Random-cluster model

On a finite graph, the random-cluster model assigns an edge configuration $\omega$ probability proportional towhere $o,c,k$ count open edges, closed edges, and open connected components.

For $q\ge 1$, the random-cluster measure lies between the two Bernoulli product measures displayed above. The upper Holley condition ratio is $q^{k(\omega \wedge \eta )-k(\eta )}\ge 1$. The lower ratio is $q^{k(\omega \vee \eta )-k(\omega )+|\eta \setminus \omega |}\ge 1$, because adding an edge reduces the component count by at most one. The comparisons persist under wired limits and give $p_c(1)\le p_c(q)\le q{p}_c(1)/(1+(q-1)p_c(1))$.

The multiplicative factor $q$ assigned to each connected component of a graph in a finite random-cluster measure. At fixed edge probability, the weight of a configuration with $k$ components includes $q^k$. For $q=1$ this reduces to independent bond percolation; increasing $q$ at fixed $p$, for $q\ge 1$, suppresses increasing events by random-cluster weight monotonicity.

The threshold for the origin to have an infinite percolation cluster under a consistently chosen infinite-volume random-cluster measure, conventionally the wired measure: $p_c(q)=\inf \{p:{\theta }^{\mathrm{w}}(p,q)>0\}$. The definition alone makes no assertion about percolation exactly at the threshold.

The finite-volume random-cluster weight monotonicity passes to consistently wired or free infinite-volume measures for finite increasing connection events. Decreasing these events to origin-to-infinity connectivity yields $\theta (p,q_1)\ge \theta (p,q_2)$ for $q_1\le q_2$. The corresponding supercritical parameter sets are nested, and hence $p_c(q_1)\le p_c(q_2)$.

For fixed $p\in (0,1)$ and $1\le q_1\le q_2$, the random-cluster measure at $q_1$ stochastically dominates that at $q_2$. In Holley's condition, Bernoulli factors cancel. Put $a=k(S)-k(S\cup T)$ and $b=k(S\cap T)-k(T)$; supermodularity of graph component count gives $b\ge a\ge 0$, and the weight ratio is $(q_2/q_1{)}^a{q}_2^{b-a}\ge 1$. The comparison also holds with the same boundary wiring in both measures.

Planar random-cluster duality relates edge odds by $[p/(1-p)][p^{\ast }/(1-p^{\ast })]=q$. The fixed point $p=p^{\ast }$ has odds $\sqrt{q}$, giving the displayed parameter. A self-dual parameter does not by itself establish a critical threshold on an arbitrary planar graph. At this parameter the symmetric cluster weight at the self-dual parameter makes the primal-dual cluster counts explicit.

The planar cluster-count identity rewrites $o(\omega )+2k(\omega )$ as $|V|-1+k(\omega )+k({\bar{\omega }}^{\ast })$. At the self-dual parameter of the random-cluster model, $p/(1-p)=\sqrt{q}$, so its weight is $(\sqrt{q}{)}^{o(\omega )+2k(\omega )}$ up to a configuration-independent factor. Absorbing $(\sqrt{q}{)}^{|V|-1}$ into normalization proves the symmetric weight, with the unbounded dual face counted.

For an edge with endpoints already connected without that edge, including any boundary wiring, the conditional open probability is $p$. Otherwise opening the edge merges two components, so the conditional probability is $p/(p+q(1-p))$. For $q\ge 1$, this conditional probability increases when more other edges are open or more boundary vertices are identified. This is the local mechanism of boundary monotonicity of the random-cluster measure.

A boundary condition for a finite-box random-cluster measure is a partition of the boundary vertices. All vertices in one block are identified before counting open components; the identifications carry no Bernoulli edge factors. The weight is $p^{o(\omega )}(1-p{)}^{|E|-o(\omega )}{q}^{k_{\xi }(\omega )}$. A coarser partition makes more identifications. General partitions need not have a planar realization.

