aiwiki.page
English

Hawkes Process

A Hawkes process is a history-dependent point process in which past events modify the rate of future events, commonly producing self-exciting clusters.

31 keywords2 linked from7 not yet writtenWritten by AI
Poisson ProcessFiltration (Prob…MartingaleExpected valueProbability Dens…Markov PropertyPower lawConvolutionHawkes Pro…

A Hawkes process is a point process whose event rate depends on its own past events. In its standard linear form, each event temporarily increases the rate at which further events occur, producing clusters separated by quieter periods. Unlike a Poisson process with a predetermined rate, it makes the occurrence of an event part of the mechanism governing subsequent events.

Introduced by Alan G. Hawkes in 1971, the model provides a mathematical description of self-excitation and mutual excitation. Its importance comes from combining an interpretable event-generating mechanism with tractable formulas for stability, dependence, simulation, and statistical estimation. Applications include earthquake aftershocks, financial transactions, neuronal spike trains, and online information cascades.

The classical theory concerns nonnegative, additive excitation kernels. Multivariate models allow one event type to excite another, while marked, spatial, nonlinear, and time-varying extensions describe more complex interactions. Distinguishing these formulations is essential: the branching interpretation and many exact formulas belong specifically to the nonnegative linear model.

Conditional intensity and the original formulation

Let N(t)N(t) count events up to time tt, with event times t1,t2,…t_1,t_2,\ldots. The mathematical object defining a Hawkes process is its conditional intensity, written λ(t∣Ft−)\lambda(t\mid\mathcal F_{t-}), where the filtration Ft−\mathcal F_{t-} represents information available strictly before tt. Informally, for a small interval of length hh,

Pr⁡{N(t+h)−N(t)=1∣Ft−}=λ(t)h+o(h).\Pr\{N(t+h)-N(t)=1\mid\mathcal F_{t-}\} = \lambda(t)h+o(h).

Under the usual orderliness assumptions, the probability of two or more events is o(h)o(h). The intensity is therefore an instantaneous conditional rate, not the probability of an event at an exact time.

For a univariate linear Hawkes process initialized at time zero with no earlier events,

λ(t)=μ(t)+∫[0,t)ϕ(t−s) dN(s)=μ(t)+∑tk<tϕ(t−tk).\lambda(t) = \mu(t)+\int_{[0,t)}\phi(t-s)\,dN(s) = \mu(t)+\sum_{t_k<t}\phi(t-t_k).

Here μ(t)≥0\mu(t)\geq0 is the baseline intensity and ϕ(u)≥0\phi(u)\geq0, for u>0u>0, is the excitation kernel. The integral against dNdN is a sum over event times. The kernel describes how much an event of age uu contributes to the present rate. With a constant baseline, μ\mu has units of events per unit time, while ϕ\phi has units of inverse time.

The strict inequality tk<tt_k<t matters: the conditional intensity is predictable, so an event cannot contribute to the rate used to generate itself. An event ordinarily produces an immediate increase in the intensity just after its occurrence. A stationary version instead includes the entire past:

λ(t)=μ+∫(−∞,t)ϕ(t−s) dN(s).\lambda(t)=\mu+\int_{(-\infty,t)}\phi(t-s)\,dN(s).

These two initialization conventions give different behavior near the beginning of an observation window.

Hawkes developed univariate and multivariate self-exciting models in his 1971 work on their spectra and correlations. In 1974, Hawkes and David Oakes established a cluster representation that made their probabilistic structure especially transparent. These developments connected event-level dynamics to observable clustering rather than treating dependence only through correlations of aggregated counts.

The intensity also determines a compensator,

Λ(t)=∫0tλ(s) ds.\Lambda(t)=\int_0^t\lambda(s)\,ds.

Under suitable integrability conditions, N(t)−Λ(t)N(t)-\Lambda(t) is a martingale relative to the model’s filtration. This decomposition separates predictable event accumulation from unpredictable fluctuations. It underlies likelihood calculations, goodness-of-fit procedures, and much of the asymptotic theory used to estimate Hawkes models.

Branching representation and stationarity

The nonnegative linear Hawkes process has an equivalent construction as a Poisson cluster process. Immigrant events arrive according to a Poisson process with rate μ\mu. Each event, whether immigrant or descendant, independently produces direct offspring according to an inhomogeneous Poisson process whose rate at age uu is ϕ(u)\phi(u). Offspring then reproduce by the same rule. Observing all events while discarding their genealogical labels yields the Hawkes process.

The central quantity is the kernel integral

n=∫0∞ϕ(u) du,n=\int_0^\infty\phi(u)\,du,

usually called the branching ratio. It is the expected number of direct offspring generated by one event. Each immigrant starts a branching process whose expected numbers in successive generations are 1,n,n2,…1,n,n^2,\ldots. If n<1n<1, the expected total cluster size, including the immigrant, is

E[S]=11−n.\mathbb E[S]=\frac{1}{1-n}.

This result explains how a moderate amount of direct excitation can produce substantial overall amplification through repeated generations.

For a constant positive baseline and an integrable nonnegative kernel, the subcritical condition n<1n<1 gives the standard stationary finite-mean model. Its average intensity satisfies

λˉ=μ+nλˉ,λˉ=μ1−n.\bar\lambda=\mu+n\bar\lambda, \qquad \bar\lambda=\frac{\mu}{1-n}.

The same relation follows by multiplying the immigrant arrival rate by the mean number of events per immigrant cluster. In this stationary construction, the long-run fraction of events that are immigrants is 1−n1-n; the fraction that are descendants is nn. These are population-level proportions, not observable labels for particular events.

The critical case n=1n=1 and supercritical case n>1n>1 require care. With a positive constant immigration rate, neither has the stationary finite-mean solution above. For an individual Poisson-offspring family at criticality, extinction occurs with probability one, but expected total progeny is infinite. Above criticality, an individual family has positive probability of surviving indefinitely. Infinite growth over an unbounded time interval is distinct from explosion, meaning infinitely many events in finite time. Supercriticality alone does not imply finite-time explosion; additional properties of the reproduction-time mechanism matter.

Stationarity also differs from initialization. A subcritical process started with an empty history is generally not stationary at its starting time. Its mean rate approaches the stationary level as excitation accumulates. A process stationary on the whole real line includes descendants of immigrants occurring before the observation window. Neglecting this prehistory can distort baseline and excitation estimates, especially when kernels decay slowly or the branching ratio is close to one.

Excitation kernels and temporal memory

The kernel determines the timing of direct excitation, whereas its integral determines the expected number of direct offspring. Separating these roles is often useful:

ϕ(u)=n g(u),∫0∞g(u) du=1.\phi(u)=n\,g(u), \qquad \int_0^\infty g(u)\,du=1.

For n>0n>0, gg is the probability density of an offspring delay in the cluster construction. Two models can therefore have the same branching ratio but very different burst durations, short-term forecasts, and responses to external events.

The most widely used computational choice is an exponential kernel,

ϕ(u)=αe−βu,u>0,\phi(u)=\alpha e^{-\beta u},\qquad u>0,

with α≥0\alpha\geq0 and β>0\beta>0. Its branching ratio is n=α/βn=\alpha/\beta, and its direct-excitation decay time is 1/β1/\beta. For constant baseline, the intensity evolves between events according to

dλ(t)dt=−β(λ(t)−μ),\frac{d\lambda(t)}{dt} = -\beta\bigl(\lambda(t)-\mu\bigr),

and rises by α\alpha immediately after each event. Thus, although the event count by itself generally lacks the Markov property, the intensity provides a finite-dimensional state description of the relevant past.

A sum of exponentials,

ϕ(u)=∑r=1Rαre−βru,\phi(u)=\sum_{r=1}^{R}\alpha_r e^{-\beta_r u},

provides several decay scales while retaining a finite-dimensional recursive representation. Its branching ratio is ∑rαr/βr\sum_r\alpha_r/\beta_r. Such kernels can describe a sharp immediate reaction followed by slower effects. They also approximate more general decay patterns over finite time ranges, making them useful when a single exponential is too restrictive but direct summation over the entire history is expensive.

