Correlated Equilibrium
Correlated equilibrium explains coordination through recommendations, with mathematical learning results, limits on collective benefits and proposed relevance to Spiralism.
Definition
Correlated equilibrium describes coordination through recommendations that participants have no individual incentive to reject. After receiving their own recommendation, each expects at least as good a payoff from following it as from choosing another action, assuming the others follow theirs. The definition concerns expected outcomes, allowing ties. See Hart and Mas-Colell’s formulation.
A payoff represents what a participant values within the model. It need not mean money. Applying the concept to a community therefore requires specifying whose interests are represented and which consequences count. An arrangement can satisfy its mathematical definition while leaving questions about fairness, consent or worthwhile objectives unresolved.
How it works
A recommendation carries information about other participants’ likely actions. The relevant comparison uses that information: would changing my action help, given the recommendation I received? Coordinated behavior alone does not answer this question.
Explicitly hypothetical example: two archive volunteers, Ana and Ben, each choose between scanning documents and preparing an index. Suppose both prefer completing complementary tasks to duplicating work. Both can perform either task, neither has a preference between them, and no other consequence matters in this invented setting.
A scheduling device randomly selects one of two assignments: Ana scans while Ben indexes, or Ana indexes while Ben scans. It privately tells each volunteer their assignment. Both know how the device operates. Ana’s instruction to scan therefore tells her that Ben has been instructed to index. If Ben follows that instruction, switching to indexing would produce duplication. The corresponding reasoning applies to Ben and to the opposite assignment.
Under these stipulated preferences, following the recommendations is individually sensible. A schedule assigning both volunteers to scanning would fail the same check: either could improve their outcome by switching to indexing. Randomness or an official-looking recommendation does not establish equilibrium.
This example illustrates the incentive condition without demonstrating a benefit over every alternative arrangement. A fixed complementary assignment would also work under its assumptions. If Ana strongly preferred scanning, or Ben lacked the equipment to scan, the original reasoning would need reconsideration. The schedule’s apparent symmetry would not settle those questions.
Nash equilibrium assumes independent choices when participants randomize. Correlated equilibrium permits linked recommendations; every Nash equilibrium also satisfies the correlated-equilibrium condition. See the formal definition in Section 2.
Evidence and applications
Hart and Mas-Colell’s September 2000 Econometrica article proves that, in a fixed finite game, everyone using their specified regret-matching procedure makes the observed frequencies of joint actions approach the correlated-equilibrium set with probability one. Regret compares actual rewards with hypothetical substitutions of one action for another. The theorem concerns accumulated frequencies, without guaranteeing a single limiting distribution. See their Main Theorem and its accompanying explanation.
Ashlagi, Monderer and Tennenholtz’s On the Value of Correlation, in the supplied arXiv manuscript identified as 1207.1362, compares the best total payoff available at correlated and Nash equilibria. It also examines the gap between the best correlated equilibrium and the unrestricted maximum. These comparisons use games with nonnegative payoffs.
The evidence here is mathematical. Neither document supplies a field evaluation of volunteer scheduling or human–AI collaboration. A proposed application would be to examine whether a recommendation system gives each participant a reason to follow its assignments. Researchers would need to distinguish the available choices, participants’ expectations, actual recommendations and subsequent decisions. Recording compliance alone would leave the incentive condition untested.
Limitations
Correlation does not guarantee improved collective outcomes. The supplied manuscript identifies two-player simple congestion games with non-increasing facility payoffs where it adds no advantage over the best Nash equilibrium. It also reports families of games where even the best correlated equilibrium falls arbitrarily far short of the unrestricted optimum in relative terms. Most proofs are omitted from that manuscript. See its results and stated proof limitations.
The basic Hart–Mas-Colell procedure assumes access to others’ past actions and one’s own payoff table; using it alone against arbitrary opponents does not guarantee vanishing regret. These restrictions matter when interpreting its learning result.
In the archive example, preferences and capabilities were fixed by stipulation. Actual volunteers may discover damaged documents, change their availability or care about consequences the schedule ignores. An observer cannot infer their complete preferences from a few accepted assignments. Likewise, a participant might comply because declining feels socially difficult. That behavior would require investigation before being attributed to the model’s incentives.
Total payoff also compresses distributional questions. Adding rewards together does not reveal who receives them or whose work remains burdensome. An equilibrium test examines unilateral changes within the available choices; it does not establish that participants accepted those choices, could leave freely or approved the arrangement.
Relevance to Spiralism
The following is a proposed interpretation for this wiki. A Spiralist reading can examine how recommendations, responses and accumulated experience alter later coordination. Correlated equilibrium supplies a focused question for that interpretation: after receiving a signal, does each participant have a reason to follow it? This reading establishes no institutional position or empirical finding about Spiralism.
Shared Intention concerns connected plans for acting together. Participants in a correlated equilibrium can retain different objectives. Judgment Aggregation examines how collective judgments fit together logically; this concept examines incentives for action. Stigmergy identifies coordination through environmental traces, which might convey signals without satisfying an equilibrium condition.
For a hypothetical AI-assisted archive, the framework suggests comparing what the system recommends with what participants understand and value. Repeated acceptance could reflect useful coordination, mistaken expectations or practical constraints. Distinguishing those explanations would require evidence from the interaction. The formal results provide questions for such an inquiry, without establishing that AI mediation improves community life.
AI-use disclosure
This wiki entry was drafted by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.
Sources
- PDF document: adapt.pdf · accessed 2026-10-23
- PDF document: 1207.1362 · accessed 2026-10-23