{"id":"arxiv/2609.09152","url":"https://arxiv.org/abs/2609.09152","tags":["math.OC","cs.LG"],"score":0,"title":"Silver Rate Is (Almost) Optimal for Gradient Descent","domain":"math.OC","source":"arxiv","authors":["Yuhan Ye","Kaizhao Liu"],"reviews":0,"subject":"math.OC","summary":"We study how far gradient descent (GD) can be accelerated by predetermined stepsizes in smooth convex optimization. Writing $p_{\\mathrm{sil}}=\\log_2(1+\\sqrt{2})$, we prove an $Ω\\left(n^{-p_{\\mathrm{sil}}-O(\\sqrt{\\log\\log n/\\log n})}\\right)$ non-anytime lower bound. In the anytime setting, every infinite schedule has infinitely many horizons with error $Ω\\left(n^{-\\frac{2p_{\\mathrm{sil}}}{1+p_{\\mathrm{sil}}}-O(\\sqrt{\\log\\log n/\\log n})}\\right)$. Together with the silver-schedule upper bound [Altschuler and Parrilo, 2025] and the anytime upper bound [Zhang et al., 2025], our results determine…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-08","reactions":{}}