Pith. sign in

REVIEW 4 major objections 7 minor 62 references

Adversarial Transform Particle Filters

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read ATPF learns a pushforward map minimizing kernel MMD between the prior ensemble and a self-normalized importance-sampling posterior estimate, combining particle-filter consistency with EnKF particle diversity.

desk verdict A sensible NLEAF variant with promising experiments, but the main convergence-rate theorem rests on an assumption that cannot hold for the implemented algorithm. read the letter →

arxiv 2502.06165 v1 pith:CYKCYP3H submitted 2025-02-10 stat.ME stat.ML

Chengxin Gong , Wei Lin , Cheng Zhang This is my paper · ORCID
classification stat.MEstat.ML MSC 62M2062F1568T07
keywords adversariallearningdataassimilationensembleKalmanfilterparticlestate-spacemodelsmaximummeandiscrepancyoptimaltransportneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes ATPF, a filtering framework for state-space models that learns a transformation pushing prior ensemble particles toward the posterior. The central move is to estimate posterior expectations by self-normalized importance sampling, exactly as a particle filter does, but instead of reweighting particles, to train a map $G_t$ minimizing the maximum mean discrepancy (MMD) between the pushed-forward prior $G_t\#\hat P_t$ and the importance-sampling estimate of $Q_t$. Kernel methods reduce the adversarial inner maximization to an explicit squared-MMD loss, and an optimal-transport penalty keeps the transformation near the identity for stability. The authors claim this combines the statistical consistency of particle filters with the particle diversity of ensemble Kalman filters, and they provide generalization bounds for both the analysis and forecast steps. If right, the method offers a nonlinear, non-Gaussian update that does not rely on linear-Gaussian assumptions or on particle weights that collapse in high dimensions.

What carries the argument

The load-bearing object is the kernel maximum mean discrepancy combined with self-normalized importance weights $w_t(x')\propto w_{t-1}(x')p_{\varepsilon_t}(y_t-H_t(x'))$. Restricting test functions to the unit ball of an RKHS turns the inner supremum into the explicit objective $E_{x,x'\sim P_t}K(G_t(x),G_t(x'))-2E_{x\sim P_t,y\sim Q_t}K(G_t(x),y)+E_{y,y'\sim Q_t}K(y,y')$, so the bi-level adversarial optimization reduces to a one-level loss estimated by V-statistics and SNIS. An optimal-transport penalty $\lambda_t E_{x\sim P_t}\|G_t(x)-x\|_2^2$ regularizes the transform, and joint training across time treats the forecast maps as activation layers of a deep network. The theory then rides on covering numbers of the test-function and transformation classes and on an oracle assumption that a transformation in a Besov space exactly pushes the prior to the posterior.

What would settle it

Run ATPF on a designed state-space model whose posterior is not in the assumed smoothness class—for example a posterior with a point mass at zero or a heavy-tailed mixture—with increasing $N_1$ and $N_2$, and check whether the empirical MMD and RMSE decay at the claimed rate or plateau at a positive floor. A plateau would indicate the oracle-pushforward assumption fails and the excess-risk bound no longer controls the error.

Watch

Extended reading notes

Core claim

The paper's central claim is that the analysis step of filtering can be solved by minimizing $\mathrm{MMD}(G_t\#\hat P_t\,\|\,Q_t)$, where $\hat P_t$ is the prior ensemble, $Q_t$ is the posterior, and $G_t$ ranges over neural-network transformations. Because the MMD only needs expectations of test functions, the posterior side can be estimated consistently by self-normalized importance sampling, preserving the particle filter's statistical guarantees; because the update moves particles rather than collapsing weights, the ensemble retains the diversity associated with the EnKF. Making the test-function class the unit ball of a reproducing kernel Hilbert space collapses the adversarial inner supremum into the closed-form squared MMD, and an optimal-transport regularizer $\lambda_t E_{x\sim P_t}\|G_t(x)-x\|_2^2$ keeps the map stable. The paper proves nonasymptotic excess-risk bounds whose three terms correspond to SNIS bias, the weighted empirical process, and the empirical process of the transformation class, with explicit rates for Gaussian and linear kernels, and derives forecast-step bounds showing the analysis error propagates through Lipschitz dynamics plus evolution noise. Experiments on a skewed nonlinear model, rainfall data, Lorenz63, and Lorenz96 report lower RMSE than EnKF, NLEAF, and particle-filter baselines, often with far fewer particles than the PF.

