Computational Implications of Reducing Data to Sufficient Statistics