{"id":"7dc10ad5-ae80-407e-8b91-435317014d60","arxiv_id":"2412.17273","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For an all-to-all balanced E/I network with n^{-1/2} synaptic scaling, the empirical measure converges to a Gaussian whose mean and variance solve explicit ODEs.","lead":"For a model brain network of n excitatory and n inhibitory neurons with random synaptic failures, the paper proves that as n grows large the population activity converges to a Gaussian state described by explicit equations. This matters because balanced excitation and inhibition is a leading theory for the brain's variability, and this gives it a rigorous mathematical foundation in a new noise regime.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5.7's quadratic-form dissipativity does not follow from the eigenvalue condition in (3.8): J_v is non-symmetric with positive ∂Fe/∂ve, so q^T J_v q can be positive even when all eigenvalues have negative real part.","rationale":"The reader correctly identified Lemma 5.7 and the dissipativity of the balanced manifold as load-bearing, but treated the eigenvalue condition in (3.8) as sufficient for the quadratic-form bound. The stress-test pass shows that the implication q^T J q ≤ -ζ|q|^2 is not a consequence of negative-real-part eigenvalues for the nonsymmetric Jacobian arising in this model; indeed the sign structure of the derivatives makes q^T J_v q positive along the pure excitatory direction for sigmoidal f's. This is an internal gap in the proof, not merely a disagreement with external consensus. The paper's framework and numerical evidence are valuable, and the hydrodynamic limit may be true, but the stated proof does not establish it under the stated hypotheses. A correct proof would need a different Lyapunov norm or an added hypothesis on the symmetric part of J_v, plus a separate derivation of covariance convergence and a corrected KMT tail estimate. I agree with the reader's CONDITIONAL assessment, but with a sharper and more specific reason: the central damping estimate in Lemma 5.7 is not merely unproven, it is false under the assumptions as written. If the authors add an explicit strong-dissipativity condition and repair the covariance argument, the theorem could become valid; without that, the current proof cannot be relied on.","tokens_in":17793,"tokens_out":13480,"duration_ms":134953,"concrete_test":"Using the Section 6 parameters (Cee=1, Cei=1.5, Cie=0.5, Cii=0.5, c's, τ=1) at both balanced initial points (ke=ki=1 and ke=1,ki=1/2), compute J_v and the symmetric part S=(J_v+J_v^T)/2, and evaluate q^T J_v q for q=(1,0). If q^T J_v q > 0 (which holds whenever ∂Fe/∂ve>0), then inequality (5.72) is violated at d_W=0, confirming that Lemma 5.7 does not follow from (3.8). Independently, check whether the stated KMT bound exp(-Cε n^{-1/2}) makes ∑_n exp(-Cε n^{-1/2}) finite; if not, the Borel-Cantelli step in Corollary 5.3 requires a stronger rate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The proof of Theorem 4.1 rests on Lemma 5.7, which asserts q_e F_e^n - q_i F_i^n ≤ -ζ/2 (q_e^2+q_i^2) + c_T (d_W(μ^e,μ_e)+d_W(μ^i,μ_i)) |q|. The proof of Lemma 5.7 justifies this by writing q^T J(bar v, μ) q ≤ -ζ |q|^2, claiming this follows from the balanced manifold being attracting. But (3.8) only requires that the eigenvalues of J_v have strictly negative real parts; it says nothing about the quadratic form. For a nonsymmetric 2x2 matrix, q^T J q is controlled by the symmetric part (J+J^T)/2, which can have a positive eigenvalue while J's eigenvalues have negative real parts. In the model with standard increasing sigmoidal f's, ∂_{ve}Fe = Cee ∫ ρ(Ke,x) f_ee'(ve+x) dx ≥ 0 and ∂_{vi}Fi = -Cii ∫ ρ(Ki,x) f_ii'(vi+x) dx ≤ 0, so the diagonal of the symmetric part has opposite signs and q^T J_v q > 0 for q=(1,0). A concrete matrix with eigenvalues -1,-1 but q^T J q = 1 > 0 at q=(1,0) is J = [[1,-2],[2,-3]]. Thus the key damping estimate can fail even when (3.8) holds, so the O(n^{-1/2}) mean fluctuations are not shown to be damped and Lemma 5.4/Theorem 4.1 are not established. Secondary issues compound this: the proof of Theorem 4.1 asserts convergence of K^n without deriving it, and W1-convergence in Lemma 5.2 does not control second moments; the KMT tail in (5.30) is stated as exp(-Cε n^{-1/2}), whose