前文链接
等价SDE
例 1：
求解Ito-SDE
d
X
t
=
[
q
f
(
X
t
)
+
1
2
f
(
X
t
)
f
′
(
X
t
)
]
d
t
+
f
(
X
t
)
d
B
t
dX_t=[qf(X_t)+\frac{1}{2}f(X_t)f'(X_t)]dt+f(X_t)dB_t
dXt=[qf(Xt)+21f(Xt)f′(Xt)]dt+f(Xt)dBt
可以知道
{
a
(
t
,
x
)
=
q
f
(
X
t
)
+
1
2
f
(
X
t
)
f
′
(
X
t
)
b
(
t
,
x
)
=
f
(
X
t
)
\begin{cases} a(t, x)=qf(X_t)+\frac{1}{2}f(X_t)f'(X_t)\\ b(t, x)=f(X_t) \end{cases}
{a(t,x)=qf(Xt)+21f(Xt)f′(Xt)b(t,x)=f(Xt)
根据转换定理
{
a
~
(
t
,
x
)
=
a
(
t
,
x
)
−
1
2
b
(
t
,
x
)
b
x
(
t
,
x
)
=
q
f
(
X
t
)
b
~
(
t
,
x
)
=
f
(
X
t
)
\begin{cases} \widetilde{a}(t, x)=a(t, x)-\frac{1}{2}b(t, x)b_x(t, x)=qf(X_t)\\ \widetilde{b}(t, x)=f(X_t) \end{cases}
{a
(t,x)=a(t,x)−21b(t,x)bx(t,x)=qf(Xt)b
(t,x)=f(Xt)
得到Stratonovich-SDE
d
X
t
=
q
f
(
X
t
)
d
t
+
f
(
X
t
)
∘
d
B
t
dX_t=qf(X_t)dt+f(X_t)\circ dB_t
dXt=qf(Xt)dt+f(Xt)∘dBt
求解analogue在
[
0
,
t
]
[0, t]
[0,t]上积分：
d
k
(
t
)
f
(
k
(
t
)
)
=
q
d
t
+
d
c
(
t
)
∫
k
(
0
)
k
(
t
)
d
k
(
t
)
f
(
k
(
t
)
)
=
q
t
+
c
(
t
)
−
c
(
0
)
\frac{dk(t)}{f(k(t))}=qdt+dc(t)\\ \int_{k(0)}^{k(t)}\frac{dk(t)}{f(k(t))}=qt+c(t)-c(0)
f(k(t))dk(t)=qdt+dc(t)∫k(0)k(t)f(k(t))dk(t)=qt+c(t)−c(0)
设
d
u
f
(
u
)
\frac{du}{f(u)}
f(u)du的原函数为
g
(
u
)
g(u)
g(u)，得到
g
(
k
(
t
)
)
−
g
(
k
(
0
)
)
=
q
t
+
c
(
t
)
−
c
(
0
)
g
(
X
t
)
−
g
(
X
0
)
=
q
t
+
B
(
t
)
g(k(t))-g(k(0))=qt+c(t)-c(0)\\ g(X_t)-g(X_0)=qt+B(t)
g(k(t))−g(k(0))=qt+c(t)−c(0)g(Xt)−g(X0)=qt+B(t)
例 2：
求解Ito-SDE
d
X
t
=
[
q
(
X
t
+
1
)
+
1
2
(
X
t
+
1
)
]
d
t
+
(
X
t
+
1
)
d
B
t
,
X
0
=
0
dX_t=[q(X_t+1)+\frac{1}{2}(X_t+1)]dt+(X_t+1)dB_t, X_0=0
dXt=[q(Xt+1)+21(Xt+1)]dt+(Xt+1)dBt,X0=0
令
f
(
x
)
=
x
+
1
f(x)=x+1
f(x)=x+1，可以得到
d
X
t
=
[
q
f
(
X
t
)
+
1
2
f
(
X
t
)
f
′
(
X
t
)
]
d
t
+
f
(
X
t
)
d
B
t
dX_t=[qf(X_t)+\frac{1}{2}f(X_t)f'(X_t)]dt+f(X_t)dB_t
dXt=[qf(Xt)+21f(Xt)f′(Xt)]dt+f(Xt)dBt
和例 1形式相同，求出
g
(
u
)
g(u)
g(u)
g
(
u
)
=
∫
1
1
+
x
d
x
=
ln
(
1
+
x
)
g(u)=\int \frac{1}{1+x}dx=\ln(1+x)
g(u)=∫1+x1dx=ln(1+x)
得到
ln
(
X
t
+
1
)
=
q
t
+
B
t
X
t
=
exp
(
q
t
+
B
t
)
−
1
\ln(X_t+1)=qt+B_t\\ X_t=\exp(qt+B_t)-1
ln(Xt+1)=qt+BtXt=exp(qt+Bt)−1
Ornstein-Uhlenbeck Process
O-U过程满足如下形式
X
t
=
X
0
+
c
∫
0
t
X
s
d
s
+
σ
∫
0
t
d
B
s
X_t=X_0+c\int_0^tX_sds+\sigma\int_0^tdB_s
Xt=X0+c∫0tXsds+σ∫0tdBs
且
X
0
X_0
X0非随机.
