acceptodds
Under review as a conference paper at ICLR 2027

Multiplayer Performative Prediction Games with Incomplete Information

Abstract

This paper investigates multiplayer performative prediction games in settings with incomplete information. We formalize these games as *performative Bayesian games*, in which players strategically choose their decisions (or predictions) that jointly shape the resulting data distributions, while each player possesses private information unknown to others. A natural adaptive strategy for players is to iteratively (1) optimize their decisions based on current posterior beliefs about others' private information under the data distribution induced by the joint decision, and (2) update those beliefs upon observing others' realized decisions. We introduce the concept of a performatively stable Bayesian equilibrium as a desirable outcome: a decision profile and associated belief system such that the deployed decisions are mutually optimal given the data distributions they induce and the converged beliefs, eliminating the need for further update of either decisions or beliefs. Our main contributions are threefold. First, we provide a formal game-theoretic definition of performative Bayesian games. Second, we derive theoretical conditions under which the natural alternating best-response-and-belief-update dynamics converge to a performatively stable Bayesian equilibrium. Third, leveraging these theoretical results, we design an algorithm that guarantees convergence to such an equilibrium and minimizes the maximum deviation from the original equilibrium when the convergence conditions are violated. Experimental evaluations on crude oil trading 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, Bayesian 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_ccRvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies multiplayer performative prediction games with incomplete information, where players have private information and their joint decisions affect the data distributions. The authors formulate the problem as a performative Bayesian game and introduce the notion of a performatively stable Bayesian equilibrium. They analyze the convergence of alternating decision optimization and belief updates under strong monotonicity and distribution sensitivity conditions. Based on these results, they propose an adaptive regularization algorithm and evaluate it in crude oil trading and ride-sharing market simulations.

Critical strengths
  1. The paper addresses an interesting extension of multiplayer performative prediction. Allowing players to have private information makes the setting more realistic and introduces additional challenges beyond the complete-information case.
  2. The combination of performative prediction and Bayesian games is well motivated. I particularly like the idea of studying decision updates and belief updates together, rather than treating the data distribution shift and information uncertainty separately.
  3. The theoretical analysis is a valuable part of the paper. The authors derive a sensitivity-based convergence condition and attempt to extend the analysis to finite-sample settings. This provides a useful starting point for understanding the interaction between performativity and incomplete information.
  4. The proposed regularization approach is relatively simple and intuitive. Adjusting the regularization strength according to distribution sensitivity is a reasonable way to stabilize the decision dynamics.
  5. The experiments cover two different market settings and several levels of distribution sensitivity. The results are encouraging, and the authors also provide detailed proofs and experimental settings in the appendix.
Critical weaknesses
  1. Bayesian equilibrium formulation. I found the equilibrium definition somewhat difficult to follow. In a Bayesian game, players generally choose strategies that depend on their private types, but Equations (2)–(5) mainly describe a fixed joint decision vector. The appendix later uses type-dependent decisions. It would help to clarify how these two representations are connected and what exactly the Bayesian equilibrium mapping operates on.
  2. Belief updates and information assumptions. Equation (7) uses a likelihood to update beliefs based on observed decisions, but I could not find a sufficiently clear explanation of how this likelihood is obtained when the distribution maps are unknown. The experimental examples also appear to rely on additional observable information to infer players' costs. Clarifying what each player knows and observes would make the incomplete-information setting more convincing.
  3. Convergence analysis. Theorem 3.7 is central to the paper, but some steps in the proof could be explained more carefully. In particular, the proof uses the proposed stable equilibrium as a fixed point before explicitly establishing its existence. It would help to clarify the assumptions that ensure the equilibrium mapping has a well-defined fixed point, especially regarding the decision space.
  4. Finite-sample guarantee. I have some concerns about Theorem 3.9. Assumption 3.8 is described as a finite-variance condition, but it actually imposes a uniform bound on gradient estimation error. The resulting bound also does not explicitly depend on the sample size. Moreover, it is not entirely clear how the proof handles the difference between the current and limiting beliefs. A more precise explanation of these points would strengthen the theoretical contribution.
  5. Connection between the algorithm and theory. Algorithm 1 appears to use individual best-response updates, while the convergence analysis assumes a Bayesian equilibrium is computed at each iteration. It is unclear whether these are equivalent in the proposed implementation. Also, the sensitivity estimate in line 10 is based on consecutive decisions, whereas Definition 3.1 requires a global bound. I would like to understand how the algorithm ensures that the theoretical condition is actually satisfied. The regularization and truthful-reporting guarantees in Theorem 4.1 and Appendix I could also be stated more carefully, since the former concerns an upper bound and the latter relies on informal incentive arguments.
  6. Experimental validation. The experiments are useful, but they rely on two specific market models with largely simulated distribution shifts. Most results focus on total revenue rather than directly measuring equilibrium convergence. Reporting belief convergence, equilibrium residuals, or estimated versus true sensitivity would provide stronger evidence for the theoretical claims. A clearer distinction between the use of real market data and simulated market dynamics would also help.
Questions for authors
  1. Could the authors clarify whether the proposed Bayesian equilibrium is defined over type-dependent strategies or fixed decision vectors? How is the likelihood in Equation (7) constructed, and what information is assumed to be observable to the players?
  2. Could the authors explain how Algorithm 1 satisfies the assumptions of Theorem 3.9 in practice? In particular, how is the Bayesian equilibrium computed at each iteration, and how can the local sensitivity estimate guarantee the required global sensitivity bound?
5 comments

We sincerely appreciate the reviewer's valuable time spent in carefully reviewing our manuscript, and the recognition of our motivation, theoretical analysis, algorithm design and experimental validation.

