第五章 Galton-Watson分枝过程
1.分枝过程模型
令
ξ
\xi
ξ是一个非负整型随机变量,分布是
P
(
ξ
=
k
)
=
p
k
,
k
≥
0
,
p
0
<
1
P(\xi=k)=p_k,k\ge 0,p_0<1
P(ξ=k)=pk,k≥0,p0<1。假设某物种繁衍的后代数服从
ξ
\xi
ξ的分布,且物种内每个个体的繁衍是独立的,第一代个体为
Z
1
Z_1
Z1,第二代为
Z
2
Z_2
Z2,自然地有
Z
n
+
1
=
∑
j
=
1
Z
n
ξ
n
,
j
Z_{n+1}=\sum_{j=1}^{Z_n}\xi_{n,j}
Zn+1=j=1∑Znξn,j
如此得到一列随机变量
Z
0
,
Z
1
,
⋯
Z_0,Z_1,\cdots
Z0,Z1,⋯,它们取非负整数值,
Z
0
=
1
Z_0=1
Z0=1,
Z
n
Z_n
Zn满足递推关系式,则这样的随机过程
Z
=
(
Z
n
,
n
≥
0
)
\boldsymbol Z=(Z_n,n\ge0)
Z=(Zn,n≥0)是Markov链,状态空间为
Z
+
\Z^+
Z+,转移概率为
p
i
j
=
P
(
∑
k
=
1
i
ξ
k
=
j
)
,
i
,
j
≥
0
p_{ij}=P\left(\sum_{k=1}^i \xi_k=j\right),\quad i, j\ge0
pij=P(k=1∑iξk=j),i,j≥0
定义了分枝过程,可以求其数字特征。假设
ξ
\xi
ξ的分布为
P
(
ξ
=
k
)
=
p
k
,
E
ξ
=
μ
,
D
ξ
=
σ
2
P(\xi=k)=p_k,E\xi=\mu,D\xi=\sigma^2
P(ξ=k)=pk,Eξ=μ,Dξ=σ2,则有以下结论:
-
E Z n = μ n EZ_n=\mu^n EZn=μn
由定义 Z 1 = d ξ Z_1\stackrel d= \xi Z1=dξ,所以 E Z 1 = E ξ = μ EZ_1=E\xi =\mu EZ1=Eξ=μ。
对 n ≥ 1 n\ge 1 n≥1,由全期望公式有
E Z n + 1 = E ( ∑ k = 1 Z n ξ n , j ) = ∑ N = 0 ∞ E ( ∑ k = 1 Z n ξ n , k ∣ Z n = N ) P ( Z n = N ) = ∑ N = 0 ∞ E ( ∑ k = 1 N ξ n , k ) P ( Z n = N ) = ∑ N = 0 ∞ μ N P ( Z n = N ) = μ E Z n \begin{aligned} EZ_{n+1}=&E\left(\sum_{k=1}^{Z_n} \xi_{n,j}\right)\\ =&\sum_{N=0}^\infty E\left(\sum_{k=1}^{Z_n}\xi_{n, k}\bigg|Z_n=N\right)P(Z_n=N)\\ =&\sum_{N=0}^\infty E\left( \sum_{k=1}^N \xi_{n,k} \right)P(Z_n=N)\\ =&\sum_{N=0}^\infty \mu NP(Z_n=N)\\ =&\mu EZ_n \end{aligned} EZn+1=====E(k=1∑Znξn,j)N=0∑∞E(k=1∑Znξn,k∣∣∣∣Zn=N)P(Zn=N)N=0∑∞E(k=1∑Nξn,k)P(Zn=N)N=0∑∞μNP(Zn=N)μEZn
由此递推关系式,可以得到 E Z n = μ n EZ_n=\mu^n EZn=μn。 -
D Z n = σ 2 μ n − 1 ( 1 + μ + ⋯ + μ n − 1 ) DZ_n=\sigma^2\mu^{n-1}(1+\mu+\cdots+\mu^{n-1}) DZn=σ2μn−1(1+μ+⋯+μn−1)
由于 Z 1 = d ξ Z_1\stackrel d= \xi Z1=dξ,所以 D Z 1 = D ξ = σ 2 DZ_1=D\xi=\sigma^2 DZ1=Dξ=σ2。
对 n ≥ 1 n\ge 1 n≥1,由全期望公式有
E ( Z n + 1 2 ) = E ( ∑ k = 1 Z n ξ k ) 2 = ∑ N = 0 ∞ E [ ( ∑ k = 1 Z n ξ k ) 2 ∣ Z n = N ] P ( Z n = N ) = ∑ N = 0 ∞ E [ ( ∑ k = 1 N ξ k ) 2 ] P ( Z n = N ) = ∑ N = 0 ∞ [ N E ξ 2 + N ( N − 1 ) ( E ξ ) 2 ] P ( Z n = N ) = ∑ N = 0 ∞ ( σ 2 + μ 2 ) N P ( Z n = N ) + ∑ N = 0 ∞ μ 2 N ( N − 1 ) P ( Z n = N ) = ( σ 2 + μ 2 ) E Z n + μ 2 E ( Z n 2 − Z n ) = σ 2 μ n + μ 2 E Z n 2 D ( Z n + 1 ) = E ( Z n + 1 2 ) − ( E Z n + 1 ) 2 = σ 2 μ n + μ 2 E ( Z n 2 ) − μ 2 μ 2 n = σ 2 μ n + μ 2 D Z n \begin{aligned} E(Z_{n+1}^2)=&E\left(\sum_{k=1}^{Z_n}\xi_{k}\right)^2\\ =&\sum_{N=0}^{\infty}E\left[\left(\sum_{k=1}^{Z_n}\xi_k\right)^2\bigg|Z_n=N\right]P(Z_n=N)\\ =&\sum_{N=0}^\infty E\left[\left(\sum_{k=1}^N\xi_k\right)^2\right]P(Z_n=N)\\ =&\sum_{N=0}^\infty [NE\xi^2+N(N-1)(E\xi)^2]P(Z_n=N)\\ =&\sum_{N=0}^\infty (\sigma^2+\mu^2)NP(Z_n=N)+\sum_{N=0}^\infty \mu^2N(N-1)P(Z_n=N)\\ =&(\sigma^2+\mu^2)EZ_n+\mu^2E(Z_n^2-Z_n)\\ =&\sigma^2\mu^n+\mu^2EZ_n^2\\ \quad\\ D(Z_{n+1})=&E(Z_{n+1}^2)-(EZ_{n+1})^2\\ =&\sigma^2\mu^n+\mu^2E(Z_n^2)-\mu^{2}\mu^{2n}\\ =&\sigma^2\mu^n+\mu^2DZ_n\\ \end{aligned} E(Zn+12)=======D(Zn+1)===E(k=1∑Znξk)2N=0∑∞E⎣⎡(k=1∑Znξk)2∣∣∣∣Zn=N⎦⎤P(Zn=N)N=0∑∞E⎣⎡(k=1∑Nξk)2⎦⎤P(Zn=N)N=0∑∞[NEξ2+N(N−1)(Eξ)2]P(Zn=N)N=0∑∞(σ2+μ2)NP(Zn=N)+N=0∑∞μ2N(N−1)P(Zn=N)(σ2+μ2)EZn+μ2E(Zn2−Zn)σ2μn+μ2EZn2E(Zn+12)−(EZn+1)2σ2μn+μ2E(Zn2)−μ2μ2nσ2μn+μ2DZn
