Abstract
In this work we study the behavior of the forward-backward (FB) algorithmwhen the proximity operator is replaced by a sub-iterative procedure toapproximate a Gaussian denoiser, in a Plug-and-Play (PnP) fashion. Inparticular, we consider both analysis and synthesis Gaussian denoisers within adictionary framework, obtained by unrolling dual-FB iterations or FBiterations, respectively. We analyze the associated minimization problems aswell as the asymptotic behavior of the resulting FB-PnP iterations. Inparticular, we show that the synthesis Gaussian denoising problem can be viewedas a proximity operator. For each case, analysis and synthesis, we show thatthe FB-PnP algorithms solve the same problem whether we use only one or aninfinite number of sub-iteration to solve the denoising problem at eachiteration. To this aim, we show that each "one sub-iteration" strategy withinthe FB-PnP can be interpreted as a primal-dual algorithm when a warm-restartstrategy is used. We further present similar results when using a Moreau-Yosidasmoothing of the global problem, for an arbitrary number of sub-iterations.Finally, we provide numerical simulations to illustrate our theoreticalresults. In particular we first consider a toy compressive sensing example, aswell as an image restoration problem in a deep dictionary framework.