stochastic calc; finance
TRANSCRIPT
-
8/12/2019 Stochastic Calc; Finance
1/23
MIDDLE EAST TECHNICAL UNIVERSITY Ankara, Turkey
INSTITUTE OF APPLIED MATHEMATICS
http://iam.metu.edu.tr
Ito-Taylor Expansions for Systems ofStochastic Differential Equations with
Numerical ApplicationsFikriye Ylmaz 1 , Hacer Oz 2 and Gerhard Wilhelm Weber 3
Abstract. Stochastic differential equations (SDEs) are playing a growing role in
financial mathematics, actuarial sciences, physics, biology and engineering. In this
paper, we focus on a numerical simulation of systems of SDEs based on the stochas-
tic Taylor series expansions. At first, we apply the vector-valued It o formula to the
systems of SDEs, then, the stochastic Taylor formula is used to get the numerical
schemes. In the case of higher dimensional stochastic processes and equations, the
numerical schemes may be expensive and take more time to compute. Iterated
integrals with several n-dimensional Brownian motions can be quite complicated.
Sometimes, such integrals may be approximated. In this paper, we deal with both
the systems with standard n-dimensional Brownian motions and the systems of
SDEs having correlated Brownian motions. The main issue is to transform the
systems of SDEs with correlated Brownian motions to the ones having independent
Wiener processes, and then, to apply the Ito formula to the transformed systems.
Moreover, explicit computations of multiple Ito integrals with several Brownian
motions are given, supported by MATLAB. The paper ends with conclusion and
outlook to future studies.
Keywords. Systems of SDEs, Ito-Taylor expansions, Correlated Brownianmotions, Vector-valued Ito formula, Numerical schemes for SDEs.
Preprint No. 2013-14
March 2013
1Gazi University, 06500 Ankara, Turkey. [email protected] of Applied Mathematics, Middle East Technical University, Ankara, Turkey.
[email protected] of Applied Mathematics, Middle East Technical University, Ankara, Turkey.
-
8/12/2019 Stochastic Calc; Finance
2/23
1 Introduction
There has been a great interest in the simulation methods of SDEs in thefields of, e.g., finance, insurance, medicine and the modern technologies[10, 11, 16, 20, 21]. As the need to take into account of uncertainty is moreand more accepted in science and the applications, SDEs are an emerg-ing subject of interest. Stochastic Taylor expansion provides a source forthe discrete-time approximation methods. One of the simplest ways to dis-cretize the process is Euler method, which approximates the integrals byusing the left-point rule. TheMilsteinscheme (1974), which has the order1.0 of strong convergence, is stronger than Euler method. By adding fur-ther stochastic integrals, with the equations and using the stochastic Taylor
expansion, more accurate schemes can be obtained.
P.E. Kloeden and E. Platen [10] have given a methodical means of derivingthe Taylor series for both Stratonovich and Ito form of a SDE. The applica-tion of Ito Taylor formula to the 1-dimensional SDEs is given explicitly in[10]. In this case, Ii1,i2,...,ik represent the Ito integral, where integration iswith respect to ds ifik = 0, or dWs ifik = 1 [22]. For example,
I010=
t0
s10
s20
ds3dWs2ds1.
Such a display of integrals can be considered as a part of the inner beauty of
stochastic dynamics. This beauty expresses itself in terms of digitalization,algebraization and automization which are not only very aesthetic indeed,but also very practical.
By recursively using the Ito formula, the obtained Taylor series can be re-lated to a tree theory. The tree expansion is given for the true solution in[3]. Runge-Kutta method has been constructed, having order 1.5, in [3, 4, 5].
In the deterministic case, the accuracy of the numerical scheme can be ob-tained by comparing the obtained result with the exact solution. In thestochastic differential equations, there are two ways to measure the accuracy
of the solution: strong and weak convergence. A time-discrete approxima-tion Xh is said to converge stronglywith order p >0 at time Tif there is apositive constantc, independent ofh, and h0> 0 such that
E(|XT Xh(T)|) chp, h (0, h0).
On the other hand, Xh is said to converge weaklywith order p >0 at timeTif there are constants c >0, independent ofh, and h0 > 0 such that
|E(XT) E(Xh(T))| chp, h (0, h0).
2
-
8/12/2019 Stochastic Calc; Finance
3/23
In this paper, in order to get numerical solutions, we shall consider Taylor
schemes that converge strongly.
Vector-valued Ito calculus may involve more than one Wiener process. Al-though it seems that the extension of one-variable SDEs to the multi-valuedSDEs is easy, the computations of iterated Ito integrals are very expensivein the numerical approximations. But, the systems of SDEs model the im-portant cases having several source of randomness and correlation; these aremain reasons of financial risk and instability.
There are not many packages with regards to the simulations of SDEs [7].The Maple package stochasticprovides a symbolic manipulation for SDEs.
