Abstract
A univariate Hawkes process is a simple point process that is self-exciting and has a clustering effect. The intensity of this point process is given by the sum of a baseline intensity and another term that depends on the entire past history of the point process. Hawkes processes have wide applications in finance, neuroscience, social networks, criminology, seismology, and many other fields. In this paper, we prove a functional central limit theorem for stationary Hawkes processes in the asymptotic regime where the baseline intensity is large. The limit is a non-Markovian Gaussian process with dependent increments. We use the resulting approximation to study an infinite-server queue with high-volume Hawkes traffic. We show that the queue length process can be approximated by a Gaussian process, for which we compute explicitly the covariance function and the steady-state distribution. We also extend our results to multivariate stationary Hawkes processes and establish limit theorems for infinite-server queues with multivariate Hawkes traffic.
Similar content being viewed by others
Notes
Market data feeds are typically composed of event messages that provide, in real time, the status of the market such as asset prices, reports of completed trades, and order activities. While an industry white paper [41] suggests that the market data traffic clearly exhibits clustering, we are not aware of academic studies or publicly available data on market data feeds.
In Lemma 2 in [58], it was proved that \({\mathbb {E}}[(N_{i}(1))^{2}]<\infty \). By the stationarity of \(N_i\) and the Cauchy–Schwarz inequality, for every positive integer t, \({\mathbb {E}}[(N_{i}(t))^{2}]={\mathbb {E}}[(\sum _{j=1}^{t}N_{i}(j-1,j))^{2}] \le t\sum _{j=1}^{t}{\mathbb {E}}[(N_{i}(j-1,j))^{2}]=t^{2}{\mathbb {E}}[(N_{i}(1))^{2}]<\infty \).
The Burkholder–Davis–Gundy inequality reads that for a local martingale \(M_{t}\) starting at 0 at \(t=0\), and \(M_{t}^{*}:=\sup _{0\le s\le t}|M_{s}|\), we have \(c_{p}{\mathbb {E}}[\langle M\rangle _{t}^{p/2}] \le {\mathbb {E}}[(M_{t}^{*})^{p}] \le C_{p}{\mathbb {E}}[\langle M\rangle _{t}^{p/2}]\), for some constant \(c_{p}<C_{p}\) depending on \(p\ge 1\) only and where \(\langle M\rangle _{t}\) is the (predictable) quadratic variation of \(M_{t}\). As a corollary, we have \({\mathbb {E}}[|M_{t}|^{p}]\le C_{p}{\mathbb {E}}[\langle M\rangle _{t}^{p/2}]\). In our application, \(p=4\) and \(M_{t}=N_{1}[s,t]-\int _{s}^{t}\lambda _{1}(v)\mathrm{d}v\), \(t\ge s\), so that \(M_{t}\) is a martingale with \(M_{s}=0\) and predictable quadratic variation \(\int _{s}^{t}\lambda _{1}(v)\mathrm{d}v\).
References
Bacry, E., Delattre, S., Hoffmann, M., Muzy, J.F.: Scaling limits for Hawkes processes and application to financial statistics. Stoch. Process. Their Appl. 123, 2475–2499 (2013)
Bacry, E., Mastromatteo, I., Muzy, J.F.: Hawkes processes in finance. Mark. Microstruct. Liq. 01, 1550005 (2015)
Blanchet, J., Chen, X., Lam, H.: Two-parameter sample path large deviations for infinite-server queues. Stoch. Syst. 4(1), 206–249 (2014)
Blundell, C., Beck, J., Heller, K.A.: Modelling reciprocating relationships with Hawkes processes. In: Pereira, F., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) Advances in Neural Information Processing Systems, Lake Tahoe, NV, pp. 2600–2608 (2012)
Billingsley, P.: Convergence of Probability Measures, 2nd edn. Wiley, New York (1999)
Bordenave, C., Torrisi, G.L.: Large deviations of Poisson cluster processes. Stoch. Models 23, 593–625 (2007)
Bowsher, C.G.: Modelling security market events in continuous time: intensity based, multivariate point process models. J. Econom. 141(2), 876–912 (2007)
Brémaud, P., Massoulié, L.: Stability of nonlinear Hawkes processes. Ann. Probab. 24, 1563–1588 (1996)
Burton, R., Waymire, E.: The central limit problem for infinitely divisible random measures. In: Taqqu, M., Eberlein, E. (eds.) Dependence in Probability and Statistics. Birkhauser, Boston (1986)
Chevallier, J.: Mean-field limit of generalized Hawkes processes. Stoch. Process. Their Appl. 127(12), 3870–3912 (2017)
Cont, R., De Larrard, A.: Order book dynamics in liquid markets: limit theorems and diffusion approximations. Available at SSRN 1757861 (2012)
Crane, R., Sornette, D.: Robust dynamic classes revealed by measuring the response function of a social system. Proc. Natl. Acad. Sci. USA 105, 15649 (2008)
Da Fonseca, J., Zaatour, R.: Hawkes process: fast calibration, application to trade clustering, and diffusive limit. J. Futur. Mark. 34(6), 548–579 (2014)
