Pith. sign in

REVIEW 1 major objections 5 minor 31 references

Conditional central limit theorems for exponential random graphs

T0 review · 1 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2506.15159 v1 pith:2XHZDFX4 submitted 2025-06-18 math.PR

Xiao Fang , Song-Hao Liu , Zhonggen Su , Xiaolin Wang This is my paper · ORCID
classification math.PR MSC 60F0505C80
keywords conditionalcentrallimittheoremexponentialrandomgraphmodeltwo-starcountsexchangeablepairslocalhigher-orderconcentrationinequalitiessubcriticalregionsubgraph
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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 $C_\varepsilon n^{-1/2+\varepsilon}$.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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 $N\widehat p_n$.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 5 minor

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.

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 (1)
  1. [Theorem 1.1 and Section 2.2] 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.
minor comments (5)
  1. [Section 2.2, equation (2.63)] 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.
  2. [Proof of Theorem 1.1, after Proposition 1.2] 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.
  3. [Theorem 1.1, opening paragraph] 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.
  4. [Conjecture 1.1] There is a duplicated 'and' in the sentence following equation (1.15): 'and and eDelta :=' should be 'and eDelta :='.
  5. [Introduction, page 1] The word 'Wassertein' in the definition of d_W should be spelled 'Wasserstein'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the conditional two-star CLT is derived from exchangeable-pair linearity conditions and independent local-limit ingredients, not from fitting or renaming its target.

full rationale

The derivation chain is self-contained and non-circular. The centering and variance in Theorem 1.1 are not fitted parameters: they are solved from exchangeable-pair identities. Specifically, the conditional mean E(eV) is obtained by taking expectations in the identity for M1,+ + M1,- and using antisymmetry (equations (2.54)-(2.55)), and the variance is obtained by matching M2,± to the required Stein-type form, yielding (2.63), which is then identified with (1.7). These are proof by-products with explicit formulas, not data-fit values. The only self-citation that is genuinely load-bearing is the edge-count Kolmogorov bound (1.12) from Fang et al. (2025), used in the proof of Proposition 1.2. That cited result is a distinct unconditional edge CLT with its own stated assumptions; it does not assert or assume the present conditional two-star CLT, so using it as an ingredient is legitimate independent support rather than circularity. The proof of Proposition 1.2 also refers to equation (2.21) from the main proof, but this is a forward reference to a Glauber-dynamics computation that does not depend on Proposition 1.2, so it creates no circular loop. The statement-level issue noted in the reader's take—that {En/N = ep_n} is empty unless N ep_n is an integer—is a well-posedness/correctness gap in the theorem's quantification over arbitrary ep_n, not a circularity. The proof itself conditions on eE = 0 at an achievable integer edge count and uses the local CLT to control P(eE = 0); for non-achievable densities the statement is vacuous or undefined. This is a mathematical rigor issue, not a self-referential reduction, and therefore does not affect the circularity score.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claim has no fitted free parameters: β and ep_n are inputs, and p is the unique fixed point of (1.3). The proof relies on prior results on subcritical ERGMs as axioms, none of which assume the conditional CLT being proved. No new entities are postulated.

assumptions (6)
  • domain assumption Unique subcritical fixed point p solving φβ(p)=p with φ'_β(p)<1 exists, as in (1.3).
    Defines the regime used throughout the paper; also ensures denominators and variances are bounded away from zero.
  • domain assumption The ERGM parameters β_2,...,β_m are positive, with H_1 an edge.
    Positivity is used to obtain positive association of edge indicators in the covariance estimates of Lemma 3.3.
  • domain assumption Ganguly and Nam (2024) estimates: conditional edge probabilities differ from unconditional by O(n^{-1}) and product expectations factorize up to O(n^{-1}) in the subcritical region, quoted as (2.31)-(2.32).
    This is the workhorse for centering ∂_ij H and for product-moment bounds in the proof of Proposition 2.2.
  • domain assumption Sambale and Sinulis (2020) Hoeffding-type decomposition holds in the Dobrushin uniqueness region, and Lemma 4.1 extends it to the subcritical region via the Poincaré inequality.
    Used to decompose ∂_ij H into dominant and remainder terms and to prove Lemmas 2.2-2.4.
  • domain assumption Fang et al. (2025): the edge count satisfies a Kolmogorov-distance normal approximation with rate n^{-1/4} in the subcritical region, stated in (1.12).
    Input to Proposition 1.2, which strengthens the result to a local CLT with rate n^{-9/8}.
  • domain assumption Reinert and Ross (2019): |p - E(E_n)/N| ≤ C n^{-1/2} in the subcritical region.
    Used as the starting bound in the bootstrap argument for Proposition 1.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional central limit theorems for exponential random graphs." pith.science (2026). https://pith.science/paper/2XHZDFX4

