On the signature of squared distance matrices of metric measure spaces

Abstract

We consider the numbers of positive and negative eigenvalues of matrices of squared distances between randomly sampled i.i.d. points in a given metric measure space. These numbers and their limits, as the number of points grows, in fact contain some important information about the whole space. In particular, by knowing them, we can determine whether this space can be isometrically embedded in the Hilbert space. We show that the limits of these numbers exist almost surely, are nonrandom and the same for all Borel probability measures of full support, and, moreover, are naturally related to the operators defining the multidimensional scaling (MDS) method. We also relate them to the signature of the pseudo-Euclidean space in which the given metric space can be isometrically embedded. In addition, we provide several examples of explicit calculations or just estimates of those limits for sample metric spaces. In particular, for a large class of countable spaces (for instance, containing all graphs with bounded intrinsic metrics), we get that the number of negative eigenvalues increases to infinity as the size of samples grows. However, we are able to provide examples when the number of samples grows to infinity and the numbers of both negative and positive eigenvalues increases to infinity, or the number of positive eigenvalues is bounded (but as large as desired), and the number of positive ones is fixed. Finally, we consider the example of the universal countable Rado graph.