Statistics › 🧠 Advanced & Bayesian Modeling

The Poisson Process, Gamma Waiting Times & Queues

One rate λ governs how many rare events arrive — and how long you wait between them.

AdvancedUpper-undergraduate
💡
The big idea: A Poisson process is the mathematics of rare, independent events sprinkled along time: calls hitting a switchboard, customers joining a queue, decays in a detector. Count them in a fixed window and you get the Poisson distribution, whose mean and variance are the same number λ. Look instead at the gaps between events and you get the Exponential; add several gaps and you get the Gamma. Together they are the engine behind queuing theory.
🎯 By the end, you'll be able to
  • State the assumptions that define a Poisson process and write down its probability mass function
  • Explain why a Poisson distribution has equal mean and variance, both λ
  • Recognize the Poisson as the limiting case of a Binomial with n large, p small, and np fixed
  • Connect counts to waiting times: Exponential gaps between events, Gamma for the time to the k-th event
  • Read a simple single-server queue in terms of an arrival rate λ and a service rate μ
📎 You should already know
  • The Binomial distribution
  • Expectation and variance
  • Probability density functions (pdf and cdf)
  • Basic calculus with exponentials and factorials

Counting rare, independent events

Many real processes are streams of rare events that arrive independently: phone calls hitting a switchboard, customers joining a checkout line, radioactive decays in a second, typos on a page, mutations along a strand of DNA. You never know when the next one lands, but over a long window they arrive at some steady average rate λ.

The Poisson distribution answers one question about such a stream: in a single fixed interval, how many events will I see? The answer is a probability for each count k = 0, 1, 2, … — and remarkably, that entire distribution is pinned down by the one number λ.

🔑 What makes a process 'Poisson'
Three assumptions define a Poisson process at rate λ: events in non-overlapping intervals are independent; the rate stays constant over time; and events happen one at a time (no simultaneous pile-ups). Under these rules, the number of events in a window of length t is Poisson-distributed with mean λt — so only the expected count is needed to fix the whole distribution.
\[ P(X=k) = \frac{e^{-\lambda}\,\lambda^{k}}{k!}, \qquad k = 0, 1, 2, \dots \]
The Poisson probability of exactly k events when the expected count is λ. The factor e^{-λ} normalizes the λ^k/k! terms so the probabilities sum to 1.

Center and spread are the same number

Compute the expectation and the variance of that PMF — both are short sums that telescope thanks to the k! in the denominator — and something unusual happens: they come out equal. The mean, the variance, and therefore the whole shape are all controlled by λ alone.

\[ \mathbb{E}[X] = \lambda, \qquad \operatorname{Var}(X) = \lambda, \qquad \operatorname{SD}(X) = \sqrt{\lambda} \]
Mean and variance both equal λ, so the standard deviation is √λ. Doubling the rate doubles the variance but multiplies the spread by only √2.

Watch the mass shift

The tool below plots the exact Poisson PMF as bars over the possible counts k. Drag the rate λ and watch three things move together: the peak slides toward larger counts, the bars spread out, and the shaded ±1 SD band (half-width √λ) widens. At small λ the distribution is lopsided and piled against 0; as λ grows it turns more symmetric and starts to resemble a bell — a preview of the normal approximation for large λ.

🎮 Interactive: the Poisson PMF LIVE
Bars show P(X=k) = e^{-λ} λ^k / k! for each count k. The vertical line marks the mean λ; the shaded band is mean ± 1 SD = λ ± √λ. The center and the width grow together, because the mean and the variance are both λ.
✨ Mean = variance is the Poisson fingerprint
For most distributions the center and the spread are independent dials. The Poisson welds them together: Var(X) = E[X] = λ. In the tool, raising λ pushes the peak right and widens the band at the same time. This gives a quick diagnostic for real data: if a count variable has variance much larger than its mean (overdispersion), a plain Poisson model is too tight, and richer models such as the negative binomial tend to fit better.

Where Poisson comes from: the Binomial limit

Why should a single number control everything? Picture chopping the interval into n tiny sub-slots, each so short it holds at most one event, with a small success probability p in each. The total count is then Binomial(n, p). Now let the slots get finer — n → ∞ and p → 0 — while holding the expected count np = λ fixed. In that limit the Binomial PMF collapses exactly onto the Poisson:

\[ \binom{n}{k}\, p^{k}(1-p)^{\,n-k} \;\xrightarrow[\; np = \lambda \;]{\; n \to \infty,\; p \to 0 \;}\; \frac{e^{-\lambda}\,\lambda^{k}}{k!} \]
As the slots become infinitely fine with np held at λ, the Binomial becomes Poisson. This is why the Poisson is the natural law for rare events over many opportunities — sometimes called the 'law of small numbers.'

