推导2

γjk\gamma_{jk}代表第j个观测值是否选择了第k个模型,1:是,0:否 γ^jk=E(γjky,θ)\hat \gamma_{jk} = E(\gamma_{jk}|y,\theta)代表在y和θ\theta已知的情况下γjk\gamma_{jk}的条件期望。

γ^jk=E(γjky,θ)=P(γjk=1y,θ)1=P(γjk=1,yjθ)kP(γjk=1,yjθ)2=P(yjγjk=1,θ)P(γjk=1θ)kP(yjγjk=1,θ)P(γjk=1θ)3=akϕ(yjθk)kakϕ(yjθk)4\begin{aligned} \hat \gamma_{jk} & = E(\gamma_{jk}|y,\theta) \\ & = P(\gamma_{jk}=1|y,\theta) && {1} \\ & = \frac{P(\gamma_{jk}=1, y_j | \theta)}{\sum_kP(\gamma_{jk}=1, y_j | \theta)} && {2} \\ & = \frac{P(y_j | \gamma_{jk}=1, \theta)P(\gamma_{jk}=1|\theta)}{\sum_kP(y_j | \gamma_{jk}=1, \theta)P(\gamma_{jk}=1|\theta)} && {3} \\ & = \frac{a_k\phi(y_j|\theta_k)}{\sum_k a_k\phi(y_j|\theta_k)} && {4} \end{aligned}

关于以上公式的说明: (1):离散变量\gamma_{jk}的取值范围为0和1。根据离散型变量的期望公式,得:

E(γjky,θ)=1P(γjk=1y,θ)+0P(γjk=0y,θ)E(\gamma_{jk}|y,\theta) = 1 * P(\gamma_{jk}=1|y,\theta) + 0 * P(\gamma_{jk}=0|y,\theta)

(2):

P(γjk=1y,θ)=P(γjk=1yj,θ)只有yjP(γjk)有关,其它y都可以忽略=P(γjk=1yj,θ)kP(γjkyj=1,θ)根据γjk的定义,kP(γjk=1)=1,yjθ无关=P(γjk=1yj,θ)P(yjθ)kP(γjkyj=1,θ)P(yjθ)分子分母同乘以P(yjθ)=P(γjk=1,yjθ)kP(γjk=1,yjθ)贝叶斯公式,P(AB)P(B)=P(A,B)\begin{aligned} P(\gamma_{jk}=1|y,\theta) & = P(\gamma_{jk}=1|y_j,\theta) & \text{只有}y_j\text{与}P(\gamma_{jk})\text{有关,其它y都可以忽略} \\ & = \frac{P(\gamma_{jk}=1|y_j,\theta)}{\sum_k P(\gamma_{jk}|y_j=1,\theta)} & \text{根据}\gamma_{jk}\text{的定义,}\sum_k P(\gamma_{jk}=1) = 1,\text{与}y_j\text{和}\theta\text{无关} \\ & = \frac{P(\gamma_{jk}=1|y_j,\theta)P(y_j|\theta)}{\sum_k P(\gamma_{jk}|y_j=1,\theta)P(y_j|\theta)} & \text{分子分母同乘以}P(y_j|\theta) \\ & = \frac{P(\gamma_{jk}=1, y_j|\theta)}{\sum_k P(\gamma_{jk}=1, y_j|\theta)} & \text{贝叶斯公式,}P(A|B)P(B) = P(A, B) \end{aligned}

(3):贝叶斯公式,P(A, B) = P(B|A)P(A) (4):P(γjk=1θ)P(\gamma_{jk}=1|\theta)代表选择模型k的概率,为ak P(yjγjk=1,θ)P(y_j | \gamma_{jk}=1, \theta)代表通过第k的模型得到yj的概率,为ϕ(yjθk)\phi(y_j|\theta_k)

Last updated