acceptodds
Withdrawn from ICLR 2027

Directional Information in Temporally Aggregated Hawkes Processes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Timestamp aggregation removes event order but can preserve directional information in event counts. We characterize this distinction in a stationary, mean-matched, feed-forward Hawkes family observed through independent windows. For multiple bins, we derive the exact first-order directional score and its finite-window information. For kernels with finite mean lag , information per unit exposure decays as when the long-window limit precedes the coarse-resolution limit. With one bin, the first-order score vanishes, but a nonzero second-order tangent combines excess dispersion with a mixed third-degree count moment. We derive the exact Gaussian testing limit at excitation strength and a statistic attaining it. The mixed term supplies a calculable fraction of local information absent from dispersion alone. Both tests retain local efficiency after estimating the common rate. We characterize which count summaries preserve the tangents and which discard their distinctions. Paired count-law experiments quantify the mixed term's contribution and show slow finite-sample convergence. Repeated calibration experiments assess the grid-specific assertion control given by test inversion. Misspecification experiments show substantial failures outside that grid. The results separate temporal ordering information, higher-order count information, and the scope of calibrated reporting.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Hawkes processes, temporal point processes, temporal aggregation, directional dependence, hypothesis testing, information-theoretic bounds, local asymptotic theory, statistical efficiency, simulation-based inference, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6AJSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how much directional information about excitation remains when a Hawkes process is observed through temporally aggregated event counts. Focusing on a stationary, mean-matched 2-variate Hawkes process, the authors derive an exact first-order directional score function for the multi-bin setting and show that the directional information per unit exposure decays cubically with the bin width under finite-mean kernels. When only a single bin is observed, the first-order directional information vanishes. Still, the authors identify a nonzero second-order tangent involving both excess dispersion and a mixed third-degree count moment, leading to a fourth-root local testing regime. They further establish locally efficient tests, study the effect of estimating the common event rate, and characterize which count summaries preserve the relevant directional information. Synthetic experiments examine the predicted local testing risks, the contribution of the mixed moment, calibrated directional reporting, and robustness under model misspecification.

Critical strengths
  • The paper provides a technically substantial theoretical characterization of directional information in temporally aggregated Hawkes processes, including an exact first-order directional score and a cubic decay law for information under coarse temporal resolution.
  • The single-bin analysis is particularly interesting: although first-order directional information vanishes, the paper identifies a nonzero second-order signal involving both dispersion and a mixed third-degree count moment. It derives the corresponding N−1/4N^{-1/4} local testing limit and an attaining statistic.
  • Extensive synthetic experiments support the theoretical results.
Critical weaknesses
  • The main weakness is the presentation. Several central concepts and notations, including the directional score, directional tangent, count law, the (±)(\pm) subscript, JJ (the abbreviation of JΔ,T(⋅)?)J_{\Delta,T}(\cdot)?), PaTP_a^T, CΔC_{\Delta}, and the equal-prior optimal testing risk, are introduced with little or no intuitive explanation. More importantly, the paper does not sufficiently articulate the practical or statistical significance of its main theoretical results. Key findings, such as the cubic decay of directional information and the local optimality of the score test, are stated formally, but their broader implications are only weakly discussed. The authors should explain more explicitly what these results enable us to do, what aspects of inference or data collection they can improve, and how they should influence the design of methods for aggregated event data. Because of the current presentation, even readers familiar with Hawkes processes may find the theoretical development unnecessarily difficult to follow. In its present form, the paper is accessible mainly to a relatively narrow audience already well versed in local asymptotic statistical theory. This is particularly unfortunate given that the underlying theoretical results appear to be quite interesting and potentially valuable.

  • The positioning with respect to prior work on directional inference from aggregated or discretized observations is incomplete. Prior studies, including (Santitissadeekorn et al., 2023/2025), have investigated network reconstruction from count data, while related work has also considered causal discovery under temporal aggregation. The authors should explain more explicitly how their local information analysis differs from these global identification, network-reconstruction, or parameter-estimation approaches. A more systematic comparison with closely related work is needed to make the precise novelty of the present contribution clearer. In addition, given the focus on binned Hawkes observations, it seems unusual not to refer to (Kirchner, 2016/2017) and clarify how the present analysis relates to that literature.

