McKean–Vlasov Mean-Field (MKV-MF)

• McKean–Vlasov mean-field is a large-population limit model where a representative particle’s dynamics depend on its own law or the conditional law in the presence of common noise.
• The framework is analyzed via PDEs, SPDEs, propagation of chaos, and gradient flows, offering insights into convergence, free-energy dissipation, and control variations.
• In neural network applications, MKV-MF recasts binary spiking neuron dynamics into a law-of-large-numbers PDMP, reducing computation from O(N²) to O(N) while capturing transient effects.

Searching arXiv for recent and foundational papers on McKean–Vlasov mean-field formulations, common noise, control, numerics, and the specific MKV-MF terminology in spiking networks. McKean–Vlasov mean-field denotes a class of large-population limits in which the dynamics of a representative component depend on its own law, or on a conditional law in the presence of common noise. In the classical setting, an interacting $N$-particle system converges, as $N\to\infty$, to a nonlinear stochastic differential equation or nonlinear Fokker–Planck equation whose coefficients depend on $\Law(X_t)$ or $\Law(X_t\mid \text{common noise})$ (Carmona et al., 2012). In a recent spiking-neural-network setting with stochastic STDP, the term “McKean–Vlasov mean-field (MKV-MF)” is used for the law-of-large-numbers limit of a network of “typical neurons,” yielding a McKean–Vlasov piecewise deterministic Markov process whose state includes an empirical distribution of incoming triplets $(V,S,W)$ (Helson et al., 2 Oct 2025). Across these settings, the subject comprises mean-field limits, propagation of chaos, nonlinear PDE and SPDE formulations, control and game-theoretic variants, thermodynamic and gradient-flow structures, long-time behavior, and numerical approximation.

1. Representative-particle formulation and mean-field limit

A basic McKean–Vlasov control model is defined by the state dynamics

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$

with cost functional

$J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$

(Carmona et al., 2012). This formulation isolates the representative agent while retaining dependence on the marginal law.

For interacting diffusions with common noise, one considers

$$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$

and the limit process satisfies

$dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$

(Mimikos-Stamatopoulos, 2023). Here the relevant law is conditional on the common-noise filtration rather than deterministic.

A non-Markovian variant with memory arises from interacting generalized Langevin systems. Under the quasi-Markovian assumption, the memory kernel is realized by finitely many auxiliary Ornstein–Uhlenbeck modes, and the $N$-particle dynamics converge to a McKean–Vlasov equation on an extended phase space $N\to\infty$0 (Duong et al., 2018). In that model the empirical measure

$N\to\infty$1

converges to a density $N\to\infty$2.

In the stochastic-STDP spiking-network setting, each neuron $N\to\infty$3 is represented by

$N\to\infty$4

where $N\to\infty$5 is the empirical distribution of incoming triplets $N\to\infty$6. The network is encoded by $N\to\infty$7, and as $N\to\infty$8 this empirical measure converges to a deterministic path $N\to\infty$9 in $\Law(X_t)$0 (Helson et al., 2 Oct 2025). The limiting typical neuron is therefore itself a McKean–Vlasov object whose state carries a law component.

A recurring consequence is propagation of chaos: finite subsystems become asymptotically independent and identically distributed in the appropriate conditional or unconditional sense (Kolokoltsov et al., 2015). This suggests that the “mean-field” terminology refers not merely to averaging, but to a precise asymptotic replacement of many-body interaction by self-consistent law dependence.

2. PDE, SPDE, and weak formulations

The mean-field limit typically admits a nonlinear forward equation. For the non-Markovian Langevin system with memory variables $\Law(X_t)$1, the generalized McKean–Vlasov equation is

$\Law(X_t)$2

where $\Law(X_t)$3 (Duong et al., 2018). The same paper gives the corresponding weak form against smooth test functions $\Law(X_t)$4.

With common noise, the law process becomes random and is described by a McKean–Vlasov SPDE. In weak form,

$\Law(X_t)$5

with

$\Law(X_t)$6

(Kolokoltsov et al., 2015). Under non-degeneracy $\Law(X_t)$7, the solution has an $\Law(X_t)$8-density $\Law(X_t)$9 satisfying

$\Law(X_t\mid \text{common noise})$0

The same common-noise framework leads, in the mean-field-game limit, to a quasi-linear infinite-dimensional backward PDE and then to a master equation involving variational derivatives $\Law(X_t\mid \text{common noise})$1 and mixed second-order derivatives $\Law(X_t\mid \text{common noise})$2 (Kolokoltsov et al., 2015). A related PDE viewpoint appears in the large-deviations analysis of empirical measures under common noise: the logarithmic transform of terminal probabilities yields an $\Law(X_t\mid \text{common noise})$3-player Hamilton–Jacobi–Bellman equation whose limit is a master or MFG-type PDE on $\Law(X_t\mid \text{common noise})$4 (Mimikos-Stamatopoulos, 2023).

In discrete state spaces $\Law(X_t\mid \text{common noise})$5, the nonlinear forward equation becomes a master equation

$\Law(X_t\mid \text{common noise})$6

which is shown to be a gradient flow of a free energy with respect to an explicit metric structure (Erbar et al., 2016). This places finite-state McKean–Vlasov equations in a variational framework parallel to continuous-space Wasserstein gradient flows.

