Nash–Lasry–Lions Equation Overview

• Nash–Lasry–Lions Equation is a formulation of equilibrium PDEs that encapsulates master equations in mean-field games and free-boundary price models.
• It provides a decoupling framework to derive Hamilton–Jacobi–Bellman and Fokker–Planck systems from large population dynamics and finite-state games.
• The equation’s versatility drives both numerical methods development and analytic proofs of uniqueness and monotonicity in diverse applications.

The expression Nash–Lasry–Lions equation is used in the supplied arXiv literature for equilibrium PDEs associated with Lasry–Lions theory and Nash consistency. In mean field theory, it denotes the master equation for a value field $U(x,m,t)$ depending on an individual state, a population law, and time, and from which the coupled Hamilton–Jacobi–Bellman and Fokker–Planck system can be recovered as a decoupling field (Bensoussan et al., 2014). In finite-state continuous-time dynamic games, the same object reduces to a nonlinear ODE system for the common value function of a symmetric Markov perfect equilibrium (Höfer et al., 28 Jul 2025). In a separate price-formation usage, the label is also applied to the Lasry–Lions free-boundary model for buyer and seller densities and an agreed price $p(t)$, obtained as the fast-trading limit of a Boltzmann-type market model (Burger et al., 2013).

1. Terminological scope and historical placement

In the supplied sources, the nomenclature is used in more than one mathematical setting. One line of work begins with the Lasry–Lions free-boundary price model proposed in 2007, where the unknowns are nonnegative buyer and seller densities together with a moving agreed price. Another line concerns Nash equilibria in large-population games, where P.-L. Lions introduced the concept of a master equation in lectures at the Collège de France, and where the equation is derived heuristically from the PDE system associated with a large but finite Nash game (Burger et al., 2013, Bensoussan et al., 2014).

The finite-player interpretation is explicit in later work on continuous-time Markov games. There, the “corresponding Nash system reduces to the Nash-Lasry-Lions equation for the common value function, also known as the master equation in the mean-field setting,” and this finite-state problem becomes a nonlinear ODE with a unique classical solution (Höfer et al., 28 Jul 2025). A common misconception is therefore avoided by the supplied literature itself: the term does not designate a single universal PDE form, but a family of Lasry–Lions equilibrium equations whose exact realization depends on whether the setting is mean field control/game theory, finite-state Nash systems, or free-boundary price formation.

2. Master-equation formulation in mean field theory

In the mean-field-game formulation of Bensoussan–Frehse–Yam, the master field is a function

$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$

with terminal condition $U(x,m,T)=h(x,m)$. Writing

$$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$

the Lasry–Lions master equation is

$$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$

Here $\delta U/\delta m$ is the $L^2$–Gâteaux derivative with respect to the measure variable, the data $f,g,h,\sigma$ are assumed at least $C^2$, and the diffusion matrix $p(t)$0 is uniformly elliptic (Bensoussan et al., 2014).

The central structural point is that the master equation encodes the full equilibrium dependence on the population distribution. If $p(t)$1 solves the Fokker–Planck equation under the feedback $p(t)$2 and one sets $p(t)$3, then $p(t)$4 satisfies the coupled HJB–FP system. In this sense, the master equation is the decoupling field for the forward–backward mean-field-game PDEs (Bensoussan et al., 2014).

A more general control-interaction version appears in the mean field games of controls framework. There the unknown is again $p(t)$5, but the equation contains intrinsic derivatives on Wasserstein space, idiosyncratic and common-noise operators,

$p(t)$6

$p(t)$7

and a lifted Hamiltonian $p(t)$8 defined through a fixed-point map in the joint state–control distribution (Jackson et al., 30 Jan 2026). This formulation makes explicit that the Nash–Lasry–Lions equation is not merely a first-order mean-field correction to HJB, but an infinite-dimensional parabolic PDE on $p(t)$9.

3. Finite-player Nash systems, probabilistic representations, and limiting consistency

The master equation is motivated by the large-$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$0 limit of finite-player Nash systems. In the $$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$1-player HJB–FP derivation, one introduces the empirical measure

$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$2

and formally identifies the player-$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$3 value by

$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$4

By translating Gâteaux derivatives with respect to the empirical measure into ordinary gradients in the player coordinates and sending $$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$5, one recovers the master equation (Bensoussan et al., 2014).