Borel-Cantelli sum diverges, so the a.s. step in Corollary 5.3 is also questionable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies an all-to-all network of n excitatory and n inhibitory neurons whose synaptic variables follow linear ODEs and receive Poisson spike trains with sigmoidal rates, with interactions scaled as n^{-1/2}. The authors define a 'balanced manifold' U by cancellation of mean excitation and inhibition together with local stability of the mean-field Jacobian, and state that if the initial empirical measures converge to Gaussian measures on U, then for times before the manifold is left, the empirical mean and variance converge almost surely to the solution of a finite ODE system (4.6)-(4.13). The proof strategy is to decompose each trajectory into a system-wide mean and fluctuations, prove Wasserstein-1 convergence of the empirical fluctuations to a Gaussian process via KMT coupling and Sanov's theorem, and then use the attracting nature of U to damp the mean fluctuations.","tokens_in":18198,"tokens_out":8239,"duration_ms":75430,"significance":"If the main theorem were fully proved, the contribution would be significant: it gives a rigorous hydrodynamic limit for a balanced network with O(n^2) Poisson noise sources and n^{-1/2} coupling, a regime distinct from typical McKean-Vlasov/Hawkes mean-field limits, and it yields closed ODEs for the mean and variance. The paper also provides numerical simulations supporting the limit. However, the current manuscript contains a central invalid estimate in Lemma 5.7, an incomplete proof of Theorem 4.1, and a Borel-Cantelli argument with a non-summable tail, so the significance is conditional on substantial revision.","major_comments":[{"comment":"The key dissipativity estimate is not established. In the proof of Lemma 5.7, after Eq. (5.80), the authors write that the quadratic form q^T J(bar v, mu) q is at most -zeta_t |q|^2 because 'the balanced manifold is by definition attracting', and then extend this to J(tilde v). However, Hypothesis (3.8) only constrains the real parts of the eigenvalues of J_v, not the symmetric part of J_v. For a non-symmetric 2x2 matrix, q^T J q is governed by (J + J^T)/2, which can have a positive eigenvalue while all eigenvalues of J have negative real parts; for example, J = [[1,-2],[2,-3]] has eigenvalues -1 and -1 but q^T J q = 1 for q = (1,0). In the present model with increasing sigmoidal f's, partial_{v_e} F_e >= 0 and partial_{v_i} F_i <= 0, so the quadratic form can indeed be positive for q = (1,0). Consequently Lemma 5.4 and the damping of the O(n^{-1/2}) mean fluctuations are not proved. The authors must either prove a uniform bound q^T J_v q <= -zeta |q|^2 from the model parameters or replace this step with a Lyapunov-function or weighted-norm argument.","section":"Section 5, Lemma 5.7 (Eqs. (5.80)-(5.81))"},{"comment":"The proof of the main theorem is incomplete. After Corollary 5.3, the text states that it suffices to show (5.20)-(5.23) and then stops; no argument is given for (5.20)-(5.21). The preceding lemmas only provide Wasserstein-1 estimates (Lemmas 5.1, 5.2, 5.4). Wasserstein-1 convergence of empirical measures to a Gaussian does not control the empirical second moment, so the convergence of hat K^n_e and hat K^n_i to K_e and K_i does not follow. A separate estimate for |hat K^n_alpha(t) - K_alpha(t)|, for example via uniform integrability or an L^2 or quadratic Wasserstein bound, is required.","section":"Section 5, Proof of Theorem 4.1 (after Corollary 5.3)"},{"comment":"The Borel-Cantelli step in the proof of Lemma 5.2 is invalid as written. The tail bound P(X^c_{epsilon,n}) <= exp(-C_epsilon n^{-1/2}) in (5.30) is not summable in n, so it cannot imply that X_{epsilon,n} holds almost surely for all large n. The standard KMT strong approximation for Poisson processes gives a much stronger, summable tail (typically exp(-c epsilon sqrt n) for fixed epsilon), so this may be a typographical error in the exponent, but the rate must be corrected because Corollary 5.3 and the 'with unit probability' statement in Theorem 4.1 depend on this step.","section":"Section 5, Lemma 5.2, Eq. (5.30)"},{"comment":"The proof of Lemma 5.6 contains a circular dependence. In (5.69), the difference between Q^n_{alpha beta}(t) and the Gaussian