1: \begin{abstract}

2: Low rank matrix recovery problems

3: %, including matrix completion and matrix sensing, 

4: appear in a broad range of applications.

5: In this work we present \texttt{GNMR} % must keep it \texttt{GNMR} rather than \GNMR for correct spacing!

6: --- an extremely simple iterative algorithm for low rank matrix recovery,  based on a Gauss-Newton linearization. 

7: %Our method can be essentially described in one line, as in each iteration it only solves a linear least squares problem. 

8: On the theoretical front, we derive recovery guarantees for \GNMR in both matrix sensing and matrix completion settings.

9: %In the matrix sensing setup, we prove that under very mild conditions, \GNMR obtains perfect recovery with a {quadratic} convergence rate.

10: Some of these results improve upon the best currently known for other methods.

11: A key property of \GNMR is that it implicitly keeps the factor matrices approximately balanced

12: throughout its iterations. 

13: %This is in contrast to recent works which explicitly add a regularization term to 

14: %keep the factor matrices approximately balanced.

15: On the empirical front, we show that for matrix completion with uniform sampling, \GNMR performs better than several popular methods, especially when given very few observations close to the information limit.

16: %This makes \GNMR one of the only methods with state of the art performance both theoretically and empirically.

17: \end{abstract}

18: