1: \begin{abstract}

2: In a Jacobi--Davidson (JD) type method for singular value decomposition (SVD) problems,

3: called JDSVD, a large symmetric and generally indefinite correction

4: equation is approximately solved iteratively at each outer iteration,

5: which constitutes the inner iterations and dominates

6: the overall efficiency of JDSVD. In this paper, a convergence

7: analysis is made on the minimal residual (MINRES) method for the correction equation.

8: Motivated by the results obtained, a preconditioned correction equation is derived that

9: extracts useful information from current searching subspaces to construct

10: effective preconditioners for the correction equation and

11: is proved to retain the same convergence of outer iterations of JDSVD.

12: The resulting method is called inner preconditioned JDSVD (IPJDSVD) method.

13: Convergence results show that MINRES for the preconditioned correction equation

14: can converge much faster when there is a cluster

15: of singular values closest to a given target, so that IPJDSVD is

16: more efficient than JDSVD. A new thick-restart IPJDSVD algorithm with

17: deflation and purgation is proposed that simultaneously accelerates the outer and inner

18: convergence of the standard thick-restart JDSVD and computes several singular

19: triplets of a large matrix. Numerical experiments justify the theory and illustrate the

20: considerable superiority of IPJDSVD to JDSVD.

21: \end{abstract}