How does Naïve Bayes relate to probabilistic decision boundaries in high-dimensional sparse spaces?

In high-dimensional sparse spaces, Naïve Bayes forms nearly linear decision boundaries dominated by rare but informative features.

Sparse high-dimensional data (e.g., text) leads to most features being zero. NB decision boundaries are primarily driven by non-zero informative features. This creates sparse linear separators in log-space, making NB highly effective despite independence violations.