Abstract
We prove weak convergence of order one for a class of exponential based integrators for SDEs with non-globally Lipschitz drift. Our analysis covers tamed versions of Geometric Brownian Motion (GBM) based methods as well as the standard exponential schemes. The numerical performance of both the GBM and exponential tamed methods through four different multi-level Monte Carlo techniques are compared. We observe that for linear noise the standard exponential tamed method requires severe restrictions on the step size unlike the GBM tamed method.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
We consider weak convergence analysis of tamed schemes for semi-linear stochastic differential equations (SDEs) of the form
with m, \(d\in \mathbb {N}\). Here \(W_t^i\) are iid Brownian motions on the probability space \((\Omega , \mathcal {F},\mathbb {P})\) with filtration \(({\mathcal {F} _ {t} }) _{t \in [0,T]}\). The nonlinearity \(F: \mathbb {R} ^d \rightarrow \mathbb {R} ^d\) only satisfies a one-sided Lipschitz condition, where as the \(g_i: \mathbb {R} ^d \rightarrow \mathbb {R} ^d\) are globally Lipschitz.
We assume throughout that the matrices A, \(B_i \in \mathbb {R} ^{d \times d }\) satisfy the following zero commutator conditions
Condition (1.2) is used to exploit the exact solution of Geometric Brownian Motion (GBM) below. By introducing the notation \(W_t:=[W^1_t,\ldots ,W^m_t]^T\),
and the matrix \(\sigma (x):=[\sigma _i(x)]\), (i.e. columns formed by \(\sigma _i\)) we can re-write (1.1) as
The schemes we examine are in the class of exponential integrators. These methods have proved to be effective schemes for many SDEs and stochastic partial differential equations (SPDEs). Originally only linear (or linearized) drift terms were exploited, see for example [3, 18, 19, 24, 27], however recently several exponential integrators that also exploit the linear terms in diffusion emerged [10, 11, 17, 32]. We are particularly interested in dealing with a one-sided Lipschitz drift term F with superlinear growth. Hutzenthaler et al. [14] showed both strong and weak divergence of Euler’s method for the general SDE (1.1) with superlinearly growing coefficients F and/or \(g_i\). Following [14] many explicit variants of Euler–Maruyama schemes that guarantee the strong convergence to the exact solution of SDE were derived, see for example [2, 15, 16, 22, 26, 30]. The most well known approach to deal with superlinearly growing coefficients is to employ “taming" to prevent the unbounded growth of numerical solutions. Although there has been much consideration of strong convergence of tamed schemes, see for example [15, 28, 29] there has been little consideration of weak convergence for SDEs [4, 5, 31, 33]. In [5], Bréhier investigated the weak error of the explicit tamed Euler scheme for SDE’s with one–sided Lipschitz continuous drift and additive noise to approximate averages with respect to the invariant distribution of the continuous time process. Bossy et al., [4], proposed and proved weak convergence of a semi-explicit exponential Euler scheme for a one-dimensional SDE with non-globally Lipschitz drift and diffusion behaving as \(x^{\alpha }\), with \(\alpha >1\). Due to the weak condition on the diffusion coefficient, their study covers regularity results for the solution of the Kolmogorov PDE commonly used in weak error analysis. In [31], Wang et al. formulated a general weak convergence theorem for one-step numerical approximations of SDEs with non-globally drift coefficients. They applied this to prove weak convergence of rate one for the tamed and backward Euler–Maruyama methods. We would like to point out that their analysis is not directly applicable to our GBM based schemes as it is not a classical one-step method but rather the composition of the GBM flow and a one-step flow. In the context of SPDEs, Cai et al. [7] constructed and analysed a weak convergence of a numerical scheme based on a spectral Galerkin method in space and a tamed version of the exponential Euler method. Below we impose conditions that are similar to those for the SPDE in [7]. We prove weak convergence for a class of exponential integrators where a form of taming is used for the one-sided Lipschitz drift term. The GBM methods exploit the exact solution of geometric Brownian motion, see [11, 12] where strong convergence of related methods were considered. Further, by taking \(A=B_i=0\) (or incorporating these terms into the nonlinearities) we simultaneously prove weak convergence for the standard exponential tamed scheme such as in [9]. Our proof is based on the Kolmogorov equation and one of the main difficulties is to take into account the stochasticity in the solution operator.
In our numerical experiments we compare different approaches to estimate the weak errors all using multi-level Monte Carlo (MLMC) techniques as reviewed in [13]. For a linear diffusion term we observe that the exponential tamed method does not perform well for larger time step sizes and hence a time step size restriction is required for MLMC techniques (for example to estimate the weak errors). This is of particular interest as tamed methods were originally introduced and strong convergence was examined precisely to control nonlinearities in the context of MLMC type simulations, see [15]. The GBM based method does not suffer in this way in our experiments for linear noise. For nonlinear diffusion both tamed based methods require a step size restriction for convergence on the MLMC techniques.
The paper is organized as follows: in Sect. 2 we state our assumptions on the drift and diffusion, present the new numerical method and state our main results. In Sect. 3 we present numerical simulations illustrating the rate of convergence using the MLMC simulations and compare the different approaches. The proofs of the main results are then given in detail in Sects. 4 and 5.
2 Setting and main results
Throughout the paper we let \({ \big \langle } \cdot , \cdot { \big \rangle }\) denote the standard inner product in \(\mathbb {R}^d\) (so \({ \big \langle } y , z { \big \rangle }=y ^\intercal z\) for \(y,z \in \mathbb {R}^d\) ) and \(\left\Vert \cdot \right\Vert \) represent both the Euclidean norm for vectors as well as the induced matrix norm. A vector \(\beta =(\beta _1,\beta _2,\ldots ,\beta _d)\) is a multiindex of order \(\vert \beta \vert =\sum _{i=1}^d \beta _i\) with nonnegative integers components. The partial derivative operator corresponding to the multiindex \(\beta \) is defined as
where \(h\in C^{\vert \beta \vert }(\mathbb {R}^d;\mathbb {R}^l)\). For a nonnegative integer j, we let \(\textbf{D}^jh(x)\) represent the jth order derivative operator applied to a function \(h\in C^j(\mathbb {R}^d;\mathbb {R}^l)\). When \(j=1\) we simply write the Jacobian as \(\textbf{D}h\).
