the branching random walk
What if branching individuals are not just counted but located in space, and each child is displaced from its parent? The branching random walk (BRW) adds geometry to the Galton-Watson tree: every particle has a position on the real line, and when it reproduces, its children are placed at positions equal to the parent's plus independent random displacements. The central questions become spatial: how far does the leftmost (or rightmost) particle travel, and how fast does the cloud of particles spread?
Start with one particle at 0. At each step, every particle dies and is replaced by offspring whose displacements (one per child) are drawn from a fixed point process L on the real line; so a particle at position x with k children produces particles at x + d_1, ..., x + d_k where (d_1, ..., d_k) is a realization of L. The key analytic object is the log-Laplace transform of the mean point process, Lambda(theta) = log E[ sum_i e^(theta d_i) ], whose Legendre-type structure governs the spatial spread by a large-deviations mechanism. The headline result (Hammersley, Kingman, Biggins) is the speed of the minimum: the leftmost particle M_n satisfies M_n / n -> gamma almost surely on survival, where gamma = - inf_(theta > 0) Lambda(theta)/theta is the asymptotic speed of the extremal front. Biggins's theorem makes this precise and identifies the regime where the additive martingale W_n(theta) = sum_(particles) e^(theta x_i - n Lambda(theta)) converges to a nontrivial limit (an x log x condition again). Finer still, the front M_n has a logarithmic correction: M_n = gamma n + c log n + O(1), and the centred minimum converges to a randomly-shifted Gumbel law — the discrete-space analogue of the branching-Brownian-motion extremal results.
Why it matters: the BRW is the discrete prototype for front propagation, extremal statistics of strongly correlated systems, and the universality class that includes branching Brownian motion, directed polymers, the discrete Gaussian free field, and (conjecturally) the maxima of the Riemann zeta function on the critical line. The speed gamma is a clean large-deviations computation, but the honest subtlety is that the naive linear prediction gamma n is wrong at second order: the log n correction (the Bramson shift) and the random Gumbel limit are genuine, hard results that require the spine/derivative-martingale machinery. And as always, the additive martingale is non-degenerate only under the appropriate integrability — speed and martingale convergence are different theorems with different (overlapping) hypotheses.
Suppose each particle has exactly 2 children, each displaced by an independent standard normal step. Then Lambda(theta) = log 2 + theta^2/2, and the front speed of the maximum is v = inf_(theta>0) Lambda(theta)/theta = sqrt(2 log 2). The rightmost of the about 2^n particles after n steps sits near v n, far ahead of a single random walk (which only reaches O(sqrt n)).
Many correlated walkers push the front to linear speed gamma n, with a universal log n correction and Gumbel fluctuations.
The leading speed gamma n is a large-deviations computation, but it is only first order: the true front is gamma n + c log n + O(1) (the Bramson logarithmic correction), and the centred extremal particle converges to a random shift of a Gumbel law, not to a constant.