Abstract
For a class of finite elements approximations for linear stochastic parabolic PDEs it is proved that one can accelerate the rate of convergence by Richardson extrapolation. More precisely, by taking appropriate mixtures of finite elements approximations one can accelerate the convergence to any given speed provided the coefficients, the initial and free data are sufficiently smooth.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
We are interested in finite elements approximations for Cauchy problems for stochastic parabolic PDEs of the form of Eq. (2.1) below. Such kind of equations arise in various fields of sciences and engineering, for example in nonlinear filtering of partially observed diffusion processes. Therefore these equations have been intensively studied in the literature, and theories for their solvability and numerical methods for approximations of their solutions have been developed. Since the computational effort to get reasonably accurate numerical solutions grow rapidly with the dimension d of the state space, it is important to investigate the possibility of accelerating the convergence of spatial discretisations by Richardson extrapolation. About a century ago Lewis Fry Richardson had the idea in [18] that the speed of convergence of numerical approximations, which depend on some parameter h converging to zero, can be increased if one takes appropriate linear combinations of approximations corresponding to different parameters. This method to accelerate the convergence, called Richardson extrapolation, works when the approximations admit a power series expansion in h at \(h=0\) with a remainder term, which can be estimated by a higher power of h. In such cases, taking appropriate mixtures of approximations with different parameters, one can eliminate all other terms but the zero order term and the remainder in the expansion. In this way, the order of accuracy of the mixtures is the exponent \(k+1\) of the power \(h^{k+1}\), that estimates the remainder. For various numerical methods applied to solving deterministic partial differential equations (PDEs) it has been proved that such expansions exist and that Richardson extrapolations can spectacularly increase the speed of convergence of the methods, see, e.g., [16, 17] and [20]. Richardson’s idea has also been applied to numerical solutions of stochastic equations. It was shown first in [21] that by Richardson extrapolation one can accelerate the weak convergence of Euler approximations of stochastic differential equations. Further results in this direction can be found in [14, 15] and the references therein. For stochastic PDEs the first result on accelerated finite difference schemes appears in [7], where it is shown that by Richardson extrapolation one can accelerate the speed of finite difference schemes in the spatial variables for linear stochastic parabolic PDEs to any high order, provided the initial condition and free terms are sufficiently smooth. This result was extended to (possibly) degenerate stochastic PDEs in to [6, 8] and [9]. Starting with [22] finite elements approximations for stochastic PDEs have been investigated in many publications, see, for example, [3, 4, 10,11,12] and [23].
Our main result, Theorem 2.4 in this paper, states that for a class of finite elements approximations for stochastic parabolic PDEs given in the whole space an expansion in terms of powers of a parameter h, proportional to the size of the finite elements, exists up to any high order, if the coefficients, the initial data and the free terms are sufficiently smooth. Then clearly, we can apply Richardson extrapolation to such finite elements approximations in order to accelerate the convergence. The speed we can achieve depends on the degree of smoothness of the coefficients, the initial data and free terms; see Corollary 2.5. Note that due to the symmetry we require for the finite elements, in order to achieve an accuracy of order \(J+1\) we only need \(\lfloor \frac{J}{2} \rfloor \) terms in the mixture of finite elements approximation. As far as we know this is the first result on accelerated finite elements by Richardson extrapolation for stochastic parabolic equations. There are nice results on Richardson extrapolation for finite elements schemes in the literature for some (deterministic) elliptic problems; see, e.g., [1, 2] and the literature therein.
We note that in the present paper we consider stochastic PDEs on the whole space \({\mathbb {R}}^d\) in the spatial variable, and our finite elements approximations are the solutions of infinite dimensional systems of equations. Therefore one may think that our accelerated finite elements schemes cannot have any practical use. In fact they can be implemented if first we localise the stochastic PDEs in the spatial variable by multiplying their coefficients, initial and free data by sufficiently smooth non-negative “cut-off” functions with value 1 on a ball of radius R and vanishing outside of a bigger ball. Then our finite elements schemes for the “localised stochastic PDEs” are fully implementable and one can show that the results of the present paper can be extended to them. Moreover, by a theorem from [6] the error caused by the localization is of order \(\exp (-\delta R^2)\) within a ball of radius \(R'<R\). Moreover, under some further constraints about a bounded domain D and particular classes of finite elements such as those described in Sects. 6.1 and 6.2, our arguments could extend to parabolic stochastic PDEs on D with periodic boundary conditions. Note that our technique relies on finite elements defined by scaling and shifting one given mother element, and that the dyadic rescaling used to achieve a given speed of convergence is similar to that of wavelet approximation. We remark that our accelerated finite elements approximations can be applied also to implicit Euler–Maruyama time discretisations of stochastic parabolic PDEs to achieve higher order convergence with respect to the spatial mesh parameter of fully discretised schemes. However, as one can see by adapting and argument from [5], the strong rate of convergence of these fully discretised schemes with respect to the temporal mesh parameter cannot be accelerated by Richardson approximation. Dealing with weak speed of convergence of time discretisations is beyond the scope of this paper.
In conclusion we introduce some notation used in the paper. All random elements are defined on a fixed probability space \((\Omega ,{\mathcal {F}},P)\) equipped with an increasing family \(({\mathcal {F}}_t)_{t\ge 0}\) of \(\sigma \)-algebras \({\mathcal {F}}_{t}\subset {\mathcal {F}}\). The predictable \(\sigma \)-algebra of subsets of \(\Omega \times [0,\infty )\) is denoted by \({\mathcal {P}}\), and the \(\sigma \)-algebra of the Borel subsets of \({\mathbb {R}}^d\) is denoted by \({{\mathcal {B}}}({\mathbb {R}}^d)\). We use the notation
for first order and second order partial derivatives in \(x=(x_1,\ldots ,x_d)\in {\mathbb {R}}^d\). For integers \(m\ge 0\) the Sobolev space \(H^m\) is defined as the closure of \(C_0^{\infty }\), the space of real-valued smooth functions \(\varphi \) on \({\mathbb {R}}^d\) with compact support, in the norm \(|\varphi |_m\) defined by
where \(D^{\alpha }=D_1^{\alpha _1}\ldots D_d^{\alpha _d}\) and \(|\alpha |=\alpha _1+\cdots +\alpha _d\) for multi-indices \(\alpha =(\alpha _1,\ldots ,\alpha _d)\), \(\alpha _i\in \{0,1,\ldots ,d\}\), and \(D_i^0\) is the identity operator for \(i=1,\ldots ,d\). Similarly, the Sobolev space \(H^m(l_2)\) of \(l_2\)-valued functions are defined on \({\mathbb {R}}^d\) as the closure of the of \(l_2\)-valued smooth functions \(\varphi =(\varphi _i)_{i=1}^{\infty }\) on \({\mathbb {R}}^d\) with compact support, in the norm denoted also by \(|\varphi |_m\) and defined as in (1.1) with \( \sum _{i=1}^{\infty }| D^{\alpha }\varphi _i(x)|^2 \) in place of \(|D^{\alpha }\varphi (x)|^2\). Unless stated otherwise, throughout the paper we use the summation convention with respect to repeated indices. The summation over an empty set means 0. We denote by C and N constants which may change from one line to the next, and by C(a) and N(a) constants depending on a parameter a.
For theorems and notations in the \(L_2\)-theory of stochastic PDEs the reader is referred to [13] or [19].
2 Framework and some notations
Let \((\Omega ,{\mathcal {F}},P,({{\mathcal {F}}}_t)_{t\ge 0})\) be a complete filtered probability space carrying a sequence of independent Wiener martingales \(W=(W^{\rho })_{\rho =1}^{\infty }\) with respect to a filtration \(({{\mathcal {F}}}_t)_{t\ge 0}\).
We consider the stochastic PDE problem
with initial condition
for a given \(\phi \in H^0=L_2({\mathbb {R}}^d)\), where
with \({\mathcal {P}}\otimes {{\mathcal {B}}}({\mathbb {R}}^d)\)-measurable real-valued bounded functions \(a^{ij}\), \(b^i\), c, and \(l_2\)-valued bounded functions \(\sigma ^{i}=(\sigma ^{i\rho })_{\rho =1}^{\infty }\) and \(\nu =(\nu ^{\rho })_{\rho =1}^{\infty }\) defined on \(\Omega \times [0,T]\times {\mathbb {R}}^d\) for \(i,j\in \{1,\ldots , d\}\). Furthermore, \(a^{ij}_t(x)=a^{j i}_t(x)\) a.s. for every \((t,x)\in [0,T]\times {\mathbb {R}}^d\). For \(i=1,2,\ldots , d\) the notation \(D_i = \frac{\partial }{\partial x_i}\) means the partial derivative in the i-th coordinate direction.
