pith. sign in

arxiv: 2606.03230 · v1 · pith:6N5MHWQ2new · submitted 2026-06-02 · 📊 stat.ME · stat.CO

Predictively-Oriented Kalman Filtering

Zheyang Shen , Gerardo Duran-Martin , Chris. J. Oates This is my paper

Pith reviewed 2026-06-28 09:03 UTC · model grok-4.3

classification 📊 stat.ME stat.CO
keywords Kalman filteringstate-space modelsmodel misspecificationpost-Bayesian inferencepredictively oriented posteriorsnonlinear filteringonline filtering
0
0 comments X

The pith

EKF-PrO performs online filtering in nonlinear state-space models without becoming overconfident under misspecification.

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

This paper develops a post-Bayesian approach to online filtering for nonlinear state-space models that avoids overconfident inferences when the dynamical model, the measurement model, or both are misspecified. It relies on predictively oriented posteriors, which concentrate only when the overall model is well-specified instead of following Bayes' theorem strictly. The central technical step is the derivation of a fast approximate linear-Gaussian update rule analogous to an iterated extended Kalman filter. The resulting EKF-PrO method requires no tunable hyperparameters and has computational cost comparable to standard filtering techniques. Performance is shown through tests on linear and nonlinear examples with systematic misspecification.

Core claim

The paper introduces EKF-PrO, a fast approximate linear-Gaussian update procedure for predictively-oriented posteriors in nonlinear state-space models. This procedure is analogous to an iterated extended Kalman filter and ensures posterior concentration occurs if and only if the overall model is well-specified, without strict adherence to Bayes' theorem and without any tunable hyperparameters.

What carries the argument

The predictively-oriented (PrO) posterior updated by an approximate linear-Gaussian rule analogous to the iterated extended Kalman filter.

If this is right

  • Posterior concentration occurs only when the state-space model is correctly specified.
  • The filter requires no hyperparameter tuning.
  • Computational cost remains comparable to that of standard extended Kalman filters.
  • The method applies directly to both linear and nonlinear models under misspecification.

Where Pith is reading between the lines

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

  • The same predictively-oriented update idea could be adapted to particle filters or other sequential Monte Carlo methods.
  • This framework may improve calibration in real-time tracking applications where models are known to be imperfect approximations.
  • Connections to other post-Bayesian online methods could yield a broader class of hyperparameter-free robust filters.

Load-bearing premise

A fast approximate linear-Gaussian update can be derived for predictively-oriented posteriors that preserves concentration only under correct specification without introducing bias or requiring tunable hyperparameters.

What would settle it

Apply EKF-PrO to data generated from a misspecified nonlinear state-space model and check whether the resulting posterior intervals achieve nominal coverage probabilities on held-out predictions.

Figures

Figures reproduced from arXiv: 2606.03230 by Chris. J. Oates, Gerardo Duran-Martin, Zheyang Shen.

Figure 1
Figure 1. Figure 1: Illustration on the Gaussian location model. Data ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tracking an object in 2D. Here the mean square error for the filtering mean of [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The misspecified Lorenz96 model. For a filtering algorithms with mean [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example of the error Lt = ∥xt − mt∥ for the location variable, as a function of time for KF and KF-PrO in the maneuver type of misspecification. The background shading denotes the existence of systematic misspecification, with white background indicating no misspecification. We can see from this plot that KF-PrO starts out similar to KF, but the propagation of overconfident predictions, and the consequent … view at source ↗
Figure 5
Figure 5. Figure 5: Tracing the mean of EKF-PrO and EKF against the ground truth state (black dotted). Even though EKF is able to capture the trajectory most of the time, it does not match EKF-PrO in accuracy, especially around the area t = 17.5. Similar to the maneuver example, we set two discrete modes, with ϕ (1) = 8 · 1, ϕ (2) ∼ N (8 · 1, I). The discrete model transition matrix is given by A =  0.995 0.005 0.005 0.995 … view at source ↗
read the original abstract

