Multigrid methods for saddle point problems: optimality systems (Q2297196)

The authors consider the following optimal control problem \[ (\bar{y},\bar{u})=\operatorname{argmin}_{(y,u)\in K}\left [ \frac{1}{2}\|y-y_d\|^2_{L_2 (\Omega)}+\frac{\beta}{2}\|u\|^2_{L_2(\Omega)}\right], \] subject to elliptic boundary value problem with a convective term \[ \int_\Omega \nabla y \cdot \nabla z dx +\int_\Omega [(\xi\cdot\nabla y)z -(\xi\cdot\nabla z)y]dx + \int_\Omega \gamma yz dx = (u.z)_{L_2(\Omega)} \quad \forall z \in H_0^1(\Omega), \] in a bounded convex polygonal domain \(\Omega\) in \(\mathbb{R}^d\) (\(d=2, 3\)). The saddle point problem, connected with the optimal control problem is discretized by \(P_1\) finite elements. Multigrid methods are developed for the resulting discrete saddle point problem, whose performance is independent of the regularization parameter \(\beta\). The authors proved that for convex domains the \(W\)-cycle algorithm is uniformly convergent for a sufficiently large number of smoothing steps with respect to mesh refinements and a regularizing parameter. The theoretical estimates and the performance of the algorithms are demonstrated by numerical results for two and three dimensional problems.