2.2 常见连续随机变量及其分布
2.2.1 连续均匀分布
f ( x ) = { 1 / ( b − 1 ) , x ∈ [ a , b ] 0 , otherwise ⇔ X ∼ U n i f o r m ( a , b ) ⇔ X ∼ U ( a , b ) w h e r e a < b f(x)=\begin{cases} 1/(b-1), & x\in [a,b]\\0,&\text{otherwise}\end{cases} \Leftrightarrow X\sim Uniform(a,b) \Leftrightarrow X\sim U(a,b) \ where\ a<b f(x)={1/(b−1),0,x∈[a,b]otherwise⇔X∼Uniform(a,b)⇔X∼U(a,b) where a<b
2.2.2 伽马分布
f
(
x
)
=
1
β
α
Γ
(
α
)
x
α
−
1
e
x
p
(
−
x
β
)
,
w
h
e
r
e
α
,
β
>
0
⇔
X
∼
G
a
m
m
a
(
α
,
β
)
⇔
X
∼
G
a
(
α
,
β
)
f(x) = \frac{1}{\beta^\alpha}{\Gamma(\alpha)} x^{\alpha-1} exp\left( -\frac{x}{\beta}\right), where\ \alpha, \beta >0 \Leftrightarrow X\sim Gamma (\alpha,\beta) \Leftrightarrow X\sim Ga(\alpha, \beta)
f(x)=βα1Γ(α)xα−1exp(−βx),where α,β>0⇔X∼Gamma(α,β)⇔X∼Ga(α,β)
其中伽马函数为
Γ
(
α
)
=
∫
0
+
∞
x
α
−
1
e
x
p
(
−
x
)
d
x
\Gamma(\alpha)=\int_0^{+\infty} x^{\alpha-1} exp(-x) dx
Γ(α)=∫0+∞xα−1exp(−x)dx;伽马函数的性质有:
Γ
(
α
+
1
)
=
α
Γ
(
α
)
\Gamma(\alpha+1) = \alpha \Gamma(\alpha)
Γ(α+1)=αΓ(α)、
Γ
(
1
)
=
Γ
(
0
)
=
0
\Gamma(1)=\Gamma(0)=0
Γ(1)=Γ(0)=0、
Γ
(
1
/
2
)
=
π
\Gamma(1/2)=\sqrt{\pi}
Γ(1/2)=π、
Γ
(
α
+
1
)
=
α
!
\Gamma(\alpha+1)=\alpha!
Γ(α+1)=α!等。
另一种写法:
f
(
x
)
=
β
α
Γ
(
α
)
x
α
−
1
e
x
p
(
−
β
x
)
,
w
h
e
r
e
α
,
β
>
0
⇔
X
∼
G
a
m
m
a
(
α
,
β
)
⇔
X
∼
G
a
(
α
,
β
)
f(x)=\frac{\beta^\alpha}{\Gamma(\alpha)} x^{\alpha-1} exp(-\beta x),where\ \alpha, \beta >0 \Leftrightarrow X\sim Gamma(\alpha, \beta) \Leftrightarrow X\sim Ga(\alpha,\beta)
f(x)=Γ(α)βαxα−1exp(−βx),where α,β>0⇔X∼Gamma(α,β)⇔X∼Ga(α,β)
其中
α
\alpha
α主要影响分布的峰起状态,称作形状参数;
β
\beta
β主要影响分布的散度状态,称作尺度参数;如果
X
i
∼
G
a
(
α
i
,
β
i
)
X_i\sim Ga(\alpha_i, \beta_i)
Xi∼Ga(αi,βi)且相互独立,则
∑
i
=
1
n
X
i
∼
G
a
(
Σ
i
=
1
n
α
i
,
β
)
\sum_{i=1}^n X_i \sim Ga(\Sigma_{i=1}^n \alpha_i,\beta)
∑i=1nXi∼Ga(Σi=1nαi,β)
2.2.3 指数分布
f
(
x
)
=
{
1
β
e
x
p
(
−
x
β
)
,
x
>
0
0
,
otherwise
⇔
X
∼
E
x
p
(
β
)
⇔
E
(
β
)
f(x)=\begin{cases}\frac{1}{\beta} exp \left(-\frac{x}{\beta}\right),&x>0\\0,&\text{otherwise}\end{cases} \Leftrightarrow X\sim Exp(\beta)\Leftrightarrow E(\beta)
f(x)={β1exp(−βx),0,x>0otherwise⇔X∼Exp(β)⇔E(β)
常用语描述“电子元器件寿命”与“两次罕见事件的间隔时间”
2.2.4 卡方分布
令伽马分布中
α
=
p
/
2
,
β
=
2
\alpha = p/2,\ \beta=2
α=p/2, β=2其中
p
p
p为正整数,会获得卡方分布
f
(
x
)
=
1
Γ
(
p
2
)
2
p
2
x
p
2
−
1
exp
(
−
x
2
)
,
0
<
x
<
∞
⇔
X
∼
χ
2
(
p
)
