Ensemble Kalman Inversion: mean-field limit and convergence analysis

Qin Li; Zhiyan Ding

REVIEW 4 major objections 4 minor 36 references

Ensemble Kalman Inversion: mean-field limit and convergence analysis

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper proves that the particle ensemble in continuous Ensemble Kalman Inversion converges, as the number of particles grows, to a Fokker-Planck equation with explicit Wasserstein-2 rates, and that in the linear Gaussian case the limit…

desk verdict The mean-field question is the right one and the linear calculation is clean, but the headline rate is unsupported: the two bootstrap inequalities live in a missing supplement and Fournier-Guillin is misquoted. read the letter →

arxiv 1908.05575 v5 pith:ZYMFG4O4 submitted 2019-08-15 math.NA cs.NAmath.PR

Zhiyan Ding , Qin Li This is my paper · ORCID

classification math.NAcs.NAmath.PR MSC 65C3560H1062F1535Q84

keywords EnsembleKalmanInversionmean-fieldlimitFokker-PlanckequationWassersteinmetricBayesianinverseproblemsstochasticdifferentialequationsconvergencerate

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 establishes the many-particle, or mean-field, limit of the continuous-time Ensemble Kalman Inversion (EKI), an algorithm widely used to sample from Bayesian posterior distributions. Its main theorem says that as the number of particles $J$ grows, the empirical ensemble converges in Wasserstein-2 distance to the solution of a Fokker-Planck equation, at rate essentially $J^{-1/2+\epsilon}$ when the parameter dimension is at most 4 and $J^{-2/L}$ in higher dimension. In the linear setting with Gaussian prior, the limiting equation at pseudo-time 1 is exactly the posterior density, so the result explains in what sense EKI produces approximate posterior samples. Why this matters: the theorem reduces a stochastic algorithm with unknown behavior to a deterministic PDE that can be studied with standard PDE tools, and it gives a precomputable cost for choosing the number of particles.

What carries the argument

The argument couples the self-interacting particle system to a 'bridge' system $\{v^j_t\}$ whose coefficients are frozen to those of the PDE solution $\rho(t,u)$; the distance between the bridge and the PDE is controlled by a classical empirical-measure concentration estimate, while the distance between the bridge and the original particles is controlled by SDE stability. The proof then applies a bootstrapping procedure: starting from boundedness of moments, two lemmas (Lemmas 7 and 8) show that an assumed decay rate $J^{-\alpha}$ for the particle difference can be tightened to $J^{-(1+\alpha)/2+\epsilon}$, and iterating to saturation yields the rates in Theorem 1.

What would settle it

Work through the deferred calculation that produces (54): if the displayed exponent cannot be obtained from the preceding moment bounds, Proposition 2 and Theorem 1 lose their proof. A complementary numerical check: for a weakly nonlinear map in dimension $L=5$, estimate $E[W_2(M_t^u,\rho(t))]$ for growing $J$ and compare the slope with $J^{-2/L}$; an empirical rate slower than $J^{-2/L}$ would contradict the claimed rate.

Watch

Extended reading notes

Core claim

Under the weakly nonlinear structural assumption $G(u)=Au+m(u)$, where the perturbation $m$ is bounded, smooth, and orthogonal in the $\Gamma^{-1}$-metric to the range of $A$, the coupled SDE system (8) has a mean-field limit: the empirical measure $M_t^u$ converges to the probability distribution induced by the strong solution $\rho(t,u)$ of the Fokker-Planck equation (9). The convergence rate is $E(W_2(M_t^u,\rho(t))) \le C_\epsilon(t)\{J^{-1/2+\epsilon}, L\le 4;\ J^{-2/L}, L>4\}$, and a separate weak-convergence statement gives the dimension-independent rate $J^{-1/2+\epsilon}$ for Lipschitz test functions. When $G$ is linear and the prior is Gaussian, $\rho(1,u)$ equals the posterior density, so stopping the algorithm at pseudo-time 1 reconstructs the target distribution.

Load-bearing premise

The load-bearing premise is that the bootstrap inequalities (54) and (59), whose derivation is deferred to supplementary calculations, actually follow from the stated assumptions, and that strong solutions to the SDE and the nonlinear Fokker-Planck equation exist in the weakly nonlinear case.

