Lecture 31 — Convergence in Distribution

1. Definitions of Convergence in Distribution

Let $X_n$ and $X$ be real-valued random variables with CDFs $F_{X_n}$ and $F_X$.

Definition 1 (CDF definition).

We say \(X_n \Rightarrow X\) if \(F_{X_n}(x) \to F_X(x) \quad\text{for all }x\text{ such that }P(X=x)=0.\)

Equivalently, \(F_X(x-) = \lim_{t\uparrow x} F_X(t),\) and the convergence is required at points of continuity of $F_X$.

Definition 2 (Convergence of expectations).

We say \(X_n \Rightarrow X\) if \(E[g(X_n)] \to E[g(X)]\) for every bounded continuous $g:\mathbb R\to\mathbb R$.

This is the Portmanteau theorem characterization.
In your notes the condition is written $g\in C_B(\mathbb R)$.


2. The Skorohod Representation Theorem (Theorem A)

(Top half of page 1; a major theorem.)

Theorem A (Skorohod).

If \(X_n \Rightarrow X,\) then there exists another probability space $(\Omega,\mathcal F,P)$ and random variables
\(Y_n \stackrel{d}{=} X_n,\qquad Y \stackrel{d}{=} X,\) such that \(Y_n \xrightarrow{\text{a.s.}} Y.\)

Thus: distributional convergence can be represented as almost sure convergence on a new space.

Your notes reference this as the bridge from Definition 1 to Definition 2:

To go from Def (1) → Def (2), use Theorem A
(because a.s. convergence plus bounded continuous $g$ → DCT).

Indeed:

  • If $Y_n\to Y$ a.s.,
  • and $g$ is bounded continuous,
  • then $g(Y_n)\to g(Y)$ a.s.,
  • and DCT yields $E[g(Y_n)]\to E[g(Y)]$.

Since $Y_n\stackrel d= X_n$ and $Y\stackrel d=X$, this gives $E[g(X_n)]\to E[g(X)]$.


3. Quantile Transformation and Coupling

This is the heart of the constructive proof of Theorem A.

Let $Y$ be a random variable with CDF $F$.
Define the quantile function: \(F^{-1}(u) = \sup\{ y : F(y) < u \},\qquad 0<u<1.\)

Your notes (page 1) draw three diagrams:

  • “Nice case”: continuous, strictly increasing CDF.
  • “Jump case”: CDF has flat pieces and jumps.
  • “Flat-line case”: intervals where the CDF is constant.

The theory must work in all three cases.

Step 1: Let $U\sim\mathrm{Unif}(0,1)$.

Define \(Y = F^{-1}(U).\)

Then (page 1 diagram reasoning):

\[P(Y\le y) = P(F^{-1}(U)\le y) = P(U \le F(y)) = F(y).\]

So $F^{-1}(U) \stackrel d= Y$.
This works even when $F$ has jumps.

Step 2: If $F_n \to F$ pointwise at continuity points of $F$, then

\(F_n^{-1}(U) \to F^{-1}(U) \quad\text{a.s.}, \qquad 0<U<1.\)

Your notes (page 1–2) emphasize:
When $F(y)>F(y-)$, i.e. no jump at $y$, the quantiles converge.

Step 3: Define the Skorohod coupling

\(Y_n = F_n^{-1}(U),\qquad Y=F^{-1}(U).\)

Then $Y_n\stackrel d= X_n$, $Y\stackrel d=X$, and $Y_n\to Y$ a.s.
This proves Theorem A.


4. Proof of Def (2) ⇒ Def (1)

(Page 2 of notes.)

Given $X_n\Rightarrow X$ in the expectation sense, take any continuity point $x$ of $F_X$.

Use a family of “hat functions” $g_{x,\varepsilon}$ drawn in your notes:

  • The upper hat:
    $g_{x,\varepsilon}(y)=1$ for $y\le x$, then decreases linearly to 0 on $[x,x+\varepsilon]$.

