Coupled Branching Random Walks

• Coupled Branching Random Walks are stochastic processes where multiple branching dynamics interact via explicit couplings, providing a framework to analyze extinction and survival behaviors.
• Finite local modifications and interaction-dependent reproduction in these models yield precise comparisons of global and local critical thresholds and phase structures.
• Diverse models, including cooperative branching with coalescence, multi-type couplings, and exact embeddings in systems like limit order books, offer key insights into anomalous spreading and capacity phenomena.

Coupled branching random walks comprise several related constructions in which branching random walk dynamics are compared, modified, or represented through an explicit coupling. In the most classical setting, two branching random walks on the same countable space are coupled by requiring that their breeding rates coincide outside a finite set, so that their trajectories agree up to the first visit to that set. In other parts of the literature, the same phrase refers to systems in which branching is itself interaction-dependent, as in cooperative branching with coalescence, to reducible multi-type branching random walks whose types force one another’s extreme-particle behavior, and to exact or asymptotic couplings between branching random walks and other stochastic systems such as limit order books, additive sums of random-walk ranges, or ancestral lineages in regulated populations.

1. Formal framework and basic notions

A standard continuous-time branching random walk (BRW) is defined on an at most countable set $X$. Each particle carries a location $x\in X$, has an exponentially distributed lifetime with parameter $1$, and, while alive, sends children from $x$ to $y$ at rate $\lambda k_{xy}$, where $K=(k_{xy})$ is a nonnegative matrix and $\lambda>0$ is a multiplicative speed parameter. Increasing $\lambda$ accelerates all breeding rates simultaneously. The model admits a standard discrete-time counterpart $(X,\mu)$ with first-moment matrix $x\in X$0, and, under irreducibility and boundedness assumptions, the continuous- and discrete-time versions have the same global and local survival behavior.

For a BRW $x\in X$1, extinction in a set $x\in X$2 is encoded by

$x\in X$3

Global extinction corresponds to $x\in X$4, while local extinction corresponds to a singleton or finite set. Strong local survival in $x\in X$5 means

$x\in X$6

whereas non-strong local survival means

$x\in X$7

The two critical parameters are the global threshold

$x\in X$8

and the local threshold

$x\in X$9

Under the hypotheses used in the comparison theory, these depend only on the irreducible class of $1$0; in an irreducible BRW one writes simply $1$1 and $1$2. A pure global survival phase is the interval $1$3 with $1$4.

Two complementary characterizations are central. The local parameter satisfies

$1$5

and also

$1$6

where

$1$7

The global parameter admits the operator form

$1$8

In the finite-perturbation setting, two irreducible BRWs are called equivalent when their rates differ only on finitely many rows:

$1$9

is finite. This notion isolates a local coupling class in which one studies how extinction probabilities and critical parameters change under finite modifications of the breeding mechanism (Bertacchi et al., 19 Jul 2025).

2. Finite-set couplings, extinction comparison, and phase structure

If two discrete-time BRWs $x$0 and $x$1 satisfy $x$2 for all $x$3, with $x$4 finite, then their paths coincide until the first visit to $x$5. This yields a comparison theorem for extinction probabilities. In particular, for all $x$6,

$x$7

where $x$8 is the probability of never visiting $x$9, and for any $y$0,

$y$1

If the two offspring laws also coincide outside $y$2, then

$y$3

The mechanism is explicit: any event involving survival in $y$4 while avoiding $y$5 has the same probability under both laws because the dynamics are identical up to the hitting time of $y$6.

For continuous-time irreducible BRWs $y$7 and $y$8 that agree outside a finite set $y$9, the comparison translates into a rigid set of alternatives for the critical parameters. Corollary 3.3 states that the following are equivalent:

1. $\lambda k_{xy}$0,
2. $\lambda k_{xy}$1,
3. $\lambda k_{xy}$2.

Thus, any finite local modification that lowers the global critical parameter necessarily collapses the pure global survival phase in the modified system: the modified BRW has $\lambda k_{xy}$3.

The possible parameter configurations inside a finite-perturbation equivalence class are exactly the following.

Case	Critical-parameter relation
1	$\lambda k_{xy}$4
2	$\lambda k_{xy}$5
3	$\lambda k_{xy}$6

A further rigidity result concerns maximality. If an irreducible BRW $\lambda k_{xy}$7 has a pure global survival phase, so $\lambda k_{xy}$8, then every finite perturbation $\lambda k_{xy}$9 satisfies

$K=(k_{xy})$0

Moreover, if $K=(k_{xy})$1 also has a pure global survival phase, then

$K=(k_{xy})$2

Within such an equivalence class, the presence of a pure phase therefore forces the global threshold of that BRW to be maximal.

