{"id":"cd7efd6c-0832-40d8-85a4-864d7d09e3b7","arxiv_id":"2411.14801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A coarse-grained Markov model of leaky integrate-and-fire networks yields an ODE system that matches finite-size spiking simulations across firing rates, attractors, and bifurcations.","lead":"Scientists reduce a spiking neural network, where neurons communicate with discrete electrical pulses, to a compact set of ordinary differential equations that track how many neurons sit in each electrical state. The reduction is built on a Markov-chain approximation and lets modelers predict firing rates, oscillations, and switches between activity patterns from synaptic parameters alone.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All Section 4 headline results are produced by the Appendix A3 scheme that does not converge to an ODE as δt→0; the derived dsODE itself is never directly simulated, so the central reduction claim is not yet validated.","rationale":"The Reader's primary weakest assumption was Assumption 1, the fast self-decorrelation of synaptic input pools. I agree that is a real limitation, and the authors are honest about its failure regimes; however, the paper's numerical comparisons sit inside the stated domain where the assumption is intended to hold, so it reads as a boundary-of-validity concern rather than a flaw in the demonstrated cases. The more load-bearing problem is that every empirical success is produced by a numerical scheme that is acknowledged not to be a discretization of the derived ODE. The dsODE may well be correct, but its behavior has not been demonstrated. This is not a demand for formal proof; it is a request for a direct numerical check that the object studied in Section 4 is the object derived in Section 3. If that check fails, the central claim that SNNs are reduced to differential equations is not established. If it passes, the conditional status can be upgraded. I therefore retain the conditional verdict, and I mark agreement as partial because the Reader identified the efficient-scheme issue as a secondary fragility rather than the primary one.","tokens_in":34409,"tokens_out":6988,"duration_ms":75022,"concrete_test":"Implement the direct turning-point dsODE exactly as in Algorithm 1, with leak terms included, and run it on the Fig. 4 tau_E sweep and the Fig. 5 S_EI sweep (for example tau_E=2 ms and S_EI=2.95), plus one bifurcation-map parameter such as S_II. Compare firing rates, limit-cycle frequency and amplitude, attractor geometry, and beat-index boundaries against both the efficient Appendix A3 scheme and the LIF network at δt=0.1 ms; also rerun the efficient scheme at δt=0.05 ms. If the direct dsODE and the A3 scheme differ by more than the reported error bars, or if halving δt materially changes the A3 scheme's firing rates or oscillation frequencies, the paper does not yet establish that the derived ODE captures the SNN dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the dsODE equations (15)-(18) quantitatively capture LIF network dynamics. The paper does not simulate those equations in any of its headline tests. Section 3.2.2 and Appendix A3 define an 'efficient numerical implementation' that replaces the piecewise-linear voltage configuration of the derived ODE with a uniform-in-bin distribution, omits leak terms during parameter exploration, and explicitly admits that it does not converge to an ODE as δt→0 because the non-differentiability of nQ(v) at state boundaries makes the recurrent flux infinitely large. Section 4 fixes M=20 and Appendix A5 states that all dsODE simulations use this efficient scheme with δt=0.1 ms. No comparison between the efficient scheme and the actual dsODE is shown; the only justification is the assertion that at δt=0.1 ms it 'closely approximates' the dsODE. Thus Figures 3-8 validate a δt-dependent numerical heuristic, not the derived differential equations. If that heuristic is inaccurate relative to the ODE, the advertised reduction to an ODE is unsupported even if the heuristic matches LIF simulations. A secondary but related gap is the unproven Markov-to-LIF convergence in Section 2.2, supported only by 'data not shown' rather than by a direct, reported comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Markov-chain approximation to leaky integrate-and-fire (LIF) spiking networks, discretizes membrane voltages and synaptic conductances into 'pending kick' pools, and then coarse-grains the Markov model under a fast self-decorrelation assumption (Assumption 1, Section 2.3). From this coarse-grained model the authors derive a system of ordinary differential equations, termed dsODEs, for the number of neurons in each voltage bin and the means and variances of the synaptic drive variables (Eqs. 15–18). The central claim is that this parameter-inherited dsODE system