predictive-kl

This paper proves that unbounded gaps between update maps can coexist with vanishing predictive KL divergence in stationary symmetric Gaussian HMMs, showing that internal computational mismatches do not necessarily imply predictive failure.