Editorial extensions

If this is right

Analysis of EKI reduces to analysis of the Fokker-Planck equation (9): well-posedness, equilibration, and long-time behavior of the PDE transfer directly to the algorithm's many-particle behavior.
For parameter dimension $L\le4$, the rate $J^{-1/2+\epsilon}$ is essentially the Monte Carlo rate, so EKI is asymptotically as efficient as an i.i.d. particle approximation of its own limit.
In the linear Gaussian case, Corollary 1 is a finite-time guarantee: for sufficiently many particles, the ensemble at pseudo-time 1 is within any prescribed $\epsilon$ of the posterior in Wasserstein-2 distance.
In the weakly nonlinear case, the limiting PDE is not the posterior; the paper derives explicit residual terms $R_1,R_2,R_3$ quantifying the deviation, so the bias of EKI as a sampling method is at least characterized.
The weak-convergence statement (Theorem 2) removes the dimension dependence for observables, which is the practically relevant notion: the ensemble approximates expectations of Lipschitz functions at rate $J^{-1/2+\epsilon}$ regardless of dimension.

Reading between the lines

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

The threshold $L\le4$ versus $L>4$ in Theorem 1 mirrors the known dimension-dependence of empirical-measure quantization in Wasserstein distance, suggesting that for $L>4$ the dominant error is the intrinsic approximation of the limiting measure by an empirical measure, not the particle interaction; if so, any sampling method that initializes from i.i.d. prior samples would face the same $J^{-2/L}
The bootstrapping mechanism that tightens polynomial decay rates is a general tool: it could be adapted to prove mean-field limits for other ensemble algorithms whose coefficients depend on empirical moments, such as ensemble Kalman samplers or consensus-based optimization methods.
The paper reveals that the nonlinear case has nonzero residuals $R_i$ separating the FP limit from the posterior; a natural next step is to compute these residuals for concrete weakly nonlinear maps and determine whether they remain $O(1)$ at $t=1$, which would quantify the sampling bias.
Because the result is proven for the continuous-time SDE, the discrete-time algorithm with step size $h=1/N$ is not directly covered; extending the proof to the joint limit $h\to0$, $J\to\infty$ would connect the theorem to the actually implemented EKI.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things to know. This is the first paper I know that takes the J→∞ limit of the EKI SDE seriously, and it identifies the right object: a Fokker-Planck equation for the law of the mean-field limit. In the linear case, the calculation showing that the FP solution interpolates from the prior at t=0 to the posterior at t=1 is concrete and, as far as I know, new. The weakly nonlinear section is honest too: the FP limit deviates from the Bayesian posterior by explicit weight terms.

What is not in good shape is the main theorem's rate. The whole proof of Proposition 2—the particle-to-particle closeness that feeds Theorem 1—rests on two inequalities, labeled (54) and (59). Each is introduced with 'With the calculation shown in Supp. A/B', and no such calculation appears anywhere in the arXiv v5 text. These are not minor technical estimates. They bound the time derivative of the ensemble moment d/dt (1/J Σ E|x̃^j|²) and mix powers of several norms with J-powers. The bootstrapping scheme in Lemma 8 uses them to sharpen α to 1/2 + α/2 − ε. Without them, the claimed E|x^j|² ≤ C_ε J^{-1+ε} has no derivation, and Theorem 1 is a sketch, not a proof.

There is also an independent rate problem. The paper cites Fournier-Guillin (their Theorem 3) as if it bounded E[W_p(ρ_N,ρ)] directly. Their theorem bounds E[W_p^p]. For W₂, that lowers the low-dimensional rate from J^{-1/2} to J^{-1/4} (and in high dimensions from J^{-2/L} to J^{-1/L}). So the 'optimal' rates in Proposition 1 are not supported by the cited result, even setting the missing supplement aside.

What deserves credit: the linear reconstruction corollary is independent of the rate analysis and appears correct. The moment lemmas are honest work. The paper is written clearly and the bootstrapping idea is sensible—it just needs to be carried out in the text.

Bottom line: this is a paper with the right question and a real partial result, but the main theorem is currently unproven as stated. If the authors supply the missing calculations and correct the rates, it would be a solid contribution. I would not accept it in this form, but I would send it to a referee rather than desk-reject: the problem is important, the approach is right, and the linear result deserves an outlet. The referee should insist on seeing Supplement A/B and a corrected citation.

