[SciPy-User] silently different results when running scipy.signal.fftconvolve on ubuntu karmic vs. lucid

[Ábel Dániel]

Hi!

About a week ago I wrote about an issue with using
scipy.signal.fftconvolve and scipy.signal.correlate2d. I hit another
issue since: apparently scipy.signal.fftconvolve gives different
results under ubuntu karmic and ubuntu lucid. I assume this is due to
some small difference in the versions of the libraries used by the
two, I alread submitted a bugreport to ubuntu's bugtracker
(https://bugs.launchpad.net/ubuntu/+source/python-scipy/+bug/705354)

Since that report didn't get any reaction, I thought I would ask here,
as well: any idea what might be causing this difference?

In addition, the fact that running the same program on two very
similar systems silently gives two very different results, I have been
thinking: is there a regression-testing suite that scipy uses? I
assume that a lot of scientists use scipy and programs based on scipy,
and having results change for no apparent reason doesn't exactly
instill confidence, and might even call into question the
reproducibility of any scientific paper based on data analysis that
used scipy.

Any thoughts?

Daniel Abel
abli at freemail.hu