I got this problem within others as homework, and I don't know how to do it. Does anyone know how to start solving it? Thank you! Any help would be appreciated. Sorry for my writing in LaTex because for some reason I can't install mathjax, and so what I see is in Latex format/display. Thanks for understanding.

Let x_1 and x_2 be independent exponentially distributed random variables with parameters $\lambda_1$ and $\lambda_2$ so that $Pr{X_i>t}=e^{-\lambda_it}$ for $t>=0$. Let $N=1$ if $X_1<X_2$, and $N=2$ if $X_2<=X_1$, $U=\min\{X_1, X_2\}=X_N$, $V=\max\{X_1,X_2\}$ and $W=V-U=|X_1-X_2|$. Show: (a). $Pr\{N=1\}=\frac{\lambda_1}{\lambda_1+\lambda_2}$ and $Pr\{N=2\}=\frac{\lambda_2}{\lambda_1+\lambda_2}$

(b). $Pr\{U>t\}=e^{-(\lambda_1+\lambda_2)t}$ for $t>=0$

(c). $N$ and $U$ are independent random variables

(d). $Pr\{W>t|N=1\}=e^{-\lambda_2t}$ and $Pr\{W>t|N=2\}=e^{-\lambda_1t}$ for $t>=0$

(e). $U$ and $W=V-U$ are independent random variables.


For a), use the law of total probability: $$ {\rm P}(X_1 < X_2 ) = \int_0^\infty {{\rm P}(X_1 < X_2 |X_2 = t)f_{X_2 } (t)\,{\rm d}t} , $$ where $f_{X_2}$ is the PDF of $X_2$.

For b), notice that $\min \{ X_1 ,X_2 \} > t$ if and only if $X_1 > t$ and $X_2 > t$ (and use the fact that $X_1$ and $X_2$ are independent).

For c), calculate ${\rm P}(\min \{X_1,X_2 \}>t , X_1 > X_2 )$ using the law of total probability, conditioning on $X_2$. You should easily find that $$ {\rm P}(\min \{X_1,X_2 \}>t , X_1 > X_2 ) = {\rm P}(\min \{X_1,X_2 \}>t ){\rm P}(X_1 > X_2 ) = \frac{{\lambda _2 }}{{\lambda _1 + \lambda _2 }}e^{ - (\lambda _1 + \lambda _2 )t}. $$

For d), note that $$ {\rm P}(|X_1 - X_2 | > t|N = 1) = \frac{{{\rm P}(X_2 - X_1 > t,X_1 < X_2 )}}{{{\rm P}(X_1 < X_2 )}} = \frac{{{\rm P}(X_2 > X_1 + t)}}{{{\rm P}(X_1 < X_2 )}}, $$ and you should easily show using the law of total probability, conditioning on $X_1$, that $$ {P(X_2 > X_1 + t)} = \frac{{\lambda _1 }}{{\lambda _1 + \lambda _2 }}e^{ - \lambda _2 t}. $$ Note: The calculation for ${\rm P}(|X_1 - X_2 | > t|N = 2)$ is completely analogous.

NOTE: Since question e) is not so easy, I give more than hints. However, try solving a significant part of it by yourself.

For e), it is straightforward to show, using that $N$ and $U$ are independent, that $$ {\rm P}(W > t | U=u) = {\rm P}(W > t | N=1, U=u){\rm P}(N=1) + {\rm P}(W > t | N=2, U=u){\rm P}(N=2). $$ For this purpose, you may replace $U=u$ by $U \in [u,u+{\rm d}u]$, where ${\rm d}u \to 0$, in order to condition on events with positive probability. Now, given $U=u$ and $N=1$, we have that $X_1 = u$ and that $X_2 - X_1$, by standard property of the exponential distribution, is exponential$(\lambda_2)$. Analogously, given $U=u$ and $N=2$, we have that $X_2 = u$ and that $X_1 - X_2$ is exponential$(\lambda_1)$. From this you should find that $$ {\rm P}(W > t | U=u) = \frac{{\lambda _1 }}{{\lambda _1 + \lambda _2 }}e^{ - \lambda _2 t} + \frac{{\lambda _2 }}{{\lambda _1 + \lambda _2 }}e^{ - \lambda _1 t} . $$ Now we are done by $$ {\rm P}(W > t ) = {\rm P}(W > t ,N = 1) + {\rm P}(W > t ,N = 2), $$ as it gives us, by virtue of a) and d), $$ {\rm P}(W > t) = \frac{{\lambda _1 }}{{\lambda _1 + \lambda _2 }}e^{ - \lambda _2 t} + \frac{{\lambda _2 }}{{\lambda _1 + \lambda _2 }}e^{ - \lambda _1 t} . $$

