{"id":"6373cd8a-a2cf-4cb0-8375-848b169bd621","arxiv_id":"2507.21469","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A tetrad-based conditional likelihood estimator for reciprocity and covariate effects in directed networks is consistent, asymptotically normal, and minimax rate-optimal under sparsity.","lead":"This paper introduces a regression model for directed networks where the tendency of two nodes to form mutual ties depends on observable pair characteristics, and proposes a conditional-likelihood estimator that removes node-specific effects by comparing four-node subgraphs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform boundedness of node heterogeneity (Assumption 2) is load-bearing for every rate in the paper, so the advertised general accommodation of node heterogeneity is not established.","rationale":"The reader's weakest_assumption correctly identifies Assumption 2. The central claim of minimax rate optimality is formulated over the parameter class Theta_n in Theorem 4, which itself imposes max_i{|alpha_i|,|beta_i|} <= C3, and the upper-bound proofs in Theorem 2 rely on the same boundedness. Thus the claim is internally consistent for that class. The load-bearing issue is that the paper's abstract and introduction advertise general node heterogeneity ('accommodates node heterogeneity,' 'unrestricted node-level heterogeneity'), while the actual theorems require uniformly bounded heterogeneity. The tetrad-event probabilities P(S_ijkl=0) ~ n^{-4a} and P(S_ijkl=±1) ~ n^{-4a+2b}, which determine every convergence rate, the CLT normalization, and the variance estimator in Theorem 3, are only valid under Assumption 2; with heterogeneity growing as C log n, these rates shift by n^{O(C)} and the entire asymptotic argument loses uniform validity. This is not a mathematical contradiction in the stated theorems, but it narrows the proven scope relative to the advertised one, and it has direct practical consequences for confidence-interval coverage on heavy-tailed networks. The proposed simulation directly tests whether Theorem 3's inference is robust to a natural violation of Assumption 2; if coverage collapses, the limitation should be prominently stated and the advertised generality softened. The reader's verdict of CONDITIONAL is appropriate; no stronger adjustment is needed because the central minimax claim is correct under its stated assumptions. No other identified concern (e.g., the Theorem 3 normalization) withstands close reading, since the n-multiplier in Theorem 3 is consistent with the unnormalized H_n and Omega_n scales in the proof.","tokens_in":48986,"tokens_out":27196,"duration_ms":304369,"concrete_test":"Simulate the model with n=200, a=0.3, b=0.1, gamma=(1,1), and draw alpha_i,beta_i i.i.d. Uniform[-C log n, C log n] with C=0.5 (violating Assumption 2) against a control with C=0; construct 95% confidence intervals from Theorem 3 across 1000 datasets. If the C=0 control gives coverage near 0.95 but the C>0 setting undercovers substantially (e.g., below 0.80) or the standardized errors fail the QQ alignment, the bounded-heterogeneity assumption is genuinely load-bearing for the advertised inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 2 bounds max_i{|alpha_i|,|beta_i|} <= C, and this bound is used throughout the proof machinery: Lemma 2's derivative bounds, the event-rate estimates P(S_ijkl=0) ~ n^{-4a} and P(S_ijkl=±1) ~ n^{-4a+2b}, the covariance orders in Proposition 2, and the normalizations in Assumptions 4 and 6 that drive Theorem 2 and Theorem 3. If alpha_i,beta_i are allowed to grow with n, e.g., max_i|alpha_i| = C log n, then a single dyad probability p_ij(1,0) ~ exp(-a log n + alpha_i + beta_j) can be n^{-a+O(C)} rather than n^{-a}; products of four such terms in Lemma 1 shift by n^{O(C)}, so the claimed uniform rates are not uniform and the U-statistic variance decomposition collapses. The model does 'assign each node its own sender and receiver parameters,' but the asymptotic and minimax claims are only proved for moderate heterogeneity. The abstract and introduction present the framework as generally accommodating node heterogeneity, which overstates the proven scope and invites application to heavy-tailed networks where the stated rates and normality will not hold. This is a scope limitation rather than an internal inconsistency, but it is the least secure support for the central claim as advertised.