Re: [eigen] Eigenvalue calculation on integer matrices

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 25.02.2014 13:19, Felix wrote:
> Hello,
>
> I noticed big errors when trying to calculate the eigenvalues of Integer
> matrices. When using SelfAdjointEigenSolver<MatrixXi> the Eigenvalues
> don't match the real ones at all. When calling eigenvalues on MatrixXi I
> just get a floating point exception. This probably has to do with
> Integer Matrices not having integer Eigenvalues. However there should be
> an implicit conversion or an error message the way it is right now if I
> hadn't compared the results to the ones from octave I would not have
> noticed that anything is wrong at all.

Yes, our algorithm uses floating point operations, and for integers they 
simply convert the result back to an integer (which usually generates 
bogus results).
I agree that we should forbid to compile decompositions which depend on 
non-supported operations.
E.g. LU/LDLT can be allowed for floating point and rational types but 
should not be allowed for integer types.
Most other decompositions are only available if sqrt (and possibly other 
functions) are available.

I thought I filed a bug for that once, but I can't find it, so I made a new:
http://eigen.tuxfamily.org/bz/show_bug.cgi?id=750

Christoph



--
----------------------------------------------
Dipl.-Inf., Dipl.-Math. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252
----------------------------------------------