id
string | text
string | image
image |
|---|---|---|
211206027/31
|
$(2)$ If
$m$ is even and $a\in\mathbb{F}_{p^{2}}\backslash\mathbb{F}_{p}$, then
\begin{align*} \begin{aligned} \frac{w_{min}}{w_{max}}&=\frac{p^{m-2}-p^{m-3}-(p-1)p^{\frac{m-4}{2}}}{p^{m-2}-p^{m-3}+(p-1)p^{\frac{m-4}{2}}}= 1-\frac{2(p-1)}{p^{\frac{m}{2}}-p^{\frac{m-2}{2}}+(p-1)} > \frac{p-1}{p}. \end{aligned} \end{align*}
$(3)$ If
$m$ is even and $a\in\mathbb{F}_{p^{m}}\backslash\mathbb{F}_{p^2}$, then
\begin{align*} \begin{aligned} \frac{w_{min}}{w_{max}}&=\frac{p^{m-2}-p^{m-3}-p^{\frac{m-2}{2}}}{p^{m-2}-p^{m-3}+p^{\frac{m-2}{2}}}= 1-\frac{2}{p^{\frac{m-2}{2}}-p^{\frac{m-4}{2}}+p} > \frac{p-1}{p}. \end{aligned} \end{align*}
Hence, for $m\ge 6$, by Lemma 5.1, $\mathcal{C}_{D_a}$ is minimal. Therefore, its dual code can be employed to construct secret sharing schemes with interesting access structures.
## Applications for $s$ -sum sets
In 1984, Courteau and Wolfmann introducted the notion of the triple-sum set [4], a natu- ral generalization of partial difference sets, which enter the study of two-weight codes [6, 21]. Recently, Shi and Sol $\acute{e}$ generalized the nonation of the triple-sum set to be $s$ -sum set $(s>1)$, and then they gave the connection among $s$ -sum sets, $s$ -strongly walk-regular graphs and three- weight codes [22, 25].
Some notations and results for the $s$ -sum set are given as follows [25].
For given integers $k$ and $s$, a nonempty set $\Omega\subsetneq\mathbb{F}_{q}^{k}$ is an $s$ -sum set if it is stable by scalar multiplication, and for any nonzero $h\in\mathbb{F}_q^{k}$, there are constants $\sigma_0$ and $\sigma_1$ such that $h$ can be written as
\begin{align*} h=\sum_{i=1}^{s}x_i~~~\big(x_i\in\Omega~\text{for}~i=1,\ldots,s\big) \end{align*}
with $\sigma_0$ times of $h\in\Omega$ and $\sigma_1$ times of $h\in\mathbb{F}_q^{k}\backslash\Omega$.
We denote $\mathcal{C}(\Omega)$ by the projective code of length $n=\frac{|\Omega|}{q-1}$ with the parity check matrix $H$, where $H$ is the $k\times n$ matrix with columns projectively non-equivalent elements of $\Omega\backslash\{0\}$.
In the following, the connection between $\Omega$ and $\mathcal{C}(\Omega)$ is given.
Lemma 5.2 (Theorem 6 in [25]) Assume that $\mathcal{C}(\Omega)^{\perp}$ is of length $n$ and has three nonzero weights $w_1<w_2=\frac{n(q-1)}{q}<w_3$, with $w_1+w_3=\frac{2n(q-1)}{q}$. Then $\Omega$ is an $s$ -set sum for any odd $s>1$.
Now, we can construct an $s$ -sum set for any odd $s>1$ from $\mathcal{C}_{D_a}$ in Theorem 3.1.
For $n=p^{m-2}$ and $D_a=\{d_1,\ldots,d_{n}\}$, let $v_1,\ldots,v_m$ be a basis of $\mathbb{F}_{p^{m}}$ over $\mathbb{F}_p$, then $\mathcal{C}_{D_a}$ has an generator matrix
\begin{align*} G=\big(\mathbf{g}_1,\ldots,\mathbf{g}_n\big), \end{align*}
where
\begin{align*} \mathbf{g}_i=\big(\mathrm{Tr}(v_1d_i^2),\ldots,\mathrm{Tr}(v_md_i^2)\big)^{\mathrm{T}}\quad(i=1,\ldots,n). \end{align*}
| |
211206027/1
|
# Constructions for several classes of few-weight linear codes and their applications
Canze Zhu and Qunying Liao ∗
Abstract. In this paper, for any odd prime $p$ and an integer $m\ge 3$, several classes of linear codes with $t$ -weight $(t=3,5,7)$ are obtained based on some defining sets, and then their complete weight enumerators are determined explicitly by employing Gauss sums and quadratic character sums. Especially for $m = 3$, a class of MDS codes with parameters $[p,3,p-2]$ are obtained. Furthermore, some of these codes can be suitable for applications in secret sharing schemes and $s$ -sum sets for any odd $s>1$.
Few-weight codes; Complete weight enumerators; Character sums; Secret sharing schemes; $s$ -sum sets
MSC (2010). 94A24, 94B05
## Introduction
Let $\mathbb{F}_{p^m}$ be the finite field with $p^m$ elements and $\mathbb{F}_{p^m}^*=\mathbb{F}_{p^m}\backslash \{0\}$, where $p$ is an odd prime and $m$ is a positive integer. An $[n,k,d]$ linear code $\mathcal{C}$ over $\mathbb{F}_p$ is a $k$ -dimensional subspace of $\mathbb{F}_p^n$ with minimum (Hamming) distance $d$ and length $n$. The dual code of $\mathcal{C}$ is defined as
$$
\mathcal{C}^{\perp}=\{\mathbf{c}^{'}\in\mathbb{F}_p^n~|~\langle\mathbf{c}^{'},\mathbf{c}\rangle=0~ \text{for any}~\mathbf{c}\in\mathcal{C}\}.
$$
Clearly, the dimension of $\mathcal{C}^{\perp}$ is $n-k$. A linear code $\mathcal{C}$ is said to be projective if the minimum distance of $\mathcal{C}^{\perp}$ is not less than $3$.
Let $A_i\ (i=0,1,\ldots,n)$ be the number of codewords with Hamming weight $i$ in $\mathcal{C}$, then the weight enumerator of $\mathcal{C}$ is defined by the polynomial
$$
1+A_1z+\cdots+A_nz^n.
$$
$\mathcal{C}$ is
called a $t$ -weight code if the number of nonzero $A_i$ in the sequence $(A_1,\ldots,A_n)$ is equal to $t$. In addition, the complete weight enumerator for a codeword $\mathbf{c}$ is the monomial
$$
w(\mathbf{c})=w_0^{t_0}w_1^{t_1}\cdots w_{p-1}^{t_{p-1}}
$$
∗ Corresponding author.
E-mail. [email protected] (Q. Liao), [email protected] (C. Zhu).
Supported by National Natural Science Foundation of China (Grant No. 12071321).
| |
211206027/3
|
The paper is organized as follows. In section 2, some related basic notations and results for character sums are given. In section 3, the complete weight enumerators for several classes of $t$ -weight $(t=3,5,7)$ linear codes are presented, especially, these $t$ -weight $(t=3,5)$ codes are projective. In section 4, the proofs for the main results are given. In section 5, some of these codes can be applicated in secret sharing schemes, and $s$ -sum sets for any odd $s>1$ is constructed from this three-weight projective codes. In section 6, we conclude the whole paper.
## Preliminaries
An additive character $\chi$ of $\mathbb{F}_{p^m}$ is a function from $\mathbb{F}_{p^m}$ to the multiplicative group $U=\{u\ |\ |u|=1,\ u\in\mathbb{C}\}$, such that $\chi(x+y)=\chi(x)\chi(y)$ for any $x,\ y\in\ \mathbb{F}_{p^m}$. For each $b\in \mathbb{F}_{p^m}$, the function
$\chi_b(x)=\zeta_p^{\mathrm{Tr}(bx)}$ $(x\in \mathbb{F}_{p^m})$
defines an additive character of $\mathbb{F}_{p^m}$. When $b=0,\ \chi_0(x)=1$ for any $x\in \mathbb{F}_{p^m}$ is called the trivial additive character of $\mathbb{F}_{p^m}$. The character $\chi:=\chi_1$ is called the canonical additive character of $\mathbb{F}_{p^m}$ and every additive character of $\mathbb{F}_{p^m}$ can be written as $\chi_b(x)=\chi(bx)$. The orthogonal property for the additive character is given by
\begin{align*} \sum_{x\in \mathbb{F}_{p^m}}\zeta_p^{\mathrm{Tr}(bx)}=\begin{cases} p^m,\quad& \text{if}~b=0;\\
0,\quad &\text{otherwise}. \end{cases} \end{align*}
We extend the quadratic character $\eta_m$ of $\mathbb{F}_{p^m}^*$ by letting $\eta_m(0)=0$, then the quadratic Gauss sums $G_m$ over $\mathbb{F}_{p^m}$ is defined as
$G_m=\sum\limits_{x\in \mathbb{F}_{p^m}}\eta_m(x)\chi(x)$.
Now, some properties for quadratic character and quadratic Gauss sums are given as follows.
Lemma For $x\in\mathbb{F}_p^*$,
\begin{align*} \eta_m(x)=\begin{cases} 1,\quad& ~2\mid m;\\
\eta_{1}(x),\quad &~\text{otherwise}. \end{cases} \end{align*}
Lemma 2.2 ([18], Theorem 5.15) For the Gauss sums $G_m$ over $\mathbb{F}_{p^m}$,
\begin{align*} G_m=(-1)^{m-1}\sqrt{-1}^\frac{(p-1)^2m}{4}p^{\frac{m}{2}}. \end{align*}
Lemma Let $f(x)=a_2x^2+a_1x+a_0\in \mathbb{F}_{p^m}[x]$ with $a_2\neq0$, then
\begin{align*} \sum_{x\in \mathbb{F}_{p^m}}\chi(f(x))=G_m\eta_m(a_2)\chi(a_0-a_1^2(4a_2)^{-1}). \end{align*}
| |
211206027/9
|
\begin{align*} W(\mathcal{C}_{D_a})=&1+\big(p^{m-2}-1\big)\prod_{\rho\in \mathbb{F}_p}w_{\rho}^{p^{m-3}}\\
&+ \frac{1}{2}\big(p^{m-1}+p^{\frac{m-4}{2}}\big) \sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}+(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}-p^{\frac{m-4}{2}}}\\
&+ \frac{1}{2}\big(p^{m-1}- p^{\frac{m-4}{2}}\big)\sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}-(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}+p^{\frac{m-4}{2}}}\\
&+ \frac{1}{2}(p-1)\big(p^{m-2}-p^{\frac{m-4}{2}}\big)w_0^{p^{m-3}+(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p^*}w_{\rho}^{p^{m-3}-p^{\frac{m-4}{2}}}\\
&+ \frac{1}{2}(p-1)\big(p^{m-2}+p^{\frac{m-4}{2}}\big)w_0^{p^{m-3}-(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p^*}w_{\rho}^{p^{m-3}+p^{\frac{m-4}{2}}}. \end{align*}
Table $4$ The weight distribution for $\mathcal{C}_{D_a}$
$(4\mid m,~a\in\mathbb{F}_{p^{2}}\backslash\mathbb{F}_p)$
\begin{tabular}{|p{6cm}<{\centering}| p{6cm}<{\centering}|}
\hline weight $w$ & frequency $A_w$ \hline $0$ & $1$ \hline $p^{m-2}-p^{m-3}-(p-1)p^{\frac{m-4}{2}}$ & $\frac{1}{2}(p-1)\big(p^{m-2}-p^{\frac{m-4}{2}}\big)$ \hline $p^{m-2}-p^{m-3}-p^{\frac{m-4}{2}}$ & $\frac{1}{2}(p-1)\big(p^{m-1}-p^{\frac{m-4}{2}}\big)$ \hline $p^{m-2}-p^{m-3}$ & $p^{m-2}-1$ \hline $p^{m-2}-p^{m-3}+p^{\frac{m-4}{2}}$ & $\frac{1}{2}(p-1)\big(p^{m-1}+p^{\frac{m-4}{2}}\big)$ \hline $p^{m-2}-p^{m-3}+(p-1)p^{\frac{m-4}{2}}$ & $\frac{1}{2}(p-1)\big(p^{m-2}+p^{\frac{m-4}{2}}\big)$ \hline
\end{tabular}
Theorem 3.5 For any integer $s$ and even $m$ with $s>2$ and $s\!\mid\! \frac{m}{2}$, if $a\!\in\!\mathbb{F}_{p^{s}}\backslash\mathbb{F}_{p^2}$, then $\mathcal{C}_{D_a}$ is a $[p^{m-2}, m, p^{m-2}-p^{m-3}-p^{\frac{m-2}{2}}]$ code with the weight distribution in Table $5$, and the complete weight enumerator is
\begin{align*} W(\mathcal{C}_{D_a})=&1+\big(p^{m-1}-p^{m-2}+p^{m-3}-1\big)\prod_{\rho\in \mathbb{F}_p}w_{\rho}^{p^{m-3}}\\
&+\frac{1}{2}\big(p^{m-3}+p^{\frac{m-6}{2}}\big)\sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}-\eta_{1}(1-\rho\gamma^{-1})p^{\frac{m-2}{2}}}\\
&+\frac{1}{2}\big(p^{m-3}- p^{\frac{m-6}{2}}\big)\sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}+\eta_{1}(1-\rho\gamma^{-1})p^{\frac{m-2}{2}}}\\
&+ \frac{1}{2}(p-1)\big(p^{m-2}-p^{\frac{m-4}{2}}\big)w_0^{p^{m-3}+(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p^{*}}w_{\rho}^{p^{m-3}- p^{\frac{m-4}{2}}}\\
&+ \frac{1}{2}(p-1)\big(p^{m-2}+p^{\frac{m-4}{2}}\big)w_0^{p^{m-3}-(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p^{*}}w_{\rho}^{p^{m-3}+p^{\frac{m-4}{2}}}\\
&+\frac{1}{2}(p-1)p^{m-2}\sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}+(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}-p^{\frac{m-4}{2}}}\\
&+\frac{1}{2}(p-1)p^{m-2}\sum_{\gamma\in\mathbb{F}_p^{*}}w_\gamma^{p^{m-3}-(p-1)p^{\frac{m-4}{2}}}\prod_{\rho\in \mathbb{F}_p\backslash\{\gamma\}}w_{\rho}^{p^{m-3}+p^{\frac{m-4}{2}}}. \end{align*}
| |
221206047/2
|
## II. M ODELING
## A. Wind field Model of Hurricane
A part of the impact of hurricanes originates from their wind speed. This wind speed is a vector field centered around the hurricane’s eye and changes as it moves inland. This wind field can be modeled as an equation with three variables: $v_{max}$, $R_{v_{max}}$, and $R_s$. Here, $v_{max}$ measures the maxi- mum sustained wind speed of the hurricane in knots, $R_{v_{max}}$ measures the distance to $v_{max}$ in nautical miles (nmi), and $R_s$ measures the radius of the hurricane from the hurricane’s eye, also known as the radius of the outermost closest isobar (roci), in nmi. Fig. 1 demonstrates the relationship between the wind speed and the distance from the hurricane’s eye, and the piecewise mathematical function that represents Fig. 1 is shown in (1).
$$
\small
\begin{aligned}
&v(x)= \begin{dcases} K \times v_{max}(1-exp{[-\Psi x]}) & 0 ~\leqslant~ x<R_{v_{max}} \\
v_{max} \exp \left[- \Lambda \left(x-R_{v_{max}}\right)\right] & R_{v_{max}} ~\leqslant~ x ~\leqslant~ R_{s} \\
0 & x>R_{s}\end{dcases} \\
&\Psi=\frac{1}{R_{v_{max}}} \ln \left(\frac{K}{K-1}\right), K>1;
\quad \Lambda = \frac{\ln \beta}{R_{s}-R_{v_{max}}}
\end{aligned}
$$
$K$ is
a known hurricane constant, $x$ is the distance from the hurricane eye, and $\beta$ is the factor that the maximum sustained wind speed will decrease at the hurricaneâ C™s boundary $(R_s)$. Using this equation, it can be assumed that the hurricane has no effect outside this boundary.
To ensure a realistic hurricane wind field, the wind param- eters in (1) are obtained using the International Best Track Archive for Climate Stewardship (IBTrACS). IBTrACS is a global collection of tropical cyclones that merges data from multiple agencies to create a database for historical hurricane parameters at various time steps [16]. The hurricane wind field is generated for each time step $t$ by obtaining $\{R_{v_{max}}^t$, $R_{s}^t$, $v_{max}^t\}$ from IBTrACS and using Eq. (1).
Fig. 1: Static gradient wind field of a typical hurricane.
## B. Storm Surge Model
In this work, we use SLOSH (Sea, Lake, and Overland Surge from Hurricanes) model to generate probabilistic storm surge scenarios. SLOSH is a numerical storm surge model that provides surge heights around the coastal regions from historical or hypothetical hurricanes based on parameters such as atmospheric pressure, hurricane tracks, forward speed, and so forth [17]. It was developed by National Oceanic and Atmo- spheric Administration (NOAA) and is currently being used by several federal agencies, including the national hurricane center (NHC) and the federal emergency management agency (FEMA), for flood advisories and evacuation. Furthermore, NHC provides a SLOSH Display Package (SDP) tool in which users can generate flooding scenarios based on the direction, forward speed, and intensity of the hurricane followed by the sea tide level [18]. The provided surge heights are above the elevation referenced in the North American Vertical Datum of 1988 (NAV88). The tool also has an added functionality to subtract land elevations so that the surge level is referenced above the ground level. Details on the SDP tool and other surge-related products from NOAA can be found at [18].
SDP contains several coastal basins on which surge scenar- ios can be created. Furthermore, SDP provides two different surge analyses based on hurricane simulations: Maximum Envelope of Water (MEOW) and Maximum of the MEOWs (MOM). The MEOWs reflect the worst-case snapshots of the storm surge for hurricanes of particular intensity and forward direction but with different landfall locations. This work uses MEOW for storm surge scenarios to identify potential flooding locations. The MEOW scenarios from SDP provide the inundation level for each defined grid in a basin. Let $\mathcal{X}_\mathcal{S}$ be the geographical coordinate of a substation $\mathcal{S}$. Then, we define $\mathcal{X}^\mathcal{B}_{\mathcal{S}, h}$ as the inundation level, $h$, for substation $\mathcal{S}$ situated at $\mathcal{X}$ for basin $\mathcal{B}$. Since the inundation level is spatially distributed, the expected value of inundation around 0.5 miles of $\mathcal{X}_\mathcal{S}$ is assumed to be $\mathcal{X}^\mathcal{B}_{\mathcal{S}, h}$.
## C. Impact Assessment Model
Combined, the hurricane and storm surge creates a spa- tiotemporal effect on the power grid. The spatiotemporal impact of hurricanes on the power grid is generally guided by the fragility curves of the transmission lines [19]. Let $\Gamma_{l}^{t,\zeta}$ be the maximum wind speed on line $l$ at time step $t$ due to a hurricane traversing in track $\zeta$. Then the outage probability of line $l$ at time step $t$ due to a hurricane traversing in track $\zeta$ is given by (2)
$$
\small
\begin{aligned}
&\mathbb{P}_{out}^{t,\zeta}(l)= \begin{dcases} 0 & \Gamma_{l}^{t,\zeta} < v_{cri} \\
\frac{\Gamma_{l}^{t,\zeta} - v_{cri}}{v_{col} - v_{cri}} & v_{cri} ~\leqslant~ \Gamma_{l}^{t,\zeta} < v_{col} \\
1 & \Gamma_{l}^{t,\zeta} ~\geqslant~ v_{col} \end{dcases} \\
\end{aligned}
$$
where $v_{cri}$ = 48.59 knots and $v_{col}$ = 106.91 knots are the critical wind speed beyond which a transmission line is affected by the hurricane and the wind speed at which the transmission line collapses, respectively [14]. If $\delta_{l}^{t,\zeta} \in \{0,1\}$ denotes the line status of line $l$ at time $t$ for hurricane in track $\zeta$, then $\delta_{l}^{t + 1,\zeta} ~\leqslant~ \delta_{l}^{t,\zeta}$. This ensures that if a line $l$ experiences
| |
221206047/4
|
is they are snapshots of the storm surge and carry no time information. However, we can assume that the components are not inundated until the hurricane approaches. Hence, we assign an activation flag for each component per basin at each time step. This is discussed in the next subsection.
## B. Power Grid Impact Assessment using MCS
This work uses a synthetic 2000-bus Texas power grid model to identify the compound probabilistic impact of hur- ricane and storm surges on the power grid [23]. The grid model maps the synthetic buses, substations, and transmission lines on the geographical footprint of Texas. There are 1250 substations and 1918 transmission lines at 4 different voltage levels — 115 kV, 161 kV, 230 kV, and 500 kV. The total generation capacity of the system is 96291.5 megawatts (MW), and there are 1125 load units with a total size of 67109.2 MW. When the hurricane moves inland, $\Gamma_{l}^{t,\zeta}$ is obtained for each $l$ and $\mathcal{X}^\mathcal{B}_{\mathcal{S}, h}$ is obtained for each $\mathcal{S}$ from the MEOW maps. It is assumed that all substations at the coast are elevated at a level of 3 ft from the ground level. Fig. 4 shows the wind field of Hurricane Harvey and MEOW maps on five different basins on the footprint of Texas.
For each $\zeta$, $\mathcal{B}$, and $t$, MCS is performed for several trials. For each trial, $\mathbb{P}_{out}^{t,\zeta}(l)$ and $\mathbb{P}^{\mathcal{S}}_{out}(\mathcal{X}^\mathcal{B}_{\mathcal{S}, h})$ is compared with a uniform random number, $r\sim U(0,1)$, to identify $\mathcal{L}^t_\mathcal{H} \cup \mathcal{L}^t_\mathcal{S}$. As discussed before, MEOW maps are time-independent. However, we define an activation flag $\mathcal{S}^{t, \zeta}_{\mathcal{B}, h} \in \{0,1\}$ such that $\mathcal{S}^{t, \zeta}_{\mathcal{B},h} = 1$ makes the substation $\mathcal{S}$ inundated with depth $h$ above the ground elevation for hurricane track $\zeta$, SLOSH basin $\mathcal{B}$, and time $t$. In this work, we do not model any drainage system, and hence it is assumed that for future time steps $t+1$ the inundation level is the same. The objective here is to identify the time stamp when the substation gets flooded to analyze the compound spatiotemporal impact of two different hazards at that time step. The information, $\mathcal{L}^t_\mathcal{H} \cup \mathcal{L}^t_\mathcal{S}$, is then sent to the power grid simulator to remove the respective branches from the system. Finally, the loss for the particular $t$ is observed as the total load disconnected from the main grid. The MCS trial is conducted until the loss converges to some value.
We assume that the inundation scenarios for each $\mathcal{B}$ and each of the hurricane tracks $\zeta$ are equally likely. If $N_\zeta$ is the total number of tracks under consideration and $N_\mathcal{B}$ is the total number of basins, then the system level load loss at each time step $t$ is given by (4)
$$
\small
loss_t = \frac{1}{N_\zeta \times N_\mathcal{B}}\sum_{\zeta = 1}^{N_\zeta}{\sum_{\mathcal{B}} {loss_{\mathcal{B}}^{t,\zeta}}}
$$
## IV. R ESULTS AND A NALYSIS
The hurricane data is obtained from IBTrACS [16], and synthetic hurricanes are generated using CLIMADA [21] in Python. SDP [18] is used to obtain the MEOW maps for Texas basins. The power grid is modeled in MATPOWER 7.1 [24], and MCS on the power grid model is conducted in MATLAB R2021a.
From the IBTrACS, the hurricane’s initial location is in the North Atlantic Ocean, which is observed at $``2017-08-16~06:00:00''$
Fig. 4: a) Wind field of Hurricane Harvey and b) Substation flooding scenario, above ground level, for Texas basins.
Fig. 5: a) Line outage probability for selected lines and the original Harvey’s track. b) Substation outage probability for all storm surge scenarios.
$``2017-08-16~06:00:00''$ and
hence creates no impact on the power grid. Thus, the simulation for the impact assessment begins at the $100^{th}$ time step, i.e., at the advisory of $``2017-08-24~12:00:00''$. To avoid any confusion, this timestamp is referred to as $t=0$ hours for the rest of the paper. Fig. 5a shows the outage probability of a set of transmission lines due to the original Hurricane Harvey track. It can be observed that with the moving nature and intensity of the hurricane, the outage probability of each line changes. The probability is maximum when the line experiences wind speed closer to $v_{col}$ and gradually decreases as the hurricane decays while moving inland. Fig. 5b represents the outage probability of substations for five different Texas basins. It is seen that the substation’s vulnerability depends on the location and basin.
It was found that 800 Monte Carlo trials are enough to achieve convergence for any simulation case. Hence, for each $\zeta$, $\mathcal{B}$, and $t$, 800 Monte Carlo trials are conducted. Fig. 6 shows the comparison of losses when considering the impact of the hurricane alone with the compound effect of the hurricane and the coastal flood for $\zeta=1$. It can be observed that
| |
221206047/1
|
# Spatiotemporal Impact Analysis of Hurricanes and Storm Surges on Power Systems
Abstract — This paper develops a spatiotemporal probabilistic impact assessment framework to analyze and quantify the com- pounding effect of hurricanes and storm surges on the bulk power grid. The probabilistic synthetic hurricane tracks are generated using historical hurricane data, and storm surge scenarios are generated based on observed hurricane parameters. The system losses are modeled using a loss metric that quantifies the total load loss. The overall simulation is performed on the synthetic Texas 2000-bus system mapped on the geographical footprint of Texas. The results show that power substation inundation due to storm surge creates additional load losses as the hurricane traverses inland.
Index Terms — Power system resilience, hurricanes, storm surge, Monte-Carlo simulation
## I. I NTRODUCTION
Hurricanes account for over a trillion dollars in economic losses and are considered to be the major reason for power outages in the US [1], [2]. During landfall, as the strong hurricane wind field traverses inland, it propels a huge water body known as a storm surge flooding the coastal regions. The storm surge is sometimes the most destructive part of a hurricane and accounts for considerable damage [3]. For instance, Hurricane Ida caused about $55 billion in damages in Louisiana alone due to wind and storm surge damage, with additional flooding damage of about$ 23 billion in the North- eastern US [4]. Almost 1.2 million customers experienced power outages across eight different states. Hurricane Ian recently had a devastating impact in Florida and is expected to have incurred billions of dollars in losses, with a peak of about 2.7 million customers in a power outage [5]. The frequency of such high-impact, low-probability (HILP) events has increased at an alarming rate, costing about $ 152.6 billion in climate- related disasters in 2021 alone in the US. Hence, there is a serious need to identify the potential impacts of these hazards on electric power systems.
Although there have been several advancements in weather prediction models, they have not been properly utilized to analyze the potential impact of upcoming natural hazards on the power grid. Such predictive information is essential to power grid planners and operators to reduce the impacts when an event is realized [6], [7]. For instance, system operators can identify potential substations that could be inundated due to storm surges and proactively disconnect them to avoid equipment damage and facilitate fast restoration. These long- term planning strategies can help planners identify vulnerable transmission lines and propose line hardening strategies. Since hurricanes and associated storm surges have compounded spatiotemporal effects, there is an urgent need for a weather- grid impact model for outage risk assessment.
Several existing works model the impacts of hurricanes on the power grid due to high-speed winds [8]. To model the spatial nature of the event, other works divided the entire power grid into zones such that each zone experiences a different wind speed [9]. However, based on the hurricane’s trajectory, each component is impacted differently as the hurricane moves inland, rendering the simplified hurricane model inaccurate for grid impact assessment. Other works model the effect of hurricanes on distribution grid [10]. Since the overall spatial exposure of a hurricane is greater than an entire distribution grid for each time step, the analysis based on the distribution system alone would not be meaningful enough. Furthermore, none of these works model the impact of storm surges on the power grid. In [11], a power system impact assessment framework is presented to identify potential mitigation strategies and enhance resilience against floods. A stochastic optimization framework for substation hardening against storm surge is presented in [12]. [13] evaluated the impacts of tropical cyclones and heatwaves on the power grid. The existing work, however, lacks the spatiotemporal impacts analysis of hurricanes characterizing the compound effects of high-speed wind and flooding.
This paper aims to develop a compound spatiotemporal ef- fect of hurricanes and storm surges on electric power systems. In this paper, we extend our previous work, spatiotemporal impact assessment of hurricanes [14], for a more realistic assessment of hurricane impacts. First, we generate dynamic hurricane scenarios based on historical hurricane tracks and obtain storm surge scenarios based on the obtained hurricane parameters closer to the landfall. Next, sequential Monte Carlo simulation (MCS) is performed for probabilistic tracks and flooding scenarios for each time step as the hurricane moves inland. Finally, we quantify the impacts using a spatiotemporal loss metric based on the compound effect of hurricanes and floods. To the authors’ knowledge, the proposed framework is the first to introduce the compound spatiotemporal impact assessment of high-speed wind and storm surges on electric power systems.
| |
221206047/3
|
an outage at time $t$, it remains out of service for the remaining duration of the hurricane.
The impact assessment from storm surges can be modeled similarly by defining the outage probability of substations. Although FEMA has provided a general fragility level for substations in their HAZUS tool [20], we relax the curve through Weibull stretched exponential function as the substa- tions around coasts have some hardening measures already in place. Hence, the outage probability of a substation $\mathcal{S}$ for a given inundation level $h$ at each basin $\mathcal{B}$ is given by (3)
$$
\small
\mathbb{P}^{\mathcal{S}}_{out}(\mathcal{X}^\mathcal{B}_{\mathcal{S}, h}) = 1 - exp\left[- {\left(\frac{\mathcal{X}^\mathcal{B}_{\mathcal{S}, h}}{a}\right)^b}\right] \\
$$
where $a\in \mathbb{R}^+$ and $b>2$ are known constants and determine the shape of the fragility curve. When a substation is flooded and deemed out-of-service, all transmission lines connected to and from the substation are disconnected. If $\delta_{l,\mathcal{S}}^{t,\zeta} \in \{0,1\}$ denotes the line status of line $l$ connected to substation $\mathcal{S}$ at time $t$ for hurricane in track $\zeta$, then $\delta_{l, \mathcal{S}}^{t + 1,\zeta} ~\leqslant~ \delta_{l,\mathcal{S}}^{t,\zeta}$. This ensures that if a line $l$ experiences an outage at time $t$ due to a storm surge, it remains out of service for the remaining event duration. When simulating storm surges as an effect of hurricanes, it is important to avoid the redundancy of impact that both transmission line and substation outages can have on the power grid. Let $\mathcal{L}^t_\mathcal{H}$ be the set of lines affected by hurricanes and $\mathcal{L}^t_\mathcal{S}$ be the set of lines out of service due to a flooded substation at time $t$. Then, the total number of lines affected by the compound effect of the hurricane and substations inundation due to storm surge is given by $\mathcal{L}^t_\mathcal{H} \cup \mathcal{L}^t_\mathcal{S}$.
## III. O VERALL F RAMEWORK
This section describes the overall approach to assess the spatiotemporal impact of hurricane and storm surges on the electric power systems; see Fig. 2. First, synthetic hurricane tracks were generated based on historical data from IBTrACS. The hurricane impact model is then used to obtain the wind speed experienced by each transmission line. Similarly, flood- ing scenarios are obtained from SLOSH basins for the given characteristic of the hurricane. The fragility function defined in (2) and (3) provides the outage probability of transmission lines due to hurricanes and the outage probability of substa- tions due to coastal flood inundation. Finally, Monte Carlo simulations (MCS) are conducted to obtain the probabilistic loss for the system. All flood scenarios are considered equally likely. The final spatiotemporal system-level loss metric for each time step is the obtained. The loss metric is the expected value of the loss obtained for the hurricane track and the flood basin for that particular time step. The following subsections detail the overall framework.
## A. Generating Hurricane and Storm Surge scenarios
The geographical data and parameters are obtained for Hurricane Harvey, that hit the coast of Texas in August 2017 using IBTrACS. Synthetic hurricane tracks are obtained by perturbing several characteristics of Hurricane Harvey, such
Fig.2: Overall architecture for generating hurricane and storm surge scenarios and grid impact assessment.
Fig. 3: Hurricane Harvey track and other synthetic tracks obtained from CLIMADA [21].
as a shift in the initial point of origin, the amplitude of trans- lational speed, bearing angle, landfall decay, etc. The synthetic tracks are obtained from the Climate adaptation (CLIMADA) tool. CLIMADA is a probabilistic natural catastrophe damage model that facilitates climate-based simulations [21]. CLI- MADA can generate synthetic tracks in IBTrACS format based on the above perturbations. To synchronize the time steps for the entire duration of the hurricanes, the data is observed every two hours, and the hurricane wind field is generated for each $t$ using (1). Fig. 3 shows the synthetic tracks generated from CLIMADA, and the track with a solid line represents the actual track of Hurricane Harvey from IBTrACS. NOAA predicts that 7 out of 10 times, the estimated hurricane track is within the cone of uncertainty [22]. Hence, five synthetic tracks are selected to be within the cone of uncertainty when forecasted over 120 hours of landfall, i.e., within 240 nmi as per [22].
For storm surge assessment, five different Texas basins are selected; $\mathcal{B} \in$ {Matagorda, Corpus, Galveston, Laguna, Sabine}. MEOW scenarios are obtained for all five basins for category 4 hurricanes moving in the northwest direction with a translational speed of 10-15 miles per hour and mean to high tide conditions. One major drawback of using MEOW maps
| |
201203841/2
|
exclude methodologies of class (i) mentioned above, as these are not generally applicable.
First, we present the different possibilities available to compute the electrostrictive coefficients as appearing in Eqs.(1). As mentioned in our review of the literature, and by reference to Eqs.(1), we can see that the elec- trostrictive coefficients may be obtained by fitting a curve of strain or stress vs electric or polarisation field. [12 – 15] However, thermodynamic considerations [10, 23, 24] re- veal that the four electrostrictive coefficients $M$, $m$, $Q$ and $q$ are related to the partial derivatives of dielectric quantities with respect to mechanical ones as follows:
$$
\begin{aligned}
&\frac{1}{\epsilon_{0}}\frac{\partial\eta_{ij}}{\partial X_{kl}} = -2Q_{ijkl} &\frac{1}{\epsilon_{0}}\frac{\partial\eta_{ij}}{\partial x_{kl}} = 2q_{ijkl} \\
&\epsilon_{0}\frac{\partial\chi_{ij}}{\partial X_{kl}} = 2M_{ijkl}
&\epsilon_{0}\frac{\partial\chi_{ij}}{\partial x_{kl}} = -2 m_{ijkl}
\end{aligned}
$$
Thus, the coefficients $Q_{ijkl}$ and $q_{ijkl}$ are given by the rate of change of the dielectric stiffness, $\eta_{ij}$ (the inverse of the dielectric susceptibility) with respect to stress, $X_{kl}$, or strain, $x_{kl}$, respectively. The coefficients $M_{ijkl}$ and $m_{ijkl}$ are given by the rate of change of the suscepti- bility, $\chi_{ij}$, with respect to stress or strain, respectively. Density functional perturbation theory (DFPT) enables the calculation of the susceptibility at different stresses or strains, avoiding the problems and disadvantages intrin- sic to performing relaxations under a finite $D$ or $E$ field. Calculations relying on Eqs. (2) have the following advan- tages over those relying on Eqs. (1): (i) The k-point grid resolution is not limited in a manner dependent on the band gap/field strength. (ii) Direct computation of hy- drostatic electrostrictive coefficients is possible through the application of hydrostatic strain/pressure. This al- lows one to calculate electrostrictive properties without breaking the intrinsic crystal symmetry, unlike with the imposition of a unidirectional field. This means that, for example, in centrosymmetric cubic crystals, all four coef- ficients ($M_{h}$, $m_{h}$, $Q_{h}$, $q_{h}$) can be computed at once with- out relaxation. (iii) The infrastructure to calculate the permitivity/susceptibility using DFPT, which has been available since the 1990s, is more established and robust than that available for energy optimisation in the pres- ence of an applied $E$ or $D$ field (iv) In piezoelectric ma- terials, the electrostrictive coefficients are still obtained directly from Eqs. (2), whereas they need to be decoupled from an often much larger piezoelectric effect if one uses Eqs. (1). [15] (v) DFPT allows for the efficient decomposi- tion of the electrostrictive tensors into an electronic and ionic part, and then the decomposition of this ionic part into contributions from each phonon mode.
To confirm these advantages and validate the method, we have calculated the electrostrictive coefficients using both the methodology proposed here (Eqs. (2)) and the one used so far (Eqs. (1)). We used the DFT pack- age ABINIT (version 8.6.1) [25], with k-point grid den- sities of 8 $\times$ 8 $\times$ 8 and a plane wave cutoff energy of 50 Ha, to ensure convergence of electrostrictive coefficients of about 1%. We used the PseudoDojo [26] normconserv- ing pseudopotentials and the exchange-correlation func- tional was treated using the generalised gradient approx- imation of Purdue, Burke, and Ernzerhof, modified for solids, PBEsol. [27] We also ensured that the linear and quadratic fittings were appropriate to the given applied strain/electric fields.
In Figure 1 we show as an example the results obtained under applied fields for rocksalt MgO. These were ob- tained by optimising both the internal atomic positions and lattice vectors at constant $E$ and $D$ fields, for panels (a) and (b), respectively. The plots evince the expected quadratic strain versus the $E$ and $P$ fields. Electrostric- tion expands MgO along the direction of the field and contracts it perpendicularly. Furthermore these plots il- lustrate that the electrostrictive strains are not volume conserving, contrary to the assumptions of Ref. 11 (for example, an electric field of magnitude 1.2 GV/m will induce a volume expansion of 0.2%). The extracted fit value obtained for $M_{11}$ of 1970 pm $^2$ /V $^2$ agrees very well with the experimental values of 2020 pm $^2$ /V $^2$ [28] (a dif- ference of 2%). We have also calculated the coefficients $q_{11}$, $q_{12}$, $m_{11}$, and $m_{12}$, by fixing the lattice constants, relaxing the internal coordinates under applied $E$ and $D$ fields, and fitting the subsequent stress dependence on $P$ or $E$ field, respectively. These results are summarised in Table. I The signs of the $q_{ij}$ and $m_{ij}$ coefficients are op- posite to those of the $Q_{ij}$ and $M_{ij}$ coefficients; this can be expected, a negative axial stress is required to prevent the system from expanding in the direction of the field, and a positive stress is required to prevent its contrac- tion perpendicular to it. To obtain experimental verifi- cation for $q_{ij}$ and $m_{ij}$, we use the hydrostatic coefficients $[Q,q,M,m]_{h}=[Q,q,M,m]_{11}+2\times[Q,q,M,m]_{12}$, and the relations: $[q,m]_{h}= -3B[Q,M]_{h}$, where $B$ is the bulk modulus. Again, good agreement is found between theory and experiment.
In Fig. 1 (c) we report the corresponding results ob- tained using the DFPT calculation of permittivity of MgO under applied hydrostatic strain and stress. We imposed strains between $\pm 0.5\%$ in steps of 0.1%. The dielectric susceptibility, stiffness (obtained by inverting the permittivity tensor), and stress at each value of the strain could then be used to determine all four hydro- static electrostrictive coefficients at once. We have also calculated the individual components of the electrostric- tive tensors using this method. In this case however, rather than a hydrostatic strain/stress as in Fig. 1 (c), an axial strain (in Voigt notation) of $x=(\alpha,0,0,0,0,0)$ or axial stress of $X=(\alpha,0,0,0,0,0)$ are imposed on the unit cell, where $\alpha$ varies between $\pm0.3$% for the strain, and $\pm2.5$ GPa for the stress. We then obtain the elec- trostriction tensor components by fitting with Eqs. 2. We found excellent agreement (within 0.1%) with the
| |
201203841/4
|
FIG. 2. Convergence of calculated electrostrictive properties of rocksalt MgO with respect to (left) k-points and (right) planewave cutoff energies. Data obtained from Eqs. (1) are given by circles and dashed lines, and data obtained from Eqs. (2) are given by squares and solid lines.
RbI, and LiCl, and difluorite HfO $_{2}$, using both Eqs.(1) and Eqs.(2). In Fig. 3 we plot the calculated $M_h$ elec- trostrictive coefficients obtained from Eqs.(1) and some experimental values, on the y-axis against the ones from Eqs.(2) on the x-axis. Thus, a given material has a fixed x position, and the vertical distance of values from the line y=x represents the extent to which the experimental values, or theoretical finite field obtained values, differ from those calculated using Eqs.(2). We observe almost an order of magnitude spread in the values of the elec- trostrictive coefficients, LiCl’s being the largest. Given the already large spread in the experimental values, the figure shows excellent agreement between the methods and the general applicability of our methodology.
With the method thus corroborated, we turn now to demonstrate how we may use it to easily obtain a more detailed understanding of electrostriction. Indeed, the permittivity tensor can be decomposed into its elec- tronic and ionic contributions, the latter can be further split into each phonon mode contributions. The elec- trostrictive tensors can be decomposed by finding the stress/strain derivatives of these individual components. We display such a decomposition for perovskite BaZrO $_{3}$ in Fig. 4, where we show the percentage contribution to the permittivity and electrostrictive coefficient $M_{h}$, of each of their constituents: the electronic part, and the three transverse optical (TO) phonon modes TO $_{1}$ (94 cm $^{-1}$, mode effective charge $\Bar{Z}^*=$ 4.0 e), TO $_{2}$ (190 cm $^{-1}$, $\Bar{Z}^*=$ 5.8 e), and TO $_{3}$ (500 cm $^{-1}$, $\Bar{Z}^*=$ 3.8 e). The figure illustrates that the permittivity has a small but non-negligible electronic component of 7.3%, whilst the largest components are the two first TO phonon modes (61.4% and 27.8%, respectively), and the smallest con- tribution of TO $_3$ mode (3.5%). For the electrostriction, we see that, unlike with the permittivity, the electronic degrees of freedom have a negligible contribution (about 0.01%, invisible on the plot). The two largest contribu- tors are from the TO $_1$ and TO $_2$ modes (80% and 19%, respectively), the TO $_{3}$ giving a small but negative contri- bution (-1%). Hence, the largest contribution to the elec-
Electrostrictive coefficients calculated using Eqs.(1), or measured (denoted by triangles), on the y-axis plotted against calculations based on Eqs.(2) on the x-axis. The line y=x corresponds to exact agreement. Ref 32 is low temper- ature epxeriment based on Eqs. (2), Refs 6, 29, and 33 are room temperature experiments using Eqs. (2), and Ref 31 is room temperature experiment based on Eqs. (1).
trostrictive response comes from the softest polar mode with large polarity, in line with the fact that the soft polar mode is extremely sensitive to pressure or strain in ferroelectric perovskites. [34, 35] We note that the four coefficients $M_{h}$, $m_{h}$, $Q_{h}$, and $q_{h}$ have the same fractional composition. This follows simply from the linearity of the derivatives with respect to stress or strain, and the equivalence of hydrostatic strain and pressure in a cubic crystal.
FIG. 4. Fractional contributions of electronic and transverse optical phonon degrees of freedom to permittivity $\epsilon$ and elec- trostrictive tensor $M_{h}$ of BaZrO $_{3}$.
In conclusion, we have compared the methodologies for the computation of the electrosctrictive response in crystals using the present common capabilities of DFT codes. While numerous previous calculations of the elec- trostrictive coefficients in the literature relied on finite field methods, here we have shown originally that calcu- lating it through the variation of the susceptibility under
| |
201203841/1
|
# Optimized methodology for the calculation of electrostriction from first-principles
Daniel S. P. Tanner, 1,2, ∗
Eric Bousquet, 2 and
Pierre-Eymeric Janolin 1
In this work we present a new method for the calculation of the electrostrictive properties of ma- terials using density functional theory. The method relies on the thermodynamical equivalence, in a dielectric, of the quadratic mechanical responses (stress or strain) to applied electric stimulus (elec- tric or polarisation fields) to the strain or stress dependence of its dielectric susceptibility or stiffness tensors. Comparing with current finite-field methodologies for the calculation of electrostriction, we demonstrate that our presented methodology offers significant advantages of efficiency, robustness, and ease of use. These advantages render tractable the highthroughput theoretical investigation into the largely unknown electrostrictive properties of materials.
Electrostriction is a nonlinear electromechanical cou- pling present in all dielectrics; it is therefore the most ubiquitous electromechanical phenomenon. Despite this, in non-centrosymmetric materials, it is often overshad- owed in amplitude by its linear counterpart: piezoelec- tricity, which constitutes the primary coupling in most electromechanical systems. Electrostrictors are neverthe- less employed in sonars [1], actuators [2, 3] and other tunable electromechanical systems. [4 – 6] There is, in ad- dition, a renewed interest in electrostriction due to the recent observation of electrostrictors that have giant elec- trostrictive coefficients 10 $^4$ times larger than the best per- ovskite materials [7 – 9].
First-principles simulations of electrostriction have not been widely pursued or studied in detail. In this letter, we survey existing methodologies and present a new route to the calculation of electrostrictive properties via den- sity functional theory (DFT), which we show to be more efficient, robust, perspicuous, and easy to apply, than previous methods.
Electrostriction describes the quadratic part of elec- tromechanical couplings that are present in any dielec- tric material. When an electric ($E$) or polarisation ($P$) field is applied to a material, a stress or a strain will be induced, depending on the boundary conditions:
$$
\begin{aligned}
x_{ij} &=d_{ijk}E_{k} + M_{ijkl}E_{k}E_{l} , \\
X_{ij} &=e_{ijk}E_{k} + m_{ijkl}E_{k}E_{l} , \\
x_{ij} &=g_{ijk}P_{k} + Q_{ijkl}P_{k}P_{l} , \\
X_{ij} &=h_{ijk}P_{k} + q_{ijkl}P_{k}P_{l} . \\
\end{aligned}
$$
Here $x_{ij}$ and $X_{ij}$ denote the strain and stress ten- sor components, respectively; $d_{ijk}$, $e_{ijk}$, $g_{ijk}$, and $h_{ijk}$ the piezoelectric tensors; and $M_{ijkl}$, $m_{ijkl}$, $Q_{ijkl}$, and $q_{ijkl}$ the electrostriction tensors. Piezoelectricity is ab- sent in centrosymmetric crystals (as well as in the non- centrosymmetric 432 space group), whereas electrostric- tion is present in all crystal classes as well as in non- crystalline materials. [10]
Previous ab initio studies of electrostriction [11 – 18] comprise three methodological classes: (i), works which impose a polarisation via a frozen polar mode, and sub- sequently determine the energy-minimising strain, or the energy coupling between strain and polarisation; (ii), works which perform DFT calculations of system prop- erties (such as stress or strain) under the condition of a fixed electric field; [12, 15] and (iii), works which use the recently developed capacity [19] to perform DFT calcula- tions under conditions of fixed displacement field. [13, 14]
The methods of class (i) utilise various assumptions which limit their applicability to a small class of materi- als. For example Wang et al. [11] impose the condition that electrostrictive strains are volume conserving, which is not true for most materials, including the BaTiO $_{3}$ they study. Of the studies which seek to parametrise Landau- Devonshire potentials [16, 17], many do not directly cal- culate the polarisation using the Berry phase technique, but rather infer it from known values of the Born effective charge tensor or spontaneous polarisations, thus neglect- ing the electronic contribution to the electrostrictive co- efficients. Furthermore, fitting to energy differences be- tween different strain states [16 – 18] results in further un- certainties due to a changing basis set, [20] and increased complexity in fitting equations.
