The false nearest neighbor method introduced by Kennel et al.. [Phys. Rev.
A 45, 3403 (1992)] is revisited and modified in order to ensure a correct d
istinction between low-dimensional chaotic data and noise. Still, correlate
d noise processes can yield vanishing percentages of false nearest neighbor
s for rather low embedding dimensions and can be mistaken for deterministic
signals. Therefore, the false nearest neighbors method has always to be co
mbined with a surrogate data test. [S1063-651X(99)08510-4].