[arXiv]score: 0.41
Zeroth-Order Algorithm Handles Heavy-Tailed Noise with Optimal Dimension Dependence
May 26, 2026
A clipped two-point gradient estimator within an online-to-nonconvex conversion framework achieves O(d^(p/2(p-1)) δ⁻¹ ε^(-(2p-1)/(p-1))) oracle complexity for finding Goldstein stationary points under heavy-tailed noise, matching best-known dimension dependence.
cs.LG
HOW THIS AFFECTS YOU
●
researcherProvides theoretical guarantees for derivative-free optimization under heavy-tailed noise with dimension scaling that matches convex lower bounds, relevant for black-box ML tuning.