{"id":"cf3dd4b9-f0bf-4424-8354-8989eb7a4ee6","arxiv_id":"2509.03772","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Four permutation tests (ridge, group LASSO, and two CCA variants) detect association between node covariates and random-dot-product-graph latent structure, with consistency theorems and cheaper computation than prior dependency tests.","lead":"This paper presents four statistical tests for whether the attributes attached to a network's nodes, such as age, protein function, or topics, line up with the network's wiring, using low-rank latent positions as the bridge. If correct, scientists get fast, assumption-light ways to decide when node features explain network structure and when the network still carries extra information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's high-dimensional guarantee for all four methods is unsupported: Theorems 3.12 and 3.13 assume fixed p, so only ridge/LASSO have p→∞ theory.","rationale":"The reader's weakest_assumption focused on the network model and the rates ξ_n, η_n, ζ_n, but the reader's rationale also flagged the fixed-p limitation via Remark 3.11. My analysis confirms that this is not a cosmetic detail: the CCA-based theorems' proofs rely on invertibility of the sample covariance of Z, specifically eΣ_Z^{-1/2}, and on p being fixed for the law-of-large-numbers steps in Lemma C.2 and Assumption 4. In the high-dimensional regime p>n used in the simulations, the unregularized CCA alignment is not even computable, so the abstract's phrase 'which we allow to be high-dimensional' cannot be read as covering the CCA methods. This does not undermine the internal validity of Theorems 3.4, 3.8, 3.12, and 3.13 under their stated assumptions; it narrows the central claim. I therefore keep the reader's conditional verdict and recommend only that the abstract be aligned with the theorem regimes and that the high-dimensional simulation analysis be labeled heuristic. No ad hominem intended; this is a scope and correctness-risk point.","tokens_in":53745,"tokens_out":20197,"duration_ms":221275,"concrete_test":"Track p=p_n through the proof of Theorem 3.12. Starting from Equations (24) and (26), replace the fixed-p assumption by p_n→∞ and attempt to control ||(1/n)Z^T M_n Z − Σ_Z|| in spectral norm; the proof requires eΣ_Z^{-1/2} to be bounded and consistent, which fails when p_n/n→∞ unless a regularized estimator is substituted and its error analyzed. Concretely: run the CCA permutation test at n=100, p=200 under H0 with unregularized eΣ_Z^{-1/2} and with the Fisher-Sun shrinkage used in Section 4; if the unregularized statistic is undefined or the level cannot be controlled under the stated assumptions, the abstract's 'all four methods' high-dimensional claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract) promises theoretical guarantees for all four methods when node covariates are high-dimensional. In the body, Section 2.2 explicitly restricts the CCA results to fixed p: 'we assume that the dimension p of the features Z is fixed with respect to n', and Remark 3.11 defers high-dimensional CCA to future work. Theorem 3.12 and Theorem 3.13 both involve CCA matrices containing eΣ_Z^{-1/2}, and Assumption 4 requires the population covariance Σ_Z to be invertible; when p_n→∞, especially p>n, eΣ_Z is singular and the proofs' spectral-norm perturbation arguments cannot be controlled. The only p→∞ theory in the paper is for ridge (Theorem 3.4/Corollary 3.5) and group LASSO (Theorem 3.8). The simulations apply CCA and netCCA at n=100, p=200 using Fisher-Sun shrinkage, which is not analyzed in the theorems. This is a scope mismatch in the paper's central claim rather than an internal contradiction inside Section 3, but it is the load-bearing limitation: the advertised high-dimensional coverage for two of the four methods is not delivered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces four (or, counting group LASSO, five) methods for testing association between observed node-level covariates and latent network structure under a low-rank latent-space model: ridge regression, (group) LASSO, CCA applied to estimated latent positions, and a regularized CCA applied directly to the adjacency matrix. The main theoretical results are Theorem 3.4 (ridge error bound), Theorem 3.8 (group-LASSO convergence rate with p allowed to grow), Theorem 3.12 (consistency of the plug-in CCA coefficient for fixed p), and Theorem 3.13 (consistency of the regularized adjacency-CCA coefficient for fixed p). The paper also contains simulations and real-data analyses using permutation tests based on these statistics.","tokens_in":53954,"tokens_out":8619,"duration_ms":100344,"significance":"The paper addresses a practically important problem and proposes computationally attractive methods. The proofs are based on standard and appropriate tools: the ridge and group-LASSO arguments follow the Buhlmann--van de Geer framework with an extra term for the estimated latent positions, and the CCA theorems reduce to singular-subspace perturbation bounds. The group-LASSO theorem, in particular, gives a concrete high-dimensional rate in terms of the embedding error xi_n and the compatibility constant. If the scope were stated accurately, the fixed-p CCA consistency results and the high-dimensional regression results would