acceptodds
Under review as a conference paper at ICLR 2027

Risk-Seeking Strategic Robustness in Games: A Substitutes-Complements Rule, and Where It Breaks

Abstract

Strategic risk aversion, formalized through a fictitious-adversary construction, has recently been shown to induce a “free lunch” in a shared-effort game: more robustness can strictly improve equilibrium welfare, in contrast to classical robust optimization. We study this phenomenon in a tractable class of two-player quadratic games, and introduce the natural risk-seeking counterpart of the construction, in which the fictitious adversary is replaced by a fictitious ally. Within this game class, we show that whether risk aversion, risk seeking, or neither attitude helps is governed by a single, classical structural property: whether the game exhibits strategic substitutes or complements, and that the losing attitude always strictly hurts welfare instead. We illustrate the theory on three examples: trading, a shared-effort aggregative game, and Bertrand price competition.

Primary area:
learning theory
Keywords:
game theory, robust optimization, multi-agent systems, strategic substitutes and complements

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qZDzOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper considers two-player quadratic games, where each player chooses the best response against a fictitious adversary that can deviate from their real opponent's action within a bounded budget. Prior work has shown that, when the fictitious adversary is risk-averse, the resulting equilibrium may admit a larger welfare. In this paper, the authors further consider fictitious adversary that is risk-seeking, a natural counterpart of risk-averse adversary, and establish the following principles to characterize welfare-improving scenarios: risk aversion improves welfare when the game is strategic substitute, while risk seeking improves welfare when the game is strategic complement. The authors further illustrate their results using specific games including trading, a share-effort aggregative game, and Bertrand price competition.

Critical strengths

This paper studies an interesting and fundamental question: whether the prior free-lunch results have more general interpretations. The non-technical parts of the paper are clearly written, and I really enjoy reading them. One may argue that the quadratic games constitute only a restricted game family, yet its special structures enable tractable analysis and hence some new insight. I find the "risk-aversion under substitutes, and risk-seeking under complements" conclusion clean and intuitive in hindsight, and I believe it's interesting to the broader community and will motivate further research (e.g., whether this conclusion holds under more general games).

Critical weaknesses

The technical expositions (Section 5) are difficult to follow. The well-posedness concept should be more properly defined and introduced, and it's hard for me to interpret all the theorem statements. In addition, it's unclear what K(λ)K(\lambda) and L(λ)L(\lambda) are.

Although the paper is rather complete as a theoretical work, some experimental results by, say, implementing this conclusion in multi-agent learning systems will be very appealing and make the conclusion more convincing.

Questions for authors

Are there more general definitions of strategic substitutes and complements in prior literature for more general games?

1 comment

We thank the reviewer for the encouraging assessment, and in particular for finding the "risk aversion under substitutes, risk seeking under complements" conclusion clean and of broader interest.

Clarity of Section 5. We agree, and will revise the section along four lines.

  • Well-posedness will be defined explicitly before Theorem 1. The robust utility involves two optimizations: the fictitious player chooses its report, and the player chooses an action. The construction is well posed when both have a unique solution: the fictitious player's problem has a minimum (adversary) or a maximum (ally), and the player's resulting robust utility is strictly concave in their own action. Theorem 1 states exactly when this holds: 1+δλ>01 + \delta\lambda > 0 for risk aversion, and λ(δ+γ2/(4∣β∣))<1\lambda\big(\delta + \gamma^2/(4|\beta|)\big) < 1 for risk seeking (this condition implies the other one in Theorem 1, which we will drop).
  • K(λ)K(\lambda) and L(λ)L(\lambda) are the intercept and slope of the robust best response, ai=K(λ)+L(λ) aja_i = K(\lambda) + L(\lambda)\,a_j, introduced in Section 5.2 with closed forms in Appendix A.1. We will move the closed forms to the main text.
  • The role of each proposition will be stated upfront. Proposition 1 shows that robustness never changes the strategic type: the best response slopes the same way as without robustness. Propositions 2 to 5 then describe, for each attitude and each type, the values of λ\lambda at which an equilibrium exists and at which best-response dynamics converge to it. Splitting this into four cases, each with its own thresholds and conditions, is a large part of what makes the section hard to follow, and the next point removes it.
  • A unifying parameter. Writing t=λt = \lambda for risk aversion and t=−λt = -\lambda for risk seeking, both attitudes are described by the same formulas. For example, the best-response slope becomes, for both attitudes, L(t)=γ/(2∣β∣)1+ktL(t) = \frac{\gamma/(2|\beta|)}{1 + k t} with k=δ+γ2/(4∣β∣)k = \delta + \gamma^2/(4|\beta|). From this single expression, existence fails only where L=1L = 1, dynamics converge exactly where ∣L∣<1|L| < 1, and the four thresholds now stated separately become one value of tt. Propositions 2 to 5 thus merge into a single proposition with one threshold and one condition, replacing the current case analysis.

