Efficient Fast Computing for U-Statistics via A-Optimal Subsampling