Seneta–Heyde Scaling in Stochastic Processes

• Seneta–Heyde scaling is a critical normalization method for branching-type martingales and chaos measures, ensuring convergence to nondegenerate limits via derivative martingales.
• It underpins various models—including branching random walks, super-Brownian motion, and Gaussian multiplicative chaos—by adapting normalization factors to specific fluctuation regimes.
• The approach employs truncated martingales and change-of-measure techniques to manage degenerate additive processes, highlighting its broad applicability in stochastic analysis.

Seneta–Heyde scaling is a critical normalization principle for branching-type martingales and multiplicative-chaos measures. It describes the regime in which the natural additive object converges to $0$, but a sharper deterministic renormalization restores a nondegenerate limit, typically proportional to a derivative martingale or derivative-normalized critical measure. In the modern literature this phenomenon appears in supercritical Galton–Watson and continuous-state branching settings, branching random walk, super-Brownian motion, homogeneous fragmentation, Gaussian multiplicative chaos, Brownian multiplicative chaos, and more recent matrix-valued branching models (Aidekon et al., 2011, Kyprianou et al., 2015, Powell, 2017, Hou et al., 2021, Grama et al., 13 Jul 2025).

1. Classical template and general meaning

The classical Seneta–Heyde theorem concerns rescaling a martingale or population size when the standard normalization is degenerate. In the super-Brownian setting with a general branching mechanism, this classical picture appears through an embedded supercritical continuous-state branching process obtained from Dynkin exit measures. For every $\lambda\in(0,\underline\lambda]$, there exists a slowly varying function $L_\lambda$, slowly varying at $0$, such that

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$

where $\Delta(\lambda)$ is a nondegenerate nonnegative random variable and $\{\Delta(\lambda)=0\}=\mathcal E$ almost surely. In the same framework, travelling waves are represented by

$$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$

and the wave tail is asymptotic to $e^{-\lambda x}L_\lambda(e^{-\lambda x})$; at the critical speed one has $L_{\underline\lambda}(x)\sim -\log x$ under the stated $\lambda\in(0,\underline\lambda]$0 condition (Kyprianou et al., 2010).

The same idea extends beyond genuine branching processes. In branching-within-branching, the contaminated-cell count $\lambda\in(0,\underline\lambda]$1 is not itself a Galton–Watson process, yet a proper Heyde–Seneta norming still exists. Under

$\lambda\in(0,\underline\lambda]$2

there is a sequence $\lambda\in(0,\underline\lambda]$3 such that

$\lambda\in(0,\underline\lambda]$4

with $\lambda\in(0,\underline\lambda]$5. This shows that Seneta–Heyde normalization can govern derived observables whose dynamics are inherited from a branching environment rather than given by a branching law directly (Alsmeyer et al., 2015).

2. Boundary branching random walk

The canonical critical formulation is the boundary case of a one-dimensional supercritical branching random walk. Under

$\lambda\in(0,\underline\lambda]$6

the additive martingale

$\lambda\in(0,\underline\lambda]$7

converges almost surely to $\lambda\in(0,\underline\lambda]$8, while the derivative martingale

$\lambda\in(0,\underline\lambda]$9

converges almost surely to a positive limit $L_\lambda$0 on survival under the stated moment assumptions. The sharp Seneta–Heyde theorem identifies the exact normalization: $L_\lambda$1 where

$L_\lambda$2

The same work proves that this convergence is not almost sure: $L_\lambda$3 The mechanism is controlled by a centered associated random walk, its renewal function $L_\lambda$4, and the identity

$L_\lambda$5

where $L_\lambda$6 and $L_\lambda$7 (Aidekon et al., 2011).

A later revisitation streamlined the proof under optimal assumptions. Its main replacements are “certain second moment estimates by truncated first moment bounds” and “ballot-type theorems for random walks by estimates coming from an explicit expression for the potential kernel of random walks killed below the origin.” It also gives “a criterion for convergence in probability of non-negative random variables in terms of conditional Laplace transforms” (Boutaud et al., 2019).

