Estimating the dimension of a manifold and finding local charts on it by using nonlinear singular value decomposition (Q2862933)

Suppose that a collection of points in \({\mathbb R}^N\) is given, which are sampled from an (unknown) embedded manifold. The authors present a method for determining the dimension of the manifold, as well as coordinate charts. In the case when the manifold is an affine subspace, it is shown how this may be accomplished using the singular value decomposition (SVD) of the \(P\times N\) matrix \(D\) whose rows are the coordinates \((x_1,\dots,x_N)\) of the \(P\) sample points. For more general manifolds, the method is extended by first choosing a neighborhood of the manifold and selecting the \(P\) sample points that lie within that neighborhood. Next, \(K\) functions \(f_k\) of \(N\) variables are chosen -- in the numerical example presented, the authors choose monomials of degree two. The matrix \(D\) is then extended to a \(P\times(N+K)\) matrix \(E\) by augmenting each row with the \(K\) additional values \(f_k(x_1,\dots,x_N)\). The SVD of the augmented matrix \(E\) now leads to a set of implicit equations for the manifold, which are then solved to obtain a local coordinate chart for the manifold.