{"id":"c7c7dffb-5308-4638-87db-9da980c1b4e5","arxiv_id":"2501.14707","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":9.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Cluster counts of the Gaussian free field on Z^d are asymptotically Gaussian in d≥4; in d=3 the limit is Gaussian except possibly at non-degenerate critical points of the cluster density, where a non-Gaussian Hermite distribution appears.","lead":"The paper proves that the number of sign and level-set clusters of the Gaussian free field on Z^d has Gaussian fluctuations for d≥4 at every non-critical level, and Gaussian or non-Gaussian fluctuations in d=3 depending on the level. It also shows sign-cluster counts fluctuate less than generic levels, a Berry-type cancellation, via a new chaos expansion for non-local functionals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's verdict of ACCEPT with moderate confidence is appropriate. The paper is long and technically dense, but the main logical structure is sound: finite chaos expansions for the cluster count, semi-localisation of the pivotal intensities via percolation-theoretic decay, and a semi-local limit theory that handles both Gaussian and non-Gaussian limits. The reader identified reliance on DGRS23 as the weakest assumption; I agree that the percolation input is the most external and least re-derived ingredient, but since it is published and the authors have isolated its role through uniform T_ell assumptions, this is not an internal flaw. My own most serious candidate was the (1-t) singularities in the joint pivotal intensity estimates used in Proposition 3.4; after re-reading the proof, the small-power extraction in Lemma 3.18 appears to keep the iterated integrals finite, so this does not rise to a load-bearing objection. I therefore recommend leaving the verdict unchanged.","tokens_in":63375,"tokens_out":58299,"duration_ms":493236,"concrete_test":"As a verification step, independently re-derive the bound on P^t_{R;>r}(x;y) in Lemma 3.18, tracking the (1-t) exponent from Proposition 3.14(3) and Proposition A.5 through the small-power extraction, and confirm that the iterated t-integral in Proposition 2.16/3.4 is finite for every m >= m0. If the exponent after the (m-1) integrations is less than -1, the tail bound in Proposition 3.4 would fail; this check would settle whether the hinge between de-pinning and tail integrability is sound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central claim is supported by a detailed and internally coherent argument: Theorem 2.9 gives a rigorous chaos expansion for level-set functionals, Propositions 3.1-3.4 semi-localise the pivotal intensities using pinned truncated arm decay, and Appendix B supplies the needed extended Breuer-Major/Dobrushin-Major theory. The main external input, DGRS23 truncated arm decay, is a published theorem, and the paper's Theorem 1.10 indicates that the argument only needs the a priori property T_ell. The most delicate internal step is the t-integrability in Proposition 3.4 after combining the (1-t) singularities from the de-pinning estimates and the Hermite-polynomial bounds; however, the small-power extraction in the proof of Lemma 3.18 appears to render the iterated integral finite. I did not find a load-bearing gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves limit theorems for the number of sign and level-set clusters of the discrete Gaussian free field (GFF) in large boxes. The main results are: (i) in dimension d ≥ 4, the centered and normalized cluster count converges to a standard Gaussian at every non-critical level (Theorem 1.1); (ii) in dimension d = 3, the limit is Gaussian except at non-degenerate critical points of the cluster density μ, where an order-2 Hermite distribution appears (Theorem 1.2); (iii) variance asymptotics for all non-critical levels, exhibiting Berry-cancellation suppression at levels where μ' = 0 (Theorem 1.4); and (iv) volume-order lower and R^{d+2}-order upper variance bounds at criticality (Theorem 1.6). The proof combines a Wiener-Itô chaos expansion for level-set functionals (Theorem 2.9), semi-localization of the chaotic components via truncated arm-decay estimates (Section 3), extensions of Breuer-Major/Dobrushin-Major theory to semi-local functionals (Appendix B), and a general modular theorem under a uniform arm-decay assumption (Theorem 1.10).","tokens_in":63476,"tokens_out":21629,"duration_ms":176447,"significance":"If correct, these results constitute the first limit theorems for the cluster count of the GFF and reveal a genuinely dimension-dependent Gaussian/non-Gaussian transition. The methodological contribution is substantial: the paper introduces a tractable chaos expansion for non-local functionals of strongly correlated Gaussian fields, links the coefficients to pivotal intensities, and shows how percolation-theoretic arm decay controls the tail of the expansion. The variance suppression for sign clusters (Berry cancellation) is a new geometric phenomenon for the GFF. The paper is honest about its limitations (e.g., Questions 1.3, 1.5, 1.7) and isolates the external DGRS23 arm-decay input cleanly. The proofs are detailed and internally coherent, with machine-checkable algebra in the appendices and clearly stated assumptions.","major_comments":[],"minor_comments":[{"comment":"In the paragraph following the mean-value theorem, the text states that taking ε→0 shows the right derivative of μ at ℓ_c is P∞(ℓ_c;0), and then concludes μ′(ℓ_c)=P∞(ℓ_c;0). However, the displayed equation in the same paragraph gives μ′(ℓ_c+ε̃)=−P∞(ℓ_c+ε̃;0), so the right derivative should be −P∞(ℓ_c;0), and the conclusion should read μ′(ℓ_c)=−P∞(ℓ_c;0), consistent with (4.1) for non-critical levels.","section":"Section 4.2 (proof of Proposition 4.1, critical levels)"},{"comment":"The sentence before (3.27) claims the terms vanish \"unless each point of x and y is within distance r of ∂Λ_R\". This is stronger than what Lemmas 3.11 and 3.17 provide (only one point in each tuple needs to be near the boundary, together with the diameter support condition), and the displayed bound (3.27) appears to contain typos: the summation index \"x2\" should presumably be \"y1\", and the summation range over boundary-layer points should be spelled out. Please clarify the argument.","section":"Proof of Proposition 3.4"},{"comment":"The displayed bound for φ_{f(x),f^t(y)}(ℓ,ℓ) is garbled (\"1√ 1 − t2Corr(f (x), f(y)\"). Please display the correct expression, for example using the bivariate normal density and the fact that Corr(f(x), f^t(y)) = tG(x−y)/G(0), so that the subsequent integration in t is transparent.","section":"Proposition 5.2"},{"comment":"The paper repeatedly writes \"Weiner-Itô\" (Abstract, Section 1.4.1, and elsewhere); this should be \"Wiener-Itô\".","section":"Throughout"},{"comment":"The statement \"Z if ℓ ∈ R \\ C′\" includes levels where μ′(ℓ)=0 with μ″(ℓ)=0 (e.g., flat critical points). The limit is indeed Gaussian in those cases, but a short remark explaining this (and pointing to the third and fourth cases of Theorem 1.4) would help readers connect the two theorems.","section":"Theorem 1.2"}],"recommendation":"minor_revision","confidential_remarks":"The paper is technically impressive and the main results appear correct. I found no load-bearing gaps; the issues are local presentation problems, including a sign error in the proof of Proposition 4.1 at the critical level (which does not affect the non-critical statements), a garbled equation in Proposition 3.4, and several typos. I recommend minor revision rather than acceptance in its current form so that the authors can correct these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is simple: this is the first paper to get actual limit theorems for the cluster count of the Gaussian free field, and the method—a chaos expansion adapted to non-local level-set functionals—is genuinely new and likely reusable. The main results (Gaussian limits for d≥4, the d=3 dichotomy, variance asymptotics, and Berry cancellation for sign clusters) are significant and clearly stated. The proof is long but structured, and I did not find a load-bearing gap. The reader's take and the stress-test both land: the arguments cohere, the pivotal-intensity machinery is well motivated, and the heavy reliance on DGRS23 truncated arm decay is explicit and properly cited.\n\nWhat the paper does especially well is the honest handling of its own boundaries. The d=3 non-Gaussian case is stated conditionally on the set C′ of non-degenerate critical points of the cluster density, and the authors explicitly say they cannot prove C′ is non-empty; the theorem remains valid either way, and they flag it as an open problem. The variance bounds at criticality are presented as bounds, not dressed up as a full result. The self-citations to BMM22/BMM24 provide context and alternative arguments, not the load-bearing step—so no circularity concern.\n\nSoft spots are proportional. The paper depends on the deep DGRS23 theorem, so if that theorem had a flaw the semi-localization would collapse; but that is true of any paper that uses a published result. The length and technical density will make referee work heavy, and the most delicate internal step—the t-integrability after combining the (1−t) singularities—looks fine but is the kind of estimate that should be checked carefully by the referee. The boundary effects in the variance constants (Remark 6.1) are a minor caveat but not a flaw.