pith. sign in

arxiv: 2603.06752 · v2 · submitted 2026-03-06 · 💻 cs.LG · cs.NA· math.NA· stat.ME· stat.ML

Latent Autoencoder Ensemble Kalman Filter for Nonlinear Data assimilation

Xin T. Tong , Yanyan Wang , Liang Yan This is my paper

Pith reviewed 2026-05-15 15:11 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NAstat.MEstat.ML
keywords data assimilationensemble Kalman filterautoencoderlatent spacenonlinear dynamicsmachine learningstate estimation
0
0 comments X

The pith

Reformulating data assimilation in a learned latent space with linear stable dynamics allows the ensemble Kalman filter to accurately handle nonlinear systems.

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

The paper introduces the latent autoencoder ensemble Kalman filter to overcome limitations of the standard ensemble Kalman filter in strongly nonlinear systems. It learns a nonlinear encoder and decoder along with a stable linear evolution operator in the latent space, creating a linear state-space model where Kalman updates can be applied directly. This approach aims to provide more accurate and stable assimilation results without significantly increasing computational cost. A reader would care because accurate data assimilation is essential for reliable predictions in fields like meteorology and oceanography where systems exhibit chaotic nonlinear behavior.

Core claim

The LAE-EnKF learns a nonlinear encoder-decoder together with a stable linear latent evolution operator and a consistent latent observation mapping, yielding a closed linear state-space model in the latent coordinates. This restores compatibility with the Kalman filtering framework, allowing both forecast and analysis steps to be carried out entirely in the latent space. Theoretical analysis establishes generalization error bounds, and experiments on nonlinear and chaotic systems show improved accuracy and stability over standard methods.

What carries the argument

The latent autoencoder that maps the high-dimensional state to a low-dimensional space where dynamics are linear and stable, enabling direct application of the ensemble Kalman filter.

If this is right

  • Forecast and analysis steps are performed entirely in latent space using linear Kalman updates.
  • The method maintains comparable computational cost to the standard EnKF.
  • It provides theoretical generalization error bounds for the latent model.
  • Demonstrates superior performance on representative nonlinear and chaotic systems.

Where Pith is reading between the lines

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

  • This framework could be extended to other data assimilation techniques beyond EnKF, such as variational methods.
  • The emphasis on stability in the latent operator may prevent filter divergence in long assimilation windows.
  • Connections to manifold learning suggest potential applications in reduced-order modeling for control systems.

Load-bearing premise

That a nonlinear encoder-decoder can be trained such that the induced latent dynamics are accurately captured by a single stable linear operator and a consistent linear observation map for the full range of states encountered in assimilation.

What would settle it

A numerical experiment on a chaotic system like the Lorenz attractor where the LAE-EnKF produces higher assimilation errors than the standard EnKF would falsify the claim of improved performance.

Figures

Figures reproduced from arXiv: 2603.06752 by Liang Yan, Xin T. Tong, Yanyan Wang.

