Lions Derivative in Wasserstein Analysis

• Lions derivative is a first-order, state-dependent differential defined for functionals on probability measures by lifting them to L² spaces.
• It serves as a canonical tool in Wasserstein calculus, playing a crucial role in mean-field games, McKean–Vlasov dynamics, and master equations.
• Higher-order Lions calculus introduces additional free variables and mixed derivatives, enabling advanced Itô formulas and numerical schemes in both finite and infinite-dimensional settings.

The Lions derivative is the differential object attached to a functional of a probability law, typically on $\mathcal P_2(\mathbb R^d)$, obtained by lifting the functional to a map on square-integrable random variables and representing the resulting Fréchet derivative as evaluation of a deterministic function at the underlying state. In contemporary Wasserstein calculus it is the standard first-order differential notion for mean-field games, McKean–Vlasov dynamics, and master equations. The same surname appears in other settings, however, and the Lions derivative must be distinguished from Lasry–Lions regularization and from J.-L. Lions’ boundary-integral formulas for reproducing kernels (Chen et al., 2017, Djitte et al., 10 Feb 2026).

1. Classical formulation on Wasserstein space

In the standard finite-dimensional formulation, one starts with a map

$f:\mathcal P_2(\mathbb R^d)\to \mathbb R$

or more generally $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$, and defines its lift on a linear Hilbert space by

$$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$

The map $f$ is then said to be Lions differentiable, or $L$-differentiable, at $\mu$ if the lifted map is Fréchet differentiable at some $X$ with $\mathcal L(X)=\mu$, and the derivative can be represented as

$D\widetilde f(X)=\partial_\mu f(\mu)(X).$

This is the basic construction used in higher-order Lions calculus and in later stochastic applications (Salkeld, 2023).

A decisive structural fact is that the derivative of a law-invariant lift is itself a function of the underlying random variable. Wu and Zhang proved, in the scalar $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$0-valued case, that if

$f:\mathcal P_2(\mathbb R^d)\to \mathbb R$1

is law invariant and continuously Fréchet differentiable, then

$f:\mathcal P_2(\mathbb R^d)\to \mathbb R$2

where $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$3 is deterministic and depends only on the law of $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$4 (Wu et al., 2017). This is the statement that makes it legitimate to regard the derivative as a function of $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$5, rather than as an arbitrary random variable.

The first-order increment formula on Wasserstein space takes the form

$f:\mathcal P_2(\mathbb R^d)\to \mathbb R$6

with $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$7 a coupling of $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$8 and $f:\mathcal P_2(\mathbb R^d)\to \mathbb R$9 (Salkeld, 2023). In this sense, the Lions derivative measures the first-order response of a functional to infinitesimal perturbations of the law implemented through perturbations of an underlying random variable.

The basic examples already exhibit the derivative’s state-dependent character. For

$f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$0

the lifted map is $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$1, and the Lions derivative is

$f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$2

For multilinear polynomial-type functionals of $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$3, the derivative is obtained by differentiating the integrand in each slot and inserting a free variable $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$4 into the corresponding position (Salkeld, 2023).

2. Relation to intrinsic, extrinsic, and density-based derivatives

A major theme of the literature is that the Lions derivative is only one member of a family of derivatives on spaces of measures. In one common formulation, the intrinsic derivative is defined directionally along transport perturbations: $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$5 with $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$6. In the formulation used for distribution-dependent SDEs, $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$7 is $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$8-differentiable if, in addition, the full Fréchet-type remainder condition in $f:\mathcal P_2(\mathbb R^e)\to \mathbb R^d$9 holds; then the intrinsic derivative and the $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$0-derivative coincide (Huang et al., 2021).

On complete Riemannian manifolds and on spaces of finite measures, the relation between derivative notions can be made especially concrete. For a sufficiently regular $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$1, the extrinsic derivative

$$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$2

coincides with the linear functional derivative, while the intrinsic derivative and the Lions derivative agree and satisfy

$$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$3

On probability space the same principle holds with the centered extrinsic derivative $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$4 in place of $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$5 (Ren et al., 2019). This yields a direct computational rule: under the stated regularity, the Lions derivative is the spatial gradient of the extrinsic derivative.

