An information-theoretic bound that sets a lower limit on classification error based on the mutual information between input and output.