> The floating point error occurred quite a bit until I caught most of
> them. I think I converted it to a RelaxError, although it sounds like
> the creation of a NaN is a much better behaviour. Those Lorentzians
> in the model-free spectral density functions are particularly bad when
> using a grid search where the correlation time starts at zero. Python
> deserves to be picked on a bit - implementing the IEEE 754 standard
> should have occurred a long ago even if that means avoiding the
> underlying C libraries.
Ah but they don't want to do this, because they would have to support
multiple archiitectures and their are time and speed tradeoffs that can
be made here by selecting the underlying compiler...
It would be pretty hard to implement! Maybe once they have Python
functioning well under Python rather than C (PyPy
http://codespeak.net/pypy/dist/pypy/doc/news.html) or under some other
language (http://www.python.org/dev/implementations.html) then maybe
they can implement the standard. For portability and not forcing a
single version of Python, prevention may be the best option - avoid
NaN at all costs. Catching it - the cure - with relax remaining
platform independent, is that possible?