Bounding the fairness and accuracy of classifiers from population statistics