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the R²-Model, a generalized p1 model for directed networks that incorporates reciprocity through a baseline parameter ρ_n and a covariate vector γ, while allowing node-specific sender and receiver heterogeneity. To avoid estimating 2n+2+q parameters, the authors propose a conditional likelihood estimator based on tetrad configurations that conditions out the degree sequence and isolates (ρ_n, γ). They prove consistency, asymptotic normality, and minimax rate optimality under sparsity assumptions, and support the results with simulations and two empirical applications. The proof strategy relies on U-statistic projections and a case-by-case analysis of tetrad dependence, with complete proofs in the supplementary material.","tokens_in":49193,"tokens_out":21774,"duration_ms":223729,"significance":"If the results hold, this is a useful and nontrivial extension of Graham's tetrad-based conditional likelihood approach to directed networks with reciprocity and covariates. The paper provides machine-checkable auxiliary material in the form of detailed proofs, and it gives a constructive proof of minimax optimality via Le Cam's two-point method, which is a strong selling point. The closed-form estimator for the p1 submodel and the real-data applications increase the paper's practical appeal. However, the theoretical guarantees are established only under a bounded-heterogeneity assumption, which is narrower than the paper's advertised 'unrestricted' heterogeneity, and the numerical studies lack a description of how the computationally heavy tetrad enumeration was performed.","major_comments":[{"comment":"The paper repeatedly claims to allow 'unrestricted node-level heterogeneity' (Section 1) and the abstract states the model 'accommodates both network sparsity and node heterogeneity,' but Assumption 2 requires max_i{|α_i|,|β_i|} ≤ C for a constant independent of n. This uniform boundedness is load-bearing: Lemma 1's event probabilities, Lemma 2's derivative bounds, Proposition 2's covariance orders, and the rates in Theorems 2 and 3 all use it. If max_i|α_i| were to grow with n, say as C log n, then the probabilities P(S_ijkl=0) ∼ n^{-4a} and P(S_ijkl=±1) ∼ n^{-4a+2b} would fail, and the U-statistic variance decomposition would collapse. The authors should either weaken the stated claims to 'moderate degree heterogeneity' or extend the theory to slowly growing heterogeneity with explicit rate conditions.","section":"Section 1, p.2; Section 2.2, Assumption 2"},{"comment":"The numerical study reports 1000 replications for n = 50, 100, 150, 200 without explaining how the estimator L_n, which sums over all C(n,4) tetrads, was computed. For n=200 there are approximately 6.5×10^7 tetrads per replication, making the reported 1000 replications (about 6.5×10^10 tetrad evaluations per setting) computationally infeasible with the described full enumeration. If a random subsample of tetrads or another approximation was used, that estimator is not covered by Theorems 1-3, and the finite-sample claims and confidence-interval coverages need to be justified for the implemented procedure. The paper should state the exact algorithm and, if subsampling is used, provide a corresponding theoretical statement.","section":"Section 3.1, simulations"},{"comment":"In the proof of Theorem 2, Case C, the text states 'ˆρ − ρ0 = ... = OP(1)' after having shown n^{4a}Ψ_{n,ρ}(ϑ0) = OP(n^{(-2+a)/2}) and that the scaled Hessian is Θ(1). This implies the difference should be OP(n^{-1+a/2}), not OP(1). This is a typographical error, but it obscures the convergence rate and should be corrected in the supplementary material.","section":"Supplementary, proof of Theorem 2, Case C"}],"minor_comments":[{"comment":"Theorem 3 uses the normalization 'n · (ˆρ_n − ρ_n0) / sqrt(Σ̂_1,1)', which is correct because the variance estimator Σ̂ automatically scales as n^{2a-b} (or n^a in the b>a case), but the statement would be clearer with explicit parentheses, e.g., 'n(ˆρ_n − ρ_{n0})'. The stray 'q' artifacts in the displayed formulas should also be cleaned.","section":"Section 2.2, Theorem 3"},{"comment":"The proof of Proposition 1 derives ˆρ_n = (1/2) log( Σ I(S=1) / Σ I(S=0) ), while the proposition states the numerator as Σ I(S=±1). The numerator should be I(S=±1) to match the first-order condition; the current text is inconsistent.","section":"Supplementary, proof of Proposition 1"},{"comment":"The covariate V_ij is defined for i<j in Assumption 1, but the model formulas use V_il and V_jk for the S=-1 configuration. Since these expressions are only