Geometric mean and geodesic regression on Grassmannians (Q472423): Difference between revisions

The main purpose of the paper is to generalize linear regression on Grassmann manifolds. A Grassmannian (\(k\) dimensional subspace of \(\mathbb{R}^n\)) is identified with the subclass of orthogonal projection matrices (idempotent symmetric \(n\times n\) matrices of rank \(k\)) equipped with the Euclidean metric. A first step is to give a formula for the geodesic distance between two points: \(d^2(P,Q)=-\frac{1}{4}\mathrm{tr}(\log^2((I-2Q)(I-2P)))\). Next, the geometric mean of a finite set of points \(\{P_i\}\) is considered. It is a solution of the equations expressing the necessary and sufficient conditions for minimizing the function \(\Phi(P)=\sum_i d^2(P,P_i)\). The ultimate fitting problem is then as follows. Given a finite set of time-labeled points \((P_i,t_i)\) on the Grassmannian, find a geodesic path \(\gamma(t)\) that minimizes \(E(\gamma)=\sum_i d^2(P_i,\gamma(t_i))\). A geodesic path through \(P\) has the form \(\gamma(t)=e^{t\Omega}Pe^{-t\Omega}\) with \(\Omega\) skew-symmetric so that the problem for \(\gamma\) can be reformulated as an optimization problem in \(P\) and \(\Omega\).