{"id":"cca4dfdb-2c6f-42e7-a5b6-bf5c2b74a40f","arxiv_id":"2508.04730","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A general core-periphery strength parameter, label recovery guarantees, and two hypothesis tests distinguish intrinsic core-periphery structure from degree-driven artifacts.","lead":"This paper builds a statistical framework for core-periphery structure in networks, defining a population parameter, proving when labels can be recovered, and introducing hypothesis tests against two null models. The methods can tell whether a network's core-periphery pattern is intrinsic or just a side effect of degree differences, and they suggest true core-periphery structure is rarer than commonly claimed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2's CL-null type-I-error proof silently invokes strong consistency (α_n/√ϱ_n = o(1/√(n log n))) and the CL ordering condition; neither is stated, so the size control is unproven as stated.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing gap: Theorem 3.2's type-I-error proof relies on a strong-consistency condition that is not among the theorem's stated assumptions. I agree this is the key issue because the CL-null test is a central novel contribution; without proven size control, the real-data conclusions about endogenous versus exogenous CP structure are not supported. I extend the reader's point by noting that strong consistency for the CL model also requires the ordering θ(k)θ(n) > θ(k+1)θ(k+2), and that the power proof imports the α_n ≥ (log(n log n))^2/n condition from Theorem 3.1. These are fixable by either adding the assumptions to Theorem 3.2 or proving the result under weaker conditions, so a conditional verdict remains appropriate. The paper's other contributions—label recovery under CP-SBM/CL/DCBM and the ER-null test—are largely unaffected by this specific gap, and the extensive simulations provide useful empirical support. My proposed simulation tests whether the missing condition is actually necessary or merely a proof artifact, which would settle whether the concern lands.","tokens_in":42157,"tokens_out":16273,"duration_ms":175377,"concrete_test":"Simulate the CL-null test with n=2000, k=n^{0.8} (so α_n=n^{-0.2}), ϱ_n=n^{-0.5}, and θ_i iid Uniform(0.5,1.5). This gives α_n/√ϱ_n = n^{0.05} → ∞, so the strong-consistency condition fails, and typically θ(k)θ(n) < θ(k+1)θ(k+2). Generate 1000 networks, run Algorithm 1 and the Theorem 3.2 rejection rule; if the empirical rejection rate is not near 0 at n=2000 and does not trend to 0 as n grows (e.g., n=1000,2000,4000), the unstated assumptions are load-bearing. If it is near 0, the proof gap is patchable and the theorem may survive with a weaker condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central hypothesis-testing claim for the Chung-Lu null (Theorem 3.2) is not established by its proof. In Supplemental Section 12, the proof of type I error begins 'Note that since α_n/√ϱ_n = o(1/√(n log n)), strong consistency is achieved as per Theorem 2.1,' then replaces T(A, hat c) with T(A, c*) and ρ(hat P, hat c) with ρ(hat P, c*). This condition appears nowhere in the theorem statement. Furthermore, Theorem 2.1's strong consistency for the CL model additionally requires θ(k)θ(n) > θ(k+1)θ(k+2), also not stated for H0. If either fails, the equality T1 = T(A,c*) with high probability breaks, and the argument bounding T1 by C1 collapses. The power part of the same proof also uses α_n ≥ (log(n log n))^2/n from Theorem 3.1 without stating it. Thus Theorem 3.2's guarantees are conditional on assumptions the reader is not told to check, directly affecting the paper's central claim that CP structure becomes testable against the CL null.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a model-agnostic population parameter rho(P,c) (Eq. 1) and its sample analogue T(A,c) (Eq. 3) to quantify core-periphery strength. It studies these under the ER, CP-SBM, Chung-Lu, and CP-DCBM models, proving label recovery guarantees (Theorem 2.1) and constructing intersection tests against ER and Chung-Lu nulls (Theorems 3.1 and 3.2) with analytic cutoffs. Simulations and thirteen real-world networks are presented. The central claim is that core-periphery structure becomes a testable statistical property, with a formal distinction between endogenous and exogenous CP structure.","tokens_in":42429,"tokens_out":9802,"duration_ms":100595,"significance":"If fully established, this is a valuable contribution: it provides the first recovery guarantees for a CP metric, analytic cutoffs for CP hypothesis tests, and a principled taxonomy of endogenous versus exogenous CP structure. The paper is ambitious and contains substantial proof effort in the supplemental materials, and the empirical evaluation is extensive and reproducible. The main theorems, however, have several load-bearing gaps that prevent the claims from being accepted as stated, particularly in the size control of the Chung-Lu null test and in the power proof of the ER test. These issues are fixable but require added assumptions or modified proof arguments.","major_comments":[{"comment":"The type I error proof for the Chung-Lu null test invokes the condition alpha_n/sqrt(rho_n) = o(1/sqrt(n log n)) to obtain strong consistency and then replaces T(A, hat c) with