Skip to content
Tutorials Center
My learning

Topics

Lesson 19 of 36 · Poisson Processes

The Poisson Process: Counting Random Events in Time

Learn what a Poisson process is, why counts in a time window follow a Poisson distribution, and check the idea with a binomial limit and a simulation.

Customers walk into a shop, emails reach your inbox, atoms decay in a sample, cars pass a sensor. Each of these is a stream of events that happen at random moments. The Poisson process is the most important mathematical model for such a stream. It is simple enough to solve exactly and realistic enough to be used everywhere, from telephone exchanges to insurance claims.

In this lesson you meet the definition, see why the Poisson distribution appears, and check it with code. Later lessons in this chapter build on it: waiting times, splitting and merging streams, and time-varying rates.

Counting processes

A counting process N(t) records how many events have happened in the time interval from 0 to t. It has three obvious properties:

  • N(0) = 0, because nothing has happened yet.
  • N(t) is a whole number.
  • N(t) never decreases as t grows.

An increment such as N(t) - N(s) (with s < t) is the number of events in the window (s, t].

Definition of a Poisson process

A counting process is a Poisson process with rate λ (λ > 0, measured in events per unit time) if:

  1. N(0) = 0.
  2. Independent increments: counts in disjoint time windows are independent random variables. What happened this morning tells you nothing about this afternoon.
  3. Stationary increments: the count in a window depends only on the window's length, not on where it starts.
  4. For a short window of length h, P(one event) ≈ λh, and P(two or more events) is negligible compared with h.

From these rules one can prove the key fact:

Text
N(t) ~ Poisson(λt)
P(N(t) = k) = e^(-λt) (λt)^k / k!

So the mean and the variance are both λt. If a help desk gets λ = 3 calls per hour, then in 2 hours you expect λt = 6 calls, with a standard deviation of √6 ≈ 2.4.

The Poisson distribution has equal mean and variance. If real count data has a variance much larger than the mean (called overdispersion), a plain Poisson process is probably too simple.

Why Poisson? The binomial limit

Cut the interval [0, t] into n tiny slices of length t/n. In each slice an event happens with probability p = λt/n, independently of the other slices (this is rule 4). The number of slices with an event is Binomial(n, p). As n grows with np = λt fixed, this binomial approaches Poisson(λt).

The next program computes both distributions from the formulas, without any library, and prints how far apart they are.

Python Binomial approaches Poisson

Press Run (or Ctrl + Enter) to see the result here.

Program input · what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The gap shrinks steadily as the slices get thinner. That is the intuition behind the name "law of rare events": many tiny chances, each very unlikely, add up to a Poisson count.

A simulation from the definition

You can build a Poisson process exactly the way the definition says: slice time, flip a biased coin per slice. This is slow compared with the clever methods of the next lesson, but it follows the rules literally.

Python Simulate by slicing time

Press Run (or Ctrl + Enter) to see the result here.

Program input · what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The mean and variance both land near 6, and the correlation between disjoint windows is near 0, as rule 2 predicts (here they are generated independently, so this shows what "independent" looks like in data).

Using the Poisson process in practice

Suppose a website receives 40 requests per minute on average (λ = 40 per minute). Then:

  • Requests in 15 seconds: N ~ Poisson(10).
  • P(no request in 3 seconds) = e^(-2) ≈ 0.135, since λt = 40 × 0.05 = 2.
  • Requests in two non-overlapping windows add up: counts in disjoint windows are independent Poisson variables, and the sum of independent Poissons is Poisson with the sum of the means.

When can you trust the model? Check three things: events arrive one at a time, there is no obvious clustering or schedule (such as buses on a timetable), and the rate does not change much during the observation. If the rate changes, move to a nonhomogeneous process.

flowchart LR
  A["Rare chance in each tiny slice"] --> B["Binomial count over many slices"]
  B --> C["Poisson(λt) count"]
  C --> D["Poisson process N(t)"]

From tiny independent chances to the Poisson process.

Check your understanding

Quick check · 4 questions

of 4 answered You scored / 4 · perfect!

  1. 1. A Poisson process has rate λ = 4 events per hour. What is the expected number of events in 30 minutes?

    The window is 0.5 hours, so the mean is λt = 4 × 0.5 = 2.
  2. 2. Which property says the count this morning gives no information about the count this afternoon?

    Independent increments means counts in disjoint windows are independent.
  3. 3. For a Poisson random variable with mean 9, what is the variance?

    A Poisson distribution has variance equal to its mean.
  4. 4. What is P(N(t) = 0) for a Poisson process with rate λ?

    Put k = 0 in the Poisson pmf: e^(-λt) (λt)^0 / 0! = e^(-λt).

Summary

  • A Poisson process with rate λ counts events with independent, stationary increments.
  • The count in a window of length t is Poisson(λt): mean and variance both λt.
  • It arises as the limit of many tiny independent chances (binomial to Poisson).
  • Check the assumptions: single events, no schedule, roughly constant rate.

Next: Interarrival and Arrival Times