Under standard regularity conditions for a parametric statistical model with likelihood p(x; θ), the variance of any unbiased estimator (θ̂) of a scalar parameter θ is lower-bounded by the reciprocal of the Fisher information: Var(θ̂) ≥ 1 / I(θ), where I(θ)=E[(∂/∂θ log p(X;θ))^2]. In the multivariate case Cov(θ̂) ≥ I(θ)^{-1} in the matrix sense. The bound quantifies a fundamental information–precision trade-off for unbiased estimators.

Principle

Fisher information measures the sensitivity of the model's likelihood to parameter changes; greater Fisher information implies a smaller lower bound on achievable unbiased estimator variance—information bounds precision.

Demonstration

Illustrative scenario → Estimating the mean μ of i.i.d. Gaussian samples with known variance σ^2: the sample mean is unbiased with Var(μ̂)=σ^2/n and Fisher information I(μ)=n/σ^2, so Var(μ̂)=1/I(μ), achieving the Cramér–Rao bound and demonstrating attainability under regularity and sufficiency.

Misapplication

Using the bound for biased estimators without accounting for bias (ignoring bias contribution to mean-squared error), applying it when regularity conditions fail (non-differentiable likelihood, parameter on the boundary), or treating it as necessarily tight for small samples; the error is misremembering scope and required conditions.

Consequence

Provides a benchmark for estimator performance and guides the design of efficient estimators (e.g., maximum likelihood achieves the bound asymptotically under regularity); it also formalizes how model information limits achievable precision for unbiased estimation.

Reversal

When estimators are biased, the mean-squared error can be below the CR bound for variance alone; in models with irregular likelihoods, discrete parameters, or failure of regularity conditions, the CR bound may not hold or require modification (e.g., van Trees, Hammersley–Chapman–Robbins bounds).

Boundary

Clearly within → parametric models with differentiable likelihoods, interior parameter values, and finite Fisher information where unbiased estimators are considered. Boundary case → small-sample regimes where asymptotic efficiency concepts may mislead. Clearly outside → nonparametric settings lacking finite-dimensional Fisher information or problems with nonregular parameters.

Semantic Tension

Bias–variance trade-off ↔ Cramér–Rao focus on unbiased variance: minimizing variance subject to unbiasedness may be suboptimal for overall MSE, so practical estimator design often balances bias to reduce MSE despite the CR bound's variance limit.

Synthesis

The Cramér–Rao bound links differential information in the likelihood to a formal lower limit on the variance of unbiased estimators and serves as both a diagnostic (how informative is the model) and a target (efficient estimators aim to attain it asymptotically). Its practical use requires checking regularity, considering bias, and recognizing alternatives when assumptions fail.