A new class of non-stationary Gaussian fields with general smoothness on metric graphs

The increasing availability of network data has driven the development of advanced statistical models specifically designed for metric graphs, where Gaussian processes play a pivotal role. While models such as Whittle-Matérn fields have been introduced, there remains a lack of practically applicable options that accommodate flexible non-stationary covariance structures or general smoothness. To address this gap, we propose a novel class of generalized Whittle-Matérn fields, which are rigorously defined on general compact metric graphs and permit both non-stationarity and arbitrary smoothness. We establish new regularity results for these fields, which extend even to the standard Whittle-Matérn case. Furthermore, we introduce a method to approximate the covariance operator of these processes by combining the finite element method with a rational approximation of the operator’s fractional power, enabling computationally efficient Bayesian inference for large datasets. Theoretical guarantees are provided by deriving explicit convergence rates for the covariance approximation error, and the practical utility of our approach is demonstrated through simulation studies and an application to traffic speed data, highlighting the flexibility and effectiveness of the proposed model class.

💡 Research Summary

The paper introduces a novel class of non‑stationary Gaussian random fields defined on compact metric graphs, extending the Whittle‑Matérn framework to allow spatially varying variance and correlation range functions (κ(s) and τ(s)) as well as arbitrary smoothness controlled by a fractional order α > ½. By formulating the field as the solution of the stochastic partial differential equation (κ² − Δ_Γ)^{α/2} τ u = W, where Δ_Γ is the Kirchhoff Laplacian on the graph, the authors obtain a flexible SPDE representation that can capture heterogeneous dependence structures and differentiable sample paths.

Theoretical contributions include: (1) precise regularity and positivity conditions (Assumptions 1 and 2) guaranteeing existence, uniqueness, and strict positive‑definiteness of the covariance function; (2) a detailed analysis of local Hölder regularity on each edge (u|_e ∈ C^{0,γ}(e) for any γ < min{α‑½, 1}) and global regularity on the whole graph, which crucially depends on the Hölder continuity of τ; and (3) a demonstration that internal vertices behave like boundary points, making τ’s global smoothness essential for overall field regularity—an effect absent in Euclidean domains.

From a computational standpoint, the authors propose a two‑stage approximation of the covariance operator. First, a finite‑element discretisation of the operator L = κ² − Δ_Γ yields a sparse matrix L_h that respects Kirchhoff vertex conditions. Second, the fractional inverse L^{‑α/2} is approximated by a rational function Σ_{k=1}^K w_k (L_h + ζ_k I)^{‑1}, where the weights w_k and shifts ζ_k are obtained via Gauss‑Legendre quadrature. This rational approximation provides a controllable error that decays algebraically with the number of terms K, while the FEM discretisation error scales with the mesh size h. The authors prove an explicit convergence bound ‖C − C_{h,K}‖_{L²→L²} ≤ C₁ h^{2p} + C₂ K^{‑β}, linking the overall error to the FEM order p and rational‑approximation order β.

These approximations are directly compatible with the R‑INLA framework, enabling fast Bayesian inference for latent Gaussian models on large graphs. The paper supplies an R package (MetricGraph) that implements the FEM assembly, rational approximation, and INLA integration, allowing practitioners to fit the proposed models to data with thousands of observations.

Empirical validation consists of two parts. In synthetic experiments, the authors vary α, κ, and τ to assess parameter recovery and predictive performance. Results show that accurate estimation of α is critical: misspecifying smoothness leads to substantial degradation in mean‑square prediction error, confirming the theoretical importance of the smoothness parameter. In a real‑world application to traffic speed measurements on a road network containing multiple edges and loops, the non‑stationary Whittle‑Matérn model outperforms a stationary Matérn baseline, reducing RMSE by more than 15 % and providing credible intervals that adapt to local variability.

In summary, the paper delivers (i) a rigorous mathematical foundation for non‑stationary Gaussian fields with arbitrary smoothness on metric graphs, (ii) novel regularity results that highlight the interplay between vertex conditions and the spatially varying τ, (iii) an efficient FEM‑plus‑rational‑approximation scheme with provable error rates, and (iv) practical tools and demonstrations that showcase the model’s applicability to large‑scale network data. This work substantially advances the statistical methodology for graph‑structured spatial data, opening avenues for sophisticated modeling in transportation, hydrology, power grids, and other domains where data naturally reside on metric graphs.