{"id":"db027a49-db1b-4d9e-8b64-c9156eab3f0a","arxiv_id":"2506.15589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A structured Koopman operator formulation for multi-agent systems with hierarchical and time-scale-separated dynamics, enabling comparison of socially optimal and Nash equilibrium control policies.","lead":"This paper builds a Koopman operator framework for multi-agent control, where each agent's nonlinear dynamics are lifted into a linear space and interactions between agents appear as explicit coupling terms. It then compares two ways of choosing control policies, one central planner and one where each agent plays a game, and quantifies the gap between them.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6) is not the Nash condition for the coupled game: it omits cross-agent adjoint terms, so the claimed optimum-vs-Nash difference is actually a centralized-vs-decoupled difference.","rationale":"Good faith: the paper builds a genuinely structured multi-agent Koopman formulation, and the algebra of Eqs. (6)-(7) is a clean observation about the lifted linear model. The reader's concern about finite-dimensional Koopman invariance is real but generic; my concern sits one level deeper. Even if the lifted model were exact, the equilibrium condition (6) is derived from a problem in which each agent optimizes its own dynamics row while treating the other agents' lifted states as exogenous. For the coupled game (1)-(2), that is not Nash: in an open-loop Nash equilibrium each player's optimization must respect all coupling because other agents' state trajectories are endogenous and respond to the player's control through the state dynamics. The missing cross term in (6) is precisely the term that couples the agents' adjoint equations in a correct Nash formulation, so the paper's central comparison (Eq. 6 vs Eq. 7) describes the gap between a social optimum and a decoupled best-response fixed point, not the optimum-vs-Nash gap claimed in the abstract. The proposed LQ test would settle whether Eq. (6) is recoverable from the full coupled Lagrangian; if not, the paper needs either a correct Nash derivation or an explicit redefinition of the equilibrium concept, and the results need to be relabeled accordingly. This is a more fundamental correction than the reproducibility issues the reader identified, though it does not invalidate the multi-agent Koopman construction itself.","tokens_in":10300,"tokens_out":19270,"duration_ms":201085,"concrete_test":"Take the two-agent lifted scalar LQ example with x1' = a x1 + c x2 + u1, x2' = d x1 + a x2 + u2 and costs x_i^2 + u_i^2. Derive the open-loop Nash stationarity condition with respect to x1 by forming each player's Lagrangian with both dynamics rows included. If the condition contains d times the multiplier on the x2 row (the analog of λ_j^T(I−K_jj)K_ji), then Eq. (6) is not the Nash condition. Optionally, solve that full two-player MCP and compare the resulting policies with the paper's Eq. (6)-based 'equilibrium' policies.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Within the lifted model, the algebra leading to Eq. (7) is sound, but the object called the Nash equilibrium is not the Nash equilibrium of the coupled game defined by (1)-(2). Eq. (6) is the stationarity condition for an optimization in which only the agent's own row of (3) appears as a constraint: the other agents' lifted states are treated as exogenous inputs, so no adjoint variables from the other dynamics rows enter. In a standard open-loop Nash equilibrium of the coupled system, each player must include the full stacked dynamics (4) as constraints, because every state trajectory is endogenous and depends on all controls through feedback. The stationarity condition with respect to ψx_i,t would then contain contributions of the form Σ_{j≠i} λ_j^{(k),T}(I−Kxx_jj)Kxx_ji from that player's multipliers on the other agents' dynamics rows, exactly the term omitted from Eq. (6). Consequently, Eq. (6) describes a fixed point of decoupled, coupling-blind best responses, not a Nash equilibrium of the original dynamic game. The numerical MCP appears to implement this decoupled version, so the reported 'equilibrium' policies should not be described as Nash equilibria of the multi-agent system.