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


Regarding the Bayesian equilibrium formulation. We thank the reviewer for this insightful question.

We acknowledge that the simplification in the original manuscript may cause confusion. In the revised manuscript, we define the equilibrium explicitly in the strategy space:

  • The strategy si:Θi→Xis _i:\Theta _i\to\mathcal{X} _i is a map from type space to decision space.
  • The Bayesian equilibrium set BEμ(X)⊆ΣBE _\mu(X)\subseteq\Sigma is computed under the data distributions induced by the joint decision XX and the belief μ\mu.
  • The fixed decision vector X=s(θ)X=s(\boldsymbol{\theta}) is the realized decisions when the players' types are θ\boldsymbol{\theta}.

Revision.

  • We have defined the equilibrium in the strategy space and updated the related definitions in Section 2 (lines 150-155, 160, 173, 177-180, 184, 198, 201-203, 227-231).

Regarding the belief update and the information structure. We are grateful for the reviewer's observation and for the chance to clarify this point.

Belief update. The reviewer is correct that the distribution maps Di(X),∀i∈[n]\mathcal{D} _i(X),\forall i\in[n] are unknown to players. ϱi(xi(t)∣h(t),θi)\varrho _i(x _i ^{(t)}\mid h ^{(t)},\theta _i) is the conditional probability of observing the decision xi(t)x _i ^{(t)} given the history h(t)=(X(0),⋯ ,X(t−1),Z(0),⋯ ,Z(t−1))h ^{(t)}=(X ^{(0)},\cdots,X ^{(t-1)},Z ^{(0)},\cdots,Z ^{(t-1)}) and the type θi\theta _i. Specifically, at time step tt, players make decisions X(t)X ^{(t)} based on the data samples Zi(t−1),∀i∈[n]Z _i ^{(t-1)},\forall i\in[n] collected from the data distributions induced by X(t−1)X ^{(t-1)}. Given this information, the players can update the belief about their types without requiring the distribution maps.

In the experiments, this likelihood is calculated with the realized cost:

  • Cournot competition on crude oil exportation. The type θi\theta _i is the average cost, and the realized cost is ci∼N(θi,ς2)c _i\sim\mathcal{N}(\theta _i,\varsigma ^2).

    1. Player ii makes decisions based on cic _i, and the other players make decisions according to the expectation of cost E[ci]=θi\mathbb{E}[c _i]=\theta _i.
    2. After observing the decision xi(t)x _i ^{(t)}, the other players j≠ij\neq i can compute the cost of player ii by ci(t)=z−b∑j≠i(qj+E[xj])−2bqi−2bxi(t)c _i ^{(t)} = z - b\sum _{j\neq i}(q _j+\mathbb{E}[x _j]) - 2bq _i - 2bx _i ^{(t)}, where z,bz,b (the parameters in demand function), and qi,qjq _i,q _j (the base export quantities) are public knowledge, and E[xj]\mathbb{E}[x _j] is the expectation of the other players' decisions under the public belief. Thus, only the cost ci(t)c _i ^{(t)} is unknown and it is recovered from the observed decision.
    3. The likelihood of the observed cost sequence is then ϱi(ci(1),…,ci(t)∣θi)=∏k=1t12πς2e−(ci(k)−θi)22ς2.\varrho _i(c _i ^{(1)},\dots,c _i ^{(t)}\mid\theta _i)=\prod _{k=1} ^t\frac{1}{\sqrt{2\pi\varsigma ^2}}e ^{-\frac{(c _i ^{(k)}-\theta _i) ^2}{2\varsigma ^2}}.
  • Bertrand competition on the ride-share market. The type θi∈R11\theta _i\in\mathbb{R} ^{11} is the average cost vector, and the realized cost is ci∼N11(θi,I11)c _i\sim\mathcal{N} _{11}(\theta _i,\mathbf{I} _{11}).

    1. Player ii makes decisions based on cic _i, and the other players make decisions according to E[ci]=θi\mathbb{E}[c _i]=\theta _i.
    2. After observing the decision xi(t)x _i ^{(t)}, the other players j≠ij\neq i can compute the cost of player ii by ci(t)=(AiT)−1((Ai+AiT)xi(t)+zi+A−iE[x−i]+AiTp1)c _i ^{(t)}=(\mathbf{A} _i ^T) ^{-1}\big((\mathbf{A} _i+\mathbf{A} _i ^T)x _i ^{(t)}+z _i+\mathbf{A} _{-i}\mathbb{E}[x _{-i}]+\mathbf{A} _i ^Tp\mathbf{1}\big), where Ai,A−i\mathbf{A} _i,\mathbf{A} _{-i} (the parameters in demand function), pp (the base price), and ziz _i (the demand parameter) are public knowledge, and E[x−i]\mathbb{E}[x _{-i}] is the expectation of the competitor's decision under the public belief. Thus, only the cost ci(t)c _i ^{(t)} is unknown and it is recovered from the observed decision.
    3. The likelihood of the observed cost sequence is then ϱi(ci(1),…,ci(t)∣θi)=∏k=1t1(2π)11/2e−12∣∣ci(k)−θi∣∣22\varrho _i(c _i ^{(1)},\dots,c _i ^{(t)}\mid\theta _i)=\prod _{k=1} ^t\frac{1}{(2\pi) ^{11/2}}e ^{-\frac{1}{2}||c _i ^{(k)}-\theta _i|| _2 ^2}.

