An unconventional robust integrator for dynamical low-rank approximation

Abstract We propose and analyse a numerical integrator that computes low-rank approximation to large time-dependent matrices are either given explicitly via their increments or the unknown solution matrix differential equation. Furthermore, is extended of tensors by Tucker fixed multilinear rank. The proposed different from known projector-splitting for dynamical approximation, but it retains important robustness small singular values has so far been only integrator. new also offers some potential advantages over integrator: It avoids backward time integration substep integrator, which potentially unstable dissipative problems. more parallelism, preserves symmetry anti-symmetry tensor when equation does. Numerical experiments illustrate behaviour