acceptodds
Under review as a conference paper at ICLR 2027

A Representer Theorem for Coarse-Grained Hawkes Processes

Abstract

Hawkes processes are a standard framework for modeling event sequences and estimating triggering kernels that characterize inter-event influences from event data. In many practical settings, however, observations are limited to event counts within fixed time bins or windows rather than exact event timestamps. This challenging observation regime has motivated growing interest in estimating triggering kernels on aggregated event data. Among the approaches developed for this setting, the recently proposed *coarse-grained Hawkes process* (CGHP) offers a parametric framework that combines computational efficiency with robustness to temporal aggregation. In this paper, we propose a nonparametric extension of CGHP within a reproducing kernel Hilbert space (RKHS) framework, enabling flexible estimation of triggering kernels from aggregated observations. By exploiting the mathematical structure of CGHP, we establish a representer theorem, a cornerstone of kernel methods, that reduces the infinite-dimensional estimation problem to a finite-dimensional one, providing the foundation for an efficient estimation algorithm. To the best of our knowledge, this is the first representer theorem for a Hawkes process model designed for aggregated event data. Experiments on synthetic and real-world datasets demonstrate that the proposed method achieves higher estimation accuracy than the state-of-the-art nonparametric approach, with particularly pronounced improvements under coarse temporal aggregation.

Primary area:
learning theory
Keywords:
Hawkes processes, kernel methods, representer theorem, aggregated data, event time series analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pCLeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies nonparametric estimation of Hawkes triggering kernels when observations consist only of event counts within fixed time bins. The authors extend coarse-grained Hawkes processes (CGHPs) to a reproducing kernel Hilbert space (RKHS), deriving a representer theorem that converts the functional optimization problem into a finite-dimensional problem involving weighted integrated kernels. Experiments on synthetic datasets, together with a social media dataset, show that the proposed method outperforms INAR under coarse temporal aggregation.

Critical strengths

The absence of exact event timestamps is an important limitation in many applications. Combining CGHP’s treatment of intra-bin interactions with RKHS-based function estimation provides a useful way to recover triggering patterns beyond restrictive parametric families.

Critical weaknesses
  1. The CGHP loss depends on the unknown functions through finitely many weighted integral functionals. When these functionals are bounded, the finite-dimensional representation under squared RKHS-norm regularization appears to follow from existing general representer theorems through an orthogonal decomposition argument. Then, Theorem 1 is primarily a specialization of an established principle to CGHP. The authors should clarify the substantive theoretical advance beyond this specialization. Moreover, Corollary 2 describes HH as positive semidefinite but directly uses H−1H^{-1}.

  2. The comparisons simultaneously change intra-bin modeling, function representation, and regularization. An ablation that removes Φ0\Phi_0 while retaining the same RKHS, regularization, and hyperparameter-selection procedure is missing. Hence, the current results do not clearly establish how much improvement comes from intra-bin modeling versus regularization.

INAR is fitted using Adam, although it admits a closed-form least-squares solution.

Questions for authors
  1. What conclusions or technical steps in Theorem 1 go beyond existing representer theorems for bounded linear functionals? Please also specify the conditions guaranteeing that HH is invertible and explain how ill-conditioning is handled in the implementation.

  2. The LDU parameterization guarantees a positive determinant, but how do you ensure nonnegative baseline intensities and nonnegative conditional intensities?

  3. Does the runtime advantage persist when INAR is implemented using its closed-form solution or an efficient linear algebra solver?

3 comments

We sincerely thank the reviewer for the valuable comments, to which we provide a detailed response below. In particular, we believe we can fully address the reviewer's primary concern about the paper's contribution by clarifying our position with respect to prior research.

Q1-a: What conclusions or technical steps in Theorem 1 go beyond existing representer theorems for bounded linear functionals?

As the reviewer correctly pointed out, the generalized representer theorem (Schölkopf et al., 2001) applies to regularized optimization problems in which the data-fitting terms depend on the latent functions via finitely bounded linear functionals. Theorem 1 is indeed a specialization of this general result to the coarse-grained Hawkes process (CGHP) setting.