be a solid contribution. However, the advertised high-dimensional guarantee for the two CCA-based methods is not delivered, and the paper does not provide test-level guarantees for the permutation procedures that form the actual inferential proposal.","major_comments":[{"comment":"The abstract claims theoretical guarantees for all four methods when node covariates are high-dimensional. This is unsupported for the CCA-based methods. Section 2.2 explicitly assumes p is fixed, Remark 3.11 states that high-dimensional CCA is deferred to future work, and Theorems 3.12 and 3.13 both use eSigma_Z^{-1/2} under Assumption 4 requiring Sigma_Z invertible. When p_n > n, eSigma_Z is singular and the proofs' spectral-norm perturbation arguments cannot be controlled. The only p -> infinity theory in the paper is for ridge (Theorem 3.4/Corollary 3.5) and group LASSO (Theorem 3.8). The abstract and Section 5 should be revised to distinguish the high-dimensional guarantees for regression methods from the fixed-p consistency results for CCA methods, or new high-dimensional CCA theory must be supplied.","section":"Abstract and Section 2.2, Eq. (12)-(14); Remark 3.11"},{"comment":"The paper's stated goal is testing, but Theorems 3.12 and 3.13 only establish that the test statistic converges in probability to the oracle population CCA coefficient. They do not establish that the permutation test controls level under H0 or has power. For the CCA-based methods, H0 is defined in Section 4.1 as zero population covariance, which is weaker than the row-exchangeability of Z given X needed for exact permutation inference. Unless the null model is taken to include independence of X and Z, or an asymptotic analysis of the permutation null distribution is supplied, the central testing claim is not proven. This is a load-bearing gap, not a presentation issue.","section":"Section 4.1; Theorems 3.12 and 3.13"},{"comment":"The simulations apply CCA and netCCA with n=100, p=200 using the Fisher-Sun shrinkage estimator, a setting not covered by Theorem 3.12 or 3.13. The paper acknowledges this in Remark 3.11, but the abstract's high-dimensional promise and the simulation section should clearly separate these heuristic demonstrations from the proved guarantees. Otherwise the reader cannot tell which simulation results are covered by theory.","section":"Section 4.2, Tables and Figures; Remark 3.11"}],"minor_comments":[{"comment":"The paper alternates between four and five methods. The abstract says four novel methods, but Section 5 summarizes five methods (ridge, LASSO, group LASSO, CCA, netCCA). This should be reconciled, e.g., by treating group LASSO as a variant of the LASSO-based approach.","section":"Abstract and Section 5"},{"comment":"There is a typo: 'constitistutes' should be 'constitutes'.","section":"Section 1, paragraph 2"},{"comment":"The bottom-left and bottom-right panels are described in the caption as 'sSNR' but appear to refer to rSNR; please check and standardize the notation.","section":"Figure 3 caption"},{"comment":"It would be helpful to add a short table or remark listing which of the six simulated scenarios are covered by which theorem, and which are outside the theorem assumptions. This would prevent the reader from thinking the high-dimensional CCA simulations are theoretically guaranteed.","section":"Remark 3.11 and Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a real kernel: the regression-based guarantees are plausible and the fixed-p CCA consistency arguments are technically sound. However, the abstract's blanket high-dimensional claim is not backed by the theorems, and the permutation-test validity is not established. Both issues are fixable by honest scope revision and by adding an exchangeability/independence condition or an asymptotic permutation-null analysis. I recommend major revision rather than rejection, because the core techniques appear correct and the gaps are in presentation and in missing test-level statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: competent and useful paper, with one real but contained flaw. The four methods (ridge, group LASSO, CCA on spectral embeddings, regularized adjacency CCA) use known pieces, but their combination for network dependency testing is new, and the theory is mostly honest.\n\nThe stress-test note is right. The abstract says the paper provides theoretical guarantees for all four methods in the high-dimensional setting. The body delivers p→∞ theory only for ridge (Thm 3.4) and group LASSO (Thm 3.8). The CCA theorems (3.12, 3.13) assume fixed p, and Remark 3.11 explicitly defers high-dimensional CCA to future work. So the headline claim overreaches for the two model-free methods. This is a scope mismatch, not a broken proof—the fixed-p arguments themselves look sound. I spot-checked the plug-in bound in Thm 3.4, the group-LASSO proof against Buhlmann–van de Geer with the extra Xhat error term, and the CCA proofs' use of Cai–Zhang and Yu et al. bounds, and found no load-bearing error.\n\nWhat is genuinely good: the plug-in idea is handled carefully. Theorems 3.4 and 3.8 account for the (2,∞) embedding error instead of treating Xhat as if it were X, and Thm 3.13 (netCCA consistency without selecting a latent dimension) is the most interesting piece. Simulations cover six scenarios including misspecified ones, level is checked under H0, and the head-to-head with Lee et al. shows comparable power at a fraction of the runtime. The limitation remarks (2.1, 2.3, 3.11) are present and honest.