Poisson Processes

A Poisson process is a model for events occurring randomly and independently in time (or space) at a constant average rate $λ$ . The number of events in any interval of length $t$ follows a Poisson distribution with mean $λ t$ , and the waiting times between events are independent exponential random variables. It applies wherever rare, memoryless events accumulate — radioactive decay, customer arrivals, photon detection, telephone calls. Introduced via Poisson's 1837 distribution and developed into a stochastic process by the early 20th century (Erlang, Lundberg).

A Poisson process ${N (t) : t \geq 0}$ is a counting process in which events arrive one at a time, independently, with a constant rate. Formally, $N (t)$ counts events in $[0, t]$ ; increments over disjoint intervals are independent, and the count over any interval of length $τ$ is Poisson-distributed with mean $λ τ$ . It is the canonical model of "complete randomness in time" and is foundational to queueing theory, reliability, insurance mathematics, and the theory of point processes. A surprising fact: if you observe a fixed number of events on $[0, T]$ , their arrival times are distributed exactly like an ordered uniform random sample — the process "forgets" its rate and looks like scattered points.

The Core Insight

Memorylessness made into a process. The defining property is that the future evolution of the process depends only on the present, not the past — and the rate never changes. This single assumption (stationary, independent increments with no simultaneous events) forces both the Poisson count distribution and the exponential inter-arrival times. The Poisson process is the unique continuous-time analogue of repeated independent coin flips made infinitely frequent and infinitely rare in just the right balance.

Formal Statement

Counting Process

A stochastic process ${N (t) : t \geq 0}$ is a counting process if $N (t)$ represents the total number of events in $[0, t]$ and satisfies:

$N (t) \geq 0$ and $N (0) = 0$ ,
$N (t)$ is integer-valued,
if $s < t$ then $N (s) \leq N (t)$ ,
for $s < t$ , $N (t) - N (s)$ equals the number of events in $(s, t]$ .

Poisson Process — Axiomatic Definition

A counting process ${N (t)}$ is a Poisson process with rate $λ > 0$ if:

$N (0) = 0$ ;
Independent increments: the numbers of events in disjoint intervals are independent;
Stationary increments: the distribution of the number of events in an interval depends only on its length;
The number of events in any interval of length $t$ is Poisson distributed:

P {N (t + s) - N (s) = n} = e^{- λ t} \frac{( λ t ) ^{n}}{n !}, n = 0, 1, 2, \dots

⟳This is the Poisson pmf with parameter

λ t

; it follows from the infinitesimal axioms below via a differential equation, shown in the proof.

Here $λ$ is the rate or intensity (events per unit time), $t$ is the interval length, and $n$ the event count.

Poisson Process — Infinitesimal (Local) Definition

An equivalent definition replaces axiom 4 with local conditions. As $h \to 0$ :

P {N (h) = 1} = λh + o (h), P {N (h) \geq 2} = o (h), P {N (h) = 0} = 1 - λh + o (h)

⟳

o (h)

denotes terms vanishing faster than

h

; these encode "at most one event in a tiny window" — orderliness — plus a constant per-unit hazard.

Equivalence of Definitions

The axiomatic definition and the infinitesimal definition are equivalent: a counting process with $N (0) = 0$ , stationary independent increments, and the orderliness conditions [2] satisfies [1], and conversely.

Inter-arrival and Waiting Times

Let $T_{1}, T_{2}, \dots$ be the inter-arrival times (time between consecutive events) and $S_{n} = \sum_{k = 1}^{n} T_{k}$ the arrival time of the $n$ -th event. Then:

T_{n} \sim iid Exponential (λ), P {T_{n} > t} = e^{- λ t}

⟳

{T_{1} > t} = {N (t) = 0}

, which has probability

e^{- λ t}

by [1]; independence and stationarity extend this to all

T_{n}

S_{n} \sim Gamma (n, λ), f_{S_{n}} (t) = λ e^{- λ t} \frac{( λ t ) ^{n - 1}}{( n - 1 )!}, t \geq 0

⟳

S_{n}

is a sum of

n

iid exponentials; convolve, or differentiate

P {S_{n} \leq t} = P {N (t) \geq n}

using [1].

Why It Holds / Derivation

Deriving the Poisson Distribution from the Infinitesimal Axioms

Let $P_{n} (t) = P {N (t) = n}$ . We derive [1] from [2].

Step 1 — Probability of zero events. Using independent and stationary increments,

P_{0} (t + h) = P {N (t) = 0, N (t + h) - N (t) = 0} = P_{0} (t) P_{0} (h) .

By [2],

P_{0} (h) = 1 - λh + o (h)

, so

P_{0} (t + h) - P_{0} (t) = P_{0} (t) (- λh + o (h)) .

