Tree-Indexed Markov Chains Overview

• Tree-indexed Markov chains are stochastic processes indexed by rooted trees, enabling Markovian dependence to flow along branch structures rather than in linear time.
• They use formalisms like the Ulam–Harris–Neveu tree and bifurcating kernels to analyze ergodicity, empirical averages, and fluctuations in complex branching models.
• This framework bridges graphical models and MCMC, introducing new threshold phenomena and phase transitions essential for advanced statistical inference on trees.

Searching arXiv for the cited tree-indexed Markov chain literature. Tree-indexed Markov chains are stochastic processes whose indices form a rooted tree rather than a line. In the Ulam–Harris–Neveu formalism, a branching Markov process $X=(X_u)_{u\in\mathbb T_\infty}$ on a Polish state space $(\mathcal X,\mathcal B)$ is specified by an initial law $v$ and a transition kernel $Q$ such that for every finite subtree $T\subset\mathbb T_\infty$ containing the root,

$$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$

(Weibel, 2024). On full binary trees, bifurcating Markov chains replace the single-child kernel by a mother-to-two-daughters kernel $P(x,dy,dz)$, while in finite-alphabet graphical models a “Markov chain on a tree” is characterized by graph-separation conditional independence (Penda et al., 2020, Bhattacharya et al., 2023). The subject therefore encompasses several closely related formalisms in which Markovian dependence is propagated along ancestral structure rather than along a one-dimensional time axis.

1. Indexing trees, notation, and canonical constructions

A standard index set is the Ulam–Harris–Neveu tree

$$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$

with root $\varnothing$, parent map $p(u)$, height $(\mathcal X,\mathcal B)$0, latest common ancestor $(\mathcal X,\mathcal B)$1, and graph distance

$(\mathcal X,\mathcal B)$2

(Weibel, 2024). For binary models one often writes

$(\mathcal X,\mathcal B)$3

so that $(\mathcal X,\mathcal B)$4 is the $(\mathcal X,\mathcal B)$5-th generation and $(\mathcal X,\mathcal B)$6 is the tree up to generation $(\mathcal X,\mathcal B)$7 (Penda et al., 2020, Penda et al., 2021). For rooted $(\mathcal X,\mathcal B)$8-ary trees,

$(\mathcal X,\mathcal B)$9

and $v$0 for the first $v$1 levels (Ban et al., 2024).

The line graph remains a limiting benchmark. In Weibel’s formulation, the finite line graph

$v$2

with edges $v$3, is the special case in which the tree-indexed process reduces to an ordinary length-$v$4 Markov chain (Weibel, 2024). This comparison is structurally important because many asymptotic and variance formulas can be read as tree analogues of classical chain results.

In bifurcating models, the transition mechanism is not a single-child kernel but a joint law for siblings. If $v$5 is the mother-to-two-daughter kernel, its one-dimensional marginals are

$v$6

and the averaged kernel

$v$7

defines the auxiliary lineage chain obtained by following a typical random branch (Penda et al., 2020, Penda et al., 2021). This auxiliary chain is central in many-to-one identities, ergodic theory, and fluctuation results.

2. Markov properties, factorization, and neighboring formalisms

For tree-indexed processes in the rooted sense, the branching-Markov property states that descendants evolve conditionally independently across different parents once the current generation is fixed. In the binary case, conditionally on the $v$8-field generated by the first $v$9 generations, the pairs $Q$0 are independent and each has conditional distribution $Q$1 (Penda et al., 2020). The associated many-to-one formulas identify expectations and second moments of generation sums in terms of iterates of the lineage kernel $Q$2 (Penda et al., 2020, Penda et al., 2021).

A different but related formalism assigns variables to the vertices of a finite tree $Q$3 and defines a Markov chain on the tree by edgewise conditional independence. For every edge $Q$4,

$Q$5

and this is equivalent to a global Markov property based on graph separation: $Q$6 (Bhattacharya et al., 2023). In the strictly positive directed case this yields the familiar factorization

$Q$7

The literature also contains a block version. An $Q$8-block Markov chain on a rooted tree satisfies

$Q$9

so that the joint law of the immediate children depends only on the parent state and is independent of the exterior of the descendant subtree (Souissi, 2020). Measures that are block Markov chains for every choice of root are classical Markov chains on the tree, and they form a strict subclass of Markov random fields; in the one-dimensional case the class of block Markov chains coincides with the class of Markov chains (Souissi, 2020).

