Poisson Distribution - College Statistics

SECTION 1

Historical Context & Motivation

The study of probability has always been driven by practical needs—gambling odds, insurance risk, quality control—but certain phenomena resisted the tools available in the early nineteenth century. Specifically, events that occurred rarely and independently within a fixed window of time or space posed a modeling challenge that the binomial distribution handled clumsily when the number of trials was very large and the probability of success very small. The Poisson distribution emerged precisely to fill this gap, offering an elegant closed-form expression for the probability of observing a given count of such events.

1837 • Poisson Publishes His Treatise

Siméon Denis Poisson introduces the distribution in Recherches sur la probabilité des jugements, deriving it as a limiting case of the binomial distribution when n → ∞ and p → 0 while np remains constant.

1898 • Bortkiewicz and the Prussian Horse Kicks

Ladislaus Bortkiewicz demonstrates the practical power of the Poisson model by fitting it to data on Prussian cavalry soldiers killed by horse kicks—an iconic example of rare-event counting that brought the distribution into mainstream statistics.

1940s • Wartime Applications

During World War II, statisticians use the Poisson distribution to model V-1 flying bomb strikes on London, confirming that the hits were randomly distributed across the city rather than deliberately targeted at specific neighborhoods.

Modern Era • Ubiquitous in Science and Industry

Today the Poisson distribution underpins models in queueing theory, telecommunications (packet arrivals), genomics (mutation counts), epidemiology (disease incidence), and machine learning (count regression), establishing it as one of the most widely applied discrete distributions.

The central question the Poisson distribution answers is deceptively simple: If events happen at a known average rate and independently of one another, what is the probability of observing exactly k events in a given interval? Answering this question rigorously requires a distribution that depends on only a single parameter—the average rate—and that assigns non-negligible probability to counts well above and below that average. The sections that follow develop this distribution from first principles, illustrate its shape, and demonstrate its applications.

SECTION 2

Core Principles & Definitions

Before applying the Poisson distribution, one must verify that the underlying random process satisfies a specific set of assumptions. These assumptions define what is known as a Poisson process, a stochastic model in which events occur continuously and independently at a constant average rate. Violating these conditions—for example, by introducing clustering or time-varying rates—invalidates the Poisson model and calls for more general alternatives such as the negative binomial or the non-homogeneous Poisson process.

Events Are Discrete Counts

The random variable X counts the number of occurrences (0, 1, 2, …) in a fixed interval of time, area, volume, or another continuous measure. The support is the set of all non-negative integers.

Constant Average Rate (λ)

The expected number of events per interval, denoted λ (lambda), is constant across all intervals of the same size. This single parameter fully determines the distribution.

Independence of Intervals

The number of events occurring in disjoint intervals are independent random variables. Knowing how many events occurred in one interval gives no information about another.

No Simultaneous Events

In an infinitesimally small sub-interval, the probability of two or more events occurring simultaneously is negligible—formally, it is o(Δt). At most one event can occur in any instant.

Mean Equals Variance

A hallmark property: E[X] = Var(X) = λ. This equality serves as a diagnostic check—if the sample variance substantially exceeds the sample mean, the data may exhibit overdispersion and a Poisson model may be inadequate.

✦

KEY TAKEAWAY

Think of the Poisson distribution as a model for counting raindrops hitting a single square of sidewalk during a steady drizzle. Each raindrop arrives independently of the others, you cannot predict the exact moment of the next drop, and on average the same number hit the square per minute. The rate parameter λ is analogous to the intensity of the rain—increase it, and the distribution shifts rightward, assigning higher probability to larger counts.

SECTION 3

Visual Explanation — Shape of the PMF

The probability mass function (PMF) of the Poisson distribution is defined only at non-negative integers, so it is best visualized as a bar chart rather than a continuous curve. The following diagram overlays three Poisson PMFs with different values of λ, illustrating how the distribution's shape evolves from strongly right-skewed at small λ to nearly symmetric as λ increases.

Poisson PMFs for λ = 2 (cyan), λ = 5 (violet), and λ = 10 (pink). As λ increases, the distribution spreads out and becomes more symmetric, approaching a bell-shaped curve. The mode of each distribution is at or near ⌊λ⌋.

Several features are immediately apparent from the diagram. When λ = 2, the distribution is strongly right-skewed, with the highest probability concentrated at k = 1 and k = 2 and a long tail extending to the right. As λ grows to 5 and then to 10, the peak flattens, the spread increases (recall that the standard deviation is √λ), and the shape becomes progressively more symmetric. This convergence toward symmetry is not merely visual; the Central Limit Theorem guarantees that a Poisson random variable with large λ is approximately normally distributed with mean λ and variance λ.

