13
30

Learning the optimal Tikhonov regularizer for inverse problems

Abstract

In this work, we consider the linear inverse problem y=Ax+ϵy=Ax+\epsilon, where A ⁣:XYA\colon X\to Y is a known linear operator between the separable Hilbert spaces XX and YY, xx is a random variable in XX and ϵ\epsilon is a zero-mean random process in YY. This setting covers several inverse problems in imaging including denoising, deblurring, and X-ray tomography. Within the classical framework of regularization, we focus on the case where the regularization functional is not given a priori but learned from data. Our first result is a characterization of the optimal generalized Tikhonov regularizer, with respect to the mean squared error. We find that it is completely independent of the forward operator AA and depends only on the mean and covariance of xx. Then, we consider the problem of learning the regularizer from a finite training set in two different frameworks: one supervised, based on samples of both xx and yy, and one unsupervised, based only on samples of xx. In both cases, we prove generalization bounds, under some weak assumptions on the distribution of xx and ϵ\epsilon, including the case of sub-Gaussian variables. Our bounds hold in infinite-dimensional spaces, thereby showing that finer and finer discretizations do not make this learning problem harder. The results are validated through numerical simulations.

View on arXiv
Comments on this paper