EFCE in Extensive-Form Games

• Extensive-form correlated equilibrium (EFCE) is a solution concept for games with imperfect information that uses a mediator to issue sequential, private recommendations at each information set.
• It ensures that following these recommendations is optimal by enforcing causal deviation constraints, preserving the game’s temporal and informational structure.
• EFCE computation leverages methods like linear programming, no-regret dynamics, and trigger-regret minimization to achieve efficient learning and equilibrium refinement.

to=arxiv_search 天天爱彩票提现json {"query":"Extensive-form correlated equilibrium EFCE arXiv", "max_results": 10} to=arxiv_search аԥсциалենք เติมเงินไทยฟรี with restriction? Extensive-form correlated equilibrium (EFCE) is a correlation concept for finite extensive-form games with imperfect information and perfect recall in which a publicly known correlation device samples a pure strategy profile and then issues private recommendations sequentially, only when players actually reach their information sets. If a player deviates, the device stops sending that player further recommendations. EFCE is defined so that following the recommendation stream is optimal against all causal deviations, making it the natural extensive-form counterpart to normal-form correlated equilibrium while preserving the temporal and informational structure of the game tree (Celli et al., 2020).

1. Definition and mediated play

An extensive-form game consists of a finite rooted tree of histories, terminal nodes $Z$, strategic players, and possibly a chance player with fixed known behavior. Imperfect information is represented by information sets $I \in \mathcal{I}_i$, and perfect recall ensures that each player remembers her own past actions and observations. A pure normal-form plan for player $i$ specifies one action at every information set $I \in \mathcal{I}_i$, and a joint plan is a profile of such plans (Celli et al., 2020).

In EFCE, the mediator samples a full joint plan $\pi = (\pi_i)_{i\in P}$ from a distribution $\mu \in \Delta(\Pi)$, but does not reveal the full plan upfront. Instead, when an information set $I$ of player $i$ is reached, the mediator privately recommends the action $\pi_i(I)$. Recommendations are therefore local and sequential. A defining feature is the unsubscribe rule: after a deviation, the player receives no further recommendations, although other players continue to do so (Marchesi et al., 2020).

A standard equivalent definition uses an extended game. Given an extensive-form game $\Gamma$ and a distribution $I \in \mathcal{I}_i$0, the extended game $I \in \mathcal{I}_i$1 first lets chance draw $I \in \mathcal{I}_i$2 according to $I \in \mathcal{I}_i$3, and then implements the recommendation process as the game unfolds. A distribution $I \in \mathcal{I}_i$4 is an EFCE if following recommendations is a Nash equilibrium of $I \in \mathcal{I}_i$5 (Marchesi et al., 2020).

This mediator model is weaker than immediate revelation: players observe only the currently relevant recommendation, not the entire contingent strategy profile. That distinction is essential. In extensive form, what can be inferred from future recommendations, and whether those recommendations remain available after a deviation, changes the deviation set and therefore the equilibrium concept itself (Zhang et al., 2024).

2. Causal deviations and obedience constraints

EFCE is most transparently characterized through trigger, or causal, deviations. Fix player $I \in \mathcal{I}_i$6, a trigger sequence $I \in \mathcal{I}_i$7, and a continuation distribution $I \in \mathcal{I}_i$8. A trigger agent follows recommendations unless player $I \in \mathcal{I}_i$9 reaches $i$0 and is recommended action $i$1; if that happens, the agent stops following recommendations and instead continues from $i$2 onward according to a plan drawn from $i$3 (Celli et al., 2020).

Let

$i$4

the probability of terminal node $i$5 when everyone follows the mediator, and

$i$6

the probability of $i$7 when player $i$8 triggers at $i$9 and then switches to $I \in \mathcal{I}_i$0 (Celli et al., 2020).

Then $I \in \mathcal{I}_i$1 is an EFCE if, for every player $I \in \mathcal{I}_i$2 and every $I \in \mathcal{I}_i$3 with $I \in \mathcal{I}_i$4,

$I \in \mathcal{I}_i$5

These are local incentive constraints conditioned on the recommendation at the trigger information set. They differ fundamentally from normal-form internal-swap constraints, because a deviation in extensive form changes the future information and recommendation process itself. Internal regret over single-action swaps is therefore insufficient; the correct deviation class must allow trigger-at-$I \in \mathcal{I}_i$6-then-arbitrary-continuation deviations (Celli et al., 2020).

A common misconception is that EFCE is simply correlated equilibrium applied to the normal-form expansion of the game. That is not the case. EFCE restricts what a player learns and when the player learns it; recommendations are causally revealed along the realized path, and a deviating player loses access to future recommendations. Those two features are exactly what the trigger-agent inequalities encode (Farina et al., 2019).

