
Maryam FazelResearch Scientist
Control And Dynamical Systems,

My PhD work concerns the problem of minimizing the rank of a matrix
over a convex set. In many engineering applications,
notions such as order, complexity, or dimension of a model or design
can be expressed as the rank of an appropriate matrix. If the set of
feasible models or designs is convex, then choosing the simplest model
can be cast as a Rank Minimization Problem (RMP). For example, a
lowrank matrix could correspond to a loworder controller for a
plant, a lowdegree statistical model fit for a random process, a
shape that can be embedded in a lowdimensional space, or a design
with a small number of components. It is thus not surprising that rank
minimization has diverse applications: following Occam's razor, we
often seek simple explanations and designs.
The RMP is known to be NPhard. If the variable in an RMP is a positive semidefinite matrix, a simple and effective relaxation commonly used is trace minimization. We prove that any general RMP involving nonpositive semidefinite or even nonsquare matrices, can be embedded in a larger, positive semidefinite RMP. This semidefinite embedding lemma is then used to extend the trace relaxation, as well as other methods using semidefiniteness, to general matrices. The generalized trace relaxation minimizes the dual of the matrix 2norm, referred to as the nuclear norm. Furthermore, we show that this relaxation is equivalent to minimizing the convex envelope of the rank function over the unit ball, thus providing theoretical support for its use. We also propose another heuristic based on (locally) minimizing the logarithm of the determinant (logdet) of the matrix, to refine the result of the trace relaxation and reduce the rank further. This heuristic is in fact equivalent to iteratively minimizing a weighted trace objective, and thus also admits an interpretation in terms of convex envelopes. Finally, applying these methods to application examples shows they work well in practice.
Related papers:
Network Utility Maximization with Nonconcave Utilities Using SumofSquares Method (M. Fazel, M. Chiang).
Application of Robust Model Validation Using SOSTOOLS to the Study of GProtein Signaling in Yeast (T.M. Yi, M. Fazel, X. Liu, T. Otitoju, A. Papachristodoulou, S. Prajna, and J. Doyle).
Stochastic Reachability Analysis in Complex Biological Networks (H. ElSamad, M. Fazel, X. Liu, A. Papachristodoulou, S. Prajna).
Amplitude and Sign Adjustment for Peak to Average Power Reduction (M. Sharif, C. Florens, M. Fazel, B. Hassibi).
Transient Analysis for Wireless Power Control (M. Fazel, D. Gayme, M. Chiang). To appear in Globecom'06.
Complexity and Robustness in Lattice Percolation Models and Linear Programs (M. Fazel, X. Liu, J. Doyle). In preparation.
Complexity in Automation of SOS Proofs: An Illustrative Example (D. Gayme, M. Fazel, J. Doyle). To appear in CDC'06.
A Convex Relaxation for Maximumpurity Quantum Encoding Channel Design (N. Yamamoto, M. Fazel). To be submitted.
A Methodology for Optimization of Shell and Tube Heat Exchangers in Series (A. Fakheri, M. Fazel).
*** Page under construction! ***
Last modified: Aug 2006