Graph Allen–Cahn Equation

• Graph Allen–Cahn equation is a discrete analogue of the classical phase-field model that replaces the Laplacian with a graph Laplacian to study phase transitions on networked structures.
• The formulation leverages variational methods via graph Ginzburg–Landau functionals and double-obstacle potentials, ensuring convergence to graph total variation and reliable threshold dynamics.
• Analytical frameworks extend to stochastic PDEs on metric graphs and studies on hyperbolic graphs, linking interface evolution with curvature motion and applications in image segmentation.

The graph Allen–Cahn equation denotes a family of discrete and network-based analogues of the classical Allen–Cahn phase-field equation in which the order parameter is defined on a graph, a metric network, or, in a different usage, on a surface represented as a graph over a planar domain. In the most direct discrete analogue, one replaces the continuum Laplacian by a graph Laplacian and studies an evolution of the form $u_t=-\epsilon^{-2}f(u)+Lu$, while the stationary variational form on an infinite graph is $\Delta_g(x)-V'(x_g)=0$. Across these settings, the subject links graph Ginzburg–Landau energies, threshold dynamics such as MBO, stochastic evolution on finite networks, and variational or asymptotic connections to graph total variation and graph mean-curvature-type motion (Olshanskii et al., 2020, Budd et al., 2019, Mramor, 2016, Kovács et al., 2020).

1. Canonical discrete formulations

On a finite weighted graph $G=(V,E)$, the basic unknown is a vertex function $u:V\to \mathbb{R}$. In the finite-graph formulation analyzed by Budd and van Gennip, the formal graph Allen–Cahn dynamics is

$\frac{du}{dt}=-\Delta u-W'(u),$

with $\Delta$ the graph Laplacian and $W$ a double-well or double-obstacle potential. In the surface-to-graph transfer described for evolving surfaces, the direct discrete analogue is stated as

$u_t=-\epsilon^{-2}f(u)+Lu,$

where $L$ is the graph Laplacian (Budd et al., 2019, Olshanskii et al., 2020).

A second canonical form is elliptic rather than evolutionary. For an infinite, connected graph $\Gamma=(G,E)$ with path metric, Mramor studies the steady Allen–Cahn equation

$\Delta_g(x)-V'(x_g)=0$0

where

$\Delta_g(x)-V'(x_g)=0$1

Here $\Delta_g(x)-V'(x_g)=0$2 is a $\Delta_g(x)-V'(x_g)=0$3 double-well potential with two non-degenerate absolute minima $\Delta_g(x)-V'(x_g)=0$4. This equation is the Euler–Lagrange equation of a discrete energy involving a local graph-gradient contribution and the potential $\Delta_g(x)-V'(x_g)=0$5, and global minimizers are defined by comparison with finitely supported perturbations (Mramor, 2016).

In the finite-graph setting, the variational structure is formulated through a graph Ginzburg–Landau functional with a graph-gradient term and a potential term. For the double-obstacle potential, minimizers are constrained to $\Delta_g(x)-V'(x_g)=0$6 at each vertex, and the functional $\Delta_g(x)-V'(x_g)=0$7-converges to graph total variation as $\Delta_g(x)-V'(x_g)=0$8 (Budd et al., 2019). This places graph Allen–Cahn dynamics within the same diffuse-interface paradigm that underlies continuum phase-field models, but with the geometry encoded combinatorially.

2. Variational dynamics and the MBO interpretation

A major development in finite-graph Allen–Cahn theory arose from applications in image segmentation and semi-supervised learning. Bertozzi and Flenner developed an algorithm based on the Allen–Cahn gradient flow of a graph Ginzburg–Landau functional, and Merkurjev, Kostić, and Bertozzi introduced a graph MBO variant. Budd and van Gennip gave a rigorous justification for using the MBO scheme in place of graph Allen–Cahn flow by adopting the double-obstacle potential and formulating the dynamics through a variational inequality and subdifferential flow (Budd et al., 2019).

For the non-differentiable double-obstacle potential, the evolution is written as

$\Delta_g(x)-V'(x_g)=0$9

with $G=(V,E)$0 constrained to $G=(V,E)$1 vertexwise. In this setting, existence is proved for any initial datum in $G=(V,E)$2, solutions with the same initial data are unique, solutions are globally Lipschitz in time, and a comparison principle holds. The graph Ginzburg–Landau functional is almost everywhere non-increasing along the flow, and a Lipschitz stability estimate is established between two solutions with different initial data (Budd et al., 2019).

