Lecture 24 — Formal SLLN Proof via Truncation and Block Arguments

This lecture continues the formal proof of the Strong Law of Large Numbers using truncation, “good” sequences, and block decomposition (the Chung–Erdős / Chandrasekharan–style argument your professor sketches).

We want to prove the SLLN in the general case:

Theorem (SLLN).
If ${X_k}_{k\ge 1}$ are iid with
\(E\vert X\vert < \infty, \qquad E[X]=\mu,\)
then
\(\frac{S_n}{n} \xrightarrow{a.s.} \mu, \qquad S_n=\sum_{k=1}^n X_k.\)

This lecture gives a classical, more delicate proof (following Durrett), using truncation and carefully selected subsequences $T_{2^n}$.


1. Reduction and Notation

WLOG, assume $X\ge 0$ a.s.
(General case: apply to $X^+$ and $X^-$ separately.)

Define:

\[X_k = X_k^+ - X_k^-, \qquad X_k^+ = \max(X_k,0), \qquad E[X^+] < \infty.\]

Let

\[\delta_n^+ = \sum_{k=1}^n X_k^+, \qquad \delta_n^- = \sum_{k=1}^n X_k^-, \qquad \delta_n = \delta_n^+ - \delta_n^-.\]

Then,

\[\frac{\delta_n}{n} = \frac{\delta_n^+}{n} - \frac{\delta_n^-}{n}.\]

If we can show both terms converge almost surely, we are done.


2. Step 1 — Truncation Removes Only Finitely Many Terms

Define truncated variables:

\[Y_k = X_k\,\mathbf{1}_{\{\vert X_k\vert \le k\}}.\]

Then:

\[P(X_k \ne Y_k) = P(\vert X_k\vert > k) = P(\vert X\vert > k).\]

Since $E\vert X\vert = \int_0^\infty P(\vert X\vert >t)\,dt < \infty$,

\[\sum_{k=1}^\infty P(\vert X\vert >k) < \infty.\]

Thus by Borel–Cantelli I,

\[P(X_k \ne Y_k \text{ i.o.}) = 0.\]

Hence

\[X_k = Y_k \quad\text{for all large } k,\ a.s.\]

So it suffices to prove the SLLN for the truncated variables $Y_k$.


3. Interlude — “Good” Sequences ${a_n}$

Your notes (bottom of page 1) define:

A sequence $a_n \uparrow \infty$ is good if
\(\sum_{n=m}^\infty a_n^{-2} \le C\, a_m^{-2},\qquad m\ge1.\)

Examples:

  • $a_n = n^p$ for any $0<p<2$,
  • more generally: if $a_n / n^{1/p}$ is non-decreasing, then $a_n$ is good.

For this proof, we will use the good sequence:

\[a_n = n.\]

4. Step 2 — Second-Moment Control

Let

\[Y_n = X\,\mathbf{1}_{\{\vert X\vert \le a_n\}},\qquad a_n = n.\]

Assume as above:

\[\sum_{n=1}^\infty P(\vert X\vert >a_n) < \infty.\]

Then we want to show:

\[\sum_{n=1}^\infty \frac{E[Y_n^2]}{a_n^2} < \infty.\]

Proof sketch (page 2):

Decompose according to the annuli $a_{m-1} < \vert X\vert \le a_m$:

\[\sum_{n=1}^\infty \frac{E[Y_n^2]}{a_n^2} = \sum_{n=1}^\infty a_n^{-2} \sum_{m=1}^n E(X^2; a_{m-1}<\vert X\vert \le a_m).\]

Swap summation order:

\[\sum_{m=1}^\infty E(X^2; a_{m-1}<\vert X\vert \le a_m) \sum_{n=m}^\infty a_n^{-2}.\]

Since ${a_n}$ is good,

\[\sum_{n=m}^\infty a_n^{-2} \le C a_m^{-2}.\]

Thus:

\[\sum_{n=1}^\infty \frac{E[Y_n^2]}{a_n^2} \le C \sum_{m=1}^\infty E\!\left(\frac{X^2}{a_m^2}; a_{m-1}<\vert X\vert \le a_m\right) \le C \sum_{m=1}^\infty m\, P(\vert X\vert >a_m) <\infty.\]

(The last inequality uses the summability assumption.)


