Applying the proposed method basically confirms the classical results indicating different types of oscillations at a specific electrode location over the sensorimotor cortex. High resolution of the method results in an increased sensitivity to ERD and ERS. For example, by comparing Figures 4 and 3, we find that the maximum ERS in the gamma band, reaching 200% on the classical estimates in Figure 3, is an order of magnitude stronger when calculated in exactly the same way from the MP estimate of energy density. This result is doubled by a more adequate choice of the frequency interval, based upon the complete high-resolution picture of energy distribution, combined with a selective estimate of energy carried by structures originating within this frequency interval (Figure 7).
Detailed time-frequency information in Figure 5 allowed to distinguish two differently reacting oscillations within the alpha band: one component (10 Hz) showed only a weak attenuation with movement, whereas the other (between 11 and 11.5 Hz) was completely blocked. The 10 Hz component was found only in the referential recording (visible in Fig. 6) whereas the component above 11 Hz was clearly represented in the local reference (see Fig. 4). This can be explained by the effect of spatial high-pass filtering. The calculation of the local average reference results in damping of widespread activities around 10 Hz and in enhancing local activities around 11 Hz. Functional separation between lower and upper mu rhythm components has been suggested in [1].
Using the proposed procedure, it was further possible to study the time-frequency microstructure of beta and gamma bursts. Presented results show consistently that the maximum of the gamma bursts occurred during maximal desynchronization of alpha and beta frequency components. In contrast to the conventional processing of the ERD/ERS by band pass filtering (Fig. 3), a more detailed structure of gamma components was obtained. The presence of movement-related gamma oscillations could even be demonstrated in referential data, when adequate time-frequency resolution was achieved.
Our experience gained with analysis of the data from other subjects indicates, that proposed method gives a generally applicable, robust and high-resolution estimates. It can be also naturally extended to a topographical analysis by comparison of results for simultaneously recorded derivations, as was presented e.g. in [14] for sleep spindles.
Software for matching pursuit with stochastic dictionaries, used in this research, is available from http://brain.fuw.edu.pl/~mp.