Figure 1
Figure 1. Figure 1: The architecture of latent autoencoder framework. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Learned latent representations for latent dimensions [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pairwise latent coordinate plots (zi , zj ) for n = 3 learned by the LAE [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Long-term prediction relative error for latent dimensions [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of the relative RMSE over discrete time steps for different methods with latent dimension n = 2. Solid curves show the mean over 10 independent runs, and shaded regions indicate the corresponding 95% confidence intervals. smoother and more coherent state evolution. These results demonstrate that enforcing linear latent dynamics improves both predictive stability and assimilation robustness, direc… view at source ↗
Figure 6
Figure 6. Figure 6: Reconstructed state trajectories in two representative coordinate pairs ( [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: True solution fields of the advection–diffusion–reaction equation at the initial time (Left) and final time T = 3 (Right). White circles indicate the observation locations. To better understand the role of different components in the proposed framework, we first conduct an ablation study on the loss function (3.6). Here, we fix the latent dimension to n = 20 and remove either the latent consistency term or… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study of the loss function (3.6) with latent dimension n = 20. Left: relative state prediction error ∥DAkE (x0)−xk∥2 ∥xk∥2 . Right: relative latent consistency error ∥AkE (x0)−E (xk)∥2 ∥E (xk)∥2 . 0 1 2 3 Time 10 1 10 0 0 1 2 3 Time 10 1 10 0 0 1 2 3 Time 10 1 10 0 EnKF AE-EnKF DAE-EnKF LAE-EnKF n = 16 n = 20 n = 24 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Time evolution of the relative RMSE for Example 5.2 under latent dimensions n = 16, 20, 24. Solid curves indicate the mean over 10 independent runs, and the shaded regions represent the corresponding 95% confidence intervals. produced by LAE-EnKF at the final time T, together with the corresponding pointwise absolute errors, for different latent dimensions [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Reconstructed fields (Top) and corresponding pointwise absolute errors (Bottom) at the final time T = 3 by LAE-EnKF for different latent dimensions n = 16, n = 20, and n = 24. 1.20 0.95 0.70 0.0 0.1 0.2 EnKF AE-EnKF DAE-EnKF LAE-EnKF [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of reconstructed fields (Top) and corresponding pointwise absolute errors (Bottom) at the final time T = 3 for latent dimension n = 20. nonlinear PDE systems. 5.3. Lorenz–96 model. We consider the chaotic Lorenz–96 system, (5.8) dxi dt = (xi+1 − xi−2) xi−1 − xi + F, i = 0, . . . , D − 1, with periodic boundary conditions xD ≡ x0, x−1 ≡ xD−1, and x−2 ≡ xD−2. In all experiments, we set D = 40 and… view at source ↗
Figure 12
Figure 12. Figure 12: Time evolution of the relative RMSE for Example 5.2 under different latent observation operators: identity operator (Left), randomly operator (Middle), and learned operator (Right). Solid curves denote the mean over 10 independent runs, and shaded regions indicate the corresponding 95% confidence intervals. The global relative error ERel,1:K is reported in each subplot [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 13
Figure 13. Figure 13: Reconstructed spatial state fields (Left) and corresponding error (Right) with [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Time evolution of a representative observed variable (Left) and an unobserved [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Reconstructed spatial state fields (Left) and corresponding error (Right) with [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Time evolution of a representative observed variable (Left) and an unobserved [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Reconstruction of representative state components under nonlinear observa￾tions. Comparison between localized EnKF and LAE-EnKF under different latent observation operators, including the identity, random, and learned operators. The global RMSE ERel,1:K is reported above each panel. the proposed approach constructs a latent representation in which the assumptions underlying Kalman filtering are approximat… view at source ↗
read the original abstract

The ensemble Kalman filter (EnKF) is widely used for data assimilation in high-dimensional systems, but its performance often deteriorates for strongly nonlinear dynamics due to the structural mismatch between the Kalman update and the underlying system behavior. In this work, we propose a latent autoencoder ensemble Kalman filter (LAE-EnKF) that addresses this limitation by reformulating the assimilation problem in a learned latent space with linear and stable dynamics. The proposed method learns a nonlinear encoder--decoder together with a stable linear latent evolution operator and a consistent latent observation mapping, yielding a closed linear state-space model in the latent coordinates. This construction restores compatibility with the Kalman filtering framework and allows both forecast and analysis steps to be carried out entirely in the latent space. Compared with existing autoencoder-based and latent assimilation approaches that rely on unconstrained nonlinear latent dynamics, the proposed formulation emphasizes structural consistency, stability, and interpretability. We provide a theoretical analysis of learning linear dynamics on low-dimensional manifolds and establish generalization error bounds for the proposed latent model. Numerical experiments on representative nonlinear and chaotic systems demonstrate that the LAE-EnKF yields more accurate and stable assimilation than the standard EnKF and related latent-space methods, while maintaining comparable computational cost and data-driven.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript proposes the Latent Autoencoder Ensemble Kalman Filter (LAE-EnKF) for nonlinear data assimilation. It learns a nonlinear encoder-decoder together with a stable linear latent evolution operator and consistent linear latent observation map, enabling the full forecast-analysis cycle to be performed in latent space. The paper asserts a theoretical analysis of learning linear dynamics on low-dimensional manifolds with generalization error bounds, and reports superior accuracy and stability over standard EnKF and other latent methods on nonlinear and chaotic systems at comparable cost.