Stable regimes require different normalizations. When the associated random walk is in the domain of attraction of a spectrally positive $L_\lambda$8-stable law with $L_\lambda$9, the critical additive martingale satisfies

$0$0

again in probability and not almost surely. In a broader $0$1-stable-spine setting, the classical derivative martingale can fail entirely: if $0$2, then $0$3 almost surely, and the correct replacement is built from the descending ladder-height renewal function. The exact martingale is

$0$4

with asymptotic power-law counterpart

$0$5

and the Seneta–Heyde scaling becomes

$0$6

These results show that the derivative object itself is model-dependent, even when the critical renormalization principle persists (He et al., 2016, Boutaud et al., 2020).

3. Continuous-time and measure-valued analogues

In one-dimensional supercritical super-Brownian motion with general branching mechanism

$0$7

the critical parameter is

$0$8

At $0$9, the additive martingale $$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$0 dies out, but

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$1

converges in probability to a positive limit that is a constant multiple of the derivative martingale limit $$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$2, with exact constant $$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$3. On the survival event one also has

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$4

The proof combines exit measures, truncated additive and derivative martingales, a change of measure driven by a positive truncated derivative martingale, a Bessel-3-type spine, and a skeleton/backbone decomposition. The paper explicitly emphasizes that the Seneta–Heyde phenomenon “is not tied to discrete genealogies” and persists in a “measure-valued superprocess setting with general branching” (Hou et al., 2021).

Homogeneous fragmentation furnishes a continuous-time analogue of branching random walk with nonnegative displacements. For

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$5

the critical point $$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$6 is characterized by

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$7

At $$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$8, one has

$$\lim_{x\to\infty} e^{-\lambda x} L_\lambda(e^{-\lambda x})\, Z_x^{c_\lambda} = \Delta(\lambda),$$9

while the derivative martingale

$\Delta(\lambda)$0

converges almost surely to a strictly positive limit. The Seneta–Heyde scaling is

$\Delta(\lambda)$1

where $\Delta(\lambda)$2. The proof is organized by truncation, many-to-one for the tagged fragment, survival asymptotics for a centered Lévy process above a barrier, and $\Delta(\lambda)$3 concentration estimates (Kyprianou et al., 2015).

4. Critical multiplicative chaos

In critical Gaussian multiplicative chaos, the positive critical approximation vanishes and the deterministic renormalization is logarithmic rather than polynomial. For a log-correlated field regularized by convolution,

$\Delta(\lambda)$4

vanishes at criticality, and the Seneta–Heyde rescaling is

$\Delta(\lambda)$5

For the $\Delta(\lambda)$6 GFF and for $\Delta(\lambda)$7-scale invariant kernels, one has

$\Delta(\lambda)$8

so derivative normalization and Seneta–Heyde renormalization produce the same critical measure up to the explicit deterministic factor $\Delta(\lambda)$9 (Powell, 2017).

Brownian multiplicative chaos exhibits the same structure in a non-Gaussian environment. At $\{\Delta(\lambda)=0\}=\mathcal E$0,

$\{\Delta(\lambda)=0\}=\mathcal E$1

and the Seneta–Heyde normalization multiplies by an extra $\{\Delta(\lambda)=0\}=\mathcal E$2: $\{\Delta(\lambda)=0\}=\mathcal E$3 This converges to a critical measure $\{\Delta(\lambda)=0\}=\mathcal E$4, while the derivative approximation $\{\Delta(\lambda)=0\}=\mathcal E$5 converges to $\{\Delta(\lambda)=0\}=\mathcal E$6, with

$\{\Delta(\lambda)=0\}=\mathcal E$7

The analysis replaces Gaussian tools by a Bessel-process representation of radial local times, stochastic-calculus identities, barrier estimates, and a continuity lemma (Jego, 2020).

A local-set formulation for the $\{\Delta(\lambda)=0\}=\mathcal E$8 GFF gives a cascade version of the same phenomenon. With first passage sets $\{\Delta(\lambda)=0\}=\mathcal E$9, the critical additive approximation

$$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$0

satisfies

$$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$1

while the derivative martingale $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$2 converges almost surely to $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$3. This realizes Liouville measure as a multiplicative cascade via local sets and adapts the Aïdékon–Shi strategy to a GFF setting (Aru et al., 2017).

