Figure 4

Classifiers for an independent arbitrary covariance Gaussian model. Classifiers for an independent arbitrary covariance Gaussian model with D=2 features and proper posteriors. Whereas the optimal Bayesian classifier (in green) is polynomial with expected true error 0.2007, the state-constrained optimal Bayesian classifier (in black) is quadratic with expected true error 0.2061 and the plug-in classifier (in red) is linear with expected true error 0.2078. These expected true errors are averaged over the posterior on the uncertainty class of states.