UA MATH564 概率论VI 数理统计基础3 卡方分布下 Cochran定理

UA MATH564 概率论VI 数理统计基础3 卡方分布下 Cochran定理

这一讲介绍多元正态随机变量的二次型的相关性质以及非常常用的Cochran定理。假设 X 1 , , X n X_1,\cdots,X_n 互相独立,记 X = [ X 1 , , X n ] N n ( a , I n ) X = [X_1,\cdots,X_n]' \sim N_n(a,I_n) a = [ a 1 , , a n ] a = [a_1,\cdots,a_n]'

多元正态随机变量二次型的分布

假设 A A n n 阶对称幂等矩阵,记 r = r a n k ( A ) r=rank(A) ,则 Y = X A X χ r , δ 2 , δ = a A a Y=X'AX\sim \chi^2_{r,\delta},\delta=a'Aa

证明
充分性。因为 A A 是幂等矩阵,因此 A A 的特征值为0或1, r = r a n k ( A ) r=rank(A) ,因此 A A 的特征值中有 r r 个1,由于 A A 是对称矩阵,因此 P , P P = P P = I n \exists P,P'P=PP'=I_n P A P = d i a g ( I r , 0 ) PAP'=diag(I_r,0) 。做正交变换 Z = P X Z=PX ,则 Z N n ( P a , I n ) Z \sim N_n(Pa,I_n) Y = X A X = Z P A P Z = Z d i a g ( I r , 0 ) Z = i = 1 r Z i 2 χ r , δ 2 δ 2 = i = 1 r E Z i 2 = ( E Z ) d i a g ( I r , 0 ) ( E Z ) = ( E Z ) P A P ( E Z ) = ( E Z ) P A P ( E Z ) = E ( P Z ) A E ( P Z ) = a A a Y=X'AX=Z'PAP'Z=Z'diag(I_r,0)Z=\sum_{i=1}^r Z_i^2 \sim \chi^2_{r,\delta} \\ \delta^2 = \sum_{i=1}^r EZ_i^2 = (EZ)'diag(I_r,0)(EZ)=(EZ)'PAP'(EZ) \\ = (EZ)'P'AP(EZ)=E(PZ)'AE(PZ)=a'Aa
必要性。因为 r = r a n k ( A ) r=rank(A) ,假设 A A 的非零特征根为 λ 1 , λ 2 , , λ r \lambda_1,\lambda_2,\cdots,\lambda_r Q , Q Q = Q Q = I n \exists Q,Q'Q=QQ'=I_n Q A Q = d i a g ( λ 1 , , λ r , 0 , , 0 ) QAQ'=diag(\lambda_1,\cdots,\lambda_r,0,\cdots,0) ,做正交变换 Z = P X Z=PX
Y = X A X = Z Q A Q Z = i = 1 r λ i Z i 2 Y = X'AX = Z'QAQ'Z = \sum_{i=1}^r \lambda_i Z_i^2

其中 Z N n ( P a , I n ) Z \sim N_n(Pa,I_n) ,记 Z i N ( c i , 1 ) , i = 1 , , n Z_i \sim N(c_i,1),i=1,\cdots,n ,可以用特征函数确定 Y Y 的分布。计算 λ j Z j 2 \lambda_j Z_j^2 的特征函数,
κ λ j Z j 2 ( t ) = E [ e i t λ j Z j 2 ] = e i t λ x 2 1 2 π e ( x c j ) 2 2 d x = 1 1 2 i λ j t e i λ j t c j 1 2 i λ j t \kappa_{\lambda_j Z_j^2}(t) = E[e^{it\lambda_j Z_j^2}] = \int_{-\infty}^{\infty} e^{it\lambda x^2} \frac{1}{\sqrt{2\pi}}e^{-\frac{(x-c_j)^2}{2}}dx= \frac{1}{\sqrt{1-2i\lambda_j t}}e^{\frac{i\lambda_j t c_j}{1-2i\lambda_j t}}

进一步计算 Y Y 的特征函数为
κ Y ( t ) = j = 1 r κ λ j Z j 2 ( t ) = j = 1 r 1 1 2 i λ j t e i λ j t c j 1 2 i λ j t \kappa_Y(t) = \prod_{j=1}^r \kappa_{\lambda_j Z_j^2}(t) = \prod_{j=1}^r \frac{1}{\sqrt{1-2i\lambda_j t}}e^{\frac{i\lambda_j t c_j}{1-2i\lambda_j t}}

λ 1 = λ 2 = = λ r = 1 \lambda_1 = \lambda_2 = \cdots = \lambda_r = 1 时,这个特征函数成为 χ r , δ 2 \chi^2_{r,\delta} 的特征函数,因此 A A 为幂等矩阵。

证毕

Cochran定理

假设 X N n ( a , I n ) X \sim N_n(a,I_n) X X = i = 1 m X A i X X'X = \sum_{i=1}^m X'A_iX ,则 X A i X χ n i , δ i 2 X'A_iX \sim \chi^2_{n_i,\delta_i} ,当且仅当 i = 1 m r a n k ( A i ) = n \sum_{i=1}^m rank(A_i)=n X A i X X'A_iX 互相独立,此时 n i = r a n k ( A i ) , δ i 2 = a A i a n_i = rank(A_i),\delta_i^2 = a'A_ia

证明
充分性。记 Q i = X A i X Q_i = X'A_iX ,假设 B B 是一个正交矩阵,则
i = 1 m Q i = X X = X I n X = X B B X = i = 1 m X A i X = X ( i = 1 m A i ) X \sum_{i=1}^m Q_i = X'X = X'I_nX = X'B'BX=\sum_{i=1}^m X'A_iX = X'\left(\sum_{i=1}^m A_i \right) X
做正交变换 Z = B X Z = BX ,则 Z N ( B a , I n ) Z \sim N(Ba,I_n) ,并且
i = 1 m Q i = Z Z = j = 1 n Z j 2 = i = 1 m j = e i 1 + 1 e i Z j 2 \sum_{i=1}^m Q_i = Z'Z = \sum_{j=1}^n Z_j^2 = \sum_{i=1}^m \sum_{j=e_{i-1}+1}^{e_i} Z_j^2
其中 e i = j = 1 i n j e_{i} = \sum_{j=1}^i n_j ,假设 B B 的选取使得
Q i = j = e i 1 + 1 e i Z j 2 = X A i X , i = 1 , 2 , , m Q_i = \sum_{j=e_{i-1}+1}^{e_i} Z_j^2 = X'A_iX,\forall i=1,2,\cdots,m
根据上面多元正态随机变量二次型的分布性质的证明, Q i χ n i , δ i 2 Q_i \sim \chi^2_{n_i,\delta_i}

必要性的证明需要另一个性质:
X A X χ m , δ 2 ,    X A 1 X χ m , δ 1 2 , A A 1 0 X ( A A 1 ) X χ m m 1 , δ 2 2 X'AX \sim \chi^2_{m,\delta},\ \ X'A_1X \sim \chi^2_{m,\delta_1}, A-A_1 \ge 0 \\ \Rightarrow X'(A-A_1)X \sim \chi^2_{m-m_1,\delta_2}

必要性就是把这个性质从两个矩阵( A A 1 , A 1 A-A_1,A_1 )推广到有限个( A 1 , , A m A_1,\cdots,A_m ),用数学归纳法就可以证明;这条性质本身可以用特征函数验证。

证毕

猜你喜欢

转载自blog.csdn.net/weixin_44207974/article/details/106727653