{"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","summary":"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…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-24","reactions":{}}