At Sun, 07 Mar 2010 00:33:38 -0500, Beng Heng, Ng wrote: > I believe there is a bug in gsl_eigen_hermv for at least one Hermitian > matrix that results in NaN values in the majority of the elements of > evec. There are some 0 elements, though just a few of them. > > There was a fix in GSL v1.13 for issues with NAN values. However, I'm > already using GSL v1.13. > > Since it is a 264x264 matrix, I've attached a sample program that reads > in the matrix from a file. I compiled it as follows: > g++ -lgsl -lgslcblas hermvbug.cpp -o hermvbug.exe > > It will read in the matrix, check that the input matrix is indeed > Hermitian, call gsl_eigen_hermv, and finally write the output complex > eigenvectors, evec, to "bugoutput.txt". > > I'm using GSL v1.13, gcc v4.3.2, Linux kernel 2.6.26-2-amd64. The > processor is Intel Q9300. > > I'll be glad to provide more information, and will definitely appreciate > any feedback regarding this issue.
Thanks for the email, sorry for the delay in replying. I think this is probably similar to bug 28767 https://savannah.gnu.org/bugs/index.php?28767 with NaNs in the SVD which is based on the same QR reduction. I made a fix for that by scaling intermediate stages, but it needs to be applied to the eigen routines as well. -- Brian Gough GNU Scientific Library - http://www.gnu.org/software/gsl/ _______________________________________________ Bug-gsl mailing list [email protected] http://lists.gnu.org/mailman/listinfo/bug-gsl
