Some negative results on extreme multivariate quantiles defined through convex optimisation

Stéphane Girard 1 Gilles Stupfler 2
1 MISTIS - Modelling and Inference of Complex and Structured Stochastic Systems
Inria Grenoble - Rhône-Alpes, LJK - Laboratoire Jean Kuntzmann, INPG - Institut National Polytechnique de Grenoble
2 School of Mathematical Sciences [ Nottingham]
Abstract : A discussion of some general properties that a notion of extreme multivariate quantile should satisfy will be given. We will then recall the concept of geometric quantile by transposing the definition of a univariate quantile as a minimiser of a cost function based on the so-called check function to the multivariate case. We shall then argue that extreme versions of these geometric quantiles are not suitable for the extreme-value analysis of a multivariate data set. A particular reason for this is that when the underlying distribution possesses a finite covariance matrix then the magnitude of these quantiles grows at a fixed rate that is independent of the distribution. We shall also discuss an extension of this negative result to the wider class of geometric M-quantiles.
Type de document :
Communication dans un congrès
10th International Conference of the ERCIM WG on Computing and Statistics, Dec 2017, London, United Kingdom. 2017
Liste complète des métadonnées
https://hal.archives-ouvertes.fr/hal-01667186
Contributeur : Stephane Girard <>
Soumis le : mardi 19 décembre 2017 - 10:48:36
Dernière modification le : mercredi 11 avril 2018 - 01:58:10

Identifiants

  • HAL Id : hal-01667186, version 1

Collections

INSMI | LJK | INRIA | INRIA2017 | UGA | LJK_PS | LJK_PS_MISTIS

Citation

Stéphane Girard, Gilles Stupfler. Some negative results on extreme multivariate quantiles defined through convex optimisation. 10th International Conference of the ERCIM WG on Computing and Statistics, Dec 2017, London, United Kingdom. 2017. 〈hal-01667186〉

Exporter

BibTeX TEI DC DCterms EndNote

Partager

Métriques

Consultations de la notice

140

Contact

support.ccsd.cnrs.fr
hal.support@ccsd.cnrs.fr
Logo CCSD