How does Naïve Bayes relate to KL divergence minimization in generative model fitting?

Naïve Bayes can be interpreted as minimizing KL divergence between empirical data distribution and a factorized generative model.

From an information-theoretic perspective, fitting Naïve Bayes is equivalent to projecting the true joint distribution onto a restricted family of distributions where features are conditionally independent given the class. This projection minimizes KL(P_data || P_NB). The independence assumption restricts the hypothesis space, making optimization tractable but introducing approximation bias when dependencies exist.