Pith. sign in

REVIEW 4 major objections 4 minor 61 references

A localized consensus-based sampling algorithm

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By dropping the Gaussian-only approximation in a new derivation of consensus-based sampling from Langevin dynamics, the paper constructs a gradient-free, affine-invariant sampler that stays exact on Gaussian targets and handles…

desk verdict A clean new derivation of a localized CBS sampler with solid Gaussian theory; the non-Gaussian claims are honestly offered as heuristic, which makes the conditional verdict fair. read the letter →

arxiv 2505.24861 v4 pith:LKDO653F submitted 2025-05-30 math.NA cs.NAmath.OC

Arne Bouillon , Alexander Bodard , Panagiotis Patrinos , Dirk Nuyens , Giovanni Samaey This is my paper
classification math.NAcs.NAmath.OC MSC 65C3562F1535Q8460H10
keywords localizedconsensus-basedsamplingBayesianinverseproblemsgradient-freeMoreauenvelopeproximaloperatoraffine-invarianceinteractingLangevindynamicsmean-fieldlimit
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's central claim is that consensus-based sampling (CBS), an ensemble sampler whose particles are pulled toward a probability-weighted mean without ever computing gradients, can be re-derived from ensemble-preconditioned Langevin dynamics, and that this derivation reveals exactly which approximation locks CBS to Gaussian targets. The derivation smooths the potential with a Moreau envelope, replaces the gradient by a proximal operator, and approximates that operator by a weighted particle mean; the final step, valid only for Gaussian targets, is precisely the one to discard. Removing it yields localized CBS, in which each particle interacts most strongly with nearby particles. The paper proves that localized CBS admits Gaussian targets as stationary distributions in the mean-field limit when a parameter $\gamma$ is set by a closed-form formula, and that the dynamics are affine-invariant, so rescaling or rotating the parameters does not change sampling performance. The payoff, demonstrated numerically, is a fully gradient-free, parallelizable sampler that handles multimodal and poorly scaled posteriors of the kind arising in Bayesian inverse problems.

What carries the argument

The load-bearing construction is the three-step reduction from the ensemble-preconditioned Langevin SDE (14) to the localized CBS SDE (28). Step one replaces the potential $V$ by its $\gamma$-rescaled Moreau envelope $\gamma V^{\kappa C_\alpha(\rho_t)}$, a smoothed version of $V$, and uses the identity $\nabla V^A(u)=A^{-1}(u-\mathrm{prox}^A_V(u))$ to convert the gradient into a proximal operator. Step two approximates the proximal operator by the finite-$\beta$ weighted mean (23), whose weights combine the distance $\|v-u\|^2_A$ with the target density $\hat\pi(v)^a$. Step three is the Gaussian-only identity (25), which with $\kappa=\beta=\alpha$ and $\gamma=2\alpha/(\alpha+1)$ turns the dynamics into classical CBS (15); localized CBS is defined by dropping this step and keeping $\kappa$ and $\beta$ independent. The Gaussian analysis of Lemma 4.1 and Corollary 4.2, which derives the mean and covariance ODEs (31) and solves for the bias-removing $\gamma$ in (36)-(38), is the mechanism that turns exactness for Gaussian targets into a tuning rule used heuristically for non-Gaussian ones.

What would settle it

Fix a non-Gaussian target with known closed-form moments, for example a two-component Gaussian mixture with known weights, run localized CBS with the Gaussian-calibrated $\gamma=(\kappa+\beta)/(\beta+1)$ and the unweighted-covariance preconditioner, and measure the Wasserstein-2 distance (or moment error) between the empirical stationary distribution and the mixture as $J$ grows and as $\beta\to\infty$, $\kappa\to0$; if the error does not decrease toward the interacting-Langevin limit, the weighted-mean proxy is not playing the role the Gaussian analysis assigns it. A second observable check of the affine-invariance claim is to apply a fixed ill-conditioned linear transformation to a Gaussian or mixture target and verify that the sampler's error is unchanged up to numerical tolerance.

