Distributed averaging integral Nash equilibrium seeking on networks

1 Introduction

• Continuous-time optimization algorithms expose control-theoretic properties and support closed-loop analysis of stability and optimality.The motivation extends to network games, where players may rely on limited or local information.Source: §1 · p. 1
• The approach in [11] uses passivity to design a consensus-based Nash-seeking algorithm that reconstructs nonlocal information through neighbor communication.The paper identifies this as its starting point for distributed Nash computation.Source: §1 · p. 1
• Online distributed integral control tunes communication-graph weights to enable convergence without algebraic-connectivity information or global parameters.The authors describe distributed averaging integral algorithms for Nash computation as new to their knowledge.Source: §1 · pp. 1–2

2 Mathematical background

• The mathematical background recalls established results and notation used in the subsequent analysis.The section introduces the game setup and Nash equilibrium before presenting the seeking dynamics.Source: §2 · p. 2, §2.1 · p. 2

2.1 Nash equilibrium problem

• Under the stated convexity and strict-monotonicity assumptions, the pseudo-gradient dynamics have a unique Nash equilibrium that is globally asymptotically stable.The text also notes that continuity of each cost function suffices for equilibrium existence.Source: §2.1 · p. 2

2.2 Nash equilibrium problem with partial information: Distributed seeking via static consensus

• With partial information, each agent maintains estimates of other players’ strategies and uses a consensus protocol to drive local copies toward their true values.This distributed method is based on neighbor communication over an undirected, connected graph.Source: §2.2 · pp. 2–3

2.3 Stability analysis

• The static-consensus dynamics in [11] converge exponentially to the unique Nash equilibrium under a strong-coupling condition, given the stated strong-monotonicity and Lipschitz assumptions.The solutions are also bounded for any initial condition.Source: §2.3 · p. 3

3 Distributed averaging integral Nash equilibrium seeking

• An integral loop tunes the control gains in place of static consensus coupling, removing the need for condition (6) on graph algebraic connectivity.The stated purpose is to compensate for potentially insufficient coupling.Source: §3 · p. 3
• The algorithm is written as coupled x, k, and ρ dynamics, with k_i denoting agent i's integrator state and γ_i a positive parameter.Agents exchange x_i, k_i, and ρ_i.Source: §3 · pp. 3–4
• The convergence analysis targets the Nash equilibrium without assuming strong coupling of the information-exchange graph.The result is established for system (9).Source: §3 · p. 4
• For system (9), under Assumptions 1(i), 2, 3, and 5, the solution is bounded and, for any initial condition, its x-component globally asymptotically converges to the Nash equilibrium.Specifically, limt→∞ x(t) = 1_N ⊗ x*.Source: §3 · p. 4

4 Projected distributed averaging integral Nash equilibrium seeking

• When Ω_i ⊂ R^n_i, a vector is a Nash equilibrium if and only if it satisfies VI(Ω, F).Under Assumptions 1(ii) and 2, the section states that a unique Nash equilibrium exists.Source: §4 · p. 4
• The proof uses a point-wise dissipation inequality to derive a Lyapunov inequality for the feedback interconnection with the distributed averaging integral controller.The cited inequality is used for the closed-loop system.Source: §4 · p. 4
• An invariance principle supports convergence analysis for the projected dynamics with Carathéodory solutions.The section first establishes forward invariance of the positive limit set when the vector field is Lipschitz on the closed convex set Ξ.Source: §4 · p. 5
• The section obtains the same convergence result as [11, Th. 5] without a lower bound on graph algebraic connectivity.Its proof uses an invariance principle for projected dynamical systems rather than Barbalat’s lemma as in [11, Proof of Th. 5].Source: §4 · p. 6
• Under Assumptions 1(ii), 3, 4, and 5, the projected system has a unique Carathéodory solution that stays in Ξ and whose x-component converges asymptotically to the Nash equilibrium.The convergence holds for any initial condition in Ξ.Source: §4 · p. 6