Daley, D.J., Vere-Jones, D.: An Introduction to the Theory of Point Processes, vol. I and II, 2nd edn. Springer, New York (2003)
Delattre, S., Fournier, N.: Statistical inference versus mean field limit for Hawkes processes. Electron. J. Stat. 10(1), 1223–1295 (2016)
Delattre, S., Fournier, N., Hoffmann, M.: Hawkes processes on large networks. Ann. Appl. Probab. 26, 216–261 (2016)
Eick, S.G., Massey, W.A., Whitt, W.: The physics of the \(M_t/G/\infty \) queue. Oper. Res. 41(4), 731–742 (1993)
Errais, E., Giesecke, K., Goldberg, L.: Affine point processes and portfolio credit risk. SIAM J. Financ. Math. 1, 642–665 (2010)
Evans, S.N.: Association and random measures. Probab. Theory Relat. Fields 86, 1–19 (1990)
Fasen, V.: Modeling network traffic by a cluster Poisson input process with heavy and light-tailed file sizes. Queueing Syst. 66(4), 313–350 (2010)
Fay, G., Gonzalez-Arevalo, B., Mikosch, T., Samorodnitsky, G.: Modeling teletraffic arrivals by a Poisson cluster process. Queueing Syst. 54(2), 121–140 (2006)
Gao, X., Zhu, L.: Limit Theorems for Linear Markovian Hawkes Processes with Large Initial Intensity. arXiv:1512.02155 (2015)
Gao, X., Zhu, L.: Large deviations and applications for Markovian Hawkes processes with a large initial intensity. Bernoulli (2018, to appear). arXiv:1603.07222v2 [math.PR]
Glynn, P.W., Szechtman, R.: Rare-event simulation for infinite server queues. In: Proceedings of the 2002 Winter Simulation Conference, pp. 416–423 (2002)
Gusto, G., Schbath, S.: FADO: a statistical method to detect favored or avoided distances between occurrences of motifs using the Hawkes’ model. Stat. Appl. Genet. Mol. Biol. 4(1), 1–28 (2005)
Hahn, M.G.: Central limit theorems in D [0, 1]. Probab. Theory Relat. Fields 44(2), 89–101 (1978)
Hawkes, A.G.: Spectra of some self-exciting and mutually exciting point processes. Biometrika 58, 83–90 (1971)
Hawkes, A.G.: Point spectra of some mutually exciting point processes. J. R. Stat. Soc. Ser. B (Methodological) 33, 438–443 (1971)
Hawkes, A.G., Oakes, D.: A cluster process representation of a self-exciting process. J. Appl. Prob. 11, 493–503 (1974)
Hewlett, P.: Clustering of order arrivals, price impact and trade path optimisation. In: Workshop on Financial Modeling with Jump processes, Ecole Polytechnique, pp. 6–8 (2006)
Hohn, N., Veitch, D., Abry, P.: Cluster processes: a natural language for network traffic. IEEE Trans. Signal Process. 51(8), 2229–2244 (2003)
Iglehart, D.L.: Limiting diffusion approximations for the many server queue and the repairman problem. J. Appl. Probab. 2(2), 429–441 (1965)
Jacod, J., Shiryaev, A.N.: Limit Theorems for Stochastic Processes, vol. 288. Springer, Berlin (2013)
Jaisson, T., Rosenbaum, M.: Limit theorems for nearly unstable Hawkes processes. Ann. Appl. Probab. 25, 600–631 (2015)
Jaisson, T., Rosenbaum, M.: Rough fractional diffusions as scaling limits of nearly unstable heavy tailed Hawkes processes. Ann. Appl. Probab. 26, 2860–2882 (2016)
Johnson, D.H.: Point process models of single-neuron discharges. J. Comput. Neurosci. 3(4), 275–299 (1996)
Jovanović, S., Hertz, J., Rotter, S.: Cumulants of Hawkes point processes. Phys. Rev. E 91, 042802 (2015)
Karabash, D., Zhu, L.: Limit theorems for marked Hawkes processes with application to a risk model. Stoch. Models 31, 433–451 (2015)
Ko, Y.M., Pender, J.: Strong Approximations for Time Varying Infinite-Server Queues with Non-Renewal Arrival and Service Processes (2016, Preprint). https://people.orie.cornell.edu/jpender/MAP_MAP_INF.pdf
Krichagina, E.V., Puhalskii, A.A.: A heavy-traffic analysis of a closed queueing system with a \(GI/\infty \) service center. Queueing Syst. 25(1), 235–280 (1997)
Low Latency market data. Corvil white paper. www.cisco.com/c/dam/en_us/solutions/industries/docs/finance/corvil_Latency_mkt_data.pdf
Lu, H., Pang, G., Mandjes, M.: A functional central limit theorem for Markov additive arrival process and its applications to queueing systems. Queueing Syst. 84(3–4), 381–406 (2016)
Ogata, Y.: Statistical models for earthquake occurrences and residual analysis for point processes. J. Am. Stat. Assoc. 83(401), 9–27 (1988)
Ozaki, T.: Maximum likelihood estimation of Hawkes’ self-exciting point processes. Ann. Inst. Stat. Math. 31(1), 145–155 (1979)
Pang, G., Talreja, R., Whitt, W.: Martingale proofs of many-server heavy-traffic limits for Markovian queues. Probab. Surv. 4, 193–267 (2007)
Pang, G., Whitt, W.: Two-parameter heavy-traffic limits for infinite-server queues. Queueing Syst. 65(4), 325–364 (2010)
