Why is the observed maximum not automatically the best unbiased estimate?

It tends to miss the unobserved upper portion of the numbering range.

Why does sample size matter?

A larger random sample is more likely to include numbers near the true maximum.

Does the classic formula apply unchanged to sampling with replacement?

No. That changes the sampling distribution and requires the corresponding analysis.

AI Glossary / Statistics and probability / Intermediate

German tank problem

Definition

The German tank problem is a statistical estimation problem in which an unknown population size is inferred from a sample of consecutive serial numbers. In its standard form, items are numbered 1 through N, and k distinct items are sampled uniformly without replacement. The observed maximum contains information about N. The name comes from wartime production estimation, but the mathematical lesson concerns inference from numbered populations under explicit sampling assumptions.

What it means

If the largest observed serial number is 42, there must be at least 42 items under the standard numbering model. However, observing no larger number does not establish that 42 is the final number. A sample usually misses some items above its observed maximum.

The problem asks how to account for that unobserved tail. The answer depends on both the largest observed number and the number of distinct sampled items. More observations make the observed maximum more informative.

The model is deliberately idealized. Real serial-number systems may contain gaps, separate factory ranges, reused identifiers, or nonuniform sampling. Those features can make the largest serial number a poor proxy for the number of existing items.

How it works

Let M be the sample maximum. Under simple random sampling without replacement from 1, …, N, E[M] = k(N + 1)/(k + 1). Rearranging gives the unbiased estimator N_hat = M(k + 1)/k − 1, equivalently M + M/k − 1.

For the standard model, the maximum’s probability is P(M = m) = choose(m − 1, k − 1)/choose(N, k). This distribution supports exact calculations and interval procedures. The maximum-likelihood estimate is instead the observed maximum itself; unbiasedness and maximum likelihood are different criteria.

A Bayesian analysis combines the sampling model with a declared prior and can produce a different estimate or interval. Whichever approach is used, retain uncertainty and verify the numbering and selection assumptions before interpreting the result as a population count.

Three practical examples

Inventory research: estimate the size of a genuinely consecutively numbered item population from an appropriate random sample when a complete register is unavailable.
Statistical education: demonstrate why a sample maximum tends to underestimate a population endpoint and how bias correction works.
Data auditing: use the model as a diagnostic thought experiment for sequential identifiers, while checking whether gaps, deletions, and selective access invalidate a direct population-size interpretation.

Think of it this way

Think of drawing numbered tickets from a bag whose final ticket number is unknown. The highest ticket you draw gives a lower bound, but usually leaves room for unseen higher tickets. The correction estimates that unseen portion from how much of the numbering range your sample has revealed.

Worked example

Suppose four distinct serial numbers, 8, 19, 31, and 42, are sampled under the standard assumptions. Here k = 4 and m = 42. The unbiased estimate is 42 + 42/4 − 1 = 51.5. The population count is an integer, but an unbiased point estimator can take a fractional value; rounding changes its exact statistical property.

To verify the bias-correction principle separately, consider a tiny population numbered 1 through 6 and samples of two distinct items. There are choose(6, 2) = 15 equally likely samples. The maximum equals 2, 3, 4, 5, or 6 in respectively 1, 2, 3, 4, or 5 samples.

Its average is (2 × 1 + 3 × 2 + 4 × 3 + 5 × 4 + 6 × 5)/15 = 70/15. Applying the correction gives (3/2)(70/15) − 1 = 6, exactly the true population size. Independent enumeration confirms the result. Unbiasedness holds across repeated samples; it does not guarantee that the estimate from any one sample equals the truth.

Illustrative case study

Fictional case: Orchard Devices investigates whether sampled warranty units can reveal production volume. An initial calculation applies the standard serial-number estimator immediately.

The analyst first checks the identifier scheme and discovers that different plants use separate ranges and that warranty returns overrepresent older units. The team therefore does not present the simple estimate as a defensible production count. It obtains plant-level records and a clearer selection model. The exercise illustrates that recognizing a familiar formula is less important than establishing that the population and sample match its assumptions.

Common misconceptions and mistakes

Treating arbitrary product codes as consecutive numbers starting at one.
Counting repeated sightings of the same item as distinct without-replacement observations.
Assuming captured, returned, or publicly visible items form a uniform random sample.
Confusing the largest identifier with the number of active items when deletions or unused numbers exist.
Calling a point estimate a hard upper bound.
Assuming an unbiased estimator is exact for every sample or must always be an integer.

Business use

The problem offers a useful example of extracting information from metadata, but practical application requires careful data provenance. A business should establish the numbering convention, population definition, duplicate handling, and sampling mechanism before using the estimator.

Report the sample maximum, sample size, method, assumptions, and uncertainty rather than only a final number. If the model is implausible, investigate a suitable extension or obtain better records. An apparently precise estimate from an inappropriate serial-number model can be less useful than an honest statement that the available identifiers do not support the desired count.

Check your understanding

Answer: 1.

What lower bound follows from observing serial number 42 under the model?

Answer: N must be at least 42. 2.

If every item is observed, what does the unbiased estimator return?

Answer: With k = M = N, it returns N. 3.

Does a serial-number gap automatically represent an unseen produced item?

Answer: No; the numbering process must be investigated.

Related terms

Order statistic
Discrete Uniform Distribution
Unbiased estimator
Confidence interval
Sample maximum
maximum likelihood
sampling without replacement
population size

Learn next

Enumerate small numbered populations to compare the sample maximum and bias-corrected estimator. Then study interval estimation and how unknown starting numbers, multiple numbering systems, or selective sampling change the problem.

Sources

Clark, Gonye, and Miller, “Lessons from the German Tank Problem”: https://arxiv.org/pdf/2101.08162 — primary derivation, assumptions, and extensions read.

Bilkent probability and statistics notes, point estimators: https://eray.bilkent.edu.tr/chapters/ch06-point-estimators.html — estimation context checked. The serial numbers, enumeration exercise, and fictional case are original.

Previous: Geospatial metadata Next: Gestalt pattern matching