Isn't the usual zero-padding in the computation of the auto-correlation function with FFT just one of many possible extrapolations of the original signal?
If I have a measured signal which has good reason for having non-vanishing mean (e.g. the signal from an accelerometer which includes a gravity offset of 9.8 m/s^2 in static equilibrium...), wouldn't it be more appropriate (to 0'th order) to extend the signal outside its measured bounds by the respective last value inside the boundaries.
To be honest, I am not at all confident that I understand the role of an offset in the original signal in general. Because it can change the shape of the autocorrelation curve almost completely.
Side note: I am using the autocorrelation function in the LPC procedure/Yule-Walker equations. Maybe someone can explain to me, how an offset influences the result, or if the result is valid at all (should the signal have zero mean for LPC to work?).