{"id":"753a8377-688b-43de-bf66-ad140900236f","arxiv_id":"2502.06671","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The CAMM model with a convex MLE achieves, under a strong and mostly unverified assumption, near-optimal recovery rates for the covariate matrix and the mixed memberships.","lead":"The paper introduces a logistic network model in which connections depend on both hidden community membership and observable node attributes, and derives error bounds for recovering the covariate effect and the memberships from a single observed network. The rates are claimed to be optimal, but the main theorems hinge on an assumption that is essentially the proof's hardest step and is only checked, with a gap, for a two-community example.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 7, the convex/nonconvex bridge, is unverified for the general CAMM model; even its only SBM check contains an explicitly flagged unjustified step ('Why???' in Appendix H.1), so the convex guarantees in Theorem 3.2/3.7 do not follow.","rationale":"The Reader's weakest_assumption pointed to Assumption 7, and I reach the same conclusion independently. The central claim—near-optimal guarantees for the convex estimator—depends on the convex/nonconvex equivalence, which is precisely what Assumption 7 supplies. The paper itself flags the condition as 'nearly necessary and sufficient' for that equivalence, so a gap here is directly load-bearing. The only verification, Proposition 3.1, is a special case with H* = 0, and it contains an explicitly marked unjustified step in Appendix H.1. The manuscript also contains statements that the condition is 'typically easy to satisfy,' but no general argument or plausible mechanism is given; the spectral operator involves unknown ground-truth objects and the covariance structure Z, so it is not verifiable from data. The rate inconsistencies between the abstract, Theorem 3.2, and Remark 3.3 are secondary; even if those were resolved, the bridge assumption would still be unproven. I therefore agree that the paper should not be accepted as claiming these optimal convex guarantees. I would not change the Reader's verdict: the framework is plausible and the nonconvex analysis is detailed, but the key transfer step is not established. The proposed concrete test would either close the gap (if the numerical checks pass and the SBM step is re-derived) or decisively show the current proof is incomplete.","tokens_in":94173,"tokens_out":3619,"duration_ms":35913,"concrete_test":"Test the bridge directly in simulation. First, reproduce Proposition 3.1 numerically: for the two-community SBM with H* = 0 and n = 100, 200, 400, form D*, compute (P D* P)^dagger vec(0, X*, Y*), and evaluate sigma_3(P^perp_Z (1/n M* ⊙ (Z Δ_H Z^T + (Δ_X Y*^T + X* Δ_Y^T)/n)) P^perp_Z). Compare the computed (Δ_X, Δ_Y) to the closed form (81); if they differ or the singular value exceeds 1−ε, the 'Why???' step is invalid. Second, simulate CAMM with H* ≠ 0 and random Z satisfying Assumption 3, solve the convex program (2), and compare it to ground-truth-initialized nonconvex gradient descent from Section 4.1. If the two solutions diverge even when Assumptions 2–6 hold numerically, Assumption 7 fails and the convex guarantees are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the convex estimator (2) achieves the stated near-optimal rates. The proof strategy transfers guarantees from a hypothetical ground-truth-initialized nonconvex gradient path to the convex solution, and the transfer is exactly Assumption 7 (Section 3.1), which requires a spectral smallness condition on a data-dependent operator built from unknown ground-truth matrices P*, D*, X*, Y*. The paper asserts this condition is 'nearly necessary and sufficient' for convex/nonconvex equivalence, but it is never proved for the general CAMM model. The only verification, Proposition 3.1, treats a two-community SBM with H* = 0, and even that proof contains an explicitly flagged unsupported assertion: after Eq. (77) in Appendix H.1 the authors write 'Why???' before claiming the solution has the block-constant form (78). Without a valid derivation of that form, the SBM check does not go through. Because Assumption 7 is stated in terms of unknown ground truth and cannot be checked from data, and because no argument shows it holds when H* is nonzero or the covariate matrix is general, the convex guarantees in Theorems 3.2 and 3.7 are not actually established. If Assumption 7 fails, the nonconvex analysis does not transfer to the convex program, and the central claim collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Covariates-Adjusted Mixed-Membership (CAMM) model, in which edge probabilities are logistic functions of a quadratic covariate form z_i^T H* z_j plus a low-rank DCMM component Gamma*. It studies the convex regularized MLE (2) and claims near-optimal rates for H*, Gamma*, and the mixed