I have $\gamma_l = Cov(r_t, r_{t-l})$ as a definition in my notes and now I need to find $gamma_l$ for a series $r_t- m = p(r_{t-1} - m) + a_t$ where $r_t$ is a linear time series with expected value $m$ and $a_t$ is a white noise.

Here is my attempt: $Cov(r_t - m, r_{t-l} - m) = E[(r_t - m)(r_{t-l} -m )] = E[p^2(r_{t-1} - m)(r_{t-l-1} -m ) + pa_t(r_{t-l-1} -m ) + pa_{t-l}(r_{t-1} - m) + a_ta_{t-l}] = p^2E[(r_{t-1} - m)(r_{t-l-1} -m )] + 0 + pE[a_{t-l}(r_{t-1} - m)] + E[a_t]E[a_{t-l}]$

Now if $l=0, = p^2E[(r_{t-1} - m)(r_{t-l-1} -m )] + pE[a_{t-l}(r_{t-1} - m)] + s_a^2 $

Otherwise, $= p^2E[(r_{t-1} - m)(r_{t-l-1} -m )] + pE[a_{t-l}(r_{t-1} - m)]$.

In either case, I'm stuck here. I am supposed to get everything in terms of $\gamma_{l-1}$ (or $\gamma_1$ in the case of 0, but I am not sure how to proceed.


Let $v_t=r_t-m$, then $v_t=pv_{t-1}+a_t$. Covariance is invariant under de-meaning: $\gamma_l=cov(v_t,v_{t-l})$. Let’s first solve $\gamma_0$:

$$\gamma_0=cov(pv_{t-1}+a_t, pv_{t-1}+a_t)=p^2\gamma_0+\sigma_a^2 \rightarrow \gamma_0=\frac{\sigma_a^2}{1-p^2}$$

Note that $cov(v_{t-1},a_t)=0$ because previous output is irrelevant of the current white noise.

For $l=1$: $$cov(v_t,v_{t-1})=cov(pv_{t-1}+a_t,v_{t-1})=pcov(v_{t-1},v_{t-1})=p\gamma_0$$

For $l=2$: $$cov(v_t,v_{t-2})=cov(pv_{t-1}+a_t,v_{t-2})=pcov(v_{t-1},v_{t-2})=p\gamma_1=p^2\gamma_0$$

If you go on like this, you’ll have $\gamma_l=p^l\gamma_0$. Since $\gamma_l=\gamma_{-l}$, by definiton, this formula generalizes into the following:



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.