Load-bearing premise

The convergence-rate results assume that for every time step the true posterior is exactly the pushforward of the estimated prior by a transformation of a specified smoothness class; if the posterior has point mass, sharp discontinuities, or severe weight collapse so that no such transformation exists, the approximation-error term does not vanish and the stated rates do not apply.

Editorial extensions

If this is right

  • A nonlinear, non-Gaussian posterior can be tracked by a transformed ensemble without resampling, so the filter avoids the particle collapse that standard PFs suffer under deterministic dynamics or high-dimensional weight degeneracy.
  • The kernel-MMD reformulation removes the inner discriminator step, so training the transform network is a single-level minimization rather than an alternating adversarial game.
  • Jointly optimizing the transforms across time stacks the forecast steps into a deep network and, in the Lorenz experiments, yields lower RMSE than sequential training.
  • The forecast-step bounds imply the analysis MMD propagates to the next prior at worst multiplied by a Lipschitz constant plus an evolution-noise term, so error does not diverge without bound under contractive dynamics.
  • With a linear kernel the loss reduces to matching the posterior mean, and the stated rate $O(d((\log N_1)/\sqrt{N_1}+1/\sqrt{N_2}))$ gives an explicit particle budget for mean tracking.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because MMD with a characteristic kernel controls weak convergence, the training objective should in principle match the whole posterior, not just its moments; the paper's comparisons emphasize RMSE, so an editorial test is to compare predictive intervals and tail quantiles on the skewed model and the rainfall model.
  • The method still inherits the particle filter's dependence on SNIS weights, and the $\chi^2(Q_t\|P_t)$ factor in the bound grows with weight collapse; a concrete extension suggested by the discussion is to introduce annealing distributions between prior and posterior and learn a sequence of smaller transforms.
  • The theoretical rates hinge on an oracle transformation of a fixed smoothness class; direct measurement of the residual MMD after training very expressive networks would separate the statistical error from the approximation error in practice.
  • The framework effectively treats filtering as distribution-to-distribution regression, so advances in distributional learning—other kernels, Sinkhorn losses, or flow-based parameterizations—could transfer directly to filtering; this is an inference, not a claim made by the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes the Adversarial Transform Particle Filter (ATPF), a filtering method that learns deterministic transformations of prior particles by minimizing an MMD-type discrepancy against a self-normalized importance sampling (SNIS) estimate of the posterior. The method is presented in three variants: a vanilla ATPF, a kernel-based KATPF using RKHS test functions, and a joint-optimization version (ATPF-JO) that trains all time-step transformations together. The theoretical section gives a decomposition of the excess risk into SNIS bias, weighted empirical-process fluctuations, and prior-ensemble approximation error; states generalization bounds for the analysis step (Theorem 2) and Lipschitz-type bounds for the forecast step (Propositions 2 and 3); and derives convergence rates under Besov-smoothness and neural-network-complexity assumptions (Propositions 4 and 5). Experiments on a nonlinear two-stage model, a rainfall observation model, and Lorenz63/Lorenz96 systems compare RMSE and posterior behavior against EnKF, NLEAF, PF, MPF, PFGR, and ETPF baselines.

Significance. If the theoretical guarantees were valid for the implemented algorithm, the paper would make a valuable contribution: it offers a principled way to combine the asymptotic consistency of importance-sampling-based filtering with the diversity-preserving transformation structure of ensemble Kalman methods, and its empirical results on nonlinear and non-Gaussian benchmarks are competitive. The paper deserves credit for spelling out a three-term excess-risk decomposition, for providing a fairly complete empirical-process analysis with explicit assumptions, and for releasing code and data. However, as detailed below, a central assumption in the convergence-rate analysis (Assumption 6) is incompatible with the finite-particle representation used in Algorithm 1, and the proof of Proposition 1 contains an algebraic error. These issues affect the main theoretical claims and require substantive revision.

