{"id":"4b19374e-0bbf-46e1-9c5e-a5c3bf7ac1d9","arxiv_id":"2412.12682","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper derives a closed-form mean field equilibrium for a controlled spiking neuron population and claims an approximate Nash equilibrium, but the consistency condition for the aggregate mean is not proven.","lead":"This paper presents a mean field game model where many neurons control their membrane potential to synchronize with the population average, and claims to construct an approximate Nash equilibrium. The main result is not established because the fixed point is solved for each neuron type separately while the actual interaction is through the global average.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1 relies on Lemma 4.1(iii), whose proof conflates the global mean m_U^* with per-type conditional means m_U^{*(p)}; the uncontrolled difference B_i(t,m_U^*)-B_p(t,m_U^{*(p)}) does not vanish in the n→∞ limit.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing gap: Lemma 4.1(iii) requires controlling B_i(t,m_U^*,m_φ^*) − B_p(t,m_U^{*(p)},m_φ^{*(p)}), where the two arguments are the global mean and the per-type conditional mean. I agree that this difference is uncontrolled and does not vanish in general. This is an internal correctness gap, not a disagreement with an external consensus: the paper's own equations (3.17), (4.1), and (4.3) define the two objects, and nothing in the manuscript shows they coincide or that their difference is small. The proof of Lemma 4.1(iii) in Appendix A only estimates empirical-measure deviations (A.5)–(A.11), which would be sufficient only in the symmetric case where all types are equal, a case not assumed. Since Lemma 4.1(iii) is the pivotal step in the proof of Theorem 4.1, the central approximate-Nash claim is not established. The reader also flagged the ill-typed implicit-function-theorem argument in Proposition 3.1; that is a separate concern, but even granting Proposition 3.1, the Theorem 4.1 proof fails at Lemma 4.1(iii). Therefore the reader's REJECT verdict remains appropriate; no verdict change is needed.","tokens_in":35527,"tokens_out":9430,"duration_ms":81643,"concrete_test":"Take μ = (δ_{p1} + δ_{p2})/2 with p1 = (u1,a1,c1) and p2 = (u2,a2,c2), u1 ≠ u2. For each p, solve the linear fixed-point ODE obtained by differentiating (3.17); compute m_U^{*(p)} and m_U^* = (m_U^{*(p1)} + m_U^{*(p2)})/2. Evaluate G(t) = (1/2)[c1^2 B_{p1}(t,m_U^*,m_φ^*) + c2^2 B_{p2}(t,m_U^*,m_φ^*)] − E_μ[c^2 B_p(t,m_U^{*(p)},m_φ^{*(p)})]. If ∫_0^T |G(t)|^2 dt > 0 for generic parameters, the expression that (A.4) claims vanishes in the n → ∞ limit instead has a nonzero limit; this directly falsifies Lemma 4.1(iii) and therefore Theorem 4.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The approximate Nash theorem (Theorem 4.1) is reduced to Lemma 4.1(iii), which asserts E[|°U^*_t − m_U^*(t)|^2] → 0, with m_U^* := E_μ[m_U^{*(p)}] from (4.1). In proving it, the auxiliary process ˜U^i is defined with the global arguments (m_U^*, m_φ^*), while ^U^i uses the μ-averaged coefficients with the per-type arguments (m_U^{*(p)}, m_φ^{*(p)}). The difference of their drifts contains, in (A.4), the term (1/n)Σ_i c_i^2 B_i(s,m_U^*,m_φ^*) − E_μ[c^2 B_p(s,m_U^{*(p)},m_φ^{*(p)})]. The proof then bounds this by empirical-measure deviations (A.6)–(A.8), which would be legitimate only if the two B's were evaluated at the same flow. They are not: B_i is evaluated at the global flow m_U^*, whereas the per-type conditional MFE uses m_U^{*(p)}. For heterogeneous types (u,a,c), m_U^{*(p)} solves (3.17) and m_U^* is its μ-average; nothing forces equality, and generically m_U^*(t) − m_U^{*(p)}(t) is O(1), independent of n. Hence the displayed difference has a deterministic nonzero limit as n → ∞. The estimate (A.6) silently replaces one argument by the other without controlling the functional difference, so Lemma 4.1(iii) is not established. Since Lemma 4.1(iii) supplies the convergence of °U^* to m_U^* used throughout the proof of Theorem 4.1, the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a finite-population stochastic game in which each neuron controls its membrane potential to track the population average, with dynamics given by a jump-diffusion SDE with mean-field drift and a linear connection kernel. The authors formulate a mean field game with type-heterogeneous agents, solve the representative agent's HJB equation in closed form (Lemma 3.1), derive a per-type fixed point for the conditional mean field