Semirandom Tree Process Overview

• Semirandom tree processes are families of stochastic tree constructions that merge randomness with constrained, monotone modifications to drive growth and inference.
• They include models such as uniform spanning tree selections, adversarial broadcast trees, continuous-time recursive partition trees, and trickle-down processes on fixed graphs.
• These models have practical applications in online graph construction, Bayesian inference, and algorithmic reconstruction, with analysis focusing on thresholds and asymptotic behaviors.

Semirandom tree process denotes a family of stochastic constructions in which tree growth, tree-indexed inference, or tree-based partitioning combines randomness with constrained local choice or monotone modification. In current usage, the term does not refer to a single canonical model. It includes, at minimum, the uniform spanning tree semi-random model in which each round offers a uniformly random spanning tree and Builder keeps one edge (Burova et al., 2022), semirandom broadcast models on Galton–Watson or regular trees followed by a monotone adversary (Moitra et al., 2015), continuous-time random recursive partition trees such as the Binary Space Partitioning-Tree process (Fan et al., 2019), and trickle-down growth processes on a fixed directed acyclic graph driven by local routing chains (Evans et al., 2010). This suggests that “semirandom tree process” is best treated as a technical umbrella for several closely related paradigms rather than as a single standardized definition.

1. Principal meanings of the term

A first meaning is online graph building from random spanning trees. In the uniform spanning tree semi-random model, the random object presented each round is a spanning tree on $[n]$, and the controlled action is the choice of one edge from that tree (Burova et al., 2022).

A second meaning is broadcast on a random tree with monotone adversarial post-processing. Here the underlying object is a broadcast process on a Galton–Watson tree or a regular tree, and the semirandomness enters through an offline adversary that may delete or replace certain subtrees but only through “helpful” monotone operations (Moitra et al., 2015).

A third meaning is random recursive partition trees. The BSP-Tree process is an almost surely right-continuous Markov jump process on polygonal partitions; it incrementally splits one current block at a time and induces a binary tree over blocks (Fan et al., 2019).

A fourth meaning is growth on a deterministic combinatorial skeleton. Trickle-down processes represent Markov chains as evolving connected subsets of a directed acyclic graph: particles are fed at a source, routed locally according to independent rules, and stop when they reach an unoccupied vertex (Evans et al., 2010).

Model family	Core random object	Representative source
Uniform spanning tree semi-random model	One uniformly random spanning tree per round	(Burova et al., 2022, Anastos et al., 30 Sep 2025)
Semirandom broadcast tree	Broadcast labels plus monotone adversary	(Moitra et al., 2015)
BSP-Tree process	Continuous-time recursive space partition tree	(Fan et al., 2019)
Trickle-down process	Randomly grown connected subset of a DAG	(Evans et al., 2010)

This multiplicity is not merely terminological. The models differ in state space, time parameter, control mechanism, and asymptotic questions: hitting times in online construction, reconstruction thresholds in inference, projectivity and self-consistency in Bayesian partition priors, and Doob–Martin boundary descriptions in Markov growth.

2. Uniform spanning-tree semi-random process

In the uniform spanning tree semi-random model, also called the semi-random tree process, the vertex set is $[n]=\{1,\dots,n\}$, the process starts from the empty graph $G_0=([n],\emptyset)$, and at round $i$ a spanning tree $T_i$ of $K_n$ is drawn uniformly at random from $\mathcal T_n$, independently of previous rounds. Builder sees the whole edge set $E(T_i)$ and chooses one edge of $T_i$ to add; choosing no edge is allowed for convenience (Burova et al., 2022). A deterministic strategy is a map

$\sigma:2^{E_n}\times \mathcal T_n\to E_n$

with $[n]=\{1,\dots,n\}$0, and the hitting time of a property $[n]=\{1,\dots,n\}$1 is

$[n]=\{1,\dots,n\}$2

The model was analyzed for classical graph properties with asymptotically sharp main terms. For minimum degree at least $[n]=\{1,\dots,n\}$3, if $[n]=\{1,\dots,n\}$4, there exists a strategy such that whp

$[n]=\{1,\dots,n\}$5

while every strategy needs at least the same main term whp (Burova et al., 2022). For Hamiltonicity, there exists a strategy such that whp

$[n]=\{1,\dots,n\}$6

for any $[n]=\{1,\dots,n\}$7. For a fixed tree $[n]=\{1,\dots,n\}$8, an $[n]=\{1,\dots,n\}$9-factor can be built in

$G_0=([n],\emptyset)$0

rounds whp, and for a fixed central graph $G_0=([n],\emptyset)$1 there is a constant $G_0=([n],\emptyset)$2 such that $G_0=([n],\emptyset)$3 whp (Burova et al., 2022).