Information structure. For player ii, the observability is shown as follows:

  • Observable information:
    • its own type θi\theta _i,
    • the decisions of all players xi,∀i∈[n]x _i,\forall i\in[n],
    • the data samples collected by all players Zi,∀i∈[n]Z _i,\forall i\in[n],
    • the history h=(X(0),X(1),⋯ ,h=(X ^{(0)},X ^{(1)},\cdots, Z(0),Z ^{(0)}, Z(1),⋯ )Z ^{(1)},\cdots) containing historical joint decisions and historical data samples of all players.
  • Unobservable information:
    • the other players' types θ−i\theta _{-i},
    • the distribution maps Di(X),∀i∈[n]\mathcal{D} _i(X),\forall i\in[n] of all players.

Revision.

  • We have clarified that ϱi\varrho _i is the conditional probability of observing player ii's decision given the type in Section 2.2 (lines 216-217).

Regarding the convergence analysis. We thank the reviewer for this helpful suggestion.

We now establish the existence of a performatively stable Bayesian equilibrium before the convergence analysis.

When α>∑iε^i2\alpha>\sqrt{\sum _i\hat\varepsilon _i ^2}, we prove the existence in the following steps:

  1. By Lemmas 3.6 and 3.7, we show that ∣∣X(t+1)−X(t)∣∣2→0||X ^{(t+1)}-X ^{(t)}|| _2\to0 almost surely as t→∞t\to\infty.
  2. Since XRPBG\mathcal{X} _{RPBG} is compact, there exists a convergent subsequence X(tk)→X∞X ^{(t _k)}\to X ^{\infty}.
  3. By ∣∣X(t+1)−X(t)∣∣2→0||X ^{(t+1)}-X ^{(t)}|| _2\to0 and the triangle inequality, we have X(tk+1)→X∞X ^{(t _k+1)}\to X ^{\infty} almost surely as k→∞k\to\infty.
  4. By X(tk+1)=BEμ(tk)(X(tk))X ^{(t _k+1)}=BE _{\mu ^{(t _k)}}(X ^{(t _k)}) and Lemmas 3.6 and 3.7, we have X(tk+1)→BEμ(∞)(X∞)X ^{(t _k+1)}\to BE _{\mu ^{(\infty)}}(X ^{\infty}) almost surely as k→∞k\to\infty.
  5. Since X(tk+1)→X∞X ^{(t _k+1)}\to X ^{\infty} and X(tk+1)→BEμ(∞)(X∞)X ^{(t _k+1)}\to BE _{\mu ^{(\infty)}}(X ^{\infty}), we have X∞=BEμ(∞)(X∞)X ^{\infty}=BE _{\mu ^{(\infty)}}(X ^{\infty}), which ensures the existence of a performatively stable Bayesian equilibrium.

Revision.

  • We have added Assumption 3.4 (affine structure of the loss) in Section 3.2 (lines 283-295).
  • We have proved the existence of a performatively stable Bayesian equilibrium in Appendix E (lines 868-939).

Regarding the finite-sample guarantee. We thank the reviewer for these insightful suggestions.

Bounded estimation error. We agree that the name "finite variance" was inaccurate. We have renamed Assumption 3.9 (original Assumption 3.8) to "Bounded estimation error", since it assumes a bound on the gradient estimation error without depending on sample size.

Sample-size dependence.

  • Theorem 3.10 (original Theorem 3.9) is independent of the sample size mm, since the bound σ\sigma in Assumption 3.9 is a uniform bound on the gradient estimation error.
  • Our additional Central Limit Theorem based analysis in Appendix G provides a convergence error bound that explicitly depends on the sample size mm. The average gradient estimation error with mm samples satisfies ∣∣∇G‾−∇G∣∣2≤δ||\overline{\nabla G}-\nabla G|| _2\le\delta with probability Fχ2(d)(mδ2/σ′)F _{\chi ^2(d)}(m\delta ^2/\sigma') under the large-sample assumption, where Fχ2(d)(⋅)F _{\chi ^2(d)}(\cdot) is the χ2\chi ^2 distribution with dd degrees of freedom and σ′\sigma' is the largest eigenvalue of the covariance matrix.

Current and limiting beliefs. We have revised the proof of Theorem 3.10 so that strong monotonicity is applied at the same belief. The gradient difference across beliefs vanishes almost surely as t→∞t\to\infty by Lemma 3.5 and the Portmanteau theorem.

Revision.

