{"id":"91352ff8-5e9a-4546-aa7b-02c08068d2b2","arxiv_id":"2507.15370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A system of Volterra-like integral equations characterizes the multi-temporal Laplace transform and two-time covariance of non-stationary multivariate Hawkes processes.","lead":"This paper derives integral equations for the multi-time Laplace transform, moments, and covariance of non-stationary multivariate Hawkes processes with time-dependent baseline and general non-negative excitation kernels. It gives statisticians and epidemiologists a toolbox for inference with self-exciting point processes, extending previous stationary or Markovian results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.6's Lebesgue decomposition misclassifies the lower-triangle ancestor–descendant covariance as diagonal singular mass; the claimed AC/singular split is not the true decomposition.","rationale":"The reader identified the immigrant-birth cluster representation as the weakest assumption, but that representation is standard and valid for the stated linear model under local integrability and non-explosiveness; it is not the central risk. The Laplace-transform system in Theorem 2.2 appears internally consistent, and the covariance renewal equations are plausible. The genuinely load-bearing defect is Proposition 4.6: the claimed Lebesgue decomposition is wrong because it relocates an absolutely continuous lower-triangle covariance contribution onto the diagonal. A single two-interval computation exposes a missing O(φ^2) term. This does not necessarily invalidate the Laplace transform or the explicit covariance formula, so a conditional verdict is appropriate: the decomposition must be corrected or removed, and the numerical and simulation sections should be reconciled with the corrected formula. Separately, Lemma 2.1's one-event probability is false and Theorem 2.3 has a sign inconsistency, but those are less central than the covariance decomposition defect.","tokens_in":32556,"tokens_out":47791,"duration_ms":534201,"concrete_test":"Take d=1, λ0(t)=φ(t)=α e^{-βt}, A=[2,3], B=[0,1] with B before A. The direct branching expansion gives Cov(N(A),N(B)) = α^2 ((1−e^{-β})/β)((e^{-2β}−e^{-3β})/β) + O(α^3), from immigrant–first-generation pairs. Proposition 4.6 predicts no O(α^2) contribution: its AC first summand vanishes for v<u, the singular part is supported on the diagonal, and the integral term is O(α^3). Computing the ratio for α=0.01, β=1 (predicted/true ≈ O(α)) would decisively falsify the decomposition. Equivalently, compute the mixed derivative of the paper's cumulative C(t1,t2) and verify that the resulting measure is not symmetric, whereas any covariance measure must be.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The cluster representation is standard for linear Hawkes processes and is not the real soft spot. The load-bearing problem is Proposition 4.6, which is advertised as a central contribution. In the univariate case d=1 with baseline λ0=φ, the displayed AC density contains the term φ(v−u)1_{v≥u} but no analogous φ(u−v)1_{u≥v} term. The missing term is absolutely continuous on the lower triangle {v<u}: it is the immigrant-at-v, descendant-at-u pair contribution. The proof instead obtains a cumulative contribution depending only on t1 for t1<t2 and classifies it as singular diagonal mass, adding a spurious term m^{j'}_l ∗ m^{l,(1)}_k(u) δ_u(dv)du. A measure supported on {v<u} can indeed have cumulative F(t1) for t1<t2 without being diagonal; its mixed derivative on the open lower triangle is the AC density φ(v)φ(u−v). Thus the paper's 'singular part' overstates the diagonal and its 'absolutely continuous part' is not symmetric and misses genuine O(φ^2) covariance between disjoint ordered intervals. This is an internal inconsistency, not merely a disagreement with the literature.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a general class of multivariate Hawkes processes with time-dependent baseline intensity and general excitation kernels, in view of applications to spatio-temporal epidemic modeling. The main results are: (i) a system of Volterra-like integral equations for the multi-temporal Laplace transform (Theorem 2.2) and a functional Laplace transform (Theorem 2.3), both derived from the immigrant-birth cluster representation; (ii) an infinite-divisibility-based derivation of equations for the probability distribution of counts at a single time (Proposition 3.5); (iii) integral equations and explicit series solutions for the first two moments, including a two-time covariance formula (Propositions 4.1-4.5); and (iv) a claimed