{"id":"cc5623d4-88e5-464b-b861-a42509980153","arxiv_id":"2505.21845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A non-asymptotic bound is proven for spectral clustering misclustering error in dependent community Hawkes models, and a fast GMM estimator for a restricted model is shown to be consistent.","lead":"This paper studies community detection in temporal networks where events between node pairs follow mutually exciting Hawkes processes. It derives non-asymptotic error bounds for spectral clustering on event counts and proposes a scalable parametric model with provably consistent estimation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's proof never justifies transferring the coordinatewise Kolmogorov bound to a spectral-norm tail, so the misclustering guarantee in Theorem 5 lacks a supported key step.","rationale":"The reader correctly identified that Proposition 1 has an unspecified constant C(n) and that the condition 'sigma > 0' in Proposition 2 is stated without proof. My independent read found a more structural gap: the proof of Theorem 3 does not show how a coordinatewise Kolmogorov-distance bound can be converted into a tail bound for the spectral norm of the count matrix. This transfer is the linchpin connecting the Hawkes-process Gaussian approximation to the misclustering rate. Without it, Theorem 3's high-probability bound is unsupported, and Theorem 5 inherits the gap. The issue is potentially fixable: one can smooth the indicator of the spectral-norm ball and use the d2 bound directly, or use an epsilon-net with an explicit dimension penalty. But as written, the argument jumps from orthant probabilities to operator norm probabilities, and no such step appears. This does not mean the theorem is false; the simulations and the general strategy are plausible, so a conditional verdict is appropriate rather than rejection. The reader's weaker concern about C(n) is also valid but secondary: even a fully explicit C(n) would not repair the missing spectral-norm transfer, so my focus differs from the reader's and agreement is only partial.","tokens_in":40658,"tokens_out":8213,"duration_ms":92312,"concrete_test":"Re-derive the transition from Proposition 2 to the probability bound in Theorem 3. Specifically, attempt to prove that for the event E = {\\|N_T - E N_T\\| > t}, the difference |P(E) - P(\\|M\\| > t)| is bounded by kappa(n) T^{-1/6} plus the Gaussian tail, using only the stated dK bound and d2 bound. If the derivation requires an epsilon-net over the unit sphere, compute the resulting net size and show whether the final constants (or kappa(n)) absorb it; if no valid derivation exists, Theorem 3 and hence Theorem 5 are unproven as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is Theorem 5, and it rests on Theorem 3's high-probability bound on \\|N_T - E N_T\\|. In the proof of Theorem 3 (Appendix A.1), the authors import Proposition 1, a smooth-Wasserstein (d2) bound, then convert it to a Kolmogorov bound dK(Y_T, G) for coordinatewise orthant probabilities via Proposition 2. The proof then states that this conversion can be used to conclude that, for any x, |P(sqrt(T)(vec(N_T)/T - R vec(mu)) > x) - P(vec(M) > x)| < kappa/T^{1/6}. But the target event {\\|N_T - E N_T\\| > a} is not a coordinatewise orthant event. It is the supremum over all unit vectors u, v of u^T (N_T - E N_T) v, an infinite union of linear events. The dK bound controls only probabilities of sets of the form {Z <= x}, and the d2 bound controls expectations of C^2 functions with bounded gradient and Hessian; neither directly controls the spectral-norm tail. The proof computes a Gaussian tail for \\|M\\|, then announces the same tail for N_T - E N_T with an extra kappa/T^{1/6} term, without showing how the approximation passes through the supremum over the unit sphere. A discretization over an epsilon-net would introduce an additional dimension-dependent factor, and the stated constants do not account for it. This is not merely an unspecified constant: the mechanism connecting Propositions 1-2 to the spectral-norm event is absent. The unverified 'sigma > 0' condition in Proposition 2 is less serious because, under mu_ij > 0 and rho(Gamma) < 1, R vec(mu) > 0 and R diag(R vec(mu)) R^T has positive diagonal, so sigma > 0 is in fact true; the real gap is the spectral-norm transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the dependent community Hawkes (DCH) family, which combines a stochastic block model on node pairs with mutually exciting multivariate Hawkes processes among pairs in the same or reciprocal block pairs. The main theoretical contribution is Theorem 5, a non-asymptotic upper bound on the misclustering rate of spectral clustering applied to the event count matrix, with explicit dependence on the number of nodes n, communities K, observation time T, and a dependence parameter γmax. The paper also proposes the self and reciprocal excitation (SR) model, a restricted version with a