from ${\mathcal{H}}$, there exist points of type $B$ or $C$. The existence of type $B$ points implies the existence of type $C$ points by Lemma \[lemG:BBC\].
\[lemG:distinct\_lines\_CCC\] Let $x$ be a point of ${\mathcal{N}}$ of type $C$ and let $L_1, L_2$ be two distinct lines of type $CCC$ through $x$. Then the ${\mathcal{V}}$-lines corresponding to $L_1$ and $L_2$ must be distinct.
Let $L_1 = \{x, y, z\}$ and $L_2 = \{x, y', z'\}$. Assume that they correspond to the same ${\mathcal{V}}$-line so that $f_y = f_{y'}$ and $f_z = f_{z'}$. Let $u := \pi(z) = \pi(z')$. Since $xzuz'$ is a quadrangle, the point $y$ must be collinear with the third point on the line $uz'$, call it $v$. Therefore, the valuations $f_v$ and $f_y = f_{y'}$ are collinear in ${\mathcal{V}}$. The collinearity of $f_v$ and $f_{y'}$ in ${\mathcal{V}}$ contradicts Lemma \[lem:ValBC\] by taking $f = f_{z'}$, $g = f_{y'}$ and $h = f_v$.
On the valuations of type $C$ we can define a subgeometry of ${\mathcal{V}}$ induced by the lines of type $ACC$ and the ordinary lines of type $CCC$. Let this subgeometry be denoted by ${\mathcal{V}}_C$ and its collinearity graph by $\Gamma_1$. Similarly, we can define a subgeometry ${\mathcal{N}}_C$ of ${\mathcal{N}}$ by taking the points of type $C$ and the lines that correspond to lines of ${\mathcal{V}}_C$. Let $\Gamma_2$ be the collinearity graph of ${\mathcal{N}}_C$. Since type $C$ points exist in ${\mathcal{N}}$, the graph $\Gamma_2$ is nonempty.
\[lemG:cover\] The graph $\Gamma_2$ is a cover of the graph $\Gamma_1$ by the map $x \mapsto f_x$.
Let $x$ be a point of type $C$ in ${\mathcal{N}}$. There are $16$ points of $f_x$-value $1$ that are not collinear with $\pi(x)$ (see Table \[tab1:HJ\]). Denote this set of $16$ points by $U$. If $u \in U$, then ${\mathrm{d}}(x,u) = 2$. Denote by $v$ a common neighbor of $x$ and $u$. Since $v \not= \pi(x)$, $v \not\in {\mathcal{H}}$ and $u = \pi(v)$. Since $u \not= \pi(x)$, the point $v$ cannot be contained on the line $x\pi(x)$, and so $xv$ is a line disjoint from ${\mathcal{H}}$. Since $\pi(x)$ and $\pi(v)=u$ are not collinear, Lemma \[lemG:B\] implies that $v$ has type $C$ and hence that $xv$ is a (necessarily ordinary) line of type $CCC$.
By Lemmas \[lem:ValC\], \[lemG:distinct\_lines\_CCC\] and Table \[tab2:HJ\] there are at most $8$ ordinary lines of type $CCC$ through $x$, each of which determines two points of the set $U$. Since $|U| = 16$, it follows that there are precisely $8$ ordinary lines of type $CCC$ through $x$ and they correspond bijectively to the $8$ ordinary ${\mathcal{V}}$-lines of type $CCC$ through $f_x$. This proves that the map $x \mapsto f_x$ is a local isomorphism between $\Gamma_2$ and $\Gamma_1$. The fact that this map is surjective now follows from the connectedness of $\Gamma_1$, see Lemma \[lemG:connected\].
\[corG:cover\] If $\Gamma_2$ is an $i$-cover of $\Gamma_1$ for some $i \geq 1$, then each valuation of type $C$ is induced by precisely $i$ type $C$ points of ${\mathcal{N}}$. As a consequence, through each point of ${\mathcal{H}}$, there are precisely $5i$ lines of type $ACC$.
By Table \[tab2:HJ\] there are precisely $5$ ${\mathcal{V}}$-lines of type $ACC$ through a given valuation of type $A$. Since each type $C$ valuation occurs exactly $i$ times, we have $5i$ type $ACC$ lines in ${\mathcal{N}}$ through a given point of ${\mathcal{H}}$.
***Remark:*** All results in this section so far are valid for a general near octagon of order $(2,t)$ that contains an isometrically embedded sub near octagon isomorphic to ${\mathsf{HJ}}$. In the following lemma, we need the fact that ${\mathcal{N}}$ has order $(2,10)$.
\[lemG:BC\_exist\_and\_unique\] If ${\mathcal{N}}$ is of order $(2,10)$, then each valuation of type $T \in \{B, C\}$ is induced exactly once by a point of ${\mathcal{N}}$.
Let ${\mathcal{N}}$ be of order $(2,10)$ and let $x$ be an arbitrary point of ${\mathcal{H}}$. Then there are exactly $11 - 5 = 6$ lines through $x$ that are not contained in ${\mathcal{H}}$, each of which has type $ACC$ or $ABB$. Since type $C$ points exist by Corollary \[corG:Cexists\], it follows from Corollary \[corG:cover\] that there are precisely $5$ lines of type $ACC$ through $x$ in ${\mathcal{N}}$, and hence the graph $\Gamma_2$ is a $1$-cover of $\Gamma_1$. Now the $6$-th line through $x$ which is not contained in ${\mathcal{H}}$ must be of type $ABB$. Therefore, through every point of ${\mathcal{H}}$ there are $5$ lines of type $ACC$ and a unique line of type $ABB$. This shows that for every valuation $f$ of type $B$, we can find the unique point of ${\mathcal{N}}$ that induces $f$ by first getting the point $y$ of ${\mathcal{H}}$ that induces the type $A$ valuation on the unique ${\mathcal{V}}$-line of type $ABB$ through $f$ (see Table \[tab2:HJ\]), and then picking the point on the unique line of type $ABB$ through $x$ in ${\mathcal{N}}$ that induces the valuation $f$.
For the rest of this section assume that ${\mathcal{N}}$ has order $(2,10)$. From Lemma \[lemG:BC\_exist\_and\_unique\] we know that both type $B$ and type $C$ points exist in ${\mathcal{N}}$ and each type $B$ or type $C$ valuation of ${\mathcal{H}}$ is induced by a unique point of ${\mathcal{N}}$. Let $x$ be a point of type $B$ in ${\mathcal{N}}$ and let $L_x$ be the unique line joining $x$ and $\pi(x)$. From Corollary \[corG:B\] it follows that every other line through $x$ gives rise to a quad in ${\mathcal{N}}$ that intersects ${\mathcal{H}}$ and contains $L_x$.
\[lemG:QuadIntersection\] Let $Q$ be a quad of ${\mathcal{N}}$ that intersects ${\mathcal{H}}$ nontrivially. Then $Q \cap {\mathcal{H}}$ is either a singleton or a line.
The proof is similar to that of Lemma \[lemHJ:QuadIntersection\].
\[lem:QuadBC\] Let $Q$ be a quad of ${\mathcal{N}}$ that is not a grid and that intersects ${\mathcal{H}}$ in a line $L$. Then there must exist points of type $B$ and points of type $C$ in $Q \setminus L$.
For the sake of contradiction assume that all points of $Q \setminus L$ are of a fixed type $T \in \{B, C\}$. Let $x$ be a point of $L$. Since $Q$ is not a grid, there exist two lines $L_1 = \{x, y, z\}$ and $L_2 = \{x, y', z'\}$ through $x$ with $y, y', z, z' \in Q \setminus L$. Let $w$ be a common neighbor of $z$ and $z'$ in $Q$ which is different from $x$. Then $w \in Q \setminus L$. From Lemma \[lemG:BC\_exist\_and\_unique\] it follows that the lines $wz$ and $wz'$ correspond to distinct ${\mathcal{V}}$-lines, which are of type $TTT$ by our assumption. Also note that $\pi(wz) = \pi(wz') = L$. This contradicts Lemma \[lem:ValB\] for $T = B$ and Lemma \[lem:ValC\] for $T = C$.
\[lemG:noQ52\] \[corG:noQ52\] There are no $Q(5,2)$-quads in ${\mathcal{N}}$ that meet ${\mathcal{H}}$ in a line.
Let $Q$ be a $Q(5,2)$-quad that meets ${\mathcal{H}}$ in a line $L$. By Lemma \[lem:QuadBC\] there is a point $x$ of type $C$ in $Q$. There is a unique line through $x$ that intersects ${\mathcal{H}}$ in $L$, and hence lies in $Q$. Every other line through $x$ which is contained in $Q$ projects to $L$. By Lemmas \[lem:ValC\] and \[lemG:distinct\_lines\_CCC\] there is at most one line of type $CCC$ through $x$ in $Q$. From Table \[tab2:HJ\] and Lemma \[lemG:BC\_exist\_and\_unique\] it follows that there is at most one line of type $BBC$ through $x$ in $Q$. Therefore, in total we have at most three lines through $x$ in $Q$ which contradicts the fact that the order of a $Q(5,2)$-quad is $(2,4)$.
\[corG:BBB\] Let $x$ be a point of type $B$ in ${\mathcal{N}}$. Then $x$ cannot be contained in two lines $L_1$, $L_2$ such that $L_1$ has type $BBB$, $L_2$ has type $BBC$ and $\pi(L_1) = \pi(L_2)$.
Let $L_1 = \{x, y, z\}$ and $L_2 = \{x, y', z'\}$ be two such lines, such that $\pi(y) = \pi(y')$ and $\pi(z) = \pi(z')$. Say $L_1$ is of type $BBC$ and $L_2$ of type $BBB$. Without loss of generality assume that $z$ is of type $C$. Then the lines $y\pi(y)$ and $y'\pi(y')$ are of type $ABB$. By Table \[tab2:HJ\] there is only one ${\mathcal{V}}$-line of type $ABB$ through a valuation of type $A$, and hence $f_y$ is equal to $f_{y'}$ or $f_{y''}$ where $y''$ is the third point (of type $B$) on the line $y'\pi(y')$. This contradicts Lemma \[lemG:BC\_exist\_and\_unique\].
\[lemG:typeB\] Let $x$ be a point of type $B$ in ${\mathcal{N}}$. Then
1. $x$ is incident with a unique line of type $ABB$ and ten lines of type $BBC$;
2. these ten type $BBC$ lines through $x$ correspond bijectively to the ten $BBC$ lines of the valuation geometry ${\mathcal{V}}$ through $f_x$, and they are partitioned into pairs by five $W(2)$-quads passing through the line of type $ABB$ through $x$.
There is a unique line through $x$ that intersects ${\mathcal{H}}$, namely the line joining $x$ and $\pi(x)$. Every other line through $x$ is of type $BBB$ or $BBC$ which is entirely contained in $\Gamma_1({\mathcal{H}})$ and is parallel to a line through $\pi(x)$ in ${\mathcal{H}}$ (see Corollary \[corG:B\]). Let $S$ denote the set of these other lines through $x$. By Lemma \[lemG:BC\_exist\_and\_unique\], distinct lines in $S$ correspond to distinct ${\mathcal{V}}$-lines. Let there be $i$ lines of type $BBB$ in $S$, with $i \leq 5$ by Table \[tab2:HJ\]. Since we cannot have two lines of type $BBC$ and $BBB$ in $S$ projecting to the same line of ${\mathcal{H}}$ by Lemma \[corG:BBB\] and since there are no $Q(5,2)$-quads by Lemma \[lemG:noQ52\], there are at most $2(5-i)$ lines of type $BBC$ in $S$, and hence in total at most $2(5 - i) + i + 1 = 11 - i$ lines through $x$. Therefore, we have $i = 0$ and each of the $5$ lines of ${\mathcal{H}}$ through $\pi(x)$ is parallel to exactly $2$ lines of $S$. This gives rise to $5$ $W(2)$-quads through the line $x\pi(x)$, that partition $S$ into pairs.
We are now ready to prove Theorem \[thm:G24\]. From Lemma \[lemG:typeB\] it follows that there are no lines of type $BBB$ in ${\mathcal{N}}$. Since each of the type $A$, $B$ and $C$ valuations is induced by a unique point of ${\mathcal{N}}$ and each ${\mathcal{V}}$-line of type $AAA$, $ABB$ and $ACC$ is induced by a unique line of ${\mathcal{N}}$, it suffices to show that also every ${\mathcal{V}}$-line of type $BBC$ and $CCC$ is induced by a unique line of ${\mathcal{N}}$, and that type $D$ points do not exist in ${\mathcal{N}}$ (we have already proved in Lemma \[lemG:noE\] that type $E$ points do not exist).
Let $\{ f, g, h \}$ be a ${\mathcal{V}}$-line of type $BBC$ where $f$ is of type $B$. Let $x$ be the unique point in ${\mathcal{N}}$ with $f_x = f$. By Lemma \[lemG:typeB\], there exists a line $L = \{x, y, z\}$ such that $f_y = g$ and $f_z = h$. This shows that each ${\mathcal{V}}$-line of type $BBC$ is induced by a necessarily unique line of ${\mathcal{N}}$.
Now, let $x$ be a point of type $C$. Since $\Gamma_2$ is a $1$-cover of $\Gamma_1$, there exist eight ordinary lines of type $CCC$ through $x$ that bijectively correspond to the eight ordinary ${\mathcal{V}}$-lines of type $CCC$ through $f_x$. By Table \[tab2:HJ\], there exists a unique ${\mathcal{V}}$-line of type $BBC$ through $f_x$, implying that in ${\mathcal{N}}$ there is a unique line $L$ of type $BBC$ through $x$. By Lemma \[lemG:typeB\] $L$ lies in a $W(2)$-quad $Q$, which must also contain the unique line of type $ACC$ through $x$. The third line in $Q$ through $x$ must be a special line of type $CCC$ as there is a unique type $BBC$ line through $x$ and none of the ordinary type $CCC$ lines through $x$ projects to a line of ${\mathcal{H}}$. Therefore, the unique special ${\mathcal{V}}$-line of type CCC through $f_x$ is induced by a line of ${\mathcal{N}}$. Since we have accounted for all $11$ lines through a point of type $C$, there cannot be any lines of type $CDD$, and hence there cannot be any points of type $D$ in ${\mathcal{N}}$. This completes the proof as we have shown that ${\mathcal{N}}$ is isomorphic to the $G_2(4)$ near octagon.
[99]{}
M. K. Bardoe. The universal embedding for the involution geometry of the Suzuki sporadic simple group. [*J. Algebra*]{} 186 (1996), 447–460.
M. K. Bardoe. The universal embedding for the involution geometry of $Co_1$. [*J. Algebra*]{} 217 (1999), 555–572.
A. Bishnoi and B. De Bruyn. On semi-finite hexagons of order $(2,t)$ containing a subhexagon. [*Ann. Comb.*]{} (2016), to appear. Preprint available at <http://arxiv.org/abs/1503.05865>.
A. Bishnoi and B. De Bruyn. A new near octagon and the Suzuki tower. [*Electron. J. Combin.*]{} 23, \#P2.35 (2016), 24 pp.
A. Bishnoi and B. De Bruyn. GAP-code for “Characterizations of the Suzuki tower near polygons”. Online available document, <http://cage.ugent.be/geometry/preprints.php>.
A. E. Brouwer. The Cohen-Tits near octagon on 315 points. <http://www.win.tue.nl/~aeb/graphs/HJ315.html>.
A. E. Brouwer, A. M. Cohen, J. I. Hall and H. A. Wilbrink. Near polygons and Fischer spaces. *Geom. Dedicata* 49 (1994), 349–368.
A. E. Brouwer, A. M. Cohen and A. Neumaier. [*Distance-regular graphs*]{}. Springer-Verlag, Berlin, 1989.
A. E. Brouwer, D. G. Fon-der-Flaass and S. V. Shpectorov. Locally co-Heawood graphs. [*Finite geometry and combinatorics*]{} (Deinze, 1992), pp. 59–68, London Math. Soc. Lecture Note Ser. 191, Cambridge Univ. Press, Cambridge, 1993.
A. E. Brouwer and H. A. Wilbrink. The structure of near polygons with quads. *Geom. Dedicata* 14 (1983), 145–176.
P. J. Cameron. Dual polar spaces. *Geom. Dedicata* 12 (1982), 75–85.
A. M. Cohen. Geometries originating from certain distance-regular graphs. [*Finite geometries and designs*]{} (Proc. Conf., Chelwood Gate, 1980), pp. 81–87, London Math. Soc. Lecture Note Ser. 49, Cambridge Univ. Press, Cambridge-New York, 1981.
A. M. Cohen and J. Tits. On generalized hexagons and a near octagon whose lines have three points. [*European J. Combin.*]{} 6 (1985), 13–27.
B. De Bruyn. *Near polygons*. Frontiers in Mathematics, Birkhäuser, Basel, 2006.
B. De Bruyn. The valuations of the near polygon $\mathbb{G}_n$. [*Electron. J. Combin.*]{} 16 (2009), Research Paper 137, 29 pp.
B. De Bruyn. The use of valuations for classifying point-line geometries. pp. 27–40 in “Groups of exceptional type, Coxeter groups and related geometries”, [*Springer Proc. Math. Stat.*]{} 82, Springer, 2014.
B. De Bruyn and P. Vandecasteele. Valuations of near polygons. [*Glasg. Math. J.*]{} 47 (2005), 347–361.
The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.7.5; 2014. (<http://www.gap-system.org>)
D. Leemans. A family of geometries related to the Suzuki tower. [*Comm. Algebra*]{} 33 (2005), 2201–2217.
A. Neumaier. Rectagraphs, diagrams, and Suzuki’s sporadic simple group. [*Algebraic and geometric combinatorics*]{}, pp. 305–318, North-Holland Math. Stud. 65, North-Holland, Amsterdam, 1982.
D. V. Pasechnik. Geometric characterization of graphs from the Suzuki chain. [*European J. Combin*]{} 14 (1993), 491–499.
S. E. Payne and J. A. Thas. [*Finite generalized quadrangles*]{}. European Mathematical Society, Zürich, 2009.
E. E. Shult and A. Yanushka. Near $n$-gons and line systems. [*Geom. Dedicata*]{} 9 (1980), 1–72.
L. H. Soicher. On simplicial complexes related to the Suzuki sequence graphs. [*Groups, combinatorics and geometry*]{} (Durham, 1990), pp. 240–248, London Math. Soc. Lecture Note Ser. 165, Cambridge Univ. Press, Cambridge, 1992.
L. H. Soicher. Three new distance-regular graphs. [*European J. Combin.*]{} 14 (1993), 501–505.
M. Suzuki. A finite simple group of order 448, 345, 497, 600. [*Symposium on Finite Groups,*]{} Brauer and Sag (eds.), pp. 113–119, Benjamin, New York, 1969.
R. A. Wilson. [*The finite simple groups*]{}. Graduate Texts in Mathematics 251, Springer, 2009.
S. Yoshiara. A lattice theoretical construction of a GAB of the Suzuki sporadic simple group. [*J. Algebra*]{} 112 (1988), 198–239.
[^1]: see <http://brauer.maths.qmul.ac.uk/Atlas/v3/exc/G24/> for the notation and description of this class
[^2]: see <http://brauer.maths.qmul.ac.uk/Atlas/v3/permrep/3Suzd2G1-p5346B0>
|
{
"pile_set_name": "ArXiv"
}
|
=22.5cm 0.3cm
-1.0cm -0.8cm -0.8cm
DESY 96-257\
CERN-TH/96-352\
hep-ph/9612334
[**Four-Jet Signal at LEP2 and Supersymmetry**]{}\
\
$^a$ Theory Division, CERN\
Geneva, Switzerland\
$^b$ Deutsches Elektronen-Synchrotron, DESY\
Hamburg, Germany\
Abstract
ALEPH has reported a significant excess of four-jet events in the LEP runs above the $Z^0$ resonance, which however has not been confirmed by the other LEP collaborations. We assume here that this excess corresponds to a physics signal and try to interpret it in the context of supersymmetric models with $R$-parity violation. Associated production of a left and right selectron can explain all the distinctive features of the ALEPH data: the value of the cross section, the dijet mass difference, the absence of bottom quarks in the final state, and the dijet charge content. Our proposed scenario makes definite predictions, which can be tested at future LEP runs at higher energies.
CERN-TH/96-352\
December 1996
Introduction
============
One of the most intriguing and controversial results of the LEP run above the $Z^0$ resonance has been the excess of four-jet events reported by ALEPH [@jet; @rag]. As the other three experimental collaborations working at LEP do not observe any anomaly in four-jet topologies, the resolution of the experimental controversy is a most urgent issue. All experimental collaborations are actively working on the question, which hopefully will be settled by further study and, most importantly, by the new runs at higher energies. From the theoretical side, we believe that it is important to investigate if the reported ALEPH data can be interpreted as a consistent physics signal. At least such a study can be used as a benchmark to compare the present results with future data at higher $\sqrt{s}$.
The ALEPH four-jet events have been selected from the data recorded at centre-of-mass energies between 130 and 172 GeV. An excess is observed in the distribution of the sum of the two dijet invariant masses contructed by pairing jets with the smallest dijet mass difference. This distribution shows a peak at $106.1\pm 0.8$ GeV, corresponding to 18 events observed with 3.1 expected from QCD background [@rag]. If interpreted as particle pair production, this corresponds to a cross section of $2.5\pm 0.7$ pb when only data with $\sqrt{s}$ in the range between 130 and 161 GeV are considered, and of $1.5\pm 0.8$ pb when all data with $\sqrt{s}$ in the range between 130 and 172 GeV are considered [@rag]. This cross section is too large for Higgs bosons or for electroweakly-interacting scalar particles, whose productions are suppressed by a factor $\beta^3$. Here $\beta$ is the final-state particle velocity in the centre-of-mass, which is, for the relevant kinematical configuration, about 0.6 at $\sqrt{s}=130$ GeV and 0.8 at $\sqrt{s}=172$ GeV. The inferred value of the cross section could be accommodated by production of fermions with electroweak couplings or of scalar particles with a substantial colour or multiplicity factor.
The dijet mass difference distribution of the selected 18 events is consistent with a value around 10 GeV [@jet; @rag]. Combining this with the information on the dijet mass sum, it can be concluded that the pair-produced particles should have masses of about 58 and 48 GeV, respectively. Pair-production of equal-mass particles is disfavoured.
At the moment little information can be extracted from angular distributions. From measurements of the “rapidity-weighted" jet charge, a variable that statistically retains information on the electric charge of the primary parton [@cha], one concludes [@jet] that the pair-produced particles have a sizeable charge. Electrically neutral particles are therefore disfavoured. Finally there is little or no presence of $b$ quarks in the final states [@jet; @rag]. This is another reason to reject the hypothesis of Higgs-boson production.
In this paper we want to study whether the ALEPH data, assumed here to correspond to a real physics signal, can be explained by pair production of a left-handed and a right-handed selectron, each particle decaying into two quarks, as an effect of $R$-parity violating interactions. Other interpretations of the four-jet events have already been presented in the literature [@oth], but to our knowledge this is the first example of a consistent picture of pair production and decay of particles with different masses, in agreement with the results of the ALEPH analysis.
Particle Production
===================
Let us start by considering the particle-production cross sections. Left-handed or right-handed selectrons can be pair-produced at LEP in any of the channels $\el \el$, $\er \er$, and $\el \er$. The interactions involved in the production of the pairs $\el \el$ or $\er \er$ always require that the incoming electron and positron have opposite helicities ([*i.e.*]{} collinear spin vectors). This means that the cross section for scalar particle production has a $\beta^3$ suppression, corresponding to a $p$-wave suppression near threshold. On the other hand, the production of $\el \er$ pairs involves the interaction of an electron and a positron with the same helicity ([*i.e.*]{} opposite spin vectors), and the cross section near threshold is proportional to $\beta$, corresponding to an $s$-wave.
The differential cross section for $\el \er$ production is (e\^+e\^-\^+\^-)= (e\^+e\^-\^+\^-) =\_[a,b=1]{}\^4 , \[cross\] A\_aN\_[a1]{}( N\_[a1]{}+) . Here $N_{a1}$ and $N_{a2}$ are the $B$-ino and $W_3$-ino components of the $a$-th neutralino with mass $M_{\chi^0_a}$. In the limit of a purely $B$-ino state, only one neutralino contributes to the sum in eq. (\[cross\]). Therefore the $U(1)$ gaugino mass $M_1$ is the most important parameter entering eq. (\[cross\]). For simplicity, we will concentrate on the case in which the lightest neutralino is a pure $B$-ino. Because of the necessary helicity flip proportional to the gaugino mass, the cross section decreases only as $M_1^{-2}$, for large $M_1$.
To reproduce the kinematical configuration suggested by the ALEPH data, we choose $m_{\el} =58$ GeV and $m_{\er} =48$ GeV. The electron sneutrino mass is then also determined by the weak $SU(2)$ relation m\_\^2 = m\_\^2+(1-\^2\_W)2 M\_Z\^2 , \[neum\] where $\tan\beta$ is the usual ratio of Higgs vacuum expectation values. We assume here that the sleptons of the second and third generations are heavier than those of the first one, and cannot have been produced at LEP. We will comment in sect. 5 on the case in which slepton masses are universal in flavour.
In fig. 1 we show the cross sections at $\sqrt{s}=130$ and 172 GeV for the different production channels $\el \er$, $\el \el$, $\er \er$, $\nl \nl$, and $\chi_1^0
\chi_1^0$ as a function of the $U(1)$ gaugino mass $M_1$, in the limit of large $\mu$ and $M_2$. The cross sections are corrected for initial-state radiation and have been generated by the numerical code SUSYXS [@man]. The important result is that, for $M_1$ less than about 100 GeV, the $\el \er$ production cross section is large, and consistent with the value suggested by the ALEPH data. Also, in the range $M_1=80$–100 GeV, all other particle–antiparticle production channels are quite suppressed. For the $\el \el$ and $\er \er$ channels, this is the result of an efficient destructive interference between the $s$-channel $\gamma$/$Z$ exchange and the $t$-channel neutralino/chargino exchange. To sufficiently suppress the $\nl$ cross section, we have to choose $m_{\nl}$ close to the upper bound determined by eq. (\[neum\]). This implies that $\tan \beta$ is close to 1, and the top Yukawa coupling is quite large. This can be made consistent with the absence of a Landau pole below the grand unification scale only if some new physics threshold exists. Indeed, the need of an effective supersymmetry-breaking scale $\Lambda_{\rm SUSY}$ much lower than the Planck scale is also suggested, in our scenario, by the presence of small slepton masses together with larger gaugino masses. In fact, for very large values of $\Lambda_{\rm SUSY}$, such hierarchy of masses would require a large amount of fine-tuning between the value of the slepton and the gaugino mass parameters at the high energy scale (for a recent discussion of the dependence of the renormalization group evolution of the scalar mass parameters on the effective supersymmetry breaking scale, see ref. [@car]).
The results presented in fig. 1 correspond to the case in which the lightest neutralino is a pure $B$-ino. Had we assumed unification of gaugino masses, and values of the higgsino mass $\mu$ not too large, then the cross section for $\el \er$ production could be larger than what is shown in fig. 1, as a consequence of the mixing between $B$-ino and $W_3$-ino, see eq. (\[cross\]). However the $\nl$ production cross section would also sizeably increase, because of the constructive interference between chargino and $Z$ exchange contributions. For instance, for large values of $|\mu|$, and $M_2 = 500 \; (300)$ GeV, the charged slepton production cross sections at $\sqrt{s} = 172$ GeV are not significantly modified, but the sneutrino cross section is enhanced from 0.4 pb to 0.51 (0.72) pb for $m_{\nl} = 53$ GeV, while for $m_{\nl} = 58$ GeV the cross section is enhanced from 0.31 pb to 0.40 (0.57) pb. Hence, values of $M_2 \simgt 500$ GeV will efficiently suppress the $\nl$ production cross section.
The differential cross section for $\el \er$ production, eq. (\[cross\]), leads to an angular distribution that is different from the usual scalar-particle pair production via gauge bosons in the $s$-channel with $d\sigma /dt \propto (ut-m^4)$. In fig. 2 we compare the two distributions as a function of the angle $\theta$ between the beam direction and one of the two dijet momenta ($0<\theta < \pi /2$). At the moment, the experimental uncertainties are too large for us to distinguish between the two distributions. If the charge of the primary parton is identified, one can measure the forward–backward asymmetry ${\cal A}_{FB}$ of the dijet system with a definite charge. In the case of ordinary scalar particle pair production, the distribution is symmetric in the forward and backward regions, and ${\cal A}_{FB}=0$. This is however not true for the distribution in eq. (\[cross\]), which produces the following integrated forward–backward asymmetry \_[FB]{}= . \[fbasy\] Here we have assumed that the $B$-ino is an approximate mass eigenstate, and defined the forward and backward regions with respect to initial- and final-state particles with the same electric charge. For $M_1=80$–100 GeV, ${\cal A}_{FB}$ is large, about 40–60% at $\sqrt{s}=130$ GeV and 30–50% at $\sqrt{s}=172$ GeV.
Particle Decay
==============
In order to generate the four-jet final state from the slepton pair, we have to introduce some $R$-parity violating interaction. The only renormalizable operator that couples quarks to leptons has the following expression in the superpotential: \_[ijk]{} L\_L\^i Q\_L\^j [|D]{}\_R\^k . \[rp\] We have used here a standard notation for lepton and quark chiral superfields, and denoted the generation indices as $i,j,k$. We assume that one of the couplings $\lambda_{ijk}$ is much larger than all the others; this coupling determines the decay mode of the lightest supersymmetric particle. As we do not want to consider top or bottom quarks in the final state, we are interested only in the couplings $\lambda_{1jk}$ with $j,k=1,2$. If in the future more experimental information on the flavour content of the jets becomes available, we will be able to further restrict the choice of the operators.
Our interpretation of the four-jet events as slepton pairs requires that the $R$-parity violating decay mode has a branching ratio close to 1. Thus it is important to compare the rate for $\el$ decay into two quarks, (\^- |[u]{}\_j d\_k )= m\_ , with the $R$-parity conserving decay rates. Indeed, $\el$ can decay into the lightest supersymmetric particle, $\er$, through neutralino exchange. In the approximation $M_{\chi^0}\gg m_{\el} , m_{\er}$, the decay widths are (\^- e\^- e\^+ \^- )= m\^3\_ F\_1 ( ) \_[a,b=1]{}\^4 , F\_1(x)=(1-x)(1+10x+x\^2)+6x(1+x)x , (\^- e\^- e\^- \^+ )= m\^5\_ F\_2 ( ) \_[a,b=1]{}\^4 , F\_2(x)=(1-x)(1-7x-7x\^2+x\^3)-12x\^2x . Also, $\el$ can decay into $\nl$ via $W$ exchange (\^- f |[f\^]{} )= N\_[ff\^]{} F\_2 ( ) , Here $N_{ff^\prime}$ is a colour factor, equal to 9, when summed over the light quarks and leptons in the final state. The decay rate for $\el^- \to e^- \nu_e
\bar{\nl}$ can be neglected, as it is suppressed by the chargino mass.
Figure 3 shows the value of $BR(\el^- \to \bar{u}_j d_k )$, as a result of a phase-space integration in the limit of a purely $B$-ino neutralino, but with no approximations on the ratio $M_{\chi^0_1}
/m_{\el}$. The $R$-parity violating mode dominates the $\el$ decays for $\lambda_{1jk}$ larger than few times $10^{-4}$. These values for the $R$-parity violating coupling constants are consistent with present bounds, as we discuss in the following.
Experimental bounds on $\lambda_{1jk}$ depend on the values of the squark masses, which mediate the effective four-fermion interactions between the leptons and quarks. We give here the bounds for a typical squark mass of 300 GeV, although the value of the squark mass does not enter into our analysis. The heavier the squarks are, the weaker the bounds on $\lambda_{1jk}$ become. From charged current universality, one finds $\lambda_{11k}<0.1$ [@bar]. From limits on $BR(K^+\to \pi^+ \nu \bar \nu )$, one finds $\lambda_{1jk}<0.03$ [@ber], although this limit depends on assumptions about the flavour structure. From radiative contributions to the electron neutrino mass, one can get significant limits only for $R$-parity violating operators that involve a third generation index [@hal]. From negative searches of neutrinoless double-$\beta$ decay, one obtains an interesting limit on $\lambda_{111}<8\times 10^{-3}$ [@ger], and a bound on the product $\lambda_{121}
\lambda_{112}<3\times 10^{-5}$ [@moh]. Experiments at HERA have set bounds on $\lambda_{1jk}$ of about $10^{-1}$ for squark masses of 200 GeV; these bounds disappear for values of the squark masses above 300 GeV [@her]. The only problematic constraint comes from cosmological considerations about the survival of a baryon asymmetry created at the very early stages of the Universe, which requires $\lambda_{1jk}<10^{-7}$ [@cos]. However, this limit does not apply to cosmological models with low-temperature baryogenesis, and can also be evaded under certain conditions [@cli].
We therefore conclude that there is a large range of $\lambda_{1jk}$ values, consistent with present bounds on $R$-parity violation, in which $\el$ decays almost entirely into a quark pair. This is true, although $\el$ is not the lightest supersymmetric particle, because the $R$-parity violating two-body decay is more important than kinematically suppressed three-body decay modes.
In our scenario, $\er$, the lightest supersymmetric particle, does not participate in the $R$-parity violating interaction in eq. (\[rp\]), which involves only quarks and left-handed leptons. Therefore the $\er$ decay will occur either via the small mixing $\phi$ between $\er$ and $\el$, (\^- |[u]{}\_j d\_k )= \_[1jk]{}\^2 \^2 m\_ , or via virtual neutralino and $\el$ exchange, (\^- e\^+e\^- |[u]{}\_j d\_k ) = G\_1( ) \_[a,b=1]{}\^4 , \[3bod1\] G\_1(x)=(4x\^2+25x+1)( ) -+x(3x+2)[Li]{}( ) , (\^- e\^-e\^- u\_j [|d]{}\_k ) = G\_2( ) \_[a,b=1]{}\^4 , \[3bod2\] G\_2(x)=(5x\^3-27x\^2-15x+1)( ) --6x\^2[Li]{}( ) .
In fig. 4 we show the $BR$($\er \to \bar{u}_j d_k $) as a function of the mixing angle $\phi$, in the limit of a purely $B$-ino neutralino. Again, although eqs. (\[3bod1\])–(\[3bod2\]) have been derived in the approximation $M_{\chi^0_a} \gg m_{\el} ,m_{\er}$, the results plotted in fig. 4 follow from a numerical integration of phase space with no restrictive assumptions.
The mixing angle $\phi$ is related to the higgsino mass $\mu$ and to the trilinear coupling $A$ by the relation = ( ) 10\^[-4]{} . Therefore the most plausible values for $\sin \phi$ lie in the range around $10^{-4}$. From fig. 4 we then infer that, in this range, $\er$ dominantly decays into two jets. In conclusion, although $\er$ does not participate in the $R$-parity violating interaction, the small left–right mixing ensures that the preferred $\er$ decay mode is into a quark pair, rather than into a phase-space suppressed four-body final state.
The $R$-parity violating coupling $\lambda_{1jk}$ does not influence the $\er$ decay branching ratio, as long as it is non-vanishing. It determines however the $\er$ lifetime, which is \_ = ( )\^2 ( )\^2 210\^[-13]{} [s]{} . For the relevant kinematical configuration, this correspond to a decay vertex displacement of about d\_ = \_( )\^2 ( )\^2 50 80 . Depending on the values of $\sin \phi$ and $\lambda_{1jk}$, this could be measured at LEP.
Prospects for LEP Searches at Higher Energies
=============================================
The best testing ground for the plausibility of the ALEPH data will come with the new LEP runs at higher $\sqrt{s}$. If the ALEPH signal is real and our interpretation correct, we should expect pair production of $\el \el$, $\er \er$, $\el \er$, and $\nl \nl$ with the rates shown in table 1. The preferred sneutrino decay mode is $\nl \to {\bar d}_j d_k$, as the $R$-parity conserving decay modes $\nl \to \nu_e e^\pm \er^\mp$ are suppressed by phase space and by the small mixing between $B$-ino and $W_3$-ino. Therefore all slepton production processes correspond to four-jet events, although the peaks in the distributions of the sum and difference of dijet masses depend on the process.
Charginos are expected to be too heavy to be produced at LEP, even if gaugino mass unification holds. There is however a chance to observe the lightest neutralino $\chi^0_1$, if the parameter $M_1$ is in the lower part of the allowed range. The $B$-ino state $\chi^0_1$ decays into two jets and an electron with more than 80% probability, or else into two jets and a neutrino. The relevant production cross sections are also shown in table 1.
Finally there are very good prospects for the discovery of the Higgs boson. Within the supersymmetric model with minimal Higgs structure, the low values of $\tan \beta$ assumed here imply that the lightest Higgs boson has Standard Model-like couplings and a mass, coming almost entirely from radiative corrections, roughly below 80 GeV [@rep].
The Case of Flavour Universality
================================
As we have mentioned before, in our analysis we have assumed that sleptons of the second and third generations are heavier than those of the first. This assumption is not inconsistent with the strong bounds on individual lepton number conservation, derived from $\mu \to e \gamma$ and similar processes. An approximate lepton flavour conservation can be the result of an alignment between leptons and sleptons, as a consequence of additional global symmetries [@nir] or of a dynamical principle [@noi].
Let us suppose now that the slepton supersymmetry-breaking masses are universal in flavour. Because of the mixing effect, we find that the lightest smuon and stau are lighter than $\er$ by an amount \_[[e]{}]{} = ( )\^2 4 [MeV]{} , \_[[e]{}]{} = ( )\^2 1 [GeV]{} . Thus the mainly right-handed stau is the lightest supersymmetric particle.
Because of the absence of $t$-channel contributions, pair productions of smuons, staus, and their corresponding sneutrinos do not suffer from destructive interference and have relatively large cross sections. Some indicative numbers are the following: For $m_{\tilde{\mu}_R} = 48$ GeV, the $\tilde{\mu}_R \tilde{\mu}_R$ production cross section is $0.55$ pb for $\sqrt{s}=130$ GeV and $0.54$ pb for $\sqrt{s}=172$ GeV. For $m_{\tilde{\tau}_R} = 47$ GeV, the $\tilde{\tau}_R \tilde{\tau}_R$ production cross section is $0.59$ pb for $\sqrt{s}=130$ GeV and $0.57$ pb for $\sqrt{s}=172$ GeV. For $m_{\tilde{\mu}_L} = 58$ GeV, the $\tilde{\mu}_L \tilde{\mu}_L$ production cross section is $0.18$ pb for $\sqrt{s}=130$ GeV and $0.43$ pb for $\sqrt{s}=172$ GeV. For $m_{\tilde{\tau}_L} = 59$ GeV, the $\tilde{\tau}_L \tilde{\tau}_L$ production cross section is $0.14$ pb for $\sqrt{s}=130$ GeV and $0.41$ pb for $\sqrt{s}=172$ GeV. Finally, for $m_{\tilde{\nu}} = 58 \; (53) $ GeV, the $\tilde{\nu} \tilde{\nu}$ production cross section is $0.21 \; (0.49)$ pb for $\sqrt{s}=130$ GeV and $0.31 \; (0.4) $ pb for $\sqrt{s}=172$ GeV.
The simultaneous presence of $R$-parity violating interactions with $\lambda_{ijk}\ne 0$, for different values of the index $i$, is severely constrained by lepton flavour-transition processes like $\mu \to e \gamma$. We are then led to assume that the second- and third-generation sleptons do not participate in the $R$-parity violating interaction, and consequently their decays have to involve real or virtual $\er$, $\el$, or $\nl$. Their signatures are therefore two jets accompanied by soft leptons or small amounts of missing energy. The presence of the leptons and/or neutrinos is a necessary feature of the transition between different generations of sleptons. Since such events have not been reported by any of the LEP experimental collaborations, we believe that the case of universality has to be rejected.
Conclusions
===========
In this paper we have assumed that the controversial ALEPH excess of four-jet events corresponds indeed to a physics signal and we have interpreted it in the context of a supersymmetric model with $R$-parity violation. If we consider non-universal mass terms for gauginos and for sleptons with different flavours, we find that $\el \er$ production can reproduce the four-jet events, while the production of other associated supersymmetric particles occurs at a much lower rate. Within an acceptable range of $R$-parity violating couplings, both $\el$ and $\er$ naturally have a decay branching ratio into two quarks very close to 1.
Our model is compatible with all discernible features emerging from the ALEPH data: the value of the cross section, the dijet mass difference, the absence of bottom quarks in the final state, and the jet charge content. It also predicts a specific angular distribution and a large forward–backward asymmetry in the jet charge. LEP runs at higher energies will be able to confirm or rule out this scenario.
We wish to thank P. Janot for very useful discussions. We also acknowledge conversations with J. Marcos and P. Morawitz. The work of S.L. is funded by a Marie Curie Fellowship (TMR-ERBFMBICT-950565).
\#1\#2\#3[[*Int. Jour. Mod. Phys.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Phys. Lett.* ]{}[**B\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Z. Phys.* ]{}[**C\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Phys. Rev. Lett.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Rev. Mod. Phys.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Phys. Rep.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Phys. Rev.* ]{}[**D\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Nucl. Phys.* ]{}[**B\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Mod. Phys. Lett.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Annu. Rev. Nucl. Part. Sci.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Sov. J. Nucl. Phys.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*JETP Lett.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Acta Phys. Polon.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Riv. Nuovo Cim.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Ann. Phys.* ]{}[**\#1 **]{}(19\#2) \#3]{} \#1\#2\#3[[*Prog. Theor. Phys.* ]{}[**\#1 **]{}(19\#2) \#3]{}
[99]{}
D. Buskulic (ALEPH Coll.), .
F. Ragusa, for the ALEPH Coll., talk at the LEPC Meeting, November 19, 1996.
D. Buskulic (ALEPH Coll.), .
V. Barger, W.Y. Keung, and R.J.N. Phillips, ;\
G.R. Farrar, ; preprint RU-96-71, hep-ph/9608387;\
A.K. Grant, R.D. Peccei, T. Veletto, and K. Wang, ;\
D. Choudhury and D.P. Roy, Phys. Rev. [**D54**]{} (1996) 6797;\
S.F. King, preprint SHEP-96-09, hep-ph/9604399;\
D.K. Ghosh, R.M. Godbole, and S. Raychaudhuri, preprint BU-TH-96-2, hep-ph/9605460;\
H. Dreiner, S. Lola, and P. Morawitz, preprint hep-ph/9606364;\
P.H. Chankowski, D. Choudhury, and S. Pokorski, preprint SCIPP-96-27, hep-ph/9606415.
M.L. Mangano , [*Event Generators for Discovery Physics*]{}, hep-ph/9602203, in Physics at LEP2, eds. G. Altarelli , CERN Report 96-01.
M. Carena, P.H. Chankowski, M. Olechowski, S. Pokorski, and C.E.M. Wagner, preprint CERN-TH-96-241, hep-ph/9612261.
V. Barger, G.F. Giudice, and T. Han, .
K. Agashe and M. Graesser, .
L.J. Hall and M. Suzuki, ;\
R.M. Godbole, P. Roy, and X. Tata, .
M. Hirsch, H.V. Klapdor-Kleingrothaus, and S.G. Kovalenko, [*Phys. Rev. Lett.* ]{}[**75**]{} (1995) 17.
K.S. Babu and R.N. Mohapatra, .
S. Aid (H1 Coll.), .
B. Campbell, S. Davidson, J. Ellis, and K.A. Olive, ;\
W. Fischler, G.F. Giudice, R.G. Leigh, and S. Paban, .
L.E. Ibañez and F. Quevedo, ;\
H. Dreiner and G.G. Ross, ;\
J.M. Cline, K. Kainulainen, and K.A. Olive, ;\
S. Davidson and R. Hempfling, preprint UCDPHY-96-27, hep-ph/9609497.
M. Carena , [*Higgs Physics*]{}, hep-ph/9602250, in Physics at LEP2, eds. G. Altarelli , CERN Report 96-01.
Y. Nir and N. Seiberg, .
S. Dimopoulos, G.F. Giudice, and N. Tetradis, .
--------------------- --------------- -------------------------------- --------------------------------
Process Particle mass $\sigma$ at $\sqrt{s}=186$ GeV $\sigma$ at $\sqrt{s}=195$ GeV
\[GeV\] \[pb\] \[pb\]
$\el \er$ $M_1=80$ 1.46 1.37
$\el \er$ $M_1=100$ 1.33 1.27
$\el \el$ $M_1=80$ 0.20 0.19
$\el \el$ $M_1=100$ 0.23 0.22
$\er \er$ $M_1=80$ 0.42 0.47
$\er \er$ $M_1=100$ 0.20 0.22
$\nl \nl$ $m_{\nl}=53$ 0.36 0.33
$\nl \nl$ $m_{\nl}=58$ 0.29 0.28
$\chi^0_1 \chi^0_1$ $M_1=80$ 0.62 0.80
--------------------- --------------- -------------------------------- --------------------------------
: Predictions for new particle production at future LEP runs. We have taken $m_{\el}=58$ GeV and $m_{\er}=48$ GeV.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We prove that the Sierpiński curve admits a homeomorphism with strong mixing properties. We also prove that the constructed example does not have Bowen’s specification property.'
address: 'National Supercomputing Centre IT4Innovations, Division of the University of Ostrava, Institute for Research and Applications of Fuzzy Modeling, 30. dubna 22, 70103 Ostrava, Czech Republic – and – Faculty of Applied Mathematics, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Kraków, Poland'
author:
- 'Jan P. Boroński'
- Piotr Oprocha
title: On dynamics of the Sierpiński carpet
---
Introduction
============
The aim of this note is to exhibit a homeomorphism of the Sierpiński curve (known as the planar universal curve or Sierpiński carpet) with some strong mixing properties. In 1993, Aarts and Oversteegen proved that the Sierpiński curve admits a transitive homeomorphism [@AO], answering a question of Gottschalk. They also showed that it does not admit a minimal one. Earlier, in 1991 Kato proved that the Sierpiński curve, does not admit expansive homeomorphisms [@K]. In [@Bis] the authors proved that the Sierpiński curve admits a homeomorphism with positive entropy. They also showed that it admits a minimal group action (by [@AO] it cannot be done using single homeomorphism). There has been quite a lot of interest in dynamical properties of the planar universal curve, also due to its occurrence as Julia sets of various complex maps (see e.g. [@D]). Nonetheless, we were unable to find any examples in the literature that would explicitly show homeomorphisms of the Sierpński curve with chaocity beyond Devaney chaos. The writing of the note was also motivated by some recent questions. During the Workshop on Dynamical Systems and Continuum Theory, at University of Vienna, in June of 2015 the following question was raised.
\[que:1\] Suppose a 1-dimensional continuum $X$ admits a mixing homeomorphism. Must $X$ be $\frac{1}{n}$-indecomposable for some $n$?
Recall that a continuum $X$ is *$\frac{1}{n}$-indecomposable*, if given $n$ mutually disjoint subcontinua of $X$ at least one of them must have empty interior in $X$. Note that the Sierpiński curve is $\frac{1}{n}$-indecomposable for no $n\in\mathbb{N}$. This is because it is locally connected, so every point has an arbitrarily small connected neighborhood. Our example is quite simple, however it relies on many nontrivial facts from topology and ergodic theory. In principle, the general strategy is very similar to the one in [@AO], but the starting point is a bit different. We start with an Anosov torus diffeomorphism, which allows us to say much more about the dynamics of the constructed map.
\[thm11\] The Sierpiński curve $S$ admits a homeomorphism $H\colon S \to S$ such that:
1. \[thm11:ber\] $H$ has a fully supported measure $\mu$, such that $(H,\mu)$ is Bernoulli,
2. \[thm11:per\] $H$ has a dense set of periodic points,
3. \[thm11:spec\] $H$ does not have specification property.
Since every Bernoulli measure is strongly mixing, and $\mu$ in Theorem \[thm11\] is fully supported, we immediately obtain the following result, answering Question \[que:1\] in the negative.
The Sierpiński curve $S$ admits a topologically mixing homeomorphism with dense set of periodic points.
By the arguments in the proof of Theorem \[thm11\], it seems very likely that Aarts-Oversteegen technique [@AO] which we employ here, will never lead to a map with specification property. This motivates the following natural question.
Does the Sierpiński curve admit a homeomorphism with the specification property?
Preliminaries
=============
By a dynamical system $(X,T)$ we mean a compact metric space $(X,d)$ with a continuous map $T\colon X\to X$.
We identify ${\mathbb{T}}^2$ with the quotient space ${\mathbb{R}}^2/{\mathbb{Z}}^2$ and denote $K^2=[-1/2,1/2]^2$.
A *Sierpiński curve* is any set homeomorphic to ${\mathbb{S}}^2 \setminus \bigcup_{i=1}^n \operatorname{int}D_i$ where
1. \[SC:1\] each $D_i$ is a disc and $D_i\cap D_j=\emptyset$ for $i\neq j$,
2. \[SC:2\] $\{D_i\}_{i=1}^\infty$ is a null sequence, i.e. the diameters of $D_i$ tend to zero, as $i\to\infty$.
3. \[SC:3\] $\bigcup_{i=1}^n D_i$ is dense in ${\mathbb{S}}^2$.
Whyburn [@W] proved that Sierpiński curve does not depend on the choice of the sequence of discs $\{D_i\}_{i=1}^\infty$, that is any two Sierpiński curves are homeomorphic.
Topological notions of mixing
-----------------------------
A dynamical system $(X,T)$ is topologically mixing if for any two nonempty open sets $U,V$ there is an $N>0$ such that $T^n(U)\cap V\neq \emptyset$ for all $n\geq N$. There are many different extensions of the above property to characterize stronger mixing in the system. From the point of view of our work the following two are very important. It is not hard to see that they imply topological mixing.
In his seminal paper [@Bowen] Bowen introduced an important, strong version of mixing, called (periodic) specification property. Let $T\colon X\to X$ be a continuous onto map. Following Bowen, we say that $(X,T)$ has the *specification property* if for any ${\varepsilon}> 0$, there is a positive integer $N=N({\varepsilon})$ such that for any integer $s\geq 2$, any $s$ points $y_1,\dots,y_s\in X$, and any sequence $0=j_1\leq k_1 < j_2 \leq k_2
< \dots < j_s \leq k_s$ of $2s$ integers with $j_{m+1} - k_m\geq N$ for $m= 1,\dots,s-1$, there is a point $x\in X$ such that, for each positive integer $m\leq s$ and all integers $i$ with $j_m \leq i \leq k_m$, the following two conditions hold: $d(T^i(x),T^i(y_m))<{\varepsilon}$. If, in addition, we can always select $x$ as a periodic point such that $T^{k_m-j_1+N}(x)=x$ then $(X,T)$ has the *periodic specification property*. Note that the problem of characterizing the relations between various types of mixing for maps in specified classes of one-dimensional continua is of high interest (e.g. see [@MH] and references therein).
Invariant measures
------------------
Let $X$ be a compact metric space with metric $d$ and let $M(X)$ be the space of Borel probability measures on $X$ equipped with the [*Lévy-Prokhorov metric*]{} $\rho$ defined by $$\rho(\mu, \nu)=\inf\{{\varepsilon}\colon \mu(A)\leq \nu(A^{{\varepsilon}})+{\varepsilon}\text{ for all Borel subsets } A\subset X\},$$ where $A^\delta=\{x : \operatorname{dist}(x,A)<\delta\}$. The topology induced by $\rho$ coincides with the weak$^*$-topology on $M(X)$. It is also well known that $(M(X), \rho)$ is a compact metric space. For a dynamical system $(X,T)$ we denote by $M_T(X)$ the set of all $T$-invariant measures. For more details on Lévy-Prokhorov metric and weak\*-topology the reader is referred to [@Huber], and basic properties related to invariant measures (ergodicity, strong mixing, Bernoulli shift) can be found in [@Walters].
Quasi-Hyperbolic Toral Automorphisms
------------------------------------
Let $A$ be a $2\times 2$ matrix with integer entries such that $|\det A|=1$. Then $A^{-1}$ also has integer entries, and so $A$ induces a homeomorphism of the 2-dimensional torus $F\colon {\mathbb{T}}^2 \to {\mathbb{T}}^2$ by $F(x)=Ax\textrm{(mod 1)}$, e.g. see [@Brin] for more details. Since $|\det A|=1$, every toral automorphism preserves Lebesgue measure. It is known that the periodic points of an ergodic toral automorphism are exactly those with rational coordinates (see [@Denker Proposition 24.7]). It was first proved by Adler and Weiss for ${\mathbb{T}}^2$ and then extended by Katznelson to each ${\mathbb{T}}^n$, that if toral automorphism is ergodic with respect to the Lebesgue measure, then it is measure-theoretically conjugate to a Bernoulli shift (e.g. see [@Denker Theorem 24.6]). Following Lind [@Lind] we say that $F$ is *quasi-hyperbolic* if $A$ does not have roots of unity as eigenvalues. In dimension $2$ every quasi-hyperbolic automorphism must be hyperbolic, that is, it does not have eigenvalues on the unit circle, and has strong specification property [@Lind].
Branched Covering from $\mathbb{T}^2$ to $\mathbb{S}^2$ {#sec:quotient}
-------------------------------------------------------
Represent ${\mathbb{T}}^2$ as $\mathbb{R}^2/\mathbb{Z}^2$. Take a quotient of $\mathbb{T}^2$ by the relation $J$, that identifies $(x,y)$ with $(-x,-y)$. $J$ induces a branched covering map $\pi \colon {\mathbb{T}}^2\to {\mathbb{S}}^2$ (see e.g. [@Walters p. 140]), which is 2-to-1 except at four branch points in $\mathbb{T}^2$ given by ${\mathcal{C}}=\{(0,0), (1/2,0), (0,1/2), (1/2,1/2)\}$. Since the relation $J$ is preserved by any toral automorphism, for every toral automorphism $F$ we obtain a factor map $G\colon{\mathbb{S}}^2\to {\mathbb{S}}^2$ such that $G\circ \pi=\pi \circ F$. Note that if $x\not\in {\mathcal{C}}\cup F^{-1}({\mathcal{C}})$ then there is an open neighborhood $U$ of $x$ such that $U$ has at most one element of any equivalence class of the relation $J$ and the same holds for $F(U)$. Then on $U$ the factor map $\pi$ is a local isometry.
Proof of Theorem \[thm11\]
==========================
Start with “Arnold’s cat map” $F\colon \mathbb{T}^2\to\mathbb{T}^2$ on the torus given by $$F(x,y)=(2x+y,x+y)\textrm{(mod 1)}.\label{ACM}$$ Clearly $F$ is hyperbolic with eigenvalues $\lambda_1=\frac{3+\sqrt{5}}{2}$ and $\lambda_2=\frac{-\sqrt{5}-1}{2}$, hence it has periodic specification property. There is also an invariant foliation of $\mathbb{T}^2$ by lines $\mathcal{L}$, that is $F(\mathcal{L})=\mathcal{L}$. Denote the Lebesgue measure on ${\mathbb{T}}^2$ by $\lambda$. $F$ has a dense set of periodic points and $({\mathbb{T}}^2,F,\lambda)$ is measure-theoretically conjugate to Bernoulli shift. Let $\pi \colon {\mathbb{T}}^2\to{\mathbb{S}}^2$ be the quotient map from Section \[sec:quotient\] and let $G$ be the induced homeomorphism of ${\mathbb{S}}^2$. Since $G$ is a factor of $F$, $({\mathbb{S}}^2,G,\mu)$ is Bernoulli with respect to a fully supported measure $\mu$ which is a push-forward of $\lambda$ by $\pi$ (see [@Walters Theorem 4.29(ii)]). Let $\mathcal{G}=\pi(\mathcal{L})$. Clearly $\mathcal{G}$ is an invariant foliation of ${\mathbb{S}}^2$; i.e. $G(\mathcal{G})=\mathcal{G}$.
Let $\mathcal{O}=\{O_n:n\in\mathbb{N}\}$ be a dense family of periodic orbits of $G$. We assume that each $O_n\cap \pi({\mathcal{C}})=\emptyset$ and that $\operatorname{Per}(G)\setminus \mathcal{O}$ is dense in ${\mathbb{S}}^2$.
We will modify $G$ inductively, blowing up consecutive periodic orbit. Take $O_1\in \mathcal{O}$, and assume that it has period $p_1$, say $O_1=\{c,G(c),\ldots,G^{p_1-1}(c)\}$. Since $\pi^{-1}(O_1)\cap {\mathcal{C}}=\emptyset$, there are open discs $D_0,\ldots, D_{p_1-1}$ such that $\pi(D_i)\cap \pi(D_j)=\emptyset$ for $i\neq j$ and $\pi$ is 1-1 on each $D_i$. Let $U_i=\pi(D_i)$. Then we have a natural foliation of each $U_i$ (induced locally from $D_i$) by $G^i(x)\cap\mathcal{G}$, such that if $L\subset U_i$ is a sufficiently short line emerging from $G^i(c)$ then $G(L)\cap U_{i+1}$ is contained in the corresponding line. In other words, we have a continuous foliation of a neighborhood of each point $G^i(c)$ into lines, and $G$ preserves these lines.
Let $\mathcal{F}_i=\{l^i_\theta: \theta\in [0,2\pi)\}$ be a foliation of a neighborhood $U_i$ of $G^i(c)$ mentioned above by the lines emanating from $G^i(c)$. We remove $O_1$ and compactify each $\operatorname{cl} U_i\setminus \{G^{i}(c)\}$ by a topological copy $S^i_c$ of the unit circle $\mathbb{S}^1$ adding, for a fixed $i$, a point $\theta_i \in S^i_c$ compactifying the leaf $l^i_\theta$. That way we obtain a $p_1$-punctured sphere $S_1$. We may easily extend $G$ to a map $H_1\colon S_1\to S_1$ by setting $H_1(\theta_i)=\theta'_{j}$, where $G(l^i_\theta)=l^{j}_{\theta'}$, $j= i+1 (\textrm{mod }p_1)$, and $\theta'_j\in S^j_c$. Clearly $H_1$ defined that way is invertible with a continuous inverse, so $H_1$ is a homeomorphism of $S_1$. Observe that the dynamics of all other points under $H_1$ in $S_1$ is exactly the same as on ${\mathbb{S}}^2$ for $G$. Hence we can repeat this procedure, puncturing $S_2$ and replacing the periodic orbit $O_2$ by a periodic sequence of circles. Inductively, we obtain a sequence of spheres with $\sum_{i=1}^n p_i$-holes $S_n$, homeomorphisms $H_n\colon S_n\to S_n$ and factor maps $\pi_{n}\colon S_n \to S_{n-1}$ which collapse newly introduced circles into points of $O_{n-1}$, where $S_0={\mathbb{S}}^2$ and $H_0=G$. In other words, $\pi_{n}$ reverts the modification made in step $n$. Clearly, each $\pi_n$ is a continuous onto map and $\pi_n \circ H_{n+1}=H_n\circ \pi_n$.
Embed each $S_n$ in ${\mathbb{S}}^2$ in a natural way, and extend $\pi_n$ to a map $\eta_n\colon {\mathbb{S}}^2\to {\mathbb{S}}^2$ in the following way. If $D$ is an open disc bounded by $S_n$ in ${\mathbb{S}}^2$ then we fix any $y\in \partial D\cap S_n$ and define $\eta_n(x)=y$ for every $x\in D$. Since $\pi_n$ collapses the circle $\partial D$ to a point, each $\eta_n$ is an almost homeomorphism; i.e. it can be obtained as a uniform sequence of homeomorphism. Denote by $S_\infty$ the inverse limit of spaces $S_n$ with bonding maps $\pi_n$ and by $Q_\infty$ the inverse limit of spheres ${\mathbb{S}}^2$ with $\eta_n$ as bonding maps; i.e. $$S_\infty=\{(z_0,z_1,\ldots):\pi_n(z_n)=z_{n-1})\},$$ $$Q_\infty=\{(z_0,z_1,\ldots):\eta_n(z_n)=z_{n-1})\}.$$ Since each $\eta_n$ is an almost homeomorphism, a result of Brown [@Brown Theorem 4] implies that $Q_\infty$ is homeomorphic to ${\mathbb{S}}^2$. Observe that if we fix any $c\in O_n$ then the set $B$ of all inverse sequences in $Q_\infty$ with $c$ on the first coordinate is homeomorphic to a disc. Simply, after dropping $n$ first coordinates we see that $B$ is an inverse limit of a disk $D$ with the identity as a unique bounding map. But $\bigcup_{n=1}^\infty O_n={\mathbb{S}}^2$, hence $S_\infty$ satisfies conditions (S\[SC:1\])–(S\[SC:3\]) and so is a Sierpiński curve. Observe that if we put $H=H_1\times H_2\times \ldots \times H_n\times \ldots$ then $H(S_\infty)=S_\infty$, hence $H$ is a homeomorphism of the Sierpiński curve. Let $M={\mathbb{S}}^2\setminus \bigcup_{n=1}^\infty O_n$ and $M_\infty=\{(z_0,z_1,\ldots)\in S_\infty : z_0\in M\}$ be the set of all inverse sequences in $S_\infty$ with the first coordinate in $M$. It follows directly from the construction that if $z\in M_\infty$ then $z=(x,x,x\ldots)$ for some $x\in M$ and $H(z)=(G(x), G(x), \ldots)$. Since periodic orbits of $G$ in $M$ are dense in ${\mathbb{S}}^2$, it is not hard to see that $H$ has a dense set of periodic points. The set $M_\infty$ is Borel, so for any Borel set $U\in S_\infty$ we can view $U\cap M_\infty$ as a Borel subset of $M$ (by projection onto the first coordinate) and so we obtain a well defined $H$-invariant Borel probability measure $\nu(U)=\mu(U\cap M_\infty)$. The measure $\mu$ is ergodic, so we have $\mu({\mathbb{S}}^2\setminus M_\infty)=0$, hence also $\nu(S_\infty\setminus M_\infty)=0$, and so $\nu$ and $\mu$ are isomorphic, in particular $(S_\infty, H, \nu)$ is measure-theoretically conjugate to a Bernoulli shift. Take any open set $U$ in $S_\infty$. We claim that $\nu(U)>0$. Indeed, the basic open sets in $S_\infty$ are given by $U_\infty=(\eta_1\circ\ldots\circ\eta_{i-1}(U_i),\ldots,\eta_{i-1}(U_i), U_i,\eta_i^{-1}(U_i),\ldots),$ where $U_i\subseteq S_i$ is open, for some $i\in\mathbb{N}$ (see e.g. Theorem 3 on p.79 in [@Engelking]). Since $S_i$ is a sphere with a finite number of holes, the Lebesgue measure of $U_i$ in $S_i$ is positive and $\nu(U_\infty)=\nu(U_\infty \cap M_\infty)=\mu(U_i\cap M)$, therefore $U_\infty$ has positive product measure. This shows that $\nu$ has full support, which completes the proof of Theorem \[thm11\].
It remains to prove . Assume on the contrary that $H$ has the specification property. Since the specification property is preserved under higher iterations, $H^{p_1}$ has the specification property, where $p_1$ is the period of $O_1$. For simplicity of notation replace $H$ by $H^{p_1}$ and $A$ by $A^{p_1}$. By [@PS Theorem 2.1] the specification property implies that for every invariant measure $\mu\in M_T(S_\infty)$ there exists a sequence of ergodic measures such that $\mu_n\to \mu$, when $n\to \infty$, in Lévy-Prohorov metric. Since we blew up a hyperbolic periodic point $c$ in $O_1$ in the first step of our construction, after passing from $A$ to the coordinates giving its Jordan form, we have locally a phase portrait (for $G$ and $H$) as on Figure \[fig1\].
![Phase portrait for a hyperbolic point before and after a “blow-up”[]{data-label="fig1"}](phase1.pdf){width="80.00000%"}
\[phase\]
Let us start with the following observation. Consider the hyperbolic linear map $f(x,y)=(ax,by)$ where $0<a<1<b$ and $ab=1$. Let $D=[-{\varepsilon},{\varepsilon}]^2$ for some small ${\varepsilon}>0$. Now let $z=(p,q)\in D$ with $|p|\geq |q|$. Next assume that the trajectory of $z$ is not fully contained in $D$. Then there exists a minimal $m\geq 1$ such that $a^m|p|\geq b^m|q|$ and $a^{m+1}|p|< b^{m+1}|q|$. Observe that $b^{2m}|q|<{\varepsilon}$ as otherwise ${\varepsilon}\leq b^{2m}|q|=a^{-m}b^m|q|\leq |p|$ and so $(p,q)\not\in D$ which is a contradiction. Now, let $v$ be a compactification of the leaf representing the stable direction for hyperbolic point $c$ and take a small neighborhood $U$ of $v$. Let $U'=\pi(U)$ where $\pi$ is the natural factor map $\pi \colon (S_\infty,H) \to ({\mathbb{S}}^2,G)$. If $U$ is sufficiently small, then $\pi(U)\subset D$ and furthermore, if $(p,q)\in \pi(U)$ then $|p|\geq |q|$.
Fix any periodic point $u\in S_\infty$, say of period $s$, and consider the invariant measure $\hat \mu =(1-\alpha) \delta_c + \frac{\alpha}{s}\sum_{i=0}^{s-1} \delta_{H^i(u)}$ with a small $\alpha$, say $\alpha<\frac{1}{10}$. Assume also that $\pi(u)\not\in D$. Take $\delta<2s \alpha$ such that $\operatorname{dist}(c, \{u,H(u),\ldots, H^{s-1}(u)\})>3{\varepsilon}$. Let $\delta$ be small enough, so that there exists an open set $c\in V$ such that $B(V,2\delta)\subset U$ and $\pi(B(u,2\delta))\cap D=\emptyset$. Denote $W=B(u,2\delta)$. There exists an ergodic measure $\hat \nu$ such that $\rho(\hat \nu,\hat \mu)<\delta$. This implies that $\hat \nu(U)\geq \hat \nu(V^\delta)\geq \hat \mu(V)-\delta \geq 4/5$ and $
\hat \nu(W)\geq \hat \mu(\{u\}^\delta)-\delta>0.$ By the Birkhoff ergodic theorem there exists $x\in S_\infty$ such that $\lim_{n\to \infty}\frac{1}{n}\{j<n: H^j(x)\in U\}=\hat \nu(U)$ and $\lim_{n\to \infty}\frac{1}{n}\{j<n: H^j(x)\in W\}=\hat \nu(W).$ Since $\hat \nu(W)>0$ there exists an increasing sequence $k_i$ such that $H^{k_i}(x)\in W$. Let us estimate the number of iterations $k_{i}\leq j <k_{i+1}$ such that $H^j(x)\in U$. Observe that if $\pi(H^{k_i}(x))\not\in D$ and $\pi(H^{k_{i+1}}(x))\not\in D$ then, by the earlier analysis, we see that no more than half of iterations $H^j(x)$ for $j=k_i+1,...,k_{i+1}$ can visit $U$. This implies that $\lim_{i\to \infty} \frac{1}{k_i}|\{j : H^j(x)\in U\}|\leq1/2$. By the choice of $x$ we obtain that $\hat \nu(U)=1/2<4/5$ which is a contradiction. This shows that $(S_\infty,H)$ does not have the specification property, completing the proof.
Acknowledgments
===============
The authors express many thanks to L‘ubomir Snoha for some helpful discussions on the properties of Besicovitch’s homeomorphism used in the construction in [@AO] and Andrzej Biś for discussion on constructions in [@Bis] and related topics.
The authors’ work was supported by the European Regional Development Fund in the IT4Innovations Center of Excellence project (CZ.1.05/1.1.00/02.0070).
J. Boroński also gratefully acknowledges the partial support from the MSK DT1 Support of Science and Research in the Moravian-Silesian Region 2014 (RRC/07/2014).
[12]{} Proc. Amer. Math. Soc. 120 (1994), no. 3, 965–968. Modelling minimal foliated spaces with positive entropy. Hokkaido Math. J. 36 (2007), no. 2, 283–310. *Periodic points and measures for Axiom A diffeomorphisms*, Trans. Amer. Math. Soc. [154]{} (1971), 377-397. . Cambridge University Press, Cambridge, 2002. *Some applications of an approximation theorem for inverse limits.* Proc. Amer. Math. Soc. 11 (1960), 478-483. *Ergodic theory on compact spaces*. Lecture Notes in Mathematics, Vol. 527. Springer-Verlag, Berlin-New York, 1976. Notices Amer. Math. Soc. 51 (2004), no. 1, 9–15. Biblioteka Matematyczna, Tom 25 Państwowe Wydawnictwo Naukowe, Warsaw, 1965. *Hierarchies of chaotic maps on continua*. Ergodic Theory Dynam. Systems 34 (2014), no. 6, 1897–1913. *Robust statistics*. Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, Inc., New York, 1981. Topology Appl. 34 (1990), 161–165. . Ergodic theory (Proc. Conf., Math. Forschungsinst., Oberwolfach, 1978), pp. 93–104, Lecture Notes in Math., 729, Springer, Berlin, 1979. . Monatsh. Math. 89 (1980), no. 2, 121–129. *Large deviations estimates for dynamical systems without the specification property. Applications to the $\beta$-shifts.* Nonlinearity 18 (2005), no. 1, 237–261. Graduate Texts in Mathematics, 79. Springer-Verlag, New York-Berlin, 1982. Fund. Math. 45 (1958) 320–324.
\[lastpage\]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The operator structures that can contribute to three-nucleon forces are classified in the $1/N_c$ expansion. At leading order in $1/N_c$ a spin-flavor independent term is present, as are the spin-flavor structures associated with the Fujita-Miyazawa three-nucleon force. Modern phenomenological three-nucleon forces are thus consistent with this ${\cal O}(N_c)$ leading force, corrections to which are suppressed by a power series in $1/N_c^2$. A complete basis of operators for the three-nucleon force, including all independent momentum structures, is given explicitly up to next-to-leading order in the $1/N_c$ expansion.'
author:
- 'Daniel R. Phillips$^{1}$'
- 'Carlos Schat$^{1,2}$'
title: 'Three-nucleon forces in the $1/N_c$ expansion'
---
Introduction
============
Over the last fifteen years advances in few-body methods and the steady increase in computational power have enabled numerically accurate calculations of few-nucleon scattering observables and the spectra of light nuclei. In the three-nucleon system such calculations show clear evidence for three-nucleon forces (3NFs) when compared with experimental data [@KalantarNayestanaki:2011wz; @Hammer:2012id]. The simplest and best known example of this is that the triton binding energy is underestimated by about 800 keV if a Hamiltonian with two-nucleon potentials alone is employed [@Fr93]. A similar underbinding occurs for other light nuclei as well [@Pudliner:1997ck; @Wiringa:2000gb; @Pieper:2002ne]. (Although, see Ref. [@Nogga:2004ab] for a study of the dependence of this conclusion on the resolution scale at which the NN potential is defined.) Indeed, the role of three-nucleon forces in the spectra of light nuclei has been a subject of intense investigation during this period (see, e.g. [@Pieper:2001mp; @Na07; @Ma12], as well as Ref. [@Hammer:2012id]). Recently, state-of-the-art treatments of the role of 3NFs in heavier nuclei show that they could play a role in determining the location of the neutron-drip line in the oxygen and calcium isotopes [@Otsuka:2009cs; @Holt:2010yb], and in extending the half life of Carbon-14 [@Holt:2009uk].
Historically, 3NFs were first derived in the classic paper of Fujita and Miyazawa [@FM57]. There, a 3NF due to the exchange of two pions was computed. This 3NF still forms a key portion of the 3NFs employed today, appearing, for example, in the Urbana three-nucleon force [@Carlson:1983kq; @CS98]: $$V_{ijk}=V_{ijk}^{2 \pi} + V_{ijk}^R \ ,
\label{eq:Urbana}$$ with [@Hu99] $$\begin{aligned}
&& V_{ijk}^{2 \pi}=\tilde{A}_{2 \pi} \frac{\sigma_i \cdot {\bf k}_1 \sigma_k \cdot {\bf k}_2}{({\bf k}_1^2 + m_\pi^2)({\bf k}_2^2 + m_\pi^2)}\nonumber\\
&& \qquad \left[(a + b \, {\bf k}_1 \cdot {\bf k}_2) \tau_i \cdot \tau_k + d \, \tau_i \cdot (\tau_j \times \tau_k) \sigma_j \cdot ({\bf k}_1 \times {\bf k}_2 )\right] \ .
\label{eq:FM}\end{aligned}$$ Here ${\bf k}_{1,2}$ are the momenta of the two pions in the exchange, $\sigma$ and $\tau$ are the usual Pauli matrices for nucleon spin and isospin, and the coefficients $a$, $b$, and $d$ represent the strength of s-wave and p-wave $\pi$N scattering. If, as was assumed by Fujita and Miyazawa, we take the p-wave pieces to arise from the spin-3/2, isospin-3/2, $\pi$N channel, where the $\Delta(1232)$ resides, we have $b=4d$.
Meanwhile, the term $V_{ijk}^R$ in Eq. (\[eq:Urbana\]) is spin and isospin independent, and produces repulsion. The strength of this term, and the overall strength of $V_{ijk}^{2
\pi}$, are adjusted so that calculations with the AV18 NN potential and this 3NF reproduce the triton binding energy and “…provide additional repulsion in hypernetted-chain variational calculations of nuclear matter near equilibrium density" [@CS98]. The combination AV18/Urbana is quite successful in describing the spectrum of nuclei up to $A=8$ [@Pudliner:1997ck; @Wiringa:2000gb]. But, it does fail to predict the correct isospin dependence of binding in these systems, and also underpredicts the spin-orbit splitting of, e.g., the 3/2$^-$ and 1/2$^-$ resonances in the A=5 system. Consequently, the Urbana 3NF has been updated to produce a set of “Illinois" potentials, which include (phenomenologically, at least) the effect of “pion ring" diagrams, and have 2–3 parameters that are tuned to reproduce levels in the spectra of nuclei up to $A \leq
8$ [@Pi01]. These potentials, when acting in concert with the AV18 NN force, do a good job of describing spectra in systems with A=9 and 10 [@Pieper:2002ne].
However, it is not obvious that the Urbana and Illinois potentials are grounded in QCD. Some of the structures are derived from diagrams involving pion exchange, but the coefficient functions in front of those structures are, in some cases, chosen for ease of numerical implementation, and given strengths which are adjusted to reproduce data. Closer connection to the chiral symmetry of QCD was sought in, e.g. the Tucson-Melbourne 3N potential, which considered the role of the $\rho$ meson, as well as the constraints of chiral symmetry on the $\pi$N amplitude which appears in the two-pion-exchange 3NF [@Co79; @CP93]. The Brazilian 3NF also attempted to impose constraints from chiral symmetry [@RIF84].
The advent of chiral perturbation theory ($\chi$PT) as a tool for analyzing nuclear forces resulted in the derivation of a 3NF which is in accord with the pattern of chiral-symmetry breaking in QCD [@vK94]. If the chiral expansion is applied directly to the 3N potential—as was done in Ref. [@vK94]—then three contributions occur at leading order (LO). They are: a short-range, spin-isospin independent piece (as in the $V_{ijk}^R$ of Eq. (\[eq:Urbana\])); a piece associated with the short-range emission of a pion by an NN pair with its subsequent absorption by the third nucleon; and a two-pion-exchange 3NF. The $\pi$N amplitude that appears in the two-pion-exchange piece of the chiral 3NF involves LECs from ${\cal L}_{\pi N}^{(2)}$: $c_1$, $c_3$, and $c_4$. The LECs $c_3$ and $c_4$ encode p-wave $\pi$N scattering, so $\chi$PT has the Fujita-Miyazawa force as one of the dominant pieces of its 3NF. (Indeed, if a variant of $\chi$PT with an explicit Delta degree of freedom is employed then the Fujita-Miyazawa 3NF occurs one order earlier than the other pieces of the chiral 3NF [@vK94; @EKM08].) The leading $\chi$PT 3NF has been used to investigate scattering in the 3N system [@Ep02], and nuclear spectra in [*ab initio*]{} calculations up to A=13 [@Na07; @Ma12]. And, as mentioned above, it has, under certain approximations to the many-body physics, been shown to improve descriptions of the binding of neutron-rich nuclei [@Otsuka:2009cs; @Holt:2010yb]. It has also been applied to obtain an equation of state for neutron-rich matter [@HS10].
In spite of these successes, puzzling discrepancies between theory and data persist. One example is the analyzing power $A_y$ in neutron-deuteron scattering at low energies, with a similar issue also occurring for neutron-${}^3$He scattering (see, e.g. Ref. [@Clegg:2009zz]). No modification of NN potentials which is consistent with the NN data and the dominance of one-pion exchange at long range seems able to explain this discrepancy, leaving the “$A_y$ puzzle" firmly in the realm of 3NFs to resolve. However, neither the model 3NFs on the market, nor the LO chiral 3NF described in the previous paragraph, can do so. Of course, extending a $\chi$PT calculation of the 3NF to higher orders in the chiral expansion might reveal the operator and mechanism (or mechanisms) which solves this problem, and work along these lines is in progress [@Be08; @Be11; @Krebs:2013kha; @Girlanda:2011fh]. But, as the chiral order increases, classifying the possible 3NF operators becomes very involved. It would be interesting to have an additional tool that could help sort out the most relevant operator structures.
The $1/N_c$ expansion of QCD can be used to provide this kind of insight [@'tHooft:1973jz; @Witten:1979kh]. This approach to the non-perturbative regime of QCD has proven very useful in the study of baryons [@Jenkins:2009wv], for reviews see [@Jenkins:1998wy; @Manohar:1998xv]. In the context of nuclear forces the $1/N_c$ expansion was first used to study the central part of the NN potential by Savage and Kaplan [@KS96], and then to analyze the complete potential, classifying the relative strengths of the central, spin-orbit and tensor forces, by Kaplan and Manohar [@KM97]. These authors analyzed the NN potential for momenta of order $N_c^0$, i.e. $p \sim \Lambda_{QCD}$, and found that it is an expansion in $1/N_c^2$. Furthermore, the $1/N_c^2 \approx 1/10$ (in our world) hierarchy between the different contributions to the NN potential is roughly borne out in the Nijm93 [@St94] NN potential. The arguments that lead to this conclusion will be recapitulated in Sec. \[sec-NN\].
In this work we extend that analysis to the three-nucleon system, classifying the possible operator structures that can contribute to a general 3NF according to a counting in $1/N_c$. We do this by computing the energy of the 3N system as $N_c \rightarrow
\infty$, starting with the Hartree expansion for the nuclear Hamiltonian in the large-$N_c$ limit [@Dashen:1994qi; @KM97]: $$\begin{aligned}
H = N_c \sum_{s, t, m} v_{stm} \left(\frac{S}{N_c}\right)^s
\left(\frac{I}{N_c}\right)^t \left(\frac{G}{N_c}\right)^m \ ,
\label{eq:Hartree}\end{aligned}$$ where we suppressed spin and isospin indices in the spin-flavor structures $O=\{S,I,G\}$ and vector indices in the coefficients $v$. These coefficients are, in fact, ${\cal O}(1)$ functions of the momenta. The explicit factors of $1/N_c$ ensure that an $m$-body interaction scales generically as $1/N_c^{m-1}$, as mandated by large-$N_c$ QCD counting [@Witten:1979kh]. Spin, isospin and vector indices are contracted so that $H$ is rotation and isospin invariant, as well as parity even and time-reversal even. In a quark-operator basis the spin-flavor structures are given by one-body operators $$\begin{aligned}
\label{qop}
S^i = q^\dagger \frac{\sigma^i}{2} q , \quad I^a = q^\dagger \frac{\tau^a}{2} q , \quad G^{ia} = q^\dagger \frac{\sigma^i \tau^a}{4} q , \end{aligned}$$ where $q^\dagger, q$ are creation and annihilation operators for the light quarks $u,d$ and $\sigma$, $\tau$ are the standard $SU(2)$ Pauli matrices acting on spin and isospin, respectively. Taken together, the 15 operators in Eq. (\[qop\]) generate the $SU(4)$ algebra $$\begin{aligned}
[S^i,S^j]=i\epsilon^{ijk} S^k \ , && \ [S^i,G^{ja}]=i\epsilon^{ijk} G^{ka}, \nonumber \\
{[}I^a,I^b]=i\epsilon^{abc} I^c \ , && \ {[}I^a,G^{ib}]=i\epsilon^{abc} G^{ic}, \nonumber \\
{[}S^i,I^a]=0 \ , && {[}G^{ia},G^{ib}]=\frac{i}{4} \delta^{ij} \epsilon^{abc} I^c + \frac{i}{4} \delta^{ab} \epsilon^{ijk} S^k.
