Let $a_1, a_2, ...$ be a sequence of positive numbers such that $\frac{\sum_{j = 1} ^ n a_j}{n} \to \infty$ as $n \to \infty$. What can we say about the behavior of $$\frac{(\sum_{j = 1} ^ n a_j)^p}{n^{p - 1} \sum_{j = 1} ^ n a_j ^ p}$$ as $n \to \infty$ for $p > 1$? If I have my way, it should go to $0$ under these conditions, but I haven't been able to show this, I suspect because I'm missing some inequality. I can use Jensen's inequality to show that the expression above is in $[0, 1]$, but I guess I need something sharper to get it to go to $0$; maybe some inequality that leaves a factor of $\sum a_j / n$ in the denominator to exploit.


Not neccesarily zero: Let $a_{j}=j$. Then $$\sum_{j\leq n}a_{j}\sim\frac{n^{2}}{2}\ \text{and}\ \sum_{j\leq n}a_{j}^{p}\sim\frac{1}{p+1}n^{p+1}$$so that $$\frac{\left(\sum_{j\leq n}a_{j}\right)^{p}}{n^{p-1}\sum_{j\leq n}a_{j}^{p}}\sim\frac{p+1}{2^{p}}.$$

Hölder Mean: We can write the terms in a slightly nicer way. Notice $$\frac{\left(\sum_{j\leq n}a_{j}\right)^{p}}{n^{p-1}\sum_{j\leq n}a_{j}^{p}}=\frac{\left(\frac{\sum_{j\leq n}a_{j}}{n}\right)^{p}}{\left(\frac{\sum_{j\leq n}a_{j}^{p}}{n}\right)}=\left(\frac{M_{1}(a_{1},\dots,a_{n})}{M_{p}(a_{1},\dots,a_{n})}\right)^{p}$$ where $M_{r}(\boldsymbol{x})$ is the generalized Hölder mean. For $p>1$ we know that $M_{1}(\boldsymbol{x})\leq M_{p}(\boldsymbol{x})$ so that the above quantity is always in $[0,1]$.

Sequence going to zero: We cannot say too much more, such as a lower bound, as it is possible to construct a sequence which goes to zero.

Let $a_{j}=e^{j}$. Then $$\sum_{j\leq n}a_{j}\sim e^{n}\ \text{and}\ \sum_{k\leq n}a_{j}^{p}\sim\frac{1}{p}e^{np}.$$ Consequently $$\frac{\left(\sum_{j\leq n}a_{j}\right)^{p}}{n^{p-1}\sum_{j\leq n}a_{j}^{p}}\sim\frac{\left(e^{n}\right)^{p}}{n^{p-1}\left(\frac{1}{p}e^{np}\right)}\sim\frac{p}{n^{p-1}}\rightarrow0.$$

Hope that helps,

  • 2
    $\begingroup$ Also perhaps of interest: $a_j = j^k$ for $k > 0$ gives $ \frac{(\sum_{j\le n} a_j)^p}{n^{p-1} \sum_{j\le n} a_j^p} \sim \frac{kp+1}{(k+1)^p}$ which could be anything in $(0,1)$. $\endgroup$ – Robert Israel Aug 7 '11 at 20:32
  • 2
    $\begingroup$ And, I think, $a_j = \ln j$ gives a limit of 1. $\endgroup$ – Robert Israel Aug 7 '11 at 20:39
  • $\begingroup$ @Robert Israel: Thanks for pointing that out! $\endgroup$ – Eric Naslund Aug 7 '11 at 20:58
  • $\begingroup$ Meh, this makes me unhappy since I was hoping to use it to prove an analogous statement about random variables that I know is true. The analogous statement is that if you replace the sequence $a_n$ by a sequence of iid positive random variables such that $X_n \notin L_1$ then the limit is $0$ almost surely. I guess I actually have to use probability theory in order to solve it :( $\endgroup$ – guy Aug 7 '11 at 21:09
  • 1
    $\begingroup$ @guy, the reason why your analogous statement for non integrable iid random variables holds is indeed truly stochastic, namely it is the fact that $X_1+\cdots+X_n$ is of the order of the maximum $M_n=\max\{X_1,\ldots,X_n\}$ up to time $n$. Similarly, $X_1^p+\cdots+X_n^p$ is of the order of $M_n^p$ hence the ratio you consider should converge almost surely to zero because of the power of $n$ in the denominator. $\endgroup$ – Did Aug 8 '11 at 10:19

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.