Additionally, \(C^{k} _{b} (\mathbb {R}^d;\mathbb {R})\) denotes of the set of k-times differentiable functions, which are uniformly continuous and bounded together with their derivatives up to k-th order.
We define the sets
Before introducing our class of numerical methods we present three results from [8] on the existence and uniquenes, bounded moments and mean-square differentiability of the exact solution to (1.4).
2.1 Preliminary results for the SDE
For an SDE such as (1.4) with globally Lipschitz drift and diffusion coefficients many classical textbooks on stochastic analysis consider the Kolmogorov PDE. However, for non-globally Lipschitz drift coefficients there are far fewer results. One key work is Cerrai [8] for the properties of the exact solution to a SDE with one-sided drift coefficient.
Assumption 1
Let \(H\ge 0\) be given. Let the functions F and \(g_i \in C^h(\mathbb {R}^d;\mathbb {R}^d)\) for some \(h\ge H\) where \(i=1,2, \ldots , m\). Define the matrix g by the columns of \(g_i\) so that \(g=[g_i]\). Additionally, assume that
-
(i)
there exists \(r \ge 0\) such that for any \(j=0,1,\ldots ,h\)
\(\sup _{y \in \mathbb {R}^d} \left\Vert D^\alpha F(y)\right\Vert (1+\left\Vert y\right\Vert ^{2r+1-j})^{-1} < \infty , \qquad \vert \alpha \vert =j\);
-
(ii)
there exists \(\rho \le r\) such that for any \(j=0,1,\ldots ,h\)
\(\sup _{y \in \mathbb {R}^d} \left\Vert D^\alpha g_i(y)\right\Vert (1+\left\Vert y\right\Vert ^{\rho -j})^{-1} < \infty ,\qquad \vert \alpha \vert =j \);
-
(iii)
for all \(p>0\) there exist \(K=K(p) \in \mathbb {R}\) such that
\( { \big \langle } y , \textbf{D} F(z) y { \big \rangle }+ p \left\Vert \textbf{D}g (z)y \right\Vert ^2 \le K \left\Vert y\right\Vert ^2, \quad \forall \ y,z \in \mathbb {R}^d.\)
Assumption 2
There exist constants \(a>0\) and \(r,\gamma ,c \ge 0\) such that for any \(y,z \in \mathbb {R}^d\)
In particular, under Assumption 1, Cerrai [8] proves the existence and uniqueness of a solution to the SDE (1.1).
Theorem 2.1
([8], Theorem 1.3.5) Suppose that Assumption 1 holds with \(H=3\). Then there exists a unique solution \(X_t\) for \(t \in [0,T]\) to the SDE (1.1) along with the following moment bound for \(p \ge 1\) and constant \(C=C(p,T)>0\)
To get order one weak convergence we need to assume bounded moments of derivatives of the exact solution to the SDE (1.1) with respect to the initial condition. By the notation \(X_t ^x\), we emphasize that the initial condition is \(X_0=x\). We denote the derivative of the exact solution with respect to the initial condition by \(\textbf{D}_{x}X_t ^x\). The following regularity result is given in [8] (see also [31, 33]).
Theorem 2.2
[8, Theorem 1.3.6] Let Assumption 1 hold with \(H=3\), Assumption 2 hold and \(X_t^x\) be the solution to (1.1). Then \(X_t^x\) is h times mean-square differentiable and for \(i=1,\ldots , h\), \( p\ge 1\) and \(t \in [0,T]\)
Assumption 3
Let the test function \(\phi : \mathbb {R}^d \rightarrow \mathbb {R}\) and \(\phi \in C^{2} _{b}(\mathbb {R}^d)\).
Before continuing we define the quantity
Theorem 2.3
[8, Theorem 1.6.2] Let Assumption 1 with \(H=3\), Assumption 2 and Assumption 3 hold. Let \(X_t\) be the solution to (1.1). Then, \(\Psi (t,x)\) defined in (2.2) is the unique classical solution to the
Kolmogorov PDE
where, with \(\mu \) and \(\sigma _i\) defined in (1.3), \(\mathcal {L}\) is given by
Remark 2.1
Rather than imposing an extra hypotheses (Hypotheses 1.3) as in [8] on the diffusion coefficient we assume in Assumption 3 that \(\phi \) is in \(C^2 _b (\mathbb {R}^d)\). Together with the mean-square differentiability of \(X_t^x\) given in Theorem 2.2, \(\Psi (t,x)\) then satisfies smoothness and boundedness conditions required in the proof of Theorem 1.6.2 in [8].
2.2 Tamed GBM method and convergence results
Our class of exponential methods takes advantage of the linear terms in (1.1) by exploiting the stochastic operator
which, under the commutativity condition (1.2), is the solution to
Given \(N\in \mathbb {N}\) and final time T we set the time step size \(\Delta t=\frac{T}{N}\). This gives the uniform time partition \(0=t_0<t_1<t_2<\cdots <t_N=T\) with \(t_n=n\Delta t\). We denote increments \(\Delta W^i_n:=W^i_{t_{n+1}} -W^i_{t_n}\).
