1
$\begingroup$

The KL divergence between two probability measures $p, q$ on a Polish space $E$ is defined as

$$D_{KL}[q \| p] = \int_E \log \left(\frac{dq}{dp}\right)dq$$

provided that $q \ll p$.

Now is it true that $D_{KL}[q_1 \| p]\leq D_{KL}[q_2 \| p] \Rightarrow D_{KL}[p\|q_1]\leq D_{KL}[p \| q_2]$ provided that $p,q_1,q_2$ are equivalent in the sense of measures?

$\endgroup$
1
$\begingroup$

The answer is no.

I'll explicitly construct a triple $(p,q_1,q_2)$ on four points that violates the relation. This can be extended to however rich a Polish space you like (as long as it has at least $4$ points) by convolving with appropriately thin densities.


Let $$p = \frac{(1,1,1,1)}{4}\\ q_1^x = \frac{(2-x, 2-x, x,x)}{4}\\ q_2 = \frac{(3, 1, 1, 1)}{6},$$ where $x \in (0,2)$ will be chosen in the following.

To construct a counterexample, we need to find an $x$ such that $D(q_1^x \|p) < D(q_2\|p)$ and $D(p\|q_1^x ) > D(p\|q_2)$. By direct computation, this boils down to finding an $x$ such that \begin{align} (2-x) \log(2-x) + x \log x &< \log(4/3)\\ 2\log(2-x) + 2\log(x) &< \log(16/27) \end{align}

The easiest way to find an appropriate $x$ is to simply use a CAS. Wolfram alpha suggests that $x = 1/2$ is good enough. Indeed, $$ (3/2) \log(3/2) +(1/2) \log(1/2) = \log(\sqrt{27}/4) \le \log(5.22/4) = \log(1.305) < \log(4/3)\\ 2\log(3/2) + 2\log(1/2) = \log(9/16) < \log(16/27),$$ where the first line uses that $\sqrt{3} < 1.74,$ and the second line follows since $9 \times 27 = 243 < 256 = 16^2$.


Process for finding the above - I started with $p$ being a uniform law because that makes $D(q\|p)$ and $D(p\|q)$ both have simple forms. $D(q_1\|p) < D(q_2\|p)$ is then equivalent to $H(q_1)>H(q_2),$ which led me to think about $q_1$ that is a little more spread out than $q_2$ - to simplify I chose $q_1$ peaked on half the space and $q_2$ on a quarter. $D(p\|q_1) > D(p\|q_2)$ is equivalent to $\sum \log(1/q_1(i)) > \sum \log(1/q_2(i))$. At this point I realised that I might as well take just $4$ points, and ran a CAS for $q_1$ and $q_2$ parametrised. Finally I just set the parameter of $q_2$ somewhere in the middle to make calculations easy.

| cite | improve this answer | |
$\endgroup$

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.