Rank-Adaptive IMEX Tensor-Train BUG Integrators for High-Dimensional Fractional Fokker–Planck Equations
Preprint
- 17 November 2026
- preprint
- Published by MDPI AG in Preprints.org
Abstract
We consider nonlinear fractional Fokker--Planck equations of McKean--Vlasov type in dimension \(D \gg 10\). Three difficulties occur at the same time. A tensor-product grid needs \(O(n^D)\) storage. The discrete Riesz operators of order \(\alpha\in(0,2)\) have condition numbers that grow like \(n^\alpha\). Finally, the curvature of the manifold of tensor trains (TT) of fixed rank behaves like \(\sigma_{\min}^{-1}\) and becomes large as soon as one singular value is small. Explicit projected integrators are then limited to steps \(\Delta t=O(\min(\sigma_{\min},n^{-\alpha}))\). The projector-splitting integrator is insensitive to small singular values, but its \(S\)-step runs backward in time, which is unstable when the operator is dissipative. We propose a rank-adaptive implicit--explicit integrator of basis-update & Galerkin (BUG) type, written directly in TT format (IMEX-TT-BUG). In each step, the separable fractional drift--diffusion part is treated implicitly in a left-to-right sweep of \(K\)-steps, and the convolution and reaction terms are treated explicitly. Each updated core enlarges a nested left basis of bond dimension at most \(2r\); a Galerkin problem for the last core is then solved in the linear space these bases span, and the result is truncated by TT-SVD. All substeps go forward in time, and the \(O((2r)^D)\) core of the Tucker version of BUG never appears. We prove a Frobenius-norm error bound in which the amplification constants depend neither on \(\sigma_{\min}^{-1}\) nor on \(\kappa_2(L)\). The bound also contains a projection residual that depends on the trajectory; one of its terms involves \(L\), and we state precisely when it stays bounded under mesh refinement. The linear part of a step costs \(O(Dnr^3+Dn\log n\,r^2N_{\mathrm{kry}})\) operations; evaluating the nonlinear term in TT format adds \(O(DnM_W^3r^6)\) with direct rounding, where \(M_W\) is the separation rank of the interaction kernel. We also describe a Julia implementation, based on circulant embedding of the Toeplitz matrices and a preallocated memory arena, designed so that the time loop performs no heap allocation. Experiments with a reference implementation in dimension up to \(20\) show first-order convergence with errors that do not depend on the mesh, stability for steps several hundred times the explicit limit, and correct damping of a mode whose singular value is \(10^{-12}\).
Keywords
All Related Versions
This publication has 14 references indexed in Scilit:
- A rank-adaptive robust integrator for dynamical low-rank approximationBIT Numerical Mathematics, 2022
- Time Integration of Tensor TrainsSIAM Journal on Numerical Analysis, 2015
- Alternating Minimal Energy Methods for Linear Systems in Higher DimensionsSIAM Journal on Scientific Computing, 2014
- A projector-splitting integrator for dynamical low-rank approximationBIT Numerical Mathematics, 2013
- Crank–Nicolson method for the fractional diffusion equation with the Riesz fractional derivativeJournal of Computational Physics, 2012
- On manifolds of tensors of fixed TT-rankNumerische Mathematik, 2011
- Tensor-Train DecompositionSIAM Journal on Scientific Computing, 2011
- Critical nonlinearity exponent and self-similar asymptotics for Lévy conservation lawsAnnales de l'Institut Henri Poincaré C, Analyse non linéaire, 2001
- On the condition numbers of large semidefinite Toeplitz matricesLinear Algebra and its Applications, 1998
- On the extreme eigenvalues of Hermitian (block) Toeplitz matricesLinear Algebra and its Applications, 1998