A nonlinear PCA based on manifold approximation (Q5943407)

The following problem is considered. Let \(X=\{ x^j\}_ {j=1,\ldots, N}\) be a set of \(N\) points to approximate in \(R^n.\) It is supposed that the data is centered. The principal components analysis builds a linear model of this set of points by approximating it with linear subspaces \(x- \sum_{k=1}^d (x, a^k) a^k =0\), where \(d\) is the dimension of the linear subspace and the set of axes \(\{ a^k\}\), \(k=1, \ldots, d\), is an orthogonal basis of this subspace. Axes are chosen to minimize the mean square distance between data and the linear subspace. The author proposes a generalization of principal components analysis. He considers the auto-associative composite (AAC) model and the auto-associative additive (AAA) model. The AAA model is a set of implicit equations \((x=F^d (\beta^d, x), d=1, \ldots,n)\) with \[ F^d(\beta^d, x) =\biggl( \sum_{k=1}^d S^{\alpha^k}, P^{a^k} \biggr) (x), \quad \beta^d =\{ (a^k, \alpha^k),\;k=1, \ldots,d \}, \] such that some properties hold true. Here \(P^a\) is the projection on the axis \([a],\) \(S^\alpha\) is the restoration function \(S^\alpha\colon t \in R \to S^\alpha (t) \in R^n\), continuously differentiable, and \(S^\alpha(0) =0.\) The author studies properties of these models. He presents an iterative algorithm for AAC model building and considers some examples. Some approximation properties of this model are derived and their consequence on the algorithm behaviour are emphasized: reduction of the mean square error at each iteration, and convergence in a finite number of steps.