{"id":"arxiv/2610.30143","url":"https://arxiv.org/abs/2610.30143","tags":["math.OC"],"score":0,"title":"The Last Iterate of the Projected Stochastic Subgradient Method Pays a Logarithm, Even in Dimension One","domain":"math.OC","source":"arxiv","authors":["Guglielmo Beretta","Tommaso Cesari","Roberto Colomboni"],"reviews":0,"subject":"math.OC","summary":"We prove that the projected stochastic subgradient method with the standard anytime stepsizes $η_t = γ/\\sqrt t$ can suffer expected last-iterate optimization error of order $(\\log n)/\\sqrt n$ already on the real line. Combined with the classical in-expectation upper bound, this establishes the sharp worst-case order of the projected stochastic subgradient method.","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-31","reactions":{}}