Near-Optimal Convergence of Accelerated Gradient Methods under Generalized and $(L_0, L_1)$-Smoothness
Abstract
We study first-order methods for convex optimization problems with functions satisfying the recently proposed -smoothness condition which generalizes the -smoothness and -smoothness. While accelerated gradient descent AGD is known to reach the optimal complexity under -smoothness, where is an error tolerance and is the distance between a starting and an optimal point, existing extensions to -smoothness either incur extra dependence on the initial gradient, suffer exponential factors in , or require costly auxiliary sub-routines, leaving open whether an AGD-type rate is possible for small-, even in the -smoothness case. We resolve this open question. Leveraging a new Lyapunov function and designing new algorithms, we achieve oracle complexity for small- and virtually any . For instance, for -smoothness, our bound is provably optimal in the small- regime and removes all non-constant multiplicative factors present in prior accelerated algorithms.
Keywords
Cite
@article{arxiv.2508.06884,
title = {Near-Optimal Convergence of Accelerated Gradient Methods under Generalized and $(L_0, L_1)$-Smoothness},
author = {Alexander Tyurin},
journal= {arXiv preprint arXiv:2508.06884},
year = {2026}
}