> > > Grr, that was a mistake. It should be 2*B+log2(N) and 2*(B-2)+log2(N)
> > > for ordinary and balanced representation, respectively.
> 
> Nitpick: my response should have used 2*(B-1)+log2(N)

I thought so, 
> > 
> > 56 bits is nearer the mark!

58 bits isn't far off either

> > I assume statistics along the lines I suggested (normal distribution)
> > explains the 4 bit discrpancy. Or is it the varying integer length of 
> > a weighted transform?

> Ernst Mayer assumed the output digits of the product had a size
> dictated by the statistics of a random walk. My understanding is that
> it's entirely possible to get a digit in the output that's larger than
> 53 bits, in which case the convolution will fail even with zero
> roundoff error.
> 
> jasonp

N steps of +/- 1 chosen at random gives a normal distribution. The maximum
possible is N but the standard deviation is SQR(N).

This suggests replacing log2(N) (depth of the butterfly tree FFT) with 
log2(N)/2.
This gives 48 bits for a roughly evens chance of a roundoff error. The extra 5
bits decrease the probability of error as desired.

Is this as plausible as it sounds?

David
_________________________________________________________________
Be one of the first to try Windows Live Mail.
http://ideas.live.com/programpage.aspx?versionId=5d21c51a-b161-4314-9b0e-4911fb2b2e6d
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to