{"id":"13299dc3-c4b5-4cb7-8c62-9bcb8231ed46","arxiv_id":"2412.03780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proposes a degree-corrected, signed block covariance model for feature clustering, with a variational EM algorithm and a proof that community memberships are consistently estimable.","lead":"This paper introduces a statistical model for grouping the columns of a data matrix when the meaningful similarity is correlation rather than distance, and each column may have its own overall activity level. The authors provide an algorithm with a consistency guarantee and demonstrate it on single-cell gene expression and stock return data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 proves consistency for the global maximizer of a simplified objective, not for the coordinate-ascent variational EM actually implemented; the paper's own random-initialization experiments show local optima, so the algorithm's consistency is unproven.","rationale":"I read the proof of Theorem 2 in good faith: Proposition 4 establishes uniqueness of the population objective, Proposition 5 gives the population gap, Proposition 6 handles uniform convergence, and the algebra in the proof of Proposition 5 is correct—the key lower bound by misclassification follows from T = E_q[(U−α)²]. I found no internal inconsistency in the asymptotic argument itself. The weakest point is exactly where the reader located it: the theorem's estimator is a global maximizer of a modified objective, while the paper's contribution is an algorithm that is not shown to reach that maximizer. The supplemental random-initialization simulation (S.1.3, Figure 7) already shows poor local optima, which makes the gap practical rather than hypothetical. This does not invalidate the model or the asymptotic analysis; it means the advertised consistency guarantee applies to a different estimator, and the manuscript should either prove a property of the algorithm's limit (for example, a local-maximum version with a favorable initialization basin) or explicitly present Theorem 2 as a property of the variational objective and treat the algorithm as heuristic. This is an addressable, non-fatal issue, so the reader's CONDITIONAL verdict is appropriate. The missing BCM/G-block comparisons in the main simulations also weaken the empirical claims, but that is secondary.","tokens_in":35583,"tokens_out":11720,"duration_ms":127422,"concrete_test":"Run a small exhaustive experiment where the global maximizer can be found exactly. Choose K=2, P=16 or 20, N=500, with parameters from the Table 1 setting; for every one of the 2^P hard label assignments c, maximize the full variational objective over q2, π, Ω, λ, and σ² by iterating the E/M updates to convergence (inner EM), and record the best objective and its ARI against the true labels. Then run Algorithm 1 from the spectral-clustering initialization and from 20 random starts on the same datasets, recording each final objective and ARI. If Algorithm 1's output has lower objective than the global maximizer, or lower ARI while the global maximizer recovers true labels, the theorem-algorithm gap is real; if it consistently reaches the global maximizer, the gap is only formal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is that Theorem 2 establishes label consistency of the global maximizer of the sample variational objective \\hat J over q1(c), q2(α), π, and Ω, but Algorithm 1 is a coordinate-ascent variational EM algorithm. No result shows that the sequence produced by Algorithm 1 converges to a global maximizer of \\hat J, or even to a stationary point with comparable objective value. Moreover, \\hat J itself differs from the algorithm's objective: \\hat J fixes λ at the moment estimator \\hat λ and omits {σ²_j}, while Algorithm 1 alternately maximizes the full variational objective J(q1,q2,Φ) over λ and σ² as well. Coordinate ascent on a nonconvex objective is only guaranteed to reach a local optimum. The paper's own Supplementary Section S.1.3 and Figure 7 demonstrate that HBCM with random initial values occasionally produces low-ARI solutions, which is direct empirical evidence that local optima with poor labels exist. Since the abstract and Section 6 claim consistent estimates from the proposed algorithm, the central guarantee and the implemented method are different objects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the heterogeneous block covariance model (HBCM), a covariance-based model for clustering P features from an N × P data matrix. The population covariance is modeled as Σ_jj' = λ_j λ_j' ω_{c_j c_j'} for j≠j' and λ_j² ω_{c_j c_j} + σ_j² for j=j', which permits signed, continuous, and heterogeneous edge weights. The authors propose a variational EM algorithm built on a second layer of latent variables α_i and a mean-field factorization q(c,α)=q1(c)q2(α). The theoretical section proves identifiability of the membership up to permutation and a scale ambiguity between λ and Ω, and establishes consistency of the community memberships for the maximizer of a sample variational objective (Theorem 2). The paper also provides simulations comparing HBCM with spectral clustering, SBM, and DC-SBM, a cross-validation procedure for selecting the number of communities, and applications to mouse embryo single-cell RNA-seq data and stock price data.","tokens_in":35840,"tokens_out":7668,"duration_ms":89120,"significance":"If the consistency guarantee covered the implemented algorithm, the paper would be a useful contribution: it extends block-model ideas to weighted, signed, covariance-type data without dichotomization, accounts