📈 economics

Robustness of Persuasion to Receiver Preferences

This paper demonstrates that Bayesian persuasion is generically both continuous and robust to infinitesimal Knightian uncertainty regarding receiver preferences, implying that while specific instances may be fragile, typical scenarios yield stable outcomes even when the sender lacks precise knowledge of those preferences.

Original authors: Ronen Gradwohl, Fengming Hu, Rann Smorodinsky

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ronen Gradwohl, Fengming Hu, Rann Smorodinsky

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 a Salesperson (the Sender) trying to convince a Judge (the Receiver) to make a specific decision. You know the true facts of the case (the "state of the world"), but the Judge does not. You can choose to share some facts, hide others, or present them in a specific way to influence the Judge's decision. This is the classic setup of Bayesian Persuasion.

In the standard textbook version of this game, everyone knows exactly how the Judge feels. The Judge has a precise "preference map" that tells them exactly which decision gives them the most happiness for every possible fact. The Salesperson knows this map perfectly and designs their pitch to get the best possible outcome.

The Big Question:
What happens if the Salesperson doesn't know the Judge's preference map perfectly? What if they only know that the Judge's happiness with a specific decision is "somewhere between 0.9 and 1.1," but they don't know the exact number?

This paper asks: Does the Salesperson's perfect strategy fall apart if they are slightly ignorant?

The authors explore two ways to look at this "slight ignorance":

1. The "Observer" vs. The "Player"

The paper distinguishes between two types of uncertainty:

Continuity (The Observer's Problem): Imagine the Salesperson does know the Judge's exact feelings, but you (the outside analyst trying to predict the outcome) don't. You only know the Judge's feelings are "roughly" in a small range.
- The Question: Can you still accurately predict what the Salesperson will get?
- The Metaphor: It's like trying to guess the winner of a race where you only know the runners' speeds are "around 10 mph." If the prediction holds up even with this fuzzy data, the model is Continuous.
Robustness (The Player's Problem): Now, imagine the Salesperson also doesn't know the Judge's exact feelings. They only know the "rough range."
- The Question: If the Salesperson has to play the game without knowing the exact rules, will they still get almost the same result as if they knew everything?
- The Metaphor: It's like playing a video game where the controls are slightly sticky or the screen is slightly blurry. If you can still win the level just as easily as if everything were perfect, the game is Robust.

2. How the Salesperson Handles Ignorance

Since the Salesperson doesn't know the exact rules, how do they decide what to do? The paper assumes they use one of two "safety-first" strategies:

The "Max-Min" Strategy: "I will choose the pitch that guarantees the best possible worst-case scenario." (I'll play it safe so I never get crushed.)
The "Regret Minimization" Strategy: "I will choose the pitch that minimizes my regret." (I want to avoid the feeling of thinking, "Oh no, if I had just said X instead of Y, I would have won big.")

3. The Main Findings

The paper delivers two surprising and comforting results:

Result A: Continuity and Robustness are Twins.
The authors prove that if the model is "Continuous" (the observer can predict the outcome), it is automatically "Robust" (the player gets the same outcome), and vice versa.

The Analogy: If you can see the finish line clearly through a slightly foggy window, the runner will also be able to see it clearly enough to win. You can't have one without the other.

Result B: It Usually Works (Genericity).
While there are rare, "knife-edge" cases where a tiny bit of ignorance causes a massive collapse in performance (like a house of cards falling with a breath), these cases are extremely rare.

The Analogy: Imagine a house of cards. Most of the time, if you blow a tiny bit of air (small ignorance), the house stays standing. There are only a few specific, weirdly balanced houses that will collapse instantly. In the real world, "typical" persuasion scenarios are stable. The Salesperson can be slightly ignorant, and they will still get almost the same reward as if they knew everything.

4. When Does It Fail? (The "Fragile" Cases)

The paper shows why it sometimes fails using a geometric metaphor.