\label{eq:algebra}\end{aligned}$$ Since we are interested in taking matrix elements between nucleon states we will indicate with $O_{\alpha}$ that the operator $O$ acts on nucleon $\alpha=1,2,3$, so that $S,I,G$ in Eq. (\[eq:Hartree\]) can be any of $S_{\alpha},I_{\alpha},G_{\alpha}$. But products of operators acting on the same nucleon in Eq. (\[eq:Hartree\]) must be reduced to a single operator. As is explained in Secs. \[sec-NN\] and \[sec-redrules\], this is achieved using the relations and reduction rules for the powers of the basic operators ${S,I,G}$ that act on the same nucleon, which are discussed in Ref. [@Dashen:1994qi]. The contributions to the 3NF that result after such reduction can be straightforwardly estimated, since matrix elements of $S$ and $I$ between nucleon states are ${\cal O}(1)$, which is in contrast to matrix elements of $G$, which are ${\cal O}(N_c)$. The leading force will thus be constructed out of $G$’s and unit operators, acting on the different nucleons. In fact, the algebra Eq. (\[eq:algebra\]) was derived in the one-nucleon sector for external nucleon momenta of order $N_c^0$, and so this conclusion holds in that kinematic regime (a similar remark applies to the NN potential derived in Ref. [@KM97]). If results for lower momenta are desired then the counting of operators obtained here can be modified accordingly. We present the analysis of leading and sub-leading 3NFs in the $1/N_c$ expansion in Sec. \[sec-3N\], and summarize our conclusions in Sec. \[sec-conclusion\].
In the large-$N_c$ limit the mass of the nucleon tends to infinity. This provides both a problem and an opportunity for computation of the nuclear potential. The opportunity arises because, in this limit, the nuclear potential can be computed as the static energy of the system in a fixed configuration in co-ordinate space (for analogous studies of heavy-quark systems on the lattice see Ref. [@NPLQCD]). This implies that the 3N potential (modulo issues of exchange diagrams, see below) obtained from our argument is local, being, e.g. a function of the Jacobi co-ordinates $r_{12}$ and $r_3$ (velocity-dependent forces arise at sub-leading orders in $1/N_c$, and lead to non-localities). The problem exists because the only measurable quantity in this infinitely-massive-nucleon limit is the total potential energy, and the large-$N_c$ analysis gives no information on the dependence of the force on $r_{12}$ and $r_3$—at least none beyond the statement that the function encoding that dependence has a size given by $N_c$ counting. Thus, since we only “measure" the total potential energy, and we cannot tell which pieces depend only on, say, $r_{12}$, we can make no [*a priori*]{} distinction between contributions to that energy from NN interactions, and contributions from 3NFs. The best we can do is to identify operator structures which occur in the 3N energy, and do not arise within the large-$N_c$ analysis of the NN potential of Ref. [@KM97].
One might be concerned that a 3NF derived from large-$N_c$ cannot be in accord with the meson-exchange picture used successfully for many years to derive NN and 3N forces. In Refs. [@Ba02; @BC02] Banerjee [*et al.*]{} and Belitsky and Cohen explored the relationship between this picture of the nuclear force and the large-$N_c$ analysis of Ref. [@KM97]. Initially it appeared that multi-meson-exchange graphs led to violations of the large-$N_c$ scaling of the NN potential: in particular to pieces of the NN potential that scaled with powers of $N_c$ larger than one. However, Ref. [@Co03] later explained this apparent discrepancy between the meson-exchange and large-$N_c$ pictures by pointing out that the potentials analyzed in Refs. [@Ba02; @BC02] were energy dependent, whereas almost all NN interactions used for phenomenological purposes are energy independent. Ref. [@Co03] concluded that an energy-independent NN potential could have $N_c$ scaling consistent with that derived in Ref. [@KM97], and so large-$N_c$ analysis is not inconsistent with a meson-exchange picture of nuclear forces for the NN case. An important point for a successful matching calculations is that the Hartree Hamiltonian Eq. (\[eq:Hartree\]) and the $SU(4)$ algebra Eq. (\[eq:algebra\]) implicitly assume the presence of the $\Delta$ resonance with $S=I=3/2$. In our discussion of the NN and NNN potentials, when taking matrix elements, we project $H$ to the nucleons-only piece of the Hilbert space. We have not performed a matching calculation to check the consistency with the meson-exchange picture for the 3N potential, but it would be a worthy subject for future study.
One might also wonder whether double counting will result if the 3N potential obtained from the large-$N_c$ analysis is used in a multi-nucleon Schrödinger equation. To address this issue we note that another assumption made in the derivation of the algebra Eq. (\[eq:algebra\]) was that meson energies are of order $\Lambda_{QCD}$. This implies that the energy of the intermediate nucleon state in the 3NF (see, e.g., Fig. \[fig:Vdir\]) must be order $\Lambda_{QCD}$ if an analysis based on this algebra is to prevail. Having states of this energy included in the computation of the nuclear potential is consistent with the insertion of the resulting nuclear force in the 3N Schrödinger equation (or, equivalently, a Faddeev equation) provided a momentum cutoff is employed there. If that momentum cutoff is above $\Lambda_{QCD}$, but below $\sqrt{N_c} \Lambda_{QCD}$, the intermediate nucleonic states with energies of order $\Lambda_{QCD}$ (i.e. momenta $\sim
\sqrt{M \Lambda_{QCD}}$) will not be accounted for by the iteration of the potential via the Schrödinger/Faddeev equation, and so should be included in the potential. The NN and 3N interactions derived here, and in Ref. [@KM97], can thus be inserted into the quantum-mechanical equation and used to compute the wave function of nuclear systems.
With the conceptual underpinning of a 3NF in large-$N_c$ QCD defined, and the circumstances under which it should be used in a Schrödinger equation for a multi-nucleon system clarified, we now turn back to the NN system, in order to explain how the corresponding analysis works in that, simpler, case.
The NN potential in the $1/N_c$ expansion: review {#sec-NN}
=================================================
Here we review the $1/N_c$ analysis of Kaplan and Manohar [@KM97] for the two-nucleon potential, setting up the notation that we will use later in Sec. \[sec-3N\] to analyze the three-nucleon force. In Ref. [@KM97] the large-$N_c$ expansion was used to analyze the object: $$U_{{\rm NN}}^A=(1-P_{12}) U \ ,$$ where $U$ is the sum of all direct diagrams, and $P_{ij}$ is the permutation operator that switches [*all*]{} quantum numbers of particles $i$ and $j$. In nuclear physics computations it is the operator $U$ which is inserted into the Schrödinger equation. The correct anti-symmetry properties of the nuclear state are then imposed by computing matrix elements only in partial waves which are allowed by the Fermi-Dirac statistics of the nucleons.
In order to discuss the momentum dependence of the potential we first define initial and final relative momenta: $${\bf p}={\bf p}_1 - {\bf p}_2 \ , \quad {\bf p}'={\bf p}'_1 - {\bf p}'_2 \ ,$$ where ${\bf p}_i ({\bf p}'_i)$ is the initial (final) momentum of the $i$-th nucleon. To simplify later analysis we also define time-reversal-odd (T-odd) and time-reversal-even (T-even) combinations of these: $${\bf p}_{\pm}={\bf p}' \pm {\bf p} \ .
\label{ppm}$$ Notice that ${\bf p}_{+}$ is T-odd and ${\bf p}_{-}$ is T-even, as initial and final states are also exchanged under time reversal. Both combinations, being vectors, are odd under parity. In $U$ only ${\bf p}_-$ enters at leading order in $N_c$ since the potential is local at this order in the $1/N_c$ expansion. Powers of ${\bf p}_+$ indicate the presence of non-locality. In a meson-exchange picture they arise due to the occurrence of relativistic corrections suppressed by $1/M_N$. Thus, each appearance of a power of ${\bf p}_+$ costs a power of $1/N_c$. Finally, energy conservation and the constraint that the external NN states in a diagram be on-shell results in $${\bf p}_+ \cdot {\bf p}_-=0 \ ,
\label{eq:onshellconstraintNN}$$ which allows to eliminate this momentum structure. In Ref. [@KM97] the potential $U$ was written as a sum of products of one-body operators, including the explicit factors of $1/N_c$ as shown in the Hartree Hamiltonian, Eq. (\[eq:Hartree\]). Isospin invariance of the interaction requires that all isospin indices are contracted.
In general, operators acting on the same nucleon with spatial or isospin indices contracted can be simplified. For instance, $G^{ia}G^{ia}$ can be reduced to the unit operator and a subleading contribution using $$G^{ia}G^{ia} = \frac{3}{16} N_c (N_c+4) \mathds{1} - \frac{1}{4} I^a I^a - \frac{1}{4} S^i
S^i \ ,
\label{eq:Casimir}$$ which is obtained from the quadratic $SU(4)$ Casimir evaluated on the completely symmetric representation $S_{N_c}$. If the spatial indices are not contracted we have the more general identity $$G^{ia} G^{ja} = \frac{1}{16} N_c (N_c+4) \delta^{ij} \mathds{1} - \frac{1}{4} \delta^{ij} S^2 +
\frac{1}{4} S^i S^j + \frac{i}{4} \epsilon^{ijk} S^k.
\label{eq:GGred}$$ The complete set of operator reduction rules can be found in [@Dashen:1994qi].
It is thus sufficient to consider structures where the contracted indices are carried by operators acting on different nucleons. For instance, the leading order of the angular momentum zero ($L=0$) component of the potential is obtained from $$\begin{aligned}
U_{L=0}^{N_c} &\subset&
N_c \sum_{{n=0}}^{N_c} u_n({\bf p}_-^2) (N_c^{-2} G_1^{ia} G_2^{ia})^{n} \ ,
\label{eq:NNLOGs}\end{aligned}$$ where $u_n({\bf p}_-^2)$ are arbitrary scalar functions of ${\bf p}_-^2$ that scale like ${\cal O}(N_c^0)$. This yields two strings of $G$’s, one on each of the two nucleons, with no contracted indices amongst the $G$’s which act on an individual nucleon. Each such string of $G$’s can, nevertheless, be reduced, because the matrix element of a general $m$-quark operator between single-baryon states scales as [@KS96; @KM97; @Dashen:1994qi] $$\begin{aligned}
\langle {\rm B}_1 |N_c^{-m} O_m | {\rm B}_1 \rangle = \frac{1}{N_c^{|I-S|}} \ .
\label{Oscaling}\end{aligned}$$ Therefore the dominant parts in the operator resulting from each string of $G$’s have $I=S$. If, in addition, we restrict ourselves to the case that the baryon is a nucleon only $(I,S)=(0,0), (1,1)$ contribute. But those $I=S=0$ and $I=S=1$ operators can, via the Wigner-Eckart theorem, be replaced by the ${\cal O}(N_c^0)$ one-body operators $\mathds{1}$ and $N_c^{-1} G^{ia}$, up to a proportionality constant that ultimately gets absorbed in the undetermined functions of momenta that appear in the large-$N_c$ NN potential. Thus, on a single-nucleon state, each string of $G$’s with uncontracted indices yields a matrix element that can be written: $$\begin{aligned}
<{\rm N} | \underbrace{G G ... G } _r|{\rm N} > = N_c^r < \mathds{1} > + N_c^{r-1} <G> + {\cal O}
(N_c^{r-2}),
\label{eq:manyG}\end{aligned}$$ where the spatial and isospin indices on the RHS of Eq, (\[eq:manyG\]) are carried by Kronecker $\delta$’s and the completely antisymmetric tensor $\epsilon$. For an example see the Appendix, in particular Eq. (\[twoGred\]).
Eqs. (\[eq:manyG\]) and (\[eq:GGred\]) show that it is enough to consider the one-quark operators $\mathds{1}$ and $N_c^{-1} G_{ia}$ acting within each nucleon to construct the leading-order spin-flavor structures. With this simple rule one obtains correctly the explicit $1/N_c$ suppression factors contained in the Hartree expression, Eq. (\[eq:Hartree\]), for the NN interaction.
The leading-order spin-flavor structures are thus $\mathds{1}_1 \mathds{1}_2$ and $G^{ia}_1 G^{ja}_2 $. The next step is to project out the different spin components of the leading-order $G_1 G_2$ tensor, namely $$\label{LOstrucs}
\quad G_1^{ia} G_2^{ia}, \quad \epsilon^{ijk} G_1^{ja} G_2^{ka}, \quad \left[G_1^{ia} G_2^{ja}\right]_2 \ ,$$ where the first two correspond to $S=0$ and $1$ respectively, and $$\left[G_1^{ia} G_2^{ja}\right]_2 \equiv G_1^{ia}
G_2^{ja} + G_1^{ja} G_2^{ia} - \frac23 \delta^{ij} G^{ka}_1 G^{ka}_2
\label{eq:spintensor}$$ is the $S=2$ component. The final step is the reduction of the operator $G$ to $\sigma^i \tau^a$ when restricted to the nucleon subspace. Table \[table-NNops\] shows a complete set of independent spin-flavor structures in the NN subspace, together with their $1/N_c$ scalings, spin content and time-reversal properties.
Each of these spin-flavor structures must then be combined with tensors formed out of the momenta ${\bf p}_-, {\bf p}_+$ to form a T-even, P-even, rotationally invariant operator. In particular, the $S=2$ structure (\[eq:spintensor\]) must be contracted with a spatial tensor of rank two. Since at LO we have a local NN potential the only possible LO tensor is ${\bf p}_-^i {\bf p}_-^j$. Meanwhile, the second ($S=1$) spin-flavor structure must be contracted with a three-vector. Parity invariance suggests ${\bf p}_- \times {\bf p}_+$ is the only possible candidate. However, ${\bf p}_- \times {\bf p}_+$ is odd under time reversal. And the constraint (\[eq:onshellconstraintNN\]) means we cannot multiply by powers of the T-odd rotational scalar ${\bf p}_+ \cdot {\bf p}_-$—at least not on-shell. Thus our $S=1$ spin-flavor structure cannot be multiplied by any combination of three-vectors that results in an overall P-even, T-even object. The operator $\epsilon^{ijk} G_1^{ja}
G_2^{ka}$ will therefore not appear in the parity-conserving, time-reversal-non-violating NN force [@Wolfenstein]. Finally, the first structure in Eq. (\[LOstrucs\]) and the unit operator are the two leading-order $S=0$, $L=0$ operators.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$O$ Order $O_{\tau \tau}$ Order $S$ T
--------------------------- ------------- ------------------------------------------------- ------------ -------- -------------------------------------------------------------------------------------------------------------------
$ \mathds{1} $ 1 $\tau_1 \cdot \tau_2$ $1/N_c^2$ 0 +
$\sigma_1 \cdot \sigma_2$ $1/N_c^2$ $\sigma_1 \cdot \sigma_2 \ \tau_1 \cdot \tau_2$ 1 0 +
$\sigma^i_1$ $1/N_c$ $\sigma^i_1 \ \tau_1 \cdot \tau_2$ $1/N_c$ 1 **[–]{}\
$\sigma^i_2$ & $1/N_c$ & $\sigma^i_2 \ \tau_1 \cdot \tau_2$ & $1/N_c$ & 1 & **[–]{}\
$(\sigma_1 \times \sigma_2)^k $ & $1/N_c^2$ & ($\sigma_1 \times \sigma_2)^k \ \tau_1 \cdot \tau_2$ & $1$ & 1 & +\
$[\sigma^i_1 \sigma^j_2]_{_2} $ & $1/N_c^2$ & $[\sigma^i_1 \sigma^j_2]_{_2} \ \tau_1 \cdot \tau_2$ & $1$ & 2 & +\
****
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Spin-flavor structures for the two-nucleon potential. The $(\sigma_1 \times \sigma_2)$ structure arises in the large-$N_c$ analysis, but its appearance in $U$ is precluded by permutation symmetry.[]{data-label="table-NNops"}
The rotational scalars formed in this way may always be multiplied by an arbitrary scalar function of ${\bf p}_-^2$. Therefore, to leading order $$\label{eq:orderNc}
U^{N_c} = N_c \left( \ U^1_{S}({\bf p}_-^2) \mathds{1} + U^2_{S}({\bf p}_-^2) \ \sigma_1 \cdot \sigma_2 \ \tau_1 \cdot \tau_2
+ U^1_{D}({\bf p}_-^2) \ [{\bf p}_- {\bf p}_-]_{_2} \cdot [\sigma_1 \sigma_2]_{_2} \ \tau_1 \cdot \tau_2 \
\right) \ ,$$ with $$[A_i B_j ]_{_2} \equiv A_i B_j + A_j B_i - \frac23 \delta_{ij} A\cdot B
\label{eq:rank2tensor}$$ the $L=2$ component of the tensor $A_i B_j$ constructed out of two vector quantities, and $U_S^{1,2}({\bf p}_-^2)$, $U_D^1({\bf p}_-^2)$ arbitrary ${\cal O}(1)$ scalar functions of ${\bf p}_-^2$. As discussed above, there are no $S=1$ terms at leading order.
Sub-leading corrections are associated with $1/N_c$-suppressed operators. Such suppression may occur for two reasons. Firstly, NN operators involving $S$ and $I$, instead of $G$, will be reduced by factors of $1/N_c$, because of the $N_c$-scaling of the nucleonic matrix elements of these operators. The second source of $1/N_c$ suppression is the appearance in expressions of the momentum ${\bf p}_+$. Time reversal and parity conservation conspire so that the expansion is in $1/N_c^2$.
With these two results regarding $1/N_c$ suppression in hand, Kaplan and Manohar concluded that the following operators give contributions to the NN potential of ${\cal
O}(1/N_c)$ (see also Table \[table-NNops\]): $$\begin{aligned}
U^{1/N_c} &=& \delta^{(2)} U^{N_c} + N_c^{-1} \left( \ U^3_{S} \ {\bf p}_+^2 \ \mathds{1} +
\ U^4_{S} \ \sigma_1 \cdot \sigma_2
+ U^5_{S} \ \tau_1 \cdot \tau_2
+ U^6_{S} \ {\bf p}_+^2 \ \sigma_1 \cdot \sigma_2 \ \tau_1 \cdot \tau_2
\right. \nonumber \\
&&
\qquad \qquad \qquad \quad + \ U^1_{P} \ ({\bf p}_+ \times {\bf p}_-) \cdot ({\bf \sigma}_1 + {\bf \sigma}_2)
+ U^2_{P} \ ({\bf p}_+ \times {\bf p}_-) \cdot ({\bf \sigma}_1 + {\bf \sigma}_2) \tau_1 \cdot \tau_2 \nonumber \\
&& \left.
\qquad \qquad \qquad \quad + \ U^2_{D} \ [{\bf p}_- {\bf p}_-]_{_2} \cdot [\sigma_1\sigma_2]_{_2} \
+ U^3_{D} \ [{\bf p}_+ {\bf p}_+]_{_2} \cdot [\sigma_1 \sigma_2]_{_2} \ \tau_1 \cdot \tau_2 \
\right) \ .
\label{eq:order1overNc}\end{aligned}$$ At this order the leading-order operators appear again, as they can also be obtained by replacing one $G^{ia}/N_c$ by $ (S^i I^a)/N_c^2$ in the Hartree Hamiltonian. We denoted this contribution by $\delta^{(2)} U^{N_c} $ in the expression above. The spin-flavor structures that appear here (and in Table \[table-NNops\]) and momentum tensors with up to four momenta can also be read off from the results for the 3NF that will be presented later in Sec. \[sec-3N\], by eliminating the third nucleon and only keeping momentum structures that depend on ${\bf p}_\pm$. Here we only show the potential up to quadratic structures in momenta (modulo arbitrary functions of ${\bf p}_-$).
Comparing Table \[table-NNops\] with Eq. (\[eq:orderNc\]) and Eq. (\[eq:order1overNc\]) one can see that the spin-flavor structures proportional to $\sigma_1
\times \sigma_2$ are missing because, as discussed for the LO case, they need to be multiplied by a T-even, P-even, $L=1$ momentum structure, which cannot be constructed in the NN case.
However, there is an additional constraint from permutation symmetry [@Wolfenstein]. For example, $\sigma_1, \sigma_2$ only appear in the $\sigma_1 + \sigma_2$ combination. The $\sigma_1 - \sigma_2$ combination is excluded by permutation symmetry, as it is T-odd and parity-even and needs to be contracted with a vector built from $\mathbf{p}_+,
\mathbf{p}_-$, where $\mathbf{p}_\pm$ are both odd under exchange of the nucleons $1,2$. For instance, if we would start from the general structure $$\begin{aligned}
U(\mathbf{p}_-^2) (\mathbf{p}_+ \times \mathbf{p}_-) \sigma_{1} + U'(\mathbf{p}_-^2) (\mathbf{p}_+ \times \mathbf{p}_-) \sigma_{2} \ \end{aligned}$$ permutation symmetry imposes $U=U'=U_P^1$ so that only the symmetric spin-flavor structure $\sigma_1+\sigma_2$ appears in Eq. (\[eq:order1overNc\]). The $\sigma_1 \times \sigma_2$ structure can also be eliminated by permutation symmetry.
In summary, to leading order (${\cal O}(N_c)$) there are two structures with $L=0$ and one with $L=2$. To subleading order (${\cal O}(1/N_c)$) and up to two momenta, there are four structures with $L=0$, two with $L=1$ and two with $L=2$.
This translates into definite scaling predictions for the different parts of the NN potential, which in the usual form is given by $$\begin{aligned}
\label{VNNpot}
V_{NN}&=&V^0_C+V^0_{SS} \ S_1 \cdot S_2 + V^0_{LS} \ L\cdot S + V^0_T \ S_{12} + V^0_Q \
Q_{12} \nonumber \\
&& + (V^1_C+V^0_{SS} \ S_1 \cdot S_2 + V^1_{LS} \ L\cdot S + V^1_T \ S_{12} + V^1_Q \
Q_{12}) \ \tau_1 \cdot \tau_2 \ . \end{aligned}$$ Here $L$ is the angular momentum operator, which is T-odd and P-even and in our notation is replaced via the Wigner-Eckart theorem by the ${\bf p}_+ \times {\bf p}_-$ structure. The quadratic spin-orbit interaction $Q_{12}$ involves four momenta in our notation and we did not include it in Eq. (\[eq:order1overNc\]).
A comparison with “experiment" can be achieved by comparing with a successful phenomenological potential. This has been done in Ref. [@KM97] using the Nijmegen potential [@St94]. The $1/N_c$ scaling of the different structures in Eq. (\[VNNpot\]) translates into a hierarchy for the functions used to parametrize the Nijmegen potential, which is well satisfied by their numerical values, as discussed in detail in Ref. [@KM97].
Although in the two-nucleon case the operator structure of the interaction is simple enough to be obtained by explicit construction, as sketched above, at this point it is useful to discuss a more systematic way of counting the number of spin-flavor structures that can contribute, something that will prove very useful in the more involved three-nucleon case. The systematic classification can be done as follows.
The number of independent spin-flavor structures $O_{IS}$ of isospin $I$ and spin $S$ that can contribute to the matrix element $\langle {\rm NN} | O_{IS} | {\rm NN} \rangle $ can be obtained by considering the decomposition of $\overline{\mathbf{R}} \otimes
\mathbf{R'}$, with $\mathbf{R},\mathbf{R'}$ the irreducible representations of spin-flavor for the two nucleons, so that the matrix element is a scalar. To obtain the possible irreps $\mathbf{R}$ we decompose the tensor product of two-nucleon states, each nucleon transforming as the fundamental representation of $SU(4)$ $$\begin{aligned}
\stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}} &\equiv& \{ p \uparrow, p \downarrow, n \uparrow, n \downarrow \} \ .\end{aligned}$$ The two-nucleon states are obtained as the decomposition of the tensor product $\mathbf{4} \otimes \mathbf{4} = \mathbf{6} \oplus \mathbf{10}$. In terms of Young tableaux $$\begin{aligned}
\stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}} \otimes \stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}}
&=&
\stackrel{\mathbf{6}}{\raisebox{-0.5cm}{\yng(1,1)}} \oplus
\stackrel{\mathbf{10}}{\raisebox{-0.1cm}{\yng(2)}} \ . \end{aligned}$$ As states and operators are labeled by their isospin and spin transformation properties, we decompose $SU(4)$ irreps in $SU(2)_I \times SU(2)_S \subset SU(4)$, labeled by $(2 I +
1, 2 S + 1)$. The result is (only $SU(4)$ irreps are in boldface): $$\begin{aligned}
\stackrel{\mathbf{10}}{\raisebox{-0.1cm}{\yng(2)}} &=&
\left(\stackrel{3}{ \raisebox{-0.1cm}{\yng(2)}} \ , \ \stackrel{3}{\raisebox{-0.1cm}{\yng(2)}} \right) \oplus
\left( \stackrel{1}{ \raisebox{-0.5cm}{\yng(1,1)}} \ , \
\stackrel{1}{\raisebox{-0.5cm}{\yng(1,1)}} \right) \ , \\
\stackrel{\mathbf{6}}{\raisebox{-0.5cm}{\yng(1,1)}} &=&
\left(\stackrel{3}{ \raisebox{-0.1cm}{\yng(2)}} \ , \ \stackrel{1}{\raisebox{-0.5cm}{\yng(1,1)}} \right) \oplus
\left(\stackrel{1}{ \raisebox{-0.5cm}{\yng(1,1)}} \ , \
\stackrel{3}{\raisebox{-0.1cm}{\yng(2)}} \right) \ . \end{aligned}$$ With this result in hand we can determine the number, and type, of spin-flavor structures that occur in $O_{IS}$. We consider the decomposition of $\overline{\mathbf{R}} \otimes \mathbf{R'}$, with $\mathbf{R},\mathbf{R'}=\mathbf{6},\mathbf{10}$ into irreps of $SU(2)_I \times SU(2)_S$. We are interested in the pieces of the direct product that yield $I=0$ operators, which are: $$\begin{aligned}
\sum_\mathbf{R,R'} \overline{\mathbf{R}} \otimes \mathbf{R'} &\supset& 4 (0,0) \oplus 6 (0,3) \oplus 2 (0,5) + \dots,\end{aligned}$$ i.e. the direct product contains four independent isoscalar structures of $S=0$, six of $S=1$ and two of $S=2$. Their explicit forms are $\mathds{1}, \sigma_1 \cdot
\sigma_2$, $\sigma_1, \sigma_2, \sigma_1 \times \sigma_2$ and $[\sigma_1 \sigma_2]_{_2}$, each of which can be multiplied by any of the two isospin invariants $\mathds{1},\tau_1 \cdot \tau_2$. The resulting spin-flavor structures are collected in Table \[table-NNops\].
This finishes the review of the NN case. We proceed now to the construction of the 3N potential.
The 3N potential in the $1/N_c$ expansion {#sec-3N}
=========================================
In this Section we will extend the analysis that we reviewed for the NN potential to the case of the 3N potential. The sum of all 3N$\rightarrow$3N diagrams can be written in operator form as: $$V_{\rm 3N}^A=(1 + P_{12} P_{23} + P_{13} P_{23})(1-P_{23}) V \ .
\label{eq:DplusE}$$ The terms in parentheses in Eq. (\[eq:DplusE\]) thus generate the exchange diagrams necessitated by the identicality of the nucleons from the operator $V$, which itself is the sum of all direct diagrams (see, e.g., Fig. \[fig:Vdir\]), and is the object that enters the Schrödinger equation in nuclear-physics computations. We will classify the structures that contribute to $V$, and derive their scaling behaviour with $N_c$.
We do this by first discussing the momenta involved, and the possible momentum structures obtainable therefrom. We then derive the LO spin-flavor structures, and count all possible spin-flavor structures. We finish with the explicit construction of the operators, including the spatial part.
Momenta and momentum structures
-------------------------------
Throughout, we work in the 3N center-of-mass frame, where: $${\bf p}_1 + {\bf p}_2 + {\bf p}_3={\bf p}_1' + {\bf p}_2' + {\bf p}_3'=0 \ .$$ Any graph can then be expressed as a function of the Jacobi momenta ${\bf p}$ and ${\bf q}$ $${\bf p}={\bf p}_1 - {\bf p}_2 \ , \qquad {\bf q}={\bf p}_3 - \left({\bf p}_1 + {\bf
p}_2\right)/2 \ ,$$ so that the 3NF can be written as a function of four three-momenta ${\bf p}$, ${\bf p}'$, ${\bf q}$, and ${\bf q}'$.
Conservation of energy yields the constraint: $${\bf p}^2 + \frac43 {\bf q}^2={\bf p}^{\prime \, 2} + \frac{4}{3} {\bf q}^{\prime \, 2}.
\label{eq:onshell}$$ In terms of the momenta with well-defined properties under time reversal, ${\bf p}_{\pm}$ as in Eq. (\[ppm\]), and the analogous ${\bf q}_{\pm}={\bf q}' \pm {\bf q}$, the constraint (\[eq:onshell\]) becomes $${\bf p}_+ \cdot {\bf p}_-=-\frac{4}{3} {\bf q}_+ \cdot {\bf q}_- \ ,$$ which will allow us to eliminate ${\bf q}_+ \cdot {\bf q}_-$ in favor of ${\bf p}_+ \cdot
{\bf p}_-$. Analysis of the contributions to Fig. \[fig:Vdir\] shows that the presence of $ {\bf p}_+ , {\bf q}_+$ comes from relativistic corrections that introduce powers of $1/M_N$, so that each power of either $ {\bf p}_+$ or $ {\bf q}_+$ is associated with a supression factor of $1/N_c$. The LO momentum structures are ${\cal O}(N_c^0)$ and depend only on $ {\bf p}_-$ and $ {\bf q}_-$. They correspond to local potentials.
In fact, since all spin-flavor structures are built out of P-even objects, parity invariance of $V$ requires that the momentum structures appearing—both at leading and sub-leading orders in $1/N_c$— must contain an even number of momenta. In Table \[table-momenta\] we show the TP properties, $L$ content and order in $1/N_c$ of the 3N-system momentum tensors which contain up to two momenta. Time-reversal-odd momentum structures only appear at subleading orders, as they must include at least either a ${\bf
p}_+$ or a ${\bf q}_+$.
$
\begin{array}{c}
\includegraphics*[width=10cm]{nnntree.eps}
\end{array}
$
---------------------------------------------------------------------------------------------------------------------------------
T P $\Pi^{\rm{TP}}$ $L$ Order
---- --- ---------------------------------------------------------------------------------------- ----------------- ----- -------
+ **[–]{} & ${\bf p}_-$ & 1 & 1\
&+ & **[–]{} & ${\bf q}_-$ & 1 & 1\
& **[–]{} & **[–]{} & ${\bf p}_+$ & 1 & $1/N_c$\
&**[–]{} & **[–]{} & ${\bf q}_+$ & 1 & $1/N_c$\
& **[–]{} & + & ${\bf p}_{+} {\bf p}_{-}$ & 0,1,2 & $1/N_c$\
& **[–]{} & + & ${\bf q}_{+} {\bf q}_{-}$ & 0,1,2 & $1/N_c$\
& **[–]{} & + & ${\bf p}_{\pm} {\bf q}_{\mp}$ & 0,1,2 &$1/N_c$, $1/N_c$\
& + & + & ${\bf p}_{\pm} {\bf q}_{\pm}$ & 0,1,2 &$1/N_c^2$, $1$\
& + & + & ${\bf p}_{\pm} {\bf p}_{\pm}$ & 0,2 &$1/N_c^2$, $1$\
& + & + & $ \ \ \ {\bf q}_{\pm} {\bf q}_{\pm} \ \ \ $ & 0,2 & $1/N_c^2$, $1$ \
******************
---------------------------------------------------------------------------------------------------------------------------------
: Tensors $\Pi^{\rm{TP}}$ constructed from up to two 3N-system momenta, together with their $\rm{T}$ and $\rm{P}$ properties and their angular momentum ($L$) content. Note that in the NN system none of the tensors involving ${\bf q}_{\pm}$ are present. The $\pm$ signs in the subscripts are always to be read as correlated, so that the last four entries in the Table each contain two possible tensors. The last column shows the $1/N_c$ order at which the corresponding momentum structure appears. []{data-label="table-momenta"}
Leading spin-flavor structures {#sec-redrules}
------------------------------
As in the NN case, leading-order spin-flavor structures are obtained from products of an arbitrary number of $G$’s, with their indices contracted in order to get isoscalar operators of spin $S=0,1,2,3$, which are the only quantum numbers relevant for isospin conserving interactions in the 3N subspace. Any spatial indices associated with the spin tensor of rank-0,1,2,3 are then contracted with a momentum tensor of the same rank to form a singlet, so that the interaction is invariant under rotations.
For example, the leading order $S=0$ 3N structures are obtained from $$\begin{aligned}
V_{L=0}^{N_c} &\subset&
N_c \sum_{{n_{12},n_{13},n_{23}} \atop n_{123}} v_{{n_{12},n_{13}\dots}} (N_c^{-2}
G_1^{ia} G_2^{ia})^{n_{12}} (N_c^{-2} G_1^{jb} G_3^{jb})^{n_{13}} \nonumber \\
& & \qquad \qquad \qquad \qquad \quad \times (N_c^{-2} G_2^{kc} G_3^{kc})^{n_{23}}
(N_c^{-3} \epsilon^{lmr} \epsilon^{def} G_1^{ld} G_2^{me} G_3^{rf} )^{n_{123}} \nonumber \\
& & \qquad \qquad \qquad + \cdots
\label{eq:LOGs}\end{aligned}$$ where the dots stand for terms with more complex index contractions. A general structure has the form $O_\alpha O_\beta O_\gamma$, with the greek index indicating the nucleon on which a particular $O$ acts. As in the NN case, products like $G_1^{ia} G_1^{ja}$ where there is at least one index contracted between operators acting on the same nucleon are not included. The structures shown in Eq. (\[eq:LOGs\]) still seem hard to reduce, but this can be achieved after taking matrix elements in the NNN subspace using Eq. (\[Oscaling\]) and Eq. (\[eq:manyG\]), as we did in the NN case. The simple rule is again that, at leading order, an arbitrary product of $G$’s can be reduced to a sum of $I=S$ operators, which for the N subspace reduce just to the unit operator and one $G$. So, as in the NN force, the LO structures are found by considering one-quark operators $\mathds{1}$ and $N_c^{-1} G^{ia}$ acting on each nucleon. This gives the explicit $1/N_c$ suppression factors that come from the spin-flavor part. Then, within the N subspace we replace $G^{ia}$ by $\sigma^i \tau^a$. Bearing in mind that spin and isospin indices should be contracted with $\delta_{ij},\delta_{ab}$ or $\epsilon_{ijk},\epsilon_{abc}$ tensors one straightforwardly obtains the leading spin-flavor structures shown in Table \[table-leadingops\].
The isospin structures are the unit operator, the three scalar products $\tau_\alpha \cdot \tau_\beta$ and a new structure that was not present in the NN case, the triple product $(\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma$. It is important to notice that the triple product of $\tau$ is time-reversal odd, as under time reversal $(\tau^1, \tau^2, \tau^3) \rightarrow (\tau^1, -\tau^2, \tau^3)$. This is in contrast to $(\sigma^1, \sigma^2, \sigma^3) \rightarrow (-\sigma^1, -\sigma^2, -\sigma^3)$. The different transformation properties of the spin and isospin operators under time reversal just reflect the fact that under time reversal spins get flipped, while protons and neutrons retain their identity and are not exchanged.
Spin content LO quark operator $\sigma\tau$-projection Multiplicity
---------------- -- --- ------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------- --------------
$S=0$ $\mathds{1}$ $\mathds{1}$ 1
$ N_c^{-2} \ G_\alpha^{ia} G_\beta^{ia} $ $ \sigma_\alpha \cdot \sigma_\beta \ \tau_\alpha \cdot \tau_\beta $ 3
$ N_c^{-3} \ \epsilon^{ijk}\epsilon^{abc} G^{ia}_\alpha G^{jb}_\beta G^{kc}_\gamma $ $(\sigma_\alpha \times\sigma_\beta ) \cdot \sigma_\gamma (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma$ 1
\[.1cm\] $S=1$ $ N_c^{-2} \ \epsilon^{ijk} G_\alpha^{ia} G_\beta^{ja} $ $(\sigma_\alpha \times \sigma_\beta) \ \tau_\alpha \cdot \tau_\beta $ 3
$ N_c^{-3} \ \epsilon^{abc} G_\alpha^{ia} G_\beta^{ib} G_\gamma^{kc} $ $ (\sigma_\alpha \cdot \sigma_\beta) \sigma_\gamma \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ 3
\[.1cm\] $S=2$ $ N_c^{-2} \ [G_\alpha^{ia} G_\beta^{ja}]_{_2} $ $ [\sigma_\alpha \sigma_\beta]_{_2} \ \tau_\alpha \cdot \tau_\beta $ 3
$ \ \ \ N_c^{-3} \ \epsilon^{abc} \Big[ (G_\alpha^{ia} G_\beta^{jb}\epsilon^{ijl}) G_\gamma^{kc}\Big]_{_2} \ \ \ $ $\ \ \ \Big[ (\sigma_\alpha \times \sigma_\beta ) \sigma_\gamma \Big]_{_2} (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma \ \ \ $ 2
\[.1cm\] $S=3$ $ N_c^{-3} \ \epsilon^{abc} \Big[ G_\alpha^{ia} G_\beta^{jb} G_\gamma^{kc}\Big]_{_3} $ $\Big[ \sigma_\alpha \sigma_\beta \sigma_\gamma \Big]_{_3} (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ 1
\[.1cm\] 17
: Leading-order quark operators and their projection on nucleon spin-isospin structures. Structures are listed according to their spin content within the nucleonic space. $\alpha, \beta, \gamma$ are a permutation of $123$, designating on which nucleon the spin and isospin operators act. The multiplicity indicates how many independent structures are generated by these permutations. The 17 leading-order structures are all parity even and time-reversal even.[]{data-label="table-leadingops"}
The last column of Table \[table-leadingops\] shows the multiplicity of each structure, obtained by running $\alpha,\beta,\gamma$ over all the permutations of $1,2,3$. For the spin-2 structures a non-trivial constraint reduces the multiplicity of the $ \Big[(\sigma_\alpha \times \sigma_\beta) \sigma_\gamma\Big]_{_2}
= \Big\{\Big[(\sigma_1 \times \sigma_2) \sigma_3\Big]_{_2},
\Big[(\sigma_1 \times \sigma_3) \sigma_2\Big]_{_2},
\Big[(\sigma_2 \times \sigma_3) \sigma_1\Big]_{_2} \Big\} $ operator structure from three to two, because Eq. (\[eq:triple\]), projected onto a symmetric and traceless rank-two tensor, gives: $$\begin{aligned}
[(A \times B) C]_{_2} + [(B \times C) A]_{_2} + [(C \times A) B]_{_2} = 0 \ .
\label{eq:triple2}\end{aligned}$$ There are 17 independent structures at leading order. They are all time-reversal even. Further details are given below, with the leading potential exhibited in Eqs. (\[LOL0\]), (\[LOL1\]), (\[LOL2\]), and (\[LOL3\]).
Counting all the spin-flavor structures
---------------------------------------
However, in order to enumerate all sub-leading structures we find it important to first generalize our counting of spin-flavor structures using $SU(4)$ irreps from the NN to the NNN case. In this way we determine the number of spin-flavor structures we expect to find once we consider all orders in $N_c$.
The number of NNN states is given by $\mathbf{4} \otimes \mathbf{4} \otimes \mathbf{4} =
\mathbf{4} \oplus \mathbf{20'} \oplus\mathbf{20'} \oplus \mathbf{20}$: $$\begin{aligned}
\stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}} \otimes \stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}} \otimes
\stackrel{\mathbf{4}}{\raisebox{-0.1cm}{\yng(1)}}
&=&
\stackrel{\mathbf{4}}{\raisebox{-0.8cm}{\yng(1,1,1)}} \oplus
\stackrel{\mathbf{20'}}{\raisebox{-0.5cm}{\yng(2,1)}} \oplus
\stackrel{\mathbf{20'}}{\raisebox{-0.5cm}{\yng(2,1)}} \oplus
\stackrel{\mathbf{20}}{\raisebox{-0.1cm}{\yng(3)}} \ . \end{aligned}$$ Decomposing these $SU(4)$ irreps into $SU(2)_I \times SU(2)_S \subset SU(4)$ we have (as above, the $SU(2)_I \times SU(2)_S$ irreps are labeled by $(2 I +
1, 2 S + 1)$ and only $SU(4)$ irreps are in boldface): $$\begin{aligned}
\stackrel{\mathbf{20}}{\raisebox{-0.1cm}{\yng(3)}} &=&
\left(\stackrel{4}{ \raisebox{-0.1cm}{\yng(3)}} \ , \ \stackrel{4}{\raisebox{-0.1cm}{\yng(3)}} \right) \oplus
\left( \stackrel{2}{ \raisebox{-0.5cm}{\yng(2,1)}} \ , \
\stackrel{2}{\raisebox{-0.5cm}{\yng(2,1)}} \right) \ , \\
\stackrel{\mathbf{20'}}{\raisebox{-0.5cm}{\yng(2,1)}} &=&
\left(\stackrel{4}{ \raisebox{-0.1cm}{\yng(3)}} \ , \ \stackrel{2}{\raisebox{-0.5cm}{\yng(2,1)}} \right) \oplus
\left(\stackrel{2}{ \raisebox{-0.5cm}{\yng(2,1)}} \ , \ \stackrel{4}{\raisebox{-0.1cm}{\yng(3)}} \right) \oplus
\left( \stackrel{2}{ \raisebox{-0.5cm}{\yng(2,1)}} \ , \
\stackrel{2}{\raisebox{-0.5cm}{\yng(2,1)}} \right) \ , \\
\stackrel{\mathbf{4}}{\raisebox{-0.8cm}{\yng(1,1,1)}} &=&
\left(\stackrel{2}{ \raisebox{-0.5cm}{\yng(2,1)}} \ , \
\stackrel{2}{\raisebox{-0.5cm}{\yng(2,1)}} \right) \ . \end{aligned}$$
The number of independent operators $O_{IS}$ of isospin $I$ and spin $S$ that can contribute to $\langle {\rm NNN} | O_{IS} | {\rm NNN} \rangle $ can now be obtained by considering the decomposition of $\overline{\mathbf{R}} \otimes \mathbf{R'}$, with $\mathbf{R},\mathbf{R'}=\mathbf{4},\mathbf{20'},\mathbf{20}$ into irreps of $SU(2)_I \times SU(2)_S$. Notice that $\mathbf{20'}$ has to be considered twice. We are interested in $I=0$ spin-flavor structures, and for those we find: $$\begin{aligned}
\sum_{R,R'} \overline{\mathbf{R}} \otimes \mathbf{R'} &\supset& 25 (0,0) \oplus 45 (0,3) \oplus 25 (0,5) \oplus 5 (0,7) + \dots\end{aligned}$$ So, there are 25 independent isoscalar structures of $S=0$, 45 of $S=1$, 25 of $S=2$ and 5 of $S=3$. There are thus 100 spin-flavor structures in total, 50 T-even and 50 T-odd. This provides an important check for the explicit construction of operators that we will describe in the next subsection. The results of that construction are shown in Tables \[table-3NL0\]–\[table-3NL3\], and we indeed find a total of 100 structures.
Explicit construction of the three-nucleon operators
----------------------------------------------------
In the following subsubsections we write down, successively, the 3N potential-energy operators which are built out of $S=0,1,2,3$ spin-flavor structures. Since we seek rotational scalars, each spin-flavor structure is coupled to a momentum structure of equal rank. We therefore use the terms “$L=a$" and “$S=a$" interchangeably when referring to the operators that appear in $V$. We present explicit expressions up to ${\cal O}(1/N_c)$.
### $L=S=0$
A complete set of spin-flavor structures in the $S=0$ sector is given by the $S^{(r)}_\xi$ listed in Table \[table-3NL0\].
Operator Order T Multiplicity
---------------------------------------------------------------------------------------------------------------------------- ----------- ----- --------------
$S^{(0)}_1 = \mathds{1} $ $1$ + 1
$S^{(0)}_{2-4} = \sigma_\alpha \cdot \sigma_\beta \ \tau_\alpha \cdot \tau_\beta $ $1$ + 3
$S^{(0)}_5 = (\sigma_\alpha \times \sigma_\beta) \cdot \sigma_\gamma \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1$ $+$ 1
$S^{(1)}_{1-3} = \sigma_\alpha \cdot \sigma_\beta \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1/N_c$ $-$ 3
$S^{(1)}_{4-6} = (\sigma_\alpha \times \sigma_\beta) \cdot \sigma_\gamma \ \tau_\alpha \cdot \tau_\beta $ $1/N_c$ $-$ 3
$S^{(2)}_{1-3} = \tau_\alpha \cdot \tau_\beta $ $1/N_c^2$ + 3
$S^{(2)}_{4-6} = \sigma_\alpha \cdot \sigma_\beta $ $1/N_c^2$ + 3
$S^{(2)}_{7-12} = \sigma_\alpha \cdot \sigma_\beta \ \tau_\beta \cdot \tau_\gamma $ $1/N_c^2$ + 6
$S^{(3)}_1 = (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1/N_c^3$ $-$ 1
$S^{(3)}_2 = (\sigma_\alpha \times \sigma_\beta) \cdot \sigma_\gamma $ $1/N_c^3$ $-$ 1
25
: $S=0$ spin-isospin structures. The order given in the second column is relative to $N_c$. The third column indicates the behaviour of each structure structure under time reversal, namely, even (+) or odd (–), and its multiplicity is given in the last column. Here $\alpha \neq \beta \neq \gamma$ label the nucleon on which each of the spin and isospin operators act. In the last line we give the total number of independent structures, obtained as the sum of $M_0^{(0)}=5, M_0^{(1)}=6, M_0^{(2)}=12$ and $M_0^{(3)}=2$.[]{data-label="table-3NL0"}
The superscript $(r)$ indicates the [*relative order*]{} in $N_c$ at which the spin-flavor structure appears for the first time (i.e. its lowest order). This corresponds to $r=s+t$ in the Hartree Hamiltonian, Eq. (\[eq:Hartree\]) and essentially counts the number of subleading operators $S$ and $I$ that contribute to the structure. The resulting contribution to the 3N force is obtained after taking into account the overall factor of $N_c$ in Eq. (\[eq:Hartree\]) and the momentum structure that combines with each spin-flavor structure to give a rotational scalar, time-reversal-even and parity-even Hamiltonian. Each occurrence of a time-reversal-odd momentum ${\bf p}_+,{\bf q}_+$ costs an additional power of $1/N_c$.
The time-reversal-even spin-flavor structures at order $N_c^0$ and order $1/N_c^2$ can be straightforwardly incorporated into the potential. They only need to be multiplied by arbitrary scalar functions of the vectors ${\bf p}_-$ and ${\bf
q}_-$. We denote the functions, which are all of ${\cal
O} (N_c^0)$, $V_X^m({\bf p}_-^2,
{\bf q}_-^2, {\bf p}_- \cdot {\bf q}_-)$, where $X$ runs over the different spin-flavor structures and $m$ enumerates functions $V$ corresponding to different momentum structures. Beyond the statement that they are ${\cal O}(N_c^0)$, the large-$N_c$ expansion sheds no light on the behavior of these functions.
With this notation the ${\cal O} (N_c)$ (leading-order) potential is: $$\begin{aligned}
\label{LOL0}
V_{ L=0}^{N_c}= N_c \sum_{\xi=1}^{M_{0}^{(0)}} V^1_{S_\xi} ({\bf p}_-^2,{\bf q}_-^2,{\bf
p}_- \cdot {\bf q}_-) S_\xi^{(0)} \ , \end{aligned}$$ with $M_0^{(0)}=5$ the number of independent, leading-order, $S=0$ spin-flavor structures (see Table \[table-3NL0\] or Table \[table-leadingops\]). In fact, their contribution once the spatial part of the 3N state is taken into account is not completely independent, since the functions $V^{1}_{S_2}$,$V^{1}_{S_3}$ and $V^{1}_{S_4}$ are related to one another by permutation symmetry, i.e. the requirement that the total force be symmetric under permutations of all particle labels. This constraint in the 3N case is, however, more complicated than in the NN case, and there seems to be no obvious simplification due to permutation symmetry.
There are thus five spin-flavor structures that contribute at leading order in $N_c$ to the $L=0$ part of the 3N potential: the identity, a $ \sigma_\alpha \cdot \sigma_\beta \
\tau_\alpha \cdot \tau_\beta $ structure, where one of the three nucleons is not involved, and the structure $(\sigma_\alpha \times \sigma_\beta) \cdot \sigma_\gamma \ (\tau_\alpha
\times \tau_\beta) \cdot \tau_\gamma$. Of these, the first two already occur in the NN potential, and as already discussed above, without knowledge of the ${\bf q}_-$ dependence in $V$, we cannot separate their appearance here from the fact that they contribute to the energy of the NN pairs in the 3N system.
We now turn our attention to sub-leading corrections to the $L=0$ 3N force. It at first appears that there are spin-flavor structures which generate contributions of relative order $1/N_c$. But in fact the resulting structures are all time-reversal odd. In consequence they must be multiplied by a time-reversal-odd dot product in order to appear in the $L=0$ component of the 3N potential. In contrast to the NN case such dot products exist in this system, e.g. ${\bf p}_+ \cdot {\bf q}_-$. But all of the T-odd ones involve either ${\bf p}_+$ or ${\bf q}_+$. Thus the first sub-leading contribution is suppressed by two powers of $1/N_c^2$ relative to leading: one because of the matrix elements of the spin-flavor structures which appear, and one because of the necessity for a $1/M_N$ factor in order to generate some non-locality and introduce ${\bf p}_+$ or ${\bf q}_+$.
At relative order $1/N_c^2$ we also have the 12 structures $S_\xi^{(2)}$ shown in Table \[table-3NL0\]. In addition, the leading structures $S_\xi^{(0)}$ can reappear, now multiplied by two of the $1/N_c$ suppressed dot products, or by one $1/N_c^2$ suppressed dot product of momenta. Using the energy-conservation and on-shell condition, Eq. (\[eq:onshell\]), at ${\cal O}(1/N_c)$ we find three momentum structures of ${\cal O}(1/N_c)$ and three structures of ${\cal O}(1/N_c^2)$, all of which involve two momenta. With four momenta there are six new structures of order ${\cal O}(1/N_c^2)$.
Lastly, we observe that operators from the LO potential occur again at this order, as they can arise via the replacement of one $G^{ia}/N_c$ by $S^i I^a/N_c^2$ in the Hartree Hamiltonian, as already discussed for the NN potential. We denote this contribution by $\delta^{(2)} V_{L=0}^{N_c} $ which stands for $$\begin{aligned}
\label{LOexpL0}
\delta^{(2)} V_{ L=0}^{N_c}= N_c^{-1}
\sum_{\xi=1}^{M_{0}^{(0)}} V^1_{S_\xi,1/N_c^2} ({\bf p}_-^2,{\bf q}_-^2,{\bf p}_- \cdot
{\bf q}_-) S_\xi^{(0)} \end{aligned}$$ where the explicit $N_c$ factors ensure that the $V^1_{S_\xi,1/N_c^2} ({\bf p}_-^2,{\bf q}_-^2,{\bf p}_- \cdot {\bf q}_-)$ are of order ${\cal O}(N_c^0)$. Eqs. (\[LOL0\]) and (\[LOexpL0\]) can be combined, with the effect that the functions $V_{S_\xi}^1$ each have their own expansion in $1/N_c^2$.
The full ${\cal O}(1/N_c)$ piece of the $L=0$ 3N potential is then: $$\begin{aligned}
V_{L=0}^{1/N_c}&=& \delta^{(2)} V_{L=0}^{N_c} + N_c \sum_{\xi=1}^{M_0^{(2)}} V^{2}_{S_\xi} S_\xi^{(2)} \nonumber\\
&&+ N_c \sum_{\xi=1}^{M_0^{(1)}} \left( V^{3}_{S_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{4,5}_{S_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right) S_\xi^{(1)} \nonumber\\
&&+ N_c \sum_{\xi=1}^{M_0^{(0)}} \left\{ \ V^{6}_{S_\xi} {\bf p}_+^2
+ V^{7}_{S_\xi} {\bf p}_+ \cdot {\bf q}_+
+ V^{8}_{S_\xi} {\bf q}_+^2 \right. \nonumber \\
&&\qquad \ \qquad + \left. V^{9}_{S_\xi} ({\bf p}_+ \cdot {\bf p}_-)^2
+ V^{10,11}_{S_\xi} ({\bf p}_\pm \cdot {\bf q}_\mp)^2 \right. \nonumber \\
&&\qquad \ \qquad + \ V^{12,13}_{S_\xi} ({\bf p}_+ \cdot {\bf p}_-)({\bf p}_\pm \cdot {\bf q}_\mp)
+ \left. V^{14}_{S_\xi} ({\bf p}_+ \cdot {\bf q}_-)({\bf q}_+ \cdot {\bf p}_-) \ \right\} S_\xi^{(0)} \ .
\label{NLOL0}\end{aligned}$$ Here $M_0^{(2)}=12$ and $M_0^{(1)}=6$ are obtained summing over the multiplicities shown in Table \[table-3NL0\]. Note that here any momentum structure that involves more powers of the ${\cal O}(N_c^0)$ momenta ${\bf p}_-$, ${\bf q}_-$, is absorbed in the ${\cal O}(N_c^0)$ scalar functions $V^{2-14}_{S_\xi} ({\bf p}_-^2,{\bf q}_-^2,{\bf p}_-
\cdot {\bf q}_-)$.
There is no correction at order $1/N_c^3$, due to the T-odd nature of the operators $S^{(3)}_\xi$ listed above, and the restrictions of parity and time-reversal invariance regarding the vector dot products which can be considered. The correction of order $1/N_c^4$ can be constructed in analogy to the results for $1/N_c^2$ given in Eq. (\[NLOL0\]). For example, the terms involving the $M_0^{(3)}=2$ operators $S_\xi^{(3)}$ are $$N_c \sum_{\xi=1}^{M_0^{(3)}} \left( V^{15}_{S_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{16,17}_{S_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
S_\xi^{(3)} \ . \\$$ Meanwhile there are many other terms at this order involving $\delta^{(4)} V_{L=0}^{N_c},
\delta^{(2)} V_{L=0}^{1/N_c} $ and various combinations of dot products of momenta and the operators $S_\xi^{(2)}$, $S_\xi^{(1)}$ and $S_\xi^{(0)}$. These are too numerous to list here, but it is clear that the expansion of the potential is in $1/N_c^2$, as was also the case for the NN potential.
Before proceeding to the construction of the $L=1,2,3$ components of the potential it is important to explain why there are no cross products in the momentum structures of the $L=0$ potential. Triple products like $({\bf p}_-\times{\bf p}_+)\cdot{\bf q}_-$ are parity odd and should appear together with another triple product giving a structure of six momenta, which can be written in terms of scalar products alone. Quadruple products like $({\bf p}_-\times{\bf p}_+)\cdot({\bf p}_-\times{\bf p}_+)$ can also be reduced to the ones already present, as they involve two contracted epsilon tensors. The identities used for this purpose can be found in Appendix \[ap-reduction\], see, specifically, Eq. (\[eq:epseps\]).
In summary, for $L=0$ there are five operators at LO, given in Eq. (\[LOL0\]), which are built out of T-even spin-isospin structures. At sub-leading order there are 57 additional operators which are built out of T-even spin-isospin structures, as well as 18 operators involving T-odd structures. These are listed in Eq. (\[NLOL0\]). Up to ${\cal
O}(1/N_c)$ we presented the explicit expressions for this total of 80 operators, out of which 17 only depend on ${\bf p}_-, {\bf q}_-$ and correspond to a local potential.
Operator Order T Multiplicity
------------------------------------------------------------------------------------------------------------------------- ----------- ----- --------------
$P^{(0)}_{1-3} = (\sigma_\alpha \times \sigma_\beta) \ \tau_\alpha \cdot \tau_\beta $ $1$ + 3
$P^{(0)}_{4-6} = (\sigma_\alpha \cdot \sigma_\beta) \sigma_\gamma \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1$ $+$ 3
$P^{(1)}_{1-3} = \sigma_\alpha $ $1/N_c$ $-$ 3
$P^{(1)}_{4-9} = \sigma_\alpha \ \tau_\alpha \cdot \tau_\beta $ $1/N_c$ $-$ 6
$P^{(1)}_{10-12} = (\sigma_\alpha \times \sigma_\beta) \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1/N_c$ $-$ 3
$P^{(1)}_{13-15} = (\sigma_\alpha \cdot \sigma_\beta) \sigma_\gamma \ \tau_\alpha \cdot \tau_\beta $ $1/N_c$ $-$ 3
$P^{(1)}_{16-21} = (\sigma_\alpha \cdot \sigma_\beta) \sigma_\gamma \ \tau_\beta \cdot \tau_\gamma $ $1/N_c$ $-$ 6
$P^{(2)}_{1-3} = \sigma_\alpha \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1/N_c^2$ + 3
$P^{(2)}_{4-6} = (\sigma_\alpha \times \sigma_\beta) $ $1/N_c^2$ + 3
$P^{(2)}_{7-12} = (\sigma_\alpha \times \sigma_\beta) \ \tau_\beta \cdot \tau_\gamma $ $1/N_c^2$ + 6
$P^{(3)}_{1-3} = \sigma_\alpha \ \tau_\beta \cdot \tau_\gamma $ $1/N_c^3$ $-$ 3
$P^{(3)}_{4-6} = (\sigma_\alpha \cdot \sigma_\beta) \sigma_\gamma $ $1/N_c^3$ $-$ 3
45
: $S=1$ spin-isospin structures, as in Table \[table-3NL0\]. In the last line we give the total number of independent structures, obtained as the sum of $M_1^{(0)}=6, M_1^{(1)}=21, M_1^{(2)}=12$ and $M_1^{(3)}=6$.[]{data-label="table-3NL1"}
Operator Order T Multiplicity
----------------------------------------------------------------------------------------------------------------------------------------- ----------- ----- --------------
$D^{(0)}_{1-3} = [\sigma_\alpha \sigma_\beta]_{_2} \ \tau_\alpha \cdot \tau_\beta $ $1$ + 3
$D^{(0)}_{4,5} = \Big[(\sigma_\alpha \times \sigma_\beta) \sigma_\gamma\Big]_{_2} \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1$ $+$ 2
\[.1cm\] $D^{(1)}_{1-3} = [\sigma_\alpha \sigma_\beta]_{_2} \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $ $1/N_c $ $-$ 3
$D^{(1)}_{4,5} = \Big[(\sigma_\alpha \times \sigma_\beta) \sigma_\gamma\Big]_{_2} \ \tau_\alpha \cdot \tau_\beta $ $1/N_c$ $-$ 2
$D^{(1)}_{6-9} = \Big[(\sigma_\alpha \times \sigma_\beta) \sigma_\gamma\Big]_{_2} \ \tau_\beta \cdot \tau_\gamma $ $1/N_c$ $-$ 4
\[.1cm\] $D^{(2)}_{1-3} = [\sigma_\alpha \sigma_\beta]_{_2} $ $1/N_c^2$ + 3
$D^{(2)}_{4-9} = [\sigma_\alpha \sigma_\beta]_{_2} \ \tau_\beta \cdot \tau_\gamma $ $1/N_c^2$ + 6
$D^{(3)}_{1,2} = \Big[(\sigma_\alpha \times \sigma_\beta) \sigma_\gamma\Big]_{_2} $ $1/N_c^3$ $-$ 2
\[.1cm\] 25
: $S=2$ spin-isospin structures, as in Table \[table-3NL0\]. In the last line we give the total number of independent structures, obtained as the sum of $M_2^{(0)}=5, M_2^{(1)}=9, M_2^{(2)}=9$ and $M_2^{(3)}=2$.[]{data-label="table-3NL2"}
---------------------------------------------------------------------------------------------------------------------------------------------------
Operator Order T Multiplicity
------------------------------------------------------------------------------------------------------------------ ----------- ----- --------------
$F^{(0)}_1 =[\sigma_\alpha \sigma_\beta \sigma_\gamma]_{_3} \ (\tau_\alpha \times \tau_\beta) \cdot \tau_\gamma $1$ $+$ 1
$
$F^{(1)}_{1-3} =[\sigma_\alpha \sigma_\beta \sigma_\gamma]_{_3} \ \tau_\alpha \cdot \tau_\beta $1/N_c$ $-$ 3
$
$F^{(3)}_1 =[\sigma_\alpha \sigma_\beta \sigma_\gamma]_{_3} $1/N_c^3$ $-$ 1
$
5
---------------------------------------------------------------------------------------------------------------------------------------------------
: $S=3$ spin-isospin structures, as in Table \[table-3NL0\]. In the last line we give the total number of independent structures, obtained as the sum of $M_3^{(0)}=1, M_3^{(1)}=3, M_3^{(2)}=0$ and $M_3^{(3)}=1$.[]{data-label="table-3NL3"}
### $L=S=1$
In contrast to the two-nucleon potential, the three-nucleon potential contains $S=L=1$ terms at leading order. That is because vector spin-flavor structures can be constructed out of two and three $G$’s, see Table \[table-leadingops\], and then contracted with the P-even, T-even cross product ${\bf p}_- \times {\bf q}_-$. No such time-reversal-even cross product exists in the NN system at leading order. The leading-order $L=1$ force is then: $$V_{L=1}^{N_c}= N_c \sum_{\xi=1}^{M_1^{(0)}} V_{P_\xi}^{1}({\bf p}_- \times {\bf q}_-) \cdot P_\xi^{(0)}.
\label{LOL1}$$ Here $M_1^{(0)}=6$ is the number of leading-order $S=1$ structures, see Table \[table-3NL1\]. Since ${\bf q}_-$ appears in the momentum structure of Eq. (\[LOL1\]) they cannot occur in the NN force, and are unambiguously the result of 3N interactions. However, we note that, once again, the functions $V^1_{P_{1-3}}$ will be related to one another through permutation symmetry, as will the functions $V^1_{P_{4-6}}$.
Operators with matrix elements suppressed by $1/N_c$ are easily obtained, see Table \[table-3NL1\] where they are listed as $P_\xi^{(1)}$, with $\xi=1,\dots ,
M_1^{(1)}$, and $M_1^{(1)}=21$. However, as is displayed in the table, these are all time-reversal odd. Thus they must be contracted with T-odd cross products, and this costs another power of $1/N_c$, since it mandates that ${\bf p}_+$ or ${\bf q}_+$ be involved. Thus the order of such contributions is $1/N_c^2$ relative to leading. At this order the T-even structures $P_\xi^{(2)}$ also appear, contracted with ${\bf p}_- \times {\bf q}_-$, the leading-order momentum structure. The $L=1$ 3N force of order $1/N_c^2$ is therefore of the form: $$\begin{aligned}
\label{NLOL1}
V_{L=1}^{1/N_c}&=& \delta^{(2)} V_{L=1}^{N_c} +
N_c \sum_{\xi=1}^{M_1^{(2)}} V_{P_\xi}^{2}({\bf p}_- \times {\bf q}_-) \cdot P_\xi^{(2)}
\nonumber\\
&&+ N_c \sum_{\xi=1}^{M_1^{(1)}} \left\{ \ V_{P_\xi}^{3} {\bf p}_+ \times {\bf p}_-
+ V_{P_\xi}^{4,5} {\bf p}_\pm \times {\bf q}_\mp
+ V_{P_\xi}^{6} {\bf q}_+ \times {\bf q}_-
\right. \nonumber\\
&& \qquad \qquad \ + \left. \left( V^{7}_{P_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{8,9}_{P_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
({\bf p}_- \times {\bf q}_-) \ \right\} \cdot P_\xi^{(1)}
\nonumber\\
&&+ N_c \sum_{\xi=1}^{M_1^{(0)}} \left\{ \ V_{P_\xi}^{10} ({\bf p}_+ \times {\bf q}_+) \right. \nonumber\\
&&\qquad \qquad \ + \left. \left( V_{P_\xi}^{11} {\bf p}_+^2
+ V_{P_\xi}^{12} {\bf p_+ \cdot q_+}
+ V_{P_\xi}^{13} {\bf q}_+^2 \right) ({\bf p}_- \times {\bf q}_-) \right. \nonumber\\
&& \qquad \qquad \ + \left. \left( V^{14}_{P_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{15,16}_{P_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
({\bf p}_+ \times {\bf p}_-) \ \right.
\nonumber\\
&& \qquad \qquad \ + \left. \left( V^{17}_{P_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{18,19}_{P_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
({\bf p}_+ \times {\bf q}_-) \ \right.
\nonumber\\
&& \qquad \qquad \ + \left. \left( V^{20}_{P_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{21,22}_{P_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
({\bf p}_- \times {\bf q}_+) \ \right.
\nonumber\\
&& \qquad \qquad \ + \left. \left( V^{23}_{P_\xi} {\bf p}_+ \cdot {\bf p}_-
+ V^{24,25}_{P_\xi} {\bf p}_\pm \cdot {\bf q}_\mp \right)
({\bf q}_+ \times {\bf q}_-)
\ \right\} \cdot P_\xi^{(0)} \nonumber\\\end{aligned}$$ In this equation $M_1^{(2)}=12$ and $M_1^{(1)}=21$. Using two momenta there is only one structure of ${\cal O}(1)$. There are four structures of ${\cal O}(1/N_c)$ and one structure of ${\cal O}(1/N_c^2)$. With four momenta there are three structures at ${\cal O}(1/N_c)$ and 15 new structures of ${\cal O}(1/N_c^2)$. Triple products can be eliminated using the identity (\[eq:triple\]).
Once again, the new spin-flavor structures that appear at ${\cal O}(1/N_c^3)$ are all T-odd. Thus they must be combined with ${\bf p}_- \times {\bf
q}_+$, or one of the three other T-odd cross products involving a ${\bf +}$ vector, to yield something appropriate for inclusion in $V_{L=1}$. The overall result is then a contribution to the 3N force of relative order $1/N_c^4$. The expansion for $V_{L=1}$ is, like that for $V_{L=0}$, an expansion in $1/N_c^2$.
In summary, for $L=1$ there are six operators at LO, given in Eq. (\[LOL1\]), which are built out of T-even spin-isospin structures. At sub-leading order there 108 new operators involving T-even spin-isospin structures, and 147 operators involving T-odd ones, see Eq. (\[NLOL1\]). Up to ${\cal
O}(1/N_c)$ we presented the explicit expressions for these 261 operators, out of which 18 involve only ${\bf p}_-, {\bf q}_-$, and so correspond to a local potential.
### $L=S=2$
The leading 3N spin-flavor structures with $S=2$ are constructed out of $G$’s and $\mathds{1}$’s as shown in Table \[table-leadingops\]. A Cartesian rank-two tensor with $S=2$, constructed out of two vector quantities $A^i$ and $B^j$, is symmetric and traceless in its two indices and will be denoted $ [ A_i B_j ]_{_2}$ (see Eq. (\[eq:rank2tensor\])). Subleading structures are obtained after introducing a growing number of $S$ and $I$ operators, following Eq. (\[eq:Hartree\]).
The complete set of resulting spin-flavor structures is displayed in Table \[table-3NL2\]. This time we have $M_2^{(0)}=5$. The five LO structures must be contracted with $L=2$ tensors constructed out of ${\bf p}_-$ and ${\bf q}_-$ to obtain the LO contribution to the 3N potential.
We reiterate that Eq. (\[eq:triple2\]) explains why there are only two structures $D^{(0)}_{4,5}$, instead of the multiplicity three in similar structures with three spin operators in the $L=1$ case. Similar reductions in multiplicity occur for sub-leading $S=2$ spin-flavor structures too.
The leading force is then: $$\begin{aligned}
\label{LOL2}
V_{L=2}^{N_c}= N_c \sum_{\xi=1}^{M^{(0)}_2} \left\{ \ V_{D_\xi}^{1} \ [{\bf p}_- {\bf p}_-]_{_2}
+ V_{D_\xi}^{2} \ [{\bf p}_- {\bf q}_-]_{_2}
+ V_{D_\xi}^{3} \ [{\bf q}_- {\bf q}_-]_{_2} \ \right\} \cdot D_\xi^{(0)} \ , \end{aligned}$$ with the sum over the $M_2^{(0)}=5$ structures as listed in Table \[table-3NL2\]. Note that the first three structures in the sum, i.e. $D_{1-3}^{(0)}$, contracted with the appropriate ${\bf p}_-$, already occur in the NN potential. Only ${\bf q}_-$ dependence in $ V_{D_{1-3}}^{1} $ would reveal it is a “true" 3N force.
Using analogous arguments to those already discussed for the $L=0$ and $1$ cases, we obtain the subleading $L=S=2$ contribution to the 3NF: $$\begin{aligned}
\label{NLOL2}
V_{L=2}^{1/N_c}&=& \delta^{(2)} V_{L=2}^{N_c} + N_c \sum_{\xi=1}^{M^{(2)}_2} \left\{ \
V_{D_\xi}^{4} \ [{\bf p}_- {\bf p}_-]_{_2}
+ V_{D_\xi}^{5} \ [{\bf p}_- {\bf q}_-]_{_2}
+ V_{D_\xi}^{6} \ [{\bf q}_- {\bf q}_-]_{_2}
\ \right\} \cdot D_\xi^{(2)} \nonumber \\
&& + N_c \sum_{\xi=1}^{M^{(1)}_2} \left\{ \
V_{D_\xi}^{7} \ [{\bf p}_+ {\bf p}_-]_{_2}
+ V_{D_\xi}^{8,9} \ [{\bf p}_\pm {\bf q}_\mp]_{_2}
+ V_{D_\xi}^{10} \ [{\bf q}_+ {\bf q}_-]_{_2}
\right. \nonumber \\
&& \ \qquad \qquad + \left(
V_{D_\xi}^{11} \ {\bf p}_+ \cdot {\bf p}_- + V_{D_\xi}^{12,13} \ {\bf p_\pm \cdot q_\mp}
\right) [{\bf p}_- {\bf p}_-]_{_2} \nonumber \\
&& \ \qquad \qquad + \left(
V_{D_\xi}^{14} \ {\bf p}_+ \cdot {\bf p}_- + V_{D_\xi}^{15,16} \ {\bf p_\pm \cdot q_\mp} \right)
[{\bf p}_- {\bf q}_-]_{_2} \nonumber \\
&& \ \qquad \qquad + \left. \left( V_{D_\xi}^{17} \ {\bf p}_+ \cdot {\bf p}_- + V_{D_\xi}^{18,19} \ {\bf p_\pm \cdot q_\mp} \right)
[{\bf q}_- {\bf q}_-]_{_2}
\ \right\} \cdot D_\xi^{(1)} \nonumber \\
&& + N_c \sum_{\xi=1}^{M^{(0)}_2} \left\{ \ V_{D_\xi}^{20} \ [{\bf p}_+ {\bf p}_+]_{_2}
+ V_{D_\xi}^{21} \ [{\bf p}_+ {\bf q}_+]_{_2}
+ V_{D_\xi}^{22} \ [{\bf q}_+ {\bf q}_+]_{_2}
\right. \nonumber \\
&&\ \qquad \qquad + \left( V_{D_\xi}^{23} \ {\bf p}_+^2 + V_{D_\xi}^{24} \ {\bf p_+ \cdot q_+} + V_{D_\xi}^{25} \ {\bf q}_+^2 \right) [{\bf p}_- {\bf p}_-]_{_2} \nonumber\\
&&\ \qquad \qquad + \left( V_{D_\xi}^{26} \ {\bf p}_+^2 + V_{D_\xi}^{27} \ {\bf p_+ \cdot q_+} + V_{D_\xi}^{28} \ {\bf q}_+^2 \right) [{\bf p}_- {\bf q}_-]_{_2} \nonumber\\
&&\ \qquad \qquad + \left( V_{D_\xi}^{29} \ {\bf p}_+^2 + V_{D_\xi}^{30} \ {\bf p_+ \cdot q_+} + V_{D_\xi}^{31} \ {\bf q}_+^2 \right) [{\bf q}_- {\bf q}_-]_{_2} \nonumber\\
&&\ \qquad \qquad + \left( V_{D_\xi}^{32} \ {\bf p_+ \cdot p_-} + V_{D_\xi}^{33,34} \ {\bf p_\pm \cdot q_\mp} \right) [{\bf p}_+ {\bf p}_-]_{_2} \nonumber\\
&&\ \qquad \qquad + \left( V_{D_\xi}^{35} \ {\bf p_+ \cdot p_-} + V_{D_\xi}^{36,37} \ {\bf p_\pm \cdot q_\mp}\right) [{\bf p}_+ {\bf q}_-]_{_2} \nonumber\\
&&\ \qquad \qquad + \left( V_{D_\xi}^{38} \ {\bf p_+ \cdot p_-} + V_{D_\xi}^{39,40} \ {\bf p_\pm \cdot q_\mp}\right) [{\bf p}_- {\bf q}_+]_{_2} \nonumber\\
&&\ \qquad \qquad + \left. \left( V_{D_\xi}^{41} \ {\bf p_+ \cdot p_-} + V_{D_\xi}^{42,43} \ {\bf p_\pm \cdot q_\mp} \right) [{\bf q}_+ {\bf q}_-]_{_2}
\ \right\} \cdot D_\xi^{(0)} \ , \end{aligned}$$ where $M^{(1)}_2=9,M^{(2)}_2=9$. In writing Eq. (\[NLOL2\]) we employed three tensors built out of two momenta at ${\cal O}(1)$, four at ${\cal
O}(1/N_c)$ and three at ${\cal O}(1/N_c^2)$; with four momenta there are 9 structures at ${\cal O}(1/N_c)$ and 21 at ${\cal O}(1/N_c^2)$. Of course, all structures are again multiplied by the usual arbitrary functions of ${\bf p}_-^2$, ${\bf q}_-^2$, and ${\bf p}_- \cdot {\bf q}_-$. As in the $L=S=0$ and $L=S=1$ cases, the next-to-next-to-leading-order contributions to $V$ appear at relative order $1/N_c^4$, as a consequence of parity and time reversal.
In summary, for this $L=S=2$ part of the 3N force there are 15 operators at LO, which are built out of T-even spin-isospin structures. These are given in Eq. (\[LOL2\]). There are 264 new operators at relative order $1/N_c^2$, 147 (117) of which are based on T-even (T-odd) spin-flavor structures, see Eq. (\[NLOL2\]). Up to ${\cal O}(1/N_c)$ we presented the explicit expressions for all of these 279 operators. Amongst these are 42 operators that only depend on ${\bf p}_-, {\bf q}_-$ and could appear in a local potential.
### $L=S=3$
These operators have no NN analog. The rank-three Cartesian tensor with $L=3$ that can be constructed out of three vectors $A,B,C$ is symmetric and traceless, namely $$\begin{aligned}
{[}A_i B_j C_k]_{_3} &=& A_i B_j C_k + A_j B_k C_i + A_k B_i C_j + A_j B_i C_k + A_i B_k C_j + A_k B_j C_i {\nonumber}\\
&& - \frac25 \delta_{ij} (A \cdot B \ C_k + A \cdot C \ B_k + B \cdot C \ A_k) {\nonumber}\\
&& - \frac25 \delta_{ik} (A \cdot B \ C_j + A \cdot C \ B_j + B \cdot C \ A_j) {\nonumber}\\
&& - \frac25 \delta_{jk} (A \cdot B \ C_i + A \cdot C \ B_i + B \cdot C \ A_i) \ . \end{aligned}$$ There is just one leading-order spin-flavor structure, three suppressed by $1/N_c$ and one suppressed by $1/N_c^3$, as shown in Table \[table-3NL3\]. Note that there are no $1/N_c^2$ structures in this case, because of isospin conservation, as three $\sigma$ operators would need to appear along with one $\tau$ to generate a structure at that order. Table \[table-3NL3\] again shows that the suppressed structures are T-odd, and so ultimately lead to contributions to the 3N force that are down by $1/N_c^2$ and $1/N_c^4$ respectively.
Because of parity conservation, we need at least four momenta to construct the $L=3$ component of the potential. There are three $L=3$, P-even, T-even, momentum tensors at ${\cal O}(1)$. Both ${\bf q}_-$ and ${\bf p}_-$ are needed to construct these, as, e.g. ${\bf q}_- {\bf q}_{-} {\bf q}_{-} {\bf q}_{-} $ only contains $L=0,2,4$ components. The leading operator is $$\begin{aligned}
\label{LOL3}
V_{L=3}^{N_c} =
N_c \left\{ \
V^{1}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf p}_- {\bf p}_-]_{_3} +
V^{2}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf p}_- {\bf q}_-]_{_3} +
V^{3}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf q}_- {\bf q}_-]_{_3}
\ \right\}
\cdot F^{(0)}_1 \ . \nonumber \\\end{aligned}$$
The subleading $L=3$ force contains $L=3$, P-even T-odd tensors involving three ${\bf q}_-$ or ${\bf p}_-$ vectors, together with one ${\bf p}_+$ or ${\bf q}_+$, contracted with the three $F_{1-3}^{(1)}$ structures. Also appearing at this order are the terms which involve the leading spin-flavor structure, contracted with P-even T-even tensors in which two of the four vectors are ${\bf p}_+$ or ${\bf q}_+$. Meanwhile, the subleading correction to the LO structure $F_1^{(0)}$ corresponds to $s=t=1$ in Eq. (\[eq:Hartree\]) and is given by $\delta^{(2)} V_{L=3}^{N_c}$. Thus, finally we obtain $$\begin{aligned}
\label{NLOL3}
V_{L=3}^{1/N_c} &=& \delta^{(2)} V_{L=3}^{N_c} \nonumber \\
&& + N_c \sum_{\xi=1}^{M^{(1)}_3} \left\{ \
V^{4}_{F_\xi} [({\bf p}_+ \times {\bf p}_-) {\bf p}_- {\bf p}_-]_{_3} +
V^{5}_{F_\xi} [({\bf p}_+ \times {\bf p}_-) {\bf p}_- {\bf q}_-]_{_3} +
V^{6}_{F_\xi} [({\bf p}_+ \times {\bf p}_-) {\bf q}_- {\bf q}_-]_{_3} \right. \nonumber \\
&& \quad \qquad + \left.
V^{7,8}_{F_\xi} [({\bf p}_\pm \times {\bf q}_\mp) {\bf p}_- {\bf p}_-]_{_3} +
V^{9,10}_{F_\xi} [({\bf p}_\pm \times {\bf q}_\mp) {\bf p}_- {\bf q}_-]_{_3} +
V^{11,12}_{F_\xi} [({\bf p}_\pm \times {\bf q}_\mp) {\bf q}_- {\bf q}_-]_{_3} \right. \nonumber \\
&& \quad \qquad + \left.
V^{13}_{F_\xi} [({\bf q}_+ \times {\bf q}_-) {\bf p}_- {\bf p}_-]_{_3} +
V^{14}_{F_\xi} [({\bf q}_+ \times {\bf q}_-) {\bf p}_- {\bf q}_-]_{_3} +
V^{15}_{F_\xi} [({\bf q}_+ \times {\bf q}_-) {\bf q}_- {\bf q}_-]_{_3}
\ \ \ \right\} \cdot F^{(1)}_\xi \nonumber \\
&&+ N_c \left\{ \ V^{16}_{F_1} [({\bf p}_+ \times {\bf p}_-) {\bf p}_+ {\bf p}_-]_{_3} \right. \nonumber \\
&& \ \qquad \ + \left.