\n\nWho is this for? Anyone working on level-set percolation for strongly correlated fields, and probabilists interested in chaos expansion techniques for non-local functionals. It deserves a serious referee and probably a revise-and-resubmit, not a desk rejection. I would not cite it in my own immediate work, but I would bring it to a reading group if someone had the patience for a long technical proof.","headline":"First limit theorems for GFF cluster counts, built on a new semi-local chaos expansion; the proof is long and leans on heavy percolation input, but the central argument holds up and the paper deserves serious refereeing.","tokens_in":713,"tokens_out":1350,"would_cite":false,"duration_ms":63811,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G60","60G15","60F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the Gaussian free field on $\\mathbb{Z}^d$, the number of sign and level-set clusters in a large box has Gaussian fluctuations in $d\\ge 4$ at every non-critical level, while in $d=3$ a non-Gaussian Hermite law appears exactly at…","keywords":["Gaussian free field","level-set clusters","sign clusters","chaos expansion","pivotal intensities","Hermite distributions","cluster density","percolation"],"falsifier":"Compute (rigorously or numerically) the cluster density $\\mu$ near level 0 in dimension $d=3$. If $\\mu''(0)\\neq 0$, the normalized sign-cluster count should converge to the order-2 Hermite law with variance $\\sim (\\beta_{3,2}(\\mu''(0))^2/2)R^4$; if $\\mu''(0)=\\mu'''(0)=0$, it should converge to a Gaussian with variance $\\sim\\sigma^2R^3$. Finding any zero of $\\mu'$ with $\\mu''\\neq 0$ would confirm that the non-Gaussian regime exists, and finding $\\mu'(\\ell_c)\\neq 0$ would confirm the critical lower bound.","tokens_in":63171,"feed_emoji":"📊","tokens_out":10553,"duration_ms":93348,"temperature":0.7,"pith_summary":"The Gaussian free field on $\\mathbb{Z}^d$ is a canonical strongly correlated random field, and this paper asks how many of its sign and level-set clusters (connected components of $\\{f>\\ell\\}$ and $\\{f<\\ell\\}$) sit inside a large box. The answer: in dimensions $d\\ge 4$, at every non-critical level the cluster-count fluctuations are asymptotically Gaussian after centering and scaling by the standard deviation. In dimension $d=3$ the limit can instead be a non-Gaussian order-2 Hermite distribution, and this happens precisely at non-degenerate critical points of the cluster density $\\mu$. The paper also determines the variance order at every non-critical level, showing that sign clusters have suppressed fluctuations compared with generic levels because $\\mu'(0)=0$ by symmetry. The proof is the first chaos expansion for a non-local functional of a strongly correlated Gaussian field, with percolation-theoretic control of the pivotal intensities in the expansion.","feed_headline":"GFF cluster counts turn Gaussian in high dimensions and twist in d=3","feed_subtitle":"A chaos-expansion proof ties the fluctuations to derivatives of the cluster density, and explains why sign clusters fluctuate less.","key_machinery":"The central object is the chaos expansion of the level-set functional, whose $m$-th coefficient is the pivotal intensity $P(y)=\\mathbb{E}[d_y\\Xi(f-\\ell)|f(y)=\\ell]\\varphi_{f(y)}(\\ell)$; by Proposition 2.7 this equals the $m$-th mixed derivative of $\\mathbb{E}[\\Xi(f-\\ell)]$. The proof's main work is semi-localisation: percolation truncated-arm decay, extended to conditioned fields by a de-pinning argument, shows $P$ can be replaced by a stationary, exponentially decaying $P_\\infty$, so each chaos becomes a sum over Wick products with rapidly decaying coefficients. The identity connecting sums of $P_\\infty$ to derivatives of $\\mu$ then converts analytic properties of $\\mu$ into variance asymptotics and limit laws. The tail of the expansion is controlled by an iterated interpolation formula that expresses the tail variance through joint pivotal intensities of a fixed order, avoiding the lack of smoothness of the level-set functional.","core_discovery":"Write $N_R(\\ell)$ for the number of bounded clusters of $\\{f>\\ell\\}$ and $\\{f<\\ell\\}$ in the box $\\Lambda_R$, and let $\\mu(\\ell)$ be the cluster density from the law of large numbers. The paper proves that for every $\\ell\\neq \\pm\\ell_c$ the $m$-th chaos $Q_m$ of $N_R$ has variance governed by the stationary pivotal intensity $P_\\infty$, in the explicit form $\\sum_{x_2,\\ldots,x_m} P_\\infty(0,x_2,\\ldots,x_m)=(-1)^m \\mu^{(m)}(\\ell)$. Hence the dominant chaos is the first with a non-vanishing derivative of $\\mu$: $\\mu'(\\ell)\\neq 0$ gives a Gaussian limit driven by the first chaos; in $d=3$, $\\mu'(\\ell)=0$ and $\\mu''(\\ell)\\neq 0$ gives a non-Gaussian order-2 Hermite limit driven by the second chaos; and if the first three derivatives vanish, the variance drops to volume order and the limit is again Gaussian. In $d\\ge 4$ all remaining chaoses are Gaussian by the high-order chaos central limit principle, so the whole count is Gaussian at every non-critical level. The sign clusters are covered because $\\mu'(0)=0$ by symmetry.","pith_inferences":["The open question of whether $\\mu''(0)$ vanishes in $d=3$ is exactly the question of whether sign clusters have a non-Gaussian limit; a computation of the cluster