pith. sign in

arxiv: 2605.16644 · v1 · pith:UOHND3XZnew · submitted 2026-05-15 · 📡 eess.SY · cs.LG· cs.SY· math.OC· stat.ML

The Score Kalman Filter

Kaito Iwasaki , Anthony Bloch , Taeyoung Lee , Maani Ghaffari This is my paper

Pith reviewed 2026-05-20 15:31 UTC · model grok-4.3

classification 📡 eess.SY cs.LGcs.SYmath.OCstat.ML
keywords score matchingStein's identitymoment closurenonlinear filteringKalman filterBayesian filteringpartition function
0
0 comments X

The pith

The Score Kalman Filter closes nonlinear moment hierarchies with score matching and Stein's identity, avoiding all partition function integrals.

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

The paper develops the Score Kalman Filter to overcome the computational barrier of partition function integrals in moment-based nonlinear filtering. Score matching reduces density reconstruction to a linear solve from the moments, and Stein's identity then uses that score to close the moment hierarchy in both prediction and update. This approach keeps all operations in linear algebra and extends the feasible dimension from four to twenty states. It demonstrates lower error than conventional filters on coupled-oscillator examples. Readers interested in scalable uncertainty propagation in engineering systems would find this relevant.

Core claim

The central discovery is that score matching on propagated moments produces parameters that, when inserted into Stein's identity, close the moment hierarchy for both the prediction step under nonlinear dynamics and the update step after nonlinear measurements, thereby completing the Bayesian filter loop through linear algebra alone and recovering the information-form Kalman filter when the models are linear.

What carries the argument

Score matching for density reconstruction from moments via a single linear solve, combined with Stein's identity applied to the fitted score to close the moment hierarchy without integration.

If this is right

  • Every step of the predict-update cycle reduces to assembling and solving linear systems whose coefficients come directly from the current moments.
  • The SKF recovers the classical information-form Kalman filter exactly when the dynamics and measurements are linear.
  • The method scales at least to twenty-dimensional state spaces on nonlinear coupled-oscillator networks.
  • It produces lower root-mean-square error than the extended, unscented, ensemble, and particle Kalman filters on the reported synthetic benchmarks.

Where Pith is reading between the lines

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

  • The linear-algebra structure may allow sparse or iterative solvers to push the dimension well beyond twenty in structured large-scale problems.
  • Replacing the current score model with a learned or parametric form could improve accuracy in strongly non-Gaussian regimes without changing the overall loop.
  • Analogous score-Steins closures might be applied to moment methods for stochastic differential equations or continuous-time filtering.

Load-bearing premise

Score matching from the propagated moments produces a density whose score function can be directly inserted into Stein's identity to accurately close the moment hierarchy for the nonlinear dynamics and measurement models.

What would settle it

In a low-dimensional nonlinear system where exact moments can be computed by direct numerical integration, check whether the SKF's predicted and updated moments match those exact values within sampling error as the moment order is increased.

Figures

Figures reproduced from arXiv: 2605.16644 by Anthony Bloch, Kaito Iwasaki, Maani Ghaffari, Taeyoung Lee.

Figure 1
Figure 1. Figure 1: SE(2) density reconstruction. The leftmost panel shows the Monte Carlo reference for the [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative LV centered moments (n=2, r=6, ¯d=1): Dynkin + Stein (blue dashed) vs. MC (black), degrees 2–6 (up to the score matching truncation order r). Error grows monotonically with degree. 7.1 SE(2) prediction with Fourier × position moments ( ¯d = 0) SE(2) rigid body kinematics in embedded coordinates (c, s, px, py) with ω=1 rad/s, v=2 m/s, heading noise σ=0.3. Drift and diffusion are linear, so ¯d… view at source ↗
Figure 3
Figure 3. Figure 3: Coupled oscillators (n=8, r=3). The SKF (blue) tracks all four positions through 25 steps, including unobserved oscillators (q2, q4) inferred from coupling alone. EKF (orange), UKF (purple), EnKF (brown), and PF (red) have larger RMSE. Multi-seed statistics are reported in Table A1. single-oscillator n=2 Duffing filter is in Appendix J.4. At N=3, 4, 5 (n=6, 8, 10, r=3) observing the odd-indexed positions, … view at source ↗
read the original abstract

