{"id":"badcf269-7793-4d96-b090-9421616f87d4","arxiv_id":"1908.05495","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A homogenization-surrogate ensemble Kalman filter converges to the exact multiscale ensemble Kalman filter at rate O(ε + h^(s+1)), with the empirical posteriors converging in Wasserstein distance.","lead":"The paper combines the ensemble Kalman filter with numerical homogenization to recover an oscillating parameter in multiscale elliptic equations from boundary measurements. It proves that the cheap homogenized version of the filter converges to the full multiscale filter as the oscillation scale and mesh size vanish, with an explicit rate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3 (all particles remain in a fixed ball around u*) has no propagation mechanism; with Gaussian data noise it fails almost surely for any fixed R, and the central convergence proof relies on it.","rationale":"The reader's weakest_assumption identified Assumption 3, and my stress-test reaches the same conclusion. The central claim of the paper is that the EnKF using the homogenized FE-HMM forward map converges to the EnKF using the exact multiscale map; this is Theorem 1, whose proof passes through Lemma 3 and Lemma 4. Both lemmas require the particles to remain in a fixed ball around the true parameter. The paper states this as Assumption 3 but provides no invariant-ball argument, and the Gaussian noise used in the algorithm makes a deterministic fixed ball impossible almost surely. A simple one-dimensional linear calculation confirms that a normally distributed observation perturbation can push a particle outside any prescribed ball with positive probability. I also noted that Lemma 4's claim that the random variables η^{(j)}_{n+1}+η are i.i.d. and independent of the ensembles is not literally correct because the same η appears at every iteration; however, under Assumption 3 the resulting bound can likely be recovered, so the primary unresolved issue remains Assumption 3. The theorem is honest in stating 'if Assumption 3 holds', so this is not a fatal inconsistency, but it is a load-bearing gap that prevents the theorem from applying to the algorithm without additional hypotheses. The reader's CONDITIONAL verdict is appropriate and I do not propose changing it.","tokens_in":24484,"tokens_out":10538,"duration_ms":117384,"concrete_test":"Consider the scalar linear test case X=Y=R, G(u)=u, Γ=1, J=2, initial ensemble {0,a} with a>0, and y=0. The sample variance is s^2=a^2/4, so the Kalman gain is K=s^2/(s^2+1)=a^2/(a^2+4)>0. The update of particle 2 is u_1^{(2)} = a(1-K)+K η_2^{(2)} with η_2^{(2)}~N(0,1). For every fixed R>0, P(|a(1-K)+K η_2^{(2)}|>R)>0, so the particle leaves B_R(0) with positive probability. This shows that Assumption 3 cannot hold almost surely with a fixed deterministic R under Gaussian data noise, and therefore the proof of Theorem 1 requires an additional bounded-noise or compactly supported prior assumption unless a different, random-radius argument is supplied.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem is conditional on Assumption 3: every particle of every ensemble must lie in a fixed ball B_R(u*) at every iteration. This assumption is used in Lemma 3 to bound C_up and C_pp and their Lipschitz constants, and in Lemma 4 to control the data-misfit factors in S1 and S2. Without it, the recursive estimate E||u^ε_{n+1}-u^0_{n+1}|| ≤ α E||u^ε_n-u^0_n|| + γ E[e(ε,u^0_n)] collapses, and with it both Proposition 1 and Proposition 2 (hence Theorem 1) collapse. The paper gives no proof that the EnKF update preserves such a ball, and in the stated Gaussian-noise setting it cannot: the update (16) contains a term K η^{(j)}_{n+1} with nonzero Kalman gain, and since η is Gaussian unbounded, for any fixed R there is positive probability that a particle leaves B_R(u*) in one step. Thus Assumption 3 is not satisfied almost surely for the algorithm as specified. A secondary but related issue is that in Lemma 4 the quantities ζ^{(j)}_{n+1}=η^{(j)}_{n+1}+η are not i.i.d. and are not independent of the ensembles, since the same η is used at every iteration; this is probably repairable using Assumption 3, but it reinforces how much of the proof leans on an invariant-ball condition that is never established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an ensemble Kalman filter (EnKF) method for inverse problems governed by multiscale elliptic PDEs, replacing the expensive multiscale forward map G^ε by a homogenized finite-element surrogate G^0_h obtained via FE-HMM. The main theoretical results are Theorem 1, an explicit-rate convergence of the EnKF ensemble with the surrogate to the ensemble with the exact multiscale