Watch

Extended reading notes

Core claim

Localized CBS is the dynamics (28), in which the drift pulls each particle toward the weighted mean $\mu_{\beta,\kappa C_{U,t}/\beta}(\rho_t;U_t)$ that stands in for a proximal operator, and the noise is scaled by a covariance-based preconditioner $C_{U,t}$. When the target is Gaussian, Lemma 4.1 gives exact ODEs for the mean and covariance of the particle law, and Corollary 4.2 shows that $\pi$ is stationary in the mean-field limit provided $\gamma$ satisfies one of the closed-form formulas (36)-(38), for the default unweighted-covariance preconditioner $\gamma=(\kappa+\beta)/(\beta+1)$, with analogous values for weighted and localized covariances, while no choice of $\gamma$ works for a constant preconditioner. Beyond Gaussians, Lemma 4.4 proves affine-invariance for the covariance-based preconditioners, Lemma 5.1 shows that re-imposing the Gaussian step (with $\kappa=\beta=\alpha$ and $\gamma=2\alpha/(\alpha+1)$) recovers classical CBS up to a known time rescaling, and the comparison in Section 5 identifies three advantages over polarized CBS: a correction term $\nabla_u\cdot C_{U,t}$ that restores consistency with interacting Langevin dynamics as $\beta\to\infty$ and $\kappa\to0$, an anisotropic distance weight that confers affine-invariance, and a decoupling of the weights in the mean from those in the covariance so that a cheap unweighted covariance suffices. Numerical experiments show the method sampling multimodal, tent-shaped, and Darcy-flow posteriors, recovering from an incorrectly scaled initial ensemble, and losing accuracy when the correction term is omitted.

Load-bearing premise

For non-Gaussian targets, the method's usefulness presupposes that the finite-$\beta$ weighted mean still behaves like the proximal operator and that the Gaussian-derived value of $\gamma$ keeps the bias acceptably small; the paper states explicitly that the stationary distribution of (28) for non-Gaussian $\pi$ and the mean-field limit remain open questions (Remark 4.3, Section 8), so the non-Gaussian results rest on the specific numerical configurations tested.

Editorial extensions

If this is right

  • Bayesian inverse problems whose posteriors are known only up to a normalizing constant can be sampled without gradients: only evaluations of the unnormalized density and the particle positions enter the update, and all particle work parallelizes trivially.
  • Affine-invariance removes the need to pre-scale or precondition the parameter space by hand: a poorly scaled target is sampled as accurately as an isotropic one, and the sampler recovers from an incorrectly scaled initial ensemble.
  • Because the correction term $\nabla_u\cdot C_{U,t}$ makes the dynamics consistent with interacting Langevin dynamics as $\beta\to\infty$ and $\kappa\to0$, localized CBS can explore multimodal targets that classical CBS, which is tied to its Gaussian mean and covariance, cannot represent.
  • Localized CBS contains classical CBS as a special case for Gaussian targets up to the time factor $c=\alpha'/(\alpha(\alpha'+1))$, so it functions as a strict generalization rather than a separate heuristic.
  • The random-batch variant with interaction fraction $\nu<1$ improves sampling of higher-dimensional multimodal targets at a modest computational premium, which runs counter to the usual expectation that subsampling only trades accuracy for speed.

Reading between the lines

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

  • The open non-Gaussian question (Remark 4.3) suggests a concrete program: compute the stationary distribution of (28) perturbatively in the small-$\kappa$, large-$\beta$ regime and check whether the Gaussian-tuned $\gamma$ minimizes the first-order bias; if so, the heuristic tuning would acquire a rigorous justification the paper does not supply.
  • The finite-particle experiments (Section 7.3) hint at an optimal-balance relation between $\beta$, $\kappa$, and the ensemble size $J$, with moderate parameters beating extreme ones when $J$ is small; an empirical or theoretical rule of the form $\beta\kappa\approx f(J,d)$ would make the method usable without hand-tuning.
  • Because the method never evaluates gradients, it should extend to non-differentiable potentials, which the authors note as future work; a direct test would be sampling from a Laplace-prior or total-variation posterior, where gradient-based samplers cannot run.
  • The correction-term study suggests a diagnostic the paper does not pursue: comparing samples with and without $\nabla_u\cdot C_{U,t}$ for a given target measures how far the Gaussian approximations are from being exact.
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 / 4 minor