V^{17,18}_{F_1} [({\bf p}_+ \times {\bf p}_-) {\bf p}_\pm {\bf q}_\mp]_{_3}
+ V^{19}_{F_1} [({\bf p}_+ \times {\bf p}_-) {\bf q}_+ {\bf q}_-]_{_3} \right. \nonumber \\
&& \ \qquad \ + \left.
V^{20,21}_{F_1} [({\bf p}_\pm \times {\bf q}_\mp) {\bf p}_+ {\bf p}_-]_{_3}
+V^{22,23}_{F_1} [({\bf p}_\pm \times {\bf q}_\mp) {\bf p}_+ {\bf q}_-]_{_3}
\right. \nonumber \\
&& \ \qquad \ + \left.
V^{24,25}_{F_1} [({\bf p}_\pm \times {\bf q}_\mp) {\bf q}_+ {\bf p}_-]_{_3}
+V^{26,27}_{F_1} [({\bf p}_\pm \times {\bf q}_\mp) {\bf q}_+ {\bf q}_-]_{_3}
\right. \nonumber \\
&& \ \qquad \ + \left.
V^{28,29}_{F_1} [({\bf q}_+ \times {\bf q}_-) {\bf p}_\pm {\bf q}_\mp]_{_3}
+V^{30}_{F_1} [({\bf q}_+ \times {\bf q}_-) {\bf q}_+ {\bf q}_-]_{_3} \ \right\}
\cdot F^{(0)}_1 \ , \end{aligned}$$ where $M^{(1)}_3=3$. There are 12 momentum structures at ${\cal O}(1/N_c)$ and 15 momentum structures at ${\cal O}(1/N_c^2)$. Here we used again Eq. (\[eq:triple2\]) to reduce the number of momentum structures. The usual arguments show that the next correction, which involves the $F_1^{(3)}$ structure, is down by $1/N_c^4$ compared to the leading contribution, again because of the need to contract $F_1^{(3)}$ with a P-even, T-odd tensor.
In summary, for $L=3$ there are three operators at LO built out of T-even spin-isospin structures, given in Eq. (\[LOL3\]), and 51 additional (15 corresponding to new T-even and 36 to T-odd structures) at relative order $1/N_c^2$, given in Eq. (\[NLOL3\]). We presented the explicit expressions for these 54 operators which occur up to ${\cal O}(1/N_c)$. Out of these three depend solely on ${\bf p}_-, {\bf q}_-$ and could be part of a local potential.
Summary and Discussion {#sec-conclusion}
======================
We have classified all the spin-flavor structures that can contribute to the three-nucleon force (3NF) and power counted these structures in the $1/N_c$ expansion. The leading-order (LO) part of the 3NF is constructed from $G^{ia}/N_c$ and the unit operator, since these are the quark operators that have nucleon matrix elements that are ${\cal O}(1)$. Isospin-invariant structures like $$\mathds{1}_\alpha \mathds{1}_\beta \mathds{1}_\gamma ,
\qquad N_c^{-2} G^{ia}_\alpha \mathds{1}_\beta G^{j a}_\gamma ,
\qquad N_c^{-3} \epsilon^{abc} G^{ia}_\alpha G^{jb}_\beta G^{kc}_\gamma,
\label{eq:LOops}$$ with $\alpha$, $\beta$, and $\gamma$ labelling the three nucleons, are the leading contributions. Contraction of these structures with spatial tensors of the appropriate rank, built from the ${\cal O}(1)$ momenta ${\bf p}_-$ and ${\bf q}_-$, together with a re-expression in terms of the angular momentum content of these structures, and use of the reduction of $G^{ia}$ to $\sigma^i \tau^a$ when restricted to the nucleon subspace, produces the LO force: $$\begin{aligned}
\label{theLO}
V_{3{\rm N}}^{N_c} &=&N_c \sum_{\xi=1}^{M_{0}^{(0)}} V^1_{S_\xi} S_\xi^{(0)}
+ N_c \sum_{\xi=1}^{M_1^{(0)}} V_{P_\xi}^{1}({\bf p}_- \times {\bf q}_-) \cdot P_\xi^{(0)} \\
&& + N_c \sum_{\xi=1}^{M^{(0)}_2} \left\{ \ V_{D_\xi}^{1} \ [{\bf p}_- {\bf p}_-]_{_2}
+ V_{D_\xi}^{2} \ [{\bf p}_- {\bf q}_-]_{_2}
+ V_{D_\xi}^{3} \ [{\bf q}_- {\bf q}_-]_{_2} \ \right\} \cdot D_\xi^{(0)} \nonumber \\
&& +
N_c \left\{ \
V^{1}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf p}_- {\bf p}_-]_{_3} +
V^{2}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf p}_- {\bf q}_-]_{_3} +
V^{3}_{F_1} [({\bf p}_- \times {\bf q}_-) {\bf q}_- {\bf q}_-]_{_3}
\ \right\}
\cdot F^{(0)}_1. \nonumber \end{aligned}$$ Here the $S,P,D,F$ spin-flavor structures are given in Tables \[table-3NL0\], \[table-3NL1\], \[table-3NL2\], \[table-3NL3\]. The corresponding multiplicities are $M_{0,1,2,3}^{(0)}=5,6,5,1 $. Including all the independent momentum structures, to leading order we have 29 operators distributed as $5,6,15$ and 3 operators with $L=0,1,2,3$ respectively.
It follows straightforwardly that our LO force contains the structures present in the Fujita-Miyazawa three-nucleon potential, Eq. (\[eq:FM\]). Indeed, the only structure beyond the Fujita-Miyazawa result is the unit operator which is added to ${V}_{ijk}^{2
\pi}$ in most modern implementations of the 3NF. Of course, models of the 3NFs contain specific predictions for the coefficient functions $V_{L_\xi}^{m}$. The large-$N_c$ expansion can say nothing about these functions beyond the statement that they should be “natural", i.e. ${\cal O}(1)$; the insights from large-$N_c$ reside in the statements regarding the overall size that different spin-isospin-momentum structures within the 3NF should have.
The LO 3NF contains spin-dependent forces, but it does not contain the spin-orbit forces that have been proposed as a solution to the $A_y$ puzzle (see, e.g. Ref. [@Ki99]). The $A_y$ puzzle is not straightforwardly resolved by $1/N_c$ power counting arguments.
Spin-orbit forces, together with several other operators, all of which we have tabulated in Section \[sec-3N\], appear at ${\cal O}(1/N_c^2)$ compared to leading. We have also shown that the next-to-next-to-leading correction to the 3NF is at order $1/N_c^4$ relative to LO. The NNN force is therefore, like the NN force, an expansion in $1/N_c^2$. We have given explicit expressions for the 674 operators that appear in the 3N potential up to (overall) order $1/N_c$ in Eqs. (\[theLO\]), (\[NLOL0\]), (\[NLOL1\]), (\[NLOL2\]) and (\[NLOL3\]).
Many of these operators involve non-localities and time-reversal-odd momentum structures. For a local 3NF only time-reversal-even momentum structures involving ${\bf p}_-$ or ${\bf q}_-$ can occur. Such structures occur in both the leading and sub-leading 3NF, but do not occur at higher orders, where the presence of at least one time-reversal-odd momentum is required. Taking into account the different momentum structures which satisfy this constraint, at relative order $1/N_c^2$ we have $12$, $12$ and 27 operators with $L=0,1,2$ respectively. No $L=3$ local operator occurs at this relative order in the expansion. Combining these operators with the 29 LO operators yields a total of 80 operators that constitute the most general basis for a local 3NF. These operators can be easily read off from Eqs. (\[theLO\]), (\[NLOL0\]), (\[NLOL1\]), (\[NLOL2\]) and (\[NLOL3\]). In a recent paper [@Krebs:2013kha] the authors needed a basis of 89 operators to obtain the most general contribution of a local 3NF. Their operator basis is somewhat different from ours, so a comparison is not immediate. An important subject for future investigation is the relation between the two sets of operators, and a determination of the minimal basis of operators for a general, local 3NF.
We have not discussed the constraints imposed by the permutation group on our analysis. In the NN case such considerations resulted in the elimination of the spin structure $\sigma_1 - \sigma_2$. In the 3N case such constraints will impose relations between the different coefficient functions we have used in our expansion. Since the 3N coefficient functions depend on three rotational scalars it seems unlikely that a general, permutation-group-based argument can be used to eliminate an operator structure from the 3NF—at least in the absence of additional assumptions about the coefficient functions themselves. A permutation-group analysis of the structures we have obtained would be a useful step towards understanding the particular 3N partial waves which the different operator structures we have obtained contribute to.
It would also be interesting to test whether adding the $1/N_c^2$ structures we have listed here to phenomenological 3NFs improves the description of few-nucleon scattering data and light-nuclear spectra. Recently developed three-body potentials like the lllinois force [@Pi01] or that derived from $\chi$PT [@Be08; @Be11; @Krebs:2013kha] include several of these structures. Matching such 3NFs to our large-$N_c$ expressions is appreciably more involved than in the NN case analyzed in Ref. [@KM97], but nevertheless, they could be matched to the list of operators presented here. This would illuminate precisely which structures are present in these particular potentials, and whether the relative size of the different contributions is well predicted by the $1/N_c$ expansion.
An explicit example of operator reduction
=========================================
For completeness, it is worth discussing an explicit example of the operator reduction Eq. (\[eq:manyG\]). Consider the case of the product of two $G$’s, which can be written as $$\begin{aligned}
G^{ia} G^{jb} &=& \frac12 \{G^{ia},G^{jb}\} + \frac12 {[}G^{ia},G^{jb}] \ .\end{aligned}$$ The commutator contains the antisymmetric terms in $(ia)$ and $(jb)$ and is suppressed by $1/N_c^2$. The symmetric part in $(ia)$ and $(jb)$ can be written as a tensor $W^{(ij),(ab)}$ which is symmetric in the spatial indices $i,j$ and symmetric in the isospin indices $a,b$ independently, and a tensor $W^{[ij],[ab]}$ that is antisymmetric in the spatial and isospin indices taken separately $$\begin{aligned}
\{G^{ia},G^{jb}\} &=& W^{(ij),(ab)} + W^{[ij],[ab]} \ ,
\label{twoGs}\end{aligned}$$ where $$\begin{aligned}
W^{(ij),(ab)} &=& \frac12 \{G^{ia},G^{jb}\} + \frac12 \{G^{ib},G^{ja}\} \ , \\
W^{[ij],[ab]} &=& \frac12 \{G^{ia},G^{jb}\} - \frac12 \{G^{ib},G^{ja}\} = \epsilon^{ijk}
\epsilon^{abc} A^{kc} \ , \end{aligned}$$ with $$\begin{aligned}
A^{kc} &=& \frac14 \epsilon^{kij} \epsilon^{abc} \{G^{ia}, G^{jb}\} \ .\end{aligned}$$ Now we can use the following identity, see Ref. [@Dashen:1994qi], to reduce the number of $G$’s by one $$\begin{aligned}
\label{epsG2}
\epsilon^{ijk} \epsilon^{abc} \{G^{ia}, G^{jb}\}= - (N_c+2) G^{kc} + \frac12 \{
S^k,I^c\} \ .\end{aligned}$$ The $W^{(ij),(ab)}$ tensor has $(I,S)=(0,0),(2,0),(0,2),(2,2)$ components. Only the $(0,0)$ component contributes in the nucleon subspace. It is obtained by contracting the indices and is $W^{ij,ab}_{(0,0)}= \frac{2}{9} \delta^{ij} \delta^{ab} G^{kc}G^{kc} $. The $SU(4)$ quadratic Casimir operator $C_2 = \frac12 S^2 + \frac12 I^2 + 2 G^{kc}G^{kc}$ evaluated in the symmetric irrep $S_N$ that corresponds to ground state nucleons gives $C_2(S_{N_c}) =
\frac{3}{8} N_c(N_c+4) \mathds{1}$ and shows explicitly that to leading order $G^{kc}G^{kc}$ can be replaced by the unit operator. The two terms in Eq. (\[epsG2\]) that enter in $W^{[ij],[ab]}$ correspond to the $(1,1)$ component, and only the first one contributes to leading order.
We obtain $$\begin{aligned}
\langle {\rm N} | N_c^{-2} G^{ia} G^{jb} |{\rm N} \rangle &=& \langle {\rm N} | \frac{N_c+4}{48 N_c} \delta^{ij} \delta^{ab} \mathds{1}
- \frac{N_c+2}{8 N_c} \epsilon^{ijk} \epsilon^{abc} \left( \frac{G^{kc}}{N_c} \right)
|{\rm N} \rangle + {\cal O}(N_c^{-2}) \ ,
\label{twoGred}\end{aligned}$$ where this is an example of Eq. (\[eq:manyG\]) that shows explicitly the tensor structure in the spatial and isospin indices.
Useful tensor identities {#ap-reduction}
========================
Here we collect a few identities involving epsilon tensors that are used to simplify the number of spin-flavor and also momentum structures. The product of two epsilon tensors can be written as $$\begin{aligned}
\label{eq:epseps}
\epsilon_{ijk} \epsilon_{lmn} &=& {\rm det}
\left[
\begin{array}{ccc}
\delta_{il} & \delta_{im} & \delta_{in} \\
\delta_{jl} & \delta_{jm} & \delta_{jn} \\
\delta_{kl} & \delta_{km} & \delta_{kn} \\
\end{array}
\right] \nonumber \\
&=& \delta_{il} ( \delta_{jm} \delta_{kn} - \delta_{jn} \delta_{km} )
- \delta_{im} ( \delta_{jl} \delta_{kn} - \delta_{jn} \delta_{kl} )
+ \delta_{in} ( \delta_{jl} \delta_{km} - \delta_{jm} \delta_{kl} ) \ .
\label{eq:doubleepsilon}\end{aligned}$$ This expression is very useful for constructing an independent set of momentum structures. Another useful identity can be obtained by contracting Eq. (\[eq:doubleepsilon\]) with $\epsilon_{ijp} $ giving $$\begin{aligned}
\delta_{kp} \epsilon_{lmn} &=& \delta_{kn} \epsilon_{lmp} + \delta_{km} \epsilon_{lpn} +
\delta_{kl} \epsilon_{pmn} \ , \end{aligned}$$ from where $$\begin{aligned}
(A \times B)^i C^j + (B \times C)^i A^j + (C \times A)^i B^j = (A \times B) \cdot C \
\delta^{ij}
\label{eq:triple}\end{aligned}$$ is obtained. This can be used to eliminate triple products from all our momentum structures and also to reduce the number of momentum or spin-flavor structures that contain a cross product.
[99]{} N. Kalantar-Nayestanaki, E. Epelbaum, J. G. Messchendorp and A. Nogga, Rept. Prog. Phys. [**75**]{}, 016301 (2012) \[arXiv:1108.1227 \[nucl-th\]\]. H.-W. Hammer, A. Nogga and A. Schwenk, Rev. Mod. Phys. [**85**]{}, 197 (2013) \[arXiv:1210.4273 \[nucl-th\]\]. J. L. Friar, G. L. Payne, V. G. J. Stoks and J. J. de Swart, Phys. Lett. B [**311**]{}, 4 (1993) \[nucl-th/9307010\]. B. S. Pudliner, V. R. Pandharipande, J. Carlson, S. C. Pieper and R. B. Wiringa, Phys. Rev. C [**56**]{}, 1720 (1997) \[nucl-th/9705009\]. R. B. Wiringa, S. C. Pieper, J. Carlson and V. R. Pandharipande, Phys. Rev. C [**62**]{}, 014001 (2000) \[nucl-th/0002022\]. S. C. Pieper, K. Varga and R. B. Wiringa, Phys. Rev. C [**66**]{}, 044310 (2002) \[nucl-th/0206061\]. A. Nogga, S. K. Bogner and A. Schwenk, Phys. Rev. C [**70**]{}, 061002 (2004) \[nucl-th/0405016\]. S. C. Pieper and R. B. Wiringa, Ann. Rev. Nucl. Part. Sci. [**51**]{}, 53 (2001) \[nucl-th/0103005\]. P. Navratil, V. G. Gueorguiev, J. P. Vary, W. E. Ormand and A. Nogga, Phys. Rev. Lett. [**99**]{}, 042501 (2007) \[nucl-th/0701038\]. P. Maris, J. P. Vary and P. Navratil, Phys. Rev. C [**87**]{}, 014327 (2013) \[arXiv:1205.5686 \[nucl-th\]\]. T. Otsuka, T. Suzuki, J. D. Holt, A. Schwenk and Y. Akaishi, Phys. Rev. Lett. [**105**]{}, 032501 (2010) \[arXiv:0908.2607 \[nucl-th\]\]. J. D. Holt, T. Otsuka, A. Schwenk and T. Suzuki, J. Phys. G [**39**]{}, 085111 (2012) \[arXiv:1009.5984 \[nucl-th\]\]. J. W. Holt, N. Kaiser and W. Weise, Phys. Rev. C [**79**]{}, 054331 (2009) \[arXiv:0901.4750 \[nucl-th\]\]. J. Fujita and H. Miyazawa, Prog. Theor. Phys. [**17**]{}, 360 (1957). J. Carlson, V. R. Pandharipande and R. B. Wiringa, Nucl. Phys. A [**401**]{}, 59 (1983). J. Carlson and R. Schiavilla, Rev. Mod. Phys. [**70**]{}, 743 (1998).
J. L. Friar, D. Huber and U. van Kolck, Phys. Rev. C [**59**]{}, 53 (1999) \[nucl-th/9809065\]. S. C. Pieper, V. R. Pandharipande, R. B. Wiringa and J. Carlson, Phys. Rev. C [**64**]{}, 014001 (2001) \[nucl-th/0102004\]. S. A. Coon, M. D. Scadron, P. C. McNamee, B. R. Barrett, D. W. E. Blatt and B. H. J. McKellar, Nucl. Phys. A [**317**]{}, 242 (1979). S. A. Coon and M. T. Pena, Phys. Rev. C [**48**]{}, 2559 (1993) \[nucl-th/9305017\]. M. R. Robilotta and M. P. Isidro Filho, Nucl. Phys. A [**414**]{}, 394 (1984). U. van Kolck, Phys. Rev. C [**49**]{}, 2932 (1994). E. Epelbaum, H. Krebs and U. -G. Meißner, Nucl. Phys. A [**806**]{}, 65 (2008) \[arXiv:0712.1969 \[nucl-th\]\]. E. Epelbaum, A. Nogga, W. Gloeckle, H. Kamada, U. G. Meißner and H. Witala, Phys. Rev. C [**66**]{}, 064001 (2002) \[nucl-th/0208023\]. K. Hebeler and A. Schwenk, Phys. Rev. C [**82**]{}, 014314 (2010) \[arXiv:0911.0483 \[nucl-th\]\]. T. B. Clegg, PoS CD [**09**]{}, 061 (2009). V. Bernard, E. Epelbaum, H. Krebs and U. -G. Meißner, Phys. Rev. C [**77**]{}, 064004 (2008) \[arXiv:0712.1967 \[nucl-th\]\]. V. Bernard, E. Epelbaum, H. Krebs and U. -G. Meißner, Phys. Rev. C [**84**]{}, 054001 (2011) \[arXiv:1108.3816 \[nucl-th\]\]. H. Krebs, A. Gasparyan and E. Epelbaum, Phys. Rev. C [**87**]{}, 054007 (2013) \[arXiv:1302.2872 \[nucl-th\]\]. L. Girlanda, A. Kievsky and M. Viviani, Phys. Rev. C [**84**]{}, 014001 (2011) \[arXiv:1102.4799 \[nucl-th\]\]. G. ’t Hooft, Nucl. Phys. B [**72**]{}, 461 (1974). E. Witten, Nucl. Phys. B [**160**]{}, 57 (1979). E. E. Jenkins, A. V. Manohar, J. W. Negele and A. Walker-Loud, Phys. Rev. D [**81**]{}, 014502 (2010) \[arXiv:0907.0529 \[hep-lat\]\]. E. E. Jenkins, Ann. Rev. Nucl. Part. Sci. [**48**]{}, 81 (1998) \[arXiv:hep-ph/9803349\]. A. V. Manohar, arXiv:hep-ph/9802419. D. B. Kaplan and M. J. Savage, Phys. Lett. B [**365**]{}, 244 (1996) \[hep-ph/9509371\]. D. B. Kaplan and A. V. Manohar, Phys. Rev. C [**56**]{}, 76 (1997) \[arXiv:nucl-th/9612021\]. V. G. J. Stoks, R. A. M. Klomp, C. P. F. Terheggen [*et al.*]{}, Phys. Rev. [**C49**]{}, 2950-2962 (1994). R. F. Dashen, E. E. Jenkins and A. V. Manohar, Phys. Rev. D [**51**]{}, 3697 (1995) \[hep-ph/9411234\]. W. Detmold, K. Orginos, M. J. Savage, Phys. Rev. [**D76**]{}, 114503 (2007). M. K. Banerjee, T. D. Cohen, B. A. Gelman, Phys. Rev. [**C65**]{}, 034011 (2002). A. V. Belitsky, T. D. Cohen, Phys. Rev. [**C65**]{}, 064008 (2002). T. D. Cohen, Phys. Rev. [**C66**]{}, 064003 (2002). L. Wolfenstein, Phys. Rev. [**96**]{}, 1654 (1954).
A. Kievsky, Phys. Rev. C [**60**]{}, 034001 (1999) \[nucl-th/9905045\].
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Patrick Hennebelle, Olivier Iffrig'
bibliography:
- 'biblio.bib'
title: Simulations of magnetized multiphase galactic disk regulated by supernovae explosions
---
Introduction
============
Star formation is a multi-scale and multi-physics problem, which is only partially understood. In particular what physical process is responsible for the relatively low star formation of the Milky-way [e.g. @zuckerman+1974; @dobbs+2013] remains a subject of controversy.
Historically three main physical processes have been emphasized, namely magnetic field [e.g. @shu+1987], turbulence [e.g. @maclow+2004], and stellar feedback [e.g. @maclow2013; @agertz+2013] which includes supernovae explosions, ionising radiation, heating by stellar radiation, stellar outflows and stellar winds. While magnetic field has been measured to have substantial intensities [@crutcher2012], it may be nevertheless too weak to reduce the star formation rate (SFR) by orders of magnitude. The effects of turbulence and feedback are not straighforward to disantangle. In particular, turbulence decays in a few crossing times [e.g. @maclow+2004] and must be fed at large scales either through galactic large scale spiral waves or by the various sources of stellar feedback. The feedback may therefore play a dual role in limiting the amount of mass that is eventually accreted by stars while in the same time triggering large scale turbulence. In any case, previous studies, which have been simulating a whole galaxy found a clear impact of the feedback onto the SFR [e.g. @tasker-bryan2006; @dubois+2008; @bournaud+2010; @kim+2011; @dobbs+2011; @tasker2011; @hopkins+2011; @renaud+2013], which is able to reduce star formation substantially, may be up to observed values. Although there is a general agreement that the simulations without feedback present too high an SFR, the amount by which it is reduced when feedback is introduced depends on which feedback is introduced and how. For example @tasker-bryan2006 who perfomed simulations with supernovae feedback found that the SFR is reduced by a factor of about 2, @tasker2011 included the UV radiation feedback found that the SFR is typically reduced by a factor of 1.5-2. Finally, @hopkins+2011 have introduced the radiative feedback assuming that the radiation of stars can efficienly communicate its momentum to the gas. They found that the SFR can be reduced by a factor on the order of 10 to 30.
In spite of these studies, the exact roles played by feedback, both for triggering the turbulence and for limiting the star formation, as well as by magnetic field are only partially understood. First of all, given the limited resolution of large scale studies (typically a few pc) the exact way feedback is applied remains partially arbitrary. In particular, the first pioneering studies which have been first focussing on kpc scales regulated by supernovae explosions [@deavillez+2005; @joung+2006], did not include self-gravity and therefore could not associate supernovae with star formation events accurately though they unanbigously show the relevance of these studies. Even when self-gravity is included, the exact influence of the choices made to determine their locations has not been explored clearly. Second of all, the influence that the magnetic field has on the star formation rate at the kpc scale is less explored. The only simulations that include both self-gravity and magnetic field which have been performed to date are presented in @Wang+2009 [@rudiger+2013]. These authors have been concluding that magnetic field reduces the star formation rate by a factor of a few. Moreover these 2 studies modeled a whole galaxy implying that the spatial resolution is necessary limited to describe the ISM structure.
In parallel to the numerical studies [@slyz+2005; @deavillez+2005; @joung+2006; @hill+2012; @kim+2011; @kim+2013; @gent+2013], a few analytical models have been developed and compared with observations and simulations [@ostriker+2010; @kim+2011; @faucher+2013]. For example, in their model, @ostriker+2010 consider vertical mechanical equilibrium between gravity and pressure (mainly thermal and kinetic) as well as thermal equilibrium in the ISM (i.e. equilibrium between heating from stars and cooling). It is then further assumed that the thermal and turbulent supports are proportional. This leads them to predict the SFR as a function of the column density through the galactic planes and the model compares well with a sample of observations [@leroy+2008] and simulations [@kim+2011]. In their models, @faucher+2013 consider a galactic disk, which is also in equilibrium along the vertical direction but assume that the galactic disk has a Toomre parameter [@toomre1964] of about $Q \simeq 1$, that is to say is in marginal equilibrium and self-regulates. They then perform an energy budget between the energy dissipated by units of time through turbulent cascade and the energy injected through stellar feedback. In both cases important assumptions are made regarding how energy and momentum are injected within the dense gas. While it is clearly unavoidable to make such assumptions, given the difficulty of the problem, it is nevertheless important to understand how exactly are momentum and energy injected, more precisely how they distribute between diffuse and dense gas. More generally, what are the uncertainties induced by our incomplete understanding of the correlation between massive stars, at the origin of most of the feedback, and the surrounding dense material ? It is the purpose of the present paper to address these issues.
In this paper we adopt a similar setup to the one adopted by @slyz+2005 [@deavillez+2005; @joung+2006; @hill+2012; @kim+2011; @kim+2013; @gent+2013], that is to say a kpc simulation of a galactic disk in which turbulence is driven by supernovae remnants. Previous studies have been finding that they can reproduce many of the interstellar medium feature such as multi-phase ISM, approximate energy equipartition between thermal, magnetic and kinetic energies, galactic outflows and formation of molecular clouds therefore demonstating the interest of performing this type of simulations. Indeed, this range of scales is a good compromise between the need for enough numerical resolution to describe sufficiently well the ISM physics and the amount of computational power available on present computers. While the results obtained previously are encouraging, only few of these works have been treating self-gravity and none of these works have been considering magnetic field and self-gravity, which is mandory for a proper physical description.
We include both self-gravity and magnetic field and we explore various schemes for the supernovae feedback going from a random distribution to a distribution in which supernovae are correlated both spatially and temporally with star formation. This spatial and temporal correlation turns out to be drastically important when self-gravity is self-consistently treated. The primary reason is that when a dense region undergoes gravitationnal collapse, it becomes largely decoupled from the surrounding medium and therefore little influenced by the supernovae which may be exploding nearby. It is only if a supernova explodes within the collapsing region and while collapsing is occuring that feedback has a significant impact and can reduce the mass that is eventually accreted.
The second section of the paper presents the numerical setup and in particular discusses the various schemes we have been developing to implement supernovae feedback. In the third section we describe the disk structure in the various models while in fourth section we investigate the properties of the multi-phase ISM in two of the simulations. In section 5 we discuss the star formation rate and the mass distribution of the sink particles formed in the different simulations. Section six concludes the paper.
Numerical setup
===============
(0,18) (0,8)[{width="9cm"}]{} (9,8)[{width="9cm"}]{} (0,0)[{width="9cm"}]{} (9,0)[{width="9cm"}]{}
(0,18) (0,0)[{width="9cm"}]{} (9,0)[{width="9cm"}]{} (0,8)[{width="9cm"}]{} (9,8)[{width="9cm"}]{}
Physical processes and initial conditions
-----------------------------------------
The physical processes and initial conditions are similar to what has been described by previous authors [e.g. @deavillez+2005; @joung+2006; @hill+2012; @kim+2013]. We consider a 1 kpc computational box in which a gravitational field is applied along the $z$-axis. Its value is identical to the choice made by @joung+2006 taken from @Kuijken+1989 and is given by $$\begin{aligned}
g(z) = - { a_1 z \over \sqrt{z^2+z_0^2} } -a_2 z,
\label{grav_pot}\end{aligned}$$ where $a_1=1.42 \times 10^{-3}$ kpc Myr$^{-2}$, $a_2= 5.49 \times 10^{-4}$ Myr$^{-2}$ and $z_0=0.18$ kpc. This gravitational field represents the contribution of the stars and dark matter in our Galaxy.\
We solve the ideal magneto-hydrodynamical (MHD) equations in the presence of self and external gravity and include cooling and heating processes relevant for the ISM. The equations are given by $$\begin{aligned}
\frac{\partial \rho}{\partial t} + { \mbox{\boldmath{$\nabla$}} } {{ \mbox{\boldmath{$\cdot$}} }}(\rho { \mbox{\boldmath{$v$}} })
&=& 0, \\
\rho \left[ \frac{\partial { \mbox{\boldmath{$v$}} }}{\partial t} + ({ \mbox{\boldmath{$v$}} } {{ \mbox{\boldmath{$\cdot$}} }}{ \mbox{\boldmath{$\nabla$}} }) { \mbox{\boldmath{$v$}} } \right] &=& -{ \mbox{\boldmath{$\nabla$}} } P +
\frac{({ \mbox{\boldmath{$\nabla$}} } {{ \mbox{\boldmath{$\times$}} }}{ \mbox{\boldmath{$B$}} }) {{ \mbox{\boldmath{$\times$}} }}{ \mbox{\boldmath{$B$}} }}{4\pi} \nonumber \\
&& - \rho {\bf g} - \rho { \mbox{\boldmath{$\nabla$}} } \phi, \label{impulsion_eq}\\
\rho \left[ \frac{\partial { \mbox{\boldmath{$e$}} }}{\partial t} + ({ \mbox{\boldmath{$v$}} } {{ \mbox{\boldmath{$\cdot$}} }}{ \mbox{\boldmath{$\nabla$}} }) { \mbox{\boldmath{$e$}} } \right] &=& - P ({ \mbox{\boldmath{$\nabla$}} } {{ \mbox{\boldmath{$\cdot$}} }}{ \mbox{\boldmath{$v$}} }) - \rho {\cal L}, \label{internal_nrj_eq} \\
\frac{\partial { \mbox{\boldmath{$B$}} }}{\partial t} &=& { \mbox{\boldmath{$\nabla$}} } {{ \mbox{\boldmath{$\times$}} }}({ \mbox{\boldmath{$v$}} }
{{ \mbox{\boldmath{$\times$}} }}{ \mbox{\boldmath{$B$}} }), \label{induction_eq} \\
\Delta \phi &=& 4 \pi G \rho, \label{poisson}\end{aligned}$$ where all notation have their usual meaning. The heating and cooling terms, which appear in the loss function ${\cal L}$, are identical to the one considered in @ah05, i.e. include UV heating due to photoelectric effect on grains, Lyman-$\alpha$, oxygen and ionised carbon cooling as well as cooling due to the recombination onto grains [see e.g. @wolfire+2003]. Note that at this stage we use a constant UV field which is not correlated to the star formation rate [e.g. @tasker2011; @kim+2013]. At temperature larger than 10$^4$ K, we use the fit provided in @joung+2006 for the cooling function inferred by @sutherland+1993. Coriolis and centrifugal forces are not included at this stage.
At the beginning of the simulation, the density distribution is given by $$\begin{aligned}
n(z) = n_0 \exp \left(- \left( { z \over z_0} \right)^2 \right),
\label{dens0}\end{aligned}$$ with $n_0=1.5$ cm$^{-3}$ and $z_0=150$ pc. The temperature is initially equal to about 8000 K which corresponds to the temperature of the warm neutral gas (WNM). A [*turbulent*]{} velocity field is generated using random phase and a Kolmogorov powerspectrum. Its rms amplitude is equal to 5 km s$^{-1}$. Finally, the magnetic field is initially aligned along the $x$-axis and is proportional to the density field. Its value in the equatorial plane is about 2.5$\mu$G for our fiducial run (later named run C1), the value of 0.5$\mu$G is also explored.
Supernovae prescriptions
------------------------
To take into account the supernovae feedback we first select a position as described below. Then we increase the thermal energy of all the cells located at a distance smaller than three grid cells from the supernova center in such a way that the thermal energy is uniform in this sphere and equal to $10^{51}$ erg. In most, but not all, of our runs we have also introduced a kinetic feedback. This is achieved by adding to the corresponding cells a radial homologous velocity field (proportional to the distance from the supernova center). The total kinetic energy is equal to 5% of the thermal energy which corresponds to the typical momentum that is injected at the end of the Sedov Phase, that is to say during the phase for which the supernova expansion remains nearly adiabatic. Indeed when the shell surrounding the supernova bubble becomes radiative, most of the energy is radiated away and the expansion proceeds at constant momentum [e.g. @chevalier77]. Fortunately enough this momentum is largely independent of the density field [@blondin+1998] even when it is highly irregular (Iffrig & Hennebelle 2014, in prep). Note that in principle, the momentum should be self-consistently generated during the Sedov expansion. However, the lack of resolution does not guarantee that this phase is well treated when the supernovae explodes in a dense region. In any case, as described below we have also performed the case without kinetic feedback for comparison. These two runs probably constitute upper and lower limits.
The spatial location of the supernovae is another important aspect. Previous authors [e.g. @deavillez+2005; @joung+2006] have been distributing them randomly or in correlation with density. These authors also tested the case where the supernovae are clustered [see e.g. @joung+2006 for a description]. In this work, four spatial and temporal supernovae distributions have been tested, hereafter scheme A, B, C and D. We recall that an important difference with these classical studies is that self-gravity is treated.
Scheme A is very similar to the scheme described for example in @joung+2006. The supernovae are distributed randomly in the $x$ and $y$-directions. To mimic the observed supernovae distribution in the Milky Way, their z-coordinate follows a Gaussian distribution of thickness equal to 150 pc. Their rate is equal to the observed galactic rate and is equal to 1/50 per year. One difference is that we use a fixed radius for the supernovae remnant rather than using a radius which enclosed a fixed mass. This implies that the timesteps in the simulation can be quite low since the temperature is higher when a supernova explodes in a diffuse medium. Another difference is that we do not redistribute the mass within the supernova radius as @joung+2006 who impose a uniform gas density inside the supernova bubbles. In principle, it could be worth testing all these choices but here we focus on a different issue.
Scheme B consists looking at the density maximum in the simulations and choosing its position for the supernova center. The rate is also equal to the galactic one. A supernovae is introduced only if the density peak is larger than $10$ cm$^{-3}$ in the simulation. However, denser gas develops rapidly in the simulation and except for the very first, the supernovae are generally associated to gas of densities $10^2-10^3$ cm$^{-3}$ which is present at all time. This scheme has the advantage to have a good spatial correlation with star formation events. However it does not have any temporal correlation since the supernovae rate is fixed and equal to the galactic one.
Scheme C and D are different and take advantage of the sink particles used in our simulations. Each time a sink particle has accreted 120 solar masses, we place a supernova in its neighborhood. This prescription is motivated by the typical abundance of stars more massive than the 8 solar masses needed to give raise to a supernova explosion. However, we do not place the supernova center directly at the sink position for various reasons. First of all, as described later the sink particles have a radius of 4 computational cells, corresponding to 16 pc with our current resolution. This number of cells for the sink radius is typical to what is usually assumed [e.g. @Krumholz+04]. By definition it sets the limit of the resolution in the simulation and inside the sink, the gas distribution is not well described. Second of all, it takes at least 4 Myr for the most massive stars to explode. During this time both the star and the cloud have evolved. For example in 10 Myr, a star which moves at a velocity of 1 km s$^{-1}$, will have cover a distance of about 10 pc. Finally, because of the other sources of feedback, massive stars may have pushed the dense gas away before supernovae explode. To test the importance of the spatial correlation between supernovae and sink particles, we have implemented two prescriptions. First (scheme C), we place randomly the supernovae within the sink particles radius. Second (Scheme D), we place them within a shell of inner radius equal to the sink radius and outer radius equal to two times this value. As will be seen later, these two schemes lead to similar but not identical results.
Another important issue is the time delay that should be also taken into account since supernovae typically explode between about 4 and 40 Myr after the formation of the massive star. Although we note that some delay is introduced with our scheme since, at least 120 solar masses of gas have to be accreted before supernovae take place, it is in most of the time shorter (10$^5-10^6$ yr) than a few Myr. However as emphasized in other studies [e.g. @matzner2002; @dale+2012; @dale+2013; @agertz+2013; @kim+2013], other sources of feedback, namely ionising radiation, winds and jets, start influencing the surrounding clouds much earlier. These sources should be taken into account as well. Since we feel it is important to go step by step, we postpone studying these effects and concentrate for now on supernovae only.
Numerical code and resolution {#code}
-----------------------------
To carry out our numerical simulations, we employed RAMSES [@teyssier2002; @fromang+2006], an adaptive mesh refinement (AMR) code that uses Godunov schemes to solve the MHD equations and the constrained transport method to ensure that ${\rm div}B$ is maintained at zero within machine accuracy. For most runs, we do not use the AMR capacity and keep the resolution fix using 256 grid points in each direction. However, we also perform a run with one more AMR level introduced when the density reaches a value of 10 cm$^{-3}$. The computational box size is equal to 1 kpc and the spatial resolution is 4 pc. This choice is dictated by the very large number of time steps ($\simeq$ 50,000-100,000), which are required for these calculations. This is because supernovae feedback introduced very high velocities of the order of few 100 km s$^{-1}$ as well as temperature as high as $10^8$ K in few cells. Moreover, we integrate far enough in order to make sure that some quasi-stationary regime has been reached. This will be assessed by verifying that the mean profile of various quantities such as densities does not vary significantly with time. It should be stressed however that since self-gravity is treated and accretion is occuring onto the sink particles, no strict stationarity can be reached.
We use periodic boundary conditions in $x$ and $y$ directions and outflow condition at the $z$ boundaries. In particular, this implies that the gas ejected from the galactic disk can escape the computational box.
In this work sink particles (implemented in the public version of RAMSES) are being used to follow the dense regions which have collapsed under the influence of self-gravity. They closely follow the implementation of @Krumholz+04. The sinks are introduced when a density threshold of 10$^3$ cm$^{-3}$ is reached. Their radius is equal to 4 grid cells. A new sink can be created only if it is not located closer than 10 grid cells from another sink. When the sinks get too close, i.e. closer than one grid cell, they get merge using a friend of friend procedure. Finally, the sinks accrete gas from surrounding cells if they are located at less than a sink radius and if the density is larger than 10$^3$ cm$^{-3}$. The sink particles interact with the surrounding medium through the gravitational field. The contribution of the sink to the gravitational potential is included using a particle-mesh approach, that is to say the mass within the sink is projected onto the grid and added to the gas density when the Poisson equation is solved.
Comparison between various setups
---------------------------------
It is worth emphazising the differences between the various setups which have been used so far, keeping in mind that given the complexity of the problem under investigation, $i)$ it is hard to include every relevant process, $ii)$ it is important to perform studies which make different choices and approximations to disantangle the effects of the different physical processes.
In the setup used by @deavillez+2005 [@joung+2006] and @hill+2012, hydro or MHD equations are solved using a 500 pc or a 1 kpc size box in the equatorial plane but a much larger scale height (typically 5 to 10 times these values). This insures a good description of the galactic wind and the disk-halo connection though large scale modes may be filtered out by the elongated box. @gent+2013 proceed somewhat similarly but also include the galactic shear in their study. In these studies the supernovae explode randomly or are correlated with density peaks. None of these studies include self-gravity.
@kim+2011 [@kim+2013] do include self-gravity and like us, consider a cubic computational box. These choices are clearly putting more emphasis on the disk itself and therefore on the star formation, which is taking place, than on the galactic outflows and the halo. There is an important difference regarding the forcing. They estimate the number of massive stars that should form in a given region rather than following the accretion onto sink particles and they consider only the mechanical feedback from supernovae.
As will be seen below, the different prescriptions leads to quite different results. In particular, the correlation between star forming gas and the feedback is a necessary condition to prevent very efficient star formation. This is however true mostly if self-gravity is included.
Runs performed
--------------
In the present paper, we perform various runs to test the influence that magnetic field has onto the galactic disk evolution and to study the influence of the various prescriptions for the supernovae feedback. In our fiducial run (later referred as C1), the magnetic intensity has an intensity in the equatorial plane of about 2.5 $\mu$G and scheme C is used for the supernovae with both thermal and kinetic feedback.
To study the influence of the magnetic field, we perform an hydrodynamical run (later referred as C2) and a run with a lower magnetisation initially equal to about 0.5 $\mu$G in the midplane (run C3). Apart from the strength of the field, these runs are identical to run C1.
To study the influence of the feedback scheme, we perform a series of runs identical to run C1 apart for the feedback scheme. First, we run two cases without any feedback, one purely hydrodynamical (NF2) and one with 2.5 $\mu$G initially (NF1). Second, we consider two cases with scheme A and B respectively (simply labelled runs A and B) and both thermal and kinetic feedback. Third we perform a simulation with supernovae scheme C but with thermal feedback only (run C4). Fourth we carry out a calculation with scheme D (run D).
Finally, to investigate the important issue of numerical resolution, we also present a run identical to run C1 but using another level of refinement leading to an effective resolution of about 2 pc. The refinement is performed when the cell density reaches a threshold of 10 cm$^{-3}$ leading to a total number cells at this level comparable to the number of cells in the non-refined runs. The results are presented in the Appendix.
Table \[run\] summarizes the runs performed in the paper and provide consistent label.
[| [l]{} | [l]{} | [l]{} | [l]{} | [l]{} | ]{} label & physics & scheme & feedback & resolution\
NF1 & MHD (2.5 $\mu$G) & - & - & 256$^3$\
NF2 & hydro & - & - & 256$^3$\
A & MHD (2.5 $\mu$G) & A & thermal+kinetic & 256$^3$\
B & MHD (2.5 $\mu$G) & B & thermal+kinetic & 256$^3$\
C1 & MHD (2.5 $\mu$G) & C & thermal+kinetic & 256$^3$\
C1b & MHD (2.5 $\mu$G) & C & thermal+kinetic & 512$^3$\
C2 & hydro & C & thermal+kinetic & 256$^3$\
C3 & MHD (0.5 $\mu$G) & C & thermal+kinetic & 256$^3$\
C4 & MHD (2.5 $\mu$G) & C & thermal & 256$^3$\
D & MHD (2.5 $\mu$G) & D & thermal+kinetic & 256$^3$\
(0,23.) (0,18)[![Mean density profile along the $z$-axis for five different models (see label) at four different timesteps. The disk profile is much thinner for runs NF1 and A than for runs C1, C2 and D.[]{data-label="rho_z"}](LS_grav_pmsink_sn_dens3/rho_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Mean density profile along the $z$-axis for five different models (see label) at four different timesteps. The disk profile is much thinner for runs NF1 and A than for runs C1, C2 and D.[]{data-label="rho_z"}](LS_grav_pmsink_sn/rho_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Mean density profile along the $z$-axis for five different models (see label) at four different timesteps. The disk profile is much thinner for runs NF1 and A than for runs C1, C2 and D.[]{data-label="rho_z"}](LS_grav_pmsink/rho_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Mean density profile along the $z$-axis for five different models (see label) at four different timesteps. The disk profile is much thinner for runs NF1 and A than for runs C1, C2 and D.[]{data-label="rho_z"}](LS_grav_pmsink_sn_dens3_bis/rho_z.jpg "fig:"){width="8cm"}]{} (0,13.5)[![Mean density profile along the $z$-axis for five different models (see label) at four different timesteps. The disk profile is much thinner for runs NF1 and A than for runs C1, C2 and D.[]{data-label="rho_z"}](LS_grav_pmsink_sn_dens3_hydro/rho_z.jpg "fig:"){width="8cm"}]{}
(0,23.) (0,18)[![Thermal pressure profile along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for run D and the smallest for run NF1 which has no feedback.[]{data-label="Pth_z"}](LS_grav_pmsink_sn_dens3/Pth_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Thermal pressure profile along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for run D and the smallest for run NF1 which has no feedback.[]{data-label="Pth_z"}](LS_grav_pmsink_sn/Pth_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Thermal pressure profile along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for run D and the smallest for run NF1 which has no feedback.[]{data-label="Pth_z"}](LS_grav_pmsink/Pth_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Thermal pressure profile along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for run D and the smallest for run NF1 which has no feedback.[]{data-label="Pth_z"}](LS_grav_pmsink_sn_dens3_bis/Pth_z.jpg "fig:"){width="8cm"}]{} (0,13.5)[![Thermal pressure profile along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for run D and the smallest for run NF1 which has no feedback.[]{data-label="Pth_z"}](LS_grav_pmsink_sn_dens3_hydro/Pth_z.jpg "fig:"){width="8cm"}]{}
(0,23.) (0,18)[![Kinetic pressure profile along the $z$-axis for five different models (see label) at four different timesteps. Run D presents the largest values while Run NF1 presents values significantly smaller than the other runs. Moreover kinetic pressure quickly drops at higher altitude for run NF1.[]{data-label="Pkin_z"}](LS_grav_pmsink_sn_dens3/Pkin_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Kinetic pressure profile along the $z$-axis for five different models (see label) at four different timesteps. Run D presents the largest values while Run NF1 presents values significantly smaller than the other runs. Moreover kinetic pressure quickly drops at higher altitude for run NF1.[]{data-label="Pkin_z"}](LS_grav_pmsink_sn/Pkin_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Kinetic pressure profile along the $z$-axis for five different models (see label) at four different timesteps. Run D presents the largest values while Run NF1 presents values significantly smaller than the other runs. Moreover kinetic pressure quickly drops at higher altitude for run NF1.[]{data-label="Pkin_z"}](LS_grav_pmsink/Pkin_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Kinetic pressure profile along the $z$-axis for five different models (see label) at four different timesteps. Run D presents the largest values while Run NF1 presents values significantly smaller than the other runs. Moreover kinetic pressure quickly drops at higher altitude for run NF1.[]{data-label="Pkin_z"}](LS_grav_pmsink_sn_dens3_bis/Pkin_z.jpg "fig:"){width="8cm"}]{} (0,13.5)[![Kinetic pressure profile along the $z$-axis for five different models (see label) at four different timesteps. Run D presents the largest values while Run NF1 presents values significantly smaller than the other runs. Moreover kinetic pressure quickly drops at higher altitude for run NF1.[]{data-label="Pkin_z"}](LS_grav_pmsink_sn_dens3_hydro/Pkin_z.jpg "fig:"){width="8cm"}]{}
(0,18.5) (0,13.5)[![Magnetic pressure profile along the $z$-axis for four different models (see label) at four different timesteps. All runs show similar values in the midplane. For runs A, C1 and D the profile tends to become shallower with time illustrating the transport of the magnetic flux toward higher altitude.[]{data-label="Pmag_z"}](LS_grav_pmsink_sn_dens3/Pmag_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Magnetic pressure profile along the $z$-axis for four different models (see label) at four different timesteps. All runs show similar values in the midplane. For runs A, C1 and D the profile tends to become shallower with time illustrating the transport of the magnetic flux toward higher altitude.[]{data-label="Pmag_z"}](LS_grav_pmsink_sn/Pmag_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Magnetic pressure profile along the $z$-axis for four different models (see label) at four different timesteps. All runs show similar values in the midplane. For runs A, C1 and D the profile tends to become shallower with time illustrating the transport of the magnetic flux toward higher altitude.[]{data-label="Pmag_z"}](LS_grav_pmsink/Pmag_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Magnetic pressure profile along the $z$-axis for four different models (see label) at four different timesteps. All runs show similar values in the midplane. For runs A, C1 and D the profile tends to become shallower with time illustrating the transport of the magnetic flux toward higher altitude.[]{data-label="Pmag_z"}](LS_grav_pmsink_sn_dens3_bis/Pmag_z.jpg "fig:"){width="8cm"}]{}
We show four timesteps for each runs. The first one is at about 25-30 Myr, the second at 40 Myr and the third at about 55-60 Myr. For the fourth we select a timestep of about 90 Myr for runs C1 and C2 and about 65-70 Myr for run A, D and NF1. The first timesteps have been chosen at the beginning of the star forming phase (see Fig. \[mass\_sink\_tot\]) and the second at about 10 Myr later because the SFR is typically close to its maximum. For the third timesteps the SFR is nearly constant with time and thus the profiles correspond to the quasi-stationnary regime. This is confirmed by the last timestes which show no significant evolution with respect to the third one although for runs A, D and NF1 the evolution is faster because accretion is higher and the profiles keep evolving rapidly at later times.
For the sake of conciseness, we will select the runs which we think are most relevant to emphasize the impact of the physics and of the schemes. When investigating the global disk structure (sect. \[global\]), we concentrate mainly on runs NF1, A, C1, C2 and D. When we investigate the multiphase ISM (sect. \[multiphase\]), we restrict to runs C1 and C2. The other runs (NF2, C3 and C4) are used to quantify the influence of the supernovae scheme on the star formation rate (sect. \[sfr\]).
Global structure {#global}
================
In this part we characterize the global structure of the galactic disks.
Qualitative description
-----------------------
Figure \[image\] shows the column density field along the $z$-axis (left panels) and $y$-axis (right panels) for MHD run C1 (upper panels) and hydrodynamical run C2 (lower panels). In both cases, the disk is clearly visible although its structure is quite irregular and varies significantly from place to place. The disk is slightly thicker in the magnetized run than in the hydrodynamical one, which is a clear consequence of the magnetic support. The column density distributions appear different with and without magnetic field. In particular, small scale fluctuations are more pronounced in the hydrodynamical run. This trend is also similar to what has been found at smaller scales [e.g. @h+2008] and has been interpreted by @h2013 to be a consequence of the Kelvin-Helmholtz instability being stabilized by the magnetic field [@ryu+2000]. As will be discussed later in the paper, this has consequences on the mass of self-gravitating objects, which form.
Figure \[image2\] shows column densities for the run without supernovae feedback, run NF1 (upper panels), and with our feedback scheme D as described above. The disk structure is very different in both cases. When no feedback is applied, very long filaments develop across the computational box. They converge towards a region where most of the mass accumulates. There are much less dense regions compared to run C1 and the disk is obviously thinner. This behaviour is very similar to what is reported in @hopkins+2011 (see for example their Fig. 2) On the other hand, scheme D leads to a disk whose structure is much more irregular than the structure of the disk obtained with scheme C although the number of supernovae and their energy are identical in the two cases. This constitutes a clear confirmation to the works of previous authors [e.g. @bournaud+2010; @dobbs+2011; @hopkins+2011] that feedback is playing a crucial role for the structure of galactic disks and (see next section) for regulating star formation. It is also clear that the correlation between the gas and the supernovae does influence significantly the galactic disk structure.
Disk density profile
--------------------
One fundamental aspect for the galactic structure is the density profile and the typical thickness of the gas distribution. In the case of the Milky Way, it has been measured (e.g. Ferrière 2001) that the different phases have different distributions. They tend to be roughly Gaussian but their thickness vary. The molecular gas has a full width at half maximum (FWHM) of about 120 pc while for the atomic gas it is about 230 pc. Both the molecular and the atomic gas have a mean density of about 0.5 cm$^{-3}$ for a total of 1 cm$^{-3}$.
Figure \[rho\_z\] shows the density distribution along the $z$-axis for five different models. The run C1 presents a density of about 3-4 cm$^{-3}$ and a FWHM of about 50 pc. This is also roughly the case for the hydrodynamical run. Comparison with run C1b shown in the appendix reveals that numerical resolution may be partly at the origin of this discrepency. In the run based on scheme D, the maximum density varies with time because of stronger accretion but is around 1-2 cm$^{-3}$ after 40 Myr. The FWHM is larger than in the previous cases and equal to about 100 pc. When no feedback is included, the disk tends to be thinner. For example at times 55.8 and 64.8 Myr, it is about 30 pc. A similar distribution is found at time 57.1 and 65.4 Myr with scheme A, which as we will see later has an accretion behaviour very similar to the run without feedback.
Pressure support
----------------
Since the disk thickness is a direct consequence of the various supports, we present the profile along $z$-axis of the three relevant pressures, namely thermal, kinetic and magnetic ones noted $P_{th}$, $P_{kin}$, $P_{mag}$ respectively. While $P_{th}$ and $P_{mag}$ have their standard definitions, $P_{kin}$ is taken as $$\begin{aligned}
P_{kin} = { \sum v_z^2 \rho dV \over \sum dV }. \end{aligned}$$
Figures \[Pth\_z\], \[Pkin\_z\] and \[Pmag\_z\] show $P_{th}$, $P_{kin}$, $P_{mag}$ respectively, as a function of altitude.
In all models, the thermal pressure ranges between a few $10^{-13}$ and a few $10^{-12}$ erg cm$^{-3}$, which corresponds to the typical pressure in the ISM [e.g. @ferriere+2001]. Its variation with the altitude, $z$, closely follows the density variation with the notable exception of runs NF1 and A (i.e. no or randomly distributed supernovae). For this latter, the thermal pressure is roughtly constant up to an altitude at which the density is about 0.1 cm$^{-3}$. This is clearly due to the an efficient production of warm and hot gas by supernovae explosions. Note that run D presents more variability than run C1. This is likely a consequence of the supernovae being less spatially correlated to the sink particles. When a supernova explodes in a dense regions, it tends to mimic what happens for most supernovae in run C1 but when it explodes in a diffuse regions, it tends to mimic run A where most supernovae explode in the WNM which has the largest volume filling factor.
The kinetic pressure is typically a few times larger than the thermal one and depending of the model, reaches values of the order of $\simeq 1-3 \times 10^{-12}$ erg cm$^{-3}$. These values are also very similar to what is reported in related studies [e.g. @joung+2009; @kim+2011]. Interestingly, the scale height of $P_{kin}$ is larger than the density scale height by a factor on the order of 1.5-2 for run C1 and up to 3-4 for run C2. In the case of scheme D, it even slightly increases with altitude. As expected, in the absence of supernovae feedback, $P_{kin}$ drops to small values rapidly.
The magnetic pressure is comparable to the thermal pressure and reaches values of the order of a few $10^{-12}$ erg cm$^{-3}$. It therefore contributes to support the galactic disk against gravity. Interestingly, while the magnetic intensity tends to decrease in the midplane as time goes on, it tends to increase with time at high altitude. This is a consequence of the generation of magnetic field through turbulence but also a consequence of the transport of the field lines by galactic outflows.
(0,9.5) (0,4.5)[![Density PDF for MHD run C1 (upper panel) and hydrodynamical run C2 (lower panel). The small drop at $n \simeq 2-3$ cm$^{-3}$ corresponds to the thermally unstable regime, which persists in spite of the strong turbulence. The two peaks correspond to the WNM and CNM phases.[]{data-label="mass_dens"}](LS_grav_pmsink_sn_dens3/mass_dens.jpg "fig:"){width="8cm"}]{} (0,0)[![Density PDF for MHD run C1 (upper panel) and hydrodynamical run C2 (lower panel). The small drop at $n \simeq 2-3$ cm$^{-3}$ corresponds to the thermally unstable regime, which persists in spite of the strong turbulence. The two peaks correspond to the WNM and CNM phases.[]{data-label="mass_dens"}](LS_grav_pmsink_sn_dens3_hydro/mass_dens.jpg "fig:"){width="8cm"}]{}
(0,9.5) (0,4.5)[![Mean temperature as a function of density for run C1 (upper panel) and C2 (lower panel). As expected the temperature lies mainly in 3 ranges of temperature namely $10^6$, $10^4$ and $10^2$ K, corresponding to the three phases of the ISM, the HIM, the WNM and the CNM.[]{data-label="T_dens"}](LS_grav_pmsink_sn_dens3/T_dens.jpg "fig:"){width="8cm"}]{} (0,0)[![Mean temperature as a function of density for run C1 (upper panel) and C2 (lower panel). As expected the temperature lies mainly in 3 ranges of temperature namely $10^6$, $10^4$ and $10^2$ K, corresponding to the three phases of the ISM, the HIM, the WNM and the CNM.[]{data-label="T_dens"}](LS_grav_pmsink_sn_dens3_hydro/T_dens.jpg "fig:"){width="8cm"}]{}
(0,18) (0,8)[{width="9cm"}]{} (9,8)[{width="9cm"}]{} (0,0)[{width="9cm"}]{} (9,0)[{width="9cm"}]{}
(0,9.5) (0,4.5)[![Root mean square value of $v_z$ as a function of $z$ (see text) for MHD run C1 and hydrodynamical run C2. Typical values are about 4-5 km s$^{-1}$ at the midplane.[]{data-label="rmsVz_z"}](LS_grav_pmsink_sn_dens3/rmsVz_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Root mean square value of $v_z$ as a function of $z$ (see text) for MHD run C1 and hydrodynamical run C2. Typical values are about 4-5 km s$^{-1}$ at the midplane.[]{data-label="rmsVz_z"}](LS_grav_pmsink_sn_dens3_hydro/rmsVz_z.jpg "fig:"){width="8cm"}]{}
(0,9.5) (0,4.5)[![Root mean square velocity as a function of density for MHD run C1 and hydrodynamical run C2. []{data-label="Vrms_dens"}](LS_grav_pmsink_sn_dens3/Vrms_dens.jpg "fig:"){width="8cm"}]{} (0,0)[![Root mean square velocity as a function of density for MHD run C1 and hydrodynamical run C2. []{data-label="Vrms_dens"}](LS_grav_pmsink_sn_dens3_hydro/Vrms_dens.jpg "fig:"){width="8cm"}]{}
(0,9.5) (0,4.5)[![Mean magnetic intensity as a function of $z$ and as a function of density. Note in particular that typical values of about 5 $\mu$G are being obtained in the midplane. Between 1 and $10^3$ cm$^{-3}$, the magnetic intensity weakly varies with the density.[]{data-label="mag"}](LS_grav_pmsink_sn_dens3/B_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Mean magnetic intensity as a function of $z$ and as a function of density. Note in particular that typical values of about 5 $\mu$G are being obtained in the midplane. Between 1 and $10^3$ cm$^{-3}$, the magnetic intensity weakly varies with the density.[]{data-label="mag"}](LS_grav_pmsink_sn_dens3/B_dens.jpg "fig:"){width="8cm"}]{}
Multiphase ISM {#multiphase}
==============
In this section, we study the density, temperature and magnetic field distribution of the gas in the simulations. For conciseness, we restrict our attention to runs C1 and C2 only stressing the most interesting differences with other runs. The comparison between runs C1 and C2 emphasizes the role that the magnetic field has to determine the characteristic of the ISM.
Density and temperature distributions
-------------------------------------
As it is the case in other works [e.g. @deavillez+2005; @joung+2006; @hill+2012], the interstellar medium which is initially uniform in density and temperature, quickly breaks up into a multiphase medium in which the density varies from less than 10$^{-3}$ cm$^{-3}$ to almost 10$^3$ cm$^{-3}$ while the temperature can be as high as 10$^7$ K and as low as a few tens of Kelvin.
Figure \[mass\_dens\] shows the mass contribution of the various gas densities in the computational box. It is dominated by the dense gas ($n > 10$ cm$^{-3}$) but a non-negligible fraction lays at lower densities, which corresponds to the WNM regime and thermally unstable gas ($0.1 < n < 10$ cm$^{-3}$). Interestingly, there is a small dip in the thermally unstable regime ($n \simeq 2-3$ cm$^{-3}$). This indicates that in spite of the relatively high level of turbulence, typical of galactic disks (see below), the 2-phase structure [@wolfire+2003] that would be obtained in a static medium, is not erased by the dynamical processes and able to persist. At early time there is a little more dense gas in the hydrodynamical case. This is due to the magnetic support that reduces the amount of self-gravitating gas. For run D (not displayed here for conciseness), the transition between WNM and cold neutral medium (CNM) is less pronounced, which is a consequence of the stronger turbulence in this run. These behaviours are reminiscent of what has been found in colliding flow calculations, which attempted to model the ISM at scales on the order of 10-50 pc [e.g. @ah05; @vazquez+2006; @h+2008; @heitsch+2008; @banerjee+2009; @ah+2010; @inoue+2012] and also in similar supernovae regulated galaxy simulations [e.g. @dib+2006; @hill+2012; @kim+2013]. In particular, in these simulations it has been found that the ISM quickly breaks up into a multi-phase, clumpy medium where the 2-phase behaviour (i.e. present an excess of gas in thermodynamical states close to the two stable branches of thermal equilibrium) is maintained even though the medium is largely turbulent.
Figure \[T\_dens\] displays the temperature distribution as a function of density. It clearly shows the existence of 3 main domains corresponding to the hot ionised medium ($T \simeq 10^6$ K), the warm neutral medium ($T \simeq 10^4$ K), and the cold neutral medium ($T<100$ K). This is in good agreement with the classical 3 phases model of the interstellar medium as early described for example in @mckee-ostriker1977. As noted by previous authors [e.g. @gazol+2001], there is gas in the thermally unstable regions ($T\simeq 10^3 K$), whose existence is permitted by the turbulent motions. The two runs present similar distributions although the transition between the warm and the cold phase (at density 1-10 cm$^{-3}$) is a little more shallow for the MHD run (C1) than for the hydrodynamical one (run C2). This is because the magnetic field contributes to the total pressure and can therefore stabilize the pieces of gas that are thermally unstable. Again in run D, the 3 regimes are less clearly separated though the global temperature range in the simulation is similar.
The spatial distribution is illustrated by Fig. \[im\_dens\_T\], which displays a cut through the equatorial plane of the density and temperature fields both for MHD run C1 and hydrodynamical run C2. As can be seen, most of the volume is found to be occupied by the warm neutral gas with temperature of the order of 10$^4$ K and densities of the order of 1 cm$^{-3}$. The hot gas, produced within supernovae explosions, occupies only a small fraction of the volume. Interestingly, the structures in the hydrodynamical and MHD runs have a slightly different shape. As already discussed in the previous section, the dense clouds in the hydrodynamical simulations tend to be more fragmented and on average slightly smaller (see also Fig. \[image\]).
Velocity dispersion
-------------------
In the supernovae regulated numerical simulations [e.g. @slyz+2005; @deavillez+2005; @joung+2006; @hill+2012; @kim+2011; @kim+2013; @gent+2013] the kinetic energy, which decays through the turbulent cascade, is replenished by the supernovae explosions. The velocity dispersion in the computational box is the result of a balance between injection and dissipation as emphasized for example in @maclow+2004, who present orders of magnitude suggesting that supernovae explosions can explain the velocity dispersion observed in the Milky Way to be of the order of 6 km s$^{-1}$.
Figure \[rmsVz\_z\] displays the rms $z$-component of the velocity field weighted by density $$\begin{aligned}
\sigma_z = \sqrt{ { \sum v_z^2 \rho dz \over \sum \rho dz } }.
\label{eq_rmsVz_z}\end{aligned}$$
Close to the equatorial plane, $\sigma _z$ is of the order of 4-5 km s$^{-1}$ for run C1 and 5-6 km $^{-1}$ for run C2. At higher altitude, the velocity dispersion increases to values of about 8-10 km s$^{-1}$ at 100 pc for run C1 and 8-15 km s$^{-1}$ for run C2. This is essentially due to the gas density getting lower at higher altitude. These values are again similar to what has been previously reported in similar studies and can be understood by relatively simple considerations. Following @maclow+2004, we can simply estimate the amount of mechanical energy which dissipates in the turbulent cascade as $\dot{E}_{diss} = M \sigma^2 / \tau$ where $M$ is the total mass of the system, $\tau$ is the crossing time and $\sigma$ the total velocity dispersion. Assuming that $\tau = h / \sigma$, where $h$ is the disk scale height, we get $\dot{E}_{diss} = M \sigma^3 / h$ where $\dot{E}_{diss}$ is the energy dissipated per units of time. The amount of energy which is injected by the supernovae into the system is simply $\dot{E}_{inj} = \epsilon \dot{N}_{sn} \times 10^{51} \, \rm{erg}$, where $\dot{N}_{sn}$ is the density of supernovae per units of time and $\epsilon$ is the efficiency at which turbulence is triggered. Equating these two rates, we get $$\begin{aligned}
\sigma = \left( {\epsilon \dot{N}_{sn} h \times 10^{51} \, \rm{erg} \over M } \right)^{1/3}.\end{aligned}$$ To estimate the value of $\sigma$, we take values typical for the Milky way. These values are also representative of our simulation parameters. We take a mass of $10^{10} M_\odot$, a frequency of supernovae $\dot{N}_{sn}=1/50$ yr$^{-1}$, a height $h=100$ pc and an efficiency $\epsilon=0.1$, we get $$\begin{aligned}
\sigma \simeq 8 \, {\rm km \, s^{-1}} \left( {\epsilon \over 0.1 } {\dot{N}_{sn} \over 1/50 {\rm yr}^{-1} } {h \over 100 {\rm pc}}
{10^{10} M _\odot \over M } \right)^{1/3}.\end{aligned}$$ This value is thus in good agreement with the velocity dispersion inferred from the simulations and from the observations. It is worth stressing that due to the weak dependence in all parameters (to the power 1/3), it is relatively unsurprising to find that the velocity dispersion generally does not undergo large variations. We note that the fluctuations in the hydrodynamical run C2 (lower panel) appear to be quite large with respect to the MHD run C1 (upper panel). This is likely a consequence of the higher SFR (see below) found in the hydronamical case. This results in a stronger feedback.
Finally, since stars mainly form in the dense gas, it is important to understand the star formation process, to know more accurately how velocity dispersion depends on the gas density. For that purpose Fig. \[Vrms\_dens\] displays the rms velocity field (taking into account the 3 components) weighted by density as a function of density in the whole computational box. As expected the velocity dispersion is weaker in the dense gas than in the diffuse one by a factor of about 2. Typical velocity dispersion in the dense gas is on the order of 4-5 km s$^{-1}$.
Distribution of magnetic intensity
----------------------------------
Figure \[mag\] shows the magnetic intensity as a function of altitude (upper panel) and density (lower panel). At the mean density of the galactic disk, i.e. $n \simeq 2-3$ cm$^{-3}$ the magnetic intensity is about 3-5 $\mu$G, which is also the mean value up to an altitude of about 100 pc. This is coherent with the values of about 5 $\mu$G reported by @heiles+2005. At higher densities, the magnetic intensity increases and reaches values of about 10$\mu$G at densities of about $10^2$ cm$^{-3}$. Note that this corresponds to a rather shallow variation of the magnetic intensity with density as observed in the diffuse gas [e.g. @troland+1986]. The exact reason of this weak correlation is most likely due, on the one hand to the Lorentz force, that resists contraction perpendicular to the field lines [@h+2000; @passot+2003]. On the other hand, it is also partly due to the turbulent diffusivity, which has also been observed to play an important role in numerical simulations . For densities below $n \simeq 10^{-2}$ cm$^{-3}$, a steep drop is observed with density. This is due to the fast expansions produced by supernovae explosions, which tend to dilute the magnetic intensity very significantly.
Star formation rate, sink mass function and outflows {#sfr}
====================================================
We now investigate the characteristics of the star formation in the simulations. This is achieved through the sink particles described in sect. \[code\]. We first quantify the total mass of the sink particles, which represents the star formation rate in the simulations. We then study the mass distribution, i.e. the sink mass function of some of our models. Finally, we study the outflows which are launched at high altitude and eventually escape the computational box.
Star formation rates
--------------------
(0,15) (0,10)[![Total mass of sink particles as a function of time for the various models. Upper panel shows the influence of the supernova feedback scheme while middle and lower panels show the influence of the magnetic field. Lower panel shows the SFR (i.e. time derivative of mass) corresponding to middle panel. In upper panel, run B corresponds to the solid curve which starts at 120 Myr. The runs with feedback present SFR that are typically 10 to 30 times smaller. Magnetic field reduces the SFR by a factor on the order of 2.[]{data-label="mass_sink_tot"}](mass_sink_mod.jpg "fig:"){width="9cm"}]{} (0,5)[![Total mass of sink particles as a function of time for the various models. Upper panel shows the influence of the supernova feedback scheme while middle and lower panels show the influence of the magnetic field. Lower panel shows the SFR (i.e. time derivative of mass) corresponding to middle panel. In upper panel, run B corresponds to the solid curve which starts at 120 Myr. The runs with feedback present SFR that are typically 10 to 30 times smaller. Magnetic field reduces the SFR by a factor on the order of 2.[]{data-label="mass_sink_tot"}](mass_sink_mag.jpg "fig:"){width="9cm"}]{} (0,0)[![Total mass of sink particles as a function of time for the various models. Upper panel shows the influence of the supernova feedback scheme while middle and lower panels show the influence of the magnetic field. Lower panel shows the SFR (i.e. time derivative of mass) corresponding to middle panel. In upper panel, run B corresponds to the solid curve which starts at 120 Myr. The runs with feedback present SFR that are typically 10 to 30 times smaller. Magnetic field reduces the SFR by a factor on the order of 2.[]{data-label="mass_sink_tot"}](sfr_mag.jpg "fig:"){width="9cm"}]{}
Figure \[mass\_sink\_tot\] shows the total mass of the sink particles as a function of time in the simulations. Upper panel shows the influence of the sink particle prescription while lower panel shows the influence of the magnetisation. Before commenting on the difference between the various models, we first discuss the main trends and numbers. For all models (except run B) accretion onto sink particles starts between 20 and 30 Myr. In about 10-20 Myr, the total accreted mass varies between a few 10$^5$ to $\simeq$10$^7$ $M_\odot$. As discussed below, these differences are due to the various feedback prescriptions and magnetisations. At later times, all models tend to reach a phase of stationary accretion at a rate which ranges from about 10$^{-2}$ to $\simeq 10^{-1}$ $M_\odot$ yr$^{-1}$. It is worth comparing these values with the typical 3 $M_\odot$ yr$^{-1}$ at which Milky way is forming stars. In order to do so, one must first correct for the volume of our computational box. Since in the Milky way most stars form inside the solar circle whose radius is about 8 kpc and since the size of the computational domain is equal to 1 kpc, a geometrical factor of $\pi \times 8^2 \simeq 200$ should be taken into account. However, it should also be accounted for the fact that the efficiency of the mass eventually accreted into the stars is only a fraction of the mass that is accreted onto the cores. This value is not known with great accuracy but has been estimated to be of the order of 1/3 . We note that the sink particles used in this study are at this stage much larger than dense molecular cores. Therefore it could be that the efficiency should be even lower than this value. Combining these two numbers, we find that for a galaxy like the Milky way, our models would predict a star formation rate ranging from about 1 to 20 $M_\odot$ yr$^{-1}$. Given the large uncertainties, the first value appears to be in reasonable agreement with the galactic one.
### Influence of feedback prescriptions
All models displayed in top panel of Fig. \[mass\_sink\_tot\] have initial conditions identical to run C1, i.e. have an initial magnetic field whose initial intensity in the midplane is about 2.5$\mu$G.
First of all, the large difference between the solid line (run NF1) and the dashed line (run C1) confirms the drastic influence of the feedback on the star formation rate which is reduced by a factor of 20-30. This constitutes a strong hint that feedback can be largely responsible to solve the long standing issue of the so-called Zuckermann-Evans catastroph [@zuckerman+1974]. If all the molecular gas of the Milky way was collapsing in a free-fall time, about 300 $M_\odot$ yr$^{-1}$ of stars would form in the Galaxy.
Second of all, when the supernovae are not correlated to density (run A) not only is the feedback unable to reduce the star formation rate but this latter is even slightly higher. This is because since most of the volume is occupied by warm gas, most of the supernovae therefore explode in low density regions. Their net effect is thus to further compress the dense gas. When the supernovae are correlated with the density peak (run B), it takes a long time before sinks can form because the dense gas is efficiently dispersed. However, once sinks start forming, since the supernovae are not correlated locally in space and in time with accretion but simply with the densest cell in the simulation, they are unable to reduce the accretion rate. Therefore star formation rates comparable to the run without feedback are obtained.
Third, the runs C4 and D show that SFR larger by a factor 2-3 are obtained when the feedback is either purely thermal or less tightly correlated to the sink particles. Given that these two aspects are largely uncertain, this illustrates the limit of this modeling and suggests that the typical accuracy of these models is at best on the order of a factor 2-3. Note that another severe source of uncertainties comes from the time at which supernovae are introduced. In particular, if a delay of tens of Myr is introduced, SFR comparable to the ones of run NF1 are obtained. This suggests that in order to get more accurate models, it is necessary to have a better description of the small scales and in particular of the formation and evolution of massive stars up to the point where they explode. Ideally, this would require running a set of specific small scale simulations to quantify more accurately the impact of the feedback.
### Influence of magnetisation
All models displayed in middle and bottom panel of Fig. \[mass\_sink\_tot\] are performed either with no feedback (runs NF1 and NF2) or with the same feedback scheme (scheme C). Different levels of magnetisation are compared.
In the hydrodynamical run C2, stars start forming a few Myr before run C1. The SFR is initially significantly reduced compared to run C2. At later time, they become however comparable. These effects are a consequence of the magnetic support which contribute to resist the gravitational contraction but also to the density PDF that is narrower in the presence of a magnetic field [@molina+2012]. A similar effect is obtained for the two runs without feedback (runs NF1 and NF2) for which it is seen that the SFR is a little higher in the hydrodynamical case than in the MHD one.
Interestingly, even when the magnetic field is rather weak (0.5 $\mu$G initially), it still has a visible impact and reduces the SFR by a factor of about $50 \%$ during the first 20 Myr after star formation has started. This is because magnetic field is quickly amplified to larger values.
This effect is quantitatively comparable to what has been inferred at smaller scales by various teams who investigate star formation in substancially magnetized, though supercritical clouds. For example @price+2008 simulate the collapse of a self-gravitating clump while @dib+2010 and @padoan+2011 perform self-gravitating, MHD calculations within periodic boxes. They all infer that magnetic field reduces the SFR by a factor of about 2. The exact reason of this lower value has not been analysed in great details so far but it is likely a consequence of the magnetic support which tends to resist gravity and the somehow narrower density PDF which tends to reduce the SFR [@hc2013].
Mass function of sink particles
-------------------------------
(0,20) (0,15)[![Sink mass spectra for four models at four timesteps. In the case with no feedback there are typically few massive sink particles while a broader distribution develops when feedback is included.[]{data-label="sink_mass_spectrum"}](LS_grav_pmsink_sn_dens3/sink_mass_spectrum.jpg "fig:"){width="9cm"}]{} (0,10)[![Sink mass spectra for four models at four timesteps. In the case with no feedback there are typically few massive sink particles while a broader distribution develops when feedback is included.[]{data-label="sink_mass_spectrum"}](LS_grav_pmsink_sn_dens3_hydro/sink_mass_spectrum.jpg "fig:"){width="9cm"}]{} (0,5)[![Sink mass spectra for four models at four timesteps. In the case with no feedback there are typically few massive sink particles while a broader distribution develops when feedback is included.[]{data-label="sink_mass_spectrum"}](LS_grav_pmsink_sn_dens3_bis/sink_mass_spectrum.jpg "fig:"){width="9cm"}]{} (0,0)[![Sink mass spectra for four models at four timesteps. In the case with no feedback there are typically few massive sink particles while a broader distribution develops when feedback is included.[]{data-label="sink_mass_spectrum"}](LS_grav_pmsink/sink_mass_spectrum.jpg "fig:"){width="9cm"}]{}
Figure \[sink\_mass\_spectrum\] shows the sink mass function for various models and at four timesteps from which one can verify that the trends discussed below are not due to a time selection. For runs C1 and C2, a large number of sinks form (about 400 and 700 respectively for run C1 and run C2). Their masses span about 3 orders of magnitude. Given the limited numerical resolution of the present study many features of the distribution must be taken with great care. In particular the peak at about 10$^3 \, M_\odot$ would certainly shift to smaller values in more resolved runs [e.g. @ha07]. There is a possible trend for a powerlaw developing at large masses (in the range $\simeq 3 \times 10^3-10^4 \, M_\odot$) with an exponent compatible with $\simeq -1$. However, the limited resolution precludes a firm conclusion. We note that in massive collapsing magnetised clumps, the number of fragments has also been found to be reduced by a factor of about two. This is a clear consequence of the cold gas being more coherent and less fragmented. As noted previously, the reason is that the magnetic field makes the flow more coherent since it tends to connect fluid particles linked by the magnetic field lines [e.g. @h2013].
The sink mass function obtained when no feedback (run NF1) is included is quite different. There are much less sink particles (about 70) and most of them have a mass larger than $10^4 \, M_\odot$. Indeed, the most massive sink particle has in this case a mass equal to a few 10$^6 M_\odot$. This behaviour is again a consequence of the absence of feedback. The gas tends to concentrate in a few locations under the influence of gravity. The sink mass function obtained for run D is inbetween the one obtained for run C1 and run NF1. This illustrates again the fact that in run D, the stellar feedback is less efficient in supporting the gas against gravitational collapse.
(0,18.5) (0,13.5)[![Mean flux of mass along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for scheme D.[]{data-label="flux_z"}](LS_grav_pmsink_sn_dens3/flux_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Mean flux of mass along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for scheme D.[]{data-label="flux_z"}](LS_grav_pmsink_sn/flux_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Mean flux of mass along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for scheme D.[]{data-label="flux_z"}](LS_grav_pmsink_sn_dens3_bis/flux_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Mean flux of mass along the $z$-axis for five different models (see label) at four different timesteps. The largest values are obtained for scheme D.[]{data-label="flux_z"}](LS_grav_pmsink_sn_dens3_hydro/flux_z.jpg "fig:"){width="8cm"}]{}
Galactic outflows
-----------------
The existence of galactic outflows is now well established [see @veilleux+2005 for a recent review] in many galaxies. The typical scale height at which these outflows are observed, is of the order of several tens of kpc, which is much larger than the scale of the present simulation. Also the box length is equal to only 1 kpc and we do not have a proper halo structure which influences the flow launching [e.g. @dubois+2008]. It is nevertheless worth to quantify them, in particular because supernovae are believed to be largely responsible of their launching.
Figure \[flux\_z\] shows the mass flux as a function of the altitude $z$ for four timesteps and four models. As can be seen it varies significantly with time and altitude from typically a few $10^{-2}$ to a few $10^{-1}$ M$_\odot$ yr$^{-1}$. Taking into account that the surface of the box is 1 kpc$^2$, this would lead to a flux of about $\pi \times 8^2 \simeq 200$ times larger for a galaxy similar to ours, i.e. a few solar mass to a few tens or solar mass per year. These values are typical of what is measured for galactic outflows [@veilleux+2005].
Another interesting trend is that the mass flux broadly correlates with the SFR (see fig. \[mass\_sink\_tot\]). For run C1 (top panel), the peak value of the mass flux is about 3 times smaller than what is obtained for run D (third panel). Comparing the mean value at the computational box edges ($z$=500 pc), the ratio between the fluxes of run C1 and D leads to somewhat larger values of about 5-10. This is likely a consequence of the dual role of supernovae explosions, which are responsible for the regulation of star formation through energy and momentum injection in the dense gas but also for the launching of the galactic outflows through injection onto the diffuse gas. Since the two processes are linked, it is expected that larger SFR lead to stronger outflows as they imply more feedback. Indeed the SFR ratio for run D and C1 is about 3 (from upper panel of Fig. \[mass\_sink\_tot\]), which is comparable with the value of 3 quoted above but a little to small to explain the second value obtained at the box edges. This may indicate that another effect must be considered. We believe that since in run D the supernovae explode further from the sink particles than in run C1, more energy and momentum tend to be injected in the diffuse gas than in run C1. Since the outflows are primarily made by diffuse fastly expanding material, it seems reasonable that the efficiency with which they are produced is higher in run C1 than in run D.