quantitatively captures firing rates, transient dynamics, partial synchrony, metastability, and bifurcation structure of finite-size LIF networks. The paper tests this claim on one-population, E-I, and two-competing-E-population architectures, and compares dsODE predictions with rate models, Fokker–Planck equations, and refractory density methods.","tokens_in":34642,"tokens_out":2837,"duration_ms":31659,"significance":"If the central claim holds, the paper provides a genuinely parameter-free reduction from spiking network dynamics to ODEs, with no fitting of ODE constants to LIF outputs, and it addresses finite-size effects that many mean-field theories neglect. The work also offers a clear conceptual chain: LIF network → Markov model → coarse-grained Markov model → dsODE, with code made available. These are substantive strengths. However, as detailed below, the validation of the central claim is currently incomplete because the headline simulations do not actually solve the derived dsODE system.","major_comments":[{"comment":"All Section 4 headline results (Figures 3–8, including the bifurcation maps and the finite-size study) are produced by the 'efficient numerical implementation' described in Section 3.2.2 and Appendix A3, not by simulating the dsODE system defined by Eqs. 15–18. As the paper itself states, this efficient scheme does not converge to an ODE as δt→0 because the recurrent flux becomes infinitely large at state boundaries. Appendix A5 fixes M=20 and δt=0.1 ms, and the only justification for identifying the efficient scheme with the dsODE is the assertion that it 'closely approximates' the dsODE at this timestep. No comparison between the efficient scheme and the actual dsODE is shown. This is a load-bearing gap: if the efficient heuristic differs substantially from the true dsODE dynamics, then the advertised reduction to an ODE is not validated by the presented simulations, even though the heuristic may match LIF networks. I request a direct comparison of the actual dsODE (using the Algorithm 1 implementation of Appendix A1/A2) with the efficient scheme and with LIF simulations for at least the standard E-I and two-population test cases, or a rigorous argument that the two schemes coincide in the parameter regime tested.","section":"Section 3.2.2 and Appendix A5"},{"comment":"The Markov model is the foundation of the entire derivation, but its convergence to the LIF model as M→∞ is supported only by the statement 'This is supported by our numerical experiments with both single neurons and networks (data not shown)'. Since this convergence is the first link in the reduction chain, and since the paper explicitly does not provide the data, the claim is not verifiable from the manuscript. Please include the numerical comparison of voltage traces, interspike interval statistics, or firing rates between the Markov model and the LIF model at increasing M, or state precisely in what weak sense convergence is expected and provide a targeted test.","section":"Section 2.2"},{"comment":"The downward-jump transition terms in Eq. 6 (the third and fourth lines, with rates proportional to DQE) are introduced with the statement 'We argue that the downward jumps are to mimic the homogenization of Eq. 1'. This is an ad hoc modeling element that is not derived from Assumption 1 or from the pending-kick process described in Section 2.2. Because these terms contribute to the flux that later becomes the diffusion term in the dsODE flux expressions (Eqs. 16c, 16d), their inclusion directly affects the central dynamics, including the firing threshold behavior and the shape of limit cycles. Please provide a derivation of these terms from the underlying Markov process, or at minimum a sensitivity analysis showing that the reported results do not depend critically on the specific form or magnitude of the downward-jump terms.","section":"Equation (6) and Section 2.3.2"},{"comment":"The derivation stacks several approximations beyond the stated 'only assumption': the Gaussian approximation of Poisson/Binomial fluxes (Section 3.1), the replacement of the detailed voltage configuration by a piecewise linear density nQ(v) whose closure P(v|v̄m) is chosen rather than derived (Section 3.2.1), and the uniform-bin approximation in the efficient numerical scheme (Appendix A3). These are legitimate modeling choices, but the abstract's claim that 'our only assumption for the Markov approximation is the fast self-decorrelation of synaptic conductances' is therefore overstated. I recommend either weakening that claim or providing a systematic error analysis showing that each subsequent approximation has a controlled effect on the quantities reported in Section 4.","section":"Sections 3.1 and 3.2"}],"minor_comments":[{"comment":"The line defining the reduced state variable uses 'M = M/L' where the left-hand side