map as ε,h→0, and Theorem 2, a corresponding L^1 weak/Wasserstein convergence of the empirical posterior measures. The paper also gives sample-size bounds for offline and online estimation of the modelling error (Theorems 3 and 4) and a numerical experiment recovering a discontinuous coefficient from boundary-flux observations.","tokens_in":24831,"tokens_out":11351,"duration_ms":118912,"significance":"If the convergence results hold, the paper makes a useful contribution by transferring the EnKF machinery to multiscale inverse problems with a rigorous homogenization/discretization error decomposition and explicit rates. The detailed proofs and the quantitative sample-size bounds for modelling-error estimation are strengths. However, the central theorems depend on an unverified and generically false invariant-ball assumption, and the noise handling in Lemma 4 contains a correlation error; these issues must be resolved before the main claims are reliable.","major_comments":[{"comment":"Assumption 3 is load-bearing but no mechanism is given for its propagation. In the update (16), the increment contains C_up(C_pp+Γ)^{-1} η^{(j)}_{n+1} with η^{(j)}_{n+1}∼N(0,Γ); whenever the empirical gain is non-zero, which is generic for a non-collapsed ensemble, any particle in B_R(u*) has positive probability of leaving B_R(u*) in one step, and an initial i.i.d. Gaussian ensemble is contained in B_R(u*) only with probability strictly less than one. The constants in Lemma 3 and the bound of the data-misfit factors in Lemma 4 (e.g., the use of ‖Gε(u*)−Gε(u^{(j)}_n)‖≤C_G R) all require this assumption, so Theorems 1 and 2 establish convergence only conditionally on an event that the algorithm does not guarantee. The paper should either prove a propagation result (for example by adding a projection onto B_R(u*) after each update) or reformulate the theorems as conditional on this event, ideally with a probability estimate.","section":"§2.3, Assumption 3; §3.1, Lemma 4"},{"comment":"The statement in the proof of Lemma 4 that ζ^{(j)}_{n+1}=η^{(j)}_{n+1}+η are i.i.d. and independent of uε_n,u0_n is not correct for the algorithm as specified. The observation y in (8) is a single fixed vector; writing y=Gε(u*)+η introduces one realization η that is reused at every iteration n and for every particle, so the variables ζ^{(j)}_{n+1} are correlated across n and (through the history of the ensembles) dependent on uε_n and u0_n. If instead η is treated as random, the same draw is used at all iterations, so i.i.d.-ness across n still fails. This invalidates the step going from E[‖uε_n−u0_n‖(C_G R+‖ζ‖)] to α1 E[‖uε_n−u0_n‖]. The argument can likely be repaired by working with fixed y and assuming a deterministic bound on ‖y−Gε(u)‖ on B_R(u*), or by conditioning on y, but as written the proof does not go through.","section":"§3.1, Lemma 4"},{"comment":"Theorem 2 inherits Assumption 3, and this is more than a technicality here: the Wasserstein distance W_{1,2} in Lemma 6 is computed on the metric space (B_R(u*),‖·‖_2), so the empirical measures μ^ε and μ^0_h must be supported in B_R(u*). Since the paper does not show that particles remain in this ball, the Bayesian convergence statement is again conditional on the unproved event of Assumption 3. In addition, the notation '{μ^ε−μ^0_h} L1 ⇀ 0' is informal because the difference of two probability measures is not a probability measure and the mode of convergence in Definition 1 is defined for random measures; the proof actually establishes E[W_{1,2}(μ^ε,μ^0_h)]→0, which should be the stated conclusion.","section":"§3.2, Theorem 2"}],"minor_comments":[{"comment":"The numerical setup is inconsistent: §5.1 fixes ε=1/64 for the data, while §5.2 reports experiments with ε=1/32 and with varying ε. Please clarify whether the true observations are recomputed for each value of ε in Figures 4–6.","section":"§5.1–§5.2"},{"comment":"The Gaussian modelling-error assumption E(u)∼N(m,Σ) with m and Σ independent of u is explicitly acknowledged to lack theoretical justification; this should be presented as a modelling assumption whose validity is heuristic, and the text should state that Theorems 1–2 do not depend on it.","section":"§4"},{"comment":"There are several typos and grammatical slips: 'a inverse problems' in the introduction, 'if this work' in Section 3, and 'witch' instead of 'which' in the proof of Theorem 4. A careful proofread is needed.","section":"§1, §3, §4"},{"comment":"The final step of Lemma 7 ('extend to all bounded continuous functions by density') is terse; since Lipschitz functions are dense in the sup norm on the compact space B_R(u*), one sentence of justification would make the argument self-contained.","section":"Appendix, Lemma 