We propose and prove weak convergence of the tamed GBM method
where \(F^{tm} _{\Delta t}\) is the taming term given by
The taming function \(\alpha (\Delta t,y)\) is assumed to satisfy for all \(y\in \mathbb {R}^d\) and \(\Delta t>0\)
where \(C>0\) is a constant independent of \(\Delta t\). The typical form of taming (e.g. [15, 28]) is to take \(\alpha (\Delta t,y)=(1+\Delta t\Vert F(y)\Vert ^p)^{-1}\), so that with \(p=1\)
Strong convergence of (2.6) with (2.9) and \(g_i\equiv 0\) was considered in [12] and the efficiency of the method was illustrated numerically. If we take \(\alpha (t,y) \equiv 1\), so that \(F^{tm} _{\Delta t}=F\) in (2.6), then we obtain one of the methods in [11] (proved to be strongly convergent with order 1/2 under global Lipschitz assumptions). Further it was shown the method is highly efficient for SDE’s with dominant linear terms and, by a homotopy approach, is competitive when applied to highly non-linear forms of (1.1). Note that it is clear from (1.1) that taking \(B_i=0\) and \(A=B_i=0\) (or incorporating these terms into the nonlinearities) we recover from (2.6) the exponential tamed and the standard tamed methods respectively.
Before we state our main results we give an integral representation of the continuous version of the numerical method. Let us define the continuous extension of numerical solution \(\bar{Y}_t \) for \(t\in [t_n,t_{n+1}]\) by
Then it is clear that \(Y_n^N=\bar{Y}_{t_n} \) for \(t_n \le t < t_{n+1}\).
Lemma 2.1
Let \( \bar{Y}_t\) be the interpolated continuous version in (2.10) of the numerical solution given in (2.6). Then the differential for this solution is given by
where \(\mu ^{tm}=\mu ^{tm}(t,t_n)\) and \(\sigma _i^{tm}=\sigma _i^{tm}(t,t_n)\) and
Proof
By the definition of the inverse GBM (see for example [20]),
We seek the appropriate \(\mu ^{tm}\) and \(\sigma _i ^{tm}\). The product rule for the Itô differential gives
On the other hand (2.10) can be written as
By comparison of (2.14) with (2.13), we find
Solving matrix equation (2.16) for \( \sigma _i^{tm}\), the commutativity conditions (1.2) and substitution into equation (2.15) to determine \(\mu ^{tm}\) gives the desired result. \(\square \)
Our main result is weak convergence of order one of the numerical scheme (2.6) and to prove this we make use of bounded moments.
Assumption 4
There exists \(K>0\) such that for all \(i=1,\ldots , m\)
Remark 2.2
The global Lipschitz property given in Assumption 4 implies boundedness of \(\textbf{D}g\). Together with Assumption 1, by the mean value Theorem, there exists \(K \in \mathbb {R}\) such that for all \(y,z \in \mathbb {R}^d\)
Theorem 2.4
Let Assumption 1 with \(H=1\) and Assumption 4 hold. Then, for \(Y_n^N\) be given by (2.6) and for all \(p \in [1,\infty )\)
The proof of this Theorem is given in Sect. 4 and follows the approach of [15]. However we need to control the stochastic operator in (2.4) and in contrast to [12] we also now need to take account of the nonlinear diffusion terms \(g_i\). The main novelty in our proof below is the interaction between these two terms.
Theorem 2.5
Let Assumption 1 with \(H=4\), Assumption 2 and Assumption 4 hold. Let \(X_T\) be the solution to (1.1). Let \(Y_N^N\) be found from (2.6) and (2.8) hold. Then, for all \(\phi :\mathbb {R}^d\rightarrow \mathbb {R}\), \( \phi \in C^{4} _{b} (\mathbb {R}^d)\) there is a constant \(C>0\), independent of \(\Delta t\) such that
Although in Theorem 2.3 we have \(\phi \in C^2 _b (\mathbb {R}^d)\), to get first order weak convergence, we impose \(\phi \in C^4 _b (\mathbb {R}^d)\) in Theorem 2.5. We prove Theorem 2.5 in Sect. 5 using the Kolmogorov equation. Once again we need to take careful account of the stochastic operator \(\mathbf {\Phi } _{t,t_0} \) from (2.4) as well as dealing with the one-sided Lipschitz drift F. Before giving the proofs we present some numerical results.
3 Numerical results
We seek to estimate numerically the weak discretization error \( \vert \mathbb {E}{\left[ \phi (X_T)\right] }-\mathbb {E}{\left[ \phi (Y_N^N)\right] }\vert \), where \(Y_N^N\) is a numerical approximation to \(X_T\) with \(\Delta t= T/N\). To illustrate the rate of convergence we need to estimate the weak error for different values of \(\Delta t\). Our aim is to examine this in the absence of an analytic solution or where the numerical solution of the Kolomogorov equation is prohibitively expensive. We also wish to illustrate the weak convergence rate of order 1 that is proved in Theorem 2.5.
In practice we take a reference numerical solution so that \(X_T\approx {\mathbb {Y}}^{N_{R}} \) with \({N_{R}}>>N\) (with \(\Delta t_R=T/N_R\)). We then estimate
Note that \({\mathbb {Y}}^{N_{R}} \) may be computed by a different method to that for \(Y_N^N\). In [21] issues in computing weak errors using MLMC methods for SPDEs are discussed with multiplicative noise and upper and lower bounds of simulation errors are obtained. However the authors did not consider the simultaneous computation of a reference solution. In [1] the MLMC method is examined where the zero solution is asymptotically mean square stable and an importance sampling technique was introduced. We observe similar stability issues below.
We briefly discuss four approaches to estimate the weak error using the multi-level Monte-Carlo technique (MLMC), see [13, 23]. We denote these methods Trad, MLMCL0, MLMC, MLMCSR and examine them numerically in our experiments. In a traditional method, denoted Trad, we estimate independently \(\mathbb {E}{\left[ \phi ({\mathbb {Y}}^{N_{R}})\right] }\) and \(\mathbb {E}{\left[ \phi (Y_N^N)\right] }\) by a MLMC method. Thus, for the reference solution we have
and for the approximation, with \(N_L:=N\)
An alternative is to exploit difference in the telescoping sums from the MLMC approach for the reference \({\mathbb {Y}}^{N_{R}} \) (3.1) and numerical approximation \(Y_N^N\) (3.2). Subtracting we get
For the coarsest level \(\ell =0\) we have a choice of estimating \(\mathbb {E}{\left[ \phi ({\mathbb {Y}}^{N_0})\right] }-\mathbb {E}{\left[ \phi (Y_{N_0}^{N_0})\right] }\) or \(\mathbb {E}{\left[ \phi ({\mathbb {Y}}^{N_0})-\phi (Y_{N_0}^{N_0})\right] }\). If the reference \({\mathbb {Y}}^{N_0}\) is found with a different method to that of \(Y^{N_0}_{N_0}\), we can expect some variance reduction using the latter method over the former (and so fewer samples required to approximate the expectation). Note that if only computing for a fixed single L then we expect further variance reduction for the second term, i.e. in estimating
However, here we wish to compute for different values of L in order to illustrate the rate of convergence. Thus, rather than recomputing (3.3) for different L we instead as follows (to avoid recomputing the MLMC estimates).
