1: \begin{abstract}

2: This paper proposes scalable and fast algorithms for solving the

3: Robust PCA problem, namely recovering a low-rank matrix with an

4: unknown fraction of its entries being arbitrarily corrupted. This

5: problem arises in many applications, such as image processing, web

6: data ranking, and bioinformatic data analysis. It was recently

7: shown that under surprisingly broad conditions, the Robust PCA

8: problem can be exactly solved via convex optimization that

9: minimizes a combination of the nuclear norm and the $\ell^1$-norm

10: . In this paper, we apply the method of augmented Lagrange

11: multipliers (ALM) to solve this convex program. As the objective

12: function is non-smooth, we show how to extend the classical

13: analysis of ALM to such new objective functions and prove the

14: optimality of the proposed algorithms and characterize their

15: convergence rate. Empirically, the proposed new algorithms can be

16: more than \emph{five times faster} than the previous

17: state-of-the-art algorithms for Robust PCA, such as the

18: accelerated proximal gradient (APG) algorithm. Moreover, the new

19: algorithms achieve higher precision, yet being less storage/memory

20: demanding. We also show that the ALM technique can be used to

21: solve the (related but somewhat simpler) matrix completion problem

22: and obtain rather promising results too. We further prove the

23: necessary and sufficient condition for the inexact ALM to converge

24: globally. Matlab code of all algorithms discussed are available at

25: \url{http://perception.csl.illinois.edu/matrix-rank/home.html}

26: \keywords{Low-rank matrix recovery or completion \and Robust

27: principal component analysis \and Nuclear norm minimization \and

28: $\ell^1$-norm minimization \and Proximal gradient algorithms \and

29: Augmented Lagrange multipliers}

30: % \PACS{PACS code1 \and PACS code2 \and more}

31: % \subclass{MSC code1 \and MSC code2 \and more}

32: \end{abstract}

33: