0
$\begingroup$

Consider iid RVs $\{X_n\}$, and $E\left[\frac{1}{X_n}\right]=+\infty$. I'm looking for an upper bound for the tail of the harmonic mean of $\{X_n\}$, i.e. I want to upper bound the following

\begin{equation*} P\left(\sum_{n=1}^N \frac{1}{X_n} > t\right) = P\left(\frac{N}{\sum_{n=1}^N \frac{1}{X_n}} < \frac{N}{t}\right). \end{equation*}

But I'm out of luck because $E\left[\sum_{n=1}^N\frac{1}{X_n}\right]=+\infty$, and I cannot apply Markov's inequality; I don't think I can invoke a Chernoff bound either.

Any ideas?

$\endgroup$
  • 1
    $\begingroup$ You may use the fact that the hyperbolic mean of the $X_n$ is lower than their arithmetic mean, and find an upper bound for the arithmetic mean... Just my 2 cents. $\endgroup$ – Nicolas FRANCOIS Jul 19 '17 at 15:11
  • $\begingroup$ @NicolasFRANCOIS you're absolutely correct, but I just realized I need to upper bound the CDF of the harmonic mean, and NOT the tail. $\endgroup$ – ToniAz Jul 19 '17 at 16:09
  • $\begingroup$ I suspect that for given $t$ and $N$ there will be particular distributions meeting your requirements for $X_n$ giving values for the probability anywhere in $(0,1]$ $\endgroup$ – Henry Jul 19 '17 at 17:40
1
$\begingroup$

If $\sum_{n=1}^N 1/X_n > t>0$ then at least one $X_n < N/t$, so $$ P\left(\sum_{n=1}^N \frac{1}{X_n} > t\right) \le P\left(\bigcup_{n=1}^N (X_n < N/t)\right) = 1 - P(X_n \ge N/t)^N $$

$\endgroup$
  • $\begingroup$ Fantastic! I established a lower bound in a similar way, using the fact that if $\forall n,\,1/X_n > t/N$, then $\sum_{n=1}^N \,1/X_n > t$. Thank you for the upper bound! $\endgroup$ – ToniAz Jul 19 '17 at 17:54

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.