Step 2 — Differential equation for $P_{0}$ . Divide by

h

and let

h \to 0

P_{0}^{'} (t) = - λ P_{0} (t) .

With

P_{0} (0) = 1

this gives

P_{0} (t) = e^{- λ t}

. Step 3 — Recursion for $n \geq 1$ . Condition on events in

(t, t + h]

. Using orderliness (at most one event in

h

P_{n} (t + h) = P_{n} (t) (1 - λh) + P_{n - 1} (t) λh + o (h) .

Rearranging and taking

h \to 0

P_{n}^{'} (t) = - λ P_{n} (t) + λ P_{n - 1} (t) .

Step 4 — Solve the recursion. Multiply by the integrating factor

e^{λ t}

\frac{d}{d t} [e^{λ t} P_{n} (t)] = λ e^{λ t} P_{n - 1} (t) .

Inductively, assuming

P_{n - 1} (t) = e^{- λ t} (λ t)^{n - 1} / (n - 1)!

, integration with

P_{n} (0) = 0

yields

e^{λ t} P_{n} (t) = \frac{( λ t ) ^{n}}{n !},

hence

P_{n} (t) = e^{- λ t} (λ t)^{n} / n!

, which is [1].

■

Inter-arrival Times are Exponential

The first arrival exceeds $t$ iff no event occurred in $[0, t]$ :

P {T_{1} > t} = P {N (t) = 0} = e^{- λ t} .

Thus

T_{1} \sim Exponential (λ)

. For

T_{2}

, condition on the first arrival at

s

. By independent and stationary increments, the process restarts at time

s

(the strong Markov / regeneration property), so the time to the next event has the same exponential law and is independent of

T_{1}

. Inductively all

T_{n}

are iid Exponential

(λ)

■

Siméon Denis Poisson | 1781–1840

Poisson was a French mathematician and physicist whose 1837 treatise Recherches sur la probabilité des jugements en matière criminelle et en matière civile introduced the distribution now bearing his name, as a limiting form of the binomial for rare events. Although he never studied the process (the temporal stochastic object), his distribution describing the count of rare events became its defining marginal law. Poisson also contributed Poisson's equation in potential theory, the Poisson bracket in mechanics, and Poisson's ratio in elasticity — a remarkable spread across pure and applied mathematics.

Building Intuition

Imagine standing under a light drizzle and counting raindrops hitting a small square of pavement. Drops fall independently; there's no "memory" — the chance of a hit in the next instant doesn't depend on how long you've waited. Over a fixed second, the average number of hits is constant. This is the Poisson process: the count over any time window is Poisson, and the gaps between drops are exponential.

Two facts crystallize the picture:

Memorylessness: $P {T > s + t ∣ T > s} = P {T > t}$ . Having waited 5 minutes for a bus with no arrival, your additional wait has the same distribution as a fresh start. The process never "ages."
Rate balance: the binomial $(n, λ t / n)$ converges to Poisson $(λ t)$ as $n \to \infty$ . Slice $[0, t]$ into $n$ tiny bins, each holding an event with probability $λ t / n$ ; infinitely many infinitely-rare trials give the Poisson law.

A Poisson process: arrival times $S_n$ (gold dots) and the staircase count $N(t)$, with exponential gaps $T_n$.

Worked Examples

Call Center Arrivals

Given: Calls arrive at a switchboard as a Poisson process with rate

λ = 3

calls per minute. (a) Probability that exactly 5 calls arrive in a 2-minute window. Mean

= λ t = 3 \times 2 = 6

. By [1]:

P {N (2) = 5} = e^{- 6} \frac{6 ^{5}}{5 !} = e^{- 6} \frac{7776}{120} = e^{- 6} (64.8) \approx 0.1606.

(b) Probability that no calls arrive in the next 40 seconds (

= 2/3

min).

λ t = 3 \times \frac{2}{3} = 2

. So

P {N = 0} = e^{- 2} \approx 0.1353.

(c) Expected waiting time until the next call:

E [T] = 1/ λ = 1/3

min

= 20

seconds.

P (a) \approx 0.161, P (b) \approx 0.135, E [T] \approx 20 s

Radioactive Decay Detection

Given: A detector registers decay events from a sample at rate

λ = 0.5

counts per second. Assume a Poisson process. (a) Expected time for the 4th count:

E [S_{4}] = n / λ = 4/0.5 = 8

s (from [4], mean of Gamma is

n / λ

). (b) Variance of the time to the 4th count:

Var (S_{4}) = n / λ^{2} = 4/0.25 = 16

s², standard deviation

4

s. (c) Probability the 1st count takes longer than 3 s:

P {S_{1} > 3} = e^{- λ \cdot 3} = e^{- 1.5} \approx 0.2231.

