I have a random variable $t$ that follows an exponential distribution with rate $\lambda_1$ in the interval $[0,t_1]$ and an exponential distribution with rate $\lambda_2$ in the interval $[t_1,\infty]$. I want to estimate the probability that there is no event in the time interval $[0,t_2]$, where $t_2>t_1$.

Here is how I have approached this problem.

$\begin{equation} P(t>t_1+t_2)\\ =\{1-\int_{-\infty}^{t_1}f_{\lambda_1}(t)\} \{1-\int_{t_1}^{t_2}f_{\lambda_2}(t)\} \end{equation} $

Is the correct way to approach this problem? Should the second integral be $\{1-\int_{0}^{t_2}f_{\lambda_1}(t)\}$?

In general, what is the procedure to find the waiting time to the 1st event in a case where the inter-arrival process is non-homogenous?

  • 1
    $\begingroup$ Suppose $\lambda_1$ happened to equal $\lambda_2$. You would likely use a simpler expression for the waiting time involving just one integral. Is your formula equivalent in that case to the simple expression? $\endgroup$
    – whuber
    Commented Aug 17, 2016 at 18:53
  • $\begingroup$ That is a good point. It is not. $\endgroup$ Commented Aug 17, 2016 at 18:58
  • $\begingroup$ OK. The simple expression can be written as an integral. There is a natural way to break it into an integral over two regions separated at the value $t_1$, which re-expresses it as a sum of two integrals. Consider why that is so and apply that insight to the case $\lambda_1 \ne \lambda_2$. $\endgroup$
    – whuber
    Commented Aug 17, 2016 at 19:02
  • 3
    $\begingroup$ Unless $\lambda_1 = \lambda_2$, the density of the "random variable" does not integrate to 1. If $\lambda_1 \ge \lambda_2$, it would integrate out to 1 if the $\lambda_1$ rate section ended at $t_1$, then there was a gap of 0 density, and the $\lambda_2$ rate section started at $t_3 = (\lambda_1/\lambda_2) t_1$. Barring that, a time-inhomogeneos Possion Process, yes, a proper random variable, no. $\endgroup$ Commented Aug 17, 2016 at 19:49
  • 1
    $\begingroup$ I think Math1000 has given a nice demonstration. It differs from your approach by recognizing you have to compute the probability by multiplying a marginal by a conditional probability rather than adding those two probabilities. $\endgroup$
    – whuber
    Commented Aug 18, 2016 at 19:44

1 Answer 1


Let $\{N(t):t\in\mathbb R_+\}$ be a (non-homogeneous) Poisson process with intensity $$\lambda(t) = \lambda_1\mathsf 1_{[0,t_1)}(t) + \lambda_2\mathsf 1_{[t_1,\infty)}(t).$$ Then for $t_2>t_1$, we have \begin{align} \mathbb P(N(t_2)=0) &= \mathbb P(N(t_1)=0,N(t_2)-N(t_1)=0)\\ &= \mathbb P(N(t_1)=0)\mathbb P(N(t_2)-N(t_1)=0)\\ &= e^{-\lambda_1 t_1}e^{-\lambda_2(t_2-t_1)}. \end{align}


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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