generalized method of moments (GMM) estimator whose identification and consistency are proven, and supplements the theory with simulations and real-data experiments on five temporal network datasets.","tokens_in":41085,"tokens_out":7341,"duration_ms":70175,"significance":"If the main result is correct, it would provide the first non-asymptotic misclustering guarantee for spectral clustering under a class of dependent Hawkes-process block models, extending prior consistency results for the conditionally independent CHIP model. The proposed SR model and its GMM estimator are computationally scalable and are shown to be competitive with much slower alternatives on real data. The proof of the GMM identification condition for the restricted SR model is a useful standalone contribution. The empirical evaluation is thorough and generally aligned with the qualitative predictions of the theory.","major_comments":[{"comment":"The proof does not justify the crucial transfer from a coordinatewise Kolmogorov bound to a spectral-norm tail. After applying Proposition 2, the proof asserts that for any x the coordinatewise orthant probabilities of sqrt(T)(vec(N_T)/T - R vec(mu)) and vec(M) differ by at most kappa/T^{1/6}, and then it applies this to the event {||M|| > a}. But {||M|| > a} is the supremum of linear events over the unit sphere, not an orthant event, and neither the d_K bound nor the d_2 bound directly controls such an event. An epsilon-net argument or another mechanism is required, and the stated constants do not account for the dimension-dependent factors that such an argument would introduce. This gap is load-bearing because Theorem 5 inherits the spectral-norm control from Theorem 3.","section":"Appendix A.1, proof of Theorem 3"},{"comment":"The algebra of the Gaussian tail bound does not deliver the displayed inequality in Theorem 3. With C1 = (1 - sigma*)^-2 (1 + gamma_max), the proof chooses a = sqrt(2 n C1^3 mu_max n log n log T) and uses E||M|| <= 2 sqrt(n C1^3 mu_max (1 + 2 log n)). This yields ||M|| <= 2 sqrt(n C1^3 mu_max (1 + 2 log n)) + n sqrt(2 C1^3 mu_max log n log T) with probability at least 1 - exp(-n log n log T), not the stated 3 sqrt(n C1^3 mu_max (1 + 2 log n) log T). For large n the former is larger by a factor on the order of sqrt(n / log T), so the final inequality as written does not follow. Either the threshold or the probability exponent must be corrected.","section":"Appendix A.1, Gaussian tail computation"},{"comment":"The statement of Proposition 2 says 'We verify that sigma > 0 and d2(Y_T, G) <= (sqrt(4 log n + 2))/(2 sigma) for sufficiently large T', but no such verification appears in the proof or elsewhere. While sigma > 0 is plausible under mu > 0 and rho(Gamma) < 1, the d2 inequality depends on the unspecified constant C(n) from Proposition 1, and the phrase 'sufficiently large T' is incompatible with Theorem 3's claim for all T > 1. The paper should either provide the verification, state a threshold T_0(n), or weaken the theorem accordingly.","section":"Proposition 2"},{"comment":"The theorems state the probability as 1 - exp(-log n log T) - kappa(n)/T^{1/6} for every T > 1, but kappa(n) is an unspecified positive function of n. For small T the second term can exceed 1, making the claimed probability negative and the statement vacuous. The non-asymptotic claim therefore needs a qualification such as T >= T_0(n), or the small-T regime must be handled separately. Relatedly, Proposition 1 is imported only for fixed n with an unspecified C(n); without explicit control of C(n), the claimed n-dependence in the misclustering bound is not fully established.","section":"Theorem 3 and Theorem 5, probability statement"}],"minor_comments":[{"comment":"The probability is written as 1 - exp(log n log T) - kappa(n)/T^{1/6}; the minus sign before exp is missing and should read 1 - exp(-log n log T) - kappa(n)/T^{1/6}.","section":"Appendix A.2, proof of Theorem 5"},{"comment":"The text says 'The Kth largest singular values of E[N_T]/T is n/K (v1 - v2)', but it should refer to the K-th singular value lambda_K, and the derivation of lambda_K from the block structure of the expected count matrix should be made explicit because it drives the K^2/(nT) rate.","section":"Appendix A.3, proof of Corollary 6"},{"comment":"The caption of Figure 4 says the MSEs 'decrease as n or T decreases'; this should read 'decrease as n or T increases' (or equivalently, 'increase as n or T decreases').","section":"Section 5.2, GMM simulation"},{"comment":"In the proof of Theorem 9, the definition of g02(., .) is written as (Lambda_ba)_0 - Lambda_ab; this appears to be a typo for (Lambda_ba)_0 - Lambda_ba, since otherwise the moment function is not centered at zero at the true parameter.","section":"Section 4, GMM setup"},{"comment":"The notation 'for any n > 1 and T > 1' is repeated in several statements, but the proof only establishes the Gaussian approximation for fixed n with a constant C(n); the paper should