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a finite-dimensional Koopman-operator representation for multi-agent systems in which each agent's lifted state evolution is written with explicit cross-agent coupling terms (Eq. (3)), and extends this representation to systems with hierarchical control and time-scale separation (Eqs. (12)-(14)). For both settings, the authors formulate a centralized optimal control problem and a purportedly noncooperative game-theoretic problem solved through a mixed complementarity problem (MCP). The central analytical claim is that the centralized social optimum and the Nash equilibrium optimality conditions differ only by the term sum_{j != i} lambda_j^T (I - Kxx_jj) Kxx_ji (Eq. (7) vs. Eq. (6)), which is interpreted as quantifying the coupling-driven divergence between the two solution concepts. A two-agent numerical example with nonlinear oscillators is used to illustrate the difference.","tokens_in":10595,"tokens_out":17748,"duration_ms":169496,"significance":"If the identification of Eq. (6) with the Nash equilibrium were correct, the paper would provide a clean way to use Koopman lifting in multi-agent dynamic games and to quantify the gap between cooperative and noncooperative outcomes. The structured lift with per-agent observables and explicit cross-agent blocks is natural and potentially useful, and the hierarchical/multi-scale extension is a sensible generalization of the author's prior work. The algebraic derivation of the centralized optimality condition (Eq. (7)) is straightforward, and the paper gives a reproducible computational pipeline (Pyomo/IPOPT/PATH) while honestly reporting model errors. However, the central game-theoretic interpretation is, in my assessment, not correct: Eq. (6) corresponds to a coupling-blind, decoupled best-response problem, not to the Nash equilibrium of the coupled dynamic game. The comparison between Eq. (7) and Eq. (6) is therefore a comparison between centralized optimization and a decentralized, coupling-ignoring equilibrium, not between a social optimum and a Nash equilibrium.","major_comments":[{"comment":"Equation (6), called the 'equilibrium optimality condition', is not the Nash equilibrium condition for the coupled dynamic game defined by (1)-(2) or by its exact Koopman lift (4). In an open-loop Nash equilibrium, each agent i optimizes its own objective with respect to its control sequence subject to the entire coupled dynamics, taking the other agents' control sequences as given. The KKT conditions for agent i's problem must therefore include adjoint equations for all stacked states, and the adjoint equation for psi_x_i,t contains terms of the form sum_{j != i} lambda_{i,j,t}^T (I - Kxx_jj) Kxx_ji, where lambda_{i,j,t} is agent i's multiplier on agent j's dynamics row. Equation (6) contains no such terms; it is the stationarity condition for a single-agent optimal control problem in which the other agents' lifted states are treated as exogenous inputs and only the agent's own row of (3) is a dynamic constraint. Consequently, the MCP solved in Section IV characterizes a fixed point of coupling-blind, decoupled best responses, not a Nash equilibrium of the coupled system. The authors should either re-derive the equilibrium conditions from the stacked dynamics (4) with a separate multiplier vector per agent, or explicitly reframe the contribution as a comparison between centralized optimization and a decoupled decentralized solution and avoid the term 'Nash equilibrium'. This is load-bearing because the paper's central claim that the optimum and equilibrium conditions differ by a single cross-coupling term depends on Eq. (6) being the Nash condition.","section":"Sec. II-A, Eqs. (4), (6), (7)"},{"comment":"The numerical demonstration does not support the game-theoretic conclusions as stated. Because the equilibrium object solved in the MCP is the decoupled fixed point described in the previous comment, the reported optimum/equilibrium differences (average RMS less than 0.005 in Sec. IV-A, roughly 0.02 in Sec. IV-B) measure a difference between centralized and decoupled solutions, not the social-optimum-versus-Nash gap. In addition, the epsilon->0 model used for control in the hierarchical case has mean RMS prediction errors of 0.152 (Agent 1) and 0.175 (Agent 2), which are not negligible relative to the reported policy differences, and no error bars or sensitivity analysis are provided for any learned-model result. The paper's own caveat that the stability analysis 'should be taken with a larger grain of salt' applies equally to the control results. The quantitative claims about agent effort/benefit asymmetries should be re-examined after correcting the equilibrium definition, or softened.","section":"Sec. IV, Tables I-II"}],"minor_comments":[{"comment":"The stated value 'epsilon = 100' is inconsistent with the epsilon -> 0 limit used throughout Sec. II-B and with the identification of y and w as fast-scale variables; if epsilon is meant to be small, the displayed value appears to be a typo (e.g., epsilon = 0.01), and if epsilon = 100 is intended, the time-scale labels 'slow' and 'fast' in the numerical experiment