本节中将探讨条件期望的一些基本性质。


命题1(特殊的条件期望)\(X\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的随机变量,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域。

(1)如果\(X\)关于\(\mathscr{G}\)可测,则\(\mathbb{E}(X|\mathscr{G})=X,\mathrm{a.s.}\)

(2)如果\(X\)\(\mathscr{G}\)独立,则\(\mathbb{E}(X|\mathscr{G})=\mathbb{E}X,\mathrm{a.s.}\)


证明(1)\(X\)\((\Omega,\mathscr{G},\mathbb{P})\)上的可测函数,且 \[ \int_{C}X\mathrm{d}\mathbb{P}=\int_{C}X\mathrm{d}\mathbb{P},\quad\forall C\in\mathscr{G}, \] 根据定义知\(\mathbb{E}(X|\mathscr{G})=X,\mathrm{a.s.}\)

(2)\(\mathbb{E} X\)\((\Omega,\mathscr{G},\mathbb{P})\)上的可测函数,且 \[ \int_C X\mathrm{d}P=\mathbb{E}(X\cdot I_C)=\mathbb{E} X\cdot\mathbb{P}(C)=\int_C\mathbb{E}X\mathrm{d}\mathbb{P}, \] 根据定义知\(\mathbb{E}(X|\mathscr{G})=\mathbb{E}X,\mathrm{a.s.}\)

在这里应用了结论:若\(X\)\(C\)独立,则\(\mathbb{E}(X\cdot I_C)=\mathbb{E}X\cdot\mathbb{P}(C)\),证明需要使用典型方法。


命题2(复合条件期望)\(X\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的随机变量,\(\mathscr{G}_1\subset\mathscr{G}_2\subset\mathscr{F}\)\(\mathscr{F}\)的子\(\sigma\)-域,则 \[ \mathbb{E}[\mathbb{E}(X|\mathscr{G}_1)|\mathscr{G}_2]=\mathbb{E}[\mathbb{E}(X|\mathscr{G}_2)|\mathscr{G}_1]=\mathbb{E}(f|\mathscr{G}_1),\quad\mathrm{a.s.}. \]


证明 一方面,根据\(\mathbb{E}(X|\mathscr{G}_1)\)关于\(\mathscr{G}_1\subset\mathscr{G}_2\)可测,应用命题1的(1)得 \[ \mathbb{E}[\mathbb{E}(X|\mathscr{G}_1)|\mathscr{G}_2]=\mathbb{E}(f|\mathscr{G}_1),\quad\mathrm{a.s.}; \] 另外一方面,根据\(\mathbb{E}[\mathbb{E}(X|\mathscr{G}_2)|\mathscr{G}_1]\)关于\(\mathscr{G}_1\)可测,且对任意的\(C\in\mathscr{G}_1\subset\mathscr{G}_2\),都有 \[ \int_C\mathbb{E}[\mathbb{E}(X|\mathscr{G}_2)|\mathscr{G}_1]\mathrm{d}\mathbb{P}=\int_C\mathbb{E}(X|\mathscr{G}_2)\mathrm{d}\mathbb{P}=\int_CX\mathrm{d}\mathbb{P}=\int_C\mathbb{E}(X|\mathscr{G}_1)\mathrm{d}\mathbb{P}, \]\(C\)的任意性得 \[ \mathbb{E}[\mathbb{E}(X|\mathscr{G}_2)|\mathscr{G}_1]=\mathbb{E}(f|\mathscr{G}_1),\quad\mathrm{a.s.}. \] 综合以上两条可知命题成立。


命题3(单调性)\(X, Y\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的随机变量,\(X\le Y, \mathrm{a.s.}\)\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,则 \[ \mathbb{E}(X|\mathscr{G})\le\mathbb{E}(Y|\mathscr{G}),\quad\mathrm{a.s.}. \] 特别地,\(|\mathbb{E}(X|\mathscr{G})|\le\mathbb{E}(|X||\mathscr{G}),\mathrm{a.s.}\)