  • Although I do not regard this as a weakness that should materially lower the score, the theoretical and empirical scope is relatively narrow. The main theoretical results are developed for a stationary, mean-matched feed-forward family, and the empirical evaluation is entirely synthetic. I believe these restrictions should be stated more explicitly as limitations in the main text, together with a brief discussion of which aspects of the theory are expected to extend to more general Hawkes models and real-world data.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Statistical implications, prior work, and presentation

Thank you for recognizing the technical contribution. We agree that its meaning and relation to earlier work should be clear before the theorem statements.

  1. A likelihood-first explanation. We observe NN independent windows, each containing mm bin-count pairs. In the forward model, type 1 excites type 2. In the reverse model, the type labels are exchanged. Both marginal mean rates are ν\nu. The signs ++/−- denote these orientations, not time reversal. Pa,±TP^T_{a,\pm} is the event-path law, CΔC_\Delta maps paths to counts, and Qa,±,ΔTQ^T_{a,\pm,\Delta} is the resulting count law.

Writing its probability mass as pa±p_a^\pm, the directional score is the difference of likelihood derivatives at zero excitation, D=∂alog⁡(pa+/pa−)∣a=0+,J=E0D2.D=\left.\partial_a\log(p_a^+/p_a^-)\right|{a=0^+},\qquad J=\mathbb E_0D^2. Thus “information” means the squared size of a specified local directional contrast. The “tangent” is its leading distributional term. RN∗R_N^* is the smallest average misclassification probability over all tests when the two orientations have equal prior probability. JJ abbreviates JΔ,T(h)J{\Delta,T}(h). These objects should be introduced together before the two regimes.

  1. Statistical and practical consequences. The cubic law concerns first-order information per unit exposure. In the long-window, coarse-resolution regime, doubling bin width divides that information rate by approximately eight. Preserving first-order local testing power then requires approximately eight times the exposure. This is an iterated-limit comparison, not a universal finite-amplitude sample-complexity law. In particular, our explicit finite-sample envelope is vacuous in the main moderate-amplitude decoder setting, as Table 2 states.

At one bin, the first-order directional signal vanishes, but dispersion and a mixed third-degree moment produce an attainable fourth-root experiment. The information split tells a representation designer what dispersion alone misses. At T=τ=1T=\tau=1, the mixed term contributes 45.51%, 17.27%, and 4.96% of the second-order information at expected counts 0.25, 1, and 4. Appendix S identifies an added coordinate that preserves the efficient statistic exactly. This guides summary design without asserting neural efficiency.

We have now tested that connection with a matched single-bin learning ablation. Across six fixed-kernel binary settings, both linear and neural decoders receive the same development data and 20,000 independent held-out datasets per setting, with three training seeds. At ν=1,N=65536\nu=1,N=65536, the added coordinate reduces neural error from 18.15% to 16.82%, a paired reduction of 1.32 percentage points (SE 0.17). At ν=0.25\nu=0.25, the corresponding reduction is 4.38 points (SE 0.24). These uncertainties are over test datasets, conditional on the fitted models. Linear decoding remains competitive: the result supports retaining the missing moment, not neural superiority.

  1. Positioning against the work you identified. The omissions are well taken. Kirchner (2016) establishes the Hawkes/INAR connection, and Kirchner (2017) develops multivariate count-based estimation through INAR/VAR approximation and conditional least squares. We instead derive local directional testing constants from the exact finite-window count law at fixed resolution.

Santitissadeekorn et al. (2023) reconstruct influence networks from counts using ensemble filtering. Santitissadeekorn, Short and Lloyd (2025) develop batch and sequential network reconstruction, including discrete-time Hawkes/Cox models. Their reconstruction procedures and our weak-excitation testing experiment answer different questions. We do not claim to introduce directional network inference from counts.

Gong et al. (2017) establish causal identifiability for temporally aggregated non-Gaussian VAR models. Fan, Gong and Zhang (2024) study when aggregation preserves or distorts causal relations in an i.i.d. setting. We should distinguish these identifiability questions from local statistical difficulty and should not equate predictive Hawkes direction with intervention-based causality. Koyama (2025) supplies the coarse-grained second-order modeling context. A dedicated comparison of observation models, targets, assumptions, and guarantees will clarify these distinctions.