The package may be downloaded fromwww.math.uni-frankfurt.de/numerik/kloeden. Many numerical schemes may be generated by using this package.However, the computations of the iterated multi-dimensional Ito integralsare not supported. The software package SDE lab generated by H. Gilsingand T. Shardlow [6] is a good source especially for 1-dimensional SDEs.
In this work, we provide a routine, supported by MATLAB, to compute theiterated multi-dimensional Ito integrals with presence of the formulationsgiven in [10]. We use the Polar Marsaglia methodto generate random vari-ables. This method, that is attributed to G. Marsaglia, is a variation ofthe Box-Muller method. It is based on choosing random points (x, y) in the
square given by1< x
-
8/12/2019 Stochastic Calc; Finance
4/23
ij =
1 ifi = j,0 ifi =j.
We shall call the process Wt:= (W1t, W2t, . . . , W
nt )
T as a correlated Brown-ian motion if
dWit dWjt =ijdt, for i, j= 1, 2, . . . , n ,
for a positive symmetric matrix = (ij)1i,jn satisfying the follows:
ii= 1, and ij =ji [1, 1] for i =j.The paper is organized as follows. Multivariable Ito calculus is reviewed
in Section 2. In Section 3, the systems of SDEs with independent Wienerprocesses and their Taylor expansions are analyzed. Then, the correlatedsystems are covered in Section 4. After obtaining discretization schemesin Section 5, the numerical results and implementation issues are given inSection 6. We conclude and give an outlook to the future studies in Section7.
2 Multi-dimensional Ito Calculus
We consider the process Xt in Rd. Let Zt be a multi-dimensional Brownian
motion, defined asZ
t= (Z
1
t, Z
2
t, . . . , Z
n
t)
T
. Then, the k
th
component of thevector-valued SDE is given by
dXkt =akdt+n
j=1
hkjdZjt , k= 1, 2, . . . , d ,
where ak(t,Xt) and hkj(t,Xt) are the drift and the diffusion coefficients, re-spectively.
We define A(t,Xt) = (a1(t,Xt), . . . , ad(t,Xt))T, Xt= (X1t, . . . , X
dt)
T, and
H(t,Xt) =
h11(t,Xt) . . . h1n(t,Xt)...
. . . ...
hd1(t,Xt) . . . hdn(t,Xt)
,
to get the following matrix formulation:
dXt= Adt+ HdZt. (1)
4
-
8/12/2019 Stochastic Calc; Finance
5/23
3 Ito-Taylor approximation for standard Brown-
ian Motions
In this section, we will assume that the Brownian motions are independent.Eqn. (1) can be written in integral form as
Xt= Xt0+
tt0
A(s,Xs)ds+
tt0
H(s,Xs)dZs. (2)
The second integral is called Ito stochastic integral, which is defined by K.Ito in 1940 [9]. This integral can be approximated by stochastic Taylormethod. Before going into the numerical schemes, we recall Ito Lemma in
several dimensions.
Theorem 3.1 (Ito Lemma in several dimensions, [18])LetXt = (X1t, . . . , X
dt)
T be a vector-valued Ito process satisfying Eqn. (1).Letg : [0,) Rd Rp be a given bounded function in C2([0,) Rd).Then,
dg(t,Xt) =g
tdt+
di=1
g
xidXit +
1
2
di,j=1
2g
xi, xjdXitdX
jt . (3)
We consider now the k th component of the system of SDEs (1):
dXkt =ak(t,Xt)dt+n
j=1
hkj(t,Xt)dZjt ,
where ak(t,Xt) =ak(t, X1t, . . . , X
dt) and hkj(t,Xt) =hkj(t, X
1t, . . . , X
dt).
For applying the Ito formula, we let g(t,Xt) = (g1(t,Xt)), . . . , gp(t,Xt))T
and Yt= g(t,Xt).
Then, the component Ykt is given by
dYkt =gk
t dt+
di=1
gk
xidXit+
1
2
di,j=1
2gk
xixjdXitdX
jt .
Equivalently,
dYkt =
gk
t +
di=1
aigk
xi+
1
2
di,j=1
np=1
hjphip2gk
xixj
dt +
dj=1
np=1
hjpgk
xjdZpt,
5
-
8/12/2019 Stochastic Calc; Finance
6/23
where all derivatives ofgk are to be evaluated in (t,Xt) and Brownian mo-
tions are independent.
The system of SDEs can be classified with respect to the shared states andBrownian motions. We give explicit formulations of each classification inthe following subsections.
3.1 Completely decoupled systems
Now, we apply the theorem to
dXkt =ak(t, Xkt)dt+hkk(t, X
kt)dZ
kt, k= 1, 2, . . . , d ,
where all equations have their own independent Brownian motions and statesthat implies d = n. Then,
dYkt = (gk
t +
di=1
aigk
xi +
1
2
di=1
hiihii2gk
xixi)dt+
di=1
hiigk
xidZit ,
where gk(t,Xt) = (X1t, . . . , X dt)
T. Herewith,
dXkt =ak(t, Xkt)dt+hkk(t, X
kt)dZ
kt. (4)
The integral form of Eqn. (4) is
Xkt =Xkt0
+
tt0
ak(s, Xks )ds+
tt0
hkk(s, Xks )dZ
ks . (5)
Before obtaining the Taylor series expansions, we define the following oper-ators:
L0 :=
t+
di=1
ai
xi+
1
2
di,j=1
np=1
hjphip2
xixj,
and
Lj :=d
p=1
hpj
xp
, for j = 1, 2, . . . , n .
