[Help-gsl] svd inaccuracy

[Evan Drumwright]

Hi All,

I've been attempting to use the GSL (specifically, the svd function) to 
compute the pseudo-inverse of a potentially singular matrix.  Unfortunately, 
the smaller singular values that I am getting (with both 
gsl_linalg_sv_decomp() or gsl_linalg_sv_decomp_jacobi()) are orders of 
magnitude off from the true result (computed and verified in octave).

I do understand the issues with numerical computing, and yet this still 
appears to be a strange issue (especially when you consider that most svd 
algorithms are quite robust).  For example, even if the matrix is not rank 
deficient, I still get poor results from the svd, enough to cause the 
pseudo-inverse to deviate significantly from the true inverse.  I was going 
to try calling svd with long double matrices to see whether this would 
improve things any, but the gsl does not seem to support this (am I wrong?).

Any suggestions?

Thanks,
Evan