Let X₁, X₂, ⋯ be a sequence of independent random variables such that
P[Xₖ = n] = 1/n and P[Xₖ = 0] = 1 − 1/n, k = 1, 2, ⋯
Check whether the weak law of large numbers (WLLN) holds for this sequence or not.
5 marks · directive: Check
The core of a good answer
Reading the printed n as the index k (P[Xₖ = k] = 1/k), E(Xₖ) = 1 and V(Xₖ) = k − 1, so V(Sₙ)/n² = (n − 1)/(2n) → 1/2 ≠ 0 and Chebyshev's condition fails; a direct argument shows P(Sₙ/n ≤ 1/2) ≥ 1/14 for every n ≥ 4, so Sₙ/n does not converge in probability to 1 and the WLLN does not hold.
An outline written by upsc.help, not an official UPSC answer; every fact checked by a separate review. How we work
What the question asks
As printed the law of Xₖ uses n while the index is k. Read literally, with n a fixed constant, the Xₖ would be i.i.d. with mean 1 and the WLLN would hold by Khinchin's theorem; the reading P[Xₖ = k] = 1/k, P[Xₖ = 0] = 1 − 1/k is used below, and an answer should state which reading it takes. The usual first step is Chebyshev's sufficient condition V(Sₙ)/n² → 0; since it fails, a complete answer must also show directly that Sₙ/n does not settle down.
Answer outline
Introduction
Sₙ = X₁ + ⋯ + Xₙ. The WLLN holds if (Sₙ − E Sₙ)/n → 0 in probability.
Step 1: Moments
- E(Xₖ) = k × (1/k) = 1; E(Xₖ²) = k² × (1/k) = k; V(Xₖ) = k − 1.
- E(Sₙ) = n; V(Sₙ) = ∑(k − 1) = n(n − 1)/2.
Step 2: Chebyshev's condition
- V(Sₙ)/n² = (n − 1)/(2n) → 1/2 ≠ 0.
- So the sufficient condition fails; this alone does not settle the question.
Step 3: Direct proof that the law fails
- Fix m = ⌊n/4⌋ and let E be the event that Xₖ = 0 for every k with m < k ≤ n.
- P(E) = ∏_{k=m+1}^{n} (1 − 1/k) = m/n (telescoping product), and m/n ≥ 1/7 for n ≥ 4.
- On E, Sₙ = X₁ + ⋯ + Xₘ, which is independent of E and has mean m; by Markov's inequality P(X₁ + ⋯ + Xₘ > n/2) ≤ 2m/n ≤ 1/2.
- So P(Sₙ/n ≤ 1/2) ≥ (m/n)(1 − 2m/n) ≥ (1/7)(1/2) = 1/14 for every n ≥ 4, while E(Sₙ)/n = 1.
- Hence P(|Sₙ/n − 1| ≥ 1/2) does not tend to 0: the WLLN fails.
Step 4: Interpretation and check
- Most of Sₙ comes from the rare large values Xₖ = k with k close to n, so Sₙ/n keeps a non-degenerate spread.
- Python simulation: for n = 1000 and 10000, V(Sₙ/n) ≈ 0.50 and P(|Sₙ/n − 1| > 0.5) ≈ 0.5, not shrinking.
Conclusion
The weak law of large numbers does not hold for this sequence: Sₙ/n does not settle down to E(Sₙ)/n = 1. The trouble is that a few rare but very large values (Xₖ = k) dominate the average.
Examples and data to cite
- Contrast: if P[Xₖ = ±k^α] = 1/2, V(Sₙ)/n² is of order n^{2α−1}, which tends to 0 for α < 1/2, so Chebyshev's condition holds there.
Facts to remember
- Chebyshev's WLLN: if the variance of the average, V(Sₙ)/n², tends to 0, then (Sₙ − E Sₙ)/n → 0 in probability.
- ∑_{k=1}^{n} (k − 1) = n(n − 1)/2.
- Telescoping product: ∏_{k=m+1}^{n} (1 − 1/k) = m/n.
- Markov's inequality: P(Y > a) ≤ E(Y)/a for Y ≥ 0, a > 0.
- Khinchin's WLLN (1929): for i.i.d. variables a finite mean suffices.