Abstract
We consider the strong convergence of the stochastic theta (ST) method for highly nonlinear hybrid stochastic differential equations with piecewise continuous arguments (SDEPCAs). There are three major ingredients. The first is the pth moment boundedness of the ST method. Second, the mean square convergence rate of the ST method for hybrid SDEPCAs is given by means of the forward–backward Euler–Maruyama method. The third ingredient is a numerical simulation, which shows the agreement with the theoretical convergence rate.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Stochastic differential equations with piecewise continuous arguments (SDEPCAs) play an important role in stochastic theory. Such models are applicable in a variety areas including biology, control science and neural networks (Li 2014; Mao et al. 2014; You et al. 2015; Xie and Zhang 2020). Component failures, changes in subsystem interconnections and sudden environmental disturbances can lead to abrupt changes of structures and parameters in many practical systems. To tackle these problems, hybrid systems driven by continuous-time Markov chains have become a powerful tool (Jobert and Rogers 2006; Smith 2002; Song and Mao 2018). Furthermore, it is well known that Markov chain can work as a stabilizing factor (Li and Mao 2020; Hu et al. 2020), that is, the whole system can be stable even though some subsystems are stable and others are unstable, such property is referred to as switching-dominated stability in Zhang et al. (2019).
Since explicit solutions are almost impossible to obtain for such systems, it becomes extremely important to solve them numerically. Finite time convergence analysis of an Euler type method for stochastic differential equations (SDEs) with Markovian switching was given in Mao et al. (2007) and Yuan and Mao (2004). It has been extended to stochastic differential delay equations (SDDEs) with Markovian switching (Li and Hou 2006; Milošević and Jovanović 2011; Zhang and Xie 2019), SDDEs with Markovian switching and Poisson jump (Li and Chang 2007; Wang and Xue 2007) and neutral SDDEs with Markovian switching (Yin and Ma 2011; Zhou and Wu 2009), etc. Numerical invariant measure of the backward Euler–Maruyama method for SDEs with Markovian switching was investigated in Li et al. (2018). It is worth mentioning that most of previous studies for hybrid systems is devoted to those equations driven by a continuous-time and homogeneous Markov chain independent of the Brownian motion, and the switching process r(t) was assumed to have a finite state space. Recently, Yin et al. extended the study to the Markov process (X(t), r(t)) by allowing the generator r(t) to depend on the current state X(t) and to have a countable state space (Yin and Zhu 2010; Nguyen and Yin 2016).
To our best knowledge, there are few works on SDEPCAs with Markovian switching. An SDEPCA belongs to the SDDEs, but the delay term is different from \(t-\tau \) and may be a discontinuous function. Moreover, although the SDEPCAs are retarded, the solutions of these equations are determined by only a finite set of initial data, rather than a function, as in the case of general SDDEs (Wiener 1993; Mao 2007). Because of these characteristics, we cannot simply generalize the properties of hybrid SDDEs to hybrid SDEPCAs.
In this work, we concentrate on the numerical solutions for highly nonlinear hybrid SDEPCAs, the stochastic theta (ST) scheme, which is an extension of the Euler–Maruyama method and the backward Euler–Maruyama, is adopted. The rest of this paper is organized as follows. Some basic notations and assumptions are introduced in Sect. 2. The ST method for hybrid SDEPCAs is established in Sect. 3. Section 4 is devoted to the pth moment boundedness of the ST method. Then we go further to reveal the strong convergence rate of the numerical method in Sect. 5. Finally, a numerical experiment is given in Sect. 6 to verify our theoretical convergence order.
2 Notations and preliminaries
Throughout this paper, unless otherwise specified, we let \((\Omega ,{\mathcal {F}},\left\{ {\mathcal {F}}_t\right\} _{t\ge 0},{\mathbb {P}})\) be a complete probability space with a filtration \(\left\{ {\mathcal {F}}_t\right\} _{t\ge 0}\) satisfying the usual conditions (i.e., it is right continuous and \({\mathcal {F}}_0\) contains all \({\mathbb {P}}\)-null sets). If A is a vector or matrix, its transpose is denoted by \(A^\mathrm{{T}}\). Let \(B(t)=(B_1(t),\dots ,B_d(t))^\mathrm{{T}}\) be a d-dimensional Brownian motion defined on the probability space. If x is a vector, \(\Vert x\Vert \) denotes its Euclidean norm. \(\langle x,y\rangle \) denotes the inner product of vectors x and y. If A is a matrix, its trace norm is denoted by \(\Vert A\Vert =\sqrt{{{\,\mathrm{trace}\,}}(A^\mathrm{{T}}A)}\). For two real numbers a and b, we will use \(a\vee b\) and \(a\wedge b\) for the \(\max \left\{ a,b\right\} \) and \(\min \left\{ a,b\right\} \), respectively. Let \({\mathcal {L}}^p_{{\mathcal {F}}_t}(\Omega ;{\mathbb {R}}^n)\) denotes the family of \({\mathcal {F}}_t\)-measurable \({\mathbb {R}}^n\)-valued random variables \(\xi \) with \({\mathbb {E}}\Vert \xi \Vert ^p<\infty \). Let \({\mathcal {L}}^p([a,b];{\mathbb {R}}^n)\) denotes the family of \({\mathbb {R}}^n\)-valued \({\mathcal {F}}_t\)-adapted processes \(\{f(t)\}_{a\le t\le b}\) such that \(\int _a^{b}\Vert f(t)\Vert ^p \mathrm{{d}}t<\infty \), a.s., \({\mathcal {L}}^p({\mathbb {R}}_+;{\mathbb {R}}^n)\) denotes the family of processes \(\{f(t)\}_{t\ge 0}\) such that for every \(T>0, \) \(\{f(t)\}_{0\le t\le T}\in {\mathcal {L}}^p([0,T];{\mathbb {R}}^n)\). \({\mathcal {M}}^p([a,b];{\mathbb {R}}^n)\) denotes the family of processes \(\{f(t)\}_{a\le t\le b}\) in \({\mathcal {L}}^p([a,b];{\mathbb {R}}^n)\) such that \({\mathbb {E}}\int _a^{b}\Vert f(t)\Vert ^p \mathrm{{d}}t<\infty \). \([\cdot ]\) denotes the greatest integer function.
