Banded Householder representation of linear subspaces (Q417442)

The main result of the paper is as follows: If \(m\geq n\), any matrix \(A\in {\mathbb R}^{m\times n}\) can be factored as \(A=G\cdot (B\,|\,0)^{\text{T}}\), where \(B\in {\mathbb R}^{n\times n}\) is square and \(G\) is a product of \(n\) reflections in hyperplanes with a banded structure. The matrix \(G\) can be stored in \(n(m-n)\) floats by using Householder vectors. Consequently, one can represent any \(n\)-dimensional subspace of \({\mathbb R}^ m\) using \(n(m-n)\) floating point numbers. This is optimal since these subspaces form a Grassmannian of dimension \(n(m-n)\). The motivating application for this banded Householder decomposition is the compression of blend shape matrices for digital characters.