Ambrosio-Figalli-Trevisan Superposition Principle

• The Ambrosio-Figalli-Trevisan principle is a representation theorem that links Eulerian measure-valued solutions of continuity or Fokker–Planck equations with Lagrangian probability measures on path space.
• It extends both first-order (ODE trajectory) and second-order (martingale problem) formulations, relaxing classical integrability conditions through radial and Lyapunov-type hypotheses.
• Its applications span singular and nonlinear Fokker–Planck equations, mean field games, and non-local Lévy-type operators, offering a unifying bridge between analytic PDE methods and probabilistic weak solutions.

Searching arXiv for core papers on the Ambrosio–Figalli–Trevisan superposition principle and related extensions. Searching arXiv for Trevisan’s multidimensional diffusion/martingale problem paper. Searching arXiv for Ambrosio/Trevisan continuity-equation and metric-space generalizations. The Ambrosio–Figalli–Trevisan superposition principle is a representation theorem that connects Eulerian measure-valued solutions of continuity or Fokker–Planck–Kolmogorov equations with Lagrangian probability measures on path space. In its stochastic form, it states that a probability solution of a Fokker–Planck equation can be realized as the family of time marginals of a probability measure on trajectories solving the associated martingale problem; in its first-order antecedent, it represents solutions of the continuity equation by ODE trajectories. In this sense, it is a bridge between analytic PDE formulations and probabilistic weak solutions of SDEs or martingale problems (Bogachev et al., 2019, Bondi et al., 31 Jul 2025).

1. Canonical formulation

The classical deterministic starting point is Ambrosio’s superposition principle for the continuity equation

$\partial_t \mu_t + \nabla \cdot (b(t,x)\mu_t) = 0,$

which represents a measure-valued solution $(\mu_t)$ as the time marginals of a probability measure on absolutely continuous curves satisfying the ODE $\dot w_t=b(t,w_t)$. Figalli and Trevisan developed the second-order stochastic version for Fokker–Planck equations associated with SDEs, thereby replacing ODE trajectories by solutions of the martingale problem (Rockner et al., 2019, Bondi et al., 31 Jul 2025).

For a Borel diffusion matrix $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$, symmetric and nonnegative definite, and a Borel drift $b(t,x)=(b^i(t,x))_{i=1}^d$, the Kolmogorov operator is

$$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$

A probability solution $(\mu_t)_{t\in[0,T]}$ of

$\partial_t\mu_t=L^*\mu_t,\qquad \mu_0=\nu,$

is a weakly continuous curve of probability measures satisfying the integral identity

$$\int_{\mathbb{R}^d} \varphi(x)\,\mu_t(dx) = \int_{\mathbb{R}^d} \varphi(x)\,\nu(dx) + \int_0^t \int_{\mathbb{R}^d} L\varphi(s,x)\,\mu_s(dx)\,ds$$

for all test functions $\varphi\in C_0^\infty(\mathbb{R}^d)$, with local integrability of the coefficients against $(\mu_t)$0 (Bogachev et al., 2019).

The superposition principle then asserts that there exists a Borel probability measure $(\mu_t)$1 on the path space $(\mu_t)$2, with canonical process $(\mu_t)$3, such that $(\mu_t)$4 for every $(\mu_t)$5, and for every $(\mu_t)$6 the process

$(\mu_t)$7

is a martingale. Thus each probability solution is generated by a solution of the martingale problem, and conversely any martingale solution yields a solution of the PDE (Bogachev et al., 2019, Shaposhnikov et al., 20 Aug 2025).

2. Integrability hypotheses and the Bogachev–Röckner–Shaposhnikov refinement

A central issue in the theory is how little can be assumed on $(\mu_t)$8 and $(\mu_t)$9. In the formulation emphasized by Bogachev–Röckner–Shaposhnikov, the classical global integrability requirement on $\dot w_t=b(t,w_t)$0 is replaced by the weaker radial condition

$\dot w_t=b(t,w_t)$1

This asks for integrability of the diffusion and of the radial component of the drift, rather than of the full drift norm, and allows $\dot w_t=b(t,w_t)$2 to have quadratic growth in $\dot w_t=b(t,w_t)$3 (Bogachev et al., 2019).

A further consequence recorded in the same work is that, assuming $\dot w_t=b(t,w_t)$4, it is sufficient to impose the one-sided bounds

$\dot w_t=b(t,w_t)$5

to obtain the superposition conclusion. This is a Lyapunov-type regime rather than a bounded-coefficient regime, and it is tailored to coercive or dissipative settings in which only radial control is available (Bogachev et al., 2019).

