{"id":"3a18f234-b0c0-4609-a6f9-9ad329384d8e","arxiv_id":"2608.12839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"An Elephant-reinforced spiking-neuron network is shown to be non-explosive and, under a dissipativity condition, contractive, while its replica mean-field version has unique solutions and only silent stationary states.","lead":"This paper introduces a mathematical model of spiking neurons in which past firing makes future firing harder through an Elephant-type memory variable. The authors prove the model is well-defined, build a mean-field approximation, and simulate how this memory produces a gradual decline and faster extinction of activity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RMF approximation of the finite network is not proven: it rests on a single unquantified numerical comparison, with no propagation-of-chaos limit established.","rationale":"After checking the proofs, I find the finite-volume non-explosion, the conditional Wasserstein contraction (Theorem 2.2), the RMF fixed-point construction (Theorem 4.1), and the invariant-measure characterization (Theorem 5.1) mathematically sound. The contraction proof's leakage and firing bounds are correct; the fixed-point contraction on B_T with constant 2T is standard; the invariant-measure argument via k and n is valid. The only place where the paper's central narrative reaches beyond its proof is the claim that the finite network is closely matched by the RMF approximation. The RMF is formulated, not derived, and the numerical evidence is a single configuration with no uncertainty quantification. This does not invalidate the theorems, but it supports the reader's conditional verdict. The proposed M-scaling test would directly check whether the replica system converges to the RMF, which is the missing link.","tokens_in":20336,"tokens_out":25227,"duration_ms":244776,"concrete_test":"Simulate the M-replica system of Section 3 for N=100 at p=0.7, alpha=0.5, gamma=0.2 with M=1, 10, 100, 1000, and compare the empirical law of a randomly selected replica of a fixed type to the nonlinear RMF solution (approximated by a high-Q particle system) over [0,100]. Measure, say, the L1 distance between mean firing-rate curves or the 1-Wasserstein distance between the tagged replica's law and the RMF law. If the distance does not decrease systematically with M, the claimed RMF approximation is not supported; if it does, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core theorems (non-explosion, conditional Wasserstein contraction, RMF existence/uniqueness, silent invariant measures) appear correct. The load-bearing weakness is the abstract's claim that finite-network dynamics are 'closely matched' by the RMF approximation. Section 4 defines the nonlinear RMF process directly, and Section 6.5 compares it with the N=100 finite network at a single parameter point (p=0.7, alpha=0.5, gamma=0.2) using Q=1000 particles and 20 realizations. No propagation-of-chaos theorem is proved for the M-to-infinity replica system, nor for any N-to-infinity limit. The reported integrated absolute difference 0.9206 over [0,100] has no confidence interval, and the Discussion concedes the match is 'qualitative,' which is weaker than 'closely matched.' Because the abstract's RMF-approximation sentence is part of the central claim, the paper should either prove a limit theorem or explicitly restrict the claim to the tested regime with uncertainty quantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a finite stochastic spiking-neuron network on a ring in which each neuron's firing threshold is raised by a reinforcement variable S_i updated by an Elephant-type rule, and it formulates a nonlinear replica mean-field (RMF) process. The analytical results are: non-explosion of the finite network (Theorem 2.1); conditional exponential contraction in W1 on a truncated potential space under a dissipativity condition (Theorem 2.2); global existence, uniqueness in law and non-explosion of the nonlinear RMF process (Theorem 4.1 and Proposition 4.4); and a characterization of all invariant measures as supported on silent configurations (Theorem 5.1). The numerical section studies firing-rate adaptation under sustained input, extinction times in the undriven network, and a comparison of finite-network and RMF trajectories.","tokens_in":20504,"tokens_out":12392,"duration_ms":124688,"significance":"If the analytical results are correct, the paper provides a clean and tractable example in which spike-history-dependent excitability is incorporated through reinforcement and can be analyzed rigorously: the non-explosion bound is explicit, the RMF existence and uniqueness proof is a genuine fixed-point construction, and the silent-invariant-measure characterization is sharp. The