2024.Q4 – Lindeberg Condition, CLT, and Variance Asymptotics

2024 Probability Prelim Exam (PDF)

Problem Statement (verbatim)

Let ${(X_k, Y_k)}_{k=1,2,\dots}$ be a sequence of pairs of random variables. Denote
\(S_n = \sum_{k=1}^n X_k,\quad T_n = \sum_{k=1}^n Y_k,\quad n = 1,2,\dots\)

a. Assume that $\sum_{k=1}^\infty P(X_k \ne Y_k) < \infty$, and let $a_n \to \infty$. Prove that if $\frac{T_n}{a_n}$ converges in distribution to $W$, then $\frac{S_n}{a_n}$ converges in distribution to $W$ as well.

b. From now on assume that ${X_k}{k=1,2,\dots}$ are independent and \(X_1 = 0,\quad X_k = \begin{cases} \pm 1 & \text{with probability } \frac{1}{2} - \frac{1}{2k^2},\\ \pm k & \text{with probability } \frac{1}{2k^2}, \end{cases} \quad k = 2,3,\dots\) Let $Y_0 = 0$, $Y_k = X_k\cdot \mathbf{1}{{X_k=\pm 1}}, k = 1,2,\dots$. Prove that \(\frac{\mathrm{Var}(S_n)}{2n} \longrightarrow 1 \quad\text{and}\quad \frac{\mathrm{Var}(T_n)}{n} \longrightarrow 1.\)

c. (i) Does the triangular array \(\Big\{\tfrac{X_k}{\sqrt{2n}}\Big\}_{k=1,\dots,n,\;n=1,2,\dots}\) satisfy Lindeberg condition? What about the triangular array \(\Big\{\tfrac{Y_k}{\sqrt{n}}\Big\}_{k=1,\dots,n,\;n=1,2,\dots}?\)

(ii) Prove that $\sqrt{\tfrac{S_n}{n}}$ converges in distribution to $N(0,1)$.


Part (a)

Claim.

Assume $\sum_{k=1}^\infty P(X_k\ne Y_k) < \infty$ and $a_n\to\infty$.
If $\dfrac{T_n}{a_n} \xrightarrow{d} W$, then $\dfrac{S_n}{a_n} \xrightarrow{d} W$ as well.

Proof.

Let \(A_k = \{X_k \ne Y_k\}.\) Since \(\sum_{k=1}^\infty P(A_k) < \infty,\) by the first Borel–Cantelli lemma we have \(P(A_k \text{ i.o.}) = 0.\) So, with probability 1, there exists a (random) index $K(\omega)$ such that
$X_k(\omega) = Y_k(\omega)$ for all $k \ge K(\omega)$.

Define the (almost surely finite) random variable \(D(\omega) = \sum_{k=1}^\infty \big(X_k(\omega) - Y_k(\omega)\big).\) This is a finite sum because all but finitely many terms are zero a.s.

For each finite $n$, \(S_n - T_n = \sum_{k=1}^n (X_k - Y_k).\) For $n \ge K(\omega)$, all terms with $k > K(\omega)$ vanish, so \(S_n(\omega) - T_n(\omega) = \sum_{k=1}^\infty (X_k(\omega) - Y_k(\omega)) = D(\omega).\) Thus a.s. we have \(S_n - T_n = \begin{cases} \text{some partial sum}, & n < K(\omega), D, & n \ge K(\omega). \end{cases}\) In particular, \(|S_n - T_n| \le |D| \quad\text{a.s. for all }n.\)

Since $a_n\to\infty$, \(\frac{|S_n - T_n|}{a_n} \le \frac{|D|}{a_n} \xrightarrow[n\to\infty]{} 0 \quad\text{a.s.}\) Hence \(\frac{S_n - T_n}{a_n} \xrightarrow{P} 0.\)

Now use Slutsky’s theorem:

  • $\displaystyle \frac{T_n}{a_n} \xrightarrow{d} W$ by assumption,
  • $\displaystyle \frac{S_n - T_n}{a_n} \xrightarrow{P} 0.$

Thus \(\frac{S_n}{a_n} = \frac{T_n}{a_n} + \frac{S_n - T_n}{a_n} \xrightarrow{d} W + 0 = W.\)

Conclusion.

Because $\sum P(X_k\ne Y_k)<\infty$, the difference $S_n - T_n$ is eventually constant a.s., and the normalized difference tends to 0. By Slutsky, $S_n/a_n$ has the same limit in distribution as $T_n/a_n$.

