We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

math.OC

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo

Mathematics > Optimization and Control

Title: A nearly linearly convergent first-order method for nonsmooth functions with quadratic growth

Abstract: Classical results show that gradient descent converges linearly to minimizers of smooth strongly convex functions. A natural question is whether there exists a locally nearly linearly convergent method for nonsmooth functions with quadratic growth. This work designs such a method for a wide class of nonsmooth and nonconvex locally Lipschitz functions, including max-of-smooth, Shapiro's decomposable class, and generic semialgebraic functions. The algorithm is parameter-free and derives from Goldstein's conceptual subgradient method.
Comments: 75 pages, 8 figures. Revised intro. New experiments. Simplified proofs. PyTorch code available here: this https URL
Subjects: Optimization and Control (math.OC)
MSC classes: 65K05, 65K10, 90C15, 90C30
Cite as: arXiv:2205.00064 [math.OC]
  (or arXiv:2205.00064v3 [math.OC] for this version)

Submission history

From: Damek Davis [view email]
[v1] Fri, 29 Apr 2022 20:18:16 GMT (1786kb,D)
[v2] Fri, 3 Jun 2022 18:57:19 GMT (1215kb,D)
[v3] Mon, 17 Jul 2023 13:45:21 GMT (5479kb,D)

Link back to: arXiv, form interface, contact.