pith. sign in

arxiv: 2407.14781 · v3 · submitted 2024-07-20 · 🧮 math.ST · cs.NA· math.AP· math.NA· math.PR· stat.TH

Bernstein-von Mises theorems for time evolution equations

Richard Nickl This is my paper

Pith reviewed 2026-05-23 22:38 UTC · model grok-4.3

classification 🧮 math.ST cs.NAmath.APmath.NAmath.PRstat.TH
keywords Bernstein-von Mises theoreminfinite-dimensional posteriornonlinear parabolic PDEGaussian process priorWasserstein distancesupremum normreaction-diffusion equationtime-dependent Schrödinger equation
0
0 comments X

The pith

Non-Gaussian posteriors on PDE trajectories are approximated by Gaussians in Wasserstein distance for the supremum norm.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that for infinite-dimensional dynamical systems governed by nonlinear parabolic PDEs, with a Gaussian process prior on the initial condition and discrete space-time observations, the resulting posterior measures on trajectory function spaces converge in Wasserstein distance to a Gaussian random function when distances are measured in the supremum norm. This holds under a general set of conditions on the system, prior, and data. For the specific case of periodic nonlinear reaction-diffusion equations with smooth compactly supported reaction functions, the limiting Gaussian is the solution of a time-dependent Schrödinger equation driven by rough Gaussian initial conditions, with an explicit covariance operator. A reader would care because the result supplies an asymptotic normality statement that justifies Gaussian-based inference procedures for Bayesian analysis of time-evolving systems in function space.

Core claim

We give a general set of conditions under which these non-Gaussian posterior distributions are approximated, in Wasserstein distance for the supremum-norm metric, by the law of a Gaussian random function. We demonstrate the applicability of our results to periodic non-linear reaction diffusion equations where the limiting Gaussian measure can be characterised as the solution of a time-dependent Schrödinger equation with rough Gaussian initial conditions whose covariance operator we describe.

What carries the argument

Wasserstein distance approximation of the posterior by a Gaussian measure on the space of trajectories, equipped with the supremum norm.

If this is right

  • Bayesian credible sets for trajectories can be constructed from the limiting Gaussian under the stated conditions.
  • The result applies directly to periodic nonlinear reaction-diffusion equations with any smooth compactly supported reaction term.
  • The limiting Gaussian covariance operator is explicitly describable from the initial Gaussian process prior.
  • The approximation holds in the supremum norm on the trajectory space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same limiting object may supply asymptotic frequentist coverage guarantees for posterior credible bands in this PDE setting.
  • The Schrödinger characterization could be used to simulate approximate posterior samples without running full MCMC on the nonlinear system.
  • The general conditions might be verifiable for other parabolic equations beyond reaction-diffusion, such as those with different nonlinearities.

Load-bearing premise

The unspecified general conditions on the dynamical system, the Gaussian process prior, and the discrete observations must hold to guarantee the infinite-dimensional approximation.

What would settle it

A concrete numerical example for a specific smooth compactly supported reaction function in which the Wasserstein distance between the posterior and the described Schrödinger-driven Gaussian fails to approach zero as the number of space-time samples grows.

read the original abstract

We consider a class of infinite-dimensional dynamical systems driven by non-linear parabolic partial differential equations with initial condition $\theta$ modelled by a Gaussian process `prior' probability measure. Given discrete samples of the state of the system evolving in space-time, one obtains updated `posterior' measures on a function space containing all possible trajectories. We give a general set of conditions under which these non-Gaussian posterior distributions are approximated, in Wasserstein distance for the supremum-norm metric, by the law of a Gaussian random function. We demonstrate the applicability of our results to periodic non-linear reaction diffusion equations \begin{align*} \frac{\partial}{\partial t} u - \Delta u &= f(u) \\ u(0) &= \theta \end{align*} where $f$ is any smooth and compactly supported reaction function. In this case the limiting Gaussian measure can be characterised as the solution of a time-dependent Schr\"odinger equation with `rough' Gaussian initial conditions whose covariance operator we describe.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper establishes Bernstein-von Mises theorems for posteriors on trajectories of infinite-dimensional dynamical systems governed by nonlinear parabolic PDEs. Under a general set of conditions on the system, prior, and observations, non-Gaussian posteriors are approximated in Wasserstein distance with respect to the supremum-norm metric by the law of a Gaussian random function. For periodic nonlinear reaction-diffusion equations with smooth compactly supported reaction term f, the limiting Gaussian is characterized explicitly as the solution of a time-dependent Schrödinger equation driven by rough Gaussian initial conditions, whose covariance operator is described.