The studies in classes (ii) and (iii) use finite field tech- niques. [19, 21, 22] While these studies do not utilise the same restrictive assumptions as those of class (i), there are yet subtleties which must be accounted for. The first of these is the entanglement of electrostriction and non-linear piezoelectricity in non-centrosymmetric crys- tals [12, 15]. More generally, all calculations under finite field have a fundamental limitation on k-points and band gap to prevent Zener breakdown. [21]
However, not all viable methods by which to compute electrostriction are represented in the literature. Here, we aim at presenting and comparing possible method- ologies to compute electrostriction from thermodynamic considerations [10, 23, 24] and DFT calculations. We
| |
201203841/3
|
FIG. 1. (a) Strain as a function of electric field. (b) Strain as a function of polarisation. Notations have been reduced to two indices: the electric indices have to be identical (i.e. only $M_{ijkk}$ with $k$ =1..3); and Voigt notation is used for the mechanical indices. (c) Variation of relative susceptibility $\chi_r$, and relative dielectric stiffness $\eta_r$ (= $\chi_r^{-1}$) with hydrostatic stress and strain. Calculated $\chi_r$ ($\eta_r$) values are given by red squares (blue circles), with a linear fitting given by the solid red (blue) line, and a quadratic fitting given by a dashed red (blue) line.
hydrostatic coefficients calculated directly from hydro- static strains/pressure. These results are summarised in Table. I, where the coefficients subscripted with ’ $h$ ’ are obtained directly from hydrostatic data, as shown in Fig. 1 (c).
\begin{tabularx}{\columnwidth}{>{\setlength\hsize{1\hsize}\centering}X c >{\setlength\hsize{1\hsize}\centering}X c}
\hline\hline
Coefficient
&
$\frac{\partial\epsilon_{0}\chi,\nicefrac{\eta}{\epsilon_{0}}}{\partial x,X}$
&
$\frac{\partial^{2}x,X}{\partial E^{2},D^{2}}$
&
Exp
\\
\hline
$m_{11}$ & -534.7 & -477.7 & - \\
$m_{12}$ & 58.3 & 16.5 & - \\
$m_{h}$ (pNV$^{-2}$) & -418.2 & -444.6 & -396.4\textsuperscript{b,d} \\ \hline
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
$M_{11}$ & 2251 & 1970 & 2020\textsuperscript{a} \\
$M_{12}$ & -682 & -508 & - \\
$M_{h}$ (pm$^{2}$V$^{-2}$) & 883 & 954 & 824\textsuperscript{b} \\ \hline
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
$q_{11}$ & -76.8 & -71.6 & - \\
$q_{12}$ & 8.4 & 2.5 & - \\
$q_{h}$ (GNm$^{2}$C$^{-2}$) & -60.0 & -66.7 & -53.7\textsuperscript{b,d} \\ \hline
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
$Q_{11}$ & 0.323 & 0.292 & 0.33\textsuperscript{c} \\
$Q_{12}$ & -0.098 & -0.075 & - \\
$Q_{h}$ (m$^{4}$C$^{-2}$) & 0.127 & 0.142 & 0.109\textsuperscript{b} \\ \hline
\hline
\end{tabularx}
TABLE I. Comparison of coefficients obtained via Eqs. (1) and Eqs. (2) for MgO. a=Ref[28]; b=Ref[29]; c=Ref[4]; d=Ref[30]. The experimental $M_h$ is that measured in Ref. [29]; values for $Q_h$, $m_h$, and $q_h$, are obtained from this $M_h$ value using the permittivity of MgO provided by Ref. [29], and the bulk modulus found in Ref.[30].
The results obtained with our methodology agree with both the ones obtained under applied fields as well as the experimental values as shown in Table I. Examin- ing the table more closely, we see that Eqs. (2) pro- duces individual tensor components which are larger in magnitude than those obtained from Eqs. (1), but that these coefficients sum up to smaller hydrostatic coeffi- cients. The largest disagreement is found for the smaller transverse coefficients, with the disagreement between the two theoretical methods for the hydrostatic coeffi- cients being less than 10%. Given that previous authors have attributed an uncertainty of 25% to the finite field method by observing differences in values that should be equal by symmetry [12], and that our method exhibits no such differences, we attribute the disagreement between the methods to shortcomings of the finite field method. Both methods also show good agreement with experi- ment, with mean absolute relative errors, over all coeffi- cients, being $<$ 8% for each method, which is reasonable agreement, accounting for the large spread found gener- ally in measurements of electrostriction. [31]
Having guarantied that the different methods to com- pute the electrostrictive coefficients give consistent re- sults, we can now compare their efficiencies. In accor- dance with its aforementioned advantages, we find that our methodology using DFPT to compute the variations of the permittivity runs significantly faster. For example, it is eight times faster to calculate the susceptibility of MgO under a 0.5% strain than to perform a relaxation under an applied field of $\sim$ 1.25 GVm $^{-1}$ with the same settings of k-point grid density and cutoff energy. Fur- thermore, the calculations using our DFPT methodology also converge faster with respect to the k-point grid and plane wave cut-off energy as evidenced in Fig. 2: at a cutoff energy of 45 Ha, all electrostrictive coefficients ob- tained via our DFPT methodology are within 0.15% of their values at 60 Ha whereas they are around 2% with the finite field method. Likewise with k-point conver- gence, compared to the 10 $\times$ 10 $\times$ 10 k-point value: about 0.2% difference with our method versus about 4% with applied fields with a k-point grid density of 4 $\times$ 4 $\times$ 4.
To generalise this validation and better corroborate the method, we have calculated electrostrictive properties for a host of materials: rocksalt MgO, LiF, NaCl, KCl, KBr,
| |
180408558/5
|
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{sx}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)
\left[\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)+\frac{\theta+h'/2}{3\mathcal{H}}
+3\mathcal{H}(c^2_{sx}-w_x)\frac{\theta_x}{k^2}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{sx}^{2})\theta _{x}+\frac{%
c_{sx}^{2}}{(1+w_{x})}k^{2}\delta _{x}+3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\frac{%
\theta _{c}-(1+c_{sx}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
3\mathcal{H}\xi\frac{\rho_x}{\rho_c}\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\delta _{c}-\delta_x-\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)-\frac{\theta+h'/2}{3\mathcal{H}}\right] , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{sx}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)
\left[\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)+\frac{\theta+h'/2}{3\mathcal{H}}
+3\mathcal{H}(c^2_{sx}-w_x)\frac{\theta_x}{k^2}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{sx}^{2})\theta _{x}+\frac{%
c_{sx}^{2}}{(1+w_{x})}k^{2}\delta _{x}+3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\frac{%
\theta _{c}-(1+c_{sx}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
3\mathcal{H}\xi\frac{\rho_x}{\rho_c}\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\delta _{c}-\delta_x-\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)-\frac{\theta+h'/2}{3\mathcal{H}}\right] , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{sx}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)
\left[\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)+\frac{\theta+h'/2}{3\mathcal{H}}
+3\mathcal{H}(c^2_{sx}-w_x)\frac{\theta_x}{k^2}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{sx}^{2})\theta _{x}+\frac{%
c_{sx}^{2}}{(1+w_{x})}k^{2}\delta _{x}+3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\frac{%
\theta _{c}-(1+c_{sx}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
3\mathcal{H}\xi\frac{\rho_x}{\rho_c}\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\delta _{c}-\delta_x-\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)-\frac{\theta+h'/2}{3\mathcal{H}}\right] , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{sx}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)
\left[\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)+\frac{\theta+h'/2}{3\mathcal{H}}
+3\mathcal{H}(c^2_{sx}-w_x)\frac{\theta_x}{k^2}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{sx}^{2})\theta _{x}+\frac{%
c_{sx}^{2}}{(1+w_{x})}k^{2}\delta _{x}+3\mathcal{H}\xi\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\frac{%
\theta _{c}-(1+c_{sx}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
3\mathcal{H}\xi\frac{\rho_x}{\rho_c}\exp\left(\frac{\rho_x}{\rho_c}-1\right)\left[\delta _{c}-\delta_x-\frac{\rho_x}{\rho_c}(\delta_x-\delta_c)-\frac{\theta+h'/2}{3\mathcal{H}}\right] , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
\begin{tabular}{c|c c}
Parameter & Prior (IDE) \hline
$\Omega_{b} h^2$ & $[0.005,0.1]$ $\tau$ & $[0.01,0.8]$ $n_s$ & $[0.5, 1.5]$ $\log[10^{10}A_{s}]$ & $[2.4,4]$ $100\theta_{MC}$ & $[0.5,10]$ $w_x$ & $-$ $\xi$ & $[0, 2]$
\end{tabular}
TABLE I: Summary of the flat priors on the parameters for the interacting model (3).
## OBSERVATIONAL DATA, FITTING TECHNIQUE AND THE RESULTS
The observational data, methodology and the results of the exponential interaction model are described in that Section.
We consider several observational data to constrain the current interaction model as follows:
- Cosmic microwave background radiation from Planck [68, 69]. The data is recognized as Planck TTTEEE+low TEB.
- Baryon acoustic oscillation (BAO) distance mea- surements from the 6dF Galaxy Survey (6dFGS) (redshift measurement at $z_{\emph{\emph{eff}}}=0.106$) [70], Main Galaxy Sample of Data Release 7 of Sloan Digital Sky Survey (SDSS-MGS) ($z_{\emph{\emph{eff}}}=0.15$) [71], CMASS and LOWZ samples from the latest Data Release 12 (DR12) of the Baryon Oscillation Spectroscopic Survey (BOSS) ($z_{\mathrm{eff%
}}=0.57$) and ($z_{\mathrm{eff}}=0.32$) [72].
- Redshift space distortion (RSD) data from CMASS sample ($z_{\mathrm{eff%
}}=0.57$) [73] and the LOWZ sample ($z_{\mathrm{eff}%
}=0.32$).
- The weak gravitational lensing (WL) data from the Canada $-$ France $-$ Hawaii Telescope Lensing Survey (CFHTLenS) [74, 75].
- Joint light curve analysis (JLA) sample [76] from in the redshift interval $z \in [0.01, 1.30]$ comprising 740 measurements.
- Latest cosmic chronometers (CC) measurements spanned in the redshift interval $0< z < 2$ [77].
- The current estimated value of the Hubble param- eter from the Hubble space telescope (HST) yieling $H_0= 73.02 \pm 1.79$ km/s/Mpc with 2.4 precision [78]. We identify this data as HST.
We use the markov chain monte carlo package cos- momc [79,80] to constrain the model. This is an efficient simulation where the convergence of the model parame- ters is based on the Gelman-Rubin statistics [81] that may result in a sufficient convergence of all model pa- rameters. The parameters space for the IDE scenario is
$$
\begin{aligned}
\mathcal{P}_2 \equiv\Bigl\{\Omega_bh^2, \Omega_{c}h^2, 100 \theta_{MC},
\tau, w_x, \xi, n_s, log[10^{10}A_S]\Bigr\},
\end{aligned}
$$
which is eight dimensional. Here, $\Omega_bh^2$, $\Omega_{c}h^2$, are the baryons and cold dark matter density respectively; $100 \theta_{MC}$, is the ratio of sound horizon to the angular di- ameter distance, $\tau$, is the optical depth; $w_x$ is the equa- tion of state parameter for dark energy; $\xi$ is the coupling strength; $n_s$, $A_S$, are respectively the scalar spectral in- dex, and the amplitude of the initial power spectrum.
Now let us come to the observational constraints on the model. To constrain the entire interacting scenario we have used four different observational data, namely,
- Planck TTTEEE $+$ lowTEB (CMB),
- CMB $+$ BAO $+$ RSD,
- CMB $+$ BAO $+$ HST,
- CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC.
| |
180408558/7
|
FIG. 3: We display the 68% and 95% confidence-level contour plots for various combinations of the model parameters for the exponential interaction model using the different combined analysis. The parameter $\Omega_{m0}$ is the present value of the total matter density parameter $\Omega_m = \Omega_b +\Omega_c$ and $H_0$ is the current value of the Hubble parameter in the units km/Mpc/s. From the upper panel one can see that the coupling strength $\xi$ is uncorrelated with the parameters $w_x$, $H_0$ and $\Omega_{m0}$, while from the lower panel one can clearly observe the existing correlations amongst the parameters $w_x$, $H_0$ and $\Omega_{m0}$.
at 95% CL, and $\xi < 0.0172$ at 99% CL, which definitely suggest for a weak interaction scenario. The suggestion of weak interaction is also followed by other observational combinations in this work. Additionally, concerning the observational constraints on the dark energy equation of state, we have some different observations. As from Ta- ble II, one can see that for the first two analyses, namely, CMB alone and CMB $+$ BAO $+$ RSD, the dark energy state parameter is found to be of quintessence type while for the remaining two analyses, its phantom character is suggested. Moreover, we note that for the analysis, CMB $+$ BAO $+$ HST, $w_x < -1$, is preferred in 68% CL. The addition of other external data sets, namely WL, JLA and CC into this data set (i.e. CMB $+$ BAO $+$ HST) shrinks the parameters space for $w_x$ constraining, $%
w_{x}=-1.0168_{-0.0331}^{+0.0407}$(at 68 CL) which shows that the quintessence regime is also not excluded but of course the dark energy state parameter is close to the cosmolog- ical constant limit, ‘$w_x = -1$’. We further note that for all the observational data sets,$w_x$is actually very close to the cosmological constant boundary ‘$w_x = -1$’. Since the coupling strength is very small and$w_x$is close to ‘$-1$’ boundary, thus, one can find that the current interaction model is quite close to that of the non-interacting$\Lambda$ CDM cosmological model.
In Fig. 4 we also show the dependence of the matter fluctuation amplitude $\sigma _{8}$ with different model parame- ters which clearly shows that $\sigma _{8}$ is correlated with the coupling strength $\xi$ and also with the CDM density pa- rameter $\Omega _{m0}$. Certainly, a higher coupling in the dark sector allows higher values of $\sigma _{8}$. One important fea- ture we observe is that, the parameter $\sigma_8$ takes larger values (for all combined analyses) in presence of an in- teraction in the dark sector while in absense of the in- teraction, $\sigma_8$ takes lower values 2. The allowance of inter- action may increase the values of $\sigma_8$, is already explored in [82]. This is the first evidence which demonstrates that the exponential interaction (3) which although al- lows a very small coupling between the dark sectors but might present a slight different behaviour compared to the non-interacting $\Lambda$ CDM cosmological model. That is
2 The estimations of $\sigma_8$ for the non-interacting $\Lambda$ CDM model using different observational data are enlisted in [5].
| |
180408558/2
|
dark matter and dark energy are of same order at current time? ” also known as coincidence problem [21]. Conse- quently, it was found that the old concept of interaction between fields [20] can explain the cosmic coincidence problem [22]. Following this, a large amount of investi- gations [23–45] have been performed. Recently a series of investigations toward the same direction comment that the astronomical data available today do not completely rule out the possibility of a non-zero interaction in the dark sector [46–59]. Additionally, some most recent ar- ticles in this context argue that the interaction in the dark sector could be a very fantastic theory that may release the tension on the local Hubble constant [54, 55], a most talkative issue in modern cosmology at present. Moreover, it has been found that the presence of interac- tion in the dark sector pushes the dark energy fluid into the phantom region [28, 35, 57, 58]. On the other hand, interaction cosmologies can describe, in a phenomenolog- ical way, the unified dark energy models, for instance see [60, 61]. Thus, the interacting models having the above features clearly demand for more investigations in recent years.
In the current work we investigate the cosmological constraints allowing an exponential interaction between dark matter and dark energy. The choice of an expo- nential interaction is indeed phenomenological, however it cannot be excluded on the basis of other interaction models that have been widely studied in the last couple of years. We consider such an interaction in order to investi- gate their ability with the observational data. For metric which describes the geometry of the universe we consider the spatially flat Friedmann-Lema ˆı tre-Robertson-Walker (FLRW) line element. Moreover, the dark components are assumed to have barotropic nature. The scenario has been fitted using the latest astronomical measurements from various data sets and the markov chain monte carlo package cosmomc has been used to extract the observa- tional constraints of the model. It is quite interesting to note that even if we allow an exponential interaction in the dark sector, the resulting scenario does not deviate much from the $\Lambda$ CDM cosmology. This might be consid- ered to be an interesting result in the field of interacting cosmologies because this reflects that although any arbi- trary choice for an interaction model can be made, but the observational data may not allow a strong interaction in the dark sector.
The presentation of the manuscript is as follows. In section 2 we describe the gravitational equations of the interacting universe at the background and perturbative levels. Section 3 describes the observational data em- ployed in this work, fitting technique, and the results of the analysis. Finally, section 4 closes the entire work with a short summary.
## GRAVITATIONAL EQUATIONS IN AN INTERACTING UNIVERSE: BACKGROUND AND PERTURBATIONS
In this section we describe the background and pertur- bation equations for the interacting dark fluids. Specif- ically, we consider a model of our universe where the total energy density of the universe is contributed by rel- ativistic (radiation) and non-relativistic species (baryons, pressureless dark matter and dark energy). The fluids are barotropic where dark matter and dark energy interact with each other while the radiation and baryons do not take part in the interaction. We denote $(p_{i},\rho _{i})$ as the pressure and energy density of the $i$ -th component of the fluid where $%
i=r,b,c,x$ respectively represent the radia- tion, baryons, pressureless dark matter and dark energy.
Now, considering a spatially flat FLRW line element for the universe with expansion scale factor $a(t)$, the con- servation equations for the interacting fluids follow
$$
\begin{aligned}
\dot{\rho}_{c}+3\frac{\dot{a}}{a}\rho _{c} &=&-Q, \\
\dot{\rho}_{x}+3\frac{\dot{a}}{a}(1+w_{x})\rho _{x} &=&Q,
\end{aligned}
$$
$$
\begin{aligned}
\dot{\rho}_{c}+3\frac{\dot{a}}{a}\rho _{c} &=&-Q, \\
\dot{\rho}_{x}+3\frac{\dot{a}}{a}(1+w_{x})\rho _{x} &=&Q,
\end{aligned}
$$
where $w_{x}=p_{x}/\rho _{x}$ is the equation of state parameter for the dark energy fluid which we assume to be con- stant. And $Q$ is the interaction rate between the dark fluids which determines the direction of energy flow be- tween them. For $Q< 0$, the energy flow takes place from DE to CDM whereas the energy flow from CDM to DE is conferred by $Q >0$. The conservation equations for radiation and baryons are the usual ones and they respec- tively take the forms $\rho _{r}=\rho _{r0}a^{-4}$, $\rho
_{b}=\rho _{b0}a^{-3}$. Here,$\rho _{i,0}$($i=r,b$) is the value of$\rho
_{i}$at current time for the$i$ -th fluid.
The Hubble equation takes the form
\begin{equation*} \left(\frac{\dot{a}}{a}\right) ^{2}=\frac{8\pi G}{3}\left(\rho _{r}+\rho _{b}+\rho _{c}+\rho _{x}\right) ~, \end{equation*}
which together with the conservation equations for all fluids ((1) and (2)) for pressureless dark matter, dark energy respectively and two for radiation and baryons), can determine the dynamics of the universe, provided the interaction rate $Q$, is supplied from outside. Technically, there is no such specific rule to select the forms for $Q$ and thus some phenomenological choices are initially made and then they are tested with the astronomical data. The well known models for the interaction rates are, $Q\propto \rho _{x}$ [67], $Q\propto \rho _{c}$, $Q\propto (\rho _{c}+\rho _{x})$, $Q\propto \dot{\rho}_{x}$, $Q \propto \rho_x^2/\rho_c$ etc.
We remark that the establishment of those interactions in the current literature followed from their agreement with the observational data and their stabilities at large- scale, and thus a new interaction appearing in the lit- erature should be equally welcomed. In this work we propose the following interaction
| |
180408558/6
|
\begin{tabular}{ccccccccccccc}
\hline\hline
Parameters & CMB & CMB+BAO+RSD & CMB+BAO+HST &
$\begin{array}[c]{c}
\text{CMB+BAO+RSD+HST}+\text{WL+JLA+CC} \end{array}$ &
\hline
$\Omega_b h^2$ & $0.02214_{- 0.00018- 0.00036}^{+ 0.00019+ 0.00034}$ & $0.02226_{- 0.00014- 0.00029}^{+ 0.00014+ 0.00029}$ & $0.02225_{- 0.00016- 0.00030}^{+ 0.00014+ 0.00031}$ & $0.02232_{- 0.00014- 0.00028}^{+ 0.00015+ 0.00026}$ $\Omega_c h^2$ & $0.1154_{- 0.0027- 0.0091}^{+ 0.0050+ 0.0077}$ & $0.1118_{- 0.0034- 0.0108}^{+ 0.0068+ 0.0085}$ & $0.1143_{- 0.0025- 0.0067}^{+ 0.0043+ 0.0057}$ & $0.1139_{- 0.0024- 0.0075}^{+ 0.0043+ 0.0061}$ $100\theta_{MC}$ & $1.04066_{- 0.00040- 0.00085}^{+ 0.00038+ 0.00091}$ & $1.04095_{- 0.00047- 0.00088}^{+ 0.00044+ 0.00092}$ & $1.04079_{- 0.00035- 0.00072}^{+ 0.00035+ 0.00070}$ &$1.04090_{- 0.00041- 0.00078}^{+ 0.00038+
0.00078}
$$
n_s$&$ 0.9715_{- 0.0043- 0.0183}^{+ 0.0059+ 0.0115}$&$ 0.9751_{- 0.0042- 0.0081}^{+ 0.0042+ 0.0083}$&$ 0.9750_{- 0.0044- 0.0088}^{+ 0.0044+ 0.0083}$&$0.9769_{- 0.0044- 0.0079}^{+ 0.0044+ 0.0083}
$$
\tau$&$ 0.073_{- 0.018- 0.038}^{+ 0.018+ 0.035}$&$ 0.075_{- 0.018- 0.034}^{+ 0.018+ 0.036}$&$ 0.081_{- 0.018- 0.036}^{+ 0.019+ 0.034}$&$0.068_{- 0.018- 0.038}^{+ 0.019+ 0.038}
$$
\mathrm{ln}(10^{10} A_s)$&$ 3.091_{- 0.037- 0.072}^{+ 0.036+ 0.068}$&$ 3.092_{- 0.035- 0.068}^{+ 0.035+ 0.069}$&$ 3.104_{- 0.036- 0.070}^{+ 0.039+ 0.066}$&$3.076_{- 0.038- 0.075}^{+ 0.037+ 0.078}
$$
w_x$&$ -0.9961_{- 0.0630- 0.1402}^{+ 0.0624+ 0.1349}$&$ -0.9756_{- 0.0377- 0.0986}^{+ 0.0574+ 0.0866}$&$ -1.0860_{- 0.0454- 0.1064}^{+ 0.0530+ 0.1110}$&$-1.0168_{- 0.0331- 0.0688}^{+ 0.0407+ 0.0664}
$$
\xi$&$ 0.0081_{- 0.0081- 0.0081}^{+ 0.0029+ 0.0161}$&$ 0.0106_{- 0.0106- 0.0106}^{+ 0.0028+ 0.0128}$&$ 0.0062_{- 0.0048- 0.0062}^{+ 0.0029+ 0.0065}$&$0.0058_{- 0.0058- 0.0058}^{+ 0.0014+ 0.0085}
$$
\Omega_{m0}$&$ 0.309_{- 0.023- 0.041}^{+ 0.022+ 0.042}$&$ 0.300_{- 0.016- 0.030}^{+ 0.017+ 0.029}$&$ 0.280_{- 0.012- 0.026}^{+ 0.016+ 0.024}$&$0.292_{- 0.011- 0.020}^{+ 0.010+ 0.020}
$$
\sigma_8$&$ 0.919_{- 0.112- 0.174}^{+ 0.050+ 0.223}$&$ 0.966_{- 0.172- 0.201}^{+ 0.057+ 0.304}$&$ 0.969_{- 0.121- 0.159}^{+ 0.053+ 0.204}$&$0.905_{- 0.098- 0.126}^{+ 0.038+ 0.183}
$$
H_0$&$ 67.01_{- 1.98- 3.56}^{+ 1.93+ 3.90}$&$ 67.04_{- 1.36- 2.267}^{+ 0.10+ 2.36}$&$ 70.02_{- 1.22- 2.61}^{+ 1.26+ 2.52}$&$68.48_{- 0.78- 1.72}^{+ 0.90+ 1.56}$ \hline\hline
\end{tabular}
TABLE II: Observational constraints at 68% ($1\protect\sigma$), 95% confidence-levels ($2%
\protect\sigma$) on the model parameters for the interacting scenario with constant dark energy equation of state have been displayed using the observational analyses shown in the table. We recall that here$%
\Omega_{m0}$is the current value of$\Omega_m (= \Omega_b +\Omega_c$).
FIG. 2: One dimensional posterior distributions of some selected parameters of the interacting model have been shown for different combined analysis employed in this work.
Using the priors for the model parameters summarized in Table I and then performing a likelihood analysis using cosmomc, in Table II, we summarize the results. In Fig. 2 we show the one-dimensional posterior distributions for some selected parameters of the interacting scenario for the employed observational analyses. Further, in Fig. 3, we display the contour plots for different combinations of the free as well as the derived parameters using different combined analysis mentioned above.
Our analyses show that the observational data allow a very small interaction in the dark sector which is con- sistent with the non-interaction limit, $\xi =0$. One strin- gent point we we notice is that, for the observational data CMB $+$ BAO $+$ HST, $\xi = 0$ is not allowed at least within 68% confidence-level (CL), but in the 95% CL, the non- interacting scenario is recovered. The lowest coupling strength as observed from Table II is attained for the fi- nal combined analysis (CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC) where $\xi =0.0058_{-0.0058}^{+0.0014}$ at 68 CL. In fact, for this particular combined analysis, $\xi < 0.0143$
| |
180408558/4
|
earlier works [66, 67] the momentum transfer potential is specialized to be the simplest physical choice, which becomes zero in the rest frame of the dark matter, that means, we have the following equation $k^2f_A=Q_A(\theta-\theta_c)$.
Now, introducing $\delta _{A}=\delta \rho _{A}/\rho _{A}$, as the density pertur- bation for the fluid $A$, and assuming no anisotropic stress (i.e. $%
\pi _{A}=0$), in the synchronous gauge, that means with the conditions$\phi
=B=0$,$\psi =\eta$, and$k^{2}E=-h/2-3\eta$), the explit perturbation equations (density and velocity perturbations) can be written as [65–67]
$$
\begin{aligned}
\delta_A^{\prime} + 3 \mathcal{H} \left(c_{sA}^2 - w_A \right) \delta_A + 9 \mathcal{H}^2 \left(1+w_A \right) \left(c_{sA}^2- c_{aA}^2 \right)\frac{\theta_A}{k^2} + \left(1+w_A \right) \theta_A -3 \left(1+w_A \right) \psi^{\prime} + \left(1+w_A \right) k^2 \left(B- E^{\prime} \right)\\ = \frac{a}{\rho_A} \left(\delta Q_A - Q_A \delta _A \right) + \frac{a Q_A}{\rho_A} \left[\phi + 3 \mathcal{H} \left(c_{sA}^2- c_{aA}^2 \right)\frac{\theta_A}{k^2}\right],\\
%
%
\theta_A^{\prime} + \mathcal{H} \left(1-3 c_{sA}^2 \right)\theta_A - \frac{c_{sA}^2}{1+w_A} k^2 \delta_A -k^2 \phi = \frac{a}{(1+w_A)\rho_A} \Bigl[\left(Q_A \theta -k^2 f_A \right) - \left(1+ c_{sA}^2 \right) Q_A \theta_A \Bigr],
\end{aligned}
$$
$$
\begin{aligned}
\delta_A^{\prime} + 3 \mathcal{H} \left(c_{sA}^2 - w_A \right) \delta_A + 9 \mathcal{H}^2 \left(1+w_A \right) \left(c_{sA}^2- c_{aA}^2 \right)\frac{\theta_A}{k^2} + \left(1+w_A \right) \theta_A -3 \left(1+w_A \right) \psi^{\prime} + \left(1+w_A \right) k^2 \left(B- E^{\prime} \right)\\ = \frac{a}{\rho_A} \left(\delta Q_A - Q_A \delta _A \right) + \frac{a Q_A}{\rho_A} \left[\phi + 3 \mathcal{H} \left(c_{sA}^2- c_{aA}^2 \right)\frac{\theta_A}{k^2}\right],\\
%
%
\theta_A^{\prime} + \mathcal{H} \left(1-3 c_{sA}^2 \right)\theta_A - \frac{c_{sA}^2}{1+w_A} k^2 \delta_A -k^2 \phi = \frac{a}{(1+w_A)\rho_A} \Bigl[\left(Q_A \theta -k^2 f_A \right) - \left(1+ c_{sA}^2 \right) Q_A \theta_A \Bigr],
\end{aligned}
$$
where the prime is the differentiation with respect to the conformal time $%
\tau$;$\mathcal{H}$is the conformal Hubble parame- ter;$c_{sA}^2$,$c_{aA}^2$, are respectively the adiabatic and physical sound velocity for the fluid$%
A$related as$c_{aA}^{2}=p_{A}^{\prime}/\rho _{A}^{\prime
}=w_{x}+w_{x}^{\prime}/(\rho _{A}^{\prime}/\rho _{A})$;$\theta =
\theta_{\mu}^{\mu}$is the volume expansion scalar. To avoid any kind of instabilities,$c_{sA}^2 \geq 0$has been as- sumed. For cold dark matter, since$w_c =0$, thus, one has$c_{sc}^2 =0$. On the other hand, for dark energy fluid we assume$c_{sx}^2 = 1$ [65–67]. Now, one can write down the density and the velocity perturbations for the dark energy and cold dark matter as
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{s,x}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&\frac{aQ}{\rho _{x}}\left[-\delta _{x}+\frac{\delta Q}{Q}+3\mathcal{H}%
(c_{s,x}^{2}-w_{x})\frac{\theta _{x}}{k^{2}}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{s,x}^{2})\theta _{x}+\frac{%
c_{s,x}^{2}}{(1+w_{x})}k^{2}\delta _{x}+\frac{aQ}{\rho _{x}}\left[\frac{%
\theta _{c}-(1+c_{s,x}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
\frac{aQ}{\rho _{c}}\left(\delta _{c}-\frac{\delta Q}{Q}\right) , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{s,x}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&\frac{aQ}{\rho _{x}}\left[-\delta _{x}+\frac{\delta Q}{Q}+3\mathcal{H}%
(c_{s,x}^{2}-w_{x})\frac{\theta _{x}}{k^{2}}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{s,x}^{2})\theta _{x}+\frac{%
c_{s,x}^{2}}{(1+w_{x})}k^{2}\delta _{x}+\frac{aQ}{\rho _{x}}\left[\frac{%
\theta _{c}-(1+c_{s,x}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
\frac{aQ}{\rho _{c}}\left(\delta _{c}-\frac{\delta Q}{Q}\right) , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{s,x}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&\frac{aQ}{\rho _{x}}\left[-\delta _{x}+\frac{\delta Q}{Q}+3\mathcal{H}%
(c_{s,x}^{2}-w_{x})\frac{\theta _{x}}{k^{2}}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{s,x}^{2})\theta _{x}+\frac{%
c_{s,x}^{2}}{(1+w_{x})}k^{2}\delta _{x}+\frac{aQ}{\rho _{x}}\left[\frac{%
\theta _{c}-(1+c_{s,x}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
\frac{aQ}{\rho _{c}}\left(\delta _{c}-\frac{\delta Q}{Q}\right) , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
$$
\begin{aligned}
\delta _{x}^{\prime} &=&-(1+w_{x})\left(\theta _{x}+\frac{h^{\prime}}{2}%
\right) -3\mathcal{H}(c_{s,x}^{2}-w_{x})\left[\delta _{x}+3\mathcal{H}%
(1+w_{x})\frac{\theta _{x}}{k^{2}}\right] -3\mathcal{H}w_{x}^{\prime}\frac{%
\theta _{x}}{k^{2}} \notag \\
&+&\frac{aQ}{\rho _{x}}\left[-\delta _{x}+\frac{\delta Q}{Q}+3\mathcal{H}%
(c_{s,x}^{2}-w_{x})\frac{\theta _{x}}{k^{2}}\right] , \\
\theta _{x}^{\prime} &=&-\mathcal{H}(1-3c_{s,x}^{2})\theta _{x}+\frac{%
c_{s,x}^{2}}{(1+w_{x})}k^{2}\delta _{x}+\frac{aQ}{\rho _{x}}\left[\frac{%
\theta _{c}-(1+c_{s,x}^{2})\theta _{x}}{1+w_{x}}\right] , \\
\delta _{c}^{\prime} &=&-\left(\theta _{c}+\frac{h^{\prime}}{2}\right) +%
\frac{aQ}{\rho _{c}}\left(\delta _{c}-\frac{\delta Q}{Q}\right) , \\
\theta _{c}^{\prime} &=&-\mathcal{H}\theta _{c},
\end{aligned}
$$
where $\delta Q/Q$ includes the perturbation term for the Hub- ble expansion rate $\delta H$. One may note that in the evolu- tion equation for $\theta _{c}^{\prime}$, no interaction term is present. This is because, since for the cold dark matter species, $c^2_{sc}=0$ has been assumed, and $k^2f_c=Q_c(\theta-\theta_c)$, thus, the term inside the third brace of the right hand side of eqn. (??) actually vanishes. Now, for the interaction model (3), the explicit evolution for density and velocity perturbations are
| |
180408558/1
|
# Cosmological constraints on an exponential interaction in the dark sector
Weiqiang Yang, 1, ∗
Supriya Pan, 2, † and
Andronikos Paliathanasis 3,4,5, ‡
Cosmological models where dark matter (DM) and dark energy (DE) interact with each other are the general scenarios in compared to the non-interacting models. The interaction is usually motivated from the phenomenological ground and thus there is no such rule to prefer a particular interaction between DM and DE. Being motivated, in this work, allowing an exponential interaction between DM and DE in a spatially flat homogeneous and isotropic universe, we explore the dynamics of the universe through the constraints of the free parameters where the strength of the interaction is characterized by the dimensionless coupling parameter $\xi$ and the equation of state (EoS) for DE, $w_x$, is supposed to be a constant. The interaction scenario is fitted using the latest available observational data. Our analyses report that the observational data permit a non-zero value of $\xi$ but it is very small and consistent with $\xi =0$. From the constraints on $w_x$, we find that both phantom ($w_x< -1$) and quintessence ($w_x> -1$) regimes are equally allowed but $w_x$ is very close to ‘ $-1$ ’. The overall results indicate that at the background level, the interaction model cannot be distinguished from the base $\Lambda$ -cold dark matter model while from the perturbative analyses, the interaction model mildly deviates from the base model. We highlight that, even if we allow DM and DE to interact in an exponential manner, but according to the observational data, the evidence for a non-zero coupling is very small.
## INTRODUCTION
According to a large number of independent astronom- ical surveys [1–5], our universe is currently expanding with an acceleration. This accelerating phase does not fit into the standard cosmological model requiring the presence of some negative pressure component fluid in the universe sector dubbed as dark energy. And from the current astronomical estimation, this so-called dark energy fluid occupies almost 68% of the total energy den- sity of the universe. The rest 32% of this energy density is filled up by a pressureless dark matter fluid (also called as cold dark matter) and baryons, radiation. The com- mon behaviour in both dark matter and dark energy is that, both are unknown to us by its origin, character, dynamics for instance. The above picture can be framed in terms of the $\Lambda$ CDM cosmology where the dark en- ergy fluid is represented by some cosmological constant, $\Lambda
>0$ and CDM is the cold dark matter. But, as well known, the problem with the cosmological constant [6] leads to several alternative models [7–17] trying to ex- plain this accelerating phase so that the observational data can match with the theoretical models at hand.
Among various cosmological models, a particular class of models where the underlying fluids may interact with each other, widely known as interacting cosmologies, gained a significant attention to modern cosmological re- search. In interacting cosmologies, usually the gravita- tional theory is assumed to be described by the General theory of Relativity and the main two fluids of the uni- verse describing its dark sector, namely, the dark matter and dark energy, are allowed to interact with each other 1. In particular, the total fluid of the dark sector is con- served. For a detailed understanding of the interacting cosmologies, we refer to some recent reviews [18, 19].
One may note that the origin of interaction was not to explain the current accelerating universe rather its pri- mary motivation was to find a possible explanation to- wards the cosmological constant problem [20] which as well known to the cosmological community, is existing since long back ago and remained silent until the dark energy era began. When the alternative $%
\Lambda $ CDM mod- els appeared in the literature, it was found that they raised a problem which asks “ why the energy densities of
1 Technically, one may allow all components of the universe to in- teract with each other, however, the most viable theory is the interaction between dark matter and dark energy, since for other components, if allowed to interact with each other may lead to some unphysical situations, like some inflexible “ fifth-force ” con- straints.
| |
180408558/3
|
$$
Q=3H\xi \rho _{x}\exp \left(\frac{\rho _{x}}{\rho _{c}}-1\right),
$$
where $\xi$ is the coupling strength of the interaction. One can see that in terms of the coincidence parame- ter $r=\rho _{c}/\rho _{x}$, the interaction (3) can be recast as $Q=3H\xi \rho _{x}\exp \left(\frac{1}{r}%
-1\right) $, and thus, for$r\rightarrow \infty$,$Q\simeq 3H\xi \rho _{x}$while for$r\rightarrow 1$,$Q\simeq 3H\xi \rho _{x}.$Those limits have been studied extensively in the bibliography, see [66] and ref- erences therein. In general, the majority of the interac- tion models are linear functions on the energy densities. There are a few nonlinear models [18] which however do not provide the linear interactions in the limit. On the other hand, for the exponential interaction (3), its linear and nonlinear behaviour are still retained. As one can see, for$r \rightarrow 1$, and$r\rightarrow \infty$it mimics the linear interac- tion scenario$Q \propto \rho_{x}$, while on the other hand, it may also provide quadratic terms in the interaction rate as the first corrections in the linear case. One can check that the Taylor series expansion of (3) around$\rho_x =0$, one gets
$$
\begin{aligned}
Q \propto \rho_x + \frac{\rho_x^2}{\rho_c}+...
\end{aligned}
$$
In Fig. 1, we describe the qualitative evolution of the exponential interaction model (3), denoted by $Q_e$ for dif- ferent values of the coupling parameter. We also made a comparison between the interaction models. From the comparison, we see that the model $Q_2$ always presents a very different behaviour in compared to the exponen- tial model as well as with other interaction models. A common behaviour we notice from the analysis is that, the exponential model (3) behaves similarly to other two interaction models ($Q_1$, $Q_3$); however, the exponential model leaves a notable deviation around a very small neighbourhood of $z =0$. We also observe from Fig. 1 that, for large redshifts the exponential interaction is only differentiated from other two interaction models ($Q_1$, $Q_3$) only for large coupling parameter.
Now, for any cosmological model, one must ensure its stability in the large scale of the universe, and thus, we need to study the perturbation equations. In order to do that, we consider the perturbed FLRW metric with scalar mode $k$ given by [62–64]
$$
\begin{aligned}
ds^{2} &=&a^{2}(\tau)\Bigg[-(1+2\phi)d\tau ^{2}+2\partial _{i}Bd\tau dx^{i}
\notag \\
&&+\Bigl((1-2\psi)\delta _{ij}+2\partial _{i}\partial _{j}E\Bigr)%
dx^{i}dx^{j}\Bigg],
\end{aligned}
$$
where $\tau$ is the conformal time and the quantities $\phi$, $B$, $\psi$, $E$, represent the gauge-dependent scalar perturbations.
The perturbation equations for the metric (5) follow
FIG. 1: We compare the exponential interaction model (3) with some known interaction models, namely, $Q_1 = 3 H \xi \rho_x$, $Q_2 = 3 H \xi \rho_c$ and $Q_3 = 3 H \xi \rho_c \rho_x/(\rho_c +\rho_x)$ for some specific values of the coupling parameter $\xi$ as $\xi = 0.001$ (upper panel), $\xi = 0.05$ (middle panel) and $\xi = 0.01$ (lower panel). We note that $Q_0 = H_0 \rho_{tot,0} = 3 H_0^3 /(8 \pi G)$ where $\rho_{tot,0}$ is the present value of the total energy density $\rho_{tot}$ of the universe, i.e. $\rho_{tot} = \left(\rho _{r}+\rho_{b}+\rho _{c}+\rho _{x}\right)$. The introduction of $Q_0$ makes the quantities $Q_i/Q_0$ ($i \in \{e, 1, 2, 3,\}$) dimensionless.
$$
\nabla _{\nu}T_{A}^{\mu \nu}=Q_{A}^{\mu},~~~~\sum\limits_{\mathrm{A}}{%
Q_{A}^{\mu}}=0,
$$
where we have used $A$ just to represent the fluid (either dark matter or dark energy); $Q_{A}^{\mu}=(Q_{A}+\delta Q_{A})u^{\mu}+F_A^{\mu}$ in which $Q_A$ is the energy transfer rate and $F_A^{\mu} = a^{-1} (0, \partial^{i}
f_A)$is the momentum density transfer relative to the four-velocity$u^{\mu
}$. Let us note that following the
| |
180408558/9
|
FIG. 5: The evolution of the Hubble rate (left panel) and the density parameters for CDM and DE (right panel) for different coupling strengths of the exponential interaction model have been shown for the parameters fixed from the mean values of the combined analysis CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC.
FIG. 6: The plots show how the coupling strength affects the CMB spectra (left panel) and the matter power spectra (right panel). We note that while drawing the plots we take the mean values of the remaining parameters from the combined analysis CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC.
## SUMMARY AND CONCLUSIONS
An interacting scenario between a pressureless dark matter and a dark energy fluid where both of them have constant barotropic state parameters, has been stud- ied. The background geometry is described by the usual FLRW line-element with no curvature.
The speciality of this work is the consideration of an exponential interaction in the dark sector, and then to see how an exponential interaction affects the entire dynam- ics of the universe as it is expected that the exponential character of the interaction rate might affect the back- ground and perturbative evolutions in an extensive way. We note that the exponential interaction is the simplest generalization of the linear interaction scenario [29, 52]. Thus, allowing such an interaction in the dark sectors, we fit the entire cosmological scenario with the markov chain monte carlo package cosmomc [79, 80] which is equipped with a converging diagnostic [81]. Interestingly enough, we find that even if we allow such an exponential na- ture of the interaction rate in the dark sector, the ob- servational data, at present, do not allow the resulting scenario beyond the $\Lambda$ CDM model at least at the back- ground level.
To analyze the model we have constrained the entire interacting scenario using different observational data. We find that the coupling strength, $\xi$, estimated by all the analyses is low and hence a weak interaction limit (i.e. $\xi \sim 0$) is suggested. We also find that, for all the analyses, $\xi =0$ can be recovered within 68% CL (for the analysis with CMB $+$ BAO $+$ HST, $\xi =0$ is re- covered at 95% CL). Thus, one can clearly see that a non-interacting $w_{x}$ CDM cosmology is positively recov- ered by the observational data. Now, concerning the dark energy state parameter, we find that its quintessen- tial and phantom characters are both allowed but in- deed, all the estimations are close to the cosmological constant boundary. In particular, for the analysis with CMB alone and CMB $+$ BAO $+$ RSD, the mean values of $w_x$ are quintessential while for the rest two analyses, that means with CMB $+$ BAO $+$ HST and the final com- bination, CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC, the dark energy state parameter exhibits its phan- tom behavior. The estimated value of the dark-energy-
| |
180408558/8
|
FIG. 4: 68% and 95% confidence level dependence of the matter fluctuation amplitude $\protect\sigma _{8}$ with various model parameters in presence of the exponential interaction in the dark sector. Here too we have shown the figures for different combined analysis as in other plots. From the above figures we find that $\sigma_8$ is uncorrelated with $w_x$, but the remaining combinations do exhibit the correlations.
something which has been derived analytically for a class of general cosmological models [83]. Moreover, in Fig. 5 we show the effects of the coupling parameter on the evolution of the Hubble rate as well as on the density parameters for DM and DE. From this figure (Fig. 5) one can see that as the coupling strength increases, the model deviates from the $%
\Lambda $ -cosmology, as expected, see again [83].
We now move to the analysis of the model at the per- turbative level. The plots have been displayed for the single analysis CMB $+$ BAO $+$ RSD $+$ HST $+$ WL $+$ JLA $+$ CC. At first we measure the effects of the coupling strength on the CMB TT and matter power spectra both shown in Fig. 6 which shows that higher coupling strength is equivalent to significant deviation from the $w_x$ CDM cosmology. The deviation is much pro- nounced from the matter power spectra (right panel of Fig. 6). However, since the estimated values of the cou- pling strength is small (see Table II), thus, it is expected that the model is close to that of the $w_x$ CDM cosmol- ogy as well as the $\Lambda$ CDM cosmology, however, practically that is not true. In order to understand that deviation, in Fig. 7, we demonstrate the relative deviation of the model from the $\Lambda$ CDM model through the CMB TT (left panel of Fig. 7) and matter power spectra (right panel of Fig. 7). In both panels, we see that the interaction model mildly deviates from the $\Lambda$ CDM cosmology and such a mild deviation is only detected from the analyses of the model at the perturbative level $-$ not from the analyses at the backgour level. That means the devia- tion, however small it is, is not detectable only from the estimations of the coupling parameter $\xi$ and the dark en- ergy equation of state, $w_x$
$-$ the
analyses at the level of perturbations are necessary.
Thus, according to the observational data employed in this work, one may notice that a nonzero value of the cou- pling parameter $\xi$ for the present exponential interaction model (3) is allowed, however, the evidence for a non-zero coupling is very small; see the one dimensional posterior distribution for $\xi$ displayed in Fig. 2. And following this, a very mild deviation of the exponential interaction model from the non-interacting $w_x$ CDM cosmology (and from the $\Lambda$ CDM cosmology too) is also allowed by the data, whilst such a deviation is only realized from the analyses at the perturbative level.
| |
230401194/5
|
$n_i \in \left\{{(-1, 1), \allowbreak (-1, 0), ..., (1,1)} \right\}$ is
a regular $3{\times}3$ kernel grid. The convolution is performed on non-uniform positions $({n_i} + \Delta {n_i})$, where ${n_i}$ may be fractional. To tackle the frac- tional values, this operation is implemented with the bilin- ear interpolation.
