5
$\begingroup$

Suppose that $f\left(x\right)\geq0$ is continuous on $\left[-\infty,\infty\right]$ and $\int_{-\infty}^{\infty}f\left(x\right)dx=1$. Is it true that $\int_{-\infty}^{\infty}\left|x\right|f\left(x\right)dx\leq\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\left|x-y\right|f\left(x\right)f\left(y\right)dxdy$?

Thanks.

$\endgroup$
4
  • 1
    $\begingroup$ Since there is no background for the problem, why do you tag "probability"? It's (classical analysis). Why don't you do the change of variable $\hat y=y-x$ yourself? Of course, you have to do it carefully by replacing the improper integral by a proper one... $\endgroup$ Jun 9, 2010 at 6:24
  • 1
    $\begingroup$ As Robin suggested, this problem could be solved by using probability and expection. I tried to integrate them by removing the absolute value and changing variables, but it seems hard. $\endgroup$
    – user4606
    Jun 9, 2010 at 7:45
  • 1
    $\begingroup$ Wadim, it's clearly a question (but not a very interesting one) about continuous probability distributions, even if not phrased in that language. It's also going to be a race between the proposer providing more context/conditions to make the question more relevant/interesting and closure. $\endgroup$ Jun 9, 2010 at 11:12
  • $\begingroup$ I rolled back the original tagging. I am convinced now that the readers of the post are more related to probability. $\endgroup$ Jun 10, 2010 at 2:49

3 Answers 3

10
$\begingroup$

As suggested by Robin, we would need to prove that if $X$ and $Y$ are independent random variables with the same distribution, the inequality $E|X|\leq E|X-Y|$ holds.

This is certainly NOT true in general; for example suppose $X$ and $Y$ have some constant non-zero value; then the RHS is 0 and the LHS is positive.

However, if you add the condition that $E(X)=0$ then the result is true. (In your notation, $\int_{-\infty}^{\infty}xf(x)=0$ ).

In that case write $p=P(X\geq 0)$, and write $X_+$ for the positive part of $X$ and $X_-\leq 0$ for the negative part of $X$.

We have $0=EX=E|X_+|-E|X_-|$, so $E|X_+|=E|X_-|$.

Hence also $E|X|=E|X_+|+E|X_-|=2E|X_+|$.

Now $E|X-Y| \geq E(|X-Y|I(Y<0)I(X\geq 0)) + E(|X-Y|I(X<0)I(Y \geq 0))$

$=2 E(|X-Y|I(Y<0)I(X\geq0))$

$=2 E(|X_+|I(Y<0)I(X\geq0)) + 2E(|Y_-|I(Y<0) I(X\geq 0))$

$=2(1-p)E(|X_+|I(X\geq 0)) + 2p E(|Y_-|I(Y<0))$

$=2(1-p)E|X_+| + 2pE|Y_-|$

$=2E|X_+|$

$=E|X|$

(using at various times the facts that $X$ and $Y$ are independent and that $X$ and $Y$ have the same distribution)

$\endgroup$
3
  • $\begingroup$ James, I vote for your scepticism! The condition $\int_{\mathbb R}xf(x)\ dx=0$ is to replace the parity of $f(x)$, so your proof is essentially the same. And I now show why this is not true in general. $\endgroup$ Jun 9, 2010 at 10:43
  • $\begingroup$ James, I can't get the voting logic: if I give my students your solution, they would clearly complain. Anyway, seeing a very strong preference of yours by the MO community, I have to delete mine. $\endgroup$ Jun 9, 2010 at 23:29
  • 1
    $\begingroup$ Hi Wadim, what would they complain about? (Maybe the fact that it's written out in terms of random variables rather than integrals; just because I'd seen Robin's answer and I was feeling lazy and it seemed easier to write it out that way... no doubt my answer could be shortened.) I think your response was fine, no need to delete it. The answer I gave was somewhat more general since it covers any case where the distribution has mean 0, not just those which are symmetric around 0. $\endgroup$ Jun 10, 2010 at 16:25
3
$\begingroup$

This can be rephrased as follows.

Let $X$ and $Y$ be independent random variables with the same, continuous, distribution. Is it true that $E(X)\le E(|X-Y|)$.

  1. Is this likely?

  2. Is it true for a discrete distribution?

  3. If one has a discrete counterexample, can one convert it to a continuous counterexample?

Added I now realise that there was some absolute value signs in the original integral (these come out badly on my screen) and I should have written $E(|X|)\le E(|X-Y|)$. Plus ca change...

$\endgroup$
1
  • $\begingroup$ Sorry, I keep imagining the question written with the other inequality! the function I wrote below (that is 1/x supported on [1,n] and normalized) gives indeed n-1 and 2n+o(n), that is, it satisfies the better inequality that you are suggesting. $\endgroup$ Jun 9, 2010 at 8:57
1
$\begingroup$

As it is mentioned in James Martin's answer, if one does not make additional assumptions, the statement is wrong.

One correct statement is: if $X$ and $Y$ are independent r.v.'s with finite expectations, and $\mathbf{E}Y=0$, then $\mathbf{E}f(X)\le \mathbf{E}f(X+Y)$ for any convex function $f$ such that these expectations are finite.

Proof: The sequence of two r.v.'s $X$ and $X+Y$ is a martingale. Therefore, the sequence $f(X), f(X+Y)$ is a submartingale, and the claim follows.

Of course, $f$ should be taken to be the absolute value function, and $Y$ should be replaced by $-Y$, if we want this statement to look more like the statement in question.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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