8 1. The Fisher Efficiency In the regular case, the differentiation and integration are interchangeable, hence differentiating in θ , we get the equation, 1 + b n (θ) = Rn ˆ n (x1,...,xn) ∂p (x1,...,xn,θ)/∂θ dx1 . . . dxn = Rn ˆ n (x1,...,xn) ∂p (x1,...,xn,θ)/∂θ p (x1,...,xn,θ) p (x1,...,xn,θ) dx1 . . . dxn = ˆ n L n (θ) = Covθ ˆ n , L n (θ) where we use the fact that L n (θ) = 0. The correlation coefficient ρn of ˆ n and Ln(θ) does not exceed 1 in its absolute value, so that 1 ρ2 n = ( Covθ ˆ n , L n (θ) ) 2 Varθ[ˆ n ] Varθ[Ln(θ)] = (1 + bn(θ))2 Varθ[ˆ n ] In(θ) . 1.4. Efficiency of Estimators An immediate consequence of Theorem 1.8 is the formula for unbiased esti- mators. Corollary 1.9. For an unbiased estimator ˆ n , the Cram´ er-Rao inequality (1.1) takes the form (1.2) Varθ ˆ n 1 In(θ) , θ Θ. An unbiased estimator θ n = θ n (X1,...,Xn) in a regular statistical experiment is called Fisher efficient (or, simply, efficient) if, for any θ Θ, the variance of θ n reaches the Cram´ er-Rao lower bound, that is, the equality in (1.2) holds: Varθ θn = 1 In(θ) , θ Θ. Example 1.10. Suppose, as in Example 1.1(a), the observations X1,...,Xn are independent N (θ, σ2) where σ2 is assumed known. We show that the sample mean ¯ n = (X1 + · · · + Xn)/n is an efficient estimator of θ. Indeed, ¯ n is unbiased and Varθ ¯ n = σ2/n. On the other hand, ln p (X, θ) = 1 2 ln(2 π σ2) (X θ)2 2σ2 and l (X , θ) = ln p (X, θ) ∂θ = X θ σ2 . Thus, the Fisher information for the statistical experiment is In(θ) = n ( l (X , θ) ) 2 = n σ4 (X θ)2 = nσ2 σ4 = n σ2 .
Previous Page Next Page