Algorithm MGB to solve highly nonlinear elliptic PDEs in $\tilde{O}(n)$ FLOPS

Abstract: We introduce Algorithm MGB (Multi Grid Barrier) for solving highly nonlinear convex Euler-Lagrange equations. This class of problems includes many highly nonlinear partial differential equations, such as $p$-Laplacians. We prove that, if certain regularity hypotheses are satisfied, then our algorithm converges in $\tilde{O}(1)$ damped Newton iterations, or $\tilde{O}(n)$ FLOPS, where the tilde indicates that we neglect some polylogarithmic terms. This the first algorithm whose running time is proven optimal in the big-$\tilde{O}$ sense. Previous algorithms for the $p$-Laplacian required $\tilde{O}(\sqrt{n})$ damped Newton iterations or more.