f(x)=\frac{1}{\Gamma\left(\frac{p}{2}\right) 2^{\frac{p}{2}}} x^{\frac{p}{2}-1} \exp \left(-\frac{x}{2}\right), 0<x<\infty \Leftrightarrow X \sim \chi^{2}(p)
f(x)=Γ(2p)22p1x2p−1exp(−2x),0<x<∞⇔X∼χ2(p)
2.2.5 正态(高斯)分布
f
(
x
)
=
1
2
π
σ
e
x
p
[
−
(
x
−
μ
)
2
2
σ
2
]
⇔
X
∼
N
(
μ
,
σ
2
)
f(x)=\frac{1}{\sqrt{2\pi}\sigma} exp\left[ -\frac{\left( x-\mu\right)^2}{2\sigma^2} \right]\Leftrightarrow X\sim \mathcal{N}\left( \mu,\sigma^2\right)
f(x)=2πσ1exp[−2σ2(x−μ)2]⇔X∼N(μ,σ2)
如果
μ
=
0
\mu=0
μ=0且
σ
2
=
1
\sigma^2=1
σ2=1,则
f
(
x
)
=
1
2
π
e
x
p
(
−
x
2
2
)
⇒
N
(
0
,
1
)
f(x)=\frac{1}{\sqrt{2\pi}} exp\left(-\frac{x^2}{2}\right)\Rightarrow \mathcal{N}(0,1)
f(x)=2π1exp(−2x2)⇒N(0,1)
2.2.6 指数幂分布
f
(
x
)
=
1
2
q
+
1
q
Γ
(
q
+
1
q
)
σ
e
x
p
[
−
(
x
−
μ
)
2
2
σ
2
]
f(x)=\frac{1}{2^{\frac{q+1}{q}} \Gamma\left(\frac{q+1}{q}\right) \sigma} exp\left[ -\frac{\left(x-\mu\right)^2}{2\sigma^2} \right]
f(x)=2qq+1Γ(qq+1)σ1exp[−2σ2(x−μ)2]
令
q
=
2
q=2
q=2可得高斯分布
2.2.7 拉普拉斯分布
在指数幂分布中令
q
=
1
q=1
q=1可得拉普拉斯分布:
f
(
x
)
=
1
4
σ
e
x
p
[
∣
x
−
μ
∣
2
σ
]
f(x)=\frac{1}{4\sigma} exp\left[ \frac{|x-\mu|}{2\sigma}\right]
f(x)=4σ1exp[2σ∣x−μ∣]
2.2.8 广义逆高斯分布
f
(
x
)
=
(
α
β
)
p
2
2
K
p
(
α
β
)
x
p
−
1
e
x
p
[
−
1
2
(
α
x
+
β
x
)
]
,
x
≥
0
f(x)=\frac{\left(\frac{\alpha}{\beta}\right)^{\frac{p}{2}}}{2K_p\left(\sqrt{\alpha \beta}\right)}x^{p-1} exp\left[ -\frac{1}{2} \left(\alpha x + \frac{\beta}{x}\right) \right],\quad x\geq 0
f(x)=2Kp(αβ)(βα)2pxp−1exp[−21(αx+xβ)],x≥0
其中
K
p
K_p
Kp表示二阶修正的贝塞尔函数
2.2.9 逆伽马分布
f ( x ) = β α Γ ( α ) x − ( α + 1 ) ( 1 − x ) β − 1 , x ≥ 0 ⇔ X ∼ I G ( α , β ) f(x)=\frac{\beta^\alpha}{\Gamma(\alpha)} x^{-(\alpha+1)} (1-x)^{\beta-1},\quad x\geq 0\Leftrightarrow X\sim IG(\alpha,\beta) f(x)=Γ(α)βαx−(α+1)(1−x)β−1,x≥0⇔X∼IG(α,β)
2.2.10 贝塔分布
f
(
x
)
=
Γ
(
α
+
β
)
Γ
(
α
)
Γ
(
β
)
x
α
−
1
(
1
−
x
)
β
−
1
,
0
<
x
<
1
f(x)=\frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)} x^{\alpha-1} (1-x)^{\beta-1},\ 0<x<1
f(x)=Γ(α)Γ(β)Γ(α+β)xα−1(1−x)β−1, 0<x<1
其中,贝塔函数为
B
e
(
α
,
β
)
=
∫
0
1
x
α
−
1
(
1
−
x
)
β
−
1
d
x
Be(\alpha,\beta)=\int_0^1 x^{\alpha-1}(1-x)^{\beta-1} dx
Be(α,β)=∫01xα−1(1−x)β−1dx
2.2.11 T分布
f ( x ) = Γ ( α + 1 2 ) Γ ( α 2 ) ( 1 + x 2 α ) α + 1 2 ⇔ X ∼ t α f(x)=\frac{\Gamma\left( \frac{\alpha+1}{2}\right)}{\Gamma\left(\frac{\alpha}{2}\right)}\left( 1+\frac{x^2}{\alpha}\right)^{\frac{\alpha+1}{2}}\Leftrightarrow X\sim t_\alpha f(x)=Γ(2α)Γ(2α+1)(1+αx2)2α+1⇔X∼tα