T(A, c*) and rho(hat P, hat c) with rho(hat P, c*). This condition is not stated anywhere in Theorem 3.2. Likewise, the CL strong-consistency result in Theorem 2.1 also requires the ordering condition theta_(k)theta_(n) > theta_(k+1)theta_(k+2), which is also not stated for the null hypothesis in Theorem 3.2. Without these assumptions, the inequality T(A, hat c) <= C1 used to control the size is not established. The theorem as stated is therefore unproven and needs either the missing conditions added or a proof that avoids the strong-consistency step.","section":"Section 3.2, Theorem 3.2; Supplemental Section 12, proof under H0"},{"comment":"The power proof for T2 asserts that 'Since alpha_n >= (log(n log n))^2/n, we have k^{1.5} rho_n log n = o(n^2 alpha_n^2 rho_n)'. This implication is false in general: if alpha_n = c (log n)^2 / n, then k = c log^2 n and the ratio equals 1/sqrt(c), a positive constant, not o(1). The dominance argument requires k >> log^2 n, i.e., alpha_n log^2 n / n -> infinity. Without this stronger condition, the proof that the third term dominates the fluctuation term in Eq. (20) fails, and the power-one claim for T2 under the stated alternatives is not justified. The condition in the theorem should be strengthened accordingly, or the proof revised.","section":"Theorem 3.1 proof, Supplemental Section 12, around Eq. (20)"},{"comment":"There is a mismatch between the theoretical optimization problem and the implemented estimator. Eq. (4) defines hat c as the argmax of T(A,c) over all labelings c, and Algorithm 1 swaps a single node's label at each step, which changes the core size. The theoretical guarantees, however, are proved only for labelings with the same core size as c*: Lemma 2 and the proof of Theorem 2.1 explicitly restrict to 'any c != c* such that the size of the core is the same for c and c*'. The paper states that k is assumed known but does not impose the constraint |c| = k in (4) or in Algorithm 1. Consequently, Theorem 2.1 as proved does not apply to the estimator actually used in simulations and data analysis. Either the optimization and algorithm should be modified to preserve core size k, or the theory must be extended to cover varying k.","section":"Section 2.4-2.5, Eq. (4) and Algorithm 1"}],"minor_comments":[{"comment":"The probability statements in (22) and (24) appear reversed: the proofs establish concentration, i.e., P[|...| <= ...] -> 1, but the displayed results state P[|...| <= ...] -> 0. Please correct these directions.","section":"Supplemental Section 12, Eqs. (22) and (24)"},{"comment":"The text 'strong consistency is achieved as per Theorem 2.3' should refer to Theorem 2.1, not Theorem 2.3.","section":"Supplemental Section 12, proof of Theorem 3.2 under H1"},{"comment":"The power proof for T2 also uses the condition alpha_n >= (log(n log n))^2/n (to claim C2 = o(rho_n)), but this condition is not listed among the three assumptions in the theorem statement. Furthermore, as noted in the second major comment, this condition is insufficient for the claimed order, so the same fix is needed here.","section":"Theorem 3.2, power part"},{"comment":"There is a duplicated 'are' in 'the sharper error bounds in Theorem 2.1 are are crucial'.","section":"Remark 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the paper gives core-periphery analysis something it has lacked—a population-level parameter, a taxonomy distinguishing endogenous from exogenous CP, consistency guarantees for label recovery under three canonical models, and analytic-cutoff intersection tests against ER and CL nulls. The label-recovery results and the ER-null test are, as far as I can tell, new and largely correct. The CL-null test, however, has a proof gap that needs to be fixed.\n\nThe good parts. The parameter ρ(P,c) is genuinely model-agnostic and connects to the BE metric, so the theory applies to practice. The taxonomy (ER at the origin, SBM endogenous, CL exogenous, DCBM both) is clear and the formalization of the \"shaded region\" in the DCBM is a nice way to state when endogenous structure dominates. Lemma 1's uniform deviation bound and Theorem 2.1's error rates are substantial; the o(α_n) bound is appropriately stronger than weak consistency, and the strong consistency condition has a clean interpretation. The intersection-test structure (T1 plus T2) is a real improvement over single-statistic tests, and the real-data example makes the point forcefully: T1 rejects for all 13 networks, but only 7 survive the T2 check, and only 4 survive the CL null. That is a useful message for practitioners. The supplement is detailed on the whole.\n\nThe soft spots. Theorem 3.2 is the load-bearing issue. The type-I-error proof replaces T(A,ĉ) with T(A,c*) by saying \"since α_n/√ϱ_n = o(1/√(n log n)), strong consistency is achieved as per Theorem 2.1.