Lecture 18 — Weak Law of Large Numbers & Triangular Arrays

Lecture 18 covers:

  • A general Weak Law of Large Numbers (WLLN) assumption:
    • $x P(\vert X\vert>x)\to 0$
  • Truncation method for proving WLLN.
  • Khinchin-type corollary.
  • Full WLLN proof for i.i.d. with finite mean.
  • Weak Law for Triangular Arrays, including the two required conditions.
  • St. Petersburg paradox as an example with infinite mean and unusual normalization.

1. A General Condition for WLLN

Let ${X_k}_{k\ge 1}$ be i.i.d.

Assume: \(x\, P(\vert X\vert > x) \xrightarrow[x\to\infty]{} 0. \tag{1}\)

Markov’s inequality gives only: \(x\, P(\vert X\vert > x) \le E\vert X\vert ,\) which is not useful by itself unless $E\vert X\vert <\infty$.
The assumption (1) is strictly stronger and is used to guarantee truncation works.


2. DCT Lemma Used for Tail Control

The notes (page 1) use:

\[\vert X\vert \mathbf{1}_{\{\vert X\vert \ge x\}} \le \vert X\vert , \quad \vert X\vert \mathbf{1}_{\{\vert X\vert \ge x\}} \xrightarrow[x\to\infty]{} 0 \ \text{a.s.}\]

By Dominated Convergence:

\[E\big[\vert X\vert \,\mathbf{1}_{\{\vert X\vert \ge x\}}\big] \xrightarrow[x\to\infty]{} 0. \tag{2}\]

Since \(E\big[\vert X\vert \,\mathbf{1}_{\{\vert X\vert \ge x\}}\big] = x\,P(\vert X\vert >x) + \int_x^\infty P(\vert X\vert >t)\,dt,\) the vanishing of (2) implies (1).

Thus the general WLLN condition (1) is equivalent to tail integrability.


3. Truncation: Constructing $X_{n,k}$

Define the truncated variables (page 1):

\[X_{n,k} = X_k\, \mathbf{1}_{\{\vert X_k\vert \le n\}}, \qquad k=1,\dots,n.\]

Let:

\[S_n = \sum_{k=1}^n X_k, \qquad S_n' = \sum_{k=1}^n X_{n,k}, \qquad m_n = E[X_{n,1}].\]

On page 1 (blue text):

“need to cut, can’t use Chebyshev directly”,
so we truncate to gain finite variance.


4. Step 1 — Show $S_n - S_n’ = o_p(n)$

Because truncation only removes the tail events:

\[P(S_n \neq S_n') \le \sum_{k=1}^n P(\vert X_k\vert > n) = n\, P(\vert X\vert >n).\]

Using assumption (1):

\[n\,P(\vert X\vert >n) = \frac{n}{n}\, nP(\vert X\vert >n) \to 0.\]

Thus:

\[\frac{S_n - S_n'}{n} \xrightarrow{p} 0. \tag{3}\]

5. Step 2 — Chebyshev Bound for Truncated Variables

\[P\left(\vert \frac{S_n'}{n} - m_n\vert > \varepsilon\right) \le \varepsilon^{-2}\operatorname{Var}\left(\frac{S_n'}{n}\right) = \frac{ \operatorname{Var}(X_{n,1}) }{n\varepsilon^2}.\]

Since $\vert X_{n,1}\vert \le n$, we compute:

\[E[X_{n,1}^2] = \int_0^\infty 2x\, P(\vert X_{n,1}\vert \ge x)\, dx. \tag{4}\]

Now:

\[P(\vert X_{n,1}\vert \ge x) = P(\vert X\vert \ge x)\mathbf{1}_{\{x\le n\}}.\]

Thus:

\[\frac{E[X_{n,1}^2]}{n} = \frac{1}{n} \int_0^n 2x\, P(\vert X\vert \ge x)\, dx. \tag{5}\]

Split the integral at a fixed $M$:

\[\frac{E[X_{n,1}^2]}{n} = \frac{1}{n} \int_0^M 2x P(\vert X\vert \ge x)\, dx + \frac{1}{n} \int_M^n 2x P(\vert X\vert \ge x)\, dx.\]
  • First term
    $\le 2M^2/n \to 0$.
  • Second term
    $\le \delta$ for large $M$ by condition (1).

Thus:

\[\frac{E[X_{n,1}^2]}{n} \longrightarrow 0.\]

Therefore:

\[P\left(\vert \frac{S_n'}{n} - m_n\vert > \varepsilon\right) \to 0. \tag{6}\]

6. Step 3 — Convergence of the Means

$m_n = E[X\,\mathbf{1}_{{\vert X\vert \le n}}]$.
By Dominated Convergence:

\[m_n \to E[X]. \tag{7}\]

7. Final WLLN Conclusion

Combine (3), (6), and (7):

\[\frac{S_n}{n} = \frac{S_n'}{n} + \frac{S_n - S_n'}{n} \xrightarrow{p} E[X].\]

Thus:

\[\boxed{ \frac{S_n}{n} \xrightarrow{p} E[X]. }\]

This is the Weak Law of Large Numbers proved via truncation.


8. Corollary — Khinchin’s Theorem

If $E\vert X\vert <\infty$ then the WLLN follows:

\[\frac{S_n}{n} \xrightarrow{p} E[X].\]

This is because (1) holds automatically for integrable random variables.


9. Weak Law for Triangular Arrays

We have a set of independent RVs in each row:

\[\{X_{n,k} : 1\le k\le n\},\]

not necessarily identically distributed.

Let $b_n \to \infty$.
Define truncated variables:

\[\bar X_{n,k} = X_{n,k}\mathbf{1}_{\{\vert X_{n,k}\vert \le b_n\}}.\]

Assumptions:

(i) Tail condition
\(\sum_{k=1}^n P(\vert X_{n,k}\vert > b_n) \to 0.\)

(ii) Variance condition
\(\frac{1}{b_n^2} \sum_{k=1}^n E[\bar X_{n,k}^2] \to 0.\)

Let:

\[S_n = \sum_{k=1}^n X_{n,k}, \qquad a_n = \sum_{k=1}^n E[\bar X_{n,k}].\]

Conclusion:

\[\boxed{ \frac{S_n - a_n}{b_n} \xrightarrow{p} 0. }\]

10. Example — St. Petersburg Paradox (page 2)

Durrett Example 2.2.7.

Let:

\[P(X = 2^k) = 2^{-k},\qquad k=1,2,\dots\]

Then:

  • $E[X] = \infty$,
  • The usual WLLN fails (normalization by $n$ breaks).

But your notes record:

\[\frac{S_n}{n\log_2 n} \xrightarrow{p} 1.\]

This shows a non-standard normalization is needed when the mean is infinite.


11. Summary

  • WLLN can be proved for i.i.d. using truncation if the tail condition
    $ x P(\vert X\vert >x)\to 0 $ holds.
  • This tail condition is equivalent to $E\vert X\vert <\infty$.
  • WLLN generalizes to triangular arrays with two conditions:
    • controlled tail probability,
    • controlled truncated second moments.
  • St. Petersburg illustrates WLLN may require nonstandard normalization.

Comments