2
$\begingroup$

The following fraction shows up when trying to show consistency of the OLS slope estimator in a simple linear regression on a log-log scale where the window of observation changes as the sample size $T$ increases and I want to find a constant $c > 0$ such that $$ \frac{\sum_{k = 1}^{N(T)} \vert x_{k, T} - \bar{x}_{N(T)} \vert}{\sum_{k = 1}^{N(T)} (x_{k, T} - \bar{x}_{N(T)})^2} \leq c $$ if $T$ is large enough, where $$ \begin{align*} x_{k, T} &= \log(T^{\varepsilon} + k - 1), \ \varepsilon \in (0, 1), \\ \bar{x}_{N(T)} &= \frac{1}{N(T)} \sum_{k = 1}^{N(T)} x_{k, T}, \\ N(T) &= (m - 1) T^{\varepsilon},\ m > 1 \end{align*} $$

So far, I have tried some form of figuring out for which $k$ it holds that $\vert x_{k, T} - \bar{x}_{N(T)} \vert \leq 1$ and then using that the absolute value is less or equal to its square which helps to bound some part of the fraction but I do not now how to deal with the remaining part.

Another idea of mine is to rewrite the fraction by multiplying both the numerator and denominator by $1 / N(T)$ such that the numerator is the mean absolute deviation of a (non-random) sample $(x_{1, T}, \ldots, x_{N(T), T})$ from its mean whereas the denominator is the mean squared deviation. I was hoping that there maybe is some knowledge about the asymptotic behavior of both numerator and denominator but so far I did not find anything. Also, things are a bit tricky because as $T$ increases the sample not only becomes larger but changes altogether.

Finally, another idea is to use the rewritten fraction from the previous approach and try to show that both numerator and denominator are asymptotically equivalent to integrals of the form $\int_{T^{\varepsilon}}^{mT^{\varepsilon}}...dx$ which are hopefully somewhat easy to evaluate. However, I am not quite sure if such an asymptotic equivalence is feasible.

$\endgroup$
2
  • $\begingroup$ Is $m$ an integer? Is $c$ allowed to depend on $m$? on $\varepsilon$? $\endgroup$ Jun 30, 2021 at 12:34
  • $\begingroup$ $c$ may depend on $\varepsilon$ and $m$. Also, $m$ does not need to be an integer. $\endgroup$
    – AlbertRapp
    Jun 30, 2021 at 13:00

1 Answer 1

1
$\begingroup$

$\newcommand{\ep}{\varepsilon}\newcommand{\num}{\operatorname{num}}\newcommand{\den}{\operatorname{den}}\newcommand{\R}{\mathbb R}$Let $$x_k:=x_{k,T}=\ln(S+k-1),\quad S:=T^\ep,\quad N:=N(T).$$ Let us ensure that $N$ is an integer, assuming, more generally, just that \begin{equation} N\sim(m-1)S. \end{equation} We have to show that eventually (that is, for all large enough $S>0$) \begin{equation} \frac{\num}{\den}\le c<\infty, \end{equation} where \begin{equation} \num:=\sum_1^N|x_k-\bar x|,\quad \den:=\sum_1^N(x_k-\bar x)^2,\quad \bar x:=\frac1N\,\sum_1^Nx_k, \end{equation} and $c$ may depend only on $m\in(1,\infty)$.

Note that $\sum_1^N|x_k-a|$ is convex in $a\in\R$. Hence, \begin{align*} \num&\le\max\Big(\sum_1^N(x_k-x_1),\sum_1^N(x_N-x_k)\Big) \\ &\le\max\Big(\sum_1^N\ln\Big(1+\frac{k-1}S\Big),\sum_1^N\ln\Big(1+\frac{N-K}S\Big)\Big) \\ &\le\max\Big(\sum_1^N\frac{k-1}S,\sum_1^N \frac{N-K}S\Big) \\ &=\frac{N(N-1)}{2S}\le\frac{(m-1)^2}{2+o(1)}\,S. \end{align*}

On the other hand, by Jensen's inequality for the concave function $\ln$, \begin{equation} \bar x=\frac1N\,\sum_1^N\ln(S+k-1)\le\ln\Big(\frac1N\,\sum_1^N(S+k-1)\Big) =\ln(S+(N-1)/2). \end{equation} So, \begin{align*} \den&\ge\sum_{3N/4\le k\le N}(x_k-\bar x)^2 \\ &\ge\sum_{3N/4\le k\le N}\ln^2\Big(1+\frac{k-1-(N-1)/2}{S+(N-1)/2}\Big) \\ &\ge\frac N{4+o(1)}\,\ln^2\Big(1+\frac{N/4}{S+(m-1)S/2}\Big) \\ \ &\sim\frac{(m-1)S}4\,\ln^2\Big(1+\frac{m-1}{2m+2}\Big). \end{align*} Thus, \begin{equation} \frac{\num}{\den}\le(2+o(1))(m-1)\Big/\ln^2\Big(1+\frac{m-1}{2m+2}\Big), \end{equation} as desired.

$\endgroup$
5
  • $\begingroup$ Great solution. Thanks. However, a few comments: How did you arrive at $N - k$ in the maximum? I could only bound $x_N - x_k$ by $\log(1 + (N - 1)/S$. In this case, it does not matter though since this bounds $\log(1 + (k - 1)/S$ from above, so we can bound the maximum accordingly. I lose a factor 1/2 though because then I only bound the sum by $N(N-1)$ instead of calculating the sum explicitly. $\endgroup$
    – AlbertRapp
    Jun 30, 2021 at 15:39
  • $\begingroup$ Are the $o(1)$ necessary? I do not understand the purpose. $\endgroup$
    – AlbertRapp
    Jun 30, 2021 at 15:40
  • $\begingroup$ When bounding the denominator, you started $k$ at $3N/4$. Heuristically, I understand that then $x_k - \bar{x} > 0$ which we need to use the inequality we found due to Jensen. However, is that clear formally? $\endgroup$
    – AlbertRapp
    Jun 30, 2021 at 15:42
  • $\begingroup$ @AlbertRapp : (i) "How did you arrive at $N-k$ in the maximum?" -- We have $x_N-x_k=\ln(1+\frac{N-K}{S+k-1})\le\ln(1+\frac{N-K}S)$. (ii) "Are the $o(1)$ necessary?" -- They are needed already because I only assume that $N\sim(m-1)S$. Also, if $3N/4\le k\le N$, then $k-1-(N-1)/2\ge N/4-1/2=N/(4+o(1))$. $\endgroup$ Jun 30, 2021 at 16:05
  • $\begingroup$ @AlbertRapp : (iii) "When bounding the denominator, you started $k$ at $3N/4$. Heuristically, I understand that then $x_k-\bar x>0$ which we need to use the inequality we found due to Jensen. However, is that clear formally?" -- Here we are using the following: if $3N/4\le k\le N$, then $x_k-\bar x\ge\ln(S+k-1)-\ln(S+(N-1)/2)=\ln(1+\frac{k-1-(N-1)/2}{S+(N-1)/2})\ge0$. $\endgroup$ Jun 30, 2021 at 16:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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