{"id":"b71b0aab-9c26-4049-9294-c598bba585ec","arxiv_id":"2501.04257","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves a Bernstein inequality and an oracle inequality for kernel estimation of the Vlasov-Fokker-Planck density, and gives a moment estimator for FitzHugh-Nagumo parameters that contains algebraic errors in its estimating equations.","lead":"This paper develops concentration inequalities and nonparametric estimators for the mean-field limit of a kinetic particle system, then proposes a six-parameter moment estimator for a network version of the FitzHugh-Nagumo neuron model. A sign error and missing factors in the displayed moment equations undermine the parameter estimator as stated, while the nonparametric results appear technically sound apart from a bandwidth typo.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.3 moment equations misstate coefficients (missing factors of k, wrong sign in the λ term), so the estimator (21) does not identify ϑ and Theorem 6 fails.","rationale":"The reader's weakest assumption targets the algebraic correctness of the moment equations, and the independent re-derivation confirms the problem: factors of k are missing on the c̄, ā, b̄, λ terms, the λ sign is reversed, and the last entry of M is misindexed as m_{k−2,k} instead of m_{k−2,0}. Since the estimator (21) and Theorem 6 depend on the identity A = Mϑ + Λ, these errors break the identification of the six parameters and invalidate the central applied claim. The reader's secondary concern that b<0 violates the Lyapunov condition is not decisive, because choosing k3 sufficiently large bounds y b2 for any real b; the algebraic mis-specification alone justifies rejection. The nonparametric results (Theorems 3–5) may be salvageable with fixes, but the advertised moment-estimation application is not supported as written.","tokens_in":24732,"tokens_out":13494,"duration_ms":102663,"concrete_test":"Recompute the moment equation for a_k(µ_T) in Section 2.3 by Itô's formula directly from (18) for k=1 and k=2, keeping all factors of k and the sign of the interaction term. If the coefficients of c̄, ā, b̄, λ differ from the displayed rows of M (i.e., c̄ coefficient k∫m_{1,k−1}, ā k∫m_{0,k−1}, b̄ −k∫m_{0,k}, λ −k∫(m_{k,0}−m_{1,0}m_{k−1,0}), and σ² coefficient (k(k−1)/2)∫m_{k−2,0}), then the identity ϑ = M^{-1}(A−Λ) is false and Theorem 6 does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central applied claim rests on the moment identification A = Mϑ + Λ in Section 2.3. Recomputing the evolution of a_k(µ_T) = ∫(x^k + y^k)µ_T via Itô's formula from (18) gives d/dt a_k(t) = k[m_{k,0} − (1/3)m_{k+2,0} − m_{k−1,1} + I m_{k−1,0} − λ m_{k,0} + λ m_{1,0}m_{k−1,0}] + k[c̄ m_{1,k−1} + ā m_{0,k−1} − b̄ m_{0,k}] + (σ²/2)k(k−1)m_{k−2,0}. In the displayed expansion in Section 2.3, the c̄, ā, b̄, and λ terms are missing the factor k, and the λ term has a plus sign between m_{k,0} and m_{1,0}m_{k−1,0} instead of a minus. The matrix row for M given just before (21) repeats the plus sign and also lists the last entry as m_{k−2,k} instead of m_{k−2,0}. Because ϑ̂_N = M̂_N^{-1}(Â_N − Λ̂_N) is built from these erroneous coefficients, the identity ϑ = M^{-1}(A−Λ) used in the proof of Theorem 6 is false, so the estimator is not consistent and the advertised parametric-rate claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a mean-field system of N interacting particles in R^d x R^d with degenerate diffusion in the first coordinate. Its three advertised contributions are: (i) a Bernstein-type concentration inequality for the empirical measure around the solution of the kinetic Vlasov-Fokker-Planck equation, extending earlier work of Della Maestra and Hoffmann; (ii) a pointwise Goldenshluger-Lepski kernel estimator of the density with an oracle inequality and minimax adaptation; and (iii) a moment estimator for the six parameters of a FitzHugh-Nagumo model for populations of neurons, with