Poisson processes work by modelling random events that arrive through continuous time at a stable average rate, with independent increments and exponentially distributed waiting times in the homogeneous case. If N(t) counts events by time t and the rate is λ, then N(t) has a Poisson distribution with mean λt, interarrival times are independent Exponential(λ) variables, and the memoryless property connects the count process to a simple continuous-time Markov chain. Poisson processes are foundational in queueing, reliability, telecommunications, insurance, epidemiology, operations research and stochastic simulation.
The key modelling idea is not merely that counts happen to look Poisson. A homogeneous Poisson process makes structural claims: events in disjoint time intervals are independent, expected count grows proportionally with interval length, and the probability of multiple arrivals in a tiny interval is negligible relative to the interval length. Those assumptions create the Poisson count law rather than being consequences that can be ignored.
Poisson processes connect two views of the same random system. The count view asks how many events occur by time t. The waiting-time view asks how long until the next event or the kth event. Moving between these views turns arrival data into rates, queueing models, reliability calculations and point-process reasoning.
1. N(t) Counts Arrivals Through Time
A counting process begins at zero and increases by integer jumps as events occur. In a simple Poisson process, jumps are almost surely one at a time.
2. Homogeneous Rate Means Expected Count Is Proportional to Time
If the rate is λ events per unit time, E[N(t)]=λt. Doubling the observation window doubles expected count under a stable homogeneous model.
3. Counts Are Poisson Distributed
N(t) has probability P(N(t)=k)=exp(−λt)(λt)^k/k!. Mean and variance both equal λt in the basic model.
4. Independent Increments Separate Disjoint Time Windows
Counts in non-overlapping intervals are independent. This is a strong modelling assumption and can fail under bursts, common shocks or feedback.
5. Stationary Increments Depend Only on Interval Length
For a homogeneous process, the distribution of N(t+h)−N(t) depends on h, not on calendar time t.
6. Tiny-Interval Behaviour Generates the Process
Over a short interval Δt, one arrival has probability approximately λΔt, zero arrivals have probability approximately 1−λΔt, and two or more arrivals have smaller order probability. These local rules generate the global count law.
7. Waiting Times Are Exponential
The time T₁ until the first arrival satisfies P(T₁>t)=exp(−λt). Interarrival times are IID Exponential(λ) in the homogeneous process.
8. Exponential Waiting Is Memoryless
Conditional on having waited s time units with no event, the remaining waiting-time distribution is unchanged. This special property creates a Markov representation.
9. The kth Arrival Time Is Gamma Distributed
The sum of k independent exponential interarrival times has a gamma/Erlang distribution. Count questions and waiting-time questions are therefore two sides of the same process.
10. Count and Waiting-Time Events Are Equivalent
The event that the kth arrival occurs by time t is exactly the event N(t)≥k. This identity translates between Poisson and gamma calculations.
11. The Process Is Markov
Given the current count, future increments are independent of the earlier arrival history under the homogeneous model. See How Markov Chains Work for the wider state-transition architecture.
12. Superposition Adds Independent Poisson Streams
Independent Poisson processes with rates λ₁,…,λₘ combine into a Poisson process with rate equal to the sum of the component rates.
13. Thinning Splits a Poisson Stream
If each arrival is independently classified into type A with probability p, the retained A events form a Poisson process of rate pλ and the complementary events form another independent Poisson process.
14. Conditional Arrival Times Are Uniformly Scattered
Conditional on N(t)=n, the unordered arrival times inside [0,t] behave like n independent Uniform(0,t) draws. Their ordered values are uniform order statistics.
15. Nonhomogeneous Poisson Processes Allow Time-Varying Rates
Replace constant λ with intensity λ(t). Expected count over an interval becomes the integral of intensity. Independent increments remain, but stationary increments generally disappear.
16. Time Transformation Can Simplify a Nonhomogeneous Process
Accumulated intensity Λ(t)=∫₀ᵗλ(s)ds acts as a transformed clock. Under standard conditions, mapping event times through accumulated intensity converts the process to unit-rate Poisson time.
17. Conditional Intensity Generalises the Idea Further
General point processes can have intensity depending on event history. Self-exciting Hawkes processes and self-correcting processes abandon independent increments while retaining an instantaneous-rate description.
18. Bursty Data Often Reject the Homogeneous Poisson Model
If events trigger more events, counts become clustered. Variance can exceed the mean and interarrival times lose the simple exponential structure.
19. Inhibition Can Produce More Regular Spacing
Some systems suppress nearby events, producing underdispersion or refractory periods. A Poisson process has no such memory.
20. Queueing Theory Uses Poisson Arrivals Carefully
The M/M/1 queue combines Poisson arrivals with exponential service and one server. The model is analytically elegant because memorylessness creates a birth–death Markov chain.
21. PASTA Connects Arrivals With Time Averages
Under appropriate Poisson-arrival conditions, Poisson arrivals see time averages: the state distribution observed by arriving customers matches the time-average state distribution.
22. Reliability Uses Poisson Counts for Some Failure Mechanisms
A constant hazard with independent events can lead to exponential waiting times and Poisson counts. Ageing components with increasing hazard violate this simple structure.
23. Insurance Uses Compound Poisson Processes
Let claim arrivals follow a Poisson process and attach a random claim severity to each arrival. The aggregate loss is a compound Poisson process, separating frequency from severity.
24. Telecommunications Models Packet or Call Arrivals
Poisson arrivals can be useful baselines when many independent low-rate sources superpose. Modern network traffic can show dependence and burstiness that require richer models.
25. Epidemiology Can Use Poisson Rates for Event Counts
Incidence counts over person-time are often modelled with Poisson regression, but contagion, heterogeneity and overdispersion can violate independent constant-rate assumptions.
26. Poisson Regression Is a Count Model, Not Automatically a Poisson Process
A cross-sectional Poisson regression specifies a conditional mean/count distribution. A temporal Poisson process additionally specifies increment and event-time structure.
27. Estimating a Homogeneous Rate Is Simple
Observe n events over total exposure T. The maximum-likelihood estimate is λ̂=n/T under the homogeneous model. Uncertainty remains, especially with small counts.
28. Exposure Must Be Measured Correctly
Ten events in ten machine-hours and ten events in ten thousand machine-hours imply radically different rates. Rate inference requires a denominator.
29. Overdispersion Is a Diagnostic Signal
The basic Poisson distribution has equal mean and variance. Substantially larger empirical variance can signal unobserved heterogeneity, clustering or model misspecification.
30. Mixtures Can Produce Negative-Binomial Counts
If Poisson rates vary across units according to a gamma distribution, marginal counts become negative binomial. Apparent overdispersion can therefore reflect hidden rate heterogeneity.
31. Censoring and Observation Windows Matter
Events outside the recording window are unobserved. Detection failures, downtime and left truncation can distort estimated rates if exposure is treated as complete.
32. Simulation Is Straightforward in the Homogeneous Case
Generate exponential interarrival times and accumulate them until the horizon is crossed, or sample the total Poisson count and then uniform order statistics conditional on that count.
See How Monte Carlo Simulation Works.
33. Thinning Simulates Nonhomogeneous Processes
Simulate a homogeneous process at an upper-bound rate and accept candidate events with time-dependent probability λ(t)/λmax. Efficiency depends on how tight the bound is.
34. What Poisson Processes Preserve
They preserve event timing, count distributions, rate, independent increments and the duality between counts and exponential waiting times under the homogeneous model.
35. What the Homogeneous Model Discards
It discards seasonality, contagion, inhibition, hidden rate heterogeneity, ageing and dependence between disjoint intervals.
36. Hostile Test: Daily Seasonality
Calls arrive far more often at noon than midnight. A single constant λ averages the day and misrepresents both peak congestion and quiet periods.
37. Hostile Test: Self-Exciting Events
One event raises short-term probability of another. Independent increments fail even if the overall average count looks Poisson-like.
38. Hostile Test: Ageing Failure Hazard
A component becomes more failure-prone as it ages. Exponential waiting and constant hazard are structurally wrong.
39. Practical Workflow
Define events and exposure, inspect time variation, test dispersion and interarrival behaviour, examine dependence, estimate rate with uncertainty, compare richer point-process models, validate peak and tail behaviour, and simulate only after the event mechanism is credible.
40. Canonical Boundary
Probability Distributions owns the Poisson count distribution broadly. Stochastic Processes owns the wider process family. This article owns Poisson arrival-time and counting-process mechanics.
Final Thought
The Poisson process is powerful because one rate parameter creates a complete relationship between counts, waiting times and independent increments. That simplicity is useful only when the event mechanism deserves it.