A random sample of size n is drawn from N(θ, σ²). Let the prior distribution of θ be N(μ, τ²), where σ², μ and τ² are known. Obtain the Bayes estimator for θ under (i) squared error loss function and (ii) absolute error loss function.
15 marks · directive: Obtain
The core of a good answer
The posterior of θ is normal with mean (nτ²x̄ + σ²μ)/(nτ² + σ²) and variance σ²τ²/(nτ² + σ²); under squared error the Bayes estimator is the posterior mean, and under absolute error it is the posterior median, which for a normal posterior is the same value θ̂ = (nτ²x̄ + σ²μ)/(nτ² + σ²).
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
The steps are: derive the posterior by multiplying likelihood and prior and completing the square in θ, then recall which posterior summary minimises each loss (mean for squared error, median for absolute error). The examiner expects the derivation, not just the conjugate formula.
Answer outline
Introduction
Likelihood depends on the data through x̄: L(θ) ∝ exp[−n(x̄ − θ)²/(2σ²)]. Prior: π(θ) ∝ exp[−(θ − μ)²/(2τ²)].
Step 1: Posterior kernel
- π(θ | x) ∝ exp{−½[n(θ − x̄)²/σ² + (θ − μ)²/τ²]}.
Step 2: Complete the square
- Coefficient of θ²: n/σ² + 1/τ² = 1/τₙ², so τₙ² = σ²τ²/(nτ² + σ²).
- Linear term gives the mean: μₙ = τₙ²(nx̄/σ² + μ/τ²) = (nτ²x̄ + σ²μ)/(nτ² + σ²).
- θ | x ~ N(μₙ, τₙ²).
Step 3: Squared error loss
- Posterior expected loss E[(θ − a)² | x] is minimised at a = E(θ | x).
- θ̂ = (nτ²x̄ + σ²μ)/(nτ² + σ²).
Step 4: Absolute error loss
- E[|θ − a| | x] is minimised at a = posterior median.
- A normal distribution is symmetric, so median = mean: θ̂ = (nτ²x̄ + σ²μ)/(nτ² + σ²).
Step 5: Interpretation
- θ̂ = w x̄ + (1 − w)μ, w = nτ²/(nτ² + σ²): a precision-weighted average of sample mean and prior mean.
- As n → ∞, w → 1 and θ̂ → x̄; as τ² → 0, θ̂ → μ.
Conclusion
Both losses give θ̂ = (nτ²x̄ + σ²μ)/(nτ² + σ²), with posterior variance σ²τ²/(nτ² + σ²). The estimator shrinks x̄ toward the prior mean μ.
Examples and data to cite
- With σ² = 4, τ² = 1, n = 4: w = 4/(4 + 4) = 1/2, so the estimate is the simple average of x̄ and μ.
Facts to remember
- Normal prior is conjugate for a normal mean with known variance.
- Posterior precision = data precision n/σ² + prior precision 1/τ².
- Squared error loss ⇒ posterior mean; absolute error loss ⇒ posterior median.
- For a symmetric unimodal posterior, mean = median = mode.