2010/7/30 Guillaume Chérel <guillaume.c.cherel@gmail.com>:
Hi,
Thanks for all your answers and the references (and yes, I have to admit that I've been a bit lazy with Goldberg's article, though it looks very thorough).
But as numpy is designed for scientific computing, is there no implementation of an "exact type" (http://floating-point-gui.de/formats/exact/) to avoid floating point issues?
I think there is a misunderstanding: the *vast* majority of scientific computing use floating point. The most commonly used way of doing "exact" computation is to do it symbolically, which is inapproriate in most practical cases. Arbitrary precision is also very different from exact precision - arbitrary precision means you are trading speed/memory for more precision, but there will still be errors. I don't really understand the exact problem you are trying to solve, but since you are approximating something, I am doubtful you need exact precision, and in that case, floating point as used in numpy (and every other software doing *numerical* computation) is likely to be enough. cheers, David