Bipolar Aggregation-Diffusion Systems

• Bipolar aggregation–diffusion systems are two-species models where each density interacts through distinct self and cross potentials, driving complex behaviors.
• They are formulated via nonlocal mean-field models, local fourth-order limits, and bipolar drift–diffusion equations to capture adhesion, segregation, and dynamic patterns.
• Analytical frameworks, including gradient flows and bifurcation theory, provide insights into phase separation, cell sorting, and stability properties in these systems.

Searching arXiv for the cited works to ground the article in current literature. arXiv search query: "A nonlocal-to-local approach to aggregation-diffusion equations" A bipolar aggregation–diffusion system is, in the contemporary PDE literature, a two-species aggregation–diffusion model in which two densities evolve under distinct self-interactions and cross-interactions, with diffusion or higher-order regularization acting simultaneously. In this sense, “bipolar” refers to the presence of two coupled components, often denoted $\rho$ and $\eta$, whose interactions may favor mixing, segregation, engulfment, or other patterned states. Recent work places such systems within a common framework spanning nonlocal mean-field models, local fourth-order short-range limits of Cahn–Hilliard or thin-film type, and bipolar drift–diffusion equations with Poisson or Riesz interactions (Falcó et al., 13 May 2025, Alves et al., 2020, Alves et al., 6 Sep 2025).

1. Conceptual scope and terminology

In the language of "A nonlocal-to-local approach to aggregation-diffusion equations" (Falcó et al., 13 May 2025), a bipolar aggregation–diffusion system is naturally interpreted as a two-species aggregation–diffusion model with possibly different self-interactions and cross-interactions. The two “poles” are the two cell types or densities, and the bipolar structure is encoded by distinct self-interaction parameters together with cross-couplings that can represent attraction, repulsion, and interfacial tension.

This two-species interpretation is not the only one in the literature. In "Lattice and Continuum Models Analysis of the Aggregation Diffusion Cell Movement" (Bao et al., 2018), the bipolar character appears instead as a sharp change of sign in an effective diffusion coefficient, separating an aggregation regime from a diffusion regime in a scalar model. In "Learning Representation over Dynamic Graph using Aggregation-Diffusion Mechanism" (Liu et al., 2021), the phrase denotes a two-phase mechanism on continuous-time dynamic graphs: aggregation as a pull update and diffusion as a push update. This suggests a terminological family rather than a single canonical definition, but the dominant usage in mathematical biology and PDE analysis is the two-species continuum one.

Within that dominant usage, bipolar aggregation–diffusion systems usually combine three ingredients: a continuity-equation structure, nonlinear mobilities such as $u_i \nabla \mu_i$, and a self-consistent potential or free-energy formulation. The resulting models belong simultaneously to the theory of aggregation equations, cross-diffusion systems, Wasserstein gradient flows, and phase-separation models.

2. Canonical continuum formulations

A standard nonlocal two-species model for densities $\rho(x,t)$ and $\eta(x,t)$ on $\Omega\subset\mathbb{R}^d$ is

$$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$

where $W_{11},W_{22}$ are self-adhesion potentials, $W_{12},W_{21}$ are cross-adhesion potentials, $*$ denotes convolution, and $\eta$0 is the strength of local repulsion or volume exclusion (Falcó et al., 13 May 2025). A common differential-adhesion parametrization is $\eta$1 with $\eta$2 purely attractive, $\eta$3, and $\eta$4.

The same paper derives a local fourth-order limit of thin-film or Cahn–Hilliard type: $\eta$5 Here $\eta$6 is the relative self-“surface tension” of $\eta$7, $\eta$8 is a cross-gradient coupling, $\eta$9 is a self-aggregation pressure for $u_i \nabla \mu_i$0, and $u_i \nabla \mu_i$1 is a cross-aggregation pressure whose sign controls attraction versus repulsion (Falcó et al., 13 May 2025).

A second major family is the bipolar drift–diffusion system

$u_i \nabla \mu_i$2

with no-flux boundary conditions and $u_i \nabla \mu_i$3. Here $u_i \nabla \mu_i$4 and $u_i \nabla \mu_i$5 are densities of positively and negatively charged carriers, $u_i \nabla \mu_i$6 is the electrostatic potential, and $u_i \nabla \mu_i$7 are monotone increasing pressure functions (Alves et al., 2020). Because $u_i \nabla \mu_i$8 is generated by $u_i \nabla \mu_i$9, the system is a bipolar aggregation–diffusion model with Coulomb-type interaction.

The Riesz generalization replaces Poisson interactions by the kernel

$\rho(x,t)$0

and yields the limiting aggregation–diffusion system

$\rho(x,t)$1

in which the drift is generated by the charge density $\rho(x,t)$2 and the two species feel the field with opposite signs (Alves et al., 6 Sep 2025).

