The Poisson Distribution

The Poisson distribution gives the probability of observing exactly $k$ events in a fixed interval (time, area, volume, or a genome), when events happen independently at a constant average rate $\lambda$ (lambda) per interval. Its probability mass function is

where $\lambda > 0$ is the only parameter, $e \approx 2.71828$ is Euler's number, and $k!$ is the factorial of $k$. Unlike the binomial distribution, which requires you to specify both a number of trials $n$ and a per-trial probability $p$, the Poisson requires only $\lambda$, the expected number of events. This economy is not an accident — it reflects the fact that Poisson variables arise in a regime where $n$ and $p$ individually become meaningless (there might be no natural "number of trials" for mutations along a genome), but their product remains a well-defined average rate. Two defining mathematical facts make the Poisson distribution instantly recognizable and testable. First, its expected value and variance are equal:

This is a strong, falsifiable prediction: if you count events in many equal-sized intervals (many plates, many quadrats, many time windows) and compute the sample mean and sample variance of the counts, a true Poisson process demands they match. Second, the Poisson distribution is closed under summation: if $X_1$ and $X_2$ are independent Poisson variables with rates $\lambda_1$ and $\lambda_2$, their sum $X_1 + X_2$ is Poisson with rate $\lambda_1 + \lambda_2$. This additivity is why you can scale up — combine plates, extend an observation window, pool genomes — and still expect a Poisson count, just with a bigger $\lambda$. The concept requires three assumptions, direct descendants of Independence Revisited: events occur independently of each other (one mutation does not make another more or less likely); the average rate $\lambda$ is constant across the interval (no hidden time-of-day or hotspot effect); and two events cannot occur at exactly the same point (in the continuous limit, "simultaneity" has probability zero). When these hold, the Poisson distribution is not merely a convenient approximation — it is the unique mathematical consequence, in the same sense that the Binomial Distribution is the unique consequence of independent Bernoulli trials.

Step 1: Start from the binomial world. Imagine a bacterial genome of $n$ possible base positions, each of which independently mutates with tiny probability $p$ during one round of replication. The number of mutations $X$ is binomial: $X \sim \text{Binomial}(n, p)$, with mean $np$ and variance $np(1-p)$.

Step 2: Take the rare-event limit. Biologically, $n$ (positions, cells, time-ticks) is enormous and $p$ (probability any one of them yields an event) is tiny. What stays finite and meaningful is the product $\lambda = np$ — the expected total count. Let $n \to \infty$ and $p \to 0$ while holding $np = \lambda$ fixed. Algebraically expanding the binomial mass function under this limit:

The combinatorial term $\binom{n}{k}p^k$ converges to $\lambda^k/k!$ because there are effectively infinitely many "slots" but each contributes an infinitesimal chance; the $(1-p)^{n-k}$ term converges to $e^{-\lambda}$ by the standard limit definition of $e$. Every trace of the artificial "number of trials" $n$ disappears — only the rate $\lambda$ remains.

Step 3: Interpret $\lambda$ as an intensity. In continuous settings — colonies growing on a plate, mutations along a chromosome, radioactive decays per minute — $\lambda$ is reinterpreted as a rate multiplied by the size of the observation window: $\lambda = r \times T$ (rate per unit time or area, times the window size). Double the plate area, double $\lambda$; double the exposure time, double $\lambda$. This proportionality is the practical handle biologists use to scale predictions.

Step 4: Assemble the distribution across the sample space. Practically, imagine dividing your interval (a genome, a time window, a Petri dish) into an enormous number of tiny, non-overlapping sub-intervals so small that each can contain at most one event. Each sub-interval is a near-independent Bernoulli trial with a vanishingly small success probability. Summing across all sub-intervals recovers exactly the binomial-to-Poisson limit derived above — this is the mechanistic picture biologists actually use: many independent "chances," each individually negligible, aggregating into a rare-but-nonzero count.

Step 5: Recognize the diagnostic signature. Because $\text{Var}(X) = E(X) = \lambda$, biologists routinely compute the index of dispersion, $D = s^2/\bar{x}$, from replicate counts. A value near 1 supports the Poisson (random, independent) model. A value well above 1 (overdispersion) signals clumping — contagious spread of disease, clustered mutational hotspots, or patchy colony growth from cell aggregates. A value well below 1 (underdispersion) signals regularity or repulsion — territorial spacing of nesting birds, or self-avoiding placement.