integral is bounded by c |bar v_beta(t) - v^n_beta(t)| + c d_W(hat mu^n_{beta,t}, nu^n_{beta,t}) + c |K^n_{alpha beta}(t) - K_{alpha beta}(t)|, and the lemma is then declared to follow. Since K^n_{alpha beta}(t) is the empirical covariance whose convergence is part of the desired conclusion, this is circular unless an independent estimate for |K^n_{alpha beta}(t) - K_{alpha beta}(t)| is supplied. This gap is connected to the previous comment and also affects the proof of Lemma 5.1.","section":"Section 5, Lemma 5.6, Eq. (5.69)"}],"minor_comments":[{"comment":"The text says 'in fact we do this in Section ' with a blank section number; the intended reference is missing.","section":"Section 3, Assumptions"},{"comment":"The statement of Lemma 5.5 contains the typo 'na^{-1} log P'; it should read 'n^{-1} log P'.","section":"Section 5, Lemma 5.5"},{"comment":"In the definition tilde u^j_{alpha,t} = tilde x^j_{alpha,t} + v_{alpha,t}, the symbol v_{alpha,t} is not defined at that point; it should presumably be v^n_{alpha}(t).","section":"Section 5, Eq. (5.38)"},{"comment":"In the intermediate value step, the text writes tilde v_e = a bar v_e(t) + (1-a) v^n_e(t) with a single parameter a for both components; since the two components may require different convex parameters, the notation should be clarified.","section":"Section 5, proof of Lemma 5.7"},{"comment":"The captions for the six panels are missing from the text, so the reader cannot tell which panel corresponds to K^n_i versus K^n_e or v^n_i versus v^n_e; please add figure captions and axis labels.","section":"Section 6, Numerical Simulations"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an interesting and timely problem, and the proposed limiting equations are plausible. However, the proof of the main theorem is currently incomplete and contains a genuinely invalid step in the derivation of the dissipativity estimate. I would encourage the authors to repair these gaps, and I would be willing to review a revised version. The paper is within the scope of the journal, but the current form is not acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nHere's my take on arXiv:2412.17273. The paper asks a good question: what is the hydrodynamic limit of an all-to-all balanced E/I network when the noise is attached to each synapse (so O(n^2) noise sources) and the interactions scale as n^{-1/2}? The limiting ODEs (4.6)-(4.15) for the means and variances are new, and the distinction from the Hawkes-process limits of Erny-Locherbach-Loukianova and Pfaffelhuber et al. is real and correctly drawn. The authors also make a sensible honest move: they restrict the statement to times before the trajectory leaves the balanced manifold U, and they verify the manifold condition numerically for their examples. The simulations look consistent with the ODEs.\n\nUnfortunately, the proof of Theorem 4.1 has load-bearing gaps. First, Lemma 5.7 claims q^T J q ≤ -ζ|q|^2 from the assumption that the eigenvalues of J have negative real parts. That is not valid for a non-symmetric matrix. The symmetric part of J can have a positive direction even when the eigenvalues are in the left half-plane; in fact the model's J has a positive diagonal entry in the first coordinate (because f_ee is increasing), so q=(1,0) gives a positive quadratic form. The simple example J = [[1,-2],[2,-3]] has eigenvalues -1,-1 but q^T J q = 1 for q=(1,0). So the damping of the O(n^{-1/2}) mean fluctuations is not established. Second, the proof of Theorem 4.1 lists (5.20)-(5.23) as what remains and then never returns to them: nothing in Section 5.1 shows that the empirical variances K^n_e, K^n_i converge to K_e, K_i. Lemma 5.2 gives only W1 convergence of the empirical measure to the Gaussian, and W1 does not control second moments. Third, the KMT tail bound in (5.30) is stated as exp(-Cε n^{-1/2}), which tends to 1, so it cannot be used for the Borel-Cantelli step in Corollary 5.3 as written.\n\nThese are not cosmetic issues. The main theorem as stated is unsupported. That said, I don't think the result is false; the heuristic is plausible and the ODEs are likely right. What is needed is a corrected dissipativity argument (e.g., a Lyapunov function adapted to the linearized flow rather than the Euclidean norm) and a real estimate on the fluctuation covariance, plus fixing the KMT exponent.