Random Cluster Model: Overview & Dynamics

• Random Cluster Model is a dependent percolation model where edge configurations are weighted by size and connectivity, unifying percolation and Potts models.
• It connects classical models (e.g., independent bond percolation, Ising) through unique formulations that emphasize connectivity inequalities and phase transition behavior.
• Efficient simulation methods and dynamic update algorithms reveal its critical scaling, mixing times, and applications on both finite graphs and mean-field limits.

The random cluster model is a dependent percolation model on a graph $G=(V,E)$ in which an edge configuration $A\subseteq E$ is weighted by both its size and its connectivity structure. In its standard form, the weight is proportional to $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$, where $k(A)$ is the number of connected components of the spanning subgraph $(V,A)$; equivalently, with $w=p/(1-p)$, the partition function is $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$. This formulation unifies ordinary bond percolation, the ferromagnetic Potts model, and spanning-tree limits, while also providing a common language for connectivity inequalities, phase transitions, and Markov-chain dynamics (Nguyen et al., 13 Jul 2025, Elçi et al., 2015).

1. Formal definition and canonical specializations

On a finite graph, the random-cluster state space is $\{0,1\}^E$, or equivalently the set of edge subsets $A\subseteq E$. The standard finite-volume measure can be written as

$$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$

with homogeneous specialization $A\subseteq E$0, or in edge-weight variables as

$A\subseteq E$1

For planar and inhomogeneous settings, it is also common to use $A\subseteq E$2, so that the measure becomes $A\subseteq E$3 under free boundary conditions (Nguyen et al., 13 Jul 2025, Bencs et al., 2022, Duminil-Copin et al., 2017).

Boundary conditions are intrinsic to the model because cluster counting depends on whether boundary vertices are regarded as distinct or identified. In the finite-volume planar formulation, free boundary conditions count clusters in the original graph, whereas wired boundary conditions identify all boundary vertices into a single vertex before counting. For $A\subseteq E$4, infinite-volume free and wired measures are obtained as weak limits from finite exhaustions (Duminil-Copin et al., 2017).

Several classical models arise as specializations or limits.

Regime	Interpretation
$A\subseteq E$5	ordinary independent bond percolation
integer $A\subseteq E$6	ferromagnetic $A\subseteq E$7-state Potts model
$A\subseteq E$8	Ising case
$A\subseteq E$9 with appropriate scaling	uniform spanning tree

The same cluster-weighting principle has a site analogue. In the site random-cluster model, each occupied-site configuration receives weight

$p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$0

where $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$1 is the number of occupied sites and $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$2 is the number of connected occupied clusters. This is the exact site-percolation counterpart of the Fortuin–Kasteleyn bond random-cluster model (Wang et al., 2014).

2. Correlation structure, pivotal edges, and connectivity inequalities

For $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$3, the random-cluster model lies in the FKG regime: positive association is known, and in particular edge events are positively correlated. By contrast, for $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$4, edge-negative correlation is conjectured rather than proved. A refined version of this question appears already at $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$5, where ordinary covariance vanishes because the model becomes independent bond percolation. The nontrivial object is the renormalized quantity

$p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$6

whose value at $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$7 admits an explicit subtraction-free combinatorial expansion indexed by “paracels” and “smoots.” This shows that the first-order $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$8 coefficient of edge covariance is highly structured even though the covariance itself is exactly zero at $p^{|A|}(1-p)^{|E\setminus A|}q^{k(A)}$9 (Nguyen et al., 13 Jul 2025).

A complementary line of analysis classifies open edges by their role in connectivity. An open edge is a bridge if deleting it increases the number of connected components, and a non-bridge otherwise. If $k(A)$0 denotes bridge density and $k(A)$1 the open-edge density, then for every finite graph and every $k(A)$2,

$k(A)$3

This exact identity ties a local pivotality observable to the thermodynamic edge density. On the square lattice at criticality it yields

$k(A)$4

The same framework also leads to exact fluctuation formulas for the number of bridges and to the notion of bridge load, which controls the point where percolation clusters are maximally fragile (Elçi et al., 2015).

Connectivity inequalities beyond standard FKG behavior are subtler. The bunkbed problem asks whether, in the bunkbed graph $k(A)$5, one always has

$k(A)$6

Major positive results extend from percolation to all random-cluster measures for complete graphs, complete bipartite graphs, and the limit $k(A)$7. At the same time, the inequality is not true in general for

$k(A)$8

The same work identifies arboreal-gas and almost-spanning-tree regimes in which positivity re-emerges, together with new correlation inequalities related to Rayleigh-type phenomena (Ayyer et al., 23 Sep 2025).

