Unbiased Estimation & UMVUE — Unbiased estimation and UMVUE concepts (Lehmann–Scheffé intro) - Statistics Graduate Level

SECTION 1

Historical Context & Motivation

The search for optimal estimators is one of the central themes in the history of mathematical statistics. When confronted with a statistical model and tasked with estimating an unknown parameter, a natural question arises: among all estimators that are unbiased — meaning their expected value equals the true parameter for every possible parameter value — can we find one with the smallest possible variance? This question drove some of the twentieth century's most elegant theoretical developments, culminating in the concept of the Uniformly Minimum Variance Unbiased Estimator (UMVUE) and the celebrated Lehmann–Scheffé theorem that provides a practical recipe for constructing one.

1922 • Fisher's Foundations

R. A. Fisher introduces the concepts of sufficiency, efficiency, and consistency in his landmark paper, laying the conceptual groundwork for optimality in estimation.

1935 • Rao–Blackwell Idea Emerges

The seeds of the Rao–Blackwell theorem take shape as C. R. Rao and David Blackwell independently show that conditioning an unbiased estimator on a sufficient statistic never increases variance.

1945–1947 • Completeness & Lehmann–Scheffé

Erich Lehmann and Henry Scheffé formalize the notion of a complete sufficient statistic and prove that any unbiased estimator based solely on such a statistic is the unique UMVUE.

1950s • Cramér–Rao Lower Bound Widely Adopted

The Cramér–Rao lower bound becomes a standard benchmark for assessing estimator efficiency, complementing the UMVUE theory by providing a variance floor for unbiased estimators.

1983 • Lehmann's Textbook Codification

Lehmann's Theory of Point Estimation unifies the UMVUE framework, making it a cornerstone of graduate-level statistical inference curricula worldwide.

The fundamental gap these developments address is straightforward: the class of unbiased estimators for a given parameter is typically vast — indeed, infinitely large. Without a principled criterion for choosing among them, a statistician might employ an unbiased estimator that, while centering on the truth on average, exhibits enormous variability from sample to sample. The UMVUE theory asks: is there an estimator in this class whose variance is simultaneously the smallest for every possible value of the parameter? The Lehmann–Scheffé theorem provides an affirmative answer in many commonly encountered models and a constructive method for finding it.

SECTION 2

Core Principles & Definitions

The UMVUE theory rests on a carefully layered set of definitions. Understanding each component — unbiasedness, sufficiency, completeness — and how they interlock is essential before one can appreciate the power of the Lehmann–Scheffé theorem. The following grid summarizes the foundational pillars.

Unbiasedness

An estimator T(X) is unbiased for g(θ) if E_θ[T(X)] = g(θ) for all θ in the parameter space. The estimator hits the target on average, regardless of where the target sits.

Sufficient Statistic

A statistic S(X) is sufficient for θ if the conditional distribution of X given S(X) does not depend on θ. Intuitively, S captures all the information the sample contains about θ — nothing is lost by reducing the data to S.

Complete Statistic

A sufficient statistic S is complete if, for any measurable function h, E_θ[h(S)] = 0 for all θ implies h(S) = 0 almost surely. Completeness prevents 'hidden' unbiased estimators of zero from lurking in the model.

UMVUE

An unbiased estimator T* is UMVUE (Uniformly Minimum Variance Unbiased Estimator) for g(θ) if Var_θ(T*) ≤ Var_θ(T) for every θ and every other unbiased estimator T.

Lehmann–Scheffé Theorem

If S is a complete sufficient statistic and T(X) is any unbiased estimator of g(θ), then h(S) = E[T | S] is the unique UMVUE of g(θ). Equivalently, any function of a complete sufficient statistic that is unbiased is automatically UMVUE.

✦

KEY TAKEAWAY

Think of a complete sufficient statistic as a lossless compression of your entire dataset: it retains every byte of information about the parameter, and the completeness condition ensures that there is exactly one way to decode an unbiased estimate from it. The Lehmann–Scheffé theorem tells you that this uniquely decoded estimate is, by necessity, the best unbiased estimate you could ever hope for — no other unbiased estimator, no matter how cleverly constructed from the raw data, can beat it in variance.

SECTION 3

Visual Explanation — The UMVUE Construction Pipeline

The diagram traces two paths to the UMVUE. The top path directly identifies a complete sufficient statistic and finds an unbiased function of it (the Lehmann–Scheffé route). The bottom path starts with any unbiased estimator, conditions on the sufficient statistic (Rao–Blackwell), and arrives at the same UMVUE when S is complete. The nested bars illustrate the monotone decrease in variance at each stage.

