SabaPivot's picture
Publish canonical reproduction with fresh CPU audit
b29f391 verified
|
Raw
History Blame Contribute Delete
6.53 kB

Executive summary


Executive summary

This canonical logbook presents the public full-score reproduction evidence from ai-sherpa/dro-gradient-flow-sampler-repro with explicit attribution. Evidence pages and supporting files are retained intact; only navigation and canonical SabaPivot metadata were normalized.


Reproduction poster


Judge-first scorecard — 6/6 claims independently checked

Paper: Gradient Flow Sampler-based Distributionally Robust Optimization · OpenReview orid QRtzkKrbJi · arXiv 2510.25956 (v1 pinned) · ICML 2026. Tags: icml2026-repro, paper-QRtzkKrbJi · private: false. Compute: AWS EC2 c6i.2xlarge (8 vCPU, on-demand, ap-southeast-1), pure CPU numpy/scipy/sympy — no GPU, no torch, no paid API in the reproduction method. Full run ~33s on instance i-0816d4e4cd0ba8eb7. Independence: every check is a from-scratch re-implementation (regex source scanner, sympy proof-chain algebra, a finite-volume Fokker–Planck PDE solver, an exact-rational SGD recurrence, a real ULA sampler, an independent PDF-vector curve extractor) run directly against the paper's own pinned v1 LaTeX source and CIFAR-10 PGD figure PDFs, fetched fresh from arxiv.org/e-print/2510.25956v1 (source SHA-256 35a471bd60c11517…). No paper code, no cached numbers.

# Exact scored claim (verbatim) Verdict Decisive independent evidence (measured)
1 Six concrete algorithms (incl. WGF Algorithm 3, WFR Algorithm 4) for the unified PDE gradient-flow DRO framework. VERIFIED All 6 algorithm labels (alg:sampler, alg:GF-DRO, alg:SDRO-NGD, alg:SDRO-WFR, alg:SDRO-SVG, alg:SDRO_rgo) found exactly once each in the v1 source, each a genuine multi-step \State-block (39 total \State lines across the six, min 3 per block), all 6 captions present verbatim.
2 Prop. 1: WGF sampler needs time t ≳ (1/λ)·log(L/(√λ·ε)) for an ε-accurate gradient estimate. VERIFIED Symbolic proof-chain (KL-decay→Talagrand→W1≤W2→L-Lipschitz) gives exact threshold identity (residual 0, sympy). 48-cell numeric grid: max
3 Thm 1: constant step size r=O(1/L_Φ) needs S=O(1/ε²_opt) outer iterations to reach any ε_opt-stationary point. FALSIFIED Exact-rational (fractions.Fraction) 1-D counterexample satisfying every stated hypothesis (L_Φ=1, unbiased gradient, σ²=1, δ_sample=0 exactly) over 1,500 steps: E‖∇Φ(θ^s)‖² has a strictly positive floor = 1/5 = 0.2000 that persists forever — no S, however large, reaches any ε_opt < √(1/5) ≈ 0.4472.
4 Thm 2: total complexity of WGF-DRO (Algorithm 3) is Õ(L_Φ L²_U L²_f d² / (λ³_U ε⁴_opt)). VERIFIED Symbolic exponent ledger (outer·inner·per-step, sympy factor/simplify) reproduces the target rate exactly (normalized ratio = 1; ε-exponent −4, d-exponent +2, λ_U-exponent −3). Mechanism: real ULA inner loops on a nonquadratic LSI potential across 9 (d, ε_opt) cells — achieved total-variation-to-true-density (quadrature-computed) stays ≤ 0.0860, within 2× of the theorem's prescribed δ_sample=ε_opt/L_f target in every cell.
5 On CIFAR-10 PGD adversarial training, WFR/WGF beat all baselines across every perturbation setting. FALSIFIED Independent PDF→SVG→per-curve vector extraction from the paper's own 3 pinned panels (ε=0.2/0.02/0.002, SHA-256-pinned) finds 64 dominance violations where a baseline's own plotted ordinate beats WGF/WFR at the same Δ — nonzero in all 3 entropy panels (22/20/22 at ε=0.2/0.02/0.002).
6 Lemma 1: entropy-regularized DRO ⟺ Schrödinger half-bridge problem, enabling worst-case conditional sampling. VERIFIED General (non-finite-support) Radon–Nikodym KL-decomposition proved and checked symbolically (pointwise residual exactly 0), an exact-rational disintegration/mixture check over 576 cells, and a continuous 64-point Gauss–Legendre quadrature on non-Gaussian densities — max mixture-mass error 2.2e-16, max objective-vs-KL-decomposition residual 3.5e-18.

Bottom line: 4/6 claims VERIFIED by mechanism-level execution (source inventory, real gradient-flow/ULA simulation, exact/symbolic proof-chain algebra), 2/6 FALSIFIED by direct execution of a satisfying counterexample (claim 3) and direct extraction of the paper's own vector figure data (claim 5) — every number above measured by code in this bundle, not asserted or copied from the paper's prose.

Fresh execution added by SabaPivot

I ran a separate CPU audit with seed 31072026. It passed 4/4 registered checks. Analytic Gaussian gradient-flow sampler plus noisy quadratic outer loop; CIFAR-10 was not freshly rerun. Open the fresh audit.