Al Chou wrote:
Uh, did we drop the idea of using the "corrected two-pass" algorithm for the
variance in the non-rolling case? I excerpted that thread below.
I was going to mention that. The discussion above regards the "non-stored vector" approach (UnivariateImpl). I submitted some patches to the "stored" classes (StoredUnivariate,AbstractStoreUnivariate) last weekend (adding percentiles -- http://issues.apache.org/bugzilla/show_bug.cgi?id=20377) and I was waiting for them to get committed before suggesting this change/submitting the patch.
Given our recent NR entanglements, we need to make sure that if and when we apply that patch we document using the original source of the formula.
Thanks for the remider.
What is relevant to the UnivariateImpl computations is Brent's suggestion re: making "running" error corrections. Have you investigated this further?
I'll pick up this patch and apply it before any of my changes to UnivariateImpl. Thanks for redirecting my attention to it, I'm still getting used to going into bugzilla.
-Mark
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