Pernice, V., Staude, B., Carndanobile, S., Rotter, S.: How structure determines correlations in neuronal networks. PLoS Comput. Biol. 85, 031916 (2012)
Reed, J., Talreja, R.: Distribution-valued heavy-traffic limits for the \(\mathit{{G/}GI}/\infty \mathit{} \) queue. Ann. Appl. Probab. 25(3), 1420–1474 (2015)
Reynaud-Bouret, P., Schbath, S.: Adaptive estimation for Hawkes processes; application to genome analysis. Ann. Stat. 38(5), 2781–2822 (2010)
Reynaud-Bouret, P., Rivoirard, V., Tuleau-Malot, C.: Inference of functional connectivity in neurosciences via Hawkes processes. In: 1st IEEE Global Conference on Signal and Information Processing (2013)
Revuz, D., Yor, M.: Continuous Martingales and Brownian Motion, 3rd edn. Springer, Berlin (1998)
Sokol, A., Hansen, N.R.: Exponential martingales and changes of measure for counting processes. Stoch. Anal. Appl. 33, 823–843 (2015)
Whitt, W.: Stochastic-Process Limits: An Introduction to Stochastic-Process Limits and Their Application to Queues. Springer, Berlin (2002)
Whitt, W.: The Infinite-Server Queueing Model: The Center of the Many-Server Queueing Universe (i.e., More Relevant Than It Might Seem). http://www.columbia.edu/~ww2040/8100S12/ISqueue021412.pdf (2012)
Zhang, X., Blanchet, J., Giesecke, K., Glynn, P.W.: Affine point processes: approximation and efficient simulation. Math. Oper. Res. 40, 797–819 (2015)
Zhu, L.: Nonlinear Hawkes Processes. Ph.D. thesis, New York University (2013)
Zhu, L.: Moderate deviations for Hawkes processes. Stat. Probab. Lett. 83, 885–890 (2013)
Zhu, L.: Central limit theorem for nonlinear Hawkes processes. J. Appl. Probab. 50, 760–771 (2013)
Zhu, L.: Limit theorems for a Cox–Ingersoll–Ross process with Hawkes jumps. J. Appl. Probab. 51, 699–712 (2014)
Zhu, L.: Large deviations for Markovian nonlinear Hawkes Processes. Ann. Appl. Probab. 25, 548–581 (2015)
Acknowledgements
We are grateful to two anonymous referees, and the Associate Editor for very careful readings of the manuscript, and helpful suggestions, that greatly improve the quality of the paper. We also thank Jim Dai for helpful comments and Junfei Huang for many useful discussions. Xuefeng Gao acknowledges support from Hong Kong RGC ECS Grant 24207015 and CUHK Direct Grants for Research with project codes 4055035 and 4055054. Lingjiong Zhu is grateful to the support from NSF Grant DMS-1613164.
Author information
Authors and Affiliations
Corresponding author
Appendices
Proofs of results in Section 3
This section collects the proofs of results in Sect. 3.
1.1 Proof of Theorem 2
Proof of Theorem 2
The proof relies on Hahn’s theorem (see Theorem 2 in [26] or Theorem 7.2.1. in [53]), and delicate estimates of moments of stationary Hawkes processes.
For the sake of simplicity, we first let \(\mu \) be a positive integer. By the immigration-birth representation, we can decompose \(N^{\mu }\) as the sum of \(\mu \) independent and identically distributed (i.i.d.) Hawkes processes \(N_i^1, i =1, 2,\ldots , \mu \), each distributed as a stationary Hawkes process with baseline intensity 1 (the superscript 1 in \(N_i^1\)) and the exciting function \(h(\cdot )\). For notational simplicity, we use \(N_i(\cdot )\) for \(N_i^1 (\cdot )\). Therefore, we have
Let \({\tilde{N}}_i ({t}):=N_i({t})-\frac{t}{1-\Vert h\Vert _{L^{1}}}\). Then, \({\tilde{N}}_i\) are i.i.d. random elements of \(D([0,\infty ),{\mathbb {R}})\) with \({\mathbb {E}}[{\tilde{N}}_i(t)]=0\) (see, for example, Eq. (9) in [27]) and \({\mathbb {E}}[({\tilde{N}}_i (t))^{2}]<\infty \) for any t (see, for example, Lemma 2 in [58]).Footnote 2
By Hahn’s theorem, since \({\tilde{N}}_{i}\) are i.i.d., as \(\mu \rightarrow \infty ,\) we have
weakly in \((D([0,\infty ),{\mathbb {R}}),J_{1})\), where G is a mean-zero almost surely continuous Gaussian process with the covariance function of \({\tilde{N}}_1\), provided that the following condition is satisfied: For every \(0<T<\infty \), there exist continuous non-decreasing real-valued functions g and f on [0, T] and numbers \(\alpha >1/2\) and \(\beta >1\) such that
and
for all \(0\le s\le t\le u\le T\) with \(u-s<1\).
Let us prove (A.1) and (A.2). For notational simplicity, we use \(N_1(a,b]\) to stand for \(N_1((a,b])\) (equivalently, \(N_1(b)-N_1(a)\)), which records the number of points of the process \(N_1\) in the interval (a, b]. We also use \(\lambda _1\) to denote the intensity process of the stationary Hawkes process \(N_1\) with baseline intensity 1. We now present a lemma which is the key to the proofs of (A.1) and (A.2). The proof is given at the end of this section.