Summary. The paper proposes a new gradient-free, parallelizable ensemble sampler called localized consensus-based sampling (localized CBS). The method is derived by starting from ensemble-preconditioned Langevin dynamics, replacing the potential by a Moreau envelope, replacing the proximal operator by a finite-beta weighted mean, and then removing only the Gaussian-motivated approximation used in standard CBS. The resulting mean-field SDE (28) is studied: Lemma 4.1 derives ODEs for the Gaussian mean and covariance, Corollary 4.2 gives explicit gamma values making a Gaussian target stationary, and Lemma 4.4 proves affine-invariance for covariance-type preconditioners. Section 6 provides particle and time discretizations, including a random-batch modification, and Section 7 reports numerical experiments on Gaussian, multimodal, tent-shaped, and Darcy-flow targets. The authors explicitly state in Remark 4.3 and Section 8 that well-posedness and the stationary distribution of the dynamics for non-Gaussian targets, as well as the mean-field limit, remain open.

Significance. If the advertised properties hold, localized CBS would be a genuinely useful contribution: it is derivative-free, parallelizable, affine-invariant, and has a transparent derivation connecting consensus-based methods to interacting Langevin dynamics. The Gaussian analysis in Section 4 is clean, the gamma values are explicit and machine-verifiable from the displayed ODEs, and the affine-invariance proof is correct. The authors are also commendably honest about the open non-Gaussian theory. However, the paper's central practical claim is non-Gaussian sampling, and that claim is currently supported only by selected numerical configurations; the default algorithm fails in one of the flagship multimodal experiments, and the successful variant relies on a random-batch heuristic with no theoretical justification. The significance of the contribution will be considerably higher if the non-Gaussian gap is closed or at least sharply characterized.

major comments (4)
  1. [Remark 4.3; Section 8] The paper's stated objective is sampling from non-Gaussian distributions (Section 1.3), yet Remark 4.3 explicitly concedes that for non-Gaussian pi the well-posedness of (28) and the relation between pi and any stationary distribution are open questions. This is load-bearing because the only evidence for the main claim is empirical. I am not asking for a full non-Gaussian convergence theorem, but the manuscript should either provide a quantitative statement for a tractable class (e.g., log-concave or weakly convex targets) or unambiguously reframe the contribution as a heuristic algorithm with Gaussian exactness as a consistency check. As written, the abstract's 'exact for Gaussian targets' can easily be read as implying more than is proven.
  2. [Section 7.2, Figure 2; Eq. (61)] In the d=10 bimodal example, the default all-to-all interaction (nu=1) fails to approximate the target, and good agreement is obtained only with the random-batch modification nu=0.5. This modification is introduced in Section 6 purely as a computational device and is given no theoretical or even heuristic explanation for why subsampling improves sampling accuracy. Since this single experiment is the main evidence for non-Gaussian performance in higher dimensions, the random-batch mechanism is load-bearing and needs analysis, or at minimum a systematic study of its effect, before the robustness claim can be accepted.
  3. [Section 7.2 and Section 7.4] The numerical comparisons do not currently establish the advertised 'improved robustness over polarized CBS'. In the multimodal experiment, polarized CBS and localized ALDI are hand-tuned per dimension (lambda=0.005 vs 0.1 and lambda=0.02 vs 0.4), while localized CBS uses fixed parameters; no error bars or run-to-run variability are reported, and samples are aggregated over 16 runs. In the Darcy flow experiment (Section 7.4), localized CBS is compared only to random-walk MCMC, not to polarized CBS or localized ALDI, so the only Bayesian inverse problem in the paper does not test the claim against the most relevant competitors. I would like to see error bars, multiple seeds, and a clearer statement of the parameter-selection protocol.
  4. [Eqs. (22) and (24); Remark 7.1] The derivation replaces the proximal operator by the finite-beta weighted mean (24), but Eq. (22) is only a Laplace-principle limit as beta tends to infinity; the paper provides no quantitative bound on the discrepancy for finite beta. Corollary 4.2 removes the Gaussian bias by choosing gamma, which is transparent parameter design rather than evidence that the finite-beta dynamics approximate pi for non-Gaussian targets. Remark 7.1 in fact shows a mean-field bias that persists as J grows, confirming that the method has an intrinsic bias that should be discussed as such rather than left as a numerical observation.
minor comments (4)
  1. [Definition 2.1; Lemma 4.4] Definition 2.1 is called a definition, but Lemma 4.4 and the surrounding text repeatedly cite it as 'Theorem 2.1'; please reconcile the cross-reference.
  2. [Lemma 5.1 proof; Corollary 4.2 proof] The text after Lemma 5.1 refers to 'Theorem 5.1', and the proof of Corollary 4.2 cites 'Theorem 4.1'; the correct labels appear to be Lemma 5.1 and Lemma 4.1, respectively.
  3. [Section 7.4; Section 8] The text cites 'Theorem 6.1' when referring to the scaling discussion, but the relevant statement is Remark 6.1; the same mislabel appears in the conclusions.
  4. [Section 2 heading] The heading 'The interacing-Langevin samplers' contains a typo; it should be 'interacting-Langevin'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Gaussian exactness is an explicit parameter choice, and non-Gaussian limitations are openly stated.

