{"id":"c3b9efae-8d97-4cf0-bf94-80d5ee2a94d0","arxiv_id":"2509.02310","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For any subcritical random connection model with integrable connection function, the cluster size distribution has an exponential tail.","lead":"This paper proves that in a subcritical random connection model, the probability a given point belongs to a cluster of size at least n decays exponentially in n, even when connections can be arbitrarily long. The result settles a sharpness question for continuum percolation with unbounded connection functions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2.13's stated infinite-volume sup is false as written; the proof gives only a fixed-box version, which is what Lemma 2.8 needs, so the gap is likely repairable.","rationale":"Reading the paper in good faith, the stochastic-comparison strategy is coherent and the truncation argument in Section 3 is plausible, so the reader's CONDITIONAL verdict is appropriate. The weakest load-bearing point is indeed the uniformity assertion in Proposition 2.13, but I would sharpen it: the issue is not merely that uniformity is delicate; the proposition as printed is false for the infinite vertex set, while the proof actually establishes only a fixed-box statement. Since Lemma 2.8 operates in a fixed box Lambda_{nR} and then sends N to infinity, the fixed-box statement is sufficient, so the error is repairable and does not undermine the main theorem unless the finite-box proof also fails. I did not find an independent reason to doubt the central claim: the stochastic domination lemmas are standard, the large-deviation estimates are plausible, and the unbounded-support reduction via cut-off connection functions is reasonable, although the proof of Lemma 3.6 via branching-process domination may need to be replaced by a direct tree-product bound for heavy-tailed g. The verdict should remain CONDITIONAL, with a required clarification and correction of Proposition 2.13.","tokens_in":15521,"tokens_out":36527,"duration_ms":364411,"concrete_test":"Replace Proposition 2.13 with the finite-volume statement: for fixed K, P_{N lambda,g}( sup_{x in eta_N^0 cap Lambda_K} P_{1/N,h}(x connects to green in G_g(omega_N cup {x})) > m_h(lambda)+delta ) -> 0. Check that the existing proof with x in Lambda_K and constants depending on K establishes exactly this, and that Lemma 2.8 only needs the case K = nR. To confirm the printed infinite-volume version is false, take d=1, g = 1_{[-1,1]}, fixed N and small h, exhibit a finite auxiliary configuration C making the conditional connection probability exceed any threshold below 1, and apply the ergodic theorem to conclude that translated bad configurations occur infinitely often, so the supremum over all x exceeds the threshold almost surely.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 2.13 is stated with sup over all x in the infinite point process eta_N^0. As written this cannot hold for fixed N whenever m_h(lambda)+delta<1: choose a finite local configuration C in a bounded region such that, given eta_N contains C around 0, the conditional probability P_{1/N,h}(0 connects to green in G_g(omega_N cup {0})) is close to 1. The cylinder event has positive probability, and by ergodicity of the Poisson process translates of C occur infinitely often almost surely, so the supremum over all x exceeds m_h(lambda)+delta almost surely. The proof, however, fixes K, takes x in Lambda_K, and uses the finite lattice L_{K,s}; every bound depends on K through |F_x|, e^{alpha lambda K^d}, and N0 = N0(K,delta,s,alpha). The final line 'for any x in eta_N^0' is therefore unsupported, and no tail estimate outside Lambda_K is supplied. Lemma 2.8 only invokes the estimate on the finite vertex set inside Lambda_{nR}, so replacing Proposition 2.13 by its finite-box analogue (sup over x in eta_N^0 cap Lambda_K with K fixed) would restore the argument. Until that replacement is made explicit, the comparison proof of Lemma 2.8 rests on a false statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves exponential decay of the cluster size tail for subcritical Poisson random connection models (RCM) under the minimal integrability assumption on the connection function, allowing unbounded support. The proof adapts Vanneuville's stochastic comparison method from Bernoulli percolation on vertex-transitive graphs to the RCM. The key new ingredient is an 'asymptotic vertex-transitivity' estimate: the subcritical RCM is represented as site percolation on a high-intensity RCM, and the conditional probability that a vertex connects to a ghost field is shown to be nearly uniform across vertices. Section 2 treats bounded connection functions via discretization, large