Point Processes and Queues

1. Martingales (鞅)

1.1 Histories of probability space

$(\Omega ,\mathcal{F},P)$ probability space
$(X_t,t\ge 0)$ stochastic process
定义. 随机变量生产的$\sigma$-代数：In association with such a process $X_t$, we define for each $t\ge 0$ a sub-$\sigma$-filed of $\mathcal{F}$, denoted ${\mathcal{F}}_t^X$, by
$${\mathcal{F}}_t^X=\sigma (X_s,s\in [0,t])$$

举例：随机变量生成的$\sigma$-代数，指的是一组特殊事件组成的集合：这些事件是否发生，可以通过随机变量的取值明确判断出来。举个例子吧。
比如今天可能下雨也可能不下，下雨时随机变量 $X=1$，反之 $X=0$。然而，下不下雨只是“今天”的一个属性，其它属性，比如我早饭吃的是火腿还是培根，也是全世界所包含的信息的一部分。
假设整个概率空间由下面这些元素组成：
[火腿，下雨]：概率0.25，X = 1
[火腿，不下雨]：概率0.25，X = 0
[培根，下雨]：概率0.25，X = 1
[培根，不下雨]：概率0.25，X = 0
这时，X生成的代数包括下面三个非空集合：
{X = 1} = {[火腿，下雨]，[培根，下雨]}
{X = 0} = {[火腿，不下雨]，[培根，不下雨]}
{X = 0或1} = {[火腿，不下雨]，[培根，不下雨]，[火腿，下雨]，[培根，下雨]}
可见不管X=1还是0，我都既有可能吃火腿，也有可能吃培根；从X的取值里，你得不到任何关于我早饭吃了什么的信息。因此“早饭吃培根”这一事件，就被排除在X生成的代数之外了。具体来说：早饭吃培根 = {{培根，下雨}，{培根，不下雨}}，不属于X生成的代数。

定义. Let $Y$ be a ${\mathbb{R}}^k$- valued process such that for each $t\ge 0$, $Y_t$ is ${\mathcal{F}}_t^X$-measurable, a well-known representation result on measurability states that
$$Y=\varphi (X_s,s\in S)$$ where $S$ is a countable subset of $[0,t]$ and $\varphi$ is a borelian function（Borel函数） from ${\mathbb{R}}^{n|S|}$ into ${\mathbb{R}}^k$.
We say $Y_t$ is adapted to ${\mathcal{F}}_t^X$, if
$$Y_t={\psi }_t(X_0^t)$$ for some ${\psi }_t$, i.e., $Y_t$ depends causally on $X_t$.

If ${\mathcal{F}}_t\supseteq {\mathcal{F}}_t^X$, then ${\mathcal{F}}_t$ is called a history of $X_t$, and $X_t$ is adapted to ${\mathcal{F}}_t$.

定义. 停时(stopping time) 一类随机时刻，指具有某种与将来无关性质的随机时刻。给定概率空间$(\Sigma ，\mathcal{F}，P)$及其滤子${\mathcal{F}}_t$，映射 τ ： Σ → T ∪ { ∞ } \tau：\Sigma→T\cup\{\infty\} τ：Σ→T∪{∞}，如果对任意的 $t\in I$， { ω : ω ∈ Σ ， τ ( ω ) ≤ t } ∈ F t \{\omega:\omega\in\Sigma，τ(ω)≤t\}\in\mathcal{F}_t {ω:ω∈Σ，τ(ω)≤t}∈Ft​，则称映射$\tau$为一个${\mathcal{F}}_t$停时。

1.2 Martingale

A history is given on the probability space $(\Omega ,\mathcal{F},P)$. A $(P,{\mathcal{F}}_t)$-martingale over $[0,c]$ is a real-valued stochastic process $X_t$ such that:
(1) $X_t$ is adapted to ${\mathcal{F}}_t$,
(2) $X_t$ is $P$-integrable, i.e., $E[|X_t|]<\infty ,\forall t\in [0,c]$
(3) for all $0\le s\le t\le c$, $E[X_t|{\mathcal{F}}_s]=X_s$, $P-a.s$ (almost surely).

1.3 Levy Formula

Let $X_t$ be a right-continuous $N_{+}$-valued ${\mathcal{F}}_t$-Markov chain which is stable and conservative and admits the $Q$-matrix $(q_{ij}:i,j\in N_{+})$. If $f$ is a nonnegative function from $N_{+}\times N_{+}$ into $R_{+}$, then for any $0\le s\le t$, the Levy formulat holds:
$$E\left[\sum _{s<u\le t}f(X_{u^{-}},X_u)|\sigma (X_s)\right]=E\left[{\int }_s^t\sum _{j\ne X_u}{q}_{X_{u}j}f(X_u,j)du|\sigma (X_s)\right]$$

1.4 Radon-Nikodyn Derivatives (likeihood ratios)

Let $P$ and $Q$ be two probability measures defined on the same measurable space $(\Sigma ,\mathcal{F})$, and let ${\mathcal{F}}_t$ be a history. For each $t\ge 0$, denote by $P_t$ and $Q_t$ the restrictions of $P$ and $Q$ to ${\mathcal{F}}_t$ respectively. Suppose that for some $c\ge 0$, $Q_c$ is absolutely continuous with respect to $P_c$. Then clearly, for all $t\in [0,c]$, $Q_t$ is absolutely continuous with respect to $P_t$. Define
$$L_t:=\frac{d{Q}_t}{d{P}_t}$$ to be the Radon-Nikodym derivation of $Q_t$ with respect to $P_t$.

Proposition. $L_t$ is a $(P,{\mathcal{F}}_t)$-martingale over $[0,c]$.

1.5 Predictability

Exercise. Given a stopping time $\tau$, let
$$I_n=\{\begin{array}{ll}1,& \text{if }n\le \tau \\ 0,& \text{if }n>\tau \end{array}$$ Show that $(I_n{)}_{n\ge 1}$ is a predictable process.

2. Girsanov Theorem

定理1. If $P$ and $Q$ are equivalent measures, and $X_t$ is an ${\mathcal{F}}_t$-adpated process then the following results hold:
$$E_Q(X_t)=E_P\left(X_t\frac{dQ}{dP}\right)$$