A central obstacle in nonlinear Bayesian filtering is representing the belief distribution. Moment-based filters address this by propagating polynomial moments and reconstructing a density from them. Recent work completes the predict-update loop via the maximum-entropy (MaxEnt) principle, but each step requires the partition function and its gradient, both $n$-dimensional integrals whose cost scales exponentially, restricting the demonstrated MaxEnt moment filtering to $n \le 4$. We avoid the partition function entirely by combining score matching with Stein's identity. In our setting, score matching reduces the density fit to a single linear solve whose coefficients are assembled directly from the propagated moments. The same parameters then drive Stein's identity to close the moment hierarchy during prediction and to recover posterior moments after each Bayesian update, keeping the full predict-update loop free of partition function evaluation. The resulting Score Kalman Filter (SKF) reduces to the classical information-form Kalman filter as a special case and performs every step through linear algebra. On nonlinear coupled-oscillator networks, the SKF runs through $n=20$ and reports lower RMSE than the EKF, UKF, EnKF, and particle-filter baselines on the tested synthetic benchmarks.

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 / 2 minor

Summary. The manuscript introduces the Score Kalman Filter (SKF) for nonlinear Bayesian filtering. It propagates polynomial moments and reconstructs the density via score matching, which reduces to a single linear solve assembled from the moments. Stein's identity is then used to close the moment hierarchy in both the prediction and update steps, avoiding any partition-function evaluation. The method is claimed to reduce exactly to the classical information-form Kalman filter in the linear-Gaussian case and is demonstrated on coupled-oscillator networks with state dimension up to n=20, reporting lower RMSE than EKF, UKF, EnKF, and particle-filter baselines on the tested synthetic benchmarks.

Significance. If the central construction is shown to close the nonlinear moment hierarchy without additional bias or truncation, the SKF would provide a scalable, partition-function-free moment filter that operates entirely through linear algebra. The explicit reduction to the information-form Kalman filter and the ability to reach n=20 are concrete strengths that distinguish it from prior MaxEnt moment filters limited to n≤4.

major comments (2)
  1. [Abstract (central construction)] The central claim that score matching on the propagated moments followed by Stein's identity exactly closes the moment hierarchy for nonlinear dynamics and measurements is load-bearing. The abstract states that the linear solve assembles coefficients directly from the moments and that the same parameters drive Stein's identity, but provides no derivation showing that the resulting score produces E[score · polynomial] expectations that match the true integrals against the unknown transition kernel when the dynamics contain quadratic or cubic nonlinearities. This must be supplied with an explicit proof or a counter-example analysis.
  2. [Abstract (experimental claims)] The reported performance advantage on the n=20 oscillator benchmarks rests on the assumption that the score-matched density yields unbiased moment closure. Without error bars, the number of Monte-Carlo trials, or the precise experimental protocol, it is impossible to judge whether the lower RMSE is statistically meaningful or an artifact of the particular synthetic instances.
minor comments (2)
  1. [Abstract] The abstract mentions that every step is performed through linear algebra; a short pseudocode or complexity table would help readers verify the claimed O(n^3) scaling.
  2. [Method description] Clarify the precise score-matching objective (e.g., whether it is the standard Fisher divergence or a moment-weighted variant) and the exact linear system that is solved for the score parameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review of our manuscript on the Score Kalman Filter. The comments identify important points regarding the theoretical justification of the moment closure and the statistical rigor of the experiments. We respond to each major comment below and outline the changes we will incorporate in the revised version.

read point-by-point responses

  1. Referee: [Abstract (central construction)] The central claim that score matching on the propagated moments followed by Stein's identity exactly closes the moment hierarchy for nonlinear dynamics and measurements is load-bearing. The abstract states that the linear solve assembles coefficients directly from the moments and that the same parameters drive Stein's identity, but provides no derivation showing that the resulting score produces E[score · polynomial] expectations that match the true integrals against the unknown transition kernel when the dynamics contain quadratic or cubic nonlinearities. This must be supplied with an explicit proof or a counter-example analysis.

