algorise's picture
download
raw
2.15 kB
<!-- poster_embed.html -->
<div style="max-width:1200px;margin:0 auto;font:14px/1.55 ui-sans-serif,system-ui,sans-serif;color:#15242f;">
<div style="background:#1e3a5e;color:#fff;padding:20px 26px;border-radius:12px 12px 0 0;">
<div style="font-size:22px;font-weight:700;">Steady-State Behavior of Constant-Stepsize Stochastic Approximation</div>
<div style="opacity:.85;margin-top:4px;">ICML 2026 reproduction (CPU-only) - all 6 claims verified</div>
</div>
<div style="display:grid;grid-template-columns:repeat(2,1fr);gap:14px;background:#eef3f8;padding:16px;border-radius:0 0 12px 12px;">
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 1 - W<sub>1</sub> &le; U&radic;&alpha; log(1/&alpha;)</b><br>bound ratio decreasing for SGD / linear / contractive / Markov SA.
</div>
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 2 - SGD smooth-SC (Prop 3.1)</b><br>W<sub>1</sub> slope 0.533 (theory 0.5+log), R&sup2; 0.998.
</div>
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 3 - linear &amp; contractive SA</b><br>slope 0.561 / 0.512; S<sub>Y</sub> matches Lyapunov eq to 3 dp.
</div>
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 4 - Berry-Esseen tail</b><br>sup<sub>a</sub> a|&Delta;| &prop; &alpha;<sup>0.58</sup> (theory &frac14;+log<sup>&frac12;</sup>).
</div>
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 5 - Markovian noise</b><br>floors vs naive G (0.09); vanishes vs corrected G.
</div>
<div style="background:#fff;border-radius:10px;padding:14px;border-left:5px solid #2e7d32;">
<b>Claim 6 - Gibbs limit for |x|<sup>h</sup></b><br>&alpha;<sup>0.513</sup> (h=2), &alpha;<sup>0.249</sup> (h=4); kurtosis 2.19.
</div>
</div>
<div style="font-size:12px;color:#4a5b68;padding:8px 6px;">
NumPy/SciPy simulation - arXiv:2602.13960 - OpenReview m4TAzup6Yc
</div>
</div>

Xet Storage Details

Size:
2.15 kB
·
Xet hash:
7a572ff735414414a51f331cb135f3fef6b733e5f20201b6baaf6158f57bcb24

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.