{"id":"87cc1d0e-0805-4809-b976-d4a2b12a7865","arxiv_id":"2608.12804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"High-temperature Ising models on graphon random graphs have Gaussian spin statistics with covariance given by the graphon resolvent, yielding functional and Sobolev-space limits.","lead":"This paper proves central limit theorems for spin fluctuations of high-temperature Ising models on large inhomogeneous random graphs, with covariance described by a resolvent of the graphon operator. The results extend to Gaussian-process limits for Bayesian neural networks and to asymptotic normality of treatment-effect estimators under network interference.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.1's MGF ratio rests on unverified lemmas from the authors' companion preprint; that is the load-bearing gap.","rationale":"The reader's weakest_assumption was the sparsity lower bound N^{2/3}θ_N ≳ 1, and the paper explicitly restricts to that regime. I agree that the abstract's 'sparse graphs' phrasing is broader than the technical condition, but the theorem statement is precise. The more load-bearing issue is that the proof of Theorem 2.1 depends on core lemmas from the authors' companion preprint [54], with one proof omitted and several results quoted verbatim. This is not an ad hominem concern; it is a missing-support concern: if the companion's exponential expansions or Gaussian-chaos limit are wrong, the MGF ratio calculation in Section 4.1 collapses and the resolvent covariance in Theorem 2.1 does not follow. The paper's own appendix labels these lemmas as coming from [54], so the dependence is explicit and verifiable. A conditional verdict is appropriate: the mathematical architecture is coherent, the applications follow from the main CLT, and the examples are consistent with known ER limits, but independent verification of the delegated lemmas (or inclusion of their proofs) is needed before full acceptance. My read does not move the reader's verdict, hence UNCHANGED.","tokens_in":45306,"tokens_out":19386,"duration_ms":175483,"concrete_test":"Obtain arXiv:2606.07065 and independently verify Lemma F.1 (used here as Lemmas E.1 and E.2) and Proposition F.1 (used as Lemma E.5), including the o(1/(Nθ_N)) remainders and the O(Q_N(σ)) bounds under θ_N = Ω(N^{-2/3}). In particular, rederive the expansion for E[T(σ)] from first principles and check that the β^4/(N^4 θ_N^3) deterministic term cancels in the MGF ratio while the stated remainder estimates hold. If these companion lemmas are correct, Theorem 2.1 is supported; if any fails, the central CLT is unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of the central Theorem 2.1 reduces the conditional MGF to the ratio ψ_N(t)=Ẑ_f^N(β,t)/Ẑ_N(β) and invokes Lemma B.2 to replace numerator and denominator by their annealed expectations. Lemma B.2 relies on Lemma E.1/E.2 for the expansion of E[T(σ)], Lemma E.3 for the two-copy linearization, and Lemma E.5 for the Gaussian-chaos limit of Y_{f,N}. The paper states that these are 'the first part of Lemma F.1 (Eq. (F.1))' and 'Proposition F.1' of the companion preprint [54], and the proof of Lemma E.3 is explicitly omitted. Lemma B.5 further cites [54] for moment bounds. Since [54] is a preprint by three of the same authors, the pivotal exponential expansions have no independent verification in this manuscript. If any of these lemmas fails at the stated threshold θ_N = Ω(N^{-2/3}), the convergence in (4.10) to the resolvent covariance does not follow. The sparsity condition is a genuine scope restriction—expected degree grows at least like N^{1/3}—but it is stated explicitly in Assumption 1(3); the abstract's 'sparse graphs' is broader-sounding, yet that is a presentation issue rather than a mathematical flaw. The delegation to [54] is therefore the load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a quenched fluctuation theory for Ising models on inhomogeneous random graphs generated by graphons, in the high-temperature regime β||T_W||_{op}<1. The central result, Theorem 2.1, is a multivariate CLT for linear statistics σ_N(f_1),...,σ_N(f_k) conditional on the random graph, with limiting covariance ⟨f_i,(I−βT_W)^{-1}f_j⟩. This drives three infinite-dimensional extensions: a Donsker-type limit for the partial-sum process S_N in D[0,1] (Theorem 2.2), a function-indexed CLT over compact classes of absolutely continuous test functions (Theorem 2.3), and convergence of the empirical spin field η_N in negative Sobolev spaces H^{-s}(0,1) for s>1/2 (Theorem 2.4). The paper also derives two applications: an infinite-width Gaussian-process limit for two-layer Bayesian neural networks with Ising-dependent output-layer signs (Theorem 3.1), and asymptotic normality of Hájek estimators under network interference when treatment assignments follow the graphon Ising model (Theorem 3.2). The proofs use a rank-one perturbation analysis of the partition function, concentration of random partition functions around annealed expectations, a 2-modified Fredholm determinant identity, and moment bounds obtained from the Lebowitz inequality.","tokens_in":45563,"tokens_out":3526,"duration_ms":40887,"significance":"If the technical lemmas from the companion preprint [54] are correct, this is a substantial contribution: it unifies and extends known results for Erdős–Rényi and Curie–Weiss Ising models to the full graphon setting, and it provides a resolvent characterization of the limiting covariance that is both explicit and checkable in examples. The paper contains several genuinely useful auxiliary results, including the operator-norm concentration bound for graphon adjacency matrices (Lemma B.5), the quenched moment bounds (Lemma B.4), and the explicit Fredholm-determinant computation (Lemma C.2). The applications are well chosen and the recovery of known special cases — for example the ER magnetization variance 1/(1−β) and the ReLU neural-network Gaussian-process kernel at β=0 — is a genuine strength. The main limitation is that the load-bearing exponential expansions and Gaussian-chaos limits (Lemmas E.1, E.2, E.3, E.5) are delegated to an unpublished companion paper by three of the same authors, so the central result is not self-contained as submitted.","major_comments":[{"comment":"The proof of Theorem 2.1 rests on Lemmas E.1, E.2, E.3, and E.5, which are quoted from the companion preprint [54] or stated with only sketched proofs. In particular, Lemma E.3 is explicitly stated with 'proof omitted', and Lemma E.5 is justified by a cumulant sketch rather than a full proof. These lemmas are exactly what converts the conditional moment-generating function ratio in (4.4)–(4.5) into the resolvent covariance appearing in (4.10). Since [54] has three authors in common with the present paper and is itself an arXiv preprint, the central claim of Theorem 2.1 is not independently verifiable from the manuscript alone. The authors should either include complete proofs of these lemmas in an appendix or restructure the paper so that the main theorem is proved from first principles.","section":"§4.1 and Appendix E"},{"comment":"The sparsity condition N^{2/3}θ_N ≳ 1 means the expected degree grows at least like N^{1/3}, so the paper covers only moderately sparse graphs, not bounded-degree sparse graphs. This restriction is explicitly stated in Assumption 1(3) and is used in Lemma B.5 via the matrix Bernstein bound (B.26), but the abstract and the introduction describe the results as covering 'both dense and sparse graphs' without this qualification. The scope statement should be revised so that readers are not led to expect applicability in the bounded-degree regime, where the proof strategy and possibly the CLT behavior are different.","section":"Assumption 1(3) and Lemma B.5"},{"comment":"Lemma B.2, which is the concentration step for the ratio of random partition functions in (4.4), relies on Lemmas E.1, E.2, E.3, and on 'the argument leading to [54, Eq. (A.51)]' and 'the proof of Lemma A.8 in [54]'. The variance bound O(1/(Nθ_N)) is the quantitative reason the conditional MGF converges to its annealed limit, so this lemma is load-bearing for Theorem 2.1. As with the companion lemmas, the dependence on [54] should be made explicit in the statement and, ideally, the proof should be included in the present paper.","section":"Lemma B.2 and Appendix B"}],"minor_comments":[{"comment":"The abstract's phrase 'encompassing both dense and sparse graphs' should be qualified by the condition N^{2/3}θ_N ≳ 1, for example by saying 'moderately sparse graphs' or by explicitly stating the sparsity range in the abstract.","section":"Abstract and §1.2"},{"comment":"The first bullet states the sparsity condition as θ_N = Ω(N^{-2/3}); this is equivalent to Assumption 1(3) but the notation 'Ω' is only introduced in §1.3, so the reader meets it slightly earlier than the notation section.","section":"§1.2, first bullet"},{"comment":"The variance computation for X_rank(1) is correct but somewhat compressed; the intermediate use of Cov(B(1), ∫_0^1 g dB) = ∫_0^1 g could be displayed for readability.","section":"Example 2"},{"comment":"The statement says 'All three infinite series appearing in the distributional limit converge in L^2 and almost surely', but the proof only sketches the cumulant computation. For a lemma that is used in the main proof, a complete proof or a precise reference to [54] is needed; as written this is a presentation gap.","section":"Lemma E.5"},{"comment":"The Ornstein–Uhlenbeck representation is a nice addition, but the claim that U_s has the law of X for each fixed s would benefit from a one-sentence justification that the Gaussian process in (F.1) has the same covariance as X, which is already established in Lemma D.1.