Why does Naïve Bayes often perform well despite the independence assumption being violated?

Naïve Bayes works well because correct class ranking matters more than accurate probability estimates.

Even when features are correlated, Naïve Bayes often produces correct decision boundaries because it relies on relative likelihoods rather than exact joint distributions. Correlations may distort probabilities but often preserve class ranking. Additionally, redundancy in features can reinforce signal strength.