I apologize for the poor title.

Let $p\in 2\mathbb N$ and $a_1,\ldots , a_n\geq 0$.

Show that $$\left (\sum a_j\right )^p \leq n^{p-1}\sum a_j^p $$

So, firstly, by rearranging the expression, it's sufficient to show that $$\left (\sum\frac{a_j}{n}\right )^p \leq \sum\frac{a_j^p}{n}. $$ It's known from theory of integrals that for any convex function $g:\mathbb R\to\mathbb R$ $$g(EX)\leq E g(X) \quad\mbox{the Jensen inequality}$$ Suppose $(\Omega,\mathcal F,\mathbb P)$ is a probability space with $X_1,\ldots, X_n$ some $\mathcal F$-measurable functions such that $EX_j \equiv a_j$. Since $p\in2\mathbb N$, among others, the mapping $x\mapsto x^p$ is convex. Define $$X := \sum \frac{X_j}{n}, $$ The inequality yields $$g(EX) = g\left (\int\sum\frac{X_j}{n}\mbox{d}\mathbb P\right ) = \left (\sum\frac{a_j}{n}\right )^p\leq \int\left (\sum\frac{X_j}{n}\right )^p\mbox{d}\mathbb P = Eg(X) $$ ..and I don't see any way to make progress or I'm missing something very trivial or is the choice of $g$ or $X$ a poor one? Are there any suggestions?


Alternatively, you can also use Holder's inequality. It states that, if $q=\frac{p}{p-1}$, then

\begin{align*} \|f g\|_{1} &\leq \|f\|_{p}\|g\|_{q} \\ \|f g \|_{1}^{p} &\leq \|f\|_{p}^{p} \|g\|_{q}^{p} \end{align*}

By comparing the expressions in the problem, we note that we want $f_{i} \cdot g_{i} = a_{i}$. In this case,

\begin{align*} \|f g\|_{1}^{p} &= \left(\sum_{i=1}^{n}{|a_i|}\right)^{p} \\ &= \left(\sum_{i=1}^{n}{a_i}\right)^{p} & a_i \geq 0 \end{align*}

Next, if we take $g_{i}=1$, obtain \begin{align*} \|g\|_{q}^{p} &= \left(\sum_{i=1}^{n}{1^{q}}\right)^{\frac{p}{q}} \\ &= n^{\frac{p}{q}} = n^{p-1} & q = \frac{p}{p-1} \end{align*}

Finally, since $f_ig_i=a_i$ and $g_i=1$, obtain $f_i=a_i$. That is, \begin{align*} \|f\|_{p}^{p} &= \sum_{i=1}^{n}{a_i^p} \end{align*} Finally, plugging in the expressions, into Holder's inequality \begin{align*} \|f g \|_{1}^{p} &\leq \|f\|_{p}^{p} \|g\|_{q}^{p} \\ \left(\sum_{i=1}^{n}{a_i}\right)^{p} &\leq n^{p-1} \sum_{i=1}^{n}{a_i^p} \end{align*}

  • $\begingroup$ Oh wow, I like this one even more :) $\endgroup$ – Alvin Lepik Nov 7 '17 at 20:28

Although your line of reasoning is good, the choice of $X$ is poor. A common trick to transform empirical means into expectations is to take the empirical distribution. For this, let $\Omega=\{a_1,\ldots,a_{n}\}$ be such that $P(\{a_i\})=n^{-1}$ and $X(\omega)=\omega$. We say that $X$ follows the empirical distribution because it takes with probability $n^{-1}$ each value in the sample.

Next, let $g(x)=x^{ p}$, for $p \geq 1$. It follows from your reasoning that

\begin{align*} \left(\sum_{i=1}^{n}{a_i \cdot n^{-1}}\right)^{p} &= g(E[X]) \\ &\leq E[g(X)] = \sum_{i=1}^{n}{a_{i}^{p} \cdot n^{-1}} \end{align*}

  • $\begingroup$ PS: There seems to be a side case when $p=0$. You can solve this one by hand, since it simplifies to $1 \leq 1$. $\endgroup$ – madprob Nov 7 '17 at 19:59
  • 1
    $\begingroup$ Yes, I understand that. I didn't consider $E[g(X)]^p \leq E[g(X)^p]$, though. Thanks for your help! $\endgroup$ – Alvin Lepik Nov 7 '17 at 20:01
  • $\begingroup$ Ops, that was a typo! I fixed now. $\endgroup$ – madprob Nov 7 '17 at 20:08

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.