Contribution à l'analyse d'algorithmes en arithmétique à virgule flottante

Floating-point arithmetic is an approximation of real arithmetic in which each operation may introduce a rounding error. The IEEE 754 standard requires elementary operations to be as accurate as possible. However, through a computation, rounding errors may accumulate and lead to totally wrong results. It happens for example with an expression as simple as ab + cd for which the naive algorithm sometimes returns a result with a relative error larger than 1. Thus, it is important to analyze algorithms in floating-point arithmetic to understand as thoroughly as possible the generated error. In this thesis, we are interested in the analysis of small building blocks of numerical computing, for which we look for sharp error bounds on the relative error. For this kind of building blocks, in base and precision p, we often successfully prove error bounds of the form α·u + o(u²) where α > 0 and u = 1/2·β1-p is the unit roundoff. To characterize the sharpness of such a bound, one can provide numerical examples for the standard precisions that are close to the bound, or examples that are parametrized by the precision and generate an error of the same form α·u + o(u²), thus proving the asymptotic optimality of the bound. However, the paper and pencil checking of such parametrized examples is a tedious and error-prone task. We worked on the formalization of a symbolicfloating-point arithmetic, over numbers that are parametrized by the precision, and implemented it as a library in the Maple computer algebra system. We also worked on the error analysis of the basic operations for complex numbers in floating-point arithmetic. We proved a very sharp error bound for an algorithm for the inversion of a complex number in floating-point arithmetic. This result suggests that the computation of a complex division according to x/y = (1/y)·x may be preferred, instead of the more classical formula x/y = (x·y)/|y|². Indeed, for any complex multiplication algorithm, the error bound is smaller with the algorithms described by the “inverse and multiply” approach.This is a joint work with my PhD advisors, with the collaboration of Claude-Pierre Jeannerod (CR Inria in AriC, at LIP).