major comments (4)
  1. [Section 4.2, Assumption 6 and Proposition 4, Eq. (25)] Assumption 6 is not compatible with the algorithm as implemented. In Section 3 and Algorithm 1, \hat P_t is the empirical distribution of the N1 forecast particles (Algorithm 1, lines 5–9), i.e., an atomic measure. For any absolutely continuous posterior Q_t—which is the situation in every Section 5 experiment—there is no deterministic map G'_t in a Besov space with G'_t#\hat P_t = Q_t, because the pushforward of an atomic measure is atomic. Consequently, inf_{G_t ∈ G_t} MMD(G_t#\hat P_t ∥ Q_t) does not vanish for finite N1, and the rate in Eq. (25) omits an irreducible N1-discretization error. If \hat P_t in Assumption 6 is instead intended to mean the true prior P_t, then the notation conflicts with the definition in Section 3 ("let \hat P_t and P_t be the estimated and the ground truth prior"), and the theoretical object no longer matches the empirical objective optimized in Eq. (12) and Algorithm 1. The analysis-step convergence claim therefore needs either an additional discretization term or a reformulation of the oracle assumption.
  2. [Appendix A, Proposition 1 proof] The proof of Proposition 1 contains an invalid equality: the displayed chain begins with d_FD(P∥Q) = sup_{f∈FD}|E_P f − E_Q f| = sup_{∥f∥_H≤1}|E_P f − E_Q f|^2, which asserts MMD = MMD^2 without a square on the left-hand side. The statement of Proposition 1 and Eq. (10) are standard, and the later derivation of the RKHS form is correct once the first equality is read as the squared MMD, but as printed the proof does not prove the proposition. Please correct the display and use consistent squared/unsquared MMD notation throughout Section 3.1 and Theorem 2.
  3. [Section 4.1, Theorem 2 and Assumption 5] The exponential-in-N2 concentration in Theorem 2 relies on Assumption 5 (strong log-concavity of P_t) through Theorem 4. In the Lorenz63 and Lorenz96 experiments the state evolution is deterministic and chaotic, so P_t is not obviously strongly log-concave, and no verification is provided. If Assumption 5 fails, the paper's own Theorem 6 supplies only a polynomial rate. The authors should either state where Assumption 5 is expected to hold or qualify Theorem 2's bound as conditional on a condition that is not verified in the main experiments.
  4. [Section 3, Eq. (9) and Algorithm 1] The theoretical analysis treats x'^j_t, j=1,...,N2, as i.i.d. draws from the true prior P_t, but Algorithm 1 maintains the x' particles through sequential reweighting, resampling, inflation, and forecast steps. After the first time step these particles are not i.i.d. from P_t; they are a sequentially dependent, resampled particle approximation. The SNIS bias and concentration bounds in Theorem 2 and Theorems 3–6 are derived under the i.i.d. assumption, so they do not directly justify the sequential filtering algorithm. The paper should either prove the bounds conditionally on the forecast ensemble and the filtering history, or explicitly state that the theory covers a single analysis step with fresh i.i.d. prior samples and that the sequential algorithm is an approximation to that setting.
minor comments (7)
  1. [Section 5.2] The phrase "95% credit intervals" should be "95% credible intervals."
  2. [Section 3, Eq. (9)] The notation w_t(x) is used for both the unnormalized product w_{t-1}(x)p_ε_t(y_t − H_t(x)) and the renormalized weights; please clarify the distinction.
  3. [Appendix A, Corollary 1 proof] There is a typo in "right-hande side" that should read "right-hand side."
  4. [Section 4.2, Eq. (22)] The three covering-number cases (γ-finite spectrum, γ-exponential decay, γ-polynomial decay) are not labeled in the display; please label them or refer to them explicitly in the text.
  5. [Appendix C] The localization description uses both l and k to denote the window size, with "2l + 1" and "2k + 1" appearing in the same paragraph; the notation should be made consistent.
  6. [Appendix B, Table 4] The arrow "↗" in Table 4 is not explained in the caption or text; please define it.
  7. [Section 5.1] The phrase "1-dimension 2-stage state-space model" is awkward; consider rewriting as "one-dimensional two-stage state-space model."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ATPF derivation is self-contained, with the SNIS posterior estimate serving as an external target and all key bounds inherited from standard external results.

