Why these closure rules matter
By now the Poisson process is familiar from three angles: as a counting process with independent stationary increments, as the limit of many rare independent chances, and as a stream whose interarrival times are independent Exponential waits. This guide is about something different — not what one Poisson process is, but how Poisson processes behave when you combine them, filter them, or peek inside them. The astonishing fact is that the family is almost indestructible: the natural operations you would want to perform on a stream of events keep landing you back inside the Poisson family.
Three closure rules carry the load, and they are the heart of this guide. Superposition: merge several independent Poisson streams and the merged stream is again Poisson, with the rates simply added. Thinning: keep each point of a Poisson stream independently with some probability and the kept points form a Poisson process again, at a smaller rate. And the uniform (order-statistics) property: condition on knowing exactly how many points landed in an interval, and those points are scattered as if you had thrown that many independent uniform darts. Each rule turns a hard probability question into an easy one, which is exactly why the Poisson process is the workhorse model for arrivals, defects, calls, photons, and rare events of every kind.
Superposition: pouring streams together
Picture a call centre with two independent lines. Sales calls arrive as a Poisson process at rate lambda_1 = 3 per hour; support calls arrive, independently, as a Poisson process at rate lambda_2 = 5 per hour. Now ignore the labels and just watch the total flow of calls. Superposition says the merged stream is itself a Poisson process, and its rate is the sum: lambda = lambda_1 + lambda_2 = 8 per hour. Nothing exotic happens at the join; the combined arrivals are as clean and memoryless as either part.
Why does the sum of the rates appear, and not something messier? Take the increment view. Over any interval the two streams contribute independent Poisson counts — say N_1 ~ Poisson(lambda_1 t) and N_2 ~ Poisson(lambda_2 t) over a window of length t. A foundational fact about the Poisson distribution is that independent Poissons add to a Poisson: N_1 + N_2 ~ Poisson((lambda_1 + lambda_2) t). The increments of the merged stream are therefore Poisson with the summed rate, and because both originals had independent increments over disjoint intervals, so does the merger. That is the whole proof — superposition is the time-process face of "independent Poissons add."
There is a beautiful bonus for any single merged arrival: which stream did it come from? The chance that the next call is a sales call is just its share of the total rate, lambda_1 / (lambda_1 + lambda_2) = 3/8. This follows from the competing-exponentials picture you met for interarrival times — the minimum of independent Exponential waits is Exponential at the summed rate, and the probability that a given one is the winner is its rate divided by the total. So superposition not only merges the timing cleanly, it splits the credit for each event in proportion to the rates.
Thinning: keeping a random subset
Thinning runs superposition in reverse. Start with one Poisson stream and, independently for each arrival, flip a biased coin: keep this point with probability p, discard it with probability 1 - p. Thinning says the kept points form a Poisson process at rate p*lambda, the discarded points form a Poisson process at rate (1 - p)*lambda — and, the surprising part, these two resulting streams are independent of each other. A stream of 8 calls per hour, of which each is genuinely urgent with probability p = 0.25, yields urgent calls as a Poisson process at rate 0.25 * 8 = 2 per hour.
Why Poisson again, and at exactly rate p*lambda? Condition on N ~ Poisson(lambda t) total points in a window. Each is kept independently with probability p, so the number kept, given N, is Binomial(N, p). Mixing a Binomial over a Poisson count is a classic computation whose answer is clean: the marginal number kept is Poisson(p*lambda*t), and the number discarded is Poisson((1 - p)*lambda*t), and the two are independent. The independence is the genuinely surprising line and it is special to the Poisson; for most starting distributions the kept and discarded counts would be entangled.
Thinning a Poisson(lambda) stream by keeping each point with prob p:
kept points ~ Poisson process, rate p * lambda
discarded points ~ Poisson process, rate (1 - p) * lambda
AND the kept and discarded streams are INDEPENDENT.
Superposition is the reverse:
Poisson(rate p*lambda) + Poisson(rate (1-p)*lambda)
--(merge independent)--> Poisson(rate lambda)Thinning is enormously useful precisely because it decouples the pieces. If 30% of website visitors arrive on mobile and arrivals are Poisson at rate lambda, then mobile visits and desktop visits are two independent Poisson processes — you can model, queue, and forecast them separately without ever worrying that a busy mobile hour drained the desktop stream. The same trick classifies insurance claims by type, photons by detector, or particles by which way they scattered, each category inheriting its own clean Poisson process.