Significance. If the stability of the fixed linear latent operator and the generalization bounds hold under the distribution shifts of the EnKF cycle, the approach could meaningfully extend Kalman filtering to strongly nonlinear regimes while preserving interpretability and computational efficiency. The emphasis on structural consistency (linear stable dynamics plus linear observations) distinguishes it from unconstrained latent-dynamics methods and could influence data-assimilation practice in chaotic systems.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (Method): The central construction assumes a single fixed stable linear operator accurately captures latent dynamics for all states visited during the EnKF forecast-analysis cycle. No analysis is given of how analysis-step distribution shifts in chaotic systems affect the validity of this fixed operator, which directly undermines the claimed stability and accuracy gains.
  2. [Theoretical analysis section] Theoretical analysis section: The stated generalization error bounds for learning linear dynamics on low-dimensional manifolds do not address propagation of approximation error through repeated EnKF iterations; without this, the bounds do not support the headline performance claim for chaotic systems.
minor comments (1)
  1. [Abstract] Abstract, final sentence: the phrase ends abruptly with 'and data-driven.' and appears truncated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below, clarifying the scope of our claims and indicating the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (Method): The central construction assumes a single fixed stable linear operator accurately captures latent dynamics for all states visited during the EnKF forecast-analysis cycle. No analysis is given of how analysis-step distribution shifts in chaotic systems affect the validity of this fixed operator, which directly undermines the claimed stability and accuracy gains.

    Authors: The latent model is trained on long trajectories sampled from the full nonlinear dynamics, which include states visited during typical EnKF forecast-analysis cycles. The enforced stability of the linear operator (via spectral radius constraint) is intended to provide robustness against moderate shifts. We acknowledge that the manuscript lacks an explicit discussion of analysis-induced distribution shifts. In the revision we will add a dedicated paragraph in Section 3 describing the training distribution relative to the assimilation loop and include new numerical experiments that vary analysis frequency and ensemble size to quantify sensitivity to such shifts. revision: partial

  2. Referee: [Theoretical analysis section] Theoretical analysis section: The stated generalization error bounds for learning linear dynamics on low-dimensional manifolds do not address propagation of approximation error through repeated EnKF iterations; without this, the bounds do not support the headline performance claim for chaotic systems.

    Authors: The generalization bounds are derived for one-step prediction error under the training distribution on the manifold and do not claim to control accumulated error over multiple closed-loop EnKF steps. The stability constraint limits exponential growth of perturbations, which is consistent with the observed performance on chaotic test cases. We agree that the current presentation does not explicitly address iterative error propagation. In the revised theoretical section we will clarify the scope of the bounds, add a remark on the role of the spectral radius in controlling multi-step error, and support this with a short numerical study of error growth under repeated application. revision: partial

Circularity Check

0 steps flagged

No circularity detected; latent linear model learned independently of assimilation loop

full rationale

The derivation chain consists of training a nonlinear encoder-decoder jointly with a stable linear latent evolution operator and linear observation map to produce a closed linear state-space model, followed by standard Kalman filtering performed entirely in that latent space. No equation reduces a claimed prediction to a fitted parameter by construction, no uniqueness theorem is imported via self-citation, and no ansatz is smuggled through prior work. The performance claims rest on separate numerical experiments on chaotic systems and stated generalization bounds for the latent model; these are external to the training procedure itself and do not collapse into the inputs by definition. The construction is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the existence of a low-dimensional manifold on which the nonlinear dynamics can be well approximated by a single stable linear operator after nonlinear encoding; this is an empirical modeling assumption rather than a derived property.

free parameters (2)
  • latent dimension
    Chosen to balance reconstruction fidelity and linear dynamics fit; value not stated in abstract.
  • stability margin / regularization weight
    Hyper-parameter that enforces stability of the learned linear operator during training.
axioms (2)
  • domain assumption The nonlinear system admits a low-dimensional manifold on which dynamics are approximately linear after a suitable nonlinear coordinate change.
    Invoked when the authors state that the latent model yields a closed linear state-space representation.
  • standard math Standard Kalman filter update equations remain optimal once the state and observation are expressed in the learned latent coordinates.
    Follows from the construction of linear latent evolution and observation maps.

pith-pipeline@v0.9.0 · 5525 in / 1459 out tokens · 51557 ms · 2026-05-15T15:11:31.471628+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.

