Approximate stationarity in disjunctive optimization: concepts, qualification conditions, and application to MPCCs
Isabella Käming, Patrick Mehlitz
TL;DR
The paper studies optimization problems with disjunctive constraints of the form $F(x)\in \Gamma$, where $\Gamma = \bigcup_{j=1}^t \Gamma_j$ with convex polyhedral $\Gamma_j$. It introduces approximate stationarity concepts—AM-stationarity and SAS-stationarity—and associated regularity notions AM-regularity and AS-regularity, along with a new subset Mangasarian–Fromovitz condition (subMFC) for orthodisjunctive problems. It shows that AM-/AS-regularity together with the respective approximate stationarity notions imply M-/S-stationarity, and that subMFC provides a practical, sequence-based route to deduce exact stationarity from approximate information. The results specialize to inequality problems and MPCCs, recovering AKKT-type conditions and Clarke/Mordukhovich/weak-stationarity notions, and open a path for algorithmic verification and extension to other disjunctive models.
Abstract
In this paper, we are concerned with stationarity conditions and qualification conditions for optimization problems with disjunctive constraints. This class covers, among others, optimization problems with complementarity, vanishing, or switching constraints, which are notoriously challenging due to their highly combinatorial structure. The focus of our study is twofold. First, we investigate approximate stationarity conditions and the associated strict constraint qualifications which can be used to infer stationarity of local minimizers. While such concepts are already known in the context of so-called Mordukhovich-stationarity, we introduce suitable extensions associated with strong stationarity. Second, a qualification condition is established which, based on an approximately Mordukhovich- or strongly stationary point, can be used to infer its Mordukhovich- or strong stationarity, respectively. In contrast to the aforementioned strict constraint qualifications, this condition depends on the involved sequences justifying approximate stationarity and, thus, is not a constraint qualification in the narrower sense. However, it is much easier to verify as it merely requires to check the (positive) linear independence of a certain family of gradients. In order to illustrate the obtained findings, they are applied to optimization problems with complementarity constraints, where they can be naturally extended to the well-known concepts of weak and Clarke-stationarity.