Why a population needs a distribution, not just a mean
Suppose each individual produces a random number of children and then disappears. The population in one generation is the sum of many independent offspring counts. Even with the same average number of children, two populations can have different chances of dying out: reproduction may be nearly certain or may alternate between none and many. We need the entire offspring distribution to calculate extinction.
This unit treats a discrete-time countable-state Markov-chain example. The reproduction assumptions below are a model: dependence between families or a changing environment requires different equations.
Definition 2.12.1
(The Galton–Watson construction)
Let for with . Each individual independently has children with probability , using the same distribution in every generation. Starting from , write
The empty sum is zero; thus zero is absorbing. The probability generating function (PGF) is defined on the whole displayed interval without a moment assumption. Its left derivative at one is the mean , possibly infinite. Extinction means for some finite .
Conditioned on , the next count is a sum of independent offspring variables; it does not depend on the earlier population history. Thus is a Markov chain on the nonnegative integers. The state must be the population count, not the generation number alone.
The fixed point selected by the process
Let be the probability that the line of one ancestor is extinct by generation .
Lemma 2.12.2
(Finite-generation extinction recursion)
The probabilities increase to the extinction probability, and
With independent initial ancestors the extinction probability is .
Since zero is absorbing, the events increase to the extinction event, so . Conditioning on the first individual's children, each of the family lines must be extinct by the remaining time; their independent probabilities multiply to , and averaging over gives . The same independence of family lines gives for ancestors.
Theorem 2.12.3
(Extinction is the smallest fixed point)
For any offspring distribution, is the smallest solution in of . If , then . Otherwise, if , then ; if , including , then .
Proof
(Monotone iteration and the tangent at one)
Continuity of the PGF on follows by dominated convergence, so . If is any fixed point, induction using monotonicity gives from ; hence . This minimality matters because one is always a fixed point.
For the classification, set . If and , then necessarily . If for every , then for . Otherwise, for ,
Here when and ; the sums are finite in the required mean-weighted sense. Thus there is no fixed point below one. The exceptional law instead produces exactly one individual forever.
If , then as , including an infinite limit, by monotone convergence applied to . Therefore for some . The iteration from zero is bounded above by that , because is increasing and ; hence . This proves all cases without assuming a finite variance.
Proof X-ray
What monotonicity and the tangent at one each decide
Two different properties of do the work. Monotonicity gives minimality: from and increasing, every iterate stays below any fixed point , so the iteration selects the smallest root even though always solves . The classification instead compares with its tangent line at one, whose slope is : the bound says exactly that for , which is why the mean rather than the whole law decides whether a root below one exists. The hypothesis enters once, to force when ; for , for every and the iteration stays at zero. In the supercritical case the argument closes by trapping the iteration below a point where , so neither a finite variance nor a finite mean is needed.
A complete calculation
Worked example 2.12.4
(Zero, one or two children)
Suppose , and . Then , and
The fixed points are and ; the extinction probability from one ancestor is , and from three independent ancestors it is . The first two finite-generation probabilities are and . These do not equal the eventual answer: more generations allow further opportunities for extinction.
Worked example 2.12.5
(Roots and finite generations)
For and , find , and the probability of extinction from two ancestors.
The equation has roots and , so . Since , . Independent initial lines give extinction probability .
Worked example 2.12.6
(A degenerate case)
Each individual has exactly two children. Find and explain why solving the fixed-point equation alone requires care.
, so . Here has roots zero and one; the smallest root is correct. Selecting one merely because it is always a root would wrongly predict extinction.
Common mistake
(Reading almost-sure extinction as a deadline)
The law has mean one and extinction probability one, yet for every finite : all individuals of generations may each have two children. Almost-sure eventual extinction does not give a deterministic deadline.
What moments can and cannot establish
Suppose . Conditional expectation yields , so induction gives
For , Markov's inequality gives , a second proof of extinction. At the bound is only one. It cannot prove survival either: the critical zero-or-two example dies out almost surely while its expected population remains one at every generation. Rare large surviving populations compensate for the increasingly likely zero population.
If additionally the offspring variance , conditioning on and using total variance gives
for , with the first expression also interpreted directly at ; then nonnegative offspring with mean zero is identically zero, so its variance is zero. These are finite-generation identities, not a proof of a limiting distribution. Infinite variance makes the displayed variance calculation unavailable even when the mean is finite.
Counterexample mode
Equal means need not give equal extinction probabilities
The deterministic law and the law both have mean one. Their extinction probabilities are respectively zero and one. The mean determines the broad classification only after the deterministic-one exception is identified; in the supercritical case the actual probability requires the PGF, not just its derivative at one.
Practice with solutions
Quick check
(Let each individual have zero or two children with equal probability. Find and , and state the eventual extinction probability.)
Use the mean and variance recursions with .
Solution · Answer
The offspring mean and variance are both one, so and . As the law is critical and not deterministically one child, extinction occurs with probability one.
Quick check
(In the zero-one-two example, what is the probability that exactly one of three initial family lines survives forever?)
Each line survives with probability , independently.
Solution · Answer
Each line survives with probability , independently. The answer is . This is not the same event as survival of the combined population, whose probability is .
Quick check
(For , and , find and explain why is a double root of .)
Factor .
Solution · Answer
Here and , so the theorem gives . Directly, , which is positive on , so there is no fixed point below one. The double root reflects tangency: the slope of at one equals , the slope of the diagonal.
Further scope
The argument gives extinction probabilities and finite-generation moments under independent identical reproduction. It does not cover changing environments, immigration, dependent families, extinction-time asymptotics or conditioning on survival.