For MLMCL0 we exploit efficiency of variance reduction on the coarsest level and so estimate \(\mathbb {E}{\left[ \phi ({\mathbb {Y}}^{N_0})-\phi (Y^{N_0}_{N_0})\right] },\) with estimations of
Where as for MLMC we estimate the weak error by (3.3) and estimate \(\mathbb {E}{\left[ \phi ({\mathbb {Y}}^{N_0})\right] }\), \(\mathbb {E}{\left[ \phi (Y^{N_0}_{N_0})\right] }\) separately. Finally we see from (3.3) that if the same numerical scheme is used to estimate both the reference solutions and \(Y_N^N\) that
This method does not allow for a more accurate (e.g. higher order) reference solution (as it uses a self reference). We call this MLMCSR and estimate
To illustrate the rate one weak convergence results of Theorem 2.5 we consider the following cubic SDE in \(\mathbb {R}^d\)
This cubic equation is often used as a test equation as solutions exhibit transitions between different phases. With \(d=1\) it is sometimes known as the Ginzburg-Landau equation, [20]. For larger d the system of SDEs can be thought of as arising from the spatial discretization of a stochastic Allen-Cahn equation, see for example [6, 7, 11, 14, 24]. We first examine the linear diffusion with \(\beta _1=0.1\), \(\beta _2=0\) and then non-linear diffusion \(\beta _1=\beta _2=0.1\). We look at dimensions \(d=1,4,10,50\) and \(d=100\). For \(d=1\) we take \(A=-4\), \(X_0=x=0.5\). For \(d\ge 4\) A is the standard tridiagonal matrix from the finite difference approximation to the Laplacian: \( A=0.5 d^{-2}\text {diag}(1,-2,1) \) and as initial data we take \(X_0=x=0.5\exp (-10(y-0.5)^2)\) where \(y=[1/d,2/d,\ldots ,(d-1)/d]^T\). We solve to \(T=1\) and as \(\phi \), we take \(\phi (y)=\Vert y\Vert ^2\). The standard exponential tamed method is used as the reference solution \({\mathbb {Y}}^{N_{R}} \) (except for MLMCSR when looking at our GBM tamed method) and we take (2.9) as our taming function. We perform 10 separate runs computing the full weak error convergence plot and use this to estimate the standard deviation.
First we examine the linear noise case with \(\beta _1=0.1\) (and \(\beta _2=0\)). Table 1 compares the efficiency of the different methods for estimating the weak error for dimensions \(d=1,4\) and 10. We restrict \(\Delta t<0.0075\) in these computations to ensure we see convergence of the exponential tamed method. We observe, as expected, that estimating the weak error directly by either MLMCL0 or MLMC is far more efficient than using the MLMC in a traditional from Trad. We also observe there is a slight advantage to using MLMCL0 compared to MLMC (due to some variance reduction on the coarsest level). Trad (GBM) looks at convergence for the GBM based methods and (Tamed) to the exponential tamed method, we observe a small overhead in the GBM method compared to Tamed. However, as we discuss below, there are advantages to the GBM approach. An alternative method would be to take a drift implicit method and to solve a nonlinear system of equations at each step. The weak errors are similar to those for GBM. Using the standard fsolve (without providing the linearization) in MATLAB leads to computational times two orders of magnitude larger (e.g. 2396 s (8.6) for MLMC0 and 2396 s (2.8) for MLMC) performed on the same cluster. For this reason we only show convergence plots for the explicit methods.
In Fig. 1 we show weak convergence for \(d=1\) and in all approaches observe the expected rate of convergence one. Furthermore in (c) and (d), where we have a direct comparison to the exponential tamed method, we see the GBM based method has a smaller error constant. We see in (d) that the two smallest step sizes (\(\Delta t=2^{-9}\) and \(\Delta t=2^{-9}\)) are too close to the reference time step to observe convergence.
In Fig. 2 we show convergence for \(d=4\) and observe the predicted convergence of rate 1. In (a), (b) and (c) we restrict the largest step so that \(\Delta t<0.02\). For the exponential tamed method it was essential to impose this restriction on the maximum step size as, although solutions remain bounded, the error is too large to observe convergence. These large solutions also lead to large variances and hence an infeasible large number of samples, (see also the discussion in [1]). This is illustrated by comparing (c) where we restrict \(\Delta t<0.02\) and (d) where \(\Delta t<0.2\). The exponential tamed method only starts to converge for \(\Delta t\le 0.02\). For the larger step sizes in (d) we see the GBM based method performs well (unlike the exponential tamed method). The small step size restriction on the standard exponential method as d increases makes it difficult to obtain a reference solution using this method using the MLMC method.
In Fig. 3 we illustrate convergence using MLMCSR for the two methods with \(d=50\) and \(d=100\). We only examine MLMCSR as for these larger values of d due to unreliable results the exponential tamed method for larger \(\Delta t\) values (as discussed for Fig. 2). We see that for \(d=50\) we require \(\Delta t<2\times 10^{-4}\) and for \(d=100\) that \(\Delta t<10^{-4}\) to obtain an approximation using the standard exponential tamed method. However, we observe that the GBM method converges with the predicted order and there is no issue with large variances.