Lemma 15
We have
As a result, for all \(0\le s\le t\le u\le T\) and \(u-s<1\), there are some constants \(c, C>0\) independent of s, t, u, such that
With Lemma 15, we are ready to prove (A.1) and (A.2). First, let us prove (A.1). It is clear from the definition of \({{\tilde{N}}}_1\) that
Using (A.4) in Lemma 15 and the fact that \(0 \le u-s<1\), we immediately obtain that (A.1) is satisfied with \(g(x)=\left( C + \frac{1}{(1- \Vert h\Vert _{L^{1}})^2} \right) x\) and \(\alpha =1.\)
Next, let us prove (A.2). Note that
Since \(0\le s\le t\le u\le T\) and \(u-s<1\), we can then infer from Lemma 15 and (A.6) that (A.2) is satisfied with \(f(x)=C'x\) for some positive constant \(C'\) (independent of u, s, t) and \(\beta =2\).
Now we have proved Theorem 2 by assuming \(\mu \) is a positive integer in our discussions. The same result holds when \(\mu \in (0, \infty )\) for \(\mu \rightarrow \infty \). Note that for \(\mu \in (0, \infty )\), by the immigration-birth representation, the process \(N^{\mu }(\cdot )\) can be decomposed as the sum of two independent stationary Hawkes processes \(N^{\lfloor \mu \rfloor }(\cdot )\) and \(N^{\mu -\lfloor \mu \rfloor }(\cdot )\), where the superscripts \(\lfloor \mu \rfloor , \mu -\lfloor \mu \rfloor \) represent the baseline intensities, respectively. Hence, to show the result holds for \(\mu \in (0, \infty )\) with \(\mu \rightarrow \infty \), it suffices to show that for any \(T>0\),
in probability as \(\mu \rightarrow \infty \). This can be easily verified since, for any \(\epsilon >0\), for sufficiently large \(\mu \), we have \(\sup _{0\le t\le T}\frac{1}{\sqrt{\mu }} \frac{(\mu -\lfloor \mu \rfloor )t}{1-\Vert h\Vert _{L^{1}}} \le \frac{1}{\sqrt{\mu }}\frac{T}{1-\Vert h\Vert _{L^{1}}}\le \frac{\epsilon }{2}\), and
as \(\mu \rightarrow \infty \). Here \(N^1\) denotes a stationary Hawkes process with a baseline intensity one and an exciting function h.
Finally, let us compute the covariance function of G, or equivalently (from Hahn’s Theorem), the covariance function of \({\tilde{N}}_1(\cdot )\). Since \({\tilde{N}}_1 (t)\) and \({\tilde{N}}_1 (s)\) have mean zero, we can compute that, for any \(t>s\),
It is clear that
In addition, we can verify that
Hence, we get
The proof is therefore complete. \(\square \)
Proof of Lemma 15
We first prove (A.3). Using the definition of the intensity \(\lambda _1\) and the simple inequality \(\left( \frac{x+y}{2}\right) ^4 \le \frac{x^4 + y^4}{2}\), we obtain that, for sufficiently small \(\delta >0\),
Note here that, under Assumption 1, we know \(h(\cdot )\) is locally bounded and Riemann integrable, hence, for sufficiently small \(\delta >0\),
Applying the Jensen’s inequality to (A.9), we get
where we have used the stationarity of \(N_1\), (A.10) and the fact that \({\mathbb {E}}[e^{\theta N_1[0,1]}]<\infty \) for sufficiently small \(\theta >0\); see, for example,
[58].
We next prove (A.4). We can directly compute that
for some positive constant C. Here, the second line follows from the martingale property, see Sect. 2; the third line follows from the stationarity of the intensity process and \({\mathbb {E}}[\lambda _{1}(0)]=\frac{1}{1-\Vert h\Vert _{L^{1}}}\); the fourth line follows from the Cauchy–Schwarz inequality, so that \(\left( \int _{s}^{u}\lambda _1(v)\mathrm{d}v\right) ^{2}\le \int _{s}^{u}1\mathrm{d}v\cdot \int _{s}^{u}(\lambda _{1}(v))^{2}\mathrm{d}v\); the fifth line is due to the fact that \(\lambda _1\) is a stationary process; and the last inequality is due to (A.3) and the fact that \(0<u-s<1\).
Finally, we prove (A.5). We can directly compute that
where we use the fact that \(N_1(t,u]-\int _{t}^{u}\lambda _1 (v)\mathrm{d}v\) is \({\mathcal {F}}_{t}\)-measurable and is a martingale with predictable quadratic variation \(\int _{t}^{u}\lambda _1 (v)\mathrm{d}v\), so that \(\left( N_1(t,u]-\int _{t}^{u}\lambda _1 (v)\mathrm{d}v\right) ^{2}-\int _{t}^{u}\lambda _1 (v)\mathrm{d}v\) is also a martingale (see Sect. 2), and the Cauchy–Schwarz inequality, so that \(\left( \int _{t}^{u}\lambda _1(v)\mathrm{d}v\right) ^{2}\le \int _{t}^{u}1\mathrm{d}v\cdot \int _{t}^{u}(\lambda _{1}(v))^{2}\mathrm{d}v\). From here, we can further estimate that
where we applied the Cauchy–Schwarz inequality to the first term in the last line in (A.13), so that \({\mathbb {E}}\left[ \int _{t}^{u}\lambda _1 (v)\mathrm{d}v\cdot (N_1 (s,t])^{2}\right] \le \left( {\mathbb {E}}\left[ \left( \int _{t}^{u}\lambda _1 (v) \mathrm{d}v\right) ^{2}\right] \right) ^{1/2} \left( {\mathbb {E}}\left[ (N_1(s,t])^{4}\right] \right) ^{1/2}\), and the simple inequality \(2(\lambda _1 (v))^{2}(N_1 (s,t])^{2}\le (\lambda _1 (v))^{4}+(N_1 (s,t])^{4}\) to the second term in the last line in (A.13). From here, we can continue that
where we used the Cauchy–Schwartz inequality \(\left( \int _{t}^{u}\lambda _1(v)\mathrm{d}v\right) ^{2}\le \int _{t}^{u}1\mathrm{d}v\cdot \int _{t}^{u}(\lambda _{1}(v))^{2}\mathrm{d}v\), and the stationarity of the Hawkes processes; see Sect. 2.