well-defined if V_ij = V_ji, the paper should state explicitly that dyadic covariates are symmetric.","section":"Section 1, model definition"},{"comment":"Proposition 2 states Var(τ_n ⊙ τ_n ⊙ Ψ_n(ϑ)) = o(1), but the proof analyzes Var(τ_n ⊙ Ψ_n(ϑ)). The extra 'τ_n ⊙' appears to be a typo and should be removed.","section":"Supplementary, Proposition 2 statement"}],"recommendation":"major_revision","confidential_remarks":"The main theoretical machinery appears sound, and the reader's concern about the 'n' normalization in Theorem 3 is not supported once one accounts for the implicit scaling of the sandwich variance estimator; I therefore do not list it as a major comment. The two substantive issues are the gap between the advertised and actual heterogeneity scope, and the unexplained computational feasibility of the simulations. Both can be addressed with a careful revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a real advance—conditional-likelihood estimation of reciprocity and covariate effects in a directed network with unrestricted node effects—and the theory is mostly credible. The paper earns a serious referee. But the abstract sells more than the assumptions deliver, and Theorem 3 as printed has a normalization slip that should be fixed before publication.\n\nWhat's new: the R2-model generalizes the p1 model by letting reciprocity depend on dyadic covariates, and the estimator extends Graham's tetrad conditioning to directed networks with two reciprocity parameters. That is not a trivial patch; the dependence structure is substantially more complex, and the U-statistic variance decomposition in the supplement is genuinely involved. The minimax lower bound via Le Cam is clean and does not depend on the estimator. Prior work (Yan and Leng 2015) was numerical, so this fills a real gap.\n\nSoft spots, in proportion: The load-bearing assumption is Assumption 2: max_i |alpha_i|,|beta_i| <= C. Every rate—event probabilities, covariance orders, normalizations—uses it. The paper says it 'accommodates node heterogeneity' and 'assigns each node its own sender and receiver parameters,' which is true, but the theory only covers moderate heterogeneity. Heavy-tailed degree networks don't satisfy this. That should be stated plainly in the abstract and introduction, not buried as 'moderate degree heterogeneity.' It's a scope limitation, not a fatal flaw.\n\nAlso, Theorem 3's printed normalization doesn't match Theorem 2 or the proof; the proof uses the correct rates. Someone should check that the displayed statement wasn't garbled. Minor: no code or data artifacts; simulations are descriptive, not reproducible. The comparison with the p1.5 MLE is informal—fine for a supplement.\n\nThe minimax claim is honest: it's rate-optimal over the parameter class Theta_n they define, and they say so. Just don't let 'among all possible estimations' escape the class constraint.\n\nWho is this for: applied network statisticians and econometricians doing directed network analysis will get value. I'd take it to a reading group. Verdict: send it to review. The central idea holds; the flaws are fixable. I'd want the authors to re-scope the claims and fix the Theorem 3 statement. If cleaned up, I'd cite it.","headline":"Solid, important extension of tetrad conditioning to directed networks with reciprocity; watch the bounded-heterogeneity assumption before you trust the advertised scope.","tokens_in":49737,"tokens_out":2540,"would_cite":true,"duration_ms":27138,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62E20","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces the R²-Model, which regresses reciprocity on dyad covariates while removing node heterogeneity by conditioning on four-node subgraphs, and proves the resulting estimator is minimax rate-optimal.","keywords":["directed networks","reciprocity","conditional likelihood","tetrad","sparse networks","node heterogeneity","asymptotic normality","minimax optimality"],"falsifier":"Simulate the model with a fixed sparsity regime ($a,b$), node sizes growing from $n=50$ to $n=200$, and $\\alpha_i,\\beta_i$ drawn so that $\\max_i|\\alpha_i|$ grows with $n$ (e.g., Pareto tails), rather than bounded as Assumption 2 requires; compare the empirical frequencies of the three tetrad types against $n^{-4a}$ and $n^{-4a+2b}$, and check whether the Theorem 3 confidence intervals for $\\rho_n$ and $\\gamma$ keep their nominal 95% coverage. A widening coverage gap as $\\max_i|\\alpha_i|$ grows would show where the minimax claim