{"id":"fb5eea48-3a90-484b-97f7-01c5bb4f032e","arxiv_id":"1908.10213","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Infinite loops in random loop models on bounded-degree graphs occur strictly later than infinite clusters in the associated Bernoulli percolation.","lead":"This paper proves that on infinite graphs with bounded degree, infinite loops in the random interchange model appear at a strictly larger parameter than infinite clusters in the coupled Bernoulli percolation. The result gives a parameter interval where percolation has infinite clusters while all loops remain finite.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3.17) in Lemma 3.11 is false: removing a type-1 neighbour interval can increase the probability that two crosses are far apart, so the proof of the uniform red-edge domination (Proposition 3.1), and hence Theorem 2.1, is not valid as written.","rationale":"The reader correctly identified Proposition 3.1 as the load-bearing assumption, but did not isolate the specific false step. My stress-test found a concrete, checkable error inside the proof of Proposition 3.1: inequality (3.17) in Lemma 3.11 is not true. This is not a matter of disagreement with the consensus; it is an internal inconsistency in a central proof step. The counterexample is elementary and does not depend on exotic graph structure. Because the uniform domination of the red-edge process is exactly what converts a barely supercritical percolation into a subcritical blue-edge process, the failure of Lemma 3.11's proof means Theorem 2.1 is unsupported as written. I am not claiming the theorem is false; the gap may be repairable with a more careful lower bound. Nevertheless, the current manuscript should not be accepted in its present form, and the verdict should move from CONDITIONAL to REJECT (or at least major revision requiring a corrected proof of Proposition 3.1).","tokens_in":15556,"tokens_out":37526,"duration_ms":384286,"concrete_test":"Fix β=10, c=1.25 (corresponding to Δ=3 in Lemma 3.11). Let S=[0,4.9]∪[5.1,10], modeling a type-1 neighbour interval [4.9,5.1] removed from the support of X_e. Compute the probability that the range of two independent uniform points on S is <c and compare with the same probability on [0,10]: the ratio is ≈0.2225/0.2344<1, directly contradicting inequality (3.17). A successful repair would replace (3.17) with a component-counting lower bound of order 1/(2(Δ−1)+1) and re-verify Lemma 3.11.","verdict_should_be":"REJECT","load_bearing_attack":"The proof of Theorem 2.1 rests entirely on Proposition 3.1, which asserts that, conditioned on every edge of a subgraph having at least one link, the red-edge process stochastically dominates a product Bernoulli measure with a uniform δ>0. A key step in Proposition 3.1 is Lemma 3.11, whose proof uses inequality (3.17): for a red neighbour e of e0, conditioning on type-1 neighbours is claimed to only make b_e−a_e smaller in probability, so that P(b_e−a_e<c | O_e) ≥ P(b_e−a_e<c | n_e=2). This monotonicity assertion is false. If a type-1 neighbour removes a small interval from the middle of [0,β], the allowed support for X_e becomes two components; two uniform points then have a strictly smaller chance of being within a fixed distance c than on the full interval, because points falling in different components are separated by the gap. For example, with β=10 and c=1.25, two independent uniform points on [0,10] have range <1.25 with probability 1−(1−1.25/10)^2=0.2344. If the support is S=[0,4.9]∪[5.1,10], the same probability is 0.5·(1−(1−1.25/4.9)^2)≈0.2225, which is strictly smaller. Thus inequality (3.17) fails. Since (3.17) is the only justification for the uniform lower bound in Lemma 3.11, the proof of Proposition 3.1 has a genuine gap. The theorem may still be true, but the argument as written does not establish it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies random loop models on connected, countably infinite graphs of uniformly bounded degree, parametrised by a time parameter beta and a cross/double-bar intensity u. The main result, Theorem 2.1, asserts that for every u in (0,1] the critical loop parameter beta_c(u) is strictly larger than the critical percolation parameter beta_c^per defined by p=1-e^{-beta}; equivalently, there is an interval of beta where infinite percolation clusters occur almost surely while