\(\varepsilon\)-optimality conditions for composed convex optimization problems (Q935086)

The aim of this paper is to provide a formula for the \(\varepsilon\)-subdifferential of a composed convex function in order to characterize the approximate solutions of an optimization problem involving composed convex functions. In this paper, \(X,Y\) are separated locally convex spaces, and \(X^*,Y^*\) denote their topological dual spaces; \(K\subset Y\) is a nonempty, closed and convex cone. Consider the functions \(f:X\to \overline{\mathbb R},\) \(g:Y\to \overline{\mathbb R}\) and \(h:X\to Y^{\bullet}\) such that \(f\) is proper, convex and lower semicontinuous, \(g\) is proper, convex, lower semicontinuous and \(K\)-increasing, while \(h\) is proper, \(K\)-convex and star \(K\)-lower semicontinuous, respectively. The main result (Theorem 2) provides the following formula for the \(\varepsilon\)-subdifferential of \(f+g\circ h,\) for any \(x\in \text{dom}(f)\cap \text{dom}(h)\cap h^{-1}(\text{dom}(g)),\) and for all \(\varepsilon\geq 0:\) \[ \partial_{\varepsilon}(f+g\circ h)(x)=\bigcup_{\varepsilon_1, \varepsilon_2,\varepsilon_3\geq 0,\;\varepsilon_1+\varepsilon_2+\varepsilon_3=1}\left\{\partial_{\varepsilon_1}f(x) +\partial_{\varepsilon_2}(\lambda h)(x):\lambda\in K^*\cap \partial_{\varepsilon_3}g(h(x)) \right\}. \] This formula holds if and only if the following regularity condition, based on the epigraphs of the conjugates of the functions involved, is fulfilled: the set \[ \text{epi}(f^*)+\bigcup_{\lambda\in \text{dom}(g^*)}\left(\text{epi}((\lambda h)^*)+(0,g^*(\lambda))\right) \tag{RC} \] is closed (here, \(l^*\) denotes the Fenchel transform of \(l\)). Denote by (P) the following optimization problem: \[ \inf_{x\in X}\left\{f(x)+(g\circ h)(x)\right\}. \] A point \(\overline{x}\) is an \(\varepsilon\)-optimal solution to (P) if \[ f(\overline{x})+(g\circ h)(\overline{x})\leq \inf_{x\in X}\left\{f(x)+(g\circ h)(x)\right\}+\varepsilon; \] equivalently, \(\overline{x}\in \partial_{\varepsilon}(f+g\circ h)(\overline{x}).\) In Theorem 4, under condition (RC), the authors provide necessary conditions and sufficient conditions in order that \(\overline{x}\) is an optimal solution of problem (P). In the last part of the paper, the results are applied to some particular cases (the composition with a linear operator \(h(x)=Ax,\) the case \(f\equiv 0,\) and the ordinary convex optimization problem), and the corresponding \(\varepsilon\)-subdifferential sum formulae are obtained.