📈 economics

Coordination Mechanisms with Partially Specified Probabilities

This paper characterizes the outcomes implementable through coordination mechanisms that disclose only partial statistical information (specifically, expectations of finitely many variables) by showing that unrestricted message spaces yield jointly coherent outcomes while canonical mechanisms require the target outcome to satisfy a specific cross-entropy condition relative to a correlated equilibrium.

Original authors: Francesco Giordano

Published 2026-05-11

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Francesco Giordano

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to coordinate a group of friends to meet up for dinner. You know everyone's favorite foods (the "payoffs"), but you don't know exactly how your friends are communicating with each other. Maybe they are all reading the same news feed, or maybe they are just guessing what the others will do.

This paper, written by Francesco Giordano, explores what happens when people have to make decisions based on incomplete information about how that information was generated. Specifically, it looks at a world where people don't know the full "recipe" for how data is created, but they do know a few key ingredients (like the average or the "moments" of the data).

Here is the breakdown of the paper's ideas using simple analogies:

1. The "Maximum Entropy" Guessing Game

In the real world, when we don't know the full story, we often fill in the blanks with our best guess. This paper assumes people use a specific, very logical way of guessing called Maximum Entropy.

The Analogy: Imagine you are a detective who finds a few clues at a crime scene. You know the suspect was tall and wore a red hat, but you don't know their exact height or the shade of red.
The Rule: The "Maximum Entropy" rule says: "Don't invent extra details. Assume the suspect is any height and any shade of red that fits the clues, but treat all those possibilities as equally likely."
The Result: If you only know the average height of a group, this rule makes you assume everyone is exactly average. If you only know the individual heights but not how they relate to each other, this rule makes you assume they are all independent of one another. This is called "correlation neglect." The players act as if their friends' signals are unrelated, even if they actually are related.

2. The "Black Box" Data Generator

The paper imagines a scenario where an "Information Provider" (like a news algorithm or a corporate analyst) generates data for players.

The Provider knows the True Process (the exact probability of every outcome).
The Players only see a Coarse Summary (e.g., "The average recommendation was X," or "These two events never happen together").
Because the Players don't see the full picture, they use the "Maximum Entropy" rule to guess the full picture.

3. The Big Discovery: We Can Do More Than Just "Correlated Equilibrium"

In traditional game theory, there is a concept called Correlated Equilibrium. Think of this as a referee who whispers a secret plan to everyone. If the referee says "Go Left," everyone goes Left, and no one wants to change their mind because they trust the referee.

The paper asks: Can we get players to coordinate on outcomes that a standard referee couldn't achieve?

The Answer is Yes.
Because players are "naive" about how the data is correlated (they assume independence when it might not exist), a clever designer can trick them into coordinating on outcomes that are usually impossible.

The "Chicken" Game Example: Imagine two drivers heading toward each other. The best outcome is for one to swerve and the other to go straight. Usually, a referee can only get them to swerve with a certain probability.
The Paper's Trick: The designer reveals just enough information to make the drivers think, "Oh, the other guy's signal is totally random and unrelated to mine." Because they assume the other guy is acting independently, they might both decide to swerve (or both go straight) in a way that a standard referee couldn't force. The paper shows that by hiding the true correlation and letting players assume no correlation, we can achieve better or different results than standard game theory allows.

4. Two Types of "Mechanisms"

The paper looks at two ways this information can be shared:

Type A: The "Open Book" (Unrestricted Messages)
If the designer can send any kind of message (not just "Go Left" or "Go Right"), the paper proves that any outcome that is logically consistent (called "jointly coherent") can be achieved. Basically, if a result is possible under some reasonable belief system, this method can make it happen.
Type B: The "Direct Recommendation" (Canonical Mechanisms)
This is when the designer just says, "You should do Action A."
Here, the paper finds a very specific mathematical rule (involving something called Cross-Entropy). It says: "You can only implement a new outcome if it sits on a specific 'level set' of the old, standard outcomes."
- The Metaphor: Imagine the standard outcomes are a mountain peak. The new outcomes you can create are like a flat plateau that touches the peak at one specific point. You can't just go anywhere; you are stuck on that specific contour line defined by how much the players' "guessing" distorts the truth.

5. Why This Matters (In Simple Terms)

This paper explains how hiding the full story can actually be a powerful tool for coordination.

In Social Media: If an algorithm shows you posts but hides how many people saw them, you might assume your friends are acting independently. This could cause a "viral" effect where everyone jumps on a trend at once, even if the trend wasn't actually that popular.
In Finance: Traders might see analyst reports but not know how much the analysts are talking to each other. If they assume the reports are independent, they might all buy or sell the same stock, creating a market bubble that a fully informed market wouldn't have.
In Business: A committee might get reports from different departments. If they don't know how much the departments share data, they might treat the reports as independent, leading to a decision that looks "safe" to them but is actually highly correlated.

Summary

The paper argues that when people don't know the full "recipe" of how information is generated, they fill in the gaps by assuming things are independent. A smart designer can use this "blind spot" to coordinate people into doing things that would be impossible if everyone knew the full truth. It turns ignorance into a coordination device.

Technical Summary: Coordination Mechanisms with Partially Specified Probabilities

Problem Statement
This paper investigates the set of economic outcomes implementable in simultaneous-move games when players possess only partial knowledge of the data-generating process (DGP) governing their information. In many strategic settings—such as financial markets with shared analyst forecasts or social media coordination via opaque algorithms—players understand the payoff structure but face uncertainty regarding the correlation of their private signals. The central question is: what outcomes can be induced when a designer discloses only coarse statistics (moments) of the underlying distribution rather than the full joint distribution?

