Properties of Euclidean and non-Euclidean distance matrices |
| |
Authors: | J.C. Gower |
| |
Affiliation: | Statistics Department Rothamsted Experimental Station Harpenden, Hertfordshire AL5 2JQ, United Kingdom |
| |
Abstract: | A distance matrix D of order n is symmetric with elements , where dii=0. D is Euclidean when the quantities dij can be generated as the distances between a set of n points, X (n×p), in a Euclidean space of dimension p. The dimensionality of D is defined as the least value of p=rank(X) of any generating X; in general p+1 and p+2 are also acceptable but may include imaginary coordinates, even when D is Euclidean. Basic properties of Euclidean distance matrices are established; in particular, when ρ=rank(D) it is shown that, depending on whether eTD?e is not or is zero, the generating points lie in either p=ρ?1 dimensions, in which case they lie on a hypersphere, or in p=ρ?2 dimensions, in which case they do not. (The notation e is used for a vector all of whose values are one.) When D is non-Euclidean its dimensionality p=r+s will comprise r real and s imaginary columns of X, and (r, s) are invariant for all generating X of minimal rank. Higher-ranking representations can arise only from p+1=(r+1)+s or p+1=r+ (s+1) or p+2=(r+1)+(s+1), so that not only are r, s invariant, but they are both minimal for all admissible representations X. |
| |
Keywords: | |
本文献已被 ScienceDirect 等数据库收录! |
|