How does Naïve Bayes relate to probabilistic decision surfaces in exponential family representations?

Naïve Bayes decision surfaces become linear in log-space when using exponential family distributions.

Exponential family distributions express likelihoods as exp(θ·T(x) - A(θ)). Taking logs converts Naïve Bayes into linear combinations of sufficient statistics. This yields linear decision boundaries in transformed feature space, explaining its surprising effectiveness in high-dimensional problems like text classification.