Poisson Processes

Understanding the Poisson process—a mathematical model for counting random events over time—is fundamental across engineering, finance, and science. It provides the rigorous framework for analyzing everything from network traffic and insurance claims to radioactive decay and customer arrivals. This model’s power lies in its elegant simplicity, yielding tractable results for complex stochastic systems, making it a cornerstone of applied probability and operations research.

Homogeneous vs. Nonhomogeneous Poisson Processes

The core model is the homogeneous Poisson process (HPP). It is defined by three key properties. First, events occur one at a time; simultaneous arrivals are negligible. Second, it possesses independent increments, meaning the number of events in non-overlapping time intervals are independent random variables. Third, it has stationary increments, where the distribution for the number of events in any interval depends only on the interval's length, not its starting time. For an HPP, the probability of observing exactly $k$ events in an interval of length $t$ is given by the Poisson distribution:

$$P(N(t)=k)=\frac{(\lambda t{)}^k{e}^{-\lambda t}}{k!},k=0,1,2,...$$

Here, $N(t)$ is the counting process, and $\lambda >0$ is the constant rate or intensity (e.g., 2 arrivals per minute). This rate is the expected number of events per unit time, so $E[N(t)]=\lambda t$.

The nonhomogeneous (or inhomogeneous) Poisson process (NHPP) generalizes this by relaxing the stationary increments assumption. Its rate is a function of time, $\lambda (t)$, called the intensity function. The expected number of events by time $t$ is given by the mean value function $\Lambda (t)={\int }_0^t\lambda (s)ds$. The probability of $k$ events in the interval $(a,b]$ becomes: $$P(N(b)-N(a)=k)=\frac{[\Lambda (b)-\Lambda (a){]}^k{e}^{-[\Lambda (b)-\Lambda (a)]}}{k!}.$$ This is crucial for modeling systems with predictable peaks and lulls, such as call center traffic during business hours or the arrival rate of patients at an emergency room over a 24-hour cycle.

Key Distributions: Interarrival and Waiting Times

Two intimately related distributions govern the timing of events. Let $T_1,T_2,...$ denote the interarrival times, the random periods between consecutive events. For a homogeneous Poisson process with rate $\lambda$, these times are independent and identically distributed (i.i.d.) exponential random variables with parameter $\lambda$. The probability density function (PDF) is $f_T(t)=\lambda e^{-\lambda t}$ for $t\ge 0$. This exponential distribution is memoryless, meaning $P(T>s+t|T>s)=P(T>t)$; the time until the next event is independent of how long you've already waited.

The $n$th waiting time (or arrival time), $S_n=T_1+T_2+...+T_n$, is the time of the $n$th event. Since $S_n$ is the sum of $n$ i.i.d. exponential($\lambda$) variables, it follows an Erlang distribution (a special case of the Gamma distribution). Its PDF is: $$f_{S_n}(t)=\frac{{\lambda }^n{t}^{n-1}{e}^{-\lambda t}}{(n-1)!},t\ge 0.$$ This relationship is powerful: the event "the $n$th arrival occurs by time $t$" ($S_n\le t$) is equivalent to the event "at least $n$ events have occurred in $[0,t]$" ($N(t)\ge n$). This duality connects continuous-time distributions ($S_n$) with discrete counting probabilities ($N(t)$).

Fundamental Properties: Superposition and Thinning

Two operations that preserve the Poisson structure are superposition and thinning, which are invaluable for modeling and decomposition.

Superposition states that if you merge independent Poisson processes, the combined process is also Poisson. Formally, if $N_1(t)\sim \text{Poisson}({\lambda }_{1}t)$ and $N_2(t)\sim \text{Poisson}({\lambda }_{2}t)$ are independent, their superposition $N(t)=N_1(t)+N_2(t)$ is a Poisson process with rate ${\lambda }_1+{\lambda }_2$. This property is extensively used in telecommunications, where aggregate packet streams from many independent users converge on a network router, forming a combined arrival process that can often be modeled as Poisson.

Thinning (or splitting) is the reverse operation. Starting with a Poisson process $N(t)$ of rate $\lambda$, classify each arrival as "Type I" with probability $p$ or "Type II" with probability $1-p$, independently of all other events. The processes of Type I and Type II arrivals, $N_1(t)$ and $N_2(t)$, are independent Poisson processes with rates $\lambda p$ and $\lambda (1-p)$, respectively. A classic application is in insurance claim modeling, where a company's total claims form a Poisson process, and each claim is independently classified as either "major" (exceeding a deductible) or "minor" based on a probability, yielding two independent streams for different risk pools.

Compound Poisson Processes

A compound Poisson process enriches the model by attaching a random "weight" to each event. It is defined as $X(t)={\sum }_{i=1}^{N(t)}{Y}_i$, where $N(t)$ is a Poisson process and $\{Y_i\}$ is an i.i.d. sequence of random variables, independent of $N(t)$. The process $X(t)$ models the cumulative value—such as total damage, cost, or data packets—up to time $t$.

Its moments can be derived using conditional expectation. If $E[Y]=\mu$ and $\text{Var}(Y)={\sigma }^2$, then: $$E[X(t)]=E[N(t)]E[Y]=\lambda t\mu ,$$ $$\text{Var}(X(t))=\lambda t({\sigma }^2+{\mu }^2).$$ This process is pivotal for modeling aggregate insurance claims over a period, where claim amounts ($Y_i$) vary randomly. It also forms the basis for fundamental queueing models, like the M/G/1 queue, where Poisson arrivals (the M) bring generally distributed service requirements (the G).

Common Pitfalls

1. Assuming Unjustified Independence: The most frequent error is assuming a process is Poisson without verifying the core properties, especially independent increments. For example, arrivals during a sale might cluster (violating independence), or the rate might vary systematically (requiring an NHPP). Always question the underlying assumptions before applying the model.
2. Confusing Arrival Rates and Probabilities: Remember that for a small time interval $\Delta t$, the probability of exactly one arrival in an HPP is approximately $\lambda \Delta t$, not equal to it. The exact probability is $e^{-\lambda \Delta t}\lambda \Delta t$, with the approximation being valid only as $\Delta t\to 0$. Using $\lambda \Delta t$ for a sizable interval will overestimate the probability.
3. Misapplying the Exponential Interarrival Result: The i.i.d. exponential interarrival times property holds only for the homogeneous Poisson process. For a nonhomogeneous process, interarrival times are neither exponential nor independent. Mistaking them as such will lead to incorrect predictions about waiting times.
4. Overlooking the Role of Thinning in Decomposition: When given a combined process and a classification probability, a common mistake is to treat the resulting sub-processes as dependent. Remember, the beauty of Poisson thinning is that it yields independent Poisson streams. Failing to use this independence can complicate analyses unnecessarily.

Summary

• The homogeneous Poisson process is characterized by independent and stationary increments with a constant rate $\lambda$, leading to a Poisson count distribution and i.i.d. exponential interarrival times.
• The nonhomogeneous Poisson process generalizes the model with a time-varying intensity function $\lambda (t)$, making it essential for modeling systems with non-constant average rates.
• Superposition (combining) and thinning (splitting) of independent Poisson processes yield new, independent Poisson processes, enabling the modeling of aggregated streams or classified events.
• A compound Poisson process sums i.i.d. random variables at each event time, providing the standard model for aggregate random quantities like total insurance claims or bulk data arrivals.
• These models are directly applied to telecommunications (traffic aggregation), insurance (claim frequency and severity), and queueing theory (modeling customer arrivals and service requirements).