metadata
license: mit
tags:
- icml2026-repro
- single-index-models
- gradient-descent
- theory-reproduction
Reproduction results — "Full-Batch Gradient Descent Outperforms One-Pass SGD" (ICML 2026 #26332)
Raw and aggregated results for an independent reproduction of arXiv:2602.02431 / OpenReview QItZDBVCT0.
All numbers were produced locally on 2x NVIDIA RTX 4000 Ada with the scripts in the
reproduction logbook (scripts/sim.py, scripts/sweep_spherical.py,
scripts/sweep_online_sgd.py, scripts/sweep_squared_gd.py, scripts/spectral_audit.py,
scripts/thm32_bound.py).
| file | contents |
|---|---|
sweep_quad.csv |
full-batch spherical GD, σ(z)=z², per-run squared overlap (Claim 1) |
sweep_trunc.csv |
full-batch spherical GD, σ(z)=min(z²,8) (Claims 2/5) |
sweep_smooth.csv |
same with the C^∞ truncation of paper eq. (3.10) (robustness) |
sweep_online_{trunc,quad}.csv |
one-pass spherical SGD baseline over a grid of η=c/d (Claim 5) |
gd_trunc_r2_{traj,summary}.csv |
squared-loss full-batch GD, r₀=d⁻² (Claims 3/4) |
gd_trunc_r15_{traj,summary}.csv |
squared-loss full-batch GD, r₀=d⁻¹⁵ (exact Theorem 4.1 setting) |
gd_quad_r2_*, gdd_* |
untruncated-activation and δ-sweep controls |
gd_trunc_eta*_* |
learning-rate sweep for the O(log d / η) phase-1 length |
audit_{spectrum,uniform_bbp,indicator}.csv |
numerical audits of the spectral statements |
thm32_bound.csv |
audit of the Theorem 3.2 deficit bound 1 − C(e^{−M/2} + (d/n)^{1/5}) |
agg_*.csv, thresholds.csv, threshold_fits.csv |
aggregated curves and log-d fits |