Poisson Process

Siméon Poisson published the distribution in 1837 inside Recherches sur la probabilité des jugements, a treatise trying to model wrongful convictions in French criminal courts. He never named it after himself. The trial-data application was a footnote; the machinery turned out to describe almost anything that arrives one-at-a-time without coordinating. One parameter, λ, rate per unit time. That is the whole model.

How it works

A counting process N(t) is Poisson with rate λ if three things hold:

N(0) = 0. The clock starts empty.
Independent increments. What happened in (0, s] tells you nothing about what happens in (s, t].
The probability of one event in a tiny window of width h is λh + o(h). Two events in the same tiny window is o(h) — vanishingly rare.

From those axioms, two facts fall out:

Counts are Poisson. P(N(t) = k) = e^(-λt) (λt)^k / k!. Mean and variance both equal λt. The collapse of mean and variance into one number is what makes the distribution easy to fit and easy to falsify.
Inter-arrival times are exponential. Gaps between events are independent, each with mean 1/λ. This is the memoryless property: a customer who has been queuing five minutes is no closer to being next than one who just arrived.

The deep claim sitting underneath all this: if you superpose many independent rare-event streams, the limit is Poisson. This is the Palm-Khintchine theorem, the arrival-process cousin of the central limit theorem.

Where it shows up

Domain	What arrives	Typical λ
Call centres	Inbound calls	100-3,000 per hour, time-varying
Retail footfall	Shoppers entering a store	20-200 per 15-min window
Particle physics	Photons hitting a CCD pixel	depends on flux, sub-second
Insurance	Claims arriving at a desk	a few per day per book
Internet	Packet arrivals at a router	millions per second
Genomics	Mutations along a DNA strand	10^-9 per base per generation
Astronomy	Gamma-ray bursts in the sky	~1 per day, all-sky

Bortkiewicz's 1898 study of Prussian cavalry officers killed by horse kicks is the textbook example. Across 14 corps over 20 years, 122 deaths. Counts per corps-year fit Poisson with λ ≈ 0.61 almost perfectly. The point is not the macabre subject. The point is that an event with no plausible mechanism for clustering, sampled across many independent units, looks Poisson.

In retail operations this is load-bearing. Footfall in 15-minute windows is the input to staffing models, queueing approximations, and conversion-rate denominators. If arrivals are Poisson with rate λ(t) varying through the day, an M/M/c queue gives you closed-form expected wait times. Get the rate wrong and the staffing model is wrong before anyone has done arithmetic. This is one place where the concept bayesian inference update — treating yesterday's λ as a prior, today's count as evidence — pays for itself.

What's contested

The independence assumption is the crack in the foundation. Real arrivals cluster. Calls arrive in bursts after a TV ad airs. Shoppers arrive in waves when a mall food court empties. Photon arrivals at low intensity are Poisson, but coherent light shows bunching (Hanbury Brown-Twiss, 1956) and antibunching for single emitters. Each of these violates independent increments.

The field has three responses:

Non-homogeneous Poisson process (NHPP). Let λ vary with time, λ(t). Fits diurnal patterns. Counts in (s, t] are Poisson with mean ∫λ(u)du. Still assumes independence — just relaxes constant rate.
Cox process (doubly stochastic). λ itself is random, drawn from another process. Captures over-dispersion: variance exceeds mean. This is what call centres actually do.
Hawkes process (self-exciting). Each event raises the rate for a short window afterward. Models bursts: earthquake aftershocks, viral retweets, market microstructure. The Hawkes-1971 paper is the canonical reference and the model has become the workhorse for limit-order-book research since roughly 2010.

A 2011 paper by Brown et al. on call-centre arrivals found that the homogeneous Poisson model is rejected at almost every level of aggregation finer than hourly. NHPP with a smooth λ(t) survives. This matters: many textbook queueing results assume homogeneous Poisson, and the staffing recommendations they produce are systematically wrong by 10-30% during peaks.

The other open question is goodness-of-fit at the tail. Poisson tails decay roughly as e^(-x log x). Real-world tails — financial losses, network traffic, earthquake magnitudes — are often power-law. Mistaking a fat-tailed process for Poisson means under-pricing rare disasters. This is the concept fat tails critique, and it is the reason actuaries pair Poisson frequency models with non-Poisson severity distributions.

Why this has to do with other realms

The exponential inter-arrival time is the same memoryless distribution that governs radioactive decay. A single uranium-238 atom has no age — its probability of decaying in the next second is identical at t=0 and at t=4.5 billion years. This is why nuclear physicists and queueing theorists share equations. The same axiom — no memory — yields the same maths.

The connection runs further. The concept shannon information of a Poisson process per unit time is λ log(e/λ) bits when properly normalised; the rate parameter is also the information rate of the underlying source. And in evolutionary biology, the concept molecular clock uses Poisson statistics on neutral mutations to date species splits — the same machinery Poisson built for trial verdicts now dates the human-chimp divergence to roughly 6 million years ago. One distribution, four sciences, one parameter each.

An open question

If real arrival processes are routinely non-Poisson at the fine grain (clustered, self-exciting, doubly-stochastic), why does the Poisson approximation work so well at coarse aggregation? The Palm-Khintchine theorem says superposition smooths things out. But the conditions it requires — many sparse independent streams — are rarely audited in practice. When does aggregation fail to rescue the model, and how would we know before the queue collapses?

Key sources

Poisson, S.D. (1837). Recherches sur la probabilité des jugements en matière criminelle et en matière civile. The original — chapter 4 has the distribution.
Feller, W. (1968). An Introduction to Probability Theory and Its Applications, Vol. I, chapters 6 and 17. The pedagogical reference; nobody explains Poisson better.
Brown, L. et al. (2005). "Statistical Analysis of a Telephone Call Center." JASA 100(469). The canonical empirical study showing where homogeneous Poisson fails.
Hawkes, A.G. (1971). "Spectra of some self-exciting and mutually exciting point processes." Biometrika 58. The self-exciting generalisation; the foundation of modern microstructure work.
Daley, D.J. and Vere-Jones, D. (2003). An Introduction to the Theory of Point Processes. The graduate-level reference for everything beyond homogeneous Poisson.

Abhishek's take

The thing that grabbed me late was that Poisson built this to model wrongful convictions and the distribution barely fits courtrooms — verdicts cluster around individual judges, panels, news cycles. The machinery works best where it was an afterthought: photons, packets, footfall. The general lesson is that powerful mathematical objects often outgrow the question that produced them, and the inventor rarely sees where they will land. The contested layer matters operationally — assuming independence when arrivals self-excite is how staffing models fail at exactly the moments you needed them to hold.

Where I've used this

Retail footfall in 15-minute windows is the daily working substrate. The honest version is non-homogeneous Poisson with a λ(t) curve that bends around mall anchor-tenant timings and weather; the textbook M/M/c is a useful first approximation and a dangerous final one. Pairing it with a Bayesian update on yesterday's residuals is what keeps the staffing math from drifting.

Abhishek