伊藤公式

引理:设X={Xt,t[0,T]}X=\{X_t,t\in[0,T] \}是如下形式的伊藤过程。

Xt=X0+0tαsdBs+0tβsdsX_t=X_0+\int_{0}^{t}\alpha_s \mathrm{d}B_s+\int_{0}^{t}\beta_s \mathrm{d}s

f(t,x)f(t,x)的偏导数\frac{\part f}{\part t},\frac{\part f}{\part x},\frac{\part^2 f}{\part x^2}都存在且连续,则:

Y={Yt,t[0,T]}={f(t,Xt),t[0,T]}Y=\{Y_t,t\in [0,T] \}=\{f(t,X_t),t\in [0,T] \}

也是伊藤过程,且:

Y_t=f(t,X_t)=f(0,X_0)+\int_{0}^{t}[\frac{\part f}{\part t}(s,X_s)+\frac{\part f}{\part x}(s,X_s)\beta_s+\frac{1}{2}\frac{\part^2 f}{\part x^2}(s,X_s)\alpha_s^2] \mathrm{d}s+\int_{0}^{t}[\frac{\part f}{\part x}(s,X_s)\alpha_s] \mathrm{d}B_s

微分形式:

\mathrm{d}Y_t=\mathrm{d} f(t,X_t)=\frac{\part f}{\part t}(t,X_t)\mathrm{d}t+\frac{\part f}{\part X}(t,X_t)\mathrm{d}X_t+\frac{1}{2}\frac{\part^2 f}{\part x^2}(t,X_t)(\mathrm{d}X_t)^2

而我们知道:

(dXt)2=(αtdBt+βtdt)2=αt2(dBt)2=αt2dt(\mathrm{d}X_t)^2=(\alpha_t \mathrm{d}B_t+\beta_t \mathrm{d}t)^2=\alpha_t^2(\mathrm{d}B_t)^2=\alpha_t^2\mathrm{d}t

所以:

\mathrm{d}Y_t=[\frac{\part f}{\part t}(t,X_t)+\frac{1}{2}\alpha_t^2\frac{\part^2 f}{\part x^2}(t,X_t)]\mathrm{d}t+\frac{\part f}{\part X}(t,X_t)\mathrm{d}X_t \\ =[\frac{\part f}{\part t}(t,X_t)+\beta_t\frac{\part f}{\part x}(t,X_t)+\frac{1}{2}\alpha_t^2\frac{\part^2 f}{\part x^2}(t,X_t)]\mathrm{d}t+\alpha_t\frac{\part f}{\part X}(t,X_t)\mathrm{d}B_t

证明思路

使用泰勒展开,一些高阶无穷小项被省略掉了,区别于传统二元函数泰勒展开,因为(dXt)2=αt2dt(\mathrm{d}X_t)^2=\alpha_t^2\mathrm{d}t,故(dXt)2(dX_t)^2项没有被省略掉。

随机微分方程(SDE)(SDE) (Stochastic Differential Equation)(Stochastic\ Differential\ Equation)

线性SDESDE

dX(t)=(c1(t)X(t)+c2(t))dt+(σ1(t)X(t)+σ2(t))dBtX0=x0\mathrm{d} X(t)=(c_1(t)X(t)+c_2(t))\mathrm{d} t+(\sigma_1(t)X(t)+\sigma_2(t))\mathrm{d} B_t \\ X_0=x_0

其中c1(t),c2(t),σ1(t),σ2(t)c_1(t),c_2(t),\sigma_1(t),\sigma_2(t)是给定的函数。

齐次线性SDESDE

dX(t)=c1(t)X(t)dt+σ1(t)X(t)dBtX0=x0>0\mathrm{d} X(t)=c_1(t)X(t)\mathrm{d} t+\sigma_1(t)X(t)\mathrm{d} B_t \\ X_0=x_0>0

求解方法如下:

​ 令Yt=lnXtY_t=\ln X_t,对YtY_t应用伊藤公式:

dY(t)=dlnX(t)=[c1(t)12σ12(t)]dt+σ1(t)dB(t)\mathrm{d} Y(t)=\mathrm{d}\ln X(t)=[c_1(t)-\frac{1}{2}\sigma_1^2(t)] \mathrm{d}t+\sigma_1(t) \mathrm{d}B(t)

所以:

lnXt=Y(t)=lnX0+0t[c1(s)12σ12(s)]ds+0tσ1(s)dBs\ln X_t=Y(t)=\ln X_0+\int_{0}^{t}[c_1(s)-\frac{1}{2}\sigma_1^2(s)] \mathrm{d} s+\int_{0}^{t} \sigma_1(s) \mathrm{d} B_s

