指数族分布(2)
指数族分布(1):
[https://blog.csdn.net/RSstudent/article/details/127465224?spm=1001.2014.3001.5501]
典型形式指数族分布在矩、累积量的计算方面存在方便之处,包括期望、方差。
定义:令
T
∈
R
s
T\in \mathbb{R}^s
T∈Rs,Moment generating function(MGF) of
T
T
T定义为
M
T
(
u
)
=
E
[
e
u
T
]
M_T(u)=E[e^{uT}]
MT(u)=E[euT]
累计生成函数CGF:
K
T
(
u
)
=
l
o
g
M
T
(
u
)
K_T(u)=logM_T(u)
KT(u)=logMT(u)
引理:
如果MGF M X ( u ) M_X(u) MX(u)和 M Y ( u ) M_Y(u) MY(u)对于随机向量 X X X和 Y Y Y有限 ,且一致在某个非空集合的内点 u u u,则 P X = P Y P_X=P_Y PX=PY
T
1
,
⋯
,
T
s
T_1, \cdots,T_s
T1,⋯,Ts的幂次的期望称为
T
T
T的矩
α
r
1
,
r
2
,
⋯
,
r
s
=
E
[
T
1
r
1
×
T
2
r
2
×
⋯
×
T
s
r
s
]
\alpha_{r_1,r_2, \cdots,r_s}=E[T_1^{r_1}\times T_2^{r_2}\times\cdots\times T_s^{r_s}]
αr1,r2,⋯,rs=E[T1r1×T2r2×⋯×Tsrs]
通过在
u
=
0
u=0
u=0点求取MGF的导数,可以获得这些矩。
定理1
若
M
T
M_T
MT在远点的某个邻域内有限,且在原点具有个各阶连续导数
α
r
1
,
r
2
,
⋯
,
r
s
=
∂
r
1
∂
u
1
r
1
⋯
∂
r
s
∂
u
s
r
s
M
T
(
u
)
∣
u
=
0
\alpha_{r_1,r_2, \cdots,r_s}=\frac{\partial^{r_1}}{\partial u_1^{r_1}}\cdots\frac{\partial^{r_s}}{\partial u_s^{r_s}}M_T(u)|_{u=0}
αr1,r2,⋯,rs=∂u1r1∂r1⋯∂usrs∂rsMT(u)∣u=0
这是矩,相应的
K
T
(
u
)
K_T(u)
KT(u)的导数称为累积量
κ
r
1
,
r
2
,
⋯
,
r
s
=
∂
r
1
∂
u
1
r
1
⋯
∂
r
s
∂
u
s
r
s
K
T
(
u
)
∣
u
=
0
\kappa_{r_1,r_2, \cdots,r_s}=\frac{\partial^{r_1}}{\partial u_1^{r_1}}\cdots\frac{\partial^{r_s}}{\partial u_s^{r_s}}K_T(u)|_{u=0}
κr1,r2,⋯,rs=∂u1r1∂r1⋯∂usrs∂rsKT(u)∣u=0
当
s
=
1
s=1
s=1的时候,
K
T
′
=
(
l
o
g
M
T
)
′
=
M
T
′
M
T
K_T^{'}=(logM_T)'=\frac{M_T'}{M_T}
KT′=(logMT)′=MTMT′,以及
K
T
′
′
=
M
T
′
′
M
T
−
M
T
′
2
M
T
2
K_T''=\frac{M_T''M_T-M_T'^2}{M_T^2}
KT′′=MT2MT′′MT−MT′2
可以发现,取导数在
u
=
0
u=0
u=0,就
M
T
′
=
E
[
T
e
u
T
]
M_T^{'}=E[Te^{uT}]
MT′=E[TeuT]
M T ′ = E [ e u T ] M_T'=E[e^{uT}] MT′=E[euT]
κ 1 = K T ′ ∣ u = 0 = E [ T ] E [ 1 ] = E [ T ] \kappa_1=K_T'|_{u=0}=\frac{E[T]}{E[1]}=E[T] κ1=KT′∣u=0=E[1]E[T]=E[T]
M T ′ ′ = E [ T 2 e u T ] M_T''=E[T^2e^{uT}] MT′′=E[T2euT]
κ 2 = E [ T 2 ] − E [ T ] 2 = V a r ( T ) \kappa_2=E[T^2]-E[T]^2=Var(T) κ2=E[T2]−E[T]2=Var(T)
定理2
设
X
X
X和
Y
Y
Y是独立随机变量。若
X
X
X和
Y
Y
Y均为正的,或者
E
∣
X
∣
E|X|
E∣X∣和
E
∣
Y
∣
E|Y|
E∣Y∣有限(Fubini定理条件),则
E
[
X
Y
]
=
E
[
X
