PhD defense Diane Gallois-Wong

Formalisation en Coq des algorithmes de filtre numérique calculés en précision finie
by Diane Gallois-Wong
Thursday 04 March 2021 at 10h00
room 445 bâtiment 650 as well as online

Abstract: Digital filters have numerous applications, from telecommunications to aerospace. To be used in practice, a filter needs to be implemented using finite precision (floating- or often fixed-point arithmetic). Resulting rounding errors may become especially problematic in embedded systems, where tight time, space, and energy constraints mean that we often need to cut into the precision of computations in order to improve their efficiency. Moreover, digital filter programs are strongly iterative: rounding errors may propagate and accumulate through many successive iterations.

As some of the application domains are critical, I study rounding errors in digital filter algorithms using formal methods to provide stronger guaranties. More specifically, I use Coq, a proof assistant that ensures the correctness of this numerical behavior analysis. I aim at providing certified error bounds over the difference between outputs from an implemented filter (computed using finite precision) and from the original model filter (theoretically defined with exact operations). Another goal is to guarantee that no catastrophic behavior (such as unexpected overflows) will occur.

Using Coq, I define linear time-invariant (LTI) digital filters in time domain. I formalize a universal form called SIF: any LTI filter algorithm may be expressed as a SIF while retaining its numerical behavior. I then prove two theorems that allow us to analyze this numerical behavior. This analysis also involves the sum-of-products algorithm used during the computation of the filter. Therefore, I formalize several sum-of-products algorithms, that offer various trade-offs between output precision and computation speed. This includes a new algorithm whose output is correctly rounded-to-nearest. I also formalize modular overflows, and prove that one of the previous sum-of-products algorithms remains correct even when such overflows are taken into account.

Jury:

  • Yves BERTOT, Directeur de recherche, Inria Sophia Antipolis-Méditerranée - Reviewer
  • Éric FERON, Professor, School of Aerospace Engineering, États-Unis - Reviewer
  • Jérôme FERET, Chargé de recherche, Inria Paris & DI-ÉNS, ÉNS/CNRS/Université PSL - Examiner
  • Florent HIVERT, Professeur, LISN, Université Paris-Saclay - Examiner
  • Mioara JOLDES, Chargée de recherche, LAAS-CNRS, CNRS Toulouse - Examiner
  • Assia MAHBOUBI, Chargée de recherche, Inria Rennes-Bretagne Atlantique, LS2N CNRS, Nantes - Examiner
  • Jean-Michel MULLER, Directeur de recherche, LIP, CNRS Lyon - Examiner
  • Sylvie BOLDO, Directrice de recherche, Inria Saclay - Île-de-France - Thesis advisor
  • Thibault HILAIRE, Maître de conférence, LIP6, Sorbonne Université - Thesis co-advisor