{"id":"7ce3bf10-95b8-49b5-bdfe-9fd4707eb95d","arxiv_id":"2505.10373","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fitness-induced two-star exponential random graph model, fit2SM, reproduces the first two moments of empirical degree distributions using only two global parameters.","lead":"This paper introduces a two-parameter network model, the fit2SM, that matches both the total number of links and the total number of two-stars in a network, which makes it reproduce the mean and variance of the degree distribution. It offers a way to generate realistic random networks from just two global quantities, with tests on interbank transaction data showing good accuracy on sparse networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The variance-reproduction claim rests on dropping the 4Var[L]/N^2 term in Eq. (6) without quantifying it; this term is likely small on the tested eMID snapshots but must be measured before the exact claim is supported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the derivation replaces Eq. (6) with Eq. (12) by dropping 4Var[L]/N^2 without quantifying it. I agree that this is the point on which the central variance-reproduction claim depends. However, a direct bound shows the omitted term is at most 4L/N^2, and on the reported sparse eMID snapshots this is typically below 1% of Var[k], so the concern is about unverified exactness rather than a demonstrated failure. The self-consistency issue raised by the reader is real but less load-bearing: Appendix D compares single-iteration and self-consistent solutions for BIC and reports relative differences below 0.002, which suggests the practical impact is small, although it does not directly test variance or spectral radius. The proprietary data limit independent reproduction, but code is available. Overall, the model is plausible and the empirical evidence supports conditional acceptance; the main missing piece is a quantitative check of the dropped term. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":19123,"tokens_out":10990,"duration_ms":118280,"concrete_test":"Using the released fit2SM code, compute for every eMID snapshot the exact expected sample variance from Eq. (6), evaluating Var[L] = Σ_{i<j} p_ij(1−p_ij) at the reported fitted probabilities, and compare it with the empirical Var[k] from Eq. (5). If the maximum relative deviation |⟨Var[k]⟩ − Var[k]| / Var[k] across all snapshots exceeds 1%, the headline claim should be qualified. The same check should be repeated on synthetic networks with small degree variance, where the bound 4L/N^2 is not automatically negligible relative to Var[k].","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (6) is exact for any dyad-independent model: ⟨Var[k]⟩ = 2⟨S⟩/N − 4Var[L]/N^2 + 2⟨L⟩/N(1 − 2⟨L⟩/N). The fit2SM constraints (15) fix only ⟨L⟩ and ⟨S⟩, so the model's expected sample variance equals the empirical Var[k] only up to the omitted term −4Var[L]/N^2. Section V dismisses this addendum as 'may be expected not to play a relevant role', but no numerical estimate of Var[L] is reported for any snapshot. Since the paper's central claim is that the fit2SM reproduces the second moment of the degree distribution, this unquantified approximation is the weakest point of the argument. A simple bound shows Var[L] = Σ_{i<j} p_ij(1−p_ij) ≤ ⟨L⟩ = L, so the omitted term is at most 4L/N^2. For the sparse eMID daily and weekly snapshots this is likely only a small fraction of Var[k], which suggests the empirical conclusions may survive, but the exactness of the variance-reproduction claim is not established. The omission is also not a purely empirical issue: for networks with low degree variance or denser topologies, the relative error 4L/(N^2 Var[k]) can become non-negligible. The paper should either compute the term directly from the fitted probabilities or explicitly state that variance reproduction holds only up to this O(4L/N^2) correction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a new exponential random graph model, the fitness-induced two-star model (fit2SM), which is designed to reproduce the first and second moments of an empirical degree distribution while remaining in the canonical (soft-constrained) framework. After deriving an exact relationship between the sample variance of the degree distribution and the number of links L and two-stars S, the authors argue that the problematic term involving Var[L] in the expected variance may be neglected, arriving at a two-parameter model whose probabilities are z s_i s_j y^{kappa_i+kappa_j}/(1+z