Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bayesian Methods for the Navier-Stokes Equations

Nicholas Polson, Vadim Sokolov

TL;DR

This work reframes the numerical solution of the incompressible Navier-Stokes equations as Bayesian posterior inference on a discretized state-space, enabling quantified uncertainty due to discretization, forcing, and modeling error. In 2D, it leverages Feynman-Kac representations to motivate Monte Carlo solvers, while in 3D it relies on particle and ensemble methods, including mean-field and McKean-Vlasov concepts, with spectral discretization to scale to high resolution and optional data assimilation for filtering. The authors introduce robust, non-Gaussian error models via normal variance-mean mixtures, notably Normal Inverse Gaussian (NIG) augmentation, to maintain conditional Gaussian updates and online parameter learning through particle learning, reducing weight collapse. They provide detailed methodological recipes, practical comparisons to EnKF and 4DVar, and numerical demonstrations that emphasize calibrated uncertainty, online learning, and robustness to outliers, with guidance for turbulent flow control and numerical weather prediction contexts.

Abstract

We develop a Bayesian methodology for numerical solution of the incompressible Navier--Stokes equations with quantified uncertainty. The central idea is to treat discretized Navier--Stokes dynamics as a state-space model and to view numerical solution as posterior computation: priors encode physical structure and modeling error, and the solver outputs a distribution over states and quantities of interest rather than a single trajectory. In two dimensions, stochastic representations (Feynman--Kac and stochastic characteristics for linear advection--diffusion with prescribed drift) motivate Monte Carlo solvers and provide intuition for uncertainty propagation. In three dimensions, we formulate stochastic Navier--Stokes models and describe particle-based and ensemble-based Bayesian workflows for uncertainty propagation in spectral discretizations. A key computational advantage is that parameter learning can be performed stably via particle learning: marginalization and resample--propagate (one-step smoothing) constructions avoid the weight-collapse that plagues naive sequential importance sampling on static parameters. When partial observations are available, the same machinery supports sequential observational updating as an additional capability. We also discuss non-Gaussian (heavy-tailed) error models based on normal variance-mean mixtures, which yield conditionally Gaussian updates via latent scale augmentation.

Bayesian Methods for the Navier-Stokes Equations

TL;DR

This work reframes the numerical solution of the incompressible Navier-Stokes equations as Bayesian posterior inference on a discretized state-space, enabling quantified uncertainty due to discretization, forcing, and modeling error. In 2D, it leverages Feynman-Kac representations to motivate Monte Carlo solvers, while in 3D it relies on particle and ensemble methods, including mean-field and McKean-Vlasov concepts, with spectral discretization to scale to high resolution and optional data assimilation for filtering. The authors introduce robust, non-Gaussian error models via normal variance-mean mixtures, notably Normal Inverse Gaussian (NIG) augmentation, to maintain conditional Gaussian updates and online parameter learning through particle learning, reducing weight collapse. They provide detailed methodological recipes, practical comparisons to EnKF and 4DVar, and numerical demonstrations that emphasize calibrated uncertainty, online learning, and robustness to outliers, with guidance for turbulent flow control and numerical weather prediction contexts.

Abstract

We develop a Bayesian methodology for numerical solution of the incompressible Navier--Stokes equations with quantified uncertainty. The central idea is to treat discretized Navier--Stokes dynamics as a state-space model and to view numerical solution as posterior computation: priors encode physical structure and modeling error, and the solver outputs a distribution over states and quantities of interest rather than a single trajectory. In two dimensions, stochastic representations (Feynman--Kac and stochastic characteristics for linear advection--diffusion with prescribed drift) motivate Monte Carlo solvers and provide intuition for uncertainty propagation. In three dimensions, we formulate stochastic Navier--Stokes models and describe particle-based and ensemble-based Bayesian workflows for uncertainty propagation in spectral discretizations. A key computational advantage is that parameter learning can be performed stably via particle learning: marginalization and resample--propagate (one-step smoothing) constructions avoid the weight-collapse that plagues naive sequential importance sampling on static parameters. When partial observations are available, the same machinery supports sequential observational updating as an additional capability. We also discuss non-Gaussian (heavy-tailed) error models based on normal variance-mean mixtures, which yield conditionally Gaussian updates via latent scale augmentation.
Paper Structure (38 sections, 3 theorems, 54 equations, 4 figures)

This paper contains 38 sections, 3 theorems, 54 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Mathematical Foundations
  3. The Navier-Stokes Equations
  4. Functional Setting for 3D Navier-Stokes
  5. Stochastic Calculus
  6. Bayesian state-space inference (filtering when data are available)
  7. A 2D prototype: stochastic representations and Monte Carlo solvers
  8. Vorticity Formulation
  9. Yudovich's Uniqueness Theorem
  10. The Feynman-Kac Theorem
  11. Analytical Solutions
  12. Bayesian numerical solvers in three dimensions
  13. Motivation: vortex stretching and model uncertainty
  14. Why closed-form stochastic representations are limited in 3D
  15. The Constantin-Iyer Representation
  16. ...and 23 more sections

Key Result

Theorem 3.1

Let $\omega_0 \in L^1(\mathbb{R}^2) \cap L^\infty(\mathbb{R}^2)$. Then the 2D incompressible Euler equations admit a unique weak solution $\omega \in L^\infty([0,T]; L^1 \cap L^\infty(\mathbb{R}^2))$ for all $T > 0$.

Figures (4)

  • Figure 1: Toy parameter-learning example illustrating a key computational advantage of particle learning. Naive sequential importance sampling (SIS) on a static parameter suffers rapid weight collapse (ESS decays), while particle learning maintains ESS by combining sufficient statistics (marginalization) with a resample--propagate (one-step smoothing) update.
  • Figure 2: Particle/ensemble Bayesian numerical solution of 2D Navier--Stokes (vorticity form) on a torus with uncertain viscosity and forcing amplitude. No observations are used: uncertainty is propagated through the solver to produce distributions over the vorticity field and kinetic energy.
  • Figure 3: Gaussian Kalman filter vs scale-augmented (NIG-motivated) Rao--Blackwellized particle filter in a 1D linear system with outliers. Heavy-tailed robustness is achieved by sampling latent inverse-Gaussian scales that inflate observation variance during outlier events.
  • Figure 4: Viability comparison on a 1D toy problem with outliers: deterministic forecast-only baseline, sliding-window 4DVar (MAP point estimate), EnKF (Gaussian uncertainty), and an NIG-motivated scale-augmented RBPF. The robust Bayesian filter adapts to outliers via latent scales and yields improved uncertainty calibration.

Theorems & Definitions (11)

  • Definition 2.1: Leray-Hopf weak solution
  • Definition 2.2: Scale mixture of normals
  • Remark 2.3
  • Definition 2.4: Particle filter
  • Theorem 3.1: Yudovich, 1963
  • Remark 3.2
  • Theorem 3.3: Feynman-Kac
  • Remark 3.4: Connection to Black--Scholes (computational viewpoint)
  • Remark 4.1
  • Theorem 4.2: flandoli1994
  • ...and 1 more