appears to be the reduced number of bins; this notation is confusing and should be clarified (e.g., write M′ = M/L).","section":"Section 3.2, second paragraph"},{"comment":"There is a typo: 'the the dsSDE system' should be 'the dsSDE system'.","section":"Section 3.2.2, first sentence"},{"comment":"In the 'Neuronal physiology' column, the row for 'τ I' is labeled 'E-synapse timescale'; this should read 'I-synapse timescale'.","section":"Table 1"},{"comment":"The refractory decay term is written as '−N RQ/τ R' but the refractory period is τ_ref, not τ R; please check the notation and ensure τ_ref is used consistently in the refractory equation.","section":"Eq. 15b"},{"comment":"The definition of the uniform interval uses 'L' both as the bin index and as the bin width in the equations; for readability, use a different symbol for the bin width (for example α).","section":"Appendix A3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious contribution with a plausible conceptual framework, and the lack of any fitting of ODE parameters to LIF outputs is a genuine strength. The primary concern is a validation gap: the actual dsODE is never simulated, and the efficient surrogate is explicitly not an ODE limit. This is fixable within the manuscript's scope by adding direct dsODE simulations and a convergence check. I also note that the manuscript is formatted for a physics/journal-of-computational-physics audience; if the editors consider scope, this is more naturally a computational neuroscience methods paper, but the methodological depth is sufficient for a journal that publishes mathematical biology. I would not reject on scope alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a serious attempt to reduce finite LIF networks to ODEs via a coarse-grained Markov chain with pending-kick pools. Second, the central validation is currently undercut by the paper's own admission: the derived dsODE system is never simulated in the headline tests. Section 4 uses the 'efficient numerical implementation' from Appendix A3, and Section 3.2.2 states that this scheme does not converge to an ODE as δt→0. So Figures 3–8 validate a δt-dependent heuristic, not the advertised differential equations.\n\nWhat is genuinely new is the derivation path: from LIF equations to a pending-kick Markov representation, then to a coarse-grained version under the fast self-decorrelation assumption, then to a deterministic ODE system. The derivation is transparent, and the ODE parameters are inherited from the spiking model rather than fitted to its outputs, so the agreement is a real reduction test. The comparisons are broad: single-population transients, E/I oscillations, metastable switching, and bifurcation maps. The authors also state the limits of Assumption 1 honestly: long synaptic time constants, low firing rates, and strongly biased heterogeneous architectures. There is working code on GitHub.\n\nThe soft spots are real. The missing direct simulation of the dsODE is load-bearing. The paper asserts that at δt=0.1 ms the efficient scheme 'closely approximates' the dsODE, but no comparison is shown; without that, the reduction claim is unsupported. Second, the Markov-to-LIF convergence is unproven and backed only by 'data not shown.' Third, the bifurcation comparisons are mostly visual; there is no quantitative error measure. Finally, the downward-jump term in Eq. 6 is admittedly ad hoc, introduced 'to mimic' homogenization, which complicates the claim that the only assumption is fast decorrelation.\n\nWho is this for? Computational neuroscientists working on population models, and anyone wanting a systematic parameter-to-dynamics map for homogeneous spiking networks. The framework is plausible and the gaps are fixable, so a serious referee is warranted. But in its current form the central claim is not yet validated. I would send it to review with the expectation of major revision: simulate the actual dsODE, compare it to the efficient scheme, and report the Markov-to-LIF convergence check. The paper's honesty makes it a candidate for revision rather than rejection.","headline":"A serious reduction attempt whose headline validation is undermined by the fact that the derived ODE system is never simulated directly; the efficient scheme used instead does not converge to those ODEs.","tokens_in":35261,"tokens_out":2629,"would_cite":false,"duration_ms":25394,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92B20","92C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"With one assumption about synaptic memory, homogeneous spiking neuron networks reduce to an ODE system that reproduces their firing rates, synchrony, and bifurcations.","keywords":["spiking neural networks","Markov models","coarse-graining","leaky integrate-and-fire","neural oscillations","finite-size effects","bifurcation