Significance. If the results hold with the stated topology, the work provides a rigorous nonparametric Bayesian justification for uncertainty quantification on space-time trajectories of nonlinear PDEs. The explicit Schrödinger-equation characterization of the limiting law in the reaction-diffusion case is a concrete strength, as it supplies an independent PDE-theoretic description of the Gaussian limit and may support future computational approximations. The choice of Wasserstein distance in the sup-norm metric is technically demanding and, if verified, yields a strong uniform approximation result.

major comments (2)
  1. [Application to periodic nonlinear reaction-diffusion equations] The central claim requires approximation in Wasserstein distance w.r.t. the supremum-norm metric on trajectories. The abstract states that the limiting Gaussian has 'rough' Gaussian initial conditions. Parabolic smoothing applies only for t>0; at t=0 the initial condition remains rough. The manuscript must therefore verify that the covariance operator of the initial condition satisfies Kolmogorov-Chentsov-type conditions ensuring almost-sure continuous sample paths on the closed space-time domain [0,T]×𝕋^d, so that the push-forward measure under the solution map is supported on C([0,T]×𝕋^d) with finite second moments. Without this verification the Wasserstein distance is either undefined or taken with respect to an incomplete metric space.
  2. [Statement of the general Bernstein-von Mises theorem] The general set of conditions under which the BvM holds is described as load-bearing for the result, yet the precise regularity, identifiability, and moment conditions on the prior, the observation operator, and the nonlinear map are not visible in sufficient detail to assess whether they guarantee the required tightness and local asymptotic normality in the infinite-dimensional setting. These conditions must be stated explicitly with references to the relevant function spaces and norms.
minor comments (1)
  1. The abstract uses the phrase 'rough' Gaussian initial conditions without a precise definition of the Sobolev or Hölder regularity; this should be clarified when the covariance operator is introduced.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below and will revise the manuscript to incorporate the requested clarifications and verifications.

read point-by-point responses

  1. Referee: [Application to periodic nonlinear reaction-diffusion equations] The central claim requires approximation in Wasserstein distance w.r.t. the supremum-norm metric on trajectories. The abstract states that the limiting Gaussian has 'rough' Gaussian initial conditions. Parabolic smoothing applies only for t>0; at t=0 the initial condition remains rough. The manuscript must therefore verify that the covariance operator of the initial condition satisfies Kolmogorov-Chentsov-type conditions ensuring almost-sure continuous sample paths on the closed space-time domain [0,T]×𝕋^d, so that the push-forward measure under the solution map is supported on C([0,T]×𝕋^d) with finite second moments. Without this verification the Wasserstein distance is either undefined or taken with respect to an incomplete metric space.

    Authors: We agree that an explicit verification of sample-path regularity is required to ensure the Wasserstein distance is well-defined on C([0,T]×𝕋^d). The general theorem is formulated on the space of continuous trajectories with the sup-norm, and the stated conditions on the prior and nonlinear map are intended to guarantee finite second moments and support on this space. For the reaction-diffusion application we will add, in the revised manuscript, a direct check that the covariance operator of the initial Gaussian process satisfies Kolmogorov-Chentsov criteria yielding a.s. Hölder-continuous paths on the compact space-time domain (including at t=0), thereby confirming that the push-forward measure lies in the desired function space. revision: yes

  2. Referee: [Statement of the general Bernstein-von Mises theorem] The general set of conditions under which the BvM holds is described as load-bearing for the result, yet the precise regularity, identifiability, and moment conditions on the prior, the observation operator, and the nonlinear map are not visible in sufficient detail to assess whether they guarantee the required tightness and local asymptotic normality in the infinite-dimensional setting. These conditions must be stated explicitly with references to the relevant function spaces and norms.

