Model reduction of large-scale systems by least squares (Q2491700): Difference between revisions

Systems with large state dimension are difficult to handle, because of obvious computational limitations. For this reason, one usually look for reduced order systems, whose response characteristic approximates the response of the original system. Section 2 of this paper reviews the basic model reduction methods for discrete-time and continuous-time single-input-single-output time-invariant systems: the balance reduction method (based on singular value decomposition) and the Krylov moment matching method. Both methods present advantages and disadvantages. Section 3 introduces a new model reduction method which combines some features of balance reduction and Krylov methods. It keeps the advantages of both: existence of global error bounds, preservation of stability, computational efficiency. The construction is performed for discrete-time systems, using least square approximation applied to the columns of the Henkel matrix and an oblique projection. It is therefore extended to continuous-time systems via bilinear transformation. The paper is concluded by comparison with other methods and numerical examples.