SECTION 4

Mathematical Framework

The mathematical elegance of the Poisson distribution lies in its derivation as a limit of the binomial distribution and in the simplicity of its single-parameter formulation. We present the probability mass function, derive the key moments, and state the moment-generating function, which is essential for proving reproductive properties.

PROBABILITY MASS FUNCTION

P(X = k) = (e^(−λ) × λᵏ) / k! for k = 0, 1, 2, …

where λ > 0 is the average number of events per interval, k is the observed count, e ≈ 2.71828 is Euler's number, and k! is the factorial of k. The factor $e^{−λ}$ ensures the probabilities sum to 1 via the Taylor expansion of e^λ.

MEAN AND VARIANCE

E[X] = λ Var(X) = λ SD(X) = √λ

The equality of the mean and variance is a defining property of the Poisson distribution. It can be derived from the PMF using the moment-generating function or by direct summation of k × P(X = k).

MOMENT-GENERATING FUNCTION

M_X(t) = exp(λ(eᵗ − 1))

The MGF is useful for proving that the sum of independent Poisson random variables is also Poisson: if X₁ ~ Poisson(λ₁) and X₂ ~ Poisson(λ₂) are independent, then X₁ + X₂ ~ Poisson(λ₁ + λ₂). This is the reproductive property of the Poisson distribution.

BINOMIAL LIMIT DERIVATION

lim (n→∞) C(n,k) × pᵏ × (1−p)^(n−k) = (e^(−λ) × λᵏ) / k! where λ = np

When the number of trials n is large and the probability of success p is small such that np = λ remains moderate, each binomial probability converges to the corresponding Poisson probability. This Poisson limit theorem justifies using the Poisson as an approximation to the binomial when n ≥ 20 and p ≤ 0.05 (or more liberally when np ≤ 10).

ℹ

Skewness & Kurtosis

The skewness of a Poisson(λ) random variable is 1/√λ and the excess kurtosis is 1/λ. Both decrease toward zero as λ → ∞, which is consistent with the distribution's convergence to normality.

SECTION 5

Cumulative Distribution & Probability Tables

In applied problems, we frequently need cumulative probabilities—questions of the form "what is the probability of observing at most k events" or "at least k events." The cumulative distribution function (CDF) is obtained by summing the PMF from 0 to k. There is no simple closed-form expression, but the sum can be evaluated using the regularized incomplete gamma function or computed numerically with statistical software. The table below provides selected Poisson probabilities for λ = 3, which is a commonly encountered rate in textbook problems.

k	P(X = k)	P(X ≤ k)	P(X > k)
0	0.0498	0.0498	0.9502
1	0.1494	0.1991	0.8009
2	0.2240	0.4232	0.5768
3	0.2240	0.6472	0.3528
4	0.1680	0.8153	0.1847
5	0.1008	0.9161	0.0839
6	0.0504	0.9665	0.0335
7	0.0216	0.9881	0.0119

Poisson probability table for λ = 3 (values rounded to four decimal places)

The Poisson CDF for λ = 3, displayed as a step function. Each filled dot marks the value of F(k) at the corresponding integer k, while the open dot indicates the left-hand limit (the CDF just before the jump). The function rises steeply near the mean and approaches 1 asymptotically.

Observe how the CDF climbs most rapidly in the neighborhood of k = 2 and k = 3, which correspond to the peak of the PMF. By k = 5, the CDF already exceeds 0.91, confirming that values above 5 are relatively improbable when λ = 3. In practice, the complement rule P(X > k) = 1 − P(X ≤ k) is frequently used to compute upper-tail probabilities—for instance, to determine whether an observed count is unusually large.

SECTION 6

Worked Example — Call Center Arrivals

A customer service call center receives an average of 4.5 calls per minute during peak hours. Assuming that calls arrive independently and at a constant average rate, find the probability that exactly 3 calls arrive in a given minute, and the probability that more than 6 calls arrive in that minute.

Poisson Probability Calculations (λ = 4.5)

1
Step 1 — Verify Assumptions & Identify Parameters

Calls arrive independently and at a constant rate, satisfying the Poisson process assumptions. The rate parameter is λ = 4.5 calls per minute. We seek P(X = 3) and P(X > 6), where X ~ Poisson(4.5).

λ = 4.5, k₁ = 3, threshold k₂ = 6
2
Step 2 — Compute P(X = 3)

Applying the PMF: P(X = 3) = $e^{−4.5}$ × 4.5³ / 3! = $e^{−4.5}$ × 91.125 / 6. First compute $e^{−4.5}$ ≈ 0.01111. Then P(X = 3) ≈ 0.01111 × 15.1875 ≈ 0.1687.

