The adaptive lasso in high-dimensional sparse heteroscedastic models