Lecture 19 — Truncation Methods and Infinite-Mean Laws of Large Numbers

Truncation arguments

  • Weak Law of Large Numbers with infinite mean
  • The St. Petersburg paradox
  • Choosing normalizing sequences $b_n$ so that $\frac{S_b}{b_n}\to 1$ in probability for heavy-tailed distributions.

1. General Truncation Inequality

Let ${X_{n,k}}_{1 \le k \le n}$ be independent for each $n$, and define truncated variables

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

Let

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

A standard inequality (via Markov + variance control on truncated part) gives:

\[P\!\left( \frac{S_n - a_n}{b_n} > \varepsilon \right) \le \underbrace{\sum_{k=1}^n P(\vert X_{n,k}\vert > b_n)}_{(I)} + \underbrace{\frac{1}{\varepsilon^2 b_n^2}\sum_{k=1}^n \mathbb{E}(X_{n,k}^2 \mathbf{1}\{\vert X_{n,k}\vert \le b_n\})}_{(II)}.\]

Conclusion.
If

  • (I) → 0, and
  • (II) → 0 as $n\to\infty$,

then

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

2. Specialization to i.i.d. Variables

If ${X_k}$ are iid with partial sums $S_n = \sum_{k=1}^n X_k$ then:

\[(I)=n\,P(\vert X\vert > b_n), \qquad (II)=\frac{n\,\mathbb{E}(X^2 \mathbf{1}\{\vert X\vert \le b_n\})}{\varepsilon^2 b_n^2},\]

and

\[a_n = n \,\mathbb{E}(X \mathbf{1}\{\vert X\vert \le b_n\}).\]

If both vanish, then

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

The problem becomes choosing a normalizing sequence $b_n$ that makes both terms go to zero.


3. The St. Petersburg Paradox

Durrett Example 2.2.7.

\[P(X = 2^k) = 2^{-k},\quad k\ge 1.\]

Then

\[E[X] = \sum_{k=1}^\infty 2^k 2^{-k} = \infty.\]

Since the mean is infinite, the usual LLN does not apply.
In fact,

\[\frac{S_n}{n} \xrightarrow{P} \infty.\]

How to prove it? Truncation.

Fix $M > 0$, and define truncated variables:

\[X^{(M)} = X \wedge M.\]

Then:

  • $X^{(M)}$ are iid,
  • $\mathbb{E}[X^{(M)}] < M < \infty$,
  • by the standard WLLN:
\[\frac{1}{n} \sum_{k=1}^n X_k^{(M)} \xrightarrow{P} \mathbb{E}[X^{(M)}].\]

Since $S_n \ge \sum_{k=1}^n X_k^{(M)}$,

\[P\!\left( \frac{S_n}{n} > \mathbb{E}[X^{(M)}] - \varepsilon \right) \to 1.\]

Now let $M\to\infty$.
By MCT:

\[\mathbb{E}[X^{(M)}] \uparrow E[X] = \infty.\]

Thus for each fixed $M$, with probability → 1, the sample average exceeds a constant that can be made arbitrarily large. Hence,

\[\frac{S_n}{n} \xrightarrow{P} \infty.\]

4. A More Refined Normalization:

Claim

\(\frac{S_n}{n\log_2 n} \xrightarrow{P} 1.\)

Your notes derive this using the truncation conditions (I), (II).

Let

\[b_n = n\log_2 n.\]

Compute the truncated mean:

\[a_n = n\, \mathbb{E}(X\,;\,X \le n\log n) = n \sum_{2^k \le n\log n} 2^k 2^{-k} = n (\log n + \log\log n).\]

Then $a_n \sim n\log n = b_n$.
So the normalization is correct.

Check (I)

\[nP(X > b_n) = n \sum_{2^k > n\log n} 2^{-k} \le \frac{2n}{n\log n} = \frac{2}{\log n}\to 0.\]

Check (II)

\[\frac{n\, \mathbb{E}(X^2; X\le b_n)}{b_n^2} \approx \frac{n\cdot 2b_n}{b_n^2} = \frac{2}{\log n} \to 0.\]

Thus both truncation conditions vanish, implying:

\[\frac{S_n}{b_n} = \frac{S_n}{n \log_2 n} \xrightarrow{P} 1.\]

This is a Weak Law of Large Numbers with infinite mean, but with a different normalizing sequence.


5. General Criterion Using $\mu(s)$

Define the truncated first moment:

\[\mu(s)= E[X\,; X\le s].\]
  • $\mu(s)$ is increasing.
  • $\lim_{s\to\infty} \mu(s) = \infty$ iff $E[X]=\infty$.

A useful fact derived in the notes:

\[\frac{\mu(s)}{s} = \frac{E[X/s\,;\,X\le s]}{1} \longrightarrow 0.\]

This uses the DCT because $X/s \le 1$ on ${X\le s}$ and $X/s\to 0$.

The normalizing sequence $b_n$

Define $b_n$ by

\[\frac{\mu(b_n)}{b_n} = \frac{1}{n}.\]

This guarantees that the truncation method will succeed and often that:

\[\frac{S_n}{b_n} \xrightarrow{P} 1.\]

Jump points

If $P(X=S_0)>0$, there may be jumps in $\mu$. The notes mention this as a warning: the minimizing $b_n$ may occur at the point mass.


6. A Feller Example

The notes conclude with another heavy-tailed example (attributed to William Feller):

\(P(X = 2^k - 1) = \frac{1}{2^k k(k+1)},\qquad k\ge 1,\) \(P(X=-1) = 1 - P(X\ge 1).\)

Using the telescoping identity

\[\frac{1}{k(k+1)} = \frac{1}{k} - \frac{1}{k+1},\]

one can show $E[X]=0$ but the variance and higher moments diverge.
A similar truncation calculation yields:

\[\frac{S_n}{n/\log_2 n} \xrightarrow{P} 1, \qquad \left(\frac{S_n}{n} \xrightarrow{P} 0\right).\]

The message: “Truncation leads to the right normalization.”

Comments