微分形式
d
X
t
=
c
X
t
+
σ
d
B
t
(1)
dX_t=cX_t+\sigma dB_t \tag{1}
dXt=cXt+σdBt(1)
离散形式
X
t
+
1
−
X
t
=
c
X
t
+
σ
(
B
t
+
1
−
B
t
)
X_{t+1}-X_t=cX_t+\sigma(B_{t+1}-B_t)
Xt+1−Xt=cXt+σ(Bt+1−Bt)
移项得到自回归模型
X
t
+
1
=
(
c
+
1
)
⏟
ϕ
X
t
+
σ
(
B
t
+
1
−
B
t
)
⏟
Z
t
∼
N
(
0
,
σ
2
)
=
ϕ
X
t
+
Z
t
X_{t+1}=\underbrace{(c+1)}_{\phi}X_t+\underbrace{\sigma (B_{t+1}-B_t)}_{Z_t\sim \mathcal{N}(0, \sigma^2)}=\phi X_t+Z_t
Xt+1=ϕ
(c+1)Xt+Zt∼N(0,σ2)
σ(Bt+1−Bt)=ϕXt+Zt
考虑线性回归方程
y
=
β
0
+
β
1
x
+
ε
y=\beta_0+\beta_1x+\varepsilon
y=β0+β1x+ε
求解方程
Y
t
=
e
−
c
t
X
t
Y_t=e^{-ct}X_t
Yt=e−ctXt，令
f
(
t
,
x
)
=
e
−
c
t
x
,
f
t
′
=
−
x
c
e
−
c
t
,
f
x
′
=
e
−
c
t
,
f
x
x
′
=
0
f(t, x)=e^{-ct}x, f'_t=-xce^{-ct}, f'_x=e^{-ct}, f'_{xx}=0
f(t,x)=e−ctx,ft′=−xce−ct,fx′=e−ct,fxx′=0
由伊藤公式得到
d
Y
t
=
d
f
(
t
,
X
t
)
=
f
t
′
d
t
+
f
x
′
d
X
t
+
1
2
f
x
x
′
d
X
(
t
)
d
X
(
t
)
=
−
c
X
t
e
−
c
t
d
t
+
e
−
c
t
(
c
X
t
d
t
+
σ
d
B
t
)
=
σ
e
−
c
t
d
B
t
(2)
\begin{aligned} dY_t&=df(t, X_t)\\ &=f'_tdt+f'_xdX_t+\frac{1}{2}f'_{xx}dX(t)dX(t)\\ &=-cX_te^{-ct}dt+e^{-ct}(cX_tdt+\sigma dB_t) \\ &=\sigma e^{-ct}dB_t \end{aligned}\tag{2}
dYt=df(t,Xt)=ft′dt+fx′dXt+21fxx′dX(t)dX(t)=−cXte−ctdt+e−ct(cXtdt+σdBt)=σe−ctdBt(2)
可以发现
d
Y
t
dY_t
dYt中没有漂移项，
Y
t
Y_t
Yt是一个martingale.
