Let $X$ be a random variable.
Prove that $g:(0,\infty)\rightarrow[0,\infty]$ which is defined by $g(t)=(E|X|^t)^{\frac 1t}$ ($E$ marks the expected value) is monotonic.

I tried experimenting with Markov's inequality but I had a hard time dealing with it since $X$ is not in a specific class of random variables.

  • $\begingroup$ Do you know Hölder's inequality? en.wikipedia.org/wiki/H%C3%B6lder%27s_inequality $\endgroup$
    – sharpe
    May 19 '15 at 14:28
  • $\begingroup$ Assuming the inequality I see why it is true. Is there a simple proof of the inequality for this specific case? $\endgroup$ May 19 '15 at 14:58
  • $\begingroup$ On the other hand, I am familiar with Jensen's inequality which can be used to prove that g is monotonic. $\endgroup$ May 19 '15 at 15:32

Fix $s\lt t$ and define $p:=t/s \gt 1$. The map $x\mapsto |x|^p$, is convex, hence for each non-negative random variable $Y$, we have $$\left(\mathbb E[Y]\right)^p\leqslant \mathbb E[Y^p].$$
Using this with $Y=|X|^s$, we derive that $$\left(\mathbb E[|X|^s]\right)^{t/s} \leqslant \mathbb E[|X|^{s\cdot t/s} ],$$ from which the wanted inequality follows.


This follows from the general Lyapunov inequality which states that $$\nu_a^{b-c}\nu_c^{a-b}\ge\nu_b^{a-c}\quad\forall a>b>c\ge0,\quad\text{where}\quad \nu_t=\mathbb{E}\left(|X|^t\right)$$

Another form of this inequality is proven here using Hölder's inequality.

Choosing $c=0$ in the generalised inequality we have,

$\nu_a^b>\nu_b^a\quad\forall a>b>0$

$\implies \nu_a^{1/a}>\nu_b^{1/b}\quad\forall a>b>0$

$\implies \nu_t^{1/t}$ is non-decreasing in $t\quad\forall t>0$.

Note that proving the generalised Lyapunov inequality is equivalent to proving $$\alpha g(a)+(1-\alpha)g(c)\ge g\left(\alpha a+(1-\alpha)c\right),$$ where $\alpha=\frac{b-c}{a-c}$ and $g(t)=\ln(\nu_t)$.

Then showing $g$ is convex (by proving $g''\ge0$) proves the inequality.


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.