{"id":"b64ac373-167b-4d34-9244-bfe3ba2c929e","arxiv_id":"1908.07203","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For two new models in which occupied sites draw line segments to their nearest occupied neighbors, the paper proves that infinite connected clusters exist for sparse sites and do not exist for dense sites, and it establishes partial bounds on the independent model's phase boundary.","lead":"This paper introduces two new models of random line segments on a lattice and determines when they form an infinite connected network. In one model, an infinite network appears when the lattice is sparsely populated and disappears when it is densely populated.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.4(i) rests on an unproved branching-process domination; the proof is explicitly deferred ('may be written out formally').","rationale":"The reader's weakest assumption identifies the same gap, and it is the most load-bearing concern because Theorem 2.4(i) is the only non-percolation result for the independent model at low lambda, and its proof is explicitly deferred. The other potential weakness (the Burton-Keane adaptation in Theorem 5.2) is also sketched, but the branching-process bound is the one the authors themselves flag as needing formalization. A fully formal proof may exist (the means and heuristics are plausible), so the appropriate action is to keep the paper under conditional acceptance pending that proof, not to reject. No independent check of the central claim's truth is proposed; the required check is completeness of the argument.","tokens_in":14222,"tokens_out":18538,"duration_ms":191463,"concrete_test":"Provide a complete proof of the claimed stochastic domination in Section 4: define the cluster exploration algorithm precisely, and prove by induction that at each visited occupied site the vector of newly discovered sites is stochastically dominated by (2d-1) independent copies of a variable that is 0 with probability 1-lambda and geometric(p) with probability lambda (and similarly for unoccupied sites, with 2(d-1) copies). If the induction cannot be carried out, exhibit a configuration or history where the conditional offspring distribution exceeds the claimed mean, which would break Theorem 2.4(i).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4's proof of Theorem 2.4(i) asserts that the blue cluster at the origin is stochastically bounded by a two-type branching process with mean offspring mu1=(2d-1)lambda/p for occupied sites and mu2=2(d-1)lambda/p for unoccupied sites, and concludes no percolation when mu1<1. No formal domination or coupling is supplied; the text says only that the argument 'may be written out formally' and that 'complications arising from the conjunction of probability and combinatorics' do not interfere. The missing step is nontrivial: one must show that, conditional on the exploration history, the number of newly discovered sites from an occupied/unoccupied site is stochastically dominated by independent mixtures of Bernoulli(lambda) and geometric(p) offspring, and that double-counting can only reduce the count. Without this, the conclusion lambda < p/(2d-1) does not follow from the reasoning given. If the domination fails, Theorem 2.4(i) is unsupported and the lower phase boundary of the independent model is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines two dependent percolation models on Z^d, d>=2, built over an occupied-site set with density p. In the one-choice model, each occupied site chooses one of its incident feasible segments uniformly and the chosen segments are blue; in the independent model, each feasible segment is blue with probability lambda independently. The authors prove partial phase diagrams. For the one-choice model, Theorem 2.1 asserts the existence of p0(d) <= p1(d) such that there is a unique infinite blue cluster for p < p0(d) and no infinite blue cluster for p > p1(d). For the independent model, Theorem 2.4 gives no percolation when lambda < p/(2d-1), percolation when lambda > c log(1/q), and percolation above a Lipschitz curve lambda_c(p) obtained by comparison with mixed site-bond percolation. The proofs use a coupling to the corrupted compass model, a branching-process upper bound, a finite-energy argument, a Burton-Keane-style uniqueness analysis adapted to the one-choice model, and a block construction with 1-dependent site percolation.","tokens_in":14377,"tokens_out":16945,"duration_ms":186721,"significance":"If the results are correct, the paper establishes rigorous, nontrivial phase transitions for two natural models of intersecting random line segments, with the counterintuitive feature that percolation occurs at low site density and fails at high density in the one-choice model. The block argument in Theorem 5.3 is quantitative and gives explicit exponential control, and the comparison with mixed percolation in Theorem 2.4(iii) is clean and parameter-free. The paper is also careful to separate proved results from conjectures and numerical estimates. The main weakness is that one of the two central lower-bound proofs is explicitly presented as a sketch rather than a formal argument, and the uniqueness proof for the one-choice model contains an abbreviated surgery step; both need to be supplied before the theorems can be regarded as fully proved.","major_comments":[{"comment":"The proof of the non-percolation region for the independent model is not complete. The text asserts that the size of the blue cluster at the origin is stochastically bounded by a two-type branching process with mean offspring mu1 = (2d-1)lambda/p for occupied sites and mu2 = 2(d-1)lambda/p for unoccupied sites, but no formal domination or coupling is provided. The manuscript states only that 'the above argument may be written out formally' and that 'there are some complications arising from the conjunction of probability and combinatorics, but these do not interfere with the conclusion.' This is a load-bearing step: if the domination fails, the conclusion lambda < p/(2d-1) does not follow. A formal proof must specify the exploration order, show that, conditional on the exploration history, the number of newly discovered sites from a site of each type is stochastically dominated by the claimed offspring distributions, and handle the dependencies created by previously counted sites and by overlapping feasible segments. In particular, the heuristic that the 'true conditional expectations will typically be less' is not by itself an inequality. This gap is especially important because part (i) is the only proof of the lower phase boundary for the independent model.","section":"Section 4, Proof of Theorem 2.4(i)"},{"comment":"The proof of uniqueness for the one-choice model is also abbreviated at a central point. After describing the construction of three paths rho1, rho2, rho3 from the segment S to three boundary sites z1,z2,z3, the text says 'It may be checked that this construction is always possible for m >= 7' and concludes that the origin is a trifurcation. This check is load-bearing for the a.s. uniqueness assertion in Theorem 2.1(i). The proof should explicitly verify that the three paths are disjoint except at their attachment points to S, that all edges on them are blue under the altered choices, that the surgery inside Dm does not create unintended blue connections, and that deleting the origin separates at least three distinct infinite components. Since the one-choice measure lacks finite energy, the usual Burton-Keane framework cannot simply be imported without this verification.","section":"Section 5, Theorem 5.2, proof of statement C"}],"minor_comments":[{"comment":"The function lambda_c(p) defined in (4.3) for the mixed percolation model is denoted by the same symbol as the conjectured critical curve in Section 2.2. The authors should state explicitly that Theorem 2.4(iii) uses the mixed-percolation critical curve, not the conjectured independent-model frontier, to avoid confusion.","section":"Section 4, Proof of Theorem 2.4(iii)"},{"comment":"The theorem statements say lambda > c log(1/q) without specifying the base of the logarithm, while the proof uses log_2(1/q). Since the constant c is absolute, the base is immaterial, but the statements should either specify a base or note that the constant absorbs the choice.","section":"Theorem 5.3 and Theorem 2.4(ii)"},{"comment":"The choice r = 1/log_2(1/q) is not generally an integer; the text says a 'small correction' is necessary and is overlooked. For a formal proof, the authors should either define r as a floor/ceiling and adjust the subsequent estimates, or include a sentence explaining that the error is absorbed by the constants c1 and c2.","section":"Equation (5.5)"},{"comment":"Reference [16] is listed as 'to appear' in the Annals of Applied Probability; please update the citation with the final publication data or a stable arXiv identifier, since Section 3 relies directly on its non-percolation criterion.","section":"References"},{"comment":"The added remark on the work of Hilário and Ungaretti is useful, but the comparison with the hexagonal-lattice threshold p_H^c should be stated with the same notational conventions, since the reader must infer that p_H^c is a bond-percolation threshold on the hexagonal lattice rather than on Z^d.","section":"Remark 2.6"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the unproved branching-process domination in Section 4; it is explicitly deferred and is the sole evidence for Theorem 2.4(i). I do not think this warrants rejection, because the claimed bound is plausible and a formal exploration argument could plausibly be written out. A second concern is that Theorem 2.1(ii) relies entirely on the corrupted-compass-model non-percolation criterion of [16], which is by one of the authors and was 'to appear' at the time of this version; the editor should confirm that this companion paper is in press and publicly available. The numerical estimates in Remarks 2.3 and 2.7 are clearly labeled as simulations and do not affect the proofs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one for the one-choice model alone: Beaton, Grimmett, and Holmes introduce two line-segment percolation models based on site percolation, and prove that the one-choice model percolates for small p and fails for large p. That inverted transition is real and fresh; I don't know of another model with that behavior. The independent model is also natural, with a partial phase diagram mixing bond and site thresholds.\n\nWhat's new: the models themselves, and the observation that blue edges become more aligned as p decreases, flipping the usual monotonicity. The proofs are mostly adaptations: the mixed percolation comparison for Theorem 2.4(iii) is elegant and rigorous; the block argument for supercriticality (Theorem 5.3) is careful and gives the right 1-dependent domination; the finite-energy computation for the independent model is solid.\n\nThe soft spot is exactly where the stress-test lands. The proof of Theorem 2.4(i) — no percolation when λ < p/(2d−1) — is not a proof. The text says the branching-process domination \"may be written out formally\" and that the combinatorial complications \"do not interfere.\" That is not a theorem as written. One needs an honest coupling: conditional on the exploration history, the number of new sites discovered from an occupied or unoccupied site is stochastically bounded by the claimed two-type offspring distribution, and no double-counting can inflate the bound. That is plausible, maybe even standard, but it is not supplied. So the lower boundary of the independent model's phase diagram is currently unsupported as stated.\n\nI'd also want the Burton-Keane adaptation in Theorem 5.2 checked closely. The outline is detailed enough that I don't suspect a hole, but the surgery on sites and directions is intricate, and \"the arguments of [4] may be adapted\" is doing some work there.\n\nThe self-citation to Hirsch–Holmes–Kleptsyn for the corrupted compass model is fine; that's a real paper and the reduction is clean.\n\nBottom line: the paper is worth a serious referee. The central phenomenon is interesting enough that the gap in 2.4(i) should not sink it, but the authors should either supply the missing formal domination or soften the claim to a conditional statement. I'd cite the one-choice theorem in my own work once the details are in place.\n\nRecommendation: send to peer review; conditional acceptance, with the missing proof required before publication.","headline":"A genuinely new pair of percolation models with a counterintuitive inverted transition; the main ideas are sound, but one theorem is explicitly left as a sketch and needs a formal proof before the claims are fully supported.","tokens_in":14901,"tokens_out":2427,"would_cite":true,"duration_ms":24649,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that in two stochastic models of blue line segments on the integer lattice in dimension at least two, sparse occupancy almost surely gives a unique infinite blue cluster while dense occupancy almost surely gives none.","keywords":["percolation","one-choice model","independent model","phase transition","blue cluster","site percolation","line segments","random geometric graph"],"falsifier":"In the independent model on $\\mathbb{Z}^2$, take $p=0.6$ and $\\lambda=0.19$, just below $p/(2d-1)=0.2$; if the largest blue cluster in boxes of side $L$ grows linearly with $L$ as $L\\to\\infty$, the no-percolation claim of Theorem 2.4(i), and the branching-process domination behind it, would be wrong.","tokens_in":14009,"feed_emoji":"🔵","tokens_out":9543,"duration_ms":91736,"temperature":0.7,"pith_summary":"This paper treats two random models built on site percolation on $\\mathbb{Z}^d$: each occupied site, or each feasible segment, chooses blue line segments connecting nearest occupied sites. Its central claim is that both models undergo a phase transition, but backwards from classical percolation: when occupied sites are sparse, an infinite blue cluster exists almost surely, and when they are