How does Naïve Bayes behave in ultra-high-dimensional regimes with heavy-tailed feature distributions?

In ultra-high-dimensional heavy-tailed regimes, Naïve Bayes can become unstable due to extreme likelihood contributions.

Heavy-tailed distributions (e.g., Pareto-like) introduce rare but extreme feature values. In Naïve Bayes, these extreme values dominate log-likelihood sums, causing unstable posterior estimates. Robust variants using clipped likelihoods or heavy-tailed distributions like Student-t improve stability.