# Proving a messy inequality

EDIT:
After much work I was able to reduce the inequality to a single variable function which I need to show is non-positive. That function is (for $0\leq p\leq\frac{1}{2}$) $$\frac{p^2(\log(p))^2 - (1-p)^2(\log(1-p))^2}{1-2p} + \log(p)\log(1-p) + p\log(p)+(1-p)\log(1-p) \leq 0$$ This looks tractable, but is quite hard to prove since neither the function nor any of its derivatives are convex/concave/monotonic. Its graph looks like

.

Any inputs welcome!

END EDIT

I'm trying to prove this messy inequality I got while working on an information theoretic problem (more details below).

Let $0 \leq p,q \leq \frac{1}{2}$.

Let $H(x) = -x\log(x) - (1-x)\log(1-x)$ be the binary entropy function.

Let $p\star q = p(1-q) +q(1-p)$ (think of this as addition of two Bernoulli random variables in $\mathbb{Z}_2$)

The inequality is: $$-\frac{(1-2q)}{(p \star q) (1- p \star q)\log\left(\frac{1-p}{p}\right)} + \frac{\log\left(\frac{1- p \star q}{p \star q}\right)}{p(1-p)\left(\log\left(\frac{1-p}{p}\right)\right)^2} - \frac{H(q)(1-2p)}{H(p)(p \star q) (1- p \star q)\log\left(\frac{1-q}{q}\right)} \leq 0$$

Some Remarks:

If we write this as $-A+B-C \leq 0$ with $A,B,C$ being the corresponding quantities above. I can show that $-A +B \geq 0$, so the $-C$ term subtracts enough to make the expression negative. The plot of $-A+B-C$ as a function of $p,q$ looks like

.

It appears to decrease in $q$ for a fixed $p$, starting at $0$ and going to $-\infty$ as $q$ goes to $\frac{1}{2}$.

Origin:
This inequality came from looking at a function $f(x,y) = H(H^{-1}(x)\star H^{-1}(y))$ where $f: [0,1]\times [0,1] \to [0,1]$. This function occasionally pops up in information theory, like when studying the binary symmetric channel or an inequality called Mrs. Gerber's Lemma. The inequality I have implies that $f(x,y)$ is concave along lines through the origin (i.e $f(x,\theta x)$ is concave in $x$). What I've written is a simplified version $f''$ along a line through the origin.

I'm at a complete loss as to how to attack this. All the plots I have show this is true, but a plot is not a proof. Any suggestions welcome!

-
If all else fails: with much work, a plot can be turned into a proof. Sample points at sufficiently fine grid, which given a crude upper bound on the derivative implies inequality nearby whenever the inequality is true with an epsilon to spare. Then check neighborhoods of the set where the equality holds. For those, compute the derivates around them, show that they point in the right direction, and bound the second derivative to show that one gets a control for big enough neighborhood of the equality case. This is a pain, but general Tarski's exponential function problem is still open. – Boris Bukh Apr 6 '12 at 11:42
@boris-bukh :Thanks for the suggestion. Do you happen to know any such sample proofs which I can read and get a better idea of how these techniques work? – VSJ Apr 6 '12 at 18:06
This can even be done in floating point to full rigor if interval arithmetic is used. But that would be a "desperate last hope" type of proof. – Brendan McKay Apr 7 '12 at 1:15
@Brendan : Ah, in that case I'll wait a bit. To be honest I'm amazed to hear that floating point arithmetic can be used to rigourously prove anything at all! I would really appreciate any references. This also raises the question if these brute force methods work, why doesn't everyone use them all the time? – VSJ Apr 7 '12 at 1:40
My comment on this answer mathoverflow.net/questions/17189/… is an example of what Boris is talking about. – David Speyer Apr 11 '12 at 21:37

I think I managed to prove the entire inequality analytically. The whole proof is a bit long to post here (about 7 pages) and involves ugly looking expressions. I'll outline the general strategy I used:

1. Start with the original inequality $$-\frac{(1-2q)}{(p \star q) (1- p \star q)\log\left(\frac{1-p}{p}\right)} + \frac{\log\left(\frac{1- p \star q}{p \star q}\right)}{p(1-p)\left(\log\left(\frac{1-p}{p}\right)\right)^2} - \frac{H(q)(1-2p)}{H(p)(p \star q) (1- p \star q)\log\left(\frac{1-q}{q}\right)} \leq 0$$ and throw all the $p\star q$ terms on one side to get $$(p \star q)(1- p\star q)\log\left(\frac{1-p \star q}{p \star q}\right) \leq(1-2q)p(1-p)\log\left(\frac{1-p}{p}\right) + \frac{p(1-p)(1-2p)\left(\log\left(\frac{1-p}{p}\right)\right)^2 H(q)}{H(p)\log\left(\frac{1-q}{q}\right)}$$
2. Now keep $p \star q$ fixed $= k$ say, and try minimise the right hand side. Notice that when $p = k, q = 0$ we have equality. This might lead us to hope that the right hand side is a decreasing function of $p$, and when $p$ becomes $k$ we have equality.

3. Now consider the partial derivative of the RHS in $p$, and try to show that this is negative. By some stroke of luck, it turns out that if we look at $\frac{\partial RHS}{\partial p}$ for a fixed $p$ and vary $k$, the value of $k$ which maximises $\frac{\partial RHS}{\partial p}$ is $k = p$. So its sufficient to prove $\frac{\partial RHS}{\partial p}|_{(p,p)} \leq 0$

4. $\frac{\partial RHS}{\partial p}|_{(p,p)}$ when evaluated gives me the second inequality (the one I added in the EDIT). Thus it suffices to show $$p^2(\log(p))^2 - (1-p)^2(\log(1-p))^2 + (1-2p)\left(\log(p)\log(1-p) + p\log(p)+(1-p)\log(1-p)\right) \leq 0$$

5. Let me call the above expression $F(p)$. I showed that if $F(p) > 0$ somewhere, then $F'''(p)$ would have atleast $2$ zeros in $[0,0.5]$. (This is easy to show.) Explicitly computing $F'''(p)$ shows that it is $-\infty$ at $0$ and strictly positive at $0.5$. So, if we happened to show that $F'''$ is concave, it would have exactly one zero, and thus contradict the result that it should have $2$, and thus prove the result.

6. Now I do something truly horrible. I explicitly compute the 5-th derivative and show that it is negative, thus $F'''$ is indeed concave.

7. The inequality $F'''''(p) \leq 0$ can be somewhat simplified to get $$P_1(p)\log(p) + P_2(p)\log(1-p) + P_3(p) \leq 0$$ where $P_1, P_2,P_3$ are some $7$-th degree polynomials.

8. I now use Sturm's theorem to show that over the range $[0,0.5]$ $P_1,P_2 \geq 0$. This, too, was quite horrible.

9. Now I use a polynomial approximation to $\log$, and use the positivity of $P_1,P_2$ to conclude $$P_1(p)\log(p) + P_2(p)\log(1-p) + P_3(p)\leq P_1(p)\left( -(1-p) - \frac{(1-p)^2}{2}\right) + P_2(p)\left(-p - \frac{p^2}{2}\right) + P_3(p)$$

10. By another stroke of luck, the RHS here factorises to give $$-24(1-p)^2 (p-0.5)^2 p^2 (p^2-p+\frac{7}{3})$$ which is clearly seen to be negative, and thus by step $9$ I can now say that $F'''''$ is negative, and $F'''$ is concave, and it has only one root in $[0,0.5]$ and thus $F$ has to be non-positive everywhere in $[0,0.5]$.

In retrospect, I guess using numerical techniques (as advised in the comments section) would have been more prudent.

-

Just take derivatives with respect to p and q. You'll find that they're never zero, and hence the maximum is on the boundary. The signs of the derivatives on the boundary will tell you where the maximum is, and this maximum with most likely be zero.

-
The derivatives in $p$ and $q$ give me really complicated expressions so that their being non-zero is not at all obvious. You are right in that the maximum of the function along the boundary is zero, this can be evaluated. – VSJ Apr 6 '12 at 18:03