(d) Probability of exactly 2 counts in a 10-second window:

λ t = 5

, so

P {N (10) = 2} = e^{- 5} \frac{5 ^{2}}{2 !} = e^{- 5} (12.5) \approx 0.0842.

E [S_{4}] = 8 s, σ_{S_{4}} = 4 s, P (S_{1} > 3) \approx 0.223, P (N = 2) \approx 0.084

Common Misconceptions

"Stationary increments mean evenly spaced events." No. Stationarity means the distribution of counts depends only on interval length, not on where the interval sits. The actual events are highly irregular — clusters and long gaps are typical, not anomalies. People routinely mistake genuine Poisson randomness for "non-random" because real Poisson data looks clumpier than the false intuition of regular spacing. Confusing the rate $λ$ with a probability.

λ

has units of events per unit time and can exceed 1. The quantity

λh

is only approximately a probability for small

h

; over a unit interval the probability of an event is

1 - e^{- λ}

, not

λ

. A process that fails orderliness. Suppose events arrive in pairs — every "trigger" deposits exactly two events simultaneously, with triggers forming a Poisson process. Then

P {N (h) = 2} \neq = o (h)

(it is

\approx λh

), violating axiom [2]. The count process is not Poisson; it is a compound Poisson process. Its variance exceeds its mean (

Var = 4 λ t > 2 λ t = E [N]

), whereas a true Poisson process always has equal mean and variance,

Var (N (t)) = E [N (t)] = λ t

Connections

The Poisson process sits at a crossroads of probability:

It is a continuous-time Markov Chain (a pure birth process) — the simplest non-trivial example, with generator concentrated on $+ 1$ jumps.
Its inter-arrival times link it to the Exponential Distribution and, via sums, the Gamma Distribution.
It is the prototypical Lévy process: stationary, independent increments — the discrete-jump counterpart of Brownian Motion.
As a model of random points, it is the foundational example in the theory of point processes and spatial statistics.
The Poisson limit theorem connects it to the Binomial Distribution (rare-event limit) and underpins its derivation.

Real-World Applications

Queueing theory / telecommunications: Erlang modeled telephone call arrivals as Poisson to size switchboards; the M/M/1 queue assumes Poisson arrivals and exponential service.
Insurance & risk theory: claim arrivals in the Cramér–Lundberg model are Poisson; ruin probabilities depend on the compound Poisson claim process.
Nuclear physics: radioactive decay counts (Geiger–Müller tubes) and photon arrivals in low-light detection are Poisson, with shot noise as the signature.
Neuroscience: spike trains of neurons are often modeled as (inhomogeneous) Poisson processes to a first approximation.
Reliability engineering: failures of components with constant hazard rate occur as a Poisson process; the exponential lifetime is its inter-arrival time.
Web/server traffic: page requests and packet arrivals are modeled as Poisson for capacity planning and load testing.
Genomics & spatial ecology: mutation locations along DNA and tree positions in a forest are modeled by spatial Poisson processes.

Historical Development

1711 — Abraham de Moivre studies rare-event approximations of binomial probabilities, foreshadowing the Poisson limit.
1837 — Siméon Denis Poisson publishes the Poisson distribution as a limit of the binomial in his work on judicial probabilities.
1898 — Ladislaus Bortkiewicz applies the distribution to rare events (famously, Prussian soldiers killed by horse kicks) in Das Gesetz der kleinen Zahlen.
1909 — Agner Krarup Erlang models telephone traffic as Poisson arrivals, founding queueing theory.
1903–1910 — Harald Cramér and Filip Lundberg develop collective risk theory using the Poisson claim process.
1910 — Ernest Rutherford and Hans Geiger empirically confirm Poisson statistics for alpha-particle emission.
1930s–40s — Andrey Kolmogorov and others formalize the Poisson process within the axiomatic theory of stochastic processes and Markov processes.
1955 — David Cox introduces doubly stochastic (Cox) processes, generalizing the Poisson process to random intensities.
1960s–70s — The general theory of point processes (Khinchin, Matthes, Mecke) places the Poisson process as the canonical "complete randomness" reference.

Subtleties & Limits

Beyond the homogeneous case. Real-world arrival rates vary with time (traffic peaks, daily cycles). The inhomogeneous Poisson process replaces

λ

with

λ (t)

, so counts become Poisson with mean

\int_{s}^{t} λ (u) d u

. Estimating

λ (t)

nonparametrically, and testing whether data are genuinely Poisson versus clustered (overdispersed) or regular (underdispersed), remain active in applied statistics. The Cox process lets

λ

itself be random, capturing burstiness; distinguishing a Cox process from a pure inhomogeneous Poisson process from a single realization is notoriously hard. Higher-dimensional analogues (spatial and spatio-temporal Poisson processes) raise deep questions in stochastic geometry, including the behavior of Poisson–Voronoi tessellations and percolation on Poisson point clouds.