Prove that there always exists a sufficient statistic for the one-parameter exponential family of distributions. Using it, show that sample total is sufficient statistic for the parameter for the Poisson model. State whether this estimator is complete.
15 marks · directive: Prove
The core of a good answer
For f(x; θ) = exp[Q(θ)T(x) + S(x) + D(θ)], the joint density factorises as exp[Q(θ)∑T(xᵢ) + nD(θ)] × exp[∑S(xᵢ)], so ∑T(Xᵢ) is always sufficient; the Poisson pmf is exp[x log θ − θ − log x!], giving T(x) = x and the sample total ∑Xᵢ as sufficient, and ∑Xᵢ ~ Poisson(nθ) is also complete.
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
Three parts: a general proof via the factorisation theorem, its application to the Poisson, and a completeness argument. For completeness, show that E_θ[g(T)] = 0 for all θ forces g ≡ 0, using the power-series form of the Poisson expectation.
Answer outline
Introduction
One-parameter exponential family: f(x; θ) = exp[Q(θ)T(x) + S(x) + D(θ)], x in a support A that does not depend on θ.
Step 1: Joint density
- L(x; θ) = ∏ f(xᵢ; θ) = exp[Q(θ)∑T(xᵢ) + nD(θ)] × exp[∑S(xᵢ)] × ∏ I_A(xᵢ).
Step 2: Factorisation
- g(t, θ) = exp[Q(θ)t + nD(θ)] with t = ∑T(xᵢ); h(x) = exp[∑S(xᵢ)] ∏ I_A(xᵢ), free of θ.
- By the Neyman factorisation theorem, ∑T(Xᵢ) is sufficient for θ, whatever the member of the family.
Step 3: Poisson
- f(x; θ) = e^{−θ}θˣ/x! = exp[x log θ − θ − log x!], x = 0, 1, 2, …
- Q(θ) = log θ, T(x) = x, S(x) = −log x!, D(θ) = −θ.
- So T = ∑Xᵢ, the sample total, is sufficient for θ.
Step 4: Completeness
- T ~ Poisson(nθ). If E_θ[g(T)] = ∑_{t≥0} g(t) e^{−nθ}(nθ)ᵗ/t! = 0 for all θ > 0,
- then ∑ g(t) nᵗ θᵗ/t! = 0 for all θ > 0; a power series vanishing on an interval has all coefficients zero, so g(t) = 0 for every t.
- Hence ∑Xᵢ is complete (and so complete sufficient).
Step 5: Consequence
- By Lehmann-Scheffé, X̄ = ∑Xᵢ/n, an unbiased function of T, is the UMVUE of θ.
Conclusion
Every one-parameter exponential family admits the sufficient statistic ∑T(Xᵢ); for the Poisson this is the sample total, which is also complete. More generally, in a full-rank exponential family whose natural parameter Q(θ) ranges over an open interval, ∑T(Xᵢ) is complete.
Examples and data to cite
- Other members: Bernoulli (T = ∑Xᵢ), normal with known variance (T = ∑Xᵢ), exponential (T = ∑Xᵢ).
Facts to remember
- Exponential family: f(x; θ) = exp[Q(θ)T(x) + S(x) + D(θ)], support independent of θ.
- Neyman-Fisher factorisation theorem.
- Sum of n independent Poisson(θ) variables is Poisson(nθ).
- Completeness: E_θ g(T) = 0 for all θ implies P(g(T) = 0) = 1.