On weak conjugacy, augmented Lagrangians and duality in nonconvex optimization (Q2216183)

A function \(f:X\to \mathbb{R}\) is weakly subdifferentiable at a point \(\bar{x}\) if there exist \((u,\alpha)\in X^*\times \mathbb{R}\) such that \(f(x)\geq \langle u,x-\bar{x}\rangle - \alpha \|x-\bar{x}\|\) for any \(x\in \mathrm{dom} f\). Then \((u,\alpha)\) is called a weak subgradient of \(f\) at \(\bar{x}\). The aim of this paper is to investigate Lagrangian duality for weakly subdifferentiable functions. The authors start by definition the weak conjugate and weak biconjugate to a function \(f\), and show that the weak biconjugate is always a minorant to \(f\). They then show that equality occurs at a point if and only if \(f\) is weakly subdifferentiable at that point (i.e., when it admits a weak subgradient). The next section of the paper introduces a dualization scheme, based on a dualising function \(\Phi:X\times Y \to \mathbb{R}\) such that \(\Phi(x,0)=x\) for all \(x\), using which they define a dual problem (using weak conjugation of \(\Phi\)) and establish weak duality of this problem and the original problem of minimising \(f\). The authors show that the sharp augmented Lagrangian introduced by Rockafellar and Wets can be derived using their scheme, and derive another scheme for the case of inequality constraints. Examples are provided throughout the section.