Conclusion
==========
We have performed a series of numerical simulations describing a galactic disk regulated by supernovae feedback at kpc scale. Our simulations include both magnetic field and self-gravity. In particular we have explored the influence of various schemes to prescribe the supernovae feedback. Our simulations reproduce many features already found by other authors such as multi-phase density and temperature distributions or velocity dispersion, typically of the order of 5 km s$^{-1}$ in the galactic plane. Our results are as follows. When the supernovae are randomly distributed they drive the interstellar turbulence but are unable to resist self-gravity efficiently and the star formation rate is as high (even slightly higher) as when no feedback is included. When supernovae are correlated to the density peaks, they efficiently limit star formation by preventing the gas to become too dense. However as time goes on, dense gas eventually develops. When sink particles are being introduced, then the star formation rate is as high as its value without feedback.
When supernoave are spatially and temporally correlated to star formation events, the star formation rate is significantly reduced by a factor of the order of ten or more. However, we find that the exact implementation of the supernovae does influence the galactic disk structure and the star formation rate significantly. In particular, if the supernoave are distributed in a shell of about 16 pc around the sink particles, the accretion rate is higher by a factor of about 3 than if they are randomly placed within a sphere of radius equal to 16 pc. In a similar way, if the feedback is purely thermal, the star formation rate is about 2 times larger than if it has 5% of kinetic feedback. This implies that a detailed knowledge of how the feedback operates on small scales is mandatory to understand its impact with sufficient precision. In particular, the correlation between the massive stars and the dense star forming gas should be determined using small scale simuations.
Magnetic field has a significant impact. It delays and reduces star formation by a factor of the order of 2. It also tends to reduce the number of star formation regions (e.g. sink particles) by a factor of about 2 therefore producing slightly bigger star forming regions. Finally, it should be kept in mind that magnetic field has an important impact on the fragmentation of massive cores that it tends to reduce significantly [@commercon+2011; @myers+2013]. This implies that more massive stars form when magnetic field is strong. Since feedback is a non-linear function of the stellar masses and since feedback influences drastically the galactic structure and evolution, it is likely the case that the impact magnetic field has on galaxy evolution is probably even larger than what is estimated here.\
\
*Acknowledgments* We thank the anonymous referee for a careful reading of the manuscript which has significantly improved the paper. This work was granted access to HPC resources of CINES under the allocation x2014047023 made by GENCI (Grand Equipement National de Calcul Intensif). PH acknowledge the finantial support of the Agence National pour la Recherche through the COSMIS project. This research has received funding from the European Research Council under the European Community’s Seventh Framework Programme (FP7/2007-2013 Grant Agreement no. 306483 and no. 291294).
Appendix: the issue of numerical convergence
============================================
(0,23.) (0,18)[![Mean density, $z$-velocity dispersion, kinetic, thermal and magnetic profile along the $z$-axis for run C1b at four different timesteps. These results should be compared with the corresponding quantities for run C1 displayed in Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\] and \[rmsVz\_z\].[]{data-label="runC1b"}](LS_grav_pmsink_sn_dens3_res2/rho_z.jpg "fig:"){width="8cm"}]{} (0,4.5)[![Mean density, $z$-velocity dispersion, kinetic, thermal and magnetic profile along the $z$-axis for run C1b at four different timesteps. These results should be compared with the corresponding quantities for run C1 displayed in Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\] and \[rmsVz\_z\].[]{data-label="runC1b"}](LS_grav_pmsink_sn_dens3_res2/Pth_z.jpg "fig:"){width="8cm"}]{} (0,9)[![Mean density, $z$-velocity dispersion, kinetic, thermal and magnetic profile along the $z$-axis for run C1b at four different timesteps. These results should be compared with the corresponding quantities for run C1 displayed in Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\] and \[rmsVz\_z\].[]{data-label="runC1b"}](LS_grav_pmsink_sn_dens3_res2/Pkin_z.jpg "fig:"){width="8cm"}]{} (0,0)[![Mean density, $z$-velocity dispersion, kinetic, thermal and magnetic profile along the $z$-axis for run C1b at four different timesteps. These results should be compared with the corresponding quantities for run C1 displayed in Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\] and \[rmsVz\_z\].[]{data-label="runC1b"}](LS_grav_pmsink_sn_dens3_res2/Pmag_z.jpg "fig:"){width="8cm"}]{} (0,13.5)[![Mean density, $z$-velocity dispersion, kinetic, thermal and magnetic profile along the $z$-axis for run C1b at four different timesteps. These results should be compared with the corresponding quantities for run C1 displayed in Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\] and \[rmsVz\_z\].[]{data-label="runC1b"}](LS_grav_pmsink_sn_dens3_res2/rmsVz_z.jpg "fig:"){width="8cm"}]{}
In order to investigate the issue of numerical convergence, we present the various profiles for run C1b, which is identical to run C1 but has an effective resolution 2 times larger. As can be seen from Fig. \[runC1b\], and Fig. \[rho\_z\], \[Pth\_z\], \[Pkin\_z\], \[Pmag\_z\], \[rmsVz\_z\] which displayed the results for run C1, the profiles present some moderate differences for the two cases implying that numerical convergence is not fully reached. In particular, the density profile is slightly less peaked for run C1b than for run C1. Similarly, the rms velocity is about 5 km s$^{-1}$ for run C1b while it is equal to about 4-4.5 km s$^{-1}$ for run C1. The pressures however present very comparable values and profiles between the two runs. Altogether, this shows that the quantities which characterize the disk structure are reasonably described at the resolution used in the paper.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Rate Adaptation (RA) is a fundamental mechanism in 802.11 systems. It allows transmitters to adapt the coding and modulation scheme as well as the MIMO transmission mode to the radio channel conditions, and in turn, to learn and track the (mode, rate) pair providing the highest throughput. So far, the design of RA mechanisms has been mainly driven by heuristics. In contrast, in this paper, we rigorously formulate such design as an online stochastic optimisation problem. We solve this problem and present ORS (Optimal Rate Sampling), a family of (mode, rate) pair adaptation algorithms that provably learn as fast as it is possible the best pair for transmission. We study the performance of ORS algorithms in both stationary radio environments where the successful packet transmission probabilities at the various (mode, rate) pairs do not vary over time, and in non-stationary environments where these probabilities evolve. We show that under ORS algorithms, the throughput loss due to the need to explore sub-optimal (mode, rate) pairs does not depend on the number of available pairs, which is a crucial advantage as evolving 802.11 standards offer an increasingly large number of (mode, rate) pairs. We illustrate the efficiency of ORS algorithms (compared to the state-of-the-art algorithms) using simulations and traces extracted from 802.11 test-beds.'
author:
- 'Richard Combes$^\dag$, Alexandre Proutiere$^{\dag,\times}$, Donggyu Yun$^\ddag$, Jungseul Ok$^\ddag$, Yung Yi$^\ddag$[^1]'
bibliography:
- 'RA.bib'
title: 'Optimal Rate Sampling in 802.11 Systems'
---
[^1]: $\dag$: KTH Royal Institute of Technology, Sweden. $\times$: INRIA/Microsoft joint research centre, Palaiseau, France. $\ddag$: Department of Electrical Engineering, KAIST, South Korea.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Here we report the fabrication and optical characterization of organic microcavities containing liquid-crystalline conjugated polymers (LCCPs): poly(9,9-dioctylfluorene-co-benzothiadiazole) (F8BT), poly(9,9-dioctylfluorene) (PFO) and poly(2,7-(9,9-dihexyl fluorene)-co-bithiophene) (F6T2) aligned on top of a thin transparent Sulfuric Dye 1 (SD1) photoalignment layer. We extract the optical constants of the aligned films using variable angle spectroscopic ellipsometry and fabricate metallic microcavities in which the ultrastrong coupling regime is manifest both for the aligned and non-aligned LCCPs. Transition dipole moment alignment enables a systematic increase in the interaction strength, with unprecedented solid-state Rabi splitting energies up to 1.80 eV for F6T2, the first to reach energies comparable to those in the visible spectrum; with an optical gap of 2.79 eV this also gives the highest-to-date organic microcavity coupling ratio, 65$\%$. We also demonstrate that the coupling strength is polarization-dependent with bright polaritons photoluminescence for TE polarization parallel to the transition dipoles and either no emission or weakly coupled emission from the corresponding TM polarization. The use of uniaxally aligned organic microcavities with switchable coupling strength offers exciting prospects for direct observations of ultrastrong coupling signatures, quantum simulation, polaritonics and condensation related phenomena.'
author:
- 'F. Le Roux'
- 'R. A. Taylor'
- 'D. D. C. Bradley'
bibliography:
- 'manuscriptACSbiblio.bib'
title: 'Exciton-Polaritons in Uniaxially Aligned Organic Microcavities'
---
The strong coupling (SC) regime in the solid-state is entered[@Weisbuch1992] when the interaction between the electric component of a confined electromagnetic field and the excitations present within a semiconductor becomes sufficiently intense that their original energy levels are replaced by so-called polariton hybrid states of light and matter, separated by a Rabi splitting energy $\hbar\Omega_{\rm R}$. Organic semiconductor Frenkel excitons are an interesting alternative to the more traditional Wannier excitons seen in III-V inorganic semiconductors for the study of exciton-polaritons thanks to their large binding energies ($E_{B}\sim0.5\pm0.25\;$eV[@Marks1994; @Alvarado1998]) which allow room-temperature observation of varied phenomena including Bose-Einstein condensation[@Plumhof2014; @Daskalakis2014], superfluidity of light[@Lerario2017] and optical logic[@Zasedatelev2019]. Their large intrinsic oscillator strengths[@Lidzey1998] combined with the small mode volumes $V_{\rm m}$ of metallic microcavities[@Schwartz2011] have enabled $\hbar\Omega_{\rm R} \geq 1\;$ eV[@Kena-Cohen2013; @Gambino2014; @Mazzeo2014; @Suzuki2019], with values up to $\hbar\Omega_{\rm R} = 1.12\;$eV [@Liu2015; @Liu2019]. This splitting is directly comparable to the exciton transition energy $\hbar\omega_{\rm ex}$ an yields normalized coupling ratios $g = \frac{\Omega_{\rm R}}{\omega_{\rm ex}}\geq20\%$, thereby crossing into ultrastrong coupling (USC), an interaction space that has received great recent attention, with attractive research perspectives and multiple emerging applications[@FriskKockum2019; @Forn-Diaz2019]. Experimental realizations of increasingly higher coupling ratios have also been reported for inorganic semiconductor based intersubband polaritons[@Askenazi2017], and other physical systems, including superconducting circuits[@Yoshihara2018], Landau polaritons[@Bayer2017] and plasmonic picocavities interacting with vibrational degrees of freedom of individual molecules[@Benz2016].
For an ensemble of organic semiconductor excitons within a cavity, $\hbar\Omega_{\rm R}$ scales with the square root of $\omega_{\rm ex}$ according to [@Ciuti2005; @George2015; @Tropf2018]:
$$\label{eq:Scaling}
\hbar\Omega_{\rm R} = 2\boldsymbol{{\mu}.E}\sqrt{\frac{N\hbar\omega_{\rm ex}}{2\epsilon_{\rm eff}V_{\rm m}}},$$
where $\boldsymbol{{\mu}}$ is the transition dipole moment, $\boldsymbol{E}$ the electric field, $N$ the number of molecules, $\hbar\omega_{\rm ex}$ the exciton transition energy, $\epsilon_{\rm eff}$ the cavity effective permittivity and $V_{\rm m}$ the cavity mode volume.
One way to increase the value of the coupling ratio has then been to work with lower energy excitons, as done by Barachati et al.[@Barachati2018], resulting in a then record $g = 62\%$. This approach is inherently accompanied by a reduction of $\hbar\Omega_{\rm R}$ compared with the use of excitons lying at higher energies with equivalent oscillator strengths. The alternative is to look to increase $\hbar\Omega_{\rm R}$; The most direct routes to achieve this include (i) increasing $N$, which although generally not straightforward can be done, for example, by reducing the bulkiness of conjugated polymer solubilizing groups[@Campoy-Quiles2008] (ii) increasing $\boldsymbol{{\mu}}$ through conformational control[@LeRoux2018; @Perevedentsev2016] or a photo-switchable configuration change[@Schwartz2011], and/or (iii) increasing $\boldsymbol{{\mu}.E}$. In the latter case, uniaxial orientation has been shown to enhance conjugated polymer thin film refractive index (and correspondingly transition dipole moment) in the direction parallel to the chain orientation axis[@Virgili2001; @Campoy-Quiles2005b] yielding an enhanced dot product for a suitable polarization of $\boldsymbol{E}$. This offers a clear route to enhancing $\hbar\Omega_{\rm R}$ that is demonstrated below, using a photoalignment process to achieve thermotropic liquid crystalline conjugated polymer (LCCP) chain orientation. Two recent reports on the coupling of liquid crystal (LC) vibrational modes[@Hertzog2017] and carbon nanotubes Wannier excitons[@Gao2018] have also shown that $\boldsymbol{{\mu}.E}$ can be maximized in this way and that polarization-dependence allows for applications discussed further in the text.
Previous approaches to LCCP orientation typically used a traditional rubbed polyimide (PI) alignment layer onto which the polymer was spin coated prior to thermal treatment[@Grell1997]. The clearing temperatures of LCCPs are relatively high ($\sim$200-300$^{\circ}$C)[@Grell1997; @GRELL1999], leading to the requirement for a high temperature stable PI, for which there are limited commercial options. Precursor route poly(p-phenylenevinylene) has also been used as an alternative rubbed alignment layer, having the advantage of temperature stability and an electronic structure that more readily permits charge injection from the underlying electrode to the LCCP[@Whitehead2000a]. Other approaches to orientation include stretching and rubbing the conjugated polymer[@Dyreklev1995], Langmuir-Blodgett deposition[@Cimrova1996], use of an aligned host matrix[@Hagler1991] or nanoimpriting[@ZijianZheng2007]. Successful fabrication of polarized light emitting diodes[@Whitehead2000; @Schmid2008; @Lee2017; @Whitehead2000; @Whitehead2000a; @Misaki2008] and polarized photoluminescence structures[@Virgili2001] has resulted, together with intrachain mobility enhanced transistors[@Sirringhaus2000]. For a variety of practical reasons, oriented LCCPs have, however, not been used before in strongly or ultrastrongly-coupled microcavities.
Non-contact photoalignment of LC mesophases has emerged[@Seki2014] as a promising alternative to rubbing-induced alignment. Among photoalignment layer materials, the azobenzene-containing Sulfuric Dye 1 (SD1) has shown high temperature stability and remarkable quality for the alignment of low molecular weight LCs[@Li2006; @Ma2015]. Recently, the orientation of poly(9,9-dioctylfluorene-co-benzothiadiazole) (F8BT) using SD1 has also been observed[@Zhang2019] in which a thermal treatment enables the orientation of the LCCP. Potential advantages of using SD1 photoalignment for photonics and polaritonics are threefold, namely that the SD1 layer can be very thin ($\leq 5\;$ nm), that it is almost transparent in the visible (peak absorption at $\sim$ 3.25 eV) and that it is a patternable process which allows a straightforward way to fabricate novel photonic structures[@Ma2015; @Ma2017].
We report here a detailed study of the use of oriented LCCP films within metallic microcavities. F8BT, poly(9,9-dioctylfluorene) (PFO) and poly(2,7-(9,9-dihexylfluorene)-co-bithiophene) (F6T2) films are oriented with SD1 photoalignment layers and in all three cases we demonstrate a systematic enhancement of $\hbar\Omega_{\rm R_{\rm TE}}$, for TE-polarized light parallel to the chain orientation direction, compared to non-aligned reference samples. The maximum $\hbar\Omega_{\rm R_{\rm TE}}= 1.80\pm 0.01\;$eV (689 nm) is for F6T2, a value that would sit within the visible spectrum. This structure also gives the largest normalized coupling ratio, $g = 65 \%$, reported to date for an organic semiconductor microcavity. Photoluminescence for TE-(parallel to the alignment direction) and corresponding TM-polarizations makes the changes in coupling strength between polarizations evident for all three polymers. Going beyond enhancement of the Rabi-splitting energy, we discuss the potential use of uniaxially aligned organic microcavities for demonstration of the elusive polaritonic NOT gate[@Sanvitto2016; @Solnyshkov2015; @Espinosa-Ortega2013], for quantum simulation through complex energy landscapes and more generally its advantages for the realization of polarization sensitive devices, lasing and condensation related phenomena.
Results And Discussion
======================
The optical constants for thin films of PFO, F8BT, F6T2 (see Methods for all fabrication protocols) were extracted using Variable Angle Spectroscopic Ellipsometry (VASE). As the polymer chains tend to lie in the plane of the film[@Campoy-Quiles2005a; @Campoy-Quiles2014; @Campoy-Quiles2005], the resulting optical constants are well fitted[@LeRoux2018; @Tropf2017] using an in-plane/out-of plane anisotropic model yielding the components $n_{\rm ord}, k_{\rm ord}, n_{\rm ex}, k_{\rm ex}$ of the complex refractive index $\tilde{n} = n + ik$. Figure \[fig:FigureIndices\] shows (in green) the in-plane optical constants ($n_{\rm ord}, k_{\rm ord}$) for F8BT (b), PFO (c) and F6T2 (d) (The complete sets of optical components extracted are available in Supplementary Information). All spectra are comprised of either one or several inhomogeneously broadened distributions ($E_{\rm X_{\rm PFO}}$ at around $3.23\;$eV, $E_{\rm X_{\rm F8BT_{1}}}$ and $E_{\rm X_{\rm F8BT_{2}}}$ at respectively $3.82\;$eV and $2.70\;$eV and $E_{X_{F6T2}}$ at $2.79\;$eV) with excitation states lying above $5\;$eV associated to ring-localized fluorene states[@Rothe2006].
Following photoalignment of an SD1 spincoated film in the in-plane y direction, the optical components $n_{x}, n_{y}, n_{z}, k_{x}, k_{y}, k_{z}$ were extracted using a biaxial anisotropic model[@Valyukh2008] and are shown in Figure \[fig:FigureIndices\](a). As expected, the alignment brings about intense $n_{y}$, $k_{y}$ optical components compared to their perpendicular counterparts $n_{x}$, $k_{x}$. Polymer layers of F8BT, PFO and F6T2 were then spincoated on top of the photoaligned SD1, thermally annealed into their respective nematic phases ($250^{\circ}$C for F8BT, $160^{\circ}$C for PFO and $220^{\circ}$C for F6T2) and subsequently quenched to room-temperature causing and effectively freezing the in-plane uniaxial alignment in the $y$ direction. The corresponding optical components of the obtained films were fitted and are shown in Figure \[fig:FigureIndices\](b),(c),(d) with $n_{y}$, $k_{y}$ substantially larger than $n_{x}$, $k_{x}$. As the oscillator strengths in the $x$ and $y$ direction $f_{x,y} \propto \int k_{x,y} (E)dE$ are proportional to the number of underlying contributing dipoles, we calculate the ratio $R= 1 - \frac{f_{x}}{f_{x} + f_{y}}$ where $f_{x,y}$ is the oscillator strength in the corresponding direction and estimate the total percentage of dipoles aligned in the $y$ direction. For all excitons, $R$ exceeds 83$\%$ (calculated for $E_{\rm X_{\rm F8BT_{1}}}$) with a maximum value of 94$\%$ (for $E_{\rm X_{\rm F6T2}}$) underlining the remarkable alignment quality.
![\[fig:FigureIndices\] (a) Optical components for a thin film of SD1 aligned along the y direction. (b),(c),(d) In green: In-plane $n_{\rm ord}$, $k_{\rm ord}$ optical components for spincoated F8BT (b), PFO (c) and F6T2 (d). The in-plane optical components of the films following alignment are shown in blue for $n_{y}$, $k_{y}$ (parallel to the alignment direction) and in red for $n_{x}$, $k_{x}$ (perpendicular to the alignment direction) for SD1 (a), F8BT (b), PFO (c) and F6T2 (d). Solid lines give the real component of the complex refractive index $\tilde{n} = n + ik$, dashed lines the imaginary component.](FigureIndices.png)
Time-integrated photoluminescence (PL) spectra were recorded at normal incidence for the three polymers and are shown in Figure \[fig:FigurePLFilms\] (see Methods for the experiment geometry). The spectra were measured at normal incidence with the collection polarizer both in vertical (blue) and horizontal (red) positions; for the aligned films, the vertical direction matched the direction of the alignment. In each case, we calculate the integrated ratio $R_{\rm VH} = \int \frac{I_{\rm V}(E)}{I_{\rm H}(E)}dE$ which reveals the presence of in-plane uniaxial alignment in the film. For F8BT, the spectrum of the non-aligned film (Figure \[fig:FigurePLFilms\] (a)) reveals an inhomogeneously broadened distribution with $S_{1}-S_{0}$ (0-0), (0-1) vibronic peaks located at $2.29\;$eV ($541\;$nm) and $2.15\;$eV ($577\;$nm). We calculate $R_{\rm VH} = 1.08$, with a deviation from unity being fully accounted for by the degree of polarization of the excitation laser beam; the polymer chains have as expected no preferential in-plane orientation. For the aligned F8BT (\[fig:FigurePLFilms\] (b)), the same spectral positions for the vibronic peaks are recorded with a difference in relative heights due to thickness variations between the aligned and non-aligned films. The integrated ratio $R_{\rm VH} = 8.3$ makes the in-plane preferential alignment of the emitting layer evident (in-depth studies relating the polarized emission to the microscopic parameters of aligned F8BT can be found in Ref 41). For non-aligned PFO (Figure \[fig:FigurePLFilms\](c)), the $S_{1}-S_{0}$ well resolved (0-0), (0-1) and (0-2) PL vibronic peaks appear at $2.88\;$eV ($430\;$nm), $2.71\;$eV ($457\;$nm) and $2.58\;$eV ($481\;$nm) with an integrated ratio $R_{\rm VH}$ that increases from 1.09 to 6.9 from non-aligned to aligned film (Figure \[fig:FigurePLFilms\](d)). As the vibronic structure is this time well resolved, we calculate the peak ratios $R_{\rm pVH}(E) = \frac{I_{\rm V}(E)}{I_{\rm H}(E)}$ for $R_{\rm pVH}(2.88) = 11.2$, $R_{\rm pVH}(2.71) = 9.5$ and $R_{\rm pVH}(2.58) = 8.2$, results comparable to those obtained previously by using rubbed PI layers[@Whitehead2000]. For non-aligned F6T2 (Figure \[fig:FigurePLFilms\](e)), the PL vibronic peaks appear at $2.28\;$eV ($544\;$nm), $2.12\;$eV ($585\;$nm) and $1.95\;$eV ($636\;$nm) and $R_{\rm VH}$ increases from 1.16 to 3.50 (Figure \[fig:FigurePLFilms\](g)), a ratio harder to calculate due to a weaker emission (in which the noise had to be removed using a fast Fourier transform algorithm) compared with F8BT and PFO.
![\[fig:FigurePLFilms\] PL spectra for non-aligned ((a),(c) $\&$ (e)) and aligned ((b),(d) $\&$ (f)) F8BT ((a) $\&$ (b)), PFO ((c) $\&$ (d)) and F6T2 ((e) $\&$ (f)) thin films. For a given polymer in a given alignment state, both polarizations spectra were normalized by the same value to allow for their comparisons. Solid blue lines show the PL collected using a polarizer in the vertical direction and red lines in the horizontal direction. In panels (b),(d) and (f): ${\mathbin{\!/\mkern-5mu/\!}}$ and $\perp$ indicate that the collection polarizer’s direction respectively matches or is perpendicular to the direction of the polymer chains in the film. A fast Fourier transform algorithm was used for F6T2 ((e) $\&$ (f)) to reduce noise in the recorded signal.](FigurePL.png)
Al-Polymer-Al microcavities were fabricated using non-aligned F8BT, PFO, F6T2 and for each cavity, a corresponding Al-Aligned(SD1-Polymer)-Al microcavity was fabricated. Angle-resolved polarized reflectivity maps were recorded for each microcavity by varying the angle $\theta$ formed between the vector normal to the microcavity plane and the incident light direction. For the aligned cavities, the measurement was performed at an angle $\Phi$ (formed between TE polarization and the polymer chain direction) equal to $0^{\circ}$. All results were analysed using a Hopfield-Agranovich Hamiltonian[@Agranovich1957; @Hopfield1958; @Ciuti2005] including either one (PFO, F6T2) or two (F8BT) separate excitons with all fitting results displayed in Table \[tab:1\]. For each microcavity, the experimental results are supported by transfer matrix reflectivity (TMR) calculations whose outputs are shown in Supporting Information. The reference results obtained for the non-aligned PFO and F8BT cavities are also shown in Supporting Information and agree with previous reports[@Tropf2017; @LeRoux2018].
Measured and fitted results for the non-aligned F6T2 cavity for TE polarization are shown in Figure \[fig:FigureTECoupling\] (a) where we observe the Lower (LP) and Upper (UP) Polaritons avoiding crossing of the exciton ($E_{\rm X_{\rm F6T2}} = 2.72\;$eV) by more than $1\;$eV, providing clear evidence of USC. Even though the layer is non-aligned, the largest Rabi splitting energy to date $\hbar\Omega_{0_{\rm TE}} = 1.38 \pm 0.01\;$eV is measured, corresponding to a coupling ratio $g = 49\%$. As expected, the results for TM polarization (Figure \[fig:FigureTMCoupling\] (a)) displays a flatter angular dispersion for the polaritons as the effective refractive index $n_{\rm eff_{\rm TM}}$ of the microcavity increases due to polarization-dependent penetration depth through the metallic mirrors[@Kena-Cohen2013; @Economou1969; @Litinskaya2012]. For this polarization, the in-plane/out-of-plane anisotropy of the active layer slightly reduces the Rabi energy down to $\hbar\Omega_{0_{\rm TM}} = 1.27 \pm 0.01\;$eV, as the contribution of the weaker out-of plane ($k_{\rm ex}$) component to the overall interaction increases along with $\theta$[@Kena-Cohen2013].
Results for the aligned cavities of PFO, F6T2 and F8BT are shown respectively in Figure \[fig:FigureTECoupling\] (b), (c) and (d) for TE polarization at $\Phi = 0^{\circ}$. In each case, USC is clearly observed with fitting results revealing a systematic increase for $\hbar\Omega_{0_{\rm TE,\Phi = 0^{\circ}}}$ of $\sim 40\%$ compared to the non-aligned cavities. As the transition dipole moment $\boldsymbol{\mu}$ rotates together with the polymer chains in the $y$-direction, the relative increase in coupling strength can be calculated using the term $\boldsymbol{{\mu}.E}$ in Equation \[eq:Scaling\] to be: $\frac{\cos(0)}{\cos(\frac{\pi}{4})} = \sqrt{2}\sim 41 \%$ for TE polarization. Different factors can explain deviations to this theoretical value: (i) albeit thin, the SD1 layer can slightly decrease the overall coupling strength by diminishing the overlap between exciton and cavity modes, (ii) all the polymer chains are not perfectly aligned in the $y$ direction with remainding optical activity in the $x$ direction (see Figure \[fig:FigureIndices\]) reducing the coupling strength, (iii) the reordering of the polymer can affect its microscopic properties. All the $\hbar\Omega_{0_{\rm TE,\Phi = 0^{\circ}}}$ values derived here, if not for the one we have just reported for non-aligned F6T2, exceed previous reports: the aligned F8BT cavity exhibits $\hbar\Omega_{01_{\rm TE,\Phi = 0^{\circ}}} = 1.18 \pm 0.01\;$eV and $\hbar\Omega_{02_{\rm TE,\Phi = 0^{\circ}}} = 1.25 \pm 0.01\;$eV, the aligned PFO cavity $\hbar\Omega_{0_{\rm TE,\Phi = 0^{\circ}}} = 1.47 \pm 0.01\;$eV and the aligned F6T2 cavity reaches a value of $\hbar\Omega_{0_{{\rm TE},\Phi = 0^{\circ}}} = 1.80 \pm 0.01\;$eV. To our knowledge, this last value is the first Rabi splitting energy to reach values comparable to photons in the visible spectrum ($\sim 689\;$nm) in a solid-state system, exceeding by more than 60$\%$ the previous record[@Liu2015; @Liu2019] and corresponds to a slightly higher coupling ratio $g \sim 65 \%$ than the one obtained in Ref 25, fact which is remarkable since this result is obtained for an exciton lying $\sim 1.6\;$ eV higher in energy.
![\[fig:FigureTECoupling\] Experimental, angle-resolved, TE-polarized reflectivity maps for microcavities containing: (a) non-aligned F6T2, (b) aligned PFO, (c) aligned F6T2 and (d) aligned F8BT. For (b), (c) $\&$ (d), the measurements were performed at $\Phi = 0^{\circ}$ (see definition in the text) as depicted in the experiment schematic shown above the maps. Overlaid solid white lines are the exciton $E_{\rm X}$ and cavity $E_{\rm C}$ modes, black dashed lines are polaritons fitted from the analytical model.](FigureTECouplingarxiV.png)
The results for TM polarization for the aligned cavities are shown in Figure \[fig:FigureTMCoupling\] (b), (c) and (d). In those measurements, the weak coupling regime is clearly observed for F6T2 (c) and F8BT (d) as the LP and UP are no longer visible and replaced by a single photonic mode. For PFO (b), the reflectivity observed does not show proper anti-crossing around the exciton ($E_{\rm X_{\rm PFO}} = 3.23\;$eV) with the PL measurements in the next section supporting the lack of evidence for SC.
![\[fig:FigureTMCoupling\] Experimental, angle-resolved, TM-polarized reflectivity maps for microcavities containing (a) non-aligned F6T2 (b) aligned PFO (c) aligned F6T2 and (d) aligned F8BT. For (b), (c) $\&$ (d), the measurements were performed at $\Phi = 0^{\circ}$ (see definition in the text) and the overlaid solid white line is the exciton $E_{\rm X}$. In (a) in addition to the exciton $E_{\rm X}$, the second white line represents $E_{\rm C}$ the cavity mode and the black dashed lines the polaritons fitted from the analytical model.](FigureTMCouplingarxiV.png)
--------------------------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
Polymer: PFO NA PFO A F8BT NA F8BT A F6T2 NA F6T2 A
$\hbar\omega_{1} ({\rm eV})$$\;$ 3.23 3.23 3.82 3.82 2.79 2.79
$\hbar\omega_{2}({\rm eV})$ $\;$ - - 2.70 2.70 - -
$\hbar\Omega_{01_{\rm TE}} ({\rm eV})$ $\;$ 1.02 $\pm$ 0.01 1.47 $\pm$ 0.01 0.84 $\pm$ 0.01 1.18 $\pm$ 0.01 1.38 $\pm$ 0.01 1.80 $\pm$ 0.01
$g_{01_{\rm TE}} \%$ $\;$ 32 46 22 31 49 65
$\hbar\Omega_{01_{\rm TM}} ({\rm eV})$ $\;$ 0.97 $\pm$ 0.01 - 0.80 $\pm$ 0.01 - 1.27 $\pm$ 0.01 -
$g_{01_{\rm TM}} \%$ $\;$ 30 - 21 - 46 -
$\hbar\Omega_{02_{\rm TE}} ({\rm eV})$$\;$ - - 0.84 $\pm$ 0.01 1.25 $\pm$ 0.01 - -
$g_{02_{\rm TE}} \%$ $\;$ - - 31 46 - -
$\hbar\Omega_{02_{\rm TM}} ({\rm meV})$$\;$ - - 0.80 $\pm$ 0.01 - - -
$g_{02_{\rm TM}} \%$ $\;$ - - 30 - - -
$ n_{\rm eff_{\rm TE}}$ $\;$ 1.68 $\pm$ 0.01 1.70 $\pm$ 0.01 1.70 $\pm$ 0.01 1.87 $\pm$ 0.01 1.64 $\pm$ 0.01 1.71 $\pm$ 0.01
$ n_{\rm eff_{\rm TM}}$$\;$ 2.33 $\pm$ 0.01 - 2.37 $\pm$ 0.01 - 2.12 $\pm$ 0.01 -
$ E_{\rm c_{\rm TE}} (0) ({\rm eV})$ $\;$ 3.13 $\pm$ 0.02 2.92 $\pm$ 0.02 2.54 $\pm$ 0.02 2.71 $\pm$ 0.02 2.48 $\pm$ 0.02 2.78 $\pm$ 0.02
$ E_{\rm c_{\rm TM}} (0)({\rm eV})$ $\;$ 3.15 $\pm$ 0.02 - 2.59 $\pm$ 0.02 - 2.50 $\pm$ 0.02 -
--------------------------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
: \[tab:1\]Extracted and pre-set parameter values for microcavities containing non-aligned and aligned polymers, modelled using a Hopfield-Agranovich Hamiltonian[@Hopfield1958; @Agranovich1957; @Ciuti2005]. Values are shown for both TE- and TM-polarization, at $\Phi = 0^{\circ}$ (see definition in the text) for the aligned cavities. The initials A and NA respectively designate aligned and non-aligned polymer layers.
Exciton oscillator 1 transition energy.
Exciton oscillator 2 transition energy.
TE-polarized Rabi energy associated with exciton 1 for $\omega_{\rm cav_{\rm TE}}=\omega_{1}$ (see definition in the text).
TE-polarized normalized coupling ratio energy associated with exciton 1 (see definition in the text).
TM-polarized Rabi energy associated with exciton 1 for $\omega_{\rm cav_{\rm TM}}=\omega_{1}$ (see definition in the text).
TM-polarized normalized coupling ratio energy associated with exciton 1 (see definition in the text).
TE-polarized Rabi energy associated with exciton 2 for $\omega_{\rm cav_{\rm TE}}=\omega_{1}$ (see definition in the text).
TE-polarized normalized coupling ratio energy associated with exciton 2 (see definition in the text).
TM-polarized Rabi energy associated with exciton 2 for $\omega_{\rm cav_{\rm TM}}=\omega_{1}$ (see definition in the text).
TM-polarized normalized coupling ratio energy associated with exciton 2 (see definition in the text).
Effective refractive index for TE polarization.
Effective refractive index for TM polarization.
TE-polarized energy of the bare cavity mode at normal incidence.
TM-polarized energy of the bare cavity mode at normal incidence.
Angle-resolved PL was recorded for each microcavity. The excitation laser used for the thin film PL measurement was focused onto the sample at an incidence of 75$^{\circ}$, the excitation geometry between films and microcavities experiments being identical. For each microcavity, the average power was kept low ($\leq 10 \mu$W) and the excitation energy was chosen to optically pump one of the intense absorptions arising from the optical transitions of the underlying polymer. For the aligned cavities, the measurement was performed at $\Phi = 0^{\circ}$.
PL intensity maps plotted by energy vs emission angle for both TE- ((a), (c) and (e))) and TM-polarized ((b), (d) and (f)) emissions from the non-aligned F8BT ((a) and (b)), PFO ((c) and (d)) and F6T2 ((e) and (f)) microcavities are shown in Figure \[fig:FigureAPLNA\]. In each case the emission is dominated by a narrow single peak originating from the LP which is relatively insensitive to angular-dispersion: a recognizable feature of USC[@Kena-Cohen2013; @Gambino2014; @Mazzeo2014] (values including peak positions and full width at half maximum (FWHM) at normal incidence as well as the angular dispersion from 0 to $60^{\circ}$ are shown in Table \[tabPL\]). The results for PFO agree with a previous report[@LeRoux2018] and the variety of polymers used here allows for emissions across the visible spectrum (in the blue at $\sim 452\;$ nm for PFO, green at $\sim 537\;$nm for F8BT and yellow/orange at $\sim 588\;$nm for F6T2).
Corresponding PL intensity maps for the aligned cavities are shown in Figure \[fig:FigureAPLA\]. TE-polarized emission ((a), (c) and (f)) resembles the one observed from non-aligned microcavities with a single peak emitted from the LP. It however differs in energy for F8BT and PFO between non-aligned and aligned cavities: the peak emission is recorded at $2.13\;$eV at normal incidence for aligned F8BT compared to $2.31\;$eV when non-aligned, and at $2.59\;$eV for aligned PFO compared to $2.74\;$eV when non-aligned. These redshifts are not the result of different thicknesses between the cavities as TMR calculations show that for each pair, the polymer layer thicknesses are comparable (the aligned F8BT is $110\;$nm-thick compared to $118\;$nm when aligned, the aligned PFO 96 nm-thick compared to 97 nm when aligned) but are direct evidence of the increased interaction strength which repels the UP and LP to respectively higher and lower energies. The emission for aligned and non-aligned F6T2 is closer in energy at normal incidence ($2.11\;$eV aligned compared with $2.10\;$eV non-aligned) and is this time the result of a much larger thickness of the non-aligned F6T2 layer (the aligned F6T2 is $94\;$nm-thick compared to $123\;$nm non-aligned), resulting in lower energy cavity modes and LP (as can be observed in Figure \[fig:FigureTECoupling\]) which makes up for the difference in interaction strength between the two microcavities.
TM-polarized measurements ((b), (d) and (f)) show no distinguishable emission for F8BT (b) and F6T2 (g) as the only photonic mode in both those cavities lies too high in energy (at $\sim 3\;$eV for F6T2 and $\sim 2.9\;$eV for F8BT) to allow any emission from the underlying polymer (Figure \[fig:FigurePLFilms\] (b) and (f) shows the emission located at energies lower than $2.6\;$eV for both F6T2 and F8BT). For the microcavity containing aligned PFO (d), a 5-times weaker, broad, angle-independent TM-polarized emission was detected (f), with two maxima at 2.88 and $2.71\;$eV coinciding with the ones from the bare film (Figure \[fig:FigurePLFilms\](d)). This broad, structured, angle-insensitive dispersion confirms that the microcavity no longer operates under USC for this polarization. The emission itself is only allowed through the photonic mode at $\sim 2.88\;$eV (see \[fig:FigurePLFilms\](b)) and even though this photonic mode overlaps with the most intense part of the bare film’s PL, the resulting signal is much weaker in TM than TE polarization as the emission in the direction perpendicular to the chain alignment is intrinsically much weaker.
![\[fig:FigureAPLNA\] Angle-resolved PL spectral intensity maps for microcavities containing non-aligned F8BT ((a) $\&$ (b)), PFO ((c) $\&$ (d)) and F6T2 ((e) $\&$ (f)), with TE ((a), (c), (e)) and TM ((b), (d), (f)) polarized spectra plotted separately. The excitation energy is overlaid in white for each measurement. Each polarization pair ((a) $\&$ (b)), ((c) $\&$ (d)), ((e) $\&$ (f)) is normalized by the same value to allow for comparison.](FigureAPLNA.png)
![\[fig:FigureAPLA\] Angle-resolved PL spectral intensity maps for microcavities containing aligned F8BT ((a) $\&$ (b)), PFO ((c) $\&$ (d)) and F6T2 ((e) $\&$ (f)), with TE ((a), (c), (e)) and TM ((b), (d), (f)) polarized spectra plotted separately. The measurements were performed at $\Phi = 0^{\circ}$ (see definition in the main text) . Each polarization pair ((a) $\&$ (b)), ((c) $\&$ (d)), ((e) $\&$ (f)) is normalized by the same value to allow for comparison.](FigureAPLA.png)
----------------------------- ------------ ------------- ------------ ------------- ------------- -------------
Polymer: PFO NA PFO A F8BT NA F8BT A F6T2 NA F6T2 A
Peak position (eV) 2.74 2.59 2.31 2.13 2.10 2.11
${\rm FWHM}_{\rm TE}$ (meV) 99 $\pm$ 1 140 $\pm$ 5 89 $\pm$ 5 127 $\pm$ 1 109 $\pm$ 1 109 $\pm$ 1
${\rm FWHM}_{\rm TM}$ (meV) 99 $\pm$ 1 - 82 $\pm$ 5 - 109 $\pm$ 1 -
TE dispersion (meV) 130 50 30 80 110 60
TM dispersion (meV) 110 - 20 - 110 -
----------------------------- ------------ ------------- ------------ ------------- ------------- -------------
: \[tabPL\] Peak positions and FWHMs at normal incidence for the TE- and TM-polarized emission s displayed in Figure \[fig:FigureAPLNA\] and \[fig:FigureAPLA\]. The angular dispersion of the emission for both polarizations from 0 to $60^{\circ}$ is also reported. The initials A and NA respectively designate aligned and non-aligned polymer layers.
Several opportunities unique to our photoalignment technique for exciton-polaritons exist: the possibility to align SD1 at mesoscopic scales[@Ma2015; @Ma2017; @Zhang2019] allows local alignment of the polymer chains, which could lead to the fabrication of the missing NOT gate in optical logic thanks to spontaneous splittings between TE and TM polarizations[@Sanvitto2016; @Espinosa-Ortega2013; @Solnyshkov2015], rapid switching of the coupling strength using LCPPs and LCs in microcavities could also be an opportunity to perform direct extraction of the USC ground state virtual contents[@Ciuti2005; @FriskKockum2019; @Forn-Diaz2019] and the fabrication of complex energy landscapes, especially when combined with the tunability of the molecular structure (for example by generating segments of $\beta$-phase in PFO[@Perevedentsev2015]) could help address challenges in quantum simulation[@Scafirimuto2018]. On a more practical standpoint, we expect that a wide breadth of polarization sensitive devices and phenomena such as Bose-Einstein condensation and exciton-polariton lasing (which has recently been demonstrated using pentafluorene [@Rajendran2019]) will take advantage of the polarization-dependent coupling (the spontaneous polarization observed during lasing and condensation could for instance be controlled by molecular alignment).
Conclusions
===========
We have fabricated organic microcavities containing LCCPS (F8BT, PFO, F6T2) aligned using a thin photoalignment layer (SD1). The USC regime was first observed for the non-aligned microcavities with a value of $\hbar\Omega_{\rm R_{F6T2}} = 1.34$ eV exceeding previous reports. The alignment then allowed for a systematic increase of the coupling strength in the direction of the alignment with giant values of $\hbar\Omega_{\rm R}$ culminating at 1.80 eV for F6T2, a value comparable to photon energies in the visible spectrum, also corresponding to the highest coupling ratio $g$ = 65$\%$ to date in the solid-state[@FriskKockum2019; @Forn-Diaz2019]. Angle-resolved PL for the TE polarization parallel to the alignment direction revealed red-shifted LP emissions compared to the non-aligned cavities, a signature of the increased interaction strength. In this geometry, the absence of polaritons in TM-polarized reflectivity and weak or no PL also demonstrated that the coupling strength was polarization dependent. By using three different polymers, we demonstrated that the alignment can be generalized to other LCPPs and that a real opportunity to reach coupling ratios close to $90\%$ exists if the alignment can be applied at lower energies. Aligning LCCPS in microcavities at microscopic scales using SD1 also offers further possibilities for the realization of polaritonic devices and rich energy landscapes.
Methods
=======
Materials
---------
The three polymers used in this study were supplied by Cambridge Display Technology (F8BT), Sumitomo Chemical (PFO and F6T2) and used as received. Their peak molecular weights were: $M_{p_{\rm PFO}}$ = $50\times 10^{3}$ g.mol$^{-1}$, $M_{p_{\rm F8BT}}$ = $77\times 10^{3}$ g.mol$^{-1}$, $M_{p_{\rm F6T2}}$ = $80\times 10^{3}$ g.mol$^{-1}$. The azo-dye photo-alignement layer SD1 was supplied by Dai-Nippon Ink and Chemicals, Japan. Anhydrous toluene (99.8$\%$), anhydrous chloroform ($\geq$99$\%$) and anhydrous 2-methoxyethanol($\geq$99.8$\%$) were purchased from Sigma-Aldrich. Solvents were used as received. For the mirror fabrication, Aluminium pellets (99.999$\%$) were purchased from Kurt J. Lesker.
Film Fabrication
----------------
The bare films (used for ellipsometry and PL) of SD1, PFO, F8BT, F6T2 were spincoated from solutions in 2-methoxyethanol (SD1 at $1$ mg.mL$^{-1}$), toluene (PFO at $18\;$mg.mL$^{-1}$ and F8BT at $18\;$mg.mL$^{-1}$) and chloroform (F6T2 at $13\;$mg.mL$^{-1}$). All solutions were prepared in an inert environment, left to stir overnight at a temperature of $55^{\circ}$C except for F6T2 in chloroform which was left stirring at room temperature. All solutions were then filtered using a $0.45 \;\mu$m PTFE filter. All samples were spincoated on top of fused silica substrates. The non-aligned polymer films were spun for $1\;$min at a speed of $2000\;$rpm with an acceleration of $1200\;$rpm.s$^{-1}$. For the aligned polymer films, the SD1 layer was first spincoated by spinning $5\;$s at $500\;$rpm (acceleration $500\;$rpm.s$^{-1}$) and $25\;$s at $2000\;$rpm (acceleration $1200\;$rpm.s$^{-1}$). The film was then annealed for 6 minutes at a temperature of $150^{\circ}$C to drive any traces of solvent away. The alignment of the SD1 layer was performed in air by exposing the sample to $5\;$mW of polarized UV light (emitted by a M365LP1 LED in front of a broadband WP25M-UB polarizer from Thorlabs) for 10 minutes; The SD1 chains aligning perpendicular to the direction of the polarized light[@Li2006; @Ma2015]. The samples were then put back in an inert environment where the polymer layer was spincoated by spinning for $1\;$min at a speed of $2000\;$rpm with an acceleration of $1500\;$rpm.s$^{-1}$. Each polymer was then thermally annealed into their respective nematic phases (160$^{\circ}$C for PFO, 220$^{\circ}$C for F6T2, 250$^{\circ}$C for F8BT) using a Linkam THMS600 heating stage with a heating rate of 30$^{\circ}$C.min$^{-1}$. This temperature was held for 10 minutes and subsequent quenching to room temperature was realized by quickly placing the sample on the metallic floor of the glovebox while applying a gentle flow of nitrogen.
Microcavity Fabrication
-----------------------
The aluminium mirrors were evaporated at a rate of $10\;$A.s$^{-1}$ at a pressure of $10^{-9}\;$mbar. For the non-aligned cavities, the spincoating conditions used on top of the bottom mirror were identical to the ones described for the bare films. For the aligned cavities, the SD1 layer was spincoated on top of the bottom mirror. The structures fabricated had low Q factors $\sim 25$, characteristic of metallic microcavities using Aluminium. The concentration of the SD1 solution in 2-methoxyethanol was increased to $3$ mg.mL$^{-1}$ as spincoating SD1 on a metallic surface results in lower thicknesses than on fused silica. A slightly thicker layer also acts as a protection layer to prevent the aluminium from reacting with the polymer upon annealing at high temperatures (using $1\;$mg.mL$^{-1}$ SD1 solution resulted in samples unfit for measurement). The sample was then annealed 6 minutes at a temperature of $150^{\circ}$C to drive any traces of solvent away and the rest of the alignment procedure was similar to the one used for the bare films. For the spincoating of the polymer layers, the solution concentrations were in some cases adjusted so as to adjust the thicknesses and therefore the cavity mode energy. The concentrations used in the non-aligned microcavities were: PFO at $19\;$mg.mL$^{-1}$ , F8BT at $20\;$mg.mL$^{-1}$ and F6T2 at $15\;$mg.mL$^{-1}$. The concentrations for the aligned microcavities were: PFO at $18\;$mg.mL$^{-1}$, F8BT at $18\;$mg.mL$^{-1}$, F6T2 at $13\;$mg.mL$^{-1}$.
Optical Characterization
------------------------
The optical constants for the non-aligned and aligned films of PFO, F6T2 and F8BT were extracted using a J.A. Woollam ESM-300 ellipsometer. For each sample, 8 reflection-geometry measurements were performed with light incident from 45$^{\circ}$ to 61$^{\circ}$ (angles of incidence are quoted relative to the plane normal) together with a normal incidence (0$^{\circ}$) transmission measurement, for the aligned films the measurement was performed at $\Phi = 0^{\circ}$. The reflectivity maps obtained in Figure \[fig:FigureIndices\] were obtained using a home-built white light reflectivity setup. The microcavities were placed at the center of a stage with two independent rotating arms. A deuterium-halogen light source (DH-2000-DUV from Ocean Optics) was coupled into a fiber whose output was collimated onto a broadband polarizer (WP25M-UB from Thorlabs) and onto the sample (final spot size $1\;$mm). The reflected light was then coupled into a second fiber placed onto the second arm and analyzed using a spectrometer (HRS 500, 150 g/mm grating with blazing wavelength at 300 nm) and CCD (Pylon-2KB CCD from Princeton Instruments). In all cases, a neat aluminium mirror with known reflectivity was used as reference. The acquisition angle was varied by $0.5^{\circ}$ steps from 10 to 70$^{\circ}$.
Photoluminescence
-----------------
All the time-integrated PL measurements were performed on the same home-built setup. The fiber coupled to the white light source was disconnected and the first arm was positioned at 75$^{\circ}$ relative to the sample plane normal. The pulsed laser beam from a supercontinuum white light laser (SuperK Extreme with its UV spectral extension unit Extend-UV, NKT Photonics) was used as excitation source and focused onto the sample (spot size $<1$ mm). The excitation energy was tuned according to the optical transitions, the incident power was kept low ($\leq 10\mu W$) with pulse widths of 20 to 30 ps and a repetition rate of $77.87\;$MHz. The broadband polarizer was placed on the collection arm, at a distance of $10\;$ cm from the sample. Two nearly closed irises ($\sim 1$cm) at a distance of $5\;$cm from each other were then placed before the coupling lens of the collection fiber in order to ensure that the collected light was emitted at the desired angle in the horizontal plane. The light was then analysed using the spectrometer and CCD described before using this time a 300 g/mm grating blazed at a wavelength of $500\;$nm. For the polymer films, the light was collected at normal incidence. For the microcavities the acquisition angle was varied by $1^{\circ}$ steps from -40 to 60$^{\circ}$.
Data Analysis
-------------
The minima of the reflectivity maps shown in Figures \[fig:FigureTECoupling\] and \[fig:FigureTMCoupling\] were analyzed using a least-square fitting algorithm for the eigenvalue problem:
$$\label{eq:8}
H_{q} \bold{v_{i,q}}= \omega_{i,q} \bold{v_{i,q}},$$
where $H_q$ is an extension of the Agranovich/Hopfield Hamiltonian containing either one[@Hopfield1958; @Agranovich1957; @Ciuti2005] (PFO, F6T2) or two excitonic resonances[@LeRoux2018] (F8BT):
$$\label{eq:9}
H_{q}=
\begin{bmatrix}
\omega_{{\rm cav},q}+2D_{q} & -i\frac{\Omega_{1,q}}{2} & -i\frac{\Omega_{2,q}}{2} & -2D_{q} & -i\frac{\Omega_{1,q}}{2} & -i\frac{\Omega_{2,q}}{2}\\
i\frac{\Omega_{1,q}}{2} & \omega_{1} & 0 & -i\frac{\Omega_{1,q}}{2} & 0 & 0\\
i\frac{\Omega_{2,q}}{2} & 0 & \omega_{2} & -i\frac{\Omega_{2,q}}{2} & 0 & 0\\
2D_{q} & -i\frac{\Omega_{1,q}}{2} & -i\frac{\Omega_{2,q}}{2} & -\omega_{{\rm cav},q} - 2D_{q} & -i\frac{\Omega_{1,q}}{2} & -i\frac{\Omega_{2,q}}{2}\\
-i\frac{\Omega_{1,q}}{2} & 0 & 0 & i\frac{\Omega_{1,q}}{2} & -\omega_{1} & 0\\
-i\frac{\Omega_{2,q}}{2} & 0 & 0 & i\frac{\Omega_{2,q}}{2} & 0 & -\omega_{2}
\end{bmatrix} ,$$
In the case of a single exciton oscillator, $H_q$ reduces to the usual 4 x 4 Hopfield-like USC matrix[@Kena-Cohen2013]. In Eq. (\[eq:8\]), (\[eq:9\]), q is the in-plane wave vector, $\omega_{{\rm cav}_q}$ the cavity mode energy, $\omega_j$ the frequency for the $j$-excitons, $\Omega_{j,q}$ is the associated Rabi frequency, and for a given angle $\theta$: $\Omega_{j,q}= \Omega_{j}(\theta) = \Omega_{0j}\sqrt{\frac{\omega_j}{\omega_{{\rm cav}}(\theta)}}$ where $\Omega_{0j}$ is the Rabi frequency on resonance for the $j$-excitons. It was shown that in metal-organic semiconductor-metal cavities $\omega_{{\rm cav}}(\theta)$ can be approximated by[@Kena-Cohen2013]:
$$\label{eq:5}
\omega_{{\rm cav}_{{\rm (TE,TM)},q}} = \omega_{{\rm cav}_{{\rm (TE,TM)}}}(\theta)=
\omega_{{\rm cav}}(0)\left(1-\frac{\sin^{2}(\theta)}{n^{2}_{{\rm eff}_{{\rm (TE,TM)}}}}\right)^{-\frac{1}{2}}$$
where $n_{\rm eff_{\rm TE,TM}}$ is polarization dependent. Finally, $D_q = \sum _{j} \frac{\Omega_{j,q}^2}{4\omega_{j}}$ is the contribution of the squared magnetic vector potential.
In order to diagonalize *H*, the polariton annihilation operators $p_{i,q} = w_{i,q}a_{q}+\sum_{j}x_{i,j,q}b_{j,q}+y_{i,q}a^{\dagger}_{-q}+\sum_{j}z_{i,j,q}b^{\dagger}_{j,-q}$ for $i\epsilon\left \{\rm LP,MP,UP\right\}$ are introduced, where $a_q$ and $a^{\dagger}_{q}$ respectively annihilate and create a photon at frequency $\omega_{{\rm cav}_q}$, $b_j$ and $b^{\dagger}_j$ respectively annihilate and create a $j$-exciton at frequency $\omega_j$. The terms $w, x, y$ and $z$ label, respectively, the photon, exciton, anomalous photon and anomalous exciton Hopfield coefficients. The eigenvalues of $H_q$ were fitted to the experimental results for each cavity, for both TE- and TM-polarization, using the $R$-minima in the 10 - 70$^{\circ}$ range.
In order to minimize the number of fitting parameters and obtain meaningful results, only $\omega_{{\rm cav}_{\rm TE,TM}}(0)$, $n_{\rm eff_{\rm TE,TM}}$ $\&$ $\Omega_{01_{\rm TE,TM}}$ were allowed to vary in fittings of the PFO and F6T2 cavities. Similarly, only $\omega_{{\rm cav}_{\rm TE,TM}}(0)$, $n_{\rm eff_{\rm TE,TM}}$, $\Omega_{01_{\rm TE,TM}}$ $\&$ $\Omega_{02_{\rm TE,TM}}$ were allowed to vary in the fitting of the F8BT cavities. For each exciton, the value of $\hbar\omega_{j}$ was set to be at the energy that corresponds to the mid-point of the integral oscillator strength for the corresponding optical transition using $\int_{E_{\rm min}}^{\hbar\omega_{j}}\epsilon(\omega)d\omega=\frac{1}{2}\int_{E_{\rm min}}^{E_{\rm max}}\epsilon(\omega)d\omega$, where $\epsilon(\omega)$ is the extinction coefficient for $X_{j}$ in the $E_{\rm min}$ to $E_{\rm max}$ energy range.
The authors thank Prof. Moritz Riede for access to his facilities and Dr Richard Hamilton for fruitful discussions. They also acknowledge funding from the University of Oxford, from the UK Engineering and Physical Sciences Research Council and the Jiangsu Industrial Technology Research Institute. F.L.R. further thanks Wolfson College and Dr Simon Harrison for the award of a Wolfson Harrison UK Research Council Physics Scholarship.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Our aim is to investigate the properties of existence and uniqueness of greedy bases in Banach spaces. We show the non-existence of greedy basis in some Nakano spaces and Orlicz sequence spaces and produce the first-known examples of non-trivial spaces (i.e., different from $c_{0}$, $\ell_{1}$, and $\ell_{2}$) with a unique greedy basis.'
address:
- |
Mathematics Department\
Universidad Pública de Navarra\
Pamplona 31006\
Spain
- |
Department of Mathematics and Computer Sciences\
Universidad de La Rioja\
Logroño 26004\
Spain
- |
Department of Mathematics\
University of South Carolina\
Columbia SC 29208\
USA
- |
Institute of Mathematics Bulgarian Academy of Sciences\
Sofia\
Bulgaria\
*currentadress*\
Department of Mathematics University of Illinois at Urbana-Champaign\
Urbana, IL 61801\
USA
author:
- 'F. Albiac'
- 'J. L. Ansorena'
- 'S. J. Dilworth'
- Denka Kutzarova
title: Existence and uniqueness of greedy bases in Banach spaces
---
[^1]
Introduction and background {#introd}
===========================
Let ${\ensuremath{\mathbb{X}}}$ be a separable (real) Banach space. One of the most important problems in the isomorphic theory dating back to Banach’s school is the study of the existence and uniqueness of Schauder bases for ${\ensuremath{\mathbb{X}}}$. The question of uniqueness is formulated in a meaningful way through the notion of equivalence of bases. Recall that two normalized (or semi-normalized) bases $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ and $({\ensuremath{\mathbf{x}}}_n)_{n=1}^{\infty}$ of ${\ensuremath{\mathbb{X}}}$ are called *equivalent* provided a series $\sum_{n=1}^{\infty}a_{n}{\ensuremath{\mathbf{e}}}_{n}$ converges if and only if $\sum_{n=1}^{\infty}a_{n}{\ensuremath{\mathbf{x}}}_{n}$ converges. This is the case if and only if the map $$\sum_{n=1}^{\infty}a_{n}{\ensuremath{\mathbf{e}}}_{n}\to \sum_{n=1}^{\infty}a_{n}{\ensuremath{\mathbf{x}}}_{n}$$ defines an automorphism of ${\ensuremath{\mathbb{X}}}$. That is, there exists a constant $C\ge 1$ such that $$\frac1C\left\Vert \sum_{n=1}^{N}a_{n}{\ensuremath{\mathbf{e}}}_{n}\right\Vert\le \left\Vert \sum_{n=1}^{N}a_{n}{\ensuremath{\mathbf{x}}}_{n}\right\Vert \le C\left\Vert \sum_{n=1}^{N}a_{n}{\ensuremath{\mathbf{e}}}_{n}\right\Vert,$$ for $a_{1},\dots,a_{N}\in \mathbb R$ and $N=1,2,\dots$. As it happens, in every infinite-dimensional Banach space with a basis there are uncountably many non-equivalent normalized bases [@PelcynskiSinger1964]. Thus in order to get a more accurate structural information on a given space using bases as a tool, one needs to restrict the discussion on their existence and uniqueness to bases with certain special properties.
The most useful and extensively studied class of special bases is that of *unconditional bases.* A basis $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ of $X$ is [*unconditional*]{} if $({\ensuremath{\mathbf{e}}}_\pi(n))_{n=1}^{\infty}$ is a basis of $X$ for any permutation $\pi$ of the indices. If a Banach space has a unique normalized unconditional basis it has to be equivalent to all its permutations, i.e., it has to be [*symmetric*]{}.
For a wide class of sequence spaces the canonical unit vector basis is the unique symmetric basis. This class contains all the Orlicz sequence spaces $\ell_{F}$ for which $\lim_{t\to 0} tF^{\prime}(t)/F(t)$ exists ([@LindTzaf1971a]), and also the Lorentz sequence spaces $d(w,p)$ where $p\ge 1$ and $w= (w_{n})_{n=1}^{\infty}\in c_{0}\setminus \ell_{1}$ is a nonincreasing sequence of positive numbers with $w_{1} =1$ (see [@ACL1973]). In particular the $\ell_{p}$ spaces for $1\le p<\infty$ have a unique symmetric basis. However, a complete classification of Banach spaces with a unique symmetric basis seems far from being achieved.
For a Banach space with a symmetric basis it is rather unusual to have a unique unconditional basis. It is well-known that $\ell_{2}$ has a unique unconditional basis [@KotheToeplitz1934], and a classic result of Lindenstrauss and Pe[ł]{}czy[ń]{}ski [@LindenstraussPel1968] asserts that $\ell_{1}$ and $c_{0}$ also have a unique unconditional basis. Lindenstrauss and Zippin [@LindenstraussZippin1969] completed the picture by showing that those three spaces are the all and only Banach spaces in which all unconditional bases are equivalent.
Once we have determined that a Banach space does not have a symmetric basis (a task that can be far from trivial) we must rethink the problem of uniqueness of unconditional basis. In fact, an unconditional non-symmetric basis admits a continuum of nonequivalent permutations (cf. [@Hennefeld1973]\*[Theorem 2.1]{}). Hence for Banach spaces without symmetric bases it is more natural to consider instead the question of uniqueness of unconditional bases up to (equivalence and) permutation, (UTAP) for short. Two unconditional bases $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ and $({\ensuremath{\mathbf{x}}}_n)_{n=1}^{\infty}$ of a Banach space $X$ are said to be [*permutatively equivalent*]{} if there is a permutation $\pi$ of $\mathbb N$ so that $({\ensuremath{\mathbf{e}}}_\pi(n))_{n=1}^{\infty}$ and $({\ensuremath{\mathbf{x}}}_n)_{n=1}^{\infty}$ are equivalent. Then we say that ${\ensuremath{\mathbb{X}}}$ has a (UTAP) unconditional basis $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ if every unconditional basis in ${\ensuremath{\mathbb{X}}}$ is permutatively equivalent to $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$.
The first movers in this direction were Edelstein and P. Wojtaszczyk, who proved that finite direct sums of $c_0$, $\ell_1$ and $\ell_2$ have a (UTAP) unconditional basis [@EdelWoj1973]. Bourgain et al. embarked on a comprehensive study aimed at classifying those Banach spaces with unique unconditional basis up to permutation, that culminated in 1985 with their [*Memoir*]{} [@BCLT1985]. They showed that the spaces $c_{0}(\ell_{1})$, $c_{0}(\ell_{2})$, $\ell_{1}(c_{0})$, $\ell_{1}(\ell_{2})$ and their complemented subspaces with unconditional basis all have a (UTAP) unconditional basis, while $\ell_{2}(\ell_{1})$ and $\ell_{2}(c_{0})$ do not. However, the hopes of attaining a satisfactory classification were shattered when they found a nonclassical Banach space, namely the $2$-convexification $\mathcal T^{(2)}$ of Tsirelson’s space having a (UTAP) unconditional basis. Their work also left many open questions, most of which remain unsolved as of today. Using completely different techniques, Casazza and Kalton solved some of these problems more recently in [@CasazzaKalton1998] by showing that the original Tsirelson space $\mathcal T$, and certain complemented subspaces of Orlicz sequence spaces have a (UTAP) unconditional basis. They also found a space with a (UTAP) unconditional basis with a complemented subspace failing to have a (UTAP) unconditional basis.
At the end of the 20th century, Banach space theorists began to feel attracted to study a brand new type of bases called greedy bases originating from nonlinear approximation and compressed sensing. Let us recall the definition. For $N=1,2,\dots,$ we consider non-linear operators $G_{N}$ defined by $$x=\sum_{n=1}^{\infty}{\ensuremath{\mathbf{e}}}_n^{\ast}(x){\ensuremath{\mathbf{e}}}_n\in X\mapsto G_N(x)=\sum_{n\in \Lambda_{N}(x)}{\ensuremath{\mathbf{e}}}_n^{\ast}(x){\ensuremath{\mathbf{e}}}_n,$$ where $\Lambda_{N}(x)$ is [*any*]{} $N$-element set of indices such that $$\min\{|{\ensuremath{\mathbf{e}}}_n^{\ast}(x)|: n \in \Lambda_{N}(x)\}\ge \max\{|{\ensuremath{\mathbf{e}}}_n^{\ast}(x)|: n\not\in \Lambda_{N}(x)\}.$$ The basis $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ is said to be [*greedy*]{} [@KoTe] if $G_N(x)$ is essentially the best $N$-term approximation to $x$ using the basis vectors, i.e., there is a constant $C$ such that for all $x\in X$ and $N\in \mathbb N$, we have $$\Vert x-G_{N}(x)\Vert\le C\inf\left\{\left\Vert x-\sum_{n\in A}\alpha_n{\ensuremath{\mathbf{e}}}_n\right\Vert : |A|=N, \alpha_n\in {\ensuremath{\mathbb{R}}}\right\}.$$ Konyagin and Temlyakov showed in [@KoTe] that greedy bases can be simply characterized as unconditional bases with the additional property of being [*democratic*]{}, i.e., for some constant $\Delta>0$ we have $$\left\Vert \sum_{n\in A}{\ensuremath{\mathbf{e}}}_n\right\Vert\le \Delta \left\Vert \sum_{n\in B}{\ensuremath{\mathbf{e}}}_n\right\Vert,$$ whenever $A$ and $B$ are finite subsets of integers of the same cardinality. Symmetric bases are trivially greedy but there exist greedy bases, such as the normalized Haar system in $L_{p}[0,1]$ for $1<p<\infty$, which are greedy but not symmetric. Thus being greedy is an intermediate property between being unconditional and being symmetric.
In this article we are concerned with investigating the novel property of uniqueness of greedy bases in Banach spaces in the expected sense. Since it is the first time this property is explicitly formulated in the literature we include its definition.
Suppose $X$ is a Banach space with a semi-normalized greedy bases $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$. The space $X$ is said to have a unique greedy basis (respectively, up to permutation) if whenever $({\ensuremath{\mathbf{x}}}_n)_{n=1}^{\infty}$ is another semi-normalized greedy basis of $X$, then $({\ensuremath{\mathbf{x}}}_n)_{n=1}^{\infty}$ is equivalent to $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$ (respectively, to a permutation of $({\ensuremath{\mathbf{e}}}_n)_{n=1}^{\infty}$.
Let us examine some possible scenarios. Firstly, Banach spaces that possess symmetric basis and have a unique unconditional basis (i.e., $\ell_{1}$, $\ell_{2}$, and $c_{0}$) also have a unique greedy basis, while if a Banach space with symmetric basis does not have a unique symmetric basis it cannot have a unique greedy basis. In between these two extremes we come across spaces like $\ell_p$ for $1<p<\infty$, $p\not=2$, with a unique symmetric basis and a wealth of permutatively non-equivalent greedy bases (cf. [@DHK2006]). The following question naturally arises:
\[problemon\] Are there any other Banach spaces aside from $c_{0}, \ell_{1},$ and $\ell_{2}$ with the property of uniqueness of greedy basis?
Secondly, in a Banach space without a symmetric basis it may happen that the natural basis of $X$ fails to be greedy. Such is the case in Bessov-type spaces like $(\oplus_{n=1}^\infty \ell_q^n)_p$ and $\ell_{p}(\ell_{q})$ for $1\le p\not=q<\infty$. This makes relevant to investigate the question of existence of greedy bases. Dilworth et al. [@DFOS2011] settled the problem of existence of greedy bases in $(\oplus_{n=1}^\infty \ell_q^n)_p$ by proving that these spaces have a greedy basis when $1<p<\infty$ and do not otherwise. In turn, Shechtmann [@Schectmann2014] showed that $\ell_{p}(\ell_q)$ fails to have a greedy basis in the nontrivial cases. As far as uniqueness is concerned, the right question to ask in Banach spaces without a symmetric basis is the uniqueness of greedy basis up to a permutation. For instance, Smela proved that the $L_{p}$-spaces for $1<p<\infty$, $p\not=2$, and $H_{1}$ have infinitely many permutatively non-equivalent greedy bases [@Smela2007] (cf. [@DHK2006]).
In the $L_{p}$-spaces ($1<p<\infty, p\not=2$) we find other interesting features that are worth it singling out. They have greedy basis (the normalized Haar system, [@Temlyakov1998]) and contain complemented subspaces such as $\ell_{p}(\ell_{2})$ and $\ell_{p}\oplus\ell_{2}$ with no greedy basis. One could argue that a reason for this pathology is that $L_{p}$ has no symmetric basis. But Pe[ł]{}czy[ń]{}ski’s universal space denoted by $U$ has a symmetric basis and clearly contains complemented subspaces with no greedy basis (see also Theorems \[ComplementedNotGreedy\] and \[ComplementedNotGreedySR\] below for more natural examples of Orlicz sequence spaces exhibiting this phenomenon).
Notice that in order to establish the existence and/or the uniqueness up to permutation of greedy basis for Banach spaces without a symmetric basis that have (UTAP) unconditional basis, the point is to determine whether the unique unconditional basis is greedy. This is the case with the aforementioned Tsirelson-type space $\mathcal T^{(2)}$, not to be confused with its close relative, the original Tsirelson space $\mathcal T$. The former provides an example of a Banach space without a symmetric basis but with a (UTAP) greedy basis.The latter, despite the fact that it has a (UTAP) unconditional basis, fails to contain any democratic basic sequences (cf. [@DKK2003]\*[Remark 5.8]{}). This observation answers in the negative what we could call [*the greedy basic sequence problem*]{} (evocative of the unconditional basic sequence problem).
Let us next briefly describe the contents of this article. For expositional ease, in the preparatory Sect. \[Prelim\] we have gathered some prerequisites on Musielak-Orlicz spaces and have included the not so well-known concepts of right/left dominant spaces and of sufficiently lattice Euclidean spaces. In Sect. \[Nakano\] we study the existence of greedy basis in complemented subspaces of Nakano spaces. In particular we see that certain Nakano spaces fail to have a greedy basis.
In Sect. \[Orlicz\] we turn to Orlicz sequence spaces with an eye to investigating the existence and uniqueness of greedy bases and we settle Problem \[problemon\].
One could argue that the spaces with a unique greedy basis in Sect. \[Orlicz\] were rigged to be close to $\ell_{1}$ and that this facilitates their having a unique greedy basis. Perhaps. We accepted the challenge and in Sections \[superrefExam\] and \[Lorentz\] we find other non-trivial examples of spaces with unique greedy basis, this time far from $\ell_{1}$. The example provided in Sect. \[superrefExam\] is even super-reflexive. In Sect. \[Lorentz\] we show that the separable parts of the weak-$\ell_p$ sequence spaces (in contrast to the result for the $\ell_p$ spaces mentioned above) have a unique greedy basis.
It should be pointed out, though, that the spaces with a unique greedy basis obtained in Sections \[Orlicz\], \[superrefExam\] and \[Lorentz\] have a symmetric basis. Thus, the main unanswered question in this line of research seems to be:
\[problemon22\] Does there exist a space with a non-symmetric unique greedy basis up to a permutation but not a unique unconditional basis up to a permutation?
In what follows we employ the standard notation and terminology commonly used in Banach space theory (see, e.g., [@AlbiacKalton2006]). A [*basis*]{} always means a Schauder basis, and all bases will be assumed to be semi-normalized, i.e., the norm of their elements is uniformly bounded above and below. Given families of positive real numbers $(\alpha_i)_{i\in I}$ and $(\beta_i)_{i\in I}$, the notation $\alpha_i\lesssim \beta_i$ for all $i\in I$ means that $\sup_{i\in I}\alpha_i/\beta_i <\infty$, while $\alpha_i\approx \beta_i$ for all $i\in I$ means that $\alpha_i\lesssim \beta_i$ and $\beta_i\lesssim \alpha_i$ for all $i\in I$. Applied to Banach spaces ${\ensuremath{\mathbb{X}}}$ and ${\ensuremath{\mathbb{Y}}}$, the notation ${\ensuremath{\mathbb{X}}}\approx {\ensuremath{\mathbb{Y}}}$ means that ${\ensuremath{\mathbb{X}}}$ and ${\ensuremath{\mathbb{Y}}}$ are isomorphic.
To quantify the democracy of a basis ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{e}}}_n)_{n=1}^\infty$ we will consider the [*upper democracy function*]{} (also called the [*fundamental function*]{}) of ${\ensuremath{\mathcal{B}}}$ in ${\ensuremath{\mathbb{X}}}$, given by $$\varphi_{u}[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](N)=\sup_{|A|\le N}\left\Vert \sum_{n\in A}{\ensuremath{\mathbf{x}}}_n\right\Vert, \qquad N=1,2,\dots,$$ and the [*lower democracy function*]{} of ${\ensuremath{\mathcal{B}}}$ in ${\ensuremath{\mathbb{X}}}$, $$\varphi_{l}[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](N)=\inf_{|A|\ge N}\left\Vert \sum_{n\in A}{\ensuremath{\mathbf{x}}}_n\right\Vert,\qquad N=1,2,\dots$$ A basis ${\ensuremath{\mathcal{B}}}$ is democratic if and only if $\varphi_u[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](N)\approx \varphi_l[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](N)$ for all $N\in{\ensuremath{\mathbb{N}}}$. Any other more specific notation will be introduced in context when needed.
Preliminaries {#Prelim}
=============
We summarize some basic facts about Musielak-Orlicz sequence spaces [@Musielak1983] adapted to our needs.
A (normalized) *convex Orlicz function* is a nonnnegative and nonzero convex function $F\colon[0,\infty)\to[0,\infty)$ such that $F(0)=0$ and $F(1)=1$.
Let $\omega$ be the set of real-valued sequences.
Given a (normalized) *convex Musielak-Orlicz sequence* $(F_n)_{n=1}^\infty$, i.e., $(F_n)_{n=1}^\infty$ are convex Orlicz functions, we define the *Musielak-Orlicz modular* $m_{(F_n)}\colon \omega \to [0,\infty]$ by $$\label{modularMuseilak}
(a_n)_{n=1}^\infty \mapsto m_{(F_n)}((a_n)_{n=1}^\infty))=\sum_{n=1}^\infty F_n(|a_n|).$$ The *Musielak-Orlicz norm* $\Vert \cdot \Vert_{(F_n)}\colon \omega \to [0,\infty]$ is then given by $$\label{normMuseilak} \alpha \mapsto \inf\{t\in(0,\infty) {\, : \,}m_{(F_n)}(\alpha/t) \le 1\},$$ and its asociated (full) *Musielak-Orlicz space* is $({\ell}_{(F_n)},\Vert \cdot \Vert_{(F_n)})$, where $$\label{fullMuseilak}
{\ell}_{(F_n)} =\{ \alpha\in\omega {\, : \,}m_{(F_n)}(\alpha/t)<\infty \text{ for some } t>0\}.$$ The space $( {\ell}_{(F_n)},\Vert \cdot \Vert_{(F_n)} )$ is a Banach space and the canonical unit vectors $({\ensuremath{\mathbf{e}}}_n)_{n=1}^\infty$ form a $1$-unconditonal basic sequence in $ {\ell}_{(F_n)}.$
The (restricted) Musielak-Orlicz sequence space, denoted by ${\mathit{h}}_{(F_n)}$, is the closed linear span of $({\ensuremath{\mathbf{e}}}_n)_{n=1}^\infty$ in ${\ell}_{(F_n)}$. We have $$\label{restrictedMuseilak}
{\mathit{h}}_{(F_n)}=\{ \alpha\in\omega {\, : \,}m_{(F_n)}(\alpha/t)<\infty \text{ for all } t>0\}.$$ Musielak proved the following result:
\[MusielakDensity\] Let $=(F_n)_{n=1}^\infty$ be a convex Musielak-Orlicz sequence. Then ${\ell}_{(F_n)}={\mathit{h}}_{(F_n)}$ if and only if there exist a positive sequence $(a_n)_{n=1}^\infty \in \ell_1$, some $0<\delta<\infty$, and a constant $1<C<\infty$ such that $$F_n(t) < \delta \Rightarrow F_n(2 t) \le C F_n(t) + a_n.$$
A similar theorem characterizes the inclusions (hence, the identifications) between different Musielak-Orlicz sequence spaces.
\[MusielakInclusion\] Let $(F_n)_{n=1}^\infty$ and $(G_n)_{n=1}^\infty$ be two convex Musielak-Orlicz sequences. Then ${\ell}_{(F_n)}\subseteq {\ell}_{(G_n)}$ if and only if there exist a positive sequence $(a_n)_{n=1}^\infty \in \ell_1$, some $\delta >0$, and positive constants $b$ and $C$ such that $$G_n(t) < \delta \Rightarrow F_n(t) \le C G_n(b t) + a_n.$$
Note that a Musielak-Orlicz norm is determined by its values on $c_{00}$. Therefore, ${\ell}_{(F_n)}\subseteq {\ell}_{(G_n)}$ if and only if ${\mathit{h}}_{(F_n)}\subseteq {\mathit{h}}_{(G_n)}$.
Given $A\subseteq {\ensuremath{\mathbb{N}}}$ we denote by ${\mathit{h}}_{(F_n)}[A]$ (respectively, ${\ell}_{(F_n)}[A]$) the complemented subspace of ${\mathit{h}}_{(F_n)}$ (respectively, complemented of ${\ell}_{(F_n)}$) consisting of all sequences of the space supported in $A$. If $\pi\colon{\ensuremath{\mathbb{N}}}\to {\ensuremath{\mathbb{N}}}$ is an injective mapping and $A=\pi({\ensuremath{\mathbb{N}}})$, the sequence $({\ensuremath{\mathbf{e}}}_{\pi(n)})_{n=1}^\infty$ is a basis of ${\mathit{h}}_{(F_n)}[A]$ isometrically equivalent to the unit vector basis of the Musielak-Orlicz space ${\mathit{h}}_{(F_{\pi(n)})}$. This elementary and useful property can be formulated in terms of direct sums: given two convex Musielak-Orlicz sequences $(F_n)_{n=1}^\infty$ and $(G_n)_{n=1}^\infty$ the unit vector basis of ${\mathit{h}}_{(F_n)}\oplus {\mathit{h}}_{(G_n)}$ is equivalent to the unit vector basis of ${\mathit{h}}_{(H_n)}$, where $$(H_n)_{n=1}^\infty=(F_1,G_1,F_2,G_2,\dots, H_{2n-1},H_{2n},\dots)$$ This, combined with the asymmetry in the canonical basis caused by the variation in $n$ of the Orlicz function $F_n$, makes of Musielak-Orlicz spaces a suitable place to search for bases which are not equivalent to their square (cf. [@CasazzaKalton1998]\*[Proposition 6.8]{}).
A Banach sequence space ${\ensuremath{\mathbb{X}}}$ is called *right dominant* if there is a constant $C$ such that whenever $(x_j)_{j=1}^N$ and $(y_j)_{j=1}^N$ in ${\ensuremath{\mathbb{X}}}$ are such that ${\mathop{\mathrm{supp}}\nolimits}x_i\cap{\mathop{\mathrm{supp}}\nolimits}x_j={\mathop{\mathrm{supp}}\nolimits}y_i\cap{\mathop{\mathrm{supp}}\nolimits}y_j=\emptyset$ for $i\not=j$ and, for $j=1,\dots, N$ we have $ \Vert x_j \Vert \le \Vert y_j \Vert$ and $\max\{n \colon n\in{\mathop{\mathrm{supp}}\nolimits}x_j \} < \min\{n \colon n\in{\mathop{\mathrm{supp}}\nolimits}y_j \}$ it implies $$\left\Vert \sum_{j=1}^N x_j \right\Vert \le C \left\Vert \sum_{j=1}^N y_j \right\Vert.$$
A Banach space ${\ensuremath{\mathbb{X}}}$ is called *sufficiently (lattice) Euclidean* if there is a constant $\lambda$ such that for any $n$ there are operators $S_n\colon X\to \ell_2^n$ and $T_n\colon \ell_2^n \to X$ such that $S_n$ and $T_n$ are a (lattice) homomorphism, $S_n\circ T_n ={\ensuremath{\mathrm{Id}}}_{\ell_2^n}$, and $\Vert S_n \Vert \Vert T_n \Vert \le \lambda$. We will say that ${\ensuremath{\mathbb{X}}}$ is *(lattice) anti-Euclidean* if it is not sufficiently (lattice) Euclidean. An unconditional basic sequence ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_{n})_{n=1}^{\infty}$ induces a lattice structure on its closed linear span ${\ensuremath{\mathbb{Y}}}=[{\ensuremath{\mathbf{x}}}_{n}\colon n\in {\ensuremath{\mathbb{N}}}]$. We say that ${\ensuremath{\mathcal{B}}}$ is lattice anti-Euclidean if ${\ensuremath{\mathbb{Y}}}$ is lattice anti-Euclidean.
