A Cool Fact about Uniform Distributions

A Cool Fact about Uniform Distributions

On Monday I received the problem of the week from Doubleclick Technical Group guru, Andrew. The problem was as follows:

If I have a hat with all the numbers from 0 to 1 in it, and I start randomly choosing numbers from it...how many times will the number I draw be the maximum value for the numbers I've already drawn when I draw a total of n numbers from the hat? It should be noted that the question is asking: on average how many times will the maximum value change.

I know the answer is obvious to most people, but I had to do a little bit of work to get it. In the end I discovered a cool fact...that Uniform distributions and Beta distributions are related!

First what we really have here is a process that selects a random variable (r.v) from a Uniform(0,1) distribution. A uniform(0,1) distribution has mean .5 and variance .25. Uniform distributions can be used in many cool ways, the least of which is simulating a coin flip.

A depiction of the probability density of a Uniform distribution

Secondly, when we talk about the maximum of a set of r.v, max{X1, X2,....,Xn}, we can also refer to the last element in the sorted, or ordered, set {X(1), X(2), .... , X(n)<----this would be the maximum}. In this case the sort goes in ascending order.

Here is where it gets interesting! (ok interesting if your into this sort of thing, and I am...so....) When we sort the random variables produced from a Uniform distribution, the ordered set takes on the appearance of a Beta distribution! Furthermore, each element in the set has it's own specific Beta distribution....i.e. the k-th element in the set {X(1), ... , X(k), ... , X(n)} has the distribution Beta(k, k+1-n)! Wow!

These are various depictions of Beta probability density functions...neat-o!

Now, we know that the Beta(k, k+1-n) distribution has a mean value of k/(k+k+1-n). This is quite useful, because in our problem above we are asking the question: how many times does the maximum change, or rather, what is the sum of the probabilities for each draw to be a maximum. We can conclude, with some thinking, that on average after n draws the maximum of a set of uniformly drawn r.v.'s is 1 - the mean value of the n-th Uniform r.v. from the ordered set. Thus the Probability that our nth draw is the new maximum is Pr(Xn = max) = 1 - mean(Beta(n, n+1-n)) = 1 - n/(n+1)! It should be noted that I am interchanging the actual value of the n-th draw with it's probability because for the r.v. to be a max it only has to fall within the previous max's value and 1....making the probability of the draw being the max and the mean value of the Beta(n, n+n+1-n) the same thing.

So... if X0, X1, ... , Xn-1 are our draws...we have the following:

Pr(X0 = max) = 1 -0 (first draw must be the maximum)

Pr(X1 = max) = 1- 1/2 (1/(1+1))

Pr(X2 = max) = 1- 2/3 = 1/3

Pr(X3 = max) = 1- 3/4 = 1/4

.

.

.

Pr(Xn-1 = max) = 1-n-1/(n)

And by summing up these probabilities, we get the expected or average amount of times that the number drawn is the maximum...

sum(1-j/(j+1)) [for j=0 to n-1] !

So for n=10,000,000 we expect to obtain about 14 new maximums along the way!

And you thought Math was useless :-) Next week if you're lucky I will discuss how most natural processes can be described by a Normal distribution...including the stock market!