The welfare result (Theorem 2) will also be stated more precisely, as discussed with Reviewer amU5.

Experiments. We see the paper's contribution as theoretical, with three worked case studies (trading, a shared-effort game, and Bertrand pricing) in which every prediction is derived in closed form and verified numerically. Learning experiments are a natural next step: Qu et al. (2026) already train collaborative agents with risk-averse robustness, and our results predict which games call for the risk-seeking counterpart instead. We leave this to follow-up work.

More general definitions of substitutes and complements. Yes. In general, a game has strategic complements when a higher action by the opponent raises the benefit of a player's own higher action, and strategic substitutes when it lowers it. This is formalized in the theory of supermodular games (Topkis, 1979; Milgrom and Roberts, 1990) and, for smooth payoffs, by the sign of the cross-partial derivative of the payoff (Bulow et al., 1985). In our quadratic games, all of these reduce to the sign of γ\gamma. This general framework, with its monotone comparative statics (Milgrom and Shannon, 1994), is the natural route to extending our pairing rule beyond quadratic games, and we will mention it in the discussion.

Reviewer_amU5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

his paper explores what happens if agents engage in “strategic risk aversion” by assuming that the other agent will behave worse or better than expected if just assuming a “best-response”. The paper considers what might happen if an agent imagines that the other agent is taking an action that leads to a deviation that it is harmful to them, or alternatively taking an action that is beneficial for them, and then optimizes accordingly.

The paper provides a mathematical framework for modelling such scenarios and reasoning about strategic “complements” and “substitutes”. The authors further characterize when the outcome improves (in terms of social welfare) over the Nash equilibrium outcome as a function of the risk aversion/seeking alone. Small simulations are also provided.

Critical strengths

I like the core problem. As the paper mentions, there has been earlier work looking at risk aversion, but understanding the impacts of risk-seeking behaviour is interesting and, to the best of my knowledge, novel. The model, while limited to two player, scaler action, bilinear quadratic games, does provide a starting point to understand the value of considering alternative models of behaviour and have it may impact the outcomes (particular in terms of welfare). This is particularly nice since we all know that sometimes Nash can result in unappealing outcomes from a social welfare perspective.

Critical weaknesses

While I believe Theorem 1 is correct, I am less confident in the conclusions of Theorem 2, which is the key result of the paper. Theorem 2 provides a characterization as to when risk-averse play or risk-seeking play can improve welfare over a NE outcome. It is not obvious to me that this is correct, and even in the examples provided in Section 6 there seems to be a contradiction. For example, in Figure 2, the figure shows that the social welfare drops below the NE welfare if \lamba>6, but I think \lambda still satisfies the constraints of Theorem 2. My hunch is that a weakened version of Theorem 2 might be true (maybe imposing some local properties or new constraints on the parameters) but not as currently stated.

Questions for authors

I would like the authors to address my question about Theorem 2.

1 comment

We thank the reviewer for the careful reading and the precise counterexample. The reviewer is right: Theorem 2 is not stated precisely enough, and the hunch is correct. The theorem identifies which attitude improves welfare, but not for which values of λ\lambda. The precise statement, consistent with Remark 2, is the following.

