High-Dimensional Variable Selection With Reciprocal L1-Regularization