acceptodds
Under review as a conference paper at ICLR 2027

Performative Mean Field Games

Abstract

This paper investigates performative prediction in large-population games where the number of players is unknown and effectively infinite. In many real-world settings, such as traders adjusting investment strategies or merchants setting prices, any single player's decision has negligible influence on the overall environment, yet the collective decision distribution of all players significantly shapes the data distributions that each player observes and makes decisions on. Existing performative prediction frameworks, which rely on discrete, finite-player models, fail to capture these dynamics, since a change in any single player's discrete decision does not meaningfully alter the population-level decision distribution and thus induces no perceptible shift in the induced data distributions. We address this challenge by introducing the *performative mean field game* framework. In this framework, a continuum of players interacts through an unknown mapping from the population decision distribution to the resulting data distributions. We define the *performatively stable mean-field equilibrium* as the fixed point where the decision distribution is exactly the aggregate of each player's optimal response to the data distribution induced by that same decision distribution, eliminating the need for further decision updates. Our main technical contributions are threefold. First, we formalize the above game and equilibrium concept. Second, we introduce a sensitivity measure in Sobolev norms that quantifies the impact of distribution shifts on individual decision-making and use it to establish theoretical convergence conditions for performatively stable mean-field equilibria. Third, building on these theoretical results, we design an adaptive algorithm that estimates sensitivity on the fly and provably converges to a performatively stable mean-field equilibrium without requiring prior knowledge of the underlying decision-to-data mapping. Experiments on quantitative trading, recommendation system, and ride-sharing market validate our theoretical results and demonstrate the effectiveness of the proposed algorithm.

Primary area:
optimization
Keywords:
Performative Prediction, Data Distribution Shift, Mean-Field Game

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZJFoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a new notion of mean field game where the cost of the agents depends on a provided state-decision probability measure and on the state-decision probability measure of the population of agents. It defines the notion of a stable equilibrium, where the provided state-decision probability measure is equal to its induced mean field game probability measure. The remainder of the paper consists of establishing assumptions under which such an equilibrium exists and can be computed via an algorithm. Lastly, one numerical experiment is performed on three home-made environments.

Critical strengths

Strength 1: Introducing distributional shift in mean field games is a natural and timely contribution.

Strength 2: The paper tries to be complete in its presentation. It introduces a new definition of mean field game, establishes assumptions under which an equilibrium exists, and provides an algorithm that is tested numerically. It also provides in appendix proofs for every statement.

Critical weaknesses

Weakness 1: The objective of the paper is to define a new notion of equilibrium, called 'Stable Performative Equilibrium'. Yet, I have the impression that this notion of equilibrium is simply a classical Nash equilibrium in an extended mean field game (a mean field game where the dynamics and reward depend on the joint state-action probability measure of the population of agents). Define the mean field game M=(Θ,X,c)\mathcal{M} = (\Theta, \mathcal{X}, c) where Θ\Theta is the state space, X\mathcal{X} is the action space, and c:Θ×X×P(Θ×X)→Rc: \Theta \times \mathcal{X} \times \mathcal{P}(\Theta\times\mathcal{X}) \to \mathbb{R} is the cost. Using the notation of the paper, let c(θ,x,m)=L(θ)(m,m)(x)c(\theta, x, m) = \mathcal{L}(\theta)(m, m)(x) for any (θ,x,m)∈Θ×X×P(Θ×X)(\theta, x, m) \in \Theta \times \mathcal{X} \times \mathcal{P}(\Theta\times\mathcal{X}). Let m∈P(Θ×X)m \in \mathcal{P}(\Theta\times\mathcal{X}). We call a policy a measurable mapping π\pi from Θ\Theta to ΔX\Delta_{\mathcal{X}}. We call the set of best responses with respect to any m∈P(Θ×X)m \in \mathcal{P}(\Theta\times\mathcal{X}) the set of policies minimizing V(π;m)=E(c(θ~,X,m))V(\pi; m) = \mathbb{E}(c(\tilde \theta, X, m)), where θ~\tilde \theta has probability distribution m(⋅×X)m(\cdot \times \mathcal{X}) and XX is a random variable distributed according to π(⋅∣θ~)\pi(\cdot\mid\tilde \theta), θ~−a.s.\tilde{\theta}-a.s.. Let us also introduce mπ,μ(dθ,dx)=μ(dθ)π(dx∣θ)m^{\pi, \mu}(d\theta,dx) = \mu(d\theta)\pi(dx\mid \theta) for any initial state distribution μ∈PΘ\mu \in \mathcal{P}_{\Theta}. We call a Nash equilibrium for the extended mean field game and initialization μ\mu any policy π\pi such that π\pi is a best response to mπ,μm^{\pi, \mu}.