P(X = 3) ≈ 0.1687
3
Step 3 — Compute P(X ≤ 6) via Cumulative Sum

To find P(X > 6), we first compute P(X ≤ 6) = Σ P(X = k) for k = 0, 1, …, 6. Evaluating each term: P(0) ≈ 0.0111, P(1) ≈ 0.0500, P(2) ≈ 0.1125, P(3) ≈ 0.1687, P(4) ≈ 0.1898, P(5) ≈ 0.1708, P(6) ≈ 0.1281. Summing: P(X ≤ 6) ≈ 0.8311.

P(X ≤ 6) ≈ 0.8311
4
Step 4 — Apply the Complement Rule

P(X > 6) = 1 − P(X ≤ 6) = 1 − 0.8311 = 0.1689. There is approximately a 16.9% chance that more than six calls arrive in any given minute.

P(X > 6) ≈ 0.1689
5
Step 5 — Interpretation

About 17% of one-minute intervals will see 3 calls arrive, and roughly the same fraction will see more than 6 calls. A call center manager might use these probabilities to determine staffing levels—ensuring enough agents are available to handle the upper-tail scenario roughly one minute out of every six.

SECTION 7

Strengths, Limitations & Comparisons

The Poisson distribution is powerful in its simplicity—one parameter, clean mathematical properties, and wide applicability—but its assumptions are restrictive. Understanding when the Poisson model is appropriate and when to reach for alternatives is a core competence in applied statistics.

Feature	Strengths	Limitations
Parameterization	Single parameter λ makes estimation (via MLE: λ̂ = x̄) trivially simple and interpretable.	Cannot independently control mean and variance—both are locked to λ.
Tail behavior	Assigns positive probability to all non-negative integers, naturally accommodating rare high counts.	Underestimates extreme counts when data are overdispersed (Var > Mean).
Independence	Independence assumption simplifies analysis and enables the reproductive (additive) property.	Fails when events cluster (e.g., earthquake aftershocks, viral social media posts).
Constant rate	Homogeneous rate simplifies computation and makes the model time-/space-invariant.	Inapplicable when rates vary over time (e.g., rush-hour traffic vs. midnight).
Approximation power	Excellent approximation to the binomial when n is large and p is small.	Approximation quality degrades when np > 10 or p is not small.

Strengths and limitations of the Poisson distribution

⚠

WHEN TO REACH FOR ALTERNATIVES

If your data show overdispersion (sample variance much larger than sample mean), consider the negative binomial distribution, which introduces a second parameter to decouple mean and variance. If the rate is not constant, a non-homogeneous Poisson process with a time-varying intensity function λ(t) may be appropriate. For data with excess zeros (more zeros than the Poisson predicts), zero-inflated Poisson (ZIP) models mix a point mass at zero with a standard Poisson component.

SECTION 8

Connections to Advanced Theory

The Poisson distribution is not an isolated construct; it occupies a central node in the web of probability distributions and stochastic processes. Its connections to the exponential distribution, the gamma distribution, and the broader family of generalized linear models make it indispensable in advanced coursework and research.

Poisson Distribution	Advanced Extension
Counts events in a fixed interval	The exponential distribution models the waiting time between successive events in the same Poisson process. If X ~ Poisson(λt), then inter-arrival times follow Exp(λ).
Single parameter λ (mean = variance)	The negative binomial generalizes the Poisson by introducing a dispersion parameter, allowing Var > Mean. It arises naturally as a Poisson–gamma mixture.
PMF with fixed λ	Poisson regression (a GLM with log link) models the rate parameter as a function of covariates: log(λᵢ) = β₀ + β₁xᵢ₁ + ⋯ + βₚxᵢₚ.
Reproductive property (sum of Poissons is Poisson)	Compound Poisson processes allow each event to carry a random "mark" (e.g., claim size in insurance), leading to aggregate loss distributions used in actuarial science.
Approximates the binomial for large n, small p	The normal approximation to the Poisson (X ≈ N(λ, λ) for large λ) is a direct consequence of the Central Limit Theorem and is used for constructing confidence intervals for rates.

How the Poisson distribution connects to advanced topics

If you continue into courses on stochastic processes, you will encounter the Poisson point process in higher dimensions—modeling, for example, the spatial distribution of trees in a forest or galaxies in a sector of the sky. The mathematical machinery you have learned here—the PMF, the mean-equals-variance property, the MGF—transfers directly to these richer settings.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A Poisson random variable has λ = 7. Without performing any calculations, explain why the probability P(X = 7) is not necessarily the maximum of the PMF, and identify what integer(s) achieve the mode. Justify your reasoning using the relationship between consecutive Poisson probabilities.

Show answer

The ratio of consecutive Poisson probabilities is P(X = k)/P(X = k − 1) = λ/k. This ratio exceeds 1 when k < λ and falls below 1 when k > λ, so the PMF increases up to k = ⌊λ⌋ and decreases thereafter. When λ is an integer, this ratio equals exactly 1 at k = λ, meaning P(X = λ) = P(X = λ − 1), so the distribution is bimodal at k = 6 and k = 7. Therefore P(X = 7) is tied for the maximum, and the PMF does not have a unique mode at k = 7.

