From: Matthias Troyer (troyer_at_[hidden])
Date: 2007-01-31 15:57:58
On Jan 31, 2007, at 4:40 PM, Topher Cooper wrote:
> At 03:58 PM 1/30/2007, you wrote:
>> For now it seems that you use the default variance implementation
>> which should be the naive estimator from the sum and sum of squares.
> Why would you supply a poor implementation as the default when the
> alternative (West's algorithm) is efficient, easily implemented and
> is much more precise? The only reasons I can think for including the
> naive algorithm at all is for those rare cases where a small increase
> in performance is more important than a potentially very large loss
> in precision or if you are using exact arithmetic (e.g., if instead
> of floating point you are using rational numbers or if all your
> values are integers).
Yes, this is the reason - maybe one could just make the more accurate
version the default?
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk