{"id":"e84dfd6b-6573-4247-b63e-a8423e7c6bb5","arxiv_id":"2608.10183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A Galves-Loecherbach spiking network with bounded elephant-style synaptic reinforcement is shown to be non-explosive, conditionally Wasserstein-contractive, and to admit a conditional replica mean-field equation under the Poisson hypothesis.","lead":"This paper defines a network of spiking neurons in which each past excitatory or inhibitory event shifts the probability of the next event, like an elephant random walk. It proves the network does not explode, gives a Wasserstein contraction for membranes sharing the same history, and writes a mean-field equation under the Poisson hypothesis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The replica mean-field equation depends on unproved Poisson Hypothesis and existence of invariant replica laws; if these fail, Eq. (7.1) need not describe the stationary limit.","rationale":"The reader's verdict and weakest-assumption identification match my read. Sections 3-6 give a coherent Lyapunov/non-explosion argument and a quenched Wasserstein contraction with explicit dissipativity conditions; I did not find an internal error there. Section 7 is honestly labeled conditional, but the central advertised mean-field derivation depends on strong hypotheses that the paper does not prove. The correct classification is CONDITIONAL rather than REJECT because the identity is a valid conditional statement: if the invariant replica laws and the stated convergence properties hold, the generator computation yields Eq. (7.1). The paper's own limitation disclosures, especially Remarks 1.1, 7.3, and 7.4, are explicit and should be preserved in any revised version. No machine-checked formalization or independent numerical verification is provided for Section 7, so the burden remains on the unproved assumptions. My concern does not change the reader's verdict.","tokens_in":31639,"tokens_out":10926,"duration_ms":123434,"concrete_test":"For the simplest nontrivial feedback case, a two-neuron network with mutual interactions, linear rates, reset values r_i, and a frozen environment, derive the stationary joint MGF as M -> infinity without imposing independence. If the limit of E[(X_1^n+z_1)e^{uX_2^m}] is not beta_1 Lambda_2(u), for instance because the joint MGF retains a nonzero mixed cumulant, then the factorization in Proposition 7.2 fails and Eq. (7.1) is not the correct stationary limit. Alternatively, simulate the M-replica system for M = 2, 4, 8, 16 and measure the normalized cross-covariance under the empirical stationary law; it should decay at rate 1/M for the Poisson Hypothesis to hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 7's advertised result is Proposition 7.2: the stationary transform satisfies Eq. (7.1) under the Poisson Hypothesis and under assumed existence, exchangeability, exponential integrability, and convergence of pi^eta_M, Lambda_M, Theta_M, and beta_M. The key step is the factorization E_{pi^eta_M}[(X_j^n+z_j)e^{u(X_i^m+z_i)}] -> beta_j Lambda_i^eta(u), used to replace the incoming replica sum by beta_j times the target marginal transform; this is exactly the Poisson-Hypothesis assertion. The paper explicitly notes in Remarks 7.3 and 7.4 that no existence or uniqueness theorem for the invariant measures, no propagation-of-chaos theorem, and no convergence theorem for the M-replica process is proved. This is not a formal inconsistency, but it is load-bearing: without a decorrelation mechanism, the limiting equation may contain extra covariance terms, and Eq. (7.1) need not describe the stationary behavior of the infinite network. The non-explosion and conditional contraction theorems are independent of this issue and appear internally sound; the concern is specifically the unconditional reading of the replica mean-field claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an infinite-dimensional Galves–Löcherbach neuronal network in which the signs of synaptic interactions are reinforced by an Elephant-random-walk-type memory mechanism. The state is the lifted PDMP Y_t = (X_t, N_t, S_t), where X_t are membrane potentials and (N_t, S_t) are reinforcement counters. The main results are: Theorem 3.8, non-explosion and finite expected total spike count under assumptions (3.1)–(3.2); Theorem 4.2, exact first-moment identities for the reinforcement variables; Theorem 5.5, a quenched Wasserstein contraction for two membrane-potential processes sharing a prescribed reinforcement profile under conditions (5.4)–(5.7); Theorem 6.3, a quantitative stability estimate when the two profiles differ under an additional bounded-rate assumption; and Proposition 7.2, a stationary replica mean-field identity (7.1) in a frozen environment, conditional on the Poisson Hypothesis and on the existence and convergence of invariant replica laws.","tokens_in":31841,"tokens_out":12402,"duration_ms":129477,"significance":"If the main results hold, the paper provides a rigorous infinite-dimensional model with complete-memory reinforcement in synaptic signs. The Lyapunov non-explosion argument and the maximal-coupling contraction estimates are detailed and internally coherent, and the paper is explicit that the Wasserstein contraction is quenched and does not apply to the full endogenous process (X, N, S). These proven parts are genuine contributions. The replica mean-field section is not a theorem about the actual infinite network: Proposition 7.2 is conditional on the Poisson Hypothesis and on unproved existence, exchangeability, and convergence assumptions for the finite-replica invariant laws, as the paper's own Remarks 7.3–7.4 acknowledge. The contribution should therefore be assessed as a rigorous non-explosion and quenched-stability analysis together with a conditional RMF calculation, not as a derivation of the stationary behavior of the infinite reinforced network.","major_comments":[{"comment":"The advertised replica mean-field identity is conditional on several unproved premises: existence of the invariant measures pi^eta_M, replica exchangeability, the Poisson Hypothesis, and convergence of Lambda_M, Theta_M, and beta_M. In particular, the key limiting step in the proof, E_{pi^eta_M}[(X_j^n+z_j)e^{u(X_i^m+z_i)}] -> beta_j Lambda_i^eta(u), is exactly the asymptotic-independence assertion of the Poisson Hypothesis and is not proved. Without a decorrelation mechanism, the limiting stationary equation could contain additional covariance terms, so Eq. (7.1) need not describe the stationary behavior of the infinite network. This is load-bearing for one of the three advertised contributions. The abstract and introduction should either state all these hypotheses explicitly or present Section 7 as a conditional identification/calculation rather than as a derivation of the replica mean-field equation.","section":"Section 7, Proposition 7.2 and Remarks 7.3–7.4"},{"comment":"The existence of the invariant probability measures pi^eta_M is assumed for the finite-replica system, but under Section 7 the firing rates are linear, phi_i(x)=x+z_i, and the global summability assumptions (3.1)–(3.2) are explicitly not required. Since the M-replica system is still infinite-dimensional, existence of an invariant measure is a nontrivial statement and is not a consequence of the non-explosion theorem proved earlier. The paper acknowledges this in Remarks 7.3–7.4, but the lack of any well-posedness or existence result means the finite-M generators and expectations used in the proof of Proposition 7.2 are only defined under an additional, unverified hypothesis. This should be stated as an open problem or addressed directly.","section":"Section 7, paragraph defining pi^eta_M"},{"comment":"The Poisson Hypothesis is used without any supporting argument specific to this Elephant-reinforced model. The replica construction routes interactions uniformly among the M replicas, but the states of different replicas remain dependent through the shared routing mechanism and through the common frozen environment. The paper does not quantify how M to infinity kills these correlations, nor does it cite a propagation-of-chaos theorem for this class of processes. Because this is the step that turns the finite-M stationary equations into the closed self-consistent equation (7.1), the authors should either provide a proof attempt, a precise conjecture with the missing estimates, or a clear statement that the derivation is formal.","section":"Section 7, Poisson Hypothesis paragraph"}],"minor_comments":[{"comment":"The phrase 'we derive the corresponding replica mean-field equation under the Poisson hypothesis' should mention the additional assumed existence and convergence of invariant replica laws; otherwise the abstract overstates the status of Eq. (7.1).","section":"Abstract and Introduction"},{"comment":"The sentence 'if the initial state is admissible, then for every time for which the process is defined as in (2.1)' is grammatically incomplete; it should say that the admissibility conditions (2.1) hold at all times for which the process is defined.","section":"Section 2, after Eq. (2.1)"},{"comment":"There are many typographical artifacts, including 'Galves–L\\\"ocherbach', 'c\\` adl\\` ag', and inconsistent spacing in displayed formulas; a careful proofreading pass is needed.","section":"Throughout"},{"comment":"Reference [53] is listed as 'arXiv, 2026' without an arXiv number or further publication data; it should be completed or cited as a preprint with an identifier.","section":"References"},{"comment":"The constant C_c is defined in Lemma 5.4 but never used afterward; either use it in the integrability estimate or remove it to avoid confusion.","section":"Lemma 5.4"},{"comment":"In the step after applying the stopped