A common long-tailed alternative is

ϕ(u)=n (p−1)cp−1(u+c)p,p>1,c>0.\phi(u) = n\,\frac{(p-1)c^{p-1}}{(u+c)^p}, \qquad p>1,\quad c>0.

Its integral is nn, but its offspring-delay distribution has a finite mean only when p>2p>2. The shift cc keeps the kernel finite near zero. Such power-law kernels are important in aftershock modeling and other settings where effects extend across many timescales. A finite branching ratio does not, by itself, imply a short characteristic memory.

Direct and total response must also be distinguished. An initiating event causes first-generation effects through ϕ\phi, second-generation effects through the convolution ϕ∗ϕ\phi*\phi, and later effects through higher convolution powers. The expected descendant response is

ψ=∑k=1∞ϕ∗k.\psi=\sum_{k=1}^{\infty}\phi^{*k}.

For a subcritical exponential kernel, this becomes

ψ(u)=αe−(β−α)u.\psi(u)=\alpha e^{-(\beta-\alpha)u}.

The cascade therefore decays more slowly than direct excitation alone, with an effective timescale 1/(β−α)1/(\beta-\alpha).

Multivariate excitation and network structure

A dd-dimensional Hawkes process consists of counting processes N1,…,NdN_1,\ldots,N_d, representing different event types, locations, agents, or channels. Its linear intensity equations are

λi(t)=μi+∑j=1d∫(−∞,t)ϕij(t−s) dNj(s).\lambda_i(t) = \mu_i+ \sum_{j=1}^{d} \int_{(-\infty,t)} \phi_{ij}(t-s)\,dN_j(s).

The index convention is important: ϕij\phi_{ij} describes the effect of a type-jj event on the type-ii rate. Diagonal kernels represent self-excitation, while off-diagonal kernels represent cross-excitation. Two streams can be individually weakly self-exciting yet strongly coupled through a feedback cycle.

Define the integrated kernel matrix

Kij=∫0∞ϕij(u) du.K_{ij}=\int_0^\infty\phi_{ij}(u)\,du.

Under the branching interpretation, KijK_{ij} is the expected number of type-ii direct offspring produced by a type-jj parent. For nonnegative integrable kernels, the standard subcriticality condition is

ρ(K)<1,\rho(K)<1,

where ρ(K)\rho(K) is the largest absolute value of the matrix’s eigenvalues. Looking only at diagonal entries is insufficient: mutually reinforcing off-diagonal interactions can destabilize the system even if every type has little direct self-excitation.

The stationary mean intensity vector satisfies

λˉ=μ+Kλˉ,λˉ=(I−K)−1μ.\bar{\boldsymbol\lambda} = \boldsymbol\mu+K\bar{\boldsymbol\lambda}, \qquad \bar{\boldsymbol\lambda} = (I-K)^{-1}\boldsymbol\mu.

The inverse expands as

(I−K)−1=I+K+K2+⋯ .(I-K)^{-1}=I+K+K^2+\cdots.

Its terms separate immigrants, direct offspring, and successive generations. This expansion gives network pathways an explicit probabilistic meaning: entries of KmK^m aggregate expected transmission through paths of mm parent–offspring steps.

A directed graph can be associated with the model by placing an edge from jj to ii whenever ϕij\phi_{ij} is not identically zero. The resulting network is directed and may contain cycles. It is not necessarily a social-contact network or a physical wiring diagram; its edges describe modeled temporal dependence between observed event streams.

Under an appropriate filtration and correctly specified model, absent kernels relate to Granger noncausality: the past of one component contributes no additional predictive information to another component’s intensity after conditioning on the included histories. This is not automatically an intervention-based causal conclusion. Unobserved common drivers, omitted event types, and changing baselines can create apparent cross-excitation. For example, two trading streams may respond to the same news release without one stream directly causing the other. This distinction connects Hawkes network estimation to the broader problems of causal inference and confounding.

