Histograms revisited: when are histograms the best approximation method for aggregates over joins?