Is every unbiased estimator asymptotically unbiased?

If it is unbiased at every applicable sample size, its bias is already zero, so yes.

Is every consistent estimator asymptotically unbiased?

Not without additional conditions controlling expectations; convergence in probability alone is insufficient.

How large must the sample be?

There is no universal number. It depends on the estimator, distribution, desired precision and assumptions.

AI Glossary / Governance, ethics and safety / Advanced

Asymptotically unbiased estimator

Definition

An asymptotically unbiased estimator is a sequence of statistical estimation rules whose bias approaches zero as the sample size grows without bound, under stated assumptions. If T_n estimates a parameter theta, the condition is E[T_n] - theta approaches zero as n increases. The expectation averages over possible samples; it does not describe the error of one observed estimate.

What it means

An estimator can systematically overestimate or underestimate a quantity in small samples while that average discrepancy becomes negligible with increasing sample size. “Asymptotically” describes the limiting behavior, not a promise that a particular available dataset is already large enough.
This differs from exact unbiasedness, where the expectation equals the parameter at every applicable sample size. It also differs from consistency, which concerns the estimate becoming concentrated near the true parameter. A vanishing average error does not by itself ensure that the estimator's variability shrinks.
Here, bias has its statistical meaning. It is not a claim about social fairness, discrimination or whether the underlying data collection is appropriate.

How it works

To assess the property, first specify the quantity being estimated and the assumptions about the data. Then derive or analyze the estimator's expectation as a function of sample size. Subtracting the true parameter gives the bias; its limit determines asymptotic unbiasedness.
For independent, identically distributed observations with finite variance sigma squared, consider the average squared deviation from the sample mean using divisor n. Its expectation is ((n - 1) / n) times sigma squared. Its bias is therefore minus sigma squared / n, which tends to zero.
The identity follows by writing the sum of squared deviations from the sample mean as the sum of squared deviations from the population mean minus n times the squared error of the sample mean. The two expected terms are n times sigma squared and sigma squared, respectively. This argument depends on the sampling assumptions and finite moments.

Three practical examples

Process monitoring: An analyst estimates measurement variance using divisor n. The finite-sample downward bias matters more with a short batch than with a large independent sample.
Simulation studies: A team generates many datasets from a known process and examines whether an estimator's average error decreases as sample size grows. Finite simulations provide evidence, not proof of a limiting claim.
Forecasting research: An analyst considers large-sample properties of a parameter estimate but separately evaluates finite-sample forecasts. A useful asymptotic result does not guarantee accurate predictions for the available history.

Think of it this way

Imagine repeated sets of measurements made with a procedure whose average offset becomes smaller as each set contains more observations. The offset may vanish in the limit while individual sets still vary. The analogy concerns the procedure's average behavior, not an assurance that every new observation improves the current estimate.

Worked example

Suppose a population has variance 9 squared units, and observations are independent and identically distributed. Estimate its variance with the sum of squared deviations from the sample mean divided by n.
For n = 4, the estimator's expectation is (3 / 4) x 9 = 6.75. Its bias is 6.75 - 9 = -2.25. For n = 20, the expectation is (19 / 20) x 9 = 8.55 and the bias is -0.45. For n = 100, the expectation is 8.91 and the bias is -0.09.
These are averages over hypothetical repeated samples, not the values every sample produces. For example, the observed values 2, 4, 6 and 8 have mean 5 and squared deviations totaling 20. The divisor-n estimate is 20 / 4 = 5.
Using divisor n - 1 instead gives 20 / 3, approximately 6.67, and yields an exactly unbiased variance estimator under these assumptions. Neither observed result must equal the population variance. Bias concerns the estimation rule across samples.

Illustrative case study

Fictional company Hazel Instruments compares variability across production batches. A report describes its divisor-n variance calculation as asymptotically unbiased. A manager interprets this as meaning the reported variance is effectively correct even for batches containing only four measurements.
The analyst explains the expectation formula and distinguishes systematic bias from sample-to-sample variation. The team reports the number of observations, uses the appropriate variance convention for its purpose and checks whether measurements are sufficiently independent for the analysis.
It also investigates process changes instead of assuming that collecting more measurements from a changing process fixes the problem. This fictional case illustrates why an asymptotic property needs a finite-sample interpretation and explicit assumptions.

Common misconceptions and mistakes

“Asymptotically unbiased means unbiased now.” The bias can be material at the actual sample size.
“It guarantees consistency.” Not alone. An estimator that always uses only the first observation can be unbiased for the population mean but retain the same variability as more observations arrive.
“More records fix any bias.” Selection problems, dependence or an incorrect model can invalidate the argument.
“The estimate must get closer with every additional observation.” A limiting statement does not require monotonic improvement along a realized sample sequence.
“Statistical unbiasedness proves fairness.” The concepts address different questions.

Business use

Treat asymptotic unbiasedness as one property of an estimation method. Ask how large the finite-sample bias and variability may be under conditions resembling the actual problem. Mean squared error, interval performance and operational consequences can matter more than a limiting label alone.
Check whether the sampling assumptions fit the records being used. Repeated measurements from the same source may not provide the independent information assumed by a simple derivation. Explain the distinction between a theoretical expectation and the number displayed in a business report.

Check your understanding

What tends to zero in the definition?

Answer: The estimator's bias, not necessarily its variance or each observed error.

Why does the divisor-n variance underestimate on average?

Answer: Estimating the mean reduces the expected squared-deviation total to (n - 1) times the population variance.

What should accompany the limiting claim?

Answer: Assumptions and evidence relevant to the available sample size.

Related terms

Learn next

Compare bias, variance and consistency using simple estimators. Then examine why mean squared error combines variance with squared bias and how finite-sample performance can differ from limiting behavior.

Sources

Glossary of Forecasting Terms, asymptotically unbiased estimator and unbiasedness. https://robjhyndman.com/mwh3/FG4.pdf MATH/STAT 355: Statistical Theory, Consistency, distinctions between expectation and convergence in probability. https://mac-stat.github.io/MathStat355Notes/consistency.html The variance derivation, numerical illustration and Hazel Instruments case are original teaching explanations under the stated assumptions.

Previous: Asymptotic normality of the maximum likelihood estimator Next: Asynchronous Advantage Actor-Critic