{"id":"arxiv/2610.41251","url":"https://arxiv.org/abs/2610.41251","tags":["math.OC","stat.ML"],"score":0,"title":"Stochastic Mirror Descent under Dependent Observations","domain":"math.OC","source":"arxiv","authors":["Gersende Fort","Anatoli Juditsky","Jiaming Liang"],"reviews":0,"subject":"math.OC","summary":"We study the non-asymptotic convergence of Stochastic Mirror Descent (SMD) algorithm when the noisy first-order oracle is driven by dependent observations. Departing from the classical i.i.d. assumption underlying most existing guarantees, we introduce a flexible ``forgetting\" condition that accommodates geometrically or polynomially dependent sequences. Under this condition we derive explicit finite-sample accuracy bounds for the averaged SMD iterates, recovering the classical $O(1/T)$ and $O(1/\\sqrt{T})$ rates in the large- and small-initial-error regimes. We instantiate the general…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-30","reactions":{}}