Pareto-optimal Nash equilibrium in capacity allocation game for self-managed networks

In this paper we introduce a capacity allocation game which models the problem of maximizing network utility from the perspective of distributed noncooperative agents. Motivated by the idea of self-managed networks, in the developed framework decision-making entities are associated with individual transmission links, deciding on the way they split capacity among concurrent flows. An efficient decentralized algorithm is given for computing strongly Pareto-optimal strategies, constituting a pure Nash equilibrium. Subsequently, we discuss the properties of the introduced game related to the Price of Anarchy and Price of Stability. The paper is concluded with an experimental study.

💡 Research Summary

The paper introduces a novel game‑theoretic framework for distributed capacity allocation in self‑managed networks, where each physical link is modeled as an autonomous decision‑making agent. Unlike traditional centralized network management, this approach assumes that links independently decide how to split their finite capacity among multiple concurrent flows that traverse them. The authors formalize the “capacity allocation game” by defining players (links), strategies (vectors of capacity shares for each flow on a link), and utilities derived from the end‑to‑end flow rates using the standard α‑fair utility family.

A key theoretical contribution is the proof that the game is a potential game. Consequently, a pure Nash equilibrium (NE) is guaranteed to exist. However, not every NE is socially efficient; some may be Pareto‑inefficient. To address this, the authors define a subset of strategies that are strongly Pareto‑optimal and show that equilibria confined to this subset retain the NE property while achieving maximal collective welfare.

The paper then proposes a fully decentralized algorithm that converges to a strongly Pareto‑optimal NE. Each link repeatedly performs a gradient‑ascent update on its local utility, using only information exchanged with neighboring links (current capacity shares and the resulting flow rates). The update step is projected onto the feasible set defined by the link’s capacity constraint, and a diminishing step‑size schedule guarantees convergence. By invoking the Karush‑Kuhn‑Tucker (KKT) conditions and Lagrange multipliers, the authors demonstrate that the limit point satisfies the optimality conditions of the global utility maximization problem, thereby bridging the gap between non‑cooperative behavior and system‑wide efficiency.

Beyond existence and convergence, the authors analyze the price of anarchy (PoA) and price of stability (PoS) of the game. PoA measures the worst‑case efficiency loss of any NE relative to the centralized optimum, while PoS captures the best‑case loss. Theoretical bounds are derived as functions of the α‑fairness parameter and network topology. Empirically, simulations on random trees, grid networks, and a real ISP topology (derived from CAIDA data) reveal that PoA never exceeds 1.2, indicating that even the least efficient equilibrium retains at least 83 % of the optimal utility. In contrast, by deliberately selecting a strongly Pareto‑optimal equilibrium, PoS approaches 1 (observed values between 0.98 and 1.00), meaning the system can operate almost as efficiently as a centrally coordinated solution.

The experimental evaluation validates the algorithm’s practicality. Across all topologies, the distributed method converges within 30–45 iterations, achieving 95 %–98 % of the utility obtained by a centralized convex optimizer. Communication overhead is modest: each iteration requires exchanging only local allocation vectors, amounting to less than 0.5 % of total network traffic. Moreover, the algorithm adapts quickly to sudden traffic shifts, re‑stabilizing within 5–10 iterations, demonstrating robustness to dynamic environments.

In conclusion, the paper offers a rigorous and scalable solution for capacity allocation in environments where centralized control is infeasible or undesirable. By framing the problem as a potential game and enforcing strong Pareto optimality, the authors reconcile individual self‑interest with collective efficiency, delivering a pure NE that is both stable and near‑optimal. The work opens several avenues for future research, including extensions to multi‑class service differentiation, handling delayed or noisy information exchanges, and integrating incentive mechanisms (e.g., blockchain‑based smart contracts) to further align autonomous link behavior with network‑wide performance goals.