The terminology is not uniform across the literature. A separate line studies Markov chains on rooted trees where the tree is the state space rather than the index set. For almost upper-directed kernels on infinite locally finite trees, every irreducible kernel has some invariant measures, and an $T\subset\mathbb T_\infty$0-invariant measure is given by a determinantal formula; recurrence and positive recurrence are then analyzed through explicit criteria and a leaf addition algorithm (Fredes et al., 2024). This is a distinct problem from labeling a tree by a Markovian field.

3. Ergodicity, empirical averages, and law-of-large-numbers phenomena

For arbitrary finite subsets $T\subset\mathbb T_\infty$1, empirical averages are defined by

$T\subset\mathbb T_\infty$2

Under a geometrical assumption requiring that two uniformly sampled vertices in $T\subset\mathbb T_\infty$3 are typically far apart, together with either tightness of the height of their latest common ancestor or a strong-ergodicity condition on the kernel $T\subset\mathbb T_\infty$4, one has

$T\subset\mathbb T_\infty$5

(Weibel, 2024). In the ergodic case with unique invariant law $T\subset\mathbb T_\infty$6 and bounded continuous $T\subset\mathbb T_\infty$7, the limit is $T\subset\mathbb T_\infty$8 (Weibel, 2024).

Several tree families satisfy the geometric hypotheses naturally. Any infinite rooted tree of degree $T\subset\mathbb T_\infty$9 satisfies the required distance condition for every $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$0 with $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$1; spherically symmetric trees with generation sets $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$2 satisfy both the distance and ancestor conditions; and for a supercritical Galton–Watson tree conditioned on non-extinction, both $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$3 and $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$4 satisfy the hypotheses almost surely (Weibel, 2024). These examples show that the ergodic theorem is not restricted to regular binary trees.

A complementary law of large numbers arises from observing the tree-indexed chain along an independent random walk on the complete binary tree. If the random-walk observation process $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$5 converges to a càdlàg Feller process $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$6, then the empirical measure process over generation $$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$7,

$$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$8

converges to

$$\Pr\bigl(X_u\in dx_u,\ u\in T\bigr) = v(dx_{\varnothing})\prod_{u\in T\setminus\{\varnothing\}}Q\bigl(x_{p(u)};dx_u\bigr)$$9

in $P(x,dy,dz)$0 (Czuppon et al., 2014). The mechanism is asymptotic independence of two uniformly chosen leaves in a deep generation, since their most recent common ancestor lies at depth $P(x,dy,dz)$1 (Czuppon et al., 2014).

The same framework yields an optimization statement relevant for Markov-chain Monte Carlo. When the underlying chain is stationary and reversible, and $P(x,dy,dz)$2 is compact self-adjoint on $P(x,dy,dz)$3, the variance of $P(x,dy,dz)$4 for an eigenfunction $P(x,dy,dz)$5 is proportional to the Hosoya–Wiener polynomial

$P(x,dy,dz)$6

Among all subtrees of fixed size $P(x,dy,dz)$7, the line graph $P(x,dy,dz)$8 uniquely minimizes $P(x,dy,dz)$9 for each nonzero eigenvalue $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$0, so the ordinary chain yields the smallest non-asymptotic variance for the empirical average in that regime (Weibel, 2024).

4. Fluctuations, concentration, and large deviations

For bifurcating Markov chains, additive functionals over generations or whole trees exhibit a three-regime phase structure governed by the comparison between the branching factor $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$1 and the ergodicity rate $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$2 of the lineage kernel. If

$$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$3

then under geometric ergodicity one has the following trichotomy: in the subcritical case $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$4, $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$5 converges to a centered Gaussian limit; in the critical case $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$6, $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$7 converges to a centered Gaussian limit with a different variance; and in the supercritical case $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$8, $$\mathbb T_\infty=\bigcup_{k=0}^\infty (\mathbb N^*)^k, \qquad (\mathbb N^*)^0=\{\varnothing\},$$9 is asymptotically governed by martingale limits rather than a Gaussian law (Penda et al., 2020, Penda et al., 2021). The proofs rely on martingale decompositions, predictable quadratic variation, and spectral projection arguments.