Reference-Based Feature Enrichment. In the presence of complex pixel displacements among frames, simple align- ment techniques [3, 4, 9] may not able to align burst fea- tures completely. Thus, to fix the remaining minor mis- alignment issues, we propose the reference-based feature enrichment (RBFE) module, shown in Fig. 2(c). RBFE enables additional interaction of aligned frames features $\bar{m}{\bar g}^b$ with the reference frame features $\bar{m}{g}^{b_r}$ to generate consolidated and refined representations. This interactive feature merging is performed via our burst feature fusion (BFF) unit as illustrated in Fig. 2(d). The BFF mecha- nism is built upon the principles of feature back projec- tion [13] and squeeze-excitation techniques [16]. Given the concatenated feature maps of the current frame and the reference frame $[\bar{m}{\bar g}^b, \bar{m}{g}^{b_r}] \in{\mathbb{R}^{1 \times 2\text{*}f \times H \times W}}$, BFF ap- plies BFA to generate representations $\bar{m}{g^b_a}$ encoding the lo- cal non-local context. Overall, BFF yields fused features $\bar{m}{g^b_f}\in{\mathbb{R}^{1 \times f \times H \times W}}$:
$$
{\bar{m}{g^b_f}} = {\bar{m}{g^b_s}} + W\left({{\bar{m}{g^b_a}}} - {\bar{m}{g^b_e}} \right),
$$
where $\bar{m}{g^b_s}{=}{W_s}\bar{m}{g^b_a}$ represents squeezed features and $\bar{m}{g^b_e}{=}W_eW_s\bar{m}{g^b_a}$ are the expanded features. $W_s$ and $W_e$ denote simple convolutions to squeeze and expand feature channels. The squeezed features $\bar{m}{g^b_s}$ poses complementary properties of multiple input features. While, $\bar{m}{g^b_e}$ is used to compute the high-frequency residue with the attentive fea- tures $\bar{m}{g^b_a}$. The aggregation of this high-frequency residual with the squeezed features $\bar{m}{g^b_s}$ helps to learn the feature fu- sion process implicitly and provides the capability to extract high-frequency complementary information from multiple inputs. While illustrated for fusing features of two frames in Fig. 2(d), the proposed BFF can be flexibly adapted to any number of inputs.
## 3.2. Image Reconstruction
Figure 2(e) illustrates the overall image reconstruction stage. To operate on bursts of arbitrary sizes, we pro- pose an adaptive burst feature pooling (ABFP) mechanism that returns the constant burst-size features. As shown in Fig. 2(f), the burst features ($B*f$) are concatenated along channel dimension followed by 1D average pooling oper- ation which adaptively pools out the burst features ($8*f$) as per the requirement. Next, the pooled burst feature maps pass through the no-reference feature enrichment (NRFE) module, shown in Fig. 2(g). The key idea of the proposed NRFE module is to pair immediate neighborhood frames along feature dimension and fuse them using the BFF mod- ule. However, doing this would limit the inter-frame com- munication to successive frames only. Therefore, we pro- pose cyclic burst sampling (CBS) that gathers the neighbor- hood frames in zigzag manner (called here as burst neigh- borhoods) such that reference frame could interact with the last frame as well via intermediate frames; see Fig. 2(h). This cyclic scheme of sampling the burst frames helps in long-range communication without increasing the compu- tational overhead unlike the existing pseudo burst tech- nique [9]. Next, the sampled neighborhood features are combined along burst dimension and processed with BFF to integrate the useful information available in multiple frames of the burst sequence.
To upscale the burst features, we adapt pixel-shuffle [32] such that the information available in burst frames is shuf- fled to increase the spatial resolution. This helps in reducing the compute cost and the overall network parameters.
## 4. Experiments and Analysis
We evaluate the performance of the proposed Burstormer on three different burst image processing tasks: (a) super-resolution (on synthetic and real burst im- ages), (b) low-light image enhancement, and (c) denoising (on grayscale and color data). Additional visual results, ablation experiments, and more details on the network and training settings are provided in the supplementary material.
Implementation Details. We train separate models for dif- ferent tasks in an end-to-end manner without pre-training any module. We pack the input mosaicked raw burst into 4-channel RGGB format. All burst frames are handled with shared Burstormer modules (FA, RBFE, BFF, NRFE) for better parameter efficiency. The following training set- tings are common to all tasks, whereas task-specific ex- perimental details are provided in their corresponding sec- tions. The EDA module of Burstormer is a 3-level encoder- decoder, where each level employs 1 FA (containing single deformable conv. layer) and 1 RBFE module. The BFF unit both in RBFE and NRFE consists of 1 BFA module. Each BFA module consists of 1 multi-dconv head transposed at- tention (MDTA) and 1 gated-Dconv feed-forward network (GDFN) [43]. In the image reconstruction stage, we use 2 NRFE modules. We train models with $L_1$ loss and Adam optimizer with the initial learning rate $1e^{-4}$ that is gradually reduced to $1e^{-6}$ with the cosine annealing scheduler [23] on four RTX6000 GPUs. Random horizontal and vertical image flipping is used for data augmentation.
## 4.1. Burst Super-resolution
We evaluate the proposed Burstormer on synthetic as well as on real-world datasets [2, 3] for the SR scale fac- tor $\times$ 4. For comparisons, we consider several burst SR ap-
| |
230401194/14
|
FigureS4.
Burst low-light image enhancement comparisons on the Sony subset of SID dataset [6].
| |
230401194/2
|
volutions [9]. However, these approaches target only the local features at a single level, while the use of global infor- mation together with multi-scale features has not been ex- tensively explored. Additionally, while aggregating multi- frame features, existing approaches either employ late fu- sion strategy [3, 4] or rigid fusion mechanism (in terms of number of frames) [9]. The former one limits the flexible inter-frame communication, while the later one limits the adaptive multi-frame processing.
In this work, we propose Burstormer for burst image pro- cessing, which incorporates multi-level local-global burst feature alignment and adaptive burst feature aggregation. In contrast to previous works [3, 4] that employ bulky pre- trained networks for explicit feature alignment, we present a novel enhanced deformable alignment (EDA) module that handles misalignment issues implicitly. Overall, the EDA module reduces noise and extracts both local and non-local features with a transformer-based attention and performs multi-scale burst feature alignment and feature enrichment which is not the case with the recent BIPNet [9].
Unlike existing approaches [3, 4, 9] which allow a one go interaction with the reference frame during alignment process, we add a new reference-based feature enrichment (RBFE) mechanism in EDA to allow a more extensive in- teraction with the reference frame. This helps in effec- tively aligning and refining burst features even in complex misalignment cases where the simple alignment approaches would not suffice. In the image reconstruction stage we pro- gressively perform feature consolidation and upsampling, while having access to the multi-frame feature information at all time. This is achieved with our no-reference fea- ture enrichment (NRFE) module. NRFE initially generates burst neighborhoods with the proposed cyclic burst sam- pling (CBS) mechanism that are then aggregated with our burst feature fusion (BFF) unit. Unlike, the existing pseudo bursts [9], the proposed burst neighborhood mechanism is flexible and enables inter-frame communication with signif- icantly less computational cost.
The key highlights of our work are outlined below:
• Our Burstormer is a novel Transformer based de- sign for burst-image restoration and enhancement that leverages multi-scale local and non-local features for improved alignment and feature fusion. Its flexible de- sign allows processing bursts of variable sizes. • We propose an enhanced deformable alignment mod- ule which is based on multi-scale hierarchical design to effectively denoise and align burst features. Apart from aligning burst features it also refines and consol- idates the complimentary burst features with the pro- posed reference-based feature enrichment module. • We propose no-reference feature enrichment module to progressively aggregate and upsample the burst features with less computational overhead. To en- able inter-frame interactions, it generates burst neigh- borhoods through the proposed cyclic burst sampling mechanism followed by the burst feature fusion.
Our Burstormer sets new state-of-the-art on several real and synthetic benchmark datasets for the task of burst super- resolution, burst low-light enhancement, and burst denois- ing. Compared to existing approaches, Burstormer is more accurate, light-weight and faster; see Fig. 1. Further, we provide detailed ablation studies to demonstrate the effec- tiveness of our design choices.
## 2. Related Work
Multi-Frame Super-Resolution. Unlike single image super-resolution, multi-frame super-resolution (MFSR) ap- proaches are required to additionally deal with the sub-pixel misalignments among burst frames caused by camera and object motions. While being computationally efficient, the pioneering MFSR algorithm [37] processes burst frames in frequency domain, often producing images with notice- able artifacts. To obtain better super-resolved results, other methods operate in the spatial domain [10, 17], exploit im- age priors [33], use iterative back-projection [31], or max- imum a posteriori framework [1]. However, all these ap- proaches assume that the image formation model, and mo- tion among input frames can be estimated reliably. Succes- sive works addressed this constraint with the joint estima- tion of the unknown parameters [11, 15]. To deal with noise and complex motion, the MSFR algorithm of [40] employs non-parametric kernel regression and locally adaptive detail enhancement model.
The DBSR algorithm [3] addresses the MFSR problem by applying the explicit feature alignment and attention- centric fusion mechanisms. However, their image warp- ing technique and explicit motion estimation may find dif- ficult in handling scenes with fast moving objects. The EBSR [25] builds on prior PCD alignment techniques [39] by aligning enhanced features specifically for the burst SR task. In addition, the BSRT [24] employs a combination of optical flow and deformable convolution for feature align- ment and utilizes a Swin Transformer [21] for feature ex- traction. More recently, BIPNet [9] was introduced to pro- cess noisy raw bursts using implicit feature alignment and pseudo-burst generation. Building on BIPNet, AFCNet [29] incorporates existing Restormer [43] to improve feature ex- traction for burst SR tasks. Despite having an effective inter-frame communication, their approach is rigid to using certain number of burst frames during alignment and fusion.
Multi-Frame Denoising. Aside from aforementioned MFSR approaches, several multi-frame methods have been developed to perform denoising [7, 14, 26, 27]. The algo- rithm of [36] leverages visually similar image block within and across frames to obtain denoised results. Other works
| |
230401194/11
|
Supplemental Material
Here we provide more details on architectural design, additional ablations, and visual comparisons for burst SR, low-light image enhancement and denoising.
## A. Network Architectural Details
In Burstormer, the EDA module is a 3-level encoder-decoder, where each level employs 1 FA (containing single de- formable conv. layer) and 1 RBFE module. In the image reconstruction stage, we use 2 NRFE modules. The BFF unit both in RBFE and NRFE consists of 1 BFA module.
Figure S1 shows the BFA module that consists of multi-dconv head transposed attention (MDTA) and gated-Dconv feed- forward network (GDFN) [43]. MDTA encodes local and non-local context, and efficient enough to be applied to high- resolution images. Whereas, GDFN performs controlled feature transformation i.e., suppressing less informative features, and allowing only the useful information to pass further through the network.
## B. Ablations on alignment and fusion modules
Table S1 compares the the properties of the proposed EDA and other existing alignment modules. Unlike existing explicit feature alignment approaches DBSR [3] and MFIR [4], the proposed EDA operates at multiple spatial scales and aligns burst features implicitly without any additional supervision. Overall, the proposed EDA module possesses required properties which makes it effective for the burst feature alignment.
Table S2 compares several feature fusion techniques. Our NRFE is flexible to taking as input the features of more than two frames. It extracts local and non-local burst features, enables long-range inter-frame interactions and aggregates the burst neighborhoods to obtain high-quality image.
## C. Additional visual results
Burst Super-resolution. Figure S2, and Figure S3 show qualitative results of competing approaches on examples from the SyntheticBurst and (real) BurstSR datasets [3] for $4\times$ SR. The reproductions of our Burstormer are more detailed, sharper than those produced by the other methods.
Burst low-light image enhancement. Figure S4 depicts that Burstormer produces images that are visually more closer to the ground-truth than the other approaches.
Burst Denoising. Figure S5 shows that the proposed Burstormer is capable of removing noise, while preserving the desired texture and structural content.
FigureS1. Burst Feature Attention (BFA) used in the proposed alignment and reconstruction stages to extract features encoding both local and non-local pixel interactions.
| |
230401194/4
|
Figure 2. Overall pipeline of the proposed Burstormer for burst image processing. Burstormer takes as input a RAW burst of degraded images and outputs a clean high-quality sRGB image. It has two main parts: enhanced deformable alignment (EDA), and the image reconstruction. EDA, labeled as (a), is a multi-scale hierarchical module that, at each level, first extracts noise-free local and non- local features with the burst feature attention (BFA), performs feature alignment (b), and finally refines and consolidates features with an additional interaction with the base frame via (c) the proposed reference-based feature enrichment (RBFE) module. RBFE further employs (d) the burst feature fusion (BFF) unit for merging features by using the back-projection and squeeze-excitation mechanisms. The aligned burst of features are then passed to the image reconstruction stage (e). Here (f) the adaptive burst pooling module transforms the input burst size (B frames) to constant 8 frames through an average pooling operator. Finally, (g) the no-reference feature enrichment (NRFE) module progressively aggregates, and upsamples the burst features to generate the final HR image.
lution [48] to align features of each burst frame to those of the reference frame. Let, $\left\{{{\bar{m}{g}^b}}: b \in [1, \ldots, B]\right\} \in \mathbb{R}^{B \times f \times H \times W}$ denotes the burst features obtained after BFA module, where $B$ denotes number of burst frames, $f$ is the number of feature channels, and $H$
$\times$ $W$ is
the spatial size. We align the features of the current frame $\bar{m}{g}^{b}$ with the ref- erence frame * $\bar{m}{g}^{b_r}$. Feature alignment module processes $\bar{m}{g}^b$ and $\bar{m}{g}^{b_r}$ via an offset convolution layer and outputs the offset $\Delta {n}$ and modulation scalar $\Delta a$ values for $\bar{m}{g}^b$. In Fig. 2 (a), for simplicity, only offset $\Delta {n}$ is shown. The aligned features ${\bar{m}{\bar g}}^b$ are computed as:
$$
{{\bar{m}{\bar g}}^b} = {W_{\text{def}}}\left({{\bar{m}{g}^b},\; \{\Delta {n},\;\Delta a\}} \right),\;
\\
\{\Delta n, \Delta a\} = {W_{\text{off}}}\left({{\bar{m}{g}^b},\;{\bar{m}{g}^{{b_r}}}} \right),
$$
where, $W_{\text{def}}(\cdot)$ and $W_{\text{off}}(\cdot)$ represent the deformable and offset convolutions, respectively. Specifically, every posi- tion $n$ on the aligned feature map ${\bar{m}{\bar g}}^b$ is calculated as:
$$
{{\bar{m}{\bar g}}^b}_n = \sum\limits_{i=1}^K {{W_{n_i}^{d}}\,\,\,{\bar{m}{y}_{\left(n + {n_i} + \Delta {n_i} \right)}^b}} \cdot \Delta {a_{n_i}},
%\vspace{-0.8em}
$$
where, $K$ =9, $\Delta a$ in range $[0, 1]$ for each $n_i \in \left\{{(-1, 1), \allowbreak (-1, 0), ..., (1,1)} \right\}$
* We consider the first burst image to be the reference frame.
| |
230401194/15
|
Noisy Input
FigureS5. Burst denoising results on burst images from the grayscale [30] and color datasets [41].
| |
230401194/13
|
FigureS3. Burst super-resolution ($4 \times$) results on BurstSR (real) dataset [3].
| |
230401194/1
|
# Burstormer: Burst Image Restoration and Enhancement Transformer
Akshay Dudhane $^1$ Syed Waqas Zamir $^2$ Salman Khan $^{1,3}$ Fahad Shahbaz Khan $^{1,4}$ Ming-Hsuan Yang $^{5,6,7}$
$^1$ Mohamed
bin Zayed University of AI $^2$ Inception Institute of AI $^3$ Australian National University $^4$ Link ¨ o ping University $^5$ University of California, Merced $^6$ Yonsei University $^7$ Google Research
On a shutter press, modern handheld cameras capture multiple images in rapid succession and merge them to gen- erate a single image. However, individual frames in a burst are misaligned due to inevitable motions and contain mul- tiple degradations. The challenge is to properly align the successive image shots and merge their complimentary in- formation to achieve high-quality outputs.
Towards this direction, we propose Burstormer: a novel transformer-based architecture for burst image restoration and enhancement. In comparison to existing works, our ap- proach exploits multi-scale local and non-local features to achieve improved alignment and feature fusion. Our key idea is to enable inter-frame communication in the burst neighborhoods for information aggregation and progressive fusion while modeling the burst-wide context. However, the input burst frames need to be properly aligned before fusing their information. Therefore, we propose an enhanced de- formable alignment module for aligning burst features with regards to the reference frame.
Unlike existing methods, the proposed alignment mod- ule not only aligns burst features but also exchanges fea- ture information and maintains focused communication with the reference frame through the proposed reference- based feature enrichment mechanism, which facilitates han- dling complex motions. After multi-level alignment and enrichment, we re-emphasize on inter-frame communica- tion within burst using a cyclic burst sampling module. Finally, the inter-frame information is aggregated using the proposed burst feature fusion module followed by pro- gressive upsampling. Our Burstormer outperforms state- of-the-art methods on burst super-resolution, burst denois- ing and burst low-light enhancement. Our codes and pre- trained models are available at https://github.com/ akshaydudhane16/Burstormer.
## 1. Introduction
In recent years, smartphone industry has witnessed a rampant growth on account of the fueling demand of smart-
Figure 1. Burst super-resolution results (Tab. 1) vs. efficiency (GFlops). Burstormer advances state-of-the-art, while being com- pute efficient and light-weight.
phones in day-to-day life. While the image quality of smart- phone cameras is rapidly improving, there are several barri- ers that hinder in attaining DSLR-like images. For instance, the physical space available in handheld devices restricts manufacturers from employing high-quality bulky camera modules. Most smartphone cameras use small-sized lens, aperture, and sensor, thereby generating images with lim- ited spatial resolution, low dynamic range, and often with noise and color distortions especially in low-light condi- tions. These problems have shifted the focus nowadays in developing computational photography (software) solutions for mitigating the hardware limitations and to approach the image quality of DSLRs.
One emerging approach to achieve high-quality results from a smartphone camera is to take advantage of burst shots consisting of multiple captures of the same scene. The burst image processing approaches aim to recover the high- quality image by merging the complementary information in multiple frames. Recent works [3, 4, 9] have validated the potential of burst processing techniques in reconstruct- ing rich details that cannot be recovered from a single im- age. However, these computationally expensive approaches are often unable to effectively deal with the inherent sub- pixel shifts among multiple frames arising due to camera and/or object movements. This sub-pixel misalignment of- ten causes blurring and ghosting artifacts in the final im- age. To tackle alignment issues, existing methods employ complex explicit feature alignment [3] and deformable con-
| |
230401194/3
|
[7, 27] extend the state-of-the-art single image denoising technique BM3D [7] to videos. The method of [22] yields favorable denoising results by employing a novel homog- raphy flow alignment technique with consistent pixel com- positing operator. In the work of [12], the authors extend single-image denoising network to multi-frame task via re- current deep convolutional neural network. The kernel pre- diction network [30] generates per-pixel kernels for fusing multiple-frames. RViDeNet [42] uses deformable convolu- tions to perform explicit frame alignment in order to pro- vide improved denoising results. The re-parametrization approach of MFIR [4] learns image formation model in deep feature space for the multi-frame denoising. BIP- Net [9] presents a novel pseudo-burst feature fusion ap- proach to perform denoising on burst frames.
Multi-Frame Low-light Image Enhancement. In low- light conditions, smartphone cameras often yield noisy and color-distorted images due to their small aperture and sen- sor pixel cavities. [6] collect a multi-frame dataset for low- light image enhancement, and present a data-driven ap- proach to learn camera imaging pipeline in order to map under-exposed RAW images directly to well-lit sRGB im- ages. The quality of output image is further improved with the perceptual loss presented by [45]. The works of [28] and [47], respectively, use residual learning framework and recurrent convolution network to obtain enhanced images from multiple degraded low-lit input frames. The two- stage approach of [18] employs one subnet for explicitly denoising multiple frames followed by the second subnet to provide us with the enhanced image. Along with super- resolution and denoising, BIPNet [9] is also capable of per- forming multi-frame low-light image enhancement. Unlike the existing multi-frame approaches, our Burstormer aligns burst features at multiple-scales and enables flexible inter- frame communication without much computational over- head. It also incorporates progressive feature merging to obtain high-quality images.
## 3. Proposed Burst Image Processing Pipeline
Burst sequences are usually acquired with handheld devices. The spatial and color misalignments among burst frames are unavoidable due to hand-tremor and cam- era/object motions. These issues negatively affect the over- all performance of the burst processing approaches. In this work, our goal is to effectively align and progres- sively merge the desired information from multiple de- graded frames to reconstruct a high-quality composite im- age. To this end, we propose Burstormer, a novel unified model for multi-frame processing where different modules jointly operate to perform feature denoising, alignment, fu- sion, and upsampling tasks. Here, we describe our method for the task of burst super-resolution, nevertheless, it is ap- plicable to different burst restoration tasks such as burst de- noising and burst enhancement (see experiments Sec. 4).
Overall Pipeline. Fig. 2 shows the overall pipeline of the proposed Burstormer. First, the RAW input burst is passed through the proposed enhanced deformable align- ment (EDA) module which extract noise-free deep features that are aligned and refined with respect to the reference frame features. Second, an image reconstruction module is employed that takes as input the burst of aligned features and progressively merges them using the proposed no ref- erence feature enrichment (NRFE) module. To obtain the super-resolved image, the upsampling operation is immedi- ately applied after each NRFE module in the reconstruction stage. Next, we describe each stage of our approach.
## 3.1. Enhanced Deformable Alignment
In burst processing, effective alignment of mismatched frames is essential as in any error arising in this stage will propagate to later stages, subsequently making the re- construction task difficult. Existing methods perform im- age alignment either explicitly [3, 4], or implicitly [9]. While, these techniques are suitable to correcting mild pixel displacements among frames, they might not adequately handle fast moving objects. In Burstormer, we propose enhanced deformable alignment (EDA) which employs a multi-scale design as shown in Fig. 2(a). Since sub-pixel shifts among frames are naturally reduced at low-spatial res- olution, using the multi-level hierarchical architecture pro- vides us with more robust alignment. Therefore, EDA starts feature alignment from the lowest level (3 $^{rd}$ in this paper) and progressively passes offsets to upper high-resolution levels to help with the alignment process. Furthermore, at each level, the aligned features are passed through the pro- posed reference-based feature enrichment (RBFE) module to fix remaining misalignment issues in burst frames by in- teracting them again with the reference frame. EDA has two key components: (1) Feature alignment, and (2) Reference- based feature enrichment.
Feature alignment. Burst images are often contaminated with random noise that impedes in finding the dense cor- respondences among frames. Therefore, before performing alignment operation, we extract noise-free burst features by using burst feature attention (BFA) module which is built upon the existing transformer block [43]. Unlike in other approaches [3, 4, 9], the BFA module encodes local and non-local context using MDTA block [43] and controls fea- ture transformation through the GDFN [43] block. Further- more, unlike existing attention techniques [21, 34, 38], BFA module is efficient enough to be applied to high-resolution images. The denoised features from BFA are passed fur- ther for alignment. Figure 2(b) shows the feature alignment (FA) module that utilizes a modulated deformable convo-
| |
220201116/20
|
Figure 9: Additional qualitative results of OTS (ours), bicubic upsample, FSSR and DASR on AIM 2019 (800 $\times$ 800 crops).
| |
220201116/18
|
Figure7: Color palettes and their variance for Test HR, LR datasets and solutions of Bicubic Upscale, OTS, FSSR, DASR methods on AIM19.
Hyperparameters. For EDSR, we set the number of resid- ual blocks to 64, the num- ber of features to 128, and the residual scaling to 1. For UNet, we set the base factor to 64. The train- ing details are given in Ta- ble 4. We provide a com- parison of the hyperparame- ters of FSSR, DASR and OTS (ours) in Table 5. In con- trast to FSSR and DASR, our method does not contain a degradation part. This helps to notably reduce the amount of tunable hyperparameters.
Optimizer. We employ Adam (Kingma&Ba, 2014).
Computational complexity. Training OTS with EDSR as the transport map and the per- ceptual transport cost on AIM 2019 dataset takes $\approx 4$ days on a single Tesla V100 GPU.
## I MAGE S UPER -R ESOLUTION OF F ACES
We conduct an experiment using CelebA (Liuetal.,2015) faces to test the applicability of OT for unpaired SR. We test our Algorithm 1 with MSE as the cost and UNet or EDSR as the transport map.
| |
220201116/17
|
see (Santambrogio,2015, 7.2). Our particular interest is to compute the optimal potential $(f^{*})^{c}$ at $\mathbb{Q}'=\mathbb{Q}$. We recall and use $\mathbb{W}_{p}^{p}(\mathbb{Q},\mathbb{Q})=0$ to derive
$$
\mathbb{W}_{p}^{p}(\mathbb{Q},\mathbb{Q})=0=
\sup_{f}\bigg[\int_{\mathcal{X}} f^{c}(y')d\mathbb{Q}'(y')+\int_{\mathcal{Y}} f(y)d\mathbb{Q}(y)\bigg].
$$
One may see that $f^{*}\equiv 0$ attains the supremum (its $c$ -transform $(f^{*})^{c}$ is also zero). Thus, if $(f^{*})^{c}\equiv 0$ is a unique potential (up to a constant), the first variation of $\mathbb{Q}'\mapsto \mathbb{W}_{p}^{p}(\mathbb{Q}',\mathbb{Q})$ at $\mathbb{Q}'\!=\!\mathbb{Q}$ vanishes.
Case 3 (Maximum Mean Discrepancy). Let $k:\mathcal{Y}\times\mathcal{Y}\rightarrow\mathbb{R}$ be a positive definite symmetric kernel. The (square) of the Maximum Mean Discrepancy between $\mathbb{Q}',\mathbb{Q}$ is given by
$$
\begin{aligned}
\text{MMD}_{k}^{2}(\mathbb{Q}',\mathbb{Q})\stackrel{\textrm{def}}{=}\int_{\mathcal{Y}\times \mathcal{Y}}k(y_{1},y_{2})d\big[(\mathbb{Q}-\mathbb{Q}')\times (\mathbb{Q}-\mathbb{Q}')\big](y_{1},y_{2})
\\
=\int_{\mathcal{Y}\times \mathcal{Y}}k(y_{1},y_{2})d(\mathbb{Q}'\times\mathbb{Q}')(y_{1},y_{2})-2\int_{\mathcal{Y}\times \mathcal{Y}}k(y_{1},y_{2})d(\mathbb{Q}'\times\mathbb{Q})(y_{1},y_{2})+\text{Const}(\mathbb{Q})
\end{aligned}
$$
see (Sejdinovic et al., 2013, Equation 3.3). The first variation of the quadratic in $\mathbb{Q}'$ term is given by $y\mapsto 2\cdot\int_{\mathcal{Y}}k(y, y_{2})d\mathbb{Q}'(y_{2})$, see (Santambrogio,2015, (cid:77) 7.2). The second term is linear in $\mathbb{Q}'$ and its first variation is simply $y\mapsto (-2)\cdot\int_{\mathcal{Y}}k(y, y_{2})d\mathbb{Q}(y_2)$. When $\mathbb{Q}'=\mathbb{Q}$, the sum of these terms is zero. That is, the first variation of the functional $\mathbb{Q}'\mapsto \text{MMD}_{k}^{2}(\mathbb{Q}',\mathbb{Q})$ vanishes at $\mathbb{Q}'=\mathbb{Q}$.
## A SSESSING THE BIAS OF METHODS ON AIM19 DATASET
We additionally demonstrate the bias issue by comparing color palettes of HR images and super- resolution results of different methods, see Figure 7. We construct palettes by choosing random image pixels from dataset images and representing them as an RGB point cloud in $[0,1]^{3}\subset\mathbb{R}^{3}$. Figure 7 shows that OTS (d) captures large contrast of HR (a) images (variance of its palette), while FSSR (e), DASR (f), Bicubic Upscale (c) palettes are less contrastive and closer to LR (b). We construct palettes 100 times to evaluate their average contrast (variance). The metric quantitatively confirms that our OTS method better captures the contrast of HR dataset, while GAN-based methods (FSSR and DASR) are notably biased towards LR dataset statistics (low contrast).
## T RAINING D ETAILS
Perceptual cost. In 7.2 we test following per- ceptual cost as $b$:
$$
\begin{aligned}b(x^{\text{up}}\hspace{-1mm},y)\hspace{-0.7mm}=\hspace{-0.7mm}\text{MSE}(x^{\text{up}}\hspace{-1mm},y)\hspace{-0.7mm}+\hspace{-0.7mm}\nicefrac{1}{3}\cdot\text{MAE}(x^{\text{up}}\hspace{-1mm},y)\hspace{-0.7mm}+\hspace{-0.7mm}
\nicefrac{1}{50} \hspace{-0.7mm}\cdot\hspace{-0.7mm} \hspace{-6mm}\sum_{k\in\{3,8,15,22\}}\hspace{-6mm}\text{MSE}\big(f_{k}(x^{\text{up}}),f_{k}(y)\big),
\end{aligned}
$$
=-1 where $f_{k}$ denotes the features of the $k$ th layer of a pre-trained VGG-16 network (Simonyan& Zisserman,2014), MAE is the mean absolute error $\text{MAE}(x,y)=\frac{\|x-y\|_{1}}{\dim (\mathcal{Y})}$.
Dynamic transport cost. In the preliminary experiments, we used bicubic upsampling as the “ $\text{Up}$" operation. Later, we found that the method works better if we gradually change the upsampling. We start from the bicubic upsampling. Every $k_{c}$ iterations of $f_{\omega}$ (see Table 4), we change the cost to $c(x,y)=b\big(T_{\theta}'(x), y\big)$, where $T_{\theta}'$ is a fixed frozen copy of the currently learned SR map $T_{\theta}$.
| |
220201116/5
|
In the following Theorem, we prove that, in general, $\mathbb{Q}^{\lambda}\neq\mathbb{Q}$ for the minimizer $\mathbb{Q}^{\lambda}$ of (6).
Theorem1 (ThedistributionsolvingtheregularizedGANproblemisalwaysbiased). Under the assumptions of Lemma 2, assume that the first variation (Santambrogio,2015, Definition7.12) of the functional ${\mathbb{Q}'\mapsto\mathcal{D}(\mathbb{Q}',\mathbb{Q})}$ at the point $\mathbb{Q}'=\mathbb{Q}$ exists and is equal to zero. This means that $\mathcal{D}(\mathbb{Q}+\varepsilon\Delta\mathbb{Q})=\mathcal{D}(\mathbb{Q}, \mathbb{Q})+ o(\varepsilon)$ for every signed measure $\Delta\mathbb{Q}$ of zero total mass and $\varepsilon\geqslant 0$ such that $\mathbb{Q}+\varepsilon\Delta\mathbb{Q}\in\mathcal{P}(\mathcal{Y})$. Then, if $\mathbb{P}\neq\mathbb{Q}$, then $\mathbb{Q}'=\mathbb{Q}$ does not deliver the minimum to $\mathcal{F}$.
Before proving Theorem 1, we highlight that the assumption about the vanishing first variation of $\mathbb{Q}'\mapsto\mathcal{D}(\mathbb{Q}',\mathbb{Q})$ at $\mathbb{Q}'=\mathbb{Q}$ is reasonable. In Appendix A, we prove that this assumption holds for the popular GAN discrepancies $\mathcal{D}(\mathbb{Q}',\mathbb{Q})$, e.g., $f$ -divergences (Nowozinetal.,2016), Wasserstein distances (Arjovskyetal.,2017), and Maximum Mean Discrepancies (Lietal.,2017).
Proof. Let $\Delta \mathbb{Q}=\mathbb{P}-\mathbb{Q}$ denote the difference measure of $\mathbb{P}$ and $\mathbb{Q}$. It has zero total mass and $\forall\varepsilon\in[0,1]$ it holds that $\mathbb{Q}+\varepsilon\Delta\mathbb{Q}=\varepsilon\mathbb{P}+(1-\varepsilon)\mathbb{Q}$ is a mixture distribution of probability distributions $\mathbb{P}$ and $\mathbb{Q}$. As a result, for all $\varepsilon\in[0,1]$, we have
$$
\begin{aligned}
\mathcal{F}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}) = \mathcal{D}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}, \mathbb{Q}) + \lambda\cdot\text{Cost}(\mathbb{P}, \mathbb{Q}+\varepsilon\Delta \mathbb{Q})
\\
=\mathcal{D}(\mathbb{Q}, \mathbb{Q}) + o(\varepsilon) +
\lambda\cdot\text{Cost}(\mathbb{P}, \varepsilon\mathbb{P}+(1-\varepsilon)\mathbb{Q})
\\
\leqslant o(\varepsilon) + \lambda\cdot\varepsilon \cdot\text{Cost}(\mathbb{P}, \mathbb{P}) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q}) =
o(\varepsilon) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})
%
\\
= \underbrace{\lambda\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})}_{=\mathcal{F}(\mathbb{Q})} - \lambda \cdot\varepsilon\cdot \underbrace{\text{Cost}(\mathbb{P}, \mathbb{Q})}_{>0} + o(\varepsilon),
\end{aligned}
$$
$$
\begin{aligned}
\mathcal{F}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}) = \mathcal{D}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}, \mathbb{Q}) + \lambda\cdot\text{Cost}(\mathbb{P}, \mathbb{Q}+\varepsilon\Delta \mathbb{Q})
\\
=\mathcal{D}(\mathbb{Q}, \mathbb{Q}) + o(\varepsilon) +
\lambda\cdot\text{Cost}(\mathbb{P}, \varepsilon\mathbb{P}+(1-\varepsilon)\mathbb{Q})
\\
\leqslant o(\varepsilon) + \lambda\cdot\varepsilon \cdot\text{Cost}(\mathbb{P}, \mathbb{P}) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q}) =
o(\varepsilon) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})
%
\\
= \underbrace{\lambda\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})}_{=\mathcal{F}(\mathbb{Q})} - \lambda \cdot\varepsilon\cdot \underbrace{\text{Cost}(\mathbb{P}, \mathbb{Q})}_{>0} + o(\varepsilon),
\end{aligned}
$$
$$
\begin{aligned}
\mathcal{F}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}) = \mathcal{D}(\mathbb{Q}+\varepsilon\Delta \mathbb{Q}, \mathbb{Q}) + \lambda\cdot\text{Cost}(\mathbb{P}, \mathbb{Q}+\varepsilon\Delta \mathbb{Q})
\\
=\mathcal{D}(\mathbb{Q}, \mathbb{Q}) + o(\varepsilon) +
\lambda\cdot\text{Cost}(\mathbb{P}, \varepsilon\mathbb{P}+(1-\varepsilon)\mathbb{Q})
\\
\leqslant o(\varepsilon) + \lambda\cdot\varepsilon \cdot\text{Cost}(\mathbb{P}, \mathbb{P}) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q}) =
o(\varepsilon) + \lambda\cdot(1-\varepsilon)\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})
%
\\
= \underbrace{\lambda\cdot \text{Cost}(\mathbb{P}, \mathbb{Q})}_{=\mathcal{F}(\mathbb{Q})} - \lambda \cdot\varepsilon\cdot \underbrace{\text{Cost}(\mathbb{P}, \mathbb{Q})}_{>0} + o(\varepsilon),
\end{aligned}
$$
where in transition from (9) to (10), we use $\mathcal{D}(\mathbb{Q},\mathbb{Q})=0$ and exploit the convexity of the OT cost (Villani,2003,Theorem4.8). In, we use $\text{Cost}(\mathbb{P},\mathbb{P})=0$. We see that $\mathcal{F}(\mathbb{Q}\!+\!\varepsilon\Delta \mathbb{Q})$ is smaller then $\mathcal{F}(\mathbb{Q})$ for sufficiently small $\varepsilon>0$, i.e., $\mathbb{Q}'\!=\!\mathbb{Q}$ does not minimize $\mathcal{F}$.
Additionally, we provide a toy example that further illustrates the issue with the bias.
Example Consider $\mathcal{X}=\mathcal{Y}=\mathbb{R}^{1}$. Let $\mathbb{P}=\frac{1}{2}\delta_{0}+ \frac{1}{2}\delta_{2}$, $\mathbb{Q}=\frac{1}{2}\delta_{1}+ \frac{1}{2}\delta_{3}$ be distributions concentrated at $\{0,2\}$ and $\{1,3\}$, respectively. Put $c(x,y)=|x-y|$ to be the content loss. Also, let $\mathcal{D}$ to be the OT cost for $|x-y|^2$. Then for $\lambda=0$ there exist two maps between $\mathbb{P}$ and $\mathbb{Q}$ that deliver the same minimal value for, namely $T(0)=1,T(2)=3$ and $T(0)=3,T(2)=1$. For $\lambda>0$, the optimal solution of the problem (5) is unique, biased and given by $T(0)=1-\frac{\lambda}{2},T(2)=3-\frac{\lambda}{2}$.
Proof. Let $T(0)=t_{0}$ and $T(2)=t_{2}$. Then $T_{\#}\mathbb{P}=\frac{1}{2}\delta_{t_{0}}+\frac{1}{2}\delta_{t_{2}}$, and now becomes
$$
\min_{t_{0},t_{2}}\bigg[\min\big\{\frac{1}{2}(t_{0}-1)^{2}+\frac{1}{2}(t_{2}-3)^{2}; \frac{1}{2}(t_{0}-3)^{2}+\frac{1}{2}(t_{2}-1)^{2}\big\}+\lambda \big\{\frac{1}{2}|0-t_0|+\frac{1}{2}|2-t_{2}|\big\}\bigg],
$$
where the second term is $\mathcal{R}_{c}(T)$ and the first term is the OT cost $\mathcal{D}(T_{\#}\mathbb{P}, \mathbb{Q})$ expressed as the minimum over the transport costs of two possible transport maps $t_0\mapsto 1; t_{2}\mapsto 3$ and $t_0\mapsto 3; t_{2}\mapsto 1$. The minimizer can be derived analytically and equals $t_{0}=1-\frac{\lambda}{2},t_{2}=3-\frac{\lambda}{2}$.
In Example 1, $T_{\#}^{\lambda}\mathbb{P}=\mathbb{Q}^{\lambda}$ never matches $\mathbb{Q}$ exactly for $\lambda>0$. In (cid:77) 7.1, we conduct an evaluation of maps obtained via minimizing objective (5) on the synthetic benchmark by (Korotinetal.,2021). We empirically demonstrate that the bias exists and it is indeed a notable practical issue.
Remarks. Throughout this section, we enforce additional assumptions on (5), e.g., we restrict our analysis to content losses $c(\cdot,\cdot)$, which are powers of Euclidean norms $\|\cdot\|^{p}$. This is needed to make the derivations concise and to be able to exploit the available results in OT. We think that the provided results hold under more general assumptions and leave this question open for future studies.
## U NBIASED O PTIMAL T RANSPORT S OLVER
In 6.1, we derive our algorithm to compute OT maps. Importantly, in (cid:77) 6.2, we detail its differences and similarities with regularized GANs, which we discussed in (cid:77) 5.
| |
220201116/2
|
with marginals $\mathbb{P}$ and $\mathbb{Q}$. For a measurable map $T:\mathcal{X}\rightarrow\mathcal{Y}$, we denote the associated push-forward operator by $T_{\#}$. The expression $\|\cdot\|$ denotes the usual Euclidean norm if not stated otherwise. We denote the space of $\mathbb{Q}$ -integrable functions on $\mathcal{Y}$ by $L^{1}(\mathbb{Q})$.
## U NPAIRED I MAGE S UPER -R ESOLUTION T ASK
In this section, we formalize the unpaired image super-resolution task that we consider (Figure 2).
Figure2: The task of super-resolution we consider.
Let $\mathbb{P}$ and $\mathbb{Q}$ be two distributions of LR and HR images, respectively, on spaces $\mathcal{X}$ and $\mathcal{Y}$, respectively. We assume that $\mathbb{P}$ is obtained from $\mathbb{Q}$ via some unknown degradation. The learner has access to unpaired random samples from $\mathbb{P}$ and $\mathbb{Q}$. The task is to fit a map ${T:\mathcal{X}\rightarrow\mathcal{Y}}$ satisfying ${T_{\#}\mathbb{P}=\mathbb{Q}}$ which inverts the degradation.
We highlight that the image SR task is theoretically ill-posed for two reasons.
- 1. Non-existence. The degradation filter may be non-injective and, consequently, non-invertible. This is a theoretical obstacle to learn one-to-one SR maps $T$. 2. Ambiguity. There might exist multiple maps satisfying ${T_{\#}\mathbb{P}=\mathbb{Q}}$ but only one inverting the degradation. With no prior knowledge about the correspondence between $\mathbb{P}$ and $\mathbb{Q}$, it is unclear how to pick this particular map.
The first issue is usually not taken into account in practice. Most existing paired and unpaired SR methods learn one-to-one SR maps $T$, see (Ledig et al., 2017; Lai etal.,2017;Weietal.,2021).
The second issue is typically softened by regularizing the model with the content loss. In the real-world, it is reasonable to assume that HR and the corresponding LR images are close. Thus, the fitted SR map $T$ is expected to only slightly change the input image. Formally, one may require the learned map $T$ to have the small value of
$$
\mathcal{R}_{c}(T)\stackrel{def}{=}\int_{\mathcal{Y}}c\big(x,T(x)\big)d\mathbb{P}(x),
$$
where $c:\mathcal{X}\times\mathcal{Y}\rightarrow\mathbb{R}_{+}$ is a function estimating how different the inputs are. The most popular example is the $\ell^{1}$ identity loss, i.e, formulation for $\mathcal{X}=\mathcal{Y}=\mathbb{R}^{D}$ and $c(x,y)=\|x-y\|_{1}$.
More broadly, losses $\mathcal{R}_{c}(T)$ are typically called content losses and incorporated into training objec- tives of methods for SR (Lugmayretal.,2019a, (cid:77) 3.4), (Kimetal.,2020, (cid:77) 3) and other unpaired tasks beside SR (Taigmanetal.,2016, (cid:77) 4), (Zhuetal.,2017, (cid:77) 5.2) as regularizers. They stimulate the learned map $T$ to minimally change the image content.
## B ACKGROUND ON O PTIMAL T RANSPORT
In this section, we give the key concepts of the OT theory (Villani,2008) that we use in our paper.
Primal form. For two distributions ${\mathbb{P}\in\mathcal{P}(\mathcal{X})}$ and ${\mathbb{Q}\in\mathcal{P}(\mathcal{Y})}$ and a transport cost $c:\mathcal{X}\times\mathcal{Y}\rightarrow\mathbb{R}$, Monge’s primal for- mulation of the optimal transport cost is as follows:
$$
\text{Cost}(\mathbb{P},\mathbb{Q})\stackrel{\text{def}}{=}\inf_{T_{\#}\mathbb{P}=\mathbb{Q}}\ \int_{\mathcal{X}} c\big(x,T(x)\big)d\mathbb{P}(x),
$$
| |
220201116/16
|
## F IRST V ARIATIONS OF GAN D ISCREPANCIES V ANISH AT THE O PTIMUM
We demonstrate that the first variation of $\mathbb{Q}'\mapsto \mathcal{D}(\mathbb{Q}',\mathbb{Q})$ is equal to zero at $\mathbb{Q}'=\mathbb{Q}$ for common GAN discrepancies $\mathcal{D}$. This suggests that the corresponding assumption of our Theorem 1 is relevant.
To begin with, for a functional $\mathcal{G}:\mathcal{P}(\mathcal{Y})\rightarrow \mathbb{R}\cup \{\infty\}$, we recall the definition of its first variation. A measurable function $\delta\mathcal{G}[\mathbb{Q}]:\mathcal{Y}\rightarrow\mathbb{R}\cup \{\infty\}$ is called the first variation of $\mathcal{G}$ at a point $\mathbb{Q}\in\mathcal{P}(\mathcal{Y})$, if, for every measure $\Delta\mathbb{Q}$ on $\mathcal{Y}$ with zero total mass ($\int_{\mathcal{Y}}1\hspace{0.5mm} d\Delta\mathbb{Q}(y)=0$),
$$
\mathcal{G}(\mathbb{Q}+\varepsilon \Delta\mathbb{Q})=\mathcal{G}(\mathbb{Q})+\varepsilon\int_{\mathcal{Y}}\delta\mathcal{G}[\mathbb{Q}](y)\hspace{0.5mm}d\Delta\mathbb{Q}(y)+ o(\varepsilon)
$$
for all $\varepsilon\geqslant 0$ such that $\mathbb{Q}+\varepsilon\Delta\mathbb{Q}$ is a probability distribution. Here for the sake of simplicity we suppressed several minor technical aspects, see (Santambrogio,2015,Definition7.12) for details. Note that the first variation is defined up to an additive constant.
Now we recall the definitions of three most popular GAN discrepancies and demonstrate that their first variation is zero at an optimal point. We consider $f$ -divergences (Nowozinetal.,2016), Wasserstein distances (Arjovskyetal.,2017), and Maximum Mean Discrepancies (Lietal.,2017).
Case 1 ($f$ -divergence). Let $f:\mathbb{R}_{+}\rightarrow \mathbb{R}$ be a convex and differentiable function satisfying $f(1)=0$. The $f$ -divergence between $\mathbb{Q}',\mathbb{Q}\in\mathcal{P}(\mathcal{Y})$ is defined by
$$
\mathcal{D}_{f}(\mathbb{Q}',\mathbb{Q})\stackrel{\normalfont{def}}{=}\int_{\mathcal{Y}}f\bigg(\frac{d\mathbb{Q}'(y)}{d\mathbb{Q}(y)}\bigg)d\mathbb{Q}(y).
$$
The divergence takes finite value only if $\mathbb{Q'}\ll \mathbb{Q}$, i.e., $\mathbb{Q'}$ is absolutely continuous w.r.t. $\mathbb{Q}$. Vanilla GAN loss (Goodfellowetal.,2014) is a case of $f$ -divergence (Nowozinetal.,2016, Table1).
We define $\mathcal{G}(\mathbb{Q}')\stackrel{\normalfont def}{=}\mathcal{D}_{f}(\mathbb{Q}',\mathbb{Q})$. For $\mathbb{Q}'=\mathbb{Q}$ and some $\Delta\mathbb{Q}$ such that $\mathbb{Q}+\varepsilon\Delta\mathbb{Q}\in\mathcal{P}(\mathcal{Y})$ we derive
$$
\begin{aligned}
\mathcal{G}(\mathbb{Q}+\varepsilon\Delta\mathbb{Q})=\int_{\mathcal{Y}}f\bigg(\frac{d\mathbb{Q}(y)}{d\mathbb{Q}(y)}+\varepsilon\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}\bigg)d\mathbb{Q}(y)=
\int_{\mathcal{Y}}f\bigg(1+\varepsilon\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}\bigg)d\mathbb{Q}(y)
\\
=\int_{\mathcal{Y}}f(1)d\mathbb{Q}(y)+\int_{\mathcal{Y}}f'(1)\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}d\mathbb{Q}(y)+o(\varepsilon)=
\mathcal{G}(\mathbb{Q})+\int_{\mathcal{Y}}f'(1)d\Delta\mathbb{Q}(y)+o(\varepsilon),
\end{aligned}
$$
$$
\begin{aligned}
\mathcal{G}(\mathbb{Q}+\varepsilon\Delta\mathbb{Q})=\int_{\mathcal{Y}}f\bigg(\frac{d\mathbb{Q}(y)}{d\mathbb{Q}(y)}+\varepsilon\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}\bigg)d\mathbb{Q}(y)=
\int_{\mathcal{Y}}f\bigg(1+\varepsilon\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}\bigg)d\mathbb{Q}(y)
\\
=\int_{\mathcal{Y}}f(1)d\mathbb{Q}(y)+\int_{\mathcal{Y}}f'(1)\frac{d\Delta\mathbb{Q}(y)}{d\mathbb{Q}(y)}d\mathbb{Q}(y)+o(\varepsilon)=
\mathcal{G}(\mathbb{Q})+\int_{\mathcal{Y}}f'(1)d\Delta\mathbb{Q}(y)+o(\varepsilon),
\end{aligned}
$$
where in transition from (18) to (19), we consider the Taylor series w.r.t. $\varepsilon$ at $\varepsilon=0$. We see that $\delta\mathcal{G}[\mathbb{Q}](y)\equiv f'(1)$ is constant, i.e., the first variation of $\mathbb{Q}'\mapsto\mathcal{D}_{f}(\mathbb{Q}',\mathbb{Q})$ vanishes at $\mathbb{Q}'=\mathbb{Q}$.
