Robust and Non-convex Principal Components
We present a mathematical analysis for non-convex robust principal components. We show that under a generic condition on inliers and outliers, the energy landscape of the underlying optimization problem is nice enough and this observation leads to performance guarantees. The robust components add interpretability over methods which try to robustly find the whole underlying subspace at once. This is a joint work with Tyler Maunu, which is different than the recent work on the Fast Median Subspace.