    Authors: We agree that the abstract is too concise to convey the full justification. The manuscript body (Sections 3.2–3.3 and 4.1) derives the score-matching linear system from the moments and shows how the resulting parameters enter Stein’s identity to produce the closed moment updates. Because Stein’s identity holds exactly for any sufficiently smooth density and the score is represented in the same polynomial basis used for the moments, the expectations E[score · p(x)] match the required integrals for polynomial nonlinearities up to cubic order without additional truncation. To make this explicit and self-contained, we will insert a new subsection (provisionally 3.4) that walks through the algebraic verification for quadratic and cubic terms, confirming that no bias is introduced beyond the score-matching approximation itself. revision: yes

  2. Referee: [Abstract (experimental claims)] The reported performance advantage on the n=20 oscillator benchmarks rests on the assumption that the score-matched density yields unbiased moment closure. Without error bars, the number of Monte-Carlo trials, or the precise experimental protocol, it is impossible to judge whether the lower RMSE is statistically meaningful or an artifact of the particular synthetic instances.

    Authors: We concur that additional experimental detail is required for reproducibility and statistical assessment. The revised manuscript will expand the numerical-results section to report: (i) 100 independent Monte-Carlo trials per filter, (ii) mean RMSE together with standard-deviation error bars, and (iii) the complete benchmark protocol, including initial-state distribution, process- and measurement-noise covariances, and the precise definition of the coupled-oscillator dynamics. These additions will allow readers to evaluate the significance of the reported improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the Score Kalman Filter derivation chain

full rationale

The derivation assembles score-matching parameters via a single linear solve whose coefficients come directly from the propagated moments, then applies Stein's identity to those parameters in order to close the moment hierarchy for the predict and update steps. This constitutes an approximation procedure whose validity rests on the quality of the score-matched density for the nonlinear push-forward, rather than any tautological reduction of a predicted quantity to the input moments by construction. The paper explicitly notes that the construction recovers the classical information-form Kalman filter as a special case, and no load-bearing self-citation, uniqueness theorem, or ansatz smuggling is invoked to justify the central steps. The method therefore remains self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard properties of score matching and Stein's identity drawn from prior literature; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Stein's identity holds for the score function obtained from the moment-based density fit
    Invoked to close the moment hierarchy during prediction and to recover posterior moments after update.

pith-pipeline@v0.9.0 · 5751 in / 1232 out tokens · 37350 ms · 2026-05-20T15:31:57.264984+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

61 extracted references · 61 canonical work pages

  1. [1]

    R. E. Kalman. A new approach to linear filtering and prediction problems.Journal of Basic Engineering (ASME Transactions, Series D), 82(1):35–45, 1960

    work page 1960

  2. [2]

    Stanley F. Schmidt. Application of state-space methods to navigation problems. InAdvances in Control Systems, volume 3, pages 293–340. Academic Press, 1966

    work page 1966

  3. [3]

    Jazwinski.Stochastic Processes and Filtering Theory

    Andrew H. Jazwinski.Stochastic Processes and Filtering Theory. Academic Press, New York, 1970

    work page 1970

  4. [4]

    Julier and Jeffrey K

    Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. InSignal Processing, Sensor Fusion, and Target Recognition VI, volume 3068 of Proceedings of SPIE, pages 182–193, 1997

    work page 1997

  5. [5]

    Wan and Rudolph van der Merwe

    Eric A. Wan and Rudolph van der Merwe. The unscented Kalman filter for nonlinear estimation. InProceedings of the IEEE Adaptive Systems for Signal Processing, Communications, and Control Symposium (AS-SPCC), pages 153–158, 2000

    work page 2000

  6. [6]

    Geir Evensen. Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics.Journal of Geophysical Research: Oceans, 99 (C5):10143–10162, 1994

    work page 1994

  7. [7]

    Analysis scheme in the ensemble Kalman filter.Monthly Weather Review, 126(6):1719–1724, 1998

    Gerrit Burgers, Peter Jan van Leeuwen, and Geir Evensen. Analysis scheme in the ensemble Kalman filter.Monthly Weather Review, 126(6):1719–1724, 1998

    work page 1998

  8. [8]