Application of Theorem 3.1 to Eqn. (5) gives
Ykt =Ykt0
+
tt0
L0gkds+n
j=1
tt0
LjgkdZjs . (6)
6
-
8/12/2019 Stochastic Calc; Finance
7/23
In Eqn. (5), firstly, gk is chosen as ak(t, Xkt) and then, the Ito formula (6)
is applied. Secondly, we let gk =hkk(t, Xkt) and use the Ito formula to get
Xkt =Xkt0
+
tt0
ak(t0, X
kt0
) +
st0
L0ak(, Xk)d
+n
j=1
st0
Ljak(, Xk)dZj
ds+
tt0
hkk(t0, X
kt0
)
+
st0
L0hkk(, Xk)d+n
j=1
st0
Ljhkk(, Xk)dZj
dZks ,
=Xkt0 + ak(t0, Xkt0
)I0+hkk(t0, Xkt0
)Ik
+
t
t0
st0
L0ak(, Xk)d+ nj=1
s
t0
Ljak(, Xk)dZj
ds (7)
+
tt0
st0
L0hkk(, Xk)d+n
j=1
st0
Ljhkk(, Xk)dZj
dZks .
We can continue with an application of Ito formula (6) to the functions L0ak,nj=1Ljak,L0hkkand
nj=1Ljhkkin (7), and then, to the functions L0L0ak,n
j=1LjL0ak,n
j=1L0Ljak,n
j,p=1LjLpak,L0L0hkk,n
j,p=1LjL0hkk,nj,p=1L0Ljhkk,
nj,p=1LjLphkk, to obtain Ito-Taylor expansion:
Xkt =Xkt0
+ ak(t0, Xkt0)I0+hkk(t0, Xkt0
)Ik
+ L0ak(t0, Xkt0)I00+n
j=1
Ljak(t0, Xkt0)Ij0
+ L0hkk(t0, Xkt0)I0k+n
j=1
Ljhkk(t0, Xkt0)Ijk
+ L0L0ak(t0, Xkt0)I000+n
j=1
LjL0ak(t0, Xkt0)Ij00
+
n
j=1
L0
Lj
ak(t0, Xk
t0)I0j0+
n
j,p=1
Lj
Lp
ak(t0, Xk
t0)Ijp0
+ L0L0hkk(t0, Xkt0)I00k+n
j,p=1
LjL0hkk(t0, Xkt0)Ij0k
+n
j,p=1
L0Ljhkk(t0, Xkt0)I0jk +n
j,p=1
LjLphkk(t0, Xkt0)Ijpk +R,
where R denotes the remainder term and Ii1,i2,...,ik represent the multiple
7
-
8/12/2019 Stochastic Calc; Finance
8/23
Ito integral, which is defined as [10]:
I =
1, ifk = 0,tt0
Ids, ifk 1 and k = 0,tt0
IdZks , ifk 1 and k 1,
for = (1, 2, . . . , k)T (N0)k with k 2, where , the multi index,
can be obtained by deleting the last component of, i.e., integration is withrespect to ds ifik = 0, or dZ
js ifik = j= 0, j = 1, . . . , n.
3.2 Systems with common states
Now, we consider the following system
dXkt =ak(t,Xt) +hkk(t,Xt)dZkt, k= 1, 2, . . . , d , (8)
where all equations may have all states in common but they are not sharingBrownian motions.It has a similar formulation with completely decoupled systems
Xkt =Xkt0
+ ak(t0,Xt0)I0+hkk(t0,Xt0)Ik
+ L0ak(t0,Xt0)I00+n
j=1
Ljak(t0,Xt0)Ij0
+ L0hkk(t0,Xt0)I0k+n
j=1
Ljhkk(t0,Xt0)Ijk
+ L0L0ak(t0,Xt0)I000+n
j=1
LjL0ak(t0,Xt0)Ij00
+
nj=1
L0Ljak(t0,Xt0)I0j0+n
j,p=1
LjLpak(t0,Xt0)Ijp0
+
L0
L0hkk(t0,Xt0)I00k+
n
j,p=1L
j
L0hkk(t0,Xt0)Ij0k
+
nj,p=1
L0Ljhkk(t0,Xt0)I0jk +n
j,p=1
LjLphkk(t0,Xt0)Ijpk +R,
where R denotes the remainder term.
