Counting many small chances
Suppose a system has many opportunities for an event to occur, but each opportunity has a small probability. We may want the probability of no failures, or of several arrivals, without adding a long list of binomial terms. A Poisson law can approximate the count, even when the individual probabilities are unequal. The two ingredients are small individual probabilities and an appropriate dependence model; merely knowing the expected count is not enough.
Let be independent indicators with , where . Set and . The mean of is exactly . A Poisson variable with parameter also has mean , and
At , interpret the law as a point mass at zero (equivalently, use in the formula).
Matching the mean suggests an approximation but does not prove it. The following bound controls the error for every event consisting of nonnegative integer counts, not just for one value of .
Theorem 30.1.1
(A quantitative rare-event approximation)
Under the independent-indicator assumptions above, if , then for every ,
The right-hand side may exceed one, in which case it remains true but gives no useful numerical information.
Proof
(Build the two counts on the same probability space)
For each independently, first draw . We construct an indicator alongside it. If , put . If , independently set with probability
This is between zero and one: . Thus . Independent construction across makes have the same distribution as . The sum is Poisson with parameter , by multiplication of the independent Poisson generating functions.
The two coordinates differ precisely when , or when and the extra Bernoulli draw gives one. Therefore
If the two sums differ, at least one pair differs, so the union bound gives . Finally, for any event , the indicators and can differ only when the sums differ. Since has the law of and has the law of ,
Proof X-ray
One coupling controls every event at once
The key move is to stop comparing two probability formulas and instead build the two counts on one probability space, so that the comparison reduces to the single number . That number does not mention , which is why the bound is uniform over all events.
Independence enters twice, and only after the coordinatewise construction. Independence of the original indicators is what makes have the law of ; independence of the makes their sum Poisson. The union bound itself needs no independence. The size of the error is decided coordinate by coordinate: a Bernoulli and a Poisson variable with the same mean can be matched except on an event of probability , and turns this into . The hypothesis is used only to keep a probability. The coupling is a proof device: it preserves each sum's law, without asserting that the events in the original system are Poisson events.
Worked example 30.1.2
(A thousand independent opportunities)
Suppose independent components each fail during a fixed interval with probability . Then and . The approximation to the probability of at least two failures is
The theorem guarantees absolute error at most for this event. This is an absolute probability error, not a relative error. The exact answer is ; the bound does not require us to evaluate it.
The theorem does not need equal probabilities, and it is worth seeing how conservative its bound can be.
Worked example 30.1.3
(Unequal probabilities)
Take independent indicators, with probability and with probability . Then and . The exact probability of no successes is
while the Poisson value is . The actual error, about , is well inside the guaranteed . The bound depends on the individual probabilities only through , so the mixture of two sizes needs no separate treatment.
Worked example 30.1.4
(An edge case where the bound is uninformative)
Let and , so with certainty and . For ,
and no event gives a larger difference, since is the total Poisson mass off the point . The theorem's bound is : true, but of no numerical use. A certain event is the opposite of a rare event, and the bound correctly refuses to call a point mass approximately Poisson.
A limiting statement and its hypotheses
Consider a sequence of models, indexed by , with independent indicators in each model and probabilities , and let be the count in model .
Corollary 30.1.5
(Poisson limit for rare independent indicators)
If and , then converges in distribution to Poisson, including the degenerate case .
Proof
(Combine the bound with a comparison of Poisson laws)
First,
For any set of nonnegative integers, the theorem compares with a Poisson variable of parameter , with error at most this sum. It remains to compare Poisson with Poisson uniformly in . For , couple a Poisson count with an independent Poisson increment; their sum is Poisson, and the two counts differ only when the increment is nonzero, which has probability . Reverse their roles when . Hence for Poisson, which in particular gives convergence in distribution.
Counterexample mode
Small marginals do not replace independence
Let be Bernoulli with success probability , and take all indicators equal to . Each individual probability tends to zero and the sum of the success probabilities is one, but the count is . Thus , whereas Poisson assigns probability to zero. Here successes occur together; the independent-indicator theorem does not apply. The equal means do not repair the dependence.
Common mistakes
Common mistake
(Reading the bound as a relative error)
In the thousand-component example the bound is . For this is a small relative error, but for a tail event it is not. The Poisson value of is , and an absolute error of up to is more than a quarter of that. For small target probabilities the bound alone does not certify the leading digit.
Common mistake
(Approximating a count of likely events by a Poisson law)
If each of independent components works with probability , the number working is not approximately Poisson: its variance is , whereas that Poisson law has variance . Apply the theorem to the failures instead. The failure count is a sum of indicators with probability , so it is close to Poisson within , and the number working is minus that count.
Exercises
Quick check
(Independent indicators have probabilities , and . Give a Poisson approximation to the probability of no successes and a bound on its absolute error.)
Compute and .
Solution · Answer
Here . The approximation is ; the error is at most . The exact probability is , so the actual error is about .
Quick check
(Let for . Find the limiting probability of exactly two successes and a bound on the absolute error of the Poisson approximation for each .)
All indicators have the same probability .
Solution · Answer
The parameter is for every , and . Thus , and for each the absolute error of the Poisson approximation is at most . For this bound is at least one and says nothing.
Quick check
(Suppose each of independent events has probability . Does the theorem show that a Poisson distribution with parameter approximates the count with error tending to zero? Explain.)
Evaluate .
Solution · Answer
No. The bound is , which does not tend to zero. The individual probabilities do tend to zero, but the total mean grows. This calculation only says that this particular bound fails to establish a vanishing error; it does not by itself prove that no better approximation is possible.