for per-feature heterogeneity, and provides detailed identifiability analysis and a computationally efficient variational EM scheme. The simulations are broad, the applications provide external validation through GICS sectors and MSigDB gene sets, and the theoretical development is largely self-contained. However, the central load-bearing issue is that Theorem 2 concerns a global maximizer of a simplified variational objective, while Algorithm 1 is a coordinate-ascent method on a different objective; the paper's own random-initialization experiments show that local optima with poor labels exist. This gap is real and currently prevents the paper from supporting the headline claim that the proposed algorithm gives provably consistent estimates.","major_comments":[{"comment":"Theorem 2 establishes label consistency for the global maximizer (q̂c, q̂2, π̂, Ω̂) of the sample objective Ĵ over the constraint sets, but Algorithm 1 is a coordinate-ascent variational EM that alternately updates q1, q2, and Φ. Coordinate ascent on a nonconvex objective is only guaranteed to reach a local stationary point, and no result in the paper shows that the trajectory of Algorithm 1 approaches a global maximizer of Ĵ or attains a comparable objective value. Supplementary Section S.1.3 and Figure 7 provide direct empirical evidence that local optima with low ARI exist even when ten random starts are used and the highest objective value is selected. Since the abstract and Section 6 attribute consistency to the proposed algorithm, the theorem and the implemented method are currently different objects; this is the main load-bearing gap that must be closed or explicitly acknowledged.","section":"§3.2, Theorem 2 and Algorithm 1"},{"comment":"The objective actually maximized by Algorithm 1, J(q1,q2,Φ) in Eq. (4), includes λ and σ² as parameters to be updated in the M-step, whereas the theoretical sample objective Ĵ fixes λ at the moment estimator λ̂_j and omits {σ²_j}. Thus even the exact global maximizer of Ĵ need not be a fixed point of the algorithm. The paper should either prove a quantitative relation between the two optimizations under the stated assumptions, or redefine the theoretical analysis so that it applies to the objective used by Algorithm 1.","section":"§2.2, Eq. (4) and §3.2 definition of Ĵ"},{"comment":"Proposition 5 assumes |α*_ik| ≤ B1 for a fixed constant B1, while the model generates α*_i ∼ N(0, Ω*). For fixed B1, max_{i,k} |α*_ik| diverges almost surely as N → ∞, so the assumed constraint is violated with probability tending to one under the paper's own data-generating process. Because Theorem 2's probability statement conditions only on c* and not on α*, the theorem as written does not cover the generative model. The authors should either state the result conditionally on a high-probability event and allow B1 to grow at an appropriate rate, or treat α* as fixed non-random and replace the Gaussian-tail argument for d_min with an assumption stated directly on the realized α*.","section":"§3.2, Proposition 5 assumptions"}],"minor_comments":[{"comment":"The constraint set C_Ω is written as a set of P × P matrices, but the model's Ω is K × K; the eigenvalue bounds should be stated for the K × K block matrix.","section":"§3.2, definition of C_Ω"},{"comment":"The proof contains the typo 'Jensen's ineqaulity'; it should read 'Jensen's inequality'.","section":"Supplement S.4, Proposition 7 proof"},{"comment":"The Tabu search reference is cited as 'Glover and Lagunas'; the correct author name is Laguna.","section":"References"},{"comment":"Proposition 4 says the objective 'is maximized by ω', but the maximization is over (q^c, μ); the symbol ω appears to be a typo and should be corrected.","section":"§3.2, Proposition 4"},{"comment":"The text says the cross-validation result is 'shown in the left panel of Figure 6', but the cross-validation plot appears as panel (a) of Figure 5; the figure reference should be corrected.","section":"§5, mouse embryo analysis"}],"recommendation":"major_revision","confidential_remarks":"The main gap between Theorem 2 and Algorithm 1 is not an artifact of the review pipeline; the manuscript's own random-initialization experiments (S.1.3, Figure 7) implicitly acknowledge that the algorithm can settle in poor local optima. I would not require a general proof of global convergence of variational EM, but the authors need to either provide a verifiable optimality condition, prove that their initialization and multiple-restart strategy attain the global maximizer with high probability, or explicitly narrow the consistency claim to the exact maximizer and present the algorithm as a heuristic with empirical support. The bounded-α* issue in Proposition 5 is also fixable but should be addressed. The paper fits the journal's scope if these issues are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine contribution, but the main theorem is proved for a global maximizer of a simplified objective, while the algorithm is a coordinate-ascent variational EM. The paper's own random-initialization experiments show low-ARI local optima, so the \"provable consistent estimates\" claim in the abstract and discussion outruns the proof.\n\nWhat's new: the HBCM extends the block covariance model with per-feature scale parameters, allowing signed and heterogeneous covariances within a community. The two-layer variational EM is a real algorithmic device with closed-form updates and O(NP) complexity. Theorem 1 on identifiability is careful and genuinely useful. The consistency proof for the global maximizer, while conditional on compact parameter sets and boundedness assumptions, is a substantial piece of asymptotic work. Simulations are extensive and cover signal strength, misspecification, heavy-tailed noise, and comparison with spectral, SBM, DC-SBM, and WGCNA. The two applications (scRNA-seq, stock sectors) provide external validation.