Matrix Branching Random Walk

• Matrix branching random walk is a model where branching dynamics are coupled with matrix products and mean operators to analyze particle evolution.
• It applies transfer operators, multitype formulations, and random environments to yield insights into spectral regimes and survival criteria.
• The method employs martingale techniques, spine decompositions, and derivative scaling to capture critical transitions and asymptotic behaviors.

Searching arXiv for recent and foundational papers on matrix branching random walk and related operator/multitype formulations. Matrix branching random walk denotes several closely related constructions in which branching dynamics are coupled to matrix-valued or matrix-governed evolution. In one line of work, it refers to a branching random walk on the semigroup of nonnegative matrices, where a particle’s position is the product of random matrices along its ancestral line and the associated scalar displacement is obtained from the logarithm of a matrix norm (Grama et al., 13 Jul 2025). In another line, it denotes multitype or operator-based branching structures whose expected evolution is encoded by mean matrices or infinite-dimensional generators (Hong et al., 2010, Hong et al., 2012, Filichkina et al., 2023, Bertacchi et al., 2015). A further, more interpretive usage appears in random-matrix theory, where a scalar branching random walk serves as a proxy model for characteristic polynomials of random unitary matrices, so that “matrix branching random walk” means a branching model whose statistics mirror a matrix object rather than a matrix-valued branching process itself (Bailey et al., 2020). Across these settings, the unifying theme is that branching is analyzed through matrix products, matrix-valued mean offspring operators, transfer operators, or matrix-recursive decompositions.

1. Matrix-valued and matrix-governed formulations

The most literal matrix branching random walk is formulated on the semigroup of nonnegative matrices. Fix $d \ge 2$, let $\mathbb{M}$ be the semigroup of $d\times d$ nonnegative allowable matrices, and let $\mathbb{M}_+\subset\mathbb{M}$ be the subset of matrices with all entries strictly positive. For $g\in\mathbb{M}$ and $x\in\mathbb{S}_+^{d-1}$, the projective action is $g\cdot x = gx/\|gx\|$, and the cocycle is $\sigma(g,x)=\log\|gx\|$. A Galton–Watson tree is decorated by i.i.d. point processes of matrices, and the position of a particle $u$ is the left product $G_u$ of matrices along its ancestral line. The associated branching Markov chain is then $\mathbb{M}$0, where $\mathbb{M}$1 and $\mathbb{M}$2 (Grama et al., 13 Jul 2025).

A broader usage treats a branching random walk as matrix-governed when its mean evolution is encoded by a matrix or operator. For a discrete-time branching random walk on a countable state space $\mathbb{M}$3, the expected number of children sent from $\mathbb{M}$4 to $\mathbb{M}$5 is $\mathbb{M}$6, defining the first-moment matrix $\mathbb{M}$7. The multidimensional generating function $\mathbb{M}$8 has Jacobian $\mathbb{M}$9 at $d\times d$0, so the linearization of the nonlinear branching dynamics near extinction is controlled by a matrix operator (Bertacchi et al., 2015). In continuous-time models with one branching center at the origin and absorption elsewhere, the first moment evolves under the operator

$d\times d$1

where $d\times d$2 is the generator of the underlying random walk, $d\times d$3 is rank one, and $d\times d$4 is the uniform killing term (Filichkina et al., 2023).

A different matrix-governed formulation appears when a random walk is re-expressed as a multitype branching process. For bounded-jump random walk on $d\times d$5, the excursion counts before the first ladder time form a 9-type non-homogeneous branching process whose expectations are propagated by level-dependent $d\times d$6 mean offspring matrices $d\times d$7 (Hong et al., 2010). For random walk on a strip in a random environment, the down-step counts between layers form an inhomogeneous multitype branching process in a random environment, with mean matrices

$d\times d$8

derived from the transition matrices $d\times d$9 and exit-probability matrices $\mathbb{M}_+\subset\mathbb{M}$0 (Hong et al., 2012).

2. Genealogical dynamics, products, and transfer operators

In the matrix-semigroup model, the central analytic object is the transfer operator

$\mathbb{M}_+\subset\mathbb{M}$1

defined for $\mathbb{M}_+\subset\mathbb{M}$2 in an interval $\mathbb{M}_+\subset\mathbb{M}$3 on continuous functions over $\mathbb{M}_+\subset\mathbb{M}$4. This operator has a positive eigenfunction $\mathbb{M}_+\subset\mathbb{M}$5, an eigenmeasure $\mathbb{M}_+\subset\mathbb{M}$6, and a dominant eigenvalue $\mathbb{M}_+\subset\mathbb{M}$7, with $\mathbb{M}_+\subset\mathbb{M}$8 (Grama et al., 13 Jul 2025). The additive martingale is

$\mathbb{M}_+\subset\mathbb{M}$9

which is the matrix analogue of Biggins’ additive martingale.

