\documentclass[11pt]{article}
\begin{document}

    \renewcommand{\baselinestretch}{2}
    \normalsize
    \renewcommand{\baselinestretch}{2}


\section {Approximating lower bounds on $S_n$.}

(This first paragraph below is to help keep notation straight, but probably
should be
removed ultimately.)

We suppose as before that
$X_i$ are iid Pareto random variables with distribution function $F(x) = 1
- x^{-\alpha}$ for
$x \geq 1$,
that $S_n = \sum \limits _{i=1} ^n X_i$, and that
$M_n = \max\{X_1, X_2, ..., X_n\}$.

An approximation for the $q$th quantile of $S_n$, for small $q$
(e.g.\ $q = 2.5\%$), may be
constructed based on the following idea.
One seeks $c_n$, where
\begin{eqnarray}
P(S_n < c_n) \approx q. \label{cn}
\end{eqnarray}
For any value $y_n$,
\begin{eqnarray}
P(S_n < c_n) &=& P(S_n < c_n | M_n \leq y_n) P(M_n \leq y_n) \nonumber\\
& & \hspace{.01in} + \hspace{.05in}
P(S_n < c_n | M_n > y_n) P(M_n > y_n). \label{mn1}
\end{eqnarray}
The term $P(S_n < c_n | M_n \leq
y_n)$ in (\ref{mn1}) can be approximated using the central
limit theorem, since the variables being summed are now truncated
and hence have finite moments.
The resulting approximation will be close for sufficiently small values
of $y_n$ and sufficiently large $n$. At the same time,
for any reasonably large value of $y_n$, and for $c_n$ satisfying
(\ref{cn}) for small $q$, the quantity
$P(S_n < c_n | M_n > y_n)$ in (\ref{mn1})
will be infinitessimal;
hence the entire final term in (\ref{mn1}) may be considered negligeable.
Thus, in approximating $c_n$, we suggest choosing an appropriate value of
$y_n$,
and considering the approximation
\begin{eqnarray}
P(S_n < c_n) & \approx & P(S_n < c_n | M_n \leq y_n) P(M_n \leq y_n)
\label{rough}\\
& \approx &  \Phi\{(c_n - n \mu_{y_n}) / (\sigma_{y_n} \sqrt{n})\} P(M_n \leq
y_n), \label{clt1}
\end{eqnarray}
where $\Phi$ is the standard normal distribution function, and
%\begin{eqnarray}
%\mu_y &:=& E[X_1 | X_1 \leq y] =
%[\alpha \beta^\alpha y^{1-\alpha} - \alpha \beta] / [1 - \alpha +
%\alpha \beta^\alpha y^{-\alpha}],\label{muy}\\
%\sigma_y^2 &:=& V[X_1 | X_1 \leq
%y] =  \frac{\alpha \beta ^ \alpha y ^{2-\alpha} -
% \alpha \beta ^2}{(2 - \alpha)(1- \beta ^\alpha y ^{-\alpha})} -
% \mu_y^2,\label{sy}
% \end{eqnarray}

\begin{eqnarray}
\mu_y &:=& E[X_1 | X_1 \leq y] = \frac{
\alpha y^{1-\alpha} - \alpha } {(1 - \alpha)(1-y^{-\alpha})},
\hspace{.5in} \alpha \neq 1, \label{muy}\\
& &  \hspace{.96in} = \frac{ln(y)}{1-y^{-1}}, \hspace{1.11in} \alpha = 1,
\nonumber\\
\sigma_y^2 &:=& V[X_1 | X_1 \leq
y] =  \frac{\alpha  y ^{2-\alpha} -
 \alpha}{(2 - \alpha)(1-  y ^{-\alpha})} -
 \mu_y^2, \hspace{.16in} \alpha \neq 2, \label{sy}\\
 & & \hspace{.96in} = \frac{2 ln(y)}{1-y^{-2}} - \mu_y^2,
 \hspace{.77in} \alpha = 2.\nonumber
 \end{eqnarray}

Note that there is a tradeoff in choosing $y_n$ in (\ref{clt1}):
if one selects too small a value of $y_n$, then the term $P(S_n < c_n |
M_n > y_n)$ in (\ref{mn1}) is not
negligeable, so the resulting approximation may not be satisfactory. On the
other hand,
if $y_n$
is too large, then the approximation of $P(S_n < c_n | M_n < y_n)$
using the central limit theorem may be unsatisfactory; this is
particularly true for small $n$.

One option is to choose some value
$p^*$ to represent the
probability $P(S_n < c_n | M_n \leq y_n)$ in (\ref{rough}).
%For instance, one might take $p^* = \sqrt{q}$.
From (\ref{cn}), one then has $q/p^* = P(M_n \leq y_n) = \{1 -
(\frac{1}{y_n})^\alpha\}^n$, and solving
this for $y_n$ one obtains
\begin{eqnarray}
y_n = \{1 - (q/p^*)^{1/n}\}^{-1/ \alpha}.
    \label{yn}
\end{eqnarray}

One may then obtain an approximation of $c_n$ by plugging this value of $y_n$
from (\ref{yn})
into the first term in (\ref{clt1}),
yielding
\begin{eqnarray}
\hat c_n &=&  \sigma_{y_n} \sqrt{n} \Phi^{-1}(p^*) + n \mu_{y_n},
\label{final}
\end{eqnarray}
where $\mu_{y_n}$ and $\sigma_{y_n}$ are given by equations
(\ref{muy}-\ref{yn}).

