Paperlane

no. 14 · Friday 13 November 2026

Robust High-Order BUG Integrators without Basis Augmentation

Shiheng Zhang

math.NA

Abstract

We develop robust high-order basis-update and Galerkin (BUG) integrators for dynamical low-rank approximation without basis augmentation or rank truncation. The basis and core updates use the same approximate increment. We establish a two-sided relaxed-exactness estimate that controls the error in the low-rank approximation by the increment error. Under the stated assumptions on the vector field and its normal component, sufficiently accurate increments yield local and global error bounds whose constants and step-size restrictions are independent of small singular values. Using explicit Runge–Kutta increments, we obtain robust orders two, three, and four for the explicit midpoint method, Heun's third-order method, and Merson's fourth-order method, respectively. Every internal stage and final approximation has rank at most the prescribed rank $r$. We also prove that these orders are retained under arbitrary basis augmentation followed by best rank-$r$ SVD truncation at the internal stages and the endpoint, without a separate assumption on the truncation error. Numerical experiments confirm the convergence orders and show comparable performance with and without augmentation.

fields
Numerical Analysis
pdf ↗ page arXiv:2611.15175 ↗