We would like to emphasize that the principal contribution of our work is not the discovery of a new representer theorem, but rather the first demonstration that the problem of estimating Hawkes processes from aggregated event data admits a representer theorem formulation, along with the development of a concrete, computationally efficient estimation algorithm. Despite its practical importance, the analysis of aggregated event data remains relatively underexplored in the machine learning community. By establishing the applicability of kernel methods, a fundamental and widely used machine learning framework, to this challenging problem, we believe our work can help draw greater attention to this important research direction. This is precisely what we intended to convey in the Abstract and Introduction: "this is the first representer theorem for a Hawkes process model designed for aggregated event data." Our intention was to highlight the novelty of applying the representer theorem in this particular setting, rather than to claim a fundamentally new representer theorem.

Nevertheless, the reviewer's comment has made us realize that our original presentation could have been misleading about the nature of our theoretical contribution. To avoid this ambiguity, we will replace the sentences immediately preceding Theorem 1 (lines 160-161) with the following:

Theorem 1 follows by specializing the generalized representer theorem (Schölkopf et al., 2001), which applies to regularized optimization problems whose data-fitting terms depend on the latent functions through finitely bounded linear functionals.

We believe this revision will clarify the precise theoretical contribution of our work and its relationship to the existing literature. We would appreciate it if the reviewer could let us know about any remaining concerns.

Q1-b: Please also specify the conditions guaranteeing that HH is invertible and explain how ill-conditioning is handled in the implementation?

We sincerely thank the reviewer for identifying this important issue. The representer theorem established in Theorem 1 holds for positive semidefinite RKHS kernels. However, the optimization formulation in Corollary 2, expressed in terms of the coarse-grained kernel coefficients Φ\Phi (or equivalently BB), involves the inverse of the matrix HH and therefore requires HH to be strictly positive definite. We acknowledge that this requirement was not sufficiently clarified in the original manuscript. We will therefore revise Corollary 2 to explicitly state that the matrix HH is assumed to be strictly positive definite, thereby ensuring its invertibility. Furthermore, to provide a theoretical justification for this assumption, we will add the following proposition, which establishes a sufficient condition for the strict positive definiteness of HH in the case of shift-invariant RKHS kernels.

Proposition. Suppose that the RKHS kernel is shift-invariant and admits a spectral density that is strictly positive almost everywhere. Then, the matrix HH defined in Eq. (12) is strictly positive definite and hence invertible.

Proof. See Appendix G in our manuscript.

Importantly, both the Gaussian kernel used in our main experiments and the Laplace kernel considered in Table D1 satisfy the condition of this proposition, since their spectral densities are strictly positive over the entire frequency domain. Therefore, the invertibility of HH is theoretically guaranteed for the RKHS kernels employed in our experiments, and no modifications to the experimental methodology or results are required.

Interestingly, the reviewer's comment has also led us to identify a promising direction for future research. In Corollary 2, we formulate the estimation problem as an optimization over the coarse-grained kernel coefficients BB. Alternatively, it is possible to formulate the optimization directly in terms of the dual coefficients AA. Specifically, by using the relationship Bij:=HAij:B_{ij:}=HA_{ij:}, the RKHS regularization term can be rewritten from ∑i,jBij:⊤H−1Bij:\sum_{i,j} B_{ij:}^{\top}H^{-1}B_{ij:} to ∑i,jAij:⊤HAij:\sum_{i,j} A_{ij:}^{\top}HA_{ij:}. This alternative formulation eliminates the need to compute H−1H^{-1} and allows the estimation problem to be formulated using positive semidefinite kernels. A key remaining challenge is to develop an efficient mechanism for ensuring that the log-determinant term, log⁡∣Id−B::1∣\log|I_d-B_{::1}| for Bij:=HAij:B_{ij:}=HA_{ij:}, remains well-defined during iterative optimization. Although a straightforward approach is to employ projected gradient descent, developing a more computationally efficient optimization strategy is an important direction for future research. We will add this discussion to the Conclusion.

Critical Weakness: An ablation that removes Φ0\Phi_0 while retaining the same RKHS, regularization, and hyperparameter-selection procedure is missing. Hence, the current results do not clearly establish how much improvement comes from intra-bin modeling versus regularization.

