Optimal Coarse Correlated Equilibria

• Optimal coarse correlated equilibria are distributions over actions that maximize social welfare within the set of CCEs, formulated via linear programming over an exponential polytope.
• They reveal a complexity trade-off where structured representations like aggregative or congestion games permit tractable algorithms, while general succinct games face NP-hardness.
• Algorithmic frameworks employing dual separation, regret minimization, and dynamic programming provide approximate solutions, even in extensive-form and mean-field settings.

Optimal coarse correlated equilibria are coarse correlated equilibria that optimize a prescribed objective over the set of all coarse correlated equilibria, most often social welfare. In an $n$-player $m$-action game with action profile set $A=A_1\times\cdots\times A_n$, a distribution $\sigma\in\Delta(A)$ is a coarse correlated equilibrium (CCE) if, for every player $p$ and every fixed deviation $j\in A_p$,

$$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$

The associated welfare objective is

$$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$

This optimization problem sits at the intersection of equilibrium computation, welfare maximization, and succinct game representation: exact CCEs are often easy to compute, but welfare-optimal CCEs can be computationally intractable, while remaining tractable in several important structured classes and alternative representations (Barman et al., 2015).

1. Normal-form formulation and optimization objectives

In normal form, an $n$-player $m$-action game is a tuple $m$0, where $m$1, each $m$2 has $m$3, and $m$4 is player $m$5's payoff. In the succinct model, one assumes an encoding $m$6 of length $m$7 and an oracle $m$8 such that

$m$9

in time $A=A_1\times\cdots\times A_n$0. Many natural games, including graphical, congestion, polymatrix, and anonymous games, admit such a representation (Barman et al., 2015).

The optimization problem is typically posed over the CCE polytope. In the notation of Jiang and Leyton-Brown, if $A=A_1\times\cdots\times A_n$1 is a distribution over pure profiles $A=A_1\times\cdots\times A_n$2, and $A=A_1\times\cdots\times A_n$3 is social welfare, then the maximum-welfare CCE problem is the linear program

$A=A_1\times\cdots\times A_n$4

This places optimal CCE computation within linear optimization over a polytope whose size is generally exponential in the normal-form representation, but whose dual admits a useful separation interpretation (Jiang et al., 2011).

Approximate variants are central in algorithmic work. A distribution $A=A_1\times\cdots\times A_n$5 is an $A=A_1\times\cdots\times A_n$6-CCE if, for every player $A=A_1\times\cdots\times A_n$7 and deviation $A=A_1\times\cdots\times A_n$8,

$A=A_1\times\cdots\times A_n$9

One also studies welfare guarantees relative to

$\sigma\in\Delta(A)$0

with an $\sigma\in\Delta(A)$1-CCE said to have welfare guarantee $\sigma\in\Delta(A)$2 if $\sigma\in\Delta(A)$3 (Barman et al., 2015).

2. Hardness of nontrivial welfare improvement in succinct games

The central negative result for optimal CCE in succinct multiplayer games is that computing any nontrivial welfare improvement can already be NP-hard. Barman and Ligett define the decision problem $\sigma\in\Delta(A)$4: given a succinct game $\sigma\in\Delta(A)$5 and a CCE $\sigma\in\Delta(A)$6 of minimum welfare, decide whether there exists a CCE $\sigma\in\Delta(A)$7 with $\sigma\in\Delta(A)$8. Their Theorem 3.1 shows that $\sigma\in\Delta(A)$9 is NP-hard. Equivalently, unless $p$0, there is no $p$1-time algorithm that, given $p$2, produces a CCE whose welfare strictly exceeds the worst possible CCE. The same paper proves analogous hardness for correlated equilibria, and shows that the hardness persists under egalitarian and Pareto objectives (Barman et al., 2015).

The approximation barrier is comparably strong. For any inverse-polynomial $p$3, it is NP-hard, even under randomized reductions, to compute an $p$4-CCE $p$5 satisfying

$p$6

The formulation given in the paper is that no $p$7-time algorithm can guarantee any nontrivial multiplicative improvement over the worst CCE, even when utilities lie in $p$8 and are encoded succinctly (Barman et al., 2015).

The proof strategy is a reduction from welfare maximization in a succinct game. Starting from a game $p$9, the construction adds to each player a “safe” action $j\in A_p$0, producing a new game $j\in A_p$1 in which the profile $j\in A_p$2 is a low-welfare CCE. Any CCE strictly better than this benchmark is then shown to encode either an original profile with welfare at least $j\in A_p$3 or a contradiction to equilibrium constraints via profitable deviations to safe actions. This construction isolates a recurring misconception: the existence of polynomial-time algorithms for computing some CCE does not imply the tractability of optimizing welfare over the CCE set (Barman et al., 2015).

3. Positive algorithmic frameworks in compact and structured games

Despite the general hardness results, several broad algorithmic frameworks identify settings in which optimal or near-optimal CCEs are tractable. Jiang and Leyton-Brown reduce the maximum-welfare CCE problem to the coarse deviation-adjusted social-welfare problem. For a nonnegative dual vector $j\in A_p$4, they define the coarse deviation-adjusted utility and welfare by

$j\in A_p$5

$j\in A_p$6

The dual separation problem is then equivalent to maximizing $j\in A_p$7 over pure profiles. Their sufficient-condition theorem states that if a compact representation admits a polynomial-time oracle for maximizing coarse deviation-adjusted social welfare, then one can compute a maximum-welfare CCE in polynomial time. They use this to prove tractability for singleton congestion games, where the oracle is implemented by a dynamic program running in $j\in A_p$8 time per oracle call (Jiang et al., 2011).

Barman and Ligett develop a complementary framework for approximate optimality. They introduce a regret vector $j\in A_p$9 whose coordinates index all unilateral deviations and a welfare-gap coordinate $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$0. For $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$1, they define the modified welfare

$$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$2

Using a generalized Blackwell approachability condition, they show that repeatedly solving approximate maximization subproblems for $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$3 yields an $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$4-CCE whose welfare is within $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$5 of $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$6. If each subproblem is solved to additive error $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$7, then $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$8 in $$\sum_{a\in A} [u_p(j,a_{-p})-u_p(a)]\cdot \sigma(a)\le 0.$$9 steps, and the output uniform mixture $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$0 satisfies $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$1. The total number of steps is $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$2 (Barman et al., 2015).

In aggregative games, this framework becomes fully polynomial. Here each payoff has the form $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$3 with $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$4, $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$5, and a Lipschitz bound in the public aggregator. The modified-welfare subproblem is solved by discretizing the aggregator grid to mesh $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$6, then using a dynamic program over players and partial sums. The resulting corollary is that in any aggregative game one can compute in $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$7 time an $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$8-CCE $$W(\sigma)=\mathbb E_{a\sim\sigma}\Big[\sum_{p=1}^n u_p(a)\Big].$$9 with

$n$0

This is one of the strongest positive results for welfare-optimal approximate CCE in succinct normal-form games (Barman et al., 2015).

4. Sequential and extensive-form variants

In extensive-form games, the relevant coarse notion is usually the normal-form coarse correlated equilibrium (NFCCE): a distribution over pure plans in the reduced normal form such that no player can improve ex ante by switching to a fixed pure plan. Celli, Marchesi, Bianchi, and Gatti show that, in two-player extensive-form games with perfect recall and no Chance, an optimal NFCCE can be computed in time polynomial in the size of the game tree. Their method uses an ellipsoid formulation whose separation oracle reduces to a polynomial-time tree computation, and they also give a more practical row-generation or column-generation approach based on pricing the same subproblem (Celli et al., 2019).

The tractability frontier is sharp. The same work proves that computing an NFCCE maximizing social welfare is NP-hard in two-player games with Chance moves, even with binary payoffs, and in general-sum games with at least three players and no Chance. Farina, Bianchi, and Sandholm refine this boundary by identifying a positive condition: in two-player perfect-recall games with public chance moves, the correlation-plan polytope $n$1 coincides with a polynomial-size mass-conservation polytope $n$2. Under this condition, optimal EFCE, EFCCE, and NFCCE can all be computed in polynomial time in the game size (Farina et al., 2020).

A later complexity classification establishes that, in multiplayer stochastic extensive-form games with perfect recall, the Threshold problem for NFCCE is NP-complete; the same is true for EFCCE and AFCCE. By contrast, the Threshold problem for normal-form correlated equilibrium (NFCE) is PSPACE-hard, even for fixed thresholds. The paper presents this as a “surprising complexity reversal”: in normal-form games, optimal correlated equilibria are computationally simpler than optimal Nash, whereas in extensive-form games computing optimal correlated equilibria is provably harder (Cheval et al., 15 Jul 2025).

The parameterized picture is also structured. For an information-complexity parameter $n$3, Zhang, Farina, and Sandholm construct a correlation DAG and show that all linear programs over $n$4 or the equilibrium polytope $n$5 can be solved in time $n$6 for $n$7. Their size bounds are

$n$8

with further simplification in games with public actions (Zhang et al., 2022).

