Information criteria can help to determine the number of sources from EEG and MEG

A problem with analyzing EEG measurements is the reconstruction of the sources that generate the observed data. Most methods used to solve this problem are based on the assumption that the number of sources is known. Most methods to determine this number are based on the eigenvalues of the covariance matrix of the measured data, the so-called principal components. The assumption is that many of the smaller eigenvalues represent only noise. The decision, where to cut off the spectrum of eigenvalues, is often taken subjectively. Taking into account the properties of the noise, several criteria will be discussed, which allow a more objective choice. The effectiveness of these criteria is assessed by means of computer simulations and the analysis of measurements. Simulations were performed by adding noise to the signals that arise from two rotating current dipoles. The simulations show that the decision which criterion to use, has to be based on the available information about the noise. Besides, it depends on whether overestimation or underestimation of the number of sources would be less harmful. The criteria were also applied to measured data.