Let \(r(t),t\ge 0,\) be a right-continuous Markov chain on the probability space taking values in a finite state space \(S=\left\{ 1,2,\dots ,N\right\} \) with generator \(\Gamma =(\gamma _{ij})_{N\times N}\) given by
where \(\Delta >0\). Here \(\gamma _{ij}\ge 0\) is the transition rate from i to j when \(i\ne j\), and
We assume that the Markov chain \(r(\cdot )\) is independent of the Brownian motion \(B(\cdot )\).
Consider the following hybrid SDEPCAs
with initial data \(x(0)=x_0\in {\mathbb {R}}^n\) and \(r(0)=i_0\in S,\) where
Let us give the definition of the solution.
Definition 1
An \({\mathbb {R}}^n\)-valued stochastic process \(\left\{ x(t)\right\} _{t\ge 0}\) is called a solution of (1) if it has the following properties:
-
(1)
\(\left\{ x(t)\right\} \) is continuous and \({\mathcal {F}}_t\)-adapted;
-
(2)
\(\left\{ f(x(t),x([t]),r(t))\right\} \in {\mathcal {L}}^1({\mathbb {R}}_+;{\mathbb {R}}^n)\), \(\left\{ g(x(t),x([t]),r(t))\right\} \in {\mathcal {L}}^2({\mathbb {R}}_+;{\mathbb {R}}^{n\times d})\);
-
(3)
Equation (1) is satisfied on each interval \([n,n+1)\subset [0,\infty )\) with integral end-points almost surely.
A solution \(\left\{ x(t)\right\} \) is said to be unique if any other solution \(\left\{ {\bar{x}}(t)\right\} \) is indistinguishable from \(\left\{ x(t)\right\} \), that is
We impose some assumptions:
Assumption 2.1
For every integer \(R\ge 1\), there exists a constant \(L(R)>0\) such that
for all \(i\in S\) and those \(x, y,{\bar{x}},{\bar{y}}\in {\mathbb {R}}^n\) with \(\Vert x\Vert \vee \Vert y\Vert \vee \Vert {\bar{x}}\Vert \vee \Vert {\bar{y}}\Vert \le R\).
Assumption 2.2
There exists a constant \(\alpha >0\) such that
Assumption 2.3
There exist constants \(L_1>0\) and \(h_1\ge 1\) such that
Assumption 2.4
There exist constants \(L_2>0\) and \(h_2\ge 1\) such that
Theorem 2.5
Under Assumptions 2.1–2.4, for any \(T>0\), there exists a unique global solution x(t) to Eq. (1) on \(t\in [0,T]\) with the initial data \(x_0\). Moreover, the solution has the property that \({\mathbb {E}}\Vert x(t)\Vert ^p<\infty \) for all \(t\in [0,T]\).
Proof
For any given \(i\in S\), we first prove that there exists a unique global solution to the SDEPCA
on \(t\in [0,T]\) with the initial data \(x_0\) and the solution has the property that \({\mathbb {E}}\Vert x(t)\Vert ^p<\infty \) for all \(t\in [0,T]\). To distinguish between the solution of (1) and that of (2), we denote the solution of (2) by y(t).
In a similar way as the proof of Theorem 3.15 in Mao and Yuan (2006), there is a unique maximal local solution y(t) exists on \([0,\eta _{e})\) under the local Lipschitz condition, where \(\eta _{e}\) is the explosion time. Then for each integer \(R\ge \Vert x_0\Vert \), define the stopping time \(\eta _R=\inf \{t\in [0,\eta _e):\Vert y(t)\Vert \ge R\}\). Clearly, \(\eta _R\) is increasing as \(R\rightarrow \infty \). We denote that \(\eta _{\infty }=\lim _{R\rightarrow \infty }\eta _R\) and \(\inf \emptyset =\infty \). Hence, \(\eta _{\infty }\le \eta _e\) almost surely. If we can obtain \(\eta _{\infty }=\infty \) almost surely, then \(\eta _e=\infty \) almost surely. In what follows, we will prove \(\eta _{\infty }=\infty \) almost surely and \({\mathbb {E}}\Vert y(t)\Vert ^p<\infty \).