A common misconception is that the McKean–Vlasov law is always deterministic. That is correct only in the absence of common noise; with Gaussian or Poissonian common noise, the limit is a random conditional law process $\Law(X_t\mid \text{common noise})$7 (Hernández-Hernández et al., 2023).

3. Structural frameworks: memory, free energy, and gradient flows

In the quasi-Markovian non-Markovian setting, the McKean–Vlasov equation can be embedded into the GENERIC structure by augmenting the state with an energy-reservoir variable $\Law(X_t\mid \text{common noise})$8. The state is $\Law(X_t\mid \text{common noise})$9, with energy and entropy functionals

$(V,S,W)$0

$(V,S,W)$1

and evolution

$(V,S,W)$2

with degeneracy conditions $(V,S,W)$3 and $(V,S,W)$4 (Duong et al., 2018). In this formulation, $(V,S,W)$5 and $(V,S,W)$6 are immediate.

For multi-species mean-field systems, under the symmetry $(V,S,W)$7, the free energy is

$(V,S,W)$8

and along solutions of the coupled McKean–Vlasov PDE system one has

$(V,S,W)$9

(Duong et al., 10 Jul 2025). The free energy therefore acts as a Lyapunov functional.

On finite spaces, the free energy

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$0

generates the gradient-flow equation

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$1

where $dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$2 is an explicit Onsager operator built from logarithmic means of detailed-balance fluxes (Erbar et al., 2016). The associated Benamou–Brenier-type distance yields a discrete $dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$3-Wasserstein geometry.

These results indicate that McKean–Vlasov dynamics are not solely probabilistic objects. They also admit thermodynamic, variational, and metric descriptions. A plausible implication is that the same law-dependent evolution can often be studied through several equivalent structures: SDE/SPDE, nonlinear Kolmogorov equation, gradient flow, or thermodynamic formalism.

4. Stationary states, phase transitions, and long-time behavior

For the extended non-Markovian McKean–Vlasov equation, any smooth stationary density has the form

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$4

and the number of stationary states is exactly the same as in the overdamped McKean–Vlasov model; in particular, the bifurcation diagram of the stationary problem is independent of the memory in the system (Duong et al., 2018). For Curie–Weiss interaction and a double-well confining potential, the same transition temperature and multiplicity of equilibria as in Dawson–Gärtner are recovered.

In the multi-species non-convex setting, stationary profiles satisfy

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$5

and one proves existence of at least one stationary solution, uniqueness in the high-noise regime, and symmetry-breaking solutions via bifurcation analysis under symmetry assumptions (Duong et al., 10 Jul 2025). In the prototypical case

$dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$6

there is a critical inverse-noise $dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$7. For $dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$8 there is a unique symmetric stationary profile, whereas for $dX_t \;=\; b\bigl(t,X_t,\Law(X_t),a_t\bigr)\,dt \;+\;\sigma\,dW_t,\qquad X_0=x_0,$9 there appear exactly three branches $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$0, two stable and one unstable (Duong et al., 10 Jul 2025).

Uniform exponential convergence can also be established in weak-interaction regimes. For mean-field interacting particle systems with confinement potential $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$1 and interaction potential $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$2, under a Dobrushin-type condition

$J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$3

one obtains

$J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$4

uniformly in $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$5, and for two solutions $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$6 of the nonlinear limit,

$J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$7

(Liu et al., 2020). The same framework yields uniform-in-time propagation of chaos and concentration inequalities with explicit constants.

Approximation of invariant measures can be performed through weighted empirical measures of the mean-field process itself or of self-interacting SDEs. Under a monotonicity condition,

$J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$8

and one derives rates for empirical approximation of $J(a)\;=\;\E\Bigl[\int_0^T f\bigl(t,X_t,\Law(X_t),\,a_t\bigr)\,dt \;+\; g\bigl(X_T,\Law(X_T)\bigr)\Bigr]$9 using frozen-law and self-interacting auxiliary processes (Du et al., 2021).

A frequent misconception is that non-convexity necessarily destroys analytically tractable long-time behavior. The cited results do not support that generalization: non-convex landscapes may still admit free-energy dissipation, bifurcation analysis, and convergence to stationary branches, though uniqueness may fail (Duong et al., 10 Jul 2025).

5. Control, mean field games, and backward systems

A central distinction in the subject is between optimal control of McKean–Vlasov dynamics and mean field games. In the mean-field-game path one fixes a deterministic flow of measures, solves the single-agent control problem, and then imposes the consistency condition $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$0. In the McKean–Vlasov-control path one takes the limit first and then optimizes over controls while internalizing the law dependence (Carmona et al., 2012). The resulting adjoint equations differ: in MKV control, the backward driver contains measure-derivative terms such as $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$1, $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$2, and $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$3 (Carmona et al., 2012).

With common noise and stochastic maximum principle, multidimensional MFGs lead to McKean–Vlasov FBSDEs of the form

$$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$4

with terminal condition $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$5 (Dianetti, 2022). Under submodularity-type structural assumptions, existence of strong solutions follows from Tarski’s fixed point theorem, and the set of solutions has a lattice structure with minimal and maximal solutions (Dianetti, 2022).

When diffusion coefficients are controlled, the absence of common noise allows characterization of the MFG equilibrium by a McKean–Vlasov second-order backward SDE. The driver is defined through a Hamiltonian supremum over relaxed drift and volatility controls, and the solution is formulated quasi-surely over a family of priors $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$6 (Barrasso et al., 2020). The paper emphasizes that diffusion control forces a genuine 2BSDE rather than an ordinary BSDE representation.

Conditional McKean–Vlasov control introduces dependence on the conditional law given survival in a domain $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$7. The controlled state solves a weak SDE up to the exit time $$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$8, and optimality is characterized by a new Pontryagin maximum principle connecting the problem to a generalized McKean–Vlasov BSDE (Carmona et al., 8 Oct 2025). Applications include a Schrödinger bridge with hard killing and the construction of equilibria in potential mean field games via McKean–Vlasov control (Carmona et al., 8 Oct 2025).

These distinctions matter conceptually. Mean field games and McKean–Vlasov control are related but not interchangeable, because the order of optimization and passage to the limit changes the adjoint structure and therefore the equilibrium equations (Carmona et al., 2012).

6. Computation, inference, and the MKV-MF neural-network instantiation

Simulation of McKean–Vlasov processes can be accelerated by multilevel Monte Carlo. For SDEs of the form

$$dX_t^{i,N} \;=\; b\bigl(X_t^{i,N},\,m^N_{t}\bigr)\,dt \;+\; dW^i_t \;+\; \sqrt{\alpha_0}\,dW^0_t, \qquad m^N_{t}=\frac1N\sum_{j=1}^N\delta_{X_t^{j,N}},$$9

a multilevel estimator uses coarse-level mean-field estimates to inform fine-level computations. In the analyzed linear case, the method achieves $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$0 error of size $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$1 with $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$2 complexity, compared with $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$3 for standard methods (Ricketson, 2015).

For density reconstruction, a stochastic particle-Euler method combined with a kernel estimator and Goldenshluger–Lepski bandwidth selection provides a nonparametric numerical approach to the nonlinear Fokker–Planck solution. The estimator

$dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$4

satisfies deviation and $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$5-error bounds, and the data-driven selector yields an oracle inequality up to a $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$6 factor (Hoffmann et al., 2023). Related nonparametric inference results establish adaptive estimators and minimax optimality for the density and drift in large-population McKean–Vlasov models (Maestra et al., 2020).

Parametric inference in the mean-field regime has also been developed through local asymptotic normality. For a $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$7-dimensional parameter $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$8 in the drift, one has $dX_t \;=\; b\bigl(X_t,\Law\,(X_t\mid W^0)\bigr)\,dt \;+\; dW_t \;+\; \sqrt{\alpha_0}\,dW^0_t$9 and LAN expansions of the log-likelihood ratio, leading to asymptotic normality and Hájek-optimality of the MLE under identifiability and non-degeneracy assumptions (Maestra et al., 2022).

Neural methods now target the fixed-point structure directly. NF-MKV Net reformulates MFG equilibria as MKV FBSDEs, couples a value network to a process-regularized normalizing flow for the law $N$0, and enforces volumetric invariance and temporal continuity through Jacobian constraints on density transfer maps (Liu et al., 29 Jan 2025). For stationary McKean–Vlasov Fokker–Planck equations with quadratic interaction, WANPM reduces the mean-field nonlinearity to a batch sample mean and reports that gradient flow through the self-consistent mean estimate is essential for uniqueness, while sufficiently large adversarial test-function frequencies are needed to avoid spurious minimizers (He et al., 17 Mar 2026).

The most explicit use of the label MKV-MF occurs in stochastic-STDP spiking networks. There, a network of $N$1 all-to-all connected binary neurons with synaptic weights $N$2 is recast in terms of a “typical neuron” with state

$N$3

where $N$4, $N$5, and

$N$6

The large-$N$7 limit is a PDMP of McKean–Vlasov type, and the authors term this analysis McKean–Vlasov mean-field (MKV-MF) (Helson et al., 2 Oct 2025). In that limit, the synaptic current is

$N$8

the membrane state jumps $N$9 at rate $N\to\infty$00, and the law evolution is written in weak form for Fréchet-differentiable test functions $N\to\infty$01 on the state space (Helson et al., 2 Oct 2025). The model is time-dependent, assumes no time-scale separation, and is described as an exact law-of-large-numbers description of the original $N\to\infty$02-particle PDMP that captures transient and heterogeneous effects at cost $N\to\infty$03 instead of $N\to\infty$04 (Helson et al., 2 Oct 2025).

This suggests two concurrent uses of “MKV-MF”: a broad one, referring to McKean–Vlasov mean-field theory in general, and a narrower one, used explicitly in recent neural-network work for a typical-neuron McKean–Vlasov PDMP representation (Helson et al., 2 Oct 2025).