7"}],"recommendation":"major_revision","confidential_remarks":"The main concerns are technical and appear fixable by adding a projection step or by restating the theorems as conditional on the invariant-ball event. The paper's reliance on its own prior works [2,3] is legitimate, but the overlap with [3] in the modelling-error section should be clarified if the paper is to be judged on novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper does something new and useful: it replaces the expensive multiscale forward solver inside an ensemble Kalman filter with a homogenized finite-element surrogate, then proves the filter error decays as ε + h^{s+1}. That's a clean, practical result. Theorems 3 and 4, which budget how many full multiscale solves are needed to estimate the modelling error, are also genuinely new and likely useful. The proofs are mostly solid; the self-citations to [2,3] set up the problem but the convergence analysis here is not circular.\n\nThe soft spots are real. The biggest is Assumption 3. All the covariance bounds and the error recursion need every particle to stay inside a fixed ball around the true parameter. The paper never shows the EnKF update preserves that ball, and with Gaussian observation noise it cannot: the update is a linear combination that includes the unbounded noise, so there is positive probability of leaving the ball in one step. The stress-test note is right. This leaves the main theorem conditional on a property the algorithm is not known to satisfy. It may be repairable—via a probabilistic radius, a bounded-noise model, or an explicit propagation estimate—but as written it is a genuine gap.\n\nSecond, Lemma 4 misstates the noise structure. The ensemble at step n depends on the fixed noise realization η from the data, so the variables ζ^(j) = η^(j)+η are not i.i.d. and not independent of the ensembles. Conditioning on η and tracking ‖η‖ would fix the computation, with slightly worse constants. Minor in the long run, but it should be corrected.\n\nThird, the Gaussian modelling error assumption is acknowledged as heuristic; fine as an assumption, but the paper would be stronger if it offered real justification beyond a one-dimensional reference. The numerical section is illustrative: no code, a few plots, no detailed convergence check against a known truth.\n\nNet: this deserves a serious referee. It is a solid contribution that needs a major revision to tighten the particle-boundedness argument. I would bring it to a reading group as an example of a clean conditional convergence theorem with a load-bearing but unverified assumption.","headline":"A genuinely new combination of EnKF with homogenization, with solid conditional theorems, but the central result leans on an unverified particle-boundedness assumption that likely fails for Gaussian noise.","tokens_in":25319,"tokens_out":5121,"would_cite":true,"duration_ms":49696,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","65N21","74Q05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that an ensemble Kalman filter can use a cheap homogenized surrogate for oscillatory multiscale PDEs and still converge to the filter built on the exact multiscale model.","keywords":["inverse problems","multiscale elliptic PDEs","homogenization","ensemble Kalman filter","Bayesian inference","modelling error","FE-HMM"],"falsifier":"Take a small ensemble (say $J=10$) with a prior of unbounded support and a large Kalman gain, run the homogenized EnKF, and record the particle norms after each update: if any particle leaves $B_R(u_*)$, Assumption 3 is violated and the hypotheses of Theorem 1 are not satisfied. Alternatively, fix $\\varepsilon$ and $h$, sample $E_i = G^\\varepsilon(u_i) - G_h^0(u_i)$ for i.i.d. prior draws, and apply a normality test to the empirical distribution; a decisive rejection would falsify the Gaussian modelling-error model of Section 4 for that problem.","tokens_in":24286,"feed_emoji":"🧮","tokens_out":8206,"duration_ms":72742,"temperature":0.7,"pith_summary":"This paper aims to establish that the ensemble Kalman filter (EnKF) for inverse problems governed by multiscale elliptic PDEs can be run with a homogenized surrogate forward map instead of the exact oscillatory solver, and still provably converge to the same result. It proves that the ensemble of recovered parameters from the surrogate-based filter converges to the ensemble from the exact multiscale filter as the oscillation scale $\\varepsilon$ and the finite-element mesh size $h$ tend to zero, with the explicit rate $C(\\varepsilon + h^{s+1})$. The same convergence is proved for the empirical posterior distributions in the Wasserstein distance. This matters because it