3. Criticality, universality, and scaling theory

A particularly sharp universality theory is available in two dimensions on isoradial graphs. For the canonical integrable edge weights determined by the rhombus angles, the random-cluster model is critical at $k(A)$9 for every $(V,A)$0. The phase transition is continuous for $(V,A)$1 and discontinuous for $(V,A)$2. In the continuous regime the model has strong RSW bounds, no infinite cluster in the wired measure, and polynomial one-arm bounds; for $(V,A)$3, the free and wired critical states differ, with exponential decay in the free state and an infinite cluster in the wired state. Conditional on existence of arm exponents, the arm-event exponents are universal across doubly periodic isoradial graphs, including triangular and hexagonal lattices (Duminil-Copin et al., 2017).

Beyond planar integrability, the model admits general exponent inequalities on arbitrary transitive weighted graphs. For $(V,A)$4, a differential inequality for the cluster-volume tail $(V,A)$5 yields

$(V,A)$6

whenever the standard susceptibility, tail, and gap exponents are well defined. The same method proves sharp subcritical decay of cluster volume: $(V,A)$7 for both free and wired infinite-volume measures, and even in infinite-range weighted settings (Hutchcroft, 2019).

On sparse regular graphs, the thermodynamic limit can be identified explicitly. If $(V,A)$8 is an essentially large girth sequence of $(V,A)$9-regular graphs, then for $w=p/(1-p)$0 and $w=p/(1-p)$1,

$w=p/(1-p)$2

The resulting free energy exhibits a phase transition at

$w=p/(1-p)$3

For $w=p/(1-p)$4, the free energy is the symmetric value $w=p/(1-p)$5; for $w=p/(1-p)$6, it is strictly larger, and for $w=p/(1-p)$7 this transition is first order. The site random-cluster model on the square lattice exhibits the same long-distance critical behavior as the bond random-cluster model: for $w=p/(1-p)$8, the measured exponents $w=p/(1-p)$9 and $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$0 agree with the Coulomb-gas predictions for the bond model, while the transition becomes first order for larger $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$1, with explicit hysteresis and double-peak signatures at $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$2 (Bencs et al., 2022, Wang et al., 2014).

4. Dynamics, mixing times, and sampling algorithms

The canonical local dynamics is single-edge heat-bath Glauber dynamics. If the selected edge is a cut edge in the current open subgraph, then its heat-bath insertion probability is

$Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$3

otherwise it is $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$4. On the $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$5 box in $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$6, for every $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$7 and every $Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$8, the mixing time is

$Z_G(q,w)=\sum_{A\subseteq E} q^{k(A)}w^{|A|}$9

and this order is optimal. The proof uses a random-cluster version of spatial mixing and a new disagreement-propagation argument tailored to long-range connectivity dependence (Blanca et al., 2015).

In mean field, the phase structure appears directly in the dynamics. On the complete graph with $\{0,1\}^E$0, the Chayes–Machta dynamics has mixing time $\{0,1\}^E$1 when $\{0,1\}^E$2, and $\{0,1\}^E$3 when $\{0,1\}^E$4. The same metastable window produces exponential slowdown for local heat-bath dynamics. For $\{0,1\}^E$5, the window collapses to the critical point; for $\{0,1\}^E$6, there is a genuine coexistence interval $\{0,1\}^E$7 reflecting first-order behavior (Blanca et al., 2014).

From an algorithmic perspective, a decisive issue is whether connectivity updates can be implemented efficiently. Sweeny’s single-bond dynamics updates the random-cluster configuration directly, but each step requires determining whether the chosen edge is pivotal for connectivity. An implementation based on dynamic connectivity data structures reduces connectivity queries to amortized $\{0,1\}^E$8 and insertions or deletions to amortized $\{0,1\}^E$9, making the algorithm asymptotically more efficient than BFS-based approaches. This is especially significant for noninteger $A\subseteq E$0 and for $A\subseteq E$1, where spin-based cluster algorithms are unavailable (Elçi et al., 2013).

On sparse random graphs, dynamics and computation separate sharply. For graphs with prescribed degree sequence and bounded average branching $A\subseteq E$2, the random-cluster Glauber dynamics mixes in optimal

$A\subseteq E$3

throughout the full high-temperature uniqueness regime $A\subseteq E$4, while Potts Glauber dynamics in the same regime can require

$A\subseteq E$5

steps. On random $A\subseteq E$6-regular graphs at all temperatures, worst-case mixing remains exponentially slow near the ordered/disordered transition, but initialized dynamics mixes rapidly within the appropriate phase: all-out initialization in the disordered phase and all-in initialization in the ordered phase. Large-$A\subseteq E$7 polymer methods also show that random-cluster Glauber and Swendsen–Wang/Chayes–Machta dynamics are exponentially slow on an open interval around the critical temperature, even though polynomial-time approximate counting and sampling remain available by non-MCMC methods (Blanca et al., 2021, Galanis et al., 2023, Helmuth et al., 2020).

5. Mean-field, tree limits, and random-graph formulations

On the complete graph $A\subseteq E$8, the mean-field random-cluster model is parametrized by $A\subseteq E$9. Its equilibrium phase transition occurs at

$$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$0

and in the ordered phase the giant-component density $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$1 is the largest $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$2 solving

$$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$3

This makes the largest component the natural scalar order parameter for both equilibrium and dynamics in mean field (Blanca et al., 2014).

