Randomize low-rank Runge-Kutta methods (Q6967915)

This interesting paper studies randomized low-rank Runge-Kutta methods. More precisely, the authors study solutions say \(B(t)\) to large-scale matrix differential equations of the form \(\dot{B}(t)=F(B(t))\), \(B(0)=B_0\in \mathbb R^{m\times n}\). It turns out that often in applications in physics, uncertainty quantification and machine learning, various types of autonomous ordinary differential equations can be naturally viewed as such a matrix differential equation for certain \(F\). When the integers \(m,n\) become large as the authors point out, the solution of such equations becomes expensive and more so, it may not even be possible to store the entire matrix \(B(t)\) explicitly. To deal with this limitation, various model order reduction techniques are typically employed. One approach is based on exploiting the (approximate) low-rank structure of \(B(t)\), which arises, for example, from smoothness properties of the underlying physical system.\N\NThe work under review, studies a new class of numerical integrators for computing low-rank approximations to solutions of a matrix differential equation for example of the type above. The authors for example, combine an explicit Runge-Kutta method with repeated randomized low-rank approximations and under suitable assumptions, establish in particular, moment and tail bounds on the error of their randomized low-rank Runge-Kutta methods. Numerical experiments support their theoretical work.