Optimal Regularization for Performative Learning