full rationale

The derivation chain is not circular. The training objective (Eq. 9/12) minimizes the MMD between a pushforward of the estimated prior and an SNIS estimate of the posterior; the target is the Bayes posterior defined by Eq. 3, not an output of the learned transformation. The main generalization bound (Theorem 2/7) is a standard empirical-process decomposition into three terms: the SNIS bias (cited to Agapiou et al. 2017), the stochastic error of the weighted empirical process, and an empirical-process term for the transformation class. None of these terms is defined in terms of the quantity being bounded, and the consistency claim is inherited from the external SNIS theory rather than manufactured inside the paper. The convergence-rate results (Propositions 4-5) rely on external covering-number and neural-network approximation theorems (Suzuki 2019; Liang 2021) and on an explicit oracle assumption (Assumption 6), not on the authors' own prior work. There are no self-citations in the reference list, so no self-citation chain can be load-bearing. The W2 surrogate equivalence in Eqs. 13-14 is imported from external work (Xu et al. 2024) and could be questioned on correctness or formality grounds, but it is not a circular reduction. The main caveat is that Assumption 6, requiring an oracle Besov pushforward map to the exact posterior, is strong and may not hold for the finite-particle implementation of Algorithm 1; however, that is an assumption-strength and correctness concern, not evidence that the paper's claims reduce to their inputs by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical entities or hidden variables. Its theoretical guarantees rest on standard smoothness and log-concavity assumptions plus one strong oracle assumption (Assumption 6). The free parameters are the OT weight, kernel bandwidth, neural network hyperparameters, and inflation coefficient, none of which are derived from first principles.

free parameters (4)
  • OT regularization weight lambda_t = 0.001, 0.0, 0.2, or tuned lambda* per experiment
    Controls the trade-off between MMD matching and displacement cost in Equation 14. Values are chosen per method and per Lorenz63 setting rather than derived from first principles.
  • Gaussian kernel bandwidth w = median pairwise L2 distance among particles
    A data-dependent bandwidth heuristic used for all Gaussian-kernel KATPF experiments. This is a chosen parameter that materially affects the MMD landscape.
  • Neural network architecture and training hyperparameters = not reported in the text
    The paper defines the network class Phi(L,W,S,R) but does not report depth, width, sparsity, learning rate alpha, or iteration limit tau for the experiments. These are essential to reproduce the results.
  • Covariance inflation parameter delta = selected by likelihood of the data
    The inflation step in Appendix C is applied to all methods and chosen via likelihood, which is a standard but still data-dependent calibration choice.
assumptions (6)
  • domain assumption Assumption 1: M_t, H_t are continuous and p_eta_t, p_epsilon_t are C^1.
    Smoothness of the model functions and noise densities is used throughout the proofs and may fail for discontinuous observation operators common in some applications.
  • domain assumption Assumption 2: weights and test functions are uniformly bounded by B.
    Boundedness is used for Hoeffding-type concentration and for the SNIS bias bound. It may be violated for heavy-tailed likelihoods.
  • domain assumption Assumption 4: test functions and weights are beta-Lipschitz.
    Lipschitzity is needed for the empirical process bounds and for the composition argument in the covering-number step.
  • domain assumption Assumption 5: the prior P_t is strongly log-concave with parameter gamma_t.
    This gives the exponential concentration used in Theorem 4. The authors note that Assumptions 4 and 5 are not necessary and provide a slower bound in Theorem 6, so this is not the load-bearing premise for consistency but is load-bearing for the fast rate.
  • ad hoc to paper Assumption 6: there exists an oracle G'_t in a Besov space such that G'_t#\hat P_t = Q_t.
    The convergence rates in Propositions 4 and 5 require the posterior to be an exact pushforward of the estimated prior by a smooth map. This is a strong structural assumption that is not derived from the state-space model and may fail for discrete or non-smooth posterior components.
  • standard math Lemma A.1 in Xu et al. (2024) establishes the equivalence between Equation 13 and Equation 14.
    The paper imports this equivalence without proof. It is likely correct, but the regularization step relies on that external result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adversarial Transform Particle Filters." pith.science (2026). https://pith.science/paper/CYKCYP3H

@misc{pith2026250206165,
  author       = {Pith},
  title        = {Pith review of: Adversarial Transform Particle Filters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CYKCYP3H}},
  note         = {Machine review of arXiv:2502.06165}
}
read the original abstract

The particle filter (PF) and the ensemble Kalman filter (EnKF) are widely used for approximate inference in state-space models. From a Bayesian perspective, these algorithms represent the prior by an ensemble of particles and update it to the posterior with new observations over time. However, the PF often suffers from weight degeneracy in high-dimensional settings, whereas the EnKF relies on linear Gaussian assumptions that can introduce significant approximation errors. In this paper, we propose the Adversarial Transform Particle Filter (ATPF), a novel filtering framework that combines the strengths of the PF and the EnKF through adversarial learning. Specifically, importance sampling is used to ensure statistical consistency as in the PF, while adversarially learned transformations, such as neural networks, allow accurate posterior matching for nonlinear and non-Gaussian systems. In addition, we incorporate kernel methods to ease optimization and leverage regularization techniques based on optimal transport for better statistical properties and numerical stability. We provide theoretical guarantees, including generalization bounds for both the analysis and forecast steps of ATPF. Extensive experiments across various nonlinear and non-Gaussian scenarios demonstrate the effectiveness and practical advantages of our method.

