Spectral substructured two-level domain decomposition methods.

Two-level domain decomposition (DD) methods are very powerful techniques for the efficient numerical solution of partial differential equations (PDEs). A two-level domain decomposition method requires two main components: a one-level preconditioner (or its corresponding smoothing iterative method), which is based on domain decomposition techniques, and a coarse correction step, which relies on a coarse space. The coarse space must properly represent the error components that the chosen one-level method is not capable to deal with. In the literature most of the works introduced efficient coarse spaces obtained as the span of functions defined on the entire space domain of the considered PDE. Therefore, the corresponding two-level preconditioners and iterative methods are defined in volume. In this paper, a new class of substructured two-level methods is introduced,for which both domain decomposition smoothers and coarse correction steps are defined on the interfaces (or skeletons). This approach has several advantages. On the one hand, the required computational effort is cheaper than the one required by classical volumetric two-level methods. On the other hand, it allows one to use some of the well-known efficient coarse spaces proposed in the literature. While analyzing in detail the new substructured methods, we present a new convergence analysis for two-level iterative methods, which covers the proposed substructured framework. Further, we study the asymptotic optimality of coarse spaces both theoretically and numerically using deep neural networks. Numerical experiments demonstrate the effectiveness of the proposed new numerical framework.