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The heavy reliance on Lemmas E.1, E.2, E.3, and E.5 from the authors' own companion preprint [54] is the main risk. Even if the lemmas are correct, the current manuscript cannot be evaluated independently without that preprint. The editor may wish to request that the authors either append full proofs of these lemmas or make the companion preprint publicly available and explicitly identify which statements are proved there. The sparsity qualification in the abstract is also worth correcting in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the right statement: quenched Gaussian fluctuations for linear spin statistics on graphon-generated graphs, with covariance given by the resolvent (I-βT_W)^{-1}, plus process-level and negative-Sobolev field limits. Second, the proof of the central CLT is not self-contained; it imports four technical lemmas from a companion preprint by three of the same authors, and if one of those fails the whole thing collapses.\n\nWhat's genuinely good: the resolvent covariance is the natural high-temperature answer, and the paper checks that it reduces to known ER and Curie–Weiss limits. The Donsker-type partial-sum limit and the H^{-s} field limit are new, as are the two applications. The neural-network limit gets a clean rank-one correction to the classical kernel; the Hajek estimator variance formula is derived transparently. The examples (two-block SBM, rank-one graphon) are worked out and consistent.\n\nWhere the soft spots are: the load-bearing problem is delegation. Lemma B.2 – which converts the conditional MGF into an annealed ratio – rests on Lemmas E.1, E.2, E.3 and E.5, which are quoted from [54]. E.3's proof is explicitly omitted. [54] is an arXiv preprint by the same authors (plus a fourth) and has not been independently vetted. That is a real circularity burden; the reader cannot check the central step without trusting the companion. It is not a fatal flaw, but it should be fixed before full acceptance.\n\nThe sparsity condition is a lesser issue. Assumption 1(3) requires θ_N = Ω(N^{-2/3}), expected degree at least of order N^{1/3}. The abstract says 'sparse graphs', which is broader than that; it is a presentation problem, not a hidden mathematical error. The paper should say 'moderately sparse' or qualify the abstract.\n\nWho it's for: probabilists and statistical mechanicians working on mean-field spin systems on networks, and statisticians interested in network-dependent treatment assignment. It deserves a serious referee. The right editorial decision is to send it to review, but to ask the authors to either incorporate the deferred lemmas or certify that [54] has been completed and posted with full proofs. Conditional on that, the result looks solid.","headline":"A strong and coherent generalization of Ising CLTs to graphon random graphs, but the main theorem's proof leans on an unreviewed companion preprint, so the result is conditional until then.","tokens_in":46105,"tokens_out":2324,"would_cite":false,"duration_ms":23844,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","60K35","82B20","05C80","60G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a quenched central limit theorem for Ising linear statistics on inhomogeneous random graphs, with covariance given by the resolvent of the graphon operator.","keywords":["Ising model","inhomogeneous random graphs","graphon","quenched central limit theorem","Gaussian process limit","negative Sobolev space","Bayesian neural networks","causal interference"],"falsifier":"Simulate the high-temperature Ising model with $W\\equiv1$ and sparsity $\\theta_N=N^{-3/4}$ (below the paper's threshold) at $\\beta=1/2$, and estimate the quenched law of $N^{-1/2}\\sum_{i=1}^N\\sigma_i$ across graphs. If its variance does not converge to $1/(1-\\beta)=2$ or the law is not asymptotically Gaussian, the $N^{2/3}\\theta_N\\gtrsim1$ condition is a genuine boundary for the Gaussian limit; if it still converges, the threshold is an artifact of the proof.","tokens_in":45073,"feed_emoji":"🧲","tokens_out":10194,"duration_ms":101734,"temperature":0.7,"pith_summary":"At high temperature, every macroscopic linear statistic of the Ising spin configuration on a graphon-generated random graph is asymptotically Gaussian, even conditional on the random graph itself. The covariance between two statistics with test functions $f$ and $g$ is $\\langle f,(I-\\beta T_W)^{-1}g\\rangle$, where $T_W$ is the integral operator of the graphon and $\\beta$ is the inverse temperature; the whole network effect is