Subspace linear prediction approach to extracting poles

It is shown that the conventional linear prediction (LP) methods (including various versions of the Prony method) and the matrix pencil method for extracting poles of data sequences can be unified under a generalized approach called the subspace linear prediction (SLP) approach. The conventional LP methods can be considered as high-order SLP methods, while the matrix pencil method is a first-order SLP method. The authors also discuss a special form of the matrix pencil method for oversampled data. It is observed that, for oversampled data, the noise sensitivity of the least-square Prony method can be significantly improved without using singular value decomposition or other subspace decomposition algorithms.<>