Nikaido–Isoda Gap Function

• The Nikaido–Isoda gap function is a scalar-valued measure that aggregates individual deviations to characterize equilibrium in Nash and generalized Nash games.
• Regularized and gradient-based variants enhance differentiability and computational tractability, aiding in optimization and algorithm design.
• Algorithmic applications leverage the gap function for finite convergence, error bounds, and reformulations in stochastic, mixed-integer, and learning-based game settings.

Searching arXiv for papers on the Nikaido–Isoda gap function and closely related formulations. arXiv search query: "Nikaido Isoda gap function generalized Nash" The Nikaido–Isoda gap function is a scalar-valued merit function for Nash equilibrium and generalized Nash equilibrium problems. In its standard form, for a strategy profile $x=(x_1,\dots,x_N)$ and an alternative profile $y=(y_1,\dots,y_N)$, it is written as

$$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$

or equivalently, with cost notation,

$$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$

The associated value or gap function is

$V(x)=\sup_{y\in X(x)}\Psi(x,y),$

with $X(x)$ denoting the feasible strategy set, which may depend on rival decisions in generalized Nash games. Across the formulations surveyed in recent work, $V(x)\ge 0$, and a feasible profile is an equilibrium if and only if the gap is zero; in generalized Nash settings this requires both $x\in X(x)$ and $V(x)=0$ (Harks et al., 2021, Sultana et al., 2023, Duguet et al., 3 Jun 2025).

1. Formal definition and equilibrium characterization

The defining interpretation of the Nikaido–Isoda construction is aggregate unilateral improvement. Each summand compares the current objective value of player $i$ at $y=(y_1,\dots,y_N)$0 with the objective value obtained when only player $y=(y_1,\dots,y_N)$1 deviates to $y=(y_1,\dots,y_N)$2 while the rivals remain fixed at $y=(y_1,\dots,y_N)$3. The resulting scalar therefore aggregates deviations by all players into a single nonnegative certificate of disequilibrium.

For generalized Nash equilibrium problems, the admissible set $y=(y_1,\dots,y_N)$4 depends on $y=(y_1,\dots,y_N)$5, because shared or coupling constraints make each player’s feasible set depend on rival strategies. In this setting, the gap function

$y=(y_1,\dots,y_N)$6

retains the exact equilibrium characterization: a profile $y=(y_1,\dots,y_N)$7 is a generalized Nash equilibrium if and only if $y=(y_1,\dots,y_N)$8 and $y=(y_1,\dots,y_N)$9. Several papers use this equivalence as the basis for optimization reformulations, convexification arguments, or algorithm design (Harks et al., 2021, Duguet et al., 3 Jun 2025, Wen et al., 17 Sep 2025).

A recurring technical point is that the zero-gap property is exact for the original NI construction, but only after feasibility is enforced. A common simplification is to speak of $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$0 alone; in generalized Nash models this omits the requirement that the profile be feasible for the strategy-dependent constraint system. The distinction is explicit in formulations based on Rosen-type jointly convex generalized Nash games and in mixed-integer generalized Nash equilibrium problems (Sultana et al., 2023, Harks et al., 2021).

2. Regularized and smooth variants

Because the unregularized value function may be nondifferentiable, several works replace $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$1 by a regularized Nikaido–Isoda function. One standard form is

$$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$2

with associated regularized gap

$$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$3

Under the convexity and regularity conditions stated in the cited works, $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$4 is differentiable, and a normalized Nash equilibrium is characterized by $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$5. The maximizer $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$6 yields explicit gradient formulas used in analysis and algorithms (Sultana et al., 2023).

Regularization also underlies error bounds. In jointly convex generalized Nash games, weak sharpness of the solution set $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$7 is characterized through

$$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$8

and, under convexity of $$\Psi(x,y)=\sum_{i=1}^N \big[\theta_i(x_{-i},x_i)-\theta_i(x_{-i},y_i)\big],$$9, linear conditioning of $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$0 is equivalent to weak sharpness. These properties lead directly to finite convergence results for proximal-point-type methods (Sultana et al., 2023).

A second regularized form appears in multi-leader–follower games: $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$1 and, for follower games parameterized by leader decisions,

$$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$2

The induced value function $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$3 is used as a differentiable surrogate for follower equilibrium constraints. The cited analysis states that, for $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$4, $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$5 is continuously differentiable even if the follower problems are nonconvex (Hori et al., 18 Jun 2026).

