Abstract:
This thesis is concerned with the computational intractabilities that arise from spectral discretizations of high-dimensional partial differential equations. Using the example of the time-dependent multi-particle Schrödinger equation, we consider a spectral Galerin approximation in space with a tensor product basis of Hermite functions. When propagating the resulting system of ordinary differential equations in time, one typically needs to evaluate matrix-vector products involving a matrix representation of the Hamiltonian operator in each time step. Since the size of this matrix equals the number of equations, which (for an unreduced basis) depends exponentially on the dimension and thus quickly becomes enormously large, this can make computations infeasible - both due to a lack of memory and due to unbearably long computation times.
We present a fast algorithm for an efficient computation of these matrix-vector products that scales only linearly with the size of the Galerkin basis - without assembling the matrix itself. Besides being a matrix-free approach, the fast algorithm is compatible with the idea of reducing the index set that underlies the basis. The computational speed-up is achieved using orthogonality of the Hermite functions in combination with their generic three-term recurrence. Briefly, these properties allow to compute the action of the matrix representations of the coordinatewise position operators on vectors in linear time. The basic idea is then to insert these coordinate matrices into a polynomial approximate of the potential. This has originally been proposed by E. Faou, V. Gradinaru, and Ch. Lubich. We modify their approach and turn it into a rigorous algorithm. For an unreduced set of basis functions, we show this tentative proceeding to be equivalent to a suitable entrywise approximation of the potential matrix by Gauß-Hermite quadrature. Reducing the index set for the basis functions yields an additional error.
We derive error estimates for all approximation steps involved. In particular, using a binary tree approach, bounds for the errors due to quadrature and index set reduction are deduced. Both errors decay spectrally if the potential is significantly smoother than the exact solution wherever the latter does not essentially vanish. Extensive numerical experiments corroborate these findings. Besides, we show performance tests comparing the fast algorithm to a matrix-free approach from the chemical literature.
Apart from the above basic form of the fast algorithm, we present applications of the general methodology to the nonlinear Schrödinger equation and, most prominently, to initial-boundary value problems. As an example, we study the acoustic wave equation with non-constant coefficients and Engquist-Majda boundary conditions, and construct efficient procedures for the different kinds of matrix-vector products together with an error analysis and numerical tests.