LAPACK Archives

[Lapack] Non-orthogonal eigenvectors from DSYEVR


Hello all,

The conversation has been moved to the forum, 
http://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=2&t=1820

I remind all of you that if you want to receive an email for each post on 
the forum, you can. There is a bunch of noise there since there is a few 
spams ... I try to forward the relevant email to the mailing-list.

Best wishes,
Julien.

On Wed, 21 Apr 2010, Pauli Virtanen wrote:

Hi,

For your information: For some matrices DSYEVR, when told to compute a
small number of eigenvalues, returns non-orthogonal eigenvectors.
Perhaps it possible for the algorithm to detect this and return with an
error?

This applies at least when using the automatically chosen ABSTOL.
However, at least for the example below, I wasn't able to find a
suitable ABSTOL for which it would work.

Maybe, this might have to do with the fact that for the example problem
below, the first eigenvalue is multiply degenerate, and the numerical
second eigenvalue is eps-close to the first one.

The following test program illustrates this:

       http://projects.scipy.org/scipy/raw-attachment/ticket/1159/test.f90

with the test data

       http://projects.scipy.org/scipy/raw-attachment/ticket/1159/L.txt

On Lapack 3.2.1 + reference BLAS, I get:

$ gfortran -o test test.f90 -llapack -lblas
$ ./test < L.txt
orth error (           2  eigs): -0.96599125529927221      HUGE ERROR
orth error (           3  eigs): -0.96599125529927232      HUGE ERROR
orth error (           4  eigs): -0.96599125529927221      HUGE ERROR
orth error (           5  eigs): -2.52928802610462387E-017
orth error (           6  eigs): -1.47893223140021721E-016
orth error (           7  eigs): -1.18177613284446356E-016
orth error (           8  eigs):   0.0000000000000000
orth error (           9  eigs): -2.84549495443531079E-016
orth error (          10  eigs): -2.93456682158620488E-016


_______________________________________________
Lapack mailing list
Lapack@Domain.Removed
http://lists.eecs.utk.edu/mailman/listinfo/lapack


<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or