Existence of greedy bases in Nakano spaces {#Nakano}
==========================================
If $1\le p<\infty$ the map $t\mapsto t^p$ is a convex Orlicz function. Given a sequence $(p_n)_{n=1}^\infty$ in $[1,\infty)$, the Musielak-Orlicz spaces defined as in and for the sequence of Orlicz functions $(F_n)_{n=1}^\infty$, where $$F_n(t)=t^{p_n}$$ are denoted them by ${\ell}_{(p_n)}$ and ${\mathit{h}}_{(p_n)}$ and are called [*Nakano spaces*]{}. The corresponding modular and norm, as defined in and , for this particular case will be denoted respectively by $m_{(p_n)}$ and $\Vert \cdot \Vert_{(p_n)}$.
These spaces were introduced by Nakano in 1950 [@Nakano1950]. In [@Nakano1951], he completely characterized the inclusions between these spaces. In 1965, Simmons [@Simmons1965] obtained a similar result in the non-locally convex setting (i.e., when $p_n\le 1$ for all $n$). We refer to [@BlascoGregori2001] for a proof of Theorem \[NakanoInclusion\] based on Theorem \[MusielakInclusion\].
\[NakanoInclusion\] Let $(p_n)_{n=1}^\infty$ and $(q_n)_{n=1}^\infty$ be two sequences in $[1,\infty)$. Then ${\ell}_{(p_n)} \subseteq {\ell}_{(q_n)} $ (with continuous inclusion) if and only if there is $0<r<1$ such that $$\sum_{q_n<p_n} r^{ p_n q_n / (p_n -q_n)} <\infty.$$
\[NakanoIdentification\] Let $(p_n)_{n=1}^\infty$ and $(q_n)_{n=1}^\infty$ be two Nakano indices. Then ${\ell}_{(p_n)} = {\ell}_{(q_n)} $ (with equivalent norms) if and only if there is $0<r<1$ such that $$\sum_{n=1}^\infty r^{ p_n q_n / |p_n -q_n|} <\infty.$$
Notice that inclusions between Nakano spaces work as in the $\ell_p$ spaces, in the sense that if $(p_n)_{n=1}^\infty$ and $(q_n)_{n=1}^\infty$ are Nakano indices satisfying $p_n\le q_n$ for all $n\in{\ensuremath{\mathbb{N}}}$, then ${\ell}_{(p_n)} \subseteq{\ell}_{(q_n)} $. We will need a more quantitative formulation of this result valid even for *finite dimensional Nakano spaces*.
\[NormOneInequality\]Let $\eta\in{\ensuremath{\mathbb{N}}}\cup\{\infty\}$. Let $(p_n)_{n=1}^\eta$, $(q_n)_{n=1}^\eta$ such that $1\le p_n\le q_n <\infty$. Then $$\left\Vert (a_n)_{n=1}^\eta \right\Vert_{(q_n)} \le \left\Vert (a_n)_{n=1}^\eta \right\Vert_{(p_n)}$$ for all $(a_n)_{n=1}^\eta$ in ${\ensuremath{\mathbb{R}}}$.
The following theorem is crucial in Casazza-Kalton’s discussion on Nakano spaces in [@CasazzaKalton1998].
\[RightDom\] Let $(p_n)_{n=1}^\infty$ be a decreasing sequence in $[1,\infty)$. Then ${\mathit{h}}_{(p_n)}$ is a right-dominant sequence space.
The spaces $\ell_p$ for $1\le p<\infty$ are Nakano spaces ${\ell}_{(p_n)}$ in which $(p_{n})$ is the constant sequence $p$, while the space $\ell_\infty$ is not, a priori, a Nakano space. However, we can state a theorem similar to Corollary \[NakanoIdentification\] that characterizes when ${\ell}_{(p_n)}$ coincides with the space $\ell_\infty$ (i.e., the corresponding Nakano space ${\mathit{h}}_{(p_n)}$ agrees with $c_0$). The proof is similar to that of Theorem \[NakanoInclusion\].
\[NakanoInftyIdentification\] Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$. Then ${\ell}_{(p_n)}=\ell_\infty$ (with equivalent norms) if and only if there is $0<r<1$ such that $ \sum_{n=1}^\infty r^{p_n} <\infty$.
Let us next enunciate a result that characterizes when ${\ell}_{(p_n)}$ and ${\mathit{h}}_{(p_n)}$ agree.
\[NakanoDensity\] Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$. Then ${\ell}_{(p_n)}={\mathit{h}}_{(p_n)}$ if and only if $\sup_n p_n<\infty$.
The same condition as in Theorem \[NakanoDensity\] allows us to characterize boundedness in Nakano spaces without appealing to the Nakano norm.
\[NakanoBoundedness\] Suppose $(p_n)_{n=1}^\infty$ is a sequence in $[1,\infty)$ such that $\sup_n p_n<\infty$. Let $\mathcal{A}$ be a subset in ${\ell}_{(p_n)}$. Then:
1. $\mathcal{A}$ is norm-bounded above (i.e., $\sup_{\alpha\in \mathcal{A}} \Vert \alpha\Vert_{(p_n)}<\infty$) if and only if it is modular-bounded above (i.e., $\sup_{\alpha\in \mathcal{A}} m_{(p_n)}(\alpha)<\infty$).
2. $\mathcal{A}$ is norm-bounded below (i.e., $\inf_{\alpha\in \mathcal{A}} \Vert \alpha\Vert_{(p_n)}>0$) if and only if it is modular-bounded below ( i.e., $\inf_{\alpha\in \mathcal{A}} m_{(p_n)}(\alpha)>0$).
Let $s=\sup_n p_n$. The result is an easy consequence of the estimates $$\min\{ \Vert \alpha \Vert_{(p_n)} , \Vert \alpha \Vert_{(p_n)}^s \} \le m_{(p_n)}(\alpha) \le \max\{ \Vert \alpha \Vert_{(p_n)} ,
\Vert \alpha \Vert_{(p_n)}^s \}$$ for any $\alpha\in\omega$.
We would like to point out the close connection between Proposition \[NakanoBoundedness\] and the equivalence between the *norm convergence* and the *modular convergence* obtained in [@Musielak1983]\*[Theorem 8.14]{}. However, Proposition \[NakanoBoundedness\] provides a formulation more fit for our purposes.
Duality in Nakano space works as expected. We refer the reader to [@Musielak1983]\*[Theorem 13.11]{} for a more general result in the setting of Musielak-Orlicz spaces.
\[NakanoDuality\] Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$. Let $(q_n)_{n=1}^\infty$ be a sequence in $(1,\infty)$ such that $1/p_n+1/q_n=1$ if $p_n>1$ and, for some $0<r<1$, $\sum_{p_n=1} r^{q_n}<\infty$. Then ${\mathit{h}}_{(p_n)}^*={\ell}_{(q_n)}$ with the natural duality pair and equivalent norms.
We are now in a position to state and prove our first results about bases in Nakano spaces.
\[clusterPointDemocracyEstimate\]Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$ and let $p$ be a cluster point of $(p_n)_{n=1}^\infty$. Denote by ${\ensuremath{\mathcal{E}}}$ the unit vector basis. Then, for all $N\in{\ensuremath{\mathbb{N}}}$, $$\varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N)\le N^{1/p} \le \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N).$$
Fix $N\in{\ensuremath{\mathbb{N}}}$. Let $r<p<s$. There is $A\subseteq{\ensuremath{\mathbb{N}}}$ such that $|A|=N$ and $r<p_n<s$ for all $n\in A$. By Lemma \[NormOneInequality\], $$N^{1/s}= \left\Vert \sum_{n\in A} {\ensuremath{\mathbf{e}}}_n\right\Vert_{s} \le \left\Vert \sum_{n\in A} {\ensuremath{\mathbf{e}}}_n\right\Vert_{(p_n)} \le \left\Vert \sum_{n\in A} {\ensuremath{\mathbf{e}}}_n\right\Vert_{r}= N^{1/r}.$$ Hence, $$\varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N) \le N^{1/r}\quad \text{and}\quad \quad N^{1/s} \le \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N).$$ Choosing $r$ and $s$ arbitrarily close to $p$ we get the desired result.
\[uvbNakanoGreedy\]Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$. Then the unit vector basis is a greedy basis for the Nakano space ${\mathit{h}}_{(p_n)}$ if and only if there is $p\in[1,\infty]$ such that ${\mathit{h}}_{(p_n)}=\ell_p$, i.e., if and only if the unit vector basis of ${\mathit{h}}_{(p_n)}$ is equivalent to the unit vector basis of $\ell_p$ ($c_0$ if $p=\infty$), in which case $\lim_n p_n=p$.
Suppose that the unit vector basis ${\ensuremath{\mathcal{E}}}=({\ensuremath{\mathbf{e}}}_n)_{n=1}^\infty$ is a greedy basis of ${\mathit{h}}_{(p_n)}$. Then, in particular, ${\ensuremath{\mathcal{E}}}$ is democratic.
The sequence $(p_n)_{n=1}^\infty$ decreases to $p\in[1,\infty)$.
Let $N\in{\ensuremath{\mathbb{N}}}$. By Lemma \[NormOneInequality\]. $$\left\Vert \sum_{n=1}^N {\ensuremath{\mathbf{e}}}_n\right\Vert_{(p_n)} \le \varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N)\le \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N) \le
\left\Vert \sum_{n=1}^N {\ensuremath{\mathbf{e}}}_n\right\Vert_{p}
= N^{1/p}.$$
Hence, by Lemma \[clusterPointDemocracyEstimate\], $$\varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N) = \left\Vert \sum_{n=1}^N {\ensuremath{\mathbf{e}}}_n\right\Vert_{(p_n)}\quad \text{and}\quad \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N)=N^{1/p}.$$
Since ${\ensuremath{\mathcal{E}}}$ is democratic, $$\inf_N \frac{ \varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N) }{ \varphi_u [{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N) } =\inf_{N\in{\ensuremath{\mathbb{N}}}} \left\{ N^{-1/p} \left \Vert \sum_{n=1}^N {\ensuremath{\mathbf{e}}}_n \right\Vert_{(p_n)} {\, : \,}N\in{\ensuremath{\mathbb{N}}}\right\}>0.$$ By Lemma \[NakanoBoundedness\], there is $0<c<1$ such that $$\sum_{n=1}^N N^{-p_n/p}\ge c, \quad N=1,2,\dots$$ Therefore, since $p_n\ge p_N$ for $n\le N$, $$N^{1-p_N/p}\ge c, \quad N=1,2,\dots$$ Hence, $$\frac{1}{p}-\frac{1}{p_N} \le \frac{p_N}{p}-1\le \log\left(\frac{1}{c}\right) \frac{1}{\log(N)}, \quad N=1,2,\dots$$ If $0<r<c$, $$\sum_{N=1}^\infty r^{p p_N/(p_N-p)}\le\sum_{N=1}^\infty N^{-\log (r)/log(c)}<\infty.$$ By Theorem \[NakanoIdentification\], ${\mathit{h}}_{(p_n)}=\ell_p$.
The sequence $(p_n)_{n=1}^\infty$ increases to $p\in[1,\infty)$.
It is similar to Case 1 and we leave the details for the reader.
The sequence $(p_n)_{n=1}^\infty$ converges to $p\in[1,\infty)$.
Consider $A_1=\{n\in{\ensuremath{\mathbb{N}}}{\, : \,}p_n\le p\}$ and $A_2=\{ n\in{\ensuremath{\mathbb{N}}}{\, : \,}p_n>p\}$. Denote ${\ensuremath{\mathbb{N}}}_j=\{ n\in{\ensuremath{\mathbb{N}}}{\, : \,}n \le |A_j|\}$ ($j=1,2$). There is an increasing bijection from ${\ensuremath{\mathbb{N}}}_1$ onto $A_1$. Appealing to the Case 2 we obtain ${\mathit{h}}_{(p_n)}[A_1]=\ell_p[A_1]$. Similarly, there is a decreasing bijection from ${\ensuremath{\mathbb{N}}}_2$ onto $A_2$. Appealing to the case (a) we obtain ${\mathit{h}}_{(p_n)}[A_2]=\ell_p[A_2]$. Hence ${\mathit{h}}_{(p_n)}={\mathit{h}}_{(p_n)}[A_1]\oplus {\mathit{h}}_{(p_n)}[A_2]=\ell_p[A_1]\oplus \ell_p[A_2]=\ell_p$.
$\lim_n p_n=\infty$.
Combining the democracy of the ${\ensuremath{\mathcal{E}}}$ with Lemma \[clusterPointDemocracyEstimate\], we obtain that $\varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_{(p_n)}](N)\lesssim 1$ for all $N\in{\ensuremath{\mathbb{N}}}$. Taking into account that ${\ensuremath{\mathcal{E}}}$ is an unconditional basis, we get ${\mathit{h}}_{(p_n)}=c_0$.
The sequence $(p_n)_{n=1}^\infty$ has no limit.
Denote $p_1=\liminf_n p_n<p_2=\limsup_n p_n$. Combining the democracy of ${\ensuremath{\mathcal{E}}}$ with Lemma \[clusterPointDemocracyEstimate\] we get $N^{1/p_1} \lesssim N^{1/p_2}$ for $N\in{\ensuremath{\mathbb{N}}}$, an absurdity.
The proof of the above theorem gives that for a monotone sequence $(p_n)_{n=1}^\infty$ converging to $p<\infty$, one has $\ell_{(p_n)}=\ell_p$ if and only if $$\sup_{n\in{\ensuremath{\mathbb{N}}}} \log(n) |p_n-p| <\infty.$$ A similar result was obtained by Simmons [@Simmons1965] in the non-locally convex setting.
\[GreedyNakano\] Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$ with $\lim_n p_n =1$.
- Any complemented greedy basic sequence in ${\ell}_{(p_n)}$ is equivalent to the unit vector basis of $\ell_1$.
- A complemented subspace of ${\ell}_{(p_n)}$ has a greedy basis if and only if it is isomorphic to $\ell_1$.
Part (b) follows readily from (a). To prove (a), let us assume without loss of generality that $p_n \searrow 1$. Then, by Theorem \[RightDom\], ${\ell}_{(p_n)}$ is right-dominant. Theorem \[NakanoIdentification\] yields that the unit vector basis of ${\ell}_{(p_n)}$ has a subsequence equivalent to the unit vector basis of $\ell_1$. In particular, $\ell_1$ is disjointly finitely representable in ${\ell}_{(p_n)}$. Taking into account [@CasazzaKalton1998]\*[Proposition 5.3]{}, we obtain that ${\ell}_{(p_n)}$ is anti-Euclidean. Let ${\ensuremath{\mathcal{B}}}$ be a complemented greedy basis in ${\ell}_{(p_n)}$. Then ${\ensuremath{\mathcal{B}}}$ is lattice anti-Euclidean. By [@CasazzaKalton1998]\*[Theorem 3.5]{}, ${\ensuremath{\mathcal{B}}}$ is equivalent to a complemented block basis of the unit vector basis of ${\ell}_{(p_n)}^M$ for some $M\in{\ensuremath{\mathbb{N}}}$.
Pick $(q_n)_{n=1}^\infty$ such that $q_n=p_m$ whenever $(m-1)M+1\le n \le mM$. Notice that ${\ell}_{(p_n)}^M={\ell}_{(q_n)}$. In particular, ${\ell}_{(p_n)}^M$ is a right-dominant sequence space. Appealing to [@CasazzaKalton1998]\*[Theorem 5.6]{} we get that ${\ensuremath{\mathcal{B}}}$ is permutatively equivalent to a subsequence of the unit vector basis of ${\ell}_{(q_n)}$. Therefore, there is an injective mapping $\pi\colon{\ensuremath{\mathbb{N}}}\to {\ensuremath{\mathbb{N}}}$ such that ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of the Nakano space ${\ell}_{(q_{\pi(n)})}$. Hence, by Theorem \[uvbNakanoGreedy\], ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $\ell_1$.
\[UniqueGreedyNakano\] Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$ such that $\lim_n p_n =1$ and $$\sum_{n=1}^\infty r^{1/|p_n-1|}=\infty \text{ for all } 0<r<1.$$ Then ${\ell}_{(p_n)}$ does not have a greedy basis.
Suppose that ${\ensuremath{\mathcal{B}}}$ is a greedy basis of ${\ell}_{(p_n)}$. By Theorem \[GreedyNakano\], ${\ell}_{(p_n)} \approx \ell_1$. Since $\ell_1$ has a unique unconditional basis (cf. [@LindenstraussPel1968]), ${\ell}_{(p_n)}=\ell_1$, in contradiction with Theorem \[NakanoIdentification\].
Next we obtain analogous results to Theorem \[GreedyNakano\] and Corollary \[UniqueGreedyNakano\] for the dual case, i.e., when $\lim_n p_n=\infty$. Although we could use similar techniques in their proofs, we will get more with simpler techniques.
Let $(p_n)_{n=1}^\infty$ be a sequence in $[1,\infty)$ such that $\lim_n p_n =\infty$.
1. Any greedy basic sequence in ${\mathit{h}}_{(p_n)}$ is equivalent to the unit vector basis of $c_0$.
2. A subspace of ${\mathit{h}}_{(p_n)}$ has a greedy basis if and only if it is isomorphic to $c_0$.
3. The space ${\mathit{h}}_{(p_n)}$ has a greedy basis if and only if ${\mathit{h}}_{(p_n)}=c_0$.
\(a) Let ${\ensuremath{\mathcal{B}}}$ be a greedy basis sequence in ${\mathit{h}}_{(p_n)}$. Since the dual space of ${\mathit{h}}_{(p_n)}$ is separable (see Theorem \[NakanoDuality\] and Theorem \[NakanoDensity\]), ${\ensuremath{\mathcal{B}}}$ is a weakly null sequence (see [@AlbiacKalton2006]\*[Proposition 3.2.7]{} and [@AlbiacKalton2006]\*[Theorem 3.3.1]{}). Applying Bessaga-Pe[ł]{}czy[ń]{}ski Selection Principle (cf. [@AlbiacKalton2006]\*[Proposition 1.3.10]{}) we get a normalized block basis of the the unit vector basis of ${\mathit{h}}_{(p_n)}$, say ${\ensuremath{\mathcal{B}}}_0=({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$, equivalent to a subbasis of ${\ensuremath{\mathcal{B}}}$. Denote $A_k={\mathop{\mathrm{supp}}\nolimits}{\ensuremath{\mathbf{x}}}_k$ and pick $n_k\in A_k$ such that $q_k:=p_{n_k}\le p_n$ for $n\in A_k$. Passing to a subsequence, if necessary, we can suppose that $\sum_{k=1}^\infty r^{q_k}<\infty$ for some $0<r<1$. By Theorem \[NakanoInftyIdentification\], ${\mathit{h}}_{(q_k)}=c_0$. Then $$\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \right\Vert_{(p_n)} \le \left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{e}}}_k \right\Vert_{(q_k)} \approx
\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{e}}}_k \right\Vert_\infty,$$ for all $(a_n)_{n=1}^\infty\in c_{00}$. Therefore, $$\varphi_u[{\ensuremath{\mathcal{B}}}, {\mathit{h}}_{(p_n)}] (N) \approx \varphi_u[{\ensuremath{\mathcal{B}}}_0, {\mathit{h}}_{(p_n)}] (N) \lesssim \varphi_u[{\ensuremath{\mathcal{E}}}, c_0] (N)=1.$$ From here, taking into account that ${\ensuremath{\mathcal{B}}}$ is an unconditional basis, we get readily that ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $c_0$.
Part (b) is an easy consequence of (a) and Part (c) follows from (b) and the uniqueness of unconditional basis in $c_0$.
Uniqueness of greedy basis in Orlicz sequence spaces {#Orlicz}
====================================================
Orlicz sequence spaces can be seen as a particular case of Musielak-Orlicz sequence spaces. Indeed, we just need to consider a sequence $(F_n)_{n=1}^\infty$ such that $F_n=F$ for all $n$ and some convex Orlicz function. We put ${\ell}_{(F_{n})}={\ell}_F$ and ${\mathit{h}}_{(F_{n})}={\mathit{h}}_F$. The identification between ${\ell}_F$ and ${\mathit{h}}_F$ is simplier than for Musielak-Orlicz sequence spaces: ${\ell}_F={\mathit{h}}_F$ if and only if $F$ satisfies the $\Delta_2$ condition at the origin, i.e., there exist constants $a\in(0,\infty)$ and $C\in(1,\infty)$ such that $F(2t) \le C F(t)$ for $t\in[0,a]$. Notice that Orlicz spaces only depend, up to an equivalent norm, of the behavior of the functions defining them at a neighborhood of the origin. To be precise, ${\ell}_{F}={\ell}_{G}$ if and only if there exist positive constants $a$ and $b$ such that $F(b t)\approx G(t)$ for all $t\in[0,a]$.
In an Orlicz space ${\mathit{h}}_F$, the unit vector basis is a $1$-symmetric basis. In particular, it is a greedy basis. Its democracy functions $\varphi_l[{\ensuremath{\mathcal{E}}},{\mathit{h}}_F](N)=\varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_F](N)=D_N$ are determined by the formula $$F\left( \frac{1}{D_N}\right)=\frac{1}{N}, \quad N=1,2,\dots$$ From here it is easy to obtain the following result.
\[SameDemocracyFunction\] Let $F$ and $G$ be two convex Orlicz functions. Then ${\mathit{h}}_F={\mathit{h}}_G$ (with equivalence of norms) if and only if $ \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_F](N)\approx \varphi_u[{\ensuremath{\mathcal{E}}},{\mathit{h}}_G](N)$ for all $N\in \mathbb N$.
Another elementary property of interest for us is that the unit vector basis of ${\mathit{h}}_F\oplus {\mathit{h}}_F$ is equivalent to the unit vector basis of ${\mathit{h}}_F$.
We will need to consider Musielak-Orlicz sequence spaces arising from the flows of an Orlicz function. To be precise, fix a convex Orlicz function $F$, and for $0<s<\infty$ consider $$F_s\colon[0,\infty)\to[0,\infty), \quad t\mapsto \frac{F(st)}{F(s)}.$$ Given a sequence $(s_n)_{n=1}^\infty$ in $(0,\infty)$ we define $${\mathit{h}}_F[s_n]:={\mathit{h}}_{(F_{s_n})}.$$ The following result, implicitly stated in [@CasazzaKalton1998], establishes the connection between this kind of Musielak-Orlicz sequence spaces and block bases in Orlicz sequence spaces.
\[BlockOrliczEquivalentMusielak\] Let $F$ be a convex Orlicz function.
1. Let ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_n)_{n=1}^\infty$ be a constant-coefficient normalized block basic sequence of the unit vector basis of ${\mathit{h}}_F$. For each $n\in \mathbb N$, denote by $N_n$ the lenght of the block ${\ensuremath{\mathbf{x}}}_n$ and let $s_n\in(0,\infty)$ be such that $N_n F(s_n) =1$. Then ${\ensuremath{\mathcal{B}}}$ is a complemented basic sequence in ${\mathit{h}}_F$ isometrically equivalent to the unit vector basis of ${\mathit{h}}_F[s_n]$.
2. Let $(s_n)_{n=1}^\infty$ be a bounded sequence of positive numbers. Then the unit vector basis of ${\mathit{h}}_F[s_n]$ is equivalent to a constant-coefficient block basis of the unit vector basis of ${\mathit{h}}_F$.
Next we focus on convex Orlicz functions that are equivalent at the origin to $t\mapsto t^p(-\log t)^{-a}$ for some $1\le p$ and $a>0$. To be precise, put $$\begin{aligned}
\label{OurOrliczFunction}
F^{p,a}(t)=\begin{cases}
e^{-pa} t^p(-\log t)^{-a} &\text{if}\; 0<t <\frac{1}{e}, \\
t^{p+a} &\text{if}\; \frac{1}{e}\le t <\infty.
\end{cases}\end{aligned}$$ Let us recall some properties of these Orlicz functions. Denote $F=F^{p,a}$.
- $F$ is multiplicatively convex, i.e., $$F(s^\theta t^{1-\theta})\le F(s)^\theta F(t)^{1-\theta} \text{ whenever }0<s,t,\theta<1.$$
- ${\ell}_F={\mathit{h}}_F$.
- If $p=1$, ${\ell}_{F}$ is anti-Euclidean (cf. [@CasazzaKalton1998]\*[Lemma 6.2]{}).
- ${\ell}_{F}$ and $\ell_p$ are the unique Orlicz spaces that are subspaces of ${\ell}_{F}$ (cf. [@LinTza1977]\*[Theorem 4.a.8]{}).
- ${\ell}_{F}$ has a unique symmetric basis (cf. [@LinTza1977]\*[Proposition 4.b.10]{}). However, ${\ell}_{F}$ does not have a unique unconditional basis (cf. [@LindenstraussZippin1969]).
\[complementedlp\] Let $F=F^{p,a}$ and suppose $\lim_n s_n=0$. It follows from [@CasazzaKalton1998]\*[Proposition 6.6]{} that the unit vector basis of ${\mathit{h}}_{F}[s_n]$ has a subbasis equivalent to the unit vector basis of $\ell_p$. Combining with Lemma \[BlockOrliczEquivalentMusielak\] we obtain that $\ell_p$ is a complemented subspace of ${\ell}_F$.
\[MusielakGreedy\] Let $p\in[1,\infty)$ and $a\in(0,\infty)$. Let $(s_n)_{n=1}^\infty$ be a sequence such that $\lim_n s_n = 0$. Consider $F=F^{p,a}$ defined as is . The following are equivalent:
1. ${\mathit{h}}_{F}[s_n]=\ell_p$, i.e., the unit vector bases of ${\mathit{h}}_{F}[s_n]$ and $\ell_p$ are equivalent.
2. The unit vector basis of ${\mathit{h}}_{F}[s_n]$ is greedy.
3. There is a constant $R>1$ such that $$\label{targetineq}
|\{ n\in {\ensuremath{\mathbb{N}}}{\, : \,}s_n \ge \exp(-2^k) \}| \le R^{2^k},\quad \forall\;k\in{\ensuremath{\mathbb{N}}}.$$
That (a) implies (b) is obvious, and that (c) implies (a) is established in [@CasazzaKalton1998]\*[Proposition 6.6(3)]{}. Let us show that (b) implies (c). Assume, without lost of generality, that $s_n\le e^{-1}$ for all $n\in{\ensuremath{\mathbb{N}}}$. Since the unit vector basis ${\ensuremath{\mathcal{E}}}$ of ${\mathit{h}}_{F}[s_n]$ has a subbasis equivalent to the unit vector basis of $\ell_p$ (see Remark \[complementedlp\]) and ${\ensuremath{\mathcal{E}}}$ is democratic, there is a constant $0<c<1$ such that $$\label{initialbound}
c N^{1/p}\leq \varphi_l[{\ensuremath{\mathcal{E}}}, {\mathit{h}}_{F}[s_n]](N), \quad N\in{\ensuremath{\mathbb{N}}}.$$ For each $k\in{\ensuremath{\mathbb{N}}}$ let $A_k=\{ n\in {\ensuremath{\mathbb{N}}}{\, : \,}s_n \ge \exp(-2^k) \}$ and put $N_k=|A_k|$. It suffices to prove for $k$ such that $cN_k\ge 1$.
By we have $$c N_k^{1/p}\le \left\Vert \sum_{n\in A_k} {\ensuremath{\mathbf{e}}}_n \right\Vert_{ {\mathit{h}}_{F}[s_n]}.$$Hence, $$\begin{aligned}
1 &\le \sum_{n\in A_k} \frac{ F(c^{-1}N_k^{-1/p} s_n)}{F(s_n)}\\
&= c^{-p} N_k^{-1} \sum_{n\in A_k}\left( \frac{-\log(s_n)}{-\log(s_n)+\log(cN_k^{1/p})}\right)^a\\
&\le c^{-p} \left(\frac{2^k}{2^k+\log(cN_k^{1/p})}\right)^a,\end{aligned}$$ which yields $$\log(cN_k^{1/p})\le (c^{-p/a}-1) 2^k.$$ Consequently, choosing $R=c^{-p} \exp\{p (c^{-p/a}-1)\}$ we obtain $
N_k \le R^{2^k}.
$
\[Identification\] Suppose $F$ is a Orlicz function. Let $(s_n)_{n=1}^\infty$ be a bounded sequence of positive numbers which does not converge to zero. There are only two mutually exclusive possibilities:
- $0<\liminf_n s_n\le \limsup_n s_n<\infty$. It is then quite obvious that ${\mathit{h}}_F[s_n]={\mathit{h}}_F$.
- $0=\liminf s_n <\limsup_n s_n<\infty$. Then ${\mathit{h}}_F[s_n] \approx {\mathit{h}}_F$ (cf. [@LinTza1977]\*[Proposition 3.a.5]{}), but this does not mean that the unit vector basis of ${\mathit{h}}_F[s_n]$ is a greedy basis. We deal with this situation in the following corollary.
\[uvbNotGreedy\] Suppose that $p\in[1,\infty)$ and that $a\in(0,\infty)$. Consider $F=F^{p,a}$ as defined in . Let $(s_n)_{n=1}^\infty$ be a sequence such that $0=\liminf_n s_n < \limsup_n s_n<\infty$. Then the unit vector basis of $\ell_F[s_n]$ is not a greedy basis.
Suppose that the unit vector basis of $\ell_F[s_n]$ is greedy. There are injective mappings $\pi,\rho\colon{\ensuremath{\mathbb{N}}}\to {\ensuremath{\mathbb{N}}}$ such that $\lim_n s _{\pi(n)} =0$ and $\inf_n s_{\rho(n)} >0$. Obviously, ${\ensuremath{\mathcal{B}}}_1=({\ensuremath{\mathbf{x}}}_{\pi(n)})_{n=1}^\infty$ is isometricaly equivalent to the unit vector basis of $\ell_F[s_{\pi(n)}]$. Hence, by Proposition \[MusielakGreedy\], ${\ensuremath{\mathcal{B}}}_1$ is equivalent to the unit vector basis of $\ell_p$. Analogously, ${\ensuremath{\mathcal{B}}}_2=({\ensuremath{\mathbf{x}}}_{\rho(n)})_{n=1}^\infty$ is isometricaly equivalent to the unit vector basis of $\ell_F[s_{\rho(n)}]$. Therefore (see Remark \[Identification\]) ${\ensuremath{\mathcal{B}}}_2$ is equivalent to the unit vector basis of ${\ell}_F$. Consequently, denoting by ${\ensuremath{\mathcal{E}}}$ the unit vector basis, $$\varphi_u[{\ensuremath{\mathcal{E}}},\ell_p] (N) \approx \varphi_u[{\ensuremath{\mathcal{B}}}_1,{\ell}_F[s_n])(N) \approx \varphi_u[{\ensuremath{\mathcal{B}}}_2,{\ell}_F[s_n])(N) \approx
\varphi_u[{\ensuremath{\mathcal{E}}},{\ell}_F](N).$$ By Lemma \[SameDemocracyFunction\], $\ell_p\approx {\ell}_F$. Then $t\approx t^p(-\log t)^{-a}$ at the origin, an absurdity.
\[MainOrlicz\] Let $a>0$ and consider $F=F^{1,a}$ as defined in . If ${\ensuremath{\mathcal{B}}}$ is a complemented greedy basic sequence in $\ell_F$, then ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis either of $\ell_1$ or of $\ell_F$. Conversely, $\ell_F$ has a complemented basic sequence equivalent to the unit vector basis of $\ell_1$.
Notice that since ${\ell}_F$ is anti-Euclidean the basic sequence ${\ensuremath{\mathcal{B}}}$ is lattice anti-Euclidean. By [@CasazzaKalton1998]\*[Theorem 3.4]{}, ${\ensuremath{\mathcal{B}}}$ is permutatively equivalent to a block basis of the unit vector basis of ${\ell}_F$. By [@CasazzaKalton1998]\*[Lemma 6.4]{}, ${\ensuremath{\mathcal{B}}}$ is permutatively equivalent to a normalized constant coefficient block basis. Let $(N_n)_{n=1}^\infty$ be the sequence of the lengths of those blocks, and choose $(s_n)_{n=1}^\infty$ such that $N_n F(s_n)=1$. By Lemma \[BlockOrliczEquivalentMusielak\], ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of ${\mathit{h}}_F[s_n]$.
Suppose that $\lim_{n} N_n=\infty$, hence $\lim_{n} s_n=0$. Then, by Proposition \[MusielakGreedy\], ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $\ell_1$.
Suppose that $\sup N_n<\infty$. Then (see Remark \[Identification\]) ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of ${\ell}_F$.
Suppose that $\liminf_n N_n <\limsup N_n=\infty$. Then, by Corollary \[uvbNotGreedy\], the unit vector basis is not greedy.
The converse is established in Remark \[complementedlp\].
Corollaries \[uniquenesslF\] and \[ExistenceComplementedlF\] are inmmediate from Theorem \[MainOrlicz\].
\[uniquenesslF\] Let $a>0$ and consider $F=F^{1,a}$ as defined in . Then $\ell_F$ has a unique greedy basis.
\[ExistenceComplementedlF\] Let $a>0$ and consider $F=F^{1,a}$ as defined in . A complemented subspace of $\ell_F$ has a greedy basis if and only if it is isomorphic either to $\ell_1$ or to $\ell_F$.
Notice that Corollary \[uniquenesslF\] solves Problem \[problemon\].
We end this section providing a much more natural example than the afore-mentioned Pe[ł]{}czy[ń]{}ski’s universal space $U$, of a space with symmetric basis having a complemented subspace without a greedy basis
\[ComplementedNotGreedy\] Let $a>0$ and consider $F=F^{1,a}$ as defined in . There is a complemented subspace of $\ell_F$ with an unconditional basis but with no greedy basis.
Use Proposition \[MusielakGreedy\] to pick a sequence $(s_n)_{n=1}^\infty$ convergent to zero such that the unit vector basis of ${\mathit{h}}_F[s_n]$ is not greedy. Notice that, by Lemma \[BlockOrliczEquivalentMusielak\], ${\mathit{h}}_F[s_n]$ is isomorphic to a complemented subspace of $\ell_F$. Since $\ell_1$ has a unique unconditional basis, ${\mathit{h}}_F[s_n]\not\approx \ell_1$.
If ${\mathit{h}}_F[s_n]$ had a greedy basis, ${\mathit{h}}_F[s_n]$ would be isomorphic to $ \ell_F$ by Corollary \[ExistenceComplementedlF\]. By [@CasazzaKalton1998]\*[Theorem 3.5]{}, the unit vector basis of $\ell_F$ would be equivalent to a complemented block basis of ${\mathit{h}}_F^M[s_n]$ for some $M\in{\ensuremath{\mathbb{N}}}$. By [@CasazzaKalton1998]\*[Lemma 6.4]{}, there would be another sequence $(s_n')_{n=1}^\infty$, decreasing to zero, such that the unit vector basis of $\ell_F$ is permutatively equivalent to the unit vector basis of ${\mathit{h}}_F[s_n']$. In particular, the unit vector basis of ${\mathit{h}}_F[s_n']$ would be greedy. But this is an absurdity because, again by Theorem \[MusielakGreedy\], we would have $\ell_F\approx {\mathit{h}}_F[s_n'] \approx \ell_1$.
A super-reflexive Banach space with a unique greedy basis {#superrefExam}
=========================================================
In this section we continue exploiting the structure of sequence Orlicz spaces. Consider $F=F^{2,1}$ as defined in . The dual of the Orlicz sequence space $\ell_F$, under the canonical dual pair, is the Orlicz sequence space $\ell_G$, where $G$ is a convex Orlicz function such that $$\label{dualfunction}
G(t)= t^2(-\log(t)) \text{ for } 0<t\le e^{-3/2}$$ (cf. [@LinTza1977]\*[Example 4.c.1]{} and [@CasazzaKalton1998]\*[Theorem 6.10]{}). Since both $F$ and $G$ satisfy the $\Delta_2$-condition at the origin it follows that $\ell_F$ and $\ell_G$ are super-reflexive (see e.g. [@LT1979 Sect. 1.f]). Moreover $\ell_G$ has Rademacher type $p$ for any $p< 2$ and cotype $2$ (cf. [@CasazzaKalton1996]). If we try to apply to $\ell_F$ the same techniques as in Sect. \[Orlicz\] we come across the obstruction that these Orlicz spaces are sufficiently Euclidean (in fact $\ell_2$ is a complemented subspace of $\ell_F$, see Remark \[complementedlp\]). In this new situation, it comes to our aid [@CasazzaKalton1996]\*[Theorem 4.3]{} (which asserts that every complemented unconditional basic sequence in $\ell_G$ is equivalent to a constant coefficient block basic sequence of the unit vector basis) and the reflexivity of $\ell_F$ (which guarantees that every semi-normalized basis in $\ell_F$ is weakly null).
\[MainSuperReflexive\] Consider $F=F^{2,1}$ as in and $G$ as in . If ${\ensuremath{\mathcal{B}}}$ is a complemented greedy basic sequence in $\ell_F$ (respectively, $\ell_G$) then ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis either of $\ell_2$ or of $\ell_F$ (respectively, $\ell_G$).
Since $\ell_G^*=\ell_F$, appealing to [@DKKT2003]\*[Theorem 5.1]{} it suffices to prove the result for a complemented greedy basic sequence ${\ensuremath{\mathcal{B}}}$ in $\ell_F$. Notice that the dual basic sequence of a constant coefficient block basic sequence is equivalent to a constant coefficient block basic sequence in the dual space. Hence, by [@CasazzaKalton1996]\*[Theorem 4.3]{}, ${\ensuremath{\mathcal{B}}}$ is permutatively equivalent to a normalized constant coefficient block basic sequence of the unit vector basis. Now, proceed as in the proof of Theorem \[MainOrlicz\].
Corollaries \[uniquenesssSRbis\] and \[ExistenceComplementedSRbis\] are the analogous ones to Corollaries \[uniquenesslF\] and \[ExistenceComplementedlF\] in the preceding section, hence they need no further explanation.
\[uniquenesssSRbis\] Consider $F=F^{2,1}$ as in and $G$ as in . Then both spaces $\ell_F$ anf $\ell_G$ have a unique greedy basis.
\[ExistenceComplementedSRbis\] Consider $F=F^{2,1}$ as in and $G$ as in . A complemented subspace of $\ell_F$ (respectively, $\ell_G$) has a greedy basis if and only if it is isomorphic either to $\ell_2$ or to $\ell_F$ (respectively, $\ell_G$).
The last result of this section would be in correspondence with Theorem \[ComplementedNotGreedy\] in Section \[Orlicz\].
\[ComplementedNotGreedySR\] Consider $F=F^{2,1}$ as in and $G$ as in . Then both spaces $\ell_F$ and $\ell_G$ have a complemented subspace with an unconditional basis but with no greedy basis.
Appealing again to [@DKKT2003]\*[Theorem 5.1]{} it suffices to prove the result for $\ell_F$. As in the proof of Theorem \[ComplementedNotGreedy\], pick a sequence $(s_n)_{n=1}^\infty$ convergent to zero such that the unit vector basis of ${\mathit{h}}_F[s_n]$ (which is isomorphic to a complemented subspace of $\ell_F$) is not greedy. Since $\ell_2$ has a unique unconditional basis, ${\mathit{h}}_F[s_n]\not\approx \ell_2$. Suppose that ${\mathit{h}}_F[s_n]$ has a greedy basis ${\ensuremath{\mathcal{B}}}$. Then, by Theorem \[MainSuperReflexive\], ${\ensuremath{\mathcal{B}}}$ must be equivalent to the unit vector basis $ \ell_F$.
By [@CasazzaKalton1998]\*[Theorem 6.6 (1)]{} there is a sequence of finite dimensional Banach spaces $(V_n)_{n=1}^\infty$ such that ${\mathit{h}}_F[s_n]\approx \ell_{2}(V_n)$. Since ${\ensuremath{\mathcal{B}}}$ is weakly null and symmetric, appealing to the Bessaga-Pe[ł]{}czy[ń]{}ski selection principle, we claim that it is equivalent to the unit vector basis of $\ell_2$. This absurdity proves the result.
Uniqueness of greedy basis in Marcinkiewicz spaces {#Lorentz}
==================================================
Throughout this section a *weight* will be a sequence of positive numbers ${\ensuremath{\textit{\textit{w}}}}= (w_n)_{n=1}^\infty$ whose *primitive weight* ${\ensuremath{\textit{s}}}=(s_n)_{n=1}^\infty$ given by $s_n = \sum_{i=1}^n w_i $, verifies the doubling condition $s_{2n} \lesssim s_n$ for all $n$. In case that ${\ensuremath{\textit{\textit{w}}}}$ decreases to zero and ${\ensuremath{\textit{s}}}$ increases to infinity, we consider the *Marcinkiewicz sequence space* $m_{\ensuremath{\textit{s}}}$, consisting of all sequences $(a_n)_{n=1}^\infty$ in $c_0$ for which the following norm is finite: $$\Vert (a_n)_{n=1}^\infty \Vert_{m_{\ensuremath{\textit{s}}}} = \sup_{n \in {\ensuremath{\mathbb{N}}}} \frac{1}{s_n} \sum_{i=1}^n a_i^*,$$ where $(a^*_n)_{n=1}^\infty$ is the decreasing rearrangement of $(|a_n|)_{n=1}^\infty$. The *separable part* of $m_{\ensuremath{\textit{s}}}$, denoted $m^0_{\ensuremath{\textit{s}}}$, is the completion of $c_{00}$ in the space $m_{\ensuremath{\textit{s}}}$.
The unit vector basis is a symmetric basis both for $m^0_{\ensuremath{\textit{s}}}$ and, under the natural duality, for its dual space, the Lorentz sequence space $d_{{\ensuremath{\textit{\textit{w}}}},1}$, whose norm is given by $$\Vert (a_n)_{n=1}^\infty \Vert_{{\ensuremath{\textit{\textit{w}}}},1} = \sum_{n=1}^\infty a_n^* w_n.$$
We say that a weight ${\ensuremath{\textit{\textit{w}}}}$ is *regular* if $$ \frac{s_n}{n} \lesssim w_n, \quad \forall\;n\in{\ensuremath{\mathbb{N}}}.$$ The regularity of the weight implies the following equivalence of quasi-norms $$\label{eq: weaknorms}
\sup_{n\in{\ensuremath{\mathbb{N}}}} \frac{a^*_n}{w_n} \lesssim \left\|\sum_{n=1}^\infty a_n {\ensuremath{\mathbf{e}}}_n\right\|_{m_{\ensuremath{\textit{s}}}} \le \sup_{n \in{\ensuremath{\mathbb{N}}}} \frac{a^*_n}{w_n}$$ for all $(a_n)_{n=1}^\infty\in c_{00}$. (Note that the right hand-side inequality in does not require regularity.) The inequalities in give us an identification between Marcinkiewicz sequence spaces and weak Lorentz spaces. In general, given a weight ${\ensuremath{\textit{\textit{v}}}}=(v_n)_{n=1}^\infty$, the space $d_{{\ensuremath{\textit{\textit{v}}}},\infty}$ consists of all sequences $ (a_n)_{n=1}^\infty\in c_0$ such that the quasi-norm $$\Vert (a_n)_{n=1}^\infty \Vert_{{\ensuremath{\textit{\textit{v}}}},\infty} = \sup_{n\in{\ensuremath{\mathbb{N}}}} \left( \sum_{i=1}^n v_i \right) a_n^*$$ is finite. So, if ${\ensuremath{\textit{\textit{w}}}}$ is a regular weight and ${\ensuremath{\textit{\textit{v}}}}=(1/w_n - 1/w_{n-1})_{n=1}^\infty$, then $m_{\ensuremath{\textit{s}}}= d_{{\ensuremath{\textit{\textit{v}}}},\infty}$. In the particular case that for some $1<p<\infty$ the weight ${\ensuremath{\textit{\textit{w}}}}$ is given by $w_n:= n^{1/p} - (n-1)^{1/p}$, so that ${\ensuremath{\textit{s}}}=(n^{1/p})_{n=1}^\infty$, then ${\ensuremath{\textit{\textit{w}}}}$ is regular and the Marcinkiewicz space $m_{\ensuremath{\textit{s}}}$ coincides with the classical weak-$\ell_q$ space $\ell_{q,\infty}$, whose natural quasi-norm is given by $$\| (a_n)_{n=1}^\infty\|_{q, \infty} = \sup_{n\in{\ensuremath{\mathbb{N}}}} a_n^*\, n^{1/q},$$ where $q = (p-1)/p$.
We will use the following result, which illustrates the connection between Lorentz sequence spaces and greedy-like bases. We would like to remark that it remains valid for quasi-greedy bases.
\[DemocracyEmbedding\] Let ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$ be a greedy basis is a Banach space ${\ensuremath{\mathbb{X}}}$. Let ${\ensuremath{\textit{\textit{w}}}}=(w_n)_{n=1}^\infty$ be a weight and denote by $(s_n)_{n=1}^\infty$ its primitive weight.
- ${\ensuremath{\mathbb{X}}}$ embeds in $d_{{\ensuremath{\textit{\textit{w}}}},\infty}$ via ${\ensuremath{\mathcal{B}}}$, i.e., $$\Vert (a_k)_{k=1}^\infty \Vert_{{\ensuremath{\textit{\textit{w}}}},\infty} \lesssim \left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k\right\Vert,\quad \forall\; (a_{k})_{n=1}^\infty \in c_{00},$$ if and only if $$s_n \lesssim \varphi_l[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](n), \quad \forall\,n\in{\ensuremath{\mathbb{N}}}.$$
- $d_{{\ensuremath{\textit{\textit{w}}}},1}$ embeds in ${\ensuremath{\mathbb{X}}}$ via the ${\ensuremath{\mathcal{B}}}$, i.e., $$\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_n\right\Vert\lesssim \left\Vert (a_k)_{k=1}^\infty \right\Vert_{{\ensuremath{\textit{\textit{w}}}},1}\,\quad\forall\;(a_k)_{k=1}^\infty\in c_{00},$$ if and only if $$\varphi_u[{\ensuremath{\mathcal{B}}},{\ensuremath{\mathbb{X}}}](n) \lesssim s_n,\quad\forall \;n\in{\ensuremath{\mathbb{N}}}.$$
Part (b) requires a little bit more work than part (a) but both parts can be obtained rewriting carefully the proof of [@AA2015]\*[Theorem2.1]{}.
The study of symmetric bases in Lorentz sequence spaces leads to consider basic sequences whose terms are equidistributed disjointly supported sequences. Two sequences $(a_n)_{n=1}^\infty$ and $(b_n)_{n=1}^\infty$ of real numbers are said to be *equidistributed* if $(a_n^{\ast})_{n=1}^{\infty}=(b_n^*)_{n=1}^{\infty}$. In this direction, Altshuler el al. proved in [@ACL1973] the following theorem. Recall that a weight $(s_n)_{n=1}^\infty$ is called *submultiplicative* if $$ \sup_{n,k>0} \frac{s_{nk}}{s_n s_k} < \infty.$$ Note that $(n^{1/p})_{n=1}^\infty$ is a submultiplicative weight for each $1<p<\infty$.
\[Altshuler\]Let ${\ensuremath{\textit{\textit{w}}}}$ be a weight decreasing to zero such that its primitive weight is submultiplicative. Then any equidistributed disjointly supported basic sequence in $d_{{\ensuremath{\textit{\textit{w}}}},1}$ is equivalent to the unit vector basis of $ d_{{\ensuremath{\textit{\textit{w}}}},1}$.
Next we prove a lemma stating that equidistributed disjointly supported basic sequences are [*sufficiently far*]{} from being equivalent to the unit vector basis of $\ell_1$, even without imposing the submultiplicative condition on the weight.
\[newlemma\] Let ${\ensuremath{\textit{\textit{w}}}}$ be a weight decreasing to zero. Let ${\ensuremath{\mathcal{B}}}$ be an equidistributed disjointly supported basic sequence in $d_{{\ensuremath{\textit{\textit{w}}}},1}$. Then there is a weight ${\ensuremath{\textit{\textit{v}}}}$ decreasing to zero such that $d_{{\ensuremath{\textit{\textit{v}}}},1}$ embeds in $d_{{\ensuremath{\textit{\textit{w}}}},1}$ via ${\ensuremath{\mathcal{B}}}$.
Let ${\ensuremath{\mathbf{y}}}=(a_n)_{n=1}^\infty$ be a decreasing sequence of non negative numbers such that each element in ${\ensuremath{\mathcal{B}}}$ is equidistributed with ${\ensuremath{\mathbf{y}}}$. Then, using Abel’s summation formula, for all $N$ we have $$\varphi_l[{\ensuremath{\mathcal{B}}},d_{{\ensuremath{\textit{\textit{w}}}},1}](N)=\sum_{j=1}^\infty a_j \sum_{n=(j-1)N+1}^{jN} w_n=\sum_{j=1}^\infty (a_j-a_{j+1}) \sum_{n=1}^{jN} w_n.$$ Therefore $$v_N:=\varphi_l[{\ensuremath{\mathcal{B}}},d_{{\ensuremath{\textit{\textit{w}}}},1}](N)-\varphi_l[{\ensuremath{\mathcal{B}}},d_{{\ensuremath{\textit{\textit{w}}}},1}](N-1)=\sum_{j=1}^\infty (a_j-a_{j+1})
\sum_{n=1+j(N-1)}^{jN} w_n.$$ Notice that $( \sum_{n=1+j(N-1)}^{jN} w_n)_{N=1}^\infty$ decreases to zero. Then, by the dominated convergence theorem, $(v_N)_{N=1}^\infty$ decreases to zero. The proof is over invoking Lemma \[DemocracyEmbedding\](b).
\[lem: uncbasicseqfact\] Let ${\ensuremath{\textit{\textit{w}}}}$ be a weight decreasing to zero such that its primitive weight ${\ensuremath{\textit{s}}}$ increases to infinity. Let ${\ensuremath{\mathcal{B}}}$ be a semi-normalized unconditional basic sequence in $d_{{\ensuremath{\textit{\textit{w}}}},1}$. Then
- ${\ensuremath{\mathcal{B}}}$ has a subsequence which is equivalent either to the unit vector basis of $\ell_1$ or to a disjointly supported equidistributed sequence.
- If ${\ensuremath{\textit{s}}}$ is submultiplicative then ${\ensuremath{\mathcal{B}}}$ has a subsequence which is equivalent either to the unit vector basis of $\ell_1$ or to the unit vector basis $d_{{\ensuremath{\textit{\textit{w}}}},1}$.
Taking into account Theorem \[Altshuler\], we need only prove (a). If ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$ is not weakly null then, by the unconditionality of ${\ensuremath{\mathcal{B}}}$ we get that it has a subsequence equivalent to the unit vector basis of $\ell_1$. So we may assume that ${\ensuremath{\mathcal{B}}}$ is weakly null. Appealing to the Bessaga-Pe[ł]{}czy[ń]{}ski selection principle, we may assume that ${\ensuremath{\mathcal{B}}}$ is a block basis of the unit vector basis. We can suppose also the that there is an infinite subset $B$ of ${\ensuremath{\mathbb{N}}}$ such that $x_{k,n}=0$ for all $k\in{\ensuremath{\mathbb{N}}}$ and $n\in B$. Consider $(B_k)_{k=1}^\infty$ a partition of ${\ensuremath{\mathbb{N}}}$ such that ${\mathop{\mathrm{supp}}\nolimits}{\ensuremath{\mathbf{x}}}_k \subseteq B_k$ and $B_k \setminus {\mathop{\mathrm{supp}}\nolimits}{\ensuremath{\mathbf{x}}}_k$ is infinite. Denote, for each $k\in{\ensuremath{\mathbb{N}}}$, ${\ensuremath{\mathbf{x}}}_k=(x_{k,n})_{n=1}^\infty$. Let $\pi_k\colon {\ensuremath{\mathbb{N}}}\to S_k$ bijective and such that the absolute value of ${\ensuremath{\mathbf{x}}}_k':=(x_{k,\pi_k(j)})_{j=1}^\infty$ is a decreasing sequence. Notice that $$|x_{k,\pi_k(j)}| \le \frac{ \Vert {\ensuremath{\mathbf{x}}}_k \Vert_{{\ensuremath{\textit{\textit{w}}}},1} }{s_j}, \quad j,k\in{\ensuremath{\mathbb{N}}}.$$ Therefore, regarding $({\ensuremath{\mathbf{x}}}_k')_{k=1}^\infty$ as a sequence of functions defined in the compact space ${\ensuremath{\mathbb{N}}}\cup\{\infty\}$, $({\ensuremath{\mathbf{x}}}_k')_{k=1}^\infty$ is an equicontinuous and uniformly bounded sequence. By the Arzelà-Ascoli theorem, passing again to a subsequence, we can suppose that there is ${\ensuremath{\mathbf{y}}}=(y_j)_{j=1}^\infty \in c_0$ such that $\lim_k {\ensuremath{\mathbf{x}}}_k'={\ensuremath{\mathbf{y}}}$ uniformly. By Fatou’s Lemma, ${\ensuremath{\mathbf{y}}}\in d_{{\ensuremath{\textit{\textit{w}}}},1}$.
For each $k\in{\ensuremath{\mathbb{N}}}$ let ${\ensuremath{\mathbf{y}}}_k=(y_{k,n})_{n=1}^\infty$ be the sequence given by $y_{k,\pi_k(j)}=y_j$ for all $j\in{\ensuremath{\mathbb{N}}}$, and $y_{k,n}=0$ if $n\notin B_k$. Let ${\ensuremath{\mathbf{z}}}_k={\ensuremath{\mathbf{x}}}_k -{\ensuremath{\mathbf{y}}}_k$. We have:
- $({\ensuremath{\mathbf{y}}}_k)_{k=1}^\infty$ is a disjointly supported sequence in $d_{{\ensuremath{\textit{\textit{w}}}},1}$,
- $\lim_k \Vert {\ensuremath{\mathbf{z}}}_k \Vert_\infty=\lim_k \Vert {\ensuremath{\mathbf{x}}}_k'-{\ensuremath{\mathbf{y}}}\Vert_\infty=0$, and
- ${\ensuremath{\mathbf{x}}}_k={\ensuremath{\mathbf{y}}}_k+{\ensuremath{\mathbf{z}}}_k$.
If $\liminf_k \|{\ensuremath{\mathbf{z}}}_k\|_{{\ensuremath{\textit{\textit{w}}}},1} = 0$, then ${\ensuremath{\mathcal{B}}}$ has a subsequence which is an arbitrarily small perturbation of a subsequence of $({\ensuremath{\mathbf{y}}}_k)_{k=1}^\infty$ and hence is equivalent to $({\ensuremath{\mathbf{y}}}_k)_{k=1}^\infty$. If $\liminf_k \|{\ensuremath{\mathbf{z}}}_k\|_{{\ensuremath{\textit{\textit{w}}}},1} > 0$ then, appealing to [@LinTza1977 Prop. 4.e.3] and passing to a subsequence, we can suppose that $({\ensuremath{\mathbf{z}}}_k)_{k=1}^\infty$ is equivalent to the unit vector basis of $\ell_1$. Therefore, $$A\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{e}}}_k\right\Vert_{1} \le \left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{z}}}_k\right\Vert_{{\ensuremath{\textit{\textit{w}}}},1}$$ for some constant $A$ and for all sequences $(a_k)_{k=1}^\infty$ in $c_{00}$. Furthermore, by Lemma \[newlemma\], there is a weight ${\ensuremath{\textit{\textit{v}}}}=(v_n)_{n=1}^\infty$ decreasing to zero and a constant $B$ such that for all $(a_k)_{k=1}^\infty \in c_{00}$, $$\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{y}}}_k\right\Vert_{{\ensuremath{\textit{\textit{w}}}},1} \le B \left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{e}}}_k\right\Vert_{{\ensuremath{\textit{\textit{v}}}},1}.$$ Since $({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$ is semi-normalized there is a constant $C$ such that for all $j\in{\ensuremath{\mathbb{N}}}$ and all $(a_k)_{k=1}^\infty \in c_{00}$, $$|a_j| \le C \left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k\right\Vert_{{\ensuremath{\textit{\textit{w}}}},1}.$$ Combining, we obtain that $({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$ is equivalent to the unit vector basis of $\ell_1$.
Indeed, let $N\in{\ensuremath{\mathbb{N}}}$ be minimal with the property that $D:=A -B v_N>0$. Let $(a_k)_{k=1}^\infty \in c_{00}$. Denoting by $(a_k^*)_{k=1}^\infty$ the decreasing rearrangement of $(|a_k|)_{k=1}^\infty$, $$\begin{aligned}
\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k\right\Vert_{{\ensuremath{\textit{\textit{w}}}},1}
&\ge A\sum_{k=1}^\infty a_k^*
- B \sum_{k=1}^\infty a_k^* v_k \\
&\ge D \sum_{k=N}^\infty a_k^*+ \sum_{k=1}^{N-1} (A- B v_k)a_k^*\\
&= D\sum_{k=1}^\infty |a_k|- B \sum_{k=1}^{N-1} (v_k-v_N)a_k^*\\
& \ge D \sum_{k=1}^\infty |a_k|- BC\sum_{k=1}^{N-1}(v_k-v_N)\left\Vert \sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k\right\Vert_{{\ensuremath{\textit{\textit{w}}}},1}.\end{aligned}$$ This yields the desired result.
\[prop: uncbasisfact\] Let ${\ensuremath{\textit{\textit{w}}}}$ be a weight decreasing to zero such that its primitive weight ${\ensuremath{\textit{s}}}=(s_n)_{n=1}^\infty$ increases to infinity. Let ${\ensuremath{\mathcal{B}}}$ be a semi-normalized unconditional basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$ such that $$\label{assumptionondemocracyf}
\varphi_u[{\ensuremath{\mathcal{B}}},d_{{\ensuremath{\textit{\textit{w}}}},1}](n)\lesssim s_n,\quad \forall n\in {\ensuremath{\mathbb{N}}}.$$ Then ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$.
Let ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_k)_{k=1}^\infty$ and ${\ensuremath{\mathbf{x}}}_k =(x_{k,n})_{n=1}^\infty$. Then $$\delta:=\inf_{k\in{\ensuremath{\mathbb{N}}}}
\| {\ensuremath{\mathbf{x}}}_k \|_\infty >0.$$ Indeed, if not, by [@LinTza1977 Prop. 4.e.3], ${\ensuremath{\mathcal{B}}}$ would have a subsequence equivalent to the unit vector basis of $\ell_1$, contradicting the assumption that holds.
Next, we claim that there exists $N \in {\ensuremath{\mathbb{N}}}$ such that, for each $n\in{\ensuremath{\mathbb{N}}}$, $|x_{k,n}| \ge \delta$ for at most $N$ values of $k$. Suppose that this is not the case. Then, for every $N \ge 1$ there exist $n := n(N)$ and $k_1<k_2<\dots<k_N$ such that $|x_{k_j,n}| \ge \delta$ for $1 \le j \le N$. Hence there exists a choice of signs $\varepsilon_j = \pm 1$ ($1 \le j \le N$) such that $$\left\|\sum_{j=1}^N \varepsilon_j {\ensuremath{\mathbf{x}}}_{k_j}\right\|_{{\ensuremath{\textit{\textit{w}}}},1} \ge \sum_{j=1}^N |x_{k_j,n}| \ge N\delta,$$ contradicting again the assumption on $\varphi_u[{\ensuremath{\mathcal{B}}},d_{{\ensuremath{\textit{\textit{w}}}},1}]$.
Hence, there is a partition $(B_j)_{j=1}^N$ of ${\ensuremath{\mathbb{N}}}$ such that for each $k \in {\ensuremath{\mathbb{N}}}$ there exists $n=n(k)$ such that $|x_{k,n(k)}| \ge \delta$ and the map $k \mapsto n(k)$ is one-one on each $B_j$.
Now we estimate from below the norm of any element $\sum_{n=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \in d_{{\ensuremath{\textit{\textit{w}}}},1}$. By the unconditionality of ${\ensuremath{\mathcal{B}}}$ (see [@LT1979]\*[Theorem 1.d.6]{}), $$\left \|\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \right\|_{{\ensuremath{\textit{\textit{w}}}},1} \approx \sum_{j=1}^N \left\|\sum_{k\in B_j}^\infty a_k {\ensuremath{\mathbf{x}}}_k\right\|_{{\ensuremath{\textit{\textit{w}}}},1}
\approx \sum_{j=1}^N \left\| \left( \sum_{k \in B_j} |a_k|^2 |{\ensuremath{\mathbf{x}}}_k|^2\right)^{1/2}\right\|_{{\ensuremath{\textit{\textit{w}}}},1}.$$ Then, using the estimate $|x_{k,n(k)}| \ge \delta$ and the symmetry of the unit vector basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$, $$\begin{aligned}
\left \|\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \right\|_{{\ensuremath{\textit{\textit{w}}}},1}
& \gtrsim \sum_{j=1}^N
\left\| \sum_{k \in B_j} |a_k| {\ensuremath{\mathbf{e}}}_{n(k)}\right\|_{{\ensuremath{\textit{\textit{w}}}},1} \\
&= \sum_{j=1}^N\left\| \sum_{k \in B_j} a_k {\ensuremath{\mathbf{e}}}_k\right\|_{{\ensuremath{\textit{\textit{w}}}},1}\\
& \approx \left\|\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{e}}}_k\right\|_{{\ensuremath{\textit{\textit{w}}}},1}.\end{aligned}$$ The upper estimate follows from Lemma \[DemocracyEmbedding\](b).
Suppose that ${\ensuremath{\textit{\textit{w}}}}$ is a regular weight decreasing to zero such that its primitive weight ${\ensuremath{\textit{s}}}$ is submultiplicative. Then $m^0_{\ensuremath{\textit{s}}}$ has a unique greedy basis.
Let ${\ensuremath{\mathcal{B}}}=({\ensuremath{\mathbf{x}}}_n)_{n=1}^\infty$ be a greedy basis for $m^0_{\ensuremath{\textit{s}}}$ and let ${\ensuremath{\mathcal{B}}}^*=({\ensuremath{\mathbf{x}}}_n^*)_{n=1}^\infty$ be its biorthogonal basic sequence in $d_{{\ensuremath{\textit{\textit{w}}}},1}$. Since $m^0_{\ensuremath{\textit{s}}}$ has a separable dual it follows from the unconditionality of the basis that ${\ensuremath{\mathcal{B}}}$ is shrinking, hence ${\ensuremath{\mathcal{B}}}^*$ is a (semi-normalized and unconditional) basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$.
Note that ${\ensuremath{\mathcal{B}}}^*$ does not contain any subsequence equivalent to the unit vector basis of $\ell_1$ since, otherwise, by duality, ${\ensuremath{\mathcal{B}}}$ would contain a subsequence equivalent to the unit vector basis of $c_0$. This in turn would imply that $\varphi_u[{\ensuremath{\mathcal{B}}},m_{\ensuremath{\textit{s}}}^0]$ is bounded, which would yield that ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $c_0$. We would infer that $m^0_s$ is isomorphic to $c_0$, which is clearly false since the unit vector basis of $m^0_s$ is not equivalent to the unit vector basis of $c_0$, and $c_0$ has a unique symmetric basis.
Therefore, by Lemma \[lem: uncbasicseqfact\](b), ${\ensuremath{\mathcal{B}}}^*$ has a subsequence equivalent to the unit vector basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$. By duality this implies that the corresponding subsequence of ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $m^0_{\ensuremath{\textit{s}}}$. Using the regularity of the weight, $$\varphi_l[{\ensuremath{\mathcal{B}}},m_{\ensuremath{\textit{s}}}^0](N)\approx
\varphi_u[{\ensuremath{\mathcal{B}}},m_{\ensuremath{\textit{s}}}^0](N)\approx \left\Vert \sum_{n=1}^N {\ensuremath{\mathbf{e}}}_n\right\Vert = \frac{N}{s_N} \gtrsim \frac{1}{w_N},$$ for all $N$.
Now, and Lemma \[DemocracyEmbedding\](a), allow us to estimate from below the norm of any element $\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \in m^0_s$. We have $$\left \|\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}_k \right \|_{m_{\ensuremath{\textit{s}}}}
\gtrsim \sup_{k\in{\ensuremath{\mathbb{N}}}} \frac{1}{w_k} a^*_k
\ge
\left\Vert (a_k)_{k=1}^\infty\right\Vert_{m_{\ensuremath{\textit{s}}}}.$$ By duality, for all $(a_n)_{n=1}^\infty \in c_{00}$ we have $$\left\|\sum_{k=1}^\infty a_k {\ensuremath{\mathbf{x}}}^*_k\right\|_{{\ensuremath{\textit{\textit{w}}}},1} \lesssim \|\ (a_k)_{k=1}^\infty \|_{{\ensuremath{\textit{\textit{w}}}},1}.$$ Appealing to Lemma \[DemocracyEmbedding\](b) we obtain that $$\varphi_u[{\ensuremath{\mathcal{B}}}^*, d_{{\ensuremath{\textit{\textit{w}}}},1}] (n)\lesssim s_n,\quad \forall n,$$ which implies by Proposition \[prop: uncbasisfact\] that ${\ensuremath{\mathcal{B}}}^* $ is equivalent to the unit vector basis of $d_{{\ensuremath{\textit{\textit{w}}}},1}$. Hence, by duality, ${\ensuremath{\mathcal{B}}}$ is equivalent to the unit vector basis of $m^0_{\ensuremath{\textit{s}}}$.
The separable part of weak-$\ell_p$ has a unique greedy basis.
Acknowledgements {#acknowledgements .unnumbered}
----------------
F. Albiac and J. L. Ansorena acknowledge the support of Spanish Research Grant MTM2014-53009-P, [*Análisis Vectorial, Multilineal, y Aplicaciones*]{}. The first-named author was also partially supported by Spanish Research Grant MTM2012-31286, [*Operators, lattices, and structure of Banach spaces*]{}. S. J. Dilworth was partially supported by the National Science Foundation under Grant Number DMS-1361461. D. Kutzarova has been partially supported by the Bulgarian National Scientific Fund under Grant DFNI-I02/10.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this note, it is proved the existence of an infinitely generated multiplicative group consisting of entire functions that are, except for the constant function $1$, hypercyclic with respect to the convolution operator associated to a given entire function of subexponential type. A certain stability under multiplication is also shown for compositional hypercyclicity on complex domains.'
address:
- 'Departamento de Análisis Matemático, Facultad de Matemáticas, Universidad de Sevilla, Avenida Reina Mercedes, 41080 Sevilla, Spain.'
- 'Instituto Universitario de Matemática Pura y Aplicada, Universitat Politècnica de València, 46022 València, Spain.'
- 'Department of Mathematics and Applied Mathematics, University of Crete, Voutes Campus, 70013 Heraklion, Crete, Greece.'
- 'IMI and Departamento de Análisis Matemático,Facultad de Ciencias Matemáticas, Plaza de Ciencias 3, Universidad Complutense de Madrid,28040 Madrid, Spain.'
author:
- 'Luis Bernal-González'
- 'J. Alberto Conejero'
- George Costakis
- 'Juan B. Seoane-Sepúlveda'
title: Multiplicative structures of hypercyclic functions for convolution operators
---
[^1]
Introduction
============
Assume that $X$ is a (Hausdorff) topological vector space over the real line ${\mathbb{R}}$ or the complex plane ${\mathbb{C}}$, and consider the vector space $L(X)$ of all operators on $X$, that is, the family of all continuous linear self-mappings $T:X \to X$. An operator $T \in L(X)$ is said to be [*hypercyclic*]{} provided that it admits a dense orbit, that is, provided that there is some vector $x_0 \in X$ (called hypercyclic for $T$) such that the orbit $\{T^n x_0: \, n \in {\mathbb{N}}\}$ of $x_0$ under $T$ is dense in $X$ (${\mathbb{N}}:= \{1,2, \dots \}$). The set of hypercyclic vectors for $T$ will be denoted by $HC(T)$. In this paper, we are concerned with the [*size*]{} of $HC(T)$, mainly from an algebraic point of view and for certain differential operators. For background on hypercyclic operators we refer the reader to the excellent books [@bayartmatheron2009; @grosseperis2011]. An account of concepts and results about algebraic structures inside nonlinear sets can be found in [@arongurariyseoane2005; @aronbernalpellegrinoseoane; @bernalpellegrinoseoane2014linear; @TAMS-Enflo; @Seo1; @Seo2; @Seo3].
.15cm
It is well known that if $X$ is an F-space (i.e., complete and metrizable) and $T$ is a hypercyclic operator then the set $HC(T)$ is [*residual,*]{} that is, it contains a dense $G_\delta$ subset of the (Baire) space $X$; we can say that $HC(T)$ is topologically large. Furthermore, for any topological vector space and any hypercyclic $T \in L(X)$, the family $HC(T)$ is algebraically large; specifically, it contains, except for $0$, a [*dense*]{} (even $T$-invariant) vector subspace of $X$ (see [@wengenroth2003]). Starting from [@montes1996], a number of criteria have been established for an operator $T \in L(X)$ to support a [*closed infinite dimensional*]{} vector subspace all of whose nonzero vectors are $T$-hypercyclic on an F-space $X$; however, not all hypercyclic operators support such a subspace (see [@bayartmatheron2009 Chapter 8] and [@grosseperis2011 Chapter 10]).
.15cm
By a domain in ${\mathbb{C}}$ we mean a nonempty connected open subset $G \subset {\mathbb{C}}$. It is well known that the space $\mathcal H (G)$ of holomorphic functions $G \to {\mathbb{C}}$ becomes an F-space when endowed with the topology of uniform convergence on compact subsets of $G$. We are mainly interested in operators defined on the space ${\mathcal{H}(\mathbb{C})}$ of entire functions ${\mathbb{C}}\to {\mathbb{C}}$. An operator $T \in L({\mathcal{H}(\mathbb{C})})$ is said to be a [*convolution operator*]{} provided that it commutes with translations, that is, $$T \circ \tau_a = \tau_a \circ T \hbox{ \ for all \ } a \in {\mathbb{C}},$$ where $\tau_a f := f(\cdot + a)$. Then $T \in L({\mathcal{H}(\mathbb{C})})$ happens to be a convolution operator if and only if $T$ is an infinite order linear differential operator with constant coefficients $T = \Phi (D)$, where $\Phi$ is an entire function with exponential type, that is, there exist constants $A,B \in (0,+\infty )$ such that $|\Phi (z)| \le A e^{B|z|}$ for all $z \in {\mathbb{C}}$. For such a function $\Phi (z) = \sum_{n \ge 0} a_nz^n$ and $f \in {\mathcal{H}(\mathbb{C})}$, we have $$\Phi (D) f = \sum_{n = 0}^\infty a_n f^{(n)} .$$ Godefroy and Shapiro [@godefroyshapiro1991] proved that every non-scalar convolution operator is hypercyclic, so covering the classical Birkhoff and MacLane results on hypercyclicity of the translation operator $\tau_a$ (take $\Phi (z) = e^{az}$, $a \ne 0$) and of the derivative operator $D: f \mapsto f'$ (take $\Phi (z) = z$), respectively. It has been proved that, for every non-scalar convolution operator $T$, the set $HC(T)$ contains, except for $0$, a closed infinite dimensional vector subspace of ${\mathcal{H}(\mathbb{C})}$; see [@menet2014; @petersson2006; @shkarin2010] and also [@aronbernalpellegrinoseoane Section 4.5] and [@grosseperis2011 Section 10.1].
.15cm
In view of the preceding paragraph, we can say that hypercyclic vectors of convolution operators are rather stable under summation and scaling. However, stability [*under multiplication*]{} does not seem to be so clear. For instance, for every $a \ne 0$, every $f \in HC(\tau_a)$ and every $n \in {\mathbb{N}}$ with $n \ge 2$ we have that $f^n \not\in HC(\tau_a)$ (see [@aronconejeroperisseoane2007 Cor. 2.4]). As a positive result, it was proved in [@aronconejeroperisseoane2007 Th. 2.3] the existence of a function $f \in {\mathcal{H}(\mathbb{C})}$ –in fact, of a residual subset of them– such that $f^n \in HC(D)$ for all $n\in{\mathbb{N}}$, see also [@aronconejeroperisseoane2007sums]. This result was extended by Bayart and Matheron who proved that there is even a residual set of functions in ${\mathcal{H}(\mathbb{C})}$ generating a hypercyclic algebra for the derivative operator, that is every non-null function in one of these algebras is hypercyclic for the operator $D$ [@bayartmatheron2009 Th. 8.26]. Recall that the algebra generated by a function $f$ is nothing but the set $\{P \circ f: \, P$ polynomial, $P(0) = 0\}$. The existence of (one-generated) algebras of hypercyclic functions for $D$ was also independently proved by Shkarin in [@shkarin2010] and extended by Bès [*et al.*]{} in [@besconejeropapathaniasiou2016] to operators of the form $P(D)$, where $P$ is a nonconstant polynomial with $P(0) = 0$. The existence of hypercyclic algebras for many convolution operators not induced by polynomials, such as $\cos(D)$, $De^D$, or $e^D-aI$, where $0<a\le 1$ was recently shown in [@besconejeropapathaniasiou2018].
.15cm
So far, the existence of multiply generated algebras of hypercyclic entire functions is unknown. The lack of knowledge about existence of families of hypercyclic functions that are stable under multiplication seems to be one of the main obstacles. Our goal in this paper is to contribute to fill in this gap. Specifically, given a nonconstant entire function $\Phi$ of subexponential type, we will construct in Section 3 an infinitely generated multiplicative group of entire functions all of whose members, except for the constant function ${\bf 1} (z) = 1$, are hypercyclic with respect to the operator $\Phi (D)$. A certain multiplicative stability is also shown for hypercyclicity with respect to composition operators. Section 2 is devoted to provide the necessary background and auxiliary results.
Non-vanishing hypercyclic entire functions
==========================================
Recall that the exponential type of a $\Phi \in {\mathcal{H}(\mathbb{C})}$ is defined as, see [@boas1954], $$\tau (\Phi ) = \limsup_{r \to \infty} {\log \max \{|\Phi (z)|: \, |z|=r\} \over r}.$$ Hence, $f$ has subexponential type if and only if, given ${\varepsilon}> 0$, there exists $A = A({\varepsilon}) \in (0,+\infty )$ such that $|f(z)| \le A \, e^{{\varepsilon}|z|}$ for all $z \in {\mathbb{C}}$. In fact, if $\Phi$ has subexponential type then $\Phi (D)$ is a well defined operator on $\mathcal{H} (G)$ for any domain $G \subset {\mathbb{C}}$ (see for instance [@berensteingay1995] or [@bernal1999 Theorem 4]). It should be underlined that if $\Phi$ is of exponential type, then the operator $\Phi (D)$ does not act on proper domains in general: take, for instance, $G = {\mathbb{D}}$ and $\Phi (z) = e^z$.
.15cm
As we have mentioned in Section 1, there is no entire function $f$ such that $f^2$ is hypercyclic for the translation operator $\tau_a = e^{aD}$ $(a \ne 0)$. If we observe that $\Phi (z) := e^{az}$ has exponential type $\tau (\Phi ) = |a| > 0$, then we may wonder whether there are entire functions $f$ such that $f$ and $f^2$ are in $HC(\Phi (D))$ if $\Phi$ has [*subexponential type,*]{} i.e., if $\tau (\Phi ) = 0$. This property happens to be true, in a strong sense, in the special case of $\Phi = P$, a nonconstant polynomial with $P(0) = 0$ [@besconejeropapathaniasiou2016]. The rate of growth of a hypercyclic function for $D$ has been optimally estimated in [@grosse1990universal; @shkarin1993].
.15cm
It is well known (see, for instance, [@ahlfors1979]) that a function $f \in {\mathcal{H}(\mathbb{C})}$ is never zero if and only if there is $g \in {\mathcal{H}(\mathbb{C})}$ such that $f = e^g$. Consistently, we can denote the family of non-vanishing entire functions by $e^{{\mathcal{H}(\mathbb{C})}}$. This family forms, trivially, a group under the pointwise multiplication. Moreover, $e^{{\mathcal{H}(\mathbb{C})}}$ is –as it is easily proven– a topological group (for background on topological groups, see for instance [@markley2010]) under the topology of uniform convergence in compacta, that is, both mappings $$(f,g) \in e^{{\mathcal{H}(\mathbb{C})}} \times e^{{\mathcal{H}(\mathbb{C})}} \mapsto fg \in e^{{\mathcal{H}(\mathbb{C})}} \hbox{ \ and \ } f \in e^{{\mathcal{H}(\mathbb{C})}} \mapsto {1 \over f} \in e^{{\mathcal{H}(\mathbb{C})}}$$ are continuous, from which one can derive that each mapping $f \in e^{{\mathcal{H}(\mathbb{C})}} \mapsto f^m \in e^{{\mathcal{H}(\mathbb{C})}}$ ($m \in {\mathbb{Z}}=$ the set of integers) is a continuous homomorphism (recall that a mapping $\Phi : G_1 \to G_2$ between two groups $G_1,G_2$ is an homomorphism whenever $\Phi (ab) = \Phi(a) \Phi (b)$ for all $a,b \in G_1$). Note that this mapping is not bijective (except for $m = \pm 1$), but if we consider the topological subgroup $e^{{\mathcal{H}(\mathbb{C})}}_{+,0} := \{f \in e^{{\mathcal{H}(\mathbb{C})}}: \, f(0) > 0\}$ then each restriction $$\Phi_m: f \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0} \mapsto f^m \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0} \quad (m \in {\mathbb{Z}}\setminus \{0\})$$ is not only continuous (and homomorphic) but also bijective, because for any $g \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ there are exactly $m$ entire functions $f$ with $f^m = g$ (necessarily, such $f$’s are non-vanishing), but only one of them satisfies $f(0) > 0$. In fact, in the following two lemmas it is shown that the group $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ also enjoys a nice topological structure and that every $\Phi_m$ is an automorphism of it.
\[Lemma: topology of e+0\] The set $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is a $G_\delta$-subset of ${\mathcal{H}(\mathbb{C})}$. In particular, it is a separable completely metrizable space, as well as a Baire space.
The second part is a consequence of the first one. Indeed, Alexandroff’s theorem (see e.g. [@oxtoby1980 pp. 47–48]) asserts that a $G_\delta$-set is homeomorphic to a complete metrizable space; separability is inherited by any subspace of a separable metrizable space; finally, any completely metrizable space is a Baire space (see e.g. [@munkres2000]). As for the first part, simply put $e^{{\mathcal{H}(\mathbb{C})}}_{+,0} = \bigcap_{k=1}^\infty A_k$, where $A_k$ is defined as $$A_k := \left\{f\in{\mathcal{H}(\mathbb{C})}:\, \inf_{|z|\leq k}|f(z)|>0, \, |{\rm Im} f(0)| < {1 \over k} \text{ and }\, {\rm Re} f(0) > 0\right\}.$$ That each $A_k$ is open follows from the facts that $\{z\in\mathbb{C}\,:\,|z| \le k\}$ is compact and that both evaluation mappings $f \in {\mathcal{H}(\mathbb{C})}\mapsto {\rm Re} f(0) \in {\mathbb{R}}$, $f \in {\mathcal{H}(\mathbb{C})}\mapsto {\rm Im} f(0) \in {\mathbb{R}}$ are continuous.
\[Lemma: onto homeomorphisms\] For each $m \in {\mathbb{Z}}\setminus \{0\}$, the mapping $\Phi_m$ is an onto homeomorphism.
The unique property to be proved is that every $\Phi_m$ has inverse continuous. Since $\Phi_{-1}$ is an onto homeomorphism, it is enough to see that, given $m \in {\mathbb{N}}$ with $m \ge 2$, the mapping $\Psi = (\Phi_m)^{-1}: e^{{\mathcal{H}(\mathbb{C})}}_{+,0} \to e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is continuous. By Lemma \[Lemma: topology of e+0\], $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is a completely metrizable topological group. Note that $\Psi$ is a homomorphism from the group $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ into itself. According to the abstract closed graph theorem (see, e.g., [@christensen Theorem 5.2]), it is enough to show that $\Psi$ has closed graph. To this end, assume that $(f_k) \subset e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is a sequence such that there are $f,g \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ with $(f_k, \Psi (f_k)) \to (f,g)$ as $k \to \infty$ in the topology product. Then $f_k \to f$ and $\Psi (f_k) \to g$. The continuity of $\Phi_m$ yields $f_k = \Phi_m \Psi (f_k) \to \Phi_m (g)$, and the uniqueness of the limit implies $f = \Phi_m (g)$ or, that is the same, $g = \Psi (f)$, which proves that the graph of $\Psi$ is closed, as required.
