Id: arxiv/2602.13620 Url: https://arxiv.org/abs/2602.13620 Tags: math.OC, math.NA Score: 0 Title: An adaptive framework for first-order gradient methods Domain: math.OC Source: arxiv Authors: Xiaozhe Hu, Sara Pollock, Zhongqin Xue, Yunrong Zhu Reviews: 0 Subject: math.OC Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-02-14 Page: https://delphi.science/r/arxiv/2602.13620 Gradient methods are widely used in optimization problems. In practice, while the smoothness parameter can be estimated utilizing techniques such as backtracking, estimating the strong convexity parameter remains a challenge; moreover, even with the optimal parameter choice, convergence can be slow. In this work, we propose a framework for dynamically adapting the step size and momentum parameters in first-order gradient methods for the optimization problem, without prior knowledge of the strong convexity parameter. The main idea is to use the geometric average of the ratios of successive…