3. Position among related correlation concepts

The main neighboring notions differ by what the mediator reveals and by when a player must decide whether to follow the recommendation stream.

Concept	Recommendation timing	Deviation timing
NFCCE	Full plan treated in normal form	Commit before seeing any recommendation
EFCCE	Sequential recommendations	Commit at an infoset before seeing that recommendation
EFCE	Sequential recommendations	Deviate after seeing the local recommendation
BCE	Sequential recommendations continue even after deviations	Counterfactual obedience at every infoset

For extensive-form coarse correlated equilibrium, the player decides at an information set whether to commit before seeing the recommended move. For normal-form coarse correlated equilibrium, the decision to commit is made once at the outset. The inclusion chain

$I \in \mathcal{I}_i$7

is established in the literature summarized in (Celli et al., 2020, Farina et al., 2019). The weakening from EFCE to EFCCE and then to NFCCE is entirely due to the weakening of the deviation class.

Behavioral correlated equilibrium (BCE) uses a different signaling rule: the mediator continues to make private recommendations even if a player has deviated earlier, and obedience is formulated through counterfactual regret at each information set. Despite the different operational model, EFCE and BCE are outcome-equivalent: every outcome distribution achievable under one is achievable under the other, and this equivalence yields a polynomial-time route to computing BCE by first computing EFCE and then converting it (Zhang et al., 2024).

The relation to immediate-revelation extensive-form correlated equilibrium is subtler. Immediate revelation exposes the entire plan at time $I \in \mathcal{I}_i$8, effectively collapsing the concept toward normal-form correlation. In complete-information extensive-form games, explicit separations show that some outcome distributions are implementable by EFCE but not by immediate-revelation EFCE, and conversely that quantum correlated equilibrium can implement outcomes not implementable by EFCE or immediate-revelation EFCE (Deckelbaum, 2011).

4. Correlation plans, sequence form, and exact computation

A central algorithmic insight is that EFCE can be represented through correlation plans over relevant sequence pairs rather than explicit distributions over exponentially many pure normal-form profiles. In two-player perfect-recall games, the von Stengel–Forges polytope imposes nonnegativity and flow-conservation constraints over relevant sequence pairs; in settings where the feasible correlation-plan polytope coincides with that relaxation, optimal correlated equilibria can be computed by linear programming (Farina et al., 2020).

For two-player games without chance, EFCE can also be cast as a bilinear saddle-point problem between a mediator choosing a correlation plan and a deviator choosing a trigger agent. This yields projected subgradient methods and exposes sequence-form structure that is hidden in the normal-form representation (Farina et al., 2019). A different exact route constructs the space of correlation plans incrementally via the scaled-extension operation, showing that a regret minimizer can be built over the correlation-plan polytope itself while maintaining feasibility of every iterate (Farina et al., 2019).

In general $I \in \mathcal{I}_i$9-player extensive-form games with chance, EFCE is computable in polynomial time via an ellipsoid-against-hope approach. The underlying linear program has exponentially many variables $\pi = (\pi_i)_{i\in P}$0 but only polynomially many constraints, and the separation oracle exploits the trigger structure of deviations (Marchesi et al., 2020). When the objective is not merely feasibility but optimality, the complexity landscape is more delicate. In two-player perfect-recall games, optimal EFCE, EFCCE, and NFCCE are polynomial-time computable when the game is triangle-free; in particular, this holds whenever all chance moves are public (Farina et al., 2020).

More recent work gives two complementary exact or near-exact approaches. Fixed-parameter algorithms use a correlation DAG whose size depends exponentially only on a parameter tied to the information structure, while two-sided column generation re-optimizes sequence-form strategies against previously added correlation plans and outperforms earlier one-sided methods in many regimes (Zhang et al., 2022). For online refinement, safe subgame resolving shows that EFCE possesses enough independence between subgames to allow local re-solving with guarantees that exploitability does not worsen relative to a blueprint correlation plan (Ling et al., 2022).

5. Learning dynamics and approximation rates

The learning-theoretic interpretation of EFCE is organized around trigger regret. For player $\pi = (\pi_i)_{i\in P}$1 and trigger sequence $\pi = (\pi_i)_{i\in P}$2, trigger regret compares the realized continuation value to the best hindsight continuation plan from $\pi = (\pi_i)_{i\in P}$3, but only on rounds in which $\pi = (\pi_i)_{i\in P}$4 would actually have triggered. If the empirical distribution of repeated play is $\pi = (\pi_i)_{i\in P}$5, then small maximum trigger regret implies that $\pi = (\pi_i)_{i\in P}$6 is an approximate EFCE; in fact, the maximum EFCE violation equals the maximum per-round trigger regret (Celli et al., 2020).

