Lagrangian methods for composite optimization

Shoham Sabach, Marc Teboulle

Research output: Chapter in Book/Report/Conference proceedingChapterpeer-review

Abstract

Lagrangian-based methods have been on the market for over 50 years. These methods are robust and often can handle optimization problems with complex geometries through efficient computational steps. The last decade of research have generated a large volume of literature on various practical and theoretical aspects of many Lagrangian-based algorithms. This chapter reviews the basic elements of Lagrangian-based methods for composite minimization in the convex and nonconvex setting. In the convex case, the focus is on global rate of convergence results, which are derived here through a novel approach and very simple proof technique. In the much harder nonconvex case, we survey a very recent methodology which allows to establish global pointwise convergence results for a broad class of genuine nonlinear composite semialgebraic problems.

Original languageEnglish
Title of host publicationProcessing, Analyzing and Learning of Images, Shapes, and Forms
Subtitle of host publicationPart 2
EditorsRon Kimmel, Xue-Cheng Tai
PublisherElsevier B.V.
Pages401-436
Number of pages36
ISBN (Print)9780444641403
DOIs
StatePublished - 2019

Publication series

NameHandbook of Numerical Analysis
Volume20
ISSN (Print)1570-8659

Keywords

  • 65K05
  • 90C06
  • 90C25
  • 90C26
  • Alternating minimization
  • Convex and nonconvex composite minimization
  • Decomposition schemes
  • Global pointwise convergence
  • Global rate of convergence analysis
  • Kurdyka–Łosiajewicz property
  • Lagrangian multiplier methods
  • Proximal multiplier algorithms
  • Semialgebraic optimization

Fingerprint

Dive into the research topics of 'Lagrangian methods for composite optimization'. Together they form a unique fingerprint.

Cite this