Other two-species variants emphasize different interaction symmetries. The predator–prey system

$\rho(x,t)$3

combines quadratic diffusion, self-attraction, predator attraction to prey, and prey repulsion from predators; the cross-potentials are mutually proportional by a negative constant $\rho(x,t)$4 (Fagioli et al., 2019). On the torus, a symmetric nonlocal model with linear diffusion,

$\rho(x,t)$5

is used for rigorous bifurcation analysis and for applications to cell–cell adhesion (Carrillo et al., 4 Jul 2025).

3. Nonlocal-to-local reduction and energetic structure

The nonlocal-to-local derivation in (Falcó et al., 13 May 2025) starts from the short-range scaling

$\rho(x,t)$6

with $\rho(x,t)$7 small, radial kernels $\rho(x,t)$8, and suitable moment assumptions. A Taylor expansion gives

$\rho(x,t)$9

where

$\eta(x,t)$0

After inserting these terms into the two-species nonlocal system and rescaling variables, one obtains the fourth-order local model. The fourth-order couplings can be organized through the surface-tension matrix

$\eta(x,t)$1

which must be positive definite, yielding

$\eta(x,t)$2

This is the local analogue of a positive-definite interfacial tensor in multicomponent Cahn–Hilliard systems (Falcó et al., 13 May 2025).

The same local model is a Wasserstein gradient flow of

$\eta(x,t)$3

with chemical potentials given by the variational derivatives and the energy dissipation identity

$\eta(x,t)$4

under the natural boundary conditions

$\eta(x,t)$5

(Falcó et al., 13 May 2025).

Bipolar drift–diffusion systems admit analogous energetic formulations. For the Poisson-coupled model,

$\eta(x,t)$6

and the formal gradient-flow representation is

$\eta(x,t)$7

(Alves et al., 2020). The Euler–Riesz limit has the corresponding interaction energy

$\eta(x,t)$8

and the paper notes that it is natural to view the limiting system as a gradient flow in product Wasserstein space (Alves et al., 6 Sep 2025).

4. Pattern formation, segregation, and stationary states

The local fourth-order two-species model in (Falcó et al., 13 May 2025) was built to reproduce Steinberg’s differential adhesion picture of cell sorting. Four canonical morphologies are identified: sorting (full segregation), partial engulfment, engulfment, and mixing. Under the parametrization $\eta(x,t)$9 with $\Omega\subset\mathbb{R}^d$0, weak cross-adhesion $\Omega\subset\mathbb{R}^d$1 corresponds to $\Omega\subset\mathbb{R}^d$2, while strong cross-adhesion $\Omega\subset\mathbb{R}^d$3 corresponds to $\Omega\subset\mathbb{R}^d$4. The limiting regimes are described as follows: $\Omega\subset\mathbb{R}^d$5 yields sorting, $\Omega\subset\mathbb{R}^d$6 yields partial engulfment, $\Omega\subset\mathbb{R}^d$7 yields engulfment, and $\Omega\subset\mathbb{R}^d$8 yields mixing (Falcó et al., 13 May 2025).

These morphologies are already encoded at the level of the free energy. When $\Omega\subset\mathbb{R}^d$9, the cross-bulk term penalizes overlap and favors segregation; when $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$0, overlap becomes energetically favorable; and $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$1 contributes a cross-gradient interfacial energy that sets the effective surface tension between phases (Falcó et al., 13 May 2025). A plausible implication is that the sign of the bulk coupling and the definiteness of the gradient part play roles analogous to bulk and surface terms in classical phase-separation theory.

The predator–prey model in (Fagioli et al., 2019) exhibits a different bipolar mechanism: self-attraction within each species, but opposite-signed cross-interaction between predators and prey. Its stationary states can consist of multiple bumps, each of which is compactly supported, symmetric around a prescribed center, and close in the small-diffusion regime to a Barenblatt-type profile

$$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$2

with an analogous expression for $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$3. The existence theorem constructs such multi-bump states from equilibria of the underlying particle system by means of the functional Implicit Function Theorem (Fagioli et al., 2019).

On the torus, "Long-time behaviour and bifurcation analysis of a two-species aggregation-diffusion system on the torus" (Carrillo et al., 4 Jul 2025) gives a fully rigorous classification of branches bifurcating from the homogeneous state $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$4. The phase ratio on a bifurcating branch determines whether the two species are in phase, with peaks at the same locations, or out of phase, with peaks of one coinciding with troughs of the other. An explicit application to cell–cell adhesion shows that stable segregation patterns can arise at the onset of cell sorting even when all interactions are purely attractive (Carrillo et al., 4 Jul 2025). This directly counters the common misconception that purely attractive interactions necessarily imply complete mixing.