infinite loops do not. The proof constructs a coupled three-colour percolation process: red edges carry exactly two crosses and no neighbouring links between them, blue edges carry at least one link but are not red, and loops are contained in blue clusters. The central technical step, Proposition 3.1, claims that the red-edge process conditioned on all edges being coloured dominates a non-degenerate product Bernoulli measure with parameter delta depending only on beta and the degree bound. From this domination the authors derive that the blue process is subcritical for beta slightly above beta_c^per, yielding the theorem. The paper also contains a corollary for Z^d, an extension to expander graphs, and a discussion of why the u=0 case is not covered.","tokens_in":15914,"tokens_out":6939,"duration_ms":74033,"significance":"If the main theorem is correct, it establishes a strict separation between loop and Bernoulli percolation on all bounded-degree graphs, extending a result previously known for regular trees and providing the first general confirmation that the inequality beta_c >= beta_c^per can be strict outside mean-field settings. The proof idea is attractive and genuinely different from earlier tree arguments: it uses a stochastic domination criterion of Liggett-Schonmann-Stacey to show that the 'cancelling' red edges are sufficiently abundant and sufficiently independent. The claimed uniform dependence of delta only on beta and Delta is exactly the right kind of statement, and Corollary 2.2 is a concrete, falsifiable consequence for Z^d. The paper is self-contained in its derivation and does not disguise fitted parameters as predictions. However, as detailed below, a key inequality in the proof of the uniform domination is false, so the central argument is not yet valid as written.","major_comments":[{"comment":"Inequality (3.17) is false. The text asserts that conditioning on type-1 neighbours of ~e only makes b_~e - a_~e smaller in probability, because X_~e becomes a Poisson point process on a subset of [0, beta] with the same intensity. But deleting an interval from the support can split the support into two components and thereby reduce the probability that two independent uniform points are within a fixed distance c. For beta=10 and c=1.25, two independent uniform points on [0,10] have range less than c with probability 1 - (1 - 1.25/10)^2 = 0.2344; on S = [0,4.9] ∪ [5.1,10] the same probability is 0.5(1 - (1 - 1.25/4.9)^2) ≈ 0.2225, which is strictly smaller. Thus conditioning on a type-1 neighbour can decrease, not increase, the left-hand side of (3.17). This inequality is the only justification for the uniform lower bound in Lemma 3.11, and Lemma 3.11 is used in the chain (3.18) to produce the positive constant in (3.19). The proof of Proposition 3.1, and therefore of Theorem 2.1, is not valid as written.","section":"§3.4, Lemma 3.11, Eq. (3.17)"},{"comment":"The claim that it suffices to prove Proposition 3.1 for u=1 because u in (0,1) 'merely decreases the intensity of red edges by a factor u^2 > 0' is not established. Under the measure mu conditioned on n_e > 0, the probability of the red event depends on u through the Poisson intensity u, through the conditioning on at least one link, and through the fact that double bars on neighbouring edges also count in the condition N_~e(a_e,b_e] = 0. A simple factor u^2 does not obviously absorb all of these effects. Since Theorem 2.1 is stated for all u in (0,1], this reduction is load-bearing and requires a proof rather than a heuristic sentence.","section":"§3.2, reduction to u=1"},{"comment":"The conditional measure mu = P_{beta,u}( . | n_e > 0 for all e in E') is not defined in the usual sense: for an infinite edge set E', the event {n_e > 0 for all e in E'} has probability zero under the product law, so conditioning on it requires a limiting or regular-conditional construction. Remark 3.2 acknowledges the issue and refers to discretisation of time and taking limits, but no such construction is actually given. Since Proposition 3.1 is used for infinite subgraphs in the proof of Theorem 2.1, this gap should be closed explicitly, for example by proving the domination first for finite subgraphs and then passing to the limit.","section":"§3.2, definition of mu; Proposition 3.1"}],"minor_comments":[{"comment":"The sentence 'cycles must be