Moderate deviations refine the central-limit regime for one-variable additive functionals of bifurcating Markov chains. Under uniform geometric ergodicity of the lineage kernel $\varnothing$0, if $\varnothing$1 with $\varnothing$2 in the subcritical case, then $\varnothing$3 satisfies a moderate deviation principle on $\varnothing$4 with speed $\varnothing$5 and quadratic rate function $\varnothing$6; in the critical case $\varnothing$7, the normalization becomes $\varnothing$8 under the second spectral gap assumption (Penda et al., 2021).

Concentration inequalities can be obtained without full spectral analysis. Under transportation cost-information assumptions $\varnothing$9 on the initial law and local branching kernel, the law of the bifurcating Markov chain on $p(u)$0 belongs to $p(u)$1, which yields Gaussian-type tail bounds for Lipschitz empirical means (Penda et al., 2015). In particular, under $p(u)$2, the generation average

$p(u)$3

satisfies a sub-Gaussian Laplace bound with variance proxy $p(u)$4, and $p(u)$5 when $p(u)$6, where $p(u)$7 is the sum of the Wasserstein contraction constants of the two marginals (Penda et al., 2015). Analogous subtree-level bounds hold for $p(u)$8, with $p(u)$9 when $(\mathcal X,\mathcal B)$00 (Penda et al., 2015).

These results show that branching geometry does not merely change constants. It produces threshold phenomena, new critical renormalizations, and non-Gaussian supercritical limits that have no analogue for ordinary one-dimensional chains with the same kernel.

5. Information-theoretic, statistical, and geometric extensions

In the information-theoretic formulation of a Markov chain on a finite tree, the shared information of the entire collection admits an explicit edgewise characterization: $(\mathcal X,\mathcal B)$01 The minimizer is the weakest edge of the tree, and this converts estimation of shared information into a best-arm identification problem over edges (Bhattacharya et al., 2023). When the joint law is unknown but the tree is known, the paper develops a uniform-sampling multiarmed bandit algorithm based on empirical mutual information and establishes error and sample-complexity bounds (Bhattacharya et al., 2023).

Statistical inference for hidden models has also been developed in the tree-indexed setting. For hidden Markov models indexed by the complete binary tree, with hidden branching Markov chain $(\mathcal X,\mathcal B)$02 and conditionally independent observations $(\mathcal X,\mathcal B)$03, the maximum-likelihood estimator based on $(\mathcal X,\mathcal B)$04 is strongly consistent and asymptotically normal under the stated dominance, Doeblin, regularity, and smoothness assumptions (Weibel, 2024). The proofs use ergodic theorems for Markov chains indexed by trees with neighborhood-dependent functions, exponential forgetting of conditional initial distributions, a score martingale-array central limit theorem, and a law of large numbers for the observed information matrix (Weibel, 2024).

Finite-state tree-indexed chains on rooted $(\mathcal X,\mathcal B)$05-trees admit a large-deviation and dimension-theoretic theory via the method of types. For a transition matrix $(\mathcal X,\mathcal B)$06 and a positive weight matrix $(\mathcal X,\mathcal B)$07, the empirical average

$(\mathcal X,\mathcal B)$08

satisfies a Cramér-type theorem along periodic subsequences, the empirical averages converge almost surely to the unique zero of the rate function in the irreducible case, and the Hausdorff dimension of the associated Markov hom tree-shift is given by a nonlinear Perron–Frobenius variational formula (Ban et al., 2024).

A continuum extension replaces the discrete rooted tree by a Lévy tree. In that setting one constructs a Markov process indexed by the Lévy tree through the snake property, defines local time at a regular and instantaneous point, proves a Poisson decomposition of excursions away from that point, and shows that the genealogy of excursions is itself encoded by a Lévy tree called the tree coded by the local time (Riera et al., 2024). This recovers, in particular, the excursion theory of Abraham and Le Gall for Brownian motion indexed by the Brownian tree (Riera et al., 2024).

6. Poisson representability and phase transitions on finite and infinite trees

A recent development studies $(\mathcal X,\mathcal B)$09-valued tree-indexed Markov chains through Poisson representations. For a rooted tree $(\mathcal X,\mathcal B)$10, parameters $(\mathcal X,\mathcal B)$11, and a parent–child edge $(\mathcal X,\mathcal B)$12, the transition matrix is

$(\mathcal X,\mathcal B)$13

equivalently

$(\mathcal X,\mathcal B)$14

with $(\mathcal X,\mathcal B)$15; along each parent-child link the state flips to an independent fresh $(\mathcal X,\mathcal B)$16 with probability $(\mathcal X,\mathcal B)$17 and otherwise retains its parent’s value (Forsström, 24 Jan 2025).

