On generalized Csiszár-Kullback inequalities (Q1596342): Difference between revisions

The authors give elementary but delicate proofs for several variants of the following result: Let \(\psi\) be a strictly convex nonnegative continuous function of a possibly unbounded interval \(J\) such that \(\psi\) tends to \(\infty\) at the boundary of \(J.\) Then there is a function \(U : [0,\infty) \times J \rightarrow \mathbb{R}\) satisfying natural monotonicity and continuity conditions such that for all probability spaces \((\Omega, {\mathcal S}, \mu)\) and all \(u \in L^ 1 (d\mu)\) with \(\int \psi (u) d\mu < \infty\) the inequality \[ \|u - \int_\Omega u\;d\mu \|_1 \leq U \biggl(\int \psi (u) d\mu, \int u d\mu\biggr) \] holds. For various \(\psi\) the function \(U\) is known explicitly. This allows in particular the \(L^1\)-distances to be estimated in terms of entropy quantities.