We sincerely thank the reviewer for this important suggestion. In response, we have conducted an additional ablation study to isolate the contribution of intra-bin modeling from that of RKHS regularization. Specifically, we compared our proposed method, CGH$_{\mathrm{rkhs}}$, with an ablated variant in which the intra-bin interaction term Φ0\Phi_0 is constrained to zero throughout the optimization, while retaining the same RKHS framework, regularization, and hyperparameter selection procedure. We evaluated the estimation error δ2\delta^2 on the gamma-mixture scenario data with T=2000T=2000 and Δ∈{1,0.5,0.2}\Delta \in \{1,0.5,0.2\}. The following table shows the results:

δ2×103\delta^2 \times 10^3Δ\Delta = 1Δ\Delta = 0.5Δ\Delta = 0.2
CGH$_{\mathrm{rkhs}}$1.551.180.98
CGH$_{\mathrm{rkhs}}$ (Φ0=0\Phi_0 = 0)2.551.330.98

The results clearly demonstrate that the contribution of intra-bin modeling depends strongly on the temporal aggregation level. For relatively large bin widths (Δ=1\Delta=1 and 0.50.5), the ablated variant with Φ0=0\Phi_0=0 exhibits higher estimation errors than the original CGH$_{\mathrm{rkhs}}.Incontrast,forthesmallerbinwidth(. In contrast, for the smaller bin width (\Delta=0.2$), the difference in estimation accuracy becomes negligible. These findings are consistent with the underlying mechanism of CGHP. When the bin width is large, intra-bin interactions contribute substantially to the observed event counts, and neglecting these interactions introduces additional estimation errors. As the bin width decreases, however, the contribution of intra-bin interactions diminishes, making their explicit modeling progressively less important. We will include these additional experimental results and the corresponding discussion in the Appendix.

Q2: how do you ensure nonnegative baseline intensities and nonnegative conditional intensities?

To ensure the non-negativity of the estimated conditional intensity, we apply post-hoc clipping to the unconstrained intensity estimator λ^(⋅)\hat{\lambda}(\cdot) as max⁡(λ^(⋅),ϵ)\max(\hat{\lambda}(\cdot),\epsilon) for 0<ϵ≪10 < \epsilon \ll 1, while the non-negativity of the estimated baseline intensity is not enforced. The treatment of the conditional intensity is described in the second paragraph of Section 4.2. Note that when the true intensity function is larger than a small positive value ϵ\epsilon, clipping the estimated intensity at ϵ\epsilon does not increase the pointwise or integrated squared estimation error.

When triggering functions are allowed to take negative values, there are two major approaches to ensuring that conditional intensities are non-negative. One is to employ nonlinear Hawkes processes with non-negative link functions (e.g., Bonnet & Sangnier, 2025), which guarantee non-negativity but may entail substantially higher computational costs. The other is to apply post hoc clipping to intensity estimates obtained from linear Hawkes processes (e.g., Kim & Iwata, 2026), thereby offering greater computational efficiency while maintaining high estimation accuracy under moderate inhibition. Our method adopts the latter approach, and its effectiveness is demonstrated in the refractory scenario (Table 3 and Figure F3). Nevertheless, developing efficient nonparametric estimation methods for aggregated event data based on nonlinear Hawkes processes, particularly in regimes dominated by strong inhibition, remains an important direction for future work. We will add this discussion to the Conclusion to clarify the scope and limitations of our approach

Q3: Does the runtime advantage persist when INAR is implemented using its closed-form solution or an efficient linear algebra solver?

First, we emphasize that we do not claim our proposed method is inherently faster than INAR. Rather, our primary claim is that our method achieves greater robustness to the bin width Δ\Delta than INAR does while remaining computationally feasible. This point is already explicitly stated in the Experiments section (lines 359–362).

To further clarify the computational trade-offs, we discuss the computational complexity of INAR in more detail. The INAR estimation problem is formulated as the minimization of the conditional least-squares objective: ∑k=1n∣Xk−Δμ−Δ∑l=1k‾−1ϕ(lΔ)Xk−l∣2\sum_{k=1}^n \bigl| X_k - \Delta \mu - \Delta \sum_{l=1}^{\overline{k}-1} \phi (l\Delta) X_{k-l} \bigr|^2, where XkX_k denotes the event-count vector in the kk-th bin. Minimizing this objective yields estimates of the baseline intensity μ\mu and the triggering-kernel values ϕ(lΔ)l=1n‾\\{\phi(l\Delta) \\}_{l=1}^{\overline{n}}. When gradient descent is employed, as in our experiments, the computational complexity of INAR is O(Pd2nn‾)\mathcal{O} (Pd^2 n\overline{n}), or O(Pd2nlog⁡n)\mathcal{O} (Pd^2 n\log n) with fast Fourier transform (FFT) acceleration, where PP denotes the number of iterations required for convergence. In comparison, the computational complexity of our proposed method with FFT is O(P(d3+d2n‾2+d2nlog⁡n)+n‾3)\mathcal{O} (P(d^3+d^2\overline{n}^2 + d^2 n\log n) + \overline{n}^3). Thus, our method generally incurs higher computational costs per iteration. Nevertheless, across the four scenarios, our method required fewer iterations on average to converge, resulting in shorter observed runtimes than the gradient-descent implementation of INAR.