Dynkin formula, the text writes 'Since the environments are prescribed, the function D is deterministic' and then replaces the expectation of the integral up to t wedge tau_R by the full integral up to t; this is only an upper bound, not an equality, and the wording should be corrected.","section":"Theorem 6.3, proof"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about the conditional status of Section 7, and the non-explosion and quenched-contraction portions appear sound. The main editorial issue is that the abstract and introduction give the replica mean-field result equal billing with the proven theorems. If the RMF part is repositioned as a conditional calculation or conjecture, with the abstract adjusted accordingly, the paper is likely publishable. The self-citations [21, 55, 56] are contextual and do not appear to carry the load-bearing content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead this paper for the non-explosion and contraction results. The elephant-style bounded reinforcement of synaptic signs in an infinite Galves–Löcherbach network is genuinely new, and the analytic core is in good shape. The paper proves global existence under explicit summability assumptions, gets exact first-moment identities, and establishes exponential contraction in a weighted Wasserstein distance for two membrane-potential processes driven by a common reinforcement profile. There is also a perturbation estimate for different reinforcement environments, with a bounded-rate assumption. These parts are carefully argued; I did not find a gap in the Lyapunov or coupling reasoning.\n\nThe soft spot is exactly where the abstract points: the replica mean-field equation in Section 7. The author is honest that it depends on the Poisson Hypothesis and on existence, exchangeability, integrability, and convergence of invariant replica laws, none of which are proved. So treat (7.1) as a formal stationary identity, not a theorem. If those assumptions fail, the equation need not describe the infinite network. The stress-test note is on target here. The paper says this itself in Remarks 7.3 and 7.4, so there is no hidden circularity—just an unproved premise that limits the scope of the mean-field claim.\n\nMinor caveats: the contraction assumptions (5.4)–(5.7) are somewhat restrictive, and the theorem is quenched, so it does not control the full endogenous process (X,N,S). The author flags this repeatedly, which I appreciate. The stability estimate also requires bounded firing rates, which is a real restriction. None of this undercuts the main theorems.\n\nCitation pattern is fine: the self-citations are background and related technology, not load-bearing. This paper is for people working on Galves–Löcherbach networks, PDMPs, or elephant-random-walk variants; those readers will find reusable tools here.\n\nOverall: a serious, honest paper. I would send it to a knowledgeable referee, with a note that the RMF section should be framed as a conditional calculation and that the referee should push on whether the Poisson Hypothesis can be verified in any nontrivial regime. I would cite the contraction tools in my own work if I were doing GL-type networks.","headline":"Solid non-explosion and quenched contraction results for a genuinely new elephant-type GL network; the replica mean-field section is a well-labeled conditional derivation, not a theorem.","tokens_in":32356,"tokens_out":2797,"would_cite":true,"duration_ms":28179,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J25","60K35","60J75","92B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"An infinite spiking-neuron network with Elephant-type reinforcement is non-explosive, contracts exponentially in a fixed environment, and satisfies a replica mean-field equation under Poisson assumptions.","keywords":["Galves–Löcherbach model","Elephant random walk","reinforcement","piecewise-deterministic Markov processes","non-explosion","Wasserstein contraction","replica mean field","interacting spiking neurons"],"falsifier":"Run two coupled copies of a small network with rates $g(x)=x$, $\\phi_i(x)=x+z_i$, and weights satisfying (5.4)–(5.7) in a common reinforcement profile; if the measured $H(X_t,\\tilde X_t)$ ever exceeds $e^{-dt}H(X_0,\\tilde X_0)$, the contraction theorem's hypotheses are insufficient or misstated. Alternatively, simulate the frozen $M$-replica system and test whether the left-hand side of (7.1), evaluated at the empirical stationary transform, converges to zero as $M$ grows; a persistent discrepancy would falsify the Poisson Hypothesis or the assumed convergence.","tokens_in":1929,"feed_emoji":"🐘","tokens_out":4649,"duration_ms":105081,"temperature":0.7,"pith_summary":"The paper introduces a spiking-neuron network in which the sign of each synaptic event, excitatory or inhibitory, is chosen by a complete-memory reinforcement rule borrowed from the Elephant