More accurate summations [was Re: This math scares me]
Edward Jason Riedy
ejr at cs.berkeley.edu
Sat Mar 17 06:17:35 CET 2001
And I mistakenly blather:

 The accumulated error in s from floatingpoint calculations is at
 most one unit in the last place of each summand x, [...]
Geez, I can't get anything right. It's not one ulp, but
rather the distance from 1.0 to the next higher floating
point number (either epsilon or 2*epsilon, depending on
whose definition you use). But self.c still gives a bound
on the total rounding error.
Jason

More information about the Pythonlist
mailing list