Phil Steitz wrote:
I looked at this some more last night and now agree that if you are
just computing SSE, scoring the data and running that one sum in a
second pass should in general be more accurate. The problem is, as
Luc pointed out, the need to store all of the data and I don't see any
way around that.
That's what I meant by "more complex".
If there are better stateless formulas, then we
should look at them.
I don't think there is a stateless formula not subject to numerical
problems in certain cases.
I am still -0 on adding a separate stateful
impl, but could be convinced if others feel differently and someone is
willing to volunteer to research, code, doc and write tests for it.
I don't think it's worth the effort. The current behaviour should be
prominently documented, of course, and if someone can come up with
a way to warn unsuspecting users that the result may be somewhat
inaccurate in case the bit cancellation is detected, that would
be great (I'd like to avoid introducing a logging mechanism).
J.Pietschmann
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]