    The invariant extended Kalman filter as a stable observer

    Axel Barrau and Silvère Bonnabel. The invariant extended Kalman filter as a stable observer. IEEE Transactions on Automatic Control, 62(4):1797–1812, 2017

    work page 2017

  9. [9]

    A code for unscented Kalman filtering on manifolds (UKF-M)

    Martin Brossard, Axel Barrau, and Silvère Bonnabel. A code for unscented Kalman filtering on manifolds (UKF-M). InIEEE International Conference on Robotics and Automation (ICRA), pages 5701–5708, 2020

    work page 2020

  10. [10]

    Max entropy moment Kalman filter for polynomial systems with arbitrary noise

    Sangli Teng, Harry Zhang, David Jin, Ashkan Jasour, Ram Vasudevan, Maani Ghaffari, and Luca Carlone. Max entropy moment Kalman filter for polynomial systems with arbitrary noise. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  11. [11]

    Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    Aapo Hyvärinen. Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    work page 2005

  12. [12]

    Moment closure—a brief review

    Christian Kuehn. Moment closure—a brief review. InControl of Self-Organizing Nonlinear Systems, Understanding Complex Systems, pages 253–271. Springer, Cham, 2016. 10

    work page 2016

  13. [13]

    On the use of the normal approximation in the treatment of stochastic processes

    Peter Whittle. On the use of the normal approximation in the treatment of stochastic processes. Journal of the Royal Statistical Society, Series B, 19(2):268–281, 1957

    work page 1957

  14. [14]

    David Levermore

    C. David Levermore. Moment closure hierarchies for kinetic theories.Journal of Statistical Physics, 83(5–6):1021–1065, 1996

    work page 1996

  15. [15]

    Hespanha

    Abhyudai Singh and João P. Hespanha. Approximate moment dynamics for chemically reacting systems.IEEE Transactions on Automatic Control, 56(2):414–418, 2011

    work page 2011

  16. [16]

    Alfred Kume and Stephen G. Walker. On Stein’s method of moments and generalized score matching.arXiv preprint arXiv:2602.06482, 2026

    work page arXiv 2026

  17. [17]

    Gillespie

    Colin S. Gillespie. Moment-closure approximations for mass-action models.IET Systems Biology, 3(1):52–58, 2009

    work page 2009

  18. [18]

    On the kinetic theory of rarefied gases.Communications on Pure and Applied Mathematics, 2(4):331–407, 1949

    Harold Grad. On the kinetic theory of rarefied gases.Communications on Pure and Applied Mathematics, 2(4):331–407, 1949

    work page 1949

  19. [19]

    Kirkwood

    John G. Kirkwood. Statistical mechanics of fluid mixtures.The Journal of Chemical Physics, 3 (5):300–313, 1935

    work page 1935

  20. [20]

    Matt J. Keeling. The effects of local spatial structure on epidemiological invasions.Proceedings of the Royal Society of London B, 266(1421):859–867, 1999

    work page 1999

  21. [21]

    Kiss, Joel C

    István Z. Kiss, Joel C. Miller, and Péter L. Simon.Mathematics of Epidemics on Networks: From Exact to Approximate Models, volume 46 ofInterdisciplinary Applied Mathematics. Springer Cham, 2017

    work page 2017

  22. [22]

    Mead and N

    Lawrence R. Mead and N. Papanicolaou. Maximum entropy in the problem of moments. Journal of Mathematical Physics, 25(8):2404–2417, 1984

    work page 1984

  23. [23]

    MePoly: Max entropy polynomial policy optimiza- tion.arXiv preprint arXiv:2602.17832, 2026

    Hang Liu, Sangli Teng, and Maani Ghaffari. MePoly: Max entropy polynomial policy optimiza- tion.arXiv preprint arXiv:2602.17832, 2026

    work page arXiv 2026

  24. [24]

    The geometry of Stein’s method of moments: A canonical decomposition via score matching.arXiv preprint arXiv:2603.12843, 2026

    Mitsuki Nagai and Keisuke Yano. The geometry of Stein’s method of moments: A canonical decomposition via score matching.arXiv preprint arXiv:2603.12843, 2026

    work page arXiv 2026

  25. [25]

