Lévy-Type Mean-Field Systems

• Lévy-type mean-field systems are models where nonlocal jump dynamics and heavy-tailed Lévy statistics are integrated with population-dependent drift and diffusion.
• They employ generator-level formulations, semigroup techniques, and Krylov estimates to derive weak existence, stability, and propagation-of-chaos results.
• These systems extend into optimization and mean-field games, demonstrating explicit feedback synthesis, control with jumps, and effective continuous-time random walk approximations.

Lévy-type mean-field systems are mean-field models in which collective interaction is coupled to Lévy-type nonlocality, either because the representative dynamics is governed by a Lévy–Khintchine generator depending on the population law, because the state equation is driven by Brownian motion plus a Poisson random measure or by additive Lévy noise, or because a deterministic long-range mean-field system displays effective truncated Lévy statistics in suitably coarse-grained variables [(Kolokoltsov et al., 2011); (Tang et al., 18 Mar 2025); (Lim et al., 4 Aug 2025); (Ye, 14 Apr 2026); (Figueiredo et al., 2012)]. Across these formulations, the mean field may enter the drift, diffusion matrix, jump kernel, cost functional, or equilibrium consistency relation, while the Lévy-type structure appears as a nonlocal generator, a compensated Poisson random measure, a general Lévy process, or a heavy-tailed continuous-time random walk description. The surrounding literature also contains adjacent frameworks—switching diffusions, adaptive networks with memory, hypergraph Vlasov limits, and Tanaka-style non-exchangeable additive-noise systems—that are structurally relevant but not, as stated, genuine Lévy-type models (Nguyen et al., 2019, Throm, 28 Jul 2025, Ayi et al., 2024, Chaintron et al., 6 Oct 2025).

1. Conceptual scope and taxonomy

Within the corpus, the most direct meaning of “Lévy-type mean-field system” is a mean-field evolution whose infinitesimal dynamics contains a Lévy–Khintchine jump term or whose stochastic forcing is specified through a Lévy process or a compensated Poisson random measure. In this sense, the topic includes controlled nonlinear Markov processes with measure-dependent Lévy kernels (Kolokoltsov et al., 2011), semigroup-based mean-field generators with drift, diffusion, and jump parts (Lim et al., 4 Aug 2025), and distribution-dependent SDEs of the form

$dX_t=b(t,X_t,\mathscr L_{X_t})\,dt+dL_t,$

with additive Lévy noise (Ye, 14 Apr 2026).

A second, narrower but still direct strand consists of optimization and game formulations. Here the mean field enters through $\mathbb E[X]$, $\mathbb E[u]$, or a stationary scalar statistic, while the noise is Brownian plus jump or purely Lévy. This includes stochastic linear-quadratic control with compensated Poisson random measures (Tang et al., 18 Mar 2025) and mean-field games with two-sided singular controls for one-dimensional Lévy processes (Oliú, 28 May 2025).

A third usage is effective rather than microscopic. The Hamiltonian Mean Field model is a deterministic long-range interacting rotor system, but its momentum-space trapping statistics are well described by a one-sided truncated Lévy distribution, so the system is “Lévy-type” at the level of coarse-grained sojourn times and continuous-time-random-walk interpretation, not at the level of a Lévy-driven state equation (Figueiredo et al., 2012).

Class	Defining structure	Representative reference
Generator-level MFG	Measure-dependent Lévy–Khintchine operator	(Kolokoltsov et al., 2011)
Semigroup mean-field model	$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$	(Lim et al., 4 Aug 2025)
DDSDE	Additive Lévy noise with law-dependent drift	(Ye, 14 Apr 2026)
Control and MFG	Poisson-jump LQ control; singular control of Lévy processes	(Tang et al., 18 Mar 2025, Oliú, 28 May 2025)
Effective Lévy dynamics	Truncated Lévy sojourn statistics in momentum space	(Figueiredo et al., 2012)

A common misconception is to identify every nonlocal or hybrid mean-field model with a Lévy-type one. The cited literature draws a sharper boundary: path-dependent memory on $C([0,T],\mathbb R^d)$, Markovian regime switching, and higher-order hypergraph interactions are nonlocal in time, environment, or interaction architecture, but not Lévy in the sense of jump generators or càdlàg state dynamics as formulated in those works (Throm, 28 Jul 2025, Nguyen et al., 2019, Ayi et al., 2024).