From counts to waiting times: Exponential and Gamma

Counting events in a window is one view; timing the gaps between them is the dual view. In a Poisson process the wait for the next event is Exponential with rate λ — its density decays as λe^{-λt}, and its mean gap is 1/λ. The Exponential is memoryless: having already waited five minutes tells you nothing about how much longer you will wait, which is exactly the 'no memory' character of independent events.

Add several gaps and you get the Gamma distribution (with an integer shape it is also called the Erlang): the time until the k-th event is a sum of k independent Exponential gaps, with mean k/λ. So one process wears three faces — Poisson for counts, Exponential for single gaps, and Gamma for cumulative waits.

These pieces snap together into queuing theory. The simplest single-server queue feeds in arrivals as a Poisson process at rate λ and lets the server finish jobs at an Exponential service rate μ. The ratio ρ = λ/μ is the utilization: when ρ < 1 the server keeps up and the line stays finite on average, but as ρ approaches 1 the expected wait blows up. That single inequality — arrivals slower than service — is the heartbeat of capacity planning for call centers, checkout lines, and servers.

\[ \underbrace{\lambda e^{-\lambda t}}_{\text{gap: Exponential}} \qquad \text{and} \qquad \underbrace{\frac{\lambda^{k}\,t^{\,k-1}\,e^{-\lambda t}}{(k-1)!}}_{\text{wait to } k\text{-th event: Gamma}}, \qquad t \ge 0 \]
Left: the Exponential density of one gap (mean 1/λ). Right: the Gamma/Erlang density of the wait until the k-th event (mean k/λ), a sum of k Exponential gaps.
📝 Worked example: A help desk receives support tickets independently at an average rate of λ = 2 per hour, well modeled as a Poisson process. What can you say about the counts and the waits?
  1. Center and spread of the hourly count are both λ = 2, so the standard deviation is \( \sqrt{2} \approx 1.41 \) tickets per hour.
  2. A fully quiet hour (zero tickets): \( P(X=0) = e^{-2}\dfrac{2^{0}}{0!} = e^{-2} \approx 0.135 \) — about a 13.5% chance.
  3. Exactly three tickets in an hour: \( P(X=3) = e^{-2}\dfrac{2^{3}}{3!} = e^{-2}\dfrac{8}{6} \approx 0.180 \).
  4. The gap to the next ticket is Exponential with mean \( 1/\lambda = 1/2 \) hour — a typical 30-minute wait between tickets.
✓ The hourly count has mean and SD of 2 and ≈ 1.41; a fully quiet hour occurs about 13.5% of the time, exactly three tickets about 18%, and the average gap between tickets is 30 minutes.

Check your understanding

1. For a Poisson distribution with rate λ, the variance is…
The Poisson fingerprint is Var(X) = E[X] = λ; the standard deviation is √λ.
2. The Poisson distribution arises as the limiting case of a Binomial(n, p) when…
Fine slots (n large), a rare success in each (p small), and a fixed expected count np = λ — the 'law of small numbers.'
3. Events follow a Poisson process with rate λ per hour. The time until the next event follows which distribution, with what mean?
Gaps between consecutive Poisson events are Exponential with mean 1/λ; the memoryless property mirrors the independence of the events.
4. In a Poisson process with λ = 1 event per interval, the probability of exactly 2 events in one interval is…
P(X=2) = e^{-1}·1^2/2! = e^{-1}/2 ≈ 0.184.
✅ Key takeaways
  • A Poisson process counts rare, independent events in a fixed interval; the count follows P(X=k) = e^{-λ} λ^k / k!.
  • Its signature is mean = variance = λ (so SD = √λ) — the center and the spread grow together with the rate.
  • It is the limit of a Binomial when n is large, p is small, and np = λ stays fixed.
  • The gaps between consecutive events are Exponential (mean 1/λ); the time until the k-th event is Gamma/Erlang (mean k/λ).
  • A simple single-server queue pairs a Poisson arrival rate λ with an Exponential service rate μ; it stays stable when λ < μ (utilization ρ = λ/μ < 1).