turns an often computationally infeasible procedure—many solves of a finely resolved oscillatory PDE inside the filter—into a cheap coarse-mesh computation whose error is quantified and controlled.","feed_headline":"Cheap homogenized solves provably match exact multiscale EnKF","feed_subtitle":"Coarse-mesh homogenized runs converge to full multiscale runs at rate C(epsilon + h^(s+1)), making costly solves unnecessary.","key_machinery":"The load-bearing mechanism is a one-step error recursion, Lemma 4: if $u_n^\\varepsilon$ and $u_n^0$ are the ensembles from the exact multiscale map and the exact homogenized map at iteration $n$, then $\\mathbb{E}[\\|u_{n+1}^\\varepsilon - u_{n+1}^0\\|] \\le \\alpha \\mathbb{E}[\\|u_n^\\varepsilon - u_n^0\\|] + \\gamma \\mathbb{E}[e(\\varepsilon,u_n^0)]$, where $e(\\varepsilon,u)$ is the mean squared forward-map difference $(1/J)\\sum_j \\|G^\\varepsilon(u^{(j)}) - G^0(u^{(j)})\\|^2$. Combining this recursion with Lemma 2 (homogenization error $\\le K\\varepsilon$) and Lemma 5 (finite-element error $\\le \\tilde K h^{s+1}$) gives the theorem. The recursion closes because Lemma 3 proves the empirical covariance matrices used in the Kalman gain are bounded and Lipschitz as functions of the ensemble, provided all particles stay in the ball of Assumption 3.","core_discovery":"The central discovery is Theorem 1: under Lipschitz regularity of the tensor and observation operators and an assumed bound on all particles, the ensembles produced by the EnKF with the homogenized finite-element surrogate $G_h^0$ and with the exact multiscale map $G^\\varepsilon$ satisfy $\\mathbb{E}[\\| u_N^\\varepsilon - u_{N,h}^0 \\|] \\le C(\\varepsilon + h^{s+1})$, where $s=\\min\\{r,q\\}$ depends on the finite-element degree and the regularity of the homogenized solution. Theorem 2 transfers this ensemble convergence to the empirical Bayesian posterior measures, which converge weakly in $L^1$ as $\\varepsilon,h\\to 0$. The proof separates homogenization error from discretization error and shows each is propagated linearly through the Kalman update, so the surrogate is not merely cheaper but faithful in the small-$\\varepsilon$, small-$h$ limit.","pith_inferences":["A direct extension of the argument would replace Assumption 3 by an algorithmic safeguard—for instance, projecting particles back into a fixed ball after each update—which would make the convergence theorem unconditional rather than conditional on the dynamics staying bounded.","The same exact-map-versus-cheap-surrogate recursion should apply to other surrogate forward maps, such as reduced-order or data-driven models, whenever the surrogate error admits a bound like Lemma 2 and the forward maps are Lipschitz; this suggests a general template for certified EnKF surrogates.","The bounds in Theorem 3 predict that the required number of full multiscale solves for modelling-error estimation shrinks quadratically (mean) or quartically (covariance) with $\\varepsilon$ and $h$, a scaling one could test empirically by varying $\\varepsilon$ at fixed noise level.","The posterior convergence is established for the weak Wasserstein metric; stronger distances would require additional assumptions, so the result should not be read as a total-variation or pointwise posterior guarantee."],"forward_implications":["Practitioners can replace each multiscale forward solve in an EnKF by a homogenized coarse-mesh solve, reducing cost by roughly the ratio of fine to coarse degrees of freedom while keeping an explicit error of order $\\varepsilon + h^{s+1}$.","The Bayesian statement implies that uncertainty quantifications obtained from the cheap filter, such as credible intervals and posterior means, are consistent, as $\\varepsilon,h\\to 0$, with those obtained from the full multiscale model.","The modelling-error analysis gives a rule for choosing how many full multiscale solves are needed in offline estimation: $N_E = O(\\eta^{-2}\\log(\\alpha^{-1})(\\varepsilon^2 + h^{2(s+1)}))$ for the mean, and a quartic analogue for the covariance.","Because the limits in $\\varepsilon$ and $h$ can be interchanged, the guarantee is insensitive to which parameter is reduced first in a convergence study.","Convergence of ensembles in the ensemble norm implies convergence of the empirical measures in the Wasserstein metric $W_{1,2}$, so point estimations and posterior distributions improve together."],"supporting_citations":[{"why":"Supplies the ensemble Kalman formulation for inverse problems and the Kalman update formula (16)–(17) that the error recursion starts from.","marker":"[11]"},{"why":"Provides the Bayesian interpretation of the EnKF ensemble and the convergence of the empirical measure to the