state explicitly which results are uniform in n and which are not.","section":"Section 3.1, Theorem 3 statement"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important problem and contains a promising theoretical framework, but the main spectral-norm guarantee currently rests on an unproved transfer from coordinatewise Gaussian approximation to a spectral-norm tail, and the Gaussian tail computation in the proof contains an algebraic inconsistency. The GMM identification and consistency results are more self-contained and may be publishable independently. I would encourage the authors to repair the proof of Theorem 3 and to state the dependence on C(n) and kappa(n) explicitly before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the main theorem is not established by the proof. The Gaussian approximation in Proposition 1 is a smooth-Wasserstein bound, and Proposition 2 only converts it to coordinatewise Kolmogorov control. Theorem 3 needs a tail bound on the spectral norm of N_T - E N_T, which is a supremum over the unit sphere, not a coordinatewise orthant event. The proof never shows how the approximation passes through that supremum. The stress-test note is correct: without an epsilon-net argument, the step is missing, and such an argument would introduce additional dimension-dependent factors. This is load-bearing because Theorem 5's misclustering rate rests entirely on Theorem 3. The claim may still be true, but it is unsupported as written.\n\nThe paper does several things well. The DCH class usefully unifies CHIP, BHM, and MULCH under one set of assumptions. The SR model is a reasonable middle ground between scalability and flexibility. The GMM identification lemma is genuinely new and proven from first principles, not imported from prior work, and the consistency proof via Newey-McFadden is plausible. The simulations and real-data experiments are extensive and line up with the qualitative prediction that stronger dependence hurts clustering. There is no circularity and the citations to the authors' own MULCH paper are appropriate for the technical imports.\n\nThe soft spots beyond the main gap are real but less severe. Proposition 2 says \"we verify\" a condition and then never carries out the verification; the σ > 0 part is likely true under the model assumptions, so this is fixable. The probability bound is stated for all T > 1, but kappa(n)/T^{1/6} can exceed 1 for small T, so the statement is not uniformly meaningful as written. The unspecified C(n) in the Gaussian approximation also obscures the claimed n-dependence. These are repair items, not reasons to reject the program.\n\nWho this is for: people working on spectral clustering for temporal networks or on moment-based estimation for multivariate Hawkes processes. The modeling and estimation half of the paper is in good shape; the clustering half needs a serious proof repair. I would send it to peer review rather than desk reject, but a referee should require a correct spectral-norm transfer or a revised theorem with an explicit epsilon-net argument before the main result is accepted.","headline":"Main misclustering theorem is not established as written because the proof never transfers the Gaussian approximation to the spectral-norm event, but the model class and GMM identification are solid and the paper deserves refereeing.","tokens_in":41567,"tokens_out":3027,"would_cite":false,"duration_ms":36689,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","60G55","62M09"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under the dependent community Hawkes model class, spectral clustering on the event-count matrix recovers the true communities with a non-asymptotic misclustering bound that decays as the observation window grows.","keywords":["temporal networks","relational events","Hawkes processes","community detection","spectral clustering","stochastic block model","generalized method of moments","non-asymptotic bounds"],"falsifier":"Simulate a DCH network with moderate n (say 100) and K = 4, run Algorithm 1 over a range of T, and compare the empirical misclustering rate against the predicted log T / T scaling: a slope far from −1 in log r versus log T, or a misclustering rate plateauing while the bound decays, would refute the Theorem 5 mechanism. A more direct test computes the Kolmogorov distance between the standardized vectorized count matrix and its approximating Gaussian for n = 50, 100, 200 and checks whether the $T^{{-1/6}}$ rate holds with a constant that does not grow with n.","tokens_in":40479,"feed_emoji":"🕸️","tokens_out":8335,"duration_ms":79539,"temperature":0.7,"pith_summary":"Temporal networks built from timestamped relational events often show community structure and mutual excitement, where one interaction triggers later interactions between other node pairs. This paper introduces the dependent community Hawkes (DCH) model class, which puts a stochastic block model on the baseline rates and a