{"id":"b2cd7e23-d1aa-4259-9bd4-12ffcf8a8f75","arxiv_id":"2602.19129","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new estimator with proven consistency and normality for latent sender/receiver positions and layer-specific connection matrices in generalized multilayer latent space models.","lead":"This paper builds a statistical model and estimation method for multilayer directed networks with binary, count, or continuous edges, giving each node separate sender and receiver roles and each layer its own interaction matrix. It proves consistency and asymptotic normality for these estimates, enabling confidence intervals and change-point tests, and demonstrates the method on world-trade data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 1.1's exact orthogonality is the load-bearing point: under local violations, Algorithm 1's column selection mixes Θ/Φ with degree factors, so the stated normal theory targets a different parameter.","rationale":"I agree with the reader's identification of Assumption 1.1 as the weakest assumption. The two-sided centering in (10) removes the intercept and degree-factor blocks only after the exact orthogonality conditions are imposed; otherwise the centered subspace is a mixture of Θ* (or Φ*) with degree-heterogeneity directions. Algorithm 1 then selects columns of the uncentered bU_m by projection onto the centered subspace, but it never projects the selected columns back, so the contamination is retained. Since Theorem 2's Gaussian limit is proved at the exactly orthogonal parameter point, even moderate departures could move the estimand and invalidate confidence regions and layer-equality tests. This is not merely a proof gap: it affects the interpretation of what the estimator targets. A dedicated local-misspecification simulation would settle the seriousness of the issue. The reader's other concerns — deferred proofs, ad hoc real-data preprocessing, and absent code — are real but secondary; they reinforce a conditional verdict rather than changing it. I therefore recommend leaving the reader's CONDITIONAL verdict unchanged.","tokens_in":21864,"tokens_out":22188,"duration_ms":217093,"concrete_test":"Simulate under model (2) with controlled misspecification: set Θ* = orthonormalize(Θ_0 + δ U*_β) and Φ* = orthonormalize(Φ_0 + δ U*_α) for δ ∈ {0, 0.01, 0.05, 0.1}, keeping all other Assumptions 1–2 satisfied. Run Algorithm 1 for n = 800, 1600 and T = 100, 200; compute ∥bΘ − Θ*R∥_{2,∞} after optimal sign/permutation alignment and the empirical coverage of 95% intervals from Corollary 1. If error or under-coverage does not vanish as n,T grow for δ > 0, Assumption 1.1 is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 1.1 (Θ* ⊥ span{1_n, U*_β}, Φ* ⊥ span{1_n, U*_α}) is what makes Lemma 1 and the two-sided centering in (10) identify Θ* and Φ* as specific columns of U_svd_1 and U_svd_2. In Algorithm 1, columns of bU_m are chosen by largest projection norm onto J_n bZ_m J^T_{n,T}; if the true Θ* has an O(δ) component along span{1_n, U*_β}, the selected columns inherit that component and the projection P_m used for selection does not remove it. The asymptotic covariance in Theorem 2 is derived at δ = 0, so confidence regions would be centered at a biased parameter, and the bias need not vanish as n,T grow. The paper provides no local-sensitivity analysis or misspecification simulation, and the real-data preprocessing (zero imports replaced by a large negative value) is far from the exact condition. This is the main reason the central claim remains conditional: consistency and normality are established for an exactly identified representative of the parameter, not for nearby data-generating processes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generalized multilayer latent space model for directed multilayer networks. Each node has two latent positions (sending and receiving), each layer has a connection matrix, and node/degree heterogeneity is modeled with low-rank factor structures. Estimation is performed by unfolding the adjacency tensor along two modes, solving constrained maximum likelihood low-rank matrix problems, selecting the latent-position columns via a two-sided centering projection, and then fusing the two unfoldings to estimate the connection matrices. The main theoretical contributions are consistency and asymptotic normality for the estimated latent positions and connection matrices, with confidence-region construction and layer-equality tests as applications. The method is validated in simulations and applied to international trade data.","tokens_in":22196,"tokens_out":17577,"duration_ms":149919,"significance":"If the theoretical results are correct, this is a valuable contribution to the multilayer network literature: it is one of the first inferential theories for nonlinear multilayer latent space models, it covers directed networks and multiple edge types, and the unfolding-and-fusion procedure is computationally more feasible than direct tensor optimization. The rates in Theorem 1 align with existing results in the COSIE and related literature. The paper also provides a clear path to confidence regions and structural-break tests. However, the central identification strategy rests on an exact orthogonality condition that is strong and not stress-tested; the practical scope of the claimed inference therefore remains conditional.","major_comments":[{"comment":"The exact orthogonality assumptions Θ*⊥span{1_n,U*_β} and Φ*⊥span{1_n,U*_α} are load-bearing. Equation (10) and Lemma 1 hold only under these conditions, and Algorithm 1 (step 4) selects columns of bU_m by projection onto the column space of J_n bZ_m J^⊤_{n,T}. If the true Θ* has an O(δ) component along span{1_n,U*_β} (and similarly for Φ*), the selected columns inherit that component and the centering does not remove it. The asymptotic covariances in Theorems 2 and 4 are derived at δ=0; the bias in the 'estimated latent positions' need not vanish as n,T grow. No local sensitivity analysis