{"id":"2492b137-fd28-47ed-9af6-48ca05d72a16","arxiv_id":"2507.15325","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Strategic robustness can be added to game theory using optimal transport ambiguity sets, yielding equilibria that interpolate between Nash and safety at no extra computational cost.","lead":"This paper introduces a new equilibrium concept, the strategically robust equilibrium, in which each player assumes the other players may deviate within an optimal-transport ball around their equilibrium play. The concept interpolates between Nash equilibria and security strategies, and the authors prove it exists and is computable under essentially the same conditions as Nash equilibrium.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the strong-duality dependence flagged by the reader is satisfied under the paper's compact-Polish/continuous-payoff assumptions, and the central existence and PPAD claims are internally consistent.","rationale":"The reader's verdict of ACCEPT is supported. The reader identified strong duality for Wasserstein distributionally robust optimization as the weakest assumption; this is a reasonable place to look, but under the paper's assumptions the duality hypotheses are satisfied. The action spaces are compact Polish, the transport cost (d_{-i})^s is lower semicontinuous because d_{-i} is a distance, and the expected-payoff loss is continuous and bounded. Thus the reformulation in Proposition 5 is sound, and the dependence of Theorem 2 and Theorem 3 on it is not a real vulnerability. I also checked the other central proof components: Theorem 1 follows from Glicksberg once ambiguity-set hemicontinuity and payoff continuity/concavity are established; Lemma 1 supplies the needed hemicontinuity for optimal-transport balls; Theorem 2's reduction to a concave game verifies continuity and concavity in each player's own strategy; and Theorem 3's use of the standard concave-game existence theorem is legitimate because the pointwise infimum of affine/concave robust payoffs preserves concavity in the player's own action. The only concrete imperfection I noticed is that Proposition 2's displayed KKT conditions omit the explicit simplex equality, but this is a presentation issue rather than a flaw in the underlying equivalence when p̄ is understood to be a strategy profile. I therefore see no reason to change the reader's verdict.","tokens_in":35224,"tokens_out":24339,"duration_ms":300854,"concrete_test":"As a verification step, re-derive Proposition 1 from finite-dimensional LP duality for a 2×2 game, bypassing the continuous Blanchet–Murthy theorem, and compare the resulting best-response value with the original max-min value by exhaustive enumeration of the simplex on a fine grid; if they match to numerical precision, the finite-action PPAD claim is independent of any continuous-duality concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I found no load-bearing defect in the central claims. The candidate weak point is the use of strong duality for the Wasserstein distributionally robust reformulation in Proposition 1/5. Under the paper's standing assumptions—compact Polish action spaces, continuous payoffs, and d_{-i} a distance—the transport cost (d_{-i})^s is lower semicontinuous and bounded, and the loss E_{a_i∼p_i} u_i(a_i, ·) is continuous and bounded. These are precisely the hypotheses of the Blanchet–Murthy duality theorem, so the dual reformulation is valid. For finite-action games the same step reduces to finite LP duality, so Theorem 2 does not inherit a continuous-duality risk. For concave games, the same conditions hold with pure centers δ_{a_-i}, and Berge's theorem plus concavity of the resulting robust payoff justify Theorem 3. The PPAD membership proof additionally checks concavity, compactness, well-boundedness, and Lipschitz representability of the surrogate payoff. I therefore see no threat to Corollary 1, Theorem 2, or Theorem 3. A minor presentation issue: Proposition 2's displayed KKT system omits the explicit simplex equality Σ_{a_i} p_i(a_i)=1, so the literal equation list would admit p=0 if read without the strategy-profile context; read as part of the LP for p_i∈Δ_i, the constraint is present. This is a notation/implementation nit, not a load-bearing flaw.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new equilibrium notion, called strategically robust equilibrium (SRE), in which each player maximizes expected payoff against the worst-case strategy profile contained in an optimal-transport ambiguity ball centered at the other players' equilibrium strategies. The main theoretical results are: (i) an existence theorem for general ambiguity sets under a hemicontinuity condition (Theorem 1), specialized to optimal-transport ambiguity sets under the same assumptions needed for mixed Nash equilibria (Corollary 1); (ii) a dual reformulation of the robust best response (Proposition 1); (iii) a PPAD membership result for computing approximate SRE in finite games (Theorem 2); (iv) an equivalent multilinear complementarity formulation, which becomes linear in two-player games (Proposition 2); and (v) existence of pure SRE in concave games together with an equivalent concave surrogate game (Theorem 3 and Proposition 3). The paper also provides numerical experiments on bi-matrix games, congestion games, and Cournot competition, illustrating