@misc{pith2026250615159,
  author       = {Pith},
  title        = {Pith review of: Conditional central limit theorems for exponential random graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XHZDFX4}},
  note         = {Machine review of arXiv:2506.15159}
}
read the original abstract

In this paper, we study the Exponential Random Graph Models (ERGMs) conditioning on the number of edges. In subcritical region of model parameters, we prove a conditional Central Limit Theorem (CLT) with explicit mean and variance for the number of two stars. This generalizes the corresponding result in the literature for the Erd\H{o}s--R\'enyi random graph. To prove our main result, we develop a new conditional CLT via exchangeable pairs based on the ideas of Dey and Terlov. Our key technical contributions in the application to ERGMs include establishing a linearity condition for an exchangeable pair involving two star counts, a local CLT for edge counts, as well as new higher-order concentration inequalities. Our approach also works for general subgraph counts, and we give a conjectured form of their conditional CLT.

Figures

Figures reproduced from arXiv: 2506.15159 by the authors.

Figure 1
Figure 1. Two graphs in ∂ijH¯A numerator of ∂ijH¯A, we obtain by (4.2) that ∂ijH¯A = Or( √ n/n) = Or(n −1/2 ). (4.5) Similarly, we collect those terms involving two additional vertices besides i and j to form ∂ijH¯B as in (2.37) and its order is ∂ijH¯B = Or(n/n2 ) = Or(n −1 ). The remaining terms are collected to be ∂ijH¯C. The sum of those terms involving three or more additional vertices besides i and j is of order Or(n −3/… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [1]

    Bauerschmidt, R., Bodineau, T., and Dagallier, B. (2024). Kawasaki dynamics beyond the uniqueness threshold. Probability Theory and Related Fields , pages 1--36

    work page 2024

  2. [2]

    Bhamidi, S., Bresler, G., and Sly, A. (2011). Mixing time of exponential random graphs. The Annals of Applied Probability , 21(6):2146--2170

    work page 2011

  3. [3]

    and Nagaraj, D

    Bresler, G. and Nagaraj, D. (2018). Optimal single sample tests for structured versus unstructured network data. In Conference On Learning Theory , pages 1657--1690. PMLR

    work page 2018

  4. [4]

    and Dembo, A

    Chatterjee, S. and Dembo, A. (2016). Nonlinear large deviations. Advances in Mathematics , 299:396--450

    work page 2016

  5. [5]

    and Diaconis, P

    Chatterjee, S. and Diaconis, P. (2013). Estimating and understanding exponential random graph models. The Annals of Statistics , 41(5):2428--2461

    work page 2013

  6. [6]

    Chen, L. H. Y., Goldstein, L., and Shao, Q.-M. (2010). Normal Approximation by Stein’s Method . Springer Science & Business Media

    work page 2010

  7. [7]

    Cook, N. A. and Dembo, A. (2024). Typical structure of sparse exponential random graph models . The Annals of Applied Probability , 34(3):2885 -- 2939

    work page 2024

  8. [8]

    Dey, P. S. and Terlov, G. (2023). Stein’s method for conditional central limit theorem. The Annals of Probability , 51(2):723--773

    work page 2023

Show all 31 references
  1. [9]

    and Fang, X

    Ding, W.-Y. and Fang, X. (2024). Second-order approximation of exponential random graph models. arXiv preprint arXiv:2401.01467

    2024 arXiv

  2. [10]

    Fang, X., Liu, S.-H., Shao, Q.-M., and Zhao, Y.-K. (2025). Normal approximation for exponential random graphs. To appear in Probability Theory and Related Fields, arXiv preprint arXiv:2404.01666v2

    2025 arXiv

  3. [11]

    and Strauss, D

    Frank, O. and Strauss, D. (1986). Markov graphs. Journal of the American Statistical Association , 81(395):832--842