The diagram above captures the essential logic of UMVUE construction. Starting from raw data X₁, …, Xₙ, one first identifies a sufficient statistic S using the factorization theorem, then verifies that S is complete. Once completeness is established, any function h(S) that is unbiased for the target parameter g(θ) is automatically the UMVUE. Alternatively, one may start with an arbitrary unbiased estimator T and apply the Rao–Blackwell technique — conditioning on S — to obtain an improved estimator E[T | S], which itself must equal the UMVUE when S is complete. The nested variance bars at the bottom reinforce a critical insight: each stage of reduction compresses the estimator's variance, and the UMVUE represents the irreducible lower bound within the unbiased class.

SECTION 4

Mathematical Framework

Formal Definitions & Theorems

UNBIASEDNESS CONDITION

E_θ[T(X)] = g(θ) for all θ ∈ Θ

T(X) is a statistic (a function of the sample), g(θ) is the estimand (the parametric function of interest), and Θ is the parameter space. The subscript θ on E indicates that the expectation is computed under the distribution indexed by θ.

SUFFICIENCY — FACTORIZATION THEOREM (FISHER–NEYMAN)

f(x | θ) = h(x) · g_θ(S(x))

A statistic S(X) is sufficient for θ if and only if the joint density (or pmf) factors into a part h(x) that depends only on the data and a part g_θ(S(x)) that depends on the data only through S. This factorization guarantees that S captures all information about θ.

COMPLETENESS

E_θ[h(S)] = 0 ∀θ ∈ Θ ⟹ P_θ(h(S) = 0) = 1 ∀θ

A sufficient statistic S is complete if the only unbiased estimator of zero based on S is the trivial function h ≡ 0. In exponential families with full-rank natural parameter spaces, the natural sufficient statistic is always complete (a result often leveraged in practice).

LEHMANN–SCHEFFÉ THEOREM

If S is complete & sufficient for θ, and h(S) is unbiased for g(θ), then h(S) is the unique UMVUE of g(θ).

Uniqueness follows directly from completeness: if h₁(S) and h₂(S) are both unbiased for g(θ), then E_θ[h₁(S) − h₂(S)] = 0 for all θ, and completeness forces h₁(S) = h₂(S) a.s. The minimum variance property follows from the Rao–Blackwell theorem: for any unbiased T, E[T | S] is a function of S that is also unbiased with Var(E[T | S]) ≤ Var(T), and by completeness it must equal h(S).

Rao–Blackwell Theorem (Supporting Result)

RAO–BLACKWELL INEQUALITY

Var_θ(E[T | S]) ≤ Var_θ(T) for all θ ∈ Θ

This follows from the law of total variance: Var(T) = Var(E[T | S]) + E[Var(T | S)]. Since E[Var(T | S)] ≥ 0, the Rao–Blackwellized estimator E[T | S] always has variance no greater than T. Equality holds if and only if T is already a function of S.

The mathematical elegance of this framework lies in the interplay between these components. The Rao–Blackwell theorem tells us how to improve estimators by conditioning, while completeness tells us when the improvement process has reached its endpoint — the unique UMVUE. Without completeness, one may still obtain a good estimator via Rao–Blackwell, but uniqueness and the guarantee of uniform minimum variance are lost.

SECTION 5

Complete Sufficient Statistics in Exponential Families

In practice, the most common setting in which the Lehmann–Scheffé theorem is applied involves exponential family distributions. A distribution belongs to a k-parameter exponential family if its density can be written as f(x | θ) = c(θ) h(x) exp(∑ η_j(θ) T_j(x)). A fundamental result states that if the natural parameter space contains an open set in ℝ^k, then the natural sufficient statistic T = (T₁, …, T_k) is complete. This single fact covers the Normal, Poisson, Exponential, Gamma, Binomial, and many other standard models.

This reference chart displays six important models. The first five possess complete sufficient statistics, enabling direct UMVUE construction. The curved exponential family example (N(θ, θ²)) illustrates a case where the natural sufficient statistic exists but is not complete because the parameter space is a one-dimensional curve in ℝ², not an open set.

The distinction between full-rank and curved exponential families is critical for determining when the Lehmann–Scheffé machinery applies. In a full-rank family, the dimension of the parameter space matches the dimension of the natural statistic, and the natural parameter space contains an open set. Under these conditions, the natural sufficient statistic is complete, and UMVUEs exist for a wide class of estimands. In curved families — such as N(θ, θ²), where the mean and variance are linked by a single parameter — the natural statistic is two-dimensional but the parameter space is a one-dimensional curve. Completeness fails, and the Lehmann–Scheffé theorem cannot be directly invoked.

