Hypergeometric Distribution (Extended)

The variance of X ~ Hypergeometric(N, K, n) is: Var(X) = n K/N (1 - K/N) ( (N - n) / (N - 1) ) = n p (1 - p) ( (N - n) / (N - 1) ). Notice the first part np(1-p) is the exact Binomial variance! The extra multiplier (N - n) / (N - 1) is the Finite Population Correction (FPC). Because n ≥ 1, (N - n) / (N - 1) < 1 whenever n > 1. Sampling without replacement exhausts the population, creating negative covariance between draws and reducing overall volatility.

In political survey sampling, if a national poll samples n=1,200 voters from a small municipality of N=5,000 citizens, the FPC factor is √((5000-1200)/4999) ≈ 0.872, shrinking the margin of error by nearly 13% compared to assuming an infinite population.

The Finite Population Correction was derived by English statistician Arthur Bowley in 1906, revolutionizing socioeconomic survey sampling by proving small samples from finite populations are more precise than previously assumed.

The FPC Shrinkage Rule: 'Binomial variance times (N - n)/(N - 1)'. When n = N, (N-N)/(N-1) = 0 (zero variance!). When N → ∞, (N-n)/(N-1) → 1 (Binomial variance).

Let us calculate step-by-step:

1. Identify parameters: Total bulbs N = 20, Defective bulbs K = 5, Non-defective bulbs N - K = 15. Sample size n = 4, Desired defectives x = 2.
2. Pick defectives: Number of ways to choose 2 defectives from 5: C(5, 2) = (5 × 4) / (2 × 1) = 10.
3. Pick non-defectives: Number of ways to choose remaining 2 non-defectives from 15: C(15, 2) = (15 × 14) / (2 × 1) = 105.
4. Favorable outcomes: 10 × 105 = 1,050.
5. Total outcomes: Number of ways to choose any 4 bulbs from 20: C(20, 4) = (20 × 19 × 18 × 17) / (4 × 3 × 2 × 1) = 4,845.
6. Final probability: P(X = 2) = (1,050) / (4,845) = 70/323 ≈ 0.2167 (21.67%).

In lottery games (such as Powerball or Texas Two Step), calculating the exact odds of matching 3 out of 5 white balls uses this exact calculation where N=69 white balls, K=5 winning numbers, n=5 picks on ticket.

This type of urn problem was first formally solved by Dutch mathematician Christiaan Huygens in 1657 in De Ratiociniis in Ludo Aleae, establishing urn modeling as the universal archetype for probability without replacement.

The Urn Calculation: 'Ways to pick good × Ways to pick bad / Total ways to pick'. (C(K, x) C(N-K, n-x)) / C(N, n).

Consider an urn with 2 Red balls and 2 Black balls:

• With Replacement: Draw 1 Red ball, put it back. The probability of Red on draw 2 is still 2/4 = 0.50. The draws are independent. Total successes follows Binomial(n, 0.5).
• Without Replacement: Draw 1 Red ball, set it aside. The urn now holds 1 Red and 2 Black balls. The probability of Red on draw 2 drops to 1/3 = 0.33! Draws are statistically dependent. Total successes follows Hypergeometric.

One physical action—putting the item back or keeping it—completely changes the mathematical governing law.

In software testing and fuzzing, sampling test inputs with replacement (bootstrapping) tests average-case behavior, while sampling without replacement (permutation testing) ensures exhaustive, deduplicated path coverage across critical corner cases.

Jacob Bernoulli spent years studying urn models in Ars Conjectandi, proving that when population N is massive compared to n (n/N → 0), sampling without replacement asymptotically mimics sampling with replacement.

Decision Rule: 'Put it back? Binomial. Keep it out? Hypergeometric'.

