1: \begin{abstract}

2:     In this paper, we study first-order methods on a large variety of

3: low-rank matrix optimization problems, whose solutions only live in a low

4: dimensional eigenspace. Traditional first-order methods depend on the eigenvalue decomposition 

5: at each iteration which takes most of the computation time. 

6: In order to reduce the cost, we propose an inexact algorithm framework

7: based on a polynomial subspace extraction.

8: The idea is to use an additional polynomial-filtered iteration to extract an approximated eigenspace, 

9: and project the iteration matrix on this subspace, followed by an optimization update.

10: The accuracy of the extracted subspace can be controlled by the degree of the polynomial filters.

11: This kind of subspace extraction also enjoys the warm start property: the subspace of the

12: current iteration is refined from the previous one. 

13: Then this framework is instantiated into two algorithms: 

14: the polynomial-filtered proximal gradient method and the

15: polynomial-filtered alternating direction method of multipliers.

16: We give a theoretical guarantee to the two algorithms that the polynomial degree is not necessarily very large.

17: They share the same convergence speed as the corresponding original methods if

18: the polynomial degree grows with an order $\Omega(\log k)$ at the $k$-th iteration.

19: If the warm-start property is considered, the degree can be reduced to a constant, independent of the

20: iteration $k$. Preliminary numerical experiments on

21: several low-rank matrix optimization problems show that the polynomial filtered algorithms usually provide multi-fold speedups.

22: 

23: \end{abstract}

24: