[SciPy-User] Information about the Numerical Stability of Scipy/Numpy

[David Cournapeau]

Hi Marcel,

On Tue, Jan 17, 2012 at 7:58 PM, Marcel Caraciolo <caraciol at gmail.com> wrote:
> Hi all,
>
> I am studying scipy and numpy and  I decide to write some routines with
> those packages for a paper submission in a scientific congress. The problem
> is the validation of the results and experiments I have to show that the
> libraries that I used (in this case Scipy and Numpy) provides numerical
> stability, otherwise the chances that my article be approved will be
> decreased.

Are you referring to a precise guideline of the publication you have
in mind, and if so, could you point to it ?

Depending on the meaning you put behind numerical stability, numpy may
or may not be stable. I would say it is not fundamentally different
than any similar numerical package (e.g. matlab, octave, etc...), as
they share a lot of the same underlying implementation for fundamental
algorithms. Incidentally, a lot of this common implementation is taken
from netlib, and the quality of the code there is variable.

If you are implementing new algorithms, I think it is fair to say that
the stability depends as much if not more from how you use a library
than the library itself.

cheers,

David