Referee Report

4 major / 4 minor

Summary. The paper analyzes the continuous-time limit of Ensemble Kalman Inversion (EKI), a particle method for Bayesian inverse problems. In the J→∞ limit, the empirical measure of the coupled SDE system (8) is claimed to converge to the solution of a nonlinear Fokker-Planck equation (9) in the Wasserstein-2 metric, with rates J^{-1/2+ε} for L≤4 and J^{-2/L} for L>4 (Theorem 1). In the linear case, the limit at t=1 is the posterior distribution (Corollary 1). The proof strategy is a Dobrushin-type comparison: first, a 'bridge' system {v^j} following the PDE flow is compared with ρ using the Fournier-Guillin empirical-measure bound (Proposition 1); then the original system {u^j} is compared with {v^j} via SDE stability and a bootstrapping argument (Proposition 2). The linear and weakly nonlinear cases are treated under assumption (6).

Significance. If fully proven, the result would be a significant contribution to the theory of EKI, providing the first quantitative mean-field limit with near-optimal rates and a rigorous finite-time posterior reconstruction in the linear case. The paper contains useful moment estimates (Lemmas 1-5) and an explicit computation showing that the candidate Gaussian transition (11) solves the FP equation in the linear case. However, the central rate estimates depend on two bootstrap inequalities whose derivations are missing, so the main theorem is not currently substantiated.

major comments (4)

[Section 6.2, Lemmas 7-8, Eqs. (54) and (59)] The inequalities (54) and (59) are the only input to the recursive rate improvement α_n = 1/2 + α_{n-1}/2 - ε in the proof of Proposition 2, but they are asserted to follow from calculations in 'Supp. A' and 'Supp. B', which do not appear in the submitted manuscript. The existing Appendices A and B contain different material (a moment-summation lemma and high-moment bounds for {u^j}), so this is not a matter of mislabeling. Without a derivation of (54) and (59), the bound E|u^j_t - v^j_t|^2 ≤ C_ε J^{-1+ε} in Proposition 2 has no proof, and Theorem 1, which is a direct triangle-inequality consequence, is unsupported.
[Section 2.2 and Theorem 1] The manuscript assumes, without proof, the existence and uniqueness of a strong solution to the Fokker-Planck equation (9) in the weakly nonlinear case; the text states that this is 'beyond the focus of the current paper'. Theorem 1 and Proposition 1 are therefore conditional on an unstated regularity hypothesis. Either prove well-posedness or state it explicitly as an assumption in the theorem statements.
[Section 3, optimality discussion after Theorem 1] The claim that the rate J^{-2/L} for L>4 is 'optimal' is not established. The cited theorem [17] gives an upper bound on the rate of convergence of empirical measures; it does not provide a matching lower bound. The optimality assertion should be removed or supported by a lower-bound argument.
[Section 5.2, Theorem 3] One possible concern about the use of the Fournier-Guillin result is that the theorem bounds E[W_p^p] rather than E[W_p]. After checking the statement in [17], I do not find this to be the case: Theorem 3 as quoted matches the standard form of Theorem 1 of [17], which bounds E(W_p) directly. The rates in Proposition 1 are therefore not affected by a square-root factor on this account.

minor comments (4)

[Section 2.1] The phrase 'It is not our intension' should be 'It is not our intention'.
[Section 4.1] In the derivation of the linear case, the notation Γ_0^{-1} is used without prior definition; presumably it denotes the inverse of the prior covariance matrix.
[Lemma 3, equation (21)] In the second inequality of (21), the expression |1/J Σ_j |q^j_t|^2 − Tr(Cov_{ρ_t})| should be written with absolute values enclosing the entire difference, e.g., |(1/J Σ_j |q^j_t|^2) − Tr(Cov_{ρ_t})|, to avoid ambiguity.
[Throughout] The companion paper [10] is cited as a source of related results; the manuscript would benefit from a sentence specifying which results are taken from [10] and which are new here.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the main theorem is a standard coupling/bootstrapping argument; the flagged gaps (missing Supp. A/B, Fournier-Guillin optimality reading) are completeness/correctness defects, not circular reductions, and the self-citation [10] is not load-bearing.

