{"id":"e808b2d3-6bd2-4226-978b-e3d3fb898f08","arxiv_id":"2412.05706","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For all d≥3 except d=6, the probability that the critical metric-graph GFF level set connects two sets across an annulus equals, up to constants, N^{(6-d)∧0} times the product of the two one-sided connection probabilities.","lead":"This paper proves a sharp up-to-constants factorization for the probability that two distant sets are connected by the critical zero level-set of the Gaussian free field on the metric graph of Z^d. The result settles the GFF analogue of a 2017 percolation conjecture and supplies the two-point input used in a companion paper to construct incipient infinite clusters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the argument is internally coherent; the main fragility is the imported sharp exponents (1.3)–(1.8), especially the d=6 gap, which the paper explicitly acknowledges.","rationale":"The reader's ACCEPT verdict with moderate confidence is appropriate. My stress-test pass did not identify a concrete internal error: the chain of estimates from the regularity propositions to the quasi-multiplicativity bounds follows logically, and the high-dimensional correction factor N^{6-d} is rigorously derived rather than imported from heuristics. The weakest point is indeed the set of imported sharp exponents, especially the d=6 slack, which the paper acknowledges and isolates. This does not undermine the main theorem for d≠6, and the d=6 statement is honest about its precision. I do not see grounds to reject or make acceptance conditional; the paper's claims match what is proved. A worthwhile verification step would be to re-derive the key loop-measure bound (6.21) and the d=6 one-arm exponent, but these are checks of prior work and internal details rather than indications of a flaw. Therefore the verdict remains UNCHANGED.","tokens_in":55318,"tokens_out":22362,"duration_ms":187843,"concrete_test":"Independently re-derive the loop-measure estimate in (6.21) from Lemma 2.9 and the Brownian excursion decomposition, checking that the factor |w1−w2|2−d |w2−w3|2−d P_{w3}(τ_{eB_{w1}(1)} < τ_{∂B(N+1)}) fully accounts for the sub-path constraint; if a missing factor is found, recompute the high-dimensional upper bound (6.25). Also cross-check the d=6 one-arm upper bound N^{-2+ς(N)} used in (1.4) against [6, Theorem 1.1]; if that bound cannot be reproduced, Theorem 1.1(d=6) weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I reviewed the proof architecture and found no internal gap that would invalidate the central claim. The lower bounds rely on harmonic averages, FKG, and Lemma 6.1, and the upper bounds on loop-soup decomposition, BKR, and stochastic domination (Lemmas 4.2, 5.1, 5.2). The regularity propositions (1.5–1.9) are used consistently, and the d≥7 correction factor N^{6-d} emerges from the annulus crossing factors in Lemma 5.1 and the boundary-to-set comparisons in Corollary 1.10; the heuristic volume-count in Remark 1.3 is only motivational and is not needed for the proof. The one genuine fragility is the dependence on the sharp one-arm and crossing exponents (1.3)–(1.8) from [5, 6, 12]. In particular, for d=6 only N^{-2} ≤ θ_6(N) ≤ N^{-2+ς(N)} is known, which is why Theorem 1.1(d=6) has N^{±o(1)} slack. This is openly stated (Remarks 1.3 and 1.7) and is a limitation, not an internal inconsistency. If any imported exponent were wrong, the high-dimensional and critical-dimensional parts of the theorem would collapse, but no evidence of such an error is present in this paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves quasi-multiplicativity for critical level-sets of the Gaussian free field on the metric graph \\widetilde{\\mathbb{Z}}^d, for all d \\ge 3. Theorem 1.1 states that the probability of connecting two sets on opposite sides of an annulus of radius N is comparable to the product of the two one-sided boundary-to-set connection probabilities, with no correction factor for 3 \\le d \\le 5, an N^{6-d} correction factor for d \\ge 7, and N^{\\pm o(1)} slack at d=6. The proof develops a body of