Suppose that $N_t$ is a Poisson process with rate $\lambda$.

  1. Find $\mathbb E\left[\frac{N^2_1}{N_3}\right]$
  2. Find $\mathbb E\left[\frac{N^2_3}{N_1}\right]$

Now, I know that if $N_t$ is a Poisson process then $N_t$ represents the number of arrivals up to time $t$. In fact, we can derive that if $N_t$ is a Poisson process then (surprise!) $N_t$ is Poisson-distributed with parameter $\lambda t$. That is, \begin{equation*} \mathbb P(N_t = k) = \frac{e^{-\lambda t}}{k!} (\lambda t)^k, \quad t \geq 0 \end{equation*}

Additionally, an important result is that the inter-arrival times of each arrival is exponentially distributed with rate $\lambda$. That is, if $T_1$ is the time until the first arrival of our Poisson process $N_t$ then \begin{equation*} \mathbb P(T_1 \leq t) = 1 - e^{\lambda t}, \quad t \geq 0 \end{equation*}

Up until now I've basically listed just a bunch of (probably useful) facts/results about Poisson processes. However, I've yet to manage to weave them into something that'll help me with a solution.

One problem I encountered is that, taking (1) as an example, $N_3 = 0$ with a nonzero probability, blowing up the expectation? Clearly this is just a heuristic argument, but we can illustrate it more clearly by attempting to solve the expectation by simply computing it:

\begin{equation*} \mathbb E\left[\frac{N^2_1}{N_3}\right] = \sum^\infty_{n = 0} \sum^\infty_{m = 0} \frac{n^2}{m} \mathbb P(N_1 = n, N_3 = m) \end{equation*}

Clearly any term with $m = 0$ will give us a singularity, and so I must have approach this poorly. It is also not obvious to me how to find the joint probability

\begin{equation*} \mathbb P(N_1 = n, N_3 = m) \end{equation*}

Thank you for any input!

Edit. I've thought about this a bit and looked through some older course material (+ the help of Henry's comment!). I forgot a nice property of Poisson processes: If $N_t$ is the number of arrivals up to time $t$ and $N_{t+s}$ the arrivals up to time $t + s$ $(t, s > 0)$ then \begin{equation*} \mathbb P(N_{t + s} - N_t = k) = e^{-\lambda t} \frac{(\lambda t)^k}{k!}, \quad k = 0, 1, 2, ... \end{equation*}

That is, $N_{t + s} - N_t \sim \text{Pois}(\lambda s)$. This property permits us to write $N_3$ in terms of $N_1$ and an independent Poisson random variable with parameter $2\lambda$: \begin{equation*} N_3 = N_1 + (N_3 - N_1) \end{equation*}

where \begin{align*} N_1 &\sim \text{Pois}(\lambda) \\ N_3 - N_1 &\sim \text{Pois}(2\lambda) \end{align*}

and the independence of $N_1$ and $N_3 - N_1$ is gained via the fact that Poisson processes have independent (non-overlapping) increments.

For brevity, denote \begin{align*} X &\stackrel{\text{def}}{=} N_1 \\ Y &\stackrel{\text{def}}{=} N_3 - N_1 \end{align*}

then $X \sim \text{Pois}(\lambda)$ and $Y \sim \text{Pois}(2\lambda)$, and $X$ and $Y$ are independent of each other. Therefore, our expectation of $N_1$ and $N_3$ becomes the expectation of independent $X$ and $Y$ \begin{align*} \mathbb E\left[\frac{N^2_1}{N_3}\right] &= \mathbb E\left[\frac{X^2}{X + Y} \right] \\ &= \sum_x \sum_y \frac{x^2}{x + y} \mathbb P(X = x, Y = y) \\ &= \sum_x \sum_y \frac{x^2}{x + y} \mathbb P(X = x)\mathbb P(Y = y) \quad \text{($X$ and $Y$ indep.)} \\ &= \sum_x \sum_y \frac{x^2}{x + y} e^{-\lambda} \frac{\lambda^x}{x!} e^{-2\lambda} \frac{(2\lambda)^y}{y!} \end{align*}

Now, I've played a bit around with these sums and I've yet to yield anything conclusive. Not to mention the issue with 0 in the denominator is still present.

