Branching Processes

Core Idea

A Galton-Watson branching process models population dynamics: each individual in generation n independently produces a random number of offspring according to a fixed distribution {p_k}, forming generation n+1. The process is classified by the mean offspring number μ = E[X]: subcritical (μ < 1), critical (μ = 1), or supercritical (μ > 1). The extinction probability q — the probability the population eventually dies out — satisfies q = G(q) where G is the probability generating function of the offspring distribution. Extinction is certain (q = 1) if and only if μ ≤ 1; when μ > 1, the population survives with positive probability 1 - q, and the normalized process W_n = Z_n/μ^n is a non-negative martingale whose convergence determines the growth rate.

Explainer

Branching processes are the fundamental stochastic model for population dynamics with random reproduction. The Galton-Watson process — introduced by Francis Galton and Henry Watson in 1874 to study the extinction of Victorian family surnames — begins with a single ancestor (Z_0 = 1) and evolves by each individual independently producing a random number of offspring drawn from a fixed distribution {p_k}_{k≥0}. The population in generation n+1 is Z_{n+1} = Σ_{i=1}^{Z_n} X_i⁽ⁿ⁾, where the X_i⁽ⁿ⁾ are i.i.d. copies of the offspring variable X. The process is Markov with state space {0, 1, 2, ...}, and 0 is an absorbing state — once the population dies, it stays dead.

The criticality classification by the mean offspring number μ = E[X] governs the qualitative behavior. When μ < 1 (subcritical), E[Z_n] = μ^n → 0 exponentially, and extinction is certain with geometrically decaying survival probability. When μ = 1 (critical), E[Z_n] = 1 for all n, but extinction is still certain (assuming Var(X) > 0) — survival probability decays as 2/(nσ²). When μ > 1 (supercritical), the population grows exponentially on the survival event, with E[Z_n] = μ^n. The extinction probability q is the smallest fixed point of the probability generating function G(s) = E[s^X] = Σ_k p_k s^k: the equation q = G(q) follows from conditioning on the first-generation size and using the independence of descendant subtrees.

The martingale connection is central. The normalized population W_n = Z_n/μ^n is a non-negative martingale: E[W_{n+1} | ℱ_n] = Z_n · μ / μ^{n+1} = W_n. By the martingale convergence theorem, W_n → W a.s. for some non-negative random variable W. The question is whether W is degenerate (W = 0 a.s.) or has a genuine positive part. In the subcritical and critical cases, W = 0 a.s. In the supercritical case, {W = 0} = {extinction}, so P(W > 0) = 1 - q. But even in the supercritical case, the limit can degenerate if the offspring distribution has too heavy a tail. The Kesten-Stigum theorem provides the sharp criterion: W is non-degenerate (equivalently, W_n converges in L¹) if and only if E[X log X] < ∞. This X log X condition is the branching-process analogue of the uniform integrability condition in general martingale convergence theory.

Extensions of the basic Galton-Watson model are numerous and important. Multi-type branching processes allow several types of individuals with type-dependent reproduction, governed by a mean matrix M whose largest eigenvalue determines criticality. Continuous-time branching processes (Bellman-Harris processes) replace discrete generations with random lifetimes, connecting to renewal theory and age-dependent models. Branching processes in random environments (BPRE) let the offspring distribution vary randomly between generations, modeling fluctuating environmental conditions. In the continuous limit, the population process converges to a continuous-state branching process (CSBP), which is a Lévy process with a specific branching structure — these connect to superprocesses and measure-valued diffusions in modern probability theory. Branching processes also appear in applications far beyond biology: nuclear chain reactions (the original motivation for the Bellman-Harris model), epidemic spreading, the cascade structure of Galton-Watson trees in combinatorics, and the analysis of recursive algorithms in computer science.