equilibrium (Proposition 3.1), and then construct a candidate approximate Nash equilibrium from the mu-average of the per-type fixed points (Section 4). The main result, Theorem 4.1, claims that this strategy vector is an epsilon_n-Nash equilibrium with epsilon_n tending to zero as the number of neurons tends to infinity.","tokens_in":35859,"tokens_out":13334,"duration_ms":112453,"significance":"If correct, Theorem 4.1 would be a valuable contribution: it would provide explicit, closed-form best responses and a computable approximate equilibrium for a biologically motivated synchronization model, going beyond existing jump-diffusion mean field game results. The paper's strengths include the closed-form Riccati solution, a genuine fixed point formulation, and a clear construction of the candidate equilibrium. However, the central approximation theorem rests on Lemma 4.1(iii), whose proof does not establish the required convergence, and the proof of Proposition 3.1 contains a serious functional-analytic error. These are load-bearing issues, not presentation problems.","major_comments":[{"comment":"Lemma 4.1(iii) is the key convergence step for Theorem 4.1: it must show that E[|\\bar U^*_t - m^*_U(t)|^2] tends to zero for the empirical mean of the system under the constructed strategies. In the decomposition (A.2), the term (A.4) contains the squared difference between (1/n)\\sum_i c_i^2 B_i(s,m^*_U,m^*_\\phi) and \\mathbb{E}_\\mu[c^2 B_p(s,m^{*(p)}_U,m^{*(p)}_\\phi)]. The bound (A.6) replaces B_i(s,m^*_U,m^*_\\phi) with B_i(s,m^{*(p)}_U,m^{*(p)}_\\phi) without estimating the functional difference induced by the argument change from m^*_U to m^{*(p)}_U. Since m^*_U is the mu-average of the per-type flows m^{*(p)}_U, this argument difference is generically O(1), independent of n, so the displayed term need not vanish. The subsequent empirical-measure estimates (A.7)-(A.8) therefore do not control the original quantity. As Lemma 4.1(iii) supplies the convergence of the empirical average to m^*_U used throughout the proof of Theorem 4.1, for example after (4.10) and in (4.13), the central claim of the paper is not established.","section":"Appendix A, Lemma 4.1(iii), Eqs. (A.2)-(A.6)"},{"comment":"The construction of the candidate equilibrium in (4.2) uses only the global flow m^*_U, while the per-type conditional equilibria in Definition 3.1 are computed with the per-type flows m^{*(p)}_U. The paper does not prove that the conditional mean of U^{*,i} under the global-flow feedback (4.2) equals m^{*(p_i)}_U; indeed, the dynamics (4.4) are driven by the global mean, not by the per-type mean. Thus m^*_U defined by (4.1) is not evidently the correct limit of the empirical average of the finite-player system under theta^*. Lemma 4.1(iii) is precisely the missing law-of-large-numbers statement, and its proof fails as described above. Without a global consistency condition or a quantitative estimate of the discrepancy between the global and per-type flows, the approximate Nash property in Theorem 4.1 is unsupported.","section":"Definition 3.1 and Section 4, Eq. (4.1)-(4.2)"},{"comment":"The proof of Proposition 3.1 attempts to apply the implicit function theorem to the mapping F:[0,T]\\times C_T\\to\\mathbb{R} defined by F(t,m_U)=\\Phi(t,m_U)-m_U(t). For fixed t, the derivative F_{m_U}(t,m^{*(p)}_U) is a bounded linear functional on C_T, not an isomorphism from C_T to C_T. The assertion that this functional is one-to-one and onto \\mathbb{R} is impossible for a nonzero functional on an infinite-dimensional space, and the equation x=\\Phi_{m_U}(t,m^*)x-\\alpha posed in C_T is not equivalent to the scalar equation F_{m_U}x=\\alpha. Consequently, the claimed application of the open mapping theorem and the conclusion m^{*(p)}_U\\in C^1_T are not justified by the given argument. The differentiability of the fixed point may be provable by another route, but the proof as written contains a genuine functional-analytic error.","section":"Proposition 3.1, proof following Eq. (3.22)"}],"minor_comments":[{"comment":"The quantification \"for any (\\theta^i)^n_{i=1}\\in A\" should read \"for any \\theta^i\\in A\"; as written it quantifies over an n-tuple while the inequality involves the single strategy \\theta^i.","section":"Section 2, Definition 2.1"},{"comment":"In the displayed definition of \\check{\\theta}^{*,j}_t, the state argument is written as \\check{U}^{*,j}_t instead of \\check{U}^j_t; this notation is inconsistent with the dynamics in (4.6).","section":"Section 