Method of types

For a finite alphabet $A$, there are at most $(n+1{)}^{|A|}$ length-$n$ types, and under an i.i.d. law $P$ the probability of the type class $T_R$ is at most $2^{-nD(R\Vert P)}$ when logarithms use base two.

An information projection of $Q$ onto a set $\mathcal{C}$ of probability distributions minimizes $D(P\Vert Q)$ over $P\in \mathcal{C}$. For a closed convex set and suitable $Q$, it obeys the Pythagorean inequality