Chapitre D'ouvrage Année : 2018

A Reduced Product of Absolute and Relative Error Bounds for Floating-Point Analysis

Résumé

Rigorous estimation of bounds on errors in finite precision computation has become a key point of many formal verification tools. The primary interest of the use of such tools is generally to obtain worst-case bounds on the absolute errors. However, the natural bound on the elementary error committed by each floating-point arithmetic operation is a bound on the relative error, which suggests that relative error bounds could also play a role in the process of computing tight error estimations. In this work, we introduce a very simple interval-based abstraction, combining absolute and relative error propagations. We demonstrate with a prototype implementation how this simple product allows us in many cases to improve absolute error bounds, and even to often favorably compare with state-of-the art tools, that rely on much more costly relational abstractions or optimization-based estimations.

Fichier principal
Vignette du fichier
jacquemin_sas18.pdf (357.83 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-02633062 , version 1 (29-02-2024)

Licence

Identifiants

Citer

Maxime Jacquemin, Sylvie Putot, Franck Védrine. A Reduced Product of Absolute and Relative Error Bounds for Floating-Point Analysis. Static Analysis 25th International Symposium, SAS 2018, Freiburg, Germany, August 29–31, 2018, Proceedings, 11002, pp.223-242, 2018, Lecture Notes in Computer Science, ⟨10.1007/978-3-319-99725-4_15⟩. ⟨hal-02633062⟩
157 Consultations
240 Téléchargements

Altmetric

Partager

  • More