{"id":"266bcc10-b75d-40f4-b7a8-fa4ec2cd4d2f","arxiv_id":"2506.01612","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For random 2-lifts of transitive graphs, the percolation critical point is continuous in the switching probability q, strictly smaller than the base threshold, and subcritical clusters decay exponentially at q=1/2.","lead":"This paper studies Bernoulli percolation on a random two-sheeted cover of a graph, where each pair of parallel edges is switched with probability q. It proves that the percolation threshold varies continuously with q, is strictly below the threshold of the base graph, and that cluster sizes decay exponentially when q=1/2.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The coupling proof of Proposition 4.1 does not justify that the enhancement variables α_u are i.i.d. Bernoulli(s); without the missing thinning argument, Theorem 1.2's comparison to pc(G,s) does not follow.","rationale":"We agree with the reader's CONDITIONAL verdict but not with the stated weakest assumption. The annealed-to-quenched transfer is automatic: Ψ(p,q) = E[θ_η(p)] ∈ {0,1}; if Ψ=1 then θ_η(p)=1 a.s., so p_c(G_q(η)) ≤ p < p_c(G) a.s. No ergodicity is needed. The actual load-bearing gap is in the construction of the enhanced percolation coupling. The α_u variables are the mechanism by which the comparison to pc(G,s) is made; if they are not i.i.d. Bernoulli(s), then C∞ is not the augmented-percolation cluster and Proposition 4.1—the only bridge from Proposition 4.2 to percolation on G_q—does not hold. The paper's assertion that the α_u are Bernoulli(s) is unsupported; the success probability is history- and p-dependent. The gap is likely fixable via the thinning argument from [MS19], since the number of edges involved is bounded by M and p is bounded below by ε, so this does not force a REJECT. We therefore keep the reader's CONDITIONAL recommendation; the preprint should be revised to supply the missing thinning construction. We also note the secondary presentation error in Proposition 4.1's implication direction (stated opposite to what is constructed and used), which should be corrected in revision.","tokens_in":22429,"tokens_out":30120,"duration_ms":304506,"concrete_test":"Repair attempt: following [MS19], at each even step of the exploration in §4.2, condition on the history and compute ρ = P(all newly s-explored edges are open and β_x=1 | history). Choose s ≤ t·ε^M (with M as in the paper) and define α_u = 1 only if an independent Bernoulli variable of parameter s/ρ is 1 and the substeps succeed; also draw independent α's for vertices never reached. Then check that the resulting (α_u) are i.i.d. Bernoulli(s) and independent of the unrevealed ω's. If this succeeds, Proposition 4.1 is valid as a repair; if ρ can be 0 or if the required thinning breaks the coupling's open-path condition, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Proposition 4.1 (§4.2, Step 2N+2, substep 6), the paper defines α_u to be 1 exactly when all the newly s-explored edges in Z(x,r) and S_{r+1/2}(u) are open and β_x=1, and then asserts that 'therefore α_u is a Bernoulli variable with parameter s and the α_u variables are independent'. This is not justified: the number N of such edges is random (it depends on the exploration history and on η), so the success probability is t·p^N, which depends on p and on the history. The marginal law of α_u is consequently a mixture, not Bernoulli(s), and independence across overlapping Z-regions is not established. This matters because C∞ is only distributed as the augmented percolation cluster if α ~ B(s)^{⊗V(G)}; without that, the conclusion p_c(G_q) ≤ p_c(G,s) in the proof of Theorem 1.2 collapses. The gap is probably repairable by the thinning argument of [MS19] (since N ≤ M and p ≥ ε, the success probability is bounded below by t·ε^M), but the preprint does not supply it. Note that the 'a.s.' transfer issue raised in the reader's weakest assumption is not the real problem: by the 0-1 law, Ψ(p,q)=1 implies E[θ_η(p)]=1, hence θ_η(p)=1 a.s., no ergodicity needed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Bernoulli percolation on random 2-lifts G_q of a transitive graph G, obtained by switching each pair of parallel edges with probability q. It proves three results: the annealed critical parameter p_c(q) is continuous on (0,1) (Theorem 1.1); for connected non-tree transitive G with p_c(G) < 1, p_c(G_q) < p_c(G) almost surely (Theorem 1.2); and the cluster-size tail decays exponentially for p < p_c(1/2) (Theorem 1.4). The continuity proof is a coupling showing 1/2-Hölder continuity, the monotonicity proof adapts the augmented-percolation method of Martineau and Severo, and the exponential decay proof follows Vaneuville's ghost-field strategy with a new exploration framework for random graphs.","tokens_in":22581,"tokens_out":29619,"duration_ms":283559,"significance":"The continuity result is a clean and likely correct application of coupling and gives a new locality-type statement for random graph limits. The exponential decay section contains a novel exploration of random graph structure and an elegant use of the Haar-measure property of Bernoulli(1/2) switching variables to prove the key