Xt=X0 exp{0t[c1(s)12σ12(s)]ds+0tσ1(s)dBs}X_t=X_0\ \exp\{\int_{0}^{t}[c_1(s)-\frac{1}{2}\sigma_1^2(s)] \mathrm{d} s+\int_{0}^{t} \sigma_1(s) \mathrm{d} B_s\}

称为广义布朗运动。

C1(s)=μ , σ1(s)=σC_1(s)=\mu\ ,\ \sigma_1(s)=\sigma 为常数,则:

Xt=X0 e(μ12σ2)t+σBtX_t=X_0\ e^{(\mu-\frac{1}{2}\sigma^2)t+\sigma B_t}

是几何布朗运动。

线性SDESDE的求解

补充知识:

若:

dX(t)=α1(t)dB(t)+β1(t)dtdY(t)=α2(t)dB(t)+β2(t)dt\mathrm{d}X(t)=\alpha_1(t)\mathrm{d}B(t)+\beta_1(t)\mathrm{d}t \\ \mathrm{d}Y(t)=\alpha_2(t)\mathrm{d}B(t)+\beta_2(t)\mathrm{d}t

则:

d(X(t)Y(t))=X(t)dY(t)+Y(t)dX(t)+dX(t)dY(t)\mathrm{d} (X(t)Y(t))=X(t)\mathrm{d}Y(t)+Y(t)\mathrm{d}X(t)+\mathrm{d}X(t)\mathrm{d}Y(t)

求解:

Y(t)Y(t)为齐次线性SDESDE的解,令X(1)=Y1(t)X^{(1)}=Y^{-1}(t),则:

dX(1)=dY1(t)=[c1(t)+σ12(t)]Xt(1)dtσ1(t)Xt(1)dBt\mathrm{d} X^{(1)}=\mathrm{d} Y^{-1}(t)=[-c_1(t)+\sigma_1^2(t)]X_t^{(1)} \mathrm{d}t-\sigma_1(t)X_t^{(1)} \mathrm{d} B_t

运用上述补充知识可得:

d(Xt(1)Xt)=[c2(t)σ1(t)σ2(t)]Xt(1)dtσ2(t)Xt(1)dBt\mathrm{d} (X_t^{(1)}X_t)=[c_2(t)-\sigma_1(t)\sigma_2(t)]X_t^{(1)} \mathrm{d}t-\sigma_2(t)X_t^{(1)} \mathrm{d} B_t

注意到Y0=1Y_0=1,对两边积分并化简可得:

Xt=Yt(X0+0t[c2(s)σ1(s)σ2(s)]Ys1ds+0tσ2(s)Ys1dBs)X_t=Y_t(X_0+\int_{0}^{t}[c_2(s)-\sigma_1(s)\sigma_2(s)]Y_s^{-1} \mathrm{d}s+\int_{0}^{t}\sigma_2(s)Y_s^{-1} \mathrm{d} B_s)

其中,YtY_t是齐次线性SDESDE的解:

Yt= exp{0t[c1(s)12σ12(s)]ds+0tσ1(s)dBs}Y_t=\ \exp\{\int_{0}^{t}[c_1(s)-\frac{1}{2}\sigma_1^2(s)] \mathrm{d} s+\int_{0}^{t} \sigma_1(s) \mathrm{d} B_s\}

解的期望与方差

对于非齐次线性SDESDE:

dX(t)=(c1(t)X(t)+c2(t))dt+(σ1(t)X(t)+σ2(t))dBtX0=x0\mathrm{d} X(t)=(c_1(t)X(t)+c_2(t))\mathrm{d} t+(\sigma_1(t)X(t)+\sigma_2(t))\mathrm{d} B_t \\ X_0=x_0

我们考虑对其解求期望。

对其积分形式两边求期望:

E[X(t)]=X0=0tE[(c1(s)X(s)+c2(s))]ds+0E[X(t)]=X_0=\int_{0}^{t}E[(c_1(s)X(s)+c_2(s))]\mathrm{d} s+0

假设c1(s),c2(s)c_1(s),c_2(s)为确定性函数,则:

E[X(t)]=X0=0t(c1(s)E[X(s)]+c2(s))ds+0E[X(t)]=X_0=\int_{0}^{t}(c_1(s)E[X(s)]+c_2(s))\mathrm{d} s+0

一阶线性常微分方程

g(t)=E[X(t)]g(t)=E[X(t)],则:

g(t)=c1(t)g(t)+c2(t)g(0)=x0g'(t)=c_1(t)g(t)+c_2(t) \\ g(0)=x_0

我们也可以直接令原方程中的σ1,σ2=0\sigma_1,\sigma_2=0,解出来的XtX_t即为期望。