a nonasymptotic deviation bound implying sqrt(N)-tightness. The proofs use a Girsanov change-of-measure argument and a Lyapunov condition in place of global Lipschitz assumptions.","tokens_in":1982,"tokens_out":2428,"duration_ms":160626,"significance":"The probabilistic and nonparametric parts of the paper are potentially valuable: the Bernstein inequality covers kinetic models with degenerate diffusion and unbounded test functions, and the oracle inequality would give adaptive pointwise estimation in a 2d-dimensional density problem. The paper also makes its moment-growth assumptions explicit and gives a self-contained proof of well-posedness. However, the applied centerpiece, the FitzHugh-Nagumo moment estimator, is built on an algebraically incorrect moment system, and the nonparametric variance penalty is mis-specified. As a result, the main advertised claims in Theorems 5 and 6 are not established as stated.","major_comments":[{"comment":"The displayed moment evolution for a_k is inconsistent with the SDE model (18). Applying Ito's formula to a_k(t) = integral of (x^k + y^k) d mu_t gives d/dt a_k(t) = k( m_{k,0} - (1/3)m_{k+2,0} - m_{k-1,1} + I m_{k-1,0} - lambda m_{k,0} + lambda m_{1,0} m_{k-1,0} ) + k( cbar m_{1,k-1} + abar m_{0,k-1} - bbar m_{0,k} ) + (sigma^2/2) k(k-1) m_{k-2,0}. The paper's expansion and the matrix row immediately before (21) do not match this identity: the displayed cbar, abar, bbar and lambda terms are missing the factor k; the lambda term is written as -lambda(m_{k,0} + m_{1,0}m_{k-1,0}) instead of -lambda m_{k,0} + lambda m_{1,0}m_{k-1,0}; the Lambda_k formula gives m_{k-1,1} the coefficient 1/3 instead of 1; and the sigma^2 entry is m_{k-2,k} instead of m_{k-2,0}. In addition, the column order in the displayed row is (I, cbar, abar, bbar, lambda, sigma^2), whereas (19) defines the parameter vector as (I, abar, bbar, cbar, lambda, sigma^2). Since the estimator (21) is defined by inverting this matrix, the identification identity theta = M^{-1}(A - Lambda) used in the proof of Theorem 6 is false for the estimator as defined, and Theorem 6 is unsupported.","section":"Section 2.3, Theorem 6"},{"comment":"The proof of Theorem 6 requires the 6x6 matrix M to be invertible with a bounded inverse uniformly over the admissible parameter range, because the estimator is M_hat_N^{-1}(A_hat_N - Lambda_hat_N) and the constants zeta_2 depend on |M^{-1}|. No argument for this invertibility is supplied. The entries of M are time-integrals of moments of the unknown law mu_t, and invertibility is not immediate from the displayed formulas; without a uniform lower bound on the smallest singular value of M, the event that M_hat_N is singular is uncontrolled. The theorem should either prove such a bound or explicitly restrict the parameter set to a region on which it holds.","section":"Section 2.3, Theorem 6"},{"comment":"The FitzHugh-Nagumo parameter set is stated as a, I, c > 0 and b in R, but for b < 0 (equivalently bbar < 0) the Lyapunov condition in Assumption 2(iii) fails. With G(x,y) = cbar x + abar - bbar y, the quantity y G(x,y) = cbar x y + abar y + |bbar| y^2 cannot be bounded above by k_3(1 + x^2 + y^2) for any finite k_3. Since Theorem 4, which is invoked in the proof of Theorem 6, is proved under Assumption 2(iii), the claimed concentration and estimation results do not cover b < 0. The authors should either restrict the statement to b > 0 or provide a separate argument for the negative-b case.","section":"Assumption 2(iii), Section 2.1"},{"comment":"The variance penalty in the nonparametric oracle inequality is mis-specified. Since K_h(x,y) = h^{-2d} K(h^{-1}x, h^{-1}y), the variance of the kernel estimator is of order N^{-1} h^{-2d}; however, (15) defines V_N^h = rho |K|^2_{L^2} (log N) N^{-1} h^{-d}. The introduction correctly states the variance order as h^{-2d} N^{-1}, and the proof of Lemma 11 obtains (N h^{2d})^{-1}. With V_N^h as