Linear-quadratic-Gaussian (LQG) procedures for controller design yield compensators of the same order as the plant. However, practical design considerations often dictate low-order controllers. This paper presents a procedure for reducing the order of LQG-designed controllers. The proposed method is based on the same performance index / used for designing the optimal regulator; the value of /for the system with the reduced- order controller is essentially identical to the value obtained with the full-order controller. It is shown that the observer normal coordinates play an important role in the proposed method. The numerical results presented illustrate the effectiveness of this method. ONTROLLER design using linear-quadratic-Gaussian (LQG) procedures yields compensators of the same order as the plant. Many physical systems, such as aeroservoelas tic systems, involve a large number of states and therefore result in high-order controllers. These high-order controllers are dif- ficult to implement and may be susceptible to reliability prob- lems. Therefore, it is imperative to find ways to reduce the order of these controllers. There are two basic approaches used to find low-order controllers: The first approach involves the approximation of the high-order system by a lower order one,1'3 and then appli- cation of the controller design procedure. In the second ap- proach, the controller for the full-order system is first de- signed, and only then are attempts made to reduce the controller order.4'7 The method proposed in this paper falls within the latter approach, and deals with a procedure for reducing the LQG-designed high-order controller into a lower order one. It differs from most other existing methods in that no attempts are made to balance the system in any form or to minimize any norm associated with the controller or the plant. The method may be reminiscent of the component cost analy- sis (CCA) method,8'9 but as will be shown later, it differs widely from it. Description of the Method General The method proposed assumes that the optimal control input u has been determined from optimal regulator theory, such that it minimizes a given quadratic performance index /. It is then assumed that all of the states, except for a single state, are used for feedback in the plant-regulator system, and J and eigenvalues for the closed-loop system are computed. This procedure is repeated, with a different single state omit- ted each time from the optimal feedback expression, until all of the states have been scanned in such a manner, and the resulting effects on / and on the eigenvalues are computed. Because / is minimal with full-state feedback, it is expected that the omission of a single state from the optimal feedback input will give rise to higher values of J. In this way, a numerical value can be attached to the importance of each state in its effect on the value of J'. Those states that affect J the most will be considered the most important states, whereas those states that have a negligible effect on /will be considered