accuracy of intrinsic SUM function

[N.M. Maclaren]

On Aug 24 2012, N.M. Maclaren wrote:
>
>I may try with some malice-aforethought data distributions, to see
>if I can make pairwise less accurate on a distribution that is likely
>to occur in practice, but not tonight.

Yup.  I can.  I can't do that to Kahan summation, of course, as that is
mathematically equivalent to using double the precision.  So the correct
statement is that pairwise is more accurate than simple summation except
in rare cases.


Regards,
Nick Maclaren.