March 27, 2015

Numerical and Computational Challenges in Science and Engineering Program

March 25, 11am, Room 210

Robert M. Freund, MIT Sloan School of Management

Two Topics on the Complexity of Convex Optimization, One Computational and One Theoretical

We consider the following quite general convex optimization problem format: $$\begin{array}{lclr} G_d: & {\rm minimize}_x & c^{T}x
\\ & \mbox{ s.t. } & Ax=b\\ & & x \in P, \\\end{array}$$ \noindent where $P$ is a closed convex set, not necessarily a cone. Linear
optimization is a special case of this format in the case when $P$ is polyhedral, and a modern complexity theory for linear optimization has been developed based on a notion of condition number $C(d)$ for the data $d=(A,b,c)$ for $G_d$. We develop some computational experience and test the practical relevance of this condition number theory, as applied to problem instances that one might encounter in practice, using the NETLIB suite of LP problem instances. We present empirical evidence that indicates that 42% of the variation in IPM iterations among the NETLIB suite problem instances is accounted for by log C(d) of the problem instances after pre-processing.

Our theoretical work concerns the complexity of convex optimization using geometry-based measures rather than algebraic or data-based measures for obtaining complexity bounds. We bound the complexity of computing an almost-optimal solution of $G_d$ in terms of natural geometry-based measures of the feasible region and the level-set of almost-optimal solutions, relative to a given {\em reference point} $x^r$ that might be close to the feasible region and/or the almost-optimal level set.

Back to Thematic Year Index