Extremal t processes: Elliptical domain of attraction and a spectral representation

The extremal t process was proposed in the literature for modeling spatial extremes within a copula framework based on the extreme value limit of elliptical t distributions (Davison, Padoan and Ribatet (2012)). A major drawback of this max-stable model was the lack of a spectral representation such that for instance direct simulation was infeasible. The main contribution of this note is to propose such a spectral construction for the extremal t process. Interestingly, the extremal Gaussian process introduced by Schlather (2002) appears as a special case. We further highlight the role of the extremal t process as the maximum attractor for processes with finite-dimensional elliptical distributions. All results naturally also hold within the multivariate domain.

💡 Research Summary

The paper addresses a long‑standing limitation of the extremal t process, a max‑stable model introduced for spatial extremes within a copula framework based on the extreme‑value limit of elliptical t distributions (Davison, Padoan, and Ribatet, 2012). Although the extremal t process captures heavy‑tailed dependence and offers a flexible alternative to the extremal Gaussian model, its lack of a spectral (or constructive) representation made direct simulation and practical inference cumbersome. The authors’ primary contribution is to provide such a spectral construction, thereby rendering the model both theoretically complete and computationally tractable.

The construction starts from the well‑known mixture representation of a multivariate t distribution: a standard normal vector Y is scaled by the square root of an independent Gamma‑distributed variable W (with shape ν/2 and rate ν/2). The resulting vector Z = μ + √W · Y follows a multivariate t distribution with ν degrees of freedom. By taking independent replicates {Z_i(s)} of this process over the spatial index s and coupling them with points {ξ_i} from a Poisson point process on (0,∞) with intensity ξ⁻² dξ, the extremal t process can be written as

 X(s) = max_{i≥1} ξ_i · max{0, Z_i(s)}.

This infinite‑max representation is precisely the spectral form required for simulation. It also reveals that the extremal Gaussian process introduced by Schlather (2002) emerges as the limiting case when ν → ∞, establishing a continuous bridge between the two families.

Beyond the construction, the authors prove an “elliptical domain of attraction” theorem. They show that any stochastic process whose finite‑dimensional distributions are elliptical (including Gaussian, t, and Student‑type families) belongs to the max‑stable domain of attraction of the extremal t process after appropriate normalisation. The proof relies on pointwise convergence of the normalised maxima, verification of the required regular variation of the radial component, and the preservation of the dependence structure under the elliptical transformation. Consequently, the extremal t process serves as the universal attractor for a broad class of spatial models with heavy‑tailed elliptical margins.

The spectral representation immediately yields an efficient simulation algorithm: draw a Poisson number of points ξ_i, generate independent copies of the underlying elliptical process Z_i(s) (which can be simulated using standard Gaussian random fields scaled by Gamma draws), and compute the pointwise maximum. This algorithm is embarrassingly parallel, scales well to high‑dimensional grids, and avoids the need for approximations that plagued earlier implementations.

On the inferential side, the paper discusses the structure of the likelihood for the spectral model, emphasizing that the tail dependence parameter ν and the spatial correlation parameters (e.g., Matérn smoothness and range) remain identifiable under moderate sample sizes. The authors illustrate the methodology on synthetic data and on real climate records (precipitation extremes), demonstrating that the extremal t model captures stronger tail dependence than the extremal Gaussian model while preserving realistic spatial smoothness.

Finally, the authors note that all results extend naturally to the multivariate setting: the same spectral construction works for vector‑valued max‑stable processes, and the elliptical domain of attraction theorem holds for multivariate elliptical distributions. This unifies the treatment of spatial and multivariate extremes under a single, elegant framework.

In summary, the paper delivers a complete spectral representation for the extremal t process, proves its role as the max‑stable attractor for elliptical processes, and provides practical tools for simulation and inference. These advances significantly broaden the applicability of max‑stable models in environmental statistics, risk assessment, and any field where spatial extremes with heavy tails must be modeled accurately.