Marking arrivals
A single stream of arrivals may contain two types, or two separate streams may feed the same system. When can their counts be treated as Poisson processes, and when are the types independent? We state the marking assumptions before calculating.
Definition 4.1.1
(Independent marking of a Poisson process)
Let be a homogeneous Poisson process of rate . Each arrival independently receives type A with a fixed probability , and type B otherwise. The marks are independent of the entire original process and of one another. Let and count the respective types up to time , so that .
For or , a rate-zero process is the identically zero process. The two independence requirements in the definition, between marks and between marks and , are both used below.
Theorem 4.1.2
(Independent splitting)
The processes and are independent homogeneous Poisson processes with rates and , respectively.
Proof
(Factor the joint counts on every interval)
On a fixed interval of length , put . Given , the chance of A marks and B marks is . Multiplying by the Poisson mass gives, for ,
This factors into Poisson masses of means and , because , including zero-mean cases under the usual convention. For any finite collection of disjoint intervals, original increments are independent and marks assigned to their arrivals are independent. The same factorization in every interval gives mutually independent A and B increments across the whole collection. Thus both are Poisson processes with the stated stationary independent increments, and their finite-dimensional distributions factor, proving independence as processes.
Proof X-ray
Independence across time needs a proof
The proof has a one-interval computation and an extension step, and the two independence hypotheses enter at different places.
- On one interval, the binomial factor uses that the marks are independent of one another and of : given , the marks are still independent trials with success probability . The key move is the cancellation of against the binomial coefficient, together with the split .
- Factoring the two counts at one time is not enough to establish process independence. The extension uses independent increments of and the independence of marks across intervals, so that the factorization holds jointly on every finite collection of disjoint intervals.
The argument closes because finite-dimensional distributions on disjoint intervals determine the law of a counting process. If the marks were allowed to depend on each other, the binomial step would already fail; the counterexample below shows this concretely.
Combining independent streams
The reverse operation merges streams. Here independence of the two streams is the hypothesis, and the conditional split given the total is the conclusion.
Theorem 4.1.3
(Superposition and the conditional split)
If are independent Poisson processes of rates , their sum is a Poisson process of rate . For , and every ,
Proof
(Multiply generating functions, then divide masses)
For a fixed interval the product of their probability generating functions is
and disjoint-interval increments remain independent, so the sum has independent Poisson increments of the stated rate. For the conditional law, divide the product of the two Poisson masses by the mass of their sum:
The conditioning event has positive probability for every under these hypotheses.
Splitting and superposition are inverse descriptions: a Poisson stream marked independently with probability is the same as the superposition of independent streams of rates and .
Worked examples
Worked example 4.1.4
(Two marked arrival types)
Suppose per hour and . Over 20 minutes the A and B means are and . Independence gives . Given exactly three total arrivals in that interval, the chance that two are type A is .
Worked example 4.1.5
(Merging two independent streams)
Let calls arrive on two independent lines at rates and per hour. Over one hour the total is Poisson with mean , so . Given four calls in that hour, the number from the first line is , so the chance that exactly one came from the first line is
Check: the conditional mean is , and averaging over the Poisson total gives , the rate of the first line.
Worked example 4.1.6
(The boundary case where every mark is A)
Take . Then and is identically zero. The joint formula reads , which with is the Poisson mass at and zero for . So is the rate-zero Poisson process, and it is independent of because a constant process is independent of everything. The theorem therefore needs no separate statement at , but only because of the convention .
When the marks are dependent
The marking hypotheses matter. Equal marginal type probabilities do not replace independent marks.
Counterexample mode
Identically distributed marks need not be independent
Claim: if each arrival individually has probability of type A, then and are independent Poisson processes of rate . Suppose instead a single fair coin is tossed once and then all arrivals are assigned A if it is heads, otherwise all are B. Each arrival individually has probability of type A, but marks are not independent. For , always while both means equal , so and the counts are not independent. Nor is Poisson: , which exceeds for because for , with equality only at . The repaired statement is the theorem: marks independent of one another and of .
Common mistakes
Common mistake
(Treating counts as independent after conditioning on the total)
For and an interval with , the unconditional counts are independent. Given , however, , so they cannot both equal one, even though each has positive conditional probability or of being one. Their conditional joint law therefore does not factor. Independence of streams and dependence after conditioning on their sum are compatible.
Common mistake
(Superposing dependent streams)
The superposition theorem needs independent streams. If , the sum takes only even values, so although its mean is positive; it is not Poisson.
Exercises
Quick check
(For rate per hour and , give the A and B rates and the expected A count in 20 minutes.)
Rates multiply by the mark probabilities.
Solution · Answer
The rates are and per hour; the expected A count in 20 minutes is .
Quick check
(With the same rates, find over 40 minutes.)
Find the two means first.
Solution · Answer
Over 40 minutes the means are and , so independence gives .
Quick check
(Given four total arrivals in an interval, find the probability of exactly one A mark when .)
Use the binomial split given the total.
Solution · Answer
Given the total, the A count is , so the probability is .
Quick check
(Why do independent types become dependent when conditioned on exactly one total arrival, for ?)
What is on that event?
Solution · Answer
Given a total of one, their sum is fixed at one. Both conditional probabilities of a count equal to one are positive, but the joint probability that both equal one is zero.
Further reading
MIT OpenCourseWare, Poisson combining and splitting develops combining and splitting of Poisson processes.