Color-Preserving Mean-Field Theory

• Color-preserving mean-field theory is a framework that retains intrinsic ‘color’ information—whether temporal noise correlations or discrete Potts labels—in the macroscopic limit.
• It applies extended state-space methods and spectral techniques to derive nonlocal Fokker–Planck equations and evaluate phase transitions in noisy or spin glass systems.
• The approach provides practical insights into stability, symmetry preservation, and critical behavior by contrasting dynamical operator theory with combinatorial free-energy methods.

Color-preserving mean-field theory denotes a class of mean-field constructions in which a notion of “color” is retained rather than integrated out. In one usage, “color” means the temporal correlation structure of colored noise: the macroscopic limit is a McKean–Vlasov or nonlocal Fokker–Planck equation posed on an extended state space that includes the noise variables themselves, so the generator of the microscopic noise survives in the mean-field description (Gomes et al., 2019). In a second usage, “color” means the discrete states of a Potts system: preservation of color symmetry means that the thermodynamic limit does not prefer any Potts color, so balanced and unconstrained free energies coincide in the high-temperature phase (Kim, 2 Mar 2026). The shared principle is that the mean-field limit is constructed so as not to erase the structural symmetry or correlation encoded by color.

1. Terminological scope and conceptual core

In interacting diffusions with colored noise, the relevant microscopic system is a Desai–Zwanzig–type model of $N$ weakly interacting particles in a bistable potential $V$, with Curie–Weiss mean-field coupling and colored additive noise. The colored noise processes are themselves specified by finite-dimensional SDEs, so the full particle system is Markov in an extended state space $(x,\mathbf y)$, where $\mathbf y$ denotes the auxiliary noise variables. The mean-field limit is therefore not a scalar equation in $x$ alone, but a joint equation for the one-particle law $\rho(x,\mathbf y,t)$ (Gomes et al., 2019).

In Potts spin glasses, the relevant microscopic system is the complete-graph mean-field Potts model with $\kappa$ colors. Here “color preservation” refers to invariance under permutations of the color labels and to the persistence of balanced empirical color frequencies in the thermodynamic limit. The paper formalizes this by saying that the model preserves color symmetry at inverse temperature $\beta$ if the balanced and unconstrained limiting free energies coincide, and at zero temperature if the balanced and unconstrained limiting ground-state energies coincide (Kim, 2 Mar 2026).

These two uses are mathematically distinct. One preserves noise color by augmenting the state space and retaining the exact noise generator; the other preserves Potts color symmetry by proving that the symmetric thermodynamic sector remains dominant. The common theme is that mean-field closure is not achieved by averaging away the color variable, but by elevating it to an explicit macroscopic descriptor.

2. Extended-state mean-field theory for colored-noise diffusions

The colored-noise construction in (Gomes et al., 2019) begins from weakly interacting particles in the symmetric double-well potential

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$

with mean-field coupling strength $\theta$. The paper considers four noise models: scalar Ornstein–Uhlenbeck noise, harmonic noise given by the first component of a two-dimensional OU process, a non-Gaussian bistable noise process with

$V$0

and a shifted, tilted non-Gaussian process

$V$1

with $V$2 chosen so that the stationary noise has mean zero (Gomes et al., 2019).

Because the colored noises are themselves Markovian after state-space augmentation, the mean-field limit is a McKean–Vlasov or nonlocal Fokker–Planck equation for $V$3 with self-consistency constraint

$V$4

The operator in the auxiliary variables is exactly the Fokker–Planck generator of the underlying noise SDE: for scalar OU noise,

$V$5

for harmonic noise,

$V$6

and for the non-Gaussian models,

$V$7

This is the precise sense in which temporal color is preserved: the evolution in $V$8 is not replaced by an effective diffusion in $V$9, and the joint law $(x,\mathbf y)$0 retains the full statistics generated by the colored process (Gomes et al., 2019).

A central structural consequence is the loss of the white-noise gradient-flow representation. In the colored setting, the stationary flux in extended space need not vanish, detailed balance is generally absent, and explicit stationary densities of Gibbs form are no longer available. The stationary mean-field problem becomes an implicit fixed-point problem for $(x,\mathbf y)$1 and its first moment in $(x,\mathbf y)$2, rather than an explicit variational formula.