are reversed.","section":"Sec. III-A, text after Eq. (38)"},{"comment":"The indexing in the quadratic cost approximation is difficult to parse: the notation [psi^i_{k,t}]^T Q_{k,ij} psi^j_{k,t} with i,j in {u,x} should be defined more explicitly, including the dimensions of the Q_{k,ij} blocks and the ordering of the indices, to avoid confusion with the cross-coupling matrices Kxx_{ij}.","section":"Sec. II-A, Eq. (5)"},{"comment":"There is a doubled word in 'and and x_{-i}' near Eq. (8), and a similar typo appears in Sec. III-A ('the the k-th variable'); a careful proofread of the manuscript is needed.","section":"Sec. II-B, text near Eq. (8)"},{"comment":"The Lyapunov equation is written with bare symbols Kii, Kij, and I; these should be decorated (e.g., Kxx_ii, Kxx_ij) to match the notation elsewhere in the paper and to avoid confusion between the identity matrix I and the index i.","section":"Sec. II-C, Eq. (26)"},{"comment":"The notation Pmax,i (elsewhere P_{max,i}) and the phrase 'evaluating the fixed k solution over a sufficiently large range of k values' are vague; please state precisely how the maximum over k is computed in the SVD-based procedure.","section":"Sec. II-C, Eqs. (27)-(28)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript depends heavily on the author's own unpublished work ([12], [14]) for two of its building blocks, which makes independent verification of some components difficult. The main issue, however, is the game-theoretic mislabeling: the 'Nash equilibrium' solved in the MCP is actually a decoupled best-response fixed point rather than the Nash equilibrium of the coupled dynamic game. This is fixable either by re-deriving the equilibrium conditions from the stacked dynamics or by reframing the paper as a centralized-vs-decoupled comparison, but it is central to the contribution. I would not recommend rejection, as the structured Koopman lifting formulation and the centralized-vs-decentralized comparison have potential value; however, the revision needs to be substantial rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: the paper's main game-theoretic claim is mislabeled. Equation (6), presented as the Nash equilibrium optimality condition, is not the stationarity condition for a Nash equilibrium of the coupled game. In an open-loop Nash equilibrium, each player optimizes subject to the full stacked dynamics (4), which produces cross-agent adjoint terms of the form sum_{j≠i} λ_j^T (I−Kxx_jj) Kxx_ji. Those terms appear in equation (7) but are absent from equation (6). So equation (6) describes a decoupled best-response equilibrium, where each agent treats the other agents' states as exogenous inputs. The paper compares a centralized optimum against that decoupled fixed point, not against a genuine Nash equilibrium of the coupled system. This is a load-bearing flaw, because the paper's central stated result is the difference between the social optimum and the Nash equilibrium.\n\nThat said, the paper does something useful. The explicit multi-agent Koopman lifting with cross-agent coupling blocks (Kxx_ij) is a clean structuring of the problem, and the algebra leading to the correct centralized condition is straightforward. The hierarchical/time-scale extension is a systematic combination of the author's earlier single-agent work with the multi-agent structure; it extends naturally and the averaged fast-scale representation is sensible. The proposed analytical tools (transient growth metrics, controllability gramians) are reasonable ways to quantify coupling and cross-scale feedback.\n\nThe numerical demonstration is the soft part. The models are learned with no error bars, and the stability enforcement used in training biases the later stability metrics; the authors do acknowledge this partially, but it should be stated more plainly. The reported 'equilibrium' policies would likely change if the true Nash conditions were imposed, and the claim that the optimum and equilibrium are nearly identical might not survive. The comparison of agent effort and benefit is interesting but is about the decoupled solution concept, not Nash.