block-structured, mutually exciting Hawkes intensity on the event streams, and asks whether spectral clustering on the directed event-count matrix still recovers the hidden communities. The main claim is that it does: the misclustering rate is bounded non-asymptotically by an explicit function of node count, community count, observation time, and the total mutual-excitation level, and this rate goes to zero as time grows. A second contribution is the self-and-reciprocal (SR) submodel with a scalable generalized method of moments estimator, whose identification condition the paper proves and whose consistency is shown as both time and block sizes grow.","feed_headline":"Spectral clustering provably recovers dependent network communities","feed_subtitle":"New bound shows misclustering shrinks as observation time grows even when dyads mutually excite.","key_machinery":"The load-bearing device is the block-diagonal excitation matrix Γ of the $n^{2}$-dimensional mutually exciting Hawkes process, reordered by block pairs, together with the row-and-column-sum identity ('stochastic equivalence') that makes every node pair in a community pair exert and absorb the same total excitation. This structure makes the expected count matrix equal to Z B Z^T, so its top singular vectors are exactly the community indicator matrix up to a transformation, giving a canonically identifiable signal. The argument then controls the noise by bounding the row sums of R = (I − Γ)^{−1} in terms of (1 − σ*)^{-2}(1 + γmax), importing a Gaussian approximation for the vectorized count matrix, and uses a Davis-Kahan/Wedin perturbation bound together with an approximate k-means step to turn the spectral-norm error into a per-node misclustering rate.","core_discovery":"On the paper's own terms, the central result is Theorem 5: if the n-node network is generated by a DCH model with spectral radius σ* < 1 and per-block excitation row sums bounded by γmax, then the misclustering rate r of the spectral clustering algorithm applied to the count matrix N_T satisfies r ≤ (log T / T) · 1440(2+ε)^2 n_max K / $λ_K^{2}$ · (1−σ*)^{−6}(1+γmax)^3 μmax (1+2 log n), with probability at least 1 − exp(−log n log T) − κ(n)/$T^{{1/6}}$. In the simplified symmetric submodel this yields r ≲ $K^{2}$ μmax log n log T / (n T), so consistency follows as T → ∞ and the required observation time grows with $K^{2}$ log n / n. The proof transfers a Gaussian approximation of the vectorized Hawkes count matrix to a spectral-norm bound on N_T minus its expectation, then converts the bound into a misclustering bound through a Davis-Kahan-type row-perturbation argument; the expected count matrix is shown to have singular vectors aligned with the community indicator matrix.","pith_inferences":["Editorial extension: the theorems transfer directly to static weighted directed networks whose edge counts are generated by a DCH process, suggesting spectral clustering accuracy in weighted networks is governed by n, T, and excitation strength, not just by edge density.","Editorial extension: the proof's reliance on an unspecified n-dependent constant in the Gaussian approximation means the true finite-sample scaling in n may be worse than the displayed (1+2 log n) factor, and an empirical check of the Gaussian approximation across n would test this directly.","Editorial extension: the identification-by-symmetry strategy behind the restricted SR model suggests a general recipe: impose a symmetric parameter constraint to reduce the parameter count to the number of first- and second-order cumulant equations, then verify identification case by case."],"forward_implications":["If the bound is right, spectral clustering is consistent for every DCH model as the observation window T grows, even though dyadic event streams are dependent.","For the simplified symmetric model, the bound shows the misclustering rate depends on K^2 μmax log n log T / (nT), making explicit how many nodes and how much observation time are needed.","The bound increases with γmax, the total mutual excitation per block pair, and blows up as the stability boundary σ* → 1, so stronger cross-dyad dependence is predicted to make communities harder to recover.","The restricted SR model's GMM estimator is consistent as T and block-pair sizes grow, and the paper proves the identification condition that prior GMM theory merely assumes.","On five real datasets the restricted SR model attains test log-likelihood and dynamic link prediction AUC competitive with the slower MULCH model while scaling to tens of thousands of nodes."],"supporting_citations":[{"why":"Supplies Proposition 1, the Malliavin-Stein Gaussian approximation for the vectorized multivariate Hawkes count process that starts the proof of Theorem 3.","marker":"Khabou (2021)"},{"why":"Provides the Kolmogorov-distance bound used in Proposition 2 to turn the smooth-Wasserstein approximation into the tail probability in Theorem 3.","marker":"Gaunt and Li (2023)"},{"why":"Defines