{"id":"5a6ddb3b-10d7-4507-992e-c370c5b9d726","arxiv_id":"2506.15159","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For subcritical exponential random graphs conditioned on edge count, the two-star count satisfies a central limit theorem with explicit mean and variance, at rate n^{-1/2+ε} in Wasserstein distance.","lead":"This paper proves that in a common family of network models, after fixing the total number of edges, the number of two-star patterns follows a normal distribution with an explicit formula for its center and spread. This gives statisticians a precise benchmark for testing whether a network really comes from such a model or from a simpler one.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The theorem conditions on {En/N = ep_n} for arbitrary ep_n in (0,1), but the event is empty unless N ep_n is an integer, so the statement is not well-posed.","rationale":"I agree with the reader that the achievability of ep_n is the main load-bearing concern. The theorem's proof is long but the central mechanism — the exchangeable-pair conditional CLT (Proposition 2.1), the linearity computation (Proposition 2.2), and the local CLT (Proposition 1.2) — appears internally consistent for integer N ep_n. I checked the algebraic steps: the sign in equation (2.54) appears to have a sign typo (the bracket should contain '-2 eq² Σ β_l s_l N ep^{e_l}' rather than '+'), but the subsequent displayed mean (2.55) is consistent with the corrected sign, so the final formulas (1.6)-(1.7) are not affected. The theorem statement would also benefit from explicitly assuming the full parameter vector lies in the compact set B, which the proof already does. None of these secondary issues changes the verdict: the manuscript should be accepted conditionally on fixing the conditioning statement and minor typos.","tokens_in":34557,"tokens_out":41077,"duration_ms":373383,"concrete_test":"Set n = 3 (so N = 3 and N ep_n = 1.5 for ep_n = 1/2); the event {En/N = 1/2} is empty and P(En = 1.5) = 0, so (1.5) is not defined. Then check whether the proof becomes valid after re-stating Theorem 1.1 with the extra assumption N ep_n ∈ ℤ or with conditioning on En = k_n for an integer k_n; the key exchangeable-pair argument (Proposition 2.2) and local CLT estimates already apply verbatim to {eE = 0} when N ep_n is integer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central statement quantifies over every ep_n in (0,1), but the event {En/N = ep_n} has probability zero unless N ep_n is an integer, because En is integer-valued. The proof conditions on eE = 0, i.e., En = N ep_n (around (2.17) in Section 2.2), and relies on the local CLT (Proposition 1.2) to ensure P(eE = 0) > 0 at integer k. For non-achievable densities the conditional law is undefined and the Wasserstein bound (1.5) is vacuous or meaningless. This is a statement-level gap rather than a flaw in the proof for achievable edge counts: if one adds 'N ep_n is an integer' or reformulates with an integer sequence k_n such that |k_n - N ep_n| ≤ 1, the machinery appears to go through unchanged.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a conditional central limit theorem for the number of two-stars in a dense exponential random graph model (ERGM) when conditioning on the total number of edges, under the subcriticality condition (1.3). The centering and scaling constants are given explicitly in terms of the conditioned edge density ep_n and the model parameters (Theorem 1.1, equations (1.6)-(1.7)). The proof develops a one-dimensional exchangeable-pair conditional CLT (Proposition 2.1), verifies the required linearity conditions for ERGMs (Proposition 2.2), and supplies supporting tools: an n^{-1} bound between the mean edge density and the subcritical fixed point (Proposition 1.1), a local CLT for edge counts with rate n^{-9/8} (Proposition 1.2), and higher-order concentration inequalities in the subcritical region (Lemmas 2.2-2.4). A conjectured extension to general subgraph counts is stated in Conjecture 1.1 with computations sketched in Appendix A.","tokens_in":34723,"tokens_out":31269,"duration_ms":294841,"significance":"If the statement gap discussed below is fixed, this is a substantial contribution. The paper generalizes the known conditional CLT for two-star counts in Erdős-Rényi graphs to ERGMs in the subcritical region, with explicit mean and variance formulas that can be used in statistical testing (Remark 1.2). The technical machinery is significant in its own right: Proposition 2.1 provides a streamlined exchangeable-pair conditional CLT, Proposition 1.1 strengthens the mean-density estimate to O(n^{-1}) and addresses a conjecture of Winstein, and Proposition 1.2 establishes a local CLT with explicit polynomial rate. The higher-order concentration inequalities are extended from the Dobrushin uniqueness region to the subcritical region via a Poincaré inequality, which is a nontrivial strengthening. The proof is detailed and largely coherent, and the conjectured general-subgraph CLT is clearly motivated. The main obstruction to the theorem as stated is the ill-posed conditioning for non-achievable edge densities, which appears to be fixable without changing the proof.","major_comments":[{"comment":"The statement of Theorem 1.1 quantifies over every ep_n in (0,1), but the conditioning event {En/N = ep_n} is empty unless N ep_n is an integer, because En is integer-valued. In that case the conditional law of Vn is undefined and the Wasserstein bound (1.5) is not a meaningful statement. The proof applies Proposition 2.1 with k=0, that is, it conditions on eE = 0 (equation (2.17)), and Proposition 1.2 only guarantees positive point mass at integer edge counts; hence the proof requires N ep_n to be an integer. This is a statement-level gap, not a defect in the exchangeable-pair argument: the proof appears to go through unchanged if one adds the assumption that N ep_n is an integer, or reformulates the theorem for an integer sequence k_n with k_n/N = ep_n and replaces ep_n by k_n/N in (1.6)-(1.7). The theorem should be corrected accordingly.","section":"Theorem 1.1 and Section 2.2"}],"minor_comments":[{"comment":"The displayed equality for sigma_V^2 in (2.63) is only asymptotically valid as written. The derivation leading to it produces the leading term N n ep^2 q^2 / D^2 plus an O(n^2) correction that is relatively O(1/n); the final formula (1.7) is correct to the precision needed for the n^{-1/2+epsilon} bound, but the exact equality in (2.63) should be replaced by an asymptotic equality or a leading-order statement.","section":"Section 2.2, equation (2.63)"},{"comment":"The proof states that P(eE = 0) = 1/sqrt(2 pi sigma_n^2) + O(n^{-9/8}), but by Proposition 1.1 the point N ep can be O(n) away from the mean mu_n = N p, so the leading point mass is 1/sqrt(2 pi sigma_n^2) times a factor exp(-t^2/2) with t = O(1). The proof only needs P(eE = 0) >= c/n and r_0 = 1 + O(n^{-1/8}), both of which remain true; this line should be corrected.","section":"Proof of Theorem 1.1, after Proposition 1.2"},{"comment":"The theorem should state explicitly that the parameter tuple (beta_{1n}, beta_{2n}, ..., beta_{mn}) is assumed to lie in the compact subset B. The proof at the start of Section 2.2 relies on this, and the constants in Propositions 1.1 and 1.2 are only uniform over B.","section":"Theorem 1.1, opening paragraph"},{"comment":"There is a duplicated 'and' in the sentence following equation (1.15): 'and and eDelta :=' should be 'and eDelta :='.","section":"Conjecture 1.1"},{"comment":"The word 'Wassertein' in the definition of d_W should be spelled 'Wasserstein'.","section":"Introduction, page 1"}],"recommendation":"major_revision","confidential_remarks":"The main theorem's quantification over arbitrary ep_n in (0,1) is the only serious obstruction I see; it appears to be a drafting issue rather than a mathematical one, since the proof machinery works for achievable integer edge counts. The authors should be asked to restate Theorem 1.1 with an integrality condition or with an integer edge-count sequence. If they do so, the paper is likely suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the theorem is new and the machinery is credible, but the statement as written conditions on an event that can be empty for many of the ep_n it quantifies over. That is a genuine statement-level bug, not a proof gap, and it looks easily fixed.\n\nWhat is actually new: Theorem 1.1 gives a Wasserstein-rate conditional CLT for the two-star count given edge density, with explicit mean and variance depending on the conditioning value. For G(n,p) this is known, but for subcritical ERGMs it is new; the unconditional two-star CLT (Mukherjee-Xu) does not imply it. The proof builds a modified exchangeable-pair conditional CLT, verifies an approximate linearity condition for the two-star count, and supplies a local CLT for edge counts plus higher-order concentration inequalities. That is a serious piece of work, and the conjectured extension to general subgraph counts is clearly labeled as conjecture.\n\nWhere the soft spots are. The conditioning statement is not well-posed as written: Theorem 1.1 quantifies over every ep_n in (0,1), but {En/N = ep_n} is empty whenever N ep_n is not an integer, because En is integer-valued. The proof actually works with {eE = 0} = {En = N ep_n} and uses the local CLT to show that event has positive probability. So the bound (1.5) is meaningful for achievable edge counts, and vacuous otherwise. This is exactly the kind of gap that a referee should catch but a revision can fix: add an integrality assumption on N ep_n, or restate with an integer sequence k_n within 1 of N ep_n. Once that is done, the proof appears to go through unchanged.\n\nMy confidence is moderate rather than high because the proof is long and some technical lemmas (Lemmas 2.2-2.4) are sketched, and the paper leans on results from the same research group, notably Fang et al. (2025) for the edge-count CLT. That is legitimate - the prior work is independent - but it means a careful referee should spot-check the borrowed bounds rather than take them on faith.\n\nBottom line: worth refereeing, and likely acceptable after the conditioning fix and some expansion of the skipped steps. I would cite it once the statement is corrected.","headline":"A genuinely new conditional CLT for two-star counts in subcritical ERGMs, with a real but fixable well-posedness gap in how the conditioning event is stated.","tokens_in":35285,"tokens_out":2045,"would_cite":true,"duration_ms":20522,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"Conditioning on the edge count makes two-star counts in exponential random graphs asymptotically normal, with an explicit, computable mean and variance and a Wasserstein error of order $n^{-1/2+\\varepsilon}$ in the subcritical parameter…","keywords":["conditional central limit theorem","exponential random graph model","two-star counts","exchangeable pairs","local limit theorem","higher-order concentration inequalities","subcritical region","subgraph counts"],"falsifier":"Take $n=4$ so $N=6$, and choose $\\widehat p_n=1/4$; then $N\\widehat p_n=1.5$ is not an integer, the event $\\{E_n/N=1/4\\}$ is empty, and the conditional distribution appearing in Theorem 1.1 is undefined, so the theorem's quantified statement cannot hold literally for that density unless the authors supply an explicit conditional extension. A concrete check is to compute the left-hand side under any proposed limiting or regularized definition of conditioning for such non-integer densities and compare it with the bound $C_\\varepsilon n^{-1/2+\\varepsilon}$.","tokens_in":34355,"feed_emoji":"📊","tokens_out":8268,"duration_ms":77657,"temperature":0.7,"pith_summary":"Exponential random graph models assign probabilities to networks by rewarding small subgraph patterns, and they are widely used in social-network analysis. This paper asks: if you condition on the exact number of edges, what is the distribution of the number of two stars (a vertex with two incident edges)? For the subcritical parameter region—where a mean-field fixed-point equation has a unique attracting solution—the answer is a normal law with explicit mean and variance. The result extends the classical conditional central limit theorem for independent-edge random graphs to models with dependent edges, at the cost of an arbitrarily small exponent loss in the error rate. Because the approximating mean and variance are explicit functions of the conditioning edge density, the theorem can be used to test whether an observed network with a given edge count is explained by a candidate exponential random graph model.","feed_headline":"Two-star counts go Gaussian once edges are fixed","feed_subtitle":"Subcritical network models: two-star counts are normal with explicit moments, enabling tests of edge-conditioned ERGMs.","key_machinery":"The load-bearing device is an exchangeable pair $(G,G')$ formed by resampling one randomly chosen edge according to its conditional law under the model. The paper verifies an approximate linearity condition: split according to whether the edge count increases or decreases by one, the expected increment of the two-star count, given the edge count and the two-star count, is approximately $-λ$ times the standardized two-star count plus controlled remainder terms. This linearity condition is combined with a local central limit theorem for the edge count (so the conditioning event $E_n=0$ has probability of order $n^{-1}$) and with new higher-order concentration inequalities that control the remainders in $L^r$ norms. A general proposition for conditional central limit theorems via exchangeable pairs converts these ingredients into the Wasserstein bound, and the explicit mean and variance in the theorem emerge as by-products of the coefficients in the linearity condition.","core_discovery":"Let $G_n$ follow the exponential random graph model with homomorphism densities for $H_1$ equal to an edge, $H_2,\\dots,H_m$ as additional small patterns, and parameters lying in a compact subset of the subcritical region (the set where the fixed-point equation $\\varphi_\\beta(a)=a$ has a unique solution $p$ with $\\varphi'_\\beta(p)<1$). Let $E_n$ be the edge count, $N=n(n-1)/2$, and let $V_n$ be the number of two stars. The paper proves that for every $\\varepsilon>0$ there is a constant $C_\\varepsilon$ such that, conditionally on $E_n/N=\\widehat p_n$, the Wasserstein distance between $(V_n-\\mu_{V_n})/\\sigma_{V_n}$ and a standard normal variable is at most $C_\\varepsilon n^{-1/2+\\varepsilon}$. The centering and scaling are explicit: $\\mu_{V_n}=N(n-2)\\widehat p_n^2+2N(1-\\widehat p_n)^2\\sum_{l=2}^m\\beta_{ln}s_l\\widehat p_n^{e_l}\\big/(1-2(1-\\widehat p_n)\\sum_{l=2}^m\\beta_{ln}s_l\\widehat p_n^{e_l-1})$ and $\\sigma_{V_n}^2=Nn\\widehat p_n^2(1-\\widehat p_n)^2\\big/(1-2(1-\\widehat p_n)\\sum_{l=2}^m\\beta_{ln}s_l\\widehat p_n^{e_l-1})^2$, where $e_l$ and $s_l$ are the numbers of edges and two stars in $H_l$. The proof routes through a new conditional central limit theorem for exchangeable pairs, a local central limit theorem for edge counts, and new higher-order concentration inequalities for subgraph counts.","pith_inferences":["A natural next test is to run the same exchangeable-pair linearity check for other sufficient-statistic conditionings, such as degree sequences or triangle counts, where the local-CLT and concentration inputs are the main prerequisites; the paper's machinery gives a template rather than a proof.","The positivity of the conditional variance $\\sigma_{V_n}^2$ at every solution with $1-\\varphi'_\\beta(p)\\ge 0$ suggests that the phase transition for edge-conditioned two-star counts, if it exists, may occur beyond the uniqueness threshold; a numerical study approaching criticality could check whether the normal approximation deteriorates only at the boundary.","The explicit $n^{-1}$ expansion of the difference between expected edge density and the fixed point, with its computed constant, could be used to construct higher-order approximations or Edgeworth corrections for edge counts, going beyond the leading central limit theorem."],"forward_implications":["In the subcritical region, conditioning on the edge density turns the two-star count into a Gaussian statistic whose mean and variance are known functions of that density and the model parameters, so the result gives the asymptotic distribution under the null for network-model tests using two-star counts.","The explicit normalization depends only on the observed edge density $E_n/N$ and the model parameters, not on the asymptotic edge density, so the theorem applies when only the realized edge count is available.","The sharp $n^{-1}$ bound between expected edge density and the mean-field fixed point, proved as Proposition 1.1, strengthens the earlier $n^{-1/2}$ bound and resolves a stated conjecture in the subcritical region.","The proof supplies a local central limit theorem for the edge count with rate $n^{-9/8}$, which is what turns conditioning on the exact edge-count event into tractable probabilities.","The approach extends to general subgraph counts; the paper gives explicit conjectured mean and variance formulas for such counts in terms of the conditional means and variances of two-star and triangle counts."],"supporting_citations":[{"why":"Supplies the exchangeable-pair conditional central limit theorem framework that the paper streamlines into Proposition 2.1.","marker":"Dey and Terlov (2023)"},{"why":"Supplies the edge-count normal approximation and Kolmogorov bound that Proposition 1.2 sharpens into a local central limit theorem.","marker":"Fang et al. (2025)"},{"why":"Supplies the Glauber-dynamics Stein method that Proposition 1.1 adapts to obtain the sharp $n^{-1}$ edge-density bound.","marker":"Reinert and Ross (2019)"},{"why":"Supplies uniformity of estimates in compact subcritical regions that the proof relies on for constants and error bounds.","marker":"Bhamidi et al. (2011)"},{"why":"Supplies the subcritical concentration, weak-dependence, and variance-inequality estimates used for the higher-order concentration and local-CLT arguments.","marker":"Ganguly and Nam (2024)"},{"why":"Supplies the Hoeffding decomposition and higher-order concentration toolbox that Lemma 4.1 extends to the subcritical region via a Poincaré-type inequality.","marker":"Sambale and Sinulis (2020)"},{"why":"Supplies the local-limit-theorem framework used to convert the Kolmogorov bound into the $n^{-9/8}$ local CLT for edge counts.","marker":"Röllin and Ross (2015)"},{"why":"Supplies the two-star-count testing baseline and the independent-edge conditional CLT that Theorem 1.1 generalizes.","marker":"Bresler and Nagaraj (2018)"},{"why":"Supplies the positive-association covariance bound used to estimate the variance of edge-flip probabilities in Lemma 3.3.","marker":"Newman (1980)"}],"fun_headline_variants":["Edge-conditioned ERGMs: two-star counts hit normal limit","Subcritical regime: conditional CLT for two-star counts","Two-star counts become Gaussian under fixed edge count","Explicit Gaussian limit for two-stars in edge-fixed ERGMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem is stated for any edge density $\\widehat p_n\\in(0,1)$, but the event $\\{E_n/N=\\widehat p_n\\}$ is empty unless $N\\widehat p_n$ is an integer, so the conditional law is undefined for non-achievable densities; without an achievability restriction or a smoothed conditional definition, the quantified statement is only meaningful for integer multiples $N\\widehat p_n$.","fun_headline_variants_meta":{"raw":{"variants":["Edge-conditioned ERGMs: two-star counts hit normal limit","Subcritical regime: conditional CLT for two-star counts","Two-star counts become Gaussian under fixed edge count","Explicit Gaussian limit for two-stars in edge-fixed ERGMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000503,"raw_usage":{"total_tokens":2523,"prompt_tokens":1079,"completion_tokens":1444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":1373}},"tokens_in":695,"tokens_out":1444,"duration_ms":10912,"temperature":1.0,"reasoning_tokens":1373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:42:04.905328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $n=4$ so $N=6$, and choose $\\widehat p_n=1/4$; then $N\\widehat p_n=1.5$ is not an integer, the event $\\{E_n/N=1/4\\}$ is empty, and the conditional distribution appearing in Theorem 1.1 is undefined, so the theorem's quantified statement cannot hold literally for that density unless the authors supply an explicit conditional extension. A concrete check is to compute the left-hand side under any proposed limiting or regularized definition of conditioning for such non-integer densities and compare it with the bound $C_\\varepsilon n^{-1/2+\\varepsilon}$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the exchangeable-pair conditional central limit theorem framework that the paper streamlines into Proposition 2.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies uniformity of estimates in compact subcritical regions that the proof relies on for constants and error bounds."},{"cited_title":"and Nam, K","cited_arxiv_id":null,"evidence_quote":"Supplies the subcritical concentration, weak-dependence, and variance-inequality estimates used for the higher-order concentration and local-CLT arguments."},{"cited_title":"and Nagaraj, D","cited_arxiv_id":null,"evidence_quote":"Supplies the two-star-count testing baseline and the independent-edge conditional CLT that Theorem 1.1 generalizes."}],"review_version":1}