Figures

Figures reproduced from arXiv: 2502.06165 by the authors.

Figure 1
Figure 1. Our parameterization methods and flow chart of filtering for SSMs. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Experimental results of a simple two-stage state-space model. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Experimental results of the rainfall model. The shaded areas represent the corresponding 95% [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Experimental results of the Lorenz63 system when [PITH_FULL_IMAGE:figures/full_fig_p032_4.png]
Figure 5
Figure 5. Figure 5: Experimental results of the Lorenz63 system when [PITH_FULL_IMAGE:figures/full_fig_p033_5.png]
Figure 6
Figure 6. Figure 6: Experimental results of the Lorenz63 system when [PITH_FULL_IMAGE:figures/full_fig_p034_6.png]
Figure 7
Figure 7. Figure 7: Experimental results of the Lorenz63 system when [PITH_FULL_IMAGE:figures/full_fig_p035_7.png]
Figure 8
Figure 8. Figure 8: Experimental results of the Lorenz63 system when [PITH_FULL_IMAGE:figures/full_fig_p036_8.png]
Figure 9
Figure 9. Figure 9: State at location 1 over time in the Lorenz96 system. [PITH_FULL_IMAGE:figures/full_fig_p037_9.png]
Figure 10
Figure 10. Figure 10: State at location 2 over time in the Lorenz96 system (Recall that observation of this location [PITH_FULL_IMAGE:figures/full_fig_p037_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 52 canonical work pages

  1. [1]

    Importance sampling: Intrinsic dimension and computational cost

    Sergios Agapiou, Omiros Papaspiliopoulos, Daniel Sanz-Alonso, and Andrew M Stuart. Importance sampling: Intrinsic dimension and computational cost. Statistical Science, pp.\ 405--431, 2017

    work page 2017

  2. [2]

    Empirical processes associated with v-statistics and a class of estimators under random censoring

    Michael G Akritas. Empirical processes associated with v-statistics and a class of estimators under random censoring. The Annals of Statistics, pp.\ 619--637, 1986

    work page 1986

  3. [3]

    Kernels for vector-valued functions: A review

    Mauricio A Alvarez, Lorenzo Rosasco, Neil D Lawrence, et al. Kernels for vector-valued functions: A review. Foundations and Trends in Machine Learning , 4 0 (3): 0 195--266, 2012

    work page 2012

  4. [4]

    An ensemble adjustment kalman filter for data assimilation

    Jeffrey L Anderson. An ensemble adjustment kalman filter for data assimilation. Monthly weather review, 129 0 (12): 0 2884--2903, 2001

    work page 2001

  5. [5]

    An adaptive covariance inflation error correction algorithm for ensemble filters

    Jeffrey L Anderson. An adaptive covariance inflation error correction algorithm for ensemble filters. Tellus A: Dynamic meteorology and oceanography, 59 0 (2): 0 210--224, 2007

    work page 2007

  6. [6]

    A non-gaussian ensemble filter update for data assimilation

    Jeffrey L Anderson. A non-gaussian ensemble filter update for data assimilation. Monthly Weather Review, 138: 0 4186--4198, 2010

    work page 2010

  7. [7]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and L \'e on Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pp.\ 214--223. PMLR, 2017

    work page 2017

  8. [8]

    A tutorial on particle filters for online nonlinear/non-gaussian bayesian tracking

    M Sanjeev Arulampalam, Simon Maskell, Neil Gordon, and Tim Clapp. A tutorial on particle filters for online nonlinear/non-gaussian bayesian tracking. IEEE Transactions on signal processing, 50 0 (2): 0 174--188, 2002

    work page 2002

Show all 62 references
  1. [9]

    Sequential data assimilation techniques in oceanography

    Laurent Bertino, Geir Evensen, and Hans Wackernagel. Sequential data assimilation techniques in oceanography. International Statistical Review, 71 0 (2): 0 223--241, 2003

