Constant Potts Model (CPM)

• Constant Potts Model is a Potts-type objective used in community detection that rewards intra-community edges and penalizes non-neighbor pairs using a tunable resolution parameter.
• The model decomposes a global Hamiltonian into local utilities, yielding a hedonic-game interpretation with convergence guarantees for better-response dynamics.
• Its fixed-q formulation applies to statistical mechanics and algebraic geometry, linking phase transitions, MLE approximations, and period integrals.

The Constant Potts Model (CPM) is a Potts-type objective used most explicitly in network community detection, where it scores a partition by rewarding intra-community edges and penalizing intra-community non-edges through a resolution parameter $\gamma$. In the recent graph-partitioning literature, CPM is formulated as a global Hamiltonian and, equivalently, as a collection of aligned local utility functions, which yields a hedonic-game interpretation with pseudo-polynomial convergence guarantees for better-response dynamics (Felipe et al., 4 Sep 2025). In a separate statistical-mechanical usage, “constant” refers to fixing the number of spin states $q$, so that one studies either the classical Potts model on a lattice or the fixed-$q$ zero locus of the multivariate Tutte polynomial (Duenas-Herrera et al., 25 Sep 2025, Marcolli et al., 2011).

1. Hamiltonian formulation in community detection

For a partition $\sigma$ of a graph with adjacency matrix $A$, the CPM Hamiltonian is

$$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$

Here, $\delta(\sigma_i=\sigma_j)$ is $1$ when nodes $i$ and $j$ belong to the same community and $q$0 otherwise, while $q$1 controls the granularity of the partition. In this formulation, node pairs contribute

$q$2

so that neighbors placed in the same community contribute $q$3, non-neighbors placed in the same community contribute $q$4, and all other pairs contribute $q$5 (Felipe et al., 4 Sep 2025).

This objective was introduced to address the resolution-limit problem present in earlier methods like modularity. Its structure makes explicit that community quality is not measured only by internal edge density: it is equally shaped by the penalty imposed on grouping non-neighboring vertices together. A plausible implication is that CPM is naturally suited to settings where the distinction between “friends” and “strangers” inside a community must be resolved at a tunable scale rather than by a fixed null model.

2. Local utility decomposition and hedonic-game interpretation

A central development is the decomposition of the global Hamiltonian into local node utilities. For node $q$6 and community $q$7, the local potential is

$q$8

where $q$9 is the number of neighbors of node $q$0 in community $q$1 and $q$2 is the number of non-neighbors of node $q$3 in community $q$4. The associated community and partition potentials are

$q$5

and

$q$6

with $q$7 the number of internal edges and $q$8 the community size (Felipe et al., 4 Sep 2025).

The key structural fact is exact alignment between local and global optimization: if a node moves to improve its own utility, the global partition potential increases by the same amount. This makes CPM an exact potential game, specifically an additively separable hedonic game with non-transferable utility. In this representation, a Nash equilibrium is a partition where no node has an incentive to move, and local optima of the global Hamiltonian coincide with equilibria of the induced game.

This equivalence yields an algorithmic guarantee. When $q$9 is rational, each better-response move increases the potential by at least $\sigma$0, the potential is bounded by $\sigma$1, and better-response dynamics converge in at most $\sigma$2 steps. For the special cases $\sigma$3 and $\sigma$4, convergence is $\sigma$5 (Felipe et al., 4 Sep 2025).

3. Resolution parameter, strict robustness, and equilibrium ranges

The parameter $\sigma$6 mediates a trade-off between maximizing intra-community neighbors and minimizing intra-community non-neighbors. In the relaxed utility

$\sigma$7

low $\sigma$8 prioritizes maximizing friends and yields larger, coarser communities, whereas high $\sigma$9 prioritizes minimizing strangers and yields finer, smaller communities or singletons (Felipe et al., 4 Sep 2025).

The strict robustness criterion strengthens this picture. A node is robust when its assigned community simultaneously maximizes neighbors and minimizes non-neighbors among all candidate communities: $A$0 Partition robustness is then

$A$1

This criterion isolates partitions in which local membership is unambiguous even before the weighted trade-off encoded by $A$2 is invoked (Felipe et al., 4 Sep 2025).

When a node faces a genuine trade-off, the “familiarity index” provides the threshold at which preference switches. For a contemplated move from community $A$3 to community $A$4,

$A$5

If $A$6, the node prefers $A$7; if $A$8, it prefers $A$9; and if $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$0 or $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$1, the choice is unambiguous (Felipe et al., 4 Sep 2025).

Two structural theorems further constrain the equilibrium landscape. If a partition is an equilibrium at $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$2 and at $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$3, then it is an equilibrium for all $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$4. In addition, a partition that is an equilibrium at $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$5 and whose communities are of equal size is an equilibrium for all $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$6, i.e. fully robust (Felipe et al., 4 Sep 2025).