or misspecification simulation is provided, and the real-data preprocessing (zero imports replaced by a large negative value) is far from the exact condition. Please add a formal perturbation analysis (e.g., characterize the bias under a vanishing orthogonality violation and give conditions under which","section":"§2.2, Assumption 1.1; §2.3, Algorithm 1"}],"minor_comments":[{"comment":"The plug-in quantities bπ_{m,i,j} are defined via [M_m(bX)]_{i,j}, but the tensor bX is not explicitly defined in the main text. Please define it, e.g., as the tensor with entries bΘ_i^⊤ bΛ_t bΦ_j + bβ_{it}+bα_{jt}.","section":"§3.1, Corollary 1"},{"comment":"The treatment of zero import values ('set as a large negative value') is arbitrary. Since the theory assumes a correctly specified Gaussian link, the real-data analysis should report sensitivity to this preprocessing choice or justify the value used.","section":"§4.2, real data"},{"comment":"The notation in the derivatives, e.g., ∂²ℓ_{m,s,j}(π^*_{m,s,j})/∂π²_{m,s,j}, is confusing because the argument is also indexed. Use a dummy variable, e.g., ∂²ℓ_{m,s,j}(π)/∂π²|π=π^*_{m,s,j}.","section":"§2.2, Equations (16)–(17)"},{"comment":"The main text defers all proofs to the supplementary material. Given the complexity of the first-order expansions (especially for Theorem 4 and Corollary 2), the supplement should be carefully checked by the editor/reviewers for the joint asymptotic covariance claims.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The contribution is significant and the central results appear credible, but the exact-orthogonality assumption is a genuine load-bearing concern. I recommend that the editor ensure the supplementary proofs of Theorem 4 and Corollary 2 are thoroughly verified, and that the authors are asked to address the sensitivity of the method to Assumption 1.1 in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a serious, technically strong paper that delivers the first asymptotic normality results for connection-matrix (core tensor) estimators in nonlinear multilayer latent space models, covering directed edges with separate sender/receiver positions and general link functions. Prior work (Zhang et al. 2020b; He et al. 2025) had consistency but not distribution theory; Arroyo et al. 2021 relied on linear models. The new results are real and the rates match existing literature.\n\nThe method is also genuinely nice. The unfolding-and-fusion approach avoids large-scale tensor optimization, which would be impractical for non-continuous edge types. The simulations cover Gaussian, Poisson, and binary edges with reasonable coverage rates, and the COW trade analysis gives interpretable latent components and change points that line up with known historical events. The paper is honest about the identifiability issue and builds the theory around it rather than ignoring it.\n\nThe main soft spot is exactly what the stress-test note flags: Assumption 1.1, the exact orthogonality of Theta* and Phi* to the degree-heterogeneity spans, is load-bearing for Lemma 1 and for the column-selection step in Algorithm 1. If it holds only approximately, the selected columns mix degree-heterogeneity directions and the asymptotic covariance is centered at a biased parameter; that bias does not vanish as n and T grow. The paper provides no local-sensitivity analysis and no misspecification simulation. For a model intended for real network data, this matters. The trade application almost certainly violates the condition, and the preprocessing choices (unreported substitution value for zero imports, unspecified country selection) make the empirical claim harder to evaluate. The change-point testing also corrects for the k1*k2 comparisons but not for the successive time points. And there is no code, which compounds the reproducibility problem.\n\nThese are addressable rather than fatal. The theory is derived carefully, the rates are consistent with the field, and the inference results are genuinely new. The main text defers all proofs to the supplement, so a referee will need to check those, but that is standard practice.\n\nWho is this for? Anyone working on multilayer network inference or latent space models. It deserves a serious referee, and I would send it to review with a request for a sensitivity analysis around Assumption 1.1, a clearer real-data preprocessing description, and ideally code. My verdict is conditional, but the condition is met by revision rather than by rethinking the core idea.","headline":"Solid first inference theory for nonlinear directed multilayer latent space models, but the load-bearing exact orthogonality assumption is under-tested and the real-data analysis has some ad hoc choices.","tokens_in":22648,"tokens_out":2816,"would_cite":true,"duration_ms":29955,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes consistency and asymptotic normality for estimated latent positions and connection matrices in a broad class of multilayer directed network models with nonlinear link functions, enabling confidence regions for node po","keywords":["multilayer network","latent space model","Tucker tensor decomposition","asymptotic normality","confidence regions","structural change detection","degree heterogeneity"],"falsifier":"Simulate a multilayer network from the model with degree factors U*_β, U*_α correlated with Θ*, Φ* (violating Assumption 1.1), fit the proposed estimator, and check whether the empirical coverage of the nominal 95% confidence regions for [Θ*]_i, [Φ*]_i, and [Λ*_t]_ij stays near 95% as n and T grow; systematic undercoverage would demonstrate that the orthogonality condition is necessary for the inference claims.","tokens_in":21735,"feed_emoji":"📊","tokens_out":5866,"duration_ms":50338,"temperature":0.7,"pith_summary":"The paper claims that for multilayer directed networks with binary, count, or continuous edges — where edge probabilities or means are linked to latent structure through nonlinear link functions — the node-level sending