As the reviewer correctly points out, the INAR objective is quadratic in its parameters, allowing its minimizer to be obtained by solving a system of linear equations. Using a direct solver, the computational complexity is O((dn‾)3+nd2n‾2)\mathcal{O} ((d\overline{n})^3+n d^2 \overline{n}^2). This approach eliminates the dependence on the iteration count PP and can therefore be substantially faster than gradient descent. However, an important consideration is that direct solvers require the factorization of a matrix of size (dn‾+1)(d\overline n + 1), which incurs cubic computational complexity. Even when dd and n‾\overline n are individually moderate, their product can become prohibitively large. For example, when d=100d = 100 and n‾=500\overline{n} = 500, the resulting matrix dimension is approximately 50,000, making dense matrix factorization computationally intensive. In contrast, gradient-based methods avoid the need for such large-scale matrix factorizations and can therefore offer greater flexibility in high-dimensional or long-support settings.

We will incorporate this discussion into the manuscript to clarify the computational trade-offs between the two approaches.

Reviewer_XyHDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper extends the coarse-grained Hawkes process (CGHP) to a nonparametric formulation. Hawkes process is used to model signals with triggering event data and coarse-grained variants are used when event data is aggregated into bins. Authors use finite-support kernels and an RKHS-norm penalty on the CGHP Gaussian loss to prove a representer theorem. Based on my understanding, optimal kernels that can fit the model are finite combinations of bin-integrated kernel functions. An FFT-based implementation is also mentioned to be feasible. Experiments on three synthetic scenarios (exponential, gamma-mixture, refractory) and a retweet dataset are provided.

Critical strengths
  • It is my first time reading a paper on coarse Hawkes processes although I am familiar with Hawkes processes.I find the problem well-motivated and solution interesting.

  • The proposed approach is technically sounds and derivations are clear.

  • Computational cot analysis suggests that the proposed technique Is faster and suitable for large scale datasets.

  • Outperforms in coarse bin settings (events' timestamps are less granular).

Critical weaknesses
  • Comparisons with other techniques (for example earlier Monte-Carlo based or pure Hawkes process variants where event count is taken into account) could be provided.

  • Larger scale realistic datasets or binned variants of them (if they have fine resolution) could be considered.

Questions for authors
  • Could you elaborate on why the technique performs weaker under fine resolution? How can this be improved?

4 comments

We sincerely thank the reviewer for the positive and constructive comments. Below, we provide detailed responses to each point raised. We believe that the reviewer's primary concern regarding the experimental evaluation of our proposed method can be fully addressed through additional experiments and further clarification of our existing results.

Comparisons with other techniques (for example earlier Monte-Carlo based or pure Hawkes process variants where event count is taken into account) could be provided.

Following the reviewer's valuable suggestion, we have included the Monte Carlo expectation-maximization (MC-EM) method proposed by Shlomovich et al. (2022a) as an additional baseline and extended our experiments to compare its performance in the exponential and gamma-mixture scenarios. The MC-EM method is a parametric approach that assumes non-negative exponential triggering kernels. We implemented it in Python based on the MATLAB code publicly released by the original authors. Following the experimental configuration in the original paper, we set both the number of Monte Carlo samples and the number of EM iterations to 10.

The results are included in Tables H1 and H2 in our manuscript (Appendix H). As shown in Table H1, the MC-EM method consistently achieves the highest estimation accuracy in the exponential scenario. Unlike CGH${\mathrm{exp}},whichreliesonauniformapproximationofeventtimeswithineachbin,theMC−EMmethodexplicitlyaccountsforuncertaintyinthelatenteventtimesthroughMonteCarlosampling.Thisenablesmoreaccurateestimation,althoughattheexpenseofsubstantiallyhighercomputationalcosts.Incontrast,TableH2showsthattheMC−EMmethodisoutperformedbythenonparametricapproachesinmanyexperimentalsettingsunderthegamma−mixturescenario.ThisbehaviorisconsistentwiththatofCGH, which relies on a uniform approximation of event times within each bin, the MC-EM method explicitly accounts for uncertainty in the latent event times through Monte Carlo sampling. This enables more accurate estimation, although at the expense of substantially higher computational costs. In contrast, Table H2 shows that the MC-EM method is outperformed by the nonparametric approaches in many experimental settings under the gamma-mixture scenario. This behavior is consistent with that of CGH{\mathrm{exp}}$ and highlights the limitations of assuming an exponential functional form when the underlying triggering kernels exhibit more complex temporal structures.

