0
$\begingroup$

I am trying to prove the following statement: $$ E[g(X)] E[X^2g(X)]\ge E[Xg(X)] E[Xg(X)] $$ where $X$ is a random variable, $E[\cdot]$ denotes the expectation operator with respect to $X$, and $g(\cdot)$ is a well-behaved,positive-valued, bounded function.

For $X$ being discrete, I have both the intuition and the proof. For simplicity, assume $X$ takes only two values $x_1$ and $x_2$, with respective probabilities $p_1$ and $p_2$. Further, let $g(x_1)=g_1$ and $g(x_2)=g_2$. Then, the left hand side of the above inequality is given as \begin{align} E[g(X)] E[X^2g(X)]&=[g_1p_1+g_2p_2]\times[{x_1}^2g_1p_1+{x_2}^2g_2p_2] \\ &=x_1^2g_1^2p_1^2+x_2^2g_1g_2p_1p_2+x_1^2g_1g_2p_1p_2+x_2^2g_2^2p_2^2. \end{align} Similarly, the right hand side is given as \begin{align} E[Xg(X)] E[Xg(X)]&=[x_1g_1p_1+x_2g_2p_2]\times[x_1g_1p_1+x_2g_2p_2] \\ &=x_1^2g_1^2p_1^2+2x_1x_2g_1g_2p_1p_2+x_2^2g_2^2p_2^2. \end{align} So the difference between these two expressions is $$ E[g(X)] E[X^2g(X)] - E[Xg(X)] E[Xg(X)] = (x_1-x_2)^2g_1g_2p_1p_2\ge0, $$ which implies that the given inequality is correct. This procedure can be applied to the case where $X$ admits $n$ discrete data points instead of just two.

My question is how to prove the same relationship for continuous random variables, i.e., show that $$ \int g(x)f(x)dx \int x^2 g(x) f(x) dx \ge \int x g(x) f(x) dx \int x g(x) f(x) dx $$ where, $f(\cdot)$ is the probability density function of $X$ and $g(\cdot)\ge0$.

$\endgroup$
4
  • $\begingroup$ In your "simple case" you seem to be assuming that $g$ takes positive values (because you want $g_1g_2\geq 0$ $\endgroup$
    – Yemon Choi
    Feb 4, 2015 at 3:26
  • 3
    $\begingroup$ If you assume that $g\geq 0$ in the general problem then the result follows immediately from the Cauchy-Schwarz inequality: $Xg(X)= Xg(X)^{1/2}\cdot g(X)^{1/2}$ $\endgroup$
    – Yemon Choi
    Feb 4, 2015 at 3:27
  • 2
    $\begingroup$ And if you don't require $g \ge 0$ then this is false, even in the discrete case. Let $X = \pm 1$ with probability 1/2 each (the Rademacher distribution) and let $g(x) = x$ on $[-1,1]$ (you can extend it to a smooth bounded function on $\mathbb{R}$ if you like). Then your equation reads $E[X] E[X^3] \ge E[X^2] E[X^2]$ which is false; the left side is 0 and the right side is 1. For an example with a continuous random variable, let $X$ be uniformly distributed on $[-1,1]$. $\endgroup$ Feb 4, 2015 at 4:19
  • $\begingroup$ Yes, I should have clearly stated that $g(\cdot) \ge 0$. Otherwise, based on the counterexamples you suggested the assertion is not necessarily true. Yemon, can you elaborate more about using the Cauchy-Schwarz inequality? My understanding is that one can show $E[X^2g(X)]\ge(E[Xg(X)^{1/2}])^2$. However, this is not exactly what I need. $\endgroup$
    – Aliveli
    Feb 4, 2015 at 14:23

3 Answers 3

3
$\begingroup$

Converting Yemon Choi's comment to an answer:

If $E[X g(X)] \le 0$, we are done because the left side of your proposed inequality is nonnegative. So assume $E[X g(X)] \ge 0$.

First, by the triangle inequality we have $$E[X g(X)] \le E [|X| g(X)].\tag{1}$$

The Cauchy-Schwarz inequality says that for any random variables $Y,Z$ we have $$E[|YZ|] \le \sqrt{E[Y^2]\, E[Z^2]}.$$ Applying this with $Y = \sqrt{g(X)}$, $Z = X \sqrt{g(X)}$ we obtain $$E[|X| g(X)] \le \sqrt{E[g(X)] \,E[X^2 g(X)]}. \tag{2}$$ Now combine (1) with (2) and square both sides.

$\endgroup$
2
  • $\begingroup$ Thanks Nate - I was in a rush earlier so you have saved me the trouble of writing it out $\endgroup$
    – Yemon Choi
    Feb 4, 2015 at 16:52
  • $\begingroup$ Thanks a lot! This is indeed a very nice answer! In my case, X only takes positive values so the proof is even easier. Thanks! $\endgroup$
    – Aliveli
    Feb 6, 2015 at 3:13
2
$\begingroup$

If $(\Omega,\mathcal F, P)$ is the probability space you are using, then $$ \mu(A) = \frac{E[A g(X)]}{E[g(X)]} $$ for $A \in \mathcal F$ is another probability measure. Your inequality follows from $$ \|X\|_2 \ge \|X\|_1 $$ using norms of the spaces $L^2(\Omega,\mathcal F, \mu)$ and $L^1(\Omega,\mathcal F, \mu)$. I leave proof of this to you...

$\endgroup$
0
$\begingroup$

You can prove this working backwards from your last expression:

$(y - x)^2 g(x) g(y) f(x) f(y) > 0$, now if you expand this term you obtain

$g(x) f(x) y^2 g(y) f(y) + x^2 g(x) f(x) g(y) f(y) \geq 2 x g(x) f(x) y g(y) f(y)$.

If you now integrate this inequality with respect to $ x$ and $y$ you get the result you want.

$\endgroup$

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.