Envelopt: Constrained Convex Composite Optimization

Abstract

We introduce Envelopt, a globally convergent iterative framework for a broad class of structured optimization problems where a smooth objective is augmented by a nonsmooth convex regularizer composed with a smooth mapping, and the variables are subject to general smooth constraints. All smooth functions may be nonconvex. The method is akin to an augmented-Lagrangian method in which partial minimization with respect to a lifting variable results in smooth subproblems involving the Moreau envelope of the nonsmooth regularizer, and the original constraints are retained explicitly. Only the proximal operator of the regularizer is required. Subproblems may be solved with off-the-shelf smooth optimization solvers. We state global convergence properties, establish that feasible limit points are asymptotically stationary, and develop an infeasibility detection mechanism. We derive worst-case iteration complexity bounds when the penalty parameter is and is not bounded away from zero. The framework subsumes the classical augmented Lagrangian method and accommodates important extensions, including stabilized formulations for degenerate problems, exact penalty methods, and conic constraints. We provide a Julia implementation, Envelopt.jl, as part of the JuliaSmoothOptimizers ecosystem. Numerical experiments with low-rank matrix completion, semidefinite programming, complementarity-constrained optimization, and nonconvex regularizers demonstrate the effectiveness and versatility of Envelopt.