boundary of limsup and liminf for sequence of averages with convergent subsequence

43 Views Asked by At

I'm having difficulty following the proof of Lemma III.1.7 in "Probability and Stochastics" (Cinlar). The lemma is:


Let $(x_n)$ be a sequence of positive numbers and denote $\overline x_n = \frac{1}{n} \sum_{i=1}^n x_i$

Let $N = (n_k)_{k \in \mathbb{N}}$ be a subsequence of $\mathbb{N}$ such that $\lim_{k \rightarrow \infty} \frac{n_{k+1}}{n_k} = r > 0$

If the sequence $(\overline x_n)$ converges along $N$ to $x$ (so the subsequence $(x_{n_k})$ converges to $x$)

Then $\frac{x}{r} \leq \lim \inf \overline x_n \leq \lim \sup \overline x_n \leq rx$


To begin with, this chain of inequalities implies that $\frac{x}{r} \leq rx$ which implies $r^2 \geq 1$. Allowing this is a small oversight with the intention that we should have $r \geq 1$, I'm having trouble further in the proof.

We begin by observing that for $n_k \leq n < n_{k+1}$, noting positivity of $x_n$ and that each summand (in the averages) occurs in the sum on the right we get

$\frac{n_k}{n_{k+1}} \overline x_{n_k} \leq \overline{x}_n \leq \frac{n_{k+1}}{n_k} \overline x_{n_{k+1}}$

Now the argument states simply that taking limits ($n \rightarrow \infty$, which implies also $k \rightarrow \infty$ to satisfy $n_k \leq n < n_{k+1}$) yields $LHS \rightarrow \frac{x}{r}$ and $RHS \rightarrow rx$; and true enough (with $r$ and $x_n$ positive). But where do the $\lim \inf$ and $\lim \sup$ fit into this?

1

There are 1 best solutions below

4
On BEST ANSWER

Basically, since $\frac{n_k}{n_{k+1}} \overline x_{n_k} \leq \overline{x}_n \leq \frac{n_{k+1}}{n_k} \overline x_{n_{k+1}}$ holds for all $n \in \mathbb{N}$ (in the sense that for a given $n$ there always exists a $k_n$ such that it holds), it'll still be true for $n$s belonging to a sequence $(m_l)_l$ such that the subsequence $(\overline{x}_{m_l})_l$ tends to $\liminf \overline{x}_n$, and same for $\limsup \overline{x}_n$ (I specify "tends" and not "converges" since at this step they could be equal to $+\infty$, at least for the latter, but that's just some detail). The existence of those subsequences is a good exercise if you haven't shown this before.

Taking the limit when $l \to \infty$ in the resulting inequalities (which I won't write because it's index-bonanza, what with all the $x$s, $k$s, $n$s, $m$s, $l$s, I prefer writing subsequences in the form $(u_{\varphi(n)})_n$ for that reason... but that's just a preference!) will get you $\frac{x}{r} \leq \lim \inf \overline x_n$ for $\liminf$ and $\lim \sup \overline x_n \leq rx$ for $\limsup$, and the middle inequality $\lim \inf \overline x_n \leq \lim \sup \overline x_n$ is true in general, thus we do have the desired result.

EDIT: To try and clarify what I said in view of your valid apprehension (I will admit what I said above is probably not clear at all), let me write down my anwser more thoroughly. I'll switch to subsequence maps for the notations, if I may:

Let $\varphi$ be the subsequence map corresponding to $N$, meaning: $\overline{x}_{\varphi(n)} \xrightarrow[n \to \infty]{} x$ and $u_n := \frac{\varphi(n+1)}{\varphi(n)} \xrightarrow[n \to \infty]{} r$.

As you said, for each $n \in \mathbb{N}$ there exists a unique $k(n) \in \mathbb{N}$ such that: $\varphi(k(n)) \leq n < \varphi(k(n)+1)$, and for that $k(n)$ we get: $$\frac{\varphi(k(n))}{\varphi(k(n)+1)} \overline x_{\varphi(k(n))} \leq \overline{x}_n \leq \frac{\varphi(k(n)+1)}{\varphi(k(n))} \overline x_{\varphi(k(n)+1)}$$ A bit more succintly, with the notation $u_n$ I introduced: $$\frac{\overline x_{\varphi(k(n))}}{u_{k(n)}} \leq \overline{x}_n \leq u_{k(n)}\, \overline x_{\varphi(k(n)+1)}$$

Now, define $\psi$ another subsequence map such that $\overline{x}_{\psi(n)} \xrightarrow[n \to \infty]{} \liminf \overline{x}$. Along that subsequence, we have: $$\frac{\overline x_{\varphi(k(\psi(n)))}}{u_{k(\psi(n))}} \leq \overline{x}_{\psi(n)} \leq u_{k(\psi(n))}\, \overline x_{\varphi(k(\psi(n))+1)}$$ Even if it means extracting a subsequence of $\psi(n)$ by taking out $n$s for which $k(\psi(n))$ are duplicates (except for one copy of course), we can assume WLOG that $k \circ \psi$ is also a subsequence map ($k$ is not a subsequence map though in general), this is possible due to the definition of the $k(n)$s.

We are on the final stretch: since $k \circ \psi$ is a subsequence map, we obtain that: $u_{k(\psi(n))} \xrightarrow[n \to \infty]{} r$, that: $\overline{x}_{\varphi(k(\psi(n)))} \xrightarrow[n \to \infty]{} x$ and that $\overline{x}_{\varphi(k(\psi(n))+1)} \xrightarrow[n \to \infty]{} x$, hence we can take the limit in the inequalities, and: $$\frac{x}{r} \leq \liminf \overline{x}_n \leq rx$$

Since the same can be done for $\limsup$, and we always have $\liminf \leq \limsup$, we get the desired result.