1: \begin{abstract}

2: The singular value decomposition (SVD) of large-scale

3: matrices is a key tool in data analytics and scientific computing.

4: The rapid growth in the size of matrices further increases the need

5: for developing efficient large-scale SVD algorithms. Randomized SVD

6: based on one-time sketching has been studied, and its potential has

7: been demonstrated for computing a low-rank SVD. Instead of exploring different single random sketching techniques, we propose a Monte Carlo type integrated SVD algorithm based on multiple random sketches.

8: The proposed integration

9: algorithm takes multiple random sketches and then integrates the results

10: obtained from the

11: multiple sketched subspaces. So that the integrated SVD can achieve

12: higher accuracy and lower stochastic variations.

13: The main component of the integration

14: is an optimization problem with a matrix Stiefel manifold

15: constraint. The optimization problem is solved using

16: Kolmogorov-Nagumo-type averages. Our theoretical analyses show that

17: the singular vectors can be induced by population averaging and

18: ensure the consistencies between the computed and true subspaces and

19: singular vectors. Statistical analysis further proves a strong Law

20: of Large Numbers and gives a rate of convergence by the Central

21: Limit Theorem. Preliminary numerical results suggest that the

22: proposed integrated SVD algorithm is promising.

23: \end{abstract}