A proximal gradient descent method for the extended second-order cone linear complementarity problem (Q2268076)

The paper of Shoahua Pan and Jein-Shan Chen Wu is located in the area of mathematical programming where the solution of complementarity conditions meets optimization theory. The interrelations between the two areas are of various nature, e.g., in terms of duality, game, equilibrium (Hamiltonian, etc.) or other geometric ways of motivation. Historically, a motivation for them consisted in corresponding optimality conditions of first order (for classical nonlinear programs underlying), which gave rise to a new solution technique for the regarded problems, exploiting the multiplicative nature of equality constraints accompanied by nonnegativity conditions on the factors. By the use of a so-called NCP function, the latter conditions becomes ``absorbed'' and, hence, disappears in the problem reformulation. In this paper, optimization theory is employed after the regarded complementarity problem has become reformulated as an optimization problem. The setting of the problems studied and solved is cone-valued, with the famous second-order (or ''icecream'' or Lorentz) cones. The present paper considers an extended second-order cone linear complemtarity problem (SOCLCP) including the following subclasses of that system: the generalized SOCLCP, the horizontal SOCLCP, the vertial SOCLCP and the mixed SOCLCP. The authors present some simple second-order cone constrained and unconstrained reformulation problems. Under mild conditions they prove the equivalence between the stationary points of these optimization problems and the solutions of the extended SOCLCP. In particular, the authors develop a proximal gradient decent method for solving the second-order cone constrained problems. This in fact simple method makes only one Euclidean projection onto second-order cones in each iteration. In this paper, a global convergence and, under a local Lipschitzian error bound assumption, a linear rate of convergence are established. Furthermore, for the unconstrained reformulations, numerical comparisons with limited-memory BFGS method are provided, verifying the effectiveness of the proposed method. This paper is well written and structured, it advances science and may serve to support practical applications and future research as well. It consists of Section 1 on introduction, Section 2 on preliminaries and examples, of Section 3 on constrained and unconstrained reformulations, of Section 4 on the solution of the SOC constrained problem, on Section 5 on numerical experience and of Section 6 on conclusions.