Key Takeaways

  • Borel–Cantelli (1): if $\sum P(A_k)<\infty$, then $A_k$ occurs only finitely many times a.s.
  • That gives “eventual equality” between two processes, so their difference becomes a finite a.s. constant.
  • Dividing a bounded (or finite a.s.) random variable by $a_n\to\infty$ gives convergence to 0, at least in probability.
  • Slutsky’s theorem: small perturbations (in probability) do not affect the distributional limit.

Part (b)

Claim.

With $S_n = \sum_{k=1}^n X_k$ and $T_n = \sum_{k=1}^n Y_k$, \(\frac{\mathrm{Var}(S_n)}{2n} \longrightarrow 1 \quad\text{and}\quad \frac{\mathrm{Var}(T_n)}{n} \longrightarrow 1.\)

Proof.

The $X_k$ are independent, so \(\mathrm{Var}(S_n) = \sum_{k=1}^n \mathrm{Var}(X_k).\)

  • For $k=1$, we are given $X_1=0$, so $\mathrm{Var}(X_1)=0$.
  • For $k\ge2$, we have \(X_k = \begin{cases} \pm 1 & \text{with prob } \frac{1}{2} - \frac{1}{2k^2},\\ \pm k & \text{with prob } \frac{1}{2k^2}. \end{cases}\) The distribution is symmetric, so $E[X_k]=0$.

    Then $$ E[X_k^2] = 1^2 \cdot 2\left(\frac{1}{2} - \frac{1}{2k^2}\right)

    • k^2 \cdot 2\left(\frac{1}{2k^2}\right) = \left(1 - \frac{1}{k^2}\right) + 1 = 2 - \frac{1}{k^2}. $$ Hence $\mathrm{Var}(X_k)=E[X_k^2]=2-\frac1{k^2}$.

So \(\mathrm{Var}(S_n) = \sum_{k=2}^n \left(2 - \frac{1}{k^2}\right) = 2(n-1) - \sum_{k=2}^n \frac{1}{k^2}.\) Rewriting, \(\mathrm{Var}(S_n) = 2n - 2 - \sum_{k=2}^n \frac{1}{k^2} = 2n - 1 - \sum_{k=1}^n \frac{1}{k^2}.\)

Now \(\lim_{n\to\infty} \sum_{k=1}^n \frac{1}{k^2} = \frac{\pi^2}{6} < \infty,\) so $$ \lim_{n\to\infty} \frac{\mathrm{Var}(S_n)}{2n} = \lim_{n\to\infty} \left( 1 - \frac{1}{2n}

  • \frac{1}{2n} \sum_{k=1}^n \frac{1}{k^2} \right) = 1. $$

Next, consider $T_n = \sum_{k=1}^n Y_k$, where \(Y_k = X_k \cdot \mathbf{1}_{\{X_k=\pm1\}}.\) Then:

  • $Y_1 = X_1\cdot 1_{{X_1=\pm1}}=0$, so $\mathrm{Var}(Y_1)=0$.
  • For $k\ge2$, $Y_k$ takes values $\pm1$ with total probability $1 - 1/k^2$, and is 0 otherwise. Again symmetry gives $E[Y_k]=0$, and \(E[Y_k^2] = 1^2 \cdot 2\left(\frac{1}{2} - \frac{1}{2k^2}\right) = 1 - \frac{1}{k^2}.\) So $\mathrm{Var}(Y_k)=1 - \frac{1}{k^2}$.

Then \(\mathrm{Var}(T_n) = \sum_{k=2}^n \left(1 - \frac{1}{k^2}\right) = (n-1) - \sum_{k=2}^n \frac{1}{k^2} = n - \sum_{k=1}^n \frac{1}{k^2}.\)

Thus \(\frac{\mathrm{Var}(T_n)}{n} = 1 - \frac{1}{n}\sum_{k=1}^n \frac{1}{k^2} \longrightarrow 1.\)

Conclusion.

The “big jumps” in $X_k$ ($\pm k$) happen rarely enough that the variance is asymptotically linear in $n$: $\mathrm{Var}(S_n)\sim 2n$, $\mathrm{Var}(T_n)\sim n$.

Key Takeaways

  • For sums of independent variables, variance is additive.
  • Symmetry makes $E[X_k]=0$ and simplifies variance to just $E[X_k^2]$.
  • Rare but large values (here $\pm k$ with prob $1/k^2$) contribute a finite adjustment to the variance growth.
  • Harmonic-type series $\sum 1/k^2$ converge, so they appear as $O(1)$ corrections.

