Convex optimization methods for model reduction

Abstract

Model reduction and convex optimization are prevalent in science and engineering applications. In this thesis, convex optimization solution techniques to three different model reduction problems are studied.Parameterized reduced order modeling is important for rapid design and optimization of systems containing parameter dependent reducible sub-circuits such as interconnects and RF inductors. The first part of the thesis presents a quasi-convex optimization approach to solve the parameterized model order reduction problem for linear time-invariant systems. Formulation of the model reduction problem as a quasi-convex program allows the flexibility to enforce constraints such as stability and passivity in both non-parameterized and parameterized cases. Numerical results including the parameterized reduced modeling of a large RF inductor are given to demonstrate the practical value of the proposed algorithm.A majority of nonlinear model reduction techniques can be regarded as a two step procedure as follows. First the state dimension is reduced through a projection, and then the vector field of the reduced state is approximated for improved computation efficiency. Neither of the above steps has been thoroughly studied. The second part of this thesis presents a solution to a particular problem in the second step above, namely, finding an upper bound of the system input/output error due to nonlinear vector field approximation. The system error upper bounding problem is formulated as an L2 gain upper bounding problem of some feedback interconnection, to which the small gain theorem can be applied. A numerical procedure based on integral quadratic constraint analysis and a theoretical statement based on L2 gain analysis are given to provide the solution to the error bounding problem. The numerical procedure is applied to analyze the vector field approximation quality of a transmission line with diodes.

(Cont) The application of Volterra series to the reduced modeling of nonlinear systems is hampered by the rapidly increasing computation cost with respect to the degrees of the polynomials used. On the other hand, while it is less general than the Volterra series model, the Wiener-Hammerstein model has been shown to be useful for accurate and compact modeling of certain nonlinear sub-circuits such as power amplifiers. The third part of the thesis presents a convex optimization solution technique to the reduction/identification of the Wiener-Hammerstein system. The identification problem is formulated as a non-convex quadratic program, which is solved by a semidefinite programming relaxation technique. It is demonstrated in the thesis that the formulation is robust with respect to noisy measurement, and the relaxation technique is oftentimes sufficient to provide good solutions. Simple examples are provided to demonstrate the use of the proposed identification algorithm.