H-Consistency Bounds: Characterization and Extensions

Neural Information Processing Systems 

These are upper bounds on the zero-one estimation error of any predictor in a hypothesis set, expressed in terms of its surrogate loss estimation error.