How to generalize Geometric ICA to higher dimensions.pdfVIP

How to generalize Geometric ICA to higher dimensions Fabian J. Theis, Elmar W. Lang Institute of Biophysics University of Regensburg, D-93040 Regensburg, Germany email: fabian.theis@mathematik.uni-regensburg.de Abstract. Geometric algorithms for linear independent component analysis (ICA) have recently received some attention due to their picto- rial description and their relative ease of implementation. The geometric approach to ICA has been proposed first by Puntonet and Prieto [6] in order to separate linear mixtures. One major drawback of geomet- ric algorithms is, however, an exponentially rising number of samples and convergence times with increasing dimensiononality thus basically restricting geometric ICA to low-dimensional cases. We propose to ap- ply overcomplete ICA to geometric ICA [7] to reduce high-dimensional problems to lower-dimensional ones, thus generalizing geometric ICA to higher dimensions. 1 Basics For m,n ∈ N let Mat(m×n) be the R?vectorspace of real m×n matrices, and Gl(n) := {W ∈ Mat(n× n) | det(W ) 6= 0} be the general linear group of Rn. In linear blind source separation (BSS), a random vector X : ? → Rm (mixed vector) originates from an independent random vector S : ? → Rn (source vector) by mixing with a mixing matrix A ∈ Mat(m × n), i.e. X = AS. Here ? denotes a fixed probability space. Only the mixed vector is known, and the task is to recover both the mixing matrix A and the source signals S. We will assume that the mixing matrix A has full rank. In the quadratic case (m = n) A then is invertible ie. A ∈ Gl(n) and S can be recovered from A by S = A?1X. This is not the true in the overcomplete case where less mixtures than sources are given (m n); then the BSS problem is ill-posed, hence further restrictions like source density assumptions have to be made. In the following we denote two matrices B, C ∈ Mat(m × n) to be equiv- alent, B ～ C, if C can be written as C = BPL with an invertible diagonal matrix (scaling matrix) L ∈ Gl(n) an