# Intuition of the convergence of sample ACF

One of the problems in Brockwell and Davis book about time series is to show that

1) if $$\begin{equation} x_t = a + b t \end{equation}$$ then the sample autocorrelation ($$\hat{\rho}(h)$$) converges to 1 as the sample size tends to infinity for $$h \geq 1$$.

2) if $$\begin{equation} x_t = a \text{cos}(\omega t) \end{equation}$$ then the sample autocorrelation ($$\hat{\rho}(h)$$) converges to $$\text{cos}(\omega h)$$ as the sample size tends to infinity, where $$a \neq 0$$ and $$\omega \in [-\pi, \pi)$$.

I can prove these two results, but what is the message this question is trying to give?

Referring to the first case, since x is a simple trend it increases linearly as a function of b. That is: $$x_t=x_{t-1}+b$$. So it is a perfectly predictable process where the first difference is deterministic and equal to b. It is a unit root process and as such it has perfect memory about all its past values because each $$x_t$$ depends on $$x_{t-1}$$ which on its turn depends on $$x_{t-2}$$, etc.. so it is just a sum of deterministic terms over time. So it has a perfect memory of all its past values. So logically the autocorr must be one at any lag.
Hint: try to write the autocorr function for a AR process and set the AR coefficuent to 1 (if you drop the white noise to substitute it with an intercept b and set the AR coeff ti 1 then you get exactly this non-stationary deterministic trend, so you get the same process!). For an AR process the autocorr at k lags is $$\phi ^{k}$$ so if $$\phi =1$$ then the autocorr will be always 1 regardless k.
• How could you derive the theoretical autocorrelation function for $X_t = a + bt$, because $E(X_t) = a + bt, \forall t$, so $cov(X_{t+h}, X_t) = 0$? – shani Aug 19 '19 at 12:34
• How it is helpful to derive the autocorrelation function of $X_t$, I think the derivation of $\phi^k$ assumes $|\phi |< 1$ – shani Aug 19 '19 at 12:40