\n\nWho is this for? Researchers working on rigorous mean-field limits of balanced networks. It deserves a serious referee and, ultimately, publication if the proof is repaired. I'd send it to review, but the current version should not be accepted without major revision.\n\nBest,\n[Your name]","headline":"A novel and well-motivated model for balanced E/I networks with synaptic noise, but the main theorem is not proven: the key dissipativity lemma is false as stated and the covariance convergence is missing.","tokens_in":18801,"tokens_out":4852,"would_cite":true,"duration_ms":46129,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60F05","60G55","92B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Balanced neuron networks converge to four exact equations","keywords":["balanced neural networks","hydrodynamic limit","excitation-inhibition balance","Poisson spiking","Gaussian fluctuations","mean-field limit","Piecewise Deterministic Markov Process","large deviations"],"falsifier":"Compute the real parts of the eigenvalues of the Jacobian $J_v$ along the ODE trajectory of Lemma 4.2; if any becomes non-negative at a time $t < \\eta$, the balanced manifold is not attracting. Simulate the $2n$-neuron system in that parameter regime: if the empirical means and covariances visibly diverge from the ODE solution before time $\\eta$, the hydrodynamic limit as stated does not hold there.","tokens_in":17538,"feed_emoji":"🧠","tokens_out":8366,"duration_ms":73364,"temperature":0.7,"pith_summary":"This paper tries to establish a hydrodynamic limit for a fully connected network of $n$ excitatory and $n$ inhibitory neurons whose interactions scale as $n^{-1/2}$, much stronger than the usual $n^{-1}$ mean-field scaling. The authors prove that, as $n \\to \\infty$ and for every time up to the moment the system leaves the balanced manifold, the empirical distribution of synaptic activities becomes Gaussian and its mean and variance obey a closed system of four ordinary differential equations. A sympathetic reader should care because this turns a popular verbal explanation for cortical variability -- the dynamic balance of excitation and inhibition -- into a precise limit theorem with explicit equations.","feed_headline":"Balanced neuron networks converge to four exact equations","feed_subtitle":"As the network grows, mean and variance of excitatory and inhibitory activity follow deterministic ODEs until balance breaks.","key_machinery":"The object carrying the argument is the balanced manifold $U$, the set of mean-covariance states $(v_e, v_i, K_e, K_i)$ at which the mean excitatory and inhibitory input fields $F_e, F_i$ vanish and the Jacobian $J_v$ of these fields has eigenvalues with negative real parts. The load-bearing step is the dissipativity estimate of Lemma 5.7, $q^T J q \\le -\\zeta_t |q|^2$, which shows that deviations of the system-wide mean from the limit are exponentially damped as long as the empirical law is near the Gaussian limit. This damping is what lets the $O(n^{-1/2})$ synaptic noise remain harmless, allowing the variance equations to close and the empirical measure to become Gaussian.","core_discovery":"The central claim is Theorem 4.1: with unit probability, for any $T < \\eta$, the empirical means $\\hat{v}^n_e, \\hat{v}^n_i$ and covariances $\\hat{K}^n_e, \\hat{K}^n_i$ converge uniformly on $[0,T]$ to the solution $(\\bar{v}_e, \\bar{v}_i, K_e, K_i)$ of the autonomous ODE system (4.6)-(4.13) constrained to the balanced manifold $U$. Along the way the empirical measure of the $2n$ synaptic variables is shown to concentrate on a Gaussian law: the means are pinned by the balance conditions $F_e = F_i = 0$, while the variances evolve through an Ornstein-Uhlenbeck-type equation driven by the limiting firing rates. The limit holds only while the balanced manifold remains attracting; at the exit time $\\eta$ the ODE system leaves $U$ and the theorem no longer applies.","pith_inferences":["An unstated consequence of the Gaussian limit is that the empirical distribution's skewness and higher cumulants should vanish as $n$ grows; one could test this directly by measuring the third and fourth empirical moments in the same simulations.","Because the stability condition is checked