My claim is the following: if π\pi is a Nash equilibrium for the extended mean field game above, with respect to some initialization μ\mu, then mπ,μm^{\pi, \mu} is a stable equilibrium in the sense you define. Conversely, any stable Nash equilibrium mm, as you define it, induces a Nash equilibrium for the extended mean field game for some initial distribution.

If I understood correctly and I am right, your notion of stable equilibrium does not define anything new. Extended mean field games have already been studied in the literature, and you should clearly state what is different in your paper, and what has been newly introduced and proven.

Weakness 2: I am not convinced that the numerical experiment proves that the algorithm converges toward a stable Nash equilibrium. Why would showing that the averaged payoff converges prove that the algorithm converges to 1) a stable Nash equilibrium under the real loss, or 2) a stable Nash equilibrium under the regularized loss? You do not report any values, such as exploitability, that would show that a deviating agent cannot obtain a better loss than the one obtained under the state-decision probability measure output by the algorithm.

Weakness 3: Definitions 2 is central to this paper. Since I read it as a theoretically oriented paper, I believe the definition would benefit from additional care and precision. In their current form, they left me with several questions, which I detail in the Questions section below.

Weakness 4: From my point of view, the second contribution of the paper, "Equilibrium Convergence Analysis", is too similar to existing results on mean field games to be a real contribution for a theoretically oriented paper. Also, the third claimed contribution, "This algorithm ensures convergence to a informatively stable mean-field equilibrium by enforcing the convergence conditions to be satisfied during a repeated performative mean field game.", is a bit misleading because the algorithm does not converge toward the stable equilibrium of the original game, but toward that of the regularized game.

Questions for authors

Question 1-2: Could you discuss weakness 1,2 ?

Question 3: In definition 2, could you clearly state how m0m^0 is taken ? Is it any Nash equilibrium ? Is it a Nash equilibrium for some fixed mm ? But then, I have the impression the set MRPMFEM_{RPMFE} should also depend on the initialization. You are taking the closure of a set. Under which topology ? You do not discuss why, could you ? What happen if MFE(mt)MFE(m^t) is empty ?

7 comments

We are sincerely grateful for the reviewer's commitment of valuable time in carefully evaluating our manuscript.

Below, we respond to the reviewer's valuable feedback. All revisions made in the manuscript are highlighted in blue for easy identification.


Regarding the relation to extended mean-field games. We thank the reviewer for the detailed reading and for the opportunity to clarify the relation to extended mean-field games.

Setting. We clarify that the data distribution map Dθ\mathcal{D} _\theta in c(θ,x,m)=Ez∼Dθ(m)[ℓ(x,z,θ,m)]c(\theta,x,m)=\mathbb E _{z\sim\mathcal D _\theta(m)}[\ell(x,z,\theta,m)] is unknown in our setting, although the construction with cost c(θ,x,m)=Lθ(m,m)(x)c(\theta,x,m)=\mathcal L _\theta(m,m)(x) is mathematically correct.

As a result, V(π;m)=E[c(θ~,X~,m)]V(\pi;m)=\mathbb E[c(\tilde\theta,\tilde X,m)] cannot be minimized directly, whereas existing extended-MFG methods minimize V(π;m)V(\pi;m) directly.

The novelty. We clarify that the main difference lies in the learning problem, where the data distribution maps Dθ\mathcal{D} _\theta are unknown, so the cost c(θ,x,m)c(\theta,x,m) cannot be evaluated directly.

We compare our setting with existing mean-field game literature:

Our settingExisting mean-field game
Cost c(θ,x,m)c(\theta,x,m)not evaluableknown
Data distribution map Dθ\mathcal D _\thetaunknownnot applicable
Equilibriumobtained by the convergence of the repeated gamesolved directly

Due to these differences, we introduce the following to solve the problem:

  1. the ε\varepsilon-sensitivity in H−sH ^{-s}, which quantifies the performative effect induced by the decision distribution;
  2. the convergence condition ε<α\varepsilon<\alpha for the repeated game;
  3. an adaptive algorithm that estimates ε\varepsilon and tunes the regularization to meet this condition.