analysis","metastability"],"falsifier":"Run a homogeneous E-I LIF network with slow synaptic time constants (for example, $\\tau_E = \\tau_I = 100$ ms) and low firing rates (below 1 Hz), and compare dsODE predictions for mean firing rate and attractor geometry against direct spiking simulations; if the dsODE fails to track the simulated dynamics while the fast-synapse case succeeds, the fast self-decorrelation assumption is the point of failure.","tokens_in":1599,"feed_emoji":"🧠","tokens_out":2785,"duration_ms":87053,"temperature":0.7,"pith_summary":"The paper claims that a homogeneous spiking neuron network can be reduced to a low-dimensional system of ordinary differential equations under a single stated assumption: that each neuron's synaptic input pool decorrelates quickly from its own history. Under that assumption, neurons in the same voltage state become interchangeable, so the network is summarized by counts of neurons per voltage bin together with the mean and variance of the pending-kick pools that carry synaptic input. The resulting discrete-state ODEs (dsODEs) are then shown to reproduce firing rates, transient and partial synchrony, gamma-band oscillations, beat patterns, metastable switching between attractors, and bifurcation structure across broad parameter sweeps of the original spiking network. A sympathetic reader would care because the framework is a candidate parameter-to-dynamics map: single-neuron physiology, coupling weights, projection probabilities, and external drive become coefficients of an ODE system. That would put regimes previously accessible only by simulation, including finite-size fluctuations, fluctuation-driven activity, and strongly recurrent coupling, within reach of dynamical systems tools.","feed_headline":"Spiking networks reduce to ODEs on one assumption","feed_subtitle":"Finite-neuron firing rates, gamma oscillations, metastable switches, and bifurcations all survive the reduction to differential equations.","key_machinery":"The load-bearing object is the coarse-grained Markov model, whose state is the vector of neuron counts in each voltage bin plus the mean and variance of the four recurrent pending-kick pools. The fast self-decorrelation assumption makes same-state neurons interchangeable, and averaging over the fast pools closes the system. Gaussian approximations to the Poisson fluxes convert the master equation into stochastic differential equations, and a heterogeneous-multiscale bin-coarsening step with a piecewise-linear voltage density turns those SDEs into the deterministic dsODEs; firing rates are the expected total flux through the threshold. The efficient numerical scheme used in the simulations approximates the voltage distribution inside each bin as uniform and applies Gaussian convolution at each timestep.","core_discovery":"The central claim is that the dsODE system, derived from the coarse-grained Markov model, is a faithful surrogate for finite-size homogeneous leaky-integrate-and-fire networks. The paper shows that the system captures high-frequency partial synchrony and metastability of finite-neuron networks, and that beyond firing rates it quantitatively captures the geometry of attractors and bifurcation structures of SNNs. The derivation sidesteps the singularity and irreversibility of spikes by treating spike emission and reset as ordinary state transitions in a Markov chain. The price is a loss of spatial resolution in membrane potential; the compensation is a closed ODE system whose variables are neuron counts per voltage bin and recurrent-drive moments, with firing rates computed as expected flux across threshold.","pith_inferences":["Beyond the paper, the same construction could be applied to other homogeneous spiking models by replacing the LIF transition rules with the corresponding conductance-based or adaptive-neuron rules, at the cost of more state variables per bin.","The Markov-chain formulation suggests that spectral data of the transition matrix, not explored here, could yield estimates of oscillation frequencies and metastable switching rates without forward simulation.","The claim that information loss is minimized is qualitative; one could make it quantitative by measuring the mutual information between the coarse-grained state and the full spiking trajectory as bin size and pool timescales vary.","A practical adaptive scheme could monitor the empirical correlation between the pending-kick pools and voltage and refine voltage bins only where the fast-decorrelation assumption is strained."],"forward_implications":["On one E-population, dsODE errors in average firing rate stay near 3 percent; on E-I networks below 6 percent; on the two-E-plus-I network below 5 percent.","Bifurcation maps across eight parameter directions show the same beat-index branches and bifurcation locations as LIF simulations, and adding noise