On some pattern-reduction matrices which appear in statistics (Q1066593)

If X is a matrix vec X is a column vector formed by the columns of X in natural order. The vec permutation matrix \(I_{(p,q)}\) is defined by its action, vec \(A_{p\times q}=I_{(p,q)}vec A'\). A (variable) matrix X has linear pattern p (denoted by \(X\in {\mathcal M}_ p)\) if it has p functionally independent elements, arranged in some order as \(vec_ p X\), and the remaining elements are linear combinations of the elements of \(vec_ p X\). Then, \(X\in {\mathcal M}_ p\) if and only if vec X\(=G_ pvec_ p X\), where \(G_ p=Q_ p(I'\!_ pA'\!_ p)'\) for some constant matrix \(A_ p\), where \(Q_ p\) is a permutation matrix. The first part of the paper presents known and new results on: (1) Basic definitions and rules for calculating the derivatives of matrix-valued functions Y(X) of a matrix argument X, using vec, Kronecker products and vec permutation matrices; (2) Derivatives of patterned matrices, and (3) properties of \(G^+_ p(A\otimes B)G_ q\), where \(G^+_ p=(G'\!_ pG_ p)^{-1}G'\!_ p.\) The final section deals with applications in statistics. Jacobians of transformations are presented, for example of \(Y=AX'B'+BXA'\) and \(Y=AXB'+BXA'\), and the construction of ellipsoidal confidence regions for covariance matrices, on the basis of maximum likelihood or robust M- estimators, is given.