A different route arises from differentiation with respect to a density on an underlying reference probability space. For

$$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$6

the derivative with respect to the density variable $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$7 is represented by a scalar function $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$8, and when $$\widetilde f(X)=f(\mathcal L(X)),\qquad X\in L^2(\Omega;\mathbb R^d).$$9 is Lions differentiable one has

$f$0

Thus the density derivative is a scalar potential whose spatial derivative recovers the Lions derivative (Buckdahn et al., 2020).

These formulations are complementary rather than redundant. The lifted $f$1-definition is canonical for $f$2; the transport-based intrinsic derivative is often better adapted to stochastic analysis; the extrinsic derivative gives a practical route on manifolds; and the density-based derivative is natural in weak formulations of mean-field control.

3. Higher-order Lions calculus

Higher-order Lions calculus differs sharply from ordinary Fréchet calculus because each new Lions derivative creates a new free variable. Already at second order, the distinction between

$f$3

is fundamental: the first differentiates in the already existing free variable, while the second differentiates again in the measure and therefore introduces an additional state variable (Salkeld, 2023).

To organize this, higher-order derivatives are indexed by partition sequences $f$4. Repeated labels mean differentiation in an already existing free variable; a new label means a new Lions differentiation. This produces a family of derivatives $f$5, rather than a single $f$6-linear tensor. The resulting Taylor expansion on $f$7 is

$f$8

with the increment terms built from couplings $f$9 and the remainder controlled by moments of $L$0 under the coupling (Salkeld, 2023).

For functions of both a Euclidean variable and a measure,

$L$1

the notation is extended by adding a symbol for differentiation in the Euclidean variable. This yields mixed derivatives such as $L$2, $L$3, and $L$4, all of which are distinct objects (Salkeld, 2023).

In rough-path and probabilistic mean-field expansions, the combinatorics of iterated Lions derivatives are encoded by Lions trees and Lions forests. These objects record which derivative slots share the same free variable and which new variables are created by subsequent Lions differentiations. The associated coupled Hopf or bialgebra structure is not a cosmetic reformulation; it is the bookkeeping mechanism needed because iterated Lions–Taylor expansions are indexed by partition structure rather than by ordinary multi-indices (Delarue et al., 2022, Salkeld, 2023).

4. Infinite-dimensional, Banach, and path-space extensions

The finite-dimensional lift-to-$L$5 picture becomes nontrivial when the state space is infinite dimensional or when the functional takes values in a Banach space. For

$L$6

with $L$7 a separable Hilbert space and $L$8 a Banach space, the lift

$L$9

remains natural, but $\mu$0 is no longer canonically representable as a random variable. The infinite-dimensional theory developed in 2024 interprets the lifted derivative as a vector measure and defines the Lions derivative as the Radon–Nikodym derivative of the associated image vector measure: $\mu$1 In that framework,

$\mu$2

which yields a canonical Banach-valued extension of the Lions derivative (Vogler et al., 2024).

This extension supports an Itô formula for law flows and a mild Itô formula for mean-field SPDEs. For a Hilbert space valued mean-field stochastic evolution equation, the regular $\mu$3-derivative enters the law-dependent drift terms of the mild Itô formula and serves as the basis for higher-order Taylor expansion and higher-order numerical schemes (Vogler et al., 2024).

A parallel development treats measures on Banach path spaces, especially

$\mu$4

There the tangent space at $\mu$5 is

$\mu$6

the intrinsic derivative is defined by pushforward under $\mu$7, and the Lions derivative is the stronger Fréchet-type notion in this Banach setting. The resulting chain rule for laws of Banach-valued random variables extends the standard identity

$\mu$8

to path-space laws and underpins later Bismut formulas for distribution-path dependent SDEs (Bao et al., 2020).