Applying Itô’s formula to \(\Vert y(t)\Vert ^p, p\ge 2\), we have
-
Take any \(t\in [0,1)\), integrating both sides of (3) from 0 to \(t\wedge \eta _R\), then
$$\begin{aligned} \begin{aligned}&\Vert y(t\wedge \eta _R)\Vert ^p\\&\quad \le \Vert y(0)\Vert ^p+p\int _0^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}\left( y^\mathrm{{T}}(s)f(y(s),y(0),i)+\frac{p-1}{2}\left\| g(y(s),y(0),i)\right\| ^2\right) \mathrm{{d}}s\\&\qquad +p\int _0^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}y^\mathrm{{T}}(s)g(y(s),y(0),i)\mathrm{{d}}B(s). \end{aligned} \end{aligned}$$Hence,
$$\begin{aligned} \begin{aligned}&{\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\\&\quad \le \Vert x_0\Vert ^p+p{\mathbb {E}}\int _0^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}\left( y^\mathrm{{T}}(s)f(y(s),x_0,i)+\frac{p-1}{2}\left\| g(y(s),x_0,i)\right\| ^2\right) \mathrm{{d}}s. \end{aligned} \end{aligned}$$By Assumptions 2.2 and 2.4, together with Young’s inequality, one has
$$\begin{aligned} \begin{aligned} {\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\le&\Vert x_0\Vert ^p+p{\mathbb {E}}\int _0^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}\Bigg (\alpha (1+\Vert y(s)\Vert ^2+\Vert x_0\Vert ^2)\\&\qquad \qquad +\frac{(p-1)L_2^2}{2}(1+\Vert y(s)\Vert +\Vert x_0\Vert ^{h_2})^2\Bigg )\mathrm{{d}}s\\ \le&\Vert x_0\Vert ^p+p{\mathbb {E}}\int _0^{t\wedge \eta _R}\bigg \{\left( \gamma _0+\alpha \Vert x_0\Vert ^2+3(p-1)L_2^2/2\Vert x_0\Vert ^{2h_2}\right) \Vert y(s)\Vert ^{p-2}\\&\qquad \qquad \qquad \qquad +\gamma _0\Vert y(s)\Vert ^{p}\bigg \}\mathrm{{d}}s\\ \le&\Vert x_0\Vert ^p+p{\mathbb {E}}\int _0^{t\wedge \eta _R}\bigg \{\gamma _0^{\frac{p}{2}}+\alpha ^{\frac{p}{2}}\Vert x_0\Vert ^p+(3(p-1)L_2^2/2)^{\frac{p}{2}}\Vert x_0\Vert ^{ph_2}\\&\qquad \qquad \qquad \qquad +(3+\gamma _0)\Vert y(s)\Vert ^{p}\bigg \}\mathrm{{d}}s\\ \le&pt\gamma _0^{\frac{p}{2}}+(1+pt\alpha ^{\frac{p}{2}})\Vert x_0\Vert ^p+pt(3(p-1)/2)^{\frac{p}{2}}L_2^p\Vert x_0\Vert ^{ph_2}\\&\qquad \qquad \qquad \qquad +(3+\gamma _0)p\int _0^{t}{\mathbb {E}}\Vert y(s\wedge \eta _R)\Vert ^{p}\mathrm{{d}}s, \end{aligned} \end{aligned}$$where \(\gamma _0=\alpha +3(p-1)L_2^2/2\). Now it can be obtained from Gronwall’s inequality that
$$\begin{aligned} \begin{aligned} {\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\le&\left( pt\gamma _0^{\frac{p}{2}}+(1+pt\alpha ^{\frac{p}{2}})\Vert x_0\Vert ^p+pt(3(p-1)/2)^{\frac{p}{2}}L_2^p\Vert x_0\Vert ^{ph_2}\right) e^{(3+\gamma _0)pt}\\ \le&\beta _0e^{(3+\gamma _0)p}, \end{aligned} \end{aligned}$$(4)where
$$\begin{aligned} \beta _0=p\gamma _0^{\frac{p}{2}}+(1+p\alpha ^{\frac{p}{2}})\Vert x_0\Vert ^p+p(3(p-1)/2)^{\frac{p}{2}}L_2^p\Vert x_0\Vert ^{ph_2}<\infty . \end{aligned}$$Thus,
$$\begin{aligned} {\mathbb {E}}\Vert y(1\wedge \eta _R)\Vert ^p=\lim _{t\rightarrow 1}{\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\le \beta _0e^{(3+\gamma _0)p}. \end{aligned}$$(5)Let \(I_G\) denote the indicator function of the set G, then
$$\begin{aligned} \beta _0e^{(3+\gamma _0)p}\ge {\mathbb {E}}\Vert y(1\wedge \eta _R)\Vert ^p\ge {\mathbb {E}}\left( \Vert y(\eta _R)\Vert ^pI_{\{\eta _R\le 1\}}\right) \ge R^p{\mathbb {P}}(\eta _R\le 1); \end{aligned}$$hence,
$$\begin{aligned} {\mathbb {P}}(\eta _R\le 1)\le \frac{\beta _0e^{(3+\gamma _0)p}}{R^p}. \end{aligned}$$Let \(R\rightarrow \infty \), we have \({\mathbb {P}}(\eta _{\infty }\le 1)=0\), which gives
$$\begin{aligned} {\mathbb {P}}(\eta _{\infty }> 1)=1, \end{aligned}$$combining (4) and (5), it can be obtained that
$$\begin{aligned} {\mathbb {E}}\Vert y(t)\Vert ^p\le \beta _0e^{(3+\gamma _0)p},\quad t\in [0,1]. \end{aligned}$$ -
For any \(t\in [1,2)\), similar to the process above, integrating both sides of (3) from 1 to \(t\wedge \eta _R\), and then taking the expectations, we can arrive at
$$\begin{aligned} \begin{aligned}&{\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\\&\quad \le \!{\mathbb {E}}\Vert y(1)\Vert ^p\!+\!p{\mathbb {E}}\int _1^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}\left( y^\mathrm{{T}}(s)f(y(s),y(1),i)+\frac{p-1}{2}\left\| g(y(s),y(1),i)\right\| ^2\right) \mathrm{{d}}s\\&\quad \le {\mathbb {E}}\Vert y(1)\Vert ^p+p{\mathbb {E}}\int _1^{t\wedge \eta _R}\Vert y(s)\Vert ^{p-2}\Bigg (\alpha (1+\Vert y(s)\Vert ^2+\Vert y(1)\Vert ^2)\\&\qquad +\frac{(p-1)L_2^2}{2}(1+\Vert y(s)\Vert +\Vert y(1)\Vert ^{h_2})^2\Bigg )\mathrm{{d}}s\\&\quad \le {\mathbb {E}}\Vert y(1)\Vert ^p+p{\mathbb {E}}\int _1^{t\wedge \eta _R}\bigg \{\left( \gamma _0+\alpha \Vert y(1)\Vert ^2+3(p-1)L_2^2/2\Vert y(1)\Vert ^{2h_2}\right) \Vert y(s)\Vert ^{p-2}\\&\qquad +\gamma _0\Vert y(s)\Vert ^{p}\bigg \}\mathrm{{d}}s\\&\quad \le {\mathbb {E}}\Vert y(1)\Vert ^p+p{\mathbb {E}}\int _1^{t\wedge \eta _R}\bigg \{\gamma _0^{\frac{p}{2}}+\alpha ^{\frac{p}{2}}\Vert y(1)\Vert ^p+(3(p-1)L_2^2/2)^{\frac{p}{2}}\Vert y(1)\Vert ^{ph_2}\\&\qquad +(3+\gamma _0)\Vert y(s)\Vert ^{p}\bigg \}\mathrm{{d}}s\\&\quad \le p(t-1)\gamma _0^{\frac{p}{2}}+(1+p(t-1)\alpha ^{\frac{p}{2}}){\mathbb {E}}\Vert y(1)\Vert ^p+p(t-1)(3(p-1)/2)^{\frac{p}{2}}L_2^p{\mathbb {E}}\Vert y(1)\Vert ^{ph_2}\\&\qquad +(3+\gamma _0)p\int _1^{t}{\mathbb {E}}\Vert y(s\wedge \eta _R)\Vert ^{p}\mathrm{{d}}s. \end{aligned} \end{aligned}$$By Gronwall’s inequality, one has
