An improved estimator of dimension and some comments on providing confidence intervals (Q1197516)

The aim of the present paper is to develop an improved estimator for the dimension of an attractor of a dynamical system, based on the Grassberger-Procaccia idea [\textit{P. Grassberger} and \textit{I. Procaccia}, Phys. Rev. Lett. 50, 346-349 (1983)] of calculating interpoint distances from a trajectory. Although it is remarkably simple and widely applied, the Grassberger-Procaccia method for estimating the dimension of an attractor has the drawback of not providing an error (or a confidence interval) for the estimate. The statistical error of the dimension estimates using the Grassberger-Procaccia method is a topic of considerable research interest, and certain results concerning the bias and the variation of estimates with the sample size have been reported [\textit{J. B. Ramsey} and \textit{H.-J. Yuan}, Phys. Lett. A 134, 287-297 (1989) and \textit{J. Theiler}, Phys. Rev. A 41, 3038-3051 (1990)]. Unlike these approaches, the paper under review is concerned with the more fundamental issue of improving the estimator of dimension. As opposed with the conventional implementations of the Grassberger-Procaccia method, the novel estimator is formulated by considering directly the distribution of the interpoint distances from a trajectory, instead of going through the step of calculating first the correlation function. The new estimator of dimension is shown to have significant advantages as compared with the conventional implementations of the Grassberger- Procaccia method. It eliminates the operator dependent characteristics (such as choosing a scaling region), it uses the maximum amount of information from a time series, and it can provide a form of confidence interval under ideal circumstances. However, the problem of providing confidence intervals is not completely solved, as for higher dimension (greater than about 3.5) attractors, the dimension estimate has a bias which is dependent on the structure of the attractor. Henceforth, to provide confidence intervals for an estimate of dimension (which is a property of the distribution of interpoint distances on the attractor), one must either know or assume the family of distributions for the interpoint distances. In order to avoid technical details, throughout this research, the author has considered a simple model of an attractor and dynamics, for which he has derived the distribution of interpoint distances. The dynamical system model which has been considered, generalizes the well-known baker's map to a system having an attractor of the form \(V\times Z\), where \(Z\) is some bounded, connected subset of a smooth manifold, and \(V\) is a Cantor-like set, that is, a \((k,\lambda)\) self-similar set.