The paper models this as a mechanism design problem where an information provider generates data according to a true DGP $(M, \eta)$ but reveals only a feedback structure $F$ (a set of random variables whose expectations are known). Faced with a set of plausible DGPs consistent with these moment constraints, players form beliefs using the maximum-entropy principle (Shannon entropy). This heuristic leads to "correlation neglect," where players treat unknown correlations as independent as possible, given the known constraints.

Methodology and Framework
The author formalizes a framework combining mechanism design with information misspecification:

Partially Specified DGP: Defined as a triple $D = (M, \eta, F)$ , where $M$ is the message space, $\eta$ is the true distribution, and $F$ is the feedback structure. Players know $\mathbb{E}_\eta[f(m)]$ for all $f \in F$ .
Belief Formation: Players do not know $\eta$ but know the set of distributions $\Delta_D$ consistent with $F$ . They select a belief $q \in \Delta_D$ by maximizing Shannon entropy: $q = \arg\max_{q' \in \Delta_D} H(q')$ .
Implementation: An outcome $\mu \in \Delta_A$ $μ \in Δ_{A}$ is implemented if there exists a DGP and a strategy profile $\sigma$ $σ$ such that:
- Players form the maximum-entropy belief $q$ .
- Strategies constitute a best response to $q$ (incentive compatibility).
- The pushforward of the true DGP via strategies yields $\mu$ ( $\mu = \eta \circ \sigma$ ).

The analysis distinguishes between arbitrary message spaces (where messages need not be action recommendations) and canonical mechanisms (where $M=A$ and strategies are obedient).

Key Results

1. General Implementation (Arbitrary Message Spaces)
The paper establishes a characterization of implementable outcomes when the message space is unrestricted.

Joint Coherence: An outcome is implementable if and only if it is jointly coherent. A strategy profile is jointly coherent if it lies in the support of some correlated equilibrium (CE). An outcome is jointly coherent if it is a randomization over such profiles.
Expansion of Correlated Equilibria: The set of implementable outcomes strictly expands the set of correlated equilibrium outcomes. While standard correlated equilibria require the true distribution to satisfy incentive constraints, partially specified DGPs allow the belief (derived via max-entropy) to satisfy these constraints, even if the true distribution does not.
Approximation: For any jointly coherent outcome, there exists a finite partially specified DGP that $\epsilon$ -implements it. If the game has rational payoffs, the set of implementable outcomes coincides exactly with the set of jointly coherent outcomes.

2. Canonical Implementation (Direct Mechanisms)
When restricting attention to canonical mechanisms (where messages are action recommendations and players obey them), the characterization becomes more specific and relies on information geometry.

Cross-Entropy Condition: An outcome $\mu$ $μ$ is directly implementable if and only if there exists a correlated equilibrium $q$ $q$ such that:
1. $\text{supp}(\mu) \subseteq \text{supp}(q)$ .
2. $E_\mu[\log q] = E_q[\log q]$ .
Geometric Interpretation: This condition implies that $\mu$ must lie on the cross-entropy level set of some correlated equilibrium $q$ that passes through $q$ itself. Geometrically, this defines a linear constraint within the simplex.
Unique CE Case: If a game admits a unique correlated equilibrium $q^*$ , the set of directly implementable outcomes is the intersection of the support of $q^*$ and the hyperplane defined by the cross-entropy equality. This set is convex.

3. Illustrative Examples
The paper provides examples demonstrating how partial disclosure can sustain outcomes unattainable by standard correlated equilibria:

Chicken Game: A disclosure policy revealing only partial support allows players to form a uniform belief over a subset of actions, sustaining a degenerate outcome (both players playing the cooperative action) that yields higher welfare than the optimal correlated equilibrium.
Coordination Game: By disclosing only marginal distributions of a perfectly anti-correlated true DGP, players (via correlation neglect) form an independent belief. This allows the implementation of an anti-coordination outcome (miscoordination) that lies strictly outside the correlated equilibrium polytope.
Matching Pennies: The paper demonstrates how expanding the message set allows for the implementation of outcomes that would otherwise be impossible under canonical constraints, illustrating the flexibility of non-canonical mechanisms.

Significance and Claims
The paper claims to integrate classical mechanism design (Myerson, 1982) with recent advances in partial specification and information misspecification. Its primary contributions are:

Behavioral Foundation for Correlation: It provides a rigorous behavioral foundation for "correlation neglect" via the maximum-entropy heuristic, showing how coarse information disclosure acts as a coordination device.
Expansion of Implementable Sets: It demonstrates that the set of implementable outcomes is larger than the set of correlated equilibria, specifically capturing outcomes that are "jointly coherent."
Information-Geometric Characterization: In the canonical case, it derives a novel condition linking the target outcome to the set of correlated equilibria via cross-entropy, quantifying the extent to which a data-generating process can be "distorted" into a belief consistent with a correlated equilibrium.

The author notes that the framework applies to complete-information games with payoff-irrelevant messages. The paper modestly suggests that extending this to Bayesian games (where messages reveal payoff-relevant states) and dynamic games with learning are natural directions for future research, rather than claiming these extensions are currently solved. The results rely on the specific assumption of maximum-entropy belief formation; the paper does not claim these results hold for other ambiguity attitudes or heuristics.