Efficient Calculation of Jacobians Using Dynamic Programming

Uwe Naumann 1
1 TROPICS - Program transformations for scientific computing
CRISAM - Inria Sophia Antipolis - Méditerranée
Abstract : The chain rule - fundamental for Automatic Differentiation (AD) - can be applied to computational graphs representing vector functions in arbitrary orders resulting in different operations counts for the calculation of their Jacobian matrices. Very few authors have looked at this interesting subject so far and there is no generally accepted terminology for dealing with these combinations of the forward and reverse modes of AD. The minimizati- on of the number of arithmetic operations required for the calculation of the complete Jacobian leads to a computationally hard combinatorial optimization problem. In this paper we will describe an approach to the solution of the edge elimination problem in computational graphs that builds on the idea of optimizing chained matrix products using dynamic programming techniques. We will discuss the theory and present some test results by regarding this approach in comparison with other methods for computing Jacobians using a minimal number of arithmetic operations.
Type de document :
RR-3689, INRIA. 1999
Liste complète des métadonnées

Littérature citée [1 références]  Voir  Masquer  Télécharger

Contributeur : Rapport de Recherche Inria <>
Soumis le : mercredi 24 mai 2006 - 11:30:33
Dernière modification le : samedi 27 janvier 2018 - 01:31:34
Document(s) archivé(s) le : dimanche 4 avril 2010 - 23:30:12



  • HAL Id : inria-00072980, version 1



Uwe Naumann. Efficient Calculation of Jacobians Using Dynamic Programming. RR-3689, INRIA. 1999. 〈inria-00072980〉



Consultations de la notice


Téléchargements de fichiers