证明：
P
t
=
e
−
c
t
P_t=e^{-ct}
Pt=e−ct时方程可以求解
设方程
Y
t
=
P
t
X
t
Y_t=P_tX_t
Yt=PtXt，令
f
(
t
,
x
)
=
P
t
x
,
f
t
′
=
P
t
′
x
,
f
x
′
=
P
t
,
f
x
x
′
′
=
0
f(t, x)=P_tx, f'_t=P'_tx, f'_x=P_t, f''_{xx}=0
f(t,x)=Ptx,ft′=Pt′x,fx′=Pt,fxx′′=0
d
Y
t
=
d
f
(
t
,
X
t
)
=
P
t
′
X
t
d
t
+
P
t
d
X
t
=
P
t
′
X
t
d
t
+
P
t
(
c
X
t
d
t
+
σ
d
B
t
)
=
(
P
t
′
+
c
P
t
)
X
t
d
t
+
P
t
σ
d
B
t
\begin{aligned} dY_t&=df(t, X_t)=P_t'X_tdt+P_tdX_t\\ &=P_t'X_tdt+P_t(cX_tdt+\sigma dB_t)\\ &=(P_t'+cP_t)X_tdt+P_t\sigma dB_t \end{aligned}
dYt=df(t,Xt)=Pt′Xtdt+PtdXt=Pt′Xtdt+Pt(cXtdt+σdBt)=(Pt′+cPt)Xtdt+PtσdBt
只有当漂移项为0时，方程可以求解，因此要求
P
t
′
+
c
P
t
=
0
P'_t+cP_t=0
Pt′+cPt=0
解得
P
t
=
e
−
c
t
P_t=e^{-ct}
Pt=e−ct
对方程
(
2
)
(2)
(2)两边进行积分
Y
t
−
Y
0
=
∫
0
t
σ
e
−
c
s
d
B
s
Y_t-Y_0=\int_0^t\sigma e^{-cs}dB_s
Yt−Y0=∫0tσe−csdBs
代入
Y
t
=
e
−
c
t
X
t
Y_t=e^{-ct}X_t
Yt=e−ctXt
e
−
c
t
X
t
=
X
0
+
∫
0
t
σ
e
−
c
s
d
B
s
X
t
=
e
c
t
X
0
+
σ
e
c
t
∫
0
t
e
−
c
s
d
B
s
e^{-ct}X_t=X_0+\int_0^t\sigma e^{-cs}dB_s\\ X_t=e^{ct}X_0+\sigma e^{ct}\int_0^t e^{-cs}dB_s
e−ctXt=X0+∫0tσe−csdBsXt=ectX0+σect∫0te−csdBs
Ito Process Distribution
伊藤过程的分布
根据
I
(
t
)
=
∫
0
t
Δ
(
u
)
d
B
(
u
)
E
[
I
(
t
)
]
=
E
[
I
(
t
)
∣
F
(
0
)
]
=
I
(
0
)
=
0
E
[
I
2
(
t
)
]
=
∫
0
t
E
[
Δ
2
(
u
)
]
d
u
\begin{aligned} &I(t)=\int_0^t\Delta(u)dB(u)\\ &\mathbb{E}[I(t)]=\mathbb{E}[I(t)\mid \mathcal{F}(0)]=I(0)=0\\ &\mathbb{E}[I^2(t)]=\int_0^t\mathbb{E}[\Delta^2(u)]du \end{aligned}
I(t)=∫0tΔ(u)dB(u)E[I(t)]=E[I(t)∣F(0)]=I(0)=0E[I2(t)]=∫0tE[Δ2(u)]du
得到
I
(
t
)
∼
N
(
0
,
∫
0
t
E
[
Δ
2
(
u
)
]
d
u
)
I(t)\sim \mathcal{N}(0, \int_0^t\mathbb{E}[\Delta^2(u)]du)
I(t)∼N(0,∫0tE[Δ2(u)]du)
一般线性SDE
一般线性SDE方程
d
X
(
u
)
=
(
a
(
u
)
+
b
(
u
)
X
(
u
)
)
d
u
+
(
γ
(
u
)
+
σ
(
u
)
X
(
u
)
)
d
W
(
u
)
dX(u)=(a(u)+b(u)X(u))du+(\gamma(u)+\sigma(u)X(u))dW(u)
dX(u)=(a(u)+b(u)X(u))du+(γ(u)+σ(u)X(u))dW(u)
其中
a
(
u
)
,
b
(
u
)
,
γ
(
u
)
,
σ
(
u
)
a(u), b(u), \gamma(u), \sigma(u)
a(u),b(u),γ(u),σ(u) are adapted to
F
(
u
)
,
u
≥
0
\mathcal{F}(u), u\geq 0
F(u),u≥0.
