5
$\begingroup$

There are two Poisson processes, process A with rate $\lambda$ and process B with rate $\mu$. These to processes (A and B) can be merged and yield again a Poisson process C with rate $\lambda + \mu$. Let D be the event that the first arrival of the merged process C is from process A. Then, the probability of event D is $$P(D)=\frac{\lambda}{\lambda + \mu}$$

Let $T$ be the time until the first arrival in the merged process C. Then the expectation of $T$ is: $$E[T]=\frac{1}{\lambda + \mu}$$

Now my question is: What is the conditional expectation of the time until the first arrival in the merged process C conditioned on the event D (that the first arrival is from the first process, i.e. what is $$E[T\mid D]?$$

Is it $E[T\mid D]=\frac{1}{\lambda}$ or $E[T\mid D]=\frac{1}{\lambda+\mu}$, and especially why?

I tried to find the PDF $f_{T\mid D}$ but I did get nowhere since it comes down to finding the probability $$P(T \geq t \,\cap D)$$ and I could not do that.

The question comes from trying problem 3 of http://ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-041sc-probabilistic-systems-analysis-and-applied-probability-fall-2013/unit-iii/lecture-15/MIT6_041SCF13_assn07.pdf

$\endgroup$
  • 1
    $\begingroup$ $f_{T|D}(t)\propto \lambda e^{-\lambda t}e^{-\mu t}$, hence $T|D\sim \exp(\lambda+\mu)$ and $E(T|D)=\frac 1{\lambda+\mu}$. $\endgroup$ – A.S. Feb 23 '16 at 18:27
  • 2
    $\begingroup$ You can also derive this simply by noting that $T|D$ and $T|D^c$ are both memoryless (and hence exponentials) and when mixed together they add up to $\exp(\lambda+\mu)$. This is only possible when they are both $\exp(\lambda+\mu)$ (otherwise tails won't match). $\endgroup$ – A.S. Feb 23 '16 at 18:40
1
$\begingroup$

In the Poisson process one can let $T_x$ be the time until the $x$th arrival and $X_t$ be the number of arrivals before time $t$, and then the event $$[T_x>t] \tag T$$ is the same as the event $$[X_t < x]. \tag X$$ So you're looking for $$\operatorname{E}(\min\{T_1^A, T_1^B\} \mid T_1^A<T_1^B) = \operatorname{E}(T_1^A\mid T_1^A<T_1^B). \tag E$$

We have a joint distribution

$$ f_{T_1^A,T_1^B} (s,t) \, ds\,dt = e^{-\lambda s} e^{-\mu t} (\lambda\,ds) (\mu\,dt) \text{ for } s>0,\ t>0 $$ and a conditional distribution given $T_1^A<T_1^B$: $$ f_{T_1^A,T_1^B\,\mid\, T_1^A<T_1^B} (s,t) \, ds\,dt = \frac{e^{-\lambda s} e^{-\mu t} (\lambda\,ds) (\mu\,dt)}{\lambda/(\lambda+\mu)} \text{ for } 0<s<t. $$ So the expected value sought in $(\text{E})$ is \begin{align} & \iint\limits_{\{(s,t)\,:\,0<s<t\}} s \, \frac{e^{-\lambda s} e^{-\mu t} (\lambda\,ds) (\mu\,dt)}{\lambda/(\lambda+\mu)} \\[10pt] = {} & (\lambda + \mu) \int_0^\infty \left( \int_0^t s e^{-\lambda s} e^{-\mu t} \,ds \right) \, (\mu\, dt) \\[10pt] = {} & (\lambda + \mu) \int_0^\infty \left( e^{-\mu t} \left( \int_0^s se^{-\lambda s} \,ds \right) \right) \, (\mu\,dt) \\[10pt] = {} & \frac{\lambda+\mu}{\lambda^2} \int_0^\infty \left( e^{-\mu t} \left( \int_0^t (\lambda s)e^{-\lambda s} (\lambda\,ds) \right) \right) \, (\mu\,dt) \\[10pt] & \text{I am omitting the rest of this because the} \\ & \text{software is refusing to let me see the preview as I work.} \end{align} The inside integral is $$ \int_0^t (\lambda s)e^{-\lambda s} (\lambda\,ds) = \int_0^{\lambda t} u e^{-u} \,du $$

