The method is based on the calculation of the
probabilities p_X in the space of Fourier-coefficients.
E.g., for class A
log(p_A)=-2*tr(D*C_A^(-1))+constant
where C_A is the averaged cross-spectrum
at a specific frequency (here, alpha in extreme
wide-band) and D is the cross-spectrum
of the single trial.
The feature space is one-dimensional:
f(D)=log(p_A)-log(p_B)
On the training set we found that the
results improve if we heavily regularize
the data by transforming them for after detrending as
x(t)=a*tanh(x(t)/a)
for each channel. We set
a=alpha*std(x)
where alpha is set to a very small number (.0001)
such that the data are essentially reduced to
the signs.