We propose a new Iteratively Reweighted Least Squares (IRLS) algorithm for the problem of recovering a matrix $X \in \mathbb{C}^{d_1 \times d_2}$ from incomplete linear observations, solving a sequence of quadratic problems. The easily implementable algorithm, which we call Harmonic Mean Iteratively Reweighted Least Squares ($\texttt{HM-IRLS}$), is superior compared to state-of-the-art algorithms for the low-rank recovery problem in several performance aspects. More specifically, the strategy $\texttt{HM-IRLS}$ uses to optimize a non-convex Schatten-p penalization to promote low-rankness carries three major strengths, in particular for the matrix completion setting. First, we show empirically that, starting from a trivial initialization, it converges globally to a low-rank minimizer of the non-convex Schatten-p objective function for a large number of interesting cases, for which other (non-)convex optimization approaches fail to recover a low-rank matrix. Secondly, as a main theoretical result, we prove that $\texttt{HM-IRLS}$, unlike other IRLS variants, exhibits a locally superlinear rate of convergence if the linear observations fulfill a suitable null space property. This rate is of order $2-p$, depending on the non-convexity parameter $p$, and thus arbitrarily close to quadratic for $p \ll 1$, which is accurately confirmed by our numerical experiments. Thirdly, our experiments show that $\texttt{HM-IRLS}$ exhibits an empirical recovery probability close to $1$ even for very close to optimal sample complexities, i.e., already for significantly fewer linear observations than any other tractable approach in the literature.

Type

Publication

In *12th International Conference on Sampling Theory and Applications (SampTA)*, Tallinn, Estonia, pp. 489-493

This work was extended to a paper that has been published in the JMLR.