Solving stochastic systems with low-rank tensor compression

For parametrised equations, which arise, for example, in equations dependent on random parameters, the solution naturally lives in a tensor product space. The application which we have in mind is a stochastic linear elliptic partial differential equation (SPDE). Usual spatial discretisation leads to a potentially large linear system for each point in the parameter space. Approximating the parametric dependence by a Galerkin ‘ansatz’, the already large number of unknowns—for a fixed parameter value—is multiplied by the dimension of the Galerkin subspace for the parametric dependence, and thus can be very large. Therefore, we try to solve the total system approximately on a smaller submanifold which is found adaptively through compression during the solution process by alternating iteration and compression. We show that for general linearly converging stationary iterative schemes and general adaptation processes—which can be seen as a modification or perturbation of the iteration—the interlaced procedure still converges. Our proposed modification can be used for most stationary solvers for systems on tensor products. We demonstrate this on an example of a discretised SPDE with a simple preconditioned iteration. We choose a truncated singular value decomposition (SVD) for the compression and give details of the implementation, finishing with some examples.