memberships Pi*: O(sqrt(log n/n)) for H* and entrywise Gamma*, O(1) for Frobenius Gamma*, and O(sqrt(log n/n)) for the membership vectors in the bounded-parameter regime. The proof strategy is to analyze a hypothetical nonconvex gradient descent initialized at the ground truth, then to transfer its guarantees to the convex solution via a debiasing argument and a bridge theorem (Theorem H.1). The transfer is asserted to hold under Assumption 7, which is described as nearly necessary and sufficient for the equivalence of convex and nonconvex solutions. Simulation results confirm the rates in one setting (r=2, p=3), and a real-data analysis of S&P 500 stocks is presented.","tokens_in":94397,"tokens_out":3124,"duration_ms":31436,"significance":"If the main theorems were fully established, the paper would provide the first optimal-rate guarantees for convex mixed-membership estimation with binary edges and covariates, extending the nonconvex-to-convex transfer framework of Chen et al. (2020) to logistic loss and to general convex losses. The paper is also praiseworthy for its explicit goal of entrywise and 2-infinity control, its use of debiasing to identify the dominant error term, and its detailed leave-one-out analysis. However, the central bridge assumption (Assumption 7) is not verified for the general model, and the only special-case verification contains an explicitly flagged unjustified step. Because the main theorems are conditional on that unverified bridge, the paper's central contribution is currently unsupported. The manuscript therefore cannot be accepted in its present form.","major_comments":[{"comment":"The proof of the convex guarantees is entirely dependent on Assumption 7, which the paper itself describes (Section 4.2 and the paragraph after Assumption 7) as nearly necessary and sufficient for the equivalence between the convex estimator and the hypothetical nonconvex solution. No general verification of Assumption 7 is provided; it is stated in terms of unknown ground-truth quantities (D*, M*, X*, Y*, P*) and cannot be checked from data. Since the claimed optimal rates in Theorems 3.2 and 3.7 are for the convex program (2), the absence of a provable bridge means the central claim is conditional on an unverified, data-independent hypothesis. This is a load-bearing gap.","section":"Section 3.1, Assumption 7; Theorems 3.2 and 3.7"},{"comment":"The only special-case verification of Assumption 7, for a two-community SBM with H*=0, is incomplete: after Eq. (77) the text reads 'Why???' before asserting that the solution must have the block-constant form (78). That block-constant form is then used to compute Delta and to verify the spectral bound sigma_3(...) < 1 - epsilon. Without a valid derivation of (78), the verification of Proposition 3.1 does not go through, and the paper's statement that Assumption 7 'holds in common settings such as stochastic block models' is unsupported. This is not a presentation issue; it is an explicit gap in the only evidence offered for the bridge assumption.","section":"Appendix H.1, proof of Proposition 3.1"},{"comment":"The bridge theorem (Theorem H.1) bounds the distance between the convex and nonconvex solutions by the projected gradient norm, but the proof of Claim H.3 repeatedly invokes Assumption 7 to control the (r+1)-th singular value of the Hessian-related matrix. No argument is given that Assumption 7 holds for the general CAMM model with nonzero H* and general covariates. The paper also claims that the analysis extends to all convex loss functions, but this claim is not substantiated beyond a remark; no formal statement or proof is given for the general loss case. These points, together, mean that the key innovation advertised in the abstract and introduction—the identification of a condition that bridges convex and nonconvex solutions—is not established for the model under study.","section":"Section 4.2 and Theorem H.1"},{"comment":"The simulation study is too narrow to validate the claimed optimal rates: it fixes r=2 and p=3, uses lambda = sqrt(n) without explaining how lambda is chosen, reports no confidence intervals or slope estimates, and does not compare against any competing estimator. More importantly, the simulations do not verify Assumption 7, so they do not probe the validity of the bridge assumption. As a result, the simulations provide only weak empirical support for the central theoretical claim, even if the rates are correct under the assumptions.","section":"Section 5 and Theorem 3.2/3.7"}],"minor_comments":[{"comment":"The paper defines vec(A) by stacking rows of A rather than columns; this is an unusual convention and should be flagged explicitly in the notation section, since readers of matrix completion papers will expect column-wise vectorization.","section":"Notation, Section 1"},{"comment":"The real-data analysis uses several ad hoc preprocessing choices (e.g., the edge threshold Sigma_ij > 0.16, lower-bound capping of covariates, and the finite-sample heuristic test based on 1000 shuffles) that are not given a formal justification. These choices are fine for an illustrative analysis, but the paper should present them as exploratory rather than as confirmatory evidence for the model.","section":"Section 6"},{"comment":"The definition of M* and Delta involves the pseudoinverse of PD*P and the ground-truth matrices; the paper should state explicitly that Assumption 7 is a condition on the true generating mechanism and not on the data. A short discussion of when one might expect it to hold (beyond the SBM example) would improve readability.","section":"Section 3.1, Assumption 7"},{"comment":"The presence of the literal string 'Why???' in a published manuscript is highly unusual. Even if the step were correct, the authors should provide the missing justification; leaving the question mark in the text signals incompleteness and is not acceptable in a formal proof.","section":"Appendix H.1, Eq. (77)"}],"recommendation":"reject","confidential_remarks":"The manuscript falls within the scope of math.ST and the authors have clearly invested substantial effort in the leave-one-out analysis and the debiasing arguments. However, the 'Why???' in Appendix H.1 is a red flag that the special-case verification of Assumption 7 is incomplete, and the dependence of the main theorems on that assumption is not otherwise addressed. In my view this is not a mere technical gap that a local revision can fix; the paper would need either a full verification of Assumption 7 under a nondegenerate covariate model, or a proof that the convex rates hold without the bridge condition. The citation pattern and related-work section are appropriate, but the manuscript appears to have been submitted before the proof was fully checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the Fan-Ge-Hou paper on covariate-adjusted mixed-membership estimation. The model is natural: logistic DCMM with node covariates and an orthogonality constraint for identifiability. That is a real contribution, and the proof machinery—nonconvex gradient descent plus a debiased Newton step to bridge to the convex solution—is serious. Extending Chen et al.'s bridging technique from MSE to logistic loss is a nontrivial technical task, and the paper treats it honestly, even flagging a gap in its own appendix with a \"Why???\".\n\nThe load-bearing problem is Assumption 7. The proof transfers guarantees from a hypothetical ground-truth-initialized nonconvex path to the convex estimator, and that transfer is exactly Assumption 7. But Assumption 7 is never verified for the general model; it is stated in terms of unknown ground-truth matrices, so it cannot be checked from data. The only special-case verification, a two-community SBM with H*=0, contains an explicitly unjustified step: the \"Why???\" before the block-constant solution form. Without Assumption 7, Theorems 3.2 and 3.7 do not follow. The paper itself says Assumption 7 is \"nearly necessary and sufficient\" for the convex/nonconvex equivalence, which makes the gap central, not cosmetic.\n\nThe rate statements are also not fully aligned: the abstract, Theorem 3.2, and Remark 3.3 present the Gamma Frobenius rate differently, and \"optimal\" is asserted without lower bounds. The promised generalization to all convex losses and ranking problems is only a sketch. These are real soft spots, though minor compared with Assumption 7.\n\nThe nonconvex analysis itself is long and appears careful; I did not find a similar gap in the local geometry or leave-one-out arguments. That part is valuable and likely salvageable.\n\nBottom line: this paper deserves a serious referee. The model is plausible, the analysis is ambitious, and the core idea could be right. But the central claim is not established as written. Reviewers should focus on Assumption 7: prove it for a meaningful class (e.g., SBM with covariates and general H*), or replace it with a checkable condition, and resolve the \"Why???\" step. If that repair succeeds, this becomes an important paper. As it stands, I would not cite it yet.","headline":"Plausible framework, unproven bridge: the convex guarantees are conditional on an unverified Assumption 7, so the main claim does not follow as written.","tokens_in":95027,"tokens_out":2352,"would_cite":false,"duration_ms":22588,"reading_group":"maybe","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":"This paper claims that a convex logistic-likelihood estimator recovers both the covariate interaction matrix and the latent mixed-membership structure of a binary network at near-optimal rates, and that the proof can be carried out by…","keywords":["mixed-membership estimation","community detection","node covariates","degree-corrected mixed membership","convex relaxation","nonconvex optimization","logistic