【FinE】Girsanov定理

Conditioned Expectation

定义：概率空间$(\Omega ,\mathcal{F},\mathbb{P})$ G 是域流$\mathcal{F}$一个sub-$\sigma$-algebra. 设随机变量$X$非负且可积，定义条件期望$\mathbb{E}[X\mid \mathcal{G}]$.
满足以下条件
(1) $\mathbb{E}[X\mid \mathcal{G}]$是$\mathcal{G}-$可测的.
(2) ${\int }_A\mathbb{E}[X\mid \mathcal{G}]d\mathbb{P}(w)=\underset{\text{Partial Averaging}}{\underbrace{{\int }_{A}X(w)d\mathbb{P}(w)}}$，对于$A\in \mathcal{G}$.
(3) 如果$X$是$\mathcal{G}-$可测的，那么
$$\mathbb{E}[XY\mid \mathcal{G}]=X\mathbb{E}[Y\mid \mathcal{G}]$$
(4) Iteration/Tower Property. 设$\mathcal{H}$是$\mathcal{G}$的sub-$\sigma$-algebra.
$$\mathbb{E}[\mathbb{E}[X\mid \mathcal{G}]\mid \mathcal{H}]=\mathbb{E}[X\mid \mathcal{H}]$$
例：
考虑Radon-Nicodym导数
$$Z(w)=\exp (-\theta X(w)-\frac{1}{2}{\theta }^2)$$
其中，$\theta$是常数. 可以进行测度转换
$$(\Omega ,\mathcal{F},\mathbb{P})\stackrel{\mathbb{Q}(A)={\int }_{A}Z(w)d\mathbb{P}(w)}{\to }(\Omega ,\mathcal{F},\mathbb{Q})$$
特别的
$$X\sim {\mathcal{N}}^{\mathbb{P}}(0,1),Y=X+\theta \sim {\mathcal{N}}^{\mathbb{P}}(\theta ,1)\stackrel{\text{测度转换}}{\to }Y\sim {\mathcal{N}}^{\mathbb{Q}}(0,1)$$

Radon-Nociodym Dervative Process

定义：
$$Z(t)=\mathbb{E}[Z\mid \mathcal{F}(t)],0\le t\le T$$
为Radon-Nicodym Derivative Process.
有如下性质：
(1) $Z(t)$是鞅过程
令$0\le s\le t$
$$\mathbb{E}[Z(t)\mid \mathcal{F}(s)]=\mathbb{E}[\mathbb{E}[Z(t)\mid \mathcal{F}(t)]\mid \mathcal{F}(s)]=\mathbb{E}[Z\mid \mathcal{F}(s)]=Z(s)$$
(2) $Z(w)$是一个R-N导数，${\mathbb{E}}^{\mathbb{Q}}[Y]={\mathbb{E}}^{\mathbb{P}}[YZ]$
$\because$
$${\mathbb{E}}^{\mathbb{P}}[YZ]={\mathbb{E}}^{\mathbb{P}}[{\mathbb{E}}^{\mathbb{P}}[YZ\mid \mathcal{F}(t)]]$$
当$Y$是$\mathcal{F}(t)$可测
$${\mathbb{E}}^{\mathbb{P}}[{\mathbb{E}}^{\mathbb{P}}[YZ\mid \mathcal{F}(t)]]={\mathbb{E}}^{\mathbb{P}}[Y{\mathbb{E}}^{\mathbb{P}}[Z\mid \mathcal{F}(t)]]={\mathbb{E}}^{\mathbb{P}}[YZ(t)]={\mathbb{E}}^{\mathbb{Q}}[Y]$$
(3) 当$0\le s\le t\le T$. $Y$是$\mathcal{F}(t)$可测的
$${\mathbb{E}}^{\mathbb{Q}}[Y\mid \mathcal{F}(s)]=\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[YZ(t)\mid \mathcal{F}(s)]$$
$\because Z(s)是\mathcal{F}(s)$可测，且根据条件期望性质，可知整体是$\mathcal{F}(s)$可测.
不妨在$\mathbb{Q}$测度下等价转换，计算条件期望，对于$\forall A\in \mathcal{F}(s)$
$${\int }_{A}Y(w)d\mathbb{Q}(w)={\int }_A\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[YZ(t)\mid \mathcal{F}(s)]d\mathbb{Q}(w)$$
计算R.H.S
$$\begin{array}{rl}& {\int }_{\Omega }{\mathbb{I}}_A(w)\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[YZ(t)\mid \mathcal{F}(s)]d\mathbb{Q}(w)\\ & ={\mathbb{E}}^{\mathbb{Q}}[\underset{整体Y}{\underbrace{{\mathbb{I}}_A(w)\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[YZ(t)\mid \mathcal{F}(s)]}}]\\ & \stackrel{逆用性质(2)}{=}{\mathbb{E}}^{\mathbb{P}}[{\mathbb{I}}_A(w){\mathbb{E}}^{\mathbb{P}}[YZ(t)\mid \mathcal{F}(s)]]\\ & \stackrel{期望迭代}{=}{\mathbb{E}}^{\mathbb{P}}[{\mathbb{I}}_A(w)YZ(t)]\\ & \stackrel{性质(2)}{=}{\mathbb{E}}^{\mathbb{Q}}[{\mathbb{I}}_A(w)Y(w)]\\ & ={\int }_{\Omega }{\mathbb{I}}_A(w)Y(w)d\mathbb{Q}(w)\\ & ={\int }_{A}Y(w)d\mathbb{Q}(w)\\ & =\text{L.H.S}\end{array}$$