4. Algorithmic behavior and empirical use in community tracking

The empirical analysis of CPM in community tracking emphasizes three metrics: efficiency, robustness, and accuracy, with accuracy measured against ground truth using the Adjusted Rand Index (ARI). In noisy or partially correct initial partitions, robust equilibria are reported to be both quickly found and highly correlated with ground truth (Felipe et al., 4 Sep 2025).

Within this setting, Leiden Phase 1—the local mover based on utility-improving node relocations—already suffices for robust, accurate tracking, with better efficiency and distributability than full Leiden including refinement and aggregation. The reported experiments indicate that CPM-based local optimization recovers accurate partitions even from noisy initializations, outperforming baselines and performing on par or better than spectral methods when the number of communities is larger than $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$7 (Felipe et al., 4 Sep 2025).

The experiments also relate robustness to partition quality. Ground-truth partitions in synthetic planted-partition networks are often fully robust, and partitions with higher robustness more often align with the planted structure. This suggests that robustness is not only a stability notion but also a practical model-selection signal for choosing $$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$8 and for selecting among equilibrium partitions.

5. Fixed-$$\mathcal{H}_{\text{CPM}}(\sigma) = - \sum_{i,j} (A_{ij} - \gamma)\delta(\sigma_i = \sigma_j).$$9 Potts formulations in statistics and algebraic geometry

In a different usage, the classical Potts model is referred to as the Constant Potts Model (CPM) when the number of categories or spin states is fixed. On an $\delta(\sigma_i=\sigma_j)$0 lattice with $\delta(\sigma_i=\sigma_j)$1 categories, the probability mass function is

$\delta(\sigma_i=\sigma_j)$2

where $\delta(\sigma_i=\sigma_j)$3 counts sites of color $\delta(\sigma_i=\sigma_j)$4 and $\delta(\sigma_i=\sigma_j)$5 counts neighboring pairs with the same color. The denominator

$\delta(\sigma_i=\sigma_j)$6

is intractable for any reasonably sized lattice, motivating MCMCMLE and related approximations (Duenas-Herrera et al., 25 Sep 2025).

This fixed-$\delta(\sigma_i=\sigma_j)$7 model inherits the classical phase-transition difficulty. For zero external field, the critical value occurs near

$\delta(\sigma_i=\sigma_j)$8

and for $\delta(\sigma_i=\sigma_j)$9 the model’s typical realizations are nearly all of one color, i.e. ground states. As summarized in the simulation study, the model is suitable for moderate correlation, but it can fit very poorly when spatial correlation is strong and multiple categories remain well represented; then MLE may fail to converge and simulated lattices become unrealistic (Duenas-Herrera et al., 25 Sep 2025).

The same fixed-$1$0 viewpoint appears in algebraic geometry through the Potts partition function

$1$1

where

$1$2

For fixed $1$3, the associated hypersurface is

$1$4

This “constant $1$5 Potts Model” exhibits arithmetic phenomena absent from simpler examples: the tetrahedron graph $1$6 provides a failure of the fibration condition, and for graphs like $1$7, already at $1$8, the point count over finite fields is not a polynomial in $1$9 even after excluding the trivial failures at primes dividing $i$0 (Marcolli et al., 2011).

The fixed-$i$1 setting also generates period integrals of the form

$i$2

For polygonal chains, these periods evaluate to combinations of multiple zeta values; silicate tetrahedral chains are presented as candidates for possible non-mixed Tate periods (Marcolli et al., 2011).

6. Terminological scope and neighboring CPM usages

A recurring source of confusion is that CPM is not unique to the Constant Potts Model. In biological physics, CPM very often denotes the Cellular Potts Model, a lattice-based model in which each site carries a cell index and the Hamiltonian may include adhesion, area, and perimeter terms. In confluent tissues, this model has been used to study equilibrium glassy dynamics, with two regimes controlled by the target perimeter $i$3, sub-Arrhenius relaxation in the low-$i$4 regime, and the absence of the rigidity transition found in vertex models (Sadhukhan et al., 2020). That literature also includes algorithmic developments such as a connectivity algorithm that forbids cell fragmentation and restores detailed balance (Durand et al., 2016), as well as GPU-based parallelization schemes that preserve local update statistics and are up to $i$5–$i$6 orders of magnitude faster than serial implementations (Sultan et al., 2023).

In integrable-model and algebraic-geometry contexts, CPM is also the standard abbreviation for the chiral Potts model. There the rapidity family of the $i$7-state model is isomorphic to the Fermat hypersurface

$i$8

with symmetry encoded by the universal CP group and its modular extension (Roan, 2013). In computational transfer-matrix work, CPM may even denote the Catalan Parallel Method, whose configuration space scales as the Catalan number $i$9 and is improved by the Parallel Family Trees strategy to a sub-Catalan space of size $j$0 (Navarro et al., 2013).

For technical writing, this suggests a simple editorial rule: “Constant Potts Model” should be interpreted from context rather than acronym alone. In network science it denotes the resolution-parameter Hamiltonian for community detection; in statistical mechanics and arithmetic geometry it can mean the fixed-$j$1 Potts model; and in large parts of biophysics and integrability, the same acronym refers to entirely different models.