$$\begin{aligned} \begin{aligned} {\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\le \beta _1e^{(3+\gamma _0)p}, \end{aligned} \end{aligned}$$(6)where
$$\begin{aligned} \beta _1=\left( p\gamma _0^{\frac{p}{2}}+(1+p\alpha ^{\frac{p}{2}}){\mathbb {E}}\Vert y(1)\Vert ^p+p(3(p-1)/2)^{\frac{p}{2}}L_2^p{\mathbb {E}}\Vert y(1)\Vert ^{ph_2}\right) <\infty . \end{aligned}$$Hence,
$$\begin{aligned} \begin{aligned} {\mathbb {E}}\Vert y(2\wedge \eta _R)\Vert ^p=\lim _{t\rightarrow 2}{\mathbb {E}}\Vert y(t\wedge \eta _R)\Vert ^p\le \beta _1e^{(3+\gamma _0)p}, \end{aligned} \end{aligned}$$(7)it gives
$$\begin{aligned} \beta _1e^{(3+\gamma _0)p}\ge {\mathbb {E}}\Vert y(2\wedge \eta _R)\Vert ^p\ge {\mathbb {E}}\left( \Vert y(\eta _R)\Vert ^pI_{\{\eta _R\le 2\}}\right) \ge R^p{\mathbb {P}}(\eta _R\le 2). \end{aligned}$$Taking \(R\rightarrow \infty \), yields
$$\begin{aligned} {\mathbb {P}}(\eta _{\infty }\le 2)\le \lim _{R\rightarrow \infty }\frac{\beta _1e^{(3+\gamma _0)p}}{R^p}=0, \end{aligned}$$which implies
$$\begin{aligned} {\mathbb {P}}(\eta _{\infty }> 2)=1, \end{aligned}$$then combining (6) and (7), it can be obtained that
$$\begin{aligned} {\mathbb {E}}\Vert y(t)\Vert ^p\le \beta _1e^{(3+\gamma _0)p},\quad t\in [1,2]. \end{aligned}$$ -
Repeating this procedure we can deduce that, for any integer \(j\ge 1\),
$$\begin{aligned} {\mathbb {P}}(\eta _{\infty }> j)=1, \end{aligned}$$and
$$\begin{aligned} {\mathbb {E}}\Vert y(t)\Vert ^p\le \beta _je^{(3+\gamma _0)p},\quad t\in [j,j+1], \end{aligned}$$(8)where
$$\begin{aligned} \beta _j=\left( p\gamma _0^{\frac{p}{2}}+(1+p\alpha ^{\frac{p}{2}}){\mathbb {E}}\Vert y(j)\Vert ^p+p(3(p-1)/2)^{\frac{p}{2}}L_2^p{\mathbb {E}}\Vert y(j)\Vert ^{ph_2}\right) <\infty . \end{aligned}$$
Since \(j\ge 1\) is an arbitrary integer, we can conclude that \(\eta _{\infty }=\infty \) almost surely and \({\mathbb {E}}\Vert y(t)\Vert ^p< \infty , \forall t\ge 0\).
Now we are in a position to prove Eq.(1) has a unique global solution x(t) and the solution has the property that \({\mathbb {E}}\Vert x(t)\Vert ^p<\infty \).
It is well known (see Anderson 2012) that almost every sample path of the Markov chain \(r(\cdot )\) is a right-continuous step function with a finite number of sample jumps in any finite subinterval of \({\mathbb {R}}_+\). Hence, there is a sequence of stopping times \(0=\tau _0<\tau _1<\cdots<\tau _k<\cdots \) such that
We first consider Eq. (1) on \(t\in [\tau _0,\tau _1)\), which becomes
with initial data \(x_0\). By the existence-and-uniqueness proof for SDEPCA (2), we know that Eq. (9) has a unique continuous solution which belongs to \({\mathcal {M}}^2([\tau _0,\tau _1);{\mathbb {R}}^n)\) and has the property that \({\mathbb {E}}\Vert x(t)\Vert ^p<\infty \). In particular, \(x(\tau _1)=\lim _{t\rightarrow \tau _1^-}x(t)\in L^2_{{\mathcal {F}}_{\tau _1}}(\Omega ;{\mathbb {R}}^n)\). We next consider Eq. (1) on \(t\in [\tau _1,\tau _2)\), which becomes
with initial data \(x(\tau _1)\) given by the solution of Eq. (9). Again we know that Eq. (10) has a unique continuous solution which belongs to \({\mathcal {M}}^2([\tau _1,\tau _2);{\mathbb {R}}^n)\) and has the property that \({\mathbb {E}}\Vert x(t)\Vert ^p<\infty \). Repeating this procedure, we see that Eq. (1) has a unique solution x(t) on [0, T] and has the property that
The proof is completed. \(\square \)
3 Stochastic theta method
To define the ST scheme, let us first explain how to simulate the discrete Markov chain \(\left\{ r_k^{\Delta },k=0,1,2,\dots \right\} \). Recall the property of the embedded discrete Markov chain:
Given a step size \(\Delta >0\), let \(r_k^{\Delta }=r(k\Delta )\) for \(k\ge 0\). Then \(\left\{ r_k^{\Delta },k=0,1,2,\dots \right\} \) is a discrete Markov chain with the one-step transition probability matrix
Hence, the discrete Markov chain \(\left\{ r_k^{\Delta },k=0,1,2,\dots \right\} \) can be simulated as follows: Let \(r_0^{\Delta }=i_0\) and compute a pseudo-random number \(\zeta _1\) from the uniform [0, 1] distribution. Define
where we set \(\sum _{j=1}^{0}{\mathbb {P}}_{i_0,j}(\Delta )=0\) as usual. In other words, we ensure that the probability of state s being chosen is given by \({\mathbb {P}}(r_1^{\Delta }=s)={\mathbb {P}}_{i_0,s}(\Delta )\). Generally, having calculated \(r_0^{\Delta },r_1^{\Delta },\dots ,r_k^{\Delta },\) we compute \(r_{k+1}^{\Delta }\) by drawing a uniform [0, 1] pseudo-random number \(\zeta _{k+1}\) and setting
This procedure can be carried out independently to obtain more trajectories.
