Rank-preserving geometric means of positive semi-definite matrices (Q1947084): Difference between revisions
From MaRDI portal
Latest revision as of 09:15, 6 July 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Rank-preserving geometric means of positive semi-definite matrices |
scientific article |
Statements
Rank-preserving geometric means of positive semi-definite matrices (English)
0 references
12 April 2013
0 references
The definition of a geometric mean of a pair of positive definite matrices has been explored by \textit{T. Ando, C.-K. Li} and \textit{R. Mathias} [Linear Algebra Appl. 385, 305-334 (2004; Zbl 1063.47013)]. Seven properties which such a mean should satisfy are listed and one particular mean defined by \(A_{1}\#A_{2} :=A_{1}^{1/2}(A_{1}^{-1/2}A_{2}A_{1}^{-1/2})^{1/2}A_{1}^{1/2}\) is well known. These definitions can be extended to semidefinite matrices by continuity, but there are some situations where we should like to define a geometric mean of two semidefinite matrices of rank \(p\) such that the mean also has rank \(p\), and this is not achieved by such a simpleminded extension. The authors propose the following definition based on the fact that the the geometric mean is a solution of a minimization problem. Let \(S^{+}(p,n)\) be the set of all semidefinite \(n\times n\) matrices of rank \(p.\) Given \(A_{1},A_{2}\in S^{+}(p,n)\) there exist \(n\times p\) matrices \(U_{i}\) with orthonormal columns and \(p\times p\) positive definite matrices \(R_{i}^{2}\) such that \(A_{i}=U_{i}R_{i}^{2}U_{i}^{T}\). These factorizations are not unique since for any \(Q_{i}\) in the orthogonal group \(O(p)\) we have alternative factorizations \(A_{i}=(U_{i}Q_{i})(Q_{i}^{T}R_{i}^{2}Q_{i} )(U_{i}Q_{i})^{T}\). However, if \(Q_{1}\) and \(Q_{2}\) are chosen as orthogonal matrices such that \(U_{1}^{T}U_{2}=Q_{1}(\cos\Sigma)Q_{2}^{T}\) where \(\Sigma\) is a diagonal matrix whose entries are the principal angles between the \(p\)-dimensional subspaces spanned by the columns of \(U_{1}\) and \(U_{2}\), then \(Y_{1}:=U_{1}Q_{1}\) and \(Y_{2}:=U_{2}Q_{2}\) are at minimum (Grassmann) distance. The proposed mean is now given by \(A_{1}\circ A_{2}:=W(R_{1} ^{2}\#R_{2}^{2})W^{T}\) where \(W\) is the Riemann (or Karcher) mean of \(Y_{1}\) and \(Y_{2}\) (\(W\) is the midpoint of a suitable geodesic from \(Y_{1}\) to \(Y_{2}\)). The authors give a geometric explanation of why the proposed mean should have desirable properties, they explain how it can be generalized to a mean of more than two elements of \(S^{+}(p,n)\), and they prove that it satisfies modified versions of Ando's seven properties. The paper concludes with a section on how the mean can be applied to filtering problems.
0 references
matrix means
0 references
geometric mean
0 references
positive semi-definite matrices
0 references
Riemannian geometry
0 references
principal angles
0 references
Karcher mean
0 references
Riemann mean
0 references
rank
0 references
factorization
0 references
orthogonal matrices
0 references
0 references
0 references
0 references