@user2974951 What I'm looking for is NOT a statistical test but a means of decomposing a complex, position-dependent set of features (i.e. effectively a "sequence logo") into a global compositional component plus residual position-dependent features, in order to aid insight/interpretability. I did some algebra work in the mean time that seems to show that the Kullback-Leibler divergence possesses the necessary additivity properties to enable this.