27 — Central Limit Theorem for Martingales

Central Limit Theorem (for Martingales)

Two proof strategies

  • Dvoretzky — uses a Taylor expansion.
  • Hille — uses characteristic functions.

This result relates closely to the autoregressive proof of CLT.


Lindeberg–Feller CLT for Martingale Differences

We have a martingale-difference triangular array:

\[\{D_{n,j},\,\mathcal{F}_{n,j}\}_{1 \le j \le n}, \qquad n=1,2,\ldots\]

with

\[E[D_{n,j}\mid \mathcal{F}_{n,j-1}]=0, \qquad E[D_{n,j}^2] < \infty .\]

Theorem (Martingale CLT)

If:

  1. Conditional variances converge to 1
    \(\sum_{j=1}^n E[D_{n,j}^2 \mid \mathcal{F}_{n,j-1}] \xrightarrow{P} 1 ,\)

  2. Lindeberg condition \(\sum_{j=1}^n E\!\left(D_{n,j}^2 \mathbf{1}_{\{|D_{n,j}|>\varepsilon\}} \mid \mathcal{F}_{n,j-1}\right) \xrightarrow{P} 0, \quad \forall \varepsilon>0,\)

then

\[S_n = \sum_{j=1}^n D_{n,j} \Rightarrow N(0,1).\]

Optional strengthening

We may add:

  1. Uniform bound (predictability condition)
    \(\sum_{j=1}^n E[D_{n,j}^2\mid\mathcal{F}_{n,j-1}] \le 2, \qquad\text{a.s. for all }n,\) which allows use of a stopping time.

Define the stopping index:

\[T_n = \max\left\{ j : \sum_{k=1}^j E(D_{n,k}^2\mid \mathcal{F}_{n,k-1}) \le 2 \right\}.\]

Then:

\[D_{n,j} \mapsto D_{n,j}\,\mathbf{1}_{\{j\le T_n\}}\]

and

\[P(T_n = n) \to 1.\]

Goal:

Show the characteristic functions converge:

\[E[e^{it S_n}] \longrightarrow e^{-t^2/2}, \qquad t\in\mathbb{R}.\]

Key decomposition (Page 1 figure)

\[E\!\left[e^{itS_n}\right] = E\!\left[ \prod_{k=1}^n \frac{e^{itD_{n,k}}}{E(e^{itD_{n,k}} \mid \mathcal{F}_{n,k-1})} \right].\]

Define

\[E(e^{itD_{n,k}}\mid\mathcal{F}_{n,k-1}) = 1 + r_{n,k},\]

where

\[r_{n,k} = E\!\left[e^{it D_{n,k}} - 1 - itD_{n,k} \mid\mathcal{F}_{n,k-1}\right].\]

Step 1: Show three key properties (Page 2)

  1. Sum of remainders approaches the Gaussian variance term \(\sum_{j=1}^n r_{n,j} \xrightarrow{P} -\frac{t^2}{2}.\)

  2. Uniform ℓ¹ bound \(\sum_{j=1}^n |r_{n,j}| \le 2t^2.\)

  3. Maximum term goes to zero \(\max_{1\le j\le n} |r_{n,j}| \xrightarrow{P} 0.\)


Step 2: Use product lemma (Lemma 8.81)

A general fact (page 2 bottom):

If a triangular array $a_{n,m}$ satisfies
1) $\sum_{m=1}^n a_{n,m}\to a$,
2) $\sup_n \sum_m |a_{n,m}| <\infty$,
3) $\max_m |a_{n,m}| \to 0$,
then
\(\prod_{m=1}^n (1 + a_{n,m}) \to e^a.\)

Applying this to $a_{n,m}=r_{n,m}$, we get

\[\prod_{k=1}^n (1 + r_{n,k}) \xrightarrow{P} e^{-t^2/2}.\]

Thus,

\[E[e^{itS_n}] \to e^{-t^2/2},\]

so $S_n \Rightarrow N(0,1)$.


Autoregressive Example

Consider the AR(1) model:

\[X_n = \theta X_{n-1} + U_n, \qquad n\ge 1, \ |\theta|<1,\]

with innovations $U_n$ i.i.d., mean $0$, variance $\sigma^2$, and $X_0$ independent of ${U_n}$.

Least-squares estimator

\[\hat{\theta}_n = \arg\min_{\theta} \sum_{k=1}^n (X_k - \theta X_{k-1})^2.\]

Solve normal equation:

\[(\alpha - \theta\beta,\beta)=0,\]

where
\(\alpha = (X_1,\ldots,X_n),\qquad \beta=(X_0,\ldots,X_{n-1}).\)

Thus

\[\hat{\theta}_n = \frac{(\alpha,\beta)}{(\beta,\beta)}.\]

Asymptotic normality

\[\sqrt n(\hat\theta_n - \theta) = \frac{\sqrt n \sum_{k=1}^n U_k X_{k-1}}{\sum_{k=1}^n X_{k-1}^2} \stackrel{d}{\longrightarrow} N\!\left(0,\ \frac{\sigma^2}{E[X_{0}^2]}\right).\]

(Derivation to be finished “on Wednesday”, per handwritten note.)


End of Lecture 28

Comments