JG1310's picture
Reproduction bundle: QOT localization bounds & affine case (kcnuX4xEpL)
5953244 verified
|
Raw
History Blame Contribute Delete
4.06 kB
metadata
license: mit
tags:
  - icml2026-repro
  - optimal-transport
  - reproduction-bundle

Reproduction bundle — Quadratically Regularized Optimal Transport: Localization Bounds and Affine Case Analysis

Paper: Quadratically Regularized Optimal Transport: Localization Bounds and Affine Case Analysis (ICML 2026, OpenReview kcnuX4xEpL, arXiv 2605.24644).

This bundle contains everything needed to re-run the reproduction described in the Trackio logbook JG1310/repro-quadratically-regularized-optimal-transport-localization-bounds-and-affine.

Both extracted claims are proven theorems. The reproduction is therefore an independent numerical audit (per the challenge guide's theory-paper clause), not a GPU benchmark:

  • Claim 1 (Theorem 3.3): directed-Hausdorff support cannot concentrate around the Monge graph faster than ε^(1/(d+2)) — a lower bound.
  • Claim 2 (Theorem 3.7): in the affine Brenier regime (incl. Gaussian→Gaussian) a sharp pointwise tube bound of order ε^(1/(d+2)) holds — a matching upper bound.

Contents

Path What it is
DERIVATIONS.md Step-by-step re-derivation of both theorems (D1–D6c), each paired with an executable check
scripts/derivation_checks.py 13 numerical audits of the algebra/lemmas (CHK-D1..D6c), CPU-only, <0.5 s
scripts/exp01_affine_scaling.py Full-scale affine-scaling diagnostic (paper §5 / Appendix B.2–B.6)
specs/exp01_affine_scaling.md Exact experiment spec (parameters, solvers, acceptance gate)
gates.py Structural acceptance gate for results/exp01.json
results/exp01.json Full-scale run output: 80 records, 8 summary rows (N=M=2000, R=10, d∈{100,200,500,1000})
results/GATE_REPORT.txt gates.py --full report (21/21 PASS)
results/derivation_checks.log derivation_checks.py output (13/13 PASS)
results/DRIVER_REPORT.json Driver status ({"exp01":"PASS"})
logs/exp01.log Verbatim stderr trace of the 800-solve full run
BRIEF_WRITER.md, STATE.md Planner's claim→evidence map and in-regime honesty notes

How to re-run

python3 -m venv .venv && . .venv/bin/activate
pip install numpy scipy joblib          # derivation_checks also uses scipy

# 1) Derivation audit (seconds, CPU): re-verifies every algebraic step of both theorems.
python3 scripts/derivation_checks.py     # expect "13/13 PASS", exit 0

# 2) Full-scale empirical diagnostic (~8.5 h on 8 cores; d-independent 2000x2000 solves).
JOB_CORES=8 python3 scripts/exp01_affine_scaling.py     # writes results/exp01.json (+ work/ checkpoints)

# 3) Structural gate on the produced results file.
python3 gates.py --full                  # expect "ALL PASS (21/21)"

# Smoke test only (minutes): a d=10, N=M=200, R=2 toy of the same code path.
python3 scripts/exp01_affine_scaling.py --toy

exp01 is checkpointed per (d, seed) unit under work/; a restart skips completed units, so a kill loses at most one in-flight unit.

Outcome

Both claims VERIFIED via the derivation audit (13/13 machine-precision checks, incl. an in-regime confirmation of the shared ε^(2/(d+2)) value-gap rate at d=1). The full-scale exp01 diagnostic reproduces the paper's qualitative Figure-1 pattern (β̂ decreasing with d, tight two-solver agreement, correct order of magnitude, monotonically increasing RelErr) with one honestly documented quantitative discrepancy: the RelErr zero-crossing occurs between d=100 and d=200 here, versus the paper's reported crossing between d=500 and d=1000. Because the theorem's strict asymptotic regime (ε ≤ ε₀ ≈ 10^(−152) at d=100) is unreachable in double precision for any tested d, the entire β̂(d) curve is pre-asymptotic — the discrepancy reflects how far pre-asymptotic effects push each implementation, not a violation of the theorems (whose correctness is established by the derivation audit).