Find your next learning step
A pointwise limit can make an estimator look excellent at one parameter value while hiding poor performance a short distance away. The local asymptotic minimax theorem changes the question. It asks for the largest risk across a shrinking neighbourhood, translates the experiment to a Gaussian shift, and supplies a lower bound that no estimator sequence can beat uniformly under the theorem’s conditions.
Open the chapter contents · Probability root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the decision, or read from Chapter 1. The chapters remain open; these grouped routes are a navigable index.
Route 1 · Define the decision
Route 2 · Work the ledgers
Route 3 · Transfer the method
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE DECISION
1. Define the local risk question
Fix a reference parameter θ₀ and examine alternatives θₙ=θ₀+h/rₙ, usually with rₙ=√n. For an estimator Tₙ of a differentiable target q(θ), scale the error as rₙ[Tₙ−q(θₙ)] and evaluate it with a declared loss ℓ.
The key object is not the risk only at h=0. It is the supremum of risk over bounded local directions, followed by the appropriate large-sample and expanding-neighbourhood limits. That uniformity blocks a procedure from claiming victory by being unusually good at a single point.
CHAPTER 2 OF 12 · DEFINE THE DECISION
2. State the Gaussian lower bound
Under local asymptotic normality, differentiability of q and the theorem’s regularity conditions, every estimator sequence has local worst-case risk at least the risk of the efficient estimator in the limiting Gaussian shift experiment. The efficient error has covariance q̇(θ₀) I(θ₀)⁻¹ q̇(θ₀)ᵀ.
A standard formulation takes a lower limit in n, a supremum over local parameters and then an expanding bound M. The exact order of limits and allowed losses belongs in the statement; silently replacing uniform local risk by pointwise risk changes the theorem.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. Work a normal-mean risk ledger
Let X₁,…,Xₙ be independent N(θ,4). The sample mean has √n(X̄−θ)∼N(0,4), so its limiting squared-error risk is 4. Its limiting absolute-error risk is E|Z|=2√(2/π), about 1.596, for Z∼N(0,4).
| Quantity | Squared loss | Absolute loss | Meaning |
|---|---|---|---|
| Efficient limit | Z∼N(0,4) | Z∼N(0,4) | Gaussian shift error |
| Benchmark risk | E(Z²)=4 | E|Z|≈1.596 | Loss changes the number |
| Local direction h | Same centred error | Same centred error | Translation handled by estimator |
| Claim justified | Uniform local benchmark | Uniform local benchmark | Not a finite-sample guarantee |
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. Understand a bowl-shaped loss
A bowl-shaped loss is nonnegative, symmetric about zero and has convex—or, in scalar language, interval-shaped—sublevel sets. Squared error, absolute error and the indicator of leaving a symmetric error interval are standard examples.
The shape matters because adding independent centred Gaussian dispersion cannot improve every such loss. Asymmetric costs, one-sided constraints and nonconvex decision regions may require a different minimax statement rather than a casual reuse of the symmetric bound.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Change the estimator and inspect uniform risk
Now use Tₙ=X̄+c/√n. Its scaled error is Z+c. Under squared loss the risk becomes 4+c²; for c=1 it is 5. A shrinkage rule may reduce risk exactly at θ₀, yet its local error law changes with h and its worst nearby risk can exceed the Gaussian benchmark.
| Procedure | Risk at θ₀ | Nearby local behaviour | Lesson |
|---|---|---|---|
| Sample mean | 4 | Stable at 4 | Attains squared-loss bound |
| Add 1/√n | 5 | Stable at 5 | Visible local bias |
| Add 1/n | Tends to 4 | Perturbation vanishes | Same first-order risk |
| Point shrinkage | Can fall below 4 | Risk spike can move with h | Pointwise gain is not uniform |
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Connect the bound to LAN
The local asymptotic normality guide explains how log-likelihood ratios converge to a Gaussian shift experiment. The minimax theorem solves the corresponding decision problem and transfers the risk benchmark back to the original sequence of experiments.
LAN is a local approximation, not a blanket declaration that every estimator is Gaussian. The estimator may be irregular, the information matrix singular or the target nondifferentiable, in which case another rate or limit experiment may be needed.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Compare convolution and minimax conclusions
The convolution-theorem guide describes the limit law of a regular estimator as efficient Gaussian error plus independent extra noise. The local minimax theorem is broader in one direction: it bounds local worst-case risk for arbitrary estimator sequences.
The two conclusions answer different questions. Convolution identifies the structure of regular limits; minimax rules out a uniformly better local risk. Neither says that the finite-sample sample mean dominates every competitor at every θ.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Use experiment comparison without overclaiming
The Le Cam-deficiency guide tracks how decision risks transfer between nearby experiments. The Blackwell-ordering guide treats exact comparison by garbling.
A minimax lower bound is not itself an ordering of experiments. It is a statement about the best attainable worst-case risk for a specified local problem, parameter set, action space and loss.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Know the limits and common misconceptions
The theorem is asymptotic, local and conditional on the declared experiment. It does not guarantee robustness to misspecification, good small-sample tails or an attainable bound when no suitable efficient estimator exists.
“Minimax” does not mean pessimistic guessing over the entire parameter space. Here the neighbourhood shrinks in the original parameter scale while expanding in the local coordinate. Pointwise convergence cannot replace that uniform construction.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Use an evidence and source routine
Record θ₀, the local rate, the LAN expansion, information matrix, target derivative, loss, action space, neighbourhood sequence, order of limits, Gaussian benchmark, candidate estimator risk and any boundary or nuisance complications.
Wellner’s University of Washington overview states the local asymptotic minimax and convolution theorems together. Pollard’s Yale notes develop asymptotic minimax bounds and bowl-shaped losses. University of Washington lecture notes place the bound inside LAN decision theory.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Why expand M after taking n large?
It first protects against every bounded local direction, then covers progressively larger parts of the Gaussian shift experiment without turning the original neighbourhood into a fixed global set.
Can the bound be smaller for absolute loss?
The numerical bound changes with the loss, as the Gaussian ledger shows; the efficient covariance benchmark is the same.
Does equality prove an estimator is best in finite samples?
No. It establishes first-order local asymptotic minimaxity under the stated loss and assumptions.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue with a changed-loss transfer check
Repeat the normal-mean ledger with variance 9. Compute the Gaussian squared-error risk and absolute-error risk. Then add 2/√n to the sample mean and compute the new squared-error risk. Explain why adding 2/n instead disappears at the √n scale.
Continue through the Probability root, the Gold Standard of Education and the How X Works hub.
Top · Contents · Previous · How X Works