Levy定理

定理：令$M(t),t\ge 0$为域流$\mathcal{F}(t)$上的鞅，假设$M(0)=0$，$M(t)$具有连续路径，并且$[M,M](t)=t$对于任意$t\ge 0$. 则$M(t)$是布朗运动.

Girsanov

定理：令$W(t)$为概率空间$(\Omega ,\mathcal{F},\mathbb{P})$ 上的布朗运动，令$\Theta (t)$ 为域流$\mathcal{F}(t)$上的Adapted Process，定义
Z ( t ) = exp ⁡ { − ∫ 0 t Θ ( u ) d W ( u ) − 1 2 ∫ 0 t Θ 2 ( u ) d u } Z(t)=\exp\{-\int_0^t\Theta(u)dW(u)-\frac{1}{2}\int_0^t\Theta^2(u)du\} Z(t)=exp{−∫0t​Θ(u)dW(u)−21​∫0t​Θ2(u)du}
令
$$\tilde{W}(t)=W(t)+{\int }_0^t\Theta (u)du$$
设$Z=Z(T)$，那么$\mathbb{E}Z=1,Z\ge 0$. 且在$\mathbb{Q}$下，对于任意$A\in \mathcal{F}$
$$\mathbb{Q}(A)={\int }_{A}Z(w)d\mathbb{P}(w)$$
则过程$\tilde{W}(t)$是布朗运动.
证明：
根据Levy定理验证二次变差$[\tilde{W},\tilde{W}](t)$.
由表达式
$$\begin{gathered} \tilde{W}(t)=W(t)+{\int }_0^t\Theta (s)ds \\ d\tilde{W}(t)=dW(t)+\Theta (t)dt \\ d\tilde{W}(t)d\tilde{W}(t)=dW(t)dW(t)=dt \end{gathered}$$
因此$[\tilde{W},\tilde{W}](t)=t$.