\n\nThe soft spots, in order. First, the gap between Theorem 2 and Algorithm 1 is not a technicality. The theorem treats the maximizer of a version of the objective that fixes lambda at a moment estimator and drops sigma-squared; Algorithm 1 maximizes the full variational objective by coordinate ascent. No result shows the algorithm's output approaches the global maximizer. Figure 7 demonstrates that random starts occasionally land in poor local optima, and spectral initialization helps but is not part of the theorem. This is addressable, but the paper should either prove a local-optimum consistency result or state clearly that consistency is for the global maximizer and the algorithm is a heuristic that appears to work well. Second, the simulations omit the most relevant baselines: BCM (Yang 2020) and G-block covariance (Bunea et al. 2020). They are only mentioned in the introduction. A comparison would directly show the added value of the heterogeneous parameters. Third, the cross-validation method for choosing K is heuristic; the paper says so, which is fine.\n\nWho is this for? Statisticians working on covariance-based clustering, weighted community detection, and gene co-expression analysis. They get a flexible model, a fast algorithm, and a solid but slightly overclaimed theory. It deserves a serious referee, and I would send it out with a request to close or honestly frame the optimization gap and to add the missing baselines.","headline":"Solid model-plus-algorithm paper, but the consistency theorem covers a global maximizer of a simplified objective, not the coordinate-ascent EM that is actually run; the gap is real and the paper's own experiments show it.","tokens_in":36333,"tokens_out":2115,"would_cite":true,"duration_ms":23730,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces the heterogeneous block covariance model for clustering features from signed, continuous weighted data and proves that the variational objective's maximizer recovers the true community memberships up to label…","keywords":["community detection","weighted network","covariance matrix","variational EM algorithm","block covariance model","label consistency","gene expression data","feature clustering"],"falsifier":"Take a small HBCM instance (e.g., $P = 12$ features, $K = 3$), enumerate all $3^{12}$ labelings to find the exact global maximizer of the estimated variational objective, and run Algorithm 1 from many random initializations; if the algorithm's output regularly has higher misclassification error than the global maximizer and does not converge to the true labels as $N$ grows to thousands, the practical claim of membership consistency would be refuted.","tokens_in":35392,"feed_emoji":"🧬","tokens_out":5656,"duration_ms":48401,"temperature":0.7,"pith_summary":"This paper introduces the heterogeneous block covariance model (HBCM), a generative model for clustering features (columns) of a data matrix when the signal lives in the covariance structure rather than in the means. The model lets pairs of features have signed, continuous covariances, and lets each feature have its own scale parameters, so features in one community can differ in how strongly they connect without losing community structure. The authors prove that the community memberships are identifiable up to label permutation, and that under mild growth conditions the maximizer of their variational objective recovers the true memberships with probability tending to one. They also supply a variational EM algorithm whose updates all have closed forms, and demonstrate on simulations and real gene-expression and stock-price data that it improves on spectral clustering baselines.","feed_headline":"Covariance model recovers true feature clusters","feed_subtitle":"Heterogeneous block covariance model plus variational EM yields provable membership consistency.","key_machinery":"The machinery is a variational expectation-maximization objective built on an augmented latent structure: the model writes each observation as $X_{ij} = \\lambda_j \\alpha_{i c_j} + \\sigma_j \\epsilon_{ij}$, where $\\alpha_i$ is a $K$-dimensional random effect with community-level covariance $\\Omega$. This second latent layer puts the complete-data likelihood in the exponential family, yielding closed-form E- and M-steps and reducing the posterior to factorized forms $q_1(c)=\\prod_j q_{1j}(c_j)$ and $q_2(\\alpha)=\\prod_i N(\\mu_i, V_i)$. The consistency proof works with the population version of the core objective, showing its unique maximizer is the true label/\\$\\alpha$ pair up to permutation, and then uses uniform convergence of the sample core objective plus a soft confusion matrix $R$ to transfer the separation to the estimated label distribution.","core_discovery":"On its own terms, the paper's central discovery is that community detection on covariance-structured data can be carried out as likelihood-based inference under a model that separates community-level covariance (a $K\\times K$ matrix $\\Omega$) from per-feature heterogeneity (scales $\\lambda_j$ and variance offsets $\\sigma_j^2$). Theorems 1 and 2 are the load-bearing results: Theorem 1 characterizes all parameter systems giving the same covariance and shows membership is uniquely defined up to label