Minimax exactness and global saddle points of nonlinear augmented Lagrangians

Abstract

We study global minimax exactness of merit functions for constrained optimization problems. This concept arises as a natural generalization of the definition of global saddle points in the unified theory of exactness of penalty and augmented Lagrangian functions. We obtain necessary and sufficient conditions for the global minimax exactness of nonlinear augmented Lagrangians in the form of the localization principle, which allow one to reduce the study of the existence of global saddle points (or the existence of solutions of the augmented dual problem) to a local analysis of sufficient optimality conditions. With the use of the localization principle, we obtain simple necessary and sufficient conditions for the existence of global saddle points of He-Wu-Meng's augmented Lagrangian for inequality constrained problems and nonlinear rescaling Lagrangians for nonconvex semidefinite programs. We also introduce and analyze a new nonlinear smooth augmented Lagrangian for constrained minimax problems and provide necessary and sufficient conditions for the existence of a global saddle point of this augmented Lagrangian, which, in particular, expose some limitations of exponential penalty function methods.