Revision.

  • We have clarified that the data distribution maps Dθ\mathcal{D} _\theta are unknown in Section 1.2 (lines 147-148, 151-153).

Regarding experimental evidence. We thank the reviewer for the comment and for the chance to strengthen our experimental evidence.

Regularized equilibrium. The algorithm targets the regularized game and converges to its performatively stable mean-field equilibrium.

Additional experimental results. We conducted additional experiments to verify that our algorithm converges to a performatively stable mean-field equilibrium, which are detailed as follows.

  • Exploitability. In Appendix J.4, we report the exploitability, i.e., the gap between the loss of the deployed decision and the lowest loss that a deviating agent can obtain given the decision distribution m(t)m ^{(t)} at each iteration:

    Eθ∼μ[EZθ∼Dθ(m(t))[ℓ(xθ⋆,Zθ,θ,m(t))+γC(xθ⋆,θ,m(t))]]−Eθ∼μ[min⁡x∈X(EZθ∼Dθ(m(t))[ℓ(x,Zθ,θ,m(t))+γC(x,θ,m(t))])],\mathbb{E} _{\theta\sim\mu}\Big[\mathbb{E} _{Z _\theta\sim\mathcal{D} _\theta(m ^{(t)})}\big[\ell(x _\theta ^{\star},Z _\theta,\theta,m ^{(t)})+\gamma C(x _\theta ^{\star},\theta,m ^{(t)})\big]\Big]-\mathbb{E} _{\theta\sim\mu}\Big[\min _{x\in\mathcal{X}}\left(\mathbb{E} _{Z _\theta\sim\mathcal{D} _\theta(m ^{(t)})}\big[\ell(x,Z _\theta,\theta,m ^{(t)})+\gamma C(x,\theta,m ^{(t)})\big]\right)\Big], where xθ⋆x _\theta ^{\star} is the deployed decision of a player with type θ\theta under Zθ∼Dθ(m(t))Z _\theta\sim\mathcal{D} _\theta(m ^{(t)}) given decision distribution m(t)m ^{(t)}.

    Figs. 12–14 in Appendix J.4 show that the exploitability approaches 00 for our method with b=1.1,10b=1.1,10 within 33 iterations, indicating that our algorithm converges to the mean-field equilibrium for the regularized game.

  • Equilibrium residual. In Appendix J.5, we further report the equilibrium residual ∥m⋆(t)−m(t)∥H−s\|m ^{\star(t)}-m ^{(t)}\| _{H ^{-s}}, where m⋆(t)m ^{\star(t)} is the mean-field equilibrium under the data distribution induced by m(t)m ^{(t)}.

    Figs. 15–17 in Appendix J.5 show that the equilibrium residual approaches 00 for our method with b=1.1,10b=1.1,10 within 33 iterations, showing that our algorithm converges to a mean-field equilibrium of the game it induces.

Revision.

  • We have provided additional experimental results of exploitability and equilibrium residual in Appendices J.4 and J.5 (lines 2024-2105).

Regarding the repeated game definition. We thank the reviewer for the careful observation and the chance to clarify these points.

m(0)m ^{(0)} in MRPMFG\mathcal{M} _{RPMFG}.

  • In the experiments, m(0)m ^{(0)} is a type-decision distribution with zero average decision. With this m(0)m ^{(0)}, the induced data Zθ∼Dθ(m(0))Z _\theta\sim\mathcal{D} _\theta(m ^{(0)}) is the raw data from the dataset.
  • m(0)m ^{(0)} can be any type-decision distribution. m(0)m ^{(0)} is not required to be a Nash equilibrium, nor a Nash equilibrium for some fixed mm.

Dependence on initialization. Yes. MRPMFG\mathcal{M} _{RPMFG} is the closure of the sequence {m(t)}\{m ^{(t)}\}, which starts at m(0)m ^{(0)}. Thus, different initializations produce different sets MRPMFG\mathcal{M} _{RPMFG}. Specifically, MRPMFG\mathcal{M} _{RPMFG} is determined by:

  • m(0)m ^{(0)}, the initial type-decision distribution,
  • the game structure, determining the strong monotonicity parameter α\alpha,
  • the data distribution maps, determining the sensitivity parameter ε\varepsilon.

