T. Serre et al., SEARCH FOR LOW-DIMENSIONAL NONLINEAR BEHAVIOR IN IRREGULAR VARIABLE-STARS - THE GLOBAL HOW RECONSTRUCTION METHOD, Astronomy and astrophysics, 311(3), 1996, pp. 833-844
We describe a powerful, recently developed method for nonlinear time-s
eries analysis. The method allows one to check whether the given signa
l has a dominant component that been generated by a low-dimensional no
nlinear dynamics. Furthermore it allows one to extract properties of t
his dynamics from the mere knowledge of the given scalar (single) obse
rved quantity. In the context of variable stars this is normally the l
uminosity of the star, or possibly its radial velocity. The method is
tailored to irregular signals and thus complements the classical techn
iques of analysis which apply only to multi-periodic signals. The ulti
mate purpose is to develop a better physical understanding of the puls
ations and to derive novel astrophysical constraints from irregular li
ght curves. Before applying the global flow reconstruction to signals
of unknown properties it is imperative to test it on a well known syst
em. For that purpose we have applied it to the well studied Rossler os
cillator which, we note, has a behavior that is similar to the one enc
ountered in the numerical model pulsations of W Virginis stars. For th
e analysis we allow ourselves only a short section of the temporal beh
avior of only one of the 3 Rossler variables to infer properties of th
e whole attractor. In order to make the test more realistic, Gaussian
noise has also been added to the Rossler oscillator data. The method i
s shown to perform very well, producing synthetic signals that, when c
haotic, are very close to the original one, with similar Fourier spect
ra. The map that is obtained from the data allows one then to quantify
the complexity of a chaotic signal, e.g. with the help of Lyapunov ex
ponents and fractal dimensions of the synthetic signals. These quantit
ies appear to be fairly robust. The minimum embedding dimension for th
e reconstruction with the first Rossler variable is found to be 3. Imp
ortantly, also, even though only one Rossler variable is assumed to be
known the method recovers the 'physical' dimension 3 of the RGssler b
and. The method works well even when large amounts of noise are added
to the signal prior to the analysis. In a twin paper we analyze the pu
lsations of a W Virginis model. Applications to observational data of
R Set and of AC Her are presented in companion papers.