Id: arxiv/2601.19381 Url: https://arxiv.org/abs/2601.19381 Tags: math.OC, cs.DC Score: 0 Title: Decentralized Nonsmooth Nonconvex Optimization with Client-Data Sampling Domain: math.OC Source: arxiv Authors: Cheng Chen, Xinyan Chen, Weiguo Gao, Luo Luo Reviews: 0 Subject: math.OC Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-01-27 Page: https://delphi.science/r/arxiv/2601.19381 This paper studies decentralized nonsmooth nonconvex optimization with Lipschitz-continuous local functions. We propose an efficient stochastic first-order method with a novel client-data sampling strategy in which the mini-batch sizes of the stochastic gradients on all clients follow a multinomial distribution. We show that our method can find a $(δ,ε)$-Goldstein stationary point on each client with the best-known dependence on $ε$~and $δ$ in both computational and communication complexity. Moreover, our results tighten the upper complexity bounds with respect to the Lipschitz parameter,…