network model","nuclear norm regularization"],"falsifier":"Solve the linear system (77) in Appendix H.1 for the two-community example and check whether the solution really has the constant-block form asserted with the annotation 'Why???'; if not, Proposition 3.1, the only stated verification of Assumption 7, fails. More generally, simulate a CAMM instance satisfying Assumptions 2–6 but with $\\sigma_{r+1}$ of the Assumption 7 matrix larger than $1-\\epsilon$, and compare the convex solution with the ground-truth-initialized gradient path; a divergence between them at the predicted $1/\\sqrt{n}$ separation would refute the equivalence step and the stated rates for the convex estimator.","tokens_in":93827,"feed_emoji":"🕸️","tokens_out":6560,"duration_ms":56170,"temperature":0.7,"pith_summary":"This paper proposes the Covariates-Adjusted Mixed Membership (CAMM) model, in which each edge of a binary network is a Bernoulli draw with logistic probability $\\exp(z_i^\\top H^* z_j + \\Gamma^*_{ij})/(1+\\exp(z_i^\\top H^* z_j + \\Gamma^*_{ij}))$, where $H^*$ is a covariate interaction matrix and $\\Gamma^*=\\Theta^*\\Pi^*W^*\\Pi^{*\\top}\\Theta^*$ is the low-rank degree-corrected mixed-membership term. The paper's claim is that the constrained convex regularized maximum-likelihood estimator, with logistic loss plus nuclear norm on $\\Gamma$ and $\\Gamma$ forced orthogonal to the covariate span, estimates $H^*$ at a near-$1/\\sqrt{n}$ Frobenius rate, $\\Gamma^*$ at a $1/\\sqrt{n}$ entrywise rate, and the mixed memberships $\\Pi^*$ at a $1/\\sqrt{n}$ row-wise rate. Because the convex program cannot be analyzed directly, the proof analyzes a hypothetical nonconvex gradient descent initialized at ground truth and transfers its guarantees through a debiasing step and an equivalence assumption linking convex and nonconvex solutions. If the transfer holds, this would be the first optimal-rate guarantee for binary-edge mixed-membership estimation with covariates.","feed_headline":"Convex MLE recovers network memberships at optimal rates","feed_subtitle":"A logistic model with covariates estimates the network component and memberships to a 1/sqrt(n) rate, if convex and nonconvex solutions…","key_machinery":"The load-bearing machinery is the triple: the convex program (2) with logistic loss and nuclear norm, the nonconvex reparametrization $\\Gamma=XY^\\top$ with $P_Z X=P_Z Y=0$, motivated by $\\|\\Gamma\\|_*=\\min_{XY^\\top=\\Gamma}\\tfrac12(\\|X\\|_F^2+\\|Y\\|_F^2)$, and the bridge theorem (Theorem H.1) that bounds the distance between the convex minimizer and the nonconvex solution by the projected gradient norm at the nonconvex point. The bridge uses a debiased estimator, namely one Newton step from the nonconvex solution, and a Taylor expansion of the logistic gradient around ground truth, with Assumption 7 giving the crucial spectral gap $\\sigma_{r+1}\\big(P_Z^\\perp \\tfrac1n M^*\\odot(\\cdot)P_Z^\\perp\\big)<1-\\epsilon$ that makes the convex solution's subgradient condition hold.","core_discovery":"The central discovery, stated as Theorem 3.2 and Theorem 3.7, is that the convex estimator (2) achieves $\\|\\hat{H}_c-H^*\\|_F \\lesssim \\lambda\\sqrt{\\mu r\\kappa}/(n\\sigma_{\\min})$, $\\|\\hat{\\Gamma}_c-\\Gamma^*\\|_F \\lesssim \\lambda\\kappa\\sqrt{\\mu r}$, $\\|\\hat{\\Gamma}_c-\\Gamma^*\\|_\\infty \\lesssim \\mu r\\kappa(\\lambda\\sigma_{\\max}/(n^2\\sqrt{\\mu r})(1+n/\\sigma_{\\max})+\\kappa\\sqrt{\\log n/n})$, and max row error $\\max_i\\|\\hat{\\pi}_i-\\pi^*_i\\|_1 \\lesssim \\lambda(\\kappa^{1.5}\\sqrt{\\mu r}+\\sqrt{\\mu}\\kappa r^{5/4})(\\mu\\kappa^{0.5}/\\beta_n+\\kappa\\mu^{1.5}\\sqrt{\\beta_n})(r/(\\sqrt{n}\\,\\bar\\theta^*_2))^2$, with $\\lambda\\gtrsim \\epsilon^{-1}(1+\\mu r\\sigma_{\\max}/n)\\sqrt{n\\log n}$. In the bounded parameter regime these become $\\sqrt{\\log n/n}$ for $H$ and entrywise $\\Gamma$, constant Frobenius error for $\\Gamma$, and $\\sqrt{\\log n/n}$ for memberships. The argument establishes a chain: a ground-truth-initialized nonconvex gradient path stays within $1/\\sqrt{n}$ of the truth; a Newton–Raphson debiased step isolates the systematic bias; and Assumption 7, controlling the $(r+1)$-th singular value of a projected weighted matrix, ensures the convex minimizer of (2) and the nonconvex solution coincide, transferring the nonconvex rates to the convex estimator.","pith_inferences":["Editorial inference: Assumption 7 implicitly requires that the covariate-adjusted logistic likelihood have a unique well-separated low-rank component after projection; a data-driven proxy could be constructed by plugging estimated parameters into the same matrix and thresholding its $(r+1)$-th singular value, which would make the theory testable on real data.","Editorial inference: if the flagged 'Why???' step in the two-community verification cannot be justified, the evidence for Assumption 7 is thinner than