Non-Gaussian Spatiotemporal Random Fields

Updated 2 May 2026

Non-Gaussian spatiotemporal random fields are probability models with joint distributions that deviate from Gaussianity, capturing heavy tails, skewness, and nonlinear dependencies.
They are constructed using frameworks like TGMRF, SPDE-driven models, and dynamic scale-mixture approaches that preserve spatial-temporal dependencies while allowing flexible marginal behaviors.
These models facilitate rigorous theoretical analysis and efficient inference, enabling practical applications in ecology, environmental sciences, and spatial extremes.

Non-Gaussian spatiotemporal random fields are probability models for spatial-temporal processes whose joint finite-dimensional distributions do not exhibit Gaussianity. Explicitly modeling non-Gaussianity is technically essential in a variety of domains where interactions, marginal distributions, or tail risks are inadequately captured by the Gaussian assumption. Advances in theory, computational methods, and applications have allowed construction and inference for broad classes of non-Gaussian fields, including non-separable, heavy-tailed, skewed, or dynamically heterogeneous models. Below, key frameworks and results are surveyed, organized across foundational principles, construction methodologies, dependence structures, limit theorems, geometric analysis, and model-based applications.

1. Frameworks and Definitions

A non-Gaussian spatiotemporal random field (STRF) is a collection $\{Y(s, t): \, s \in D \subset \mathbb{R}^d,\, t\in T\}$ such that for at least some finite set of space-time indices, the joint law of $\{Y(s_i, t_i)\}$ differs from the multivariate normal. Typically, these fields retain second-order structure—mean and covariance—while introducing higher-order dependencies, marginal non-normality, and possibly nonlinearity or asymmetry.

Several foundational paradigms are prevalent:

Subordination: Nonlinear transformations of latent Gaussian fields, $Y(x, t) = g(Z(x, t))$ , induce non-Gaussianity while preserving (through copula mechanisms) the latent dependency structure (Angulo et al., 2022).
Scale-mixing: The marginal variance or scale process, possibly correlated and dynamic, modulates Gaussian noise to yield heavy-tailed or skewed fields. Log-Gaussian and other mixture structures fall in this class (Fonseca et al., 2021).
SPDE-driven models: Solutions to stochastic partial differential equations driven by non-Gaussian noises (e.g., Normal-inverse Gaussian, Generalized Asymmetric Laplace) extend classic Matérn fields into the non-Gaussian regime (Lindgren et al., 2021).
Lancaster-Sarmanov expansions: Joint densities of fields expanded in orthogonal polynomial bases (Hermite for Gaussian-based, Laguerre for gamma/chisquare), with nonlinear subordination controlling higher-order structure (Angulo et al., 2022).

2. Main Construction Methodologies

2.1 Transformed Gaussian Markov Random Fields (TGMRF)

The TGMRF framework provides a general recipe for non-Gaussian, non-separable STRFs (Azevedo et al., 2020):

Start with a latent space-time Gaussian Markov random field, $Z \sim N_{nT}(0, Q^{-1}(\theta))$ , where $\theta$ encapsulates spatial, temporal, and cross-dependence parameters.
Transform $Z$ componentwise by $Y_{it} = g_{it}(Z_{it}) = F_{it}^{-1}[\Phi(Z_{it})]$ for desired marginal CDFs $F_{it}$ .
Typical mappings include gamma-quantile, log-normal, beta, or Weibull.
The dependence among $\{Y_{it}\}$ is governed by the Gaussian copula structure of $Z$ , while arbitrary univariate marginals are imposed.

2.2 SPDE and Non-Gaussian Noise

Non-Gaussian STRFs can be defined as solutions to SPDEs driven by non-Gaussian independently scattered random measures (e.g., NIG, GAL). Discretization yields tractable conditional-Gaussian representations (Lindgren et al., 2021):

$\{Y(s_i, t_i)\}$ 0, with $\{Y(s_i, t_i)\}$ 1 a differential operator (e.g., Matérn type), and $\{Y(s_i, t_i)\}$ 2 noise process with chosen non-Gaussian law.
After finite element discretization and introduction of latent variance/scale variables, conditionally Gaussian procedures enable efficient inference.

2.3 Dynamic Scale-Mixture Models

Dynamic non-Gaussian models leverage scale-mixture forms:

$\{Y(s_i, t_i)\}$ 3, where $\{Y(s_i, t_i)\}$ 4 is Gaussian, $\{Y(s_i, t_i)\}$ 5 is a spatiotemporally correlated scale process (decomposed as $\{Y(s_i, t_i)\}$ 6 with log-Gaussian or dynamic linear modeling for the components (Fonseca et al., 2021).

3. Dependence Structures and Non-separability

Constructing non-Gaussian STRFs with realistic dependence properties requires careful specification of covariance or precision structures.