posterior, which Theorem 2 extends to the multiscale setting.","marker":"[17]"},{"why":"Establishes that a homogenized surrogate solves multiscale inverse problems under Tikhonov regularization, the point-estimate idea generalized here to the EnKF.","marker":"[2]"},{"why":"Develops Bayesian numerical homogenization and modelling-error estimation for multiscale inverse problems, and supplies the experimental setup used in Section 5.","marker":"[3]"},{"why":"Defines the FE-HMM used to build the homogenized discrete forward map $G_h^0$ in the algorithm.","marker":"[4]"},{"why":"Provides the homogenization theory guaranteeing existence of $A_u^0$ and weak convergence $p^\\varepsilon \\to p^0$ used in Lemma 2.","marker":"[8]"},{"why":"Gives the $O(\\varepsilon)$ convergence rate for homogenized solutions used in Lemma 2's rate bound $K\\varepsilon$.","marker":"[12]"},{"why":"Supplies the finite-element a priori error estimate used in Lemma 5 to obtain the $h^{s+1}$ term.","marker":"[7]"},{"why":"Introduces the offline modelling-error formulation, model discrepancy plus noise, used in Section 4.","marker":"[6]"},{"why":"Provides the dynamic, online updating strategy for the modelling-error distribution that the paper adapts.","marker":"[5]"}],"fun_headline_variants":["Cheap EnKF converges to full multiscale","Surrogate Kalman filter exact in the limit","Multiscale inversion with surrogate EnKF","EnKF + homogenization: proven convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 3: every particle of the ensemble stays, at every iteration, inside a fixed ball $B_R(u_*)$ around the true parameter; the paper gives no mechanism by which the EnKF updates would guarantee this, and without it the covariance estimates and the error recursion fail.","fun_headline_variants_meta":{"raw":{"variants":["Cheap EnKF converges to full multiscale","Surrogate Kalman filter exact in the limit","Multiscale inversion with surrogate EnKF","EnKF + homogenization: proven convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001093,"raw_usage":{"total_tokens":4520,"prompt_tokens":853,"completion_tokens":3667,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3609}},"tokens_in":469,"tokens_out":3667,"duration_ms":25395,"temperature":1.0,"reasoning_tokens":3609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:12:01.659310+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small ensemble (say $J=10$) with a prior of unbounded support and a large Kalman gain, run the homogenized EnKF, and record the particle norms after each update: if any particle leaves $B_R(u_*)$, Assumption 3 is violated and the hypotheses of Theorem 1 are not satisfied. Alternatively, fix $\\varepsilon$ and $h$, sample $E_i = G^\\varepsilon(u_i) - G_h^0(u_i)$ for i.i.d. prior draws, and apply a normality test to the empirical distribution; a decisive rejection would falsify the Gaussian modelling-error model of Section 4 for that problem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ensemble Kalman formulation for inverse problems and the Kalman update formula (16)–(17) that the error recursion starts from."},{"cited_title":"Schillings and A","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian interpretation of the EnKF ensemble and the convergence of the empirical measure to the posterior, which Theorem 2 extends to the multiscale setting."},{"cited_title":"Abdulle and A","cited_arxiv_id":null,"evidence_quote":"Establishes that a homogenized surrogate solves multiscale inverse problems under Tikhonov regularization, the point-estimate idea generalized here to the EnKF."},{"cited_title":"Uncertain","cited_arxiv_id":null,"evidence_quote":"Develops Bayesian numerical homogenization and modelling-error estimation for multiscale inverse problems, and supplies the experimental setup used in Section 5."},{"cited_title":"Cioranescu and P","cited_arxiv_id":null,"evidence_quote":"Provides the homogenization theory guaranteeing existence of $A_u^0$ and weak convergence $p^\\varepsilon \\to p^0$ used in Lemma 2."},{"cited_title":"Moskow and M","cited_arxiv_id":null,"evidence_quote":"Gives the $O(\\varepsilon)$ convergence rate for homogenized solutions used in Lemma 2's rate bound $K\\varepsilon$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the finite-element a priori error estimate used in Lemma 5 to obtain the $h^{s+1}$ term."},{"cited_title":"Calvetti, O","cited_arxiv_id":null,"evidence_quote":"Introduces the offline modelling-error formulation, model discrepancy plus noise, used in Section 4."},{"cited_title":"Calvetti, M","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic, online updating strategy for the modelling-error distribution that the paper adapts."}],"review_version":1}