Previous slide Next slide Back to the first slide View text version


Of course while we are at it we may chose to optimize over the tangent space (what will be called V).

The tangent space is a sub-space of the image space. So we are looking for a matrix which is not full rank.

The non-full rank matrix that will do the best job of approximating the full space is built from the first few eigenvectors of the covariance matrix.