Logistic Branching Process

• The logistic branching process is a density-dependent model where linear reproduction is counterbalanced by a quadratic death term, ensuring regulated population sizes.
• It encompasses continuous-time birth–death processes, Feller diffusions, and spatial/multitype systems, linking deterministic and stochastic modeling approaches.
• The analysis covers extinction guarantees, quasi-stationarity through conditioning, and genealogical limits that reveal both Kingman and multiple-merger coalescent behaviors.

The logistic branching process is a density-dependent branching model in which linear reproduction is counterbalanced by a quadratic competition term, so that growth is supercritical at low density and suppressed at high density. In a standard continuous-time birth–death formulation, a population of size $i$ gives birth at rate $b\,i$ and dies at rate $d\,i+c\,i(i-1)$; in large-population scalings the competition coefficient is replaced by $c/N$, yielding a typical equilibrium size of order $N$. The same mechanism appears in discrete-time branching models with carrying capacity, in Feller diffusions with logistic drift, in continuous-state branching processes with quadratic competition, and in spatial and multitype systems whose genealogies exhibit both classical and multiple-merger coalescent limits (Etheridge et al., 2013, Garrett et al., 5 Sep 2025, Foucart, 2021).

1. Canonical formulations

A basic logistic branching process is a continuous-time Markov chain on $\mathbb Z_+$ with transition rates

$q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$

or, equivalently,

$$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$

Here births are linear in the current population size, whereas competition contributes a quadratic death term. This model has been proposed for numbers of individuals in populations competing for some resource, and also for numbers of species (Etheridge et al., 2013).

In the large-carrying-capacity setting used for genealogical limits, one writes $X_t^N\in\{0,1,2,\dots\}$ and

$$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$

The logistic regulation is then the term $b\,i$0: each ordered pair of individuals contributes to density-dependent mortality at rate $b\,i$1. At the deterministic level,

$b\,i$2

so the carrying-capacity scale is

$b\,i$3

This formulation makes explicit that the equilibrium population size is of order $b\,i$4 (Garrett et al., 5 Sep 2025).

The literature also uses discrete-time analogues. In population-size-dependent branching processes, individuals reproduce i.i.d. according to an offspring law $b\,i$5 that depends on the current population size $b\,i$6, with mean $b\,i$7. Logistic behavior is encoded by choosing $b\,i$8 for $b\,i$9 and $d\,i+c\,i(i-1)$0 for $d\,i+c\,i(i-1)$1. Two standard examples are Beverton–Holt,

$d\,i+c\,i(i-1)$2

and Ricker,

$d\,i+c\,i(i-1)$3

Equivalent deterministic-control branching processes encode the same mean through a control function $d\,i+c\,i(i-1)$4 when $d\,i+c\,i(i-1)$5 (Braunsteins et al., 2023).

Setting	State space	Logistic mechanism
Birth–death chain	$d\,i+c\,i(i-1)$6	death rate $d\,i+c\,i(i-1)$7
Large-$d\,i+c\,i(i-1)$8 chain	$d\,i+c\,i(i-1)$9	death rate $c/N$0
Feller diffusion	$c/N$1	drift $c/N$2
Logistic CSBP	$c/N$3	generator term $c/N$4

2. Deterministic, diffusion, and pathwise limits

For the continuous-time birth–death process with carrying capacity parameter $c/N$5, the rescaled population $c/N$6 converges, as $c/N$7, to the deterministic logistic equation

$c/N$8

This law-of-large-numbers regime isolates the macroscopic carrying-capacity effect and suppresses stochastic fluctuations (Cheek, 2020).

A diffusion-scale version is Feller’s branching diffusion with logistic growth,

$c/N$9

This process admits a Ray–Knight representation in terms of a reflected Brownian motion $N$0 whose drift is affine linear in the local time accumulated at its current level. In one formulation,

$N$1

and the stopped local-time profile $N$2 has the law of the logistic diffusion started from $N$3. The microscopic interpretation uses a “pecking order”: when contemporaries compete, the left-most wins and the right-most dies, which turns competition into a local-time drift in the exploration process (Le et al., 2013, Pardoux et al., 2013).

Lamperti-type representations extend this picture. In a Brownian environment,

$N$4

is put in one-to-one correspondence with a generalized Ornstein–Uhlenbeck process $N$5 through a random time change. For logistic continuous-state branching processes with quadratic competition, a generalized Ornstein–Uhlenbeck process again underlies the construction, now combined with branching mechanisms of Lévy–Khintchine type. These pathwise representations are central in extinction, conditioning, and boundary analyses because they transfer questions about the branching process to one-dimensional diffusions or time-changed Lévy-driven dynamics (Leman et al., 2019, Foucart et al., 2024).

3. Extinction, quasi-stationarity, and conditioned dynamics

A basic feature of the classical logistic birth–death process is that the absorbing state $N$6 is reached almost surely in finite time, even when the intrinsic growth rate $N$7 is positive. The quadratic death term dominates at high densities, makes the chain non-explosive, and nevertheless forces eventual extinction. This sharply distinguishes logistic branching from supercritical linear branching, where positive growth can sustain survival with positive probability (Etheridge et al., 2013).

Conditioning on non-extinction therefore becomes a central object. For survival up to a fixed time $N$8, the process admits a finite-time $N$9-transform based on a Wright–Fisher-type diffusion $\mathbb Z_+$0 with

$\mathbb Z_+$1

and

$\mathbb Z_+$2

The conditioned jump rates are

$\mathbb Z_+$3

The Yaglom law is described through the generating function $\mathbb Z_+$4, which solves

$\mathbb Z_+$5

with $\mathbb Z_+$6, $\mathbb Z_+$7, and $\mathbb Z_+$8. As $\mathbb Z_+$9, the conditioned process converges to a $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$0-process with time-homogeneous rates $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$1; this $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$2-process is positive recurrent and reversible, and its time reversal supports an explicit joint generator for total population size and sample genealogy (Etheridge et al., 2013).

For logistic continuous-state branching processes, the conditioning problem takes a different form. Under suitable assumptions, extinction is equivalent to finiteness of the total progeny

$q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$3

Conditioning on non-extinction is implemented by requiring $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$4 to exceed arbitrarily large exponential random variables, which leads to a Doob $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$5-transform with an explicit excessive function $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$6. The conditioned process superposes the original logistic branching dynamics with a density-dependent immigration term, but it still has finite lifetime almost surely: depending on the logarithmic moment of the Lévy measure, the lifetime is either a killing time or a continuous explosion time (Foucart et al., 2024).

Boundary classification for logistic continuous-state branching with competition is controlled by dual diffusions $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$7 and $q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$8, defined by

$q_{i,i+1}=b\,i,\qquad q_{i,i-1}=d\,i+c\,i(i-1),$9

Laplace duality $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$0 and Siegmund duality $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$1 yield criteria for extinction, explosion, regular reflection at $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$2, and the law of the local time accumulated at $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$3 (Foucart, 2021).

4. Genealogies and coalescent limits

The genealogical analysis of logistic branching has recently become precise in the large-equilibrium regime. For the scaled process with typical size of order $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$4, one samples a fixed number of individuals and traces their ancestry backward after the time acceleration

$$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$5

The limit depends on the tail of the offspring distribution, equivalently on the probability that a single individual produces many surviving offspring in a short time interval (Garrett et al., 5 Sep 2025).

The proof uses a modified lookdown construction. Individuals are assigned levels $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$6; for each pair of levels $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$7 there is a Poisson clock of rate $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$8, and when it rings one occupied level dies, encoding logistic death. Births from level $$L f(i)= b\,i\bigl[f(i+1)-f(i)\bigr]+\bigl(d\,i+c\,i(i-1)\bigr)\bigl[f(i-1)-f(i)\bigr].$$9 occur as Poisson processes, individuals at levels above a random insertion point are pushed upward, and a new individual is inserted. Tightness of empirical measures, identification of the limiting martingale problem, and control of large birth sweeps then determine the ancestral partition limit (Garrett et al., 5 Sep 2025).

Three regimes emerge. If the offspring distribution has finite second moment, the ancestral process converges to the Kingman coalescent, with only pairwise mergers surviving. If the tail is regularly varying with index $X_t^N\in\{0,1,2,\dots\}$0,

$X_t^N\in\{0,1,2,\dots\}$1

the limit is the $X_t^N\in\{0,1,2,\dots\}$2 coalescent with

$X_t^N\in\{0,1,2,\dots\}$3

At the boundary case $X_t^N\in\{0,1,2,\dots\}$4, the limit is the Bolthausen–Sznitman coalescent, corresponding to $X_t^N\in\{0,1,2,\dots\}$5. In all three cases the merger rates are

$X_t^N\in\{0,1,2,\dots\}$6

and the partition-valued genealogy converges in $X_t^N\in\{0,1,2,\dots\}$7 (Garrett et al., 5 Sep 2025).

These results correct a common simplification. Logistic regulation stabilizes the population size around a carrying capacity of order $X_t^N\in\{0,1,2,\dots\}$8, and classical neutral reasoning would therefore suggest Kingman genealogies; however, if reproduction is sufficiently bursty, multiple mergers survive the large-$X_t^N\in\{0,1,2,\dots\}$9 limit. The predicted genetic signatures include an excess of very short internal branches and high-frequency derived variants (Garrett et al., 5 Sep 2025).