Forward citations

Cited by 3 Pith papers

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

  1. One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators

    stat.ML 2026-05 unverdicted novelty 7.0

    A single neural operator can approximate the map from arbitrary joint densities to their conditionals, backed by new continuity results and illustrated on Gaussian mixtures.

  2. One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators

    stat.ML 2026-05 unverdicted novelty 6.0

    A single neural operator can approximate the map from joint densities to conditional densities to arbitrary accuracy, with a proof based on continuity of the conditioning operator and a demonstration on Gaussian mixtures.

  3. FLUID: Flow-based Unified Inference for Dynamics

    stat.ML 2026-04 unverdicted novelty 6.0

    FLUID uses a recurrent encoder to create a fixed-size summary of observations, then learns coupled forward and backward flows to approximate filtering distributions and recover smoothing paths for nonlinear dynamics, ...

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · cited by 2 Pith papers · 1 internal anchor

  1. [1]

    Introduction.Data assimilation (DA) optimally integrates dynamical model predictions with noisy and incomplete observations to estimate the evolving state of complex systems. In practice, it combines heterogeneous data sources, such as in-situ measurements, satellite retrievals, and radar observations, with either physics-based or data-driven models of th...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    For ease of presentation, we adopt the following notation

    Data Assimilation.In this section, we introduce the problem formulation and the EnKF framework. For ease of presentation, we adopt the following notation. Lower-case letters denote scalars, bold lower-case letters denote vectors, and upper-case letters denote matrices. Calligraphic letters represent manifolds, sets, and function classes, while script lett...

    work page

  3. [3]

    Latent-Space Ensemble Kalman Filter.As discussed in Section 2, the main limitation of the ensemble Kalman filter in nonlinear settings is not the Kalman update itself, but the mismatch between its linear–Gaussian structure and the non- linear geometry of the physical state space. Rather than modifying the Kalman update, we adopt a representation-driven pe...

    work page

  4. [4]

    Theoretical Analysis.In this section, we provide a performance guarantee of the proposed LAE. Our LAE is motivated by the manifold hypothesis [ 42, 26], which suggests that although the full-order statesx ∈R D is high-dimensional, the set of dynamically attainable states concentrates on a low-dimensional geometric structure. In particular, we assume that ...

    work page

  5. [5]

    Numerical Results.In this section, we present numerical experiments to assess the performance of the proposed LAE-EnKF in nonlinear data assimilation problems. The experiments are designed to evaluate three key aspects of the method: assimilation accuracy, robustness over time, and the benefit of enforcing linear dy- namics in a learned latent space. All ...

    work page arXiv 2027

  6. [6]

    1 N NX i=1 ∥bG(x i)−x + i ∥2 2 # +E S E(x,x+)∼PX,X+ ∥bD◦ bE(x)−x∥ 2 2 −C 1 ES

    Conclusion.This paper introduced the latent autoencoder ensemble Kalman filter (LAE-EnKF), a structure-preserving framework for data assimilation in nonlinear and partially observed systems. By learning a nonlinear encoder–decoder pair together with a stable linear latent dynamical model and a unified observation embedding, LAE-ENKF FOR NONLINEAR DATA ASS...

    work page

  7. [7]

    Akbari, P

    S. Akbari, P. H. Dabaghian, and O. San,Blending machine learning and sequential data assimilation over latent spaces for surrogate modeling of Boussinesq systems, Physica D: Nonlinear Phenomena, 448 (2023), p. 133711

    work page 2023

  8. [8]

    Al-Ghattas, J

    O. Al-Ghattas, J. Bao, and D. Sanz-Alonso,Ensemble Kalman filters with resampling, SIAM/ASA Journal on Uncertainty Quantification, 12 (2024), pp. 411–441

    work page 2024

  9. [9]