remaining-graph comparison; this is a valuable contribution in its own right. The strict monotonicity theorem, if established, would be a substantial extension of the Martineau–Severo result to covers without bounded fibers. However, the proof of Theorem 1.2 contains a gap in the construction of the enhanced percolation variables: the α_u variables are not shown to be i.i.d. Bernoulli(s), and their independence from the exploration is not proved. This gap is likely repairable by a thinning argument, but it is load-bearing for the main theorem as written.","major_comments":[{"comment":"The proof asserts in substep 6 that α_u is Bernoulli with parameter s and that the α_u's are independent, and the final paragraph of the proof uses this to conclude that C∞ has the law of augmented percolation with parameters (p,s). This does not follow from the construction: the success of substeps 3 and 4 requires all of a random number N of s-explored edges to be open, so the success probability is t·p^N (with N depending on p, η, and the exploration history). The marginal law of α_u is therefore a mixture rather than Bernoulli(s), and no independence across u is established. Without α ~ B(s)^{⊗V(G)} independent of ω, the comparison p_c(G_q) ≤ p_c(G,s) in the proof of Theorem 1.2 collapses. A thinning argument (for instance using the uniform bound M from property (D) and the lower bound p ≥ ε to choose s = t ε^M) is needed, but the manuscript does not supply it.","section":"4.2, Step 2N+2, substeps 4–6"},{"comment":"The variable β_x imported from Lemma 4.7 is a deterministic function of the switching configuration η (via the variables O_i), and the exploration process reveals information about η. The proof does not show that β_x is independent of the σ-field generated by the earlier exploration, nor that the events {β_x=1} for different vertices are conditionally independent given the history. These properties are required for the claimed joint law of the α_u variables and for the identification of C∞ as an augmented-percolation cluster.","section":"4.2, substep 5"},{"comment":"The theorem is a quenched statement (the inequality holds a.s. for the random graph G_q), but the proof establishes the annealed inequality p_c(q) ≤ p_c(G,s) < p_c(G). The paper never explains why the quenched critical parameter p_c(G_q(η)) is almost surely equal to the annealed p_c(q). The equality follows from Ψ(p,q)=E[θ_η(p)] together with the Kolmogorov 0-1 law and monotonicity in p, but this argument should be written out because it is essential to the a.s. formulation.","section":"Theorem 1.2 and Section 4"}],"minor_comments":[{"comment":"In the proof of Theorem 1.4, the last display and the concluding sentence give C = 1/(1−2m_h(p′)), but the computation yields C = 1/(1−m_h(p′)); the bound is ψ_n(p) ≤ (1/(1−m_h(p′))) e^{-hn}.","section":"Proof of Theorem 1.4"},{"comment":"In the proof of Lemma 5.10, the sentence 'Let e be a type (3) edge' refers to a type that has not been defined; from the previous paragraph this should read 'type (1)'.","section":"Proof of Lemma 5.10"},{"comment":"In Lemma 5.7 and its proof, the event 'A ∩ Expl_k' is written without specifying that A is a subset of {0,1}^E × F × {0,1}^V while Expl_k is a subset of {0,1}^E × F; the text should clarify that the projection of A onto the first two coordinates is taken.","section":"Lemma 5.7"},{"comment":"The proof of Proposition 2.4 invokes a 'factorization theorem' without naming it; this is an orbit–stabilizer count and should be stated as such.","section":"Proof of Proposition 2.4"},{"comment":"The first sentence of the introduction misspells 'Hammersley' as 'Hammersely'.","section":"Introduction"},{"comment":"In Remark 1.3, the phrase 'G_q consists essentially of two copies of G in parallel' is true for trees but deserves a proof or a more precise statement: every 2-lift of a tree is isomorphic to two disjoint copies of G.","section":"Remark 1.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its limitations and credits related work appropriately. The continuity and exponential decay sections are likely correct and well within the journal's scope. The strict monotonicity section has a repairable but significant gap; I recommend a major revision rather than rejection, as the missing thinning argument and independence proof seem feasible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a genuinely interesting paper with a real gap in its central strict-monotonicity proof. The continuity theorem is clean, the exponential-decay section is ambitious, and the model is natural. But the proof of Proposition 4.1 does not actually show that the enhancement variables α_u are i.i.d. Bernoulli(s), and Theorem 1.2 leans on that. I'd send it to a serious referee, but I would not accept it as-is.\n\nThe three results are new. Continuity of p_c(q) via the coupling in Lemma 3.2 is a nice locality statement for a random graph family; the argument is self-contained and looks correct. Exponential decay at q=1/2 follows Vaneuville's ghost-field method, with a serious extra twist (Lemma 5.10) about the remaining graph; the coupling proof is sketched but the idea is plausible, and the q=1/2 restriction is honest. The strict monotonicity is the weak point. The proof imports [MS19]'s enhanced percolation but must replace deterministic bounded fibers with probabilistic control, which is a genuinely new difficulty.