Due to the substantial computational cost of the MC-EM method, completing the full experimental evaluation within the limited rebuttal period is challenging. We therefore currently report the means and variances of the estimation errors over 10 independent trials. We intend to extend the evaluation to 100 trials and report the corresponding results in the final version of the paper. Note that because the MC-EM method assumes non-negative triggering kernels, we excluded it from the refractory scenario, which involves inhibitory interactions.

We believe that these additional comparisons provide a more comprehensive evaluation of our proposed method and further clarify the trade-offs between parametric and nonparametric approaches to estimating Hawkes processes from aggregated event data.

Larger scale realistic datasets or binned variants of them (if they have fine resolution) could be considered.

We appreciate the reviewer's suggestion regarding the importance of evaluating the proposed method on large-scale real-world data. In fact, we have already conducted an experiment on a large-scale real-world social networking service (SNS) dataset containing 104,833 events, as reported in Section 4.2 of the manuscript. The experimental results demonstrate the effectiveness of our proposed method on real-world data and confirm its advantages over the existing nonparametric baseline. We would be grateful if the reviewer could revisit Section 4.2 and let us know whether this experiment adequately addresses the intended concern or whether a different aspect of large-scale evaluation was needed.

Could you elaborate on why the technique performs weaker under fine resolution? How can this be improved?

We sincerely thank the reviewer for raising this important question. We are also particularly interested in understanding this phenomenon and have already discussed it in the manuscript (lines 413-418). One promising hypothesis is that the difference in baseline intensity estimation between our proposed method and INAR may account for their relative estimation accuracy at fine temporal resolutions. Specifically, our method follows the CGHP framework, in which the baseline intensity μ\mu is estimated approximately using a relationship derived from the stationarity assumption of the Hawkes process, rather than being explicitly optimized within the likelihood-based estimation procedure. In contrast, INAR estimates the baseline intensity jointly with the other model parameters through its estimation objective. As the bin width Δ\Delta decreases, the advantages of CGHP over INAR in accounting for intra-bin interactions become less pronounced. Under such conditions, the difference in baseline intensity estimation may become a more influential factor, potentially explaining why INAR achieves slightly higher estimation accuracy in some fine-resolution settings.

A promising way to address this limitation is to extend our formulation to jointly optimize the baseline intensity and the triggering kernels. Specifically, the CGHP loss in Eq. (6) could be reformulated to explicitly incorporate the baseline intensity μ\mu. The functional optimization problem in Eq. (10) could then be extended to jointly estimate μ\mu and the untruncated triggering kernels φ\varphi. Importantly, because the dependence of the objective on the triggering functions would still be characterized through finitely many bounded linear functionals, an analogous representer theorem would remain applicable under the corresponding regularity conditions. Consequently, the resulting infinite-dimensional optimization problem could again be reduced to a finite-dimensional one, involving the joint optimization of d2nd^2 n dual coefficients and the dd-dimensional baseline intensity vector μ\mu. We regard this extension as a particularly promising direction for future research, as it may further improve estimation accuracy at fine temporal resolutions while preserving the flexibility of the RKHS-based nonparametric formulation. In the present work, however, we deliberately adopted the original CGHP framework to maintain consistency with its underlying estimation methodology. We therefore did not introduce an additional formulation that explicitly optimizes μ\mu.

We will add this discussion to the manuscript.

I thank the authors for the rebuttal. Authors address my questions regarding additional baselines and clarify the points related to realistic datasets. I raise the rating to 3.

Thank you very much for taking the time to carefully read our rebuttal and reconsider your assessment. We sincerely appreciate your thoughtful response and constructive feedback. Your comments have been very helpful in improving the paper.

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

This paper studies learning multivariate Hawkes processes from aggregated event data. It extends the recently developed coarse-grained Hawkes process (CGHP), which assumes parametric triggering functions, by modeling the triggering functions in an RKHS instead. The method is evaluated on synthetic and real-world datasets.

