Representational limitations of invertible models