For the nonlinear diffusion, with \(\beta _1=\beta _2=0.1\) in (3.4), we illustrate convergence in Fig. 4 (a) MLMCSR and (b) Trad. We have taken \(d=4\) and observe weak convergence of rate one. Both the methods MLMC and MLMCL0 also show rate one convergence with the GBM based method having a smaller error constant. For this nonlinear noise, for larger values of \(\Delta t\), although the taming ensures solutions remain bounded the variance does not reduce. This is now for both methods.
4 Proof of bounded moments
We let \( \Delta \textbf{W}_{k}:=(\Delta W_k^1,\Delta W_k^2,\ldots ,\Delta W_k^m)^T.\) For the drift we define the notation
where \(F^{tm} _{\Delta t}\) is defined in (2.7). We can use \(\mathcal {N}\) to re-write (2.6) as
We now prove the boundedness of the operator \(\mathbf {\Phi } _{t,t_0} \) in stochastic \(L^p\) spaces.
Lemma 4.1
Suppose \(p \ge 2\), \(0\le s<t\le T \) and let \(\mathbf {\Phi } _{t,s} \) be given in (2.4) with \(t_0=s\). Then,
-
(i)
For any \(\mathcal {F}_{s}\) measurable random variable v in \(L^p (\Omega ,\mathbb {R}^d )\)
$$\begin{aligned} \left\Vert \mathbf {\Phi } _{t,s} v\right\Vert _ {L^p(\Omega ,\mathbb {R}^d)} \le \exp { \left( \left( \left\Vert A\right\Vert +\frac{p-1}{2}\sum _{i=1} ^m \left\Vert B_i\right\Vert ^2\right) (t-s)\right) } \left\Vert v\right\Vert _ {L^p(\Omega ,\mathbb {R}^d)} . \end{aligned}$$ -
(ii)
There are constants \(\kappa _{1,2}>0\), independent of \((t-s)\) such that
$$\begin{aligned} \mathbb {E}\left[ \left\Vert \mathbf {\Phi } _{t,s} \right\Vert ^p \right] \le \kappa _1\exp (\kappa _2 (t-s)). \end{aligned}$$
Proof
For the proof of (i), see [12].
To show (ii) we recall that for a standard Gaussian variable \(z \sim N(0,1)\) and \(\alpha >0\)
From (2.4) we have that
By taking expectation of both sides in (4.3) and considering the independence and distribution of random values \(\vert W_t^i -W_{s}^i \vert \) and using (4.2) for \(i=1,2,\ldots , m\), we have
By the boundedness of the function \(\text {erf}\), the positive constants \(\kappa _1\) and \(\kappa _2\) are determined in terms of \(p, \left\Vert A\right\Vert \) and \(\left\Vert B_i\right\Vert \). \(\square \)
For the proof of Theorem 2.4, we adapt the approach given in [15]. In [12] a similar approach was taken with linear diffusion and so the additional element below is dealing with the nonlinear diffusion terms \(g_i\). We start by introducing appropriate sub events of \(\Omega \). We let \(\Omega _0^N:=\Omega \), then for \(n\in \mathbb {N}_N ^+\)
where the parameter r is as defined in Assumption 1. The dominating stochastic process \(D_n^N\) is defined with \(D_0^N:=\left( \lambda + \left\Vert x\right\Vert \right) e^{\lambda }\) and for \(n \in \mathbb {N}_{N}^+\) by
where
Here \(\lambda _0,\ldots ,\lambda _4\) are constants defined by
where the constant \(K>0\) denotes a constant that arises from the constants in Assumption 1, 4 and (2.17). The first result shows we can dominate the numerical solution on the set \(\Omega _n^N\).
Lemma 4.2
Let Assumption 1 hold with \(H=1\). Let \(Y_n^N\) be given by (2.6) and let (2.8) hold. For all \(n \in \mathbb {N}_{N}\) we have the pathwise inequality
Proof
On \(\Omega _{n+1}^N\) by construction \(\left\Vert \Delta \textbf{W}_n\right\Vert \le 1\) for \(n\in \mathbb {N}_{N-1}\). Therefore, \(\left\Vert \mathbf {\Phi } _{t_{n+1},t_n} \right\Vert < \infty \) for \(n \in \mathbb {N}_{N-1}\). We prove the Lemma on two subsets of \(\Omega _{n+1}^N\)
First, on \(S^{(1)}_{n+1}\), we have from (2.6) and the triangle inequality that
Since \(\left\Vert Y_n^N\right\Vert \le 1\), \(\left\Vert \Delta \textbf{W}_n\right\Vert \le 1\) on \(S^{(1)}_{n+1}\), and by the taming inequality \(\alpha (\Delta t,y)\le 1 \) from (2.8), we have that
Adding and subtracting F(0), \(\sum _{i=1}^mB_ig_i(0)\), as well as \(g_i(0)\) for \(i=1,\ldots , m\), and then applying the triangle inequality we get
By the polynomial growth condition on \(\textbf{D} F\) given in Assumption 1, the global Lipschitz condition on \(g_i\) (Assumption 4), and using that \(\Delta t\le T\)
On \(S^{(1)}_{n+1}\), since \(\left\Vert Y_n^N\right\Vert \le 1\) and \( \left\Vert \Delta \textbf{W}_n\right\Vert \le 1 \), we have
To bound \(S^{(2)}_{n+1}\), we start from (2.6) by squaring the norm
Applying the Cauchy–Schwarz and Arithmetic–Geometric inequalities
Let us consider \(\mathcal {N}(Y_n^N)\) (see (4.1)). Following the argument in [15, (40) in Proof of Lemma 3.1] we have
Further, by Assumption 4, the global Lipschitz property of the function \(g_i\)
and
By definition of \(\lambda \), we have
Therefore the linear growth of the second term on the RHS of (4.5)
on \(S^{(2)}_{n+1}\) is obtained.