证明 对任意的\(C\in\mathscr{G}\),都有 \[ \int_C\mathbb{E}(X|\mathscr{G})\mathrm{d}\mathbb{P}=\int_CX\mathrm{d}\mathbb{P}\le\int_C Y\mathrm{d}\mathbb{P}=\int_C\mathbb{E}(Y|\mathscr{G})\mathrm{d}\mathbb{P}, \]\(C\)的任意性得 \[ \mathbb{E}(X|\mathscr{G})\le\mathbb{E}(Y|\mathscr{G}),\quad\mathrm{a.s.}. \] 另外,注意到\(-X\le|X|, X\le|X|\), 因此 \[ \begin{cases} -\mathbb{E}(X|\mathscr{G})\le\mathbb{E}(|X||\mathscr{G}),&\mathrm{a.s.}, \\ \mathbb{E}(X|\mathscr{G})\le\mathbb{E}(|X||\mathscr{G}),&\mathrm{a.s.}. \end{cases} \] 综合以上两式,有\(|\mathbb{E}(X|\mathscr{G})|\le\mathbb{E}(|X||\mathscr{G}),\mathrm{a.s.}\)


命题4(可加性)\(X, Y\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的随机变量,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,\(a, b\in\mathbb{R}\),且\(a\mathbb{E}X+b\mathbb{E}Y\)存在,则 \[ \mathbb{E}(aX+bY|\mathscr{G})=a\mathbb{E}(X|\mathscr{G})+b\mathbb{E}(Y|\mathscr{G}),\quad\mathrm{a.s.}. \]


证明\(\mathbb{E}(aX+bY)=a\mathbb{E}X+b\mathbb{E}Y\)存在,知\(\mathbb{E}(aX+bY|\mathscr{G})\)有定义。对任意的\(C\in\mathscr{G}\),都有 \[ \begin{aligned} \int_C(aX+bY)\mathrm{d}\mathbb{P}&=a\int_CX\mathrm{d}\mathbb{P}+b\int_CY\mathrm{d}\mathbb{P}\\ &=a\int_C\mathbb{E}(X|\mathscr{G})\mathrm{d}\mathbb{P}+b\int_C\mathbb{E}(Y|\mathscr{G})\mathrm{d}\mathbb{P}\\ &=\int_C(a\mathbb{E}(X|\mathscr{G})+b\mathbb{E}(Y|\mathscr{G}))\mathrm{d}\mathbb{P}. \end{aligned} \] 因此,\(\mathbb{E}(aX+bY|\mathscr{G})=a\mathbb{E}(X|\mathscr{G})+b\mathbb{E}(Y|\mathscr{G}),\mathrm{a.s.}\)


命题3和4都是期望\(\mathbb{E}(\cdot)\)的性质向条件期望\(\mathbb{E}(\cdot|\mathscr{G})\)的推广。联想到期望所满足的单调收敛定理、Fatou引理及Lebesgue控制收敛定理,我们接下来证明这些定理对于条件期望也是成立的。


定理5(单调收敛定理)\(\{X_n\}\)\(X\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的积分存在的随机变量,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,若\(0\le X_n\uparrow X,\mathrm{a.s.}\),则 \[ 0\le\mathbb{E}(X_n|\mathscr{G})\uparrow\mathbb{E}(X|\mathscr{G}),\quad\mathrm{a.s.}. \]


证明 由命题3得\(\{\mathbb{E}(X_n|\mathscr{G})\}\)单调递增,且根据\(\mathbb{E}(X_n|\mathscr{G})\)关于\(\mathscr{G}\)可测知\(\displaystyle\lim_{n\to\infty}\mathbb{E}(X_n|\mathscr{G})\)关于\(\mathscr{G}\)可测。对任意的\(C\in\mathscr{G}\),应用单调收敛定理得 \[ \int_C\lim_{n\to\infty}\mathbb{E}(X_n|\mathscr{G})\mathrm{d}\mathbb{P}=\lim_{n\to\infty}\int_C\mathbb{E}(X_n|\mathscr{G})\mathrm{d}\mathbb{P}=\lim_{n\to\infty}\int_CX_n\mathrm{d}\mathbb{P}=\int_CX\mathrm{d}\mathbb{P}, \] 这便说明了\(\mathbb{E}(X_n|\mathscr{G})\uparrow\mathbb{E}(X|\mathscr{G}),\mathrm{a.s.}\)


定理6(Fatou引理)\(\{X_n\}\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的积分存在的随机变量序列,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,若\(X_n\geq 0,\mathrm{a.s.}\),则 \[ \mathbb{E}\left(\liminf_{n\to\infty}X_n|\mathscr{G}\right)\le\liminf_{n\to\infty}\mathbb{E}(X_n|\mathscr{G}),\quad\mathrm{a.s.}. \]


