Cramèr-Rao Inequality “The Strategy”: • Suppose T is an unbiased estimator of a real-valued parametric function γ(θ) (i.e., Eθ (T ) = γ(θ), any θ ∈ Θ) and we wish to know if T is the UMVUE for γ(θ). • Suppose further that we know of a function of θ, say c(θ), where it holds that Varθ (T1 ) ≥ c(θ) for any unbiased estimator T1 of γ(θ) and for any θ ∈ Θ. • If you compute Varθ (T ) and find Varθ (T ) = c(θ) for all θ ∈ Θ, then you know T is the UMVUE!! Sometimes you can find such a lower bound c(θ) by the Cramèr-Rao Inequality Theorem (Cramèr-Rao Inequality): Let f (x1 , x2 , . . . , xn |θ), θ ∈ Θ, be the joint pdf/pmf of X1 , X2 , . . . , Xn . Assume that 1. Θ is an open subset of R 2. A ≡ {(x1 , x2 , . . . , xn ) : f (x1 , x2 , . . . , xn |θ) > 0} does not depend on θ 3. d f (x1 , x2 , . . . , xn |θ) exists on Θ, for all (x1 , x2 , . . . , xn ) ∈ A dθ 4. For any estimator T ∗ = T ∗ (X1 , X2 , . . . , Xn ) with Eθ (T ∗ )2 < ∞ for all θ, it holds that Z d f (x1 , x2 , . . . , xn |θ) T ∗ (x1 , x2 , . . . , xn ) dx1 dx2 . . . dxn if Xi ’s continuous ∗ dθ d Eθ (T ) A X = d f (x1 , x2 , . . . , xn |θ) dθ T ∗ (x1 , x2 , . . . , xn ) if Xi ’s discrete dθ (x1 ,x2 ,...,xn )∈A µ 5. For all θ ∈ Θ, 0 < In (θ) ≡ Eθ d log f (X1 , X2 , . . . , Xn |θ) dθ ¶2 <∞ Then, for any unbiased estimator T of γ(θ), it holds that (γ 0 (θ))2 ´2 = ³ In (θ) d log f (X1 ,X2 ,...,Xn |θ) (γ 0 (θ))2 Varθ (T ) ≥ Eθ for all θ ∈ Θ (1) dθ where γ 0 (θ) ≡ d γ(θ)/dθ is assumed to exist on Θ. Remarks: • The right-hand side of (1) is called the Cramèr-Rao Lower Bound. • Conditions 1 - 5 in the Theorem are called the “Cramèr-Rao Regularity Condtions” They are satisfied if X1 , X2 , . . . , Xn are a random sample from the 1-parameter exponential family. Eg., Binomial(n, θ), Poisson(θ), Geometric(θ), N(θ, σ 2 ), N(µ, θ), gamma(α, θ), gamma(θ, β), etc. 1 • In (θ) is called the Fisher Information number (for size n sample) • If X1 , X2 , . . . , Xn are iid with common pdf/pmf f (x|θ), then µ ¶2 d log f (X1 |θ) In (θ) = nI1 (θ), where I1 (θ) = Eθ dθ (2) and I1 (θ) represents the Fisher information for one observation. d2 f (x1 , x2 , . . . , xn |θ) exits on Θ, for all (x1 , x2 , . . . , xn ) ∈ A, then dθ2 ¶2 µ 2 µ ¶ d log f (X1 , X2 , . . . , Xn |θ) d log f (X1 , X2 , . . . , Xn |θ) = −Eθ In (θ) = Eθ . dθ dθ2 • If If, in addition, X1 , X2 , . . . , Xn are iid with common pdf/pmf f (x|θ), then we have µ ¶2 µ 2 ¶ d log f (X1 |θ) d log f (X1 |θ) In (θ) = nI1 (θ) where I1 (θ) = Eθ = −Eθ dθ dθ2 Proof of Equation (2)/continuous case. Just using definitions, for any sample size n, µ ¶ d log f (X1 , X2 , . . . , Xn |θ) Eθ dθ Z d log f (x1 , x2 , . . . , xn |θ) = f (x1 , x2 , . . . , xn |θ)dx1 , . . . , dxn dθ ZA d f (x1 , x2 , . . . , xn |θ) f (x1 , x2 , . . . , xn |θ) = dx1 , . . . , dxn derivative of log dθ f (x1 , x2 , . . . , xn |θ) A Z d f (x1 , x2 , . . . , xn |θ) = 1· dx1 , . . . , dxn dθ A d Eθ (1) = by condition (4) of Theorem with T ∗ = 1 dθ d = 1=0 dθ so that µ In (θ) = = = = = = = ¶2 d log f (X1 , X2 , . . . , Xn |θ) Eθ dθ µ ¶ d log f (X1 , X2 , . . . , Xn |θ) Varθ dθ Ã ! n Qn d X Varθ log f (Xi |θ) since f (x1 , x2 , . . . , xn |θ) = i=1 f (xi |θ) dθ i=1 ! Ã n X d log f (Xi |θ) Varθ dθ i=1 ¶ µ n X d log f (Xi |θ) sum of independent variables Varθ dθ i=1 µ ¶ d log f (X1 |θ) nVarθ iid variables dθ µ ¶2 d log f (X1 |θ) nEθ = nI1 (θ) dθ 2