A modification of the Kumaresan-Tufts methods for estimating rational impulse responses

This correspondence presents a modification of the Kumaresan-Tufts algorithm for estimating the parameters of rational impulse responses in additive white noise. The modified algorithm uses the forward and backward predictors to select the signal poles, and then combines the two sets of poles to form the final estimates. The modified algorithm exhibits an improved robustness and a smaller bias in some test cases.