full rationale

The derivation chain is self-contained. Localized CBS dynamics (28) are obtained from preconditioned Langevin dynamics (14) through three explicitly stated approximations: replacing the potential by a Moreau envelope, replacing the proximal operator by a weighted mean, and assuming a Gaussian target; the Gaussian assumption is then deliberately omitted to obtain the non-Gaussian variant. The Gaussian exactness in Corollary 4.2 is not a hidden prediction: the parameter gamma is solved from the stationarity condition d/dt Sigma_t = 0, so the statement 'pi is stationary for Gaussian targets' is by explicit parameter design, and the paper states that this gamma is used for non-Gaussian targets only as a heuristic. There are no self-citation loops: the external results used (Laplace principle, Moreau calculus, CBS analysis) are standard and do not presuppose localized CBS. The limitations that could be mistaken for missing support are explicitly acknowledged rather than hidden: Remark 4.3 says the non-Gaussian stationary distribution and well-posedness are open, Remark 6.1 says high-dimensional performance is unanalyzed, Section 8 lists incomplete theory including the mean-field limit, and the random-batch heuristic is presented as an empirical observation. These are scope restrictions and correctness risks, not circular reasoning. The numerical comparisons are experiments, not outputs forced by the derivation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters gamma, beta, kappa, lambda, and nu are hand-tuned or derived to make the Gaussian case exact, and the non-Gaussian behavior depends on them heuristically.

free parameters (5)
  • gamma = (kappa+beta)/(beta+1) for unweighted covariance; (alpha+1)^-1 times that for weighted covariance
    Chosen so that a Gaussian target is stationary (Corollary 4.2); used as a heuristic for non-Gaussian targets in experiments.
  • beta = 5, 10, 3 in numerical experiments
    Finite approximation of the Laplace principle limit; controls localization versus bias in the weighted mean approximation.
  • kappa = 0.01, 0.03, 0.2 in numerical experiments
    Moreau envelope smoothing scale; smaller values make the envelope closer to the original potential, but increase finite-particle sensitivity.
  • lambda = 0.1, 0.5, infinity in localized-preconditioner experiments
    Width of the distance kernel in the preconditioner C_alpha,lambda; only used when the localized weighted covariance is selected.
  • nu = 0.5 or 1 (random-batch parameter)
    Fraction of particle interactions retained after random subsampling; nu=1 is full interaction and nu=0.5 improved d=10 multimodal results.
assumptions (5)
  • standard math Laplace principle justifies replacing the proximal operator by a weighted mean (Eq. 22).
    Used as the second approximation step in Section 3; the finite-beta error is not quantified for non-Gaussian targets.
  • domain assumption The potential V is differentiable and L-weakly convex, and C_alpha(rho_t) is positive definite.
    Assumed at the start of Section 3 for the Moreau-envelope gradient formula and to ensure the dynamics are well-defined.
  • domain assumption Gaussian target and initial distributions for the analytic results in Section 4.3.
    The exactness result and the gamma formulas in Corollary 4.2 are derived only under pi = N(m, Sigma) and rho_0 Gaussian.
  • domain assumption Propagation of chaos / mean-field limit for the particle discretization (Section 6).
    The finite-particle system replaces rho_t by its empirical measure; convergence is cited from [57, 15] but not proven for localized CBS.
  • standard math Moreau envelope gradient identity (Eq. 19).
    Standard result from Rockafellar and Wets, used to eliminate the gradient from the Langevin dynamics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A localized consensus-based sampling algorithm." pith.science (2026). https://pith.science/paper/LKDO653F