Case 2 (Wasserstein distance). If in OT for- mulation (3) the cost function $c(x,y)$ equals $\|x-y\|^p$ with $p\geqslant1$, then $\big[\text{Cost}(\mathbb{P}, \mathbb{Q})\big]^{1/p}$ is called the Wasserstein distance ($\mathbb{W}_p$). Gener- ative models which use $\mathbb{W}_{p}^{p}$ as the discrepancy are typically called the Wasserstein GANs (WGANs). The most popular case is $p=1$ (Arjovskyetal.,2017;Gulrajanietal.,2017), but more general cases appear in related work as well, see (Liuetal.,2019;Mallastoetal., 2019).
The first variation of $\mathcal{G}(\mathbb{Q}')\stackrel{\normalfont \textrm{def}}{=}\mathbb{W}_{p}^{p}(\mathbb{Q}',\mathbb{Q})$ at a point $\mathbb{Q}'$ is given by $\mathcal{G}[\mathbb{Q}'](y)=(f^{*})^{c}(y)$, where $f^{*}$ is the optimal dual potential (provided it is unique up to a constant) in (4) for a pair $(\mathbb{Q}',\mathbb{Q})$,
| |
220201116/6
|
\begin{algorithm}[t!]
\SetInd{0.5em}{0.3em}
{
\SetAlgorithmName{Algorithm}{empty}{Empty}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{distributions
$\mathbb{P},\mathbb{Q}$ accessible by samples; mapping network $T_{\theta}:\mathcal{X}\rightarrow\mathcal{Y}$;\\ potential $f_{\omega}:\mathcal{X}\rightarrow\mathbb{R}$;
transport cost $c:\mathcal{X}\times\mathcal{Y}\rightarrow\mathbb{R}$; number $K_{T}$ of inner iters;\\
}
\Output{approximate OT map $(T_{\theta})_{\#}\mathbb{P}\approx \mathbb{Q}$\;}
}
\Repeat{not converged}{
Sample batches $X\sim\mathbb{P}$, $Y\!\sim\! \mathbb{Q}$\;
$\mathcal{L}_{f}\leftarrow \frac{1}{|X|}\sum\limits_{x\in X}f_{\omega}\big(T_{\theta}(x)\big)- \frac{1}{|Y|}\sum\limits_{y\in Y}f_{\omega}(y)$\;
Update $\omega$ by using $\frac{\partial \mathcal{L}_{v}}{\partial \omega}$\;
\For{$k_{T} = 1,2, \dots, K_{T}$}{
Sample batch $X\sim \mathbb{P}$\;
${\mathcal{L}_{T}=\frac{1}{|X|}\sum\limits_{x\in X}\big[c
\big(x, T_{\theta}(x)\big)-f_{\omega}\big(T_{\theta}(x)\big)\big]}$\;
Update $\theta$ by using $\frac{\partial \mathcal{L}_{T}}{\partial \theta}$\;
}
}
\caption{OT solver to compute the OT map between $\mathbb{P}$ and $\mathbb{Q}$ for transport cost $c(x,y)$.}
\end{algorithm}
OT solver to compute the OT map between $\mathbb{P}$ and $\mathbb{Q}$ for transport cost $c(x,y)$.
\begin{algorithm}[t!]
\SetInd{0.5em}{0.3em}
{
\SetAlgorithmName{Algorithm}{empty}{Empty}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{distributions
$\mathbb{P},\mathbb{Q}$ accessible by samples; mapping network $T_{\theta}:\mathcal{X}\rightarrow\mathcal{Y}$;\\ potential $f_{\omega}:\mathcal{X}\rightarrow\mathbb{R}$;
transport cost $c:\mathcal{X}\times\mathcal{Y}\rightarrow\mathbb{R}$; number $K_{T}$ of inner iters;\\
}
\Output{approximate OT map $(T_{\theta})_{\#}\mathbb{P}\approx \mathbb{Q}$\;}
}
\Repeat{not converged}{
Sample batches $X\sim\mathbb{P}$, $Y\!\sim\! \mathbb{Q}$\;
$\mathcal{L}_{f}\leftarrow \frac{1}{|X|}\sum\limits_{x\in X}f_{\omega}\big(T_{\theta}(x)\big)- \frac{1}{|Y|}\sum\limits_{y\in Y}f_{\omega}(y)$\;
Update $\omega$ by using $\frac{\partial \mathcal{L}_{v}}{\partial \omega}$\;
\For{$k_{T} = 1,2, \dots, K_{T}$}{
Sample batch $X\sim \mathbb{P}$\;
${\mathcal{L}_{T}=\frac{1}{|X|}\sum\limits_{x\in X}\big[c
\big(x, T_{\theta}(x)\big)-f_{\omega}\big(T_{\theta}(x)\big)\big]}$\;
Update $\theta$ by using $\frac{\partial \mathcal{L}_{T}}{\partial \theta}$\;
}
}
\caption{OT solver to compute the OT map between $\mathbb{P}$ and $\mathbb{Q}$ for transport cost $c(x,y)$.}
\end{algorithm}: distributions P, Q accessiblebysamples;mappingnetwork T θ: X →Y; potential f ω: X → R;transportcost c: X ×Y → R;number K T ofinneriters; Output: approximateOTmap (T θ) # P ≈ Q; repeat Samplebatches X ∼ P, Y ∼ Q; L f ← 1 | X | (cid:80) x ∈ X f ω (cid:0) T θ (x) (cid:1) − 1 | Y | (cid:80) y ∈ Y f ω (y); Update ω byusing ∂ L v ∂ω; for k T =1, 2,...,K T do Samplebatch X ∼ P; L T = 1 | X | (cid:80) x ∈ X (cid:2) c (cid:0) x,T θ (x) (cid:1) − f ω (cid:0) T θ (x) (cid:1)(cid:3); Update θ byusing ∂ L T ∂θ; until notconverged;
## M INIMAX O PTIMIZATION A LGORITHM
We derive a minimax optimization problem to recover the optimal transport map from $\mathbb{P}$ to $\mathbb{Q}$. We expand the dual form (4). To do this, we first note that
$$
\begin{aligned}
\int_{\mathcal{X}}f^{c}(x) d\mathbb{P}(x)=\int_{\mathcal{X}}\inf_{y\in\mathcal{Y}}\left\lbrace c(x,y)-f(y)\right\rbrace d\mathbb{P}(x)=
\inf_{T:\mathcal{X}\rightarrow\mathcal{Y}}\int_{\mathcal{X}}\left\lbrace c\big(x,T(x)\big)-f\big(T(x)\big)\right\rbrace d\mathbb{P}(x).
\end{aligned}
$$
Here we replace the optimization over points $y\in\mathcal{Y}$ with an equivalent optimization over the functions ${T:\mathcal{X}\rightarrow\mathcal{Y}}$. This is possible due to the Rockafellar interchange theorem (Rockafellar, 1976,Theorem3A). Substituting (11) to (4), we have
$$
\begin{aligned}
\text{Cost}(\mathbb{P},\mathbb{Q})=\sup_{f}\inf_{T:\mathcal{X}\rightarrow\mathcal{Y}}\bigg[\int_{\mathcal{Y}} f(y)d\mathbb{Q}(y)+
\int_{\mathcal{X}}\left\lbrace c\big(x,T(x)\big)-f\big(T(x)\big)\right\rbrace d\mathbb{P}(x)\bigg]
\vspace{-1.7mm}
\end{aligned}
$$
We denote the expression under the $\sup\inf$ by $\mathcal{L}(f,T)$. Now we show that by solving the saddle point problem (12) one can obtain the OT map $T^{*}$.
Lemma3 (OTmapssolvethesaddlepointproblem). Assume that the OT map $T^{*}$ between $\mathbb{P},\mathbb{Q}$ for cost $c(x,y)$ exists. Then, for every optimal potential $f^{*}\in\mathop{\mathrm{argsup}}_{f}\big[\inf_{T:\mathcal{X}\rightarrow\mathcal{Y}}\mathcal{L}(f,T)\big]$ of (12),
$$
T^{*}\in\mathop{\mathrm{arginf}}_{T:\mathcal{X}\rightarrow\mathcal{Y}}\int_{\mathcal{X}}\left\lbrace c\big(x,T(x)\big)-f\big(T(x)\big)\right\rbrace d\mathbb{P}(x).
$$
Proof. Since $f^{*}$ is optimal, we have $\inf_{T:\mathcal{X}\rightarrow\mathcal{Y}}\mathcal{L}(f^{*},T)=\text{Cost}(\mathbb{P},\mathbb{Q}).$ We use $T^{*}_{\#}\mathbb{P}=\mathbb{Q}$ and the change of variables $y=T^{*}(x)$ to derive $\int_{\mathcal{X}}f^{*}\big(T^{*}(x)\big)d\mathbb{P}=\int_{\mathcal{Y}}f^{*}(y)d\mathbb{Q}$. Substituting this equality into, we obtain $\mathcal{L}(f^{*},T^{*})=\int_{\mathcal{X}}c\big(x,T^{*}(x)\big)d\mathbb{P}(x)=\text{Cost}(\mathbb{P},\mathbb{Q}),$ i.e., holds.
Lemma 3 states that one can solve a saddle point problem (12), obtain an optimal pair $(f^{*},T^{*})$, and use $T^{*}$ as an OT map from $\mathbb{P},\mathbb{Q}$. For general $\mathbb{P},\mathbb{Q}$, the $\mathop{\mathrm{arginf}}_{T}$ set for an optimal $f^{*}$ might contain not only OT map $T^{*}$ but other functions as well. However, our experiments ((cid:77) 7) show that this is not a serious issue in practice. To solve the optimization problem (12), we approximate the potential $f$ and map $T$ with neural networks $f_{\omega}$ and $T_{\theta}$, respectively. We train the networks with stochastic gradient ascent-descent by using random batches from $\mathbb{P},\mathbb{Q}$.
The practical optimization procedure is detailed in Algorithm 1. We call this procedure an Optimal Transport Solver (OTS).
## R EGULARIZED GAN S VS. O PTIMAL T RANSPORT S OLVER
In this subsection, we discuss similarities and differences between our optimization objective (12) and the objective of regularized GANs (5). We establish an intriguing connection between GANs that use integral probability metrics (IPMs) as $\mathcal{D}$. A discrepancy $\mathcal{D}\!:\!\mathcal{P}(\mathcal{Y})\!\times\! \mathcal{P}(\mathcal{Y})\!\rightarrow\!\mathbb{R}_{+}$ is an IPM if
| |
220201116/4
|
Xieetal.,2019;Bousquetetal.,2017;Balajietal.,2020) optimize complex GAN objectives such as (5) and provide biased solutions ((cid:77) 5, (cid:77) 7.1). For a comprehensive overview of dual-form methods, we refer to (Korotinetal.,2021). The authors conduct an evaluation of OT methods for the quadratic cost $c(x,y)=\|x-y\|^{2}$. According to them, the best performing method is $\lfloor \text{MM:R}\rceil$. It is based on the variational reformulation of (4), which is a particular case of our formulation (12). Extensions of $\lfloor\text{MM:R}\rceil$ appear in.
## B IASED O PTIMAL T RANSPORT IN GAN S
Figure4: Illustration of Lemma 1. The solution $T^{\lambda}$ of (5) is an OT map from $\mathbb{P}$ to $T^{\lambda}_{\#}\mathbb{P}$. In general, $T^{\lambda}_{\#}\mathbb{P}\neq \mathbb{Q}$ (Thm.).
In this section, we establish connections between GAN methods regularized by content losses (1) and OT. A common approach to solve the unpaired SR via GANs is to define a loss function $\mathcal{D}:\mathcal{P}(\mathcal{Y})\times\mathcal{P}(\mathcal{Y})\rightarrow\mathbb{R}_{+}$ and train a generative neural network $T$ via minimizing
$$
\inf_{T:\mathcal{X}\mapsto\mathcal{Y}}\big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda \mathcal{R}_{c}(T)\big].
$$
The term $\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})$ ensures that the generated distribution $T_{\#}\mathbb{P}$ of SR images is close to the true HR distribution $\mathbb{Q}$; the second term $\mathcal{R}_{c}(T)$ is the content loss (1). For convenience, we assume that $\mathcal{D}(\mathbb{Q}, \mathbb{Q})=0$ for all $\mathbb{Q}\in\mathcal{P(Y)}$. Two most popular examples of $\mathcal{D}$ are the Jensen – Shannon divergence (Goodfellowetal.,2014), i.e., the vanilla GAN loss, and the Wasserstein-1 loss (Arjovsky &Bottou,2017). In unpaired SR methods, the optimization objectives are typically more complicated than (5). In addition to the content or identity loss (1), several other regularizations are usually introduced, see
(cid:77)
4.
For a theoretical analysis, we stick to the basic formulation regularized with generic content loss (5). It represents the simplest and straightforward SR setup. We prove the following lemma, which connects the solution $T^{\lambda}$ of (5) and optimal maps for transport cost $c(x,y)$.
Lemma1 (ThesolutionoftheregularizedGANisanOTmap). Assume that $\lambda>0$ and the minimizer $T^{\lambda}$ of (5) exists. Then $T^{\lambda}$ is an OT map between $\mathbb{P}$ and $\mathbb{Q}^{\lambda}\stackrel{\text{\normalfont def}}{=}T^{\lambda}_{\#}\mathbb{P}$ for cost $c(x,y)$, i.e., it minimizes
\begin{equation*} \inf_{T_{\#}\mathbb{P}=\mathbb{Q}^{\lambda}}\mathcal{R}_{c}(T)=\inf_{T_{\#}\mathbb{P}=\mathbb{Q}^{\lambda}}\int_{\mathcal{X}}c\big(x,T(x)\big)d\mathbb{P}(x). \end{equation*}
Proof. Assume that $T^{\lambda}$ is not an optimal map between $\mathbb{P}$ and $T^{\lambda}_{\#}\mathbb{P}$. Then there exists a more optimal $T^{\dagger}$ satisfying ${T^{\dagger}_{\#}\mathbb{P}=T_{\#}^{\lambda}\mathbb{P}}$ and $\mathcal{R}_{c}(T^{\dagger})<\mathcal{R}_{c}(T^{\lambda})$. We substitute this $T^{\dagger}$ to and derive
$$
\begin{aligned}
\mathcal{D}(T^{\dagger}_{\#}\mathbb{P},\mathbb{Q})+\lambda \mathcal{R}_{c}(T^{\dagger})=\mathcal{D}(T^{\lambda}_{\#}\mathbb{P},\mathbb{Q})+
\lambda \mathcal{R}_{c}(T^{\dagger})<
\mathcal{D}(T^{\lambda}_{\#}\mathbb{P},\mathbb{Q})+\lambda \mathcal{R}_{c}(T^{\lambda}),
\end{aligned}
$$
which is a contradiction, since $T^{\lambda}$ is a minimizer of (5), but $T^{\dagger}$ provides the smaller value.
Our Lemma 1 states that the minimizer $T^{\lambda}$ of a regularized GAN problem is always an OT map between $\mathbb{P}$ and the distribution $\mathbb{Q}^{\lambda}$ generated by the same $T^{\lambda}$ from $\mathbb{P}$. However, below we prove that $\mathbb{Q}^{\lambda}\neq\mathbb{Q}$, i.e., $T^{\lambda}$ does not actually produce the distribution of HR images (Figure 4). To begin with, we state and prove the following auxiliary result.
Lemma 2 (Reformulation of the regularized GAN via distributions). Under the assumptions of Lemma 1, let $\mathcal{X}=\mathcal{Y}$ be a compact subset of $\mathbb{R}^D$ with negligible boundary. Let $\mathbb{P}\!\in\! \mathcal{P(X)}$ be absolutely continuous, $\mathbb{Q}\!\in\!\mathcal{P(Y)}$ and $c(x,y)\!=\!\|x-y\|^{p}$ with $p>1$. Then is equivalent to
$$
\inf_{\mathbb{Q}'\in\mathcal{P}(\mathcal{Y})}\mathcal{F}(\mathbb{Q}')\stackrel{\text{\normalfont{def}}}{=}\!\inf_{\mathbb{Q}'\in\mathcal{P}(\mathcal{Y})}\big[\mathcal{D}(\mathbb{Q}', \mathbb{Q})+\lambda\cdot\text{\normalfont{Cost}}(\mathbb{P}, \mathbb{Q}')\big],
$$
and the solutions of (5) and (6) are related as $\mathbb{Q}^{\lambda}=T^{\lambda}_{\#}\mathbb{P}$, where $\mathbb{Q}^{\lambda}$ is the minimizer of (6).
Proof. We derive
$$
\begin{aligned}
\inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda \mathcal{R}_{c}(T) \big]
= \inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda \int_{\mathcal{X}}c\big(x,T(x)\big)d\mathbb{P}(x)\big]= \\
\inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda\cdot \text{Cost} (\mathbb{P}, T_{\#}\mathbb{P})\big]
= \inf_{\mathbb{Q}'\in\mathcal{P}(\mathcal{Y})}\big[\mathcal{D}(\mathbb{Q}', \mathbb{Q})+\lambda\cdot \text{Cost}(\mathbb{P}, \mathbb{Q}')\big].
\vspace{-2mm}
\end{aligned}
$$
$$
\begin{aligned}
\inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda \mathcal{R}_{c}(T) \big]
= \inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda \int_{\mathcal{X}}c\big(x,T(x)\big)d\mathbb{P}(x)\big]= \\
\inf_{T:\mathcal{X}\mapsto\mathcal{Y}} \big[\mathcal{D}(T_{\#}\mathbb{P},\mathbb{Q})+\lambda\cdot \text{Cost} (\mathbb{P}, T_{\#}\mathbb{P})\big]
= \inf_{\mathbb{Q}'\in\mathcal{P}(\mathcal{Y})}\big[\mathcal{D}(\mathbb{Q}', \mathbb{Q})+\lambda\cdot \text{Cost}(\mathbb{P}, \mathbb{Q}')\big].
\vspace{-2mm}
\end{aligned}
$$
In transition from (7) to (8), we use the definition of OT cost (2) and our Lemma 1, which states that the minimizer $T^{\lambda}$ of (5) is an OT map, i.e., $\int_{\mathcal{X}}c\big(x,T^{\lambda}(x)\big)d\mathbb{P}(x)=\text{Cost}(\mathbb{P},T^{\lambda}_{\#}\mathbb{P})$. The equality in (8) follows from the fact that $\mathbb{P}$ is abs. cont. and $c(x,y)=\|x-y\|^{p}$: for all $\mathbb{Q}'\in\mathcal{P}(\mathcal{Y})$ there exists a (unique) solution $T$ to the Monge OT problem (2) for $\mathbb{P}, \mathbb{Q}'$ (Santambrogio,2015,Thm. 1.17).
| |
220201116/10
|
(Weietal.,2021) methods. FSSR method is the winner of AIM 2019 Challenge; DASR is a current state-of-the- art method for unpaired image SR. Both methods utilize the idea of frequency separation and solve the problem in two steps. First, they train a network to generate LR images. Next, they train a super-resolution network using generated pseudo-pairs. Differently to FSSR, DASR also employs real-world LR images for training SR network taking into consideration the domain gap between generated and real-world LR images. Both methods utilize several losses, e.g., adversarial and perceptual, either on the entire image or on its high/low frequency components. For testing FSSR and DASR, we use their official code and pretrained models.
Implementation details. We train the networks using 128 $\times$ 128 HR and 32 $\times$ 32 LR random patches of images augmented via random flips and rotations. We conduct separate experiments using EDSR as the transport map and either MSE or perceptual cost.
Metrics. We calculate PSNR, SSIM, LPIPS, FID. FID is computed on $32\hspace{-0.7mm}\times\hspace{-0.7mm}32$ patches of LR test images upsampled by the method in view w.r.t. random patches of test HR. We use 50k patches to compute FID. The other metrics are computed on the entire upsampled LR test and HR test images.
\begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular} \begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular} \begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular} \begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular} \begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular} \begin{tabular}{c| c c c c}
\rightarrowprule
\textbf{Method} & \textbf{FID} $\downarrow$ & \textbf{PSNR} $\uparrow$ & \textbf{SSIM} $\uparrow$ & \textbf{LPIPS} $\downarrow$\midrule
\makecell{Bicubic upsample} & 178.59 & {\color{LimeGreen}22.39} & {\color{LimeGreen}0.613} & 0.688\midrule
\makecell{OTS (ours, MSE)} & 139.17 & 19.73 & 0.533 & 0.456\midrule
\makecell{OTS (ours, VGG)} & {\color{blue}89.04} & \underline{20.96} & {\color{blue}0.605} & {\color{blue}0.380}\midrule
FSSR & {\color{LimeGreen}53.92} & 20.83 & 0.514 & \underline{0.390}\midrule
DASR & \underline{124.09} & {\color{blue}21.79} & 0.577 & {\color{LimeGreen}0.346}\bottomrule
\end{tabular}
Table3: Comparison of OTS (ours) with FSSR, DASR on AIM19 dataset. The 1st, 2nd, 3rd best results are highlighted in green, blue and underlined, respectively.
=-1 Experimental results are given in Table 3, Figure 6. The results show that the usage perceptual cost function in OTS boosts performance. According to FID, OTS with perceptual cost function beats DASR. On the other hand, it outperforms FSSR in PSNR, SSIM and, importantly, LPIPS. Note that bicubic upsample outperforms all the methods, according only to PSNR and SSIM, which have issues stated in (cid:77) 7.1. According to visual analysis, OTS deals better with noise artifacts. Additional results are given in Appendix E. We also demonstrate the bias issue of FSSR and DASR in Appendix B.
## D ISCUSSION
Significance. Our analysis connects content losses in GANs with OT and reveals the bias is- sue. Content losses are used in a wide range of tasks besides SR, e.g., in the style transfer and domain adaptation tasks. Our results demon- strate that GAN-based methods in all these tasks may a priori lead to biased solutions. In certain cases it is undesirable, e.g., in medical applications (Bissotoetal.,2021). Failing to learn true data statistics (and learning biased ones instead), e.g., in the super-resolution of MRI images, might lead to a wrong diagnosis made by a doctor due to SR algorithm drawing inexistent details on the scan. Thus, we think it is essential to emphasize and alleviate the bias issue, and provide a way to circumvent this difficulty.
=-1 Potential Impact. We expect our OT approach to improve the existing applications of image super-resolution. Importantly, it has less hyperparameters, uses smaller number of neural networks than many existing methods (see Table 5 in Appendix C for comparison), and is end-to-end — this should simplify its usage in practice. Besides, our method is generic and presumably can be applied to other unpaired learning tasks as well. Studying such applications is a promising avenue for the future work.
Limitations. Our method fits a one-to-one optimal mapping (transport map) for super- resolution which, in general, might not exist. Besides, not all optimal solutions of our op- timization objective are guaranteed to be OT maps. These limitations suggest the need for
| |
220201116/21
|
Figure 10: Additional qualitative results of OTS (ours), bicubic upsample, FSSR and DASR on AIM 2019. The sizes of crops on the 1st and 2nd images are 350 $\times$ 350 and 800 $\times$ 800, respectively.
| |
220201116/1
|
# A N O PTIMAL T RANSPORT P ERSPECTIVE ON U NPAIRED I MAGE S UPER -R ESOLUTION
M. Gazdieva ∗ L. Rout † A. Korotin $^*$ A. Kravchenko ‡ A. Filippov § E. Burnaev $^*$
=-1 Real-world image super-resolution (SR) tasks often do not have paired datasets, which limits the application of supervised techniques. As a result, the tasks are usually approached by unpaired techniques based on Generative Adversarial Net- works (GANs), which yield complex training losses with several regularization terms, e.g., content or identity losses. We theoretically investigate optimization problems which arise in such models and find two surprizing observations. First, the learned SR map is always an optimal transport (OT) map. Second, we theo- retically prove and empirically show that the learned map is biased, i.e., it does not actually transform the distribution of low-resolution images to high-resolution ones. Inspired by these findings, we propose an algorithm for unpaired SR which learns an unbiased OT map for the perceptual transport cost. Unlike the existing GAN-based alternatives, our algorithm has a simple optimization objective reduc- ing the need for complex hyperparameter selection and an application of additional regularizations. At the same time, it provides a nearly state-of-the-art performance on the large-scale unpaired AIM19 dataset.
## I NTRODUCTION
Figure1: Super-resolution of a squirrel using Bicubic upsample, OTS (ours) and DASR (Weietal.,2021) methods (4 $\times$ 4 upsample, 370 $\times$ 800 crops).
The problem of image super-resolution (SR) is to reconstruct a high-resolution (HR) image from its low-resolution (LR) counterpart. In many modern deep learning approaches, SR networks are trained in a supervised manner by using synthetic datasets containing LR-HR pairs (Limetal.,2017, 4.1); (Zhangetal.,2018b, (cid:77) 4.1). For example, it is common to create LR images from HR with a simple downscaling, e.g., bicubic (Ledigetal.,2017, (cid:77) 3.2). However, such an artificial setup barely represents the practical setting, in which the degradation is more sophisticated and unknown (Maeda, 2020). This obstacle suggests the necessity of developing methods capable of learning SR maps from unpaired data without considering prescribed degradations.
Contributions. We study the unpaired image SR task and its solutions based on Generative Adversar- ial Networks (Goodfellowetal.,2014,GANs) and analyse them from the Optimal Transport (OT, see (Villani,2008)) perspective.
- =-1 We investigate the GAN optimization objectives regularized with content losses, which are common in unpaired image SR methods ((cid:77) 5, (cid:77) 4). We prove that the solution to such objectives is always an optimal transport map. We theoretically and empirically show that such maps are biased ((cid:77) 7.1), i.e., they do not transform the LR image distribution to the true HR image distribution. We provide an algorithm to fit an unbiased OT map for perceptual transport cost ((cid:77) 6.1) and apply it to the unpaired image SR problem ((cid:77) 7.2). We establish connections between our algorithm and regularized GANs using integral probability metrics (IPMs) as a loss (
(cid:77)
- 6.2).
Our algorithm solves a minimax optimization objective and does not require extensive hyperparameter search, which makes it different from the existing methods for unpaired image SR. At the same time, the algorithm provides a nearly state-of-art performance in the unpaired image SR problem ((cid:77) 7.2).
Notation. We use $\mathcal{X},\mathcal{Y}$ to denote Polish spaces and $\mathcal{P}(\mathcal{X}), \mathcal{P}(\mathcal{Y})$ to denote the respective sets of probability distributions on them. We denote by $\Pi(\mathbb{P},\mathbb{Q})$ the set of probability distributions on $\mathcal{X} \times \mathcal{Y}$
[email protected]
∗ Skolkovo Institute of Science and Technology, Moscow, Russia. Corr. to: † Indian Space Research Organisation, Ahmedabad, India ‡ University of Oxford, Oxford, UK § Huawei Noah’s Ark Lab, Moscow, Russia
| |
220201116/3
|
where the minimum is taken over the measurable functions (transport maps) $T:\mathcal{X}\rightarrow\mathcal{Y}$ that map $\mathbb{P}$ to $\mathbb{Q}$, see Figure 3a. The optimal $T^{*}$ is called the optimal transport map.
Note that (2) is not symmetric, and this formulation does not allow mass splitting, i.e., for some $\mathbb{P},\mathbb{Q}$ there may be no map $T$ that satisfies $T_{\#}\mathbb{P}=\mathbb{Q}$. Thus, (Kantorovitch,1958) proposed the relaxation:
$$
\text{Cost}(\mathbb{P},\mathbb{Q})\stackrel{\text{def}}{=}\inf_{\pi\in\Pi(\mathbb{P},\mathbb{Q})}\int_{\mathcal{X}\times \mathcal{Y}}c(x,y)d\pi(x,y),
$$
where the minimum is taken over the transport plans $\pi$, i.e., the measures on $\mathcal{X}\times\mathcal{Y}$ whose marginals are $\mathbb{P}$ and $\mathbb{Q}$ (Figure 3b). The optimal $\pi^{*}\in\Pi(\mathbb{P},\mathbb{Q})$ is called the optimal transport plan.
With mild assumptions on the transport cost $c(x,y)$ and distributions $\mathbb{P}$, $\mathbb{Q}$, the minimizer $\pi^{*}$ of (3) always exists (Villani, 2008, Theorem 4.1) but might not be unique. If $\pi^{*}$ is of the form $[\text{id} , T^{*}]_{\#} \mathbb{P}\in\Pi(\mathbb{P},\mathbb{Q})$ for some $T^*$, then $T^{*}$ is an optimal transport map that minimizes (2).
(a) Monge’s formulation of OT.
(b) Kantorovich’s formulation of OT.
Figure3: Monge’s and Kantorovich’s formulations of Optimal Transport.
Dual form. The dual form (Villani,2003) of OT cost (3) is as follows:
$$
\begin{aligned}\text{Cost}(\mathbb{P}, \mathbb{Q})=
\sup_{f}\bigg[\int_{\mathcal{X}} f^{c}(x)d\mathbb{P}(x)+\int_{\mathcal{Y}} f(y)d\mathbb{Q}(y)\bigg];
\end{aligned}
$$
here $\sup$ is taken over all $f\!\in\! \mathcal{L}^{1}(\mathbb{Q})$, and ${f^{c}(x)\!=\!\inf\limits_{y\in\mathcal{Y}}\!\big[c(x,y)\!-\!f(y)\!\big]}$ is the $c$ - transform of $f$.
## R ELATED W ORK
U NPAIRED I MAGE S UPER -R ESOLUTION. =-1 Existing approaches to unpaired image SR mainly solve the problem in two steps. One group of approaches learn the degra- dation operation at the first step and then train a super-resolution model in a super- vised manner using generated pseudo-pairs, see (Bulatetal.,2018;Fritscheetal.,2019). Another group of approaches (Yuanetal.,2018;Maeda,2020) firstly learn a mapping from real-world LR images to “ clean ” LR images, i.e., HR images, downscaled using predetermined (e.g., bicubic) operation, and then a mapping from “ clean" LR to HR images. Most methods are based on CycleGAN (Zhuetal.,2017), initially designed for the domain transfer task, and utilize cycle-consistency loss. Methods are also usually endowed with several other losses, e.g. content (Kim et al., 2020, (cid:77) 3), identity (Wangetal.,2021,
(cid:77)
3.2) or perceptual (Jietal.,2020, (cid:77) 3.4).
O PTIMAL T RANSPORT IN G ENERATIVE M ODELS. The majority of existing OT- based generative models employ OT cost as the loss function to update the gener- ative network, e.g., see (Arjovsky et al., 2017). These methods are out of scope of the present paper, since they do not com- pute OT maps. Existing methods to com- pute the OT map approach the primal (2), (3) or dual form (4). Primal-form methods (Luetal.,2020;
| |
220201116/9
|
Figure5: Comparison of OTS (ours), regularized IPM GAN on the Wasserstein-2 benchmark. The 1st line shows blurry faces $x\sim\mathbb{P}$, the 2nd line, clean faces $y=T^{*}(x)$, where $T^{*}$ is the OT map from $\mathbb{P}$ to $\mathbb{Q}$. Next lines show maps from $\mathbb{P}$ to $\mathbb{Q}$ fitted by the methods.
Figure6: Qualitative results of OTS (ours), bicubic upsample, FSSR and DASR on AIM 2019 dataset (350 $\times$ 350 crops).
and WGAN-GP solely struggles to fit a good restora- tion map. Even for the best performing $\lambda=10^{2}$ all metrics are notably worse than for OTS. Impor- tantly, OTS decreases the burden of parameter searching as there is no parameter $\lambda$.
## L ARGE -S CALE E VALUATION
For evaluating our method at a large-scale, we employ the dataset by (Lugmayretal.,2019b) of AIM 2019 Real-World Super-Resolution Challenge (Track 2). The train part contains 800 HR images with up to 2040 pixels width or height and 2650 unpaired LR images of the same shape. They are constructed using artificial, but realistic, image degradations. We quantitatively evaluate our method on the validation part of AIM dataset that contains 100 pairs of LR-HR images.
Baselines. We compare OTS on AIM dataset with the bicubic upsample, FSSR (Fritscheetal.,2019) and DASR
| |
220201116/8
|
\begin{tabular}{@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}|@{\hskip1.9pt}c@{\hskip1.9pt}}\rightarrowprule
\multirow{2}{*}{\makecell{\textit{Metrics}/ \textit{Method}}} & \multicolumn{8}{c|}{\textbf{Regularized} \textbf{IPM GAN} (WGAN-GP, $\lambda_{\text{GP}}=10$)} & \multirow{2}{*}{\makecell{\textbf{OTS}(ours)}} \cline{2-9}
& $\lambda=0$ & $\lambda=10^{-1}$ & $\lambda=10^{0}$ & $\lambda=10^{1}$ & $\lambda=10^{2}$ & $\lambda=10^{3}$ & $\lambda=10^{4}$ & $\lambda=10^{5}$ & \midrule
$\mathcal{L}^{2}\text{-UVP}\downarrow$ & $25.2\%$ & $16.7\%$ & $17.7\%$ & $12.0\%$ & $\textbf{4.0}\%$ & $14.0\%$ & $28.5\%$ & $30.5\%$ & $\textbf{1.4}\%$\textit{FID}$\downarrow$ & $57.24$ & $46.23$ & $40.04$ & $42.89$ & $\textbf{24.25}$ & $187.95$ & $332.7$ & $334.7$ & $\mathbf{15.65}$ \textit{PSNR}$\uparrow$ & $17.90$ & $19.76$ & $19.34$ & $20.81$ & $\textbf{25.58}$ & $19.91$ & $16.90$ & $16.52$ & $\mathbf{30.02}$ \textit{SSIM}$\uparrow$ & $0.565$ & $0.655$ & $0.656$ & $0.689$ & $\textbf{0.859}$ & $0.702$ & $0.520$ & $0.498$ & $\mathbf{0.933}$\textit{LPIPS}$\downarrow$ & $0.135$ & $0.093$ & $0.099$ & $0.081$ & $\textbf{0.031}$ & $0.172$ & $0.429$ & $0.446$ & $\mathbf{0.013}$ \bottomrule
\end{tabular}
Table2: Quantitative evaluation of restoration maps fitted by the regularized IPM GAN, OTS (ours) using the Wasserstein-2 images benchmark (Korotinetal.,2021).
the potential $f_{\omega}$. In (cid:77) 7.1, where input and output images have the same size, we use UNet 1 (Ronneberger et al., 2015) as a transport map $T_{\theta}$. In (cid:77) 7.2, the LR input images are $4\times 4$ times smaller than HR, so we use EDSR network (Limetal.,2017).
=-1 Transport costs. In 7.1, we use the mean squared error (MSE), i.e., ${c(x,y)=\frac{\|x-y\|^{2}}{\dim (\mathcal{Y})}}$. It is equivalent to the quadratic cost but is more convenient due to the normalization. In 7.2, we consider $c(x,y)=b(\text{Up}(x),y)$, where $b$ is a cost between the bicubically upsampled LR image $x^{\text{up}}= \text{Up}(x)$ and HR image $y$. We test $b$ defined as $\text{MSE}$ and the perceptual cost using features of a pre-trained VGG-16 network (Simonyan&Zisserman,2014), see Appendix C for details.
## A SSESSING THE B IAS IN R EGULARIZED GAN S
In this section, we empirically confirm the insight of 5 that the solution $T^{\lambda}$ of (5) may not satisfy $T^{\lambda}_{\#}\mathbb{P}=\mathbb{Q}$. Note if $T^{\lambda}_{\#}\mathbb{P}=\mathbb{Q}$, then by our Lemma 1, we conclude that $T^{\lambda}\equiv T^{*}$, where $T^{*}$ is an OT map from $\mathbb{P}$ to $\mathbb{Q}$ for $c(x,y)$. Thus, to access the bias, it is reasonable to compare the learned map $T^{\lambda}$ with the ground truth OT map $T^{*}$ for $\mathbb{P}$, $\mathbb{Q}$.
=-1 For evaluation, we use the Wasserstein-2 benchmark (Korotinetal.,2021). It provides high- dimensional continuous pairs $\mathbb{P}$, $\mathbb{Q}$ with an analytically known OT map $T^{*}$ for the quadratic cost ${c(x,y)=\|x-y\|^{2}}$. We use their “ Early" images benchmark pair. It simulates the image deblurring setup, i.e., $\mathcal{X}=\mathcal{Y}$ is the space of $64\times 64$ RGB images, $\mathbb{P}$ is blurry faces, $\mathbb{Q}$ is clean faces satisfying $\mathbb{Q}=T^{*}_{\#}\mathbb{P}$, where $T^{*}$ is an analytically known OT map, see the 1st and 2nd lines in Figure 5.
=-1 To quantify the learned maps from $\mathbb{P}$ to $\mathbb{Q}$, we use PSNR, SSIM, LPIPS (Zhangetal.,2018a), FID (Heuseletal.,2017) metrics. Similar to (Weietal.,2021), we use the AlexNet-based (Krizhevsky etal.,2012) LPIPS. FID and LPIPS are practically the most important since they better correlate with the human perception of the image quality. We include PSNR, SSIM as popular evaluation metrics, but they are known to badly measure perceptual quality (Zhangetal.,2018a;Nilsson&Akenine- Möller,2020). Due to this, higher PSNR, SSIM values do not necessarily mean better performance. We calculate metrics using scikit-image for SSIM and open source implementations for PSNR 2, LPIPS 3 and FID 4. In this section, we additionally use the $\mathcal{L}^{2}\text{-UVP}$ (Korotinetal.,2021,
(cid:77)
4.2) metric.
On the benchmark, we compare OTS (12) and IPM GAN (5). We use MSE as the content loss $c(x,y)$. In IPM GAN, we use the Wasserstein-1 ($\mathbb{W}_{1}$) loss with the gradient penalty ${\lambda_{\text{GP}}=10}$ etal.,2017) as $\mathcal{D}$. We do $10$ discriminator updates per $1$ generator update and train the model for 15K generator updates. For fair comparison, the rest hyperparameters match those of our algorithm. We train the regularized WGAN-GP with various coefficients of content loss $\lambda\in \{0, 10^{-1}, \dots, 10^{5}\}$ and show the learned maps $T^{\lambda}$ and the map $\hat{T}$ obtained by OTS in Figure 5.
Results. The performance of the regularized IPM GAN significantly depends on the choice of the content loss value $\lambda$. For high values $\lambda\geqslant 10^{3}$, the learned map is close to the identity as expected. For small values $\lambda\leqslant 10^{1}$, the regularization has little effect,
| |
220809342/20
|
\begin{align*} \ensuremath{\mathbf{y}}_l^* \left(P\left(\sum_{i\in A_j} a_{l,\nu_j(i)} \ensuremath{\mathbf{x}}_{\nu_j(i)}\right)\right) &=\sum_{i\in A_j} \left\lvert \ensuremath{\mathbf{z}}_l^*(\ensuremath{\mathbf{x}}_{\nu_j(i)}) \, \ensuremath{\mathbf{x}}_{\nu_j(i)}^*(\ensuremath{\mathbf{y}}_l)\right\rvert\\
&\geqslant \left\lvert \ensuremath{\mathbf{z}}_l^*(\ensuremath{\mathbf{x}}_{\nu_j(l)}) \, \ensuremath{\mathbf{x}}_{\nu_j(l)}^*(\ensuremath{\mathbf{y}}_l)\right\rvert\\
&\geqslant \delta. \end{align*}
Consequently, if $C_u$ is as in (2.3) and $C_r$ is the truncation quasi-greedy constant of $\ensuremath{\mathcal{Y}}$ (see (2.2)), we have
$$
\left\lVert \ensuremath{\mathbbm{1}}_{A_j}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{Y}}] \right\rVert \leqslant \frac{C_u C_r}{\delta} \left\lVert P\left(\sum_{i\in A_j} \varepsilon_{l,\nu_j(i)} \ensuremath{\mathbf{x}}_{\nu_j(i)}\right)\right\rVert, \quad j=1,\dots, K.
$$
If we put $a=\sup_{n\in\ensuremath{\mathcal{N}}} \left\lVert \ensuremath{\mathbf{x}}_n^*\right\rVert$, $b=\sup_{i\in\ensuremath{\mathcal{M}}} \left\lVert \ensuremath{\mathbf{y}}_i\right\rVert$, and let $\kappa$ be the optimal constant such that
$$
\left\lVert \sum_{j=1}^K y_j\right\rVert \leqslant \kappa \sum_{j=1}^K \left\lVert y_j\right\rVert, \quad y_j\in\ensuremath{\mathbb{Y}},
$$
and $C_s$ be as in (2.4), we obtain
\begin{align*} \left\lVert \ensuremath{\mathbbm{1}}_{A}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{Y}}] \right\rVert &=\left\lVert \sum_{j=1}^{K}\ensuremath{\mathbbm{1}}_{A_j}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{Y}}]\right\rVert\\
&\leqslant\kappa\sum_{j=1}^{K}\left\lVert \ensuremath{\mathbbm{1}}_{A_j}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{Y}}]\right\rVert\\
&\leqslant \kappa K\frac{C_uC_r}{\delta}\left\lVert P\right\rVert abC_s\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m).\qedhere \end{align*}
Theorem 3.12. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a basis $\ensuremath{\mathcal{X}}$. As- sume that:
- (a) Either $\ensuremath{\mathcal{X}}$ is unconditional and strongly absolute, or (b) there is a non-decreasing doubling sequence $\ensuremath{\mathbf{s}}=(s_m)_{m=1}^\infty$ such that $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}] \gtrsim \ensuremath{\mathbf{s}}$ and
$$
\sum_{m=1}^\infty \frac{1}{s_m}<\infty.
$$
Suppose that $\ensuremath{\mathcal{Y}}$ is a complemented truncation quasi-greedy basic se- quence of $\ensuremath{\mathbb{X}}$. Then
$$
\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\lesssim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. If (a) holds, we let $\ensuremath{\mathbb{B}}$ be the sequence space induced by the semi- normalized basis $\ensuremath{\mathcal{X}}$. If (b) holds, we take $\ensuremath{\mathbb{B}}=d_{1,\infty}(\ensuremath{\mathbf{w}})$, where $\ensuremath{\mathbf{w}}$ is the weight with primitive weight $\ensuremath{\mathbf{s}}$, and we appeal to Theorem 3.3 to claim
| |
220809342/18
|
\begin{align*} \left\lVert \ensuremath{\mathbbm{1}}_A[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\right\rVert &\approx \Ave_{\varepsilon_i=\pm1} \left\lVert \sum_{i\in A}\varepsilon_i\, \ensuremath{\mathbf{y}}_i\right\rVert\\
&\gtrsim \Ave_{\varepsilon_i=\pm1}\left\lVert \sum_{i\in A}\varepsilon_i\, \ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \left(\sum_{i\in A}\left\lvert \ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)\right\rvert^2\right)^{1/2}\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \left(\sum_{i\in A}\left\lvert f_i\right\rvert^2\right)^{1/2}\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \max_{i\in A}\left\lvert f_i\right\rvert\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \ensuremath{\mathbbm{1}}_{\Omega_A}[\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}}, \ensuremath{\mathbb{B}}]\right\rVert. \end{align*}
Now the statement follows since $\left\lvert \Omega_A\right\rvert\geqslant \lceil m/K \rceil$.
Theorem 3.8. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a strongly absolute semi-normalized unconditional basis $\ensuremath{\mathcal{X}}$ which induces an $L$ -convex lat- tice structure on $\ensuremath{\mathbb{X}}$. Suppose that $\ensuremath{\mathcal{Y}}$ is a UCC basic sequence with the HBEP in $\ensuremath{\mathbb{X}}$. Then there is a constant $c>0$ such that
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}](m) \gtrsim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](\lceil cm\rceil), \quad m\in\ensuremath{\mathbb{N}}.
$$
Proof. Just apply Lemma 3.7 with $\ensuremath{\mathbb{B}}$ the sequence space induced by the semi-normalized basis $\ensuremath{\mathcal{X}}$.
Theorem 3.9. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a truncation quasi- greedy basis $\ensuremath{\mathcal{X}}$. Let $\ensuremath{\mathbf{s}}=(s_m)_{m=1}^\infty$ be a non-decreasing doubling sequence of positive scalars such that $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}] \gtrsim \ensuremath{\mathbf{s}}$ and
$$
\sum_{m=1}^\infty \frac{1}{s_m}<\infty.
$$
Suppose that $\ensuremath{\mathcal{Y}}$ is a UCC basic sequence with the HBEP in $\ensuremath{\mathbb{X}}$. Then
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}](m) \gtrsim s_m,\quad m\in\ensuremath{\mathbb{N}}.
$$
Proof. By Theorem 3.3, Theorem 2.6, and Lemma 2.7 we can apply Lemma with $\ensuremath{\mathbb{B}}=d_{1,\infty}(\ensuremath{\mathbf{w}})$, where $\ensuremath{\mathbf{w}}$ is the weight with primitive weight $\ensuremath{\mathbf{s}}$.
Corollary 3.10. Let $\ensuremath{\mathcal{X}}$ be a squeeze-symmetric basis of a quasi-Banach space $\ensuremath{\mathbb{X}}$. Suppose that
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
| |
220809342/17
|
3.1.
## Lower estimates for democracy functions
. A subtle, yet im- portant, obstruction to apply Lemma 3.5 to basic squences $\ensuremath{\mathcal{Y}}$ in a quasi-Banach space $\ensuremath{\mathbb{X}}$ is that the coordinate functionals associated to $\ensuremath{\mathcal{Y}}$ are not defined on $\ensuremath{\mathbb{X}}$ but on the closed subspace of $\ensuremath{\mathbb{X}}$ generated by $\ensuremath{\mathcal{Y}}$, denoted by $\ensuremath{\mathbb{Y}}$. If $\ensuremath{\mathbb{X}}$ is locally convex, the Hahn-Banach theorem comes to our aid: any bounded linear functional on $\ensuremath{\mathbb{Y}}$ extends to a bounded linear functional on $\ensuremath{\mathbb{X}}$ without increasing its norm. However, there are important spaces, such as Hardy spaces, that are not locally convex and so this extension cannot be taken for granted. This situation mo- tivated Day [14] to define the Hahn-Banach Extension Property (HBEP for short). We say that $\ensuremath{\mathbb{Y}}$ has the HBEP in $\ensuremath{\mathbb{X}}$ if there is a constant $C$ such that for every $f^*\in\ensuremath{\mathbb{Y}}^*$ there is $g^*\in\ensuremath{\mathbb{X}}^*$ such that $g^*|_\ensuremath{\mathbb{Y}}=f^*$ and $\left\lVert g^*\right\rVert \leqslant C\left\lVert f^*\right\rVert$. Needless to say, $\ensuremath{\mathbb{X}}$ has the HBEP in $\ensuremath{\mathbb{X}}$. For the pur- poses of this paper, it suffices to keep in mind that any complemented subspace of a quasi-Banach space $\ensuremath{\mathbb{X}}$ has the HBEP in $\ensuremath{\mathbb{X}}$. We say that the basic sequence $\ensuremath{\mathcal{Y}}$ has the HBEP in $\ensuremath{\mathbb{X}}$ if $\ensuremath{\mathbb{Y}}$ does.
