The schemes for the generation of surrogate data in order to test the null
hypothesis of linear stochastic process undergoing nonlinear static transfo
rm are investigated as to their consistency in representing the null hypoth
esis. In particular, we pinpoint some important caveats of the prominent al
gorithm of amplitude adjusted Fourier transform surrogates (AAFT) and compa
re it to the iterated AAFT, which is more consistent in representing the nu
ll hypothesis. It turns out that in many applications with real data the in
ferences of nonlinearity after marginal rejection of the null hypothesis we
re premature and have to be reinvestigated taking into account the inaccura
cies in the AAFT algorithm, mainly concerning the mismatching of the linear
correlations. In order to deal with such inaccuracies, we propose the use
of linear together with nonlinear polynomials as discriminating statistics.
The application of this setup to some well-known real data sets cautions a
gainst the use of the AAFT algorithm. [S1063-651X(99)02509-X].