1
$\begingroup$

Given

  • $X_i$ are independent random variables.
  • $|X_i| < 1$
  • $E[X_i] = 0$
  • $X = \sum_i^n X_i$
  • $var(X)=\sigma$

Prove:

$$ E(X^p)^{1/p} = O(\sqrt{p}\sigma +p)$$ for all even p

Things I've tried:

First note, that all terms with more than p/2 terms are 0 (since $E[X_i] = 0$).

Furthermore, note that $\sum_i X^{4} \leq \sum_i X^2$

So this ends up being some way to count the various terms involving exactly $t$ variables. I don't know how to count this. What should I try?

$\endgroup$
4
  • $\begingroup$ Where does this question come from? $\endgroup$
    – Igor Rivin
    Aug 22, 2012 at 15:41
  • 1
    $\begingroup$ Look at Lemma 4, page 172, in Billingsley's book Convergence of probability measures. Lemma is for p=4, but it works for all even p. If I am not mistaken, this lemma gives the bound you are looking for. $\endgroup$ Aug 22, 2012 at 16:08
  • $\begingroup$ @waler: Can you identify which section that's from? In my copy of the second edition of that book, there is no Lemma 4 on or near page 172. $\endgroup$ Aug 22, 2012 at 18:04
  • $\begingroup$ @Mark Meckes: 1968 edition, chapter 4. dependent variables, section 20: mixing processes, paragraph on moment inequalities. $\endgroup$ Aug 22, 2012 at 18:40

1 Answer 1

2
$\begingroup$

Take a loot at R. Ibragimov and Sh. Sharakhmetov, The Exact Constant in the Rosenthal Inequality for Random Variables with Mean Zero, Theory Probab. Appl., 46(1), 127–132. (6 pages) Read More: http://epubs.siam.org/doi/abs/10.1137/S0040585X97978762

Abstract Let $\xi_1, \ldots, \xi_n$ be independent random variables with ${\bf E}\xi_i=0,$ ${\bf E}|\xi_i|^t<\infty$, $t>2$, $i=1,\ldots, n,$ and let $S_n=\sum_{i=1}^n \xi_i.$ In the present paper we prove that the exact constant ${\overline C}(2m)$ in the Rosenthal inequality $$ {\bf E}|S_n|^t\le C(t) \max \Bigg(\sum_{i=1}^n{\bf E}|\xi_i|^t,\ \Bigg(\sum_{i=1}^n {\bf E}\xi_i^2\Bigg)^{t/2}\Bigg) $$ for $t=2m,$ $m\in {\bf N},$ is given by $$ \overline C(2m)=(2m)! \sum_{j=1}^{2m} \sum_{r=1}^j \sum \prod_{k=1}^r \frac {(m_k!)^{-j_k}} {j_k!}, $$ where the inner sum is taken over all natural $m_1 > m_2 > \cdots > m_r > 1$ and $j_1, \ldots, j_r$ satisfying the conditions $m_1j_1+\cdots+m_rj_r=2m$ and $j_1+\cdots+j_r=j$. Moreover $$ \overline C(2m)={\bf E}(\theta-1)^{2m}, $$ where $\theta $ is a Poisson random variable with parameter 1.

In your case, one gets $$ {\bf E}|S_n|^{2m} \le \overline C(2m) \max( n, (var S_n)^{m})= \overline C(2m) \max( n, n^{m}c^m). $$ where $c=var(\xi_i)$. Thus for large $m$, $$ \Bigl( {\bf E}|S_n|^{2m}\Bigr)^{\frac 1{2m}} \le (\overline C(2m))^{1/(2m)} \sqrt{n} c. $$ Finally, they cite earlier papers where it was shown that $$ \overline C(t)= O( t/\log t). $$ Ans since is $(t/\log t)^{1/t}<\sqrt{t}$ for $t>2$ we are done.

$\endgroup$
3
  • 3
    $\begingroup$ That is a nice paper but is overkill for what the OP asks. For the $t/ \log t$ result, see the first paper by Schechtman, Zinn, and myself. It also contains an easy proof of the bound $t/ (\log t)^{1/2}$. $\endgroup$ Aug 22, 2012 at 17:11
  • $\begingroup$ Johnson, W. B.; Schechtman, G.; Zinn, J. Best constants in moment inequalities for linear combinations of independent and exchangeable random variables. Ann. Probab. 13 (1985), no. 1, 234–253. $\endgroup$ Aug 22, 2012 at 17:14
  • $\begingroup$ @Bill Johnson: absolutely, severe overkill. result must follow also from some easier inequalities as well. $\endgroup$ Aug 22, 2012 at 18:36

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.