The corresponding many-to-one formula replaces a sum over particles by expectation under a tilted Markov random walk $g\in\mathbb{M}$0. Under $g\in\mathbb{M}$1, the pair $g\in\mathbb{M}$2 is a Markov random walk on $g\in\mathbb{M}$3, and the cocycle $g\in\mathbb{M}$4 supplies the additive component. This is the technical mechanism by which products of random matrices are converted into one-particle dynamics amenable to renewal analysis and martingale methods (Grama et al., 13 Jul 2025).

Operator formulations in non-matrix-valued models have a parallel structure. In the single-source branching random walk with absorption, the first moment satisfies

$g\in\mathbb{M}$5

so $g\in\mathbb{M}$6 evolves by the semigroup $g\in\mathbb{M}$7. After the change of variables $g\in\mathbb{M}$8, the spectral analysis reduces to the rank-one perturbation $g\in\mathbb{M}$9 (Filichkina et al., 2023). In multidimensional generating-function language, the full nonlinear dynamics are encoded by $x\in\mathbb{S}_+^{d-1}$0, while the mean matrix $x\in\mathbb{S}_+^{d-1}$1 controls spectral quantities such as

$x\in\mathbb{S}_+^{d-1}$2

which govern local and global survival in important classes of branching random walks (Bertacchi et al., 2015).

This suggests a common pattern: matrix branching random walk theory is organized around a reduction from a branching tree of matrix products or multitype transitions to transfer operators, mean matrices, or Markov additive processes. The scalar displacement is retained, but only after augmenting the state by projective direction, type, or environment.

3. Martingales, spine constructions, and critical scaling

The matrix-semigroup model supports a full analogue of classical spine theory. In the boundary case, one assumes the existence of $x\in\mathbb{S}_+^{d-1}$3 such that

$x\in\mathbb{S}_+^{d-1}$4

Then the additive martingale $x\in\mathbb{S}_+^{d-1}$5 converges almost surely to zero on survival, while a derivative martingale can be defined by

$x\in\mathbb{S}_+^{d-1}$6

where $x\in\mathbb{S}_+^{d-1}$7 is the corrector associated with the centered Markov random walk under $x\in\mathbb{S}_+^{d-1}$8 (Grama et al., 13 Jul 2025).

The size-biased measure constructed from $x\in\mathbb{S}_+^{d-1}$9 yields a spine $g\cdot x = gx/\|gx\|$0 such that the process along the spine has the same law as $g\cdot x = gx/\|gx\|$1 under $g\cdot x = gx/\|gx\|$2. Off-spine subtrees remain independent copies of the original matrix branching random walk, while the spine offspring law is size-biased. This is the matrix analogue of the Lyons spinal decomposition and is the basis for the convergence theory of $g\cdot x = gx/\|gx\|$3 and $g\cdot x = gx/\|gx\|$4 (Grama et al., 13 Jul 2025).

Under Conditions A1*, A2, A3, A4, and A5 in that work, the derivative martingale converges almost surely to a nonnegative limit $g\cdot x = gx/\|gx\|$5, and with the extra moment condition it is strictly positive on the survival event. The boundary-case Seneta–Heyde scaling then takes the form

$g\cdot x = gx/\|gx\|$6

where $g\cdot x = gx/\|gx\|$7 is the variance of the centered Markov random walk under $g\cdot x = gx/\|gx\|$8 (Grama et al., 13 Jul 2025).

A distinct martingale regime appears in the scalar non-boundary branching random walk conditioned on large martingale limit. There the additive martingale $g\cdot x = gx/\|gx\|$9 converges almost surely and in mean to a non-degenerate limit $\sigma(g,x)=\log\|gx\|$0, while the derivative martingale $\sigma(g,x)=\log\|gx\|$1 converges in $\sigma(g,x)=\log\|gx\|$2 for every $\sigma(g,x)=\log\|gx\|$3. The right tail satisfies

$\sigma(g,x)=\log\|gx\|$4

and conditioned on $\sigma(g,x)=\log\|gx\|$5, the branching random walk viewed from the minimum converges in law in the vague sense (Chen et al., 2024). This is not a matrix model, but it provides a template for what a non-boundary matrix theory would plausibly seek.

4. Spectral regimes, critical parameters, and asymptotics

In the operator model with one branching center and absorption at every site, the spectral picture is explicit. The perturbed operator $\sigma(g,x)=\log\|gx\|$6 has a unique isolated positive eigenvalue $\sigma(g,x)=\log\|gx\|$7 when $\sigma(g,x)=\log\|gx\|$8, where

$\sigma(g,x)=\log\|gx\|$9

for transient random walks and $u$0 in the recurrent case. The eigenvalue is characterized by

$u$1

and the top eigenvalue of the mean evolution operator $u$2 is $u$3 (Filichkina et al., 2023).

