Convergence, Part II
Last modified — 21 Jun 2026
By the end of this lecture, students are anticipated to be able to:
A sequence of random variables \(X_1,X_2,\ldots,X_n,\ldots\) with CDFs \(F_n\) converges in distribution to a random variable \(X\) with CDF \(F\) if, for all \(t\) at which \(F\) is continuous, \[\lim_{n \to \infty} F_n(t) = F(t).\]
If there exists \(s>0\) such that for all \(t \in (-s,s)\) \(m_{X_n}(t) \to m_X(t)\), then \(X_n \overset{d}{\to}X\).
Let \(U \sim {\mathrm{Unif}}(0,1)\), and let \(U_n\sim {\mathrm{Unif}}(0,1)\) all independent. Define
\[X_n = U_n + B_n\]
where \(B_n\sim {\mathrm{Bern}}(1/n)\) are independent Bernoullis, also independent of \(U, U_1, U_2, \ldots\).
Then \(X_n\overset{d}{\to}U\) but \(X_n\) does NOT converge in probability to \(U\).
We have that, for all \(t\), \[m_{X_n}(t) = m_{U_n}(t) m_{B_n}(t) = \frac{e^t - 1}{t} (1-\frac{1}{n} + \frac{1}{n}e^t) \to \frac{e^t - 1}{t} = m_U(t).\]
….continued….
However, \[\begin{aligned} \mathbb{P}(|X_n - U| > \epsilon) &= \mathbb{P}\left(|U_n + B_n - U| > \epsilon\right) \\ &= (1-1/n)\mathbb{P}\left(|U_n - U| > \epsilon\right) + (1/n)\mathbb{P}\left(|U_n +1- U| > \epsilon\right) \\ &= (1-1/n)(1-\epsilon)^2 + (1/n)a \quad\quad\text{for some $a\in[0,1]$}\\ &\rightarrow (1-\epsilon)^2 \neq 0. \end{aligned}\]
Let \(X_n\sim \mathcal{N}(0, 1 + 1/n)\) for all \(n\), mutually independent. Show that \(X_n \overset{d}{\to}Z\sim\mathcal{N}(0,1)\) by examining the moment generating functions.
Hint: Recall that the MGF of \(\mathcal{N}(\mu, \sigma^2)\) is \(m(t) = e^{\mu t + \sigma^2 t^2/2}\).
Please take 10 minutes to fill out the course evaluation. This will:
Help inform future course offerings (I’m teaching this again in Fall)
Provide feedback on my own teaching on where to improve
Help me stay employed in this economy :–)
I want you to fill this out regardless of how you feel about this course. Constructive feedback is welcome - rude comments about things I cannot change are not welcome. Keep it honest but professional - thank you!
The following implications hold for any sequence of random variables \(X_1, X_2, \ldots\) and any random variable \(X\): \[ X_n \overset{a.s.}{\to}X \Longrightarrow X_n \overset{p}{\to}X \Longrightarrow X_n \overset{d}{\to}X. \]
Interpreting convergence
Let \(U_1, U_2, \ldots\) be i.i.d. \({\mathrm{Unif}}(0,1)\) random variables. Define \(Y_n = \max\{U_1, \ldots, U_n\}\).
Show that \(n(1-Y_n) \overset{d}{\to}{\mathrm{Exp}}(1)\).
Hints:
Start by finding the CDF \(F_{Y_n}(t)\) of \(n(1-Y_n)\).
Recall that the CDF of \({\mathrm{Exp}}(1)\) is \(F(t) = 1 - e^{-t}\).
Let \(X_n\sim \mathcal{N}(0, 1 + 1/n)\), mutually independent, for all \(n\). Does \(X_n \overset{p}{\to}Z\sim\mathcal{N}(0,1)\)? Justify your answer.
Let \(X_{1},X_{2}, \dots, X_n,\ldots\) be i.i.d random variables with finite mean \(\mu\) and variance \(\sigma^2\).
Then, \[ \frac{\sqrt{n}(\overline{X}_n-\mu )}{\sigma } \overset{d}{\to}\mathcal{N}\left( 0,1\right). \]
Interpretation
Probability statements about \(\overline{X}_n\) can be approximated using a Normal distribution. It’s the probability statements that we are approximating, not the random variable itself.
Define \[Z_n = \frac{\sqrt{n}(\overline{X}_n-\mu )}{\sigma }.\]
You should not say things like:
\[\overline{X}_n \overset{d}{\to}\mathcal{N}(\mu, \sigma^2/n).\]
In many situations, the exact distribution of \(\overline{X}_n\), \(\mathbb{P}(\overline X_n \leq x)\), is hard to determine exactly.
The CLT allows us to approximate this value by
\[\mathbb{P}(\overline X_n \leq x) \approx \Phi\Big ( \frac{\sqrt{n}(x-\mu)}{\sigma}\Big )\] with a respectable precision when \(n\) is large.
Use the CLT to determine how many measurements the astronomer should make if they want the probability of a mismeasurement larger than 1 light year to be no more than 0.01?
Chebyshev’s inequality suggests \[n \ge 400\] independent observations.
CLT suggests \[n \ge 27\] independent observations
Both are correct, but the CLT is more precise.
To be fair, it used more information (the asymptotic distribution of the sample mean), which may or may not be accurate.
Chebyshev’s doesn’t use any approximation, it’s a guarantee.
Let \(Z_n = \frac{\sqrt{n}(\overline{X}_n-\mu )}{\sigma}\) as before.
Define \(Y_i = (X_i - \mu)/\sigma\) for all \(i\).
Suppose that \(Y_i\) has moment generating function \(m_Y(t)\).
Therefore, the moment generating function of \(Z_n\) is \(m_Y(t/\sqrt{n})^n\).
Now, \(m'_Y(0) = \mathbb{E}[Y_i] = 0\) and \(m''_Y(0) = \mathbb{E}[Y_i^2] = 1\).
By Taylor’s theorem, for all \(t\), \[\begin{aligned} m_Y(t) &= m_Y(0) + m'_Y(0)t + \frac{1}{2}m''_Y(0)t^2 + \cdots\\ &= 1 + 0 + \frac{t^2}{2} + \frac{t^3}{3!}m'''_Y(0) + \cdots \\ &= 1 + \frac{t^2}{2} + \frac{t^3}{3!}m'''_Y(0)+ \cdots.\\ \Longrightarrow \quad m_{Z_n}(t) &= m_Y(t/\sqrt{n})^n \ = \left(1 + \frac{\frac{t^2}{2} + \frac{t^3}{3!n^{1/2}}m'''_Y(0) + \cdots}{n}\right)^n \to e^{t^2/2} = m_Z(t). \end{aligned}\]
[We used the fact that \(\lim_{n \to \infty} (1 + a_n/n)^n = e^a\) when \(a_n \to a\).]
The CLT states that when \(n\) is large, the distribution of \[\frac{\overline{X}_n - \mu }{\sigma /\sqrt{n}} \text{ \ \ is \ approximately \ } \mathcal{N}\left( 0,1\right).\]
This implies that when \(n\) is large, we can also say something about the distribution of \(S_n = \sum_{i=1}^{n}X_{i}\).
\[\begin{aligned} 1-\Phi(z) &= \lim_{n\rightarrow\infty}\mathbb{P}\left( \frac{\overline X_n - \mu} {\sigma/\sqrt{n}} > z \right)\\ &=\lim_{n\rightarrow\infty}\mathbb{P}\left( \frac{(n\overline X_n - n\mu)} {n\sigma/\sqrt{n}} > z \right)\\ &=\lim_{n\rightarrow\infty}\mathbb{P}\left( \frac{S_n - n\mu} {\sqrt{n}\sigma} > z \right) \end{aligned}\]
The daily sales on any given day of a restaurant is a random variable with mean of $2500 and standard deviation of $500.
Assume that daily sales are independent random variables.
Give an approximate value of the probability that the total sale for the 30 days will be over $80,000.
Leave your answer in terms of \(\Phi\) (the CDF of a standard Gaussian).
The final page of your exam will also contain common distributions and general mean/variances (download the sheet here)
Stat 302 - Winter 2025/26