1: \begin{abstract}

2: 	Low rank matrix approximation is a popular topic in machine learning.

3: 	In this paper, we propose a new algorithm for this topic by minimizing the least-squares estimation  over the Riemannian manifold of fixed-rank matrices.

4: 	The algorithm is an adaptation of classical gradient descent within the 

5: 	framework of optimization on manifolds. 

6: 	In particular, we reformulate an unconstrained optimization problem on a low-rank manifold into a differential dynamic system.

7: 	We develop a  splitting numerical integration method by applying  a splitting integration scheme to the dynamic system.

8: 	%Although our formulation leads to a nonconvex optimization problem, 

9: 	We conduct the  convergence analysis of our splitting numerical integration algorithm.

10: 	It can  be guaranteed that the error between the recovered matrix and true result is monotonically decreasing in the Frobenius norm.

11: 	Moreover, our splitting numerical integration can be adapted into matrix completion scenarios.

12: 	Experimental results show that our approach has good scalability for large-scale problems with  satisfactory accuracy.

13: \end{abstract}

14: