Second-order optimality conditions in minimax optimization problems (Q1949567): Difference between revisions

A finite-dimensional minimax problem \(\min_{x} \sup_{y\in Y} f(x,y)\) s.t. \(g(x)\in D\), \(h(x)=0\) with \(x\in \mathbb R^n\), \(Y\subset \mathbb R^m\) compact, \(D\subset \mathbb R^m\) closed with nonempty interior, \(f(\cdot,y),y\in Y\); \(g\); \(h\) \(C(1,1)\) functions is reformulated and dealt with cone constraint as infinite programming problem (MP): \(\min r\) s.t. \(f(x,\cdot)-r \in -C_{+}(Y)\), \(g(x)\in D\), \(h(x)=0\). Here \(C_{+}(Y)\) is the order cone of continuous functions from \(Y\) in \(\mathbb R^m\). Using second-order lower and upper Dini directional derivatives approximations of Clarke's Hessian for \(C(1,1)\) functions and a generalized Taylor's expansion of order 2 for vector-valued mappings is given. Second-order necessary conditions for (MP) are derived in three looks: as an infinite programming problem with multiplier in the adjoint spaces (Theorem 4.1), as a representation of the multiplier in infinite space by a probability measure (Theorem 4.2) with integration and finally using [\textit{H. Kawasaki}, Math. Program., Ser. A 49, No. 2, 213--229 (1990; Zbl 0726.90075)] Lemma 4.5 similar to semi-infinite programming (SIP, Theorem 4.3) (in the last representation the authors should reduce the number \(s\) to \(n+1\) using [loc. cit.] Lemma 4.6). Main tools are first- (second-) order inner and outer tangent cones (sets), regularity conditions as Mangasarian-Fromovitz-constraint qualification (MFCQ), Abadie condition related to the image spaces (here additionally necessary for second-order conditions in the \(C(1,1)\) case) and a KKT-like condition with multipliers for optimality conditions in tangent set form (Lemma 3.2). For SIP-representation second-order sufficient conditions are derived without using a regularity condition. All results are proved in detail. The paper is nearly self-contained.