[SciPy-dev] Strange bug in linalg

Norbert Nemec Norbert.Nemec.list at gmx.de
Tue Nov 16 04:26:54 CST 2004


On Monday 15 November 2004 19:11, Fernando Perez wrote:
> Norbert Nemec wrote:
> > On Monday 15 November 2004 13:34, Robert Kern wrote:
> >>Norbert Nemec wrote:
> >>>Hi there,
> >>>
> >>>the following code shows a really strange bug:
> >>>
> >>>-----------------------------
> >>>
> >>>>from numarray import *
> >>>>from scipy.linalg import *
> >>>>print array([[1.0]])
> >>>>print eig(array([[1.0,1.0j],[-1.0j,1.0j]]))
> >>
> >>Scipy does not currently work with numarray. My first guess is that that
> >>is causing the crash.
> >
> > Large parts do already work with numarray. The longterm goal is also set
> > to support numarray (As I understand, probably as an alternative to
> > Numeric). Therefore, I would assume that those points that do not work
> > yet should be identified and either be fixed or marked as ToDo.
>
> This one _might_ not be a scipy problem.  I don't have time to track it
> right now, but google for Andrew Straw in the numarray list for some recent
> messages regarding floating point exceptions.  He was running into similar
> issues, and it turned out to be a very obscure glibc bug.  You might want
> to look into that before spending too much more time on it from the python
> side, since it was quite tricky to track down, and _not_ the kind of thing
> you'd think of first.

Thanks! That seems to be the answer. 

For those interested in details, see:
	http://aspn.activestate.com/ASPN/Mail/Message/numpy-discussion/2207861

Now, I'm only wondering how long it will take Debian to upgrade the glibc. :-(

(Actually, I don't even have a simple workaround. Even converting the 
numarray.array to Numeric.array before calculating the eigenvalues does not 
prevent the crash...)

Ciao,
Nobbi

-- 
_________________________________________Norbert Nemec
         Bernhardstr. 2 ... D-93053 Regensburg
     Tel: 0941 - 2009638 ... Mobil: 0179 - 7475199
           eMail: <Norbert at Nemec-online.de>




More information about the Scipy-dev mailing list