4
$\begingroup$

Is there any global constant $C$ such that $$C<\frac{\sum_{i=1}^{n}x_{i}\log x_{i}-x_{i}+(1-x_{i})\log(1-x_{i})}{\sum_{i=1}^{n}x_{i}}+\log(\sum_{i=1}^{n}x_{i})-\log(\sum_{i=1}^{n}x_{i}^{2})$$ for all vectors $x\in\mathbb{R}^n$ such that $0<x_i<1$ for all $i$? The main sticking point seems to be that the last term is not convex, so i can't just fix $\sum_i x_i $ and apply symmetry.

$\endgroup$
1
  • $\begingroup$ Have you studied the monoticity, that is the derivative, if you vary one $x_i$? $\endgroup$
    – hänsel
    Dec 22, 2017 at 11:05

2 Answers 2

4
$\begingroup$

Now I think that no, even if we remove the negative summands $-x_i+(1-x_i)\log(1-x_i)$. Note that our inequality becomes homogeneous, so we may forget that $x_i$ are less than 1. Choose $x_i=1+t_i$ so that $\sum t_i=0$, then the inequality rewrites as $$ \log\frac{\sum x_i^2}{\sum x_i}=\log\left(1+\frac{t_1^2+\dots+t_n^2}n\right)<-C+\frac1n\sum (1+t_i)\log(1+t_i). $$ Imagine that $t_1=n^{3/4}$ and other $t$'s are equal to $-t_1/(n-1)$. Then LHS is large while RHS is bounded.

$\endgroup$
4
  • $\begingroup$ The fraction $(\sum x_i\log x_i)/(\sum x_i)$ is not homogeneous. $\endgroup$
    – GH from MO
    Dec 22, 2017 at 13:50
  • 1
    $\begingroup$ Of course it is not, but the inequality at whole is: when we replace all $x_i$ to $A\cdot x_i$, RHS does not change. $\endgroup$ Dec 22, 2017 at 13:54
  • $\begingroup$ Good point, +1 from me! $\endgroup$
    – GH from MO
    Dec 22, 2017 at 14:00
  • $\begingroup$ Thanks @FedorPetrov, those are both very clear explanations. One follow-up, do you think it's possible to say anything about the necessary conditions for unboundness in terms of the "spread" of the $x_i$'s? Like, there has to be a small subset of $x_i$'s that's very large? $\endgroup$ Dec 22, 2017 at 21:25
4
$\begingroup$

Your expression is not bounded from below by an absolute constant (independent of $n$). To see this, consider the example $x_1=\dots=x_{n-1}=\epsilon$ and $x_n=1-\epsilon$. For this example, the left hand side equals $$ n\cdot\frac{\epsilon\log\epsilon+(1-\epsilon)\log(1-\epsilon)}{1+(n-2)\epsilon}-1+\log\bigl(1+(n-2)\epsilon\bigr)-\log\bigl(1-2\epsilon+n\epsilon^2\bigr).$$ For $\epsilon:=1/n$, the first term asymptotically $-(\log n)/2$, while the other three terms are bounded, so the expression tends to negative infinity.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.