After explaining how to simulate the discrete Markov chain, we can now define the ST approximate solution to Eq. (1). Let \(\Delta =1/m\) be a given step size with integer \(m\ge 1\), and let the gridpoints \(t_k\) be defined by \(t_k=k\Delta (k\in {\mathbb {N}})\). Since for arbitrary \(k\in {\mathbb {N}}\), there exist \(s\in {\mathbb {N}}\) and \(l=0,1,2,\dots ,m-1\) such that \(k=sm+l\), the adaptation of the ST method to (1) leads to a numerical process of the following type by setting \(X_0=x(0)=x_0, r_0^{\Delta }=r(0)=i_0\),
for \(s\in {\mathbb {N}}\) and \(l=0,1,2,\dots ,m-1\), where \(\Delta B_{sm+l}=B(t_{sm+l+1})-B(t_{sm+l})\), \(\theta \in [0,1]\) is a free parameter that is specified a priori. \(X_{sm+l}\) is an approximation to the exact solution \(x(t_{sm+l})\).
Since the ST scheme is semi-implicit when \(\theta \ne 0\), the first item that need to be considered is the existence and uniqueness of solutions of these equations. In that sense, we will employ the one-sided Lipschitz condition in the first argument of the function f, which is given in the following.
Assumption 3.1
There exists a positive constant L such that
for all \(x,{\bar{x}}, y\in {\mathbb {R}}^n\).
Remark 1
By Lemma 3.1 in Mao and Szpruch (2013), it is obvious to obtain that the ST method has a unique solution if \(\Delta <\frac{1}{L\theta }\). In the rest of this paper, we always assume that \(\Delta <\frac{1}{L\theta }\).
To implement numerical scheme (11), we define a map F, let
then we can represent (11) as follows:
for \(l=0,1,\dots , m-2,\)
for \(l=m-1\).
According to Assumption 3.1, there exists an inverse mapping \(F^{-1}\) and the solution to (11) can be represented in the following form:
for \(l=0,1,\dots , m-2,\)
for \(l=m-1\). Clearly, \(X_{sm+l+1}\) is \({\mathcal {F}}_{t_{sm+l+1}}\)-measurable.
4 pth moment boundedness of the ST method
Throughout this section, we fix \(T>0\) be arbitrary and show that the pth moment of the ST method is bounded. The following lemma shows that to guarantee the boundedness of moments for \(X_{sm+l}\) it is enough to bound the moments of \(F(X_{sm+l})\), where \(F(X_{sm+l})\) is defined by (12).
Lemma 4.1
Suppose that Assumption 2.2 holds. Let \(\delta \) be any given constant with \(1-4\alpha \theta \Delta \ge \delta >0\). Then for any \(p\ge 2\),
Moreover,
Proof
Using Assumption 2.2, we can arrive at
which implies
then applying the inequality \((x+y+z)^{\frac{p}{2}}\le 3^{\frac{p}{2}-1}(x^{\frac{p}{2}}+y^{\frac{p}{2}}+z^{\frac{p}{2}}),\forall x,y,z>0\), and the fact that \(1-4\alpha \theta \Delta \ge \delta \), we obtain
Furthermore, if \(l=0\), one can get that
from (15) directly, substituting (18) into (16), then
which gives
In particular, it from (18) that
\(\square \)
In what follows, for notational simplicity, we use the convention that C represents a generic positive constant independent of \(\Delta \), the value of which may vary with each appearance. For example, \(C=C+C\) and \(C=C\times C\) are understood in an appropriate sense. Moreover, we may give specific expressions of C when needed. Let us begin to establish the fundamental result of this paper that reveals the boundedness of the pth moment for the ST scheme.