encoded by this resolvent. From this finite-dimensional theorem the paper derives a functional limit for the partial-sum process, a function-indexed central limit theorem, and convergence of the full empirical spin field in negative Sobolev spaces. The same results give explicit infinite-width Gaussian-process limits for Bayesian neural networks with Ising-coupled output signs and asymptotic normality for Hajek treatment-effect estimators under network interference.","feed_headline":"Graphon resolvent controls spin-field Gaussian limits","feed_subtitle":"Graphon structure enters only through a resolvent covariance; neural-net and causal limits follow.","key_machinery":"The load-bearing computation is the conditional moment-generating function of $\\sigma_N(f)^2$, written as a ratio of random partition functions. Concentration results inherited from the companion paper reduce this ratio to an annealed expectation, and a rank-one determinant identity collapses the infinite eigenvalue product to $e^{t\\int f^2}\\sqrt{\\det_2(I-\\beta T_W)/\\det_2(I-\\beta T_W-2t|f\\rangle\\langle f|)}=1/\\sqrt{1-2t\\langle f,(I-\\beta T_W)^{-1}f\\rangle}$. This 2-modified Fredholm determinant identity is what turns the graphon's spectrum into the resolvent covariance; ferromagnetic correlation bounds and Kolmogorov tightness then lift the finite-dimensional statement to process and Sobolev-field convergence.","core_discovery":"Conditional on the graph adjacency matrix $A_N$, the vector $(\\sigma_N(f_1),\\ldots,\\sigma_N(f_k))$ converges weakly in probability to $\\mathcal{N}_k(0,\\Sigma)$ with $\\Sigma_{ij}=\\langle f_i,(I-\\beta T_W)^{-1}f_j\\rangle$, for any Riemann-integrable $f_1,\\ldots,f_k$ and $\\beta\\|W\\|_{\\mathrm{op}}<1$, $N^{2/3}\\theta_N\\gtrsim 1$, $\\theta_N\\to\\theta$. The same resolvent covariance drives the partial-sum limit $X(t)=B(t)+\\sum_i((1-\\beta\\lambda_i)^{-1/2}-1)\\Phi_i(t)Z_i$ and the $H^{-s}(0,1)$-valued limit of the empirical spin field $\\eta_N$ for every $s>1/2$. Explicit examples for two-block and rank-one graphons show how community structure and latent profiles imprint finite-rank corrections on the Gaussian fluctuations, and these examples feed the neural-network and causal-inference applications.","pith_inferences":["An open question left by the paper is whether the same resolvent covariance persists when $\\theta_N$ decays faster than $N^{-2/3}$; the proof's concentration estimates degrade exactly at that threshold, so the Gaussian limit in bounded-degree sparse graphs is not established here.","Because the covariance is written through $(I-\\beta T_W)^{-1}$, the same fluctuation formalism should carry over to other mean-field spin systems whose partition functions admit the analogous rank-one perturbation identity, with the Ising-specific ferromagnetic bounds replaced by the corresponding correlation inequalities.","The quenched statement suggests a practical route to conditional inference on a single large network: estimate the graphon from the observed adjacency structure and use the resolvent covariance for approximate confidence sets, rather than resampling new networks."],"forward_implications":["For an Erdos-Renyi graph ($W\\equiv1$), the scaled magnetization converges to $\\mathcal{N}(0,(1-\\beta)^{-1})$ and the partial-sum process converges to $B(t)+((1-\\beta)^{-1/2}-1)tB(1)$, recovering and extending the known dense-graph result.","For a two-block stochastic block model, the covariance is $\\langle f_i,f_j\\rangle+\\frac{\\beta(p+q)}{2-\\beta(p+q)}m_i^+ m_j^+ + \\frac{\\beta(p-q)}{2-\\beta(p-q)}m_i^- m_j^-$, so the two community eigenmodes appear as a rank-two correction to the independent-spin covariance.","The empirical spin field converges in $H^{-s}(0,1)$ for every $s>1/2$, so the full random field rather than only its projections has a Gaussian scaling limit.","A two-layer Bayesian neural network whose output signs follow the Ising measure has infinite-width Gaussian-process limits with kernel $C_a(x,y)+(q_\\beta-1)m_a(x)m_a(y)$, a rank-one departure from the classical neural-network Gaussian-process kernel.","The Hajek estimator of the direct average treatment effect under graphon-Ising treatment assignment is asymptotically normal with variance $\\kappa_2+\\kappa_1^2(q_\\beta-1)$, extending the Curie-Weiss analysis to inhomogeneous networks."],"supporting_citations":[{"why":"Supplies the core concentration and remainder lemmas that make the partition-function ratio in the proof of Theorem 2.1 converge.","marker":"[54]"},{"why":"Provides the trace-ideal and 2-modified Fredholm determinant identities used to convert the eigenvalue product into the resolvent covariance.","marker":"[66]"},{"why":"Provides the rank-one determinant identity $\\det(I+|f\\rangle\\langle g|)=1+\\langle