  • We have renamed the assumption to "Bounded estimation error" (Assumption 3.9) in Section 3.3 (line 346).
  • We have revised the proof to account for the gradients under different beliefs in Appendix F (lines 1066-1092).
Reviewer_uZvpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies competition among players whose decisions affect the data distributions they face and reveal information about their private types. At each round, a Bayesian equilibrium is computed under the current distributions, and observed decisions are used to update beliefs. The equilibrium concept pairs a limiting belief with a decision profile that remains optimal under the distribution generated by those decisions. The convergence analysis relies on posterior convergence and a bound on the sensitivity of equilibrium decisions to changes in the underlying distributions. The paper establishes sufficient conditions for convergence, examines the effects of sampling error, and introduces adaptive quadratic regularization to improve stability. Numerical experiments in Cournot and Bertrand competition illustrate the method.

Critical strengths

Introducing private information is a natural extension of multiplayer performative prediction. In the market settings considered in the paper, a firm's decisions can affect market conditions while also revealing information about its costs to competitors. This motivates studying how beliefs and decisions evolve together. The formulation also allows uncertainty to persist in the limiting belief when different types cannot be distinguished from observed behavior. Lemma 3.6 relates the stability of the game under a fixed belief to the strength of the distribution feedback. The calculation in Appendix E further shows how a contraction argument can accommodate belief changes that vanish over time, without requiring a specific rate of posterior convergence. This is a useful observation, although the convergence result still requires establishing continuity of the full Bayesian equilibrium.

Critical weaknesses

The convergence guarantee for the sampled adaptive algorithm needs substantial revision. Under the implementation described in Appendix H, line 10 of Algorithm 1 estimates sensitivity by dividing differences in empirical gradients by changes in consecutive decisions. When successive decisions become close, sampling noise can keep the numerator away from zero while the denominator approaches zero. As a result, the sensitivity estimate and the regularization coefficient can become arbitrarily large, even with a constant data distribution and uniformly bounded gradient errors. Compactness of the decision set does not prevent this behavior. Theorem 3.9 establishes an asymptotic error neighborhood for a fixed game, but does not establish convergence of the sampled adaptive algorithm. Such a result would need to account for noise in the sensitivity estimate, changes in the regularization coefficient, and the equilibrium, if any, approached by the resulting process.

The main Bayesian convergence argument also has gaps. A Bayesian equilibrium specifies a strategy for each possible type, so the analysis must distinguish these strategies from the actions actually observed. In Appendix C, the opponents' equilibrium decisions are held fixed as beliefs change, and the resulting individual minimizers are then treated as the equilibrium of the modified game. This does not establish continuity of the joint equilibrium, since the opponents' strategies may also change with beliefs. Appendix F, Eq. (65), applies strong monotonicity to gradients evaluated under different beliefs without accounting for the change in beliefs. These issues are particularly important because Zhong et al. (2026a) already develops the sensitivity measure and adaptive regularization approach for the complete-information setting. A precise formulation of Bayesian strategies and a complete proof of equilibrium continuity are therefore needed to establish the claimed extension.

Questions for authors

The reporting argument in Appendix I needs to account for firms' incentives. In Cournot competition, stronger common regularization can reduce total output and increase a reporting firm's profit under its original payoff function. A firm may therefore benefit from overstating its sensitivity. The paper should either establish that truthful reporting is in each firm's interest or explicitly assume that the reported sensitivities can be verified.

The scope of Theorem 4.1 also needs clarification. Appendix J compares equilibria while holding the distribution and belief fixed, and its optimality argument concerns a particular upper bound on the distance between these equilibria. This is narrower than the claim in the abstract about the maximum deviation from the original equilibrium. A comparison of the final performative equilibria must also account for the distributions and beliefs reached by the original and regularized processes. Without this additional analysis, it is unclear whether the stated bound extends to the final equilibria.

5 comments

We are sincerely grateful for the reviewer's commitment in evaluating our manuscript and for recognizing the novelty of our research problem and theoretical results.

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


Regarding the sensitivity estimate. We thank the reviewer for this detailed comment and the chance to clarify the implementation details.

Algorithm 1 is an adaptive implementation that uses the convergence condition of Theorem 3.10 (original Theorem 3.9) to set the regularization strength. We estimate the sensitivity during the repeated game and tune the regularization accordingly. 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.

  • Denominator in ε^i\hat{\varepsilon}_i estimation. We have added a small ν>0\nu>0 to the denominator of the estimate, i.e., ∣∣X(t)−X(t−1)∣∣2+ν||X^{(t)}-X^{(t-1)}||_2+\nu, so the estimate does not grow without bound when consecutive decisions become close. We have also reported the value of ν\nu in the manuscript.

  • The fixed game in the algorithm. Since the estimate ε^i\hat{\varepsilon}_i is a running maximum, the regularization parameter γ=max⁡(0,c∑kε^k2−ψ)\gamma=\max(0,c\sqrt{\sum_k\hat\varepsilon_k^2}-\psi) increases as the estimate increases. When γ\gamma is sufficiently large to satisfy the convergence condition, the decisions converge. Because of the ν\nu in the denominator, the estimate ε^i\hat{\varepsilon}_i does not grow without bound once the decisions converge. Then γ\gamma and the regularized losses become fixed. This fixed game meets the conditions of Theorem 3.10 under the estimate ε^i\hat{\varepsilon}_i.