A later result establishes the bounded-degree spanning-graph threshold. If $G_0=([n],\emptyset)$4 is an $G_0=([n],\emptyset)$5-vertex graph with maximum degree $G_0=([n],\emptyset)$6, then in the semirandom tree process Builder has a strategy guaranteeing that with high probability after

$G_0=([n],\emptyset)$7

rounds the constructed graph contains a copy of $G_0=([n],\emptyset)$8; this is asymptotically optimal in $G_0=([n],\emptyset)$9 because Builder can take at most one edge per round and $i$0 (Anastos et al., 30 Sep 2025).

The analysis uses spanning-tree-specific probability estimates derived from the Aldous–Broder construction and Moon’s rooted-forest count. Two key lemmas are

$i$1

for any edge set $i$2, and

$i$3

which support the greedy, bridging, and final repair phases of the construction (Anastos et al., 30 Sep 2025). A common misconception is that the richness of a whole random spanning tree per round trivializes the process. The optimal $i$4 bound shows instead that the model remains tightly constrained by the one-edge-per-round bottleneck.

3. Broadcast trees with monotone adversaries

In the semirandom broadcast framework, the base model is the binary broadcast process on a tree. The principal case is a Galton–Watson tree with Poisson offspring,

$i$5

root spin $i$6 uniform, and per-edge flip probability $i$7, so that

$i$8

The edge correlation is $i$9, and in the pure random model the Kesten–Stigum threshold is

$T_i$0

Partial recovery means that some estimator of the root has success probability bounded away from $T_i$1 as the depth $T_i$2 (Moitra et al., 2015).

The semirandom version adds a monotone adversary after the broadcast instance is drawn. In the cutting semirandom model, the adversary sees the full labeled tree and may delete any number of edges between nodes with different labels; when an edge is cut, the corresponding subtree is removed and not revealed. In the strong semirandom model, for each edge where the child spin differs from the parent spin, the adversary may replace the entire rooted subtree at that child by any tree it likes (Moitra et al., 2015).

The central threshold result is that semirandom modifications can make reconstruction strictly harder. For every $T_i$3 there exists $T_i$4 such that

$T_i$5

so the random broadcast tree is reconstructible, yet under the cutting semirandom model partial recovery is information-theoretically impossible: $T_i$6 The mechanism is a structured adversary that cuts certain degree-2 correction patterns, reducing the effective branching factor from $T_i$7 to $T_i$8 and pushing the relevant condition below threshold (Moitra et al., 2015).

The same work shows a robustness contrast between algorithms. Majority on leaves is optimal at the Kesten–Stigum threshold in the pure model but is not robust to semirandom perturbations. Recursive majority, by contrast, remains valid under a strong monotone adversary on $T_i$9-ary trees. If

$K_n$0

then recursive majority succeeds under the strong semirandom adversary; asymptotically,

$K_n$1

A common misconception is that “helpful” adversarial changes must improve inference. Here, monotone changes that strengthen assortativity can nonetheless destroy the local probabilistic structure exploited by optimal algorithms, and can even shift the information-theoretic threshold inward (Moitra et al., 2015).

4. Geometric partition trees and the BSP-Tree process

The Binary Space Partitioning-Tree process is a continuous-time stochastic process on partitions of a bounded convex polygon $K_n$2. A partition is a finite collection of convex polygonal blocks whose union is $K_n$3, and the process

$K_n$4

is an almost surely right-continuous Markov jump process on the space of such partitions. It is piecewise constant: at each jump time, a single current block is split into two blocks, inducing a binary tree over blocks (Fan et al., 2019).

The jump mechanism is geometric. If the current partition has blocks $K_n$5, then the waiting time to the next cut is exponential with rate equal to the sum of current block perimeters,

$K_n$6

so the probability that the next cut occurs in block $K_n$7 is proportional to $K_n$8 (Fan et al., 2019). In the uniform-direction version, after choosing a block, one samples a direction $K_n$9 with density proportional to the projected length $\mathcal T_n$0, then samples a cut location uniformly along the projected segment. Proposition 2 gives

$\mathcal T_n$1

and the resulting joint measure on lines is uniform over all cut lines intersecting the block.

The process extends beyond isotropic cuts by introducing a direction weight $\mathcal T_n$2, yielding

$\mathcal T_n$3

The choice $\mathcal T_n$4 gives isotropic cuts; the choice $\mathcal T_n$5 recovers axis-aligned Mondrian cuts (Fan et al., 2019). The process is self-consistent: restricting a BSP-Tree partition generated on a larger convex polygon to a convex sub-polygon yields the same distribution as generating the process directly on the sub-polygon. This projectivity enables construction on $\mathcal T_n$6 via the Kolmogorov Extension Theorem.

Inference is performed by Conditional-Sequential Monte Carlo, treating the tree as the latent trajectory. The likelihood depends on the application: blockwise Dirichlet–Multinomial models for toy labeling and blockwise Beta–Bernoulli models for relational data. Empirically, the paper reports clear inferential improvements over the standard Mondrian process and related methods; in relational modeling on five social networks, BSP-RM achieves the highest AUC, improving by about $\mathcal T_n$7 over the best baselines (Fan et al., 2019).