paper is also transparent about the heuristic nature of the power-law fits and about the fact that the RMF process is not proved to be a limit of the finite network. The numerical implementation is reproducible, with scripts and fixed seeds supplied, which strengthens the empirical part. The main gap is not in the proofs of the four theorem-level results, but in the abstract's broader claim that the RMF approximation is closely matched to finite-network dynamics.","major_comments":[{"comment":"The paper does not prove any propagation-of-chaos or mean-field limit connecting the finite network, or the M-replica system of Section 3, to the nonlinear RMF process of Section 4. Section 6.5 compares the finite network with an RMF particle system at a single parameter point (N=100, Q=1000, p=0.7, alpha=0.5, gamma=0.2, 20 realizations), reporting an integrated absolute difference of 0.9206 over [0,100] with no confidence interval. The abstract statement that finite-network dynamics are closely matched by the RMF approximation is therefore not supported to the claimed strength, and the Discussion itself later says only qualitative approximation. Please either state and prove a limit theorem, or restrict the abstract and conclusions to qualitative agreement in the tested regime with the single-parameter limitation and the uncertainty made explicit.","section":"Abstract and Section 6.5"},{"comment":"Theorem 2.2 is stated under the assumption sum_i |lambda_i(t,V) - lambda_i(t,Vhat)| <= L d(V,Vhat), but the paper never verifies this Lipschitz condition for the hard-threshold rate lambda_i(t,V) = 1_{V_i > alpha S_i^+(t)} used throughout. When the reinforcement history S(t) is common to both copies, one can take L=1: an index contributes to the left-hand side only if V_i and Vhat_i lie on opposite sides of alpha S_i^+(t), in which case |V_i - Vhat_i| >= 1. The theorem should state this bound explicitly and replace the abstract condition by the explicit dissipativity requirement gamma > M+2, since as written the central stability result is conditional on an unverified hypothesis.","section":"Theorem 2.2"},{"comment":"The proof of Theorem 2.2 says 'Couple leakages with the same clocks. If site i leaks in both copies...', but the leakage rate is state-dependent, gamma 1_{V_i>0}, so a common clock should be described as a rate-gamma clock per site that resets whichever copy has positive potential. The intended inequality is correct under this description, but the current wording is inaccurate when exactly one copy has V_i>0.","section":"Section 2.2, leakage coupling"}],"minor_comments":[{"comment":"The state space E = N x Z x N^* uses N^* without definition; please state explicitly that N = {0,1,2,...} and N^* = {1,2,...}.","section":"Section 3"},{"comment":"In the proof of Theorem 5.1, the sentence 'only the leakage transition contributes to the generator' is terse: after lambda=0 pi_i-almost surely and a_i(pi)=0, the firing and input terms vanish, and the displayed expression follows; a sentence making this explicit would aid the reader.","section":"Section 5.1"},{"comment":"Table 3 gives median extinction time with an interquartile range but reports 'Mean total firings' without dispersion; state the number of simulations and, ideally, give a measure of variability for the mean total firings as well.","section":"Table 3"},{"comment":"The integrated absolute difference 0.9206 is reported without a confidence interval or sensitivity to the bin width and to the number of realizations; a bootstrap interval or a short sensitivity check would make the numerical comparison more informative.","section":"Section 6.5"}],"recommendation":"major_revision","confidential_remarks":"The analytical core is solid and the paper is publishable in principle. The main risk is overclaiming the RMF approximation: no propagation-of-chaos theorem is proved, and the abstract is stronger than what Section 6.5 and the Discussion support. A revision that restricts the mean-field claims to the tested regime and adds the explicit L=1 verification in Theorem 2.2 would address the load-bearing issues; I would not require a full propagation-of-chaos theorem for acceptance if the claims are appropriately qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the RMF well-posedness and the invariant-measure theorem, not for the numerics. The core proofs are clean and correct. Non-explosion is immediate from a bounded total rate; the Wasserstein contraction is a standard synchronous-coupling argument under a Lipschitz condition on the firing rate; and the fixed-point proof for the nonlinear process is a textbook contraction on short time intervals, extended by concatenation. The invariant-measure theorem is the neatest part: any invariant law must have zero firing rate, hence V=0 a.s., because K grows at each firing and is never reduced. That argument is worth having.