    1986

  4. [12]

    and Nam, K

    Ganguly, S. and Nam, K. (2024). Sub-critical exponential random graphs: concentration of measure and some applications. Transactions of the American Mathematical Society , 377(04):2261--2296

    2024

  5. [13]

    and Wiroonsri, N

    Goldstein, L. and Wiroonsri, N. (2018). Stein’s method for positively associated random variables with applications to the I sing and voter models, bond percolation, and contact process. Ann. Inst. H. Poincar\'e Probab. Statist. , 54:385--421

    2018

  6. [14]

    and Sambale, H

    G \"o tze, F. and Sambale, H. (2019). Higher order concentration in presence of P oincar \'e -type inequalities. In High Dimensional Probability VIII: The Oaxaca Volume , pages 55--69. Springer

    2019

  7. [15]

    G \"o tze, F., Sambale, H., and Sinulis, A. (2019). Higher order concentration for functions of weakly dependent random variables. Electronic Journal of Probability , 24:1--19

    2019

  8. [16]

    Holland, P. W. and Leinhardt, S. (1981). An exponential family of probability distributions for directed graphs. Journal of the American Statistical Association , 76(373):33--50

    1981

  9. [17]

    Janson, S. (1990). A functional limit theorem for random graphs with applications to subgraph count statistics. Random Structures & Algorithms , 1:15--37

    1990

  10. [18]

    and Xu, Y

    Mukherjee, S. and Xu, Y. (2023). Statistics of the two star ERGM . Bernoulli , 29(1):24--51

    2023

  11. [19]

    Newman, C. M. (1980). Normal fluctuations and the FKG inequalities. Communications in Mathematical Physics , 74(2):119--128

    1980

  12. [20]

    and Ross, N

    Reinert, G. and Ross, N. (2019). Approximating stationary distributions of fast mixing G lauber dynamics, with applications to exponential random graphs. The Annals of Applied Probability , 29(5):3201--3229

    2019

  13. [21]

    and Ross, N

    R \"o llin, A. and Ross, N. (2015). Local limit theorems via L andau-- K olmogorov inequalities. Bernoulli

    2015

  14. [22]

    and Sinulis, A

    Sambale, H. and Sinulis, A. (2020). Logarithmic S obolev inequalities for finite spin systems and applications. Bernoulli , 26(3):1863--1890

    2020

  15. [23]

    A., Pattison, P

    Snijders, T. A., Pattison, P. E., Robins, G. L., and Handcock, M. S. (2006). New specifications for exponential random graph models. Sociological methodology , 36(1):99--153

    2006

  16. [24]

    Stein, C. (1986). Approximate C omputation of E xpectations. IMS

    1986

  17. [25]

    and Faust, K

    Wasserman, S. and Faust, K. (1994). Social Network Analysis: Methods and Applications . Cambridge University Press

    1994

  18. [26]

    Winstein, V. (2025). Concentration via metastable mixing, with applications to the supercritical exponential random graph model. arXiv preprint arXiv:2503.07571

    2025

  19. [27]

    Local limit theorems via Landau–Kolmogorov inequalities[J]

    Röllin A, Ross N. Local limit theorems via Landau–Kolmogorov inequalities[J]. 2015

    2015

  20. [28]

    Approximating stationary distributions of fast mixing Glauber dynamics, with applications to exponential random graphs[J]

    Reinert G, Ross N. Approximating stationary distributions of fast mixing Glauber dynamics, with applications to exponential random graphs[J]. 2019

    2019

  21. [29]

    Normal fluctuations and the FKG inequalities[J]

    Newman C M. Normal fluctuations and the FKG inequalities[J]. Communications in Mathematical Physics, 1980, 74(2): 119-128

    1980

  22. [30]

    Sub-critical Exponential random graphs: Concentration of measure and some applications[J]

    Ganguly S, Nam K. Sub-critical Exponential random graphs: Concentration of measure and some applications[J]. arXiv preprint arXiv:1909.11080, 2019

    1909 arXiv

  23. [31]

    Computations for conj1 sec:multi Step 1

    Sambale H, Sinulis A .Logarithmic Sobolev inequalities for finite spin systems and applications[J].Bernoulli, 2020(3): 1863-1890. Computations for conj1 sec:multi Step 1. We first sketch a computation leading to a multivariate conditional CLT for the joint distribution of cent...

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.