Multilevel preconditioned augmented Lagrangian techniques for \(2\)nd order mixed problems (Q1919123)

The following classic boundary value problem is studied: \[ -\text{div} (a\nabla u)= f\text{ in } \Omega\subset \mathbb{R}^2; \quad u=g\text{ on } \Gamma_D \subset \partial \Omega, \quad \langle a\nabla u,\nu \rangle= 0\text{ on } \Gamma_N \subset \partial \Omega; \tag{1} \] \(f\in L^2 (\Omega)\) and \(a\in L^\infty (\Omega)\) with \(\overline \alpha\geq a(x)\geq\underline\alpha >0\) a.e. in \(\Omega\). Problem (1) is set up as the mixed variational problem which, in turn, is given an operator form: \[ A{\mathbf j} +B^*u=, \quad B{\mathbf j} =f. \tag{2} \] The augmented Lagrangian approach assumes then the form: \[ (A+rB^* i^{-1} B){\mathbf j} +B^*u= g+rB^*i^{-1}f, \quad B{\mathbf j} =f \tag{3} \] where \(r>0\) is the augmented Lagrangian parameter and \(i^{-1}\) denotes the inverse of the Riesz map \(i\). We observe that the exact solution of (2) can be recovered from (3) no matter what the value of \(r\) might be. The augmented Lagrangian is carefully discussed. It combines the finite element method with multilevel splitting. Numerical examples confirm the efficiency of the scheme proposed.