Global Optimality in Bivariate Gradient-based DAG Learning