Generalized Nash Equilibrium (GNE) Explained

• Generalized Nash Equilibrium is a framework for games where each player's feasible strategies depend on the decisions of others.
• It extends classical Nash equilibrium by incorporating shared and interdependent constraints, allowing realistic modeling of resource limitations and market dynamics.
• Algebraic approaches, including Carathéodory-based KKT representation and Moment–SOS relaxation, enable global optimization and complete GNE enumeration despite exponential branch complexity.

A generalized Nash equilibrium (GNE) extends the Nash equilibrium concept to games where each player’s feasible strategy set may depend on the strategies chosen by other players, typically due to shared, coupled, or interdependent constraints. Classic Nash equilibrium applies to games where each player minimizes their own cost (or maximizes payoff) over a fixed feasible set; in contrast, the GNE framework accommodates more complex and realistic modeling of resource limitations, market-clearing, network flows, competitive allocation, and hierarchical policy games found in economic, engineering, and multi-agent systems.

1. Mathematical Definition and Characterizations

Consider a game with $N$ players, where each player $i$ controls a decision variable $x_i \in \mathbb{R}^{n_i}$ and the joint strategy is $x = (x_1, ..., x_N) \in \mathbb{R}^n$. Player $i$ minimizes a cost (or maximizes a payoff) $f_i(x)$ over a feasible set $X_i(x_{-i})$ that depends on the strategies of all other players $x_{-i} = (x_1, ..., x_{i-1}, x_{i+1}, ..., x_N)$. Formally,

$$X_i(x_{-i}) = \{ x_i \mid A_i x_i - b_i(x_{-i}) \ge 0 \}$$

with $A_i$ constant and $i$0 a vector of polynomials in $i$1 (Choi et al., 2024).

A vector $i$2 is a GNE if, for every $i$3,

$i$4

That is, no player can decrease their cost by unilaterally deviating, considering the resulting changes in their own feasible set.

The first-order (KKT) conditions for GNEs incorporate Lagrange multipliers for constraints. For smooth $i$5 and representing the constraints as $i$6, for each $i$7: $i$8 with $i$9 (Choi et al., 2024). The set of all solutions to the aggregate KKT system is the set of "KKT points," which includes all GNEs.

2. Carathéodory-Based KKT Set Representation and Branch Decomposition

A key innovation for GNEPs with quasi-linear constraints (constraints that are linear in local variables but may be polynomial in the variables of other players) is the use of Carathéodory's theorem to represent the set of feasible multipliers. For each player $x_i \in \mathbb{R}^{n_i}$0, any Lagrange multiplier $x_i \in \mathbb{R}^{n_i}$1 can be chosen to have at most $x_i \in \mathbb{R}^{n_i}$2 nonzero entries.

For any $x_i \in \mathbb{R}^{n_i}$3 with $x_i \in \mathbb{R}^{n_i}$4 and $x_i \in \mathbb{R}^{n_i}$5, define the submatrix $x_i \in \mathbb{R}^{n_i}$6 and $x_i \in \mathbb{R}^{n_i}$7. The basic KKT solution must satisfy: $x_i \in \mathbb{R}^{n_i}$8 with all other components of $x_i \in \mathbb{R}^{n_i}$9 zero. We can explicitly represent

$x = (x_1, ..., x_N) \in \mathbb{R}^n$0

and combine these to define, for each multi-index or "branch" $x = (x_1, ..., x_N) \in \mathbb{R}^n$1,

$x = (x_1, ..., x_N) \in \mathbb{R}^n$2

The entire KKT set is then

$x = (x_1, ..., x_N) \in \mathbb{R}^n$3

where $x = (x_1, ..., x_N) \in \mathbb{R}^n$4 is the set of all such multi-indices across players (Choi et al., 2024).

