Convergence Theorems
Published in Fall, 2024
Click to show/hide solution
By triangle inequality, we have $$ \vert Y_n \vert \leq \vert Y_n - \mathbb{E} Y_n \vert + \vert \mathbb{E} Y_n \vert = \vert Y_n - \mathbb{E} Y_n \vert + \frac{1}{n}. $$ Then, $\forall \epsilon >0$, we have $$ \begin{align*} \mathbb{P}\{ \vert X_n - X \vert \geq \epsilon \} & = \mathbb{P} \{ \vert Y_n \vert \geq \epsilon \}\\ & \leq \mathbb{P} \left\{ \vert Y_n -\mathbb{E} Y_n \vert + \frac{1}{n} \geq \epsilon \right\} \\ & = \mathbb{P} \left\{ \vert Y_n - \mathbb{E} Y_n \vert \geq \epsilon - \frac{1}{n} \right\} \\ & \leq \frac{\mathbf{Var}(Y_n)}{\left( \epsilon - \frac{1}{n} \right)^2} \hspace{2cm} \text{By Chebyshev's Inequality} \\ & = \frac{\sigma^2}{n \left( \epsilon - \frac{1}{n} \right)^2} \to 0, \hspace{1cm} \text{as $n \to +\infty$}. \end{align*} $$ Therefore we conclude that $X_n \overset{P}{\to} X$.
As we mentioned previously, convergence in probability is stronger than convergence in distribution. That is, if $X_n \overset{P}{\to} X$ then $X_n \overset{d}{\to} X$. However the converse is not necessarily true. To see this, let $X_1,X_2,\cdots,X_3$ be a sequence of i.i.d $Bernoulli(0.5)$ random variables, Let also $X \sim Bernoulli(0.5)$ be independent from the $X_i$’s, then $X_n \to X$, however $X_n$ does not converge in probability to $X$ since $\vert X_n - X \vert$ is in fact also a $Bernoulli(0.5)$ random variable and
\[\mathbb{P}\{ \vert X_n - X \vert \geq \epsilon \} = 0.5\]A special case in which the converse true is when $X_n \overset{d}{\to} c$ where $c$ is a constant. In this case convergence in distribution implies convergence in probability.
Click to show/hide proof
Suppose $X_n$ has distribution function $F_X$, then by definition we know that $\lim_{n \to +\infty} F_X = c$. Now for anny $k>0$ we have $$ \begin{align*} \mathbb{P}\{ \vert X_n - c \vert < k \} &= \mathbb{P} \{ -k < X_n - c < k\} \\ &= \mathbb{P}\{ -k+c < X_n < k+c \}\\ &= F_{X}(k+c) - F_X(-k+c) \\ &=1 \end{align*} $$ thus $X_n \overset{P}{\to} c$.
In general, the reverse of this theorem is not true!
Click to show/hide solution
(a) : First we find $\mathbb{E} X$, since those are $i.i.d$, they must have the same $\mathbb{E} X$ value: $$ \begin{align*} \mathbb{E} X &= \int_{-\infty}^{+\infty} x f_X(x) dx \\ & = \int_{\theta}^{+\infty} x e^{-(x - \theta)} dx\\ & = -xe^{-(x - \theta)} \Bigg\vert_{\theta}^{+\infty} - \int_{\theta}^{+\infty} -e^{-(x - \theta)} dx \\ & = \theta + 1. \end{align*} $$ Then, we find out that what we want to show is just $\overline{X}_n \overset{P}{\to} \mathbb{E} X$, to further prove this we consider for any $\epsilon>0$, the term $\mathbb{P} \{ \vert \overline{X}_n - \mathbb{E} X \vert >\epsilon \}$, and by Chevbeshev's inequality, $$ \begin{align*} \mathbb{P}\{ \vert \overline{X}_n - \mathbb{E} X \vert > \epsilon \} &\leq \frac{\mathbf{Var}(\overline{X}_n)}{\epsilon^2} \\ & = \frac{\frac{1}{n} \mathbf{Var}(X)}{\epsilon^2}\\ & = \frac{\mathbf{Var}(X)}{\epsilon^2} \cdot \frac{1}{n} \to 0, \text{as $n \to +\infty$}. \end{align*} $$ So we have proved that $\overline{X}_n \overset{P}{\to} 1+\theta$. (b) : By