The free terms f and \(g=(g^{\rho })_{\rho =1}^{\infty }\) are \({\mathcal {P}}\otimes {{\mathcal {B}}}({\mathbb {R}}^d)\)-measurable functions on \(\Omega \times [0,T]\times {\mathbb {R}}^d\), with values in \({\mathbb {R}}\) and \(l_2\) respectively. Let \(H^{m}(l_2)\) denote the \(H^m\) space of \(l_2\)-valued functions on \({\mathbb {R}}^d\). We use the notation \(|\varphi |_m\) for the \(H^m\)-norm of \(\varphi \in H^m\) and of \(\varphi \in H^m(l_2)\), and \(|\varphi |_0\) denotes the \(L_2\)-norm of \(\varphi \in H^0=L_2\).
Let \(m\ge 0\) be an integer, \(K\ge 0\) be a constant and make the following assumptions.
Assumption 2.1
The derivatives in \(x\in {\mathbb {R}}^d\) up to order m of the coefficients \(a^{ij}\), \(b^i\), c, and of the coefficients \(\sigma ^{i}\), \(\nu \) are \({\mathcal {P}}\otimes {{\mathcal {B}}}({\mathbb {R}})\)-measurable functions with values in \({\mathbb {R}}\) and in \(l_2\)-respectively. For almost every \(\omega \) they are continuous in x, and they are bounded in magnitude by K.
Assumption 2.2
The function \(\phi \) is an \(H^m\)-valued \({{\mathcal {F}}}_0\)-measurable random variable, and f and \(g=(g^{\rho })_{\rho =1}^{\infty }\) are predictable processes with values in \(H^{m}\) and \(H^m(l_2)\), respectively, such that
Assumption 2.3
There exists a constant \(\kappa >0\), such that for \((\omega ,t,x)\in \Omega \times [0,T] \times {\mathbb {R}}^d\)
For integers \(n\ge 0\) let \({\mathbb {W}}^n_2(0,T)\) denote the space of \(H^n\)-valued predictable processes \((u_t)_{t\in [0,T]}\) such that almost surely
Definition 2.1
A continuous \(L_2\)-valued adapted process \((u_t)_{t\in [0,T]}\) is a generalised solution to (2.1)–(2.2) if it is in \({\mathbb {W}}^1_2(0,T)\), and almost surely
for all \(t\in [0,T]\) and \(\varphi \in C_0^{\infty }\), where \(D^{*}_{i}:=-D_{i}\) for \(i\in \{1,2,\ldots ,d\}\), and (, ) denotes the inner product in \(L_2\).
For \(m\ge 0\) set
Then the following theorem is well-known (see, e.g., [19]).
Theorem 2.1
Let Assumptions 2.1, 2.2 and 2.3 hold. Then (2.1)–(2.2) has a unique generalised solution \(u=(u_t)_{t\in [0,T]}\). Moreover, \(u\in {\mathbb {W}}^{m+1}_2(0,T)\), it is an \(H^{m}\)-valued continuous process, and
where C is a constant depending only on \(\kappa \), d, T, m and K.
The finite elements we consider in this paper are determined by a continuous real function \(\psi \in H^1\) with compact support, and a finite set \(\Lambda \subset {\mathbb {Q}}^d\), containing the zero vector, such that \(\psi \) and \(\Lambda \) are symmetric, i.e.,
We assume that \(|\psi |_{L_1}=1\), which can be achieved by scaling. For each \(h\ne 0\) and \(x\in {\mathbb {R}}^d\) we set \(\psi ^h_x(\cdot ):=\psi ((\cdot -x)/h)\), and our set of finite elements is the collection of functions \(\{\psi ^{h}_x:x\in {\mathbb {G}}_h\}\), where
Let \(V_h\) denote the vector space
where \(\ell _2({\mathbb {G}}_h)\) is the space of functions U on \({\mathbb {G}}_h\) such that
Definition 2.2
An \(L_2({\mathbb {R}}^d)\)-valued continuous process \(u^h=(u^h_t)_{t\in [0,T]}\) is a finite elements approximation of u if it takes values in \(V_h\) and almost surely
for all \(t\in [0,T]\) and \(\psi _x^h\) is as above for \(x\in {\mathbb {G}}_h\). The process \(u^h\) is also called a \(V_h\)-solution to (2.8) on [0, T].
Since by definition a \(V_h\)-valued solution \((u^h_t)_{t\in [0,T]}\) to (2.8) is of the form
we need to solve (2.8) for the random field \(\{U^h_t(y):y\in {\mathbb {G}}_{h}, t\in [0,T]\}\). Remark that (2.8) is an infinite system of stochastic equations. In practice one should “truncate” this system to solve numerically a suitable finite system instead, and one should also estimate the error caused by the truncation. We will study such a procedure and the corresponding error elsewhere.
Our aim in this paper is to show that for some well-chosen functions \(\psi \), the above finite elements scheme has a unique solution \(u^h\) for every \(h\ne 0\), and that for a given integer \(k\ge 0\) there exist random fields \(v^{(0)}\), \(v^{(1)}\),...,\(v^{(k)}\) and \(r_k\), on \([0,T]\times {\mathbb {G}}_h\), such that almost surely
where \(v^{(0)}\),..., \(v^{(k)}\) do not depend on h, and there is a constant N, independent of h, such that
for all \(|h|\in (0,1]\) and some \(m> \frac{d}{2}\).
To write (2.8) more explicitly as an equation for \((U^h_t(y))_{y\in {\mathbb {G}}_h}\), we introduce the following notation:
where \(\psi _{\lambda }:=\psi ^1_{\lambda }\), and \({\mathbb {G}}:={\mathbb {G}}_1\).
Lemma 2.2
For \(\alpha , \beta \in \{1,\ldots , d\}\) and \(\lambda \in {\mathbb {G}}\) we have:
Proof
Since \(\psi (-x)=\psi (x)\) we deduce that for any \(\alpha \in \{1,\ldots , d\}\) we have \(D_\alpha \psi (-x)= - D_\alpha \psi (x)\). Hence for any \(\alpha , \beta \in \{1,\ldots , d\}\) and \(\lambda \in {\mathbb {G}}\), a change of variables yields
this concludes the proof. \(\square \)
To prove the existence of a unique \(V_h\)-valued solution to (2.8), and a suitable estimate for it, we need the following condition.
Assumption 2.4
There is a constant \(\delta >0\) such that
Remark 2.1
Note that since \(\psi \in H^1\) has compact support, there exists a constant M such that
Remark 2.2
Due to Assumption 2.4 for \(h\ne 0\), \(u:=\sum _{y\in {\mathbb {G}}_h}U(y)\psi ^h_y\), \(U=(U(y))_{y\in {\mathbb {G}}_h}\in \ell _2({\mathbb {G}}_h)\) we have
Clearly, since \(\psi \) has compact support, only finitely many \(\lambda \in {\mathbb {G}}\) are such that \((\psi _\lambda ,\psi )\ne 0\); hence
where N is a constant depending only on \(\psi \).
By virtue of this remark for each \(h\ne 0\) the linear mapping \({\varvec{\Phi }}_h\) from \(\ell _2({\mathbb {G}}_h)\) to \(V_h\subset L_2({\mathbb {R}}^d)\), defined by
is a one-to-one linear operator such that the norms of U and \({\varvec{\Phi }}_hU\) are equivalent, with constants independent of h. In particular, \(V_h\) is a closed subspace of \(L_2({\mathbb {R}}^d)\). Moreover, since \(D_i\psi \) has compact support, (2.12) implies that
where N is a constant depending only on \(D_i\psi \) and \(\delta \). Hence for any \(h>0\)
with a constant \(N=N(\psi ,d,\delta )\) which does not depend on h.
Theorem 2.3
Let Assumptions 2.1 through 2.4 hold with \(m=0\). Then for each \(h\ne 0\) Eq. (2.8) has a unique \(V_h\)-solution \(u^h\). Moreover, there is a constant \(N=N(d,K,\kappa )\) independent of h such that
for all \(h\ne 0\), where \(\pi ^h\) denotes the orthogonal projection of \(H^0=L_2\) into \(V_h\).
Proof
We fix \(h\ne 0\) and define the bilinear forms \(A^h\) and \(B^{h\rho }\) by
for all \(u,v\in V_h\). Using Assumption 2.1 with \(m=0\), by virtue of (2.13) we have a constant \(C=C(|h|,K,d,\delta , \psi )\), such that
Hence, identifying \(V_h\) with its dual space \((V_h)^{*}\) by the help of the \(L_2({\mathbb {R}}^d)\) inner product in \(V_h\), we can see there exist bounded linear operators \({\mathbb {A}}^h_s\) and \({\mathbb {B}}^{h\rho }_s\) on \(V_h\) such that
and for all \(\omega \in \Omega \) and \(t\in [0,T]\). Thus (2.8) can be rewritten as
which is an (affine) linear SDE in the Hilbert space \(V_h\). Hence, by classical results on solvability of SDEs with Lipschitz continuous coefficients in Hilbert spaces we get a unique \(V_h\)-solution \(u^h=(u^h_t)_{t\in [0,T]}\). To prove estimate (2.14) we may assume \(E{\mathcal {K}}_0^2<\infty \). By applying Itô’s formula to \(|u^h|_0^2\) we obtain
with
Owing to Assumptions 2.1, 2.2 and 2.3, by standard estimates we get
with a constant \(N=N(K,\kappa ,d)\); thus from (2.16) using Gronwall’s lemma we obtain
with a constant \(N=N(T,K,\kappa ,d)\). One can estimate \(E\sup _{t\le T} |u^h_t|_0^2\) also in a standard way. Namely, since
with a constant \(N=N(K,d)\), by the Davis inequality we have
Taking supremum in t in both sides of (2.16) and then using (2.17), (2.18) and (2.19), we obtain estimate (2.14). \(\square \)
Remark 2.3
An easy computation using the symmetry of \(\psi \) imposed in (2.6) shows that for every \(x\in {\mathbb {R}}^d\) and \(h\ne 0\) we have \(\psi ^{-h}_x=\psi ^h_x\). Hence the uniqueness of the solution to (2.8) proved in Theorem 2.3 implies that the processes \(u^{-h}_t\) and \(u^h_t\) agree for \(t\in [0,T]\) a.s.