Once MRPMFG\mathcal{M} _{RPMFG} is determined, our theoretical results hold.

Topology of MRPMFG\mathcal{M} _{RPMFG}. The closure in the repeated game definition is taken in the weak$ ^{\star}$ topology of Pμ(Θ×X)\mathcal{P} _\mu(\Theta\times\mathcal{X}).

  • In this topology, the expected loss EZθ∼Dθ(mˉ)[ℓ(x,Zθ,θ,m)]\mathbb{E} _{Z _\theta\sim\mathcal{D} _\theta(\bar{m})}\big[\ell(x,Z _\theta,\theta,m)\big] is continuous in mm for bounded continuous ℓ\ell. By Berge's Maximum Theorem, the best-response correspondence is continuous with respect to mm.
  • In this topology, the type-decision distribution set Pμ(Θ×X)\mathcal{P} _\mu(\Theta\times\mathcal{X}) is compact in the weak$ ^{\star}$ topology, which ensures that the equilibrium set MRPMFG\mathcal{M} _{RPMFG} is compact in the weak$ ^{\star}$ topology.

Emptiness of MFE(m(t))MFE(m ^{(t)}). We clarify that MFE(m(t))MFE(m ^{(t)}) cannot be empty under our conditions. By Lemma 1, under α\alpha-strong monotonicity and joint continuity of loss function ℓ\ell, MFE(m)MFE(m) exists and is unique for every m∈MRPMFGm\in\mathcal{M} _{RPMFG}.

Revision.

  • We have clarified the weak$ ^{\star}$ topology in which MRPMFG\mathcal{M} _{RPMFG} is closed in Section 2.2 (lines 207-208).

Regarding our contributions. We thank the reviewer for the comment and the opportunity to clarify our contributions.

Contribution 2. We emphasize that the convergence analysis differs from existing results on mean field games.

  • Setting. The data distribution maps Dθ\mathcal{D} _\theta are unknown, so the loss function cannot be minimized directly. A detailed comparison with existing MFG methods is in our reply to the first point.

  • Iteration dynamics. The iteration dynamics in our work differ from those in existing MFG work, as detailed below.

Our workExisting results
Each iterationEquilibrium m(t+1)m ^{(t+1)} under the data distribution Dθ(m(t))\mathcal{D} _\theta(m ^{(t)})Best response m(t+1)m ^{(t+1)} to the previous type-decision distribution m(t)m ^{(t)}
Lipschitz constantEZθ∼Dθ(mˉ)[ℓ(x,Zθ,θ,m)]\mathbb{E} _{Z _\theta\sim\mathcal{D} _\theta(\bar{m})}\big[\ell(x,Z _\theta,\theta,m)\big] with respect to mˉ\bar{m}ℓ(x,θ,m)\ell(x,\theta,m) with respect to mm

Contribution 3. We clarify that our algorithm converges to the equilibrium of a regularized game, as stated in Section 4.

We agree that the description of contribution 3 in Section 1.1 could be clearer, and we have updated this point.

Revision.

  • We have clarified that our algorithm converges to the equilibrium of a regularized game in Section 1.1 (line 116).

I thank the authors for their answers and the additional experiments. However, many of my concerns remain.

1. Relation to extended MFG.

Firstly, the rebuttal agrees with the link with extended MFG (by the way, I think you must explicitly give the mathematical link as I made, in the appendix, for clarity). The claimed novelty is that Dθ\mathcal{D}_\theta is unknown. To my understanding, this is still not a novelty. It seems that it is possible to write this problem under a standard learning model-free mean field games. Guo et al, 2019 already study MFGs with unknown rewards and dynamics. More precisely, the performative MFG looks to be a special case of the general MFG (GMFG) in which the reward depends on the joint population state-action distribution and is unknown a priori. One can compltetly write the performative MFG as a GMFG, and apply Guo et al, 2019 results. Therefore, my concern remains.

