Finite-Sample Bias Propagation in Autoregressive Estimation With the Yule–Walker Method

The Yule-Walker (YW) method for autoregressive (AR) estimation uses lagged-product (LP) autocorrelation estimates to compute an AR parametric spectral model. The LP estimates only have a small triangular bias in the estimated autocorrelation function and are asymptotically unbiased. However, using them in finite samples with the YW method for AR estimation can give a strong distortion in the weak parts of the power spectral density. The distortion is shown to be influential in an example without strong spectral peaks. The true biased AR model, which is computed by applying the triangular bias to the true autocorrelation function, has an infinite order. A new objective measure is introduced to determine the smallest sample size for which the unbiased asymptotic theory can be considered as a fair approximation.Multi-Scale PhysicsApplied Science