Multivariate Nonparametric Estimation of the Pickands Dependence Function using Bernstein Polynomials

G Marcon, Simone A. Padoan, Philippe Naveau, Pietro Muliere, Johan Segers

Abstract

Many applications in risk analysis, especially in environmental sciences, require the estimation of the dependence among multivariate maxima. A way to do this is by inferring the Pickands dependence function of the underlying extreme-value copula. A nonparametric estimator is constructed as the sample equivalent of a multivariate extension of the madogram. Shape constraints on the family of Pickands dependence functions are taken into account by means of a representation in terms of a specific type of Bernstein polynomials. The large-sample theory of the estimator is developed and its finite-sample performance is evaluated with a simulation study. The approach is illustrated by analyzing clusters consisting of seven weather stations that have recorded weekly maxima of hourly rainfall in France from 1993 to 2011.

Introduction: The Challenge of Multivariate Extremes

In risk analysis, understanding the co-occurrence of extreme events—such as simultaneous flooding across multiple river basins—is critical. This requires modeling the dependence structure of multivariate extreme values. The mathematical core of this structure is the Pickands dependence function ($A$). For $A$ to be valid, it must satisfy strict shape constraints: it must be convex and bounded within a specific simplex.

Traditional nonparametric estimators often produce "noisy" results that violate these constraints, leading to physically impossible predictions. Furthermore, many existing methods suffer from the "curse of dimensionality," becoming computationally intractable as the number of variables increases beyond simple pairs.

Methodology: The BP-MD Approach

The authors propose a novel two-step estimator called the Bernstein Polynomial Madogram (BP-MD):

  1. Multivariate Madogram: The method begins with a multivariate extension of the madogram, a spatial dependence measure based on the difference between the maximum and the average of variables. This provides an initial "pilot" estimate that is easy to compute but mathematically unconstrained.
  2. Bernstein-Bézier Projection: To ensure the estimate is valid, the pilot estimate is projected onto a space of Bernstein polynomials. These polynomials are uniquely suited for enforcing convexity and boundary conditions.
  3. Optimization: The projection is framed as a quadratic programming problem, allowing for efficient computation even in higher dimensions (e.g., $d=7$).

Results and Discussion

Simulation studies using models like the Hüsler–Reiss and Extremal Skew-t demonstrated that the BP-MD estimator consistently outperforms raw estimators in terms of Mean Integrated Squared Error (MISE). By imposing shape constraints, the estimator achieves higher accuracy and produces smoother, more realistic dependence maps, particularly with smaller sample sizes.

Real-World Application: French Rainfall Patterns

The methodology was applied to analyze weekly maxima of hourly rainfall from 49 weather stations in France (1993–2011). The stations were grouped into seven clusters of seven stations each to analyze regional dependence.

Spatial distribution of the 49 weather stations in France, organized into seven clusters for multivariate analysis.

The analysis revealed a distinct latitudinal gradient in extreme rainfall dependence. In Northern France, extremes are often driven by large-scale frontal systems, leading to higher spatial dependence. In contrast, Southern France experiences more localized, intense convective storms, resulting in a different dependence structure. The BP-MD method successfully captured these nuances while maintaining the required mathematical properties of the Pickands function.

Conclusion

The use of Bernstein polynomials provides a robust framework for nonparametric estimation of multivariate dependence. By combining the flexibility of the madogram with the structural rigor of shape-constrained polynomials, the BP-MD estimator offers a scalable and physically consistent tool for environmental risk assessment. Future work will focus on optimizing the selection of the polynomial degree ($k$) and further improving computational scaling for even higher-dimensional datasets.

Reference: Marcon, G., Padoan, S. A., Naveau, P., Muliere, P., & Segers, J. (2016). Multivariate Nonparametric Estimation of the Pickands Dependence Function using Bernstein Polynomials. arXiv:1405.5228v3.

Summary updated

Abstract

Many applications in risk analysis, especially in environmental sciences, require the estimation of the dependence among multivariate maxima. A way to do this is by inferring the Pickands dependence function of the underlying extreme-value copula. A nonparametric estimator is constructed as the sample equivalent of a multivariate extension of the madogram. Shape constraints on the family of Pickands dependence functions are taken into account by means of a representation in terms of a specific type of Bernstein polynomials. The large-sample theory of the estimator is developed and its finite-sample performance is evaluated with a simulation study. The approach is illustrated by analyzing clusters consisting of seven weather stations that have recorded weekly maxima of hourly rainfall in France from 1993 to 2011.
BibTeX
@article{marcon2016multivariate,
  author = {G Marcon and Simone A. Padoan and Philippe Naveau and Pietro Muliere and Johan Segers},
  title = {{Multivariate Nonparametric Estimation of the Pickands Dependence Function using Bernstein Polynomials}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2016},
  doi = {10.1016/j.jspi.2016.10.004},
}

Loading related content...

Figure