dz02sinum

8/2/2019 dz02sinum

1/25

NUMERICAL APPROXIMATION OF SOME LINEAR STOCHASTIC

PARTIAL DIFFERENTIAL EQUATIONS DRIVEN BY SPECIAL

ADDITIVE NOISES

QIANG DU AND TIANYU ZHANG

SIAM J. NUMER. ANAL. c 2002 Society for Industrial and Applied MathematicsVol. 40, No. 4, pp. 14211445

Abstract. This paper is concerned with the numerical approximation of some linear stochasticpartial differential equations with additive noises. A special representation of the noise is consid-ered, and it is compared with general representations of noises in the infinite dimensional setting.Convergence analysis and error estimates are presented for the numerical solution based on the stan-dard finite difference and finite element methods. The effects of the noises on the accuracy of theapproximations are illustrated. Results of the numerical experiments are provided.

Key words. stochastic partial differential equation, additive noise, finite difference method,finite element method, convergence, error estimate

AMS subject classifications. 65M65, 65C30, 35R60, 60H15

PII. S0036142901387956

1. Introduction. In recent years, it has been increasingly acceptable to adoptSDE models as an essential component in the analysis of complex phenomena such aswave propagation [19], climate change [22], turbulence [21, 24], and phase transition[9, 16, 18]. The initial value and boundary value problems of stochastic partial differ-ential equations (SPDEs) have been studied theoretically in, for example, [5, 6, 8, 10,33]. Various numerical methods and approximation schemes for SDEs have also beendeveloped, analyzed, and tested [1, 2, 4, 7, 12, 13, 14, 15, 20, 25, 27, 29, 28, 31, 34, 35].

For a given physical system, many different stochastic perturbations may be con-sidered. Generically speaking, noise may enter the physical system either as temporalfluctuations of internal degrees of freedom or as random variations of some externalcontrol parameters; internal randomness often reflects itself in additive noise terms,while external fluctuations gives rise to multiplicative noise terms [18]. The main aimof this paper is to study the properties of some standard numerical approximationsto the linear SPDEs for the random field u = u(x, t) driven by an additive noise:

du = Audt + dW, x , t > 0.(1.1)

Here, is a bounded spatial domain and A is a linear second order elliptic operatorwith deterministic coefficients, which is defined on a space of functions satisfyingcertain boundary conditions. W represents an infinite dimensional Brownian motion.We also consider the related time-independent equation

Au = g + W, x ,(1.2)

Received by the editors April 13, 2001; accepted for publication (in revised form) April 25, 2002;published electronically October 23, 2002. This work was partially supported by the State MajorBasic Research Project G199903280 and by NSF grant DMS-0196522.

http://www.siam.org/journals/sinum/40-4/38795.htmlDepartment of Mathematics, Penn State University, University Park, PA 16802, and Department

of Mathematics, Hong Kong University of Science and Technology, Hong Kong ([email protected]).Department of Mathematics, Hong Kong University of Science and Technology, Hong Kong.

Current address: Department of Mathematics, University of Minnesota, Minneapolis, MN 55455([email protected]).

1421

8/2/2019 dz02sinum

2/25

1422 QIANG DU AND TIANYU ZHANG

where g is a given deterministic function and W denote a one-parameter family noise.The additive noises may appear in various forms, ranging from the space time whitenoise to colored noises generated by some infinite dimensional Brownian motion witha prescribed covariance operator [6, 28]. Once the equation is reformulated into a

weak form [5], the usual Galerkin finite element methods can be constructed andalso analyzed using standard techniques. A priori error estimates of the numericalsolution depend on the regularity of the solutions of the original SPDE. Such regularityresults are often much harder to establish than their deterministic counterpart [5, 33].In fact, if dW corresponds to the Brownian white noise, then the regularity estimatesare usually very weak, and they lead to very low order error estimates [1, 7, 13].On the other hand, if the noise is more regular, then it becomes possible to gethigher order of error estimates for the numerical solution. In recent years, studies ofmodels with colored noises and their numerical approximation have started to receivemore attention; see [28] for an example of physical application and the recent works[26, 14] for works related to stochastic ordinary differential equations (SODEs) andthe time discretization. In the present work, we provide the connections between thediscrete realizations of noises in different formulations of some SPDEs. Moreover, we

illustrate how the error analysis of the standard finite element and finite differenceapproximations depends on the noises used in the model and the approximation. Inorder to present a simple analysis, in this paper we focus on the case = (0 , 1)and Au = uxx bu with the homogeneous Dirichlet boundary condition and b being adeterministic coefficient, though much of our results can be readily extended to higherspatial dimensions and more general second order elliptic operators. For most of thediscussion, we also try to present our results in simple finite element terminology thatis familiar to people working on the numerical approximations of deterministic PDEsso that it is easy to be understood even for readers who are not necessarily expertson SDEs.

The paper is organized as follows. We first describe the various forms of thenoises and their discrete representations. Next, we discuss some convergence resultsfor standard finite element and finite difference approximations. The models used are

one dimensional linear stochastic elliptic and parabolic equations, and the results areestablished for noises given in general forms, which include the spatial or space timewhite noises as special cases. Then numerical results are presented to support thetheoretical analysis. Finally, some concluding remarks are given. The details of theproofs are provided in the appendix.

2. The representation of random noises. To study the accuracy of the dis-crete approximations, it is useful to first consider the properties of the noises whichdrive the stochastic equations and the discrete representations of the noises.

Following [1], we regularize the noise through discretization. Let {xi = ih}n0 bea partition of [0, 1] with h = 1/n. We begin with W(x) being the standard one-parameter family Brownian white noise that satisfies

E(W(x) W(x)) = (x x),(2.1)where denote the usual Dirac -function and Ethe expectation. A piecewise constantapproximation of the one-parameter white noise is given by [1]

dWn(x)dx

= cn

nj=1

jj(x),(2.2)

8/2/2019 dz02sinum

3/25

NUMERICAL APPROXIMATION OF STOCHASTIC PDEs 1423

0 0.2 0.4 0.6 0.8 12

1

0

1

0 0.2 0.4 0.6 0.8 110

5

0

5

0 0.2 0.4 0.6 0.8 110

0

10

0 0.2 0.4 0.6 0.8 120

0

20

0 0.2 0.4 0.6 0.8 120

0

20

0 0.2 0.4 0.6 0.8 150

0

50

0 0.2 0.4 0.6 0.8 1100

0

100

0 0.2 0.4 0.6 0.8 1100

0

100

n=4 n=8

n=16 n=32

n=64 n=128

n=256 n=512

Fig. 2.1. Piecewise constant approximation for the noise dWn(x)/dx = (1/

h)nj=1 jj(x).

where cn = h1/2 =

n and, for j = 1, 2, . . . , N , j N(0, 1) is independently and

identically distributed (iid),

hj =

xj+1xj

dW(x) , and j(x) =

1, xj x < xj+1,0 otherwise.

The discrete analogue of (2.1) for the piecewise constant approximation is given by

EdWn(x)dx dWn(x)

dx = h1 ifxj x, x < xj+1 for some j,

0 otherwise.

Hence,

limn

E

dWn(x)

dx dWn(x)

dx

= (x x).

