Dynamic Quadratic Cheap Talk and Signaling Games.

Serkan Sarıtaş, Serdar Yüksel, Sinan Gezici

Abstract

Simultaneous (Nash) and sequential (Stackelberg) equilibria of two-player dynamic quadratic cheap talk and signaling game problems are investigated under a perfect Bayesian formulation. For the dynamic scalar cheap talk, a zero-delay communication setup is considered for i.i.d. and Markov sources; it is shown that the final stage equilibrium is always quantized and under further restrictive conditions the equilibria for all time stages are quantized. Contrarily, the Stackelberg equilibria are always fully revealing for both scalar and multi-dimensional sources. In the dynamic signaling game where the transmission of a Gauss-Markov source over a memoryless Gaussian channel is considered, affine policies constitute an invariant subspace under best response maps for both scalar and multi-dimensional sources under Nash equilibria; whereas the Stackelberg equilibria always admit linear policies for scalar sources but such policies may be non-linear for multi-dimensional sources. A dynamic programming formulation is presented for multi-dimensional sources for optimal linear encoding policies, and conditions under which the Stackelberg equilibria are non-informative are derived.

Introduction

In classical communication theory, the sender (encoder) and receiver (decoder) are typically assumed to be cooperative agents sharing a common goal of perfect information transfer. However, in many strategic environments—such as economic signaling or strategic sensors in smart grids—the objectives of the two players may be misaligned. This paper addresses the dynamic (multi-stage) versions of Cheap Talk and Signaling Games, where a sender with a specific bias attempts to influence a receiver's actions over time.

Block diagrams for (a) 2-stage cheap talk and (b) 2-stage signaling game models.

While single-stage games often result in quantized communication—where the sender provides vague categories rather than precise data—this study investigates how these interactions evolve when the information source follows a Markov process or a sequence of independent variables.

Methodology and Models

The research utilizes a Perfect Bayesian Equilibrium framework to analyze two primary equilibrium concepts:

  1. Nash Equilibrium: A simultaneous game where neither the encoder nor the decoder has a leadership advantage.
  2. Stackelberg Equilibrium: A sequential game where the encoder (leader) commits to a strategy first, and the decoder (follower) reacts.

These concepts are applied to two communication scenarios. In Cheap Talk, communication is costless, and the only penalty arises from the misalignment of final objectives. In Signaling Games, communication is costly, often modeled as a power constraint on the transmitted signal. The authors employ mathematical techniques such as invariant subspace analysis and dynamic programming to derive optimal policies for both scalar and multi-dimensional sources.

Analysis of Equilibria

The Persistence of Quantization in Nash Cheap Talk

A key finding is that in Nash cheap talk, the inherent mistrust between players necessitates quantized equilibria. For scalar i.i.d. sources, the final stage of any multi-stage game must be quantized. Under specific conditions, this quantization persists across all time stages, preventing perfectly clear communication when players move simultaneously.

The Power of Commitment in Stackelberg Games

In contrast, the ability of a sender to commit to a policy beforehand in a Stackelberg game effectively resolves the misalignment. The authors prove that Stackelberg cheap talk equilibria are fully revealing for both scalar and multi-dimensional sources. By committing to a specific encoding rule, the sender forces a situation where the receiver can perfectly decode the information, achieving the same efficiency as a cooperative system.

Affine Invariance and Linearity

In Gaussian signaling games, the study identifies that affine policies (linear transformations with a constant offset) form an invariant subspace for Nash equilibria. While linear policies are shown to be optimal for scalar sources in Stackelberg signaling games, they may not be optimal for multi-dimensional sources. The paper provides recursive formulas for computing optimal linear encoding policies in these multi-dimensional settings.

Informativeness Thresholds

The research also establishes thresholds for informativeness in signaling games. If the cost of signaling is too high relative to the variance of the source, the only possible equilibrium is "babbling," where the encoder's signal contains no useful information and is ignored by the decoder.

Conclusion

This study provides a rigorous theoretical framework for understanding strategic communication in dynamic settings. While the Stackelberg model offers a path to fully revealing communication through commitment, the Nash model highlights the fundamental limitations imposed by misaligned incentives. Future work remains in identifying optimal non-linear policies for multi-dimensional signaling games and exploring the impact of noisy feedback channels.

Summary updated

Abstract

Simultaneous (Nash) and sequential (Stackelberg) equilibria of two-player dynamic quadratic cheap talk and signaling game problems are investigated under a perfect Bayesian formulation. For the dynamic scalar cheap talk, a zero-delay communication setup is considered for i.i.d. and Markov sources; it is shown that the final stage equilibrium is always quantized and under further restrictive conditions the equilibria for all time stages are quantized. Contrarily, the Stackelberg equilibria are always fully revealing for both scalar and multi-dimensional sources. In the dynamic signaling game where the transmission of a Gauss-Markov source over a memoryless Gaussian channel is considered, affine policies constitute an invariant subspace under best response maps for both scalar and multi-dimensional sources under Nash equilibria; whereas the Stackelberg equilibria always admit linear policies for scalar sources but such policies may be non-linear for multi-dimensional sources. A dynamic programming formulation is presented for multi-dimensional sources for optimal linear encoding policies, and conditions under which the Stackelberg equilibria are non-informative are derived.
BibTeX
@article{sartas2020dynamic,
  author = {Serkan Sarıtaş and Serdar Yüksel and Sinan Gezici},
  title = {{Dynamic Quadratic Cheap Talk and Signaling Games.}},
  journal = {Automatica},
  year = {2020},
  doi = {10.1016/j.automatica.2020.108883},
}
Figure