An approach to decrease the size of very large data sets in which the data are first 'binned' and then statistics such as the mean and variance/covariance are calculated on each bin. These statistics are then used to obtain a new sample in each bin to construct a reduced data set with the similar statistical properties to original one.