Chinese Restaurant Process Overview

• Chinese Restaurant Process is a sequential stochastic process that models random partitions by assigning each new observation (customer) to an existing cluster (table) proportionally to its size, introducing a rich-get-richer effect.
• It underpins Dirichlet process mixtures by providing an exchangeable partition structure, enabling flexible clustering and mixture modeling in Bayesian nonparametrics.
• Advanced extensions include two-parameter, hierarchical, temporal, and distance-dependent variants that adapt the basic model for structured, dynamic data.

The Chinese restaurant process (CRP) is a sequential stochastic process for random partitions in which observations are represented as customers and clusters as tables. In its standard form, a new customer joins an existing table with probability proportional to that table’s current occupancy and starts a new table with probability proportional to a concentration parameter. In Bayesian nonparametrics, this is the exchangeable partition distribution obtained by marginalizing the random probability measure in a Dirichlet process mixture, and it serves as a foundational prior for clustering, mixture modeling, and several structured extensions including hierarchical, temporal, and distance-dependent models (Poux-Médard et al., 2021).

1. Sequential construction and basic properties

In the standard CRP, if customer $i$ arrives after $i-1$ customers have already been seated, and table $k$ currently has $n_k$ customers, then

$$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$

Equivalently, existing tables are weighted by their current sizes and a new table is weighted by $\alpha$. This is the classic rich-get-richer mechanism: larger clusters are more likely to attract future observations (Poux-Médard et al., 2021).

The process is exchangeable in the sense that the probability of a partition depends only on cluster sizes, not on the order in which observations are considered. That exchangeable partition structure is what makes the CRP the canonical partition prior in Dirichlet process mixtures. A standard consequence is that the expected number of occupied clusters after $n$ observations is

$$\mathbb E[K_n] = \sum_{i=1}^n \frac{\alpha}{\alpha+i-1} \sim \alpha \log n,$$

so the number of occupied clusters grows logarithmically in sample size (Poux-Médard et al., 2021).

A common misconception is that the CRP is intrinsically a model of “true” cluster number. The underlying literature distinguishes density estimation from clustering: the CRP is attractive as a flexible prior over partitions, but under Dirichlet process mixtures it can continue to introduce additional occupied components as $n$ grows. This makes the process natural for unbounded latent structure, but not automatically parsimonious when the inferential target is a finite, stable clustering (Lu et al., 2018).

2. Partition law, Dirichlet process, and stick-breaking

A standard Dirichlet process mixture model can be written as

$$G \sim \mathrm{DP}(\alpha, G_0), \qquad \theta_i \mid G \sim G, \qquad x_i \mid \theta_i \sim f(x_i \mid \theta_i).$$

Because draws from a Dirichlet process are discrete almost surely, ties among the $i-1$0 induce a random partition of the observations; after integrating out $i-1$1, that partition law is the CRP (Poux-Médard et al., 2021).

The induced exchangeable partition probability function can be written directly. If $i-1$2 is a partition of $i-1$3, then

$i-1$4

Jeffrey Miller gave an elementary derivation of this law from Sethuraman’s stick-breaking representation, avoiding the usual indirect appeal to abstract Dirichlet process properties. In that construction,

$i-1$5

and labels $i-1$6 are sampled i.i.d. from $i-1$7; forgetting the numeric labels and retaining only the induced partition yields exactly the CRP law above (Miller, 2018).

This equivalence is conceptually important because the two most common views of the Dirichlet process are then seen as complementary rather than competing. Stick-breaking is constructive and ordered; the CRP is partition-valued and exchangeable. The former is often preferable for direct simulation of random measures, while the latter is preferable for collapsed partition-based inference.

3. Two-parameter extensions and permutation representations

A major extension replaces the one-parameter seating rule by the two-parameter $i-1$8-seating rule. If $i-1$9 has $k$0 blocks of sizes $k$1, then customer $k$2 joins existing table $k$3 with probability

$k$4

and starts a new table with probability

$k$5

with admissible parameter range $k$6, $k$7. When $k$8 and $k$9, this reduces to the Ewens setting (Garza et al., 2024).

This two-parameter construction is closely tied to the Pitman–Yor process. In one common parameterization, if

$n_k$0

then the induced partition has exchangeable partition probability function

$n_k$1

The corresponding sequential seating rule is

$n_k$2

and the ordinary Dirichlet-process CRP is recovered when $n_k$3 (Lawless et al., 2018).

The CRP also has a classical permutation interpretation. If each occupied table is viewed as a round table, the cyclic order of customers around that table defines a cycle of a permutation; tables correspond to cycles and table sizes correspond to cycle lengths. In the Ewens case, the induced partition law is Ewens’ sampling formula and the induced random permutation is distributed according to Ewens measure on the symmetric group. This links CRP theory directly to cycle counts, random permutation matrices, and spectral statistics (Garza et al., 2024).

4. Hierarchical, recurrent, and structured extensions

The nested Chinese restaurant process (nCRP) generalizes the CRP from partitions to trees. It assigns probability distributions to infinitely-deep, infinitely-branching trees and can be used as a prior distribution in a Bayesian nonparametric model of document collections. In the associated topic-hierarchy model, documents are modeled as paths down a random tree, and the preferential attachment dynamics leads to clustering according to sharing of topics at multiple levels of abstraction (0710.0845).

A second major development is the hierarchical Chinese restaurant process (HCRP), or Chinese restaurant franchise, which is the collapsed predictive representation of the hierarchical Dirichlet process. In the infinite hidden Markov model, each latent state is a restaurant, transition occurrences are customers, and tables serve dishes corresponding to next states. This representation is statistically attractive because transition measures can be collapsed out, but exact inference becomes difficult when several coupled draws must be resampled jointly; the restricted collapsed draw sampler was introduced precisely to make accurate simultaneous sampling in HCRP-based iHMMs feasible (Makino et al., 2011).