The examples sharpen these abstract alternatives. On the homogeneous tree $K=(k_{xy})$3 with nearest-neighbor rates $K=(k_{xy})$4 for neighbors,

$K=(k_{xy})$5

so a pure global survival phase exists. If one adds a loop of rate $K=(k_{xy})$6 at the origin, then for small $K=(k_{xy})$7 the parameters remain unchanged; for intermediate $K=(k_{xy})$8 one has $K=(k_{xy})$9 while $\lambda>0$0 decreases monotonically; and for

$\lambda>0$1

one gets $\lambda>0$2, with both tending to $\lambda>0$3 as $\lambda>0$4. By contrast, on $\lambda>0$5 with symmetric nearest-neighbor rates $\lambda>0$6 for $\lambda>0$7, one has $\lambda>0$8, and finite local modifications cannot create a pure global survival phase if the base process is quasi-transitive with finite degrees. The same paper also exhibits equivalence classes in which the supremum of $\lambda>0$9 is finite but not attained, and others in which the class supremum is infinite (Bertacchi et al., 19 Jul 2025).

3. Interaction-driven coupled systems: cooperative branching, coalescence, and cobra walks

A distinct meaning of coupled branching random walk appears when branching is not independent per particle but depends on local particle configurations. In the one-dimensional cooperative branching–coalescent, the state space is $\lambda$0, so at most one particle occupies each site. Particles perform nearest-neighbor random walks with rate $\lambda$1 to each neighbor, while branching occurs only from pairs of neighboring occupied sites: if $\lambda$2, an offspring is created at $\lambda$3 with rate $\lambda$4, and symmetrically at $\lambda$5 with rate $\lambda$6. Coalescence is built into the update $\lambda$7, so collisions merge particles rather than increase occupancy. The generator splits as

$\lambda$8

with the random-walk and cooperative-branching parts given explicitly in terms of the local updates $\lambda$9 and $(X,\mu)$0.

This system is attractive, admits an upper invariant law $(X,\mu)$1, and has a phase transition. There exist critical parameters $(X,\mu)$2 and $(X,\mu)$3 such that $(X,\mu)$4 for $(X,\mu)$5, while $(X,\mu)$6 is nontrivial for $(X,\mu)$7, and the process dies out for $(X,\mu)$8 but survives for $(X,\mu)$9. The numerics reported in the paper suggest

$x\in X$00

For $x\in X$01, any finite nonempty initial state almost surely ends in a single particle eventually:

$x\in X$02

A further hallmark is slow relaxation. The spectral gap is zero in the subcritical regime, and in one dimension the density started from the fully occupied state decays algebraically as $x\in X$03 rather than exponentially. For $x\in X$04, upper and lower bounds of order $x\in X$05 are proved both for the particle density and for the probability that a process started from two neighboring particles still has more than one particle at time $x\in X$06 (Sturm et al., 2013).

Coalescing-branching random walks on finite graphs, or cobra walks, give a discrete-time graph-theoretic variant. In a $x\in X$07-cobra walk on a finite connected simple graph $x\in X$08, the state at time $x\in X$09 is an active set $x\in X$10. Each active vertex $x\in X$11 chooses $x\in X$12 neighbors independently and uniformly with replacement; all chosen vertices become candidates for the next round, and duplicates coalesce:

$x\in X$13

For $x\in X$14, the resulting cover-time bounds are substantially smaller than worst-case simple-random-walk cover times. On the grid $x\in X$15, the cover time is $x\in X$16 for any fixed dimension $x\in X$17. On an $x\in X$18-vertex $x\in X$19-regular graph with conductance $x\in X$20, the cover time is

$x\in X$21

and on an arbitrary connected $x\in X$22-vertex graph,

$x\in X$23

These systems are coupled branching random walks in the literal sense that branching creates parallel exploratory trajectories while coalescence prevents uncontrolled growth and feeds back on future exploration (Mitzenmacher et al., 2016).

4. Multi-type coupling and anomalous spreading

In single-type branching random walks on $x\in X$24, reproduction is encoded by a point process $x\in X$25 of offspring displacements. If $x\in X$26 denotes the rightmost particle in generation $x\in X$27 and

$x\in X$28

then the asymptotic speed of the maximum is

$x\in X$29

equivalently

$x\in X$30

where $x\in X$31 is the convex dual. Under the standard exponential-moment assumptions, $x\in X$32 almost surely.

In an irreducible multi-type BRW, the one-generation transforms form the matrix

$x\in X$33

whose Perron–Frobenius spectral radius gives

$x\in X$34

The speed formula becomes

$x\in X$35

The reducible case is qualitatively different. For a two-type system with $x\in X$36 able to produce $x\in X$37 but not conversely, let $x\in X$38 and $x\in X$39 be the convex duals of the diagonal transforms, and let $x\in X$40 be their convex minorant. Then the speed of type $x\in X$41 started from type $x\in X$42 is

$x\in X$43

This $x\in X$44 can exceed both single-type speeds $x\in X$45 and $x\in X$46, producing anomalous spreading. The paper’s branching Brownian motion example makes this explicit: if $x\in X$47 is binary branching Brownian motion with variance $x\in X$48 and split rate $x\in X$49, then $x\in X$50; if $x\in X$51 has variance $x\in X$52, split rate $x\in X$53, and always produces two $x\in X$54 together with an independent $x\in X$55 with probability $x\in X$56, then $x\in X$57 as well, but

$x\in X$58

The same variational structure mirrors the linear-determinacy formula for coupled Fisher–KPP systems,

$x\in X$59

and the probabilistic interpretation identifies the mechanism: mass produced by one type at intermediate locations allows another type to realize faster extremes than it could in isolation (Biggins, 2010).