deviations, and a finite-box stochastic comparison lemma. Section 3 extends to unbounded connection functions by truncation, a dominated convergence argument, and a limiting procedure. The main theorems are Theorem 1.3 (exponential decay for all subcritical intensities) and Theorem 1.5 (a quantitative stochastic comparison bound involving the ghost-field intensity).","tokens_in":15696,"tokens_out":11822,"duration_ms":105342,"significance":"If correct, Theorem 1.3 is a substantial advance: it gives the first exponential decay result for the RCM with long-range (unbounded-support) connection functions under essentially optimal assumptions. The proof is self-contained modulo standard external results (phase transition, branching-process domination, large deviations), and the constants are existential rather than fitted. The 'asymptotic vertex-transitivity' idea is novel and plausibly useful for transferring other discrete percolation results to continuum models. The paper is clearly written and carefully distinguishes bounded-support and unbounded-support cases. The main issues are localized to the statement and proof of Proposition 2.13, and are repairable without changing the overall strategy.","major_comments":[{"comment":"Proposition 2.13 is false as stated if η_N is the infinite Poisson process (or a process on a box larger than Λ_K). The supremum is taken over all x in η_N^0, but the proof only controls x in Λ_K, and all bounds depend on K. For a fixed N, by ergodicity of the Poisson process, local configurations in a bounded region that force the conditional connection probability close to 1 occur infinitely often almost surely, so the supremum over all x exceeds m_h(λ)+δ almost surely whenever m_h(λ)+δ<1. Lemma 2.8 only needs the estimate on the finite vertex set inside Λ_{nR}, so the proposition should be restated with the supremum over x in η_N^0∩Λ_K (with η_N a Poisson process in Λ_K, or with an explicit tail estimate outside Λ_K). This is a load-bearing gap because Lemma 2.8 relies on the estimate, but the repair is local.","section":"Section 2.2, Proposition 2.13"},{"comment":"The proof bounds the conditional probability that x is connected to a green vertex in the discretized graph Ĝ_s(ω_N∪{x}), but the proposition statement concerns the original graph G_g(ω_N∪{x}). The transfer between Ĝ_s and G_g via Lemma 2.12 is not written. Without an explicit argument that the two conditional probabilities differ by a quantity tending to zero in probability, the final conclusion does not follow from the displayed inequality. This is repairable by adding a short coupling argument, but it must be made explicit.","section":"Section 2.2, proof of Proposition 2.13"},{"comment":"Lemma 2.6 is the core stochastic-comparison tool, but its proof is only a one-line reference to [12], asserting that the arguments transfer to site percolation. Since the setting here involves site percolation with a ghost field, and the exploration and pivotal conditions are stated for a joint law, the adaptation is not completely immediate. Please provide a full proof, or at least a detailed statement of how the proof of [12, Lemma 8] is modified to this setting. This is important for the reader to verify the central comparison step.","section":"Section 2.1, Lemma 2.6"}],"minor_comments":[{"comment":"In the final step of the proof, after taking N→∞, the continuity of λ↦P^n_{λ,g}(A) is invoked to let δ↓0. Please add a one-sentence justification, e.g., by writing the probability as a finite sum/integral over the bounded box and using the explicit Poisson distribution.","section":"Section 2.1, proof of Lemma 2.8"},{"comment":"Theorem 1.3 is derived from Theorem 1.5 by referring to [12] without further detail. Since the derivation is short, please include it explicitly so the paper is self-contained.","section":"Section 3"},{"comment":"In Definition 2.15, H_N^λ(z) is defined as |η∩Λ_s(z)| without specifying η; it should be made explicit that η is the high-intensity process η_N from the surrounding context. A small clarification would prevent ambiguity.","section":"Section 2.2, Definition 2.15"},{"comment":"The notation η_N^0 is used in Proposition 2.13 and elsewhere without always specifying whether η_N is restricted to a box. Once the finite-box version is adopted, please define η_N^0 consistently in each statement.