Some limit theorems for locally stationary Hawkes processes

Thomas Deschatre 1 , Pierre Gruet 1 , Antoine Lotz 1,2,† · 2025

Read Some limit theorems for locally stationary Hawkes processes by Thomas Deschatre 1 , Pierre Gruet 1 , Antoine Lotz 1,2,† (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-17245.txt.

Introduction

Thomas Deschatre, Pierre Gruet, and Antoine Lotz prove limit theorems for a class of Hawkes processes whose birth rate shifts over time, in work from twenty twenty-five. Hawkes processes model events that trigger more events, like trades after trades or quakes after quakes. Locally stationary forms let the trigger strength drift with time.

The team proves a law of large numbers and a functional central limit theorem for multi type versions. Non standard integral forms in the model block classic proofs, so martingale tools are reshaped with fresh ideas. The results feed financial statistics directly. As a case, closed form prices show how liquidity limits distort prices. The work firms up long run math for time varying event systems.

Catalog record

CREATORThomas Deschatre 1 , Pierre Gruet 1 , Antoine Lotz 1,2,†
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE89 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-17245

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Some limit theorems for locally stationary Hawkes processes Thomas Deschatre 1 , Pierre Gruet 1 , Antoine Lotz 1,2,† Abstract . We prove a law of large numbers and functional central limit theorem for a class of multivariate Hawkes processes with time-dependent reproduction rate. We address the difficulties induced by the use of non-convolutive Volterra processes by recombining classical martingale methods introduced in Bacry et al. [ 3 ] with novel ideas proposed by Kwan et al. [ 19 ] . The asymptotic theory we obtain yields useful applications in financial statistics. As an illustration, we derive closed-form expressions for price distortions under liquidity constraints. Mathematics Subject Classification (2020) : 60F05, 60G55, 62M10, 62P05. Keywords : Hawkes processes. Locally stationary processes. Limit Theorems. 1. Introduction 1.1. Motivation and relation to other works Hawkes processes are a class of self-interacting point processes. Their applications span numerous fields ranging from neuroscience [ 24 ] [ 5 ] to financial econometrics [ 4 ] [ 13 ] . In all generality, a multivariate point process (\boldsymbol{N}_{t})=(N_{1,t},\ldots,N_{p,t}) records discrete event-times (t^{k}_{i}) occurring in continuous time along p components according to (1) N_{k,t}=\sum_{i=1}^{\infty}\mathbb{1}_{\{t^{k}_{i}\leq t\}}. The law of such process is characterised by its predictable intensity \lambda_{t}=(\lambda_{1,t},\cdots,\lambda_{p,t}) , informally understood as an instantaneous probability of event-time arrival (2) \mathbb{P}\big[N_{k,t}\textup{ jumps in }[t,t+\mathop{}\!\mathrm{d}t]\big\lvert\mathcal{F}_{t}\big]=\lambda_{k,t}\mathop{}\!\mathrm{d}t, where (\mathcal{F}_{t}) is a history of the process. The intensity of the Hawkes process typically takes the form \lambda_{k,t}=\Phi_{k}\Big[\mu_{k}+\sum_{l=1}^{p}\int_{0}^{t}\varphi_{kl}(t-s)\mathop{}\!\mathrm{d}N_{l,s}\Big], where the \Phi_{k}\colon\mathbb{R}\mapsto\mathbb{R}^{+} are activation functions , the \mu_{k}\in\mathbb{R} baseline intensities and the function \varphi=(\varphi_{kl})\colon[0;\infty)\mapsto\mathbb{R}^{p\times p} the kernel of the process. We refer to section 2 for a rigorous definition. When p=1 and \Phi_{1}(x)=x , the dynamic of event-times arrival is that of a population process with Poisson immigration at intensity \mu_{1} and Galton-Watson reproduction at rate \lVert\varphi\rVert_{L^{1}(\mathop{}\!