Theorem 4.2
Let Assumptions 2.2–2.4 and 3.1 hold, and \(\theta \ge 0.5\). Then for any \(p\ge 2\), the ST scheme (11) has the following property:
Proof
For any \(s\in {\mathbb {N}}, l=0,1,\dots ,m-2\), using Assumption 2.2 and \(\theta \ge 0.5\), we have
where
Then we can infer that
substituting (18) and (19) into the last equation and recalling that \(1-4\alpha \theta \Delta \ge \delta \), we acquire
Thus, according to the inequality \(\left( \sum _{i=1}^k\vert a_i\vert \right) ^p\le k^{p-1}\sum _{i=1}^k\vert a_i\vert ^p\), \(p\ge 1\), one can arrive at
where \(C_1=5^{\frac{p}{2}-1}\left( \frac{2\alpha }{\delta }\right) ^{\frac{p}{2}}, C_2=5^{\frac{p}{2}-1}\left( 1+\frac{2\alpha }{\delta ^2}\right) ^{\frac{p}{2}}, C_3=5^{\frac{p}{2}-1}\), and we use the fact that \(l\Delta \le 1, l=0,1,\dots ,m-2.\)
Since \(\Delta B_{i}\) is \({\mathcal {F}}_{t_{i}}\)-independent, with the help of Hölder’s inequality and Assumption 2.4, we have
combining (20) and (21), we can obtain that
where
According to the definition of \(\Delta N_{sm+l}\) and \(F(X_{sm+l})\), using the time discrete Burkholder–Davis–Gundy type inequality, we yield
Similar to (24) and (25), one can get that
By substituting (20) and (27) into (26), we have
where
It follows from (23), (25) and (28) that
where we set \(\sum _{i=0}^{-1}{\mathbb {E}}\Vert F(X_{sm+i+1})\Vert ^p=0\) as usual. Here \(C_{12}=C_1+C_3C_4C+C_3C_{8}\), \(C_{13}=C_2+C_3C_5C+C_3C_{9}\), \(C_{14}=C_3C_6C+C_3C_{10}\), \(C_{15}=C_1+C_3C_7C+C_3C_{11}\). Using the discrete Gronwall inequality (Theorem 2.5 in Mao and Yuan 2006), we obtain
By induction, we divide the proof into several steps to show
Step 1. For \(s=0, l=0,1,\dots , m-2\), (29) implies
Noting that \(F(X_0)=X_0-f(X_0,X_0,i_0)\Delta , X_0=x_0\), using Assumption 2.3, we can easily get that
Substituting the last equations into (30) we have
and it follows from Lemma 4.1 that
Repeating the procedures as discussed above, we can also get that
Next we show that \({\mathbb {E}}\Vert X_m\Vert ^p\le C\) and \({\mathbb {E}}\Vert F(X_m)\Vert ^p\le C\). Applying (12), (13) and (14) again, by Assumptions 2.3 and 2.4, one has
hence by (31)–(33), we infer that
Moreover, repeating the process (15)–(17), one can get that
Using \(F(X_{sm+l})=X_{sm+l}-\theta f(X_{sm+l},X_{sm},r_{sm+l}^{\Delta })\Delta \) again, we obtain
Step 2. For \(s=1\), according to (29), we have
which gives
then it follows from Lemma 4.1 that
Adopting the same procedures as in the Step 1, we can arrive at \({\mathbb {E}}\Vert X_{2m}\Vert ^p\le C\), then \({\mathbb {E}}\Vert F(X_{2m})\Vert ^p\le C\) follows from (12).
Step 3. For \(s\in \{2,3,\dots ,[T]\}\), the following assertion can be proved in the same way as shown before, for any fixed T, there exists a constant C independent of \(\Delta \) such that
Combining Steps 1–3, we can get that
for any fixed T. The proof is completed. \(\square \)
5 Rate of strong convergence
It is convenient to work with a continuous extension of a numerical method here, because the continuous extension enables us to use the powerful continuous-time stochastic analysis to formulate theorems on numerical approximations. For this purpose, we introduce a new numerical scheme, which is called the forward–backward Euler–Maruyama (FBEM) scheme, to help us get a well-defined continuous-time numerical approximation.
First we compute the discrete values \(X_{sm+l}\) from the ST method, then we define the discrete FBEM scheme on \([s,s+1)\subset [0,\infty ),s\in {\mathbb {N}}\) by
where \(l=0,1,\dots , m-1, {\hat{X}}_0=X_0=x(0)\).
Let \(X(t)=X_{sm+l}\), \({\bar{X}}(t)={\hat{X}}_{sm+l}\), \(r^{\Delta }(t)=r_{sm+l}^{\Delta }\) for \(t\in [t_{sm+l},t_{sm+l+1})\), and the continuous FBEM scheme is defined by
on each interval \([t_{sm+l},t_{sm+l+1})\). We would like to remark that the continuous and discrete FBEM schemes coincide at the grid points, that is, \({\hat{X}}(t_{sm+l})={\hat{X}}_{sm+l}\).
Now we impose some stronger versions of Assumptions 2.2–2.4 to get the convergence rate.
Assumption 5.1
For any constant \(K>0\), there exists a positive constant \(K_1\) such that
for all \(i\in S\) and \(x, {\bar{x}},y, {\bar{y}}\in {\mathbb {R}}^n\).
Assumption 5.2
There exist constants \(K_2>0\) and \(\rho _1\ge 1\) such that
for all \(i\in S\) and \(x,y,{\bar{x}},{\bar{y}}\in {\mathbb {R}}^n.\)
Assumption 5.3
There exist constants \(K_3>0,K_4>0\) and \(\rho _2\ge 1\) such that
for all \(i\in S\) and \(x, y, {\bar{x}},{\bar{y}}\in {\mathbb {R}}^n\).
Assumption 5.4
There exists a positive constant \(K_5\) such that
for all \(i\in S\).
Remark 2
Assumptions 5.1–5.4 imply Assumptions 2.2–2.4. Suppose that Assumptions 5.1–5.4 hold, we can easily get that
which is Assumption 2.2, and
which is Assumption 2.3, as well as
which is Assumption 2.4, where \(K_6=K_3+2K_4+K_5\).
Under the assumptions above, we can determine the rate of strong convergence for ST scheme (11) to the solution of (1). First, we need some lemmas.