This observation led to the first uncoupled no-regret dynamics converging to EFCE in multiplayer general-sum perfect-recall games. The ICFR algorithm decomposes trigger regret into local regrets at information sets, using internal regret when an infoset is reachable under the sampled plan and external regret otherwise. The conference version established almost-sure convergence of empirical play to the EFCE set (Celli et al., 2020), and the journal version sharpened this to an $\pi = (\pi_i)_{i\in P}$7-approximate EFCE with high probability after $\pi = (\pi_i)_{i\in P}$8 repetitions, while keeping per-iteration computation polynomial in the size of the game tree (Farina et al., 2021).

Subsequent work connected EFCE learning to normal-form $\pi = (\pi_i)_{i\in P}$9-regret minimization. For the family of causal deviations, $\mu \in \Delta(\Pi)$0-Hedge in the exponentially large normal form is exactly equivalent to online mirror descent in sequence form with trigger-dilated regularizers. This equivalence produces polynomial-time EFCE learning directly in extensive form and yields the first sharp bandit-feedback rate

$\mu \in \Delta(\Pi)$1

for EFCE-regret, matching the information-theoretic lower bound up to polynomial factors in the horizon (Bai et al., 2022).

A separate line introduces $\mu \in \Delta(\Pi)$2-EFCE, where players may observe and deviate from recommended actions for $\mu \in \Delta(\Pi)$3 times. EFCE is recovered at $\mu \in \Delta(\Pi)$4, and the notion becomes stricter as $\mu \in \Delta(\Pi)$5 increases. In particular, the bandit-feedback algorithm for $\mu \in \Delta(\Pi)$6-EFCE yields, when specialized to $\mu \in \Delta(\Pi)$7, the first sample-efficient algorithm for learning EFCE from bandit feedback, with episode complexity $\mu \in \Delta(\Pi)$8 (Song et al., 2022).

The fastest current no-regret dynamics for EFCE use predictions and a refined perturbation analysis of the fixed points associated with trigger deviation functions. When all players follow these accelerated dynamics, the empirical correlated distribution of play is an $\mu \in \Delta(\Pi)$9-approximate EFCE, improving on the earlier $I$0 rate (Anagnostides et al., 2022).

6. Refinements, robustness, applications, and open directions

Although EFCE is the standard extensive-form correlation concept, it inherits the off-path weaknesses of Nash-based notions in sequential games. Because recommendations cease after a deviation, EFCE guarantees best-response behavior along the equilibrium path but does not ensure sequential rationality at contingencies made reachable by small mistakes. Extensive-form perfect correlated equilibrium (EFPCE) addresses this by imposing trembling-hand perfection in the extended game: recommendations must remain credible under independent mistakes at each information set. The refinement relation is

$I$1

and, perhaps surprisingly, EFPCE can still be computed in polynomial time in general $I$2-player extensive-form games with chance (Marchesi et al., 2020).

Empirical benchmark work has shown that EFCE supports qualitatively sequential mediated behavior that does not arise in one-shot correlation. In Battleship, welfare-improving EFCEs use passcodes and punishments to deter deviations. In Sheriff of Nottingham, social-welfare-maximizing EFCEs exhibit non-monotonic and discontinuous responses to parameter changes, and additional bargaining rounds can act as passcodes that stabilize high-welfare mediated behavior (Farina et al., 2019). These examples underscore that EFCE is not merely a larger feasible set than Nash equilibrium; it changes the strategic use of information revelation over time.

The scope of exact and approximate EFCE computation has also expanded beyond tree-form games. In two-player turn-taking finite-horizon stochastic games in succinct graph form, there is an algorithm for approximately computing an optimal EFCE up to machine precision with runtime polynomial in the game size and in $I$3, as well as a polynomial-time algorithm for Stackelberg extensive-form correlated equilibrium (Zhang et al., 2024). This suggests that extensive-form correlation remains tractable in some graph-form stochastic settings where tree expansion would be prohibitive.

Open directions remain. The literature explicitly identifies compact correlated representations for EFPCE, broader trembling-hand refinements of correlated solution concepts, stronger optimistic rates for EFCE learning, extensions beyond perfect recall, and wider generalizations of efficient computation beyond triangle-free or public-chance settings as unresolved problems (Marchesi et al., 2020, Bai et al., 2022, Farina et al., 2020).