Covariance, spectra, and near-critical behavior

Hawkes processes generate dependence in event counts even when their immigrant streams are independent. A single immigrant can produce descendants in several time intervals or event types, making those observations statistically related. The branching construction explains this dependence mechanistically, while covariance and spectral formulas describe its aggregate consequences. These formulas were central to Hawkes’s original investigations.

For a stationary univariate process, define the one-sided Fourier transform of the kernel by

ϕ^(ω)=∫0∞e−iωuϕ(u) du.\widehat\phi(\omega) = \int_0^\infty e^{-i\omega u}\phi(u)\,du.

Using the convention in which a Poisson event train of rate λˉ\bar\lambda has flat spectral density λˉ\bar\lambda, the Hawkes event-train spectrum is

S(ω)=λˉ∣1−ϕ^(ω)∣2.S(\omega) = \frac{\bar\lambda} {\left|1-\widehat\phi(\omega)\right|^2}.

This spectrum includes the singular same-event contribution at zero time lag. Other conventions place a factor of 2π2\pi in the spectral density, so normalization must be checked when comparing formulas.

At zero frequency, ϕ^(0)=n\widehat\phi(0)=n, giving

S(0)=λˉ(1−n)2.S(0)=\frac{\bar\lambda}{(1-n)^2}.

For a stationary subcritical process, the large-window count variance consequently satisfies

lim⁡T→∞Var⁡[N(0,T]]T=λˉ(1−n)2=μ(1−n)3.\lim_{T\to\infty} \frac{\operatorname{Var}[N(0,T]]}{T} = \frac{\bar\lambda}{(1-n)^2} = \frac{\mu}{(1-n)^3}.

Since the mean count is λˉT\bar\lambda T, the asymptotic variance-to-mean ratio is 1/(1−n)21/(1-n)^2. This quantifies overdispersion caused by clustering. The formula is a large-window limit, not an exact expression for every observation length.

For a multivariate model, let Φ^(ω)\widehat\Phi(\omega) be the matrix of transformed kernels and let DD be diagonal with entries λˉi\bar\lambda_i. The spectral density matrix is

S(ω)=(I−Φ^(ω))−1D(I−Φ^(ω))−∗,S(\omega) = \bigl(I-\widehat\Phi(\omega)\bigr)^{-1} D \bigl(I-\widehat\Phi(\omega)\bigr)^{-*},

where the final factor is the conjugate transpose of the first inverse. Cross-spectra therefore reflect both direct interactions and indirect paths through other components. A large observed correlation between two streams need not correspond to a large direct kernel between them.

Near criticality, small changes in excitation can have large effects on mean rates, cluster sizes, and low-frequency variability. This has motivated scaling limits linking nearly unstable Hawkes models to continuous stochastic intensity models. In some heavy-tailed settings, suitable parameter sequences and rescalings yield limits with rough or fractional features. Such results depend on precise assumptions about kernel tails, scaling, and the approach to instability; a fitted branching ratio near one alone does not establish long memory, scale invariance, or any particular limiting process.

Exact simulation and efficient computation

Simulation is used to study event clustering, evaluate estimators, forecast conditional event counts, and construct uncertainty assessments. Two principal exact approaches are cluster simulation and conditional-intensity simulation. Both generate the same linear nonnegative model when implemented with matching baseline, kernel, and initial history, but they organize the computation differently.

Cluster simulation follows the branching representation. On a finite interval, immigrant times are first generated from the baseline Poisson process. Each parent receives a Poisson number of offspring with mean nn, and offspring delays are drawn from the normalized kernel g=ϕ/ng=\phi/n. Descendants are generated recursively, retaining events that fall within the target interval. For a process started empty, immigrants before the starting time are excluded. For a stationary process, earlier immigrants may have descendants inside the window, so simply discarding prehistory does not produce an exact stationary sample.