归纳得到 D ( Z n ) = σ 2 μ n − 1 ( 1 + μ + ⋯ + μ n − 1 ) D(Z_n)=\sigma^2\mu^{n-1}(1+\mu+\cdots+\mu^{n-1}) D(Zn)=σ2μn−1(1+μ+⋯+μn−1)。
虽然分枝过程的数字特征可以求出,但其具体分布却不易求得。
2.生成函数
对于非负整型随机变量
ξ
\xi
ξ,概率分布列为
P
(
ξ
=
k
)
=
p
k
P(\xi=k)=p_k
P(ξ=k)=pk,则定义其生成函数为
ϕ
ξ
(
s
)
=
E
s
ξ
=
∑
k
=
0
∞
p
k
s
k
,
0
≤
s
≤
1
\phi_\xi (s)=Es^{\xi}=\sum_{k=0}^\infty p_ks^k,\quad 0\le s\le 1
ϕξ(s)=Esξ=k=0∑∞pksk,0≤s≤1
生成函数具有一些性质:
-
ϕ ( 1 ) = 1 , 0 ≤ ϕ ( s ) ≤ 1 \phi(1)=1,0\le \phi(s)\le 1 ϕ(1)=1,0≤ϕ(s)≤1;
-
ϕ ( s ) \phi(s) ϕ(s)在 [ 0 , 1 ] [0,1] [0,1]上一致连续;
-
如果 E ξ k < ∞ E\xi^k<\infty Eξk<∞,那么 ϕ ( s ) \phi(s) ϕ(s)在 [ 0 , 1 ] [0,1] [0,1]上 k k k次可微,特别当 E ξ 2 < ∞ E\xi^2<\infty Eξ2<∞时 ϕ ′ ( 1 ) = E ξ , ϕ ′ ′ ( 1 ) = E ξ 2 − E ξ \phi'(1)=E\xi,\phi''(1)=E\xi^2-E\xi ϕ′(1)=Eξ,ϕ′′(1)=Eξ2−Eξ;
-
ϕ ( s ) \phi(s) ϕ(s)在 s = 0 s=0 s=0处无穷次可微,且
p k = ϕ ( k ) ( 0 ) k ! , ∀ k ≥ 0 p_k=\frac{\phi^{(k)}(0)}{k!},\quad \forall k\ge0 pk=k!ϕ(k)(0),∀k≥0 -
如果 ξ , η \xi,\eta ξ,η独立,都是非负整型随机变量,那么 ξ + η \xi+\eta ξ+η的生成函数为 ϕ s + t ( s ) = ϕ s ( s ) ϕ η ( s ) \phi_{s+t}(s)=\phi_s(s)\phi_\eta(s) ϕs+t(s)=ϕs(s)ϕη(s)。
现在求Galton-Watson分枝过程的生成函数,记
ξ
\xi
ξ的生成函数为
ϕ
(
s
)
\phi(s)
ϕ(s),
Z
n
Z_n
Zn的生成函数为
ϕ
n
(
s
)
\phi_n(s)
ϕn(s),则由于
Z
1
=
d
ξ
Z_1\stackrel d= \xi
Z1=dξ,有
ϕ
1
(
s
)
=
ϕ
(
s
)
\phi_1(s)=\phi(s)
ϕ1(s)=ϕ(s),并且
ϕ
2
(
s
)
=
E
s
Z
2
=
E
[
E
(
s
Z
2
∣
Z
1
=
N
)
]
=
∑
N
=
0
∞
E
(
s
∑
k
=
1
N
ξ
k
∣
Z
1
=
N
)
P
(
Z
1
=
N
)
=
∑
N
=
0
∞
(
E
s
∑
k
=
1
N
ξ
k
)
P
(
Z
1
=
N
)
=
∑
N
=
0
∞
[
ϕ
(
s
)
]
N
P
(
Z
1
=
N
)
=
E
(
ϕ
(
s
)
)
ξ
=
ϕ
(
ϕ
(
s
)
)
\begin{aligned} \phi_2(s)=&Es^{Z_2}\\ =&E[E(s^{Z_2}|Z_1=N)]\\ =&\sum_{N=0}^\infty E(s^{\sum\limits_{k=1}^N \xi_k}|Z_1=N)P(Z_1=N)\\ =&\sum_{N=0}^\infty (Es^{\sum\limits_{k=1}^N\xi_k})P(Z_1=N)\\ =&\sum_{N=0}^\infty [\phi(s)]^N P(Z_1=N)\\ =&E(\phi(s))^\xi\\ =&\phi(\phi(s)) \end{aligned}
ϕ2(s)=======EsZ2E[E(sZ2∣Z1=N)]N=0∑∞E(sk=1∑Nξk∣Z1=N)P(Z1=N)N=0∑∞(Esk=1∑Nξk)P(Z1=N)N=0∑∞[ϕ(s)]NP(Z1=N)E(ϕ(s))ξϕ(ϕ(s))
以此类推,可以得到
ϕ
n
+
1
(
s
)
=
ϕ
(
ϕ
n
(
s
)
)
=
ϕ
n
(
ϕ
(
s
)
)
\phi_{n+1}(s)=\phi(\phi_n(s))=\phi_n(\phi(s))
ϕn+1(s)=ϕ(ϕn(s))=ϕn(ϕ(s)),这就是分枝过程的生成函数。得到高阶分枝过程的生成函数后,比较
s
k
s^k
sk的系数即可得到
p
k
p_k
pk的值。
3.生存与灭绝概率
灭绝概率:记
α
n
=
P
(
Z
n
=
0
)
\alpha_n=P(Z_n=0)
αn=P(Zn=0),这里
(
α
n
,
n
≥
1
)
(\alpha_n,n\ge 1)
(αn,n≥1)为单调不减非负有界数列,存在极限即灭绝概率记作
τ
\tau
τ,即
τ
=
lim
n
→
∞
α
n
,
0
≤
τ
≤
1.
\tau=\lim_{n\to \infty }\alpha_n,\quad 0\le \tau \le 1.
τ=n→∞limαn,0≤τ≤1.
当
μ
=
E
ξ
<
1
\mu=E\xi<1
μ=Eξ<1时,有
P
(
Z
n
>
0
)
=
P
(
Z
n
≥
1
)
≤
E
Z
n
=
μ
n
P(Z_n>0)=P(Z_n\ge 1)\le EZ_n=\mu^n
P(Zn>0)=P(Zn≥1)≤EZn=μn,故
τ
=
1
\tau=1
τ=1,也就是繁衍均值
<
1
<1
<1时物种以概率1灭绝。
假设
Z
n
Z_n
Zn的生成函数为
ϕ
n
(
s
)
\phi_n(s)
ϕn(s),因为
ϕ
n
(
0
)
=
α
n
\phi_n(0)=\alpha_n
ϕn(0)=αn,所以有
α
n
=
ϕ
n
(
s
)
=
ϕ
(
ϕ
n
−
1
(
s
)
)
=
α
n
−
1
\alpha_n=\phi_n(s)=\phi(\phi_{n-1}(s))=\alpha_{n-1}
αn=ϕn(s)=ϕ(ϕn−1(s))=αn−1,令
n
→
∞
n\to \infty
n→∞,有
τ
=
ϕ
(
τ
)
\tau=\phi(\tau)
τ=ϕ(τ),因此灭绝概率一定满足方程
s
=
ϕ
(
s
)
s=\phi(s)
s=ϕ(s)
如果这个方程只有一个解,那必定是
s
=
1
s=1
s=1,也就是物种以概率1灭绝。如果
μ
>
1
\mu>1
μ>1,则
τ
\tau
τ为方程的最小正解,且
0
<
τ
<
1
0<\tau<1
0<τ<1。因此有如下定理:
-
对于 p 0 > 0 p_0>0 p0>0的分枝过程,设 E ξ = μ E\xi=\mu Eξ=μ,则:
如果 μ ≤ 1 \mu\le 1 μ≤1,那么 τ = 1 \tau =1 τ=1;
如果 μ > 1 \mu>1 μ>1,那么 τ \tau τ是方程 ϕ ( s ) = s \phi(s)=s ϕ(s)=s的最小正解,且 0 < τ < 1 0<\tau<1 0<τ<1。