    2003

  2. [10]

    Adaptive sampling with the ensemble transform kalman filter

    Craig H Bishop, Brian J Etherton, and Sharanya J Majumdar. Adaptive sampling with the ensemble transform kalman filter. part i: Theoretical aspects. Monthly weather review, 129 0 (3): 0 420--436, 2001

    2001

  3. [11]

    B. P. Carlin, N. G. Polson, and D. S. Stoffer. A monte carlo approach to nonnormal and nonlinear state-space modeling. Journal of the American Statistical Association, 87: 0 493--500, 1992

    1992

  4. [12]

    On the statistical properties of generative adversarial models for low intrinsic data dimension

    Saptarshi Chakraborty and Peter L Bartlett. On the statistical properties of generative adversarial models for low intrinsic data dimension. arXiv preprint arXiv:2401.15801, 2024

    2024 arXiv

  5. [13]

    Sequential data assimilation with a nonlinear quasi-geostrophic model using monte carlo methods to forecast error statistics

    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 0 (C5): 0 10143--10162, 1994

    1994

  6. [14]

    The ensemble kalman filter: Theoretical formulation and practical implementation

    Geir Evensen. The ensemble kalman filter: Theoretical formulation and practical implementation. Ocean dynamics, 53: 0 343--367, 2003

    2003

  7. [15]

    Variational wasserstein gradient flow

    Jiaojiao Fan, Qinsheng Zhang, Amirhossein Taghvaei, and Yongxin Chen. Variational wasserstein gradient flow. arXiv preprint arXiv:2112.02424, 2021

    2021 arXiv

  8. [16]

    Data assimilation for the geosciences: From theory to application

    Steven J Fletcher. Data assimilation for the geosciences: From theory to application. Elsevier, 2022

    2022

  9. [17]

    Covariance tapering for interpolation of large spatial datasets

    Reinhard Furrer, Marc G Genton, and Douglas Nychka. Covariance tapering for interpolation of large spatial datasets. Journal of Computational and Graphical Statistics, 15 0 (3): 0 502--523, 2006

    2006

  10. [18]

    Gamerman

    D. Gamerman. Markov chain monte carlo for dynamics generalized linear models. Biometrika, 85: 0 215--227, 1998

    1998

  11. [19]

    C. J. Geyer. Markov chain M onte C arlo maximum likelihood. Interface Proceedings, 1991

    1991

  12. [20]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  13. [21]

    Novel approach to nonlinear/non-gaussian bayesian state estimation

    Neil J Gordon, David J Salmond, and Adrian FM Smith. Novel approach to nonlinear/non-gaussian bayesian state estimation. In IEE proceedings F (radar and signal processing), volume 140, pp.\ 107--113. IET, 1993

    1993

  14. [22]

    A Distribution-Free Theory of Nonparametric Regression

    L \'a szl \'o Gy \"o rfi, Michael Kohler, Adam Krzyżak, and Harro Walk. A Distribution-Free Theory of Nonparametric Regression. 2002. URL https://api.semanticscholar.org/CorpusID:43315484

    2002

  15. [23]

    Advances in importance sampling

    Timothy Classen Hesterberg. Advances in importance sampling. Stanford University, 1988

    1988

  16. [24]

    Efficient data assimilation for spatiotemporal chaos: A local ensemble transform kalman filter

    Brian R Hunt, Eric J Kostelich, and Istvan Szunyogh. Efficient data assimilation for spatiotemporal chaos: A local ensemble transform kalman filter. Physica D: Nonlinear Phenomena, 230 0 (1-2): 0 112--126, 2007

    2007

  17. [25]

    The variational formulation of the fokker--planck equation

    Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker--planck equation. SIAM journal on mathematical analysis, 29 0 (1): 0 1--17, 1998

    1998

  18. [26]

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

    1960

  19. [27]

    Understanding the ensemble kalman filter

    Matthias Katzfuss, Jonathan R Stroud, and Christopher K Wikle. Understanding the ensemble kalman filter. The American Statistician, 70 0 (4): 0 350--357, 2016

    2016

  20. [28]

    Ensemble kalman methods for high-dimensional hierarchical dynamic space-time models