Along the paired attitude (risk aversion under substitutes, risk seeking under complements), as λ\lambda increases from 00, the equilibrium moves monotonically away from Nash, in the direction of the joint optimum. Welfare therefore increases up to an optimal point λ⋆\lambda^\star, where the equilibrium reaches the joint optimum exactly, and then decreases. Since welfare is a concave quadratic in the action, it falls below the Nash level beyond a second point λ0>λ⋆\lambda_0 > \lambda^\star. Both points are explicit (upper sign for risk aversion, lower sign for risk seeking): λ⋆=∓2γ+2δ,λ0=∓2(2β+γ)4βδ+βγ+δγ,\lambda^\star = \mp\frac{2}{\gamma + 2\delta}, \qquad \lambda_0 = \mp\frac{2(2\beta + \gamma)}{4\beta\delta + \beta\gamma + \delta\gamma}, each valid when positive and inside the well-posed range. If λ⋆\lambda^\star is not (which happens exactly when γ\gamma and γ+2δ\gamma + 2\delta have opposite signs), welfare increases over the whole range; if only λ0\lambda_0 is not, welfare stays above Nash over the whole range. The unpaired attitude, instead, lowers welfare for every λ\lambda starting from Nash. These statements hold on the range of λ\lambda connected to Nash, that is, up to the first singularity of the equilibrium path (Propositions 3 and 5).

In the trading example of Figure 2, λ⋆=2\lambda^\star = 2 and λ0=6\lambda_0 = 6: welfare peaks at λ=2\lambda = 2 and drops below Nash for λ>6\lambda > 6, exactly as the reviewer observed. The same holds in our other two examples (λ⋆=1\lambda^\star = 1, λ0=12/7\lambda_0 = 12/7 for the aggregative game; λ⋆=2\lambda^\star = 2, λ0=3\lambda_0 = 3 for Bertrand).

We will restate Theorem 2 and its proof accordingly. Its main message, that the sign of γ\gamma decides which attitude can improve on Nash, is unchanged; the revision adds where it does so and how much robustness is best.

Reviewer_Fow1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper builds on the recent strategic risk-aversion framework and introduces the notion of risk-seeking equilibrium, where in contrast to risk-aversion (where one models a fictitious adversary), one models a fictitious ally. The paper then develops theory for the scalar two player bilinear quadratic static game setting, (i) developing equilibrium existence conditions for risk seeking and sharpening existing results for risk aversion, (ii) characterizing behavior of best response dynamics for risk-seeking/aversion equilibrium as a function of agent payoff parameters, and (iii) when risk-seeking/aversion outperform traditional Nash equilibrium payoffs. Finally, the paper verifies their theory on three scalar examples.

Critical strengths
  • The notion of risk-seeking is quite interesting and can be applicable to several multi-agent problems where agents are not necessarily adversarially aligned. This theoretical framework can help bring new perspectives to such scenarios.
Critical weaknesses
  • The paper only considers static scalar quadratic games. If I understand correctly, there should not be bottlenecks in extending the presentation to multi-dimensional problems and make the setting a bit more general.
  • With due respect, the paper can do a better job of motivating risk-seeking with specific applications in the introduction. Given that the notion of risk-seeking is a core contribution of this paper, this can help the paper appeal to researchers beyond the immediate games community. Currently, the introduction gives the impression that risk seeking simply is the mathematical complement to risk-aversion.
Questions for authors
  1. Can the results extend to non-scalar quadratic games, and potentially to linear quadratic dynamic games where the optimal control is known to be linear?
2 comments

We thank the reviewer for the positive assessment and constructive suggestions.

Motivation for risk seeking. We agree, and will add a paragraph to the introduction. The key point is that risk seeking is needed in practice, since many interactions have the opposite structure from those where risk aversion helps. What matters is whether one player's effort makes the other's effort more or less worthwhile, and this varies across applications, often within the same one:

  • Two AI agents sharing a task. If they do overlapping work, the more one agent does, the less the other's extra effort adds (strategic substitutes). If they build parts that only work together, such as code and the tests that check it, each agent's effort makes the other's more valuable (strategic complements).
  • Protecting a shared network. If security depends on the total effort, each party is tempted to rely on the others (substitutes). If it depends on the weakest link, one party's investment is wasted unless the other invests too (complements).
  • Supply chains. Two suppliers serving the same buyer reduce each other's returns, since the more one supplies, the less the other's extra output is worth (substitutes), while a supplier's extra capacity is only useful if the manufacturer builds matching capacity (complements).