4, after Eq. (4.6)"},{"comment":"There are several typos: \"c`adl`ag\" should be \"c\\`adl\\`ag\", \"Cauchy-Schwartz\" should be \"Cauchy-Schwarz\", \"hods true\" should be \"holds true\", and \"empicical\" should be \"empirical\".","section":"Throughout the manuscript"},{"comment":"The first term on the right-hand side of (A.6) is written with the dummy variable p appearing both in the empirical sum and in the integral against \\mu, which obscures the substitution of arguments discussed in the major comments.","section":"Appendix A, Eq. (A.6)"}],"recommendation":"reject","confidential_remarks":"The gap in Lemma 4.1(iii) is not a local omission: it reflects a mismatch between the per-type consistency condition and the global flow used in the proposed strategies, so the candidate m^*_U may not be the true limit of the empirical mean for heterogeneous populations. The implicit function theorem issue in Proposition 3.1 is also serious but appears more readily repairable. The closed-form HJB analysis is sound and could serve as the basis of a revised paper if the consistency condition is reworked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a genuinely nice model: controlled Galves-Löcherbach-type spiking neurons with linear chemical and electrical coupling, and the authors do the work of solving the representative neuron's LQ control problem in closed form. The reduction of the consistency condition to a one-dimensional fixed point for linear coupling is neat, and Lemma 3.1 is careful and looks correct. That part is publishable on its own.\n\nThe problem is the main theorem. Theorem 4.1 claims an εn-Nash equilibrium, and its proof reduces to Lemma 4.1(iii), which states that the empirical mean of the constructed strategies converges to the global mean m_U*. The proof of that lemma introduces two auxiliary processes with different drift coefficients and tries to show their difference vanishes. The key estimate, around equation (A.4), bounds the difference between (1/n)Σ c_i^2 B_i(s, m_U*, m_φ*) and E_μ[c^2 B_p(s, m_U*^(p), m_φ*^(p))]. But the first B is evaluated at the global mean flow, the second at the per-type conditional mean flow. For heterogeneous neurons those flows are different; m_U* − m_U*^(p) is generically O(1) in n, not small. The proof silently replaces one argument with the other in (A.6) without controlling the functional difference. So the bound has a deterministic nonzero limit and the argument collapses. The stress-test note has this right: the gap is load-bearing, not a typo.\n\nThere is also the implicit function theorem argument in Proposition 3.1, which is ill-typed as written. F maps [0,T]×C_T into R, and the way the partial derivatives and the inverse are stated does not align with the standard IFT setup. This might be repairable by a direct regularity argument, but the current text is not convincing.\n\nThe model and the closed-form best response are worth something, and the approximate Nash claim might be true with a corrected consistency argument—say, by genuinely controlling the sensitivity of B to the mean flow. But as it stands, the central result is unsupported. I would not accept this in current form; the fix is substantial, not editorial. I would, however, send it to a referee if you think the model is worth the effort, because the core idea is sound and a repair may be possible.","headline":"A nice closed-form MFG model for spiking neurons, but the approximate Nash theorem rests on an unproven estimate that conflates global and per-type mean flows.","tokens_in":36415,"tokens_out":4743,"would_cite":false,"duration_ms":44020,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49L12","49N80","92C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes a mean-field game limit for controlling neuronal synchrony and proves a closed-form strategy is an approximate Nash equilibrium for the finite n-player game.","keywords":["nervous system","spiking activity","mean field game","mean field equilibrium","approximate Nash equilibrium","optimal control","neuronal synchronization","jump-diffusion"],"falsifier":"A direct test is to simulate the $n$-player game with two neuron types far apart in their parameters (for example, half with small $a$ and half with large $a$, with well-separated initial potentials) and check whether $E[|\\frac{1}{n}\\sum_{i=1}^n U^{*,i}_t - m^*_U(t)|^2]$ goes to zero and whether $J_i(\\theta^*) - \\inf_{\\theta^i} J_i(\\theta^i, \\theta^{*,-i})$ tends to zero; if either stays bounded away from