Let the population Ω have size |Ω| = N. It is partitioned into two disjoint subsets: S (Successes, |S| = K) and F (Failures, |F| = N - K). An experiment consists of selecting a subset A ⊂ Ω of size |A| = n, where every subset of size n has equal probability of being chosen. The random variable X(A) = |A ∩ S| measures the number of successes in sample A. Our objective is to compute the probability distribution P(X = x) for all permissible values of x.

In jury discrimination court cases (e.g. Castaneda v. Partida in the US Supreme Court), statistical experts model the selection of n=100 grand jurors from a county of population N with K minority residents using Hypergeometric partitioning to establish racial disparity.

English mathematician Thomas Bayes and later Karl Pearson analyzed partitioned urn models to formalize inductive inference and statistical hypothesis testing on social data.

The Partition Picture: 'K in group 1, N-K in group 2. Draw n total: x comes from group 1, n-x comes from group 2'.

To construct a sample containing exactly x successes and (n-x) failures, nature must complete two independent sub-tasks:

1. Sub-task 1: Choose x items from the K available successes. The number of distinct ways to do this is C(K, x).
2. Sub-task 2: Choose the remaining (n - x) items from the (N - K) available failures. The number of distinct ways is C(N-K, n-x).

By the Fundamental Rule of Counting (Multiplication Principle), any choice from Sub-task 1 can be paired with any choice from Sub-task 2. Therefore, the total count of favorable samples is: C(K, x) × C(N-K, n-x).

In poker hand probabilities, calculating the number of Full Houses (3 of one rank, 2 of another) uses this exact multiplication principle: C(4, 3) ways to choose 3 Aces multiplied by C(4, 2) ways to choose 2 Kings.

Blaise Pascal and Pierre de Fermat established the multiplication principle of independent combinatorial choices in their foundational 1654 letters on gambling odds.

The Twin Combinations: 'Choose successes: C(K, x). Choose failures: C(N-K, n-x). Multiply them together'.

Under uniform random sampling without replacement, every one of the C(N, n) possible samples of size n is equally likely to be selected. Therefore, by Laplace's classical definition of probability: P(X = x) = (Number of favorable outcomes) / (Total number of possible outcomes). From Screen 5, the number of favorable samples is C(K, x) C(N-K, n-x). The total number of ways to pick any n items from the population of N items is C(N, n). Dividing numerator by denominator yields the Hypergeometric PMF: P(X = x) = (C(K, x) C(N-K, n-x)) / (C(N, n)). That is the complete, rigorous derivation!

In clinical genetics, Fisher's Exact Test uses this exact Hypergeometric ratio to evaluate whether gene expression mutations are significantly associated with disease phenotypes in a 2x2 contingency table.

Sir Ronald A. Fisher introduced Fisher's Exact Test in 1935 in his famous 'Lady Tasting Tea' experiment, using the Hypergeometric PMF to prove Muriel Bristol could genuinely taste whether tea or milk was added first.

Top-to-Bottom Balance: Look at the tops of the combinations: K + (N - K) = N. Look at the bottoms: x + (n - x) = n. The tops add to the bottom top; the bottoms add to the bottom bottom!

The formal specification of the Hypergeometric distribution is: P(X = x) = (C(K, x) C(N-K, n-x)) / (C(N, n)). However, combinations C(a, b) are only non-zero when 0 ≤ b ≤ a. This imposes four simultaneous physical inequalities on x:

1. x ≥ 0 (cannot draw negative successes)
2. x ≤ K (cannot draw more successes than exist in population)
3. n - x ≥ 0 ⟹ x ≤ n (cannot draw more successes than sample size)
4. n - x ≤ N - K ⟹ x ≥ n - (N - K) (cannot draw more failures than exist in population)

Combining these yields the compact support constraint: max(0, n - (N - K)) ≤ x ≤ min(n, K).

In election audit recounts (e.g. Risk-Limiting Audits), election officials sample paper ballot batches without replacement, using these exact support bounds to calculate whether a discrepancy could overturn an election outcome.

