Discrete Mean Field Games on Finite Graphs as Initial Value Optimization

Haomin Zhou; Yang Xiang; Yaxin Feng

arxiv: 2604.05685 · v1 · submitted 2026-04-07 · 🧮 math.NA · cs.NA· math.OC

Discrete Mean Field Games on Finite Graphs as Initial Value Optimization

Yaxin Feng , Yang Xiang , Haomin Zhou This is my paper

Pith reviewed 2026-05-10 18:53 UTC · model grok-4.3

classification 🧮 math.NA cs.NAmath.OC

keywords mean field gamesfinite graphsinitial value optimizationHamilton-Jacobi equationcontinuity equationneural networkNash equilibriumpotential games

0 comments

The pith

Potential mean field games on finite graphs reduce to optimizing only the initial condition of the Hamilton-Jacobi equation under coupled ODE constraints.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that potential mean field games on finite graphs, which involve infinite non-cooperative agents optimizing on graph nodes, can be solved by reducing the problem to a finite-dimensional optimization over the initial value of the backward Hamilton-Jacobi equation. This initial condition serves as the sole variable, with the full coupled system of forward continuity and backward Hamilton-Jacobi equations enforced as dynamic constraints through ODE integration. This Graph MFG-IV approach is a reduced-order model that avoids time-discretizing infinite-dimensional paths and shrinks the search space compared to path-wise settings. A neural network then solves this optimization. Readers would care because it offers a way to find Nash equilibria in interactive agent systems on graphs by optimizing far fewer variables.

Core claim

For potential mean field games on finite graphs, the Nash equilibrium is characterized by a coupled ODE system of the discrete forward continuity equation and the discrete backward Hamilton-Jacobi equation. The paper shows this can be recast as an initial value finite-dimensional optimization problem, Graph MFG-IV, in which the initial condition of the Hamilton-Jacobi equation is the unique variable, and the coupled system acts as the ODE integrator constraint. This formulation has a much smaller searching space than the general path-wise problem setting and avoids time-discretization of the infinite-dimensional path.

What carries the argument

The Graph MFG-IV formulation, where the initial condition of the Hamilton-Jacobi equation is optimized as the sole variable while the coupled Hamilton-Jacobi and continuity equations are integrated as the dynamics constraint.

If this is right

The search space reduces from the full path to a single initial condition vector.
Time-discretization of paths is avoided, removing one source of approximation.
Nash equilibria are recovered exactly for potential games via the ODE integrator.
Neural networks can solve the resulting finite-dimensional constrained problem directly.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The reduction may extend to other mean field game types that admit variational or potential formulations.
ODE solvers could improve scaling to graphs with many nodes or states.
The method resembles parameter-fitting in dynamical systems where equilibria are encoded in initial conditions.

Load-bearing premise

The potential structure allows recovering the exact Nash equilibrium by optimizing solely the initial condition of the Hamilton-Jacobi equation while integrating the coupled system, without missing equilibria or introducing errors.

What would settle it

Solving the optimization on a small graph with a known Nash equilibrium and verifying whether the output satisfies the original coupled forward-backward equations or recovers the known solution.

Figures

Figures reproduced from arXiv: 2604.05685 by Haomin Zhou, Yang Xiang, Yaxin Feng.

**Figure 2.** Figure 2: Result of the dynamics of population density [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** Figure 3: Result of the dynamics of population density [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

**Figure 4.** Figure 4: Result of the dynamics of population density [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

**Figure 5.** Figure 5: Result of the dynamics of population density [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: Result of the dynamics of population dynamics [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

**Figure 7.** Figure 7: Result of the dynamics of population dynamics [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗

read the original abstract

In this paper, we propose an initial value fomulation of the discrete mean field games on finite graphs (Graph MFG), and design a neural network based approach to solve it. Graph MFG describes infinite, non-cooperative and interactive homogeneous agents move on node states through the edges to optimize their own goals. Nash Equilibrium of the Graph MFG is characterized by a coupled ordinary differential equations (ODE) system, including the discrete forward continuity equation and the discrete backward Hamilton-Jacobi equation. In this paper, we mainly focus on the potential mean field games (Potential MFG) on finite graphs, which has an infinite-dimensional constrained optimization structure. We reformulate Potential MFG as an initial value finite-dimentional optimization problem with dynamics constrains, names Graph MFG-IV. Specifically, the initial condition of the Hamilton-Jacobi equation is regarded as the unique variable, constrained by the coupled Hamilton-Jacobi and continuity equation system as the ODE integrator. This formulation is a reduced-order model, which avoids time-discretization of the infinite-dimensional path and has a much smaller searching space than the general path-wise problem setting. We design a neural network-based approach to solve the Graph MFG-IV problem.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper reformulates potential MFG on graphs as an initial-value optimization over only the HJ initial condition, which shrinks the search space, but the equivalence to the original BVP is not derived and no validation is shown.