In Figure 2.1, some sample realizations of the piecewise constant approximationof one-parameter white noise are illustrated for various values of n. (The randomnumbers are generated using MATLAB.) We note that similar discussions can beeasily generalized to the space time two-parameter family white noises.

2.1. Noises in abstract forms. The SPDEs driven by the white noise oftenhave poor regularity estimates. In the physical world, to take into account the shortand long range correlations of the stochastic effects, both white noise and colorednoises may be considered. There are many situations where colored noises modelthe reality more closely, and there are also instances where the important stochasticeffects are the noises acting on a few selected frequencies.

In general, we may use an abstract formulation of the infinite dimensional noise:

8/2/2019 dz02sinum

4/25


0 0.2 0.4 0.6 0.8 10

0.5

1

0 0.2 0.4 0.6 0.8 10

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 10

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 10.5

0

0.5

1

n=4 n=8

n=16 n=32

n=64 n=128

n=256 n=512

0 0.2 0.4 0.6 0.8 10

1

2

3

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 11

0

1

2

0 0.2 0.4 0.6 0.8 10

2

4

0 0.2 0.4 0.6 0.8 12

0

2

4

0 0.2 0.4 0.6 0.8 11

0

1

2

n=4 n=8

n=16 n=32

n=64 n=128

n=256 n=512

Fig. 2.2. Noises by Fourier modesn

k=1 kk

2sin kx with k =1

2k(left) and k =

1

k3/2

(right).

W(x) =

k=1

kkk(x),(2.3)

where the random variable k N(0, 1) is iid for any k, the deterministic functions{k(x)} form an orthonormal basis of L2(0, 1) or its subspace, and the coefficients{k} are to be chosen to ascertain the convergence of the series in the mean squaresense with respect to some suitable norms.

One of the examples is given by the Fourier modes k(x) =

2sin kx whichforms a basis of H10 (0, 1). According to the different decay rates of the coefficients,the noises may display quite different pictures. The pictures in Figure 2.2 and theleft two columns of Figure 2.3 provide sample realizations of noises having forms(2.3) in the Fourier basis with coefficients k = 2

k, k3/2, and k1/2, respectively.Clearly, the realizations give trajectories that look smoother than the ones for the

white noise. It can also be seen that the faster the coefficients k decay, the smootherthe noise trajectory dWn/dx looks, which reflects stronger spatial correlation sincethe noises are heavily concentrated near a few low frequencies. On the other hand,if the coefficients decay sufficiently slowly, then the trajectory can clearly resemblethat of a white noise away from the boundary. In fact, it is well known that forspatially uncorrelated white noises, their Fourier coefficients are independent of thefrequencies, and they stay at a constant value.

In the analysis and numerical examples given in later sections, the noises givenin terms of the Fourier modes are used. The Fourier modes provide one of manypossible representations of noises where the smoothness of the noise trajectories arerelated to the decay of the coefficients in the representation. Another illustrativeexample is to define the noise in terms of the lowest order wavelet basis. We includethe discussion here for comparison. Let be the wavelet functionand be the scaling

function[32]. Let j denote the dilation index and k denote the translation index, andj,k(x) = 2

j/2(2jxk). The discrete noise formulated in the wavelet basis is given as

WJ(x) = c(x) +J1j=0

2j1k=0

dj,kjkj,k(x).(2.4)

Here, J is the highest level to be considered, and , jk N(0, 1) are iid. In the

8/2/2019 dz02sinum

5/25


0 0.2 0.4 0.6 0.8 10

2

4

0 0.2 0.4 0.6 0.8 15

0

5

0 0.2 0.4 0.6 0.8 12

0

2

4

0 0.2 0.4 0.6 0.8 15

0

5

10

0 0.2 0.4 0.6 0.8 15

0

5

10

0 0.2 0.4 0.6 0.8 15

0

5

10

0 0.2 0.4 0.6 0.8 15

0

5

10

0 0.2 0.4 0.6 0.8 110

0

10

n=4 n=8

n=16 n=32

n=64 n=128

n=256 n=512

0 0.2 0.4 0.6 0.8 10

2

4

0 0.2 0.4 0.6 0.8 12

0

2

4

0 0.2 0.4 0.6 0.8 15

0

5

0 0.2 0.4 0.6 0.8 12

0

2

4

0 0.2 0.4 0.6 0.8 12

0

2

4

0 0.2 0.4 0.6 0.8 15

0

5

0 0.2 0.4 0.6 0.8 15

0

5

0 0.2 0.4 0.6 0.8 12

0

2

4

J=2 J=3

J=4 J=5

J=6 J=7

J=8 J=9

Fig. 2.3. Noises byn

k=11

k1/2k

2sin kx (left) and (x) +

J1j=0

2j1k=0

12j

jkj,k(x) (right).

simplest case, we may take the Haar wavelet

(x) = 1, 0 x < 1/2,1, 1/2 x < 1,

0 otherwise

and (x) = 1, 0 x < 1,

0 otherwise.

The right two columns of Figure 2.3 show sample realizations of noises taking the form

(2.4) with c = 1, dj,k = 2j . The correlation of the noise (2.4) E(dWJ(x)dx dWJ(x

)dx ) is

given by

E

dWJ(x)

dx dWJ(x

)dx

= c2(x)(x) +

J1j=0

2j1k=0

d2j,kj,k(x)j,k(x) .(2.5)

If in (2.2) n = 2J and h = 2J, then the piecewise constant approximation of the

white noise may also be represented using the wavelet Haar basis. In fact, let k(x)be characteristic function of interval [kh, (k + 1)h]; then

dWn(x)dx

=1h

2J1k=0

kk(x) = (x) +

J1j=0

2j1l=0

j,lj,l(x).

Here, = 2J/22J1

k=0 k N(0, 1) and

j,l = 2(jJ)/2

(l+1)2Jj1k=l2Jj

(1)[k/2Jj1]k

N(0, 1)are iid. Corresponding to (2.5), c = dj,k = 1 so that (2.5) leads again to (2.1). Nat-urally, when higher order wavelets are used [32, 30], we may expect to have discretenoises that are smoother spatially than the ones represented by the Haar basis whenthe high frequency coefficients enjoy fast decay properties. Comparing with Fouriermodes, wavelet functions may also have compact support; thus, on the one hand,the noises in wavelet basis can closely resemble spatially uncorrelated white noises,while on the other hand they can also be used conveniently to simulate noise moreconcentrated on certain frequencies as well as certain spatial regions.

8/2/2019 dz02sinum

6/25


In summary, different forms to represent the various noises are discussed in thissection. Similar discussion can be carried out in more than one space dimensionand for noises parameterized by both time and space variables. Such discussions arerelevant to the numerical study of SDEs as the solutions of the stochastic equations

that use noises with better regularity become more regular themselves and thus mayallow higher order numerical approximations.

3. Numerical method and error analysis. In [1], approximations of SPDEswith the additive space time white noise term discretized by the piecewise constantrandom process have been studied. Here, we follow roughly the same route, thoughmore general types of noises are used. We show how the accuracy is affected by thecorrelation or the smoothness of the noises.

We divide the discussion into two parts, starting with the simplest one dimensionalelliptic equation (boundary value problem of a SODE) and then moving to a parabolicequation in one space dimension and in time (initial boundary value problem of aSPDE). In the set-up of the problems, noises represented in general basis are used,but in the analysis we specialize in using the Fourier modes as the basis of choice tosimplify the discussion.