The results in this section rely on the following lemma.
Lemma 3.7. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in\ensuremath{\mathcal{N}}}$. Suppose that $\ensuremath{\mathbb{X}}$ embeds in an $L$ -convex sequence space $\ensuremath{\mathbb{B}}$ via $\ensuremath{\mathcal{X}}$ and that the unit vector system of $\ensuremath{\mathbb{B}}$ is strongly absolute. Then, for every UCC basic sequence $\ensuremath{\mathcal{Y}}=(\ensuremath{\mathbf{y}}_i)_{i\in \ensuremath{\mathcal{M}}}$ with the HBEP in $\ensuremath{\mathbb{X}}$ there is a constant $c$ such that
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}](m)\gtrsim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}},\ensuremath{\mathbb{B}}](\lceil cm\rceil), \quad m\in\ensuremath{\mathbb{N}}.
$$
Proof. Choose for each $i\in\ensuremath{\mathcal{M}}$ an extension $\ensuremath{\mathbf{z}}_i^*$ of the coordinate func- tional $\ensuremath{\mathbf{y}}_i^*$ in such a way that $a=\sup_{i\in \ensuremath{\mathcal{M}}} \left\lVert \ensuremath{\mathbf{y}}_i\right\rVert \left\lVert \ensuremath{\mathbf{z}}_i^*\right\rVert <\infty$, and pick $K$ and $\delta$ as in Lemma 3.5. For each $i\in \ensuremath{\mathcal{M}}$, let $f_i\in\ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}}$ be given by
$$
f_i=\left(\ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{x}}_n) \, \ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\right)_{n\in\ensuremath{\mathcal{N}}}.
$$
Notice that, for each $A\subseteq\ensuremath{\mathcal{M}}$ finite, $\max_{i\in A} \left\lvert f_i\right\rvert\geqslant \delta$ on the set
$$
\Omega_A=\{n \in \ensuremath{\mathcal{N}} \colon \left\lvert \ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{x}}_n) \, \ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert\geqslant \delta \mbox{for some} i\in A\}.
$$
Fix $m\in\ensuremath{\mathbb{N}}$. Let $A\subseteq \ensuremath{\mathcal{M}}$ be a finite set with $\left\lvert A\right\rvert\geqslant m$. Using Lemma 2.4 we obtain
\begin{align*} \left\lVert \ensuremath{\mathbbm{1}}_A[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\right\rVert &\approx \Ave_{\varepsilon_i=\pm1} \left\lVert \sum_{i\in A}\varepsilon_i\, \ensuremath{\mathbf{y}}_i\right\rVert\\
&\gtrsim \Ave_{\varepsilon_i=\pm1}\left\lVert \sum_{i\in A}\varepsilon_i\, \ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \left(\sum_{i\in A}\left\lvert \ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)\right\rvert^2\right)^{1/2}\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \left(\sum_{i\in A}\left\lvert f_i\right\rvert^2\right)^{1/2}\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \max_{i\in A}\left\lvert f_i\right\rvert\right\rVert_\ensuremath{\mathbb{B}}\\
&\gtrsim \left\lVert \ensuremath{\mathbbm{1}}_{\Omega_A}[\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}}, \ensuremath{\mathbb{B}}]\right\rVert. \end{align*}
| |
220809342/19
|
Let $\ensuremath{\mathcal{Y}}$ is a UCC basic sequence with the HBEP in $\ensuremath{\mathbb{X}}$. Then,
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\gtrsim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. Notice that $\ensuremath{\mathbf{s}}:=\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is doubling (see [5] * S8) and that, using democracy, $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\gtrsim\ensuremath{\mathbf{s}}$. Hence, we can apply Theorem 3.9.
3.2.
## Upper estimates for democracy functions
. Our results here heavily depend on the complementability of the closed subspaces of $\ensuremath{\mathbb{X}}$ generated by the basic sequences we tackle. A basic sequence in $\ensuremath{\mathbb{X}}$ that spans a complemented subspace is said to be a complemented basic sequence of $\ensuremath{\mathbb{X}}$.
Lemma 3.11. Suppose that a quasi-Banach space $\ensuremath{\mathbb{X}}$ embeds via a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in\ensuremath{\mathcal{N}}}$ in a sequence space $\ensuremath{\mathbb{B}}$ whose unit vector system is strongly absolute. Let $\ensuremath{\mathcal{Y}}=(\ensuremath{\mathbf{y}}_i)_{i\in\ensuremath{\mathcal{M}}}$ be a complemented basic sequence of $\ensuremath{\mathbb{X}}$. If $\ensuremath{\mathcal{Y}}$ is truncation quasi-greedy then
$$
\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\lesssim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. Let $P\colon\ensuremath{\mathbb{X}}\rightarrow\ensuremath{\mathbb{Y}}$ be a bounded linear projection, where $\ensuremath{\mathbb{Y}}$ is the closed subspace of $\ensuremath{\mathbb{X}}$ generated by $\ensuremath{\mathcal{Y}}$. Set
$$
\ensuremath{\mathbf{z}}_i^*=\ensuremath{\mathbf{y}}_i^*\circ P, \quad i\in I,
$$
where $(\ensuremath{\mathbf{y}}_i^*)_{i\in\ensuremath{\mathcal{M}}}$ in $\ensuremath{\mathbb{Y}}^*$ are the coordinate functionals of $\ensuremath{\mathcal{Y}}$. We have $\ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{y}}_i)=\ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{y}}_i)$ and $\left\lVert \ensuremath{\mathbf{z}}_i^*\right\rVert \left\lVert \ensuremath{\mathbf{y}}_i \right\rVert \leqslant \left\lVert P\right\rVert \left\lVert \ensuremath{\mathbf{y}}_i^*\right\rVert \left\lVert \ensuremath{\mathbf{y}}_i \right\rVert$ for every $i\in I$. Hence, by Lemma 3.5, there are $K\in\ensuremath{\mathbb{N}}$ and $\delta>0$ such that, if we put
$$
\Omega_i=\{n \in \ensuremath{\mathcal{N}} \colon \left\lvert \ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{x}}_n) \, \ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert\geqslant \delta\}, \quad i\in \ensuremath{\mathcal{M}},
$$
then $\left\lvert A\right\rvert\leqslant K \left\lvert \cup_{i\in A} \Omega_i\right\rvert$ for all $A\subseteq \ensuremath{\mathcal{M}}$ finite. Moreover, by Lemma, $\left\lvert \Omega_i\right\rvert<\infty$ for all $i\in\ensuremath{\mathcal{M}}$. Thus, by Theorem 2.9, there are a partition $(\ensuremath{\mathcal{M}}_j)_{j=1}^K$ of $\ensuremath{\mathcal{M}}$ and one-to-one maps $\nu_j\colon \ensuremath{\mathcal{M}}_j\rightarrow\ensuremath{\mathcal{N}}$ such that
$$
\left\lvert \ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{x}}_{\nu_j(i)}) \, \ensuremath{\mathbf{x}}_{\nu_j(i)}^*(\ensuremath{\mathbf{y}}_i)\right\rvert\geqslant \delta, \quad 1\leqslant j \leqslant K, \; i\in \ensuremath{\mathcal{M}}_j.
$$
Pick $m\in\ensuremath{\mathbb{N}}$, and let $A\subseteq \ensuremath{\mathcal{M}}$ be such that $\left\lvert A\right\rvert\leqslant m$. Set $A_j=A\cap\ensuremath{\mathcal{M}}_j$ for $j=1$,..., $K$. Set also
$$
a_{i,n}= \left\lvert \ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert \sgn(\ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{x}}_n)), \quad i\in \ensuremath{\mathcal{M}}, \; n\in\ensuremath{\mathcal{N}}.
$$
For each $l\in A_j$ we have
\begin{align*} \ensuremath{\mathbf{y}}_l^* \left(P\left(\sum_{i\in A_j} a_{l,\nu_j(i)} \ensuremath{\mathbf{x}}_{\nu_j(i)}\right)\right) &=\sum_{i\in A_j} \left\lvert \ensuremath{\mathbf{z}}_l^*(\ensuremath{\mathbf{x}}_{\nu_j(i)}) \, \ensuremath{\mathbf{x}}_{\nu_j(i)}^*(\ensuremath{\mathbf{y}}_l)\right\rvert\\
&\geqslant \left\lvert \ensuremath{\mathbf{z}}_l^*(\ensuremath{\mathbf{x}}_{\nu_j(l)}) \, \ensuremath{\mathbf{x}}_{\nu_j(l)}^*(\ensuremath{\mathbf{y}}_l)\right\rvert\\
&\geqslant \delta. \end{align*}
| |
220809342/5
|
semi-normalized, i.e.,
$$
0<\underset{_}{\text{inf}}{n\in \ensuremath{\mathcal{N}}} \left\lVert \ensuremath{\mathbf{x}}_n\right\rVert\leqslant \sup_{n\in \ensuremath{\mathcal{N}}} \left\lVert \ensuremath{\mathbf{x}}_n\right\rVert<\infty,
$$
and $\ensuremath{\mathcal{X}}^*$ is a basic sequence called the dual basis of $\ensuremath{\mathcal{X}}$. Note that semi-normalized Schauder bases are a particular case of bases.
Given a linearly independent family of vectors $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_ {n\in \ensuremath{\mathcal{N}}}$ in $\ensuremath{\mathbb{X}}$ and scalars $\gamma=(\gamma_n)_{n\in \ensuremath{\mathcal{N}}}\in\ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}}$, we consider the map
$$
S_\gamma=S_\gamma[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\colon {\mathfrak{sp}}_{2n}(\ensuremath{\mathbf{x}}_n \colon n\in\ensuremath{\mathbb{N}}) \rightarrow \ensuremath{\mathbb{X}},
\quad \sum_{n\in \ensuremath{\mathcal{N}}} a_n\, \ensuremath{\mathbf{x}}_n \mapsto \sum_{n\in \ensuremath{\mathcal{N}}} \gamma_n\, a_n \, \ensuremath{\mathbf{x}}_n.
$$
The family $\ensuremath{\mathcal{X}}$ is an unconditional basis of $\ensuremath{\mathbb{X}}$ if and only if it generates the whole space $\ensuremath{\mathbb{X}}$ and $S_\gamma$ is well-defined and bounded on $\ensuremath{\mathbb{X}}$ for all $\gamma\in\ell_\infty$, in which case, the uniform boundedness principle yields
$$
K_{u}=K_{u}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]:=\sup_{\left\lVert \gamma\right\rVert_\infty\leqslant 1} \left\lVert S_\gamma\right\rVert<\infty.
$$
If $\ensuremath{\mathcal{X}}$ is an unconditional basis, $K_u$ is called its unconditional basis constant. Now, given $A\subseteq \ensuremath{\mathbb{N}}$, we define the coordinate projection onto $A$ (with respect to the sequence $\ensuremath{\mathcal{X}}$) as
$$
S_A=S_{\gamma_A}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}],
$$
where $\gamma_A=(\gamma_n)_{n\in \ensuremath{\mathcal{N}}}$ is the family defined by $\gamma_n=1$ if $n\in A$ and $\gamma_n=0$ otherwise. It is known (see, e.g., [5] * Theorem 2.10) that $\ensuremath{\mathcal{X}}$ is an unconditional basis if and only if it generates $\ensuremath{\mathbb{X}}$ and it is suppression unconditional, i.e.,
$$
\sup \{\left\lVert S_A\right\rVert\colon A\subseteq\ensuremath{\mathbb{N}} \mbox{finite}\}<\infty.
$$
Unconditional bases (indexed on the set $\ensuremath{\mathbb{N}}$ of natural numbers) are a particular case of Schauder bases, and so semi-normalized uncondi- tional bases are a particular case of bases.
2.2.
## Quasi-greedy bases
. Given a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ of a quasi- Banach space $\ensuremath{\mathbb{X}}$, with dual basis $\ensuremath{\mathcal{X}}^*=(\ensuremath{\mathbf{x}}_n^*)_{n\in \ensuremath{\mathcal{N}}}$, the coefficient trans- form
\begin{equation*} \ensuremath{\mathcal{F}}\colon \ensuremath{\mathbb{X}} \rightarrow \ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}} , \quad f\mapsto (\ensuremath{\mathbf{x}}_n^*(f))_{n\in \ensuremath{\mathcal{N}}} \end{equation*}
is a bounded linear operator from $\ensuremath{\mathbb{X}}$ into $c_0$. Thus, for each $m\in\ensuremath{\mathbb{N}}$ there is a unique $A=A_m(f)\subseteq \ensuremath{\mathbb{N}}$ of cardinality $\left\lvert A\right\rvert=m$ such that whenever $i\in A$ and $j \in \ensuremath{\mathbb{N}}\setminus A$, either $\left\lvert a_i\right\rvert>\left\lvert a_j\right\rvert$ or $\left\lvert a_i\right\rvert=\left\lvert a_j\right\rvert$ and $i<j$. The $m$ th greedy approximation to $f\in\ensuremath{\mathbb{X}}$ with respect to the basis $\ensuremath{\mathcal{X}}$ is
$$
\ensuremath{\mathcal{G}}_m(f)=\ensuremath{\mathcal{G}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](f):=S_{A_m(f)}(f).
$$
| |
220809342/7
|
Semi-normalized unconditional bases are a special kind of quasi- greedy bases, and although the converse is not true in general, quasi- greedy bases always retain in a certain sense a flavour of unconditional- ity. For example, truncation quasi-greedy bases of quasi-Banach spaces are unconditional for constant coefficients (UCC, for short) [5] * Propo- sition 4.16. This means that there is a constant $C\geqslant 1$ such that $\left\lVert \ensuremath{\mathbbm{1}}_{\varepsilon,A}\right\rVert\leqslant \left\lVert \ensuremath{\mathbbm{1}}_{\varepsilon,B}\right\rVert$ whenever $A$, $B$ are finite subsets of $\ensuremath{\mathbb{N}}$ with $A\subseteq B$ and $\varepsilon\in\ensuremath{\mathbb{E}}^B$. If a basis $\ensuremath{\mathcal{X}}$ is UCC then there is another constant $C_u\geqslant 1$ such that
$$
\left\lVert \ensuremath{\mathbbm{1}}_{\delta,A}\right\rVert\leqslant C_u\left\lVert \ensuremath{\mathbbm{1}}_{\varepsilon,A}\right\rVert
$$
for all finite subsets $A$ of $\ensuremath{\mathbb{N}}$ and all choices of signs $\delta$ and $\varepsilon\in\ensuremath{\mathbb{E}}^A$ (see [5] * Lemma3.2; for a detailed discussion on the unconditionality-related properties enjoyed by truncation quasi-greedy bases in quasi-Banach spaces we refer to [5] * Section 3).
2.4.
## Democracy functions
. Given a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ of a quasi- Banach space $\ensuremath{\mathbb{X}}$ and $A\subseteq\ensuremath{\mathcal{N}}$ finite, we set $\ensuremath{\mathbbm{1}}_A=\ensuremath{\mathbbm{1}}_{\varepsilon,A}$, where $\varepsilon =1$ on $A$. The basis $\ensuremath{\mathcal{X}}$ is said to be democratic if there is a constant $D\geqslant 1$ such that
$$
\left\lVert \ensuremath{\mathbbm{1}}_A\right\rVert\leqslant D \left\lVert \ensuremath{\mathbbm{1}}_B\right\rVert
$$
for any two finite subsets $A$ and $B$ of $\ensuremath{\mathcal{N}}$ with $\left\lvert A\right\rvert\leqslant \left\lvert B\right\rvert$. The lack of democracy of a basis $\ensuremath{\mathcal{X}}$ exhibits some sort of asymmetry. To measure how much a basis $\ensuremath{\mathcal{X}}$ deviates from being democratic, we consider its upper democracy function, also known as its fundamental function,
$$
\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}}, \ensuremath{\mathbb{X}}](m): = \ensuremath{\mathbf{\varphi_u}}(m)=\sup_{\left\lvert A\right\rvert\leqslant m}\left\lVert \ensuremath{\mathbbm{1}}_A\right\rVert,\quad m\in\ensuremath{\mathbb{N}},
$$
and its lower democracy function,
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}}, \ensuremath{\mathbb{X}}](m):= \ensuremath{\mathbf{\varphi_l}}(m)=\underset{_}{\text{inf}}{\left\lvert A\right\rvert\geqslant m}\left\lVert \ensuremath{\mathbbm{1}}_A\right\rVert, \quad m\in\ensuremath{\mathbb{N}}.
$$
Notice that given a basis $\ensuremath{\mathcal{X}}$ there is a constant $C_s$, depending only on the modulus of concavity of the space $\ensuremath{\mathbb{X}}$, such that
$$
\left\lVert \sum_{n\in A} a_n\, \ensuremath{\mathbf{x}}_n\right\rVert \leqslant C_s \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m),\quad \left\lvert A\right\rvert\leqslant m,\; \left\lvert a_n\right\rvert\leqslant 1.
$$
In the particular case that the basis $\ensuremath{\mathcal{X}}$ is UCC the following hold:
$\ensuremath{\mathbf{\varphi_l}}(m)\lesssim\ensuremath{\mathbf{\varphi_u}}(m)$ for
- $m\in\ensuremath{\mathbb{N}}$; $\underset{_}{\text{inf}}{\left\lvert A\right\rvert= m}\left\lVert \ensuremath{\mathbbm{1}}_A\right\rVert\lesssim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}}, \ensuremath{\mathbb{X}}](m)$ for $m\in\ensuremath{\mathbb{N}}$; and
| |
220809342/12
|
rearrangement $(a_m)_{m=1}^\infty$ satisfies
$$
\left\lVert f\right\rVert_{\infty,\ensuremath{\mathbf{w}}}:=\sup_{m\in\ensuremath{\mathbb{N}}} s_m\, a_m<\infty.
$$
We must pay attention to whether $\ensuremath{\mathbf{s}}$ is doubling, i.e., whether $\ensuremath{\mathbf{s}}$ satisfies the condition
$$
s_{2m}\lesssim s_m, \quad m\in\ensuremath{\mathbb{N}}.
$$
The $L$ -convexity of the space will play a key role as well.
Theorem 2.6 ([12, Theorem 2.2.16] and [3, Theorem 6.1]). Let $\ensuremath{\mathbf{s}}$ be the primitive weight of a weight $\ensuremath{\mathbf{w}}$.
- (i) The space $(d_{1,\infty}(\ensuremath{\mathbf{w}}), \left\lVert \cdot\right\rVert_{\infty,\ensuremath{\mathbf{w}}})$ is quasi-normed if and only if $\ensuremath{\mathbf{s}}$ is doubling. Moreover, (ii) if $\ensuremath{\mathbf{s}}$ is doubling, then $d_{1,\infty}(\ensuremath{\mathbf{w}})$ is an $L$ -convex symmetric sequence space.
Suppose $\ensuremath{\mathcal{X}}$ is a truncation quasi-greedy basis of a quasi-Banach space $\ensuremath{\mathbb{X}}$. Then, regardless of whether $\ensuremath{\mathcal{X}}$ is democratic or not, the mere definition of lower democracy function yields a constant $C$ such that
$$
\sup_m a_m\, \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m) \leqslant C \left\lVert f\right\rVert, \quad f\in\ensuremath{\mathbb{X}},
$$
where $(a_m)_{m=1}^\infty$ is the non-increasing rearrangement of $f$.
We point out that, since $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is not necessarily doubling (see [43]), inequality (2.7) might not hold for an embedding of $\ensuremath{\mathbb{X}}$ into a se- quence space; thus we will need to appeal to the following consequence of (2.7).
Lemma 2.7 (see [5, Corollary 9.13]). Let $\ensuremath{\mathcal{X}}$ be a truncation quasi- greedy basis of a quasi-Banach space $\ensuremath{\mathbb{X}}$. Let $\ensuremath{\mathbf{w}}$ be a weight whose prim- itive weight $\ensuremath{\mathbf{s}}=(s_m)_{m=1}^\infty$ is doubling. Then, $\ensuremath{\mathbb{X}} \stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow d_{1,\infty}(\ensuremath{\mathbf{w}})$ if and only if $\ensuremath{\mathbf{s}} \lesssim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$.
2.7.
## Banach envelopes
. When dealing with a quasi-Banach space $\ensuremath{\mathbb{X}}$ it is often convenient to know which is the ‘smallest’ Banach space containing $\ensuremath{\mathbb{X}}$. Formally, the Banach envelope of a quasi-Banach space $\ensuremath{\mathbb{X}}$ consists of a Banach space $\widehat{\ensuremath{\mathbb{X}}}$ together with a linear contraction $J_\ensuremath{\mathbb{X}}\colon\ensuremath{\mathbb{X}} \rightarrow \widehat{\ensuremath{\mathbb{X}}}$ satisfying the following universal property: for every Ba- nach space $\ensuremath{\mathbb{Y}}$ and every linear contraction $T\colon\ensuremath{\mathbb{X}} \rightarrow\ensuremath{\mathbb{Y}}$ there is a unique linear contraction $\widehat{T}\colon \widehat{\ensuremath{\mathbb{X}}}\rightarrow \ensuremath{\mathbb{Y}}$ such that $\widehat{T}\circ J_\ensuremath{\mathbb{X}}=T$. Pictorially we have
| |
220809342/14
|
for every $F\subseteq I$ finite. Then, there are a partition $(I_j)_{j=1}^K$ of $I$ and one-to-one maps $\nu_j\colon I_j\rightarrow \ensuremath{\mathcal{N}}$ for $j=1$,, $K$ such that $\nu_j(i)\in \ensuremath{\mathcal{N}}_i$ for each $i\in I$.
## 3. Strongly absolute bases and their effect on the democracy functions of quasi-greedy basic sequences
Loosely speaking, one could say that strongly absolute bases are ‘purely non-locally convex’ bases, in the sense that if a quasi-Banach space $\ensuremath{\mathbb{X}}$ has a strongly absolute basis then $\ensuremath{\mathbb{X}}$ is far from being a Banach space. The strong absoluteness of a basis was identified and coined by Kalton et al. in [30], as the crucial differentiating feature of unconditional bases in quasi-Banach spaces. Here we work with a slightly different but equivalent definition.
Definition An unconditional basis $\ensuremath{\mathcal{B}}=(\ensuremath{\mathbf{b}}_n)_{n \in \ensuremath{\mathcal{N}}}$ of a quasi- Banach space $\ensuremath{\mathbb{X}}$ is strongly absolute if for every constant $R>0$ there is a constant $C>0$ such that
$$
\sum_{n \in \ensuremath{\mathcal{N}}} \left\lvert \ensuremath{\mathbf{b}}_n^*(f)\right\rvert \left\lVert \ensuremath{\mathbf{b}}_n\right\rVert\leqslant \max\left\{C \sup_{n \in \ensuremath{\mathcal{N}}} \left\lvert \ensuremath{\mathbf{b}}_n^*(f)\right\rvert \left\lVert \ensuremath{\mathbf{b}}_n\right\rVert, \frac{\left\lVert f\right\rVert}{R}\right\}, \quad f\in\ensuremath{\mathbb{X}}.
$$
By definition, if we rescale a strongly absolute basis we obtain an- other strongly absolute basis. A normalized unconditional basis $\ensuremath{\mathcal{B}}$ of $\ensuremath{\mathbb{X}}$ is strongly absolute if $\ensuremath{\mathbb{X}}\stackrel{\ensuremath{\mathcal{B}}}\hookrightarrow\ell_1$, and $\ensuremath{\mathbb{X}}$ is “ far from $\ell_1$, ” in the sense that whenever the quasi-norm of a vector in $\ensuremath{\mathbb{X}}$ and the $\ell_1$ -norm of its coordinate vector are comparable then the $\ell_{\infty}$ -norm of its coordinates is comparable to both quasi-norms. We refer to [3] for a list of quasi- Banach spaces with a strongly absolute unconditional basis. Some of those spaces will appear in Section 4, but for the time being we recall two different ways to find strongly absolute bases.
Theorem 3.2 ([3, Proposition 6.5]). Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a semi-normalized unconditional basis $\ensuremath{\mathcal{B}}$. Suppose that
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{B}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
Then $\ensuremath{\mathcal{B}}$ is strongly absolute.
Theorem 3.3 ([3, Proposition6.2]). Let $\ensuremath{\mathbf{w}}$ be a weight whose primitive weight $\ensuremath{\mathbf{s}}=(s_m)_{m=1}^\infty$ is doubling. Then the unit vector system of $d_{1,\infty}(\ensuremath{\mathbf{w}})$ is strongly absolute if and only if $\sum_{m=1}^\infty 1/s_m<\infty$.
| |
220809342/24
|
Corollary Let $0<p<1$ and $d\in\ensuremath{\mathbb{N}}$. Suppose $\ensuremath{\mathcal{X}}$ is a truncation quasi-greedy basis of a complemented subspace of $H_p(\ensuremath{\mathbb{D}}^d)$. Then for $m\in\ensuremath{\mathbb{N}}$ we have
$$
m^{1/p}(1+\log m)^{-\alpha} \lesssim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},H_p(\ensuremath{\mathbb{D}}^d)](m) \lesssim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},H_p(\ensuremath{\mathbb{D}}^d)](m) \lesssim m^{1/p},
$$
where $\alpha=(d-1)(1/2-1/p)$. Consequently,
$$
\ensuremath{\mathbf{\mu}}_m[\ensuremath{\mathcal{X}},H_p(\ensuremath{\mathbb{D}}^d)]\lesssim (1+\log m)^{\alpha},\quad m\in\ensuremath{\mathbb{N}}.
$$
Proof. Just combine Theorem 4.1 with equations (4.1) and (4.3).
Corollary Let $0<p\leqslant 1$ and $d\in\ensuremath{\mathbb{N}}$. Suppose that $\ensuremath{\mathcal{X}}$ is a truncation quasi-greedy basis of a complemented subspace of $H_p(\ensuremath{\mathbb{D}}^d)$. Then
$$
\ensuremath{\mathbf{L}}_m[\ensuremath{\mathcal{X}}, H_{p}(\ensuremath{\mathbb{D}}^d)] \lesssim (1+\log m)^{\alpha}, \quad m\in\ensuremath{\mathbb{N}},
$$
where $\alpha=\max\{1/p,(d-1)(1/2-1/p)\}$.
Proof. Combine (4.2), Corollary 4.7, and [6] * Theorem 5.1.
## 5. Further applications
Apart from the spaces $H_p(\ensuremath{\mathbb{D}})$ and $\ell_p$ for $0<p<1$, Corollary 4.2 also applies to Fefferman-Stein’s real Hardy spaces $H_p(\ensuremath{\mathbb{R}}^d)$ for $d\in\ensuremath{\mathbb{N}}$. More generally, applying Corollary 4.2 with $\ensuremath{\mathcal{X}}$ a suitable wavelet basis gives that, if $0<p<1$, $0<p\leqslant q\leqslant\infty$, $s\in\ensuremath{\mathbb{R}}$ and $d\in\ensuremath{\mathbb{N}}$, all quasi-greedy bases of the homegeneous and the inhomegeneous Triebel-Lizorkin spaces $\ring{F}_{p,q}^s(\ensuremath{\mathbb{R}}^d)$ and $F_{p,q}^s(\ensuremath{\mathbb{R}}^d)$ and their complemented subspaces are democratic, with fundamental function of the same order as $(m^{1/p})_{m=1}^\infty$ (see *).
Corollary 4.2 also applies to the $p$ -convexified Tsirelson space $\ensuremath{\mathcal{T}}^{(p)}$, $0<p<1$.
The fundamental function of the unit vector system of the Lorentz sequence spaces $\ell_{p,q}$, $0< q\leqslant \infty$, is equivalent to $(m^{1/p})_{m=1}^\infty$. We infer from Corollary 4.2 that all quasi-greedy bases of $\ell_{p,q}$ (we take its separable part if $q=\infty$) are democratic with fundamental function equivalent to $(m^{1/p})_{m=1}^\infty$.
An important class of sequence space whose unit vector system is not democratic are mixed-norm sequence spaces
$$
\ell_p\oplus\ell_q,\;
\ell_p(\ell_q),\;
\ell_q(\ell_p),\;
\left(\bigoplus_{n=1}^\infty \ell_q^n\right)_{\ell_p},\;
\left(\bigoplus_{n=1}^\infty \ell_p^n\right)_{\ell_q}, \quad 0<q<p.
$$
The lower democracy function of all these spaces is $(m^{1/p})_{m=1}^\infty$, whereas their upper lower democracy function is $(m^{1/q})_{m=1}^\infty$. In the case when
| |
220809342/2
|
to image and signal processing, numerical computation, or compressed sensing, to name but a few.
The simplicity in the implementation of the TGA and its connections with the geometry of the space attracted the attention of researchers with a more classical Banach space theory background and as a result, the last two decades have a seen a great progress in the functional analytic aspects of greedy approximation theory.
However, both from the abstract point of view of functional analy- sis as well as its applications, the development of a parallel theory of greedy bases for non-locally convex spaces was left out of the game, not because these spaces lack intrinsic interest but because of the absence of the foundational results that would cement this new ramification of the theory. It can be conceded that the locally convex case is more often used, especially in practical numerical computations, in large part due to the fact that convex algorithms are easy to implement. Neverthe- less, there exist very well established scales of spaces in the non-locally convex setting that arise naturally in analysis and have been exten- sively studied, where it is necessary to do approximation theory. Take for instance the Hardy spaces of analytic functions on various domains in $\ensuremath{\mathbb{C}}^{n}$ (see [18]), Bergman spaces of analytic functions on various do- mains in $\ensuremath{\mathbb{C}}^{n}$ (see [24]), Fefferman-Stein real Hardy spaces (see [20]), Besov, Sobolev, and Triebel-Lizorkin spaces (see [39]). Apart from those spaces, approximation theory in non-locally convex spaces plays a very important role in problems that arise in diverse areas such as approximation spaces, solutions of PDE’s with data in Hardy spaces or Besov spaces that are not Banach spaces [25], approximation spaces and wavelet numerical methods [13], layer potentials and boundary- value problems for second order elliptic operators with data in Besov Spaces, as in [9]. Approximation theory in non-locally convex spaces appears also naturally when studying interpolation problems even in the framework of Banach spaces. For example, the weak Lorentz space $L_{1,\infty}$ plays a key role in Marcinkiewicz interpolation theorem [10].
As was recently shown in [5], the main types of bases that are of in- terest in greedy approximation in the setting of Banach spaces, namely greedy, almost greedy, and quasi-greedy bases, are suitable as well for the use of the TGA in $p$ -Banach spaces for $p<1$. The article [5] was the springboard for subsequent research of different aspects related to the greedy algorithm in $p$ -Banach spaces for $p<1$ (see AAW2021b, AAW2021). Our aim in this paper is to continue investigating the con- nection between quasi-greedy bases and their democracy functions in $p$ -Banach spaces, initiated in [7], with an eye to the qualitative and
| |
220809342/23
|
Proof. Assume by contradiction that $\ensuremath{\mathcal{Y}}$ has a squeeze-symmetric ba- sis $\ensuremath{\mathcal{Y}}$. Then $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\approx\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]$ is doubling. By Theorem 4.4 we have $\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \approx\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ and $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \approx\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$, which leads to the absurdity $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$.
Corollary Let $0<p<1$ and $d\in\ensuremath{\mathbb{N}}$. If $d\geqslant 2$ the space $H_p(\ensuremath{\mathbb{D}}^d)$ has no sequeeze-symmetric bases. In particular, $H_p(\ensuremath{\mathbb{D}}^d)$ has no almost greedy bases.
Proof. Let $\ensuremath{\mathbf{h}}=(h_m)_{m=1}^\infty$ be as in (4.1). Since $\ensuremath{\mathbf{h}}$ is doubling and non- equivalent to $(m^{1/p})_{m=1}^\infty$, the result follows from Theorem 4.5.
We close this section with a quantitative estimate for the perfor- mance of the thresholding greedy algorithm implemented in Hardy spaces. To put this in context, we recall that in order to analyze the efficiency of the greedy algorithm with respect a basis $\ensuremath{\mathcal{X}}$ of a quasi- Banach space $\ensuremath{\mathbb{X}}$, it is customary to consider, for each $m\in\ensuremath{\mathbb{N}}$, the smallest constant $C\geqslant 1$ such that
$$
\left\lVert f- \ensuremath{\mathcal{G}}_m(f) \right\rVert \leqslant C \left\lVert f- z\right\rVert
$$
for all $f\in \ensuremath{\mathbb{X}}$ and all $m$ -term linear combinations $z$ of vectors from $\ensuremath{\mathcal{X}}$. This constant is called the $m$ th Lebesgue constant, and it is denoted by $\ensuremath{\mathbf{L}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$. The growth of the Lebesgue constants of bases in Banach spaces has been studied in GHO2013,BBGHO2018. We point out that the relation between the Lebesgue constants, the conditionality param- eters, and the democratic deficiency parameters established in [11] still holds in the non-locally convex setting. To be precise, if we put
\begin{equation*} \ensuremath{\mathbf{k}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}] =\sup_{\left\lvert A\right\rvert\leqslant m} \left\lVert S_A[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\right\rVert, \quad m\in\ensuremath{\mathbb{N}}, \end{equation*}
and
$$
\ensuremath{\mathbf{\mu}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]=\sup\limits_{\left\lvert A\right\rvert=\left\lvert B\right\rvert\leqslant m}\frac{\left\lVert \ensuremath{\mathbbm{1}}_A[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\right\rVert{\left\lVert \ensuremath{\mathbbm{1}}\right\rVert_B[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]}}, \quad m\in\ensuremath{\mathbb{N}}
$$
then for any basis $\ensuremath{\mathcal{X}}$ of a quasi-Banach space $\ensuremath{\mathbb{X}}$ we have
$$
\ensuremath{\mathbf{L}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx \max\{\ensuremath{\mathbf{k}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}] , \ensuremath{\mathbf{\mu}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\} , \quad m\in\ensuremath{\mathbb{N}}.
$$
Moreover, if $\ensuremath{\mathcal{X}}$ is UCC, then
$$
\ensuremath{\mathbf{\mu}}_m[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx\sup\limits_{l\leqslant m}\frac{\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](l)}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](l)}, \quad m\in\ensuremath{\mathbb{N}}.
$$
(see [4] * S1).
| |
220809342/16
|
$J\colon\ensuremath{\mathbb{B}}\rightarrow\widehat{\ensuremath{\mathbb{X}}}$ such
that $J(\ensuremath{\mathcal{F}}(f))=J_\ensuremath{\mathbb{X}}(f)$ for all $f\in\ensuremath{\mathbb{X}}$. Let $C_1$ denote its norm, and let $C_2$ be the norm of the coefficient transform $\ensuremath{\mathcal{F}}$ with respect to $\ensuremath{\mathcal{X}}$, regarded as an operator from $\ensuremath{\mathbb{X}}$ to $\ensuremath{\mathbb{B}}$. Set
\begin{align*} C_3&=\sup_{i\in A} \left\lVert \ensuremath{\mathbf{y}}_i\right\rVert \left\lVert \ensuremath{\mathbf{y}}_i^*\right\rVert <\infty,\; \mbox{and}\\
C_4&=\sup_{n\in \ensuremath{\mathcal{N}}} \left\lVert \ensuremath{\mathbf{x}}_n\right\rVert \left\lVert \ensuremath{\mathbf{x}}_n^*\right\rVert<\infty. \end{align*}
If we set
$$
\ensuremath{\mathbf{z}}_i=\ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)\in\ensuremath{\mathbb{B}}\; \mbox{and}\; \ensuremath{\mathbf{z}}_i^*= \widehat{\ensuremath{\mathbf{y}}_i^*}\circ J\in\ensuremath{\mathbb{B}}^*,\quad i\in \ensuremath{\mathcal{M}},
$$
then we have
\begin{align*} \ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{z}}_i)&=\widehat{\ensuremath{\mathbf{y}}_i^*}\circ J\circ \ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i)=\ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{y}}_i)=1\; \mbox{and}\\
\left\lVert \ensuremath{\mathbf{z}}_i\right\rVert \left\lVert \ensuremath{\mathbf{z}}_i^*\right\rVert &\leqslant \left\lVert \ensuremath{\mathcal{F}}\right\rVert_{\ensuremath{\mathbb{X}}\rightarrow\ensuremath{\mathbb{B}}} \left\lVert \ensuremath{\mathbf{y}}_i\right\rVert \left\lVert J\right\rVert\left\lVert \widehat{\ensuremath{\mathbf{y}}_i^*}\right\rVert \leqslant C_1 C_2 \left\lVert \ensuremath{\mathbf{y}}_i\right\rVert \left\lVert \ensuremath{\mathbf{y}}_i^*\right\rVert\leqslant C_1C_2C_3 \end{align*}
for all $i\in A$. Moreover,
\begin{align*} \ensuremath{\mathbf{e}}_n^*(\ensuremath{\mathbf{z}}_i)&=\ensuremath{\mathbf{e}}_n^*(\ensuremath{\mathcal{F}}(\ensuremath{\mathbf{y}}_i))=\ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\; \mbox{and}\\
\ensuremath{\mathbf{z}}_i^*(\ensuremath{\mathbf{e}}_n)&=\widehat{\ensuremath{\mathbf{y}}_i^*}(J(\ensuremath{\mathcal{F}}(\ensuremath{\mathbf{x}}_n)))=\widehat{\ensuremath{\mathbf{y}}_i^*}(J_\ensuremath{\mathbb{X}}(\ensuremath{\mathbf{x}}_n))=\ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{x}}_n) \end{align*}
for all $i\in A$ and $n\in\ensuremath{\mathcal{N}}$.
Set $a=C_1 C_2 C_3$ and an arbitrary $C>1$. We pick $\delta>0$ as in Lemma 3.4 with respect to the unit vector system of $\ensuremath{\mathbb{B}}$. We infer that
$$
\left\lvert A\right\rvert\leqslant C \sum_{n\in \Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{X}})} \left\lvert \ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{x}}_n) \, \ensuremath{\mathbf{x}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert\leqslant C C_3 C_4 \left\lvert \Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{X}})\right\rvert.\qedhere
$$
The following elementary lemma puts an end to the auxiliary results of this preparatory section.
Lemma 3.6. Let $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in\ensuremath{\mathcal{N}}}$ be a basis a quasi-Banach space $\ensuremath{\mathbb{X}}$. The set
$$
\{n\in\ensuremath{\mathcal{N}} \colon \left\lvert f^*(\ensuremath{\mathbf{x}}_n) \, \ensuremath{\mathbf{x}}_n^*(f)\right\rvert\geqslant \delta\}
$$
is finite for any given $f\in\ensuremath{\mathbb{X}}$, $f^*\in\ensuremath{\mathbb{X}}^*$, and $\delta>0$.
Proof. We need to prove that $(f^*(\ensuremath{\mathbf{x}}_n)\,\ensuremath{\mathbf{x}}_n^*(f))_{n\in\ensuremath{\mathcal{N}}}\in c_0(\ensuremath{\mathcal{N}})$. But this follows from the facts that $(\ensuremath{\mathbf{x}}_n^*(f))_{n\in\ensuremath{\mathcal{N}}}\in c_0(\ensuremath{\mathcal{N}})$ and $(f^*(\ensuremath{\mathbf{x}}_n))_{n\in\ensuremath{\mathcal{N}}}\in \ell_\infty(\ensuremath{\mathcal{N}})$.
The machinery developed above will permit us to obtain estimates for democracy functions of basic sequences in a quasi-Banach space that embeds in $\ell_1$ via a basis that is far from the canonical $\ell_1$ -basis in a sense that will be made clear in place. We divide these estimates in two, depending on whether they involve lower or upper democracy functions.
| |
220809342/11
|
are the restriction to $\ensuremath{\mathbb{B}}$ of the coordinate functionals $(\ensuremath{\mathbf{e}}_n^*)_{n\in \ensuremath{\mathcal{N}}}$ defined as
$$
\ensuremath{\mathbf{e}}_n^*(f)=a_n, \quad f=(a_n)_{n\in\ensuremath{\mathcal{N}}}\in\ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}}.
$$
And, conversely, every semi-normalized unconditional basis $\ensuremath{\mathcal{X}}$ of a quasi-Banach space $\ensuremath{\mathbb{X}}$ becomes normalized and $1$ -unconditional after a suitable renorming of $\ensuremath{\mathbb{X}}$; this way we can associate a sequence space with $\ensuremath{\mathcal{X}}$.
2.6.
## Embeddings via bases and squeeze-symmetric bases
. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ and let $(\ensuremath{\mathbb{B}},\left\lVert \cdot\right\rVert_\ensuremath{\mathbb{B}})$ be a sequence space on $\ensuremath{\mathcal{N}}$. Let us recall the following terminology, which we borrow from [1].
- (a) We say that $\ensuremath{\mathbb{B}}$ embeds in $\ensuremath{\mathbb{X}}$ via $\ensuremath{\mathcal{X}}$, and put $\ensuremath{\mathbb{B}}\stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow \ensuremath{\mathbb{X}}$, if there is a constant $C$ such that for every $g\in\ensuremath{\mathbb{B}}$ there is $f\in\ensuremath{\mathbb{X}}$ such that $\ensuremath{\mathcal{F}}(f)=g$, and $\left\lVert f\right\rVert\leqslant C\left\lVert g\right\rVert_{\ensuremath{\mathbb{B}}}$.
- (b) We say that $\ensuremath{\mathbb{X}}$ embeds in $\ensuremath{\mathbb{B}}$ via $\ensuremath{\mathcal{X}}$, and put $\ensuremath{\mathbb{X}}\stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow\ensuremath{\mathbb{B}}$, if there is a constant $C$ such that $\ensuremath{\mathcal{F}}(f)\in \ensuremath{\mathbb{B}}$ with $\left\lVert \ensuremath{\mathcal{F}}(f)\right\rVert_{\ensuremath{\mathbb{B}}}\leqslant C \left\lVert f\right\rVert$ for all $f\in \ensuremath{\mathbb{X}}$.
The sequence space $\ensuremath{\mathbb{B}}$ is said to be symmetric if $f_\pi:=(a_{\pi(n)})_{n\in \ensuremath{\mathcal{N}}}\in\ensuremath{\mathbb{B}}$ and $\left\lVert f_\pi\right\rVert_\ensuremath{\mathbb{B}}=\left\lVert f\right\rVert_\ensuremath{\mathbb{B}}$ for all $f=(a_n)_{n\in \ensuremath{\mathcal{N}}}\in\ensuremath{\mathbb{B}}$ and every permutation $\pi$ of $\ensuremath{\mathcal{N}}$.
Loosely speaking, by squeezing the space $\ensuremath{\mathbb{X}}$ between two symmetric sequence spaces we obtain qualitative estimates on the symmetry of the basis in $\ensuremath{\mathbb{X}}$. Thus, a basis $\ensuremath{\mathcal{X}}$ is said to be squeeze-symmetric if there are symmetric sequence spaces $\ensuremath{\mathbb{S}}_1$ and $\ensuremath{\mathbb{S}}_2$ on $\ensuremath{\mathcal{N}}$, which are close to each other in the sense that
$$
\left\lVert \ensuremath{\mathbbm{1}}_A[\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}},\ensuremath{\mathbb{S}}_1]\right\rVert \approx \left\lVert \ensuremath{\mathbbm{1}}_A[\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}},\ensuremath{\mathbb{S}}_2]\right\rVert, \quad \left\lvert A\right\rvert<\infty,
$$
such that $\ensuremath{\mathbb{S}}_1\stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow \ensuremath{\mathbb{X}} \stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow \ensuremath{\mathbb{S}}_2$. A basis of a quasi-Banach space is squeeze- symmetric if and only if it is truncation quasi-greedy and democratic (see [5] * Lemma 9.3 and Corollary 9.15). In paticular, almost greedy bases are squeeze-symmetric.
Embeddings involving weighted Lorentz sequence spaces play an im- portant role in greedy approximation theory using bases. For our pur- poses here, it will be sufficient to deal with weak Lorentz spaces.
Let $\ensuremath{\mathbf{w}}=(w_n)_{n=1}^\infty$ be a weight, i.e., a sequence of nonnegative numbers with $w_1>0$, and let $\ensuremath{\mathbf{s}}=(s_m)_{m=1}^\infty$ be its primitive weight defined by $s_m=\sum_{n=1}^m w_n$. The weighted Lorentz sequence space (on the countable set $\ensuremath{\mathcal{N}}$) $d_{1,\infty}(\ensuremath{\mathbf{w}})$ consists of all functions $f\in c_0(\ensuremath{\mathcal{N}})$ whose non-increasing
| |
220809342/6
|
Note that the operators $(\ensuremath{\mathcal{G}}_m)_{m=1}^{\infty}$ defining the greedy algorithm on $\ensuremath{\mathbb{X}}$ with respect to $\ensuremath{\mathcal{X}}$ are not linear nor continuous. The basis $\ensuremath{\mathcal{X}}$ is said to be quasi-greedy if there is a constant $C\geqslant 1$ such that
$$
\left\lVert \ensuremath{\mathcal{G}}_m(f)\right\rVert\leqslant C\left\lVert f\right\rVert, \quad f\in\ensuremath{\mathbb{X}}, \, m\in\ensuremath{\mathbb{N}}.
$$
Equivalently, by [42] * Theorem 1 (cf. [5] * Theorem 4.1), these are pre- cisely the bases for which the greedy algorithm converges, i.e.,
$$
\lim_{m\rightarrow\infty} \ensuremath{\mathcal{G}}_{m}(f)=f, \quad f\in \ensuremath{\mathbb{X}}.
$$
2.3.
