Aldous–Pitman Fragmentation Process

• The Aldous–Pitman fragmentation process is a stochastic model that describes mass partitions derived from recursive random cuts on continuum random trees and their discrete approximants.
• It uses Poissonian cuts to induce self-similar fragmentations with Markov properties, linking to Lévy processes and Poisson–Dirichlet distributions.
• This process has significant implications in probabilistic combinatorics, scaling limits, and the study of coalescent structures in random trees.

The Aldous–Pitman fragmentation process is a fundamental stochastic model describing the evolution of mass partitions derived from recursive random cutting of continuum random trees (CRTs) or their discrete approximants. Originally formulated by Aldous and Pitman, this process encodes the law governing how a tree, or its scaling limit, breaks apart as Poissonian cuts are applied along its skeleton, and captures universal phenomena via limiting invariance principles, explicit connections with Lévy processes, Poisson–Dirichlet distributions, and minimal combinatorial structures. Its generalization to stable Lévy trees and associated partition structures forms a central object in the modern theory of self-similar fragmentation.

1. Construction and Limit Processes

The Aldous–Pitman fragmentation emerges in the scaling limits of recursive edge removal on discrete trees. On a uniform random tree (Cayley tree) $T_n$ with $n$ labelled vertices, assign i.i.d. $\mathrm{U}[0,1]$ weights to edges. For threshold $u\in[0,1]$, removing edges with weight exceeding $u$ induces a random forest with fragment sizes $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$, ordered and padded with zeros. As $u$ increases, this defines a fragmentation process interpolating from the full tree to singletons. After appropriate time and space rescaling, $F_n(u)$ converges as $n\to\infty$ to a limiting process $F^{(2)}(t)$, which is the fragmentation of the Brownian continuum random tree (CRT) cut along its skeleton by a homogeneous Poisson process in space-time intensity $n$0 (Ojeda et al., 2020, Kortchemski et al., 2023).

For general $n$1, replacing the uniform tree by a critical Galton–Watson tree with offspring law in the domain of attraction of a stable law of index $n$2, the edge-removal process, after normalization, converges to the fragmentation of the $n$3-stable Lévy tree. The limiting process is thus parametrized by $n$4 and can be realized via Poissonian marks along the skeleton of a compact $n$5-tree with length measure $n$6 (Ojeda et al., 2020).

2. Self-Similarity, Dislocation Measure, and Markov Property

The Aldous–Pitman fragmentation is a Markov process $n$7 in the simplex of ranked, summable sequences. For the Brownian CRT, each fragment of mass $n$8 splits at total rate $n$9 (self-similarity index $\mathrm{U}[0,1]$0), and the law of the fragment sizes upon splitting (the dislocation measure) is supported on two-block partitions with Beta$\mathrm{U}[0,1]$1 distribution. The generator on test functions $\mathrm{U}[0,1]$2 on decreasing mass-partitions $\mathrm{U}[0,1]$3 is

$\mathrm{U}[0,1]$4

In the general $\mathrm{U}[0,1]$5-stable case, the index is $\mathrm{U}[0,1]$6, and the dislocation measure is the law of the jumps of a pure-jump subordinator with Lévy measure $\mathrm{U}[0,1]$7; for $\mathrm{U}[0,1]$8, this recovers the Brownian case (Ojeda et al., 2020, Thévenin, 2019, Ho et al., 2018).

3. Excursion, Lamination, and Lévy Process Representations

The evolution of the ranked component masses has an encoding by Brownian (or more generally, stable) excursions with deterministic drift. For the Brownian case, if $\mathrm{U}[0,1]$9 is a standard excursion, then at time $u\in[0,1]$0 the process $u\in[0,1]$1, and the open intervals of constancy of $u\in[0,1]$2 correspond to the fragment lengths $u\in[0,1]$3.

This admits a geometric representation via laminations: to any excursion function $u\in[0,1]$4, construct a lamination of the closed unit disk $u\in[0,1]$5 by relating points $u\in[0,1]$6 when $u\in[0,1]$7, and considering the closure of chords between $u\in[0,1]$8 and $u\in[0,1]$9. Face masses in the limiting lamination coincide with the normalized fragment sizes (Thévenin, 2019). The process can be explicitly constructed for the $u$0-stable case using a spectrally positive $u$1-stable Lévy process with Laplace exponent $u$2.

4. Cut-Tree and Genealogical Encoding

The Aldous–Pitman fragmentation process admits a canonical genealogical encoding via the "cut-tree" construction. For the Brownian CRT, sample i.i.d. leaves $u$3 and define the time $u$4 when the component containing $u$5 and $u$6 is separated by a cut. A pseudo-metric is defined by

$u$7

forming a real tree, the cut-tree, itself a Brownian CRT in distribution. The original CRT and its cut-tree are related via a measurable mapping in the Gromov–Hausdorff–Prokhorov topology (Broutin et al., 2014, Kortchemski et al., 2023). Genealogical information for the fragments is thus encoded in this structure, with each subtree corresponding bijectively to a fragment.

5. Invariance Principles and Universality

Discrete Galton–Watson trees conditioned to have $u$8 vertices, when equipped with i.i.d. edge-weights and subjected to the edge-deletion process, converge (under proper rescaling determined by the stable law index) to the continuum stable Lévy tree fragmentation. The main invariance principle states

$u$9

with $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$0 chosen via the scaling limits of the conditioned random walks encoding the trees (Ojeda et al., 2020). This result recovers Aldous–Pitman as the special case $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$1 and establishes universality of the process across broad classes of discrete random tree models.

6. Connections with Partition Structures, Poisson–Dirichlet, and Mittag-Leffler Laws

The law of the ranked fragment sizes at a (possibly exponential) random time is given by a Poisson–Dirichlet $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$2 mixture, itself related to generalized Mittag-Leffler functions and mixed-Poisson waiting time frameworks. For instance, starting from total mass $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$3 and fragmenting for exponential time $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$4, the mass-partition’s law is expressed via

$F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$5

where $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$6 involves the stable law density and the Prabhakar–Mittag-Leffler function. For $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$7, this specializes to classic Aldous–Pitman formulas involving Hermite functions and the Beta$F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$8 splitting law (Ho et al., 2018).

7. Duality, Additive Coalescent, and Reversibility

Reversing the process in time yields an additive coalescent: as time decreases, clusters (masses) merge pairwise at rates proportional to their sums. Dual constructions exist—such as the shuffle transform, which reconstructs the original CRT from its cut-tree, providing a law-level inverse to the cutting operation. This reversibility underlies connections between fragmentation and coalescent processes, and between the Aldous–Pitman process and combinatorial operations like minimal factorizations of the $F_n(u)=(F_{n,1}(u),F_{n,2}(u),\ldots)$9-cycle (in the Brownian case), making the process a central bridge between stochastic geometry, combinatorics, and probability (Broutin et al., 2014, Thévenin, 2019, Kortchemski et al., 2023).