\n\nThe numerical part is reproducible: code, seeds, and outputs are promised, and the AIC bootstrap comparison is honest. The authors disclose that the power-law/exponential comparison is finite-time and not asymptotic, and that adaptation is a consequence of the Elephant rule with p>1/2 and positive initial S. So the model is what it says: a parsimonious description, not a mechanism.\n\nWhere the abstract oversells: 'closely matched' is too strong. Section 6.5 uses one parameter set (N=100, Q=1000, p=0.7), 20 realizations, and reports an integrated absolute difference of 0.9206 over [0,100] with no uncertainty. The Discussion says 'qualitative approximation.' Those are in tension. There is no propagation-of-chaos theorem; the RMF is defined directly, not derived as a limit. I would ask the authors to soften the abstract or quantify the match with confidence intervals and state that it is qualitative and parameter-dependent.\n\nThe conditional contraction is more limited than it first appears: it holds on a truncated potential space and conditionally on a common reinforcement history. The Lipschitz constant L is left abstract, though for the hard-threshold indicator one can take L=1 trivially. A one-line remark would help. This is a minor fix.\n\nThe citation pattern is appropriate: the model is attributed to the earlier preprint by one of the authors, and the RMF framework to Baccelli and Taillefumier. The paper is honest about the fact that the invariant measures are silent, which is a useful negative result for anyone hoping to get active stationary states from cumulative reinforcement.\n\nOverall, the mathematical core holds up. This is a qualifying contribution for a math.PR audience interested in mean-field limits of reinforced neuron models. Not a landmark, but rigorous, readable, and reproducible. I would send it to a serious referee; after the abstract is toned down and the RMF-comparison claim qualified, it should be accepted.","headline":"The mathematics is solid; the abstract oversells the RMF match.","tokens_in":21070,"tokens_out":5904,"would_cite":true,"duration_ms":51637,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J25","60J27","60K35","92B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A bounded spiking-neuron network with Elephant-type reinforcement memory is non-explosive, and its replica mean-field process has only silent stationary states.","keywords":["spiking neural networks","elephant random walk","reinforcement memory","replica mean-field","Wasserstein contraction","invariant measures","spike-frequency adaptation","extinction times"],"falsifier":"Run the finite ring network and the RMF particle approximation under parameter regimes far from Section 6.5, for example with $p<1/2$ or with weak leakage and large $\\alpha$, and compare the mean firing-rate curves by integrated absolute difference over a long time window; if the difference grows with network size or replica count instead of staying small, the claimed mean-field match fails. A direct check of Theorem 5.1 would search, numerically or analytically, for an invariant measure of the nonlinear process with positive firing probability—none should exist because $K$ grows without bound at every firing.","tokens_in":20112,"feed_emoji":"🧠","tokens_out":8945,"duration_ms":81680,"temperature":0.7,"pith_summary":"This paper introduces a finite stochastic spiking-neuron network in which each firing updates a signed reinforcement variable, and positive reinforcement raises the neuron's firing threshold; the memory rule comes from the Elephant random walk, so past spiking biases future excitability. The paper's central claim is that this history-dependent model is tractable: the finite process never explodes, its truncated membrane-potential dynamics contract exponentially in the 1-Wasserstein distance whenever leakage dominates the firing expansion, and the associated nonlinear replica mean-field process has a unique global solution whose stationary laws can be characterized. It also proves that every invariant probability measure of that mean-field process is supported on silent states, so undriven activity is always transient. Numerical experiments show that Elephant memory produces a $p$-dependent decline in firing rate under constant input, that the decline is better described by a power law than by a single exponential over the simulated interval, and that the replica mean-field process closely tracks the finite network's transient. If correct, the paper supplies a rigorous example of spike-history-dependent excitability with adaptation-like behaviour emerging from a single cumulative memory variable.","feed_headline":"Spiking