McKean-Vlasov Fokker-Planck Equations

• McKean-Vlasov Fokker-Planck equations are nonlinear forward equations that characterize the evolution of probability laws in systems with mean-field interactions.
• They link probabilistic representations from McKean-Vlasov SDEs to nonlinear, measure-valued PDEs and SPDEs through superposition principles.
• Functional-analytic frameworks and numerical methods ensure well-posedness and uniqueness even under complex feedback and stochastic noise conditions.

McKean-Vlasov Fokker-Planck equations are nonlinear forward equations for the evolution of probability laws of stochastic systems whose coefficients depend on the current law, the conditional law, or a density evaluated at the current state. In the basic diffusion setting, they are the Fokker-Planck equations associated with McKean-Vlasov stochastic differential equations (SDEs); in broader settings they appear as measure-valued partial differential equations, stochastic partial differential equations (SPDEs), integro-differential equations with jumps, or Fokker-Planck equations posed on spaces of probability measures. The subject links nonlinear PDE, stochastic analysis, martingale problems, optimal transport, nonlinear semigroup theory, filtering, mean-field limits, and control. A recurrent theme is the equivalence—often through superposition principles—between a nonlinear PDE or SPDE for laws and a probabilistic representation by a McKean-Vlasov process (Ren et al., 2019).

1. Core definition and canonical forms

The standard McKean-Vlasov SDE on $\mathbb{R}^d$ has the form

$$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$

and its time marginals $\mu_t$ solve a nonlinear Fokker-Planck equation

$$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$

with $a=\sigma\sigma^\top$, in weak form or in distributional form depending on regularity (Ren, 2021). In this sense, the nonlinearity is not only through the unknown density but through the law itself, which may enter coefficients globally, through convolution kernels, or pointwise in density.

A widely studied subclass is the nonlinear Fokker-Planck equation

$$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$

or, in Nemytskii-type form,

$$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$

Here the drift and diffusion coefficients depend on $u(t,x)$ pointwise, so that the corresponding SDE coefficients are singular functionals of the time marginal density rather than smooth functionals of the law (Barbu et al., 2019, Grube, 2024).

Another central class is the aggregation-diffusion or granular-media equation,

$$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$

which is the mean-field Fokker-Planck equation associated with interaction drift

$$b(x,\rho)= -\nabla V(x)-\int_{\mathbb{R}^n}\nabla W(x-y)\rho(y)\,dy.$$

For the quadratic interaction kernel $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$0, the nonlinearity reduces to dependence on the mean $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$1, and the stationary equation admits a self-consistent Boltzmann form (He et al., 17 Mar 2026).

The literature therefore uses a single label—McKean-Vlasov Fokker-Planck equation—for several related but distinct objects: deterministic PDEs on $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$2, stochastic Fokker-Planck equations for conditional laws, and lifted Fokker-Planck equations on measure spaces (Lacker et al., 2020, Feng et al., 18 Jun 2025). A common misconception is that the forward equation is always a classical deterministic PDE on Euclidean space. The conditional and measure-space formulations show that this is false in the presence of common noise, filtering structure, or law-valued state variables.

2. Probabilistic representations and superposition principles

The most basic link between McKean-Vlasov SDEs and their Fokker-Planck equations is furnished by Itô’s formula: any sufficiently regular weak solution of the SDE yields a weak solution of the corresponding forward equation for its marginals (Barbu et al., 2019). The inverse direction is subtler. It requires a superposition principle asserting that a sufficiently integrable solution of the nonlinear Fokker-Planck equation can be realized as the marginal law flow of a weak solution of a McKean-Vlasov SDE. This principle is central in several modern treatments (Qiao, 2020, Barbu et al., 2022).

A systematic linearization viewpoint lifts the nonlinear equation on $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$3 to a linear Fokker-Planck equation on $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$4, where $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$5 denotes a space of probability measures. In that framework, the lifted generator $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$6 acts on test functions on the product of state space and measure space, using the intrinsic gradient $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$7 on the tangent bundle over $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$8. The diffusion generated by $$dX_t=b\bigl(t,X_t,\mu_t\bigr)\,dt+\sigma\bigl(t,X_t,\mu_t\bigr)\,dW_t, \qquad \mu_t=\mathrm{Law}(X_t),$$9 is intrinsically related to the McKean-Vlasov SDE, and restricted well-posedness of the nonlinear equation together with its linearized version implies restricted well-posedness of the McKean-Vlasov equation and a Markov property for the laws of solutions (Ren et al., 2019). This formulation is particularly useful when the dependence on the law is merely measurable or of Nemytskii type.

With common noise, the appropriate object is no longer the unconditional law $\mu_t$0 but the conditional law $\mu_t$1. In that setting the conditional time marginals satisfy a nonlinear second-order SPDE, while the laws of the conditional time marginals satisfy a Fokker-Planck equation on the space of probability measures (Lacker et al., 2020). Two superposition principles are available there: any solution of the SPDE can be lifted to a solution of the conditional McKean-Vlasov SDE, and any solution of the Fokker-Planck equation on $\mu_t$2 can be lifted to a solution of the SPDE (Lacker et al., 2020). The 2025 extension to general conditional McKean-Vlasov systems makes this three-level correspondence explicit between the conditional McKean-Vlasov SDE, a nonlinear Zakai equation, and an infinite-dimensional conditional Fokker-Planck equation, so that weak well-posedness of one level implies weak well-posedness of the others (Feng et al., 18 Jun 2025).

This hierarchy clarifies a second common misconception: the nonlinear forward equation is not merely a byproduct of the SDE. In several works it is treated as an autonomous analytical object whose well-posedness can imply existence and uniqueness for the stochastic system itself (Barbu et al., 2022, Feng et al., 18 Jun 2025).

3. Well-posedness, uniqueness, and functional-analytic frameworks

A major branch of the theory studies deterministic nonlinear Fokker-Planck equations by nonlinear semigroup methods. For

$\mu_t$3

on $\mu_t$4, suitable assumptions on $\mu_t$5, $\mu_t$6, and $\mu_t$7 imply that the equation generates a unique flow $\mu_t$8 in $\mu_t$9 as a mild solution in the sense of nonlinear semigroup theory; the same flow is unique in a class of $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$0 distributional solutions, and right-differentiability in the $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$1 norm can be established under additional coercivity (Barbu et al., 2022). The same semigroup viewpoint yields existence and uniqueness for initial data in $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$2, preservation of nonnegativity and mass, and extension to bounded measures under additional assumptions such as $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$3 (Barbu et al., 2020).

In the nondegenerate monotone-diffusion case, uniqueness of distributional solutions can be proved by an $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$4 energy method. For

$$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$5

if $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$6 is monotone and uniformly coercive and $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$7 is sufficiently regular, any two bounded distributional solutions with the same initial data coincide (Barbu et al., 2019). The key estimate controls the difference in $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$8 using monotonicity of $$\partial_t\rho_t = -\nabla\!\cdot\!\bigl(b(t,\cdot,\rho_t)\rho_t\bigr) +\tfrac12\sum_{i,j}\partial_{x_i}\partial_{x_j}\bigl(a_{ij}(t,\cdot,\rho_t)\rho_t\bigr),$$9 and a drift estimate that is absorbed by the coercive diffusion term. This uniqueness transfers to weak uniqueness in law for the associated McKean-Vlasov SDE (Barbu et al., 2019).

The degenerate case requires different balance conditions. In one formulation, $a=\sigma\sigma^\top$0 may vanish on sets of positive measure, but monotonicity is retained and the growth of $a=\sigma\sigma^\top$1 is tied to $a=\sigma\sigma^\top$2; under these assumptions the operator driving the nonlinear Fokker-Planck equation is $a=\sigma\sigma^\top$3-accretive in $a=\sigma\sigma^\top$4, uniqueness holds among broad classes of distributional solutions, and weak uniqueness follows for the corresponding McKean-Vlasov SDE (Barbu et al., 2022). In another formulation with Nemytskii-type coefficients, a unique strong solution to the associated degenerate McKean-Vlasov SDE can be constructed once the PDE solution $a=\sigma\sigma^\top$5 is fixed, using pathwise uniqueness for the fixed-law SDE and a Yamada-Watanabe argument (Grube, 2024).

