Local Linear Convergence of Inertial Forward-Backward Splitting for Low Complexity Regularization
Abstract
In this abstract, we consider the inertial Forward-Backward (iFB) splitting method and its special cases (Forward-Backward/ISTA and FISTA). Under the assumption that the non-smooth part of the objective is partly smooth relative to an active smooth manifold, we show that iFB-type methods (i) identify the active manifold in finite time, then (ii) enter a local linear convergence regime that we characterize precisely. This gives a grounded and unified explanation to the typical behaviour that has been observed numerically for many low-complexity regularizers, including 1 , 1,2-norms, total variation (TV) and nuclear norm to name a few. The obtained results are illustrated by concrete examples.
Domains
Image Processing [eess.IV]
Origin : Files produced by the author(s)
Loading...