Revisiting Frank-Wolfe for Structured Nonconvex Optimization

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Maskan, Hoomaan, Hou, Yikun, Sra, Suvrit, Yurtsever, Alp
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866915640837668864
author Maskan, Hoomaan
Hou, Yikun
Sra, Suvrit
Yurtsever, Alp
author_facet Maskan, Hoomaan
Hou, Yikun
Sra, Suvrit
Yurtsever, Alp
contents We introduce a new projection-free (Frank-Wolfe) method for optimizing structured nonconvex functions that are expressed as a difference of two convex functions. This problem class subsumes smooth nonconvex minimization, positioning our method as a promising alternative to the classical Frank-Wolfe algorithm. DC decompositions are not unique; by carefully selecting a decomposition, we can better exploit the problem structure, improve computational efficiency, and adapt to the underlying problem geometry to find better local solutions. We prove that the proposed method achieves a first-order stationary point in $O(1/ε^2)$ iterations, matching the complexity of the standard Frank-Wolfe algorithm for smooth nonconvex minimization in general. Specific decompositions can, for instance, yield a gradient-efficient variant that requires only $O(1/ε)$ calls to the gradient oracle. Finally, we present numerical experiments demonstrating the effectiveness of the proposed method compared to other projection-free algorithms.
format Preprint
id arxiv_https___arxiv_org_abs_2503_08921
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Revisiting Frank-Wolfe for Structured Nonconvex Optimization
Maskan, Hoomaan
Hou, Yikun
Sra, Suvrit
Yurtsever, Alp
Optimization and Control
Machine Learning
90C26
We introduce a new projection-free (Frank-Wolfe) method for optimizing structured nonconvex functions that are expressed as a difference of two convex functions. This problem class subsumes smooth nonconvex minimization, positioning our method as a promising alternative to the classical Frank-Wolfe algorithm. DC decompositions are not unique; by carefully selecting a decomposition, we can better exploit the problem structure, improve computational efficiency, and adapt to the underlying problem geometry to find better local solutions. We prove that the proposed method achieves a first-order stationary point in $O(1/ε^2)$ iterations, matching the complexity of the standard Frank-Wolfe algorithm for smooth nonconvex minimization in general. Specific decompositions can, for instance, yield a gradient-efficient variant that requires only $O(1/ε)$ calls to the gradient oracle. Finally, we present numerical experiments demonstrating the effectiveness of the proposed method compared to other projection-free algorithms.
title Revisiting Frank-Wolfe for Structured Nonconvex Optimization
topic Optimization and Control
Machine Learning
90C26
url https://arxiv.org/abs/2503.08921