Hence, to bound \({\mathbb {E}}\left[ (N_1(t,u])^{2}(N_1(s,t])^{2}\right] \), we need to estimate \({\mathbb {E}}\left[ (N_1 (s,t])^{4}\right] \). We can compute that (explanations follow below)
The first inequality in (A.16) uses the inequality \((\frac{x+y}{2})^{4}\le \frac{x^{4}+y^{4}}{2}\). The second inequality in (A.16) uses the martingality of \(N_1 [s,t]-\int _{s}^{t}\lambda _1 (v) \mathrm{d}v\) with the predictable quadratic variation \(\int _{s}^{t}\lambda _1 (v) \mathrm{d}v\) and the Burkholder–Davis–Gundy inequality, where \({\bar{C}}>0\) is a constant from the Burkholder–Davis–Gundy inequality.Footnote 3 The third inequality in (A.16) uses Jensen’s inequality, so that
and
Finally, the last equality in (A.16) is due to stationarity of the intensity process; see Sect. 2.
Combining (A.15), (A.16) and (A.3), we deduce that (A.5) holds. The proof is thus complete. \(\square \)
1.2 Proof of Proposition 3
Proof of Proposition 3
We first prove Part (b), then prove Parts (a) and (c).
To prove Part (b), we recall that
Denote \(H(t):=\int _t^{\infty }h(s)\mathrm{d}s\). By integrating Eq. (A.17) on both sides, we get
Since \(\Vert h\Vert _{L^{1}}<1\), there exists some \(\epsilon >0\) such that \(\Vert h\Vert _{L^{1}}+\epsilon <1\). In addition, note that H(M) is decreasing in M to 0 as \(M\rightarrow \infty \). Hence, for sufficiently large M we have \(H(M)\le \epsilon \). This implies
Note that \(\text {Var}(N^{1}(t))<\infty \) (see, for example, Lemma 2 in [58]) and hence
Moreover, \(\int _{0}^{t_{2}}\phi (t_{2}-t_{1})\mathrm{d}t_{1}=\int _{0}^{t_{2}}\phi (t_{1})\mathrm{d}t_{1}\) is non-decreasing in \(t_{2}\), and \(\int _{0}^{t}\int _{0}^{t_{2}}\phi (t_{1})\mathrm{d}t_{1}\mathrm{d}t_{2}<\infty \) for every t. This implies that \(\int _{0}^{t_{2}}\phi (t_{1})\mathrm{d}t_{1}<\infty \) for every \(t_{2}\). Thus, \(\int _{0}^{M}\phi (v)\mathrm{d}v<\infty \). Hence, it follows that
To establish the second part of Part (b), we first note from the definition of K(t) in (3.1) that
The differentiability of \(K(\cdot )\) is due to the integrability of \(\phi \) as given in (A.19). Since \(\phi \) is nonnegative, it follows that \(K'(\cdot )\) is non-decreasing. Hence, \(K(\cdot )\) is convex. In addition, we note that \(\int _{0}^{t}\phi (t-t_{1})\mathrm{d}t_{1}=\int _{0}^{t}\phi (t_{1})\mathrm{d}t_{1} \le \Vert \phi \Vert _{L^{1}}< \infty \) for all t. Thus, \(K(\cdot )\) is Lipschitz continuous.
