6
$\begingroup$

The definition of a compound Poisson process and its characteristic function I have are the following:

Let $\lambda>0$ and $N\sim\text{Poisson}(\lambda T)$. Also, $\{X_i\}_{i=1}^N$ are i.i.d. and independent of $N$. And $\{U_i\}_{i=1}^N$ are i.i.d., $U_i\sim\text{Uniform}([0,T])$, and independent from $X_i,N$. Define: $$ Y_t\equiv\sum_{i=1}^N\mathbb{1}_{\{U_i\leq t\}}X_i, 0\leq t\leq T $$ Then $Y_t$ is a compound Poisson process with intensity parameter $\lambda$ and jump pdf $f(x)$.

The characteristic function of $Y_1$ is: $$ \mathbb{E}(e^{iuY_1})=e^{\lambda\int(e^{ix}-1)f(x)dx} $$

Note that the characteristic function I quoted above is for $Y_1$, not $Y_t$. I am trying to show the equality above. I currently have:

$$ \begin{align} \mathbb{E}(e^{iuY_1})&=\sum_nP(N=n)\mathbb{E}(e^{iuY_1}\mid N=n)\\ &=\sum_nP(N=n)\prod_{j=1}^n\mathbb{E}(e^{iu\mathbb{1}_{\{U_j\leq 1\}}X_j})\quad\text{(by independence)}\\ &=\sum_n P(N=n)\prod_{j=1}\int e^{iux}f(x)dx\quad\text{(by uniform)} \end{align} $$

I am not sure how to proceed. Any tips? Thanks for helping! :D

$\endgroup$
  • $\begingroup$ As a hint, it looks like you can combine the Poisson and Uniform parts of your overall process into one Poisson process. That will greatly simplify things, as you will then just have a sum of i.i.d. random variates with the number of elements of the sum distributed Poisson... $\endgroup$ – jbowman Aug 23 '16 at 19:52
  • $\begingroup$ @jbowman thank you for the tip, I will try to combine them! $\endgroup$ – Guilherme Salomé Aug 23 '16 at 19:58
  • $\begingroup$ If you don't see how to do it, I'll help out more later, here's a followup hint. Look carefully at the definition of $Y_t$ and how $U_i$ is used. The question is really very well done, the essential fact that makes it simple is not immediately obvious. $\endgroup$ – jbowman Aug 23 '16 at 20:04
  • $\begingroup$ @jbowman ok! I'm trying to solve it by myself, so the tip helps a lot. I already found something new using that. I will update it when I understand it better. Thank you very much! :D $\endgroup$ – Guilherme Salomé Aug 23 '16 at 20:08
  • $\begingroup$ @jbowman I have been trying to solve it for the past few hours, but I think I got very stuck, and I feel like I am wasting time. I found this on combining poisson and uniform. It seemed related to your hint, but I couldn't prove the theorem, even for the case $n=1$. I then proceeded to trying to find the pdf of $Y_1$, but it involves a bunch of convolutions, so I don't think that's the way you intended me to follow. $\endgroup$ – Guilherme Salomé Aug 24 '16 at 0:17
3
$\begingroup$

I was missing the knowledge of the exponential series: $$ e^x=1+x+\frac{x^2}{2!}+\frac{x^3}{3!}+\frac{x^4}{4!}+\cdots $$ I also made a mistake when I separated the uniform in the expectation. Fixing these problems: $$ \begin{align} \mathbb{E}(e^{iuY_1})&=\sum_nP(N=n)\mathbb{E}(e^{iuY_1}\mid N=n)\\ &=\sum_nP(N=n)\prod_{j=1}^n\mathbb{E}(e^{iu\mathbb{1}_{\{U_j\leq 1\}}X_j})\quad\text{(by independence)}\\ &=\sum_n P(N=n)\left(\mathbb{E}(e^{iu\mathbb{1}_{\{U_1\leq 1\}}X_1})\right)^n\quad\text{(by i.i.d.)}\\ &=\sum_{n=0}^\infty\frac{(\lambda T)^n e^{-(\lambda T)}}{n!}\left(\mathbb{E}(e^{iu\mathbb{1}_{\{U_1\leq 1\}}X_1})\right)^n\quad\text{(by Poisson)}\\ &=e^{-(\lambda T)}\cdot e^{(\lambda T)\mathbb{E}(e^{iu\mathbb{1}_{\{U_1\leq 1\}}X_1})}\quad\text{(by the exponential series)} \end{align} $$ We can calculate the expectation by conditioning on the uniform: $$ \mathbb{E}(e^{iu\mathbb{1}_{\{U_1\leq 1\}}X_1})=\frac{T-1}{T}+\frac{1}{T}\int e^{iux}f(x)dx $$ Substituting and doing some algebra we get the answer: $$ \begin{align} \mathbb{E}(e^{iuY_1})&=e^{\lambda \int (e^{iux}-1)f(x)dx} \end{align} $$

