Zero-Error Nash Equilibrium Coordination

• Zero-Error Nash Equilibrium Coordination is the phenomenon where strategic interactions achieve an exact Nash equilibrium without miscoordination by leveraging precise strategy marginalization or pure equilibrium paths.
• The topic examines diverse models—from two-player zero-sum and coordination games to Bayesian and networked scenarios—highlighting structural conditions required for literal zero error.
• Mechanism design, distributed optimization, and quantum extensions are explored as actionable approaches to overcome learning noise and communication constraints for exact equilibrium coordination.

Zero-Error Nash Equilibrium Coordination denotes a family of exact-coordination phenomena in which strategic interaction reaches Nash-equilibrium behavior without residual exploitability, miscoordination, or stage-wise equilibrium error. In two-player zero-sum games, the most direct formulation is that the marginal strategies of any coarse correlated equilibrium already form a Nash equilibrium, so convergence of no-external-regret learning to coarse correlated equilibrium yields empirical marginals whose exploitability goes to zero; in the approximate case, an $\epsilon$-coarse correlated equilibrium yields a $2\epsilon$-Nash equilibrium (2304.07187). In the broader literature, closely related formulations include deterministic convergence to pure equilibria in coordination games, zero probability of miscoordination under communication-proof equilibria, asymptotically exact convergence to an optimal Nash equilibrium in distributed aggregative games, and type-by-type realization of a stage-game Nash equilibrium in Bayesian games (Avramopoulos, 2016, Heller et al., 2020, Ma et al., 2022, Ambuj et al., 3 Sep 2025).

1. Formal scope and variant notions

The expression is not tied to a single formalism. In the cited literature, it appears in several mathematically distinct settings that share the requirement of exact equilibrium coordination rather than merely high expected utility or small average regret.

Setting	Zero-error notion	Representative guarantee
Two-player zero-sum normal-form games	Exact Nash marginals from correlated play	CCE marginals are NE; $\epsilon$-CCE implies $2\epsilon$-NE
Coordination and stag-hunt mechanisms	No-loss path to a pure equilibrium	Strictly profitable or non-harmful unilateral deviations lead to adoption
Anti-coordination with public signals	Zero collision conditional on the signal	Each signal realizes a PSNE and induces an efficient CE
Cheap-talk coordination with private values	Zero miscoordination	$\Pr(M)=0$ under communication-proof equilibrium
Noisy global games	Coordination efficiency equal to one	Impossible under strictly positive observation noise
Bayesian games with entanglement	Pathwise equilibrium realization for every type profile	$\Pr[a(t)=a^*(t)]=1$ for each realized type profile

In the zero-sum formulation, the object of interest is exploitability: the best-response payoff against one player’s marginal minus the worst-case payoff guaranteed by the other player’s marginal. In deterministic coordination games, the focus shifts to pure-strategy stability and to improvement paths that never require a loss-making step. In communication models, the relevant error is miscoordination probability. In Bayesian games with private information, the strongest requirement is pathwise: for every realized type profile, the realized action profile must be a Nash equilibrium of the induced complete-information stage game (2304.07187, Avramopoulos, 2016, Cigler et al., 2014, Heller et al., 2020, Vasconcelos et al., 2023, Ambuj et al., 3 Sep 2025).

These variants are not equivalent. Some are exact in the limit of learning dynamics, some are exact at every realized state or type profile, and some are exact only under special structural assumptions such as zero-sum payoffs, potential structure, or access to a correlation device. A central theme is therefore not merely convergence to equilibrium, but the structural conditions under which equilibrium coordination can be exact.

2. Zero-sum games: coarse correlated equilibrium already determines Nash equilibrium

For a finite two-player zero-sum game with action sets $A_1,A_2$, payoff functions $u_1,u_2$ satisfying $u_2(a_1,a_2)=-u_1(a_1,a_2)$, and mixed strategies $x\in\Delta(A_1)$, $2\epsilon$0, a distribution $2\epsilon$1 is a coarse correlated equilibrium if, for each player $2\epsilon$2 and unilateral deviation $2\epsilon$3,

$2\epsilon$4

Let $2\epsilon$5 and $2\epsilon$6 be the marginals. The key theorem states that, in two-player zero-sum games, the marginals of any CCE form a Nash equilibrium (2304.07187).

The proof is driven by a marginalization lemma. For any fixed deviation $2\epsilon$7 of player 1,

$2\epsilon$8

and analogously for any fixed $2\epsilon$9 of player 2,

$\epsilon$0

Combining these identities with the CCE inequalities and the zero-sum relation yields the saddle-point inequalities

$\epsilon$1

so $\epsilon$2 is a saddle point with value $\epsilon$3, hence a Nash equilibrium (2304.07187).