    Matthias Katzfuss, Jonathan R Stroud, and Christopher K Wikle. Ensemble kalman methods for high-dimensional hierarchical dynamic space-time models. Journal of the American Statistical Association, 115 0 (530): 0 866--885, 2020

    2020

  21. [29]

    Introduction to kalman filter and its applications

    Youngjoo Kim, Hyochoong Bang, et al. Introduction to kalman filter and its applications. Introduction and Implementations of the Kalman Filter, 1: 0 1--16, 2018

    2018

  22. [30]

    A note on importance sampling using standardized weights

    Augustine Kong. A note on importance sampling using standardized weights. University of Chicago, Dept. of Statistics, Tech. Rep, 348: 0 14, 1992

    1992

  23. [31]

    Recursive monte carlo filters: algorithms and theoretical analysis

    Hans R K \"u nsch. Recursive monte carlo filters: algorithms and theoretical analysis. Annals of Statistics, pp.\ 1983--2021, 2005

    1983

  24. [32]

    Bellman filtering and smoothing for state--space models

    Rutger-Jan Lange. Bellman filtering and smoothing for state--space models. Journal of Econometrics, 238 0 (2): 0 105632, 2024

    2024

  25. [33]

    A moment matching ensemble filter for nonlinear non-gaussian data assimilation

    Jing Lei and Peter Bickel. A moment matching ensemble filter for nonlinear non-gaussian data assimilation. Monthly Weather Review, 139 0 (12): 0 3964--3973, 2011

    2011

  26. [34]

    How well generative adversarial networks learn distributions

    Tengyuan Liang. How well generative adversarial networks learn distributions. Journal of Machine Learning Research, 22 0 (228): 0 1--41, 2021

    2021

  27. [35]

    Stein variational gradient descent: A general purpose bayesian inference algorithm

    Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016

    2016

  28. [36]

    Implicit regularization in nonconvex statistical estimation: Gradient descent converges linearly for phase retrieval and matrix completion

    Cong Ma, Kaizheng Wang, Yuejie Chi, and Yuxin Chen. Implicit regularization in nonconvex statistical estimation: Gradient descent converges linearly for phase retrieval and matrix completion. In International Conference on Machine Learning, pp.\ 3345--3354. PMLR, 2018

    2018

  29. [37]

    Concentration inequalities for log-concave distributions with applications to random surface fluctuations

    Alexander Magazinov and Ron Peled. Concentration inequalities for log-concave distributions with applications to random surface fluctuations. The Annals of Probability, 50 0 (2): 0 735--770, 2022

    2022

  30. [38]

    Fastslam 2.0: An improved particle filtering algorithm for simultaneous localization and mapping that provably converges

    Michael Montemerlo, Sebastian Thrun, Daphne Koller, Ben Wegbreit, et al. Fastslam 2.0: An improved particle filtering algorithm for simultaneous localization and mapping that provably converges. In IJCAI, volume 3, pp.\ 1151--1156, 2003

    2003

  31. [39]

    Rao-blackwellised particle filtering for dynamic bayesian networks

    Kevin Murphy and Stuart Russell. Rao-blackwellised particle filtering for dynamic bayesian networks. In Sequential Monte Carlo methods in practice, pp.\ 499--515. Springer, 2001

    2001

  32. [40]

    Improving regularised particle filters

    Christian Musso, Nadia Oudjane, and Francois Le Gland. Improving regularised particle filters. In Sequential Monte Carlo methods in practice, pp.\ 247--271. Springer, 2001

    2001

  33. [41]

    Merging particle filter for sequential data assimilation

    Shinya Nakano, Genta Ueno, and Tomoyuki Higuchi. Merging particle filter for sequential data assimilation. Nonlinear Processes in Geophysics, 14 0 (4): 0 395--408, 2007

    2007

  34. [42]

    Diffusion models are minimax optimal distribution estimators

    Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pp.\ 26517--26582. PMLR, 2023

    2023

  35. [43]

    Optimal approximation of piecewise smooth functions using deep relu neural networks

    Philipp Petersen and Felix Voigtlaender. Optimal approximation of piecewise smooth functions using deep relu neural networks. Neural Networks, 108: 0 296--330, 2018

    2018

  36. [44]

    A localized particle filter for high-dimensional nonlinear systems

    Jonathan Poterjoy. A localized particle filter for high-dimensional nonlinear systems. Monthly Weather Review, 144 0 (1): 0 59--76, 2016

    2016

  37. [45]

