Extreme geometric quantiles in a multivariate regular variation framework

Stéphane Girard, Gilles StupflerExtremes 2015 · DOI

TL;DR

This paper extends the concept of extreme quantiles to multivariate settings by introducing "extreme geometric quantiles" within a multivariate regular variation framework. This work is crucial for understanding the tails of multivariate distributions, especially when traditional integrability assumptions are violated, and provides a theoretical foundation for robust statistical inference in extreme value analysis.

What Problem Does This Paper Address?

Understanding extreme events is critical in many fields, from finance to environmental science. While univariate extreme value theory is well-developed, extending these concepts to multivariate distributions has been challenging due to the lack of a universally accepted definition of multivariate quantiles and tails. This paper tackles this by focusing on "geometric quantiles," which generalize univariate quantiles by considering both direction and magnitude. Specifically, it addresses the behavior of these geometric quantiles when their "order" approaches the extreme tail of the distribution (i.e., when the norm of the quantile vector approaches 1), particularly in cases where the data's tails are "heavy" (violating standard integrability conditions).

Methods

The authors build upon the existing framework of geometric quantiles and introduce a multivariate regular variation condition (denoted as (Ma)) to model the tail behavior of multivariate distributions. This condition allows for a more general class of heavy-tailed distributions than previously considered. They derive asymptotic expansions for the direction and magnitude of extreme geometric quantiles under this framework. Key techniques include:

  • Multivariate Regular Variation: A generalization of the concept of regular variation to higher dimensions, used to describe the tail behavior of probability density functions.
  • Asymptotic Analysis: Deriving the limiting behavior of geometric quantiles as their order approaches 1.
  • Lemmas and Propositions: Establishing several analytical results (Lemmas and Propositions) that are crucial for proving the main theorems, particularly concerning the properties of regularly varying functions and their inverses.
  • Numerical Illustrations: Using a Pareto distribution model to numerically compare the derived asymptotic approximations with exact computations of extreme geometric quantiles.

Datasets

Not discussed in the paper. The paper focuses on theoretical derivations and uses a simulated Pareto distribution for numerical illustrations.

Key Insights and Ideas

  • Bridging Univariate and Multivariate Extremes: The paper shows that under certain conditions (specifically, when the tail index $\alpha$ is large enough), the behavior of multivariate extreme geometric quantiles mirrors that of univariate extreme quantiles, where the magnitude is directly influenced by the tail decay of the distribution.
  • Role of the Tail Index ($\alpha$): The parameter $\alpha$ (the index of multivariate regular variation) plays a critical role in determining the asymptotic behavior of extreme geometric quantiles. The paper highlights how different ranges of $\alpha$ lead to distinct asymptotic behaviors and requires different analytical approaches.
  • Generalization Beyond Integrability: The work extends the analysis of extreme geometric quantiles to cases where the second moment (and even the first moment) of the random vector might not be finite, which is a significant step beyond previous results that relied on such integrability assumptions.
  • Connection to Weissman Estimators: The asymptotic expansions derived open the door to developing multivariate versions of Weissman-type estimators for extreme quantiles, allowing for extrapolation from intermediate quantiles to extreme ones.

Key Results

The paper presents several key theorems and corollaries that provide asymptotic expansions for the magnitude and direction of extreme geometric quantiles under different conditions of multivariate regular variation (characterized by the tail index $\alpha$).

  • For $\alpha \in (0, 2)$: The magnitude of the extreme geometric quantile behaves roughly like $(1-\lambda)^{-1/\alpha}$ as $\lambda \uparrow 1$, directly reflecting the tail behavior of the distribution.
  • For $\alpha = 1$ and $\alpha = 2$: These are identified as critical values where the asymptotic behavior changes, requiring specific analytical treatments.
  • Numerical results: Figures 2-6 demonstrate that the asymptotic approximations generally align well with numerically computed quantiles, especially for larger values of $\alpha$. The agreement is better for larger $\alpha$, and the shape of the iso-quantile curves can be significantly influenced by $\alpha$.

Limitations and Future Work

  • Computational Complexity: While the paper provides theoretical results and numerical illustrations, the actual computation of extreme geometric quantiles can be computationally intensive, especially in high dimensions.
  • Estimation of $\alpha$: The paper mentions that the estimation of the tail index $\alpha$ is a separate challenge that would be addressed in future research.
  • Asymptotic Distribution of Estimators: The paper suggests that studying the asymptotic distribution of estimators for extreme geometric quantiles (like the proposed sample counterpart of the minimization problem) is a promising avenue for future work.
  • High-Dimensional Cases: The paper focuses on theoretical results, and their practical application in very high-dimensional settings might require further investigation into computational feasibility and robustness.

Summary updated

BibTeX
@inproceedings{girard2015extreme,
  author = {Stéphane Girard and Gilles Stupfler},
  title = {{Extreme geometric quantiles in a multivariate regular variation framework}},
  booktitle = {Extremes},
  year = {2015},
  doi = {10.1007/s10687-015-0226-0},
}
Figure