    Al-Ghattas and D

    O. Al-Ghattas and D. Sanz-Alonso,Non-asymptotic analysis of ensemble Kalman updates: effective dimension and localization, Information and Inference: A Journal of the IMA, 13 (2023), p. iaad043

    work page 2023

  10. [10]

    Amendola, R

    M. Amendola, R. Arcucci, L. Mottet, C. Q. Casas, S. Fan, C. Pain, P. Linden, and Y.-K. Guo,Data assimilation in the latent space of a convolu- tional autoencoder, in Computational Science, Springer, 2021, pp. 373–386

    work page 2021

  11. [11]

    Arasaratnam and S

    I. Arasaratnam and S. Haykin,Cubature Kalman filters, IEEE Transactions on automatic control, 54 (2009), pp. 1254–1269

    work page 2009

  12. [12]

    Arcucci, J

    R. Arcucci, J. Zhu, S. Hu, and Y.-K. Guo,Deep data assimilation: Inte- grating deep learning with data assimilation, Applied Sciences, 11 (2021), p. 1114

    work page 2021

  13. [13]

    Azencot, N

    O. Azencot, N. B. Erichson, V. Lin, and M. Mahoney,Forecasting sequential data using consistent Koopman autoencoders, in Proceedings of the 37th International Conference on Machine Learning, vol. 119 of Proceedings of Machine Learning Research, 2020, pp. 475–485

    work page 2020

  14. [14]

    F. Bao, Z. Zhang, and G. Zhang,An ensemble score filter for tracking high-dimensional nonlinear dynamical systems, Computer Methods in Applied Mechanics and Engineering, 432 (2024), p. 117447

    work page 2024

  15. [15]

    Bocquet, A

    M. Bocquet, A. Farchi, and Q. Malartic,Online learning of both state and dynamics using ensemble Kalman filters, Foundations of Data Science, 3 (2021), pp. 305–330

    work page 2021

  16. [16]

    Brajard, A

    J. Brajard, A. Carrassi, M. Bocquet, and L. Bertino,Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the Lorenz 96 model, Journal of Computational Science, 44 (2020), p. 101171

    work page 2020

  17. [17]

    S. L. Brunton, M. Budi ˇsi´c, E. Kaiser, and J. N. Kutz,Modern Koopman theory for dynamical systems, SIAM Review, 64 (2022), pp. 229–340

    work page 2022

  18. [18]

    Buehner, P

    M. Buehner, P. L. Houtekamer, C. Charette, H. L. Mitchell, and B. He,Intercomparison of variational data assimilation and the ensemble Kalman filter for global deterministic NWP. Part I: Description and single-observation experiments, Monthly Weather Review, 138 (2010), pp. 1550–1566

    work page 2010

  19. [19]

    Buizza, C

    C. Buizza, C. Quilodr ´an Casas, P. Nadler, J. Mack, S. Marrone, Z. Titus, C. Le Cornec, E. Heylen, T. Dur, L. Baca Ruiz, C. Heaney, J. A. D´ıaz Lopez, K. S. Kumar, and R. Arcucci,Data learning: Integrating data assimilation and machine learning, Journal of Computational Science, 58 (2022), p. 101525

    work page 2022

  20. [20]

    J. A. Carrillo, F. Hoffmann, A. M. Stuart, and U. Vaes,The mean-field ensemble Kalman filter: Near-Gaussian setting, SIAM Journal on Numerical Analysis, 62 (2024), pp. 2549–2587

    work page 2024

  21. [21]

    E. Che, J. Dong, and X. T. Tong,Stochastic gradient descent with adaptive data, arXiv:2410.01195, (2024)

    work page arXiv 2024

  22. [22]

    M. Chen, H. Jiang, W. Liao, and T. Zhao,Nonparametric regression on low-dimensional manifolds using deep ReLU networks: function approximation and statistical recovery, Information and Inference: A Journal of the IMA, 11 LAE-ENKF FOR NONLINEAR DATA ASSIMILATION31 (2022), pp. 1203–1253

    work page 2022

  23. [23]