Setting	Complexity statement	Source
Two-player EFG, no Chance	optimal NFCCE computable in polynomial time	(Celli et al., 2019)
Two-player EFG, public chance	optimal EFCE, EFCCE, and NFCCE computable in polynomial time	(Farina et al., 2020)
Multiplayer stochastic EFG, Threshold-NFCCE	NP-complete	(Cheval et al., 15 Jul 2025)
Multiplayer perfect-recall EFG, Threshold-NFCE	PSPACE-hard	(Cheval et al., 15 Jul 2025)
Parameter $n$9 from information structure	$m$0 algorithms	(Zhang et al., 2022)

5. Mean-field formulations and moderator-driven optimality

In mean field games, coarse correlation is introduced through a moderator who randomizes over recommendations and mean-field flows. In the linear-quadratic framework of Campi, Cannerozzi, and Cartellier, a moderator draws a random pair $m$1, where $m$2 is a recommended control and $m$3 is a possibly stochastic flow of mean fields. A player who commits uses $m$4, while a deviating player chooses $m$5 without knowledge of $m$6 beyond its law. The correlated flow is a CCE if it satisfies optimality,

$m$7

and consistency,

$m$8

The paper derives the best deviation in feedback form, proposes an ansatz for $m$9 involving an $m$00-measurable device-randomizer $m$01, and gives explicit conditions under which the resulting CCE strictly outperforms the mean-field Nash solution while remaining below the mean-field control optimum (Campi et al., 2023).

The emission-abatement example makes these comparisons concrete. With $m$02, $m$03, $m$04, $m$05, and $m$06, the tractable subclass $m$07 uses linear flows $m$08, $m$09, and $m$10. For the choice $m$11, $m$12, the final mean abatement is $m$13, whereas Nash–MFG yields $m$14 and mean-field control yields $m$15. The payoffs are reported as $m$16, $m$17, and $m$18, so CCE versus Nash gives welfare up by $m$19 and final abatement up by $m$20 (Campi et al., 2023).

A more recent formulation introduces optimal CCE explicitly in continuous-time mean field games. Campi, Cannerozzi, and Tzouanas define a moderator’s criterion

$m$21

which may differ from the representative player’s objective. They then encode the problem through an occupation-measure linear program over a compact convex set $m$22, define LP-CCE via external-regret inequalities $m$23 for all admissible deviations $m$24, and prove existence of an optimal LP-CCE under assumptions $m$25–$m$26. The associated Lagrangian formulation,

$m$27

leads to a primal-dual no-regret algorithm whose averaged iterate satisfies a primal-dual gap of $m$28, with both $m$29 and $m$30 (Campi et al., 18 Jun 2026).

6. Welfare benchmarks, learning connections, and conceptual boundaries

The welfare meaning of coarse correlation depends strongly on the domain. In single-item first-price auctions with full information, every correlated equilibrium is outcome-equivalent to a mixture of pure Nash equilibria, so social welfare equals $m$31 and revenue is at least $m$32. Coarse equilibria are weaker: welfare can drop to $m$33 of optimum, and revenue can fall to $m$34. Under a no-overbidding restriction, the welfare guarantee improves to

$m$35

and this bound is tight (Feldman et al., 2016).

These auction results clarify a common interpretive point. CCE is often attractive because it corresponds to ex ante deviations only, and therefore to weaker informational and communication requirements than correlated equilibrium. In sequential games, NFCCE is explicitly described as the “least-communication” form of ex ante correlation, since only one message of a full plan is sent before play (Celli et al., 2019). That informational economy, however, does not by itself imply strong welfare performance: in normal-form succinct games one cannot guarantee any nontrivial multiplicative improvement over the worst CCE unless $m$36 (Barman et al., 2015).

The learning perspective explains why the concept remains central despite these hardness barriers. In full-information general-sum Markov games, Mao, Wei, and Luo show that stage-based optimistic-follow-the-regularized-leader, together with appropriate value updates, finds $m$37-approximate CCE. Their returned policy $m$38 satisfies

$m$39

and setting $m$40 recovers the static normal-form case (Mao et al., 2024). A different frontier arises in quantum black-box models: Li, Wang, and Zhang give a quantum algorithm for $m$41-CCE in $m$42-player, $m$43-action normal-form games with query complexity $m$44, together with an $m$45 quantum lower bound for fixed $m$46. This does not optimize welfare, but it sharpens the complexity of equilibrium computation itself (Li et al., 19 Oct 2025).

Taken together, these results show that optimal coarse correlated equilibria do not admit a uniform complexity classification. In succinct normal-form games, welfare optimization is obstructed by hardness even for nontrivial improvement over the worst equilibrium. In compact representations with suitable separation or dynamic programming structure, exact optimality is attainable. In extensive-form games, the frontier is governed by information structure, chance observability, and parameterization. In mean-field models, moderator-driven objectives and no-regret or LP formulations permit a distinct notion of optimality. This suggests that the substantive content of “optimal CCE” is inseparable from representation, deviation model, and objective choice.