    Authors: We acknowledge that the general conditions should be stated with greater explicitness and with direct references to the underlying function spaces. In the revised manuscript we will list the precise assumptions on the prior (Gaussian process regularity in appropriate Sobolev or Besov spaces), the observation operator, and the nonlinear map, including the required regularity, identifiability, and moment conditions. These will be tied explicitly to the norms and spaces used in the proofs of tightness and local asymptotic normality. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation self-contained in PDE and measure theory

full rationale

The paper states a general theorem giving conditions for Wasserstein approximation of posteriors by a Gaussian law on trajectories equipped with the sup-norm, then specializes to periodic nonlinear reaction-diffusion equations where the limit is identified as the solution operator applied to a rough Gaussian initial condition whose covariance is described. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear; the limiting object is obtained from the linearized PDE (Schrödinger) and the prior covariance, both external to the target BvM statement. The result is therefore not equivalent to its inputs by construction and rests on independent analytic and probabilistic arguments.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper relies on standard assumptions from Bayesian nonparametrics and PDE theory; no new free parameters or invented entities are introduced in the abstract.

axioms (3)
  • domain assumption The initial condition is modeled by a Gaussian process prior probability measure.
    Stated in the abstract as the modeling choice for θ.
  • domain assumption Discrete samples of the state of the system evolving in space-time are given.
    The observation model assumed for obtaining the posterior.
  • domain assumption f is any smooth and compactly supported reaction function for the reaction-diffusion equation.
    Specific condition for the applicability demonstration.

pith-pipeline@v0.9.0 · 5704 in / 1419 out tokens · 24164 ms · 2026-05-23T22:38:12.795420+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Data assimilation with the 2D Navier-Stokes equations: Optimal Gaussian asymptotics for the posterior measure

    math.ST 2025-07 unverdicted novelty 7.0

    Proves functional Bernstein-von Mises theorem establishing Gaussian asymptotics in supremum norm for posteriors arising from Gaussian-process priors on initial data in 2D Navier-Stokes data assimilation.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bandeira, Antoine Maillard, Richard Nickl, and Sven Wang

    Afonso S. Bandeira, Antoine Maillard, Richard Nickl, and Sven Wang . On free energy barriers in Gaussian priors and failure of cold start MCMC for high- dimensional unimodal distributions. Philos. Trans. Roy. Soc. A , 381(2247), 2023

    work page 2023

  2. [2]

    On the computation al complexity of MCMC-based estimators in large samples

    Alexandre Belloni and Victor Chernozhukov. On the computation al complexity of MCMC-based estimators in large samples. Ann. Statist. , 37(4):2011–2055, 2009

    work page 2011

  3. [3]

    Oates, Mark Girolami, Michael A

    Fran¸ cois-Xavier Briol, Chris J. Oates, Mark Girolami, Michael A. O sborne, and Dino Sejdinovic. Probabilistic integration: a role in statistical comput ation? Statist. Sci. , 34(1):1–22, 2019

    work page 2019

  4. [4]

    Carrillo, F

    J.A. Carrillo, F. Hoffmann, A.M. Stuart, and U. Vaes. Statistical a ccuracy of approximate filtering methods. arXiv, 2024

    work page 2024

  5. [5]

    Nonparametric Bernstein–von Mises Theorems in Gaussian white noise

    Isma¨ el Castillo and Richard Nickl. Nonparametric Bernstein–von Mises Theorems in Gaussian white noise. Ann. Statist. , 41(4):1999–2028, 2013

    work page 1999

  6. [6]

    On the Bernstein–von Mises phe nomenon for nonparametric Bayes procedures

    Isma¨ el Castillo and Richard Nickl. On the Bernstein–von Mises phe nomenon for nonparametric Bayes procedures. Ann. Statist. , 42(5):1941–1969, 2014

    work page 1941

  7. [7]

    A Bernstein–von Mises the orem for smooth functionals in semiparametric models

    Isma¨ el Castillo and Judith Rousseau. A Bernstein–von Mises the orem for smooth functionals in semiparametric models. Ann. Statist. , 43(6):2353–2383, 2015

    work page 2015

  8. [8]

    Multiscale Bayesian survival analysis

    Isma¨ el Castillo and St´ ephanie van der Pas. Multiscale Bayesian survival analysis. Ann. Statist. , 49(6):3559–3582, 2021. 50

    work page 2021

  9. [9]

    Navier-Stokes equations

    Peter Constantin and Ciprian Foias. Navier-Stokes equations . Chicago Lectures in Mathematics. University of Chicago Press, Chicago, IL, 1988

    work page 1988

  10. [10]

    S. L. Cotter, M. Dashti, J. C. Robinson, and A. M. Stuart. Bay esian inverse problems for functions and applications to fluid mechanics. Inverse Problems , 25(11):115008, 43, 2009

    work page 2009

  11. [11]