5. Structural proof mechanisms

Across models, Seneta–Heyde scaling is typically proved by replacing the original additive object with a truncated version that excludes trajectories crossing a lower barrier, then tilting the law by a positive derivative-type martingale. In branching random walk this produces $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$4 and $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$5, with a spine distributed as the associated random walk conditioned to stay in $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$6; the key ratio asymptotic is

$$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$7

in probability (Aidekon et al., 2011).

The same architecture reappears in super-Brownian motion through exit measures from

$$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$8

truncated martingales $$\Phi_{c_\lambda}(x) = -\log \mathbb E\Big[\exp\{-e^{-\lambda x}\Delta(\lambda)\}\Big],$$9 and $e^{-\lambda x}L_\lambda(e^{-\lambda x})$0, and the change of measure

$e^{-\lambda x}L_\lambda(e^{-\lambda x})$1

under which the spine is Bessel-3-type and immigration occurs both continuously and via the Lévy measure $e^{-\lambda x}L_\lambda(e^{-\lambda x})$2 (Hou et al., 2021).

In critical GMC, the main tools are cutoff measures $e^{-\lambda x}L_\lambda(e^{-\lambda x})$3 and $e^{-\lambda x}L_\lambda(e^{-\lambda x})$4, a rooted or Peyrière measure $e^{-\lambda x}L_\lambda(e^{-\lambda x})$5, and asymptotic Bessel behavior of the scale process; the core comparison is the ratio asymptotic

$e^{-\lambda x}L_\lambda(e^{-\lambda x})$6

for each fixed $e^{-\lambda x}L_\lambda(e^{-\lambda x})$7 and $e^{-\lambda x}L_\lambda(e^{-\lambda x})$8 (Powell, 2017). In critical Brownian multiplicative chaos, analogous “good events” and Bessel-bridge interpolation fields provide $e^{-\lambda x}L_\lambda(e^{-\lambda x})$9 and $L_{\underline\lambda}(x)\sim -\log x$0 control in the absence of Kahane’s inequality (Jego, 2020). For fragmentation, the common elements are again truncation, many-to-one, one-particle barrier asymptotics, and $L_{\underline\lambda}(x)\sim -\log x$1 concentration (Kyprianou et al., 2015).

6. Convergence mode, universality, and recent extensions

A recurrent misconception is that Seneta–Heyde normalization should yield an almost sure asymptotic. The canonical results say otherwise. In boundary branching random walk,

$L_{\underline\lambda}(x)\sim -\log x$2

and in critical super-Brownian motion,

$L_{\underline\lambda}(x)\sim -\log x$3

Thus the normalization identifies a first-order probabilistic scale, not a pathwise limit theorem (Aidekon et al., 2011, Hou et al., 2021).

At the same time, the scope of the phenomenon is unusually broad. The same critical structure occurs with $L_{\underline\lambda}(x)\sim -\log x$4 in finite-variance branching random walk, $L_{\underline\lambda}(x)\sim -\log x$5 or $L_{\underline\lambda}(x)\sim -\log x$6 in stable-spine variants, $L_{\underline\lambda}(x)\sim -\log x$7 in homogeneous fragmentation and super-Brownian motion, and $L_{\underline\lambda}(x)\sim -\log x$8 in critical multiplicative chaos (He et al., 2016, Boutaud et al., 2020, Kyprianou et al., 2015, Powell, 2017). This suggests that Seneta–Heyde scaling is best viewed as a critical normalization principle whose exact form is dictated by the fluctuation theory of the tilted one-particle process.

Recent work extends this structure to matrix branching random walks on the semigroup of nonnegative matrices. Under general assumptions it derives an analogue of Biggins’ martingale convergence theorem for the additive martingale $L_{\underline\lambda}(x)\sim -\log x$9, a spinal decomposition theorem, convergence of the derivative martingale $\lambda\in(0,\underline\lambda]$00, and a Seneta–Heyde scaling theorem in the boundary case, together with explicit duality results for the renewal measure of centered Markov random walks (Grama et al., 13 Jul 2025). The direction of development is therefore not merely toward new proofs, but toward new state spaces and new one-particle fluctuation theories.