Multiscale Stick-Breaking Construction

• Multiscale stick-breaking construction is a stochastic process that generates random probability measures through hierarchical, recursive mass allocations on tree structures.
• It generalizes classical stick-breaking by splitting mass across all branches, yielding uniform cluster sizes and enabling flexible, nonparametric Bayesian density modeling.
• The approach supports advanced posterior inference using techniques like slice sampling and Pólya–Gamma augmentation, enhancing scalability and local adaptivity.

A multiscale stick-breaking construction is a stochastic process for generating random probability measures or random partitions, characterized by hierarchical or recursive allocation of mass across several scales or resolutions. Unlike classical one-sided stick-breaking, the multiscale variant organizes splitting across all branches of a binary or general tree structure, enabling increasingly fine partitions and supporting modeling at multiple levels of granularity. This paradigm has been foundational in nonparametric Bayesian density modeling, mixture modeling, and the study of random combinatorial structures.

1. General Dyadic-Tree and Multiscale Stick-Breaking Formalism

Let $\tau$ be a finite or infinite full binary tree of depth $L$ (possibly $L=\infty$), with nodes indexed by binary strings $\varepsilon \in \{0,1\}^{\leq L}$. Each internal node $\varepsilon$ (i.e., $|\varepsilon| < L$) is assigned a split variable $$V_\varepsilon \sim \mathrm{Beta}(a_\varepsilon, b_\varepsilon)$$. The root is $\varepsilon = \emptyset$. In the standard construction:

• When $\varepsilon$ splits, fraction $V_\varepsilon$ proceeds to its left child $L$0 and $L$1 to right child $L$2.
• The stick-mass at a particular leaf $L$3 is:

$L$4

• These weights satisfy $L$5 where $L$6 is the set of binary strings indexing the leaves (Horiguchi et al., 2022).

This framework generalizes to structures beyond binary trees, such as general branching or multinomial trees, and is the basis for hierarchical allocations in mixture models and random measures.

2. Comparison to Classical Stick-Breaking and Balanced vs. Lopsided Constructions

Classical Sethuraman stick-breaking is a "lopsided" construction, corresponding to a tree where, at each stage, only the rightmost branch continues to split. The generative formula is:

$L$7

where $L$8. This induces a strong stochastic ordering: $L$9 is usually largest, followed by $L=\infty$0, etc.

In contrast, the balanced, or "multiscale," construction splits each remaining piece at every scale in a fully dyadic tree. Each leaf weight is a product of exactly $L=\infty$1 independent splits, not a random number (as in lopsided stick-breaking). This structure yields clusters or atoms of more uniform size and tunes prior correlations more flexibly, allowing vanishing cross-covariate dependencies at fine scales (Horiguchi et al., 2022).

A summary comparison:

Construction	Tree Type	Weight Formula
Classical (lopsided)	Right-deep	$L=\infty$2
Multiscale (balanced)	Full binary	$L=\infty$3 allocations along path

3. Limit Laws, Multiscale Partitions, and Connection to Permutations

Multiscale stick-breaking extends to combinatorial constructions. For instance, in the "square-cutting" or two-dimensional case, consider a random permutation $L=\infty$4 that permutes $L=\infty$5 blocks of size $L=\infty$6 and within each block. As $L=\infty$7, the normalized cycle-lengths of $L=\infty$8 converge in law to a partition generated by a recursive two-dimensional stick- (or square-) cutting process (Tung, 23 Jan 2025). The limiting partition $L=\infty$9, constructed via an infinite array of $\varepsilon \in \{0,1\}^{\leq L}$0, satisfies a self-similar distributional identity:

$\varepsilon \in \{0,1\}^{\leq L}$1

where $\varepsilon \in \{0,1\}^{\leq L}$2 and $\varepsilon \in \{0,1\}^{\leq L}$3 is a Poisson–Dirichlet (stick-breaking) partition. The largest block's tail is described via the Dickman function convolution equation.

4. Multiscale Stick-Breaking in Bayesian Nonparametric Mixture Models

Multiscale stick-breaking underpins various nonparametric priors for densities and measures. In the multiscale Bernstein polynomial (msBP) approach (Canale et al., 2014), the infinitely-deep binary tree index $\varepsilon \in \{0,1\}^{\leq L}$4 is used: each node carries a kernel (e.g., a Beta or Gaussian), a stopping probability $\varepsilon \in \{0,1\}^{\leq L}$5, and a branch variable $\varepsilon \in \{0,1\}^{\leq L}$6. Weights are:

$\varepsilon \in \{0,1\}^{\leq L}$7

with $\varepsilon \in \{0,1\}^{\leq L}$8 depending on ancestral branching (right or left). The induced density is

$\varepsilon \in \{0,1\}^{\leq L}$9

and the prior mass decays geometrically with depth, promoting local adaptivity and full support under mild hyperparameter choices.