2.2.12 初识共轭先验分布
例1
设
X
∼
B
e
r
n
o
u
l
l
i
(
θ
)
,
0
<
θ
<
1
,
θ
∼
B
e
t
a
(
α
,
β
)
X\sim Bernoulli(\theta),0<\theta<1,\theta\sim Beta(\alpha,\beta)
X∼Bernoulli(θ),0<θ<1,θ∼Beta(α,β),则
p
(
θ
∣
x
)
∼
p
(
x
∣
θ
)
p
(
θ
∣
α
,
β
)
=
θ
x
(
1
−
θ
)
1
−
x
Γ
(
α
+
β
)
Γ
(
α
)
Γ
(
β
)
θ
α
−
1
(
1
−
θ
)
β
−
1
=
∣
G
a
m
m
a
(
α
+
β
)
Γ
(
α
)
Γ
(
β
)
θ
α
+
x
−
1
(
1
−
θ
)
β
−
x
∼
B
e
t
a
(
α
+
x
,
b
e
t
a
−
x
+
1
)
\begin{align*} &p( \theta|x)\sim p(x|\theta) p(\theta|\alpha,\beta)\\=&\theta^x (1-\theta)^{1-x}\frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\theta^{\alpha-1} (1-\theta)^{\beta-1}\\=&\frac{|Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\theta^{\alpha+x-1}(1-\theta)^{\beta-x}\\&\sim Beta(\alpha+x, beta-x+1) \end{align*}
==p(θ∣x)∼p(x∣θ)p(θ∣α,β)θx(1−θ)1−xΓ(α)Γ(β)Γ(α+β)θα−1(1−θ)β−1Γ(α)Γ(β)∣Gamma(α+β)θα+x−1(1−θ)β−x∼Beta(α+x,beta−x+1)
显然,后延分布
p
(
θ
∣
x
)
p(\theta|x)
p(θ∣x)与先验分布
p
(
x
∣
t
h
e
t
a
)
p(x|theta)
p(x∣theta)都是贝塔分布,是共轭的。
例2
设
X
∼
N
(
0
,
σ
2
)
X\sim\mathcal{N}(0,\sigma^2)
X∼N(0,σ2),令
λ
=
σ
2
\lambda=\sigma^2
λ=σ2,则
f
X
(
x
)
=
1
2
π
λ
−
1
/
2
e
x
p
(
x
2
2
λ
)
f_X(x)=\frac{1}{\sqrt{2\pi}}\lambda^{-1/2} exp\left( \frac{x^2}{2\lambda}\right)
fX(x)=2π1λ−1/2exp(2λx2);另设
λ
∼
G
a
(
λ
∣
γ
,
α
2
)
\lambda \sim Ga\left( \lambda | \gamma,\frac{\alpha}{2}\right)
λ∼Ga(λ∣γ,2α)。
显然,
λ
∼
λ
γ
−
1
e
x
p
e
x
p
(
−
α
λ
2
)
\lambda \sim \lambda^{\gamma-1} exp exp\left( -\frac{\alpha \lambda}{2}\right)
λ∼λγ−1expexp(−2αλ),则
p
(
λ
∣
x
)
∝
p
(
x
∣
λ
)
p
(
λ
∣
γ
,
α
)
∝
λ
γ
−
3
2
e
x
p
[
−
1
2
(
x
2
λ
+
α
λ
)
]
p(\lambda|x)\propto p(x|\lambda)p(\lambda|\gamma,\alpha)\propto \lambda^{\gamma-\frac{3}{2}} exp\left[-\frac{1}{2}\left(\frac{x^2}{\lambda} + \alpha \lambda\right)\right]
p(λ∣x)∝p(x∣λ)p(λ∣γ,α)∝λγ−23exp[−21(λx2+αλ)]属于广义逆高斯分布,未与先验分布(伽马分布)形成共轭,于是考虑更换先验分布条件为
λ
∼
I
G
(
τ
,
β
2
)
\lambda \sim IG\left(\tau,\frac{\beta}{2}\right)
λ∼IG(τ,2β),则有
p
(
λ
∣
x
)
∝
p
(
x
∣
λ
)
p
(
λ
∣
τ
,
β
2
)
∝
λ
−
(
τ
+
2
2
)
e
x
p
(
−
x
2
+
β
2
λ
)
∼
I
G
(
τ
+
1
2
,
x
2
+
β
)
p(\lambda|x)\propto p(x|\lambda) p\left( \lambda|\tau,\frac{\beta}{2}\right) \propto \lambda^{-\left(\tau+\frac{2}{2}\right)} exp\left(-\frac{x^2+\beta}{2\lambda}\right)\sim IG\left( \tau+\frac{1}{2}, x^2+\beta\right)
p(λ∣x)∝p(x∣λ)p(λ∣τ,2β)∝λ−(τ+22)exp(−2λx2+β)∼IG(τ+21,x2+β)与先验分布形成共轭。