Further remarks on Kahan summation with decreasing ordering
Résumé
We consider Kahan's compensated summation of $n$ floating-point numbers ordered as $|x_1| \ge \cdots \ge |x_n|$,
and show that in IEEE arithmetic a large relative error can occur for a dimension as small as $n=4$. This answers a question raised in particular by Priest and Higham.
Origine : Fichiers produits par l'(les) auteur(s)