6
$\begingroup$

This is essentially what Exercise 5.4 in Boucheron, Lugosi, Massart Concentration Inequalities boils down to: For real $a,b$ and $0<p<1$,

\begin{align*} &pa^2\log( \frac{a^2}{b^2+pa^2-pb^2}) + (1-p)b^2\log(\frac{b^2}{b^2+pa^2-pb^2}) \\ &\le \frac{p(1-p)(a-b)^2}{1-2p}\log\frac{1-p}p . \end{align*} This is supposed to be provable by elementary means and in finite time. Any ideas?

$\endgroup$
4
  • $\begingroup$ Scale things so that 1=b^2 + pa^2 - pb^2 ? Gerhard "Or Make Something Else One" Paseman, 2017.03.25. $\endgroup$ Mar 26, 2017 at 0:53
  • $\begingroup$ I don't see how that's possible, as it requires that $a=\sqrt{(1-b^2+pb^2)/b}$, which need not necessarily hold. $\endgroup$ Mar 26, 2017 at 8:36
  • 2
    $\begingroup$ Replace a and b by tA and tB, and pick t so that the denominator of the original log term is equal to t ^2. Then cancel the t^2 terms everywhere. Gerhard "Make Me One With Everything" Paseman, 1017.03.26. $\endgroup$ Mar 26, 2017 at 14:55
  • 1
    $\begingroup$ A couple other thoughts: interchanging $a,b$ is equivalent to interchanging $p, 1-p$ so it is sufficient to assume $0 < p < 1/2$. Also, both sides approach 0 as $p \to 0$, so it could be helpful to compare their derivatives with respect to $p$. After making Gerhard's transformation, the left side is linear in $p$. $\endgroup$ Mar 26, 2017 at 15:40

3 Answers 3

1
$\begingroup$

By scaling if necessary, we may assume without loss of generality that $a^2p + b^2\bar p = 1$ Substituting $u = a^2p$, we can rewrite the 1-dimensional inequality as \begin{align*} f(u) := u\log \frac{u}{p} + \bar u \log \frac{\bar u}{\bar p} - c(p) (\sqrt{\bar u p} - \sqrt{u \bar p})^2 \le 0. \end{align*} We calculate the first two derivatives of $f(u)$. They are \begin{align*} f'(u) &= \log \frac{u}{\bar u} - \log \frac{p}{\bar p} - c(p)(1-2p) + c(p)(1-2u)\sqrt{\frac{p\bar p}{u\bar u}} \end{align*} and \begin{align*} f''(u) &= \frac{1}{u\bar u} - c(p)\frac{\sqrt{p\bar p}}{2(u\bar u)^{3/2}}. \end{align*} We first claim that $f''$ goes to $+\infty$ at 0 and 1, and precisely 2 zeros between 0 and 1 (they may be repeated). The first part is easy to check. To check the number of zeros, we need $2\sqrt{u\bar u} = c(p) \sqrt{p\bar p}$. As $u$ ranges from $[0,1]$, the left hand side ranges from $[0, 1]$ as well. It is not difficult to check that the right hand side is $\le 1$ (for example, by differentiating it and checking that the derivative is 0 exactly at $p=1/2$, which is a maxima where the value is 1).

Having established this, we verify that $f'$ has the following zeros: $p, 1/2, \bar p$, and moreover, $f'$ is $+\infty$ at $0$ and $1$. Since $f''$ has exactly two zeros, $f'$ cannot have any more zeros. Recall that the points where $f'$ is 0 is precisely the list of maxima or minima of $f$. Moreover, the shape of $f''$ ensures that the local maxima are at $p$ and $\bar p$ whereas the local minimum is at $1/2$. Evaluating $f$ at the local maxima, we see that \begin{align*} f(p) = 0, \end{align*} and \begin{align*} f(\bar p) = \bar p \log \frac{\bar p}{p} + p \log \frac{p}{\bar p} - (\bar p - p) \log \frac{\bar p}{p} = 0. \end{align*} Thus, the maximum of $f$ is 0, and the result is proved.

$\endgroup$
2
$\begingroup$

Let $X : \Omega \to \{0,1\}$ be a Bernoulli random variable wich takes value $0$ with probability $p$ and value $1$ with probability $q$, $p+q=1, p,q \geq 0$. Your inequality is the claim that $$ \mathbb{E} f^{2}(X) \ln f^{2}(x) - \mathbb{E}f^{2}(X) \ln \mathbb{E} f^{2}(X) \leq \frac{\ln p - \ln q}{p -q}\mathbb{E} |Df(X)|^{2} $$ for all $f : \{-1,1\} \to \mathbb{R}$ where in this simple case
$$ |Df(x)|^{2}=|f(1)-f(0)|^{2}. $$

For the proof of this and its consequences (Gaussian measure $p=q=2$, Poisson measure arbitrary $p,q$) see the paper of S. Bobkov and M. Ledoux, On Modified Logarithmic Sobolev Inequalities for Bernoulli and Poisson Measures

These kind of two-point inequalities are quite subtle and sometimes hard to prove, but if applied properly they can give interesting and unexpected applications for Gaussian inequalities and not only.

