I have to prove that, when bounding tail probabilities of a nonnegative random variable, the moments method is always better than the classical Chernoff method. In mathematical language, I have to prove that for every $t>0$, $$\inf_{n \in \mathbb{N}} \frac{\mathbf{E}[X^n]}{t^n} \leqslant \inf_{s>0} \frac{\mathbf{E}[e^{sX}]}{e^{st}}$$ This result seems to be classical and can apparently be found in a paper by Phillips and Nelsen (1995) available here for the modic sum of 14$. So instead of buying it, I tried to work it out, but the proof I have seems to be rather short for a 5-pages paper... Can somebody check my proof ?

Let $s>0$. Then using power series expansion and monotone convergence theorem, we have $\mathbf{E}[e^{sX}] = \mathbf{E}\left[\sum_{n \geqslant 0}\frac{s^n X^n }{n!} \right] = \sum_n \frac{s^n \mathbf{E}[X^n]}{n!} $. Therefore,

$$ \frac{\mathbf{E}[e^{sX}]}{e^{st}} = \frac{\sum_n \frac{s^n \mathbf{E}[X^n]}{n!}}{\sum_n \frac{s^n tX^n}{n!}} = \lim_{n \to \infty} \frac{1 + s \mathbf{E}[X] + ... + \frac{s^n \mathbf{E}[X^n]}{n!}}{1 + st + ... + \frac{s^n t^n }{n!}}$$

A classical inequality (often called Cauchy's third inequality) states that for every nonnegative numbers $a_1,...,a_n,b_1,...,b_n$ one has $$ \frac{a_1 + ... + a_n}{b_1 + ... + b_n} \geqslant \min_{k \leqslant n} \frac{a_k}{b_k}$$

Therefore, $$ \frac{1 + s \mathbf{E}[X] + ... + \frac{s^n \mathbf{E}[X^n]}{n!}}{1 + st + ... + \frac{s^n t^n }{n!}} \geq \min_{k \leq n} \frac{s^k \mathbf{E}[X^k] k!}{k! s^k t^k} = \min_{k \leq n}\frac{\mathbf{E}[X^k]}{t^k}$$

And finally, taking the limit we have, for every $s>0$ : $$ \frac{\mathbf{E}[e^{sX}]}{e^{st}} \geq \inf_{n \in \mathbb{N}}\frac{\mathbf{E}[X^n ]}{n!}$$ which is what we wanted to prove.

  • $\begingroup$ Sure, this works. Note that the paper by Phillips and Nelsen (1995) might not contain just the proof. $\endgroup$
    – Did
    May 10, 2015 at 10:06
  • $\begingroup$ Yes, but having no access to it I couldn't see if their proof was 5 pages long or 1 pages long with extra results. $\endgroup$ May 10, 2015 at 10:09

1 Answer 1


As mentioned in comments by Did, this method works and is the method of proof used by the authors.

A open link to the paper is here. The

The paper contains several common distributions (gamma, exponential, Poisson, etc) as examples and is not entirely proof.


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.