Conditional on $N(t)$, given some $\lambda(t)$ characterizing some Nonhomogenous poisson point process, the distribution of an arrival time $t_i$ is $\lambda(t_i)/\int_{A}\lambda\left(t\right)dt$ where A is the space over which the point process is defined.

Can anyone explain the derivation here?

Thank you

  • $\begingroup$ what's that $\beta$? And what do you mean by distribution? $\endgroup$
    – Taylor
    Jul 13 '16 at 4:51
  • $\begingroup$ I removed beta, it's unnecessary. Re distribution: In a regular (homogenous poisson process), conditional on the number of arrivals $N(t)$ , the distribution of the arrival times themselves are uniformly distruted over $[0,t]$. In a nonhomogenous poisson process, they are distributed as above. Why? $\endgroup$ Jul 13 '16 at 15:48
  • $\begingroup$ Do you want a heuristic / intuitive or mathematical explanation? $\endgroup$
    – jbowman
    Jul 13 '16 at 18:26
  • $\begingroup$ Both would be ideal. I have vague intuition already $\endgroup$ Jul 13 '16 at 19:08
  • $\begingroup$ I don't think your statement is entirely correct. The $i$th arrival time $t_i$ does not have pdf $\lambda(t)/\int_A\lambda(u)du$ but is instead distributed as the $i$th order statistic of $n$ iid random variables, each having this density. $\endgroup$ Jun 19 '19 at 12:18

Here's the situation for the first arrival time. Let $T_1$ be the first arrival time. Sorry for making this capitalized...I like random variables to be capital letters. Pick any $t_1 < t$, where $t$ is the end time.

Also let $\Lambda(t') = \int_0^{t'}\lambda(s) ds$, $p = \Lambda(t_1) / \Lambda(t)$, and $r = s-1$.

\begin{align*} P(T_1 \le t_1 | N(t) = n) &= \sum_{s=1}^n\frac{P(N(t_1) = s, N(t) =n)}{P(N(t) = n)}\\ &= \sum_{s=1}^n\frac{P(N(t_1) = s)P( N(t)-N(t_1) =n - s)}{P(N(t) = n)}\\ &= \sum_{s=1}^n \frac{e^{-\Lambda(t_1)}\Lambda(t_1)^s e^{-[\Lambda(t)-\Lambda(t_1)] } [\Lambda(t)-\Lambda(t_1)]^{n-s} n! }{s!(n-s)! e^{-\Lambda(t)}\Lambda(t)^n } \\ &= \sum_{s=1}^n {n \choose s} \frac{\Lambda(t_1)^s [\Lambda(t)-\Lambda(t_1)]^{n-s} }{ \Lambda(t)^n }\\ &= \sum_{s=1}^n {n \choose s} p^s (1-p)^{n-s} \\ &= \frac{np}{(r+1)} \sum_{r=0}^{n-1} {n-1 \choose r} p^r (1-p)^{n-1-r} \\ &= \frac{np}{(r+1)} \end{align*}

You're looking for $f_{T_1|N(t)=n}(t_1)$, and that's $\frac{d}{dt_1}\frac{np}{(r+1))}.$ When I take the derivative I get


which is sort of like your answer. In the special case that $n=1$, then $s$ would have to be $1$, and thus $r$ would be $0$. In that case we would get the answer you mentioned in the question.

As far as intuition goes, there's higher density in places where the rate function $\lambda$ is higher. And the denominator is just a normalizing constant.


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.