This blogpost discusses a beautiful way for solving the following interesting problem.
Example 1. If are independent and identically distributed random variables taking values in the set of real numbers then show that |
The above result follows from the positive semi-definiteness of Here we discuss an alternate solution, which is due to professor Fedor Petrov.
Solution with comments (thanks to professor Fedor Petrov). Notice that the problem doesn’t assume finiteness or even existence of the mean of Echoing on the standard practice of checking the case of infinite means first, we first get rid of the case when (and hence ) has an infinite expected value. We leave this as an exercise for the readers.
Exercise 2.
Prove the claim in Example 1 in the case when
So it suffices to consider the case when has a finite expected value. Keeping aside this simplification for some time, let us look at the problem in a different way. The problem presents a claim about two iid random variables and This might suggest that it might be reasonable to rewrite the problem as a claim about the common law of Letting denote the common law of we can rewrite the claim in Example 1 as follows.
For any arbitrary probability distribution on the real numbers one has whenever and are independent random variables each having law |
Echoing on the general problem solving strategy of finding a sufficient criteria, one may like to find a claim which implies (1). A naive start would be to guess that the random variable is non-negative with probability one. But this is not true.
Exercise 3.
Show that there exists iid random variables such that
However we have
Indeed, for any real numbers one has The next natural step would be to take expectations on both sides of (2) and use the iid-ness of to get which is not quite the same as what we wanted but only slightly different from it.
Let us amplify the above argument to get a proof. The key role will be played by the following somewhat popular inequality.
Claim 4. For any natural number and any arbitrary reals one has |
Applying the inequality (3) to the sample points for jointly independent random variables having law and take expectation on both sides we get that
thanks to Exercise 2 we see that sending gives us the desired claim.
So it suffices to prove Claim 4. A standard way is to induce on
Exercise 5.
Show (3) when and notice that there is nothing to prove if Now, assuming the claim true for every $n \le m$ for some we aim to prove the claim for To this end, replace each by for a fixed and notice that the right hand side of (3) remains unchanged; optimize in by minimizing the left hand side of (3) and use the induction hypothesis to get the claim.
This completes the proof.
Remark 6.
An interesting fact is that the inequality (1) holds even for iid vector-valued random variables and but the proof is slightly more difficult. It involves a trick of writing magnitudes for vectors in terms of inner products with random unit vectors. Here is a quick summary of this trick, by professor Terence Tao.
A retrospective analysis of the above solution, shows us that, in some sense, we removed probability from the picture (in the sense of finding a deterministic problem which implies the original claim made for random variables). This is an example of a more general strategy of de-probabilizing a problem. We shall discuss more about this matter in a later blogpost.
Also the fact that we are using Euclidean norm plays a crucial role in Example 1. For more details on it, see this post by professor Russ Lyons. This point will be discussed in more details in a later blogpost.