A universal convergence theorem for primal-dual penalty and augmented Lagrangian methods

Abstract: We present a so-called universal convergence theorem for inexact primal-dual penalty and augmented Lagrangian methods that can be applied to a large number of such methods and reduces their convergence analysis to verification of some simple conditions on sequences generated by these methods. If these conditions are verified, then both primal and dual convergence follow directly from the universal convergence theorem. This theorem allows one not only to derive standard convergence theorems for many existing primal-dual penalty and augmented Lagrangian methods in a unified and straightforward manner, but also to strengthen and generalize some of these theorems. In particular, we show how with the use of the universal convergence theorem one can significantly improve some existing results on convergence of a primal-dual rounded weighted $\ell_1$-penalty method, an augmented Lagrangian method for cone constrained optimization, and some other primal-dual methods.