Revision.

  • We have updated Algorithm 1 by adding ν\nu to the denominator in Section 4 (lines 381, 397).
  • We have reported the value of ν\nu for our experiment in Section K.1 (line 1572)
  • We have clarified that the algorithm is an adaptive implementation that uses the convergence condition of Theorem 3.10 to set the regularization strength in Section 4 (lines 418, 421).
  • We have clarified that the regularized game in our algorithm refers to the regularized game with the converged γ\gamma in Section 4 (lines 414-416).

Regarding the strategy formulation. We thank the reviewer for these careful comments, and we have strengthened the proof on these points.

Strategy space. In the revised manuscript, we define the equilibrium as a strategy profile s=(s1,…,sn)s=(s _1,\dots,s _n) with the map si:Θi→Xis_i: \Theta_i \to \mathcal{X} _i, and the Bayesian equilibrium as BEμ(X)⊆ΣBE _\mu(X) \subseteq \Sigma with Σ\Sigma the strategy space. Under α\alpha-strong monotonicity (Assumption 3.3), the best response is unique. By the additional affine structure (Assumption 3.4), the best response depends on other players' strategies only through the expected decision vector sˉ−i=Eθ−i∼μ[s−i(θ−i)]\bar{s} _{-i}=\mathbb{E} _{\theta _{-i} \sim \mu}[s _{-i}(\theta _{-i})], so the equilibrium BEμ(X)BE _\mu(X) is determined by the expected decision vector sˉ=(sˉ1,…,sˉn)\bar{s}=(\bar{s} _1,\dots,\bar{s} _n) of all players.

Continuity of the joint equilibrium. In the revised Appendix C, we analyze the equilibrium strategy profile ss, which is determined by the expected joint decision sˉμ=Eθ∼μ[s(θ)]=(sˉ1,μ,⋯ ,sˉn,μ)\bar{s} _\mu= \mathbb{E} _{\boldsymbol{\theta} \sim \mu}[ s (\boldsymbol\theta)] = \left(\bar{s} _{1,\mu},\cdots,\bar{s} _{n,\mu}\right), where sˉi,μ\bar{s} _{i,\mu} is the expected decision of player ii. The decisions of the other players change as the belief changes. By the α\alpha-strong monotonicity (Assumption 3.3) and the affine structure (Assumption 3.4), the expected decision satisfies:

sˉi,μ=Eθi∼μ[BRi(sˉ−i,μ,θi)],\bar{s} _{i,\mu}=\mathbb{E} _{\theta_i\sim\mu}\bigl[BR _i(\bar{s} _{-i,\mu},\theta _i)\bigr],

where BRi(sˉ−i,θi)BR_i(\bar{s} _{-i},\theta _i) is player ii's best response to the other players' expected decision sˉ−i=Eθ−i∼μ[s−i(θ−i)]\bar{s} _{-i}=\mathop{\mathbb{E}} _{\theta _{-i}\sim\mu}\left[s _{-i}(\theta _{-i})\right].

Then, the expected decision vector sˉμ\bar{s}_\mu is a fixed point that determines the equilibrium strategy profile ss.

We establish the continuity of the expected decision vector sˉμ\bar{s}_\mu in the belief:

  • The Berge's maximum theorem gives the continuity of BRi(sˉ−i,θi)BR _i(\bar{s} _{-i},\theta _i) in sˉ−i\bar{s} _{-i}.
  • The Portmanteau theorem gives the convergence of Eθi∼μ[BRi(sˉ−i,θi)]\mathbb{E} _{\theta _i\sim\mu}[BR _i(\bar s _{-i},\theta _i)] as μ(t)→μ(∞)\mu ^{(t)}\to\mu ^{(\infty)} for each fixed sˉ−i\bar s _{-i}.
  • The compactness of XRPBG\mathcal{X} _{RPBG} together with the uniqueness (α\alpha-strong monotonicity) gives sˉμ(t)→sˉμ(∞)\bar s _{\mu ^{(t)}}\to\bar s _{\mu ^{(\infty)}}.

Gradients under different beliefs. In the revised Appendix F, the strong monotonicity is applied only at the fixed belief μ(t)\mu^{(t)}.

By Lemma 3.5 and the Portmanteau theorem, together with the continuity of the gradients ∇xiℓi\nabla _{x _i}\ell _i in the type space, the gradient difference under different beliefs ∣∣∇XPSGμ(t)(XPS)−∇XPSGμPS(XPS)∣∣2→0||\nabla _{X^{PS}}\mathcal{G} _{\mu ^{(t)}}(X ^{PS})-\nabla _{X ^{PS}}\mathcal{G} _{\mu ^{PS}}(X ^{PS})|| _2 \to 0 almost surely as t→∞t\to\infty.

Revision.

  • We have defined the equilibrium in the strategy space and updated the related definitions in Section 2 (lines 150-155, 160, 173, 177-180, 184, 198, 201-203, 227-231).
  • We have added Assumption 3.4 (affine structure of the loss) in Section 3.2 (lines 283-295).
  • We have revised the proof to ensure the continuity of the equilibrium in Appendix C (lines 786-817).
  • We have revised the proof to account for the gradients under different beliefs in Appendix F (lines 1066-1092).

Regarding the reporting incentives. We thank the reviewer for this insightful question.

We acknowledge that truthful reporting of ε^i\hat{\varepsilon}_i is an assumption, and we agree that a player may gain by misreporting and deviating from the algorithm in some specific games, e.g., Cournot competition.

We therefore explicitly assume that the reported sensitivities are truthful. We do not assume that the reports can be verified, because the true sensitivity depends on private information that is unobservable to the other players. This assumption is motivated by the intuition that, without fixing a specific game structure, truthful reporting does not systematically reduce the players' payoffs.

Revision.

  • We have clarified the ε^i\hat{\varepsilon}_i reporting assumption in the footnote of Section 4 (line 431) and in Appendix I (lines 1452-1454, 1458-1459, 1466-1469).

Regarding the scope of Theorem 4.1. We appreciate the comment and the opportunity to clarify.

Abstract. The original description "the maximum deviation from the original equilibrium" in the Abstract may lead to misunderstanding. We have revised the statement to "the upper bound on the distance between the regularized and original equilibria".

Theorem 4.1. We clarify that Theorem 4.1 provides the upper bound on the distance between the equilibria computed using the regularized losses and the original losses, under the same data distributions and the same belief, given the regularization parameter γ\gamma.

  • Account for the distributions and beliefs. The bound provided by Theorem 4.1 compares the original and regularized equilibria in each iteration of our algorithm, where the data distributions and the belief are the same.

  • Extension to the final equilibria. Theorem 4.1 can be applied to the performatively stable Bayesian equilibrium achieved by our algorithm, where the upper bound on the distance between the regularized and original equilibria is calculated for the data distribution induced by the performatively stable Bayesian equilibrium and the converged belief.

Revision.

  • We have replaced "the maximum deviation from the original equilibrium" by "the upper bound on the distance between the regularized and original equilibria" in the Abstract (lines 27-28) and in Section 1.1 (lines 103-104).
  • We have clarified that the regularized and original games in Theorem 4.1 are compared under the same data distributions and the same belief in Section 4 (lines 427-428).
Reviewer_MbW8OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies performative prediction games with incomplete information. It formalizes the framework of Performative Bayesian Prediction Games, an extension of Performative Prediction Games where each player has her own private information. The players' loss functions depend on their joint actions XX, private information θ\theta, and data ZZ drawn from some unknown distribution D(X)D(X).

Their main result concerns repeated games. At each iteration tt, the players play a Bayesian equilibrium XtX^t under the data distribution D(Xt−1)D(X^{t-1}) induced by the previous joint decision and a belief over private information updated according to Bayes' rule. The authors show that this process converges to an equilibrium when the gradients of the expected losses are ϵ\epsilon-Lipschitz in the joint decision XX that induces the distributions, for sufficiently small ϵ\epsilon.

When this assumption is violated, they provide an algorithm that converges to an approximate equilibrium by dynamically regularizing the loss functions. Some experiments validate the results.

Critical strengths

The setting is a natural extension of the performative games studied by Zhong et al. (2026a). The paper extends the analysis of the previous work to incomplete-information settings.

Critical weaknesses
  • The main paper states the main results, but it does not provide a clear idea of the techniques employed to prove them. Since the results mirror those of Zhong et al. (2026a), it is unclear which technical challenges arise from the Bayesian extension and how they are solved.

  • I find Section 4 unclear. Since the regularizers γt,i\gamma_{t,i} change over time, it is unclear which regularized game the authors refer to. I also do not think that Theorem 3.9 can be used to prove convergence (Line 348), as it was stated for fixed loss functions and an exact ϵ^\hat\epsilon parameter, while the algorithm adopts time-varying regularized losses and estimates ϵ^\hat\epsilon.

  • I believe the argument provided in Appendix I is wrong. Lines 1363–1368 basically state that truthful reporting is not an equilibrium; assuming that the other players report truthfully, player ii gains by misreporting and deviating from the algorithm. Overall, the entire argument in Appendix I is informal.

  • The notation is unclear. The paper writes the expected loss Li(X)\mathcal{L}_i(X) as if it depended on a single joint profile, but it actually depends directly on a profile XX and indirectly on another profile YY that induces the data distribution.

Questions for authors
  • What are the main differences with respect to the analysis by Zhong et al. (2026a)?
  • Can the authors address my doubts regarding Section 4 (see weaknesses)?
  • In Example 1.1, why is the demand qiq_i different from ziz_i? Can you provide an example where a firm observes some stochastic parameter ziz_i, and then the demand is a deterministic function of this parameter and the price, rather than directly observing the demand?
4 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 difference with Zhong et al. (2026a). We appreciate the comment and the opportunity to clarify the difference from related works.

In short, the main difference is that we consider incomplete information, where the decision updates are entangled with the belief updates during a repeated game. Under this setting, the decision and the belief change simultaneously, each affecting the other’s dynamics. In the complete-information setting, Zhong et al. (2026a) only analyze the dynamics of the decision.

This entanglement creates two technical challenges that do not appear in Zhong et al. (2026a):

  1. the joint sequence of decisions and beliefs must be shown to converge, and
  2. the Bayesian equilibrium map must be shown to be continuous with respect to the belief.

These challenges require four tools that are not used in the complete-information analysis:

  1. The martingale convergence theorem for belief convergence (Appendix A, A.1–A.3);
  2. The Berge's maximum theorem for continuity of the strategy si(θi)s_i(\theta_i) in the type θi\theta_i (Appendix A, A.4);
  3. The Portmanteau theorem for continuity of the equilibrium map BEμ(X)BE_\mu(X) in the belief μ\mu (Appendix A, A.5);
  4. The equicontinuity property for uniform convergence (Appendix A, A.8–A.9).

This leads to the following differences:

Our WorkZhong et al. (2026a)
InformationPrivate type θ\theta, belief μ\mu over typesComplete information
Equilibrium objectStrategy profile s:Θ→Xs:\Theta\to\mathcal{X} and belief μ\muDecision profile XX
Convergence analysisdecision gap(t)≤β⋅decision gap(t−1)+belief-induced decision shift(t−1)\text{decision gap}^{(t)} \le \beta\cdot\text{decision gap}^{(t-1)} + \text{belief-induced decision shift}^{(t-1)}; belief gap(t)→0\text{belief gap}^{(t)} \rightarrow 0decision gap(t)≤β⋅decision gap(t−1)\text{decision gap}^{(t)} \le \beta \cdot \text{decision gap}^{(t-1)}


Regarding the algorithm in Section 4. We thank the reviewer for the opportunity to clarify this point. The regularized game referred to in the analysis is the game obtained once the regularization parameter γ\gamma has converged.

Algorithm 1 is an adaptive implementation that uses the convergence condition of Theorem 3.10 to set the regularization strength. Because ε^i\hat{\varepsilon}_i is maintained as a running maximum, γ\gamma is non-decreasing. Once γ\gamma is large enough for the decisions to stabilize, the regularized losses become fixed and the setting of Theorem 3.10 applies. 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.

  • Estimated ε^i\hat{\varepsilon}_i. If the estimate ε^i\hat{\varepsilon}_i is small, the regularization is insufficient and the players' decisions fluctuate. In this case, the running maximum in Algorithm 1 keeps increasing the estimate ε^i\hat{\varepsilon}_i as long as the regularization is insufficient. Once the regularization is sufficiently large, the decisions converge, and the estimate ε^i\hat{\varepsilon}_i stabilizes.

  • Regularization parameter γ\gamma.
    In each iteration, γ\gamma is calculated by γ=max⁡(0,c∑i=1nε^i2−ψ).\gamma = \max\left(0,c \sqrt{\sum_{i=1}^n\hat{\varepsilon}_i^2}-\psi\right). When the estimate ε^i\hat{\varepsilon}_i stabilizes, the regularization parameter γ\gamma does not change over time.

  • Regularized game in the algorithm. The regularized game refers to the game with the converged γ\gamma in the repeated game. Specifically, when γ\gamma converges, the regularized loss functions are fixed.

Revision.

  • We have clarified that the algorithm is an adaptive implementation that uses the convergence condition of Theorem 3.10 to set the regularization strength in Section 4 (lines 418, 421).
  • We have clarified that the regularized game in our algorithm refers to the regularized game with the converged γ\gamma in Section 4 (lines 414-416).

Regarding Appendix I. We appreciate the insightful observation and the opportunity to revise this point.

The original version may have presented truthful reporting as a proven property. We acknowledge that truthful reporting of ε^i\hat{\varepsilon}_i is an informal assumption, motivated by the following intuition.

  • If some or all of the players report a small ε^i\hat{\varepsilon}_i, the regularization may be insufficient for the game to converge, causing all players' payoffs to fluctuate.
  • If some or all of the players report a large ε^i\hat{\varepsilon}_i, all players' regularization may grow, but the players' payoffs may not rise.

Thus, without fixing a specific game structure, we assume that players truthfully report ε^i\hat{\varepsilon}_i, since truthfully reporting does not reduce the players' payoffs in general.

We acknowledge that a player may gain by misreporting and deviating from the algorithm in some specific games.

Revision.

  • We have clarified that truthful reporting of ε^i\hat{\varepsilon}_i is an assumption in the footnote of Section 4 (line 431) and in Appendix I (lines 1452-1454, 1458-1459, 1466-1469).

Regarding the notation of Li(X)\mathcal{L}_i(X). We appreciate the comment.

In Eqs. (2) and (3), XX is the joint decision that induces the data distribution Di(X)\mathcal{D}_i(X) appearing in the loss function Li(X)\mathcal{L}_i(X), which follows the decision-dependent-game literature (Narang et al., 2023; Zhong et al., 2026a). The notation Li(X)\mathcal{L}_i(X) indicates the data distribution under which player ii optimizes, facilitating the analysis of the performative effect.

We clarify that, for a performatively stable Bayesian equilibrium, the joint decision XX both induces the data distribution Di(X)\mathcal{D}_i(X) and is a Bayesian equilibrium under that distribution.

