Independent Sums and Orlicz Norms Let $X_{i}$ be a collection of iid random variables of cardinality $n$, and let $S_{n}=\frac{1}{\sqrt{n}}\sum_{i=1}^{n}X_{i}.$
Let $|| X||:=\inf_{B}\{E[\exp(X/B)-1]\leq 1\}$. This is the so-called sub-exponential norm.
What can be said about $|| S_{n}||$ in terms of $||X_{i}||$?
In the $L^{2}$ norm, we have $||S_{n}||_{L^{2}}=||X_{i}||_{L^{2}}$ by choice of normalizing constant.
 A: *

*For the equality $\|S_n\|_{L^2}=\|X_i\|_{L^2}$ you need the zero-mean condition -- that $EX_i=0$.


*Let $X,X_1,\dots,X_n$ be any random variables with the same norm: $\|X\|=\|X_1\|=\cdots=\|X_n\|$.
Then, by the norm inequality for the sub-exponential norm $\|\cdot\|$,
\begin{equation*}
    \|S_n\|\le\frac1{\sqrt n}\,\sum_{i=1}^n \|X_i\|=\sqrt n\,\|X\|. \tag{1}\label{1}
\end{equation*}


*Let us show that the trivial upper bound $\sqrt n\,\|X\|$ on $\|S_n\|$ in \eqref{1} cannot be asymptotically improved (for $n\to\infty$), even if it is assumed that $X,X_1,\dots,X_n$ are iid zero-mean random variables. Toward this end, assume that
\begin{equation*}
    P(X=-a)=p=1-P(X=1),
\end{equation*}
where
\begin{equation*}
    a:=n^2,\quad p:=\frac1{a+1},\quad n\to\infty. 
\end{equation*}
Then $EX=0$ and, for each real $B>0$,
\begin{equation*}
    Ee^{X/B}=pe^{-a/B}+(1-p)e^{1/B}\to e^{1/B}.   
\end{equation*}
Equating now $e^{1/B}$ with $2$, we see that
\begin{equation*}
    \|X\|\to\frac1{\ln2}. \tag{2}\label{2}
\end{equation*}
Further, for each real $c>0$,
\begin{equation*}
    Ee^{S_n/(c\sqrt n)}=(pe^{-n^2/(cn)}+(1-p)e^{1/(cn)})^n
    =(1+(1+o(1))/(cn))^n\to e^{1/c}.   
\end{equation*}
Equating now $e^{1/c}$ with $2$, we see that
\begin{equation*}
    \|S_n\|\sim\frac{\sqrt n}{\ln2},  
\end{equation*}
so that, in view of \eqref{2},
\begin{equation*}
    \|S_n\|\sim\sqrt n\,\|X\|,  
\end{equation*}
as claimed.