3. Branchwise Polynomial Optimization and Moment–SOS Relaxation

The GNE computation reduces to solving, for each $x = (x_1, ..., x_N) \in \mathbb{R}^n$5, a branchwise constrained polynomial optimization problem. For each $x = (x_1, ..., x_N) \in \mathbb{R}^n$6, define a generic positive-definite quadratic form $x = (x_1, ..., x_N) \in \mathbb{R}^n$7 (where $x = (x_1, ..., x_N) \in \mathbb{R}^n$8 is the vector of all monomials to the relevant degree $x = (x_1, ..., x_N) \in \mathbb{R}^n$9). The problem is

$i$0

where constraints in $i$1 are all polynomial (equalities and inequalities).

The Moment–SOS hierarchy offers a mechanism for solving or certifying infeasibility of such polynomial programs. Let $i$2 be a truncated moment sequence, and $i$3 the moment matrix at order $i$4. For constraints $i$5 and $i$6, the primal moment SDP is: $i$7 For large enough $i$8 and under Archimedeanity, this converges to the global optimum and exposes global minimizers via the flat extension condition.

Each detected $i$9 is tested for genuine GNE status; for nonconvex games, this requires further best-response checks, solvable again via SOS relaxations (Choi et al., 2024).

4. Branch Enumeration Algorithm and Complexity

The full procedure is as follows:

1. For each player $f_i(x)$0, compute $f_i(x)$1 and all subsets $f_i(x)$2 as above. Enumerate all $f_i(x)$3.
2. For each $f_i(x)$4, construct the explicit branch program (as above) and solve it by Moment–SOS hierarchy.
3. For each solution in a non-empty $f_i(x)$5, verify whether it is a genuine GNE (for nonconvex games).
4. To find more GNEs in $f_i(x)$6, repeat Moment–SOS optimization with a constraint $f_i(x)$7 for some small $f_i(x)$8 until infeasible.
5. Collect and return the union of all found GNEs.

Complexity grows rapidly with the number of branches, $f_i(x)$9, and with the number and size of SDP constraints as functions of $X_i(x_{-i})$0 and relaxation degree (Choi et al., 2024).

5. Representative Numerical Example

In a nonconvex two-player case with $X_i(x_{-i})$1, player objectives given by polynomials (e.g., $X_i(x_{-i})$2), and five quasi-linear constraints per player, there are $X_i(x_{-i})$3 branches.

Using Algorithm 4.5 and Moment–SOS relaxations up to order $X_i(x_{-i})$4, the unique GNE was found within $X_i(x_{-i})$5 seconds, and initial detection required only $X_i(x_{-i})$6 seconds. A classical homotopy approach produced thousands of complex solutions (KKT points), most non-real or non-GNE (Choi et al., 2024).

6. Scope, Limitations, and Algebraic-Geometric Contributions

This Carathéodory–branch–moment/SOS approach is applicable to any GNEP with polynomial objectives and local constraints quasi-linear in own variables but polynomial in other players’ variables, encompassing both convex and nonconvex games as well as both finitely and infinitely many GNEs.

The approach partitions the search space into finite semi-algebraic sets ("branches"), eliminates multiplier variables via explicit linear algebra, and solves higher-level polynomial optimization problems globally via the Moment–SOS hierarchy, with algebraic guarantees of completeness or nonexistence.

Practical bottlenecks may arise from the exponential branch count and the scaling of SDP dimensions. For moderate-scale problems, however, this methodology yields complete algebraic-geometric certificates and full solution enumeration (Choi et al., 2024).

7. Connections, Extensions, and Related Work

This algebraic-geometric method complements homotopy continuation, direct KKT system solving, and other global optimization techniques (e.g., polyhedral homotopy and general Moment–SOS approaches for polynomial GNEPs (Lee et al., 2022, Nie et al., 2022)). The explicit Carathéodory decomposition and branchwise pLME elimination uniquely exploit the quasi-linear structure to reduce scaling and optimize over much lower-dimensional sets than the full KKT system.

Comparable rational-function techniques and feasible extensions have been developed for general rational GNEPs (Nie et al., 2021). Variations of distributed and operator-theoretic GNE seeking in monotone settings adopt wholly different algorithmic, not algebraic, frameworks (Cenedese et al., 2020, Yi et al., 2017).

This branch–Moment–SOS paradigm thus establishes a robust, constructive toolkit for global enumeration/certification of all GNEs in the quasi-linear constraint regime (Choi et al., 2024).