Gaussian discriminant analysis where the covariant matrices are the same for all classes, but the means are different. Reduces the number of parameters, which may prevent Overfitting.

The decision boundaries are linear unlike in GDA

See here

The predictivion distribution $p(y|x)$ has the form of Logistic regression, i.e. a Softmax with a linear model. The difference is in the training as we have a prior, which we fit also.

null subspace algorithm