read the letter

The core of this paper is a reformulation that turns potential mean field games on finite graphs into an initial value optimization. They set the initial condition of the backward Hamilton-Jacobi equation as the only free variable and then integrate the coupled forward continuity equation and backward HJ equation using an ODE solver to enforce the dynamics. This is new because it avoids optimizing over the entire time path and instead works with a much smaller finite-dimensional space. The potential structure is key here, as it allows the Nash equilibrium to be found this way without the usual full boundary value setup. The paper does a good job laying out why this reduced-order model has a smaller search space than the general path-wise problem. It also sketches a neural network approach to solve the resulting optimization with dynamics constraints, which could be efficient for larger graphs. The main soft spot is the lack of justification for why this initial-value setup will satisfy the terminal condition of the original system. The standard characterization is a coupled ODE boundary value problem, and while the potential property might help, the abstract gives no derivation showing equivalence or that no equilibria are missed. The stress test note on this point seems to hold based on what's presented. There are no numerical experiments, error bounds, or details on how the neural network handles the integration and optimization. This makes it difficult to judge if the method works in practice or if the ODE integrator introduces significant errors. This kind of work is for specialists in numerical methods for mean field games and optimal control on graphs. Someone building solvers for applied problems in transportation or control might find the reduced dimension useful once the details are filled in. I would recommend sending it to peer review. The idea has enough substance to warrant referee input on the equivalence proof and on adding validation, even if the current draft is preliminary.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes reformulating potential mean field games on finite graphs (Graph MFG) as an initial-value finite-dimensional optimization problem (Graph MFG-IV). The sole decision variable is the initial condition of the backward Hamilton-Jacobi equation; the coupled forward continuity and backward HJ ODE system is treated as a dynamics constraint and integrated exactly by an ODE solver. A neural-network approach is designed to solve the resulting reduced-order problem, which is claimed to avoid time-discretization of infinite-dimensional paths and to possess a much smaller search space than general path-wise formulations.

Significance. If the claimed equivalence between the initial-value optimization and the original Nash-equilibrium BVP holds without approximation error or missing solutions, the reduction would constitute a useful dimension-reduction technique for potential discrete MFGs on graphs, leveraging the variational structure to replace trajectory optimization by a shooting problem over initial data. The neural-network solver could then offer a practical computational pathway. At present, however, the absence of any supporting derivation, convergence theory, or numerical evidence prevents assessment of whether these advantages are realized.

major comments (2)

[Abstract] Abstract: The central reduction asserts that, for potential MFG, optimizing only the initial condition u(0) of the backward HJ equation while integrating the coupled forward-backward ODE system recovers the exact Nash equilibrium. No derivation is supplied showing that the Euler-Lagrange conditions of this initial-value problem are equivalent to those of the original two-point boundary-value problem, nor how the terminal condition u(T,x)=g(x) is automatically satisfied for arbitrary u(0). This equivalence is load-bearing for the entire claim.
[Abstract] Abstract: The manuscript states that the formulation is a 'reduced-order model' that 'avoids time-discretization' and yields a 'much smaller searching space,' yet supplies neither an explicit statement of the objective functional being minimized nor any error analysis for the ODE integrator or the neural-network optimizer. Without these, it is impossible to verify that the computed u(0) indeed yields a consistent equilibrium.

minor comments (1)

[Abstract] Abstract contains typographical errors: 'fomulation' should read 'formulation', 'finite-dimentional' should read 'finite-dimensional', and 'names' should read 'named'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and valuable comments on our manuscript. We address each major comment below and plan to incorporate revisions to strengthen the presentation of the equivalence and the formulation details.

read point-by-point responses

Referee: [Abstract] Abstract: The central reduction asserts that, for potential MFG, optimizing only the initial condition u(0) of the backward HJ equation while integrating the coupled forward-backward ODE system recovers the exact Nash equilibrium. No derivation is supplied showing that the Euler-Lagrange conditions of this initial-value problem are equivalent to those of the original two-point boundary-value problem, nor how the terminal condition u(T,x)=g(x) is automatically satisfied for arbitrary u(0). This equivalence is load-bearing for the entire claim.