f,g\\rangle$ used in Lemma C.2.","marker":"[41]"},{"why":"Establishes the dense Erdos-Renyi magnetization CLT that the paper's $W\\equiv1$ example recovers as a special case.","marker":"[37]"},{"why":"Supplies the Gaussian-domination and correlation inequalities needed for the quenched moment and tightness bounds.","marker":"[6]"},{"why":"Gives the spectral bound on two-point correlations used in Lemma B.4's quenched fourth-moment estimate.","marker":"[65]"},{"why":"Matrix Bernstein inequality used to control the operator-norm deviation of the random adjacency kernel and justify the sparsity condition.","marker":"[70]"},{"why":"Defines the neural-network Gaussian-process kernel that the Ising-dependent application modifies by a rank-one term.","marker":"[55]"},{"why":"Gives the Curie-Weiss Hajek estimator limit that Theorem 3.2 extends to graphon-Ising treatment assignment.","marker":"[16]"}],"fun_headline_variants":["Graphon resolvent sets Ising spin-field Gaussian limits","Ising CLTs on inhomogeneous graphs via graphon resolvent","Graphon operator drives spin limits and neural-net applications","Quenched spin-field limits from a single resolvent covariance","Resolvent covariance sets Ising scaling limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof requires the random graph to stay moderately dense: $N^{2/3}\\theta_N$ must remain bounded below by a positive constant, so the average degree grows at least like $N^{1/3}$; if sparsity is stronger, the concentration estimates that make the partition-function ratio converge can break down.","fun_headline_variants_meta":{"raw":{"variants":["Graphon resolvent sets Ising spin-field Gaussian limits","Ising CLTs on inhomogeneous graphs via graphon resolvent","Graphon operator drives spin limits and neural-net applications","Quenched spin-field limits from a single resolvent covariance","Resolvent covariance sets Ising scaling limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00051,"raw_usage":{"total_tokens":2480,"prompt_tokens":943,"completion_tokens":1537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":1457}},"tokens_in":559,"tokens_out":1537,"duration_ms":14573,"temperature":1.0,"reasoning_tokens":1457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:56:05.565830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the high-temperature Ising model with $W\\equiv1$ and sparsity $\\theta_N=N^{-3/4}$ (below the paper's threshold) at $\\beta=1/2$, and estimate the quenched law of $N^{-1/2}\\sum_{i=1}^N\\sigma_i$ across graphs. If its variance does not converge to $1/(1-\\beta)=2$ or the law is not asymptotically Gaussian, the $N^{2/3}\\theta_N\\gtrsim1$ condition is a genuine boundary for the Gaussian limit; if it still converges, the threshold is an artifact of the proof.","supporting_citations":[{"cited_title":"Ising Models on Inhomogeneous Random Graphs: Inference, Local Asymptotic Minimaxity, and Limit of Experiments","cited_arxiv_id":"2606.07065","evidence_quote":"Supplies the core concentration and remainder lemmas that make the partition-function ratio in the proof of Theorem 2.1 converge."},{"cited_title":"From Painlev´ e to Zakharov–Shabat and beyond: Fredholm determi- nants and integro-differential hierarchies.Journal of Physics A: Mathematical and Theoretical, 54(3):035001, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the rank-one determinant identity $\\det(I+|f\\rangle\\langle g|)=1+\\langle f,g\\rangle$ used in Lemma C.2."},{"cited_title":"Fluctuations of the magnetiza- tion for Ising models on dense Erd¨ os–R´ enyi random graphs.Journal of Statistical Physics, 177(1):78–94, 2019","cited_arxiv_id":null,"evidence_quote":"Establishes the dense Erdos-Renyi magnetization CLT that the paper's $W\\equiv1$ example recovers as a special case."},{"cited_title":"Bhattacharya and Sumit Mukherjee","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-domination and correlation inequalities needed for the quenched moment and tightness bounds."},{"cited_title":"Spectral bounds for the Ising ferromag- net on an arbitrary given graph.Journal of Statistical Mechanics: Theory and Experiment, 2017(5):053403, 2017","cited_arxiv_id":null,"evidence_quote":"Gives the spectral bound on two-point correlations used in Lemma B.4's quenched fourth-moment estimate."},{"cited_title":"Neal.Bayesian Learning for Neural Networks, volume 118 ofLecture Notes in Statistics","cited_arxiv_id":null,"evidence_quote":"Defines the neural-network Gaussian-process kernel that the Ising-dependent application modifies by a rank-one term."},{"cited_title":"Cattaneo, Yihan He, and Ruiqi Rae Yu","cited_arxiv_id":null,"evidence_quote":"Gives the Curie-Weiss Hajek estimator limit that Theorem 3.2 extends to graphon-Ising treatment assignment."}],"review_version":1}