A parameterization of anisotropic Gaussian fields with penalized complexity priors

1. Problem Statement

Gaussian Random Fields (GFs) are fundamental in spatial modeling, often represented as solutions to Stochastic Partial Differential Equations (SPDEs) to enable computationally efficient inference via Gaussian Markov Random Fields (GMRFs). While isotropic models (where correlation depends only on distance) are common, many real-world phenomena exhibit anisotropy (direction-dependent correlation).

The paper identifies two critical challenges in existing anisotropic modeling approaches:

1. Non-identifiability: Previous parameterizations (e.g., by Higdon et al., Paciorek & Schervish, Fuglstad et al.) often suffer from non-identifiability. Multiple parameter combinations yield the same covariance matrix, leading to multi-modal likelihoods and preventing continuous recovery of parameters, especially when extending to spatially varying models.
2. Prior Sensitivity: Under in-fill asymptotics, the likelihood provides limited information about the covariance structure. Consequently, the choice of prior distribution significantly impacts the posterior. There is a lack of principled, weakly informative priors that effectively penalize model complexity (overfitting) for anisotropic parameters.

2. Methodology

A. Identifiable Parameterization of Anisotropy

The authors propose a novel, smooth, and invertible parameterization for the anisotropy matrix $H$ in the SPDE:
$$(\kappa^2 - \nabla \cdot H \nabla) \frac{u}{\sigma_u} = \sqrt{4\pi}\kappa \dot{W}$$
where $H$ is a symmetric positive-definite matrix with determinant 1.

• The Approach: Instead of parameterizing $H$ directly or using angle-based vectors that suffer from sign ambiguity (e.g., $v$ and $-v$ producing the same $H$), the authors utilize the eigensystem of $H$.
• The "Half-Angle" Transformation: They define a parameter vector $v \in \mathbb{R}^2$ such that the eigenvector of $H$ corresponds to a "half-angle" of $v$. Specifically, if $v = (v_1, v_2)$, the eigenvector direction is derived from the complex number representation of $v$ halved.
• Result: This mapping is injective (one-to-one) and smooth. It ensures that the parameter space covers all valid anisotropy matrices without redundancy, resolving the multi-modality issue found in previous works. The parameters $(\kappa, v, \sigma_u)$ have clear geometric interpretations: $\kappa$ controls the length scale, $\sigma_u$ the marginal variance, and $v$ the direction and magnitude of anisotropy.

B. Penalized Complexity (PC) Priors

The authors construct PC priors for the parameters $(\kappa, v)$. PC priors are designed to penalize the complexity of a model relative to a simple base model.

• Base Model ($M_0$): The limit where the field becomes uncorrelated white noise (or a constant field depending on the limit definition), corresponding to $\kappa \to 0$ and $v \to 0$.
• Distance Metric: Standard Kullback-Leibler (KL) divergence is infinite for Gaussian fields with different spectral densities. The authors define a Sobolev seminorm based on the difference in rescaled spectral densities to measure the distance $d(\kappa, v)$ between the complex model and the base model.
  $$D^2(M_A, M_B) = \left( \int_{\mathbb{R}^2} \|2\pi\xi\|^4 \left( \frac{S_A(\xi)}{\sigma_A^2} - \frac{S_B(\xi)}{\sigma_B^2} \right)^2 d\xi \right)^{1/2}$$
• Prior Construction: The prior is constructed such that the distance $d(\kappa, v)$ follows an exponential distribution. This results in a joint prior that:
  1. Pushes the correlation range ($\rho \propto 1/\kappa$) toward infinity (isotropic, uncorrelated limit).
  2. Pushes the anisotropy magnitude ($\|v\|$) toward zero (isotropy).
  3. Is weakly informative, controlled by hyperparameters that can be set based on desired quantiles (e.g., "there is only a 5% chance the field is 10 times more correlated in one direction").
• Sampling: The authors derive a transformation to generate samples from this prior efficiently using standard Gaussian random variables.

3. Key Contributions

1. Identifiable Parameterization: A mathematically rigorous, smooth, and invertible parameterization for 2D anisotropic SPDEs that eliminates sign ambiguity and ensures unique parameter recovery.
2. PC Prior Formulation: The derivation of PC priors for anisotropic parameters using a Sobolev-based distance metric, providing a principled way to penalize complexity and avoid overfitting.
3. Computational Efficiency: The parameterization allows for efficient sampling and integration with existing SPDE/GMRF frameworks (e.g., INLA).
4. Validation: Comprehensive simulation studies and a real-world application demonstrating the superiority of these priors over non-informative alternatives.

4. Results

Simulation Study

The authors compared the proposed PC priors against Exponential-Gaussian (EG), Uniform, and Beta priors.

• Performance: The PC and EG priors (which are structurally similar) significantly outperformed Uniform and Beta priors.
• Complexity Penalization: PC priors effectively constrained the posterior to simpler models when data was scarce, resulting in narrower credible intervals and lower model complexity (distance to the base model).
• Calibration: The PC priors provided well-calibrated posterior distributions, whereas non-informative priors led to overfitting and poor coverage in low-data regimes.

Real-World Application: Norwegian Precipitation

The model was applied to annual precipitation data in Southern Norway ($N=233$).

• Anisotropy Detection: The anisotropic model detected significant anisotropy, estimating that precipitation is approximately 1.64 times more correlated in the North-South direction than East-West.
• Predictive Performance:
  • In low-data regimes (sub-sampled data), the anisotropic model with PC priors outperformed the isotropic model in terms of RMSE, CRPS, and DSS scores.
  • In high-data regimes, the performance gap narrowed, and the isotropic model became competitive, suggesting that with sufficient data, the likelihood dominates the prior.
• Spatial Variation: The results hinted that a non-stationary model might be even better, as the anisotropy appeared to vary across the region (e.g., different behavior near the coast vs. inland).

5. Significance

This work bridges a critical gap in spatial statistics by providing a robust framework for modeling anisotropy that is both computationally feasible and statistically principled.

• Solving Identifiability: It resolves long-standing issues with parameter non-identifiability in anisotropic SPDEs, making these models more reliable for Bayesian inference.
• Regularization: It introduces a rigorous method (PC priors) to regularize complex anisotropic models, preventing overfitting when data is limited—a common scenario in spatial applications.
• Future Directions: The authors note that while the current work focuses on stationary fields, the parameterization can be extended to non-stationary settings (where parameters vary in space) and higher dimensions, though the "half-angle" construction requires generalization for $d > 2$. They also identify the extension to heavy-tailed fields (Confluent Hypergeometric class) as a promising future avenue.

In conclusion, the paper advocates for the use of informative, complexity-penalizing priors in spatial Gaussian modeling, demonstrating that such approaches yield more reliable inferences and better predictive performance, particularly in data-scarce environments.