---
title: Biggins Martingale in Branching Random Walks
url: https://www.emergentmind.com/topics/biggins-martingale
type: topic
---

# Biggins Martingale in Branching Random Walks

The **Biggins martingale** is the additive martingale associated with a supercritical branching random walk, defined by an exponential tilt of the generation-\(n\) point process and a normalization by its Laplace transform. In the real-parameter setting it is typically written
\[
W_n(t)=e^{-n\psi(t)}\sum_{|x|=n} e^{-tV(x)},
\qquad
\psi(t)=\log E\Big[\sum_{|x|=1}e^{-tV(x)}\Big],
\]
while for a complex parameter \(\lambda\) it takes the form
\[
W_n(\lambda)=m(\lambda)^{-n}\sum_{|u|=n}e^{-\lambda S(u)},
\qquad
m(\lambda)=E\Big[\sum_{|u|=1}e^{-\lambda S(u)}\Big].
\]
It is a central object in the study of branching random walks, where it governs growth, change of measure, extremal behavior, and stochastic fixed-point equations; later work extends the same structure to complex parameters, branching Lévy processes, and matrix branching random walks [1102.0217], [1903.00524], [1712.04769], [2507.09737].

## 1. Definition and basic framework

A branching random walk on \(\mathbb R\) starts with one ancestor at the origin. Each particle produces offspring whose locations relative to the parent are given by a point process, and the offspring then reproduce independently according to the same law. If \(V(x)\) denotes the position of a particle \(x\) in generation \(|x|=n\), then the additive martingale at parameter \(t\) is
\[
W_n(t)=e^{-n\psi(t)}\sum_{|x|=n}e^{-tV(x)},
\qquad
\psi(t)=\log E\Big[\sum_{|x|=1}e^{-tV(x)}\Big].
\]
When the critical parameter is normalized to \(t=1\) and \(\psi(1)=0\), this reduces to
\[
W_n=\sum_{|x|=n}e^{-V(x)}.
\]
Conditioning on the first \(n\) generations shows that \((W_n(t))_{n\ge 0}\) is a martingale with respect to the natural filtration, and for \(t=1\) one has \(E[W_n]=1\) [1102.0217].

The same construction admits a complex extension. For \(\lambda=\theta+i\gamma\) with \(m(\lambda)\neq 0\),
\[
Z_n(\lambda)=\frac{1}{m(\lambda)^n}\sum_{|u|=n}e^{-\lambda S(u)}
\]
is a complex-valued martingale with mean \(1\). The real and complex theories share the same branching structure, but the complex case loses positivity and therefore requires substantially finer moment and spectral arguments [1903.00524].

The martingale can also be embedded in a weighted branching process. Writing weights \(T_j\) on offspring and
\[
L(\emptyset)=1,\qquad L(vi)=T_i(v)L(v),
\]
one obtains the generation sum
\[
Z_n=\sum_{|v|=n}L(v).
\]
For branching random walks with complex parameter \(\lambda\), the identification
\[
T_j=\mathfrak m(\lambda)^{-1}e^{-\lambda S(j)}
\]
gives
\[
Z_n=W_n(\lambda),
\]
so Biggins’ martingale is a particular instance of a weighted branching martingale and of the smoothing-transform framework [1804.02209].

## 2. Convergence, non-degeneracy, and \(L^p\) behavior

For real parameters, classical convergence questions ask whether the additive martingale converges almost surely, whether the limit is nontrivial, and under which conditions convergence holds in \(L^p\). In the real-parameter case recovered in the complex-parameter study, \(Z_n(\theta)\) converges in \(L^p\) for \(p>1\) if and only if
\[
E[Z_1(\theta)^p]<\infty,
\qquad
\frac{m(p\theta)}{m(\theta)^p}<1,
\]
which is the standard Biggins-type criterion for the nonnegative martingale [1903.00524].

For genuinely complex parameters, the picture splits according to whether \(|m(\lambda)|=m(\theta)\) or \(0<|m(\lambda)|<m(\theta)\). In the first case, the martingale reduces essentially to the real one. In the second, necessary and sufficient conditions for \(L^p\)-convergence are more involved. For \(p\ge 2\), one obtains a criterion in terms of the first-step moment \(E|Z_1(\lambda)|^p\), a further condition \(E[Z_1(2\theta)]^{p/2}<\infty\) when \(p>2\), and a spectral inequality involving \(m(2\theta)\), \(m(p\theta)\), and \(|m(\lambda)|\) [1903.00524]. For \(p\in(1,2)\), the theory requires either a second-moment assumption or a regularly varying tail assumption for \(|Z_1(\lambda)|\), together with a Seneta–Heyde-type analysis of the corresponding real martingale [1903.00524].