Lebesgue decomposition of the covariance measure into singular and absolutely continuous parts (Proposition 4.6). The paper also contains numerical discretization schemes, illustrations of the moment equations, and simulated trajectories for a two- and four-dimensional Hawkes model.","tokens_in":32798,"tokens_out":27825,"duration_ms":256670,"significance":"If the main formulas are correct, the paper provides a substantial extension of known Laplace-transform and moment results for Hawkes processes from stationary/single-time settings to non-stationary, multivariate, multi-temporal settings. The derivations are self-contained and based on the standard cluster representation, and several formulas reduce to classical special cases, which is a genuine strength. The explicit Volterra-like systems and series solutions for the mean and covariance could serve as a basis for moment-based inference in epidemic applications. However, the advertised Lebesgue decomposition of the covariance measure (Proposition 4.6) is incorrect, and a stated lemma on one-event probabilities (Lemma 2.1) is false; these issues affect the reliability of parts of the manuscript as it stands.","major_comments":[{"comment":"The asserted formula P(Ñ^{j'}(t)=e_i) = (∫_0^t φ^{j'}_i(s)ds) exp(−Σ_j ∫_0^t φ^{j'}_j(s)ds) is false for j'=0:d. The event that the total Hawkes count equals one requires not only exactly one ancestor event but also that the Hawkes process ignited by that ancestor has zero events up to the time horizon. The proof's statement that {Ñ^{j'}(t)=e_i} corresponds exactly to the Poisson event {N^0_i(t)=1} is therefore incorrect. The correct one-event probability is p̃^{j'}_{e_i}(t) = p̃^{j'}_0(t) ∫_0^t p̃^i_0(t−u) φ^{j'}_i(u) du (obtainable from Eq. (15) with l=e_i), which reduces to the displayed formula only when the excitation φ is zero. The same erroneous expression is used in the example following Proposition 3.5, where p̃^{j'}_1(t)=e^{−Φ_{j'}(t)}Φ_{j'}(t) is claimed. This must be corrected; although Theorem 2.2 itself does not appear to rely on Eq. (6), the lemma is stated as a result and the error propagates to the illustrations.","section":"Lemma 2.1, Eq. (6)"},{"comment":"The claimed Lebesgue decomposition of the two-time covariance measure is not the correct decomposition. In the proof, the term (h̃^{j'}_j ∗ [M̃^{l,(1)}_k(·)1(·)])(t1,t2) equals ∫_0^{t1} h(u) M(t1−u) du for t1<t2, and the authors conclude that it is supported on the diagonal because it depends only on t1. However, the mixed derivative of this function on the open lower triangle {t2<t1} is the absolutely continuous density h(t2) m(t1−t2) (with m=M'), not a diagonal mass; the diagonal contribution from this term is proportional to M(0), which is zero for the cumulative mean functions considered here. Consequently, the absolutely continuous part stated in Proposition 4.6 contains the term m̃^{j',(1)}_k(u) m̃^{k,(1)}_l(v−u) for v>u but lacks its symmetric counterpart m̃^{j',(1)}_l(v) m̃^{l,(1)}_k(u−v) for u>v, and the singular part incorrectly includes the convolution term m̃^{j',(1)}_l ∗ m̃^{l,(1)}_k(u) δ_u(dv)du. The resulting covariance measure is not symmetric and does not match the known second-order structure of linear Hawkes processes (e.g., in the univariate stationary case the off-diagonal density should contain both μ h(v−u) and μ h(u−v)). Since Proposition 4.6 is advertised as a central contribution, this error requires a corrected derivation of the decomposition.","section":"Proposition 4.6, Section 4.3"}],"minor_comments":[{"comment":"In the exponent of the discrete Laplace transform approximation, the term e^{a_j} should be e^{−a_j} to match the continuous-time formula and the preceding cases for m=1 and m=2.","section":"Eq. (45)"},{"comment":"There are typographical issues in the proof: 'Λ0(tk) − Λ(tk−1)' is missing a subscript on the second term, and 'e^{−a[k,n] j } L̃^{j,(n+1−k), (t − u)[k,n]}' contains a misplaced comma that should be removed.","section":"Proof of Theorem 2.2"},{"comment":"The abstract contains the sentence 'We also provides illustrative simulations', which should be 'We also provide...'.","section":"Abstract and Introduction"},{"comment":"The definition of R̃^j_{k,l}(u,v) uses '(Id + M̃^{(1)}(u))' while the notation elsewhere is e_j + M̃^{j,(1)}(u); the matrix notation should be made consistent.","section":"Section 4.2.4"},{"comment":"The numerical scheme is presented without a convergence or error analysis; a brief statement on the order of approximation or a reference to a fixed-point convergence argument would be useful.