{"id":"arxiv/2606.10110","url":"https://arxiv.org/abs/2606.10110","tags":["math.OC"],"score":0,"title":"The Dual Averaging Power-Prox Method with Application to Heavy-Tail Incremental Gradient","domain":"math.OC","source":"arxiv","authors":["Yuan Gao","Jeremy Rack","Sebastian U. Stich"],"reviews":0,"subject":"math.OC","summary":"We study finite-sum composite optimization under two departures from classical stochastic gradient descent theory that are central in practice: incremental gradient access and heavy-tailed gradient noise. Specifically, we consider fixed cyclic passes over component gradients and assume that, at the optimum, component gradients have a bounded $q$-th centralized moment for some $q\\in(1,2]$. This setting is much closer to modern ML training practice than the assumptions used in classical SGD theory, yet its theoretical understanding remains limited. We propose a Dual Averaging Power-Prox method…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-08","reactions":{}}