{"id":"arxiv/2610.23301","url":"https://arxiv.org/abs/2610.23301","tags":["math.OC"],"score":0,"title":"Acceleration Survives Heavy Tails and Generalized Smoothness","domain":"math.OC","source":"arxiv","authors":["Taha El Bakkali El Kadi","Eduard Gorbunov"],"reviews":0,"subject":"math.OC","summary":"We establish high-probability accelerated convergence guarantees for convex $(L_0,L_1)$-smooth optimization with unbiased stochastic gradients under heavy-tailed noise, assuming a central $α$-th noise moment bounded by $σ^α$, $α\\in(1,2]$. Such guarantees are known under classical $L$-smoothness, but extending them through localization over an $O(R)$-radius ball may introduce an $e^{O(L_1R)}$ factor. We avoid this dependence in two ways. First, a two-phase scheme uses Clip-SGD to reach a low-curvature regime, after which clipped stochastic similar triangles (Clip-SSTM) attains, with…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-16","reactions":{}}