Pith. sign in

REVIEW 2 major objections 6 minor 295 references

The paper claims that the mean of the discretization error in numerical ODE solutions can be inferred from noisy observations—including its direction, not just its magnitude—by treating that mean as a latent state and filtering it with the

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:33 UTC pith:LEEZGJWA

load-bearing objection Clean conditional theorems but the experiments don't satisfy the theorem's hypotheses; the mean-error idea is a real but incremental extension. the 2 major comments →

arxiv 2607.26552 v1 pith:LEEZGJWA submitted 2026-07-29 math.NA cs.NAstat.ME

Bayesian Inference of Discretization Error Means in ODEs via Ensemble Kalman Filtering

Shoji Toyota , Yuto Miyatake This is my paper
classification math.NA cs.NAstat.ME MSC 65L0565L7062F15
keywords discretization errorordinary differential equationsEnsemble Kalman FilterMarkov priorBayesian inferencestate-space modellocal errorglobal error
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a Bayesian framework that models the difference between an ODE's true and numerical solutions as a Gaussian random variable with an unknown mean, called the discretization error mean. It shows that a Markov prior built from the classical local-error/global-error recursion makes this mean inferable from observations through a linear-Gaussian state-space model, so the Ensemble Kalman Filter applies directly. The prior depends on the solver's step size, and the paper proves that under natural conditions the prior converges to zero in probability at rate O_p(h^{min(a,b)}), matching classical error order. Experiments on the pendulum, FitzHugh–Nagumo, and Lorenz-96 systems show the inferred means track the true error, including its sign, allowing numerical solutions to be corrected toward the exact solution.

Core claim

The central claim is that the mean of the discretization error—not just its variance—is identifiable from observational data under a Markov prior derived from the recursion G(t_{i,j+1}) ≈ (I + h∇f)G(t_{i,j}) + L(t_{i,j}). Because the observation equation can be rewritten as y_t = Hμ_t + ε_t with Gaussian noise, the latent error mean can be filtered with the Ensemble Kalman Filter. Theorem 3.1 establishes that, if the multiplier M in the prior satisfies E∥I−M∥_op ≤ C_M h and the solver has local order a, then the prior mean converges to zero in probability at rate O_p(h^{min(a,b)}). The paper argues this recovers both magnitude and direction of the error, which variance-only approaches cannot

What carries the argument

The key object is the Markov prior (15), μ_{i,j+1} = M_{i,j} μ_{i,j} + L̃(t_{i,j}), which encodes how global errors accumulate from local errors. The matrix M is drawn from a distribution P and is meant to approximate I + h∇f; in the experiments it is mI with m∼N(1,β²). Together with the observation model y_t = Hμ_t + ε_t, this prior turns the problem into an EnKF-ready linear-Gaussian state-space model. Theorem 3.1 provides the step-size asymptotics that connect the prior to classical order theory.

Load-bearing premise

The load-bearing premise behind the convergence-rate theorem (Theorem 3.1, Condition (I)) is that the random multiplier M in the prior satisfies E∥I−M∥_op ≤ C_M h, shrinking to the identity as the step h goes to zero; the Gaussian prior m∼N(1,β²) used in the Section 4 experiments has E|1−m| = β√(2/π), a constant independent of h, so Condition (I) is not met for fixed β>0 and the theorem does not directly cover the experimental prior.

What would settle it

Compute the exact discretization error for a system with a known reference solution (e.g., the pendulum with a fine reference solve), run the EnKF estimator with the paper's prior at h=0.05, and check whether the posterior mean follows the sign and magnitude of the true error. The central claim fails if the inferred mean does not track the true error trajectory beyond the observation-noise level. To test the theorem instead, replace the prior with one whose spread around the identity is O(h), and check whether the posterior error shrinks at rate h^{min(a,b)} as h→0.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can infer the actual direction of discretization error from noisy data and shift the numerical trajectory toward the true solution.
  • The linear-Gaussian formulation scales to higher-dimensional systems where particle-filter alternatives suffer degeneracy, as the Lorenz-96 experiment shows.
  • The prior's convergence rate O_p(h^{min(a,b)}) gives a principled step-size-aware prior for Bayesian ODE workflows.
  • The method extends variance-only discretization-error quantification by enabling mean correction, not just uncertainty inflation.
  • Hyperparameter selection via marginal likelihood tends to choose α=1, matching the condition E[M]≈I and giving a default calibration.

