本文的目的在于提供0.5,1和1.5阶强SDE数值格式的推导和内容,所有推导基于 I t o − T a y l o r Ito-Taylor Ito−Taylor展开,由于国内外网站缺少关于强SDE数值阶的总结,笔者在此特作总结,为使用SDE数值格式的读者提供帮助,本文需要读者预先已经知道了有关Brown-motion的基本性质和随机微分的基本性质,否则格式推导会看不懂。如果想要实际应用直接看数值格式即可。
1. SDE强收敛阶判断定理
本文考虑如下SDE:
d
X
=
b
(
x
)
d
t
+
σ
(
x
)
d
W
.
dX=b(x)dt+\sigma(x)dW.
dX=b(x)dt+σ(x)dW.
使用数值SDE在时间
t
t
t下以
x
x
x为初值的近似解定义为
X
‾
t
,
x
(
t
+
h
)
\overline{X}_{t,x}(t+h)
Xt,x(t+h),其中
t
+
h
≥
t
t+h\geq t
t+h≥t.真解为:
X
t
,
x
(
t
+
h
)
X_{t,x}(t+h)
Xt,x(t+h).若满足如下定理:
定理1:
若存在
p
1
,
p
2
p_1,p_2
p1,p2,满足:
E
∣
X
t
,
x
(
t
+
h
)
−
X
‾
t
,
x
(
t
+
h
)
∣
≤
k
(
1
+
∣
x
∣
2
)
h
p
1
.
E |X_{t,x}(t+h)-\overline{X}_{t,x}(t+h)|\leq k(1+|x|^2)h^{p_1}.
E∣Xt,x(t+h)−Xt,x(t+h)∣≤k(1+∣x∣2)hp1.
[
E
∣
X
t
,
x
(
t
+
h
)
−
X
‾
t
,
x
(
t
+
h
)
∣
2
]
1
2
≤
k
(
1
+
∣
x
∣
2
)
1
2
h
p
2
.
[E |X_{t,x}(t+h)-\overline{X}_{t,x}(t+h)|^2]^{\frac{1}{2}}\leq k(1+|x|^2)^{\frac{1}{2}}h^{p_2}.
[E∣Xt,x(t+h)−Xt,x(t+h)∣2]21≤k(1+∣x∣2)21hp2.
且满足
p
1
≥
p
2
+
1
2
p_1\geq p_2+\frac{1}{2}
p1≥p2+21,
p
2
≥
1
2
p_2\geq\frac{1}{2}
p2≥21.则可认为强收敛阶为
p
2
−
1
2
p_2-\frac{1}{2}
p2−21.
该定理的证明笔者在此不做推导,感兴趣的读者自行阅读其他文献或教材,这一般都是有的,这里我们直接利用它进行以下的
I
t
o
−
T
a
y
l
o
r
Ito-Taylor
Ito−Taylor展开。
2. I t o − T a y l o r Ito-Taylor Ito−Taylor展开
我们考虑:
d
x
=
b
(
x
)
d
t
+
σ
(
x
)
d
W
.
dx=b(x)dt+\sigma(x)dW.
dx=b(x)dt+σ(x)dW.
若有一个非常光滑的
f
(
x
)
f(x)
f(x),根据
I
t
o
Ito
Ito公式,会有:
d
f
(
x
)
=
[
b
(
x
)
f
x
(
x
)
+
1
2
σ
2
f
x
x
(
x
)
]
d
t
+
[
σ
(
x
)
f
x
(
x
)
]
d
W
.
df(x)=[b(x)f_x(x)+\frac{1}{2} \sigma^2 f_{xx}(x)]dt+[\sigma(x)f_x(x)]dW.
df(x)=[b(x)fx(x)+21σ2fxx(x)]dt+[σ(x)fx(x)]dW.
这两个微分算子可以被定义为:
L
0
=
b
(
x
)
d
d
x
+
1
2
σ
2
(
x
)
d
2
d
x
2
L^0=b(x)\frac{d}{dx}+\frac{1}{2}\sigma^2(x)\frac{d^2}{dx^2}
L0=b(x)dxd+21σ2(x)dx2d2
L
1
=
σ
(
x
)
d
d
x
L^1=\sigma(x)\frac{d}{dx}
L1=σ(x)dxd
则对两边取积分会有:
f
(
x
t
)
=
f
(
x
t
0
)
+
∫
t
0
t
L
0
f
(
x
s
)
d
s
+
∫
t
0
t
L
1
f
(
x
s
)
d
W
s
.
f(x_t)=f(x_{t0})+\int_{t_0}^t L^0f(x_s)ds+\int_{t_0}^tL^1f(x_s)dW_s.
f(xt)=f(xt0)+∫t0tL0f(xs)ds+∫t0tL1f(xs)dWs.
注意到:
∫
t
0
t
L
0
f
(
x
s
)
d
s
=
L
0
f
(
x
t
0
)
(
t
−
t
0
)
+
∫
t
0
t
[
L
0
f
(
x
s
)
−
L
0
f
(
x
t
0
)
]
.
\int_{t_0}^t L^0f(x_s)ds=L^0 f(x_{t_0})(t-t_0)+\int_{t_0}^t[L^0f(x_s)-L^0f(x_{t_0})].
∫t0tL0f(xs)ds=L0f(xt0)(t−t0)+∫t0t[L0f(xs)−L0f(xt0)].
∫
t
0
t
L
1
f
(
x
s
)
d
W
s
=
L
1
f
(
x
t
0
)
(
W
t
−
W
t
0
)
+
∫
t
0
t
[
L
1
f
(
x
s
)
−
L
1
f
(
x
t
0
)
]
d
W
s
.
\int_{t_0}^t L^1f(x_s)dWs=L^1 f(x_{t_0})(W_t-W_{t_0})+\int_{t_0}^t[L^1f(x_s)-L^1f(x_{t_0})]dWs.
∫t0tL1f(xs)dWs=L1f(xt0)(Wt−Wt0)+∫t0t[L1f(xs)−L1f(xt0)]dWs.
注意到:
∫
t
0
t
[
L
0
f
(
x
s
)
−
L
0
f
(
x
t
0
)
]
=
∫
t
0
t
∫
t
0
s
d
[
L
0
f
(
x
s
1
)
]
d
s
\int_{t_0}^t[L^0f(x_s)-L^0f(x_{t_0})]=\int_{t_0}^t\int_{t_0}^s d[L^0f(x_{s_1})]ds
∫t0t[L0f(xs)−L0f(xt0)]=∫t0t∫t0sd[L0f(xs1)]ds
∫
t
0
t
[
L
1
f
(
x
s
)
−
L
1
f
(
x
t
0
)
]
=
∫
t
0
t
∫
t
0
s
d
[
L
1
f
(
x
s
1
)
]
d
W
s
\int_{t_0}^t[L^1f(x_s)-L^1f(x_{t_0})]=\int_{t_0}^t\int_{t_0}^s d[L^1f(x_{s_1})]dWs
∫t0t[L1f(xs)−L1f(xt0)]=∫t0t∫t0sd[L1f(xs1)]dWs
使用
I
t
o
Ito
Ito公式我们可以知道:
d
[
L
0
f
(
x
)
]
=
[
L
0
L
0
f
]
d
t
+
[
L
1
L
0
f
]
d
W
.
d[L^0f(x)]=[L^0L^0f]dt+[L^1L^0f]dW.
d[L0f(x)]=[L0L0f]dt+[L1L0f]dW.
d
[
L
1
f
(
x
)
]
=
[
L
0
L
1
f
]
d
t
+
[
L
1
L
1
f
]
d
W
.
d[L^1f(x)]=[L^0L^1f]dt+[L^1L^1f]dW.
d[L1f(x)]=[L0L1f]dt+[L1L1f]dW.
我们便可以得到:
f
(
x
t
)
=
f
(
x
t
0
)
+
L
0
f
(
x
t
0
)
(
t
−
t
0
)
+
L
1
f
(
x
t
0
)
(
W
t
−
W
t
0
)
+
R
‾
f(x_t)=f(x_{t0})+L^0 f(x_{t_0})(t-t_0)+L^1 f(x_{t_0})(W_t-W_{t_0})+\overline{R}
f(xt)=f(xt0)+L0f(xt0)(t−t0)+L1f(xt0)(Wt−Wt0)+R
其中:
R
‾
=
∫
t
0
t
∫
t
0
s
L
0
L
0
f
(
s
1
)
d
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
1
L
0
f
(
s
1
)
d
W
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
0
L
1
f
(
s
1
)
d
s
1
d
W
s
+
∫
t
0
t
∫
t
0
s
L
1
L
1
f
(
s
1
)
d
W
s
1
d
W
s
\overline{R}=\int_{t_0}^t\int_{t_0}^s L^0L^0f(s_1)ds_1ds+\int_{t_0}^t\int_{t_0}^s L^1L^0f(s_1)dWs_1ds+\int_{t_0}^t\int_{t_0}^s L^0L^1f(s_1)ds_1dWs+\int_{t_0}^t\int_{t_0}^s L^1L^1f(s_1)dWs_1dWs
R=∫t0t∫t0sL0L0f(s1)ds1ds+∫t0t∫t0sL1L0f(s1)dWs1ds+∫t0t∫t0sL0L1f(s1)ds1dWs+∫t0t∫t0sL1L1f(s1)dWs1dWs
当然可以继续展开下去。
我们在这里只关心
f
(
x
)
=
x
f(x)=x
f(x)=x的情况,因此这会导出本文的主题0.5(Euler-Maruyama), 1(Milstein), 和1.5 阶强Stochastic Differential Equation格式。
3、强格式1: 0.5阶强SDE (Euler-Maruyama)
令
f
(
x
)
=
x
f(x)=x
f(x)=x,显然地,我们会得到:
L
0
f
=
L
0
x
=
b
(
x
)
L^0f=L^0x=b(x)
L0f=L0x=b(x),
L
1
f
=
L
1
x
=
σ
(
x
)
L^1f=L^1x=\sigma(x)
L1f=L1x=σ(x).
则根据:
f
(
x
t
)
=
f
(
x
t
0
)
+
L
0
f
(
x
t
0
)
(
t
−
t
0
)
+
L
1
f
(
x
t
0
)
(
W
t
−
W
t
0
)
+
R
‾
f(x_t)=f(x_{t0})+L^0 f(x_{t_0})(t-t_0)+L^1 f(x_{t_0})(W_t-W_{t_0})+\overline{R}
f(xt)=f(xt0)+L0f(xt0)(t−t0)+L1f(xt0)(Wt−Wt0)+R
代入
f
(
x
)
=
x
f(x)=x
f(x)=x可得:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
+
R
‾
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})+\overline{R}
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)+R
此时
R
‾
\overline{R}
R为:
R
‾
=
∫
t
0
t
∫
t
0
s
L
0
b
(
x
s
1
)
d
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
1
b
(
x
s
1
)
d
W
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
0
σ
(
x
s
1
)
d
s
1
d
W
s
+
∫
t
0
t
∫
t
0
s
L
1
σ
(
x
s
1
)
d
W
s
1
d
W
s
\overline{R}=\int_{t_0}^t\int_{t_0}^s L^0b(x_{s_1})ds_1ds+\int_{t_0}^t\int_{t_0}^s L^1b(x_{s_1})dWs_1ds+\int_{t_0}^t\int_{t_0}^s L^0\sigma(x_{s_1})ds_1dWs+\int_{t_0}^t\int_{t_0}^s L^1\sigma(x_{s_1})dWs_1dWs
R=∫t0t∫t0sL0b(xs1)ds1ds+∫t0t∫t0sL1b(xs1)dWs1ds+∫t0t∫t0sL0σ(xs1)ds1dWs+∫t0t∫t0sL1σ(xs1)dWs1dWs
称下式为Euler-Maruyama格式:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)需要用到如下随机微分性质:
E
[
d
W
s
2
]
=
d
s
E[dWs^2]=ds
E[dWs2]=ds
根据定理和随机微分的性质可知余项
R
‾
\overline{R}
R中满足:
p
1
=
2
p_1=2
p1=2 (具有
d
W
s
dWs
dWs会使得阶数为0),
p
2
=
1
p_2=1
p2=1(一个
d
s
ds
ds贡献一个阶,一个
d
W
s
dWs
dWs贡献0.5个阶), 满足定理格式,此时强收敛阶为
p
2
−
0.5
=
0.5
p_2-0.5=0.5
p2−0.5=0.5阶,则该格式具有0.5阶强收敛性。
4、强格式2/3: 1阶强SDE (Milstein)和1.5阶强SDE.
注意到余项:
R
‾
=
∫
t
0
t
∫
t
0
s
L
0
b
(
x
s
1
)
d
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
1
b
(
x
s
1
)
d
W
s
1
d
s
+
∫
t
0
t
∫
t
0
s
L
0
σ
(
x
s
1
)
d
s
1
d
W
s
+
∫
t
0
t
∫
t
0
s
L
1
σ
(
x
s
1
)
d
W
s
1
d
W
s
\overline{R}=\int_{t_0}^t\int_{t_0}^s L^0b(x_{s_1})ds_1ds+\int_{t_0}^t\int_{t_0}^s L^1b(x_{s_1})dWs_1ds+\int_{t_0}^t\int_{t_0}^s L^0\sigma(x_{s_1})ds_1dWs+\int_{t_0}^t\int_{t_0}^s L^1\sigma(x_{s_1})dWs_1dWs
R=∫t0t∫t0sL0b(xs1)ds1ds+∫t0t∫t0sL1b(xs1)dWs1ds+∫t0t∫t0sL0σ(xs1)ds1dWs+∫t0t∫t0sL1σ(xs1)dWs1dWs
我们想要提升强格式阶,那么显然此时若能够提升
p
2
p_2
p2的阶就可以了,注意到限制了
p
2
p_2
p2的阶为
R
‾
\overline{R}
R中这一项:
G
=
∫
t
0
t
∫
t
0
s
L
1
σ
(
x
s
1
)
d
W
s
1
d
W
s
G=\int_{t_0}^t\int_{t_0}^s L^1\sigma(x_{s_1})dWs_1dWs
G=∫t0t∫t0sL1σ(xs1)dWs1dWs那么,我们需要对其进行再一次的展开,此时根据
I
t
o
Ito
Ito公式,我们可以将它展开为:
G
=
L
1
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
W
s
1
d
W
s
+
∫
t
0
t
∫
t
0
s
∫
t
0
s
1
L
0
L
1
σ
(
x
s
2
)
d
s
2
d
W
s
1
d
W
s
+
∫
t
0
t
∫
t
0
s
∫
t
0
s
1
L
1
L
1
σ
(
x
s
2
)
d
W
s
2
d
W
s
1
d
W
s
G=L^1\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s dWs_1dWs+\int_{t_0}^t\int_{t_0}^s\int _{t_0}^{s_1}L^0L^1\sigma(x_{s_2})d{s_2}dWs_1dWs+\int_{t_0}^t\int_{t_0}^s\int _{t_0}^{s_1}L^1L^1\sigma(x_{s_2})d{Ws_2}dWs_1dWs
G=L1σ(xt0)∫t0t∫t0sdWs1dWs+∫t0t∫t0s∫t0s1L0L1σ(xs2)ds2dWs1dWs+∫t0t∫t0s∫t0s1L1L1σ(xs2)dWs2dWs1dWs
此时可以得到Milstein格式为:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
+
L
1
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
W
s
1
d
W
s
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})+L^1\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s dWs_1dWs
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)+L1σ(xt0)∫t0t∫t0sdWs1dWs这即:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
+
σ
′
(
x
t
0
)
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
W
s
1
d
W
s
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})+\sigma^{'}(x_{t_0})\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s dWs_1dWs
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)+σ′(xt0)σ(xt0)∫t0t∫t0sdWs1dWs其中:
∫
t
0
t
∫
t
0
s
d
W
s
1
d
W
s
=
1
2
(
W
t
−
W
t
0
)
2
−
1
2
(
t
−
t
0
)
\int_{t_0}^t\int_{t_0}^s dWs_1dWs=\frac{1}{2}(W_t-W_{t_0})^2-\frac{1}{2}(t-t_0)
∫t0t∫t0sdWs1dWs=21(Wt−Wt0)2−21(t−t0)这即可以得到最终格式为:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
+
1
2
σ
′
(
x
t
0
)
σ
(
x
t
0
)
[
(
W
t
−
W
t
0
)
2
−
(
t
−
t
0
)
]
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})+\frac{1}{2}\sigma^{'}(x_{t_0})\sigma(x_{t_0})[(W_t-W_{t_0})^2-(t-t_0)]
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)+21σ′(xt0)σ(xt0)[(Wt−Wt0)2−(t−t0)]
此时注意到
p
1
=
2
p_1=2
p1=2不变,因为提升了
p
2
=
3
2
p_2=\frac{3}{2}
p2=23,满足定理要求,因此该格式收敛阶为1.下面继续提升,注意到此时若想提升强收敛阶,很明显提升
p
2
p_2
p2是不够的了,因为此时
p
1
p_1
p1也需要被提升,但是很明显的是在
R
‾
\overline{R}
R中有
p
2
=
3
2
p_2=\frac{3}{2}
p2=23的项两个:
∫
t
0
t
∫
t
0
s
L
0
σ
(
x
s
1
)
d
s
1
d
W
s
\int_{t_0}^t\int_{t_0}^s L^0\sigma(x_{s_1})ds_1dWs
∫t0t∫t0sL0σ(xs1)ds1dWs和
∫
t
0
t
∫
t
0
s
L
1
b
(
x
s
1
)
d
W
s
1
d
s
\int_{t_0}^t\int_{t_0}^s L^1b(x_{s_1})dWs_1ds
∫t0t∫t0sL1b(xs1)dWs1ds则他们也需要被包含在内才可以提升该格式的收敛阶,因此在
M
i
l
s
t
e
i
n
Milstein
Milstein的基础上,我们继续延拓格式
T
T
T为:
T
=
M
i
l
s
t
e
i
n
+
K
T=Milstein+K
T=Milstein+K
K
=
L
0
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
W
s
+
L
1
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
W
s
1
d
s
+
L
0
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
s
+
(
L
1
)
2
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
∫
t
0
s
1
d
W
s
2
d
W
s
1
d
W
s
K=L^0\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^sds_1dWs+L^1b(x_{t_0})\int_{t_0}^t\int_{t_0}^s dWs_1ds+L^0b(x_{t_0})\int_{t_0}^t\int_{t_0}^s ds_1ds+(L^1)^2\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s\int _{t_0}^{s_1}d{Ws_2}dWs_1dWs
K=L0σ(xt0)∫t0t∫t0sds1dWs+L1b(xt0)∫t0t∫t0sdWs1ds+L0b(xt0)∫t0t∫t0sds1ds+(L1)2σ(xt0)∫t0t∫t0s∫t0s1dWs2dWs1dWs我们仅考虑从
n
n
n到
n
+
1
n+1
n+1会发生什么,令
△
t
=
t
n
+
1
−
t
n
\triangle t=t_{n+1}-t_n
△t=tn+1−tn,
△
W
=
W
t
n
+
1
−
W
t
n
\triangle W=W_{t_{n+1}}-W_{t_n}
△W=Wtn+1−Wtn
(
L
1
)
2
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
∫
t
0
s
1
d
W
s
2
d
W
s
1
d
W
s
=
(
L
1
)
2
σ
(
x
t
0
)
[
1
6
△
W
2
−
1
2
△
t
]
△
W
(L^1)^2\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s\int _{t_0}^{s_1}d{Ws_2}dWs_1dWs=(L^1)^2\sigma(x_{t_0})[\frac{1}{6}\triangle W^2-\frac{1}{2}\triangle t]\triangle W
(L1)2σ(xt0)∫t0t∫t0s∫t0s1dWs2dWs1dWs=(L1)2σ(xt0)[61△W2−21△t]△W这一点使用
I
t
o
Ito
Ito公式很容易证明。
另一方面令:
△
Z
=
∫
t
n
t
n
+
1
∫
t
n
s
d
W
s
1
d
s
=
∫
t
n
t
n
+
1
[
t
n
+
1
−
s
]
d
W
s
\triangle Z=\int_{t_n}^{t_{n+1}}\int_{t_n}^s dWs_1ds=\int_{t_n}^{t_{n+1}}[t_{n+1}-s] dW_s
△Z=∫tntn+1∫tnsdWs1ds=∫tntn+1[tn+1−s]dWs根据随机微分的性质,显然
△
Z
∼
N
(
0
,
∫
t
n
t
n
+
1
[
t
n
+
1
−
s
]
2
d
s
)
=
N
(
0
,
1
3
△
t
3
)
\triangle Z \sim N(0,\int_{t_n}^{t_{n+1}}[t_{n+1}-s]^2 ds)=N(0,\frac{1}{3} \triangle t^3)
△Z∼N(0,∫tntn+1[tn+1−s]2ds)=N(0,31△t3)注意到:
E
[
△
Z
△
W
]
=
∫
t
n
t
n
+
1
[
t
n
+
1
−
s
]
d
s
=
1
2
△
t
2
E[\triangle Z \triangle W]=\int_{t_n}^{t_{n+1}}[t_{n+1}-s] ds=\frac{1}{2} \triangle t^2
E[△Z△W]=∫tntn+1[tn+1−s]ds=21△t2因此:
E
[
(
△
Z
−
1
2
△
W
)
△
W
]
=
0
E[(\triangle Z-\frac{1}{2} \triangle W)\triangle W]=0
E[(△Z−21△W)△W]=0这样找到了独立的样本,且注意到此时根据随机微分的性质可得到:
(
△
Z
−
1
2
△
W
)
~
N
(
0
,
1
12
△
t
3
)
(\triangle Z-\frac{1}{2} \triangle W)~N(0,\frac{1}{12}\triangle t^3)
(△Z−21△W)~N(0,121△t3)注意到:
∫
t
0
t
∫
t
0
s
d
s
1
d
W
s
=
△
t
△
W
−
△
Z
\int_{t_0}^t\int_{t_0}^sds_1dWs=\triangle t \triangle W-\triangle Z
∫t0t∫t0sds1dWs=△t△W−△Z那么显然地,我们可以根据如下方法来给出1.5阶格式,若我们有
G
1
G_1
G1,
G
2
G_2
G2两个独立的同分布正态分布变量
G
1
~
N
(
0
,
1
)
,
G
2
~
N
(
0
,
1
)
G_1~N(0,1),G_2~N(0,1)
G1~N(0,1),G2~N(0,1)那么根据我们的推导
K
=
L
0
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
W
s
+
L
1
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
W
s
1
d
s
+
L
0
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
s
+
(
L
1
)
2
σ
(
x
t
0
)
∫
t
0
t
∫
t
0
s
∫
t
0
s
1
d
W
s
2
d
W
s
1
d
W
s
K=L^0\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^sds_1dWs+L^1b(x_{t_0})\int_{t_0}^t\int_{t_0}^s dWs_1ds+L^0b(x_{t_0})\int_{t_0}^t\int_{t_0}^s ds_1ds+(L^1)^2\sigma(x_{t_0})\int_{t_0}^t\int_{t_0}^s\int _{t_0}^{s_1}d{Ws_2}dWs_1dWs
K=L0σ(xt0)∫t0t∫t0sds1dWs+L1b(xt0)∫t0t∫t0sdWs1ds+L0b(xt0)∫t0t∫t0sds1ds+(L1)2σ(xt0)∫t0t∫t0s∫t0s1dWs2dWs1dWs
K
=
L
0
σ
(
x
t
0
)
[
△
t
△
W
−
△
Z
]
+
L
1
b
(
x
t
0
)
△
Z
+
L
0
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
s
+
(
L
1
)
2
σ
(
x
t
0
)
[
1
6
△
W
2
−
1
2
△
t
]
△
W
K=L^0\sigma(x_{t_0})[\triangle t \triangle W-\triangle Z]+L^1b(x_{t_0})\triangle Z+L^0b(x_{t_0})\int_{t_0}^t\int_{t_0}^s ds_1ds+(L^1)^2\sigma(x_{t_0})[\frac{1}{6}\triangle W^2-\frac{1}{2}\triangle t]\triangle W
K=L0σ(xt0)[△t△W−△Z]+L1b(xt0)△Z+L0b(xt0)∫t0t∫t0sds1ds+(L1)2σ(xt0)[61△W2−21△t]△W其中:
△
W
=
△
t
G
1
.
\triangle W= \sqrt{\triangle t} G_1.
△W=△tG1.
△
Z
=
1
2
△
t
△
W
+
1
2
3
△
t
3
2
G
2
\triangle Z= \frac{1}{2} \triangle t\triangle W+\frac{1}{2 \sqrt{3}} \triangle t^{\frac{3}{2}} G_2
△Z=21△t△W+231△t23G2此时
K
K
K中所有成分可求,这样的1.5阶格式为:
x
t
=
x
t
0
+
b
(
x
t
0
)
(
t
−
t
0
)
+
σ
(
x
t
0
)
(
W
t
−
W
t
0
)
+
1
2
σ
′
(
x
t
0
)
σ
(
x
t
0
)
[
(
W
t
−
W
t
0
)
2
−
(
t
−
t
0
)
]
+
K
x_t=x_{t_0}+b(x_{t_0})(t-t_0)+\sigma(x_{t_0})(W_t-W_{t_0})+\frac{1}{2}\sigma^{'}(x_{t_0})\sigma(x_{t_0})[(W_t-W_{t_0})^2-(t-t_0)]+K
xt=xt0+b(xt0)(t−t0)+σ(xt0)(Wt−Wt0)+21σ′(xt0)σ(xt0)[(Wt−Wt0)2−(t−t0)]+K
K
=
L
0
σ
(
x
t
0
)
[
△
t
△
W
−
△
Z
]
+
L
1
b
(
x
t
0
)
△
Z
+
L
0
b
(
x
t
0
)
∫
t
0
t
∫
t
0
s
d
s
1
d
s
+
(
L
1
)
2
σ
(
x
t
0
)
[
1
6
△
W
2
−
1
2
△
t
]
△
W
K=L^0\sigma(x_{t_0})[\triangle t \triangle W-\triangle Z]+L^1b(x_{t_0})\triangle Z+L^0b(x_{t_0})\int_{t_0}^t\int_{t_0}^s ds_1ds+(L^1)^2\sigma(x_{t_0})[\frac{1}{6}\triangle W^2-\frac{1}{2}\triangle t]\triangle W
K=L0σ(xt0)[△t△W−△Z]+L1b(xt0)△Z+L0b(xt0)∫t0t∫t0sds1ds+(L1)2σ(xt0)[61△W2−21△t]△W当然,读者可以自行推导更高阶的
I
t
o
−
T
a
y
l
o
r
Ito-Taylor
Ito−Taylor展开,笔者在这里不过多介绍。