Part (c)(i)

We consider two triangular arrays:

  • $\xi_{n,k} = \dfrac{X_k}{\sqrt{2n}},\quad k=1,\dots,n,$
  • $\eta_{n,k} = \dfrac{Y_k}{\sqrt{n}},\quad k=1,\dots,n.$

Let \(s_n^2 := \sum_{k=1}^n \mathrm{Var}(\xi_{n,k}), \quad t_n^2 := \sum_{k=1}^n \mathrm{Var}(\eta_{n,k}).\)

From part (b), \(s_n^2 = \frac{\mathrm{Var}(S_n)}{2n} \to 1, \quad t_n^2 = \frac{\mathrm{Var}(T_n)}{n} \to 1.\)

Recall Lindeberg’s condition:
For a triangular array ${\zeta_{n,k}}$ with total variance $v_n^2=\sum_k\mathrm{Var}(\zeta_{n,k})$, \(\forall \varepsilon>0:\quad \frac{1}{v_n^2} \sum_{k=1}^n E\big[\zeta_{n,k}^2\mathbf{1}_{\{|\zeta_{n,k}|>\varepsilon\}}\big] \longrightarrow 0.\)


Claim (1).

The $X_k$-array $\xi_{n,k} = X_k/\sqrt{2n}$ does not satisfy Lindeberg’s condition.

Proof.

For large $n$, the total variance $s_n^2\to1$, so we can ignore the normalization factor (it stays bounded away from 0 and ∞).

Fix $\varepsilon>0$. We examine \(\sum_{k=1}^n E\left[\xi_{n,k}^2 \mathbf{1}_{\{|\xi_{n,k}|>\varepsilon\}}\right] = \sum_{k=1}^n \frac{1}{2n} E\left[X_k^2 \mathbf{1}_{\{|X_k| > \varepsilon\sqrt{2n}\}}\right].\)

For each $k\ge2$, the “big” values of $X_k$ are $\pm k$, occurring with total probability $1/k^2$; otherwise, $ X_k =1$. For large $n$, the condition $ X_k > \varepsilon\sqrt{2n}$ will be satisfied precisely when:
  • $X_k = \pm k$ and
  • $k > \varepsilon\sqrt{2n}$.

Thus, for large $n$, \(E\left[X_k^2 \mathbf{1}_{\{|X_k| > \varepsilon\sqrt{2n}\}}\right] = k^2\cdot P(|X_k|=k,\;k>\varepsilon\sqrt{2n}) = k^2\cdot \frac{1}{k^2}\mathbf{1}_{\{k>\varepsilon\sqrt{2n}\}} = \mathbf{1}_{\{k>\varepsilon\sqrt{2n}\}}.\)

Hence \(\sum_{k=1}^n E\left[\xi_{n,k}^2 \mathbf{1}_{\{|\xi_{n,k}|>\varepsilon\}}\right] \approx \frac{1}{2n}\sum_{k>\varepsilon\sqrt{2n}}^n 1 = \frac{1}{2n} \big(n - \lfloor \varepsilon\sqrt{2n}\rfloor\big) \longrightarrow \frac12 \ne 0.\)

Dividing by $s_n^2\to1$ does not change this limit. Therefore the Lindeberg expression does not go to 0, so the Lindeberg condition fails for ${\xi_{n,k}}$.

Conclusion (1).

The triangular array ${X_k/\sqrt{2n}}$ fails Lindeberg’s condition because the rare “big jumps” of size $k$ contribute a non-vanishing amount to the Lindeberg sum.


Claim (2).

The $Y_k$-array $\eta_{n,k} = Y_k/\sqrt{n}$ does satisfy Lindeberg’s condition.

Proof.

Recall $Y_k \in {-1,0,1}$. Therefore, \(|\eta_{n,k}| = \frac{|Y_k|}{\sqrt{n}} \le \frac{1}{\sqrt{n}}.\)

Fix $\varepsilon>0$. For all sufficiently large $n$, \(\frac{1}{\sqrt{n}} < \varepsilon \quad\Rightarrow\quad |\eta_{n,k}| \le \varepsilon\quad\text{for all }k=1,\dots,n.\) This means the indicator $\mathbf{1}{{|\eta{n,k}|>\varepsilon}}$ is identically zero for all large $n$. Thus \(\sum_{k=1}^n E\left[\eta_{n,k}^2 \mathbf{1}_{\{|\eta_{n,k}|>\varepsilon\}}\right] = 0\) for large enough $n$, and after dividing by $t_n^2\to1$ we still get 0.

