Document Type



Doctor of Philosophy


Industrial Engineering

First Adviser

Curtis, Frank E.

Other advisers/committee members

Curtis, Frank E.; Nocedal, Jorge; Robinson, Daniel P.; Terlaky, Tamas; Zuluaga, Luis F.


Primal-dual active-set (PDAS) methods are developed for solving quadratic optimization problems (QPs). Such problems arise in their own right in optimal control and statistics–two applications of interest considered in this dissertation–and as subproblems when solving nonlinear optimization problems. PDAS methods are promising as they possess the same favorable properties as other active-set methods, such as their ability to be warm-started and to obtain highly accurate solutions by explicitly identifying sets of constraints that are active at an optimal solution. However, unlike traditional active-set methods, PDAS methods have convergence guarantees despite making rapid changes in active-set estimates, making them well suited for solving large-scale problems.Two PDAS variants are proposed for efficiently solving generally-constrained convex QPs. Both variants ensure global convergence of the iterates by enforcing montonicity in a measure of progress. Besides identifying an estimate set estimate, a novel uncertain set is introduced into the framework in order to house indices of variables that have been identified as being susceptible to cycling. The introduction of the uncertainty set guarantees convergence of the algorithm, and with techniques proposed to keep the set from expanding quickly, the practical performance of the algorithm is shown to be very efficient. Another PDAS variant is proposed for solving certain convex QPs that commonly arise when discretizing optimal control problems. The proposed framework allows inexactness in the subproblem solutions, which can significantly reduce computational cost in large-scale settings. By controlling the level inexactness either by exploiting knowledge of an upper bound of a matrix inverse or by dynamic estimation of such a value, the method achieves convergence guarantees and is shown to outperform a method that employs exact solutions computed by direct factorization techniques.Finally, the application of PDAS techniques for applications in statistics, variants are proposed for solving isotonic regression (IR) and trend filtering (TR) problems. It is shown that PDAS can solve an IR problem with n data points with only O(n) arithmetic operations. Moreover, the method is shown to outperform the state-of-the-art method for solving IR problems, especially when warm-starting is considered. Enhancements to themethod are proposed for solving general TF problems, and numerical results are presented to show that PDAS methods are viable for a broad class of such problems.