## Truncation quasi-greedy bases
. Another family of nonlinear operators of key relevance in the study of the greedy algorithm in a quasi-Banach space $\ensuremath{\mathbb{X}}$ with respect to a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ is the sequence $(\ensuremath{\mathcal{R}}_m)_{m=1}^{\infty}$ of restricted truncation operators defined as follows. Let
$$
\ensuremath{\mathbb{E}}=\{\lambda\in\ensuremath{\mathbb{F}} \colon \left\lvert \lambda\right\rvert=1\}.
$$
Given $A\subset\ensuremath{\mathcal{N}}$ finite and $\varepsilon=(\varepsilon_n)_{n\in A}\in\ensuremath{\mathbb{E}}^A$ we set
$$
\ensuremath{\mathbbm{1}}_{\varepsilon,A}= \sum_{n\in A} \varepsilon_n \,\ensuremath{\mathbf{x}}_n,
$$
and given $f\in\ensuremath{\mathbb{X}}$ we define $\varepsilon(f)\in\ensuremath{\mathbb{E}}^{\ensuremath{\mathcal{N}}}$ by
$$
\varepsilon(f)=(\sgn(\ensuremath{\mathbf{x}}_n^*(f)))_{n\in\ensuremath{\mathcal{N}}},
$$
where, as is customary, $\sgn(\cdot)$ denotes the sign function, i.e., $\sgn(0)=1$ and $\sgn(a)=a/\left\lvert a\right\rvert$ if $a\in\ensuremath{\mathbb{F}}\setminus\{0\}$. For $m\in \ensuremath{\mathbb{N}}$, the $m$ th-restricted truncation operator $\ensuremath{\mathcal{R}}_m\colon \ensuremath{\mathbb{X}} \rightarrow \ensuremath{\mathbb{X}}$ is the map
$$
\ensuremath{\mathcal{R}}_m(f)=\min_{n\in A_m(f)} \left\lvert \ensuremath{\mathbf{x}}_n^*(f)\right\rvert \ensuremath{\mathbbm{1}}_{\varepsilon(f),A_m(f)}, \quad f\in\ensuremath{\mathbb{X}},
$$
and the basis $\ensuremath{\mathcal{X}}$ is said to be truncation quasi-greedy if
$$
C_r:=\sup_{m\in\ensuremath{\mathbb{N}}}\left\lVert \ensuremath{\mathcal{R}}_m\right\rVert<\infty.
$$
For the sake of generality, most of our results below will be stated and proved for truncation quasi-greedy bases (or even for bases fulfill- ing weaker unconditionality conditions such as being UCC). However, the uneasy reader can safely replace “ truncation quasi-greedy basis ” with “ quasi-greedy basis ” and can rest assured of the validity of the corresponding statements thanks to the following theorem.
Theorem 2.1 ([5, Theorem 4.13]). If $\ensuremath{\mathcal{X}}$ is a quasi-greedy basis in a quasi-Banach space $\ensuremath{\mathbb{X}}$ then $\ensuremath{\mathcal{X}}$ is truncation quasi-greedy.
| |
220809342/4
|
Question 3.8 from [7] in the case when $d=1$. As far as the Hardy spaces $H_p(\ensuremath{\mathbb{D}}^d)$ for $d\geqslant 2$ is concerned, Wojtaszczyk had shown that its canonical basis (which is unconditional) is not democratic. Here, our contribution consists of proving that, actually, no quasi-greedy basis of these multivariate Hardy spaces is democratic. Let us point out that the approach used to prove that quasi-greedy bases in $\ell_p$ are democratic falls short for the Hardy spaces, since the linear structure of the latter spaces is far more complex that the former. The new techniques that we had to develop to tackle the problem critically depend on the convexity properties of a quasi-Banach space regarded as a quasi-Banach lattice with the structure induced by its (unique) unconditional basis.
Our research suggests the pattern that if a quasi-Banach space $\ensuremath{\mathbb{X}}$ (locally convex or otherwise) has a unique unconditional basis (up to equivalence and permutation) which is democratic (hence greedy) then all quasi-greedy bases of $\ensuremath{\mathbb{X}}$ are democratic (hence almost greedy).
We also obtain valuable information about the democracy functions of quasi-greedy bases $\ensuremath{\mathcal{X}}$ of other classical quasi-Banach spaces $\ensuremath{\mathbb{X}}$ with a unique unconditional basis. For instance, we prove that the mixed- norm sequence spaces $\ell_q\oplus\ell_p$ for $0<p<q<1$ have no almost greedy bases.
We close this introduction with a brief description of the contents of the paper. Section 3 includes our advances in the theory of quasi-greedy bases. Previously, in Section 2 we set the terminology that we will use and gather some background results that we will need. In Section 4 we provide applications to important spaces in functional analysis and approximation theory.
## 2. Terminology and background
Throughout this paper we use standard facts and notation from Banach spaces and approximation theory (see, e.g., [8]). The reader will find the required specialized background and notation on greedy-like bases in quasi-Banach spaces in [5]. Nonetheless, we record the notation that is most heavily used.
2.1.
## Bases in quasi-Banach spaces
. Throughout this paper, a ba- sis of a quasi-Banach space $\ensuremath{\mathbb{X}}$ over the real or complex field $\ensuremath{\mathbb{F}}$ will be a norm-bounded countable family $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ which generates the entire space $\ensuremath{\mathbb{X}}$, and for which there is a (unique) norm-bounded family $\ensuremath{\mathcal{X}}^*=(\ensuremath{\mathbf{x}}_n^*)_{n\in \ensuremath{\mathcal{N}}}$ in the dual space $\ensuremath{\mathbb{X}}^*$ such that $(\ensuremath{\mathbf{x}}_n, \ensuremath{\mathbf{x}}_n^*)_{n\in \ensuremath{\mathcal{N}}}$ is a biorthogonal system. A basic sequence will be a sequence in $\ensuremath{\mathbb{X}}$ which is a basis of its closed linear span. If $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in \ensuremath{\mathcal{N}}}$ is a basis, then it is
| |
220809342/15
|
We point out that the fundamental function of the unit vector sys- tem of $d_{1,\infty}(\ensuremath{\mathbf{w}})$ is the primitive weight of $\ensuremath{\mathbf{w}}$. Hence, the ‘if’ part of Theorem 3.3 can be directly derived from Theorem 3.2.
Our goal in this section is to determine how the fact that a quasi- Banach space $\ensuremath{\mathbb{X}}$ has a strongly absolute basis $\ensuremath{\mathcal{B}}=(\ensuremath{\mathbf{b}}_n)_{n\in\ensuremath{\mathcal{N}}}$ affects the democracy functions of quasi-greedy bases in $\ensuremath{\mathbb{X}}$. To that end, if $\ensuremath{\mathcal{Y}}=(\ensuremath{\mathbf{y}}_i)_{i\in\ensuremath{\mathcal{M}}}$ is another basis of $\ensuremath{\mathbb{X}}$, we must estimate the size of $\left\lVert \sum_{i\in A}\ensuremath{\mathbf{y}}_{i}\right\rVert$ for any $A\subset \ensuremath{\mathcal{M}}$ finite in terms of the democracy functions of $\ensuremath{\mathcal{B}}$. The following lemma from [7] highlights the role played by the strongly absoluteness of the basis in that it makes possible picking large coefficients from the vectors of $\ensuremath{\mathcal{Y}}$. The “ large coefficient ” technique was introduced by Kalton in [26] in his study of the uniqueness of unconditional basis in non-locally convex Orlicz sequence spaces.
Given $\delta>0$ and a finite family $\ensuremath{\mathcal{S}}=(\ensuremath{\mathbf{y}}_i,\ensuremath{\mathbf{y}}_i^*)_{i\in A}$ we consider the set of indices
\begin{equation*} \Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{B}})=\{n \in \ensuremath{\mathcal{N}} \colon \left\lvert \ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{b}}_n) \, \ensuremath{\mathbf{b}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert\geqslant \delta \mbox{for some} i\in A\}, \end{equation*}
where $(\ensuremath{\mathbf{b}}_n^{\ast})_{n\in \ensuremath{\mathcal{M}}}$ is the dual basis of $\ensuremath{\mathcal{B}}$. The explicit definition of these sets goes back to the work of Wojtaszczyk on uniqueness of uncondi- tional structure of quasi-Banach spaces [41].
Lemma 3.4 ([7, Lemma 3.3]). Let $\ensuremath{\mathbb{B}}$ be a quasi-Banach space with a strongly absolute basis $\ensuremath{\mathcal{B}}$. Then, given $a\in(0,\infty)$ and $C\in(1,\infty)$, there is $\delta>0$ such that, whenever $\ensuremath{\mathcal{S}}=(\ensuremath{\mathbf{y}}_i,\ensuremath{\mathbf{y}}_i^*)_{i\in A}$ is a finite family in $\ensuremath{\mathbb{B}}\times\ensuremath{\mathbb{B}}^*$ with $\ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{y}}_i)=1$ and $\left\lVert \ensuremath{\mathbf{y}}_i\right\rVert\left\lVert \ensuremath{\mathbf{y}}_i^*\right\rVert\leqslant a$ for all $i\in A$, we have
$$
\left\lvert A\right\rvert\leqslant C \sum_{n\in \Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{B}})} \left\lvert \sum_{i\in A} \ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{b}}_n) \, \ensuremath{\mathbf{b}}_n^*(\ensuremath{\mathbf{y}}_i)\right\rvert.
$$
Since we will also come across with sets $\Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{X}})$ associated to con- ditional bases $\ensuremath{\mathcal{X}}$ of a quasi-Banach space $\ensuremath{\mathbb{X}}$, in order to make headway we will need the following alteration of Lemma 3.4.
Lemma 3.5. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a basis $\ensuremath{\mathcal{X}}=(\ensuremath{\mathbf{x}}_n)_{n\in\ensuremath{\mathcal{N}}}$. Suppose that $\ensuremath{\mathbb{X}}$ embeds via $\ensuremath{\mathcal{X}}$ in a sequence space $\ensuremath{\mathbb{B}}$ whose unit vector system is strongly absolute. Then, given $a\in(0,\infty)$ there are positive scalars $K$ and $\delta$ such that, whenever $\ensuremath{\mathcal{S}}=(\ensuremath{\mathbf{y}}_i,\ensuremath{\mathbf{y}}_i^*)_{i\in A}$ is a finite family in $\ensuremath{\mathbb{X}}\times\ensuremath{\mathbb{X}}^*$ with $\ensuremath{\mathbf{y}}_i^*(\ensuremath{\mathbf{y}}_i)=1$ and $\left\lVert \ensuremath{\mathbf{y}}_i\right\rVert\left\lVert \ensuremath{\mathbf{y}}_i^*\right\rVert\leqslant a$ for all $i\in A$, we have
$$
\left\lvert A\right\rvert\leqslant K \left\lvert \Omega_\delta(\ensuremath{\mathcal{S}},\ensuremath{\mathcal{X}})\right\rvert.
$$
Proof. The hypothesis implies that the space $\ensuremath{\mathbb{B}}$ embeds continuously into $\ell_1$. Terefore, by Lemma 2.8, there is a bounded linear map $J\colon\ensuremath{\mathbb{B}}\rightarrow\widehat{\ensuremath{\mathbb{X}}}$
| |
220809342/10
|
Lemma 2.4. Let $\ensuremath{\mathbb{X}}$ be an $L$ -convex quasi-Banach lattice. Then for every $0<r<\infty$ there is constant $C$ such that
$$
\left\lVert \left(\sum_{i\in A} \left\lvert f_i\right\rvert^2\right)^{1/2}\right\rVert
\leqslant C\left(\Ave_{\varepsilon_i=\pm 1} \left\lVert \sum_{i\in A} \varepsilon_i\, f_i\right\rVert^r\right)^{1/r},\quad \left\lvert A\right\rvert<\infty,\; f_i\in\ensuremath{\mathbb{X}}.
$$
Proof. Since that mapping
$$
r\mapsto \left(\Ave_{\varepsilon_i=\pm 1} \left\lVert \sum_{i\in A} \varepsilon_i\, f_i\right\rVert^r\right)^{1/r}
$$
is increasing, by Theorem 2.3, we can assume the $\ensuremath{\mathbb{X}}$ is an $r$ -convex Banach lattice. Then, by Theorem 2.2,
\begin{align*} \left\lVert \left(\sum_{i\in A} \left\lvert f_i\right\rvert^2\right)^{1/2}\right\rVert &\leqslant T_r \left\lVert \left(\Ave_{\varepsilon_i=\pm 1}\left\lvert \sum_{i\in A} \varepsilon_i\, f_i\right\rvert^r\right)^{1/r}\right\rVert\\
&\leqslant T_r M^{(r)}(\ensuremath{\mathbb{X}}) \left(\Ave_{\varepsilon_i=\pm 1} \left\lVert \sum_{i\in A} \varepsilon_i\, f_i\right\rVert^r\right)^{1/r} \end{align*}
for every finite family $(f_i)_{i\in A}$ in $\ensuremath{\mathbb{X}}$.
Remark 2.5. In light of Khintchine-Kahane-Kalton’s inequalities (see [27] * Theorem 2.1) the index $r$ in Lemma 2.4 is irrelevant. From an opposite perspective, we point out that the proof of Lemma 2.4 does not depend on Khintchine-Kahane-Kalton’s inequalities.
We are interested in quasi-Banach lattices of functions defined on a countable set $\ensuremath{\mathcal{N}}$. The term sequence space will apply to a quasi-Banach space $\ensuremath{\mathbb{B}}\subseteq \ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}}$ such that:
$\ensuremath{\mathbf{e}}_n:=(\delta_{i,n})_{i\in \ensuremath{\mathcal{N}}}$ is
- a norm-one vector in $\ensuremath{\mathbb{B}}$ for all $n\in \ensuremath{\mathcal{N}}$; and
- if $f\in \ensuremath{\mathbb{X}}$ and $g\in\ensuremath{\mathbb{F}}^\ensuremath{\mathcal{N}}$ satisfy $\left\lvert g\right\rvert\leqslant\left\lvert f\right\rvert$, then $g\in\ensuremath{\mathbb{X}}$ and $\left\lVert g\right\rVert\leqslant \left\lVert f\right\rVert$.
That way, $\ensuremath{\mathbb{B}}$ becomes a quasi-Banach lattice with the natural order. In this particular case $r$ -sums take a more workable form: given $f_i=(a_{i,n})_{n\in \ensuremath{\mathcal{N}}}$ for $i\in A$,
$$
\left(\sum_{i\in A}\left\lvert f_i\right\rvert^r\right)^{1/r} = \left(\left(\sum_{i\in A} \left\lvert a_{i,n}\right\rvert^r\right)^{1/r}\right)_{n\in \ensuremath{\mathcal{N}}}.
$$
There is a close relation between sequence spaces and unconditional bases. If $\ensuremath{\mathbb{B}}$ is a sequence space, the unit vector system $\ensuremath{\mathcal{E}}_\ensuremath{\mathcal{N}}=(\ensuremath{\mathbf{e}}_n)_{n\in \ensuremath{\mathcal{N}}}$ is a $1$ -unconditional basic sequence of $\ensuremath{\mathbb{B}}$ whose coordinate functionals
| |
220809342/13
|
$$
\xymatrix{
\widehat{\ensuremath{\mathbb{X}}}\ar[drr]^{\widehat{T}} & \\
\ensuremath{\mathbb{X}} \ar[u]^{J_\ensuremath{\mathbb{X}}} \ar[rr]_T && \ensuremath{\mathbb{Y}}.
}
$$
If a Banach space $\ensuremath{\mathbb{V}}$ and a bounded linear map $J\colon \ensuremath{\mathbb{X}}\rightarrow \ensuremath{\mathbb{V}}$ are such that $\widehat{J}\colon \widehat{\ensuremath{\mathbb{X}}}\rightarrow \ensuremath{\mathbb{V}}$ is an isomorphism, we say that $\ensuremath{\mathbb{V}}$ is an isomorphic rep- resentation of the Banach envelope of $\ensuremath{\mathbb{X}}$ via $J$. For instance, given $0<p<1$, the Bergman space $A^1_{1-p/2}$ is an isomorphic represen- tation of the Banach envelope of $H_p(\ensuremath{\mathbb{D}})$ via the inclusion map (see DRS1969,Shapiro1976).
If a quasi-Banach space $\ensuremath{\mathbb{X}}$ has a basis $\ensuremath{\mathcal{X}}$, and $\ensuremath{\mathbb{X}}\stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow \ell_1$, then $\ell_1$ is an isomorphic representation of the Banach envelope of $\ensuremath{\mathbb{X}}$ via the coefficient transform ([2] * Proposition 2.10). The following lemma is an immediate consequence of this and so we leave its verification to the reader.
Lemma 2.8. Let $\ensuremath{\mathcal{X}}$ be a basis of a quasi-Banach space $\ensuremath{\mathbb{X}}$, and let $\ensuremath{\mathbb{B}}$ be a sequence space. Suppose that $\ensuremath{\mathbb{X}}\stackrel{\ensuremath{\mathcal{X}}}\hookrightarrow \ensuremath{\mathbb{B}}$ and that $\ensuremath{\mathbb{B}}\subseteq\ell_1$ continuously. Then the envelope map $J_\ensuremath{\mathbb{X}}\colon\ensuremath{\mathbb{X}}\rightarrow\widehat{\ensuremath{\mathbb{X}}}$ factors through the coefficient trans- form regarded as map from $\ensuremath{\mathbb{X}}$ into $\ensuremath{\mathbb{B}}$, i.e., there is a bounded linear map $J\colon\ensuremath{\mathbb{B}}\rightarrow\widehat{\ensuremath{\mathbb{X}}}$ such that $J_\ensuremath{\mathbb{X}}=J\circ \ensuremath{\mathcal{F}}$.
2.8.
## The Marriage Lemma
. A classical problem in combinatorics is to determine whether a given family $(\ensuremath{\mathcal{N}}_i)_{i\in I}$ of subsets of a set $\ensuremath{\mathcal{N}}$ admits a one-to-one map $\nu\colon I\rightarrow\ensuremath{\mathcal{N}}$ such that $\nu(i)\in \ensuremath{\mathcal{N}}_i$ for all $i\in I$. A necessary condition for the existence of such a map is
$$
\left\lvert F\right\rvert\leqslant\left\lvert \bigcup_{i\in F} \ensuremath{\mathcal{N}}_i\right\rvert, \quad F\subseteq I,\; \left\lvert F\right\rvert<\infty.
$$
P. Hall proved in [23] that (2.8) is also sufficient, provided the set $I$ of indices and all the sets $\ensuremath{\mathcal{N}}_i$ are finite. Subsequently, M. Hall [22] ex- tended this result to the case when $I$ is not necessarily finite. Here, we will use a generalization by Wojtaszczyk [41] of the latter result which was effectively used in his study of the uniqueness of unconditional bases in quasi-Banach spaces.
Theorem Let $\ensuremath{\mathcal{N}}$ be a set and $(\ensuremath{\mathcal{N}}_i)_{i\in I}$ be a family of finite subsets of $\ensuremath{\mathcal{N}}$. Let $K\in\ensuremath{\mathbb{N}}$. Suppose that
$$
\left\lvert F\right\rvert\leqslant K\left\lvert \bigcup_{i\in F} \ensuremath{\mathcal{N}}_i\right\rvert
$$
| |
220809342/21
|
that the unit vector system of $\ensuremath{\mathbb{B}}$ is strongly absolute. This way, in both cases an application of Lemma 3.11 yields the desired result.
## 4. Applications to quasi-greedy bases in Hardy spaces
We are now in a position to apply the results of the previous section to the case when $\ensuremath{\mathbb{X}}$ is the Hardy space $H_p(\ensuremath{\mathbb{D}})$ for $0<p<1$. Our main result will be a ready consequence of the following theorem.
Theorem 4.1. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a truncation quasi- greedy basis $\ensuremath{\mathcal{X}}$ such that $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is doubling and
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
Suppose that $\ensuremath{\mathcal{Y}}$ is a complemented truncation quasi-greedy basic se- quence of $\ensuremath{\mathbb{X}}$. Then
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\lesssim \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \lesssim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\lesssim \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. Just combine Theorem 3.9 with Theorem 3.12.
Corollary 4.2. Let $\ensuremath{\mathcal{X}}$ be a squeeze-symmetric basis of a quasi-Banach space $\ensuremath{\mathbb{X}}$. Suppose that
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
Let $\ensuremath{\mathcal{Y}}$ is be a complemented truncation quasi-greedy basis of $\ensuremath{\mathbb{X}}$. Then $\ensuremath{\mathcal{Y}}$ is democratic, and
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx \ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \approx \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]\approx \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. Since $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]\approx \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$, $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is doubling, and
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
Thus, the result follows from Theorem 4.1.
Corollary Let $0<p<1$. If $\ensuremath{\mathcal{Y}}$ is a truncation quasi-greedy basis of a complemented subspace of $H_p(\ensuremath{\mathbb{D}})$ then $\ensuremath{\mathcal{Y}}$ is democratic with $\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}}, H_p(\ensuremath{\mathbb{D}})](m)\approx m^{1/p}$ for $m\in\ensuremath{\mathbb{N}}.$ In particular, all quasi-greedy bases of $H_p(\ensuremath{\mathbb{D}})$ are almost greedy.
| |
220809342/22
|
Proof. The basis $\ensuremath{\mathcal{H}}$ of $H_p(\ensuremath{\mathbb{D}})$ constructed in [40] is unconditional (hence truncation quasi-greedy) and democratic with
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{H}},H_p(\ensuremath{\mathbb{D}})](m) \approx m^{1/p}\approx \ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{H}},H_p(\ensuremath{\mathbb{D}})](m),\quad m\in \ensuremath{\mathbb{N}},
$$
(see [30]). Now our claim follows from Theorem 4.2.
We note that for $d>1$ the canonical unconditional basis $\ensuremath{\mathcal{H}}^d$ of the Hardy space $H_p(\ensuremath{\mathbb{D}}^d)$, constructed from the canonical unconditional basis $\ensuremath{\mathcal{H}}$ of $H_p(\ensuremath{\mathbb{D}})$ by means of tensor products, inherits the uncondi- tionality from $\ensuremath{\mathcal{H}}$, but not its democracy. Indeed, for every $d\in\ensuremath{\mathbb{N}}$ we have
$$
\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{H}}^d,H_p(\ensuremath{\mathbb{D}}^d)](m) \approx m^{1/p}, \quad m\in\ensuremath{\mathbb{N}},
$$
whereas
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{H}}^d,H_p(\ensuremath{\mathbb{D}}^d)](m) \approx h_m:=m^{1/p}(1+\log m)^{(d-1)(1/2-1/p)},\; m\in\ensuremath{\mathbb{N}},
$$
(see [38] * S4 and [42] * S4).
Theorem 4.4. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a truncation quasi- greedy basis $\ensuremath{\mathcal{X}}$. Suppose that $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is doubling with
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
If $\ensuremath{\mathcal{Y}}$ is a truncation quasi-greedy basis of $\ensuremath{\mathbb{X}}$, then
$$
\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \approx\ensuremath{\mathbf{\varphi_u}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Moreover, if $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]$ is doubling then
$$
\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}] \approx\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}].
$$
Proof. Apply Theorem 4.1 to obtain the estimates given there. Using Theorem with $\ensuremath{\mathbf{s}}=\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ and the roles of $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ swapped, we deduce the equivalence between the upper democracy functions. If $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]$ is doubling, we can apply Theorem with $\ensuremath{\mathbf{s}}=\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{Y}},\ensuremath{\mathbb{X}}]$ and the roles of $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ swapped. This gives the equivalence between the lower democracy functions.
Theorem 4.5. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach space with a truncation quasi- greedy basis $\ensuremath{\mathcal{X}}$. Suppose that $\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}]$ is doubling with
$$
\sum_{m=1}^\infty \frac{1}{\ensuremath{\mathbf{\varphi_l}}[\ensuremath{\mathcal{X}},\ensuremath{\mathbb{X}}](m)}<\infty.
$$
If $\ensuremath{\mathcal{X}}$ is not democratic then $\ensuremath{\mathbb{X}}$ has no squeeze-symmetric bases.
| |
220809342/1
|
# DEMOCRACY OF QUASI-GREEDY BASES IN $\mathbf p$ -BANACH SPACES WITH APPLICATIONS TO THE EFFICIENCY OF THE THRESHOLDING GREEDY ALGORITHM IN THE HARDY SPACES $\mathbf{H_{p}(\ensuremath{\mathbb{D}}^{d})}$
FERNANDO ALBIAC,
JOS ´ E L. ANSORENA, AND
GLENIER BELLO
Abstract. We use new methods, specific of non-locally convex quasi-Banach spaces, to investigate when the quasi-greedy bases of a $p$ -Banach space for $0<p<1$ are democratic. The novel tech- niques we obtain permit to show in particular that all quasi-greedy bases of the Hardy space $H_p(\ensuremath{\mathbb{D}})$ for $0<p<1$ are democratic while, in contrast, no quasi-greedy basis of $H_p(\ensuremath{\mathbb{D}}^d)$ for $d\geqslant 2$ is, solving thus a problem that was raised in [7]. Applications of our results to other spaces of interest both in functional analysis and approx- imation theory are also provided.
## 1. Introduction
The formal development of a theory of greedy bases was spurred at the turn of the 21st century by the work of Konyagin and Temlyakov on the efficiency of the Thresholding Greedy Algorithm (TGA for short) in Banach spaces [32]. The TGA simply takes $m$ terms with maximum absolute values of coefficients from the expansion of a signal (function) relative to a fixed representation system (a basis). Different greedy algorithms originate from different ways of choosing the coefficients of the linear combination in the $m$ -term approximation to the signal. Another name, commonly used in the literature for $m$ -term approxi- mation is sparse approximation. Sparse approximation of functions is a powerful analytic tool which is present in many important applications
Key words and phrases. quasi-greedy basis, democratic basis, quasi-Banach spaces, Hardy spaces.
Operators, lattices, and structure of Banach spaces An ´a lisis Vectorial, Multilineal y Aproximaci ´ n
The first-named author acknowledges the support of the Spanish Ministry for Science and Innovation under Grant PID2019-107701GB-I00 for. F. Albiac and J. L Ansorena were supported by the Spanish Ministry for Science, Innovation, and Universities under Grant PGC2018- 095366-B-I00 for o. G. Bello was sup- ported by National Science Centre, Poland grant UMO-2016/21/B/ST1/00241.
| |
220809342/3
|
quantitative study of the efficiency of the greedy algorithm in Hardy spaces $H_{p}$ and other important non-locally convex quasi-Banach spaces.
Both quasi-greedy and democratic bases were introduced by Konya- gin and Temlyakov back in 1999 in their seminal paper [32]. These two special types of bases can nowadays be regarded in hindsight as the pillars that sustain the entire theory of greedy approximation in Banach spaces using bases. Certainly, apart from the pioneering char- acterization of greedy bases as those bases that are simultaneously unconditional and democratic from [32] this claim is supported by the subsequent characterization of almost greedy bases in Banach spaces as those bases that are at the same time quasi-greedy and democratic [16].
Thus, although a priori, being almost greedy is more restrictive for a basis than being quasi-greedy, there exist spaces whose geometry forces quasi-greedy basis to be democratic. This is what happens for example with $c_0$, $\ell_2$, and $\ell_1$. The specific techniques used to prove that property in each space exhibit the critical structural aspect involved in making quasi-greedy basis be democratic. In $\ell_2$ the decisive ingredient is the fact that its Rademacher type and cotype are $2$ ([42]). In the space $\ell_1$ it is crucial that it is a GT-space [17], while for $c_0$ what matters is that its dual is a GT-space [15].
The situation is different in $\ell_{p}$ when $p\in(1,2)\cup(2,\infty)$. Indeed, these spaces have unconditional bases (in particular, quasi-greedy bases) that are not democratic: To see this, one just has to remember that when $p\in(1,2)\cup(2,\infty)$, the space $\ell_{p}$ is isomorphic to $(\bigoplus_{n=1}^\infty\ell_2^n)_{\ell_p}$, whose canonical basis is unconditional but not democratic.
The study of greedy-like bases in non-locally convex spaces sprang from the paper [5], where it is proved that the aforementioned charac- terizations of greedy bases and almost greedy bases remain valid in this more general framework. In this spirit, the authors of [7] ventured out beyond the “ psychological” border of the index $p=1$ and proved that all quasi-greedy bases in the spaces $\ell_{p}$ for $0<p<1$ are democratic. The non-locally convex nature of these spaces required the introduction of new techniques in order to determine how the geometry of the space shapes the structure of their quasi-greedy bases. The results obtained in [7] heavily rely on the theory of $\ensuremath{\mathscr{L}}_p$ -spaces for $0<p<1$ developed by Kalton in [29].
In this paper we investigate the democracy of quasi-greedy bases in other classical non-locally convex quasi-Banach spaces. Our approach permits to obtain, for instance, that all quasi-greedy bases of the Hardy space $H_p(\ensuremath{\mathbb{D}})$ for $0<p<1$ (as well as all quasi-greedy bases of its complemented subspaces) are democratic, solving thus in the positive
| |
220809342/25
|
$p<1$,
we apply Theorem 4.5 to obtain that no space from the list in (5.1) has an almost greedy basis. This partially solves [5] * Problem 13.8.
| |
220809342/9
|
This way, the quasi-Banach lattice $\ensuremath{\mathbb{X}}$ is $r$ -convex with $M^{(r)}(\ensuremath{\mathbb{X}})\leqslant C<\infty$ if and only if
$$
\left\lVert \left(\Ave_{i\in A} (\left\lvert f_i\right\rvert^r\right)^{1/r}\right\rVert \leqslant C \left(\Ave_{i\in A} (\left\lVert f_i\right\rVert^r\right)^{1/r}, \quad \left\lvert A\right\rvert<\infty,\; f_i\in\ensuremath{\mathbb{X}}.
$$
Defining $r$ -sums and $r$ -averages in quasi-Banach lattices allows us to state a lattice-valued version of Khintchine’s inequalities:
Theorem 2.2. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach lattice. For each $0<r<\infty$ there are constants $T_r$ and $C_r$ such that for any finite family $(f_i)_{i\in A}$ in $\ensuremath{\mathbb{X}}$,
\begin{equation*} \frac{1}{C_r} \left(\Ave_{\varepsilon_i=\pm 1} \left\lvert \sum_{i\in A} \varepsilon_i\, f_i\right\rvert^r\right)^{1/r} \leqslant \left(\sum_{i\in A} \left\lvert f_i\right\rvert^2\right)^{1/2} \leqslant T_r \left(\Ave_{\varepsilon_i=\pm 1} \left\lvert \sum_{i\in A} \varepsilon_i\, f_i\right\rvert^r\right)^{1/r}. \end{equation*}
Proof. Just apply the functional calculus $\tau$ described in [33] * Theorem 1.d.1 to the functions $f$, $g\colon\ensuremath{\mathbb{R}}^n \rightarrow \ensuremath{\mathbb{R}}$ given by
$$
f((x_i)_{i=1}^n)=\left(\sum_{i=1}^n \left\lvert x_i\right\rvert^2\right)^{1/2},
\quad g((x_i)_{i=1}^n)= \left(\Ave_{\varepsilon_i=\pm 1} \left\lvert \sum_{i=1}^n \varepsilon_i\, x_i\right\rvert^{r}\right)^{1/r},
$$
and use Khintchine’s inequalities [31].
If $\ensuremath{\mathbb{X}}$ is a $r$ -convex quasi-Banach lattice, then $\ensuremath{\mathbb{X}}$ is a $\min\{r,1\}$ -convex quasi-Banach space. The converse does not hold, to the extent that there are quasi-Banach lattices that are not $r$ -convex for any $r>0$ (see [28] * Example 2.4). Theorem 2.2 from the same paper characterizes quasi-Banach lattices with some nontrivial convexity as those that are $L$ -convex. A quasi-Banach lattice $\ensuremath{\mathbb{X}}$ is said to be $L$ -convex if there is $\varepsilon>0$ so that whenever $f$ and $(f_i)_{i\in A}$ in $\ensuremath{\mathbb{X}}$ satisfy $0\leqslant f_i\leqslant f$ for every $i\in A$, and $(1-\varepsilon)\left\lvert A\right\rvert f\leqslant \sum_{i\in A} f_i$ we have $\varepsilon \left\lVert f \right\rVert\leqslant \max_{i\in A} \left\lVert f_i\right\rVert$.
The aforementioned Theorem 2.2 from [28] also gives the following.
Theorem 2.3. Let $\ensuremath{\mathbb{X}}$ be a quasi-Banach lattice and let $0<r<\infty$. If $\ensuremath{\mathbb{X}}$ is lattice $r$ -convex, then it is lattice $s$ -convex for every $0<s<r$.
Quoting Kalton from [28], $L$ -convex lattices behave similarly to Ba- nach lattices in many respects. The following result, which generalizes to quasi-Banach lattices [34] * Lemme 5, is in this spirit.
| |
220809342/8
|
$\ensuremath{\mathcal{X}}$ is
- democratic if and only $\ensuremath{\mathbf{\varphi_u}}(m)\lesssim\ensuremath{\mathbf{\varphi_l}}(m)$ for $m\in\ensuremath{\mathbb{N}}$, in which case it is super-democratic, i.e., there is a constant $D\geqslant 1$ such that
\begin{equation*} \left\lVert \ensuremath{\mathbbm{1}}_{\varepsilon,A}\right\rVert \leqslant D \left\lVert \ensuremath{\mathbbm{1}}_{\delta,B}\right\rVert \end{equation*}
- for any two finite subsets $A$, $B$ of $\ensuremath{\mathbb{N}}$ with $\left\lvert A\right\rvert\leqslant\left\lvert B\right\rvert$, any $\varepsilon\in\ensuremath{\mathcal{E}}^A$ and any $\delta\in\ensuremath{\mathcal{E}}^B$.
Here and throughout this paper, the symbol $\alpha_j\lesssim \beta_j$ for $j\in \ensuremath{\mathcal{N}}$ means that there is a positive constant $C$ such that the families of non-negative real numbers $(\alpha_j)_{j\in \ensuremath{\mathcal{N}}}$ and $(\beta_j)_{j\in \ensuremath{\mathcal{N}}}$ are related by the inequality $\alpha_j\leqslant C\beta_j$ for all $j\in \ensuremath{\mathcal{N}}$. If $\alpha_j\lesssim \beta_j$ and $\beta_j\lesssim \alpha_j$ for $j\in \ensuremath{\mathcal{N}}$ we say $(\alpha_j)_{j\in \ensuremath{\mathcal{N}}}$ are $(\beta_j)_{j\in \ensuremath{\mathcal{N}}}$ are equivalent, and we write $\alpha_j\approx \beta_j$ for $j\in \ensuremath{\mathcal{N}}$.
2.5.
## Quasi-Banach lattices
. Let $0<r\leqslant\infty$. A quasi-Banach space $\ensuremath{\mathbb{X}}$ is (topologically) $r$ -convex or $r$ -normable if there is a constant $C\geqslant 1$ such that
$$
\left\lVert \sum_{i\in A} f_i\right\rVert\leqslant C
\left(\sum_{i\in A}\left\lVert f_i\right\rVert^r\right)^{1/r}
,\quad \left\lvert A\right\rvert<\infty,\, f_i\in\ensuremath{\mathbb{X}}.
$$
If a quasi-Banach space is $r$ -convex then $r\leqslant 1$. Conversely, by the Aoki-Rolewicz theorem any quasi-Banach space is $r$ -convex for some $r\in(0,1]$. In turn, any $r$ -convex quasi-Banach space $\ensuremath{\mathbb{X}}$ becomes an $r$ - Banach space under a suitable renorming, i.e., $\ensuremath{\mathbb{X}}$ can be endowed with an equivalent quasi-norm satisfying (2.5) with $C=1$.
The existence of a lattice structure in $\ensuremath{\mathbb{X}}$ leads to a (related but) different notion of convexity. A quasi-Banach lattice $\ensuremath{\mathbb{X}}$ is said to be $r$ -convex ($0<r<\infty$) if there is a constant $C$ such that
$$
\left\lVert \left(\sum_{i\in A}\left\lvert f_i\right\rvert^r\right)^{1/r}\right\rVert\leqslant C \left(\sum_{i\in A}\left\lVert f_i\right\rVert^r\right)^{1/r}, \quad \left\lvert A\right\rvert<\infty,\; f_i\in\ensuremath{\mathbb{X}}.
$$
where the lattice $r$ -sum $\left(\sum_{i\in A}\left\lvert f_i\right\rvert^r\right)^{1/r}\in\ensuremath{\mathbb{X}}$ is defined unambiguously exactly as for the case of Banach lattices (cf. [33] * pp. 40–41 and [36]). If $\ensuremath{\mathbb{X}}$ is an $r$ -convex quasi-Banach lattice, we will denote by $M^{(r)}(\ensuremath{\mathbb{X}})$ the smallest constant $C$ such that (2.6) holds.
We will also consider lattice averages and use them to reformulate lattice convexity in those terms. Given $0<r\leqslant \infty$, the $r$ -average of a finite family $(f_i)_{i\in A}$ in a quasi-Banach lattice $\ensuremath{\mathbb{X}}$ is defined as
$$
\left(\Ave_{i\in A} \left\lvert f_i\right\rvert^r\right)^{1/r}=\left(\frac{1}{\left\lvert A\right\rvert}\sum_{i\in A} \left\lvert f_i\right\rvert^r \right)^{1/r}=\left\lvert A\right\rvert^{-1/r} \left(\sum_{i\in A}\left\lvert f_i\right\rvert^r\right)^{1/r}.
$$
| |
220602985/5
|
for GEBD, which regards each video clip as a whole. Specifically, given a video clip of arbitrary length, we first use conventional CNN backbone to extract the 2D feature representation for each frame and get the frame sequence, i.e., $V = \{I_t\}_{t=1}^T$, where $I_t \in \mathbb{R}^C$ and $T$ is the length of the video clip. Then the structured partition of sequence (SPoS) mechanism is employed to re-partition input frame sequence $\{I_t\}_{t=1}^T$ and provide structured context for each candi- date frame. The Transformer encoder blocks [?] are then used to learn the high level representation of each local sequence. After that, we compute the group similarities to capture temporal changes and a following lightweight fully convo- lutional network [?] (FCN) is used to recognize different patterns of the grouped 2D similarity maps. We will introduce the details of each module in the following sections. The overall architecture of proposed method is presented in Figure 1.
## Structured Context Transformer
The existence of an event boundary in a video clip implies that there is a visual content change at that point, thus it is very difficult to infer the boundary from one single frame. As a result, the key clue for event boundary detection is to localize changes in the temporal domain. Modeling in temporal domain has long been explored by different approaches, including LSTM [?], Transformer [?], 3D Convolutional Neural Network [?], etc. Transformer [?] has recently demonstrated promising results on both natural language processing (NLP) tasks and computer vision tasks. Despite its success, it is difficult to apply Transformer directly to the GEBD task due to its quadratic computational complexity of self-attention. The computation cost and memory consumption increase dramatically as the length of video increases. Previous methods [?,?] regard each individual frame as one sample and its nearby frames are fed into network together to provide temporal information for this frame. This method introduces redundant computation in adjacent frames since each frame is fed into the network as input for multiple times. In this paper, we seek to explore a more general and efficient temporal representation for GEBD task.
Structured Given the video snippet $V = \{I_t\}_{t=1}^T$, where $T$ is the time span of the video snippet and can be any arbitrary length, $I_t \in \mathbb{R}^C$ is the feature vector of frame $t$, which is generated from ResNet50 [?] backbone followed by a global average pooling layer, our goal is to obtain $K$ ad- jacent frames before candidate frame $I_t$ and $K$ adjacent frames after candidate frame $I_t$, where $K$ is the adjacent window size. We term this local sequence cen- tred with candidate frame $I_t$ as structured context for frame $I_t$. To accomplish this while enabling feature sharing and maintaining efficiency and parallelism, we propose the novel Structured Partition of Sequence (SPoS) mechanism. Specifi- cally, we first pad video $V = \{I_t\}_{t=1}^T$ with $\operatorname{ceil}(\frac{T}{K}) \cdot K - T$ zero vectors at the end of the frame sequence so that the new video length $T'$ is divisible by $K$. Then given the padded video $V' \in \mathbb{R}^{T' \times C}$, we split it into $K$ slices where each slice $S_k$ ($k$ is the slice number, starts from $0$) is responsible to provide structured context frames for all $[k::K]$ th frames (i.e., all frames that start from $k$ with a step of
| |
220602985/7
|
$I_{\leftarrow t}$ and
$I_{t\rightarrow}$ in a one-to-one manner, which is the key for accurate boundary detection. Our SPoS differs from Swin-Transformer [?] in that Swin-Transformer makes each frame able to attend very distant frames (i.e., tend to learn a global representation) due to its stacked shifted windows design. This is deleterious for boundary detection as very distant frames may cross multiple boundaries and thus provides less useful information. Another advantage of SPoS is that we can model these structured sequences using any sequential modeling method without considering computation complexity, due to its local shared and parallel nature, which can be computed in linear time to video length.
Encoding with Transformer. We use Transformer to model the structured context information. Given structured context features $I_{\leftarrow t} \in \mathbb{R}^{K \times C}, I_{t\rightarrow} \in \mathbb{R}^{K \times C}$ of frame $I_t \in \mathbb{R}^C$, we first concatenate them in the temporal dimension to obtain context sequence $\mathbf{I}_t$ for frame $I_t$, i.e.,
$$
\mathbf{I}_t = [I_{\leftarrow t}, I_t, I_{t\rightarrow}]
$$
where $\mathbf{I}_t \in \mathbb{R}^{L \times C}$, $L=2K+1$ and $[\cdot, \cdot, \cdot]$ denotes the concatenating operation. Then to model temporal information, we adapt a 6-layer Transformer [?] block to processing the context sequence $\mathbf{I}_t$ to get temporal representation $\mathbf{x}_t \in \mathbb{R}^{L \times C}$ within this structured context window. Unlike other methods [?,?] where the computation of multi-head self attention (MSA) is based on global video frames sequence, our MSA computation is based only on the local temporal window. The computational complexity of the former is quadratic to video length $T$, i.e., $4TC^2 + 2T^2C$, and the computational complexity of our method is linear when $K$ is fixed (set to 8 by default, i.e., $L=17$), i.e., $4TC^2 + 2L^2TC$. Global self- attention computation is generally unaffordable for a large video length $T$, while our local structured based self-attention is scalable.
## Group Similarity
The event boundaries of the GEBD task could be located at the moments where the action changes (e.g., Run to Jump), the subject changes (e.g., a new per- son appears), or the environment changes (e.g., suddenly become bright), for example. We experimentally observed that the frames within an adjacent local window provide more cues for event boundary detection than distant frames. This is consistent with human’s intuition since the change of visual content can be regarded as an event boundary only in a short time period. Based on this observation, we can model local temporal information naturally based on struc- tured context features extracted in Section 3.1.
The Transformer block aims at discovering relationships between frames and giving high level representation of frames sequence. However, event boundaries emphasize the differences between adjacent frames and neural networks tend to take shortcuts during learning [?]. Thus classifying these frames directly into boundaries may lead to inferior performance due to non-explicit cues. Based on this intuition, we propose to guide classification with feature similarity of each frame pair in the structured temporal window $\mathbf{x}_t \in \mathbb{R}^{L \times C}$. Instead of performing
| |
220602985/14
|
Table 5: Effect of model width $C$.
\begin{tabular}{c|cccc}
\rightarrowprule
$C$ & 0.05 &0.25 & 0.5 & Average \midrule
$128$&0.775 &\textbf{0.895} &\textbf{0.913} &\textbf{0.881} $256$&\textbf{0.777} &\textbf{0.895} &0.911 & \textbf{0.881} $512$&0.774 &0.892 &0.908 & 0.879 $768$&0.768 &0.887 &0.904 &0.875 $1024$&0.770 &0.889 &0.905 &0.876 \bottomrule
\end{tabular}
Table 6: Effect of number of groups $G$.
\begin{tabular}{c|cccc}
\rightarrowprule
$G$ & 0.05 &0.25 & 0.5 & Average \midrule
$1$&0.761 &0.871 &0.887 &0.861 $2$&0.769 &0.891 &0.907 & 0.877 $4$&0.777 &0.895 &0.911 & 0.881 $8$&\textbf{0.778} &\textbf{0.896} &\textbf{0.913} &\textbf{0.882} $16$&0.777 &\textbf{0.896} &0.912 &0.881 \bottomrule
\end{tabular}
our method is effective to different metrics, and we use cosine metric in our experiments.
Table 7: Effect of $\texttt{similarity-function}(\cdot, \cdot)$ in Equation 3.
\begin{tabular}{l|cccc}
\rightarrowprule
Function & 0.05 &0.25 & 0.5 & Average \midrule
Chebyshev&0.770 &0.887 &0.905 &0.872 Manhattan&0.774 &0.894 &0.907 &0.878 Euclidean&0.776 &\textbf{0.895} &0.910 &\textbf{0.881} Cosine &\textbf{0.777} &\textbf{0.895} &\textbf{0.911} &\textbf{0.881} \bottomrule
\end{tabular}
Loss ablations. GEBD task can be regarded as a framewise binary classifica- tion (boundary or not) after capturing temporal context information. We train our model with binary cross entropy (BCE) loss and mean squared error (MSE) loss with turning Gaussian smoothing (introduced in section 3.3) on and off. As shown in Table 8, Gaussian smoothing can improve the performance in both settings, which shows its effectiveness. We attribute this improvement to two aspects: 1) Consecutive frames have similar feature representation in the latent space thus consecutive frames are always tend to output closer responses, hard labels violate this rule and lead to poor convergence. 2) Annotations of GEBD are ambiguous in nature and Gaussian smoothing prevents the network from be- coming overconfident. We use “ BCE + Gaussian ” setting in all our experiments.
Table 8: Effect of loss function.
\begin{tabular}{ccc|cccc}
\rightarrowprule
BCE & MSE &Gaussian & 0.05 &0.25 & 0.5 & Average \midrule
&\checkmark& &0.758 &0.881 &0.899 &0.865 &\checkmark&\checkmark &0.771 &0.893 &0.910 &0.877 \checkmark &&&0.763 &0.887 &0.905 &0.872 \checkmark &&\checkmark &\textbf{0.777} &\textbf{0.895} &\textbf{0.911} &\textbf{0.881} \bottomrule
\end{tabular}
## Conclusions
| |
220602985/2
|
Fig.1: Overview architecture of the proposed method. The proposed method can predict all boundaries of video sequences in a single forward pass with high efficiency. We use a CNN backbone to extract the 2D features of each video frame. These features are then pooled into the vectors and converted into a se- quence. The structured partition of sequence (SPoS) mechanism is employed to re-partition input frame sequence and provides structured context for each candidate frame. Based on this structured context, the transformer encoder blocks are used to learn the high level representations of each local sequence, which have linear computational complexity with respect to the video length and enable feature sharing. After that, we compute the group similarities to en- code frame differences and use a lightweight fully convolutional network (FCN) is predict event boundaries based on the computed 2D grouped similarity maps.
employed for each video to label the boundaries based on predefined principles. These characteristics differentiate GEBD from the previous video localization tasks [?] by several high-level causes, for example, 1) Change of Subject, i.e., new subject appears or old subject disappears, 2) Change of Action, i.e., an old action ends, or a new action starts, 3) Change in Environment, i.e., significant changes in color or brightness of the environment, 4) Change of Object of Inter- action, i.e., the subject starts to interact with a new object or finishes with an old object. The aforementioned factors make GEBD to be a more challenging task compared to video localization.
Solving GEBD task is not trivial since detecting event boundaries highly rely on temporal context information. Existing methods tackle this problem by processing each frame individually [?,?,?] or computing the global self-similarity matrix and using extra parsing algorithm to find boundary patterns based on the self-similarity matrix [?,?]. The methods in the first category introduce sub- stantial redundant computations of adjacent frames in a video sequence when predicting boundaries and have to solve the class imbalance issue of event bound- aries. The methods in the second category have quadratic computation complex-
| |
220602985/11
|
Fig.4: Example qualitative results on Kinetics-GEBD validation split. Compared with PC [?], our SC-Transformer can generate more accurate boundaries which are consistent with ground truth.
