Fast and flexible inference for spatial extremes

Peng Zhong, Scott A. Sisson, Boris BérangerarXiv 2024 · DOI · arXiv

TL;DR

This paper introduces a theoretical framework to make modeling spatial extremes (like massive floods or heatwaves) both faster and more flexible. By establishing conditions for "continuous exponent measures," the authors derive two new models—the skewed Brown-Resnick and truncated extremal-$t$—and demonstrate a spectral likelihood inference method that is up to 470 times faster than traditional approaches while capturing complex, non-stationary patterns.

What Problem Does This Paper Address?

Predicting extreme events across a map is a high-stakes challenge for risk management and environmental science. Statistically, we use "max-stable" or "$r$-Pareto" processes to model how an extreme event at one location relates to another.

However, researchers face a "curse of dimensionality." When you try to model thousands of locations simultaneously, the math becomes a "combinatorial explosion" of terms. To stay computationally feasible, most researchers stick to the Brown-Resnick model because it possesses a specific mathematical property: a continuous exponent measure. Other flexible models, like the extremal-$t$, have "mass on the boundaries" (discontinuities), which forces researchers to use slow, approximate methods like pairwise composite likelihoods that often miss higher-order spatial dependencies.

Methods

The authors tackle this by first proving Theorem 1, which provides the "holy grail" condition: a simple probabilistic check to see if a model's intensity function concentrates all its mass in the interior of its domain (ensuring continuity).

Using this theorem, they build two new tools:

  1. The Skewed Brown-Resnick Model: An extension of the classic model that incorporates "skewness" via spline functions. This allows the model to handle non-stationarity—where the strength and direction of dependence change depending on where you are on the map.
  2. The Truncated Extremal-$t$ Model: A clever modification of the popular extremal-$t$ model that "trims" the distribution to remove the problematic boundary mass, making it compatible with fast inference.

For estimation, they leverage spectral likelihoods. Unlike standard methods that look at pairs of points, this looks at the "intensity" of the entire extreme event, which is statistically more efficient and computationally a breeze.

Datasets

The paper validates these models using:

  • Synthetic Data: Extensive simulations to compare speed and bias against traditional "pairwise" methods.
  • Florida Rainfall: Radar measurements from the Tampa Bay area (1995–2019) across a 2km grid, totaling over 139,000 images with 4,449 spatial observations each.
  • Red Sea Surface Temperature: 31 years of annual maxima (1985–2015) recorded at 1,043 gridded locations.

Key Insights and Ideas

  • The "Interior" is Key: The authors show that if a model's underlying stochastic process has a zero probability of hitting zero at any location (conditional on other locations), the resulting extreme value model will be "continuous." This is the secret sauce for fast math.
  • Flexibility without the Lag: Traditionally, adding "non-stationarity" (making the model more realistic) meant making it slower. The skewed Brown-Resnick model breaks this trade-off.
  • Better Simulations: The paper provides a new rejection sampling algorithm for $r$-Pareto processes that is significantly more efficient than previous state-of-the-art methods, especially as the number of locations increases.

Key Results

  • Massive Speedups: In simulations, the spectral likelihood approach was 68 to 476 times faster than the standard pairwise composite likelihood.
  • Superior Accuracy: In the Florida rainfall application, the skewed model achieved a lower AIC (Akaike Information Criterion), proving it better captured the local, anisotropic (direction-dependent) rainfall patterns that the standard Brown-Resnick model ignored.
  • Scalability: The authors successfully fit models to over 1,000 locations in the Red Sea dataset in under an hour—a task that typically bogs down standard spatial extreme models.

Limitations and Future Work

The primary limitation is that spectral likelihood estimates are biased. Because they rely on a Poisson point process approximation, you need to set a very high "extremeness threshold" to keep the bias low. If the threshold is too high, you lose data; if it's too low, your estimates drift.

The authors suggest that future work could explore "censored" versions of these likelihoods to further mitigate bias in lower dimensions. They also see a major opportunity to apply these flexible dependence structures to extremal graphical models, which are currently limited by the same restrictive assumptions this paper has begun to dismantle.

Summary updated

Abstract

Statistical modelling of spatial extreme events has gained increasing attention over the last few decades with max-stable processes, and more recently $r$-Pareto processes, becoming the reference tools for the statistical analysis of asymptotically dependent data. Although inference for r-Pareto processes is easier than for max-stable processes, there remain major hurdles for their application to high dimensional datasets within a reasonable timeframe. In addition, both approaches have almost exclusively focused on the Brown-Resnick model, for its Gaussian foundations, and for the continuity of its exponent measure. In this paper, we derive a class of models for which this continuity property holds and present the skewed Brown-Resnick model, an extension of the Brown-Resnick that allows for non-stationarity in the dependence structure, and the truncated extremal-t model, a refinement of the well-known extremal-$t$ model. We use an inference methodology based on the intensity function of the process which is derived from the exponent measure, and demonstrate the statistical and computational efficiency of this approach. Applications to two real-world problems illustrate valuable gains in modelling flexibility as well as appealing computational gains over reference methodologies.
BibTeX
@misc{zhong2024fast,
  author = {Peng Zhong and Scott A. Sisson and Boris Béranger},
  title = {{Fast and flexible inference for spatial extremes}},
  year = {2024},
  eprint = {2407.13958},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.13958},
}
Figure