Nonparametric Bayesian inference for reversible multi-dimensional diffusions

Matteo Giordano, Kolyan RayarXiv 2020 · DOI · OpenReview

Abstract

We study nonparametric Bayesian models for reversible multi-dimensional diffusions with periodic drift. For continuous observation paths, reversibility is exploited to prove a general posterior contraction rate theorem for the drift gradient vector field under approximation-theoretic conditions on the induced prior for the invariant measure. The general theorem is applied to Gaussian priors and $p$-exponential priors, which are shown to converge to the truth at the minimax optimal rate over Sobolev smoothness classes in any dimension.

Introduction

Diffusion processes, characterized by a combination of deterministic drift and stochastic noise, are fundamental in modeling complex systems in physics, chemistry, and biology. A critical subset of these is reversible diffusions, where the dynamics are time-invariant. In such systems, the drift vector field is the gradient of a scalar "potential" function $B$. While estimation in one dimension is well-understood, multi-dimensional cases ($d \ge 2$) present significant theoretical hurdles, as standard tools like local times do not generalize easily. This paper addresses the challenge of placing a prior directly on the potential $B$ and proves that the resulting posterior distribution concentrates around the true potential as observation time $T$ increases.

Potential-Modelling Approach

The authors propose a potential-modelling framework. Instead of placing a prior on the drift vector field itself, they model the scalar potential function $B$. This approach inherently enforces the physical property of reversibility, transforming a vector estimation problem into a simpler scalar one.

Illustrative example of a spatially heterogeneous potential (left) and its corresponding gradient vector field (right).

To handle the nonlinearity of the relationship between the potential and the observed path, the authors utilize rescaled priors, adjusting variance based on the time horizon $T$ and expected smoothness $s$. They evaluate two primary classes of priors:

  • Gaussian Process Priors: Ideal for smooth, regular potentials.
  • $p$-exponential Priors: Heavy-tailed priors (e.g., Laplace-type) that effectively capture "spiky" or spatially inhomogeneous potentials where the landscape changes abruptly.

Theoretical Insights

A major contribution of this work is linking Bayesian inference for diffusions to nonlinear inverse problems. By adapting tools from the Bayesian Inverse Problems literature, the authors establish that the posterior contracts at the rate $T^{-s/(2s+d)}$. This is the minimax optimal rate for estimating an $s$-smooth function in $d$ dimensions, demonstrating that the Bayesian approach achieves the highest possible theoretical efficiency.

Furthermore, the authors develop new plug-in tests based on the invariant measure $\mu_B \propto e^{2B}$. By employing elliptic Partial Differential Equation (PDE) theory, they establish a rigorous link between the potential and this measure, facilitating the proof of convergence.

Key Results

The core theoretical findings are summarized in Theorems 2.1 and 2.4, which prove that for both Gaussian and $p$-exponential priors, the posterior probability of the error exceeding a threshold $M T^{-s/(2s+d)}$ vanishes as $T \to \infty$. Additionally:

  • Posterior Mean Consistency: The posterior mean is shown to converge at the same optimal rate.
  • Computational Efficiency: For Gaussian priors in continuous observation models, the posterior remains Gaussian (conjugacy), allowing for efficient sampling and implementation.

Limitations and Future Directions

While providing a robust theoretical foundation, the study assumes the potential is periodic (defined on a torus) to simplify PDE estimates. Future work may extend these results to potentials on unbounded spaces ($\mathbb{R}^d$). Additionally, the current framework requires prior knowledge of the smoothness parameter $s$ for correct scaling; developing adaptive priors that learn smoothness from data remains an open challenge. Finally, the model assumes a constant and known diffusivity, suggesting another avenue for complexity in future research.

Summary updated

Abstract

We study nonparametric Bayesian models for reversible multi-dimensional diffusions with periodic drift. For continuous observation paths, reversibility is exploited to prove a general posterior contraction rate theorem for the drift gradient vector field under approximation-theoretic conditions on the induced prior for the invariant measure. The general theorem is applied to Gaussian priors and $p$-exponential priors, which are shown to converge to the truth at the minimax optimal rate over Sobolev smoothness classes in any dimension.
BibTeX
@misc{giordano2020nonparametric,
  author = {Matteo Giordano and Kolyan Ray},
  title = {{Nonparametric Bayesian inference for reversible multi-dimensional diffusions}},
  year = {2020},
  eprint = {2012.12083},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.12083},
}
Figure