The Stable Case: Imagine the Judge's preferences create a "safe zone" for a decision that is wide and open. If the Salesperson nudges the facts slightly, the Judge still stays in that safe zone. The outcome doesn't change.
The Fragile Case: Imagine the Judge's preference for a specific decision exists only on a single, razor-thin line (or a single point). If the Salesperson's ignorance shifts the facts even a tiny bit off that line, the Judge switches to a completely different decision, and the Salesperson loses everything.
- The Paper's Verdict: These "razor-thin line" scenarios are mathematically possible but statistically unlikely to happen by accident in a real-world setting.

Summary

The paper concludes that Bayesian Persuasion is surprisingly sturdy.
Even if the person trying to persuade you doesn't know your preferences perfectly, and even if they have to guess how to play to avoid regret, they will almost always get a result that is very close to what they would have gotten with perfect knowledge. The "perfect information" assumption isn't as fragile as one might fear; for most real-world situations, a little bit of ignorance doesn't break the system.

Technical Summary: Robustness of Persuasion to Receiver Preferences

1. Problem Statement

The paper investigates the robustness of the Bayesian persuasion framework (Kamenica and Gentzkow [15]) to uncertainty regarding the receiver's preferences. Standard Bayesian persuasion assumes that the sender and receiver share common knowledge of the receiver's utility function. However, in practice, the sender (and the modeler) may possess only imprecise knowledge of these preferences.

The authors distinguish between two conceptually distinct notions of robustness:

Continuity: The modeler lacks precise knowledge of the receiver's preferences (modeled as infinitesimal, non-probabilistic/Knightian uncertainty), but the sender possesses complete knowledge. The model is continuous if the modeler's predictions (based on the "true" benchmark model) remain accurate despite this ignorance.
Robustness: The sender also lacks precise knowledge of the receiver's preferences. The sender must choose a signal policy without knowing the exact receiver type, only that it lies within a small interval. The sender is assumed to behave according to either max-min utility (maximizing worst-case expected utility) or regret minimization (minimizing worst-case regret). The model is robust if the outcome under this ignorance is close to the benchmark outcome where all information is common knowledge.

The central question is whether the standard result—that the sender's maximal utility equals the concavification of their indirect utility function—holds when receiver preferences are subject to small, non-probabilistic uncertainty.

2. Methodology

Model Setup

The authors extend the standard Bayesian persuasion model $G = \{\Omega, A, \mu_0, u, v\}$ to a $\delta$ -BP model.

State Space: $\Omega$ (finite), prior $\mu_0$ .
Actions: $A$ (finite).
Uncertainty: The receiver's utility $u_\theta(a, \omega)$ for any action $a$ and state $\omega$ is unknown but lies within an interval $[u_{a,\omega}, \bar{u}_{a,\omega}]$ of length $\delta$ . The set of possible receiver types $\Theta$ corresponds to all possible utility profiles within these intervals.
Sender's Behavior under Ignorance:
- Max-min: The sender chooses $\pi'$ to maximize $\min_{\theta \in \Theta} E v_\theta(\pi')$ .
- Regret-minimization: The sender chooses $\pi^*$ to minimize $\max_{\theta \in \Theta} [E v_\theta(\text{opt}(\theta)) - E v_\theta(\pi)]$ .

Technical Tools

Geometric Analysis of Best-Reply Regions: The authors analyze the stability of the receiver's best-reply correspondence $B_\theta(a) = \{\mu : a \in \text{opt}_A^\theta(\mu)\}$ . They utilize properties of outer semicontinuity (the set does not expand discontinuously) and inner semicontinuity (the set does not shrink discontinuously) under small perturbations of utility parameters.
Adjustment Signal Policy: A technical device (inspired by Kamenica and Gentzkow [15] and Lipnowski et al. [19]) is employed. If a policy is optimal for a reference type $\theta_0$ , the authors construct a "perturbed" policy for a nearby type $\theta$ by adjusting posterior beliefs and adding a counterweighting belief to maintain the prior constraint. This allows bounding the utility loss due to type mismatch.
Lexicographic Perturbation: To handle cases where the receiver is indifferent among multiple actions (duplicate actions), the authors construct "pseudo-types" that impose specific tie-breaking rules (e.g., forcing the receiver to choose the sender's least preferred action among indifferent options) to demonstrate necessity.