To prove rate of convergence results we introduce more conditions on \(\psi \) and \(\Lambda \).
Notation Let \(\Gamma \) denote the set of vectors \(\lambda \) in \({\mathbb {G}}\) such that the intersection of the support of \(\psi _{\lambda }:=\psi ^1_{\lambda }\) and the support of \(\psi \) has positive Lebesgue measure in \({\mathbb {R}}^d\). Then \(\Gamma \) is a finite set.
Assumption 2.5
Let \(R_\lambda \), \(R^i_\lambda \) and \(R^{ij}_\lambda \) be defined by (2.11); then for \(i,j,k,l\in \{1,2,\ldots ,d\}\):
where
and for sets of indices A and B the notation \(\delta _{A,B}\) means 1 when \(A=B\) and 0 otherwise.
Note that if Assumption 2.5 holds true, then for any family of real numbers \(X_{ij,kl}, i,j,k,l\in \{1,\ldots , d\}\) such that \(X_{ij,kl}=X_{ji,kl}\) we deduce from the identities (2.22) that
Our main result reads as follows.
Theorem 2.4
Let \(J\ge 0\) be an integer. Let Assumptions 2.1 and 2.2 hold with \(m>2J+\frac{d}{2}+2\). Assume also Assumption 2.3 and Assumptions 2.4 and 2.5 on \(\psi \) and \(\Lambda \). Then expansion (2.9) and estimate (2.10) hold with a constant \(N=N(m, J,\kappa ,K,d,\psi ,\Lambda )\), where \(v^{(0)}=u\) is the solution of (2.1) with initial condition \(\phi \) in (2.2). Moreover, in the expansion (2.9) we have \(v^{(j)}_t=0\) for odd values of j.
Set
with \(\bar{J}:= \lfloor \frac{J}{2} \rfloor \), \((c_0,\ldots ,c_{\bar{J}}) =(1,0\ldots ,0)V^{-1}\), where \(V^{-1}\) is the inverse of the \((\bar{J}+1)\times (\bar{J}+1)\) Vandermonde matrix
We make also the following assumption.
Assumption 2.6
Corollary 2.5
Let Assumption 2.6 and the assumptions of Theorem 2.4 hold. Then
for \(|h|\in (0,1]\), with a constant \(N=N(m,K,\kappa ,J,T,d,\psi ,\Lambda )\) independent of h, where u is the solution of (2.1)–(2.2).
3 Preliminaries
Assumptions 2.1, 2.2 and 2.4 are assumed to hold throughout this section. Recall that \(|\cdot |_{0,h}\) denotes the norm, and \((\cdot ,\cdot )_{0,h}\) denotes the inner product in \(\ell _2({\mathbb {G}}_h)\), i.e.,
for functions \(\varphi _1, \varphi _2 \in \ell _2({\mathbb {G}}_h)\).
Dividing by \(|h|^d\), it is easy to see that the Eq. (2.8) for the finite elements approximation
can be rewritten for \((U^h_t(x))_{x\in {\mathbb {G}}_h}\) as
\(t\in [0,T],\,x\in {\mathbb {G}}_h\), where
and for functions \(\varphi \) on \({\mathbb {R}}^d\)
with
Remark 3.1
Notice that due to the symmetry of \(\psi \) and \(\Lambda \) required in (2.6), Eq. (3.1) is invariant under the change of h to \(-h\).
Remark 3.2
Recall the definition of \(\Gamma \) introduced before Assumption 2.5. Clearly
i.e., the definition of \({\mathcal {I}}^h\), \({\mathcal {L}}^h_t \) and \({\mathcal {M}}^{h,\rho }_t \) does not change if the summation there is taken over \(\lambda \in {\mathbb {G}}\). Owing to Assumption 2.1 with \(m=0\) and the bounds on \(R^{\alpha \beta }_\lambda \), the operators \({\mathcal {L}}^h_t\) and \({\mathcal {M}}^{h,\rho }_t\) are bounded linear operators on \(\ell _2({\mathbb {G}}_h)\) such that for each \(h\ne 0\) and \(t\in [0,T]\)
for all \(\varphi \in \ell _2({\mathbb {G}}_h)\), with a constant \( N_h =N(|h|,K,d,\psi ,\Lambda )\). One can similarly show that
with a constant \(N=N(K,d,\Lambda ,\psi )\) independent of h. It is also easy to see that for every \(\phi \in L_2\) and \(\phi ^h\) defined as in (3.2) we have
with a constant \(N=N(d, \Lambda ,\psi )\) which does not depend on h; therefore
Lemma 3.1
The inequality (3.6) implies that the mapping \({\mathcal {I}}^h\) is a bounded linear operator on \(\ell _2({\mathbb {G}}_h)\). Owing to Assumption 2.4 it has an inverse \(({\mathcal {I}}^h)^{-1}\) on \(\ell _2({\mathbb {G}}_h)\), and
Proof
For \(\varphi \in \ell _2({\mathbb {G}}_h)\) and \(h \ne 0\) we have
Together with (3.6), this estimate implies that \({\mathcal {I}}^h\) is invertible and that (3.7) holds. \(\square \)
Remark 3.2 and Lemma 3.1 imply that Eq. (3.1) is an (affine) linear SDE in the Hilbert space \(\ell _2({\mathbb {G}}_h)\), and by well-known results on solvability of SDEs with Lipschitz continuous coefficients in Hilbert spaces, Eq. (3.1) has a unique \(\ell _2({\mathbb {G}}_h)\)-valued continuons solution \((U_t)_{t\in [0,T]}\), which we call an \(\ell _2\)-solution to (3.1).
Now we formulate the relationship between Eqs. (2.8) and (3.1).
Theorem 3.2
Let Assumption 2.4 hold. Then the following statements are valid.
(i) Let Assumptions 2.1 and 2.2 be satisfied with \(m=0\), and
be the unique \(V_h\)-solution of (2.8); then \((U^h_t)_{t\in [0,T]}\) is the unique \(\ell _2\)-solution of (3.1).
(ii) Let Assumption 2.1 hold with \(m=0\). Let \(\Phi \) be an \(\ell _2({\mathbb {G}}_h)\)-valued \({\mathcal {F}}_0\)-measurable random variable, and let \(F=(F_t)_{t\in [0,T]}\) and \(G^{\rho }=(G^{\rho }_t)_{[0,T]}\) be \(\ell _2({\mathbb {G}}_h)\)-valued adapted processes such that almost surely
Then Eq. (3.1) with \(\Phi \), F and \(G^{\rho }\) in place of \(\phi ^h\), \(f^h\) and \(g^{\rho ,h}\), respectively, has a unique \(\ell _2({\mathbb {G}}_h)\)-solution \(U^h=(U^h_t)_{t\in [0,T]}\). Moreover, if Assumption 2.3 also holds then
with a constant \(N=N(K,d,\kappa ,\delta ,\Lambda , \psi )\) which does not depend on h.
Proof
(i) Substituting (3.8) into Eq. (2.8), then dividing both sides of the equation by \(|h|^d\) we obtain Eq. (3.1) for \(U^h\) by simple calculation. Hence by Remark 3.2 we can see that \(U^h\) is the unique \(\ell _2({\mathbb {G}})\)-valued solution to (3.1).
To prove (ii) we use Remark 3.1 on the invertibility of \({\mathcal {I}}^h\) and a standard result on solvability of SDEs in Hilbert spaces to see that Eq. (3.1) with \(\Phi \), F and \(G^{\rho }\) has a unique \(\ell _2({\mathbb {G}})\)-valued solution \(U^h\). We claim that \(u^h_t(\cdot )=\sum _{y\in {\mathbb {G}}} U^h_t (y)\psi ^h_{y}(\cdot )\) is the \(V_h\)-valued solution of (2.8) with
and
respectively. Indeed, (3.3) yields
In the same way we have
which proves the claim. Using Remarks 2.2 and 3.1 we deduce
with a constant \(N=N(\psi ,\Lambda )\). Hence by Theorem 2.3
with \(N=N(K,T,\kappa ,d,\psi ,\Lambda ,\delta )\) independent of h, which by virtue of Remark 2.2 implies estimate (3.9). \(\square \)
4 Coefficients of the expansion
Notice that the lattice \({\mathbb {G}}_h\) and the space \(V_h\) can be “shifted” to any \(y\in {\mathbb {R}}^d\), i.e., we can consider \({\mathbb {G}}_h(y):={\mathbb {G}}_h+y\) and
Thus Eq. (2.8) for \(u^h=\sum _{x\in {\mathbb {G}}_h(y)}U(x)\psi ^h_x\) should be satisfied for \(\psi _x\), \(x\in {\mathbb {G}}_h(y)\). Correspondingly, Eq. (3.1) can be considered for all \(x\in {\mathbb {R}}^d\) instead of \(x\in {\mathbb {G}}_h\).