The approximate statement is equally sharp at the level proved in the source. If $\epsilon$4 is an $\epsilon$5-CCE, then

$\epsilon$6

and the marginals form a $\epsilon$7-Nash equilibrium. The left-hand side is exactly the exploitability gap, so approximate coarse correlated play translates directly into a quantitative equilibrium guarantee (2304.07187).

The Rock–Paper–Scissors example makes the mechanism transparent. The correlated distribution

$\epsilon$8

places all mass on ties. Its marginals are uniform, so $\epsilon$9 and similarly for $2\epsilon$0. The expected payoff under $2\epsilon$1 is $2\epsilon$2, and every unilateral deviation against the uniform marginal also yields $2\epsilon$3. Thus $2\epsilon$4 is a CCE, and its marginals are exactly the well-known mixed Nash equilibrium of Rock–Paper–Scissors (2304.07187).

This equivalence is special to two-player zero-sum structure. The argument uses $2\epsilon$5 to flip the player-2 CCE inequality into the right-hand saddle condition, and the source states explicitly that the implication “CCE marginals $2\epsilon$6 NE” does not generally hold in general-sum or multiplayer games (2304.07187).

3. Mechanism design and dynamic equilibrium selection in coordination games

In stag hunts, zero-error coordination is formulated as a path from defection to cooperation consisting only of unilateral deviations that are either strictly profitable or non-harmful. The base game has strategy set $2\epsilon$7 for each player, with $2\epsilon$8 denoting adopt or cooperate and $2\epsilon$9 denoting defect; it has two strict pure Nash equilibria, universal adoption and universal defection, plus a mixed equilibrium of low predictive value (Avramopoulos, 2016).

The insurance mechanism enlarges the strategy set to $\Pr(M)=0$0, where $\Pr(M)=0$1 means “buy insurance and adopt.” The basis payoffs for $\Pr(M)=0$2 remain unchanged. If coordination fails, the insured payoff exceeds the constant payoff $\Pr(M)=0$3 from defection; if coordination succeeds, insured adoption is strictly worse than uninsured adoption because of the premium. This yields strict iterated dominance: $\Pr(M)=0$4 strictly dominates $\Pr(M)=0$5, and after eliminating $\Pr(M)=0$6, $\Pr(M)=0$7 strictly dominates $\Pr(M)=0$8. The induced game is solvable by strict iterated dominance to universal adoption without insurance, which is therefore the unique pure Nash equilibrium. From any state, there exists a path of strictly profitable unilateral deviations to $\Pr(M)=0$9: first switch all $\Pr[a(t)=a^*(t)]=1$0-players to $\Pr[a(t)=a^*(t)]=1$1, then all $\Pr[a(t)=a^*(t)]=1$2-players to $\Pr[a(t)=a^*(t)]=1$3. The source describes this as a “zero-error” adoption path because no player ever loses along the path (Avramopoulos, 2016).

The election mechanism instead enlarges the strategy set to $\Pr[a(t)=a^*(t)]=1$4, where $\Pr[a(t)=a^*(t)]=1$5 means “vote, adopt if the outcome is positive, defect otherwise,” and $\Pr[a(t)=a^*(t)]=1$6 means “vote, adopt irrespective of the vote outcome.” It is solvable by weak iterated dominance to profiles consisting of $\Pr[a(t)=a^*(t)]=1$7 and $\Pr[a(t)=a^*(t)]=1$8, and the induced game is weakly ordinally acyclic. All strongly maximal equilibria entail adoption, while universal defection remains a weak Nash equilibrium of reduced predictive value. The relevant zero-error statement is weaker than in the insurance mechanism but still exact at the level of path properties: from any state with defectors, there exists a path of non-harmful unilateral deviations to adoption, so no player needs to take a risky step (Avramopoulos, 2016).

A different dynamic selection mechanism appears in a 2-by-3 embedding of a 2-by-2 coordination game. Player B acquires a third action $\Pr[a(t)=a^*(t)]=1$9, and the payoff perturbation is chosen so that

$A_1,A_2$0

With the additional condition

$A_1,A_2$1

the two-population replicator dynamics on the modified 2-by-3 game satisfy a global selection result: for every initial condition in the interior of the state space, the $A_1,A_2$2-limit set is the singleton $A_1,A_2$3. The third action is not used at the limit, but its existence destabilizes the undesired equilibrium and eliminates asymptotic mixed-strategy miscoordination. The paper’s interpretation is that the existence of a new action, although unused, allows both players to coordinate on one Nash equilibrium of the original game (Castro, 2023).

These constructions show that exact coordination need not arise only from equilibrium refinement. It can also be engineered by enlarging the action space or by adding opt-in institutions that reshape profitable-deviation geometry while leaving the original coordination objective intact.

