Bayesian Spatiotemporal Wombling

Technical Summary: Bayesian Spatiotemporal Wombling

Problem Statement
Statistical inference regarding rates of change in spatial random fields is critical for scientific discovery, a concept historically referred to as "wombling" (Womble, 1951). While existing methods effectively identify zones of rapid change in purely spatial settings (often as curves), they face significant limitations when applied to spatiotemporal data. In spatiotemporal contexts, boundaries tracking rapid change evolve over time, forming surfaces rather than static curves. Furthermore, many scientific phenomena (e.g., neural activity or pollutant dispersion) require inference on higher-order derivatives (such as curvature) rather than just first-order gradients. Current approaches, including machine learning clustering, image segmentation, and level set methods, often lack fully model-based uncertainty quantification or are restricted to detecting static regions rather than evolving boundaries. There is a need for a framework that formalizes "wombling" surfaces in space-time, handles higher-order differential processes, and provides rigorous Bayesian inference with uncertainty quantification.

Methodology
The authors develop a fully model-based Bayesian inferential framework for spatiotemporal wombling. The methodology proceeds through several key stages:

1. Spatiotemporal Derivative Processes: The framework defines a parent Gaussian process (GP), $Z(s, t)$, and induces higher-order derivative processes using tensor product operators ($\nabla^{\otimes r}$). The authors establish the distribution theory for these processes, deriving the joint distribution of the parent process and its derivatives (up to fourth order) to ensure valid probabilistic inference. This relies on specific covariance kernels (e.g., non-separable Matérn kernels) that guarantee the existence of required spectral moments.
2. Wombling Measures on Surfaces: The authors define "wombling" surfaces as 2-dimensional compact Riemannian manifolds in $\mathbb{R}^3$ (space $\times$ time) that enclose regions of rapid change. They formalize wombling measures as integrals of directional differential processes over these surfaces. Specifically, they define total and average wombling measures ($\Gamma(C)$ and $\bar{\Gamma}(C)$) based on the normal derivatives of the process across the surface boundary.
3. Triangulation and Computation: To avoid computationally expensive high-dimensional quadrature required for integrating over arbitrary surfaces, the authors propose a triangulation approach. The spatiotemporal surface is approximated by a union of parametric triangular planes. This discretization allows the wombling measures to be computed as sums over triangles.
4. Bayesian Inference: The framework employs a hierarchical spatiotemporal model where observations are modeled as a linear trend plus a GP and white noise. Posterior inference is conducted via Markov Chain Monte Carlo (MCMC) methods (Gibbs and Metropolis-Hastings). Crucially, the authors derive closed-form expressions or reduce the dimensionality of integrals (from 4D to 2D or 1D) for the cross-covariance matrices between observations and the wombling measures, enabling efficient posterior sampling of the wombling statistics.

Key Contributions

• Formalization of Spatiotemporal Wombling: The paper extends the concept of wombling from static spatial curves to evolving spatiotemporal surfaces, providing a rigorous mathematical definition for boundaries that change over time.
• Higher-Order Inference: Unlike previous works focusing primarily on first-order gradients, this framework supports inference on higher-order rates of change, including spatiotemporal curvature and mixed derivatives.
• Computational Efficiency: By utilizing surface triangulation and deriving specific algebraic simplifications for the cross-covariance of wombling measures (particularly under Gaussian and Matérn kernels), the authors avoid the prohibitive cost of high-dimensional numerical integration.
• Uncertainty Quantification: The framework provides full posterior distributions for wombling measures, offering uncertainty quantification for the location and magnitude of rapid change zones, a feature often missing in machine learning or level-set approaches.

Results
The methodology is validated through comprehensive simulation experiments and real-world applications:

• Simulations: Using synthetic data with known true derivatives, the authors demonstrate that the framework accurately recovers spatiotemporal differential processes and wombling measures. Coverage probabilities for 95% highest posterior density (HPD) intervals are assessed, showing the method's reliability in estimating local geometry and surface boundaries.
• Neuroimaging Application: The framework is applied to EEG data from alcoholic and control subjects to detect predisposition to alcoholism. The analysis reveals significant differences in neural activity dynamics between the two groups. Specifically, the control group exhibited significant spatiotemporal changes in the occipital and temporal regions, whereas the alcoholic group showed reduced activity and different temporal rates of change. The study highlights that purely spatial wombling fails to capture the extent of variability present in the temporal dimension, underscoring the necessity of the spatiotemporal approach.
• Additional Applications: The paper notes successful application to environmental data, including precipitation analysis in Northern California and PM2.5 levels during Canadian wildfires (detailed in the supplementary materials), demonstrating the method's versatility across domains.

Significance and Claims
The authors claim that this work represents a "landmark" in wombling literature by formalizing the concept of wombling surfaces within a spatiotemporal setting. The significance lies in the ability to:

1. Quantify Evolution: Move beyond static boundaries to analyze how zones of rapid change evolve over time.
2. Integrate Geometry and Statistics: Combine differential geometry (curvatures, fluxes) with rigorous Bayesian statistics to provide uncertainty-aware inference.
3. Scientific Discovery: Offer a tool to discern subtle differences in complex systems (such as neural predispositions or environmental hazards) that might be missed by methods ignoring temporal dynamics or lacking uncertainty quantification.

The paper concludes that while the current framework is robust, future work could extend these methods to non-Euclidean domains (manifolds) for modern neuroimaging and integrate wombling into predictive stacking frameworks. The authors emphasize that their approach provides a principled, model-based alternative to heuristic methods for detecting and characterizing dynamic boundaries in scientific data.