To determine the coefficients \((v^{(j)})_{j=1}^{k}\) in the expansion (2.9) we differentiate formally (3.1) in the parameter h, j times, for \(j=1,2,\ldots ,J\), and consider the system of SPDEs we obtain for the formal derivatives
where \(D_h\) denotes differentiation in h. To this end given an integer \(n\ge 1\) let us first investigate the operators \({\mathcal {I}}^{(n)}\), \({\mathcal {L}}^{(n)}_t\) and \({\mathcal {M}}^{(n)\rho }_t\) defined by
for \(\varphi \in C_0^{\infty }\).
Lemma 4.1
Let Assumption 2.1 hold with \(m\ge n+l+2\) for nonnegative integers l and n. Let Assumption 2.5. also hold. Then for \(\varphi \in C_0^{\infty }\) and \(t\in [0,T]\) we have
with a constant \(N=N(K,d,l, n,\Lambda ,\Psi )\) which does not depend on h.
Proof
Clearly, \({\mathcal {I}}^{(n)} =\sum _{\lambda \in \Gamma } R_{\lambda }\partial ^n_{\lambda }\varphi \), where
This shows the existence of a constant \(N=N(\Lambda , \psi ,d,n)\) such that the first estimate in (4.3) holds. To prove the second estimate we first claim the existence of a constant \(N=N(K,d,l,\Lambda ,\psi )\) such that
for
Recall the definition of \(R^{ij}_\lambda \) given in (2.11). To prove (4.5) we write \(\Phi _t(h,x)=\sum _{i=1}^3 \Phi ^{(i)}_t(h,x)\) for \(h\ne 0\) with
for
and
where \(D_{kl}:=D_kD_l\). Here we used Taylor’s formula
with \(n=1\) and \(f(h):=a^{ij}_t(x+h\lambda )\).
Note that Lemma 2.2 and (2.20) in Assumption 2.5 imply
To rewrite \(\Phi ^{(2)}_t(h,x)\) note that \(S^{ij}_{-\lambda } = -S^{ij}_{\lambda }\); indeed since \(\psi (-x)=\psi (x)\) the change of variables \(y=-z\) implies that
Furthermore, an obvious change of variables, (4.8) and Lemma 2.2 yield
This implies
Note that since \(a^{ij}_t(x)=a^{ji}_t(x)\), we deduce
for \(\partial _\lambda F \) defined by (4.4). Thus the Eqs. (4.8) and (4.9) imply
Furthermore, the definition of \(\Phi ^{(3)}_t(h,x)\) yields
Using Assumption 2.1 and Remark 2.1, this completes the proof of (4.5).
Taylor’s formula (4.6) with \(n=0\) and \(f(h):=b^i_t(x+h z )\) implies
with
Using Lemma 2.2 and computations similar to those used to prove (4.5) we deduce that
which yields
Furthermore, the definition of \(\Phi ^{(5)}(h,x)\) implies
This implies the existence of a constant \(N=N(K,d,l,\Lambda ,\psi )\) which does not depend on h such that
Finally, let
Then we have
so that
for some constant N as above.
Since \({\mathcal {L}}_t^h \varphi (x)=\Phi _t(h,x)+{\tilde{\Phi }}_t(h,x)+\Phi ^{(6)}_t(h,x)\), the inequalities (4.5), (4.11) and (4.12) imply that \({\mathcal {L}}^{(n)}_t\) satisfies the estimate in (4.3); the upper estimates of \({\mathcal {M}}^{(n),\rho }_t\) can be proved similarly. \(\square \)
For an integer \(k\ge 0\) define the operators \({\hat{L}}_t^{(k)h}\), \({\hat{M}}_t^{(k)h,\rho }\) and \({\hat{I}}^{(k)h}\) by
where \({\mathcal {L}}^{(0)}_t:={\mathcal {L}}_t\), \({\mathcal {M}}^{(0),\rho }_t:={\mathcal {M}}^{\rho }_t\), and \({\mathcal {I}}^{(0)}\) is the identity operator.
Lemma 4.2
Let Assumption 2.1 hold with \(m\ge k+l+3\) for nonnegative integers k and n. Let Assumption 2.5 also hold. Then for \(\varphi \in C_0^{\infty }\) we have
for a constant N which does not depend on h.
Proof
We obtain the estimate for \({\hat{L}}_t^{(k)h}\) by applying Taylor’s formula (4.6) to \(f(h):= \Phi ^{(i)}_t(h,x)\) for \(i=1,\ldots , 6\) defined in the proof of Lemma 4.1, and by estimating the remainder term
using the Minkowski inequality. Recall that \({\mathcal {L}}_t\varphi (x) = {\mathcal {L}}^{(0)}_t\varphi (x)\). Using Assumption 2.5 we prove that \({\mathcal {L}}_t^{(0)} \varphi (x)= \lim _{h\rightarrow 0} {\mathcal {L}}_t^h \varphi (x)\). We have \({\mathcal {L}}^{h}_t\varphi (x) = \sum _{i=1}^6 \Phi ^{(i)}_t(h,x)\) for \(h\ne 0\). The proof of Lemma 4.1 shows that \({\tilde{\Phi }}^{(i)}_t(0,x) :=\lim _{h\rightarrow 0} \Phi ^{(i)}_t(h,x)\) exist and we identify these limits. Using (4.7), (4.10) and (2.24) with \(X_{ij,kl}=a^{ij}_t(x) D_{kl}\varphi (x)\) (resp. \(X_{ij,kl}=\partial _k a^{ij}_t(x) \partial _l \varphi (x)\)) we deduce
which implies that \({\tilde{\Phi }}^{(1)}_t(0,x)+{\tilde{\Phi }}^{(2)}_t(0,x) =D_i\big (a^{ij}_t D_j\varphi \big )(x)\). The first identity in (2.23) [resp. (2.21), the second identity in (2.23) and the first identity in (2.20)] imply
This completes the identification of \({\mathcal {L}}_t\) as the limit of \({\mathcal {L}}_t^h\). Using once more the Minkowski inequality and usual estimates, we prove the upper estimates of the \(H^l\) norm of \(\hat{L}^{(k)h}_t\varphi \). The other estimates can be proved similarly. \(\square \)
Assume that Assumption 2.2 is satisfied with \(m\ge J+1\) for an integer \(J\ge 0\). A simple computation made for differentiable functions in place of the formal ones introduced in (4.1) shows the following identities
where \(\partial _z^i \varphi \) is defined iterating (4.4), while \(\phi ^h\), \(f_t^h\) and \(g_t^{h,\rho }\) are defined in (3.2). Set
Lemma 4.3
Let Assumption 2.1 holds with \(m\ge l+J+1\) for nonnegative integers J and l. Then there is a constant \(N=N(J,l,d,\psi )\) independent of h such that
Proof
Clearly, it suffices to prove the estimate for \({\hat{\phi }}^{(J)h}\), and we may assume that \(\phi \in C_0^{\infty }\). Applying Taylor’s formula (4.6) to \(f(h)=\phi ^h(x)\) for the remainder term we have
Hence by Minkowski’s inequality and the shift invariance of the Lebesgue measure we get
with a constant \(N=N(J,m,d,\psi )\) which does not depend on h. \(\square \)
Differentiating formally Eq. (3.1) with respect to h at 0 and using the definition of \({\mathcal {I}}^{(i)}\) in (4.2), we obtain the following system of SPDEs:
for \(i=1,2,\ldots ,J\), \(t\in [0,T]\) and \(x\in {\mathbb {R}}^d\), where \({\mathcal {L}}^{(0)}_t={\mathcal {L}}_t\), \({\mathcal {M}}^{(0)\rho }_t={\mathcal {M}}^{\rho }_t\), and \(v^{(0)}=u\) is the solution to (2.1)–(2.2).
Theorem 4.4
Let Assumptions 2.1 and 2.2 hold with \(m\ge J+1\) for an integer \(J\ge 1\). Let Assumptions 2.3 through 2.5 be also satisfied. Then (4.15)–(4.16) has a unique solution \((v^{(0)},\ldots ,v^{(J)})\) such that \(v^{(n)}\in {\mathbb {W}}^{m+1-n}_2(0,T) \) for every \(n=0,1,\ldots ,J\). Moreover, \(v^{(n)}\) is a \(H^{m-n}\)-valued continuous adapted process, and for every \(n=0,1,\ldots ,J\)
with a constant \(N=N(m,J,d,T,\Lambda ,\psi ,\kappa )\) independent of h, and \({\mathfrak {K}}_m\) defined in (2.5).