Lemma 5.5
Let Assumptions 5.1–5.4 hold, and \(\theta \ge 0.5\). Then for any \(p\ge 2\) and sufficiently small step size \(\Delta \), \({\hat{X}}_{sm+l}\) and \(X_{sm+l}\) obey
Proof
For any \(s\in {\mathbb {N}}, l = 0,1,\dots ,m-1\), summing up both schemes of the discrete FBEM (34) and ST (11), respectively, we have
then we can infer that
Thus, we obtain
According to Assumptions 5.2, 5.4 and the inequality \((\vert a\vert +\vert b\vert )^p\le 2^{p-1}(\vert a\vert ^p+\vert b\vert ^p), p\ge 1\), we can acquire
Similarly, one can also get that
Substituting (39) and (40) into (38), with the help of Theorem 4.2, for \(s\in [0,[T]]\), we yield
and the assertion follows. \(\square \)
Lemma 5.6
Let Assumptions 5.1–5.4 hold, and \(\theta \ge 0.5\). Then for any \(p\ge 2\),
Proof
It follows from Theorem 4.2 and Lemma 5.5 that
Moreover, according to (35) and Hölder’s inequality,
it follows that
Similar to (39), from Theorem 4.2 we can get that
According to Remark 2 and Theorem 1.7.2 in Mao (2007), together with Theorem 4.2 again, we can infer that
Substituting (41), (43) and (44) into (42), one has \({\mathbb {E}}\left\{ \sup _{0\le t\le T}\Vert {\hat{X}}(t)\Vert ^p\right\} \le C\). \(\square \)
Lemma 5.7
Let Assumptions 5.1–5.4 hold, and \(\theta \ge 0.5\), then for any \(p\ge 2\),
Proof
For any \(t\in [0,T]\), there always exist \(s\in {\mathbb {N}}\) and \(l\in \{0,1,\dots , m-1\}\) such that \(t\in [t_{sm+l}, t_{sm+l+1})\); hence,
Applying Hölder’s inequality and Theorem 1.7.1 in Mao (2007), for \(t\in [t_{sm+l}, t_{sm+l+1})\),
similar to (39), by Assumptions 5.2–5.4 and Theorem 4.2, we can easily get
which means
By substituting the last equation and (36) into (45), the desired assertion follows. \(\square \)
Lemma 5.8
Let Assumptions 5.1–5.4 hold, and \(\theta \ge 0.5\), then for any \(s\in {\mathbb {N}}\),
Proof
Using Assumption 5.2, we have
where in the last step we use the fact that \(X(u)=X_{sm+l},X(s)=X_{sm}\) and \(I_{\left\{ r(u)\ne r(t_{sm+l})\right\} }\) when \(t_{sm+l}< u< t_{sm+l+1}\) are conditionally independent with respect to the \(\sigma \)-algebra generated by \(r(t_{sm+l})\). By the property of Markov chain, one has
where \({\bar{C}}=\max _{1\le i\le N}(-\gamma _{ii})\). Hence,
Then using Theorem 4.2, one can get that
The second inequality can also be proved similarly. \(\square \)
Theorem 5.9
Under Assumptions 5.1–5.4, and \(\theta \ge 0.5\), the continuous FBEM method (35) strongly converges to the solution of hybrid SDEPCAs (1), that is
Proof
Let \(e(t)={\hat{X}}(t)-x(t)\), \(e_{\Delta }(t)={\hat{X}}(t)-X(t)\). For any \(t\in [0,T]\), there always exists \(s\in {\mathbb {N}}\) such that \(t\in [s,s+1)\), hence
For any \(t\in [s,s+1)\), it follows from (34) and (35) that
then according to the generalised Itô formula (Mao and Yuan 2006), one has
Hence, it is easy to see that
where
Applying the Burkholder–Davis–Gundy inequality and \(2ab\le a^2+b^2\), we obtain
Substituting the last equation into (48), we get
Using Assumption 5.2, Lemma 5.8, Hölder’s inequality, Theorem 4.2, Lemmas 5.6, and 5.7, one can acquire that
According to Lemma 5.8, Assumption 5.3, Theorem 4.2, Lemmas 5.6 and 5.7, we yield
Substituting \(J_1\) and \(J_2\) into (49), using Assumption 5.1, then
For \(s=0, t\in [0,1)\), (50) implies
then according to the Gronwall inequality and the continuity of \(\Vert e(u)\Vert ^2\), we have
In particular, we know that \({\mathbb {E}}\Vert e(1)\Vert ^2\le C\Delta \).
For \(s=1, t\in [1,2)\), (50) implies
using the Gronwall inequality and the continuity of \(\Vert e(u)\Vert ^2\) once more, we can also get
in particular, \({\mathbb {E}}\Vert e(2)\Vert ^2\le C\Delta \).
Repeating the same procedures, for any \(s\in [0,[T]]\), we always have
substituting this inequality into (47), which gives
The proof is completed. \(\square \)
We are now ready to formulate the main theorem of this paper.
Theorem 5.10
Let Assumptions 5.1–5.4 hold, and \(\theta \ge 0.5\), there exists a positive constant C, independent of \(\Delta \), such that the ST method (11) strongly converges to the solution of hybrid SDEPCAs (1), that is
Proof
It is apparent from the triangle inequality that
then the assertion follows from Lemma 5.7 and Theorem 5.9. \(\square \)
6 Numerical simulation
In this section, we consider the following scalar nonlinear hybrid SDEPCA
where \(f:{\mathbb {R}}\times {\mathbb {R}}\times S\rightarrow {\mathbb {R}}\),
and \(g:{\mathbb {R}}\times {\mathbb {R}}\times S\rightarrow {\mathbb {R}}\),
with the initial conditions \(x_0=1\) and \(i_0=1\in S=\{1,2\}\). Here B(t) is a scalar Brownian motion on \((\Omega ,{\mathcal {F}},\{{\mathcal {F}}_t\}_{t\ge 0},{\mathbb {P}})\), and r(t) is a right-continuous Markov chain taking values in S with the generator \(\Gamma =(\gamma _{ij})_{2\times 2}=\begin{bmatrix}-1 &{} 1 \\ 2 &{} -2 \end{bmatrix}.\)
By a straight calculation, one has
and
which means the coefficients satisfy Assumption 5.1. Similarly we can also verify that the coefficients satisfy other conditions of Theorem 5.10. We generate 2000 different discretized Brownian paths and use the numerical solution of the backward EM method with step-size \(\Delta =2^{-15}\) as the “exact solution”.