dense, it disappears almost surely. For the one-choice model, the paper proves rigorous upper and lower bounds on this transition; for the independent model it proves a non-percolation region, two percolation regions, and uniqueness of the infinite cluster whenever one exists. A reader should care because the blue segments' density is essentially independent of $p$, so the transition is driven by geometric alignment and long-range correlations rather than by the number of blue edges.","feed_headline":"Sparse blue segments percolate; dense ones do not","feed_subtitle":"Two random line-segment models show a phase transition driven by alignment, not by how many edges are blue.","key_machinery":"The argument is carried by four main mechanisms. The upper bound for the one-choice model uses a coupling in which every blue edge is contained in a 'corrupted compass model' of turquoise edges; non-percolation of that model for large $p$ is read off from the eigenvalues of an explicit $3\\times 3$ matrix. The non-percolation bound for the independent model rests on a two-type branching process whose mean offspring are $(2d-1)\\lambda/p$ for occupied sites and $2(d-1)\\lambda/p$ for unoccupied sites, with extinction implying $\\lambda < p/(2d-1)$. Uniqueness of infinite clusters is obtained through the finite-energy property for the independent model and through a surgical adaptation of the classical trifurcation argument for the one-choice model, which lacks finite energy. Percolation at small $p$ is proved by a block construction on boxes of side $6r$ in a two-dimensional sublattice, with $r\\approx 1/\\log_2(1/q)$; 'good' blocks dominate a one-dependent site percolation process, and an infinite cluster of good blocks forces an infinite blue cluster.","core_discovery":"At the paper's core are Theorem 2.1 and Theorem 2.4. Place each site of $\\mathbb{Z}^d$, $d\\ge 2$, in the occupied state with probability $p$, and connect each occupied site to the next occupied site in each coordinate direction. In the one-choice model each occupied site makes exactly one of its $2d$ incident segments blue; in the independent model each feasible segment is blue with probability $\\lambda$. The central discovery is that the one-choice model has a phase transition in $p$: a unique infinite blue cluster exists almost surely for small $p$, and no infinite blue cluster exists almost surely for $p$ close to $1$. For the independent model, the paper proves a nontrivial $(p,\\lambda)$ phase diagram: no infinite blue cluster when $\\lambda < p/(2d-1)$, a unique infinite blue cluster when $\\lambda > c\\log(1/q)$ with $q=1-p$, and a unique infinite blue cluster above a Lipschitz curve $\\lambda_c(p)$ that interpolates between the site and bond percolation thresholds. In both models the marginal probability that a given edge is blue is constant in $p$, so the phase transition is not a density effect.","pith_inferences":["A natural next step, not taken in the paper, would be to test numerically whether the conjectured monotonicity of the origin's infinite-cluster probability in $p$ holds; if it does, the transition is a sharp order-disorder transition rather than merely two disconnected existence ranges.","The same inversion mechanism, long straight segments at low occupancy versus short aligned pieces at high occupancy, should appear in continuum line-segment models built from Poisson points, so the two lattice models can serve as a tractable proxy for those systems.","The two-type branching-process comparison suggests a sharper sufficient condition for non-percolation than $\\lambda < p/(2d-1)$: the full extinction criterion for a two-type branching process gives a curve in $(p,\\lambda)$ that could be checked against simulations.","Because the marginal edge density is constant in $p$, these models give a clean way to quantify alignment as an order parameter; one could ask whether a suitably defined collinearity index jumps at the percolation threshold."],"forward_implications":["For the one-choice model on $\\mathbb{Z}^d$, $d\\ge 2$, there is a provable interval of low occupancy with a unique infinite blue cluster and a provable interval of high occupancy with none; the conjectured critical point for $d=2$ lies near $0.505$, and for $d=3$ near $0.862$.","For the independent model, the phase diagram contains at least three proven regions: no infinite cluster for $\\lambda < p/(2d-1)$, and a unique infinite cluster for $\\lambda > c\\log(1/q)$ or above the Lipschitz frontier $\\lambda_c(p)$.","When $p=1$, the independent model reduces to bond percolation, so the critical curve satisfies $\\lambda_c(1)=p_{\\mathrm{bond}}^c$ and $\\lambda_c(p_{\\mathrm{site}}^c)=1$; if $\\lambda_c$ is continuous near $p=1$, the critical value approaches the bond percolation threshold as $p\\to 1$.","The independent model on a $d$-dimensional lattice, restricted to a two-dimensional sublattice, is exactly a two-dimensional independent model; hence percolation proved in two dimensions lifts to higher dimensions.","In both models, the probability an individual edge is blue is independent of $p$, so the existence of a phase transition shows that large-scale connectivity is governed by correlations and segment alignment, not by edge density."],"supporting_citations":[{"why":"Supplies the corrupted compass model and the eigenvalue criterion proving that the turquoise-edge process does not percolate for large p; this yields the upper bound for the one-choice model.","marker":"[16]"},{"why":"Supplies the classical lower-bound method for bond percolation on Z^d that the two-type branching process argument elaborates.","marker":"[13]"},{"why":"Provides the uniqueness machinery for infinite clusters that is adapted to prove almost-sure uniqueness in both models.","marker":"[4]"},{"why":"Establishes the Lipschitz critical curve for mixed percolation that Theorem 2.4(iii) uses to locate a percolation region.","marker":"[6]"},{"why":"Gives the domination theorem for one-dependent site percolation used to produce an infinite cluster of good blocks.","marker":"[17]"},{"why":"Gives the extinction criterion for multi-type branching processes invoked in the non-percolation bound.","marker":"[2]"},{"why":"Supplies the block-argument methodology for turning finite-box percolation events into an infinite cluster.","marker":"[3]"},{"why":"Supplies the supercritical-phase framework that the block comparison uses.","marker":"[11]"}],"fun_headline_variants":["Sparse sites percolate; dense sites don't","Alignment, not density, drives percolation transition","Infinite blue cluster only when sites are sparse","Both models: phase transition in p, not in edge count"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The non-percolation bound for the independent model rests on the paper's assertion, not fully written out, that the blue cluster is no larger in distribution than a branching process with the stated average offspring counts; if that comparison is wrong, the bound $\\lambda < p/(2d-1)$ does not follow from the proof.","fun_headline_variants_meta":{"raw":{"variants":["Sparse sites percolate; dense sites don't","Alignment, not density, drives percolation transition","Infinite blue cluster only when sites are sparse","Both models: phase transition in p, not in edge count"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000603,"raw_usage":{"total_tokens":2800,"prompt_tokens":918,"completion_tokens":1882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1817}},"tokens_in":534,"tokens_out":1882,"duration_ms":16783,"temperature":1.0,"reasoning_tokens":1817,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:23:44.911510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the independent model on $\\mathbb{Z}^2$, take $p=0.6$ and $\\lambda=0.19$, just below $p/(2d-1)=0.2$; if the largest blue cluster in boxes of side $L$ grows linearly with $L$ as $L\\to\\infty$, the no-percolation claim of Theorem 2.4(i), and the branching-process domination behind it, would be wrong.","supporting_citations":[{"cited_title":"Hirsch, M","cited_arxiv_id":null,"evidence_quote":"Supplies the corrupted compass model and the eigenvalue criterion proving that the turquoise-edge process does not percolate for large p; this yields the upper bound for the one-choice model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical lower-bound method for bond percolation on Z^d that the two-type branching process argument elaborates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the uniqueness machinery for infinite clusters that is adapted to prove almost-sure uniqueness in both models."},{"cited_title":"Chayes and R","cited_arxiv_id":null,"evidence_quote":"Establishes the Lipschitz critical curve for mixed percolation that Theorem 2.4(iii) uses to locate a percolation region."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the domination theorem for one-dependent site percolation used to produce an infinite cluster of good blocks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the extinction criterion for multi-type branching processes invoked in the non-percolation bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the block-argument methodology for turning finite-box percolation events into an infinite cluster."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the supercritical-phase framework that the block comparison uses."}],"review_version":1}