@misc{pith2026250524861,
  author       = {Pith},
  title        = {Pith review of: A localized consensus-based sampling algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKDO653F}},
  note         = {Machine review of arXiv:2505.24861}
}
read the original abstract

We propose a localized consensus-based method for sampling from non-Gaussian distributions, a task that frequently arises when solving Bayesian inverse problems. Our method arises from an alternative derivation of consensus-based sampling (CBS). Starting from ensemble-preconditioned Langevin dynamics, we replace the potential by its Moreau envelope -- a smoother approximation -- in order to replace the gradient in the Langevin equation with a proximal operator. We then approximate this operator by a weighted mean. In the limit of infinitely smoothing the potential to a quadratic function, this procedure recovers the standard CBS dynamics. In addition, outside this limit, we retrieve a refined variant of polarized CBS. We call the resulting algorithm localized consensus-based sampling, since particles interact more with nearby particles than with faraway ones. Our method is affine-invariant, exact for Gaussian targets in the mean-field limit, and demonstrates improved robustness over polarized CBS in numerical experiments. Like other consensus-based methods, localized CBS is gradient-free and easily parallelizable.

Figures

Figures reproduced from arXiv: 2505.24861 by the authors.

Figure 1
Figure 1. Localized CBS samples for a one-dimensional Gaussian target distribution [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Marginalized samples for a 1- or 10-dimensional multimodal target distribution. [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Marginalized samples for the poorly scaled distribution with potential [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Localized CBS samples for (70) with various preconditioners and initial distributions 20 [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Localized and polarized CBS samples for (70) correction term is omitted. We also apply polarized CBS, which does not have a similar cor￾rection term, with α = 10 and λ ∈ {0.001, 0.002, 0.005}. We use N = 1000 timesteps, J = 200 particles, and an initial covariance Σ0 =…
Figure 6
Figure 6. Figure 6: Wasserstein-2 error between the localized CBS samples and the target distribution [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: MCMC and localized CBS samples of the first 8 KL terms in the [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 58 canonical work pages

  1. [1]

    Al-Ghattas and D

    O. Al-Ghattas and D. Sanz-Alonso , Non-asymptotic analysis of ensemble Kalman updates: Effective dimension and localization , Inf. Inference J. IMA, 13 (2024), p. iaad043

    work page 2024

  2. [2]

    Besag , Discussion of the Paper by Grenander and Miller , J

    J. Besag , Discussion of the Paper by Grenander and Miller , J. R. Stat. Soc. B, 56 (1994), pp. 591--592

    work page 1994

  3. [3]

    Brooks, A

    S. Brooks, A. Gelman, G. Jones, and X.-L. Meng , eds., Handbook of Markov Chain Monte Carlo , Chapman and Hall/CRC , New York, May 2011

    work page 2011

  4. [4]

    Bungert, T

    L. Bungert, T. Roith, and P. Wacker , Polarized consensus-based dynamics for optimization and sampling , Math. Program., (2024)

    work page 2024

  5. [5]

    J. A. Carrillo, Y.-P. Choi, C. Totzeck, and O. Tse , An analytical framework for consensus-based global optimization method , Math. Models Methods Appl. Sci., 28 (2018), pp. 1037--1066

    work page 2018

  6. [6]

    J. A. Carrillo, F. Hoffmann, A. M. Stuart, and U. Vaes , Consensus-based sampling , Stud. Appl. Math., 148 (2022), pp. 1069--1140

    work page 2022

  7. [7]

    N. K. Chada, Y. Chen, and D. Sanz-Alonso , Iterative ensemble Kalman methods: A unified perspective with some new variants , Found. Data Sci., 3 (2021), pp. 331--369

    work page 2021

  8. [8]

    S. L. Cotter, G. O. Roberts, A. M. Stuart, and D. White , MCMC Methods for Functions : Modifying Old Algorithms to Make Them Faster , Stat. Sci., 28 (2013), pp. 424--446

    work page 2013

Show all 61 references
  1. [9]

    T. Cui, K. J. H. Law, and Y. M. Marzouk , Dimension-independent likelihood-informed MCMC , J. Comput. Phys., 304 (2016), pp. 109--137

