Scaling invariance and adaptivity (Q5954978)

The authors show that scaling invariance has the unique property amongst general transformations, that the properties of discretization and invariance commute. They consider the autonomous differential equation \(d\mathbf u /dt=\mathbf f (\mathbf u)\). It is supposed the independent variable \(t\) to be itself a function of a fictive computational variable \(\tau \) such that \(dt/d\tau =g(\mathbf u)\). Then using Sundman transformation one gets the transformed equation \(d\mathbf u /d\tau =g(\mathbf u)\mathbf f (\mathbf u)\). This is applied in order to seek to choose new coordinates so that the transformed equation in the transformed variables is in some ways easier to solve either analytically or numerically. In particular, singular solutions in the variable \(t\) should ideally become regular in the rescaled variable \(\tau \). It turns out that the most significant feature is that the resulting numerical solutions accurately reproduce self-similar solutions for arbitrary large times with a discretization error that does not grow with time. This feature occurs both with a priori and a posteriori scaling strategies. Some significant questions exist. For instance, what is the effect of using a rescaling strategy based on successive halving of the step-size? Furthermore, what is the best adaptive approach in problems (such as Hamiltonian or reversible problems) for which scaling invariance is only part of the underlying geometry of the solution?