On constrained maximum likelihood estimation with non-i.i.d. observations