\" That condition appears nowhere in the theorem statement. For the CL null, Theorem 2.1's strong consistency also requires θ_(k)θ_(n) > θ_(k+1)θ_(k+2), and that ordering condition is absent from the statement of Theorem 3.2 as well. Without those assumptions, the equality T1 = T(A,c*) with high probability is not established, and the bound by C1 collapses. The power part uses α_n ≥ (log(n log n))^2/n without stating it. This is repairable—either add the assumptions to the theorem or find a proof that avoids strong consistency—but as it stands the CL-null test's guarantees are conditional on unstated conditions. Two smaller points: the theory is for the exact argmax over fixed-k labelings, while the implementation is a greedy algorithm; the gap is mentioned nowhere. And the real-data implementation uses plug-in estimates for ϱ_n and cutoffs, so the asymptotic thresholds should be treated as approximate. Neither is disqualifying.\n\nBottom line: I'd send this to a serious referee. The label-recovery contribution alone is enough to merit careful review, and the testing framework is worth pursuing despite the missing assumptions. For a reading group, the taxonomy and the intersection-test idea are definitely discussable. I'd cite it if I did CP inference.","headline":"A genuine step forward for CP inference—label recovery and the ER-null test are real advances, but the CL-null test (Theorem 3.2) is not proven as stated.","tokens_in":42883,"tokens_out":4469,"would_cite":true,"duration_ms":43195,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62F03","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single population parameter makes core-periphery structure a statistically testable property, and the maximizing sample metric recovers the true labels exactly under a sparsity condition.","keywords":["core-periphery structure","random graph models","stochastic block model","Chung-Lu model","degree-corrected stochastic block model","hypothesis testing","label recovery","network inference"],"falsifier":"Simulate a Chung-Lu network with no endogenous CP structure, choosing the core size $k$ so that $\\alpha_n/\\sqrt{\\varrho_n}$ is not $o(1/\\sqrt{n\\log n})$, and run the intersection test of Theorem 3.2; if the rejection rate under this null does not converge to zero, the claimed size control fails in that regime.","tokens_in":41975,"feed_emoji":"🕸️","tokens_out":7611,"duration_ms":83026,"temperature":0.7,"pith_summary":"This paper tries to put core-periphery structure on the same statistical footing as community structure: it defines a population-level measure of CP strength, shows that its sample version recovers the true core and periphery labels, and builds hypothesis tests with proven error guarantees. The parameter is model-agnostic, applying to any edge-independent random graph, and the paper studies it under the Erdős–Rényi, stochastic block, Chung–Lu, and degree-corrected stochastic block models. If the results hold, analysts can decide not only whether a network has a core-periphery pattern, but whether that pattern is intrinsic to the generative mechanism or merely an artifact of degree heterogeneity. Applied to thirteen real networks, the tests suggest that statistically significant core-periphery structure is somewhat rare once degree variation is accounted for.","feed_headline":"One metric recovers true core-periphery labels exactly","feed_subtitle":"New intersection tests separate genuine core-periphery wiring from degree-variation artifacts.","key_machinery":"The central object is the population parameter $\\rho(P,c)$, a correlation-like measure between centered edge probabilities and the core-periphery indicator $\\Delta_{ij}=c_i+c_j-c_ic_j$, with the sample version $T(A,c)$ replacing $P$ by the adjacency matrix $A$. This sample metric coincides with the Borgatti–Everett template-matching metric, so the new inference attaches statistical meaning to an established descriptive quantity. The proof machinery combines Bernstein-type concentration bounds, union bounds over labelings, and counting lemmas that relate the population gap $\\rho(P,c^*)-\\rho(P,c)$ to the misclassification fraction $\\xi_n(c)$; the hypothesis tests use analytic cutoffs $C_1$ and $C_2$ rather than bootstrap thresholds.","core_discovery":"The paper's central claim is that the strength of a core-periphery structure can be quantified at the level of the data-generating mechanism by the parameter $\\rho(P,c)$, the normalized centered expected edge count falling on core-core and core-periphery pairs, and that the sample analogue $T(A,c)$ obtained by maximizing over labelings is a statistically valid estimator of the labels. Under the CP-SBM, Chung-Lu, and degree-corrected SBM with the stated separation conditions, Theorem 2.1 gives a misclassification bound of order $o(\\alpha_n)$, and under the additional condition $\\alpha_n/\\sqrt{\\varrho_n}=o(1/\\sqrt{n\\log n})$ the labels are recovered exactly with probability tending to one. The paper further claims that the intersection test based on $T_1(A)=\\max_c T(A,c)$ and $T_2(A)=\\hat p_{11}-\\hat p_{12}$ drives type I error to zero under the Erdős–Rényi and Chung-Lu nulls and power to one under the specified alternatives, thereby distinguishing endogenous CP structure from exogenous structure induced by degree heterogeneity.","pith_inferences":["A natural extension, not developed in the paper, is to adapt $\\rho(P,c)$ to weighted or directed networks by replacing the adjacency indicator with a suitable edge-weight summary; the population parameter itself does not depend on the graph being binary.","The rarity of significant CP structure in the thirteen networks suggests that many reported cores in applied work may be assortative or disassortative community