*2.3. Covariance Matrix Regularization*

Although the spatiotemporal beamformer, in theory, achieves optimal separation between target and non-target classes, in analogy to linear discriminant analysis [27], it does not always perform well on unseen data. The main challenge is to find a good estimator for the inverse covariance matrix *C*−<sup>1</sup> since the real underlying covariance matrix generating the data is, in principle, unknown.
