Coordinate descent accelerations for signal recovery on scale-free graphs based on total variation minimization

Abstract

We extend our previous work on learning smooth graph signals from a small number of noisy signal samples. Minimizing the signal's total variation amounts to a non-smooth convex optimization problem. We propose to solve this problem using a combination of Nesterov's smoothing technique and accelerated coordinate descent. The resulting algorithm converges substantially faster, specifically for graphs with vastly varying node degrees (e.g., scale-free graphs).