A learning bound showing that an algorithm's error is close to the best possible error achievable by any method with access to the true data distribution.