Can anyone offer some guidance on proving the following inequality? Define $\Lambda_1(a)=-a\log a$ and $\Lambda_2(a,b)=-(a+b)\log(a+b)$. Then if $a$, $b$, $c$, and $d$ are non-negative numbers summing to one, the following holds: \begin{align} \Lambda_2(a,b)+\Lambda_2(b,c)+\Lambda_2(c,d)+\Lambda_2(d,a)\geq \Lambda_1(a)+\Lambda_1(b)+\Lambda_1(c)+\Lambda_1(d). \end{align} I've tested a bunch of cases in Mathematica, so I'm pretty certain it's true. The concavity of $\Lambda_1$ gives an upper bound on the left-hand-side, so that doesn't seem to be the right approach. It's also straightforward to show that equality holds if $a=d$ and $b=c$, but I don't think that has much to do with the general case. I assume this would follow quickly from the right log inequality, so even just the name of such an inequality would be helpful.

  • $\begingroup$ Are $a,b,c,d$ in $[0,1]$? $\endgroup$
    – Alex R.
    Jul 21 '15 at 0:53
  • $\begingroup$ Yes, they all are. $\endgroup$
    – Jeffrey
    Jul 21 '15 at 1:10
  • $\begingroup$ Started with $\log(a^2+b^2+c^2+d^2) \ge a \log(a) + b \log(b) + c \log(c) + d \log(d)$ but didn't get anywhere. Also perhaps simplify notation by $\Lambda_2(a,b)=\Lambda_1(a+b)$ $\endgroup$
    – karakfa
    Jul 21 '15 at 1:59

By weighted AM/GM, \begin{align*} &\left(\frac{(a+b)(d+a)}{a}\right)^a \left(\frac{(b+c)(a+b)}{b}\right)^b \left(\frac{(c+d)(b+c)}{c}\right)^c \left(\frac{(d+a)(c+d)}{d}\right)^d \\ &\le a\cdot\frac{(a+b)(d+a)}{a} +b\cdot\frac{(b+c)(a+b)}{b} +c\cdot\frac{(c+d)(b+c)}{c} +d\cdot\frac{(d+a)(c+d)}{d} \\ &= (a+b)(d+a) +(b+c)(a+b) +(c+d)(b+c) +(d+a)(c+d) \\ &= \big((a+b)+(c+d)\big) \big((d+a)+(b+c)\big) \\ &= 1 \end{align*} Rearranging, $$ \frac1{a^a b^b c^c d^d} \le \frac1{(a+b)^{a+b} (b+c)^{b+c} (c+d)^{c+d} (d+a)^{d+a}} $$ Taking logs yields the desired inequality.

  • $\begingroup$ Very nice application +1! $\endgroup$
    – Macavity
    Jul 21 '15 at 3:31

Since $a,b,c,d$ are nonnegative and sum to 1, consider them as probabilities in some sample space with outcomes $A,B,C,D$ respectively having probabilities $a,b,c,d$.

Without loss of generality, let $U$ be the event $A\cup B$, and $V$ be the event $B\cup C$. Let $H(X)=\sum_{i=1}^m -p_i \log p_i$ be the standard entropy functional for a discrete random variable $X$ with p.m.f. $\{p_1,p_2,...p_m\}$. Then, the LHS of the original problem is $H(\mathbf{1}_U)+H(\mathbf{1}_V)$ while the RHS is $H(\mathbf{1}_U,\mathbf{1}_V)$ according to standard notation.

From elementary information theory (via Jensen's inequality primarily) we have $H(\mathbf{1}_U)+H(\mathbf{1}_V) = H(\mathbf{1}_U,\mathbf{1}_V) + I(\mathbf{1}_U;\mathbf{1}_V)$ where $I(\mathbf{1}_U;\mathbf{1}_V)\ge 0$ is the mutual information between $\mathbf{1}_U$ and $\mathbf{1}_V$. The inequality in the original question follows.


Just a trial...

From the definitions follow

$$ \Lambda_2(a,b) = \Lambda_1(a+b). $$

So need to show that

$$ \log\big( a^a b^b c^c d^d \big) - \log\big( (a+b)^{a+b} (b+c)^{b+c}(c+d)^{c+d}(d+a)^{d+a} \big) \ge 0. $$

First we rewrite this as $$ (abcd) \log\big( abcd \big) - (a+b)(b+c)(c+d)(d+a) \log\big( (a+b)(b+c)(c+d)(d+a) \big) \ge 0. $$

Let us write $$ p = abcd, \hspace{2em} q = (a+b)(b+c)(c+d)(d+a). $$

We have $$ q = (a+b)(b+c)(c+d)(d+a) < (a+b+c+d)^4 = p^4, $$ but also $$ p < q. $$

Due to the summation to one, we have $$ a+b+c+d = 1 \Rightarrow p = abcd \le 4^{-4}. $$

Note that $x \log(x)$ is decreasing for $0 \le x < e^{-1}$.

So clearly we have $$ 0 < p < q < 4^{-1} < e^{-1}. $$

And consequently we obtain $$ p \log\big(p\big) \ge q \log\big( q \big). $$


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.