Research Codes

• Soft-Thresholding and Max-Det Matrix Completion for Graphical Lasso. In our ICML 2018 paper, we describe an algorithm that recovers the graphical lasso estimator of Friedman et al. (2008) in $$O(n^2/p + n)$$ time over $$p$$ parallel processors. Given the sample matrix $$\mathbf{X}=[\mathbf{x}_{1},\ldots,\mathbf{x}_{N}]_{i=1}^{N}$$, we first compute the soft-thresholded sample covariance matrix $$C_{\lambda}$$ in $$O(n^2/p)$$ time and $$O(n\cdot N)$$ memory. Then, we solve an associated maximum determinant matrix completion problem using $$C_{\lambda}$$ as input in $$O(n)$$ time and memory. [Associated paper].
• Modified SeDuMi. A modified version of SeDuMi with identical syntax for large-and-sparse low-rank semidefinite programs. If the $$n \times n$$ solution matrix is low-rank and if the problem data is "sufficiently sparse", then each SeDuMi iteration costs $$O(n^{3})$$ time and $$O(n^{2})$$ memory. Lorentz cone is not (yet) supported. [Associated paper].
• fastlitz - Realistic litz wire characterization. Simulates litz wires in three-dimensions using the PEEC formulation. Computes the resistance of a wire, and the induced eddy loss when exposed to a magnetic field. Comes with a nice graphical user interface. [Associated paper] UPDATE 04/14: Removed dependencies on MATLAB Statistics Toolbox.