3
$\begingroup$

Suppose $X$ is a random real variable with zero mean and finite second moment $\langle X^2\rangle$. Under what conditions can we give a bound (upper/lower) for the third moment $\langle X^3\rangle$?

$\endgroup$
  • 1
    $\begingroup$ What kinds of "conditions" do you have in mind? You need something, because the third moment can be undefined (consider the Student $t$ distribution with 3 df, for instance), it can be infinite, and it can be negatively infinite: there are no general bounds. $\endgroup$ – whuber Mar 2 at 15:05
  • 1
    $\begingroup$ That actually answers my question, thanks @whuber $\endgroup$ – becko Mar 2 at 23:50
8
$\begingroup$

There is no such universal bound.

Here is a cute fact to put this question in perspective:

Given real numbers $k^\prime \gt k \ge 0,$ there exist random variables $X$ that have a finite $k^\text{th}$ moment but infinite $k^{\prime\,\text{th}}$ moment.

It doesn't matter whether we consider these raw moments or central moments: see Mathematical statistics (moments vs. central moments).

To prove the fact, let $X$ be supported on the positive real numbers with a density proportional to $(1 + x^2)^{-(k^\prime+1) /2}.$ Since $k^\prime \gt 0,$ this density integrates to a finite number, showing such distributions exist.

Because $1+x^2 \gt x^2,$

$$\mu_{k}(X) \propto \int_0^\infty \frac{x^k\,\mathrm{d}x}{(1 + x^2)^{(k^\prime+1)/2}} \le \int_0^\infty x^{k - k^\prime - 1}\,\mathrm{d}x = \frac{1}{k^\prime - k} \le \infty$$

but

$$\mu_{k^\prime}(X) \propto \int_0^\infty \frac{x^{k^\prime}\,\mathrm{d}x}{(1 + x^2)^{(k^\prime+1)/2}} \gt \int_1^\infty \frac{x^{k^\prime}\,\mathrm{d}x}{(2 x^2)^{(k^\prime+1)/2}} \propto \int_1^\infty \frac{\mathrm{d}x}{x} = \lim_{x\to\infty}\log(x)$$

diverges, QED. The inequality in the last line (a) dropped the area from $0$ to $1$ and (b) uses $2x^2 \ge 1 + x^2$ when $x \ge 1.$

The question concerns the case $k^\prime=3$ and $k=2.$ To apply the foregoing analysis, let $\sigma^k$ be the given $k^{\text{th}}$ moment and let $X$ be the random variable described above. Then the $k^\text{th}$ moment of $$ \frac{\sigma\,X}{\mu_k(X)^{1/k}} $$ equals $\sigma^k$ but its $k^{\prime\,\text{th}}$ moment is infinite.

| cite | improve this answer | |
$\endgroup$

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.