Globally Optimal Robust Matrix Completion Based on M-Estimation

Robust matrix completion allows for estimating a low-rank matrix based on a subset of its entries, even in presence of impulsive noise and outliers. We explore recent progress in the theoretical analysis of non-convex low-rank factorization problems to develop a robust approach that is based on a fast factorization method. We propose an algorithm that uses joint regression and scale estimation to compute the estimates. We prove that our algorithm converges to a global minimum with random initialization. An example function for which the guarantees hold is the pseudo-Huber function. In simulations, the proposed approach is compared to state-of the art robust and nonrobust methods. In addition, its applicability to image inpainting and occlusion removal is demonstrated.