Cluster updates and critical slowing down

For one positive coupling K, rewrite its Boltzmann factor as a constant times a mixture of an absent bond and a present bond that requires equal spins. The probability parameter is one minus the exponential of minus twice K. Given the spins, every parallel nearest-neighbor pair receives an independent bond with that probability, while antiparallel neighbors cannot be bonded. This conditional independence applies to bond variables on a fixed spin assignment; the spins themselves remain correlated in equilibrium.

Given a bond assignment, every connected component must have a common spin sign. At zero field, the two signs have equal conditional weight, independently for different components. Alternating these two conditional draws is a Gibbs update of an enlarged spin-bond measure whose spin marginal is the original Ising distribution. This establishes equilibrium correctness through an explicit auxiliary measure, rather than assuming that any visually coherent domain can be flipped freely. The bond probability is neither the equilibrium probability that two arbitrary spins agree nor a tunable acceptance fraction.

A Swendsen–Wang sweep builds all eligible bonds, finds their connected components, and assigns a fresh random sign to each component. A Wolff update instead chooses a random seed and grows one component with the corresponding bond rule before reversing it. Random seeding biases the selected component toward larger clusters, an essential part of that algorithm’s probability law. Treating a uniformly chosen cluster from a stored list as the same move would require a separate balance argument.

One single-cluster move need not touch a whole lattice. Comparing its autocorrelation in moves directly with that of a local algorithm in sweeps can therefore be misleading. Record average sites visited, cluster-construction cost, and time per effective measurement. Critical slowing is often summarized by a power of correlation length or lattice size at criticality, but its dynamic exponent belongs to the algorithm, observable, and chosen work convention. Cluster updates can dramatically reduce it for suitable models without providing a universal zero exponent or curing every slow mode.

After independent signs are assigned to the components, the conditional expectation of a spin product is one if its sites share a component and zero otherwise. Averaging the connectivity indicator can therefore estimate the correlation with less sign-assignment noise. This statement averages over the conditional cluster signs, not over an arbitrary nonequilibrium spin configuration’s original product. It is an example of replacing a noisy measurement by its conditional expectation, with the full equilibrium average preserved.

The update and the measurement must be matched carefully. Single-cluster estimators can need size or seed-probability factors different from all-cluster estimators. Warmup, finite-size effects, and autocorrelation analysis remain necessary after an algorithmic improvement. At a first-order transition or in a frustrated model, barriers outside this simple cluster representation may remain severe. The binomial instrument below counts independent candidate bond decisions on a fixed eligible set; it does not identify connected components, update a lattice, or measure a critical exponent.