Critical strengths

The paper addresses a relevant problem: learning Hawkes processes from aggregated event data. It is fairly well written and easy to follow. My main concerns are the level of novelty and contribution, and the scalability of the method.

Critical weaknesses

The work builds on CGHP, which approximates the discrete-time count process with an AR model. Assuming the triggering functions lie in an RKHS is a valid extension, but a fairly straightforward one. Given the well-known representer theorem, adding a Tikhonov (RKHS-norm) penalty to the loss L_CG naturally yields a similar result, so Theorem 1 is essentially trivial. Also, triggering functions are commonly assumed to be non-negative, but the method imposes no such constraint; the functions are only assumed to lie in an RKHS.

In Corollary 2, the matrix H defined in (12) is described as SPD, but invertibility is not shown. Positive definiteness is not established. (If H is only symmetric positive semi-definite, invertibility does not follow.)

How crucial is the assumption of equal-length bins Delta? What happens when the bins have different lengths, as is often the case in practice?

The method appears to become computationally inefficient as d and n grow, and the experiments are low-dimensional (d = 2 or 3).

What is the average number of events per bin for the different values of Delta used in the experiments? I would expect the method to be quite sensitive to Delta

Tables 1, 2, and 3 do not show that the RKHS approach substantially improves the accuracy of the learned triggering functions. Simply fitting exponential kernels seems to do a decent job.

Questions for authors

Please see my comments.

5 comments

We sincerely thank the reviewer for the thoughtful and constructive comments. We provide detailed responses to each point below. In particular, we believe that the reviewer's primary concern about the paper's contribution can be fully addressed by clarifying how our work is positioned within the existing literature.

Given the well-known representer theorem, adding a Tikhonov (RKHS-norm) penalty to the loss L_CG naturally yields a similar result, so Theorem 1 is essentially trivial.

As the reviewer correctly pointed out, the generalized representer theorem (Schölkopf et al., 2001) applies to regularized optimization problems in which the data-fitting terms depend on the latent functions via finitely bounded linear functionals. Theorem 1 is indeed a specialization of this general result to the coarse-grained Hawkes process (CGHP) setting.

We would like to emphasize that the principal contribution of our work is not the discovery of a new representer theorem, but rather the first demonstration that the problem of estimating Hawkes processes from aggregated event data admits a representer theorem formulation, along with the development of a concrete, computationally efficient estimation algorithm. Despite its practical importance, the analysis of aggregated event data remains relatively underexplored in the machine learning community. By establishing the applicability of kernel methods, a fundamental and widely used machine learning framework, to this challenging problem, we believe our work can help draw greater attention to this important research direction. This is precisely what we intended to convey in the Abstract and Introduction: "this is the first representer theorem for a Hawkes process model designed for aggregated event data." Our intention was to highlight the novelty of applying the representer theorem in this particular setting, rather than to claim a fundamentally new representer theorem.

Nevertheless, the reviewer's comment has made us realize that our original presentation could have been misleading about the nature of our theoretical contribution. To avoid this ambiguity, we will replace the sentences immediately preceding Theorem 1 (lines 160-161) with the following:

Theorem 1 follows by specializing the generalized representer theorem (Schölkopf et al., 2001), which applies to regularized optimization problems whose data-fitting terms depend on the latent functions through finitely bounded linear functionals.

We believe this revision will clarify the precise theoretical contribution of our work and its relationship to the existing literature.

triggering functions are commonly assumed to be non-negative, but the method imposes no such constraint

First, we would like to clarify that the existing state-of-the-art nonparametric approach, INAR, also does not enforce non-negativity of the estimated triggering kernels. Indeed, the original INAR paper (Kirchner, 2017) explicitly reports instances in which the estimated triggering kernels take negative values. Thus, the absence of an explicit non-negativity constraint is not a "limitation" specific to our proposed method. More importantly, we wish to emphasize that allowing negative triggering function values is not inherently a drawback. In fact, recent research on Hawkes processes has increasingly focused on modeling inhibitory interactions, which naturally require triggering kernels that can take negative values.