the theorems assume; the main results would still hold conditional on Assumption 7, but the claim that the assumption is 'nearly necessary and sufficient' and 'easy to satisfy' would remain unsubstantiated.","Editorial inference: the same debiased-Newton bridge should adapt to logistic matrix completion and to networks with missing edges, where the noise is heteroskedastic in the same way; testing those settings would show how far the 'any convex loss' claim reaches."],"forward_implications":["If Theorems 3.2 and 3.7 are correct, the CAMM estimator achieves the minimax-type $1/\\sqrt{n}$ rates jointly for covariate effects and mixed memberships in binary networks, so covariates no longer need to be set aside or treated as noise.","Entrywise control of $\\hat\\Gamma_c$ makes the subsequent vertex-hunting step (Mixed-SCORE style Algorithm 1) recover $\\Pi^*$ at $1/\\sqrt{n}$, which is the rate needed for inferential use of membership profiles.","The convex–nonconvex transfer is proved for the logistic loss and claimed to extend to every convex loss, so pairwise-comparison ranking problems with covariates could inherit the same guarantees.","The simulations and the S&P 500 analysis indicate the model is tractable in practice: a Nesterov-accelerated nuclear-norm program plus vertex hunting, with covariates explaining about 59% of the fitted interaction in the stock network."],"supporting_citations":[{"why":"Supplies the leave-one-out nonconvex analysis and the convex-to-nonconvex transfer framework that this paper adapts from MSE loss to logistic loss.","marker":"Chen et al. (2020)"},{"why":"Provides the Mixed-SCORE vertex-hunting procedure used as Algorithm 1 to reconstruct memberships from $\\hat\\Gamma_c$.","marker":"Jin et al. (2017)"},{"why":"Supplies the DCMM identifiability proof structure and the technical lemmas about eigen-ratios and vertex hunting used in Theorem 3.7.","marker":"Jin et al. (2023)"},{"why":"Justifies replacing the nuclear norm by the balanced $\\tfrac12(\\|X\\|_F^2+\\|Y\\|_F^2)$ reparametrization in the nonconvex problem.","marker":"Srebro and Shraibman (2005)"},{"why":"Gives the pseudoinverse perturbation bound used to control the difference between the debiased and ground-truth Newton steps.","marker":"Stewart (1977)"},{"why":"Provides the rotation-alignment and matrix-sign lemmas used to keep the nonconvex iterates aligned with the truth.","marker":"Ma et al. (2018)"},{"why":"Supports the spectral-regularization view of nuclear-norm minimization used in the convex formulation.","marker":"Mazumder et al. (2010)"},{"why":"Supplies the Davis–Kahan-type rotation control used in the membership reconstruction proof.","marker":"Yan et al. (2024)"}],"fun_headline_variants":["Convex MLE hits optimal rates with covariate-adjusted networks","Covariate-boosted MLE nails network membership recovery","Optimal mixed-membership estimation via convex and nonconvex bridge","Network memberships recovered at 1/sqrt(n) rate by convex MLE","Assumption bridges convex to nonconvex for optimal network fits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 7, that a certain matrix built from unknown ground-truth parameters has its $(r+1)$-th singular value below $1-\\epsilon$; this is what transfers the nonconvex guarantees to the convex estimator, it cannot be checked from data, and its only verification is a two-community example with an explicitly flagged unjustified step.","fun_headline_variants_meta":{"raw":{"variants":["Convex MLE hits optimal rates with covariate-adjusted networks","Covariate-boosted MLE nails network membership recovery","Optimal mixed-membership estimation via convex and nonconvex bridge","Network memberships recovered at 1/sqrt(n) rate by convex MLE","Assumption bridges convex to nonconvex for optimal network fits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2945,"prompt_tokens":1193,"completion_tokens":1752,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":809,"completion_tokens_details":{"reasoning_tokens":1660}},"tokens_in":809,"tokens_out":1752,"duration_ms":12487,"temperature":1.0,"reasoning_tokens":1660,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:45:24.081281+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the linear system (77) in Appendix H.1 for the two-community example and check whether the solution really has the constant-block form asserted with the annotation 'Why???'; if not, Proposition 3.1, the only stated verification of Assumption 7, fails. More generally, simulate a CAMM instance satisfying Assumptions 2–6 but with $\\sigma_{r+1}$ of the Assumption 7 matrix larger than $1-\\epsilon$, and compare the convex solution with the ground-truth-initialized gradient path; a divergence between them at the predicted $1/\\sqrt{n}$ separation would refute the equivalence step and the stated rates for the convex estimator.","supporting_citations":[],"review_version":1}