    Cheng, J

    S. Cheng, J. Chen, C. Anastasiou, P. Angeli, O. K. Matar, Y.-K. Guo, C. C. Pain, and R. Arcucci,Generalised latent assimilation in heterogeneous reduced spaces with machine learning surrogate models, Journal of Scientific Computing, 94 (2023), p. 11

    work page 2023

  24. [24]

    Cheng, I

    S. Cheng, I. C. Prentice, Y. Huang, Y. Jin, Y.-K. Guo, and R. Arcucci, Data-driven surrogate model with latent data assimilation: Application to wildfire forecasting, Journal of Computational Physics, 464 (2022), p. 111302

    work page 2022

  25. [25]

    Cheng, Y

    S. Cheng, Y. Zhuang, L. Kahouadji, C. Liu, J. Chen, O. K. Matar, and R. Arcucci,Multi-domain encoder–decoder neural networks for latent data assimilation in dynamical systems, Computer Methods in Applied Mechanics and Engineering, 430 (2024), p. 117201

    work page 2024

  26. [26]

    Cloninger and T

    A. Cloninger and T. Klock,A deep network construction that adapts to intrinsic dimensionality beyond the domain, Neural Networks, 141 (2021), pp. 404– 419

    work page 2021

  27. [27]

    O. G. Ernst, B. Sprungk, and H.-J. Starkloff,Analysis of the ensemble and polynomial chaos Kalman filters in Bayesian inverse problems, SIAM/ASA Journal on Uncertainty Quantification, 3 (2015), pp. 823–851. [22]G. Evensen,Data assimilation: the ensemble Kalman filter, Springer, 2009

    work page 2015

  28. [28]

    S. J. Julier and J. K. Uhlmann,New extension of the Kalman filter to nonlinear systems, in Signal processing, sensor fusion, and target recognition VI, vol. 3068, 1997, pp. 182–193. [24]E. Kalnay,Atmospheric Modeling, Data Assimilation and Predictability, Cam- bridge University Press, 2002

    work page 1997

  29. [29]

    Kelly, A

    D. Kelly, A. J. Majda, and X. T. Tong,Concrete ensemble Kalman filters with rigorous catastrophic filter divergence, Proceedings of the National Academy of Sciences, 112 (2015), pp. 10589–10594

    work page 2015

  30. [30]

    Y. Khoo, X. T. Tong, W. Wang, and Y. Wang,Temporal label recovery from noisy dynamical data, arXiv:2406.13635, (2024)

    work page arXiv 2024

  31. [31]

    Kirszbraun, ¨Uber die zusammenziehende und Lipschitzsche Transformatio- nen, Fundamenta Mathematicae, 22 (1934), pp

    M. Kirszbraun, ¨Uber die zusammenziehende und Lipschitzsche Transformatio- nen, Fundamenta Mathematicae, 22 (1934), pp. 77–108

    work page 1934

  32. [32]

    J. M. Lee,Riemannian manifolds: an introduction to curvature, vol. 176, Springer Science & Business Media, 2006

    work page 2006

  33. [33]

    Z. Li, B. Dong, and P. Zhang,Latent assimilation with implicit neural representations for unknown dynamics, Journal of Computational Physics, 506 (2024), p. 112953

    work page 2024

  34. [34]

    Z. Li, B. Dong, and P. Zhang,State-observation augmented diffusion model for nonlinear assimilation with unknown dynamics, Journal of Computational Physics, 539 (2025), p. 114240

    work page 2025

  35. [35]

    H. Liu, A. Havrilla, R. Lai, and W. Liao,Deep nonparametric estimation of intrinsic data structures by chart autoencoders: Generalization error and robustness, Applied and Computational Harmonic Analysis, 68 (2024), p. 101602

    work page 2024

  36. [36]

    N. Liu, S. Liu, X. T. Tong, and L. Jiang,Estimate of Koopman modes and eigenvalues with Kalman filter, SIAM Journal on Scientific Computing, 48 (2026), pp. A512–A539

    work page 2026

  37. [37]