Proof
The proof is based on an induction argument. We can solve this system consecutively for \(i=1,2,\ldots ,J\), by noticing that for each \(i=1,2,\ldots ,k\) the equation for \(v^{(i)}\) does not contain \(v^{(n)}\) for \(n=i+1,\ldots ,J\). For \(i=1\) we have \(v^{(1)}_0 = \phi ^{(1)}\) and
i.e.,
with
By virtue of Theorem 2.1 this equation has a unique solution \(v^{(1)}\) and
Clearly, Lemma 4.1 implies
with a constant \(N=N(d,K,\Lambda ,\psi ,m)\) which does not depend on h. Hence for \(m\ge 1\)
Let \(j\ge 2\). Assume that for every \(i<j\) the equation for \(v^{(i)}\) has a unique solution such that (4.15) holds and that its equation can be written as \(v^{(i)}_0=\phi ^{(i)}\) and
with adapted processes \({\bar{f}}^{(i)}\) and \({\bar{g}}^{(i) \rho }\) taking values in \(H^{m-i-1}\) and \(H^{m-i}\) respectively, such that
with a constant \(N=N(K,J,m,d,T,\kappa ,\Lambda ,\psi )\) independent of h. Hence
Then for \(v^{(j)}\) we have
with
Note that \(|f^{(j)}_t|_{m-1-j}\le N|f_t|_{m-1}\) ; by virtue of Lemma 4.1, and by the inequalities (4.18) and (4.19) we have
where \(N=N(K,J,d,T,\kappa ,\psi ,\Lambda )\) denotes a constant which can be different on each occurrence. Consequently,
and we can get similarly
Hence (4.20) has a unique solution \(v^{(j)}\), and Theorem 2.1 implies that the estimate (4.17) holds for \(v^{(j)}\) in place of \(v^{(n)}\). This completes the induction and the proof of the theorem. \(\square \)
Recall that the norm \(|\cdot |_{0,h}\) has been defined in (2.7).
Corollary 4.5
Let Assumptions 2.1 and 2.2 hold with \(m>\frac{d}{2}+J+1\) for an integer \(J\ge 1\). Let Assumptions 2.3 through 2.5 be also satisfied. Then almost surely \(v^{(i)}\) is continuous in \((t,x)\in [0,T]\times {\mathbb {R}}^d\) for \(i\le J\), and its restriction to \({\mathbb {G}}_h\) is an adapted continuous \(\ell _2({\mathbb {G}}_h)\)-valued process. Moreover, almost surely (4.15)–(4.16) hold for all \(x\in {\mathbb {R}}^d\) and \(t\in [0,T]\), and
for some constant \(N=N(m,J,d,T,\Lambda ,\psi ,\kappa )\) independent of h.
One can obtain this corollary from Theorem 4.4 by a standard application of Sobolev’s embedding of \(H^m\) into \(C^{2}_b\) for \(m>2+d/2\) and by using the following known result; see, for example [7], Lemma 4.2.
Lemma 4.6
Let \(\varphi \in H^m\) for \(m>d/2\). Then there is a constant \(N=N(d,\Lambda )\) such that
where I denotes the Sobolev embedding operator from \(H^m\) into \(C_b({\mathbb {R}}^d)\).
5 Proof of theorem 2.4
Define a random field \(r^h\) by
where \((v^{(1)},\ldots ,v^{(J)})\) is the solution of (4.15) and (4.16).
Theorem 5.1
Let Assumptions 2.1 and 2.2 hold with \(m>\frac{d}{2}+2J+2\) for an integer \(J\ge 0\). Let Assumptions 2.3 through 2.5 be also satisfied. Then \(r^h\) is an \(\ell _2({\mathbb {G}}_h)\)-solution of the equation
where \(F^h\) and \(G^h\) are adapted \(\ell _2({{\mathbb {G}}}_h)\)-valued such that for all \(h\ne 0\) with \(|h|\le 1\)
where \(N=N(m, K,J,d,T,\kappa ,\Lambda ,\psi )\) is a constant which does not depend on h.
Proof
Using (5.1), the identity \(u^h_t(x)= U^h_t(x)\) for \(x\in {{\mathbb {G}}}_h\) which is deduced from Assumption 2.6 and Eq. (3.1), we deduce that for \(x\in {{\mathbb {G}}}_h\),
Taking into account Corollary 4.5, in the definition of \(d v^{(i)}_t(x) \) in (4.15) we set
where \(B(i)_t\) (resp. \(\sigma (i)^\rho _t\)) contains the operators \({{\mathcal {L}}}^{(j)}\) (resp. \({{\mathcal {M}}}^{(j)\rho }_t\)) for \(0\le j\le i\) while \(F(i)_t\) (resp. \(G(i)^\rho _t\)) contains all the free terms \(f^{(j)}_t\) (resp. \(g^{(j)\rho }_t\)) for \(1\le j\le i\). We at first focus on the deterministic integrals. Using the recursive definition of the processes \(v^{(i)}\) in (4.15), it is easy to see that
In the sequel, to ease notations we will not mention the space variable x. Using the expansion of \({{\mathcal {L}}}_t^h\), \({{\mathcal {I}}}^h\) and the definitions of \(\hat{L}_t^{(J),h}\) and \(\hat{I}^{(J),h}\) in (4.13), the expansion of \(f^h_t\) and the definition of \(\hat{f}^{(J)h}_t\) given in (4.14) together with the definition of \(d v^{(i)}_t \) in (5.6), we deduce
where
Equation (4.15) implies
Using the recursive Eq. (5.7) we deduce that for every \(h>0\) and \(t\in [0,T]\),
A similar computation based on (5.8) implies
In \({{\mathcal {T}}}^h_t(2)\) all terms have a common factor \(h^{J+1}\). We prove an upper estimate of
for \(0\le i,j\le J\). Let I denote the Sobolev embedding operator from \(H^k\) to \(C_b({\mathbb {R}}^d)\) for \(k>d/2\). Lemma 4.6, inequalities (4.3) and (4.17) imply that for \(k>d/2\),
where the constant N does not depend on h and changes from one upper estimate to the next. Clearly, for \(0\le i,j\le J\) with \(i+j\ge J+1\), we have \(i+j+k+1>2J+1+\frac{d}{2}\). Similar computations prove that for \(i,j\in \{0,\ldots , J\}\) with \(i+j\ge J+1\) and \(k>\frac{d}{2}\),
These upper estimates imply the existence of some constant N independent of h such that for \(|h|\in (0,1]\) and \(k>\frac{d}{2}\)
We next find an upper estimate of the \(|\cdot |_{0,h}\) norm of both terms in \({{\mathcal {T}}}_t^h(3)\). Using Lemmas 4.6, 4.2 and (4.17) we deduce that for \(k>\frac{d}{2}\)
where N is a constant independent of h with \(|h|\in (0,1]\). Furthermore, similar computations yield for \(k>\frac{d}{2}\) and \(|h|\in (0,1]\)
Hence we deduce the existence of a constant N independent of h such that for \(|h|\in (0,1]\),
where \(k>\frac{d}{2}\).
We next compute an upper estimate of the \(|\cdot |_{0,h}\) norm of \({{\mathcal {T}}}^h_t(5)\). All terms have a common factor \(h^{(J+1)}\). Recall that \({{\mathcal {I}}}^{(0)}=Id\). The induction Eq. (5.8) shows that \(F(i)_t\) is a linear combination of terms of the form \(\Phi (i)_t:=\big ( {{\mathcal {I}}}^{(a_1)} \big )^{k_1} \ldots \big ( {{\mathcal {I}}}^{(a_i)} \big )^{k_i} f_t\) for \(a_p, k_p\in \{0,\ldots , i\}\) for \(1\le p\le i\) with \(\sum _{p=1}^i a_p k_p=i\), and of terms of the form \(\Psi (i)_t:=\big ( {{\mathcal {I}}}^{(b_1)} \big )^{l_1} \ldots \big ( {{\mathcal {I}}}^{(b_{i-j})} \big )^{l_{i-j}} f^{(j)}_t\) for \(1\le j\le i\), \(b_p, l_p\in \{0,\ldots , i-j\}\) for \(1\le p\le i-j\) with \(\sum _{p=1}^{i-j} b_p l_p+j=i\). Using Lemmas 4.6 and 4.1 we obtain for \(k>\frac{d}{2}\), \(i,j=1, \ldots J\)
A similar computation yields
These upper estimates imply that for \(k>\frac{d}{2}\), there exists some constant N independent on h such that for \(|h|\in (0,1)\)
We finally prove an upper estimate of the \(|\cdot |_{0,h}\)-norm of both terms in \({{\mathcal {T}}}_t^h(6)\). Using Lemmas 4.6 and 4.3, we obtain for \(k>\frac{d}{2}\),
where N is a constant which does not depend on h. Furthermore, Lemmas 4.6 and 4.2 yield for \(k>\frac{d}{2}\) and \(|h|\in (0,1]\),
for some constant N independent of h. Hence we deduce that for some constant N which does not depend on h and \(k>\frac{d}{2}\), we have for \(|h|\in (0,1]\)
Similar computations can be made for the coefficients of the stochastic integrals. The upper bounds of the corresponding upper estimates in (5.11) and (5.12) are still valid because the operators \({{\mathcal {M}}}^\rho _t\) are first order operators while the operator \({{\mathcal {L}}}_t\) is a second order one. This implies that all operators \({{\mathcal {M}}}^{h,\rho }_t\), \({{\mathcal {M}}}^{(i)\rho }_t\) and \(\hat{M}^{(J)h}_t\) contain less derivatives than the corresponding ones deduced from \({{\mathcal {L}}}_t\).