令
Z
(
u
)
=
exp
{
∫
t
u
σ
(
v
)
d
W
(
v
)
+
∫
t
u
(
b
(
v
)
−
1
2
σ
2
(
v
)
)
d
v
⏟
Q
(
u
)
}
Y
(
u
)
=
x
+
∫
t
u
a
(
v
)
−
σ
(
v
)
γ
(
v
)
Z
(
v
)
d
v
+
∫
t
u
γ
(
v
)
Z
(
v
)
d
W
(
v
)
Z(u)=\exp\{\underbrace{\int_t^u\sigma(v)dW(v)+\int_t^u(b(v)-\frac{1}{2}\sigma^2(v))dv}_{Q(u)}\} \\ Y(u)=x+\int_t^u\frac{a(v)-\sigma(v)\gamma(v)}{Z(v)}dv+\int_t^u\frac{\gamma(v)}{Z(v)}dW(v)
Z(u)=exp{Q(u)
∫tuσ(v)dW(v)+∫tu(b(v)−21σ2(v))dv}Y(u)=x+∫tuZ(v)a(v)−σ(v)γ(v)dv+∫tuZ(v)γ(v)dW(v)
初值
Z
(
t
)
=
1
Z(t)=1
Z(t)=1.
d
Z
(
u
)
=
d
e
Q
(
u
)
dZ(u)=de^{Q(u)}
dZ(u)=deQ(u)，
Q
(
u
)
Q(u)
Q(u)是一个Ito Process，令
f
(
t
,
x
)
=
e
x
,
f
t
=
0
,
f
x
=
e
x
,
f
x
x
=
e
x
f(t, x)=e^x, f_t=0, f_x=e^x, f_{xx}=e^x
f(t,x)=ex,ft=0,fx=ex,fxx=ex
d
Z
(
u
)
=
d
e
Q
(
u
)
=
d
f
(
t
,
Q
(
u
)
)
=
e
Q
(
u
)
d
Q
(
u
)
+
1
2
e
Q
(
u
)
d
Q
(
u
)
d
Q
(
u
)
=
Z
(
u
)
[
σ
(
u
)
d
W
(
u
)
+
(
b
(
u
)
−
1
2
σ
2
(
u
)
)
d
u
+
1
2
Z
(
u
)
σ
2
(
u
)
d
u
]
=
Z
(
u
)
[
σ
(
u
)
d
W
(
u
)
+
b
(
u
)
d
u
]
d
Y
(
u
)
=
a
(
u
)
−
σ
(
u
)
γ
(
u
)
Z
(
u
)
d
u
+
γ
(
u
)
Z
(
u
)
d
W
(
u
)
\begin{aligned} dZ(u)&=de^{Q(u)}=df(t, Q(u))\\ &=e^{Q(u)}dQ(u)+\frac{1}{2}e^{Q(u)}dQ(u)dQ(u) \\ &=Z(u)[\sigma(u)dW(u)+(b(u)-\frac{1}{2}\sigma^2(u))du+\frac{1}{2}Z(u)\sigma^2(u)du]\\ &=Z(u)[\sigma(u)dW(u)+b(u)du]\\ dY(u)&=\frac{a(u)-\sigma(u)\gamma(u)}{Z(u)}du+\frac{\gamma(u)}{Z(u)}dW(u) \end{aligned}
dZ(u)dY(u)=deQ(u)=df(t,Q(u))=eQ(u)dQ(u)+21eQ(u)dQ(u)dQ(u)=Z(u)[σ(u)dW(u)+(b(u)−21σ2(u))du+21Z(u)σ2(u)du]=Z(u)[σ(u)dW(u)+b(u)du]=Z(u)a(u)−σ(u)γ(u)du+Z(u)γ(u)dW(u)
证明
X
(
u
)
=
Y
(
u
)
Z
(
u
)
X(u)=Y(u)Z(u)
X(u)=Y(u)Z(u)
d
X
(
u
)
=
d
[
Y
(
u
)
Z
(
u
)
]
=
Z
(
u
)
d
Y
(
u
)
+
Y
(
u
)
d
Z
(
u
)
+
d
Y
(
u
)
d
Z
(
u
)
=
(
a
(
u
)
−
σ
(
u
)
γ
(
u
)
)
d
u
