The decoupling of different length scales makes it possible to have a self-consistent theory that only describes the relevant length scales for a given problem.
In particular, the decoupling of the class conditional feature distributions means that each distribution can be independently estimated as a one-dimensional distribution.