A population model
Definition 4.2.1
(Galton–Watson process)
Start with one individual, . Each individual independently has a nonnegative integer number of children with the same distribution , and different individuals and generations use independent offspring counts. The size of generation is .
In particular is absorbing: once no individuals remain, none can be born. The independence and identical-distribution assumptions are part of the model, not automatic features of real populations; actual offspring counts may be dependent or limited by resources.
Definition 4.2.2
(Probability generating function)
The probability generating function of the offspring law is for , with also at .
It is continuous on this interval, since the tail of the series is uniformly bounded by the corresponding probability tail, and it is nondecreasing because every coefficient is nonnegative. Define and let be the probability of eventual extinction. We will show how iterating gives extinction by a specified generation and how its smallest fixed point describes eventual extinction.
Finite time and the fixed point
Theorem 4.2.3
(Extinction probability is the smallest fixed point)
The probabilities satisfy and , they increase to , and is the smallest solution of in .
Proof
(Condition on the first generation)
Condition on the number of children of the initial individual. Given , their descendant trees evolve independently and each is distributed as a fresh copy of the whole process. All trees must be extinct after another generations, so
The events increase with , because is absorbing. By continuity of probability from below,
Letting in and using continuity of gives . If is any fixed point, then , and if then because is nondecreasing. By induction for every , hence .
The root always exists, but choosing it without checking for a smaller root can give the wrong answer. Starting instead with independent initial individuals gives eventual extinction probability , since all independent family trees must die out.
Mean offspring and the threshold
Assume now that . Conditioning on gives , so . Finite expectation is justified inductively; it does not by itself determine extinction.
Theorem 4.2.4
(Extinction threshold)
Suppose . If , then . If and , then . If , then , so survival has positive probability.
Proof
(Compare the graph of the generating function with the diagonal)
If , Markov's inequality applied to the integer-valued gives , so .
If and , note that for and integer , , with strict inequality when . The law must assign positive probability to some : otherwise almost surely and mean forces almost surely. Hence, for ,
There is no fixed point below , so the previous theorem gives .
If , the left derivative of at is . This follows from monotone convergence applied to , which increases to as ; when the inner sum is empty and equals . Consequently for sufficiently close to . Since , the continuous function has a zero in (possibly itself). Therefore the smallest fixed point satisfies .
Proof X-ray
What the mean controls in each case
The subcritical case does not use fixed points at all: and Markov's inequality are enough, and the step uses that is integer-valued. The other two cases use the previous theorem to turn a question about eventual extinction into a question about where the graph of meets the diagonal on . The mean is the slope of at . A slope above one forces just left of , while puts the graph on or above the diagonal at ; the intermediate value theorem supplies a smaller root. A slope equal to one only gives the tangent-line bound , and strictness comes from mass on two or more children. That is the single place where the hypothesis enters.
The threshold is about the probability of eventual extinction, not about whether the expected population grows.
Counterexample mode
Mean one without the nondegeneracy hypothesis
Claim: every offspring law with mean gives extinction with probability one. Take . Then , , and every point of is a fixed point; the smallest is . Indeed for every , so . The claim fails, and the critical case of the theorem needs : some individual must have a positive chance of two or more children, which by mean one also forces a positive chance of none.
Common mistake
(Taking the root at one)
The equation always has the root , and for a law with and mean above one it also has a smaller root. Extinction is the smallest root, because climbs from and can never pass a fixed point. Reporting for would ignore the root that the iterates actually approach.
Common mistake
(Growing means do not exclude extinction)
A mean is compatible with a positive extinction probability. In the example below the expected size is , yet the population dies out with probability ; the mean is large because surviving populations become large, not because extinction is impossible.
Worked examples
Worked example 4.2.5
(Zero or two children)
Suppose and . Then and . Extinction by generations 1 and 2 has probabilities
The fixed-point equation has roots and ; the smallest is . Thus survival probability is , even though extinction remains possible and . With two independent initial individuals extinction probability is .
Worked example 4.2.6
(A critical geometric law)
Let for . The series gives , and . The fixed-point equation is , so is the only fixed point and , as the critical case predicts. The iterates are , , . In general , which one checks by induction: . Hence : extinction is certain but slow, unlike the geometric decay of the subcritical bound.
Worked example 4.2.7
(An edge case with no childless individuals)
Let , so . Then and . The fixed points solve , namely and , and the smallest is . Directly, every individual has at least one child, so for all and for every . This is the case “possibly itself” in the supercritical proof: and the root below is the endpoint .
The explorer below iterates from for five offspring laws, marks every fixed point of in with the smallest highlighted, and reads off for initial individuals.
Exercises
Quick check
(Let and . Find and the eventual extinction probability. Give a bound on .)
Solution · Answer
, so . Markov's inequality yields .
Quick check
(Let . Explain why the mean equals one but eventual extinction still has probability one. Contrast this with the deterministic one-child law.)
Solution · Answer
The mean is and , so the nondegenerate critical case applies and . Directly, and becomes . For almost surely the population is always one and .
Quick check
(For the zero-or-two example, compute the probability of extinction by generation 3 and the probability of extinction **first occurring** at generation 2.)
Solution · Answer
. Since extinction events are nested, first extinction at generation 2 has probability .