When negative triggering functions are allowed, the essential challenge is how to ensure that the resulting intensity function remains non-negative. There are two major approaches to addressing this issue. The first approach is to introduce a nonlinear Hawkes process with a non-negative link function (e.g., Bonnet & Sangnier, 2025). This provides a principled way to guarantee the non-negativity of the intensity function, although it can entail substantially higher computational costs. The second approach is to apply post-hoc non-negative clipping to the estimated intensity function (e.g., Kim & Iwata, 2026). This approach facilitates computationally efficient estimation while potentially maintaining high estimation accuracy, particularly when inhibitory effects are moderate. Our proposed method follows the second approach, and we have explicitly evaluated its effectiveness in the presence of moderate inhibitory interactions. We refer the reviewer to Table 3 and Figure F3 for the refractory scenario, which demonstrate our method's ability to recover inhibitory triggering kernels.

We will incorporate this discussion into a new Section 2.5 to clarify the treatment of negative triggering kernels and further highlight the scope and contribution of our approach.

In Corollary 2, the matrix H defined in (12) is described as SPD, but invertibility is not shown. Positive definiteness is not established.

We sincerely thank the reviewer for identifying this important issue. The representer theorem established in Theorem 1 holds for positive semidefinite RKHS kernels. However, the optimization formulation in Corollary 2, expressed in terms of the coarse-grained kernel coefficients Φ\Phi (or equivalently BB), involves the inverse of the matrix HH and therefore requires HH to be strictly positive definite. We acknowledge that this requirement was not sufficiently clarified in the original manuscript. We will therefore revise Corollary 2 to explicitly state that the matrix HH is assumed to be strictly positive definite, thereby ensuring its invertibility. Furthermore, to provide a theoretical justification for this assumption, we will add the following proposition, which establishes a sufficient condition for the strict positive definiteness of HH in the case of shift-invariant RKHS kernels.

Proposition. Suppose that the RKHS kernel is shift-invariant and admits a spectral density that is strictly positive almost everywhere. Then, the matrix HH defined in Eq. (12) is strictly positive definite and hence invertible.

Proof. See Appendix G in our manuscript.

Importantly, both the Gaussian kernel used in our main experiments and the Laplace kernel considered in Table D1 satisfy the condition of this proposition, since their spectral densities are strictly positive over the entire frequency domain. Therefore, the invertibility of HH is theoretically guaranteed for the RKHS kernels employed in our experiments, and no modifications to the experimental methodology or results are required.

Interestingly, the reviewer's comment has also led us to identify a promising direction for future research. In Corollary 2, we formulate the estimation problem as an optimization over the coarse-grained kernel coefficients BB. Alternatively, it is possible to formulate the optimization directly in terms of the dual coefficients AA. Specifically, by using the relationship Bij:=HAij:B_{ij:}=HA_{ij:}, the RKHS regularization term can be rewritten from ∑i,jBij:⊤H−1Bij:\sum_{i,j} B_{ij:}^{\top}H^{-1}B_{ij:} to ∑i,jAij:⊤HAij:\sum_{i,j} A_{ij:}^{\top}HA_{ij:}. This alternative formulation eliminates the need to compute H−1H^{-1} and allows the estimation problem to be formulated using positive semidefinite kernels. A key remaining challenge is to develop an efficient mechanism for ensuring that the log-determinant term, log⁡∣Id−B::1∣\log|I_d-B_{::1}| for Bij:=HAij:B_{ij:}=HA_{ij:}, remains well-defined during iterative optimization. Although a straightforward approach is to employ projected gradient descent, developing a more computationally efficient optimization strategy is an important direction for future research. We will add this discussion to the Conclusion.

How crucial is the assumption of equal-length bins Delta? What happens when the bins have different lengths, as is often the case in practice?

We thank the reviewer for raising this interesting and practically important question. Both our proposed method and INAR, a leading existing nonparametric approach, assume equally spaced bins. Extending these methods to irregularly spaced bins is a nontrivial problem. In particular, both methods exploit the fact that bin boundaries are aligned across all dimensions. This structure enables event counts to be represented as vectors and facilitates concise, computationally efficient estimation algorithms based on vector and matrix operations. When bin widths or boundaries vary, this convenient structure is no longer directly available. Consequently, preserving computational efficiency becomes a major challenge in extending these methods to irregularly binned data. Some existing parametric approaches based on computationally expensive Monte Carlo methods (e.g., Shlomovich et al., 2022a) have been reported to accommodate irregular binning. However, developing an efficient nonparametric estimation method for irregularly binned Hawkes process data remains an open challenge for the broader research field, rather than a limitation unique to our approach. We regard this as an important direction for future work.

