skip to content

Geometric numerical integration for optimisation

Abstract: In this thesis, we study geometric numerical integration for the optimisation of various classes of functionals. Numerical integration and the study of systems of differential equations have received increased attention within the optimisation community in the last decade, as a means for devising new optimisation schemes as well as to improve our understanding of the dynamics of existing schemes. Discrete gradient methods from geometric numerical integration preserve structures of first-order gradient systems, including the dissipative structure of schemes such as gradient flows, and thus yield iterative methods that are unconditionally dissipative, i.e. decrease the objective function value for all time steps. We look at discrete gradient methods for optimisation in several settings. First, we provide a comprehensive study of discrete gradient methods for optimisation of continuously differentiable functions. In particular, we prove properties such as well-posedness of the discrete gradient update equation, convergence rates, convergence of the iterates, and propose methods for solving the discrete gradient update equation with superior stability and convergence rates. Furthermore, we present results from numerical experiments which support the theory. Second, motivated by the existence of derivative-free discrete gradients, and seeking to solve nonsmooth optimisation problems and more generally black-box problems, including for parameter optimisation problems, we propose methods based on the Itoh--Abe discrete gradient method for solving nonconvex, nonsmooth optimisation problems with derivative-free methods. In this setting, we prove well-posedness of the method, and convergence guarantees within the nonsmooth, nonconvex Clarke subdifferential framework for locally Lipschitz continuous functions. The analysis is shown to hold in various settings, namely in the unconstrained and constrained setting, including epi-Lipschitzian constraints, and for stochastic and deterministic optimisation methods. Building on the work of derivative-free discrete gradient methods and the concept of structure preservation in geometric numerical integration, we consider discrete gradient methods applied to other differential systems with dissipative structures. In particular, we study the inverse scale space flow, linked to the well-known Bregman methods, which are central to variational optimisation problems and regularisation methods for inverse problems. In this setting, we propose and implement derivative-free schemes that exploit structures such as sparsity to achieve superior convergence rates in numerical experiments, and prove convergence guarantees for these methods in the nonsmooth, nonconvex setting. Furthermore, these schemes can be seen as generalisations of the Gauss-Seidel method and successive-over-relaxation. Finally, we return to parameter optimisation problems, namely nonsmooth bilevel optimisation problems, and propose a framework to employ first-order methods for these problems, when the underlying variational optimisation problem admits a nonsmooth structure in the partial smoothness framework. In this setting, we prove piecewise differentiability of the parameter-dependent solution mapping, and study algorithmic differentiation approaches to evaluating the derivatives. Furthermore, we prove that the algorithmic derivatives converge to the implicit derivatives. Thus we demonstrate that, although some parameter tuning problems must inevitably be treated as black-box optimisation problems, for a large number of variational problems one can exploit the structure of nonsmoothness to perform gradient-based bilevel optimisation.