In the case that the feature computes the autocorrelation function, there are many equivalent parametric representations of the AR parameters including reflection coefficients (RC), auto-regressive (AR) coefficients, and auto-correlation function (ACF). All of them are related by 1:1 transformations. We will discuss these in Chapter 9.

For now, we stay with the AR coefficients, which are well known [25]. These consist of the innovation variance $ \sigma^2$ and the AR coefficients $ {\bf a}=a_1,a_2 \ldots a_P$ . There are few non-linear problems with ML solutions available without iterative solutions. But, the AR coefficients, which are obtained by the Levinson algorithm, are surprisingly the same parameters that maximize the likelihood function. The ACF and AR coefficients are obtained from each other by 1:1 transformation, so the ACF also serves as the ML estimate.

Baggenstoss 2017-05-19