3. Correlation time, white-noise limits, and phase transitions

The small-correlation-time analysis in (Gomes et al., 2019) rescales the noise as

$(x,\mathbf y)$3

with $(x,\mathbf y)$4 controlling the correlation time and $(x,\mathbf y)$5 chosen so that

$(x,\mathbf y)$6

The paper reports

$(x,\mathbf y)$7

Accordingly, $(x,\mathbf y)$8 functions as the correlation time: small $(x,\mathbf y)$9 yields weakly colored noise close to white noise, while larger $\mathbf y$0 corresponds to longer memory (Gomes et al., 2019).

In the limit $\mathbf y$1, the interacting particle system converges to the classical white-noise model, and the mean-field equation reduces to

$\mathbf y$2

In that white-noise setting, the effective potential is

$\mathbf y$3

and the stationary density has the explicit form

$\mathbf y$4

The classical bifurcation is a pitchfork: for high temperature there is a single stable equilibrium with $\mathbf y$5, and at a critical inverse temperature $\mathbf y$6 two stable nonzero equilibria $\mathbf y$7 appear (Gomes et al., 2019).

The color-preserving theory shows that the same qualitative picture persists for symmetric colored noises—OU, harmonic, and the symmetric non-Gaussian bistable model—but with a correlation-time-dependent critical temperature $\mathbf y$8. For the asymmetric NS noise, the two stable branches become unequal and separated. The paper further identifies the order of the first nonzero correction to the stationary $\mathbf y$9-marginal in powers of $x$0: OU and B have first correction of order $x$1, NS has first correction of order $x$2, and harmonic noise has first correction of order $x$3 (Gomes et al., 2019). For OU noise, the expansion is given explicitly as

$x$4

with $x$5 chosen to preserve normalization.

The perturbative self-consistency relation

$x$6

yields an approximate bifurcation diagram and a criterion for the shifted critical point. For symmetric colored noises, increasing $x$7 shifts the bifurcation to lower $x$8 than in the white-noise case, while harmonic noise remains closer to the white-noise limit because its corrections start only at order $x$9 (Gomes et al., 2019). A plausible implication is that the phase boundary is controlled not merely by noise intensity but by the detailed form of the noise generator and stationary law.

4. Hermite spectral computation and Monte Carlo validation

To solve both linear and nonlinear, local and nonlocal Fokker–Planck equations without assuming gradient structure, (Gomes et al., 2019) develops a Hermite spectral method. In one dimension, the white-noise Fokker–Planck operator is transformed and expanded in weighted Hermite functions. In the colored setting, the method uses tensorized Hermite bases in the extended variables; for an OU-type $\rho(x,\mathbf y,t)$0 system the basis takes the form

$\rho(x,\mathbf y,t)$1

and the approximation is

$\rho(x,\mathbf y,t)$2

The transformed operator becomes a sparse matrix because differentiation and polynomial multiplication act locally under Hermite recursion (Gomes et al., 2019).

For the McKean–Vlasov term, the order parameter is computed from the spectral coefficients,

$\rho(x,\mathbf y,t)$3

and the mean-field drift induces a quadratic nonlinearity in the coefficient ODEs. The implementation uses either Runge–Kutta 45 for transients or a semi-implicit scheme. Several numerical details are structurally important: rectangular index sets are necessary to capture the white-noise limit $\rho(x,\mathbf y,t)$4 correctly, triangular sets can introduce artificial drift or “parasitic bias” for non-Gaussian noise, and the Hermite scaling parameters are tuned to the width of the stationary law (Gomes et al., 2019).

The paper validates the spectral solver against Monte Carlo simulation of the microscopic particle system using Euler–Maruyama, with $\rho(x,\mathbf y,t)$5 to resolve the fast noise dynamics. The reported comparisons show excellent agreement for small and moderate correlation times. For OU noise, spectral solutions, perturbative predictions, and Monte Carlo estimates agree well for $\rho(x,\mathbf y,t)$6, while the asymptotics degrade for larger $\rho(x,\mathbf y,t)$7 but the spectral and Monte Carlo results remain consistent. For harmonic noise, spectral and Monte Carlo agreement is excellent for small $\rho(x,\mathbf y,t)$8, with slight discrepancy at large $\rho(x,\mathbf y,t)$9 for $\kappa$0, attributed to limited spectral resolution. For the non-Gaussian B and NS models, the spectral method matches the asymptotics, and the Monte Carlo results are mainly reported qualitatively (Gomes et al., 2019).

These computations are integral to the color-preserving construction rather than an auxiliary check: because the extended-space stationary problem is non-gradient and generally lacks a closed-form invariant measure, numerical resolution of the joint density $\kappa$1 is part of the theory itself.

