Measures minimizing regularized dispersion

Abstract : We consider a continuous extension of a regularized version of the minimax, or dispersion , criterion widely used in space-filling design for computer experiments and quasi-Monte Carlo methods. We show that the criterion is convex for a certain range of the regularization parameter (depending on space dimension) and give a necessary and sufficient condition characterizing the optimal distribution of design points. Using results from potential theory, we investigate properties of optimal measures. The example of design in the unit ball is considered in details and some analytic results are presented. Using recent results and algorithms from experimental design theory, we show how to construct optimal measures numerically. They are often close to the uniform measure but do not coincide with it. The results suggest that designs minimizing the regularized dispersion for suitable values of the regularization parameter should have good space-filling properties. An algorithm is proposed for the construction of n-point designs.
Type de document :
Article dans une revue
Journal of Scientific Computing, Springer Verlag, In press
Liste complète des métadonnées

Littérature citée [22 références]  Voir  Masquer  Télécharger

https://hal.archives-ouvertes.fr/hal-01864088
Contributeur : Luc Pronzato <>
Soumis le : mercredi 29 août 2018 - 13:46:15
Dernière modification le : mercredi 12 septembre 2018 - 01:16:14

Fichier

dispersion-opt-measures_Spring...
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01864088, version 1

Collections

Citation

Luc Pronzato, Anatoly Zhigljavsky. Measures minimizing regularized dispersion. Journal of Scientific Computing, Springer Verlag, In press. 〈hal-01864088〉

Partager

Métriques

Consultations de la notice

19

Téléchargements de fichiers

8