3.1. One dimensional elliptic equation with noise. We now consider theSDE (1.2); that is,

u(x) + bu(x) = g(x) + W(x), 0 < x < 1,u(0) = u(1) = 0,

(3.1)

where W(x) denotes the noise, g(x) is a given deterministic term, and b = b(x) is agiven deterministic coefficient.

As in [1], we may first replace W(x) by a finite dimensional noise Wn(x) and let undenote the solution of the corresponding equation. We then numerically approximatethe equation associated with Wn(x) and let u

hn denote the numerical solution.

If the noise W(x) in (3.1) is the white noise, Wn(x) is the piecewise constant

approximation (2.2), and the Galerkin finite element method with piecewise constantbasis is applied to (3.1), the error estimate is given by [1]

Eu unL2 C h,Eun uhnL2 C h3/2,Eu uhnL2 C h.

Due to the poor regularity of the solution, it is seen that, even with higher orderfinite elements, the order of error estimates does not improve. With colored noises,the order of approximation may increase with better regularity on the solution andthe use of higher order elements. As an illustration, we consider the following noise:

W(x) = k=1

kkk(x),(3.2)

where {k} are random variables satisfyingk N(0, 1) and cov(k, l) = E(kl) = qkl,

with {k} to be chosen.

8/2/2019 dz02sinum

7/25


Let {nk }k=1 approach {k}k=1 as n in some appropriate sense; then anapproximation of W(x) is

Wn

(x) =

k=1

2nk

k

k

(x)sin kx.

The definition of noise term leads to the following stochastic integral for f L2(0, 1):

S =

10

f(x)dW(x) =k=1

kfkk,

Sn =

10

f(x)dWn(x) =

k=1

nkfkk,

where fk =

1

0 f(x)k(x)dx. That is, S and Sn are random variables having thedistribution

S N

0,

k=1

l=1

klfkflqkl

,

Sn N

0,k=1

l=1

nknl fkflqkl

,

provided the double sum is convergent.For convenience, we introduce the following notation:

n = (n1 ,

n2 , . . . ,

nk , . . . )

T,

= (1, 2, . . . , k, . . . )T

are infinite column vectors. For two vectorsn and

f , we use

nf to denote the

componentwise product

nf = (n1 f1,

n2 f2, . . . ,

nk fk, . . . )

T.

Let Q be the covariance matrix of random fields {k}, namely, Q is the infinitematrix (operator) with entries Q = (qkl)

k,l=1. For an integer s, let Qs be the infinite

matrix with entries Qs = ((kl)sqkl)

k,l=1. It is easy to see both Q and Qs are positive

semidefinite. Define the weighted semi-inner products of the vectors and as

,

Q =

T

Q

=

k=1

l=1 klqkl,, Qs = T Qs =

k=1

l=1

kl(kl)sqkl.

The seminorms induced by the above semi-inner products are

2Q = , Q and 2Qs = , Qs .

8/2/2019 dz02sinum

8/25


Note that Q0 = Q. Using the above notation,

S N

0, f2Q

, Sn N

0, nf2Q

.

The difference between S and Sn is given by

E|S Sn|2 = E|k=1

(nk k)fkk|2 =f nf2

Q.

Equation (3.1) can be written in a weak form or an integral form. Both formsare equivalent as shown in [3]. In fact, the solution of (3.1) is a stochastic processu = u(x) which satisfies the weak formulation

10

u(x)(x)dx +

10

bu(x)(x)dx =

10

g(x)(x)dx +

10

(x)dW(x)(3.3)

for

C2(0, 1)

C0(0, 1). The integral form is

u(x) +

10

b k(x, y) u(y)dy =

10

k(x, y)g(y)dy +

10

k(x, y)dW(y).(3.4)

Here, k(x, y) = x y xy is the Greens function associated with the elliptic equationv(x) = (x), v(0) = v(1) = 0 so that v(x) = 10 k(x, y)(y)dy. (x y means thesmaller one of x and y.) In the present investigation, it is assumed the coefficient b is

small enough so that 2 =10

10 b

2k2(x, y)dxdy < 1. We note that this condition isprimarily needed in the case of b < 0; such a restriction can be lifted for b > 0, andthe conclusions given later remain valid.

We now substitute dW(y) by dWn(y) in (3.4) to obtain the following equation:

un(x) + 10 b k(x, y) un(y)dy = 1

0 k(x, y)g(y)dy + 1

0 k(x, y)dWn(y).(3.5)

Thus, un(x) satisfy the two-point boundary value problem

un(x) + bun(x) = g(x) + Wn(x), un(0) = un(1) = 0.(3.6)

The following theorem shows that un indeed approximates u, the solution of(3.4). In order to illustrate the higher order of convergence for more regular noises,we specialize our discussion to the choice of {k(x) =

2sin kx}, that is, noises

represented by the Fourier modes.Theorem 3.1. For Wn(x) =

k=1

nkkk(x) and k(x) =

2sin kx, if un

andu are the solutions of (3.5) and(3.4), respectively, then, for some constant C > 0,

Eu unL2 C1 n Q1 ,where < 1 is defined as before.

Proof. Let en(x) = u(x) un(x) and

F(x) =

10

k(x, y)dW(y) 10

k(x, y)dWn(y).

8/2/2019 dz02sinum

9/25


Subtracting (3.5) from (3.4), we have

en(x) = 10

b k(x, y) en(y)dy + F(x).

By Holders inequality, it is easy to show that10

e2n(x)dx 210

e2n(y)dy + 2

10

F2(x)dx

1/2 10

e2n(y)dy

1/2+

10

F2(x)dx,

where 2 =10

10

b2k2(x, y)dxdy and it is assumed that < 1. Taking expectations

on both sides, letting en = E(10 e

2n(x)dx) and Gn = E(

10 F

2(x)dx) and using theBurkholderGundy-type inequality (EX)2 E(X2), we get

en(1 2) 2

en

Gn Gn 0.(3.7)

This implies en

Gn(1 ).(3.8)

Now let us estimate Gn.

Gn = E

10

F2(x)dx

=

10

E

k=1

(nk k)fk(x)k2

dx

=

10

f(x) nf(x)2Q

dx,

wheref(x) = (f1(x), f2(x), . . . , f k(x), . . . )

T and fk(x) = 1

0 k(x, y)k(y)dy. Since

k(x, y) = x y xy, direct calculation gives that, for any x [0, 1],

|fk(x)| =1

0

k(x, y)k(y)dy

= 10

k(x, y)

2sin kydy

ck ,which implies that, for x [0, 1],f(x) nf(x)

Q C

nQ1

for some constant C > 0. Hence,

Gn C

n

2

Q1.

Combining the above inequality with (3.8), we get

Eu unL2

Eu un2L2 =

en C1

n Q1

.

This proves the theorem.We now state a bound on Wn(x) in the following lemma.

8/2/2019 dz02sinum

10/25


Lemma 3.1. For Wn(x) =

k=1 nkkk(x) and k(x) =

2sin kx, if s 0 is

an integer, then

E

WnHs

C

k=1(nkks)21/2

,

provided that the right-hand side is convergent.

Proof. First,

ds

dxs

dWn

dx

=

k=1

2nk k(k)

s sin

s

2+ kx

.

Since {sin(s2 + kx)} are orthogonal on [0, 1], we have

E

ds

dxs

dWn

dx

2

L2

= E

10

k=1

2nkk(k)

s sin

s

2+ kx

2dx

= E k=1

(nk )22k(k)2s c k=1

(nk ks)2

for some constant c > 0. The above inequality also implies that, for any r s,

E

drdxr

dWndx

2L2

E dsdxs

dWn

dx

2L2

.

Hence,

EWnHs

EWn2Hs C k=1

(nkks)2

1/2for some constant C > 0.

Concerning the above lemma, we note that similar lower bound can also be es-tablished. Moreover, the results may be established for the case s < 0 as well.

We now consider a standard finite element approximation of un. From the weakformulation (3.3), un satisfies1

0

un(x)dx + b

10

un(x)(x)dx =

10

g(x)(x)dx +

10

(x)dWn(x)(3.9)

for (x) H10 (0, 1). By the LaxMilgram theorem, there exists a unique solutionun H10 (0, 1) to (3.9). For convenience, we consider the same partition of [0, 1]:0 = x1 < x2 < < xn+1 = 1 with xi = (i1)h and h = 1/n. IfVh0 (0, 1) denotes thefinite element subspace ofH10 (0, 1), and {j(x)}Nj=1 forms a basis ofVh0 (0, 1), the finiteelement solution of (3.9) is uhn

Vh0 (0.1) that satisfies (3.9) for all (x)

Vh0 (0, 1).

Thus, uhn(x) = Nl=1 ull(x) satisfies the following linear system for j = 1, 2, . . . , N :Nl=1

ul

10

l(x)j(x) + b

Nl=1

ul

10

l(x)j(x)dx

=

10

g(x)j(x)dx +k=1

nk k

10

j(x)k(x)dx,(3.10)

8/2/2019 dz02sinum

11/25


where k N(0, 1). The solution uhn is clearly well defined.The following lemma gives the standard finite element error estimates of (3.9) in

the pathwise sense.Lemma 3.2. If Vh0 (0, 1) contain all piecewise polynomials of degree r in H

10 (0, 1),

and un H10 (0, 1) H

r+1

(0, 1), then

un uhnL2 + h|un uhn|H1 Chr+1unHr+1 Chr+1g + WnHr1(3.11)

for some constantC > 0.Furthermore, combining Theorem 3.1 and Lemma 3.2, an estimate on E(u

uhnL2) follows from the triangle inequality.Theorem 3.2. Let u and uhn be the solution of (3.3) and (3.10), respectively; if

the hypothesis in Lemma 3.2 is satisfied, then the error estimate is

Eu uhnL2 Cn

Q1

+ hr+1gHr1 + hr+1EWnHr1

Cn Q1 + hr+1gHr1 + hr+1 k=1

(nkkr1)21/2(3.12) for some generic constantC > 0.

Numerical examples are given in a later section to provide an illustration of thespecific order of error estimates one can get based on the above theorem.

Remark3.1. The same idea can be applied to two dimensional elliptic equations ina rectangular domain, namely, by representing the two dimensional noise as the combi-nations of the tensor products ofk(x), similar to how Theorem 3.2 can be obtained.

3.2. Parabolic equation in one spatial dimension. Let 2W

tx denote a spacetime noise term and g be a deterministic function; we now consider the linear stochas-tic equations of the form

ut (t, x)

2ux2 (t, x) + bu(t, x) =

2Wtx(t, x) + g(t, x), t > 0,

u(0, x) = u0(x), 0 x 1,u(t, 0) = u(t, 1) = 0, t 0,

(3.13)

where the coefficient b, for simplicity, is assumed to be a constant.The weak formulation of (3.13) is1

0

u(t, x)(x)dx t0

10

u(s, x)d2

dx2dxds +

t0

10

bu(s, x)(x)dxds

=

1

0

u0(x)(x)dx +

t

0 1

0

(x)dW(s, x) +

t

0 1

0

g(s, x)(x)dxds(3.14)

for C2[0, 1] C0[0, 1] . The integral formulation of (3.13) is

u(t, x) +

t0

10

Gts(x, y)bu(x, y)dyds =10

Gt(x, y)u0(y)dy(3.15)

+

t0

10

Gts(x, y)dW(s, y) +t0

10

Gts(x, y)g(s, y)dyds,

8/2/2019 dz02sinum

12/25


where Gt(x, y) = 2

m=1 sin mx sin mye(m)2t is the fundamental solution of

vt(t, x) vxx(t, x) = 0, v(0, x) = (x), v(t, 0) = v(t, 1) = 0,

so that v(t, x) = 10 Gt(x, y)(y)dy .Using the same idea as that in the previous section, we represent the noise as2W

tx=

k=1

k(t) k(t)k(x),(3.16)

where k(t) is a continuous function, k(t) is the derivative of standard Wienerprocess, and k(x) =

2sin kx. Now define a partition of [0, T] [0, 1] by

rectangles [ti, ti+1] [xj , xj+1] for i = 1, 2, . . . , I and j = 1, 2, . . . , n, whereti = (i 1)t, xj = (j 1)h, t = T /I, and h = 1/n. A sequence of noise whichapproximates the noise is defined as

2Wn

tx =

k=1

n

k (t)k(x)

I

i=1

1

t kii(t),(3.17)

where i(t) is the characteristic function for the ith time subinterval and

ki =1t

ti+1ti

dk(t) N(0, 1).

Replacing k(t) by nk (t), we get the discretization in the x-direction, and replacing

k(t) byI

i=11t

kii(t) we get the discretization in the t-direction. Then2Wntx is

substituted for 2W

tx in (3.15) to get the following equation:

un(t, x) + t

0 1

0 Gts(x, y)bun(s, y)dyds = 1

0 Gt(x, y)u0(y)dy(3.18)

+

t0

10

Gts(x, y)dWn(s, y) +t0

10

Gts(x, y)g(s, y)dyds;

that is, un is the solution of the equationunt (t, x)

2unx2 (t, x) + bun(t, x) =

2Wntx (t, x) + g(t, x), t > 0,

un(0, x) = u0(x), 0 x 1,un(t, 0) = un(t, 1) = 0, t 0.

(3.19)

Now we assume that

T010t010

G2ts(x, y)b2dydsdxdt = 2 < 1.

Then, under proper assumptions on {k(t)} and {nk (t)}, un approximates u, thesolution of (3.15), as illustrated in the next theorem.

Theorem 3.3. Let {k(t)} and its derivative be uniformly bounded by

|k(t)| k, |k(t)| k t [0, T],

8/2/2019 dz02sinum

13/25


and the coefficients {nk (t)} are constructed such that|k(t) nk (t)| nk , |nk (t)| nk , |nk (t)| nk t [0, T]

with positive sequences{

n

k}being arbitrarily chosen,

{n

k }and

{n

k }being related to

{nk k} and {k}. Let un(t, x) and u(t, x) be the solution of (3.18) and (3.15),respectively; then, for some constants C > 0, independent of t and h,

Eu un2L2 C

(1 )2k=1

(nk)

2

2(k)2+ [k4(nk )

2 + (nk )2](t)2

,(3.20)

provided that the infinite series are all convergent.