random walk, while the event's amplitude stays bounded. It proves three types of results: the full Markovian lift of the process never explodes and has finite expected spike count on compact time intervals; two membrane-potential systems sharing the same reinforcement profile converge exponentially in a weighted 1-Wasserstein distance; and, under the Poisson Hypothesis together with assumed stationary replica laws, the stationary law of a frozen-environment system satisfies an explicit functional equation. A sympathetic reader would care because the reinforcement mechanism is a minimal way to give neurons history-dependent synaptic plasticity while preserving the analytical tools of piecewise-deterministic Markov processes and mean-field approximations.","feed_headline":"Infinite neuron networks with elephant memory never explode","feed_subtitle":"Reinforced spiking networks stay finite-time, contract in fixed environments, and satisfy an explicit mean-field equation.","key_machinery":"The load-bearing object is the lifted piecewise-deterministic Markov process $Y_t=(X_t,N_t,S_t)$ with generator (2.2), where the reinforcement law is $Q_i(+1)=\\frac12+\\frac{2p-1}{2}\\frac{S_i}{N_i}$ for $N_i\\ge1$—the sign-probability analogue of the Elephant random walk's drift. Non-explosion is carried by the Lyapunov function $h(x)=\\sum_i \\ell_i x_i$ and a localized Dynkin argument; contraction is carried by a maximal coupling of firing clocks that uses identical reinforced signs at simultaneous spikes, measured in $H(x,y)=\\sum_i q_i|x_i-y_i|$, with the dissipativity condition $d=k_1\\inf_i(a_i-c_i/q_i)>0$; and the mean-field identity is carried by the moment-generating transform $\\Lambda_i^\\eta(u)=E^\\eta[e^{u(X_i+z_i)}]$ together with the truncated transform $\\Theta_i^\\eta(u,v)$ for inhibitory jumps through the positive-part map. These three mechanisms are what make the infinite-dimensional system finite, contractive, and self-consistent.","core_discovery":"The central claim is that adding Elephant-type complete-memory reinforcement to the Galves–Löcherbach model does not destroy its tractability. Theorem 3.8 proves non-explosion of the minimal cadlag solution and $E[N^{\\mathrm{sp}}[0,T]]<\\infty$ under assumptions (3.1)–(3.2). Theorem 4.2 gives exact first-moment equations for the reinforcement counter $N^j$ and signed balance $S^j$, namely $\\frac{d}{dt}E[N^j_t]=E[\\Lambda_j(X_t)]$ and $\\frac{d}{dt}E[S^j_t]=(2p-1)E[r_j(N_t,S_t)\\Lambda_j(X_t)]$, the continuous-time analogue of the Elephant drift. Theorem 5.5 proves that, in a fixed reinforcement environment, $W_{1,H}(\\mu P^{\\eta}_{0,t},\\nu P^{\\eta}_{0,t})\\le e^{-dt}W_{1,H}(\\mu,\\nu)$ under conditions (5.4)–(5.7). Theorem 6.3 extends this to two different environments with an integrated discrepancy term. Finally, Proposition 7.2 identifies the conditional stationary replica mean-field equation (7.1) satisfied by the limiting transform under the Poisson Hypothesis and the existence of invariant replica laws.","pith_inferences":["If the Poisson Hypothesis is eventually proved in this setting, equation (7.1) would become a genuine closed mean-field description; however the presence of $\\Theta_i^\\eta$ shows that inhibitory truncation prevents closure in the transform alone, so a full solution would require an additional boundary distribution.","The quenched contraction result suggests a separation of time scales: conditional on a frozen synaptic history, potentials mix at rate $d$, while the reinforcement variables evolve on a slower incoming-spike scale; the paper does not establish this two-scale property.","The bounded-amplitude reinforcement can be read as a minimal plasticity rule that changes only sign bias, and the non-explosion proof relies essentially on this boundedness; relaxing to unbounded amplitudes would require a new Lyapunov argument."],"forward_implications":["The minimal cadlag process is globally well defined and has finite expected spike count on bounded time intervals, so the model supports simulation and further statistical analysis.","In a fixed reinforcement environment, two membrane-potential systems with the same initial distance converge exponentially at rate $d$, giving a quantitative synchronization time scale conditional on the shared history.","For two different reinforcement environments with uniformly bounded rates and bounded discrepancy, the Wasserstein distance is controlled by an exponentially decaying initial term plus an accumulated discrepancy term, so the potential dynamics depend continuously on the excitation–inhibition environment.","Under the Poisson Hypothesis and the assumed invariant replica measures, the stationary transform of the infinite network satisfies equation (7.1), which can be used to compute rate functions and moment relations.","The exact moment identities (4.5)–(4.6) show that the mean reinforcement balance