跳至主要内容

一致性(Consistency)

當我們獲得的數據量越來越多時,一個好的估計方法應該要能更準確的估計出真實的數據分布。這種性質稱為一致性。

Definition

ε>0\forall \varepsilon >0 ,we say δnnPη(θ)\delta_n \xrightarrow[n\to\infty]{P} \eta(\theta) if

Pθ(δnη(θ)ε)n1P_\theta(|\delta_n-\eta(\theta)|\le \varepsilon) \xrightarrow[n\to\infty]{} 1

也就是 δn\delta_n 以幾率收斂到 η(θ)\eta(\theta)

Theorem

δn\delta_nη(θ)\eta(\theta) 的一致估計,則對於任何符合以下要求的數列 an,bna_n, b_n

limnan=1limnbn=0\lim_{n\to\infty} a_n = 1 \quad\lim_{n\to\infty} b_n = 0

    δn=anδn+bn\implies\delta^*_n=a_n\delta_n+b_n 也是 η(θ)\eta(\theta) 的一致估計。

點估計中,有很多一致性估計。所以我們應該要捨棄那些不一致的估計方法。

Note:

ε>0\forall\varepsilon>0

Pθ(δnη(θ)>ε)=Pθ(δnη(θ)2>ε2)Eθ(δnη(θ))2ε2Chebyshev’s inequalityP_\theta(|\delta_n-\eta(\theta)|>\varepsilon) = P_\theta(|\delta_n-\eta(\theta)|^2>\varepsilon^2) \le \frac{E_\theta(\delta_n-\eta(\theta))^2}{\varepsilon^2} \quad\text{Chebyshev's inequality}     Eθ(δnη(θ))2n0\implies E_\theta(\delta_n-\eta(\theta))^2 \xrightarrow[n\to\infty]{} 0

Note that δn=Var(δn)+Bias2(δn)\delta_n=Var(\delta_n)+Bias^2(\delta_n)

i.e. If Eθ(δn)η(θ)E_\theta(\delta_n)\to\eta(\theta) and Var(δn)0Var(\delta_n)\to 0 then δn\delta_n is a consistent estimator of η(θ)\eta(\theta).

i.e. 方差會隨樣本數增加而趨近於 0 的無偏估計方法是一致的。


EX: X1,,XniidE(Xi)=θX_1, \cdots, X_n \overset{\text{iid}}{\sim}E(X_i)=\theta.

let δ\delta^* be the UMVUE of θ\theta, i.e. Var(δ)Var(δ)Var(\delta^*)\le Var(\delta) for all unbiased estimator δ\delta.

    δ\implies \delta^* is consistent.

E.g. δ=Xˉ    Var(Xˉ)=σ2n0\delta=\bar{X}\implies Var(\bar{X})=\frac{\sigma^2}{n}\to 0

大數法則

Theorem
  1. X1,,XnX_1,\cdots,X_n are iid with E(Xi)=μE(X_i)=\mu     XˉPμ\implies \bar{X}\xrightarrow{P}\mu, i.e. Xˉ\bar{X} is consistent.
  2. XnPc    g(Yn)Pg(c),gX_n\xrightarrow{P}c\implies g(Y_n)\xrightarrow{P}g(c), \forall g continuous. e.g. Xˉ2Pμ2\bar{X}^2\xrightarrow{P}\mu^2
  3. XnPc,YnPd    Xn+YnPc+dX_n\xrightarrow{P}c, Y_n\xrightarrow{P}d\implies X_n+Y_n\xrightarrow{P}c+d

EX: X1,,XniidE(Xi)=μ,Var(Xi)=σ2<X_1,\cdots,X_n\overset{\text{iid}}{\sim}E(X_i)=\mu, Var(X_i)=\sigma^2<\infty

    X12++Xn2iidE(Xi2)=Var(Xi)+E(Xi)2=σ2+μ2\implies X_1^2+\cdots+X_n^2\overset{\text{iid}}{\sim}E(X_i^2)=Var(X_i)+E(X_i)^2=\sigma^2+\mu^2

By LLN, X2ˉ=1ni=1nXi2Pσ2+μ2\bar{X^2}=\frac{1}{n}\sum_{i=1}^nX_i^2\xrightarrow{P}\sigma^2+\mu^2 and XˉPμ    Xˉ2Pμ2\bar{X}\xrightarrow{P}\mu\implies \bar{X}^2\xrightarrow{P}\mu^2

    S2=1ni=1n(XiXˉ)2=1ni=1nXi2Xˉ2Pσ2\implies S_*^2=\frac{1}{n}\sum_{i=1}^n(X_i-\bar{X})^2=\frac{1}{n}\sum_{i=1}^nX_i^2-\bar{X}^2\xrightarrow{P}\sigma^2 nn1n1    S2=nn1S2Pσ2\because \frac{n}{n-1}\xrightarrow[n\to\infty]{}1\implies S^2=\frac{n}{n-1}S_*^2\xrightarrow{P}\sigma^2

以上並沒有假設 XiX_i 的分佈,只要 E(Xi)=μ,Var(Xi)=σ2<E(X_i)=\mu, Var(X_i)=\sigma^2<\infty 即可。