An Infeasible Method with Feasibility Safeguard for Nonsmooth Composite Optimization Over Manifolds

Abstract: In this paper, we consider nonsmooth composite optimization over compact embedded submanifolds defined by nonlinear equality constraints. We propose a feasibility-safeguarded inexact proximal linearized method (FSIPL), which allows infeasible iterates while keeping them within a prescribed bounded neighborhood of the manifold. Each iteration approximately solves a strongly convex proximal linearized subproblem, performs a correction step to reduce constraint violation, and uses a merit-function-based nonmonotone backtracking line search to select stepsizes and accept trial iterates. The feasibility safeguard, incorporated into both correction and line search, controls infeasibility and makes the boundedness needed in the analysis a consequence of the algorithmic design. We prove finite termination of backtracking, subsequential convergence to stationary points, and an $O(\varepsilon^{-2})$ outer iteration complexity bound. Under a Kurdyka--Łojasiewicz assumption on a suitable auxiliary function, we further establish full-sequence convergence. Numerical results on sparse PCA and sparse spectral clustering illustrate its efficiency.