    Overview of global data assimilation developments in numerical weather-prediction centres

    Florence Rabier. Overview of global data assimilation developments in numerical weather-prediction centres. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography, 131 0 (613): 0 3215--3233, 2005

    2005

  38. [46]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 0 (2): 0 3, 2022

    2022 arXiv

  39. [47]

    A nonparametric ensemble transform method for bayesian inference

    Sebastian Reich. A nonparametric ensemble transform method for bayesian inference. SIAM Journal on Scientific Computing, 35 0 (4): 0 A2013--A2024, 2013

    2013

  40. [48]

    Venezuelan rainfall data analysed by using a bayesian space--time model

    Bruno Sanso and Lelys Guenni. Venezuelan rainfall data analysed by using a bayesian space--time model. Journal of the Royal Statistical Society: Series C (Applied Statistics), 48 0 (3): 0 345--362, 1999

    1999

  41. [49]

    Log-concavity and strong log-concavity: a review

    Adrien Saumard and Jon A Wellner. Log-concavity and strong log-concavity: a review. Statistics surveys, 8: 0 45, 2014

    2014

  42. [50]

    Approximation theorems of mathematical statistics

    Robert J Serfling. Approximation theorems of mathematical statistics. John Wiley & Sons, 2009

    2009

  43. [51]

    Shephard and M

    N. Shephard and M. Pitt. Likelihood analysis of non-gaussian measurement time series. Biometrika, 84: 0 653--667, 1997

    1997

  44. [52]

    Maximum mean discrepancy

    Alexander J Smola, A Gretton, and K Borgwardt. Maximum mean discrepancy. In 13th international conference, ICONIP, pp.\ 3--6, 2006

    2006

  45. [53]

    Obstacles to high-dimensional particle filtering

    Chris Snyder, Thomas Bengtsson, Peter Bickel, and Jeff Anderson. Obstacles to high-dimensional particle filtering. Monthly Weather Review, 136 0 (12): 0 4629--4640, 2008

    2008

  46. [54]

    High-dimensional ensemble kalman filter with localization, inflation, and iterative updates

    Hao-Xuan Sun, Shouxia Wang, Xiaogu Zheng, and Song Xi Chen. High-dimensional ensemble kalman filter with localization, inflation, and iterative updates. Quarterly Journal of the Royal Meteorological Society, 150 0 (765): 0 4870--4884, 2024

    2024

  47. [55]

    Adaptivity of deep relu network for learning in besov and mixed smooth besov spaces: optimal rate and curse of dimensionality

    Taiji Suzuki. Adaptivity of deep relu network for learning in besov and mixed smooth besov spaces: optimal rate and curse of dimensionality. In International Conference on Learning Representations, volume 7, 2019

    2019

  48. [56]

    High-Dimensional Statistics: A Non-Asymptotic Viewpoint

    Martin Wainwright. High-Dimensional Statistics: A Non-Asymptotic Viewpoint. 2019. ISBN 9781108498029. doi:10.1017/9781108627771

    2019 doi

  49. [57]

    The implicit and explicit regularization effects of dropout

    Colin Wei, Sham Kakade, and Tengyu Ma. The implicit and explicit regularization effects of dropout. In International Conference on Machine Learning, pp.\ 10181--10192. PMLR, 2020

    2020

  50. [58]

    Bayesian learning via stochastic gradient langevin dynamics

    Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pp.\ 681--688. Citeseer, 2011

    2011

  51. [59]

    A note on the particle filter with posterior gaussian resampling

    Xiaozhen Xiong, Ionel Michael Navon, and Bahri Uzunoglu. A note on the particle filter with posterior gaussian resampling. Tellus A: Dynamic Meteorology and Oceanography, 58 0 (4): 0 456--460, 2006

    2006

  52. [60]

    Normalizing flow neural networks by jko scheme

    Chen Xu, Xiuyuan Cheng, and Yao Xie. Normalizing flow neural networks by jko scheme. Advances in Neural Information Processing Systems, 36, 2024

    2024

  53. [61]

    On function approximation in reinforcement learning: Optimism in the face of large state spaces

    Zhuoran Yang, Chi Jin, Zhaoran Wang, Mengdi Wang, and Michael I Jordan. On function approximation in reinforcement learning: Optimism in the face of large state spaces. arXiv preprint arXiv:2011.04622, 2020

    2011 arXiv

  54. [62]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.