\n\nThe math in Sections II-A and II-B is internally consistent. The citation pattern is also reasonable; self-citations are to prior building blocks. The central conceptual error is what needs fixing. I would send this to peer review, because the structural lifting and the derivation of the correct centralized conditions are worth having, and correcting the Nash conditions is achievable either by deriving the true open-loop Nash conditions or by explicitly relabeling the concept as a decoupled best-response equilibrium and studying its relationship to Nash. As is, the paper would mislead readers who take the game-theoretic claims at face value.","headline":"The Koopman multi-agent lifting is worth having, but the paper calls a decoupled best-response equilibrium 'Nash' and misses the cross-agent adjoint terms, so the central comparison is mislabeled.","tokens_in":11122,"tokens_out":7261,"would_cite":true,"duration_ms":72281,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37N35","91A10","93C10","37M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that in a Koopman-lifted multi-agent system, the social optimum and the Nash equilibrium differ by exactly one cross-agent coupling term, making the source of the gap explicit.","keywords":["multi-agent systems","Koopman operator","optimal control","game theory","Nash equilibrium","time scale separation","hierarchical control","mixed complementarity problem"],"falsifier":"Build a two-agent system whose coupling is strong and nonlinear enough that no finite Koopman dictionary closes exactly, lift it with a finite set of observables, and compute the difference between the centralized optimum and the Nash equilibrium predicted by Eq. (7) versus Eq. (6). Then solve the true social optimum and Nash equilibrium of the original nonlinear game numerically; if the predicted gap is not close to the true gap, the claim that the coupling term carries the divergence is falsified.","tokens_in":10083,"feed_emoji":"🎮","tokens_out":7795,"duration_ms":66947,"temperature":0.7,"pith_summary":"This paper develops a Koopman-operator formulation of multi-agent nonlinear dynamics in which each agent's state is lifted through its own observables and every other agent's state enters as an input. The structured lift yields a finite-dimensional linear model with explicit cross-agent coupling matrices $K^{xx}_{ij}$. The author then writes the optimality conditions for centralized control (the social optimum) and for general-sum game play (the Nash equilibrium) on the same lifted model. The two conditions are identical except for the term $\\sum_{j\\ne i}\\lambda_{j,t}^T(I-K^{xx}_{jj})K^{xx}_{ji}$, so the gap between the two solution concepts is pinned to the cross-agent couplings. A two-agent numerical demonstration shows the gap is small but nonzero and that agent effort and payoff split asymmetrically.","feed_headline":"A single coupling term sets the gap between Nash play and the optimum","feed_subtitle":"A structured Koopman lift of multi-agent dynamics makes the divergence explicit and diagnosable.","key_machinery":"The central object is the structured multi-agent Koopman lift of Eq. (3): each agent has its own observable set $\\psi_i^x$ and the effect of other agents is folded in as an input through matrices $K^{xx}_{ij}$, with control entering through $K^{xu}_i$. The companion piece is the mixed complementarity problem (MCP) formed from the lifted dynamics and the agents' quadratic costs, whose stationarity conditions are compared with those of the centralized quadratic program to expose the coupling term. The machinery works by making cross-agent terms explicit enough to be analyzed with linear-systems tools such as transient growth bounds, controllability gramians, and SVD-based perturbation maxima.","core_discovery":"The central discovery is a multi-agent generalization of the Koopman lift in which the dynamics of agent $i$ are written as $\\psi_i^x(x_{i,t+1}) = K^{xx}_{ii}(\\psi_i^x(x_{i,t}) - \\sum_{j\\ne i}K^{xx}_{ij}\\psi_j^x(x_{j,t}) - K^{xu}_i\\psi_i^u(u_{i,t})) + \\sum_{j\\ne i}K^{xx}_{ij}\\psi_j^x(x_{j,t}) + K^{xu}_i\\psi_i^u(u_{i,t})$. This structure keeps each agent's own dynamics on the diagonal block $K^{xx}_{ii}$ and treats every other agent's lifted state as an additive input. Comparing the Nash first-order condition (Eq. 6) with the centralized optimum condition (Eq. 7), the only difference is the term $\\sum_{j\\ne i}\\lambda_{j,t}^T(I-K^{xx}_{jj})K^{xx}_{ji}$. The paper establishes that, in this lifted linear model, all divergence between decentralized strategic play and a centrally computed optimum is carried by the cross-agent coupling matrices $K^{xx}_{ji}$, and it extends the same comparison to the $\\epsilon\\to 0$ limit of hierarchical, time-scale-separated multi-agent systems.","pith_inferences":["The norm of $\\sum_{j\\ne i}\\lambda_j^T(I-K^{xx}_{jj})K^{xx}_{ji}$ could serve as a practical diagnostic: when it is small, solving the cheaper centralized optimum and using it as a warm start for the game would save the hour-or-more MCP solve the paper reports.","Because the gap term is weighted by the dual variables $\\lambda_j$, the optimum-versus-Nash divergence is not a fixed property of the dynamics but depends on the cost structure and constraints, so the same physical coupling can be strategically