In continuous time, branching Lévy processes provide the natural analogue of branching random walks. Their Biggins martingale is
\[
W_t(\theta)=e^{-t\kappa(\theta)}(Z_t,e_\theta)
          =e^{-t\kappa(\theta)}\sum_{u\in\mathcal N(t)}e^{\theta X_u(t)},
\]
where \(\kappa\) is the cumulant generating function expressed through the branching Lévy triplet \((\sigma^2,a,\Lambda)\). A version of Biggins’ theorem holds in this setting: \(W_t(\theta)\) is uniformly integrable if and only if
\[
\theta\kappa'(\theta)<\kappa(\theta)
\]
and
\[
\int_{\mathcal P}(x,e_\theta)\big(\log (x,e_\theta)-1\big)^+\,\Lambda(dx)<\infty;
\]
otherwise \(W_\infty=0\) almost surely [1712.04769]. For \(1<p\le 2\), later work gives ultimate necessary and sufficient conditions for \(L_p\)-convergence in terms of
\[
\kappa(p\theta)<p\kappa(\theta)
\]
and an explicit integrability condition on the offspring Lévy measure [1811.08721].

A recurrent misunderstanding is to regard nontrivial martingale convergence as generic. The literature instead separates a supercritical \(L^1\)-convergent regime, a critical boundary regime where the additive martingale degenerates, and heavy-tail or complex regimes where different normalizations become necessary [1712.04769], [1102.0217].

## 3. Boundary case and derivative martingale

The most delicate real-parameter regime is the **boundary case** in the sense of Biggins–Kyprianou. After normalization to the critical parameter \(t=1\), it is characterized by
\[
E\Big[\sum_{|x|=1}e^{-V(x)}\Big]=1,
\qquad
E\Big[\sum_{|x|=1}V(x)e^{-V(x)}\Big]=0,
\]
together with
\[
\sigma^2:=E\Big[\sum_{|x|=1}V(x)^2e^{-V(x)}\Big]\in(0,\infty).
\]
In this regime the additive martingale
\[
W_n=\sum_{|x|=n}e^{-V(x)}
\]
still converges almost surely, but the limit is degenerate:
\[
W_n\to 0
\quad \text{a.s.}
\]
on the survival event [1102.0217], [1606.03211].

The nontrivial object is the **derivative martingale**
\[
D_n=\sum_{|x|=n}V(x)e^{-V(x)}.
\]
Biggins and Kyprianou proved that under the boundary assumptions and logarithmic moment conditions,
\[
D_n\to D_\infty
\quad \text{a.s.},
\]
where \(D_\infty\) is finite and strictly positive on nonextinction [1102.0217], [1606.03211]. In this setting the derivative martingale is literally the derivative, up to sign, of the additive martingale with respect to the parameter at the critical point.

Chen identified a sharp necessary and sufficient condition for the derivative martingale limit to be nontrivial. With
\[
Y=\sum_{|u|=1}e^{-V(u)},
\qquad
Z=\sum_{|u|=1}V(u)^+e^{-V(u)},
\]
the limit \(D_\infty\) is nontrivial if and only if
\[
E\bigl[Z\log_+ Z + Y(\log_+ Y)^2\bigr]<\infty.
\]
This is a Kesten–Stigum-like criterion for the boundary derivative martingale and pins down exactly when the critical Mandelbrot cascade does not collapse [1402.5864].

A second major result is the **Seneta–Heyde scaling**. Under the boundary assumptions and the moment conditions
\[
E\Big[\sum_{|x|=1}V(x)^2e^{-V(x)}\Big]<\infty,
\qquad
E[X\log_+^2X]<\infty,
\qquad
E[\widetilde X\log_+\widetilde X]<\infty,
\]
with
\[
X=\sum_{|x|=1}e^{-V(x)},
\qquad
\widetilde X=\sum_{|x|=1}V(x)^+e^{-V(x)},
\]
Aïdékon and Shi proved
\[
n^{1/2}W_n
\xrightarrow{P^*}
\left(\frac{2}{\pi\sigma^2}\right)^{1/2}D_\infty.
\]
The convergence is in probability under the survival-conditioned law \(P^*\), not almost surely; in fact,
\[
\limsup_{n\to\infty} n^{1/2}W_n=\infty
\quad P^*\text{-a.s.}
\]
This separates the additive and derivative martingales sharply: the additive martingale dies out, but at rate \(n^{-1/2}\) it is asymptotically proportional to the derivative martingale [1102.0217].