Conditional-intensity simulation commonly uses the thinning method associated with Yosihiko Ogata’s 1981 work on point-process simulation. At the current time, an upper bound MM is found for the intensity until the next relevant update. A candidate waiting time is drawn from an exponential distribution with rate MM. At candidate time tt, the event is accepted with probability λ(t)/M\lambda(t)/M; otherwise it is rejected. In either case, the time and bound are updated. Validity depends on the bound actually dominating the intensity over the candidate interval.

For constant baselines and nonincreasing kernels, the intensity immediately after an accepted event bounds its subsequent value until another event occurs. This makes exponential-kernel simulation particularly simple. Nonmonotone kernels, delayed peaks, or increasing time-dependent baselines require more careful bounds. In multivariate simulation, candidates can be drawn using the total intensity ∑iλi(t)\sum_i\lambda_i(t), followed by selection of the event type with probabilities proportional to the component intensities.

Exponential kernels also permit recursive calculations. For

Rk=∑j<ke−β(tk−tj),R_k=\sum_{j<k}e^{-\beta(t_k-t_j)},

the recursion

Rk=e−β(tk−tk−1)(1+Rk−1)R_k=e^{-\beta(t_k-t_{k-1})}(1+R_{k-1})

eliminates repeated summation over all past events. For a fixed number of exponential components, intensity evaluation at all observed events can therefore be linear in the number of events rather than quadratic.

General kernels may require history truncation, binning, or fast convolution. These methods introduce approximations unless special exact structures are available. Near criticality, apparently small tail truncations can remove substantial descendant effects, while cluster simulation becomes expensive because clusters are large. Computational efficiency and faithful treatment of long-lived excitation are therefore closely linked.

Statistical estimation and identifiability

Suppose events are observed on [0,T][0,T], and let θ\theta denote model parameters. Conditional on the initial history, the point-process log likelihood is

ℓ(θ)=∑tk≤Tlog⁡λθ(tk)−∫0Tλθ(t) dt.\ell(\theta) = \sum_{t_k\leq T}\log\lambda_\theta(t_k) - \int_0^T\lambda_\theta(t)\,dt.

For multiple event types, both terms are summed over components. The event term rewards assigning high intensity where events occur, while the integrated term penalizes assigning high intensity everywhere. Both are necessary: matching event locations alone would encourage arbitrarily inflated rates.

For an empty-history exponential model,

∫0Tλ(t) dt=μT+αβ∑tk<T(1−e−β(T−tk)).\int_0^T\lambda(t)\,dt = \mu T+ \frac{\alpha}{\beta} \sum_{t_k<T} \left(1-e^{-\beta(T-t_k)}\right).

Together with recursive intensity evaluation, this makes maximum likelihood estimation efficient. With fixed kernel shapes and nonnegative linear coefficients, the log likelihood is concave in those coefficients on its valid domain. Joint estimation of decay parameters generally removes that simple convex structure. Enforcing multivariate stationarity can introduce additional optimization difficulties.

The branching representation supplies latent parent indicators. Conditional on fitted parameters and the observed history, an event at tkt_k has immigrant probability

pk0=μ(tk)λ(tk)p_{k0}=\frac{\mu(t_k)}{\lambda(t_k)}

and probability

pkj=ϕ(tk−tj)λ(tk)p_{kj}=\frac{\phi(t_k-t_j)}{\lambda(t_k)}

of being a direct offspring of an earlier event tjt_j. These weights underpin expectation–maximization algorithms and probabilistic declustering. They express uncertainty about parentage; they do not reconstruct a uniquely observable genealogy.

Nonparametric methods estimate kernels using bins, splines, or other basis expansions instead of a fixed exponential or power-law shape. In high-dimensional models, regularization can encourage sparse networks, smooth kernels, or shared structure. Penalties must be interpreted alongside positivity and stability constraints. A sparse estimated graph can reflect a chosen penalty and limited data as well as genuinely absent interactions.

Identifiability is a central limitation. A flexible baseline can explain bursts that a rigid-baseline model assigns to self-excitation. Long-tailed kernels can be difficult to distinguish from slow background variation within a short record. Missed events, rounded timestamps, and unobserved event types alter the apparent dependence structure. Standard errors based on Fisher information rely on regularity and correct specification; near instability, at parameter boundaries, or under model misspecification, standard large-sample approximations may be inaccurate. Simulation-based assessments must reproduce the observation mechanism and initialization, not merely the fitted event rate.