aligned or misaligned by design.","The same lifted model should support Stackelberg (leader-follower) equilibria as mathematical programs with equilibrium constraints, which the author notes but does not develop beyond the Nash case.","In the numerical examples, one agent consistently spends more control effort and still does worse than the other; reading which couplings dominate the $K^{xx}_{ij}$ matrices would allow the system designer to rebalance incentives."],"forward_implications":["If the finite-dimensional Koopman model is accurate, the same lifted matrices $K^{xx}_{ij}$ serve both centralized and game-theoretic control, making the two solution concepts computationally comparable at the same fidelity.","When $(I-K^{xx}_{jj})K^{xx}_{ji}=0$ for all $j\\ne i$, the optimum and equilibrium optimality conditions coincide, so agents that do not affect one another in the lifted dynamics will reach the social optimum by playing their own game.","In the $\\epsilon\\to 0$ limit of hierarchical, time-scale-separated systems, the effective slow-scale matrices $B^{xx}_i$ and $B^{xu}_i$ retain the same coupling structure, so the optimum-versus-Nash comparison carries over with a substantially smaller optimization problem.","The transient growth metrics and controllability gramians computed from the structured model identify when cross-agent feedback amplifies disturbances and when coupling makes the combined system more controllable than either agent separately."],"supporting_citations":[{"why":"Establishes the Koopman operator framework: nonlinear dynamics become linear on the space of observables.","marker":"[1]"},{"why":"Surveys finite-dimensional Koopman approximation, the basis for the matrices used here.","marker":"[2]"},{"why":"Shows how controlled Koopman models feed linear/quadratic MPC, which the multi-agent formulation extends.","marker":"[3]"},{"why":"Provides the prior single-agent hierarchical control and time scale separation formulation that Section II-B extends to multi-agent systems.","marker":"[14]"},{"why":"Supplies the stability-assuming (or enforcing) Koopman formulation modified in Eq. (3) to separate agents.","marker":"[15]"},{"why":"Defines mixed complementarity problems, the mechanism used to combine agents' optimality conditions for the Nash equilibrium.","marker":"[16]"}],"fun_headline_variants":["Coupling term alone defines Nash-optimum gap","One coupling term separates Nash and social optimum","Koopman lift: one term sets Nash gap","Nash vs optimum gap reduces to a single coupling term","Cross-agent coupling fixes Nash-optimum gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that the chosen finite set of lifted coordinates (observables) exactly represents the true nonlinear dynamics over the operating region; when that representation is only approximate, the computed optimum and Nash solutions are solutions of the approximate model, not of the original multi-agent game.","fun_headline_variants_meta":{"raw":{"variants":["Coupling term alone defines Nash-optimum gap","One coupling term separates Nash and social optimum","Koopman lift: one term sets Nash gap","Nash vs optimum gap reduces to a single coupling term","Cross-agent coupling fixes Nash-optimum gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000999,"raw_usage":{"total_tokens":4275,"prompt_tokens":1037,"completion_tokens":3238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":3164}},"tokens_in":653,"tokens_out":3238,"duration_ms":23173,"temperature":1.0,"reasoning_tokens":3164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:32:03.731545+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a two-agent system whose coupling is strong and nonlinear enough that no finite Koopman dictionary closes exactly, lift it with a finite set of observables, and compute the difference between the centralized optimum and the Nash equilibrium predicted by Eq. (7) versus Eq. (6). Then solve the true social optimum and Nash equilibrium of the original nonlinear game numerically; if the predicted gap is not close to the true gap, the claim that the coupling term carries the divergence is falsified.","supporting_citations":[{"cited_title":"Time scale separation and hierarchical control with the koopman operator,","cited_arxiv_id":null,"evidence_quote":"Provides the prior single-agent hierarchical control and time scale separation formulation that Section II-B extends to multi-agent systems."},{"cited_title":"A tutorial review of complementarity models for decision-making in energy markets,","cited_arxiv_id":null,"evidence_quote":"Defines mixed complementarity problems, the mechanism used to combine agents' optimality conditions for the Nash equilibrium."}],"review_version":2}