5.1 On the proposed distributed averaging algorithm

• The distributed integral action averages local decisions, with state k tuning the coupling input u.The alternative controller uses the local average ρ_i only, but requires a different Lyapunov function and additional boundedness analysis.Source: §5.1 · p. 6
• Implementing (7) requires exchanging k_iρ_i in addition to x_i; the alternative uses ρ_i only.The alternative also requires a different Lyapunov function and additional technical results for boundedness.Source: §5.1 · p. 6
• A disturbance-robust internal-model controller is suggested by, but investigation of that extension is left for future work.The passage says such a suitably modified version could still guarantee convergence despite an additive exosystem-generated disturbance.Source: §5.1 · p. 6

5.2 On the algebraic connectivity

• Under Lipschitz continuity, [11, Th. 3] relaxes the algebraic-connectivity condition but requires a global gain larger than a bound.The gain is global to all agents; the proposed approach removes the need for a global parameter.Source: §5.2 · p. 6
• For constrained games, Theorem 3 relaxes the connectivity requirement relative to [11, Th. 5].For the unconstrained case, the proposed approach also removes the need for a global parameter.Source: §5.2 · p. 6

6 Conclusion

• The paper considers distributed Nash-equilibrium computation without requirements on graph algebraic connectivity or high-gain assumptions.The conclusion leaves Nash games with coupling constraints for future research.Source: §6 · p. 6
• The problem is solved using dynamic consensus and pseudo-gradient dynamics rather than static consensus.The analysis and design use passivity and an invariance principle for projected dynamical systems.Source: §6 · p. 6
• Nash games with coupling constraints are left as future research.Source: §6 · p. 6

A Proof of Theorem 1

• The proof selects a Lyapunov function W and splits x into disagreement and average components to bound its derivative.The disagreement component is x⊥=(P_N ⊗ I_M)x and the average component is x∥=1_N ⊗ avg(x).Source: A Proof of Theorem 1 · pp. 6–7
• Lipschitz continuity and strong monotonicity bound the pseudo-gradient contribution, while the coupling terms combine into −xᵀ(LK*L ⊗ I_M)x.The proof uses the Lipschitz property of F and the strong monotonicity assumption.Source: A Proof of Theorem 1 · p. 7
• Choosing k* so M ≻ 0 yields the derivative inequality used to establish boundedness and convergence.The proof obtains the inequality by substituting bounds into (A.1).Source: A Proof of Theorem 1 · p. 7
• Radial unboundedness of W and the inequality imply bounded solutions that converge to the largest invariant set where disagreement vanishes and avg(x)=x*.On that set, x=1_N ⊗ x*, so the state converges to 1_N ⊗ x*.Source: A Proof of Theorem 1 · p. 7

B Proof of Theorem 2

• Compactness, convexity, and continuous differentiability make f Lipschitz on K ∩ Ψ, supporting existence of a unique trajectory that remains there.The trajectory is a Carathéodory solution defined on [0, ∞).Source: B Proof of Theorem 2 · p. 7
• On the compact invariant set, V is bounded below and nonincreasing, so it converges and the trajectory has a nonempty limit set.The proof uses invariance of the set and the Lyapunov derivative inequality.Source: B Proof of Theorem 2 · p. 7
• The limit set lies in the largest invariant set where ∇V(y)ᵀΠ_K(f(y))=0.The proof uses constancy of V on the limit set and invariance of limit-set solutions.Source: B Proof of Theorem 2 · pp. 7–8

• M ≻ 0 for sufficiently large k*, using Lemma 4 and Assumption 5 as in the proof of Theorem 1.Source: C Proof of Theorem 3 · p. 8
• For fixed k*, compactness and Lipschitz continuity yield a unique Carathéodory solution that remains in Ξ∩Ψ and converges to the largest invariant set where ˙W = 0.Source: C Proof of Theorem 3 · p. 8
• On that invariant set, x = 1_N ⊗ x*, which yields the theorem’s conclusion.Source: C Proof of Theorem 3 · p. 8