Covariance-Based Rational Approximations of Fractional SPDEs for Computationally Efficient Bayesian Inference

David Bolin, Alexandre B. Simas, Zhen XiongJ. Comput. Graph. Stat 2024 · DOI

TL;DR

This paper introduces a novel method for approximating fractional Stochastic Partial Differential Equations (SPDEs), enabling computationally efficient Bayesian inference for complex spatial models. By directly approximating the covariance operator using a finite element method and rational approximations, the authors achieve a numerically stable Gaussian Markov Random Field (GMRF) representation compatible with popular Bayesian inference tools like R-INLA.

What Problem Does This Paper Address?

Modeling large spatial datasets often involves complex Gaussian random fields. The standard SPDE approach provides efficient approximations but is limited to specific smoothness parameters. Extending this to arbitrary smoothness (fractional SPDEs) has been challenging for Bayesian inference, as existing methods don't yield GMRFs, which are crucial for efficient computation with tools like R-INLA. This gap hinders the application of flexible fractional SPDE models in many real-world scenarios.

Methods

The core of this work is a covariance-based rational approximation of fractional SPDEs. Instead of approximating the solution of the SPDE, the authors directly approximate its covariance operator, which is a fractional power of a differential operator. This is achieved by:

  1. Finite Element Method (FEM): Discretizing the spatial domain and representing the field using FEM basis functions.
  2. Rational Approximation: Approximating the fractional power of the differential operator using rational functions (ratios of polynomials). This is done by splitting the fractional power and applying a rational approximation to the resulting terms.
  3. GMRF Representation: The key innovation is that this approximation results in a representation that is a sum of independent GMRFs with sparse precision matrices. This structure is directly compatible with the Integrated Nested Laplace Approximation (INLA) method.

The authors also provide a rigorous theoretical analysis of the convergence rate of their approximation.

Datasets

  • Precipitation Data: The paper uses a dataset of standardized annual precipitation anomalies from weather stations across the contiguous United States for the year 1962. This dataset was previously used in studies on covariance tapering methods.

Key Insights and Ideas

  • Direct Covariance Approximation is Key: Approximating the covariance operator directly, rather than the SPDE solution, is crucial for achieving a GMRF representation compatible with INLA.
  • Rational Approximation for Fractional Powers: Rational approximations are a powerful tool for handling fractional powers of differential operators, which arise in fractional SPDEs.
  • Numerical Stability: The proposed covariance-based method demonstrates better numerical stability compared to operator-based rational approximations, especially for higher orders.
  • Bridging the Gap to INLA: The GMRF representation is the critical link that allows flexible fractional SPDE models to be used within the efficient INLA framework, enabling estimation of all model parameters, including smoothness.

Key Results

  • Theoretical Convergence: The paper provides an explicit rate of convergence for the covariance approximation in the L2-norm.
  • Empirical Accuracy: Numerical experiments show that the covariance-based method achieves similar or better accuracy than existing methods, including the standard INLA approximation for stationary models.
  • Improved Predictive Performance: In an application to precipitation data, the fractional model (using the proposed method) outperformed an integer model, indicating the benefit of allowing non-integer smoothness.
  • Computational Efficiency: The GMRF representation allows for computationally efficient Bayesian inference, as demonstrated by the R package rSPDE and its integration with R-INLA.

Limitations and Future Work

  • Prior Specification for Smoothness: The current version of rSPDE offers truncated log-normal and beta priors for the smoothness parameter. Future work could involve developing more systematic ways to choose these priors, potentially using penalized complexity (PC) priors.
  • Spatio-Temporal Extensions: The authors suggest extending the proposed method to spatio-temporal SPDE models.
  • Numerical Stability for High Orders: While generally stable, the paper notes potential numerical instabilities in operator-based approximations for higher orders, which the covariance-based method mitigates.

Summary updated

BibTeX
@article{bolin2024covariance,
  author = {David Bolin and Alexandre B. Simas and Zhen Xiong},
  title = {{Covariance-Based Rational Approximations of Fractional SPDEs for Computationally Efficient Bayesian Inference}},
  journal = {J. Comput. Graph. Stat},
  year = {2024},
  doi = {10.1080/10618600.2023.2231051},
}
Figure