As a practical workaround, both INAR and our method can be applied to irregularly binned data through a preprocessing step. Specifically, one can assume that events are uniformly distributed within each original bin and use this assumption to impute event counts on a newly defined regular grid. The resulting regularly binned data can then be analyzed using either method. We will add this discussion to the Conclusion to clarify the current scope of our method and possible extensions to irregular binning.

The method appears to become computationally inefficient as d and n grow, and the experiments are low-dimensional (d = 2 or 3).

We appreciate the reviewer's concern regarding computational scalability. However, we would like to distinguish between scalability with respect to the number of bins, nn, and scalability with respect to the dimensionality, dd. First, as explicitly analyzed in Section 2.4, our method has linear or nearly linear computational complexity with respect to the number of bins nn, under a fixed effective kernel support. This favorable dependence on nn is comparable to that of INAR. We suspect that the distinction between the number of bins nn and the effective kernel support size n‾\overline{n} may not have been sufficiently clear in our presentation. We will clarify their respective roles in the complexity analysis in Section 2.4.

On the other hand, we agree with the reviewer that our method is less scalable with respect to the dimensionality dd. Specifically, its computational complexity grows cubically with dd. Therefore, the current formulation is primarily intended for problems of up to hundreds of dimensions. For comparison, INAR has quadratic complexity, O(d2)\mathcal{O}(d^2), with respect to dd. Although this is more favorable than our method's cubic scaling, INAR also faces substantial computational challenges in high-dimensional settings. More generally, methods that estimate all pairwise triggering functions nonparametrically inevitably involve at least a quadratic number of interactions in dd. Thus, achieving computational efficiency in high-dimensional nonparametric Hawkes process estimation is a fundamental challenge. We will expand Section 2.4 to make these computational trade-offs more explicit.

Furthermore, in response to the reviewer's suggestion, we are currently preparing additional experiments in a higher-dimensional setting. We expect to report the results within the next few days and would appreciate the reviewer's patience in the meantime.

What is the average number of events per bin for the different values of Delta used in the experiments? I would expect the method to be quite sensitive to Delta

We thank the reviewer for suggesting a quantitative investigation of sensitivity to the bin width Δ\Delta. In fact, one of the central findings consistently supported by Tables 1-4 is that our proposed method substantially mitigates the sharp deterioration in estimation accuracy exhibited by INAR as the bin width increases. In other words, our method is less sensitive to the temporal aggregation level Δ\Delta than the existing nonparametric baseline. To quantify this observation more clearly, we conducted an additional sensitivity analysis. Specifically, for each experimental setting in Tables 1–4, we fitted a log-linear regression model relating the bin width Δ\Delta to the estimation error δ2\delta^2 or the test log-likelihood llll: log⁡δ2=alog⁡Δ+b\log \delta^2 = a\log\Delta + b or log⁡ll=alog⁡Δ+b\log ll = a\log\Delta + b. We then used the absolute value of the estimated regression coefficient, ∣a∣|a|, as a measure of sensitivity to the bin width, where a smaller value of ∣a∣|a| indicates that the model is less sensitive to Δ\Delta. The results are summarized below.

Table 1: Sensitivity ∣a∣|a|

T = 1000T = 2000T = 10000
INAR0.280.420.76
CGH$_{\mathrm{rkhs}}$0.110.140.30

Table 2: Sensitivity ∣a∣|a|

T = 1000T = 2000T = 10000
INAR0.200.380.89
CGH$_{\mathrm{rkhs}}$0.110.220.60

Table 3: Sensitivity ∣a∣|a|

T = 1000T = 2000T = 10000
INAR0.210.310.66
CGH$_{\mathrm{rkhs}}$0.190.180.19

Table 4: Sensitivity ∣a∣|a|

INAR0.29
CGH$_{\mathrm{rkhs}}$0.19

Remarkably, our proposed method exhibits a smaller sensitivity coefficient than INAR across all experimental settings considered. These additional results provide quantitative evidence that our method is consistently more robust to changes in the level of temporal aggregation. We believe this analysis further strengthens one of the key advantages of our proposed approach, and we will include the results in the Appendix. For completeness, we also report the average number of events per bin for the different values of Δ\Delta, as requested by the reviewer. The following results correspond to T=1000T=1000 in Tables 1–3.

ExperimentΔ=1\Delta=1Δ=0.5\Delta=0.5Δ=0.2\Delta=0.2Δ=0.1\Delta=0.1
Table 12.331.160.470.23
Table 22.441.220.490.24
Table 30.310.150.060.03

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.