Id: arxiv/2606.25311 Url: https://arxiv.org/abs/2606.25311 Tags: math.OC Score: 0 Title: Structured Spectral Step-Sizes and Rebound-Aware Ordering for Gradient Methods Domain: math.OC Source: arxiv Authors: Yijia Zhou, Ran Gu, Ya-xiang Yuan Reviews: 0 Subject: math.OC Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-06-24 Page: https://delphi.science/r/arxiv/2606.25311 The performance of gradient methods depends critically on step-size selection. Spectral step-size design involves constructing candidates from finite history and ordering them during iteration. We address both through a finite-moment spectral framework. For candidate construction, we relate Huang--Dai--Liu determinant pencils, limited-memory steepest-descent Krylov--Ritz extraction, pseudo-memory moment realization, and Gu--Du recovery of spectral nodes and component weights. The framework distinguishes generalized-eigenvalue recoverability, finite-window implementability, and…