Non-separable Precision Matrices: For TGMRFs, the precision matrix is $\{Y(s_i, t_i)\}$ 7, embedding spatial ( $\{Y(s_i, t_i)\}$ 8), temporal ( $\{Y(s_i, t_i)\}$ 9), and space-time ( $Y(x, t) = g(Z(x, t))$ 0) interactions in a flexible, sparse GMRF framework (Azevedo et al., 2020).
SPDE-induced Dependence: Operator $Y(x, t) = g(Z(x, t))$ 1 in the SPDE form sets marginal smoothness, range, and anisotropy. Non-stationarity is encoded via spatially (or temporally) varying coefficients in $Y(x, t) = g(Z(x, t))$ 2. On manifolds, $Y(x, t) = g(Z(x, t))$ 3 is replaced by Laplace–Beltrami operators (Lindgren et al., 2021).
Lancaster–Sarmanov Expansions: Expansion coefficients in the orthogonal polynomial basis relate higher-order dependencies for pairs at distance $Y(x, t) = g(Z(x, t))$ 4 via function $Y(x, t) = g(Z(x, t))$ 5, generalizing the covariance function.

Non-separability is realized structurally, e.g., by including terms like $Y(x, t) = g(Z(x, t))$ 6 in the GMRF precision or by using non-separable diffusion terms in SPDEs.

4. Limit Theorems and Asymptotic Behavior

The asymptotics of integrated or smoothed functionals of non-Gaussian STRFs reveal profound differences from classical Gaussian fields.

4.1 Subordinated Fields and Noncentral Limit Theorems

If $Y(x, t) = g(Z(x, t))$ 7 is a non-linear transformation of an LRD (long-range dependent) Gaussian field, the integrated field can exhibit non-Gaussian limit laws (Ruiz-Medina, 11 Feb 2026).

Example: For $Y(x, t) = g(Z(x, t))$ 8 with Hermite rank 2, e.g., $Y(x, t) = g(Z(x, t))$ 9, normalized integrals converge weakly to Rosenblatt-type distributions (second Wiener chaos) under LRD scaling.
Both pure-point (manifolds) and continuous (Euclidean) spectral cases are covered, with explicit scaling exponents dictating convergence rates.

4.2 Information Decay and Structural Complexity

Shannon mutual information between distant sites decays as $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 0, with decay accelerated by the polynomial rank $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 1 of the subordinating function and the LRD exponent $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 2. Generalizations to Rényi MI yield decay rate $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 3 for order $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 4 (Angulo et al., 2022). This quantifies macroscale “diversity loss” and structural complexity.

5. Geometric and Topological Descriptors

Beyond classical cumulants and spectra, non-Gaussianity is directly expressed in the stochastic geometry and topology of fields (Beuman et al., 2012):

Critical Point Statistics: Imbalance in densities of maxima and minima ( $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 5), created by local or nonlocal nonlinearities.
Minkowski Functionals: Corrections to area, boundary, and Euler characteristic of excursion sets, scaling with field cumulants.
Umbilics and Curvature: Fraction of monstar-type umbilical points shifts proportionally to higher-order derivatives (e.g., $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 6 from cubic terms).

For small non-Gaussianity, explicit perturbative expansions relate these geometric functionals to the cumulants generated by nonlinear perturbations or dynamical mechanisms (e.g., KPZ dynamics).

6. Statistical Inference and Model Implementation

Inference for non-Gaussian STRFs is computationally tractable via conditional-Gaussian construction, sparse representations, and data augmentation algorithms:

Likelihood-based: Exploiting sparsity in the latent GMRF precision or SPDE operators, using MCMC, EM, or INLA-type approaches (Azevedo et al., 2020, Lindgren et al., 2021).
Augmentation: Data-augmentation (e.g., via latent scale variables) for efficient Gibbs or Metropolis-Hastings block updates (Fonseca et al., 2021).
Model Selection: Comparison via predictive scores (WAIC, LPML, DIC, scoring rules) and diagnostic checks (over/under-dispersion, autocorrelation).
Software: Packages such as TGMRF, R-INLA, rSPDE, and ngme encapsulate many of these methods (Lindgren et al., 2021, Azevedo et al., 2020).

7. Empirical Applications and Illustrative Studies

Non-Gaussian STRF methods have been applied to ecological data, environmental processes, and spatial extremes:

TGMRF in Ecology: The abundance of Nenia tridens across $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 7 sites and $Z \sim N_{nT}(0, Q^{-1}(\theta))$ 8 years was analyzed using gamma and log-normal marginals in a spatial-temporal TGMRF. Gamma-shape and gamma-scale models outperformed log-normal, with critical environmental covariates identified through the hierarchical model fit (Azevedo et al., 2020).
Dynamic Non-Gaussian Models: Maximum temperature and ozone processes were more accurately predicted with scale-mixture dynamic models, yielding improved uncertainty quantification compared to Gaussian models (Fonseca et al., 2021).
Macroscopic Complexity: Power-law decay of mutual information in subordinated fields quantifies diversity and memory for spatial and spatiotemporal models, with explicit rate dependence on functional rank and LRD exponent (Angulo et al., 2022).

In summary, non-Gaussian spatiotemporal random fields encompass a diverse array of theoretical constructs and model structures, enabling flexible modeling of complex spatial-temporal phenomena with heavy tails, skewness, asymmetry, and non-separable dependencies. The joint evolution of theoretical tools—subordination, SPDEs, TGMRFs and geometric quantification—has yielded a robust modeling toolkit, amenable both to rigorous limit theory and scalable inferential procedures for scientific applications (Azevedo et al., 2020, Angulo et al., 2022, Beuman et al., 2012, Ruiz-Medina, 11 Feb 2026, Lindgren et al., 2021, Fonseca et al., 2021).