We next provide a proof of Part (a), which will be useful in proving Part (c). Write \(N^1\) for a stationary Hawkes process with baseline intensity \(\mu =1\). From the Bartlett spectrum for the stationary Hawkes process (see [27] or [14]), we know that
where \({\hat{\psi }}(\omega ) = \int _{{\mathbb {R}}} e^{i\omega t} \psi (t) \mathrm{d}t\). Note that
and
We can compute that
Therefore,
Notice that, for any t,
and \(\int _{{\mathbb {R}}}\frac{\sin ^{2}\left( \frac{1}{2}\omega \right) }{\left( \frac{1}{2}\omega \right) ^{2}}\mathrm{d}\omega =2\pi \). Thus, \(\frac{\sin ^{2}\left( \frac{1}{2}\omega \right) }{\left( \frac{1}{2}\omega \right) ^{2}} \frac{1}{|1-{\hat{h}}(\frac{\omega }{t})|^{2}} \le \frac{\sin ^{2}\left( \frac{1}{2}\omega \right) }{\left( \frac{1}{2}\omega \right) ^{2}} \frac{1}{(1-\Vert h\Vert _{L^{1}})^{2}}\), which is integrable. On the other hand, for every \(\omega \), \(\lim _{t\rightarrow \infty }{\hat{h}}(\frac{\omega }{t}) ={\hat{h}}(0)=\Vert h\Vert _{L^{1}}\). Therefore, by the dominated convergence theorem, we obtain
We now prove Part (c). This requires a more delicate analysis of the Bartlett spectrum. Write \({\overline{z}}\) for the complex conjugate of a complex number z. We can obtain from (A.20) and \({\hat{h}}(0)=\Vert h\Vert _{L^{1}}\) that
where
We claim that \(f(\omega )\in L^{1}({\mathbb {R}})\), i.e., f is integrable on the real line. To see this, notice first that
and thus \(\int _{|\omega |\ge \epsilon }|f(\omega )|\mathrm{d}\omega <\infty \) for any \(\epsilon >0\). Moreover, by L’Hôpital’s rule, we can check that
Since
we apply the L’Hôpital’s rule again and get
Note that the first and second derivatives of \({\hat{h}}(\omega )\) and \(\overline{{\hat{h}}(\omega )}\) are well-defined due to the assumption \(\int _{0}^{\infty }t^{2}h(t)\mathrm{d}t<\infty \), and are given by
We deduce from the above that \(f(\omega )\in L^{1}({\mathbb {R}})\), and then the Riemann–Lebesgue theorem gives
which implies that the limit of the real part is also zero:
Hence, we conclude from (A.21) that
where the fact that this constant is negative follows from the observation that, for each \(\omega ,\)
Hence we complete the proof of Part (c). \(\square \)
1.3 Proof of Proposition 4
Proof of Proposition 4
We first prove that G has stationary increments. One can directly verify this fact by noting that, for \(\tau >0, s \ge 0,\) \(G(s+ \tau )- G(s)\) is a mean-zero Gaussian random variable, with variance given by
Using (3.1) and (3.3), it is easily checked that \(\text{ Var }(G(s +\tau )-G(s))= K(\tau )= \text{ Var }(G(\tau ))\), which is independent of s.
We next show that the Gaussian process G is not Markovian unless \(h\equiv 0\). To see this, recall (see, for example, Revuz and Yor [51, p.86]) that a centered Gaussian process \(\Upsilon \) with covariance function \(\Gamma (s,t):={\mathbb {E}}[\Upsilon _s \Upsilon _t]\) is Markovian if and only if
for every \(0 \le s< t < u\). Given the covariance function of G in (3.3), one can directly check that (A.22) does not hold for any nonzero exciting function h.
Finally, we prove that the paths of G are Hölder continuous of order \(\gamma \) for every \(\gamma <\frac{1}{2}\). To see this, note that \(G(s+ \tau )- G(s)\) is a mean-zero Gaussian random variable with variance \(K(\tau )\), which implies that, for \(p>0\),
where \(C={\mathbb {E}}|Z|^p\) and Z follows a standard normal distribution. By the Lipschitz property of K in Proposition 3, we infer from the Kolmogorov–Chentsov theorem that the sample paths of G are Hölder continuous with order less than \(\frac{1}{2}\). \(\square \)
Proof of results in Section 5
1.1 Proof of Theorem 12
Proof of Theorem 12
For notational simplicity, we write, for each t,
to stand for a k-dimensional stationary Hawkes processes where the intensity is given in (5.1) with \(\mu =1\). It should be self-evident that here the notation \(N^{i}\) stands for the i-th process (in Appendix A we used this notation to represent a univariate Hawkes process with baseline intensity i). Let us define
and let \(\tilde{{\mathbb {N}}}_{j}, j =1, 2, \ldots ,\) be independent copies of \(\tilde{{\mathbb {N}}}\), where \({\mathbb {E}} [\tilde{{\mathbb {N}}}(t)] =0\) for each t. We obtain from the immigration-birth representation of multivariate Hawkes processes that
where, as in the proof of Theorem 2, it suffices to establish the weak convergence of the sequence \(\left( \mathbb {{{\hat{N}}}}^{(\mu )} \right) \) for positive integer-valued \(\mu \).
We first establish the tightness of the sequence of processes \(\left( \mathbb {{{\hat{N}}}}^{(\mu )} \right) \). We use the tightness criteria in [33, Chapter VI. Theorem 4.1] and verify the three conditions there. Condition (i) trivially holds. To verify Condition (ii) and (iii), it suffices to check the following two conditions: for every \(0<T<\infty \), there exist some positive constants \(C_{1},C_{2}\) such that
and
for all \(0\le s\le t\le u\le T\) with \(u-s<1\), where the notation \(|| \cdot ||\) stands for the usual Euclidean norm of a vector in \({\mathbb {R}}^k. \) To see this, first notice that using the Markov inequality, it is straightforward to verify that (B.3) implies Condition (ii) in [33, Chapter VI. Theorem 4.1]. In addition, following the proof of Theorem 2 in [26] (the processes considered there are real-valued, but the argument in that proof also works for \({\mathbb {R}}^k\)-valued processes), one can immediately deduce that (B.4) implies Condition (iii) in [33, Chapter VI. Theorem 4.1].
We now prove (B.3) and (B.4). As the dimension k of the multivariate Hawkes process \({\mathbb {N}}\) is finite, in order to prove (B.3) and (B.4), it suffices to check that for every \(0<T<\infty \), there exist some positive constants \(C_{1},C_{2}\) such that, for all \(0\le s\le t\le u\le T\) with \(u-s<1\) and every \(1\le i,j\le k\),
and
We next prove (B.5) and (B.6). Similarly to (A.12), we can compute that
for some positive constant \(C_{1}\), provided that \({\mathbb {E}}[(\lambda ^{i}(0))^{2}]<\infty \).