Where Pith is reading between the lines

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

  • A prior with variance scaling like h—so that Condition (I) of Theorem 3.1 holds exactly—would put the experimental setup under the theorem's hypotheses; one could test whether the fixed-β Gaussian prior is the reason observed rates might differ from O(h).
  • The same Markov-prior idea could be coupled with parameter estimation, jointly inferring numerical error and model parameters instead of assuming the solver is accurate.
  • The recursion behind the prior could be adapted to stochastic differential equation integrators or spatial discretizations of PDEs, wherever local errors accumulate through a linearized flow.
  • Jointly modeling mean and variance in a single state-space model, as the conclusion hints, may yield more honest credible intervals than mean-only correction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper proposes a Bayesian framework for quantifying the mean discretization error in numerical ODE solutions from noisy observations. It models x(t_i)-x_{t_i} ~ N(μ_i, Σ), places a Markov prior on μ derived from local-error accumulation, and performs inference with the Ensemble Kalman Filter (EnKF) with fixed-lag smoothing. The main theoretical results are Theorem 2.1 (the EnKF ensemble moments converge to the true Gaussian posterior moments in the large-ensemble limit) and Theorem 3.1 (the proposed Markov prior converges to zero at rate O_p(h^{min(a,b)}) under Conditions (I)-(IV)). Numerical experiments on the pendulum, FitzHugh-Nagumo, and Lorenz-96 systems illustrate error-direction recovery and predictive correction of the numerical solution.

Significance. If the results hold, the framework is a useful extension of the discretization-error-variance literature: modeling the mean rather than only the covariance provides directional information about discretization error, and the linear-Gaussian observation form allows EnKF rather than particle filtering, with potential high-dimensional advantages. The paper contains self-contained proofs of both theorems, and Theorem 3.1 gives a clean asymptotic justification of the prior in the idealized setting. The experimental evidence is suggestive but, as discussed below, is not currently tied to the theorem's hypotheses.

major comments (2)
  1. [Appendix E, Table 3] Condition (I) of Theorem 3.1 requires E_{M~P_λ} ||I-M||_op ≤ C_M h. In all experiments M = mI with m ~ N(α,β^2), so E||I-M||_op = E|1-m| = β√(2/π). For β = 0.3 this is ≈ 0.239, which is independent of h; no h-dependent scaling is introduced. Moreover p(μ_{t0}) = N(0,I) gives E||μ_{t0}|| constant, contradicting Condition (III) for any b > 0. Consequently Theorem 3.1 does not apply to the experimental configuration, and the statement in §4 that α=1 is 'consistent with the theoretical implication' conflates E[M]=I with the operator-norm deviation condition. The observed performance in Figures 2-6 cannot be attributed to the convergence theorem. Please either use h-dependent priors (e.g., β = O(h) and p_h(μ_0) with scale h^b) so that Conditions (I) and (III) hold, or expressly restrict the theoretical claims to the idealized setting.
  2. [Appendix E, Table 3] The EnKF-vs-PF comparison is not controlled: EnKF hyperparameters were selected by maximizing the marginal likelihood, while PF was assigned fixed (α,β,γ)=(1,1,1) because the computed log-marginal likelihood overflowed. Thus Table 3 and Figure 6 may reflect hyperparameter mismatch rather than a property of the filter. To support the claimed high-dimensional advantage, either tune PF by the same or an equivalent criterion, or report both filters with identical hyperparameters.
minor comments (6)
  1. [§3.3] Condition (II) is stated as L(t_i,j) ≤ C_L h^{a+1} without a norm; the proof uses the norm, so add ||·||.
  2. [§4] The 'ensemble-based estimate of the marginal likelihood' used for hyperparameter selection is not defined. Please provide the estimator or a reference, especially since hyperparameter choice drives the experiments.
  3. [Algorithm 1] The return line contains a typo ('aensembles') and the output set notation is confusing; please clarify which ensembles are returned.
  4. [§2.3.2, Eq. (9)] The observation noise ε_t has nonzero mean Hx_t; the text correctly notes this, but an explicit zero-mean reformulation would help readability.
  5. [Figures 2-5] Please describe how the pointwise credible intervals are computed from the ensemble/smoother samples.
  6. [References] Several related works are cited as preprints; ensure full publication data are provided where available.

