scipy.signal.convolve and related functions give drastically different results on karmic vs. lucid

Binary package hint: python-scipy

Running scipy.signal.convolve, scipy.signal.fftconvolve and scipy.signal.correlate2d on 2d arrays (from a grayscale image) gives drastically different results when run on karmic versus run on lucid.

There are no error messages reported, the output is simply different.

Considering that for example the correctness of scientific papers might depend on scipy functions working correctly and reproducibly, such silent change of behaviour is a pretty serious issue.

I will attach a .zip file containing the output on these two systems. The contents of the zip is also available at
https://hal.elte.hu/~abeld/scipy_convolve_bug/

Note that the email thread at http://news.gmane.org/gmane.comp.python.scientific.user might be relevant (where the problem was apparently caused by silent overflows) although the issue in this case is different.

ProblemType: Bug
Architecture: amd64
CheckboxSubmission: 2190597fb4a1561ac0df4777f2ef1081
CheckboxSystem: bb422ca46d02494cdbc459927a98bc2f
Date: Thu Jan 20 11:35:09 2011
DistroRelease: Ubuntu 9.10
InstallationMedia: Ubuntu 9.10 "Karmic Koala" - Release amd64 (20091027)
NonfreeKernelModules: fglrx
Package: python-scipy 0.7.0-2
ProcEnviron:
 PATH=(custom, user)
 LANG=en_US.UTF-8
 SHELL=/bin/bash
ProcVersionSignature: Ubuntu 2.6.31-22.70-generic
SourcePackage: python-scipy
Uname: Linux 2.6.31-22-generic x86_64