  1. Scope and extensions. We retain the restrictions to a stationary, mean-matched feed-forward family and synthetic evaluation. Theorem 6 handles an unknown common rate, not unequal unknown rates or arbitrary kernels. Conditional-score projection is a general principle, but feedback, nuisance parameters, or higher dimension can alter the cancellations and information constants. We do not claim that the cubic or fourth-root laws transfer unchanged. Real-data validation and a sharp many-bin transition remain open extensions of this work, not consequences of the present theorems.

Thank you for the detailed response and additional experimental results. I appreciate the authors' efforts to address my concerns.

The explanations of the statistical implications of the theoretical results, particularly the cubic information decay and the role of the mixed third-degree moment, are convincing. The additional ablation study provides meaningful empirical support for the theory's practical relevance. I also find the clarification of the differences from prior work satisfactory.

However, my primary concern remains the manuscript's readability and accessibility, which the other reviewers have also identified as major weaknesses. While the response provides helpful clarifications, it is difficult to assess whether this concern has been adequately addressed without seeing how these explanations are integrated into the revised manuscript. In particular, the key concepts, notation, and practical implications should be clearly introduced and discussed in the main text.

Reviewer_rwRAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies different forms of information which describe the directional excitation direction of Hawkes processes in the count-aggregated setting. Hawkes processes are self-exciting. A bivariate Hawkes process can model "directional effects", where (only) one dimension influences the other. In the count-aggregated setting, the paper looks at quantities associated with gradients (and higher order gradients). They study the first order gradient; and the second order gradient when the entire window of observation is count-aggregated.

Critical strengths

To the best of my knowledge, the results concerning the directional information are new and could be useful. These results could possibly be leveraged for nice statistical tests.

Critical weaknesses

I have 3 main issues with the current papers: the presentation/clarity, the lack of motivation, and limitations of the model setting itself.

Firstly, the presentation is very poor in my opinion. The paper has a very opaque structure and usage of language. Perhaps the biggest problem is that (at least from my readings) the notion of "direction information" is never explicitly defined in any formal sense. It took quite sometime before I identified that it corresponded to differences of likelihood gradients after looking at Theorem 1. Perhaps I could have made an inference that the "score" in score information referred to the Fisher score (however, in ML there are other quantities called the score, eg, the Stein score), but the paper never even talks about likelihoods around that section. The whole concept of a "forward" and "reverse" model was not clear in the writing of the paper. I haven't heard of this terminology before. Furthermore, the ±\pm notation is used without any explanation (which I believe is the distinguish between forward and reverse excitation directions). The interchanged usage of "tangent", "information", and "direction" is also extremely confusing. There are also many instances where the notation is used without clear definitions. Some I could infer from context (CΔC_\Delta, Φ\Phi), but I was still confused by other notation (ZθbZ_{\theta b} in (19), JJ in Theorem 7).

Second, there is a lack of motivation on why we are studying this problem in the first place. Part of the problem is the presentation where everything is so opaque. But I still think a better job could be done for the paper's approach. For example: Why don't we just fit a fully self-exciting bivariate Hawkes process (with same kernel) and examine fitted parameter differences to infer "direction"? What is the point of examining (and fitting with a Hawkes process) when the entire observation interval is count-aggregated? It further doesn't help that Sec 3 & 4 do not discuss the significance of the results. The examples with specific kernels are appreciated, but this was not sufficient to understand why these result mattered.

Lastly, the setting is a bit limited. There is a specific structured of excitation specified in (1). More flexible variants could be envisioned, e.g., multiple dimensions d>2d > 2, add at least self-excitation within the same dimension. Its also hard to see why the analysis of a "single bin" matters. In this scenario, shouldn't we just fit a Poisson process? I think this should be consistent with Rizoiu et al. , 2022.

Questions for authors

Why don't we just fit a fully self-exciting bivariate Hawkes process (with same kernel) and examine fitted parameter differences to infer "direction"?

What is the point of examining (and fitting with a Hawkes process) when the entire observation interval is count-aggregated?

2 comments
AuthorsOpenReview ↗

Clarifying the count experiment and the Poisson comparison

Thank you. We accept that the exposition assumes too much familiarity with both Hawkes processes and local statistical experiments. We address the model, the meaning of information, and the Poisson comparison explicitly below.

  1. The model and the statistical target. Let Ni(A)N_i(A) count events of type ii in interval AA. In the forward model, type 1 is a stationary Poisson source of rate ν\nu. Each source event increases the future type-2 intensity by ah(t−s)ah(t-s). Type 2 also has an independent background of rate ν(1−a)\nu(1-a). The reverse model exchanges the type labels. Thus both types have mean rate ν\nu in both models. The signs ++ and −- label these two orientations. They do not mean reversal of time, and there is no self-excitation in this family.