So the Lindeberg condition holds for ${\eta_{n,k}}$.

Conclusion (2).

The triangular array ${Y_k/\sqrt{n}}$ satisfies Lindeberg’s condition because the individual normalized terms become uniformly small as $n\to\infty$.

Key Takeaways

  • Lindeberg condition detects whether rare large jumps are negligible under the chosen normalization.
  • For $X_k/\sqrt{2n}$, the rare ±k jumps are still “large enough” to violate Lindeberg.
  • For $Y_k/\sqrt{n}$, the values are always in $[-1,1]$, so after dividing by $\sqrt{n}$ they are uniformly tiny, and the Lindeberg condition is automatically satisfied.

Part (c)(ii)

Claim.

\(\sqrt{\frac{S_n}{n}} \;\xrightarrow{d}\; N(0,1).\)

(Equivalently, $\frac{S_n}{\sqrt{n}} \xrightarrow{d} N(0,1)$.)

Proof.

We already know from part (b) that \(\mathrm{Var}(T_n) \sim n,\) and from part (c)(i) that the triangular array \(\eta_{n,k} = \frac{Y_k}{\sqrt{n}}\) satisfies Lindeberg’s condition, with total variance $t_n^2 \to 1$.

By the Lindeberg–Feller CLT for triangular arrays, \(\frac{T_n}{\sqrt{n}} \xrightarrow{d} N(0,1).\)

Now write \(S_n = T_n + D_n,\quad\text{where}\quad D_n := \sum_{k=2}^n X_k \mathbf{1}_{\{|X_k|=k\}}\) collects the “big jump” contributions.

Note \(P(|X_k|=k) = \frac{1}{k^2},\) so \(\sum_{k=2}^\infty P(|X_k|=k) = \sum_{k=2}^\infty \frac{1}{k^2} < \infty.\) By the Borel–Cantelli lemma (1), \(P(|X_k|=k \text{ i.o.}) = 0.\) Thus only finitely many of the events ${|X_k|=k}$ occur a.s., and the series \(D_\infty := \sum_{k=2}^\infty X_k \mathbf{1}_{\{|X_k|=k\}}\) converges a.s. to a finite (random) limit. In particular, $D_n\to D_\infty$ a.s., hence \(\frac{D_n}{\sqrt{n}} \xrightarrow{P} 0.\)

Now \(\frac{S_n}{\sqrt{n}} = \frac{T_n}{\sqrt{n}} + \frac{D_n}{\sqrt{n}}.\)

We already have $\frac{T_n}{\sqrt{n}} \xrightarrow{d} N(0,1)$ and $\frac{D_n}{\sqrt{n}} \xrightarrow{P} 0$. By Slutsky’s theorem, \(\frac{S_n}{\sqrt{n}} \xrightarrow{d} N(0,1).\)

Conclusion.

Even though $X_k$ occasionally takes large values ±k, those large jumps occur so rarely that after normalization by $\sqrt{n}$, they vanish in probability. The main contribution is from the “±1” part captured by $T_n$, and thus $S_n/\sqrt{n}$ also converges in distribution to $N(0,1)$.

Key Takeaways

  • Strategy:
    1. Isolate the “nice” part $T_n$ that satisfies Lindeberg and has variance $\sim n$.
    2. Show the “bad” part $D_n$ contributes only a negligible term after normalization.
  • Rare but large values can be handled with Borel–Cantelli and then Slutsky.
  • This is a classic pattern:
    • decompose $S_n = \text{nice part} + \text{rare spikes}$,
    • prove CLT for the nice part,
    • show rare spikes are negligible.

Global Key Takeaways for Question 4

  • Part (a): Using Borel–Cantelli to show eventual equality of two processes, then Slutsky to transfer limits.
  • Part (b): Computing variances explicitly to see linear growth, with a convergent correction term.
  • Part (c)(i): Understanding Lindeberg’s condition as “big jumps vanish in aggregate” under normalization.
  • Part (c)(ii): Combining Lindeberg–Feller CLT, Borel–Cantelli, and Slutsky to handle heavy tails in a controlled way.

This problem is a great template for:

  • how to compare two sums,
  • how to separate “nice” and “rare spike” parts of a sequence,
  • and how to assemble multiple probabilistic tools into one coherent CLT proof.

Comments