1. Definition and Core Properties

1.1 Counting process framework

A Poisson process is a stochastic counting model that tracks how many random events have occurred by a time \(t\). It is represented by a nondecreasing integer-valued process \(N(t)\), where \(N(0)=0\) and increments correspond to additional events appearing after time \(t\).

1.2 Stationary increments

The process has stationary increments when the distribution of the number of events in an interval depends only on the interval length, not on its starting time. In the homogeneous (time-constant rate) case, this means that the law of \(N(t+s)-N(s)\) is determined by \(t\) alone.

1.3 Independent increments

Independent increments means that counts in disjoint time intervals are independent random variables. For example, the numbers of events in \((0,s)\) and \((s,t)\) are independent. This independence is central to many derivations and computations for the model.

1.4 Poisson distribution of event counts

For a fixed time length \(t\), the increment count \(N(t+s)-N(s)\) follows a Poisson distribution with a mean proportional to \(t\). Under the homogeneous assumption, one writes \[ P(N(t)=k)=e^{-\lambda t}\frac{(\lambda t)^k}{k!}, \quad k=0,1,2,\dots \] where \(\lambda>0\) is the constant rate.

1.5 No-simultaneous-events assumption

The model is also described by a “no multiple events at exactly the same time” requirement, formalized so that simultaneous occurrences have probability zero. Informally, in a very small interval, at most one event is likely, and the probability of two or more events is negligible of higher order.

2. Mathematical Formulation

2.1 Intensity (rate) parameter \(\lambda\)

The parameter \(\lambda\) is the intensity of the process. It represents the average number of events per unit time in the homogeneous setting. When \(\lambda\) increases, event counts grow and waiting times shrink.

2.2 Notation for \(N(t)\)

The standard notation \(N(t)\) denotes the cumulative number of events occurring up to time \(t\). For increments over an interval \([s,s+t]\), one uses \(N(s+t)-N(s)\), which counts events appearing within that window.

2.3 Transition probabilities

Transition probabilities describe the chance of moving from count \(i\) at time \(s\) to count \(j\) at time \(s+t\). With the Poisson increment property, for \(m=j-i\ge 0\), \[ P(N(s+t)=j\mid N(s)=i)=P(N(t)=m)=e^{-\lambda t}\frac{(\lambda t)^m}{m!}. \]

2.4 Increment distributions over time intervals

More generally, for any \(t\ge 0\), \[ N(t)\sim \text{Poisson}(\lambda t), \] and for disjoint intervals, corresponding increments are independent. These two ideas—Poisson-distributed increments and their independence—fully specify the homogeneous Poisson process under standard formulations.

3. Interarrival Times

3.1 Waiting-time distribution

Interarrival times measure the time between consecutive event occurrences. Let \(T_1\) be the time of the first event, and \(T_n\) be the waiting time between the \((n-1)\)th and \(n\)th events; equivalently, the arrival times are \(S_n=T_1+\cdots+T_n\).

3.2 Exponential distribution characterization

In a homogeneous Poisson process, each interarrival time is exponentially distributed with rate \(\lambda\): \[ P(T_1>t)=e^{-\lambda t}, \quad t\ge 0, \] and the density is \(\lambda e^{-\lambda t}\). Moreover, the interarrival times are independent and identically distributed under the homogeneous assumption.

3.3 Memoryless property and intuition

The exponential distribution is characterized by memorylessness: \[ P(T_1>t+s\mid T_1>s)=P(T_1>t). \] In the process interpretation, conditioning on having waited \(s\) without an event does not change the remaining waiting-time law.

3.4 Relation between counts and waiting times

Counts and waiting times are linked by the threshold relationship: the event count exceeds \(k-1\) by time \(t\) exactly when the \(k\)th arrival occurs by time \(t\). Using arrival times \(S_k\), \[ \{N(t)\ge k\}=\{S_k\le t\}. \] This provides a bridge between Poisson counting formulas and exponential (and related gamma/Erlang) waiting-time computations.

