{"id":"arxiv/2605.02127","url":"https://arxiv.org/abs/2605.02127","tags":["math.OC","cs.LG"],"score":0,"title":"A Parameter-Free First-Order Algorithm for Non-Convex Optimization with $\\tilde{\\mkern1mu O}(ε^{-5/3})$ Global Rate","domain":"math.OC","source":"arxiv","authors":["Sichao Xiong","Sadok Jerad","Coralia Cartis"],"reviews":0,"subject":"math.OC","summary":"We introduce PF-AGD, the first parameter-free, deterministic, accelerated first-order method to achieve $O(ε^{-5/3}\\log(1/ε))$ oracle complexity bound when minimizing sufficiently smooth, non-convex functions; this is the best-known bound for first-order methods on smooth non-convex objectives. Unlike existing methods possessing this rate that require a priori knowledge of smoothness constants, we use an adaptive backtracking scheme and a gradient-based restart mechanism to estimate local curvature. This yields a practical algorithm that matches best-known theoretical rates. Empirically,…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-05-04","reactions":{}}