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 Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-07-31 Page: https://delphi.science/r/arxiv/2610.30143 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.