Poisson-representability is formulated on a countable index set $(\mathcal X,\mathcal B)$18. A $(\mathcal X,\mathcal B)$19-valued process $(\mathcal X,\mathcal B)$20 belongs to the class $(\mathcal X,\mathcal B)$21 if there exists a measure $(\mathcal X,\mathcal B)$22 on nonempty subsets of $(\mathcal X,\mathcal B)$23 such that, for a Poisson point process $(\mathcal X,\mathcal B)$24 on $(\mathcal X,\mathcal B)$25 with intensity $(\mathcal X,\mathcal B)$26,

$(\mathcal X,\mathcal B)$27

Equivalently, the unique signed intensity measure satisfies

$(\mathcal X,\mathcal B)$28

and $(\mathcal X,\mathcal B)$29 precisely when this $(\mathcal X,\mathcal B)$30 is nonnegative on all finite subsets (Forsström, 24 Jan 2025).

For finite trees that are not simple paths, representability is controlled by two thresholds depending only on the leaf boundary size $(\mathcal X,\mathcal B)$31 and maximal degree $(\mathcal X,\mathcal B)$32. The paper defines $(\mathcal X,\mathcal B)$33 from complementary Bell numbers and $(\mathcal X,\mathcal B)$34 from the largest negative root of $(\mathcal X,\mathcal B)$35, and proves:

• if $(\mathcal X,\mathcal B)$36 and $(\mathcal X,\mathcal B)$37, then $(\mathcal X,\mathcal B)$38;
• if $(\mathcal X,\mathcal B)$39 and $(\mathcal X,\mathcal B)$40, then $(\mathcal X,\mathcal B)$41;
• if $(\mathcal X,\mathcal B)$42 and $(\mathcal X,\mathcal B)$43, then $(\mathcal X,\mathcal B)$44;
• if $(\mathcal X,\mathcal B)$45 and $(\mathcal X,\mathcal B)$46, then $(\mathcal X,\mathcal B)$47.

Thus every non-path finite tree exhibits a genuine phase transition in $(\mathcal X,\mathcal B)$48 for Poisson-representability (Forsström, 24 Jan 2025).

The infinite case is not degenerate. For the “octopus” tree $(\mathcal X,\mathcal B)$49, consisting of one central vertex of degree $(\mathcal X,\mathcal B)$50 with $(\mathcal X,\mathcal B)$51 infinite rays attached, there is again a phase transition solely in $(\mathcal X,\mathcal B)$52: if $(\mathcal X,\mathcal B)$53 then $(\mathcal X,\mathcal B)$54 for all $(\mathcal X,\mathcal B)$55; there exists $(\mathcal X,\mathcal B)$56, independent of $(\mathcal X,\mathcal B)$57, such that if $(\mathcal X,\mathcal B)$58 then $(\mathcal X,\mathcal B)$59 for all $(\mathcal X,\mathcal B)$60; and for $(\mathcal X,\mathcal B)$61 the critical value is exactly $(\mathcal X,\mathcal B)$62, independent of $(\mathcal X,\mathcal B)$63 (Forsström, 24 Jan 2025).

The proofs are combinatorial and rely on Möbius inversion for the signed intensity measure, an explicit inclusion–exclusion formula for $(\mathcal X,\mathcal B)$64, Taylor expansion around $(\mathcal X,\mathcal B)$65 and $(\mathcal X,\mathcal B)$66, and restriction lemmas extended to signed measures (Forsström, 24 Jan 2025). The same paper also sharpens earlier results by showing that any signed measure $(\mathcal X,\mathcal B)$67 arising from a Markov field has support only on connected subsets, and by extending line-chain formulas to arbitrary tree-indexed chains (Forsström, 24 Jan 2025).

Tree-indexed Markov chains are therefore not a single theorem but a broad research area in which branching geometry, ancestral overlap, and local kernel structure govern ergodic averages, fluctuations, concentration, inference, information measures, and representability. The common theme is that the tree replaces linear time by a partially ordered genealogy, and this change introduces new thresholds, new factorization phenomena, and new links to graphical models, large deviations, and random-tree geometry.