written, balancing V_N^h against a squared bias of order h^{2 beta} gives a squared error of order N^{-2 beta/(2 beta + d)}, which is not the announced minimax rate N^{- beta/(beta + d)} for a 2d-dimensional Holder class. The definition of V_N^h should be h^{-2d}, and the bandwidth calibration and proofs should be made consistent with that choice.","section":"Theorem 5, Eq. (15)"}],"minor_comments":[{"comment":"There are several typographical and grammatical errors, for example 'on that may serve', 'an optimally estimate', and 'The versatiliy of such models go way beyond'.","section":"Abstract and Section 1.1"},{"comment":"The sentence 'Thje constants c1, c2 and c3 ...' has a typo ('Thje') and should be rewritten for clarity.","section":"Section 2.2"},{"comment":"The notation is inconsistent in places: 'aik(mu_T)' mixes italic and Roman letters, 'm_{k,0}(mus)' is a typo for m_{k,0}(mu_s), and the estimator is sometimes written as mu_hat^N_GL(t0,x0) without the y0 argument.","section":"Section 2.3"},{"comment":"In Step 2 of the proof of Theorem 5, the notation '(t0, x0y0)' should be '(t0, x0, y0)', and the role of the constant rho_1 mentioned in the theorem statement is not defined consistently with rho in (15).","section":"Section 3.4"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the probabilistic half of this paper is a legitimate extension of Della Maestra–Hoffmann, but the advertised FitzHugh–Nagumo moment estimator is built on algebraically wrong moment equations, so Theorem 6 does not hold as stated. I would not accept the paper in this form.\n\nWhat is new and good: Theorem 4 gives a Bernstein inequality for the empirical measure of a kinetic McKean–Vlasov system with degenerate diffusion and locally Lipschitz coefficients, under a Lyapunov condition instead of global Lipschitz assumptions. That is a real, useful extension, and the proof strategy (Girsanov change of measure plus a relative-entropy bound) looks coherent. Theorem 5's oracle inequality for the Goldenshluger–Lepski kernel estimator is also a plausible adaptation of the DMH framework to the 2d-dimensional density.\n\nThe soft spots are in Section 2.3. I recomputed d/dt of a_k from (18), and the algebra check is right. The c̄, ā, b̄, and λ contributions are missing the factor k, and the λ term has the wrong sign between m_{k,0} and m_{1,0}m_{k−1,0}. The matrix row repeats the sign error and lists the last entry as m_{k−2,k} instead of m_{k−2,0}; Λ_k also has a sign error on m_{k−1,1}. Because ϑ̂_N is defined as M̂^{-1}(Â−Λ̂), the identity ϑ = M^{-1}(A−Λ) is false and the proof of Theorem 6 collapses. These are not cosmetic typos: the estimator will converge to the wrong parameter.\n\nThere are also smaller issues in Theorem 5's proof: V_N^h is defined with h^{-d} but the calculation yields h^{-2d}, and the lower endpoint of H_N makes (N h^{2d})^{-1} ~ N/(log N)^4, so the claim max (N h^{2d})^{-1} ≲ 1 is false. These look fixable by changing the bandwidth exponents.\n\nBottom line: the nonparametric results deserve a referee's time, and the moment-estimation idea is worth salvaging after a careful rewrite. But as submitted, the main applied theorem is wrong. This is a revise-and-resubmit candidate at best, not an accept.","headline":"The Bernstein and oracle inequalities are a real extension of Della Maestra–Hoffmann, but the FitzHugh–Nagumo moment estimator is built on algebraically wrong moment equations, so Theorem 6 fails as stated.","tokens_in":25545,"tokens_out":5982,"would_cite":false,"duration_ms":51180,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62M05","60J80","60J20","92D25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Empirical moments of a neuron population identify the six FitzHugh-Nagumo parameters at the optimal $\\sqrt{N}$ rate.","keywords":["FitzHugh-Nagumo model","mean-field interacting particle systems","McKean-Vlasov equations","Vlasov-Fokker-Planck