Clustering extremal dependence structures via Wasserstein distances between angular measures

summary

The gist

Compound extreme events cause damage out of proportion to their frequency, and grouping random vectors by their joint extremal behaviour is a recurring task in climate science, hydrology…

In short

The method compares complete angular measures of extreme events using optimal transport via 1-Wasserstein distance. This nonparametric approach groups random vectors by their joint extremal behavior without needing a fitted model. It recovers latent partitions through clustering, providing a robust way to compare complex dependence structures.

Key concepts

Angular Measure
This characterizes how the magnitude of extreme events is distributed across different variables, derived from multivariate extreme-value theory. It involves decomposing the vector into radial (magnitude) and angular (distribution) components to capture the full shape of joint extremes.
1-Wasserstein Distance
This distance metric is used to compare the empirical angular measures after transforming them into a linear Euclidean space of Aitchison coordinates. It measures the minimum 'cost' to transport one set of observations (angular measures) into another, effectively comparing their underlying dependence structures.
Aitchison Coordinates
This transformation maps simplex-valued observations—which inherently live on a constrained space—into an unconstrained linear Euclidean space. This process removes the constraint while preserving the relative information between different extreme event patterns, making comparison mathematically tractable.
Consistency of Clustering
The method demonstrates that hierarchical clustering and K-medoids based on the resulting dissimilarity matrix reliably recover the true latent partition when certain conditions are met. This ensures that grouping based on these distances accurately reflects the underlying extremal dependence structure.

Terminology used across episodes

This episode discusses

The paper

Clustering extremal dependence structures via Wasserstein distances between angular measures · Read on arXiv

K. GASSER, J. SEGERS, F. RAGONE

Institute of Statistics, Biostatistics and Actuarial Sciences, LIDAM, UCLouvain · Department of Mathematics, KU Leuven and LIDAM/ISBA, UCLouvain · School of Computing and Mathematical Sciences, University of Leicester

Grouping random vectors by their joint extremal behaviour is a recurring task in climate science, hydrology, and finance. Classical clustering is driven by typical observations, while existing extremal procedures either summarize dependence by a single coefficient or discretize the simplex in an ad hoc way. We propose a dissimilarity that compares complete angular measures by optimal transport: each angular measure is estimated above a high radial threshold, mapped from the simplex into a Euclidean space of Aitchison coordinates, and compared by the 1-Wasserstein distance. The methodology is fully nonparametric and requires no fitted model. The dissimilarities feed any distance-based clustering algorithm. Under intermediate-threshold asymptotics and mild moment and separation conditions, hierarchical clustering and partitioning around medoids based on them recovers the latent partition with probability tending to one. In simulations, the proposed approach matches established competitors when clusters differ mainly in the strength of extremal dependence and performs substantially better when equal-strength clusters differ in the shape of their angular measures, especially when asymmetry drives the separation. Applications to rain--snowmelt--runoff extremes in Swiss catchments and to extremal co-movement between international equity markets and the US market reveal hydrologically and economically interpretable groups.

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Clustering extremal dependence structures via Wasserstein distances between angular measures".

Jane: Here's the gist of the paper. Compound extreme events cause damage out of proportion to their frequency, and grouping random vectors by their joint extremal behaviour is a recurring task in climate science, hydrology, and finance The proposed method compares complete angular measures by optimal transport using the 1-Wasserstein distance This approach is fully nonparametric and requires no fitted model Under intermediate-threshold asymptotics and mild moment and separation conditions, hierarchical clustering and partitioning around medoids based on them recovers the latent partition with probability tending to one The resulting procedure compares complete extremal dependence structures without summary coefficients, ad hoc discretization of the simplex, or fitted parametric or generative models.

Tom: The methodology relies on estimating angular measures from empirical observations and comparing them using Wasserstein distances The angular measure is characterized by the radial–angular decomposition from multivariate extreme-value theory The radial component is defined by the L1-norm of the transformed vector, and the angular component describes how this magnitude is distributed across different variables.

Jane: The empirical angular measure is defined using empirical probability integral transforms applied to rank observations above a high radial threshold.

Tom: Transformation to Euclidean Space To compare the angular measures, the authors map them into a linear Euclidean space of Aitchison coordinates using the centred log-ratio transformation This transformation removes the simplex constraint and maps angular observations to an unconstrained Euclidean space while preserving their relative information.

Title and authors: Jane: The paper title itself, "Clustering extremal dependence structures via Wasserstein distances between angular measures," tells you exactly what the method is doing—it's using optimal transport distance to compare how different variables behave in their extremes. But who wrote this? Tom, can you tell us a bit about the authors and what their expertise brings to this specific kind of problem?

Tom: Kasser and Segers are from institutions like LIDAM at UCLouvain and KU Leuven, with expertise spanning statistics, biostatistics, and mathematics. They bring that background in rigorous statistical learning—they understand how to build models on top of these complex geometric comparisons.

