其他分享
首页 > 其他分享> > 统计推断(三) Exponential Family

统计推断(三) Exponential Family

作者:互联网

1. Exponential family

2. Sufficient statistics

2.1 Non-Bayesian case

Theorem 1(likelihood characterization):

t(y)t(y)t(y) is sufficient w.r.t p(y;x)p(y;x)p(y;x)      py(y;x)pt(t(y);x)\iff \ \frac{p_{y}(y;x)}{p_t(t(y);x)}⟺ pt​(t(y);x)py​(y;x)​ doesn’t depend on x, for all x and y

Proof:omit…

Theorem 2(Neyman Factorization theorem):

t(y)t(y)t(y) is sufficient w.r.t p(y;x)p(y;x)p(y;x)      a(,)b()使  p(y;x)=a(t(y),x)b(y)\iff \ 存在a(\cdot,\cdot)和b(\cdot)使得 \ \ p(y;x)=a\left(t(y),x\right) \cdot b(y)⟺ 存在a(⋅,⋅)和b(⋅)使得  p(y;x)=a(t(y),x)⋅b(y)

Proof:omit…

Theorem:complete \Longrightarrow⟹ minimal

Proof:假设 t 为complete,s 为 minimal,存在 s=g(t)s=g(t)s=g(t),E[t]=E[E[ts=s]]E[t]=E\left[E\left[t|s=s\right]\right]E[t]=E[E[t∣s=s]]

E[ts=s]=f(s)=f(g(t))=f~(t)E[t|s=s]=f(s)=f(g(t))=\tilde{f}(t)E[t∣s=s]=f(s)=f(g(t))=f~​(t)

ϕ(t)=tf~(t)\phi(t)=t-\tilde{f}(t)ϕ(t)=t−f~​(t),有 E[ϕ(t)]=0E[\phi(t)] = 0E[ϕ(t)]=0

根据 complete 的定义,有 ϕ(t)0t=f~(t)=f(s)\phi(t)\equiv0 \Longrightarrow t = \tilde{f}(t)=f(s)ϕ(t)≡0⟹t=f~​(t)=f(s)

故 t 也是 minimal

2.2 Bayesian case

Theorem(Belief characterization):

t(y)t(y)t(y) is sufficient w.r.t p(y,x)p(y,x)p(y,x)      p(xy)=p(xt(y))\iff \ p(x|y)=p(x|t(y))⟺ p(x∣y)=p(x∣t(y)), for all x and y

Proof:omit…

Theorem(Neyman Factorization theorem):

t(y)t(y)t(y) is sufficient w.r.t p(y,x)p(y,x)p(y,x)      p(yx)=p(t(y)x)p(yt(y))\iff \ p(y|x)=p(t(y)|x)\cdot p(y|t(y))⟺ p(y∣x)=p(t(y)∣x)⋅p(y∣t(y)), for all x and y

Proof:omit…

3. Conjugate priors

Bonennult 发布了37 篇原创文章 · 获赞 27 · 访问量 2万+ 私信 关注

标签:frac,Family,Exponential,ln,py,cdot,推断,alpha,dot
来源: https://blog.csdn.net/weixin_41024483/article/details/104165233