1
$\begingroup$

In order to prove asymptotic normality of the time series OLS estimator (this context is not important), I have done the following:

Assumptions:

• $\mathrm{var}\left(y_{t}\right)=\gamma_{0}\quad\forall t$

• $ \mathrm{cov}\left(y_{t},y_{t-j}\right)=\gamma_{j}\in\mathbb{R}^{k}\quad\forall j$

• $\underset{j\rightarrow\infty}{\lim}\mathrm{cov}\left(y_{t},y_{t-j}\right)=\gamma_{j}=0$

• ${\sum}\left|\gamma_{j}\right|<\infty$

Then:

$$\begin{align} L&= \underset{{\scriptstyle T\rightarrow\infty}}{\lim}\mathrm{\mathrm{var}}\left(\sqrt{T}\left(T^{-1}\sum_{t}y_{t}-\mu\right)\right) \\ &= \underset{{\scriptstyle T\rightarrow\infty}}{\lim}\mathrm{\mathrm{var}}\left(T^{-1/2}\sum_{t}y_{t}\right) \\ &=\underset{{\scriptstyle T\rightarrow\infty}}{\lim}T^{-1}\mathrm{\mathrm{var}}\left(\sum_{t}y_{t}\right) \\ &=\underset{{\scriptstyle T\rightarrow\infty}}{\lim}T^{-1}\sum_{t}\sum_{s}\mathrm{cov}\left(y_{t},y_{s}\right) \\ &=\underset{{\scriptstyle T\rightarrow\infty}}{\lim}T^{-1}\left[T\cdot\gamma_{0}+2\cdot\left(T-1\right)\cdot\gamma_{1}+2\cdot\left(T-2\right)\cdot\gamma_{2}+...+2\cdot\gamma_{T-1}\right] \\ &=\underset{{\scriptstyle T\rightarrow\infty}}{\lim}T^{-1}\left[T\gamma_{0}+2\sum_{1\leq j\leq T-1}\left(T-j\right)\gamma_{j}\right] \\ &=\underset{{\scriptstyle T\rightarrow\infty}}{\lim}\gamma_{0}+2\sum_{1\leq j\leq T-1}\left(1-\dfrac{j}{T}\right)\gamma_{j} \end{align}$$ I'm assuming the last step equals: $\gamma_{0}+2\sum_{j=1}^{\infty}\gamma_{j}$

since concluding from here is rather easy. The problem being that I do not think it's the case, since j also converges to infinity.

Am I wrong? If so, why?

$\endgroup$
  • $\begingroup$ It's definitely true if $\sum j\gamma_j$ converges. $\endgroup$ – David Kleiman Feb 2 '16 at 15:35
  • $\begingroup$ Said sum converges (as $\sum{|\gamma_{j}|}$ does), as assumed. Could you please expose why? $\endgroup$ – rsm Feb 2 '16 at 15:37
  • $\begingroup$ The sum does not necessarily converge. Take $\gamma_j = 1/j^2$ . (In this case it your statement is true anyway). $\endgroup$ – David Kleiman Feb 2 '16 at 15:57
  • $\begingroup$ If you just multiply out the terms, you get $2(\sum\gamma_j \gamma_j - T^{-1}\sum j\gamma_j)$, so all you need is that $T^{-1}\sum j\gamma_j = 0$. $\endgroup$ – David Kleiman Feb 2 '16 at 15:59
  • $\begingroup$ So what you really need is that $\sum j\gamma_j$ is asymptotically less than $T$. $\endgroup$ – David Kleiman Feb 2 '16 at 16:01
1
$\begingroup$

Ok, so after a long battle with this one I think what you have written is true. Note that, since $\sum|\gamma_j|$ converges, we can find $N\in\mathbb{N}$ such that:

$$ \sum_{j=N}^T j| \gamma_j| \leq \sum_{j=N}^T j \frac cj = cT \;\;\;\; \forall c\in \mathbb{R} \setminus \{0\}. $$

This gives that

$$ \lim_{T\to \infty}T^{-1}\sum_{j=N}^T j |\gamma_j| \leq c\;\;\;\; \forall c\in \mathbb{R}\setminus \{0\}, $$

which implies that

$$ \lim_{T\to \infty}T^{-1}\sum_{j=1}^T j |\gamma_j| = 0. $$

Finally,

$$ \underset{{\scriptstyle T\rightarrow\infty}}{\lim}\gamma_{0}+2\sum_{1\leq j\leq T-1}\left(1-\dfrac{j}{T}\right)\gamma_{j} = {\lim_{T\to\infty}}\gamma_{0}+2\sum_{j=1}^\infty\gamma_j - 2T^{-1}\sum_{j=1}^\infty j\gamma_j \\ ={\lim_{T\to\infty}}\gamma_{0}+2\sum_{j=1}^\infty\gamma_j $$

$\endgroup$
  • $\begingroup$ Perfect! Thank you. Notice that you meant that c is non-zero, if you want to edit. $\endgroup$ – rsm Feb 2 '16 at 17:37
  • $\begingroup$ Yes thank you. I forgot to mention that. $\endgroup$ – David Kleiman Feb 2 '16 at 17:38
  • $\begingroup$ Can you explain how did you get your first inequality. Thanks $\endgroup$ – chandu1729 Feb 2 '16 at 18:21
  • $\begingroup$ @chandu1729 I edited my answer to fix up some details. Since the series $\sum |\gamma_j|$ converges, its tail is less than any constant $c$ times the tail of the harmonic series (which diverges). $\endgroup$ – David Kleiman Feb 2 '16 at 19:19
  • $\begingroup$ But how does that imply your inequality. Its not clear. I feel your arguments are wrong. Please be more rigorous $\endgroup$ – chandu1729 Feb 2 '16 at 19:30

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.