5. Bismut formulas and McKean–Vlasov applications

One of the most active uses of the Lions derivative is the analysis of nonlinear semigroups generated by McKean–Vlasov SDEs. For

$\mu$9

where $X$0 solves a distribution-dependent SDE with initial law $X$1, the problem is to differentiate $X$2 with respect to $X$3. In the nondegenerate setting, the 2018 Bismut formula gives a stochastic representation

$X$4

and analogous formulas are obtained in degenerate Hamiltonian settings via Malliavin divergence (Ren et al., 2018). These formulas yield explicit derivative estimates and total variation bounds for $X$5.

For singular drifts, Huang–Song–Wang developed a Bismut formula for intrinsic/Lions derivatives of distribution-dependent SDEs whose drift is only Dini continuous in the spatial variable. Their method combines a distribution-dependent Zvonkin transform with Malliavin calculus, and under an additional integrability assumption upgrades intrinsic differentiability to full $X$6-differentiability (Huang et al., 2021).

A later development addresses singular interaction kernels of convolution type,

$X$7

for which coefficient-level Lions differentiability may fail. In that setting the derivative object is the intrinsic derivative of the semigroup

$X$8

and the resulting Bismut formula is obtained without assuming that the drift coefficient itself is Lions differentiable in the measure variable (Ren, 10 Apr 2026). This marks a conceptual shift: semigroup differentiability in the initial law can persist beyond coefficient-level Lions differentiability.

For distribution-path dependent SDEs on $X$9, the Banach-space Lions derivative supports exact Bismut formulas in additive and certain multiplicative noise settings, and asymptotic Bismut formulas when path-dependent diffusion prevents an exact representation (Bao et al., 2020). Across these settings, the Lions derivative is the measure-space analogue of the spatial gradient in classical Bismut–Elworthy–Li theory.

6. Distinctions, misconceptions, and scope of the term

The expression “Lions derivative” is often conflated with objects that are mathematically unrelated. The first major source of confusion is Lasry–Lions regularization. The 2017 paper on discounted Hamilton–Jacobi equations studies intrinsic Lasry–Lions approximation

$\mathcal L(X)=\mu$0

selects distinguished supergradients of viscosity solutions, and analyzes generalized characteristics, but it explicitly does not treat derivatives with respect to probability measures, lifted derivatives on $\mathcal L(X)=\mu$1, Wasserstein differential calculus, or $\mathcal L(X)=\mu$2 (Chen et al., 2017).

The second source of confusion is the “Lions’ type formula” in PDE/RKHS theory. In the 2026 paper on fractional harmonic functions, this expression refers to a boundary-integral formula for a reproducing kernel,

$\mathcal L(X)=\mu$3

originating with J.-L. Lions’ work on harmonic-function RKHSs. It is unrelated to differentiation on Wasserstein space (Djitte et al., 10 Feb 2026).

Even within measure calculus, the Lions derivative should not be identified with every first-order measure derivative. The density-based derivative $\mathcal L(X)=\mu$4 is scalar-valued and defined only up to additive constants, whereas the Lions derivative is the vector-valued quantity obtained after spatial differentiation: $\mathcal L(X)=\mu$5 under the stated regularity assumptions (Buckdahn et al., 2020). Likewise, intrinsic differentiability and $\mathcal L(X)=\mu$6-differentiability are closely related but not interchangeable in every framework; the latter includes the Fréchet-type remainder condition in the tangent norm (Huang et al., 2021).

In modern usage, therefore, the Lions derivative denotes the state-dependent derivative of a functional of a probability measure, canonically represented as $\mathcal L(X)=\mu$7, together with its higher-order extensions $\mathcal L(X)=\mu$8, $\mathcal L(X)=\mu$9, and their mixed iterates. Its role is foundational in Wasserstein analysis, but its precise realization depends on the ambient setting: lifted Fréchet derivatives on $D\widetilde f(X)=\partial_\mu f(\mu)(X).$0, transport derivatives on measure space, vector-measure Radon–Nikodym densities in infinite dimensions, or path-space derivatives on Banach-valued laws.