Recovery of Low-Rank Matrices under Affine Constraints via a Smoothed Rank Function

TL;DR

This paper introduces SRF, a novel algorithm that uses a smoothed rank function approximation to improve low-rank matrix recovery under affine constraints, outperforming nuclear norm minimization especially near the minimal data threshold.

Contribution

It proposes a non-convex optimization scheme with a progressive approximation of the rank function, backed by theoretical convergence guarantees and superior empirical performance.

Findings

SRF can recover matrices not recoverable by nuclear norm minimization.

The algorithm achieves higher accuracy in matrix completion near the minimal number of observed entries.

Theoretical convergence to the minimum rank solution is established.

Abstract

In this paper, the problem of matrix rank minimization under affine constraints is addressed. The state-of-the-art algorithms can recover matrices with a rank much less than what is sufficient for the uniqueness of the solution of this optimization problem. We propose an algorithm based on a smooth approximation of the rank function, which practically improves recovery limits on the rank of the solution. This approximation leads to a non-convex program; thus, to avoid getting trapped in local solutions, we use the following scheme. Initially, a rough approximation of the rank function subject to the affine constraints is optimized. As the algorithm proceeds, finer approximations of the rank are optimized and the solver is initialized with the solution of the previous approximation until reaching the desired accuracy. On the theoretical side, benefiting from the spherical section property, we will show that the sequence of the solutions of the approximating function converges to the minimum rank solution. On the experimental side, it will be shown that the proposed algorithm, termed SRF standing for Smoothed Rank Function, can recover matrices which are unique solutions of the rank minimization problem and yet not recoverable by nuclear norm minimization. Furthermore, it will be demonstrated that, in completing partially observed matrices, the accuracy of SRF is considerably and consistently better than some famous algorithms when the number of revealed entries is close to the minimum number of parameters that uniquely represent a low-rank matrix.