The semi-discrete scheme is an implicit Euler time discretization: $G=(V,E)$3 Its special role is that, when $G=(V,E)$4, the update coincides with the graph MBO scheme. In the standard threshold form, one first evolves by graph heat flow, $G=(V,E)$5, and then thresholds,

$G=(V,E)$6

Budd and van Gennip further prove that, as $G=(V,E)$7 with $G=(V,E)$8 fixed, the semi-discrete iterates converge to the Allen–Cahn trajectory, and that the discrete Lyapunov functional associated with the scheme also $G=(V,E)$9-converges to graph total variation (Budd et al., 2019).

This finite-graph theory is significant because it supplies a rigorous PDE-style justification for a widely used thresholding algorithm. It also clarifies that the graph Allen–Cahn equation with double-obstacle potential is not merely a heuristic relaxation of graph cut problems, but a well-posed gradient-flow system with comparison, monotonicity, and variational convergence properties.

3. Minimal solutions on hyperbolic graphs

On infinite graphs, the Allen–Cahn equation has been studied as a variational problem at infinity. Mramor considers an infinite, locally uniformly bounded, connected graph that is Gromov-hyperbolic, admits a boundary at infinity $u:V\to \mathbb{R}$0, and satisfies additional visual and isoperimetric assumptions. Under these conditions, the graph can be compactified to $u:V\to \mathbb{R}$1 with the visual metric, and the Allen–Cahn equation becomes a discrete boundary-value problem with prescribed asymptotic phases (Mramor, 2016).

The main existence theorem asserts that there are non-constant, uniformly bounded global minimizers with prescribed asymptotic behaviours. More precisely, if the boundary at infinity is partitioned into appropriate regions $u:V\to \mathbb{R}$2, then there exists a global minimizer $u:V\to \mathbb{R}$3 such that $u:V\to \mathbb{R}$4 converges uniformly to $u:V\to \mathbb{R}$5 or $u:V\to \mathbb{R}$6 along geodesic rays headed toward points of $u:V\to \mathbb{R}$7 or $u:V\to \mathbb{R}$8, respectively. In the terminology of the paper, this solves a minimal Dirichlet problem at infinity for the discrete Allen–Cahn equation (Mramor, 2016).

The assumptions are geometric rather than purely analytic. They include Gromov-hyperbolicity, uniform local finiteness, a visual property, and an isoperimetric inequality with exponent $u:V\to \mathbb{R}$9. The paper notes that these are natural analogues of geometric and analytic properties on negatively curved manifolds and are automatically satisfied on Cayley graphs of hyperbolic groups (Mramor, 2016).

In phase-field terms, these minimizers describe energy-minimising steady-state phase transitions whose “phase at infinity” is prescribed by asymptotic directions. This makes the hyperbolic-graph theory conceptually different from finite-graph diffusion-and-threshold dynamics: the emphasis is not time evolution or clustering, but the existence of globally minimizing stationary interfaces shaped by the large-scale negative curvature of the graph.

4. Stochastic Allen–Cahn equations on finite networks

A distinct but closely related setting is the stochastic Allen–Cahn equation on a finite network represented by a finite graph. Kovács and Sikolya consider a finite, connected graph with edges $\frac{du}{dt}=-\Delta u-W'(u),$0 and vertices $\frac{du}{dt}=-\Delta u-W'(u),$1, where each edge is parameterized by $\frac{du}{dt}=-\Delta u-W'(u),$2. Functions on the graph are tuples $\frac{du}{dt}=-\Delta u-W'(u),$3, and the main state space is the Banach space $\frac{du}{dt}=-\Delta u-W'(u),$4 of functions continuous on each edge and across vertices (Kovács et al., 2020).

On each edge, the evolution is a stochastic PDE with multiplicative Gaussian noise: $\frac{du}{dt}=-\Delta u-W'(u),$5 where the Allen–Cahn nonlinearity is classically $\frac{du}{dt}=-\Delta u-W'(u),$6. The noise is edge-wise, multiplicative, and driven by independent space-time Gaussian white noises $\frac{du}{dt}=-\Delta u-W'(u),$7 (Kovács et al., 2020).

The vertex conditions are essential. Continuity at each vertex is imposed by

$\frac{du}{dt}=-\Delta u-W'(u),$8

and a Kirchhoff-type law balances weighted fluxes: $\frac{du}{dt}=-\Delta u-W'(u),$9 The first condition enforces continuity across the whole network, while the second is a generalized Kirchhoff flux-conservation law with an additional interaction encoded by the symmetric, negative semi-definite, diagonally dominant matrix $\Delta$0 (Kovács et al., 2020).