  • $\begingroup$ @Shai Covo: I have a quick question. If X~B(p,N) and N~P(k), what is Pr(X=x)? I think is Poisson distribution with parameter kp, but I don't know how to start it. Thanks. $\endgroup$ – kira Feb 23 '11 at 5:01
  • 1
    $\begingroup$ @kira: Start with the law of total probability: $P(X = j) = \sum\limits_{i = 0}^\infty {P(X = j|N = i)P(N = i)} = \sum\limits_{i = j}^\infty {P(X = j|N = i)P(N = i)} $. $\endgroup$ – Shai Covo Feb 23 '11 at 5:51
  • 1
    $\begingroup$ And then try showing that $P(X = j) = \frac{{e^{ - k} }}{{j!}}(\frac{p}{{1 - p}})^j \sum\limits_{i = 0}^\infty {\frac{1}{{i!}}(1 - p)^{i + j} k^{i + j} }$. This should lead you straight to the result you mentioned. $\endgroup$ – Shai Covo Feb 23 '11 at 6:06

This is to show that one can solve (e) and similar questions in a fully automatized way.

One is asked to show that $U$ and $W$ are independent, let us be more ambitious and try to compute the distribution of $(U,W)$. If this distribution is a product, we are done.

Question How to compute the distribution of any random variable $Z$?
Answer By writing $E(\varphi(Z))$ as the integral of $\varphi$ with respect to a measure $\mu$, for every (bounded measurable) function $\varphi$. Then $\mu$ is the distribution of $Z$.

And this is actually often quite easy to do...

Let us see what happens for $Z=(U,W)$. The first step is to replace $Z$ by a function of $(X_1,X_2)$, in the case at hand, $$ \varphi(Z)=\varphi(\min\{X_1,X_2\},\max\{X_1,X_2\}-\min\{X_1,X_2\}). $$ Now the RHS is a (quite ugly) function of $(X_1,X_2)$ but this does not matter. The only important thing is that the distribution of $(X_1,X_2)$ has density $f_1(x_1)f_2(x_2)$, hence, like for any function of $(X_1,X_2)$, by definition of the distribution of $(X_1,X_2)$, one has $$ E(\varphi(Z))=\int\varphi(\min\{x_1,x_2\},\max\{x_1,x_2\}-\min\{x_1,x_2\})f_1(x_1)f_2(x_2)\mathrm{d}x_1\mathrm{d}x_2. $$ Up to this point, everything is general. Now one begins to use the max/min thing. This forces us to decompose the integral into two parts, one for the domain where $x_1\le x_2$ and the other for the domain $x_1>x_2$. This decomposition yields $E(\varphi(Z))$ as $(*)+(**)$ with $$ (*)=\int\varphi(x_1,x_2-x_1)f_1(x_1)f_2(x_2)\mathbf{1}_{x_1<x_2}\mathrm{d}x_1\mathrm{d}x_2, $$ and $$ (**)=\int\varphi(x_2,x_1-x_2)f_1(x_1)f_2(x_2)\mathbf{1}_{x_1>x_2}\mathrm{d}x_1\mathrm{d}x_2. $$ Recall that our goal is to write $E(\varphi(Z))$ as $$ (o)=\int\varphi(u,w)\mathrm{d}\mu(u,w), $$ for a given measure $\mu$. Let us rewrite $(*)$ and $(**)$ with this goal in mind. The changes of variables $[u=x_1,w=x_2-x_1]$ in $(*)$ and $[u=x_2,w=x_1-x_2]$ in $(**)$ lead to $$ (*)=\int\varphi(u,w)f_1(u)f_2(u+w)\mathbf{1}_{w>0}\mathrm{d}u\mathrm{d}w, $$ and to $$ (**)=\int\varphi(u,w)f_1(u+w)f_2(u)\mathbf{1}_{w>0}\mathrm{d}u\mathrm{d}w. $$ Comparing $(o)$ with $(*)+(**)$, one sees that the only way to make them equal for every $\varphi$ is that $\mathrm{d}\mu(u,w)=g(u,w)\mathrm{d}u\mathrm{d}w$ with $$ g(u,w)=[f_1(u)f_2(u+w)+f_1(u+w)f_2(u)]\mathbf{1}_{w>0}. $$ We are done and two things are to be noted: first, all these steps are fully automatic, and second, the formula for $g$ is valid for any $(U,W)$ based on independent $X_1$ and $X_2$ with densities $f_1$ and $f_2$.

In the case at hand, $f_i(x)=\lambda_i\mathrm{e}^{-\lambda_i x}$ for $x>0$, hence, for every $u>0$ and $w>0$, $$ g(u,w)=\lambda_1\lambda_2\mathrm{e}^{-(\lambda_1+\lambda_2)u}[\mathrm{e}^{-\lambda_1w}+\mathrm{e}^{-\lambda_2w}]. $$ The function $g(u,w)$ is a product $g_1(u)g_2(w)$ hence $(U,W)$ is independent.

And naturally, this proves simultaneously that the functions $g_1$ and $g_2$ are the densities of the distributions of $U$ and $W$, up to multiplicative positive constants.


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.