{"id":"arxiv/2604.08236","url":"https://arxiv.org/abs/2604.08236","tags":["math.OC"],"score":0,"title":"Improved Convergence for Decentralized Stochastic Optimization with Biased Gradients","domain":"math.OC","source":"arxiv","authors":["Qing Xu","Yiwei Liao","Wenqi Fan","Xingxing You","Songyi Dian"],"reviews":0,"subject":"math.OC","summary":"Decentralized stochastic optimization has emerged as a fundamental paradigm for large-scale machine learning. However, practical implementations often rely on biased gradient estimators arising from communication compression or inexact local oracles, which severely degrade convergence in the presence of data heterogeneity. To address the challenge, we propose Decentralized Momentum Tracking with Biased Gradients (Biased-DMT), a novel decentralized algorithm designed to operate reliably under biased gradient information. We establish a comprehensive convergence theory for Biased-DMT in…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-04-09","reactions":{}}