The derivative martingale also governs extreme-value asymptotics. Its tail satisfies
\[
\lim_{x\to\infty} x\,P(D_\infty\ge x)=c_{D_\infty}>0,
\]
and the proof relates \(D_\infty\) to the global minimum \(M\) of the branching random walk through the factorization
\[
D_\infty=e^{-M}\,\mathfrak D^M.
\]
This connects large derivative-martingale values to very deep minima of the branching random walk [1606.03211].

## 4. Complex parameters, smoothing transforms, and absolute continuity

For complex \(\lambda\), Biggins’ martingale is
\[
W_n(\lambda)=\mathfrak m(\lambda)^{-n}\sum_{|v|=n}e^{-\lambda S(v)}.
\]
In the weighted branching formulation, the corresponding limit \(Z\) solves the **complex smoothing equation**
\[
Z\overset{d}{=}\sum_{j\ge1}T_j Z_j,
\]
where \(Z_j\) are i.i.d. copies of \(Z\), independent of the weights. The characteristic function then satisfies
\[
\phi(\xi)=E\Big[\prod_{j=1}^N \phi(\overline{T_j}\xi)\Big].
\]
This places the limit of Biggins’ martingale within the general theory of stochastic fixed points and operator-stable decompositions [1804.02209].

A structural theorem recalled in the same setting states that, under assumptions \((A1)\)–\((A3)\) and technical hypotheses with \(\alpha\neq1\), any fixed point law can be written
\[
X\overset{d}{=}Y_W+xZ,
\]
where \(W\) is the nonnegative martingale limit associated with \(\sum |L(v)|^\alpha\), \(Z\) is the martingale limit \(\lim Z_n\), and \(Y_W\) is a Lévy component with operator-stable invariance. In finite-variance branching-random-walk applications the Lévy part vanishes, leaving \(X=xZ\); in that sense the limit of Biggins’ martingale is the martingale-type solution of the smoothing transform [1804.02209].

Damek and Mentemeier gave a simple criterion for the **absolute continuity** of the limit law. In the genuinely complex case, assume:
\[
N>0\ \text{a.s.},\quad
\alpha\in(1,2),\quad
Z_n\to Z\ \text{a.s. and in }L^1,\quad
\operatorname{supp}(Z)\not\subset\mathbb R,
\]
together with
\[
E[N^2]<\infty,
\qquad
E\Big[N\sum_{j=1}^N \log_+|T_j|\Big]<\infty.
\]
Then the law of \(Z\) is absolutely continuous on \(\mathbb C\). If \(\operatorname{supp}(Z)\subset\mathbb R\), the condition \(E[N]<\infty\) suffices in place of the stronger complex assumption [1804.02209].

For branching random walks, the specialization \(T_j=\mathfrak m(\lambda)^{-1}e^{-\lambda S(j)}\) identifies this result directly with the limit \(W(\lambda)\) of Biggins’ martingale. The paper emphasizes that “for given values of \(\lambda\), the assumptions of Theorem \(\ref{thm:continuity of Z}\) are readily checked,” and in the symmetric binary example computes explicitly
\[
\mathfrak m(\lambda)=2\cosh(\lambda),
\qquad
m(s)=\frac{2}{2^s}\frac{\cosh(s\,\Re(\lambda))}{|\cosh(\lambda)|^s}.
\]
Thus, under suitable spectral and logarithmic moment conditions, the limit law of Biggins’ martingale with complex parameter has a density with respect to Lebesgue measure on \(\mathbb C\) [1804.02209].

## 5. Fluctuation theory

Beyond convergence, a substantial literature analyzes the rate and mode of convergence of Biggins’ martingale to its limit. In the square-integrable regime, Iksanov and Kabluchko proved a functional central limit theorem for the tail process. Under the normalization \(m(1)=1\), the assumptions
\[
m(2)<1,
\qquad
\sigma^2=\mathrm{Var}(W_1(1))<\infty
\]
imply
\[
\left(
\frac{W_\infty(1)-W_{n+r}(1)}{m(2)^{(n+r)/2}}
\right)_{r\in\mathbb N_0}
\xrightarrow{d}
\left(\sqrt{v^2W_\infty(2)}\,U_r\right)_{r\in\mathbb N_0},
\]
where
\[
v^2=\frac{\sigma^2}{1-m(2)},
\]
and \((U_r)\) is a stationary centered Gaussian sequence with covariance
\[
\operatorname{Cov}(U_r,U_s)=m(2)^{|r-s|/2}.
\]
They also proved a law of the iterated logarithm:
\[
\limsup_{n\to\infty}
\frac{W_\infty(1)-W_n(1)}{m(2)^{n/2}\sqrt{2\log n}}
=
\sqrt{2}\,v\,\sqrt{W_\infty(2)}
\quad \text{a.s.},
\]
with the corresponding negative liminf [1507.08458].