+
γ
(
u
)
d
W
(
u
)
+
Y
(
u
)
Z
(
u
)
⏟
X
(
u
)
[
b
(
u
)
d
u
+
σ
(
u
)
d
W
(
u
)
]
+
γ
(
u
)
σ
(
u
)
d
u
=
[
a
(
u
)
−
σ
(
u
)
γ
(
u
)
+
X
(
u
)
b
(
u
)
+
γ
(
u
)
σ
(
u
)
]
d
u
+
[
γ
(
u
)
+
σ
(
u
)
X
(
u
)
]
d
W
(
u
)
=
[
a
(
u
)
+
X
(
u
)
b
(
u
)
]
d
u
+
[
γ
(
u
)
+
σ
(
u
)
X
(
u
)
]
d
W
(
u
)
\begin{aligned} dX(u)&=d[Y(u)Z(u)]\\ &=Z(u)dY(u)+Y(u)dZ(u)+dY(u)dZ(u)\\ &=(a(u)-\sigma(u)\gamma(u))du+\gamma(u)dW(u)+\underbrace{Y(u)Z(u)}_{X(u)}[b(u)du+\sigma(u)dW(u)]+\gamma(u)\sigma(u)du\\ &=[a(u)-\sigma(u)\gamma(u)+X(u)b(u)+\gamma(u)\sigma(u)]du+[\gamma(u)+\sigma(u)X(u)]dW(u)\\ &=[a(u)+X(u)b(u)]du+[\gamma(u)+\sigma(u)X(u)]dW(u) \end{aligned}
dX(u)=d[Y(u)Z(u)]=Z(u)dY(u)+Y(u)dZ(u)+dY(u)dZ(u)=(a(u)−σ(u)γ(u))du+γ(u)dW(u)+X(u)
Y(u)Z(u)[b(u)du+σ(u)dW(u)]+γ(u)σ(u)du=[a(u)−σ(u)γ(u)+X(u)b(u)+γ(u)σ(u)]du+[γ(u)+σ(u)X(u)]dW(u)=[a(u)+X(u)b(u)]du+[γ(u)+σ(u)X(u)]dW(u)
证明唯一解，使用Lipchitz条件验证
d
X
(
t
)
=
[
a
(
t
)
+
b
(
t
)
X
(
t
)
]
⏟
m
(
t
,
x
)
:
=
a
(
t
)
+
b
(
t
)
x
d
t
+
[
γ
(
t
)
+
σ
(
t
)
X
(
t
)
]
⏟
n
(
t
,
x
)
:
=
γ
(
t
)
+
σ
(
t
)
x
d
W
(
t
)
dX(t)=\underbrace{[a(t)+b(t)X(t)]}_{m(t, x):=a(t)+b(t)x}dt+\underbrace{[\gamma(t)+\sigma(t)X(t)]}_{n(t, x):=\gamma(t)+\sigma(t)x}dW(t)
dX(t)=m(t,x):=a(t)+b(t)x
[a(t)+b(t)X(t)]dt+n(t,x):=γ(t)+σ(t)x
[γ(t)+σ(t)X(t)]dW(t)
由Lipchitz条件
∣
m
(
t
,
x
)
−
m
(
t
,
y
)
∣
+
∣
n
(
t
,
x
)
−
n
(
t
,
y
)
∣
≤
K
∣
x
−
y
∣
K
≥
∣
b
(
t
)
∣
+
∣
σ
(
t
)
∣
|m(t, x)-m(t, y)|+|n(t, x)-n(t, y)|\leq K|x-y|\\ K\geq |b(t)|+|\sigma(t)|
∣m(t,x)−m(t,y)∣+∣n(t,x)−n(t,y)∣≤K∣x−y∣K≥∣b(t)∣+∣σ(t)∣
因为
∣
b
(
t
)
∣
,
∣
σ
(
t
)
∣
|b(t)|, |\sigma(t)|
∣b(t)∣,∣σ(t)∣在
[
0
,
t
]
[0, t]
[0,t]上是有界的，所以Lipchitz条件是满足的.
扫一扫
2342
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