zero for increasing $n$, Theorem 4.1 fails.","tokens_in":35279,"feed_emoji":"🧠","tokens_out":10598,"duration_ms":84868,"temperature":0.7,"pith_summary":"The paper studies a large population of neurons whose membrane potentials evolve as controlled jump-diffusions coupled through the population average, with spiking events modeled by Poisson random measures and a linear connection strength between neurons. Each neuron wants to choose an external stimulus so that its own potential tracks the population average, which makes the problem a finite-player stochastic game. The paper's aim is to show that this game has a tractable mean-field limit: the limiting control problem has a closed-form solution, the consistency condition that pins down the mean field reduces to a unique fixed point, and the resulting mean-field equilibrium gives an explicit strategy that is an approximate Nash equilibrium for the finite population. If the construction is right, it provides a rigorous, computable design rule for synchronizing neuronal populations by external stimulation, with an error that vanishes as the number of neurons grows.","feed_headline":"Mean-field equilibrium gives near-optimal control of neuron synchrony","feed_subtitle":"As the population grows, the gap between the closed-form strategy and each neuron's best response shrinks to zero.","key_machinery":"The load-bearing machinery is the representative-neuron HJB equation with the quadratic ansatz $V^{(p)}(t,x) = A_p(t)(x - m_U^{(p)}(t))^2 + B_p(t)(x - m_U^{(p)}(t)) + C_p(t)$, which produces an explicit optimal feedback control $\\theta^{*, (p)}(t,x) = (-cA_p(t) - \\rho/2)(x - m_U^{(p)}(t)) - \\frac{c}{2} B_p(t, m_U^{(p)}, m_\\varphi^{(p)})$. The consistency condition $m_U^{(p)}(t) = E[U_t^{*, (p)}]$, with the linear connection strength making $m_\\varphi^{(p)}$ a deterministic function of $m_U^{(p)}$, becomes a single fixed point equation; existence and uniqueness are obtained by a local contraction argument on $C_T$ and then the implicit function theorem to upgrade the solution to $C^1_T$. The finite-player strategy uses the type-averaged mean field $m^*_U(t) = \\int_O m^{*,(p)}_U(t)\\,\\mu(dp)$ and the same coefficients, and the approximate-Nash proof compares the finite game to the mean-field control problem term by term.","core_discovery":"The central claim is Theorem 4.1: for the finite $n$-neuron game with dynamics (2.1) and costs (2.2), the strategy vector $\\theta^*$ defined in (4.2) from the mean-field equilibrium is an $\\epsilon_n$-Nash equilibrium, meaning $J_i(\\theta^*) - \\epsilon_n \\le \\inf_{\\theta^i} J_i(\\theta^i, \\theta^{*,-i})$ for every $i$ with $\\epsilon_n \\to 0$. The proof works by showing that, under the mean-field strategy, the empirical average of the potentials converges to the type-averaged mean field $m^*_U(t) = E_\\mu[m^{*,(p)}_U(t)]$, so each neuron's game cost approaches the representative-neuron cost; the closed-form best response to the mean field then becomes approximately optimal in the finite game.","pith_inferences":["Extension: the same contraction-and-implicit-function route should work for connection strengths that are small perturbations of the linear case, since the uniqueness argument is stable under such perturbations.","Extension: an explicit bound on $\\epsilon_n$ would let a practitioner choose the population size needed to guarantee a given approximation tolerance; the paper does not compute this rate.","Extension: the closed-form structure suggests a model calibration experiment: fit the parameters $(a,c,\\beta,\\gamma,\\rho)$ to recorded spike trains and compare the predicted optimal stimulus profile with the stimulus that minimizes empirical variance in closed-loop simulation."],"forward_implications":["If Theorem 4.1 is correct, an explicit feedback law exists for synchronizing a neuronal population: each neuron's stimulus is a linear function of its deviation from the mean-field trajectory, with coefficients fixed by Riccati solutions and the unique fixed point.","The same mean-field equilibrium gives a prediction for the emergent population dynamics: the average membrane potential follows $m^*_U$, so the design tells an experimenter what synchronized trajectory to expect.","Because the equilibrium is unique, numerical implementations do not face a selection problem among multiple consistent mean fields.","The approximate-Nash guarantee means that for any finite $n$, no neuron can improve its cost by more than $\\epsilon_n$ by deviating from the