German Title: Berechenbarkeit und Fraktale Dimension

PDF, English
Download (888Kb)  Terms of use 
Abstract
This thesis combines computability theory and various notions of fractal dimension, mainly Hausdorff dimension. An algorithmic approach to Hausdorff measures makes it possible to define the Hausdorff dimension of individual points instead of sets in a metric space. This idea was first realized by Lutz (2000). Working in the Cantor space of all infinite binary sequences, we study the theory of Hausdorff and other dimensions for individual sequences. After giving an overview over the classical theory of fractal dimension in Cantor space, we develop the theory of effective Hausdorff dimension and its variants systematically. Our presentation is inspired by the approach to algorithmic information theory developed by Kolmogorov and his students. We are able to give a new and much easier proof of a central result of the effective theory: Effective Hausdorff dimension coincides with the lower asymptotic algorithmic entropy, defined in terms of Kolmogorov complexity. Besides, we prove a general theorem on the behavior of effective dimension under rexpansive mappings, which can be seen as a generalization of Hölder mappings in Cantor space. Furthermore, we study the connections between other notions of effective fractal dimension and algorithmic entropy. Besides, we are able to show that the set of sequences of effective Hausdorff dimension s has Hausdorff dimension s and infinite sdimensional Hausdorff measure (for every 0<s<1). Next, we study the Hausdorff dimension (effective and classical) of objects arising in computability theory. We prove that the upper cone of any sequence under a standard reducibility has Hausdorff dimension 1, thereby exposing a Lebesgue nullset that has maximal Hausdorff dimension. Furthermore, using the behavior of effective dimension under rexpansive transformations, we are able to show that the effective Hausdorff dimension of the lower cone and the degree of a sequence coincide. For manyone reducibility, we prove the existence of lower cones of nonintegral dimension. After giving some natural' examples of sequences of effective dimension 0, we prove that every effectively closed set A of positive Hausdorff dimension admits a computable, surjective mapping onto Cantor space. We go on to study the complex interrelation between algorithmic entropy, randomness, effective Hausdorff dimension, and reducibility more closely. For this purpose we generalize effective Hausdorff dimension by introducing the notion of strong effective Hausdorff measure 0. We are able to show that not having strong effective Hausdorff measure 0 does not necessarily allow to compute a MartinLöf random sequence, a sequence of highest possible algorithmic entropy. Besides, we show that a generalization of the notion of effective randomness to noncomputable measures yields a very coarse concept of randomness in the sense that every noncomputable sequence is random with respect to some measure. Next, we introduce Schnorr dimension, a notion of dimension which is algorithmically more restrictive than effective dimension. We prove a machine characterization of Schnorr dimension and show that, on the computably enumerable sets, Schnorr Hausdorff dimension and Schnorr packing dimension do not coincide, in contrast to the case of effective dimension. We also study subrecursive notions of effective Hausdorff dimension. Using resourcebounded martingales, we are able to transfer the use of rexpansiveness to the resourcebounded case, which enables us to show that the SmallSpan Theorem does not hold for dimension in exponential time E. Finally, we investigate the effective Hausdorff dimension of sequences against which no computable nonmonotonic betting strategy can succeed. Computable nonmonotonic betting games are a generalization of computable martingales, and it is a major open question whether the randomness notion induced by them is equivalent to MartinLöf randomness. We are able to show that the sequences which are random with respect to computable nonmonotonic betting games have effective Hausdorff dimension 1, which implies that, from the viewpoint of algorithmic entropy, they are rather close to MartinLöf randomness.
Item Type:  Dissertation 

Supervisor:  AmbosSpies, Prof. Dr. Klaus 
Date of thesis defense:  22 December 2004 
Date Deposited:  09 Jun 2005 14:52 
Date:  2004 
Faculties / Institutes:  The Faculty of Mathematics and Computer Science > Department of Computer Science 
Subjects:  510 Mathematics 
Controlled Keywords:  Fraktale Dimension, HausdorffMaß, Berechenbarkeit, Entropie <Informationstheorie> 
Uncontrolled Keywords:  Algorithmische Informationstheorie , ZufälligkeitFractal Dimension , Hausdorff Measures , Computability , Entropy , Algorithmic Information Theory , Randomness 