Abstract:
This thesis is concerned with the adaptive approximation of the stochastic Landau-Lifshitz-Gilbert equation (SLLG), which models the dynamics of a ferromagnetic body at elevated temperatures. The SLLG is a non-linear stochastic partial differential equation which possesses an inherent non-convex side constraint. In a first step, space-time and statistical adaptivity is addressed to a convection-dominated SPDE with linear drift, and the stochastic harmonic map heat flow (HMHF) to the unit sphere. Secondly, these adaptive concepts are applied to the SLLG. The latter two equations possess a weak martingale solution, rather than a probabilistically strong solution as for the first problem; however our concept of space-time and statistical adaptivity is based on distributions rather than single trajectories, and therefore is applicable also there. The thesis is splitted in three parts. The first part is concerned with ordinary differential equations (ODEs) in which we focus on local extrapolation methods to adapt the local time step size. In the second part, we repeat in the literature already existing adaptive time stepping methods based on the explicit Monte Carlo Euler method for weak approximation of the SDE. These strategies are limited to lower order systems of SDEs due to the use of Kolmogorov’s backward equation. Inspired by this, we perform a local non-parametric high-dimensional density estimation based on Monte-Carlo (MC) sampling. In the third part we propose a new adaptive time stepping method to numerically solve the SLLG where local step sizes are chosen in regard of the distance between empirical laws of current Euler iterates, and extrapolated data. This histogram-based estimator uses a data-driven partitioning of the high-dimensional state space, and efficient sampling by bootstrapping to save computer resources. Time adaptivity is then complemented by a local refinement/coarsening strategy of the spatial mesh via a stochastic version of the Zienkiewicz-Zhu estimator. Computational experiments compare the efficiency of the proposed adaptive space-time and statistical strategies of already in the literature existing stable discretizations of the SLLG. Of particular interest is the choice of the distance to measure closeness of subsequent laws (in time and space), having an possible impact on the empirical variance of standard estimators; especially in the case of discrete blow-up dynamics of the SLLG.