Salvato in:
Dettagli Bibliografici
Autori principali: Muşat, Andreea-Alexandra, Boumal, Nicolas
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2507.13804
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912766437097472
author Muşat, Andreea-Alexandra
Boumal, Nicolas
author_facet Muşat, Andreea-Alexandra
Boumal, Nicolas
contents It is known that gradient descent (GD) on a $C^2$ cost function generically avoids strict saddle points when using a small, constant step size. However, no such guarantee existed for GD with a line-search method. We provide one for a modified version of the standard Armijo backtracking method with generic, arbitrarily large initial step size. The proof underlines the double role of the Luzin $N^{-1}$ property for the iteration maps, and allows to forgo the habitual Lipschitz gradient assumption. We extend this to the Riemannian setting (RGD), assuming the retraction is real analytic (though the cost function still only needs to be $C^2$). In closing, we also improve guarantees for RGD with a constant step size in some scenarios.
format Preprint
id arxiv_https___arxiv_org_abs_2507_13804
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Gradient descent avoids strict saddles with a simple line-search method too
Muşat, Andreea-Alexandra
Boumal, Nicolas
Optimization and Control
Numerical Analysis
Dynamical Systems
90C30 (Primary) 65K05, 37C75, 58K05 (Secondary)
It is known that gradient descent (GD) on a $C^2$ cost function generically avoids strict saddle points when using a small, constant step size. However, no such guarantee existed for GD with a line-search method. We provide one for a modified version of the standard Armijo backtracking method with generic, arbitrarily large initial step size. The proof underlines the double role of the Luzin $N^{-1}$ property for the iteration maps, and allows to forgo the habitual Lipschitz gradient assumption. We extend this to the Riemannian setting (RGD), assuming the retraction is real analytic (though the cost function still only needs to be $C^2$). In closing, we also improve guarantees for RGD with a constant step size in some scenarios.
title Gradient descent avoids strict saddles with a simple line-search method too
topic Optimization and Control
Numerical Analysis
Dynamical Systems
90C30 (Primary) 65K05, 37C75, 58K05 (Secondary)
url https://arxiv.org/abs/2507.13804