permutation, and Theorem 2 states that, as $N$ and $P$ grow with $\\log(P)/N = o(1)$ and under separation conditions on the canonical parameters, the maximizer of the variational objective satisfies $P( \\min_{\\tilde{I}^c \\in \\mathcal{E}_{I^c}} 1 - \\operatorname{Tr}(R(\\hat{q}^c, \\tilde{I}^c)) \\ge \\epsilon \\mid c^*) \\to 0$ for every $\\epsilon > 0$. In words, the optimal variational label distribution becomes a Dirac measure on the true memberships.","pith_inferences":["The paper's theory covers the global maximizer of the variational objective; a natural test is whether the coordinate-ascent algorithm from random starts reliably reaches that maximizer, since local optima would break the link between Theorem 2 and Algorithm 1.","The same augmented-latent-variable trick could be applied to other profile-likelihood problems with intractable label sums, such as extensions of the latent block model to row and column clustering simultaneously.","If Theorem 2 extends to growing $K$, the model could be used for hierarchical community detection in large correlation networks such as genome-wide co-expression or financial sector graphs."],"forward_implications":["Signed, continuous edge weights can be clustered directly without thresholding or taking absolute values, so negatively correlated features can share a community.","For gene-expression studies, the model gives a principled replacement for ad-hoc correlation-network thresholds and WGCNA-style pipelines.","The algorithm is polynomial-time per iteration ($O(NP)$), so feature clustering scales to thousands of features.","The cross-validation procedure in Algorithm 2 provides an ARI-based rule for choosing the number of communities $K$.","Because the membership is scale invariant, standard preprocessing of columns does not change the recovered clustering."],"supporting_citations":[{"why":"Defines the degree-corrected stochastic block model whose per-node heterogeneity motivates the $\\lambda_j$ parameters.","marker":"Karrer and Newman, 2011"},{"why":"Gives consistency guarantees for the degree-corrected block model, the template for the paper's label-consistency theorem.","marker":"Zhao et al., 2012"},{"why":"Establishes minimax-optimal recovery for a G-block covariance model that HBCM extends with heterogeneous scales.","marker":"Bunea et al., 2020"},{"why":"Provides the profile-likelihood approach for the block covariance model that motivates the likelihood formulation.","marker":"Yang, 2020"},{"why":"Supplies the variational inference framework used to make the EM updates tractable.","marker":"Blei et al., 2017"},{"why":"Provides the spectral clustering algorithm used for initialization and as the main comparison baseline.","marker":"Ng et al., 2002"},{"why":"Justifies treating the EM algorithm as coordinate ascent, which underlies the variational EM derivation.","marker":"Neal and Hinton, 1998"},{"why":"Benchmark method (WGCNA) used in simulations and real-data comparisons.","marker":"Langfelder and Horvath, 2008"}],"fun_headline_variants":["HBCM: consistent clusters from signed weighted networks","Covariance community detection via variational EM","Signed edges? HBCM still finds true communities","Heterogeneous block covariance model: provable membership"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof shows that the global maximizer of the variational objective recovers the true labels, but Algorithm 1 is a coordinate-ascent variational EM that is only guaranteed to reach a local optimum, so the consistency guarantee may not cover the labels the algorithm actually outputs.","fun_headline_variants_meta":{"raw":{"variants":["HBCM: consistent clusters from signed weighted networks","Covariance community detection via variational EM","Signed edges? HBCM still finds true communities","Heterogeneous block covariance model: provable membership"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1949,"prompt_tokens":921,"completion_tokens":1028,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":967}},"tokens_in":537,"tokens_out":1028,"duration_ms":10790,"temperature":1.0,"reasoning_tokens":967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:05:49.136748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small HBCM instance (e.g., $P = 12$ features, $K = 3$), enumerate all $3^{12}$ labelings to find the exact global maximizer of the estimated variational objective, and run Algorithm 1 from many random initializations; if the algorithm's output regularly has higher misclassification error than the global maximizer and does not converge to the true labels as $N$ grows to thousands, the practical claim of membership consistency would be refuted.","supporting_citations":[{"cited_title":"Consistency of community detection in networks under degree-corrected stochastic block models","cited_arxiv_id":"1110.3854","evidence_quote":"Gives consistency guarantees for the degree-corrected block model, the template for the paper's label-consistency theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes minimax-optimal recovery for a G-block covariance model that HBCM extends with heterogeneous scales."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the profile-likelihood approach for the block covariance model that motivates the likelihood formulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies treating the EM algorithm as coordinate ascent, which underlies the variational EM derivation."},{"cited_title":"and Horvath, S","cited_arxiv_id":null,"evidence_quote":"Benchmark method (WGCNA) used in simulations and real-data comparisons."}],"review_version":1}