中心极限定理
令随机变量 Z ∼ N ( 0 , 1 ) Z \sim \mathcal N(0,1) Z∼N(0,1),对于独立同分布的随机变量(i.i.d.) X 1 , X 2 , … , X n X_1, X_2, \ldots, X_n X1,X2,…,Xn,令 S n = ∑ i X i S_n = \sum_{i} X_i Sn=∑iXi, Z n = S n − E [ S n ] var [ S n ] Z_n=\frac{S_n - \mathbb E[S_n]}{\sqrt {\text{var}[S_n]}} Zn=var[Sn]Sn−E[Sn],当 n → ∞ n \rightarrow \infty n→∞时,有
Z n → Z ∀ u Pr [ Z n ≤ u ] → Pr [ Z ≤ u ] Z_n \rightarrow Z \\ \forall u \ \ \text{Pr}[Z_n \leq u] \rightarrow \text{Pr}[Z \leq u] Zn→Z∀u Pr[Zn≤u]→Pr[Z≤u]
或者可以描述为
∀ ϵ > 0 , ∃ N ∈ N , for n > N , u ∈ R ∣ Pr [ Z n ≤ u ] − Pr [ Z ≤ u ] ∣ < ϵ \forall \epsilon > 0, \exist N \in \mathcal N, \text{ for } n>N, u \in \R \\ |\text{Pr}[Z_n \leq u] - \text{Pr}[Z \leq u]| < \epsilon ∀ϵ>0,∃N∈N, for n>N,u∈R∣Pr[Zn≤u]−Pr[Z≤u]∣<ϵ
然而,在设计和分析算法时,为了性能保证,我们需要知道算法的收敛率(convergence rate),这个时候中心极限定理就不再适用,这就需要Berry-Esseen中心极限定理来具体化误差边界(error bound)。
Berry-Esseen中心极限定理
对于独立的随机变量 X 1 , X 2 , … , X n X_1, X_2, \ldots, X_n X1,X2,…,Xn,不失一般性,令 E [ X i ] = 0 \mathbb E[X_i] = 0 E[Xi]=0, var [ X i ] = σ 2 \text{var}[X_i] = \sigma^2 var[Xi]=σ2,并且 ∑ i = 1 n σ 2 = 1 \sum_{i=1}^n \sigma^2 = 1 ∑i=1nσ2=1,让 S = X 1 + … + X n S= X_1 + \ldots + X_n S=X1+…+Xn(注意到 E [ S ] = 0 \mathbb E[S]=0 E[S]=0, var [ S ] = 1 \text{var}[S]=1 var[S]=1),那么
∀ u ∈ R ∣ Pr [ S ≤ u ] − Pr ⏟ Z ∼ N ( 0 , 1 ) [ Z ≤ u ] ∣ ≤ O ( 1 ) ⋅ β , where β = ∑ i = 1 n E ∣ X i ∣ 3 \forall u \in \R \ \ \ \ \Big | {\text{Pr}[S \leq u] - \underbrace{\text{Pr}}_{Z \sim\mathcal N(0,1)} [Z \leq u] } \Big | \leq \mathcal O(1) \cdot \beta, \ \ \ \ \text{where } \beta=\sum_{i=1}^n \mathbb E |X_i|^3 ∀u∈R ∣∣∣Pr[S≤u]−Z∼N(0,1) Pr[Z≤u]∣∣∣≤O(1)⋅β, where β=i=1∑nE∣Xi∣3
注意1:
O
(
1
)
→
0.5514
\mathcal O(1) \rightarrow 0.5514
O(1)→0.5514 [2]
注意2:Berry-Esseen中心极限定理不要求随机变量
X
i
,
∀
i
X_i,\ \forall i
Xi, ∀i是同分布的(identical),但是变量之间要求独立。
举例:以AMP最初的假设为例,令随机变量
X i = { + 1 N w . p . 1 2 − 1 N w . p . 1 2 X_i=\left\{ \begin{aligned} & +\frac{1}{\sqrt N} \ \ \ \ w.p. \frac{1}{2} \\ & -\frac{1}{\sqrt N} \ \ \ \ w.p. \frac{1}{2} \\ \end{aligned} \right. Xi=⎩⎪⎪⎨⎪⎪⎧+N1 w.p.21−N1 w.p.21
显然, E [ X i ] = 0 \mathbb E[X_i] = 0 E[Xi]=0, var [ X i ] = 1 N \text{var}[X_i] = \frac{1}{N} var[Xi]=N1,并且 ∑ i = 1 n var [ X i ] = 1 \sum_{i=1}^n \text{var}[X_i] = 1 ∑i=1nvar[Xi]=1,考虑三阶矩, E ∣ X i ∣ 3 = 1 N 3 / 2 \mathbb E |X_i|^3=\frac{1}{N^{3/2}} E∣Xi∣3=N3/21,因此 β = 1 N \beta = \frac{1}{\sqrt N} β=N1,根据Berry-Esseen中心极限定理,可以有
∀ u ∈ R ∣ Pr [ S ≤ u ] − Pr ⏟ Z ∼ N ( 0 , 1 ) [ Z ≤ u ] ∣ ≤ 0.5514 n \forall u \in \R \ \ \ \ \Big | {\text{Pr}[S \leq u] - \underbrace{\text{Pr}}_{Z \sim\mathcal N(0,1)} [Z \leq u] } \Big | \leq \frac{0.5514}{\sqrt n} ∀u∈R ∣∣∣Pr[S≤u]−Z∼N(0,1) Pr[Z≤u]∣∣∣≤n0.5514
其中 0.5514 n \frac{0.5514}{\sqrt n} n0.5514给出了收敛率的表征。
关于Berry-Esseen中心极限定理的证明可以看我写的另一篇博客Berry-Esseen中心极限定理。
参考
[1] http://yuanz.web.illinois.edu/teaching/IE498fa19/lec_01_02.pdf
[2] I. G. Shevtsova, “On the absolute constants in the Berry–Esseen inequality and its structural and nonuniform improvements”, Inform. Primen., 7:1 (2013), 124–125