We distinguish the two roles in the analysis of the repeated game (Definition 2.1), where the equilibrium decision X(t)X^{(t)} is computed under the data distributions induced by the previous decision X(t−1)X^{(t-1)}, i.e., X(t)=s(t)(θ)∈BEμ(θ∣h(t))(X(t−1))X^{(t)}=s^{(t)}(\boldsymbol\theta)\in BE_{\mu(\boldsymbol{\theta}|h^{(t)})}(X^{(t-1)}).

Revision.

  • We have clarified that XX in the notation Li(X)\mathcal{L}_i(X) and the data distribution map Di(X)\mathcal{D}_i(X) is the joint decision that induces the data distribution in Section 2.1 (lines 166-167).
  • We have clarified the roles of XX and YY in ∇yiLi(X)\nabla_{y_i}\mathcal{L}_i(X) in Section 3.1 (lines 251-254).

References

Adhyyan Narang, Evan Faulkner, Dmitriy Drusvyatskiy, Maryam Fazel, and Lillian J Ratliff. Multiplayer performative prediction: Learning in decision-dependent games. Journal of Machine Learning Research, 24(202):1–56, 2023.

Guangzheng Zhong, Yang Liu, and Jiming Liu. Towards performatively stable equilibria in decision-dependent games for arbitrary data distribution maps. Machine Learning, 115(8):197, 2026a.


Regarding Example 1.1. We thank the reviewer for this question.

In the Bertrand game, the demand function qi=zi+aixi+a−ix−iq_i = z_i + a_i x_i + \mathbf{a} _{-i}x _{-i} describes the influence of the firms' prices on the demand of each firm, and the intercept ziz_i represents the market size of this category of goods.

A standard Bertrand game assumes this market size is fixed. However, in practice the market size may change with the prices set by all firms, which is the performative effect we model. Two examples illustrate this:

  • Substitutable consumption (e.g., the ride-sharing service in our experiments): if all ride-sharing platforms raise their prices, the market size may shrink, since consumers switch to the alternatives (e.g., taxis, buses, and the metro).
  • Conspicuous consumption: when the overall price level rises, the good becomes a better way to show off wealth, which may enlarge the market size.

Revision.

  • We have clarified in Example 1.1 that ziz_i is the market size in Section 1 (line 62).

I thank the authors for their response.

I believe that the convergence of Algorithm 1 requires a formal proof. The convergence of the estimates ϵ^i\hat \epsilon_i and therefore of γ\gamma is not proved, and Theorem 4.1 does not establish a property of the algorithm itself. Notably, since it analyzes the games over fixed beliefs and a single strategy profile, it seems the same result proved by Zhong et al. (2026a) (Theorem 3). A clean result would need to analyze the convergence of the algorithm and γ\gamma, the distance between the equilibrium of the original game the one found by the algorithm, and the error over a finite number of iterations.

Furthermore, I am not convinced by the relationship between the expected decision vector sˉμ\bar s_\mu and the strategy profile the players converge to (both in Section 3 and in Section 4). Lines 294-295 seem to "change" definition 2.2, so that BEμ(X)BE_\mu(X) should be interpreted as the expected decision vector. However, each player knows their own type and plays xi=si(θi)x_i = s_i(\theta_i), thus I would not expect them to converge to the average decision, as player ii is not playing in expectation over θi\theta_i.

Finally, I am not convinced by the definition of the stochastic quantity in the example. Thinking about a ride-sharing service, I would say that it is natural to measure the demand qiq_i (by simply observing how many people by the service), and then compute the loss based on it (I also believe that Zhong et al. (2026a) used this model). Instead, the example assumes that the market size is directly measurable, and the demand is a deterministic function of it and the prices.

We thank the reviewer for the careful reading of our response.

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 Algorithm 1. We thank the reviewer for this comment.

  • Convergence of Algorithm 1. We agree with the reviewer that a formal convergence proof of Algorithm 1 is not established in this paper. The effectiveness of our algorithm is validated experimentally.

  • Theorem 4.1. We agree that this result follows the same argument as Theorem 3 of Zhong et al. (2026a), and we have revised it as Corollary 4.1.

  • Algorithm guarantees. We thank the reviewer for this suggestion. The convergence of the algorithm and of γ\gamma, the distance between the equilibrium of the original game and the regularized game, and the error over a finite number of iterations will be established in future work.

Revision.

  • We have revised Theorem 4.1 as Corollary 4.1 in Section 4 (lines 425–437).

Regarding the realized-action dynamics. We thank the reviewer for this careful observation. We will address this issue in future work.


Regarding the stochastic quantity. We thank the reviewer for this comment. The reviewer is correct that our example and the ride-sharing experiment differ from those of previous work (Zhong et al., 2026a; Narang et al., 2023). We consider a Bertrand competition in which the players make their decisions based on a publicly known deterministic price function. In this case, the market size can be calculated from the observed demand, zi=qi−aixi−a−ix−iz _i = q _i-a _i x _i - a _{-i} x _{-i}, and is known to all players.


Reference.

G. Zhong, Y. Liu, and J. Liu. Towards performatively stable equilibria in decision-dependent games for arbitrary data distribution maps. Machine Learning, 115(8):197, 2026a.

A. Narang, E. Faulkner, D. Drusvyatskiy, M. Fazel, and L. J. Ratliff. Multiplayer performative prediction: Learning in decision-dependent games. Journal of Machine Learning Research, 24(202):1–56, 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.