3.3 General case
We apply the Theorem 3.1 to the following system:
8
-
8/12/2019 Stochastic Calc; Finance
9/23
dXkt =ak(t,Xt) +n
j=1
hkj(t,Xt)dZjt , k= 1, 2, . . . , d , (9)
where all equations may share both Brownian motions and states.
Then, similarly,
Xkt =Xkt0
+
tt0
ak(s,Xs)ds+n
j=1
tt0
hkj(s,Xs)dZjs . (10)
Again, we first choosegk :=ak(t,Xt) to apply the Ito formula, and then, welet gk :=hkj(t,Xt) to get
Xkt =Xkt0
+
tt0
ak(t0,Xt0) +
st0
L0ak(,X)d
+n
l,j=1
st0
Ljak(, X)dZj
ds+n
j=1
tt0
hkj(t0,Xt0)
+
st0
L0hkj(,X)d+n
l=1
st0
Llhkj(, X)dZl
dZjs ,
We do the same things as in the completely decoupled case, but only differ-ence is that ak and hkj depend on (t,Xt). Then,
Xkt =Xkt0
+ ak(t0,Xt0)I0+n
j=1
hkj(t0,Xt0)Ij
+ L0ak(t0,Xt0)I00+n
j=1
Ljak(t0,Xt0)Ij0
+n
j=1
L0hkj(t0,Xt0)I0j+n
l,j=1
Llhkj(t0,Xt0)Ilj
+ L
0
L0ak(t0,Xt0)I000+
n
j=1L
j
L0ak(t0,Xt0)Ij00 (11)
+n
j=1
L0Ljak(t0,Xt0)I0j0+n
j,p=1
LjLpak(t0,Xt0)Ijp0
+n
j=1
L0L0hkj(t0,Xt0)I00j+n
j,l=1
LlL0hkj(t0,Xt0)Il0j
+n
j,p,l=1
L0Llhkj(t0,Xt0)I0lj +n
j,p,l=1
LlLphkj(t0,Xt0)Ilpj+R.
9
-
8/12/2019 Stochastic Calc; Finance
10/23
4 Ito-Taylor approximation for Correlated Brow-
nian Motions
In this section, we assume that Brownian motions are correlated, so weuse W instead ofZ in order to point out the difference from the standardBrownian motions. Now, we consider the system
dXkt =ak(t,Xt) +
n1
hkj(t,Xt)dWjt, k= 1, 2, . . . , d= n,
where dWit dWjt =ijdt.
Then, we write the correlation matrix as:
=
1 12 . . . 1n21 1 . . . 2n
... ...
. . . ...
n1 n2 . . . 1
, ij =ji [1, 1].
Here, is positive symmetric matrix which means that = T and
ni,j=1
ijxixj 0,
for all x= (x1, . . . , xn)T Rn.
By using standard Linear Algebra, one can find an n n matrix B =(bij)1i,jn such that
= BBT.
Moreover, using Cholesky Decomposion, we can take B as an upper (orlower) triangular matrix.
Correlated Brownian motions can be interpreted as
Wt= BZt,
where Wt = (W1t, . . . , W
nt )
T is a standard n-dimensional Brownian motionand the Brownian motions, Wit for i = 1, 2, . . . , n, are correlated.
In componentwise, notation,
Wit =n
j=1
bijZjt , i= 1, 2, . . . , n .
10
-
8/12/2019 Stochastic Calc; Finance
11/23
Example 1. We consider 2-dimensional version ofweakly-coupled Ornstein-
Uhlenbeck model:
dX1t =1(1 X1t)dt+1dW1t,dX2t =2(2 X2t)dt+2dW2t, (12)
wheredW1tdW2t =dt,all coefficients are real and positive, and (1, 1).
Then, the correlation matrix can be written as:
=
1 1
=
1 0
1 2
1
0
1 2
,
by Cholesky Decomposition [2].
Thus,
Wt = BZt=
1 0
1 2
Z1tZ2t
.
So, Eqn. (12) becomes
dX1t =1(1 X1t)dt+1dZ1t,dX2t =2(2 X2t)dt+2dZ1t +2
1 2dZ2t. (13)
In integral form,
X1t =X1t0
+ 11
tt0
ds 1 tt0
X1s ds+1
tt0
dZ1s ,
X2t =X2t0
+ 22
tt0
ds 2 tt0
X2s ds+2
tt0
dZ1s +2
1 2 tt0
dZ2s .
Applying the procedure in the previous sections, we get
X1t =X1t0
+ 1(1 X1t0)I0+1I1 21(1 X1t0)I00 11I10+31(1 X1t0)I000+211I100+R,
X2t =X2t0 + 2(2 X2t0)I0+2I1+21 2I2 22(2 X2t0)I00 22I10 22
1 2I20
+ 32(2 X2t0)I000+222I100+
1 2222I200+R.
Example 2. We state the 2-dimensional version ofstrongly-coupled Ornstein-Uhlenbeck model
dX1t = (11X1t 12X2t)dt+1dW1t,dX2t = (21X1t 22X2t)dt+2dW2t, (14)
11
-
8/12/2019 Stochastic Calc; Finance
12/23
where dW1tdW2t =dt and 1, 2= 0.
We obtain the following system in terms of standard Brownian motions:
dX1t = (11X1t 12X2t)dt+1dZ1t,dX2t = (21X1t 22X2t)dt+2dZ1t +2
1 2dZ2t. (15)
In a similar way, we obtain:
X1t =X1t0 (11X1t0+12X2t0)I0+1I1 (111+212)I10
+ 11(11X1t0
+ 12X2
t0) +12(21X
1
t0+22X
2
t0)I00
212
1 2I20
(211+1221)(11X1t0
+12X2t0
)
+ (1112+1222)(21X1t0
+ 22X2t0
)
I000
+
1(
211+1221) +2(1112+1222)
I100
+ 2(1112+1222)
1 2I200+R,
and
X2t =X2t0 (21X1t0+ 22X2t0)I0+2I1+2
1 2I2
(121+222)I10+
21(11X1t0
+12X2t0
)
+ 22(21X1t0
+ 22X2t0
)
I00 222
1 2I20 [(2111+2221)(11X1t0+12X2t0) + (2112+ 222)(21X
1t0
+ 22X2t0
)]I000+ [1(1121+2221)
+ 2(2112+222)]I100+2(1221+
222)
1 2I200+R.
12
-
8/12/2019 Stochastic Calc; Finance
13/23
5 Discretization schemes with Strong Taylor ap-
proximations
We represent some numerical approximations of Ito integrals. Let denotethe increments, then:
Ij =
n+1n
dZjs = Zj =Zjn+1Zjn ,
Ij0=
n+1n
sn
dZjuds= W ,
I0j =
n+1n
sn
dudZjs = (Zj) W ,
Ijj = 12
(Zj)2 ,
Ijj0= I0jj =Ij0j =1
6
(Zj)2 ,Ijjj =
1
6
(Zj)3 3(Zj),Ij00= I0j0= I00j =
1
62Zjs ,
where the Wand Zj are Gaussian random variables with Zj N(0, ),W N(0, 1
33) and E
ZjW) = 1
22.
Now, we shall use the above relations to propose some strong approxima-tions.
5.1 The Euler-Maruyama Scheme
The simplest example of a strong Taylor approximation Y of the solutionof Eqn. (9) is the Euler-Maruyama or Euler method attaining the order ofstrong convergence 0.5. The kth component of the Euler scheme is of theform
Ykn+1= Ykn +ak +
nj=1
hkjZj, for k = 1, 2, . . . , d . (16)
In the cases where drift and difussion coefficients are nearly constant, thismethod generally gives us good numerical results. However, when the coeffi-cients are nonlinear the method can provide a poor estimate of the solution.So, higher-order schemes should be used to obtain more satisfactory schemes.
Let us consider the system (15) by taking 11 = 21 = 22 = 1, 12 =2, 1 = 2= 1 and = 0.6 to get
dX1t = (X1t 2X2t)dt+dZ1t,dX2t = (X1t X2t)dt+ 0.6dZ1t + 0.8dZ2t. (17)
13
-
8/12/2019 Stochastic Calc; Finance
14/23
Euler scheme reads the system (17) as:
Y1n+1= Y1n + (Y2n 2Y1n ) + Z1,
Y2n+1= Y2n + (Y2n Y1n ) + 0.6Z1 + 0.8Z2.
5.2 The Milstein Scheme
Milstein scheme has the order of strong convergence 1.0. By including an ad-ditional term from Ito-Taylor expansion (11), we obtain theMilstein schemewhose k th component is of the form
Ykn+1= Ykn +ak +
n
j=1
hkjZj +
n
j1,j2=1
Lj1hkj2Ij1j2, for k = 1, 2, . . . , d .
(18)We note that Milstein scheme is identical to the Euler scheme when thediffusion term does not contain an Xt term.
Applying the scheme to the system (17), we obtain the Milstein scheme:
Y1n+1= Y1n + (Y2n 2Y1n ) + Z1,
Y2n+1= Y2n + (Y2n Y1n ) + 0.6Z1 + 0.8Z2.
5.3 The Order 1.5 Strong Taylor Scheme
We can get more accurate strong Taylor schemes by including further multi-ple stochastic integrals from the stochastic Taylor approximation (11). Thekth component ofthe order 1.5 strong Taylor schemeis given by
Ykn+1= Ykn + ak +
1
2L0ak
2 +n
j=1
(hkjZj +L0hkjI0j+L
jakIj0
+n
j1,j2=1
Lj1hkj2Ij1j2+n
j1,j2,j3=1
Lj1Lj2hkj3Ij1j2j3 (19)
for k = 1, 2, . . . , d .
Applying this method to the system (17), we get:
Y1n+1= Y1n + (Y2n 2Y1n ) +
1
2(3Y2n + 5Y
1n )
2 + Z1 2.6I10 0.8I20,
Y2n+1= Y2n + (Y2n Y1n ) +
1
2(2Y2n + 3Y
1n )
2 + 0.6Z1
+0.8Z2 1.6I10 0.8I20.