Secondly, the paper's conditions make the extended game strongly monotone. Let $F(m) := \mathcal{L}_ {\theta}(m, m) $. The convergence condition α>ε\alpha > \varepsilon is a sufficient condition for (α−ε)(\alpha - \varepsilon)-strong Lasry–Lions monotonicity of the extended game. Uniqueness within MRPMFG then follows from the classical monotonicity argument. Moreover, fictitious play (Perrin et al., 2020) only require the cost x↦cθ(x,mt)x \mapsto c_\theta(x, m_t) at the current population mtm_t, which can be estimated from samples of Dθ(mt)\mathcal{D}_\theta(m_t) since ℓ\ell is known. Their convergence guarantees are stated for exact cost evaluation, which is the setting of Theorem 2 of the paper; a stochastic version would correspond to Theorem 3, for which Guo et al., 2019 provides a sample-based guarantee.

2. Numerical Experiments.

The reported exploitability includes γC\gamma C, with a different γ\gamma for each method (γ=0\gamma = 0 for b=0b = 0, γ=1\gamma = 1 for fixed regularization, data-driven for b=1.1b = 1.1 and b=10b = 10). Each curve therefore measures the distance to the equilibrium of a different game, and the curves are not comparable. More importantly, my question was whether a deviating agent can improve its loss in the game defined in the paper, i.e. under ℓ\ell, not under ℓ+γC\ell + \gamma C. I do not think your new experiments answer my question.

The main text still reports average payoff, which is not an equilibrium criterion: a higher payoff does not indicate an equilibrium. It is not evidence of convergence to an equilibrium and should not be the main experiment

3. Definitions.

The rebuttal confirms that $\mathcal{M} _ {RPMFG} $ depends on m(0)m^{(0)}. In Algorithm 1 it also depends on the sequence of γ\gamma, since the trajectory, and hence the set, changes with γ\gamma. This dependence should appear in the notation, e.g. MRPMFG(m(0),γ)\mathcal{M}_{RPMFG}(m^{(0)}, \gamma).

The uniqueness argument (Eqs. 31–32) invokes α\alpha-strong monotonicity for two arbitrary equilibria m1,m2∈Pμ(Θ×X)m_ 1, m_ 2 \in \mathcal{P}_ \mu(\Theta \times \mathcal{X}). However, strong monotonicity is only assumed on MRPMFG\mathcal{M}_{RPMFG}, which is itself defined through the (assumed unique) sequence of equilibria. As written, uniqueness is not established. The rebuttal's answer to the emptiness question relies on this lemma.

Summary.

I appreciate the clarification that Algorithm 1 converges to the stable equilibrium of a regularized game. But the regularized equilibrium can differ substantially from the original one, and its distance to the original equilibrium is neither bounded theoretically nor measured empirically. By contrast, Guo et al., 2019 has a convergence guarantee to an ϵ\epsilon-Nash equilibrium of the original game under the same type of oracle. In my view, this limits the significance of contribution 3. I also still consider contribution 2 to be close to known results on monotone mean field games.

References

  • X. Guo, A. Hu, R. Xu, and J. Zhang. Learning mean-field games. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
  • S. Perrin, J. Pérolat, M. Laurière, M. Geist, R. Élie, and O. Pietquin. Fictitious play for mean field games: continuous time analysis and applications. In Advances in Neural Information Processing Systems (NeurIPS), 2020.
Reviewer_TUPXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies games where individual players have little influence, but their collective decisions change the data everyone observes. It introduces a population-level performative prediction framework, gives conditions under which repeated equilibrium updates converge, and extends the analysis to finite samples. It also proposes an adaptive regularization algorithm and evaluates it in trading, recommendation, and ride-sharing examples.

Critical strengths
  • The population-level perspective addresses a meaningful problem. The distinction between negligible individual influence and substantial collective influence is useful for modeling large markets. Bringing this feedback into a mean-field framework is a worthwhile research direction.

  • The conditional convergence result gives a clear takeaway. Repeated updates converge when the game’s stabilizing effect, measured by strong monotonicity, outweighs the effect of changing data distributions. The finite-sample bound also explains how estimation errors affect the final accuracy. These are useful results even independently of the proposed adaptive algorithm.

Critical weaknesses
  • The adaptive algorithm’s convergence claim is not sufficiently established. Appendix F argues that the sensitivity estimate will increase until the convergence condition holds, but does not prove that the observed ratios provide sufficient control over later updates. It also treats a converging regularization coefficient as though the game had already become fixed. These steps need a separate argument, especially with sampling noise. This affects a central claimed contribution.

