Index: The Book of Statistical ProofsGeneral Theorems ▷ Information theory ▷ Kullback-Leibler divergence ▷ Convexity

Theorem: The Kullback-Leibler divergence is convex in the pair of probability distributions $(p,q)$, i.e.

\[\label{eq:KL-conv} \mathrm{KL}[\lambda p_1 + (1-\lambda) p_2||\lambda q_1 + (1-\lambda) q_2] \leq \lambda \mathrm{KL}[p_1||q_1] + (1-\lambda) \mathrm{KL}[p_2||q_2]\]

where $(p_1,q_1)$ and $(p_2,q_2)$ are two pairs of probability distributions and $0 \leq \lambda \leq 1$.

Proof: The Kullback-Leibler divergence of $P$ from $Q$ is defined as

\[\label{eq:KL} \mathrm{KL}[P||Q] = \sum_{x \in \mathcal{X}} p(x) \cdot \log \frac{p(x)}{q(x)}\]

and the log sum inequality states that

\[\label{eq:logsum-ineq} \sum_{i=1}^n a_i \log \frac{a_i}{b_i} \geq \left( \sum_{i=1}^n a_i \right) \log \frac{\sum_{i=1}^n a_i}{\sum_{i=1}^n b_i}\]

where $a_1, \ldots, a_n$ and $b_1, \ldots, b_n$ are non-negative real numbers.

Thus, we can rewrite the KL divergence of the mixture distribution as

\[\label{eq:KL-conv-qed} \begin{split} &\mathrm{KL}[\lambda p_1 + (1-\lambda) p_2||\lambda q_1 + (1-\lambda) q_2] \\ \overset{\eqref{eq:KL}}{=} &\sum_{x \in \mathcal{X}} \left[ \left[ \lambda p_1(x) + (1-\lambda) p_2(x) \right] \cdot \log \frac{\lambda p_1(x) + (1-\lambda) p_2(x)}{\lambda q_1(x) + (1-\lambda) q_2(x)} \right] \\ \overset{\eqref{eq:logsum-ineq}}{\leq} &\sum_{x \in \mathcal{X}} \left[ \lambda p_1(x) \cdot \log \frac{\lambda p_1(x)}{\lambda q_1(x)} + (1-\lambda) p_2(x) \cdot \log \frac{(1-\lambda) p_2(x)}{(1-\lambda) q_2(x)} \right] \\ = &\lambda \sum_{x \in \mathcal{X}} p_1(x) \cdot \log \frac{p_1(x)}{q_1(x)} + (1-\lambda) \sum_{x \in \mathcal{X}} p_2(x) \cdot \log \frac{p_2(x)}{q_2(x)} \\ \overset{\eqref{eq:KL}}{=} &\lambda \, \mathrm{KL}[p_1||q_1] + (1-\lambda) \, \mathrm{KL}[p_2||q_2] \end{split}\]

which is equivalent to \eqref{eq:KL-conv}.


Metadata: ID: P148 | shortcut: kl-conv | author: JoramSoch | date: 2020-08-11, 07:30.