{"id":"496b655f-a197-4cc2-b7eb-ad30531dbf79","arxiv_id":"2607.29296","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A provably consistent top-down procedure that recovers latent hierarchical clusters of networks from hierarchical distance matrices.","lead":"A new algorithm clusters collections of networks and simultaneously recovers the tree-like hierarchy among the clusters. It splits networks with spectral clustering and uses a two-sample test to decide when to stop, without knowing the number of clusters in advance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof is invalid: the constructed B = κ11^T + L_D has all-ones as dominant eigenvector, so Lemma 1 cannot yield the claimed sign-separating bipartition.","rationale":"The paper's central claim is that NHCL exactly recovers the latent hierarchy when D is an HDM (Theorem 1 / Corollary 1) and consistently in the empirical setting (Theorems 3–4). The most load-bearing concern I find is not the separation assumption itself — which is explicitly acknowledged and even mildly relaxed in Section 4.3 — but the proof of Theorem 1. The construction of B = κ11^T + L_D is intended to make the off-diagonal blocks negative and the diagonal blocks positive so that Lemma 1 applies. However, adding κ11^T inevitably makes the all-ones vector an eigenvector with eigenvalue κm, and this eigenvalue is typically the largest. The proof's assertion that this cannot be the dominant eigenvector because its signs are not mixed is logically backwards: Lemma 1's conclusion is exactly what must be proved, and in this case it is false. This is a concrete, internally checkable error, not a matter of external consensus. If Theorem 1 is unproven, the recursive exact-recovery guarantee and the theoretical foundation for the empirical consistency theorem are unsupported. That said, the theorem may be true and fixable by analyzing the largest eigenvector of L_D restricted to the subspace orthogonal to 1; the simulations are also suggestive. The reader already assigned CONDITIONAL, partly citing a spectral simplicity issue, so my concern reinforces that verdict rather than moving it. I therefore recommend UNCHANGED (i.e., keep CONDITIONAL), with the understanding that the proof of Theorem 1 must be corrected before the central claim can be regarded as established.","tokens_in":29065,"tokens_out":18901,"duration_ms":183385,"concrete_test":"Verify the proof step directly on the minimal HDM example: take m=2, D = [[0,a],[a,0]], choose κ=a/2, and compute the dominant eigenvector of B = κ11^T + L_D. It will be (1,1)^T, not (1,-1)^T, so Lemma 1's conclusion fails for this B. To test whether Theorem 1 itself might still be true, exhaustively enumerate small HDM matrices (e.g., two clusters of sizes 2 and 2, all within-cluster entries ≤ β0, all between-cluster entries ≥ α1 > β0, with large β1 allowed), and for each compute the dominant eigenvector of L_D after removing the all-ones direction. Check whether its sign pattern always equals the true bipartition. If any counterexample appears, Theorem 1 is false; if none appears, the theorem may be salvageable but the current proof remains incorrect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central exact-recovery result (Theorem 1 / Corollary 1) rests on the proof in Appendix S.2. The authors define B = κ11^T + L_D with κ ∈ (β0, α1), claim B has a Positive-Negative block structure, and invoke Lemma 1 to assert that the dominant eigenvector of B has strictly positive entries in one cluster and strictly negative in the other. But B1 = κm1, so the all-ones vector is an eigenvector of B with eigenvalue κm. Since L_D is positive semidefinite, all other eigenvalues of B are those of L_D on the subspace orthogonal to 1. For any nonempty split, 1 is not orthogonal to 1, so the dominant eigenvector of B is typically 1 itself — not the sign-separating vector. The proof then says 'the signs of 1 are the same, implying u1 ≠ 1,' which is exactly the opposite of what follows: if the dominant eigenvector were 1, its signs are indeed all the same. This is not a minor gap: Lemma 1 as stated is false for such B. For example, with D = [[0,a],[a,0]], L_D has dominant eigenvector [1,-1]^T, but B = κ11^T + L_D with κ=a/2 has dominant eigenvector [1,1]^T. Thus the proof of Theorem 1 does not go through, and the paper does not provide an alternative argument showing that the largest eigenvector of L_D (in the subspace orthogonal to 1) has the required sign pattern. Since Corollary 1 and the theoretical basis of Theorem 4 depend on Theorem 1, the central claim is not established by the submitted proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Hierarchical Distance Matrix (HDM), a recursive separation condition on population-level distance matrices that encodes a latent binary tree over populations of networks. It proposes NHC-TST, a top-down algorithm that recursively partitions networks by spectral clustering and uses a graph-based two-sample test as a data-driven stopping rule. The main theoretical claims are: (i) Theorem 1/Corollary 1, exact recovery of each population-level bipartition and hence the full hierarchy by the Laplacian spectral method NHCL when the true distance matrix is an HDM; (ii) Theorem 2/Corollary 2, a parallel guarantee for the distance-matrix method NHCD; (iii) Theorem 3, consistency of NHCL with an estimated distance