Spatial Pattern versus Spatial Filter (Interpretation)

A spatial filter is the weight vector that is multiplied against the channels to extract a component (CSP filters, LDA weights, xDAWN filters, decoder weights all are filters). A spatial pattern is the reverse map: how a given source actually projects onto the sensors. These are not the same object, and the recurring error in the literature is to plot filter weights as if they were a scalp topography of the underlying brain source.

The distinction matters because a filter can place large weights on channels that carry no signal of interest, purely to subtract a correlated noise source; interpreting such a weight as neural activity at that electrode is simply wrong. To recover something physiologically interpretable, the extraction filter must be converted into its associated forward pattern, which for a linear model is obtained by multiplying the data covariance by the filter (and normalizing by the component's covariance). Only the pattern, not the filter, licenses statements about where the signal comes from.