pith. sign in

arxiv: 2407.11518 · v2 · pith:BYCN7LRTnew · submitted 2024-07-16 · 📊 stat.ML · cs.LG· stat.OT

Ensemble Transport Filter via Optimized Maximum Mean Discrepancy

Dengfei Zeng , Lijian Jiang This is my paper

Pith reviewed 2026-05-23 22:58 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.OT
keywords ensemble filtertransport mapmaximum mean discrepancyparticle filterdata assimilationvariance penaltyhigh-dimensional problemsposterior approximation
0
0 comments X

The pith

A transport map optimized by maximum mean discrepancy with a variance penalty reconstructs the particle filter analysis step for high-dimensional assimilation.

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

The paper seeks to rebuild the analysis step of particle filtering as a direct transport of prior particles to posterior particles. It does this by solving an optimization problem whose loss is the maximum mean discrepancy between the transported particles and a reference posterior, plus a variance penalty term. The goal is to keep the accurate posterior estimates that particle filters provide while making the method usable in high-dimensional settings where standard particle filters break down. If the optimization succeeds, the resulting filter would match informative statistics more reliably than plain maximum mean discrepancy and outperform the ensemble Kalman filter on the tested assimilation tasks.

Core claim

The analysis step of the particle filter is recast as an optimization problem that finds a transport map minimizing the maximum mean discrepancy between the empirical distribution of transported prior particles and the reference posterior, with an added variance penalty that emphasizes highly informative statistics; the resulting map produces posterior particles that inherit the accuracy of particle filtering while remaining tractable in high dimensions.

What carries the argument

The transport map obtained by minimizing maximum mean discrepancy loss augmented with a variance penalty term, which aligns expectation information of the approximated posterior with that of the reference posterior.

If this is right

  • The method retains the accurate posterior estimation property of particle filtering.
  • The variance penalty improves robustness by guiding the map toward highly informative statistics.
  • The approach extends particle filtering to high-dimensional assimilation problems.
  • Numerical examples demonstrate better performance than the ensemble Kalman filter.

Where Pith is reading between the lines

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

  • The same optimized-transport construction could be applied to other sequential estimation tasks where particle methods are accurate but dimensionally limited.
  • Replacing the variance penalty with alternative regularizers might further stabilize the optimization in even higher dimensions.
  • The explicit focus on matching expectations of informative statistics suggests the method could integrate with dimension-reduction techniques used in ensemble methods.

Load-bearing premise

The optimization problem using the maximum mean discrepancy loss function with variance penalty can be reliably solved to produce a transport map that accurately approximates the posterior distribution from prior particles in high-dimensional settings.

What would settle it

Numerical experiments in a high-dimensional assimilation problem where the posterior particles produced by the optimized transport map show large discrepancy from the reference posterior or yield worse state estimates than the ensemble Kalman filter.

Figures

Figures reproduced from arXiv: 2407.11518 by Dengfei Zeng, Lijian Jiang.