full rationale

The proof chain is not circular. Theorem 1 is obtained by the triangle inequality from Proposition 1 (empirical measure of the bridge {v^j} versus the Fokker-Planck solution rho) and Proposition 2 (the two particle systems {u^j} and {v^j}): the former applies the external Fournier-Guillin rate Theorem 3 to i.i.d. samples with moments bounded in Lemmas 1-3, and the latter is a bootstrap whose base alpha_0=0 uses the boundedness proved in Lemmas 4-5. The only serious gaps are inequalities (54) and (59), introduced in Lemmas 7-8 with the phrases 'With the calculation shown in Supp. A' and 'With the calculation in Supp. B and Lemma 7'; no Supp. A/B appears in the v5 text. Missing derivations are an omitted-proof/completeness defect, not a circular step, because the text never assumes the target rate J^{-1+epsilon} in order to prove it. Corollary 1 is a direct verification that the Gaussian bridge mu(t,u) in (11) balances the terms of (9); it does not presuppose Theorem 1. The linear-case rho(t)=mu(t) check is a calculation, not a fit. The authors' self-citation [10] is used only in the introduction when listing EKS and as a pointer to further nonlinear discussion in Section 4.2, so it is not load-bearing. The statement that J^{-2/L} is 'the best possible' rate reads the Fournier-Guillin upper bound as a lower bound, which is a correctness concern stemming from an external theorem, not a circular reduction. Overall, no prediction reduces by construction to a fitted input or to a self-citation; score 2 reflects only the minor non-load-bearing self-citation.

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

The central claim rests on a structural assumption about G, on the existence of solutions to an SDE and a nonlinear PDE that is not proved, and on external theorems (Fournier-Guillin) which are misquoted. There are no fitted numerical parameters and no invented physical entities.

assumptions (5)

domain assumption Weakly nonlinear structure (6): G = A u + m(u), with Range(m) ⊥_{Γ^{-1}} Range(A) and |m| + |∇m| ≤ M.
This confines the analysis to a restricted class of nonlinear maps; it is used throughout to eliminate cross terms and to bound the noise terms. It does not cover generic nonlinear forward maps.
domain assumption Existence and uniqueness of strong solution to the SDE system (8) for the weakly nonlinear case.
The paper states in Section 2.2 that wellposedness for the nonlinear case 'should work' but is 'beyond the focus'; the mean-field limit proof operates on this solution.
domain assumption Existence and uniqueness of a strong solution to the nonlinear Fokker-Planck equation (9).
Theorem 1 refers to ρ(t,u) as the strong solution, but the paper provides no existence or uniqueness proof for this nonlinear FP equation.
domain assumption Prior density µ0 is C^2 and has finite moments of all orders.
Stated in Theorem 1 and used in the moment bound lemmas (Lemma 2, Lemma 5).
domain assumption Initial particles are drawn i.i.d. from µ0 and the measurement noise is η ~ N(0, Γ).
Standard EKI Bayesian setup; defines the initial condition and the noise term in the SDE.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ensemble Kalman Inversion: mean-field limit and convergence analysis." pith.science (2026). https://pith.science/paper/ZYMFG4O4

@misc{pith2026190805575,
  author       = {Pith},
  title        = {Pith review of: Ensemble Kalman Inversion: mean-field limit and convergence analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZYMFG4O4}},
  note         = {Machine review of arXiv:1908.05575}
}

abstract

Ensemble Kalman Inversion (EKI) has been a very popular algorithm used in Bayesian inverse problems. It samples particles from a prior distribution, and introduces a motion to move the particles around in pseudo-time. As the pseudo-time goes to infinity, the method finds the minimizer of the objective function, and when the pseudo-time stops at $1$, the ensemble distribution of the particles resembles, in some sense, the posterior distribution in the linear setting. The ideas trace back further to Ensemble Kalman Filter and the associated analysis, but to today, when viewed as a sampling method, why EKI works, and in what sense with what rate the method converges is still largely unknown. In this paper, we analyze the continuous version of EKI, a coupled SDE system, and prove the mean field limit of this SDE system. In particular, we will show that 1. as the number of particles goes to infinity, the empirical measure of particles following SDE converges to the solution to a Fokker-Planck equation in Wasserstein 2-distance with an optimal rate, for both linear and weakly nonlinear case; 2. the solution to the Fokker-Planck equation reconstructs the target distribution in finite time in the linear case.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 34 canonical work pages

[17]

Fournier and A

N. Fournier and A. Guillin. On the rate of convergence in wasserstein distance of the empirical measure. Probability Theory and Related Fields , 162(3):707–738, Aug 2015

work page 2015
[1]