The same smoothing logic appears in stochastic Nash equilibrium under uncertainty, where a regularized value function $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$6 is constructed so that it is continuously differentiable and admits a Lipschitz gradient, enabling projected gradient methods with sampling and inexact inner solves (Marrinan et al., 27 Oct 2025).

3. Reformulations in generalized Nash and mixed-integer games

A major line of recent work uses the NI gap to reformulate equilibrium computation as optimization. One direct formulation is

$$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$7

which yields a bilevel min–max problem. In mixed-integer generalized Nash equilibrium problems, this reformulation is the basis of an exact branch-and-cut method. An epigraph form introduces playerwise best-response values $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$8 through auxiliary variables $$\Psi(x,y)=\sum_{i\in N}\pi_i(x)-\sum_{i\in N}\pi_i(y_i,x_{-i}).$$9, and node relaxations solve a high-point problem whose positive optimal value certifies absence of equilibrium in the branch (Duguet et al., 3 Jun 2025).

A related nonconvex perspective treats NI minimization as a minimum disequilibrium problem. In that formulation, the minimum of the gap function measures minimum aggregate disequilibrium and reduces to traditional equilibrium when equilibrium exists. The cited work connects this construction with bilevel feasibility, semi-infinite programming, Lagrangian duality, and cutting-plane or constraint-generation methods, explicitly without convexity assumptions and with applicability to mixed-integer player problems (Harwood et al., 2021).

For generalized Nash equilibrium problems with mixed-integer variables, convexification based on the Nikaido–Isoda function provides another reformulation route. A feasible strategy profile is an equilibrium for the original instance if and only if it is an equilibrium for a convexified instance and the convexified cost functions coincide with the initial ones. The same work derives characterizations for when convexification leads to a jointly constrained or jointly convex generalized Nash equilibrium problem, and for quasi-linear models the convexified formulation reduces the game to a standard optimization problem (Harks et al., 2021).

Under uncertainty, NI-based reformulation is also used to convert a generalized Nash game with shared distributionally robust chance constraints over a Wasserstein ball into a deterministic mixed-integer nonlinear program. In the quadratic case discussed in the cited paper, the resulting MINLP has the distinctive property that the integer and continuous variables are decoupled in both the objective function and the constraints (Wen et al., 17 Sep 2025).

4. Algorithmic roles: pruning, cuts, descent, and finite convergence

The NI gap serves not only as a characterization device but also as an algorithmic progress measure. In the branch-and-cut method for mixed-integer generalized Nash equilibrium problems, the node problem minimizes a relaxation of the NI-based high-point formulation. If the optimal value is strictly positive, then the corresponding branch contains no equilibrium and can be pruned. If an integer-feasible point with value zero is found, it is a Nash equilibrium. When a node solution is not an equilibrium, non-NE cuts derived from the best-response structure encoded in $V(x)=\sup_{y\in X(x)}\Psi(x,y),$0 remove that point without excluding equilibria; finite termination and correctness are proved (Duguet et al., 3 Jun 2025).

In the regularized setting, the proximal point algorithm is analyzed through $V(x)=\sup_{y\in X(x)}\Psi(x,y),$1. If $V(x)=\sup_{y\in X(x)}\Psi(x,y),$2 is linearly conditioned or if the set of normalized Nash equilibria is weakly sharp, then any sequence generated by the stated proximal-point scheme reaches the equilibrium set in finitely many iterations. For a quadratic class, the cited work also gives an estimate

$V(x)=\sup_{y\in X(x)}\Psi(x,y),$3

for the number of iterations after which all subsequent iterates belong to $V(x)=\sup_{y\in X(x)}\Psi(x,y),$4 (Sultana et al., 2023).

For nonconvex differentiable games, the original NI function is replaced by the Gradient-based Nikaido–Isoda function

$V(x)=\sup_{y\in X(x)}\Psi(x,y),$5

where $V(x)=\sup_{y\in X(x)}\Psi(x,y),$6 changes only player $V(x)=\sup_{y\in X(x)}\Psi(x,y),$7’s component by a single gradient step. This variant vanishes exactly at stationary Nash points, provides explicit error bounds relating $V(x)=\sup_{y\in X(x)}\Psi(x,y),$8 to $V(x)=\sup_{y\in X(x)}\Psi(x,y),$9, and supports gradient descent with sublinear convergence to a first-order stationary point. In bilinear min–max games and multi-player quadratic games, the cited paper states that the GNI function is convex, yielding linear convergence to a Nash equilibrium when one exists (Raghunathan et al., 2019).