s_i s_j y^{kappa_i+kappa_j}), with z and y fixed by the constraints <L>=L* and <S>=S*. The model is tested on the eMID interbank network at five temporal aggregations, with claims that it reproduces the degree variance, improves spectral-radius estimation at daily and weekly scales, and outperforms the UBCM and dcGM under BIC on sparse snapshots. The manuscript also contains a negative result on the mean-field degree-corrected two-star model, which degenerates when both degree and two-star constraints are imposed.","tokens_in":19428,"tokens_out":4201,"duration_ms":47497,"significance":"If the central claim is established, the fit2SM is a genuinely useful minimal model: it reproduces the first two moments of the degree distribution with only two global parameters and arbitrary node fitnesses, it is fast to solve, and it appears to improve on both the UBCM and the dcGM for sparse snapshots. The algebraic derivation connecting the degree variance to L and S is clean, the numerical solver reproduces the target constraints to high precision (Table I), and the code is publicly available. The spectral-radius and BIC comparisons are also welcome because they test consequences of the second moment rather than merely re-fitting it. However, the headline variance-reproduction claim is only approximate as stated, and the approximation is never quantified; this limits the significance of the paper unless the authors either measure the omitted term or explicitly reframe the claim as an approximation with a bounded error.","major_comments":[{"comment":"The central claim that the fit2SM reproduces the sample variance is exact only if the term -4Var[L]/N^2 is negligible, but this is asserted rather than demonstrated. For any dyad-independent model, Eq. (6) gives <Var[k]> = 2<S>/N - 4Var[L]/N^2 + 2<L>/N(1 - 2<L>/N). Since the constraints in Eq. (15) fix only <L> and <S>, the model's expected variance differs from the empirical sample variance by -4Var[L]/N^2. The sentence 'such an addendum may, thus, be expected not to play a relevant role' is not a quantitative argument. A simple bound is Var[L] = sum_{i<j} p_ij(1-p_ij) <= <L> = L, so the omitted term is at most 4L/N^2; for the tested sparse eMID snapshots this may indeed be small, but the paper never reports its value. I request that the authors compute Var[L] from the fitted probabilities for the reported snapshots, and either state the resulting relative error explicitly or rewrite the claim as holding up to an O(4L/N^2) correction.","section":"Section V, Eqs. (6)-(12)"},{"comment":"The headline empirical support for variance reproduction consists of only two weekly snapshots in Fig. 1. Given that <Var[k]> is determined, up to the omitted Var[L] term, by the two constraints in Eq. (15), agreement in two cases is more a consistency check of the approximation than an independent model test. The manuscript should systematically report, for all snapshots and temporal aggregations, the empirical Var[k] versus the model's <Var[k]>, together with the estimated omitted term 4Var[L]/N^2 or its relative contribution. Without this, the statement that the fit2SM 'correctly reproduces' the sample variance is not supported by the presented evidence.","section":"Fig. 1 and Section V"},{"comment":"The model as implemented uses kappa_i fixed at the dcGM values from the single-iteration initialization, as Eq. (26) and Appendix D make clear, so the reported probabilities are not the self-consistent solutions of the non-linear system advertised in Eqs. (14)-(17). The authors do compare single-iteration and self-consistent solutions for BIC in Appendix D, but not for the variance-reproduction or spectral-radius claims. Please clarify which version of the model is used for each reported result, and show that the single-iteration choice does not materially affect the variance-reproduction and spectral-radius conclusions, not only the BIC values.","section":"Eqs. (14), (17)-(18) and Appendix D"}],"minor_comments":[{"comment":"The bullet on the monthly time-scale contains an apparent contradiction: it says BIC_fit2SM < BIC_UBCM on 33.94% of monthly snapshots but then says BIC_fit2SM < BIC_UBCM on 97.78% of monthly snapshots since 2009. Please rephrase to distinguish the full-sample result from the post-2009 subsample.","section":"Section VI C"},{"comment":"The iteration indices are confusing: Eq. (16) uses t for the z and y updates, while Eq. (17) uses e for epochs; the relationship between these two iteration levels should be defined explicitly.","section":"Eqs. (16)-(17)"},{"comment":"The