next up previous contents
Next: Sample frequency and length Up: PRACTICAL ASPECTS Previous: Stationarity   Contents


Noise, drift and hum usually affect the estimation of the dimension and entropy. Therefore it is tempting to use filters. There are at least two arguments against this [Mayer-Kress, 1987,Broomhead and King, 1986]. First, a filter introduces a new set of equations, which may cause the estimated dimension to increase. Second, noise may induce a clear cross-over behaviour of the correlation integrals at the noise level. For distances less than the noise level, the estimated dimension will increase with the embedding dimension, while for distances greater than the noise level, the estimated dimension may saturate. This allows the identification of the beginning of the scaling region.