    S. L. Cotter, G. O. Roberts, A. M. Stuart, and D. White. MCMC methods for functions: modifying old algorithms to make them faster. Statist. Sci. , 28(3):424– 446, 2013

    work page 2013

  12. [12]

    E. B. Davies. Spectral theory and differential operators , volume 42 of Cambridge Studies in Advanced Mathematics . Cambridge University Press, Cambridge, 1995

    work page 1995

  13. [13]

    Bayesian numerical analysis

    Persi Diaconis. Bayesian numerical analysis. In Statistical decision theory and related topics, IV, Vol. 1 , pages 163–175. Springer, New York, 1988

    work page 1988

  14. [14]

    Richard M. Dudley. Real analysis and probability . Cambridge University Press, Cambridge, 2002

    work page 2002

  15. [15]

    Lawrence C. Evans. Partial differential equations , volume 19. American Mathe- matical Society, Providence, RI, second edition, 2010

    work page 2010

  16. [16]

    Evensen, F.C

    G. Evensen, F.C. Vossepoel, and J. van Leeuwen. Data assimilation fundamen- tals. Springer textbooks in Earth sciences, Geography and Environme nt. Springer, Cham, 2022

    work page 2022

  17. [17]

    Gerald B. Folland. Real analysis. John Wiley & Sons, Inc., New York, 1999

    work page 1999

  18. [18]

    On the Bernstein-von Mises theorem with infinit e-dimensional parameters

    David Freedman. On the Bernstein-von Mises theorem with infinit e-dimensional parameters. Ann. Statist. , 27(4):1119–1140, 1999

    work page 1999

  19. [19]

    van der Vaart

    Subhashis Ghosal and Aad W. van der Vaart. Fundamentals of Nonparametric Bayesian Inference. Cambridge University Press, New York, 2017

    work page 2017

  20. [20]

    Mathematical foundations of infinite-dimensional statistical models

    Evarist Gin´ e and Richard Nickl. Mathematical foundations of infinite-dimensional statistical models. Cambridge University Press, New York, 2016

    work page 2016

  21. [21]

    Statistical algorithms for low- frequency diffu- sion data: A PDE approach

    Matteo Giordano and Sven Wang. Statistical algorithms for low- frequency diffu- sion data: A PDE approach. arXiv, 2024

    work page 2024

  22. [22]

    Stuart, and Sebastian J

    Martin Hairer, Andrew M. Stuart, and Sebastian J. Vollmer. Spe ctral gaps for a Metropolis-Hastings algorithm in infinite dimensions. Ann. Appl. Probab. , 24(6):2455–2490, 2014

    work page 2014

  23. [23]

    Atmospheric modelling, data assimilation, and predictabi lity

    Eugenia Kalnay. Atmospheric modelling, data assimilation, and predictabi lity. Cambridge University Press, Cambridge, UK, 2003

    work page 2003

  24. [24]

    de Laplace

    Pierre-Simon M. de Laplace. Theorie analytiques des probabilit´ es. Courcier, Paris, 1812

    work page

  25. [25]

    Data assimilation , vol- ume 62 of Texts in Applied Mathematics

    Kody Law, Andrew Stuart, and Konstantinos Zygalakis. Data assimilation , vol- ume 62 of Texts in Applied Mathematics . Springer, Cham, 2015. A mathematical introduction

    work page 2015

  26. [26]

    Majda and John Harlim

    Andrew J. Majda and John Harlim. Filtering complex turbulent systems . Cam- bridge University Press, Cambridge, 2012

    work page 2012

  27. [27]

    Approximate inertial manifolds for reaction-diff usion equations in high space dimension

    Martine Marion. Approximate inertial manifolds for reaction-diff usion equations in high space dimension. J. Dynam. Differential Equations , 1(3):245–267, 1989

    work page 1989

  28. [28]

    Paternain

    Fran¸ cois Monard, Richard Nickl, and Gabriel P. Paternain. Consistent inversion of noisy non-Abelian X-ray transforms. Comm. Pure Appl. Math. , 74(5):1045–1099, 51 2021

    work page 2021

  29. [29]

    Paternain

    Fran¸ cois Monard, Richard Nickl, and Gabriel P. Paternain. Sta tistical guaran- tees for Bayesian uncertainty quantification in nonlinear inverse pr oblems with Gaussian process priors. Ann. Statist. , 49(6):3255–3298, 2021

    work page 2021

  30. [30]

