抄録
Burg's algorithm and forward-backward linear prediction algorithm (FBLP algorithm) are effective for short data records in autoregrresive spectrum analysis. But Burg's algorithm has several problems, including spectral line splitting and biases in frequency estimates. Burg's algorithm minimizes the sum of the forward and backward prediction powers subject to the constraint that the AR parameter estimates satisfy the Levinson recursion. The constraint causes the problems, but ensures the stable prediction filter. While FBLP algorithm minimizes it with respect to all AR paramters. This algorithm is the procedure in which the Levinson recursion constraint is removed, and includes less biases in frequency estimates and absence of spectral line splitting, however cannot guarantee the stable prediction filter.
In this paper, we propose a new algorithm possessing the advantage of Burg's algorithm and improving the estimates and show that spectral line splitting cannot be observed by using this algorithm.
