Incomplete Information Robustness

This paper introduces a notion of robustness for belief-invariant Bayes correlated equilibria in incomplete information games and establishes that a generalized potential function provides a sufficient condition for such robustness, implying that in supermodular potential games, robust equilibria coincide with Bayes Nash equilibria.

The Gist

Imagine you are a weather forecaster trying to predict tomorrow's weather. You have a model based on the data you can see: temperature, humidity, and wind speed. But you know your model isn't perfect. Maybe there's a hidden correlation between the clouds and the wind that you missed, or maybe your sensors are slightly off.

In the world of economics and game theory, this "weather forecaster" is an analyst trying to predict how people (players) will behave in a strategic situation, like a business negotiation or a traffic jam. The paper by Stephen Morris and Takashi Ui asks a crucial question: If our model of the world is slightly wrong, can we still trust our predictions?

Here is a breakdown of their findings using simple analogies.

1. The Problem: The "Hidden Correlation"

In the real world, people often have secret connections or shared information that an outside observer (the analyst) doesn't see.

• The Analyst's View: "Player A and Player B both think it's going to rain, so they will both bring umbrellas."
• The Reality: Maybe Player A and Player B are actually siblings who secretly texted each other, or maybe they both saw the same cloud formation that the analyst missed.

The paper argues that if an analyst assumes players are acting purely on their own private thoughts (what economists call "Bayes Nash Equilibrium"), they might be wrong. Why? Because in a slightly different version of reality (a "nearby game"), players might be coordinating their actions in a way that looks like magic to the analyst, but is actually just a result of hidden signals.

2. The Solution: The "Belief-Invariant" Crystal Ball

The authors propose a new way to make predictions called Belief-Invariant Bayes Correlated Equilibrium (BIBCE).

Think of this as a Magic Correlation Device. Imagine a referee who whispers a suggestion to each player.

• If the referee tells Player A, "Go left," and Player B, "Go right," they do it.
• The Catch: The referee's whisper must not change what the players believe about the world. If Player A hears "Go left," they shouldn't suddenly think, "Oh, Player B must know something I don't!" The suggestion must be "belief-invariant."

The paper shows that if you look for outcomes that are Robust (meaning they hold up even if the analyst's model is slightly tweaked or if hidden correlations exist), you should look for these "Magic Correlation" outcomes, not just the standard "everyone acts alone" outcomes.

3. The "Potential Function": The Mountain Climber

How do you find these robust outcomes? The authors use a mathematical tool called a Generalized Potential Function.

The Analogy: Imagine the game is a landscape with hills and valleys.

• The Players: Hikers trying to find the highest peak (the best outcome).
• The Potential Function: A giant map that shows the height of the terrain.
• The Rule: If the players are smart, they will naturally try to climb to the highest point on this map.

The paper proves a powerful result: If a game has a "Potential Map" (a potential game), the outcomes that maximize the height of this map are the most robust. Even if the hikers have slightly different maps or hidden paths, they will still end up at the same peak.

4. The Big Surprise: "Doing Nothing" vs. "Coordinating"

In many classic games, the "best" outcome is when everyone acts rationally on their own. But this paper finds something surprising in incomplete information games:

• The Standard Prediction (BNE): "I think you will do X, so I will do Y." (This is fragile; if my guess about your guess is wrong, the whole thing collapses).
• The Robust Prediction (BIBCE): "We are both following a hidden script that tells us to coordinate, even if we don't know exactly why."

The Motivating Example:
The authors use a game where two people need to coordinate.

• In the "standard" model, there are infinite ways they could coordinate, and none of them are stable. If you tweak the model slightly, the prediction changes completely.
• In the "robust" model, there is one specific outcome where they coordinate perfectly. This outcome survives even when the model is tweaked. It turns out this robust outcome looks like a "correlated equilibrium" (a hidden script), not a standard equilibrium.

5. When Does "Acting Alone" Work?

The paper asks: "Is there ever a case where the standard 'act alone' prediction is robust?"

• Answer: Yes, but only in very specific, "supermodular" games (games where your best move gets better if your opponent's move gets better, like investing in a stock when everyone else is investing).
• In these specific cases, if there is a unique "best" outcome that maximizes the potential map, then that outcome is robust. But in most other games, you must rely on the "hidden script" (BIBCE) to get a reliable prediction.

