Zeroth-Order Non-smooth Non-convex Optimization via Gaussian Smoothing

Anik Kumar Paul, Shalabh Bhatnagar · 2025

Read Zeroth-Order Non-smooth Non-convex Optimization via Gaussian Smoothing by Anik Kumar Paul, Shalabh Bhatnagar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-11073.txt.

Description

Paul and Bhatnagar optimize the hardest kind of objective: rough, kinked, and offering no gradients, only noisy values. Such problems arise whenever a system can be tested but not differentiated. The classic tools stall because nonsmooth functions lack the Taylor expansions that smooth analysis leans on.

Their method blurs the objective with Gaussian noise to approximate its Clarke subgradients. It then runs two coupled update loops: a fast one tracks the approximate subgradient, and a slow one moves the solution. They prove the expectation of the smoothed subgradient sits within a bounded error of the true one, and that the iterates converge, almost surely, near the set of stationary points. The authors call it the first gradient-free method with this guarantee for constrained nonsmooth, nonconvex problems.

Catalog record

CREATORAnik Kumar Paul, Shalabh Bhatnagar
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-11073

Browse the full library · About the Tower of Babel