An LQP-based descent method for structured monotone variational inequalities (Q609244)

This paper presents a descent type method for solving a class of structured variational inequalities (SVIs) by improving the method presented by \textit{B. S. He, Y. Xu} and \textit{X. M. Yuan} [Comput. Optim. Appl. 35, No.~1, 19--46 (2006; Zbl 1121.90119)]. The authors consider the structured variational inequality of the form \[ \text{Find }u^* \in \Omega\text{ such that }(u-u^*) F(u^*) \geq 0 \quad \forall u \in \Omega,\tag{1} \] where \[ u =\binom{x}{v} ,\quad F(u) = \binom{f(x) + Ay}{-A^Tx+b},\quad \Omega =\mathbb R^n_+ \times {y},\tag{2} \] \(A\in\mathbb R^{n \times m}\), \(b \in\mathbb R^m\), \({y}\) is a nonempty closed convex subset of \(\mathbb R^m\) and \(f\) is a continuous and monotone mapping from \(\mathbb R^n\) into itself (here, \(F(u)\) is monotone, too). It is derived that a solution of the SVI (1)--(2) can be approached by the iterative sequence \(\{ u^{K+1} = (x^{K+1}, y^{K+1}) \in \mathbb R^n_{++} \times {y}\}\), where \(x^{K+1}\) and \(y^{K+1}\) are the solutions of the problems: \[ \beta_K [f(x) + Ay] + x - (1-\mu)x^K - \mu X^2_K x^{-1} =0\tag{3a} \] \[ (y' - y)^T \{ (y-y^K) - \frac{1}{\nu} \beta_K (A^T x-b)\} \geq 0 \quad \forall y' \in {y}.\tag{3b} \] Here, \(\beta_K \geq \beta >0\), \(\mu \in (0,1)\), \(\nu >0\), \(X_K=\text{diag}(x^K_1, x^K_2, \dots,x^K)\) and \(x^{-1}\) is an \(n\)-vector whose \(j\)-th element is \(1/x_j\). The solution of the mixed system (3) is very difficult because this mixed system, which consists of a system of equations and a variational inequality, involves the overlapped variables \(x\) and \(y\) that need to be solved simultaneously. This paper develops a logarithmic-quadratic proximal (LQP) method-based descent method for solving the SVI (1)--(2) by improving the method presented by He, Yu, and Yuan [loc. cit.]. For the current iterate \(u^K = (x^K, y^K)\), the prediction step generates a trial value of the new iterate \(u^{K+1}\), denoted by \(\widetilde{u}^K = (\widetilde{x}^K, \widetilde{y}^K)\) via solving (3a) approximately under practical inexactness restriction and computing a projection onto \({y}\). Then, \(u^{K+1}\) is obtained by correcting \(\widetilde{u}^K\) with an additional projection onto \({y}\) and some trivial computations at the correction step. Main result: The new descent method is presented. Some contractive properties of the new method are first analysed. In particular, the strategy of determining the optimal step-sizes of the descent method is investigated. The global convergence of the new method is proved. The authors apply the new method to solve some traffic equilibrium problems, and compare it with the method presented by He, Yu, and Yuan [loc. cit.]. Finally, numerical results and some conclusions are also given.