    2016

  2. [10]

    Dashti, K

    M. Dashti, K. J. H. Law, A. M. Stuart, and J. Voss , MAP estimators and their consistency in Bayesian nonparametric inverse problems , Inverse Probl., 29 (2013), p. 095017

    2013

  3. [11]

    Del Moral, A

    P. Del Moral, A. Doucet, and A. Jasra , Sequential Monte Carlo Samplers , J. R. Stat. Soc. B, 68 (2006), pp. 411--436

    2006

  4. [12]

    Dembo and O

    A. Dembo and O. Zeitouni , Large Deviations Techniques and Applications , vol. 38 of Stochastic Modelling and Applied Probability , Springer, Berlin, Heidelberg, 2010

    2010

  5. [13]

    Ding and Q

    Z. Ding and Q. Li , Ensemble Kalman inversion: Mean-field limit and convergence analysis , Stat. Comput., 31 (2021), p. 9

    2021

  6. [14]

    Ding and Q

    Z. Ding and Q. Li , Ensemble Kalman Sampler : Mean-field Limit and Convergence Analysis , SIAM J. Math. Anal., 53 (2021), pp. 1546--1578

    2021

  7. [15]

    T. J. Dodwell, C. Ketelsen, R. Scheichl, and A. L. Teckentrup , A Hierarchical Multilevel Markov Chain Monte Carlo Algorithm with Applications to Uncertainty Quantification in Subsurface Flow , SIAM/ASA J. Uncertainty Quantif., 3 (2015), pp. 1075--1108

    2015

  8. [16]

    O. R. A. Dunbar, A. B. Duncan, A. M. Stuart, and M.-T. Wolfram , Ensemble Inference Methods for Models With Noisy and Expensive Likelihoods , SIAM J. Appl. Dyn. Syst., 21 (2022), pp. 1539--1572

    2022

  9. [17]

    Evensen , Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics , J

    G. Evensen , Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics , J. Geophys. Res. Oceans, 99 (1994), pp. 10143--10162

    1994

  10. [18]

    Fornasier, T

    M. Fornasier, T. Klock, and K. Riedl , Consensus- Based Optimization Methods Converge Globally , SIAM J. Optim., 34 (2024), pp. 2973--3004

    2024

  11. [19]

    Garbuno-Inigo , F

    A. Garbuno-Inigo , F. Hoffmann, W. Li, and A. M. Stuart , Interacting Langevin Diffusions : Gradient Structure and Ensemble Kalman Sampler , SIAM J. Appl. Dyn. Syst., 19 (2020), pp. 412--441

    2020

  12. [20]

    Garbuno-Inigo , N

    A. Garbuno-Inigo , N. N \"u sken, and S. Reich , Affine Invariant Interacting Langevin Dynamics for Bayesian Inference , SIAM J. Appl. Dyn. Syst., 19 (2020), pp. 1633--1658

    2020

  13. [21]

    N. J. Gerber, F. Hoffmann, and U. Vaes , Mean-field limits for Consensus-Based Optimization and Sampling , arXiv:2312.07373, (2024)

    2024

  14. [22]

    Goodman and J

    J. Goodman and J. Weare , Ensemble samplers with affine invariance , Commun. Appl. Math. Comput. Sci., 5 (2010), pp. 65--80

    2010

  15. [23]

    Greengard , An ensemblized Metropolized Langevin sampler , Master's thesis, Courant Institute, New York University, (2015)

    P. Greengard , An ensemblized Metropolized Langevin sampler , Master's thesis, Courant Institute, New York University, (2015)

    2015

  16. [24]

    W. K. Hastings , Monte Carlo Sampling Methods Using Markov Chains and Their Applications , Biometrika, 57 (1970), pp. 97--109

    1970

  17. [25]

    Herrmann, M

    L. Herrmann, M. Keller, and C. Schwab , Quasi- Monte Carlo Bayesian estimation under Besov priors in elliptic inverse problems , Math. Comput., 90 (2021), pp. 1831--1860

    2021

  18. [26]

    P. L. Houtekamer and H. L. Mitchell , Data Assimilation Using an Ensemble Kalman Filter Technique , Mon. Weather Rev., 126 (1998), pp. 796--811