The proof of Theorem 3.3 is given in the appendix.Remark 3.2. The assumption on being small is not crucial; some generaliza-

tions can be made without this assumption, for example when b < 0.Now we consider the approximation of un. In particular, we use a finite element

discretization with respect to the x variable and an implicit difference method in the

t variable. Since un satisfies the weak formulation,10

un(t, x)(x)dx +

t0

10

unx

(s, x)d

dx(x)dxds +

t0

10

bun(s, x)(x)dxds

=

10

u0(x)(x)dx +

t0

10

(x)dWn(s, x) +

t0

10

g(s, x)(x)dxds(3.21)

for H10 (0, 1). Meanwhile, the semidiscretization in space leads only to the followingproblem: find un(t, ) H10(0, 1), t (0, T), such that1

0

unt

dx +

10

unx

xdx +

10

bundx =

10

g +

2Wntx

dx(3.22)

with 10

un(0, x)(x)dx =

10

u0(x)(x)dx

for all H10 (0, 1), t (0, T).The finite element discretization of (3.22) is to find uhn(t, ) Vh0 (0, 1), t (0, T),

such that 10

uhnt

dx +

10

uhnx

xdx +

10

buhndx =

10

g +

2Wntx

dx(3.23)

with

10 uhn(0, x)(x)dx = 1

0 u0(x)(x)dx

for all Vh0 (0, 1), t (0, T). Here, Vh0 (0, 1) denote the finite element subspace ofH10 (0, 1). By using the expression

uhn(t, x) =n1l=1

ul(t)l(x), t (0, T),

8/2/2019 dz02sinum

14/25


(3.23) leads to a system of ODEs for ul(t), l = 1, . . . , n1. Using the backward-Eulermethod to solve this ODE system yields the following numerical scheme:

n1

l=1 (ui+1,l ui,l)1

0

l(x)j(x)dx + tn1

l=1 ui+1,l 1

0

l(x)j(x)dx

+ bt

n1l=1

ui+1,l

10

l(x)j(x)dx

=

ti+1ti

10

g(s, x)j(x)dxds +

ti+1ti

10

j(x)dWn(s, x)(3.24)

for j = 1, 2, . . . , n 1, i = 1, 2, . . . , I where ui,l ul(ti) . Let

uhn(ti, x) =

n1l=1

ui,ll(x).

For simplicity, we now focus on the case of using the continuous piecewise linear finiteelement in the spatial discretization. The following pathwise error estimate can befound in Theorem 8.2 of [17]:

un(tm, ) uhn(tm, )L2(3.25)

C

1 + logtmt

maxim

titi1

unt (, )L2

d + maxttm

h2un(t, )H2

.

The following lemma gives estimates of the terms on the right-hand side of (3.25).Lemma 3.3. Let un be the solution of (3.15) with g C2([0, T] [0, 1]), u0

C2[0, 1], and nk (t) has the bound given in Theorem 3.3. Let the constantb be suitablysmall. Then, if t 1/(2|b|), the following inequalities hold for some constant c,independent of t and h:

E

titi1

unt (, )L2

d c

(t)2 + tk

k2(nk )2 +k

(tnk )2

1/2(3.26)

and

Eun(t, )H2 c

1 +1

t

k

k2(nk )2

1/2.(3.27)

The proof of Lemma 3.3 is given in the appendix.Combining Lemma 3.3 and inequality (3.25), we have the following theorem.Theorem 3.4. Assume that the conditions in Lemma 3.3 hold; then

Eun(tm, ) uhn(tm, )L2 c

1 + logtmt

1/2

(t)2 + tk

k2(nk )2 +k

(tnk )2 +

h4

t

k

k2(nk )2

1/2 for some constant c.

8/2/2019 dz02sinum

15/25


The error Eu(tm, ) uhn(tm, )L2 can be obtained by applying the triangleinequality to the results of Theorems 3.3 and 3.4.

Remark 3.3. Note that when applied to the case of white noise, that is, k(t) = 1for all k, we may take nk =

nk = 1,

nk = 0 for k N, and nk = nk = 0, nk = 1

for k > N, where N as n ; then, after simplification, the estimates in theabove theorems give

Eu(tm, ) uhn(tm, )L2 c

1 + logtmt

1/21

N1/2+ (t)1/2N3/2 +

h2N3/2

()1/2

so that h = O(t)1/2 and N = O(h1/2) = O((t)1/4) give a best order of (t)1/8

or h1/4, up to a logarithmic factor, for Eu(tm, )uhn(tm, )L2 . This is indeed a verylow order convergence estimate as was expected [1]. In the next section, however, wepresent a few examples with colored noises for which the above theorems allow muchbetter estimates on the order of the approximations.

Remark 3.4. The estimate on the order of convergence in the time step size is seento be at best O(

t), which is largely due to the fact that we restricted our attention

to the case where {k(t)} in (3.16) correspond to the derivatives of the Wiener processwith t being the parameter. In many physical applications, other processes may alsobe used [11]. One may also naturally consider more general formulation for the noiseterms {k(t)} like what is used for dW/dx in (3.2). In the case where {k} are moreregular in time, better error estimates may be obtained using similar techniques.

Discussions and extensions to higher space dimensions can be found in [36].

4. Numerical results for some model equations.

4.1. One dimensional elliptic equation. We now study two cases of the onedimensional elliptic equation with noise described in the previous section. We demon-strate that for different forms of coefficient {nk }, different rates of convergence are tobe obtained.

Case 1. Let the random variables

{k

}be iid, namely,

qkl = E(kl) = kl =

1 ifk = l,

0 ifk = l, k =1

k3/2, nk =

k, k n,0, k > n.

Then

n Q1

=

k=n+1

1

k3/2 1

k

21/2 1

n2.

From Lemma 3.1, we have, for some generic constant C > 0,

E

WnL2 C

k=1

(n

k )2

1/2

C

k=1

1

k31/2

= C.

In other words, Wn L2(0, 1); this means that, in Theorem 3.2, r = 1. If thepiecewise linear finite element basis is used, and g L2(0, 1), the following errorestimate yields

E(u uhnL2) C(n2 + h2g + WnL2) C h2 .

8/2/2019 dz02sinum

16/25


Thus, asymptotically, we have a second order convergence rate in h for the expectationof the L2 error.

Case 2. Now let us consider using different coefficients {nk } which yield highorder convergence results for high order finite element spaces. Still let

qkl = E(kl) = kl = 1 ifk = l,

0 ifk = l, k =1

k7/2, nk =

k, k n,0, k > n.

Then n Q1

=

k=n+1

1

k7/2 1

k

21/2 1

n4.

From Lemma 3.1, we have

EWnH2 C k=1

(nkk2)2

1/2 C

k=1

1

k7/2k221/2

= C.

In other words, Wn H2(0, 1); this means that, in Theorem 3.2, r = 3. If we usethe cubic spline finite element basis, and assume that g is bounded in H2(0, 1), thefollowing error estimate yields

E(u uhnL2) C(n4 + h4g + WnH2) C h4

for some constant C that depends only on g. Note that such a high order cannot beachieved if we have adopted a white noise [1].