Finally, we will use in the next section the following theorem, that is contained in [@bernal1997 Theorem 5]. This theorem is, in turn, an extension of a result about inherited $D$-hypercyclicity due to Herzog [@herzog1994].
\[Th: Bernal 1997\] If $\Phi$ is a nonconstant entire function of subexponential type, then the set $HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}$ is residual in $e^{{\mathcal{H}(\mathbb{C})}}$.
Additional statements on zero-free $D$-hypercyclic entire functions can be found in [@bernalbonillacostakis2012].
Groups of hypercyclic functions
===============================
Firstly, we show that the family of entire functions $f$ that are hypercyclic with respect to certain convolution operators (including $D$) satisfying $f(0) > 0$ is topologically large.
\[Prop: f(0) > 0 residual\] If $\Phi$ is a nonconstant entire function of subexponential type, then the set $HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is residual in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$.
According to Birkhoff transitivity theorem (see e.g. [@grosseperis2011]), if $T_n : X \to Y$ $(n \in {\mathbb{N}})$ is a sequence of continuous mappings between two Hausdorff topological spaces, with $X$ Baire and first-countable and $Y$ second-countable, then the set $\mathcal U = \mathcal U ((T_n))$ of points $x \in X$ whose orbit $\{T_n x : \, n \in {\mathbb{N}}\}$ is dense in $Y$ is a $G_\delta$ subset of $X$; consequently, $\mathcal U$ is residual as soon as it is dense. By Theorem \[Th: Bernal 1997\], the set $HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}$ is residual in $e^{{\mathcal{H}(\mathbb{C})}}$. It is easy to see that the mapping $$\Pi : f \in e^{{\mathcal{H}(\mathbb{C})}} \mapsto {|f(0)| \over f(0)} \cdot f \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$$ is continuous and onto. Then it takes dense sets into dense sets. In particular, the image $\Pi (HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}})$ is dense in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$. Since any nonzero scalar multiple of a hypercyclic vector is also hypercyclic, we obtain $\Pi (HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}) \subset HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$. But $\mathcal U := HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ equals the set of elements $f \in X := e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ whose orbit under the sequence $(T_n) := ((\Phi (D))^n|_{e^{{\mathcal{H}(\mathbb{C})}}_{+,0}})$ is dense in $Y := {\mathcal{H}(\mathbb{C})}$ (note that $X$ is Baire and first-countable by Lemma \[Lemma: topology of e+0\]). Therefore $\mathcal U$ is a $G_\delta$ dense subset of $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$, hence residual.
Our main result (Theorem \[Th: groups in HC(Phi(D))\]) will be deduced as a consequence of the following, more abstract, assertion.
\[Th: groups in residual subsets\] Let $X$ be a separable infinite dimensional F-space. Assume that $X$ is also a unitary commutative linear algebra, and that the corresponding multiplication law $(f,g) \in X \times X \mapsto f * g \in X$ is continuous. Suppose also that $Z$ is a subset of $X$ fulfilling the following conditions:
1. $Z$ is a $G_\delta$-subset of $X$.
2. $(Z,*)$ is a topological group.
3. The mappings $u \in Z \mapsto u^k \in Z$ $(k \in {\mathbb{N}})$ are onto homeomorphisms.
4. No nonempty relatively open subset of $Z$ is contained in a countable dimensional vector subspace of $X$.
Then for each residual subset $\mathcal R$ of $Z$ there exists a subset $\mathcal G \subset Z$ satisfying the following properties:
1. $\mathcal G$ is a subgroup of $Z$.
2. $\mathcal G$ is dense in $Z$.
3. $\mathcal R \supset \mathcal G \setminus \{{\bf e}\}$, where ${\bf e}$ denotes the unit element of $X$.
4. $\mathcal G$ is infinitely generated in a strong sense, namely, the algebra generated by $\mathcal G$ is infinitely generated.
By Alexandroff’s theorem and by the fact that $X$ is a separable F-space, we have (thanks to (i)) that $Z$ is a completely metrizable space that, in addition, is second-countable. Then $Z$ is a Baire space and there is a countable open basis $\{G_n\}_{n \ge 1}$ for the topology of $Z$. From (ii) and (iii) one takes out that every mapping $\Psi_k : u \in Z \mapsto u^k \in Z$ $(k \in {\mathbb{Z}}\setminus \{0\})$ is an onto homeomorphism. Moreover, it follows from (ii) that, for every $v \in {\mathbb{Z}}$, the multiplication mapping $M_{v} : u \in Z \mapsto v * u \in Z$ is also an onto homeomorphism.
Fix a residual subset $\mathcal R$ of $Z$. Then, for each $k \in {\mathbb{Z}}\setminus \{0\}$, the set $\Psi_k^{-1} (\mathcal R)$ is residual in $Z$. Then the countable intersection $$\mathcal R_0 := \bigcap_{k \in {\mathbb{Z}}\setminus \{0\}} \Psi_k^{-1} (\mathcal R)$$ is also residual –and, in particular, dense– in $Z$. Let us set $L_0 := {\rm span} \{{\bf e}\}$, which equals the algebra generated by ${\bf e}$. Since $L_0$ is a finite dimensional subspace, it is closed in $X$. Then $G_1 \setminus L_0$ is a nonempty (otherwise, $G_1 \subset L_0$, contradicting (iv)) open set in the topology of $Z$. Then there exists $v_1 \in G_1 \cap \mathcal R_0$.
Let us proceed by induction. Assume that, for some $n \in {\mathbb{N}}$, the vectors $v_1, \dots ,v_n$ have been selected. Let $L_n$ denote the linear algebra generated by $v_0,v_1, \dots ,v_n$, that is, $L_n = {\rm span} \{v_1^{j_n} * \cdots * v_n^{j_n}: \, (j_1, \dots ,j_n) \in {\mathbb{N}}_0^n\}$, where ${\mathbb{N}}_0 := {\mathbb{N}}\cup \{0\}$. Observe that $L_n$ is a countable dimensional vector subspace of $X$. Since $X$ is a Baire space, if follows that $L_n$ is an $F_\sigma$ set with empty interior. Now, we set $$\mathcal R_n := \bigcap_{k \in {\mathbb{Z}}\setminus \{0\}} \Psi_k^{-1} \bigg( \bigcap_{(k_1, \dots ,k_n) \in {\mathbb{Z}}^n}
M^{-1}_{v_1^{k_1} * \cdots * v_n^{k_n}} (\mathcal R) \bigg) .$$ Observe that the choice $(k_1, \dots ,k_n) = (0, \dots ,0)$ gives $\mathcal R_n \subset \mathcal R_0$. Since the $\Psi_k$’s as well as the $M^{-1}_{v_1^{k_1} * \cdots * v_n^{k_n}}$’s are onto homeomorphisms and countable intersections of residual sets are residual, we get that each $\mathcal R_n$ is residual in $Z$. Assume, by way of contradiction, that $G_{n+1} \setminus L_n$ is of first category (in the sense of Baire) in $Z$. Then $G_{n+1} = (G_{n+1} \setminus L_n) \cup (G_{n+1} \cap L_n)$ and, by the assumption (iv), $L_n \cap Z$ is a (relatively $F_\sigma$) set with empty interior in $Z$, so of first category in $Z$. This implies that its subset $G_{n+1} \cap L_n$ is also of first category in $Z$, hence $G_{n+1}$ is, which is absurd because $Z$ is a Baire space. Thus, $G_{n+1} \setminus L_n$ is of second category (in the sense of Baire) in $Z$. As $\mathcal R_n$ is residual in $Z$, we have $\mathcal R_n \cap (G_{n+1} \setminus L_n) \ne \varnothing$. Therefore, we can select $v_{n+1} \in \mathcal R_n \cap (G_{n+1} \setminus L_n)$ and the induction procedure is finished.
Define $\mathcal G$ as the group generated by $\{v_n\}_{n \in {\mathbb{N}}}$, that is $$\mathcal G = \big\{ v_1^{k_1} * \cdots * v_n^{k_n} : \, (k_1, \dots ,k_n) \in {\mathbb{Z}}^n, \, n \in {\mathbb{N}}\big\} .$$ Since each $v_n$ belongs to $Z$ and $v_n \in G_n$ $(n \in {\mathbb{N}})$, we obtain conclusions (a) and (b). Property (c) is derived from the fact that, by construction, each combination $v_1^{k_1} * \cdots * v_n^{k_n} \in \mathcal R$ as soon as $(k_1, \dots ,k_n) \in {\mathbb{Z}}^n \setminus \{(0, \dots ,0)\}$. Finally, property (d) is true because the algebra generated by $\mathcal G$ contains the algebra $\mathcal A := \bigcup_{n \ge 0} L_n$, and $\mathcal A$ is infinitely generated. Indeed, assume that $\mathcal A$ is generated by some finite subset $F \subset \mathcal A$. Since $L_n \subset L_{n+1}$ $(n \ge 0)$, there is $N \in {\mathbb{N}}$ such that $F \subset L_N$. Then $L_N$ generates $\mathcal A$. But $L_N$ is itself an algebra, so $\mathcal A = L_N$, which is absurd because $v_{N+1} \in \mathcal A \setminus L_N$. Hence $\mathcal A$ cannot be finitely generated, which concludes the proof.
\[Th: groups in HC(Phi(D))\] Assume that $\Phi$ is a nonconstant entire function of subexponential type. Consider the convolution operator $T = \Phi (D)$ associated to $\Phi$. Then there exists an infinitely generated multiplicative group $\mathcal G \subset \mathcal H ({\mathbb{C}})$ such that every member of $\mathcal G$ –except for [**1**]{}– is $\Phi (D)$-hypercyclic. Moreover, $\mathcal G$ is a dense subgroup of $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ and the algebra generated by $\mathcal G$ is infinitely generated.
It suffices to apply Theorem \[Th: groups in residual subsets\] with the following “characters”: Take $X := {\mathcal{H}(\mathbb{C})}$, $Z := e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$, $* :=$ the usual multiplication, ${\bf e} := {\bf 1}$ and $\mathcal R := HC(\Phi (D)) \cap e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$. Now, $Z$ is a topological group for which the mappings $f \mapsto f^k$ $(k \in {\mathbb{N}})$ are onto homeomorphisms by Lemma \[Lemma: onto homeomorphisms\], $Z$ is a $G_\delta$ subset of $X$ by Lemma \[Lemma: topology of e+0\], and $\mathcal R$ is residual in $Z$ by Proposition \[Prop: f(0) > 0 residual\].
Now, assume that $O$ is a nonempty open subset in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ and that, by way of contradiction, $Y$ is a countable dimensional space of ${\mathcal{H}(\mathbb{C})}$ with $O \subset Y$. Then ${\rm span} \{h_n\}_{n \ge 1} = Y$ for certain $h_n \in {\mathcal{H}(\mathbb{C})}$ $(n \ge 1)$. Setting $Y_n:= {\rm span} \{h_1, \dots ,h_n\}$ we get $Y = \bigcup_{n \ge 1} Y_n$. But each $Y_n$ is a finite dimensional subspace, hence $\sigma$-compact, and so $Y$ is. In other words, there are countably many compact sets $K_n \subset e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ $(n \in {\mathbb{N}})$ such that $Y = \bigcup_{n \ge 1} K_n$. Therefore each set $Z_n := e^{{\mathcal{H}(\mathbb{C})}}_{+,0} \cap K_n$ is compact (hence closed) in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ and $O \subset \bigcup_{n \ge 1} Z_n$. Since $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ is a completely metrizable space, Baire’s category theorem tells us that at least one $Z_m$ has nonempty interior in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$. Now, the group structure of $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ entails that the function [**1**]{} possesses a compact neighborhood $W$ in $e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$.
Now, consider the vector space $S := \{g \in {\mathcal{H}(\mathbb{C})}: \, {\rm Im} \, g(0) = 0\}$ endowed with the topology inherited from ${\mathcal{H}(\mathbb{C})}$. Trivially, $S$ is a topological group for the operation “+” and the same topology. For every $f \in e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ there is a unique $g \in S$ such that $f = e^g$. Hence the mapping $T : f = e^g \in (e^{{\mathcal{H}(\mathbb{C})}}_{+,0}, \cdot ) \mapsto g \in (S,+)$ is an algebraic group isomorphism. But it is in fact a topological isomorphism. Indeed, $T^{-1} : g \mapsto e^g$ is trivially continuous (any superposition mapping $g \in {\mathcal{H}(\mathbb{C})}\mapsto \varphi \circ g \in {\mathcal{H}(\mathbb{C})}$, with $\varphi \in {\mathcal{H}(\mathbb{C})}$, is continuous), and $T$ is continuous because it is continuous at the neutral element [**1**]{}. This means that if a sequence $(f_n = e^{g_n}) \subset e^{{\mathcal{H}(\mathbb{C})}}_{+,0}$ (with $g_n = u_n + iv_n \in S$, so that $v_n(0) = 0$ for all $n$) satisfies $f_n \to {\bf 1}$ then $g_n = u_n + iv_n = T(f_n) \to T({\bf 1}) = 0$ uniformly on compacta. This, in turn, follows by invoking the continuity of the principal branch $\log_p w := {\rm ln} \, |w| + i \, {\rm arg}_p \, w$ of the logarithm on ${\mathbb{C}}\setminus (-\infty ,0]$. Indeed, since $1 \in {\mathbb{C}}\setminus (-\infty ,0]$, for given $R > 0$ we have that $f_n (K) \subset {\mathbb{C}}\setminus (-\infty ,0]$ eventually, where $K = \{z: \, |z| \le R\}$. Then $\log_p (f_n) \to log_p 1 =0$ uniformly on $K$. But $\log_p (f_n (z)) = u_n (z) + i \arg_p (e^{iv_n (z)}) = u_n (z) + i(v_n(z) + 2 k_n \pi )$ for certain $k_n \in {\mathbb{Z}}$ not depending on $z$. Hence $u_n \to 0$ and $v_n + 2k_n\pi \to 0$ uniformly on $K$. In particular, $2k_n\pi = v_n(0) + 2k_n\pi \to 0$, so $k_n \to 0$ and, since $k_n \in {\mathbb{Z}}$, we get $k_n = 0$ eventually. This implies $v_n \to 0$ uniformly on $K$, and so does $g_n = u_n + iv_n$ as $n \to \infty$, as required. We have proved that $T$ is a homeomorphism. Thus $T(W)$ is a compact neighborhood of $0$ in $S$. But Riesz’s theorem (see, e.g., [@horvath1966 page 147, Theorem 3]) implies that ${\rm dim} \, (S) < \infty$, which is absurd because $S$ contains all functions $z^n$ $(n \in {\mathbb{N}})$. This is the desired contradiction.
Hence $O$ is not contained in any countable dimensional subspace of ${\mathcal{H}(\mathbb{C})}$. Consequently, conditions (i)–(iv) in Theorem \[Th: groups in residual subsets\] are fulfilled and the conclusion follows.
Another important class of operators is the one formed by the composition operators. If $G \subset {\mathbb{C}}$ is a domain and $\varphi : G \to G$ is a holomorphic self-map of $G$, the mapping $$C_\varphi : f \in \mathcal (G) \mapsto f \circ \varphi \in H(G)$$ is well defined, linear and continuous, and it is called the [*composition operator*]{} with symbol $\varphi$. By ${\rm Aut} (G)$ we denote the group of automorphisms of $G$, i.e., the family of all bijective holomorphic functions $G \to G$. The translations $\tau_a$ $(a \in {\mathbb{C}})$ form a special instance when $G = {\mathbb{C}}$ and $\varphi (z) := z + a$. We have already seen that, even for translations, the set of hypercyclic functions is not stable under products ($f^2$ is never $\tau_a$-hypercyclic, whatever $f$ is). Nevertheless, we obtain a reasonably degree of stability if the products do not allow repetition of factors. This can be done for many composition operators, as the following theorem –with which we conclude this paper– shows. Recall that a domain $G \subset {\mathbb{C}}$ is called simply connected if it lacks holes, that is, if ${\mathbb{C}}_\infty \setminus G$ is connected.
Let $G \subset {\mathbb{C}}$ be a simply connected domain. Assume that $\varphi \in {\rm Aut} (G)$ and $\varphi$ has no fixed points. Then for every $f \in HC(C_\varphi )$ there is a family $\mathcal F \subset H(G)$ satisfying the following properties:
1. $f \in \mathcal F \subset HC(C_\varphi )$.
2. $\mathcal F$ is dense in $H(G)$.
3. For each finite nonempty subfamily $\mathcal F_0 \subset \mathcal F$, one has $\prod_{g \in \mathcal F_0} g \in HC(C_\varphi )$.
4. $\mathcal F$ is a linearly free system.
Since $\varphi$ has no fixed points, the operator $C_\varphi$ is mixing (see [@grosseperis2011 Theorem 4.37]). In the context of F-spaces (as $H(G)$ is), this is equivalent to the fact that, for every (strictly increasing) subsequence $(n_k) \subset {\mathbb{N}}$, the set $\mathcal U ((C_\varphi^{n_k}))$ of functions having dense $(C_\varphi^{n_k})$-orbit is dense (even residual) in $H(G)$. Of course, mixing property implies hypercyclicity. Observe that $(C_\varphi)^m = C_{\varphi^m}$ for $m \in {\mathbb{N}}$, where $\varphi^1 := \varphi$ and $\varphi^{m+1} := \varphi \circ \varphi^m$.
Fix $f \in HC(C_\varphi )$ as well as a countable topological basis $(G_n)$ for the topology of $H(G)$ (recall that $H(G)$ is metrizable and separable, hence second countable). We set $f_0 := f$ and $G_0 := H(G)$. Then there is a sequence $(p(1,k)) \subset {\mathbb{N}}$ such that $f_0 \circ \varphi^{p(1,k)} \to 1$ $(k \to \infty )$ uniformly on each compact subset of $G$. Since the set $\mathcal U ((C_\varphi^{p(1,k)}))$ is dense in $H(G)$, there exists $f_1 \in G_1 \cap \mathcal U ((C_\varphi^{p(1,k)}))$. In particular, for the constant function $z \mapsto 1$, there is a subsequence $(p(2,k)) \subset (p(1,k))$ satisfying $f_1 \circ \varphi^{p(2,k)} \to 1$ $(k \to \infty )$ in $H(G)$. Of course, $f_0 \circ \varphi^{p(2,k)} \to 1$ in $H(G)$.
By following this procedure, we can obtain recursively a countable family $\mathcal F := \{f_n\}_{n \ge 0} \subset H(G)$ as a well as countable family $\{(p(n,k))_{k \ge 1}\}_{n \ge 0}$ of strictly increasing sequences of natural numbers, satisfying, for each $n \ge 0$, the following:
- $f_n \in G_n \cap \mathcal U ((C_\varphi^{p(n,k)}))$,
- $(p(n+1,k)) \subset (p(n,k))$, and
- $f_j \circ \varphi^{p(n+1,k)} \to 1$ $(k \to \infty )$ in $H(G)$ for all $j = 0, \dots ,n$.
Since $(G_n)$ is a topological basis and $f_n \in G_n$, the set $\mathcal F$ is dense in $H(G)$. By construction, $f \in \mathcal F$ and, since $HC(C_\varphi ) \supset \mathcal U ((C_\varphi^{p(n,k)}))$ for every $n$, we get $\mathcal F \subset HC(C_\varphi )$. Now, if $\varnothing \ne \mathcal F_0 \subset \mathcal F$ with $\mathcal F_0$ finite, we can write $\mathcal F_0 = \{f_{m(1)}, \dots , f_{m(N)}\}$, with $0 \le m(1) < m(2) < \cdots < m(N)$. Let $$h := \prod_{g \in \mathcal F_0} g = \prod_{j=1}^N f_{m(j)}.$$ Trivially, $h \in HC(C_\varphi )$ if $N=1$. If $N > 1$ then $h = h_0 \cdot f_{m(N)}$, where $h_0 := \prod_{j=1}^{N-1} f_{m(j)}$. Fix $F \in H(G)$. As $f_{m(N)} \in \mathcal U ((C_\varphi^{p(m(N),k)}))$, there is a subsequence $(\nu_k ) \subset (p(m(N),k))$ such that $f_{m(N)} \circ \varphi^{\nu_k} \to F$ $(k \to \infty )$ uniformly on compacta in $G$. But $f_{m(j)} \circ \varphi^{\nu_k} \to 1$ $(k \to \infty )$ uniformly on compacta in $G$ for each $j \in \{1,\dots ,N-1\}$. This entails $h_0 \circ \varphi^{\nu_k} \to 1$ in $H(G)$. Thus $$h \circ \varphi^{\nu_k} = (h_0 \circ \varphi^{\nu_k}) \cdot (f_{m(N)} \circ \varphi^{\nu_k}) \longrightarrow 1 \cdot F = F \hbox{ \ in \ } H(G).$$ Consequently, $h \in HC(C_\varphi )$. So far, (a), (b) and (c) have been proved.
In order to prove (d), assume, by way of contradiction, that there is some $N$-tuple $(c_0,c_1, \dots ,c_N) \in {\mathbb{C}}^{N+1} \setminus \{(0, \dots , 0)\}$ such that $c_0f_0 + \dots + c_Nf_N = 0$, where, without loss of generality, we can suppose $c_N \ne 0$. Then $c_0 (f_0 \circ \varphi^{p(N,k)}) + \dots + c_N (f_N \circ \varphi^{p(N,k)}) = 0$ for all $k \ge 1$. Letting $k \to \infty$ we get $c_0 \cdot 1 + \cdots + c_{N-1} \cdot 1 + c_N (f_N \circ \varphi^{p(N,k)}) \to 0$, so $$f_N \circ \varphi^{p(N,k)} \longrightarrow - \sum_{j=0}^{N-1} c_j c_N^{-1} \hbox{ \ in \ } \mathcal H (G) \hbox{ \ as \ } k \to \infty,$$ which is absurd because $\{f_N \circ \varphi^{p(N,k)}\}_{k \ge 1}$ is dense in $\mathcal H (G)$. The theorem is now totally proved.
[**Acknowledgments.**]{}
The authors are indebted to the referee for helpful comments and suggestions, specially for nice simplifications of the proofs of some results.
The first author has been supported by the Plan Andaluz de Investigación de la Junta de Andalucía FQM-127 Grant P08-FQM-03543 and by MEC Grant MTM2015-65242-C2-1-P. The second author has been supported by MEC Grant MTM2016-75963-P. The second and third authors were also supported by Generalitat Valenciana, Project GV/2010/091, and by Universitat Politècnica de València, Project PAID-06-09-2932. The fourth author has been supported by Grant MTM2015-65825-P.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'G.Principe [^1]'
- 'D. Malyshev'
- 'J. Ballet'
- 'S. Funk'
bibliography:
- 'lowe\_cat\_papers.bib'
date: 'Received March 27, 2018; accepted June 12, 2018'
title: 'The first catalog of [*Fermi*]{}-LAT sources below 100 MeV'
---
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Gas inflows fueling AGN are now traceable at high-resolution with ALMA and NOEMA. Dynamical mechanisms are essential to exchange angular momentum and drive the gas to the super-massive black hole. While at 100pc scale, the gas is sometimes stalled in nuclear rings, recent observations reaching 10pc scale (50mas), inside the sphere of influence of the black hole, may bring smoking gun evidence of fueling, within a randomly oriented nuclear molecular disk. AGN feedback is also observed, in the form of narrow and collimated molecular outflows, which point towards the radio mode, or entrainment by a radio jet. Precession has been observed in a molecular outflow, indicating the precession of the radio jet. One of the best candidates for precession is the Bardeen-Petterson effect at small scale, which exerts a torque on the accreting material, and produces an extended disk warp. The misalignment between the inner and large-scale disk, enhances the coupling of the AGN feedback, since the jet sweeps a large part of the molecular disk.'
---
Introduction
============
It is now well established that there exists a tight relation between the mass of the central supermassive black hole, and the bulge mass, or central velocity dispersion, which has been interpreted as a co-evolution of galaxies and black holes [e.g., @Kormendy2013; @Heckman2014]. This co-evolution might be due to a common feeding mechanism, either through mergers or cosmic gas accretion followed by secular evolution, as recenty reviewed by [@Storchi2019], and/or to AGN feedback mechanisms, regulating the star formation in the galaxy host [e.g., @Fabian2012; @Morganti2018].
The new frontier in this domain is to understand in more details the feeding and feedback mechanisms at the highest possible resolution, in the complex circumnuclear region, surrounding the black hole, with the help of multi-wavelength observations [@Ramos2017]. A new view is emerging, where the absorbing material is not due to the long-expected dusty torus [@Hoenig2019]. VLT Interferometer (VLTI) observations showed that the dust on parsec scales is not mainly in a thick torus, but instead in a polar structure, forming like a hollow cone, perpendicular to a thin disk [e.g., @Asmus2016; @Asmus2019]. The circumnuclear region, as a transition between the Broad Line Region (BLR) of the accretion disk, and the Narrow Line Region (NLR), is complex and clumpy, and contains both inflowing material in a thin disk, where millimeter lines have been found, with also H$_2$O masers, and an outflowing component, in the perpendicular direction [e.g., @Cicone2014; @Santi2016].
In the following, I review ALMA observations at high-angular resolution of the molecular gas, revealing nuclear trailing spiral features, that explain the feeding of the central black hole, through exchange of angular momentum. ALMA observations have also revealed outflows, some being extremely collimated in a molecular jet. These outflows must be due to the radio mode of AGN feedback, even when no radio jet has yet been detected. ALMA has also revealed in most nearby Seyferts the existence of molecular circum-nuclear disks, misaligned with the large-scale disks, and with decoupled kinematics. We identify these parsec-scale structures as molecular tori, able to obscure the central accretion disks. Several mechanisms are reviewed to explain the misalignments.
![ALMA observations of the Seyfert-1 galaxy NGC 1566. Left is a zoomed 8”x8” region of the CO(3-2) intensity map, showing the nuclear trailing spiral (1”=35pc). Right, is the model rotation curve from NIR images, represented schematically, with the circular velocity (black) epicyclic frequency $\kappa$ (red), and corresponding $\Omega-\kappa$/2 curve (blue) within the central kpc. The contribution of a super-massive black hole in the nucleus with MBH = 8.3 × 10$^6$ M$_\odot$ has been included. From [@Combes2014; @Combes2019].[]{data-label="fig1"}](combes-fig1.png){width="99.00000%"}
Feeding the monster
===================
To fuel the central black hole, the main problem is to transfer the angular momentum of the gas outwards. This can be done by the gravity torques exerted by bars on the gas in spiral arms [@Santi2005]. Torques are positive outside corotation, and the gas is driven outward to accumulate in a ring at the Outer Lindblad Resonance (OLR). Inside corotation, torques are negative, and gas is driven inward, to pile up in a nuclear ring at the Inner Lindblad resonance (ILR).
What happens inside the ILR depends on the winding sense of orbits there. The gas is orbiting in elliptic streamlines, which gradually tilt by 90$^\circ$ at each resonance and wind up in spiral structures. The precession rate of these elliptical orbits is equal to $\Omega-\kappa$/2, with $\Omega$ the rotation frequency = V/r, and $\kappa$ the epicyclic frequency. Usually, inside ILR, and far from the black hole, $\Omega-\kappa$/2 increases with radius, and the spiral is leading. The torque of the bar is positive, and the gas is driven back to the ILR. But near the massive black hole, the precessing frequency $\Omega-\kappa$/2 is decreasing with radius, and the spiral is trailing. The gas can then fuel the AGN [@Buta1996].
ALMA has the resolution to enter the sphere of influence of the black hole, and a trailing nuclear spiral was first seen in NGC 1566 [@Combes2014]. This nuclear spiral is located well inside the r=400pc ring, corresponding to the ILR of the bar. Fig. \[fig1\] shows the nuclear spiral in the CO(3-2) line (left), and at right the rotation curve and corresponding frequencies, derived from the stellar potential traced by near-infrared images. The precessing rate $\Omega-\kappa$/2 increases towards the center, inside 50pc, due to a black hole of mass 8.3 × 10$^6$ M$_\odot$. Such trailing nuclear spirals have been found also in NGC 613 [@Audibert2019] and in NGC 1808 (Audibert et al. 2020, in prep.).
AGN feedback: jets and winds
============================
Molecular outflows are now commonly observed as AGN feedback [@Cicone2014]. If NGC 1566 does not reveal any outflow, both inflow and outflow can be observed simultaneously, as in NGC 613, where a very short (23pc) and small velocity (300km/s) outflow is detected on the minor axis, parallel to the VLA radio jet [@Audibert2019]. A very small molecular outflow is also seen in NGC 1433, along the minor axis, cf Fig. \[fig2\]. This might be the smallest molecular outflow in a nearby Seyfert galaxy, and could be associated to a past radio jet [e.g., @Combes2013; @Smajic2014].
In these nearby low-luminosity AGN, which accrete far below the Eddington limit, the main mechanism to drive molecular outflows is the radio mode, i.e. entrainement by the radio jets. In some more luminous cases, where L approaches L$_{Edd}$/100, there could be both the radio mode, and winds generated by radiation pressure (either in the ionized gas, or on dust). This might be the case of the prototypical Seyfert-2 NGC 1068, where there is clearly a molecular outflow parallel to the radio jet, sweeping part of the galactic disk [@Santi2014]. The jet is not perpendicular to the plane, due to the misalignment of the accretion disk with the plane. The ALMA observations of the various CO rotational lines reveal clearly a molecular torus, almost edge-on, and a molecular flow in the perpendicular direction, aligned with the polar dust. The molecular disk appears warped and tilted with respect to the H$_2$O maser disk [@Santi2016].
![ALMA CO(3-2) observations of the Seyfert-2 NGC 1433: Left, the velocity field, with the color bar labelled in km/s; The thin line indicates the minor axis (PA=109$^\circ$). Right, spectrum summing the blue and red-shifted components close to the center, along the minor axis. The systemic velocity is V$_{sys}$= 1075 km/s. The mass in the outflow is 3.6 10$^6$ M$_\odot$. From [@Combes2013; @Combes2019].[]{data-label="fig2"}](combes-fig2.png){width="99.00000%"}
The lenticular galaxy NGC 1377 is an exceptional case, with a very thin and highy collimated molecular outflow, in the absence of any detected radio jet [@Aalto2016]. The molecular outflow changes sign along the flow, on each side of the galaxy. This means that the jet is almost in the plane of the sky, and that a slight precession of only 10$^\circ$ is able to tip the jet from redshifted to blue-shifted and back. Such a precession is observed in micro-quasars jets in the Milky Way, for instance SS433 [@Mioduszewski2005]. But this can be attributed to the companion star. Here there must exist another origin of the precesion, which could be relativistic (see next section).
A precessing molecular outflow model is compatible with the data [@Aalto2016]. The flow is launched close to the center (r $<$ 10pc). A radio jet must exist at a low level, or has existed in a recent past.
Molecular tori: misalignment
============================
With the high spatial resolution of ALMA, it was possible to unveil circumnuclear disks in the CO emission, towards nearby Seyferts. These happen to be misaligned to the large-scale disks, and kinematically decoupled. We call them molecular tori, they exist in 7 out of the 8 cases observed [@Combes2019]. The average radius of the molecular tori is 18 pc, with a median at 21 pc. Their average molecular mass is M(H$_2$) = 1.4 10$^7$ M$_\odot$, and in average their inclination on the plane of the sky is 29$^\circ$ different from their galactic disk.
These molecular tori are clearly within the sphere of influence of their black holes, and can serve to measure their mass, provided that their inclination is sufficient [@Combes2019]. This has been done also for more massive early-type galaxies, by the WISDOM project [@Davis2018].
We can invoke at least three mechanisms of misalignment between the large-scale galactic disks and the molecular tori and/or accretion disks. One of them is the radiation-driven warping instability [@Pringle1996]. A tilted optically thick disk, which absorbs the radiation from the central AGN, receives in each point some momentum from the radiation, but no torque, because of the radial direction. But then it re-radiates perpendicularly to its orientation, and this produce torques, which maintain and amplify the warping. Assuming the luminosity is powered by accretion eliminates the unknown viscosity parameter $\alpha$. The instability occurs for radii R $>$ 0.1pc M$_{BH}$/(10$^8$M$_\odot$). The efficiency of the mechanism was tested by simulations, both in the case of retrograde and prograde precession with respect to the disk rotation [@Maloney1997]. A second meschanism is the magnetic instability, and consequent torques but compatibility with AGN observations is contrived, it is more adapted to accretion disks around magnetic stars [@Pfeiffer2004].
A third mechanism uses the Bardeen-Petterson effect [@Bardeen1975], due to Lense-Thirring precession. The accretion disk has a random orientation, generally not aligned with the black hole spin. The relativistic frame dragging effect induces a precession, which tends to align the inner parts of the accretion disk with the black hole equator. The disk develops a warping up to distances 10$^2$ to 10$^4$ Schwarzschild radius R$_s$. The precession of the disk and its warp can be seen from inner to outer disk, up to 1 pc M$_{BH}$/(10$^9$M$_\odot$). According to the amplitude of viscosity, one can distinguish two regimes: the diffusion, when $\alpha >$ H/R, where H is the height of the disk, and the regime of bending waves, when $\alpha <$ H/R [@Papaloizou1983]. In the first case, the disk is warping smoothly and continuously, while in the second case, the disk can break in several rings, with different inclinations and precessing rates. Then the various rings, with differential precession, collide, and drive the gas to fuel the AGN more quickly. This regime has been simulated by [@Nealon2015]. Some works found that the alignment of accretion disks with the black hole equator, through the Bardeen-Petterson effect, was inefficient [@Zhuravlev2014; @Banerjee2019a; @Banerjee2019b]. More precisely, according to some viscous parameters (parallel of perpendicular to the disk), and viscosity generated by magnetized turbulence, the disk near the black hole can retain its inital inclination, instead of aligning, cf. Fig. \[fig3\]. Then precessing jets can be launched, perpendicular to the disk, but not aligned to the black hole spin [@Liska2018; @Liska2019].
![Misalignment of the accreting material near a black hole: Left: radial profiles of the disk tilt-angle (with respect to the black hole spin), for several values of the parameter a, dimensionless angular momentum $J$ of the black hole $M$, 0 $<$ a $<$ 1 (a= c$J$/(G$M^2$)). The initial tilt-angle is 5$^\circ$. The radius is in unit of the gravitational radius R$_g$ = G$M$/c$^2$. From [@Banerjee2019a]. Right: Simulation of the Bardeen-Petterson effect, in a disk initially inclined by 60$^\circ$ with respect to the black hole spin. The Lense-Thirring precession causes the disk to break in a few tilted rings. Adapted from [@Nealon2015].[]{data-label="fig3"}](combes-fig3.png){width="99.00000%"}
A manifestation of these warping instabilities is the observation of the warped maser disks. Water masers in the prototypes NGC 4258 [@Herrnstein1999] and NGC 1068 [@Gallimore2004] are detected on 0.3-0.8 pc warped discs. These observations are best represented by the Lense-Thirring effect and/or the radiation driven warps. These perturbations also heat the disk. The fitting of the observations has been done for NGC 4258 by [@Martin2008], and for NGC 1068 by [@Caproni2006]. In NGC 1068, where the Bardeen-Petterson mechanism gives the best fit with the observations, the disk is aligned with the black hole spin, until the radius R$_{BP}$ = 10$^{-5}$ to 10$^{-4}$ pc, then warps. For one of the best fit models the alignment time-scale is 7580 yr, the misalignment angle 40$^\circ$ and the velocity of the jet 0.17c. The shape and precession of the pc and kpc-scale jet is also fitted, following [@Wilson1987], in addition to the warped H$_2$O maser disk by [@Gallimore2004].
How can the gas accreted from the galactic disk be so misaligned with the disk itself? First, the potential in the center is almost spherical, and the disk very thick with respect to the parsec-scales in question, and second, star formation feedback constantly ejects some gas out of the plane, which rains down in a fountain effect at a random orientation, sometimes in a polar ring [@Renaud2015; @Emsellem2015].
Summary
=======
Thanks to the high resolution provided by ALMA on the molecular gas, it is now possible to better understand the AGN fueling mechanisms. If at large scale, the primary bars can drive the gas towards the 100pc scales, in ILR rings, the nuclear bars act on a trailing nuclear spiral to drive the gas towards the black hole, when the circumnuclear gas enters its sphere of influence. Inside the nuclear spiral, ALMA has revealed the existence of morphologically and kinematically decoupled circum-nuclear disks, or molecular tori.
In some less frequent cases, we can see both AGN fueling, and molecular outflows, through AGN feedback. This can occur via entrainment by radio jets, or through disk winds (or both). The radio mode is distinguished by extremely thin and collimated molecular jets, sometimes precessing, as are the entraining radio jets.
To explain the precession, and the misalignment of the circum-nuclear disks, we can invoke at small scale the Bardeen-Petterson effect, which produces torques in the accreting material, to align it with the black hole spin, and then induces a warping up to a fraction of parsec-scale. The black hole is likely to be fueled via several accretion episodes, coming from the large-scale galacic disk, but also the fountain gas, ejected above the plane through supernovae feedback.
2016, *A&A*, 590, A73
2019, *MNRAS*, 489, 2177
2016, *Ap. J.*, 822, 109
2019, *A&A*, 632, A33
2019b, *MNRAS*, 487, 3488
2019a, *Ap. J.*, 870, 95
1975, *Ap. J.*, 195, L65
1996, *Fund. Cosmic Phys.*, 17, 95 . 2006, *Ap. J.*, 638, 120
2014, *A&A*, 562, A21
2013, *A&A*, 558, A124
2014, *A&A*, 565, A97
2019, *A&A*, 623, A79
2018, *MNRAS*, 473, 3818
2015, *MNRAS*, 446, 2468
2012, *ARAA*, 50, 455
2004, *Ap. J.*, 613, 794
2005, *A&A*, 441, 1011
2014, *A&A*, 567, A125
2016, *Ap. J.*, 823, L12
2014, *ARAA*, 52, 589
1999, *Nature*, 400, 539
2019, *Ap. J.*, 884, 171
2013, *ARAA*, 51, 511
2018, *MNRAS*, 474, L81
2019, *MNRAS*, 487, 550
1997, *Ap. J.*, 491, L43
2008, *MNRAS*, 387, 830
2005, *ASPC*, 340, 281
2018, *A&ARv*, 26, 4
2015, *MNRAS*, 448, 1526
1983, *MNRAS*, 202, 1181
2004, *Ap. J.*, 604, 766
1996, *MNRAS*, 281, 357
2017, *NatAs*, 1, 679
2015, *MNRAS*, 454, 3299
2014, *A&A*, 567, A119
2019, *NatAs*, 3, 48
1987, *Ap. J.*, 319, 105
2014, *Ap. J.*, 796, 104
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the four-fold anisotropy of in-plane ferromagnetic resonance (FMR) field $H_r$, found in a square lattice of circular Permalloy dots when the interdot distance $a$ gets comparable to the dot diameter $d$. The minimum $H_r$, along the lattice $\langle11\rangle$ axes, and the maximum, along the $\langle10\rangle$ axes, differ by $\sim$ 50 Oe at $a/d$ = 1.1. This anisotropy, not expected in uniformly magnetized dots, is explained by a non-uniform magnetization $\bm(\br)$ in a dot in response to dipolar forces in the patterned magnetic structure. It is well described by an iterative solution of a continuous variational procedure.'
author:
- 'G.N. Kakazei,$^{1,2}$ Yu.G. Pogorelov,$^3$ M.D. Costa,$^4$ T. Mewes,$^{5}$ P.E. Wigen,$^2$ P.C. Hammel,$^2$ V.O. Golub,$^{1}$ T. Okuno,$^6$ V. Novosad$^7$'
title: 'Origin of four-fold anisotropy in square lattices of circular ferromagnetic dots'
---
ł ø § Ø
\#1[(\[\#1\])]{}
Magnetic nanostructures are of increasing interest for technological applications, such as patterned recording media [@moser], or magnetic random access memories [@allwood]. One of the most important issues for understanding their collective behavior is the effect of long-range dipolar interactions between the dots [@demokritov]. For the single-domain magnetic state of a dot, the simplest approximation is that dots are uniformly magnetized and interactions only define relative orientation of their magnetic moments [@guslienko]. If so, the system of dipolar coupled dots in a square lattice should be magnetically isotropic.
However, in all known experimental studies of closely packed arrays of circular dots, a four-fold anisotropy (FFA) was found, either by Brillouin light scattering [@mathieu], ferromagnetic resonance (FMR) [@jung] or magnetization measurements (from hysteresis loops) [@natali; @zhu]. It is important to note that FFA exists in both unsaturated samples and saturated ones (i.e. above vortex annihilation point on the hysteresis loop). Hence it cannot be only associated with vortex formation suggested in Ref. [@natali]. It was instead qualitatively related to stray fields from unsaturated parts of magnetization inside the dots [@mathieu]. However no quantitative description of FFA in such systems was given up to now. So the aim of this study is to explain quantitatively the deviations from isotropy in terms of modified demagnetizing effect in a patterned planar system at decreasing inter-dot distance, from the limit of isolated dot to that of continuous film. The choice of *X*-band FMR techniques for this study has an advantage in eliminating possible interference from domain (vortex) structure [@kakazei03]. The variational theoretical analysis is followed by micromagnetic simulations.
Permalloy (Py) dots were fabricated with electron beam lithography and lift-off techniques, as explained elsewhere [@novosad]. The dots of thickness $t$ = 50 nm and diameter $d$ = 1 $\mu$m were arranged into square arrays with the lattice parameter $a$ (center to center distance) varying from 1.1 $\mu$m to 2.5 $\mu$m. The dimensions were confirmed by atomic force microscopy and scanning electron microscopy. Room temperature FMR studies were performed at 9.8 GHz using a standard *X*-band spectrometer. The dependence of the FMR field $H_r$ on the azimuthal angle $\varphi_H$ of applied field $\mathbf{H}$ with respect to the lattice \[10\] axis for almost uncoupled dots ($a$ = 2.5 $\mu$m) is shown in Fig. \[fig1\]a. Only a weak uniaxial anisotropy of $H_r(\varphi_H)$ is present here, which can be fitted by the simple formula $H_{r}(\varphi_H) =
H_{r,av} + H_{2}\cos2\varphi_H$. For the $a$ = 2.5 $\mu$m sample, we found the average peak position $H_{r,av} \approx 1.13$ kOe and the uniaxial anisotropy field $H_{2} \approx 5$ Oe. The latter value remains the same for the rest of our samples, so this uniaxial anisotropy is most probably caused by some technological factors.
With decreasing distance $a$ between dots, two changes are observed in the $H_r(\varphi_H)$ dependence. First, $H_{r,av}$ decreases to $\approx 1.09$ kOe at $a$ = 1.1 $\mu$m (Fig. \[fig1\]b). Second, a four-fold anisotropy (FFA) is detected in the samples with $a\leq$1.5 $\mu$m by pronounced minima of $H_{r}(\varphi_H)$ at $\varphi_H$ close to the lattice $\langle11\rangle$ axes. This behavior is fitted by $H_r(a,\varphi_H) = H_{r,av}(a) + H_4(a)
\cos4\varphi_H + H_2 \cos2\varphi_H$, as shown in Fig. \[fig1\]b. The interdot distance dependence of $H_{r,av}$ and FFA field $H_{4}$ is shown in Fig. \[fig2\]. Also such anisotropy is detected in the FMR linewidth, smaller for $\langle11\rangle$ than for $\langle10\rangle$ case (reaching $\sim$30% at $a/d$ = 1.1).
The FFA effect, which could not arise in uniformly in-plane magnetized cylindrical dots, is evidently related to a non-uniform distribution of the magnetization $\bm(r,\varphi,z)$ (in cylindric coordinates $0 \leq r \leq R = d/2, 0 \leq \varphi < 2\pi, 0\leq z
\leq t$). A similar effect was discovered using Brillouin light scattering [@mathieu] and magnetization reversal [@natali; @zhu] in such systems under weak enough external fields, which displace vortices in each dot. This can be modeled by displacements of two oppositely in-plane magnetized uniform domains [@guslienko01]. But in the presence of external fields strong enough to observe FMR, one has to assume a continuous (and mostly slight) deformation of $\mathbf m(r,\varphi)$. The simplest model for such deformation uses a variational procedure with respect to a single parameter [@metlov]. However, as will be shown below, the non-uniform magnetic ground state of this coupled periodic system results from a rather complicate interplay between intra-dot and inter-dot dipolar forces, which requires a more general variational procedure.
Assuming fully planar and $z$-independent dot magnetization with the 2D Fourier amplitudes $\bm_\bg = \int \textrm{e}^{i\bg \cdot \br}
\bm(\br)d\br$, the total (Zeeman plus dipolar) magnetic energy (per unit thickness of a dot) can be written as (see Appendix)
E = - \_0 + \_[0]{} |\_|\^2,
where $f(u)=1-\left(1-\textrm{e}^{-u}\right)/u$ [@guslienko] and the vectors of the 2D reciprocal lattice are $\varphi_H$-dependent: $\bg = (2\pi/a) (n_1 \cos\varphi_H - n_2 \sin
\varphi_H, n_1 \sin \varphi_H + n_2 \cos \varphi_H)$ (for $\mathbf{H}\parallel x$ and integer $n_{1,2}$). The variation of exchange energy at deformations on the scale of whole sample is of the order of stiffness constant ($\sim 10^{-6} \textrm{ erg/cm}$ for Py) and it can be neglected beside the variation $\sim H M_s d^2
\sim 10^{-2} \textrm{ erg/cm}$ of terms included in Eq. \[eq1\]. If the dot magnetization has constant absolute value: $\bm (\br) =
M_s \left (\cos \varphi(\br), \sin \varphi(\br) \right)$, its variation: $\d \bm (\br) = \hat{\bz} \times \bm (\br)
\d\varphi(\br)$ (where $\hat{\bz}$ is unit vector normal to plane), is only due to the angle variation $\d \varphi (\br)$. Using the Fourier transform $\d\bm_\bg = \hat{\mathbf{z}} \times
\sum_{\bg^\prime} \bm_{\bg - \bg^\prime}\d\varphi_{\bg^\prime}$ in the condition $\d E = 0$ leads to the equilibrium equation for the Fourier amplitudes:
m\_[,y]{} & = & \_[\^0]{} (\_[\^]{} \^)\
& & (\_[- \^]{} \^) .
It can be suitably solved by iterations:
m\^[(n)]{}\_[,y]{} &=& \_[\^0]{} (\^[(n-1)]{}\_[\^]{} \^)\
& & (\^[(n-1)]{}\_[- \^]{} \^),
starting from uniformly magnetized dots as zeroth iteration: $m^{(0)}_{\bg,y} = 0$, $m^{(0)}_{\bg,x} = 2\pi R M_s
J_1(g R)/g$ (with the Bessel function $J_1$). Already the 1st iteration (including the inverse Fourier transform):
\^[(1)]{}\_y()& =& -(d-2r) \_[0]{}\
& &J\_1(g R) (),
(with the Heavyside $\theta$ function) reveals the FFA behavior, due to the rotationally non-invariant product $g_x g_y $. The calculated maximum variation of $\varphi(\br) = \arcsin [m_y
(\br)/M_s]$ in $\langle10\rangle$ field geometry is $\sim60\%$ bigger than in the $\langle11\rangle$ geometry (Fig. \[fig3\], upper row). This expected behavior persists upon further iterations. Our analytic approach was checked, using the micromagnetic OOMMF code [@donahue] on a 9$\times$9 array of considered disks (Fig. \[fig3\], lower row) at standard values of $M_{s} = 0.83$ kOe and exchange stiffness $1.3 \cdot 10^{-6}$ erg/cm [@kakazei04] for Py. The distributions obtained in this way for the central disk in the array are within $\sim10\%$ to the analytic results of the 1st iteration.
{width="8cm"}
![ Density plots of the equilibrium magnetization angle $\varphi(\mathbf r)$ for two field geometries, $\varphi_H = 0$ ($\langle10\rangle$) and $\varphi_H = \pi/4$ ($\langle11\rangle$). Upper row: calculated from the sum, Eq. \[eq5\], over 100$\times$100 sites of reciprocal lattice at parameter values $a =
1.1\, \mu$m, $H=$ 1.1 kOe, $M_s=$ 0.83 kOe. Lower row: micromagnetic calculation by OOMMF code for the central disc in the 9$\times$9 array.](Fig4.eps){width="8cm"}
The FMR precession of $\bm(\br)$ is defined by the internal field $\mathbf{H}_i(\br) = \mathbf{H} + \mathbf{h}(\br)$ through the local dipolar field
h\_z()&=&-\_m\_[,z]{}(),\
h\_()&=&-\_[,0]{} \_[,]{}(),
($\a,\b=x,y$). The 1st iteration for $\mathbf{h}(\br)$ corresponds to the zeroth iteration for $\bm(\br) =
(M_s,\m_y,\m_z)$, which now includes the uniform FMR amplitudes $\m_y,\,\m_z$. Then the local demagnetizing factors $N_x(\br) = -
h_x(\br) / M_s,\,N_{y,z}(\br) = -h_{y,z}(\br)/\m_{y,z}$ define the local FMR field $H_r(\br)$:
H\_r() &=&\
&-& M\_s\[N\_z()+N\_y()-2N\_x()\]/2
(here $H_{0}\approx 3.3$ kOe). The average FMR field is defined by the isotropic averaged demagnetizing factors
$$\begin{aligned}
\overline{N}_{x,y}& = &\left(\pi R^2 \right)^{-1} \int_{r < R}
N_{x,y}
\left(\br\right)d\br \nonumber\\
&=& (8\pi^2/a^2)\sum_{\bg\neq 0}f(g t)J_1^2(g R)/g^2, \label{eq8}\end{aligned}$$
and $\overline{N}_z = 4\pi-2\overline{N}_x$. At $a\to\infty$, they tend to the single dot values [@joseph] which are for $t/R=0.1$: $N_{x,y}^{(d)}\approx0.776$ and $N_z^{(d)}
\approx 11.01$. Using $N_i^{(d)}$ instead of $N_i(\br)$ in Eq. \[eq7\] accurately reproduces the single dot FMR limit $H_{r}^{(d)} \approx 1.15$ kOe (estimated from Fig. \[fig2\]b). Otherwise, for decreasing interdot distance, $a\to d$, the 1st iteration values, Eq. \[eq8\], used in Eq. \[eq7\] well describe the tendency of $H_{r,av}(a)$ towards the continuous film limit $H_r^{(f)}=\sqrt{H_0^2+4\pi^2M_s^2}-2\pi M_s\approx0.96$ kOe (Fig. \[fig2\]a).
Finally, by calculating the true local FMR fields $H_{r}(\br)$ from Eqs. \[eq6\] and \[eq7\], the field dependent absorption is obtained as $I(H)\propto\int_{r<R} \d(H-H_{r}(\br))d\br$. Then the FMR fields, $H_r$ defined from maximum of $I(H)$ in two geometries, display FFA in a good agreement with the experimental data (Fig. \[fig2\]). This effect is due to the fact that stronger deformation of magnetization stronger suppresses the demagnetizing effect (the differences $N_z-N_{x,y}$) and thus enhances $H_r$. Also it produces a bigger spread of local resonance fields $H_r(\br)$ and thus broadens the FMR line, again in agreement with our observations.
In conclusion, it is shown that under in-plane magnetic fields, $\mathbf{H}$, even strong enough for FMR, the dipolar coupling in a dense lattice of circular magnetic dots is able to produce a continuous deformation of the dot magnetization, strongest for the field orientation along lattice axes.
Work at ANL was supported by the U.S. Department of Energy, BES Materials Sciences under Contract No. W-31-109-ENG-38; MDC was supported by FCT (Portugal) and the European Union, through POCTI (QCA III) grant No. SFRH/BD/7003/2001.
[10]{} A. Moser, K. Takano, D.T. Margulies, M. Albrecht, Y. Sonobe, Y. Ikeda, S. Sun, and E.E Fullerton, J. Phys. D: Applied Physics **35**, R157 (2002); S. Sun, D. Weller, J. Magn. Soc. Jpn. 25, 1434 (2001). D.A. Allwood, Gang Xiong, M.D. Cooke, C.C. Faulkner, D. Atkinson, N. Vernier, and R. P. Cowburn, Science **296**, 2003 (2002). S.O. Demokritov, B. Hillebrands, and A.N. Slavin, Phys. Rep. **348**, 441 (2001). K.Yu. Guslienko and A.N. Slavin, J. Appl. Phys. **87**, 6337 (2000); J. Magn. Magn. Mat. **215**, 576 (2000). C. Mathieu, C. Hartmann, M. Bauer, O. Büttner, S. Riedling, B. Roos, S.O. Demokritov, and B. Hillebrands, Appl. Phys. Lett. **70**, 2912 (1997). S. Jung, B. Watkins, L. DeLong, J.B. Ketterson, and V. Chandrasekhar, Phys. Rev. B **66**, 132401 (2002). M. Natali, A. Lebib, Y. Chen, I.L. Prejbeanu, and K. Ounadjela, J. Appl. Phys. **91**, 7041 (2002). X. Zhu, P. Grutter, V. Metlushko, and B. Ilic, Appl. Phys. Lett. **80**, 4789 (2002). G.N. Kakazei, P.E. Wigen, K.Y. Guslienko, R.W. Chantrell, N.A. Lesnik, V. Metlushko, H. Shima, K. Fukamichi, Y. Otani, and V. Novosad, J. Appl. Phys. **93**, 8418 (2003). V. Novosad, K. Yu. Guslienko, H. Shima, Y. Otani, S. G. Kim, K. Fukamichi, N. Kikuchi, O. Kitakami, and Y. Shimada, Phys. Rev. B 65, 060402 (2002). K.Yu. Guslienko, Phys. Lett. **278**, 293 (2001). K.L. Metlov, Phys. Stat. Sol. (a) **189**, 1015 (2002); K.L. Metlov and K.Yu. Guslienko, Phys. Rev. B **70**, 052406 (2004). M.J. Donahue and D.G. Porter, URL: http://math.nist.gov/oommf G.N. Kakazei, P.E. Wigen, K.Y. Guslienko, V. Novosad, A.N. Slavin, V.O. Golub, N.A. Lesnik, and Y. Otani, Appl. Phys. Lett. **85**, 443 (2004). R.I. Joseph and E. Schlömann, J. Appl. Phys. **36**, 1579 (1965).
Appendix
========
For fully planar and *z*-independent dot magnetization, the dipolar energy per unit thickness of a dot in the lattice is:
E\_d & = & 1 [2t]{} \_[-t/2]{}\^[t/2]{} dz \_[-t/2]{}\^[t/2]{} dz\^\_c dd\^\_[,]{} m\_()\
& & ,
where the 2D integrations $\int_c d\br$ and $\int d\br$ are respectively over the unit cell and over the entire plane. It can be also presented as
E\_d & = & 1 [2t]{} \_[-t/2]{}\^[t/2]{} dz \_c d\_ m\_() h\_(,z)\
& =& 1 [4t a\^2]{} \_[-t/2]{}\^[t/2]{} dz \_[-]{}\^ dq [e]{}\^[-i q z]{} \_[,]{} m\_[,]{} h\_[,,q]{},
where the Fourier amplitudes of the dipolar field are:
h\_[,,q]{} & = & \_c d\_[-]{}\^ dz\^\^[i (+ qz\^)]{} h\_(,z\^)\
& = & \_c d\_[-]{}\^ dz\^\^[i ( + q z\^)]{} d\^\_[-t/2]{}\^[t/2]{} dz\^\
& & \_ .
To calculate them, we express the lattice magnetization $m_\b(\br^\prime)$ through its Fourier amplitudes:
$$m_\b(\br^\prime) = \frac 1 {a^2} \sum_{\bg^\prime} {\rm e}^{-i \bg^\prime
\cdot \br^\prime} m_{\b,\bg^\prime},$$
and then introduce the factor ${\rm e}^{i(\bg^\prime \cdot \br -
qz^{\prime\prime})}$ into the integral in $d\br^\prime dz^{\prime}$, and the compensating factor ${\rm e}^{-i(\bg^\prime \cdot \br -
qz^{\prime\prime})}$ into the integral in $d\br dz^{\prime\prime}$. Then the spatial integrations in $E_d$ are done accordingly to the formulas:
&&\_c d\^[i (- \^) ]{} = a\^2 \_[,\^]{},\
&&\_[-t/2]{}\^[t/2]{} dz \_[-t/2]{}\^[t/2]{} dz\^[e]{}\^[i q (z\^[ ]{} - z)]{} = 4 [q\^2]{} \^2 2,\
&& \_[-]{}\^ dz\^d\^\^[i(- \^) + i q(z\^- z\^)]{}\
&& 1\
& & = .
Finally, the momentum integration
$$\int_{-\infty}^{\infty} \frac{\sin^2 (qt/2)} {q^2({g}^2 +
q^2)} dq = \frac{\pi t}{2 {g}^2} f(g t)$$
leads to the result included in Eq. \[eq1\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a problem: to find the minimum number of correlation function measurements that are required to determine whether a given quantum system is entangled or separable. In particular, the case when restricted knowledge of only zero or non-zero correlations is available is examined. Through the concrete investigation on a pure state for two particles, each particle having two basis states (the $2\times 2$ system), we show that the minimum number of measurements for entanglement detection with certainty is three. This involves four quantum observables. We next consider the comparable problem applied to the mixed density matrix. The mixed quantum state case appears to require more detailed information, which we illustrate by studying the concrete example of the Werner state.'
author:
- |
Toru Ohira[^1]\
Graduate School of Mathematics, Nagoya University, Nagoya, Japan
title: Minimal Correlation Measurements for Entanglement Detection
---
Introduction
============
Correlation measurements are a standard tool to investigate quantum systems. To determine whether the quantum system is separable or entangled, correlation measurements have been actively performed alongside of various theoretical investigations. (e.g.,[@bell; @clauser; @aspect; @leggett; @vedral; @altepeter; @sakai]). For example, a recent work by Fujikawa et. al.[@fujikawa2] analyzed experimental results with the view of associating separability with zero–correlations, which is one of many active endeavors for obtaining conditions for separabilities of density matrices(e.g.,[@werner; @peres; @horodecki])
Here, we try to be as simplistic as possible and consider the approach of utilizing correlation functions that appear in traditional theories and measurements to detect entanglements. In particular, we base our analysis on achieving zero–correlations among observables measured on a composite quantum system. In other words, we assume that we need only measure the existence of correlations (either zero or non-zero). We do not require information on the exact values of non-zero correlation functions. It should be noted that this constraint simplifies the analysis of measurements. This approach is different from other more elaborate mathematical methods for entanglement detections or separability conditions where more detailed information on correlation functions and/or the density matrix is necessary ([@horodecki2; @guhne] for overviews).
The main question we proposed to ask with the above setup is the following: what is the minimum number of such correlation measurements we need to perform to detect entanglements with certainty? To gain insights toward this question, we start with the simplest example of the two-particle (bipartite) $2\times 2$ system. It is found that if the system is in a pure quantum state, three correlation measurements involving four observables are sufficient to distinguish the entangled state from the separable one with certainty.
For mixed states, however, the answer to the above question is not obtained. We illustrate the difficulty by considering the density matrix called the Werner state and show that the correlation structure is the same regardless of whether there is entanglement or separability. Particularly, in this case, zero correlations appear for entangled states just in the same way as the separable states. This implies that non-zero correlations do not certify the existence of entanglement nor does zero correlation entail separability.
We end the paper with discussions on extensions of this line of approach to higher dimensions as well as general questions on concepts of correlations and entanglements in capturing quantum nature through a brief mention of the quantum pigeon hole effect.
Main Question
=============
We consider a typical situation of correlation measurements on pure or mixed quantum systems. Let us start with the simplest case of a bipartite system consisting of two quantum particles $A$ and $B$. We can measure observable operators $\mathcal{X}$ and $\mathcal{Y}$ defined as $\mathcal{X} = {\mathcal{Q}_A}\otimes{\bf{1}}_B$ and $\mathcal{Y} = {\bf{1}}_A\otimes\mathcal{R}_B$. They can be measured independently meaning measurement only on particle $A$, or only on particle $B$. Or, they can be measured at the same time. Through these measurements, we obtain three expectation values, $\bra{\psi}\mathcal{X}\ket{\psi}, \bra{\psi}\mathcal{Y}\ket{\psi}$, and $\bra{\psi}\mathcal{X}\mathcal{Y}\ket{\psi}$, from which we can compute the correlation (covariance) as $$c(\mathcal{X}, \mathcal{Y}) = \bra{\psi}\mathcal{X}\mathcal{Y}\ket{\psi} - \bra{\psi}\mathcal{X}\ket{\psi}\bra{\psi}\mathcal{Y}\ket{\psi}.$$
As is well known, if the system is described by a density matrix $\rho_{AB}$, we can generalize the above using its trace and its partial traces $\rho_A$ and $\rho_B$ to describe the quantum state and compute the correlation function in the following manner. $$c(\mathcal{X}, \mathcal{Y}) = Tr(\rho_{AB}\mathcal{X}\mathcal{Y}) - Tr(\rho_{A}\mathcal{X})Tr(\rho_{B}\mathcal{Y}).$$
We also mention that a density matrix describing a pure quantum system is separable if it can be written as $$\rho_{AB}=\rho_A\otimes\rho_B,$$ otherwise it is entangled. For a density matrix describing a mixed state, the separability is defined as $$\rho_{AB}=\sum_i p_i\rho^i_{A}\otimes\rho^i_{B}$$ with non-negative real $\{ p_i \}$ such that $\sum_i p_i = 1$. If such decomposition is not possible, it is entangled.
With the set up above our main question can be phrased as
[**[Main Question]{}**]{}
What is the minimum number of correlation measurements we need to perform on a given quantum state to determine with certainty whether it is entangled or separable?
We assume that we can design experiments to measure observables $\mathcal{X}$ and $\mathcal{Y}$ to obtain the above mentioned-expectation values. The question is to find the minimum number of measurements of correlation functions so that we can be certain to determine whether the given system is in a separable or an entangled state. We also assume that the correlation measurement results are binary: either zero or non-zero. That is to say, we do not need to know the values of any non–zero correlation function. We will see that this leads to a relatively simple measurement analysis. It should be noted that, with this restriction, our approach is different from much investigated separability criteria and entanglement detections, where more information from quantum systems is required[@horodecki2; @guhne]. In the following, we illustrate how much we can gain insight into quantum systems even with these restricted correlation measurements.
Analysis with the $2\times 2$ system
====================================
To gain insight, we proceed to consider the simplest bipartite system of two 2-state particles ($2\times 2$ system), such as two spin 1/2 particles or two qubits systems(e.g.,[@wootters; @kummer; @abouraddy; @jchen]).
Our analysis will employ the density matrix, whether or not the system is in a pure or mixed state. It is known that any density matrix for $2\times2$ systems can be written using the Pauli matrices as follows.
$$\rho_{AB}={1\over 4}({\bf{1}_A}\otimes {\bf{1}_B}+{\vec{a}}\cdot{\vec{\mathcal{\sigma}}} \otimes {\bf{1}_B} + {\bf{1}_A}\otimes {\vec{b}}\cdot{\vec{\mathcal{\sigma}}} +\sum_{ij} F_{ij}{\mathcal{\sigma}}_i\otimes {\mathcal{\sigma}}_j)
\label{density}$$
where ${\bf{1}}$ is the $2\times2$ identity matrix, ${\vec{a}},{\vec{b}}$ are vectors consists of 3 real numbers ($\cdot$ is the inner product), and $F_{ij}$ are real number elements of a $3\times3$ matrix ${\mathcal{F}}$, and ${\vec{\mathcal{\sigma}}} = ({\mathcal{\sigma}_x}, {\mathcal{\sigma}_y},{\mathcal{\sigma}_z})$ is the vector with the Pauli matrices. $${\mathcal{\sigma}_x}=
\begin{bmatrix}
0 & 1 \\
1 & 0
\end{bmatrix}
, \quad
{\mathcal{\sigma}_y}=
\begin{bmatrix}
0 & -i \\
i & 0
\end{bmatrix}, \quad
{\mathcal{\sigma}_z}=
\begin{bmatrix}
1 & 0 \\
0 & -1
\end{bmatrix}.\nonumber$$
Now, the two quantum observable operators ${\mathcal{Q}_A}, {\mathcal{R}_B}$ can be also expressed using the Pauli matrices up to a scale factor as $${\mathcal{Q}_A} = {1\over 2}({\bf{1}_A} + {\vec{x}}\cdot{\vec{\mathcal{\sigma}}}), \quad {\mathcal{R}_B} = {1\over 2}({\bf{1}_B} + {\vec{y}}\cdot{\vec{\mathcal{\sigma}}})
\label{obs}$$ where ${\vec{x}},{\vec{y}}$ are three dimensional real vectors.
Then, for operators $\mathcal{X} = {\mathcal{Q}_A}\otimes{\bf{1}}_B$ and $\mathcal{Y} = {\bf{1}}_A\otimes\mathcal{R}_B$, we can calculate the expectation values as follows; $$\langle \mathcal{X}\mathcal{Y} \rangle = {\mathrm{Tr_{AB}}}[\rho_{AB}\mathcal{X}\mathcal{Y}]={1\over 4}(1+
{\vec{a}}\cdot{\vec{x}} + {\vec{b}}\cdot{\vec{y}} + {\vec{x}}\cdot{\mathcal{F}}\cdot{\vec{y}}),\nonumber$$ and $$\langle \mathcal{X} \rangle ={1\over 2}(1 + {\vec{a}}\cdot{\vec{x}}),\quad
\langle \mathcal{Y} \rangle ={1\over 2}(1 + {\vec{b}}\cdot{\vec{y}}). \nonumber$$ This leads to $$\langle \mathcal{X}\mathcal{Y} \rangle - \langle \mathcal{X} \rangle\langle \mathcal{Y} \rangle = {1\over 4}({\vec{x}}\cdot{\mathcal{F}}\cdot{\vec{y}} - ({\vec{a}}\cdot{\vec{x}})({\vec{b}}\cdot{\vec{y}})) = {1\over 4}{\vec{x}}\cdot({\mathcal{F}} - {\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}})\cdot{\vec{y}},$$ where ${\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}}$ is the outer product of ${\vec{a}}, {{\vec{b}}}$. By defining a real $3\times3$ matrix $\mathcal{C}$ as $$\mathcal{C} = {\mathcal{F}} - {\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}},
\label{corr}$$ $$c(\mathcal{X}, \mathcal{Y}) = {1\over 4}{\vec{x}}\cdot \mathcal{C} \cdot{\vec{y}}.
\label{corrv}$$ Thus, the nature of the correlation function depends on that of the matrix $\mathcal{C}$. Particularly, we note the following[@cheng], which generalizes our previous analogous results[@ohira1; @ohira2] for any pure bipartite $2\times2$ state vectors.
[**[Theorem 1]{}**]{}: For any bipartite $2\times2$ system (regardless of separable or entangled, or, pure or mixed), there always exits a pair of observables $\mathcal{X}, \mathcal{Y}$ that gives a zero–correlation, $c(\mathcal{X}, \mathcal{Y}) = 0$.
[**[Proof]{}**]{}: From the setup above, zero–correlation is achieved when ${\vec{x}}\cdot \mathcal{C} \cdot{\vec{y}} = 0$. This is simply an orthogonality relation of two real vectors ${\vec{x}}$ and $\mathcal{C} \cdot{\vec{y}}$ in three-dimensional space. Such a pair of ${\vec{x}}, {\vec{y}}$ can always be found for any real $3\times3$ matrix $\mathcal{C}$. (Q.E.D)
This proof shows that a zero–correlation observable pair can not only be found for any $\rho_{AB}$, but also there are uncountably many such pairs. It also turns out that, for pure states, the special nature of the matrix $\mathcal{C}$ allows us to answer our main question of distinguishing entangled states.
Pure States
-----------
For the case of pure states, the density matrix has special properties, which in turn constrain the nature of the matrix $\mathcal{C}$.
[**[Theorem 2]{}**]{}: For a pure state, the rank of the $3\times3$ matrix $\mathcal{C}$ is given as follows
\(a) For a separable state $rank(\mathcal{C}) = 0$, that is $\mathcal{C}$ is the zero matrix.
\(b) For an entangled state $rank(\mathcal{C}) = 3$, that is $\mathcal{C}$ is a regular matrix.
[**[Proof]{}**]{}: First, we list some known properties[@jchen] when the density matrix in (\[density\]) describes a pure state.
\(i) ${\mathcal{F}}\cdot \vec{b} = \vec{a}, \quad {\mathcal{F}}^{\hspace{0.05cm} \mathsf{T}}\cdot \vec{a} = \vec{b}$.
\(ii) $0 \leq \begin{Vmatrix} \vec{a} \end{Vmatrix} = \begin{Vmatrix} \vec{b} \end{Vmatrix} \leq 1$ ($\vec{a}$ and $\vec{b}$ has the same length).
\(iii) $Det({\mathcal{F}}) = \begin{Vmatrix} \vec{a} \end{Vmatrix}^2 - 1 = \begin{Vmatrix} \vec{b} \end{Vmatrix}^2 -1$.
(iv)The density matrix describes a pure and separable state if and only if $\begin{Vmatrix} \vec{a} \end{Vmatrix} = \begin{Vmatrix}\vec{b} \end{Vmatrix} = 1$.
\(a) We note that any density matrix describing a pure separable state can be written as $$\rho^{sep}_{AB} = {1\over 2}({\bf{1}_A} + {\vec{a}}\cdot{\vec{\mathcal{\sigma}}})\otimes {1\over 2}({\bf{1}_B} + {\vec{b}}\cdot{\vec{\mathcal{\sigma}}})
\label{sep}$$ By expanding this, we can immediately see with (\[density\]) that $${\mathcal{F}} = {\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}}.$$ Hence, by the definition of the correlation matrix (\[corr\]), $$\mathcal{C} = {\mathcal{F}} - {\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}} = \mathcal{O}.$$
\(b) We consider two possible cases.
For ${\vec{a}} = {\vec{b}} = {\vec{0}}$: By the definition (\[corr\]) of the correlation matrix $\mathcal{C} = \mathcal{F}$, and therefore by the property (iii), we have $$Det({\mathcal{C}}) = Det({\mathcal{F}}) = - 1$$ This shows that the correlation matrix is regular and thus has $rank(\mathcal{C}) = 3$.
For ${\vec{a}} \neq {\vec{0}}$: By the property (ii), we also have ${\vec{b}} \neq {\vec{0}}$. Then, by (\[corr\]) and the property (i), for any ${\vec{b}} \neq {\vec{0}}$, $$\mathcal{C}\cdot {\vec{b}} = {\mathcal{F}}\cdot {\vec{b}} - (\begin{Vmatrix} \vec{b} \end{Vmatrix}^2){\vec{a}}
= (1 - \begin{Vmatrix} \vec{b} \end{Vmatrix}^2) {\vec{a}}.$$ The right-hand side of this equation is a non-zero vector by the property (iv) for the entangled case. This means that the only solution for the matrix equation $$\mathcal{C}\cdot {\vec{s}} = {\vec{0}}$$ is ${\vec{s}} = {\vec{0}}$, which, in turn, shows that the correlation matrix is regular and thus has $rank(\mathcal{C}) = 3$. (Q.E.D)
This theorem provides an answer to our main question to find the minimum number of correlation measurements for entanglement detection.
[**[Theorem 3]{}**]{}: For pure bipartite $2\times2$ quantum systems, the minimum number of the correlation measurements for entanglement detection with certainty is three.
[**[Proof]{}**]{}: From Eq. (\[corrv\]), and Theorem 2(a), the correlation function $c(\mathcal{X}, \mathcal{Y})$ is $0$ for the separable case for any pair of $\mathcal{X}, \mathcal{Y}$. Now, for the entangled state, let us choose $\mathcal{Y}$ to have any value other than the identity (no measurement). This is equivalent to fixing a real three-dimensional vector $\vec{y}(\neq 0)$. Then, the correlation function is $0$ if and only if the vector $\vec{x}$ defining $\mathcal{X}$ lies in the plane perpendicular to ${\vec{y'}}\equiv \mathcal{C} \cdot{\vec{y}}$ that is a non-zero vector by Theorem 2(b) for the entangled case. Thus, if we prepare thee observables $\mathcal{X}_a, \mathcal{X}_b, \mathcal{X}_c$ defined by linearly independent $\vec{x_a}, \vec{x_b}, \vec{x_c}$, at least one of them gives non-zero correlation with respect to the fixed $\mathcal{Y}$. (The perpendicular plane can only accommodate up to two linearly independent real three-dimensional vectors.) This will identify the entangled states, distinguishing them from separable ones. (Q.E.D)
[**[Examples]{}**]{}: To illustrate the above theorem, we discuss here two examples of pure entangled states.
\(i) The first example is the singlet state, which is also one of the Bell states. $$\ket{\Psi^{-}}= {1\over\sqrt{2}}(\ket{a_1}\otimes\ket{b_2} - \ket{a_2}\otimes\ket{b_1})
\equiv {1\over\sqrt{2}}(\begin{bmatrix}
1 \\
0
\end{bmatrix}_A \otimes
\begin{bmatrix}
0 \\
1
\end{bmatrix}_B
-
\begin{bmatrix}
0 \\
1
\end{bmatrix}_A
\otimes
\begin{bmatrix}
1 \\
0
\end{bmatrix}_B).
\label{bell2}$$
The corresponding density matrix is given as follows in the notation above: $$\rho_{AB} ={1\over 4}({\bf{1}_A}\otimes {\bf{1}_B} + \sum_{i} (-1){\mathcal{\sigma}}_i\otimes {\mathcal{\sigma}}_i)
= {1\over 2}\begin{bmatrix}
0,& 0,& 0,& 0 \\
0,& 1,& -1,& 0 \\
0,& -1,& 1,& 0 \\
0,& 0,& 0,& 0
\end{bmatrix}
\label{density-singlet}$$ Thus, in this case, the correlation matrix is particularly simple as $$\mathcal{C} = \mathcal{F}
= \begin{bmatrix}
-1,& 0,& 0 \\
0,& -1, & 0 \\
0,& 0, & -1
\end{bmatrix}
\label{cmatrix-singlet}$$ Thus, the correlation function (\[corrv\]) is $-{1\over 4}\vec{x}\cdot\vec{y}$, which is zero for any pair of observable $\mathcal{X}, \mathcal{Y}$ defined by a pair of orthogonal vectors $\vec{x}, \vec{y}$. Therefore, for a given $\mathcal{Y}$, it suffices to prepare and measure three observables operators $\mathcal{X}_a, \mathcal{X}_b, \mathcal{X}_c$ defined by three linearly independent vectors $\{\vec{x_a}, \vec{x_b}, \vec{x_c}\}$: at least one of them will give a non-zero value, showing that this is an entangled state.
\(ii) The second example is from [@jchen]. The entangled state vector is given as $$\begin{aligned}
\ket{\Phi}&=&{1\over\sqrt{3}}(\ket{a_1}\otimes\ket{b_1} +\ket{a_1}\otimes\ket{b_2}+ \ket{a_2}\otimes\ket{b_2})\nonumber\\
&\equiv& {1\over\sqrt{3}}(\begin{bmatrix}
1 \\
0
\end{bmatrix}_A \otimes
\begin{bmatrix}
1 \\
0
\end{bmatrix}_B
+
\begin{bmatrix}
1 \\
0
\end{bmatrix}_A
\otimes
\begin{bmatrix}
0 \\
1
\end{bmatrix}_B
+
\begin{bmatrix}
0 \\
1
\end{bmatrix}_A
\otimes
\begin{bmatrix}
0 \\
1
\end{bmatrix}_B).
\label{chen}\end{aligned}$$ The corresponding density matrix is given as $$\rho_{AB}=\ket{\Phi}\bra{\Phi} = {1\over 9}\begin{bmatrix}
1,& 2,& 0,& 2 \\
2,& 4,& 0,& 4 \\
0,& 0,& 0,& 0 \\
2,& 4,& 0,& 4
\end{bmatrix}
\label{density-then}$$ This density matrix can be written in the form of (\[density\]) with $${\vec{a}}={1\over 3}
\begin{bmatrix}
2 \\
0 \\
1
\end{bmatrix}
, \quad
{\vec{b}}={1\over 3}
\begin{bmatrix}
2 \\
0 \\
-1
\end{bmatrix}, \quad
{\mathcal{F}} =
\begin{bmatrix}
2 & 0 & -2\\
0 & -2 & 0 \\
2 & 0 & 1
\end{bmatrix}.\nonumber$$ This yields the correlation matrix as $$\mathcal{C} = {\mathcal{F}} - {\vec{a}}\cdot{{\vec{b}^{\hspace{0.05cm} \mathsf{T}}}}
= {2\over 9}
\begin{bmatrix}
1 & 0 & -2\\
0 & -3 & 0 \\
2 & 0 & 2
\end{bmatrix}.$$ If we set $${\vec{y}}=
\begin{bmatrix}
1 \\
0 \\
0
\end{bmatrix},\quad
{\vec{y'}} = \mathcal{C} \cdot{\vec{y}}={2\over 9}
\begin{bmatrix}
1 \\
0 \\
2
\end{bmatrix}.$$ Thus, the correlation function is zero for any operator $\mathcal{X}$ defined by a vector ${\vec{x}}$ that resides on the plane spanned by $$\{
\begin{bmatrix}
2 \\
0 \\
-1
\end{bmatrix},
\begin{bmatrix}
0 \\
1 \\
0
\end{bmatrix}
\}$$ Thus, again, as we prepare three observable operators defined by three linearly independent vectors $\{\vec{x_a}, \vec{x_b}, \vec{x_c}\}$, we can be sure that at least one of them will yield a non-zero correlation.
Mixed State
-----------
As has been known, if the system is in a mixed state, the situation is more complex. It is also the case for our approach here. We illustrate this by considering a concrete example of the Werner state[@werner].
The Werner state for bipartite $2\times2$ quantum systems is described by the following density matrix $$\rho^{W}_{AB}={{1-\xi}\over 4}{\bf{1}_A}\otimes {\bf{1}_B} + \xi\ket{\Psi^{-}}\bra{\Psi^{-}}
= {1\over 4}\begin{bmatrix}
1-\xi,& 0,& 0,& 0 \\
0,& 1+\xi,& -2\xi,& 0 \\
0,& -2\xi,& 1+\xi,& 0 \\
0,& 0,& 0,& 1-\xi
\end{bmatrix},$$ where $\ket{\Psi^{-}}$ is the singlet state (\[bell2\]), to which it reduces with $\xi=1$. Thus, it can be viewed as a generalization of the singlet state. It can be rewritten as follows $$\rho^{W}_{AB}={{1}\over 4}({\bf{1}_A}\otimes {\bf{1}_B} +(-\xi) \sum_{i}{\mathcal{\sigma}}_i\otimes {\mathcal{\sigma}}_i),
\label{mixdensity}$$ which is of the form (\[density\]), with $\vec{a}=\vec{b}=\vec{0}$, $F_{ij}= -\delta_{ij}\xi$. This state has been studied extensively (e.g., [@peres; @hiroshima; @azuma] and known to be separable for $0 \leq \xi \leq {1\over 3}$, and entangled for ${1\over 3} < \xi \leq 1$.
