MarxistLeninist/repro-lTqCHLllix-artifacts

Reproduction: Online Learning with Recency: Algorithms for Sliding-window Streaming Multi-armed Bandits

• OpenReview ID: lTqCHLllix
• arXiv: 2606.08977
• Challenge: ICML 2026 Reproduction Challenge
• Compute: CPU only, ~5 min total, $0.00

Overview

An independent CPU-scale numerical audit of the paper's five main theoretical claims. All experiments are deterministic (seeded) and reproducible with only numpy and matplotlib.

Per the challenge guide, theorem claims are reproduced as numerical audits: implement the setup, check the stated equalities and inequalities hold, and include controls that relax the theorem's conditions so the property is shown to degrade. This is not a proof replacement. Where the paper quantifies over all algorithms, what is audited is the paper's hard-instance mechanism against a specific natural policy class.

Claims reproduced

Claim	Source	Verdict	Evidence
C1	Thm 2 (weak eps-PE)	verified (correctness + space); sample complexity not tested	success 1.000 >= 1-delta=0.95; peak memory <= ceil(2/eps) = O(1/eps); mutation control falls to 0.460 at 1000x budget cut. Pull count is fixed by construction, so matching the theoretical count is tautological.
C2	Thm 1 (Omega(W) space)	verified (audit, one policy class)	error > 0 for every m < W, exactly 0 at m = W; error at fixed m/W flat in W (0.732 / 0.741 / 0.745 for W = 8/16/32) => threshold linear in W
C3	Lem 4.1/4.2 (strong vs weak)	partial	at bucket width eps, weak all-times success 0.900 -> 0.067 vs strong 1.000 -> 0.250 as n/W goes 10 -> 320; weak per-query failure 2x strong throughout. Neither meets 1-delta: a flat per-query floor (0.0010) reveals eps-discretization error that more samples cannot remove.
C4	Thm 4 (regret/memory)	positive half verified; lower-bound half inconclusive	retain-best = 0 regret at every budget incl. memory 1; FIFO = rate ~0.5 at every budget incl. memory = W. The instance cannot test the lower bound: a 1-slot policy is optimal there.
C5	Thm 3 (epoch regret)	upper bound verified; lower bound vacuous	fitted exponent 0.561 (all) / 0.441 (tail) vs theory 0.5; R_T >= T/(64W^2) satisfied with 2.6e3-6.5e4 slack

Nothing in this reproduction contradicts the paper. Every claim is either supported or untested; none is falsified.

Corrections to the first pass

An earlier pass reported 4 verified / 1 partial. Re-analysis found two defects in that analysis, both of which changed a verdict:

• C4 retention bug. regret_everlasting_fifo evicts via store.pop(0). Arm 0 is the everlasting best arm and the first element, so it is discarded as soon as the buffer fills at every memory budget, mem = W included. The memory-W run gives rate 0.4986, not 0; the "0 regret" figure came from the separate regret_everlasting_oracle. Sweeping two eviction rules over the same budgets shows regret here is set by the retention rule, not the memory budget.
• C5 units error. lower_bound_rate_thm3 returns T/(64W^2), a bound on total regret, but it was compared against rate = R_T/T. Corrected on totals the bound holds everywhere; the apparent violation was a units mismatch, not a finding.
• C3 masking. The first pass used buckets of width eps/2, finer than Algorithm 1 specifies, which drove both budgets to 1.000 success at all n/W and hid the separation.

Structure

src/
  swmab.py           core algorithms and simulation framework
  experiment.py      first-pass consolidated experiment
  reanalysis.py      corrected C4/C5 analysis + C1 mutation control
  c3_control.py      C3 at both bucket widths
  plot_results.py    figures from experiment.py output
  plot_reanalysis.py figures from the corrected analyses
  build_logbook.py   builds the Trackio logbook from stored JSON
  make_bundle.py     assembles this bundle with a hashed manifest
tests/
  smoke.py           fast sanity checks
outputs/
  results.json       first-pass results
  reanalysis.json    corrected C4/C5 + C1 control
  c3_control.json    C3 at both bucket widths
figures/
  fig_c*.png         first-pass figures
  fig_r*.png         corrected-analysis figures

Quick start

pip install numpy matplotlib
export OMP_NUM_THREADS=1     # tiny workloads; BLAS threads make them much slower

python tests/smoke.py          # ~10 s
python src/experiment.py       # ~90 s   -> outputs/results.json
python src/reanalysis.py       # ~87 s   -> outputs/reanalysis.json
python src/c3_control.py       # ~129 s  -> outputs/c3_control.json
python src/plot_results.py     # figures/fig_c*.png
python src/plot_reanalysis.py  # figures/fig_r*.png

Every figure is regenerated from the stored JSON, so plots and tables cannot drift apart. Seeds are fixed per trial; reruns reproduce the reported numbers exactly.

Dependencies

Python >= 3.10, numpy, matplotlib.

License

Reproduction artifacts are provided for academic evaluation under the ICML 2026 Reproduction Challenge terms.