$\endgroup$
2
$\begingroup$

Here's another approach that uses a common trick with characteristic functions to avoid having to work out the sums / integrals.

I'll set $\lambda = 1$ without loss of generality, it simplifies notation and can be put back in in obvious ways in what follows. This means all the "$\lambda$"s below are not related to the $\lambda$ in the problem statement, until the very end where I include it again.

First, note that the definition of $Y_t$ involves the sum of $X_i$ corresponding to $U_i \leq t$. This can be thought of as summing "observed" $X_i$, where an $X_i$ is "observed" with a probability $p = t/T$ that is the same across all $i$.. The number of "observed" $X_i$, label it $n$, is therefore distributed Poisson$(pT)$, which is the same as Poisson$(t)$.

The proof of this is straightforward. The number of observed $X_i$, label it $n$, conditional upon $N$ is clearly distributed Binomial$(N, t/T)$. Now, let's look at the characteristic function (ch.f.) of the Binomial distribution:

$\phi_{n|N}(i\theta) = (1-p+p\text{e}^{i\theta})^N$

We will want to integrate out $N$ w.r.t. the Poisson distribution to get the ch.f. of $n$. The simple way to do this is to note that:

$\phi_{n|N}(i\theta) = \exp(N*\log(1-p+p\text{e}^{i\theta}))$

Writing out the integration (summation) gives us:

$\phi_n(i\theta) = \sum_N \exp(N*\log(1-p+p\text{e}^{i\theta})) p(N|\lambda)$

Looking at this, we can see this will have the same form as the ch.f. of a Poisson distribution ($\exp(\lambda(\text{e}^{i\theta}-1))$), just with $\log(1-p+p\text{e}^{i\theta})$ substituted in wherever $i\theta$ appears in the ch.f. Making this substitution gives us:

$\phi_n(i\theta) = \exp(\lambda \text{e}^{\log(1-p+p\text{e}^{i\theta})} - \lambda)$

which quickly reduces to:

$\phi_n(i\theta) = \exp(\lambda(1-p+p\text{e}^{i\theta}) - \lambda)$

which can be rearranged to:

$\phi_n(i\theta) = \exp(p\lambda(\text{e}^{i\theta}-1))$

which is the ch.f. of a Poisson variate with mean $p\lambda$. Substituting $t/T$ for $p$ and $T$ for $\lambda$ gives us the result.

On to step 2. Now we have the ch.f. of the number of elements in the sum $n$. Let's define $\phi_Y(i\theta)$ as the ch.f. of $Y_t$, $\phi_\Sigma(i\theta)$ as the ch.f. of the sum of $n$ $X_i$ and $\phi_X(i\theta)$ as the ch.f. of a single $X_i$. Since the elements are i.i.d., we know that, conditional upon $n$,

$\phi_\Sigma(i\theta) = \phi_X^n(i\theta) = \exp\{n \log \phi_X(i\theta)\}$

We can apply exactly the same approach as above to integrate out $n$:

$\phi_Y(i\theta) = \sum_n \exp\{n \log \phi_X(i\theta)\} p(n | t)$

where we know that $p(n|t)$ is a Poisson distribution. This will be the ch.f. of a Poisson$(t)$ distribution with $\log \phi_X(i\theta)$ substituted for $i\theta$:

$\phi_Y(i\theta) = \exp\{t(\phi_x(i\theta)-1)\}$

Adding the $\lambda$ from the original problem statement gives the answer:

$\phi_Y(i\theta) = \exp\{t\lambda(\phi_x(i\theta)-1)\}$

$\endgroup$

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.