Simple sequential quadratically constrained quadratic programming feasible algorithm with active identification sets for constrained minimax problems (Q2250067)

The maximum \(F\) of \(m\) smooth functions \(f_i\) over \(\mathbb R^n\) is minimized w.r.t. \(m'\) smooth inequality constraints \(g_j(x)\leq 0\). Under the MFCQ for the constraints, \(C^1\) property of all functions, the well-known KKT conditions with multipliers for the \(f_i\) and \(g_j\) are formulated. Index sets for nearly active sets for \(f_{i_0}\leq\max_i f_i(x)\) and \(g_j(x)\leq 0\) are introduced in analogy to [\textit{D.-L. Han} et al., Nonlinear Anal., Theory Methods Appl., Ser. A, Theory Methods 74, No. 9, 3022--3032 (2011; Zbl 1235.90150)]. The main feasible descent direction is determined by a constrained quadratic problem with the objective \(\gamma_0 z+0.5 d^TB_k d\) and as inequality constraints the first-order approximations of the functions \(f_i\) and \(g_j\) at \(x^k\) with a right hand side linear perturbation \(F(x^k)+\gamma_0 z\), resp. a quadratic perturbation \(\gamma_j\theta_k z-\epsilon_kz^2\), where the indices \(i,j\) belongs to the nearly active set at \(x^k\). \(B_k\) is a positively definite approximation of the Hessian of the Lagrangian using \(f_i\) and \(g_j\). \(z\) controls the stationarity of \(x^k\). In order to avoid the Maratos effect, a higher-order correction \(\tilde d\) of the main direction is calculated by some improved technique of \textit{J.-B. Jian} et al. [Appl. Math. Optim. 56, No. 3, 343--363 (2007; Zbl 1144.90025)]. Some Armijo search is done along the parabula \(x^k+td+t^2\tilde d\). The multipliers are updated by the Lagrange multipliers of the quadratic problem. If the second-order correction does not exist, it is set to zero. They prove under the uniform positively definiteness of \(B_k\), \(C\geq \theta_k+\varepsilon_k\geq c>0\) and MFCQ for the active constraints \(g_i\) that each accumulation point of the sequence \(x^k\) is stationary for an arbitrary starting point (global convergence, Theorem 3.1). Under the common second-order assumptions as \(f_i,g_j\) in \(C^2\), LICQ, strong second-order sufficient conditions and strict complementarity w.r.t. the \(g_j\) at a limit point \(x^*\) and two suggested weakened Newton like conditions w.r.t. \(B_k\), they state the superlinear convergence of the \(x^k\) to \(x^*\) (Theorem 4.3). For the proof of this theorem they show some intermediate results and refer to [\textit{J.-B. Jian} et al., J. Comput. Appl. Math. 205, No. 1, 406--429 (2007; Zbl 1149.90148), Theorem 4.3]. Numerical experiments with Powell's modification of BFGS show the suggested method is well applicable for small- and mid-scaled problems. However, there is no comparison with former methods discussed in the introduction. It remains open, as the authors remarks, which updates for \(B_k\) satisfy the weakened Newton-like conditions but do not the Newton-like ones.