numerically rather than proved, a natural extension is to map out which parameter regions make $\\zeta_t$ positive before $\\eta$; those regions should show the predicted breakdown of the Gaussian description.","The same balance-damping mechanism could plausibly carry over to networks with sparse random connectivity, where the effective interaction strength per neuron would need to be rescaled; the paper's all-to-all proof would need a new argument for the mean-field approximation of the firing-rate sums."],"forward_implications":["The population-level mean activity $\\hat{v}^n_e, \\hat{v}^n_i$ concentrates on the ODE mean $\\bar{v}_e, \\bar{v}_i$ uniformly up to time $T < \\eta$, so macroscopic activity is deterministic in the infinite-size limit.","The empirical covariances converge to $K_e, K_i$, so trial-to-trial variability at the population level is described by just two variance equations and not by $2n$ coupled random trajectories.","The limiting law is Gaussian, meaning the balanced state produces exactly the irregular, asynchronous fluctuation picture that the balanced-network theory was created to explain.","For $t \\ge \\eta$ the theorem gives no prediction; the authors conjecture an abrupt, discontinuous change in activity when balance breaks."],"supporting_citations":[{"why":"Supplies the balanced-network paradigm that motivates the model and the notion of cancellation between excitation and inhibition.","marker":"[57, 58]"},{"why":"Provides the time-rescaled representation of Poisson processes used to express each synapse's spike count in terms of independent counting processes.","marker":"[5]"},{"why":"Gives the strong approximation of counting processes by Brownian motion that underlies the Gaussian fluctuation argument.","marker":"[34, 35]"},{"why":"Supplies Sanov's theorem, used to prove the large-deviation bound that forces the empirical measure toward the Gaussian law.","marker":"[24]"},{"why":"Provides the slow-fast framework for stochastic dynamics near a strongly attracting manifold, used to justify the balanced-manifold reduction.","marker":"[11]"},{"why":"Supports the assertion that large drift forces the solution onto the attracting manifold, the template for damping the mean deviations.","marker":"[33]"}],"fun_headline_variants":["Balanced neural nets: exact ODEs in Gaussian limit","Hydrodynamic limit of balanced neurons is Gaussian ODEs","Balanced excitation and inhibition yield exact network ODEs","Balanced networks avoid blow-up, hit Gaussian ODE limit","Balanced neural networks converge to four deterministic ODEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the balanced state keeps pulling the network back toward it for the whole time interval: if that restoring influence ever vanishes or reverses before the time horizon, the proof of the hydrodynamic limit collapses, and the paper only checks this numerically for its examples.","fun_headline_variants_meta":{"raw":{"variants":["Balanced neural nets: exact ODEs in Gaussian limit","Hydrodynamic limit of balanced neurons is Gaussian ODEs","Balanced excitation and inhibition yield exact network ODEs","Balanced networks avoid blow-up, hit Gaussian ODE limit","Balanced neural networks converge to four deterministic ODEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3500,"prompt_tokens":980,"completion_tokens":2520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2447}},"tokens_in":596,"tokens_out":2520,"duration_ms":17987,"temperature":1.0,"reasoning_tokens":2447,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:38:54.095915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the real parts of the eigenvalues of the Jacobian $J_v$ along the ODE trajectory of Lemma 4.2; if any becomes non-negative at a time $t < \\eta$, the balanced manifold is not attracting. Simulate the $2n$-neuron system in that parameter regime: if the empirical means and covariances visibly diverge from the ODE solution before time $\\eta$, the hydrodynamic limit as stated does not hold there.","supporting_citations":[{"cited_title":"Large Deviations Techniques and Applications 2nd Edition","cited_arxiv_id":null,"evidence_quote":"Supplies Sanov's theorem, used to prove the large-deviation bound that forces the empirical measure toward the Gaussian law."}],"review_version":1}