networks with Elephant memory settle into silence","feed_subtitle":"A reinforced threshold makes firing decline like adaptation and drives undriven networks to extinction faster.","key_machinery":"The load-bearing mechanism is the Elephant reinforcement update inside the firing rule. When a neuron fires, its signed memory $S$ changes by $\\pm 1$ with probabilities $q_\\pm(s,k)=1/2 \\pm (2p-1)s/(2k)$, its counter $K$ increments, and the new threshold is $\\alpha S^+$; firing itself occurs at rate $\\mathbf{1}\\{V_i>\\alpha S_i^+\\}$. This single cumulative variable converts spike history into excitability. The bounded indicator rate keeps the total intensity at most $(\\gamma+1)N$, giving non-explosion. The contraction proof couples two potential copies with common leakage clocks and common firing clocks, and uses the truncation level $M$ to bound the distance increase caused by an asynchronous firing by $M+2$; together with the Lipschitz estimate on firing-rate differences, this yields the generator inequality that produces the exponential Wasserstein decay. For the mean-field side, the key identity is the self-consistency equation $\\beta_i(t)=P(V_i(t)>\\alpha S_i^+(t))$, and well-posedness follows from a fixed-point map on measurable intensity paths that is a contraction on intervals shorter than $1/2$.","core_discovery":"On the paper's own terms, the discovery is that Elephant-type reinforcement can be embedded in a bounded hard-threshold spiking network while preserving rigorous control. The total event rate is bounded by $(\\gamma+1)N$, proving non-explosion; when potentials are truncated at level $M$ and two copies share the same reinforcement history, the coupled generator inequality $\\mathcal{L} d \\leq -\\delta d$ with $\\delta = \\gamma - (M+2)L$ yields exponential contraction $W_1 \\leq e^{-\\delta t}W_1$ under the dissipativity condition $\\gamma > (M+2)L$. The replica mean-field dynamics, in which representative neurons receive Poisson inputs at self-consistent rates $\\beta_i(t)=P(V_i(t)>\\alpha S_i^+(t))$, admit a unique global solution obtained by a fixed-point contraction on short time intervals, and spatial homogeneity is preserved. The invariant-measure theorem states that the only stationary laws of the nonlinear process are supported on configurations with zero membrane potential, hence with zero firing rate; no active stationary state exists. The numerical results showing $p$-dependent adaptation, threshold growth, and shorter extinction for larger $p$ are presented as evidence that the reinforcement mechanism reproduces activity-dependent response modulation.","pith_inferences":["The numerical match between finite network and RMF suggests a propagation-of-chaos theorem may hold for this bounded-rate ring model; proving it would turn the approximation claim into a mathematical limit.","A natural extension is to introduce decay or reset of the reinforcement variable $S$; the invariant-measure theorem suggests such forgetting is necessary for active stationary regimes, and could be tested by adding a small $S$-decay and checking whether nontrivial stationary measures reappear.","The conditional contraction result does not by itself control the reinforcement variables; averaging over histories could reveal whether exponential stability survives unconditionally or only within the dissipative regime, which a direct simulation of unconditioned Wasserstein trajectories would test.","The finite-window power-law comparison does not establish an asymptotic power law; computing the large-time exponent from the moment equations, for example from $d/dt\\,E[V]$ and $d/dt\\,E[S]$, could distinguish true power-law relaxation from a slow exponential."],"forward_implications":["Under the explicit condition $\\gamma>(M+2)L$, two copies of the network with identical reinforcement history converge in Wasserstein distance at rate $\\delta$, so leakage acts as a quantitative stabilizer of the potential dynamics.","The global existence and uniqueness of the replica mean-field process make the model usable for further analytic study, and ring symmetry reduces it to a single representative neuron with input rate $2\\beta(t)$.","Because all invariant measures of the undriven nonlinear process are silent, any persistent firing in the model must be transient or externally driven; sustained activity requires additional structure such as input or memory decay.","In the driven simulations, larger $p$ yields stronger adaptation-like decline, higher effective thresholds, and shorter extinction times in the undriven setting, so the memory parameter directly controls the duration and intensity of network activity.","The power-law fits dominate single-exponential fits by AIC in all bootstrap resamples over the simulated