robustness and the reported 'coordination via robustification' effect.","tokens_in":35523,"tokens_out":24761,"duration_ms":284478,"significance":"If the results hold, this is a substantial contribution: it bridges distributionally robust optimization and equilibrium analysis, and it shows that a robust non-Nash equilibrium notion can have the same existence guarantees and the same worst-case computational complexity as mixed Nash equilibrium. The proofs are detailed and use standard tools (Glicksberg's fixed point theorem, Berge's maximum theorem, Wasserstein DRO duality, and recent concave-game complexity results). The paper ships reproducible code for the numerical experiments, which is a concrete strength. I also checked the main technical risk identified by the reader—the dependence of Proposition 1/5 on strong duality for Wasserstein distributionally robust optimization—and under the paper's standing assumptions (compact Polish spaces, continuous payoffs, and a metric ground distance) the Blanchet-Murthy duality hypotheses are satisfied, so this is not a defect. The only free parameters are the radius epsilon and the order s; no ad-hoc assumptions are introduced.","major_comments":[],"minor_comments":[{"comment":"The displayed optimality system omits the primal feasibility constraint sum_{a_i} p_i(a_i)=1, so the literal condition would admit p_i=0. Please add the simplex equality to the statement of Proposition 2 or explicitly include 'together with p_i in Delta_i' in the equivalence, matching the LP (7) from which the KKT conditions are derived.","section":"Proposition 2, equations (8)-(9)"},{"comment":"In the proof of continuity of the product map, the second displayed integral is written as int phi_1 dmu_n again, but it should be int phi_2 dnu_n; this is a typographical error in an otherwise correct argument.","section":"Appendix A.2, proof of Lemma 2"},{"comment":"The constraint is stated as f_i^k(a_{-i}) <= 0, but the conjugate terms and the surrounding discussion require f_i^k to be a function of the agent's own action a_i (or of the relevant block a_j for j != i); please correct the index.","section":"Corollary 2"},{"comment":"The use of Fearnley et al. (2022, Theorem E.2) to approximate the surrogate payoff by a linear arithmetic circuit is only acknowledged through a footnote about an additional precision term; please state explicitly how the approximation accuracy is chosen relative to delta so that the final delta-approximate SRE guarantee in the chain of inequalities remains valid.","section":"Theorem 2 proof"},{"comment":"The type-s Wasserstein order s appears in the computational statements without an input convention; please specify that s is rational (or integer) so that quantities such as epsilon^s and d_{-i}(.,.)^s are polynomial-time computable in the bit complexity model.","section":"Theorem 2 and Lemma 3"}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a real contribution. The paper defines strategically robust equilibria as fixed points of best responses against Wasserstein ambiguity sets centered at the equilibrium itself, which cleanly interpolates between Nash and security strategies. The main theorems hold up: existence under the same assumptions as Nash, PPAD membership, and pure existence for concave games. I think the reader's skepticism about the strong-duality step is misplaced—the stress-test note is right that Blanchet–Murthy applies under their compact-Polish/continuous-payoff assumptions, and in finite games it just reduces to LP duality.\n\nWhat's genuinely new: the equilibrium notion itself, plus the observation that optimal transport balls are hemicontinuous in their center (Lemma 1), which opens the door to a fixed-point existence proof. The PPAD membership argument via concave games is non-obvious and carefully executed; Remark 2 honestly explains why the standard pure-deviation trick fails for SRE, which is a good sign they understand their own construction.\n\nWhat it does well beyond the theorems: the computational reformulations are useful, and the experiments, including the pedestrian and Cournot examples, are illustrative. The 'coordination via robustification' effect is presented as an empirical discovery, not a theorem, which is the right epistemic register.\n\nSoft spots, in proportion: the KKT system in Proposition 2 omits the explicit simplex constraint Σ p_i = 1. Read in isolation, the equation list would admit the zero vector. The context (the LP with p_i ∈ Δ_i) supplies the missing constraint, so it's cosmetic, but it should be fixed. The abstract's 'no additional computational cost' is a bit promotional: PPAD membership says no harder than Nash in complexity-class terms, but LCP solvers can still be nasty; a more precise phrasing would be 'same complexity class as Nash.' That's a wording issue, not a load-bearing flaw.