4. Distribution of the Number of Events

4.1 Probability of exactly k events in time t

For any \(t\ge 0\) and integer \(k\ge 0\), \[ P(N(t)=k)=e^{-\lambda t}\frac{(\lambda t)^k}{k!}. \] This expresses the likelihood of observing \(k\) arrivals in a fixed time horizon when the average rate is constant.

4.2 Mean and variance of N(t)

The expectation and variance coincide for the Poisson distribution: \[ \mathbb{E}[N(t)]=\lambda t,\qquad \mathrm{Var}(N(t))=\lambda t. \] This “equidispersion” property—equal mean and variance in the homogeneous model—is often used diagnostically in applied settings.

4.3 Moment generating function and useful identities

The moment generating function of \(N(t)\) is \[ M_{N(t)}(\theta)=\mathbb{E}[e^{\theta N(t)}]=\exp\!\big(\lambda t(e^\theta-1)\big). \] From this, one can derive moments and factorial moments, and also obtain identities useful for summing independent increments.

4.4 Approximations and limiting behavior

For large \(\lambda t\), the Poisson distribution can be approximated by a normal distribution with mean \(\lambda t\) and variance \(\lambda t\). When \(\lambda t\) is small, probabilities are dominated by the low-count terms (notably \(k=0\) and \(k=1\)), reflecting the rare-event behavior in short intervals.

5. Timing of Events: Order Statistics View

5.1 Event times as random points

Condition on having \(n\) events by time \(t\). Under the Poisson process structure, those event times behave like random points scattered over the interval \([0,t]\), with a joint distribution that can be characterized using order statistics.

5.2 Distribution of the nth arrival time

The time of the \(n\)th arrival, \(S_n\), follows an Erlang (gamma) distribution with shape \(n\) and rate \(\lambda\): \[ f_{S_n}(s)=\frac{\lambda^n s^{n-1}e^{-\lambda s}}{(n-1)!},\quad s\ge 0. \] This arises because \(S_n\) is a sum of \(n\) i.i.d. exponential interarrival times.

5.3 Conditional structure given N(t)

Given \(N(t)=n\), the \(n\) ordered arrival times \((S_1,\dots,S_n)\) have a distribution equivalent to the order statistics of \(n\) independent uniform random variables on \([0,t]\). This conditional uniformity simplifies many probability calculations about event ordering within a fixed horizon.

5.4 Uniform order statistics characterization

In the conditional-uniform framework, the gaps between consecutive order statistics have a Dirichlet-related structure. Marginally, for \(k=1,\dots,n\), the \(k\)th arrival time scaled by \(t\), i.e. \(S_k/t\), follows a Beta distribution with parameters \(k\) and \(n-k+1\).

6. Simulation and Practical Generation

6.1 Simulating interarrival times

A common simulation approach generates i.i.d. exponential waiting times and cumulatively sums them until the desired time horizon is reached. If \(U\sim \text{Uniform}(0,1)\), then an exponential variable can be obtained as \[ T=-\frac{1}{\lambda}\ln U. \] Repeated generation yields sample paths of event times.

6.2 Simulating event counts in fixed windows

Alternatively, one may sample \(N(t)\) directly from the Poisson distribution for each interval of interest. For simulation over multiple disjoint windows, counts can be generated independently (consistent with the independent increment property), then optionally refined into event times using the conditional uniform order-statistics method.

6.3 Thinning approach overview (for inhomogeneous generalizations)

For models where the event rate varies with time, a technique called thinning is often used. One simulates a higher-rate homogeneous process and then independently keeps each candidate event with probability proportional to the ratio of the true instantaneous rate to the upper bound rate. Kept events form the desired non-homogeneous process, while rejected events are discarded.

6.4 Validation checks for simulated paths

After generating sample paths, typical checks include:

  • empirical mean and variance of counts matching \(\lambda t\),
  • comparison of interarrival histograms to the exponential density,
  • verification of independence between increments across disjoint windows,
  • reproduction of conditional uniformity of arrival times given \(N(t)\).