subsets of percolation clusters' uses 'cycles' where 'loops' is meant; the subsequent notation L(v) ⊆ C^B(v) is clear, but the terminology should be corrected.","section":"§3.2"},{"comment":"The promise that conditioning on probability-zero events can be avoided by discretising time and taking limits is not followed by an actual limiting argument in the proof; this is related to major comment 3 and should be addressed.","section":"Remark 3.2"},{"comment":"The proof by contradiction would be easier to verify if the dependence of epsilon on the outer configuration were made explicit; currently it is not fully clear why the two smallness conditions on the two terms in (3.9) cannot hold simultaneously for all epsilon > 0.","section":"Lemma 3.9"},{"comment":"The notation 'η 3.9 ≤' in the displayed inequality is unclear; the intended use of Lemma 3.9 should be written as an explicit factor eta on the left-hand side.","section":"Eq. (3.13)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is genuine and lands directly on Eq. (3.17). The main theorem may still be true, but the current proof has a substantial gap in the central stochastic domination argument. I recommend major revision rather than rejection because the overall strategy is promising and the gap is localised, but the authors must either repair Lemma 3.11 or provide an alternative uniform lower bound for Proposition 3.1, and they must justify the u-reduction and the infinite conditioning more carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline is that this paper likely proves the right result — β_c(u) > β_c^per for all bounded-degree graphs with p_c<1 — and introduces a genuinely new technique. The theorem is a real advance: previous strict inequalities were known on trees and specific high-dimensional graphs; here the argument works for arbitrary bounded-degree graphs, including Z^d. The red-edge enhancement idea, where edges with exactly two crosses and empty neighbourhoods are treated as cancellations, is the right tool.\n\nWhat the paper does well: the reduction to a stochastic domination statement for red edges is clean, and the local proof via pivotal edges and a spatial Markov property is clever. The discussion of the u=0 case and the conjecture for diverging degree graphs is honest and puts the result in context.\n\nThe soft spots: the proof of Proposition 3.1 is compressed. Lemma 3.11, the key local bound, is not written carefully. The claim that 'type 0 neighbours are not admissible' is false: a neighbour of ~e can be pivotal for R~e and have R~e=1, with a configuration where the two points of ~e straddle the neighbour's interval. The proof also says type 1 neighbours only make b~e−a~e smaller because X~e becomes a Poisson process on a subset; this ignores that R~e=1 additionally conditions the two points to lie in the same component of that subset. Once that conditioning is included, the inequality (3.17) is true — in fact it follows from the simple fact that for any interval of length L≤β, the probability that two uniform points are within c is at least the value for the full interval. The stress-test's counterexample to (3.17) misses this same-component conditioning and is incorrect. So the gap is in the exposition, not the mathematics. The reduction u∈(0,1]→u=1 is also sketched too briefly; a coupling argument shows the red-edge probability scales by u^2, but that deserves to be written out.\n\nWho this is for: anyone working on random loop models, interchange processes, or quantum spin representations. The result settles a natural conjecture and sets up a research direction.\n\nRecommendation: send it to review. The theorem is important and likely correct, and the proof strategy is sound. A referee should ask for a rewrite of Lemma 3.11 and the u-reduction, but this is not desk-reject material.","headline":"Likely-correct strict inequality for loop vs percolation thresholds on bounded-degree graphs, with a proof that is intricate and needs referee work, but the stress-test's counterexample does not land.","tokens_in":16428,"tokens_out":34639,"would_cite":true,"duration_ms":306093,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82B43","60G55","82B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Infinite loops arise strictly later than infinite percolation clusters","keywords":["random loop models","random interchange