5. Step 3 — Block Averages

Define partial sums of truncated variables:

\[T_n = \sum_{k=1}^n Y_k.\]

We will show for a geometric subsequence $n=\alpha^m$:

\[\frac{T_{\alpha^m} - E[T_{\alpha^m}]}{\alpha^m} \xrightarrow{a.s.} 0.\]

Your notes use $\alpha=2$.
Write $2^m$ as the block endpoint and define the block deviations:

\[U_m = \sum_{k=2^m+1}^{2^{m+1}} \frac{Y_k - E[Y_k]}{2^m}.\]

Then:

\[\frac{T_{2^{m+1}} - E[T_{2^{m+1}}]}{2^{m+1}} = \sum_{j=1}^{m} U_j.\]

So it suffices to show:

\[\sum_{j=1}^\infty E[U_j^2] < \infty.\]

Compute $E[U_m^2]$

From page 2 of your notes:

\[E[U_m^2] \le 4 \sum_{k=2^m+1}^{2^{m+1}} \frac{\operatorname{Var}(Y_k)}{k^2} \le 4 \sum_{k=1}^\infty \frac{E[Y_k^2]}{k^2}.\]

But Step 2 showed the latter sum is finite.
Thus $\sum_m E[U_m^2]<\infty$.

By the Kolmogorov convergence criterion (or BC I applied to $U_m^2$):

\[U_m \to 0 \quad a.s.\]

Therefore:

\[\frac{T_{2^m} - E[T_{2^m}]}{2^m} \xrightarrow{a.s.} 0.\]

6. Step 4 — Identify the Limit

We already have:

\(E[Y_k] = E(X; \vert X\vert \le k) \xrightarrow{k\to\infty} E[X] = \mu\) by Dominated Convergence.

Thus:

\[\frac{E[T_{2^m}]}{2^m} = \frac{1}{2^m} \sum_{k=1}^{2^m} E[Y_k] \xrightarrow{m\to\infty} \mu.\]

Combine with the previous step:

\[\frac{T_{2^m}}{2^m} = \frac{E[T_{2^m}]}{2^m} + \frac{T_{2^m}-E[T_{2^m}]}{2^m} \xrightarrow{a.s.} \mu.\]

7. Step 5 — Extend From the Subsequence to All $n$

For $2^m \le n \le 2^{m+1}$,

\[\frac{T_{2^m}}{2^{m+1}} \le \frac{T_n}{n} \le \frac{T_{2^{m+1}}}{2^m}.\]

Since both endpoints converge almost surely to $\mu$, the middle term must as well.

Thus:

\[\boxed{ \frac{T_n}{n} \xrightarrow{a.s.} \mu. }\]

Finally, since $X_k=Y_k$ eventually almost surely,

\[\frac{S_n}{n} = \frac{T_n}{n} + o(1) \xrightarrow{a.s.} \mu.\]

This completes the proof of the SLLN.


Cheat-Sheet Summary — Lecture 24

  • Use truncation $Y_k = X_k \mathbf{1}_{{\vert X_k\vert \le k}}$.
    BC I shows truncation changes only finitely many terms.

  • Introduce good sequences ${a_n}$ to guarantee
    \(\sum E[Y_n^2]/a_n^2 < \infty.\)

  • For the geometric subsequence $n=2^m$, define block deviations
    \(U_m = \sum_{k=2^m+1}^{2^{m+1}} (Y_k - E[Y_k]) / 2^m.\)

  • Show $\sum E[U_m^2] < \infty$ so $U_m\to 0$ a.s.

  • Conclude
    \(\frac{T_{2^m}}{2^m} \to \mu.\)

  • Squeeze all intermediate $n$, giving
    \(\frac{S_n}{n} \to \mu \quad \text{a.s.}\)

Comments