Limits on the majority vote accuracy in classifier fusion