Is anything known about the behavior of the function $$f(x)=\prod_{i=1}^n x_i^{x_i}$$ on the standard simplex, i.e. the set $\{x\in\mathbb{R}^n:\sum_{i=1}^n x_i=1, x_i\geq0\}$? I ask because I have done a lot of numerical experiments that suggest that $$1\leq\left(\prod_{i=1}^{n}x_{i}^{x_{i}}\right)\left(\sum_{i=1}^{n}x_{i}^{1/2}\right)^{2}\leq 2$$ for all $n$ and all $x$ on the standard simplex, and I have no clue how to prove this. Of course, one could also re-write this more compactly as $$1\leq\left(\prod_{i=1}^{n}x_{i}^{x_{i}}\right)\|x\|_{1/2}\leq 2$$but this still doesn't seem particularly helpful to me.

  • 1
    $\begingroup$ Is this the sort of idel curiosity that lead to the discovery of so many things, or a question motivated by a "real problem"? (I am asking out of idle curiosity because I can't image how and why anyone would think of these inequalities.) $\endgroup$ – Andrej Bauer Apr 12 '16 at 8:27
  • $\begingroup$ @AndrejBauer Definitely a real problem, but it's a long story. I am hoping it is true because it allows me to describe the probability distribution of a sum of random variables whose cdfs have a certain form. $\endgroup$ – Jennifer Gao Apr 12 '16 at 8:32
  • $\begingroup$ You assume $0^0=1$, right? $\endgroup$ – joro Apr 12 '16 at 9:22
  • 1
    $\begingroup$ If you use $\log\sum_i \sqrt{x_i} = \log\sum_i x_i \frac{1}{\sqrt{x_i}}\geq \sum_i x_i \log(\frac{1}{\sqrt{x_i}})=-2\sum_ix_i\log x_i$ (where the inequality is Jensen's) you get $\sum_i \sqrt{x_i}\geq \sqrt{\prod_i x_i^{-2x_i}} $. Which is enough for the lower bound, because $\prod_i x_i^{x_i}<1$. $\endgroup$ – martin cripps Apr 12 '16 at 9:27
  • $\begingroup$ Sorry too many square roots in the last inequality - it should read $\sum_i\sqrt{x_i}\geq \prod_ix_i^{-2x_i}$! $\endgroup$ – martin cripps Apr 12 '16 at 9:46

The upper bound is not correct. E.g., let $n=101$, $x_1=\dots=x_{100}=1/1000$, $x_{101}=9/10$. Then $$\left(\prod_{i=1}^{n}x_{i}^{x_{i}}\right)\left(\sum_{i=1}^{n}x_{i}^{1/2}\right)^{2}>7>2.$$

More generally, take any $a\in(0,1)$ and any natural $k$, and then let $n=k+1$, $x_1=\dots=x_k=a/k$, $x_{k+1}=1-a$. Then $$\ln\left[\left(\prod_{i=1}^{n}x_{i}^{x_{i}}\right)\left(\sum_{i=1}^{n}x_{i}^{1/2}\right)^{2}\right]=a\ln\frac ak+(1-a)\ln(1-a)+2\ln(\sqrt{ak}+\sqrt{1-a})\to\infty$$ as $k\to\infty$.

| cite | improve this answer | |
  • 2
    $\begingroup$ Thanks, you taught me a valuable lesson about sampling from the boundary of a simplex :) $\endgroup$ – Jennifer Gao Apr 12 '16 at 16:37
  • $\begingroup$ I am glad this was useful. $\endgroup$ – Iosif Pinelis Apr 12 '16 at 16:58

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.