Sparse regular variation

Nicolas Meyer, Olivier WintenbergerarXiv 2019 · OpenReview

Abstract

Regular variation provides a convenient theoretical framework to study large events. In the multivariate setting, the dependence structure of the positive extremes is characterized by a measure—the spectral measure—defined on the positive orthant of the unit sphere. This measure gathers information on the localization of extreme events and often has a sparse support since severe events do not simultaneously occur in all directions. However, it is defined through weak convergence which does not provide a natural way to capture this sparsity structure. In this paper, we introduce the notion of sparse regular variation which allows to better learn the dependence structure of extreme events. This concept is based on the Euclidean projection onto the simplex for which efficient algorithms are known. We prove that under mild assumptions sparse regular variation and regular variation are two equivalent notions and we establish several results for sparsely regularly varying random vectors. Finally, we illustrate on numerical examples how this new concept allows one to detect extremal directions.

Introduction

In risk management, climate science, and finance, understanding extreme events is paramount. In Multivariate Extreme Value Theory (EVT), the dependence between rare events is typically described by a "spectral measure" on a unit sphere. In high-dimensional settings, this measure is often sparse—meaning that when an extreme event occurs, it usually only affects a small subset of variables simultaneously. For instance, a massive storm might hit several coastal cities but not an entire continent.

Standard methods estimate this by examining the "angles" of extreme vectors ($X/|X|$). However, this approach has a significant flaw: mathematical convergence often fails on the low-dimensional boundaries where sparsity resides. Existing workarounds, such as the DAMEX algorithm, use "thickened" rectangles to capture these boundaries, but these require sensitive hyperparameters and can be computationally intensive.

Sparse Regular Variation and Euclidean Projection

The core innovation presented by Meyer and Wintenberger is the shift from radial normalization to Euclidean projection. Instead of simply dividing a vector by its norm, the authors propose projecting the vector $X/t$ onto the simplex. Unlike simple scaling, Euclidean projection can set small components exactly to zero, naturally capturing the sparsity of extreme events.

Euclidean projection of vectors onto the 2D simplex.

The authors define "sparse regular variation" (SRV) based on this projection $\pi$. They establish a "Main Theorem" showing that under mild conditions, SRV is mathematically equivalent to standard regular variation. This allows for the introduction of a new angular limit vector $Z$. While the traditional spectral vector $\Theta$ might contain many negligible non-zero values, $Z$ is naturally sparse, serving as a superior proxy for identifying truly dependent variables during extremes.

The geometry of the unit simplex in three dimensions.

Algorithmic Implementation and Results

The paper provides a threshold-based algorithm to identify "maximal directions"—the largest groups of variables likely to be extreme simultaneously. Because Euclidean projection can be computed in $O(d)$ expected time, the method is highly scalable.

The framework was validated using sophisticated simulated datasets where ground truth for extremal directions is known:

  • Asymptotic Independence: In a 40-dimensional Gaussian-based model, the method correctly identified all 40 axes as the only extremal directions with nearly zero errors, significantly outperforming the DAMEX algorithm.
  • Dependent Clusters: In a 50-dimensional model with 20 specific lower-dimensional clusters, the algorithm successfully recovered the clusters out of $2^{50}-1$ possibilities, demonstrating high precision and scalability.
  • Hierarchical Structures: A 60-dimensional case confirmed the algorithm's robustness in handling complex dependence structures.

Conclusion

Sparse regular variation offers a natural "denoiser" for high-dimensional extremes. By pushing non-extremal values to zero through Euclidean projection, it highlights true extremal directions without complex topological fixes. While the choice of the threshold parameter $p$ remains somewhat heuristic, the method's stability and theoretical equivalence to classical regular variation mark a significant advancement in multivariate extreme value theory. Future work will focus on deeper interpretations of non-maximal directions and applications to real-world high-dimensional datasets.

Summary updated

Abstract

Regular variation provides a convenient theoretical framework to study large events. In the multivariate setting, the dependence structure of the positive extremes is characterized by a measure - the spectral measure - defined on the positive orthant of the unit sphere. This measure gathers information on the localization of extreme events and has often a sparse support since severe events do not simultaneously occur in all directions. However, it is defined through weak convergence which does not provide a natural way to capture this sparsity structure.In this paper, we introduce the notion of sparse regular variation which allows to better learn the dependence structure of extreme events. This concept is based on the Euclidean projection onto the simplex for which efficient algorithms are known. We prove that under mild assumptions sparse regular variation and regular variation are two equivalent notions and we establish several results for sparsely regularly varying random vectors. Finally, we illustrate on numerical examples how this new concept allows one to detect extremal directions.
BibTeX
@misc{meyer2019sparse,
  author = {Nicolas Meyer and Olivier Wintenberger},
  title = {{Sparse regular variation}},
  year = {2019},
  eprint = {1907.00686},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.00686},
}

Loading related content...

Figure