Moreover, similarly to the derivations in (A.13), (A.14) and (A.15), we have
Similarly to (A.16) in the proof of Theorem 2, we have
Hence, we obtain
for some positive constant \(C_{2}\), provided that \({\mathbb {E}}[(\lambda ^{j}(0))^{4}]<\infty \) for \(1 \le j \le k\).
It remains to prove that \({\mathbb {E}}[(\lambda ^{j}(0))^{4}]<\infty \) for \(1 \le j \le k\), as this implies \({\mathbb {E}}[(\lambda ^{j}(0))^{2}]<\infty \). Similarly to the derivations in (A.9)–(A.11) in the proof of Theorem 2, since \(h_{ij}\) are locally bounded and Riemann integrable by Assumption 1, for sufficiently small \(\delta >0\) we obtain
for some positive constant C. Thus, it remains to show that for every \(1\le \ell \le k\), \({\mathbb {E}}\left[ \left( N^{\ell }(0,1]\right) ^{4}\right] <\infty \). It suffices to show that there exists some constant \(c_{\ell }>0\) so that \({\mathbb {E}}[e^{c_{\ell }N^{\ell }(0,1]}]<\infty \). Let us define \({\mathbb {E}}^{\emptyset }\) as the expectation under which the process \({\mathbb {N}}=(N^{1},\ldots ,N^{k})\) (with slight abuse of notation) is a multivariate Hawkes process starting from an empty history, that is, \(\lambda ^{i}(t)=p_{i}+\sum _{j=1}^{k}\int _{0}^{t-}h_{ij}(t-s)N^{j}(\mathrm{d}s)\). For any \(\psi _{i}>0\), \(e^{\sum _{i=1}^{k}\psi _{i}N^{i}(t)-\sum _{i=1}^{k}(e^{\psi _{i}}-1)\int _{0}^{t}\lambda ^{i}(s)\mathrm{d}s}\) is a martingale (see, for example, [52]), and thus
which implies that
Since the spectral radius of the matrix \({\mathbb {H}}= (\Vert h_{ij}\Vert _{L^{1}})_{1 \le i,j \le k}\) is strictly less than 1, we know that \(({\mathbb {I}}-{\mathbb {H}})^{-1}\) exists and \(({\mathbb {I}}-{\mathbb {H}})^{-1}=\sum _{n=0}^{\infty }{\mathbb {H}}^{n}\). Thus, for any fixed positive column vector \(m=(m_{1},\ldots ,m_{k})^{t}\in {\mathbb {R}}_{>0}^{k}\), we have \((({\mathbb {I}}-{\mathbb {H}})^{-1}m)_{i}>0\) for every i, where \((({\mathbb {I}}-{\mathbb {H}})^{-1}m)_{i}\) is the i-th component of the vector \(({\mathbb {I}}-{\mathbb {H}})^{-1}m\). Let \(\psi _{i}=\epsilon (({\mathbb {I}}-{\mathbb {H}})^{-1}m)_{i}\), where \(\epsilon >0\) is sufficiently small so that we can find some constant C(m) that depends on m such that
Hence, we deduce from (B.7) that there exists \(c_{i}=\epsilon m_{i}-C(m)\epsilon ^{2}>0\) such that
In particular, for every \(1\le \ell \le k\), we have
Since the linear Hawkes process (either with an empty history or the stationary version) is associated, we then deduce that, for positive integer t,
Hence, by the ergodicity of the Hawkes processes with an empty history where the exciting function satisfies Assumption 1, see, for example, [8], we obtain
where we recall that \({\mathbb {E}}\) is the expectation under which the Hawkes process is stationary. Hence, we have proved that there exists some constant \(c_{\ell }>0\) such that \({\mathbb {E}}[e^{c_{\ell }N^{\ell }(0,1]}]<\infty \) for each \(1 \le l \le k\).
Now we have proved the tightness of the sequence \(\left( \mathbb {{{\hat{N}}}}^{(\mu )} \right) \), we next show that the finite-dimensional distributions of the sequence of processes \(\left( \mathbb {{{\hat{N}}}}^{(\mu )} \right) \) converge in distribution to those of the limiting process \({\mathbb {G}}\) as \(\mu \rightarrow \infty \). To this end, we note that one can readily compute the covariance function of \(\tilde{{\mathbb {N}}}\) as in the univariate case [see Eqs. (A.7)–(A.8)], and find that, for \(t\ge s\),
Hence, in view of (B.2) and (B.3), the weak convergence of the finite-dimensional distributions of this sequence \(\left( \mathbb {{{\hat{N}}}}^{(\mu )} \right) \) immediately follows from the central limit theorem for sums of i.i.d. random vectors and the Cramér–Wold device (see, for example, Section 4.3.2 in [53]).
Finally, as the covariance of \({\mathbb {G}}\) is the same as that of \(\tilde{{\mathbb {N}}}\), we immediately infer from (B.3) that, for \(s<u,\)
which implies that the limiting Gaussian process \({\mathbb {G}}\) has continuous sample paths [51, p.37]. The proof is therefore complete. \(\square \)
1.2 Proof of Proposition 13
Proof of Proposition 13
To prove Proposition 13, we rely on [40]. For notational simplicity, we prove the joint weak convergence of \(({\mathbb {X}}_1^{\mu }, \ldots , {\mathbb {X}}_k^{\mu })\) as \(\mu \rightarrow \infty \) for the case \(k=2.\) The general case \(k \ge 2\) follows similarly.