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main result is significant and the overall strategy is sound. The principal issue is Proposition 2.13, which is overstated and as written false for an infinite point process; however, the finite-box version needed by Lemma 2.8 is what the proof actually establishes, so the gap is repairable within the manuscript's scope. I recommend major revision rather than rejection. The novelty of the asymptotic transitivity method is real, and the citation practice seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — quick take on Higgs, arXiv:2509.02310. The headline is that the paper proves something new: exponential decay of the cluster size in the subcritical random connection model under only integrability of the connection function. That was previously known only for bounded support (Küpper–Penrose), and the long-range case was genuinely open until now. The strategy is to adapt Vanneuville's stochastic-comparison proof by constructing the subcritical RCM as site percolation on a much denser RCM, then showing the dense graph is \"asymptotically vertex-transitive\". That transfer idea is the real contribution and I expect it will be reused.\n\nThe paper is mostly well argued. Section 3, where they pass from cut-off to full connection functions, is clean: Proposition 3.3 with the dominated-convergence argument is convincing, and Lemma 3.6 is a nice use of the branching-process domination. Lemma 2.16 and Lemma 2.18 are standard large-deviation estimates done carefully. The citation pattern is honest: the bounded-support prior is cited, Vanneuville is cited, and the author explicitly notes that no long-range proof exists elsewhere.\n\nNow the soft spots. The biggest issue is Proposition 2.13. As stated, it claims a uniform bound on the supremum over all x in the infinite point process η_N^0. That is not what the proof gives and, as the stress-test note says, the infinite supremum is false as written for fixed N: with positive probability there are local configurations anywhere that make the conditional green-connection probability close to 1. The proof only establishes the bound for x in a fixed box Λ_K, with constants depending on K. Fortunately Lemma 2.8 only needs the finite-box version, so the argument can be repaired by restating the proposition with sup over x in η_N^0 ∩ Λ_K and carrying K through. Until that is made explicit, the proof of Lemma 2.8 relies on a false statement. This is a real flaw in the manuscript, but a repairable one.\n\nA second, smaller gap: Lemma 2.6 is asserted with \"little more than verifying\" and no proof. I believe it, but a referee will want the verification, especially since the exploration in Definition 2.4 is a bit under-specified for random graphs. Also the transfer from the discretized graph Ĝ_s back to G_g in Proposition 2.13 is somewhat implicit; it relies on Lemma 2.12 and a monotonicity step that should be spelled out.\n\nOn balance: the central theorem is very likely true, the method is good, and the gaps are presentational/fixable. As submitted I would not accept it, but I would certainly send it to a serious referee rather than desk-reject. The right venue is a probability journal. If I were refereeing, I would ask for the finite-box reformulation of Proposition 2.13 and a proof of Lemma 2.6 before recommending acceptance.","headline":"New and likely correct exponential-decay result for subcritical RCM with unbounded connections, held back by an over-stated key proposition that needs a finite-box fix.","tokens_in":16280,"tokens_out":2752,"would_cite":true,"duration_ms":25504,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60D05","82B43"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that in the subcritical random connection model, the probability that the origin's cluster has size at least n decays exponentially in n, for any radial, non-increasing connection function with finite positive integral —…","keywords":["random connection model","continuum percolation","exponential decay","subcritical phase","stochastic comparison","asymptotic transitivity","ghost field","long-range connections"],"falsifier":"Simulate ψ_n(λ) for a subcritical intensity with an unbounded connection function such as g(x) = c(1 + ||x||)^{-d-ε} in dimension 2 at λ = 0.9 λ_c; if log ψ_n(λ) + cn does not stay bounded for some c > 0 as n grows, Theorem 1.3 would fail. Alternatively, test Proposition 2.13 directly: over many samples of the Nλ-intensity RCM, compute the maximum over vertices of the conditional ghost-connection probability; if this maximum exceeds m_h(λ) + δ for large N with non-negligible frequency, the uniform bound collapses.","tokens_in":15239,"feed_emoji":"📉","tokens_out":7678,"duration_ms":71935,"temperature":0.7,"pith_summary":"The paper establishes exponential decay of the cluster-size distribution throughout the subcritical phase of the Poisson random connection model (RCM), under only the minimal assumption that the connection function has finite positive integral. Previously such exponential tails were known only when the