5. Trickle-down growth on a deterministic skeleton

Trickle-down processes provide a general Markovian framework for semirandom tree growth on a fixed directed acyclic graph $\mathcal T_n$8 with unique root $\mathcal T_n$9. Particles are fed in one by one at the root, successive particles proceed along directed edges according to local routing rules, and each particle comes to rest once it encounters an unoccupied vertex. When $E(T_i)$0 is a rooted tree such as $E(T_i)$1 or $E(T_i)$2, the state can be identified with a finite rooted subtree containing the root (Evans et al., 2010).

The formalism assigns to each vertex $E(T_i)$3 a Markov routing chain $E(T_i)$4 on states $E(T_i)$5, where $E(T_i)$6 is the successor set. The local transition always increments exactly one coordinate, and the global configuration is assembled by clock variables $E(T_i)$7 that count how many particles have passed through $E(T_i)$8 after its first occupation. The resulting global chain $E(T_i)$9 is Markov, transient, and strictly increasing in the natural partial order (Evans et al., 2010).

A fundamental structural fact is the product form of hitting probabilities and Martin kernels. If $T_i$0 are global states, then

$T_i$1

and the Martin kernel factorizes accordingly: $T_i$2 This reduction allows global asymptotics to be read from local routing chains (Evans et al., 2010).

The framework encompasses a broad class of tree-growth models. For the binary search tree process, each $T_i$3 is a two-color Pólya urn. Its boundary is identified with probability measures on $T_i$4, and the extended Martin kernel is

$T_i$5

where $T_i$6 is the mass of the cylinder rooted at $T_i$7 (Evans et al., 2010). The same architecture also captures digital search trees, random recursive trees and their nested Pitman two-parameter Chinese restaurant generalizations, Mallows $T_i$8 tree processes, $T_i$9-binomial tree processes, and the Luczak–Winkler Catalan tree process. In these models the boundary object is either a probability measure on path space or an infinite tree with a single spine, and the tail $\sigma:2^{E_n}\times \mathcal T_n\to E_n$0-field is generated by this limit object (Evans et al., 2010).

6. Unifying themes, related directions, and misconceptions

Outside explicitly tree-based papers, semirandom models are often defined as a random instance followed by a constrained adversary. One formulation allows an adversary to decrease $\sigma:2^{E_n}\times \mathcal T_n\to E_n$1 for any $\sigma:2^{E_n}\times \mathcal T_n\to E_n$2, where $\sigma:2^{E_n}\times \mathcal T_n\to E_n$3 is the planted set, and a slightly more permissive variant allows the adversary to decrease $\sigma:2^{E_n}\times \mathcal T_n\to E_n$4 off the planted set and increase $\sigma:2^{E_n}\times \mathcal T_n\to E_n$5 on the planted set (Brennan et al., 2019). This broader formulation helps explain why semirandom tree models so often emphasize monotone cutting, deletion, or rewriting rather than arbitrary corruption.

Tree-indexed recursive processes provide another nearby perspective. Hipster random walks are defined on the complete rooted infinite binary tree by local random functions $\sigma:2^{E_n}\times \mathcal T_n\to E_n$6: if the two child values differ, one child value is chosen uniformly; if they are equal, a random step $\sigma:2^{E_n}\times \mathcal T_n\to E_n$7 is added. Under iid integer input at the leaves, the one-point marginal laws satisfy deterministic nonlinear recurrences. In the totally asymmetric $\sigma:2^{E_n}\times \mathcal T_n\to E_n$8-lazy case,

$\sigma:2^{E_n}\times \mathcal T_n\to E_n$9

and in the symmetric case

$[n]=\{1,\dots,n\}$00

with the proofs proceeding through discrete analogues of Burgers and porous medium equations (Addario-Berry et al., 2019). Although this literature uses a different name, it exhibits the same combination of local random gates, deterministic recursive structure, and tree-indexed asymptotics.

A common misconception is that semirandom tree process denotes one universally accepted object. The literature instead contains at least four technically distinct families: online spanning-tree offer processes, broadcast trees with monotone adversaries, geometric recursive partition trees, and trickle-down growth on deterministic skeletons. Another misconception is that semirandomness invariably simplifies inference or construction. In broadcast trees, monotone “helpful” perturbations can make reconstruction strictly harder than in the pure random model (Moitra et al., 2015); in the online spanning-tree model, even globally informative tree offers do not beat the asymptotic $[n]=\{1,\dots,n\}$01 lower bound for embedding a bounded-degree spanning graph (Anastos et al., 30 Sep 2025).

This suggests a unifying description: a semirandom tree process is a tree-valued or tree-indexed stochastic system with a fixed combinatorial or geometric support, a local random update rule or monotone adversarial modification, and a global evolution analyzed through thresholds, projective consistency, or boundary objects. Under that description, the term links online graph construction, broadcasting and reconstruction, Bayesian partition priors, and Markov tree growth, while preserving the technical differences that drive each subliterature.