\n\nSoft spots, in proportion: (1) the abstract; (2) no code or data shipped; (3) the simulation's latent-dimension selection method is not named, and the high-dimensional simulations (n=100, p=200) apply Fisher–Sun shrinkage to CCA, which is outside the theorems—so the high-dimensional empirical claims are only loosely anchored to the theory. All three are fixable in revision.\n\nThe central argument holds up. Applied network scientists who want a fast dependency test get real value here, and a serious referee should engage. I'd send to peer review with a required revision aligning the abstract to the theorem regimes, naming the dimension-selection procedure, and ideally shipping code.","headline":"Solid, useful network dependency testing toolkit with honest proofs; the abstract overclaims high-dimensional coverage for the CCA-based methods, which are fixed-p only.","tokens_in":54561,"tokens_out":3354,"would_cite":true,"duration_ms":33160,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62F03","05C80","62H12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that four low-cost tests—ridge/group-LASSO regression and two canonical-correlation variants—reliably detect whether node-level covariates are associated with the latent structure of a low-rank network, even when the covar","keywords":["network dependency testing","random dot product graph","canonical correlation analysis","spectral embedding","high-dimensional covariates","permutation tests","group LASSO","ridge regression"],"falsifier":"Fix a stochastic blockmodel with d=2 true latent dimensions and Bernoulli edges, set the covariates independent of X, and run the spectral-embedding CCA permutation test with d incorrectly chosen as 3; if rejection rates at nominal 5% exceed binomial sampling error substantially as n grows, Assumption 2's correct-dimension premise fails and Theorem 3.12 no longer applies. Similarly, generate weighted RDPG edges with heavy-tailed noise and check whether the deviation of the regularized adjacency-CCA statistic from the oracle stays within the eta_n, zeta_n rates claimed in Theorem 3.13.","tokens_in":53517,"feed_emoji":"🕸️","tokens_out":8073,"duration_ms":82528,"temperature":0.7,"pith_summary":"The paper shows how to test whether node-level covariates are correlated with network structure without observing the latent variables that drive the network. Its central claim is that plugging spectral estimates of latent positions into ridge or group-LASSO regression, or into canonical correlation analysis, yields statistics that converge to the oracle quantities one would get from the true latent positions. The CCA guarantees hold for both the estimated-latent-position version and a regularized version that works directly on the adjacency matrix, avoiding the need to choose the latent dimension. All four methods run on matrix-vector products and leading singular values, so they scale to large networks and high-dimensional feature vectors. Simulations and two real-data analyses indicate the tests control their level and detect association under linear, nonlinear, and assortative-mixing generative models.","feed_headline":"Four fast tests check whether node traits track network structure","feed_subtitle":"Spectral embeddings and canonical correlation keep the tests consistent even with high-dimensional covariates.","key_machinery":"The central object is the CCA alignment score rho(X,Z), the leading singular value of the normalized cross-covariance matrix Sigma_X^{-1/2} Sigma_{XZ} Sigma_Z^{-1/2}, and its plug-in analogues CCA(hat X, Z) and CCA_gamma(A,Z). The workhorse is the two-to-infinity norm spectral embedding guarantee (Assumption 2): hat X recovers X up to an orthogonal rotation with row-wise error xi_n = o(1). Because the CCA coefficient is invariant to orthogonal rotation of either data block, the unknown rotation cancels; sin-theta subspace perturbation bounds then transfer row-wise embedding error into singular-value error. For the adjacency CCA, the paper proves a pseudo-inverse lemma showing that the nonzer","core_discovery":"The central discovery is that network dependency testing can be carried out consistently on estimated latent positions rather than true ones. The paper proves (Theorem 3.12) that |rho(hat X, Z) - rho(X,Z)| = oP(1) when spectral embeddings recover the latent positions with (2,infinity)-norm error xi_n = o(1); the unknown orthogonal rotation that makes hat X an estimate of X cancels because the CCA coefficient is a subspace-alignment quantity. It proves further (Theorem 3.13) that canonical correlation between the adjacency matrix itself and Z, with a ridge-style regularization gamma I on the network sample covariance, also tracks the oracle coefficient at rate OP(eta_n/sqrt(gamma n) + |gamma|","pith_inferences":["A testable extension the paper leaves open is replacing the sample covariance of Z in CCA(hat X, Z) with sparse or regularized CCA when p >> n; simulations suggest power would then survive high-dimensional regimes where dense CCA currently loses to LASSO.","The proof of Theorem 3.13 relies only on concentration