Bayesian Persuasion with a Risk-Conscious Receiver

TL;DR

This paper explores Bayesian persuasion where the receiver's preferences are based on CVaR, revealing new insights into the structure and computational complexity of optimal signaling under risk-sensitive criteria.

Contribution

It introduces a polynomial-size linear program for CVaR-based persuasion and characterizes the boundary between tractable and NP-hard cases.

Findings

CVaR preferences alter the incentive structure in persuasion.

An exact polynomial-size LP can optimize signals under CVaR.

Certain risk representations make the persuasion problem NP-hard.

Abstract

We study Bayesian persuasion when the receiver evaluates actions by reward-side Conditional Value-at-Risk (CVaR) rather than expected utility. CVaR preferences break the standard action-based direct-recommendation reduction: merging signals that recommend the same action can change the receiver's tail-risk ranking and destroy incentive compatibility. We show that this failure does not imply intractability in the explicit finite-state model. Each CVaR action value is max-affine in the posterior, and refining recommendations by the active affine piece yields an active-facet revelation principle and an exact polynomial-size linear program. We further identify a representation boundary: listed polyhedral risks remain tractable by the same LP, whereas succinctly represented facet families make exact persuasion NP-hard. Finally, we give a finite-precision approximation scheme for risk preferences determined by finitely many stable posterior statistics.