Let pa+(z)p_a^+(z) and pa−(z)p_a^-(z) be the probability masses of one observed vector of bin counts. At a=0a=0 they coincide. The directional score and information are D(z)=∂alog⁡pa+(z)pa−(z)∣a=0+,J=E0[D(Z)2].D(z)=\left.\partial_a\log\frac{p_a^+(z)}{p_a^-(z)}\right|_{a=0^+},\qquad J=\mathbb E_0[D(Z)^2]. Here “score” means a likelihood derivative with respect to excitation strength, not a derivative with respect to the observation as in a Stein score. DD is the difference of the two parameter scores, and JJ is the squared size of that directional contrast, not the full Fisher information for all Hawkes parameters. A “tangent” is the leading term by which the two observed laws separate as excitation grows from zero. Theorem 1 computes that term. Theorem 5 computes the second-order term when the first vanishes.

For the remaining notation: CΔC_\Delta maps an event path to its bin counts. Pa,±TP^T_{a,\pm} denotes the exact-time window law, and Qa,±,ΔTQ^T_{a,\pm,\Delta} its count law after applying this map. Φ\Phi is the standard normal distribution function. ZθbZ_{\theta b} in Eq. (19) is the bbth independently simulated calibration dataset under candidate θ\theta, with the same sample size and observation design as the data. In Theorem 7, JJ abbreviates JΔ,T(h)J_{\Delta,T}(h). These definitions belong before their first use.

  1. Why a single bin is not a Poisson estimation problem. “Single bin” means one count pair from each of NN independent stationary windows, not one count pair in total. Conditional on the complete source process, the target count is Poisson with a random mean. Marginalizing the source does not preserve a Poisson distribution. In fact, EX=EY=νT,Var⁡(X)=νT,Var⁡(Y)=νT+a2νβ2,\mathbb E X=\mathbb E Y=\nu T,\quad \operatorname{Var}(X)=\nu T,\quad \operatorname{Var}(Y)=\nu T+a^2\nu\beta_2, where β2=∫RpT(s)2ds>0\beta_2=\int_{\mathbb R}p_T(s)^2ds>0 and pT(s)=∫0Th(t−s)1s<tdtp_T(s)=\int_0^T h(t-s)\mathbf1{s<t}dt. The reverse orientation swaps these variances. A Poisson model preserving only the means discards this difference and the mixed count moment used by the optimal local test.

This distinction is consistent with Rizoiu et al. (2022, Section 4.1), rather than contradicted by it. Their Mean Behavior Poisson Process is a surrogate with corresponding parameters, not an assertion that the aggregated Hawkes count law is Poisson. Their paper explicitly distinguishes the Hawkes and Poisson likelihoods.

  1. Why not just fit a larger Hawkes model? Fitting self-excitation and both cross-excitation kernels is a useful estimation problem. Kirchner (2017) and Santitissadeekorn et al. (2025) provide relevant count-based approaches. Our question is different: how accurately can any method distinguish two specified directions from the available counts, and what statistic attains that accuracy? An estimated parameter difference alone does not establish that testing limit or its dependence on timestamp resolution.

The restricted family isolates this question without allowing different mean rates to reveal the answer. It is not a claim that recurrent or higher-dimensional models are unnecessary, or that our constants apply to them.

  1. Why the results matter. The cubic law quantifies loss of first-order information per observation time, with the long-window limit taken first. With one bin, temporal antisymmetry disappears but count fluctuations retain second-order direction. Accordingly, the attainable local excitation scales are N−1/2N^{-1/2} and N−1/4N^{-1/4}. The paper supplies the attaining tests, not only statistics that might later be used for testing. It also identifies which summaries preserve those tests and why marginal dispersion alone omits information.

We agree that these consequences should accompany Sections 3 and 4. The conclusions remain limited to the stated family, local regimes, and independent windows. Synthetic evidence is not real-data validation. We would welcome clarification of any remaining mathematical soundness concern after distinguishing the full count law from its Poisson mean approximation.

Thank you for the response.

Marking that your second response (2) answers my thought about Poisson processes, the question did not make sense in hindsight.