Summary for the Everyday Person

Imagine you are trying to predict the outcome of a chaotic traffic jam.

• Old Way: You assume every driver is a genius who only looks at their own GPS. You predict a mess because everyone is guessing what others will do.
• New Way (This Paper): You realize drivers might be reacting to subtle, shared cues (like a police officer's hand signal or a radio broadcast) that you can't see. You predict that they will coordinate to clear the jam in a specific way.
• The Takeaway: If you want a prediction that won't break when you realize you missed a tiny detail in the model, you shouldn't assume people are acting in isolation. You should assume they are following a robust, coordinated pattern that maximizes the overall "goodness" of the situation, even if that pattern involves hidden coordination.

In short: To make predictions that survive reality checks, stop assuming people are isolated islands. Assume they are part of a coordinated system, and look for the outcome that makes the whole system work best. That is the only prediction that is truly "robust."

Technical Summary

Here is a detailed technical summary of the paper "Incomplete Information Robustness" by Stephen Morris and Takashi Ui (February 2026).

1. Problem Statement and Motivation

The Core Challenge:
When an analyst models a strategic situation as a game of incomplete information, they must specify the players' "types." However, types can be decomposed into two components:

1. Belief Hierarchies: The infinite sequence of beliefs about the state, beliefs about others' beliefs, etc. (Mertens-Zamir hierarchies).
2. Correlating Devices: Mechanisms that may create correlations between players' types that are not captured by their individual belief hierarchies (e.g., duplicated belief hierarchies).

The analyst faces two epistemic limitations:

• They may only have approximate knowledge of the true belief hierarchies.
• They lack information regarding the correlation structure (i.e., whether types are duplicated or correlated in ways unknown to the analyst).

The Question:
If the true game differs slightly from the analyst's model (due to unknown correlations or slight belief discrepancies), which predictions about player behavior remain justifiable? Specifically, the authors seek to identify outcomes that are robust to small perturbations in the information structure, including the potential existence of hidden correlating devices.

Motivating Example:
The authors illustrate that standard Bayes Nash Equilibria (BNE) are often not robust. In a specific coordination game, the analyst's model has many BNEs. However, a nearby game (an "elaboration" with a specific correlation structure) has a unique BNE that is qualitatively different from any BNE in the analyst's model. The only outcome consistent with the nearby game is a Bayes Correlated Equilibrium (BCE) that is belief-invariant (BIBCE). This suggests that BIBCE, rather than BNE, is the appropriate solution concept for robust prediction under incomplete information.

---

2. Methodology and Framework

The paper develops a rigorous framework to define and analyze robustness in incomplete information games.

A. Key Definitions

• Elaboration: A game $(\bar{T}, \Theta, \bar{\pi}, u)$ is an elaboration of the analyst's model $(T, \Theta, \pi, u)$ if it is isomorphic to the conjunction of the original game and a belief-invariant communication rule. In an elaboration, players may have duplicated types or correlated types, but their belief hierarchies over the state $\Theta$ remain identical to the original model.
• $\varepsilon$-Elaboration: A "nearby" game where the belief hierarchies are approximately those of an elaboration. Formally, with probability at least $1-\varepsilon$, players have types that map to the original types such that their beliefs and payoffs are close to the original model.
• Belief-Invariant Bayes Correlated Equilibrium (BIBCE): A decision rule where:
  1. Obedience: Players maximize expected utility given the recommendation.
  2. Belief-Invariance: The recommendation does not reveal additional information about opponents' types or the state beyond what the player already knows.
  • Significance: BIBCE encompasses all outcomes achievable by BNE in any elaboration of the model.

B. The Robustness Criterion

A set of BIBCE is defined as robust if, for every sufficiently small $\varepsilon > 0$, every $\varepsilon$-elaboration possesses a BIBCE that is close (in distribution) to some BIBCE in the set.

• If a single BIBCE is robust, it is the only justifiable prediction.
• The authors show that the set of BNE is generally not robust because nearby games may introduce correlations that eliminate certain BNEs while preserving BIBCEs.