Let \(\epsilon \) and \(\eta \) denote the errors in mean square,
We calculate the errors in mean square \(\epsilon (1),\epsilon (2), \epsilon (3)\) and \(\eta (2), \eta (3)\) with step sizes \(2^{-6},2^{-7},2^{-8},2^{-9},2^{-10}\), respectively. The log–log mean square error plots corresponding to those chosen values of \(\Delta \) and \(\theta \) are given in Figs. 1 and 2. It is well known that the slope of a line in the log–log error plot implies the order of convergence for the numerical method. Graphically, the mean square error lines’ slopes are close to the reference lines’ slope. Therefore, it can be seen from Figs. 1 and 2 that the order of convergence in mean square for the ST method is close to 0.5.
References
Anderson WJ (2012) Continuous-time Markov chains: an applications-oriented approach. Springer Science & Business Media, Berlin
Hu J, Liu W, Deng F, Mao X (2020) Advances in stabilization of hybrid stochastic differential equations by delay feedback control. SIAM J Control Optim 58(2):735–754
Jobert A, Rogers LCG (2006) Option pricing with Markov-modulated dynamics. SIAM J Control Optim 44(6):2063–2078
Li X (2014) Existence and exponential stability of solutions for stochastic cellular neural networks with piecewise constant argument. J Appl Math 2014:145061
Li R, Chang Z (2007) Convergence of numerical solution to stochastic delay differential equation with Poisson jump and Markovian switching. Appl Math Comput 184(2):451–463
Li R, Hou Y (2006) Convergence and stability of numerical solutions to SDDEs with Markovian switching. Appl Math Comput 175(2):1080–1091
Li X, Mao X (2020) Stabilisation of highly nonlinear hybrid stochastic differential delay equations by delay feedback control. Automatica 112:108657
Li X, Ma Q, Yang H, Yuan C (2018) The numerical invariant measure of stochastic differential equations with Markovian switching. SIAM J Numer Anal 56(3):1435–1455
Mao X (2007) Stochastic differential equations and applications. Woodhead Publishing, Cambridge
Mao X, Szpruch L (2013) Strong convergence rates for backward Euler–Maruyama method for non-linear dissipative-type stochastic differential equations with super-linear diffusion coefficients. Stochastics 85(1):144–171
Mao X, Yuan C (2006) Stochastic differential equations with Markovian switching. Imperial College Press, London
Mao X, Yuan C, Yin G (2007) Approximations of Euler–Maruyama type for stochastic differential equations with Markovian switching, under non-lipschitz conditions. J Comput Appl Math 205(2):936–948
Mao X, Liu W, Hu L, Luo Q, Lu J (2014) Stabilization of hybrid stochastic differential equations by feedback control based on discrete-time state observations. Syst Control Lett 73:88–95
Milošević M, Jovanović M (2011) A Taylor polynomial approach in approximations of solution to pantograph stochastic differential equations with Markovian switching. Math Comput Model 53(1):280–293
Nguyen DH, Yin G (2016) Modeling and analysis of switching diffusion systems: past-dependent switching with a countable state space. SIAM J Control Optim 54(5):2450–2477
Smith DR (2002) Markov-switching and stochastic volatility diffusion models of short-term interest rates. J Bus Econ Stat 20(2):183–197
Song M, Mao X (2018) Almost sure exponential stability of hybrid stochastic functional differential equations. J Math Anal Appl 458(2):1390–1408
Wang L, Xue H (2007) Convergence of numerical solutions to stochastic differential delay equations with Poisson jump and Markovian switching. Appl Math Comput 188(2):1161–1172
Wiener J (1993) Generalized solutions of functional differential equations. World Scientific, Singapore
Xie Y, Zhang C (2020) Compensated split-step balanced methods for nonlinear stiff SDEs with jump-diffusion and piecewise continuous arguments. Sci China Math 63:2573–2594
Yin B, Ma Z (2011) Convergence of the semi-implicit Euler method for neutral stochastic delay differential equations with phase semi-Markovian switching. Appl Math Model 35(5):2094–2109
Yin G, Zhu C (2010) Hybrid switching diffusions: properties and applications. Springer, New York
You S, Liu W, Lu J, Mao X, Qiu Q (2015) Stabilization of hybrid systems by feedback control based on discrete-time state observations. SIAM J Control Optim 53(2):905–925
Yuan C, Mao X (2004) Convergence of the Euler–Maruyama method for stochastic differential equations with Markovian switching. Math Comput Simul 64(2):223–235
Zhang C, Xie Y (2019) Backward Euler–Maruyama method applied to nonlinear hybrid stochastic differential equations with time-variable delay. Sci China Math 62:597–616
Zhang Y, Li R, Huo X (2019) Switching-dominated stability of numerical solutions for hybrid neutral stochastic differential delay equations. Nonlinear Anal Hybrid Syst 33:76–92
Zhou S, Wu F (2009) Convergence of numerical solutions to neutral stochastic delay differential equations with Markovian switching. J Comput Appl Math 229(1):85–96
Acknowledgements
This work was supported by National Natural Science Foundation of China (Grant nos. 12071101 and 11671113).
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Pierre Etore.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Zhang, Y., Song, M., Liu, M. et al. Strong convergence rate of the stochastic theta method for nonlinear hybrid stochastic differential equations with piecewise continuous arguments. Comp. Appl. Math. 41, 372 (2022). https://doi.org/10.1007/s40314-022-02089-6
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40314-022-02089-6
Keywords
- Stochastic differential equations with piecewise continuous arguments (SDEPCAs)
- Stochastic theta (ST) method
- Forward–backward Euler–Maruyama (FBEM) method
- Convergence rate