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The Laplace transform and moment equation results in Sections 2 and 4 appear to be largely sound, but the manuscript contains two serious mathematical errors: the false one-event probability in Lemma 2.1 and the incorrect Lebesgue decomposition in Proposition 4.6. Since Proposition 4.6 is advertised as a central contribution, the current version cannot be accepted. The correct decomposition appears to be obtainable by recomputing the mixed derivative of the diagonal-convolution solution, so the issues are fixable within the paper's scope. I would encourage the editor to request a revision that corrects these points and checks the propagation of the Lemma 2.1 error into the examples."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the multi-time Laplace transform equations (Theorems 2.2 and 2.3) for non-stationary multivariate Hawkes processes appear to be a genuine, correct extension of the single-time formula, worked out carefully through the cluster representation. Second, the advertised Lebesgue decomposition of the two-time covariance in Proposition 4.6 is wrong, and the error is not cosmetic. The AC density they display is not symmetric, and the 'singular' part contains mass that actually lives off the diagonal. In the univariate case, the term they classify as diagonal — the one coming from \\tilde h * [M(u)1(v)] — has a mixed derivative that includes h(v)m(u-v) on the open lower triangle {v<u}, an absolutely continuous ancestor–descendant contribution. So the claimed AC/singular split misclassifies genuine O(phi^2) covariance between disjoint ordered intervals as diagonal mass. The covariance measure as a whole (Prop 4.5) may still be right, but the decomposition and its proof need to be redone.\n\nThe paper does good things. The system of Volterra-like equations for multi-time Laplace transforms is a real extension beyond the single-time results of Hawkes–Oakes and others, and it is derived cleanly. The explicit series solution for the two-time covariance (Prop 4.5) is useful. The mean equations are standard but clearly presented. The numerical scheme is sensible in outline, but the example parameters are not fully reported, so the figures are not reproducible.\n\nOne more soft spot: Lemma 2.1's formula for the probability of exactly one event is wrong — it ignores the possibility that the single immigrant produces no descendants by time t. That lemma is not load-bearing, but it should be corrected.\n\nBottom line: this paper deserves a serious referee, but the referee should ask for a major revision. The Laplace transform results can stand, but Proposition 4.6 must be fixed or removed, and Lemma 2.1 corrected. With those changes it would be a solid contribution for people who need multi-time characteristics of non-stationary multivariate Hawkes processes in epidemic modeling or point-process inference.","headline":"Solid multi-time Laplace transform results, but the advertised covariance Lebesgue decomposition is misclassified and needs a real fix.","tokens_in":33335,"tokens_out":7189,"would_cite":false,"duration_ms":75960,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","60E10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A general multivariate Hawkes process with time-dependent baseline and non-Markovian excitation has a closed-form multi-temporal Laplace transform, from which exact two-time covariances and count probabilities follow.","keywords":["Hawkes process","multi-temporal Laplace transform","characteristic function","infinite divisibility","covariance structure","non-stationary point process","spatio-temporal epidemic model","Volterra integral equations"],"falsifier":"Take $d=1$ with $\\lambda_0(t)=1+\\sin t$ and $\\phi(t)=\\alpha e^{-\\beta t}$, simulate many independent paths, and compare the empirical covariance density off the diagonal with the absolutely continuous part predicted by Proposition 4.6; a systematic mismatch beyond Monte Carlo error would refute the covariance formula. The diagonal mass $\\delta_{k=l}m^{(1)}_k + m^{(1)}_l * m^{(1)}_k$ predicted by the singular part can be checked in the same simulation.","tokens_in":32376,"feed_emoji":"🦠","tokens_out":15858,"duration_ms":152508,"temperature":0.7,"pith_summary":"This paper aims to give a complete distributional calculus for a broad class of multivariate Hawkes processes in which the baseline intensity varies with time and the excitation matrix is an arbitrary non-negative matrix function, so the model is usually non-Markovian. Its central result is a recursive system of Volterra-like integral equations for the joint Laplace transform of the counts at arbitrarily many