SECTION 6

Worked Example — Finding the UMVUE for the Poisson Mean

Suppose X₁, X₂, …, Xₙ are independent and identically distributed as Poisson(λ), where λ > 0 is unknown. We wish to find the UMVUE of g(λ) = λ.

UMVUE of λ in the Poisson Model

1
Step 1 — Write the Joint PMF and Identify a Sufficient Statistic

The joint pmf of the sample is f(x₁, …, xₙ | λ) = ∏ (e^−λ λ^xᵢ / xᵢ!) = e^−nλ · λ^∑xᵢ / ∏(xᵢ!). By the factorization theorem, the joint pmf factors into g_λ(S) = e^−nλ · λ^S (depending on the data through S = ∑Xᵢ) and h(x) = 1/∏(xᵢ!) (depending only on the data).

S = ∑Xᵢ is sufficient for λ. Moreover, S ~ Poisson(nλ).
2
Step 2 — Verify Completeness

The Poisson(λ) family is a one-parameter exponential family with natural parameter η = log λ and natural sufficient statistic T = ∑Xᵢ. The natural parameter space is {η ∈ ℝ : η = log λ, λ > 0} = ℝ, which is an open set in ℝ¹. By the completeness theorem for full-rank exponential families, S = ∑Xᵢ is a complete sufficient statistic for λ.

S = ∑Xᵢ is complete and sufficient for λ.
3
Step 3 — Find an Unbiased Function of S

We need a function h(S) such that E[h(S)] = λ for all λ > 0. Since E[S] = nλ, we have E[S/n] = λ. Thus h(S) = S/n = X̄ is unbiased for λ.

h(S) = S/n = X̄ is unbiased for λ.
4
Step 4 — Apply the Lehmann–Scheffé Theorem

Since S is complete and sufficient, and h(S) = X̄ is an unbiased estimator of λ that depends on the data only through S, the Lehmann–Scheffé theorem immediately implies that X̄ is the unique UMVUE of λ. No other unbiased estimator — whether based on S or on the full sample — can have smaller variance at any value of λ.

X̄ is the unique UMVUE of λ, with Var(X̄) = λ/n.
5
Step 5 — Verify Against the Cramér–Rao Lower Bound

As a sanity check, the Fisher information for a single Poisson(λ) observation is I(λ) = 1/λ. The CRLB for unbiased estimators of λ based on n observations is 1/(nI(λ)) = λ/n. Since Var(X̄) = λ/n, the UMVUE attains the CRLB — confirming that it is efficient in the Cramér–Rao sense as well.

Var(X̄) = λ/n = CRLB, so X̄ is an efficient estimator.

💡

Extension: UMVUE of e⁻λ

Consider estimating g(λ) = e^−λ = P(Xᵢ = 0). Start with the unbiased estimator T = 𝟙(X₁ = 0). Rao–Blackwellize by computing E[T | S] = P(X₁ = 0 | ∑Xᵢ = s) = C(n−1, s)/C(n, s) × ((n−1)/n)⁰ = ((n−1)/n)^s. Since S is complete sufficient, h(S) = ((n−1)/n)^S is the UMVUE of e^−λ. This example illustrates the Rao–Blackwell + Lehmann–Scheffé approach for non-linear functions of the parameter.

SECTION 7

Strengths, Limitations, and Comparisons

The UMVUE framework, anchored by the Lehmann–Scheffé theorem, is a powerful theoretical tool — but like any optimality criterion, it operates under specific assumptions and has recognized limitations. A balanced understanding of where the theory shines and where it falls short is essential for any practicing statistician.

Aspect	Strength	Limitation
Optimality	Provides the unique estimator with the smallest variance among all unbiased estimators, uniformly over the entire parameter space.	Restricts attention to the unbiased class; biased estimators (e.g., James–Stein, ridge regression) can dominate UMVUEs in terms of mean squared error.
Constructive	The Lehmann–Scheffé theorem gives a clear recipe: find a complete sufficient statistic, then find any unbiased function of it.	Finding the unbiased function h(S) with E[h(S)] = g(θ) can be algebraically challenging, especially for nonlinear g.
Uniqueness	Completeness guarantees uniqueness: there is at most one unbiased estimator in the class of functions of S, eliminating ambiguity.	When completeness fails (e.g., curved families, non-identifiable models), the theorem provides no guidance, and a UMVUE may not exist.
Applicability	Covers all full-rank exponential families and several non-exponential models (e.g., Uniform(0, θ)), which account for the majority of classical parametric models.	Non-parametric and semiparametric settings generally lack complete sufficient statistics, limiting the scope of the theory.
Risk criterion	Variance is a natural and interpretable measure of precision for unbiased estimators.	Mean squared error (MSE = Bias² + Variance) is often a more appropriate criterion; insisting on zero bias can inflict an MSE penalty. The Hodges estimator exemplifies superefficiency violations.