    Eustice, and Jessy W

    Ross Hartley, Maani Ghaffari, Ryan M. Eustice, and Jessy W. Grizzle. Contact-aided invariant extended Kalman filtering for robot state estimation.The International Journal of Robotics Research, 39(4):402–430, 2020

    work page 2020

  26. [26]

    Equivariant filter (EqF).IEEE Transactions on Automatic Control, 68(6):3501–3512, 2023

    Pieter van Goor, Tarek Hamel, and Robert Mahony. Equivariant filter (EqF).IEEE Transactions on Automatic Control, 68(6):3501–3512, 2023

    work page 2023

  27. [27]

    Williams

    Ashkan Jasour, Allen Wang, and Brian C. Williams. Moment-based exact uncertainty propa- gation through nonlinear stochastic autonomous systems.arXiv preprint arXiv:2101.12490, 2021

    work page arXiv 2021

  28. [28]

    Moment-based Kalman filter: Nonlinear Kalman filtering with exact moment propagation

    Yutaka Shimizu, Ashkan Jasour, Maani Ghaffari, and Shinpei Kato. Moment-based Kalman filter: Nonlinear Kalman filtering with exact moment propagation. InIEEE International Conference on Robotics and Automation (ICRA), pages 3948–3954. IEEE, 2023

    work page 2023

  29. [29]

    GMKF: Generalized moment Kalman filter for polynomial systems with arbitrary noise.arXiv preprint arXiv:2403.04712, 2024

    Sangli Teng, Harry Zhang, David Jin, Ashkan Jasour, Maani Ghaffari, and Luca Carlone. GMKF: Generalized moment Kalman filter for polynomial systems with arbitrary noise.arXiv preprint arXiv:2403.04712, 2024

    work page arXiv 2024

  30. [30]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021

    work page 2021

  31. [31]

    Score matching based assumed density filtering with the von Mises-Fisher distribution

    Mario Bukal, Ivan Markovi´c, and Ivan Petrovi´c. Score matching based assumed density filtering with the von Mises-Fisher distribution. In20th International Conference on Information Fusion (FUSION), 2017. 11

    work page 2017

  32. [32]

    Gaunt, Babette Picker, and Yvik Swan

    Bruno Ebner, Adrian Fischer, Robert E. Gaunt, Babette Picker, and Yvik Swan. Stein’s method of moments.Scandinavian Journal of Statistics, 52(4):1594–1624, 2025

    work page 2025

  33. [33]

    Popov, Kristen Michaelson, Felipe Giraldo-Grueso, Dalton Durant, Simone Servadio, and Uwe D

    Renato Zanetti, Andrey A. Popov, Kristen Michaelson, Felipe Giraldo-Grueso, Dalton Durant, Simone Servadio, and Uwe D. Hanebeck. A survey of nonlinear estimation filters.Journal of Advances in Information Fusion, 2025. Accepted for publication

    work page 2025

  34. [34]

    Guy Revach, Nir Shlezinger, Xiaoyong Ni, Adrià López Escoriza, Ruud J. G. van Sloun, and Yonina C. Eldar. KalmanNet: Neural network aided Kalman filtering for partially known dynamics.IEEE Transactions on Signal Processing, 70:1532–1547, 2022

    work page 2022

  35. [35]

    Wainwright and Michael I

    Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and variational inference.Foundations and Trends in Machine Learning, 1(1–2):1–305, 2008

    work page 2008

  36. [36]

    Edwin T. Jaynes. Information theory and statistical mechanics.Physical Review, 106(4): 620–630, 1957

    work page 1957

  37. [37]

    Charles Stein. A bound for the error in the normal approximation to the distribution of a sum of dependent random variables.Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability, 2:583–602, 1972

    work page 1972

  38. [38]

    Springer, Berlin, 6th edition, 2003

    Bernt Øksendal.Stochastic Differential Equations: An Introduction with Applications. Springer, Berlin, 6th edition, 2003

    work page 2003

  39. [39]

    Bar-Itzhack, and Yaakov Oshman

    Daniel Choukroun, Itzhack Y . Bar-Itzhack, and Yaakov Oshman. Novel quaternion Kalman filter.IEEE Transactions on Aerospace and Electronic Systems, 42(1):174–190, 2006

    work page 2006

  40. [40]

