Universal Minimality in TV Regularization with Rectilinear Anisotropy

Abstract: In this article, we study theoretical aspects of the $\ell^{1}$-anisotropic Rudin-Osher-Fatemi (ROF) model on hyperrectangular domains of arbitrary dimension. Our main result states that the solution of the ROF model minimizes a large class of convex functionals, including all $L^p$-norms, over a certain neighbourhood of the datum. This neighbourhood arises naturally in the dual formulation and is given by the TV subdifferential at zero, scaled by the regularization parameter and translated by the datum. In order to prove the main result, we exploit the recently established fact that, for piecewise constant datum, the ROF model reduces to a finite-dimensional problem. We then show a finite-dimensional version of the main result, before translating it back to the original setting and finishing the proof by means of an approximation argument.