Temporal and recurrent variants alter the exchangeable CRP prior into a history-sensitive one. In the temporally reweighted CRP mixture, the assignment probability is the usual CRP term multiplied by a cohesion function of recent lagged observations,

$n_k$4

so regime membership depends on both cluster occupancy and recent temporal context. The authors explicitly characterize this as a non-exchangeable extension for time series (Saad et al., 2017).

Related dynamic constructions use decayed historical counts. The recurrent CRP assigns document $n_k$5 at epoch $n_k$6 to event $n_k$7 with probability proportional to

$n_k$8

for existing events, and $n_k$9 for a new event, allowing clusters to persist, evolve, die out, and be born over time (Wei et al., 2017). The Dynamic Multi-Relational CRP generalizes the predictive law further by replacing global counts with relation-specific neighbor counts and mixing several such relation-conditioned CRPs through user-specific influence weights, so that world-wide, user-specific, network, and geographic effects can all act simultaneously and evolve across epochs (Lakkaraju et al., 2012).

5. Non-exchangeable modifications and application-driven adaptations

One strand of work modifies the CRP specifically to control the strength of the rich-get-richer effect. The powered Chinese restaurant process replaces the linear weight $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$0 by $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$1, $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$2, yielding

$$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$3

This suppresses the persistence of tiny spurious clusters and discourages opening new clusters when a few large clusters already dominate, but it deliberately abandons exchangeability; order dependence is mitigated in practice by random permutation during Gibbs sampling (Lu et al., 2018). A related powered Dirichlet-process formulation makes the same point in finite-dimensional form: replacing counts by $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$4 separates the baseline tendency to create new clusters from the strength of preferential attachment itself (Poux-Médard et al., 2021).

Another influential generalization is the distance dependent CRP (ddCRP), which replaces direct table assignment by customer-to-customer links: $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$5 Clusters are the connected components of the induced link graph. This breaks exchangeability in a controlled way and makes spatial, temporal, or covariate structure explicit. The ddCRP has been used for image segmentation and object proposal generation, where superpixels link preferentially to adjacent, similar superpixels and posterior samples of segmentations induce ranked object proposals (Lauri et al., 2017). In the non-conjugate setting, ddCRP inference becomes trans-dimensional because link changes alter the number of connected components; a reversible-jump MCMC framework has been developed to handle exactly that difficulty (Marsh et al., 15 May 2026).

Not all descendants of the CRP preserve the original probabilistic semantics. In cognate clustering, a CRP-inspired threshold-free algorithm reassigns each word to the cluster with maximum average similarity unless even the best average similarity is below $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$6, in which case a new cluster is created; the paper is explicit that this is motivated by CRP rather than being a full Bayesian CRP posterior sampler (Rama, 2016). The “Chinese Restaurant Game” likewise keeps the restaurant metaphor but replaces exogenous seating probabilities by strategic Bayesian agents whose utilities decrease with crowding, thereby modeling social learning with negative network externality rather than exchangeable random partitioning (Wang et al., 2011).

6. Limit theorems and asymptotic regimes

The asymptotic theory of CRP-type partitions depends sharply on parameter regime. In the generalized two-parameter CRP with $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$7, $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$8, the number of blocks grows polynomially rather than logarithmically: $$P(z_i = k \mid z_{1:i-1}) = \frac{n_k}{\alpha + i - 1}, \qquad P(z_i = \text{new} \mid z_{1:i-1}) = \frac{\alpha}{\alpha + i - 1}.$$9 The number $\alpha$0 of blocks of size $\alpha$1 concentrates around

$\alpha$2

with coefficient asymptotics $\alpha$3. This gives a finite-$\alpha$4 concentration theory for power-law block-size profiles in the Pitman–Yor regime (Pereira et al., 2018).

For the one-parameter CRP, recent work has refined classical fixed-$\alpha$5 Poisson asymptotics by encoding full small-block growth histories as point measures. After linear time scaling, the point measure of blocks tracked until they reach size $\alpha$6 converges to an inhomogeneous Poisson random measure on

$\alpha$7

with intensity

$\alpha$8

This framework yields functional limit theorems in the Skorokhod topology for singleton counts, the first singleton, and related small-block characteristics (Galganov et al., 8 Feb 2025).

The permutation viewpoint gives a complementary limit theory. For CRP-induced permutations in the $\alpha$9-seating regime with $n$0, weighted cycle-count sums

$n$1

satisfy a functional central limit theorem after normalization by $n$2, with a random Gaussian limit scaled by the $n$3-diversity. This sharply contrasts with the Ewens case $n$4, where small cycle counts converge to independent Poisson variables (Garza et al., 2024).

A different asymptotic picture appears in the disordered CRP, where each table carries an i.i.d. random weight and customers join table $n$5 with probability proportional to $n$6. In that model every fixed table is microscopic,

$n$7

the largest table captures almost all mass in probability,

$n$8

and the largest two tables capture almost all mass almost surely,

$n$9

This breaks the exchangeable Poisson–Dirichlet picture and replaces it with an out-of-equilibrium localization phenomenon driven by table-specific fitness (Björnberg et al., 2023).

The CRP has therefore become both a specific exchangeable partition law and a broader modeling template. In its original form it is the predictive partition distribution of the Dirichlet process; in its two-parameter forms it links to Poisson–Dirichlet laws and random permutations; in hierarchical and recurrent forms it supports trees, franchises, temporal regimes, and structured dependence; and in generalized or non-exchangeable forms it becomes a controlled mechanism for altering preferential attachment, incorporating distances, or adapting cluster creation to domain-specific constraints.