breaks.","tokens_in":48735,"feed_emoji":"🔁","tokens_out":10382,"duration_ms":100656,"temperature":0.7,"pith_summary":"The paper proposes a regression model for reciprocity in directed networks: the tendency of a pair of nodes to form mutual ties is written as $\\rho_{ij} = \\rho_n + V_{ij}^{\\top}\\gamma$, where $\\rho_n$ is a baseline that may change with network size and $V_{ij}$ are observed dyad covariates. The model generalizes the classical $p_1$ model by keeping unrestricted sender- and receiver-specific effects for every node while allowing for network sparsity. The authors show that conditioning on four-node subgraphs (tetrads) that share the same degree sequence removes all node-specific nuisance parameters, leaving a likelihood that depends only on $(\\rho_n, \\gamma)$; the resulting estimator is consistent, asymptotically normal, and attains the minimax rate of convergence over the sparsity-indexed parameter class. If these claims hold, this is the first rigorous and rate-optimal way to regress reciprocity on covariates in sparse directed networks.","feed_headline":"Optimal rate proven for covariate-driven reciprocity","feed_subtitle":"Conditioning on four-node subgraphs removes node effects, and proven bounds show no estimator can do better.","key_machinery":"The load-bearing object is the tetrad statistic $S_{ijkl}\\in\\{0,+1,-1\\}$: three directed configurations on four nodes that are edge-rewirings of one another and therefore share the same in- and out-degree sequence. Conditioning on $S_{ijkl}\\in\\{0,\\pm1\\}$ cancels the density parameter and all $2n$ node heterogeneity terms, leaving a three-outcome likelihood with probabilities $1/(1+r)$, $f/(1+r)$, $g/(1+r)$ that depend only on $\\rho_n$ and $\\gamma$—this is the mechanism that isolates the reciprocity effects. The convergence rates are set by the sparsity indices $a$ and $b$: the three tetrad events have probabilities of order $n^{-4a}$ and $n^{-4a+2b}$, so the score and Hessian need different normalizations in different regimes, and asymptotic normality is proved by viewing the score as a $U$-statistic and showing its Hájek projection onto dyads dominates the variance.","core_discovery":"The paper's central claim is that the reciprocity parameters can be estimated at the statistically optimal rate despite the presence of $2n$ nuisance parameters and sparsity. Explicitly, with sparsity indices $a,b$ defined by $\\mu_n = -a\\log n$ and $\\rho_n = b\\log n$, the tetrad-based conditional likelihood estimator satisfies $\\sqrt{n^{2-2a+\\min\\{a,b\\}}}(\\hat{\\rho}_n - \\rho_{n0}) \\to N(0,\\sigma_\\rho^2)$ and $\\sqrt{n^{2-2a+b}}(\\hat{\\gamma}-\\gamma_0) \\to N(0,\\Sigma_\\gamma)$, and Theorem 4 proves that no estimator over the parameter class can achieve mean squared error smaller than $n^{-2+2a-\\min\\{a,b\\}}$ for $\\rho_n$ or $n^{-2+2a-b}$ for $\\gamma$. The same construction gives a feasible inference procedure, Theorem 3, which does not require knowing $a$ or $b$.","pith_inferences":["The same tetrad-conditioning design should transfer to models with edgewise homophily covariates, because the sufficiency lemma only requires configurations sharing the same degree sequence; a direct implementation would give inference on the homophily slope without node-effect estimation.","A simulation with heavy-tailed $\\alpha_i,\\beta_i$ (for example, Pareto-distributed) would likely show tetrad frequencies and coverage degrading, marking the empirical boundary of Assumption 2; this is a concrete testable extension the paper does not run.","Although the theory is stated for a single network snapshot, the log-odds structure suggests the estimator could be adapted to panel or multi-layer directed networks by treating each layer's degree sequences as the conditioning statistic—an extension the authors do not pursue.","The closed-form estimator of Proposition 1 (a log ratio of tetrad counts) hints at an interpretation of the R²-Model as comparing two oriented 4-cycles; this connection suggests the methodology could be exported to any dyadic outcome model where a sufficient statistic can be frozen by degree-preserving rewirings."],"forward_implications":["Practitioners can test whether reciprocity varies with covariates using confidence intervals computed from Theorem 3 without knowing the sparsity parameters $a$ and $b$.","In the classical $p_1$ model (no covariates), the estimator reduces to a closed-form log-ratio of tetrad counts, providing a simple method-of-moments-style estimator of baseline