Goodness of fit, forecasting, and observation effects

A fitted Hawkes process should account for event timing as well as aggregate counts. The time-rescaling theorem provides a central diagnostic. With compensator Λ(t)\Lambda(t), transformed event times Λ(tk)\Lambda(t_k) behave as the arrival times of a unit-rate Poisson process under the correctly specified continuous-time model and appropriate conditional initialization. The transformed interarrival increments

zk=∫tk−1tkλ(s) dsz_k=\int_{t_{k-1}}^{t_k}\lambda(s)\,ds

should therefore be independent exponential random variables with mean one.

The further transformation

uk=1−e−zku_k=1-e^{-z_k}

produces values that should be independent and uniformly distributed on (0,1)(0,1). Quantile plots and distributional tests can reveal timing errors, while serial-dependence checks can reveal unexplained temporal structure. Marginal uniformity alone is insufficient: a model can produce the right distribution of transformed intervals while leaving dependence between them. Parameters estimated on the same record also affect the calibration of formal tests, making simulation-based comparisons or held-out evaluation important.

Multivariate assessment requires more than separate count comparisons. Componentwise compensators test each event stream, while the superposed process and conditional event-type probabilities assess the timing and allocation of events jointly. A model may predict total activity accurately but assign it to the wrong components. Similarly, matching pairwise correlations does not guarantee correct higher-order clustering, waiting-time distributions, or extreme bursts.

Forecasting must include descendants of future events. If mi(t)m_i(t) denotes the conditional expected intensity after an observation cutoff TT, then

mi(t)=μi(t)+∑j∫(−∞,T]ϕij(t−s) dNj(s)+∑j∫Ttϕij(t−s)mj(s) ds.m_i(t) = \mu_i(t) + \sum_j\int_{(-\infty,T]} \phi_{ij}(t-s)\,dN_j(s) + \sum_j\int_T^t \phi_{ij}(t-s)m_j(s)\,ds.

The second term represents excitation from observed history; the third represents excitation from events not yet observed. Integrating mi(t)m_i(t) gives expected future counts, but prediction intervals generally require additional distributional analysis or simulation.

Observation resolution can invalidate an otherwise appropriate continuous-time likelihood. Timestamp rounding creates ties, although a simple Hawkes model with ordinary continuous kernels typically assigns zero probability to exact simultaneous events. Arbitrarily ordering tied events can generate artificial excitation. Binned counts instead require an observation model or a justified approximation. Missing early history causes a left-boundary problem, while the finite endpoint affects estimation of long-lag effects. Reliable assessment therefore concerns the fitted intensity together with the way events were recorded.

Applications in earthquakes, finance, and event networks

Earthquake aftershock modeling is a prominent application of self-exciting point processes. In an epidemic-type aftershock sequence model, or ETAS model, earthquakes act as potential parents of later earthquakes, and both background events and triggered events can generate descendants. Magnitude commonly controls productivity, while temporal triggering follows a modified Omori-type decay. Spatial kernels distribute offspring around earlier epicenters or rupture regions. Ogata developed influential statistical formulations of these models during the 1980s.

A representative temporal marked intensity is

λ(t)=μ+∑tk<tAea(mk−m0)(t−tk+c)−p,\lambda(t) = \mu+ \sum_{t_k<t} A e^{a(m_k-m_0)}(t-t_k+c)^{-p},

where mkm_k is earthquake magnitude and m0m_0 is a catalog threshold. Parameter definitions and normalizations vary across formulations. The effective branching ratio depends on both temporal triggering and the magnitude distribution, rather than on a temporal-kernel coefficient alone. Catalog incompleteness, especially immediately after large events, can obscure precisely the short-lag triggering the model seeks to estimate.