    Hall.Lie Groups, Lie Algebras, and Representations: An Elementary Introduction, volume 222 ofGraduate Texts in Mathematics

    Brian C. Hall.Lie Groups, Lie Algebras, and Representations: An Elementary Introduction, volume 222 ofGraduate Texts in Mathematics. Springer, 2nd edition, 2015

    work page 2015

  41. [41]

    Chirikjian.Stochastic Models, Information Theory, and Lie Groups, Volume 2: Analytic Methods and Modern Applications

    Gregory S. Chirikjian.Stochastic Models, Information Theory, and Lie Groups, Volume 2: Analytic Methods and Modern Applications. Birkhäuser, Boston, 2012

    work page 2012

  42. [42]

    Lasserre

    Jean B. Lasserre. Global optimization with polynomials and the problem of moments.SIAM Journal on Optimization, 11(3):796–817, 2001. doi: 10.1137/S1052623400366802

    work page doi:10.1137/s1052623400366802 2001

  43. [43]

    Pablo A. Parrilo. Semidefinite programming relaxations for semialgebraic problems.Mathe- matical Programming, 96(2):293–320, 2003. doi: 10.1007/s10107-003-0387-5

    work page doi:10.1007/s10107-003-0387-5 2003

  44. [44]

    Measuring sample quality with Stein’s method

    Jackson Gorham and Lester Mackey. Measuring sample quality with Stein’s method. In Advances in Neural Information Processing Systems (NeurIPS), 2015

    work page 2015

  45. [45]

    Minimum Stein discrepancy estimators

    Alessandro Barp, François-Xavier Briol, Andrew Duncan, Mark Girolami, and Lester Mackey. Minimum Stein discrepancy estimators. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    work page 2019

  46. [46]

    Gaunt, Fatemeh Ghaderinezhad, Jackson Gorham, Arthur Gretton, Christophe Ley, Qiang Liu, Lester Mackey, Chris J

    Andreas Anastasiou, Alessandro Barp, François-Xavier Briol, Bruno Ebner, Robert E. Gaunt, Fatemeh Ghaderinezhad, Jackson Gorham, Arthur Gretton, Christophe Ley, Qiang Liu, Lester Mackey, Chris J. Oates, Gesine Reinert, and Yvik Swan. Stein’s method meets computational statistics: A review of some recent developments.Statistical Science, 38(1):120–139, 2023

    work page 2023

  47. [47]

    Riemannian score-based generative modelling

    Valentin De Bortoli, Emile Mathieu, Michael Hutchinson, James Thornton, Yee Whye Teh, and Arnaud Doucet. Riemannian score-based generative modelling. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    work page 2022

  48. [48]

    Uncertainty propagation for general stochastic hybrid systems on compact Lie groups.SIAM Journal on Applied Dynamical Systems, 21(3):2215–2240, 2022

    Weixin Wang and Taeyoung Lee. Uncertainty propagation for general stochastic hybrid systems on compact Lie groups.SIAM Journal on Applied Dynamical Systems, 21(3):2215–2240, 2022

    work page 2022

  49. [49]

    René Thom.Structural Stability and Morphogenesis. W. A. Benjamin, 1975

    work page 1975

  50. [50]

    R. E. Kalman. Contributions to the theory of optimal control.Boletín de la Sociedad Matemática Mexicana, 5(2):102–119, 1960

    work page 1960

  51. [51]

    D. Q. Mayne, J. B. Rawlings, C. V . Rao, and P. O. M. Scokaert. Constrained model predictive control: Stability and optimality.Automatica, 36(6):789–814, 2000. 12

    work page 2000

  52. [52]

    Ames, Xiangru Xu, Jessy W

    Aaron D. Ames, Xiangru Xu, Jessy W. Grizzle, and Paulo Tabuada. Control barrier function based quadratic programs for safety critical systems.IEEE Transactions on Automatic Control, 62(8):3861–3876, 2017. 13 A Background on exponential families and maximum entropy This section collects standard results on exponential families [ 35, 36] that are used in ou...

