1: \begin{abstract}

2: Nonconvex regularizers have been successfully used in

3: low-rank matrix

4: learning. 

5: In this paper,

6: we extend 

7: this to the more challenging problem of

8: low-rank tensor completion.

9: Based on the proximal average algorithm,

10: we develop an efficient solver

11: that avoids expensive 

12: tensor

13: folding and unfolding.

14: A special ``sparse plus

15: low-rank" structure, which is essential for fast computation of individual proximal

16: steps, 

17: is maintained

18: throughout the iterations.

19: We also incorporate adaptive momentum to further speed up empirical convergence.

20: Convergence results to critical points are provided

21: under smoothness and Kurdyka-Lojasiewicz conditions.

22: Experimental results on a number of synthetic and real-world data sets show that the proposed 

23: algorithm is more 

24: efficient in both

25: time and space,

26: and is also more accurate

27: than existing approaches.

28: \end{abstract}