Optimal Signaling with Mismatch in Priors of an Encoder and Decoder.

Ertan Kazıklı, Serkan Sarıtaş, Sinan Gezici, Serdar Yüksel

Abstract

We consider communications between an encoder and a decoder, viewed as two decision makers, which have subjective beliefs on the probabilistic model of the source distribution. Even though the decision makers employ the same cost function, induced expected costs are different from the perspective of the encoder and decoder due to their subjective probabilistic beliefs, which requires a game theoretic treatment. Depending on the commitment nature of the encoder to its policies, we analyze this signaling game problem under Nash and Stackelberg equilibrium concepts. In particular, we consider a communication scenario through a Gaussian noise channel with a power constrained encoder. We show that the Stackelberg equilibrium cost of the encoder is upper semi-continuous, under the Wasserstein metric, as encoder's prior approaches to decoder's prior and in the particular case of Gaussian subjective priors it is also lower semi-continuous, which proves the robustness of the equilibrium around the team setup. We further prove that the optimality of affine policies for Gaussian signaling under Stackelberg equilibria breaks down due to the presence of prior mismatch. We also investigate the informativeness of Stackelberg equilibria under affine policy restriction when there is prior mismatch and show that under certain conditions the equilibria become non-informative, that is information transmission ceases to exist. For the Nash setup, we provide necessary and sufficient conditions under which there exist informative affine Nash equilibria. Furthermore, we show that there exist fully informative Nash and Stackelberg equilibria for the cheap talk problem (i.e., no additive noise term and no power constraint at the encoder) as in the team theoretic setup under an absolute continuity condition.

Introduction

In classical communication theory, the encoder and decoder are typically assumed to be perfectly aligned, sharing an identical probabilistic model of the source. However, in many practical scenarios, these agents may have subjective beliefs—or "prior mismatch"—regarding the source distribution. This paper explores the strategic implications of such disagreements, treating the communication process as a signaling game between two decision-makers with differing perspectives on the underlying probability space.

Problem Formulation and Equilibrium Concepts

The study focuses on two primary communication models: the signaling game, which includes additive noise and power constraints, and the "cheap talk" scenario, where communication is noiseless and unconstrained.

The basic signaling game and cheap talk models used in the paper.

To analyze these interactions, the authors employ two equilibrium concepts:

  1. Stackelberg Equilibrium: The encoder (leader) commits to a signaling policy, and the decoder (follower) optimally responds. This is common in scenarios where the encoder's strategy is transparent.
  2. Nash Equilibrium: Both parties choose their strategies simultaneously, representing a non-cooperative environment where neither can commit beforehand.

Robustness and the Wasserstein Metric

A central question is whether the system's performance is robust to small discrepancies in priors. By employing the Wasserstein metric to measure the distance between the encoder's and decoder's subjective distributions, the authors prove that the Stackelberg equilibrium cost is upper semi-continuous. In the specific case of Gaussian priors, the cost is fully continuous, ensuring that small model perturbations do not lead to catastrophic performance loss.

A suboptimal scheme involving a coupling channel used to prove the upper semi-continuity of the encoder's cost.

The Breakdown of Affine Optimality

In standard team-theoretic Gaussian communication, affine (linear) policies are known to be optimal. However, this paper demonstrates that prior mismatch breaks this optimality. Under a Stackelberg setup with mismatched Gaussian priors, non-linear strategies—such as quantization or "binning"—can outperform the best possible affine policy. This discovery challenges the conventional reliance on linear models in mismatched environments.

Informativeness and "Babbling" Equilibria

The authors also investigate the conditions under which meaningful information transmission occurs. They identify thresholds where the prior mismatch or signaling costs become so high that the only stable equilibrium is a "non-informative" (or babbling) one, where the encoder's signal carries no useful information for the decoder. Conversely, for cheap talk, fully informative communication remains possible as long as the subjective priors are mutually absolutely continuous.

Conclusion

This research highlights the critical impact of subjective beliefs on optimal signaling. While communication systems exhibit a degree of robustness to prior mismatch, the loss of affine optimality necessitates more sophisticated, non-linear strategic approaches. Future work may extend these findings to non-quadratic costs and dynamic, multi-stage signaling games.

Summary updated

Abstract

We consider communications between an encoder and a decoder, viewed as two decision makers, which have subjective beliefs on the probabilistic model of the source distribution. Even though the decision makers employ the same cost function, induced expected costs are different from the perspective of the encoder and decoder due to their subjective probabilistic beliefs, which requires a game theoretic treatment. Depending on the commitment nature of the encoder to its policies, we analyze this signaling game problem under Nash and Stackelberg equilibrium concepts. In particular, we consider a communication scenario through a Gaussian noise channel with a power constrained encoder. We show that the Stackelberg equilibrium cost of the encoder is upper semi continuous, under the Wasserstein metric, as encoder's prior approaches to decoder's prior and in the particular case of Gaussian subjective priors it is also lower semi continuous, which proves the robustness of the equilibrium around the team setup. We further prove that the optimality of affine policies for Gaussian signaling under Stackelberg equilibria breaks down due to the presence of prior mismatch. We also investigate the informativeness of Stackelberg equilibria under affine policy restriction when there is prior mismatch and show that under certain conditions the equilibria become non-informative, that is information transmission ceases to exist. For the Nash setup, we provide necessary and sufficient conditions under which there exist informative affine Nash equilibria. Furthermore, we show that there exist fully informative Nash and Stackelberg equilibria for the cheap talk problem (i.e., no additive noise term and no power constraint at the encoder) as in the team theoretic setup under an absolute continuity condition.
BibTeX
@article{kazkl2022optimal,
  author = {Ertan Kazıklı and Serkan Sarıtaş and Sinan Gezici and Serdar Yüksel},
  title = {{Optimal Signaling with Mismatch in Priors of an Encoder and Decoder.}},
  journal = {IEEE Transactions on Automatic Control},
  year = {2022},
  doi = {10.1109/tac.2022.3142121},
}
Figure