Nonlinear Self-Excited Hawkes Processes

• Nonlinear self-excited Hawkes processes are point processes where the intensity is a nonlinear, history-dependent function, enabling complex modeling of event clustering with both excitation and inhibition.
• They exhibit diverse regimes—from sublinear to explosive—with rigorous limit theorems such as Gaussian fluctuations and large deviations underpinning their statistical behavior.
• Applications in neuroscience, finance, seismology, and network science benefit from advanced inference techniques like Bayesian nonparametrics and neural surrogates for flexible model estimation.

A nonlinear self-excited Hawkes process is a class of point processes in which the conditional intensity at time $t$ is a nonlinear, history-dependent functional of previous events. These processes generalize classical linear Hawkes models by allowing nonlinear or state-dependent "link" functions, richer memory kernels, excitation and inhibition, and complex interactions in multivariate and marked extensions. Nonlinear self-excited Hawkes processes capture a broad array of phenomena in fields such as neuroscience, finance, seismology, and network science, where event clustering, feedback, and power-law statistics are prevalent.

1. Formal Definition and Existence Theory

Let $\{N_t\}_{t\ge0}$ denote a simple point process adapted to its natural filtration. The canonical univariate nonlinear Hawkes process is specified by a link (rate) function $\phi:\mathbb{R}_+\to\mathbb{R}_+$ and an excitation kernel $h:[0,\infty)\to[0,\infty)$, yielding the conditional intensity

$$\lambda_t = \phi\left(\int_0^{t-} h(t-s)\,dN_s\right),$$

with the process satisfying the compensator condition $$\mathbb{E}[N(a,b]|\mathcal{F}_a] = \mathbb{E}[\int_a^b \lambda_s\,ds|\mathcal{F}_a]$$ for all $a < b$.

For existence and uniqueness, standard assumptions are:

• $h$ is nonnegative, càdlàg, decreasing, and integrable ($\|h\|_1 < \infty$);
• $\phi$ is nonnegative, increasing, and Lipschitz with constant $a$ such that $a \|h\|_1 < 1$.

Under these, the process is stationary, ergodic, and nonexplosive (Zhu, 2012, Zhu, 2013).

Extensions include:

• Multivariate versions, with intensity for each $k$ given by

$$\lambda^k_t = \phi_k\left(\nu_k + \sum_{l=1}^K \int_{-\infty}^{t^-} h_{lk}(t-s) dN^l_s\right),$$

supporting both excitation ($h_{lk} > 0$) and inhibition ($h_{lk} < 0$) (Sulem et al., 2021, Chen et al., 2017).

• Marked and path-dependent processes, where the conditional intensity is determined by the history of event times and marks, potentially with highly nonlinear structure (Clark et al., 28 May 2025).

Recent results relax the global Lipschitz requirement on $\phi$, establishing existence under continuity and subcritical linear growth: $\Phi(x) \leq C + Lx$ with $L\|h\|_1 < 1$ (Robert et al., 2022).

Processes with superlinear activations (e.g., polynomial $\phi(x) = a x^k,\,k>1$) are explosive: there is no stationary solution, and events may accumulate in finite time (Robert et al., 2022).

2. Regimes, Dynamical Properties, and Phase Transitions

Nonlinear self-excited Hawkes processes display rich dynamical regimes determined by growth rates of $\phi$ and the integrated kernel $m_1 = \|h\|_1$:

• Sublinear: $\lim_{z\to\infty} \phi(z)/z = 0$. The process is stationary and exhibits standard LLN, CLT, and LDP.
• Subcritical linear: $\phi(z) \sim z$ for large $z$, $m_1 < 1$. The process is stationary. Mean event rate is $\mu = \phi(0)/(1-m_1)$.
• Critical: $\phi(z) \sim z$, $m_1 = 1$. Stationarity fails, with quantities such as $N_t$ growing polynomially.
• Supercritical: $\phi(z) \sim z, m_1 > 1$ or $\phi(z)/z \to \infty$. Exponential or explosive growth, no nontrivial stationary law.
• Explosive: Sufficiently fast-growing $\phi$ (e.g., superlinear) cause clustering to coalesce events within finite time (Zhu, 2013, Robert et al., 2022).

In multivariate and high-dimensional cases, excitation and inhibition create further complexity but analogous spectral-radius conditions ensure stationarity (Chen et al., 2017, Sulem et al., 2021).

3. Limit Theorems: Gaussian Fluctuations and Large Deviations

Under regularity and stability, the stationary nonlinear Hawkes process exhibits principled asymptotic behavior:

• Functional Central Limit Theorem (FCLT): For large observation windows $T$,

$$X^T_t = \frac{N_{Tt}-\mu T t}{\sqrt{T}},\qquad 0\le t\le1$$

converges weakly in Skorokhod topology to a Gaussian process with variance

$$\sigma^2 = \operatorname{Var}(N[0,1]) + 2\sum_{j=1}^\infty \operatorname{Cov}(N[0,1], N[j,j+1]).$$

In the linear case, this reduces to $\sigma^2 = \nu/(1-\|h\|_1)^3$ (Zhu, 2012).

• Law of the Iterated Logarithm (Strassen's Principle): Sample paths exhibit oscillations bounded by

$$\limsup_{T\to\infty} \frac{N_{Tt}-\mu\,Tt}{\sqrt{2T\log\log T}} = \sigma \sqrt{t}\quad \text{in } C([0,1])$$

(Zhu, 2012).

