Consider the generalized (power) mean of positive numbers $a_1, \dotsc, a_n$ $$M_p(a_1, \dotsc, a_n)=\left(\frac{a_1^p + \dotsb + a_n^p}{n}\right)^{1/p}\qquad p\in \mathbb{R}$$

where for $p=0$ we use the geometric mean. The generalized mean inequality says that $$ p < q \implies M_p(a_1, \dotsc, a_n) \leq M_q(a_1, \dotsc, a_n),$$

with equality holding iff $a_1 = \dotsb = a_n$. On Wikipedia it says that one can prove this by differentiating with respect to $p$ and using Jensen's inequality, and noting that $\partial_p M_p(a_1, \dotsc, a_n)>0$. When I do so, I get:

$$\partial_p \left(\left(\frac{a_1^p + \dotsb + a_n^p}{n}\right)^{1/p}\right)\\= \left(\frac{a_1^p + \dotsb + a_n^p}{n}\right)^{1/p} \partial _p \left( \frac1p \log (a_1^p + \dotsb + a_n^p) - \frac1p \log n \right),$$ which is positive iff $\partial _p \left( \frac1p \log (a_1^p + \dotsb + a_n^p) - \frac1p \log n \right)$ is positive.

$$\partial _p \left( \frac1p \log (a_1^p + \dotsb + a_n^p) - \frac1p \log n \right) \\= \left(\frac{-1}{p^2}\right) \log (a_1^p + \dotsb + a_n^p)+ \frac1p \frac{a_1^p \log a_1+ \dotsb + a_n^p \log a_n }{a_1^p + \dotsb + a_n^p}+\frac{1}{p^2}\log n.$$

I'm trying to show that this is positive. When I think about what Jensen's equality would tell me, I note that $$\frac{a_1^p\log a_1 + \dotsb + a_n^p \log a_n}{a_1^p + \dotsb + a_n^p} \\ \leq \log \left( \frac{a_1^p}{a_1^p + \dotsb + a_n^p}a_1 + \dotsb + \frac{a_n^p}{a_1^p + \dotsb + a_n^p}a_n \right) \\ = \log \left( \frac{a_1^{p+1} + \dotsb + a_n^{p+1}}{a_1^p + \dotsb + a_n^p} \right),$$ or we could try to bring the factor $\frac1p$ of the term in question into the mix and make a similar estimate.

Anyone have an idea?

  • 1
    $\begingroup$ I don't see why one would differentiate at all, Jensen's inequality directly yields the result. $\endgroup$ – Daniel Fischer Jan 22 '14 at 20:08
  • $\begingroup$ @DanielFischer Wanting to treat $p>q>0$, $p>0>q$ and $0>p>q$ simultaneously would be one reason, although the computation is indeed ugly. $\endgroup$ – Alex Shpilkin Jan 31 at 1:46

I agree with Daniel Fischer's comment: it's easier to use Jensen's directly (as here, of which the present question is a special case, since averages are a special case of integrals).

But maybe you really want to prove the derivative wrt $p$ is positive; after all, this gives more information than the function being strictly increasing. Then do the following: normalize $a_i$ (multiplying them by the same positive constant) so that $\sum a_i^p =1$. The scary inequality $$ \left(\frac{-1}{p^2}\right) \log (a_1^p + \dotsb + a_n^p)+ \frac1p \frac{a_1^p \log a_1+ \dotsb + a_n^p \log a_n }{a_1^p + \dotsb + a_n^p}+\frac{1}{p^2}\log n\ge 0$$ suddenly simplifies to $$ \frac1p (a_1^p \log a_1+ \dotsb + a_n^p \log a_n )+\frac{1}{p^2}\log n\ge 0$$ Writing $b_i=a_i^p$ simplifies it further: $$ \sum b_i \log b_i \ge \log (1/n) \tag{1}$$ with equality iff all $b_i$ are equal. You may have seen (1) in the context of entropy. Jensen's inequality yields (1) because the function $x\mapsto x\log x$ is convex on $[0,1]$.

x log x


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.