Next: Corollary 5 Up: The Asymptotic Frequency Distribution Previous: Corollary 4

Theorem 5 (modified from Gnedenko (1943), pp. 438-439)

Let R and S be real numbers such that

R = e-S,

and let $n_1 < n_2 < n_3 < \ldots$ be any subsequence of the positive integers. For x = x0 and for any finite real constants anj > 0 and bnj, $j = 1,2,3,\ldots$,

 
 \begin{displaymath}
\lim_{j \to \infty} F^{n_j} (a_{n_j} x_0 + b_{n_j} ) = R\end{displaymath} (29)

if and only if

 
 \begin{displaymath}
\lim_{j\to\infty} n_j \left[ 1 - F (a_{n_j} x_0 + b_n) \right] = S.\end{displaymath} (30)

Proof: By Taylor's Theorem [Landau (1951), p. 120, theorem 178],

\begin{displaymath}
\log y = - (1-y) \left\{ 1 - \frac{1-y}{2q^2} \right\}, \qquad y < q <
1.\end{displaymath}

Then,

 
 \begin{displaymath}
\begin{array}
{r@{}l}
\log F^{n_j} & (a_{n_j} x_0 + b_{n_j})...
 ...frac{1 - F (a_{n_j} x_0 + b_{n_j} )}{2q^2} \right\},\end{array}\end{displaymath} (31)

where F(anj x0 + bnj ) < q < 1.


(a) Suppose

\begin{displaymath}
\lim_{j \to \infty} F^{n_j} (a_{n_j} x_0 + b_{n_j}) = R.\end{displaymath}

If $R \neq 0$, then by Eq. (29),

\begin{displaymath}
\lim_{j\to\infty} F(a_{n_j} x_0 + b_{n_j} ) = 1.\end{displaymath}

Hence, taking the limit in Eq. (31) gives

 
 \begin{displaymath}
\lim_{j\to\infty} \log F^{n_j} ( a_{n_j} x_0 + b_{n_j}) = -
\lim_{j\to\infty} n_j \left[ 1-F (a_{n_j} x_0 + b_{n_j} ) \right]\end{displaymath} (32)

or $- \log R = S$. If R = 0 and $\lim_{j \to \infty} F(a_{n_j} x_0 + b_{n_j} ) = 1$,the above argument still yields Eq. (32).

If R = 0 and $\overline{\lim} _{j\to\infty} F (a_{n_j} x_0 +
b_{n_j}) < 1$, then immediately

\begin{displaymath}
\begin{array}[t]
{c} \underline{\lim} \\  \scriptstyle j \to...
 ...1 - F (a_{n_j} x_0 + b_{n_j} ) \right] = -\infty = -\log R = S.\end{displaymath}

Hence, in all cases,

 
 \begin{displaymath}
\lim_{j \to \infty}
 n_j \left[ 1 - F (a_{n_j} x_0 + b_{n_j} ) \right]
= -\log R = S.\end{displaymath} (33)


(b) Now, suppose that

\begin{displaymath}
\lim_{j\to\infty} n_j \left[ 1-F (a_{n_j}
x_0 + b_{n_j}) \right] = S.\end{displaymath}

If $S < \infty$, then $\lim_{j \to \infty} F(a_{n_j} x_0 + b_{n_j} ) = 1$, and by Eq. (31) as $n \to
\infty$,

\begin{displaymath}
\lim_{j\to\infty} \log F^{n_j} (a_{n_j} x_0 + b_{n_j}) = -
\...
 ...to\infty} n_j \left[ 1 - F (a_{n_j} x_0 + b_{n_j}) \right] =
-S\end{displaymath}

or

 
 \begin{displaymath}
\lim_{j\to\infty} F^{n_j} (a_{n_j} x_0 + b_{n_j}) = e^{-S}.\end{displaymath} (34)

If $S = \infty$ and $\lim_{j \to \infty} F(a_{n_j} x_0 + b_{n_j} ) = 1$, the reasoning leading to Eq. (34) still holds and Eq. (34) is true. If $S = \infty$ and $\lim_{j\to\infty} F
(a_{n_j} x_0 + b_{n_j} ) < 1$, then

\begin{displaymath}
\begin{array}[t]
{c} \overline{\lim} \\  \scriptstyle j\to\infty\end{array} F^{n_j} (a_{n_j} x_0 + b_{n_j} ) = 0 = e^{-S}.\end{displaymath}

Hence, in every case,

\begin{displaymath}
\lim_{j\to\infty} F^{n_j} (a_{n_j} x_0 + b_{n_j} ) = e^{-S} = R.\end{displaymath}



 
Next: Corollary 5 Up: The Asymptotic Frequency Distribution Previous: Corollary 4

Leon Borgman
3/10/1998