Circularity Check

0 steps flagged

No significant circularity: the Markov prior is derived from classical local-error analysis, Theorem 3.1 is proved from explicit conditions, and self-citations are not load-bearing.

full rationale

The central derivation is self-contained. The proposed Markov prior (15) is motivated by the classical global-error propagation identity derived in Section 3.1 via the mean-value theorem (Eq. 14), not by the paper's target result. Theorem 3.1 is proved in Appendix C directly from Conditions (I)-(IV) using Markov's inequality and a recursive error bound; no step of the proof assumes the posterior being estimated or the numerical results being claimed. The experiments compare the inferred discretization-error mean against a highly accurate independent reference solution, so the demonstration is not constructed from the fitted posterior itself. There is repeated self-citation to Toyota and Miyatake (2025), but the cited prior construction and Theorem 4.1 are re-derived here with proof, so the citation is not load-bearing. A separate concern--that the experimental prior M=mI with m~N(alpha,beta^2) has E|1-m| constant in h and p(mu_0)=N(0,I), so Conditions (I) and (III) of Theorem 3.1 are not satisfied, making the Section 4 claim 'consistent with the theoretical implication' unsupported--is an applicability/correctness gap, not a circular reduction: Theorem 3.1 is a conditional statement and the empirical validation has independent content through the reference-solution comparison.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The work introduces no new physical or ontological entities; the discretization error mean is a statistical parameter. The free parameters (α,β,γ) are fitted to the observation trajectory via marginal likelihood. The axioms are a mix of standard ODE/observation assumptions and theorem-specific conditions (I)-(IV), with Condition (I) and (III) not met by the experimental setup.

free parameters (3)
  • α (mean of m in M=mI) = 1.0 (pendulum, FHN, Lorenz-96)
    Chosen by maximizing the ensemble-based marginal likelihood over a grid α∈{−1.4,...,1.4}; optimal at 1.0 in all experiments.
  • β (std of m in M=mI) = 0.3 (pendulum), 0.3 (FHN), 1.0 (Lorenz-96)
    Chosen by marginal likelihood; controls the randomness in the prior transition M=mI.
  • γ (scale of discretization error covariance Σ=γI) = 0.5 (pendulum), 4.0 or reported 0.4 (FHN), 1.0 (Lorenz-96)
    Chosen by marginal likelihood; affects the observation noise covariance HΣH^T+Γ.
axioms (7)
  • domain assumption The exact solution x(t) satisfies the ODE dx/dt=f(x) with continuous f.
    Standard well-posedness assumption in numerical analysis; used throughout the paper.
  • domain assumption Observations follow y*_t = Hx(t)+ε_t with ε_t∼N(0,Γ), Γ known positive definite.
    Equation (2), Section 2.1; the observation model is assumed known.
  • ad hoc to paper Discretization error is Gaussian: x(t_i)−x_{t_i} ∼ N(μ_i, Σ).
    Equation (3), Section 2.2; this is a modeling assumption central to the framework, not derived from first principles.
  • ad hoc to paper Condition (I): E||I−M||_op ≤ C_M h.
    Assumed in Theorem 3.1; not satisfied by the experimental M=mI with fixed β because E|1−m| is constant in h, making the theorem inapplicable to the experiments.
  • domain assumption Condition (II): local error L(t_i,j) ≤ C_L h^{a+1} for a solver of order a.
    Standard numerical analysis assumption for consistent methods; used in the proof of Theorem 3.1.
  • ad hoc to paper Condition (III): E||μ_t0|| ≤ C_μ h^b.
    Assumed in Theorem 3.1; experiments use p(μ_t0)=N(0,I), independent of h, so b=0 and the theorem yields no nontrivial convergence.
  • ad hoc to paper Condition (IV): M_i,j is independent of μ_i,j.
    Assumed in Theorem 3.1 proof; likely true by construction in the experiments, but not explicitly verified.