  • The novelty relative to existing performative games needs a more precise explanation. The introduction claims that finite-player models cannot represent collective distribution feedback. This seems too strong: a joint decision vector also determines an empirical decision distribution, which could feed into the data map. The continuum formulation may offer important advantages, but this observation leaves the claimed modeling separation unclear, and hence part of the novelty argument unclear.

Questions for authors
  • What ensures that Algorithm 1’s sensitivity estimate becomes sufficient for convergence? Can the argument explicitly account for the changing regularization coefficient and estimation noise, rather than applying a fixed-game theorem?

  • Why cannot the finite-player data map discussed in the introduction depend on the empirical distribution of players’ decisions? What is the main new technical result enabled by the continuum formulation?

  • Section 4 states that the proposed conditional-mean congestion term is 1-strongly monotone. Since different distributions can share the same conditional means, what restrictions on the equilibrium distribution family make this statement valid beyond the specific examples?

2 comments

We thank the reviewer for the careful reading and for the opportunity to respond to the points raised.

Below, we respond to the reviewer's valuable feedback. All revisions made in the manuscript are highlighted in blue for easy identification.


Regarding the convergence of the adaptive algorithm (Question 1). We thank the reviewer for these careful observations.

The algorithm as an implementation. We clarify that our algorithm is an adaptive implementation based on the convergence condition of Theorem 3 to set the regularization strength. The algorithm estimates ε\varepsilon based on Definition 3 and tunes the regularization parameter γ\gamma to meet the condition α=γ+ψ>ε\alpha=\gamma+\psi>\varepsilon for the estimated ε\varepsilon. We do not claim a fully rigorous end-to-end convergence proof that simultaneously accounts for sampling noise in the sensitivity estimate and the time-varying regularization coefficient; the algorithm is designed so that the conditions of the theorem are eventually satisfied.

The sensitivity estimate. If the estimate ε\varepsilon is small, the regularization is insufficient and the players' decision distribution fluctuates. In this case, the running maximum in our algorithm tends to increase the estimate ε\varepsilon when the convergence is not achieved. Once the regularization is sufficiently large, the decisions converge, and the estimate ε\varepsilon stabilizes.

The regularized game. Because ε\varepsilon is maintained as a running maximum, the regularization parameter γ=max⁡(0,bε−ψ)\gamma=\max(0,b \varepsilon-\psi) is non-decreasing. Once γ\gamma is large enough for the decisions to stabilize, the regularized losses become fixed over time. In this case, the algorithm aims to achieve a performatively stable mean-field equilibrium of this fixed regularized game.

Revision.

  • We have clarified that the algorithm is an implementation based on Theorem 3 in Section 4 (line 348).
  • We have clarified that the regularized game in our algorithm refers to the regularized game with the stabilized γ\gamma in Section 4 (line 349).

Regarding the novelty relative to existing performative games (Question 2). We thank the reviewer for the chance to clarify this point.

Key challenges in existing performative games. In existing performative games, the decisions are formulated as a vector x1,…,xnx _1,\dots,x _n. We model players as a continuum and the decision as a distribution mm, which avoids the following obstacles:

  • Unknown number of players. When the number of players nn is unknown, the finite-player data distribution maps over joint decisions (Di(x1,…,xn)\mathcal{D} _i(x _1,\dots,x _n)) or over the empirical decision distribution (Di(1n∑i=1nδxi)\mathcal{D} _i(\frac{1}{n}\sum _{i=1} ^n\delta _{x _i})) cannot be constructed, so the performative game cannot be modeled correctly.
  • Large number of players. When the number of players nn is large, the complexity of equilibrium computation increases exponentially, and computing an equilibrium becomes intractable.

Main new technical results. The continuum formulation requires us to analyze the decision distribution under the performative effect, yielding the following results.

Continuum formulationTechnical result
define ε\varepsilon-sensitivity in the Sobolev spacequantify the performative effect induced by arbitrary decision distributions (Dirac, softmax, Gaussian)
analyze the dynamics of the type-decision distribution in the negative-order Sobolev normestablish convergence to a performatively stable mean-field equilibrium (Theorems 2 and 3)

Regarding the 1-strongly monotone term C(x,θ,m)C(x,\theta,m). We thank the reviewer for this careful observation. We acknowledge that considering only the conditional mean for the term C(x,θ,m)C(x,\theta,m) may not be enough for complex game structures.