back to form the dsSDE recovers noise-induced extra branches.","Multistability appears as coexisting attractors: a limit cycle and a fixed point for the synchronous/homogeneous bistable regime, and a saddle-induced switch for biased competition between two E populations.","Metastable switching times are predicted in the tested cases, for example 15.4 ms versus 19.3 ms and 100.4 ms versus 132.2 ms for the E1/E2 competition.","Finite-size effects survive: with coupling scaled by $400/N$, dsODEs track firing rates and synchrony for $N > 400$, and the dsSDE improves predictions at $N = 40$-$60$."],"supporting_citations":[{"why":"Supplies the leaky-integrate-and-fire network formulation that the Markov approximation discretizes.","marker":"[36]"},{"why":"Supplies the cortical parameter set used in all simulations and comparisons.","marker":"[8]"},{"why":"Justifies replacing fast pending-kick pools by their expectations in the coarse-grained model.","marker":"[69]"},{"why":"Supplies the beat-index definition and multi-band oscillation phenomenology used as the bifurcation testbed.","marker":"[66]"},{"why":"Grounds the gamma-oscillation behavior in E-I networks that the dsODE is asked to reproduce.","marker":"[63]"},{"why":"Provides the refractory density method used as a comparison baseline.","marker":"[43]"},{"why":"Provides the finite-size Fokker-Planck baseline for firing-rate comparisons.","marker":"[57]"},{"why":"Supplies the ergodicity result for Markovian spiking network models invoked for invariant measures and large-deviation analysis.","marker":"[81]"}],"fun_headline_variants":["One assumption turns spiking networks into ODEs","Spiking neural networks become ODEs with minimal loss","Finite SNNs reduced to ODEs via Markov approximation","Neural spikes to equations: one assumption does it","SNN dynamics captured by ODEs with one assumption"],"cache_read_input_tokens":37248,"weakest_assumption_plain":"The whole construction rests on the premise that each neuron's synaptic input pool forgets its own history quickly enough that all neurons in the same voltage state behave identically in distribution.","fun_headline_variants_meta":{"raw":{"variants":["One assumption turns spiking networks into ODEs","Spiking neural networks become ODEs with minimal loss","Finite SNNs reduced to ODEs via Markov approximation","Neural spikes to equations: one assumption does it","SNN dynamics captured by ODEs with one assumption"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1394,"prompt_tokens":919,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":395}},"tokens_in":535,"tokens_out":475,"duration_ms":4362,"temperature":1.0,"reasoning_tokens":395,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:52:00.953500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a homogeneous E-I LIF network with slow synaptic time constants (for example, $\\tau_E = \\tau_I = 100$ ms) and low firing rates (below 1 Hz), and compare dsODE predictions for mean firing rate and attractor geometry against direct spiking simulations; if the dsODE fails to track the simulated dynamics while the fast-synapse case succeeds, the fast self-decorrelation assumption is the point of failure.","supporting_citations":[{"cited_title":"Fast global oscillations in networks of integrate-and-fire neurons with low firing rates","cited_arxiv_id":null,"evidence_quote":"Grounds the gamma-oscillation behavior in E-I networks that the dsODE is asked to reproduce."},{"cited_title":"Towards a theory of cortical columns: From spiking neurons to interacting neural populations of finite size","cited_arxiv_id":null,"evidence_quote":"Provides the refractory density method used as a comparison baseline."},{"cited_title":"Self-consistent stochastic dynamics for finite-size networks of spiking neurons","cited_arxiv_id":null,"evidence_quote":"Provides the finite-size Fokker-Planck baseline for firing-rate comparisons."},{"cited_title":"How well do reduced models capture the dynamics in models of interacting neurons?","cited_arxiv_id":null,"evidence_quote":"Supplies the ergodicity result for Markovian spiking network models invoked for invariant measures and large-deviation analysis."},{"cited_title":"Neuronal dynamics: From single neurons to networks and models of cognition","cited_arxiv_id":null,"evidence_quote":"Supplies the leaky-integrate-and-fire network formulation that the Markov approximation discretizes."},{"cited_title":"Averaging for Markov Chains","cited_arxiv_id":null,"evidence_quote":"Justifies replacing fast pending-kick pools by their expectations in the coarse-grained model."},{"cited_title":"Multi-band oscillations emerge from a simple spiking network","cited_arxiv_id":"2206.14942","evidence_quote":"Supplies the beat-index definition and multi-band oscillation phenomenology used as the bifurcation testbed."}],"review_version":1}