These weakened hypotheses are significant for nonlinear and singular applications because they align with the structure of a priori estimates obtained from logarithmic or polynomial Lyapunov functions. Later reviews present the superposition principle precisely as one of the standard tools for singular McKean–Vlasov equations, together with Zvonkin transformation, Markov marginal uniqueness, and stochastic sewing (Bondi et al., 31 Jul 2025).

3. Martingale problems, PDE uniqueness, and the scope of the principle

The principle is best understood as an equivalence mechanism between an analytic object and a probabilistic object. On the analytic side lies the weak solution of the forward Kolmogorov equation; on the probabilistic side lies the martingale problem for the generator $\dot w_t=b(t,w_t)$6. In the local diffusion case, the representation takes place on $\dot w_t=b(t,w_t)$7; in the jump case it takes place on the Skorokhod space $\dot w_t=b(t,w_t)$8. In both settings, the time marginals of the path-space law reproduce the prescribed curve of measures (Rockner et al., 2019).

A standard corollary is a one-way uniqueness transfer: if the martingale problem MP$\dot w_t=b(t,w_t)$9 has existence and uniqueness, then the Fokker–Planck equation has at most one probability-valued solution. This is frequently used when the PDE is in divergence form with measurable coefficients and direct analytic uniqueness is difficult (Bondi et al., 31 Jul 2025).

At the same time, the superposition principle is existential rather than intrinsically well-posed. It yields a probabilistic representation whenever the hypotheses are met, but it does not by itself guarantee uniqueness of the SDE or martingale-problem solution. Extra assumptions, such as local nondegeneracy and local Lipschitz regularity of the diffusion matrix or other well-posedness criteria for the martingale problem, are needed for uniqueness. This distinction is explicit in applications to mean field games, where existence of a representation is enough for equilibrium construction, while uniqueness of the SDE is not essential (Shaposhnikov et al., 20 Aug 2025).

A related misconception is that the principle is only a probabilistic restatement of an already well-posed SDE. In fact, the direction it is most often used in is the opposite one: start from a measure-valued solution of a PDE, often constructed by compactness or Lyapunov methods, and only then lift it to a path-space measure. This reversal is precisely what makes the principle effective for rough coefficients and weak compactness arguments (Bogachev et al., 2019).

4. Nonlinear Fokker–Planck equations, McKean–Vlasov dynamics, and mean field games

In nonlinear settings the principle is typically applied after freezing or linearizing the dependence on the law. For a nonlinear Fokker–Planck equation of McKean–Vlasov type,

$A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$0

one first fixes a solution $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$1 or $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$2, interprets the coefficients $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$3, $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$4 as given, and applies the linear superposition principle to the corresponding frozen equation. This yields a weak solution of the associated McKean–Vlasov SDE whose law density is exactly the prescribed PDE solution (Bondi et al., 31 Jul 2025).

This mechanism is explicit in the review of singular McKean–Vlasov equations, where bounded measurable coefficients and a nonlinear Fokker–Planck solution $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$5 imply the existence of a weak solution to

$A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$6

such that $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$7 is the law density of $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$8. In singular-distributional settings, the principle is applied at the level of regularized linearized equations and then combined with stability of rough martingale problems (Bondi et al., 31 Jul 2025).

A closely related use appears in mean field games with unbounded coefficients. There the equilibrium is first constructed analytically as a solution $A(t,x)=(a^{ij}(t,x))_{i,j=1}^d$9 of a nonlinear controlled Fokker–Planck equation

$b(t,x)=(b^i(t,x))_{i=1}^d$0

and only afterwards represented probabilistically through the SDE

$b(t,x)=(b^i(t,x))_{i=1}^d$1

In that framework the AFT principle is the device that turns a PDE-defined equilibrium into a Nash-type mean field game equilibrium of controlled SDEs (Shaposhnikov et al., 20 Aug 2025).