First, as in the proof of Theorem 3 in [40], we obtain that, for \(i=1, 2\),
where, for \(t \ge 0,\)
Here \(\zeta _{ij}\) are all independent and uniformly distributed random variables on [0, 1] and the service times \(\eta _{ij}=F_i^{-1}(\zeta _{ij})\), where \(F_i^{-1}(x):= \inf \{y: F_i(y) \ge x \}\). For each fixed i, it was proved in [40] (see (6.1)–(6.3) there) that the following weak convergence of processes hold:
In addition, there is clearly a joint weak convergence of the left-hand sides of (B.8)–(B.10) to the right-hand sides [40]. Now, by the hypothesis, the number of customers in the system at time zero, \({\mathbb {Q}}^{\mu }_i(0)\) for \(i=1, 2\), as well as their respective service requirements \({{\bar{\eta }}}_{ij}\) for \(i=1, 2,\) are mutually independent. Moreover, the arrivals of new customers and the service requirements of those new customers are independent of the initial number of customers \({\mathbb {Q}}^{\mu }(0)\) and their service times. Hence, in order to prove the joint weak convergence of \(({\mathbb {X}}_1^{\mu }, {\mathbb {X}}_2^{\mu })\), it suffices to prove the weak convergence of \((M^{\mu }_{11}, M^{\mu }_{12}, M^{\mu }_{21}, M^{\mu }_{22})\) as \(\mu \rightarrow \infty .\)
To this end, let us define
Note that by Theorem 12, we have that the sequence of processes \(\left( \frac{{\mathbb {N}}^{(\mu )}}{\mu }\right) \) converges in distribution to a deterministic limit process \(\omega \), where \(\omega (t): =at\) for each \(t \ge 0.\) As \(\omega \) has continuous paths and the Skorohod \(J_1\) topology relativized to the space of continuous functions coincides with the uniform topology there [5, p.124], we obtain that, for each \(T>0\), as \(\mu \rightarrow \infty \),
Then, using a similar argument as in the proof of Lemma 5.3 in [40], we can establish that, for each \(T>0\) and \(\epsilon >0\),
In addition, using integration by parts, we can write \((M^{\mu }_{11}, M^{\mu }_{21}) = \left( g_1( \mathbb {{{\hat{N}}}}_1^{(\mu )} ), g_2 (\mathbb {{{\hat{N}}}}_2^{(\mu )} ) \right) \), where \( \mathbb {{{\hat{N}}}}_i^{(\mu )} (s) := \frac{{\mathbb {N}}_i^{(\mu )}(s)-{{\bar{\lambda }}}_i s}{\sqrt{\mu }}\) for each \(s \ge 0\), \(g_i: D([0,\infty ),{\mathbb {R}}) \rightarrow D([0,\infty ),{\mathbb {R}})\) is defined by
and \(g_i\) is continuous at points \(x(\cdot ) \in C([0,\infty ),{\mathbb {R}})\). See the proof of Lemma 3.3 in [40]. Now in Theorem 12 we have established that \(\left( \mathbb {{{\hat{N}}}}_1^{(\mu )}, \mathbb {{{\hat{N}}}}_2^{(\mu )} \right) \) converges in distribution to the Gaussian process \(({\mathbb {G}}_1, {\mathbb {G}}_2)\) under the Skorohod \(J_1\) topology, where the limiting Gaussian process has continuous paths; it then immediately follows that
where
Furthermore, as the service processes of each class i customer are independent, we deduce that the two processes \(U_1^\mu \) and \(U_2^\mu \) are independent for each \(\mu \), which further implies that \({{\tilde{M}}}^{\mu }_{12}\) and \({{\tilde{M}}}^{\mu }_{22}\) are two independent processes. By Lemma 3.1 of [40], we have \(U_i^\mu \Rightarrow U_i\) in \(D([0,\infty ),D[0,1])\) for each i as \(\mu \rightarrow \infty \). Hence, we deduce from Lemma 5.3 of [40] that
where
Then, we can obtain from (B.12), (B.13) and the independence of the service processes and the arrival processes of each class of customers that
Together with (B.11), which implies that \(\left( {M}^{\mu }_{12} - {\tilde{M}}^{\mu }_{12}, {M}^{\mu }_{22} - {\tilde{M}}^{\mu }_{22} \right) \Rightarrow (0, 0)\), we infer that the process \((M^{\mu }_{11}, M^{\mu }_{12}, M^{\mu }_{21}, M^{\mu }_{22})\) converges in distribution to \((M_{11}, M_{12}, M_{21}, M_{22})\) as \(\mu \rightarrow \infty \). Therefore, we obtain the weak convergence of \(({\mathbb {X}}_1^{\mu }, {\mathbb {X}}_2^{\mu })\) to the desired limit process \(({\mathbb {X}}_1, {\mathbb {X}}_2)\). The proof is completed. \(\square \)
Rights and permissions
About this article
Cite this article
Gao, X., Zhu, L. Functional central limit theorems for stationary Hawkes processes and application to infinite-server queues. Queueing Syst 90, 161–206 (2018). https://doi.org/10.1007/s11134-018-9570-5
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-018-9570-5