An Adaptive-rank Approach with Greedy Sampling for Multi-scale BGK Equations

In this paper, we propose a novel adaptive-rank method for simulating multi-scale BGK equations, based on a greedy sampling strategy. The method adaptively selects important rows and columns of the solution matrix and updates them using a local semi-Lagrangian solver. An adaptive cross approximation then reconstructs the full solution matrix. This extends our prior semi-Lagrangian adaptive-rank framework, developed for the Vlasov-Poisson system, to nonlinear collisional kinetic equations. Unlike step-and-truncate low-rank integrators, our greedy sampling approach avoids explicit low-rank decompositions of nonlinear terms, such as the local Maxwellian in the BGK operator. To ensure conservation, we introduce a locally macroscopic conservative correction that implicitly couples the kinetic and macroscopic systems, enforcing mass, momentum, and energy conservation. Through asymptotic analysis, we show that this correction preserves the full-grid scheme's asymptotic behavior, and that the proposed method is conditionally asymptotic-preserving in the low-rank setting. A key advantage of our approach is its use of a local semi-Lagrangian solver, which allows large time steps. This flexibility is retained in the macroscopic solver using high-order stiffly accurate diagonally implicit Runge-Kutta methods. The resulting nonlinear systems are solved efficiently using a Jacobian-free Newton-Krylov method, avoiding the need for preconditioning at modest CFL numbers. Each nonlinear iteration provides a self-consistent correction to a provisional kinetic solution, which serves as a dynamic closure for the macroscopic model. Numerical results demonstrate the method's accuracy in capturing shocks and its robustness across mixed-regime problems with wide-ranging Knudsen numbers.