Using the expansion (5.5), the upper estimates (5.9)–(5.14) for the coefficients of the deterministic and stochastic integrals, we conclude the proof. \(\square \)
We now complete the proof of our main result.
Proof of Theorem 2.4
By virtue of Theorems 3.2 and 5.1 we have for \(|h|\in (0,1]\)
Using Remark 3.1 we have \(U^{-h}_t=U^h_t\) for \(t\in [0,T]\) a.s. Hence from the expansion (2.9) we obtain that \(v^{(j)}=-v^{(j)}\) for odd j, which completes the proof of Theorem 2.4. \(\square \)
6 Some examples of finite elements
In this section we propose three examples of finite elements which satisfy Assumptions 2.4, 2.5 and 2.6.
6.1 Linear finite elements in dimension 1
Consider the following classical linear finite elements on \({\mathbb {R}}\) defined as follows:
Let \(\Lambda = \{-1, 0, 1\}\); clearly \(\psi \) and \(\Lambda \) satisfy the symmetry condition (2.6). Recall that \(\Gamma \) denotes the set of elements \(\lambda \in {{\mathbb {G}}}\) such that the intersection of the support of \(\psi _\lambda :=\psi ^1_\lambda \) and of the support of \(\psi \) has a positive Lebesgue measure. Then \(\Gamma =\{ -1,0,1\}\), the function \(\psi \) is clearly non negative, \(\int _{{\mathbb {R}}} \psi (x) dx =1\), \(\psi (x)=0\) for \(x\in {\mathbb {Z}}\setminus \{0\}\) and Assumption 2.6 clearly holds.
Simple computations show that
Hence \(\sum _{\lambda \in \Gamma }R_\lambda =1\). Furthermore, given any \(z=(z_n)\in \ell _2({\mathbb {Z}})\) we have using the Cauchy-Schwarz inequality:
Hence Assumption 2.4 is satisfied. Easy computations show that for \(\epsilon \in \{-1,1\}\) we have
Hence \(\sum _{\lambda \in \Gamma }R^{11}_\lambda =0\), which completes the proof of (2.20). Furthermore, \(\sum _{\lambda \in \Gamma } \lambda R^1_\lambda =1\), which proves (2.21) while \(\sum _{\lambda \in \Gamma } \lambda ^2 R^{11}_\lambda = 2\), which proves (2.22).
Finally, we have for \(\epsilon \in \{-1,1\}\)
This clearly implies \(\sum _{\lambda \in \Gamma } Q^{11,11}_\lambda =0\) and \(\sum _{\lambda \in \Gamma } {\tilde{Q}}^{11}_\lambda =0\), which completes the proof of (2.23); therefore, Assumption 2.5 is satisfied.
The following example is an extension of the previous one to any dimension.
6.2 A general example
Consider the following finite elements on \({\mathbb {R}}^d\) defined as follows: let \(\psi \) be defined on \({\mathbb {R}}^d\) by \(\psi (x)=0\) if \(x\notin (-1,+1]^d\) and
The function \(\psi \) is clearly non negative and \( \int _{{\mathbb {R}}^d} \psi (x) dx =1\). Let \(\Lambda = \{ 0, \, \epsilon _k e_k, \, \epsilon _k\in \{-1,+1\}, \, k=1,\ldots ,d\}\). Then \(\psi \) and \(\Lambda \) satisfy the symmetry condition (2.6). Furthermore, \(\psi (x)=0\) for \(x\in {\mathbb {Z}}^d\setminus \{0\}\); Assumption 2.6 clearly holds.
These finite elements also satisfy all requirements in Assumptions 2.4–2.5. Even if these finite elements are quite classical in numerical analysis, we were not able to find a proof of these statements in the literature. To make the paper self-contained the corresponding easy but tedious computations are provided in an Appendix.
6.3 Linear finite elements on triangles in the plane
We suppose that \(d=2\) and want to check that the following finite elements satisfy Assumptions 2.4–2.6. For \(i=1,\ldots , 6\), let \(\tau _i\) be the triangles defined as follows:
Let \(\psi \) be the function defined by:
It is easy to see that the function \(\psi \) is non negative and that \(\int _{{\mathbb {R}}^2} \psi (x) dx=1\). Set \(\Lambda = \{ 0, e_1, -e_1, e_2, -e_2\}\); the function \(\psi \) and the set \(\Lambda \) fulfill the symmetry condition (2.6). Furthermore, \(\Gamma = \big \{ \epsilon _1 e_1 + \epsilon _2 e_2 : (\epsilon _1, \epsilon _2)\in \{-1,0,1\}^2 , \; \epsilon _1\epsilon _2\in \{0,1\} \big \} \). Hence \(\psi \) satisfies Assumption 2.6.
For \(\mathbf{i}=(i_1,i_2)\in {\mathbb {Z}}^2\), let \(\psi _\mathbf{i}\) the function defined by
For \(\gamma =1, 2,\ldots , 6\), we denote by \(\tau _\gamma (\mathbf{i}) =\big \{ (x_1,x_2) : (x_1,x_2) - \mathbf{i} \in \tau _\gamma \big \}\). Then
Easy computations show that for \(\mathbf{i}\in {\mathbb {Z}}^2\), and \(\mathbf{k} \in \{ \mathbf{i}+\lambda : \lambda \in \Gamma \}\)
and \((\psi _\mathbf{i} , \psi _\mathbf{j}) =0\) otherwise. Thus
which proves the first identity in (2.20). First we check that given any \(\alpha \in (0,1)\) by Cauchy–Schwarz inequality we have some positive constants \(C_1\) and \(C_2\) such that
for all \((U_\mathbf{i})\in \ell _2({\mathbb {Z}}^2)\). Hence, by taking \(\alpha \in (0,1)\) such that \(1-C_1 \alpha -C_2 \alpha ^2 \ge \frac{1}{2}\), we see that Assumption 2.4 is satisfied.
We next check the compatibility conditions in Assumption 2.5. Easy computations prove that for \(k=1,2\) and \(l\in \{1,2\}\) with \(l\ne k\), \(\epsilon _k, \epsilon _l \in \{-1,1\}\) we have
while
Hence for any \(k,l=1,2\) and \(l\ne k\) we have
This completes the proof of equation \(\sum _{\lambda \in \Gamma } R^{ij}_\lambda =0\) and hence of Eq. (2.20). Furthermore, given \(k,l=1,2\) with \(k\ne l\) we have for \(\alpha =k\) or \(\alpha =l\):
The last identities come from the fact that \((D_k\psi , D_l\psi _{\epsilon e_k})\) , \((D_k\psi , D_l\psi _{\epsilon e_l})\) or \((D_k\psi , D_l\psi _{\epsilon (e_1+e_2)}\) agree for \(\epsilon =-1\) and \(\epsilon =1\). This completes the proof of Eq. (2.22).
We check the third compatibility condition. Using Lemma 2.2 we deduce for \(k,l=1,2\) with \(k\ne l\) and \(\epsilon \in \{-1,+1\}\)
Therefore, using Lemma 2.2 we deduce that
This completes the proof of Eq. (2.21).
Let us check the first identity in (2.23). Recall that
and suppose at first that \(i=j\). Then we have for \(k\ne i\), \(\alpha \ne i\), \(k\ne l\) and \(\epsilon \in \{-1,+1\}\)
Suppose that \(i\ne j\); then for \(k\ne l\) and \(\epsilon \in \{-1,+1\}\) we have
The above equalities prove \(\sum _{\lambda \in \Gamma } Q^{ij,kl}_\lambda =0\) for any choice of \(i,j,k,l=1,2\). Hence the first identity in (2.23) is satisfied.
We finally check the second identity in (2.23). Recall that \({\tilde{Q}}_\lambda ^{i,k}=\int _{{\mathbb {R}}^2} z_k D_i\psi _\lambda (z) \psi (z) dz\). For \(i=k\in \{1,2\}\), \(j\in \{1,2\}\) with \(i\ne j\) and \(\epsilon \in \{-1,+1\}\) we have
Hence \(\sum _{\lambda \in \Gamma } {\tilde{Q}}_\lambda ^{i,i}=0\). Let \(i\ne k\); then for \(\epsilon \in \{-1,+1\}\) we have
Hence \(\sum _{\lambda \in \Gamma } {\tilde{Q}}_\lambda ^{i,k}=0\) for any choice of \(i,k=1,2\), which concludes the proof of (2.23) Therefore, the function \(\psi \) defined by (6.4) satisfies all Assumptions 2.4–2.6.