## Main Results
Kinetics-GEBD. Table 1 illustrates the results of our models on Kinetics- GEBD validation set. Our method surpasses all previous methods in all Rel.Dis. threshold settings, demonstrating the effectiveness of structured partition of sequence and group similarity. Compared to the method PC [?], our method achieves 15.2% absolute improvement with $5.7\times$ faster running speed (i.e., 10.8ms per frame vs 1.9ms per frame). Compared to DDM-Net [?], we also achieve 1.3% absolute improvement. Since DDM-Net is not open sourced yet we are not able to compare runtime speed with our method. However it is worth noting that DDM-Net leverage the same input representation as PC [?], i.e., each frame and it’s adjacent frames are fed into network individually, which introducing many redundant computations. For example, given a video clip of length 100 and the window is set to 11 as mentioned in their paper, they have to process $1,100$ frames individually to get all boundary predictions for this single video. Our method is independent of video length and can get all boundary predictions in a single forward pass by just feeding the necessary 100 frames. The example qualitative results on Kinetics-GEBD are shown in Figure 4.
TAPOS. We also conduct experiments on TAPOS dataset [?]. TAPOS dataset contains Olympics sport videos with 21 actions and is not suitable for GEBD task. Following [?], we re-purpose TAPOS for GEBD task by trimming each
| |
220602985/6
|
Fig.2: Illustration of proposed structured partition of sequence (SPoS). To obtain adjacent $K$ frames before candidate frame $I_t$ (denoted as $I_{\leftarrow t}$) and $K$ frames after $I_t$ (denoted as $I_{t \rightarrow}$), we split the input video sequence into $K$ slices. Each slice $S_k$ is responsible to produce adjacent frames $I_{\leftarrow t}$ and $I_{t \rightarrow}$ for the frames of specific indices (i.e., all frames that start from $k$ with a step of $K$). All video frames can be covered within all $K$ slices and can be efficiently processed in parallel. Our SPoS differs from Swin-Transformer [?] and 1D CNNs in that Swin-Transformer tends to learn a global representation after several stacks and is not structured and 1D CNNs actually make candidate frame $I_t$ attend to adjacent frames in a Gaussian distribution manner [?], respectively.
$K$).
In this way, all video frames can be covered within all $K$ slices and these $K$ slices can be efficiently processed in parallel.
In each frame slice $S_k$, we obtain structured context for frame $I_t$ in two directions, i.e., $K$ frames before frame $I_t$ and $K$ frames after frame $I_t$. We implement this through efficient memory view method provided by modern deep learning framework. Specifically, To obtain structured context frames $I_{\leftarrow t} \in \mathbb{R}^{K \times C}$ before frame $I_t$, we replicate the first frame of the padded video sequence $V'$
$K-k$ times
and concatenate to the beginning of video $V'$ and the last $K-k$ frames of video $V'$ are dropped, and thus the number of frames is kept and still divisible by $K$. We denote this shifted video sequence as $V'_\leftarrow \in \mathbb{R}^{T' \times C}$. Then we view $V'_\leftarrow \in \mathbb{R}^{T' \times C}$ as $V_\leftarrow \in \mathbb{R}^{N \times K \times C}$, where $N = T' / K$ denotes the number of processed frames in slice $S_k$. In this way, we obtain the left structured context frames for all $N$ frames (i.e., all $[k::K]$ th frames of origin video $V$). Similarly, to obtain structured context frames $I_{t\rightarrow} \in \mathbb{R}^{K \times C}$ after frame $I_t$, we replicate the last frame of the padded video sequence $k+1$ times and concatenate to the ending of video $V'$ and the first $k+1$ frames of video $V'$ are also dropped to keep the number of frames. In this way, we obtain the right structured context frames $V_\rightarrow \in \mathbb{R}^{N \times K \times C}$ for all $N$ frames. Finally, we can obtain all temporal context frames by repeating $K$ times for $K$ slices, and each frame $I_t$ is represented by its adjacent frames in a local window.
A key design element of our structured partition of sequence (SPoS) is its shared structured context information. We term this context information “ struc- tured ” since SPoS maps each candidate frame $I_t$ to individual frame sequences
| |
220602985/4
|
## Related Works
Generic Event Boundary Detection (GEBD). The goal of GEBD [?] is to localize the taxonomy-free event boundaries that break a long event into several short temporal segments. Different from TAL, GEBD only requires to predict the boundaries of each continuous segments. The current methods [?,?,?] all follow the similar fashion in [?], which takes a fixed length of video frames before and after the candidate frame as input, and separately determines whether each candidate frame is the event boundary or not. Kang et al. [?] propose to use the temporal self-similarity matrix (TSM) as the intermediate representation and use the popular contrastive learning method to exploit the discriminative features for better performance. Hong et al. [?] use the cascade classification heads and dynamic sampling strategy to boost both recall and precision. Rai et al. [?] attempt to learn the spatiotemporal features using a two stream inflated 3D convolutions architecture.
Temporal Action Localization (TAL). TAL aims to localize the action segments from untrimmed videos. More specifically, for each action segment, the goal is to detect the start point, the end point and the action class it be- longs to. Most approaches could be categorised into two groups, A two-stage method [?,?,?,?,?] and a single-stage method [?,?,?,?,?,?,?,?]. In a two-stage method setting, the first stage generates action segment proposals. The actionness and the type of action for each proposal are then determined by the second stage, along with some post-processing methods such as grouping [?] and Non-maximum Suppression (NMS) [?] to eliminate redundant proposals. For one-stage meth- ods, the classification is performed on the pre-defined anchors [?,?] or video frames [?,?]. Even though TAL task has some similarity to GEBD task, there is no straightforward way to directly apply these methods on the GEBD dataset. Since GEBD requires event boundaries to be taxonomy-free and continuous, which is different from the TAL settings.
Transformers. Transformer [?] is a prominent deep learning model that has achieved superior performance in various fields, such as natural language pro- cessing (NLP) and computer vision (CV). Despite it’s success, the computational complexity of its self-attention is quadratic to image size and hard to applied to high-resolution images. To address this issue, [?] proposes a hierarchical Trans- former whose representation is computed with shifted windows and has linear computational complexity with respect to image size. In this paper, we show that these Transformer variants are not suitable for GEBD.
## Method
The existing methods [?,?,?] formulates the GEBD task as binary classification, which predict the boundary label of each frame by considering the temporal context information. However, it is inefficient because the redundant computa- tion is conducted while generating the representations of consecutive frames. To remedy this, we propose an end-to-end efficient and straightforward method
| |
220602985/15
|
In this work, we presented SC-Transformer which is a fully end-to-end method for generic event boundary detection. Structured partition of sequence mecha- nism is proposed to provide structured context information for GEBD task and Transformer encoder is adapted to learn high-level representation. Then group similarity and FCN are used to exploit discriminative features to make accurate predictions. Gaussian kernel is used to preprocess the ground-truth annotations to speed up training process. The proposed method achieves start-of-the-art re- sults on the challenging Kinetics-GEBD and TAPOS datasets with high running speed. We hope our method can inspire future work.
| |
220602985/1
|
# Structured Context Transformer for Generic Event Boundary Detection
Congcong Li, Xinyao Wang, Dexiang Hong, Yufei Wang, Libo Zhang, Tiejian Luo, Longyin Wen
Abstract. Generic Event Boundary Detection (GEBD) aims to detect moments where humans naturally perceive as event boundaries. In this paper, we present Structured Context Transformer (or SC-Transformer) to solve the GEBD task, which can be trained in an end-to-end fashion. Specifically, we use the backbone convolutional neural network (CNN) to extract the features of each video frame. To capture temporal context information of each frame, we design the structure context transformer (SC-Transformer) by re-partitioning input frame sequence. Note that, the overall computation complexity of SC-Transformer is linear to the video length. After that, the group similarities are computed to capture the differences between frames. Then, a lightweight fully convolutional network is used to determine the event boundaries based on the grouped similarity maps. To remedy the ambiguities of boundary annotations, the Gaussian kernel is adopted to preprocess the ground-truth event bound- aries to further boost the accuracy. Extensive experiments conducted on the challenging Kinetics-GEBD and TAPOS datasets demonstrate the effectiveness of the proposed method compared to the state-of-the-art methods. Keywords: GEBD, Structured Context Transformer, Group Similarity
## Introduction
Video has accounted for a large part of human’s life in recent years. Aided by the rapid developments of hardware, video understanding has witnessed an explosion of new designed architectures [?,?,?,?,?,?] and datasets [?,?,?,?,?]. The cognitive science [?] suggests that humans naturally divide video into meaningful units. To enable machines to develop such ability, Generic Event Boundary Detection [?] (GEBD) is recently proposed which aims at localizing the moments where humans naturally perceive event boundaries.
Event boundaries in the GEBD task are taxonomy-free in nature and can be seen as a new attempt to interconnect human perception mechanisms to video understanding. Annotators are required to localize boundaries at “ one level deeper ” granularity compared to the video-level event. To remedy the ambigui- ties of event boundaries based on human perception, five different annotators are
| |
220602985/3
|
ity regarding to the length of input videos due to computation of self-attention globally and the usage of the extra parsing algorithm to predict boundaries.
To that end, we proposed an end-to-end method to predict all boundaries of video sequences in a single forward pass of the network with high efficiency. The overall architecture of the proposed method is shown in Figure 1. Specifically, the Structured Context Transformer (SC-Transformer) is designed for GEBD based on the designed structured partition of sequence (SPoS) mechanism, which has linear computational complexity with respect to input video length and enables feature sharing by design. Structured partition of sequence (SPoS) mechanism brings the local feature sequences for each frame in a one-to-one manner, which is termed as structured context. We also find that 1D CNNs actually make the candidate frames attend to adjacent frames in a Gaussian distribution manner [?], which is not optimal for boundary detection as adjacent frames are equally important. Our proposed SC-Transformer can learn a high level representation for each frame within its structured context which is critical for boundary de- tection. After that, we use the group similarity to exploit discriminative features to encode the differences between frames. The concept of groups as a dimen- sion for model design has been widely studied, including Group Convolutions [?,?], Group Normalization [?], Multi-head self attention [?], etc. However, to the best of our knowledge, there is still no study on the grouped similarity learning. Previous methods [?,?,?] actually compute similarity matrix on one dimension group. Our proposed group similarity allows the network to learn a varied set of similarities and we find it is effective for GEBD. Following the group similarity, a lightweight fully convolutional network [?] (FCN) is used to predict event boundaries. Note that, to speed up the training phase, the Gaussian kernel is used to preprocess the ground-truth event boundaries. Extensive experiments conducted on two challenging Kinetics-GEBD and TAPOS datasets demonstrate the effectiveness of the proposed method compared to the state-of-the-art meth- ods. Specifically, compared to DDM-Net [?], our method produces 1.3% absolute improvement. Meanwhile, compared to PC [?], our method achieves 15.2% ab- solute improvement with $5.7\times$ faster running speed. We also conduct several ablation studies to analyze the effectiveness of different components in the pro- posed method. We hope the proposed method can inspire future work.
The main contributions of this paper are summarized as follows. (1) We propose the structured context transformer for GEBD, which can be trained in an end-to-end fashion. (2) To capture differences between frames, we compute the group similarities to exploit the discriminative features to encode the differences between frames and use a lightweight FCN to predict the event boundaries. (3) Several experiments conducted on two challenging Kinetics-GEBD and TAPOS datasets demonstrate the effectiveness of the proposed method compared to the state-of-the-art methods.
| |
220602985/9
|
this vector is used for downstream classification:
$$
\begin{aligned}
\begin{array}{c}
s_t = \operatorname{FCN}(\mathbf{S}_t) \\
h_t = \texttt{average-pool}(s_t)
\end{array}
\end{aligned}
$$
where $s_t \in \mathbb{R}^{C \times L \times L}$ indicates the intermediate representation, $h_t \in \mathbb{R}^C$. The design principle of this module is extremely simple: computing group similarity patterns within local structured context based on previously encoded and using a small FCN to analyse the patterns.
## Optimization
Our SC-Transformer and group similarity module are fully end-to-end, lightweight and in-place i.e. no dimension change between input and output. Therefore they can be directly used for further classification which is straightforward to imple- ment and optimize. The video frame sequence $V = \{I_t\}_{t=1}^T$ is represented by $\mathbf{V} = \{h_t\}_{t=1}^T$ after group similarity module, i.e., $\mathbf{V} \in \mathbb{R}^{T \times C}$. Then we stack 3 layers of 1D convolutional neural network to predict boundary scores. We use a single binary cross entropy loss to optimize our network.
GEBD is a taxonomy-free task and interconnects the mechanism of human perception to deep video understanding. The event boundary labels of each video are annotated by around 5 different annotators to capture human perception dif- ferences and therefore ensure diversity. However, this brings ambiguity of anno- tations and is hard for network to optimize, which may lead to poor convergence. To solve this issue and prevent the model from predicting the event boundaries too confidently, we use the Gaussion distribution to smooth the ground-truth boundary labels and obtain the soft labels instead of using the “ hard labels ” of boundaries. Specifically, for each annotated boundary, the intermediate label of the neighboring position $t'$ is computed as:
$$
\mathcal{L}_{t'}^t = \exp\Big(-\frac{(t-t')^2}{2\sigma^2} \Big)
$$
where $\mathcal{L}_{t'}^t$ indicates the intermediate label at time $t'$ corresponding to the anno- tated boundaries at time $t$. We set $\sigma =1$ in all our experiments. The final soft labels are computed as the summation of all intermediate labels. Finally, binary cross entropy loss is used to minimize the difference between model predictions and the soft labels.
## Experiments
We show that our method achieves competitive results compared to previous methods in quantitative evaluation on Kinetics-GEBD [?] and TAPOS [?]. Then, we provide a detailed ablation study of different model design with insights and quantitative results.
Dataset. We perform experiments on both Kinetics-GEBD dataset [?] and TAPOS dataset [?]. Kinetics-GEBD dataset consists of $54,691$ videos and $1,290,000$
| |
220602985/8
|
Fig.3: Visualization of grouped similarity maps $\mathbf{S}_t$, $G=4$ in this example. First row indicates that there is a potential boundary in this local sequence while the second row shows no boundary in this sequence. We can also observe slightly different patterns between the same group, which may imply that each group is learning in a different aspect.
similarity calculation with all $C$ -dimensional channels, we found it beneficial to split the channels into several groups and calculate the similarity of each group independently. The concept of groups as a dimension for model design has been more widely studied, including Group Convolutions [?,?], Group Normalization [?], Multi-Head Self Attention [?], etc. However, to the best of our knowledge, there is still no study on similarity learning with grouping. Formally, given $\mathbf{x}_t \in \mathbb{R}^{L \times C}$, we first split into G groups:
$$
\mathbf{x'}_t = \texttt{reshape}(\mathbf{x}_t)
$$
where $\mathbf{x'}_t \in \mathbb{R}^{L \times G \times C'}$ and $C' = C / G$. Then the group similarity map $\mathbf{S}_t$ is calculated using the grouped feature:
$$
\mathbf{S}_t = \texttt{similarity-function}(\mathbf{x'}_t, \mathbf{x'}_t)
$$
where $\mathbf{S}_t \in \mathbb{R}^{G \times L \times L}$, and $\texttt{similarity-function}(\cdot, \cdot)$ can be cosine similarity or euclidean similarity. As the group similarity map $\mathbf{S}_t$ contains the similarity patterns (efficient score of each frame pair, i.e., high response value when two frames are visually similar), it shows different patterns (as shown in Figure 3) in different sequences, which are critical for boundary detection. To keep our model as simple as possible, we use a 4-layer fully convolutional network [?] to learn the similarity patterns, which we found it work very well and efficient enough. Then we average pool the output of FCN to get a vector representation $h_t$, and
| |
230409012/5
|
Figure 3. Heuristic process of modeling GUI-AG (c) based on given layout (b) and associated screenshot (a). All nodes represent components and all arrows represent possible relationships. We will keep only one inside relation among all children from a component (in blue) randomly, while the others are not used as input (in gray). The other relations are only possible between components that are inside the same parent and we keep only one relation between them.
GUI-AG from a layout by parsing a layout description. First, to reduce the size of the GUI-AG in order to have a smaller input in the transformer, we keep randomly only one inside relation within a group of components that are children to the same parent. This step is beneficial due to memory limitations when the input is too large. To illustrate this step, in Figure 3 we notice that CONTAINER[2] has two children, DATE PICKER[3] and NAVIGATION[6], but only one inside relationship is kept as input to the model and the other one is removed. This processing method does not lose information since all components that do not have an inside relationship are considered at the same level and are implicitly inside the parent component. We also randomly choose a sequence of components inside the layout and determine relationship between each pair of components within the sequence. More precisely, as shown in Figure 3, DATE PICKER[3] and NAVIGATION[6] are inside the same parent; in this situation, we randomly choose the sequence in the set of children [NAVIGATION[6], DATE PICKER[3]] and then add the relation between each component of the sequence to form triplets ([NAVIGATION[6], below, DATE PICKER[3]] in this case). By doing so, we get a simplified input that captures most information from the graph, as we will see in subsection 3.2. It is to note however, that we may miss some relationships. For example, if there are three components (a,b,c) and we keep only two relationships (a-b, b-c), the relation between a-c can be uncertain in some cases. Finally, in GUI-AGs, the relationships are reversible; e.g., if in the layout from Figure 3 the NAVIGATION[6] is below DATE PICKER[3], the reverse, DATE PICKER[3] is above NAVIGATION[6], must also be true. Hence, we can keep only one relationship between them.
## 3.2. Object/Relation Predictor
To construct the input for our transformer model, we first convert the GUI-AG into a sequence of relationship triplets $s_i$. We refer to this sequence of relationships as $S = \{s_1, s_2,... s_T\}$ where $T$ is the number of relationships. Relationships are separated by a special token SEP, and a special token CLS is used at the beginning of the entire sequence $S$.
Instead of using directly the sequence $S$ as input in the Object/Relation Predictor, we embed $s_{1:T}$ into $e_{1:T}^P$, which allows to take into consideration different features: word em- bedding $e_{1:T}^w$ that allows to identify the class of the object (e.g."button" or"container") or the relation (e.g."inside" or"right), object ID embedding $e_{1:T}^o$ to differentiate instances of the same object, relationship ID embedding $e_{1:T}^s$ to separate each relationship, type of word
| |
230409012/7
|
model a distribution in order to use Gaussian Mixture Models (GMM) [19]. We use this instead of directly predicting the bounding box from layout-aware representation in order to have a generative ability. Given a bounding box distribution, the bounding box $\hat{b}_t$ will be sampled from the posterior distribution $p_{\theta_t}(\hat{b}_t|c_t)$ knowing $c_t$. It can be described as follows:
$$
p_{\theta_t}(\hat{b}_t|c_t) = \sum^K_{i=1} \pi_i\mathcal{N}(\hat{b}_t;\theta_{t,i}),
$$
where $i$ indicates the $i$ -th distribution out of $K$ multivariate normal distributions, $\theta_{t,i}$ are the parameters of each distribution defined by $(\mu_{t,i}^x, \mu_{t,i}^y, \sigma_{t,i}^x, \sigma_{t,i}^y, \rho_{t,i}^{xy})$ where $\mu$, $\sigma$ and $\rho$ denote respectively the mean, standard deviation and the correlation coefficient, $\pi_i$ is a magnitude factor, and $\mathcal{N}$ is the multivariate normal distribution.
To define the objective function of the generator $\mathcal{L}_{gen}$, we start by defining the box reconstruction loss $\mathcal{L}_{box}$, which maximizes the log-likelihood of the generated GMM to fit the training data where the ground-truth bounding boxes are denoted as $b_t=(x_t, y_t, w_t, h_t)$.
$$
\mathcal{L}_{box} = -\frac{1}{K}\log (\sum^K_{i=1} \pi_i\mathcal{N}(b_t;\theta_{t,i})).
$$
To avoid the over-fitting with this loss function, the GMM distributions are fitted to a multivariate normal distribution $Q$ using a Kullback-Leibler (KL) divergence loss:
$$
\mathcal{L}_{KL} = \sum^K_{i=1} D_{KL}(P_i||Q_i).
$$
Finally, a relation consistency loss $\mathcal{L}_{rel}$ is also used since the two previous losses focuses only on the bounding boxes. It is given by:
$$
\mathcal{L}_{rel} = \frac{1}{N} \sum (\Delta \hat{b}_t - \hat{b}^{rel}_t)^2,
$$
where $N$ denotes the number of relationships in S. It calculates the Mean Square Error (MSE) between the box disparity of the relation we get, i.e. $\hat{b}_t^{rel}$ which is the predicted bounding box for the predicate, and the corresponding box disparity we calculate from the object and the subject $\Delta \hat{b}_t = (x_{t+1} - x_{t-1}, y_{t+1} - y_{t-1})$. The layout generator is trained by minimizing the weighted sum of losses using
$$
\mathcal{L}_{gen} = \lambda_{box} \mathcal{L}_{box} + \lambda_{KL} \mathcal{L}_{KL} + \lambda_{rel} \mathcal{L}_{rel},
$$
where $\lambda_{box}$, $\lambda_{KL}$ and $\lambda_{rel}$ are weighting factors for each corresponding loss.
## 3.4. Layout refiner
Since the bounding boxes are generated sequentially, they require refinement in the layout in order to consider the semantic $c_{1:T}$ and the bounding box $\hat{b}_{1:T}$. This is done in the layout refiner using the Visual-Textual Co-Attention (VT-CAtt) [6], which predicts the residual $\Delta \hat{b}_{1:T}$ for updating the bounding boxes. The objective function of this module $\mathcal{L}_{ref}$ is defined with multiple losses. The first one is a regression loss $\mathcal{L}_{reg}$ between predicted bounding boxes $b'_{1:T}$ by the layout refiner and the ground truth bounding boxes $b_{1:T}$. Another new and task- specific loss that we implemented is the overlap between children loss $\mathcal{L}_{CC}$, which aims to minimize the overlap of components that share the same parent in the GUI interface. This is
| |
230409012/12
|
## 5. Conclusion
This work propose a transformer-based model that generates a GUI layout from a given GUI-AG. Our approach is the state-of-the-art in quantitative performance across several metrics and in visual quality. We saw that using attention provides a higher performance than using graph convolution network in capturing semantic of the GUI-AG. The new components from our model compared to LayoutTransformer bring also more understanding in GUI layout constraints. This work also introduce new loss functions and evaluation metrics specific to this task of GUI layout generation. Future work is to generate GUI from the layouts to complete the GUI design pipeline.
| |
230409012/2
|
approach for generating GUI layouts from GUI-AGs. Transformer networks have recently achieved state-of-the-art results in a wide range of natural language processing and computer vision tasks. They are particularly well-suited for generating GUI layouts from GUI-AGs, as they can effectively capture the dependencies between elements in the GUI-AG and generate GUI layouts that reflect these dependencies. Our transformer-based model takes as input a series of tokens that express the GUI-AG relationships; it then outputs a realistic GUI layout. We demonstrate the effectiveness of our approach through a series of experiments on real-world datasets [3, 4] using metrics specific to this task. Results indicated that our approach produces the most relevant GUI layouts in regard to specified requirements.
Our contributions can be summarized with the following:
- We propose a new transformer-based method to generate GUI layouts from GUI- AGs that takes into account the GUI design constraints. Our experiment demon- strates that it better captures the intended GUI layout compared to previously proposed methods [5, 6];
- We introduce new loss functions and metrics for quantitative measurement of the quality of generated GUI layouts; Those metrics can be used for future work on similar tasks.
- To encourage reproduction or replication of our study, the source code of our ex- periment is publicly available with a pre-trained model at https://github.com/ dysoxor/GUILGET.
Figure 1. GUI layout generation goal is to generate a realistic GUI layout (b) from a given GUI-AG (a).
## 2. Related Work
GUI generation is an emerging but yet under-explored computer vision application, es- pecially for GUI generation based on GUI-AGs. First studies on generating GUI designs automatically adopt generative adversarial network, such as GUIGAN [7] and GANSpira- tion [8]. However, those approaches respectively do not generate new components style and do not consider any specification of the design – what components should be included or the relation between them. GUIGAN is based on a sequence of subtrees, which are hierarchical tree structures made of components, which is used as input and produces GUI by reusing different components based on their style without having the ability to produce new com- ponents. On the other hand, GANSpiration produces new design examples from existing screenshots or random vectors from the latent space representing the screenshots. To the best of our knowledge, no existing GUI generation methods explicitly consider the specifica- tions about the content, including the relationship between components. However, there are
| |
230409012/11
|
to always align components whether the complexity is low or high, and generated bounding boxes have almost the same similarity in distribution with the ground truth distribution.
## 4.4. Qualitative results
Figure 5. Qualitative comparison between our model, LayoutTransformer and SG2IM. The same input is given for the three models. The input from the first row has a low complexity with 3 unique component types while the second input has a larger complexity with 5 unique component types.
Figure 5 shows two examples produced by our model, LayoutTransformer, and SG2IM. Examples were chosen manually based on number of unique component types inside the GUI layout. We show results for low complexity (3-4 unique component types) and medium complexity (5-7 unique component types). We do not show GUI With large complexity (8 or more unique component types) as it is harder to analyze visually because of the larger number of components. The results of Figure 5 are aligned with the quantitative results from subsection 4.3. Indeed, we can see that SG2IM has a poor semantic understanding of relations for both cases in Figure5; it also struggles to place components inside their parent as we can observe in the second row from Figure 5 where all components that are supposed to be inside the CONTAINER[3] are not and the container which is supposed to be below the MAP[2] is actually entirely inside it instead. We can note however that the sizes of bounding boxes and their alignments are realistic. The LayoutTransformer shows a better understanding of relations but is not able in the second case to place components inside its parent as exemplified by BUTTON[4] and BUTTON[5] that are outside the CONTAINER[3] in the second row from Figure 5. In contrast, our model respects all the given constraints and placed correctly buttons inside the container. Also, GUILGET generates plausible bounding boxes even though the generated layouts are not aligned in the way it is in the ground truth. However, information from GUI-AG is not complete enough to reproduce the same alignment. Adding global positioning constraints on components to the GUI-AG could be an interesting avenue to investigate.
| |
230409012/6
|
embedding $e_{1:T}^t$ to identify parts of a relationship (subject = 1, predicate = 2, object = 3) for example the part of the the relationship"button inside container" are"subject predicate object" and instead of writting the plain text, we rather associate an ID to each part of the relationship to differentiate them, and parent ID embedding $e_{1:T}^p$, which is a feature that allows for each of the component to know its parent. Those features are concatenated to form the input embedding for the object/relation predictor. It is given by
$$
e_{1:T}^P = [e_{1:T}^w \bigoplus e_{1:T}^o \bigoplus e_{1:T}^s \bigoplus e_{1:T}^t \bigoplus e_{1:T}^p].
$$
The object/relation predictor learns to produce three different outputs: (1) the contex- tualized feature vectors $f_{1:T}$, (2) the size vectors $s_{1:T}$, and (3) the position vectors $p_{1:T}$. The contextualized feature vectors $f_{1:T}$ describe objects, relations and their context with features from the input. In order to capture conceptually diverse embedding and exploiting the co-occurrence among objects, predicates and parents, we follow the technique used in BERT [18] and mask randomly words from the input that must be predicted by the ob- ject/relation predictor. The size vectors $s_{1:T}$ are predictions of the bounding boxes size for each object made by the object/relation predictor. It is used as indicator later in the GUI layout generator to generate final bounding boxes size. The position vectors $p_{1:T}$ are predictions of bounding boxes position for each object.
Finally, in order to compute the objective function to train this part of the network, we need to predict $\hat{e}^P_{1:T}$ from the features $f_{1:T}$ using a single linear layer. Indeed, since there is not ground truth for $f_{1:T}$, we predict $\hat{e}^P_{1:T}$ from $f_{1:T}$ and match it with $e^P_{1:T}$ to minimize the reconstruction error. The objective function for training the module is composed of cross-entropy losses $\mathcal{L}_{predSem}$, given by
$$
\mathcal{L}_{predSem} = CrossEntropy(e^P_t,\hat{e}^P_t),
$$
for the matching word, object ID, type of word, parent ID which are all extracted from the input GUI-AG, and regression losses given by
$$
\mathcal{L}_{predBox} = Regression(s_t,\hat{s}_t) + Regression(p_t,\hat{p}_t),
$$
which are computed on predicted positions $\hat{p}_{1:T}$ and sizes $\hat{s}_{1:T}$ with their corresponding ground truth positions and sizes. The total loss of the predictor $\mathcal{L}_{pred}$ is a combination of the two losses and given by
$$
\mathcal{L}_{pred} = \mathcal{L}_{predSem} + \mathcal{L}_{predBox}.
$$
## 3.3. Layout generator
The goal of the layout generator module is to produce layout-aware representations $c_{1:T}$ and bounding boxes $\hat{b}_{1:T}$. This module is made of transformer decoder layers and interprets jointly and sequentially contextual features $f_{1:T}$, predicted bounding box sizes $s_{1:T}$ and predicted bounding box positions $p_{1:T}$ that are computed by the object/relation predictor module. The three given inputs are concatenated and expressed as $e^G_{1:T}$. After that it is translated into diverse bounding box output $\hat{b}_{1:T}$. A bounding box is described by its top- left corner position in a 2D Cartesian coordinate system and its size in terms of width and height, i.e., $\hat{b}_t = (x_t, y_t, w_t, h_t)$, and there is a bounding box produced for each subject, predicate and object. The bounding box of the predicate is the difference between the position of the object and the one of the subject, i.e., $\hat{b}_t = (x_{t+1} - x_{t-1}, y_{t+1} - y_{t-1})$.
To produce sequentially the bounding boxes $\hat{b}_t$, the features from the input $e^G_t$ are also concatenated with the previously produced bounding box $\hat{b}_{t-1}$. This input is not directly translated into the bounding box but to a layout-aware representation $c_t$ that is used to
| |
230409012/4
|
Figure2. Architecture of our model with three main components. Obj/Rel Predictor $\mathcal{P}$ takes as input an embedding $e^P$ which is a concatenation of several embeddings that describe different information about the given GUI-AG, and it produces contextual features $f$. Layout Generator $\mathcal{G}$ takes as input contextual features $f$, predicted sizes $s$ and predicted positions $p$ to translate it into a layout-aware representation $c$ and bounding boxes $b$. Layout Refiner $\mathcal{R}$ uses co-attention module with predicted bounding boxes $b$ and layout-aware representation $c$ to improve the layout.
to make predictions. GUI-AGs can be then viewed as a natural language input to the transformer since it is a logical sequence of relationships.
Our proposed method is based on the LayoutTransformer (LT-net) [6], a transformer- based model that aims to generate diverse layouts from scene graphs of images. It consists of (1) an object/relation predictor $\mathcal{P}$ that encodes the input scene graph into contextual features $f$ using transformer encoder layers, (2) a layout generator $\mathcal{G}$ that, from contextual features $f$, generates bounding boxes $b$ with distributions matching a learned Gaussian dis- tribution model and layout-aware representations $c$ with transformer decoder layers, and (3) a layout refiner $\mathcal{R}$ made of a co-attention module. Our architecture is presented in Figure2. Compared to LT-net, we introduced several improvements for GUI layout generation that will be presented in the following.
## 3.1. Building a GUI arrangement graph
GUIs are made of two types of components: (1) widgets (e.g. button, pictogram, text), which are leaf nodes in a GUI-AG representation and do not contain any other component, and (2) spatial layouts (e.g. container, list item, toolbar), which are intermediate nodes that allow to organize the structure of widgets [7]. This tree structure does not exist in images and it changes the way GUI-AGs should be modeled and processed compared to scene graphs for images. In contrast to our work, most of current works that are done in layout generation for GUI ignore the tree structure and only consider widgets to remove the complexity of organizing those widgets inside layouts.
GUI-AGs are directed graphs made of relationships that are triplets of subject-predicate- object [5]. A GUI does not require as much variety of relations and objects as images. However, it has complexity since GUI layouts have more rules and principles to be learned than in a layout from an image [2]. To train a neural network model, GUI-AGs from ground truth layouts are required. We define five types of possible predicate specifying the relationship among GUI components: left, right, top, bottom and inside. We build a
| |
230409012/10
|
to have a high CCS if CPI and GUI-AGC are not also high. The LayoutTransformer model has more understanding of predicates but it is still worse than with our model. We can also see that there is a negative correlation between CCS and CPI – if the model learns to place components inside its parent, there are more possible overlaps between components inside a layout. We want both of these metrics to be similarly high to respect both of those GUI design constraints as we can observe it in the GT data, where all metrics related to GUI design constraint are close to 1. W bbox metric is similar for all of the models which is understandable since our model and LayoutTransformer model generate bounding boxes based on distribution of bounding boxes from the training GUI layouts. On the other hand, SG2IM is not a generative model and aims to predict bounding boxes which leads to learn the most common sizes and positions for different types of component.
## 4.3.1. Influence of UI category
Figure4. Quantitative evaluation on different screen categories (a) from CLAY dataset [3]. Evaluation metrics are applied on all 27 screen categories separately. (b) shows the influence of number of unique type components on evaluation metrics.
In order to see if the screen category has an impact on the performance, we conducted an experiment where we compute each evaluation metric for each category separately. Figure4 (a) summarizes the results. Overall, our model yields similar performances among different app categories. This is a conclusive result which shows that our model is not biased toward certain types of screen categories and is able to produce equally good GUI layout for any of the category.
## 4.3.2. Influence of UI complexity
Similarly as with the previous experiment, we want to understand the influence of the UI complexity (indicated by the number of unique component types in the UI [8]) on the performance. Figure4 (b) shows that with smaller number of unique component types most of the evaluation metrics are better; in other words, our model achieved better performances when the UI is less complex. Particularly, the GUI-AGC metric is inversely proportional to the number of unique component types in the UI. For both CCS and CPI metrics, these results are expected due to the fact that with more diverse components that have various sizes and position standards, it becomes harder to organize all elements inside spatial layouts. There are however two metrics performing equally well over all number of unique component types, which are the Alignment and W bbox metrics. This shows that our model succeeds
| |
230409012/1
|
# GUILGET: GUI Layout GEneration with Transformer
\begin{tabular}{cc}
Andrey Sobolevsky\upstairs{\affilone,*}, Guillaume-Alexandre Bilodeau\upstairs{\affilone}, Jinghui Cheng\upstairs{\affilone}, Jin L.C. Guo\upstairs{\affiltwo}
[0.25ex]
{\small \upstairs{\affilone} Polytechnique Montréal} {\small \upstairs{\affiltwo} McGill University} \end{tabular}
Sketching out Graphical User Interface (GUI) layout is part of the pipeline of de- signing a GUI and a crucial task for the success of a software application. Arranging all components inside a GUI layout manually is a time-consuming task. In order to assist designers, we developed a method named GUILGET to automatically generate GUI layouts from positional constraints represented as GUI arrangement graphs (GUI- AGs). The goal is to support the initial step of GUI design by producing realistic and diverse GUI layouts. The existing image layout generation techniques often cannot in- corporate GUI design constraints. Thus, GUILGET needs to adapt existing techniques to generate GUI layouts that obey to constraints specific to GUI designs. GUILGET is based on transformers in order to capture the semantic in relationships between ele- ments from GUI-AG. Moreover, the model learns constraints through the minimization of losses responsible for placing each component inside its parent layout, for not letting components overlap if they are inside the same parent, and for component alignment. Our experiments, which are conducted on the CLAY dataset, reveal that our model has the best understanding of relationships from GUI-AG and has the best performances in most of evaluation metrics. Therefore, our work contributes to improved GUI layout generation by proposing a novel method that effectively accounts for the constraints on GUI elements and paves the road for a more efficient GUI design pipeline.
Keywords: Graphical User Interface, GUI arrangement graphs, deep learning, trans- former, generative model, GUI layout
## 1. Introduction
The design of Graphical User Interface (GUI) is an important aspect that affects the success of many software applications. The first step for the GUI designers is often sketching out the interface layout with wireframes, based on design constraints such as users’ needs and software requirements [1]. These GUI layouts define the visual arrangement of elements in a user interface, such as buttons, text fields, and containers. Creating these layouts and variations of them manually, however, can be time-consuming. In this paper, we explore an automated technique that can support designers creating GUI layouts. In our approach, we capture the design constraints that the designers have to consider through GUI arrangement graphs (GUI-AGs) and generate graphical user interface layouts from those constraints. GUI-AGs specify elements required in the UI design and the relationships among them, as illustrated in Figure 1. We use GUI-AGs because they can be used to describe the requirements with an explicit definition of components that are part of the screen and the definition of the visual relations between those components. These graph models offer the flexibility to automatically create layout variations by modifying relations in the graph.
There are several technical challenges for achieving automatically generatation of GUI layouts from GUI-AGs. One challenge is to accurately capture the logical and semantic rela- tionships between GUI elements, such as hierarchical structures and functional dependencies. Another challenge is to generate visually appealing and functional layouts that adhere to design principles and constraints [2]. In addition, the generation process should be efficient to be used in design workflow. To address these challenges, we propose a transformer-based
| |
230409012/3
|
several works focusing on image and layout generation for more general domains that can be inspirational for GUI generation [9, 10]. For example, PasteGAN [9] generates images from scene graphs and image crops by taking into account the semantic relationships between objects and their visual appearance. Another method, proposed by Zhao et al. [10], uses a transformer to generate image tokens from a given scene graph and then decodes those tokens into a plain image using a VQGAN decoder [11]. In comparison, our work aims to produce first GUI layouts, which is usually the first step in the GUI design process, rather than directly generating the GUI design. Our work can be used as a part of a GUI design generation pipeline.
Our concept of GUI-AG is inspired by scene graph, a widely used structure in computer graphics. The task of generating layouts from scene graphs is studied in computer graphic, but also requires knowledge from natural language processing (NLP) to process the textual input and understand the interactions between each element of the scene graph. Scene graphs are commonly used to represent the structure and relationships between graphical elements in a scene, and layout generation aims to determine the position and size of these elements in the final layout.
There have been several approaches proposed for generating layouts from scene graphs in the literature. For example, SG2IM by Johnson et al. [5] adopted a graph convolutional network (GCN) to process the scene graph. This approach, however, is unable to capture semantic equivalence in graphs. The work of Herzig et al. [12] tackle this issue by learning canonical graph representations from the input scene graph, and it allows to represent more complex scene graphs. Another way of processing scene graphs is by using transformer-based model such as the ones presented in [6, 13]. Transformers have shown great performances in processing textual components and learning relations between them. The method presented by Gupta et al. [13] uses self-attention to learn contextual relationships between layout element. In practice, it takes layout elements as input and produces the next layout elements as output. So it has no constraints and the model produces random new layouts. On the other hand, in the work LayoutTransformer by Yang et al. [6], the transformer-based model takes as input constraints in the form of scene graphs and predicts the layout based on relationships.
Many works have been done to generate layouts with different type of input from scene graphs. LayoutVAE [14], a variational autoencoder (VAE)-based framework for generating stochastic scene layouts, is a model that takes as input a set of labels representing the entities that must appear in the layout. It does not take into account any positional constraints, neither the amount of each label as with scene graphs. The Neural Design Network (NDN) [15] combines a GCN with a conditional VAE to generate layouts based on constraints given as input. In NDN, the scene graph is generated based on the amount of components there are and some desired positional constraints. Another work uses a generative adversarial network (GAN) [16] to generate layouts from an input of randomly-placed 2D elements. It uses self-attention modules to refine their labels and geometric parameters jointly to produce a layout.
In contrast to previous works, we propose a method that takes into account the specifica- tions of a UI design using GUI-AGs that are considering the amount of each UI component and component position relative to other components.
## 3. Method
Our work is based on the use of transformers [17] that have been shown to be state- of-the-art in multiple natural language processing tasks. A transformer uses the concept of self-attention, which allows to give different weights depending on the part of the input
| |
230409012/8
|
specific to design principles in GUI since we do not want the components to overlap because it will hide some components on the final interface. This loss is given by
$$
\mathcal{L}_{CC} = \frac{C_1 \cap C_2}{\min(C_1, C_2)},
$$
where $C$ designate the area of a children. Another principle to follow is that a children must be inside its parent. To enforce this principle, we define the overlap between children and parent loss $\mathcal{L}_{CP}$ and we express it as
$$
\mathcal{L}_{CP} = 1 - \frac{C \cap P}{C},
$$
where $C$ is the area of the children and $P$ is the area of its parent that is defined in the input GUI-AG. The objective function on which the layout refiner is trained is a weighted sum of those losses, that is
$$
\mathcal{L}_{ref} = \lambda_{reg} \mathcal{L}_{reg} + \lambda_{CC} \mathcal{L}_{CC} + \lambda_{CP} \mathcal{L}_{CP}.
$$
where $\lambda_{reg}$, $\lambda_{CC}$ and $\lambda_{CP}$ are weighting factors for each corresponding loss.
## 4. Experiments
In our experiment, we validated our proposed method by generating layouts from GUI- AGs. We aim to evaluate how close the generated layouts are to the ground truth layouts based on the graphs associated to them.
## 4.1. Dataset
We tested our method on the CLAY dataset [3], which is a UI design dataset. UI layouts in RICO dataset [4] are often noisy and have visual mismatches hence CLAY is a dataset that improves RICO by denoising UI layouts. It contains 59,555 human-annotated screen layouts, based on screenshots and layouts from RICO. A total of 24 component categories (e.g. Image, button, text) are available in layouts and 5 predicate categories (above, below, right, left, inside) are considered in GUI-AGs.
By observing data from the CLAY dataset, we decided to remove several irrelevant GUIs and their layouts. Firstly, we removed GUIs that contain two or less types of components. Those screenshots are usually not representing an application GUI but rather GUIs with, for instance, full screen image and video screenshot that do not contain useful information for our task, as there is a lack of component and interactions (predicates) between components. Then, we also removed screens that contain only a navigation bar or popup for example, for the same reason. Also usually in the dataset there are several screenshots associated to the same application but some contain only the navigation bar and others have the navigation bar and also some content. So we removed the former to avoid overfitting. In practice, we achieved that by removing GUIs in which components cover less than 25% of the total area of the screen.
## 4.2. Evaluation metrics
To evaluate the generated layouts with respect to the ground truth, we used the following metrics.
CP Inclusion (CPI) is a metric that captures the overlapping of children with its parent. The metric is computed as $1-\mathcal{L}_{CP}$ (see Equation 3.11). Hence, the goal of this
| |
170805157/20
|
Fig.B.2. The neutralising gas current $J$ (black contours) and the drift current $J_{\rm dri}$ (red for positive current and blue for negative), as a function of electric field $E$ and dust metallicity $Z_1$. The total current of the solids is separating to the left of the thick yellow zero-line and neutralising to the right of the zero-line. The drift current is at best 2-3 orders of magnitude smaller than the neutralising gas current.
Fig.B.3. Neutralising gas current and drift current for a model with re- duced $^{26}$ Al abundance (10% of nominal value) and decreased gas den- sity ($\rho_{\rm g}=10^{-8}\,{\rm kg\,m^{-3}}$). Both the neutralising gas current and the break- down value of the electric field are a factor 10 lower than in Figure B.2, but the drift current is still at least a factor 10 too small to compete with the neutralising current.
In conclusion, it is very hard to obtain charge-separating ra- dial drift currents that are higher than the neutralising current. An additional concern about chondrule formation by lightning in overdense filaments is the time-scale for which these high den- sities can be sustained. Simulations of planetesimal formation by the streaming instability demonstrate that the particle densi- ties can reach very high values of several thousand times the gas density during the collapse (Johansen et al., 2015). This phase is nevertheless short, on the order of a few orbital periods, and hence can not lead to widespread thermal processing of the par- ticles involved in the collapse. Therefore we consider in Section 3 a potentially much power powerful source of currents in pro- toplanetary discs, namely the flux of positrons emanating from dense pebble regions.
## AppendixC: Adsorption rates
The cross section $\sigma_{dj}$ for collisions between a charged species $j$ and a dust grain of type $d$ with charge $q_d$, radius $R_d$ and surface potential $\phi_d=(4 \pi \epsilon_0)^{-1} q_d/R_d$ is given by (see e.g. Shukla &
$$
\sigma_{dj} = \pi R_d^2 \mathopen{}\mathclose\bgroup\originalleft[1 - \frac{2 (\pm e) \phi_d}{m_j v_j^2} \aftergroup\egroup\originalright]
\, .
$$
Here $m_j$ and $v_j$ are the mass and speed of the charged species and the $\pm$ symbol refers to ions of charge $+e$ and electrons of charge $-e$, respectively. The index $j$ can now be used to sum over a velocity distribution. We denote the electron temperature by $T_{\rm e}$ and the ion temperature by $T_{\rm i}$. These temperatures can obtain values di ff erent from the thermal temperature $T$ due to heating by the electric field, see discussion in Section A.4. For the electrons we assume that the mean speed is negligible com- pared to the thermal speed (Okuzumi&Inutsuka,2015;Mori& Okuzumi,2016). The absorption rates of electrons are therefore given by expressions that are independent of the mean speed,
$$
\begin{aligned}
K_{d{\rm e}} &=& \pi R_d^2 \sqrt{\frac{8 k_{\rm B} T_{\rm e}}{\pi m_{\rm e}}}
\mathopen{}\mathclose\bgroup\originalleft(1 + \frac{e \phi_d}{k_{\rm B} T_{\rm e}} \aftergroup\egroup\originalright) \quad {\rm
for\,\,\phi_{\it d}>0} \, , \\
K_{d{\rm e}} &=& \pi R_d^2 \sqrt{\frac{8 k_{\rm B} T_{\rm e}}{\pi m_{\rm e}}}
\exp \mathopen{}\mathclose\bgroup\originalleft(\frac{e \phi_d}{k_{\rm B} T_{\rm e}} \aftergroup\egroup\originalright) \quad {\rm
for\,\,\phi_{\it d}<0} \, .
\end{aligned}
$$
$$
\begin{aligned}
K_{d{\rm e}} &=& \pi R_d^2 \sqrt{\frac{8 k_{\rm B} T_{\rm e}}{\pi m_{\rm e}}}
\mathopen{}\mathclose\bgroup\originalleft(1 + \frac{e \phi_d}{k_{\rm B} T_{\rm e}} \aftergroup\egroup\originalright) \quad {\rm
for\,\,\phi_{\it d}>0} \, , \\
K_{d{\rm e}} &=& \pi R_d^2 \sqrt{\frac{8 k_{\rm B} T_{\rm e}}{\pi m_{\rm e}}}
\exp \mathopen{}\mathclose\bgroup\originalleft(\frac{e \phi_d}{k_{\rm B} T_{\rm e}} \aftergroup\egroup\originalright) \quad {\rm
for\,\,\phi_{\it d}<0} \, .
