On tensor decompositions, maximum likelihood estimation, and causal inference