The finite element method (3.10) is implemented for (3.1) with g(x) = 2+bxbx2and the noise W as defined in section 3. The exact solution of (3.1) is given byu = ud + us, where ud and us correspond to the deterministic and the stochasticparts. Moreover, ud(x) = x(1 x) and

us(x) = k=1

2kb + (k)2

k sin kx .

The numerical solution is calculated for n = 4, 8, 16, 32, 64, 128 (h = 1/n being thelength of the subintervals). For each n, 10,000 runs are performed with differentsamples of the noise, uuhnL2 is calculated for each sample, and the averaged valueEu uhnL2 is calculated.

For Case 1, we let b = 0.5, k = k3/2, and we use the continuous piecewise

linear finite element space. The left picture in Figure 4.1 gives the decay of error.The horizontal axis denotes log10 n, and the vertical axis denotes log10 Eu uhnL2 .The slope of the error curve is nearly 2, in agreement with the theoretical result.

As for Case 2, we let b = 0.5, k = k7/2, and we use the finite element space

consisting of piecewise cubic splines. The right picture in Figure 4.1 gives the decay

of error. The slope of the error curve is now nearly 4, also in agreement with thetheoretical result.

4.2. Parabolic equation in one spatial dimension. Now consider a specialcase of parabolic equation described in the previous section. Let

k(t) =cos t

k3, nk (t) =

k(t), k n,

0, k > n,

8/2/2019 dz02sinum

17/25


0.5 0 0.5 1 1.5 2 2.5 3

6

5.5

5

4.5

4

3.5

log(n)

log(En)

2 1 0 1 2 3 4 5

10

9

8

7

6

5

log(En

)

log(n)

Fig. 4.1. The error decay with k = k3/2 and k7/2.

and the upper bounds nk , nk ,

nk given in Theorem 3.3 can be chosen as

n

k = 0, k n,1k3 , k > n, nk = nk = 1k3 .Backward-Euler in time with the piecewise linear finite element in space approxima-tion (3.24) was tested for the numerical solution of problem (3.14) with

g(t, x) = 10(1 + b)x2(1 x)2et 10(2 12x + 12x2)et .

We use b = 0.5 and T = 1 . In the absence of noise term, the exact solution is

u(t, x) = ud(t, x) = 10etx2(1 x)2 , with u0(x) = 10x2(1 x)2 .

The exact value of Eu(1, 0.5) is about 1.699.In theory, using the above definitions, we have

k=1

(nk)2

2(k)2

k=n+1

1

k8 1

n7= h7,(4.1)

k=1

k4(nk )2 + (nk )

2) k=1

1

k2+

1

k3

C,(4.2)

k=1

(nk )2

k=1

(knk )2 C .(4.3)

From Theorems 3.3 and 3.4, we have

Eu unL2 c(h7 + (t)2)1/2,

Eun(tm, ) uhn(tm, )L2 c1 + log tmt1/2(t)1/2 + h2(t)1/2 .Hence,

Eu(tm, ) uhn(tm, )L2 c

1 + logtmt

1/2(t)1/2 +

h2

(t)1/2

.(4.4)

8/2/2019 dz02sinum

18/25


Table 4.1

E(uhn(1, 0.5)) and E(uhn(1, 0.5))

2 by the backward-Euler finite element scheme.

h t E(uhn(1, 0.5)) E(uhn(1, 0.5))

2 E(n/2,I) var(n/2,I)

.25 .25 1.5268 2.3495 .0061 .9830

.25 .125 1.6147 2.6301 -.0217 1.0141

.25 .0625 1.6599 2.7826 -.0166 1.0079

.25 .03125 1.6821 2.8586 .0083 .9908

.25 .01563 1.6976 2.9142 -.0086 .9750

.125 .25 1.5198 2.3283 .0045 .9697

.125 .125 1.6071 2.6059 -.0014 1.0097

.125 .0625 1.6529 2.7569 -.0238 .9780

.125 .03125 1.6777 2.8432 .0002 .9829

.125 .01563 1.6912 2.8910 .0006 .9687

.0625 .25 1.5193 2.3263 -.0006 1.0182

.0625 .125 1.6043 2.5963 -.0069 .9886

.0625 .0625 1.6519 2.7539 .0124 .9852

.0625 .03125 1.6758 2.8372 -.0110 .9908

.0625 .01563 1.6888 2.8825 .0069 .9962

.03125 .25 1.5198 2.3277 -.0163 .9650

.03125 .125 1.6044 2.5971 -.0217 .9527

.03125 .0625 1.6503 2.7497 -.0071 .9984

.03125 .03125 1.6731 2.8281 .0044 .9765

.03125 .01563 1.6855 2.8724 -.0101 1.0479

.01563 .25 1.5181 2.3230 -.0166 .9918

.01563 .125 1.6067 2.6041 -.0134 .9667

.01563 .0625 1.6500 2.7482 -.0114 1.0067

.01563 .03125 1.6749 2.8336 -.0170 1.0365

.01563 .01563 1.6851 2.8704 -.0067 .9872

In the actual implementation, different values of t and h were used. For each pair{t, h}, 10,000 runs are performed with different sample of noise, and the ensembleaverages are calculated. The numerical results of E(uhn(1, 0.5)) and E(u

hn(1, 0.5))

2

are presented in Table 4.1.The computational results converge as t and h approach to 0. From the table,

it can be observed that, for fixed h, the results converge faster as t decreases, butfor fixed t the convergence is less transparent as h decreases. This can be explainedby the error estimate (4.4), which is bounded by (t)1/2 + h2(t)1/2. If t and hare of the same order, the t term dominates in the estimate.

The numerical accuracy is also affected by the random number generators usedin the different realizations. (The particular generator used in our implementation

is obtained using MATLAB.) For comparison, the last two columns of Table 4.1 listthe mean and variance of n/2,I. We see that, for the relatively larger magnitude of

E(n/2,I), the error of E(uhn(1, 0.5)) turns out to be larger as well.

Additional numerical examples can be found in [36].

5. Conclusion. In this paper, the numerical approximations of SDEs with dif-ferent noise realizations are considered. In many instances of stochastic modeling, thenoises may indeed be represented in various forms, with some emphasis on the correla-

8/2/2019 dz02sinum

19/25


tion in space and time, while others exhibit the correlation in frequency or spectrum.Our study indicates that the accuracy of the numerical approximation depends on theform of the underlying noise. Both rigorous error estimates and experimental resultsare provided in our paper.

Throughout our discussion, simple linear equations in one space dimension areused for the purpose of illustrations. We note that much of our consideration can begeneralized to stochastic elliptic and parabolic equations in higher space dimensions.For the case of a simple two dimensional square domain, related discussions have beenprovided in [36]. By confining the theoretical analysis to the one space dimension here,some tedious technical details and complicated expressions are avoided.

Naturally, it will be very interesting to study the similar problems for nonlinearSDEs, which actually motivated the present investigation. It is hopeful that suchstudies may lead to a better understanding of the behaviors of the discretization errorand the modeling error in conducting numerical simulations of nonlinear stochasticdynamics for practical problems [9, 18, 28].

Appendix.

Proof of Theorem 3.3.Step 1. First, we verify the existence of such {nk (t)}. Since {k(t)} are continuouson interval [0, T], by the Weierstrass approximation theorem, for an arbitrary sequencenk , where n is a fixed number, k = 1, 2, . . . , there exists a sequence of polynomial{Pnk (t)} such that

|k(t) Pnk (t)| nkT

t [0, T] .

