Cramér–Rao Estimator Efficiency¶

Compare an unbiased scalar estimator's variance with its regular-model Cramér–Rao information bound, under explicit conditions.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13104
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomains: Estimation Theory, Parametric Inference → Experimental Design & Statistics

Core Idea¶

Cramér–Rao estimator efficiency compares an unbiased scalar estimator's variance with the inverse-Fisher-information lower bound for the same regular model, parameter and sample. If $I_n(\theta)>0$ is information in the sample and $T$ has positive finite variance, the pointwise score is $e_\theta(T)=[1/I_n(\theta)]/\operatorname{Var}_\theta(T)$. Under the Rao–Cramér conditions, $0<e_\theta(T)\le1$; equality means this estimator attains the bound at that parameter value. Uniform equality over the parameter space is a stronger claim. The bound need not be attainable, so a subunit ratio is not automatically removable variance slack.[^ref-f2ba320a2b47]

This is a narrow reusable metric, not the whole frozen Wikipedia page “Efficiency (statistics).” Requested “Efficient estimator” and “Relative efficiency” remain separate unresolved identity senses despite redirects; test, design and asymptotic efficiencies are not absorbed. The broad Efficiency slug already belongs to a live prime.[^ref-9e103cd82d91]

Scope of Application¶

The ratio requires an unbiased estimator, regular scalar likelihood, positive finite information and variance, and the support/differentiation conditions that justify the information inequality. MIT's normal-location and Bernoulli-proportion examples satisfy these conditions; its Uniform$[0,\theta]$ example has parameter-dependent support and breaks the naive bound. Biased estimators require attention to MSE or a different bound, and a vector or asymptotic problem needs its own formulation.[^{ref-f2ba320a2b47][}ref-0437a4304ab7]

Clarity¶

“Efficient” is not a context-free compliment. This score distinguishes a proved lower bound, attainment by a particular estimator, and broader optimality under a chosen loss. An estimator can be minimum-variance among unbiased rules without attaining a loose Cramér–Rao bound; a biased rule with different MSE cannot be ranked by this narrow ratio alone. Relative efficiency compares two procedures and may exceed one depending on numerator convention; it is not this information-bound ratio.[^{ref-f2ba320a2b47][}ref-9e103cd82d91]

The proposed strict DAG prerequisites are live Estimator and Fisher information. Live Efficiency is related but not asserted as a parent: a theorem lower bound is not necessarily a feasible alternative on that prime's dominated-slack frontier.

Manages Complexity¶

Within one regular experiment, the score compresses an estimator's precision comparison to the model's lower bound and the rule's actual variance. It reveals when an estimator discards useful information: in both worked models, using all $n$ observations scores one, while using only the first of those same $n$ observations scores $1/n$. The compression hides bias, model misspecification, robustness and bound attainability, which must be restored before any broader decision.[^{ref-f2ba320a2b47][}ref-0437a4304ab7]

Abstract Reasoning¶

Declare the likelihood, scalar target, sample size and unbiased estimator. Verify support and differentiation regularity; compute information for the entire sample and the estimator's variance under the same $\theta$. Apply the lower-bound theorem, form the ratio, and distinguish a pointwise from a whole-model equality claim. If support changes with $\theta$, stop: the simple information expression may be invalid rather than an estimator having paradoxical efficiency above one.[^ref-f2ba320a2b47]

Knowledge Transfer¶

For IID $N(\mu,\sigma^2)$ observations with known $\sigma^2$, the sample mean has variance and bound $\sigma^2/n$, hence score one; $X_1$ from the same $n$ observations scores $1/n$. For IID Bernoulli$(p)$ trials with $0<p<1$, the sample proportion has variance and bound $p(1-p)/n$, again score one; using only $X_1$ scores $1/n$. The continuous and discrete carriers differ, but model, estimator, information bound, variance and ratio map exactly. That mapping does not transfer to test power, design optimality or asymptotic mean-versus-median comparisons without changing the identity.[^{ref-f2ba320a2b47][}ref-9e103cd82d91]

[^ref-f2ba320a2b47]: Anna Mikusheva, 14.381 Statistical Method in Economics, Lecture 6: Efficient Estimators, Rao-Cramer Bound, MIT OpenCourseWare original instructor notes (2018), PDF pp.3–6. [^ref-0437a4304ab7]: STATS 200: Introduction to Statistical Inference, Lecture 29, Stanford original course notes, PDF slides 16–20. [^ref-9e103cd82d91]: Charles J. Geyer, Statistics 5102 course slides, University of Minnesota original instructor material, PDF slides 48–56 and 244–245.

Relationships to Other Abstractions¶

Current abstraction Cramér–Rao Estimator Efficiency Domain-specific

Parents (2) — more general patterns this builds on

Cramér–Rao Estimator Efficiency presupposes Estimator Domain-specific

This score rates the variance of a specified statistical estimator.
Cramér–Rao Estimator Efficiency presupposes Fisher information Domain-specific

The score's regular-model benchmark is inverse Fisher information.

Hierarchy paths (2) — routes to 2 parentless roots

Cramér–Rao Estimator Efficiency → Estimator

Show alternative path (1)

Neighborhood in Abstraction Space¶

Cramér–Rao Estimator Efficiency sits in a sparse region of the domain-specific corpus (73^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Learning & Model Failure Modes (41 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08