Now, if we apply our procedure, the matrix $\mathcal{C}$ is a simple diagonal matrix, which again reduces to that of the singlet state with $\xi=1$. $$\mathcal{C} = \mathcal{F}
= (-\xi)\begin{bmatrix}
1,& 0,& 0 \\
0,& 1,& 0 \\
0,& 0,& 1
\end{bmatrix}
\label{cmatrix-werner}$$ Hence,the correlation function is also simply given as $$c(\mathcal{X}, \mathcal{Y}) = -{\xi\over 4}{\vec{x}}\cdot {\vec{y}}.$$ This shows that the $c(\mathcal{X}, \mathcal{Y})=0$ if and only if non-zero real three-dimensional vectors are orthogonal regardless of the value of $\xi$. Thus, for the Werner state, the zero-correlation condition is the same both for separable and entangled situations. Zero correlations do not entail separability, nor do non-zero correlations mean entanglements. This is in contrast to the pure state case where zero-correlations can distinguish between the separable and entangled states. In general, for the mixed states, it appears that we need to analyze further the values of the correlation function with various measurements and apply other criteria to detect entanglements.
Discussion
==========
\(i) Generalization
For pure $n \times n$ states, we expect analogous properties to hold based on the natural extension of the expression of the density matrix in higher dimensions. Details, however, are yet to be carefully examined. Hence, we keep it as the following conjecture.
[**[Conjecture]{}**]{}: For a bipartite $n \times n$ pure system, the following holds.
\(1) The analogously defined correlation matrix $C$ has a rank of either $0$ for separable states, or $n^2 -1$ for entangled states.
\(2) The minimum number of the correlation function measurements for entanglement detection with certainty is $n^2 -1$.
For a multi-party system involving more than two, we cannot anticipate any analogous theorem. It is not clear even whether there exist entangled states that never admit zero correlations.
\(ii) Limitations
As we have noted, this approach of correlation measurements particularly based on zero-correlation has not been found to be a useful way to distinguish separability from entanglement for mixed states in general. Even the existence of uncountably many zero-correlations does not certify separability, nor non-zero correlations entail entanglement. It appears that we need to gain more detailed information on the values of the correlation functions, or each element of the density matrix in question, from which more elaborate mathematical tools can be applied for entanglement determination.
\(iii) Connections
As is well known, for a bipartite $2 \times 2$ system, we have inequalities to detect quantum entanglement. For example, the CHSH inequality involves four correlation functions and four observables. The connection to our theorem, which also involves four observables, but three correlation functions, is not clear at this point, but it may be an interesting topic to explore.
\(iv) Prospects
Thus, the concept of correlations as we have set it up above is not enough to capture the nature of quantum entanglements in general. Further, we would like to point out that the concept of entanglement itself may not be enough to describe quantum correlations in general. This is indicated by the recently proposed “quantum pigeonhole effect” by Aharonov et. al. [@aharonov].
Classically the pigeonhole principle states that if we have more pigeons placed in fewer number of holes, at least one hole must have multiple pigeons together. The analogous system is considered in quantum mechanics with three two-state quantum particles (pigeons), each in a superposition of two (hole) states (a quantum $3\times2$ system). Computing correlations with cleverly chosen different pre- and post-selected product (non-entangled) states, surprisingly, show that no pair of particles can be in the same quantum (hole) state. This means that the pigeonhole principle in some cases breaks down in quantum mechanics. It also shows there are new aspects of quantum entanglement which are not apparent in product states. The experimental work with three single photons transmitted through two polarization channels indicates that this quantum pigeonhole effect is real [@chen]. Concepts, which may be a generalization of entanglement, to capture this type of quantum correlations are yet to be explored.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author would like to thank Philip M. Pearle, Professor Emeritus of Hamilton College, for his comments and encouragements. Also, comments by Dr. Shuming Cheng on Theorem 1, and by Profs. Chigaku Itoi and Shinichi Deguchi of Nihon University are acknowledged as very constructive for this work. This work was supported by funding from Ohagi Hospital, Hashimoto, Wakayama, Japan, and by Grant-in-Aid for Scientific Research from Japan Society for the Promotion of Science No.19H01201.
[9]{}
J. S. Bell, Physics, [**1**]{}, 195 (1964).
J.F. Clauser, M.A. Horne, A. Shimony, and R.A. Holt, Phys. Rev. Lett., [**23**]{}, 880 (1969).
A. Aspect, P. Grangier, and G. Roger, Phys. Rev. Lett., [**47**]{}, 460 (1981).
A. J. Leggett, and Anupam Garg, Phys. Rev. Lett., [**54**]{}, 857 (1985).
V. Vedral, M. B. Plenio, M. A. Pippin, and P. L. Knight, Phys. Rev. Lett., [**78**]{}, 2257 (1997).
J. B. Altepeter, E. R. Jeffrey, P. G. Kwiat, S. Tanizilli, N. Gisin, and A. Acin, Phys. Rev. Lett., [**95**]{}, 033601 (2005).
H. Sakai, T. Saito, T. Ikeda, K. Itoh, T. Kawahata, H. Kuboki, Y. Maeda, N. Matsui, C.Rangacharyulu, M. Sasano, Y. Satou, K. Sekiguchi, K. Suda, A. Tamio, T. Uesaka, and K. Yako, Phys. Rev. Lett., [**97**]{}, 150405 (2006).
K. Fujikawa, C. H. Oh, K. Umetsu and S. Yu, Annals of Physics, [**368**]{}, 248 (2016).
R. F. Werner, Phys. Rev. A, [**40**]{}, 4277 (1989).
A. Peres, Phys. Rev. Lett., [**77**]{}, 1413 (1996).
M. Horodecki, P. Horodecki, and R. Horodecki, Rev. Lett. A, [**223**]{}, 1 (1996).
R. Horodecki, P. Horodecki, M. Horodecki, and K. Horodecki, Rev. Mod. Phys., [**81**]{}, 865 (2009).
O. G[ü]{}hne, G. T[ú]{}th, Physics Reports, [**474**]{}, 1 (2009).
W. K. Wootters, Phys. Rev. Lett., [**80**]{}, 2245 (1998).
H. J. Kummer, Int. J. of Theor. Phys., [**40**]{}, 1071 (2001).
A. F. Abouraddy, R. E. A. Saleh, A. V. Sergienko, and M. C. Fesch, Phys. Rev. A, [**64**]{}, 050101(R) (2001).
J. L Chen, L. Fu, A. A. Ungar, and X, G Zhao, Phys. Rev. A, [**65**]{}, 044303 (2002).
S. Cheng (Private communication).
T. Ohira, Prog. Theor. Phys., [**2018-8**]{}, 083A02 (2018).
T. Ohira, Prog. Theor. Phys., [**2020-1**]{}, 013A01 (2020).
T. Hiroshima and S. Ishizaka, Phys. Rev. A, [**62**]{}, 044302 (2000).
H. Azuma and M. Ban, Phys. Rev. A, [**73**]{}, 032315 (2006).
Y. Aharonov, F. Colombo, S. Popescu, I. Sabadini, D. C. Struppa and J. Tollaksen, Proc. Natl. Acad. Sci. USA, [**113**]{}, 532 (2016).
M. Chen, C. Liu, Y. Luo, H. Huang, B. Wang, X. Wang, L. Li, N. Liu, and C. Lu and J. Pan, Proc. Natl. Acad. Sci. USA, [**116**]{}, 1549 (2019).
[^1]: The author is also affiliated with Future Value Creation Research Center, Graduate School of Informatics, Nagoya University, and with Mathematical Science Team, RIKEN Center for Advanced Intelligence Project.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An approach using topology reveals a new understanding and knowledge of laser particle acceleration. Laser pulse irradiation on a thin-foil target is examined using two-dimensional particle-in-cell simulations. Through topology, laser ion acceleration schemes are classified depending on whether the deformed target shape after laser irradiation is homeomorphic with respect to the initial target. High energy ions are generated with high-efficiency when the deformed target shape is non-homeomorphic. In this case, the target thickness greatly affects the obtained ion energy. However, when the deformed target shape is homeomorphic, the obtained ion energy is low, and the effect of the target thickness is small.'
author:
- Toshimasa Morita
title: Topological approach for laser acceleration investigation
---
Introduction
============
Recently, significant progress has been achieved in compact laser systems, and laser ion acceleration is a compelling application of high-power compact lasers [@BWE; @DNP]. If a compact laser system is able to generate ions with sufficiently high energy and quality, low-cost compact accelerators would become feasible. However, the ion energies achieved thus far are insufficient for some applications, such as hadron therapy [@CLK; @SNV]. In addition, such applications require the ion beam to have a narrow energy spread, high quality, and a sufficient number of ions [@ROT; @ESI; @BEE]. Therefore, it is important to study the best conditions for generating ions with higher energies and of higher quality [@BWP; @DL; @HSM; @PPM; @PRK; @SPJ; @Toncian; @TM1; @TM2]. Moreover, in laser ion accelerator technologies, the required energy and number of ions vary depending on the application, i.e., it is necessary to establish a method to generate specified particular amount of energy and number of ions through laser ion acceleration. Therefore, what is important for laser acceleration investigation from now on is research on methods to control the generated ion energy and number. To this end, we need to understand the laser particle acceleration phenomenon in sufficient detail. In this paper, we have added a new understanding to laser acceleration using the fundaments of topology.
Topology is used in almost all the fields of mathematics. It is one of the fundamental theories that support not only the field of geometry but also general mathematics. Additionally, it is one of the fundamental theories of physics and engineering, which are based on mathematics. In this paper, we study the laser particle acceleration phenomenon from the viewpoint of topology.
In Sec. \[topo\], how to treat the laser acceleration phenomena with topology is shown. In Secs. \[h-para\] and \[h-resu\], we present some considerations of laser particle acceleration using the hydrogen foil target. In Sec. \[ch2\], the polyethylene foil target case is shown. In the final section, the main results of our study are summarized.
Topology and laser acceleration {#topo}
===============================
First, we consider the target before laser irradiation, and the “target” after laser irradiation which is precisely composed a cloud of ions and electrons. This can be regarded as a mapping $f$ from the target before laser irradiation, $X$, to that after laser irradiation, $Y$, (Fig.\[fig:fig-mp\](a)). $$f:X \rightarrow Y.
\label{map}$$ It is assumed that no nuclear reaction occurs; therefore, an ion and electron do not disappear or split into two or generate. Therefore,
1. The ions and electrons in the target before laser irradiation are always present somewhere in the target after irradiation; i.e., one-to-one mapping $=$ injection, $$\forall x,x' \in X, x \neq x' \Rightarrow f(x) \neq f(x').
\label{bijec}$$
2. All the ions and electrons present in the target after laser irradiation come from the target before irradiation; i.e., onto mapping $=$ surjection, $$f(X)=Y.
\label{srjec}$$
Therefore, the laser particle acceleration process is understood as a bijective, that is injective and surjective, mapping $f$.
![ (a) The laser particle acceleration process is understood as a bijective mapping $f:X \rightarrow Y$. (b) $A$, $B$, and $C$ are homeomorphic ($A\approx B\approx C$). $D$, $E$, and $F$ are non-homeomorphic to $A$ and to each other ($A \not\approx D \not\approx E \not\approx F$). []{data-label="fig:fig-mp"}](fig-1.pdf){width="9.0cm"}
In topology, if the mapping $f$ is bijection and continuous, and $f^{-1}$ is also continuous, $X$ and $Y$ are called topologically isomorphic (homeomorphic) and are denoted by $X \approx Y$. $f$ is a continuous map, which means that the mapping $f$ is continuous at each point of $X$. $f$ is continuous at point $x_0$ is written as $$\forall \epsilon >0, \exists\delta >0 \mid d(x,x_0)<\delta \Rightarrow d(f(x),f(x_0))<\epsilon,
\label{conti}$$ where $d(x, x_0)$ is the distance between $x$ and $x_0$. This implies that $f$ does not leap at $x_0$ with a slight $x$ change. It means that any continuous curve is transferred to a continuous curve. Or, it can be defined using $\epsilon$-neighborhood at point $x_0$ of $X$, $B(x_0;\epsilon)=\{x \in X \mid d(x,x_0) < \epsilon\}$. If $$\forall \epsilon >0, \exists\delta >0 \mid f(B(x_0;\delta)) \subset B(f(x_0);\epsilon)
\label{neigh}$$ is satisfied at point $x_0 \in X$, $f$ is continuous at $x_0$. This implies that the mapping, which transfers nearby objects to the ones that are still nearby without cutting or overlapping, is homeomorphism.
Figures $A, B$, and $C$ in Fig.\[fig:fig-mp\](b) are homeomorphic, $A\approx B\approx C$. On the other hand, figures $D, E$, and $F$ are non-homeomorphic to $A$ and non-homeomorphic to each other, $A \not\approx D \not\approx E \not\approx F$.
In this paper, we recognize the target deforming in the acceleration process as shape changes, i.e., we see the target as a figure in the topology. The correspondence from the target shape before deformation (the initial target) to the deformed shape by laser irradiation (the deformed target) is regarded as a mapping. Determining whether two shapes are homeomorphic is a basic problem of topology. In this paper, we consider laser particle acceleration by focusing on whether the target shapes before and after laser irradiation is homeomorphic.
In the acceleration process, when an ion does not pass among the surrounding ions and its positional relationship with the surrounding ions does not differ, i.e., if an ion near other ions in the initial target is still located near those ions after laser irradiation, the target deformation is a continuous map and the targets before and after laser irradiation are homeomorphic, $X \approx Y$. However, they are not homeomorphic, $X \not\approx Y$, in the acceleration process such that some ions pass through other area ions. Therefore, it is possible to see that the acceleration process is different $\Rightarrow$ it is different in topology, i.e., it is possible to know and understand the difference of the acceleration scheme from the difference of the topology of the target. Additionally, ions that are located in a narrow area, $U(x;\delta)$, within the initial target, become ions with almost the same energy, i.e., the number and energy of the obtained ions are considered to be functions of the coordinate $x$ in the initial target, and by studying its mapping, it is possible to acquire in-depth understanding of the laser acceleration phenomenon.
Simulation parameters {#h-para}
=====================
The simulations were performed with a parallelized electromagnetic code based on the PIC method [@CBL]. The parameters used in the simulations of the hydrogen foil target are shown below. An idealized model, in which a Gaussian linear polarized laser pulse is normally incident on a foil target represented by a collisionless plasma, is used. The electron density, i.e., proton density, is $n_{e}=5\times 10^{22}$ cm$^{-3}$. The foil target has a thickness $1.0, 2.0,$ and $3.0 \mu$m in each case. The total number of quasiparticles is $5\times 10^{8}$ in the case of $1.0 \mu$m thickness, and it is two times for the $2.0 \mu$m thickness case and three times for the $3.0 \mu$m case. The number of grid cells is equal to $24000 \times 24000$ along the $X$ and $Y$ axes, respectively. Correspondingly, the simulation box size is $112\mu$m$\times 112\mu$m. The laser propagation direction is set to be along the $X$ direction, and the electric field is oriented in the $Y$ direction. The boundary conditions for particles and fields are periodic in transverse $Y$ directions and absorbing at the boundaries of the computation box along the $X$ axis. The laser-irradiated side surface of the foil is placed at $X=44 \mu$m, and the center of the laser pulse is located $20 \mu$m behind it. A $xy-$coordinate system is used throughout the text and figures. The origin of the coordinate system is located at the center of the laser-irradiated surface of the initial target, and the directions of the $x$ and $y$ axes are same as those of the $X$ and $Y$ axes, respectively. Therefore, the $x$ axis denotes the direction perpendicular to the target surface, and the $y$ axes lie in parallel to the target surface.
We are particularly interested in the intensity $I \sim 1\times 10^{22}$ W/cm$^{2}$, laser power $P \sim 1$ PW, pulse duration $\sim 30$ fs full width at half maximum (FWHM), and laser energy $ \mathcal{E}_\mathrm{las}\sim 20$ J, which has been achieved recently in compact lasers [@Kiri]. Therefore, the intensity is varied around $I=1\times 10^{22}$ W/cm$^{2}$. Herein, the shape of the laser pulse, which is determined by pulse duration and spot size, is the same in all cases. When the peak intensity is $I$, the distribution of the laser intensity $I^*$ is represented by $I^*(y,z,t)=I \cdot \psi (y, z, t)$. We changed $I$, although the function $\psi(y, z, t)$ which indicates the shape of laser pulses is the same in all cases. The laser energy is $$\mathcal{E}_\mathrm{las} = \int_{-\infty}^{\infty}\int_{-\infty}^{\infty}
\int_{-\infty}^{\infty} I^*(y,z,t)dydzdt \nonumber \\
= I\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}
\psi(y,z,t)dydzdt=I\cdot V^*,
\label{els}$$ where $V^*=\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty} \psi(y,z,t)dydzdt$. $V^*$ is the “volume” of the distribution function $\psi$ in the $y, z, t$ space and is the same in all cases under our conditions. Therefore, the laser energy, $\mathcal{E}_\mathrm{las}$, changes linearly with changes in laser intensity, $I$. In the following, if there is an expression of change in the laser intensity, it also implies change in laser energy. Moreover, the same is true for laser power. In this paper, we show a topological consideration for the target deformation, especially noticing the deformation in the target thickness direction. Therefore, in the simulation, the target needs to be sufficiently finely divided in term of thickness. We use a foil target consisting of hydrogen, since foil is the simplest, and hydrogen is the simplest and has the highest possibility of generating high energy ion [@TM1]. Foil thicknesses take $1-3\mu$m, which is possible to generate high energy protons in the above laser conditions [@TM1]. Therefore, $1 \mu$m length, which is the thinnest thickness, must be sufficiently finely divided for the simulation. Also, a large simulation area is required, since hydrogen occur a strong Coulomb explosion and is distributed in a wide area. That is, the number of grid cells needs to be very large. Therefore, we performed two-dimensional (2D) simulations.
The laser pulse has $30$ fs in duration and focused on a spot size of $2.5 \mu$m (FWHM) in all cases. The laser peak intensity is variable, $I=6.4 \times 10^{19} - 2.3 \times 10^{22}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las} = 0.1-53$ J). At $I=1\times 10^{22}$ W/cm$^{2}$, the laser peak power, $P$, is $783$ TW and the laser energy, $\mathcal{E}_\mathrm{las}$, is $25$ J, and the laser pulse with dimensionless amplitude $a_0=q_eE/m_{e}\omega c=72$.
Simulation results {#h-resu}
==================
Simulation results using a hydrogen foil target are shown. The change of the obtained proton energy and spatial distribution of protons, electrons, are described by the intensity change, i.e., energy and power also change, upon a fixed laser pulse shape, i.e., spot size and pulse duration. Three different foil thickness ware analyzed.
![ Spatial distribution of particles (protons and electrons) and electric field magnitude when $I=1 \times 10^{22}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las}=25$ J) and the target thickness $\ell=1\mu$m. The time when the center of the laser pulse reaches the surface$^-$ of the target is assumed to be $t = 0$. For protons, the color corresponds to their energy. []{data-label="fig:fig-te"}](fig-2.pdf){width="8.0cm"}
Figure \[fig:fig-te\] shows the particle distribution and the electric field magnitude when a laser pulse of $I=1 \times 10^{22}$ W/cm$^2$, $\mathcal{E}_\mathrm{las}=25$ J, is normally incident on the hydrogen foil target of thickness $\ell = 1 \mu$m. This is the laser parameter we are most interested in, given the fact this intensity and energy are currently realized and used in actual applications development. Here, the time when the center of the laser pulse, where $I$ is the strongest, reaches the position of the laser-irradiated surface of the initial target is assumed to be $t = 0$. This is because, the position of the laser pulse can be immediately recognized from the time $t$, since we consider while laser pulse and target are interacting in this paper. That is, more than half of the laser pulse remains without interaction with the target when $t<0$, and more than half of interaction have been done at $t \geq 0$. The simulation start time is $t=-67$ fs. The initial shape of the laser pulse and the target is shown at $t=-67$ fs in Fig. \[fig:fig-te\]. The laser pulse is defined on the $-x$ side of the target, and it propagates in the $+x$ direction. Hereinafter, the laser-irradiated surface of the target, $-x$ side surface, is referred to as surface$^-$ and the opposite side surface, the $+x$ side surface, is called surface$^+$. At $t=-17$ fs, the head part of the laser pulse interacts with the target, and mainly the electrons are pushed out in the $+x$ direction. At this time, the center of the laser pulse is located at $x=-5 \mu$m, which is behind the surface$^-$, and many parts of the laser pulse have still not interacted with the target. At $t=33$ fs, part of the laser pulse is reflected by the target while another part is transmitted the target. If there is no interaction with the target, the center position of the laser pulse is at $x=10 \mu$m, and it has already passed the initial target position. Therefore, the strong interaction between the laser and the target has almost finished. As $t=83, 133, 183$ fs, the target expands greatly with time. The maximum proton energy at $t=183$ fs is $330$ MeV, and these appear on the outermost part of the $+x$ side of the expanded target. Proton energy gradually decreases as it goes from there in the $-x$ direction up to $x>0$. The distribution shape of protons is spherical in $+x$ side area and mountainous-shaped in $-x$ side area. The farthest position of the proton on the $+x$ side is $36 \mu$m, and on the $-x$ side is $-20 \mu$m. The $+x$ side expansion is 1.8 times longer than that of the $-x$ side, i.e., the target expands greatly in the laser propagating direction. Note that it is necessary to pay attention to the fact that the results shown in this paper are for 2D simulation, and the obtained ion energies are evaluated higher than in 3D simulation [@TM3].
![ Spatial distribution of protons in the cases of $I=1 \times 10^{22}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las}=25$ J) (a), and $I=6 \times 10^{19}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las}=0.1$ J) (b). The color corresponds to their initial positions. The initial region is determined by slicing the foil into five regions (see zoomed figure at $t=0$ in (a)), and each region is numbered with the number shown in the figure. (a) The ion distributions are non-homeomorphic, and (b) these are homeomorphic. []{data-label="fig:fig-ti"}](fig-3.pdf){width="8.0cm"}
In Fig. \[fig:fig-ti\](a), the same result as in Fig. \[fig:fig-te\] is shown with a different way. However, only protons are shown here. Protons are color-coded by their initial position (see zoomed target at $t=-67$ fs in Fig. \[fig:fig-ti\](a)). The initial foil is sliced evenly into five in the thickness direction, $x$ direction, and each region of protons is of a different color. This makes it possible to know where the protons in a deformed target come from on the initial target. Each region is called region-1, region-2,.., region-5, sequentially from the $-x$ side. As shown at the final time $t=183$ fs, we can see that the protons near the $+x$ side outermost part come from near the opposite side, surface$^-$. That is, the target before laser irradiation and the target after irradiation are non-homeomorphic. Furthermore, by looking at this together with Fig. \[fig:fig-te\], it can be seen that the high energy protons are the protons near the laser-irradiated surface, surface$^-$. Also, at an early time, $t=33$ fs, the protons near the surface$^-$ have already appeared on the $+x$ side target outermost part, indicating that its distribution is formed at the early stage of the acceleration process.
Figure \[fig:fig-ti\](b) shows the result of $I=6 \times 10^{19}$ W/cm$^2$, $\mathcal{E}_\mathrm{las}=0.1$ J, in the same way as in Fig. \[fig:fig-ti\](a). The times are also the same as in Fig. \[fig:fig-ti\](a). The figures in the dashed-and-dotted frame are enlarged views including the center of the target. At weak laser intensity, i.e., power and energy, the protons after laser irradiation do not distribute like those in Fig. \[fig:fig-ti\](a). During the acceleration process each region is distributed in the order of the initial target at all times, without the protons near the surface$^-$ going out to the region near surface$^+$. That is, both the target before laser irradiation and the target after irradiation are homeomorphic. And, in Figs. \[fig:fig-ti\](a) and \[fig:fig-ti\](b) with different laser intensities, energies, the topology of the target after laser irradiation is different. Here, we define the terms used in this paper. When the deformed target by laser irradiation is homeomorphic with the initial target, we call this acceleration ‘homeomorphic acceleration’ (HA). Since the initial target and the targets after laser irradiation in Fig. \[fig:fig-ti\](b) are homeomorphic, this is HA. On the contrary, when the deformed target by laser irradiation is non-homeomorphic with the initial target, as shown in Fig. \[fig:fig-ti\](a), we call that acceleration ‘non-homeomorphic acceleration’ (NHA). Also, when the target after deformation is homeomorphic with the initial target, the distribution after acceleration is called homeomorphic distribution (H distribution). When it is non-homeomorphic, this is called non-homeomorphic distribution (NH distribution). When the positions of the ions (protons) change vigorously and the ions (protons) of region-1, the red part, are widely distributed in the $+x$ side farthest area of the target, like $t=183$ fs in Fig. \[fig:fig-ti\](a), we call this strong NHA. When it is NHA, but not change as vigorously, we call this weak NHA. In addition, at $t=183$ fs in Fig. \[fig:fig-ti\](b), looking closely at the center part of the target around $x=y=0$, the protons near the surface$^-$ are slightly existed in the region-5 protons, where around $x=2 \mu$m, although the number is very small. We look at this from a wider point of view and express it as HA. Depending on the degree, it is expressed as almost HA.
![ The energy spectrum of the proton beam in the case of Fig. \[fig:fig-te\] at $t=183$ fs. Hydrogens of region-1, presenting a laser irradiation surface, become the highest energy, and next highest energy region is region-2. []{data-label="fig:fig-es"}](fig-4.pdf){width="8.0cm"}
Figure \[fig:fig-es\] shows the energy spectra of the protons at $t=183$ fs in the case shown in Fig. \[fig:fig-te\] (also Fig. \[fig:fig-ti\](a)). Since the protons that we use for some applications are protons accelerated in the $+x$ direction, i.e., laser propagated direction, here we show only the result of protons accelerated in the $+x$ direction. The energy spectra for all the protons, represented by a dotted line, and each region of protons by a solid color-coded line for each region, are shown. All the high energy protons are from region-1 which has the laser irradiation surface, i.e., surface$^-$, and next highest energy region is region-2 which is the next region to it. The difference between the maximum proton energies of these two regions is quite large, at about $100$ MeV. That is, the high energy protons all come from near surface$^-$. The maximum proton energy of region-1 is prominently high. The differences in the maximum proton energies of regions 2–5 is not very different in comparison with that between them and in region-1.
![ Maximum proton energy, $\mathcal{E}_{p}$, in each foil thickness as a function of square root of the laser intensity, $\sqrt{I}$, (also $\sqrt{\mathcal{E}_\mathrm{las}}$). The maximum proton energy rises in proportion to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, and is not greatly related to the thickness in low $I$. In the region above a certain $I$, it rises in the second-order to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, and greatly affects the thickness. []{data-label="fig:fig-ip"}](fig-5.pdf){width="8.0cm"}
Figure \[fig:fig-ip\] shows the maximum proton energy at each laser intensity $I$, and also laser energy $\mathcal{E}_\mathrm{las}$. The results of three foil thickness cases of $1, 2, 3 \mu$m are shown. The laser intensity, $I$, on the bottom horizontal axis is indicated by its square root, $\sqrt{ I}$. The upper horizontal axis shows the laser energy, $\mathcal{E}_\mathrm{las}$. The inset shows the zoom of $0 \leq \sqrt{I} \leq 0.5 \times 10^{11} \sqrt{\mathrm{W/cm}^2}$. The stronger $I$, $\mathcal{E}_\mathrm{las}$, the higher the energy of the protons are generated. At low $I$, $\mathcal{E}_\mathrm{las}$, the maximum proton energy rises in proportion to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$. The dotted lines are the linear fitting of the part where the proton energy is proportional to $\sqrt{I}$. The points used for the linear fitting are the values of $\sqrt{I} \leq 0.24 \times 10^{11} \sqrt{\mathrm{W/cm}^2}$ for $1 \mu$m thickness, $\sqrt{I} \leq 0.56 \times 10^{11} \sqrt{\mathrm{W/cm}^2}$ for $2 \mu$m thickness, and $\sqrt{I} \leq 0.72 \times 10^{11} \sqrt{\mathrm{W/cm}^2}$ for $3 \mu$m thickness. The proportionality continues with the stronger $I$ as the foil thickness increases. Moreover, the obtained proton energy is not greatly related to foil thickness in the proportional region. The proton energy in this region is almost same at each thickness, although the foil thickness is up to three times different. For example, at a point of intensity indicated by $\textcircled{\scriptsize A}$, $\sqrt{I}=0.24 \times 10^{11} \sqrt{\mathrm{W/cm}^2}$, in the figure, the proton energies, which are generated from $1, 2$ and $3 \mu$m thickness are $31, 28$, and $27$ MeV, respectively, and are approximately the same. As $\sqrt{I}$ increases, the rise of proton energy changes from a linear to a second-order steep rise curve. The solid lines in the figure are the second-order fitting lines with the proton energy. The proton energy rises in second-order with $\sqrt{I}$, that is, linearly with $I$. This second-order rise starts earlier with a thinner thickness of foil. In this region, the obtained proton energy differs greatly depending on the difference in foil thickness. For example, in the proton energy at point $\textcircled{\scriptsize C}$, the thickness of $1$ and $2 \mu$m are in the second-order increasing region and the thickness of $3 \mu$m is in the linearly increasing region. The proton energies are $240, 133, 93$ MeV for $1, 2$, and $3 \mu$m, respectively. The maximum difference is $2.6$ times. It can be said that, the obtained proton energy greatly different due to the difference in the target thickness in the region above a certain intensity. On the other hand, the target thickness does not significantly influence the obtained proton energy in the region of low laser intensity.
![ The distributions of protons for each target thickness and for each laser intensity (energy). The protons are color-coded by their initial positions. The white parts of the background are NHA, and the light blue parts are HA or almost HA. Where the proton energy rises in the second-order of $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, they are the NHA regime. []{data-label="fig:fig-ii"}](fig-6.pdf){width="8.0cm"}
The distributions of protons at $t = 183$ fs for each target thickness of $1, 2, 3 \mu$m and for the laser intensity of each $\textcircled{\scriptsize A}$, $\textcircled{\scriptsize B}$, $\textcircled{\scriptsize C}$, and $\textcircled{\scriptsize D}$, are shown in Fig. \[fig:fig-ii\]. The protons are color-coded by their initial positions, as in Fig. \[fig:fig-ti\]. The $1 \mu$m thickness case of the laser intensity $\textcircled{\scriptsize D}$ is the same as $t=183$ fs in Fig. \[fig:fig-ti\](a). These are almost HA at all thicknesses in the laser $\textcircled{\scriptsize A}$, and are all strong NHA in the laser $\textcircled{\scriptsize D}$. We can see that the points where proton energy raise in a linear manner with $I$ (second-order with $\sqrt{I}$), i.e., $1 \mu$m-$\textcircled{\scriptsize B}$ $\textcircled{\scriptsize C}$, $2 \mu$m-$\textcircled{\scriptsize C}$ $\textcircled{\scriptsize D}$, and $3 \mu$m-$\textcircled{\scriptsize D}$, are strong NHA, by looking at it together with Fig. \[fig:fig-ip\]. That is, it can be said that the obtained proton energy rapidly rises in proportion to the laser intensity $I$ (laser energy $\mathcal{E}_\mathrm{las}$) when a strong NHA occurs. On the other hand, where the obtained prototype energy is proportional to $\sqrt{I}$. i.e., $1 \mu$m-$\textcircled{\scriptsize A}$, $2 \mu$m-$\textcircled{\scriptsize A}$ $\textcircled{\scriptsize B}$, and $3 \mu$m-$\textcircled{\scriptsize A}$ $\textcircled{\scriptsize B}$ $\textcircled{\scriptsize C}$, these are almost HA.
The proton energy rises in the second-order of $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, in the NHA regime, and rises in proportion to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, in the HA regime. Moreover, the target thickness strongly affects the obtained proton energy in the NHA regime, and the obtained proton energy is not significantly related to the target thickness in the HA regime. From another point of view, we consider the change in proton energy when the target thickness is gradually reduced at a certain laser intensity, energy. When the target is thick, it is the HA regime, and even if the thickness is changed making it slightly thinner, the obtained proton energy does not change much up to a certain thickness. However, when it is less than a certain thickness, it becomes the NHA regime, the produced ion energy becomes rapidly higher, and the ion energy becomes higher as the thickness becomes thinner.
The NHA regime is considered in detail below. The case of thickness $\ell=1 \mu$m, and $I=1 \times 10^{22}$ W/cm$^2$, $\mathcal{E}_\mathrm{las} =25$ J, which is the $\textcircled{\scriptsize D}$ case laser, is used for the considerations. The $1 \mu$m thickness is selected because it generates higher energy protons than the other thicknesses at laser intensities of $\textcircled{\scriptsize A} - \textcircled{\scriptsize D}$. The selection of $I=1 \times 10^{22}$ W/cm$^2$ is due to the fact that it is NHA for all $1, 2$, and $3 \mu$m thickness cases and this intensity has recently been achieved. As shown in Fig. \[fig:fig-ti\](a), we see that the NH distribution has already appeared, i.e., the formation processes has ended, at the initial time $t=33$ fs. Therefore, we investigate in detail the movement and location of electrons and ions at the initial time, $t < 0$ fs. Below, since the contrast between negatively charged particles, i.e., electrons, and positively charged particles, i.e., ions, in the target are important, we generally refer to as ions without writing protons. The laser acceleration phenomenon is shown in the time interval during which the laser pulse and the target interact, that is, in the time about the laser pulse duration. Since the pulse width $=30$ fs, the results over a very short time are shown below.
![ The target and the laser pulse (isosurface for value $a_0=2$) of the $I=1 \times 10^{22}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las}=25$ J), $\ell=1 \mu$m case at early simulation times $t=-67$, $-57$, and $-47$ fs. The center and FWHM points of the laser pulse are indicated by ’$+$’ and ’$\bigcirc$’, respectively. The start time of the interaction between the laser and the target is approximately $t=-47$ fs. []{data-label="fig:fig-ini1"}](fig-7.pdf){width="7.5cm"}
Figure \[fig:fig-ini1\] shows the laser pulse and the target at time $t=-67,-57$, and $-47$ fs. It is shown by zooming the area near the initial laser pulse and the part of the target where the laser irradiates. That is, the $x$ direction is in the range of $x=-35$ to $10 \mu$m and the $y$ direction is in the range of $y \simeq -10$ to $10 \mu$m. In this study, since the interaction between the target and the laser pulse at the initial time is important, the laser pulse is placed relatively far away from the target so that there is no interaction at the simulation start time, $t=-67$ fs. At $t=-57$ fs, the interaction between the laser pulse and the target has scarcely begun to occur yet. At $t=-47$ fs, the center of the laser pulse is at $x=-14 \mu$m and FWHM position on the side closer to the target is at $x=-9.5 \mu$m. The distance from this FWHM position to the surface$^-$ is also about FWHM. At this time, the interaction between the laser pulse and the target has started slightly. Therefore, it is assumed that the time when the interaction starts between the laser pulses and the target is $t \approx -47$ fs, and in the flowing, we investigate within approximately the pulse width $=30$ fs from this start time. That is, the acceleration process from $t=-47$ to $-7$ fs, which is the time just before the center of the laser pulse reaches the target, is considered in detail.
![ (a) Spatial distribution of electrons and ions (protons), and electric field magnitude (isosurface for value $a_0=2$). Ions are color-coded by their initial position. (b) Electrons are color-coded. In the inset: the color corresponds to the ions’ energy. []{data-label="fig:fig-ini_ie"}](fig-8.pdf){width="8.5cm"}
Figures \[fig:fig-ini\_ie\], \[fig:fig-ini\_i\], and \[fig:fig-ini\_e\] show the distributions of ions, i.e., protons, and electrons at the initial time $t \leq -7$ fs. An overview is shown schematically in Fig. \[fig:fig-ini\_ie\] and detailed distributions are shown in Figs. \[fig:fig-ini\_i\], \[fig:fig-ini\_e\].
Figures \[fig:fig-ini\_ie\](a) and (b) are the same results of the distribution of the ions and the electrons, except for their colors. Ions in Fig. \[fig:fig-ini\_ie\](a) and electrons in Fig. \[fig:fig-ini\_ie\](b) are color-coded into five colors, according to their initial positions. The figures at $t=-67$ fs are in the initial state. The laser pulse is also drawn in Fig. \[fig:fig-ini\_ie\](a), the center of it is indicated by ’$+$’, and the $+x$ side FWHM point is indicated by a ’$\bigcirc$’. These figures are zoomed views around $x=y=0$, the $y$ direction displayed area length is $10 \mu$m which is four times the laser spot diameter. First, electrons in the target are pushed out to the $+x$ side of the target by the laser pulse, and then the ions expand gradually. Let us first consider movement and distribution of the ions in Fig. \[fig:fig-ini\_ie\](a). As shown at $t=-17$ fs, when the $+x$ side FWHM point of the laser pulse reaches the surface$^-$, some of the ions of region-1, red part, have already come into region-3, the yellow area. That is, NHA has already started at $t \sim -17$ fs. At $t=-7$ fs, when the $+x$ side FWHM point has passed through the target and the pulse center has not yet been reached, some of the ions of region-1 reach region-5, the light blue area. That is NHA. At this time, region-5 expands greatly, and the other regions move in the $+x$ direction while maintaining almost the initial thickness. The inset in Fig. \[fig:fig-ini\_ie\] shows the ion distribution at $t=-7$ fs, the color corresponds to their energy. By looking at it together with $t=-7$ fs in Fig. \[fig:fig-ini\_ie\](a), we see that some of the ions of region-1, which separated from the other region-1 ions, are positioned near the center of the target and have higher energy than the other ions in the target. The energy of these ions is around $40$ MeV, and that of the other regions ions around them is around $10$ MeV. That is, at this time, some of the region-1 ions form a high energy bunch, and this moves and accelerates in the $+x$ direction and passes through the other regions.
Next, the movement and distribution of the electrons are considered (Fig. \[fig:fig-ini\_ie\](b)). The start time of the movement of the electrons is much earlier than that of the ions, and many electrons appear in the $+x$ side area of the target at $t=-37$ fs. These electrons are of region-1, the red area. The electrons near the surface$^-$ are also accelerated first, in a similar way to the ions, and pass through the electrons in the other regions and exiting on the $+x$ side of the target. Thereafter, the electrons of the other regions are also accelerated in the $+x$ direction, and the electrons of each region are almost evenly mixed, at $t=-7$ fs. The movements of the ions and electrons of each region in the early simulation time are shown in more detail below.
![ (a) Spatial distribution of electrons and ions (protons). Ions are color-coded according to the region of their initial position. (b) Distributions of the number of ions for each initial region in the $x$ direction. The number of ions per unit length along $x$, $n_{i}$, is normalized by its initial value, $n_0$. Some of the reagion-1 ions, red color ions, are passing through regions 2–4 and are going toward the $+x$ side, opposite side, surface. []{data-label="fig:fig-ini_i"}](fig-9.pdf){width="7.5cm"}
First, we show the movement of ions, i.e., protons. Figure \[fig:fig-ini\_i\] shows the spatial distributions of the ions and electrons, and the distributions of the number of ions in the $x$ direction. Figure \[fig:fig-ini\_i\](a) shows the distribution of the ions, color-coded by region, and the distribution of the electrons with a single color, dark blue, and the laser pulse. Although these are the same results as those in Fig. \[fig:fig-ini\_ie\](a), the area around the target center is zoomed more in here. The times are $t =-47$ to $-7$ fs, which are the same times as Fig. \[fig:fig-ini\_ie\]. Figure \[fig:fig-ini\_i\](b) shows the distributions of the number of ions in the $x$ direction, and this is shown in different colors for each region. Here, the ions in the range of $y=-1.25$ to $1.25 \mu$m are counted. The number density of all the ions, which is not divided into regions, is indicated by a dotted black line, and the initial distributions of the ions of each region are indicated by a dotted line according to each color. The density is shown as a ratio to the initial density, $n_i/n_0$. The initial number density of the ions, $n_{i0}$, and that of the electrons, $n_{e0}$, are equal in the hydrogen target, $n_{i0}=n_{e0}=n_0$. Although there is still no great change from the initial state at $t=-47$ fs, a slightly high-density area of region-1 ions are generated near the surface$^-$ (Fig. \[fig:fig-ini\_i\](b)). At $t=-37$ fs, particularly noticeable high-density peaks appear near the surface$^-$ in the ion distribution. This peak of ion density of region-1 moves in the $+x$ direction over time. Then, at $t=-27$ fs, this peak becomes even higher, with the maximum point of density being $3.8$ times the initial density. In the spatial distribution of ions (Fig. \[fig:fig-ini\_i\](a)), the boundaries of each ion region are clear in $t \le -27$ fs, and the ions of each region are scarcely mixed. At $t=-27$ fs, the surface$^-$ moves slightly in the $+x$ direction, and the ions near the surface$^+$ expands slightly to the $+x$ side area of the target. At $t=-17$ fs, some of the ions of region-1 have entered region-3, and the high-density peak of it begins to split into two. At $t=-7$ fs, some of the region-1 ions pass through regions 2–4 and reach the region-5, and region-1 splits into two parts. The peak density of the region-2 ions appears at $t=-17$ fs after the peak of the region-1 becomes smaller. It is region-3 that forms the next peak to this (see $t=-7$ fs). That is, the ions in the target form a high-density peak in order, from the surface$^-$ side to the $+x$ side region, over time. These peaks move in the $+x$ direction. However, the ions of region-5 do not form a high-density peak like the other regions, and expand greatly in the $+x$ direction over time.
![ (a) Spatial distribution of electrons and ions. Electrons are color-coded according to the region of their initial position. (b) Distributions of the number of electrons for each initial region in the $x$ direction. The number of electrons per unit length along $x$, $n_{e}$, is normalized by its initial value, $n_0$. The electrons near the laser irradiation surface are pushed away in the $+x$ direction, and the electrons present inside the target move in the $-x$ direction and flow into there, and then next these flowing electrons are pushed away. []{data-label="fig:fig-ini_e"}](fig-10.pdf){width="7.5cm"}
Similar figures are shown for the electrons in Fig. \[fig:fig-ini\_e\]. Figure \[fig:fig-ini\_e\](a) shows the spatial distributions of the electrons, color-coded by region, and the distribution of the ions in a single color, namely dark blue. Fig. \[fig:fig-ini\_e\](b) shows the distributions of the number of electrons in the $x$ direction. The way to make these figures and display times are the same as for the ion (Fig. \[fig:fig-ini\_i\]). The distributions of the electrons of each region have already changed at $t=-37$ fs, although in contrast, the ion distributions maintain almost the initial distributions at this time (see Fig. \[fig:fig-ini\_i\]). At $t=-47$ fs, a slightly high-density area of electrons of region-1 appears near the surface$^-$, in a way to similar to the ions. Looking at the boundary between region-1, the red line, and region-2, the orange line, of the number density distribution, the electrons of region-1 enter deeply into region-2, and even reach region-3. Some of the electrons of region-2 move a little in the $-x$ direction, and enter slightly into region-1. That is, some of the electrons of region-1 move in the $+x$ direction, and some of the electrons of region-2 move in the opposite, $-x$, direction. This is explained as follows. The laser pulse interacts first with the region-1 electrons which are located closest to it. In this strong interaction between the laser pulse and the region-1 electrons, the electrons near to the surface$^-$ gain great momentum in the $+x$ direction by the ponderomotive force of the laser pulse, and move in the $+x$ direction. At this time, the ions which are much heavier than the electrons hardly move and, as a result, the area where those electrons are located has a $+$charge. The electrons of region-2 which have still almost no momentum, and are nearly stationary, flow into this $+$charge area by Coulomb force. At $t=-37$ fs in Fig. \[fig:fig-ini\_e\](a), many electrons of region-1 appear on the $+x$ side in the outer area of the target. Looking at the density distribution at this time (Fig. \[fig:fig-ini\_e\](b)), the $-x$ side surface of the electrons of region-1 move in the $+x$ direction, and a peak of density appears at a position slightly $+x$ side from this surface. The region-1 electrons are distributed in the outer area on the $+x$ side of the target, while reducing the number density in the $+x$ direction from that peak. On the other hand, in contrast, at this time many electrons of region-2 move from the initial position to the $-x$ direction, and the main distribution area moves close to the initial region-1 area. The peak of the number density is also located in the initial area of region-1, and is at the same positon as that of the region-1 electrons. In the vicinity of those peak positions, the electrons of regions-1 and 2 are approximately equally mixed. At this time, the electrons near the $-x$ side boundaries of regions-3, 4, and 5 also move in the $-x$ direction, and the amount of movement is larger for the region on the $-x$ side. The reason for the movement is the same as for the movement in region-2 at $t=-47$ fs mentioned above, i.e., Coulomb force on the $+$charge near the surface$^-$. At $t=-27$ fs, the distribution areas of the electrons of regions-1, 2, 3, and 4 are approximately the same, and the peaks also appear at approximately the same positions. That is, they are almost equally mixed. The number density of the peak of all the electrons is about $3.5$ times the initial density and is the highest of all the times. On the other hand, region-5 still has an independent distribution at this time.
To summarize the above, the electrons located near the laser irradiation area of the target form a peak of electron density, i.e., high-density electron bunch, by the ponderomotive force of the laser pulse, and that peak gains momentum in the $+x$ direction and moves in the $+x$ direction. Some of those electrons gain greater momentum and move further in the $+x$ direction. After many electrons have moved, the heavy ions are left, and a large $+$charge occurs there. The electrons that are located at a position away from the laser irradiation surface, and are still mostly stationary, move in the $-x$ direction and flow into this area as they are attracted to this charge. These flowing electrons also form a new electron peak and are accelerated in the $+x$ direction by the ponderomotive force. A cycle occurs in which the electrons are pushed away by the laser pulse in the $+x$ direction and the electrons that are located further away flow into there, and then next these flowing electrons are pushed away. The electrons in each region gradually become uniformly mixed, and the kinetic energy of the electrons of all regions increases. We call this mechanism ‘stir-up electron heating’ (SEH). This phenomenon starts early in the laser ion acceleration process, and the SEH process speed is much faster than the ion acceleration process speed. From $t=-17$ fs, the difference in the electron distribution by region disappears, and all the regions are almost equally mixed. The peak of the all the electron densities is located near the center of the initial target in a thickness direction at $t=-17$ fs. Over time: $t=-17$ to $-7$ fs, all the electrons and the peak of the electron density move in the $+x$ direction, and the peak of the electron density gradually decreases.
The positional relationship between the ions and electrons is considered. As shown at $t=-27$ fs in Fig. \[fig:fig-ini\_e\](a), the electrons near the surface$^-$ move in the $+x$ direction, and the ions are left there (the dark blue area). That is, a layer of ions with a strong $+$charge is formed there. Looking at the number density distribution of all the electrons at this time, the black dotted line in Fig. \[fig:fig-ini\_e\](b), a high peak of the electron density is present next to the $+x$ side of that ion layer, and the $+x$ outside area of this peak remain almost its initial density. That is, a layer of electrons with a strong $-$charge is formed. Therefore, two layers, a thin ion layer with a strong $+$charge and a thin electron layer with strong $-$charge, are formed adjacent to each other, and these are strongly pulled toward each other. This electron layer is pushed strongly in the $+x$ direction by the ponderomotive force of the laser pulse. Therefore, although this high-density electron layer receives a strong Coulomb force in the $-x$ direction, by the $+$charge ion layer, it does not move in the $-x$ direction due to the laser pulse coming from the $-x$ side. Then, the ion layer near to the surface$^-$ is pulled strongly by the electron layer which moves fast in the $+x$ direction, and is strongly accelerated in the $+x$ direction. That is, the ions of region-1 near surface$^-$ are strongly accelerated in the $+x$ direction. At this time, since the density distributions of all the electrons in the other areas are almost initial density, such acceleration has not appeared for the ions of the other regions. Although a peak of electron density appears near the boundary between the initial regions-2 and 3 at $t=-17$ fs, and near initial regions-3 and 4 at $t=-7$ fs, those peaks are much smaller than that of $t=-27$ fs (Fig. \[fig:fig-ini\_e\](b)). Therefore, region-3, 4, and 5 ions do not experience such strong acceleration as region-1 ions. This is the reason why the ions of region-1 obtain the highest energy and are distributed on the outermost $+x$ side surface of the target. That is, the laser pulse forms a very high-density layer of electrons on the laser irradiation area and if this peak density is highest during the acceleration process, NH distribution occurs. Next, the positional relationship between the density distributions of the ions and electrons and the electric field is shown.
![ (a) The $x$-component of the electric field, $E_x(x)$, on the $x$ axis near the target. Distributions of the number of electrons (red dotted lines), ions (light blue lines), and the initial distributions (black dotted lines) which are normalized by those initial values, $\tilde{N}=N/n_0$. The peaks of electron and ion number densities move in the $+x$ direction. The electron number density peaks are slightly ahead of those of the ions. The $E_x(x)$ peak appears at the peak point of the number density. Some region-1 ions always position at a high $E_x(x)$ point. (b) The initial target area in the $x$ direction is $x= 0-1 \mu$m. $E_x$ is the electric field of the charged cylinder, and $E_0=\rho\ell/2\epsilon_0$. []{data-label="fig:fig-ini_ei"}](fig-11.pdf){width="17.0cm"}
Figure \[fig:fig-ini\_ei\](a) shows the $x$-component of the electric field, $E_x(x)$, on the $x$ axis around the target, $-0.5\mu$m $< x < 1.5 \mu$m, and the number of density distribution of electrons and ions at the early simulation times $t=-47$ to $-7$ fs, which are the same times as in Figs. \[fig:fig-ini\_i\] and \[fig:fig-ini\_e\]. The number of density distribution lines of ions and electrons are the same as the lines of all the ions and electrons in Figs. \[fig:fig-ini\_i\], \[fig:fig-ini\_e\]. The initial number of density distributions of ions and electrons are indicated by black dotted lines. The initial target area is $x= 0-1 \mu$m (see Fig. \[fig:fig-ini\_ei\](b)). The density distributions of electrons and ions are approximately the same at all times, however, the electron density peaks are slightly ahead of those of the ions in the $+x$ direction in detail, and the peaks of the electric field are also present there in a similar shape. These peaks move in the $+x$ direction over time. At $t=-47$ fs, although the electron and ion distributions remain almost its initial state, there are a very small peaks near the surface$^-$. The electric field, $E_x(x)$, is almost zero throughout, although very small vibrations occur in this part. At $t=-37$ fs, the distributions of electrons and ions have a sharp peak near the surface$^-$ (the area slightly inside the target). And, at the same position, $E_x (x)$ also has a peak of positive values. That is, a force of acceleration in the $+x$ direction is acting on this peak of ions. At $t=-27$ fs, the peaks of the density distribution of electrons and ions are even higher, and these peaks move further in the $+x$ direction than the previous time. And also, $E_x(x)$ has a high peak at that position. The peak position of the electron density is slightly shifted in the $+x$ direction from that of the ion in $-37 \le t \le -17$ fs. Therefore, in this time, the bunch of electrons which has a $-$charge attracts the ion with a $+$charge by the Coulomb force. These electrons form a high-density bunch and are accelerated in the $+x$ direction by the ponderomotive force, and the ions are pulled by this bunch of electrons and are accelerated in the $+x$ direction. Although the accelerating electron bunch is located slightly ahead of the accelerating ions in the accelerating direction, i.e., $+x$ direction, the distance is very short. This is known as radiation pressure acceleration (RPA). In $t \le -17$ fs, the electric field has approximately $+$values in the viewed area, and the strong $+x$ direction electric fields appear at the density peak position of ions. That is, the ions in those peaks have been continually and strongly accelerated in the $+x$ direction during this time. As time goes by: $t=-17$ to $-7$ fs, the peaks of electrons and ions move further in the $+x$ direction and those peak densities gradually decrease, and the sharp peaks become wide. In the electric field too, the narrow peak of $E_x(x)$ becomes wider and those are distributed over a wide area.
At $t=-7$ fs, $E_x(x)$ has a slightly different distribution than the other times. At the previous times, the electric fields are almost in $+$value across the display area and, moreover, sharp peaks are present. On the other hand, at this time, there is a wide high $+$value part and a wide low $-$value part, and those have approximately the same wide and absolute values. $E_x(x)$ has a $+$value on the $+x$ side of the peaks of the density of ions and electrons, and has a $-$value on that $-x$ side. The absolute value of $E_x(x)$ gradually decreases as the distance from the target surfaces, both the $+$ side and $-$ side, increases. This electric field distribution, especially on the $-x$ side, is that of the charged disk and cylinder. That is, an ion acceleration scheme due to the electric field of the charged disk has appeared strongly from this time. Therefore, at this time, a change in the acceleration scheme from the RPA to the acceleration by the charged disk starts. The ions are most strongly accelerated at the target surface, on both $+$ and $-$ sides, in charged disk acceleration, as shown below. Some ions of region-1 are around $x=1.0 \mu$m, as shown at $t=-7$ fs in Fig. \[fig:fig-ini\_i\], and they are moving in the $+x$ direction at about twice the velocity of the region-5 ions present around them. Then after this, these region-1 ions reach the outermost $+x$ side of the target. Therefore, at the time when the RPA has almost finished, and the charged disk acceleration starts and becomes the main acceleration scheme, the region-1 ions are positioned with the velocity, i.e., energy, at the surface$^+$ which is the most effectively accelerated position thereafter. Then, these ions are accelerated most effectively by charged disk acceleration, and obtain more energy, becoming maximum energy ions. Here, the target consists only of hydrogen; therefore, the acceleration by the charged disk $=$ the acceleration by Coulomb explosion of the target.
To summarize the above. In the initial stage of the acceleration process, the ions near the laser irradiation surface are strongly accelerated by RPA in the laser traveling direction, which is the direction toward the inside of the target. These accelerated ions pass through the ions and electrons in the other areas which are almost charge neutral, and are always located at the position where the high $E_x(x)$ occurs which moves at high speed in the $+x$ direction, and keep receiving strong acceleration, and become increasingly high in energy. Then, those ions appear as high energy on the opposite side surface of the target, i.e., surface$^+$. And then, those ions are effectively and further accelerated by the Coulomb explosion, which becomes the main acceleration scheme around this time. As a result those ions are at maximum energy. This is NHA. The ions located near the surface$^-$ obtain both RPA and Coulomb explosion effects, most effectively compared to the other ions in the target in NHA. The obtained energy of those ions is by PRA $+$ Coulomb explosion. NHA is caused by RPA and subsequent Coulomb explosion of the target. This is the reason for the occurrence of NHA.
The accelerated ion energy, $\mathcal{E}_i$, is proportional to the laser energy, $\mathcal{E}_\mathrm{las}$, in RPA [@ESI]. That is proportional to $\sqrt{\mathcal{E}_\mathrm{las}}$ by Coulomb explosion accelerating, as shown below. Since NHA $=$ RPA $+$ Coulomb explosion, $\mathcal{E}_i$(NHA) $=k_\mathrm{R} \mathcal{E}_\mathrm{las} + k_\mathrm{C} \sqrt{\mathcal{E}_\mathrm{las}}$, where $k_\mathrm{R}$ and $k_\mathrm{C}$ are proportionality constants; therefore, the accelerated ion energy is proportional to $\mathcal{E}_\mathrm{las}$, $I$, in NHA.
Next, HA is considered. This is considered a state in which some electrons are removed from the area within the laser spot diameter of the foil target by laser irradiation. At this time, the foil target is assumed as a charged disk or cylinder with a diameter equal to the laser spot size. The electric field of the charged cylinder is considered. It is assumed that the $x$ axis is in the cylinder height direction, and the origin is at the center of the cylinder whose radius is $R$ and height is $\ell$. The $x$-component of the electric field on the $x$ axis in the charged cylinder is written as [@TM3] $$E(x)= \frac{\rho\ell}{2\epsilon_0} \bigl[a(x)+b(x) \bigr],
\label{exab}$$ where $\rho$ is the charge density, $\epsilon_0$ is the vacuum permittivity, and $$a(x)=
\sqrt{\Bigl(\frac{x}{\ell}-\frac{1}{2}\Bigr)^2+\Bigl(\frac{R}{\ell}\Bigr)^2}
-\sqrt{\Bigl(\frac{x}{\ell}+\frac{1}{2}\Bigr)^2+\Bigl(\frac{R}{\ell}\Bigr)^2},
\label{exa}$$ $b(x)$ in the cylinder, $-\ell/2 \leq x \leq \ell/2$, is $$b(x)=
\frac{2x}{\ell}.
\label{exb}$$ The function $E(x)$ passes through the origin, $E(0)=0$, and a monotonically increasing function in $-\ell/2 \leq x \leq \ell/2$ (see Fig. \[fig:fig-ini\_ei\](b)). Therefore, in this charged target, $E(x)$ has $+$value for the $+x$ side region from the center and has $-$value for the $-x$ side region, and the absolute values increase as it goes to both surfaces from the center. In this electric field, ions in the target are accelerated more strongly toward the outside of the target as their position go from the center to the surfaces of the target, which are both surface$^-$ and surface$^+$. That is, as the initial position of the ion locates away from the target center in the $+x$ direction, the ion is strongly accelerated in the $+x$ direction, and also strongly accelerated in the $-x$ direction as it locates away in the $-x$ direction. The shape of this electric field does not change even if the target expands. That is, in this situation, as the position of the ion is closer to the surface of the target, the ion is strongly accelerated during the Coulomb explosion process. This is known as a Coulomb explosion acceleration scheme. This situation occurs when the laser intensity, energy, is relatively weak, or the number of electrons in the thickness direction is great relative to the laser intensity, energy. This is HA.
The relation between laser energy, $\mathcal{E}_\mathrm{las}$, and obtained ion energy, $\mathcal{E}_{i}$, in the disk target is written as [@MBEKK] $$\mathcal{E}_\mathrm{las}=
\frac{\pi \epsilon_0 R\mathcal{E}_{i}^2}{q_{i}^2},
\label{elas}$$ where $q_{i}$ is the ion charge. A model is used in which an ion is placed on the opposite side surface of the laser irradiation surface and the ion is accelerated from the stationary state by the electric field of the charged disk. Hence, it is an HA situation. We solve the equation for $\mathcal{E}_{i}$, then we get $\mathcal{E}_{i} = k \sqrt{\mathcal{E}_\mathrm{las}}$, where $k = q_{i} / \sqrt{\pi\epsilon_0 R}$ and it is the same value in all cases in our simulations. Furthermore, according to Eq. (\[els\]), $\mathcal{E}_\mathrm{las}=I \cdot V^*$ and $V^*$ is a constant value. Therefore, in HA, the obtained ion energy is proportional to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$. The surface charge density, $\rho\ell$, is important for obtained ion energy in HA, as shown in Eq. (\[exab\]).
Whether the target is thick or thin, if the number of electrons taken out per target surface unit area are the same, then $\rho\ell$ will be the same and there will be no difference in the obtained ion energy. Therefore, the difference in target thickness does not significantly affect the obtained ion energy in HA.
NHA is due to the acceleration scheme being RPA $+$ Coulomb explosion, and HA is due to the mainly Coulomb explosion only being dominant in that.
-CH$_2$- foil target {#ch2}
====================
In the previous section, it is shown that NHA appears when the high energy laser irradiates on the hydrogen foil target. Here, we show that NHA does not occur only in the hydrogen target. The simulation results using a polyethylene (-CH$_2$-) foil target are shown. The polyethylene (-CH$_2$-) target is chosen because -CH$_2$- is easy to handle since the material is solid at room temperature, and can produce high energy protons as well as the hydrogen target [@TM2].
The parameters used in simulation of the polyethylene (-CH$_2$-) foil target are given. Since many parameters are the same as the hydrogen foil target case (Sec. \[h-resu\]), the different points of their conditions are shown below. The laser intensity and energy chosen are $I = 1 \times 10^{22}$ W/cm$^2$ and $\mathcal{E}_\mathrm{las}=25$ J, respectively, which are the conditions shown the results in detail in the previous section. The other laser conditions are the same as in the previous section. The foil thickness is set to be a thickness that gives the same areal electron number density, $\int n_{e} dx$, as $1.0 \mu$m hydrogen foil. Here, the foil thickness is simply set to $0.1 \mu$m, although it is strictly $0.16 \mu$m. The ionization state of a carbon ion is assumed to be $Z_{i}=+6$. The electron density is $n_{e}=3\times 10^{23}$ cm$^{-3}$. Therefore, the proton density is $n_{e}/4$ and the carbon ion density is $n_{e}/8$. The total number of quasiparticles is $1\times 10^{9}$. The number of grid cells is equal to $55000\times55000$ along the $X$, and $Y$ axes, respectively. Correspondingly, the simulation box size is $102 \mu$m $\times 102 \mu$m. The laser-irradiated side surface of the foil is placed at $X=40 \mu$m, and the center of the laser pulse is located $20 \mu$m behind it.
![ Spatial distribution of particles and electric field magnitude of each time in the -CH$_2$- foil target case which is $I=1 \times 10^{22}$ W/cm$^2$ ($\mathcal{E}_\mathrm{las}=25$ J) and $\ell=0.1\mu$m. (a) For protons, the color corresponds to their energy, (b) the color corresponds to their initial positon. The protons on outermost part of the $+x$ side and spread around the area on the $+x$ side of the accelerated proton cloud come from near surface$^-$. The NHA regime also appears in the -CH$_2$- target. (c) For carbon ions, the color corresponds to their initial position. A few carbon ions from near to surface$^-$ appear in the outermost $+x$ direction part of the carbon ion cloud. []{data-label="fig:fig-ch"}](fig-12.pdf){width="8.0cm"}
Figure \[fig:fig-ch\](a) shows the result of the -CH$_2$- foil target case at each time. The protons are color-coded by their energy. The initial state is shown at $t=-67$ fs. At $t=33$ fs, part of the laser pulse is reflected by the target while another part is transmitted the target. -CH$_2$- near the laser irradiation area is separated into a carbon ions region and a protons region. The carbon ion cloud distributes near the center of the expanded target, and the proton cloud distributes around it, so as to cover it. The target is greatly expanded at $t=183$ fs, and the proton cloud in particular is largely expanded in the laser propagation direction. This distribution of protons is similar to the results in the case of hydrogen foil. The maximum proton energy at $t=183$ fs is $332$ MeV, which is almost the same as the hydrogen foil case, and those appear on the outermost part of the $+x$ side of the expanded target. Proton energy gradually decreases as goes in the $-x$ direction from there in the range of $x>0$. The maximum carbon ion energy is $54$ MeV/u, which is about 1/6 of the maximum proton energy. Although Fig. \[fig:fig-ch\](b) is the same result as Fig. \[fig:fig-ch\](a), the protons are color-coded according to their initial positions. The initial foil is sliced evenly into five in the thickness direction, i.e., $x$ direction, and each region of protons has a different color, although the carbons are not color-coded (see zoomed initial target in Fig. \[fig:fig-ch\](b)). At $t=183$ fs, it shows that the protons near by the $+x$ side outermost part come from near the opposite, i.e., the $-x$, side surface. That is, the high energy protons are the protons that are initially located near the laser-irradiated surface, surface$^-$. That is the NHA regime too. Also, the protons near the surface$^-$ have already appeared on the outermost part of the $+x$ side of the target, at $t=33$ fs, indicating that this distribution is formed at the early stage of the acceleration process. These are the same as in the hydrogen foil case. Fig. \[fig:fig-ch\](c) is the same result as Fig. \[fig:fig-ch\](a)(b), although the carbon ions are color-coded according to their initial positions. Although the carbon ions are almost HA, we can see that some carbon ions from region-1 appear at the outermost area of the carbon ion cloud in the $+x$ direction.
NHA is not a special phenomenon that occurs only in hydrogen foil targets, but also occurs in polyethylene (-CH$_2$-) foil.
Conclusions
===========
Particle acceleration driven by a laser pulse irradiating a thin-foil target is investigated with the help of 2D PIC simulations. We considered the laser particle acceleration from a topology viewpoint. The acceleration schemes are classified into homeomorphic acceleration (HA) and non-homeomorphic acceleration (NHA), according to whether the spatial distribution of accelerated ions is homeomorphic to the initial target or not. Ion acceleration by RPA $+$ Coulomb explosion is classified into NHA, and the acceleration by mainly only Coulomb explosion is classified into HA, under the conditions of our simulations. In the NHA regime, the obtained proton energy rises in proportion to the laser intensity $I$, laser energy $\mathcal{E}_\mathrm{las}$, and the foil thickness strongly affects the obtained ion energy. On the other hand, in the HA regime, the obtained proton energy rises in proportion to $\sqrt{I}$, $\sqrt{\mathcal{E}_\mathrm{las}}$, and the foil thickness does not affect this much. The HA regime is not efficient in producing high energy ions. In order to generate the highest energy ions as possible at a given laser capacity, i.e., generate high energy ions efficiently, it is necessary to produce the NHA regime. Moreover, it is important to use the optimal target thickness in this regime. We have found that the laser acceleration schemes are classified and elucidated using topology.
Acknowledgments {#acknowledgments .unnumbered}
===============
The computations were performed using the ICE X supercomputer at JAEA Tokai.
[99]{}
S. V. Bulanov, J. J. Wilkens, T. Esirkepov, G. Korn, G. Kraft, S. D. Kraft, M. Molls, and V. S. Khoroshkov: Phys. - Usp. **57**, 1149 (2014)
H. Daido, M. Nishiuchi, and A. Pirozhkov: Rep. Prog. Phys. **75**, 056401 (2012)
E. L. Clark, K. Krushelnick, J. R. Davies, M. Zepf, M. Tatarakis, F. N. Beg, A. Machacek, P. A. Norreys, M. I. K. Santala, I. Watts and A. E. Dangor: Phys. Rev. Lett. **84**, 670 (2000).
R. A. Snavely, M. H. Key, S. P. Hatchett, T. E. Cowan, M. Roth, T. W. Phillips, M. A. Stoyer, E. A. Henry, T. C. Sangster, M. S. Singh, S. C. Wilks, A. MacKinnon, A. Offenberger, D. M. Pennington, K. Yasuike, A. B. Langdon, B. F. Lasinski, J. Johnson, M. D. Perry and E. M. Campbell: Phys. Rev. Lett. **85**, 2945 (2000).
M. Roth, T. E. Cowan, M. H. Key, S. P. Hatchett, C. Brown, W. Fountain, J. Johnson, D. M. Pennington, R. A. Snavely, S. C. Wilks, K. Yasuike, H. Ruhl, F. Pegoraro, S. V. Bulanov, E. M. Campbell, M. D. Perry, and H. Powell, Phys. Rev. Lett. **86**, 436 (2001).
T. Esirkepov, M. Borghesi, S. V. Bulanov, G. Mourou, and T. Tajima, Phys. Rev. Lett. **92**, 175003 (2004).
S. V. Bulanov, E. Yu. Echkina, T. Zh. Esirkepov, I. N. Inovenkov, M. Kando, F. Pegoraro, and G. Korn, Phys. Rev. Lett. **104**, 135003 (2010).
J. Badziak, E. Woryna, P. Parys, K. Yu. Platonov, S. Jabloński, L. Ryć, A. B. Vankov, and J. Woowski, Phys. Rev. Lett. **87**, 215001 (2001).
T. Esirkepov, S. V. Bulanov, K. Nishihara, T. Tajima, F. Pegoraro, V. S. Khoroshkov, K. Mima, H. Daido, Y. Kato, Y. Kitagawa, K. Nagai, and S. Sakabe, Phys. Rev. Lett. **89**, 175003 (2002).
M. Hohenberger, D. R. Symes, K. W. Madison, A. Sumeruk, G. Dyer, A. Edens, W. Grigsby, G. Hays, M. Teichmann, and T. Ditmire, Phys. Rev. Lett. **95** 195003 (2005).
F. Peano, F. Peinetti, R. Mulas, G. Coppa, and L. O. Silva, Phys. Rev. Lett. **96**, 175002 (2006).
A. P. L Robinson, A. R. Bell, and R. J. Kingham, Phys. Rev. Lett. **96**, 035005 (2006).
H. Schwoerer, S. Pfotenhauer, O. Jäckel, K.-U. Amthor, B. Liesfeld, W. Ziegler, R. Sauerbrey, K. W. D. Ledingham, and T. Esirkepov, Nature (London) **439**, 445 (2006).
T. Toncian, M. Borghesi, J. Fuchs, E. d’Humières, P. Antici, P. Audebert, E. Brambrink, C. A. Cecchetti, A. Pipahl, L. Romagnani, and O. Willi, Science [**312**]{}, 410 (2006).
T. Morita, Phys. Plasmas **20**, 093107 (2013).
T. Morita, Phys. Plasmas **21**, 053104 (2014).
C. K. Birdsall and A. B. Langdon, *Plasma Physics via Computer Simulation* (McGraw-Hill, New York, 1985).
H. Kiriyama et al., Opt. Lett. [**43**]{}, 2595 (2018).
T. Morita, Phys. Plasmas **24**, 083104 (2017).
T. Morita, S. V. Bulanov, T. Zh. Esirkepov, J. Koga, and M. Kando, J. Phys. Soc. Jpn. **81**, 024501 (2012).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'J. P. Ruf'
- 'H. Paik'
- 'N. J. Schreiber'
- 'H. P. Nair'
- 'L. Miao'
- 'J. K. Kawasaki'
- 'J. N. Nelson'
- 'B. D. Faeth'
- 'Y. Lee'
- 'B. H. Goodge'
- 'B. Pamuk'
- 'C. J. Fennie'
- 'L. F. Kourkoutis'
- 'D. G. Schlom'
- 'K. M. Shen'
title: 'Strain-stabilized superconductivity'
---
**Superconductivity is among the most fascinating and well-studied quantum states of matter. Despite over 100 years of research, a detailed understanding of how features of the normal-state electronic structure determine superconducting properties has remained elusive. For instance, the ability to deterministically enhance the superconducting transition temperature by design, rather than by serendipity, has been a long sought-after goal in condensed matter physics and materials science, but achieving this objective may require new tools, techniques and approaches. Here, we report the first instance of the transmutation of a normal metal into a superconductor through the application of epitaxial strain. We demonstrate that synthesizing [RuO$_2$ ]{}thin films on (110)-oriented [TiO$_2$ ]{}substrates enhances the density of states near the Fermi level, which stabilizes superconductivity under strain, and suggests that a promising strategy to create new transition-metal superconductors is to apply judiciously chosen anisotropic strains that redistribute carriers within the low-energy manifold of $d$ orbitals.**
In typical weak-coupling theories of superconductivity, the effective attraction $V$ between electrons is mediated by the exchange of bosons having a characteristic energy scale $\omega_{\text{B}}$, and superconductivity condenses below a transition temperature [$T_c$ ]{}parameterized as [@carbotte_properties_1990]:
$$T_{c} \sim \omega_{\text{B}}\exp\left(-\frac{1}{N(E_F)V}\right) = \omega_{\text{B}}\exp\left(-\frac{1 + \lambda}{\lambda - \mu^{*}}\right)$$
where $N(E_F)$ is the density of states (DOS) near the Fermi level, $\lambda$ is the electron-boson coupling strength, and $\mu^{*}$ is the Coulomb pseudopotential that describes the residual Coulomb repulsion between quasiparticles [^1].
Experimental methods that boost [$T_c$ ]{}are highly desired from a practical perspective. Furthermore, by analyzing how these available knobs couple to the normal-state properties on the right side of equation (1), one can envisage engineering the electronic structure and electron-boson coupling to optimize [$T_c$]{}. For example, increasing $N(E_F)$ is a frequently suggested route towards realizing higher [$T_c$]{}, but how to achieve this for specific materials often remains unclear.
Chemical doping and hydrostatic pressure have been the most common knobs used to manipulate superconductivity. Unfortunately, doping has the complication of introducing substitutional disorder, whereas pressure studies are incompatible with most probes of electronic structure. Moreover, because large pressures are usually required to appreciably increase [$T_c$]{} [@hamlin_superconductivity_2015], pressure-enhanced superconductivity exists transiently—typically in different structural polymorphs than at ambient conditions—rendering it inaccessible for applications.
An alternative strategy for controlling superconductivity is epitaxial strain engineering. This approach is static, disorder-free, allows for the use of sophisticated experimental probes [@burganov_strain_2016], and enables integration with other materials in novel artificial interfaces and device structures [@ohtomo_high-mobility_2004; @kawasaki_rutile_2018]. To date, epitaxial strain has only been used to modulate [$T_c$ ]{}in known superconductors [@lock_penetration_1951; @locquet_doubling_1998; @*si_strain_1999; @*si_epitaxial-strain-induced_2001; @*bozovic_epitaxial_2002]. In this letter, we report the first example of a new superconductor created through epitaxial strain, starting from a compound, [RuO$_2$]{}, previously not known to be superconducting [^2]. By comparing the results of angle-resolved photoemission spectroscopy (ARPES) experiments with density functional theory (DFT) calculations, we show that splittings between the effective low-energy $d$ orbital degrees of freedom in [RuO$_2$ ]{}respond sensitively to appropriate modes of strain, and we discuss how this approach may open the door to strain tuning of superconductivity in other materials.
{width="7in"}
Bulk [RuO$_2$ ]{}crystallizes in the ideal tetragonal rutile structure (space group $\#136$, $P4_2/mnm$) with lattice constants at 295 K of $(a=4.492 \text{ {\text{\normalfont\AA}}}, c=3.106 \text{ {\text{\normalfont\AA}}})$ [@berlijn_itinerant_2017]. [RuO$_2$ ]{}thin films in distinct epitaxial strain states were synthesized using oxide molecular-beam epitaxy by employing different orientations of isostructural [TiO$_2$ ]{}substrates, $(a=4.594 \text{ {\text{\normalfont\AA}}}, c=2.959 \text{ {\text{\normalfont\AA}}})$ [@burdett_structural-electronic_1987]. As shown in Fig.1a, the surfaces of (101)-oriented substrates are spanned by the $[\overline{1}01]$ and $[010]$ lattice vectors of [TiO$_2$]{}, which ideally impart in-plane tensile strains on [RuO$_2$ ]{}(at 295 K) of $+0.04\%$ and $+2.3\%$, respectively. On [TiO$_2$]{}(110), the lattice mismatches with [RuO$_2$ ]{}are larger: $-4.7\%$ along $[001]$ and $+2.3\%$ along $[1\overline{1}0]$. Figure 1b shows electrical resistivity $\rho(T)$ measurements for [RuO$_2$ ]{}films, along with results for bulk [RuO$_2$ ]{}single crystals from Ref. [@lin_low_2004]. To compare with bulk [RuO$_2$]{}, for the thin-film samples we plot the geometric mean of $\rho$ along the two in-plane directions; the intrinsic resistivity anisotropy is known to be small (Supplementary Fig.S1).
$\rho(T)$ data for the lightly strained [RuO$_2$]{}/[TiO$_2$]{}(101) sample—henceforth referred to as [RuO$_2$(101)]{}—are nearly indistinguishable from bulk, exhibiting metallic behavior with a low residual resistivity $\rho(0.4\text{ K}) < 2\text{ } \mu\Omega\text{-cm}$. In contrast, a clear superconducting transition is observed for the more heavily strained [RuO$_2$]{}/[TiO$_2$]{}(110) sample—referred to as [RuO$_2$(110)]{}—at $T_{c} = 2.0 \pm 0.1$ K. Magnetoresistance measurements with $H_{\perp}$ applied along $[110]$ (the out-of-plane direction) show a suppression of [$T_c$ ]{}with increasing fields and an extrapolated value of $H_{c\perp}(T \rightarrow 0 \text{ K}) = 13.3 \pm 1.5$ kOe, corresponding to an average in-plane superconducting coherence length of $\xi(T \rightarrow 0 \text{ K}) = 15.8 \pm 0.9$ nm (Supplementary Fig.S2). In Fig.1d, we show a $V(I)$ curve measured on a lithographically patterned resistivity bridge at $T/T_{c} = 0.3$, from which we extract a critical current density $J_c = (9.5 \pm 1.2) \times 10^4$ A/$\text{cm}^2$. This large value of $J_c$ (over one order of magnitude larger than values reported on typical elemental superconductors with comparable [$T_c$]{}s) indicates that the superconductivity does not arise from a filamentary network, structural defects, minority phases, or from the substrate-film interface, which would all yield much smaller values of $J_c$.
{width="7in"}
In order to disentangle the effects of strain from other possible sources of superconductivity, we compare [RuO$_2$ ]{}films as functions of strain and film thickness, $t$. In Fig.2a, we plot x-ray diffraction (XRD) data from similar-thickness films of [RuO$_2$(101) ]{}and [RuO$_2$(110)]{}, showing that the bulk-averaged crystal structures of the films are strained as expected along the out-of-plane direction based on their net in-plane lattice mismatches with [TiO$_2$]{}. The primary 101 and 202 film peaks of [RuO$_2$(101) ]{}are shifted to larger angles than bulk [RuO$_2$]{}, corresponding to a $1.1 \%$ compression of $d_{101}$, while Nelson-Riley analysis of the primary 110, 220 and 330 (not shown) peak positions for [RuO$_2$(110) ]{}evidence a $2.0 \%$ expansion of $d_{110}$ relative to bulk. In Fig.2b,c, we plot resistivity data showing that reducing $t$ in [RuO$_2$(110) ]{}decreases [$T_c$]{}, characteristic of numerous families of thin-film superconductors [@pinto_dimensional_2018; @meyer_strain-relaxation_2015], with [$T_c$ ]{}dropping below our experimental threshold (0.4 K) between $t = 11.5 \text{ nm}$ and $5.8 \text{ nm}$. This suppression of [$T_c$ ]{}with thickness indicates superconductivity is *not* confined near the substrate-film interface, so possible interfacial modifications of the crystal structure [@gozar_high-temperature_2008], carrier density [@he_phase_2013], substrate-film mode coupling [@lee_interfacial_2014], and non-stoichiometry in the films or substrates [@paik_transport_2015; @yoshimatsu_superconductivity_2017] can all be eliminated as potential causes of superconductivity. These conclusions are also supported by the facts that superconductivity is not observed in [RuO$_2$(101) ]{}films, nor in bare [TiO$_2$ ]{}substrates treated in an identical fashion to the [RuO$_2$ ]{}films. Finally, in Fig.2d we show a scanning transmission electron microscopy (STEM) image of a superconducting [RuO$_2$(110) ]{}sample, which confirms uniform growth of the film over lateral length scales exceeding those expected to be relevant for superconductivity (*e.g.*, $\xi$), with no evidence of minority phases and a chemically abrupt interface between [RuO$_2$ ]{}and [TiO$_2$ ]{}(Supplementary Figs.S5-S7).
We believe the thickness dependence of [$T_c$ ]{}results primarily from the competition between: 1) an intrinsic strain-induced enhancement of [$T_c$ ]{}that should be maximized for thinner, commensurately strained [RuO$_2$(110) ]{}films, versus 2) disorder-induced suppressions of [$T_c$ ]{}that become amplified in the ultrathin limit (see $\rho_{0}$ versus $t$ in Fig.2c). While the thinnest films experience the largest substrate-imposed strains, stronger disorder scattering (likely from interfacial defects) reduces [$T_c$ ]{}below our detection threshold. Films of intermediate thickness ($t \approx 10 - 25\text{ nm}$) have lower residual resistivities and higher [$T_c$]{}s, but do exhibit signatures of partial strain relaxation. Nevertheless, a detailed analysis of misfit dislocations by STEM and XRD reciprocal-space mapping (Supplementary Figs.S3-S10) indicates that these films are largely structurally homogeneous and, on average, much closer to commensurately strained than fully relaxed. Finally, in much thicker samples ($t = 48$ nm) where a more significant volume fraction of the film should be relaxed, the strain is further released by oriented micro-cracks that make these samples spatially inhomogeneous and cause anisotropic distributions of current flow, preventing reliable resistivity measurements (Supplementary Fig.S11).
{width="7in"}
Having established the strain-induced nature of the superconductivity in [RuO$_2$(110)]{}, we now explore its underlying origin using a combination of DFT and ARPES. In Fig.3a, we present the electronic structure of commensurately strained [RuO$_2$(110) ]{}calculated by DFT$+U$ ($U$ = 2 eV), following the methods of Berlijn *et al.* [@berlijn_itinerant_2017]. Despite being constructed of RuO$_6$ octahedra having the same $4d^4$ electronic configuration as in (Ca,Sr,Ba)RuO$_3$, the electronic structure of [RuO$_2$ ]{}is markedly different from that of perovskite-based ruthenates. These distinctions arise from a sizable ligand-field splitting of the $t_{2g}$ orbitals, such that the most natural description of the low-energy electronic structure is in terms of states derived from two distinct types of orbitals: [$d_{||}$ ]{}and [$(d_{\text{xz}}$, $d_{\text{yz}})$]{}, as illustrated by plots of Wannier functions in Fig.3 [@goodenough_two_1971; @eyert_embedded_2000]. Viewed in the band basis in Fig.3a, the differentiation in $\mathbf{k}$-space between these orbitals becomes apparent: the near-$E_F$ [$d_{||}$ ]{}states form mostly flat bands concentrated around the $k_{001}=\pi/c$ (Z-R-A) plane, whereas the [$(d_{\text{xz}}$, $d_{\text{yz}})$ ]{}states form more isotropically dispersing bands distributed uniformly throughout the Brillouin zone.