These results delimit an important distinction. The classical NI gap is exact for equilibrium; the gradient-based NI construction is exact for first-order stationarity and becomes exact for Nash equilibrium only under additional convexity in each player’s own variable (Raghunathan et al., 2019, Dou et al., 2019).

5. Learning, uncertainty, and stochastic games

Recent work extends NI-based methodology into learning settings where equilibrium data are expensive or unavailable. In multiparametric generalized Nash equilibrium problems, the NI gap is used directly as a training loss for a neural network approximation of the solution mapping. Rather than regressing on equilibrium solutions, the method requires only best-response value data. To avoid bilevel optimization during training, each agent’s optimal best-response cost is approximated by a value-function surrogate, producing a single-level learning problem (Bemporad et al., 27 May 2026).

In stochastic games with independent chains, the NI gap is used as an averaged convergence metric over occupancy measures. For a profile $X(x)$0, the total Nikaido–Isoda gap is

$X(x)$1

and one cited paper states that if this quantity is less than $X(x)$2, then $X(x)$3 is an $X(x)$4-Nash equilibrium. Dual averaging, mirror descent, and decentralized mirror descent algorithms are then analyzed in terms of the averaged NI gap rather than direct convergence in policy space. The resulting guarantees are almost sure, in expectation, or with high probability, and the stronger rates rely on additional structure such as social concavity or variational stability (Etesami, 2022, Qin et al., 2023).

For continuous games without pure equilibria, a mixed-strategy extension uses pushforward measures and a mixed-strategy generalization of the Gradient-based NI function, denoted MC-GNI. Here the functional objective measures local regret in mixed-strategy space, and gradient descent on neural parameterizations converges to a stationary Nash point under the convexity assumption on payoff functions (Dou et al., 2019).

In bilevel reinforcement learning over regularized zero-sum Markov games, the NI gap

$X(x)$5

is zero if and only if $X(x)$6 is a saddle point of the lower-level game. PANDA augments the upper-level objective by $X(x)$7, thereby avoiding upper-level hypergradients and second-order information. The cited analysis proves convergence to stationary points without convexity assumptions on either level and gives $X(x)$8 iterations and $X(x)$9 sample complexity for reaching an $V(x)\ge 0$0-stationary point (Zheng et al., 26 May 2026).

6. Scope, limitations, and related concepts

The principal strength of the Nikaido–Isoda gap function is that it does not require monotonicity of the concatenated gradient map or potentiality of the game to formulate equilibrium search as optimization. This property is emphasized in stochastic Nash equilibrium under uncertainty, where minimizing a suitably defined NI value function replaces assumptions standard in variational inequality or potential-game approaches (Marrinan et al., 27 Oct 2025).

At the same time, exactness depends on the variant in use. For the classical and regularized NI value functions, zero gap characterizes equilibrium subject to feasibility. For gradient-based or mixed-strategy gradient-based variants, the primary object is first-order stationarity; the cited works explicitly add convexity assumptions when promoting stationary Nash points to Nash equilibria (Raghunathan et al., 2019, Dou et al., 2019). In generalized Nash equilibrium problems under uncertainty, one cited paper further requires a suitable regularity requirement under which a stationary point of the regularized value function is a Nash equilibrium (Marrinan et al., 27 Oct 2025).

A second limitation is computational. The original NI function requires solving, for each $V(x)\ge 0$1, a supremum or best-response problem over $V(x)\ge 0$2, which can be nonconvex, semi-infinite, or mixed-integer. Much of the recent literature can be read as a sequence of devices for managing that burden: regularization to obtain differentiability, convexification to transfer equilibrium to a tractable surrogate, branch-and-cut and cutting planes for exact mixed-integer computation, value-function surrogates for learning, and sampling or stochastic approximation for uncertain objectives (Harks et al., 2021, Duguet et al., 3 Jun 2025, Bemporad et al., 27 May 2026, Marrinan et al., 27 Oct 2025).

Finally, the NI framework supports broader disequilibrium notions when equilibrium may fail to exist. The minimum disequilibrium formulation interprets the minimum of the gap as minimum aggregate opportunity cost and reduces to equilibrium when equilibrium exists. This suggests a broader role for the Nikaido–Isoda function: not only as a zero-finding certificate for equilibrium, but also as a quantitative organizing principle for optimization, approximation, and learning in nonconvex, mixed-integer, stochastic, and hierarchical games (Harwood et al., 2021).