The sign of $u$4 classifies the process. If $u$5, then for each $u$6,

$u$7

If $u$8, then

$u$9

If $G_u$0, then all moments decay exponentially at rate $G_u$1 (Filichkina et al., 2023). This is the infinite-dimensional matrix analogue of the spectral-radius criterion in multitype branching.

For discrete- and continuous-time branching random walks with generating function $G_u$2 and first-moment matrix $G_u$3, local survival is determined by $G_u$4, while global survival for $G_u$5-BRWs is equivalent to $G_u$6. In continuous time, the critical parameters satisfy

$G_u$7

for $G_u$8-BRWs, and at $G_u$9 there is global extinction (Bertacchi et al., 2015).

A matrix-recursive asymptotic regime also occurs in the random-walk-on-a-strip representation. There, transience to the right is equivalent to negativity of the Lyapunov exponent

$\mathbb{M}$00

with $\mathbb{M}$01. The condition $\mathbb{M}$02 ensures that the branching structure embedded in leftward layers is subcritical enough for the walk to escape to $\mathbb{M}$03 (Hong et al., 2012).

5. Multitype branching structure inside random walks and random environments

The phrase matrix branching random walk is also used informally for systems where a non-branching random walk can be encoded by an intrinsic multitype branching structure. In the bounded-jump random walk on $\mathbb{M}$04, excursion counts before the first ladder time $\mathbb{M}$05 are collected into a 9-dimensional population vector

$\mathbb{M}$06

and $\mathbb{M}$07 is a 9-type non-homogeneous branching process with one immigrant at level 1. Its mean evolution is governed by level-dependent $\mathbb{M}$08 matrices $\mathbb{M}$09, and the first ladder time is represented by

$\mathbb{M}$10

Hence $\mathbb{M}$11 is a linear functional of a vector-valued branching process (Hong et al., 2010).

The same paper introduces $\mathbb{M}$12 recursion matrices

$\mathbb{M}$13

which encode exit probabilities from intervals. The mean offspring parameters in the 9-type process are rational functions of these products $\mathbb{M}$14 (Hong et al., 2010). In a homogeneous environment, spectral decomposition of $\mathbb{M}$15 yields explicit formulas such as

$\mathbb{M}$16

where $\mathbb{M}$17 is one eigenvalue of $\mathbb{M}$18.

For random walk on a strip in a random environment, an analogous intrinsic branching structure exists at the level of layers. If $\mathbb{M}$19 counts downward steps from layer $\mathbb{M}$20 to $\mathbb{M}$21, then conditional expectations satisfy

$\mathbb{M}$22

and the first hitting time of layer 1 has the exact decomposition

$\mathbb{M}$23

The expected hitting time is

$\mathbb{M}$24

with $\mathbb{M}$25 (Hong et al., 2012).

These models are not matrix-valued branching random walks in the same sense as the semigroup-of-matrices construction. Rather, they are matrix branching representations: the original walk is scalar, but the hidden branching mechanism is multitype and is propagated by products of random mean matrices.

6. Random-matrix analogues, fixed-point geometry, and broader interpretations

A distinct interpretation appears in the study of moments of moments of characteristic polynomials of Haar unitary matrices. On a binary tree of depth $\mathbb{M}$26, let

$\mathbb{M}$27

and define

$\mathbb{M}$28

Then the moments of moments are

$\mathbb{M}$29

The paper derives the recursion

$\mathbb{M}$30

and proves the asymptotic trichotomy

$\mathbb{M}$31

Under the identification $\mathbb{M}$32, the powers of $\mathbb{M}$33 match those conjectured or proved for moments of moments of CUE characteristic polynomials (Bailey et al., 2020).

That work states explicitly that the model is not matrix-valued. The characteristic polynomial is the matrix object, while the branching random walk is a scalar proxy for the logarithm of that object. In this sense, “matrix branching random walk” means a branching model faithfully reproducing the moments-of-moments exponent structure of a genuine random matrix model (Bailey et al., 2020).

The generating-function perspective adds another layer to the terminology. For a general branching random walk, the multidimensional generating function $\mathbb{M}$34 can have uncountably many fixed points, and a fixed point need not be an extinction probability, even in the irreducible case (Bertacchi et al., 2015). This shows that matrix or operator control of first moments does not determine the full nonlinear fixed-point geometry. A plausible implication is that any general theory of matrix branching random walk must distinguish carefully between linear spectral data, nonlinear extinction structures, and martingale or spine limits, because the first of these does not automatically encode the other two.