process","bond percolation","critical parameter","stochastic domination","bounded degree","phase transition"],"falsifier":"Fix $u=1$ on $\\mathbb{Z}^2$ and simulate the loop model at $\\beta=-\\ln(1-p_c(\\mathbb{Z}^2))$: if infinite loops occur with positive probability at that $\\beta$, Theorem 2.1 is false. Alternatively, construct a sequence of admissible boundary conditions on the two-neighbourhood of an edge $e_0$ for which the conditional probability that $e_0$ is red tends to zero; this would contradict Proposition 3.1 and break the proof.","tokens_in":15354,"feed_emoji":"🔁","tokens_out":9596,"duration_ms":88466,"temperature":0.7,"pith_summary":"The paper proves that on every connected, countably infinite graph with uniformly bounded vertex degree, a natural random loop model (including the random interchange process) develops infinite loops only at a strictly larger parameter $\\beta_c(u)$ than the parameter $\\beta_c^{\\mathrm{per}}$ at which the coupled Bernoulli bond-percolation model develops infinite clusters. Equivalently, there is an interval of $\\beta$ values in which infinite percolation clusters exist with positive probability while every loop visits only finitely many vertices almost surely. The result holds for all loop-type parameters $u \\in (0,1]$ and gives, on $\\mathbb{Z}^d$ for $d\\ge2$, exponential decay of loop sizes throughout the gap interval. It settles, for bounded-degree graphs, the question of whether the natural lower bound $\\beta_c \\ge \\beta_c^{\\mathrm{per}}$ can be strict, in contrast with graphs of diverging degree where equality is conjectured and known in several mean-field cases.","feed_headline":"Infinite loops arise later than infinite percolation clusters","feed_subtitle":"On integer lattices and every bounded-degree graph, a beta gap separates infinite clusters from infinite loops.","key_machinery":"The carrying object is the red-edge process $(R_e)$, a local configuration event on each edge. Proposition 3.1 shows that, conditioned on every edge of any subgraph carrying at least one link, the law of $(R_e)_{e\\in E'}$ stochastically dominates a product Bernoulli measure with parameter $\\delta>0$ depending only on $\\beta$ and the degree bound $\\Delta$, uniformly over all admissible boundary conditions. The proof of Proposition 3.1 uses a spatial Markov property to reduce arbitrary conditioning to a two-layer neighbourhood of the edge, a classification of pivotal neighbouring edges, and a final reduction to the law of a Poisson point process on a set of intervals of length at least $\\beta/2$; this yields a strictly positive uniform lower bound on the conditional probability that a given edge is red.","core_discovery":"Theorem 2.1: for all countably infinite connected graphs $G$ of uniformly bounded degree with $p_c(G,\\mathrm{bond})<1$ and all $u\\in(0,1]$, the strict inequality $\\beta_c(u)>\\beta_c^{\\mathrm{per}}$ holds, where $\\beta_c^{\\mathrm{per}}=-\\ln(1-p_c)$. The proof couples the loop configuration to bond percolation with $p=1-e^{-\\beta}$ and colours each edge red, blue, or uncoloured: an edge is red when it carries exactly two crosses and no neighbouring edge has a link between them; blue when it carries at least one link but is not red. Loops can travel only along blue edges, so if red edges are abundant enough and form an independent-like set, deleting them from a barely supercritical percolation cuts every infinite cluster while the original percolation still percolates. This yields a gap interval on every bounded-degree graph, and in particular on $\\mathbb{Z}^d$ an interval with infinite percolation clusters but exponentially small loops.","pith_inferences":["The uniform domination produced by Proposition 3.1 is quantitative, so the same construction could in principle yield explicit lower bounds on the size of the gap $\\beta_c(u)-\\beta_c^{\\mathrm{per}}$ on lattices, although the paper only proves positivity.","The red-edge event is an 'essential enhancement' with a built-in independence mechanism; this suggests that similar local cancellation events could be designed for other loop-weighted models once a comparison model is identified, a direction the paper mentions but does not develop beyond $\\theta>1$ in an 'appropriate sense'.","The failure of the argument at $u=0$ points to a qualitative difference between two-cross