Strengths and limitations of the UMVUE / Lehmann–Scheffé framework

⚖️

PERSPECTIVE

The UMVUE represents the best you can do if you commit to playing by the rules of unbiasedness. However, modern statistical theory has shown that relaxing the unbiasedness constraint — accepting a small bias in exchange for a large variance reduction — can yield estimators with smaller overall mean squared error. The UMVUE framework remains indispensable as a theoretical benchmark and in settings where unbiasedness is a regulatory or scientific requirement (e.g., clinical trials, survey estimation).

SECTION 8

Connection to Advanced Theory

The UMVUE theory sits at the beginning of a rich chain of ideas in decision-theoretic statistics. Understanding where it connects to more advanced concepts provides context for both its enduring importance and its limitations.

UMVUE Framework	Advanced Extension
Unbiasedness as the optimality class constraint	Decision theory uses general risk functions R(θ, δ) = E[L(θ, δ(X))]; unbiasedness is one special constraint, but admissibility, minimaxity, and Bayes optimality offer alternative frameworks.
Cramér–Rao lower bound as variance floor	The Hájek–Le Cam local asymptotic minimax theory generalizes efficiency bounds to semiparametric and non-regular models via the asymptotic information bound.
Complete sufficient statistic	Ancillarity, Basu's theorem, and the theory of minimal sufficiency refine the interplay between sufficient and ancillary statistics, especially when completeness fails.
UMVUE achieves minimum variance uniformly	Stein's paradox (1956) shows that in dimension ≥ 3, the sample mean is inadmissible under MSE loss — the James–Stein estimator dominates it despite being biased. This challenges the UMVUE's status as 'optimal'.
Lehmann–Scheffé for parametric models	In nonparametric estimation, U-statistics and V-statistics provide UMVUE-like constructions for symmetric functions of the population distribution.

UMVUE theory vs. advanced extensions

Looking ahead, a course in decision theory will formalize the notion of admissibility: an estimator is admissible if no other estimator has smaller risk for every θ. While every UMVUE is admissible within the unbiased class, it need not be admissible in the broader class of all estimators. Stein's phenomenon is perhaps the most dramatic illustration: when estimating a multivariate normal mean vector of dimension three or higher under squared error loss, the sample mean (the UMVUE) is inadmissible. This profound result reshaped the field and motivated the development of shrinkage estimation and empirical Bayes methods.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

Explain why completeness of a sufficient statistic S is necessary for the Lehmann–Scheffé theorem. Specifically, what could go wrong if S is sufficient but not complete?

Show answer

Without completeness, there may exist a non-trivial function h(S) such that E_θ[h(S)] = 0 for all θ, meaning h(S) is an unbiased estimator of zero with non-zero variance. If T*(S) is an unbiased estimator of g(θ), then T*(S) + h(S) is also an unbiased estimator of g(θ) based on S, and these two estimators are distinct. Consequently, we lose uniqueness: there are multiple unbiased functions of S, and there is no guarantee that any one of them uniformly minimizes variance. Completeness eliminates these 'ghost' estimators of zero, ensuring that the unbiased function of S is unique — and therefore automatically the UMVUE by the Rao–Blackwell argument.

Question 2: BASIC CALCULATION

Let X₁, …, Xₙ be iid Bernoulli(p), 0 < p < 1. Show that X̄ = (1/n)∑Xᵢ is the UMVUE of p, and compute its variance.

Show answer

The Bernoulli distribution is a one-parameter exponential family with natural statistic S = ∑Xᵢ ~ Binomial(n, p) and natural parameter η = log(p/(1−p)) ∈ ℝ. Since the natural parameter space is ℝ (an open set in ℝ¹), S is a complete sufficient statistic. The estimator X̄ = S/n satisfies E[X̄] = p for all p ∈ (0, 1), so it is unbiased. By the Lehmann–Scheffé theorem, X̄ is the unique UMVUE of p. Its variance is Var(X̄) = Var(S)/n² = np(1−p)/n² = p(1−p)/n. This also matches the Cramér–Rao lower bound since I(p) = 1/[p(1−p)] and the CRLB equals 1/(nI(p)) = p(1−p)/n.

