We introduce a new approach for sparse decomposition, based on a geometrical interpretation of sparsity. By sparse decomposition we mean finding sufficiently sparse solutions of underdetermined linear systems of equations. This will be discussed in the context of Blind Source Separation (BSS). Our problem is then underdetermined BSS where there are fewer mixtures than sources. The proposed algorithm is based on minimizing a family of quadratic forms, each measuring the distance of the solution set of the system to one of the coordinate subspaces (i.e. coordinate axes, planes, etc.). The performance of the method is then compared to the minimal 1-norm solution, obtained using the linear programming (LP). It is observed that the proposed algorithm, in its simplest form, performs nearly as well as LP, provided that the average number of active sources at each time instant is less than unity. The computational efficiency of this simple form is much higher than LP. For less sparse sources, performance gains over LP may be obtained at the cost of increased complexity which will slow the algorithm at higher dimensions. This suggests that LP is still the algorithm of choice for high-dimensional moderately-sparse problems. The advantage of our algorithm is to provide a trade-of between complexity and performance.
[1]
Michael A. Saunders,et al.
Atomic Decomposition by Basis Pursuit
,
1998,
SIAM J. Sci. Comput..
[2]
Yuanqing Li,et al.
Analysis of Sparse Representation and Blind Source Separation
,
2004,
Neural Computation.
[3]
Christian Jutten,et al.
Sparse ICA via cluster-wise PCA
,
2006,
Neurocomputing.
[4]
Barak A. Pearlmutter,et al.
Blind source separation by sparse decomposition
,
2000,
SPIE Defense + Commercial Sensing.
[5]
Michael Elad,et al.
A generalized uncertainty principle and sparse representation in pairs of bases
,
2002,
IEEE Trans. Inf. Theory.
[6]
Xiaoming Huo,et al.
Uncertainty principles and ideal atomic decomposition
,
2001,
IEEE Trans. Inf. Theory.
[7]
D. Donoho.
For most large underdetermined systems of linear equations the minimal 𝓁1‐norm solution is also the sparsest solution
,
2006
.
[8]
Mineichi Kudo,et al.
Performance analysis of minimum /spl lscr//sub 1/-norm solutions for underdetermined source separation
,
2004,
IEEE Transactions on Signal Processing.
[9]
Michael Zibulevsky,et al.
Underdetermined blind source separation using sparse representations
,
2001,
Signal Process..