  • $\begingroup$ I would agree that $0$ in the denominator with non-zero probability blows up the expectation. $\endgroup$ – Henry Feb 16 '17 at 23:52
  • $\begingroup$ For the first problem, you can fudge it by noting that if $N_3=0$ then of necessity $N_1=0$ as well, and can treat the $0^2/0$ expression as zero. Still hate that. For the second problem, even that fudge is impossible. $\endgroup$ – Mark Fischler Feb 16 '17 at 23:56
  • $\begingroup$ Given memorylessness of a Poisson process $\mathbb P(N_1 = n, N_3 = m)$ is the probability that a Poisson distributed random variable with parameter $\lambda$ takes the value $n$ multiplied by the probability that a Poisson distributed random variable with parameter $2\lambda$ takes the value $m-n$ $\endgroup$ – Henry Feb 17 '17 at 1:27

The counts of events in disjoint regions are independently distributed.

$$\mathsf P(N_1=n, N_3=m)~=~\mathsf P(N_1=n)~\mathsf P(N_3-N_1=m-n)$$

The count of arrival after time 1 until time 3 is Poisson distributed rate $(3-1)\lambda$.


So putting it all together:

$$\mathsf P(N_1=n, N_3=m)~=~\dfrac{\lambda^m 2^{m-n}e^{-3\lambda}}{n!(m-n)!} \;\mathbf 1_{(n,m)\in\Bbb N^2, 0\leq n\leq m, }$$

(NB: the support is $\{(n,m)\in\Bbb N^2: 0\leq n \leq m\}$ because there cannot be less arrivals in the whole three minutes as there were in the first one minute.)

And in your equation

$$\begin{align}\mathsf E(N_1^2/N_3) & = \sum_{n=0}^\infty\sum_{m=n}^\infty \dfrac{n^2}{m}\dfrac{\lambda^m2^{m-n}e^{-3\lambda}}{n!(m-n)!} \\ & = \sum_{n=1}^\infty\sum_{m=n}^\infty \dfrac{n^2}{m}\dfrac{\lambda^m2^{m-n}e^{-3\lambda}}{n!(m-n)!} \tag{$\star$}\end{align}$$

Which works if we allow, as mentioned in the post comments, that $\lim_{n\to 0}n^2/n = 0$

However, yes, unfortunately the same convenience does not help avoid blowout for the other expectation.

$$\begin{align}\mathsf E(N_3^2/N_1) & = \sum_{m=0}^\infty\sum_{n=0}^m \dfrac{m^2}{n}\dfrac{\lambda^m2^{m-n}e^{-3\lambda}}{n!(m-n)!} \\ & = \sum_{m=1}^\infty\sum_{n=0}^m \dfrac{m^2}{n}\dfrac{\lambda^m2^{m-n}e^{-3\lambda}}{n!(m-n)!} \tag{$\star$}\end{align}$$

  • $\begingroup$ Thank you! It seems I forgot to take care about the indices of our inner sum (starting at $m = n$). How do we deal with division by zero for $n = m = 0$? $\endgroup$ – JonMann Feb 17 '17 at 1:33
  • $\begingroup$ We don't. The second expectation is indefinite. $\endgroup$ – Graham Kemp Feb 17 '17 at 1:43
  • $\begingroup$ Thank you! All is clear after having gone through your edit. Is there an obvious way to evaluate these sums? I'm satisfied leaving it in such a form since the question is essentially answered at this point. However, wouldn't it be nice if it was indeed convergent to something elegant? $\endgroup$ – JonMann Feb 17 '17 at 1:43
  • $\begingroup$ Just to wrap up a few lingering questions: Regarding your final point, that we can't do the same trick to avoid blowout. Can we not swap the sums (via Fubini's Theorem) $\sum^\infty_{m = 0} \sum^m_{n = 0} = \sum^\infty_{n = 0} \sum^\infty_{m = n}$, and then perform the same trick as before? Also, is it just a typo that we have written $e^{-m\lambda}$ instead of $e^{-3\lambda}$? $\endgroup$ – JonMann Feb 17 '17 at 2:16
  • $\begingroup$ We still get an infinitude of non-zero numbers divided by zero. $$\sum_{m=0}^\infty m^2 \sum_{n=0}^m \frac{1}{n}\times\text{stuff} = \sum_{n=0}^\infty \frac 1n \sum_{m=n}^\infty m^2\times\text{stuff}$$ If the "stuff" does not cancel the $n$ in the denominator, the bloat cannot be eliminated (that is: the series will not converge to a finite value). $\endgroup$ – Graham Kemp Feb 17 '17 at 3:15

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.