证明 对任意的\(C\in\mathscr{G}\),应用Fatou引理得 \[ \int_C\mathbb{E}\left(\liminf_{n\to\infty}X_n|\mathscr{G}\right)\mathrm{d}\mathbb{P}=\int_C\liminf_{n\to\infty}X_n\mathrm{d}\mathbb{P}\le\liminf_{n\to\infty}\int_CX_n\mathrm{d}\mathbb{P}, \] 这便说明了\(\displaystyle\mathbb{E}\left(\liminf_{n\to\infty}X_n|\mathscr{G}\right)\le\liminf_{n\to\infty}\mathbb{E}(X_n|\mathscr{G}),\mathrm{a.s.}\)


定理7(Lebesgue控制收敛定理)\(\{X_n\}\)\(X\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的积分存在的随机变量,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,若\(\displaystyle\lim_{n\to\infty}X_n=X\),且存在\(Y\in L^1(\Omega)\),使得对任意的\(n\geq 1\),都有\(|X_n|\le Y,\mathrm{a.s.}\),则 \[ \lim_{n\to\infty}\mathbb{E}(X_n|\mathscr{G})=\mathbb{E}(X|\mathscr{G}),\quad\mathrm{a.s.}. \]


证明 由命题3知 \[ |\mathbb{E}(X_n|\mathscr{G})|\le\mathbb{E}(|X_n||\mathscr{G})\le\mathbb{E}(Y|\mathscr{G})\in L^1(\Omega). \] 对任意的\(C\in\mathscr{G}\),应用Lebesgue控制收敛定理得 \[ \int_C\lim_{n\to\infty}\mathbb{E}(X_n|\mathscr{G})\mathrm{d}\mathbb{P}=\lim_{n\to\infty}\int_C\mathbb{E}(X_n|\mathscr{G})\mathrm{d}\mathbb{P}=\lim_{n\to\infty}\int_CX_n\mathrm{d}\mathbb{P}=\int_C X\mathrm{d}\mathbb{P}, \] 这便说明了\(\displaystyle\lim_{n\to\infty}\mathbb{E}(X_n|\mathscr{G})=\mathbb{E}(X|\mathscr{G}),\mathrm{a.s.}\)


定理5、6和7的证明过程说明了,条件期望是离不开期望的。在处理条件期望的问题时,常常需要借助条件期望的定义转化为期望。最后,作为定理5的应用,我们引出如下的重要结论,其在随机过程中有应用。


命题8\(X, Y\)是概率空间\((\Omega,\mathscr{F},\mathbb{P})\)上的随机变量,\(\mathscr{G}\)\(\mathscr{F}\)的子\(\sigma\)-域,\(X\)\(XY\)的积分存在,且\(Y\)关于\(\mathscr{G}\)可测,则 \[ \mathbb{E}(XY|\mathscr{G})=Y\cdot\mathbb{E}(X|\mathscr{G}),\quad\mathrm{a.s.}. \]


证明 用典型方法。


例9\(\{(X_n,\mathscr{F}_n),n\geq0\}\)是带流的随机过程,也即对任意的\(n\geq0\)\(X_n\)是关于\(\mathscr{F}_n\)可测的随机变量,且\(\mathscr{F}_n\subset\mathscr{F}_{n+1}\)。设其期望为0,方差为\(\sigma^2<\infty\)。若其是一个鞅差过程,也即对任意的\(n\geq0\),都有 \[ \mathbb{E}(X_{n+1}|\mathscr{F}_n)=0,\quad\mathrm{a.s.}, \] 则其一定是一个白噪声过程,也即对任意的\(n\geq0\),都有 \[ \mathrm{Cov}(X_n,X_{n+1})=0. \] 这是因为,\(X_n\)是关于\(\mathscr{F}_n\)可测的,从而 \[ \mathbb{E}(X_nX_{n+1}|\mathscr{F}_n)=X_n\cdot\mathbb{E}(X_{n+1}|\mathscr{F}_n)=0,\quad\mathrm{a.s.}, \] 对上式两边取期望得 \[ \mathrm{Cov}(X_n,X_{n+1})=\mathbb{E}[\mathbb{E}(X_nX_{n+1}|\mathscr{F}_n)]=0. \]