Spatio-Spectral Decomposition (SSD)

SSD is a spatial filter that maximizes the power of an oscillation in a target frequency band relative to the power in immediately adjacent flanking bands, rather than relative to a broadband or class distinction. Concretely it solves a generalized eigenvalue problem between the covariance of the signal band-passed to the band of interest and the covariance of the signal in the neighbouring side-bands, so the leading components are those whose activity is spectrally peaked exactly where a genuine rhythm sits and flat where only broadband noise sits.

The result is a small set of virtual channels with substantially improved signal-to-noise ratio for that rhythm, which makes SSD a valued preprocessing step before CSP, before connectivity estimation, or before extracting a narrowband envelope. Because it uses only the spectral contrast and no labels, it is unsupervised and does not overfit to a classification target.