The one-sided Lipschitz condition on F (2.17) and the Cauchy–Schwarz inequality give that
where we have used that \(1\le \Vert Y_n^N\Vert \). Therefore by the inequalities (4.6), (4.7), Cauchy–Schwarz inequality for the term \({ \big \langle } Y_n^N , \sum _{i=1}^m B_i g_i (Y_n^N) { \big \rangle }\) and \(\lambda >1\)
As a result, since \(\Delta t =T/N\), \(\alpha (\Delta t,Y_n^N) < 1 \), by (2.8), and \((4T^2+2T)\le \sqrt{\lambda }\), we see that on \(S^{(2)}_{n+1}\) (4.5) becomes
Now we carry out an induction argument on n. The case for \(n=0\) is obvious by initial condition on \(\Omega _0^N=\Omega \). Let \(l \in \mathbb {N}_{N-1}\) and assume \(\left\Vert Y_{n} ^N (\omega )\right\Vert \le D_n ^N(\omega )\) holds for all \(n \in \mathbb {N}_{l}\) where \(\omega \in \Omega _n ^N\). We now prove that
For all \( \omega \in \Omega _{l+1} ^N\), we have by the induction hypothesis \(\left\Vert Y_{n} ^N (\omega )\right\Vert \le D_n ^N(\omega )\le N^{1/(4r)}\), \(n\in \mathbb {N}_l\) and \(\Omega _{l+1} ^N \subseteq \Omega _{n+1} ^N\). For any \( \omega \in \Omega _{l+1} ^N\), \(\omega \) belongs to \(S_{l+1} ^{(1)}\) or \(S_{l+1} ^{(2)}\). For the inductive argument we define a random variable
as in [15]. This definition implies that \(1 \le \left\Vert Y_n^N (\omega )\right\Vert \le N^ {1/(4r)}\) for all \(n \in \lbrace \tau _{l+1} ^N (\omega )+1, \tau _{l+2} ^N (\omega ),\ldots , l \rbrace \). By the bound in (4.8)
By considering (4.4), the following completes the induction step and proof
\(\square \)
Lemma 4.3
For all \(p \ge 1\), \( \sup _{N \in \mathbb {N} } \mathbb {E}\left[ \sup _{ n \in \mathbb {N}_{N} } \vert D_n^N \vert ^p \right] < \infty . \)
Proof
By two applications of Hölder’s inequalities, we have
Then by (4.3) and since \(\lambda \) and x are deterministic we get
The second and third terms are shown to be bounded in [15, Lemma 3.5]. On the other hand, the boundedness of the last term on the RHS is due to the \(\Delta t\) dependence of the upper bound of \(\mathbb {E}\left[ \exp (\left\Vert B_i\right\Vert \vert \Delta W^i_k \vert ) \right] \) as shown in the proof of Lemma 4.1 (ii). \(\square \)
Lemma 4.4
Let \(\bar{Y}_t \) be given by (2.10). For all \(p \ge 2\) and \(t \in (t_n,t_{n+1}]\), there exists \(K=K(p,T)>0\) such that
Proof
Note that \(\bar{Y}_t \) is the solution to a linear Itô SDE
with initial condition \(\bar{Y}_{t_n} \) on the interval \([t_n,t]\) by Lemma 2.1. After adding and subtracting the terms \(\sum _{i=1}^{m} \int _{t_n}^t \mathbf {\Phi } _{s,t_n} g_i(0) {{{\textrm{d}}}W}_{s}^i\), we have by Jensen inequality for sums and integrals as well as the Burkholder–Davis–Gundy inequality (see [25])
Using Lemma 4.1, that \( \left\Vert \alpha (\Delta t,\bar{Y}_{t_k} )F(\bar{Y}_{t_k} ) \Delta t\right\Vert \le 1\) and the global Lipschitz property of \(g_i\), the desired inequality is achieved from an application of the continuous Gronwall Lemma. \(\square \)
We now use Lemmas 4.1–4.4 to prove the scheme has bounded moments.
4.1 Proof of Theorem 2.4: bounded moments
Proof
Consider the Itô equation for the continuous extension of numerical solution given in Lemma 2.1 on [0, t]
By adding and subtracting the terms \(\sum _{i=1}^m \sum _{k=0}^{n-1} \int _{t_k}^{t_{k+1}} \mathbf {\Phi } _{s,t_k} g_i(0) {{{\textrm{d}}}W}_{s}^i\) and using the triangle inequality, we have
By Lemma 4.1, that \( \left\Vert \alpha (\Delta t,\bar{Y}_{t_k} )F(\bar{Y}_{t_k} ) \Delta t\right\Vert \le 1 \) in (2.8), along with the Burkholder–Davis–Gundy inequality we have
By taking the square of both sides and applying the Jensen inequalities for sums and integrals
Substituting (4.9) into (4.10) and rewriting the resulting integral inequality in discrete form, we have
By the discrete Gronwall inequality
Following the bootstrap argument presented in [15, Lemma 3.9] to deal with the term KN on the RHS, we get
and
The boundedness of the term in right hand side of (4.12) is proved in Lemma 4.3. Hence the proof is complete. \(\square \)
5 Proof of weak convergence
We now take C as a generic constant that is independent of the time step size \(\Delta t\).
Lemma 5.1
Let the Assumptions of Theorem 2.3 hold. For each \(T>0\), there exists \(C>0\) such that for \(0 \le t \le T\)
Proof
By interchanging the derivative and expectation (see discussion in [4, sections 5.1 and 5.2]), the chain rule and the definition of \(\Psi \) in (2.2)
Then by the deterministic Cauchy–Schwarz and triangle inequality
where on the last step we used the stochastic Cauchy–Schwarz inequality. The fact that \(\phi \in C^2_b(\mathbb {R}^d)\) and Theorem 2.2 complete the proof. \(\square \)
Considering the definition of the function \(\Psi \) in (2.2), we define
We point out the inequality (5.1) implies, for u defined in (5.2), that
where \(\xi \) is a \(\mathbb {R}^d\)-valued random variable. We are now in a position to prove weak convergence.
