Consider a Poisson process with rate $\lambda$ in a given time interval $[0,T]$. The inter-arrival time between successive arrivals is negative exponential distributed with mean $\frac{1}{\lambda}$ such that $X_1 >0$, and $\sum_{i=1}^\text{Last} X_i < T$, where $X$ represents inter-arrival time.

What about the distribution of time between Last arrival and ending time $T$? Is it also negative exponential distributed and has a mean value of $\frac{1}{\lambda}$? Can we study time segment $[0,T]$ of Poisson process in the backward direction too? In the forward direction, time between $t=0$ and first arrival is negative exponential distributed. In the backward direction, Last arrival is the first arrival and is the time between $t=T$ and Last is also negative exponential distributed. Is there any way to justify this? or some reference?

  • $\begingroup$ The conditional distribution of the time until the first arrival, given that the number of arrivals before time $T$ is exactly $k,$ has a scaled beta distribution. $\endgroup$ May 13 '18 at 11:08
  • $\begingroup$ You've altered the question, so that it now includes the csae where the number of arrivals during $[0,T]$ is $0.$ In that case, how will you define the time from the "last" arrival until $T\text{?} \qquad$ $\endgroup$ May 13 '18 at 11:39
  • $\begingroup$ when there is no arrival occurs between $[0,T]$, in that case, we assume that time from the "last" arrival until T is exactly $=T$ and it has a probability $e^{-\lambda T}$. $\endgroup$ May 13 '18 at 11:54
  • $\begingroup$ Do you likewise assume, for the same reason, that in that case the time until the first arrival is $T \text{ ?} \qquad$ $\endgroup$ May 13 '18 at 12:06
  • $\begingroup$ Unfortunately not. Except my assumption for last time as $T$ when you pointed case there is no arrival in $[0,T]$, all rest things are completely according to Poisson. I mean inter-arrival time strictly follow negative exponential distribution. $\endgroup$ May 13 '18 at 12:13

Here's an answer to the question as last modified.

First suppose that on the entire real line, not just on $[0,\infty),$ we assign to each interval $(a,b)$ a random variable $N_{(a,b)}$ for which

  • $\Pr(N_{(a,b)} = n) = \dfrac{(\lambda(b-a))^n e^{-\lambda(b-a)}}{n!}$ for $n=0,1,2,3,\ldots,$ and
  • For intervals $A,B,C,\ldots$ no two of which intersect each other, $N_A,N_B,N_C,\ldots$ are independent.

Then let $X_1$ be the time of the first arrival after time $0,$ let $X_2$ be the time from then until the next arrival after that, and so on. Then we have $$ \Pr(X_n > x) = e^{-\lambda x} \text{ for } x\ge0. $$ Now consider the time $Y$ from the last arrival before time $T>0$ until time $T.$ For $x\ge0$ we have $$ \Pr(Y>x) = \Pr(\text{no arrivals during } [T-x,T]) = \frac{(\lambda x)^0 e^{-\lambda x}}{0!} = e^{-\lambda x}. $$ I.e. it has the same distribution that any one of the inter-arrival times has.

And then if you want to truncate it at $x=T,$ as suggested in comments under the question, you can modify that accordingly.

  • $\begingroup$ These $X_n$ and $Y$ are exponentially distributed. But do these have expectation value $E[X]=\frac{1}{\lambda}$? Dont we need an additional $\lambda$ with this $e^{-\lambda x}$? $\endgroup$ May 14 '18 at 2:06
  • $\begingroup$ @Hallian1990 : The probability density function of a random variable $X$ is $x\mapsto\lambda e^{-\lambda x}$ for $x\ge0,$ (with $\lambda$ appearing as a factor of the function as a whole) if and only if $\Pr(X>x) = e^{-\lambda x}$ for $x\ge0$ (without $\lambda$ appearing as a factor of this function as a whole). $\qquad$ $\endgroup$ May 14 '18 at 2:34
  • $\begingroup$ Thanks. Moreover, I was just thinking that, is not it the case that we already assume Y to be exponentially distributed before proving it? $\endgroup$ May 14 '18 at 7:07
  • $\begingroup$ @Hallian1990 : No, I'm not assuming $Y$ is exponentially distributed; I'm proving it by recalling that the number of arrivals during a certain time has a Poisson distribution. $\endgroup$ May 14 '18 at 8:15
  • $\begingroup$ Dear @Michael. I am sorry for reaching you too much. I agree and that is correct explanation for proving it. However, I am not sure if we truncate at $x=T$, how the expressions would like like or remains same. Or I have to divide above two probabilities i.e., $P(X_n >x)$ and $P(Y >x)$ with $P(x<=T)=1-e^{-\lambda T}$ in order to normalize. Is it correct? $\endgroup$ May 14 '18 at 11:59

What appears below answers the question as originally posted, where it was stated that the number of arrivals during $[0,T]$ is equal to a given number $k.$

Suppose the number of arrivals during $[0,T]$ is exactly $k\ge 1.$ You've called the inter-arrival times $X_i,$ so that $X_1$ is the time of the first arrival, $X_2+X_2$ is the time of the second arrival, and so on. For $t\ge 0,$ let $N_t$ be the number of arrivals before time $t,$ so you're saying $N_T=k.$

What, then, is the conditional distribution of $X_1$ given that $N_T=k$?

\begin{align} & \Pr(X_1\le t\mid N_T=k) = \Pr(N_t\ge 1 \mid N_T=k) = 1-\Pr(N_t=0\mid N_T=k) \\[10pt] = {} & 1 - \frac{\Pr(N_t=0\ \&\ N_T=k)}{\Pr(N_T=k)} = 1-\frac{\Pr(N_t=0 \ \&\ N_T -N_t =k)}{(\lambda T)^k e^{-\lambda T} /k!} \\[10pt] = {} & 1-\frac{\Pr(N_t=0)\Pr(N_T-N_t = k) }{(\lambda T)^k e^{-\lambda T} /k!} = 1 - \frac{e^{-\lambda t} \cdot (\lambda(T-t))^k e^{-\lambda(T-t)}/k!}{(\lambda T)^k e^{-\lambda T} /k!} \\[10pt] = {} & 1 - \frac{(T-t)^k}{T^k}. \end{align}

(This does not depend on $\lambda.$)

In a similar way you can show that $T - \sum_{i=1}^k X_i$ has the same conditional distribution given that $N_t=k.$


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.