Loop-Cluster Monte Carlo Algorithm for Classical Statistical Models

Lei Zhang, Manon Michel, Eren Metin Elçi, Youjin DengarXiv 2019 · arXiv

Abstract

We introduce a joint model of bond-occupation variables interacting with so-called $q$-flow variables. It leads us to formulate a Loop-Cluster (LC) Monte Carlo algorithm which passes back and forth between the Fortuin-Kasteleyn (FK) bond representation of the $q$-state Potts model and the so-called $q$-flow representation. Together with the Swendsen-Wang (SW) cluster method, the LC algorithm couples the spin, FK and $q$-flow representations of the Potts model. As a result, a single Markov-chain simulation can use an arbitrary combination of the SW, worm, LC and other algorithms, and simultaneously sample physical quantities in any representation. Generalizations to real value $q \geq 1$ and to a single-cluster version are also obtained. Investigation of dynamic properties is performed for $q=2$, $3$ on both the complete graph and toroidal grids of dimension $2\leq d\leq 5$. Our numerical results suggest that the LC algorithm and its single-cluster version are in the same dynamic universality class as the SW and the Wolff algorithm, respectively. Finally, it is shown, for the Potts model undergoing continuous phase transition, the $q$-flow clusters, defined as sets of vertices connected via non-zero flow variables, are fractal objects.

Introduction: The Unified Trinity

The Potts model is a cornerstone of statistical mechanics, used to describe phenomena ranging from magnetism to the behavior of foams. Historically, physicists have studied it using three distinct representations: the original Spin version (discrete states), the Fortuin-Kasteleyn (FK) Bond representation (geometric clusters), and the $q$-flow (Loop) representation (closed loops and flow conservation).

While the mapping between spins and bonds has been understood for decades via the Edwards-Sokal joint model, the connection to the loop representation remained an open question. This fragmentation meant researchers often had to choose a single representation for a simulation, even if a different one was better suited for measuring specific physical properties. This paper introduces the "missing link"—the Loop-Cluster (LC) joint model—which unifies these three perspectives into a single theoretical framework.

The relationship between spin, bond (FK), and loop (q-flow) representations, bridged by the Swendsen-Wang and Loop-Cluster algorithms.

The Loop-Cluster (LC) Algorithm

The authors derive the LC Algorithm from a mathematical framework where bond-occupation variables and $q$-flow variables interact directly. This Monte Carlo scheme allows a simulation to "pass back and forth" between bonds and loops:

  1. Flow-to-Bond: Given a set of loops (flows), bonds are placed locally. If a loop flows through an edge, that edge must be occupied by a bond.
  2. Bond-to-Flow: Given a cluster of bonds, the algorithm constructs a spanning tree and utilizes Kirchhoff’s conservation laws to randomly assign valid flows within that cluster.

This unification enables a single Markov-chain simulation to use an arbitrary combination of the Swendsen-Wang, worm, and LC algorithms, allowing for the simultaneous sampling of physical quantities in any representation.

Efficiency and Dynamic Universality

A critical question for any new Monte Carlo algorithm is its efficiency, particularly near phase transitions where "critical slowing down" can occur. The authors tested the LC algorithm's efficiency by calculating integrated autocorrelation times—a measure of how quickly a simulation explores the state space—across various geometries, including toroidal grids ($d=2$ to $5$) and complete graphs.

Ratio of autocorrelation times (R) between the LC and SW algorithms as a function of system size (L) across different dimensions.

As shown in the numerical results, the ratio of autocorrelation times between the LC and SW algorithms converges to a constant as system size increases. This parity demonstrates that the LC algorithm belongs to the same "dynamical universality class" as the gold-standard Swendsen-Wang and Wolff algorithms, making it a highly efficient tool for studying critical phenomena.

Fractal Geometry of Flow Clusters

Beyond algorithmic efficiency, the LC model provides deep insights into the fractal geometry of physical clusters. The authors discovered that $q$-flow clusters—sets of vertices connected via non-zero flow variables—are fractal objects related to the "backbones" of FK clusters.

By introducing a new parameter ($q_F$), they constructed a hierarchy of fractal objects. For the 2D Ising model ($q=2$), they calculated the fractal dimension $D_{qF} \approx 1.375$, which matches the known external perimeter dimension ($11/8$). This provides a new numerical and theoretical pathway to characterize cluster backbones, a problem that has remained challenging for decades.

Conclusion

The Loop-Cluster algorithm represents a significant step forward in the study of classical statistical models. By bridging the gap between spins, bonds, and loops, it not only provides a versatile new simulation tool but also opens a new window into the fractal nature of phase transitions. Future work may leverage Conformal Field Theory (CFT) to further explore the algebraic symmetries governing these newly discovered fractal hierarchies.

Summary updated

Abstract

We introduce a joint model of bond-occupation variables interacting with so-called $q$-flow variables. It leads us to formulate a Loop-Cluster (LC) Monte Carlo algorithm which passes back and forth between the Fortuin-Kasteleyn (FK) bond representation of the $q$-state Potts model and the so-called $q$-flow representation. Together with the Swendsen-Wang (SW) cluster method, the LC algorithm couples the spin, FK and $q$-flow representations of the Potts model. As a result, a single Markov-chain simulation can use an arbitrary combination of the SW, worm, LC and other algorithms, and simultaneously sample physical quantities in any representation. Generalizations to real value $q \geq 1$ and to a single-cluster version are also obtained. Investigation of dynamic properties is performed for $q=2$, $3$ on both the complete graph and toroidal grids of dimension $2\leq d\leq 5$. Our numerical results suggest that the LC algorithm and its single-cluster version are in the same dynamic universality class as the SW and the Wolff algorithm, respectively. Finally, it is shown, for the Potts model undergoing continuous phase transition, the $q$-flow clusters, defined as sets of vertices connected via non-zero flow variables, are fractal objects.
BibTeX
@misc{zhang2019loop,
  author = {Lei Zhang and Manon Michel and Eren Metin Elçi and Youjin Deng},
  title = {{Loop-Cluster Monte Carlo Algorithm for Classical Statistical Models}},
  year = {2019},
  eprint = {1909.02719},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1909.02719},
}

Loading related content...

Figure