Order theoretic classification of monotone equivariant cluster methods (Q1065040)

The author considers an ordinal generalization of the Jardine-Sibson model for cluster analysis [\textit{N. Jardine} and \textit{R. Sibson}, Mathematical Taxonomy (1971; Zbl 0322.62065)]. Let M, N denote finite posets, and R the nonnegative reals. A cluster method is a map F: Res(M,R)\(\to Res(N,R)\) where Res(X,R) denotes the set of residuated maps from the poset X into R. Note that \(f\in Res(X,R)\) means that the preimage under f of a principal ideal of R is a principal ideal of X. There is a natural Galois connection between sets of cluster methods and subsets of Res(R): For any family of cluster methods \({\mathcal F}\), let \(\alpha\) (\({\mathcal F})\) denote the set of all \(\theta\in Res(R)\) for which \(F\circ \theta =\theta \circ F\) for every \(F\in {\mathcal F}\); for any subset \({\mathcal O}\) of Res(R), let \(\beta\) (\({\mathcal O})\) denote the set of all cluster methods F for which \(F\circ \theta =\theta \circ F\) for every \(\theta\in {\mathcal O}\). A cluster method F is said to be monotone equivariant if \(F\in \beta (Aut(R))\), where Aut(R) denotes the set of order automorphisms of R. The paper concerns itself with the proof of the rather startling and deep result that there are only 5 classes of Galois closed cluster methods containing the class of monotone equivariant methods. These classes are explicitly described, as are the corresponding classes of Galois closed residuated maps. Connections are then made with the original Jardine- Sibson model. There is not sufficient space in a review such as this to give detailed descriptions of the results, but it can be noted that the class of ''flat'' cluster methods emerges as a closed class, with Res(R) the associated closed class of mappings; furthermore, the class of closed maps associated with the monotone equivariant methods is shown to be the class of injective residuated maps on R. This is an important paper that provides the beginnings of a classification of cluster methods in terms of the confidence that can be placed on the values of the input dissimilarity coefficient.