regularity results for point-to-set and boundary-to-set connection probabilities (Propositions 1.5\\,--\\,1.9 and Corollary 1.10), using the loop-soup isomorphism, FKG, the strong Markov property, a BKR inequality for glued loops, and a loop decomposition. Theorem 1.11 establishes the lower bound on the critical cluster volume exponent for 3 \\le d \\le 5.","tokens_in":55584,"tokens_out":33070,"duration_ms":283893,"significance":"The result is significant: it provides the first quasi-multiplicativity theorem for a non-Bernoulli, strongly correlated percolation model in this generality, and it gives concrete support to the Basu\\,--\\,Sapozhnikov conjecture on correction factors in high dimensions. The proof architecture is coherent, with lower bounds via harmonic averages and FKG, and upper bounds via loop-soup decomposition, stochastic domination, and BKR. The d \\ge 7 correction factor is derived from the annulus-crossing factors in Lemma 5.1 and the boundary-to-set comparisons in Corollary 1.10, rather than from the heuristic volume count in Remark 1.3. The paper is also transparent about its main limitation: the sharp one-arm and crossing exponents (1.3)\\,--\\,(1.8) are imported from earlier work, and the d=6 statement inherits the unresolved gap in \\theta_6. The d=6 result is therefore conditional and weaker than the low- and high-dimensional statements, but this is explicitly acknowledged in Remarks 1.3 and 1.7.","major_comments":[{"comment":"The displayed inequality for the low-dimensional upper bound contains a factor N^{\\varsigma_d(N)} that does not follow from the cited Corollary 1.10. For d=6, Corollary 1.10 with M_1=N and M_2=CN gives P(A_2 \\leftrightarrow \\partial B(CN)) / P(A_2 \\leftrightarrow \\partial B(N)) \\le C N^{4\\varsigma(N)}, and Lemma 5.2 therefore yields an upper bound of order N^{4\\varsigma(N)} times the product, not N^{\\varsigma(N)}. This is a local gap: the proof as written appears to establish the d=6 part of Theorem 1.1 only with N^{4\\varsigma(N)} on the right-hand side of (1.12). Since N^{4\\varsigma(N)} is still N^{o(1)}, the main qualitative claim is unaffected, but the stated exponent should be corrected or an additional comparison supplied.","section":"Section 6.2, Eq. (6.18)"}],"minor_comments":[{"comment":"The notation should use \\partial B(M) rather than B(M) in the connection probabilities appearing in the ratios. In case (a) of Proposition 1.9, the event A \\leftrightarrow B(M) is trivial because A is already contained in B(N) \\subset B(M), so the displayed ratio cannot have its stated meaning unless the target is \\partial B(M).","section":"Proposition 1.9 and Corollary 1.10"},{"comment":"The proof of the lower bound in Proposition 1.9 produces the factor M^{-\\varsigma_d(M)} in the final inequality, while the statement in (1.19) writes M^{-\\varsigma_d(N)}. Since M \\ge N in the relevant regime this is a harmless weakening, but the intended statement should be clarified.","section":"Section 5.1, Eq. (5.5)"},{"comment":"The application of Lemma 5.3 should specify that item (b) is being used, with A = \\partial B(N) and M = c_9 N, and should note the condition M \\le C_{13}^{-1} N^{1-\\varsigma_d(N)}. As written, the citation is ambiguous because Lemma 5.3 has two very different hypotheses.","section":"Lemma 7.1, Eq. (7.5)"},{"comment":"There are numerous small typos and formatting infelicities, for example 'separartely' in Section 5.3 and the hard-to-read display around (2.34). A careful copyedit would improve readability.","section":"General"}],"recommendation":"minor_revision","confidential_remarks":"The paper is a substantial contribution and I believe the central argument is sound. The only substantive point I would ask the authors to address is the d=6 upper-bound factor in (6.18); if the exponent is corrected to N^{4\\varsigma(N)}, the theorem remains meaningful and the paper can be accepted after a local revision. The reliance on the authors' earlier sharp exponent results [5,6] is openly documented and not a circularity concern, but it does mean the d=6 part should be presented as conditional on (1.4)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a substantial piece of probability theory. Cai and Ding prove quasi-multiplicativity for critical level sets of the metric graph GFF: no correction for 3 ≤ d ≤ 5, a correction factor N^{6-d} for d ≥ 7, and upper/lower bounds within N^{±o(1)} at d = 6. If correct, this is the first quasi-multiplicativity theorem for this model, and it supplies the key input for their companion IIC construction.