14
-
8/12/2019 Stochastic Calc; Finance
15/23
5.4 The Order 2.0 Strong Taylor Scheme
The numerical results which was given by the order 2.0 strong Taylor schemeis better than other three method. Thekth component ofthe order 2.0 strongTaylor schemetakes the form
Ykn+1= Ykn + ak +
1
2L0ak
2 +n
j=1
(hkjZj +L0hkjI0j+L
jakIj0)
+
nj1,j2=1
Lj1hkj2Ij1j2+ L
0Lj1hkj2I0j1j2+Lj1L0hkj2Ij10j2
+ Lj1Lj2akIj1j20+n
j1,j2,j3=1
Lj1Lj2hkj3Ij1j2j3 (20)
+n
j1,j2,j3,j4=1
Lj1Lj2Lj2hkj4Ij1j2j3j4 ,
for k = 1, 2, . . . , d.
Finally, for the Eqn. (17) this formulation can be reduced to:
Y1n+1= Y1n + (Y2n 2Y1n ) +
1
2(3Y2n + 5Y
1n )
2 + Z1 2.6I10 0.8I20,
Y2n+1= Y
2n + (Y
2n Y
1n ) +
1
2 (2Y2n + 3Y
1n )
2
+ 0.6Z1
+0.8Z2 1.6I10 0.8I20.
6 Numerical Results and Implementation Details
In this section, we consider numerical examples for both the systems withindependent and correlated Brownian motions. In order to implement thediscrete scheme, we use MATLAB. There are many documentations thatdescribes the main features of MATLAB commands related to SDE. Somenumerical interpretations can be found in the SDEs MATLAB packages [7].
However, numerical examples implemented in MATLAB are mostly in the1-dimensional case. Some coupled SDEs are considered, but having sym-metric coefficients allowing easy computations that arise from multiple It ointegrals. Our first example demonstrates the triple SDEs having indepen-dent Brownian motions.
Example Run 1. The system of SDE consisting of three equationsproposed in Hofmann, Platen and Schweizer [8] is considered as:
15
-
8/12/2019 Stochastic Calc; Finance
16/23
dX
1
t =X1
tX2
tdZ1
t,dX2t = (X2t X3t)dt+ 0.3X2tdZ2t,dX3t =
1
(X2t X3t)dt,
where X1t, X2t and X
3t represent the asset price, the instantenous volatil-
ity, and the averaged volatility, respectively. As in [7], Milstein scheme isobtained as:
dX1n+1= X1n + X
1nX
2nZ
1n+
1
2X1n(X
2n)
2{(Z1n)2 }
+ 0.3X1nX2n
tn+1
tn
t
tn
dZ2sdZ1s ,
dX2n+1= X2n (X2n X3n) + 0.3X2nZ2n+ 0.045X2n{(Z1n)2 },
dX3n+1= dX3n +
1
(X2n X3n).
We take = 1, X10 = 1, X20
= 0.1, X20 = 0.1 and T= 1; is considered as29.
The scheme has the double integraltn+1tn
ttn
dZ2sdZ1s . In [7], this integral is
approximated by Euler method. Although it is a bit challenging, we computesuch integrals by using the following formulations from [10] as follows:
Ip0
= , Ipj =
j, Ip00
= 1
22,
Ipj0=1
2
j+aj0
, Ip0j =
1
2
j aj0
.
Here,
aj0=
1
2
p
r=1
1
r
jr
2pjp ,
Ipj1j2 =1
2j1j2
1
2
(aj20j1 aj10j2) + Apj1j2,
Apj1j2 = 1
2
pr=1
1
r(j1rj2r j2rj1r),
with
16
-
8/12/2019 Stochastic Calc; Finance
17/23
j = 1
Wj, jr =
2
rajr , jr =
2
rbjr ,
jp = 1
p
r=p+1
ajr , p = 1
12 1
22
pr=1
1
r2,
where j = 1, 2, . . . , mand r = 1, 2, . . . , p, for a positive number p chosen asfollows:
p= p() K2
,
for an appropriate positive constant K to ensure the convergence order ofthe numerical scheme.
Here, we note thatjr , jr and jp are independent Gaussian random vari-ables. We use the Polar Marsaglia Methodto generate the pairs of randomvariables.The following lines show the implementation of this method in MATLAB:
%Polar Marsaglia Method
function [z1,z2]= Polar
l=0.5;while l>0
u1 = rand;u2 = rand;
v1 = 2*u1 - 1;v2 = 2*u2 - 1;
V = (v1.*v1)+(v2.*v2);
if (V0)
break;
end
end
z1 = v1.*sqrt(-2*log(V)./V);
z2 = v2.*sqrt(-2*log(V)./V);
We can compute the iterated integraltn+1tn
ttn
dZ2s dZ1s as:
%Approximation of I_ij
function I_ij=ito_ij(p,Delta,G1,G2,mu1_j,mu2_j,ro)
a_ij=0;
for i=1:p
[zeta1, zeta2]= Polar;[eta1 ,eta2 ]=Polar;
a_ij=a_ij+(1/i)*(zeta1*(sqrt(2)*G2+eta2)...
