Statistics › Section D · Discrete random variables › Topic D.6

The Poisson distribution

How many emails land in an hour, typos hide in a chapter, customers arrive in a minute? Counts of events in a fixed window, with no natural upper limit, follow the Poisson distribution - the model of rare happenings at a known rate.

Counting events per interval

How many emails land in an hour? How many typos hide in a chapter, accidents on a stretch of road, customers at a counter in a minute? These are counts of events in a fixed interval of time or space, with no natural upper limit - and the Poisson distribution is their model. Where the binomial counts successes in a fixed number of trials, the Poisson counts happenings in a fixed window.

The Poisson setting

Two characteristics define it. 1. The events occur in a fixed interval of time or space - an hour, a chapter, a mile - with a known average rate. 2. Occurrences are independent of each other and of the time since the last event: one arrival neither encourages nor discourages the next. There is no fixed number of trials and no per-trial success probability - just a rate and a window in which to count. The count \(X\) has no upper limit in principle: 50 emails in an hour is possible, merely unlikely.

The formula

The probability of exactly \(k\) events, when the average is \(\lambda\) per interval, is

P(X = k) = \frac{e^{-\lambda}\lambda^k}{k!}

Here \(e \approx 2.71828\) is the natural base, and \(k!\) means \(k(k-1)\cdots 1\), with \(0! = 1\) by convention. The formula balances two forces: larger \(k\) makes \(\lambda^k\) grow, but \(k!\) grows faster, pulling the probabilities back down for large counts. Everything depends on one number - the rate \(\lambda\).

Mean equals variance

The Poisson distribution has one parameter, \(\lambda\), and it plays both roles: the mean count per interval and the variance. Busier intervals are also noisier ones - a counter averaging \(\lambda = 10\) swings more than one averaging 2, since the standard deviation is \(\sqrt{\lambda}\). This mean-equals-variance signature doubles as a quick diagnostic: if a data set's variance far exceeds its mean, the events are clustering, and the plain Poisson model is too simple for them.

Worked example: calls at a hotline

A support hotline averages \(\lambda = 3\) calls per hour. 1. \(P(0) = e^{-3} \approx 0.0498\) - fully quiet hours are rare. 2. \(P(1) = 3e^{-3} \approx 0.149\). 3. \(P(2) = \frac{9e^{-3}}{2} \approx 0.224\). 4. \(P(3) = \frac{27e^{-3}}{6} \approx 0.224\) - identical to \(P(2)\): the two highest bars. 5. Check: each step multiplies the previous probability by \(\frac{\lambda}{k}\), and \(\frac{3}{3} = 1\), so for a whole-number \(\lambda\) the counts \(\lambda - 1\) and \(\lambda\) always tie for the peak. These four values sum to about 0.647, so roughly 35% of hours bring 4 or more calls.

Worked example: typos in a chapter

An editor knows a chapter averages \(\lambda = 2\) typos. a) \(P(\text{no typos}) = e^{-2} \approx 0.135\). b) \(P(\text{at least one}) = 1 - 0.135 = 0.865\) - most chapters need correcting. c) \(P(\text{exactly 4}) = \frac{2^4 e^{-2}}{4!} = \frac{16}{24}e^{-2} \approx 0.090\) - a heavy-typo chapter, unusual but hardly impossible. Check: part (b) is the complement rule working with the Poisson formula, and \(0.135 + 0.865 = 1\).

Scaling λ to the interval

\(\lambda\) is a rate attached to a specific interval, and it must match the question's window. Three calls per hour means \(\lambda = 6\) for a two-hour block, \(\lambda = 1.5\) for a half-hour, and \(\lambda = 72\) for a 24-hour day. Rescaling is multiplication by the interval ratio - never recompute from scratch. Forgetting it is the most common practical error: \(P(\text{no calls in 2 hours})\) is \(e^{-6} \approx 0.0025\), not \(e^{-3}\), because the expected count doubles with the window.

The binomial connection

The Poisson distribution is the binomial's limit for rare events: when \(n\) is large and \(p\) is small, the binomial count is well approximated by a Poisson with \(\lambda = np\). A website's 1,000 pages each carry a 0.002 chance of a broken link, so \(np = 2\): \(P(\text{no broken links})\) is \(0.998^{1000} \approx 0.135\) by the binomial and \(e^{-2} \approx 0.135\) by the Poisson - agreeing to three decimals. The approximation trades an awkward binomial for a one-parameter Poisson, and it explains why rare-event counts so often look Poisson in the wild.

Common mistakes to avoid

First, using a \(\lambda\) that does not match the interval - scale the rate before computing. Second, treating the count as bounded: the Poisson has no upper limit, so very large counts still carry small probabilities. Third, assuming the variance differs from the mean - they are equal, and data that break this pattern warn against the model. Fourth, applying Poisson to events that cluster or encourage each other: independence of occurrences is a condition, not a bonus. Fifth, forgetting \(0! = 1\) when computing \(P(X = 0)\). Sixth, reaching for the binomial-to-Poisson approximation when \(p\) is not small.

Key takeaways

The Poisson distribution counts events in a fixed interval of time or space at a known average rate \(\lambda\), with occurrences independent of each other and of the time since the last event. \(P(X = k) = \frac{e^{-\lambda}\lambda^k}{k!}\); the mean and the variance both equal \(\lambda\); and the count has no upper limit. The rate must match the interval, so \(\lambda\) scales with the window. For large \(n\) and small \(p\), the binomial approaches a Poisson with \(\lambda = np\) - the rare-events limit that ties this topic back to the binomial.

Flashcards

Flashcard

Poisson setting

Events per fixed interval at a known average rate

Counts with no upper limit.

Test your understanding

Question 1 of 5

Which scenario is a Poisson setting?

Question 2 of 5

A chapter averages λ = 2 typos. What is P(no typos)?

Question 3 of 5

A Poisson variable has λ = 3. Its mean and variance are:

Question 4 of 5

A hotline averages 3 calls per hour. What is λ for a 2-hour shift?

Question 5 of 5

At λ = 2 typos per chapter, what is P(at least one typo)?

← Previous topicNext topic →