\n\nWhat is genuinely new: the factorization theorem (Theorem 1.1) and the N^{6-d} correction. Earlier work [5,6] determined one-arm and crossing exponents but did not prove factorization for general sets. The paper also proves point-to-set and boundary-to-set regularity propositions (1.5–1.9) that are useful in their own right, and it sharpens the volume exponent for 3 ≤ d ≤ 5.\n\nThe proof is long but structurally sound. Lower bounds use FKG and harmonic averages; upper bounds use the loop soup, BKR, and a tree expansion. I did not find an internal gap. The authors are explicit about what they import: the sharp one-arm and crossing exponents (1.3)–(1.8) from [5,6,12]. For d = 6, only N^{-2} ≤ θ_6 ≤ N^{-2+ς(N)} is known, which is exactly why the d = 6 statement carries N^{o(1)} slack. That is a stated limitation, not a hidden flaw. If any imported exponent were wrong, the high-dimensional and critical-dimensional conclusions would collapse, but there is no evidence of that here.\n\nThe main soft spot is the d = 6 case: the result is not exact quasi-multiplicativity but matching up to N^{±o(1)}. Applications needing sharp d = 6 bounds would have to live with that. There are also occasional typos in intermediate estimates, but nothing that appears to affect the proofs.\n\nThis paper deserves a serious referee. I would bring it to a reading group and cite it in my own work.","headline":"Proves quasi-multiplicativity for metric graph GFF level sets with the expected N^{6-d} correction in high dimensions, leaving the d=6 case honestly with N^{o(1)} slack.","tokens_in":56115,"tokens_out":2045,"would_cite":true,"duration_ms":19600,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60G60","60J65","82B43"],"pacs":[],"model":"deepseek-v4-flash","headline":"Quasi-multiplicativity holds for critical GFF level-sets: annulus connections factorize as products of boundary connections, with correction N^{6-d} in high dimensions and N^{o(1)} slack at d=6.","keywords":["metric graph Gaussian free field","critical level-set percolation","quasi-multiplicativity","one-arm probability","loop soup","connection probability","volume exponent","incipient infinite cluster"],"falsifier":"Determine the exact order of the one-arm probability $\\theta_6(N)$ at dimension 6. The paper's d=6 bounds have $N^{-10\\varsigma(N)}$ to $N^{\\varsigma(N)}$ slack precisely because only $N^{-2}\\le\\theta_6(N)\\le N^{-2+\\varsigma(N)}$ is known, so a sharper value outside this range would contradict (1.12). Alternatively, in $d\\ge7$ take $A_1=B(R_1)$ and $A_2=\\partial B(R_2)$ with distances violating $R_1\\lesssim N^{2/(d-4)}$ and $R_2\\gtrsim N^{(d-4)/2}$; the claimed $N^{6-d}$ factor fails there, and Remark 1.4 identifies these conditions as sharp.","tokens_in":55116,"feed_emoji":"🔗","tokens_out":14098,"duration_ms":117964,"temperature":0.7,"pith_summary":"At the critical level $h=0$ of the Gaussian free field on the metric graph $\\widetilde{\\mathbb{Z}}^d$, the probability that clusters connect two sets across an annulus of radius $N$ is, up to a dimension-dependent factor, the product of the probabilities that each set reaches the nearer side of the annulus. The paper proves this quasi-multiplicativity in all dimensions $d\\ge 3$: no correction for $3\\le d\\le 5$, a correction $N^{6-d}$ for $d\\ge7$, and two bounds differing by $N^{o(1)}$ at the critical dimension $d=6$. The result matters because this factorization is the input used in a companion paper to construct the incipient infinite cluster for this model, and because it tests the analogy between metric-graph GFF clusters and Bernoulli percolation. As by-products, point-to-set connection probabilities