Figure 1
Figure 1. Figure 1: Directed probability graph of dynamical model (2.1) and observation operator (2.2) in the form of state-space model 2.2. Sequential Bayesian Filtering. Define the sequence of model states and obser￾vations by x0:K = {x0, x1, · · · , xK} and y0:K = {y1 , y2 , · · · , yK}, respectively. The procedure of sequential Bayesian filtering is to estimate the πXk|Y 1:k iteratively. As￾suming the prior of the initial… view at source ↗
Figure 2
Figure 2. Figure 2: Propagation of particles between distributions in ensemble-based filtering For the convenience of presentation, we denote Xˆ k as the random varible of forecast distribution πXk|Y 1:k−1 . From the point of view of transport, the ensemble-based filter has two transport processes of distribution, which are denoted as push-forward operators M♯ and L♯ shown in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Posterior approximation of EnTranF with Linear transport map (Panel (a)-(c)) and Nonlinear transport map (Panel (d)-(f )) when taking different bandwidths of MMD loss. The example considered the influence of bandwidth at different scales (10−3 , 100 and 103 ) on EnTranF. We select Gaussian kernel for MMD loss with three different bandwidth values 10−3 , 100 , and 103 . Linear and Nonlinear transport maps a… view at source ↗
Figure 4
Figure 4. Figure 4: 2-d Histogram of ensemble particles generated by EnKF, EnTranF and EnTranFp. Red ”x” represents the MAP estimation of the corresponding method. model and observation operator are given by dx = (x − x 3 )dt + γdWt (4.4) , z = 0.1x 2 (4.5) + sin(x) + ϵ, where Wt is a standard brownian motion and ϵ is Gaussian noise with ϵ ∼ N (0, σ2 ). We take γ = 0.8 and σ = 0.5. In this example, the fourth-order Runge-Kutt… view at source ↗
Figure 5
Figure 5. Figure 5: compares the average RMSE between EnTranF and EnKF with different observation intervals and ensemble sizes. 0.1 0.2 0.3 0.4 0.5 Observation Interval tob 2.8 3.0 3.2 3.4 3.6 3.8 4.0 4.2 Average RMSE 1e 1 EnTranF(L-L) EnTranF(N-L) EnTranF(L-G) EnTranF(N-G) EnKF 10 406080100 200 300 400 Ensemble Size N 3.8 3.9 4.0 4.1 4.2 4.3 Average RMSE 1e 1 EnTranF(L-L) EnTranF(N-L) EnTranF(L-G) EnTranF(N-G) EnKF [PITH_FU… view at source ↗
Figure 6
Figure 6. Figure 6: Take posterior mean of PF with 10,000 particles as reference, Average RMSE of Double-Well system for ensemble size N = 400 as a function of observation interval (left), and for observation interval ∆tob = 0.5 as a function of ensemble size (rightl). panel of [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Left Panel: Spread of Double-Well system for observation interval ∆tob = 0.5 as a function of ensemble size. Right panel: Average coverage probability of Double-Well system for observation interval ∆tob = 0.5 as a function of ensemble size 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: EnTranF with variance penalty: Average RMSE of Double-Well system for ensemble size N = 400 as a function of observation interval (Panel (a)), and for observation interval ∆tob = 0.5 as a function of ensemble size (Panel (b)). Spread of Double-Well system for observation interval ∆tob = 0.5 as a function of ensemble size (Panel (c)). Average coverage probability of Double-Well system for observation interv… view at source ↗
Figure 9
Figure 9. Figure 9: Average RMSE of Lorenz’63 system for ensemble size N = 400 as a function of observation interval (Left), and for observation interval ∆tob = 0.5 as a function of ensemble size (Right). EnTranF(L-G) have 25.30% and 18.63% improvement on RMSE, respectively, when tak￾ing observation interval as ∆tob = 0.5. A well-chosen kernel function can make EnTranF more robust in systems with strong nonlinearity. The righ… view at source ↗
Figure 10
Figure 10. Figure 10: Spread of Lorenz’63 system for ensemble size N = 400 as a function of observation interval (Left), and for observation interval ∆tob = 0.5 as a function of ensemble size (Right). 0.1 0.2 0.3 0.4 0.5 Observation Interval tob 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 Average RMSE ETranFp(L-L) ETranFp(N-L) ETranFp(L-G) ETranFp(N-G) EnKF 10 406080100 200 300 400 Ensemble Size N 2.0 2.5 3.0 3.5 4.0 4.5… view at source ↗
Figure 11
Figure 11. Figure 11: EnTranF with variance penalty: Average RMSE of Lorenz’63 system for ensemble size N = 400 as a function of observation interval (Left), and for observation interval ∆tob = 0.5 as a function of ensemble size (Right). tion. Here, EnTranFp(L-G) is 21.07% better than EnKF on RMSE at observation inter￾val ∆tob = 0.5, while EnTranFp(N-G) is 27.39% better than EnKF. From panel (b) of [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 12
Figure 12. Figure 12: EnTranF with variance penalty: Spread of Lorenz’63 system for ensemble size N = 400 as a function of observation interval (Left), and for observation interval ∆tob = 0.5 as a function of ensemble size (Right). 4.4. Lorenz’96 System. The Lorenz96 model is a simplified mathematical model that describes the nonlinear dynamical behavior in atmospheric circulation systems. It consists of a set of coupled one-d… view at source ↗
Figure 13
Figure 13. Figure 13: State estimation and absolute error of Lorenz’96 system for observation interval ∆tob = 0.1 and ensemble size N = 2000. 0 100 200 300 400 500 600 700 800 Time 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 RMSE EnKF EnTranF (a) RMSE 0 100 200 300 400 500 600 700 800 Time 0.70 0.75 0.80 0.85 0.90 0.95 Spread EnKF EnTranF (b) Ensemble Spread [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: RMSE and Spread of Lorenz96 system for observation interval ∆tob = 0.1 and ensemble size N = 2000. ensemble transport filter inherited the accurate estimation of the posterior distribution from the particle filter. To improve the robustness of MMD, we introduced a variance penalty term to guide the prioritized optimization of high-informative statistics in the 25 [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
read the original abstract

