A general duality approach to \(I\)-projections (Q1907633)

The authors consider the problem \[ \inf_{x\in S} f(x), \] where \(x\in L_1(Q)\), \[ f(x)= \begin{cases} \int x\ln x dQ\quad & \text{if } x\geq 0,\;\int x dQ= 1\\ + \infty\quad & \text{elsewhere}\end{cases} \] and \(Q\) is a probability measure defined on a \(\sigma\)-field of subsets of a sample space \(\Omega\). For the above-mentioned problem the generalized dual optimization problem is constructed. Contrary to the standard Fenchel duality, the dual problem introduced in the paper is posed over the space \(\overline M(\Omega, {\mathcal F})\) of all extended valued \(\mathcal F\)-measurable functions instead of the dual space \((L_1(Q))^*\). One of the main results is the following theorem. Suppose that \(S\) is a nonempty subset of \(L_1(Q)\) and \(y_0\) is a solution to \[ \inf_{y\in S^\oplus} \int e^y dQ< \infty, \] where \(S^\oplus= \{y\in \overline M(\Omega, {\mathcal F}): \int xy dQ\geq 0\) for all \(x\in S\}\). Then \(x_0= e^{y_0}/ \int e^{y_0} dQ\) is the solution to \(\inf_{x\in S} f(x)\), if \(x_0\in S\). If \(S\) is either convex closed and contained in \(\text{dom}(f)\) or a closed convex cone, then \(x_0\in S\).