reciprocity.","When reciprocity is strong ($b > a$), the covariate slopes are estimated at a strictly faster rate than the baseline reciprocity, because comparisons between the two $\\pm1$ tetrad configurations carry information about $\\gamma$ only.","The minimax lower bounds apply to all estimators, not just conditional likelihood ones, so no alternative estimator can beat the proposed rates in the moderate-heterogeneity regime."],"supporting_citations":[{"why":"Supplies the $p_1$ dyad-multinomial model that the R²-Model generalizes and defines the reciprocity parameter being estimated.","marker":"Holland and Leinhardt (1981)"},{"why":"Introduces tetrad conditioning for degree-heterogeneous network formation, the estimation strategy this paper extends to directed networks with reciprocity.","marker":"Graham (2017)"},{"why":"Establishes asymptotics for directed exponential random graph models with increasing bi-degree sequence in the no-reciprocity case, the benchmark the paper's theory improves upon.","marker":"Yan et al. (2016)"},{"why":"Prior work introducing the effective-sample-size view of reciprocity in directed networks and the $p_{1.5}$ model used as a mis-specified comparison in simulations.","marker":"Feng and Leng (2025)"},{"why":"Simulation study of the $p_1$ model; provides the maximum-likelihood baseline whose theory is incomplete.","marker":"Yan and Leng (2015)"},{"why":"States Le Cam's two-point method, the lemma used to prove the minimax lower bounds in Theorem 4.","marker":"Wainwright (2019)"},{"why":"Provides Pinsker's inequality and the minimax framework used in the lower-bound argument.","marker":"Tsybakov (2009)"},{"why":"Linked in Proposition 1, where the no-covariate estimator reduces to a log-ratio of tetrad counts resembling their method-of-moments estimator.","marker":"Chang et al. (2024)"}],"fun_headline_variants":["Optimal reciprocity rate proven despite nuisance parameters","Sparse networks: reciprocity estimated at minimax rate","Tetrad-based conditional likelihood hits reciprocity bound","Minimax bound proven for reciprocity in directed networks","Optimal reciprocity estimation via conditional likelihood"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory assumes every node's sender and receiver effects are bounded by a constant independent of $n$; if a network has hubs or extremely inactive nodes, the tetrad event probabilities and all rates built on them no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Optimal reciprocity rate proven despite nuisance parameters","Sparse networks: reciprocity estimated at minimax rate","Tetrad-based conditional likelihood hits reciprocity bound","Minimax bound proven for reciprocity in directed networks","Optimal reciprocity estimation via conditional likelihood"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001001,"raw_usage":{"total_tokens":4233,"prompt_tokens":942,"completion_tokens":3291,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":3222}},"tokens_in":558,"tokens_out":3291,"duration_ms":26489,"temperature":1.0,"reasoning_tokens":3222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:43:24.552602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the model with a fixed sparsity regime ($a,b$), node sizes growing from $n=50$ to $n=200$, and $\\alpha_i,\\beta_i$ drawn so that $\\max_i|\\alpha_i|$ grows with $n$ (e.g., Pareto tails), rather than bounded as Assumption 2 requires; compare the empirical frequencies of the three tetrad types against $n^{-4a}$ and $n^{-4a+2b}$, and check whether the Theorem 3 confidence intervals for $\\rho_n$ and $\\gamma$ keep their nominal 95% coverage. A widening coverage gap as $\\max_i|\\alpha_i|$ grows would show where the minimax claim breaks.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes asymptotics for directed exponential random graph models with increasing bi-degree sequence in the no-reciprocity case, the benchmark the paper's theory improves upon."},{"cited_title":"and Leng, C","cited_arxiv_id":null,"evidence_quote":"Prior work introducing the effective-sample-size view of reciprocity in directed networks and the $p_{1.5}$ model used as a mis-specified comparison in simulations."},{"cited_title":"and Leng, C","cited_arxiv_id":null,"evidence_quote":"Simulation study of the $p_1$ model; provides the maximum-likelihood baseline whose theory is incomplete."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Pinsker's inequality and the minimax framework used in the lower-bound argument."}],"review_version":1}