4. Graphs, resources, and networked coordination

In coordination games on undirected graphs, players are nodes, actions are colors, and each player’s payoff is the number of neighbors choosing the same color: $A_1,A_2$4 For the unweighted model, pure Nash equilibria and 2-equilibria always exist, while a 3-equilibrium need not exist. Strong equilibria exist for several special classes, including pseudoforests, color forests, color complete graphs, and games with only two colors. The efficiency results are sharply nonuniform: the price of anarchy for pure Nash equilibria is unbounded, whereas for $A_1,A_2$5-equilibria it lies between

$A_1,A_2$6

with the upper bound tight for $A_1,A_2$7, i.e. for strong equilibria. The decision problem of whether a given profile is a $A_1,A_2$8-equilibrium is co-NP complete (Apt et al., 2015).

For weighted directed graph coordination games, the payoff of node $A_1,A_2$9 is

$u_1,u_2$0

Here exact coordination is tied to existence of polynomial-length improvement or coalitional-improvement paths. On weighted DAGs, every Nash equilibrium is a strong equilibrium, and both improvement and c-improvement paths have length at most $u_1,u_2$1. On simple cycles with restricted bonuses or edge weights, and on open chains of simple cycles, the sources provide explicit polynomial bounds for reaching NE or SE. For two-color games, both improvement and c-improvement paths have length at most $u_1,u_2$2. At the same time, the limits are strong: without the stated restrictions, a Nash equilibrium may not even exist, and the problem of determining the existence of a pure Nash equilibrium or a strong equilibrium is NP-complete already for unweighted graphs with no bonuses (Apt et al., 2019, Simon et al., 2016).

Anti-coordination produces a different exactness notion. In decentralized channel allocation with $u_1,u_2$3 identical agents and $u_1,u_2$4 orthogonal resources, a public coordination signal $u_1,u_2$5 is used to index a learned action table $u_1,u_2$6. For each signal value $u_1,u_2$7, agents learn a pure-strategy Nash equilibrium of the one-shot channel allocation game, so that

$u_1,u_2$8

The resulting repeated outcome is an efficient correlated equilibrium: each signal realizes perfect anti-coordination conditional on the signal, and social welfare equals $u_1,u_2$9 per signal when $u_2(a_1,a_2)=-u_1(a_1,a_2)$0. Fairness is quantified by Jain’s index

$u_2(a_1,a_2)=-u_1(a_1,a_2)$1

which tends to $u_2(a_1,a_2)=-u_1(a_1,a_2)$2 as $u_2(a_1,a_2)=-u_1(a_1,a_2)$3 increases (Cigler et al., 2014).

Ride sharing games furnish a positive and negative externality model in which zero-error coordination is available only under a restrictive deterministic class. Under common action sets, high capacity $u_2(a_1,a_2)=-u_1(a_1,a_2)$4, first-fit seat allocation, first-fit linear allocation, one vehicle $u_2(a_1,a_2)=-u_1(a_1,a_2)$5, and a disjoint set of necessary and sufficient paths, the game has the finite improvement property and admits a pure Nash equilibrium with ordinal potential

$u_2(a_1,a_2)=-u_1(a_1,a_2)$6

Deterministic round-robin best-response therefore terminates at a pure Nash equilibrium. By contrast, in the Bayesian version with signaling, optimal coordination is formulated as a Bayes-correlated equilibrium design problem. Signaling improves expected social cost and the price of anarchy, but the optimal recommendation rule is generally randomized, so it does not produce deterministic zero-error equilibrium selection (Iwase et al., 2016).

Across these models, exact coordination is therefore structurally heterogeneous. It may mean coalition-robust color matching, polynomial-time local search to a pure equilibrium, zero collisions under a public signal, or deterministic best-response convergence under an ordinal potential.

5. Communication, information constraints, and impossibility results

In two-player coordination with private values and pre-play cheap talk, zero-error is literally zero miscoordination. Each player chooses $u_2(a_1,a_2)=-u_1(a_1,a_2)$7 or $u_2(a_1,a_2)=-u_1(a_1,a_2)$8, with utility

$u_2(a_1,a_2)=-u_1(a_1,a_2)$9

and similarly for player 2. The key characterization states that three properties are equivalent: mutual-preference consistency, coordination, and binary communication on the one hand, and strong or weak communication-proof equilibrium on the other. Under any communication-proof equilibrium, players never miscoordinate,

$x\in\Delta(A_1)$0

they play their jointly preferred outcome whenever one exists, and they communicate only the ordinal part of their preferences. The set of communication-proof equilibria is indexed by the left tendency parameter $x\in\Delta(A_1)$1, which determines conflict resolution when preferences differ (Heller et al., 2020).