]
E
[
Y
]
E[XY]=E[X]E[Y]
E[XY]=E[X]E[Y]
利用上述定理,可以将上面的结论拓展到
n
n
n个随机向量的和的情况。
设
T
=
Y
1
,
⋯
,
Y
n
T=Y_1, \cdots, Y_n
T=Y1,⋯,Yn,且
Y
i
∈
R
s
Y_i\in\mathbb{R}^s
Yi∈Rs的独立变量。由于
M
T
(
u
)
=
E
[
e
u
1
Y
1
×
⋯
×
e
u
n
Y
n
]
M_T(u)=E[e^{u_1Y_1}\times\cdots\times e^{u_nY_n}]
MT(u)=E[eu1Y1×⋯×eunYn]
利用定理2,
(
12
)
=
M
Y
1
(
u
)
×
⋯
×
M
Y
n
(
u
)
(12)=M_{Y_1}(u)\times\cdots\times M_{Y_n}(u)
(12)=MY1(u)×⋯×MYn(u)
考虑累积量生成函数,取对数
K
T
(
u
)
=
K
Y
1
(
u
)
+
⋯
+
K
Y
n
(
u
)
K_T(u)=K_{Y_1}(u)+\cdots+K_{Y_n}(u)
KT(u)=KY1(u)+⋯+KYn(u)
因此,
T
T
T的累积量就等于相应的
Y
1
,
⋯
,
Y
n
Y_1,\cdots,Y_n
Y1,⋯,Yn的累积量之和。
考察典型形式指数族分布的矩母函数MGF,
E
η
e
u
T
(
X
)
=
∫
x
e
u
T
(
x
)
e
η
T
(
x
)
−
A
(
η
)
h
(
x
)
d
μ
(
x
)
=
e
A
(
u
+
η
)
−
A
(
η
)
∫
x
e
(
u
+
η
)
T
(
x
)
−
A
(
u
+
η
)
h
(
x
)
d
μ
(
x
)
\begin{aligned} E_\eta e^{uT(X)}&=\int_x e^{uT(x)}e^{\eta T(x)-A(\eta)}h(x)d\mu(x)\\ &=e^{A(u+\eta)-A(\eta)}\int_xe^{(u+\eta)T(x)-A(u+\eta)}h(x)d\mu(x)\\ \end{aligned}
EηeuT(X)=∫xeuT(x)eηT(x)−A(η)h(x)dμ(x)=eA(u+η)−A(η)∫xe(u+η)T(x)−A(u+η)h(x)dμ(x)
发现后面凑成了一个典型形式指数族分布,积分为1.因此,典型形式指数族分布的矩母函数的表达式为
e
A
(
u
+
η
)
−
A
(
η
)
e^{A(u+\eta)-A(\eta)}
eA(u+η)−A(η)
对应的累积量生成函数为
A
(
u
+
η
)
−
A
(
η
)
A(u+\eta)-A(\eta)
A(u+η)−A(η)
利用定理1,对
u
u
u求导并使之等于0:
∂
(
A
(
u
+
η
)
−
A
(
η
)
)
∂
u
∣
u
=
0
=
∂
A
(
u
+
η
)
∂
(
u
+
η
)
∣
u
=
0
=
∂
A
(
η
)
∂
(
η
)
\begin{aligned} \frac{\partial (A(u+\eta)-A(\eta))}{\partial u}|_{u=0} = \frac{\partial A(u+\eta)}{\partial (u+\eta)}|_{u=0}=\frac{\partial A(\eta)}{\partial (\eta)} \end{aligned}
∂u∂(A(u+η)−A(η))∣u=0=∂(u+η)∂A(u+η)∣u=0=∂(η)∂A(η)
依此类推,
κ
r
1
,
r
2
,
⋯
,
r
s
=
∂
r
1
∂
η
1
r
1
⋯
∂
r
s
∂
η
s
r
s
K
T
(
η
)
\kappa_{r_1,r_2, \cdots,r_s}=\frac{\partial^{r_1}}{\partial \eta_1^{r_1}}\cdots\frac{\partial^{r_s}}{\partial \eta_s^{r_s}}K_T(\eta)
κr1,r2,⋯,rs=∂η1r1∂r1⋯∂ηsrs∂rsKT(η)
对于指数族分布来说,考虑当
s
=
1
s=1
s=1的时候。由累积量生成函数和矩母函数之家牛的关系,
M
=
e
K
M=e^K
M=eK,进而
M
′
=
K
′
e
K
M
′
′
=
K
′
′
e
K
+
(
K
′
)
2
e
K
M'=K'e^K\ M''=K''e^K+(K')^2e^K
M′=K′eK M′′=K′′eK+(K′)2eK
在0处取值,则
E
[
T
]
=
k
1
,
E
[
T
2
]
=
k
2
+
k
1
2
E[T]=k_1,E[T^2]=k_2+k_1^2
E[T]=k1,E[T2]=k2+k12