[music-dsp] floating point running sum optimization

[Ross Bencina]

koen vos wrote:
> On Thu, 27 Nov 2003 14:59:57 +1100, "Ross Bencina" wrote:
> > In terms of maintaining stability I'm expecting to have to use the
> > overlapped periodic recalculation version I posted the other day.
>
> Yes? The method I posted remains stable over arbitrary long time, and is
also
> much more accurate. Basically, I found precisely what you were looking
for..:)

Thanks for your explanations of the other issues. I don't understand how you
can make the claim that your noise cancellation method "remains stable over
arbitrary long time" when you are throwing out information  - "your" noise
cancellation approach may reduce the error (which is a great thing), but it
cannot eliminate it. Perhaps you are making assumptions about the
distribution of input data? I cannot see how to use your method alone and
not expect it to accumulate error eventually (all be it at a much slower
rate to "my" method).

Best wishes

Ross.