On convergence rate of the augmented Lagrangian algorithm for nonsymmetric saddle point problems (Q557929)

The paper is concerned with the solution of the saddle point problem \[ \begin{pmatrix} A & L^T \\ L &0 \end{pmatrix} \begin{pmatrix} c \\ \lambda \end{pmatrix} = \begin{pmatrix} F \\ G \end{pmatrix}. \] The matrix \(A\) is changed to \(A+\varepsilon^{-1}L^TL\) in the augmented Lagrangian algorithm. The application of Uzawa's algorithm to the latter is equivalent to an iteration with the matrix \[ \begin{pmatrix} A & L^T \\ L & \varepsilon I\end{pmatrix}. \] The authors study the solvability for this matrix independently of the original problem. It is assumed that the symmetric part of \(A\) is positive definite on the kernel of \(L\); cf. \textit{F. Brezzi}'s famous theory [Rev. Franc. Automat. Inform. Rech. Operat. 8, 129--151 (1974; Zbl 0338.90047)] and the generalization to unsymmetric matrices. It is shown that the convergence rate of Uzawa's iteration is bounded by \(C\varepsilon\). Therefore one has convergence for sufficiently small \(\varepsilon\). This type of problem typically arises in discretizations of the Navier-Stokes equations. In that case the estimates have to be done uniformly for a family of discretization spaces.