Related genealogical limits appear in other logistic regimes. If one samples while a density-dependent branching population is still growing superlinearly, the coalescent tree converges to that of a density-independent supercritical branching process, yielding a limiting tree that is universal over exponential, logistic, and Gompertz-type growth models (Cheek, 2020). In two-type logistic branching with mutation and selection, the near-capacity regime converges to a Gillespie–Wright–Fisher diffusion, and the corresponding backward object is an Ancestral Selection Graph whose lineage-counting process is the moment dual of the limiting diffusion (Pra et al., 16 Jul 2025).

5. Multitype and spatial generalizations

In the two-type logistic model with selection, the state is $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$0, births are governed by type-specific offspring measures $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$1, mutations occur at rates $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$2, and each individual suffers logistic death proportional to the total population divided by $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$3. After rescaling by $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$4 in space and time, the frequency process converges to a Wright–Fisher-type SDE with selection and mutation. In the simplified case $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$5 and $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$6,

$$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$7

and the backward lineage count $$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$8 has generator

$$(\mathcal L^N f)(n) = n\lambda\bigl[f(n+1)-f(n)\bigr] + n\mu\bigl[f(n-1)-f(n)\bigr] + \frac{c}{N}n(n-1)\bigl[f(n-1)-f(n)\bigr].$$9

The moment duality is

$b\,i$00

This gives a direct connection between logistic population regulation, selection, and ancestral branching–coalescing structure (Pra et al., 16 Jul 2025).

A mean-field spatial logistic branching system on $b\,i$01 carries particles that reproduce at rate $b\,i$02, die logistically at rate $b\,i$03 at a site with $b\,i$04 particles, and migrate at rate $b\,i$05. Starting from a single particle, one obtains two time windows: an initial collision-free phase described by a Crump–Mode–Jagers process of colonies, and an emergence–equilibration window at time $b\,i$06, when the occupied-site density becomes order $b\,i$07. In the associated Fisher–Wright diffusion with rare mutation, this particle system appears as a dual genealogical structure, so that the logistic branching model and the selected diffusion are, in the paper’s phrase, “two sides of the very same coin” (Dawson et al., 2010).

Spatial logistic branching on lattices exhibits hydrodynamic and fluctuation limits. In the weak-competition regime $b\,i$08, the rescaled density field

$b\,i$09

converges to the Fisher–Kolmogorov–Petrovsky–Piskunov equation

$b\,i$10

and the non-equilibrium fluctuation field converges to a generalized Ornstein–Uhlenbeck process with deterministic heterogeneous coefficients (Tendron, 2024). For ancestral lineages in time-reversed logistic branching random walks, a quenched central limit theorem holds in a high-density regime, established through coarse graining, regeneration times, and coupling arguments (Birkner et al., 2024). In two dimensions, the pair-coalescence time of two ancestral lineages sampled from the stationary regime scales like $b\,i$11, matching the asymptotics of the stepping-stone model and recovering Malécot’s continuous-space formula for identity by descent (Birkner et al., 2024).

6. Model comparison and extinction-time asymptotics

Logistic growth can be encoded either through population-size-dependent offspring laws or through controlled branching. For deterministic-control branching processes, exact equivalence with a population-size-dependent branching process holds if and only if a divisibility condition is satisfied: for each

$b\,i$12

the offspring law $b\,i$13 must be $b\,i$14-divisible. In particular, equivalence holds when $b\,i$15 is infinitely divisible, such as Poisson or Negative-Binomial, or when $b\,i$16 for all relevant $b\,i$17. Even without exact equivalence, first and second moments can be matched by

$b\,i$18

and under regularity assumptions the one-step total-variation distance is bounded by $b\,i$19, with finite-horizon path laws becoming asymptotically indistinguishable as the initial population grows (Braunsteins et al., 2023).

For the logistic birth–death process on $b\,i$20 with

$b\,i$21

the extinction time $b\,i$22 admits a full asymptotic classification. Subcritical regimes yield classical branching-process extinction laws, linear-diffusive scaling, or Gumbel limits depending on the size of $b\,i$23 and the initial condition. Critical regimes lead to the diffusion limit

$b\,i$24

with $b\,i$25 converging to the hitting time of $b\,i$26. Supercritical regimes split into a threshold regime, where there is a positive limiting chance of rapid extinction before reaching the endemic level, and a metastable regime, where the chain approaches $b\,i$27 and then survives for an exponentially long time. In that metastable regime,

$b\,i$28

and

$b\,i$29

(Foxall, 2018).

Taken together, these results show that “logistic branching process” names a broad class of density-regulated branching models rather than a single canonical process. Across discrete, diffusion, continuous-state, multitype, and spatial settings, the common structural feature is quadratic competition; the main differences lie in scaling, state space, and genealogical regime. A plausible implication is that the appropriate logistic model is determined less by the word “logistic” itself than by the intended asymptotic question: carrying-capacity approximation, extinction and quasi-stationarity, diffusion approximation, or ancestral structure.