cancellation (where two crosses on the same edge undo a connection) and double-bar dynamics, where no such cancellation exists; testing numerically whether the gap closes as $u\\to0$ on $\\mathbb{Z}^2$ would isolate which feature of the model is responsible for the strict inequality."],"forward_implications":["On $\\mathbb{Z}^d$, $d\\ge2$, there exist $0<\\beta_1<\\beta_2<\\infty$ such that for every $\\beta\\in(\\beta_1,\\beta_2)$ the loop sizes satisfy $P_{\\beta,u}(|L(0)|=k)\\le ae^{-bk}$ while the coupled bond percolation with $p=1-e^{-\\beta}$ has an infinite cluster with positive probability (Corollary 2.2).","For sequences of $d$-regular expander graphs with diverging girth, macroscopic loops appear strictly later than macroscopic percolation clusters (Corollary 4.1).","The strict inequality holds for all $u\\in(0,1]$, while the case $u=0$ remains open and appears to require new tools for dependent percolation (Section 4.4).","For graphs of diverging vertex degree, including complete graphs, hypercubes, and Hamming graphs, known results give equality $\\beta_c=\\beta_c^{\\mathrm{per}}$, so the strict gap is specific to bounded degree (Conjecture 4.2 and the discussion around it)."],"supporting_citations":[{"why":"Supplies the domination-by-product-measures criterion: a uniform lower bound on conditional probabilities implies stochastic domination by a product Bernoulli measure.","marker":"[LSS97, Lemma 1.1]"},{"why":"Shows that deleting a positive fraction of open edges from a supercritical percolation need not destroy all infinite clusters, motivating the need for red edges to be close to independent.","marker":"[AG91]"},{"why":"Provides corrections to the essential-enhancements argument, referenced when the paper explains why independence among deleted edges is not automatic.","marker":"[BBR14]"},{"why":"Proves the sharp phase transition for the random stirring model on trees, the prior result that Theorem 2.1 extends to general bounded-degree graphs.","marker":"[Ham15]"},{"why":"Defines the loop critical parameter and proves its well-definedness on connected graphs; also supplies the tree case for $u=1$.","marker":"[Ang03]"},{"why":"Extends the critical-parameter analysis on trees to $u\\in[0,1]$, providing the benchmark that the strict gap refines.","marker":"[BU18a]"}],"fun_headline_variants":["Infinite loops need strictly more randomness than bond percolation","Beta gap on bounded-degree graphs: percolation beats loops","Without infinite loops, percolation clusters still form on Z^d","Loop percolation threshold strictly above bond percolation threshold","Strict gap: loop percolation occurs after percolation on bounded-degree graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on the claim that conditioned on every edge of a subgraph carrying at least one link, the red-edge indicators stochastically dominate a product Bernoulli measure with a strictly positive parameter $\\delta$ that is uniform over all boundary conditions; if that uniform domination failed, deleting red edges from a barely supercritical percolation could leave an infinite cluster of blue edges.","fun_headline_variants_meta":{"raw":{"variants":["Infinite loops need strictly more randomness than bond percolation","Beta gap on bounded-degree graphs: percolation beats loops","Without infinite loops, percolation clusters still form on Z^d","Loop percolation threshold strictly above bond percolation threshold","Strict gap: loop percolation occurs after percolation on bounded-degree graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3285,"prompt_tokens":931,"completion_tokens":2354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":2261}},"tokens_in":547,"tokens_out":2354,"duration_ms":17108,"temperature":1.0,"reasoning_tokens":2261,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:51:33.329805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $u=1$ on $\\mathbb{Z}^2$ and simulate the loop model at $\\beta=-\\ln(1-p_c(\\mathbb{Z}^2))$: if infinite loops occur with positive probability at that $\\beta$, Theorem 2.1 is false. Alternatively, construct a sequence of admissible boundary conditions on the two-neighbourhood of an edge $e_0$ for which the conditional probability that $e_0$ is red tends to zero; this would contradict Proposition 3.1 and break the proof.","supporting_citations":[],"review_version":1}