In this paper, we present a new ensemble-based filter method by reconstructing the analysis step of the particle filter through a transport map, which directly transports prior particles to posterior particles. The transport map is constructed through an optimization problem described by the Maximum Mean Discrepancy loss function, which matches the expectation information of the approximated posterior and reference posterior. The proposed method inherits the accurate estimation of the posterior distribution from particle filtering while gives an extension to high dimensional assimilation problems. To improve the robustness of Maximum Mean Discrepancy, a variance penalty term is used to guide the optimization. It prioritizes minimizing the discrepancy between the expectations of highly informative statistics for the reference posteriors. The penalty term significantly enhances the robustness of the proposed method and leads to a better approximation of the posterior. A few numerical examples are presented to illustrate the advantage of the proposed method over ensemble Kalman filter.

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

1 major / 2 minor

Summary. The manuscript proposes an Ensemble Transport Filter that reconstructs the analysis step of the particle filter via a transport map optimized to minimize a Maximum Mean Discrepancy (MMD) loss between the mapped prior ensemble and a reference posterior, augmented by a variance penalty term for robustness. The method is presented as inheriting the accuracy of particle filtering while extending applicability to high-dimensional assimilation problems, with numerical examples claimed to show advantages over the ensemble Kalman filter.

Significance. If the optimization of the transport map via MMD plus variance penalty can be shown to produce accurate posterior approximations reliably, the approach would offer a useful ensemble method for nonlinear and non-Gaussian data assimilation. The reported numerical examples on low-to-moderate dimensional problems indicate that the variance penalty improves stability relative to plain MMD, which is a concrete strength of the work.

major comments (1)
  1. [Numerical examples] Numerical examples section: the claim that the method 'gives an extension to high dimensional assimilation problems' is not supported by the presented experiments, which are restricted to low-to-moderate dimensional test problems. This directly affects the central claim of broader applicability beyond standard particle filters.
minor comments (2)
  1. [Abstract] Abstract: the variance penalty coefficient is introduced without discussion of its selection or sensitivity; explicit guidance would improve reproducibility.
  2. [Method] Method description: clarify how the 'reference posterior' is constructed in the numerical examples, as this is central to the MMD objective.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below regarding the numerical examples and associated claims.

read point-by-point responses

  1. Referee: [Numerical examples] Numerical examples section: the claim that the method 'gives an extension to high dimensional assimilation problems' is not supported by the presented experiments, which are restricted to low-to-moderate dimensional test problems. This directly affects the central claim of broader applicability beyond standard particle filters.

    Authors: We agree that the numerical experiments are restricted to low-to-moderate dimensions and do not empirically demonstrate performance in truly high-dimensional regimes. The manuscript's claim of extension to high-dimensional assimilation is motivated by the formulation (transport map optimization without importance weights, thereby sidestepping the degeneracy that limits standard particle filters), but this remains a theoretical motivation rather than a validated result. To align the claims with the evidence, we will revise the abstract, introduction, and conclusions to state that the approach offers a framework with potential applicability to higher-dimensional problems, while explicitly noting that current validation is limited to moderate-dimensional test cases. We will also add a remark on the need for future high-dimensional benchmarks. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's central construction defines a transport map via direct optimization of an MMD loss plus variance penalty term applied to prior and reference posterior particles. This is an explicit algorithmic definition built from standard MMD properties and ensemble filtering concepts, without any reduction of outputs to fitted inputs by construction, self-definitional loops, or load-bearing self-citations. No equations or steps in the provided description equate a claimed prediction to its own inputs. The derivation remains self-contained against external benchmarks such as known MMD definitions and particle filter properties.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach relies on standard assumptions from optimal transport and kernel methods for using MMD as a discrepancy measure; no free parameters or invented entities are explicitly detailed in the abstract.

free parameters (1)
  • variance penalty coefficient
    The weight balancing the variance penalty term against the MMD loss is a tunable hyperparameter required for the optimization.
axioms (1)
  • domain assumption Maximum mean discrepancy provides a suitable metric for matching expectations between approximated and reference posteriors in the transport optimization.
    Invoked as the core loss function for constructing the transport map.

