Sliding Modes in Solving Convex Programming Problems (Q4388929)

The authors consider a regular convex programming problem in \(R^n\). That means that the objective functional is a \(C^1\) convex function and the admissible set is characterized by a finite set of inequalities for convex functions, also of class \(C^1\), whose gradients are linearly independent on the boundary. Introducing a nonsmooth penalty function the original constrained problem is transformed into a problem of unconstrained minimization. The solution to the latter one is sought as an equilibrium point of Filippov's trajectories of the differential inclusion generated by the subdifferential of the penalized functional. Such a method was earlier applied in the theory of neural networks without rigorous theoretical analysis. The main result of the authors is a rigorous proof of the convergence. The sliding trajectories arising on the boundary of the admissible set, where the penalty term is not differentiable, are also carefully analyzed. Two illustrative examples are provided.