New recovery based a posteriori error estimators∗