Vandermonde's Identity (∑_(x) C(K, x) C(N-K, n-x) = C(N, n)), proven by Alexandre-Théophile Vandermonde in 1772, proves that summing the Hypergeometric PMF over this exact support equals identically 1.

Support Sandwich: 'x is trapped between the maximum forced successes and the minimum available successes'. max(0, n+K-N) ≤ x ≤ min(n, K).

Here is the master reference card for the Hypergeometric distribution:

• Parameters: N ∈ N^+ (population size), K ∈ 0, ṡ, N (population successes), n ∈ 1, ṡ, N (sample size).
• Support: x ∈ max(0, n - (N - K)), ṡ, min(n, K) .
• PMF: P(X = x) = (C(K, x) C(N-K, n-x)) / (C(N, n)).
• Mean: E[X] = n K/N = n p.
• Variance: Var(X) = n p (1-p) ((N - n) / (N - 1)).
• Binomial Limit: As N, K → ∞ with K/N → p, Hypergeometric(N, K, n) → Binomial(n, p).

Quality assurance managers use this cheat sheet to design sampling tables: if batch size N is over 10 times the sample size n, they safely switch software calculation routines from slow Hypergeometric factorials to lightning-fast Binomial approximations.

This master synthesis was formalized in the 1930s by Jerzy Neyman in his foundational works on sampling human populations and confidence intervals.

Hypergeometric Profile: 'Big N population, K successes, draw n, count x. Mean is np, Variance is np(1-p) with FPC brake'.

In textbook probability, replacement seems like an arbitrary choice. In applied engineering, it is dictated by physics:

• Destructive Testing: Crash testing automobiles, lighting matchsticks, testing artillery shells, or measuring blood chemical levels destroys or alters the item. The item CANNOT be replaced. Sampling is strictly WITHOUT replacement ⟹ Hypergeometric.
• Non-Destructive Testing: Scanning barcodes on packages, measuring external dimensions with lasers, or querying read-only database records leaves items pristine. Items can be sampled with replacement ⟹ Binomial.

Thus, destructive physical inspection is the natural domain of Hypergeometric probability.

In missile defense manufacturing (Lockheed Martin, Raytheon), testing solid rocket propellant burn rates destroys the motor casing. Reliability certification must use Hypergeometric statistics without replacement to calculate missile arsenal readiness.

The British Royal Navy and Woolwich Arsenal in the 18th century first commissioned mathematicians to calculate powder keg sample failure rates without wasting their limited stockpiles.

Destruction Rule: 'If testing destroys the item, replacement is impossible ⟹ Hypergeometric'. Non-destructive allows Binomial.

Let us examine the mathematical limit as population size N → ∞ while keeping success proportion p = K/N constant:

1. The FPC factor: lim_(N → ∞) (N - n) / (N - 1) = lim_(N → ∞) (1 - n/N) / (1 - 1/N) = 1. Thus, the Hypergeometric variance converges exactly to the Binomial variance np(1-p).
2. The PMF ratio: By expanding factorials, each term (K - i) / (N - j) → K/N = p. The combinatorial formula (C(K, x) C(N-K, n-x)) / (C(N, n)) converges point-wise to C(n, x) p^x (1-p)^(n-x).

The 5% Rule of Thumb: If n/N ≤ 0.05 (the sample is less than 5% of the population), statisticians routinely replace Hypergeometric with Binomial to simplify calculation.

In national political polling (e.g. Gallup, Pew Research), surveying n=1,500 voters from the US voting population of N=160,000,000 gives n/N ≈ 0.000009. Because n/N ≪ 0.05, pollsters use simple Binomial formulas rather than Hypergeometric factorials.

This asymptotic convergence was proven by Siméon Denis Poisson in 1837 as part of his broader investigation into limiting forms of discrete probability.

The 5% Threshold: 'If sample n is under 5% of N, Binomial is your friend'. FPC ≈ 1, dependence vanishes.