1: \begin{abstract}%   <- trailing '%' for backward compatibility of .sty file

2: Nonconvex regularization has been popularly used in low-rank matrix learning.

3: However, extending  it for low-rank tensor learning

4: is still computationally expensive.

5: %as multiple tensor folding/unfolding are required.

6: To address this problem,

7: %we propose in this paper an efficient algorithm with convergence and statistical guarantees.  First,

8: we develop an efficient solver for use with a nonconvex extension of the

9: overlapped nuclear norm regularizer.

10: Based on the proximal average algorithm,

11: the proposed algorithm can avoid expensive 

12: tensor folding/unfolding operations. 

13: A special ``sparse plus low-rank" structure

14: is maintained throughout the iterations, and

15: allows fast computation of the individual proximal steps.

16: Empirical convergence is further improved

17: with the use of adaptive momentum.

18: We  provide convergence guarantees to critical points on 

19: smooth

20: losses 

21: 	and also on objectives satisfying

22: 	the Kurdyka-{\L}ojasiewicz condition.

23: While the optimization problem is nonconvex and nonsmooth,

24: we show that

25: its critical points still have good statistical performance

26: on the tensor completion problem.  

27: Experiments on various synthetic and real-world data sets show that the proposed algorithm is efficient

28: in both time and space and more accurate than the existing state-of-the-art.

29: \end{abstract}