{"id":"7b766cb3-e914-41bf-ae62-8934a5a39e26","arxiv_id":"1908.05575","paper_version":5,"verdict":"REJECT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Ding and Li analyze the mean-field limit of the continuous-time Ensemble Kalman Inversion SDE and claim optimal Wasserstein-2 convergence rates to a Fokker-Planck equation, but the rates and a key proof step are unsupported.","lead":"This paper claims to prove that Ensemble Kalman Inversion (EKI), a popular method for Bayesian inverse problems, converges in the large-ensemble limit to a Fokker-Planck equation, and that in the linear case this equation reconstructs the posterior at finite time. The paper is a theoretical analysis; its main quantitative convergence rates rest on a misquoted external theorem and on calculations relegated to a missing supplement.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing Supp. A/B derivations for bootstrap inequalities (54) and (59) leave Proposition 2 and hence Theorem 1 unproven.","rationale":"The reader's weakest_assumption pinpoints the same missing calculation; my independent reading confirms that (54) and (59) are the load-bearing input. I considered whether the Fournier-Guillin rate issue was more central, but Proposition 1 follows from Theorem 3 once q is chosen large, so the rate objection is less decisive. The missing supplement is a verification gap, not a refutation: if the inequalities are true, the bootstrap fixed point yields the claimed J^{-1+ε} particle closeness and the theorem could be repaired by adding the supplement. But as submitted, the proof is not complete, so the REJECT verdict stands without change.","tokens_in":28799,"tokens_out":9879,"duration_ms":85666,"concrete_test":"Independently derive inequality (54) by applying Itô's formula to |x̃^j|^2 using the difference of equations (52) and (53), and verify that every term on the right-hand side matches (54)'s J-powers and ε-exponents, especially the leading J^{-1/4} factor multiplying the (E|x^1|^2)^{1-ε} terms. If the calculation cannot be reproduced, or if any exponent differs, the bootstrap rate in Proposition 2 is invalid. As a quicker check, request Supp. A/B from the authors; absence of a derivable supplement would confirm the v5 proof is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2, the engine of the main theorem, rests entirely on the bootstrap inequalities (54) and (59). In Lemma 7, inequality (54) is introduced with the phrase 'With the calculation shown in Supp. A'; in Lemma 8, inequality (59) is introduced with 'With the calculation in Supp. B and Lemma 7'. No Supp. A or Supp. B appears in arXiv:1908.05575v5. These are not trivial estimates: they are Itô-expansion bounds on d(1/J Σ_j E|x̃^j|^2)/dt that mix powers of E|x^1|^2, E|p^1|^2, E|p̃^1|^2 with J-powers and an ε-exponent, and they are the sole input to the recursive rate improvement α_n = 1/2 + α_{n-1}/2 - ε. Without (54) and (59), the bound E|u^j_t - v^j_t|^2 ≤ C_ε J^{-1+ε} in Proposition 2 has no derivation, and Theorem 1, which is a direct triangle-inequality consequence of Propositions 1 and 2, collapses. The paper contains no proof, no reference to a public supplement, and no sketch; the central quantitative claim is therefore not verifiable from the current text. A secondary gap, also noted in the paper itself (Section 2.2), is that strong well-posedness of the nonlinear FP equation (9) is assumed rather than proved, with only the linear case referenced from [3]; however, even granting well-posedness, the missing bootstrap calculation is fatal to the proof as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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).","tokens_in":29151,"tokens_out":13198,"duration_ms":113058,"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":[{"comment":"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":"Section 6.2, Lemmas 7-8, Eqs. (54) and (59)"},{"comment":"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":"Section 2.2 and Theorem 1"},{"comment":"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":"Section 3, optimality discussion after Theorem 1"},{"comment":"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.","section":"Section 5.2, Theorem 3"}],"minor_comments":[{"comment":"The phrase 'It is not our intension' should be 'It is not our intention'.","section":"Section 2.1"},{"comment":"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.","section":"Section 4.1"},{"comment":"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.","section":"Lemma 3, equation (21)"},{"comment":"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.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle to acceptance is the missing supplement containing the derivations of (54) and (59). If the authors can provide those calculations and they are correct, the paper could be publishable. Please also ask the authors to make the well-posedness assumption explicit in the theorem statements. I would not recommend rejection at this stage, provided the missing derivations are supplied and the optimality claim is either justified or removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat 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.\n\nThere 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.\n\nWhat 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.\n\nBottom 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.","headline":"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.","tokens_in":29655,"tokens_out":5874,"would_cite":false,"duration_ms":52818,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C35","60H10","62F15","35Q84"],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["Ensemble Kalman Inversion","mean-field limit","Fokker-Planck equation","Wasserstein metric","Bayesian inverse problems","stochastic differential equations","convergence rate"],"falsifier":"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.","tokens_in":28618,"feed_emoji":"🎲","tokens_out":14177,"duration_ms":118356,"temperature":0.7,"pith_summary":"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.","feed_headline":"EKI particles converge to a PDE limit at a proven rate","feed_subtitle":"It proves when and how fast the ensemble approximates the posterior in linear and weakly nonlinear inverse problems.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"supplies the empirical-measure Wasserstein concentration estimate (cited as Theorem 3) that controls the distance between the bridge particle system and the Fokker-Planck solution.","marker":"[17]"},{"why":"provides the well-posedness and moment-bound framework for the ensemble Kalman inversion SDE that the a-priori estimates in Lemma 4 imitate.","marker":"[3]"},{"why":"introduces the EKI algorithm and the claim that stopping at pseudo-time 1 approximates the posterior, which the linear-case corollary proves.","marker":"[22]"},{"why":"formally derives the continuum-limit SDE and the Fokker-Planck description for inverse problems, the starting point of the paper's analysis.","marker":"[33]"},{"why":"connects EKI to the ensemble Kalman filter framework and motivates finite-time posterior reconstruction.","marker":"[30]"}],"fun_headline_variants":["EKI mean-field limit proven with optimal rates","EKI hits posterior exactly at pseudo-time 1","Mean-field limit and convergence proven for EKI","Particle-to-PDE convergence rate proven for EKI","EKI's mean-field limit: proven and quantified"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["EKI mean-field limit proven with optimal rates","EKI hits posterior exactly at pseudo-time 1","Mean-field limit and convergence proven for EKI","Particle-to-PDE convergence rate proven for EKI","EKI's mean-field limit: proven and quantified"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2886,"prompt_tokens":974,"completion_tokens":1912,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1837}},"tokens_in":590,"tokens_out":1912,"duration_ms":11930,"temperature":1.0,"reasoning_tokens":1837,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:12:26.311600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Fournier and A","cited_arxiv_id":null,"evidence_quote":"supplies the empirical-measure Wasserstein concentration estimate (cited as Theorem 3) that controls the distance between the bridge particle system and the Fokker-Planck solution."},{"cited_title":"Bloemker, C","cited_arxiv_id":null,"evidence_quote":"provides the well-posedness and moment-bound framework for the ensemble Kalman inversion SDE that the a-priori estimates in Lemma 4 imitate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the EKI algorithm and the claim that stopping at pseudo-time 1 approximates the posterior, which the linear-case corollary proves."},{"cited_title":"Schillings and A","cited_arxiv_id":null,"evidence_quote":"formally derives the continuum-limit SDE and the Fokker-Planck description for inverse problems, the starting point of the paper's analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"connects EKI to the ensemble Kalman filter framework and motivates finite-time posterior reconstruction."}],"review_version":1}