statement that numerical errors 'never exceed O(10^-1)' is imprecise given that Table I lists maximum relative errors around 10^-9 to 10^-11; please state the actual error ranges.","section":"Appendix D"},{"comment":"The main text reports the S=aL^b fit with a~0.51, b~1.58, while Appendix D reports aggregation-specific values such as a_daily~0.36, a_yearly~0.69; please reconcile the two presentations so that the reader understands whether the main-text values are averages of the aggregation-specific fits or a single pooled fit.","section":"Section VII and Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the core idea is publishable, but I would not accept it in the current form. The main issue is the unquantified neglect of Var[L] in the variance-reproduction claim; I would like to see the authors either compute the term for their snapshots or explicitly state the approximation with a bound. The single-iteration versus self-consistent resolution also needs to be clarified for the central claims, not just for BIC. The manuscript is otherwise careful and the code release is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper gives you a genuinely new model, fit2SM, with two global parameters that jointly reproduce the number of links L and the number of two-stars S, and thereby the degree variance to good approximation. It also contains a clean negative result: the mean-field degree-corrected two-star model degenerates, which is worth knowing. The numerical solver is fast and accurate (Table I), and the empirical evaluation on eMID across five aggregation levels is thorough: BIC, spectral radius, QQ plots, and isolated-node counts. They ship code, and they even check single-iteration versus self-consistent solutions (differences around 0.1% in BIC), which I appreciate.\n\nThe soft spots are real but not fatal. The central variance-reproduction claim rests on dropping the -4Var[L]/N^2 term in Eq. (6), with a hand-wave that it 'may be expected not to play a relevant role'. On the actual eMID snapshots it is in fact negligible: since Var[L] ≤ L, the term is ≤ 4L/N^2, which is about 0.08 on the sparsest daily snapshot and 0.8 on the densest yearly snapshot, versus degree variances of order 100 to 10,000. So the approximation is excellent, but they should quantify it instead of asserting it, and should soften the sentence that says the model 'ensures' the variance is reproduced. That is a one-paragraph fix.\n\nThe second soft spot is that the implemented model fixes κ at the dcGM values (single iteration), so the solved probabilities are not the fully self-consistent solution of the non-linear system. They test this and find negligible differences, so it is a minor issue, but it belongs in the main text rather than buried in an appendix.\n\nThird, the data are proprietary, so the empirical results cannot be independently reproduced. That is a limitation, not a defect, but it does mean the general claims rest on one dataset. The 'S = aL^b' scaling is also only tested on two systems.\n\nWho is this for? Network-reconstruction people, and anyone who needs degree variance for epidemic thresholds or systemic risk. It is a solid contribution with a clear scope and honest limitations. I would send it to referees.\n\nRecommendation: accept after minor revision, with the Var[L] bound added and the wording of the variance-reproduction claim made precise.","headline":"A useful two-parameter canonical ERG that reproduces degree variance on sparse financial networks; the main approximation is unquantified but actually tiny on their data.","tokens_in":20006,"tokens_out":3147,"would_cite":true,"duration_ms":33107,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.Fb","02.50.Tt"],"model":"deepseek-v4-flash","headline":"A two-parameter canonical model reproduces the sample variance of empirical degree distributions where standard linear exponential random graph models cannot.","keywords":["exponential random graphs","two-star model","degree variance","network reconstruction","canonical ensembles","fitness model","interbank networks","spectral radius"],"falsifier":"On a sparse eMID snapshot, sample many configurations from the fitted fit2SM, measure the variance of the total number of links directly, and compare $4\\,\\mathrm{Var}[L]/N^2$ with the observed degree variance; if the ratio is not negligible, the central variance-reproduction claim fails. Repeating the comparison with fully self-consistent $\\kappa$ values from Eq. (17) would check whether the single-iteration