The opposite conclusion emerges in stochastic coordination under noisy observations. In homogeneous coordination games with binary actions, state

$x\in\Delta(A_1)$2

and omniscient benchmark action $x\in\Delta(A_1)$3, coordination efficiency is

$x\in\Delta(A_1)$4

Zero-error coordination is the case $x\in\Delta(A_1)$5. The paper proves the universal information-theoretic bound

$x\in\Delta(A_1)$6

For any strictly positive observation noise $x\in\Delta(A_1)$7, the conditional entropy term is positive, so zero-error coordination is impossible under any policy profile, including Bayesian Nash equilibrium policies. The same source also reports a nontrivial trade-off: in numerical experiments, Nash equilibrium policies can yield higher expected utility than certainty-equivalent policies while exhibiting lower coordination efficiency (Vasconcelos et al., 2023).

Approximate learning in multi-team zero-sum potential team games similarly falls short of exactness. Team-Fictitious Play and its independent-update variant converge almost surely to near team-Nash equilibrium, with

$x\in\Delta(A_1)$8

for Team-FP and

$x\in\Delta(A_1)$9

for Independent Team-FP, where $2\epsilon$00 as $2\epsilon$01. The irreducible term $2\epsilon$02 reflects entropy smoothing, so exact zero-error convergence is not established under the paper’s assumptions (Donmez et al., 2024).

These impossibility results delimit the domain of exact coordination. Positive observation noise, strategic nonstationarity, and entropy-based stabilization can all block literal zero error even when the equilibrium approximation is quantitatively strong.

6. Computational selection, distributed optimization, and quantum extensions

Zero-error coordination can also be posed as an exact computation problem. In integer programming games, each player chooses an integer vector from a finite feasible set and receives payoff $2\epsilon$03. The unilateral regret at profile $2\epsilon$04 is

$2\epsilon$05

and a pure Nash equilibrium is exactly a zero-regret profile. The ZERO Regrets framework builds a lifted MIP model, solves a master problem over the lifted feasible region, and uses an equilibrium separation oracle to add equilibrium inequalities of the form

$2\epsilon$06

Because the equilibrium closure coincides with the lifted perfect equilibrium formulation, the cutting-plane procedure can compute, enumerate, and optimize over pure Nash equilibria, including welfare-maximizing equilibria, and can certify nonexistence when the master becomes infeasible (Dragotto et al., 2021).

A distributed counterpart appears in monotone aggregative games with social-cost selection. Let

$2\epsilon$07

and define the optimal Nash equilibrium problem

$2\epsilon$08

The distributed coordination algorithm augments the Nash operator with a Tikhonov term $2\epsilon$09 and uses dynamical averaging to track $2\epsilon$10 and the aggregate social gradient. Under monotonicity and Lipschitz continuity of $2\epsilon$11, strong convexity of $2\epsilon$12, a connected undirected communication graph, and step sizes

$2\epsilon$13

the iterates converge to the unique optimal Nash equilibrium: $2\epsilon$14 Here zero-error means asymptotically exact convergence to the optimal element of the Nash set, not merely convergence to some equilibrium (Ma et al., 2022).

The strongest pathwise formulation occurs in Bayesian games with incomplete information and no communication after types are received. A finite Bayesian game admits zero-error Nash equilibrium coordination if, for every type profile $2\epsilon$15, the players output an action profile $2\epsilon$16 that equals some Nash equilibrium $2\epsilon$17 of the induced stage game with probability $2\epsilon$18: $2\epsilon$19 The cited quantum results identify games in which such coordination is impossible for any classical local-hidden-variable strategy, yet achievable with entanglement. In the tripartite game $2\epsilon$20, a GHZ state together with $2\epsilon$21 measurements yields

$2\epsilon$22

so every realized type profile leads with certainty to an equilibrium action triple. In the two-player game $2\epsilon$23, a stronger requirement is met by every two-qubit pure entangled state except the maximally entangled one: disallowed action pairs have zero probability, while every designated equilibrium action has strictly positive probability (Ambuj et al., 3 Sep 2025).

The same source also derives a noise-robust near-zero-error region. Under depolarizing noise on the shared state and measurements in the two-player game $2\epsilon$24, the equilibrium-violation probability is

$2\epsilon$25

Any classical deterministic strategy has error probability at least $2\epsilon$26 for $2\epsilon$27, so the quantum advantage persists whenever

$2\epsilon$28

This extends zero-error coordination from learning and mechanism design into the domain of nonlocal correlation resources (Ambuj et al., 3 Sep 2025).

Taken together, these computational and quantum formulations show that exact equilibrium coordination can be realized by exact oracle-based separation, by asymptotically exact distributed regularization, or by nonclassical correlation devices that satisfy a stage-wise equilibrium constraint impossible under classical shared randomness.