Poisson Distribution

Model random events in time and compare the Poisson with a binomial that has many trials

Intermediate

Poisson PMF, λ = 4.0(tap a bar to pick k)

0.000.050.100.150.2002468101214k (number of events)
X = kX < kX > k
Probabilities
P(X = 4)
0.19537
P(X ≤ 4)
0.62884
P(X > 4)
0.37116
Mean = variance = λ
4.0σ = 2.000

P(X = 4) = e−4.0 · 4.04 / 4!

Random arrivals in time

Events arrive one at a time with exponential gaps of mean 1/λ = 0.25. The count in each unit interval is a draw from Poisson(λ). The histogram below collects those counts.

Speed

Counts per interval(blue: observed share, red: Poisson PMF)

0.000.050.100.150.2002468101214k (number of events)
Observed vs theory
Intervals counted
0
Mean count
–λ = 4.0
Variance of counts
–λ = 4.0

Binomial(n, λ/n) → Poisson(λ)

Split the interval into n slots, each holding an event with probability λ/n. The count is Binomial(n, λ/n), and as n grows it converges to Poisson(λ).

Slots n
0.000.100.2002468101214k (number of events)
Binomial(10, 0.400)Poisson(4.0)
How close?
Total variation distance
0.1267bound λ²/n = 1.600
Binomial variance n·p(1 − p)
2.400Poisson: 4.0
ndistance
50.3951
100.1267
200.0560
500.0209
1000.0102
10000.0010

When counts are Poisson

The Poisson distribution models how many times an event happens in a fixed window when events occur independently at a steady average rate λ. Unlike the binomial, there is no fixed number of trials and no upper limit: X can be 0, 1, 2, 3, … without end, although large counts quickly become very unlikely.

The link to the binomial explains where the formula comes from. Cut the window into n tiny slots, each so short that it holds at most one event, with probability λ/n. The count is then Binomial(n, λ/n). Letting n → ∞ gives the Poisson PMF, because C(n, k)(λ/n)ᵏ → λᵏ/k! and (1 − λ/n)ⁿ⁻ᵏ → e−λ. That is why the Poisson is also called the law of rare events.

Seen in time, the same process has a second description: the gaps between events are independent exponential random variables with mean 1/λ. The timeline in the visualizer generates arrivals that way and then counts them interval by interval.

Counts in a fixed window

X counts events in a fixed stretch of time, length, area or volume: emails per hour, typos per page, raindrops per tile.

One parameter, λ

λ is the average number of events per window. It is both the mean and the variance of the count, so the spread grows like √λ.

Independent, one at a time

Events happen independently at a constant average rate, and two events never land at exactly the same instant.

Rescaling the window

If a 1-hour window has rate λ, a t-hour window has rate λt. Counts in non-overlapping windows are independent.

Key formulas

Probability mass function

P(X = k) = e−λ λk / k!

λ
average number of events per window (λ > 0)
k
number of events, 0, 1, 2, …

Mean, variance and cumulative probability

E[X] = λ,   Var(X) = λ,   P(X ≤ k) = e−λ Σi=0k λi / i!

Poisson approximation to the binomial

Binomial(n, p) ≈ Poisson(np) when n is large and p is small;  dTV ≤ np²

d_TV
total variation distance: the largest possible difference in the probability of any event

Waiting times

gap between events ~ Exponential(λ),   P(no event in time t) = e−λt

Typical Poisson counts

SituationWindowTypical λ
Calls to a help lineper hour≈ 4
Radioactive decays detected by a Geiger counterper seconddepends on the source, often 1 to 100
Goals in a soccer matchper team per match≈ 1.3
Typos in a bookper page≈ 0.2
Meteors in a meteor showerper minute≈ 1

How to use this visualizer

  1. 1
    Set the rate λ. The PMF updates, and the side panel shows that the mean and variance are both λ.
  2. 2
    Pick a count k with the slider or by tapping a bar. The panel shows P(X = k), the cumulative P(X ≤ k) and the tail P(X > k).
  3. 3
    Press Play in the arrivals section. Orange dots are events; when the red cursor crosses an interval boundary, that interval’s count is added to the histogram.
  4. 4
    Press Add 1000 intervals and compare the blue bars with the red PMF lines, and the sample mean and variance with λ.
  5. 5
    In the last section, step n from small to large. The binomial bars close in on the Poisson bars and the total variation distance shrinks roughly like 1/n.

Common misconception

Random does not mean evenly spaced. In a Poisson process events cluster and leave long gaps purely by chance, so seeing three calls in a minute and then none for ten minutes is not evidence that something changed. Over-dispersed data, where the variance is clearly larger than the mean, is the real sign that a plain Poisson model does not fit.

Practice problems

Problem 1: Calls to a help line

A help line receives on average 4 calls per hour, and calls arrive as a Poisson process. Find the probability of exactly 2 calls in an hour, of at most 2 calls in an hour, and of no calls at all in a 30-minute break.
Show solution

X ~ Poisson(λ = 4) for one hour

P(X = 2) = e⁻⁴ · 4² / 2! = 8e⁻⁴ = 8 × 0.018316 ≈ 0.1465

P(X ≤ 2) = e⁻⁴ (1 + 4 + 8) = 13e⁻⁴ ≈ 0.2381

For 30 minutes the rate is λt = 4 × 0.5 = 2

P(no calls) = e⁻² ≈ 0.1353

0.1465, 0.2381 and 0.1353

Problem 2: Defective items

A machine makes defective items with probability 1/500, independently. In a batch of 1000 items, use the Poisson approximation to find the probability of at most 1 defective item, and compare it with the exact binomial answer.
Show solution

Exact model: X ~ Binomial(1000, 0.002), so λ = np = 2

Poisson: P(X ≤ 1) ≈ e⁻² (1 + 2) = 3e⁻² ≈ 0.40601

Exact: P(X = 0) = 0.998¹⁰⁰⁰ ≈ 0.13506

P(X = 1) = 1000 × 0.002 × 0.998⁹⁹⁹ ≈ 0.27067

P(X ≤ 1) ≈ 0.40573

Difference ≈ 0.0003, well within the bound np² = 1000 × 0.002² = 0.004

Poisson 0.4060 vs exact 0.4057

Related topics

Discussion

Questions, ideas or a different approach? Join the conversation.