In many other $d^4$ ruthenates (such as Sr$_2$RuO$_4$ and Ca$_2$RuO$_4$), static mean-field electronic structure calculations (such as DFT$+U$) often predict quantitatively incorrect effective masses [@mravlje_coherence-incoherence_2011; @*mackenzie_quantum_1996; @*tamai_high-resolution_2019; @ricco_situ_2018; @*sutter_orbitally_2019] or even qualitatively incorrect ground states [@sutter_hallmarks_2017] because these approaches neglect local atomic-like (Hund’s rule) spin correlations that strongly renormalize the low-energy quasiparticle excitations. Therefore, it is imperative to compare DFT calculations for [RuO$_2$ ]{}with experimental data, to establish the reliability of any theoretically predicted dependence of the electronic structure on strain. The left half of Fig.3b shows the Fermi surface of [RuO$_2$(110) ]{}measured with He-I$\alpha$ (21.2 eV) photons at 17 K, which agrees well with a non-magnetic DFT+$U$ simulation of the Fermi surface at a reduced out-of-plane momentum of $k_{110} = -0.2 \pm 0.2 \text{ } \pi/d_{110}$ (right half of Fig.3b). In Fig.3c,d, we plot energy versus momentum spectra acquired along the dashed lines in Fig.3b: in Fig.3c, the spectrum is dominated by the flat [$d_{||}$ ]{}bands around a binding energy of 300 meV, whereas in Fig.3d the [$(d_{\text{xz}}$, $d_{\text{yz}})$ ]{}bands are steeply dispersing and can be tracked down to several hundred meV below $E_F$, both of which are well reproduced by DFT$+U$ calculations. The agreement between the experimental and DFT band velocities is consistent with recent ARPES studies of Ir-doped [RuO$_2$ ]{}single crystals [@jovic_dirac_2018] and with earlier specific heat measurements of the Sommerfeld coefficient in bulk [RuO$_2$]{}, which suggested a modest momentum-averaged mass renormalization of $\gamma_{\text{exp.}} = 1.45\gamma_{\text{DFT}}$ [@passenheim_heat_1969; @glassford_electron_1994]. The fact that the true electronic structure of [RuO$_2$ ]{}can be well accounted for by DFT+$U$ allows us to utilize such calculations to understand how epitaxial strains can be employed to engineer features of the electronic structure to enhance the instability towards superconductivity.
{width="7in"}
In Fig.4a, we show the strain dependence of the DFT band structure (left panel) and DOS (right panel) for [RuO$_2$(110)]{}, [RuO$_2$(101)]{}, and bulk [RuO$_2$]{}. While the results for [RuO$_2$(101) ]{}are almost identical to bulk, the results for [RuO$_2$(110) ]{}exhibit significant differences: the large [$d_{||}$]{}-derived peak in the DOS (centered around a binding energy of 800 meV for bulk) is split into multiple peaks for [RuO$_2$(110)]{}, several of which are shifted closer to the Fermi level, thereby increasing $N(E_F)$. In our studies, we found that this strain-dependent trend was robust against details of the DFT calculations, such as whether $U$ was finite (Supplementary Fig.S12). In order to determine whether this strain dependence of $N(E_F)$ is realized in experiment, we compared the electronic structure of a thin (7 nm) highly strained [RuO$_2$(110) ]{}film with a much thicker (48 nm) strain-relaxed [RuO$_2$(110) ]{}film. The surface lattice constants of the 48 nm thick film were closer to bulk [RuO$_2$ ]{}than the 7 nm thick film (Supplementary Fig.S14), so we expect that the surface electronic structure probed by ARPES of the thicker film to be more representative of bulk [RuO$_2$]{}. Comparisons between the [RuO$_2$(110) ]{}and [RuO$_2$(101) ]{}surfaces are less straightforward, since different parts of the three-dimensional Brillouin zone are sampled by ARPES (Supplementary Fig.S15). Figure 4b shows $E(k)$ spectra side-by-side for the 7 nm (left) and 48 nm (right) films of [RuO$_2$(110) ]{}along the same cut (c) through $\mathbf{k}$-space from Fig.3 where the photoemission intensity is dominated by [$d_{||}$ ]{}initial states. The higher levels of strain present at the film surface for the 7 nm thick sample cause a substantial shift of the flat bands towards $E_F$ by $120 \pm 20$ meV relative to the more relaxed 48 nm thick sample. Integrating the ARPES data over the full measured region of $\mathbf{k}$-space for both samples gives the average energy distribution curves plotted in the right half of Fig.4b, which show that spectral weight near $E_F$ is enhanced as the [$d_{||}$ ]{}states move towards $E_F$, in qualitative agreement with the trend predicted by DFT. Our results show that the primary electronic effect of epitaxial strains in [RuO$_2$(110) ]{}is to alter the relative occupancies of the [$d_{||}$ ]{}and [$(d_{\text{xz}}$, $d_{\text{yz}})$ ]{}orbitals as compared with bulk, and to push a large number of states with [$d_{||}$ ]{}character closer to $E_F$, which enhances $N(E_F)$ and likely [$T_c$]{}.
Observations of Fermi-liquid-like quasiparticles near $E_F$ that scatter at higher energies primarily via their interaction with phonons [@glassford_electron_1994], along with the fact that superconductivity in [RuO$_2$(110) ]{}persists in the dirty limit (Supplementary Fig.S9), are both consistent with conventional Cooper pairing, suggesting that calculations assuming an electron-phonon mechanism may be enlightening. We performed DFT-based Migdal-Eliashberg calculations of [$T_c$ ]{}for bulk [RuO$_2$ ]{}and strained [RuO$_2$(110) ]{}that indeed indicate epitaxial strain can enhance [$T_c$ ]{}by several orders of magnitude. For bulk [RuO$_2$]{}, we find that the empirical Coulomb pseudopotential must satisfy $\mu^{*} > 0.30$ to be compatible with the experimentally measured upper bound on $T_{c} < 0.3 \text{ K}$. For this range of $\mu^{*}$, [$T_c$ ]{}for [RuO$_2$(110) ]{}can be as high as $7 \text{ K}$ (Supplementary Fig.S16). A robust strain-induced enhancement of the electron-phonon coupling $\lambda_{\text{el-ph}}$ boosts [$T_c$ ]{}by a factor of 20 (for $\mu^{*} = 0.30$), and this ratio becomes even larger for higher values of $\mu^{*}$ (*e.g.* for $\mu^{*} = 0.37$, $T_{c}(110)/T_{c}(\text{bulk}) = 5 \text{ K}/5 \text{ mK}$). Although these estimations of [$T_c$ ]{}are broadly consistent with our experimental findings, conventional superconductivity in [RuO$_2$ ]{}remains a working hypothesis until measurements of the order parameter are possible.
We believe our results demonstrate that a promising strategy to create new transition-metal superconductors is to apply judiciously chosen anisotropic strains that modulate degeneracies among $d$ orbitals near $E_F$. Many classic studies of conventional superconductors that have nearly-free-electron $(s,p)$-derived states spanning $E_F$ show *decreases* in [$T_c$ ]{}under hydrostatic pressure, due to lattice stiffening [@smith_will_1967]. In a limited number of elements where [$T_c$ ]{}actually *increases* under pressure, electron transfer between $s \rightarrow d$ orbitals has been suggested as causing the enhanced [$T_c$]{}s [@hamlin_superconductivity_2015]; a drawback of this approach, however, is that large pressures of $\approx 10$ GPa are typically required to, *e.g.*, double [$T_c$]{}. More recently, measurements on single crystals of the unconventional superconductor Sr$_2$RuO$_4$ have shown that uniaxial pressures of only $\approx 1\text{ GPa}$ can boost [$T_c$ ]{}by more than a factor of two [@steppke_strong_2017]. Independent of the underlying mechanism, it appears that anisotropic strains may prove much more effective than hydrostatic pressure for tuning superconductivity in multi-orbital systems, as shown here for [RuO$_2$]{}, as well as in Sr$_2$RuO$_4$.
Sizable coupling between the lattice and electronic degrees of freedom in rutile-like crystal structures has been well established both theoretically [@eyert_embedded_2000] and experimentally in VO$_{2}$, where strain-induced variations in the orbital occupancies can modify the metal-insulator transition by $\delta T_{\text{MIT}} \approx 70 \text{ K}$ [@muraoka_metalinsulator_2002; @*aetukuri_control_2013]. Therefore, it may be promising to explore other less strongly correlated (*i.e.*, $4d$ and $5d$) rutile compounds such as MoO$_2$ for strain-stabilized superconductivity, instead of employing chemical doping [@alves_unconventional_2010; @*alves_superconductivity_2012; @*parker_evidence_2014]. Finally, since [RuO$_2$]{}/[TiO$_2$]{}(110) is the first known stoichiometric superconductor within the rutile family, further optimization of the superconductivity may enable the creation of structures that integrate superconductivity with other functional properties that have been extensively studied in other rutile compounds, such as high photocatalytic efficiency, half-metallic ferromagnetism, and large spin Hall conductivities.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank Y. Li for assistance with electrical transport measurements. This work was supported through the National Science Foundation \[Platform for the Accelerated Realization, Analysis, and Discovery of Interface Materials (PARADIM)\] under Cooperative Agreement No. DMR-1539918, NSF DMR-1709255, the Air Force Office of Scientific Research Grant No. FA9550-15-1-0474, and the Department of Energy (Award No. DE-SC0019414). This research is funded in part by the Gordon and Betty Moore Foundation’s EPiQS Initiative through Grant No. GBMF3850 to Cornell University. This work made use of the Cornell Center for Materials Research (CCMR) Shared Facilities, which are supported through the NSF MRSEC Program (No. DMR-1719875). The FEI Titan Themis 300 was acquired through NSF-MRI-1429155, with additional support from Cornell University, the Weill Institute, and the Kavli Institute at Cornell. Device fabrication and substrate preparation were performed in part at the Cornell NanoScale Facility, a member of the National Nanotechnology Coordinated Infrastructure (NNCI), which is supported by the NSF (Grant No. ECCS-1542081).
Author contributions {#author-contributions .unnumbered}
====================
H.P., N.J.S., and H.P.N. synthesized the samples by MBE. J.P.R., H.P., N.J.S., and H.P.N. characterized the samples by XRD. J.P.R., L.M., and Y.L. characterized the samples by electrical transport; L.M. patterned resistivity bridges on select films. J.P.R., J.K.K., J.N.N., and B.D.F. characterized the samples by ARPES and LEED. B.H.G. characterized the samples by STEM. J.P.R. performed DFT calculations of the electronic structure and B.P. performed DFT-based calculations of the electron-phonon coupling. L.F.K., D.G.S., and K.M.S. supervised the various aspects of this project. J.P.R. and K.M.S. wrote the manuscript with input from all authors.
Competing interests {#competing-interests .unnumbered}
===================
The authors declare no competing interests.
Data availability statement {#data-availability-statement .unnumbered}
===========================
The datasets generated during and analysed during the current study are available from the corresponding author on reasonable request.
Methods {#methods .unnumbered}
=======
Film synthesis {#film-synthesis .unnumbered}
--------------
Epitaxial thin films of [RuO$_2$ ]{}were synthesized on various orientations of rutile [TiO$_2$ ]{}substrates using a GEN10 reactive oxide molecular-beam epitaxy (MBE) system (Veeco Instruments). Prior to growth, [TiO$_2$ ]{}substrates (Crystec, GmbH) were cleaned with organic solvents, etched in acid, and annealed in air to produce starting surfaces with step-terrace morphology, following the methods in Ref. [@yamamoto_preparation_2005]. Elemental ruthenium (99.99% purity, ESPI Metals) was evaporated using an electron-beam evaporator in background oxidant partial pressures of $1 \times 10^{-6} - 5 \times 10^{-6}$ Torr of distilled ozone ($\approx$ 80% O$_3$ + 20% O$_2$) at substrate temperatures of $250 - 400^{\circ}$ C, as measured by a thermocouple. Reflection high-energy electron diffraction was used to monitor the surface crystallinity of the films *in situ* and showed characteristic oscillations in intensity during most of the Ru deposition, indicating a layer-by-layer growth mode following the initial nucleation of several-monolayer-thick [RuO$_2$ ]{}islands [@he_versatile_2015].
Film characterization {#film-characterization .unnumbered}
---------------------
The crystal structures of all [RuO$_2$ ]{}thin-film samples were characterized via lab-based x-ray diffraction (XRD) measurements with Cu-K$\alpha$ radiation (Rigaku SmartLab and Malvern Panalytical Empyrean diffractometers). Four-point probe electrical transport measurements were conducted on greater than 20 [RuO$_2$ ]{}films from 300 K down to a base temperature of 0.4 K using a Physical Properties Measurement System equipped with a He-3 refrigerator (Quantum Design). All [RuO$_2$ ]{}/ [TiO$_2$]{}(110) samples were superconducting with [$T_c$]{}s ranging from $0.5 - 2.4$ K, except for ultrathin films with residual resistivities $\rho_0 \gtrapprox 40 \text{ } \mu\Omega\text{-cm}$, as described in the main text and supplemental information. A systematic investigation of how the crystal structures, structural defects and superconducting properties (*e.g.* [$T_c$ ]{}and [$H_c$]{}) of the films are correlated—and how these vary under different film synthesis conditions—is beyond the scope of the present work and will be published elsewhere.
A subset of films studied by XRD and transport were also characterized *in situ* by angle-resolved photoemission spectroscopy (ARPES) and low-energy electron diffraction (LEED). For these measurements, films were transferred under ultrahigh vacuum immediately following growth to an analysis chamber with a base pressure of $5\times10^{-11}$ Torr equipped with a helium plasma discharge lamp, a hemispherical electron analyzer (VG Scienta R4000), and a four-grid LEED optics (SPECS ErLEED 150).
A subset of films studied by XRD and transport were also imaged using cross-sectional scanning transmission electron microscopy. Cross-sectional specimens were prepared using the standard focused ion beam (FIB) lift-out process on a Thermo Scientific Helios G4 X FIB. High-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) images were acquired on an aberration-corrected FEI Titan Themis at 300 keV with a probe convergence semi-angle of 21.4 mrad and inner and outer collection angles of 68 and 340 mrad.
Electronic structure calculations {#electronic-structure-calculations .unnumbered}
---------------------------------
Non-magnetic density functional theory (DFT) calculations for the electronic structure of [RuO$_2$ ]{}were performed using the [Quantum ESPRESSO]{} software package [@giannozzi_quantum_2009; @*giannozzi_advanced_2017] with fully relativistic ultrasoft pseudopotentials for Ru and O [@dal_corso_pseudopotentials_2014]. We represented the Kohn-Sham wavefunctions in a basis set of plane waves extending up to a kinetic energy cutoff of 60 Ry, and used a cutoff of 400 Ry for representing the charge density. Brillouin zone integrations were carried out on an $8\times8\times12$ $k$-mesh with 70 meV of Gaussian smearing. Perdew, Burke and Ernzerhof’s parametrization of the generalized gradient approximation was employed as the exchange-correlation functional [@PBE_1996], supplemented by an on-site correction of $+U_{\text{eff}} = U - J = 2$ eV within spheres surrounding the Ru sites, following Ref. [@berlijn_itinerant_2017].
After obtaining self-consistent Kohn-Sham eigenstates via DFT, we used the pw2wannier and Wannier90 codes [@Wannier90_2014] to construct 20 Wannier functions spanning the manifold of eigenstates surrounding $E_F$ (20 = 10 Kramers-degenerate $d$-orbitals per Ru atom $\times$ 2 Ru atoms per unit cell). Following Ref. [@Eyert_2002], to account for the non-symmorphic space group symmetries of rutile crystal structures, we referenced the trial orbitals employed in the Wannierisation routine to locally rotated coordinate systems centered on the two Ru sites within each unit cell. Orbital designations employed in the main text such as [$d_{||}$ ]{}and [$(d_{\text{xz}}$, $d_{\text{yz}})$ ]{}refer to projections onto this basis of Wannier functions. The more computationally efficient Wannier basis was used to calculate quantities that required dense $k$ meshes to be properly converged, such as the the projected Fermi surface in Fig.3 ($51 \times 51 \times 51$ $k$-mesh) and the near-$E_F$ density of states traces in Fig.4 ($32 \times 32 \times 48$ $k$-meshes).
Because the [RuO$_2$ ]{}samples studied in this work are thin films subject to biaxial epitaxial strains imposed by differently oriented rutile TiO$_2$ substrates, we performed DFT + Wannier calculations of the electronic structure for several different crystal structures of [RuO$_2$ ]{}as summarized in the supplemental information (Supplementary Table 1). We used the ISOTROPY software package [@isotropy] to study distortions of the parent tetragonal rutile crystal structure that are induced in biaxially strained thin films. Crystal structures and Wannier functions were visualized using the VESTA software package [@momma_vesta_2011].
Electron-phonon coupling calculations {#electron-phonon-coupling-calculations .unnumbered}
-------------------------------------
To generate the inputs required for the electron-phonon coupling calculations described below, first-principles electronic structure and phonon calculations were performed using the [Quantum ESPRESSO]{} software package with norm-conserving pseudopotentials and plane-wave basis sets [@giannozzi_quantum_2009; @*giannozzi_advanced_2017]. Here we employed a kinetic energy cutoff of 160 Ry, an electronic momentum $k$-point mesh of $16\times16\times24$, $20 \text{ meV}$ of Methfessel-Paxton smearing for the occupation of the electronic states, and a tolerance of $10^{-10} \text{ eV}$ for the total energy convergence. The generalized gradient approximation as implemented in the PBEsol functional [@perdew_restoring_2008] was employed as the exchange-correlation functional. For the Wannier interpolation, we used an interpolating electron-momentum mesh of $8\times8\times12$ and a phonon-momentum mesh of $2\times2\times3$. Results for bulk [RuO$_2$ ]{}were calculated using the crystal structure that minimizes the DFT-computed total energy with the PBEsol functional: $(a=4.464 \text{ {\text{\normalfont\AA}}}, c=3.093 \text{ {\text{\normalfont\AA}}})$ and $x_{\text{oxygen}} = 0.3062$. Results for strained [RuO$_2$(110) ]{}were calculated by changing the lattice constants of this “bulk” crystal structure by $+2.3\%$ along $[1\overline{1}0]$, $-4.7\%$ along $[001]$, $+2.2\%$ along $[110]$, and setting $x_{\text{oxygen}} = y_{\text{oxygen}} = 0.2996$. The lattice constant along $[110]$ and internal coordinates of this simulated [RuO$_2$(110) ]{}structure were determined by allowing the structure to relax so as to (locally) minimize the DFT-computed total energy.
Electron-phonon coupling calculations were performed using the EPW code [@ponce_epw_2016], using an interpolated electron-momentum mesh of $32\times32\times48$ and an interpolated phonon-momentum mesh of $8\times8\times12$. The isotropic Eliashberg spectral function $\alpha^2 F(\omega)$ and total electron-phonon coupling constant $\lambda_{\text{el-ph}}$ (integrated over all phonon modes and wavevectors) were calculated with a phonon smearing of 0.2 meV. From the calculated $\alpha^2 F(\omega)$ and $\lambda_{\text{el-ph}}$, we estimated the superconducting transition temperature using the semi-empirical McMillan-Allen-Dynes formula [@mcmillan_transition_1968; @*allen_transition_1975]:
$$T_c = \frac{\omega_{\log}}{1.2} \exp\left[ - \frac{1.04 (1 + \lambda_{\text{el-ph}} ) } { \lambda_{\text{el-ph}} - \mu^{*} (1 + 0.62 \lambda_{\text{el-ph}} ) } \right]$$
[^1]: For simplicity, we assume that all of the non-isotropic $\mathbf{q}$- and $\mathbf{k}$-dependencies that appear in a more realistic formulation of the Cooper pairing problem have been averaged away. Note that within the range of validity of equation (1)—*viz.*, $1 >> \lambda > \mu^{*}$—increasing $\lambda$ (increasing $\mu^{*}$) generally enhances (suppresses) [$T_c$]{}, respectively, assuming that superconductivity remains the dominant instability.
[^2]: In principle, assuming that all Fermi liquids are eventually unstable towards superconductivity at low enough temperatures and magnetic fields (including internal fields arising from magnetic impurities), this is not strictly a change in the ground state of the system. However, for our purposes, extremely low temperatures and fields below what are experimentally achievable can be thought of as effectively zero, justifying the use of phrases such as “strain-induced superconductivity” interchangeably with “huge enhancement of critical temperature”.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Generative adversarial networks (GANs) are increasingly attracting attention in the computer vision, natural language processing, speech synthesis and similar domains. Arguably the most striking results have been in the area of image synthesis. However, evaluating the performance of GANs is still an open and challenging problem. Existing evaluation metrics primarily measure the dissimilarity between real and generated images using automated statistical methods. They often require large sample sizes for evaluation and do not directly reflect human perception of image quality.
In this work, we describe an evaluation metric we call **Neuroscore**, for evaluating the performance of GANs, that more directly reflects psychoperceptual image quality through the utilization of brain signals. Our results show that Neuroscore has superior performance to the current evaluation metrics in that: (1) It is more consistent with human judgment; (2) The evaluation process needs much smaller numbers of samples; and (3) It is able to rank the quality of images on a per GAN basis.
A convolutional neural network (CNN) based **neuro-AI interface** is proposed to predict Neuroscore from GAN-generated images directly without the need for neural responses. Importantly, we show that including neural responses during the training phase of the network can significantly improve the prediction capability of the proposed model. Materials related to this work are provided at *<https://github.com/villawang/Neuro-AI-Interface>.*
address:
- 'V-SENSE, School of Computer Science and Statistics, Trinity College Dublin, Dublin 1, Ireland'
- 'Intel Labs, Beijing, China'
- 'Insight Centre for Data Analytics, Dublin City University, Dublin 9, Ireland'
author:
- Zhengwei Wang
- Qi She
- 'Alan F. Smeaton'
- 'Tomás E. Ward'
- Graham Healy
bibliography:
- 'ref.bib'
title: |
Synthetic-Neuroscore: Using A Neuro-AI Interface for Evaluating\
Generative Adversarial Networks
---
=1
Neuroscore, Generative adversarial networks, Neuro-AI interface, Brain-computer interface.
Introduction
============
There is a growing interest in studying generative adversarial networks (GANs) in the deep learning community [@goodfellow2014generative; @wang2019generative]. Specifically, GANs have been widely applied to various domains such as computer vision [@karras2018style], natural language processing [@fedus2018maskgan], speech synthesis [@donahue2018synthesizing] and time series synthesis [@brophy2019quick]. Compared with other deep generative models (e.g. variational autoencoders (VAEs)), GANs are favored for effectively handling sharp estimated density functions, efficiently generating desired samples and eliminating deterministic bias. Due to these properties GANs have successfully contributed to plausible image generation [@karras2018style], image to image translation [@zhu2017unpaired], image super-resolution [@ledig2017photo], image completion [@yu2018generative] etc.
However, three main challenges currently in research into GANs could be: (1) Mode collapse – the model cannot learn the distribution of the full dataset well, which leads to poor generalization ability; (2) Difficult to train – it is non-trivial for the discriminator and generator in a GAN to achieve Nash equilibrium [@heusel2017gans] during training; (3) Hard to evaluate – the evaluation of GANs can be considered as an effort to measure the dissimilarity between the real distribution $p_{r}$ and the generated distribution $p_{g}$. Unfortunately, the accurate estimation of $p_{r}$ is intractable. Thus, it is challenging to have a good estimation of the correspondence between $p_{r}$ and $p_{g}$. Challenges (1) and (2) are more concerned with computational aspects where much research has been carried out to mitigate these issues [@li2015generative; @salimans2016improved; @arjovsky2017wasserstein]. Challenge (3) is similarly fundamental, however limited literature is available and most of the current metrics only focus on measuring the dissimilarity between training and generated images. A more meaningful GAN evaluation metric that is consistent with human perceptions is paramount in helping researchers to further refine and design better GANs.
Although some evaluation metrics, e.g., Inception Score (IS), Kernel Maximum Mean Discrepancy (MMD) and Fréchet Inception Distance (FID), have already been proposed [@salimans2016improved; @heusel2017gans; @borji2018pros], they have a number of limitations. Firstly, these metrics do not agree with human perceptual judgments and human rankings of GAN models. A small artifact in images can have a large effect on the decision made by a machine learning system [@koh2017understanding], whilst the intrinsic image content does not change. In this aspect, we consider human perception to be more robust to adversarial images samples when compared to a machine learning system. Secondly, these metrics require large sample sizes for evaluation [@empirical-study; @salimans2016improved] and acquiring large-scale samples for evaluation sometimes is not realistic in real-world applications since generating them may be time-consuming. Finally, the existing metrics are not able to rank individual GAN-generated images by their quality i.e., metrics are generated on a collection of images rather than on a single image basis. The within-GAN variances are crucial because they can provide an insight into the variability of that GAN.
The current literature demonstrates that a CNN is able to predict neural responses in the inferior temporal cortex in an image recognition task [@yamins2014performance; @yamins2016using] via invasive BCI techniques [@waldert2016invasive]. The ways in which a CNN can be used to predict a neural response with a non-invasive BCI aspect is still an open question. Figure \[chap07-fig:invasive-noninvasive\] illustrates a schematic of different mesoscopic and macroscopic neural measurement techniques using invasive and non-invasive approaches. In this schematic, only EEG (Electroencephalography) is non-invasively measured from the human scalp [@mouraux2008across]. Other types of neural dynamics such as ECoG and LFP are measured invasively, which requires electrodes to be implanted. Compared to invasively measured neural dynamics, EEG pros are that it is a simple measurement, a non-painful experience during recording, easier to get ethics approval for and more easily generalized to real-world applications. However, EEG suffers challenges such as low signal quality (i.e., low SNR), low spatial resolution (interesting neural activities can span all of the scalp and are thus difficult to localise), all of which makes predicting EEG responses challenging.
![Schematic of different types of recorded neural signals (illustrated in (a)) via invasive and non-invasive measurements (illustrated in (b)). Figure from [@lago2017flexible].[]{data-label="chap07-fig:invasive-noninvasive"}](invasive_non_invasive.png){width="\textwidth"}
With the success achieved by deep neural networks (DNNs) in areas including computer vision and natural language processing, the operation and functionality of DNNs and its connection with the human brain has been extensively studied and investigated in the literature [@cichy2016comparison; @cichy2019deep; @groen2018distinct; @kuzovkin2018activations; @tu2018relating; @batista2018deep; @yamins2016using; @kriegeskorte2015deep; @kheradpisheh2016deep]. In this research area, the convolutional neural network (CNN) is widely studied and compared with the visual system in the human brain because both are hierarchical systems and the processing steps are similar. For example in an object recognition task, both CNNs and humans recognize an object by progressively extracting higher-level representations of the visual input through a hierarchy where successive layers operate on the inputs of the proceeding layers e.g., certain patterns of basic shapes, edges and colors as input can be determined at higher levels of the hierarchy to be a particular complex object composed of the inputs. Work reported in [@yamins2016using] outlines a CNN approach to delving even more deeply into understanding the development and organization of sensory cortical processing. It has recently been demonstrated that a CNN is able to reflect the spatio-temporal neural dynamics in the human brain visual processing area [@cichy2016comparison; @tu2018relating; @kuzovkin2018activations]. Despite much work carried out to reveal the similarity between CNNs and brain systems, research on interactions between CNNs and neural dynamics is limited.
In [@yamins2014performance] the authors demonstrate that a CNN matched with neural data recorded from the inferior temporal cortex of a human subject [@chelazzi1993neural] has high performance in an object recognition task. Given the evidence above that a CNN is able to predict neural responses in the brain, we explore the use of CNNs to predict P300 [@polich2007updating; @carrillo2000effect] amplitudes in this paper. This type of model can then produce (synthetic) EEG feedback for different types of GANs.
By applying advanced statistical and machine learning techniques to non-invasive EEG, better source localization and reconstruction becomes possible. Our previous work [@wang2018spatial; @wang2018review] demonstrated the effectiveness of using spatial filtering approaches for reconstructing P300 source ERP signals. Remaining low SNR issues can be further remedied by averaging EEG trials. Based on this evidence, we explore the use of DNNs to predict a metric we call Neuroscore [@wang2018use], when neural information is available via EEG.
In this work, we describe a metric called Neuroscore to evaluate the performance of GANs, which is derived from a neurophysiological response recorded via non-invasive electroencephalography (EEG). We demonstrate and validate a neural-AI interface (as seen in Figure \[fig:neuron\_AI\_interface1\]), which uses neural responses as supervisory information to train a CNN. The trained CNN model is then able to predict Neuroscore for images without requiring the corresponding neural responses. We test this framework via three models: Shallow convolutional neural network, Mobilenet V2 [@sandler2018mobilenetv2] and Inception V3 [@szegedy2016rethinking].
![Schematic of a neuro-AI interface. Stimuli (image stimuli used in this work) are simultaneously presented to an AI system and to participants. Participants’ neural responses are transferred to the AI system as supervised information for assisting the AI system to make decision.[]{data-label="fig:neuron_AI_interface1"}](NAI.pdf){width="100.00000%"}
In outline, Neuroscore is calculated via measurement of the P300, an event-related potential (ERP) present in EEG, via a rapid serial visual presentation (RSVP) paradigm. The P300 and RSVP paradigm are mature techniques in the brain-computer interface (BCI) community and have been applied in a wide variety of tasks such as image search [@gerson2006cortically], information retrieval [@mohedano2015exploring], and others. The unique benefit of Neuroscore is that it more directly reflects human perceptual judgment of images, which is intuitively more reliable compared to conventional metrics in the literature [@borji2018pros]. In summary, our contributions are two-fold:
- We combine human perception research with GANs and deep learning research. This represents a new avenue of investigation in the development of better GANs technologies.
- We propose a type of neuro-AI interface and training strategy to generalize the use of Neuroscore, which can be directly used for GAN evaluations without recording EEG. This enables our Neuroscore to be more widely applied to real-world scenarios, with a new measure we name synthetic-Neuroscore.
Related Work
============
Three well-known metrics are compared with Neuroscore in this paper.
Inception Score (IS)
--------------------
Inception Score is the most widely used metric in the literature [@salimans2016improved; @empirical-study; @borji2018pros]. It uses a pre-trained Inception network [@szegedy2016rethinking] as an image classification model $\mathcal{M}$ to compute $$\label{eq:IS-formula}
\mathrm{IS} = \exp\left(\mathbb{E}_{\bm{\mathrm{x}}\sim p_{g}} \left[\mathrm{KL}\left(p_{\mathcal{M}}\left(\mathrm{y} \vert \bm{\mathrm{x}}\right)\vert \vert p_{\mathcal{M}}(\mathrm{y})\right)\right]\right)$$ where $p_{\mathcal{M}}(\mathrm{y} \vert \bm{\mathrm{x}})$ is the label distribution of $\bm{\mathrm{x}}$ that is predicted by the model $\mathcal{M}$ and $p_{\mathcal{M}}(\mathrm{y})$ is the marginal probability of $p_{\mathcal{M}}(\mathrm{y} \vert \bm{\mathrm{x}})$ over the probability $p_{g}$. A larger inception score will have $p_{\mathcal{M}}(\mathrm{y} \vert \bm{\mathrm{x}})$ close to a point mass and $p_{\mathcal{M}}(\mathrm{y})$ close to uniform, which indicates that the Inception network is very confident that the image belongs to a particular ImageNet category [@deng2009imagenet] where all categories are equally represented. This suggests the generative model has both high quality and diversity.
Kernel Maximum Mean Discrepancy (MMD)
-------------------------------------
MMD is a method for comparing two distributions, in which the test statistic is the largest difference in expectations over functions in the unit ball of a reproducing kernel Hilbert space [@gretton2012kernel]. MMD is computed as $$\label{eq:MMD-formula}
\begin{aligned}
\mathrm{MMD}^{2}(p_{r}, p_{g}) = \mathbb{E}_{ {\substack{\bm{\mathrm{x}}_{r}, \bm{\mathrm{x}}_{r}^\top \sim p_{r} \\ {\bm{\mathrm{x}}_{g}, \bm{\mathrm{x}}_{g}^\top \sim p_{g} } } } } {\left[k(\bm{\mathrm{x}}_{r}, \bm{\mathrm{x}}_{r}^\top) - 2k(\bm{\mathrm{x}}_{r}, \bm{\mathrm{x}}_{g}) + k(\bm{\mathrm{x}}_{g}, \bm{\mathrm{x}}_{g}^\top)\right]}
\end{aligned}$$ It measures the dissimilarity between $p_{r}$ and $p_{g}$ for some fixed kernel function $k$, such as a Gaussian kernel [@li2015generative]. A lower MMD indicates that $p_{g}$ is closer to $p_{r}$, showing the GAN has better performance.
Fréchet Inception Distance (FID)
--------------------------------
FID uses a feature space extracted from a set of generated image samples by a specific layer of the Inception network [@heusel2017gans]. The feature space is modelled via a multivariate Gaussian by the mean $\bm{\mathrm{\mu}}$ and covariance $\bm{\mathrm{\Sigma}}$. FID is computed as $$\label{eq:FID-formula}
\mathrm{FID}(p_{r}, p_{g}) = \vert \vert \bm{\mathrm{\mu}}_{r} - \bm{\mathrm{\mu}}_{g} \vert \vert_{2}^{2} + \mathrm{Tr}\left(\bm{\mathrm{\Sigma}}_{r} + \bm{\mathrm{\Sigma}}_{g} - 2(\bm{\mathrm{\Sigma}}_{r}\bm{\mathrm{\Sigma}}_{g}\right)^{\frac{1}{2}})$$ Similar to MMD, lower FID is better, corresponding to more similar real and generated samples as measured by the distance between their activation distributions.
For Inception Score, the score is calculated through the Inception model [@szegedy2016rethinking]. It has been shown that Inception Score is very sensitive to the model parameters [@barratt2018note]. Even the score produced by the same model trained using different libraries (e.g., Tensorflow, Keras, PyTorch) differ a lot from each other. It also requires a large sample size for the accurate estimation for $p_{\mathcal{M}}(\mathrm{y})$. FID and MMD both measure the similarity between training images and generated images based on the feature space [@empirical-study], since the pixel representations of images do not naturally support for meaningful Euclidean distances to be computed [@forsyth2003modern]. The main concern about these two methods is whether the distributional characteristics of the feature space exactly reflect the distribution for the images [@koh2017understanding].
We list the supported features of Neuroscore and traditional metrics in Table \[tab:metric-compare\]. Neuroscore can not only evaluate image quality as can the other metrics, but also have $3$ unique characteristics, which will be demonstrated in Section \[results\].
[Feature]{} [IS]{} [MMD]{} [FID]{} [**Neuroscore**]{}
---------------------------- -------- --------- --------- --------------------
[Evaluate image quality]{}
[Consistent with human]{}
[Small sample size]{}
[Rank images]{}
: Comparison between Neuroscore and other metrics.[]{data-label="tab:metric-compare"}
Preliminaries
=============
Generative Adversarial Networks
-------------------------------
A generative adversarial network (GAN) has two components, the discriminator $D$ and the generator $G$. Given a distribution $\pmb{z} \sim p_{\pmb{z}}$, $G$ defines a probability distribution $p_{g}$ as the distribution of the samples $G(\pmb{z})$. The objective of a GAN is to learn the generator’s distribution $p_{g}$ that approximates the real data distribution $p_{r}$. Optimization of a GAN is performed with respect to a joint loss for $D$ and $G$ as $$\label{eq:GAN-formula}
\min \limits_{G} \max \limits_{D} \mathbb{E}_{\bm{\mathrm{x}} \sim p_{r}} \mathrm{log}[D(\bm{\mathrm{x}})] + \mathbb{E}_{\bm{\mathrm{z}} \sim p_{\bm{\mathrm{z}}}} \mathrm{log}\left[1 - D(G(\bm{\mathrm{z}}))\right]$$
P300 (or P3) Component and Preprocessing {#sec:P300}
----------------------------------------
In neuroscience, the P300 ERP component refers to a voltage change measured on the scalp which arises from current flow changes in the brain in response to a target stimulus [@polich2007updating], that can be measured with EEG. It reflects a participant’s attention, which can be modulated by the specific instruction given to a participant. It has been shown in previous studies that real face stimuli generate larger P300/LPP potentials than non-real face stimuli such as cartoon face images [@schindler2017differential; @ling2012comparative; @zhao2019event]. Furthermore, the P300/LPP increases linearly with face realism, reflecting increased activity in visual and parietal cortex for more realistic faces[@schindler2017differential]. The P300 response elicited by a target stimulus is typically evident between $300$ – $600$ ms post stimulus presentation depending on the type of task. EEG is normally recorded by using multiple channels e.g.. 32 channels, which makes it difficult to estimate the P300 source amplitude. We use an LDA beamformer [@treder2016lda; @wang2018spatial] to reconstruct the P300 source signal from the recorded raw EEG epochs.
Briefly, given a target EEG epoch $\bm{\mathrm{X}}_{i} \in \mathbb{R}^{C \times T}$ and a standard EEG epoch[^1] $\bm{\mathrm{K}}_{i} \in \mathbb{R}^{C \times T}$ ($C$ is the number of channels and $T$ is time points in each EEG epoch). The optimization problem for the LDA beamformer is to find a projection vector $\bm{\mathrm{w}} \in \mathbb{R}^{C \times 1}$ that solves the optimization problem: $$\label{eq:beam_cost_function}
\min \limits_{\bm{\mathrm{w}}} \bm{\mathrm{w}}^\top \bm{\mathrm{\Sigma}} \bm{\mathrm{w}} \hspace{5pt} \mathrm{s.t.} \bm{\mathrm{w}}^\top\bm{\mathrm{p}}=1
$$ where $\bm{\mathrm{\Sigma}} \in \mathbb{R}^{C \times C}$ is the EEG epoch covariance matrix ($\bm{\mathrm{\Sigma}} = \frac{1}{N} \sum_{i=1}^{N}\bm{\mathrm{X}}_{i}\bm{\mathrm{X}}_{i}^\top$, $N$ is number of trials) and $\bm{\mathrm{p}}\in \mathbb{R}^{C \times 1}$ is the spatial pattern difference between target and standard condition [@treder2016lda]. The closed-form solution is $$\label{eq:w}
\bm{\mathrm{w}}=\bm{\mathrm{\Sigma}}^{-1} \bm{\mathrm{p}} ( \bm{\mathrm{p}}^\top \bm{\mathrm{\Sigma}}^{-1} \bm{\mathrm{p}}) ^{-1}$$ The source signal of each single-trial $\bm{\mathrm{s}}$ can be obtained as $$\label{eq:S_P300}
\bm{\mathrm{s}}= \bm{\mathrm{w}}^\top \bm{\mathrm{X}}_{i} =
(\bm{\mathrm{p}}^\top \bm{\mathrm{\Sigma}}^{-1} \bm{\mathrm{p}} )^{-1} \bm{\mathrm{p}}^\top \bm{\mathrm{\Sigma}}^{-1} \bm{\mathrm{X}}_{i}$$ where $\bm{\mathrm{s}} \in \mathbb{R}^{1 \times T}$. Hence, LDA beamformer enables transformation of multi-channel EEG epochs to single-channel EEG epochs facilitating more robust measurement of the P300 and its amplitude.
Methodology
===========
Data Acquisition and Experiment
-------------------------------
We used three GAN models to generate synthetic images of faces: DCGAN [@radford2015unsupervised], BEGAN [@berthelot2017began] and progressive growing of GANs (PROGAN) [@karras2017progressive] with sample outputs shown in Figure \[fig-face-example\].
Image streams in the experiment contain generated images from DCGAN, BEGAN and PROGAN, as well as real face (RFACE) images and non-face category images. RFACE images were sampled from the CelebA dataset [@liu2015faceattributes]. Non-face category (standard images) were sampled from the ImageNet dataset [@deng2009imagenet], similar to those used in other RSVP experiments such as [@healy2017eeg; @nails]. EEG data for 12 participants was gathered. Data collection was carried out with approval from Dublin City University Research Ethics Committee (REC/2018/115). Each participant completed two types of task which we call the behavioral experiment (BE) task and the rapid serial visual presentation (RSVP) task. The sequence of blocks presented in the experiment was: BE $\rightarrow$ RSVP $\rightarrow$ BE $\rightarrow$ RSVP $\rightarrow$ BE. The presented images were randomly shuffled (across and within blocks), meaning the appearance of face images could not be predicted ahead of time by a participant i.e., they occurred at random times but always in the same quantity.
The objective of the BE task was to record participants’ responses to each type of image category while the RSVP task was to record EEG when participants were viewing the rapid presentation of images. The ultimate goal of this study was to compare whether the EEG responses in the RSVP task were consistent with participants’ responses in the BE task.
An example of the RSVP experimental protocol is shown in Figure \[fig:rsvp\_paradigm\].
![An example of RSVP experimental protocol used in this work. A rapid image stream containing targets and standards (non-target) is presented to participants at 4 Hz (4 images per second) presentation rate.[]{data-label="fig:rsvp_paradigm"}](rsvp_paradigm.png){width=".7\textwidth"}
The RSVP task contained 26 blocks. Each RSVP block contained 240 images (6 images for each face category thus 24 face targets in total and 216 non-face images), thus there were 6,240 images (624 face targets / 5,616 non-face images) available for each participant. In the RSVP task, image streams were presented to participants at a 4 Hz presentation rate. Participants in RSVP blocks were asked to search for real face (RFACE) images[^2]. This instruction to participants was constructed so that they would maintain attention to detect face images (from all GAN types), and furthermore focus their attention to what they perceived as real face images [@carrillo2000effect]. Details of the experiment can be found in [@wang2018use].
EEG was recorded from participants in both the BE and RSVP tasks along with timestamp information for image presentation and behavioural responses (via a photodiode and hardware trigger) to allow for precise epoching of the EEG signals for each trial [@wang2016investigation]. EEG data was acquired using a 32-channel BrainVision actiCHamp at 1,000 Hz sampling frequency, using electrode locations as defined by the International 10-20 system. To enhance the low signal-to-noise ratio of the acquired EEG, pre-processing is required. Pre-processing typically involves re-referencing, filtering the signal (by applying a bandpass filter to remove environmental noise or to remove activity in non-relevant frequencies), epoching (extracting a time epoch typically surrounding the stimulus onset) and trial/channel rejection (to remove those containing artifacts). In this work, a common average reference (CAR) was utilized and a bandpass filter (i.e., 0.5-20 Hz) was applied prior to epoching. EEG data was then downsampled to 250 Hz. Only trials where behavioral responses occurred between 0 and 1 second after the presentation of a stimulus were used. Trial rejection was carried out to remove those trials containing noise such as eye-related artifacts (via a peak-to-peak amplitude threshold across all electrodes).
Neuroscore
----------
We used a rapid serial visual presentation (RSVP) paradigm [@wang2016investigation; @wang2018review; @healy2020experiences] to elicit the P300 ERP. Our experimental procedure is illustrated in our previous published work [@wang2018use]. We average the single-trial P300 amplitude (as Neuroscore) to mitigate the background EEG noise [@polich2007updating], which renders a stable measurement of the EEG response to a typical type of stimulus. In general, our Neuroscore is calculated via two steps: (1) Reconstruct the P300 source signal from the raw EEG; (2) Average the P300 amplitude of each reconstructed single-trial source signal across trials (see Algorithm \[al:calculate-NS\]).
- $\bm{\mathrm{X}} \in \mathbb{R}^{N \times C \times T}$ is the EEG signal corresponding to the target stimulus, where $N$ is the number of target trials, $C$ is the number of channels, and $T$ is the number of time points.
- $\bm{\mathrm{K}} \in \mathbb{R}^{M \times C \times T}$ is the EEG signal corresponding to the standard stimulus, $M$ is number of standard trials, $C$ is number of channels, $T$ is number of time points. The target and standard EEG trials are already explained in section \[sec:P300\].
Neuroscore $\bm{\mathrm{\Sigma}} = \frac{1}{N} \sum_{i=1}^{N}\bm{\mathrm{X}}_{i}{\bm{\mathrm{X}}_{i}}^\top + \frac{1}{M} \sum_{i=1}^{M}\bm{\mathrm{K}}_{i}{\bm{\mathrm{K}}_{i}}^\top$ $\bm{\mathrm{p}}=\frac{1}{N}\sum_{i=1}^{N}\bm{\mathrm{X}}_{i, t_{i}} - \frac{1}{M} \sum_{i=1}^{M}\bm{\mathrm{K}_{i, t_{i}}}$ $t_{optimal}$=$\mathrm{argmin}_{t_{i}} \mathrm{J}$ $\bm{\mathrm{w}}_{optimal}$=$\mathrm{W}_{t_{optimal}}$ $\mathrm{t}_{P300}$=\[$\mathrm{t}_{optimal}$ - $100$ ms, $\mathrm{t}_{optimal}$ + $100$ ms\] *This is time window being detected for P300.* $\mathrm{Neuroscore}$ = $\dfrac{1}{N}\sum_{i=1}^{N}\mathrm{A}_{i}$
The proposed Neuroscore reflects a human’s perceptual response to different GANs via EEG measurements, thus it is consistent with human perceptual judgment on GANs. Figure \[fig:Neuroscore\_performance\] demonstrates the performance of Neuroscore calculated from a human neural response. In Figure \[fig:Neuroscore\_performance\](a), it can be seen that different image categories activate different P300 responses. Figure \[fig:Neuroscore\_performance\](b) illustrates a strong correlation between Neuroscore and human judgment (BE accuracy)[^3]. These results demonstrate that Neuroscore reflects human judgment perception. More details can be found in our previous work [@wang2018use].
Neuro-AI Interface
------------------
We propose a neuro-AI interface in order to generalize the use of Neuroscore. This kind of framework interfaces between neural responses and AI systems (a CNN is used in this study), which use neural responses as supervised information to train a CNN. The trained CNN is then used for generating a **synthetic-Neuroscore** given images generated by one of the popular GAN models i.e., average the outputs of corresponding images. Figure \[fig:inspiration\].
{width="1.\textwidth"}
demonstrates the schematic of the neuro-AI interface used in this work[^4]. Flow 1 shows that the image processed by a human being’s brain produces a single-trial P300 source signal for each input image. Flow 2 in Figure. \[fig:inspiration\] demonstrates a CNN with included EEG signals during the training stage. The convolutional and pooling layers process the image similarly as retina has done [@mcintosh2016deep]. *It should be noted that a CNN model is trained by using single images with their corresponding **single-trial** EEG information (including single-trial P300 signal and single-trial P300 amplitude[^5]*. Fully connected layers (FC) 1 – 3 aim to emulate the brain’s functionality that produces the EEG signal. The yellow dense layer in the architecture aims to predict the single-trial P300 source signal at $400$ – $600$ ms in response to each image input. In order to help the model make a more accurate prediction for the single-trial P300 amplitude for the output, the single-trial P300 source signal at $400$ – $600$ ms is fed to the yellow dense layer to learn parameters for the previous layers in the training step. The model was then trained to predict the single-trial P300 source amplitude (the red point shown in signal-trail P300 source signal of Figure \[fig:inspiration\]).
Training Details {#sec:trianing-details}
----------------
Mobilenet V2, Inception V3 and Shallow network (architecture of Shallow network refers to Figure \[fig:SCNN\_architecture\])
![Shallow network architecture used in this work.[]{data-label="fig:SCNN_architecture"}](SCNN_architecture.png){width="\textwidth"}
were explored in this work, where in flow 2 we use these three network bones such as Conv1-pooling layers. For Mobilenet V2 and Inception V3, we used ImageNet pre-trained parameters up to the FC 1 (as shown in Figure \[fig:inspiration\]). Table \[tab:FC\] shows the FC layers details of three networks. Due to no pretrained parameters in the Shallow net, only three FC layers are contained in order to avoid overfitting. We trained parameters from FC 1 to FC 4 for Mobilenet V2 and Inception V3. $\bm{{\mathrm{\theta}}}_{1}$ is used to denote the parameters from FC 1 to FC 3 and $\bm{{\mathrm{\theta}}}_{2}$ indicates the parameters in FC 4. For the Shallow model, parameters up to FC 2 represent $\bm{{\mathrm{\theta}}}_{1}$ and parameters in FC 3 indicate $\bm{{\mathrm{\theta}}}_{2}$.
------------- -------------- ------------- ----------- ---------
Model FC 1 FC 2 FC 3 FC 4
Shallow net (1024, 512) (512, 50) (50, 1) NA
Mobilenet (1792, 896) (896, 448) (448, 50) (50, 1)
Inception (2048, 1024) (1024, 512) (512, 50) (50, 1)
------------- -------------- ------------- ----------- ---------
: FC layers details of three networks investigated in this study.[]{data-label="tab:FC"}
We added EEG to the model because we first want to find a function $f(\chi) \to \bm{\mathrm{s}}$ that maps the images space $\chi$ to the corresponding single-trial P300 source signal $\bm{\mathrm{s}}$. This prior knowledge can help us to predict the single-trial P300 amplitude in the second learning stage.
We compared the performance of the models with, without EEG signals and with randomized EEG signals for training. We defined two stage $\mathrm{loss}$ function ($\mathrm{loss}_{1}$ for a single-trial P300 source signal in the $400$ – $600$ ms time window and $\mathrm{loss}_{2}$ for single-trial P300 amplitude) as $$\begin{aligned}
\mathrm{loss_{1}}(\bm{\mathrm{\theta}}_{1}) &= \frac{1}{N} \sum_{i=1}^{N} \lVert \bm{\mathrm{S}}_{i}^{true} - \bm{\mathrm{S}}_{i}^{pred}(\bm{\mathrm{\theta}}_{1})\rVert ^{2}_{2}\\
\mathrm{loss_{2}}(\bm{\mathrm{\theta}}_{1},\bm{\mathrm{\theta}}_{2}) &= \frac{1}{N} \sum_{i=1}^{N}(\mathrm{y}_{i}^{true} - \mathrm{y}_{i}^{pred}(\bm{\mathrm{\theta}}_{1},\bm{\mathrm{\theta}_{2}}))^{2}
\end{aligned}$$ where $\bm{\mathrm{S}}_{i}^{true} \in \mathbb{R}^{1 \times T}$ is the single-trial P300 signal in the $400$ - $600$ ms time window to the presented image, $\mathrm{y}_{i}$ refers to the single-trial P300 amplitude for each image, and $N$ refers to the batch size. In this case, we trained 20 epochs with batch size equaling to 256. An Adam optimizer with default hyperparameters was used and learning rate is 0.001.
The training of the models without using EEG signal is straightforward, models were trained directly to minimize $\mathrm{loss}_{2}(\bm{{\mathrm{\theta}}}_{1},\bm{{\mathrm{\theta}}}_{2})$ by feeding images and the corresponding single-trial P300 amplitude. In this case, training is an end-to-end process i.e., from an image to single-trial a P300 amplitude without considering stage 1. The reason that we do this is to investigate the significance of adding single-trial P300 signal as supervisory information to the network. Training with EEG information is explained in Algorithm \[al:training\]
**Input:** Images and averaged P300 signal $\bm{\mathrm{S}}_{i}^{true}$.
**Input:** Images and single-trial P300 amplitude $\mathrm{y}_{i}^{true}$.
and visualized in the “Flow 2" of Figure \[fig:inspiration\] with two stages. Stage 1 learns parameters $\bm{{\mathrm{\theta}}}_{1}$ to predict P300 source signal while stage 2 learns parameters $\bm{\mathrm{\theta}}_{2}$ to predict single-trial P300 amplitude with $\bm{{\mathrm{\theta}}}_{1}$ fixed.
Results
=======
EEG and Model Performance {#sec:Models-performance}
-------------------------
#### Individual Participant Performance
Three models have been validated for each individual participant as shown in Figure \[fig:model\_loss\_individual\].
It can be seen that all three models trained with EEG outperform the models trained without EEG. **In other words, we show that including EEG/P300 time series signals as supervisory information to the yellow dense layer yields an improvement in performance as seen in Figure \[fig:inspiration\]**. with smaller error and standard deviation across almost all individual subjects. For those cases where the reverse is true (7 from 36 have better or equal performance without EEG), this might result from the number of EEG trials for an individual participant not being sufficient enough for training of deep networks to learn the mapping function $f(\chi)$ from image to EEG.
Error mean (std)
-- ------------------ --------------------
Shallow-EEG 0.151 ($\pm$0.245)
Shallow 0.428 ($\pm$0.623)
Mobilenet-EEG 0.155 ($\pm$0.235)
Mobilenet 0.437 ($\pm$0.589)
Inception-EEG 0.157 ($\pm$0.487)
Inception 0.462 ($\pm$0.932)
: Details of error mean and standard deviation for Figure \[fig:model\_loss\_individual\].[]{data-label="tab:model_loss_individual"}
#### Cross Participant Performance
We evaluated the cross participant performance of our approach by pooling trials across participants to see if the use of pooled trials produced a smaller error. In this case, the number of EEG trials across participants is 6012. We split data into training and testing as 2:1 in which there are 4008 trials for training and 2004 trials for testing. All trials are randomly shuffled and we repeat this process for 20 times in order to get a more robust result.
Table \[tab:model\_loss\_cross\] shows the error for each model with the EEG signal, with a randomized EEG signal **within each type of GAN** and without an EEG signal. All models with EEG signals perform better than models without EEG signals, with much smaller errors and standard deviation.
Error mean(std)
-- ----------------------------------- ------------------------
Shallow-EEG **0.209 ($\pm$0.102)**
Shallow-EEG$\mathrm{_{random}}$ 0.348 ($\pm$0.114)
Shallow 0.360 ($\pm$0.183)
Mobilenet-EEG **0.198 ($\pm$0.087)**
Mobilenet-EEG$\mathrm{_{random}}$ 0.404 ($\pm$0.162)
Mobilenet 0.366 ($\pm$0.261)
Inception-EEG **0.173 ($\pm$0.069)**
Inception-EEG$\mathrm{_{random}}$ 0.392 ($\pm$0.057)
Inception 0.344 ($\pm$0.149)
: Errors for $9$ models across the 12 participants (“\*-EEG" indicates models are trained with paired EEG, “\*-EEG$\mathrm{_{random}}$" refers to EEG trials which are randomized in the $\mathrm{loss}_{1}$ **within each type of GAN**). Results are averaged by shuffling training/testing sets $20$ times.[]{data-label="tab:model_loss_cross"}
Adding the EEG signal to the intermediate layer reduces error in all three models (as the same error is shown in Figure \[fig:model\_loss\_individual\]), namely $0.151$, $0.168$ and $\textbf{0.171}$ for Shallow, Mobilenet, and Inception respectively. This indicates that the Inception model benefits most when adding EEG signal into the training stage. The performance of models with the EEG signal is ranked as Inception-EEG followwd by Mobilenet-EEG, and Shallow-EEG, which indicates that deeper neural networks may achieve better performance in this task. We used the randomized EEG signal here as a baseline to determine the efficacy of adding the EEG signal to produce better Neuroscore output. When randomizing the EEG signal, it shows that the error for each three model increases significantly. For Mobilenet and Inception, the error with the randomized EEG signal is even higher than those without the EEG signal in the training stage, demonstrating that EEG information in the training stage is crucial to each model.
Figure \[fig:cross-subject-correlation\] shows that the models with EEG information have a stronger correlation between synthetic-Neuroscore and Neuroscore. The cluster (blue, orange, and green circles) for each category of the model trained with EEG (left column) is more separable than the cluster produced by model without EEG (right column). This indicates that when with EEG is used in training models Neuroscore is more accurate and that Neuroscore is able to rank the performances of different GANs, which cannot be achieved with other metrics [@borji2018pros].
![Scatter plot of synthetic-Neuroscore (vertical axis) and Neuroscore (horizontal) for $6$ models (Shallow, Mobilenet, Inception with and without EEG signals for training) across participants, with $20$ times repeated shuffling training and testing set. Each circle represents the cluster for a specific category. Small triangle markers inside each cluster correspond to each shuffling process. The dot at the center of each cluster is the mean.[]{data-label="fig:cross-subject-correlation"}](cross_subject_correlation.png){width=".7\textwidth"}
Neuroscore Aligns with Human Perception {#sec:compare-to-other-metrics}
---------------------------------------
Figure \[fig:Neuroscore\_performance\](b) shows the correlation between Neuroscore and human judgment (BE accuracy) according to three GANs: BEGAN, DCGAN, and PROGAN. The statistical test demonstrates the strong correlation between those two variables. This indicates that Neuroscore can be used to evaluate GANs as it reflects human perceptual judgment. A number of previous studies have noted that increasing task difficulty reduces the amplitude of the P300 [@kim2008influence; @marathe2013novel; @senkowski2002effects; @scharinger2017comparison]. It may be the case that the larger P300 amplitudes observed for the PROGAN images indicate that these face images were easier to detect compared to the images from the other GANs. For example, DCGAN images tended to contain far more visual aberrations and other inherent artefacts that would impede their detection [@wolfe2010reaction]. It has also been noted in another prior study that increased sensory evidence results in shorter reaction times and larger component amplitudes in temporal and spatial regions coinciding with those examined in our work [@philiastides2014human]. Another prior study explains larger P300 amplitudes for real face images resulting from enhanced perceptual processing [@schindler2017differential]. In effect, larger average P300/LPP amplitudes for a particular GAN type are indicative that its images are perceived as being real faces.
We have already demonstrated that the Neuroscore derived from raw EEG is consistent with human perception [@wang2018use]. We will now demonstrate the same property of synthetic-Neuroscore predicted from the neuro-AI interface. We compare the synthetic-Neuroscore with three widely used evaluation metrics. The ultimate goal of GANs is to generate images that are indistinguishable from real images by human beings. Therefore, consistency between an evaluation metric and human perception is a critical requirement for the metric to be considered good. Table \[tab:traditional-method-score\] shows the comparison between synthetic-Neuroscore and three traditional scores. To be consistent with all the scores (smaller score indicates better GAN), we used 1/IS and 1/synthetic-Neuroscore for comparisons in Table \[tab:traditional-method-score\]. It can be seen that people rank the GAN performance as PROGAN $>$ BEGAN $>$ DCGAN. All three synthetic-Neuroscores produced by the three models with EEG are consistent with human judgment while the other three conventional scores are not (they all indicate that DCGAN outperforms BEGAN).
DCGAN BEGAN PROGAN
-- ----------------- --------------- --------------- --------------
[0.44]{} [0.57]{} [0.42]{}
[0.22]{} [0.29]{} [0.12]{}
[63.29]{} [83.38]{} [34.10]{}
1/Shallow-EEG [**1.60**]{} [**1.39**]{} [**1.14**]{}
1/Mobilenet-EEG [**1.71**]{} [**1.29**]{} [**1.20**]{}
1/Inception-EEG [**1.51**]{} [**1.34**]{} [**1.24**]{}
[**0.995**]{} [**0.824**]{} [**0.705**]{}
: Three conventional scores: Inception Score (IS), Maximum Mean Discrepancy (MMD), Fréchet Inception Distance (FID), and synthetic-Neuroscore produced by three models with EEG for each GAN category. A lower score indicates better performance of the GAN. Neuroscore is consistent with human judgments. Bold text indicates the consistency with human judgment (BE) accuracy.[]{data-label="tab:traditional-method-score"}
Synthetic-Neuroscore Needs Far Fewer Samples
--------------------------------------------
The number of samples needed for evaluation of a GAN is crucial in real-world applications considering computational efficiency and efforts needed for labeling and annotation. Traditional metrics need a large sample size to capture the underlying statistical properties of the real and generated images [@salimans2016improved; @empirical-study]. In practice, we should prefer a metric that is robust when dealing with small sample sizes i.e., where small sample sizes can produce good estimates. Figure \[fig:gen-test\](b) shows that synthetic-Neuroscore converges stably at around $20$ presentations of a specific image (for signal-enhancement purposes), which is far fewer than the thousands of images required by traditional methods [@borji2018pros; @empirical-study]. This is due to the fact that the LDA-beamformed single-trial P300 amplitude becomes stable when as few as dozens of EEG trials corresponding to one category are available [@mouraux2008across].
![Synthetic-Neuroscore for different evaluated sample sizes for each type of GAN. 200 repeated measurements have been made by randomly shuffling image samples.[]{data-label="fig:gen-test"}](sample_test_NS_dist.png){width=".6\textwidth"}
Synthetic-Neuroscore Can Rank Images
------------------------------------
Another property of using synthetic-Neuroscore is the ability to indicate the quality of an individual image. Traditional evaluation metrics are unable to score each individual image for two reasons. Firstly they need large-scale samples for evaluation and secondly most methods (e.g., MMD and FID) evaluate GANs based on the dissimilarity between real images and generated images so they are not able to score the generated images individually. For our proposed method, the score of each single image can also be evaluated as a synthetic single-trial P300 amplitude measurement. We demonstrate in Figure \[fig:rank-single-img\] how the predicted single-trial P300 amplitude conveys perceptual quality at the level of individual images.
![P300 for each single image predicted by the proposed neuro-AI interface in our paper. Higher predicted P300 indicates better image quality.[]{data-label="fig:rank-single-img"}](ns_changing_image.pdf){width=".7\textwidth"}
This property provides synthetic-Neuroscore with a novel capability for tracking variations in image output quality within a typical GAN. Although synthetic-Neuroscore and IS are both generated from deep neural networks, synthetic-Neuroscore is more suitable than IS for evaluating GANs as it is a direct reflection of human perception and fewer sample images are required for evaluation. This has benefits in terms of improved explanation of output than that offered by IS. For example low ranked images can be selected at evaluation time to illustrate cases where the GAN under evaluation is performing poorly.
Conclusions
===========
In this paper, we outline a metric for evaluating the quality of the outputs from GANs called Neuroscore. Furthermore, we describe a neuro-AI interface to calculate a synthetic-Neuroscore for evaluating GAN performance that only requires EEG signals as supervisory information during model training. Three deep network architectures are explored and our results demonstrate that including neural responses during the training phase of the neuro-AI interface improves its accuracy even though neural measurements are absent when evaluating on a test set. This means that human subjects are not actually needed to evaluate the output from a test GAN, their neural responses are needed only when training the model that produces a synthetic-Neuroscore.
We compared our synthetic-Neuroscore measure to three traditional evaluation metrics and demonstrated the unique advantages of synthetic-Neuroscore, that it is consistent with human perception, that it requires far fewer image samples for calculation and that it can rank individual images in terms of quality, within a specific GAN.
In this work, we demonstrated the use of CNNs to synthesize the neural response. More complicated neural architectures such as mixture of CNNs and recurrent neural networks can be investigated in future work when more EEG data is available.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is funded as part of the Insight Centre for Data Analytics which is supported by Science Foundation Ireland under Grant Number SFI/12/RC/2289.
[^1]: A target EEG epoch is an EEG trial (time duration 0 – 1 s) which corresponds to a target stimulus i.e., DCGAN, BEGAN, PROGAN and RFACE images in this study. A standard/non-target EEG epoch is an EEG trial which corresponds to a non-target images i.e., non-face image in this work.
[^2]: P300 responses were elicited for all GAN image categories e.g., while DCGAN had almost perfect behavioral accuracy labelled as being ‘fake’, DCGAN images still elicited a P300.
[^3]: BE accuracy is the recorded accuracy (calculated as the number of correctly labeled images divided by the total number of images) in the behavioral experiment. Normalized BE accuracy is calculated by subtracting the average accuracy (across GAN types for that participant) from BE accuracy.
[^4]: We understand that a human being’s brain system is much more complex than demonstrated in this work and that information flow in the brain is not one-directional [@she2016evaluating; @she2018reduced]. Our framework can be further extended to be more biologically plausible.
[^5]: Single-trial P300 amplitude refers the maximum value in the 400 ms – 600 ms time window of a single-trial EEG signal. Details can be referred to our previous work [@wang2018use].). The averaged output of a trained model in terms of one image category can be represented as the **synthesized Neuroscore** (we refer to it as **synthetic-Neuroscore** in this paper)
|
{
"pile_set_name": "ArXiv"
}
|
Introduction
============
Global symmetries of the standard model (SM), like baryon and lepton numbers, can be broken by quantum gravity effects [@qg]. In order to avoid an unacceptably large proton decay rate induced by this symmetry breaking we can elevate the conservation of baryon number to a local symmetry [@mu:1]. Moreover, there are string models that also gauge baryon number to protect the proton from decaying through higher-dimensional operators. In particular, models with low string scale, $E_{\rm string} \sim {\cal O}$(TeV), generally exhibit baryon number as a local symmetry [@st:low].
In this work, we analyze the signals of a light gauge boson coupling to baryon number at the HERA and Tevatron colliders. We assumed that the $U(1)_B$ symmetry is spontaneously broken, giving a mass ($m_B$) to $\gamma_B$, and that the mixing between $\gamma_B$ and the electroweak gauge bosons is negligible [@mu:2]. In the absence of mixing, the $\gamma_B$ boson always decays into quark pairs; its signature is thus the presence of 2 jets with an invariant mass close to $m_B$. A previous analysis [@mu:3] studied its effect on $Z$-pole physics and constrained its coupling to be $\alpha_B \lesssim
0.2$ for masses $m_B \le m_Z$. Moreover, $\Upsilon(1S)$ decays are modified by the $\gamma_B$ boson, leading to stronger constraints on $\alpha_B$ for $m_B \lesssim 30$ GeV [@mu:3].
In our analysis we concentrate on the decay $\gamma_B \to b \bar{b}$ in order to reduce the QCD backgrounds. Using a muon trigger, we show that the study of the production of $b \bar{b}$ pairs at the (upgraded) Tevatron can exclude $\gamma_B$ bosons with masses ($m_B$) in the range $40 \lesssim m_B \lesssim 300$ GeV for couplings $\alpha_B \gtrsim 2 \times 10^{-2}$ ($3 \times 10^{-3}$). A displaced vertex trigger could increase the sensitivity at the upcoming Main Injector (MI) run by another factor of three. We also use the available search by the CDF collaboration for particles decaying into $b \bar{b}$ pairs [@cdf:bb]. Since this search uses a jet trigger which needs to be pre-scaled if the jet energies are not very large, the resulting limit turns out to be weaker than the sensitivity limit of current data based on a muon trigger. We also analyze $\gamma_B$ production at HERA, however, the potential bounds derived from this study are weaker than the $Z$–pole or Tevatron ones due to the limited integrated luminosity.
In principle, $\gamma_B$ boson exchange in the $t-$channel can give rise to events presenting a large rapidity gap between two jets [@mu:3]. Recently DØ studied the gap fraction as a function of $E_T$ [@jill]. Our fit to this data in terms of $\gamma_B$ exchange gives $\gamma_B$ masses very close to zero and $\alpha_B
\simeq 0.06$. However, this region is already excluded by the analysis of $\Upsilon(1S)$ decays [@mu:3]. Moreover, the HERA data on rapidity gaps can only be explained for $m_B \simeq 0$ and very large $\alpha_B$. Therefore, the present data on the production of rapidity gaps cannot be explained by the existence of a $\gamma_B$ boson.
In our analyses we assumed that the interaction between quarks and $\gamma_B$ is described by the effective Lagrangian $${\cal L}_{\text{eff}} = \frac{1}{3} \sqrt{4 \pi \alpha_B}~ \bar{q}
\gamma^\mu q~ B_\mu \;\; ,$$ where we denote the quark fields by $q$ and the $\gamma_B$ field by $B_\mu$. Neglecting the fermions masses, the $\gamma_B$ width is given by $$\Gamma_B = \frac{N}{9}~ \alpha_B~ m_B \;\; ,$$ where $N$ is the number of quark pairs to which $\gamma_B$ can decay.
Direct Searches at Tevatron and HERA
====================================
In hadronic collisions $\gamma_B$ bosons can be produced in the $s$-channel via quark–antiquark annihilation, leading to two-jet events. In order to reduce the large QCD backgrounds we focused on $\gamma_B$ decays into $b
\bar{b}$ pairs ($p \bar{p} \to q \bar{q} \to \gamma_B \to b \bar{b} $). The main backgrounds to this process are QCD $b \bar{b}$ production, mistagged QCD jets, and $Z$ production for $m_B$ close to $M_Z$. Since the two-jet event rate is too large for the present data acquisition system to be able to analyze all events, we need to choose a trigger to select a subset of the data. In our analyses of the Run I potential for $\gamma_B$ searches, we considered only events containing a muon (from $b$ decay) with $p_T^{\mu} >
7.5$ GeV and $ |\eta_\mu| < 0.9$ (2 for the upgraded Tevatron) [@cdf:zbb]. There is one useful byproduct of this requirement: together with the requirement that both $b$’s be vertex tagged, the presence of this muon reduces the mistag background to a negligible level. Moreover, we also demanded that
- the jets should have $p_T > 15$ GeV;
- the jets should be separated by $\Delta R > 0.4$.
With these requirements the $b$ tag efficiency is 0.24. In our analyses we evaluated the scattering amplitudes of the signal and backgrounds using the package Madgraph [@madg], taking into account the interference between $\gamma_B$ and gluon$/Z$ exchange diagrams.
We considered that a point of the $(\alpha_B , m_B)$ plane is within the reach of an experimental search, if the predicted signal has a $3\sigma$ significance when we restrict ourselves to $b \bar{b}$ invariant masses in the range $ m_B - 10 \ {\rm GeV} < m_{b\bar{b}} <
m_B + 10$ GeV. We chose this rather high confidence level for an exclusion limit since our analysis does not allow for experimental resolutions or efficiencies (other than the $b-$tagging efficiency). We note, however, that CDF recently found [@cdf:zbb] preliminary evidence for $Z \to b \bar{b}$ decays in their Run I data sample using cuts very similar to the ones applied by us. We show in Fig. \[fig:bb\]a the region in the $(\alpha_B , m_B)$ plane which could be excluded by the CDF Run I data, [*i.e.*]{} for an integrated luminosity of 110 pb$^{-1}$ and a $b$-tagging efficiency of 24%, if no signal is found. In this figure we also display the effect of having a larger integrated luminosity (2 fb$^{-1}$) and an extended rapidity acceptance for the muons ($|\eta_\mu | <2$) at the MI. Notice that our $p_T$ and $\Delta R$ cuts constrain the invariant mass of the $b \bar{b}$ pair in two-jet events to be larger than 30 GeV.
In order to explore smaller $\gamma_B$ masses we considered its production in association with a jet. The processes that we analyzed are $$\begin{aligned}
p \bar{p} & \to \gamma_B g &\to b \bar{b} g \; ,
\\
p \bar{p} & \to \gamma_B q &\to b \bar{b} q \; ,\end{aligned}$$ where $q$ can be any quark or antiquark. We present in Fig.\[fig:bb\]b the potential limits on $\alpha_B$ and $m_B$ originating from the study of $b$-$\bar{b}$-jet production for Run I and at the MI. In this analysis we applied the same cuts for the $b
\bar{b}$ system and required the extra jet to have $p_T > 10 $ GeV and to be separated from the $b$ jets by $\Delta R > 0.4$.
It is also possible to search for $\gamma_B$ using the Run I data but triggering on jets with a minimum $E_T$ [@cdf:bb]. This choice of trigger requires pre-scaling, which leads to small effective integrated luminosities at low values of $E_T$. Using the CDF excluded production cross sections for particles decaying into $b \bar{b}$ pairs we obtained the limits on $\gamma_B$ shown in Fig.\[fig:cdf\]. We emphasize that the bounds shown in this figure are directly based on an experimental analysis, including all resolution and efficiency effects. As we can see, for $m_B$ up to a few hundred GeV the limits on $\gamma_B$ from this search are weaker than the ones that should be obtainable using the muon trigger, if no signal is found.
The Tevatron experimental collaborations are studying the possibility of triggering events exhibiting displaced vertices for the upcoming Main Injector run. We access the impact of this trigger on the searches for $\gamma_B$ eliminating the cuts on the muon coming from $b$ decays and introducing the QCD mistag background with a rate of 1% per jet. All other cuts are unchanged. Fig. \[fig:dis\] contains the region of the $(\alpha_B , m_B)$ plane that can be probed at the MI with this new trigger. As we can see from this figure, this trigger can increase the sensitivity of the Tevatron for $\gamma_B$ searches.
At HERA, $\gamma_B$ bosons can be produced via the hadronic content of the photon: $$\gamma p \to q \bar{q} \to \gamma_B \to q^\prime \bar{q^\prime} \; .$$ However, the two-jet signature of $\gamma_B$ is immersed in a large background from resolved photons. It turns out that the signal can not be observed even for the largest couplings allowed by the $Z$ physics ($\alpha_B \simeq 0.2$) for the presently available luminosities. In order to observe a $\gamma_B$ signal for this large couplings one would need an integrated luminosity of at least 250–500 pb$^{-1}$ depending on $m_B$. Therefore, the bounds on $\gamma_B$ from HERA are much weaker than the $Z$ pole or Tevatron ones.
Rapidity gap analysis
=====================
Experiments at HERA and the Tevatron have observed events containing two jets with no hadronic activity between them. These occur with a frequency of order of one percent in hadron–hadron collisions [@cdf-prl]. This is just one example of an interaction mediated by the exchange of the “Pomeron”, a state which carries no net color. Since $\gamma_B$ is a color singlet, it can also give rise to rapidity gap events [@mu:3]. In this case the fraction of events presenting rapidity gaps as well as their kinematical distributions are determined by $\alpha_B$ and $m_B$, which allows us to constrain these parameters. Here we extract the bounds on $\gamma_B$ from the study of rapidity gaps assuming that these are only due to $\gamma_B$ exchange in the $t-$channel.
The DØ Collaboration has recently measured the production cross section of hard jets separated by a rapidity gap as a function of transverse momentum and gap size [@jill]. This data indicates that a large fraction of the gap events originates from quark–quark collisions [@two], a feature that is present in the $\gamma_B$ exchange. In order to obtain bounds on $\gamma_B$ from this data, we evaluated $$F_{\text{gap}} (E_T) = \frac{ d \sigma_B / d E_T}
{d \sigma_{\text{total}} / d E_T} \; ,$$ where $\sigma_B$ and $\sigma_{\text{total}}$ are the $\gamma_B$ contribution and the total cross section for the production of jet pairs, respectively.
In our analysis, we fixed the value of $m_B$ and determined $\alpha_B$ in order to fit the experimental $E_T$ spectrum, using the cuts and $E_T$ bins defined in Ref. [@jill]. We exhibit in Fig.\[tev:gap\] the region in the $(\alpha_B , m_B)$ plane obtained from the fits to the data. Although the $\chi^2$ distribution as a function of $\alpha_B$ has a well-defined minimum for all values of $m_B$, the quality of the fit is good only at small $m_B$. This can be seen from Fig. \[fit:qu\], which shows the general trend of the predicted $E_T$ distribution as $m_B$ increases. Therefore, $\gamma_B$ exchange cannot be the sole source of rapidity gap events at the Tevatron since the low mass allowed region in Fig. \[tev:gap\] has already been ruled out by the direct $\gamma_B$ search in $\Upsilon$ decays [@mu:3]. Furthermore, for larger $\gamma_B$ masses the fitted values of $\alpha_B$ fall well within the region that can be probed by the the direct search for $\gamma_B$ using Run I data, see Fig.\[fig:bb\]. It is interesting to notice that introducing a gap survival probability [@bj] $P_s < 1$ only worsens the problem since this will require larger values of $\alpha_B$ to fit the data, $\alpha_B \propto 1/\sqrt{P_s}$.
Hard jets separated by a rapidity gap have also been observed in photoproduction events at the $ep$ collider HERA [@hera]. The ZEUS Collaboration measured that approximately 10% of the two-jet events with $p_T > 6$ GeV and $\Delta \eta > 3$ present a rapidity gap [@hera]. Assuming that these events stem from $\gamma_B$ exchange in the $t$–channel we can constrain the mass and coupling of $\gamma_B$. We present in Table \[gap:hera\] the values of $\alpha_B$ that lead to the observed gap fraction of 0.1 at large rapidity separation, where we imposed the cuts of Ref. [@hera]. Here it is also clear that these events cannot be explained solely as being due to $\gamma_B$ exchange.
Conclusions
===========
In this work we demonstrated that the presently available Tevatron data can be used to rule out the existence of bosons coupling to baryon number for masses in the range 40–300 GeV and $\alpha_B
\gtrsim 0.02$, if no signal is found. In the near future, the Tevatron experiments should increase this sensitivity to $\alpha_B \gtrsim
0.003$ at the Main Injector. These bounds would preclude $\gamma_B$ boson exchange as a significant source of events with rapidity gaps between hard jets.
Acknowledgments
===============
We would like to thank T. Stelzer for helping us to introduce the $\gamma_B$ into the package Madgraph. This work was supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), by Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP), and by Programa de Apoio a Núcleos de Excelência (PRONEX).
[99]{}
S. Giddings and A. Strominger, Nucl. Phys. [**B307**]{} (1988) 854; S. Coleman, Nucl. Phys. [**B336**]{} (1988) 643.
H. Murayama and D. B. Kaplan, Phys. Lett. [**B336**]{} (1994) 221; V. Ben-Hamo and Y. Nir, Phys. Lett. [**B339**]{} (1994) 77; A. E. Faraggi, Nucl. Phys. [**B428**]{} (1994) 111.
G. Shiu and S.-H. Henry Tye, preprint CNLS 98/1561 (hep-th/9805157).
For a model with a small mixing between $\gamma_B$ and electroweak bosons see C. D. Carone and H. Murayama, Phys. Rev. [ **D52**]{} (1995) 484.
C. D. Carone and H. Murayama, Phys. Rev. Lett. [**74**]{} (1995) 3122.
F. Abe [*et al.*]{}, CDF Collaboration, hep-ex/9809022.
J. Perkins (DØ Collaboration), [*Proceedings of the 5th International Workshop on Deep Inelastic Scattering and QCD*]{}, Chicago, Illinois, 1997; FERMILAB-Conf-97/250-E. See also G. Snow, contribution to the [*Proceedings of the International Conference on High Energy Physics 1998*]{}.
T. Dorigo for the CDF Collaboration, hep-ex/9806022.
W. Long and T. Stelzer, Comput. Phys. Commun. [**81**]{} (1994) 357.
S. Abachi [*et al.*]{} (DØ Collaboration), Phys. Rev. Lett. [**72**]{} (1994) 2332; F. Abe [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. [**74**]{} (1995) 855; Phys. Rev. Lett. [**80**]{} (1998) 1156; K. Goulianos (CDF Collaboration), Proceedings of the LAFEX International School on High Energy Physics (LISHEP-98), Rio de Janeiro, Brazil, 1998, FERMILAB-CONF-98/118-E.
O. J. P. Éboli, E. M. Gregores, and F. Halzen, Phys. Rev. [**D58**]{} (1998) 114005.
J. D. Bjorken, Int. J. Mod. Phys. [**A7**]{} (1992) 4189; Phys. Rev. [**D47**]{} (1993) 101; preprint SLAC-PUB-5823 (1992).
M. Derrick [*et al.*]{}, ZEUS Collaboration, Phys. Lett. [ **B369**]{} (1996) 55; T. Ahmed, [*et al.*]{}, H1 Collaboration, Nucl. Phys. [**B435**]{} (1995) 3.
$m_B$ (GeV) $\alpha_B$
------------- ------------
10. 1.05
20. 2.84
30. 4.59
: Values of $\alpha_B$ needed to explain the formation of rapidity gaps in photoproduction events at HERA for several $\gamma_B$ masses.[]{data-label="gap:hera"}
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.