Re: [Bug-gsl] GSL: NaN results in SVD calculation (gsl_linalg_SV_decomp)

[Brian Gough]

At Tue, 26 Jan 2010 11:01:13 +0100,
Bruno Daniel wrote:
> Recently, I encountered a problem in the SVD routine of GSL version 1.9:
> Certain matrices A with very many zeros lead to an SVD with NaN ("not a
> number") results in the vector of singular values S as well as in the left and
> right singular vectors.
> 
> I tracked the error down to the function trailing_eigenvalue in svdstep.c:
> If tb, tab and dt are all zero, the calculated eigenvalue is undefined:
> 
> mu = tb - (tab * tab) / (dt + hypot (dt, tab)) = 0 - 0 / 0;
>  
> This NaN result is then propagated in the function qrstep via create_givens to
> all the other results.
> 

Thanks for the bug report and example program.  I've confirmed the bug
(it also occurs in the latest release, 1.13).

I've logged it at http://savannah.gnu.org/bugs/index.php?28767
and am looking into it.

-- 
Brian Gough

GNU Scientific Library -
http://www.gnu.org/software/gsl/