Title: Every critical point of an L0 composite minimization problem is a local minimizer

Abstract: Nowadays, L0 optimization model has shown its superiority when pursuing sparsity in many areas. For this nonconvex problem, most of the algorithms can only converge to one of its critical points. In this paper, we consider a general L0 regularized minimization problem, where the L0 ''norm'' is composited with a continuous map. Under some mild assumptions, we show that every critical point of this problem is a local minimizer, which improves the convergence results of existing algorithms. Surprisingly, this conclusion does not hold for low rank minimization, a natural matrix extension of L0 ''norm'' of a vector.