Bergemann and S

K. Bergemann and S. Reich. A localization technique for ensemble kalman ﬁlters. Quarterly Journal of the Royal Meteorological Society, 136(648):701–707, 2010

work page 2010
[2]

Bergemann and S

K. Bergemann and S. Reich. A molliﬁed ensemble kalman ﬁlter. Quarterly Journal of the Royal Meteorological So- ciety, 136(651):1636–1643, 2010

work page 2010
[3]

Bloemker, C

D. Bloemker, C. Schillings, P. Wacker, and S. Weissmann. Well posedness and convergence analysis of the ensemble kalman inversion. Inverse Problems , 2019

work page 2019
[4]

Blomker, C

D. Blomker, C. Schillings, and P. Wacker. A strongly con- vergent numerical scheme from ensemble kalman inver- sion. SIAM Journal on Numerical Analysis , 56(4):2537– 2562, 2018

work page 2018
[5]

Bolley, J

F. Bolley, J. A. Ca˜ nizo, and J. A. Carrillo. Stochas- tic mean-ﬁeld limit: Non-Llipschitz forces and swarming. Mathematical Models and Methods in Applied Sciences , 21(11):2179–2210, 2011

work page 2011
[6]

J. A. Ca˜ nizo, J. A. Carrillo, and J. Rosado. A well- posedness theory in measures for some kinetic models of collective motion. Mathematical Models and Methods in Applied Sciences , 21(03):515–539, 2011

work page 2011
[7]

Craig and A

K. Craig and A. Bertozzi. A blob method for the ag- gregation equation. Mathematics of Computation , 85, 05 2014

work page 2014

Show all 36 references

[8]

Dashti and A

M. Dashti and A. M. Stuart. The Bayesian Approach to Inverse Problems . Springer International Publishing, Cham, 2017

2017
[9]

De Moral

P. De Moral. Feynman-Kac Formulae: Genealogical and Interacting Particle Approximations . Springer-Verlag, 2004

2004
[10]

Ding and Q

Z. Ding and Q. Li. Ensemble kalman sampling: mean- ﬁeld limit and convergence analysis. arXiv: 1910.12923 , 2019

1910 arXiv
[11]

Z. Ding, Q. Li, and J. Lu. Ensemble kalman inversion for nonlinear problems: weights, consistency, and vari- ance bounds, 2020

2020
[12]

Doucet, N

A. Doucet, N. de Freitas, and N. Gordon. An Introduction to Sequential Monte Carlo Methods . Springer New York, New York, NY, 2001

2001
[13]

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 Un- certainty Quantiﬁcation , 3(1):823–851, 2015

2015
[14]

G. Evensen. Sequential data assimilation with a nonlin- ear quasi-geostrophic model using monte carlo methods to forecast error statistics. Journal of Geophysical Re- search: Oceans, 99(C5):10143–10162, 1994

1994
[15]

G. Evensen. The ensemble kalman ﬁlter: theoretical for- mulation and practical implementation. Ocean Dynam- ics, 53(4):343–367, Nov 2003

2003
[16]

G. Evensen. Data Assimilation-The Ensemble Kalman Filter. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2006

2006
[18]

Garbuno-Inigo, F

A. Garbuno-Inigo, F. Hoﬀmann, W. Li, and A. M. Stuart. Interacting langevin diﬀusions: Gradient structure and ensemble kalman sampler. arXiv:1903.08866, 2019

1903 arXiv
[19]

M. Ghil, S. Cohn, J. Tavantzis, K. Bube, and E. Isaac- son. Applications of Estimation Theory to Numerical Weather Prediction. Springer New York, New York, NY, 1981

1981
[20]

Herty and G

M. Herty and G. Visconti. Kinetic methods for inverse problems. Kinetic & Related Models , 12:1109, 2019

2019
[21]

P. L. Houtekamer and Herschel L. Mitchell. A sequential ensemble kalman ﬁlter for atmospheric data assimilation. Monthly Weather Review , 129(1):123–137, 2001

2001
[22]

M. A. Iglesias, K. Law, and A. M. Stuart. Ensemble kalman methods for inverse problems. Inverse Problems , 29(4):045001, Mar 2013

2013
[23]

Lange and W

T. Lange and W. Stannat. On the continuous time limit of the ensemble kalman ﬁlter, 2019

2019
[24]

