{"id":"arxiv/2609.06555","url":"https://arxiv.org/abs/2609.06555","tags":["math.OC"],"score":0,"title":"Localize, Restart, Accelerate: Stochastic Optimization under Generalized Smoothness","domain":"math.OC","source":"arxiv","authors":["Darina Dvinskikh","Alexander Gasnikov","Aleksandr Lobanov","Ilgam Latypov"],"reviews":0,"subject":"math.OC","summary":"We study stochastic convex optimization under asymmetric \\((L_0,L_1)\\)-generalized smoothness, a model motivated by machine-learning objectives whose local curvature may grow with the gradient norm. We assume an unbiased first-order oracle with additive norm-sub-Gaussian noise. Acceleration is difficult in this setting because momentum may enter regions of much larger curvature, while stochastic gradients cannot reliably certify an unrestricted trajectory. We propose \\textsf{ARC-SG}, a two-phase accelerated method: Phase~I reduces excessively large gradients using a…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-06","reactions":{}}