4.5 Article

Global rates of convergence for nonconvex optimization on manifolds

Journal

IMA JOURNAL OF NUMERICAL ANALYSIS
Volume 39, Issue 1, Pages 1-33

Publisher

OXFORD UNIV PRESS
DOI: 10.1093/imanum/drx080

Keywords

complexity; gradient descent; trust-region method; Riemannian optimization; optimization on manifolds

Funding

  1. Fonds Speciaux de Recherche (FSR) at UCLouvain
  2. Chaire Havas 'Chaire Economie et gestion des nouvelles donnees'
  3. ERC Starting Grant SIPA
  4. Inria
  5. ENS
  6. National Science Foundation [DMS-1719558]
  7. Natural Environment Research Council [NE/L012146/1]
  8. NERC [NE/L012146/1] Funding Source: UKRI

Ask authors/readers for more resources

We consider the minimization of a cost function f on a manifold M using Riemannian gradient descent and Riemannian trust regions (RTR). We focus on satisfying necessary optimality conditions within a tolerance epsilon. Specifically, we show that, under Lipschitz-type assumptions on the pullbacks of f to the tangent spaces of M, both of these algorithms produce points with Riemannian gradient smaller than epsilon in O(1/epsilon(2)) iterations. Furthermore, RTR returns a point where also the Riemannian Hessian's least eigenvalue is larger than -epsilon in O(1/epsilon(3)) iterations. There are no assumptions on initialization. The rates match their (sharp) unconstrained counterparts as a function of the accuracy epsilon (up to constants) and hence are sharp in that sense. These are the first deterministic results for global rates of convergence to approximate first- and second-order Karush-Kuhn-Tucker points on manifolds. They apply in particular for optimization constrained to compact submanifolds of R-n, under simpler assumptions.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.5
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available