3. Key Contributions and Results

3.1 Equivalence of Continuity and Robustness

Theorem 1 establishes a fundamental equivalence:

A Bayesian persuasion model is continuous if and only if it is robust (under both max-min and regret-minimization criteria).
This implies that if a modeler cannot accurately predict the outcome due to small ignorance (lack of continuity), then the sender, acting under that same ignorance, will also fail to achieve the benchmark outcome (lack of robustness).

3.2 Characterization of Robustness

The paper provides necessary and sufficient conditions for continuity and robustness based on the geometry of the receiver's best-reply regions at the reference type $\theta_0$ .

Sufficient Conditions (Proposition 1): A model is robust if there exists an optimal signal policy at $\theta_0$ $θ_{0}$ such that for every posterior belief $\mu_i$ $μ_{i}$ in its support, the induced action $a_i$ $a_{i}$ satisfies:
1. Property U1: The action is not "duplicate" for the receiver (i.e., there is no other action $b$ such that $u_{\theta_0}(a, \omega) = u_{\theta_0}(b, \omega)$ for all $\omega$ ), OR if duplicates exist, the sender is indifferent between them at that belief ( $U^S_1$ ).
2. Property U2: The best-reply region $B_{\theta_0}(a_i)$ has full dimension ( $N-1$ ).
Necessary Conditions (Proposition 2): If these conditions fail for every optimal policy, the model is neither continuous nor robust. Specifically, the sender's utility drops by a constant amount bounded away from zero as $\delta \to 0$ .

Mechanism of Failure:

Violation of U2 (Low-dimensional regions): If an optimal policy relies on a belief where the receiver is indifferent among actions in a way that creates a lower-dimensional best-reply set (e.g., a single point), small perturbations can cause this action to disappear entirely from the receiver's best replies.
Violation of U1 (Duplicate actions): If the receiver has duplicate actions (indifferent between them) and the sender strictly prefers one, small perturbations can break the tie in the sender's disfavor. The sender cannot guarantee the preferred action without knowing the exact tie-breaking rule.

3.3 Genericity

Theorem 2 states that continuity and robustness hold generically.

Treating utility values as random variables drawn uniformly from $[0, 1]$ , the set of models satisfying the robustness conditions has measure 1.
This result relies on the fact that the condition "every receiver-optimal action is uniquely optimal at some belief" (Lipnowski et al. [19]) is generic, which implies the satisfaction of U1 and U2 for the actions induced by an optimal policy.

4. Significance and Claims

The paper claims that while specific, "knife-edge" instances of Bayesian persuasion are fragile to uncertainty about receiver preferences, typical instances are robust.

Theoretical Implication: The standard concavification result is not an artifact of the strong assumption of common knowledge regarding preferences, provided the model is not structurally degenerate (i.e., does not rely on low-dimensional indifference sets or sender-preferred tie-breaking among duplicates).
Methodological Contribution: The work bridges the gap between "continuity" (modeler's perspective) and "robustness" (sender's strategic perspective) in information design, showing they are equivalent under infinitesimal Knightian uncertainty.
Behavioral Assumptions: The results hold regardless of whether the ignorant sender adopts a max-min or regret-minimization strategy, unifying these two decision-theoretic approaches in the context of persuasion.

The authors explicitly note that their results apply to models with finite message spaces and extend to large message spaces (Appendix A). They do not claim that all persuasion models are robust, but rather that the set of non-robust models is a set of measure zero in the space of utility profiles.