Abstract: In this paper an attempt is made to probe into both the mechanism and some new computational methods of model-reduction. The principle of an optimal allotment of information resources in a reduced-order model and the technique of separate treatment for middle-frequency and low-frequency ranges are presented. Several methods and algorithms based upon these concepts are offered. Results of computation show convinoingly that these new methods of model-reduction are superior to the current methods from the point-of-view of a better similitude to the original model over a wide range of frequency. They are also simple to use and versatile in engineering application.