window, indicating the cumulative reinforcement rule can produce extended temporal response without multiple adaptation variables."],"supporting_citations":[{"why":"supplies the interacting-spiking-neuron framework with variable-length memory that this model adapts.","marker":"[32]"},{"why":"introduces the Elephant random walk whose next increment depends on accumulated past, the origin of the reinforcement rule.","marker":"[66]"},{"why":"furnishes the replica mean-field philosophy and the self-consistency structure used for the nonlinear process.","marker":"[6, 7]"},{"why":"gives the replica mean-field construction for neuronal networks with excitatory and inhibitory activity, the immediate RMF machinery applied here.","marker":"[58]"},{"why":"defines the prior Elephant-reinforced model with the (V,S,K) state that the present bounded-rate version simplifies and analyses.","marker":"[57]"},{"why":"provides the numerical extinction-time study that the present extinction simulations complement and compare with.","marker":"[65]"},{"why":"documents spike-frequency adaptation in real neurons and motivates interpreting firing-rate decline as history-dependent excitability.","marker":"[9]"},{"why":"supplies experimental evidence that spike thresholds shift with firing history, the biological basis for the reinforced threshold.","marker":"[31]"}],"fun_headline_variants":["Elephant reinforcement silences spiking networks","Rigorous convergence in Elephant-memory spiking nets","Replica mean-field captures finite Elephant spiking dynamics","Only silent states are stationary in Elephant spiking nets","Elephant memory provably silences finite spiking networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The replica mean-field process is not proved to be the large-$N$ limit of the finite network; its accuracy rests on the single numerical comparison in Section 6.5, so the approximation could fail outside the tested parameter range.","fun_headline_variants_meta":{"raw":{"variants":["Elephant reinforcement silences spiking networks","Rigorous convergence in Elephant-memory spiking nets","Replica mean-field captures finite Elephant spiking dynamics","Only silent states are stationary in Elephant spiking nets","Elephant memory provably silences finite spiking networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3250,"prompt_tokens":895,"completion_tokens":2355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2279}},"tokens_in":511,"tokens_out":2355,"duration_ms":18109,"temperature":1.0,"reasoning_tokens":2279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:13:29.877900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the finite ring network and the RMF particle approximation under parameter regimes far from Section 6.5, for example with $p<1/2$ or with weak leakage and large $\\alpha$, and compare the mean firing-rate curves by integrated absolute difference over a long time window; if the difference grows with network size or replica count instead of staying small, the claimed mean-field match fails. A direct check of Theorem 5.1 would search, numerically or analytically, for an invariant measure of the nonlinear process with positive firing probability—none should exist because $K$ grows without bound at every firing.","supporting_citations":[{"cited_title":"Galves, E","cited_arxiv_id":null,"evidence_quote":"supplies the interacting-spiking-neuron framework with variable-length memory that this model adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the Elephant random walk whose next increment depends on accumulated past, the origin of the reinforcement rule."},{"cited_title":"Papageorgiou, Replica Mean Field limits for neural networks with excitato ry and inhibitory activity","cited_arxiv_id":null,"evidence_quote":"gives the replica mean-field construction for neuronal networks with excitatory and inhibitory activity, the immediate RMF machinery applied here."},{"cited_title":"Elephant-Reinforced Galves--L\\\"ocherbach Networks","cited_arxiv_id":"2608.10183","evidence_quote":"defines the prior Elephant-reinforced model with the (V,S,K) state that the present bounded-rate version simplifies and analyses."},{"cited_title":"Romaro, F","cited_arxiv_id":null,"evidence_quote":"provides the numerical extinction-time study that the present extinction simulations complement and compare with."},{"cited_title":"Benda, A","cited_arxiv_id":null,"evidence_quote":"documents spike-frequency adaptation in real neurons and motivates interpreting firing-rate decline as history-dependent excitability."},{"cited_title":"Fontaine, J","cited_arxiv_id":null,"evidence_quote":"supplies experimental evidence that spike thresholds shift with firing history, the biological basis for the reinforced threshold."}],"review_version":1}