This paper presents a post-Bayesian approach to online filtering in nonlinear state-space models, capable of avoiding over-confident inferences in settings where either the dynamical model, the measurement model, or both, could be misspecified. This is addressed using predictively oriented (PrO) posteriors, an emerging paradigm in which learning (i.e., posterior concentration) occurs if and only if the overall model is well-specified, without strict adherence to Bayes' theorem. As the characterisation of PrO posteriors is challenging, our main technical contribution is a fast approximate linear-Gaussian update procedure, analogous to an (iterated) extended Kalman filter. The methodology, which we call EKF-PrO, has no tunable hyper-parameters and has a computational cost comparable to that of existing filtering methods. Performance is empirically assessed on a range of linear and non-linear applications, in which the state-space model is systematically misspecified.

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

0 major / 2 minor

Summary. The paper introduces EKF-PrO, a hyperparameter-free approximate linear-Gaussian update for predictively-oriented (PrO) posteriors in nonlinear state-space models. It enables online filtering that avoids over-confident inferences under misspecification of the dynamical or measurement model (or both), with the property that posterior concentration occurs if and only if the overall model is correctly specified. The method is constructed via moment-matching analogous to an iterated extended Kalman filter, has computational cost comparable to existing filters, and is assessed empirically on systematically misspecified linear and nonlinear examples.

Significance. If the derivation and experiments hold, the work is significant for providing a practical, tuning-free post-Bayesian filtering procedure that inherits the predictive-orientation property by construction. The absence of free parameters, the explicit moment-matching construction, and the validation on misspecified models are notable strengths that address a common practical limitation of standard Kalman-type filters.

minor comments (2)
  1. [Abstract] Abstract: the phrase 'post-Bayesian approach' is used without a one-sentence reminder of how PrO posteriors differ from standard Bayesian updating; a brief parenthetical would improve accessibility.
  2. [Experiments] The empirical section would benefit from an explicit statement of the misspecification levels (e.g., parameter deviation magnitudes) used in the linear and nonlinear test cases to allow direct replication.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript, the recognition of its practical contributions, and the recommendation to accept. No major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper defines EKF-PrO as a new approximate linear-Gaussian update for predictively-oriented (PrO) posteriors in nonlinear state-space models, with the concentration property (only under correct specification) inherited from the PrO paradigm described as emerging and external. The derivation is framed as analogous to iterated EKF without tunable hyperparameters, and no equations or steps reduce by construction to fitted inputs, self-definitions, or self-citation chains. The approximation preserves the target property without redefining inputs as outputs or smuggling ansatzes. This is a standard non-circular technical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the existence and properties of predictively-oriented posteriors (an emerging paradigm) and on the feasibility of a fast linear-Gaussian approximation that inherits those properties. No free parameters are introduced by the paper itself.

axioms (1)
  • domain assumption PrO posteriors concentrate if and only if the overall model is well-specified
    This is the defining property of the PrO paradigm invoked to motivate the work.

pith-pipeline@v0.9.1-grok · 5682 in / 1102 out tokens · 27819 ms · 2026-06-28T09:03:35.980548+00:00 · methodology

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

60 extracted references · 3 linked inside Pith

  1. [1]

    Agamennoni, J

    G. Agamennoni, J. I. Nieto, and E. M. Nebot. Approximate inference in state-space models with heavy-tailed noise. IEEE Transactions on Signal Processing, 60 0 (10): 0 5024--5037, 2012

    2012

  2. [2]

    P. G. Bissiri, C. C. Holmes, and S. G. Walker. A general framework for updating belief distributions. Journal of the Royal Statistical Society Series B: Statistical Methodology, 78 0 (5): 0 1103--1130, 2016

    2016

  3. [3]

    C. G. Boncelet and B. W. Dickinson. An approach to robust K alman filtering. In IEEE Conference on Decision and Control, pages 304--305, 1983

    1983

  4. [4]

    Boustati, O

    A. Boustati, O. D. Akyildiz, T. Damoulas, and A. Johansen. Generalised B ayesian filtering via sequential M onte C arlo. Advances in Neural Information Processing Systems, 34: 0 418--429, 2020