We have now clarified that the term C(x,θ,m)C(x,\theta,m) is required to be 1-strongly monotone on MRPMFG\mathcal{M} _{RPMFG}, which contains the decision distributions that can arise as equilibria of the repeated game and their closure. With this design, we only need to construct the term C(x,θ,m)C(x,\theta,m) based on the game structure for the given Sobolev order ss.

For example, we can construct the strongly monotone term C(x,θ,m)C(x,\theta,m) given that the equilibria m(⋅∣θ)m(\cdot|\theta) during the repeated game are the following distributions:

  • Dirac measure δx⋆(θ)\delta _{x ^\star(\theta)}.

    C(x,θ,m)=x⋅x⋆(θ).C(x,\theta,m)=x\cdot x ^\star(\theta).

  • Dirac measure δx⋆(θ)\delta _{x ^\star(\theta)} and Gaussian distribution N(x⋆(θ),σ2)\mathcal{N}(x ^\star(\theta),\sigma ^2) that may share the same conditional mean.

    C(x,θ,m)=x∫Xx m(dx∣θ)+x2∫Xx2 m(dx∣θ).C(x,\theta,m)=x\int _{\mathcal X}x\,m(dx|\theta)+x ^2 \int _{\mathcal X}x ^2\,m(dx|\theta).

  • Arbitrary distributions.

    C(x,θ,m)=∫Θ×X[1(2π)dΘ×X∫RdΘ×X(1+∥ξ∥22+∥η∥22)−sei(ξ(θ−θ′)+η(x−x′)) dξdη]m(dθ′,dx′).C(x,\theta,m)=\int _{\Theta\times\mathcal X}\left[\frac{1}{(2\pi) ^{d _{\Theta\times\mathcal X}}}\int _{\mathbb R ^{d _{\Theta\times\mathcal X}}}\big(1+\|\xi\| ^2 _2+\|\eta\| ^2 _2\big) ^{-s}e ^{i(\xi(\theta-\theta')+\eta(x-x'))}\,d\xi d\eta\right]m(d\theta',dx').

Revision.

  • We have clarified that the term CC is required to be 1-strongly monotone on MRPMFG\mathcal{M} _{RPMFG} in Section 4 (lines 352-353).
Reviewer_kauhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies a special class of mean field games (MFG) with expected utility (EU) agents over population-dependent distributions. The paper formalizes the game and an equilibrium concept, and proposes an adaptive algorithm to calculate the equilibrium under certain assumptions.

Critical strengths
  • The paper is clear and easy to follow.
  • The problem is well-motivated.
  • The project is complete, with game formulation, equilibrium analysis, algorithm design, and empirical validation.
Critical weaknesses

My biggest concern regarding the paper is its significance. Based on my reading, the proposed game is a mean-field game (MFG) instantiated with expected utility agents. Therefore, the existing theory on general MFGs applies, and I do not see many novel or significant results for this specific class.

Specifically, MFGs are by definition "performative." Claims such as "a player’s current decision may no longer be optimal under the induced data distribution [...], necessitating further decision updates" and "Performative stability requires that a decision (or a joint decision of all players) is optimal (or an equilibrium) under the data distribution(s) it induces" are strictly covered by the definition of an MFG where the data distribution is determined by the population distribution. Formally, Eq. (2) is an MFG with the utility function: U(x,θ,m)=−EZθ∼Dθ(m)ℓ(x,Zθ,θ,m).U(x,\theta,m) = -\mathbb E_{Z_\theta\sim \mathcal D_\theta(m)} \ell (x, Z_\theta, \theta, m).

Furthermore, I note the following direct correspondences: Definition 1 is simply fixed-point iteration in an MFG; Definition 2 is the standard mean-field equilibrium; Definition 3 is the Lipschitzness of the utility function with respect to the mean field; Definition 4 is the standard monotonicity definition; and Theorem 1 is the common contraction property under monotonicity and Lipschitzness.

The related work discussion does not clearly distinguish the proposed framework from standard MFGs---what are not covered by general MFG methods and what are unique under the proposed formulation.

In summary, standard MFGs model agents' utility as dependent on the population (the mean field). While the authors introduce an intermediate data distribution that is determined by the population and influences utility, there does not appear to be anything fundamentally distinct about this formulation, nor does it yield novel theoretical results.

