On Sun, Oct 29, 2006 at 05:58:35PM +0000, david eddy wrote: > Putting it through the FFT mill and getting an exact 48 bit integer is > IMHO quite another:)
For the record, the relative error in a Cooley-Tukey FFT is bounded by O(ε log N), (ε being 0.5 * 10^-49 here, AFAICS, and N being the transform length); the average is about O(ε √log N). I'm not sure exactly which of these estimates are actually used for Prime95, but I remember it was all hashed out in a message to this list from Brian several years ago when George asked, so it's all in good hands. :-) /* Steinar */ -- Homepage: http://www.sesse.net/ _______________________________________________ Prime mailing list [email protected] http://hogranch.com/mailman/listinfo/prime