observation times, extending the classical single-time formula. From this system the paper derives explicit formulas for the first two moments, including an expression for the covariance at two distinct times, and an analytic decomposition of that covariance into a singular diagonal part and an absolutely continuous part. These formulas matter because they make exact distributional characteristics available for spatio-temporal epidemic models and other non-stationary self-exciting processes, which can support inference and process-limit studies outside stationary or Markovian settings.","feed_headline":"Exact multi-time transform for non-stationary Hawkes processes","feed_subtitle":"The new integral equations yield two-time covariances and count probabilities for spatio-temporal epidemic models.","key_machinery":"The machinery is the immigrant-birth (cluster) representation of a linear Hawkes process: an ancestor Poisson process with intensity $\\lambda_0$ generates first-generation events, and each such event independently generates its own full Hawkes process with kernel given by the corresponding row of $\\phi$. Together with infinite divisibility, this turns the Laplace transform into a product of exponentials and yields the Volterra-like system (7). For moments, the workhorse is the resolvent-style fundamental series $\\sum_{r\\ge0}\\phi^{(*r)}$, whose diagonal-supported bivariate extension $\\widehat{\\phi}(du,dv)=\\phi(u)\\delta_u(dv)du$ carries the two-time covariance solution (43). A log-Laplace coefficient recursion from Proposition 3.1 converts these transforms into equations for the count probabilities (Proposition 3.5).","core_discovery":"On the paper's own terms, the discovery is that the $n$-time Laplace transform of a $d$-dimensional non-stationary Hawkes process driven by (2) satisfies the recursive system (7), in which the transform at later times is expressed through the $d$ basic single-type transforms under the excitation kernel. Because each time slice contributes an independent Poisson ancestor layer, the finite-dimensional distributions are fully determined by this system. The paper then solves the corresponding second-order equations in closed series form: the two-time covariance is given by Equation (43) as a convolution of the diagonal-supported measure $\\widehat{\\phi}$ with the resolvent series $\\sum_{r\\ge1}\\phi^{(*r)}$, and Proposition 4.6 splits that covariance measure into a singular part supported on the diagonal and an absolutely continuous part with explicit densities. This extends known single-time and stationary multivariate formulas to a general non-Markovian, time-dependent-baseline setting.","pith_inferences":["The explicit two-time covariance could support minimum-contrast estimators for baseline and excitation parameters without forcing the process into a Markovian state representation; this is an estimation route the paper mentions but does not develop.","The covariance decomposition suggests that diagonal fluctuations are governed by first-order mean intensities, while off-diagonal dependence is carried by the absolutely continuous part; this distinction could simplify asymptotic variance calculations for non-stationary Hawkes processes.","The same log-Laplace recursion used for count probabilities should extend to higher-order multi-time cumulants, yielding exact skewness and kurtosis under the model rather than only mean and covariance.","A direct consistency check is available by taking a constant baseline and translation-invariant excitation, where the new series should reproduce the stationary cumulant formulas from earlier work."],"forward_implications":["The multi-temporal Laplace transform in Equation (7) gives a recursive way to compute finite-dimensional distributions of non-stationary multivariate Hawkes processes, not just single-time marginals.","Means and the two-time covariance can be evaluated through convergent series based on $\\sum_{r\\ge0}\\phi^{(*r)}$, so neither Markovian nor stationary assumptions are needed for these characteristics.","The covariance measure of each component pair splits into a diagonal singular part and an absolutely continuous part with explicit densities, enabling variance and cross-correlation computations at arbitrary time pairs.","The grid-based convolution scheme approximates Laplace transforms and moments, providing numerical building blocks that can be used for estimation and calibration of epidemic models.","Because $\\lambda_0$ is allowed to be time-dependent, seasonal or multi-introduction epidemic models fall directly within the scope of the formulas."],"supporting_citations":[{"why":"supplies