References
Asadzadeh, M., Schatz, A., Wendland, W.: A new approach to Richardson extrapolation in the finite element method for second order elliptic problems. Math. Comput. 78(268), 1951–1973 (2009)
Blum, H., Lin, Q., Rannacher, R.: Asymptotic error expansion and Richardson extrapolation for linear finite elements. Numer. Math. 49, 11–37 (1986)
Brzeźniak, Z., Carelli, E., Prohl, A.: Finite-element-based discretizations of the incompressible Navier–Stokes equations with multiplicative random forcing. IMA J. Numer. Anal. 33–3, 771–824 (2013)
Carelli, E., Prohl, A.: Rates of convergence for discretizations of the stochastic incompressible Navier–Stokes equations. SIAM J. Numer. Anal. 50–5, 2467–2496 (2012)
Davie, A.J., Gaines, J.G.: Convergence of numerical schemes for the solution of parabolic stochastic partial differential equations. Math. Comput. 70, 121–134 (2001)
Gerencsér, M., Gyöngy, I.: Localization errors in solving stochastic partial differential equations in the whole space. Math. Comput. 86(307), 2373–2397 (2017)
Gyöngy, I., Krylov, N.V.: Accelerated finite difference schemes for linear stochastic partial differential equations in the whole space. SIAM J. Math. Anal. 43–5, 2275–2296 (2010)
Hall, E.J.: Accelerated spatial approximations for time discretized stochastic partial differential equations. SIAM J. Math. Anal. 44, 3162–3185 (2012)
Hall, E.J.: Higher order spatial approximations for degenerate parabolic stochastic partial differential equations. SIAM J. Math. Anal. 45(4), 2071–2098 (2013)
Hausenblas, E.: Finite element approximation of stochastic partial differential equations driven by poisson random measures of jump type. SIAM J. Numer. Anal. 46–1, 437–471 (2008)
Kovács, M., Larsson, S., Lindgren, F.: Strong convergence of the finite element method with truncated noise for semilinear parabolic stochastic equations with additive noise. Numer. Algorithms 53, 309–320 (2010)
Kovács, M., Printems, J.: Strong order of convergence of a fully discrete approximation of a linear stochastic Volterra type evolution equation. Math. Comput. 83, 2325–2346 (2014)
Krylov, N.V., Rozovskii, B.L.: On Cauchy problem for linear stochastic partial differential equations. Math. USSR Izvestija 11–4, 1267–1284 (1977)
Lemaire, V., Pagès, G.: Multi-level Richardson–Romberg extrapolation. Bernoulli 23–4A, 2643–2692 (2017)
Malliavin, P., Thalmaier, A.: Numerical error for SDE: Asymptotic expansion and hyperdistributions. C. R. Math. Acad. Sci. Paris 336, 851–856 (2003)
Rannacher, R.: Richardson extrapolation for a mixed finite element approximation of a plain blending problem (Gesellschaft für angwandte Mathematik and Mechanik, Wissenschaftliche Jahrestagung, West Germany, Apr. 1–4, 1986) Zeitschrift für andgewandte Mathematik und Mechanik, vol. 67–5 (1987)
Rannacher, R.: Richardson Extrapolation with Finite Elements, Numerical Techniques in Continuous Mechanics, pp. 90–101. Springer, Berlin (1987)
Richardson, L.F.: The approximate arithmetical solution by finite differences of physical problems including differential equations, with an application to the stresses in a masonry dam. Philos. Trans. R. Soc. A 210, 307–357 (1911)
Rozovskii, B.L.: Stochastic Evolution Systems. Linear Theory and Applications to Nonlinear Filtering. Kluwer, Dordrecht (1990)
Sidi, A.: Practical Extrapolation Methods. Cambridge University Press, Cambridge (2003)
Talay, D., Tubaro, L.: Expansion of the global error for numerical schemes solving stochastic differential equations. Stoch. Anal. Appl. 8, 94–120 (1990)
Walsh, J.B.: Finite element methods for parabolic stochastic PDE’s. Potential Anal. 23(1), 1–43 (2005)
Yan, Y.: Galerkin finite element methods for stochastic parabolic partial differential equations. SIAM J. Numer. Anal. 43(4), 1363–1384 (2005)
Acknowledgements
This work started while István Gyöngy was an invited professor at the University Paris 1 Panthéon Sorbonne. It was completed when Annie Millet was invited by the University of Edinburgh. Both authors want to thank the University Paris 1, the Edinburgh Mathematical Society and the Royal Society of Edinburgh for their financial support. The authors want to thank the anonymous referees for their careful reading and helpful remarks.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
The aim of this section is to prove that the example described in Sect. 6.2 satisfies Assumptions 2.4 and 2.5.
For \(k=1,\ldots , d\), let \(e_k\in {\mathbb {Z}}^d\) denote the k-th unit vector of \({\mathbb {R}}^d\); then \({\mathbb {G}}={\mathbb {Z}}^d\) and
For fixed \(k=1, \ldots , d\) (resp. \(k\ne l \in \{1, \ldots , d\}\)) let
Note that in the particular case \(d=1\), the functions \(\psi \) gives rise to the classical linear finite elements. Then for \(\mathbf{i} \in {{\mathbb {Z}}}^d\), we have for \(k=0, 1, \ldots , d\):
Furthermore, given \(k=0, 1, \ldots , d\), there are \(2^k\) elements \(\mathbf{i}\in {\mathbb {Z}}^d\) such that \(\sum _{l=1}^d |i_l|=k\). Therefore, we deduce
which yields the first compatibility conditon in (2.20).
We at first check that Assumption 2.4 holds true, that is
for some \(\delta >0\). For \(U\in \ell _2({{\mathbb {Z}}}^d)\) and \(k=1, \ldots , d,\) let \(T_k U=U_{e_k}\), where \(e_k\) denotes the k-th vector of the canonical basis.
For \(U\in \ell _2({\mathbb {Z}}^d)\) we have
Given \(\alpha \in (0,1)\) if we let
restricting the above integral on the set \([0,\alpha ]^d\), expanding the square and using the Cauchy Schwarz inequality we deduce the existence of some constants \(C(\gamma _1,\gamma _2, \gamma _3)\) defined for \(\gamma _i\in \{0,1, \ldots , d\}\) such that
where \(C_l\) are some positive constants. Choosing \(\alpha \) small enough, we have \( \big | \sum _\mathbf{i} U_\mathbf{i} \psi _\mathbf{i}\big |_{L^2}^2 \ge \frac{\alpha ^d}{2} |U|_{\ell _2({{\mathbb {Z}}}^d)}^2\), which implies the invertibility Assumption 2.4.
We now prove that the compatibility Assumption 2.5 holds true. For \(l=1, \ldots , d\), \(n=0, \ldots , d-1\):
For \(n=1, \ldots , d-1\) and \(k_1<k_2<\cdots < k_n\) with \(k_r\in {{\mathcal {I}}}(l)\), where \({{\mathcal {I}}}(l)\) is defined in (7.1), let
Then \(| \Gamma _l (k_1, \ldots , k_n)|=2^n\) while \(| \Gamma _l (l;k_1, \ldots , k_n)|=2^{n+1}\). For \(l=1, \ldots , d\), the identities (7.3) and (7.4) imply
This proves the second identity in (2.20) when \(i=j\). Furthermore, (7.3) implies
Furthermore, given \(k\ne l \in \{1, \ldots , d\}\),
Indeed, for \(n=1, \ldots , d-1\), \(k_1<\cdots < k_n \) where \(k_r\in {{\mathcal {I}}}(l)\) and at least one of the indices \(k_r\) is equal to k for \(r=1, \ldots ,n\), given \(\lambda \in \Gamma _l(k_1, \ldots , k_n)\) we have using (7.3) and (7.4)
This proves the second identity in (2.22) when both derivatives agree.
Also note that for \(k\ne l \in \{1, \ldots , d\}\) we have \(\sum _{\lambda \in \Gamma } R^{kl}_\lambda =0\). Indeed, for \(\lambda \) as above
while \(R^{kl}_\lambda =0\) for other choices of \(\lambda \in \Gamma \).