are shown to decay like harmonic functions and obey a Harnack inequality, and the cluster-volume tail is determined for every dimension except 6.","feed_headline":"Annulus connections factorize for critical GFF level-sets","feed_subtitle":"The chance two clusters meet across an N-annulus is the product of two boundary probabilities, times N^{6-d} when d>6.","key_machinery":"The central object is the annulus connection probability at level $0$, together with the comparison identity that replaces it by the product of two boundary-to-set probabilities. Two complementary mechanisms carry the proof. For the lower bound, the machinery is the sign-cluster exploration: after exposing the cluster of each target set up to an intermediate boundary, the harmonic average of the boundary values (a weighted sum of GFF boundary values using Brownian hitting probabilities) quantifies how positive the cluster is, and formula (2.30), giving the conditional probability of connection as $1-e^{-2\\sum K(z_1,z_2)\\phi_{z_1}\\phi_{z_2}}$, turns a lower bound on the two harmonic averages into a lower bound on the full connection probability. For the upper bound, the machinery is the loop-soup representation: clusters of the GFF level-set are the same as clusters of the critical loop soup, so crossing loops can be split into forward and backward crossing paths; stochastic domination controls rare large loops, and the BKR inequality bounds the probability that two connecting events are certified by disjoint loop collections. In $d\\ge7$ the correction $N^{6-d}$ comes from estimating the typical number $N^{d-6}$ of macroscopic loop clusters in the annulus and the chance that the two explored clusters attach to the same one.","core_discovery":"Let $P_D$ denote the law of the GFF conditioned to vanish on $D$. The central claim is that for $A_1$ well inside $\\widetilde B(N)$ and $A_2$ well outside it, $$P_{D_1\\cup D_2}(A_1\\overset{\\ge0}{\\leftrightarrow}A_2)\\asymp $N^{{(6-d)\\wedge 0}}$\\, P_{D_1}(A_1\\overset{\\ge0}{\\leftrightarrow}\\partial B(N))\\,P_{D_2}(A_2\\overset{\\ge0}{\\leftrightarrow}\\partial B(N))$$ for every $d\\ne6$, where the factor is $1$ for $3\\le d\\le5$ and $N^{6-d}$ for $d\\ge7$; at $d=6$ the ratio is bounded between $N^{-10\\varsigma(N)}$ and $N^{\\varsigma(N)}$. The lower and upper bounds use different representations of the same objects. For the lower bound one explores the sign cluster of each target set, extracts a positive harmonic average on the explored boundary, and applies the exact formula for the conditional connection probability of two sets with nonnegative boundary values. For the upper bound one passes to the loop-soup representation of the clusters, decomposes crossing loops, applies the BKR inequality to make the two boundary connections disjoint, and in high dimensions estimates the chance that the two explored clusters meet the same one of about $N^{d-6}$ macroscopic clusters in the annulus.","pith_inferences":["Editorial inference: For critical Bernoulli percolation in $d\\ge7$, the natural analogue suggested by this result is quasi-multiplicativity with correction $N^{6-d}$, matching the conjecture quoted in the introduction and making the failure of the no-correction form quantitative.","Editorial inference: At $d=6$, the paper's conjecture that $\\theta_6(N)\\asymp N^{-2}\\log^\\delta N$ implies the ratio in (1.12) should tend to $0$ at a polylogarithmic rate, so the critical dimension is the place where a nontrivial logarithmic correction should appear.","Editorial inference: The harmonic-like regularity results for point-to-set probabilities rely mainly on Green's function decay $|x-y|^{2-d}$, so a version of the same argument should work on other transient metric graphs, making the factorization portable beyond $\\mathbb{Z}^d$."],"forward_implications":["For $3\\le d\\le5$, a connection across the annulus costs exactly the product of the two boundary connections, with no correction; this is the factorization that the incipient-infinite-cluster construction requires.","For $d\\ge7$, the extra factor $N^{6-d}$ is necessary and