  • The lower hat:
    $g_{x,-\varepsilon}(y)=1$ for $y\le x-\varepsilon$, linear to 0 on $[x-\varepsilon,x]$.

These functions are bounded and uniformly continuous.

Then:

\[P(X_n \le x) \le E[g_{x,\varepsilon}(X_n)] \to E[g_{x,\varepsilon}(X)] \le P(X\le x+\varepsilon).\]

Similarly, \(P(X\le x-\varepsilon) \le \liminf_n P(X_n\le x).\)

Since $F_X$ is continuous at $x$, \(P(X\le x-\varepsilon) \uparrow P(X\le x), \qquad P(X\le x+\varepsilon)\downarrow P(X\le x),\) letting $\varepsilon\to0$ gives \(P(X_n\le x)\to P(X\le x),\) which is Definition 1.


5. Examples of Convergence in Distribution

Example 1: constant sequences

(Page 2.)

Let $Z_n \equiv a_n$, $Z\equiv a$.
Then \(Z_n \Rightarrow Z \iff a_n\to a.\)

Example 2: a.s. convergence implies distributional convergence

If $X_n\to X$ almost surely, then $X_n\Rightarrow X$.

Example 3: Geometric$(\lambda/n)$ ⇒ Exponential$(\lambda)$

(Page 2–3.)

Let $Y_n \sim \mathrm{Geom}(\lambda/n)$, so \(P(Y_n > k) = (1-\lambda/n)^k.\) Then for fixed $t\ge 0$:

\(P\!\left(\frac{Y_n}{n} > t\right) = P(Y_n > nt) = (1-\lambda/n)^{nt} \to e^{-\lambda t} = P(Y > t)\) where $Y\sim\mathrm{Exp}(\lambda)$.

Therefore: \(\frac{Y_n}{n} \Rightarrow \mathrm{Exp}(\lambda).\)

Example 4: First repeated value among iid uniform ${1,\dots,N}$

(Page 3; the “first tie” example.)

Let $X_1,X_2,\dots,X_n$ be iid uniform on ${1,\dots,N}$.
Let \(T_N = \min\{ n\ge 2 : X_n = X_i \text{ for some }i<n\}.\)

Then \(P(T_N>n) = \frac{N (N-1)\cdots (N-n+1)}{N^n} = \prod_{m=2}^n \left(1-\frac{m-1}{N}\right).\)

Let $n=\lfloor y\sqrt{N}\rfloor$.
Then (page 3 derivation):

\[P\!\left(\frac{T_N}{\sqrt{N}} > y\right) = \prod_{m=2}^{y\sqrt N} \left(1 - \frac{m-1}{N}\right) \to \exp\!\left( -\frac{y^2}{2} \right) = P(Y>y),\qquad y\ge 0.\]

Therefore: \(\frac{T_N}{\sqrt N} \Rightarrow Y\) where $P(Y>y)=e^{-y^2/2}$.
Equivalently, $Y^2\sim \chi^2_1$.


Cheat-Sheet Summary — Lecture 31

  • Two equivalent definitions of convergence in distribution:
    1. CDF convergence at continuity points.
    2. Convergence of expectations of bounded continuous functions.
  • Skorohod Representation Theorem: $X_n\Rightarrow X$ ⇒ there exist $Y_n\stackrel d=X_n$ with $Y_n\to Y$ a.s.
  • Quantile coupling: $F^{-1}(U)$ generates a variable with CDF $F$.
    Used to build the Skorohod representation.
  • Hat functions $g_{x,\varepsilon}$ used to prove Definition 2 ⇒ Definition 1.
  • Examples:
    • Constants: $Z_n\Rightarrow a\iff a_n\to a$
    • a.s. convergence ⇒ distribution convergence
    • Geometric$(\lambda/n)$ scaled ⇒ Exponential($\lambda$)
    • First repeat among iid uniforms: $T_N/\sqrt N\Rightarrow Y$ with tail $e^{-y^2/2}$.

Comments