前文链接
Compensated Poisson Process
补偿泊松过程
M
(
t
)
M(t)
M(t)表达式如下
M
(
t
)
=
N
(
t
)
−
λ
t
M(t)=N(t)-\lambda t
M(t)=N(t)−λt
M
(
t
)
M(t)
M(t) is a Martingale.
Poisson Process的性质为
E
[
N
(
t
)
∣
F
(
0
)
]
=
E
[
N
(
t
)
]
=
λ
t
≠
N
(
0
)
\mathbb{E}[N(t)\mid \mathcal{F}(0)]=\mathbb{E}[N(t)]=\lambda t\neq N(0)
E[N(t)∣F(0)]=E[N(t)]=λt=N(0)
为了构造一个鞅过程,需要加入补偿项
λ
t
\lambda t
λt.
证明
M
(
t
)
M(t)
M(t)是鞅过程.
令
s
≤
t
s\leq t
s≤t
E
[
M
(
t
)
∣
F
(
s
)
]
=
E
[
N
(
t
)
−
λ
t
∣
F
(
s
)
]
=
E
[
N
(
t
)
∣
F
(
s
)
]
−
λ
t
=
E
[
N
(
t
)
−
N
(
s
)
+
N
(
s
)
∣
F
(
s
)
]
−
λ
t
=
E
[
N
(
t
)
−
N
(
s
)
∣
F
(
s
)
]
+
E
[
N
(
s
)
∣
F
(
s
)
]
−
λ
t
=
λ
(
t
−
s
)
+
N
(
s
)
−
λ
t
=
N
(
s
)
−
λ
s
=
M
(
s
)
\begin{aligned} \mathbb{E}[M(t)\mid \mathcal{F}(s)]&=\mathbb{E}[N(t)-\lambda t\mid \mathcal{F}(s)]=\mathbb{E}[N(t)\mid \mathcal{F}(s)]-\lambda t \\ &=\mathbb{E}[N(t)-N(s)+N(s)\mid \mathcal{F}(s)]-\lambda t\\ &=\mathbb{E}[N(t)-N(s)\mid \mathcal{F}(s)]+\mathbb{E}[N(s)\mid \mathcal{F}(s)]-\lambda t\\ &=\lambda(t-s)+N(s)-\lambda t\\ &=N(s)-\lambda s\\ &=M(s) \end{aligned}
E[M(t)∣F(s)]=E[N(t)−λt∣F(s)]=E[N(t)∣F(s)]−λt=E[N(t)−N(s)+N(s)∣F(s)]−λt=E[N(t)−N(s)∣F(s)]+E[N(s)∣F(s)]−λt=λ(t−s)+N(s)−λt=N(s)−λs=M(s)
即
M
(
t
)
M(t)
M(t)是一个鞅过程.
Compound Poisson Process
令
Q
(
t
)
=
∑
i
=
1
N
(
t
)
Y
i
Q(t)=\sum\limits_{i=1}^{N(t)} Y_i
Q(t)=i=1∑N(t)Yi,
N
(
t
)
N(t)
N(t)是一个强度参数为
λ
\lambda
λ的Poisson Process,令
Y
1
,
Y
2
,
…
Y_1, Y_2, \dots
Y1,Y2,… 为一个i.i.d的序列,均值
E
Y
i
=
β
\mathbb{E}Y_i=\beta
EYi=β. 并且
{
Y
i
}
i
=
1
∞
\{Y_i\}_{i=1}^{\infty}
{Yi}i=1∞和
N
(
t
)
N(t)
N(t)相独立.
计算数学期望
∵
E
[
Q
(
t
)
∣
N
(
t
)
=
k
]
=
k
β
\because\mathbb{E}[Q(t)\mid N(t)=k]=k\beta
∵E[Q(t)∣N(t)=k]=kβ
由期望迭代性质
E
[
E
[
Q
(
t
)
∣
N
(
t
)
]
]
=
E
[
Q
(
t
)
]
=
∑
k
=
0
∞
k
β
P
(
N
(
t
)
=
k
)
=
∑
k
=
1
∞
k
β
(
λ
t
)
k
k
!
e
−
λ
t
=
∑
k
=
1
∞
β
(
λ
t
)
k
(
k
−
1
)
!
e
−
λ
t
=
β
λ
t
∑
k
=
1
∞
(
λ
t
)
l
l
!
e
−
λ
t
⏟
p
d
f
:
P
o
i
s
s
o
n
(
λ
t
)
=
β
λ
t
\begin{aligned} \mathbb{E}[\mathbb{E}[Q(t)\mid N(t)]]&=\mathbb{E}[Q(t)]=\sum_{k=0}^\infty k\beta\mathbb{P}(N(t)=k)\\ &=\sum_{k=1}^\infty k\beta\frac{(\lambda t)^k}{k!}e^{-\lambda t}\\ &=\sum_{k=1}^\infty \beta\frac{(\lambda t)^k}{(k-1)!}e^{-\lambda t}\\ &=\beta\lambda t\underbrace{\sum_{k=1}^\infty\frac{(\lambda t)^l}{l!}e^{-\lambda t}}_{pdf:Poisson(\lambda t)}\\ &=\beta\lambda t \end{aligned}
E[E[Q(t)∣N(t)]]=E[Q(t)]=k=0∑∞kβP(N(t)=k)=k=1∑∞kβk!(λt)ke−λt=k=1∑∞β(k−1)!(λt)ke−λt=βλtpdf:Poisson(λt)
k=1∑∞l!(λt)le−λt=βλt
直观意义上,
N
(
t
)
N(t)
N(t)表示跳跃的次数,期望跳跃次数为
λ
t
\lambda t
λt,
Y
i
Y_i
Yi表示每次跳跃的距离,期望为
β
\beta
β单位长度,总距离期望为
λ
t
β
\lambda t\beta
λtβ.
性质:设计补偿过程
Q
(
t
)
−
β
λ
t
Q(t)-\beta\lambda t
Q(t)−βλt是一个martingale.
证明:
E
[
Q
(
t
)
−
β
λ
t
∣
F
(
s
)
]
=
E
[
Q
(
t
)
−
Q
(
s
)
∣
F
(
s
)
]
+
E
[
Q
(
s
)
∣
F
(
s
)
]
−
β
λ
t
\mathbb{E}[Q(t)-\beta\lambda t\mid \mathcal{F}(s)]=\mathbb{E}[Q(t)-Q(s)\mid\mathcal{F}(s)]+\mathbb{E}[Q(s)\mid \mathcal{F}(s)]-\beta\lambda t
E[Q(t)−βλt∣F(s)]=E[Q(t)−Q(s)∣F(s)]+E[Q(s)∣F(s)]−βλt
根据定义
∵
E
[
Q
(
t
)
−
Q
(
s
)
∣
F
(
s
)
]
=
∑
i
=
1
N
(
t
)
Y
i
−
∑
i
=
1
N
(
s
)
Y
i
=
∑
N
(
s
)
+
1
N
(
t
)
Y
i
∣
F
(
s
)
∴
E
[
Q
(
t
)
−
β
λ
t
∣
F
(
s
)
]
=
E
[
Q
(
t
)
−
Q
(
s
)
]
+
Q
(
s
)
−
β
λ
t
=
β
λ
t
−
β
λ
s
+
Q
(
s
)
−
β
λ
t
=
Q
(
s
)
−
β
λ
s
\begin{aligned} &\because\mathbb{E}[Q(t)-Q(s)\mid\mathcal{F}(s)]=\sum_{i=1}^{N(t)}Y_i-\sum_{i=1}^{N(s)}Y_i=\sum_{N(s)+1}^{N(t)}Y_i \Big | \mathcal{F}(s)\\ &\therefore \mathbb{E}[Q(t)-\beta\lambda t\mid\mathcal{F}(s)]=\mathbb{E}[Q(t)-Q(s)]+Q(s)-\beta\lambda t\\ &=\beta\lambda t-\beta\lambda s+Q(s)-\beta\lambda t \\ &=Q(s)-\beta\lambda s \end{aligned}
∵E[Q(t)−Q(s)∣F(s)]=i=1∑N(t)Yi−i=1∑N(s)Yi=N(s)+1∑N(t)Yi∣∣∣F(s)∴E[Q(t)−βλt∣F(s)]=E[Q(t)−Q(s)]+Q(s)−βλt=βλt−βλs+Q(s)−βλt=Q(s)−βλs
即
Q
(
t
)
−
β
λ
t
Q(t)-\beta\lambda t
Q(t)−βλt是鞅过程.