The same equilibrium structure admits a probabilistic reformulation. In the McKean–Vlasov FBSDE approach, one freezes a candidate flow $$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$6 and applies the stochastic maximum principle, obtaining a forward state equation and a backward adjoint equation coupled through the optimal control

$$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$7

Imposing the consistency condition $$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$8 yields a McKean–Vlasov FBSDE whose solution generates the mean-field equilibrium. Under the assumptions listed in the paper, the resulting distributed strategies form an $$U=U(x,m,t), \qquad x\in\mathbb R^n,\quad m\in L^2(\mathbb R^n)\cap\mathcal P(\mathbb R^n),\quad t\in[0,T],$$9-Nash equilibrium for the $U(x,m,T)=h(x,m)$0-player game with

$U(x,m,T)=h(x,m)$1

and under stronger moment assumptions the rate improves to $U(x,m,T)=h(x,m)$2 (Carmona et al., 2012).

In finite-state continuous-time dynamic games, the abstraction becomes especially explicit. For a symmetric game of $U(x,m,T)=h(x,m)$3 identical players on $U(x,m,T)=h(x,m)$4, the empirical distribution of the untagged players lies in

$U(x,m,T)=h(x,m)$5

and the common value function satisfies the $U(x,m,T)=h(x,m)$6-NLL system

$U(x,m,T)=h(x,m)$7

Under the convexity assumption on the running cost, this is a standard finite-dimensional ODE with continuous right-hand side, so Picard–Lindelöf yields a unique classical solution and the induced feedback is the unique symmetric Markov perfect equilibrium (Höfer et al., 28 Jul 2025).

4. Monotonicity, regularity, and well-posedness theory

The analytic status of the master equation has evolved substantially across the supplied literature. In the 2014 treatment, the presentation is entirely formal: the existence of all derivatives in $U(x,m,T)=h(x,m)$8 is assumed, and no full analytic proof of well-posedness is given, although the linear–quadratic setting yields explicit formulas through Riccati equations and hence a global classical solution in that special case (Bensoussan et al., 2014).

A major organizing principle for well-posedness is monotonicity. In the mean field games of controls setting, Jackson–Mészáros distinguish displacement semi-monotonicity and Lasry–Lions monotonicity. The latter is stated as

$U(x,m,T)=h(x,m)$9

$$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$0

for all $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$1 and $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$2. Under only assumptions on the Lagrangian and terminal cost, with either displacement-semi-monotonicity or Lasry–Lions monotonicity, they obtain unique classical solutions of the master equation. In the LL-monotone case, the result holds for arbitrary time horizon $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$3 and includes common noise with constant intensity (Jackson et al., 30 Jan 2026).

Volatility control introduces a sharper difficulty because the master equation becomes nonlinear in both $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$4 and $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$5. Mou–Zhang–Zhou therefore introduce second-order monotonicity conditions. For Lasry–Lions monotonicity, the paper proves a notable equivalence: if $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$6, then standard first-order LL monotonicity is equivalent to second-order LL monotonicity, whereas the analogous implication fails for displacement monotonicity. Under regularity and non-degeneracy assumptions, the paper proves local well-posedness of the full master equation, and under separability,

$$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$7

together with Lasry–Lions monotonicity, it proves global existence and uniqueness of a classical solution for any time horizon (Mou et al., 13 Mar 2025).

These results clarify another recurrent misconception: monotonicity is not a merely technical convenience. In the supplied papers it is the mechanism that supports uniqueness of equilibria, propagation of a priori bounds, and global solvability of the infinite-dimensional PDE.

5. Numerical approximation and iterative solvers

Two computational traditions appear in the supplied literature: monotone discretization of mean-field-game PDEs and fixed-point iteration for finite-state NLL systems.