On trees, the wired model exhibits a richer low-temperature structure. For the infinite $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$4-regular wired tree, the paper on low-temperature uniqueness introduces the message recursion

$$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$5

with homogeneous map $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$6, $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$7. It proves that for every $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$8, $$\phi_{G,p,q}(\omega)\propto q^{k(\omega)}\prod_{e\in E} p_e^{\omega_e}(1-p_e)^{1-\omega_e},$$9, and $A\subseteq E$00, there is a unique random-cluster Gibbs measure on the infinite $A\subseteq E$01-regular wired tree. In the same regime, random-cluster Glauber dynamics on finite wired trees mixes in

$A\subseteq E$02

and on random $A\subseteq E$03-regular graphs the same near-linear behavior holds for $A\subseteq E$04 (Blanca et al., 22 Apr 2026).

Large-$A\subseteq E$05 random regular graphs exhibit explicit phase coexistence. For $A\subseteq E$06, $A\subseteq E$07 sufficiently large, and a random $A\subseteq E$08-regular graph, the pressure per vertex exists, the intermediate region is exponentially negligible, and the disordered and ordered phases dominate on opposite sides of the transition. Away from criticality, the local weak limits are the free and wired random-cluster measures on the infinite $A\subseteq E$09-regular tree. At criticality, the two phase weights converge jointly to

$A\subseteq E$10

for a positive nonconstant random variable $A\subseteq E$11, giving a finite-size scaling description of coexistence on random graphs (Helmuth et al., 2020).

External fields can also be incorporated. For $A\subseteq E$12, $A\subseteq E$13, and field $A\subseteq E$14, the partition function

$A\subseteq E$15

admits a rank-2 approximation and an exact reformulation as an extended Ising model

$A\subseteq E$16

This yields a Bethe lower bound for the pressure in general and, on random $A\subseteq E$17-regular graphs, equality of quenched and annealed pressures together with a first-order transition line

$A\subseteq E$18

A plausible implication is that mixed-sign vertex fields $A\subseteq E$19 are the main remaining obstruction to a full random-graph theory beyond the regular or high-field cases (Can et al., 22 Mar 2025).

6. Generalizations beyond bond models on fixed finite graphs

The random-cluster principle extends in several directions by replacing either the microscopic degrees of freedom or the notion of “cluster.” The site random-cluster model replaces occupied bonds by occupied sites and keeps the cluster fugacity $A\subseteq E$20; on the square lattice it has a dedicated cluster Monte Carlo algorithm based on color-assignation and Swendsen–Wang ideas, and its critical exponents coincide with those of the bond model (Wang et al., 2014).

A higher-dimensional extension is the plaquette random-cluster model. For a finite cubical complex $A\subseteq E$21, the $A\subseteq E$22-dimensional model assigns to an $A\subseteq E$23-dimensional percolation subcomplex $A\subseteq E$24 the weight

$A\subseteq E$25

Here graph connectivity is replaced by $A\subseteq E$26-dimensional cohomology. The model is dual to a $A\subseteq E$27-dimensional plaquette random-cluster model under

$A\subseteq E$28

and it has free and wired infinite-volume measures, stochastic monotonicity, FKG positive association, and uniqueness except possibly at countably many $A\subseteq E$29 (Duncan et al., 2024).

A continuum version is obtained by weighting the stationary Boolean model in $A\subseteq E$30 by the number of connected components $A\subseteq E$31 of the germ–grain structure. With Poisson reference intensity $A\subseteq E$32, the formal density is $A\subseteq E$33. In the infinite-volume regime, existence is proved for bounded radii and all $A\subseteq E$34, and also for integrable radii $A\subseteq E$35 when $A\subseteq E$36. In the non-integrable regime $A\subseteq E$37, the model is non-unique for small $A\subseteq E$38 when $A\subseteq E$39 is an integer larger than $A\subseteq E$40, via a continuum Fortuin–Kasteleyn representation through a Widom–Rowlinson model (Dereudre et al., 2015).

The random-cluster viewpoint can even be transported to continuous-spin models. For the Villain model and, more indirectly, the XY model, one can define bond or cable-system cluster objects whose connectivity events control Fourier observables. In the Villain case, with boundary condition $A\subseteq E$41 on $A\subseteq E$42, the paper proves

$A\subseteq E$43

and for the second harmonic,

$A\subseteq E$44

This suggests a broadened notion of random-cluster representation in which cluster geometry encodes spin correlations even when there is no classical FK partition function of the form $A\subseteq E$45 (Dubédat et al., 2022).

Taken together, these developments show that the random cluster model is not merely a single lattice measure but a general connectivity-weighted paradigm. Depending on context, the relevant “cluster” may be a graph component, a topological homology class, a connected germ–grain component, or an auxiliary reflection-invariant object. This suggests that the lasting content of the model is the coupling between local occupation variables and a global notion of connectedness, rather than any one particular choice of state space.