C. Generalized Potential Functions

To derive sufficient conditions for robustness, the authors introduce Generalized Potential Functions (extending Morris and Ui, 2005).

• Domain: Defined on the Cartesian product of the state space $\Theta$ and a covering $\mathcal{A}$ of the action space. A covering consists of subsets of actions (signals) rather than single actions.
• Mechanism: A generalized potential function $F$ maps a profile of action subsets and a state to a real number.
• GP-Maximizing BIBCE: A BIBCE where players choose actions from the subsets recommended by a communication rule that maximizes the expected value of the generalized potential function $F$.

---

3. Key Contributions and Results

Main Theorem (Theorem 1)

Statement: If a non-redundant incomplete information game admits a generalized potential function $F$, then the set of GP-maximizing BIBCE is non-empty and robust.

Implications:

1. Existence: Robust outcomes always exist in games with generalized potentials.
2. Uniqueness: If the game has a unique BIBCE (which implies it is the unique GP-maximizer), that equilibrium is robust.
3. Contrast with Complete Information: In complete information games (Kajii and Morris, 1997), a unique potential-maximizing Nash Equilibrium is robust. Here, the authors show that in incomplete information, the robust outcome is a BIBCE, which may not be a BNE.

Application to Potential Games (Corollary 3)

If the game is a standard Potential Game (admitting a potential function $v$), the set of BIBCE that maximize the expected value of $v$ is robust.

• Example: In the motivating coordination game, the BIBCE described in Table 2 is the unique potential-maximizing BIBCE and is therefore robust, even though no BNE is robust.

Supermodular Games and BNE Robustness (Section 6)

The paper investigates when a Bayes Nash Equilibrium (BNE) itself can be robust.

• Proposition 1: In supermodular potential games, if a P-maximizing BIBCE is unique, it is a BNE. Thus, in supermodular potential games with a unique potential-maximizing BNE, that BNE is robust.
• Binary-Action Supermodular Games: The authors apply Monotone Potential Functions (Morris and Ui, 2005).
  • They establish a link between the robustness of the "all-1" action profile and the existence of a monotone potential function satisfying specific conditions (G1, G2).
  • Result: For generic binary-action supermodular games, the "all-1" profile is robust if and only if it is a monotone potential maximizer.
  • This extends the "critical path theorem" results from complete information (Oyama and Takahashi, 2020) to incomplete information settings.

---

4. Technical Significance and Contributions

1. Shift from BNE to BIBCE: The paper fundamentally shifts the paradigm for robustness analysis in incomplete information. It demonstrates that BNE is too fragile because it cannot account for hidden correlations (duplicated types) in nearby games. BIBCE is the correct equilibrium concept for robustness because it internalizes all possible belief-invariant correlations.
2. Generalization of Potential Theory: The authors successfully extend the theory of potential games to the realm of incomplete information and belief-invariant correlations. They prove that the "potential-maximizing" logic, which guarantees robustness in complete information, holds for BIBCE in incomplete information, provided one uses the generalized potential framework.
3. Bridging Literature: The work connects three major strands of literature:
   • Robustness of Equilibrium: Extending Kajii and Morris (1997) to incomplete information.
   • Correlated Equilibrium: Utilizing the framework of Bergemann and Morris (2013, 2016) and Liu (2015) regarding belief-invariance.
   • Supermodular Games: Extending results on monotone potentials (Oyama and Takahashi, 2020) to settings with incomplete information and correlated types.
4. Methodological Rigor: The proof techniques involve constructing sequences of $\varepsilon$-elaborations, utilizing fixed-point theorems (Kakutani-Fan-Glicksberg) on spaces of decision rules, and employing tightness arguments in probability spaces to handle the convergence of equilibria as perturbations vanish.

5. Conclusion

The paper concludes that for an analyst with imperfect knowledge of belief hierarchies and correlation structures, the only justifiable predictions are those that are robust. The set of GP-maximizing BIBCE provides a sufficient condition for such robustness. In many cases, particularly in supermodular games, this robust outcome coincides with a specific BNE, but in general, the robust prediction is a correlated equilibrium that cannot be sustained as a BNE in the analyst's specific model. This highlights the necessity of considering correlated deviations and belief-invariance when modeling strategic interactions under uncertainty.