Bortkiewicz's cavalry deaths (1898). The statistician Ladislaus Bortkiewicz used Prussian army records to test whether rare, essentially random events — a soldier being fatally kicked by a horse — followed a Poisson distribution. Across 10 army corps observed over 20 years (200 corps-years of data), he tallied the number of deaths per corps per year. The average rate was $\lambda \approx 0.61$ deaths per corps-year. Comparing the observed frequency of 0, 1, 2, 3, 4+ deaths per corps-year against the Poisson formula's predictions, he found startlingly close agreement:

This is one of the earliest and cleanest empirical demonstrations that a purely mathematical limit (binomial to Poisson) accurately describes a genuinely rare, quasi-random biological/social event. It established the Poisson distribution's credibility as a real-world model, not just an algebraic trick — and it is the direct conceptual ancestor of using Poisson models for rare-disease counts and mutation counts.

The Luria–Delbrück fluctuation test (1943). Salvador Luria and Max Delbrück addressed a foundational question in bacterial genetics: do mutations conferring resistance to bacteriophage arise spontaneously, before selection is applied (Darwinian, undirected), or are they induced by the phage itself in response to the challenge (Lamarckian, directed)? If mutations arise only after exposure to the phage — each bacterium independently "deciding" to mutate with the same small probability at the moment of exposure — resistant colony counts across replicate cultures should follow a Poisson distribution: mean roughly equal to variance, little extreme scatter. Luria and Delbrück grew many small parallel bacterial cultures from a common ancestor, then plated each culture on phage-laden agar and counted resistant colonies. If mutations were induced at plating, every culture should show a similar, Poisson-distributed number of resistant colonies. Instead they observed wild "jackpots": most cultures had few or no resistant colonies, but occasionally a culture produced enormous numbers of resistant colonies — far more than a Poisson distribution with the same mean could plausibly generate. The variance vastly exceeded the mean. This fluctuation pattern is exactly what is predicted if mutations arise spontaneously and randomly during ordinary growth, before phage exposure: a mutation occurring early in a culture's growth gets copied into all its descendants, producing a jackpot; a mutation occurring late produces only a few resistant cells. The distribution of final colony counts becomes far more variance-inflated than Poisson — described instead by the Luria–Delbrück distribution. The experiment used the Poisson distribution as the null model to be broken: because the data were so wildly overdispersed relative to Poisson expectations, Luria and Delbrück could reject the induced-mutation hypothesis and support spontaneous, pre-existing mutation, a landmark result confirmed independently by Newcombe's replica-plating experiment (1949) and the Lederbergs' replica plating (1952). This is a superb illustration of the Poisson distribution's evidentiary role: it is precise enough that violations of it become discoveries in their own right.

Example 1 — mutation counts across a bacterial chromosome. Suppose a particular class of point mutation arises independently at each of $n = 4{,}500{,}000$ base pairs in E. coli's genome during one round of replication, each site with probability $p = 2 \times 10^{-9}$ per replication (a realistic order of magnitude for base-substitution mutation rates). The expected number of such mutations per genome per replication is

Because $n$ is enormous and $p$ is tiny, $X$, the mutation count per genome, is Poisson with $\lambda = 0.009$. What is the probability a daughter cell acquires at least one such mutation?

So roughly 0.9% of replications yield at least one mutation of this class — about 1 in 112 divisions. Now suppose we grow a culture until it has produced $10^8$ independent replication events (cell divisions) — Poisson additivity says the total expected mutation count across all of them is $\lambda_{\text{total}} = 0.009 \times 10^8 = 900{,}000$ mutations pooled across the population lineage, though any single lineage still carries a Poisson(0.009) count per division. This is exactly the reasoning underlying the Luria–Delbrück logic: rare per-genome events, summed over enormous numbers of independent replication opportunities, generate a predictable population-level mutational supply — and it's the departure from this Poisson expectation (jackpots) that revealed the timing of mutation.

