next up previous contents
Next: Reconstruction of phase space Up: PRACTICAL ASPECTS Previous: Sample frequency and length   Contents

Theiler correction

Autocorrelations (and interpolation) may cause severe underestimation of the dimension. Theiler [Theiler, 1986] suggests to discard samples that are close to the reference point up to the correlation time (measured in samples). This procedure seems to work even though it does not exclude all correlations. So it may be useful when the time series is rather short and decimation is unacceptable. Smith [Smith, 1988] states that this correction should not be used if one wishes to distinguish the scaling of ``true'' noise from fluctuations due to finite $N$ for small $r$.