Entropy minimization with lattice bounds (Q1335038)

The paper deals with the entropy minimization problem with a finite set of linear constraints and lattice bounds \(\alpha\) and \(\beta\). The solutions to the problem of minimizing a convex integral objective function subject to a finite number of linear constraints and requiring that the feasible functions lie in a strip \([\alpha,\beta]\) when \(\alpha\) and \(\beta\) are extended real-valued measurable functions are characterized. More specifically, suppose \((K,\mu)\) is a complete finite measure space, \(\alpha\) and \(\beta\) two extended real-valued measurable functions, \(\alpha(t)< \beta(t)\) for almost all \(t\in K\), \(\{\psi_ i\}^ n_{i=1}\) a collection of functions in \(L^ \infty(K,\mu)\) and suppose \(\phi: \mathbb{R}\to (-\infty,\infty]\) is a lower semicontinuous essentially smooth proper function strictly convex on its domain with (necessarily smooth) finite conjugate \(\phi^*\). The authors characterize solutions in \(L^ 1(K,\mu)\) to \[ \inf\left\{ \int_ K \phi(x(t))d\mu(t): \alpha\leq x\leq \beta,\;\int_ K x(t)\psi_ i(t) d\mu(t)= b_ i,\;i\leq i\leq n\right\},\tag{P} \] where \(b= (b_ 1,\dots,b_ n)\) is a given vector in \(\mathbb{R}^ n\). Using the duality theory of the first two authors [Math. Program. 57 B, No. 1, 15-48, 49-83 (1992; Zbl 0778.90049 and Zbl 0778.90050)]. It is shown that the solutions are of usual form, but truncated where they leave the strip. Applications that fit into the form of (P) can be found in convex interpolation, crystallography, digital signal processing, information theory etc.