More singular regimes are also accessible. Nonlinear Fokker-Planck equations with singular integral drifts can be treated in $a=\sigma\sigma^\top$6 if the diffusion is uniformly elliptic, the singular kernel satisfies stated integrability assumptions, and the operator is shown to be quasi-$a=\sigma\sigma^\top$7-accretive (Barbu, 2024). Critical Lorentz kernels yield another endpoint theory: for divergence-free $a=\sigma\sigma^\top$8, the equation

$a=\sigma\sigma^\top$9

admits narrowly continuous weak solutions, and uniqueness can hold in a Krylov class, with optimality in dimensions $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$0 demonstrated by nonuniqueness in the supercritical regime (Röckner et al., 20 May 2025). A plausible implication is that the forward equation can remain well posed at scaling-critical regularity even when direct SDE methods are delicate.

4. Conditional laws, stochastic forward equations, and nonclassical noises

When common noise is present, the forward equation becomes stochastic. For conditional McKean-Vlasov diffusions with jumps,

$$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$1

with $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$2, the conditional law satisfies the stochastic Fokker-Planck PIDE

$$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$3

or, in dual form,

$$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$4

for $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$5 (Agram et al., 2021). If $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$6 is absolutely continuous, the measure-valued equation reduces to an SPDE for the density $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$7 with transport, diffusion, jump, and common-noise terms (Agram et al., 2021).

Brownian sheets produce a qualitatively different forward equation. For the hyperbolic McKean-Vlasov SPDE

$$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$8

with $$\partial_tu-\Delta B(u)+\nabla\!\cdot\!\bigl(b^*(x,u)\bigr)=0, \qquad b^*(x,r)=b(x,r)\,r,$$9, the law process takes values in a weighted Sobolev-Fourier space of measures, and the resulting forward equation contains not only first- and second-order terms but two-parameter correction terms of order up to four arising from the two-parameter Itô formula (Agram et al., 2024). Even in the constant-coefficient case the differential equation for $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$0 contains a fourth derivative term proportional to $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$1 (Agram et al., 2024). This shows that “Fokker-Planck” need not mean a second-order parabolic PDE in the classical sense.

Fractional Brownian motion yields another non-Markovian variant. For

$$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$2

the measure-valued forward equation becomes

$$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$3

and, in the absolutely continuous case, the density solves a classical-looking PDE with a time-dependent diffusion coefficient induced by the fractional-noise operator $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$4 (Labed et al., 2024). For $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$5 the Brownian coefficient is recovered; for $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$6 the diffusion term becomes time dependent and encodes long memory (Labed et al., 2024).

These examples show that the forward equation adapts to the stochastic calculus underlying the particle system. Brownian motion, common noise, jumps, Brownian sheets, and fractional Brownian motion all generate distinct forward structures rather than minor variations of a single PDE.

5. Singular phenomena, stationary states, and long-time behavior

McKean-Vlasov Fokker-Planck equations need not be globally regular. In hitting-time models, nonlocal feedback through absorption can produce blow-up, and avoiding blow-up becomes a central analytical question. For

$$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$7

the density $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$8 on $$\partial_t u = -\nabla_x\!\cdot\!\bigl[E(x)\,b(u)\,u\bigr] +\Delta_x\!\bigl[\beta(u)\bigr].$$9 satisfies

$u(t,x)$0

and a Stefan transform converts the problem into a one-phase supercooled Stefan problem (Bayraktar et al., 2020). Comparison principles then yield a no-blow-up criterion: if

$u(t,x)$1

the weak solution exists globally and $u(t,x)$2 satisfies an explicit bound (Bayraktar et al., 2020). For a logarithmic feedback model, a different transform removes the nonlocal term and a relative-entropy estimate gives global existence when $u(t,x)$3 is sufficiently large and $u(t,x)$4 is sufficiently small (Bayraktar et al., 2020). A common misconception is that mean-field diffusion automatically regularizes the dynamics; these examples show that nonlinear feedback through hitting probabilities can instead create singular behavior.

