Globally Optimal On-line Learning Rules