Learning Generalized Nash Equilibria in Non-Monotone Games with Quadratic Costs

We study generalized Nash equilibrium (GNE) problems in games with quadratic costs and individual linear equality constraints. Departing from approaches that require strong monotonicity and/or shared constraints, we reformulate the KKT conditions of the (generally non-monotone) games into a tractable convex program whose objective satisfies the Polyak-Lojasiewicz (PL) condition. This PL geometry enables a distributed gradient method over a fixed communication graph with global geometric (linear) convergence to a GNE. When gradient information is unavailable or costly, we further develop a zero-order fully distributed scheme in which each player uses only local cost evaluations and their own constraint residuals. With an appropriate step size policy, the proposed zero-order method converges to a GNE, provided one exists, at rate O(1/t).

💡 Research Summary

The paper tackles the problem of computing a Generalized Nash Equilibrium (GNE) in games where each player’s cost is quadratic and each player is subject to its own linear equality constraints. Unlike most existing works that rely on strong monotonicity of the game mapping or on shared coupling constraints, the authors consider a more general setting: the pseudo‑gradient may be non‑monotone and constraints are individual.

The key technical contribution is a reformulation of the KKT conditions of the GNE problem into a single convex optimization problem. By exploiting the quadratic structure, the authors show that the resulting objective can be written as a squared norm ‖Gz + e‖², where z concatenates all primal and dual variables. This representation immediately implies that the objective satisfies the Polyak‑Łojasiewicz (PL) condition with constant µ_F = 2σ_min⁺(G)². The PL condition is weaker than strong convexity/monotonicity but still guarantees linear (geometric) convergence of gradient descent when the gradient is Lipschitz continuous.

Armed with the PL property, the authors first propose a distributed first‑order algorithm. Players communicate over a fixed, connected graph and each updates its local primal‑dual block using its own partial gradient ∇_i f_i(z). By choosing a constant stepsize α ∈ (0, 2/(L+µ_F)), where L is the Lipschitz constant of ∇F, the iterates converge globally at a linear rate (1 − αµ_F). This extends the fast convergence results that previously required strong monotonicity and shared constraints.

Recognizing that gradient information may be unavailable or expensive, the paper then develops a fully zero‑order (gradient‑free) scheme. Each player can only evaluate its local cost J_i at a queried joint action and observe its own constraint residual A_i x − b_i. Using independent Gaussian directions ξ_i^x, ξ_i^λ and η_i, each player constructs four query points around the current iterate. Two‑point finite‑difference estimators are applied to the local Lagrangian L_i and the constraint violation term c_i, yielding scalar directional increments Δ. By forming specific combinations S₁ and S₂ of these increments and aggregating them across the network (via a central aggregator that sums S₂ − d S₁ and the Δ for c_i), each player obtains unbiased stochastic estimates ζ_xi and ζ_λi of the partial derivatives ∂F/∂x_i and ∂F/∂λ_i.

The stochastic estimators have bounded variance that can be controlled by the smoothing parameters σ (perturbation magnitude) and δ (finite‑difference step). Consequently, the zero‑order update z_{t+1}=z_t − α_t ζ_t behaves like stochastic gradient descent on a PL‑satisfying function. By employing a diminishing stepsize α_t = c/(t+1), the authors prove an O(1/t) convergence rate in expectation for the objective value, which matches the optimal rate for strongly convex stochastic optimization.

Theoretical results are summarized in two main theorems: (1) linear convergence of the distributed gradient method under constant stepsize; (2) O(1/t) convergence of the zero‑order method under a suitable diminishing stepsize schedule. Both rely on the PL condition and on standard assumptions about the communication graph (connectedness) and the boundedness of the stochastic gradient estimators.

Numerical experiments on a synthetic power‑market model with ten players, each having five‑dimensional strategies and distinct linear constraints, illustrate the practical performance. The first‑order method reaches machine‑precision accuracy within 20–30 iterations, confirming the predicted geometric rate. The zero‑order method, while requiring more iterations (≈200–300) due to stochastic estimation noise, still follows the O(1/t) decay and converges to the same equilibrium, demonstrating robustness when gradients are unavailable.

The paper also discusses limitations. The need for a global aggregator to sum the locally computed scalar quantities introduces extra communication overhead, which may be undesirable in fully decentralized settings. Moreover, the zero‑order estimator’s variance scales with the problem dimension, potentially limiting scalability to very high‑dimensional games. Future research directions suggested include communication‑efficient aggregation (e.g., gossip or compression), extensions to inequality or nonlinear coupling constraints, and adaptive tuning of the smoothing parameters to improve empirical performance.

In summary, this work provides a novel pathway to compute GNEs in non‑monotone quadratic games with individual constraints, leveraging a PL‑based convex reformulation to achieve fast linear convergence with first‑order information and an optimal O(1/t) rate with only function‑value feedback. The results broaden the applicability of distributed equilibrium‑learning algorithms beyond the restrictive monotonicity and shared‑constraint regimes that dominate the current literature.