Pruning kernel density estimators