Abstract
We introduce an iterative optimization scheme for convex objectivesconsisting of a linear loss and a non-separable penalty, based on theexpectation-consistent approximation and the vector approximate message-passing(VAMP) algorithm. Specifically, the penalties we approach are convex on alinear transformation of the variable to be determined, a notable example beingtotal variation (TV). We describe the connection between message-passingalgorithms -- typically used for approximate inference -- and proximal methodsfor optimization, and show that our scheme is, as VAMP, similar in nature tothe Peaceman-Rachford splitting, with the important difference that stepsizesare set adaptively. Finally, we benchmark the performance of our VAMP-likeiteration in problems where TV penalties are useful, namely classification intask fMRI and reconstruction in tomography, and show faster convergence thanthat of state-of-the-art approaches such as FISTA and ADMM in most settings.