5. Extensions beyond local finite-dimensional diffusions

One major extension replaces local diffusions by Lévy-type generators. For operators of the form

$b(t,x)=(b^i(t,x))_{i=1}^d$2

where $b(t,x)=(b^i(t,x))_{i=1}^d$3 is a non-local Lévy-type term built from a family of Lévy measures $b(t,x)=(b^i(t,x))_{i=1}^d$4, Röckner–Xie–Zhang prove a superposition principle on the càdlàg path space $b(t,x)=(b^i(t,x))_{i=1}^d$5. Under growth and integrability conditions adapted to diffusion, drift, and jump intensity, every weak solution of

$b(t,x)=(b^i(t,x))_{i=1}^d$6

is represented by a martingale solution $b(t,x)=(b^i(t,x))_{i=1}^d$7 with $b(t,x)=(b^i(t,x))_{i=1}^d$8. In this non-local setting the principle yields the equivalence between well-posedness of the Fokker–Planck equation and well-posedness of the martingale problem for the Lévy-type operator (Rockner et al., 2019).

The same paper applies the non-local principle to the fractional porous media equation

$b(t,x)=(b^i(t,x))_{i=1}^d$9

rewriting it as a non-local Fokker–Planck equation with state-dependent Lévy measure and deriving a probabilistic representation by a distribution-dependent jump SDE. This places fractional porous media dynamics within the same PDE-to-path-space paradigm as classical diffusion-driven nonlinear Fokker–Planck equations (Rockner et al., 2019).

A different extension concerns metric spaces without smooth structure and without reference measure. Stepanov and Trevisan formulate continuity equations in duality with an algebra of observables $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$0, encode vector fields as derivations, and prove that a narrowly continuous solution curve $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$1 can be represented by a measure on $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$2 concentrated on absolutely continuous curves satisfying the associated ODE in observables. They also identify a hierarchy of superposition principles,

$$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$3

linking decomposition of acyclic metric currents, continuity equations, and absolutely continuous curves in Wasserstein spaces (Stepanov et al., 2015).

Recent work on local 1-dimensional currents pushes the geometric side further: every one-dimensional locally normal metric current admits an integral representation through currents associated to curves with locally finite length, generalizing the Paolini–Stepanov theorem from Ambrosio–Kirchheim normal currents to the locally finite-mass regime. Although this is a geometric-current analogue rather than a diffusion theorem, it lies in the same superposition tradition of representing Eulerian or current-type objects by pathwise constituents (Ambrosio et al., 23 Mar 2025).

6. Limitations, singular regimes, and contemporary uses

The finite-dimensional diffusion principle does not extend verbatim to all infinite-dimensional settings. On separable Hilbert spaces, a full norm-topology version fails in general: there exist Fokker–Planck–Kolmogorov solutions whose marginals cannot be realized by norm-continuous paths. The available result is therefore a restricted superposition principle for subclasses of solutions obtained as limits of Galerkin approximations under coercivity, Lyapunov, and compactness assumptions, and this restricted theory also has a nonlinear extension (Dieckmann, 2020).

At the opposite end of the spectrum, the classical absolutely continuous continuity-equation framework is too narrow for $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$4-BV curves with singular flux. In that case the flux decomposes as $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$5, and $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$6 may be non-trivial even for Lipschitz curves in $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$7. The modern remedy is an augmented phase-space construction, together with a minimal singular-flux selection and a BV superposition principle that represents the pair $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$8 by BV curves carrying explicit jump trajectories. This extends the superposition paradigm from absolutely continuous transport to BV evolutions with jump structure (Almi et al., 18 Jun 2025).

The principle is also increasingly used as a construction tool in interacting-particle systems with singular drift. For multilevel Dyson Brownian motions, the superposition principle of Figalli and Trevisan is used to lift a measure-valued solution $$L u(t,x) = \sum_{i,j=1}^d a^{ij}(t,x)\,\partial_{x_i}\partial_{x_j} u(t,x) + \sum_{i=1}^d b^i(t,x)\,\partial_{x_i} u(t,x).$$9 of the associated Fokker–Planck equation to a martingale solution of the multilevel SDE, thereby obtaining existence, stochastic differential equation representation, and uniqueness in distribution for all $(\mu_t)_{t\in[0,T]}$0 (Budway et al., 2024).

Across these variants, the unifying content remains unchanged: a continuity or Fokker–Planck equation is not merely a marginal law evolution but an encoded pathwise dynamics. The Ambrosio–Figalli–Trevisan superposition principle makes that encoding explicit by replacing an Eulerian family of measures with a measure on trajectories that solves the appropriate ODE, martingale problem, or jump martingale problem, and it does so under hypotheses far weaker than those required for classical pathwise SDE theory (Bogachev et al., 2019, Rockner et al., 2019).