Answers are grounded in this paper and cite their sources.
In the numerical study of dynamical systems, there are applications where equations or good initial guesses are not available, and a data-driven approach based on a set of observations of the system is useful. These observations take the form of a cloud of points in phase space. The observations lie on a smooth manifold in several important situations like stable invariant manifolds, some attractors, and parts of bifurcation surfaces. This paper describes an algorithm, based on multiparameter continuation methods, for finding a smooth manifold from a cloud of points that samples the manifold. The algorithm represents the manifold as an atlas of charts, with the distance between chart centers controlled by a local length scale. It is iterative, selecting a point from the boundary of the atlas at each step to add a new chart. Three variations are described, which use different definitions of the boundary of a set of points: nearest neighbor continuation, which is used by most manifold learning algorithms; shell continuation; and polyhedral continuation. Part of the result is chart adjacency information, which can be used to construct a mesh on the manifold with the chart centers as mesh. The results are also in a form that allows traditional manifold learning techniques for constructing low dimensional embeddings to be applied. There are many parallels between continuation and machine learning. This viewpoint provides a context for thinking about manifold learning.
No takes yet. Share an insight, caveat, or question.
Henderson et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: