Convergence Theorems

Published in Fall, 2024

Definition. Let $\{ F_n \}$ be a sequence of distribution functions, if there exists a distribution function $F$ such that $$ \lim_{n \to +\infty} F_n(x) = F(x) $$ at every point $x$ which $F$ is continuous, we say that $F_n$ converges in law (or weakly) to $F$, and we write $F_n \overset{\omega}{\to} F$; Let $\{ X_n \}$ be a sequence of random variables and $\{ F_n \}$ be the corresponding distribution functions, we say that $X_n$ converges in distribution (or law) to $X$ if there exists an random variable $X$ with distribution function $F$ such that $F_n \overset{\omega}{\to} F$. We write $X_n \overset{L}{\to} X$.
Theorem. Let $\{ X_n \}, X$ be continuous random variables such that $\lim_{n \to +\infty} f_n(x) = f(x)$ for a.e $x \in \mathbb{R}$ where $f_n, f$ are the probability density functions of $X_n$ and $X$ respectively, then $X_n \overset{L}{\to} X$.
Definition. Let $\{ X_n \}$ be a sequence of random variables defined on some probability space $(\Omega,\mathcal{F},\mathbb{P})$, we say that the sequence $\{ X_n \}$ converges in probability to the random variable $X$ if $\forall \epsilon >0$, $$ \lim_{n \to +\infty} \mathbb{P} \{ \vert X_n - X \vert > \epsilon \} = 0 $$ And we denote by $X_n \overset{P}{\to} X$.
Example. Let $X$ be a random variable and $X_n = X + Y_n$, where $$ \mathbb{E}Y_n = \frac{1}{n} \hspace{0.4cm} \mathbf{Var}(Y_n) = \frac{\sigma^2}{n} $$ where $\sigma >0$ is a constant. Show that $X_n \overset{P}{\to} X$.
Click to show/hide solution
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.

Theorem. If $X_n \overset{d}{\to} c$ where $c$ is a constant, then $X_n \overset{P}{\to} c$.
Click to show/hide proof
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 $X_n \overset{P}{\to} c$.
Corollary. 1. $X_n \overset{P}{\to} X \Longleftrightarrow X_n - X \overset{P}{\to} 0$; 2. If $X_n \overset{P}{\to} X$ and $X_n \overset{P}{\to} Y$, then $\mathbb{P}\{ X = Y\} = 1$.
Corollary. If $X_n \overset{P}{\to} X$, then $X_n - X_m \overset{P}{\to} 0$. If $X_n \overset{P}{\to} X$ and $Y_n \overset{P}{\to} Y$, then $$ X_n \pm Y_n \overset{P}{\to} X \pm Y $$
Continuous Mapping Theorem. Let $X_n \overset{P}{\to} X$ and $g$ is a continuous function defined on $\mathbb{R}$, then $g(X_n) \overset{P}{\to} g(X)$.
Theorem. If $X_n \overset{P}{\to} X$, then $X_n \overset{L}{\to} X$.

In general, the reverse of this theorem is not true!

Slutsky's Theorem. If $\vert X_n - Y_n \vert \overset{P}{\to} 0$ and $Y_n \overset{P}{\to} Y$, then $X_n \overset{P}{\to} Y$.
Cramer's Theorem. If $X_n \overset{L}{\to} X, Y_n \overset{P}{\to} c$ where $c$ is a constant, then 1. $X_n \pm Y_n \overset{L}{\to} X \pm c$; 2. $X_n Y_n \overset{L}{\to} cX$ if $c\neq 0$ and $X_nY_n \overset{P}{\to} 0$ if $c=0$; 3. $\displaystyle{\frac{X_n}{Y_n} \overset{L}{\to} \frac{X}{c}}$ if $c \neq 0$.
Example. Let $X_1,X_2,\cdots$ be a sequence of $i.i.d$ random variables with common probability density function $f(x) = e^{-(x-\theta)}$ for $x \geq \theta$ and $=0$ otherwise. Let $\overline{X}_n$ denote the sample mean given by $\overline{X}_n = \frac{1}{n} \sum_{i=1}^n X_i$. (a) Show that $\overline{X}_n \overset{P}{\to} 1+\theta$; (b) Show that for any sequence $Y_1,Y_2,\cdots,Y_n$ of random variables and a constant $c$, if $ Y_n \overset{L}{\to} c$, then $Y_n \overset{P}{\to} c$; (c) Show that $\min(X_1,X_2,\cdots,X_n) \overset{P}{\to} \theta$.
Click to show/hide solution
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), $Y_n \overset{P}{\to} \theta$.
Example. Let $X_1,X_2,\cdots,X_n$ be $i.i.d$ random variables with common mean $0$ and variance $1$, suppose $\beta_4 = \mathbb{E} X^4 < +\infty$. Find the limiting distribution of the random variable $$ W_n = \sqrt{n} \frac{\sum_{i=1}^n X_i}{\sum_{i=1}^n X_i^2}. $$
Click to show/hide solution
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 $$ \frac{U_n}{V_n} \overset{L}{\to} N(0,1). $$
Exercise. Let $X_1,\cdots$ be a sequence of $i.i.d$, $U[0,1]$ random variables, and define $$ Y_n = \min\{ X_1,\cdots, X_n\} $$ Then prove the following results separately: (a) $Y_n \overset{d}{\to} 0$; (b) $Y_n \overset{P}{\to} 0$; (c) $Y_n \overset{a.s}{\to} 0$; (d) $Y_n \overset{L^r}{\to} 0$ for all $r \geq 1$.

Recommended citation: Jiajun Zhang, (2024) Convergence Theorems