Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

Convex Approximation and Optimization with Applications in Magnitude Filter Design and Radiation Pattern Synthesis

Peter William Kassakian

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2006-64
May 18, 2006

http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-64.pdf

Using convex optimization to help solve nonconvex problems in engineering is an area of intense research activity. In this thesis we study a specific nonconvex optimization problem called magnitude least-squares that has applications primarily in magnitude filter design. Solving the problem is difficult because of the existence of many local minima. We study it in depth, deriving methods for its approximate solution, proving equivalences among differing formulations, relating it to other well-studied problems, and proving estimates of the quality of the solutions obtained using the methods. We discover structure in the problem that distinguishes it from some more general problems of the same algebraic form. The structure is related to the fact that the variables in the problem are complex-valued. We exploit this structure when proving bounds on the quality of solutions obtained using semidefinite relaxation. In addition to a detailed and generally abstract study of this specific optimization problem, we solve several practical problems in signal processing. Some of the application examples serve to illustrate the applicability of the magnitude least-squares problem, and include multidimensional magnitude filter design, magnitude filter design for nonlinearly delayed tapped filters, and spatial filtering using arbitrarily positioned array elements. We also present several application examples that illustrate the modeling capabilities of convex optimization. We use least-squares techniques to reason about the capabilities of clustered arrays of loudspeakers to accurately synthesize radiation patterns. We also provide an elegant convex optimization-based procedure for designing linear-phase audio equalizers.

Advisor: Laurent El Ghaoui and David L. Wessel


BibTeX citation:

@phdthesis{Kassakian:EECS-2006-64,
    Author = {Kassakian, Peter William},
    Title = {Convex Approximation and Optimization with Applications in Magnitude Filter Design and Radiation Pattern Synthesis},
    School = {EECS Department, University of California, Berkeley},
    Year = {2006},
    Month = {May},
    URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-64.html},
    Number = {UCB/EECS-2006-64},
    Abstract = {Using convex optimization to help solve nonconvex problems in engineering is an area of intense research activity. In this thesis we study a specific nonconvex optimization problem called magnitude least-squares that has applications primarily in magnitude filter design. Solving the problem is difficult because of the existence of many local minima. We study it in depth, deriving methods for its approximate solution, proving equivalences among differing formulations, relating it to other well-studied problems, and proving estimates of the quality of the solutions obtained using the methods. We discover structure in the problem that distinguishes it from some more general problems of the same algebraic form. The structure is related to the fact that the variables in the problem are complex-valued. We exploit this structure when proving bounds on the quality of solutions obtained using semidefinite relaxation.

In addition to a detailed and generally abstract study of this specific optimization problem, we solve several practical problems in signal processing. Some of the application examples serve to illustrate the applicability of the magnitude least-squares problem, and include multidimensional magnitude filter design, magnitude filter design for nonlinearly delayed tapped filters, and spatial filtering using arbitrarily positioned array elements. We also present several application examples that illustrate the modeling capabilities of convex optimization. We use least-squares techniques to reason about the capabilities of clustered arrays of loudspeakers to accurately synthesize radiation patterns. We also provide an elegant convex optimization-based procedure for designing linear-phase audio equalizers.}
}

EndNote citation:

%0 Thesis
%A Kassakian, Peter William
%T Convex Approximation and Optimization with Applications in Magnitude Filter Design and Radiation Pattern Synthesis
%I EECS Department, University of California, Berkeley
%D 2006
%8 May 18
%@ UCB/EECS-2006-64
%U http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-64.html
%F Kassakian:EECS-2006-64