There is one similar two-point inequality related to complex hypercontractivity on the discrete cube which is still open and if somebody will be interested I can mention it.

UPDATE:

Let $1< p \leq q < \infty$, and $|z|\leq 1, z \in \mathbb{C}$. The following conditions are equivalent:

(i) For all $a,b \in \mathbb{C}$ we have $$ \left( \frac{|a+bz|^{q}+|a-bz|^{q}}{2} \right)^{1/q} \leq \left(\frac{|a+b|^{p}+|a-b|^{p}}{2} \right)^{1/p} $$

(ii) for all $w \in \mathbb{C}$ we have $$ (q-2)(\Re\; wz)^{2} + |wz|^{2} \leq (p-2)(\Re w)^{2}+ |w|^{2} $$

Remark: (i) $\Rightarrow$ (ii), follows by Taylor's formula (take $a=1$ and $b \to 0$). The implication (i) $\Rightarrow$ (ii) is open for $3/2 <p\leq q<2$ (and its dual part $2<p\leq q <3$). For the remaining part of exponents this is the result of F. Weissler Two-point inequalities, the Hermite semigroup and the Gauss-Weierstrass semigroup

By the way, the particular case $z=i\sqrt{p-1}$, $1<p\leq 2$, and $q=p/(p-1)$ gives (after proper application of Minkowski's inequality and CLT) Hausdorff--Young inequality with sharp constants proved by W. Beckner.

$\endgroup$
4
  • $\begingroup$ Yes, Bobkov's proof was given in the proof of Theorem 2.2.8 here: pdfs.semanticscholar.org/a3a2/… $\endgroup$ Mar 26, 2017 at 19:39
  • $\begingroup$ The proof given here seems simpler (or at least shorter), but that's a judgement call. My original goal of derive this inequality "for free" as an immediate consequence of Kearns-Saul has still not been achieved. $\endgroup$ Mar 26, 2017 at 19:40
  • $\begingroup$ ... oh and by all means, please do share the conjectured inequality! $\endgroup$ Mar 26, 2017 at 19:41
  • $\begingroup$ Paata, I am interested $\endgroup$ Mar 26, 2017 at 20:31
1
$\begingroup$

After Gerhard Paseman's and Nate Eldredge's suggestions, the problem is reduced to showing that $$ f_{a,b}(p)= %A^2 p Log[A^2/1] p a^2 \log(a^2) + % (1-p) b^2\log(b^2) - \frac{p(1-p)(a-b)^2}{1-2p}\log\frac{1-p}p \le0 $$ for $a,b>0$ and $0<p<1/2$. The claim is obviously true for $p\to0$ and less obviously true for $p\to1/2$ (this is the content of Theorem 5.1 in the Boucheron, Lugosi, Massart book). Furthermore, a straightforward calculation yields $$ f''_{a,b}(p) = \frac{ (a - b)^2 (1 - 2 p - 2 (1 - p) p \log[(1-p)/p)] }{ (1 - p) p (1 - 2 p)^3 }, $$ which we claim is nonnegative for $0<p<1/2$. Once the latter claim is established, we have that $f_{a,b}(p)\le0$ for $p=0,1/2$ and is convex on $[0,1/2]$, so it is nonpositive on the whole interval.

Proving the nonnegativity of the second derivative is quite straightforward, if somewhat tedious. The denominator in the expression for $f''_{a,b}$ is obviously positive, so it suffices to consider just the numerator (and disregard the $(a-b)^2$ factor).

The expression in question, $1 - 2 p - 2 (1 - p) p \log[(1-p)/p)]$, is $1$ at $p=0$, is $0$ at $p=1/2$, and has derivative $-2 (1 - 2 p) \log(1/p - 1)$, which is clearly negative.

$\endgroup$
4
  • $\begingroup$ Thanks for the help, everyone! As a methodological point, note that the trick of differentiating w.r.t. $p$ was employed in this proof of the Kearns-Saul inequality, Theorem 3.2 in projecteuclid.org/euclid.ecp/1465315542 $\endgroup$ Mar 26, 2017 at 16:42
  • $\begingroup$ I actually spent some time trying to obtain the present inequality directly from Kearns-Saul, but couldn't see it. Would love for someone to point out a way to do it, if there is one! $\endgroup$ Mar 26, 2017 at 16:42
  • 1
    $\begingroup$ I think this solution is not correct. One may scale $a$ and $b$ for a fixed $p$, but as $p$ changes the scaling stops working. In particular if we assume $pa^2+(1-p)b^2 = 1$, then as $p \to 0$ the claim in the above proof need not be true if $b > 1$, for instance. $\endgroup$
    – VSJ
    Nov 22, 2021 at 4:28
  • $\begingroup$ Good catch! @VSJ $\endgroup$ Nov 22, 2021 at 10:15

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.