\n\nBottom line: this paper deserves a serious referee. I'd send it out; the right venues are EC, Mathematics of OR, or Operations Research. Recommended: accept after minor revision. I'd be happy to cite it.","headline":"A genuinely new equilibrium concept with solid existence and PPAD membership; the strong-duality worry doesn't land, and the only real issue is a cosmetic gap in the displayed KKT conditions.","tokens_in":36021,"tokens_out":2131,"would_cite":true,"duration_ms":24083,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A10","91A80","49Q22","90C47"],"pacs":[],"model":"deepseek-v4-flash","headline":"Strategically robust equilibria exist wherever Nash equilibria do, and computing them is no harder than computing a Nash equilibrium.","keywords":["strategically robust equilibrium","optimal transport","Wasserstein ambiguity set","distributionally robust optimization","Nash equilibrium","security strategy","PPAD complexity","concave games"],"falsifier":"Take a two-player continuous-action game (e.g., compact intervals) with a ground cost d that satisfies the triangle inequality but is not lower semicontinuous, and for a fixed opponent strategy $p_{-i}$ compute both the primal value $\\min_{\\sigma \\in \\text{ball}} U_i(p_i, \\sigma)$ and the dual expression in Proposition 1; if the difference is nonzero, the dual reformulation fails and with it the arguments for existence and PPAD membership in the continuous setting. Alternatively, search over finite 3x3 games and a fine grid of $\\varepsilon$ for a game where the linear complementarity problem of Proposition 2 has no solution, which would contradict Corollary 1.","tokens_in":1903,"feed_emoji":"🛡️","tokens_out":4337,"duration_ms":70878,"temperature":0.7,"pith_summary":"The paper introduces a new equilibrium notion, the strategically robust equilibrium (SRE), in which each agent maximizes expected payoff against the worst-case distribution over others' actions within a tunable Wasserstein ball centered at the equilibrium itself. The paper claims that SREs exist under exactly the same assumptions that guarantee mixed Nash equilibria, that computing a delta-approximate SRE lies in PPAD (no harder than Nash), and that for concave games a pure SRE always exists. These results matter because they suggest decision-makers can hedge against misspecification, bounded rationality, or out-of-equilibrium play at no extra computational cost, and experiments show the robust equilibria often yield higher payoffs for all agents.","feed_headline":"Robust play against uncertainty costs no more than Nash","feed_subtitle":"A Wasserstein ball around equilibrium play gives worst-case protection, stays in PPAD, and often raises payoffs for all players.","key_machinery":"The central object is the optimal-transport (Wasserstein) ambiguity set $B_i^\\varepsilon(p_{-i}) = \\{\\sigma_{-i} : W_s(\\sigma_{p_{-i}}, \\sigma_{-i}) \\le \\varepsilon\\}$, a ball of distributions around the product of the others' mixed strategies. It interpolates between Nash equilibria ($\\varepsilon=0$) and security strategies ($\\varepsilon\\to\\infty$) while remaining hemicontinuous (Lemma 1), which is what guarantees existence under minimal assumptions. The argument then uses strong duality for Wasserstein distributionally robust optimization to reformulate the strategically robust best response as a single convex program (Proposition 1), and reduces the SRE problem to a concave game so that known PPAD complexity results for concave games apply.","core_discovery":"The authors prove that strategically robust equilibria based on optimal transport ambiguity sets exist under the same assumptions as mixed Nash equilibria (Corollary 1), and that computing a delta-approximate SRE is in PPAD, so it is no harder than computing a Nash equilibrium (Theorem 2). For concave games, pure SREs exist (Theorem 3) and can be computed as Nash equilibria of a surrogate concave game with augmented action space. The key step is dualizing the worst-case distribution in a Wasserstein ball, which turns each agent's max-min best response into a single convex program; this dual also reveals that robustness acts like a regularization of payoffs, which in examples leads to coordination and higher equilibrium payoffs for every agent.","pith_inferences":["The family of SREs parameterized by $\\varepsilon$ defines a homotopy from Nash to security equilibria, which could serve as an equilibrium-selection device in games with multiple Nash equilibria.","Because robustness behaves like payoff regularization, SREs could be used to design taxes or prices in congestion games that improve efficiency without needing an accurate model of player behavior.","The duality-based reformulation relies on strong duality for Wasserstein DRO; for continuous action spaces the paper applies it without verifying all technical conditions, and a nonzero duality gap would break the 'no extra computational cost' claim.","The coordination-via-robustification effect is demonstrated numerically; a theoretical characterization of game classes in which robustness raises all players' payoffs would be a natural next step."],"forward_implications":["Wherever mixed Nash equilibria exist, strategically robust equilibria exist too, so robustness can be added without losing existence guarantees.","Computing a $\\delta$-approximate strategically robust equilibrium is in PPAD, so algorithms and solvers used for Nash equilibria (e.g., LCP, homotopy methods) can be reused.","The tunable parameter $\\varepsilon$ lets an agent or designer move continuously between Nash play and security play, giving a principled way to choose the level of worst-case protection.","In concave games, pure SREs exist and robustness appears as a regularization term in payoffs, e.g., inflating marginal costs in Cournot competition.","In experiments, strategic robustness often