LAPACK Archives

[Lapack] INFO=N+2 in xDGGES


attached is the zip file similar to one I sent to the lapack forum
describing another issue. It contains two matrices for dgges for which I
spotted the error.

The two matrices correspond to the first order approximation of 2-country
version Global Intergrated Monetary and Fiscal model developed in research
department in IMF.

Interestingly, I also observed, that I can solve the problem for lapack
3.1.1 and later, but earlier versions of lapack find the solution. The issue
raised by me a few days ago in the forum.

I am really inclined to believe that some change from 3.0 to 3.1.1
compromised numerical stability, sorry for being so bold, but this is
already second observation for a very different matrix. Will examine this
problem in more depth if time allows.

cheers and many thanks for your great work!

Ondra K.

On Thu, Mar 18, 2010 at 8:55 PM, James Demmel <demmel@Domain.Removed>wrote:

Thanks also for your bug report!
Do you have input data that caused this bug to occur,
that you could share with us to try to debug it?
Or did you just notice it by reading the code?
Or did you discover it another way?
Jim Demmel

julie langou wrote:

Thanks for the bug report. I have added your bug report to the LAPACK
Errata file. See: , bug0048. I
confirmed the problem and we will fix it before next release.
Best wishes,
On Mar 2, 2010, at 1:24 AM, Ondra Kamenik wrote:

 Dear lapack developers,

I believe I found a minor little small bug in xDGGES, which should return
info=n+3 if xTGSEN fails to swap eigenvalues at all, and return info=n+2 if
swapping slightly changes eigenvalues so that the selected eigenvalues are
not only in the upper part.

I am referening to dgges.f in lapack 3.2.1 which i downloaded recently.

Indeed, on line 444 the routine sets info=n+3 upon error in dtgsen.
However, then it checks whether the selected eigenvalues are only in the
upper part. If dtgsen has failed, it is almost certain that it is the case
that the selected eigenvalues are not only in the upper part, and info=n+2
is set. In this way user does not get info=n+3, but only info=n+2.

Many thanks for your great job.

Ondra K.

Lapack mailing list
Lapack@Domain.Removed <mailto:Lapack@Domain.Removed>

Julie Langou; Research Associate in Computer Science
Innovative Computing Laboratory;
University of Tennessee from Denver, Colorado ;-)
julie@Domain.Removed <mailto:julie@Domain.Removed>;<><>


Lapack mailing list

-------------- next part --------------
An HTML attachment was scrubbed...
-------------- next part --------------
A non-text attachment was scrubbed...
Type: application/zip
Size: 20912 bytes
Desc: not available
Url :

<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