The uniform property: where the points land
Now condition on the count. Suppose you are told that in the interval [0, T] exactly one Poisson arrival occurred. Where in that interval did it happen? The answer is the cleanest possible: it is uniformly distributed on [0, T]. No part of the interval is favoured — given that a single bus came in the past hour, it was equally likely at any minute. There is no "buses cluster near the end" or "things settle down later"; conditional on the count, the position carries no information about timing beyond being somewhere in the window.
The full statement is the order-statistics property: given that exactly n arrivals fell in [0, T], their unordered locations are distributed exactly like n independent Uniform(0, T) points, and the actual arrival times are then the order statistics of those — the same n uniform points simply sorted into increasing order. So to simulate the times of a Poisson process inside a fixed window, you do not need to step through it: draw N ~ Poisson(lambda*T), then throw N independent uniform darts on [0, T] and sort them. The uniform conditional property turns the dynamic process into a static sprinkle.
Here is the property earning its keep on a tiny problem. Suppose 4 customers arrived in a shop during a one-hour window, modelled as Poisson. What is the chance all four arrived in the first half hour? Conditional on the count of 4, each arrival is independently uniform on the hour, so each lands in the first half with probability 1/2. By independence the chance all four did is (1/2)^4 = 1/16. We computed a genuine probability about timing without touching lambda at all — the rate cancelled out the moment we conditioned on the count, which is the practical signature of the uniform property.
Bending the rate and the jumps: non-homogeneous and compound
Real arrivals rarely keep one fixed rate all day — restaurants fill at lunch, websites peak in the evening. The non-homogeneous Poisson process keeps independent increments but lets the rate vary in time, lambda(t). The count in an interval is now Poisson with mean equal to the integrated rate over that interval: the mean over [a, b] is the area under lambda(t) from a to b, written as the integral of lambda(t) dt. Everything else survives — superposition and thinning still hold — and the uniform property generalizes beautifully: given n points in a window, they are independent but distributed according to the shape of lambda(t) (heavier where the rate is higher), not flat-uniform.
There is a lovely link back to thinning here. A non-homogeneous Poisson process can be generated by starting with a constant-rate process at the peak rate and then thinning it with a position-dependent keep-probability that traces the desired shape — the famous thinning algorithm for simulation. So the rate-bending and the random-dropping ideas are two views of the same machinery, which is why the rate is the single dial that governs the whole picture.
The other generalization changes not when events happen but how much each one carries. In a compound Poisson process each Poisson arrival comes attached to an independent random size — claim amounts on an insurance arrival, dollars on a sale, customers in each arriving group. The total accumulated by time t is S_t = X_1 + X_2 + ... + X_{N_t}, a random sum of N_t independent jumps where N_t is itself Poisson. By the random-sum formulas (the law of total expectation and total variance), its mean is E[S_t] = lambda*t * E[X] and its variance is Var(S_t) = lambda*t * E[X^2]. The number of events stays Poisson, but the running total is no longer integer-valued or Poisson — it inherits a richer shape from the jump distribution.
Where the magic stops
These rules are powerful, so it is worth being precise about their boundaries. Every one of them rests on the same two pillars as the Poisson process itself: independence and the right increment structure. Superposition needs the merged streams to be independent of one another; thinning needs each point kept or dropped independently of the others and of the underlying process; the uniform property is a conditional statement that holds because, given the count, the points genuinely carry no further structure. Remove any of these and you leave the Poisson family — the conclusions are not approximately true, they are simply false.
Two honest misconceptions to bury. First, the closure rules do not say arrivals avoid clustering — quite the opposite. Because increments are independent and memoryless, Poisson arrivals genuinely do bunch up by chance; a quiet stretch makes the next arrival no less likely, which is the memoryless flavour of the gambler's fallacy applied to time. The uniform property even quantifies this: independent uniform darts on an interval routinely land in visible clumps, so apparent clustering is the signature of true randomness, not evidence against it. Second, thinning's independence of the kept and discarded streams is special to the Poisson; do not assume two sub-categories of any arrival process are independent just because you split them.
With superposition, thinning, and the uniform property in hand, you can manipulate Poisson streams almost algebraically — adding rates, splitting by probability, and freezing counts to expose hidden uniformity. The next guide lifts all of this off the time line and onto the plane: the spatial Poisson point process and its marked cousin, where the very same closure rules govern stars in the sky, trees in a forest, and defects on a silicon wafer.