Example 2 — surveillance for a disease cluster. A public-health office in a region of 500,000 people has tracked a rare congenital anomaly for 15 years, observing on average 3 cases per year (background rate). This year, one particular town within the region — population 20,000, about 4% of the regional population — reports 4 cases. Is this town a statistically meaningful cluster, or plausible background noise? Expected cases in that town, scaled by its population share, assuming the background rate applies uniformly:

Under $H_0$: town count $X \sim \text{Poisson}(0.12)$. The probability of observing 4 or more cases by chance is

Computing term by term: $P(0) = e^{-0.12} = 0.8869$, $P(1) = 0.8869 \times 0.12 = 0.1064$, $P(2) = 0.1064 \times 0.12/2 = 0.00639$, $P(3) = 0.00639 \times 0.12/3 = 0.000256$. Summing: $0.8869 + 0.1064 + 0.00639 + 0.000256 = 0.99995$, so

Observing 4 cases when only 0.12 are expected is extraordinarily improbable under the pure-chance Poisson model — a p-value around $5\times10^{-5}$. This kind of calculation is the statistical backbone of disease-cluster investigations: it does not prove a specific environmental cause, but it flags the town as a strong candidate for follow-up (checking for a shared exposure, reporting bias, or a genuine local risk factor) rather than dismissing it as noise.

Siméon Denis Poisson (1781–1840) was a French mathematician and physicist who derived the distribution that bears his name in 1837, in a work on the probability of judgments in criminal and civil trials — not biology at all. Poisson was interested in how rare, essentially independent events (wrongful convictions, jury errors) accumulate over many opportunities, and he derived the limiting distribution as the number of trials grows large while their probability shrinks. He had no inkling that a century later his formula would become indispensable for counting radioactive decays, bacterial mutations, and disease clusters — a striking case of a purely mathematical result finding its deepest empirical vindication in biology and physics long after its invention. The distribution's biological credibility was cemented by two later figures. Ladislaus Bortkiewicz (1868–1931), a Russian-born economist and statistician working in Germany, used Prussian cavalry death records to give the first celebrated empirical demonstration that Poisson's formula accurately predicts real-world rare-event frequencies. Decades later, Salvador Luria (1912–1991) and Max Delbrück (1906–1981), physicists-turned-biologists at the heart of the emerging field of molecular biology, used departures from the Poisson distribution to settle one of genetics' foundational disputes — whether mutations are spontaneous or induced by selection — winning them a share of the 1969 Nobel Prize in Physiology or Medicine (with Alfred Hershey) for discoveries concerning the replication mechanism and genetic structure of viruses.

Genomics and mutation surveillance. Poisson (and its overdispersed cousin, negative binomial) models underlie tools that call somatic mutations and structural variants from sequencing read counts, and they estimate baseline mutation rates in cancer genomics, where unusually high local mutation counts ("kataegis," mutational hotspots) are flagged precisely because they violate Poisson expectations.

Microbiology and colony counting. Colony-forming unit (CFU) counts on agar plates, viral plaque assays, and Poisson-based dilution methods (most-probable-number and limiting-dilution assays) all rely on the Poisson formula to back-calculate concentrations of microbes or infectious particles from countable plate data.

Epidemiology and public health. Disease-cluster investigations, as in the worked example above, use Poisson tests to decide whether an observed local excess of a rare disease (cancer clusters, birth defects, rare infections) exceeds background chance before committing resources to environmental or genetic investigation.

Pharmacology and toxicology. Rare adverse drug events reported to surveillance systems (like the FDA's Adverse Event Reporting System) are modeled with Poisson-based disproportionality statistics to flag drug-event pairs occurring more often than chance would predict.

Ecology and conservation. Counts of rare species per quadrat, camera-trap detections of endangered animals per unit time, and species-abundance surveys use Poisson (or its overdispersed extensions) to estimate population densities and detect non-random habitat clustering, informing conservation prioritization.

Astro- and radiobiology. Radiation dosimetry and cosmic-ray background counts use Poisson statistics to set detection thresholds distinguishing genuine signal from stochastic background noise — directly relevant to radiobiological risk assessment. Understanding the Poisson distribution is also a direct stepping stone toward Sampling Distributions: it is the first distribution students meet whose shape emerges purely from a limiting argument rather than direct enumeration, previewing the logic used later for the sampling distribution of the mean and other statistics built from many small, independent contributions.

Ask the tutor: How would the calculations change if the events I'm counting weren't truly independent — say, mutations that make nearby mutations more likely?