The problem attacked in this paper is to obtain the smallest model set which is described by nominal model, parametric uncertainty bound, unstructured uncertainty bound and consistent with a number of noise-free input-output data, under the condition that the denominator of the nominal model is prescribed. For the compatibility with the available robust control theory, the unstructured uncertainty is measured by H infinity-norm, while the parametric uncertainty is by parameter variation interval. It is shown that the unstructured uncertainty bound will generally increase if the nominal model error uncertainty is completely regarded as unstructured. Moreover, the identification problem is reduced to a convex optimization problem which is computationally tractable.