    2020

  5. [5]

    P. G. Chang, G. Dur \'a n-Mart \' n, A. Shestopaloff, M. Jones, and K. P. Murphy. Low-rank extended K alman filtering for online learning of neural networks from streaming data. In Conference on Lifelong Learning Agents, pages 1025--1071. PMLR, 2023

    2023

  6. [6]

    Chazal, H

    C. Chazal, H. Kanagawa, Z. Shen, A. Korba, and C. J. Oates. A computable measure of suboptimality for entropy-regularised variational objectives. arXiv preprint arXiv:2509.10393, 2025

    arXiv 2025

  7. [7]

    S.-Y. Chen. Kalman filter for robot vision: A survey. IEEE Transactions on Industrial Electronics, 59 0 (11): 0 4409--4420, 2011

    2011

  8. [8]

    Dahan, G

    Y. Dahan, G. Revach, J. Dunik, and N. Shlezinger. Bayesian KalmanNet : Q uantifying uncertainty in deep learning augmented K alman filter. IEEE Transactions on Signal Processing, 73: 0 2558--2573, 2025

    2025

  9. [9]

    S. Das. Robust State Estimation Methods for Robotics Applications. PhD thesis, West Virginia University, 2023

    2023

  10. [10]

    Duran-Martin, M

    G. Duran-Martin, M. Altamirano, A. Y. Shestopaloff, L. S \'a nchez-Betancourt, J. Knoblauch, M. Jones, F.-X. Briol, and K. Murphy. Outlier-robust K alman filtering through generalised B ayes. In Proceedings of the 41st International Conference on Machine Learning, 2024

    2024

  11. [11]

    Duran-Martin, L

    G. Duran-Martin, L. S \'a nchez-Betancourt, A. Shestopaloff, and K. Murphy. A unifying framework for generalised B ayesian online learning in non-stationary environments. Transactions on Machine Learning Research, 2025

    2025

  12. [12]

    Durbin and S

    J. Durbin and S. J. Koopman. Time Series Analysis by State Space Methods. Oxford University Press, 2012

    2012

  13. [13]

    G. Evensen. The ensemble K alman filter: T heoretical formulation and practical implementation. Ocean Dynamics, 53 0 (4): 0 343--367, 2003

    2003

  14. [14]

    Fitzgerald

    R. Fitzgerald. Divergence of the K alman filter. IEEE Transactions on Automatic Control, 16 0 (6): 0 736--747, 2003

    2003

  15. [15]

    A. Gelb. Applied Optimal Estimation. MIT Press, 1974

    1974

  16. [16]

    Gneiting and A

    T. Gneiting and A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102 0 (477): 0 359--378, 2007

    2007

  17. [17]

    Huang, Y

    Y. Huang, Y. Zhang, N. Li, and J. Chambers. A robust G aussian approximate filter for nonlinear systems with heavy tailed measurement noises. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4209--4213, 2016

    2016

  18. [18]

    Huang, Y

    Y. Huang, Y. Zhang, N. Li, Z. Wu, and J. A. Chambers. A novel robust student's t-based K alman filter. IEEE Transactions on Aerospace and Electronic Systems, 53 0 (3): 0 1545--1554, 2017

    2017

  19. [19]

    Huang, F

    Y. Huang, F. Zhu, G. Jia, and Y. Zhang. A slide window variational adaptive K alman filter. IEEE Transactions on Circuits and Systems II: Express Briefs, 67 0 (12): 0 3552--3556, 2020

    2020

  20. [20]

    Ito and K

    K. Ito and K. Xiong. Gaussian filters for nonlinear filtering problems. IEEE Transactions on Automatic Control, 45 0 (5): 0 910--927, 2002

    2002

  21. [21]

    Jankowiak, G

    M. Jankowiak, G. Pleiss, and J. Gardner. Parametric G aussian process regressors. In International Conference on Machine Learning, pages 4702--4712. PMLR, 2020

    2020

  22. [22]

