Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Oikonomidis, Konstantinos, Quan, Jan, Antonakopoulos, Kimon, Silveti-Falls, Antonio, Cevher, Volkan, Patrinos, Panagiotis
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911673711853568
author Oikonomidis, Konstantinos
Quan, Jan
Antonakopoulos, Kimon
Silveti-Falls, Antonio
Cevher, Volkan
Patrinos, Panagiotis
author_facet Oikonomidis, Konstantinos
Quan, Jan
Antonakopoulos, Kimon
Silveti-Falls, Antonio
Cevher, Volkan
Patrinos, Panagiotis
contents In this work, we develop proximal preconditioned gradient methods with a focus on spectral gradient methods providing a proximal extension to the Muon and Scion optimizers. We introduce a family of stochastic algorithms that can handle a wide variety of convex and nonconvex constraints and study its convergence under heavy-tailed noise, through a novel analysis tailored to the geometry of the proposed methods. We further propose a variance-reduced version, which achieves faster convergence under standard noise assumptions. Finally, we show that the polynomial iterations used in Muon are more accurately captured by a nonlinear preconditioner than by the ideal matrix sign, leading to a convergence analysis that more faithfully reflects practical implementations.
format Preprint
id arxiv_https___arxiv_org_abs_2605_11850
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives
Oikonomidis, Konstantinos
Quan, Jan
Antonakopoulos, Kimon
Silveti-Falls, Antonio
Cevher, Volkan
Patrinos, Panagiotis
Optimization and Control
Machine Learning
In this work, we develop proximal preconditioned gradient methods with a focus on spectral gradient methods providing a proximal extension to the Muon and Scion optimizers. We introduce a family of stochastic algorithms that can handle a wide variety of convex and nonconvex constraints and study its convergence under heavy-tailed noise, through a novel analysis tailored to the geometry of the proposed methods. We further propose a variance-reduced version, which achieves faster convergence under standard noise assumptions. Finally, we show that the polynomial iterations used in Muon are more accurately captured by a nonlinear preconditioner than by the ideal matrix sign, leading to a convergence analysis that more faithfully reflects practical implementations.
title Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives
topic Optimization and Control
Machine Learning
url https://arxiv.org/abs/2605.11850