1: \begin{abstract}

2: A Cross-Product Free (CPF) Jacobi-Davidson (JD) type method is

3: proposed  to compute a partial generalized singular value decomposition

4: (GSVD) of a large regular matrix pair $(A,B)$. It

5: implicitly solves the mathematically equivalent generalized eigenvalue

6: problem of $(A^TA,B^TB)$ but does not explicitly

7: form the cross-product matrices and thus avoids the possible accuracy loss

8: of the computed generalized singular values and generalized singular vectors.

9: The method is an inner-outer iteration method, where the expansion

10: of the right searching subspace forms the inner iterations that approximately

11: solve the correction equations involved and

12: the outer iterations extract approximate GSVD components with respect to the

13: subspaces.

14: Some convergence results are established for the inner and outer iterations,

15: based on some of which practical stopping criteria are designed for the inner

16: iterations. A thick-restart CPF-JDGSVD algorithm with deflation is developed to

17: compute several GSVD components.

18: Numerical experiments illustrate the efficiency of the algorithm.

19: \end{abstract}

20: