An Adaptive-Rank Approach with Greedy Sampling for Multi-Scale BGK Equations
William A. Sands, Jing-Mei Qiu, Daniel Hayes, Nanyi Zheng · 2025
Read An Adaptive-Rank Approach with Greedy Sampling for Multi-Scale BGK Equations by William A. Sands, Jing-Mei Qiu, Daniel Hayes, Nanyi Zheng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-17191.txt.
Introduction
William Sands, Jing-Mei Qiu, Daniel Hayes, and Nanyi Zheng proposed a faster simulation method in 2025 for multi-scale BGK equations, which model gases through colliding particles. Such simulations store a large solution matrix, and low-rank methods cut the cost by tracking only the most important rows and columns. The authors extend their earlier adaptive-rank framework from a collisionless plasma system to these nonlinear collisional equations.
Their greedy sampling strategy picks the important rows and columns directly, then updates them with a local solver and rebuilds the full solution by cross approximation. That avoids taking apart nonlinear terms such as the local equilibrium distribution. A conservative correction step keeps mass, momentum, and energy exactly accounted for, and the authors show through asymptotic analysis that the method keeps the right behavior in limiting regimes.
Catalog record
| CREATOR | William A. Sands, Jing-Mei Qiu, Daniel Hayes, Nanyi Zheng |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 83 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-17191
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Adaptive-Rank Approach with Greedy Sampling for Multi-Scale BGK Equations William A. Sands, Jing-Mei Qiu, Daniel Hayes, Nanyi Zheng Abstract In this paper, we propose a novel adaptive-rank method for simulating multi-scale BGK equations which is based on a greedy sampling technique. The method adaptively identifies important rows and columns of the solution matrix, and reduces computational complexity by updating only the selected rows and columns through a local solver. Once updated, the solution at selected rows and columns is used together with an adaptive cross approximation to reconstruct the entire solution matrix. The approach extends the semi-Lagrangian adaptive-rank approach, introduced in our previous work for the Vlasov-Poisson system [ 39 ] in several ways. Unlike the step-and-truncate low-rank integrators discussed in [ 13 ] , the greedy sampling technique considered in this paper avoids the need for explicit low-rank decompositions of nonlinear terms, such as the local Maxwellian in the BGK collision operator. We enforce mass, momentum, and energy conservation by introducing a new locally macroscopic conservative correction, which implicitly couples the kinetic solution to the solution of the corresponding macroscopic system. Using asymptotic analysis, we show that the macroscopic correction preserves the asymptotic properties intrinsic to the full-grid scheme and that the proposed method in the low-rank setting possesses a conditionally asymptotic-preserving property. Another unique advantage of our approach is the use of a local semi-Lagrangian solver, which permits large time steps compared to Eulerian schemes. This flexibility is retained in the macroscopic solver by employing high-order stiffly-accurate diagonally implicit Runge-Kutta methods. The resulting nonlinear macroscopic systems are solved efficiently using a Jacobian-free Newton-Krylov method, which eliminates the need for preconditioning at modest CFL numbers. Each iteration of the nonlinear macroscopic solver provides a self-consistent correction to a provisional low-rank kinetic solution, which is then used as a dynamic closure for the macroscopic system. Numerical results demonstrate the efficacy of the proposed method in capturing shocks and discontinuous solution structures. We also highlight its performance in a challenging mixed-regime problem, where the Knudsen number spans multiple orders of magnitude. Keywords : Sampling, adaptive-rank, cross approximation, multi-scale methods, BGK equation, conservation laws 1 Introduction The Bhatnagar–Gross–Krook (BGK) equation [ 3 ] is a popular model used to study the dynamics of rarefied gases when the mean-free path is large compared to characteristic length scales, causing fluid models such as the Euler and Navier Stokes equations to break down. It is a nonlinear relaxation model that approximates the collision dynamics of the well-known Boltzmann collision operator which requires the evaluation of a five-dimensional integral that is prohibitively expensive due to its intrinsic nonlinearity and high-dimensionality. In contrast, the BGK operator is a simpler nonlinear algebraic operator; yet it retains key properties of the Boltzmann operator, such as entropy dissipation and the preservation of collision invariants (e.g., mass, momentum, and energy) and various multiscale phenomena. As a result, the BGK equation serves as a valuable prototype for studying more complex collision operators. Low-rank tensor approximations in the context of kinetic simulations have attracted significant interest in recent years [ 13 ] . Such methods compress the time-dependent numerical solution using low-rank decomposition techniques developed for matrices and tensors, and are a promising tool to address the challenges associated with the curse of dimensionality. The primary research directions in this area can be interpreted as either dynamic low-rank (DLR) or step-and-truncate (SAT) methods. DLR methods [ 25 , 14 , 11 , 12 , 4 , 2 ] project the continuous equation onto a low-rank manifold using the low-rank bases from the solution, and then discretize the resulting equations to obtain evolution equations for the bases. In contrast, SAT methods [ 26 , 9 , 16 , 15 , 33 ] first discretize the kinetic equation and dynamically evolve a low-rank representation of the solution using tensor linear algebra. One of the challenges associated with the BGK collision operator in the context of low-rank methods is the need to evaluate a local Maxwellian, which generally does not admit a low-rank decomposition. To avoid these challenges, some of the existing low-rank methods for the BGK equation in the DLR [ 12 , 2 ] framework use a multiplicative splitting that is guided by the structure of the solution obtained from Chapman-Enskog analysis. Interpolatory low-rank methods, also known as CUR or cross approximation [ 34 ] , or an oblique projection in the DLR framework [ 7 , 8 ] , employ a sampling strategy based on the discrete empirical interpolation method (DEIM) that allows one to avoid forming explicit low-rank decompositions. This is particularly advantageous for the evaluation of the local Maxwellian in the BGK equation, where such methods were recently used to investigate compression of the entire six-dimensional distribution function [ 8 ] . In a recent work [ 39 ] , the adaptive cross approximation of matrices [ 34 ] was used to construct a non-split semi-Lagrangian adaptive-rank (SLAR) method that was applied to the nonlinear Vlasov-Poisson system [ 39 ] . Compared with existing low-rank approaches, the sampling-based adaptive-rank methods uses information from a residual to adaptively identify critical rows and columns that need to be updated. Once the rows and columns are identified, the low-rank approximation is then updated using low-cost recursive rank-one corrections. These rows and columns are a subset of the entire matrices, with a computational complexity similar to that of a low-rank solver. The method is also demonstrated to be highly flexible in the development of numerical schemes, can achieve high-order convergence in spatial and temporal discretizations, and can achieve the mass conservation property via a locally macroscopic conservation (LoMaC) type correction, similar to those in [ 15 , 17 ] . The goal of this work is to develop a new adaptive-rank approach for the multi-scale BGK equation that employs a greedy sampling strategy. The local solver used in this work is built upon a SL finite difference method [ 27 ] , which achieves high-order accuracy in both temporal and spatial discretizations and allows for rather large time steps with numerical stability. In [ 10 ] , such a local solver was shown to be asymptotic preserving (AP) and asymptotically accurate (AA) for the BGK equation in the semi-discrete sense. This work is built upon such a local solver, but with the greedy sampling technique and the SLAR method from our recent work [ 39 ] , which was shown to be flexible in its implementation, robust in its numerical performance, and efficient with a low-rank complexity. A key advantage of the adaptive-rank approach considered in this work is that explicit low-rank decomposition of local Maxwellians are no longer necessary, which greatly simplifies the construction of low-rank solvers for the BGK equation. This work also introduces a new LoMaC correction which can simultaneously achieve mass, momentum, and energy conservation. The original version of LoMaC proposed in [ 15 ] considered explicit schemes for the macroscopic system. To eliminate the CFL restriction from the macroscopic system, we consider implicit discretizations that couple the macroscopic and microscopic systems in order to define a dynamic and self-consistent correction of the low-rank kinetic solution. The approach we use is inspired by HOLO methods [ 36 , 35 ,