Adaptive integration of nonlinear evolution equations on tensor manifolds (Q2161542)

In this paper, adaptive step-truncation algorithms are developed, selecting the tensor rank at each time step based on desired accuracy and stability constraints. These methods can be implemented easily as they rely only on arithmetic operations between tensors, which efficient and scalable parallel algorithms can perform. After defining a new criterion for tensor rank adaptivity based on local error estimates, the authors prove convergence of a wide range of rank-adaptive step-truncation algorithms. Furthermore, a connection between rank-adaptive step-truncation methods and rank-adaptive dynamical tensor approximation is established. Numerical results are presented and discussed for rank shock problem, two and four-dimensional Fokker-Planck equations using adaptive Euler, adaptive midpoint, and two-step rank-adaptive Adams-Bashforth method.