2. Generator and equation-level formulations

The most general generator-level formulation in the corpus appears in the mean-field game setting with nonlinear Markov processes. For class $i$, the controlled one-particle generator is

$$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$

where

$$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$

The mean-field law enters the drift, diffusion coefficient, and Lévy kernel, and the forward equation is the weak kinetic equation

$$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$

coupled to a backward HJB equation (Kolokoltsov et al., 2011).

A closely related but more explicitly transport-based formulation is the Lévy-type mean-field generator on $\mathbb R^d$,

$\mathbb E[X]$0

with

$\mathbb E[X]$1

$\mathbb E[X]$2

$\mathbb E[X]$3

The associated nonlinear mean-field evolution is

$\mathbb E[X]$4

and McKean–Vlasov diffusion is recovered by setting $\mathbb E[X]$5 (Lim et al., 4 Aug 2025).

At the SDE level, the additive-noise McKean–Vlasov formulation is

$\mathbb E[X]$6

where $\mathbb E[X]$7 is a $\mathbb E[X]$8-dimensional Lévy process with Lévy–Itô decomposition

$\mathbb E[X]$9

Its generator is

$\mathbb E[u]$0

so the law dependence is confined to the drift and the Lévy nonlocality is additive and law-independent (Ye, 14 Apr 2026).

These formulations show that “mean field” and “Lévy-type” interact at two distinct levels. The first is coefficient dependence on $\mathbb E[u]$1. The second is the analytic structure of the generator, which may include nonlocal jump terms. The literature represented here treats both simultaneously.

3. Well-posedness, semigroup stability, and propagation of chaos

The semigroup-based framework in the Lévy application of the abstract mean-field theory is organized around a mean-field generator

$\mathbb E[u]$2

together with a transport cost $\mathbb E[u]$3. A central device is the generator metric

$\mathbb E[u]$4

where $\mathbb E[u]$5 is the Bures–Wasserstein distance on covariance matrices and $\mathbb E[u]$6 is a Lévy–Wasserstein transport cost between Lévy measures. Under the Lipschitz criterion

$\mathbb E[u]$7

the nonlinear mean-field problem has a unique $\mathbb E[u]$8-stable solution and the associated $\mathbb E[u]$9-particle system satisfies pointwise propagation-of-chaos bounds in $A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$0 (Lim et al., 4 Aug 2025).

The same work derives the estimate

$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$1

and, in an average-form Lévy-driven model, the sharper rate

$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$2

This makes the propagation-of-chaos problem reducible to empirical-measure approximation under the chosen modulus $A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$3. The paper explicitly treats McKean–Vlasov diffusion as a special case, but its novelty lies in the jump part and in the generator-level treatment of drift, diffusion, and Lévy kernel within a unified metric (Lim et al., 4 Aug 2025).

Weak existence under low regularity is addressed for additive Lévy-driven DDSDEs by a different route. The key analytic estimate is a Krylov-type bound: for semimartingales of the form

$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$4

one has

$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$5

Using this estimate, approximation by bounded Wasserstein-Lipschitz drifts $A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$6, Aldous tightness, Prokhorov compactness, and Skorokhod representation, weak solutions are obtained for

$A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$7

under $A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$8-type drift control and semigroup assumptions on the Lévy process (Ye, 14 Apr 2026).

The nonlinear Markov process approach to mean-field games furnishes a parallel law-of-large-numbers statement. The $A(\mu)=A^\nabla(\mu)+A^\Delta(\mu)+A^J(\mu)$9-particle generator acting on measure functionals admits an expansion

$C([0,T],\mathbb R^d)$0

more precisely with a second-variational correction of order $C([0,T],\mathbb R^d)$1, and the resulting approximation of the limiting kinetic flow has rate $C([0,T],\mathbb R^d)$2. This rate subsequently feeds into the game-theoretic $C([0,T],\mathbb R^d)$3-equilibrium estimate (Kolokoltsov et al., 2011).

The analytical picture is therefore stratified. Semigroup transport methods yield stability and chaos estimates when the Lévy-type generator is already available. Krylov estimates and tightness establish weak existence under singular drift in additive-noise models. Generator expansions and nonlinear martingale problems control mean-field limits in game settings.

4. Control, optimization, and mean-field games with jumps

In stochastic control, the clearest direct Lévy-type mean-field model in the corpus is the general mean-field stochastic linear-quadratic problem with random coefficients and jump noise. The state equation is

$C([0,T],\mathbb R^d)$4

and the quadratic cost depends on $C([0,T],\mathbb R^d)$5, $C([0,T],\mathbb R^d)$6, $C([0,T],\mathbb R^d)$7, and $C([0,T],\mathbb R^d)$8. The optimal control is characterized in explicit state-feedback form by a stochastic Riccati equation with jumps, an auxiliary linear BSDE with jumps, and a deterministic operator equation for the mean processes. The final control law is

$C([0,T],\mathbb R^d)$9

where $i$0, $i$1, and $i$2 are built from the coefficients, the Riccati solution, and the affine BSDE terms (Tang et al., 18 Mar 2025).

In mean-field games, the representative-agent generator may itself be Lévy–Khintchine. The HJB equation takes the form

$i$3

with

$i$4

and the forward consistency equation is

$i$5

For $i$6 classes of agents, the population law is vector-valued, $i$7, and the coupling acts through the empirical measure. Under the stated smoothing, Lipschitz, and variational-differentiability assumptions, the mean-field solution produces a perfect $i$8-equilibrium with

$i$9

This is a generator-level generalization from diffusion models to jump-diffusions, stable-like processes, pure jump processes, and other Lévy-type Feller dynamics (Kolokoltsov et al., 2011).

A separate singular-control strand studies a one-dimensional Lévy process controlled by two-sided bounded-variation controls: $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$0 The mean-field coupling is not through the full law but through the scalar statistic

$$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$1

with discounted and ergodic criteria built from a running cost $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$2 and proportional control costs $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$3. For fixed $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$4, the optimal control is reflection on an interval $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$5, characterized by an adjoint Dynkin game and the barrier identities

$$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$6

Discounted and ergodic MFG equilibria are obtained by Brouwer fixed points; equilibrium controls are of double-reflection type; an Abelian limit connects discounted and ergodic equilibria; and nonuniqueness is explicitly exhibited by counterexample (Oliú, 28 May 2025).

These control and game formulations show that Lévy-type mean-field systems are not restricted to passive evolution equations. They also support explicit feedback synthesis, free-boundary characterizations, random-measure equilibria, and $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$7-approximation statements.

5. Emergent Lévy statistics, CTRW structure, and weak ergodicity breaking

A different manifestation of the topic appears in the Hamiltonian Mean Field model

$$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$8

This is a fully coupled long-range rotor system, mean-field because each rotor interacts equally with every other one under the $$A^i[t,\mu,u]f(z)=\big(h_i(t,z,\mu,u),\nabla f(z)\big)+L_i[t,\mu]f(z),$$9 Kac normalization. In the $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$0 limit the dynamics is governed by the Vlasov equation, while at finite $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$1 fluctuations act as weak collisions (Figueiredo et al., 2012).

The key observation is that, after coarse-graining momentum space into cells of width $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$2, particle motion can be interpreted as a continuous-time random walk between cells. If $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$3 is the $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$4-th sojourn time in cell $$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$5 and

$$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$6

then the empirical sojourn-time law is well described by a truncated one-sided Lévy distribution,

$$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$7

with power-law tail

$$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$8

The sums

$$L_i[t,\mu]f(z) = \frac12\big(G_i(t,z,\mu)\nabla,\nabla\big)f(z) + \big(b_i(t,z,\mu),\nabla f(z)\big) + \int_{\mathbb R^d} \Big( f(z+y)-f(z)-(\nabla f(z),y)\mathbf 1_{B_1}(y) \Big)\,\nu_i(t,z,\mu,dy).$$9

display stable-law scaling

$$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$0

and fitted exponents remain far below the Gaussian value $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$1. Representative values reported in the paper are $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$2 and $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$3 as $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$4 increases from $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$5 to $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$6 (Figueiredo et al., 2012).

The ergodic implication is weak ergodicity breaking. Time averages are weighted by residence times $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$7, and when residence times are Lévy-distributed with $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$8, the distribution of time-averaged observables remains broad rather than collapsing to a delta function. Because the Lévy law is truncated, asymptotic ergodicity is recovered for finite $$\frac{d}{dt}(g,\mu_t)=\big(A[t,\mu_t,\gamma(t,\cdot)]g,\mu_t\big),$$9, but only beyond a crossover time tied to the Lévy-to-Gaussian transition. Numerically, the truncation length scales approximately as

$\mathbb R^d$0

both at thermodynamic equilibrium and in a homogeneous waterbag quasi-stationary state, implying that the ergodicity-restoration time diverges with system size (Figueiredo et al., 2012).

This result is central for terminology. The model is not a Lévy-driven McKean–Vlasov equation, yet it is a paradigmatic mean-field Hamiltonian system with effective Lévy-type trapping dynamics in momentum space. The Lévy character lies in waiting-time statistics and CTRW interpretation, not in an explicit Lévy–Khintchine state generator.

6. Adjacent frameworks, boundaries of the label, and emerging directions

Several papers in the corpus are best understood as structurally adjacent rather than directly Lévy-type. Mean-field interactions with Markovian switching involve diffusions whose coefficients jump with a common finite-state chain,

$\mathbb R^d$1

and the empirical limit is the random conditional law

$\mathbb R^d$2

This is a hybrid càdlàg mean-field system and is conceptually relevant to common-jump environments, but it does not contain a Lévy measure or state jumps in $\mathbb R^d$3 (Nguyen et al., 2019).

Adaptive/co-evolving network limits are deterministic path-dependent mean-field systems. After elimination of the weight dynamics,

$\mathbb R^d$4

the limit is a graphon-indexed path-space flow on $\mathbb R^d$5. The nonlocality is memory in time rather than jump nonlocality in state space, and the framework excludes Poisson random measures, càdlàg trajectories, and Lévy generators as stated (Throm, 28 Jul 2025).

Higher-order hypergraph mean-field limits likewise remain outside the Lévy class. The microscopic model is a deterministic ODE with group interactions of all orders,

$\mathbb R^d$6

and the limit is a UR-hypergraphon Vlasov equation with infinitely many interaction orders. The nonlocality is purely interaction-theoretic; there is no Lévy generator, jump process, or fractional operator (Ayi et al., 2024).

Tanaka-style fixed-point formulations for non-exchangeable network systems are also additive-noise and continuous-path in their developed form: $\mathbb R^d$7 The papers explicitly work on $\mathbb R^d$8, not on càdlàg path spaces, and derive Brownian Fokker–Planck closures rather than integro-differential Lévy equations. This suggests a possible extension to additive Lévy inputs by replacing continuous path spaces with $\mathbb R^d$9, but that extension is not carried out (Chaintron et al., 6 Oct 2025).

These boundaries clarify the term. In this literature, a mean-field model is directly Lévy-type when jumps or Lévy–Khintchine nonlocality occur in the actual state dynamics or generator. It is only indirectly relevant when the model instead features switching, memory, higher-order interaction architecture, or continuous-path additive-noise abstractions.

A plausible implication is that current work is converging on three complementary directions. One is generator-level unification, exemplified by the semigroup/transport framework and the metric $\mathbb E[X]$00 on Lévy triplets (Lim et al., 4 Aug 2025). A second is low-regularity existence theory for additive-noise McKean–Vlasov equations via semigroup smoothing and Krylov estimates (Ye, 14 Apr 2026). A third is the extension of game and control formulations—from linear-quadratic feedback synthesis to singular-control free-boundary problems—to explicitly jump-driven mean-field systems [(Tang et al., 18 Mar 2025); (Oliú, 28 May 2025); (Kolokoltsov et al., 2011)]. Taken together, these works define Lévy-type mean-field systems as a family of models in which population dependence and nonlocal stochastic dynamics are coupled at the level of generators, equations, equilibria, or effective transport statistics.