3stylelife

This blog has closed. Follow me to 3stylelife. Thanks - Barry
Share/Save/Bookmark Subscribe

Tuesday, May 27, 2008

Random Expectations

A friend of mine recently posed an interesting question to me. We can easily see that if we choose a random number between 0 and 1, the expected value should be one-half. Now, in this case, the only number is the same as the largest number, so that the expected value of the largest number is also one-half.

So what occurs if we choose two random numbers? Can we determine the expected value of the largest number in this case? What if there are N numbers? I work through a pretty direct interpretation for the N=2 case, then find a general approach for arbitrary N.

Random Expectations

Given the result of N/(N+1), I noticed that this is equal to 1 minus the integral from 0 to 1 of x^N. This would be a really quick solution, but I can't seem to find an interpretation of the situation which lends itself nicely to that expression. Perhaps it's a coincidence, but who knows?

4 comments:

Efrique said...
This comment has been removed by the author.
Efrique said...
This comment has been removed by the author.
Efrique said...

Your pdf does things rather the hard way.

Consider:

[let "_n" stand for "subscript n" and "^n" for "superscript n"]

Let X_1, X_2 ..., X_n be independent standard uniform random variables.
Let Y = max(X_1, X_2 ..., X_n)

P(Y <= y) = P(X_1 <= y) x P(X_2 <= y) x ... x P(X_n <= y)

= y^n (0 < y < 1)

This is the distribution function of the maximum

Hence the density for y is the derivative of this, which is n y^(n-1), (0 < y < 1).

And therefore the expectation is the integral of n.y^n over the same interval.


You may find the wikipedia entry on Order statistics useful (though it could be better written).

[I considered whether the integral for the expectation of the minimum would help (since by symmetry the expectation of the maximum would be 1 minus that). It does yield 1/(N+1), as it should, but I don't see an obvious direct argument that relates the integral for it to the integral for x^N.]

Then I remembered that the expectation of a random variable is equal to the integral of its survivor function. And that seems to yield what you want:

Let S(y) = 1 - F(y) (where F is the distribution function of Y).

That is E(Y) is the integral from 0 to 1 of S(y) dy

= integral from 0 to 1 of 1 - y^n dy

= integral{_0^1} 1 dy - integral{_0^1} y^n dy

= 1 - integral {0^1} y^n dy

which looks like the desired result.

So, no, not exactly a coincidence.

fashionablemathematician said...

Really nice analysis efrique, the survivor function stuff is something I hadn't run across before. Thanks for the information!