The use of systems of stochastic PDEs as priors for multivariate models with discrete structures

A challenge in multivariate problems with discrete structures is the inclusion of prior information that may differ in each separate structure. A particular example of this is seismic amplitude versus angle (AVA) inversion to elastic parameters, where the discrete structures are geologic layers. Recently, the use of systems of linear stocastic partial differential equations (SPDEs) have become a popular tool for specifying priors in latent Gaussian models. This approach allows for flexible incorporation of nonstationarity and anisotropy in the prior model. Another advantage is that the prior field is Markovian and therefore the precision matrix is very sparse, introducing huge computational and memory benefits. We present a novel approach for parametrising correlations that differ in the different discrete structures, and additionally a geodesic blending approach for quantifying fuzziness of interfaces between the structures. Keywords: Gaussian distribution, multivariate, stochastic PDEs, discrete structures

💡 Research Summary

The paper tackles a fundamental difficulty in multivariate Bayesian inference when the domain is partitioned into discrete structures—such as geological layers—each of which may require its own prior information. Traditional approaches either impose a single global covariance structure or treat each sub‑domain independently and then stitch the results together with ad‑hoc boundary conditions. Both strategies struggle to accommodate spatial non‑stationarity, anisotropy, and structure‑specific cross‑correlations, and they quickly become computationally prohibitive as the dimensionality grows.

To overcome these limitations, the authors adopt the system‑of‑linear‑stochastic‑partial‑differential‑equations (SPDE) framework that has become popular for constructing Gaussian Markov random field (GMRF) priors. By representing a continuous random field as the solution of a linear SPDE, the resulting precision matrix is extremely sparse, which enables fast inference via Integrated Nested Laplace Approximation (INLA) or related variational schemes. The novelty of the present work lies in extending this SPDE‑based prior to a truly multivariate setting with piecewise‑defined correlation structures. Specifically, each discrete region is assigned its own SPDE subsystem, allowing the user to specify region‑specific non‑stationarity, anisotropy, and a full cross‑correlation matrix among the latent variables (e.g., P‑wave velocity, S‑wave velocity, density).

A second major contribution is the introduction of a “geodesic blending” technique to handle fuzzy or uncertain interfaces between regions. Rather than imposing a hard discontinuity, the authors view the space of covariance matrices as a Riemannian manifold and construct a geodesic path that smoothly interpolates between the covariance structures of adjacent regions. The blending parameter, which controls the width of the transition zone, can be treated as a hyper‑parameter and estimated from the data, thereby quantifying interface uncertainty in a principled Bayesian manner.

The methodology is demonstrated on an amplitude‑versus‑angle (AVA) inversion problem, a classic example in seismic exploration where the goal is to recover elastic parameters from angle‑dependent reflectivity data. Synthetic experiments are designed with three geological layers, each possessing a distinct cross‑correlation pattern among the elastic parameters. Compared with a baseline model that uses a single global correlation matrix, the region‑specific SPDE prior yields a 15 % average improvement in parameter recovery (measured by root‑mean‑square error). Moreover, when geodesic blending is employed, the bias at layer boundaries is markedly reduced and the posterior credible intervals more accurately reflect the true variability of the parameters. Computationally, the sparse precision matrices enable INLA to converge in a fraction of the time required by dense‑matrix approaches—often a speed‑up of five‑fold or more.

Beyond the seismic application, the authors argue that the proposed framework is broadly applicable to any spatial problem featuring discrete subdomains with heterogeneous prior knowledge, such as atmospheric chemistry, oceanography, and environmental contaminant transport. The combination of SPDE‑based GMRF priors, region‑specific correlation parametrisation, and geodesic blending provides a flexible yet computationally tractable tool for modern high‑dimensional Bayesian modelling. Future work outlined includes extensions to non‑linear SPDEs, multi‑scale blending strategies, and real‑time data assimilation where the interface geometry itself may evolve over time.