5. Analytical results, relaxation limits, and singular regimes

For the local fourth-order cell-adhesion system, the modeling paper summarizes that global weak solutions are known under suitable structural conditions, including positive definiteness of the surface-tension matrix and subcritical parameter regimes. It also records several open problems: uniqueness of weak solutions is not known, existence of energy minimizers in the whole space is open because of potential mass loss at infinity, and characterization of steady states and their stability remains analytically difficult (Falcó et al., 13 May 2025).

The Poisson-coupled bipolar fluid model provides a different analytical route. The scaled bipolar Euler–Poisson system with friction,

$$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$5

converges in the high-friction regime to the bipolar drift–diffusion system. The proof is based on a relative energy functional

$$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$6

and the estimate

$$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$7

which gives weak–strong stability and convergence on finite time intervals (Alves et al., 2020).

The Riesz generalization establishes the analogous high-friction limit for a bipolar Euler–Riesz system with nonlocal interaction kernel $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$8. Its main theorem states that, under the stated conditions on $$\begin{aligned} \frac{\partial \rho}{\partial t} &= \nabla\cdot\left(\rho\nabla \left(W_{11}*\rho + W_{12}*\eta + \epsilon(\rho + \eta) \right)\right),\ \frac{\partial \eta}{\partial t} &= \nabla\cdot\left(\eta\nabla \left(W_{21}*\rho + W_{22}*\eta + \epsilon(\rho + \eta) \right)\right), \end{aligned}$$9, $W_{11},W_{22}$0, and sufficiently small interaction strength $W_{11},W_{22}$1, the relative energy $W_{11},W_{22}$2 satisfies

$W_{11},W_{22}$3

so well-prepared dissipative weak solutions converge to the strong solution of the bipolar aggregation–diffusion limit (Alves et al., 6 Sep 2025). This extends the Poisson case from Newtonian to general Riesz interactions.

A different singular regime appears in the lattice-to-continuum scalar model of (Bao et al., 2018). The continuum equation

$W_{11},W_{22}$4

has $W_{11},W_{22}$5 in an aggregation region $W_{11},W_{22}$6 and $W_{11},W_{22}$7 in a diffusion region $W_{11},W_{22}$8. For initial data entirely in the diffusion region, the model is globally well posed and converges exponentially to the spatial average; for initial data entirely in the aggregation region, the paper proves nonexistence of weak solutions on any $W_{11},W_{22}$9 (Bao et al., 2018). This sharp contrast highlights how bipolar aggregation–diffusion behavior can reflect either two interacting species or two incompatible parabolic regimes.

6. Related formulations and cross-disciplinary extensions

The two-species torus model of (Carrillo et al., 4 Jul 2025) makes the spectral structure of bipolar aggregation–diffusion especially explicit. For Fourier mode $W_{12},W_{21}$0, the linearized eigenvalues are

$W_{12},W_{21}$1

and bifurcation occurs when $W_{12},W_{21}$2. The paper then applies Crandall–Rabinowitz theory to classify all branches emerging from the homogeneous state, including pitchfork bifurcations with explicit formulas for branch curvature and stability exchange. This places bipolar aggregation–diffusion squarely within rigorous nonlocal bifurcation theory.

The scalar sign-changing model of (Bao et al., 2018) shows that “bipolar” can also describe a one-species system whose effective transport alternates between forward and backward diffusion. At the discrete level, the lattice dynamics remain bounded for all time, the forward or diffusion region

$W_{12},W_{21}$3

is monotonically expanding, and in the pure diffusion regime the discrete solution converges to the average density. The continuum and lattice descriptions therefore diverge most sharply on the aggregation side.

Outside PDE theory, the phrase is used in graph representation learning for a two-phase update cycle on continuous-time dynamic graphs. There the aggregation phase updates interacting nodes by pulling information from neighbors, while the diffusion phase updates neighboring nodes by pushing messages outward from the interacting nodes: $W_{12},W_{21}$4 This mechanism was inserted into DyRep and LDG and shown to improve dynamic link prediction on Social Evolution and Github, but it is conceptually distinct from PDE-based aggregation–diffusion systems (Liu et al., 2021).

Across these settings, the unifying idea is the competition or coupling between concentration-inducing and spreading-inducing mechanisms. In the PDE literature, that competition is encoded by nonlocal interaction energies, pressure laws, and gradient-flow dissipation; in graph learning, by alternating pull and push operators. The mathematically central notion, however, remains the two-species continuum system with self- and cross-interactions, whose current theory connects nonlocal models, local fourth-order limits, drift–diffusion equations, bifurcation analysis, and cell-sorting phenomena (Falcó et al., 13 May 2025).