raises payoffs for all agents (coordination via robustification) and lowers the price of anarchy in congestion games."],"supporting_citations":[{"why":"Supplies the strong-duality theorem for Wasserstein distributionally robust optimization that converts each agent's max-min best response into a single convex program.","marker":"Blanchet and Murthy 2019"},{"why":"Provides the PPAD complexity result for concave games that the paper uses to show SRE computation is no harder than Nash.","marker":"Papadimitriou et al. 2023"},{"why":"Establishes existence of pure Nash equilibria in concave games, which the SRE pure-existence result builds on.","marker":"Rosen 1965"},{"why":"Supplies the fixed-point theorem used in the existence proof for strategically robust equilibria.","marker":"Glicksberg 1952"},{"why":"Provides foundational definitions and properties of optimal transport and Wasserstein distances used to construct the ambiguity sets.","marker":"Villani 2009"},{"why":"Gives the convex reformulation of the surrogate payoff used in Corollary 2 for concave games.","marker":"Zhen et al. 2025, Theorem 2"},{"why":"Establishes PPAD-completeness of Nash equilibrium, serving as the baseline for the paper's PPAD-membership claim.","marker":"Daskalakis et al. 2009"},{"why":"Shows how to approximate the Lipschitz payoff by a linear arithmetic circuit, a step in the proof of Theorem 2.","marker":"Fearnley et al. 2022, Theorem E.2"}],"fun_headline_variants":["Robust play: no extra cost, often better payoffs","Robust equilibria: Nash cost, often higher payoff","Worst-case play via optimal transport: no extra cost vs Nash","Robust game theory: same complexity as Nash, often better outcomes"],"cache_read_input_tokens":38144,"weakest_assumption_plain":"The central reformulation assumes strong duality holds for the Wasserstein distributionally robust optimization problem inside each agent's best response, requiring the ground cost to be lower semicontinuous and the action space to be Polish; the paper applies this to continuous action spaces without fully verifying these conditions, and a failure of duality would break the existence-by-same-assumptions and no-extra-cost claims.","fun_headline_variants_meta":{"raw":{"variants":["Robust play: no extra cost, often better payoffs","Robust equilibria: Nash cost, often higher payoff","Worst-case play via optimal transport: no extra cost vs Nash","Robust game theory: same complexity as Nash, often better outcomes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3128,"prompt_tokens":934,"completion_tokens":2194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":2122}},"tokens_in":550,"tokens_out":2194,"duration_ms":20488,"temperature":1.0,"reasoning_tokens":2122,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:36:07.210625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-player continuous-action game (e.g., compact intervals) with a ground cost d that satisfies the triangle inequality but is not lower semicontinuous, and for a fixed opponent strategy $p_{-i}$ compute both the primal value $\\min_{\\sigma \\in \\text{ball}} U_i(p_i, \\sigma)$ and the dual expression in Proposition 1; if the difference is nonzero, the dual reformulation fails and with it the arguments for existence and PPAD membership in the continuous setting. Alternatively, search over finite 3x3 games and a fine grid of $\\varepsilon$ for a game where the linear complementarity problem of Proposition 2 has no solution, which would contradict Corollary 1.","supporting_citations":[{"cited_title":"Mathematics of Operations Research 44(2):565--600","cited_arxiv_id":null,"evidence_quote":"Supplies the strong-duality theorem for Wasserstein distributionally robust optimization that converts each agent's max-min best response into a single convex program."},{"cited_title":"Proceedings of the 24th ACM Conference on Economics and Computation, 1045, EC '23 (New York, NY, USA: Association for Computing Machinery)","cited_arxiv_id":null,"evidence_quote":"Provides the PPAD complexity result for concave games that the paper uses to show SRE computation is no harder than Nash."},{"cited_title":"Econometrica: Journal of the Econometric Society 520--534","cited_arxiv_id":null,"evidence_quote":"Establishes existence of pure Nash equilibria in concave games, which the SRE pure-existence result builds on."},{"cited_title":"Proceedings of the American Mathematical Society 3(1):170--174","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-point theorem used in the existence proof for strategically robust equilibria."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides foundational definitions and properties of optimal transport and Wasserstein distances used to construct the ambiguity sets."},{"cited_title":"Operations Research 73(2):862--878","cited_arxiv_id":null,"evidence_quote":"Gives the convex reformulation of the surrogate payoff used in Corollary 2 for concave games."},{"cited_title":"Communications of the ACM 52(2):89--97","cited_arxiv_id":null,"evidence_quote":"Establishes PPAD-completeness of Nash equilibrium, serving as the baseline for the paper's PPAD-membership claim."},{"cited_title":"Journal of the ACM 70(1):1--74","cited_arxiv_id":null,"evidence_quote":"Shows how to approximate the Lipschitz payoff by a linear arithmetic circuit, a step in the proof of Theorem 2."}],"review_version":1}