Principal components along quiver representations (Q6047299)

The paper under review deals with representations of a quiver (=directed graph, with loops and multiple edges allowed) \(Q\). It has two main goals. The first is to give a practical way of describing the space of sections \(\Gamma(Q;\mathbf{A}_{\bullet})\) of a quiver representation \(\mathbf{A}_{\bullet}\) (that is, the assignment of a vector space \(\mathbf{A}_v\) to any vertex \(v\) of \(Q\) and a linear map for any arrow). A section of \(\mathbf{A}_{\bullet}\) is the data of a vector at each vertex that is compatible with the linear maps between vertices. The paper gives algorithms to perform the operations considered, with complexity estimates. The second goal is to give an answer to a variational problem over \(\Gamma(Q;\mathbf{A}_{\bullet})\), called Principal Component Analysis (PCA). In this situation, the base field is the field of real numbers, \(\mathbb{R}\). For a given set of vectors in \(\mathbf{A}_{\bullet}\), one can form the \emph{covariance matrix} \(S\) (Definition 6.1). Then, for each \(r\), there is the optimisation problem. \[ \max_X\mathrm{tr}(X^{\mathrm{T}}ST) \text{ with } X^{\mathrm{T}}X=\mathrm{id}_r. \] The restriction here is that the columns of \(X\) have to be sections of \(\mathbf{A}_{\bullet}\) (making the connection with the first problem). The solution to the first question proceeds in two steps. First, the authors pass from any quiver with any representation to an acyclic quiver with a representation. Then, a procedure is described to make this quiver a rooted tree and to modify the representation accordingly. These two operations preserve the space of sections. The solution to the second question relies on a generalisation of the singular value theorem. The paper uses interesting decompositions and properties of quivers: \begin{itemize} \item[1.] Strongly-connected quivers (\S2), \item[2.] Ear decomposition of a quiver (Definition 2.1), \item[3.] Arborescence induced by an ear decomposition (Definition 2.4), \item[4.] Acyclic reduction (Definition 3.1), \item[5.] Acyclification of a representation (Definition 3.4), \item[6.] Arboreal replacement (\S4). \end{itemize} There are applications of the results to other fields, such as biology, with gene expression measurements for example, and in the emerging field of learning quiver representations.