\mathrm{d}s)}=\int_{0}^{\infty}\varphi(s)\mathop{}\!\mathrm{d}s (see Hawkes & Oakes [ 16 ] ). Under the condition \lVert\varphi\rVert_{L^{1}(\mathop{}\!\mathrm{d}s)}<1 , the process is known to reach stationarity as T\to\infty , yielding useful approximations and convergence guarantees in statistical inference procedure (see e.g Ogata [ 23 ] ). It proves limiting in practice however, as the distribution of the phenomenon at hand often fluctuates in time. Generalisations of the Hawkes process have thus arisen in an effort to accommodate time- dependent dynamics. The process is then most often coerced to live on a compact timeframe [0,T] and depends on the double index (t,T) . Chen & Hall [ 9 ] first considered an univariate linear Hawkes process with varying baseline t\in[0,T]\mapsto\mu(\frac{t}{T})\in[0,\infty) , allowing the rate of cluster formation to oscillate in time. Roueff et al. [ 26 ] [ 25 ] further generalised the model, allowing \varphi to depend on \frac{t}{T} as well – resulting in the so-called locally stationary Hawkes process. This paper explores asymptotic properties of a class of locally stationary Hawkes processes with intensities (3) \lambda^{T}_{k,t}=\mu_{k}\big(\frac{t}{T}\big)+\int_{0}^{t}g\big(\frac{t}{T}\big)\varphi_{kl}(t-s)\mathop{}\!\mathrm{d}N_{l,s},\hskip 5.69046ptt\in[0,T], where, for any k,l=1\ldots p , \mu_{k}\colon[0,1]\mapsto[0,\infty) , g\colon[0,1]\mapsto[0,\infty) and \varphi_{kl}\colon[0,\infty)\mapsto[0,\infty) . The asymptotic behaviour of the process is well-known in the standard case where g is constant. Functional limit theorems were first proved by Bacry et al. [ 3 ] for the multivariate linear Hawkes process with constant baseline (\mu_{k}) , and later extended by Deschatre & Gruet [ 13 ] to varying baselines. Their results take the form of a uniform convergence for the normalised process T^{-1}(N_{Tu})_{u\in[0,1]} and of the weak convergence of the rescaled process T^{-\frac{1}{2}}(N_{Tu}-\mathbb{E}[N_{Tu}])_{u\in[0,1]} towards a diffusive limit. Analogous results were proven under the mild finite first moment assumption \int_{0}^{\infty}t\lVert\varphi\rVert(t)\mathop{}\!\mathrm{d}t<\infty for non linear activation functions and a constant baseline, by Zhu [ 27 ] in the univariate strictly self-exciting case \varphi>0 and by Cattiaux et al. [ 8 ] in the more general multivariate and possibly inhibiting case. Strong mixing rates covering the varying baseline case were moreover obtained by Cheysson et al. [ 10 ] under a similar assumption. The asymptotic theory of the varying baseline case is completed by the three recent works of Kwan et al. [ 19 ] [ 21 ] [ 20 ] , whom derive general ergodic theorems for statistical inference motives, and introduce a general approximation strategy for non stationary point processes. Regarding varying kernels finally, a central limit Theorem was obtained by Fierro et al. [ 14 ] in the context of a cluster-like construction of (N_{t}) , wherein the kernel is replaced at each generation by a new function. As for locally stationary Hawkes processes, neither functional approximations nor ergodic Theorems are known, owing perhaps to the difficulty associated with treating non-convolutive Volterra processes. 1.2. Main contribution We aim to reconcile the model ( 3 ) with the general asymptotic theory for Hawkes processes. Our main results in section 3 are a functional law of large number (Theorem 1 ) and a central limit Theorem (Theorem 2 ). Theorem 1 states the uniform convergence \sup_{u\in[0,1]}\Big\lVert\frac{1}{T}N_{Tu}-\int_{0}^{u}(\boldsymbol{Id}-\boldsymbol{K}(x))^{-1}\mu(x)\mathop{}\!\mathrm{d}x\Big\rVert\xrightarrow[T\to\infty]{}0, where \boldsymbol{K}(x)=g(x)\int_{0}^{\infty}\varphi(s)\mathop{}\!\mathrm{d}s , and can be regarded as a generalisation of the law of large numbers of Bacry et al. [ 3 ] and of the ergodic Theorem of Kwan et al. [ 21 ] . Theorem 2 establishes the weak convergence in Skorokhod topology of the rescaled process T^{-\frac{1}{2}}\{N_{Tu}-\mathbb{E}\big[N_{Tu}]\} , u\in[0,1] , towards the diffusive limit (4) \int_{0}^{u}(\boldsymbol{Id}-\boldsymbol{K}(x))^{-1}\boldsymbol{\Sigma}^{\nicefrac{{1}}{{2}}}(x)\mathop{}\!\mathrm{d}W_{x}, where (W_{x}) is a standard Brownian motion, and \boldsymbol{\Sigma} maps x\in[0,1] to the diagonal matrix with coefficients (\boldsymbol{Id}-\boldsymbol{K}(x))^{-1}\mu(x) . This again extends the limit Theorems of Bacry et al. [ 3 ] , and further generalises their prior extension by Deschatre & Gruet [ 13 ] . Our framework however differs from these two prior approaches, in that the kernel \kappa(t,s)=g(\nicefrac{{t}}{{T}})\varphi(t-s) escapes the convolution-specific techniques on which they depend. We rely instead on the same broad principle as Kwan, Chen & Dunsmuir [ 21 ] . More specifically, while the technical basis for our proofs differs fundamentally from the coupling-like arguments in [ 21 ] , we find that their approximation strategy based on a grid refining at rate o(T) can be repurposed to produce asymptotic results for the class of non-convolutive Volterra resolvent associated with our kernel. We isolate in Proposition 3 this deterministic result, from which Theorem 1 deduces. Obtaining Theorem 2 is more involved. The key to the proof consists in remarking the sampling scheme in Theorem 3 of Bacry et al. [ 3 ] refines at the exact same speed as the approximation grid of the kcd approach. This analogy proves quite productive, as it enables a careful recombination of the two approaches from which Theorem 2 follows. 2. Locally stationary Hawkes processes 2.1. Definition. We work on a rich enough probability space (\Omega,\mathcal{F},\mathbb{P}) where a family of integer-valued random measures \boldsymbol{N}=(N_{1},\cdots,N_{p}) on the real Borel sets \mathcal{B}(\mathbb{R}) are defined by N_{k}(C)=\sum_{n=1}^{\infty}\mathbb{1}_{\{t_{n}^{k}\in C\}}, where (t^{1}_{n})_{n\in\mathbb{N}^{\star}},\ldots,(t^{p}_{n})_{n\in\mathbb{N}^{\star}} are random sequences with values in [0,\infty] , such that 0<t^{k}_{n}<t^{k}_{n+1} for every n\in\mathbb{N}^{\star} and k=1\ldots p . The multivariate point process \boldsymbol{N}_{t}=(N_{1,t},\cdots,N_{p,t}) with counting measures \boldsymbol{N} is then defined by N_{k,t}=N_{k}([0,t]) consistently with ( 1 ). We will also write \mathop{}\!\mathrm{d}N_{k,s} for the measure N_{k}(\mathop{}\!\mathrm{d}s) . The history of \boldsymbol{N}_{t} is the filtration (\mathcal{F}_{t}) generated by the process until t , that is \mathcal{F}_{t}=\sigma\{N(C)\hskip 2.84544pt\lvert\hskip 2.84544ptC\subset[0,t]\cap\mathcal{B}(\mathbb{R})\} . For any (\mathcal{F}_{t}) -progressively-measurable process \lambda_{t}=(\lambda_{1,t},\cdots,\lambda_{p,t}) , \boldsymbol{N}_{t} is said to admit (\lambda_{t}) as its intensity if, for any t>s>0 , k=1\ldots p , (5) \mathbb{E}\big[N_{k}((s,t])\big\lvert\mathcal{F}_{t}\big]=\mathbb{E}\Big[\int_{s}^{t}\lambda_{k,u}\mathop{}\!\mathrm{d}u\Big\lvert\mathcal{F}_{t}\Big]. Up to some extension

Read the complete text