8
$\begingroup$

Given two probability distributions $P$ and $Q$ defined over a finite set $\mathcal{X}$, one can define the KL divergence between $P$ and $Q$ as $$D(P||Q):=\sum_{x\in \mathcal{X}}P(x)\log\frac{P(x)}{Q(x)}.$$ It is known that KL divergence is not a distance (not symmetric and also does not satisfy the triangle inequality).

Let $\lambda\in (0, \frac{1}{2})$ and let $R_{\lambda}:=\lambda P+(1-\lambda)Q$. Two distributions $P$ and $Q$ defined over $\mathcal{X}=\{\pm k,\pm(k-1), \dots, \pm 2, \pm 1\}$, is said to be symmetry of each other if $P(x)=Q(-x)$ for $x\in \mathcal{X}$.

I want to show that the following inequality always hold for symmetric distributions $P$ and $Q$ $$\frac{D(P||R_{1-\lambda})}{D(P||R_{\lambda})}< \frac{\log(1-\lambda)}{\log(\lambda)}.$$

It is easy to see that $\lambda\mapsto D(P||R_{\lambda})$ is convex and tends to zero as $\lambda\to 1$ and hence it is strictly decreasing on $[0, 1]$ and similarly $\lambda\mapsto D(P||R_{1-\lambda})$ is convex and strictly increasing and therefore $\lambda\mapsto \frac{D(P||R_{1-\lambda})}{D(P||R_{\lambda})}$ is strictly increasing. The following is the plot of two sides of above inequality when $$Q=[0.06, 0.03, 0.05, 0.04, 0.02, 0.01, 0.04, 0.03, 0.02 , 0.04, 0.04, 0.09, 0.03, 0.25, 0.15, 0.1]$$ and $$P=0.1, 0.15, 0.25, 0.03, 0.09, 0.04, 0.04, 0.02, 0.03, 0.04, 0.01, 0.02, 0.04, 0.05, 0.03, 0.06]$$ Red is the plot of the left-hand side and blue represents the right-hand side. enter image description here

Thanks to Iosif Pinelis, now we know that the above inequality does not hold for arbitrary $P$ and $Q$.

I will be very thankful if you can let me know of any reference related to this problem.

$\endgroup$

1 Answer 1

8
$\begingroup$

The inequality $$\frac{D(P||R_{1-\lambda})}{D(P||R_{\lambda})} < \frac{\log(1-\lambda)}{\log\lambda} \quad\text{for all $\lambda\in(0,1/2)$}$$ is not true in general. E.g., take $P=\frac1{100}\,(57,42,1)$, $Q=\frac1{100}\,(41, 12, 47)$, and $\lambda=\frac1{10}$. Then the left-hand side of this inequality and its right-hand side are $0.0537\dots$ and $0.0457\dots$, respectively.

However, it seems likely that this inequality holds if it is additionally assumed that $P(x)=Q(-x)$ for all $x$.


Addendum

The previous answer was given not assuming the "symmetry" condition.
Given now this condition, write

$$D(P||R_t)=\sum_{|x|=1}^k P(x)\ln\frac{P(x)}{t P(x)+(1-t)P(-x)}$$ $$=\sum_{x=1}^k \Big(P(x)\ln\frac{P(x)}{t P(x)+(1-t)P(-x)} +P(-x)\ln\frac{P(-x)}{t P(-x)+(1-t)P(x)}\Big) $$ $$=\sum_{x=1}^k P(x)F(t,r(x))\,\ln\frac1t $$ for $t\in(0,1)$, where $r(x):=P(-x)/P(x)$ and $$F(t,u):=\frac1{\ln(1/t)}\Big(\ln\frac1{t +(1-t)u} +u\ln\frac1{t +(1-t)/u}\Big). $$ Here we are assuming that $P(x)>0$ for all $x$; otherwise, extend by the continuity.

So, the problem is reduced to showing that $$(*)\qquad \Delta(t,u):=F(t,u)-F(1-t,u)>0$$ for $t\in(0,1/2)$ and $u\in(0,1)\cup(1,\infty)$ (clearly, $\Delta(t,1)=0$). By plotting, inequality $(*)$ appears quite likely, and it should be not very hard to prove it by calculus, as the problem now involves only two real variables, $t$ and $u$.

Here the graphs of $\Delta(t,u)$ and $\text{sgn}\,\Delta(t,u)$ are shown for $t\in(0,1/2)$ and $0<u<5$: enter image description here

Addendum 2

It is indeed not very hard to show that $(*)$ is true. Note that $\Delta(t,1/u)=\Delta(t,u)/u$ for $u>0$. So, without loss of generality $u>1$. That is, it suffices to show that $$h(u):=\Delta(t,u)\ln(1-t)\ln t $$ is positive for $t\in(0,1/2)$ and $u>1$. Let $$h_2(u):=h''(u)\frac{(1 -t + t u)^2 (t + (1- t) u)^2 u}{1 + u} $$ $$= t^2 [(1- t)^2 -(1 - 4 t + 2 t^2) u + (1- t)^2 u^2] \ln t -(1- t)^2 (t^2 (u-1)^2 + u) \ln(1 - t) . $$ So, $h_2$ is a quadratic polynomial in $u$, with the coefficient of $u^2$ equal to
$-(1- t)^2 t^2 \ln(-1+1/t)<0$ for $t\in(0,1/2)$. So, $h_2$ is concave and $h_2(\infty-)=-\infty$.

Next, $h_2(1)=g(t):=t^2 \ln t-(1- t)^2 \ln(1-t)$, $g(0+)=g(1/2)$ and $g''(t)=-2\ln(-1+1/t)<0$, so that $g$ is concave and hence $h_2(1)>0$ for $t\in(0,1/2)$. Because $h_2$ is concave and $h_2(\infty-)=-\infty$, it follows that $h_2(u)$ and hence $h''(u)$ switch its sign exactly once, from $+$ to $-$, as $u$ increases from $1$ to $\infty$. So, for some real $c=c(t)>1$, the function $h$ is convex on $[1,c]$ and concave on $[c,\infty)$. At that, $h(1)=h'(1)=0$ and $h(\infty-)=\infty$. So, $h>0$ on $(1,\infty)$. That is, indeed $\Delta(t,u)>0$ for $t\in(0,1/2)$ and $u>1$. \qed

$\endgroup$
6
  • $\begingroup$ thanks for your counterexample. I will edit the question to make it just for the symmetric distributions. $\endgroup$ Nov 3, 2015 at 17:35
  • $\begingroup$ I have added a reduction concerning the symmetric case. $\endgroup$ Nov 3, 2015 at 22:18
  • $\begingroup$ I have completed the proof in the symmetric case as well. $\endgroup$ Nov 4, 2015 at 5:10
  • $\begingroup$ thanks a lot for your detailed answer. In fact, your addendum 1 was exactly what I had two days ago, but I was stuck in the computation you added. It is very smart. $\endgroup$ Nov 4, 2015 at 15:27
  • $\begingroup$ can you please let me know why $h_2(+\infty)=-\infty$? $\endgroup$ Nov 5, 2015 at 0:08

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.