Authors: The referee correctly identifies that a detailed derivation of the equivalence is missing from the current manuscript. We will add a dedicated subsection in the revision that derives the first-order optimality conditions for the initial-value problem and demonstrates their equivalence to the original two-point BVP for potential MFGs. Regarding the terminal condition, it is not satisfied for arbitrary u(0); rather, the optimization over u(0) is designed such that the minimizer enforces u(T) = g through the variational structure of the potential game. We will clarify this point explicitly. revision: yes
Referee: [Abstract] Abstract: The manuscript states that the formulation is a 'reduced-order model' that 'avoids time-discretization' and yields a 'much smaller searching space,' yet supplies neither an explicit statement of the objective functional being minimized nor any error analysis for the ODE integrator or the neural-network optimizer. Without these, it is impossible to verify that the computed u(0) indeed yields a consistent equilibrium.

Authors: We agree that the objective functional should be stated explicitly. In the revised manuscript, we will include the precise mathematical expression for the objective in the abstract and introduction, which is the potential energy functional integrated along the trajectories defined by the ODE system. For error analysis, while the theoretical formulation assumes exact integration via the ODE solver, we will add a remark on the numerical errors introduced by the integrator and the neural network approximation, including references to standard error bounds for ODE solvers and neural network optimization in similar contexts. revision: yes

Circularity Check

0 steps flagged

Reformulation of potential MFG as initial-value optimization follows directly from the known coupled ODE characterization without self-referential reduction.

full rationale

The paper starts from the standard fact that potential MFG equilibria satisfy a coupled forward continuity / backward HJ ODE system on the graph. It then treats the initial value of the backward equation as the sole decision variable and enforces the dynamics via an ODE integrator. This is a direct parameterization of the known BVP into a shooting-style constrained optimization; the abstract and description contain no fitted parameters renamed as predictions, no self-citation used to justify the equivalence, and no ansatz smuggled in. The reduction in search space is a mechanical consequence of the variational structure already present in the input characterization, not a circular re-derivation of the same object.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the standard characterization of potential MFG equilibria via coupled forward-backward ODEs and on the assumption that numerical integration of that system is accurate enough to serve as a constraint; no new entities are postulated.

axioms (2)

domain assumption Nash equilibrium of the Graph MFG is characterized by the coupled discrete forward continuity equation and discrete backward Hamilton-Jacobi equation.
This is the starting point invoked for the reformulation into Graph MFG-IV.
standard math Solutions to the coupled ODE system exist and can be integrated numerically with sufficient accuracy.
Required for the initial-value problem with dynamics constraints to be well-posed.

pith-pipeline@v0.9.0 · 5515 in / 1417 out tokens · 49207 ms · 2026-05-10T18:53:48.668766+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We reformulate Potential MFG as an initial value finite-dimensional optimization problem... the initial condition of the Hamilton-Jacobi equation is regarded as the unique variable, constrained by the coupled Hamilton-Jacobi and continuity equation system as the ODE integrator.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat recovery theorems unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Nash equilibrium... satisfies the discrete continuity equation and the discrete HJ equation system

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Achdou, J

[1]Y. Achdou, J. Han, J.-M. Lasry, P.-L. Lions, and B. Moll,Income and wealth distribution in macroeconomics: A continuous-time approach, The review of eco- nomic studies, 89 (2022), pp. 45–86. This manuscript is for review purposes only. 20YAXIN FENG, YANG XIANG, AND HAOMIN ZHOU [2]A. Aurell, R. Carmona, G. Dayanikli, and M. Lauriere,Optimal incentives t...

work page 2022
[2]

Benamou and Y

[4]J.-D. Benamou and Y. Brenier,A computational fluid mechanics solution to the monge-kantorovich mass transfer problem, Numerische Mathematik, 84 (2000), pp. 375–393. [5]T. Cabannes, M. Lauriere, J. Perolat, R. Marinier, S. Girgin, S. Perrin, O. Pietquin, A. M. Bayen, E. Goubault, and R. Elie,Solving n-player dynamic routing games with congestion: a mean...

work page arXiv 2000
[3]

Cardaliaguet and A

[7]P. Cardaliaguet and A. Porretta,An introduction to mean field game theory, in Mean Field Games: Cetraro, Italy 2019, Springer, 2021, pp. 1–158. [8]R. Carmona,Applications of mean field games in financial engineering and economic theory, arXiv preprint arXiv:2012.05237, (2020). [9]R. Carmona and M. Laurière,Convergence analysis of machine learning algor...

work page arXiv 2019
[4]