    Bernstein–von Mises theorems for statistical inv erse problems I: Schr¨ odinger equation.J

    Richard Nickl. Bernstein–von Mises theorems for statistical inv erse problems I: Schr¨ odinger equation.J. Eur. Math. Soc. (JEMS) , 22(8):2697–2750, 2020

    work page 2020

  31. [31]

    Bayesian non-linear statistical inverse problems

    Richard Nickl. Bayesian non-linear statistical inverse problems . Zurich Lectures in Advanced Mathematics. European Mathematical Society (EMS) p ress, Berlin, 2023

    work page 2023

  32. [32]

    Consistent inference for diffusions from low frequ ency measure- ments

    Richard Nickl. Consistent inference for diffusions from low frequ ency measure- ments. Annals of Statistics , 52:519–549, 2024

    work page 2024

  33. [33]

    Paternain

    Richard Nickl and Gabriel P. Paternain. On some information-th eoretic aspects of non-linear statistical inverse problems. In ICM—International Congress of Math- ematicians. Vol. 7. Sections 15–20 , pages 5516–5538. EMS Press, Berlin, [2023]

    work page 2023

  34. [34]

    Pavliotis, and Kolyan Ray

    Richard Nickl, Grigorios A. Pavliotis, and Kolyan Ray. Bayesian non parametric inference in McKean-Vlasov models. arXiv, 2024

    work page 2024

  35. [35]

    Bernstein-von Mises theorems for statistical inverse problems II: compound Poisson processes

    Richard Nickl and Jakob S¨ ohl. Bernstein-von Mises theorems for statistical inverse problems II: compound Poisson processes. Electron. J. Stat. , 13(2):3513–3571, 2019

    work page 2019

  36. [36]

    Richard Nickl and Edriss S. Titi. On posterior consistency of dat a assimilation with Gaussian process priors: the 2D Navier-Stokes equations. Annals of Statistics, to appear, 2023

    work page 2023

  37. [37]

    On polynomial-time computation of h igh- dimensional posterior measures by Langevin-type algorithms

    Richard Nickl and Sven Wang. On polynomial-time computation of h igh- dimensional posterior measures by Langevin-type algorithms. J. Eur. Math. Soc. (JEMS), 26:1031–1112, 2024

    work page 2024

  38. [38]

    Probabilistic forecasting and Bayesian data assimilation

    Sebastian Reich and Colin Cotter. Probabilistic forecasting and Bayesian data assimilation. Cambridge University Press, New York, 2015

    work page 2015

  39. [39]

    Robinson

    James C. Robinson. Infinite-dimensional dynamical systems . Cambridge Texts in Applied Mathematics. Cambridge University Press, Cambridge, 2001

    work page 2001

  40. [40]

    Andrew M. Stuart. Inverse problems: a Bayesian perspective . Acta Numer. , 19:451–559, 2010

    work page 2010

  41. [41]

    Sznitman

    A.-S. Sznitman. Topics in propagation of chaos. In ´Ecole d’ ´Et´ e de Probabilit´ es de Saint-Flour XIX—1989 , volume 1464 of Lecture Notes in Math. , pages 165–251. Springer, Berlin, 1991

    work page 1989

  42. [42]

    Michael E. Taylor. Partial differential equations I. Basic theory. Springer, New York, 2011

    work page 2011

  43. [43]

    Infinite-dimensional dynamical systems in mechanics and ph ysics, volume 68 of Applied Mathematical Sciences

    Roger Temam. Infinite-dimensional dynamical systems in mechanics and ph ysics, volume 68 of Applied Mathematical Sciences . Springer-Verlag, New York, second edition, 1997

    work page 1997

  44. [44]

    A. W. van der Vaart. Asymptotic statistics . Cambridge Univ. Press, Cambridge, 1998

    work page 1998

  45. [45]

    A. W. van der Vaart and J. H. van Zanten. Rates of contractio n of posterior distributions based on Gaussian process priors. Ann. Statist. , 36(3):1435–1463, 2008. 52

    work page 2008

  46. [46]

    Springer-Verlag, Berlin, 2009

    C´ edric Villani.Optimal transport, volume 338 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 2009. Department of Pure Mathematics & Mathematical Statistics University of Cambridge , Cambridge, UK Email: nickl@maths.cam.ac.uk 53

    work page 2009