On the estimation of frequent itemsets for data streams: theory and experiments