pith-pipeline@v1.3.0-daily-deepseek · 19907 in / 17024 out tokens · 157682 ms · 2026-08-01T13:33:38.890550+00:00 · methodology

0 comments
read the original abstract

We propose a Bayesian framework to quantify discretization errors in numerical solutions of ODE models based on observational data. The discretization error is modeled as a random variable, and its mean-referred to as the discretization error mean-is inferred from the observations. By introducing a Markov prior on the temporal evolution of the discretization error mean, we formulate the problem as a state-space model with a linear Gaussian observation process, which enables efficient inference via the Ensemble Kalman Filter. We also propose a specific form of a Markov prior motivated by classical discretization error analysis, in which global errors accumulate from local errors. The proposed prior depends on the step size of a numerical solver, and we establish its convergence rate in probability as the step size tends to zero. Numerical experiments on the pendulum system and the FitzHugh-Nagumo model demonstrate the effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 2607.26552 by Shoji Toyota, Yuto Miyatake.

Figure 1
Figure 1. Figure 1: Reformulation of the objective in (7) as a state-space model. The latent transition follows the Markov prior p(µti+1 | µti ), and the observation model is N [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Discretization error quantification results for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Discretization error quantification results for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Discretization error quantification results for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Discretization error quantification results for [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Discretization error quantification results for the 8-dimensional Lorenz–96 model. The posterior mean and 95% credible [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

295 extracted references · 1 canonical work pages

  1. [1]

    doi:10.1137/140951758 , journal =

    An Adaptive Shifted Power Method for Computing Generalized Tensor Eigenpairs , author =. doi:10.1137/140951758 , journal =

    doi:10.1137/140951758

  2. [2]

    Proceedings of the royal society of london

    A contribution to the mathematical theory of epidemics , author=. Proceedings of the royal society of london. Series A, Containing papers of a mathematical and physical character , volume=. 1927 , publisher=

    1927

  3. [3]

    Propagating Model Uncertainty through Filtering-based Probabilistic Numerical

    Yao, Dingling and Tronarp, Filip and Bosch, Nathanael , booktitle =. Propagating Model Uncertainty through Filtering-based Probabilistic Numerical

  4. [4]

    2013 , publisher=

    Matrix computations , author=. 2013 , publisher=

    2013

  5. [5]

    Analysis scheme in the ensemble

    Burgers, Gerrit and Jan van Leeuwen, Peter and Evensen, Geir , journal=. Analysis scheme in the ensemble

  6. [6]

    Adaptive Probabilistic

    Kr\". Adaptive Probabilistic. Proceedings of the First International Conference on Probabilistic Numerics , year =

  7. [7]

    Nature , year=

    Fluctuations in the Abundance of a Species considered Mathematically , author=. Nature , year=

  8. [8]

    2010 , publisher=

    Modern control engineering , author=. 2010 , publisher=

    2010

  9. [9]

    Biophysical journal , volume=

    Impulses and physiological states in theoretical models of nerve membrane , author=. Biophysical journal , volume=. 1961 , publisher=

    1961

  10. [10]

    and Richardson, S

    Gilks, W.R. and Richardson, S. and Spiegelhalter, D. , interhash =

  11. [11]