计算$\mathbb{E}Z$
由
Z ( t ) = exp ⁡ { − ∫ 0 t Θ ( u ) d W ( u ) − 1 2 ∫ 0 t Θ 2 ( u ) d u ⏟ Ito Process } Z(t)=\exp\{\underbrace{-\int_0^t\Theta(u)dW(u)-\frac{1}{2}\int_0^t\Theta^2(u)du}_{\text{Ito Process}}\} Z(t)=exp{Ito Process −∫0t​Θ(u)dW(u)−21​∫0t​Θ2(u)du​​}
令$f(t,x)=e^x,f_t=0,f_x=f_{xx}=e^x$
$$dZ(t)=e^{X(t)}dX(t)+\frac{1}{2}{e}^{X(t)}dX(t)dX(t)=Z(t)-\Theta (t)dW(t)-\frac{1}{2}{\Theta }^2(t)dt+\frac{1}{2}Z(t){\Theta }^2(t)dt$$
化简得到
$$dZ(t)=-\Theta (t)Z(t)dW(t)$$
即漂移项(drift term)为0. 所以$Z(t)$是一个$\mathbb{P}-$martingale($\mathbb{P}$测度下的鞅过程).
积分得到
$$Z(t)=Z(0)-{\int }_0^t\Theta (u)Z(u)dW(u)$$
计算期望
E [ Z ] = E [ Z ( T ) ] = E [ Z ( T ) ∣ F ( 0 ) ] = Z ( 0 ) = exp ⁡ { 0 } = 1. \mathbb{E}[Z]=\mathbb{E}[Z(T)]=\mathbb{E}[Z(T)\mid \mathcal{F}(0)]=Z(0)=\exp\{0\}=1. E[Z]=E[Z(T)]=E[Z(T)∣F(0)]=Z(0)=exp{0}=1.
证明$\tilde{W}(t)$是$\mathbb{Q}-$martingale.
$${\mathbb{E}}^{\mathbb{Q}}[\tilde{W}(t)\mid \mathcal{F}(s)]$$
对于$0\le s\le t$，根据性质(3)转换为$\mathbb{P}$测度下
$${\mathbb{E}}^{\mathbb{Q}}[\tilde{W}(t)\mid \mathcal{F}(s)]=\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[\tilde{W}(t)Z(t)\mid \mathcal{F}(s)]$$
需要证明$\tilde{W}(t)Z(t)$是$\mathbb{P}-$martingale.
计算微分
$$\begin{array}{rl}d(\tilde{W}(t)Z(t))& =Z(t)d\tilde{W}(t)+\tilde{W}(t)dZ(t)+dZ(t)d\tilde{W}(t)\\ & =Z(t)[dW(t)+\Theta (t)dt]+\tilde{W}(t)[-\Theta (t)Z(t)dW(t)]+[-\Theta (t)Z(t)dW(t)][dW(t)+\Theta (t)dt]\\ & =Z(t)(1-\tilde{W}(t)\Theta (t))dW(t)\end{array}$$
可以发现$d(\tilde{W}(t)Z(t))$中没有漂移项，因此$\tilde{W}(t)Z(t)$是$\mathbb{P}-$martingale.
根据鞅的性质可得
$$\mathbb{E}[\tilde{W}(t)\mid \mathcal{F}(s)]=\frac{1}{Z(s)}{\mathbb{E}}^{\mathbb{P}}[\tilde{W}(t)Z(t)\mid \mathcal{F}(s)]=\frac{1}{Z(s)}\tilde{W}(s)Z(s)=\tilde{W}(s)$$
即$\tilde{W}(t)$是$\mathbb{Q}-$matingale.

Original Version of Girsanov

$$dX(t)=f(X(t))dt+\sigma (X(t))dW(t)$$
令$f^{\ast }(x)$是一个新的drift function，假设$\frac{f^{\ast }(x)-f(x)}{\sigma (x)}$有界.
定义测度${\mathbb{P}}^{\ast }$
d P ∗ d P ( w ) ∣ F t = exp ⁡ { − 1 2 ∫ 0 t ( f ∗ ( x s ) − f ( x s ) σ ( x s ) ) 2 d s + ∫ 0 t f ∗ ( x s ) − f ( x 0 ) σ ( x 0 ) d W s } \frac{d\mathbb{P}^*}{d\mathbb{P}}(w)\bigg |_{\mathcal{F}_t}=\exp\{-\frac{1}{2}\int_0^t(\frac{f^*(x_s)-f(x_s)}{\sigma(x_s)})^2ds+\int_0^t\frac{f^*(x_s)-f(x_0)}{\sigma(x_0)}dW_s\} dPdP∗​(w)∣∣∣∣​Ft​​=exp{−21​∫0t​(σ(xs​)f∗(xs​)−f(xs​)​)2ds+∫0t​σ(x0​)f∗(xs​)−f(x0​)​dWs​}
那么，${\mathbb{P}}^{\ast }$和概率测度$\mathbb{P}$等价.
并且，$W^{\ast }$过程满足
$$d{W}^{\ast }=-\frac{f^{\ast }(x_t)-f(x_t)}{\sigma (x_t)}dt+d{W}_t$$
是概率测度${\mathbb{P}}^{\ast }$下的布朗运动并且
$$d{X}_t=f^{\ast }(X_t)dt+\Delta (X_t)d{W}^{\ast }(t)$$

参考资料

Girsanov part 5&6 Jerry Xu