定理 1:令
Q
(
t
)
Q(t)
Q(t)为一个compound Poisson Process 并且
0
=
t
0
<
t
1
<
…
,
t
n
0=t_0<t_1<\dots, t_n
0=t0<t1<…,tn为给定时间节点,那么增量
Q
(
t
1
)
−
Q
(
t
0
)
,
Q
(
t
2
)
−
Q
(
t
1
)
,
…
,
Q
(
t
n
)
−
Q
(
t
n
−
1
)
Q(t_1)-Q(t_0), Q(t_2)-Q(t_1), \dots, Q(t_n)-Q(t_{n-1})
Q(t1)−Q(t0),Q(t2)−Q(t1),…,Q(tn)−Q(tn−1)
为平稳且独立的,
Q
(
t
j
)
−
Q
(
t
j
−
1
)
Q(t_j)-Q(t_{j-1})
Q(tj)−Q(tj−1)具有和
Q
(
t
j
−
t
j
−
1
)
Q(t_j-t_{j-1})
Q(tj−tj−1)相同的分布.
定理 2(分解定理):令
y
1
,
y
2
,
…
,
y
m
y_1, y_2, \dots, y_m
y1,y2,…,ym为非零有限集合,并且
P
(
y
1
)
,
P
(
y
2
)
,
…
,
P
(
y
m
)
P(y_1), P(y_2), \dots, P(y_m)
P(y1),P(y2),…,P(ym)是和为1的正数.
令
Y
1
,
Y
2
,
…
,
Y
n
Y_1, Y_2, \dots, Y_n
Y1,Y2,…,Yn为i.i.d的随机变量序列,并且
P
{
Y
i
=
y
m
}
=
P
(
y
m
)
,
m
=
1
,
2
,
…
,
M
\mathbb{P}\{Y_i=y_m\}=P(y_m), m=1, 2, \dots, M
P{Yi=ym}=P(ym),m=1,2,…,M令
N
(
t
)
N(t)
N(t)为一个Poisson Process并且定义Compound Poisson Process
Q
(
t
)
=
∑
i
=
1
N
(
t
)
Y
i
Q(t)=\sum_{i=1}^{N(t)} Y_i
Q(t)=i=1∑N(t)Yi
对于
m
=
1
,
2
,
…
,
M
m=1, 2, \dots, M
m=1,2,…,M,令
N
m
(
t
)
N_m(t)
Nm(t)表示
Q
Q
Q过程中跳跃的距离时
y
m
y_m
ym的次数,那么
N
(
t
)
=
∑
m
=
1
M
N
m
(
t
)
Q
(
t
)
=
∑
m
=
1
M
y
m
N
m
(
t
)
N(t)=\sum_{m=1}^MN_m(t) \\ Q(t)=\sum_{m=1}^My_mN_m(t)
N(t)=m=1∑MNm(t)Q(t)=m=1∑MymNm(t)
过程
N
1
,
N
2
,
…
,
N
m
N_1, N_2, \dots, N_m
N1,N2,…,Nm是独立的Poisson Process,每个过程的强度参数为
λ
p
(
y
m
)
\lambda p(y_m)
λp(ym).
解析:该定理从另一个角度观察了Compound Poisson Process,分拆了强度参数,可以计算出
∑
i
=
1
M
λ
p
(
y
m
)
=
λ
∑
i
=
1
m
p
(
y
m
)
=
λ
\sum_{i=1}^M\lambda p(y_m)=\lambda\sum_{i=1}^mp(y_m)=\lambda
i=1∑Mλp(ym)=λi=1∑mp(ym)=λ