The system is reformulated as the stochastic Cauchy problem

$\Delta$1

in the Banach space $\Delta$2. Under the stated assumptions, if $\Delta$3 for $\Delta$4, then there exists a unique global mild solution

$\Delta$5

and the solution has pathwise continuity on the graph. Under further assumptions, additional space-time regularity in fractional domain or interpolation spaces is proved (Kovács et al., 2020).

This network theory shows that “graph Allen–Cahn equation” may refer not only to ODEs on vertex sets, but also to edge-wise stochastic PDEs on metric graphs with transmission conditions. The analytical framework is correspondingly different: semigroup methods in $\Delta$6, analytic contractive semigroups, and stochastic evolution in UMD Banach spaces replace the finite-dimensional or variational methods typical of vertex-based models.

5. Interface limits, total variation, and curvature motion

One of the main conceptual themes in Allen–Cahn theory is the relation between diffuse interfaces and curvature-driven motion. In the continuum surface setting, formal inner–outer expansion for the Allen–Cahn equation on a deforming surface yields the sharp-interface law

$\Delta$7

where $\Delta$8 is the material normal velocity of the interface and $\Delta$9 is its geodesic curvature. The same work states that the continuous Allen–Cahn equation on a surface has direct analogues on graphs when the surface Laplacian is replaced by a graph Laplacian, and that analogous graph-curvature motion may be investigated through interface localization and thresholding ideas (Olshanskii et al., 2020).

On finite graphs, Budd and van Gennip establish a variational route toward that correspondence. They prove that the graph Ginzburg–Landau functional with double-obstacle potential $W$0-converges to graph total variation, and they translate to the graph setting two comparison principles used by Chen and Elliott in the continuum double-obstacle Allen–Cahn to mean-curvature-flow theory. They present results toward proving a link between double-obstacle graph Allen–Cahn flow and mean curvature flow on graphs, while also noting that a fully rigorous convergence result analogous to the continuum case is not yet fully proved (Budd et al., 2019).

A related geometric development appears in the degenerate area-preserving surface Allen–Cahn equation. There the sharp-interface limit is area-preserving geodesic curvature flow, and the numerical algorithm is implemented in a graph formulation of the surface together with adaptive finite elements and semi-implicit time discretization. Numerical solutions of the sharp-interface limit are also considered in a graph formulation as benchmark solutions (Benes et al., 2023).

Taken together, these results show that curvature motion enters graph Allen–Cahn theory in several layers. On finite combinatorial graphs, the strongest rigorous statements are presently variational and comparison-based. On graph-formulated surface problems, the curvature law is derived by matched asymptotics and used as a benchmark for numerics. The shared principle is that diffuse-interface energies converge toward interfacial objects of total-variation or curvature type, even though the exact form of “graph mean curvature flow” depends strongly on the underlying notion of graph.

6. Terminological scope and neighboring graph formulations

A recurrent source of confusion is that the word “graph” is used in several mathematically distinct senses. In one sense, it means a combinatorial graph with a vertex function and a graph Laplacian, as in finite-graph Allen–Cahn flow and hyperbolic-graph minimizers (Budd et al., 2019, Mramor, 2016). In a second sense, it means a metric graph or network, where one solves edge-wise PDEs coupled by continuity and Kirchhoff-type laws at vertices (Kovács et al., 2020). In a third sense, it refers to a surface represented as the graph of a height function over a planar domain, so that the Allen–Cahn equation is written in graph coordinates rather than on a discrete graph (Benes et al., 2023).

The surface-graph usage is technically important because it changes the PDE but not its ambient continuum nature. If

$W$1

then the Laplace–Beltrami operator becomes a variable-coefficient elliptic operator in the planar coordinates, and the degenerate area-preserving surface Allen–Cahn equation can be discretized on $W$2 by adaptive finite elements with a semi-implicit time scheme (Benes et al., 2023). This is a graph formulation in the sense of differential geometry, not a graph Laplacian model.

An additional neighboring usage appears in work on Allen–Cahn equations on catenoids. There the manifold

$W$3

is a catenoid expressed by graphs, and the equation

$W$4

admits ancient solutions with multiple transition layers and asymptotically Toda-type interaction of layer locations (Gkikas, 2024). This belongs to Allen–Cahn theory on graph-defined manifolds rather than to graph Laplacian dynamics.

Precision about this terminology is essential for reading the literature. The combinatorial-graph, metric-network, and graph-surface theories share phase-field language, double-well structure, heteroclinic profiles, and curvature heuristics, but they are analytically non-equivalent models with different operators, function spaces, and notions of interface.