Poisson Distribution

The Poisson distribution models the number of events occurring in a fixed interval of time or space, given that those events happen independently of one another and at a known constant average rate — customer arrivals per hour, typos per page, server errors per minute, radioactive decay events per second. It's named after Siméon Denis Poisson, who first published the distribution in 1837 in the context of modeling the probability of wrongful criminal convictions in French courts, well before its now-dominant applications in queuing theory, reliability engineering, and operational monitoring existed.

The Poisson distribution is defined by a single parameter, λ (lambda), representing the average rate of events over the interval in question — unlike the binomial distribution's two parameters (number of trials and success probability per trial), the Poisson distribution requires no fixed number of discrete trials at all, which is precisely what makes it the appropriate choice for modeling events that can occur any number of times within a continuous interval, like phone calls arriving at a call center, rather than events bounded by a known, fixed number of opportunities to occur. This distinction matters in practice: a process like "number of defective items in a batch of exactly 100" fits the binomial model naturally (a fixed number of trials), while "number of customer support tickets filed in a day" fits the Poisson model instead, since there's no natural upper bound on how many tickets could theoretically be filed.

A distinctive and practically important property of the Poisson distribution is that its mean and its variance are mathematically equal, both equal to λ — a property that's a direct consequence of its mathematical derivation, and one that provides a useful real-world diagnostic: if observed count data shows a variance substantially larger than its mean (called overdispersion), that's a strong signal the Poisson model's underlying assumptions don't actually hold for the process generating the data, typically because the true event rate isn't actually constant (it varies across different subgroups or time periods being lumped together) or because the events aren't genuinely independent (one event makes a related event more likely, as with contagious disease spread or clustered customer complaints following a single underlying incident). When overdispersion is detected, the negative binomial distribution — a related but more flexible distribution with an extra parameter specifically to allow variance greater than the mean — is the standard modeling alternative reached for instead.

At low values of λ, the Poisson distribution is heavily right-skewed, with most of its probability mass concentrated at zero and small counts and a long thin tail extending toward larger, increasingly rare values — a shape that fits real low-rate-event data (rare manufacturing defects, rare server crashes, rare customer complaints) far better than a normal distribution would, since a normal distribution assumes symmetry and permits negative values, neither of which makes sense for a count of discrete events that can never go below zero. As λ grows large, the Poisson distribution increasingly approximates a normal distribution, the same convergence pattern the binomial distribution exhibits as its own parameters grow, both ultimately traceable to the central limit theorem operating on sums of many small, independent contributions.

The Poisson distribution underlies queuing theory, the mathematical study of waiting lines, which models customer or request arrivals as a Poisson process specifically to answer practical operational questions — how many staff or servers are needed to keep average wait times below a target, how a call center's staffing should scale with expected call volume — making it one of the more directly operationally useful named distributions in applied business and engineering analytics, beyond its purely statistical and theoretical interest.

Read the Poisson distribution as the standard model for counting rare, independent events over a continuous interval with no natural upper bound on how many could occur: its single parameter λ simultaneously determines both its mean and variance, a mathematical property that doubles as a built-in diagnostic for detecting when the model doesn't actually fit (via overdispersion), and its central role in queuing theory is a direct, practical application of exactly the kind of "how many events in this interval" question the distribution was built to answer.