SVDIFP.m: - A matlab program that computes a few smallest or largest singular values of a large m by n matrix A:

A u = sigma v; and A* v = sigma u

where A is an m by n matrix, sigma is a singular value, u a corresponding right singular vector, and v a corresponding left singular vector. It computes the singular values of A through computing the eigenvalues of A'A using the inverse free preconditioned Krylov subspace method (EIGIFP). The basic SVDIFP algorithm is described in

Q. Liang and Q. Ye, Computing Singular Values of Large Matrices With Inverse Free Preconditioned Krylov Subspace Method.

The program is intended for efficiently computing a few clustered extreme singular values through preconditioning.

Download:

The program is distributed at GitHub under the GNU General Public License. You can also download the main routine SVDIFP.m here. You may also want to download RIF.c, which constructs the robust incomplete factorization (RIF) for preconditioning (see M. Benzi and M. Tuma, A Robust Preconditioner with Low Memory Requirements for Large Sparse Least Squares Problems, SIAM J. Scientific Computing, 25 (2003), 499-512). The construction of the preconditioner is the most time consuming part of the method. Both MATLAB implementation (within svdifp.m) and C implementation are provided. To use RIF.c, invoke MATLAB's mex functionality by

mex -largeArrayDims RIF.c

Our RIF routines are based on the implementation of rifnrmex.m by Benzi and Tuma. When an RIF executable file is not present, the built-in MATLAB implementation will be used.

SVDIFP is developed by Qiao Liang and Qiang Ye. This work is supported in part by NSF grant DMS-1317424.

Return to Ye's Homepage.