A naive choice for $p^*$ is $\sqrt{q}$; this seems to balance the
aforementioned tradeoff, since then $P(S_n < c_n | M_n \leq y_n)
= P(M_n \leq y_n) = \sqrt{q}$.

The values reported in column 3 of Table 1 reflect the approximation $\hat
c_n$ in (\ref{final}),
with $p^* = \sqrt{q}$ and $y_n$, $\mu_{y_n}$, and $\sigma_{y_n}$ given by
(\ref{muy}-\ref{yn}). However, plots of ideal choices of $p^*$
versus $q$, $n$, and $\alpha$ suggest that instead, the choice
\begin{eqnarray}
  p^* = 0.136 + 0.235 q + q^2 + 0.0066 \mathop{min}(n,10) - 0.05
  \mathop{max}(a,1) \label{final2}
\end{eqnarray}
provides a better approximation. Performance of the approximation $\tilde
c_n$ resulting
from the use of (\ref{final}) with (\ref{final2}) for $p^*$ and
$y_n$, $\mu_{y_n}$, and $\sigma_{y_n}$ given by
(\ref{muy}-\ref{yn}) is shown in column 4 of Table 1. 

Although the
slightly simpler $\hat c_n$ approximates the lower quantile
quite well, in most cases there is substantial improvement from the use of the
approximation $\tilde c_n$ which employs (\ref{final2}). This is 
also displayed graphically in Figures 1 and 2 below. Figure 1 shows 
how the $0.02$ quantile $c_n$ and the two approximations 
$\hat c_n$ and $\tilde c_n$ vary with $n$, for $\alpha = 2/3$. One sees 
that the two approximations match the true quantile quite closely. 
Figure 2 highlights the  
error rates 
for the two approximations $\hat c_n$ and $\tilde c_n$ as a function 
of $n$, again 
for $\alpha = 2/3$.  
Table 1 and Figures 1 and 2 
show the error in the approximations for $q = 0.02$, but results for
other small values of $q$ are rather similar. \\[.2in]


{\bf Figure Captions:}\\

Figure 1: Quantile $c_n$ (solid curve), along with approximations
$\hat c_n$ (dotted) and $\tilde c_n$ (dashed) as functions of $n$, for
$\alpha = 2/3$ and $q = 0.02$. For each $n$, the values of $c_n$ shown 
is the empirical $0.02$ quantile from $10$ million simulations of 
$S_n$.\\

Figure 2: Error rates for the two approximations $\hat c_n$ 
(dotted) and $\tilde c_n$ (dashed)
as functions of $n$, for 
$\alpha = 2/3$ and $q = 0.02$. 
Percentage error rates are calculated as $100 (\hat c_n - c_n)/c_n$ and 
$100 (\tilde c_n - c_n)/c_n$, where each
value of $c_n$ is the empirical $0.02$ quantile taken from $10$ million simulations of 
$S_n$.\\

\pagebreak

The table below is for $q = 0.02$. Error rates are rounded to 4 decimal places.\\

\renewcommand{\baselinestretch}{1}
    \normalsize
    \renewcommand{\baselinestretch}{1}

\begin{tabular}{|c|c|c|c|}
\hline
\Large $\alpha$ & \Large $n$ & \Large $(\hat{c}_n-c_n)/c_n$
& \Large $(\tilde{c}_n-c_n)/c_n$\\
\hline
\normalsize
0.5 & 2 & 0.0157 & -0.0023\\
0.5 & 5 & 0.0120 & 0.0029\\
0.5 & 10 & 0.0079 & 0.0041\\
0.5 & 20 & 0.0052 & -0.0018\\
0.5 & 50 & 0.0040 & -0.0080\\
0.5 & 100 & 0.0035 & -0.0021\\
\hline
0.67 & 2 & 0.0019 & -0.0005\\
0.67 & 5 & 0.0012 & -0.0030\\
0.67 & 10 & -0.0050 & 0.0045\\
0.67 & 20 & -0.0019 & 0.0027\\
0.67 & 50 & -0.0010 & 0.0013\\
0.67 & 100 & 0.0005 & -0.0004\\
\hline
1 & 2 & -0.0110 & 0.0007\\
1 & 5 & -0.0056 & 0.0000\\
1 & 10 & -0.0006 & 0.0023\\
1 & 20 & 0.0019 & 0.0007\\
1 & 50 & -0.0149 & -0.0015\\
1 & 100 & -0.0076 & -0.0014\\
\hline
1.5 & 2 & 0.0003 & 0.0035\\
1.5 & 5 & 0.0045 & 0.0032\\
1.5 & 10 & -0.0182 & -0.0043\\
1.5 & 20 & -0.0085 & -0.0022\\
1.5 & 50 & 0.0017 & 0.0049\\
1.5 & 100 & 0.0059 & 0.0046\\
\hline
    \end{tabular}




\end{document}