implementation masks the discrepancy.","tokens_in":18854,"feed_emoji":"📊","tokens_out":11309,"duration_ms":102740,"temperature":0.7,"pith_summary":"The paper sets out to show that a random-graph model with only two global parameters can reproduce both the mean and the sample variance of an empirical degree distribution, something standard linear exponential random graph models cannot do. The second moment matters because it controls epidemic thresholds, consensus times, and spectral early-warning indicators such as the spectral radius. The authors show that the natural alternatives fail: forcing both the degree sequence and the two-star count microcanonically makes the model deterministic, and the mean-field degree-corrected two-star model degenerates. Their proposed fit2SM instead adds a nonlinear two-star constraint on top of the strength-based Gravity Model structure, enforced on average; on the eMID interbank network it reproduces the degree variance, gives the best spectral-radius errors at daily and weekly scales, and wins model selection on sparse snapshots.","feed_headline":"Two parameters now reproduce a network's degree variance","feed_subtitle":"It also beats standard models at spectral-radius reconstruction on daily and weekly interbank snapshots.","key_machinery":"The load-bearing identity is $\\mathrm{Var}[k] = \\frac{2S}{N} + \\frac{2L}{N}(1-\\frac{2L}{N})$, which translates the variance of the degree distribution into the total number of links $L$ and the total number of two-stars $S$, where a two-star is a pair of links that share a common node. The mechanism that enforces it is the softened link probability $p_{ij} = \\frac{z s_i s_j y^{\\kappa_i+\\kappa_j}}{1 + z s_i s_j y^{\\kappa_i+\\kappa_j}}$: replacing observed degrees by expected degrees $\\kappa_i$ keeps the model canonical and node-heterogeneous while avoiding the degeneracy that makes the mean-field degree-corrected two-star model collapse to deterministic edges. The two global parameters $z$ and $y$ are then fixed by requiring the expected link count and expected two-star count to equal their empirical values.","core_discovery":"The central claim is that a deliberately minimal, canonical model -- the fitness-induced two-star model (fit2SM) -- reproduces both the first and second moments of empirical degree distributions while keeping the explanatory power of the density-corrected Gravity Model. Its link probability is $p_{ij} = \\frac{z s_i s_j y^{\\kappa_i+\\kappa_j}}{1 + z s_i s_j y^{\\kappa_i+\\kappa_j}}$, where $s_i$ are node strengths used as exogenous fitnesses, $\\kappa_i$ are expected degrees, and the two parameters $z,y$ solve the coupled moment equations $\\langle L\\rangle = L$ and $\\langle S\\rangle = S$. The variance identity $\\mathrm{Var}[k] = \\frac{2S}{N} + \\frac{2L}{N}(1-\\frac{2L}{N})$ shows why matching the two-star count $S$ in expectation is sufficient for matching the variance, up to the neglected $\\frac{4\\mathrm{Var}[L]}{N^2}$ fluctuation term. In tests on the eMID interbank market, the model reproduces the empirical degree variance in snapshots where the UBCM systematically overestimates it and the dcGM over- or underestimates it, and it yields the smallest errors in spectral-radius reconstruction on daily and weekly aggregations.","pith_inferences":["The same construction should extend to higher-order nonlinear constraints, such as a Strauss model constraining triangle counts, giving a three-parameter canonical model for clustering or higher moments; the paper only hints at this direction.","Because the implementation freezes $\\kappa$ at dcGM expected degrees, the reported probabilities solve a linearized system; testing the fully self-consistent equations would show whether variance reproduction survives exact iteration.","The fit2SM needs only strengths and two global parameters, so it is a candidate reconstruction tool for other sparse economic networks, such as trade, input-output, or payment systems, where counterparty links are unobserved but total activity per node is known."],"forward_implications":["On sparse snapshots the UBCM overestimates the degree variance and the dcGM can over- or underestimate it, while the fit2SM reproduces it, so variance-based diagnostics become reliable under a soft-constrained model.","At daily and weekly time scales the fit2SM reproduces the spectral radius more accurately than both benchmarks, with average absolute errors of 0.79 ± 0.51 and 0.97 ± 0.66.","Model selection favors the fit2SM over the UBCM and the dcGM on sparse daily and weekly