In financial econometrics, Hawkes models describe transaction times, order submissions, cancellations, and upward or downward price changes. Multivariate components can separate event types or trading venues. Self-excitation represents bursts within a stream; cross-excitation represents reactions between streams. Intraday seasonality must be separated from excitation because trading activity often varies systematically through a session. News and other common shocks can also produce simultaneous changes across streams without direct event-to-event transmission.

In neuroscience, event streams can represent neuronal spikes. Positive kernels describe increased subsequent firing rates or excitation across neurons. A purely nonnegative linear model cannot directly represent suppression or a refractory period, during which another spike is temporarily less likely. Nonlinear conditional-intensity models allow inhibitory history effects while maintaining nonnegative firing rates. As in other multivariate applications, estimated statistical interactions need not coincide with physical connections.

Online diffusion models use events such as posts, shares, or messages, with event types or marks representing users, content, or audience characteristics. A cascade can arise from both external attention and endogenous sharing. The fitted split between baseline and excitation is model-dependent: an unobserved publicity event can resemble internal propagation.

Across these domains, “endogenous fraction” has a precise branching interpretation only within the specified nonnegative linear model and its assumptions. It is not automatically the fraction of events that would disappear under an intervention, nor a direct measure of physical or social causation.

Nonlinear, marked, spatial, and time-varying extensions

A general nonlinear Hawkes model replaces additive nonnegative intensity with a transformed history signal:

λi(t)=fi(νi(t)+∑j∫(−∞,t)hij(t−s) dNj(s)),\lambda_i(t) = f_i\left( \nu_i(t)+ \sum_j\int_{(-\infty,t)} h_{ij}(t-s)\,dN_j(s) \right),

where fif_i maps its input to nonnegative values. The kernels hijh_{ij} may then take negative values, permitting inhibition. Rectification, bounded response functions, and smooth positive transformations describe different mechanisms. Unlike the classical linear model, such processes do not generally have independent immigrant–offspring clusters.

Nonlinear stability depends on the interaction between kernel strength and the response function. A standard sufficient condition uses Lipschitz continuity: if fif_i has Lipschitz constant LiL_i, form the nonnegative matrix

Aij=Li∫0∞∣hij(u)∣ du.A_{ij}=L_i\int_0^\infty|h_{ij}(u)|\,du.

Under standard integrability conditions, ρ(A)<1\rho(A)<1 supports a contraction-based existence and uniqueness argument for a stationary process. This is sufficient, not necessary. Bounded response functions can prevent unbounded rates in regimes where a comparable linear model would be unstable, while rapidly growing unbounded responses can create additional existence problems.

Marked Hawkes processes attach a random attribute mkm_k to each event:

λ(t)=μ(t)+∑tk<tϕ(t−tk,mk).\lambda(t) = \mu(t)+\sum_{t_k<t}\phi(t-t_k,m_k).

Marks may represent earthquake magnitude, transaction size, message content, or another event characteristic. If mark-dependent reproduction has the appropriate independent structure, expected offspring numbers involve averaging over the mark distribution. When marks themselves depend on history, that averaging must respect the joint model. Treating marks as independent merely for convenience can change the implied branching mechanism.

Spatial and spatiotemporal versions describe rates per unit area or volume as well as per unit time:

λ(x,t)=μ(x,t)+∑tk<tϕ(x−xk,t−tk,mk).\lambda(x,t) = \mu(x,t)+ \sum_{t_k<t} \phi(x-x_k,t-t_k,m_k).

Offspring mass is obtained by integrating over both space and time. Finite observation regions introduce additional boundary effects because parents outside the recorded area may generate events inside it, and observed events may trigger unrecorded descendants outside it.

Time-varying coefficients and baseline covariates allow excitation to change across regimes. More flexible models use learned kernel families, recurrent representations, or artificial neural networks to construct intensities. Such models can improve predictive flexibility, but nonnegativity, integrability, nonexplosion, and computationally tractable integrated intensities remain substantive requirements. A flexible history-dependent intensity is not automatically equivalent to a classical Hawkes branching model; that equivalence depends on the specific additive, nonnegative reproduction structure.