Conservation properties of iterative methods for implicit discretizations of conservation laws (Q2161820)

The authors study conservation properties of some iterative methods applied to implicit finite volume discretizations. In fact conservation is one of the main properties which should be satisfied by a numerical scheme. It is shown that Newton's method, the Richardson iteration, Krylov subspace methods, coarse grid corrections using agglomeration, as well as ERK (Implicit Runge-Kutta) pseudo-time iterations preserve the global conservation of a given scheme if the initial guess has correct mass. However, the Jacobi and Gauss-Seidel methods do not, unless the linear system being solved possesses particular properties. The stronger requirement of local conservation has been investigated for ERK pseudo-time iterations. It is also shown that local conservation is preserved for implicit Euler finite volume schemes. However, the resulting modified numerical flux may be inconsistent with the governing conservation law. An extension of the Lax-Wendroff theorem is proved. Such extension reveals that numerical solutions based on these methods converge to weak solutions of a modified conservation law where the flux function is multiplied by a particular constant. This constant depends on the choice of RK method but is independent of both the conservation law and the discretization. Consistency is maintained by ensuring that this constant equals unity and a strategy for achieving this is presented. Numerical tests support the theoretical findings and suggest that the results hold even for systems of conservation laws in multiple dimensions. In addition to this, experiments with GMRES (Generalized Minimal Residual) method suggest that they suffer from a similar modification as pseudo-time iterations, but that this is automatically accounted for after a finite number of iterations. Similar experiments with coarse grid corrections based on agglomeration indicate no presence of such modifications.