snapshots, with the advantage shrinking on denser quarterly and yearly snapshots.","As a generative model, the fit2SM can replace the true network topology for computing spectral early-warning z-scores, including in the pre-crisis period.","Because the two-star count is empirically related to the link count by $S \\approx 0.51 L^{1.58}$, the variance constraint can be imposed even when the two-star count is not directly observable."],"supporting_citations":[{"why":"Supplies the eMID transaction data and the spectral-radius early-warning methodology used in the empirical tests.","marker":"[8]"},{"why":"Establishes the maximum-entropy exponential random graph framework and the mean-field approximation the paper builds on.","marker":"[10]"},{"why":"Gives the analytical maximum-likelihood machinery used to fit the UBCM and related canonical models.","marker":"[12]"},{"why":"Defines the density-corrected Gravity Model whose link probability the fit2SM generalizes and which serves as the main benchmark.","marker":"[27]"},{"why":"Provides the two-star model that the fit2SM is a fitness-induced variant of, including its homogeneous mean-field solution.","marker":"[39]"},{"why":"Supplies the fixed-point iteration scheme used to solve the coupled equations for z and y.","marker":"[40]"},{"why":"Defines the Chung-Lu model used as a controlled benchmark connecting spectral radius to the first two degree moments.","marker":"[41]"}],"fun_headline_variants":["Two-parameter model reproduces degree mean and variance","Fit2SM: two parameters match network degree variance","Two parameters now fit degree distribution's first two moments","Minimal canonical model captures degree variance exactly","Fitness-induced two-star model reproduces degree spread"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the fluctuation in the total number of links, divided by $N^2$, is negligible when computing the degree variance; if that fluctuation is not tiny, the fit2SM's predicted variance misses the observed one by exactly that amount.","fun_headline_variants_meta":{"raw":{"variants":["Two-parameter model reproduces degree mean and variance","Fit2SM: two parameters match network degree variance","Two parameters now fit degree distribution's first two moments","Minimal canonical model captures degree variance exactly","Fitness-induced two-star model reproduces degree spread"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000872,"raw_usage":{"total_tokens":3769,"prompt_tokens":934,"completion_tokens":2835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2771}},"tokens_in":550,"tokens_out":2835,"duration_ms":21283,"temperature":1.0,"reasoning_tokens":2771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:10:12.119342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a sparse eMID snapshot, sample many configurations from the fitted fit2SM, measure the variance of the total number of links directly, and compare $4\\,\\mathrm{Var}[L]/N^2$ with the observed degree variance; if the ratio is not negligible, the central variance-reproduction claim fails. Repeating the comparison with fully self-consistent $\\kappa$ values from Eq. (17) would check whether the single-iteration implementation masks the discrepancy.","supporting_citations":[{"cited_title":"& Squartini, T","cited_arxiv_id":null,"evidence_quote":"Supplies the eMID transaction data and the spectral-radius early-warning methodology used in the empirical tests."},{"cited_title":"& Newman, M","cited_arxiv_id":null,"evidence_quote":"Establishes the maximum-entropy exponential random graph framework and the mean-field approximation the paper builds on."},{"cited_title":"& Garlaschelli, D","cited_arxiv_id":null,"evidence_quote":"Gives the analytical maximum-likelihood machinery used to fit the UBCM and related canonical models."},{"cited_title":"& Gabrielli, A","cited_arxiv_id":null,"evidence_quote":"Defines the density-corrected Gravity Model whose link probability the fit2SM generalizes and which serves as the main benchmark."},{"cited_title":"& Newman, M","cited_arxiv_id":null,"evidence_quote":"Provides the two-star model that the fit2SM is a fitness-induced variant of, including its homogeneous mean-field solution."},{"cited_title":"J., Cimini, G","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-point iteration scheme used to solve the coupled equations for z and y."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Chung-Lu model used as a controlled benchmark connecting spectral radius to the first two degree moments."}],"review_version":1}