equation","Bernstein concentration inequality","Goldenshluger-Lepski estimator","moment estimation","nonparametric minimax estimation"],"falsifier":"Pick a parameter value inside the stated range, simulate the FitzHugh-Nagumo McKean-Vlasov flow at high accuracy, and recompute the six population moment identities $A=M\\vartheta+\\Lambda$ symbolically; any coefficient mismatch in the rows of $M$, such as on the $\\lambda$ or $\\sigma^2$ terms, or any permitted parameter choice with $\\det M=0$, would make the estimator converge to the wrong vector and would falsify Theorem 6.","tokens_in":24484,"feed_emoji":"🧠","tokens_out":12504,"duration_ms":106133,"temperature":0.7,"pith_summary":"The paper develops a statistical inference program for a mean-field system of $N$ interacting particles in which one component is diffusive and the other is purely transported, a setting that includes kinetic models and the FitzHugh-Nagumo model of neuron populations. It proves a Bernstein concentration inequality for the empirical measure around the solution of the limiting Vlasov-Fokker-Planck equation, and uses it to build a Goldenshluger-Lepski kernel estimator of the density with an oracle inequality and minimax optimality. The central result for applications is the estimation of the six parameters $\\vartheta=(I,\\bar a,\\bar b,\\bar c,\\lambda,\\sigma^2)$ from six empirical moments of the population at a fixed time, with exponential nonasymptotic deviation bounds. If the moment equations are right, these parameters are recovered at the parametric $\\sqrt{N}$ rate, the fastest possible for a regular parametric model.","feed_headline":"Six FitzHugh-Nagumo parameters estimated at optimal rate","feed_subtitle":"Empirical moments of a neuron population give root-N-accurate estimates of all six model parameters.","key_machinery":"The argument is carried by three objects. The first is a Bernstein inequality for $\\nu_N-\\nu$: for test functions $\\varphi$ it gives $P_N(|\\int\\varphi\\,d(\\nu_N-\\nu)|\\ge\\gamma)\\le c_1\\exp(-c_2 N\\gamma^2/(|\\varphi|^2_{L^2(\\nu)}+|\\varphi|_\\infty\\gamma))$, obtained by a Girsanov change of measure from an independent-particle system and valid for unbounded polynomial test functions. The second is the moment system $A=M\\vartheta+\\Lambda$, whose rows are time integrals of moments of the flow $\\mu_t$; inverting its empirical version defines $\\hat\\vartheta_N$. The third is the Goldenshluger-Lepski selection rule, which chooses a bandwidth $\\hat h_N$ by balancing the empirical bias term $A^N_h$ against the variance term $V^N_h\\asymp h^{-2d}N^{-1}\\log N$.","core_discovery":"The paper's central statistical claim is that the parameter vector of the mean-field FitzHugh-Nagumo model is identifiable from a six-by-six linear system in moments of the flow, and that the plug-in estimator $\\hat\\vartheta_N = \\hat M_N^{-1}(\\hat A_N - \\hat\\Lambda_N)$ satisfies the concentration bound $P_N(|\\hat\\vartheta_N-\\vartheta|\\ge\\gamma) \\le \\zeta_1 \\exp(-\\zeta_2 N \\min(\\gamma,1)^2/(1+\\max(\\gamma,1)))$. This gives tightness of $\\sqrt{N}(\\hat\\vartheta_N-\\vartheta)$ under $P_N$, hence estimation at the parametric rate. In the same framework, the adaptive kernel estimator $\\hat\\mu^N_{GL}$ achieves an oracle inequality matching the best bias-variance trade-off over a bandwidth grid, which yields the minimax pointwise rate $N^{-\\beta/(\\beta+d)}$ up to logarithms for a $2d$-dimensional density of Hölder smoothness $\\beta$.","pith_inferences":["The same moment-inversion scheme should transfer to other mean-field models whose drift is polynomial and linear in the parameters, once the corresponding matrix is checked for invertibility.","Because the probability measure $\\rho(dt)$ in Theorem 4 is arbitrary, a discretized version of the moment integrals over a fine time grid should inherit similar bounds, although the paper does not quantify the discretization