Here $\xi ⪯\zeta$ means each block of $\xi$ lies in a block of $\zeta$, and the measure order is stochastic domination of probability measures. The random-cluster single-edge conditional probability is increasing in both exterior open edges and wiring for $q\ge 1$. Couple two heat-bath Markov chains by identical update edges and uniform random variables, starting from ordered states. Order persists, and convergence of the finite chains to their stationary laws proves the displayed inequality. At $q=1$, the law does not depend on the boundary partition.

All boundary vertices are identified into one vertex for component counting. This is the most wired random-cluster boundary condition on a fixed boundary, and maximizes increasing-event probabilities when $q\ge 1$ by boundary monotonicity of the random-cluster measure.

For $q\ge 1$, finite random-cluster measures with all boundary graph vertices identified have a weak infinite-volume limit of a sequence. On a regular lattice it is invariant under translations and lattice reflections and has positive association of the random-cluster model. Its conditional single-edge open probability lies between $p/(p+q(1-p))$ and $p$. Connection events can be approximated by increasing finite-volume path events, so their positive-association inequalities pass to the limit of a sequence.

Positive association of random variables and translation invariance give $C_{m+n}\ge C_m{C}_n$ for axis connection probabilities. The Fekete lemma applied to $-\log C_n$ proves the displayed limit of a sequence and the bound $C_n\le e^{-n\alpha }$. The lower single-edge conditional bound gives $C_n\ge [p/(p+q(1-p)){]}^n$, so the decay rate is finite and nonnegative. It may be zero.

Choose a graph vertex $x=(n,x_2,\dots ,x_d)$ on the boundary of a lattice box. Translation and reflection invariance identify the probabilities of $0\leftrightarrow x$ and $x\leftrightarrow 2n{e}_1$. Their intersection implies $0\leftrightarrow 2n{e}_1$, so positive association of random variables gives $C_{2n}\ge \varphi (0\leftrightarrow x{)}^2$. A union bound and lattice symmetries choose such an $x$ with probability at least the box-boundary connection probability divided by the number of boundary graph vertices. This polynomial factor disappears in the exponential rate, identifying boundary and axial inverse correlation lengths.

The boundary partition has singleton blocks, so no different boundary vertices are identified. Open component counting is the ordinary counting in the finite graph, including isolated vertices. This is the least wired random-cluster boundary condition.

For $0\le p\le 1$ and $q\ge 1$, the FKG inequality implies that increasing functions of a random-cluster configuration have nonnegative covariance. The conclusion can fail for $0<q<1$.

On a finite connected graph, fixing $p=1/2$ and sending $q\downarrow 0$ makes the random-cluster model converge to the uniform law on connected spanning subgraphs: the factor $q^{k(\omega )}$ selects the minimum possible component count $k=1$, while all surviving configurations have equal remaining weight.

If $p,q\downarrow 0$ with $q/p\to 0$, the random-cluster model on a finite connected graph converges to the uniform spanning tree. Relative to a tree, an acyclic configuration with $k>1$ components pays a factor $(q/p{)}^{k-1}$, while every additional cycle pays a factor asymptotic to $p$.

The Edwards-Sokal coupling assigns one common random spin to every random-cluster component. For $q=2$ it couples the random-cluster model to the Ising model.

Given a random-cluster configuration, its components receive independent uniform colors under the Edwards-Sokal coupling. Two vertices have equal Potts spins with probability $1$ if connected and $1/q$ otherwise. Averaging proves the identity. With a wired component fixed to color $1$, the same argument gives normalized spontaneous magnetization $(q{\pi }^1({\sigma }_0=1)-1)/(q-1)={\varphi }^{\mathrm{w}}(0\leftrightarrow \infty )$ in infinite volume. Thus its onset is the wired random-cluster critical probability, transformed to ${\beta }_c=-\log (1-p_c)$ in this convention.

A heat-bath Markov chain repeatedly chooses one coordinate and redraws it from its conditional distribution given all other coordinates. The target distribution is reversible and stationary for these updates.