A related construction, the tree-structured stick-breaking process (TSSBP) (Adams et al., 2010), uses nested stick-breakings: a stop-vs-continue break per node (Beta random variable) interleaved with GEM-type splits among children, forming random measures on trees of potentially infinite depth and width.

Alternatively, the multiscale mixture model of (Stefanucci et al., 2020) employs a similar dyadic tree: each node gets "stop-here" $\varepsilon$0 and branching $\varepsilon$1 variables (Beta-distributed) with flexible scale-dependent parameterization to control mass allocation and smoothness. Kernel parameters are drawn hierarchically with location and scale adapting to tree scale.

5. Hierarchical and ψ-Stick-Breaking: Multiscale in Related NPN Mixtures

The ψ-stick-breaking construction (Soriano et al., 2017) introduces an explicit coarse-to-fine allocation for modeling related samples. For $\varepsilon$2 samples, total mass is first split as $\varepsilon$3 (shared components) and $\varepsilon$4 (idiosyncratic components) by a $\varepsilon$5 randomization:

• Shared atoms: allocated by stick-breaking on $\varepsilon$6.
• Sample-specific atoms: allocated by individual stick-breaking steps on $\varepsilon$7. This hierarchy can be generalized to more than two levels (e.g., group-sample multi-levels), leading to arbitrarily deep multiscale mixtures.

Structurally, this resembles a multiscale stick-breaking on an additive partition of mass, whereas msBP and TSSBP allocate recursively via multiplicative partitioning on trees.

6. Posterior Inference Methodologies for Multiscale Stick-Breaking

Posterior inference leverages the recursive structure for efficient computation:

• Slice sampling (Walker’s algorithm and extensions) enables truncation-free inference by augmenting allocations with latent variables $\varepsilon$8 to sample in the (potentially infinite) tree (Canale et al., 2014, Adams et al., 2010, Stefanucci et al., 2020).
• The Pólya–Gamma augmentation facilitates tractable posterior computation in covariate-dependent multiscale stick-breaking, allowing binary regression updates at each internal node, typically reducing per-iteration cost to $\varepsilon$9 for $|\varepsilon| < L$0 data and $|\varepsilon| < L$1 clusters in fully balanced trees (Horiguchi et al., 2022).
• Conjugate updates for stick-length and kernel parameters exploit the independence built into the tree, with Beta posteriors for $|\varepsilon| < L$2 and $|\varepsilon| < L$3, and conjugate distributions for kernel hyperparameters (e.g., normals for means, inverse gammas for variances in Gaussian kernels).

A typical Gibbs or blocked sampler cycles through:

1. Allocation of latent path or node for each observation.
2. Updates for stick-breaking and branching variables node-wise.
3. Updates for kernel or emission parameters.
4. (Where applicable) updating hyperparameters (e.g., via Gamma or Metropolis–Hastings steps).

7. Theoretical and Practical Properties

Multiscale stick-breaking priors possess several key properties:

• Partition of Unity: The sum of allocated masses is a.s. 1 due to the recursive tree structure and appropriately chosen Beta parameters (Stefanucci et al., 2020, Canale et al., 2014).
• Full Support and Local Adaptivity: Under mild conditions, multiscale priors are dense in $|\varepsilon| < L$4 on the space of densities, with the scale-dependent allocation allowing for locally varying smoothness.
• Prior Specification: Hyperparameters at each scale (e.g., those defining distributions for $|\varepsilon| < L$5) control smoothness and mass concentration: small values keep mass coarse (smooth densities), large values increase granularity (wiggly densities).
• Prior Correlation Structure: Balanced (multiscale) constructions permit the induced measures or clustering functions to become arbitrarily weakly correlated as depth increases, in contrast to lopsided constructions which maintain a baseline correlation (Horiguchi et al., 2022).
• Self-Similarity and Recursion: Both the random measures and induced partitions inherit a fundamental multiplicative or convolutional self-similarity, leading to integral or difference-delay equations for functionals of interest (e.g., the Dickman function for largest block size distribution in partitions) (Tung, 23 Jan 2025).
• Scalability: Exploitation of tree sparsity and blocking in inference algorithms allows practical scalability to high-dimensional or large-sample scenarios.

• (Horiguchi et al., 2022) A tree perspective on stick-breaking models in covariate-dependent mixtures
• (Tung, 23 Jan 2025) Cutting a unit square and permuting blocks
• (Canale et al., 2014) Multiscale Bernstein polynomials for densities
• (Adams et al., 2010) Tree-Structured Stick Breaking Processes for Hierarchical Data
• (Soriano et al., 2017) Mixture modeling on related samples by $|\varepsilon| < L$6-stick breaking and kernel perturbation
• (Stefanucci et al., 2020) Multiscale stick-breaking mixture models