Questions for authors

See "Critical Weaknesses"

3 comments

We thank the reviewer for the careful reading of our manuscript and for the opportunity to clarify the significance of our work.

Below, we respond to the reviewer's valuable feedback. All revisions made in the manuscript are highlighted in blue for easy identification.


Regarding the problem setting. While we agree with the reviewer that the utility can be mathematically written as U(x,θ,m)=−EZθ∼Dθ(m)[ℓ(x,Zθ,θ,m)]U(x,\theta,m)=-\mathbb{E} _{Z _\theta\sim\mathcal{D} _\theta(m)}\big[\ell(x,Z _\theta,\theta,m)\big], we clarify that our novelty lies in solving this problem under unknown data distribution shifts.

  • Unknown Dθ\mathcal{D} _\theta: The main difficulty is that the distribution maps Dθ\mathcal{D} _\theta are unknown, so U(x,θ,m)U(x,\theta,m) cannot be computed directly.

  • Non-equilibrium with fixed distributions. If we solve the mean field game under a fixed data distribution, with the loss E[ℓ(x,Zθ,θ,m)]\mathbb{E}\big[\ell(x,Z _\theta,\theta,m)\big], each player's decision may no longer be optimal under the data distributions induced by the decision distribution.


Regarding the definitions. We clarify that Definitions 1 and 3 are different from those in existing mean-field game research.

  • Definition 1. It is not the traditional best-response fixed-point iteration in MFG, where players make the best response to the previous deployed decision distribution. In our definition, at each step, players form a type-decision distribution m(t+1)m ^{(t+1)} that is a mean-field equilibrium under the data distribution Dθ(m(t))\mathcal{D} _\theta(m ^{(t)}), meaning that each player makes their best response to m(t+1)m ^{(t+1)} rather than to the deployed decision distribution m(t)m ^{(t)}.

  • Definition 3. It is not the Lipschitz constant of the utility function with respect to the mean field. In our notation Lθ(mˉ,m)\mathcal{L} _\theta(\bar{m},m), the first argument mˉ\bar{m} is the mean field that induces the data distribution Dθ(mˉ)\mathcal{D} _\theta(\bar{m}). Definition 3 bounds the sensitivity of the loss with respect to mˉ\bar{m} under a fixed mean-field mm. In this work, we use ε\varepsilon to quantify that sensitivity, thus avoiding the need to explicitly estimate the unknown and arbitrarily complex distribution maps Dθ\mathcal{D} _\theta.


Regarding the related work. We thank the reviewer for the comment and for the opportunity to clarify the relation to related work.

The key difference is that we study the convergence conditions for an equilibrium in a repeated game, and we use a scalar ε\varepsilon that measures the sensitivity of the dynamics. The detailed comparison is as follows:

Our workExisting model-free MFG learning, e.g., [1-2]
Condition to estimateA scalar ε\varepsilon: The Lipschitz sensitivity of the loss with respect to the decision distribution that induces the data distributions.A value or QQ-function: An approximation of the solution over population distributions.
ApproachThe repeated game: A mean-field equilibrium under the current data distributions at each iteration.Explicit exploration: Decisions that are not best responses.
ApplicabilityUnder Definitions 3 and 4, the equilibrium can be achieved by estimating a scalar ε\varepsilon.Relies on the capacity of function approximation, which becomes challenging when the transition dynamics are complex.

Revision.

  • We have discussed the related work on mean field game in Section 1.2 (lines 144-147).
  • We have clarified the definition of the repeated performative mean field game (Definition 1) in Section 2.2 (lines 201-203).

References

[1] Berkay Anahtarci, Can Deha Kariksiz, and Naci Saldi. Q-learning in regularized mean-field games. Dynamic Games and Applications, 13:89–117, 2023.

[2] Chenyu Zhang, Xu Chen, and Xuan Di. Stochastic semi-gradient descent for learning mean field games with population-aware function approximation. In ICLR, 2025.

I thanks the authors for their rebuttal. I share the same concerns as Reviewer ZJFo. Although the rebuttal clarifies some details, my major concern remains.

We thank the reviewer for carefully reading our rebuttal and for the response.

We have responded to these concerns in our reply to Reviewer ZJFo.

Specifically, in our work, the game dynamics (the repeated game) and the condition to be estimated (a scalar ε\varepsilon) differ from those of model-free mean-field games.

We thank the reviewer for the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.