Jingwen Fan, Tanxing Wang, Deren Han, Escaping strict saddle points in nonsmooth nonconvex optimization with proximal alternating linearized minimization
Full Text: PDF
DOI: 10.23952/jnva.11.2027.2.03
Volume 11, Issue 2, 1 April 2027, Pages 271-295
Abstract. This paper establishes the saddle-point escaping theory for the proximal alternating linearized minimization (PALM) algorithm in the context of nonsmooth and nonconvex optimization. Under the assumptions of bounded smoothness, weakly convexity, and a separable active manifold, we prove that the PALM can almost surely escape strict saddle points, provided that the coefficients of the proximal regularization terms are sufficiently large. We explicitly derive a lower bound for these coefficients, which is a constant depending on problem-specific parameters at the optimal point. The numerical experiments are displayed to show the theoretical results. The work provides the theoretical guarantee for the PALM to escape the strict saddle points and find superior solutions in challenging nonsmooth, nonconvex, and nonseparable problems.
How to Cite this Article:
J. Fan, T. Wang, D. Han, Escaping strict saddle points in nonsmooth nonconvex optimization with proximal alternating linearized minimization, J. Nonlinear Var. Anal. 11 (2027), 271-295.