K. J. H. Law, H. Tembine, and R. Tempone. Determinis- tic mean-ﬁeld ensemble kalman ﬁltering. SIAM Journal on Scientiﬁc Computing , 38(3):A1251–A1279, 2016

2016
[25]

Le Gland, V

F. Le Gland, V. Monbet, and V. Tran. Large sample asymptotics for the ensemble kalman ﬁlter. Handbook on Nonlinear Filtering , 2011

2011
[26]

Liu and D

Q. Liu and D. Wang. Stein variational gradient de- scent: A general purpose bayesian inference algorithm. In Advances in Neural Information Processing Systems 29, pages 2378–2386. 2016

2016
[27]

J. Lu, Y. Lu, and J. Nolen. Scaling limit of the stein vari- ational gradient descent: The mean ﬁeld regime. SIAM Journal on Mathematical Analysis , 51(2):648–671, 2019

2019
[28]

Y. Lu, J. Lu, and J. Nolen. Accelerating langevin sam- pling with birth-death, 2019

2019
[29]

Pavon, E

M. Pavon, E. G. Tabak, and G. Trigila. The data-driven schroedinger bridge. arXiv:1806.01364, Jun 2018

2018 arXiv
[30]

S. Reich. A dynamical systems framework for intermit- tent data assimilation. BIT Numerical Mathematics , 51(1):235–249, Mar 2011

2011
[31]

S. Reich. Data assimilation: The schrdinger perspectiv e. Acta Numerica, 28:635711, 2019

2019
[32]

C. P. Robert and G. Casella. Monte Carlo Statistical Methods. 2nd ed. Springer, New York, 2004

2004
[33]

Schillings and A

C. Schillings and A. M. Stuart. Analysis of the ensem- ble kalman ﬁlter for inverse problems. SIAM J. Numer. Anal, 55(3):1264–1290, 2017

2017
[34]

Schillings and A

C. Schillings and A. M. Stuart. Convergence analysis of ensemble kalman inversion: the linear, noisy case. Appli- cable Analysis , 97(1):107–123, 2018

2018
[35]

A. M. Stuart. Inverse problems: A bayesian perspective. Acta Numerica, 19:451559, 2010

2010
[36]

Sznitman

A. Sznitman. Topics in propagation of chaos. In Ecole d’Et´ e de Probabilit´ es de Saint-Flour XIX — 1989, pages 165–251. Springer Berlin Heidelberg, 1991. A Moments bound of summation of indepedent mean-zero random variables In this section, we prove a lemma which is used in ...

1989

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [17]

Fournier and A

N. Fournier and A. Guillin. On the rate of convergence in wasserstein distance of the empirical measure. Probability Theory and Related Fields , 162(3):707–738, Aug 2015

work page 2015

[2] [1]

Bergemann and S

K. Bergemann and S. Reich. A localization technique for ensemble kalman ﬁlters. Quarterly Journal of the Royal Meteorological Society, 136(648):701–707, 2010

work page 2010

[3] [2]

Bergemann and S

K. Bergemann and S. Reich. A molliﬁed ensemble kalman ﬁlter. Quarterly Journal of the Royal Meteorological So- ciety, 136(651):1636–1643, 2010

work page 2010

[4] [3]

Bloemker, C

D. Bloemker, C. Schillings, P. Wacker, and S. Weissmann. Well posedness and convergence analysis of the ensemble kalman inversion. Inverse Problems , 2019

work page 2019

[5] [4]

Blomker, C

D. Blomker, C. Schillings, and P. Wacker. A strongly con- vergent numerical scheme from ensemble kalman inver- sion. SIAM Journal on Numerical Analysis , 56(4):2537– 2562, 2018

work page 2018

[6] [5]

Bolley, J

F. Bolley, J. A. Ca˜ nizo, and J. A. Carrillo. Stochas- tic mean-ﬁeld limit: Non-Llipschitz forces and swarming. Mathematical Models and Methods in Applied Sciences , 21(11):2179–2210, 2011

work page 2011

[7] [6]

J. A. Ca˜ nizo, J. A. Carrillo, and J. Rosado. A well- posedness theory in measures for some kinetic models of collective motion. Mathematical Models and Methods in Applied Sciences , 21(03):515–539, 2011

work page 2011

[8] [7]

Craig and A

K. Craig and A. Bertozzi. A blob method for the ag- gregation equation. Mathematics of Computation , 85, 05 2014

work page 2014