For evolutive mean field games on the two-torus $$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$8 with periodic boundary conditions, Achdou–Camilli–Capuzzo-Dolcetta study the PDE system

$$A\varphi(x)=-\sum_{i,j=1}^n \partial_{x_i x_j}(a_{ij}(x)\varphi(x)),\qquad H(x,m,p)=\inf_{v\in\mathbb R^d}\{f(x,m,v)+p\cdot g(x,m,v)\},$$9

with mixed terminal–initial conditions. Their fully implicit–explicit finite-difference scheme uses a numerical Hamiltonian $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$0 satisfying monotonicity, consistency, $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$1 regularity, and convexity, and a discrete transport term that is the adjoint in $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$2 of the linearization in $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$3 of the Hamiltonian term. For nonlocal smoothing couplings and $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$4, the interpolants $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$5 converge respectively uniformly and strongly in $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$6, and in $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$7; related convergence results are proved for $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$8 and for local couplings $$-\partial_tU(x,m,t)+A\,U(x,m,t) +\int_{\mathbb R^n}\frac{\delta U}{\delta m}(x,m,t;\xi)\, \Bigl(A^*m(\xi)+\nabla_\xi\!\cdot\!\bigl(G(\xi,m,\nabla_xU(\xi,m,t))\,m(\xi)\bigr)\Bigr)\,d\xi =H\bigl(x,m,\nabla_xU(x,m,t)\bigr).$$9 (Achdou et al., 2012).

For finite-state NLL equations, Höfer–Laurière–Soner–Yan propose Picard and weighted Picard iterations. Starting from an initial population control $\delta U/\delta m$0, each iteration solves HJB$\delta U/\delta m$1 backward for $\delta U/\delta m$2, computes the best response

$\delta U/\delta m$3

and updates

$\delta U/\delta m$4

Theorem 3.1 gives uniform convergence to the unique solution of the $\delta U/\delta m$5-NLL system, Proposition 3.2 gives a geometric rate for $\delta U/\delta m$6, and pure Picard exhibits super-exponential decay via factorial estimates. The paper also gives two implementation routes: an ODE-based HJB solver and a Monte Carlo plus empirical risk minimization method using neural nets with soft-plus output for jump rates (Höfer et al., 28 Jul 2025).

The numerical experiments illustrate stability issues that are analytically meaningful. In the two-state synchronization example with thermal noise, direct RK45 applied to the full $\delta U/\delta m$7-NLL system becomes unstable, while Picard–HJB remains stable and correctly locates the discontinuity at $\delta U/\delta m$8. In the botnet-defense example, where the state-space size grows like $\delta U/\delta m$9, only the neural-network Picard scheme is used; its long-run distributions match the mean-field benchmarks reported in the paper (Höfer et al., 28 Jul 2025).

6. Free-boundary price-formation model and the fast-trading limit

In the price-formation usage, Lasry and Lions consider two nonnegative densities,

$L^2$0

together with a moving agreed price $L^2$1. The free-boundary system is

$L^2$2

where the transaction rate is

$L^2$3

and the free boundary is the unique zero of $L^2$4, namely $L^2$5 (Burger et al., 2013).

Burger–Caffarelli–Markowich–Wolfram derive this system as the fast-collision limit of the Boltzmann-type model

$L^2$6

with nonnegative Schwarz initial data. The local collision rate is $L^2$7, and each transaction shifts the buyer price down by $L^2$8 and the seller price up by $L^2$9. Defining $f,g,h,\sigma$0, one shows that $f,g,h,\sigma$1 and $f,g,h,\sigma$2 as Radon measures. Under the well-prepared hypothesis

$f,g,h,\sigma$3

the limiting measure collapses onto a single Dirac line,

$f,g,h,\sigma$4

and the limiting pair $f,g,h,\sigma$5 solves exactly the Lasry–Lions free-boundary system (Burger et al., 2013).

The convergence theorem is strong in the natural parabolic sense: if $f,g,h,\sigma$6 are nonnegative and well prepared, then

$f,g,h,\sigma$7

as $f,g,h,\sigma$8 (Burger et al., 2013). The same paper analyzes the fast initial layer by setting $f,g,h,\sigma$9 and $C^2$0, dropping diffusion at leading order, and proving convergence toward a well-prepared pair $C^2$1 with $C^2$2. Numerically, the collision volume $C^2$3 shrinks to a narrow spike converging to $C^2$4, and a double limit $C^2$5, $C^2$6 with $C^2$7 produces a different hydrodynamic model with internal layers of width $C^2$8 in the Lasry–Lions profile (Burger et al., 2013).

The price-formation model and the master-equation framework are mathematically different objects, but the supplied literature places them under a related Lasry–Lions equilibrium perspective: both describe how a macroscopic equilibrium structure emerges from microscopic interaction, whether through trade collisions concentrating on a free boundary or through Nash consistency in a large population of strategic agents.