Lu: From my perspective as someone focused on representation learning, I'm most interested in the idea of mapping the simplex into a linear Euclidean space using Aitchison coordinates because that transformation is what makes comparing these angular measures feasible in that space. It unconstrains the problem while preserving the relative information between observations.

Meng: But what does this mean for practical application? If we are trying to classify rare, high-impact events in climate science or finance, how much more accurate is this geometric comparison compared to just looking at pairwise tail dependence coefficients?

Tom: Well, the simulation studies show that the Wasserstein–Aitchison method consistently outperforms partition-based methods and conditional extremes clustering across a wide range of extremal dependence structures. The authors found that this gain is largest precisely in situations where the latent groups cannot be separated by their pairwise tail dependence coefficients alone.

Jane: So, if we look at international equity markets, for instance, this suggests that just looking at how two stocks move together isn't enough; you need to capture the whole structure of extreme co-movement across all variables simultaneously to find meaningful clusters.

Lalam: I can confirm that the paper shows this method yields interpretable partitions in applications like separating Swiss catchments into different snowmelt regimes or grouping Western developed markets versus Asian and emerging markets based on their co-movement with the US market.

Tom: And they found that the agreement between this Wasserstein–Aitchison approach and conditional extremes clustering supports the robustness of the extremal-dependence structure they identified in those applications.

The paper's summary: Jane: So we’ve talked about how it works, but let's get into what they actually claim this paper achieves. Can you summarize for us in plain language what the main contribution of "Clustering extremal dependence structures via Wasserstein distances between angular measures" is?

Tom: The central point is that compound extreme events cause damage that's out of proportion to how often they happen, and grouping random vectors by their joint extremal behavior is a recurring task across climate science, hydrology, and finance. This paper proposes a dissimilarity measure that compares complete angular measures by optimal transport.

Lu: Essentially, they estimate these angular measures above a high radial threshold from the empirical observations, map them into Aitchison coordinates to get rid of the simplex constraint, and then compute the one-Wasserstein distance between these empirical measures in that linear space.

Jane: It seems like they are bypassing the need to fit any kind of generative model by comparing these complete angular measures directly through this transport distance. They are focusing on the geometry of dependence rather than just fitting a summary statistic for each variable.

Meng: So it’s not about predicting the next extreme event, but rather finding a way to organize and categorize existing extreme observations in a way that respects the complex dependencies between them. That shifts the focus from forecasting to structure discovery.

Tom: Right. And they establish consistency for clustering procedures like hierarchical clustering and global K-medoids, showing that under certain conditions—intermediate-threshold asymptotics and mild moment and separation conditions—these methods recover the true latent partition with probability tending to one.

Jane: That consistency result is key because it gives confidence that the clustering you get from this geometric comparison method is actually finding the structure they are looking for in real data, not just some random grouping.

The paper's improvements: Tom: The authors point out a few specific areas where this approach improves upon existing techniques. First, it removes the need for summary coefficients and ad hoc discretization of the simplex when comparing dependence structures. That’s a big deal because those methods are often overly simplistic.

Lu: They also address the issue with conditional extremes clustering, which requires fitting a conditional model for every random vector, noting that this can become computationally demanding for large collections of datasets. Their optimal transport framework offers an alternative structure discovery path.

Meng: From a deployment standpoint, if the method is non-parametric and doesn't require fitting a complex generative model to learn the angular measure—like the adversarial network they mention—that significantly lowers our operational hurdle for applying it to real traffic or market data.

Jane: And they also provide a way to distinguish between different angular measures that might share the same pairwise tail-dependence coefficients but differ in their orientation and higher-order dependence structure, allowing for finer discrimination of complex patterns. That means they can tell subtle differences apart.

Tom: Yes, that’s the ability to distinguish structure beyond just pairwise relationships. And they map the simplex into a linear Euclidean space using Aitchison coordinates, which decompresses the information and makes those differences between distributions easier to detect computationally and conceptually.

Conclusion: Jane: So we've covered how this method uses optimal transport to compare angular measures, its consistency for clustering, and its performance in simulations, especially showing it beats existing methods when pairwise dependence isn't enough. What is the final word on what this paper means for us as a community?

Tom: It means we have a way to cluster random vectors by their joint extremal behavior without needing any fitted model or ad hoc discretization of the simplex. This approach, called "Clustering extremal dependence structures via Wasserstein distances between angular measures," offers a complete comparison of dependence that is robust.

Lu: I think the most important part is that it provides an interpretable geometry for distinguishing bivariate extremal dependence structures when you look at complete trivariate angular measures. That gives us a better way to see the structure.

Meng: For deployment, if we can rely on this non-parametric framework to find groups in complex data like climate models or financial markets, we have a tool that doesn't break down as easily as methods that require fitting conditional models for every vector.

Lalam: I’m confident the results hold up because they show clear geographical and economic interpretations in their applications, and the agreement with other clustering methods validates the robustness of this identified extremal-dependence structure.

Tom: It's a solid contribution because it provides a complete comparison without needing heavy parametric assumptions or complex model fitting for every single event type. That's what we need when we are dealing with rare, high-impact events.

More episodes

← Home