Let

nk (t) =

t0

Pnk (s)ds + k(0),

and we have

|k(t) nk (t)| =t

0

(k(s) Pnk (s))ds nk .

By the triangle inequality,

|nk (t)| |k(t)| + nk k + nk = nk ,

|nk (t)| = |Pnk (t)| |k(t)| +nkT

k + nk

T= nk .

Step 2. Let en(t, x) = u(t, x) un(t, x) and

F(t, x) = t0

10

Gts(x, y)dW(s, y) t0

10

Gts(x, y)dWn(s, y),

en = E

T0

10

e2n(t, x)dxdt,

Fn = E

T0

10

F2(t, x)dxdt.

8/2/2019 dz02sinum

20/25


Subtracting (3.18) from (3.15), and applying similar manipulation as that in section3, we get

Eu un2L2 = en Fn

(1

)2.

To estimate Fn, we introduce an intermediate noise form

2Wntx

=k=1

nk (t) k(t)k(x),

that is, a noise discretized only in the x-direction. Let

F1(t, x) =

t0

10

Gts(x, y)dW(s, y) t0

10

Gts(x, y)dWn(s, y),

F2(t, x) =

t0

10

Gts(x, y)dWn(s, y) t0

10

Gts(x, y)dWn(s, y);

then

F(t, x) = F1(t, x) + F2(t, x),

Fn = E

T0

10

F2(t, x)dxdt 2

E

T0

10

F21 (t, x)dxdt + E

T0

10

F22 (t, x)dxdt

.

Taking advantage of the orthogonality of {sin kx} on the interval [0, 1], we have

F1(t, x) =k=1

2sin kxe(k)

2t

t0

(k(s) nk (s))e(k)2sdk(s).

Since k(t) is the standard Wiener process,

E

T0

10

F21 (t, x)dxdt =

k=1T0

e2(k)2t

t0

e2(k)2s(k(s) nk (s))2ds

dt

k=1

(nk)2

T0

e2(k)2t

t0

e2(k)2sds

dt C1

k=1

(nk )2

2(k)2.

Using

ki =1t

ti+1ti

dk(t),

we have

F2(t, x) =

k=1

ke(k)2t

t0

e(k)2snk (s)dk(s)

t0

e(k)2s

nk (s)

Ii=1

1t kii(s)ds

=k=1

ke(k)2t

It1i=1

ti+1ti

e(k)

2snk (s) 1

t

ti+1ti

e(k)2snk (s)ds

dk(s)

+

ttIt

e(k)

2snk (s) 1

t tIt

ttIt

e(k)2snk (s)ds

dk(s)

,

8/2/2019 dz02sinum

21/25


where It is the integer such that tIt < t tIt+1 . Then

E

T

0 1

0

F22 (t, x)dxdt

=k=1

T0

e2(k)2t

(t)2

It1i=1

ti+1ti

ti+1ti

(e(k)2snk (s) e(k)

2snk (s))ds2

ds

+

ttIt

(t

t tIt)2ti+1

ti

(e(k)2snk (s) e(k)

2snk (s))ds

2ds

dt.

For s, s [ti, ti+1], using the smoothness assumption on k(t), we get

|e(k)2snk (s) e(k)2snk (s)|

|e(k)2s e(k)2s|nk (s) + e(k)2s|nk (s) nk (s)|

(k)2

e(k)2ti+1

n

k (s)t + e(k)2ti+1

n

k (i)t e(k)2ti+1((k)2nk + nk )t.

Here, ti i ti+1. Without loss of generality, we assume t = tIt+1; then

E

T0

10

F22 (t, x)dxdt

k=1

T0

e2(k)2t

Iti=1

ti+1ti

1

(t)2(e(k)

2ti+1((k)2nk + nk )(t)

2)2ds

dt

C

k=1T

0

e2(k)2t

It

i=1 ti+1

ti

(e2(k)2ti+1((k)4(nk )

2 + (nk )2)(t)2ds dt

Ck=1

Iti=1

T0

e2(k)2te2(k)

2ti+1dt(k4(nk )2 + (nk )

2)(t)3

Ck=1

Iti=1

(k4(nk )2 + (nk )

2)(t)3

Ck=1

(k4(nk )2 + (nk )

2)(t)2.

The last inequality comes fromIt

i=1, 1 I = 1/t. The theorem is nowproved.

Proof of Lemma 3.3. In general, by applying the same technique as in the proofof Theorem 3.1, we may first estimate

E

t0

10

u2n(t, x)dxdt c

1 Et0

10

[(Gt(x, y)u0(y))2 + (Gts(x, y)g(s, y))2]dyds

+c

1 Et0

10

t0

10

Gts(x, y)dWn(s, y)2

dxdt.

8/2/2019 dz02sinum

22/25


Next, one may differentiate (3.18) to get

unt

(t, x) =

t

0 1

0

tGts(x, y)bun(s, y)dyds +

1

0

tGt(x, y)u0(y)dy

+t0

10

tGts(x, y)g(s, y)dyds +

t0

10

tGts(x, y)dWn(s, y).

Then one may estimate Etiti1

10 (

unt (t, x))

2dxdt using the above equation. Simi-

larly, one may estimate E10 (

2unx2 (t, x))

2dxdt.Since we have assumed that b is a constant, we now provide a simpler estimate

which, in spirit, is similar to the estimate derived from the integral formulation.

Let g(t, x) =

k gk(t)k(x), un(t, x) =

k u(n)k (t)k(x), u0(x) =

k ukk(x);

then

t

u(n)k (t) + (k

22 + b)u(n)k (t) = gk(t) +

nk (t)

t i kii(t).Thus, for t [ti1, ti),

u(n)k (t) = e

((k)2+b)tuk +t0

e((k)2+b)(ts)gk(s)ds

+i

j=1

t0

e((k)2+b)(ts)

nk (s)

tkjj(s)ds.

This leads to

u(n)k (t) = e((k)2+b)tuk + t

0

e((k)2+b)(ts)gk(s)ds

+i

j=1

kjt

tjtj1

e((k)2+b)(ts)nk (s)ds,

where tl = tl for l < i and ti = t. It follows that

E

u(n)k (t)

2 cu2ke2((k)

2+b)t + cT

t0

e2((k)2+b)(ts)g2k(s)ds

+c

t

i

j=1tj

tj

1

e((k)2+b)(ts)nk (s)ds

2

for some constant c.Since u0 C2[0, 1] and g C2([0, T] [0, 1]), we have, for some constant c > 0,

k

(k)4

u2ke2((k)2+b)t +

t0

e2((k)2+b)(ts)g2k(s)ds

c.

8/2/2019 dz02sinum

23/25


Using the bounds on nk and the fact that b is a small constant, we have

i

l=1

tl

tl1

e((k)2+b)(ts)nk (s)ds

2

c(nk )2t0

e2((k)2+b)(ts)ds

c(nk )

2

(k)2 + b.

Thus, for small b, we have

Eun(, )H2 ( Eun(, )2H2)1/2

c

1 +1

t

k

k2(nk )2

1/2

for some constant c > 0. This proves the inequality (3.27).For (3.26), we have

tu(n)k (t) = ((k)2 + b)u(n)k (t) + gk(t) +

nk (t)t

i

kii(t).