5. Intersection-equivalence, additive ranges, and capacity

Another branch of the subject couples a critical branching random walk to additive geometry built from independent simple random walks. Let $x\in X$60 denote the range of a critical BRW on $x\in X$61 started at $x\in X$62, with critical offspring distribution of mean $x\in X$63 and finite positive variance $x\in X$64. Conditioning the genealogy to survive forever yields an infinite invariant tree, or sin-tree, consisting of a spine with independent critical Bienaymé–Galton–Watson trees attached to each spine vertex. In dimensions $x\in X$65, its Green’s function is of order $x\in X$66, the convolution of the simple-random-walk Green’s function with itself.

For a finite set $x\in X$67, Zhu’s branching capacity is

$x\in X$68

and it is comparable to the capacity associated with the kernel $x\in X$69:

$x\in X$70

Thus $x\in X$71, and since $x\in X$72 decays like $x\in X$73, the branching capacity is equivalent to a $x\in X$74 capacity up to constants.

The central intersection-equivalence theorem states that in $x\in X$75 there exists $x\in X$76 such that for any finite $x\in X$77 and any $x\in X$78 sufficiently far from $x\in X$79,

$x\in X$80

where $x\in X$81 denotes the range of a simple random walk and the two copies are independent. The coupling here is not pathwise identity but equivalence of distant hitting probabilities through a common potential kernel.

The additive model extends to $x\in X$82 independent simple random walks. Writing $x\in X$83 for their ranges and $x\in X$84, one has for $x\in X$85 an almost-sure law of large numbers

$x\in X$86

and distant hitting probabilities satisfy

$x\in X$87

For $x\in X$88, this places the infinite invariant critical BRW and the Minkowski sum of two independent random-walk ranges in the same potential-theoretic class (Asselah et al., 2023).

6. Exact couplings to other stochastic systems and genealogical environments

A particularly explicit coupling appears in a one-sided limit order book model. The state space is the set of finite point measures on $x\in X$89. For a book $x\in X$90, the price is the rightmost order,

$x\in X$91

with $x\in X$92 when the book is empty. At each discrete time step, if the book is empty it resets to $x\in X$93; if it is nonempty, then with probability $x\in X$94 a limit order arrives at price $x\in X$95, and with probability $x\in X$96 a market sell removes one order at $x\in X$97. The dynamics are

$x\in X$98

The coupling enriches the state space by assigning a genealogy to orders: any order placed while $x\in X$99 is declared a child of the order at $1$00. With colors green, red, and white to encode currently present, removed, and not-yet-added nodes, the book becomes the green point measure of a pre-generated random tree. The offspring number is geometric,

$1$01

with generating function

$1$02

and one-generation cumulant

$1$03

If $1$04 is the random labeled tree and $1$05 the tree operator that either activates the next white child of the front or turns the front red, then

$1$06

where $1$07 is the first empty-book time. The long-term price behavior is then read from BRW barrier arguments. Assuming $1$08 and $1$09 finite: if $1$10, then $1$11 almost surely; if $1$12 and $1$13, setting

$1$14

one has $1$15 almost surely when

$1$16

and $1$17 almost surely when

$1$18

(Simatos, 2012).

A different coupling principle arises in logistic branching random walks with local competition. Here the population process is a discrete-time Markov chain $1$19 with update

$1$20

where $1$21 are independent Poisson processes and

$1$22

Reproduction at a site depends on the whole local neighborhood through the competition kernel $1$23, so the branching is locally regulated rather than independent. Under the high-density/weak-competition regime of Assumption 4.2, the process survives with positive probability and has a unique nontrivial invariant extremal distribution $1$24.

Sampling a uniformly chosen ancestor produces a random walk in the time-reversal of this dynamic environment. Its quenched transition probabilities are

$1$25

so the ancestral walk is a spatial embedding of genealogy in a coupled BRW background. The paper proves, under the abstract coarse-graining assumptions verified for logistic branching random walks, a strong law of large numbers with speed $1$26 and an annealed central limit theorem with nontrivial covariance. In the logistic case this yields, conditional on $1$27, that the ancestral walk is almost surely defined for all times and

$1$28

under the annealed law (Birkner et al., 2015).

Taken together, these works show that coupling in branching random walks is not a single device but a family of comparison principles. It can mean equality of dynamics outside a finite set, interaction-dependent reproduction with coalescence, reducible forcing between particle types, potential-theoretic equivalence of hitting probabilities, or exact embedding of an external system into a BRW genealogy. Across these settings, the recurrent themes are extinction probabilities, front speeds, capacity, regeneration, and the way local modifications propagate into global phase structure.