\end{aligned}
$$
The adsorption rate of ions must be calculated by considering the separate flux contributions from the ion temperature $T_{\rm i}$ and the ion speed along the electric field, $\langle {\boldsymbol v}_{\rm i} \rangle$. The adsorption rate is calculated through the integral
$$
\begin{aligned}
K_{d{\rm i}} &\equiv& \pi a^2 \int d^3 v_{\rm i} \,\, v_{\rm i}
\mathopen{}\mathclose\bgroup\originalleft(1-\frac{2e\phi_d}{m_{\rm i} v_{\rm i}^2}\aftergroup\egroup\originalright) f_{\rm i}({\bm v}_{\rm i})
\\
&=& \pi a^2 \int_{v_0}^\infty dv_{\rm i} \,\, 4\pi v_{\rm i}^3
\mathopen{}\mathclose\bgroup\originalleft(1-\frac{2e\phi_d}{m_{\rm i} v_{\rm i}^2}\aftergroup\egroup\originalright) \\
&& \qquad \qquad \qquad \qquad \times \int_{\cos\theta=-1}^{\cos\theta=1}
d(\cos\theta) f_{\rm i} (v_{\rm i}, \theta) \, .
\end{aligned}
$$
We then insert a Maxwellian distribution function that includes a mean velocity $\langle {\boldsymbol v}_{\rm i} \rangle$,
$$
\begin{aligned}
f_{\rm i}({\bm v}_{\rm i},\theta) &=& \mathopen{}\mathclose\bgroup\originalleft(\frac{m_{\rm i}}{2\pi k_{\rm B} T_{\rm
i}}\aftergroup\egroup\originalright)^{3/2} \exp\mathopen{}\mathclose\bgroup\originalleft[-\frac{m_{\rm i}}{2k_{\rm B} T_{\rm i}}({\bm v}_{\rm
i}-\langle{\bm v}_{\rm i}\rangle)^2\aftergroup\egroup\originalright] \\
&=& \mathopen{}\mathclose\bgroup\originalleft(\frac{m_{\rm i}}{2\pi k_{\rm B} T_{\rm i}}\aftergroup\egroup\originalright)^{3/2}
\exp\mathopen{}\mathclose\bgroup\originalleft[-\frac{m_{\rm i}}{2k_{\rm B} T_{\rm i}}(v_{\rm i}^2 + |\langle{\bm v}_{\rm
i}\rangle|^2 - 2v_{\rm i}|\langle{\bm v}_{\rm i}\rangle|\cos\theta)\aftergroup\egroup\originalright]
\, . \\
\end{aligned}
$$
Here $v_{\rm i} = |{\bm v}_{\rm i}|$ is the magnitude of ${\bm v}_{\rm i}$ and $\theta$ is the angle between ${\bm v}_{\rm i}$ and $\langle{\bm v}_{\rm
i}\rangle$. We denote by$v_0$the collision speed below which the e ff ective collision cross section$\pi R_d^2 [1-2e\phi_d/(m_{\rm i} v_{\rm i}^2)]$ is zero. That gives
$$
\begin{aligned}
v_0 &=& \sqrt{\frac{2e\phi_d}{m_{\rm i}}} \qquad {\rm for}\,\,\phi_d>0 \, , \\
v_0 &=& 0 \qquad \qquad \quad {\rm for}\,\, \phi_d<0 \, .
\end{aligned}
$$
$$
\begin{aligned}
v_0 &=& \sqrt{\frac{2e\phi_d}{m_{\rm i}}} \qquad {\rm for}\,\,\phi_d>0 \, , \\
v_0 &=& 0 \qquad \qquad \quad {\rm for}\,\, \phi_d<0 \, .
\end{aligned}
$$
| |
170805157/18
|
Fig.A.2. The electron and ion densities (topleft; blue lines indicate electrons and red lines ions), the pebble charge (topright, blue lines indicate negative charge, red lines positive charge, black lines the charge relative to a model that neglects charge transfer in dust collisions), the electron and ion currents (bottom left) and the total current (bottom right). The electron and ion critical fields are marked in blue and red dotted lines, while the break down electric field is shown with black dotted lines (we show both the nominal value and a value ten times lower relevant when taking into account the distribution of electron energies). The positive pebble charging is mainly due to the emission of secondary electrons as positrons penetrate dust surfaces. The equilibrium charge increases with increasing field strength, since the field accelerates electrons and ions to high speeds, thus increasing the collision rate with the charged pebbles.
rent density is in good agreement with Figure A.2. For the elec- trons we obtain the limiting expression
$$
\begin{aligned}
J_{{\rm e},\infty} &=& \sqrt{\frac{\pi m_{\rm e}}{3 m_{\rm n}}} \frac{e \zeta
n_{\rm n}}{\sigma_1 n_1} = \sqrt{\frac{\pi m_{\rm e}}{3 m_{\rm n}}} J_{{\rm
i},\infty} \\
&\approx& 1.9 \times 10^{-16} \,{\rm C\,m^{-2}\,s^{-1}} \mathopen{}\mathclose\bgroup\originalleft(
\frac{Z_1}{100} \aftergroup\egroup\originalright)^{-1} \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m} \aftergroup\egroup\originalright) \, .
\end{aligned}
$$
This is also in excellent agreement with Figure A.2 until $Z\approx
100$. Beyond this mass loading the electron production rate rises due to release of secondary electrons. The electron current is nevertheless less important than the ion current in setting the overall current density at all relevant values of the mass loading. Importantly, the limiting values of the ion and electron currents are independent of the gas density and hence the results pre- sented in Figure A.2 are relatively independent of the distance from the star, except for a small e ff ect of the disc temperature on the conductivities.
## B.4. Charge flux
Figure A.2 shows that the charge on large pebbles, $q_2$, is posi- tive for high mass loading. This is due to the emission of sec- ondary electrons as a positron passes through a particle surface. Pebbles are large enough to completely stop positrons. Therefore a positron will on the average pass through $\lambda_2/\lambda_1$ small dust grains before it is absorbed by a pebble, with $\lambda_1$ and $\lambda_2$ indi- cating the mean free path of the positron relative to dust and pebbles. The relative production rate of secondary electrons is
$$
f_{12} = \frac{N_{\rm s1}}{N_{\rm s2}} \frac{\lambda_2}{\lambda_1}
= \frac{N_{\rm s1}}{N_{\rm s2}} \frac{A_1}{A_2} \, .
$$
Here $A_1/A_2$ is the total surface area of dust grains relative to the total surface area of pebbles within a given volume. The total adsorption rate of secondary electrons by the two dust com- ponents is also proportional to $A_i$, Therefore, secondary electron release only yields net charging if $N_{\rm s1}$ is not equal to $N_{\rm s2}$. In
| |
170805157/17
|
Fig.A.1. The ionisation rate $\zeta$ versus the dust-to-gas ratio. The initial rise of the ionisation rate is due to the increased density of $^{26}$ Al, while the plateau occurs when the rise is balanced by an increased absorp- tion rate of positrons in dust grains. For the production of secondary electrons, which occurs in collisions with dust, we have converted to an e ff ective rate on the gas particles for comparison. Secondary electrons come to dominate after $Z \approx 1000$.
of energy from the mean motion along the electric field to ran- dom motion after collisions with neutrals. The increased random motion in term decreases both the number densities and mean current speed of electrons and ions. Following Mori&Okuzumi (2016) we write the electron temperature as
$$
T_{\rm e} = T \mathopen{}\mathclose\bgroup\originalleft[\frac{1}{2} + \sqrt{\frac{1}{4}+\frac{2}{3} \mathopen{}\mathclose\bgroup\originalleft(
\frac{E}{E_{\rm crit}} \aftergroup\egroup\originalright)^2} \aftergroup\egroup\originalright] \, .
$$
The critical electric field for electron heating $E_{\rm crit}$ is
$$
E_{\rm crit} = \sqrt{\frac{6 m_{\rm e}}{m_{\rm n}}} \frac{k_{\rm B} T}{e
\ell_{\rm e}} \, .
$$
Here $m_{\rm e}$ and $m_{\rm n}$ are the electron and neutral masses, re- spectively. The mean free path of the electrons is simply $\ell_{\rm e} =
(n_{\rm n} \sigma_{\rm en})^{-1}$with$\sigma_{\rm en}=10^{-19}$m$^2$.
The ions are heated at higher values of the electric field. Mori &Okuzumi(2016) give an expression for the ion temperature,
$$
T_{\rm i} = T \mathopen{}\mathclose\bgroup\originalleft[1 + 7.6 \times 10^{-7} \mathopen{}\mathclose\bgroup\originalleft(\frac{T}{100\,{\rm K}}
\aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(\frac{E}{E_{\rm crit}} \aftergroup\egroup\originalright)^2 \aftergroup\egroup\originalright] \, .
$$
The ion temperature increases as the square of the electric field, while the electron temperature only increases linearly with the electric field in the limit of high field strength.
## AppendixB: Charge equilibrium
The direct temporal integration to equilibrium of the charg- ing equations presented in Appendix A is made very expen- sive by the large range of time-scales present in the problem. Particularly at high particle mass loading and strong electric fields, relevant for the heating of chondrules, the time-scale for collisions between electrons and dust particles is very low. Therefore we search for the equilibrium condition by solving di- rectly for $\dot{n}_{\rm e}=\dot{n}_{\rm i}=\dot{q}_1=\dot{q_2}=0$. Unfortunately, none of these dynamical equations are analytically solvable, due to the com- plicated expressions for the adsorption rates $K_{d{\rm e}}$ and $K_{d{\rm i}}$ (see Appendix C).
## B.1. Numerical solution method
One of the four variables in the solution vector $(n_{\rm e},n_{\rm i},q_1,q_2)$ can be eliminated by using the charge conservation condition from equation (A.6). Our algorithm for finding the equilibrium state consists of stepping through $n_{\rm i}$ from $10^{-30} n_{\rm n}$ to $10^{-5} n_{\rm n}$. For each value of $n_{\rm i}$ we find $n_{\rm e}$ that satisfies $\dot{n}_{\rm
e}=0$. This requires addi- tional knowledge of$q_1$and$q_2$. The value of$q_1$, given$n_{\rm e}$and$n_{\rm i}$, is calculated from$\dot{q}_1=0$and$q_2$is calculated straightforwardly from the charge conservation equation. We then minimise the value of$\dot{n}_{\rm i}$along the curve$\dot{n}_{\rm
e}=0$. All zero points are found by bisection to a tolerance level of$10^{-20}$. We implemented the al- gorithm in Fortran 90 and compiled with quadruple precision in order to avoid underflows in the very large terms that are present in the dynamical equations.
We consider electric field strengths $E$ between $10^{-4}\,{\rm V\,m^{-1}}$ and $10^3\,{\rm V\,m^{-1}}$ and a density of small particles relative to the gas density $Z_1$ between $10^{-3}$ and $10^4$. The density of pebbles is fixed at a 10 times higher value, $Z_2 = 10 Z_1$. The radius of small dust is set to 1 $\mu$ m and the radius of pebbles to 1 mm; both components are assumed spherical with an internal density of $\rho_\bullet = 10^3\,{\rm kg\,m^{-3}}$. For each value of $E$ and $Z_1$ we solve the charge equilibrium equations to obtain $n_{\rm e}$, $n_{\rm i}$, $q_1$ and $q_2$.
## B.2. Results
The equilibrium state is plotted in Figure A.2. We show the elec- tron and ion densities, the charge on the macroscopic pebbles and the current densities of electrons and ions, all as functions of the electric field and the density of small dust grains. We also indicate the critical electric fields for electron heating (blue dot- ted line) and ion heating (red dotted line) as well as the break down electric field to trigger lightning discharge (black dashed lines), for the nominal value of Desch&Cuzzi(2000) as well as for a 10 times lower value that may reflect the actual breakdown field strength [equation (1)]. Both ion and electron densities fall after reaching their respective critical electric field strengths, due to the increase in collision rates with dust. The ion density falls faster than the electron density as the ions interact weakly with neutrals and hence are accelerated to very high speeds that re- sult in high collision rates with the dust particles. The electron density falls much more slowly when the mass loading reaches $Z_1\sim100$ (corresponding to a total metallicity of $Z \sim 10^3$, see Figure A.1). This is due to the release of secondary electrons as positrons pass through lots of dust grains at those densities.
## B.3. Limiting values for electron and ion currents
Okuzumi & Inutsuka (2015) showed that the heating of ions causes the ion current density to plateau out at
$$
\begin{aligned}
J_{{\rm i},\infty} &=& \frac{e \zeta n_{\rm n}}{\sigma_1 n_1} = \frac{e \zeta
m_1}{Z_1 \sigma_1 m_n} \\ &\approx& 1.1 \times 10^{-14}\,{\rm
C\,m^{-2}\,s^{-1}} \mathopen{}\mathclose\bgroup\originalleft(\frac{Z_1}{100} \aftergroup\egroup\originalright)^{-1} \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm
\mu m} \aftergroup\egroup\originalright) \, .
\end{aligned}
$$
Here $\zeta=1.7 \times 10^{-17}\,{\rm s^{-1}}$ is the ionisation rate by positrons from $^{26}$ Al, valid when $Z \gg 1$ (see Figure A.1). The limiting ion cur-
| |
170805157/19
|
Fig.B.1. Diagram of the charge fluxes between the five compo- nents neutrals, electrons, ions, dust, pebbles and positrons for $E=10\,{\rm V\,m^{-1}}$ and $Z_1=100$. The fluxes are normalised by the charge flux to dust and pebbles by ionisation. The abbreviations are: IO = ionisation, AD = adsorption, SE = secondary electrons, CO = dust col- lisions, DE = nuclear decay and positron emission, ST = stopping of positrons in solids, AN = annihilation of positrons in gas.
Appendix we show that $N_{\rm s1}=0.04$ and $N_{\rm s2}=0.08$. This asymmetry in the secondary electron production yields net pos- itive charging of the pebbles at the expense of the dust grains. The resulting pebble charge can be very high, up to $10^5$ electron charges for strong electric fields. The charging of pebbles takes much higher values when the charge flux from collisions with small dust is ignored (as indicated by the black contour lines in the upper right plot of Figure A.2).
In Figure B.1 we show the charge flux for the six compo- nents of the charging model (including the neutrals), normalised to the charge flux to the dust and pebble component by positron ionisation, for $Z_1=100$ and $E=10\,{\rm
V\,m^{-1}}$. Charging by decay of$^{26}$Al (DE) and stopping of positrons (ST) are in close balance for both dust and pebbles, with just a small fraction of positrons stopped by the gas (AN). The production rate of secondary elec- trons (SE) by the dust is less than 100 times (i.e., the ratio of total dust surface to total pebble surface) that of the pebbles, due to the lower e ffi ciency of secondary electron production of the dust. This asymmetry of secondary electron production is indeed the source of pebble charging, as the pebbles must charge positively to pull in additional electrons that can compensate for the loss of secondary electrons. Charge transfer in dust collisions (CO) is another source of pebble flux on the pebbles; as is evident in Figure A.2 this dust flux reduces the pebble charge by approxi- mately a factor ten at$(E,Z_1)=(10\,{\rm
V\,m^{-1}},100$).
## B.5. Drift current
The large, positive charging of the pebbles implies that spatially separating pebbles from gas and dust will lead to the build-up of an electric field. The question is whether the charge separation is continuously neutralised by the electron / ion current or whether it builds up rapidly enough to necessitate neutralisation by light- ning discharge. We compare in Figure B.2 the drift current to the gas current. We calculate the drift current from
$$
J_{\rm dri}(E,Z_1) = n_1 e q_1 v_{1x} + n_2 e q_2 v_{2x} + n_{\rm i} e u_x -
n_{\rm e} e u_x \, ,
$$
where $v_{1x}$, $v_{2x}$ and $u_x$ are the pebble, dust and gas velocities in the radial direction. These velocities result from both radial drift, as the pressure gradient on the gas forces large pebbles to drift outwards at the cost of small dust and gas drifting inwards, as well as the electric acceleration of the dust and pebbles. The equilibrium drift velocities are found by finding the dynamical equilibrium of the gas velocity $(u_x,u_y)$ and the particle velocities $(v_{1x},v_{1y})$, $(v_{2x},v_{2y})$ in the local frame corotating with the proto- planetary disc at the Keplerian frequency $\varOmega$. This equilibrium is given by the condition
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
$$
\begin{aligned}
\dot{u}_x = 0 &=& 2 \varOmega u_y - \frac{Z_1}{\tau_1} (u_x-v_{1x}) -
\frac{Z_2}{\tau_2} (u_x-v_{2x}) - 2 \varOmega \Delta v \, , \\
\dot{u}_y = 0 &=& -\frac{1}{2} \varOmega u_x - \frac{Z_1}{\tau_1}
(u_y-v_{1y}) - \frac{Z_2}{\tau_2} (u_y-v_{2y}) \, , \\
\dot{v}_{1x} = 0 &=& 2 \varOmega v_{1y} - \frac{1}{\tau_1} (v_{1x}-u_x) +
\frac{q_1}{m_1} E_x \, , \\
\dot{v}_{1y} = 0 &=& -\frac{1}{2} \varOmega v_{1x} - \frac{1}{\tau_1}
(v_{1y}-u_y) \, , \\
\dot{v}_{2x} = 0 &=& 2 \varOmega v_{2y} - \frac{1}{\tau_1} (v_{2x}-u_x) +
\frac{q_2}{m_2} E_x \, , \\
\dot{v}_{2y} = 0 &=& -\frac{1}{2} \varOmega v_{2x} - \frac{1}{\tau_1}
(v_{2y}-u_y) \, .
\end{aligned}
$$
Here $\tau_1$ and $\tau_2$ are the friction times of dust and pebbles, respec- tively, and $\Delta v$ denotes the sub-Keplerian speed resulting from the radial pressure gradient (Nakagawa et al., 1986). In Figure B.2 we overplot contours of the drift current on the neutralising gas current. The drift current is separating (opposite sign as the elec- tric field) to the left of the zero-line neutralising (same sign as the electric field) to the right of the zero-line. The zero-line of the drift current is reached well before the breakdown value of the electric field at high dust mass loadings. This implies that the mobility of the solids leads to neutralisation of the charge separation well before lightning discharge can occur.
In Figure B.3 we present a model where the abundance of $^{26}$ Al has been reduced by a factor 10 relative to the nominal value. This leads to a corresponding decrease in the gas con- ductivity (see equation B.1). We also reduced the gas density by a factor 10 to mimic conditions at the late evolution stages of a protoplanetary disc after a few Myr of gas accretion onto the star. The pebble charge is actually not reduced by the de- crease of the positron production, since the equilibrium is set by the relative e ffi ciency of secondary electron production be- tween small dust grains and macroscopic pebbles. The drift cur- rent is also relatively una ff ected by the change of gas density, since the reduction in the number density of the solids is bal- anced by their higher mobility relative to the gas. The lowered gas density decreases both the critical electric field strengths for electron and ion heating as well as the breakdown electric field. The results nevertheless show that the zero-drift current remains relatively unchanged, although the lowered breakdown value of the electric field has approached the zero-drift current by almost a factor 10. A further reduction in the gas conductivity and the breakdown electric field by an order of magnitude could open a pathway to separating charges faster than neutralisation. However, such low gas densities would imply particle densities much lower than those required for chondrule formation.
| |
170805157/5
|
Inserting the ionisation rate from equation (10), we find the limiting values of the ion current, for low and high total particle mass loading, as
$$
\begin{aligned}
J_{\rm i,\infty} &=& 1.9 \times 10^{-11}\,{\rm C\,m^{-2}\,s^{-1}} \times
[(Z/Z_1)/10] \\
& & \times \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \quad {\rm for}\, Z \ll 1
\label {eq:JiZl} \\
J_{\rm i,\infty} &=& 1.9 \times 10^{-12}\,{\rm C\,m^{-2}\,s^{-1}} \times
Z_1^{-1} \\
& & \times \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \quad {\rm for}\, Z \gg 1
\label {eq:JiZh}
\end{aligned}
$$
$$
\begin{aligned}
J_{\rm i,\infty} &=& 1.9 \times 10^{-11}\,{\rm C\,m^{-2}\,s^{-1}} \times
[(Z/Z_1)/10] \\
& & \times \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \quad {\rm for}\, Z \ll 1
\label {eq:JiZl} \\
J_{\rm i,\infty} &=& 1.9 \times 10^{-12}\,{\rm C\,m^{-2}\,s^{-1}} \times
Z_1^{-1} \\
& & \times \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \quad {\rm for}\, Z \gg 1
\label {eq:JiZh}
\end{aligned}
$$
We here assumed that the adsorption rate of ions is dominated by the dust component, i.e. that adsorption onto pebbles can be ignored, and scaled the result to dust grains of radius $R_1 = 1\,{\rm \mu m}$ and material density $\rho_\bullet = 10^3\,{\rm kg\,m^{-3}}$. In equation (17) we scaled the ratio of the total mass loading of particles, $Z=Z_1+Z_2$, to the mass loading in dust, $Z_1$, to 10. The ions experience significant heating, and obtain their plateau current density value, above the critical electric field strength of (Okuzumi&Inutsuka,2015)
$$
\begin{aligned}
E_{\rm crit,i} &=& \frac{m_{\rm i} m_{\rm n} \sqrt{3 k_{\rm B} T}}{(m_{\rm
i}+m_{\rm n})^{3/2} e \Delta t_{\rm i}} \\
&\approx& 3.6 \times 10^{-1}\,{\rm V\,m^{-1}} \mathopen{}\mathclose\bgroup\originalleft(\frac{T}{180\,{\rm K}}
\aftergroup\egroup\originalright)^{1/2} \mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm g}}{10^{-7}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \,
.
\end{aligned}
$$
This critical field strength is 1-2 orders of magnitude smaller than the breakdown strength (equation 1) and has the same lin- ear scaling with gas density. Hence the ions are well in their heated regime at the electric field strengths relevant for lightning discharge.
Electrons have larger inertia than neutrals and hence the elec- tron kinetic energy is dominated by the random motion. The critical electric field for electron heating $E_{\rm crit}$ is (Okuzumi &
$$
\begin{aligned}
E_{\rm crit} &=& \sqrt{\frac{6 m_{\rm e}}{m_{\rm n}}} \frac{k_{\rm B} T}{e
\ell_{\rm e}} \\
&\approx& 1.9 \times 10^{-3}\,{\rm V\,m^{-1}} \mathopen{}\mathclose\bgroup\originalleft(
\frac{T}{180\,{\rm K}} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm g}}{10^{-7}\,{\rm
kg\,m^{-3}}} \aftergroup\egroup\originalright) \, .
\end{aligned}
$$
Here $\ell$ is the mean free path of the electrons to collide with neutrals. The limiting values of the electron current at strong electric field
$$
J_{{\rm e},\infty} = \sqrt{\frac{\pi m_{\rm e}}{3 m_{\rm n}}} \frac{e \zeta
n_{\rm n}}{\sigma_1 n_1} = \sqrt{\frac{\pi m_{\rm e}}{3 m_{\rm n}}} J_{\rm
i,\infty} \approx 0.017 J_{\rm i,\infty} \, .
$$
Hence the electron current density plateaus at a value ap- proximately 60 times higher lower than the ion current density. We can therefore neglect the electron current contribution to the neutralising current at the breakdown value of the electric field.
## 2.4. Sedimentation current
Charge separation can be driven by aerodynamical size sorting, if small and large solid particles are charged oppositely. Such op- posite charging can e.g. be obtained by triboelectric charging in collisions between small dust and large pebbles. Desch&Cuzzi (2000) showed that triboelectric e ff ects can charge pebbles up to approximately $10^4$ electron charges. The charge separation in particle components can then be transferred to a spatial charge separation by the di ff erential motion of the particles through the gas. Charged pebbles sedimenting towards the mid-plane carry a current this way. The sedimentation current is
$$
J_{\rm set} = n_2 v_2 q_2 \, .
$$
Here $n_2 = Z_2 \rho_{\rm g}/m_2$ is the pebble number density, $v_2$ is the terminal speed of the pebbles and $q_2$ is the pebble charge. The terminal speed at a height above the mid-plane of one scale- height, $z=H$, is given by
$$
v_2 = \frac{R_2 \rho_\bullet}{c_{\rm s} \rho_{\rm g}} \varOmega^2 H \, .
$$
Here $\varOmega$ denotes the Keplerian frequency at the distance $r$ from the star and $\varOmega^2 H$ is the gravitational acceleration towards the mid-plane. This expression simplifies to
$$
\begin{aligned}
v_2 &=& R_2
\varOmega \frac{\rho_\bullet}{\rho_{\rm g}} = 0.5\,{\rm m\,s^{-1}}\, \mathopen{}\mathclose\bgroup\originalleft(
\frac{R_2}{1\,{\rm mm}} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_\bullet}{10^3\,{\rm
kg\,m^{-3}}} \aftergroup\egroup\originalright) \times \\
&& \mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm g}}{10^{-7}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{-1} \mathopen{}\mathclose\bgroup\originalleft(
\frac{r}{2.5\,{\rm AU}} \aftergroup\egroup\originalright)^{-3/2} \, .
\end{aligned}
$$
That yields a sedimentation current density of
$$
\begin{aligned}
J_{\rm set} &=& \frac{3 Z_2 \varOmega q_2}{4 \pi R_2^2} = 9.7 \times
10^{-20}\,{\rm C\,m^{-2}\,s^{-1}} \mathopen{}\mathclose\bgroup\originalleft(\frac{Z_2}{0.005} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{q_2}{10^4 e} \aftergroup\egroup\originalright) \times \\
&& \mathopen{}\mathclose\bgroup\originalleft(\frac{R_2}{1\,{\rm mm}} \aftergroup\egroup\originalright)^{-2} \mathopen{}\mathclose\bgroup\originalleft(\frac{r}{2.5\,{\rm AU}}
\aftergroup\egroup\originalright)^{-3/2} \, .
\end{aligned}
$$
This is more than eight orders of magnitude lower than the neu- tralising ion current for $Z_1=0.005$ (equation 17). The sedimen- tation current would increase at a higher pebble mass loading $Z_2$, but this is not compatible with the assumption that pebbles fall at the terminal speed at one scale-height above the mid- plane. Increasing the pebble charge or decreasing the pebble size and semi-major axis would also boost the current, but overall it seems hopeless to gain the eight orders of magnitude needed to fight the neutralising gas current.
## 2.5. Radial drift current
Particles also separate aerodynamically by radial drift caused by the radial pressure support of the gas (Weidenschilling, 1977; Nakagawaetal.,1986). Tanakaetal.(2005) provide general so- lutions for the equilibrium velocities of gas and multiple particle components that couple via drag forces. The radial velocity of pebbles relative to the smaller dust grains embedded in the gas becomes
$$
v_{\rm dri} = \frac{2 {\rm St}_2}{Z_2} \Delta v \, .
$$
Here ${\rm St}_2 = R_2 \rho_\bullet/(H \rho_{\rm g})$ is the pebble Stokes number and $\Delta v$ is the sub-Keplerian speed parameter (Youdin & Johansen, 2007). Tanakaetal.(2005) introduced the inverse Stokes num- ber $\varGamma = {\rm St}^{-1}$ and we derived equation (26) in the limit $\varGamma_1 \gg 1$, $\varGamma_2 \gg 2$, $\varGamma_1 \gg
\varGamma_2$and$Z_2 \gg 1$. The relative drift speed at high$Z$is dominated by the outwards motion of gas and dust pushed through the almost stationary pebble component, but given charge neutrality$q_2 n_2 = -(q_1 n_1 + n_{\rm i} -
n_{\rm e})$we can use$q_2 n_2$ to describe the charge density of the current. Inserting typical parameters for the pebbles and the gas we obtain
$$
\begin{aligned}
v_{\rm dri} &=& 0.064\,{\rm m\,s^{-1}}
Z_2^{-1}
\mathopen{}\mathclose\bgroup\originalleft(\frac{R_2}{1\,{\rm mm}} \aftergroup\egroup\originalright) \times \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^{-3}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \times \\
&&
\mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm
g}}{10^{-7}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{-1} \mathopen{}\mathclose\bgroup\originalleft(\frac{\Delta v}{50\,{\rm
m\,s^{-1}}} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(\frac{r}{2.5\,{\rm AU}} \aftergroup\egroup\originalright)^{5/4} \, .
\end{aligned}
$$
The low mass loading limit ($Z_2 \ll 1$) yields the same value and scaling, but without the dependence on $Z_2$. The drift current
| |
170805157/7
|
Here $T_\beta=0.66\,{\rm MeV}$ is the kinetic energy of the positrons, $e$ is the positron charge and $E_{\rm BD}$ is the breakdown value of the electric field. Using the expression for the breakdown electric field from equation (1) we obtain
$$
W^\dagger = 66\,{\rm km}\,\mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm g}}{10^{-7}\,{\rm kg\,m^{-3}}}
\aftergroup\egroup\originalright)^{-1} \mathopen{}\mathclose\bgroup\originalleft(\frac{f_{\rm BD}}{0.1} \aftergroup\egroup\originalright)^{-1} \, .
$$
Here $f_{\rm BD}$ defines the ratio of the actual breakdown field to the nominal value given in Desch & Cuzzi (2000). A one- dimensional pebble filament should have a width comparable to $W$ for e ffi cient conversion of positron energy into lightning discharge; broader filaments do not reach the breakdown field strength at all and narrower filaments reach the breakdown field strength over such short length-scales that positrons only de- posit a minor fraction of their energy in the electric field. Two- dimensional pebble sheets, on the other hand, build up an electric field that has a constant value up to a height comparable to the extent of the sheet itself. Hence a pebble sheet loads energy into the electric field e ffi ciently independently of the width, for any width less than $W^\dagger$. Reaching the breakdown electric field, ions and electrons accelerated by the electric field to neutralise the pebble region must now release an energy similar to the origi- nal kinetic energy of the positrons. This way the energy of the positron emission becomes converted to kinetic energy and heat in the lightning discharge current.
The density of the pebble structure does not a ff ect the criti- cal width found in equation (34). However, the highest positron current is obtained when the column density of the pebble struc- ture is similar to the stopping column of the positrons. Structures of lower column density exhibit a lower current density than the characteristic value given in equation (30), while structures of higher column density self-absorb most of the emitted positrons and therefore have a low e ffi ciency in converting the positron emission to charge separation. The highest positron current com- bined with the highest charging e ffi ciency are therefore obtained when the pebble density obeys
$$
\rho_{\rm p} W = \varSigma_\beta \, .
$$
That gives a pebble density of
$$
\rho_{\rm p} = \frac{\varSigma_\beta}{W} = 5 \times 10^{-5}\,{\rm
kg\,m^{-3}}\, \mathopen{}\mathclose\bgroup\originalleft(\frac{W}{66\,{\rm km}} \aftergroup\egroup\originalright)^{-1} \, .
$$
Higher pebble densities yield ine ffi cient charging and lower pebble densities result in a reduction of the positron current. Combining with the critical width for conversion of positron ki- netic energy to electric potential, from equation (34), yields the optimal solids-to-gas ratio for a pebble filament and the mini- mum solids-to-gas ratio for a pebble sheet as
$$
\frac{\rho_{\rm p}}{\rho_{\rm g}} \approx 500 \frac{f_{\rm BD}}{0.1} \, .
$$
Lower solids-to-gas ratios can still yield e ffi cient conversion of positron energy to heating, but the current emanating from pebble structures of lower densities will be lower than the char- acteristic current. The characteristic current is nevertheless quite high, at $J_\beta=3.6\times10^{-14}\,{\rm C\,m^{-2}\,s^{-1}}$, so even a factor 10 reduction yields currents in excess of the neutralising current at $Z_1=500$.
## 3.3. Simulations of nuclear battery effect
The calculations presented in the previous subsection do not cap- ture the details of the positron emission and absorption process.
Fig.1. The charging rate by positron release in an axisymmetric pebble filament of half-width 20 km. The filament has a peak density of 1000 times the gas density and the background dust density is 0.1 times the gas density. The colored contours show the negative charging rate and the white contour lines show the dust-to-gas ratio. The charging rate has been obtained by a Monte Carlo calculation of the emission and absorp- tion of $10^7$ positrons. A hundred randomly picked positron trajectories are shown in green, released at the green dot and absorbed at the red dot (which happens mostly outside of the plotted region). The filament charges rapidly negative as most of the released positrons are absorbed in the surrounding gas.
Fig.2. The current density of the positrons as a function of the solids-to- gas ratio in the particle filament shown in Figure 1. The current density is above $2 \times 10^{-14}\,{\rm
C\,m^{-2}\,s^{-1}}$through the entire filament. The right axis shows the rate of change of the electric field. The breakdown electric field will be reached in the filament after a characteristic time-scale of only approximately$10^3$ s.
| |
170805157/12
|
when stirred by the streaming instability (Carrera et al., 2015). We consider three values of $f_{\rm g}$ – 1, 0.1 and 0.01 – to parameterise the time evolution of the protoplanetary disc.
The column density of pebbles is thick to positron stopping in the interior of the disc and thin in the exterior. The transi- tion point is at approximately 2.5 AU for a gas density that is 10% of the nominal value of $10^{-7}\,{\rm
kg\,m^{-3}}$ at 2.5 AU. A mid-plane layer thin to positron stopping is ideal for the nuclear battery ef- fect, as the positrons from the entire mid-plane layer, irrespective of any substructure, will be absorbed by the gas outside of the mid-plane. The mid-plane layer is nevertheless much wider, by two orders of magnitude at 2.5 AU, than the optimal width for loading positron energy into the electric field. Therefore sub- structure of width 1% of the mid-plane layer width must be present to reach the breakdown value of the electric field before the positrons are stopped by the electric field. This implies that the mid-plane first charges up by sending positrons to the gas above and below the mid-plane layer, but that the subsequent trapping of positrons in the mid-plane layer charges up smaller and smaller structures until the breakdown value of the electric field is reached. Simulations of the streaming instability show that filaments form on many scales within the mid-plane layer 2015). It remains to be seen whether this density spectrum is prone to charging to the breakdown value of the electric field; we plan to present such calculations in a future publication.
The Roche density in the protoplanetary disc is higher in- side of 3 AU than the constraint proposed by Cuzzi&Alexander (2006) to suppress isotopic fractionation in the heating and sta- bilise the molten chondrules against sublimation. The stricter conditions proposed by Alexanderetal.(2008) to main the high abundance of Na within the chondrules can only be achieved be- low the Roche density well inside of 1 AU in the disc.
## 5. Heating by lightning discharge in a dense pebble region
We consider in this section the heating e ff ect of lightning dis- charge into a dense region of pebbles and dust grains. The solid particles act like a thick, insulating blanket that facilitates high temperatures and long cooling times for the relatively moder- ate discharge powers expected in protoplanetary disc conditions. The discharge is assumed to have a total length $h$ into a cylin- drical pebble filament of width $R_*$. The cylindrical symmetry simplifies the radiative transfer equations significantly. Radiative transfer within a disc-like geometry is more complex, but we ex- pect that heating by lightning penetration into a circumplanetes- imal disc will be similar to the cylindrical geometry, as long as the region heated by lightning maintains approximately cylindri- cal shape.
## 5.1. Radiative equilibrium
We develop the analytical and numerical heating model around a discrete radiative-transfer approach where the cylinder is di- vided into annuli of width $\Delta R$ corresponding to an optical depth interval of unity. The optical depth with respect to radiation is dominated by the small dust component. The mean free path of photons, with wavelengths comparable to or smaller than the dust so that we can consider geometric absorption, is
$$
\begin{aligned}
\Delta R &=& \frac{1}{n_1 \sigma_1} = \frac{(4/3) R_1 \rho_\bullet}{\rho_1}
\\ &\approx& 1.3\,{\rm m}\,\mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{\rm \mu m}\aftergroup\egroup\originalright)
\mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_1}{\rm 10^{-3}\,{\rm kg\,m^{-3}}}\aftergroup\egroup\originalright)^{-1}\, .
\end{aligned}
$$
The outer edge of the $i$ th cylindrical annulus is denoted $R_i = i \Delta R$. We assume that each cylindrical ring radiates like a black body whose radiation is absorbed entirely in its two neighbour- ing annuli. That gives the temperature evolution equation for ring $i$ as
$$
\begin{aligned}
\rho_{\rm p} c_{\rm p} V_i \dot{T}_i &=& -2 \pi (R_{i-1} + R_i) h \sigma_{\rm
SB} T_i^4 + 2 \pi R_{i-1} h \sigma_{\rm SB} T_{i-1}^4 \\ & & + 2 \pi
R_i h \sigma_{\rm SB} T_{i+1}^4 \, .
\end{aligned}
$$
Here $\rho_{\rm p}=\rho_1+\rho_2$ is the total density of solid particles, $c_{\rm p}$ is the specific heat capacity of the particles, $V_i$ is the volume of annulus $i$, $\sigma_{\rm SB}$ is the Stefan-Boltzmann constant and $T_i$ is the temperature of annulus $i$. We then introduce the source function $S_i = \sigma_{\rm SB}
T_i^4$. The energy flux over the edge of annulus$i$ is
$$
\mathcal{F}_i = 2 \pi R_i h S_i - 2 \pi R_i h S_{i+1} \, .
$$
In equilibrium this equals the energy flux release by the light- ning, $\mathcal{F}_i = I E h$. Here $I$ is the current carried by the lightning and $E$ the breakdown electric field. The continuous equation for the energy flux is
$$
\mathcal{F} = -2 \pi R h \Delta R \frac{\partial S}{\partial R} = I E h \, .
$$
We define the optical depth $\tau$ measured inwards from the edge of the pebble region, with a total optical depth of $\tau_\star$. The optical depth thus relates to the radial coordinate $R$ through $R = \Delta R (\tau_\star-\tau)$. Rewriting equation (50) in terms of optical depth $\tau$ we get
$$
\mathcal{F} = 2 \pi h \Delta R (\tau_\star-\tau) \Delta R \frac{\partial S}{\Delta
R \partial \tau} = I E h \, .
$$
Rearranging we find
$$
(\tau_\star-\tau) \frac{\partial S}{\partial \tau} = \frac{I E}{2 \pi \Delta R} \, .
$$
Integration yields
$$
S(\tau) = C - \frac{I E}{2 \pi \Delta R} \ln(\tau_\star-\tau)
$$
The constant $C$ is fixed by requiring that all the released en- ergy is radiated outwards at the $\tau=0$ surface, corresponding to the requirement $ \tau_\star
S(1/2) = I E / (2 \pi \Delta R)$in the outermost annu- lus with$\tau=1/2$. This gives the final source function expression
$$
S(\tau) = \frac{I E}{2 \pi \Delta R} \mathopen{}\mathclose\bgroup\originalleft[\frac{1}{\tau_\star} +
\ln(\tau_\star-1/2)-\ln(\tau_\star-\tau) \aftergroup\egroup\originalright] \, .
$$
The temperature at optical depth $\tau=\tau_\star-1/2$, the approx- imate optical depth of the first annulus, in the limit $\tau_\star \gg 1$, is
$$
\begin{aligned}
T_\star &=& 1754\,{\rm K} \, \mathopen{}\mathclose\bgroup\originalleft(\frac{I}{10^5\,{\rm A}} \aftergroup\egroup\originalright)^{1/4}
\mathopen{}\mathclose\bgroup\originalleft(\frac{\rho_{\rm g}}{10^{-7}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{1/4} \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_\bullet}{10^3\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{-1/4} \\ & &
\times \mathopen{}\mathclose\bgroup\originalleft(\frac{R_1}{10^{-6}\,{\rm m}} \aftergroup\egroup\originalright)^{-1/4} \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_1}{10^{-3}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{1/4} \mathopen{}\mathclose\bgroup\originalleft(
\frac{\ln[2 \tau_\star]}{\ln[200]} \aftergroup\egroup\originalright)^{1/4} \, .
\end{aligned}
$$
We used here a typical current carried by terrestrial lightning and landed fortuitously in the ballpark temperature for chon- drule heating. The characteristic density of protoplanetary discs is many orders of magnitude lower than the density of the ter- restrial atmosphere. The discharge process may nevertheless be
| |
170805157/14
|
Fig.7. Heating rate (red) and cooling rate (blue) of the pebble fila- ment heated by the lightning discharge shown in Figure 6. The curves show, from top to bottom, the heating and cooling rate at optical depth $\tau=0,50,100,150,200$. The cooing rates lie at a few times $10^3\,{\rm K\,hr^{-1}}$ in the first 100 seconds after heating terminates at $t_{\rm heat}=2\times10^{3}\,{\rm s}$ and subsequently fall to below $10^3\,{\rm K\,hr^{-1}}$ over the next 1000 s. These cool- ing rates are in good agreement with the upper range of cooling rates inferred from chondrule textures (Hewins&Radomsky,1990).
$T$ through
the relations
$$
\begin{aligned}
T &=& \frac{e}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e \le e_{\rm s} \, , \\
T &=& T_{\rm s} + \frac{T_{\rm l}-T_{\rm s}}{\rho_2 [\mathcal{L} + c_{\rm p}
(T_{\rm l}-T_{\rm s})]} (e - e_{\rm s}) \quad {\rm for} \quad e_{\rm s} < e
\le e_{\rm l} \, , \\
T &=& \frac{(e-\rho_2 \mathcal{L})}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e >
e_{\rm l} \, .
\end{aligned}
$$
$$
\begin{aligned}
T &=& \frac{e}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e \le e_{\rm s} \, , \\
T &=& T_{\rm s} + \frac{T_{\rm l}-T_{\rm s}}{\rho_2 [\mathcal{L} + c_{\rm p}
(T_{\rm l}-T_{\rm s})]} (e - e_{\rm s}) \quad {\rm for} \quad e_{\rm s} < e
\le e_{\rm l} \, , \\
T &=& \frac{(e-\rho_2 \mathcal{L})}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e >
e_{\rm l} \, .
\end{aligned}
$$
$$
\begin{aligned}
T &=& \frac{e}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e \le e_{\rm s} \, , \\
T &=& T_{\rm s} + \frac{T_{\rm l}-T_{\rm s}}{\rho_2 [\mathcal{L} + c_{\rm p}
(T_{\rm l}-T_{\rm s})]} (e - e_{\rm s}) \quad {\rm for} \quad e_{\rm s} < e
\le e_{\rm l} \, , \\
T &=& \frac{(e-\rho_2 \mathcal{L})}{\rho_2 c_{\rm p}} \quad {\rm for} \quad e >
e_{\rm l} \, .
\end{aligned}
$$
Here $T_{\rm s}$ is the solidus temperature (taken to be 1500 K), $T_{\rm l}$ is the liquidus temperature (taken to be 2000 K), and the internal energies at solidus and liquidus, respectively, are defined as
$$
\begin{aligned}
e_{\rm s} &=& \rho_2 c_{\rm p} T_{\rm s} \, , \\
e_{\rm l} &=& \rho_2 c_{\rm p} T_{\rm l} + \rho_2 \mathcal{L} \, .
\end{aligned}
$$
$$
\begin{aligned}
e_{\rm s} &=& \rho_2 c_{\rm p} T_{\rm s} \, , \\
e_{\rm l} &=& \rho_2 c_{\rm p} T_{\rm l} + \rho_2 \mathcal{L} \, .
\end{aligned}
$$
Absorption and release of latent heat does not change the equi- librium temperature profile, as given in equation (54), but delays both the initial heating and the subsequent cooling.
The width of the lightning discharge channel must be scaled properly to the protoplanetary disc conditions. Hor Robertson(1996) argued that the discharge channel has a width of 1,000-10,000 times the mean free path of the electrons. This scaling relation yields
$$
W_{\rm dis} = 334\,{\rm m}\, \mathopen{}\mathclose\bgroup\originalleft(\frac{f_{\rm dis}}{10^3} \aftergroup\egroup\originalright) \mathopen{}\mathclose\bgroup\originalleft(
\frac{\rho_{\rm g}}{10^{-7}\,{\rm kg\,m^{-3}}} \aftergroup\egroup\originalright)^{-1}
$$
Here $f_{\rm dis}$ denotes the discharge width in units of the mean free path of electrons. We use here $f_{\rm dis}=10^3$ and note that the equilibrium temperature solution of equation (54) is still valid outside the discharge channel, while the temperature within the discharge channel is lower than the equilibrium solution as the decreased heat flux necessitates a lower temperature gradient.
In Figure 6 we show the evolution of the temperature pro- file of the pebble filament for selected times. We assume that the lightning discharge heats for a time-scale of $t_{\rm
heat}=2\times10^3\,{\rm s}$. The temperature climbs steadily during this time, but only ap- proaches the analytical equilibrium solution (equation 54) out- side the discharge channel. The temperature is relatively con- stant within the discharge channel itself and significantly below the analytical solution due to the lower energy flux. Cooling oc- curs relatively isothermally over the pebble filament. In Figure 7 we show the heating and cooling rate at selected regions in the pebble filament as a function of time. Heating is initially rapid, but slows down as the equilibrium temperature is approached. The cooling rate lies at a few times$10^3\,{\rm K\,hr^{-1}}$in the first few hundred seconds after heating terminates and subsequently falls to a few times$10^2\,{\rm K\,hr^{-1}}$ as the temperature drops to 1000 K. These cooling times lie towards the upper end of the values in- ferred from chondrule textures (Hewins & Radomsky, 1990). The cooling time of the filament should broadly follow the heat- ing time, so following equation (60) lower cooling rates could be achieved in filaments of large optical depth or lower gas density.
## 6. Summary
We have demonstrated in this paper that the decay energy of the short-lived radionuclide $^{26}$ Al can be harvested to drive light- ning discharge in protoplanetary discs. Positrons released from dense pebble structures load their energy into the electric field that builds up by the negative charging of these structures. This process is akin to a direct-charging nuclear battery where a ra- dioactive isotope charges a capacitor. The total energy available in the kinetic energy of the released positrons is comparable to the energy required to melt all the solid material in a protoplan- etary disc. Such an immense energy released in lightning dis- charge may be observable in nearby protoplanetary discs with e.g. the ALMA telescopes (Muranushietal.,2015).
The neutralising gas current, driven with an e ffi ciency (con- ductivity) that is primarily set by positron ionisation of hydro- gen molecules, can no longer compete with the positron current when the dust-to-gas ratio of small grains is higher than a few tens. This leaves lightning discharge as the natural path to charge neutralisation, releasing the energy stored in the electric field in thin channels. Eisenhouretal.(1994) demonstrated that heating by infrared laser light yields melting primarily of pebbles larger than 0.1 mm, due to the inability of smaller pebble aggregates in absorbing infrared photons. In this picture, the largest chon- drules represent the largest available chondrule precursors peb- bles, while the lack of small chondrules in most chondrites re- flects ine ffi cient heating of small pebbles by the heat front prop- agating from the lightning channel.
We find that the optimal conditions for heating pebbles by lightning discharge arise in dense particle discs that orbit around planetesimals and protoplanets. The high dust content of such discs implies a low conductivity and the rapid transition from dense pebble environment to the surrounding protoplanetary disc dominated by gas is ideal for driving an outwards directed cur- rent of positrons. Circumplanetesimal discs are appealing from a cosmochemical point of view also, since the prevalent condi- tions there match the high dust densities required for chondrule melting and high abundance of volatiles (Cuzzi & Alexander, 2006; Alexander et al., 2008). The formation and dynamics of such circumplanetesimal discs is nevertheless not well under- stood and should be the focus of future research.
The sedimented mid-plane layer of pebbles in the protoplan- etary disc can also become charged by the emission of positrons, if the mass loading in small dust grains is high enough in the mid-plane (larger than 10 – 50) to reduce the neutralising current.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.