Non-Convex Self-Concordant Functions: Practical Algorithms and Complexity Analysis

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Goldfarb, Donald, Lai, Lexiao, Lin, Tianyi, Zhang, Jiayu
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908938208804864
author Goldfarb, Donald
Lai, Lexiao
Lin, Tianyi
Zhang, Jiayu
author_facet Goldfarb, Donald
Lai, Lexiao
Lin, Tianyi
Zhang, Jiayu
contents We extend the standard notion of self-concordance to non-convex optimization and develop a family of second-order algorithms with global convergence guarantees. In particular, two function classes -- \textit{weakly self-concordant} functions and \textit{$F$-based self-concordant} functions -- generalize the self-concordant framework beyond convexity, without assuming the Lipschitz continuity of the gradient or Hessian. For these function classes, we propose a regularized Newton method and an adaptive regularization method that achieve an $ε$-approximate first-order stationary point in $O(ε^{-2})$ iterations. Equipped with an oracle capable of detecting negative curvature, the adaptive algorithm can further attain convergence to an approximate second-order stationary point. Our experimental results demonstrate that the proposed methods offer superior robustness and computational efficiency compared to cubic regularization and trust-region approaches, underscoring the broad potential of self-concordant regularization for large-scale and neural network optimization problems.
format Preprint
id arxiv_https___arxiv_org_abs_2511_15019
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Non-Convex Self-Concordant Functions: Practical Algorithms and Complexity Analysis
Goldfarb, Donald
Lai, Lexiao
Lin, Tianyi
Zhang, Jiayu
Optimization and Control
90C26, 90C30, 65K05, 49M15
We extend the standard notion of self-concordance to non-convex optimization and develop a family of second-order algorithms with global convergence guarantees. In particular, two function classes -- \textit{weakly self-concordant} functions and \textit{$F$-based self-concordant} functions -- generalize the self-concordant framework beyond convexity, without assuming the Lipschitz continuity of the gradient or Hessian. For these function classes, we propose a regularized Newton method and an adaptive regularization method that achieve an $ε$-approximate first-order stationary point in $O(ε^{-2})$ iterations. Equipped with an oracle capable of detecting negative curvature, the adaptive algorithm can further attain convergence to an approximate second-order stationary point. Our experimental results demonstrate that the proposed methods offer superior robustness and computational efficiency compared to cubic regularization and trust-region approaches, underscoring the broad potential of self-concordant regularization for large-scale and neural network optimization problems.
title Non-Convex Self-Concordant Functions: Practical Algorithms and Complexity Analysis
topic Optimization and Control
90C26, 90C30, 65K05, 49M15
url https://arxiv.org/abs/2511.15019