error.","The $2d$-dimensional minimax rate warns of the curse of dimensionality, so exploiting the kinetic structure to estimate lower-dimensional summaries is a natural next step.","A natural follow-up is to turn the tightness in Theorem 6 into a central limit theorem with an explicit covariance matrix, which would produce calibrated confidence intervals; the paper stops at tightness."],"forward_implications":["If Theorem 6 holds, all six parameters of the neuron-network model are estimable at the $\\sqrt{N}$ rate from a single snapshot of population moments at time $T$.","The concentration inequality in Theorem 4 gives nonasymptotic control for unbounded polynomial test functions, so moment statistics of arbitrarily high order fall under the same exponential bounds.","Theorem 5 makes the kernel density estimator adaptive: the data choose the bandwidth, and the price compared with the best oracle choice is only logarithmic.","The theory covers degenerate kinetic equations, where the $Y$ component has no diffusion, under Lyapunov-type growth instead of global Lipschitz assumptions, which applies to the cubic FitzHugh-Nagumo drift."],"supporting_citations":[{"why":"Provides the Girsanov change-of-measure strategy and the Bernstein-inequality framework that Theorem 4 extends.","marker":"[DMH22]"},{"why":"Proposes the kinetic mean-field FitzHugh-Nagumo model whose parameters the paper estimates.","marker":"[MQT16]"},{"why":"Supplies the fixed-point argument used to prove existence and uniqueness of the McKean-Vlasov solution.","marker":"[L+18]"},{"why":"Defines the Goldenshluger-Lepski bandwidth-selection machinery behind Theorem 5's oracle inequality.","marker":"[GL08, GL11, GL14]"},{"why":"Contains the Bernstein inequality for independent variables on which the concentration proof is built.","marker":"[BLM13]"},{"why":"Guarantees the solution has a continuous density, the target of pointwise estimation in Theorem 5.","marker":"[KMM10]"},{"why":"Gives the minimax lower-bound theory that identifies $N^{-\\beta/(\\beta+d)}$ as optimal for $2d$-dimensional Hölder densities.","marker":"[Tsy09]"}],"fun_headline_variants":["Six FitzHugh-Nagumo parameters estimable at root-N","Neuron population moments give root-N FHN parameter estimates","Six neural parameters identified at optimal rate from moments","FHN parameter vector recovered at parametric rate","Mean-field neuron model: six parameters at root-N accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the six displayed moment equations are algebraically correct and that the $6\\times6$ matrix $M$ built from them is invertible with bounded inverse over the stated parameter range; the paper states this system without proving the algebra or the invertibility.","fun_headline_variants_meta":{"raw":{"variants":["Six FitzHugh-Nagumo parameters estimable at root-N","Neuron population moments give root-N FHN parameter estimates","Six neural parameters identified at optimal rate from moments","FHN parameter vector recovered at parametric rate","Mean-field neuron model: six parameters at root-N accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3365,"prompt_tokens":947,"completion_tokens":2418,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2340}},"tokens_in":563,"tokens_out":2418,"duration_ms":19966,"temperature":1.0,"reasoning_tokens":2340,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:39:53.164844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a parameter value inside the stated range, simulate the FitzHugh-Nagumo McKean-Vlasov flow at high accuracy, and recompute the six population moment identities $A=M\\vartheta+\\Lambda$ symbolically; any coefficient mismatch in the rows of $M$, such as on the $\\lambda$ or $\\sigma^2$ terms, or any permitted parameter choice with $\\det M=0$, would make the estimator converge to the wrong vector and would falsify Theorem 6.","supporting_citations":[],"review_version":1}