17
-
8/12/2019 Stochastic Calc; Finance
18/23
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
1.4
t
X
X1
X2
X3
Figure 1: Numerical result of Example Run1 with Milstein approximation.
-zeta2*(sqrt(2)*G1+eta1));
end
I_ij=a_ij*Delta/(pi);
I_ij=I_ij+Delta*(G1*G2/2+sqrt(ro)*(mu1_j*G2-mu2_j*G1));
In Figure 1, we give the numerical result of Example Run1. We have thesame results as in [7].
Example Run 2. (Correlated Brownian motions) We recall the strongly-coupled Ornstein-Uhlenbeck process (14):
dX1
t = (11X1
t 12X2
t)dt+1dW1
t,dX2t = (21X1t 22X2t)dt+2dW2t,
where dW1tdW2t =dt and the transformed form (17) is
dX1t = (X1t 2X2t)dt+dZ1t,dX2t = (X1t X2t)dt+ 0.6dZ1t + 0.8dZ2t.
Our Taylor scheme with order 1/2 gives
18
-
8/12/2019 Stochastic Calc; Finance
19/23
Y1n+1= Y1n + (Y2n 2Y1n ) +
1
2(3Y2n + 5Y
1n )
2 + Z1 2.6I10 0.8I20,
Y2n+1= Y2n + (Y2n Y1n ) +
1
2(2Y2n + 3Y
1n )
2 + 0.6Z1
+0.8Z2 1.6I10 0.8I20.
We compute the integrals I20, I10 as in the following lines:
%Approximation of I_j0 and I_0j
function [I_10,I_20]=ito_j0(p,Delta,G1,G2,mu1_j,mu2_j,ro)
a_10=0;a_20=0;for i=1:p
[eta1,eta2]=Polar;
a_10=a_10+(1/i)*eta1;
a_20=a_20+(1/i)*eta2;
end
I_10=a_10*(1/pi)*sqrt(Delta*2)+2*sqrt(Delta*ro)*mu1_j;
I_10=(1/2)*Delta*I_10+(1/2)*Delta*sqrt(Delta)*G1;
I_20=a_20*(1/pi)*sqrt(Delta*2)+2*sqrt(Delta*ro)*mu2_j;
I_20=(1/2)*Delta*I_20+(1/2)*Delta*sqrt(Delta)*G2;
The main file can be run as:
clf
randn(state,1)
T = 1; Delta = 2^(-9); delta = Delta^2;
L = T/Delta; K = Delta/delta;
X1 = zeros(1,L+1); X2 = zeros(1,L+1);
X1(1) = 1;X2(1) = 0.1;
p=2;ro=0;
for i=1:p
ro=ro+1/(i*i);ro=(pi*pi)/6-ro;ro=ro/(2*pi*pi);
endfor j = 1:L
G1 = randn; G2 = randn;
Winc2 = sqrt(Delta)*G2;Winc1 = sqrt(Delta)*G1;
[mu1, mu2 ]=Polar;
[I10,I20]=ito_j0(p,Delta,G1,G2,mu1,mu2,ro);
X1(j+1) = X1(j) +(-2*X1(j)-X2(j))*Delta +...
(0.5)*(3*X2(j)+5*X1(j))*Delta^2+Winc1-(2.6)*I10-(0.8)*I20;
X2(j+1) = X2(j) + (-X2(j)-X1(j))*Delta+(0.5)*(2*X2(j)...
19
-
8/12/2019 Stochastic Calc; Finance
20/23
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
ItoTaylor approximation for X1
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
Exact numerical simulation for X1
Figure 2: Comparison of exact numerical solution of Run2 with TaylorScheme of order 1/2 for X1.
+3*X1(j))*Delta^2+ (0.6)*Winc1 +(0.8)*Winc2-(1.6)*I10-(0.8)*I20;
end
plot([0:Delta:T],X1,r-), hold on
plot([0:Delta:T],X2,bl--)
xlabel(t,FontSize,16), ylabel(X,FontSize,16)
legend(X^1,X^2)
Exact solution of the system (14) is obtained in matrix formulation as:
Xt = X0exp(tA) + t0
exp((s t)A)BdZt, (21)
where A = 1 2
1 1
, and B = 1 0
0.6 0.8
.
We perform the exact numerical simulation for the system (14). We firstcompute the matrix multiplications in the Eqn. (21), and then approximatecomponentwise. In Figures 2-3, we compare the obtained results. It can beseen easily that they are almost the same.