    S. Liu, S. Reich, and X. T. Tong,Dropout ensemble Kalman inversion for high dimensional inverse problems, SIAM Journal on Numerical Analysis, 63 (2025), pp. 685–715

    work page 2025

  38. [38]

    Mandel, L

    J. Mandel, L. Cobb, and J. D. Beezley,On the convergence of the ensemble Kalman filter, Applications of Mathematics, 56 (2011), pp. 533–541. 32X. T. TONG, Y. WANG AND L. YAN

    work page 2011

  39. [39]

    Nayak, A

    I. Nayak, A. Chakrabarti, M. Kumar, F. L. Teixeira, and D. Goswami, Temporally-consistent Koopman autoencoders for forecasting dynamical systems, Scientific Reports, 15 (2025), p. 22127

    work page 2025

  40. [40]

    Peyron, A

    M. Peyron, A. Fillion, S. G¨urol, V. Marchais, S. Gratton, P. Boudier, and G. Goret,Latent space data assimilation by using deep learning, Quarterly Journal of the Royal Meteorological Society, 147 (2021), pp. 3759–3777

    work page 2021

  41. [41]

    Reich and C

    S. Reich and C. Cotter,Probabilistic forecasting and Bayesian data assimila- tion, Cambridge University Press, 2015

    work page 2015

  42. [42]

    Rozet and G

    F. Rozet and G. Louppe,Score-based data assimilation, in Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 40521–40541

    work page 2023

  43. [43]

    Sanz-Alonso and N

    D. Sanz-Alonso and N. Waniorek,Long-time accuracy of ensemble Kalman filters for Chaotic dynamical systems and machine-learned dynamical systems, SIAM Journal on Applied Dynamical Systems, 24 (2025), pp. 2246–2286

    work page 2025

  44. [44]

    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

  45. [45]

    Takens,Detecting strange attractors in turbulence, in Dynamical Systems and Turbulence, Warwick 1980, Berlin, Heidelberg, 1981, Springer Berlin Heidelberg, pp

    F. Takens,Detecting strange attractors in turbulence, in Dynamical Systems and Turbulence, Warwick 1980, Berlin, Heidelberg, 1981, Springer Berlin Heidelberg, pp. 366–381

    work page 1980

  46. [46]

    J. B. Tenenbaum, V. d. Silva, and J. C. Langford,A global geometric framework for nonlinear dimensionality reduction, Science, 290 (2000), pp. 2319– 2323

    work page 2000

  47. [47]

    X. T. Tong, A. J. Majda, and D. Kelly,Nonlinear stability and ergodicity of ensemble based Kalman filters, Nonlinearity, 29 (2016), pp. 657–691

    work page 2016

  48. [48]

    X. T. Tong, A. J. Majda, and D. Kelly,Nonlinear stability of the ensemble Kalman filter with adaptive covariance inflation, Communications in Mathematical Sciences, 14 (2016), pp. 1283–1313

    work page 2016

  49. [49]

    A. W. v. d. Vaart and J. A. Wellner,Weak convergence and empirical processes: with applications to statistics, Mathematics and Statistics, Mathematics and Statistics (R0), 1996

    work page 1996

  50. [50]

    Wang and L

    Y. Wang and L. Yan,Data-driven operator inference for parameter estimation in nonlinear partial differential equations, Journal of Computational Physics, 544 (2026), p. 114442

    work page 2026

  51. [51]

    Y. Wang, L. Yan, and T. Zhou,Deep learning-enhanced reduced-order ensem- ble Kalman filter for efficient Bayesian data assimilation of parametric PDEs, Computer Physics Communications, 311 (2025), p. 109544

    work page 2025

  52. [52]

    Yarotsky,Error bounds for approximations with deep ReLU networks, Neural Networks, 94 (2017), pp

    D. Yarotsky,Error bounds for approximations with deep ReLU networks, Neural Networks, 94 (2017), pp. 103–114

    work page 2017

  53. [53]

    J. Zhu, S. Hu, R. Arcucci, C. Xu, J. Zhu, and Y.-k. Guo,Model error correction in data assimilation by integrating neural networks, Big Data Mining and Analytics, 2 (2019), pp. 83–91

    work page 2019