    1998

  19. [27]

    D. Z. Huang, T. Schneider, and A. M. Stuart , Iterated Kalman methodology for inverse problems , J. Comput. Phys., 463 (2022), p. 111262

    2022

  20. [28]

    M. A. Iglesias, K. J. H. Law, and A. M. Stuart , Ensemble Kalman methods for inverse problems , Inverse Probl., 29 (2013), p. 045001

    2013

  21. [29]

    S. Jin, L. Li, and J.-G. Liu , Random Batch Methods ( RBM ) for interacting particle systems , J. Comput. Phys., 400 (2020), p. 108877

    2020

  22. [30]

    J. P. Kaipio and E. Somersalo , Statistical and Computational Inverse Problems , vol. 160 of Applied Mathematical Sciences , Springer, New York, NY, 2005

    2005

  23. [31]

    M. Ko , S. Weissmann, and J. Zech , On the mean field limit of consensus based methods , arXiv:2409.03518, (2024)

    2024 arXiv

  24. [32]

    N. B. Kovachki and A. M. Stuart , Ensemble Kalman inversion: A derivative-free technique for machine learning tasks , Inverse Probl., 35 (2019), p. 095005

    2019

  25. [33]

    Y. T. Lee, R. Shen, and K. Tian , Structured Logconcave Sampling with a Restricted Gaussian Oracle , in Proceedings of Thirty Fourth Conference on Learning Theory , PMLR, July 2021, pp. 2993--3050

    2021

  26. [34]

    Leimkuhler, C

    B. Leimkuhler, C. Matthews, and J. Weare , Ensemble preconditioning for Markov chain Monte Carlo simulation , Stat. Comput., 28 (2018), pp. 277--290

    2018

  27. [35]

    M. B. Lykkegaard, T. J. Dodwell, C. Fox, G. Mingas, and R. Scheichl , Multilevel Delayed Acceptance MCMC , SIAM/ASA J. Uncertainty Quantif., 11 (2023), pp. 1--30

    2023

  28. [36]

    Y.-A. Ma, T. Chen, and E. B. Fox , A complete recipe for stochastic gradient MCMC , in Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, NIPS '15, Cambridge, MA, USA, Dec. 2015, MIT Press, pp. 2917--2925

    2015

  29. [37]

    Metropolis, A

    N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller , Equation of State Calculations by Fast Computing Machines , J. Chem. Phys., 21 (1953), pp. 1087--1092

    1953

  30. [38]

    P. D. Miller , Applied Asymptotic Analysis , vol. 75 of Graduate studies in mathematics, American Mathematical Soc., 2006

    2006

  31. [39]

    Morzfeld, X

    M. Morzfeld, X. T. Tong, and Y. M. Marzouk , Localization for MCMC : Sampling high-dimensional posterior distributions with local structure , J. Comput. Phys., 380 (2019), pp. 1--28

    2019

  32. [40]

    N \"u sken and S

    N. N \"u sken and S. Reich , Note on Interacting Langevin Diffusions : Gradient Structure and Ensemble Kalman Sampler by Garbuno-Inigo , Hoffmann , Li and Stuart , arXiv:1908.10890, (2019)

    2019 arXiv

  33. [41]

    Osher, H

    S. Osher, H. Heaton, and S. Wu Fung , A Hamilton-Jacobi-based proximal operator , Proc. Natl. Acad. Sci. U.S.A., 120 (2023), p. e2220469120

    2023

  34. [42]

    E. Ott, B. R. Hunt, I. Szunyogh, A. V. Zimin, E. J. Kostelich, M. Corazza, E. Kalnay, D. Patil, and J. A. Yorke , A local ensemble Kalman filter for atmospheric data assimilation , Tellus A: Dyn. Meteorol. Oceanogr., 56 (2004), pp. 415--428

    2004

  35. [43]

    G. A. Pavliotis, A. M. Stuart, and U. Vaes , Derivative- Free Bayesian Inversion Using Multiscale Dynamics , SIAM J. Appl. Dyn. Syst., 21 (2022), pp. 284--326

    2022

  36. [44]

    Pereyra , Proximal Markov chain Monte Carlo algorithms , Stat

    M. Pereyra , Proximal Markov chain Monte Carlo algorithms , Stat. Comput., 26 (2016), pp. 745--760

    2016

  37. [45]