    work page 2017

  53. [53]

    = 1 2P = Ω 2 .(A.27) Hence, score matching recovers the information vector η and the precision matrix Ω exactly from the first two moments. 20 D.4 Prediction step reduces to Riccati For affine dynamics dx= (Ax+b)dt+h dW with H= 1 2 hh⊤ constant, the score PDE (A.24) with S=−Ω , T= 0 , JX =A , ∇(∇ ·X) = 0 reduces to ∂ts=−A ⊤s+ Ω(Ax+b)−2ΩHs . Substitutings=...

    work page

  54. [54]

    This is the unclosed term that the Stein closure (Section 4) resolves

    + σ2 2 b(b−1)x a 1xb−2 2 =a x a−1 1 xb+1 2 −bδ x a 1xb 2 −bα x a+1 1 xb−1 2 −bβ x a+2 1 xb−1 2 + σ2 2 b(b−1)x a 1xb−2 2 .(A.35) All terms have degree |α|=a+b except xa+2 1 xb−1 2 , which has degree |α|+ 1 . This is the unclosed term that the Stein closure (Section 4) resolves. 23 G.2 Lotka-Volterra (stochastic predator-prey) SDE.Statex= (x 1, x2)(prey, pr...

    work page

  55. [55]

    This is the highest excess degree among all experimental systems and requirestwoStein closure layers

    = 2. This is the highest excess degree among all experimental systems and requirestwoStein closure layers. Generator.Applied toϕ α =x a 1xb 2: Aϕα =a(x 1 −x 3 1)x a−1 1 xb 2 +b(x 2 −x 3 2)x a 1xb−1 2 + σ2 2 a(a−1)x a−2 1 xb 2 +b(b−1)x a 1xb−2 2 .(A.45) Taking expectations: ˙mα =a m α −a m α+2e1 +b m α −b m α+2e2 + σ2 2 a(a−1)m α−2e1 +b(b−1)m α−2e2 .(A.46)...

    work page

  56. [56]

    G.6 SE(2) kinematics SDE.State x= (c, s, p x, py) with c= cosθ , s= sinθ , c2 +s 2 = 1

    that couple to the closure at degreeK. G.6 SE(2) kinematics SDE.State x= (c, s, p x, py) with c= cosθ , s= sinθ , c2 +s 2 = 1 . The SDE in embedded coordinates is dc= −ωs− σ2 2 c dt−σs dW, ds= ωc− σ2 2 s dt+σc dW, dpx =vc dt, dp y =vs dt,(A.50) whereωis the angular velocity,vis the forward speed, andσis the heading noise intensity. Drift and diffusion cla...

    work page

  57. [57]

    Moment propagation via Dynkin operates on monomial moments (the generator acts naturally on monomials)

    work page

  58. [58]

    Before score matching: convert monomial moments to the chosen basis using (A.73), or equivalently assemble A, b directly via (A.68)–(A.69) in centered/scaled monomial coordinates (which approximate orthogonality)

    work page

  59. [59]

    For Stein closure during propagation, convert to monomialλvia (A.72) if needed

    After score matching: λ is in the working basis. For Stein closure during propagation, convert to monomialλvia (A.72) if needed

    work page

  60. [60]

    extended dynamical system

    For measurement update: the conjugate addition λ+ =λ − +λ lik is performed in whichever basis both are expressed in. H.4 Centered coordinate transformation Given raw moments mα =E[x α] and mean µi =m ei, the centered moments mz α =E[(x−µ) α] are obtained via the multinomial expansion: mz α = X β≤α α β (−µ)α−β mβ,(A.76) where α β =Q i αi βi and µα−β =Q i µ...

    work page arXiv 2079

  61. [61]

    well-posedness

    produces a parabolic plume that the EKF (Gaussian, symmetric ellipse) cannot represent. SKF propagates 286 centered moments (K=10) and reconstructs the non-Gaussian marginal, matching the 500k-particle MC reference. Moment accuracy: E[z2 3] within 1.3% and E[z3 3] within 3.2% of MC at T=3 (Appendix J.2). Timing: SKF 8 s, MC (500k) 122s, EKF<0.1s. SE(2) lo...

    work page