However, similar to other reviewers, believe that the largest weakness of the paper remains: its presentation. In fact, on checking on the second response, I found another notational piece of confusions in the overloading of NN vs N1N_1. It is rather hard to see how well the manuscript can be revised to improve clarity with the current response. It is also hard to see how the motivation would be incorporated into the paper. I'll note here that I understand the purely "mathematical explanation" of the results (response (4)), but how this is useful in a practical sense isn't clear from the current version of the paper.

Reviewer_QP7ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper looks at the setting with two processes, where one is dependent on the other. It asks whether you can tell which of the two processes is dependent when only binned counts are seen (discretising the time scale and aggregating within each step). In the model, one process is Poisson, each of its events briefly raises the event rate of the other, and both have the same mean rate. With multiple bins, the paper shows that directional information per unit time decays like 1/Δ31/\Delta^3 as the bins get wider. With a single bin, there is no timing information left, but direction is still detectable: the dependent process's counts vary more than a Poisson process's would, and this extra variation is tied to the other process's count. For both cases, it derives tests that are asymptotically optimal when the excitation is weak. It also trains a small neural network (MLP) on summary statistics of the counts and shows by example that these summaries cannot recover the optimal single-bin test. All experiments use simulated data. The multi-bin test reaches its predicted error rate. The single-bin test approaches its rate slowly but comes close to a numerical likelihood. Two simpler decoders (a Gaussian-moment fit and a coarse-grained method that needs no training) are at least as accurate as the MLP at every bin width.

Critical strengths

The paper shows that the direction of dependence can still be detected from a single bin, where all timing information is lost. The signal comes from count dispersion and from how that dispersion relates to the other process's count. For two or more bins per window, it shows that directional information per unit of observation time decays like 1/Δ31/\Delta^3 as the bin width Δ\Delta grows. For both cases it derives tests that are asymptotically optimal under weak excitation, with explicit limiting error rates. Simulations support these rates, though the single-bin test approaches its rate slowly.

Critical weaknesses

1. The setting is narrow and the learning part doesn't add much.
The theory covers two processes where one excites the other, with a known kernel, and all experiments are on simulated data. The only tests outside this model are small changes to the simulation (Table 6), and the results are mixed. For example, if the true rate is ν=2\nu = 2 rather than 1, the main procedure reports the direction incorrectly 50% of the times. The guarded version on the other hand predicts on direction altogether. The MLP doesn't add much either. Two simpler decoders (Gaussian-moment and coarse-grained) are at least as accurate at every bin width (Table 24). The nine summary features it uses miss the mixed term of the single-bin test. The paper proposes a tenth feature that keeps it (App. S) but never trains a decoder with it. Further, I don't see any decoder run on single-bin data. The confidence sets use a standard method (Dalmasso et al., 2020). So the contribution is mainly the theory for this specific model.

2. The paper is hard to follow.
I had trouble following the setup. Symbols like N1N_1 and CΔC_\Delta are used without being defined, and the whole model is given in two short paragraphs at the start of Section 2. A reader who doesn't already know Hawkes processes will be lost here. Related work is a single paragraph, which says what is not new but does not explain clearly how the results differ from the cited work. I'd suggest defining the model properly and expanding the related-work paragraph into a section.

Parts of the appendix read like notes on earlier versions of the work rather than a description of the study. They refer to things that are never described, e.g. "the referee's numerical checks" (p. 54), checkpoints "unchanged from the original study" (p. 52), records that "are now saved" (p. 56), and scripts named run_round2_paired.py and verify_round2.py. This text should be rewritten or removed.

Questions for authors

Clarification questions

  1. Page 2 says the contribution is the exact directional tangents and their testing constants, and also that the lag weights (Koyama, 2025) and the fourth-root rate (Rotnitzky et al., 2000) are known. Which steps in the proofs of Theorems 1 and 5 are new? For example, is the symmetric-mixture argument on p. 5 new?

  2. No decoder is run on single-bin data. All decoder experiments have at least four bins per window. Could you run the nine- and ten-feature decoders with a single bin? This experiment would show whether the single-bin theory matters for the learned decoders at all.

Suggestions for future work

  • A real-data example.
  • Tests that work with an unknown kernel or unequal mean rates.
1 comment
AuthorsOpenReview ↗

Novelty, the single-bin ablation, and scope

Thank you for identifying the main distinction and for the concrete questions about novelty and the missing ablation. We agree that the paper should stand on its statistical contribution, not on a claim that an MLP improves on every count-based estimator.

