Learning with Known Operators reduces Maximum Training Error Bounds