4
$\begingroup$

Let $\phi$ be an $N$-function, (i.e. $\phi : \mathbb{R}_{\geq 0} \to \mathbb{R}_{\geq 0}$ is convex and satisfies $\lim_{t \to 0} \frac{\phi(t)}{t} = 0, \lim_{t\to \infty} \frac{\phi(t)}{t} = \infty$).

We can define the associated Luxemburg norm on the appropriate subspace of $\mathbb{R}$-valued random variables (Orlicz space $L_\phi$), by $\|Z\|_{\phi} := \inf \{ \lambda : \mathbb{E} \phi(|Z|/\lambda) \leq 1\}$. The space $L_\phi$ is just a space of all random variables for which this infimum is finite.

For example, when $\phi(t) = t^p$ for $p > 1$, this is just the $L_p$ norm, and when $\phi(t) = e^{t^2} -1$ the corresponding Orlicz space is the space of sub-gaussian random variables.

Question

Is the following statement true: Consider a sequence $Z_1, Z_2, \ldots Z_n$ of independent random variables with $\mathbb{E} Z_i=0$ and $\|Z_i\|_\phi \leq 1$ for all $i$. Then $$ \|\frac{1}{n} \sum_{i \leq n} Z_i\|_\phi \leq \lambda_\phi(n)$$ for some $\lambda_\phi$ s.t. $\lambda_\phi(n) \to 0$ as $n\to \infty$.

A bit more context

I am mostly interested in the scenario in which $\phi$ grows only slightly faster than linear, say $\phi(t) = t \ln(1+t)$.

Together with Markov inequality, this would imply that for any $\varepsilon, \delta$, there is some $n_\phi(\varepsilon, \delta)$ , s.t. $\mathbf{Pr}(|\sum_{i\leq n} \frac{1}{n} X_i| > \varepsilon) \leq \delta$. Law of large numbers asserts that as long as the first moment is bounded, the sample average $\sum \frac{1}{s} X_i$ converges to $0$ - and I hope to be able to quantify the speed of convergence under just slightly stronger assumption.

Note that this is true in $L_p$ spaces: for any $1< p \leq 2$, we have $\|\frac{1}{n}\sum_i Z_i\|_p \lesssim n^{1/p - 1}$, so the interesting case is $\phi$ growing slower than $t^{1+\gamma}$ for any $\gamma$.

$\endgroup$

1 Answer 1

3
$\begingroup$

In general, the answer is no. Moreover, the answer is no even if \begin{equation} \phi(t)=t\ln(1+t). \tag{1} \end{equation}

Indeed, suppose that $P(Z_i=0)=1-2p$ and $P(Z_i=b)=p=P(Z_i=-b)$ for all $i$, where \begin{equation*} p:=\frac1{2\phi(b)}, \end{equation*} $\phi$ is as given by (1), and $b$ is a large enough positive real number so that $p\in(0,1/2)$.

Then for all $i$ we have $EZ_i=0$ and $E\phi(|Z_i|)=1$, so that $\|Z_i\|_\phi\le1$. On the other hand, for all real $c>0$ and all natural $n\ge2$ \begin{equation*} \begin{aligned} &E\phi\Big(\Big|\frac1n\sum_{i=1}^n Z_i\Big|/c\Big) \\ &\ge\sum_{i=1}^n \phi\Big(\frac b{cn}\Big)P(|Z_i|=b,\ Z_j=0\ \forall j\ne i) \\ &=n \phi\Big(\frac b{cn}\Big)2p(1-2p)^{n-1} \\ &=\frac{2pb}c\,\ln\Big(1+\frac b{cn}\Big)(1-2p)^{n-1}\to\frac1{2c}>1 \end{aligned} \end{equation*} as $n\to\infty$, if $b=n^2$ and $c\in(0,1/2)$. So, for all large enough $n$ we have $E\phi\big(\big|\frac1n\sum_{i=1}^n Z_i\big|/c\big)>1$ and hence $\|\frac1n\sum_{i=1}^n Z_i\|_\phi\ge c$ and hence \begin{equation*} \Big\|\frac1n\sum_{i=1}^n Z_i\Big\|_\phi\not\to0 \end{equation*} as $n\to\infty$.


More generally, the answer will remain no if $\phi(t)=t \ell(t)$, where $\ell$ is any function such that $\ell(t)$ is slowly varying as $t\to\infty$. Yet more generally, the answer will remain no if $\phi(t)=t L(t)$, where $L$ is any function such that $\sup\limits_{K\in(0,\infty)}\limsup\limits_{t\to\infty}\dfrac{L(Kt)}{L(t)}<\infty$.

$\endgroup$
1
  • $\begingroup$ Thanks, that's a really clean example, $\endgroup$ Nov 22, 2021 at 0:52

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.