    Parallel-in-time probabilistic numerical

    Bosch, Nathanael and Corenflos, Adrien and Yaghoobi, Fatemeh and Tronarp, Filip and Hennig, Philipp and S. Parallel-in-time probabilistic numerical. Journal of Machine Learning Research , volume=

  12. [12]

    and Nørsett,

    Hairer, E. and Nørsett,. Solving Ordinary Differential Equations I: Nonstiff Problems , edition =

  13. [13]

    Statistics and Computing , volume =

    Convergence rates of. Statistics and Computing , volume =. 2020 , slug =

    2020

  14. [14]

    Schober, M. and S. A probabilistic model for the numerical solution of initial value problems , journal =. 2018 , slug =

    2018

  15. [15]

    Statistics and Computing , volume=

    Random time step probabilistic methods for uncertainty quantification in chaotic and geometric numerical integration , author=. Statistics and Computing , volume=. 2020 , publisher=

    2020

  16. [16]

    J. Mo. On Bayesian methods for seeking the extremum and their application , booktitle =

  17. [17]

    J. Mo. The application of Bayesian methods for seeking the extremum , journal =

  18. [18]

    J. Mo. Bayesian Approach to Global Optimization , publisher =

  19. [19]

    Snoek and H

    J. Snoek and H. Larochelle and R. P. Adams , title =. NeurIPS , volume =

  20. [20]

    2008 , isbn =

    Iserles, Arieh , title =. 2008 , isbn =

    2008

  21. [21]

    Hennig and M

    P. Hennig and M. Kiefel , title =. Journal of Machine Learning Research , volume =

  22. [22]

    Mahsereci and P

    M. Mahsereci and P. Hennig , title =. NeurIPS , volume =

  23. [23]

    Diaconis , title =

    P. Diaconis , title =. Statistical Decision Theory and Related Topics IV , volume =

  24. [24]

    O'Hagan , title =

    A. O'Hagan , title =. Bayesian Statistics 4: Proc. 4th Valencia International Meeting , pages =

  25. [25]

    Fisher and C

    M. Fisher and C. Oates and C. Powell and A. Teckentrup , title =. Proceedings of Machine Learning Research (PMLR) , volume =

  26. [26]

    Prüher and S

    J. Prüher and S. Särkkä , title =. IEEE 26th International Workshop on Machine Learning for Signal Processing , year =

  27. [27]

    Gessner and J

    A. Gessner and J. Gonzalez and M. Mahsereci , title =. Proceedings of Machine Learning Research (PMLR) , volume =

  28. [28]

    Karvonen and S

    T. Karvonen and S. Särkkä and C. J. Oates , title =. Statistics and Computing , volume =

  29. [29]

    H. R. Chai and R. Garnett , title =. Proceedings of Machine Learning Research (PMLR) , volume =

  30. [30]

    Jagadeeswaran and F

    R. Jagadeeswaran and F. J. Hickernell , title =. Statistics and Computing , volume =

  31. [31]

    Karvonen and S

    T. Karvonen and S. Särkkä , title =. IEEE 27th International Workshop on Machine Learning for Signal Processing , year =

  32. [32]

    Karvonen and C

    T. Karvonen and C. J. Oates and S. Särkkä , title =. NeurIPS , volume =

  33. [33]

    Osborne and R

    M. Osborne and R. Garnett and S. Roberts and C. Hart and S. Aigrain and N. Gibson , title =. Proceedings of Machine Learning Research (PMLR) , volume =

  34. [34]

    Xi and F.-X

    X. Xi and F.-X. Briol and M. Girolami , title =. Proceedings of Machine Learning Research (PMLR) , volume =

  35. [35]

    Nonlinear Processes in Geophysics , volume=

    Merging particle filter for sequential data assimilation , author=. Nonlinear Processes in Geophysics , volume=. 2007 , publisher=

    2007

  36. [36]

    Monthly Weather Review , volume=

    Particle filtering in geophysical systems , author=. Monthly Weather Review , volume=

  37. [37]

    Briol and C

    F.-X. Briol and C. J. Oates and M. Girolami and M. A. Osborne and D. Sejdinovic and others , title =. NeurIPS , volume =

  38. [38]

    Gunter and M

    T. Gunter and M. A. Osborne and R. Garnett and P. Hennig and S. J. Roberts , title =. NeurIPS , volume =

  39. [39]

    Kennedy , title =

    M. Kennedy , title =. Statistics and Computing , volume =

  40. [40]

    Nonlinear Processes in Geophysics , volume=

    Comparison of local particle filters and new implementations , author=. Nonlinear Processes in Geophysics , volume=. 2018 , publisher=

    2018

  41. [41]

    The Annals of Applied Probability , number =

    Alexandros Beskos and Dan Crisan and Ajay Jasra , title =. The Annals of Applied Probability , number =

  42. [42]

    Statistics and computing , volume=

    Sampling from multimodal distributions using tempered transitions , author=. Statistics and computing , volume=. 1996 , publisher=

    1996

  43. [43]

    Monthly Weather Review , volume=

    Implicit particle methods and their connection with variational data assimilation , author=. Monthly Weather Review , volume=

  44. [44]

    2019 , publisher=

    Probability: theory and examples , author=. 2019 , publisher=

    2019

  45. [45]

    Communications in Applied Mathematics and Computational Science , volume=

    Implicit particle filters for data assimilation , author=. Communications in Applied Mathematics and Computational Science , volume=. 2010 , publisher=

    2010

  46. [46]

    State-Space Models: Applications in Economics and Finance , pages=

    A survey of implicit particle filters for data assimilation , author=. State-Space Models: Applications in Economics and Finance , pages=. 2013 , publisher=

    2013

  47. [47]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Sequential monte carlo samplers , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2006 , publisher=

    2006

  48. [48]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Particle filters for high-dimensional geoscience applications: A review , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2019 , publisher=

    2019

  49. [49]

    O'Hagan , title =

    A. O'Hagan , title =. Journal of Statistical Planning and Inference , volume =

  50. [50]

    Larkin , title =

    F. Larkin , title =. Rocky Mountain Journal of Mathematics , volume =

  51. [51]

    C. E. Rasmussen and Z. Ghahramani , title =. NeurIPS , volume =

  52. [52]

    Briol and C

    F.-X. Briol and C. J. Oates and M. Girolami and M. A. Osborne and D. Sejdinovic and others , title =. Statistical Science , volume =

  53. [53]

    Calcolo , volume=

    Randomised one-step time integration methods for deterministic operator differential equations , author=. Calcolo , volume=. 2022 , publisher=

    2022

  54. [54]

    Statistics and Computing , volume=

    Statistical analysis of differential equations: introducing probability measures on numerical solutions , author=. Statistics and Computing , volume=. 2017 , publisher=

    2017

  55. [55]

    Statistics and Computing , volume=

    Strong convergence rates of probabilistic integrators for ordinary differential equations , author=. Statistics and Computing , volume=. 2019 , publisher=

    2019

  56. [56]

    and Kersting, H

    Tronarp, F. and Kersting, H. and S. Probabilistic Solutions To Ordinary Differential Equations As Non-Linear. Statistics and Computing , volume =. 2019 , slug =

    2019

  57. [57]

    Tronarp, F. and S. Bayesian. Statistics and Computing , volume =. 2021 , slug =

    2021

  58. [58]

    2001 , publisher=

    Sequential Monte Carlo methods in practice , author=. 2001 , publisher=

    2001

  59. [59]

    Gordon and D.J

    N.J. Gordon and D.J. Salmond and A.F.M. Smith , title =. IEE Proceedings F (Radar and Signal Processing) , volume =

  60. [60]

    Monte Carlo Filter and Smoother for Non-

    Genshiro Kitagawa , journal =. Monte Carlo Filter and Smoother for Non-

  61. [61]

    On the global error of discretization methods for ordinary differential equations , year =

