Can lambda be smaller than one?

Yes. An expected count such as 0.2 is valid even though observed counts are integers.

Does a Poisson model require rare events?

Not as a mathematical definition. Rarity is relevant to certain approximations and applications.

What happens when lambda is zero?

The distribution is degenerate at zero: the count is always zero under that model.

AI Glossary / Statistics and probability / Intermediate

Poisson distribution

Definition

The Poisson distribution is a probability distribution for a count taking the values 0, 1, 2 and so on. Its parameter lambda is the expected count in the specified exposure, such as an hour, a square meter or a batch. The probability of count k is exp(-lambda) times lambda^k divided by k factorial. For a Poisson random variable, the mean and variance both equal lambda. The distribution is useful for count modeling, but an outcome being a count does not by itself make a Poisson model appropriate.

What it means

A Poisson model describes variation around an expected number of events. If the expected hourly count is three, it does not predict exactly three events every hour. Zero, two, four or more events remain possible, with probabilities determined by the distribution. A homogeneous Poisson process provides one familiar justification: events occur with a constant rate and independent counts in disjoint intervals. The distribution can also be used as a conditional model in regression, where different observations have different expected counts based on exposure and predictors.

How it works

Define precisely what is counted and the exposure over which counting occurs. Specify or estimate lambda, then calculate individual probabilities or cumulative probabilities as needed. For independent observations with equal exposure and a common Poisson mean, the sample mean estimates lambda by maximum likelihood. Unequal exposures require an appropriate adjustment; a rate per unit exposure is not interchangeable with a count mean. Examine whether the observed variation and dependence are compatible with the model. Heterogeneous rates, bursts, event clustering, excess zeros or detection errors can require a richer explanation and a different statistical model.

Three practical examples

Operations: model incoming requests in a short interval when a suitable arrival process is plausible.
Inspection: model defect counts over comparable inspected areas, distinguishing the count of defects from the number of items with at least one defect.
AI monitoring: model event counts conditional on traffic exposure and relevant predictors, then assess whether dependence or changing rates makes the simple Poisson assumptions inadequate.

Think of it this way

Think of a counter watched for a specified length of time. The parameter sets its expected increase, while the distribution describes random variation around that expectation. Doubling the observation time under a constant-rate process doubles the expected count; it does not force the realized count to double on every occasion.

Worked example

Suppose an artificial homogeneous Poisson process has an average rate of six requests per hour. In a half-hour window, lambda is 6 times 0.5 = 3. The probability of exactly two requests is exp(-3) times 3^2/2!, approximately 0.2240. The probability of at least four is 1 minus the probabilities of zero through three, which equals 1 - exp(-3)(1 + 3 + 4.5 + 4.5), approximately 0.3528. For a fifteen-minute window, lambda is 1.5 and the probability of zero requests is exp(-1.5), approximately 0.2231. In the half-hour window the variance is 3 and the standard deviation is sqrt(3), approximately 1.732. This original calculation assumes a constant rate and the stated process model. Real traffic with bursts or time-of-day changes may have substantially different probabilities even when its overall average is six per hour.

Illustrative case study

In a fictional case, Atlas Support uses a Poisson model for short-window arrival counts. Initial averages look reasonable, but analysts discover that requests arrive in clusters after scheduled notifications. They compare variability within comparable periods and examine dependence across windows. The team revises its forecasting model rather than treating every count above the Poisson expectation as an anomaly. Its report retains the simple model as a baseline and explains why the production process needs a more detailed treatment of timing.

Common misconceptions and mistakes

Lambda is the expected count for a particular exposure, not necessarily a rate with units of events per hour. The variance equals the mean, while the standard deviation is the square root of the mean. Poisson probabilities concern integer counts; fractional counts are not possible outcomes even though lambda can be fractional. Independence of disjoint counts is a process assumption, not something guaranteed by choosing the distribution name. A count with a fixed finite number of opportunities may instead call for a binomial model. Overdispersion can arise from omitted heterogeneity or dependence, so it should be investigated rather than handled mechanically.

Business use

Businesses can use Poisson probabilities to plan capacity, assess event frequencies and establish a transparent baseline for monitoring. Match the exposure units to the operational question and avoid mixing quiet and busy periods without accounting for the difference. If estimated probabilities influence staffing or alerts, evaluate them against new observations and examine calibration of tail events. Poisson regression can include an exposure offset, but its mean-variance assumptions still deserve attention. Decisions should account for uncertainty in the estimated rate as well as randomness in future counts.

Check your understanding

A constant rate is eight events per hour. What Poisson mean applies to a fifteen-minute interval?

Answer: 2, because fifteen minutes is one quarter of an hour. The interval must be specified before calculating probabilities.

Related terms

Count data
Poisson regression
Poisson process
event rate
exposure
overdispersion
negative binomial distribution

Learn next

Learn how independent Poisson counts add and how exposure enters a regression model, then compare Poisson assumptions with models for heterogeneous or clustered counts.

Regression model

Sources

NIST/SEMATECH, Poisson Distribution: https://itl.nist.gov/div898/handbook/eda/section3/eda366j.htm NIST/SEMATECH, Counts Control Charts: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc331.htm

Previous: Poisson-Dirichlet distribution Next: Poisson equation