\n\nThe specific problem is in Step 2N+2, substep 6: the paper asserts that α_u is Bernoulli(s) with fixed s and that the α_u are independent. The definition sets α_u=1 exactly when a random number N of newly s-explored edges are open and β_x=1. N depends on the exploration history and on η, so the success probability is t·p^N, a mixture, not a fixed Bernoulli(s). Independence across overlapping Z-regions is also not established. Without α ~ B(s)^{⊗V(G)}, C∞ is not distributed as the augmented percolation cluster, and the inequality p_c(G_q) ≤ p_c(G,s) collapses. The reader's note about the 'a.s.' transfer is not the real issue—the 0-1 law plus the annealed inequality gives the quenched statement directly, no ergodicity needed. The real problem is the α construction.\n\nThat said, the gap is probably repairable. In [MS19] the analogous step uses a thinning argument: if the success probability is bounded below and the variables are locally dependent, you can extract a genuine Bernoulli(s) family stochastically dominated by the original. The paper needs to spell that out; it doesn't. Given that the rest of the strategy is sound and the missing argument is standard but nontrivial, I'd call this conditional rather than a reject. The paper is honest about what it does and doesn't prove—Question 5.14 is explicitly left open.\n\nWho is this for? People working on percolation on graphs, especially locality and covering maps. It deserves a serious referee, and I'd be happy to see a revised version with the thinning argument filled in. For my own work I wouldn't cite Theorem 1.2 yet, though I'd consider citing the continuity result if I needed it.","headline":"Genuinely new results and a natural model, but the strict-monotonicity proof has a load-bearing gap in the construction of the enhancement variables; likely fixable, so conditional rather than reject.","tokens_in":23243,"tokens_out":2372,"would_cite":false,"duration_ms":23723,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","05C80","82B43"],"pacs":[],"model":"deepseek-v4-flash","headline":"Random double covers of non-tree transitive graphs strictly lower the percolation threshold.","keywords":["random 2-lift","Bernoulli percolation","percolation threshold","critical parameter","transitive graph","exponential decay","enhanced percolation","gauge transformation"],"falsifier":"Run two independent high-accuracy simulations of bond percolation on large finite toroidal 2-lifts of the square lattice at the same $q\\in(0,1)$, and compare the two finite-size estimates of the quenched critical threshold; if they do not converge to a common value, the almost-sure constancy that Theorem 1.2 needs is false. A direct refutation of the strict inequality would be a single transitive non-tree graph with $p_c(G)<1$ and one $q\\in(0,1)$ for which the limiting threshold satisfies $p_c(G_q)\\ge p_c(G)$.","tokens_in":22089,"feed_emoji":"🕸️","tokens_out":11521,"duration_ms":108443,"temperature":0.7,"pith_summary":"Percolation on a random 2-lift of a graph $G$ takes two copies of $G$ and crosses each pair of parallel edges with probability $q$. This paper proves that for every connected, transitive, infinite graph $G$ that is not a tree and has $p_c(G)<1$, the random lift $G_q$ has a strictly smaller percolation threshold than $G$, for every $q\\in(0,1)$, almost surely. It also proves that the threshold $p_c(q)$ is continuous in $q$ on $(0,1)$ and that at $q=1/2$ the tail of the cluster of a vertex decays exponentially for every $p<p_c(1/2)$. The point of the model is that $G_q$ is neither transitive nor equipped with bounded fibers, so establishing these phase-transition facts tests whether percolation tools built for transitive graphs survive probabilistic geometry.","feed_headline":"Random double covers lower percolation thresholds","feed_subtitle":"For any nontrivial random double cover of a transitive graph, percolation begins earlier than on the original graph.","key_machinery":"The load-bearing object is the random 2-lift $G_q=(V\\times\\{0,1\\},E(\\eta))$, with independent Bernoulli switching variables $\\eta_e$; a pair of parallel edges is crossed exactly when $\\eta_e=1$. The proofs work by couplings rather than by analyzing one realization's geometry: first, a coupling between percolation on $G_q$ and enhanced percolation on $G$, in which a fully open ball of radius $r$ can, with auxiliary probability $s$, annex its boundary sphere; second, Bernoulli variables $\\beta_x$ that probabilistically certify that the two lifts of a vertex $x$ are connected within $D$ steps, replacing the deterministic bounded-fiber condition used in [MS19]; and third, at $q=1/2$, an exploration that reveals edges without always revealing both endpoints, together with a coupling of the 'remaining graph' after the origin's cluster is deleted. The $q=1/2$ step uses gauge transformations $\\phi_S$ and the fact that the switching law is the Haar measure on $\\mathbb{Z}/2\\mathbb{Z}$, whose convolution with any explored geometry remains uniform; this