So,

E

titi1

tu(n)k (s)

2ds cE

titi1

((k22+b)u(n)k (s))

2ds+c

titi1

[g2k(s)+(nk (s))

2]ds.

Thus,

Etiti1

unt (, )L2 d tEti

ti1

unt (, )2

L2

d1/2

c

(t)2 + tk

k2(nk )2 +k

(tnk )2

1/2.

This proves (3.26).

Acknowledgments. The authors would like to thank Weinan E, Max Gun-zburger, and Zhimin Zhang for interesting discussions. The authors also want tothank them and an anonymous referee for providing useful references.

REFERENCES

[1] E. Allen, S. Novosel, and Z. Zhang, Finite element and difference approximation of somelinear stochastic partial differential equations, Stochastics Stochastics Rep., 64 (1998),pp. 117142.

[2] J.F. Bennaton, Discrete time Galerkin approximation to the nonlinear filtering solution, J.Math. Anal. Appl., 110 (1985), pp. 364383.

[3] R. Buckdahn and E. Pardoux, Monotonicity methods for white noise driven quasilinearSPDEs, in Diffusion Processes and Related Problems in Analysis, I, M. Pinsky, ed.,Birkhauser Boston, Boston, MA, 1990, pp. 219233.

8/2/2019 dz02sinum

24/25


[4] P.L. Chow, J.L. Jiang, and J.L. Menaldi, Pathwise convergence of approximation solutionsto Zakais equation in a bounded domain, in Stochastic Partial Differential Equations andApplications, G. Da Prato and L. Tubaro, eds., Longman Scientific and Technical, Harlow,UK, 1992, pp. 111123.

[5] G. Da Prato and L. Tubaro, Stochastic Partial Differential Equations and Applications,

Longman Scientific and Technical, Harlow, UK, 1992.[6] G. Da Prato and J. Zabczyk, Stochastic Equations in Infinite Dimensions, Cambridge Uni-

versity Press, Cambridge, UK, 1992.[7] A. Davie and J. Gaines, Convergence of numerical schemes for the solution of the parabolic

stochastic partial differential equations, Math. Comp., 70 (2001), pp. 121134.[8] D.A. Dawson, Stochastic evolution equations, Math. Biosci., 154 (1972), pp. 187316.[9] J. Deang, Q. Du, and M. Gunzburger, Thermal fluctuations of superconducting vortices,

Phy. Rev. B, 64 (2001), pp. 5250652510.[10] A. Etheridge, Stochastic Partial Differential Equations, Cambridge University Press, Cam-

bridge, UK, 1995.[11] R. Fox, Second order algorithm for the numerical integration of colored noise problems, Phys.

Rev. A(3), 43 (1991), pp. 26492654.[12] J.G. Gaines, Numerical Experiments with S(P)DEs, in Stochastic Partial Differential Equa-

tions, London Math. Soc. Lecture Note Ser. 216, Cambridge University Press, Cambridge,UK, 1995, pp. 5571.

[13] I. Gyongy, Lattice approximations for stochastic quasi-linear parabolic partial differential

equations driven by space-time white noise II, Potential Anal., 11 (1999), pp. 137.[14] E. Hausenblas, Error analysis for approximation of stochastic differential equations drivenby Poisson random measures, SIAM J. Numer. Anal., 40 (2002), pp. 87113.

[15] D.J. Higham, Mean-square and asymptotic stability of the stochastic theta method, SIAMJ. Numer Anal, 38 (2000), pp. 753769.

[16] M. Ibanes, J. Garcia-Ojalvo, R. Toral, and J.M. Sancho, Noise-induced phase separation:Mean-field results, Phys. Rev. E(3), 60 (1999), pp. 35973605.

[17] C. Johnson, Numerical Solution of Partial Differential Equations by the Finite ElementMethod, Cambridge University Press, Cambridge, UK, 1994.

[18] T. Kamppeter, F.G. Mertens, E. Moro, A. Sanchez, and A.R. Bishop, Stochastic vor-tex dynamics in two-dimensional easy-plane ferromagnets: Multiplicative versus additivenoise, Phys. Rev. B, 59 (1999), pp. 1134911357.

[19] J.B. Keller, Stochastic equations and wave propagation in random media, Proc. Sympos.Appl. Math., 16 (1964), pp. 145170.

[20] P. Kloeden and E. Platen, Numerical Solution of Stochastic Differential Equations,Springer-Verlag, New York, 1992.

[21]L. Machiels, and M.O. Deville

, Numerical simulation of randomly forced turbulent flows, J.Comput. Phys., 145 (1998), pp. 246279.[22] A. Majda, I. Timofeyev, and E. Eijnden, Models for stochastic climate prediction, Proc.

Natl. Acad. Sci. USA, 96 (1996), pp. 1468714691.[23] G.N. Milstein, E. Platen, and H. Schurz, Balanced implicit methods for stiff stochastic

systems, SIAM J. Numer. Anal., 35 (1998), pp. 10101019.[24] E.A. Novikov, Functionals and the random-force method in turbulence theory, Soviet Phys.

JETP, 20 (1965), pp. 12901294.[25] E. Platen, An Introduction to Numerical Methods for Stochastic Differential Equations, Acta

Numer. 8, Cambridge University Press, Cambridge, UK, 1999, pp. 197246.[26] P. Protter and D. Talay, The Euler scheme for Levy driven stochastic differential equations,

Ann. Probab., 25 (1997), pp. 393423.[27] Y. Saito and T. Mitsui, Stability analysis of numerical schemes for stochastic differential

equations, SIAM J. Numer. Anal., 33 (1996), pp. 22542267.[28] J. Sancho, J. Garcia-Ojalvo, and H. Guo, Non-equilibrium Ginzburg-Landau model driven

by colored noise, Phys. D., 113 (1998), pp. 331337.[29] T. Shardlow, Numerical methods for stochastic parabolic PDEs, Numer. Funct. Anal. Optim.,

20 (1999), pp. 121145.[30] G. Strang, Wavelets and dilation equations: A brief introduction, SIAM Rev., 31 (1989),

pp. 614627.[31] D. Talay, Simulation and numerical analysis of stochastic differential systems, in Effective

Stochastic Analysis, P. Kree and W. Wedig, eds., Springer-Verlag, Berlin, 1988.[32] R.T. Ogden, Essential Wavelets for Statistical Applications and Data Analysis, Birkhauser

Boston, Boston, MA, 1997.[33] J.B. Walsh, An Introduction to Stochastic Partial Differential Equations, Lecture Notes in

8/2/2019 dz02sinum

25/25


Math. 1180, Springer-Verlag, Berlin, 1986, pp. 265439.[34] M. Werner and P. Drummond, Robust algorithms for solving stochastic partial differential

equations, J. Comput. Phys., 132 (1997), pp. 312326.[35] H. Yoo, Semi-discretization of stochastic partial differential equations on R1 by a finite-

difference method, Math. Comp., 69 (2000), pp. 653666.

[36] T. Zhang, Numerical Approximations of Stochastic Partial Differential Equations, M. Philthesis, Hong Kong University of Science and Technology, Hong Kong, 2000.

dz02sinum

Documents