sufficient, and the heuristic picture is that the annulus contains about $N^{d-6}$ macroscopic clusters, so the two sides must choose the same one.","At $d=6$, quasi-multiplicativity holds only up to $N^{o(1)}$; a sharper one-arm exponent would convert this into a definite polylogarithmic correction.","Point-to-set connection probabilities behave like harmonic functions: they decay like $|x-y|^{2-d}$, satisfy a Harnack inequality, and are stable under adding zero boundary conditions.","The cluster-volume tail satisfies $\\nu_d(M)\\asymp M^{-(d-2)/(d+2)}$ for $3\\le d\\le5$ and $\\nu_d(M)\\asymp M^{-1/2}$ for $d>6$, leaving only $d=6$ open."],"supporting_citations":[{"why":"Supplies the sharp one-arm exponent for $d>6$ and the tree-expansion and convolution estimates used for the high-dimensional correction.","marker":"[5]"},{"why":"Provides the one-arm and crossing probabilities for $3\\le d\\le6$, including the $d=6$ bounds with $\\varsigma(N)$, that Theorem 1.1 imports.","marker":"[6]"},{"why":"Gives the concurrent $d=3$ one-arm bound that completes the low-dimensional exponent in (1.3).","marker":"[12]"},{"why":"Formulates the quasi-multiplicativity conjecture for Bernoulli percolation and the incipient-infinite-cluster application that frame the theorem.","marker":"[2]"},{"why":"Establishes the isomorphism theorem identifying GFF sign clusters with loop-soup clusters, on which the upper-bound argument rests.","marker":"[15]"},{"why":"Supplies formula (2.30) for the conditional probability of two sets being connected under nonnegative boundary values, used in the lower bound.","marker":"[16]"},{"why":"Provides the BKR inequality for infinite spaces that the disjoint-cluster upper bound invokes.","marker":"[1]"}],"fun_headline_variants":["Critical GFF level-sets: annulus connection chance factorizes","Quasi-multiplicativity proven for metric graph GFF clusters","Annulus crossing probabilities factor up to N^(6-d) for GFF","GFF level-set clusters: annulus connections nearly factorize","GFF annulus crossings: product law up to N^{6-d}"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof imports the sharp one-arm and crossing exponents for the critical GFF level-set as theorems rather than rederiving them, and the d=6 statement inherits the uncertainty in that dimension's one-arm probability.","fun_headline_variants_meta":{"raw":{"variants":["Critical GFF level-sets: annulus connection chance factorizes","Quasi-multiplicativity proven for metric graph GFF clusters","Annulus crossing probabilities factor up to N^(6-d) for GFF","GFF level-set clusters: annulus connections nearly factorize","GFF annulus crossings: product law up to N^{6-d}"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001167,"raw_usage":{"total_tokens":5005,"prompt_tokens":1295,"completion_tokens":3710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":911,"completion_tokens_details":{"reasoning_tokens":3617}},"tokens_in":911,"tokens_out":3710,"duration_ms":23632,"temperature":1.0,"reasoning_tokens":3617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:26:24.974206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Determine the exact order of the one-arm probability $\\theta_6(N)$ at dimension 6. The paper's d=6 bounds have $N^{-10\\varsigma(N)}$ to $N^{\\varsigma(N)}$ slack precisely because only $N^{-2}\\le\\theta_6(N)\\le N^{-2+\\varsigma(N)}$ is known, so a sharper value outside this range would contradict (1.12). Alternatively, in $d\\ge7$ take $A_1=B(R_1)$ and $A_2=\\partial B(R_2)$ with distances violating $R_1\\lesssim N^{2/(d-4)}$ and $R_2\\gtrsim N^{(d-4)/2}$; the claimed $N^{6-d}$ factor fails there, and Remark 1.4 identifies these conditions as sharp.","supporting_citations":[{"cited_title":"Basu and A","cited_arxiv_id":null,"evidence_quote":"Formulates the quasi-multiplicativity conjecture for Bernoulli percolation and the incipient-infinite-cluster application that frame the theorem."},{"cited_title":"Arratia, S","cited_arxiv_id":null,"evidence_quote":"Provides the BKR inequality for infinite spaces that the disjoint-cluster upper bound invokes."}],"review_version":1}