In a heavy-tailed regime, the fluctuations cease to be Gaussian. If \(W_1(0)\) belongs to the domain of normal attraction of an \(\alpha\)-stable law, with
\[
P(W_1(0)>x)\sim c x^{-\alpha},
\qquad
\alpha\in(1,2),
\]
and
\[
K:=\frac{m(\alpha\theta)}{m(\theta)}<1
\]
in the notation of the paper, then the normalized tail process satisfies
\[
\big(K^{-(n-r)/\alpha}(W(0)-W_{n-r}(0))\big)_{r\in\mathbb N_0}
\xRightarrow[\mathrm{f.d.d.}]{}
\big(W(0)^{1/\alpha}U_r\big)_{r\in\mathbb N_0},
\]
where \((U_r)\) is a stationary autoregressive process of order one with \(\alpha\)-stable marginals. This gives a stable analogue of the Gaussian tail-process limit theorem [1709.07362].

For complex parameters, the fluctuation picture is richer still. One study identifies **three different regimes**. First, for parameters with small absolute values, fluctuations are Gaussian and the limit laws are scale mixtures of the real or complex standard normal laws. Second, there is a region of parameter space in which the fluctuations are determined by the extremal positions in the branching random walk. Third, on a critical region, typically on the boundary of the set of parameters for which the martingales converge to non-degenerate limits, the fluctuations are stable-like and the limit laws are those of randomly stopped Lévy processes satisfying invariance properties similar to stability [1806.09943].

The same work also makes explicit that in the extremal regime the relevant normalization involves the minimal-position scale, while on the stable boundary \(\partial\Lambda^{(1,2)}\) the fluctuations are described by a Lévy process \(X\) evaluated at a random time proportional to the derivative-martingale limit \(D_\infty\). This links additive-martingale fluctuations to both extremal process theory and smoothing-transform tail theory [1806.09943].

## 6. Continuous-time and higher-dimensional generalizations

The continuous-time analogue of Biggins’ martingale arises in **branching Lévy processes**. Such a process is determined by a characteristic triplet \((\sigma^2,a,\Lambda)\), and the additive martingale is
\[
W_t(\theta)=e^{-t\kappa(\theta)}(Z_t,e_\theta).
\]
A branching-Lévy version of Biggins’ theorem gives necessary and sufficient conditions for a nondegenerate \(L^1\) limit, and an \(L^p\)-criterion for \(p\in(1,2]\) is later derived through Lévy-type perpetuities [1712.04769], [1811.08721]. A 2025 study continues this line by analyzing moment properties of \(W_t(\theta)\) and \(W_\infty(\theta)\), the tail behavior of \(W_\infty(\theta)\), and central limit theorems for \(W_t(\theta)-W_\infty(\theta)\) [2509.09188].

The same general program extends beyond scalar spatial positions. In a **matrix branching random walk** on the semigroup of nonnegative matrices, the analogue of Biggins’ additive martingale is
\[
W_n(s)=\frac{1}{\mathfrak m(s)^n}\sum_{|u|=n}e^{-sS_u}\,r_s(X_u),
\]
where \(r_s\) is the positive eigenfunction of the transfer operator \(P_s\) and \(S_u\) is the logarithmic cocycle of the matrix product along the branch. Under general assumptions, a full analogue of Biggins’ martingale convergence theorem holds: for fixed \(x\),
\[
E_x(W_\infty(s))=1
\iff
W_\infty(s)>0
\iff
\mathfrak M(s)>s\mathfrak M'(s)
\ \text{and}\ 
E_x\big(W_1(s)\log^+W_1(s)\big)<\infty.
\]
The matrix theory also has a derivative martingale
\[
D_n=\sum_{|u|=n}(S_u+\ell_\alpha(X_u))e^{-\alpha S_u}r_\alpha(X_u),
\]
which converges to \(D_\infty\), and a Seneta–Heyde scaling
\[
\sqrt n\,W_n
\xrightarrow{\mathbf P_x}
\left(\frac{2}{\pi\sigma_\alpha^2}\right)^{1/2}D_\infty.
\]
This generalization replaces scalar spatial homogeneity by spectral theory for products of random nonnegative matrices and a centered Markov random walk on projective space [2507.09737].

A plausible implication is that Biggins’ martingale should be viewed less as a single formula than as a template: a normalized multiplicative martingale built from a branching system and a spectral transform. In scalar branching random walks, complex weighted branching processes, branching Lévy processes, and matrix branching random walks alike, the decisive structures are the same: a Laplace or transfer operator, a martingale change of measure, a boundary regime with vanishing additive martingale, and a derivative martingale that captures the critical first-order mass [1804.02209], [1712.04769], [2507.09737].

Source: https://www.emergentmind.com/topics/biggins-martingale