[12]S.-N. Chow, W. Huang, Y. Li, and H. Zhou,Fokker–planck equations for a free energy functional or markov process on a graph, Archive for Rational Mechanics and Analysis, 203 (2012), pp. 969–1008. [13]S.-N. Chow, W. Li, C. Mou, and H. Zhou,Dynamical schrödinger bridge problems on graphs, Journal of Dynamics and Differential Equations, 34 (2022), pp. 2511–

work page 2012
[5]

[14]S.-N. Chow, W. Li, and H. Zhou,Entropy dissipation of fokker-planck equations on graphs, arXiv preprint arXiv:1701.04841, (2017). [15]S.-N. Chow, W. Li, and H. Zhou,A discrete schrödinger equation via optimal transport on graphs, Journal of Functional Analysis, 276 (2019), pp. 2440–2469. [16]S.-N. Chow, W. Li, and H. Zhou,Wasserstein hamiltonian flows...

work page arXiv 2017
[6]

[21]D. A. Gomes, J. Mohr, and R. R. Souza,Discrete time, finite state space mean field games, Journal de mathématiques pures et appliquées, 93 (2010), pp. 308–328. [22]D. A. Gomes, J. Mohr, and R. R. Souza,Continuous time finite state mean field games, Applied Mathematics & Optimization, 68 (2013), pp. 99–143. [23]J. L. Gross and J. Yellen,Handbook of gra...

work page 2010
[7]

Guéant,From infinity to one: The reduction of some mean field games to a global control problem, arXiv preprint arXiv:1110.3441, (2011)

[24]O. Guéant,From infinity to one: The reduction of some mean field games to a global control problem, arXiv preprint arXiv:1110.3441, (2011). [25]O. Guéant,Existence and uniqueness result for mean field games with congestion effect on graphs, Applied Mathematics & Optimization, 72 (2015), pp. 291–303. [26]O. Guéant, J. Lasry, and P. Lions,Mean field gam...

work page arXiv 2011
[8]

INITIAL VALUE LEARNING BASED MFGS ON GRAPHS21 [27]W

This manuscript is for review purposes only. INITIAL VALUE LEARNING BASED MFGS ON GRAPHS21 [27]W. Hamilton, Z. Ying, and J. Leskovec,Inductive representation learning on large graphs, Advances in neural information processing systems, 30 (2017). [28]J. Han, R. Hu, and J. Long,Learning high-dimensional mckean–vlasov forward- backward stochastic differentia...

work page arXiv 2017
[9]

[36]A. T. Lin, S. W. Fung, W. Li, L. Nurbekyan, and S. J. Osher,Alternating the population and control neural networks to solve high-dimensional stochastic mean-field games, Proceedings of the National Academy of Sciences, 118 (2021), p. e2024713118. [37]G.-H. Liu, T. Chen, O. So, and E. Theodorou,Deep generalized schrödinger bridge, Advances in Neural In...

work page 2021
[10]

[38]Z. Liu, B. Wu, and H. Lin,A mean field game approach to swarming robots control, in 2018 Annual American Control Conference (ACC), IEEE, 2018, pp. 4293–4298. [39]J. Maas,Gradient flows of the entropy for finite markov chains, Journal of Functional Analysis, 261 (2011), pp. 2250–2292. [40]A. Mielke,A gradient structure for reaction–diffusion systems an...

work page 2018
[11]

[41]A. Roy, C. Singh, and Y. Narahari,Recent advances in modeling and control of epidemics using a mean field approach, S¯ adhan¯ a, 48 (2023), p

work page 2023
[12]

Ruthotto, S

[42]L. Ruthotto, S. J. Osher, W. Li, L. Nurbekyan, and S. W. Fung,A machine learning framework for solving high-dimensional mean field game and mean field control problems, Proceedings of the National Academy of Sciences, 117 (2020), pp. 9183–9193. [43]F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monf ardini, The graph neural network model, ...

work page arXiv 2020

[1] [1]

Achdou, J

[1]Y. Achdou, J. Han, J.-M. Lasry, P.-L. Lions, and B. Moll,Income and wealth distribution in macroeconomics: A continuous-time approach, The review of eco- nomic studies, 89 (2022), pp. 45–86. This manuscript is for review purposes only. 20YAXIN FENG, YANG XIANG, AND HAOMIN ZHOU [2]A. Aurell, R. Carmona, G. Dayanikli, and M. Lauriere,Optimal incentives t...

work page 2022

[2] [2]

Benamou and Y

[4]J.-D. Benamou and Y. Brenier,A computational fluid mechanics solution to the monge-kantorovich mass transfer problem, Numerische Mathematik, 84 (2000), pp. 375–393. [5]T. Cabannes, M. Lauriere, J. Perolat, R. Marinier, S. Girgin, S. Perrin, O. Pietquin, A. M. Bayen, E. Goubault, and R. Elie,Solving n-player dynamic routing games with congestion: a mean...

