Introduction

Populations of neurons can form complex neuronal networks whereby feedback loops are responsible for the generation of oscillatory activity and determine its resonance frequency. In general, the frequency becomes slower and the amplitude larger with increasing extent of synchronized neuronal assemblies [12]. Even if only a small fraction of the total population of neurons is synchronized, the resulting oscillations may well overweight the rest of not-synchronized neurons [4].

Because of the great complexity, reciprocal interactions between excitatory and inhibitory mechanisms and dynamics of neural networks in sensorimotor areas, not only one but a great variety of oscillations in a broad frequency range exists. These networks can be in different states depending on the level of sensorimotor information processing. So for example during immobility or rest, a 10-Hz Rolandic rhythm (mu rhythm) dominates in most cases [5]. In the phase of planning e.g. a voluntary finger movement, the mu rhythm desynchronizes [9]. This means that more and more patches of neurons start to work independently and the number of coherently activated neurons as well as the amplitudes of the oscillations decrease. Shortly before the actual movement begins, small groups of neurons, becoming synchronously activated, may result in oscillations around 40 Hz (gamma activity). After termination of movement, oscillations around 20 Hz are found. Both the oscillations in beta and gamma bands are embedded in the desynchronized 10-Hz [8,9].

For a better understanding of cortical functioning and the significance of oscillations with distinct frequencies, it is important to investigate frequency, latency and duration of oscillatory bursts as accurately as possible. The gamma oscillations are of special importance because they may serve as cortical information carrier and are essential for establishing of rapid coupling between spatially separated cell assemblies [12].

Traditionally, the time course of frequency-specific changes of energy was quantified by band-pass filtering of signals in a chosen frequency band. Drawbacks of this approach include a priori choice of the frequency bands (usually by certain trial-and-error procedure), low resolution, and low selectivity to phenomena originating in given frequency band. This paper introduces a complete methodology for a high-resolution presentation and parametrization of event-related EEG in the time-frequency plane. Estimates of energy density are based upon the matching pursuit algorithm [6] with recently introduced stochastic dictionaries [3] (a brief comparison with some other time-frequency distributions is presented on a simulated signal). Owing to the direct parametrization of signal's structures, this approach allows also for construction of selective estimators of energy carried by structures originating in a given frequency band.