[Paper Review] Exact Matrix Completion via Convex Optimization

Emmanuel J. Candès, Benjamin Recht|ArXiv.org|May 29, 2008

Sparse and Compressive Sensing Techniques27 references57 citations

TL;DR

This paper establishes that most low-rank matrices can be exactly recovered from a nearly minimal number of randomly sampled entries by solving a convex optimization problem that minimizes the nuclear norm. The key result shows that with high probability, exact recovery is possible when the number of sampled entries exceeds $ C n^{1.2} r /log n $, where $ n $ is the matrix dimension and $ r $ is the rank.

ABSTRACT

We consider a problem of considerable practical interest: the recovery of a data matrix from a sampling of its entries. Suppose that we observe m entries selected uniformly at random from a matrix M. Can we complete the matrix and recover the entries that we have not seen? We show that one can perfectly recover most low-rank matrices from what appears to be an incomplete set of entries. We prove that if the number m of sampled entries obeys m >= C n^{1.2} r log n for some positive numerical constant C, then with very high probability, most n by n matrices of rank r can be perfectly recovered by solving a simple convex optimization program. This program finds the matrix with minimum nuclear norm that fits the data. The condition above assumes that the rank is not too large. However, if one replaces the 1.2 exponent with 1.25, then the result holds for all values of the rank. Similar results hold for arbitrary rectangular matrices as well. Our results are connected with the recent literature on compressed sensing, and show that objects other than signals and images can be perfectly reconstructed from very limited information.

Motivation & Objective

• To address the fundamental problem of recovering a low-rank matrix from a small, uniformly random subset of its entries.
• To determine the minimal number of sampled entries required for exact recovery with high probability.
• To establish conditions under which nuclear norm minimization reliably recovers the original matrix.
• To extend the theory of compressed sensing to matrix recovery, showing that matrices—like signals—can be reconstructed from incomplete information.
• To provide theoretical guarantees for matrix completion in practical scenarios such as recommender systems and sensor network localization.

Proposed method

• Propose a convex optimization program that minimizes the nuclear norm of the matrix subject to fitting the observed entries.
• Leverage the duality between the nuclear norm and the spectral norm to derive recovery guarantees via random matrix theory.
• Use the noncommutative Khintchine inequality and decoupling techniques to bound the operator norm of random matrix perturbations.
• Apply a concentration inequality for Rademacher chaos processes to control the deviation of the sampling operator from its expectation.
• Establish that the sampling operator satisfies a restricted isometry property (RIP) for low-rank matrices with high probability.
• Use the duality of the nuclear norm and the spectral norm to show that the solution to the convex program is unique and equal to the original matrix.

Experimental results

Research questions

• RQ1 Can a low-rank matrix be exactly recovered from a small, random subset of its entries using a convex optimization method?
• RQ2 What is the minimal number of sampled entries required to ensure high-probability exact recovery of a rank-$ r $ matrix?
• RQ3 Does nuclear norm minimization serve as a reliable convex surrogate for rank minimization in matrix completion?
• RQ4 How does the sampling mechanism (uniform random) affect the recovery performance and theoretical guarantees?
• RQ5 Can the theory of compressed sensing be extended to matrix recovery problems beyond signal and image reconstruction?

Key findings

• Exact recovery of most $ n imes n $ matrices of rank $ r $ is possible with high probability when the number of sampled entries $ m $ satisfies $ m /geq C n^{1.2} r /log n $ for some absolute constant $ C $.
• The nuclear norm minimization program recovers the original matrix exactly when the sampling is uniform and the number of entries meets the threshold above.
• The recovery guarantee holds for all ranks when the exponent $ 1.2 $ is replaced with $ 1.25 $, extending the result to arbitrary rank values.
• The method is robust to noise and applies to rectangular matrices, not just square ones.
• The theoretical framework connects matrix completion to compressed sensing, showing that low-rank matrices can be reconstructed from incomplete data, similar to sparse signals.
• The analysis relies on concentration of measure and operator norm bounds via the noncommutative Khintchine inequality and decoupling techniques.