条件独立5条重要性质及其证明
条件独立
设 $V = {V_1, V_2, \dots}$ 表示变量的有限集合。设 $P(⋅)$ 是 $V$ 中变量的联合概率分布函数。$X, Y, Z, W$ 表示 $V$ 中变量的子集,即$X, Y, Z, W \in V$。当给定 $Z$ 时,如果
$$
P(x \mid y, z) = P(x \mid z)\quad\quad P(y, z) \gt 0
$$
则 $X, Y$ 条件独立。
我们用符号 $(X {\perp!!!\perp} Y \mid Z)$ 表示条件独立,即
$$
(X {\perp!!!\perp} Y \mid Z) \iff P(x \mid y, z) = P(x \mid z)
$$
条件独立有5条重要的性质:
对称性:$(X {\perp!!!\perp} Y \mid Z) \implies (X {\perp!!!\perp} Y \mid Z)$
$$
\tag{1}(X {\perp!!!\perp} Y \mid Z) \iff P(x \mid y,z)=P(x \mid z)
$$
$$
\tag{2}(Y {\perp!!!\perp} X \mid Z) \iff P(y \mid x,z)=P(y \mid z)
$$
利用乘法公式,1式和2式可以改写为:
$$
\tag{3}\frac {P(x,y,z)} {\boxed{P(y,z)}} = \frac {\boxed{P(x,z)}} {P(z)}
$$
$$
\tag{4}\frac {P(x,y,z)} {P(x,z)} = \frac {P(y,z)} {P(z)}
$$
3式方框标注部分位置对换一下,就是4式。$\quad\blacksquare$
分解性:$(X {\perp!!!\perp} YW \mid Z) \implies (X {\perp!!!\perp} Y \mid Z)$
$$
\tag{1}(X {\perp!!!\perp} YW \mid Z) \iff P(x \mid y,z,w) = P(x \mid z)
$$
$$
\tag{2}(X {\perp!!!\perp} Y \mid Z) \iff P(x \mid y,z) = P(x \mid z)
$$
根据公式:$P(A \mid K) = \displaystyle{\sum_{i} P(A \mid B_i, K)P(Bi \mid K)}$
$$
\tag{3}P(x \mid y,z) = \sum{w \in W} P(x \mid y,z,w)P(w \mid y,z)
$$
根据1式,$P(x \mid y,z,w)=P(x \mid z)$,带入3式得:
$$
\tag{4}P(x \mid y,z) = \sum{w \in W} P(x \mid z)P(w \mid y,z) = P(x \mid z) \sum{w \in W}P(w \mid y,z)
$$
因为对任意$\forall w \in W$的情况都取到了,所以$\displaystyle{\sum_{w \in W}P(w \mid y,z)}$求和的结果就是1:
$$
\therefore \tag{5}P(x \mid y,z) = P(x \mid z) \quad \blacksquare
$$
弱连性:$(X {\perp!!!\perp} Y W\mid Z) \implies (X {\perp!!!\perp} Y \mid ZW)$
$$
\tag{1}(X {\perp!!!\perp} YW \mid Z) \iff P(x \mid y,z,w)=P(x \mid z)
$$
$$
\tag{2}(X {\perp!!!\perp} Y \mid ZW) \iff P(x \mid y,z,w)=P(x \mid z,w)
$$
根据公式:$P(A \mid K) = \displaystyle{\sum_{i} P(A \mid B_i, K)P(Bi \mid K)}$
$$
\tag{3}P(x \mid z,w) = \sum{y \in Y} P(x \mid y,z,w)P(y \mid z,w)
$$
根据1式$P(x \mid y,z,w)=P(x \mid z)$,带入3式得:
$$
\tag{4}P(x \mid z,w) = \sum{y \in Y} P(x \mid z)P(y \mid z,w) = P(x \mid z) \sum{y \in Y} P(y \mid z,w)
$$
因为对任意$\forall y \in Y$的情况都取到了,所以$\displaystyle\sum_{y \in Y} {P(y \mid z,w)} = 1$,
$$
\therefore P(x \mid z,w) = P(x \mid z) = P(x \mid y,z,w) \quad \blacksquare
$$
缩并性:$(X {\perp!!!\perp} Y \mid Z) \& (X {\perp!!!\perp} W \mid ZY) \implies (X {\perp!!!\perp} YW \mid Z)$
$$
\begin{cases}
(X {\perp!!!\perp} Y \mid Z) &\iff P(x \mid y,z)=P(x \mid z) &\text{(1)}\
(X {\perp!!!\perp} W \mid ZY) &\iff P(x \mid y,z,w)=P(x \mid y,z)&\text{(2)}
\end{cases}
$$
要证明:
$$
\tag{3}(X {\perp!!!\perp} YW \mid Z) \iff P(x \mid y,z,w)=P(x \mid z)
$$
1式左边和2式右边相同,两式一合并:
$$
\tag{4}P(x \mid y,z,w) = P(x \mid z)
$$
4式就是要证明的3式。$\quad \blacksquare$
相交性:$(X {\perp!!!\perp} W \mid ZY) \& (X {\perp!!!\perp} Y \mid ZW) \implies (X {\perp!!!\perp} YW \mid Z)$
$$
\begin{cases}
(X {\perp!!!\perp} W \mid ZY) \iff P(x \mid y,z,w)=P(x \mid y,z) &\text{(1)}\
(X {\perp!!!\perp} Y \mid ZW) \iff P(x \mid y,z,w)=P(x \mid z,w) &\text{(2)}
\end{cases}
$$
要证明:
$$
\tag{3}(X {\perp!!!\perp} YW \mid Z) \iff P(x \mid y,z,w)=P(x \mid z)
$$
1式2式左边相同,两式一合并:
$$
\tag{4}P(x \mid y,z) = P(x \mid z,w)
$$
利用乘法公式,4式可写做:
$$
\tag{5}\frac {P(x,y,z)} {\boxed{P(y,z)}} = \frac {\boxed{P(x,z,w)}} {P(z,w)}
$$
交换5式方框部分,并对$Y$进行边缘化:
$$
\tag{6}\frac {\displaystyle\sum{Y} P(x,y,z)} {P(x,z,w)} = \frac {\displaystyle\sum{Y} P(y,z)} {P(z,w)}
$$
根据全概率公式,得:
$$
\tag{7}\frac {P(x,z)} {\boxed{P(x,z,w)}} = \frac {\boxed{P(z)}} {P(z,w)}
$$
交换方框部分,得
$$
\tag{8}\frac {P(x,z)} {P(z)} = \frac {P(x,z,w)} {P(z,w)}
$$
将8式写成条件概率
$$
\tag{9}P(x \mid z)=P(x \mid z,w)
$$
9式和2式一合并,即可得:$P(x \mid y,z,w)=P(x \mid z)$,这就是我们要证明的3式。$\quad \blacksquare$