5.1 Proof of Theorem 2.5: weak convergence
Proof
Recall the definition of \(\Psi (t,x)\) from (2.2). Applying the Itô formula on \([t_n,t_{n+1})\) for \(\Psi (t,\bar{Y}_t )\) with \(\bar{Y}_t \) given by (2.11) gives
where
and \(\mu ^{tm}\) and \(\sigma _i ^{tm}\) are defined in (2.12). Now we define the error function e by
By definition of u in (5.2) and by a telescopic sum, we have
An application of Itô ’s formula for \(u(s,\bar{Y}_s )\), using (5.4) on each of the subintervals, and the zero expectation of Itô integrals gives
On the other hand, \(u(s,\bar{Y}_s )=\Psi (T-s, \bar{Y}_s ) \) satisfies the Kolmogorov PDE (2.3)
So we have
Consider the integrand
Adding and subtracting the terms \(\textbf{ D} u(s,\bar{Y}_s ) F^{tm} _{\Delta t}(\bar{Y}_{t_k} ) \), \(\textbf{D} u(s,\bar{Y}_s )F(\bar{Y}_{t_k} )\) and defining
such that \(\Theta _1^k(t_k, \bar{Y}_{t_k} )=\Theta _2^k(t_k, \bar{Y}_{t_k} )\), \(\Theta _3^k(t_k, \bar{Y}_{t_k} )=\Theta _4^k(t_k, \bar{Y}_{t_k} )\), we have
where
In terms of these functions, the error (5.5) is given by
We now seek to bound each \(\left| \mathbb {E}\left[ T_i^k \right] \right| \) by a term of order \((s-t_k)\) or \(\Delta t\) on the subinterval \([t_k,t_{k+1}]\), for \(i=1,\ldots ,7\).
Starting with \(T_1^k\), the Itô equation for \(\mathbf {\Phi } _{} \) in (2.5) for \(\mathcal {F} _ {t_k}\) measurable, \(\mathbb {R}^d\)-valued random variables \(v\in L^2 (\Omega ,\mathbb {R}^d)\) gives
By the zero expectation of Itô integrals and Fubini’s Theorem,
Taking \(v=F^{tm} _{\Delta t}(\bar{Y}_{t_k} )\), boundedness of the operator \(\mathbf {\Phi } _{} \) from Lemma 4.1 and bounded moments of numerical solution in Theorem 2.4 and the inequality (5.3), we have
For \(T_2^k\), by the definition of \(F^{tm} _{\Delta t}\) in (2.8) we have
where the inequality \( \alpha (\Delta t,\bar{Y}_{t_k} ) < 1\) from (2.8), bounded moments of numerical solution from Theorem 2.4 and the inequality (5.3) are used.
We now consider \(T_3^k\). Itô ’s formula for \(F_{j}(\bar{Y}_s )\) around \(\bar{Y}_{t_k} \) where \(F_{j}:\mathbb {R}^d \rightarrow \mathbb {R}\) refers to the jth component of the vector valued function F for \(j=1,\ldots ,d\) gives
where
By zero expectation of the Itô integral and Jensen’s inequality for the integral and expectation, we have
By polynomial growth assumption on the derivatives of F and bounded numerical moments, we conclude that
Now consider the term \(T_4^k\). By applying the Itô formula to the function \(\Theta _1^k(t,\bar{Y}_t )\) over the interval \([t_k,s]\) and taking expectation, we have
The integrand is given by
By the continuity of all the derivatives of u, the condition that \(\phi \in C^4_b(\mathbb {R}^d)\) and bounded moments of the numerical method in Theorem 2.4 we have, similar to the arguments in [33, Appendix B], that \(\mathbb {E}\left[ (\frac{\partial }{\partial r} + \mathcal {\hat{L}}_k)\Theta _1^k(r,\bar{Y}_r ) \right] \) is uniformly bounded on \([t_k,s]\). Therefore we conclude
Similar arguments can be applied to \(T_5^k,T_6^k,T_7^k\) as well. Finally, by Fubini’s theorem for integrals
we obtain the desired order for the weak error of the scheme. \(\square \)
References
Ableidinger, M., Buckwar, E., Thalhammer, A.: An importance sampling technique in Monte Carlo methods for SDEs with a.s. stable and mean-square unstable equilibrium. J. Comput. Appl. Math. 316, 3–14 (2017). https://doi.org/10.1016/j.cam.2016.08.043
Beyn, W.J., Isaak, E., Kruse, R.: Stochastic C-Stability and B-Consistency of explicit and implicit Euler-type schemes. J. Sci. Comput. 67(3), 955–987 (2016). https://doi.org/10.1007/s10915-015-0114-4
Biscay, R., Jimenez, J., Riera, J., Valdes, P.: Local linearization method for the numerical solution of stochastic differential equations. Ann. Inst. Stat. Math. 48(4), 631–644 (1996)
Bossy, M., Jabir, J.F., Martinez, K.: On the weak convergence rate of an exponential Euler scheme for SDEs governed by coefficients with superlinear growth. Bernoulli 27(1), 312–347 (2021)
Bréhier, C.E.: Approximation of the invariant distribution for a class of ergodic sdes with one-sided lipschitz continuous drift coefficient using an explicit tamed Euler scheme. ESAIM PS 27, 841–866 (2023). https://doi.org/10.1051/ps/2023017
Bréhier, C.E., Goudenège, L.: Weak convergence rates of splitting schemes for the stochastic Allen–Cahn equation. BIT Numer. Math. 60(3), 543–582 (2020)
Cai, M., Gan, S., Wang, X.: Weak convergence rates for an explicit full-discretization of stochastic Allen–Cahn equation with additive noise. J. Sci. Comput. 86(3), 1–30 (2021)
Cerrai, S.: Second Order PDE’s in Finite and Infinite Dimension: A Probabilistic Approach, 1st edn. Springer, Berlin (2001)
Chen, Z., Gan, S., Wang, X.: A full-discrete exponential Euler approximation of the invariant measure for parabolic stochastic partial differential equations. Appl. Numer. Math. 157, 135–158 (2020). https://doi.org/10.1016/j.apnum.2020.05.008
Debrabant, K., Kværnø, A., Mattsson, N.C.: Runge–Kutta Lawson schemes for stochastic differential equations. BIT Numer. Math. 61(2), 381–409 (2021)
