Chapter 2 — Poisson Processes
This page develops the mathematics of Poisson processes and keeps the mathematical discussion separate from the Python API and worked examples.
1. Introduction
A Poisson process is introduced as a continuous-time process with non-negative integer values, used to describe the times at which random events occur: calls, customer arrivals, failures, particle emissions, and similar phenomena.
2. Counting process
Let \(N(t)\) be the number of occurrences observed in \(]0,t]\). A process \((N(t))_{t\ge0}\) is a counting process if
and if \(N(t)-N(s)\) represents the number of occurrences on \(]s,t]\).
3. First definition of the Poisson process
A counting process is a Poisson process with rate \(\lambda>0\) if:
- \(N(0)=0\) almost surely;
- the increments are independent;
- the increments are stationary;
- \(P(N(h)=1)=\lambda h+o(h)\);
- \(P(N(h)\ge2)=o(h)\).
Here \(o(h)/h\to0\) as \(h\to0\).
4. Second definition
An equivalent formulation is: \(N(0)=0\), the increments are independent, and for every interval length \(t\),
This law does not depend on \(s\), which gives stationary increments.
5. Equivalence of the definitions
Let
The analysis over \([t,t+h]\) gives
and, for \(n\ge1\),
Solving these equations gives
Therefore \(N(t)\sim\mathcal P(\lambda t)\) and
6. Inter-occurrence times
If
are the occurrence times, the inter-arrival times
form an i.i.d. sequence with exponential distribution of parameter \(\lambda\):
7. Conditional occurrence times
Conditionally on \(N(s)=1\), the time of the unique occurrence in \([0,s]\) is uniform:
More generally, conditionally on \(N(s)=k\), the \(k\) occurrence times in \([0,s]\) have the distribution of the order statistics of a sample of \(k\) independent uniform random variables on \([0,s]\).
8. Superposition
If \(N^{(1)}\) and \(N^{(2)}\) are independent Poisson processes with rates \(\lambda_1\) and \(\lambda_2\), then
is again a Poisson process, with rate
9. Splitting / thinning
Starting from a Poisson process of rate \(\lambda\), assign each occurrence independently to type 1 with probability \(p\), or to type 2 with probability \(1-p\). The two resulting processes are independent and have rates
10. Non-homogeneous Poisson process
In the non-homogeneous case, the rate depends on time: \(\lambda=\lambda(t)\). Increments are therefore no longer stationary. Introduce the cumulative mean function
Locally,
This model is suited to phenomena whose occurrence rate varies over time.
11. Key results
Increments over disjoint intervals are independent; in the homogeneous case, their law depends only on the interval lengths. Conditionally on a fixed number of occurrences, the occurrence times are distributed as uniform order statistics.