#training-error

A binary classifier uses one Gaussian per class with covariance fixed to $\sigma^2 I$; only the class means and the class priors are fit by maximum likelihood. Claim: for any three distinct training points, choosing $\sigma^2$ small enough makes the training error zero.