matrix under uniform link-probability estimation error; and (iv) Theorem 4, exact recovery with probability tending to 1 for the full NHCL-TST procedure. The paper also reports extensive simulations and an application to global migration networks.","tokens_in":29497,"tokens_out":19942,"duration_ms":213621,"significance":"If the theoretical claims are established, the HDM framework is a genuinely useful formalization of latent hierarchy for network populations, and the proposed recursive, test-based procedure is an appealing alternative to flat clustering methods. The paper's strengths include a clearly stated structural model, explicit separation assumptions, a data-driven stopping rule that avoids prespecifying the number of clusters, and a broad simulation study with a hierarchy-aware cophenetic evaluation metric. The migration application is also interpretable. The central proofs, however, currently contain load-bearing gaps: the proof of Theorem 1 does not rigorously transfer the sign pattern from an auxiliary matrix to the Laplacian, and the proof of Theorem 2 contains an invalid inequality in the Rayleigh-quotient argument. These gaps appear repairable with existing ingredients, so the paper is promising but not yet ready in its present form.","major_comments":[{"comment":"The proof constructs B = κ11^T + L_D and invokes Lemma 1 to give the dominant eigenvector u1 a mixed sign pattern. To conclude that u1 is the dominant eigenvector of L_D, the proof must show that u1 lies in the subspace orthogonal to 1. The text argues: 'Observe that the signs of 1 are the same, implying u1 ≠ 1. Consequently, u1 must belong to the set of remaining eigenvectors orthogonal to 1.' The inference from u1 ≠ 1 to u1 ⊥ 1 requires that the eigenvalue κm of B is distinct from λ1(B) and that the dominant eigenspace is simple; this is not established. The gap is repairable: since κ ∈ (β0, α1), the matrix SBS with S = diag(I_{m0}, -I_{m1}) is entrywise positive, so Perron-Frobenius gives a simple dominant eigenvector; alternatively Lemma 5 gives λ1(L_D) ≥ mα1 > κm. But as written, Theorem 1, Corollary 1, and the population-level basis of Theorem 4 are not proven.","section":"Appendix S.2 (proof of Theorem 1)"},{"comment":"The central displayed inequality in the mixed-sign case and its predecessor replace sums of products |u_i||u_j| by cardinalities of sign classes. For the positive within-block terms, the upper bound by β0 times the product of counts is valid because each |u_i| ≤ 1. But for the between-block terms, the proof effectively lower-bounds Σ_{i∈I+,j∈J+} |u_i||u_j| by m_{0+} m_{1+} (and similarly for the other cross term), which is not valid for a unit vector; such a sum can be arbitrarily small. Consequently the derived condition α1 > β0 + M is not established by the displayed algebra. Since Theorem 2 is the exact-recovery guarantee for the NHCD variant and underpins Corollary 2, this needs a rigorous argument or a corrected sufficient condition.","section":"Appendix S.3 (proof of Theorem 2)"},{"comment":"Lemma 5, used repeatedly in the proof of Theorem 3 and in any repair of Theorem 1, asserts after constructing the HCMs Hα and Hβ that Hα,ij ≤ Dij ≤ Hβ,ij for all pairs. This entrywise comparison is not immediate from Definition 4, which specifies only the constant between-block scalars at each internal node and does not explicitly define the entries of the diagonal blocks for pairs that split at lower levels. The comparison can likely be made correct by defining Hβ,ij as the maximal between-block constant on the path where i,j first separate and Hα,ij analogously, but the proof as written omits this construction. Because the lemma is load-bearing, it should be stated and proved explicitly.","section":"Appendix S.4 (Lemma 5 and Definition 4)"}],"minor_comments":[{"comment":"The pseudocode is ambiguous: line 7 tests a p-value but then line 8 calls BIPARTITION again although a partition was already produced in line 3. Please clarify whether BIPARTITION simply returns the existing split or refers to a different procedure, and make the stopping logic explicit.","section":"Algorithm 1, lines 7–8"},{"comment":"The displayed bounds contain a typographical ambiguity: 'c1√mc2 ≤ |u_i| ≤ c2/2√m' should presumably read c1/(c2√m) ≤ |u_i| ≤ c2/(2√m).","section":"Lemma 6, Eq. (S.9)"},{"comment":"The notation 'lim_{n→∞}^{m_x=o(n^ε)} P(E_T)=1' is malformed; E_T is not defined. Also, internal nodes with δ_x = 1 cannot occur at depth L, so the summation over l = 0,...,L should be adjusted or explained.","section":"Proof of Theorem 4, final line"},{"comment":"Assumption 7 (λ1(L_D) > deg_i(D) for all i) is used in the proof of Lemma 6 and Theorem 3, but no discussion is given of when it is satisfied. A brief comment that it is a mild regularity condition, or a sufficient condition, would improve the paper.","section":"Assumption 7"},{"comment":"The reference list contains formatting inconsistencies, e.g. 'A. Spielman, D. (2009). Lecture Notes on Random Matrix Theory.'","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The HDM idea and the recursive testing framework are original and potentially publishable. The main obstacle is the rigor of the exact-recovery proofs: Theorem 1 has a fixable spectral-theoretic gap, while Theorem 2's proof contains a more serious invalid inequality. I recommend major revision rather than rejection because the gaps appear repairable with ingredients already in the paper (Perron-Frobenius, HCM bounds, and a reworked Rayleigh-quotient argument). The simulations and application are extensive and support the practical interest of the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The bottom line: this is a real contribution, and the central proof concern raised in the stress-test note does not hold up. The paper deserves a serious referee, with a request for code and data.\n\nWhat's new: Hierarchical Distance Matrices give a clean formal target for latent tree structure in populations of networks, and the NHC-TST algorithm—spectral bipartitioning plus a two-sample-test stopping rule—is a genuine construction, not an incremental tweak. The theoretical package (exact recovery at the population level, consistency with plug-in distance estimates) is ambitious and mostly delivered. The adapted CPCC for evaluating dendrogram recovery is a useful idea. Simulation evidence is strong across dense and sparse settings, and the non-uniform simulation in Section 4.3 is honest: it shows the method still works when the HDM assumption is mildly violated.\n\nOn the stress-test note: its counterexample is incorrect. For D = [[0,a],[a,0]], B = κ11^T + L_D with κ=a/2 has eigenvalues 2a and a; the dominant eigenvector is [1,-1], not [1,1]. More generally, the proof excludes the all-ones eigenvector because κm < λ_1(L_D) (Lemma 5 gives λ_1(L_D) ≥ mα1 and κ < α1). So the claimed fatal gap is not there.\n\nWhere the paper is genuinely soft: (1) Lemma 1 requires strictly positive/negative entries in the dominant eigenvector, but HDM allows zero within-cluster distances; zero entries in the eigenvector could break the strict sign pattern. Fixable by a perturbation argument or a slightly weaker conclusion. (2) The simplicity of the dominant eigenvalue is not explicitly established; repeated eigenvalues would require more care. Both are refinements, not fatal flaws. (3) The stopping-rule consistency in Theorem 4 leans on the specific two-sample test from Chen et al. (2024)—if that test's assumptions fail, the adaptive depth selection is shakier. (4) No code or data are provided; for a simulation-heavy paper, that is a real handicap for verification.\n\nWho should read it: anyone working on network-valued data, spectral clustering, or hierarchical mixture models. The HDM framework will likely be a useful reference point.\n\nRecommendation: send it to peer review. Ask the authors for code/data, a careful treatment of zero within-cluster distances, and a proof or reference for Lemma 1's simplicity condition. The core result stands.","headline":"New HDM framework and NHC-TST algorithm with sound central theory; the stress-test note's counterexample is wrong, and remaining concerns are patchable. Deserves peer review.","tokens_in":29933,"tokens_out":7084,"would_cite":true,"duration_ms":68988,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that when the population-level distance matrix of a collection of networks satisfies a recursive separation condition, a spectral top-down clustering algorithm exactly recovers the latent binary tree of clusters, and a two","keywords":["Hierarchical Distance Matrix","network clustering","spectral clustering","two-sample test","latent hierarchy","graphon estimation","multiple networks","top-down clustering"],"falsifier":"Generate a population of networks from an unbalanced tree where a deep split has a larger population distance than a shallow split (violating HDM monotonicity), e.g., the mirrored version of Section 4.3, and check whether the dominant Laplacian eigenvector's signs separate the true root branches; the paper's own Theorem 1 would not apply, and the split should fail at the root.","tokens_in":28934,"feed_emoji":"🌳","tokens_out":6079,"duration_ms":57782,"temperature":0.7,"pith_summary":"The authors introduce the Hierarchical Distance Matrix, a structural condition on the population-level pairwise distances between networks: at every internal node of an unknown binary tree, every between-branch distance must exceed every within-branch distance. They show that under this condition, the sign pattern of the dominant eigenvector of the associated Laplacian encodes the correct bipartition, so recursively splitting via spectral clustering exactly reconstructs the full hierarchy. In the empirical setting where distances are estimated from link probabilities, they prove sign recovery with high probability and, when a graph-based two-sample test is used as a stopping rule, exact recovery of the hierarchy with probability tending to one. A reader should care because this gives a formal, provably consistent way to recover latent hierarchical organization among populations of networks, without the user specifying the number of clusters or tree depth in advance.","feed_headline":"Spectral splits exactly recover latent network trees","feed_subtitle":"Recursive separation of distances makes a spectral top-down method provably consistent, with no pre-specified cluster count.","key_machinery":"The central object is the Hierarchical Distance Matrix (HDM): a symmetric population-level distance matrix that, after a suitable permutation, can be recursively block-partitioned so that the off-diagonal (between-branch) block strictly dominates the diagonal (within-branch) blocks at every internal node. The