Fig. 1: Methodological steps to extract the states’ temporal, spectral, and spatial profiles.

a Temporal dimension. Starting from the states’ time courses—the posterior probabilities—we epoched the time series with respect to the task information (stimulus onset) to define the trials using a [−0.2, 1.2] s window. We ran the generalized linear model (GLM) to statistically evaluate the increased or decreased modulation of the statewise fractional occupancy (activation) level compared to the baseline. We report the states’ average response across conditions (provided by the constant regressor). The straight lines in the plot indicate the time points where the state of the same color is significantly activated or deactivated (permutation test, number of permutations = 1000, p < 0.025, correction for multiple comparisons via maximum statistics). The significantly modulated task-relevant states are states 1, 2, 3, and 5. b Spectral dimension. We weighted the MEG recordings by the states’ time courses and used a multitaper to compute the spectral density of the weighted MEG data for each subject and state separately. From this, we extracted the power spectral density (PSD) over the brain, which constitutes the spatial map of activation of a state, and the coherence across regions which constitutes the phase-coupling network of a state. We reported the plot of the statewise PSD averaged across regions over the broad frequency spectrum (1–40 Hz)—the bold lines display the mean across subjects, and the lighter area includes the standard deviation across subjects. The same plot is reproduced for the phase-coupling averaged over all the connections in the broadband spectrum (1–40 Hz)—the bold lines show the mean coherence across subjects, and the area represents the standard deviation.