Le vendredi 08 décembre 2006 à 09:25 +0000, Paul A Bristow a écrit :
I strongly support thinking about (computational) uncertainty estimates.
(Uncertainty seems to be the 'modern' term for what was called error).
But I am not clear exactly how to achieve this. We've got Interval but it's doubtful if this is the right weapon to use.
Interval as a generic notion is the right tool for bounding errors and representing uncertain inputs. Most papers on "reliable computing" are related to intervals (the name may change for higher dimensions, like ellipsoids, but they are still convex sets). Then there is the matter of representing intervals. With lower-upper bounds as in Boost.Interval, the representation is fine as long as the precision of the intervals is "small". For precisions of thousands of bits, a midpoint-radius representation is more efficient, as the radius can be stored in reduced form to improve performance and memory footprint. Best regards, Guillaume