1: \begin{abstract}

2:    This paper introduces \ourmethod,

3:    a new first-order method for solving large-scale semidefinite programs (SDPs) with bounded domain.

4:    \ourmethod is an inexact augmented Lagrangian (AL) method where the AL subproblems

5:    are solved by a novel hybrid low-rank (HLR) method.

6:    The recipe behind HLR is based on two key ingredients:

7:    1) an adaptive inexact proximal point method with inner acceleration;

8:    2) Frank-Wolfe steps to escape from spurious local stationary points.

9:    In contrast to the low-rank method of Burer and Monteiro,

10:    \ourmethod finds a near-optimal solution (with provable complexity bounds)

11:    of SDP instances satisfying strong duality.

12:    Computational results comparing \ourmethod to state-of-the-art solvers on several large SDP instances

13:    arising from 

14:    %maximum cut, 

15:    maximum stable set, phase retrieval, and matrix completion,

16:    show that the former finds highly accurate solutions in substantially less CPU time than the latter ones.

17:    For example, in less than 20 minutes, \ourmethod can solve a maximum stable set SDP instance with dimension pair $(n,m)\approx (10^6,10^7)$ within 

18:    $10^{-5}$ relative precision.

19: 

20:    \vgap

21: 

22: \noindent

23:    {\bf Keywords:} semidefinite programming, augmented Lagrangian, low-rank methods, proximal point method, Frank-Wolfe method, iteration complexity, adaptive method, global convergence rate

24: 

25:    % surpassing state-of-the-art SDP solvers.

26:    % Our experiments include SDP instances with up to trillions of scalar variables.

27: \end{abstract}

28: