On Saturday, 18 October 2014 at 12:48:00 UTC, eles wrote:
This might simply be a case biased in favor of pairwise summation, as numbers ar symmetric.
I haven't used pairwise summation though. The best solution is to partition based on exponent and use accurate integer arithmetics. The first function on the link I posted above is accurate in the partitions, but inaccurate of the summation of the partitions… :-/
But yes, you can create a lot worse datasets. Kahan only recover a few of the lost bits due to rounding.
