Nonparametric Bayesian Inference for Stochastic Reaction-Diffusion Equations

Randolf Altmeyer, Sascha GaudlitzarXiv 2025 · DOI · arXiv

Abstract

This paper establishes a rigorous theoretical framework for the Bayesian nonparametric estimation of nonlinear reaction functions in stochastic reaction-diffusion equations (SPDEs). By leveraging the infinite-dimensional Girsanov theorem, the authors demonstrate that the posterior distribution, based on a Gaussian wavelet prior, contracts around the true reaction function at the minimax optimal rate as the spatial domain grows. The analysis is further supported by a semiparametric Bernstein–von Mises theorem, providing a frequentist justification for Bayesian uncertainty quantification in this context.

Introduction

Stochastic Reaction-Diffusion equations are fundamental tools in the natural sciences, used to model complex phenomena ranging from protein dynamics within cellular environments to heat dissipation in turbulent media. These systems are typically described by the SPDE:

$$dX_t(y) = \Delta X_t(y) dt + f(X_t(y)) dt + dW_t(y)$$

where $\Delta$ represents the diffusion operator, $dW$ is space-time white noise, and $f$ is the reaction function. While the diffusion component is often dictated by known physical properties, the reaction function $f$—which describes local interactions—is frequently unknown. Traditional statistical approaches often rely on parametric models, assuming $f$ belongs to a specific family of functions. However, this paper addresses the more challenging nonparametric problem: recovering the entire functional form of $f$ from noisy, spatially distributed data without restrictive structural assumptions.

Methodology

The Bayesian Framework

The authors employ a nonparametric Bayesian approach using Gaussian wavelet priors. By representing the unknown function $f$ as a wavelet series expansion with Gaussian coefficients, the model's complexity can adaptively scale with the available data.

A critical advantage of this formulation is the tractability of the likelihood. Using the infinite-dimensional Girsanov theorem, the authors show that the log-likelihood for the SPDE is quadratic with respect to $f$. Because the prior is Gaussian and the likelihood is quadratic, the resulting posterior distribution is also Gaussian. This conjugacy is computationally significant, as it allows for direct calculation of the posterior mean and variance through linear algebra, bypassing the need for computationally intensive Markov Chain Monte Carlo (MCMC) simulations.

Theoretical Analysis

The study focuses on the "growing domain" asymptotic, where the system is observed over an increasingly large spatial domain ($\lambda \to \infty$). To prove the convergence of the posterior, the authors utilize sophisticated tools from stochastic analysis, including:

  • Malliavin Calculus: Specifically, the Clark-Ocone formula is used to derive sub-Gaussian concentration bounds for spatio-temporal averages.
  • Spatial Ergodicity: The proofs demonstrate that a single snapshot of a sufficiently large system provides enough information to identify the underlying reaction physics, a property analogous to time-ergodicity in classical stationary processes.

Key Results

Minimax Optimality and Contraction Rates

The paper proves that the posterior distribution contracts around the true reaction function at the rate of $\lambda^{-\beta/(2\beta+1)}$, where $\beta$ denotes the smoothness of the function. This rate is mathematically optimal (minimax) according to information theory, meaning no other estimator can converge faster.

Bernstein–von Mises Theorem

The authors establish a semiparametric Bernstein–von Mises (BvM) theorem. This result ensures that for linear functionals of the reaction function (such as its average value over a specific interval), the Bayesian credible intervals behave like frequentist confidence intervals. This provides a formal justification for using Bayesian uncertainty quantification in scientific applications.

Numerical Validation

The theory was validated using numerical simulations of a stochastic Allen-Cahn equation, featuring a cubic reaction function $f(x) = -(x^3 - 9x)$. Even with a single realization of the SPDE over a domain size of $\lambda = 50$ and a short time horizon ($T=1$), the method successfully recovered the cubic shape of the reaction term, demonstrating the practical utility of the "large domain" advantage.

Discussion and Limitations

Information and Occupation Time

A significant insight from the research is that posterior uncertainty is inversely proportional to the "occupation time"—the duration the system spends in a particular state. If the system rarely reaches certain concentrations, the model correctly identifies higher uncertainty for the reaction function at those values, providing a natural mechanism for reliability assessment.

Future Directions

While the paper provides a foundational theory for 1D domains, several challenges remain:

  1. Higher Dimensions: Extending the results to 2D or 3D is non-trivial. In higher dimensions, space-time white noise becomes significantly "rougher," requiring the application of advanced mathematical frameworks like regularity structures.
  2. Discrete Observations: The current theory assumes continuous observation in space and time. Future work must account for the discretization error inherent in real-world sensor data.
  3. Unknown Diffusion: The model currently assumes the diffusion operator is known. Developing a framework to simultaneously estimate both diffusion and reaction terms remains an open problem.

Conclusion

This work bridges the gap between high-level stochastic analysis and practical statistical inference. By proving that Bayesian methods can optimally recover nonlinear laws in SPDEs, the authors provide a robust mathematical basis for data-driven discovery in physics and biology.

Summary updated

Abstract

We consider the Bayesian nonparametric estimation of a nonlinear reaction function in a reaction-diffusion stochastic partial differential equation (SPDE). The likelihood is well-defined and tractable by the infinite-dimensional Girsanov theorem, and the posterior distribution is analysed in the growing domain asymptotic. Based on a Gaussian wavelet prior, the contraction of the posterior distribution around the truth at the minimax optimal rate is proved. The analysis of the posterior distribution is complemented by a semiparametric Bernstein--von Mises theorem. The proofs rely on the sub-Gaussian concentration of spatio-temporal averages of transformations of the SPDE, which is derived by combining the Clark-Ocone formula with bounds for the derivatives of the (marginal) densities of the SPDE.
BibTeX
@misc{altmeyer2025nonparametric,
  author = {Randolf Altmeyer and Sascha Gaudlitz},
  title = {{Nonparametric Bayesian Inference for Stochastic Reaction-Diffusion Equations}},
  year = {2025},
  eprint = {2507.06857},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.06857},
  howpublished = {arXiv},
}

Loading related content...

Figure