    Niesen, J , journal =. On the global error of discretization methods for ordinary differential equations , year =

  62. [62]

    Advances in Neural Information Processing Systems (NeurIPS) , title =

    Wenger, Jonathan and Pleiss, Geoff and Pf. Advances in Neural Information Processing Systems (NeurIPS) , title =

  63. [63]

    Proceedings of the 39th International Conference on Machine Learning , year =

    Fenrir: Physics-Enhanced Regression for Initial Value Problems , author =. Proceedings of the 39th International Conference on Machine Learning , year =

  64. [64]

    Bayesian Analysis , volume=

    Modelling Pathwise Uncertainty of Stochastic Differential Equations Samplers via Probabilistic Numerics , author=. Bayesian Analysis , volume=. 2025 , publisher=

    2025

  65. [65]

    Statistics and Computing , volume=

    Probabilistic linear solvers: a unifying view , author=. Statistics and Computing , volume=. 2019 , publisher=

    2019

  66. [66]

    A Bayesian conjugate gradient method (with discussion) , volume =

    Cockayne, Jon and Oates, Chris J and Ipsen, Ilse CF and Girolami, Mark , bibtex_show =. A Bayesian conjugate gradient method (with discussion) , volume =. Bayesian Analysis , number =

  67. [67]

    Probabilistic Linear Solvers for Machine Learning , year =

    Wenger, Jonathan and Hennig, Philipp , bibtex_show =. Probabilistic Linear Solvers for Machine Learning , year =. Advances in Neural Information Processing Systems (NeurIPS) , code =

  68. [68]

    and Kersting, Hans P

    Hennig, Philipp and Osborne, Michael A. and Kersting, Hans P. Probabilistic Numerics: Computation as Machine Learning. 2022

    2022

  69. [69]

    Adaptive Checkpoint Adjoint Method for Gradient Estimation in Neural

    Zhuang, Juntang and Dvornek, Nicha and Li, Xiaoxiao and Tatikonda, Sekhar and Papademetris, Xenophon and Duncan, James , booktitle =. Adaptive Checkpoint Adjoint Method for Gradient Estimation in Neural. 2020 , editor =

    2020

  70. [70]

    The Thirteenth International Conference on Learning Representations , year=

    Compositional simulation-based inference for time series , author=. The Thirteenth International Conference on Learning Representations , year=

  71. [71]

    FNOPE: Simulation-based inference on function spaces with Fourier Neural Operators , journal =

  72. [72]

    , edition =

    Butcher, John C. , edition =

  73. [73]

    Data Assimilation: the Ensemble

    Evensen, Geir , year=. Data Assimilation: the Ensemble

  74. [74]

    and Sullivan, T

    Cockayne, Jon and Oates, Chris J. and Sullivan, T. J. and Girolami, Mark , title =. SIAM Review , volume =

  75. [75]

    Understanding the ensemble

    Katzfuss, Matthias and Stroud, Jonathan R and Wikle, Christopher K , journal=. Understanding the ensemble. 2016 , publisher=

    2016

  76. [76]

    2023 , publisher=

    Bayesian Filtering and Smoothing , author=. 2023 , publisher=

    2023

  77. [77]

    Transactions on Machine Learning Research , issn=

    Numerically Robust Fixed-Point Smoothing Without State Augmentation , author=. Transactions on Machine Learning Research , issn=. 2025 , url=

    2025

  78. [78]

    Japan Journal of Industrial and Applied Mathematics , author =

    Quantifying uncertainty in the numerical integration of evolution equations based on. Japan Journal of Industrial and Applied Mathematics , author =

  79. [79]

    Toyota, Shoji and Miyatake, Yuto , journal=. Joint

  80. [80]

    SIAM/ASA Journal on Uncertainty Quantification , volume =

    Matsuda, Takeru and Miyatake, Yuto , title =. SIAM/ASA Journal on Uncertainty Quantification , volume =

Showing first 80 references.