Question 2: BASIC CALCULATION

A website experiences an average of 2 server errors per hour. Assuming errors follow a Poisson distribution, compute the probability that exactly 4 errors occur in a given hour.

Show answer

With λ = 2 and k = 4, we apply the PMF: P(X = 4) =

e^{−2}

× 2⁴ / 4! =

e^{−2}

× 16 / 24 = 0.13534 × 0.66667 ≈ 0.0902. There is approximately a 9.02% chance of observing exactly 4 server errors in a one-hour window. This result makes intuitive sense: 4 is double the mean of 2, so while not impossible, it is moderately unlikely.

Question 3: INTERMEDIATE

A hospital emergency department sees an average of 6 trauma cases per 8-hour shift. What is the probability that during a 4-hour period, at most 2 trauma cases are admitted? State any necessary adjustments to the rate parameter.

Show answer

The rate for an 8-hour shift is 6, so the rate for a 4-hour period is λ = 6 × (4/8) = 3 trauma cases per 4 hours. We need P(X ≤ 2) = P(X = 0) + P(X = 1) + P(X = 2) with λ = 3. Computing: P(0) =

e^{−3}

≈ 0.0498, P(1) =

e^{−3}

× 3 ≈ 0.1494, P(2) =

e^{−3}

× 9/2 ≈ 0.2240. Therefore P(X ≤ 2) ≈ 0.0498 + 0.1494 + 0.2240 = 0.4232. There is roughly a 42.3% chance of at most 2 trauma cases in a 4-hour period. The key step was rescaling λ proportionally to the interval length.

Question 4: APPLIED

A manufacturing plant produces semiconductor chips, and defects occur at an average rate of 1.5 per 100 cm² of wafer surface. A particular chip occupies 20 cm². (a) What is the probability that a randomly selected chip has zero defects? (b) If the plant produces 500 chips, how many are expected to be defect-free?

Show answer

(a) The rate per 20 cm² is λ = 1.5 × (20/100) = 0.3 defects per chip. P(X = 0) =

e^{−0.3}

× 0.3⁰ / 0! =

e^{−0.3}

≈ 0.7408. So approximately 74.1% of chips will be defect-free. (b) The expected number of defect-free chips in a batch of 500 is 500 × 0.7408 ≈ 370.4, so roughly 370 chips are expected to pass a zero-defect inspection. This calculation illustrates how the Poisson model converts a defect rate per unit area into a yield estimate—a standard application in semiconductor quality control.

Question 5: CRITICAL THINKING

An analyst collects data on the number of customer complaints received per day at a retail chain over 200 days. The sample mean is 3.8 complaints per day, and the sample variance is 8.1. (a) Assess whether a Poisson model is appropriate. (b) If it is not, propose an alternative model and explain how it addresses the observed discrepancy. (c) Describe a formal statistical test the analyst could perform to evaluate the goodness-of-fit of the Poisson model.

Show answer

(a) For a Poisson distribution, the mean and variance should be approximately equal. Here the sample variance (8.1) is more than twice the sample mean (3.8), indicating substantial overdispersion. The Poisson model is likely inappropriate. (b) The negative binomial distribution is a natural alternative because it introduces a dispersion parameter that allows the variance to exceed the mean: Var(X) = μ + μ²/r, where r controls the degree of overdispersion. This extra flexibility can accommodate the clustering or heterogeneity in complaint rates that the Poisson model cannot capture. (c) The analyst could perform a chi-squared goodness-of-fit test: compute expected Poisson counts for each value of k using λ̂ = 3.8, then compare observed and expected frequencies via the statistic Σ (Oᵢ − Eᵢ)² / Eᵢ, which follows χ² with (number of bins − 2) degrees of freedom. A simpler preliminary check is the dispersion test, which compares (n − 1)s²/x̄ to a χ²(n−1) distribution under the null hypothesis that the data are Poisson.

SUMMARY

Poisson Distribution — Summary

Summary

The Poisson distribution models the count of independent events occurring at a constant average rate λ within a fixed interval. Its PMF, P(X = k) = $e^{−λ}$ × λᵏ / k!, depends on a single parameter and yields the distinctive property that mean equals variance. The distribution is right-skewed for small λ and approaches normality as λ grows, reflecting the Central Limit Theorem.

Key applications include modeling arrivals in queueing systems, defect counts in manufacturing, disease incidence in epidemiology, and rare events in general. The Poisson arises as a limiting case of the binomial when n is large and p is small (np = λ), and it connects forward to the exponential distribution (inter-arrival times), negative binomial (overdispersed counts), and Poisson regression in generalized linear models. Always verify the mean-equals-variance assumption before applying the Poisson model to real data.