and subspace perturbation, not on Bernoulli edges, so the same consistency should transfer to generalized random dot product graphs and graph-root models, as the paper anticipates.","The group-LASSO covariance test derived in Appendix F is a standalone contribution to multivariate sparse regression: it yields a closed-form first-entry alpha and test statistic without cross-validation, independent of the network setting.","Since the plug-in CCA error is driven by xi_n, the practical bottleneck is spectral dimension selection (Remark 2.3); improving automated rank selection would strengthen all four tests more than improving covariate regularization."],"forward_implications":["Network scientists can test whether node attributes track latent structure without strong parametric assumptions, using spectral embedding plus CCA and permutation p-values.","Ridge and group-LASSO methods remain consistent when p grows with n, with rates slowed only by the spectral embedding error xi_n, so covariate selection on network-linked data is feasible.","The regularized adjacency CCA avoids choosing the latent dimension d, replacing a model-selection problem with a regularization parameter gamma that has a wide safe range.","The methods are computationally cheaper than diffusion-map distance-correlation baselines, relying only on matrix-vector products and leading singular values.","The theorems give formal level control for the permutation tests under the null in the low-rank model, and power follows against linear association (and empirically under nonlinear misspecification)."],"supporting_citations":[{"why":"Defines the random dot product graph model and the spectral embedding estimators whose row-wise recovery rates feed Assumptions 1 and 2.","marker":"Athreya et al. 2018"},{"why":"Introduces adjacency spectral embedding, the estimator used to build hat X in the plug-in statistics.","marker":"Sussman et al. 2012"},{"why":"Supplies the binary-network two-to-infinity norm rate xi_n = sqrt(d/n log^2 n) used in Theorem 3.1 for Assumption 2.","marker":"Lyzinski et al. 2017"},{"why":"Supplies sub-gamma weighted-network rates for eta_n, xi_n, and zeta_n used in Assumptions 1-3 and Theorems 3.12-3.13.","marker":"Levin et al. 2022"},{"why":"Provides the sin-theta singular-subspace perturbation bounds used to transfer row-wise embedding error into CCA singular-value error.","marker":"Cai and Zhang 2018"},{"why":"Provides the Davis-Kahan variant used in Theorem 3.12 to bound sin Theta(W_hat, W).","marker":"Yu et al. 2014"},{"why":"Establishes the two-to-infinity norm geometry and subspace perturbation machinery used throughout the CCA proofs.","marker":"Cape et al. 2019"},{"why":"Provides the LASSO oracle-inequality and compatibility-constant framework adapted in Theorem 3.8 for the group LASSO estimator.","marker":"Bühlmann and van de Geer 2011"},{"why":"Defines the covariance test statistic used for the LASSO-based permutation tests, whose group-LASSO analogue the paper derives.","marker":"Lockhart et al. 2014"},{"why":"Provides the diffusion-map distance-correlation baseline that the paper compares against in power and runtime experiments.","marker":"Lee et al. 2019"}],"fun_headline_variants":["Four tests reveal if node traits match network shape","Quick CCA and spectral checks for network–feature ties","Provable tests for high-dimensional node-covariate correlation","Four ways to test network structure against node features","Spectral embeddings and CCA for network dependency tests"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"All guarantees hang on the assumption that the observed network is truly a low-rank latent-position network whose edges are independent conditional on the latent positions and concentrate as required, with the latent dimension d known: if edges are dependent or heavy-tailed, or d is misspecified, the plug-in statistics need not converge to the oracle correlation.","fun_headline_variants_meta":{"raw":{"variants":["Four tests reveal if node traits match network shape","Quick CCA and spectral checks for network–feature ties","Provable tests for high-dimensional node-covariate correlation","Four ways to test network structure against node features","Spectral embeddings and CCA for network dependency tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000409,"raw_usage":{"total_tokens":1920,"prompt_tokens":671,"completion_tokens":1249,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":1173}},"tokens_in":415,"tokens_out":1249,"duration_ms":12942,"temperature":1.0,"reasoning_tokens":1173,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:43:03.673514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a stochastic blockmodel with d=2 true latent dimensions and Bernoulli edges, set the covariates independent of X, and run the spectral-embedding CCA permutation test with d incorrectly chosen as 3; if rejection rates at nominal 5% exceed binomial sampling error substantially as n grows, Assumption 2's correct-dimension premise fails and Theorem 3.12 no longer applies. Similarly, generate weighted RDPG edges with heavy-tailed noise and check whether the deviation of the regularized adjacency-CCA statistic from the oracle stays within the eta_n, zeta_n rates claimed in Theorem 3.13.","supporting_citations":[],"review_version":1}