Refine
Has Fulltext
- yes (6)
Is part of the Bibliography
- yes (6)
Document Type
- Doctoral Thesis (6)
Language
- English (6) (remove)
Keywords
- Newton-Verfahren (6) (remove)
Institute
Many optimization problems for a smooth cost function f on a manifold M can be solved by determining the zeros of a vector field F; such as e.g. the gradient F of the cost function f. If F does not depend on additional parameters, numerous zero-finding techniques are available for this purpose. It is a natural generalization however, to consider time-dependent optimization problems that require the computation of time-varying zeros of time-dependent vector fields F(x,t). Such parametric optimization problems arise in many fields of applied mathematics, in particular path-following problems in robotics, recursive eigenvalue and singular value estimation in signal processing, as well as numerical linear algebra and inverse eigenvalue problems in control theory. In the literature, there are already some tracking algorithms for these tasks, but these do not always adequately respect the manifold structure. Hence, available tracking results can often be improved by implementing methods working directly on the manifold. Thus, intrinsic methods are of interests that evolve during the entire computation on the manifold. It is the task of this thesis, to develop such intrinsic zero finding methods. The main results of this thesis are as follows: - A new class of continuous and discrete tracking algorithms is proposed for computing zeros of time-varying vector fields on Riemannian manifolds. This was achieved by studying the newly introduced time-varying Newton Flow and the time-varying Newton Algorithm on Riemannian manifolds. - Convergence analysis is performed on arbitrary Riemannian manifolds. - Concretization of these results on submanifolds, including for a new class of algorithms via local parameterizations. - More specific results in Euclidean space are obtained by considering inexact and underdetermined time-varying Newton Flows. - Illustration of these newly introduced algorithms by examining time-varying tracking tasks in three application areas: Subspace analysis, matrix decompositions (in particular EVD and SVD) and computer vision.
In the generalized Nash equilibrium problem not only the cost function of a player depends on the rival players' decisions, but also his constraints. This thesis presents different iterative methods for the numerical computation of a generalized Nash equilibrium, some of them globally, others locally superlinearly convergent. These methods are based on either reformulations of the generalized Nash equilibrium problem as an optimization problem, or on a fixed point formulation. The key tool for these reformulations is the Nikaido-Isoda function. Numerical results for various problem from the literature are given.
In this thesis we consider a reactive transport model with precipitation dissolution reactions from the geosciences. It consists of PDEs, ODEs, algebraic equations (AEs) and complementary conditions (CCs). After discretization of this model we get a huge nonlinear and nonsmooth equation system. We tackle this system with the semismooth Newton method introduced by Qi and Sun. The focus of this thesis is on the application and convergence of this algorithm. We proof that this algorithm is well defined for this problem and local even quadratic convergent for a BD-regular solution. We also deal with the arising linear equation systems, which are large and sparse, and how they can be solved efficiently. An integral part of this investigation is the boundedness of a certain matrix-valued function, which is shown in a separate chapter. As a side quest we study how extremal eigenvalues (and singular values) of certain PDE-operators, which are involved in our discretized model, can be estimated accurately.
Consider the situation where two or more images are taken from the same object. After taking the first image, the object is moved or rotated so that the second recording depicts it in a different manner. Additionally, take heed of the possibility that the imaging techniques may have also been changed. One of the main problems in image processing is to determine the spatial relation between such images. The corresponding process of finding the spatial alignment is called “registration”. In this work, we study the optimization problem which corresponds to the registration task. Especially, we exploit the Lie group structure of the set of transformations to construct efficient, intrinsic algorithms. We also apply the algorithms to medical registration tasks. However, the methods developed are not restricted to the field of medical image processing. We also have a closer look at more general forms of optimization problems and show connections to related tasks.
Applications in various research areas such as signal processing, quantum computing, and computer vision, can be described as constrained optimization tasks on certain subsets of tensor products of vector spaces. In this work, we make use of techniques from Riemannian geometry and analyze optimization tasks on subsets of so-called simple tensors which can be equipped with a differentiable structure. In particular, we introduce a generalized Rayleigh-quotient function on the tensor product of Grassmannians and on the tensor product of Lagrange- Grassmannians. Its optimization enables a unified approach to well-known tasks from different areas of numerical linear algebra, such as: best low-rank approximations of tensors (data compression), computing geometric measures of entanglement (quantum computing) and subspace clustering (image processing). We perform a thorough analysis on the critical points of the generalized Rayleigh-quotient and develop intrinsic numerical methods for its optimization. Explicitly, using the techniques from Riemannian optimization, we present two type of algorithms: a Newton-like and a conjugated gradient algorithm. Their performance is analysed and compared with established methods from the literature.