Such diagnostics help detect coding errors and confirm that assumptions align with the simulation method.

7.1 Non-homogeneous Poisson process (time-varying rate)

A non-homogeneous Poisson process allows the intensity to depend on time, \(\lambda(t)\). In that case, counts in an interval have a Poisson distribution with mean equal to the integral of the rate over the interval. Independence of increments remains, but increments are no longer stationary.

7.2 Compound Poisson process (random jump sizes)

In a compound Poisson process, each event triggers a random “jump” of a certain size (e.g., claims, packet sizes, or workloads). The total accumulated amount by time \(t\) is the sum of these random jump sizes over a Poisson number of events, linking Poisson arrivals with severity distributions.

7.3 Marked Poisson process (events with attributes)

A marked Poisson process attaches additional random attributes (“marks”) to each event, such as magnitudes, categories, or locations. One studies not only when events occur, but also how their associated features distribute and potentially depend on each other.

7.4 Connection to renewal processes

Poisson processes are closely related to renewal theory. While renewal processes use i.i.d. interarrival times in general, a Poisson process corresponds to the special case where interarrival times are exponential. This special choice yields the memoryless property and produces the Poisson count distribution.

8. Multidimensional/Spatial Perspectives

8.1 Poisson process over space

The concept extends beyond time to space (or space-time). For a bounded region, a Poisson process can model the number of randomly located events (e.g., incidents or particles) occurring within that region.

8.2 Independent increments in regions

For spatial versions, one partitions space into disjoint regions and assumes counts in those regions are independent. The process typically also requires a notion of intensity measure so that expected counts scale with region size or other geometric features.

8.3 Intensity measure concept (general idea)

Instead of a single constant rate, spatial models use an intensity measure \(\Lambda(A)\) for a region \(A\). The number of events in \(A\) then follows a Poisson distribution with mean \(\Lambda(A)\). This framework provides a flexible way to represent varying density across space.

9. Applications and Modeling Use-Cases

9.1 Queueing and arrivals

In queueing systems, Poisson arrivals represent customers arriving randomly at a steady average rate. When service times have additional assumptions, the resulting queue dynamics can become analytically tractable, such as in the classic M/M/1 setting.

9.2 Reliability and failures

In reliability contexts, Poisson modeling can represent random failure occurrences over time, particularly when failures are treated as independent and “memoryless” at the event level. This yields exponential waiting times for the next failure in the simplest homogeneous case.

9.3 Telecommunications and network traffic

For network traffic, Poisson processes are used as baseline models for call arrivals, message occurrences, or event-like interactions. While real traffic may deviate from homogeneity, the Poisson model remains a common starting point due to its mathematical convenience.

9.4 Operations and incident counting

Organizations often track incident reports, system errors, or operational tickets. Poisson processes can model the count of such events over intervals, enabling forecasts, alert thresholds, and uncertainty quantification under the assumption of a roughly constant average rate.

10. Common Derivations and Theorems

10.1 Deriving Poisson increment probabilities from axioms

A classical derivation starts from axioms such as independent increments, stationarity, and the small-interval behavior where the probability of one event is proportional to interval length and multiple events are negligible. Under these conditions, one obtains the Poisson distribution for counts in fixed time lengths.

10.2 Deriving exponential interarrival times

From the “no simultaneous events” assumption and the Poisson distribution of counts in short intervals, one can show that the probability of observing no events up to time \(t\) equals \(e^{-\lambda t}\). This leads directly to the exponential distribution for the first waiting time, and by independence of increments it extends to subsequent interarrival times.

10.3 Conditional distribution results

Conditioning on \(N(t)=n\) yields strong structural conclusions: the arrival times become distributed like uniform order statistics. These results follow from the exchangeability of event locations within the interval and from the consistency of Poisson increments across subintervals.

10.4 Useful properties for inference and prediction

For inference, the parameter \(\lambda\) can be estimated from observed counts using standard Poisson likelihood methods. For prediction, the model supplies probabilities of future event counts over future time horizons, leveraging independent increments to combine information from different observed intervals.