structure rather than true core-periphery structure, though this is an interpretation of the data analysis rather than a proven general claim.","A configuration-model null that preserves exact degrees would sit between the Erdős–Rényi and Chung-Lu nulls; building the same intersection test around it could clarify whether the authors' Chung-Lu threshold is too conservative for practitioners.","The size control of the Chung-Lu test in the proof replaces the estimated labels with the true labels using a strong-consistency condition that is not listed among the theorem's stated assumptions; whether this condition is theoretically necessary remains open."],"forward_implications":["Maximizing the sample metric becomes a consistent way to find core-periphery labels in stochastic block, Chung-Lu, and degree-corrected block models, with an explicit error rate controlled by the signal-to-noise ratio.","The two-stage testing recipe gives a practical decision rule: reject the Erdős–Rényi null to detect any CP-like structure, then reject the Chung-Lu null to conclude that the structure is stronger than degree heterogeneity alone can explain.","Analytic rejection thresholds make the significance tests scalable to large networks, removing the computational bottleneck of bootstrap-based p-values.","Exact label recovery is guaranteed only when the core is small enough relative to network density, so the theorem identifies the regime in which estimated core labels can be treated as trustworthy."],"supporting_citations":[{"why":"Defines the template-matching metric that $T(A,c)$ coincides with, linking the new inference to established core-periphery practice.","marker":"Borgatti and Everett (2000)"},{"why":"Supplies the Erdős–Rényi null model under which $\\rho(P,c)=0$ for every labeling.","marker":"Erdős and Rényi (1959)"},{"why":"Defines the degree-heterogeneity model whose induced structure is the paper's exogenous CP and the second null model.","marker":"Chung and Lu (2002)"},{"why":"Defines the degree-corrected stochastic block model used as the CP-DCBM alternative combining endogenous and exogenous structure.","marker":"Karrer and Newman (2011)"},{"why":"Sets the consistency framework for community detection on which the label recovery results are modeled.","marker":"Bickel and Chen (2009)"},{"why":"Provides the greedy optimization algorithm used to approximate the label maximizing $T(A,c)$.","marker":"Yanchenko (2022)"},{"why":"Motivates the need for null models that account for degree heterogeneity rather than relying only on the Erdős–Rényi model.","marker":"Kojaku and Masuda (2017)"},{"why":"Supplies the sparsity parameterization and consistency conventions adopted in Theorem 2.1.","marker":"Zhao et al. (2012)"}],"fun_headline_variants":["Core-periphery strength quantified with exact label recovery","Intersection tests separate true core-periphery from artifacts","Model-agnostic parameter for core-periphery structures","Exact label recovery in core-periphery networks","Statistical inference for core-periphery presence and strength"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main premise is that the true core is separated from the periphery by a large enough edge-probability gap and is small enough relative to network density; additionally, the Chung-Lu test's error-rate claim depends on a label-perfect-recovery condition that the theorem statement does not list among its assumptions.","fun_headline_variants_meta":{"raw":{"variants":["Core-periphery strength quantified with exact label recovery","Intersection tests separate true core-periphery from artifacts","Model-agnostic parameter for core-periphery structures","Exact label recovery in core-periphery networks","Statistical inference for core-periphery presence and strength"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1514,"prompt_tokens":973,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":463}},"tokens_in":589,"tokens_out":541,"duration_ms":6480,"temperature":1.0,"reasoning_tokens":463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:13:30.425556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a Chung-Lu network with no endogenous CP structure, choosing the core size $k$ so that $\\alpha_n/\\sqrt{\\varrho_n}$ is not $o(1/\\sqrt{n\\log n})$, and run the intersection test of Theorem 3.2; if the rejection rate under this null does not converge to zero, the claimed size control fails in that regime.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the template-matching metric that $T(A,c)$ coincides with, linking the new inference to established core-periphery practice."},{"cited_title":"and Newman, M","cited_arxiv_id":null,"evidence_quote":"Defines the degree-corrected stochastic block model used as the CP-DCBM alternative combining endogenous and exogenous structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the greedy optimization algorithm used to approximate the label maximizing $T(A,c)$."},{"cited_title":"and Masuda, N","cited_arxiv_id":null,"evidence_quote":"Motivates the need for null models that account for degree heterogeneity rather than relying only on the Erdős–Rényi model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sparsity parameterization and consistency conventions adopted in Theorem 2.1."}],"review_version":1}