Question 3: INTERMEDIATE

Let X₁, …, Xₙ be iid Uniform(0, θ), θ > 0. It is known that X₍ₙ₎ = max(X₁, …, Xₙ) is a complete sufficient statistic for θ. Find the UMVUE of θ and verify that it is unbiased.

Show answer

The density of X₍ₙ₎ is f(m | θ) = n mⁿ⁻¹ / θⁿ for 0 < m < θ. Therefore E[X₍ₙ₎] = ∫₀^θ m · n mⁿ⁻¹ / θⁿ dm = (n/(n+1)) θ. Since E[X₍ₙ₎] = nθ/(n+1) ≠ θ, the order statistic is biased. To correct, define h(X₍ₙ₎) = ((n+1)/n) X₍ₙ₎. Then E[h(X₍ₙ₎)] = ((n+1)/n) · (n/(n+1)) θ = θ, confirming unbiasedness. Since X₍ₙ₎ is complete sufficient and h(X₍ₙ₎) is unbiased, the Lehmann–Scheffé theorem guarantees that ((n+1)/n) X₍ₙ₎ is the unique UMVUE of θ.

Question 4: APPLIED

In a quality control setting, items are inspected and the number of defects per unit follows a Poisson(λ) distribution. A random sample of n = 25 units yields ∑Xᵢ = 40 total defects. Use the UMVUE to estimate (a) the average defect rate λ and (b) the probability of a defect-free unit, P(X = 0) = e^−λ.

Show answer

(a) The UMVUE of λ in the Poisson model is X̄ = S/n = 40/25 = 1.6 defects per unit. (b) As derived in the worked example extension, the UMVUE of e^−λ is ((n−1)/n)^S = (24/25)⁴⁰ = 0.96⁴⁰ ≈ 0.1986. For comparison, the plug-in MLE estimate is e^−1.6 ≈ 0.2019. The UMVUE is slightly smaller because ((n−1)/n)^S corrects for the finite sample size, yielding an unbiased estimate rather than the MLE's asymptotically justified approximation.

Question 5: CRITICAL THINKING

Consider X₁, …, Xₙ iid N(θ, θ²) where both the mean and variance are determined by the single parameter θ > 0. The natural sufficient statistic is (∑Xᵢ, ∑Xᵢ²), which is two-dimensional, but the parameter space is the one-dimensional curve {(θ, θ²) : θ > 0} in ℝ². Explain why the Lehmann–Scheffé theorem cannot be directly applied here, and discuss what alternative approaches one might consider for finding an optimal unbiased estimator of θ.

Show answer

This model is a curved exponential family: the natural parameter space is the one-dimensional curve η = (θ/θ², −1/(2θ²)) = (1/θ, −1/(2θ²)) for θ > 0, which is a one-dimensional manifold in ℝ² — it does not contain an open set. The completeness theorem for exponential families requires the natural parameter space to contain an open set in ℝ^k, so the two-dimensional sufficient statistic (∑Xᵢ, ∑Xᵢ²) is not complete. Without completeness, the Lehmann–Scheffé theorem fails: there may be multiple unbiased functions of the sufficient statistic, none of which is guaranteed to have uniformly minimum variance. In fact, one can construct non-trivial unbiased estimators of zero based on (∑Xᵢ, ∑Xᵢ²) in this model. Alternative approaches include: (1) direct computation of variance for specific unbiased estimators and comparing them case-by-case, (2) checking whether any unbiased estimator attains the Cramér–Rao lower bound, (3) using the theory of U-statistics if the estimand is a symmetric function, or (4) abandoning unbiasedness and pursuing admissible estimators under MSE via decision-theoretic or Bayesian methods.

SUMMARY

Lesson Summary

Summary

This lesson developed the theory of unbiased estimation and the construction of Uniformly Minimum Variance Unbiased Estimators (UMVUEs). An estimator is unbiased if its expectation equals the target parameter for every possible parameter value. The Rao–Blackwell theorem shows that conditioning any unbiased estimator on a sufficient statistic never increases variance. The Lehmann–Scheffé theorem provides the definitive result: if a complete sufficient statistic S exists and one can find an unbiased function h(S), then h(S) is the unique UMVUE.

In practice, full-rank exponential families (Normal, Poisson, Binomial, Exponential, Gamma, and others) always possess complete sufficient statistics, making the Lehmann–Scheffé machinery directly applicable. Curved exponential families and models with restricted parameter spaces may lack completeness, in which case UMVUEs may not exist. The framework remains a cornerstone of statistical inference theory, providing both a benchmark for evaluating estimators and a bridge to advanced topics including admissibility, Stein's paradox, and decision-theoretic optimality.