Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

If I have $n$ independent, identically distributed uniform $(a,b)$ random variables, why is this true: $$ \max(x_i) | \min(x_i) \sim \mathrm{Uniform}(\min(x_i),b) $$

I agree that the probability density function of $\max(x_i) | \min(x_i)$ must be non-zero only in the range $[\min(x_i), b]$.

I also find it reasonable that the posterior distribution is uniform, but I cannot see why it has to be so.

Also, I am confused how to derive this algebraically. Let $X$ be the max of $n$ iid uniform random variables. Let $Y$ be the min of $n$ iid uniform random variables. If I do $$ f_{X|Y}(x|y) = \frac{f_{X,Y}(x,y)}{f_Y(y)}, $$ $f_Y(y)$ is known. But the joint distribution is unknown ($X$ and $Y$ are not independent).

Thanks.

share|improve this question
    
The result to be proved is correct if $n=2$, but not if $n$ is any other number. –  Michael Hardy Jul 29 '11 at 22:02
add comment

2 Answers

up vote 4 down vote accepted

Your claimed result is not true, which probably explains why you're having trouble seeing it.

For simplicity I'll let $a = 0, b = 1$. Results for general $a$ and $b$ can be obtained by a linear transformation.

Let $X_1, \ldots, X_n$ be independent uniform $(0,1)$; let $Y$ be their minimum and let $X$ be their maximum. Then the probability that $X \in [x, x+\delta x]$ and $Y \in [y, y+\delta y]$, for some small $\delta x$ and $\delta y$, is

$$ n(n-1) (\delta x) (\delta y) (x-y)^{n-2} $$

since we have to choose which of $X_1, \ldots, X_n$ is the smallest and which is the largest; then we need the minimum and maximum to fall in the correct intervals; then finally we need everything else to fall in the interval of size $x-y$ in between. The joint density is therefore $f_{X,Y}(x,y) = n(n-1) (x-y)^{n-2}$.

Then the density of $Y$ can be obtained by integrating. Alternatively, $P(Y \ge y) = (1-y)^n$ and so $f_Y(y) = n(1-y)^{n-1}$.

The conditional density you seek is then $$ f_{X|Y}(x|y) = {n(n-1) (x-y)^{n-2} \over n(1-y)^{n-1}} == {(n-1) (x-y)^{n-2} \over (1-y)^{n-1}}. $$ where of course we restrict to $x > y$.

For a numerical example, let $n = 5, y = 2/3$. Then we get $f_{X|Y}(x/y) = 4 (x-2/3)^3 / (1/3)^4 = 324 (x-2/3)^3$ on $2/3 \le x \le 1$. This is larger near $1$ than near $2/3$, which makes sense -- it's hard to squeeze a lot of points in a small interval!

The result you quote holds only when $n = 2$ -- if I have two IID uniform(0,1) random variables, then conditional on a choice of the minimum, the maximum is uniform on the interval between the minimum and 1. This is because we don't have to worry about fitting points between the minimum and the maximum, because there are $n - 2 = 0$ of them.

share|improve this answer
    
In other words, it's a beta distribution. –  Michael Hardy Jul 29 '11 at 22:01
    
Hi Michael Lugo, note that your expression for $f_Y(y)$ does not seem to be the right one. $f_Y(y)$ should represent the probability density function for the minimum of n i.i.d. uniform (0,1) random variables. Your expression is the maximum of n i.i.d. uniform (0,1) random variables. $f_Y(y) = n \left(1-y \right)^{n-1}$. Then, we get $f_{X|Y}(x|y) = \frac{n(n-1)}{n} (x-y)^{n-2} \frac{1}{(1-y)^{n-1}}$ This conditional p.d.f. is equal to the p.d.f. for the minimum of ($n-1$) i.i.d. uniform(y,1) random variables, which is what we expect. –  jrand Jul 30 '11 at 2:06
    
Your explanation helped. Thanks. There is also a nice explanation in "Mathematical Statistics and Data Analysis" by Rice, Pg. 104-106, in the section "Extrema and Order Statistics". –  jrand Jul 30 '11 at 2:39
    
jrand, you're right. I kept changing my mind on whether I wanted X to be the minimum and Y to be the maximum, or vice versa. –  Michael Lugo Jul 30 '11 at 14:39
add comment

Hmm, I wonder if this question is from one of my statistics students, since I just used this in last night's lecture.

As the previous respondent pointed out, what you say is false. What is true is that, given $Y =$ min $X_i$, max $X_i$ has as its conditional distribution that of the maximum of $n - 1$ independent $U(Y,b)$ random variables.

Here is an applied probability approach: Given $Y$, we know all the $X_i \ge Y$ and $X_i = Y$ for one $i$. This leaves $n-1$ random variables that each have the conditional distribution of a $U(a,b)$ restricted to $[Y,b]$.

If you really want to do this algebraically, let $Z =$ max $X_i$, the joint density is the mixed second partial of $P${$Y \le y, Z \le z$}$ = P${$Z \le z$}$ - (1 -P${$Y > y$}$)$.

share|improve this answer
    
The last equation of your post is false. –  Did Aug 28 '11 at 20:18
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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