pith-pipeline@v0.9.0 · 5675 in / 1152 out tokens · 24268 ms · 2026-05-23T22:58:54.129571+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

34 extracted references · 34 canonical work pages

  1. [1]

    Brajard, A

    J. Brajard, A. Carrassi, M. Bocquet, and L. Bertino , Combining data assimilation and machine learning to infer unresolved scale parametrization, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 379 (2021), p. 20200086

    work page 2021

  2. [2]

    Carrassi, M

    A. Carrassi, M. Bocquet, L. Bertino, and G. Evensen , Data assimilation in the geosciences: An overview of methods, issues, and perspectives , WIREs Climate Change, 9 (2018), p. e535

    work page 2018

  3. [3]

    Chattopadhyay, E

    A. Chattopadhyay, E. Nabizadeh, E. Bach, and P. Hassanzadeh , Deep learning-enhanced ensemble-based data assimilation for high-dimensional nonlinear dynamical systems , Journal of Com- putational Physics, 477 (2023), p. 111918

    work page 2023

  4. [4]

    A. J. Chorin and X. Tu , Implicit sampling for particle filters , Proceedings of the National Academy of Sciences, 106 (2009), pp. 17249–17254

    work page 2009

  5. [5]

    T. A. El Moselhy and Y. M. Marzouk , Bayesian inference with optimal maps , Journal of Computa- tional Physics, 231 (2012), pp. 7815–7850

    work page 2012

  6. [6]

    Evensen , The ensemble kalman filter: Theoretical formulation and practical implementation , Ocean Dynamics, 53 (2003), pp

    G. Evensen , The ensemble kalman filter: Theoretical formulation and practical implementation , Ocean Dynamics, 53 (2003), pp. 343–367

    work page 2003

  7. [7]

    Evensen , Sampling strategies and square root analysis schemes for the enkf , Ocean Dynamics, 54 (2004), pp

    G. Evensen , Sampling strategies and square root analysis schemes for the enkf , Ocean Dynamics, 54 (2004), pp. 539–560

    work page 2004

  8. [8]

    Evensen , The ensemble kalman filter for combined state and parameter estimation , IEEE Control Systems, 29 (2009), pp

    G. Evensen , The ensemble kalman filter for combined state and parameter estimation , IEEE Control Systems, 29 (2009), pp. 83–104

    work page 2009

  9. [9]

    F archi and M

    A. F archi and M. Bocquet , Review article: Comparison of local particle filters and new implementa- tions, Nonlinear Processes in Geophysics, 25 (2018), pp. 765–807

    work page 2018

  10. [10]

    F archi, P

    A. F archi, P. Laloyaux, M. Bonavita, and M. Bocquet , Using machine learning to correct model error in data assimilation and forecast applications , Quarterly Journal of the Royal Meteorological Society, 147 (2021), pp. 3067–3084

    work page 2021

  11. [11]

    Gretton, K

    A. Gretton, K. Borgwardt, M. Rasch, B. Sch ¨olkopf, and A. Smola , A kernel method for the two-sample-problem, in Advances in Neural Information Processing Systems, vol. 19, MIT Press, 2006

    work page 2006

  12. [12]

    D. J. Higham. , An algorithmic introduction to numerical simulation of stochastic differential equations , SIAM Review, 43 (2001), pp. 525–546

    work page 2001

  13. [13]

    Hoang, S

    T.-V. Hoang, S. Krumscheid, H. G. Matthies, and R. Tempone , Machine learning-based condi- tional mean filter: A generalization of the ensemble kalman filter for nonlinear data assimilation , Foundations of Data Science, 5 (2023), pp. 56–80

    work page 2023

  14. [14]

    Jiang and N

    L. Jiang and N. Liu , Correcting noisy dynamic mode decomposition with kalman filters , Journal of Computational Physics, 461 (2022), p. 111175

    work page 2022

  15. [15]

    Kawabata and G

    T. Kawabata and G. Ueno, Non-gaussian probability densities of convection initiation and development investigated using a particle filter with a storm-scale numerical weather prediction model , Monthly Weather Review, 148 (2020), pp. 3–20. 26

    work page 2020

  16. [16]

    K. Law, A. Stuart, and K. Zygalakis , Data Assimilation: A Mathematical Introduction , no. volume 62 in Texts in Applied Mathematics, Springer, Cham Heidelberg New York Dordrecht London, 2015

    work page 2015

  17. [17]