definition, $Y_n \overset{L}{\to} c$ implies $\lim_{n \to +\infty} F_{Y_n} = c$ where $c$ is a constant and $F_{Y_n}$ is the distribution function of the random variable $Y_n$, and we have $$ \begin{align*} \mathbb{P} \{ \vert Y_n - c \vert < \epsilon \} &= \mathbb{P}\{ -\epsilon < Y_n - c < \epsilon \} \\ &= \mathbb{P}\{ -\epsilon + c < Y_n < \epsilon + c \} \\ &= F_{Y_n}(\epsilon + c) - F_{Y_n}(-\epsilon + c) \\ & \overset{\text{as $n \to +\infty$}}{\to} 1 \end{align*} $$ Thus we have proved that $Y_n \overset{P}{\to} c$. (c) : Define $X_0 = \min(X_1,X_2,\cdots,X_n)$, then by definition, $$ \begin{align*} \mathbb{P}(X_0 \geq x) \implies \mathbb{P}(X_1 \geq x, X_2 \geq x, \cdots, X_n \geq x) &= \left( \int_{x}^{+\infty} e^{-(t - \theta)} dt \right)^n \\ &=e^{-n(x - \theta)}, \text{if $x > \theta$}. \end{align*} $$ Thus we have $$ f_{Y_0}(x) = \begin{cases} 1 : x \leq \theta \\ e^{-n(y - \theta)} : y > \theta \end{cases} $$ and the distribution function is $$ F_{Y_0}(x) = \begin{cases} 0 : x \leq \theta \\ e^{-n(x - \theta)} : x > \theta \end{cases} $$ as $n \to +\infty$, $F_{Y_0}(x) \to \begin{cases} 0 : x \leq \theta \\ 1: y > \theta \end{cases}$ and hence we have $Y_n \overset{L}{\to} \theta$ where $\theta$ is a constant, and thus by (b), $Y_n \overset{P}{\to} \theta$.
Click to show/hide solution
Note that $$ W_n = \sqrt{n} \frac{\sum_{i=1}^n X_i}{\sum_{i=1}^n X_i^2} = \frac{\frac{\sum X_i}{\sqrt{n}}}{\frac{\sum X_i^2}{n}} $$ We define $$ U_n = \frac{\sum_{i=1}^n X_i}{\sqrt{n}}, V_n = \frac{\sum_{i=1}^n X_i^2}{n} $$ Then, according to the properties of $X_i$, we know that $\mu = 0, \sigma = 1$, and the Central Limit Theorem states that $$ \frac{\sum_{i=1}^n X_n - \mu}{\sigma / \sqrt{n}} \to N(0,1) $$ If we define $\overline{X} = \frac{1}{n} \sum_{i=1}^n X_i$, then $$ U_n = \frac{\sum_{i=1}^n X_i}{\sqrt{n}} = \frac{n\overline{X}}{\sqrt{n}} = \frac{\overline{X}}{1/\sqrt{n}} = \frac{\overline{X} - n\mu}{\sigma \sqrt{n}} \to N(0,1) $$ So we have $U_n \overset{L}{\to} N(0,1)$ by Central Limit Theorem. Furthermore, if we define $Y_i = X_i^2$, then $V_n = \overline{Y}_n$, and we know that $$ \mathbb{E}(Y) = \mathbb{E}(X^2) = \mathbf{Var}(X) + (\mathbb{E} X)^2 = 1, \mathbf{Var}(Y) = \mathbb{E}(X^4) - (\mathbb{E}(X^2))^2 = \beta_4 - 1 < +\infty $$ And then we have $$ \mathbb{E}(\overline{Y}_n) = \mathbb{E}\left( \frac{Y_1+\cdots+Y_n}{n} \right) = \frac{1}{n} \sum_{i=1}^n \mathbb{E} Y_i = \mathbb{E} Y = 1 $$ and $$ \mathbf{Var}(\overline{Y}_n) = \mathbf{Var}\left( \frac{1}{n} \sum_{i=1}^n Y_i \right) =\frac{1}{n^2} \mathbf{Var}(Y_1 + \cdots+Y_n) = \frac{1}{n^2} n \mathbf{Var}(Y_i) = \frac{\mathbf{Var}(Y)}{n}. $$ So Chevbeshev's inequality states that for any $k >0$, $$ \begin{align*} \mathbb{P}\{ \vert \overline{Y}_n - \mathbb{E} \overline{Y}_n \vert \geq k \} &\leq \frac{\mathbf{Var}(\overline{Y}_n)}{k^2}\\ & = \frac{\mathbf{Var}(Y)}{n k^2} \leq \frac{\beta_4}{n k^2} \to 0, \text{as $n \to +\infty$}. \end{align*} $$ So we have that $\overline{Y}_n \overset{P}{\to} 1$, and by Cramer's theorem, we have $$ \frac{U_n}{V_n} \overset{L}{\to} N(0,1). $$
Recommended citation: Jiajun Zhang, (2024) Convergence Theorems