  1. What is new in Theorems 1 and 5? The aggregation weights themselves are not new. Theorem 1 derives the directional likelihood score of the exact observed-count experiment, including the unobserved stationary past. Projecting the fine-time scores onto counts reveals that the within-bin terms are identical under the two orientations and cancel. Orthogonality then gives the exact finite-window information constant. The finite-mean-kernel cubic law and the attaining test follow from this calculation, rather than from an approximate binned likelihood.

Theorem 5 identifies the next directional term when the first-order contrast vanishes. D(2)D^{(2)} combines dispersion with a mixed third-degree moment, with information J(2)=β22/T2+χ22/(νT3)J^{(2)}=\beta_2^2/T^2+\chi_2^2/(\nu T^3). The two integrals distinguish stationary-past from observed-parent contributions. We derive the exact optimal risk Φ(−γ2J(2)/2)\Phi(-\gamma^2\sqrt{J^{(2)}}/2) at aN=γN−1/4a_N=\gamma N^{-1/4} and a test attaining it. The fourth-root rate alone is classical (Rotnitzky et al., 2000).

The symmetric-mixture device is not claimed as a new general asymptotic method. It handles a specific obstacle: when the shared first derivative is nonzero, as for exponential kernels, the fourth-root alternatives are not contiguous to the product Poisson null. Appendix K controls the third derivative of the observed square-root density and establishes the expansion needed for comparison through the moving mixture. The model-specific calculation and its justified testing limit, not symmetrization alone, are the contribution.

  1. The requested single-bin learning experiment. We have now fitted both nine- and ten-feature decoders at one bin. This is a binary orientation experiment with fixed known kernel and non-directional parameters. We use T=τ=1T=\tau=1, ν∈0.25,1,4\nu\in{0.25,1,4}, and N∈4096,65536N\in{4096,65536}, with amplitudes on Theorem 5's local curve. Each setting has 800 development datasets per orientation, an 80/20 training/validation split, independent calibration, and 20,000 independent test datasets sampled from the numerical stationary count law. Linear and neural decoders use identical data and three training seeds. The only added input is the mixed-moment coordinate.

At ν=1,N=65536\nu=1,N=65536, neural error falls from 18.15% to 16.82%, a paired reduction of 1.32 percentage points (SE 0.17). Linear error falls from 18.07% to 16.74%. At ν=0.25,N=65536\nu=0.25,N=65536, neural error falls from 23.58% to 19.20%, a reduction of 4.38 points (SE 0.24). These SEs average the three fitted models' losses within each test dataset and measure test-sampling uncertainty, not independent training-population uncertainty.

The added feature lowers mean error for both decoder classes in all six settings. Linear decoding remains competitive. This supports the theorem's prescription for what to retain, not neural superiority or a guarantee that training attains the local optimum. It also goes beyond the summary-collision example, which establishes a representational distinction but cannot substitute for training.

  1. Narrow scope and misspecification. The family is deliberately mean-matched and feed-forward so that direction is not revealed by different marginal means. We do not claim general network recovery. Theorem 6 permits estimation of the common rate for the score tests, but does not recalibrate a decoder trained at a fixed rate. Thus Table 6's 50% wrong-report rate at ν=2\nu=2, and the guard's complete abstention, are genuine failures of that frozen candidate family, not successful off-grid inference. We retain them. Unknown unequal rates, flexible unknown kernels, and real-data validation remain limitations.

We also agree that test inversion is established methodology (Dalmasso et al., 2020), and simulation-trained Hawkes summary inference has direct precedent in Lambe et al. (2026). Koyama (2025) provides the coarse-grained modeling precedent. Our comparison should identify these roles explicitly rather than present the MLP or calibration wrapper as the novelty.

  1. Presentation and appendix. We accept the criticism of the compressed setup. Ni(A)N_i(A) counts type-ii events in interval AA, CΔC_\Delta maps event paths to bin counts, and ++/−- exchange the source and target labels, not the direction of time. These definitions should precede the likelihood and score, with related work in a separate subsection. The cleaned manuscript removes the cited revision-history passages, filenames, and procedural commentary. The supplement is described through its scientific contents. We appreciate the opportunity to make the contribution substantially easier to assess.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.