PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Kolmogorov's Structure functions and model selection
Nikolai K. Vereshchagin and Paul M.B. Vitanyi
IEEE Transactions Information Theory Volume To appear, 2003.

Abstract

In 1974 Kolmogorov proposed a non-probabilistic approach to statistics, an individual combinatorial relation between the data and its model, expressed by the so-called ``structure function'' of the data. We show that the structure function determines all stochastic properties of the data in the sense of determining the best-fitting model at every model-complexity level. A consequence is this: minimizing the data-to-model code length (finding the ML estimator or MDL estimator), in a class of contemplated models of prescribed maximal (Kolmogorov) complexity, {\em always} results in a model of best fit, irrespective of whether the source producing the data is in the model class considered. In this setting, code minimization {\em always} separates optimal model information from the remaining accidental information, and not only with high probability. The function that maps the maximal allowed model complexity to the goodness-of-fit (expressed as minimal ``randomness deficiency'') of the best model cannot itself be monotonically approximated. However, the shortest one-part or two-part code above can---implicitly optimizing this elusive goodness-of-fit. We show that---within the obvious constraints---every graph is realized by the structure function of some data. We determine the (un)computability properties of the various functions contemplated and of the ``algorithmic minimal sufficient statistic.''

Postscript - Requires a viewer, such as GhostView
EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Computational, Information-Theoretic Learning with Statistics
Learning/Statistics & Optimisation
ID Code:126
Deposited By:Paul Vitányi
Deposited On:27 May 2004