At the opposite end of the spectrum, stationary solutions and phase transitions are classical themes. For the Vlasov-Fokker-Planck equation

$u(t,x)$5

stationary solutions have the Gibbs form

$u(t,x)$6

and correspond to critical points of a free-energy functional (Duong et al., 2015). In one dimension with $u(t,x)$7, there exists a critical value $u(t,x)$8 such that above $u(t,x)$9 the stationary solution is unique and symmetric, while below $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$0 there are exactly three stationary solutions: one symmetric and two asymmetric (Duong et al., 2015). This is a precise phase-transition statement within the McKean-Vlasov Fokker-Planck framework.

Ergodic behavior can also be formulated at the lifted level. In the time-homogeneous setting of the linearization on $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$1, ergodicity of the diffusion generated by the lifted operator can be characterized by asymptotic properties of the coupled nonlinear Fokker-Planck equation (Ren et al., 2019). For coupled McKean-Vlasov equations with jumps, superposition principles, construction of associated space-distribution valued Markov processes, and ergodicity are studied together, with exponential ergodicity obtained as a by-product for a class of McKean-Vlasov SDEs with jumps (Qiao, 2020).

6. Structural properties, applications, and numerical methods

The forward equation inherits order-theoretic properties from the underlying McKean-Vlasov dynamics. Order preservation and positive correlation for nonlinear Fokker-Planck equations can be characterized through the coefficients of the associated McKean-Vlasov SDE. Under explicit monotonicity and structural conditions on the drift $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$2 and diffusion matrix $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$3, the path-law and one-time marginals preserve stochastic order, and FKG-type positive correlation can be propagated by the nonlinear semigroup (Ren, 2021). This recovers linear diffusion criteria in the law-independent case and extends them to the mean-field setting.

Control theory is a major application. For conditional McKean-Vlasov jump diffusions, the stochastic Fokker-Planck PIDE for the conditional law makes the pair $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$4 Markovian, which in turn allows formulation of an infinite-dimensional Hamilton-Jacobi-Bellman equation involving derivatives with respect to the measure variable (Agram et al., 2021). The same work applies the framework to linear-quadratic optimal control and optimal consumption problems (Agram et al., 2021). In the conditional setting without jumps, mimicking theorems show that the conditional time marginals of an Itô process can be emulated by a conditional McKean-Vlasov SDE with Markovian coefficients, providing a route from open-loop to Markovian controls (Lacker et al., 2020).

Hydrodynamic and kinetic limits supply another source of applications. The aggregation-diffusion equation

$$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$5

appears as the overdamped limit of kinetic Vlasov-Fokker-Planck equations with singular interactions. Introducing a coarse-graining map and an intermediate system, one obtains explicit Wasserstein bounds between the kinetic spatial marginal and the limiting McKean-Vlasov equation, with an error of order $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$6 in the scaling parameter $$\partial_t \rho = \nabla\!\cdot[(\nabla V)\rho] + \nabla\!\cdot[(\nabla W*\rho)\rho] + (\sigma^2/2)\Delta \rho,$$7 (Choi et al., 2020). This places the nonlinear Fokker-Planck equation as the effective diffusive law of a higher-dimensional kinetic system.

Recent numerical work treats the stationary McKean-Vlasov mean-field Fokker-Planck equation by the Weak Adversarial Neural Pushforward Method. For the quadratic interaction kernel the mean-field nonlinearity reduces to the batch sample mean, the stationary weak form is enforced through adversarial plane-wave test functions, and a one-dimensional linear McKean-Vlasov benchmark recovers the exact Gaussian stationary distribution with small absolute errors after training (He et al., 17 Mar 2026). Two practical points are emphasized there: gradient flow through the self-consistent mean estimate is essential for uniqueness, and sufficiently large initialization of adversarial frequencies is needed to avoid spurious minimizers (He et al., 17 Mar 2026). This suggests that even in apparently simple mean-field Fokker-Planck problems, self-consistency constraints remain numerically delicate.

Across these directions, McKean-Vlasov Fokker-Planck equations function as the forward-law description of mean-field stochastic dynamics, but the term encompasses a broad family of analytical objects. Depending on the noise, interaction, and observation structure, the forward equation may be deterministic or stochastic, local or nonlocal, Euclidean or measure-valued, second order or higher order, and may exhibit uniqueness, degeneracy, criticality, blow-up avoidance, phase transition, or ergodicity within a unified mean-field framework (Ren et al., 2019, Feng et al., 18 Jun 2025).