Erdogan, U., Lord, G.J.: A new class of exponential integrators for SDEs with multiplicative noise. IMA J. Numer. Anal. 39(2), 820–846 (2018). https://doi.org/10.1093/imanum/dry008
Erdogan, U., Lord, G.J.: Strong convergence of a GBM based tamed integrator for SDEs and an adaptive implementation. J. Comput. Appl. Math. 399, 113704 (2022). https://doi.org/10.1016/j.cam.2021.113704
Giles, M.B.: Multilevel Monte Carlo methods. Acta Numer. 24, 259–328 (2015). https://doi.org/10.1017/S096249291500001X
Hutzenthaler, M., Jentzen, A., Kloeden, P.E.: Strong and weak divergence in finite time of Euler’s method for stochastic differential equations with non-globally Lipschitz continuous coefficients. Proc. R. Soc. A 467(2130), 1563–1576 (2011). https://doi.org/10.1098/rspa.2010.0348
Hutzenthaler, M., Jentzen, A., Kloeden, P.E.: Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients. Ann. Appl. Probab. 22(4), 1611–1641 (2012). https://doi.org/10.1214/11-AAP803
İzgi, B., Çetin, C.: Semi-implicit split-step numerical methods for a class of nonlinear stochastic differential equations with non-Lipschitz drift terms. J. Comput. Appl. Math. 343, 62–79 (2018). https://doi.org/10.1016/j.cam.2018.03.027
Jentzen, A., Röckner, M.: A Milstein scheme for SPDEs. Found. Comput. Math. 15(2), 313–362 (2015)
Jimenez, J., Shoji, I., Ozaki, T.: Simulation of stochastic differential equations through the local linearization method. A comparative study. J. Stat. Phys. 94(3–4), 587–602 (1999)
Jimenez, J.C., Carbonell, F.: Convergence rate of weak local linearization schemes for stochastic differential equations with additive noise. J. Comput. Appl. Math. 279, 106–122 (2015). https://doi.org/10.1016/j.cam.2014.10.021
Kloeden, P., Platen, E.: Numerical Solution of Stochastic Differential Equations. Stochastic Modelling and Applied Probability. Springer, Berlin (2011)
Lang, A., Petersson, A.: Monte Carlo versus multilevel Monte Carlo in weak error simulations of SPDE approximations. Math. Comput. Simul. 143, 99–113 (2018). https://doi.org/10.1016/j.matcom.2017.05.002
Li, X., Mao, X., Yin, G.: Explicit numerical approximations for stochastic differential equations in finite and infinite horizons: truncation methods, convergence in pth moment and stability. IMA J. Numer. Anal. 39(2), 847–892 (2018). https://doi.org/10.1093/imanum/dry015
Lord, G.J., Powell, C.E., Shardlow, T.: An introduction to computational stochastic PDEs. Cambridge University Press, Cambridge (2014). https://doi.org/10.1017/CBO9781139017329
Lord, G.J., Rougemont, J.: A numerical scheme for stochastic PDEs with Gevrey regularity. IMA J. Numer. Anal. 24(4), 587–604 (2004)
Mao, X.: Stochastic Differential Equations and Applications. Elsevier, Amsterdam (2007)
Mao, X.: The truncated Euler-Maruyama method for stochastic differential equations. J. Comput. Appl. Math. 290, 370–384 (2015). https://doi.org/10.1016/j.cam.2015.06.002
Mora, C.M.: Weak exponential schemes for stochastic differential equations with additive noise. IMA J. Numer. Anal. 25(3), 486–506 (2005). https://doi.org/10.1093/imanum/dri001
Sabanis, S.: A note on tamed Euler approximations. Electron. Commun. Probab. 18, 1–10 (2013)
Sabanis, S.: Euler approximations with varying coefficients: the case of superlinearly growing diffusion coefficients. Ann. Appl. Probab. 26(4), 2083–2105 (2016). https://doi.org/10.1214/15-AAP1140
Tretyakov, M.V., Zhang, Z.: A fundamental mean-square convergence theorem for SDEs with locally Lipschitz coefficients and its applications. SIAM J. Numer. Anal. 51(6), 3135–3162 (2013). https://doi.org/10.1137/120902318
Wang, X., Zhao, Y., Zhang, Z.: Weak error analysis for strong approximation schemes of SDEs with super-linear coefficients. IMA J. Numer. Anal. (2023). https://doi.org/10.1093/imanum/drad083
Yang, G., Burrage, K., Komori, Y., Burrage, P., Ding, X.: A class of new Magnus-type methods for semi-linear non-commutative Ito stochastic differential equations. Numer. Alg. 1–25 (2021)
Zhao, Y., Wang, X.: Weak approximation schemes for SDEs with super-linearly growing coefficients. Appl. Numer. Math. (2024). https://doi.org/10.1016/j.apnum.2024.01.003
Acknowledgements
The first author is supported by The Scientific and Technological Research Council of Turkey (TUBITAK) 2219-International Postdoctoral Research Fellowship Program, Grant Number 1059B192202051. We are grateful to two anonymous reviewers for their comments and suggestions.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
Not applicable.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The first author is supported by The Scientific and Technological Research Council of Turkey (TUBITAK) 2219-International Postdoctoral Research Fellowship Program, Grant Number 1059B192202051.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Erdoğan, U., Lord, G.J. Weak convergence of tamed exponential integrators for stochastic differential equations. Bit Numer Math 64, 29 (2024). https://doi.org/10.1007/s10543-024-01029-6
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10543-024-01029-6
Keywords
- SDEs
- Exponential integrator
- Geometric brownian motion
- Weak convergence
- Tamed Euler–Maruyama methods
- Multi-level Monte Carlo method