work page arXiv 2000

[3] [3]

Cardaliaguet and A

[7]P. Cardaliaguet and A. Porretta,An introduction to mean field game theory, in Mean Field Games: Cetraro, Italy 2019, Springer, 2021, pp. 1–158. [8]R. Carmona,Applications of mean field games in financial engineering and economic theory, arXiv preprint arXiv:2012.05237, (2020). [9]R. Carmona and M. Laurière,Convergence analysis of machine learning algor...

work page arXiv 2019

[4] [4]

[12]S.-N. Chow, W. Huang, Y. Li, and H. Zhou,Fokker–planck equations for a free energy functional or markov process on a graph, Archive for Rational Mechanics and Analysis, 203 (2012), pp. 969–1008. [13]S.-N. Chow, W. Li, C. Mou, and H. Zhou,Dynamical schrödinger bridge problems on graphs, Journal of Dynamics and Differential Equations, 34 (2022), pp. 2511–

work page 2012

[5] [5]

[14]S.-N. Chow, W. Li, and H. Zhou,Entropy dissipation of fokker-planck equations on graphs, arXiv preprint arXiv:1701.04841, (2017). [15]S.-N. Chow, W. Li, and H. Zhou,A discrete schrödinger equation via optimal transport on graphs, Journal of Functional Analysis, 276 (2019), pp. 2440–2469. [16]S.-N. Chow, W. Li, and H. Zhou,Wasserstein hamiltonian flows...

work page arXiv 2017

[6] [6]

[21]D. A. Gomes, J. Mohr, and R. R. Souza,Discrete time, finite state space mean field games, Journal de mathématiques pures et appliquées, 93 (2010), pp. 308–328. [22]D. A. Gomes, J. Mohr, and R. R. Souza,Continuous time finite state mean field games, Applied Mathematics & Optimization, 68 (2013), pp. 99–143. [23]J. L. Gross and J. Yellen,Handbook of gra...

work page 2010

[7] [7]

Guéant,From infinity to one: The reduction of some mean field games to a global control problem, arXiv preprint arXiv:1110.3441, (2011)

[24]O. Guéant,From infinity to one: The reduction of some mean field games to a global control problem, arXiv preprint arXiv:1110.3441, (2011). [25]O. Guéant,Existence and uniqueness result for mean field games with congestion effect on graphs, Applied Mathematics & Optimization, 72 (2015), pp. 291–303. [26]O. Guéant, J. Lasry, and P. Lions,Mean field gam...

work page arXiv 2011

[8] [8]

INITIAL VALUE LEARNING BASED MFGS ON GRAPHS21 [27]W

This manuscript is for review purposes only. INITIAL VALUE LEARNING BASED MFGS ON GRAPHS21 [27]W. Hamilton, Z. Ying, and J. Leskovec,Inductive representation learning on large graphs, Advances in neural information processing systems, 30 (2017). [28]J. Han, R. Hu, and J. Long,Learning high-dimensional mckean–vlasov forward- backward stochastic differentia...

work page arXiv 2017

[9] [9]

[36]A. T. Lin, S. W. Fung, W. Li, L. Nurbekyan, and S. J. Osher,Alternating the population and control neural networks to solve high-dimensional stochastic mean-field games, Proceedings of the National Academy of Sciences, 118 (2021), p. e2024713118. [37]G.-H. Liu, T. Chen, O. So, and E. Theodorou,Deep generalized schrödinger bridge, Advances in Neural In...

work page 2021

[10] [10]

[38]Z. Liu, B. Wu, and H. Lin,A mean field game approach to swarming robots control, in 2018 Annual American Control Conference (ACC), IEEE, 2018, pp. 4293–4298. [39]J. Maas,Gradient flows of the entropy for finite markov chains, Journal of Functional Analysis, 261 (2011), pp. 2250–2292. [40]A. Mielke,A gradient structure for reaction–diffusion systems an...

work page 2018

[11] [11]

[41]A. Roy, C. Singh, and Y. Narahari,Recent advances in modeling and control of epidemics using a mean field approach, S¯ adhan¯ a, 48 (2023), p

work page 2023

[12] [12]

Ruthotto, S

[42]L. Ruthotto, S. J. Osher, W. Li, L. Nurbekyan, and S. W. Fung,A machine learning framework for solving high-dimensional mean field game and mean field control problems, Proceedings of the National Academy of Sciences, 117 (2020), pp. 9183–9193. [43]F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monf ardini, The graph neural network model, ...

work page arXiv 2020