    Jazwinski

    A. Jazwinski. Stochastic Processes and Filtering Theory. Academic Press, 1970

    1970

  23. [23]

    Jones, P

    M. Jones, P. Chang, and K. Murphy. Bayesian online natural gradient (BONG) . Advances in Neural Information Processing Systems, 37: 0 131104--131153, 2024

    2024

  24. [24]

    S. J. Julier. A general method for approximating non-linear transformations of probability distributions. 1996. URL http://www.robots.ox.ac.uk/\ siju/work/work.html

    1996

  25. [25]

    R. E. Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82 0 (Series D): 0 35--45, 1960

    1960

  26. [26]

    C. D. Karlgaard. Nonlinear regression H uber-- K alman filtering and fixed-interval smoothing. Journal of Guidance, Control, and Dynamics, 38 0 (2): 0 322--330, 2015

    2015

  27. [27]

    R. Kelly. Reducing geometric dilution of precision using ridge regression. IEEE Transactions on Aerospace and Electronic Systems, 26 0 (1): 0 154--168, 2002

    2002

  28. [28]

    Knoblauch, J

    J. Knoblauch, J. Jewson, and T. Damoulas. An optimization-centric view on B ayes' rule: Reviewing and generalizing variational inference. Journal of Machine Learning Research, 23 0 (132): 0 1--109, 2022

    2022

  29. [29]

    Kuleshov, N

    V. Kuleshov, N. Fenner, and S. Ermon. Accurate uncertainties for deep learning using calibrated regression. In International Conference on Machine Learning, 2018

    2018

  30. [30]

    Lai and Y

    J. Lai and Y. Yao. Predictive variational inference: Learn the predictively optimal posterior distribution. arXiv preprint arXiv:2410.14843, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    N. Laird. Nonparametric maximum likelihood estimation of a mixing distribution. Journal of the American Statistical Association, 73 0 (364): 0 805--811, 1978

    1978

  32. [32]

    Lambert, S

    M. Lambert, S. Bonnabel, and F. Bach. The recursive variational Gaussian approximation (R-VGA) . Statistics and Computing, 32 0 (1): 0 10, 2022

    2022

  33. [33]

    D. C. Liu and J. Nocedal. On the limited memory BFGS method for large scale optimization. Mathematical Programming, 45 0 (1): 0 503--528, 1989

    1989

  34. [34]

    G. Liu. Data quality problems troubling business and financial researchers: A literature review and synthetic analysis. Journal of Business & Finance Librarianship, 25 0 (3-4): 0 315--371, 2020

    2020

  35. [35]

    Q. Liu, M. A. Fisher, Z. Shen, K. Tant, X. Zhao, A. Curtis, and C. J. Oates. Detecting model misspecification in B ayesian inverse problems via variational gradient descent. arXiv preprint arXiv:2512.01667, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    E. N. Lorenz. Predictability of Weather and Climate, chapter Predictability: A problem partly solved. Cambridge University Press, 1996

    1996

  37. [37]

    Masegosa

    A. Masegosa. Learning under model misspecification: A pplications to variational and ensemble methods. Advances in Neural Information Processing Systems, 34, 2020

    2020

  38. [38]

    P. S. Maybeck. Stochastic Models, Estimation, and Control. Academic Press, 1982

    1982

  39. [39]

    McLatchie, B.-E

    Y. McLatchie, B.-E. Cherief-Abdellatif, D. T. Frazier, and J. Knoblauch. Predictively oriented posteriors. arXiv preprint arXiv:2510.01915, 2025

    arXiv 2025

  40. [40]

    R. Mehra. Approaches to adaptive filtering. IEEE Transactions on Automatic Control, 17 0 (5): 0 693--698, 2003

    2003

  41. [41]

    J. W. Miller. Asymptotic normality, concentration, and coverage of generalized posteriors. Journal of Machine Learning Research, 22 0 (168): 0 1--53, 2021

    2021

  42. [42]

