Re: bug in make check

[John Ketchum <johnk at qualcomm dot com>]

As I said in a previous response that got bounced by the redhat mailer 
daemon, one of the problems that is causing the eigen tests to fail is a 
problem with the call to

gsl_matrix_complex_const_column

in the test routine, which is supposed to retrieve an eigenvector from the 
matrix of eigenvectors to test whether the eigenvector and associated 
eigenvalue are solutions to  m*v = lambda*v.   However,

gsl_matrix_complex_const_column

doesn't do what it is supposed to do, and what it returns is not actually a 
column vector of the matrix of eigenvectors.  However, if the lines

       gsl_vector_complex_const_view vi = 
gsl_matrix_complex_const_column(evec, i);
       gsl_vector_complex_memcpy(x, &vi.vector);

are replaced with

       for (j=0; j<N; ++j) 
gsl_vector_complex_set(x,j,gsl_matrix_complex_get(evec,j,i));

the eigenvector is correctly copied into the vector x, and the tests 
pass.  (I have attached a source file where this change has been made).

Unfortunately, this is not the end of the story.  Even though the tests 
pass, inspection of the values of the elements of the eigenvalues and 
eigenvectors, and comparison with independently obtained eigenvalues and 
eigenvectors for the same matrix, reveals that all four eigenvalues are 
correct, but two of the eigenvectors have incorrect values in some of the 
elements. These two eigenvectors should be as follows

         0                       0
         0.5257                  0.8507
         0                       0
         -0.8507                 0.5257

However, the eigenvectors returned by gsl_eigen_hermv have values of NaN 
where the non-zero elements should be.  The tests pass only because of the 
specific properties of the test matrix, not because they are correct solutions.

I haven't had a chance to look into where the problem occurs in 
gsl_eigen_hermv, but when I run the test routine with the modification 
described above, using gsl libraries built with gcc 2.95, the eigenvector 
values appear to be correct.



At 06:41 PM 4/15/2002 +0100, Brian Gough wrote:
>John Ketchum writes:
>  > Here is a paired-down version of the "test.c" in the eigen 
> subdirectory of
>  > the source tree.  This checks the eigenvalues of a single 4x4 complex
>  > matrix, When gsl-1.1 is compiled with gcc 2.96-98, a number of the checks
>  > fail-- some of the failures appear trivial, i.e. a very small number with
>  > not quite the right value, but there are also some clearly catastrophic
>  > failures.
>
>A program which shows the bug indenpedently of gsl would be ideal --
>stepping through the existing program should show the point at which
>the eigenvector data becomes corrupted within the library.

John Ketchum
Qualcomm Inc.
9 Damonmill Square Suite 2A
Concord, MA 01742
Office: 60915
Direct: 781-276-0915
Fax: 781-276-0901
Mobile: 617-312-7023

Attachment: test.c
Description: Text document