Inference on extremal dependence in a latent Markov tree model attracted to a H\"usler-Reiss distribution

Stefka Asenova, Gildas Mazo, Johan SegersarXiv 2020 · arXiv

Introduction

Predicting the propagation of extreme events, such as catastrophic floods, through complex networks is a critical task for risk management. While graphical models like Markov trees are powerful tools for representing conditional independence, they often conflict with the mathematical requirements of multivariate extreme value theory. Specifically, standard max-stable distributions typically lack the conditional independence properties that make tree structures computationally efficient. This paper addresses this gap by developing a framework for the tree-structured Hüsler–Reiss distribution, particularly in scenarios where data is missing at key network junctions.

The Hüsler–Reiss Model and Identifiability

The authors utilize the Hüsler–Reiss distribution, where extremal dependence is characterized by a variogram matrix. In a tree-structured model, the dependence between any two nodes is determined by the sum of edge-specific parameters along the unique path connecting them.

A central challenge in real-world networks is the presence of latent variables—junctions or nodes where no sensors are present. The paper establishes a rigorous "identifiability criterion": the parameters associated with a latent node can be uniquely recovered if and only if that node has a degree of at least three.

Figure 3: A latent node (shaded) with degree 3 allows for the identification of its surrounding edge parameters through the observed variables $X_1, X_3, X_4$.

As illustrated in Figure 3, a junction connecting three observed branches provides sufficient "triangulation" to solve for the hidden parameters. If a latent node only connects two branches (degree 2), it acts as a simple relay, and the individual edge parameters cannot be distinguished from their sum.

Application to the Seine River Network

The theoretical framework was applied to a dataset of high water levels from the Seine river network in France. The network includes five measurement stations: Paris, Meaux, Melun, Nemours, and Sens.

Figure 1: The Seine river network showing measurement stations (black dots) and latent nodes without data (red circles 2 and 5).

As shown in Figure 1, the network contains two critical unobserved junctions (nodes 2 and 5). Both nodes satisfy the identifiability criterion (degree $\ge 3$), allowing the authors to estimate the full set of edge parameters despite the missing data.

Estimation and Key Findings

The authors compared three estimation techniques: the Method of Moments (MME), Maximum Composite Likelihood (CLE), and Pairwise Extremal Coefficients (ECE). In both simulations and the Seine case study, MME and CLE demonstrated superior performance in terms of lower bias and variance.

The results highlight two major insights:

  1. Flow-Connectedness: Dependence between extreme water levels is driven more by hydrological "flow-connectedness" than by simple Euclidean distance. For example, Sens and Paris (flow-connected but distant) exhibited stronger tail dependence than Sens and Nemours (closer but not flow-connected).
  2. Model Robustness: The structured Hüsler–Reiss model provided an excellent fit for the observed extremal dependence, confirming that the path-sum variogram structure accurately captures the geometry of river networks.

Conclusion

This research provides a robust mathematical foundation for modeling extremes in tree-structured networks with latent variables. By identifying the "degree $\ge 3$" rule, the authors enable more accurate risk assessment in partially monitored systems, such as river basins or infrastructure networks, where installing sensors at every junction is often impractical.

Summary updated

Abstract

A Markov tree is a probabilistic graphical model for a random vector by which\nconditional independence relations between variables are encoded via an undirected tree\nand each node corresponds to a variable. One possible max-stable attractor for such a\nmodel is a Hüsler-Reiss extreme value distribution whose variogram matrix inherits its\nstructure from the tree, each edge contributing one free dependence parameter. Even if\nsome of the variables are latent, as can occur on junctions or splits in a river network, the\nunderlying model parameters are still identifiable if and only if every node corresponding\nto a missing variable has degree at least three. Three estimation procedures, based on the\nmethod of moments, maximum composite likelihood, and pairwise extremal coeffcients,\nare proposed for usage on multivariate peaks over thresholds data. The model and the\nmethods are illustrated on a dataset of high water levels at several locations on the Seine\nnetwork. The structured Hüsler-Reiss distribution is found to fit the observed extremal\ndependence well, and the fitted model confirms the importance of \now-connectedness\nfor the strength of dependence between high water levels, even for locations at large\ndistance apart.
BibTeX
@misc{asenova2020inference,
  author = {Stefka Asenova and Gildas Mazo and Johan Segers},
  title = {{Inference on extremal dependence in a latent Markov tree model attracted to a H\textbackslash\{\}"usler-Reiss distribution}},
  year = {2020},
  eprint = {2001.09510},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.09510},
}
Figure