Lecture 17 — Weak Law of Large Numbers & Bernstein Polynomials

This lecture covers:

  • Weak Law of Large Numbers (WLLN)
  • Khinchin’s theorem (variance-over-scaling implies convergence in probability)
  • Two classical examples:
    • The Coupon Collector Problem (Example 2.2.3 in Durrett)
    • Bernstein Polynomials (Example 2.2.1 in Durrett), which gives a probabilistic proof of the Weierstrass Approximation Theorem

(Chapter 2.2, Durrett)


1. Weak Law of Large Numbers (WLLN)

Let $X_n \to X$ “in probability” mean:

\[X_n \xrightarrow{p} X \quad \Longleftrightarrow \quad \forall \varepsilon>0,\ P(|X_n - X| > \varepsilon) \to 0.\]

Almost sure convergence implies convergence in probability:

\[X_n \xrightarrow{\text{a.s.}} X \quad \Rightarrow \quad X_n \xrightarrow{p} X.\]

Thus:

\[\text{SLLN} \Rightarrow \text{WLLN}.\]

2. Variance Condition for Convergence in Probability (Khinchin-type Lemma)

Let ${Y_n}$ be random variables and ${b_n}$ a nonzero scaling sequence.

Assume \(\frac{\operatorname{Var}(Y_n)}{b_n^2} \longrightarrow 0.\)

Then:

\[\frac{Y_n - \mathbb{E}[Y_n]}{b_n} \xrightarrow{L^2+p} 0.\]

Proof idea (as written in your notes)

Chebyshev:

\[P\left(\left|\frac{Y_n - \mathbb{E}(Y_n)}{b_n}\right| > \varepsilon\right) \le \frac{\operatorname{Var}(Y_n)}{\varepsilon^2 b_n^2} \longrightarrow 0.\]

Thus:

\[\frac{Y_n - \mathbb{E}[Y_n]}{b_n} \xrightarrow{p} 0.\]

This lemma is the core tool for both examples in the lecture.


3. Coupon Collector Problem (Example 2.2.3 Durrett)

Let $X_1, X_2, \ldots$ be i.i.d. uniform on ${1,\dots,n}$.
Let $\tau_k^{(n)}$ be the number of draws needed to obtain k distinct coupons.

Define:

\[X_{n,k} = \tau_k^{(n)} - \tau_{k-1}^{(n)} \quad (k=1,\dots,n),\]

the waiting time to obtain the $k$-th new coupon.

Given you have $k-1$ coupons, the chance the next draw is new is

\[p_k = \frac{n-k+1}{n}.\]

Thus:

\[X_{n,k} \sim \text{Geometric}(p_k),\qquad \mathbb{E}[X_{n,k}] = \frac{1}{p_k},\quad \operatorname{Var}(X_{n,k}) = \frac{1-p_k}{p_k^2}.\]

The total time to collect all $n$ coupons is:

\[T_n = \tau_n^{(n)} = \sum_{k=1}^n X_{n,k}.\]

Expectation

\(\mathbb{E}[T_n] = \sum_{k=1}^n \frac{1}{p_k} = n \sum_{k=1}^n \frac{1}{n-k+1} = n \sum_{j=1}^n \frac{1}{j} \sim n \log n.\)

Variance bound (from the handwritten notes)

\[\operatorname{Var}(T_n) = \sum_{k=1}^n \operatorname{Var}(X_{n,k}) \le \sum_{k=1}^n \frac{n}{(n-k+1)^2} = n \sum_{j=1}^n \frac{1}{j^2} \le n\cdot \frac{\pi^2}{6} \le C n.\]

Your notes simplify the bound further and simply record
$\operatorname{Var}(T_n) \le n^2$.
Either bound suffices for WLLN scaling.

Apply Khinchin’s lemma

Set $b_n = n \log n$. Then:

\[\frac{\operatorname{Var}(T_n)}{(n\log n)^2} \longrightarrow 0.\]

Thus:

\[\frac{T_n - \mathbb{E}[T_n]}{n\log n} \xrightarrow{p} 0.\]

Since $\mathbb{E}(T_n)/(n\log n)\to 1$,

\[\boxed{ \frac{T_n}{n\log n} \xrightarrow{p} 1. }\]

This is the WLLN for the coupon collector.


4. Bernstein Polynomials (Example 2.2.1 Durrett)

Goal: For any continuous $f:[0,1]\to\mathbb{R}$ and any $\varepsilon>0$,
find a polynomial $P_n$ such that

\[\max_{0\le x\le 1} |f(x) - P_n(x)| < \varepsilon.\]

This is the Weierstrass Approximation Theorem, proved probabilistically via Bernstein polynomials.


4.1 Definition

Fix $x\in[0,1]$.
Let $Y_{x,n} \sim \text{Binomial}(n,x)$.

Define the Bernstein polynomial:

\[P_n(x) = \mathbb{E}\!\left[f\!\left(\frac{Y_{x,n}}{n}\right)\right] = \sum_{k=0}^n f\left(\frac{k}{n}\right)\binom{n}{k} x^k (1-x)^{n-k}.\]

4.2 Convergence argument

Write:

\[|P_n(x) - f(x)| = \left|\mathbb{E}\big[f(Y_{x,n}/n) - f(x)\big]\right|.\]

Split expectation (this is the “split expectation trick” on page 2):

\[\mathbb{E}\big|f(Y_{x,n}/n) - f(x)\big| = \mathbb{E}\big[\ |f(Y_{x,n}/n) - f(x)|\ \mathbf{1}_{\{|Y_{x,n}/n - x|<\delta\}}\big] + \mathbb{E}\big[\ |f(Y_{x,n}/n) - f(x)|\ \mathbf{1}_{\{|Y_{x,n}/n - x|\ge\delta\}}\big].\]

Uniform continuity of $f$

Since $f$ is continuous on a compact interval, it is uniformly continuous:

\[|t-s| < \delta \ \Rightarrow\ |f(t)-f(s)| < \varepsilon.\]

Thus the first term is < $\varepsilon$.

Second term: use Chebyshev

\[|f(Y_{x,n}/n)-f(x)| \le 2\|f\|_\infty,\]

and

\[P(|Y_{x,n}/n - x| \ge \delta) \le \frac{\operatorname{Var}(Y_{x,n}/n)}{\delta^2} = \frac{x(1-x)}{n\,\delta^2} \longrightarrow 0.\]

Thus:

\[\mathbb{E}\big|f(Y_{x,n}/n) - f(x)\big| \le \varepsilon + 2\|f\|_\infty\cdot \frac{x(1-x)}{n\delta^2} \to \varepsilon.\]

Since $\varepsilon>0$ was arbitrary, this proves:

\[P_n(x)\to f(x)\quad\text{uniformly on }[0,1].\]

5. Summary

  • WLLN: Convergence in probability is implied by almost sure convergence; SLLN ⇒ WLLN.
  • Khinchin lemma: If $\operatorname{Var}(Y_n)/b_n^2\to0$, then $(Y_n - EY_n)/b_n \to 0$ in probability.
  • Coupon collector:
    \(T_n / (n\log n) \xrightarrow{p} 1.\)
  • Bernstein polynomials: Using binomial random variables and uniform continuity,
    \(P_n(x) \to f(x)\) uniformly on $[0,1]$, proving polynomial approximation.

Comments