connection function had bounded support; the new result covers long-range connection functions, where a cluster can have large diameter while still being very small in size. The proof works by constructing the subcritical RCM as Bernoulli site percolation on a much higher-intensity RCM, then showing that the high-intensity graph becomes 'asymptotically vertex-transitive': probabilities of events around each vertex converge to a common value uniformly over vertices. This transfers a stochastic-comparison method, originally developed for Bernoulli percolation on vertex-transitive graphs, to continuum percolation, and the author expects the method to apply to other continuum models such as the Poisson Boolean model.","feed_headline":"Cluster sizes decay exponentially in subcritical connection models","feed_subtitle":"New proof covers connection functions with unbounded range, previously known only for bounded support.","key_machinery":"The central objects are the ghost field, an independent set of green vertices of intensity 1 − $e^{{-h}}$ that detects large clusters; the construction of the subcritical RCM by keeping each vertex of an RCM of intensity Nλ with probability 1/N; an exploration-and-pivotality comparison lemma showing that conditioning on the ghost field avoiding the origin's cluster changes the percolation law no more than lowering the retention probability by a factor close to 1 − m_h(λ); and the 'asymptotic vertex-transitivity' estimate (Proposition 2.13), which bounds uniformly over all vertices the conditional probability of connecting to the ghost field by m_h(λ) + δ for all large N. The uniformity is achieved by discretising space into small boxes, proving large-deviation estimates for the numbers of vertices and edges in each box, and comparing the thinned high-intensity model with a Poisson-λ reference model via the Harris–FKG inequality.","core_discovery":"Theorem 1.3 states that for every connection function g with 0 < ∫g < ∞ and every intensity λ below the critical intensity λc(g), there exist constants c, C > 0 such that the probability the origin's cluster has at least n vertices is at most C $e^{{-cn}}$ for all n. The engine is a stochastic comparison: the law of the RCM at intensity λ′ = λ(1 − m_h(λ)) is stochastically dominated by the law of an intensity-λ RCM conditioned on a 'ghost field' of green vertices avoiding the origin's cluster, which yields ψ_n(λ′) ≤ (1 − m_h(λ))^{-1} ψ_n(λ) $e^{{-hn}}$. Iterating this relation gives the exponential tail. For bounded connection functions the comparison is proved directly using the asymptotic-transitivity estimate; the general case follows by truncating g at a large radius, proving the truncated models converge to the original via a dominated-convergence argument, and then taking limits.","pith_inferences":["The uniform asymptotic-transitivity estimate likely extends to any translation-invariant point process with finite intensity, so the same two-step argument (high-intensity approximation plus uniform ghost-field control) may prove exponential cluster-volume decay for the Poisson Boolean model as well.","One could test sharpness numerically: for a heavy-tailed g with finite integral, the exponential rate c(λ) in Theorem 1.3 is expected to vanish as λ approaches λc, and the comparison inequality might give explicit bounds on how it vanishes.","The method suggests a general principle: if a continuum percolation model can be approximated by thinning a high-intensity version that becomes 'asymptotically transitive', exponential decay transfers from the transitive discrete setting without needing exact symmetry of the underlying graph.","A direct check of Proposition 2.13 would be to simulate the Nλ-intensity RCM with retention 1/N, compute the maximum over vertices of the conditional probability of connecting to the ghost field, and verify it converges to m_h(λ) as N grows; failure for some connection function would pinpoint exactly where the uniform bound breaks."],"forward_implications":["In every subcritical phase λ < λc, the probability that a cluster contains n vertices is exponentially small, so large clusters are rare even when connection edges are unbounded.","The exponential tail holds in every dimension; in dimension one, the result is unconditional whenever the connection function has finite integral, even if the subcritical condition is trivial.","The stochastic comparison of Theorem 1.5 gives a quantitative relation between cluster-size probabilities at different intensities, which can be used to bound probabilities near criticality.","The truncation argument transfers the result from bounded-support connection functions to long-range ones, and the author notes the same