I am well aware (as for instance discussed here https://math.stackexchange.com/questions/910846/is-it-true-in-general-that-e1-x-1-ex) that for an arbitrary random variable $X$ it does not hold that $\mathbb{E}[1/X]=1/\mathbb{E}[X]$. However, when one cannot calculate $\mathbb{E}[1/X]$ directly, is there a good (rigorous) way to approximate $\mathbb{E}[1/X]$ (ideally, regardless of the distribution of $X$)?

  • 3
    $\begingroup$ What information do you have available regarding $X$? If you have central moments up to some order, and they fall off quickly compared to $\mu^n$, then you can use a series expansion of $\frac{1}{X}$ around $X = \mu$, for example. $\endgroup$ – student Jan 9 at 11:27
  • $\begingroup$ Unfortunately I do not have much information about $X$ - I only really know its expectation, which is why I am primarily interested in approximations that are independent of its distribution. However any inputs or ideas would be helpful. $\endgroup$ – Silvia Jan 9 at 12:59
  • 2
    $\begingroup$ Pick e.g. $X \sim \Gamma(\frac{\mu \nu}{\mu\nu - 1}, \frac{\nu}{\mu\nu - 1})$, then $\mathrm{E}[X] = \mu$ but $\mathrm{E}[1/X] = \nu$. $\endgroup$ – student Jan 9 at 13:54
  • $\begingroup$ @student, are you trying to imply that no approximations in term of E[X] can exist? $\endgroup$ – Silvia Jan 11 at 10:25
  • $\begingroup$ Yes, and even in the bounded case you can find such examples using basically any two-parameter distribution, e.g. the beta distribution with $\alpha = \frac{\mu (\nu-1)}{\mu\nu-1}, \beta = \frac{(1 - \mu)(\nu-1)}{\mu\nu-1}$. The bounds from Iosif's answer are likely as good as it gets. $\endgroup$ – student Jan 11 at 11:37

Assume that $X\ge0$. Then, by Jensen's inequality for the convex function $x\mapsto1/x$, \begin{equation} E\frac1X\ge\frac1\mu, \end{equation} where $\mu:=EX$. Thus, we have the lower bound $\frac1\mu$ on $E\frac1X$; this bound is exact, as it is attained when $X$ is a constant.

On the other hand, it is easy to see that there is no finite upper bound on $E\frac1X$ in terms of $\mu$ only. Indeed, for any real $\mu>0$, let $P(X=0)=1/2=P(X=2\mu)$. Then $E\frac1X=\infty$ while $EX=\mu$. Even if we assume that $X>0$, still there is no finite upper bound on $E\frac1X$ in terms of $\mu$ only. Indeed, for any real $\mu>0$ and any $a\in(0,\mu)$, let $P(X=a)=1/2=P(X=2\mu-a)$. Then $EX=\mu$ while $E\frac1X=\frac1{2a}+\frac1{2(2\mu-a)}\to\infty$ as $a\downarrow0$.

If $X$ is known to be bounded away from $0$, so that $X\ge a$ for some real $a>0$, then trivially $1/X\le1/a$ and hence \begin{equation} E\frac1X\le \frac1a. \end{equation} The latter upper bound on $E\frac1X$ is exact here for each feasible value of $\mu=EX$, even though this bound does not even involve $\mu$. Indeed, take any real $\mu\ge a$. For any real $u>\mu$, let $P(X=u)=p=1-P(X=a)$, where $p:=\frac{\mu-a}{u-a}\in[0,1)$, so that $EX=\mu$. On the other hand, letting $u\to\infty$, we have $p\to0$ and hence $E\frac1X=\frac1a\,(1-p)+\frac1u\,p\to\frac1a$, so that the upper bound $\frac1a$ on $E\frac1X$ is attained in the limit.

Finally, if $X$ is known to be bounded away both from $0$ and $\infty$, so that $a\le X\le A$ for some positive real $a,A$ such that $a<A$, then, by the convexity of the function $x\mapsto1/x$, we have $\frac1X\le\frac{A+a-X}{Aa}$ and hence \begin{equation} E\frac1X\le\frac{A+a-\mu}{Aa}. \end{equation} The latter bound on $E\frac1X$ is again exact, as it is attained when $P(X=A)=p=1-P(X=a)$, where $p:=\frac{\mu-a}{A-a}\in[0,1]$, so that $EX=\mu$.

  • $\begingroup$ Thank you for your answer. Unfortunately my X is not bound away from zero - it runs between $0$ and $\sqrt{2}$ (included), although for practical purposes I can consider it strictly between $0$ and $\sqrt{2}$, however I still do not have a non-zero lower bound. $\endgroup$ – Silvia Jan 11 at 10:27
  • $\begingroup$ @Silvia : Then, as shown in my answer, there is no finite upper bound on $E\frac1X$ in terms of $\mu$ only. That is, to be able to have a finite upper bound, you need to know more than just $\mu$. $\endgroup$ – Iosif Pinelis Jan 11 at 14:06

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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