2
$\begingroup$

I have two poisson-processes, I have seen a mathematical proof that they are independent, and offcourse they must be independent since the proof is in several textbooks. But logically I can not understand why they are independent, for me, it is very logical that they are dependent.

The processes are defined like this:

First we have a Poisson-process $\{N(t), t \ge 0\}$ with parameter $\lambda$. And we defined two new processees like this: At every instance an event occur in the original process, we define this event as type 1, with probability p, and type 2 with probability 1-p. It can then be showed that:

$\{N_1(t),t \ge 0\}$ and $\{N_2(t), t \ge 0\}$ are both poisson processes, and it can be showed that their parameters are $\lambda p$ and $\lambda (1-p)$. But now comes what I don't get, they are also independent!

I mean, lets say you have p=0.5 and you are at point t* in time. Lets also say that you are given that $N_2(t*)=1000$, why does this value of 1000 not change the probability of $P(N_1(t*) \ge 500)$? Isn't it logical that the more type 2 events that happen, it is more likely that a lot of type 1 events have happened?

$\endgroup$
  • $\begingroup$ Keyword: thinning (of point processes). $\endgroup$ – Did Mar 9 '14 at 21:08
  • $\begingroup$ Maybe you could use simulation. Try using a low overall rate with 50-50 split. Can you show a dependence? $\endgroup$ – soakley Mar 9 '14 at 22:20
1
$\begingroup$

Consider the following two scenarios:

1) You have two independent Poisson processes $N_1(t)$ and $N_2(t)$ with rates $\lambda p$ and $\lambda (1-p)$. $N(t) = N_1(t) + N_2(t)$ is their sum. This is a Poisson process with rate $\lambda$.

2) You have a Poisson process $N(t)$ with rate $\lambda$. Each time an event occurs, a coin is tossed: with probability $p$ the event is assigned type 1, otherwise it is assigned type 1. $N_1(t)$ and $N_2(t)$ are the numbers of type 1 and type 2 events occurring up to time $t$. These are Poisson processes of rates $\lambda p$ and $\lambda (1-p)$ respectively.

Do you see that the two scenarios are describing exactly the same situation?

$\endgroup$
  • $\begingroup$ Hello, I see that in both cases we get two Poisson-processess and that they in 1) are independent. But is it obvious that they are independent in 2)? I mean if the coin gives type 1, then we know that at this moment type 2 did not happen. Why is this information not useful? $\endgroup$ – user119615 Mar 9 '14 at 21:56
  • $\begingroup$ In scenario (1), with probability $1$ the two processes never have an event at the same time. When an event of the combined process $N(t)$ does occur, it has probability $p$ of coming from process $1$ and $1-p$ of coming from process $2$. The processes in scenario 2 have exactly the same joint distribution as the processes in scenario 1. $\endgroup$ – Robert Israel Mar 10 '14 at 0:18
1
$\begingroup$

Hint: We get the independence result if we can show that $$\mathbb{E}\left[\mathrm e^{\mathrm i\left(aN_1(s)+bN_2(t)\right)} \right] = \mathbb{E}\left[\mathrm e^{\mathrm i aN_1(s)} \right]\mathbb{E}\left[\mathrm e^{\mathrm i bN_2(t)} \right], $$ for all $s\leq t$ and $a,b$. Once the case $s=t$ is done, the case $s<t$ is formal, based on conditioning and increment independence properties (of Poisson processes themselves, parent and thinnings, in this case).

Case $s=t$ (I couldn't find a link to it) is based on the typical calculation: $$\mathbb{E}\left[\mathrm e^{\mathrm i\left(aN_1(t)+bN_2(t)\right)} \right]=\sum_{k=0}^{\infty} \mathbb{E}\left[\mathrm e^{\mathrm i\left(aN_1(t)+bN_2(t)\right)} | N(t)=k\right]P\left(N(t)=k\right)$$ $$ =\sum_{k=0}^{\infty} \mathbb{E}\left[\mathrm e^{\mathrm i(a-b)\sum_{i=1}^{k}Z_i}\right]e^{\mathrm i bk}\frac{(\lambda t)^k}{k!}\mathrm e^{-\lambda t}$$ $$= \sum_{k=0}^{\infty} \frac{\left(\left(1-p+p\mathrm e^{\mathrm i(a-b)}\right)\lambda te^{\mathrm i b}\right)^k}{k!}\mathrm e^{-\lambda t}$$ $$= \mathrm e^{\lambda t\left(\left(1-p+p\mathrm e^{\mathrm i(a-b)}\right)\mathrm e^{\mathrm i b}-1\right)}=\mathrm e^{p\lambda t\left(\mathrm e^{\mathrm i a}-1\right)}\mathrm e^{(1-p)\lambda t\left(\mathrm e^{\mathrm i b}-1\right)} =\mathbb{E}\left[\mathrm e^{\mathrm i aN_1(t)} \right]\mathbb{E}\left[\mathrm e^{\mathrm i bN_2(t)} \right],$$ where $Z_i$ are the iid $p$-Bernoulli variables defining the thinning: $$ N_1(t) = \sum_{i=1}^{N(t)} Z_i,$$ $$ N_2(t) = \sum_{i=1}^{N(t)} \left(1-Z_i\right).$$

$\endgroup$
0
$\begingroup$

The probability of $N_1$ given $N_2$ is the probability of $N_1$ and $N_2$, divided by the probability of $N_2$. Since $N$ is Poisson with mean $\lambda t$, and the probability of $N_2$ given $N$ is binomial with probability $1-p$, the probability of $N_2$ is Poisson with mean $(1-p)\lambda t$. The probability of $N_1$ and $N_2$ is the probability of $N_1$ and $N=N_1+N_2$, or equivalently, the probability of $N=N_1+N_2$ multiplied by the probability of $N_1$ given $N=N_1+N_2$, which is binomial with probability $p$. With some algebra, you can find

$$ P(N_1|N_2)=\frac{P_{Poi}(N_1+N_2;\lambda t)}{P_{Poi}(N_2;(1-p)\lambda t)}P_B(N_1;N_1+N_2,p)={P_{Poi}(N_1;p\lambda t)}=P(N_1) $$

showing $N_1$ is independent of $N_2$. The converse is easily shown by swapping $N_1$ and $N_2$, and $p$ and $1-p$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.