UA MATH563 概率論的數學基礎 中心極限定理14 Kolmogorov maximal inequality

一個不願透露姓名的孩子發表於2020-12-28

UA MATH563 概率論的數學基礎 中心極限定理14 Kolmogorov maximal inequality

這一講介紹一個有用的不等式,它給出了獨立隨機變數的和的最值的tail probability的階。

Kolmogorov maximal inequality
假設 X 1 , ⋯   , X n X_1,\cdots,X_n X1,,Xn是獨立的隨機變數,並且 E X i = 0 , V a r X i < ∞ EX_i=0,Var X_i<\infty EXi=0,VarXi<,則
P ( max ⁡ 1 ≤ k ≤ n ∣ S k ∣ ≥ x ) ≤ V a r ( S n ) x 2 P(\max_{1 \le k \le n} |S_k| \ge x) \le \frac{Var(S_n)}{x^2} P(1knmaxSkx)x2Var(Sn)

其中
S k = ∑ i = 1 k X i S_k = \sum_{i=1}^k X_i Sk=i=1kXi

說明
與Chebyshev不等式的對比:
P ( ∣ S n ∣ ≥ x ) ≤ V a r ( S n ) x 2 P(|S_n| \ge x) \le \frac{Var(S_n)}{x^2} P(Snx)x2Var(Sn)

顯然Kolmogorov maximal inequality比Chebyshev不等式更強,雖然它們提供一樣的上界,但Chebyshev不等式只討論前 n n n項和,Kolmogorov maximal inequality討論的是前 n n n個部分和的最大值;但是需要注意的是Chebyshev不等式不要求獨立性,但Kolmogorov maximal inequality是要求的。


證明
考慮事件 { max ⁡ 1 ≤ k ≤ n ∣ S k ∣ ≥ x } \{\max_{1 \le k \le n}|S_k| \ge x\} {max1knSkx},我們可以做一個分解
{ max ⁡ 1 ≤ k ≤ n ∣ S k ∣ ≥ x } = ⨆ k ≥ 1 A k \{\max_{1 \le k \le n}|S_k| \ge x\} = \bigsqcup_{k \ge 1}A_k {1knmaxSkx}=k1Ak

其中
A k = { ∣ S k ∣ ≥ x , ∣ S j ∣ < x , ∀ j < k } A_k = \{|S_k| \ge x,|S_j|<x,\forall j <k\} Ak={Skx,Sj<x,j<k}

計算
V a r ( S n ) = E [ S n 2 ] = ∫ Ω S n 2 d P ≥ ∫ ⨆ k = 1 n A k S n 2 d P = ∑ k = 1 n ∫ A k S n 2 d P = ∑ k = 1 n ∫ A k [ S k 2 + 2 S k ( S n − S k ) + ( S n − S k ) 2 ] d P ≥ ∑ k = 1 n ∫ A k [ S k 2 + 2 S k ( S n − S k ) ] d P Var(S_n)=E[S_n^2] = \int_{\Omega} S^2_n dP \ge \int_{\bigsqcup_{k=1}^{n}A_k}S_n^2dP \\ = \sum_{k=1}^{n} \int_{A_k}S_n^2dP = \sum_{k=1}^{n} \int_{A_k}[S_k^2+2S_k(S_n-S_k)+(S_n-S_k)^2]dP \\ \ge \sum_{k=1}^{n} \int_{A_k}[S_k^2+2S_k(S_n-S_k)]dP Var(Sn)=E[Sn2]=ΩSn2dPk=1nAkSn2dP=k=1nAkSn2dP=k=1nAk[Sk2+2Sk(SnSk)+(SnSk)2]dPk=1nAk[Sk2+2Sk(SnSk)]dP

考慮第二項, 2 S k 1 A k 2S_k1_{A_k} 2Sk1Ak σ ( { X 1 , ⋯   , X k } ) \sigma(\{X_1,\cdots,X_k\}) σ({X1,,Xk})中可測, S n − S k S_n-S_k SnSk σ ( { X k + 1 , ⋯   , X n } ) \sigma(\{X_{k+1},\cdots,X_n\}) σ({Xk+1,,Xn})中可測,也就是說 2 S k 1 A k 2S_k1_{A_k} 2Sk1Ak S n − S k S_n-S_k SnSk獨立,並且 E [ S n − S k ] = 0 E[S_n-S_k]=0 E[SnSk]=0,所以
∫ A k 2 S k ( S n − S k ) d P = ∫ 2 S k 1 A k d P ∫ ( S n − S k ) d P = ∫ 2 S k 1 A k d P E [ S n − S k ] = 0 \int_{A_k}2S_k(S_n-S_k)dP=\int 2S_{k}1_{A_k}dP \int (S_n-S_k)dP \\ = \int 2S_{k}1_{A_k}dP E[S_n-S_k]=0 Ak2Sk(SnSk)dP=2Sk1AkdP(SnSk)dP=2Sk1AkdPE[SnSk]=0

因此
V a r ( S n ) ≥ ∑ k = 1 n ∫ A k S k 2 d P ≥ ∑ k = 1 ∞ ∫ A k x 2 d P = x 2 ∑ k = 1 n P ( A k ) = x 2 P ( ⨆ k ≥ 1 A k ) = x 2 P ( max ⁡ 1 ≤ k ≤ n ∣ S k ∣ ≥ x ) Var(S_n) \ge \sum_{k=1}^{n} \int_{A_k}S_k^2dP \ge \sum_{k=1}^{\infty} \int_{A_k}x^2dP=x^2 \sum_{k=1}^{n}P(A_k) \\ = x^2P(\bigsqcup_{k \ge 1}A_k )=x^2P(\max_{1 \le k \le n}|S_k| \ge x) Var(Sn)k=1nAkSk2dPk=1Akx2dP=x2k=1nP(Ak)=x2P(k1Ak)=x2P(1knmaxSkx)

相關文章