We now study the case of mixed derivatives. Given \(k\ne l \in \{1, \ldots , d\}\) recall that \({{\mathcal {I}}}(k,l)=\{ 1, \ldots , d\} \setminus \{k,l\}\). Then for \(k\ne l \in \{1, \ldots , d\}\) and \(\mathbf{i} \in {\mathbb {Z}}^d\) we have for \(n=0, \ldots , d-2\)
For \(n=1, \ldots , d-2\) and \(k_1<\cdots <k_n\) with \(k_r\in {{\mathcal {I}}}(k,l)\) for \(r=1,\ldots , n\), set
For \(n=0\) there is no such family of indices \(k_1<\cdots <k_n\) and we let \(\Gamma _{k,l}(\emptyset )=\{0\}\). Thus for \(n=0, \ldots , d-2\), \(| \Lambda _{k,l}(k_1,\ldots , k_n)|=2^n\). Using the identities (7.5)–(7.7) we deduce
This completes the proof of the second identity in (2.20) when \(i\ne j\), and hence (2.20) holds true. Furthermore, the identities (7.6) and (7.7) imply for \(i\ne j \in \{1, \ldots , d\}\) and \(\{i,j\}=\{k,l\}\)
Equation (7.5) proves that \((D_k \psi , D_l \psi _\lambda )=0\) if \(|\lambda _k \lambda _l|\ne 1\). Hence using (7.8) we deduce that for any \(r=1, \ldots , d\),
Let \(r\in {{\mathcal {I}}}(k,l)\) and for \(n=1, \ldots , d-3\), let \(k_1<\cdots < k_n\) be such that \(k_j\in \{1, \ldots , d\} \setminus \{k,l,r\}\) and \(\lambda = \sum _{j=1}^n \epsilon _{k_j} e_{k_j}\) for \(\epsilon _{k_j}\in \{-1,1\}\), \(j=1, \ldots , n\). Then for any choice of \(\epsilon _k\) and \(\epsilon _l\) in \(\{-1,1\}\) the equalities (7.6) and (7.7) imply that
This clearly yields that for \(r\in {{\mathcal {I}}}(k,l)\) we have
Finally, given \(n=2, \ldots , d\) and \(k_1< \cdots <k_n\) where the terms \(k_j\in {{\mathcal {I}}}(k,l)\), then given any choice of \(\epsilon _k\) and \(\epsilon _l\) in \(\{-1,1\}\), the value of \((D_k \psi , D_l \psi _{\epsilon _k e_k + \epsilon _l e_l + \lambda })\) does not depend on the value of \(\lambda \in \Gamma _{k,l}(k_1,\ldots , k_n)\). Therefore, if we fix \(r_1\ne r_2\) in the set \({{\mathcal {I}}}(k,l)\), for fixed n there are as many choices of indices \(k_1<\cdots <k_n\) such that \(\epsilon _{r_1} \epsilon _{r_2}=1\) that of such indices such that \(\epsilon _{r_1} \epsilon _{r_2}=-1\). This proves
which completes the proof of the first identity in (2.22) for mixed derivatives; hence (2.22) holds true.
We now check the compatibility condition (2.21). Fix \(j\in \{1, \ldots , d\}\); then
while
For \(n=1, \ldots , d-1\) and \(k_1<\cdots <k_n\) where the indexes \(k_r\), \(r=1, \ldots , n\) are different from j we have for any \(\lambda \in \Gamma _j(k_1,\ldots , k_n)\)
while
and
Note that the number of terms \( (D_j\psi , \psi _{\epsilon _l e_l+\lambda }) \) with \(\epsilon _l=-1\) or \(\epsilon _l =+1\) is equal to \({d-1 \atopwithdelims ()n} 2^n\). Therefore, the identities (7.9)–(7.13) imply that for any \(j=1, \ldots , d\) we have
This proves (2.21) when \(i=k\).
Let \(k\ne j \in \{1,\ldots , d\}\) and given \(n=1,\ldots , d-1\) let \(k_1<\cdots < k_n\) be indices that belong to \({{\mathcal {I}}}(j)\) such that one of the indices \(k_r, r=1,\ldots , n\) is equal to k. Given any \(\lambda \in \Gamma _j(k_1,\ldots , k_n)\) we deduce that
This completes the proof of the identity (2.21).
In order to complete the proof of the validity of Assumption 2.5, it remains to check that the identities in (2.23) hold true. Recall that for \(\lambda \in \Gamma \) and \(i,j,k,l\in \{1,\ldots ,d\}\) we have
For \(p=1,\ldots , 4\), \(n=1,\ldots , d-p\) and \(i_1,\ldots , i_p \in \{1, \ldots ,d\}\) with \(i_1,\ldots , i_p\) pairwise different let
and \( {\mathcal {I}}_0(i_1,\ldots , i_p)=\{0\}\).
First suppose that \(i=j\).
First let \(k=l=i\); then for \(n=0,\ldots , d-1\) and \( \mu \in {\mathcal {I}}_n(i)\) we have
Let \(k=l\) with \(k\ne i\); then then for \(n=0,\ldots , d-1\) and \(\mu \in {\mathcal {I}}_n(i)\) we have
Let \(l=i\) and \(k\ne i\); then for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1,+1\} \) and \( \mu \in {\mathcal {I}}_n(i,k)\) we have
A similar result holds for \(k=i\) and \(l\ne i\). Furthermore, \(Q_\lambda ^{ii,ki}=0\) if \(\lambda \) is not equal to \(\mu + \epsilon e_i + e_k\) or \(\mu + \epsilon e_i - e_k\) with \(\mu \in {\mathcal {I}}_n(i,k)\) for some n.
Let \(k\ne l\) with \(k\ne i\) and \(l\ne i\); then for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1,+1\}\) and \(\mu \in {\mathcal {I}}_n(k,l)\) we have
while \(Q_\lambda ^{ii,kl}=0\) if \(\lambda \) is not equal to \(\mu + \epsilon e_i + e_k\) or \(\mu + \epsilon e_i - e_k\) with \(\mu \in {\mathcal {I}}_n(i,k)\) for some n.
We now suppose that \(i\ne j\).
First suppose that \(k=i\) and \(l=j\); then for \(n=0,\ldots , d-1\) and \(\mu \in {\mathcal {I}}_n(i)\) we have
Let \(k=l=i\); then for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1+1\}\) and \(\mu \in {\mathcal {I}}_n(i,j)\) we have
while \(Q_\lambda ^{ij,ii}=0\) is \(\lambda \) is not equal to \(\mu +\epsilon e_i +e_j\) or \(\mu + \epsilon e_i-e_j\) where \(\mu \in {\mathcal {I}}_n(i,j)\) for some n. A similar result holds exchanging i and j for \(k=l=j\).
Let \(k=l\) with \(k\not \in \{i,j\}\) and \(l\not \in \{i,j\}\); then for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1,+1\}\) and \(\mu \in {\mathcal {I}}_n(i,j)\) we have
while \(Q_\lambda ^{ij,kk}=0\) if \(\lambda \) is not equal to \(\mu +\epsilon e_i +e_j\) or \(\mu + \epsilon e_i-e_j\) with \(\mu \in {\mathcal {I}}_n(i,j)\) for some n.
Let \(l=i\) and \(k\not \in \{i,j\}\); then for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1+1\}\) and \(\mu \in {\mathcal {I}}_n(i,k)\) we have
while \(Q_\lambda ^{ij,ki}=0\) is \(\lambda \) is not equal to \(\mu +\epsilon e_i +e_k\) or \(\mu + \epsilon e_i-e_k\) where \(\mu \in {\mathcal {I}}_n(i,k)\) for some n. A similar result holds exchanging i and j for \(k=l=j\).
Finally, let \(k\ne l\) with \(k\not \in \{i,j\}\) and \(l\not \in \{i,j\}\); then for \(n=0,\ldots , d-4\), \(\epsilon _i, \epsilon _j, \epsilon _k \in \{-1+1\}\) and \(\mu \in {\mathcal {I}}_n(i,j,k,l)\) we have
while \(Q_\lambda ^{ij,kl}=0\) is \(\lambda \) is not equal to \(\mu + \epsilon _i e_i +\epsilon _j e_j + \epsilon _k e_k +e_l\) or \(\mu + \epsilon _i e_i +\epsilon _j e_j + \epsilon _k e_k -e_l\) where \(\mu \in {\mathcal {I}}_n(i,j,k,l)\) for some n. These computations complete the proof of the first identity in (2.23). Recall that for \(i,k\in \{1,\ldots , d\}\) and \(\lambda \in \Gamma \) we let
Let \(k=i\); for \(n=0,\ldots , d-1\) and \(\mu \in {\mathcal {I}}_n(i)\) we have
Let \(k\ne i\); for \(n=0,\ldots , d-2\), \(\epsilon \in \{-1,0, +1\}\) and \(\mu \in {\mathcal {I}}_n(i,k)\) we have
while \({\tilde{Q}}_\lambda ^{i,k}=0\) if \(\lambda \) is not equal to \(\mu +\epsilon e_i +e_k\) or \(\mu + \epsilon e_i-e_k\) where \(\mu \in {\mathcal {I}}_n(i,k)\) for some n. This completes the proof of the second identity in (2.23); therefore Assumption 2.5 is satisfied for these finite elements. This completes the verification of the validity of Assumptions 2.4–2.5 for the function \(\psi \) defined by (6.2).
Rights and permissions
About this article
Cite this article
Gyöngy, I., Millet, A. Accelerated finite elements schemes for parabolic stochastic partial differential equations. Stoch PDE: Anal Comp 8, 580–624 (2020). https://doi.org/10.1007/s40072-019-00154-6
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40072-019-00154-6