the CHIP model, the conditional-independence baseline whose consistency result the paper extends to dependent DCH models, and the moment-estimation template for SR.","marker":"Arastuie et al. (2020)"},{"why":"Supplies the spectral clustering consistency proof pattern and perturbation argument that Theorem 5 adapts to dependent, weighted count matrices.","marker":"Lei and Rinaldo (2015)"},{"why":"Gives the row-normalization misclustering bound and the Davis-Kahan-type result used to turn spectral-norm error into the per-node error rate.","marker":"Rohe et al. (2016)"},{"why":"Defines the MULCH model, contributes the row-sum propositions and the Z B Z^T block structure for expected counts that Lemma 4 relies on.","marker":"Soliman et al. (2022)"},{"why":"Establishes the integrated-cumulant identities relating first and second moments of a multivariate Hawkes process to M and Γ, the basis of the SR GMM estimator.","marker":"Achab et al. (2018)"},{"why":"Supplies the GMM consistency theorem whose conditions, including the identification condition, are verified in Theorem 9.","marker":"Newey and McFadden (1994)"}],"fun_headline_variants":["Spectral clustering provably handles dependent Hawkes networks","Guaranteed community detection in mutually exciting temporal networks","New bound: spectral clustering succeeds on dependent network data","Spectral clustering robust to dependences in event counts","Provable misclustering rate for spectral clustering on Hawkes models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the imported Gaussian approximation of the whole vector of event counts, stated only for fixed n with an unspecified constant C(n), together with a small-variance condition that the paper says it verifies but never proves; if C(n) grows quickly with n, or the condition fails at the T values of interest, the claimed non-asymptotic n-dependence of the misclustering bound is not established.","fun_headline_variants_meta":{"raw":{"variants":["Spectral clustering provably handles dependent Hawkes networks","Guaranteed community detection in mutually exciting temporal networks","New bound: spectral clustering succeeds on dependent network data","Spectral clustering robust to dependences in event counts","Provable misclustering rate for spectral clustering on Hawkes models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1406,"prompt_tokens":1009,"completion_tokens":397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":625,"tokens_out":397,"duration_ms":4615,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:22:07.973612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a DCH network with moderate n (say 100) and K = 4, run Algorithm 1 over a range of T, and compare the empirical misclustering rate against the predicted log T / T scaling: a slope far from −1 in log r versus log T, or a misclustering rate plateauing while the bound decays, would refute the Theorem 5 mechanism. A more direct test computes the Kolmogorov distance between the standardized vectorized count matrix and its approximating Gaussian for n = 50, 100, 200 and checks whether the $T^{{-1/6}}$ rate holds with a constant that does not grow with n.","supporting_citations":[{"cited_title":"Malliavin-Stein method for the multivariate compound Hawkes process","cited_arxiv_id":"2109.07749","evidence_quote":"Supplies Proposition 1, the Malliavin-Stein Gaussian approximation for the vectorized multivariate Hawkes count process that starts the proof of Theorem 3."},{"cited_title":"Bounding kolmogorov distances through wasserstein and related integral probability metrics","cited_arxiv_id":null,"evidence_quote":"Provides the Kolmogorov-distance bound used in Proposition 2 to turn the smooth-Wasserstein approximation into the tail probability in Theorem 3."},{"cited_title":"Chip: A hawkes process model for continuous-time networks with scalable and consistent estimation","cited_arxiv_id":null,"evidence_quote":"Defines the CHIP model, the conditional-independence baseline whose consistency result the paper extends to dependent DCH models, and the moment-estimation template for SR."},{"cited_title":"Consistency of spectral clustering in stochastic block models","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral clustering consistency proof pattern and perturbation argument that Theorem 5 adapts to dependent, weighted count matrices."},{"cited_title":"Co-clustering directed graphs to discover asymmetries and directional communities","cited_arxiv_id":null,"evidence_quote":"Gives the row-normalization misclustering bound and the Davis-Kahan-type result used to turn spectral-norm error into the per-node error rate."},{"cited_title":"Uncovering causality from multivariate hawkes integrated cumulants","cited_arxiv_id":null,"evidence_quote":"Establishes the integrated-cumulant identities relating first and second moments of a multivariate Hawkes process to M and Γ, the basis of the SR GMM estimator."}],"review_version":1}