next up previous
Next: Flatness bias of AAFT Up: Fourier based surrogates Previous: Fourier based surrogates

Rescaled Gaussian linear process

The two null hypotheses discussed so far (independent random numbers and Gaussian linear processes) are not what we want to test against in most realistic situations. In particular, the most obvious deviation from the Gaussian linear process is usually that the data do not follow a Gaussian single time probability distribution. This is quite obvious for data obtained by measuring intervals between events, e.g. heart beats since intervals are strictly positive. There is however a simple generalisation of the null hypothesis that explains deviations from the normal distribution by the action of an invertible, static measurement function:
 equation1033
We want to regard a time series from such a process as essentially linear since the only nonlinearity is contained in the -- in principle invertible -- measurement function tex2html_wrap_inline1908.

Let us mention right away that the restriction that tex2html_wrap_inline1908 must be invertible is quite severe and often undesired. The reason why we have to impose it is that otherwise we couldn't give a complete specification of the process in terms of observables and constraints. The problem is further illustrated in Sec. 7.1 below.

The most common method to create surrogate data sets for this null hypothesis essentially attempts to invert tex2html_wrap_inline1908 by rescaling the time series tex2html_wrap_inline1972 to conform with a Gaussian distribution. The rescaled version is then phase randomised (conserving Gaussianity on average) and the result is rescaled to the empirical distribution of tex2html_wrap_inline1972. The rescaling is done by simple rank ordering. Suppose we want to rescale the sequence tex2html_wrap_inline1972 so that the rescaled sequence tex2html_wrap_inline1978 takes on the same values as some reference sequence tex2html_wrap_inline1980 (e.g. draws from a Gaussian distribution). Let tex2html_wrap_inline1980 be sorted in ascending order and tex2html_wrap_inline1984 denote the ascending rank of tex2html_wrap_inline1986, e.g. tex2html_wrap_inline1988 if tex2html_wrap_inline1986 is the 3rd smallest element of tex2html_wrap_inline1972. Then the rescaled sequence is given by
 equation1035
The amplitude adjusted Fourier transform (AAFT) method has been originally proposed by Theiler et al. [6]. It results in a correct test when N is large, the correlation in the data is not too strong and tex2html_wrap_inline1908 is close to the identity. Otherwise, there is a certain bias towards a too flat spectrum, to be discussed in the following section.


next up previous
Next: Flatness bias of AAFT Up: Fourier based surrogates Previous: Fourier based surrogates

Thomas Schreiber
Mon Aug 30 17:31:48 CEST 1999