Eigendecomposition versus singular value decomposition in adaptive array signal processing (Q1180373)

The scope of this paper is to illustrate alternative algorithms for eigendecomposition (ED) of the autocorrelation matrix and for the singular value decomposition (SVD) of the data matrix. The paper has three major parts: 1) The data model, where the authors formulate the hypothesis underlying the theory. The concepts of array manifold and signal and noise subspaces are introduced. This gives a precise framework for the algorithms. A few considerations on the MUSIC and min-norm methods are presented. 2) The algorithms ED and SVD in order to perform signal or noise determination. The first algorithm (ED) is reduced to the computation of the spectral decomposition of the sum of a diagonal matrix with a Hermitian dyad. The eigenvectors of this matrix are easily computed with only one real and one complex operation by component and the eigenvalues are the zeros of a rational function. The second algorithm (SVD) assumes the decomposition of the sum of a diagonal matrix with a dyad and the singular values are founded by solving a bilinear equation. Once a singular value has been determined, the corresponding singular vector is easily computed, with three complex operations by component. 3) The simulation results where the min-norm method and its weighted version is selected to illustrate the performance of the algorithms. The ED/SVD trade-off is discussed and one concludes that the ED of the autocorrelation matrix is preferable.