Variable projection (VP) gained popularity as a technique for solving nonlinear least squares problems (NNLS) min_{x,y} ||F(x)y - b||^2. The VP approach is an itegrated algorithm in x that `projects out' the variable y at each iteration. The NLLS problem class had a range of applications, and we show that the 'projection' approach generalizes to a very broad setting, retaining the original flavour, with applications to nuisance parameter estimation, kernel learning, and non-smooth optimization.
Additional Information
Location: ESB 4133 Aleksandr Aravkin, University of Washington