evolves like a continuous-time Elephant random walk with state-dependent incoming rate $\\Lambda_j(X_t)$."],"supporting_citations":[{"why":"Supplies the Elephant random walk whose complete-memory sign mechanism is transferred to synaptic interactions.","marker":"[62]"},{"why":"Introduces the Galves–Löcherbach spiking-neuron model that the paper extends with reinforcement.","marker":"[30]"},{"why":"Provides the piecewise-deterministic Markov process framework and generator formalism used for the lifted process.","marker":"[26]"},{"why":"Establishes infinite Galves–Löcherbach networks under assumptions beyond uniform summability, the setting Theorem 3.8 extends.","marker":"[56]"},{"why":"Supplies the replica mean-field limit methodology used in Section 7.","marker":"[6]"},{"why":"Continues the pair-replica mean-field methodology underpinning the Poisson-hypothesis derivation.","marker":"[7]"},{"why":"Gives the preceding replica mean-field derivation for excitatory and inhibitory networks that the conditional RMF identity generalizes.","marker":"[55]"},{"why":"Provides numerical simulations of finite Elephant-reinforced networks against which the analytic results can be checked.","marker":"[53]"}],"fun_headline_variants":["Elephant-reinforced spiking nets: no explosion","Spiking networks with elephant memory: finite and tractable","Reinforced Galves-Löcherbach: contraction and mean-field","Complete-memory reinforcement yields non-explosion in spiking nets","Elephant memory keeps spiking networks from exploding"],"cache_read_input_tokens":34560,"weakest_assumption_plain":"The replica mean-field equation stands on two unproved assumptions: the Poisson Hypothesis and the existence and convergence of invariant measures for the finite-replica system; if either fails, the equation need not describe the stationary behaviour of the infinite network.","fun_headline_variants_meta":{"raw":{"variants":["Elephant-reinforced spiking nets: no explosion","Spiking networks with elephant memory: finite and tractable","Reinforced Galves-Löcherbach: contraction and mean-field","Complete-memory reinforcement yields non-explosion in spiking nets","Elephant memory keeps spiking networks from exploding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001051,"raw_usage":{"total_tokens":4384,"prompt_tokens":882,"completion_tokens":3502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3419}},"tokens_in":498,"tokens_out":3502,"duration_ms":24933,"temperature":1.0,"reasoning_tokens":3419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:12:17.528414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run two coupled copies of a small network with rates $g(x)=x$, $\\phi_i(x)=x+z_i$, and weights satisfying (5.4)–(5.7) in a common reinforcement profile; if the measured $H(X_t,\\tilde X_t)$ ever exceeds $e^{-dt}H(X_0,\\tilde X_0)$, the contraction theorem's hypotheses are insufficient or misstated. Alternatively, simulate the frozen $M$-replica system and test whether the left-hand side of (7.1), evaluated at the empirical stationary transform, converges to zero as $M$ grows; a persistent discrepancy would falsify the Poisson Hypothesis or the assumed convergence.","supporting_citations":[{"cited_title":"Najman, I","cited_arxiv_id":null,"evidence_quote":"Provides numerical simulations of finite Elephant-reinforced networks against which the analytic results can be checked."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Elephant random walk whose complete-memory sign mechanism is transferred to synaptic interactions."},{"cited_title":"Galves, E","cited_arxiv_id":null,"evidence_quote":"Introduces the Galves–Löcherbach spiking-neuron model that the paper extends with reinforcement."},{"cited_title":"Davis,Piecewise-deterministic Markov processes: a general class of non-diffusion stochastic models","cited_arxiv_id":null,"evidence_quote":"Provides the piecewise-deterministic Markov process framework and generator formalism used for the lifted process."},{"cited_title":"Papageorgiou,Interacting systems of infinite spiking neurons with weights beyond uniform summability","cited_arxiv_id":null,"evidence_quote":"Establishes infinite Galves–Löcherbach networks under assumptions beyond uniform summability, the setting Theorem 3.8 extends."},{"cited_title":"Baccelli, T","cited_arxiv_id":null,"evidence_quote":"Supplies the replica mean-field limit methodology used in Section 7."},{"cited_title":"Baccelli, T","cited_arxiv_id":null,"evidence_quote":"Continues the pair-replica mean-field methodology underpinning the Poisson-hypothesis derivation."},{"cited_title":"Papageorgiou,Replica Mean Field limits for neural networks with excitatory and inhibitory activity","cited_arxiv_id":null,"evidence_quote":"Gives the preceding replica mean-field derivation for excitatory and inhibitory networks that the conditional RMF identity generalizes."}],"review_version":1}