In this paper a Riemannian geometric framework is given for the minimum sensitivity design problem and its solution using a natural optimization criterion. The theory is then applied to the case of linear systems to generate a class of minimum sensitivity realizations related to the so-called balanced realizations. In particular, conditions are given which are applicable in the cases of fixed point and floating point implementations.