5. Potts spin glasses and preservation of color symmetry

In the Potts-spin-glass setting of (Kim, 2 Mar 2026), the configuration space is

$\kappa$2

with empirical color-frequency vector

$\kappa$3

Balanced configurations satisfy

$\kappa$4

The Hamiltonian is

$\kappa$5

with i.i.d. standard normal couplings, partition function

$\kappa$6

and free energy

$\kappa$7

The model is color-symmetric because the Hamiltonian depends on $\kappa$8 only through equality indicators and is invariant in law under permutations of the color labels (Kim, 2 Mar 2026).

The paper defines color symmetry preservation at inverse temperature $\kappa$9 by

$\beta$0

and at zero temperature by equality of the balanced and unconstrained limiting ground-state energies. Its main high-temperature result for $\beta$1 introduces the threshold

$\beta$2

and proves that for $\beta$3,

$\beta$4

The paper identifies this as the replica-symmetric solution of the balanced model with constant order-parameter path

$\beta$5

so all colors remain equivalent in the high-temperature phase (Kim, 2 Mar 2026).

For $\beta$6, the argument is different. Using the correspondence with the SK model and its gauge symmetry, the paper proves that unbalanced configurations occur with exponentially small probability at all temperatures $\beta$7. Specifically,

$\beta$8

for every $\beta$9, which implies color symmetry preservation for all $V(x)=\frac{x^4}{4}-\frac{x^2}{2},$0 in the two-color case (Kim, 2 Mar 2026).

The meaning of “color-preserving” here is therefore thermodynamic and symmetry-theoretic. The mean-field phase is called color-preserving not because auxiliary color variables are added to the state space, but because the limiting free energy is attained in the balanced sector and does not select any preferred Potts color.

6. Technical mechanisms, comparisons, and open directions

The Potts-spin-glass proof in (Kim, 2 Mar 2026) relies on balanced centering, overlap matrices, and a second moment method. The centered Hamiltonian is

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$1

A key point is that this centering does not change the free energies, because the difference from the uncentered Hamiltonian is a $V(x)=\frac{x^4}{4}-\frac{x^2}{2},$2-independent Gaussian constant. The covariance is then expressed in terms of the overlap matrix

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$3

with

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$4

For balanced configurations,

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$5

The second moment ratio is therefore controlled by the competition between the energetic term $V(x)=\frac{x^4}{4}-\frac{x^2}{2},$6 and the entropy term given by the large-deviation rate function

$V(x)=\frac{x^4}{4}-\frac{x^2}{2},$7

with local quadratic control supplied by a KL-divergence expansion and the far-from-uniform regime controlled using the Ellis–Wang analysis of the non-disordered mean-field Potts model (Kim, 2 Mar 2026).

The colored-noise theory in (Gomes et al., 2019) relies on a different mechanism. There is no balanced combinatorial sector and no free-energy reduction to a symmetry-constrained variational problem. Instead, the technical core is the exact embedding of the noise dynamics into the McKean–Vlasov state space, perturbative expansion in the correlation-time parameter $V(x)=\frac{x^4}{4}-\frac{x^2}{2},$8, and sparse Hermite discretization of a non-gradient extended Fokker–Planck operator. The preservation principle is therefore dynamical and operator-theoretic rather than combinatorial and entropic.

The contrast between the two theories clarifies a recurring ambiguity in the term “color.” In (Gomes et al., 2019), color means nonzero temporal autocorrelation and the decisive object is the generator $V(x)=\frac{x^4}{4}-\frac{x^2}{2},$9. In (Kim, 2 Mar 2026), color means one of $\theta$0 Potts labels and the decisive object is the balanced overlap sector around $\theta$1. The same phrase nonetheless signals an analogous methodological refusal to replace structured microscopic information by an indiscriminate mean-field average.

Several open problems are explicit in the supplied material. For the colored-noise diffusions, existence, uniqueness, stability, and bifurcation analysis for the full non-gradient McKean–Vlasov equation remain largely open, as do generalized Langevin models, multiplicative colored noise, and efficient high-dimensional spectral solvers (Gomes et al., 2019). For Potts spin glasses with $\theta$2, the exact critical temperature of color-symmetry breaking is not identified; the paper proves high-temperature preservation, cites symmetry breaking for certain large-$\theta$3 regimes, and conjectures breaking at zero temperature for all $\theta$4 (Kim, 2 Mar 2026). Taken together, these works show that “color-preserving mean-field theory” is best understood not as a single formalism but as a research program in which mean-field limits are designed to retain, rather than homogenize away, a distinguished correlation or symmetry structure.