    R. S. Liptser and A. N. Shiryaev , Statistics of Random Processes II: Applications, vol. 6 of Stochastic Modelling and Applied Probability, Springer Berlin Heidelberg, Berlin, Heidelberg, 2001

    work page 2001

  18. [18]

    R. S. Liptser and A. N. Shiryayev , Statistics of Random Processes I: General Theory , Springer New York, New York, NY, 1977

    work page 1977

  19. [19]

    Liu and L

    N. Liu and L. Jiang , Perron–frobenius operator filter for stochastic dynamical systems , SIAM/ASA Journal on Uncertainty Quantification, 12 (2024), pp. 182–211

    work page 2024

  20. [20]

    A. J. Majda and J. Harlim , Filtering Complex Turbulent Systems , Cambridge University Press, 1 ed., Feb. 2012

    work page 2012

  21. [21]

    Mandel, L

    J. Mandel, L. Cobb, and J. D. Beezley , On the convergence of the ensemble kalman filter , Applica- tions of Mathematics, 56 (2011), pp. 533–541

    work page 2011

  22. [22]

    Monge , M´ emoire sur la th´ eorie des d´ eblais et des remblais, Mem

    G. Monge , M´ emoire sur la th´ eorie des d´ eblais et des remblais, Mem. Math. Phys. Acad. Royale Sci., (1781), pp. 666–704

    work page

  23. [23]

    Pulido and P

    M. Pulido and P. J. V an Leeuwen , Sequential monte carlo with kernel embedded mappings: The mapping particle filter, Journal of Computational Physics, 396 (2019), pp. 400–415

    work page 2019

  24. [24]

    Ramdas, S

    A. Ramdas, S. Jakkam Reddi, B. Poczos, A. Singh, and L. W asserman , On the decreasing power of kernel and distance based nonparametric hypothesis tests in high dimensions , Proceedings of the AAAI Conference on Artificial Intelligence, 29 (2015)

    work page 2015

  25. [25]

    Rebeschini and R

    P. Rebeschini and R. V an Handel , Can local particle filters beat the curse of dimensionality? , The Annals of Applied Probability, 25 (2015)

    work page 2015

  26. [26]

    Richard M , Real Analysis and Probability, Chapman and Hall/CRC, 2002

    D. Richard M , Real Analysis and Probability, Chapman and Hall/CRC, 2002

    work page 2002

  27. [27]

    Sakov, D

    P. Sakov, D. S. Oliver, and L. Bertino , An iterative enkf for strongly nonlinear systems , Monthly Weather Review, 140 (2012), pp. 1988–2004

    work page 2012

  28. [28]

    Spantini, R

    A. Spantini, R. Baptista, and Y. Marzouk , Coupling techniques for nonlinear ensemble filtering , SIAM Review, 64 (2022), pp. 921–953

    work page 2022

  29. [29]

    P. J. V an Leeuwen , Particle filtering in geophysical systems , Monthly Weather Review, 137 (2009), pp. 4089–4114

    work page 2009

  30. [30]

    Villani , Optimal Transport:Old and New , vol

    C. Villani , Optimal Transport:Old and New , vol. 338 of Grundlehren Der Mathematischen Wis- senschaften, Springer Berlin Heidelberg, Berlin, Heidelberg, 2009

    work page 2009

  31. [31]

    W ang, W

    Z. W ang, W. Xing, R. Kirby, and S. Zhe , Physics informed deep kernel learning , in Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, PMLR, May 2022, pp. 1206–1218

    work page 2022

  32. [32]

    H. Yan, Z. Li, Q. W ang, P. Li, Y. Xu, and W. Zuo , Weighted and class-specific maximum mean dis- crepancy for unsupervised domain adaptation, IEEE Transactions on Multimedia, 22 (2020), pp. 2420– 2433

    work page 2020

  33. [33]

    Zhang, Z

    S. Zhang, Z. Liu, X. Zhang, X. Wu, G. Han, Y. Zhao, X. Yu, C. Liu, Y. Liu, S. Wu, F. Lu, M. Li, and X. Deng , Coupled data assimilation and parameter estimation in coupled ocean–atmosphere models: A review , Climate Dynamics, 54 (2020), pp. 5127–5144

    work page 2020

  34. [34]

    M. Zhu, P. J. V an Leeuwen, and J. Amezcua , Implicit equal-weights particle filter, Quarterly Journal of the Royal Meteorological Society, 142 (2016), pp. 1904–1919. 27

    work page 2016