Spatial Filter Regularization

Spatial filters such as CSP are computed from covariance matrices, and covariance estimated from a handful of noisy trials in a space of tens to hundreds of channels is badly conditioned, so the resulting filters overfit and generalize poorly. Regularization stabilizes the estimate by biasing it toward a simpler or more plausible solution. The most common form is shrinkage of the covariance matrix toward a scaled identity, with the mixing weight set analytically by the Ledoit-Wolf formula so no cross-validation search is needed.

Beyond shrinkage, regularization can inject prior information: penalties that pull filters toward spatial smoothness, toward channel sparsity for electrode selection, or toward covariance structure borrowed from other subjects or sessions (composite and regularized CSP for transfer learning). All of these trade a small increase in bias for a large reduction in variance, which is precisely the right trade when trials are scarce, the defining condition of a short BCI calibration.