the cluster/population representation (4) that all Laplace-transform and moment proofs start from.","marker":"[18]"},{"why":"gives the single-time multivariate characteristic-function formula that Theorem 2.2 extends to multiple times.","marker":"[11]"},{"why":"provides exact and asymptotic analysis of general multivariate Hawkes processes that the multi-time extension refines.","marker":"[26]"},{"why":"extends the cluster representation to marked Hawkes processes, justifying its use in the multivariate marked setting.","marker":"[25]"},{"why":"derives multi-time second- and third-order moments for stationary multivariate Hawkes processes, which Proposition 4.5 extends beyond stationarity.","marker":"[24]"},{"why":"gives a probability generating functional for stationary spatial Hawkes processes analogous to Theorem 2.3.","marker":"[31]"},{"why":"yields the order-statistics fact that Poisson-timed events are i.i.d., used in the Laplace-transform proofs.","marker":"[10]"},{"why":"establishes that the conditional intensity characterizes the law of a multivariate point process, used to identify the process distribution.","marker":"[21]"}],"fun_headline_variants":["Exact Laplace transform for multivariate Hawkes processes","Non-stationary Hawkes: closed-form multi-time covariances","Decomposing Hawkes covariance into singular and continuous parts","General Hawkes processes: explicit characteristic function","Unifying multi-time transforms for non-Markovian Hawkes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formulas assume the linear cluster representation of Hawkes processes, in which the process decomposes into an ancestor Poisson process and independent Hawkes subprocesses generated by each ancestor event, and the paper relies on this representation without re-proving it in full non-stationary, multivariate generality; if the excitation matrix is not integrable or the intensity is nonlinear, the integral equations no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["Exact Laplace transform for multivariate Hawkes processes","Non-stationary Hawkes: closed-form multi-time covariances","Decomposing Hawkes covariance into singular and continuous parts","General Hawkes processes: explicit characteristic function","Unifying multi-time transforms for non-Markovian Hawkes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1517,"prompt_tokens":1021,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":415}},"tokens_in":637,"tokens_out":496,"duration_ms":5745,"temperature":1.0,"reasoning_tokens":415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:34:26.794286+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $d=1$ with $\\lambda_0(t)=1+\\sin t$ and $\\phi(t)=\\alpha e^{-\\beta t}$, simulate many independent paths, and compare the empirical covariance density off the diagonal with the absolutely continuous part predicted by Proposition 4.6; a systematic mismatch beyond Monte Carlo error would refute the covariance formula. The diagonal mass $\\delta_{k=l}m^{(1)}_k + m^{(1)}_l * m^{(1)}_k$ predicted by the singular part can be checked in the same simulation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the cluster/population representation (4) that all Laplace-transform and moment proofs start from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the single-time multivariate characteristic-function formula that Theorem 2.2 extends to multiple times."},{"cited_title":"Exact and Asymptotic Analysis of General Multivariate Hawkes Processes and Induced Population Processes","cited_arxiv_id":"2106.03560","evidence_quote":"provides exact and asymptotic analysis of general multivariate Hawkes processes that the multi-time extension refines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"extends the cluster representation to marked Hawkes processes, justifying its use in the multivariate marked setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"derives multi-time second- and third-order moments for stationary multivariate Hawkes processes, which Proposition 4.5 extends beyond stationarity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives a probability generating functional for stationary spatial Hawkes processes analogous to Theorem 2.3."},{"cited_title":"David and H.N","cited_arxiv_id":null,"evidence_quote":"yields the order-statistics fact that Poisson-timed events are i.i.d., used in the Laplace-transform proofs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes that the conditional intensity characterizes the law of a multivariate point process, used to identify the process distribution."}],"review_version":1}