GSVD {PEIP}

Generalized SVD

Description

Wrapper for generalized svd from LAPACK

Usage

GSVD(A, B)

Arguments

A

Matrix, see below

B

Matrix, see below

Details

The A and B matrices will be, A=U*C*t(X) and B=V*S*t(X), respectively.

Since PEIP is based on a book, which is iteslef based on MATLAB routines, the convention here follows the book. The R implementation uses LAPACK and wraps the function so the output will comply with the book. See page 104 of the second edition of the Aster book cited below. That said, the purpose is to find an inversion of the form Y = t(A aB), where a is a regularization parameter, B is smoothing matrix and A is the design matrix for the forward problem. The input matrices A and B are assumed to have full rank, and p = rank(B). The generalized singular values are then gamma = lambda/mu, where lambda = sqrt(diag(t(C)*C) ) and mu = sqrt(diag(t(S)*S) ).

Value

U

m by m orthogonal matrix

V

p by p orthogonal matrix, p=rank(B)

X

n by n nonsingular matrix

C

singular values, m by n matrix with diagonal elements shifted from main diagonal

S

singular values, p by n diagonal matrix

Note

Requires R version of LAPACK. The code is a wrapper for the dggsvd function in LAPACK. The author thanks Berend Hasselman for advice and help preparing this function.

Author(s)

Jonathan M. Lees<[e-mail]>

References

Aster, R.C., C.H. Thurber, and B. Borchers, Parameter Estimation and Inverse Problems, Elsevier Academic Press, Amsterdam, 2005.

See Also

flipGSVD

Examples

# Example from NAG F08VAF

A <- matrix(1:15, nrow=5,ncol=3)
 B <- matrix(c(8,1,6,
               3,5,7,
               4,9,2), nrow=3, byrow=TRUE)

z <- GSVD(A,B)
C <- z$C
S <- z$S
sqrt(diag(t(C) %*% C)) / sqrt(diag(t(S) %*% S))
testA = A - z$U %*% C %*% t(z$X)
testB = B - z$V %*% S %*% t(z$X)

print(testA)
print(testB)