ZBDSQR - compute the singular value decomposition (SVD) of a real N-by-N (upper or lower) bidiagonal matrix B

SUBROUTINE ZBDSQR(

UPLO, N, NCVT, NRU, NCC, D, E, VT, LDVT, U, LDU, C, LDC, RWORK, INFO )

CHARACTER UPLO

INTEGER INFO, LDC, LDU, LDVT, N, NCC, NCVT, NRU

DOUBLE PRECISION D( * ), E( * ), RWORK( * )

COMPLEX*16 C( LDC, * ), U( LDU, * ), VT( LDVT, * )

ZBDSQR computes the singular value decomposition (SVD) of a real N-by-N (upper or lower) bidiagonal matrix B: B = Q * S * P' (P' denotes the transpose of P), where S is a diagonal matrix with non-negative diagonal elements (the singular values of B), and Q and P are orthogonal matrices.

The routine computes S, and optionally computes U * Q, P' * VT, or Q' * C, for given complex input matrices U, VT, and C.

See "Computing Small Singular Values of Bidiagonal Matrices With Guaranteed High Relative Accuracy," by J. Demmel and W. Kahan, LAPACK Working Note #3 (or SIAM J. Sci. Statist. Comput. vol. 11, no. 5, pp. 873-912, Sept 1990) and

"Accurate singular values and differential qd algorithms," by B. Parlett and V. Fernando, Technical Report CPAM-554, Mathematics Department, University of California at Berkeley, July 1992 for a detailed description of the algorithm.

UPLO (input) CHARACTER*1

= 'U': B is upper bidiagonal;

= 'L': B is lower bidiagonal.

N (input) INTEGER

The order of the matrix B. N >= 0.

NCVT (input) INTEGER

The number of columns of the matrix VT. NCVT >= 0.

NRU (input) INTEGER

The number of rows of the matrix U. NRU >= 0.

NCC (input) INTEGER

The number of columns of the matrix C. NCC >= 0.

D (input/output) DOUBLE PRECISION array, dimension (N)

On entry, the n diagonal elements of the bidiagonal matrix B. On exit, if INFO=0, the singular values of B in decreasing order.

E (input/output) DOUBLE PRECISION array, dimension (N)

On entry, the elements of E contain the offdiagonal elements of of the bidiagonal matrix whose SVD is desired. On normal exit (INFO = 0), E is destroyed. If the algorithm does not converge (INFO > 0), D and E will contain the diagonal and superdiagonal elements of a bidiagonal matrix orthogonally equivalent to the one given as input. E(N) is used for workspace.

VT (input/output) COMPLEX*16 array, dimension (LDVT, NCVT)

On entry, an N-by-NCVT matrix VT. On exit, VT is overwritten by P' * VT. VT is not referenced if NCVT = 0.

LDVT (input) INTEGER

The leading dimension of the array VT. LDVT >= max(1,N) if NCVT > 0; LDVT >= 1 if NCVT = 0.

U (input/output) COMPLEX*16 array, dimension (LDU, N)

On entry, an NRU-by-N matrix U. On exit, U is overwritten by U * Q. U is not referenced if NRU = 0.

LDU (input) INTEGER

The leading dimension of the array U. LDU >= max(1,NRU).

C (input/output) COMPLEX*16 array, dimension (LDC, NCC)

On entry, an N-by-NCC matrix C. On exit, C is overwritten by Q' * C. C is not referenced if NCC = 0.

LDC (input) INTEGER

The leading dimension of the array C. LDC >= max(1,N) if NCC > 0; LDC >=1 if NCC = 0.

RWORK (workspace) DOUBLE PRECISION array, dimension (4*N)

INFO (output) INTEGER

= 0: successful exit

< 0: If INFO = -i, the i-th argument had an illegal value

> 0: the algorithm did not converge; D and E contain the elements of a bidiagonal matrix which is orthogonally similar to the input matrix B; if INFO = i, i elements of E have not converged to zero.

TOLMUL DOUBLE PRECISION, default = max(10,min(100,EPS**(-1/8)))

TOLMUL controls the convergence criterion of the QR loop. If it is positive, TOLMUL*EPS is the desired relative precision in the computed singular values. If it is negative, abs(TOLMUL*EPS*sigma_max) is the desired absolute accuracy in the computed singular values (corresponds to relative accuracy abs(TOLMUL*EPS) in the largest singular value. abs(TOLMUL) should be between 1 and 1/EPS, and preferably between 10 (for fast convergence) and .1/EPS (for there to be some accuracy in the results). Default is to lose at either one eighth or 2 of the available decimal digits in each computed singular value (whichever is smaller).

MAXITR INTEGER, default = 6

MAXITR controls the maximum number of passes of the algorithm through its inner loop. The algorithms stops (and so fails to converge) if the number of passes through the inner loop exceeds MAXITR*N**2.