1: \begin{abstract}

2:     The cross-product matrix-based CJ-FEAST SVDsolver

3:     proposed previously by the authors is shown to

4:     compute the left singular vector possibly much less accurately than

5:     the right singular vector and may be numerically backward unstable

6:     when a desired singular value is small.

7:     In this paper, an alternative

8:     augmented matrix-based CJ-FEAST SVDsolver is considered to

9:     compute the singular triplets of a large matrix $A$ with the

10:     singular values in an interval $[a,b]$ contained

11:     in the singular spectrum.

12:     The new CJ-FEAST SVDsolver is a subspace iteration applied to an

13:     approximate spectral projector of the augmented matrix

14:     $[0, A^T; A, 0]$ associated with the eigenvalues in $[a,b]$,

15:     and constructs approximate left and right

16:     singular subspaces with the desired singular values independently,

17:     onto which $A$ is projected to obtain the

18:     Ritz approximations to the desired singular triplets.

19:     Compact estimates are given for the accuracy of the approximate spectral

20:     projector, and a number of convergence results are established.

21:     The new solver is proved to be always numerically backward stable.

22:     A convergence comparison of the cross-product and augmented matrix-based

23:     CJ-FEAST SVDsolvers is made,

24:     and a general-purpose choice strategy between the two solvers is proposed

25:     for the robustness and overall efficiency. Numerical experiments confirm

26:     all the results.

27: \end{abstract}