Continuous Relaxation of Discontinuous Shrinkage Operator: Proximal Inclusion and Conversion

Abstract

We present a principled way of deriving a continuous relaxation of a given discontinuous shrinkage operator, which is based on a couple of fundamental results. First, the image of a point with respect to the ``set-valued'' proximity operator of a nonconvex function is included by that for its lower semicontinuous (l.s.c.) 1-weakly-convex envelope. Second, the ``set-valued'' proximity operator of a proper l.s.c. 1-weakly-convex function is converted, via double inversion, to a ``single-valued'' proximity operator which is Lipschitz continuous. As a specific example, we derive a continuous relaxation of the discontinuous shrinkage operator associated with the reversely ordered weighted $\ell_1$ (ROWL) penalty. Numerical examples demonstrate potential advantages of the continuous relaxation.