We have realised and experimented a new method for spectral analysis for unevenly sampled signals based on three phases: preprocessing, extraction of principal eigenvectors and estimate of signal frequencies. This is done, respectively, by input normalization, nonlinear PCA neural networks, and the Multiple Signal Classificator algorithm. First of all, we have shown that neural networks are a valid tool for spectral analysis.

However, above all, what is really important is that neural networks, as realised in our neural estimator system, represent a new tool to face and solve a problem tied with data acquisition in many scientific fields: the unevenly sampling scheme.

Experimental results have shown the validity of our method as an alternative
to Periodogram, and in general to classical spectral analysis, mainly in
presence of few input data, few a priori information and high error
probability. Moreover, for unevenly sampled data, our system offers great
advantages with respect to *P*. First of all, it allows us to use a simple and
direct way to solve the problem as shown in all the experiments with
synthetic and Cepheid's real signals. Secondly, it is insensitive to the
frequency interval: for example, if we expand our interval in the SU Cygni
light curve, while our system finds the correct frequency, the Lomb's *P*
finds many spurious frequencies, some of them greater than the confidence
threshold, as shown in Figs. 33 and 34.

Furthermore, when we have a multifrequency signal, we can use our system also if we do not know the frequency number. In fact, we can detect one frequency at each time and continue the processing after the cancellation of the detected periodicity by IIR filtering.

A point worth of noting is the failure to find the right frequency in the
case of eclipsing binary for both our method and Lomb's *P*. Taking account of
the morphology of eclipsing light curve with two minima, this fact can not
be of concern because in practical cases the important thing is to have a
first guess of the orbital frequency. Further refinement will be possible
through a wise planning of observations. In any case we have under study
this point to try to find a method to solve the problem.

The authors would like to thank Dr. M. Rasile for the experiments related to the model selection and an unknown referee for his comments who helped the authors to improve the paper.

Copyright The European Southern Observatory (ESO)