Lecture 37 — Continuity Theorem

This lecture has two major parts:

  1. Completing the Inversion Formula for characteristic functions (CFs), including boundary cases.
  2. Proving the Continuity Theorem:
    convergence of characteristic functions implies convergence in distribution.

1. Inversion Formula Revisited

From Lecture 36, let $X$ be a real-valued r.v. with characteristic function $\varphi_X(t)$, and let
$U\sim \mathrm{Unif}(a,b)$, independent of $X$.
Set $Y = X - U$.

Then:

\[\varphi_Y(t) = \varphi_X(t)\,\varphi_{-U}(t), \qquad \varphi_{-U}(t) = \frac{e^{-ita}-e^{-itb}}{it(b-a)}.\]

Main inversion identity (page 1):

\(\frac{b-a}{2\pi} \int_{-T}^{T} \varphi_X(t)\,\varphi_{-U}(t)\,dt \;\xrightarrow[T\to\infty]{ } P(a<X<b) +\frac{P(X=a)+P(X=b)}{2}. \tag{1}\)

This recovers the mass in $(a,b)$ plus half–mass at the endpoints.


1.1 Boundary case: recovering the mass at a point

Page 1 shows the special case $a=b=x$.
Using the formula with $a=b=x$ (interpreting via limiting uniform variables):

\[\lim_{T\to\infty} \frac{1}{2\pi} \int_{-T}^{T} e^{-itx}\varphi_X(t)\,dt = P(X=x). \tag{2}\]

This is the Dirichlet kernel limit: \(\frac{\sin(Ty)}{\pi y} \to \delta_0(y),\) as indicated by the sketch on page 1 (“hope $x-a=0$”).

Thus (2) recovers point mass at any $x\in\mathbb{R}$.


1.2 When $X$ has a density

Page 1–2:
If $\int_{-\infty}^\infty \vert \varphi_X(t)\vert \,dt <\infty$, then $X$ has a PDF $f_X$, with:

\[f_X(x) = \frac{1}{2\pi} \int_{-\infty}^\infty e^{-itx}\varphi_X(t)\,dt. \tag{3}\]

Moreover, the truncated integral approximates the density:

\[\lim_{T\to\infty} \frac{1}{2\pi} \vert \int_{-T}^{T} e^{-itx}\varphi_X(t)\,dt\vert = f_X(x).\]

(Your notes show the inequality bounding the tail by $(1/2T)\int \vert \varphi_X\vert \to0$.)


2. Continuity Theorem (Durrett 3.3.6)

This is the heart of Lecture 37, written across both pages.

Continuity Theorem: Part (i)

If $X_n\Rightarrow X$, then for every $t\in\mathbb{R}$:

\[\varphi_{X_n}(t) \to \varphi_X(t). \tag{4}\]

This follows from boundedness of $e^{itX_n}$ and dominated convergence.


Continuity Theorem: Part (ii)

Let $\varphi_n(t)$ be the CFs of $X_n$.
Assume:

  1. $\varphi_n(t)\to g(t)$ for every $t\in\mathbb R$,
  2. $g$ is continuous at 0.

Then:

  • ${X_n}$ is tight,
  • $g(t)$ is the CF of some $X$,
  • and
    \(X_n \Rightarrow X. \tag{5}\)

This is the most powerful implication:
Pointwise convergence of CFs + continuity at 0 ⇒ weak convergence.

Durrett Probability 4.1e - Theorem 3.3.6 Continuity Theorem

Let $\mu_n,1\ge n\ge\infty$ be probability measures with characteristic functions $\varphi_n$. (i) If $\mu_n\Rightarrow \mu_\infty$ then $\varphi_n(t)\to \varphi_\infty(t)$ for all $t$. (ii) If $\varphi_n(t)$ converges pointwise to a limit $\varphi(t)$ that is continuous at $0$, then the associated sequence of distributions $\mu_n$ is tight and converges weakly to the measure $\mu$ with characteristic function $\varphi$.


3. Proof Sketch of the Continuity Theorem

The handwritten proof on pages 1–2 gives the core ideas:
(1) show tightness; (2) identify the limit law using inversion;
(3) show convergence on a subsequence gives same CF, therefore whole sequence converges.


3.1 Key Lemma on Page 2

For any $u>0$,

\[\frac{1}{u}\int_{-u}^{u} \big(1-\varphi(t)\big)\,dt = 2 - \frac{1}{u}\int_{-u}^{u}\varphi(t)\,dt. \tag{6}\]

Since $\vert \varphi(t)\vert \le 1$, the right-hand side is bounded by 2.

Further in the notes (page 2):

\[\vert X\vert \ge a \ \Rightarrow\ 1 - \varphi(t) \ge c\,P(\vert X\vert \ge a) \quad\text{for suitable }c>0. \tag{7}\]

The intuition is that if $\vert X\vert $ is large, the oscillation in $e^{itX}$ forces the average in (6) up.

Combining (6)–(7), for each $\varepsilon>0$ one finds an $M$ such that:

\[\frac{1}{u}\int_{-u}^u \big(1-g(t)\big)\,dt <\varepsilon \quad\Rightarrow\quad P(\vert X\vert >M) < \varepsilon. \tag{8}\]

This step (page 2 boxed conclusion):

\[P(\vert X_n\vert \ge M) < \varepsilon \quad\text{uniformly in }n,\]

is exactly tightness.

Thus the sequence of distributions is tight.


3.2 Extracting a subsequence

By tightness + Helly–Prokhorov (Lecture 33),
we may extract $X_{n_k}\Rightarrow X$ for some subsequence.

Then by part (i):

\[\varphi_{n_k}(t) \to \varphi_X(t). \tag{9}\]

But the hypothesis says $\varphi_{n_k}(t)\to g(t)$.
Thus:

\[g(t)=\varphi_X(t).\]

So $g$ is a characteristic function.

Since every convergent subsequence has the same limit law, the whole sequence converges:

\[X_n \Rightarrow X.\]

4. Summary of the Logic

The Continuity Theorem ties everything together:

\[X_n\Rightarrow X \quad\Longleftrightarrow\quad \varphi_{X_n}(t)\to\varphi_X(t)\;\forall t.\]

More generally:

  • Given just CF convergence,
    if the limit $g(t)$ is continuous at 0, then it is a valid CF of some $X$, and

    \[X_n \Rightarrow X.\]

This is the foundation of all the CF-based limit theorems, especially the CLT, stable laws, and the Lévy–Khintchine formula.


Cheat-Sheet Summary — Lecture 37

  • Inversion formula recovers probabilities: \(\frac{b-a}{2\pi}\int_{-T}^{T}\varphi_X(t)\varphi_{-U}(t)\,dt \to P(a<X<b)+\tfrac12(P(X=a)+P(X=b)).\)
  • If $\int \vert \varphi_X\vert !<\infty$, then $X$ has a density and
    \(f_X(x)=\frac{1}{2\pi}\int e^{-itx}\varphi_X(t)\,dt.\)

  • Continuity Theorem:
    • $X_n\Rightarrow X \implies \varphi_{X_n}\to\varphi_X$.
    • If $\varphi_n\to g$ pointwise and $g$ is continuous at $0$,
      then ${X_n}$ is tight, $g$ is a CF, and $X_n\Rightarrow X$.
  • The key lemma:
    \(\frac{1}{u}\!\int_{-u}^{u}(1-\varphi(t))dt = 2 - \frac{1}{u}\!\int_{-u}^{u}\varphi(t)dt.\)

    Used to force tightness.

Comments