carrying mechanism is the spectral sign check: for the Laplacian L_D = diag(D1)-D, the dominant eigenvector has entries whose signs exactly separate the two branches; the proof uses a positive-negative block matrix lemma showing that this eigenvector has a strict sign pattern, plus Davis-Kahan bounds to transfer the result to estimated distances.","core_discovery":"The paper's central claim is that a population of networks generated from K latent link-probability matrices arranged on an unknown binary tree is exactly recoverable — both cluster labels and the tree topology — whenever the population-level distance matrix D is a Hierarchical Distance Matrix. This means that after permuting rows and columns, every internal submatrix can be split into two diagonal blocks and an off-diagonal block such that the off-diagonal entries are strictly larger than all entries inside the diagonal blocks, recursively down to the leaves. Under that condition the sign pattern of the dominant eigenvector of the Laplacian L_D = diag(D1)-D coincides with the first bipartit","pith_inferences":["A direct corollary of the proof structure is that any consistent estimator of the population distance matrix — not only Frobenius distances of link probabilities — can be plugged into the same spectral splitting scheme; the framework is tied to the distance matrix, not to graphon estimation per se.","The HDM condition resembles a 'tree metric with layer-separated ranges' and is likely satisfied when signal components decay geometrically down a tree, suggesting the method transfers to hierarchical stochastic block models, topic hierarchies, and other nested-network generative models beyond the two simulation families.","The finite-sample bottleneck is the two-sample test's power at internal nodes; since the stopping rule must reject at each true internal split, the effective sample size per node limits recoverable depth — a user-facing trade-off the paper does not fully explore.","Section 4.3's mild HDM violation still yields near-perfect CPCC, hinting that the strict monotonicity condition may be stronger than necessary; an explicit tolerance theorem for 'approximately HDM' matrices would be a natural next step."],"forward_implications":["Theoretical exact recovery: if D is an HDM, NHCL recovers every bipartition exactly (Theorem 1), and recursively the full tree (Corollary 1); NHCD requires the additional separation conditions in Theorem 2/Corollary 2.","Statistical consistency: with a uniformly consistent estimator of link probabilities (e.g., neighborhood smoothing), the empirical eigenvector's sign pattern matches the population one with probability at least (1-n^{-ε})^m, making NHCL asymptotically exact (Theorem 3).","Fully adaptive hierarchy: NHC-TST requires no prior knowledge of the number of clusters or tree depth; the two-sample test decides where to stop, and Theorem 4 shows the full hierarchy is recovered with probability at least 1 - Σ(δ_x=1)(m_x n^{-ε} + β_x) - 2L α_n, which tends to 1 under asymptotic power one.","Empirical superiority: in simulations with hierarchical smooth graphons and hierarchical SBMs, NHC-TST variants achieve near-perfect cluster membership (NMI) and topology recovery (adapted CPCC) across dense and sparse regimes, outperforming existing flat and hierarchical network-clustering baselines.","Interpretable multi-resolution structure: on the global migration dataset, the method recovers a temporal hierarchy that isolates the COVID-19 crisis and distinguishes post-COVID recovery from the pre-pandemic regime, structure invisible to flat clustering."],"fun_headline_variants":["Exact network tree recovery via spectral splits","Data-driven tree clustering of networks, no depth needed","Hierarchical distance matrices enable provable network clustering","Two-sample tests guide spectral splits for network trees","Migration data reveals hidden network hierarchies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that at every internal node of the latent tree, every distance between networks on different branches is strictly larger than every distance within either branch (the HDM monotonicity condition); if this recursive separation fails at any node, the spectral sign pattern need not align with the true bipartition and the exact-recovery guarantees collapse.","fun_headline_variants_meta":{"raw":{"variants":["Exact network tree recovery via spectral splits","Data-driven tree clustering of networks, no depth needed","Hierarchical distance matrices enable provable network clustering","Two-sample tests guide spectral splits for network trees","Migration data reveals hidden network hierarchies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1052,"prompt_tokens":693,"completion_tokens":359,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":437,"tokens_out":359,"duration_ms":4346,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:46:30.428886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a population of networks from an unbalanced tree where a deep split has a larger population distance than a shallow split (violating HDM monotonicity), e.g., the mirrored version of Section 4.3, and check whether the dominant Laplacian eigenvector's signs separate the true root branches; the paper's own Theorem 1 would not apply, and the split should fail at the root.","supporting_citations":[],"review_version":1}