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.

Consider a random variable $Z$ defined as the sum of two independent copies of the uniform random variable on $[0,1]$. Let $Y$ be the minimum of $n$ independent copies of $Z$. What is the expectation and standard deviation of $Y$?

In case an exact answer will be hard to come by, I would be really happy with estimates which are correct up to a constant.

share|improve this question
2  
en.wikipedia.org/wiki/… –  Rahul Jun 23 '11 at 0:01
    
@Mike: Point taken, but I feel an answer should contain some explanation and elaboration (e.g. specialization to the 1st order statistic, the expectation and variance of the corresponding beta distribution, and so on) which I haven't the energy for at the moment. If someone else wants to take this, flesh it out, and post it as a good answer, please go ahead. –  Rahul Jun 23 '11 at 0:08
    
To clarify: while its not hard to derive an expression for the distribution of $Y$, I'm having trouble working out the integrals for its mean and variance. –  robinson Jun 23 '11 at 0:48

2 Answers 2

up vote 2 down vote accepted

This problem is more difficult than I had thought at first. :)

I'm assuming you know that the pdf of $Z$ has the triangular shape defined by $f(z) = z$, if $0 \leq z \leq 1$, and $f(z) = 2-z$ if $1 \leq z \leq 2$. Thus $P(Z > z) = 1 - z^2/2$, if $0 \leq z \leq 1$, and $P(Z > z) = 2 - 2z + z^2/2$, if $1 \leq z \leq 2$.

Then, as you note, $E[Y] = \int_0^2 P(Z > z)^n dz = \int_0^1 (1 - z^2/2)^n dz + \int_1^2 (2 - 2z + z^2/2)^n dz$. The second integrand factors into $2^{-n} (2-z)^{2n}$, which means the second integral evaluates to $2^{-n}/(1+2n)$.

To evaluate the first integral, it helps to consider $\int_0^\sqrt{2} (1 - z^2/2)^n dz$. Since $1-z^2/2$ is decreasing, the error introduced by extending the region of integration like this is smaller than $2^{-n}(\sqrt{2} - 1) \leq 2^{-(n+1)}$. Then, use the transformation $t=z^2/2$ suggested by Emre. This yields $$\int_0^\sqrt{2} (1 - z^2/2)^n dz = \frac{1}{\sqrt{2}}\int_0^1 t^{-1/2} (1 - t)^n dt = \frac{1}{\sqrt{2}} B(1/2,n+1) = \frac{1}{\sqrt{2}} \frac{\Gamma(1/2) \Gamma(n+1)}{\Gamma(n+3/2)}$$ $$= \frac{\Gamma(1/2) \Gamma(n+1)}{\sqrt{2}(n+1/2)\Gamma(n+1/2)},$$ where $B(x,y)$ is the Beta function, as noted by Emre.

Now, $\Gamma(1/2) = \sqrt{\pi}$, and (see, for example, here) $$\frac{\Gamma(n+1)}{\Gamma(n+1/2)} = \frac{{4^n }}{{{2n \choose n}}\sqrt \pi},$$ which gives $$E[Y] = \frac{{4^n }}{\sqrt{2}(n+1/2){{2n \choose n}}} + O(2^{-n}).$$

You can get a simpler expression while losing some precision by using Gautschi's inequality. This gives us $$\sqrt{n} < \frac{\Gamma(n+1)}{\Gamma(n+1/2)} < \sqrt{n+1} \Rightarrow \frac{\Gamma(n+1)}{\Gamma(n+1/2)} = \sqrt{n} + O\left(\frac{1}{\sqrt{n}}\right).$$ This implies $$E[Y] = \frac{\sqrt{\pi n}}{\sqrt{2} (n+1/2)} + O\left(\frac{1}{n^{3/2}}\right) = \frac{\sqrt{\pi}}{\sqrt{2n}} + O\left(\frac{1}{n^{3/2}}\right).$$


For the standard deviation of $Y$, use $SD[Y] = \sqrt{Var[Y]}$, $Var[Y] = E[Y^2] - E[Y]^2$, and $E[Y^2] = 2\int_0^2 z (P(Z > z))^n dz$. (For a proof of the last, see here.) The calculation of the integral for $E[Y^2]$ is much easier than that for $E[Y]$ and comes to $$E[Y^2] = \frac{2 + 2^{-n+1} + 4 n}{1 + 3 n + 2 n^2} = \frac{2}{n+1} + O\left(2^{-n}\right).$$ For example, if we use this with the simpler approximation for $E[Y]$, we obtain $$SD[Y] = \frac{\sqrt{2 - \pi/2}}{\sqrt{n}} + O\left(\frac{1}{n^{3/2}}\right).$$

share|improve this answer
    
OMG, this is awesome. I was trying to do these calculations but I got stuck. Your answer looks like it took a lot of work - thanks so much!!!! –  robinson Jun 23 '11 at 22:23
    
@robinson: You're quite welcome. Trying to figure out problems like this is fun for me. :) And it wasn't as much work as it might look like; I've worked with a lot of this stuff before. The key insight (thanks to Emre) was to transform the difficult integral into a beta function. –  Mike Spivey Jun 24 '11 at 1:48

Hint: $P(Y > x) \equiv P(Z_1 > x) \dots P( Z_n > x)$.

(If the smallest element is greater than some number, so must every single element be greater than that number. The product follows from the independence assumption.)

share|improve this answer
    
Perhaps I'm doing it wrong, but I got stuck trying to work out the integrals that arise from doing this. Namely, using $E[Y]=\int P(Z>x)^n$, led me to trying to compute $\int_0^1 (1-(1/2)x^2)^n$, which I'm not sure how to do. –  robinson Jun 23 '11 at 0:46
    
Let $t=x^2/2$ and compare with the definition of the beta function. –  Emre Jun 23 '11 at 1:09

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.