price-of-anarchy analysis

This work demonstrates that the static equilibrium and Price of Anarchy (PoA) framework overlooks dynamic imbalances inherent in multi-agent learning, leading to fragile efficiency guarantees. Adopting a dynamical systems perspective and integrating differential geometry, game theory, and regret minimization, the study reveals that Nash, correlated, and coarse correlated equilibria are typically unstable saddle points or even globally repelling sets in the underlying topology. It proves that under positive affine costs, the PoA is unbounded and that discrete-time learning dynamics can induce Li-Yorke chaos. Moreover, classical PoA bounds fail to hold under dynamic learning, as time-averaged efficiency deteriorates exponentially with the polynomial degree—scaling as $2^p$—and even swap-regret minimizing algorithms may generate macroscopic chaotic behavior.

This study addresses a fundamental limitation in traditional price-of-anarchy (PoA) analysis: its reliance on interpersonal utility comparability, which compromises invariance of efficiency measures under monotonic transformations of utility representations. To resolve this, the authors work within the cardinal non-comparable (CNC) utility framework and introduce weighted Nash welfare as a normative aggregation function. They propose, for the first time, a multiplicative smoothness condition compatible with the multiplicative structure of Nash welfare, enabling a PoA analysis that preserves representation invariance. By leveraging multiplicative envelope and geometric closure techniques, they establish a streamlined proof framework for singleton welfare games, naturally extending the results to coarse correlated equilibria and no-regret learning dynamics. The analysis yields invariant PoA bounds, revealing that the true efficiency loss in decentralized systems fundamentally hinges on assumptions about utility comparability.

Traditional Price of Anarchy (PoA) relies on precise cost values, yet in socio-technical systems (e.g., traffic control), costs often merely represent agents’ ordinal preferences—rendering their scale and zero point arbitrary. Consequently, PoA is not invariant under affine cost transformations, undermining its robustness. Method: We introduce Invariant PoA, the first efficiency metric for mechanisms grounded in social choice theory, axiomatically defined to be fully invariant under affine cost transformations. Leveraging a comparability-ranking framework, we characterize the class of social welfare functions satisfying this invariance and prove their uniqueness. Results: Empirical evaluation on real-world networks—including Zurich’s traffic system—demonstrates that conventional PoA exhibits estimation errors up to several-fold across different cost representations, whereas Invariant PoA yields consistent, robust welfare assessments. This provides a theoretically sound and practically applicable efficiency benchmark for policy design.

This paper investigates the intrinsic connection between social welfare (efficiency) and the computational tractability of Nash equilibria. For large-scale games, low-sensitivity games, and two-player zero-sum games, it introduces a unified analytical framework based on smoothness—establishing, for the first time, its equivalence to the Minty property in optimization. Methodologically, the approach integrates no-regret learning, clairvoyant mirror descent, and game-optimization interplay techniques to achieve fast, decentralized convergence to Nash equilibria. The main contributions are threefold: (1) It proves that strong smoothness simultaneously ensures both approximate optimality of social welfare and global convergence to coarse correlated equilibria; (2) it derives a tighter lower bound on social welfare than classical smoothness guarantees; and (3) it achieves, for the first time across broad classes of practical games, dual optimality—i.e., both computational efficiency in equilibrium computation and provable efficiency guarantees for social welfare.

This study addresses the inefficiency of the standard proportional allocation mechanism in online auto-bidding advertising, which incurs a price of anarchy (PoA) of up to 2 with respect to liquid welfare under pure Nash equilibria. The paper establishes, for the first time, a tight theoretical bound of PoA = 2 for this mechanism and introduces a novel payment rule that leverages duality theory and KKT conditions to refine equilibrium analysis. This improved mechanism reduces the PoA to $1 + \frac{O(1)}{n-1}$, asymptotically approaching full efficiency (PoA → 1) as the number of participants grows. By surpassing the theoretical limitations of traditional proportional mechanisms, the proposed approach significantly enhances overall system efficiency and demonstrates strong generality and practical potential.

This study addresses the verification of correlated equilibria and their subgame-perfect refinements in concurrent reachability games. Motivated by the safety and scalability requirements of multi-agent systems, it introduces—for the first time—the notion of subgame-perfect correlated equilibrium and conducts a systematic investigation integrating formal verification, game-theoretic analysis, computational complexity theory, and compact representations via Bayesian networks. The main contribution lies in establishing that, under standard representations, verifying correlated equilibria is P-complete, whereas its subgame-perfect refinement admits a solution in O(log²n) space, revealing a counterintuitive complexity separation. However, this gap vanishes when equilibria are represented using Bayesian networks, where both problems exhibit comparable computational complexity.

This work addresses the limitation of traditional equilibrium concepts, which only guard against unilateral deviations and fail to account for profitable coordinated deviations by coalitions. We propose a novel framework of multilateral stable equilibria that explicitly quantifies coalition deviation incentives as an optimizable objective. By minimizing the average (or weighted average, or maximum) gain achievable by any deviating coalition, our approach ensures both the existence and computability of equilibria. Integrating game-theoretic modeling, complexity lower-bound analysis, and matching algorithm design, we establish computational hardness results for both average- and maximum-gain objectives and present efficient algorithms that match these lower bounds. The framework is successfully applied to compute the Exploitability Welfare Frontier—the maximal social welfare attainable under a given level of exploitability.