20
-
8/12/2019 Stochastic Calc; Finance
21/23
0 0.2 0.4 0.6 0.8 11.5
1
0.5
0
0.5
ItoTaylor approximation for X2
0 0.2 0.4 0.6 0.8 11.5
1
0.5
0
0.5
Exact numerical simulation for X2
Figure 3: Comparison of exact numerical solution of Run2 with TaylorScheme of order 1/2 for X2.
7 Conclusion and Outlook
This work have totally covered the numerical solutions of SDEs by means ofIto expansions. We obtained the stochastic Taylor series expansions of thesystems os SDEs with both correlated and independent Wiener processes.In the case of correlated processes, we first transformed the system into onehaving independent Brownian motions. And then, we obtained the Ito Tay-lor series expansions. Also, we performed the computations of the iteratedIto integrals with several Wiener processes by using MATLAB.
As a future work, the relation between the Ito Taylor approximations andstochastic control problems can be considered. Furthermore, as for ap-
plication to the more complicated systems of SDEs, e.g., Heston model,may be handled by means of Ito Taylor formula. A connection betweenthe iterated Ito integrals and Malliavin calculus should be investigated[1, 12, 13, 14, 15, 17, 19]. A detailed package for the systems of SDEs,which has a higher-order convergence, can be generated via MATLAB.
References
[1] Bell, D., 2007, The Malliavin Calculus, Dover.
21
-
8/12/2019 Stochastic Calc; Finance
22/23
[2] Brummelhuis, R. Mathematical Methods Lecture Notes 6, Depart-
ment of Economics, Mathematics and Statistics, Birkbeck, Universityof London, 2009.
[3] Burrage, K. and Burrage, P.M., 1999, High strong order methods fornon-commutative stochastic ordinary differential equation system andthe Magnus formula, Physica D 133, 34-48.
[4] Burrage, K. and Burrage, P.M., 2000, Order conditions of stochasticRunge-Kutta methods by B-series, SIAM J. Numer. Anal. 38, 1626-1646.
[5] Burrage, P.M., 1999, Runge-Kutta Methods for Stochastic Differential
Equations, PhD Thesis, Department of Mathematics, University ofQueensland, Australia.
[6] Gilsing, H., and Shardlow, T., SDELab: A package for solving stochas-tic differential equations in MATLAB, Journal of Computational andApplied Mathematics 205 (2007), no. 2, 1002-1018.
[7] Higham, D. J., and Kloeden, P. E., MAPLE and MATLAB for Stochas-tic Differential Equations in Finance, 2002, in Programming Languagesand Systems in Computational Economics and Finance, 233-270.
[8] Hoffman, N., Platen, E., and Schweizer, M., 1992, Option pricing under
incompleteness and stochastic volatility, J. Mathematical Finance, 2,153-187.
[9] Ito K., 1944, Stochastic Integral, Proc. Imperial Acad. Tokyo, 20, 519-524
[10] Kloeden, P.E. and Platen, E., 1992, Numerical Solution of StochasticDifferential Equations, Springer-Verlag, Berlin.
[11] Kunita, H., 1990, Stochastic Flows and Stochastic Differential Equa-tions, Vol 24 of Cambridge Studies in Advanced Mathematics, Cam-bridge University Press.
[12] Kusuoka, S. and Stroock, D., 1982, Applications of Malliavin CalculusI, Stochastic Analysis, Proceedings Taniguchi International Sympo-sium Katata and Kyoto, 271-306.
[13] Kusuoka, S. and Stroock, D., 1985, Applications of Malliavin CalculusII, J. Faculty Sci. Uni. Tokyo Sect. 1A Math., 32, 1-76.
[14] Kusuoka, S. and Stroock, D., 1987, Applications of Malliavin CalculusIII, J. Faculty Sci. Univ. Tokyo Sect. 1A Math., 34, 391-442.
22
-
8/12/2019 Stochastic Calc; Finance
23/23
[15] Malliavin, P., Thalmaier, A., 2005, Stochastic Calculus of Variations
in Mathematical Finance, Springer.
[16] Milstein, G.N., 1995, Numerical Integration of Stochastic DifferentialEquations, Kluwer, Dordrecht/Boston/London.
[17] Nualart, D., 2006, The Malliavin calculus and related topics (Secondedition ed.), Springer-Verlag.
[18] ksendal, B., 2000, Stochastic Differential Equations, An Introductionwith Applications, 5th edition
[19] ksendal, B., 1997, An Introduction To Malliavin Calculus With Ap-
plications To Economics, Thesis, Dept. of Mathematics, University ofOslo.
[20] Platen, E., 1999, An introduction to numerical methods for stochasticdifferential equations, Acta Numerica, 8, 197-246.
[21] Rumelin, W., 1982, Numerical treatment of stochastic differentialequations, SIAM J. Numer. Anal. 19, 604-613.
[22] Ylmaz, F., Oz, H., Weber, G.W., 2012, Change of Time Methodand Stochastic Taylor Expansion with Computation of Expectation,Book Chapter, Modeling, Optimization, Dynamics and Bioeconomy,
Springer, to appear.
23