density near zero would settle it, since the paper reduces the law to this derivative.","The same pivotal-intensity machinery should transfer to other semi-local functionals such as the volume of the unbounded component above the critical level; monotonicity of that functional suggests only Gaussian limits there.","For covariance decays $\\alpha<d-2$, the mechanism predicts Hermite limits of every order $2\\le m<d/\\alpha$ with additional boundary contributions, so the Gaussian/non-Gaussian split found here is the $\\alpha=d-2$ case of a broader classification."],"forward_implications":["For $d\\ge 4$, at every non-critical level including sign clusters, $(N_R(\\ell)-\\mathbb{E}N_R)/\\sqrt{\\operatorname{Var}N_R}$ converges to a standard Gaussian.","For $d=3$, non-Gaussian order-2 Hermite fluctuations occur exactly at levels where $\\mu''(\\ell)\\neq 0$ and $\\mu'(\\ell)=0$; whether the sign level $\\ell=0$ falls there is left open.","The variance asymptotics give explicit orders: for example, in $d\\ge 5$, $\\operatorname{Var}N_R(\\ell)\\sim \\beta_{d,1}(\\mu'(\\ell))^2R^{d+2}$ if $\\mu'(\\ell)\\neq 0$ and $\\sim\\sigma^2R^d$ otherwise, with a logarithmic factor in the $d=4$ borderline case.","At the critical level the variance lies between $c_1R^d$ and $c_2R^{d+2}$, and if $\\mu'(\\ell_c)\\neq 0$ the lower bound is actually $\\beta_{d,1}(\\mu'(\\ell_c))^2R^{d+2}(1+o(1))$.","The method applies to other level-set functionals of the same field and to any stationary Gaussian field with covariance $|x|^{-(d-2)}$ plus an i.i.d. component, wherever uniform truncated arm decay holds."],"supporting_citations":[{"why":"Supplies the super-polynomial truncated arm decay for off-critical GFF level sets, the percolation input that makes semi-localisation of pivotal intensities possible.","marker":"[DGRS23]"},{"why":"Establishes real-analyticity of the cluster density away from critical levels, underlying the definition of the critical sets $C$ and $C'$ and the finite-level structure.","marker":"[PS22]"},{"why":"Classical non-central limit theorem for chaos of Gaussian fields, extended here to semi-local functionals and to the Hermite distributions that appear for $m(d-2)<d$.","marker":"[DM79]"},{"why":"Classical central limit theorem for nonlinear functionals of Gaussian fields, supplying the Gaussian limits for high-order chaoses and the moment method.","marker":"[BM83]"},{"why":"Gives the general variance upper bound and covariance arguments adapted in the critical-level bound of Theorem 1.6.","marker":"[BMM24b]"},{"why":"Gives the lower variance bound at levels with $\\mu'(\\ell)\\neq 0$, used in the critical case and in the variance analysis.","marker":"[BMM22]"},{"why":"Interpolation formula for covariances of smooth Gaussian functionals, iterated to control the tail of the chaos expansion.","marker":"[Cha08]"},{"why":"Fourth-moment theorem used to justify Gaussian convergence of fixed-order chaoses.","marker":"[NP05]"}],"fun_headline_variants":["GFF cluster counts: Gaussian in d≥4, non-Gaussian in d=3","Sign clusters of GFF show Berry cancellation, lower fluctuations","Chaos expansion yields Gaussian and Hermite limits for cluster counts","Gaussian free field clusters: d≥4 Gaussian, d=3 level-dependent","Cluster count fluctuations tied to first nonvanishing derivative of density"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof leans on the sharp percolation estimate that, at every non-critical level, the probability of a bounded cluster of the GFF excursion set reaching radius $R$ decays faster than any polynomial; without this truncated arm decay, the pivotal intensities cannot be shown to localise and the variance asymptotics collapse.","fun_headline_variants_meta":{"raw":{"variants":["GFF cluster counts: Gaussian in d≥4, non-Gaussian in d=3","Sign clusters of GFF show Berry cancellation, lower fluctuations","Chaos expansion yields Gaussian and Hermite limits for cluster counts","Gaussian free field clusters: d≥4 Gaussian, d=3 level-dependent","Cluster count fluctuations tied to first nonvanishing derivative of density"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1530,"prompt_tokens":987,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":603,"tokens_out":543,"duration_ms":5142,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:52:39.627082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute (rigorously or numerically) the cluster density $\\mu$ near level 0 in dimension $d=3$. If $\\mu''(0)\\neq 0$, the normalized sign-cluster count should converge to the order-2 Hermite law with variance $\\sim (\\beta_{3,2}(\\mu''(0))^2/2)R^4$; if $\\mu''(0)=\\mu'''(0)=0$, it should converge to a Gaussian with variance $\\sim\\sigma^2R^3$. Finding any zero of $\\mu'$ with $\\mu''\\neq 0$ would confirm that the non-Gaussian regime exists, and finding $\\mu'(\\ell_c)\\neq 0$ would confirm the critical lower bound.","supporting_citations":[],"review_version":1}