and receiving latent positions and the layer-specific connection matrices can be estimated with known convergence rates and asymptotically normal distributions. This turns the multilayer latent space model from a tool for point estimation into one that supports confidence intervals for latent positions and hypothesis tests for structural equality between layers. The authors build their estimator by unfolding the network tensor along two modes, fitting two low-rank matrices by constrained maximum likelihood, then fusing the singular vectors to recover the connection matrices. Simulations and a trade-network application suggest the Gaussian approximations are reasonable in practice, and the method detects structural breaks such as the early-1990s collapse of the Soviet bloc.","feed_headline":"Estimator yields Gaussian limits for multilayer network structure","feed_subtitle":"Handles binary, count, and continuous edges, enabling layer-equality tests and structural change detection.","key_machinery":"The load-bearing device is the 'unfolding and fusion' estimator. The network tensor is unfolded along its two node modes into matrices that, after centering with J_n, have column spaces spanned by the latent positions Θ and Φ. Two constrained maximum-likelihood problems estimate these low-rank matrices; a two-sided centering operator removes the degree-heterogeneity terms; projecting and selecting the largest columns recovers the latent positions; and the connection matrices (the Tucker core) are obtained by a fusion step, M_1(Λ̂) = (V̂_1^c)ᵀ (I_T ⊗ Φ̂)/n. The centering operator is what makes the estimate of Θ and Φ immune to the intercept factors, and the fusion structure is what makes the","core_discovery":"The paper proposes a generalized multilayer latent space model in which each directed edge y_ijt is drawn from an exponential-family distribution with a signal x_ijt = θ_iᵀ Λ_t φ_j + β_it + α_jt, where θ_i and φ_j are sending and receiving latent positions and Λ_t is a layer-specific connection matrix. The signal, arranged as a tensor, admits a Tucker-type decomposition with the latent positions as loading matrices and the connection matrices as the core, plus degree-heterogeneity terms. The central claim is that both the latent positions and the connection matrices can be estimated by an 'unfolding and fusion' procedure — replacing one hard tensor optimization with two low-rank matrix likel","pith_inferences":["The orthogonality assumption (Assumption 1.1) that the latent positions are exactly orthogonal to the degree-heterogeneity column spaces is likely to be violated in real data; the paper does not quantify the resulting bias, and a sensitivity analysis with softly violated orthogonality would be a natural next test.","Because the estimator is based on two unfoldings rather than a joint tensor fit, it could be extended to missing layers or partially observed edges without changing the machinery.","The asymptotic normality of connection matrices opens the door to simultaneous confidence bands for the entire sequence Λ_t, which could be used for online change-point detection in streaming network settings.","The fusion step is a simple multiplication, so the approach could be adapted to weighted networks with arbitrary exponential-family noise; the theory only needs the log-likelihood's smoothness conditions."],"forward_implications":["For multilayer networks with binary, count, or continuous edge types, one can construct confidence regions for each node's latent position (up to sign) using the plug-in asymptotic covariance.","One can test whether two network layers share the same connection matrix — element-wise or globally with multiplicity control — enabling structural-change detection in dynamic networks.","The estimator avoids solving a non-convex tensor problem; it uses only two low-rank matrix optimizations, so it scales to larger networks than direct tensor methods.","The convergence rates match or improve on existing results in special cases: for a single layer, the uniform rate is O_p(n^{-1/2} log n), slightly better than previous latent-space results."],"fun_headline_variants":["Multilayer latent space model achieves asymptotic normality","Unfolding and fusion yields Gaussian limits for multilayer nets","Inference for multilayer networks with mixed edge types","Layer-equality tests via generalized latent space model","Consistent estimation and Gaussian inference for multilayer networks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true sending and receiving latent positions are exactly orthogonal to the column spaces of the layer-wise out-degree and in-degree heterogeneity factors; if that orthogonality only holds approximately, the estimated latent positions can mix in the degree directions and the derived normality may fail.","fun_headline_variants_meta":{"raw":{"variants":["Multilayer latent space model achieves asymptotic normality","Unfolding and fusion yields Gaussian limits for multilayer nets","Inference for multilayer networks with mixed edge types","Layer-equality tests via generalized latent space model","Consistent estimation and Gaussian inference for multilayer networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1001,"prompt_tokens":729,"completion_tokens":272,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":199}},"tokens_in":473,"tokens_out":272,"duration_ms":3418,"temperature":1.0,"reasoning_tokens":199,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:43:51.412011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a multilayer network from the model with degree factors U*_β, U*_α correlated with Θ*, Φ* (violating Assumption 1.1), fit the proposed estimator, and check whether the empirical coverage of the nominal 95% confidence regions for [Θ*]_i, [Φ*]_i, and [Λ*_t]_ij stays near 95% as n and T grow; systematic undercoverage would demonstrate that the orthogonality condition is necessary for the inference claims.","supporting_citations":[],"review_version":1}