Now recall that $$ \int_0^\infty s e^{-\alpha s}\,ds = \frac 1 {\alpha^2} \int_0^\infty (\alpha s) e^{-\alpha s}\,(\alpha\,ds) = \frac 1 {\alpha^2} \int_0^\infty u e^{-u}\,du = \frac 1 {\alpha^2}. $$

$[\cdots]$

This may bear simplification and may even bear correction of details near the end$\,\ldots\,$I'll be back$\,\ldots$

$\endgroup$
  • 1
    $\begingroup$ I am slowing editing details while the software refuses to allow me to see the preview. $\endgroup$ – Michael Hardy Feb 22 '16 at 19:57
  • $\begingroup$ Thanks for your help! Well, it took me some time to evaluate the integral. After simplification I got the result: $(E)=\frac{1}{\lambda+\mu}$. So, it is the same as the unconditioned expectation of the first arrival in the merged process. In the solutions - ocw.mit.edu/courses/electrical-engineering-and-computer-science/… - this result was simply stated. Is there a way to get this result without evaluating the integral, which is quite tedious? I have no idea.. $\endgroup$ – user50224 Feb 23 '16 at 12:32
  • 1
    $\begingroup$ I had some details wrong and have edited them; I'm looking at this again$\,\ldots\qquad$ $\endgroup$ – Michael Hardy Feb 23 '16 at 18:24
  • 1
    $\begingroup$ I had $s$ as an upper bound of integration at one place where I clearly should have had $t$. Here's a bit more detail; maybe I'll add more later: $$\int_0^t (\lambda s)e^{-\lambda s} (\lambda\,ds) = \int_0^{\lambda t} u e^{-u} \, du = \int u (e^{-u}\,du) = \overbrace{\int u\,dv = uv - \int v\,du}^\text{integration by parts}$$ $$= \left[ -ue^{-u} \vphantom{\frac 1 1} \right]_0^{\lambda t} - \int_0^{\lambda t} -e^{-u} \, du = 1 - e^{-\lambda t} - \lambda t e^{-\lambda t}$$ $\endgroup$ – Michael Hardy Feb 23 '16 at 18:32
  • $\begingroup$ "This may bear simplification" Indeed! But, to tell the truth, the whole post is slightly surprising... If I understand correctly, you still did not evaluate integral (E)? So we have a case of accepted answer that does not answer the question. $\endgroup$ – Did Nov 1 '16 at 18:07
1
$\begingroup$

This is to elaborate on comments by A. S, and explain how to compute $$P(T \geq t \,\cap D) $$ for $t>0$. Consequently, we will have $T|D$ is memoryless. So is $T|D^c$, by the same argument.

We may rewrite the problem as follows: Let $X$, $Y$ be independent with $$X \sim \mathrm{Exp}(\lambda), \ Y\sim\mathrm{Exp}(\mu).$$ Let $D$ be the event that $X<Y$, and $T=\mathrm{min}(X,Y)$. We will show that $$P(T\geq t \cap D)=P(X\geq t)P(Y\geq t)P(D).$$ Consider the conditional probability $$P(X<Y | X\geq t, Y\geq t)=P(T\geq t \cap D)/P(X\geq t, Y\geq t)$$ By the memoryless property, the distribution of $X-t$ and $Y-t$ given that $X\geq t$, $Y\geq t$ are identical to that of $X$ and $Y$. Also note that $X<Y$ if and only if $X-t < Y-t$. This shows that the conditional probability is indeed $P(D)$. i.e. $$P(X<Y | X\geq t, Y\geq t)=P(D).$$ By independence of $X$ and $Y$, we have $P(X\geq t, Y\geq t) = P(X\geq t)P(Y\geq t)$. Combining everything together, we obtain that $$P(T\geq t \cap D)=P(X\geq t)P(Y\geq t)P(D)= e^{-(\lambda+\mu) t} \frac{\lambda}{\lambda+\mu}.$$

$\endgroup$

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.