proposed strategy."],"supporting_citations":[{"why":"Provides the interacting spiking-neuron dynamics with membrane-potential jumps and mean-field coupling that the model extends.","marker":"Galves and Löcherbach (2013)"},{"why":"Supplies the network formulation of spiking neurons as interacting processes that motivates the finite-player structure.","marker":"Galves and Löcherbach (2016)"},{"why":"Gives the linear connection strength $\\varphi(x) = kx + \\ell$, which reduces the second consistency condition to the first.","marker":"Cormier et al. (2020)"},{"why":"Supplies the variance-minimization objective for neuronal membrane potential that the cost function extends.","marker":"Feng and Tuckwell (2003)"},{"why":"Establishes the mean-field-game approximation principle for large-population stochastic games.","marker":"Huang et al. (2006)"},{"why":"Introduces the fixed-point consistency formulation that defines a mean field equilibrium.","marker":"Lasry and Lions (2007)"},{"why":"Provides the controlled-jump-diffusion MFG framework and the construction of an approximate Nash equilibrium transfer.","marker":"Benazzoli et al. (2020)"},{"why":"Supplies the implicit function theorem used to upgrade the fixed point to $C^1_T$ regularity.","marker":"Deimling (2013)"}],"fun_headline_variants":["Mean-field equilibrium yields near-optimal neuron synchrony","As neuron count grows, mean-field control becomes near-optimal","Mean-field strategy is asymptotically Nash for spiking neurons","Neuron spiking: closed-form control approaches best response at scale","Large neural populations: mean-field equilibrium nearly optimal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-type mean-field equilibria stay close enough to the type-averaged trajectory for the error between a neuron's local mean-field coefficients and the global ones to vanish as the population grows; if neuron types are widely separated, that closeness is not automatic.","fun_headline_variants_meta":{"raw":{"variants":["Mean-field equilibrium yields near-optimal neuron synchrony","As neuron count grows, mean-field control becomes near-optimal","Mean-field strategy is asymptotically Nash for spiking neurons","Neuron spiking: closed-form control approaches best response at scale","Large neural populations: mean-field equilibrium nearly optimal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1368,"prompt_tokens":810,"completion_tokens":558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":477}},"tokens_in":426,"tokens_out":558,"duration_ms":5168,"temperature":1.0,"reasoning_tokens":477,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:52:44.932643+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to simulate the $n$-player game with two neuron types far apart in their parameters (for example, half with small $a$ and half with large $a$, with well-separated initial potentials) and check whether $E[|\\frac{1}{n}\\sum_{i=1}^n U^{*,i}_t - m^*_U(t)|^2]$ goes to zero and whether $J_i(\\theta^*) - \\inf_{\\theta^i} J_i(\\theta^i, \\theta^{*,-i})$ tends to zero; if either stays bounded away from zero for increasing $n$, Theorem 4.1 fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the interacting spiking-neuron dynamics with membrane-potential jumps and mean-field coupling that the model extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the network formulation of spiking neurons as interacting processes that motivates the finite-player structure."},{"cited_title":"Tanr\\'e, and R","cited_arxiv_id":null,"evidence_quote":"Gives the linear connection strength $\\varphi(x) = kx + \\ell$, which reduces the second consistency condition to the first."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variance-minimization objective for neuronal membrane potential that the cost function extends."},{"cited_title":"Malhamé, and P.E","cited_arxiv_id":null,"evidence_quote":"Establishes the mean-field-game approximation principle for large-population stochastic games."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the fixed-point consistency formulation that defines a mean field equilibrium."},{"cited_title":"Campi, and L","cited_arxiv_id":null,"evidence_quote":"Provides the controlled-jump-diffusion MFG framework and the construction of an approximate Nash equilibrium transfer."},{"cited_title":"(2013): Nonlinear Functional Analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the implicit function theorem used to upgrade the fixed point to $C^1_T$ regularity."}],"review_version":1}