Under strategic substitutes, agents overdo it at Nash, and existing work shows that risk aversion corrects this. Under strategic complements, agents underdo it, and our results show that risk aversion makes things worse: a pessimistic agent expects less from its partner and contributes even less. What helps there is risk seeking: an agent that expects slightly more from its partner contributes more, which in turn makes its partner's contribution more valuable, and can reach the best joint outcome at an explicit level λ⋆\lambda^\star. Since both kinds of interaction are common, a theory limited to risk aversion covers only half of them; risk seeking completes it.

Multi-dimensional games. With vector actions and matrices B,Γ,ΔB, \Gamma, \Delta, well-posedness and the affine best response carry over directly. When the matrices are symmetric and commute, the game separates exactly into scalar games of our form, and all our results apply direction by direction. In general, we conjecture that the pairing rule holds with the sign of γ\gamma replaced by the definiteness of Γ\Gamma: if the game has strategic substitutes in every direction (Γ≺0\Gamma \prec 0), risk aversion should help, and if it has complements in every direction (Γ≻0\Gamma \succ 0), risk seeking should help, while with mixed directions neither attitude needs to dominate. Convergence is harder: the reaction slope becomes a matrix whose eigenvalues need not move monotonically in λ\lambda, and different directions can stop converging at different λ\lambda, so our single-crossing arguments do not carry over. We will add this as a discussion and open problem.

LQ dynamic games. Velasevic et al. (2026) extend the risk-averse construction to LQ dynamic games, where equilibrium policies stay linear and solve coupled Riccati equations. Risk seeking fits the same machinery through our sign change, with a per-stage well-posedness condition analogous to Theorem 1. Whether the pairing rule survives over a horizon is open; we will state it as future work.

Thank you for your response, I have no more questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for their careful reading. We uploaded a revised version; all new or changed text is in blue, and the main text uses the additional page.

  1. Welfare theorem (amU5). Theorem 2 now states that, for the attitude matching the game, welfare increases up to an optimal weight λ⋆\lambda^\star, decreases after it, and falls below the Nash level beyond a second threshold λ0\lambda_0, both in closed form; the other attitude decreases welfare. All statements hold on an explicitly defined range (Definition 6). The examples report λ⋆\lambda^\star and λ0\lambda_0; for instance, in the trading example welfare peaks at λ=2\lambda = 2 and falls below Nash for λ>6\lambda > 6. Abstract, introduction and discussion are adjusted accordingly.
  2. Clarity (qZDz). Well-posedness is formally defined (Definition 5); Theorem 1 gives one condition per attitude, with KK, LL and the equilibrium in numbered equations. The former Propositions 1 to 5 are merged into a single proposition. Assumptions are stated once, the definition of substitutes and complements is simplified, and the proofs in the appendix are rewritten step by step, with all intermediate computations shown and one signed parameter for both attitudes. The proof of Theorem 2 now follows a single argument, based on where the equilibrium lies between Nash and the joint optimum. New figures support the statements:
    • Figure 1 shows where each attitude is well posed (Theorem 1);
    • Figure 2 shows the best-response slope, with the regions of existence and convergence (Proposition 1);
    • Figure 3 shows where welfare is above or below Nash, with λ⋆\lambda^\star, λ0\lambda_0 and the three examples (Theorem 2);
    • the example figures now mark λ⋆\lambda^\star and λ0\lambda_0, and a new example in the appendix (Example 1, Figure 6) illustrates a boundary case of Proposition 1.
  3. Motivation and scope (Fow1). The introduction explains why risk seeking is needed, since substitutes and complements often arise within the same domain (e.g., quantity versus price competition). A new paragraph in the Discussion covers vector actions (as a conjecture), dynamic linear-quadratic games, and empirical tests with learned agents.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.