    K. B. Petersen and M. S. Pedersen , The Matrix Cookbook , 2006

    2006

  38. [46]

    Pinnau, C

    R. Pinnau, C. Totzeck, O. Tse, and S. Martin , A consensus-based model for global optimization and its mean-field limit , Math. Models Methods Appl. Sci., 27 (2017), pp. 183--204

    2017

  39. [47]

    Reich and S

    S. Reich and S. Weissmann , Fokker-- Planck Particle Systems for Bayesian Inference : Computational Approaches , SIAM/ASA J. Uncertainty Quantif., 9 (2021), pp. 446--482

    2021

  40. [48]

    Riedl, T

    K. Riedl, T. Klock, C. Geldhauser, and M. Fornasier , Gradient is All You Need ? , arXiv:2306.09778, (2023)

    2023

  41. [49]

    G. O. Roberts and R. L. Tweedie , Exponential convergence of Langevin distributions and their discrete approximations , Bernoulli, 2 (1996), pp. 341--363

    1996

  42. [50]

    R. T. Rockafellar and R. J. B. Wets , Variational Analysis , vol. 317 of Grundlehren der mathematischen Wissenschaften , Springer, Berlin, Heidelberg, 1998

    1998

  43. [51]

    S \"a rkk \"a and A

    S. S \"a rkk \"a and A. Solin , Applied Stochastic Differential Equations , Institute of Mathematical Statistics Textbooks , Cambridge University Press, Cambridge, 2019

    2019

  44. [52]

    Schillings and A

    C. Schillings and A. M. Stuart , Analysis of the Ensemble Kalman Filter for Inverse Problems , SIAM J. Numer. Anal., 55 (2017), pp. 1264--1290

    2017

  45. [53]

    Schillings, C

    C. Schillings, C. Totzeck, and P. Wacker , Ensemble- Based Gradient Inference for Particle Methods in Optimization and Sampling , SIAM/ASA J. Uncertainty Quantif., 11 (2023), pp. 757--787

    2023

  46. [54]

    Seelinger, A

    L. Seelinger, A. Reinarz, L. Rannabauer, M. Bader, P. Bastian, and R. Scheichl , High performance uncertainty quantification with parallelized multilevel Markov chain Monte Carlo , in Proceedings of the International Conference for High Performance Computing , Networking , Sto...

    2021

  47. [55]

    A. M. Stuart , Inverse problems: A Bayesian perspective , Acta Numer., 19 (2010), pp. 451--559

    2010

  48. [56]

    Sznitman , Topics in propagation of chaos , in Ecole d' Et \'e de Probabilit \'e s de Saint-Flour XIX --- 1989, D

    A.-S. Sznitman , Topics in propagation of chaos , in Ecole d' Et \'e de Probabilit \'e s de Saint-Flour XIX --- 1989, D. L. Burkholder, E. Pardoux, A.-S. Sznitman, and P.-L. Hennequin, eds., Berlin, Heidelberg, 1991, Springer, pp. 165--251

    1989

  49. [57]

    M. K. Titsias and O. Papaspiliopoulos , Auxiliary Gradient-Based Sampling Algorithms , J. R. Stat. Soc. B, 80 (2018), pp. 749--767

    2018

  50. [58]

    X. T. Tong and M. Morzfeld , Localized ensemble kalman inversion , Inverse Probl., 39 (2023), p. 064002

    2023

  51. [59]

    Wacker , Perspectives on locally weighted ensemble Kalman methods , arXiv:2402.00027, (2024)

    P. Wacker , Perspectives on locally weighted ensemble Kalman methods , arXiv:2402.00027, (2024)

    2024 arXiv

  52. [60]

    Wagner, I

    F. Wagner, I. Papaioannou, and E. Ullmann , The Ensemble Kalman Filter for Rare Event Estimation , SIAM/ASA J. Uncertainty Quantif., 10 (2022), pp. 317--349

    2022

  53. [61]

    S. Yang, Y. Chen, E. Bernton, and J. S. Liu , On parallelizable Markov chain Monte Carlo algorithms with waste-recycling , Stat. Comput., 28 (2018), pp. 1073--1081

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.