approach would transfer exponential decay from finite-range to long-range percolation on lattices."],"supporting_citations":[{"why":"Supplies the stochastic-comparison method and the ghost-field derivation that the paper adapts to site percolation on the RCM.","marker":"[12]"},{"why":"Established the previous exponential-decay result for bounded-support connection functions, the baseline that Theorem 1.3 extends.","marker":"[7]"},{"why":"Provides the phase-transition theorem and the branching-process domination used in the integrability lemma for the truncation argument.","marker":"[10]"},{"why":"Introduced the idea of realising a lower-intensity model by site percolation on a high-intensity RCM, which the construction uses.","marker":"[1]"},{"why":"Supplies the multivariate Mecke equation used to prove the convergence of truncated models.","marker":"[8]"},{"why":"Provides the Poisson-binomial approximation used in the large-deviation estimates for box vertex counts.","marker":"[9]"},{"why":"Cramér's theorem large-deviations bounds are applied to control the discretised box and edge counts.","marker":"[2]"}],"fun_headline_variants":["Subcritical random connection model: exponential cluster decay","Exponential decay of cluster size in subcritical connection model","Cluster size decays exponentially in subcritical RCM","Random connection model: exponential tail for origin cluster size"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on the asymptotic-vertex-transitivity estimate: in the very high-intensity random connection model with thinning, every vertex's conditional probability of connecting to the ghost field is eventually bounded by the common limiting value m_h(λ) plus a small error, for all vertices simultaneously — this uniformity, not implied by the theorem's assumptions, is the load-bearing new step that replaces exact transitivity.","fun_headline_variants_meta":{"raw":{"variants":["Subcritical random connection model: exponential cluster decay","Exponential decay of cluster size in subcritical connection model","Cluster size decays exponentially in subcritical RCM","Random connection model: exponential tail for origin cluster size"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1355,"prompt_tokens":874,"completion_tokens":481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":490,"tokens_out":481,"duration_ms":4524,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:39:43.155557+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate ψ_n(λ) for a subcritical intensity with an unbounded connection function such as g(x) = c(1 + ||x||)^{-d-ε} in dimension 2 at λ = 0.9 λ_c; if log ψ_n(λ) + cn does not stay bounded for some c > 0 as n grows, Theorem 1.3 would fail. Alternatively, test Proposition 2.13 directly: over many samples of the Nλ-intensity RCM, compute the maximum over vertices of the conditional ghost-connection probability; if this maximum exceeds m_h(λ) + δ for large N with non-negligible frequency, the uniform bound collapses.","supporting_citations":[{"cited_title":"Exponential decay of the volume for Bernoulli percolation: a proof via stochastic comparison","cited_arxiv_id":"2304.12110","evidence_quote":"Supplies the stochastic-comparison method and the ghost-field derivation that the paper adapts to site percolation on the RCM."},{"cited_title":"Largest component and sharpness in continuum percolation","cited_arxiv_id":"2407.10715","evidence_quote":"Established the previous exponential-decay result for bounded-support connection functions, the baseline that Theorem 1.3 extends."},{"cited_title":"Continuum Percolation","cited_arxiv_id":null,"evidence_quote":"Provides the phase-transition theorem and the branching-process domination used in the integrability lemma for the truncation argument."},{"cited_title":"Noise sensitivity in continuum percolation","cited_arxiv_id":null,"evidence_quote":"Introduced the idea of realising a lower-intensity model by site percolation on a high-intensity RCM, which the construction uses."},{"cited_title":"Lectures on the Poisson process , volume 7 of Institute of Mathematical Statistics Textbooks","cited_arxiv_id":null,"evidence_quote":"Supplies the multivariate Mecke equation used to prove the convergence of truncated models."},{"cited_title":"An approximation theorem for the Poisson binomial distri- bution","cited_arxiv_id":null,"evidence_quote":"Provides the Poisson-binomial approximation used in the large-deviation estimates for box vertex counts."},{"cited_title":"Large deviations, volume 14 of Fields Institute Mono- graphs","cited_arxiv_id":null,"evidence_quote":"Cramér's theorem large-deviations bounds are applied to control the discretised box and edge counts."}],"review_version":1}