The gap between an estimator's actual error and the Bayes-risk floor, representing the portion reducible by better data or algorithms.