1: \begin{abstract}

2: Matrix and tensor completion 

3: aim to recover a low-rank matrix / tensor from limited observations and have been

4: commonly used in 

5: applications

6: such as recommender systems and multi-relational data mining.

7: A state-of-the-art matrix completion algorithm is Soft-Impute, which exploits the

8: special ``sparse plus low-rank" structure of the matrix iterates to allow efficient SVD in each iteration.

9: Though Soft-Impute is a proximal algorithm,

10: it is generally believed that acceleration

11: destroys the special structure and is thus not useful.

12: In this paper, 

13: we show that Soft-Impute can indeed be accelerated without comprising this structure.

14: To further reduce the iteration time complexity,

15: we propose an approximate singular value thresholding scheme based on the power method.

16: Theoretical analysis shows that the proposed algorithm still enjoys the fast $O(1/T^2)$

17: convergence rate of accelerated proximal algorithms.

18: We also extend the proposed algorithm to tensor completion

19: with the scaled latent nuclear norm regularizer.

20: We show that a similar ``sparse plus low-rank'' structure also exists, leading to low iteration complexity 

21: and fast $O(1/T^2)$ convergence rate.

22: Besides, 

23: the proposed algorithm can be further extended to 

24: nonconvex low-rank regularizers,

25: which have better empirical performance than the convex nuclear norm regularizer.

26: Extensive experiments demonstrate that the proposed algorithm is much faster than Soft-Impute and other state-of-the-art matrix and tensor completion algorithms.

27: \end{abstract}