Id: arxiv/2611.21844 Url: https://arxiv.org/abs/2611.21844 Tags: math.OC, cs.LG, stat.ML Score: 0 Title: Bias-Reduced and Parallel Heavy-Tailed Stochastic Convex Optimization in $\ell_p$ Geometries Domain: math.OC Source: arxiv Authors: Zhengyun Nie, Anant Shyam, Brian Bullins Reviews: 0 Subject: math.OC Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-11-16 Page: https://delphi.science/r/arxiv/2611.21844 We consider the fundamental problem of parallel stochastic convex optimization (SCO), whose aim is to approximately minimize a convex objective when given the ability to query a moment-bounded stochastic gradient oracle multiple times in parallel per round. Whereas most previous results have focused on the variance-bounded Euclidean ($\ell_2$-norm) setting, meaning the objective is Lipschitz with respect to the $\ell_2$-norm over an $\ell_2$-bounded feasible set in $\mathbb{R}^d$, in this work we study parallel SCO in $\ell_p$ settings for $2 \le p 0$ with at most $\tilde{O}(d^{\frac{p}{2p +…