|
On 11/30/10 8:48 AM, steve donovan wrote:
On Tue, Nov 30, 2010 at 6:33 PM, Roberto Ierusalimschy <roberto@inf.puc-rio.br> wrote:Something like linterval? (http://www.tecgraf.puc-rio.br/~lhf/ftp/lua/#linterval)Yes, that's exactly it. The source of all wisdom has some very readable paragraphs: http://en.wikipedia.org/wiki/Interval_arithmetic
What it fails to mention is that intervals tend to blow up faster than actual errors for most complex computations.
So, for example, if you have to invert a matrix, the interval result winds up with gigantic values and huge intervals while the standard binary floating point result winds up within delta of the actual value most of the time.
-a