All possible computed results in correct floating-point summation (Q1118976)

The aim of this paper is to characterize the set of possible floating- point results, all equally representative of the exact algebraic results, for the case of summation algorithms. It is assumed that the set of all computed results is generated by perturbation in correct floating-point arithmetic; the errors due to the nonassociativity of the approached operators are not considered. Separate analyses are provided for the arithmetic summation, binary summation and sums of positive and negative numbers. The results obtained are compared with the estimates yielded by the classical error analysis in floating-point arithmetic. Numerical examples are given.