6
$\begingroup$

I saw this question on a friend's problem set (I mean it, not MY homework) and kept thinking about it, although without success.

Let $X$ be a a random variable such that $ -K < X < K$ for some $ K >0$. Also, let $E(X) = 0$. Now, prove that $E(e^{X}) < \frac{1}{2}(e^{-K} + e^{K})$.

I'm trying to use Jensen's inequality and so on but I made no progress. Also I tried fooling around with some smart squares but couldn't solve it this way either. Any ideas?

$\endgroup$
  • $\begingroup$ Interesting. For a uniform distribution on $[-K,K]$, you'll get $E(e^X) = \frac12(e^K-e^{-K})$, and for a distribution concentrated at $\pm K$ you'll get $E(e^X) = \frac12(e^K+e^{-K})$. I don't see a general proof yet. $\endgroup$ – Ted Shifrin Apr 11 '18 at 21:09
  • 1
    $\begingroup$ If ever $X\stackrel d=-X$ then $E(e^X)=E(e^{BX})$ where $B$ is independent of $X$ with $P(B=1)=P(B=-1)=\frac12$ hence $E(e^X)=\sum\limits_n\frac1{n!}E(B^n)E(X^n)=\sum\limits_n\frac1{(2n)!}E(X^{2n})\leqslant\sum\limits_n\frac1{(2n)!}K^{2n}=\frac12(e^K+e^{-K})$. But if the distribution of $X$ is not symmetric... $\endgroup$ – Did Apr 11 '18 at 21:13
10
$\begingroup$

If $X$ only takes values in $[-K,K]$ then it is true, for all convex $\phi$, that $\phi(X) \leqslant \frac{(\phi(K)-\phi(-K))}{2K}X+\frac{\phi(-K)+\phi(K)}{2}$. This is because the right hand is the equation of the line from $(-K,\phi(-K))$ to $(K,\phi(K))$

Setting $\phi=\exp$ and taking expectations in the above inequality gives $\mathbb{E}(e^X)\leqslant \frac{1}{2}(e^{K}+e^{-K})$

To get the strict inequality, notice that if $\phi$ is strictly convex and $X$ takes values in $(-K,K)$ all inequalities above become strict. (Strict inequalities stay strict by taking expectations because positive variables have positive expectations)

Jensen's inequality's proof looks a lot like the argument above, but gives a minorant of $\phi(X)$ by a line, instead of a majorant like it is the case here.

$\endgroup$
  • 1
    $\begingroup$ Fine argument (+1). It would be better to mention strict convexity to get an strict inequality as in the question. $\endgroup$ – Shashi Apr 11 '18 at 21:53
  • $\begingroup$ Good point, thanks ! $\endgroup$ – Sergio Apr 11 '18 at 22:55
  • $\begingroup$ This is definitely a great way to look into this problem, thanks a lot @Sergio! $\endgroup$ – Raul Guarini Apr 12 '18 at 17:01
1
$\begingroup$

Here is a messy way to solve it. First we prove it when $X$ is given by choosing an element of the list $(x_1,\dots,x_n)$ uniformly at random. The list may have repeats. The general result then follows since any random variable can be approximated by a discrete one.

The discrete problem is equivalent to the following maximization problem:

 Maximize: $E[e^X]=\frac1n\sum\limits_{i=1}^n e^{x_i}$
 Subject to: $\sum\limits_{i=1}^n x_i=0$, and for all $i$, $-K\le x_i\le K$.

Notice I replaced the condition $-K<X<K$ with $-K\le X\le K$ so that the region is compact, ensuring the maximum is attained.

I claim that $(x_1,\dots,x_n)$ does not obtain the maximum as long as there exist two numbers $x_i$ and $x_j$ which are in the interior $(-K,K)$. This is because given two such numbers, increasing the larger of these by $\epsilon$ and decreasing the smaller by $\epsilon$ will preserve $\sum\limits_{i=1}^n x_i$, but will increase $\frac1n\sum\limits_{i=1}^n e^{x_i}$ (prove this!).

Therefore, if $(x_1,\dots,x_n)$ attains the maximum, then all but at most one of the variables are equal to $\pm K$. It is easy to see that the only way to satisfy this is to have an equal number of variables equal to $+K$ and $-K$, with the leftover variable (if any) equaling zero. This results in a maximum value of at most $\frac1n(\lfloor n/2\rfloor e^{-K}+\lfloor n/2\rfloor e^K)\le \frac12(e^K+e^{-K})$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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