    W. R. Morningstar, A. Alemi, and J. V. Dillon. PACm-Bayes : N arrowing the empirical risk gap in the misspecified B ayesian regime. In International Conference on Artificial Intelligence and Statistics, 2022

    2022

  43. [43]

    K. P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023

    2023

  44. [44]

    Nogueira

    F. Nogueira. Bayesian Optimization : Open source constrained global optimization tool for Python , 2014. URL https://github.com/bayesian-optimization/BayesianOptimization

    2014

  45. [45]

    Nurminen, T

    H. Nurminen, T. Ardeshiri, R. Piché, and F. Gustafsson. Robust inference for state-space models with skewed measurement noise. IEEE Signal Processing Letters, 22 0 (11): 0 1898--1902, 2015

    1902

  46. [46]

    Piché, S

    R. Piché, S. Särkkä, and J. Hartikainen. Recursive outlier-robust filtering and smoothing for nonlinear systems using the multivariate student-t distribution. In IEEE International Workshop on Machine Learning for Signal Processing, 2012

    2012

  47. [47]

    M. Roth, E. \"O zkan, and F. Gustafsson. A student's t filter for heavy tailed process and measurement noise. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5770--5774. IEEE, 2013

    2013

  48. [48]

    M. Roth, T. Ardeshiri, E. \"O zkan, and F. Gustafsson. Robust B ayesian filtering and smoothing using S tudent's-t distribution. arXiv preprint arXiv:1703.02428, 2017 a

    Pith/arXiv arXiv 2017

  49. [49]

    M. Roth, G. Hendeby, C. Fritsche, and F. Gustafsson. The ensemble K alman filter: A signal processing perspective. Journal on Advances in Signal Processing, 2017: 0 1--16, 2017 b

    2017

  50. [50]

    Schlee, C

    F. Schlee, C. Standish, and N. Toda. Divergence in the K alman filter. AIAA Journal, 5 0 (6): 0 1114--1120, 1967

    1967

  51. [51]

    Z. Shen, J. Knoblauch, S. Power, and C. J. Oates. Prediction-centric uncertainty quantification via MMD . In International Conference on Artificial Intelligence and Statistics, 2025

    2025

  52. [52]

    Sheth and R

    R. Sheth and R. Khardon. Pseudo- B ayesian learning via direct loss minimization with applications to sparse G aussian process models. In Symposium on Advances in Approximate Bayesian Inference, 2020

    2020

  53. [53]

    Särkkä and L

    S. Särkkä and L. Svensson. B ayesian Filtering and Smoothing (2nd edition) . Cambridge University Press, 2023

    2023

  54. [54]

    P. D. Tao and L. H. An. Convex analysis approach to DC programming: T heory, algorithms and applications. Acta Mathematica Vietnamica, 22 0 (1): 0 289--355, 1997

    1997

  55. [55]

    S. Thrun. Probabilistic robotics. Communications of the ACM, 45 0 (3): 0 52--57, 2002

    2002

  56. [56]

    J.-A. Ting, E. Theodorou, and S. Schaal. Learning an outlier-robust K alman filter. In European Conference on Machine Learning, 2007

    2007

  57. [57]

    J. d. Vilmarest and O. Wintenberger. Viking: V ariational B ayesian variance tracking. Statistical Inference for Stochastic Processes, 27 0 (3): 0 839--860, 2024

    2024

  58. [58]

    H. Wang, H. Li, J. Fang, and H. Wang. Robust G aussian K alman filter with outlier detection. IEEE Signal Processing Letters, 25 0 (8): 0 1236--1240, 2018

    2018

  59. [59]

    Wenzel, K

    F. Wenzel, K. Roth, B. Veeling, J. Swiatkowski, L. Tran, S. Mandt, J. Snoek, T. Salimans, R. Jenatton, and S. Nowozin. How good is the B ayes posterior in deep neural networks really? In International Conference on Machine Learning, 2020

    2020

  60. [60]

    A. L. Yuille and A. Rangarajan. The concave-convex procedure. Neural Computation, 15 0 (4): 0 915--936, 2003

    2003