Probability & Statistics · Poisson distribution

Poisson distribution

The probability of exactly k events in a fixed interval when events arrive one at a time at a constant average rate, with counts in non-overlapping intervals independent and λ expected in this one: P(X = k) = e^(−λ)·λᵏ/k!. Cut the interval into n slots, each with chance λ/n, and let n grow: the binomial distribution turns into it.

Poisson distribution in this visualization

The Poisson distribution describes how many times something happens in a fixed stretch of time, length or area — calls to a help line in an hour, typos on a page — when the events come one at a time at a steady average rate and the counts in non-overlapping stretches are independent. It has a single parameter, the expected count λ\lambda (the rate times the length of the stretch), and P(X=k)=e−λλk/k!P(X = k) = e^{-\lambda}\lambda^k/k! for k=0,1,2,…k = 0, 1, 2, \dots. Here the interval is cut into nn slots, each holding an event with chance λ/n\lambda/n: the count is then binomial, and as nn doubles its red bars settle onto the Poisson outline while the expected count stays λ\lambda.

Where the formula comes from

With p=λ/np = \lambda/n, the binomial probability splits into three factors:(nk)pk(1−p)n−k=n(n−1)⋯(n−k+1)nk\binom{n}{k}p^k(1-p)^{n-k} = \frac{n(n-1)\cdots(n-k+1)}{n^k}⋅λkk!⋅(1−λn)n−k{}\cdot\frac{\lambda^k}{k!}\cdot\Bigl(1-\frac{\lambda}{n}\Bigr)^{n-k}. The first factor compares the ordered ways to choose kk different slots with nkn^k; for a fixed kk it tends to 1. The middle factor does not depend on nn at all. The last says that all the other slots stay empty, and it tends to e−λe^{-\lambda}: it is the compound-interest limit (1+r/n)n→er(1 + r/n)^n \to e^r with r=−λr = -\lambda, and the leftover (1−λ/n)−k(1 - \lambda/n)^{-k} tends to 1. Multiplying the three limits gives the Poisson formula. For a fixed λ\lambda, once nn is large the largest gap between the two distributions shrinks roughly like 1/n1/n, which is why the record’s ratios settle near 2.

Mean and variance are both λ

With nn slots the mean is n⋅λ/n=λn\cdot\lambda/n = \lambda for every nn, and the variance is np(1−p)=λ(1−λ/n)np(1-p) = \lambda(1 - \lambda/n), a little less than λ\lambda. The factor 1−λ/n1 - \lambda/n tends to 1 as nn grows, so the Poisson distribution has variance λ\lambda as well, and standard deviation λ\sqrt{\lambda}.

The shape

Each Poisson bar is λ/(k+1)\lambda/(k+1) times the one before it: P(X=k+1)=P(X=k)⋅λ/(k+1)P(X = k+1) = P(X = k)\cdot\lambda/(k+1). While that factor is above 1 the bars rise, and once it drops below 1 they fall. So for λ<1\lambda < 1 the most likely count is 0 and the bars fall from the start; otherwise the peak is at ⌊λ⌋\lfloor \lambda \rfloor, and when λ\lambda is a whole number the counts λ−1\lambda - 1 and λ\lambda are equally likely. As λ\lambda grows the distribution becomes nearly symmetric and bell-shaped, for reasons the central limit theorem explains.

Things to notice

  • λ need not be a whole number: an average of 0.5 events per interval is as good a model as 3.
  • An average of λ does not make λ events the usual outcome: with λ = 3, exactly 3 happen only about 22% of the time.
  • “One at a time” is part of the model. The slots build it in, since a slot holds at most one event; events that arrive in batches give a different count.

Controls

  • Click a column, or drag the triangle under the bars, to read that count’s three factors.

Related