• Large Deviation Principles (LDP): Under broad conditions (sublinear or subcritical regimes), the empirical event rate $N_t / t$ satisfies an LDP at speed $t$ with rate function obtained via a variational/Legendre transform of the log moment generating function (Zhu, 2011, Zhu, 2013, Gao et al., 2017). In Markovian cases (kernel as sum of exponentials), explicit variational formulas are available.
• Moderate Deviations: Scaling between CLT and LDP, moderate deviations are governed by explicit quadratic rate functions (Gao et al., 2017).
• Non-Markovian Field and Master Equation Approaches: Infinite-dimensional Markov embeddings yield closed master equations and functional Hamilton–Jacobi PDEs, allowing analysis of multifractality and tail behaviors (Kanazawa et al., 2020, Kanazawa et al., 2021).

4. Power-Law Distributions and Nonlinear Mechanisms

A salient property of broad classes of nonlinear Hawkes processes is the emergence of power-law (heavy-tailed) intensity and event count distributions in stationarity.

If the intensity map $g(T)$ is fast-accelerating (e.g., $g(T) = \lambda_0 \exp(\beta T)$ or $T^n$ with $n > 2$), and marks are two-sided with nonpositive mean and sufficiently fast decreasing tails, the stationary intensity distribution satisfies: $$P_{\text{st}}(\lambda) \sim \lambda^{-2} \quad \text{(Zipf's law, mean-zero marks)}.$$ More generally, $P_{\text{st}}(\lambda) \sim \lambda^{-1-\mu}$, with the exponent depending on distributional parameters (Kanazawa et al., 2021).

This mechanism fundamentally differs from branching-process-driven power laws, arising instead from the nonlinear feedback plus heavy-tailed, sign-ambiguous marks. The same scaling appears in bursty phenomena in seismology, finance, and critical networks.

5. Inference: Bayesian, Frequentist, and Nonparametric Approaches

Inference in nonlinear Hawkes models is challenged by the loss of linear structure and the lack of branching representations. Recent methods include:

• Bayesian Nonparametric Estimation: Hierarchical priors on baseline rates and kernels (e.g., splines, Gaussian Processes), with MCMC or variational inference. Posterior contraction rates are established under mild entropy and prior-mass conditions. Granger-causality (network) estimation is consistent (Sulem et al., 2021, Malem-Shinitski et al., 2021).
• Graph Recovery: Edges in the interaction graph are estimated via spike-and-slab or variable selection priors, with guaranteed consistency and control of false discovery rate (Sulem et al., 2021).
• Neural and GP Surrogates: Feed-forward neural networks (NNNH) or nonparametric GP surrogates model kernels and intensity maps directly from data, enabling flexible capture of both excitation and inhibition (Joseph et al., 2023, Malem-Shinitski et al., 2021).
• Likelihood Optimization: For processes with Markovian structure (e.g., exponential kernels), the log-likelihood and gradients have efficient recursions, facilitating scalable MLE or variational methods (Quayle et al., 30 Jul 2025, Zhou et al., 2021).
• State-Augmentation and Thinning Algorithms: For simulation and inference, thinning (Ogata) schemes are adapted to nonlinear cases, sometimes with path-dependent or latent state augmentations (Clark et al., 28 May 2025, Zhou et al., 2021).

6. Applications, Empirical Features, and Extensions

Nonlinear self-excited Hawkes processes have been successfully deployed in diverse scientific domains:

• Neuroscience: Modeling spike trains with excitation and inhibition, revealing mechanisms of synchrony and variable-length memory corresponding to neuronal refractoriness (Quayle et al., 30 Jul 2025, Chen et al., 2017).
• Finance: Interpreting clustering of trades, volatility bursts, and power-law returns via nonlinear mechanism and fat-tailed intensity statistics (Kanazawa et al., 2021, Kanazawa et al., 2020).
• Seismology: ETAS-type models use broad nonlinear Hawkes frameworks to model aftershock distributions and multifractality (Kanazawa et al., 2020, Kanazawa et al., 2021).
• Network Science: Path-dependent nonlinear marked Hawkes models describe time-varying, feedback-driven network growth, including social contact networks with higher-order influence mechanisms (Clark et al., 28 May 2025).
• High-Dimensional Data: Scalable inference (e.g., neural/deep models, sparsity-aware Bayesian schemes) supports applications to tens or hundreds of interacting streams (Joseph et al., 2023, Sulem et al., 2021, Chen et al., 2017).

Extensive simulation and empirical studies support the empirical validity and flexibility of nonlinear Hawkes processes in these domains, with parameter recovery and causal graph estimation evaluated via held-out likelihood, cross-validation, and hypothesis testing (Quayle et al., 30 Jul 2025, Joseph et al., 2023, Sulem et al., 2021).

7. Open Problems and Extensions

Key open directions and recent advances include:

• Relaxation of stability and regularity conditions, notably via the Palm space and Markov-chain perspectives, admitting more general nonlinearities (Robert et al., 2022).
• Generalization to variable memory, network-dependent memories, and marked processes (Quayle et al., 30 Jul 2025, Clark et al., 28 May 2025).
• Quantitative understanding of explosive regimes and scaling laws near criticality, including transcritical bifurcations in the critical manifold (Robert et al., 2022, Kanazawa et al., 2020).
• Full characterization of weak dependence, concentration inequalities, and non-asymptotic error in high-dimensional settings (Chen et al., 2017).
• Unified field-theoretic treatments allowing extensions to spatially distributed, multifractal, or path-dependent nonlinear settings (Kanazawa et al., 2020).

These advances continue to expand the theoretical reach and practical impact of nonlinear self-excited Hawkes processes in modern applied probability and stochastic modeling.