is why the exponential-decay coupling is proved only at $q=1/2$.","core_discovery":"The central claim is that the random choice of which edges cross between the two floors of a double cover lowers the percolation threshold strictly, in every nontrivial switching regime, and that the usual sharp phase transition persists in this non-transitive random setting. Theorem 1.2 states that if $G$ is connected, transitive, infinite, not a tree, and $p_c(G)<1$, then $p_c(G_q)<p_c(G)$ almost surely for every $q\\in(0,1)$; the hypothesis that $G$ is not a tree is necessary, because a tree has two parallel copies with the same threshold. Theorem 1.1 says $q\\mapsto p_c(q)$ is continuous on $(0,1)$ (in fact locally $1/2$-Hölder continuous), and Theorem 1.4 gives exponential decay of the cluster size at $q=1/2$, with a quenched version following for almost every realization of the random lift.","pith_inferences":["Editorial inference: the probabilistic twin-distance variables $\\beta_x$ are the only place where the proof of Theorem 1.2 uses the special structure of a 2-lift; the same scheme should yield strict monotonicity for random $n$-lifts with i.i.d. permutation switching, provided the cycle-certificate lemmas generalize.","Editorial inference: the Borel–Cantelli argument that turns the annealed exponential bound into a quenched one is a general template: any uniform annealed exponential tail bound would transfer to almost every realization in the same way.","Editorial inference: the open question of exponential decay for $q\\neq 1/2$ could be probed numerically by measuring cluster tails on large finite 2-lifts of the square lattice at $q=1/4$; the paper predicts the decay should hold, while Lemma 5.10 may fail."],"forward_implications":["For every non-tree transitive graph with $p_c(G)<1$, an arbitrarily small switching probability already makes percolation easier: $p_c(G_q)<p_c(G)$ for every $q\\in(0,1)$.","Because $q=0$ gives two disjoint copies of $G$, continuity at $0$ means the threshold drops immediately as soon as switching is turned on.","At $q=1/2$, the subcritical regime is genuinely exponentially decaying: $\\mathbb{P}(|C_o|\\ge n)\\le Ce^{-cn}$ for all $p<p_c(1/2)$, and the same holds for almost every realization of the random lift.","Continuity of $q\\mapsto p_c(q)$ in $(0,1)$ is a locality-type result for a family of random, generally non-transitive graphs: the critical parameter varies continuously when the law of the random graph is deformed."],"supporting_citations":[{"why":"Supplies the enhanced-percolation framework and the strict monotonicity theorem for bounded-fiber covers, which Proposition 4.1 adapts to the random 2-lift.","marker":"[MS19]"},{"why":"Establishes weak monotonicity $p_c(G_q)\\le p_c(G)$ for covers, the starting point of Theorem 1.2, and the general percolation-on-transitive-graphs setup.","marker":"[BS96]"},{"why":"Provides the stochastic-comparison strategy for exponential decay that Section 5 adapts to random graphs with the new exploration and Lemma 5.10.","marker":"[Van25]"},{"why":"Introduces enhanced percolation, the mechanism behind the strict monotonicity coupling.","marker":"[AG91]"},{"why":"Records counterexamples for unimodular random graphs that delineate the scope: the theorems here cannot simply extend to all random graph limits.","marker":"[BPT17]"}],"fun_headline_variants":["Random double covers lower percolation threshold","Percolation starts earlier on random double covers","Random 2-lifts drop percolation threshold","Strictly earlier percolation on random 2-lifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The almost-sure conclusion in Theorem 1.2 rests on the unstated premise that the percolation threshold of the random 2-lift is the same for essentially every realization of the switching randomness, so that the average strict inequality proved in Section 4 can be applied to each typical random graph; if different realizations had different thresholds, strict inequality could hold on average but fail almost surely.","fun_headline_variants_meta":{"raw":{"variants":["Random double covers lower percolation threshold","Percolation starts earlier on random double covers","Random 2-lifts drop percolation threshold","Strictly earlier percolation on random 2-lifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1882,"prompt_tokens":877,"completion_tokens":1005,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":954}},"tokens_in":493,"tokens_out":1005,"duration_ms":9049,"temperature":1.0,"reasoning_tokens":954,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:43:21.725257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run two independent high-accuracy simulations of bond percolation on large finite toroidal 2-lifts of the square lattice at the same $q\\in(0,1)$, and compare the two finite-size estimates of the quenched critical threshold; if they do not converge to a common value, the almost-sure constancy that Theorem 1.2 needs is false. A direct refutation of the strict inequality would be a single transitive non-tree graph with $p_c(G)<1$ and one $q\\in(0,1)$ for which the limiting threshold satisfies $p_c(G_q)\\ge p_c(G)$.","supporting_citations":[],"review_version":1}