Output-Constrained Lossy Source Coding With Application to Rate-Distortion-Perception Theory

Li Xie, Liangyan Li, Jun Chen, Zhongshan Zhang

TL;DR

This paper bridges the gap between traditional lossy source coding and the emerging field of rate-distortion-perception (RDP) coding. It establishes theoretical limits for RDP coding, particularly for Gaussian sources, by leveraging insights from output-constrained source coding, offering a more nuanced understanding of how to balance compression rate, distortion, and perceptual quality.

What Problem Does This Paper Address?

Traditional lossy source coding aims to minimize distortion for a given compression rate. However, for applications like image and video compression, minimizing pixel-wise distortion doesn't always align with human perception of quality. Rate-Distortion-Perception (RDP) theory attempts to address this by incorporating a "perception" constraint alongside rate and distortion. While RDP theory has been proposed, its precise information-theoretic limits, especially concerning the role of common randomness between the encoder and decoder, remain an active area of research. This paper specifically tackles the challenge of characterizing these limits, particularly when the reconstructed data must adhere to a specific marginal distribution (output-constrained coding) and when common randomness is limited.

Methods

The core of this paper lies in connecting the information-theoretic limits of RDP coding to those of output-constrained lossy source coding. The authors achieve this by:

  • Formulating RDP as Output-Constrained Coding: They show that RDP coding with a marginal-based perception constraint is equivalent to output-constrained source coding where the reconstruction distribution is restricted to those that satisfy the perception measure.
  • Analyzing Gaussian Sources: The paper focuses on the analytically tractable case of Gaussian sources and squared error distortion. This allows for explicit characterizations of the rate-distortion-perception function.
  • Leveraging Perception Measures: They investigate two common perception measures: the Kullback-Leibler (KL) divergence and the squared quadratic Wasserstein distance. They derive bounds for the RDP function using these measures.
  • Deriving Single-Letter Characterizations: The authors establish single-letter characterizations for the distortion-rate function under output constraints, which are then used to derive bounds on the RDP function.
  • Investigating the Impact of Common Randomness: The paper explicitly analyzes how the availability of common randomness (quantified by $R_c$) affects the achievable distortion-rate-perception tradeoff.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis of source coding limits rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • RDP as a Constrained Optimization Problem: The paper elegantly frames RDP coding as a constrained optimization problem within the framework of output-constrained source coding. This provides a powerful analytical tool.
  • Gaussianity Simplifies Perception Tradeoffs: For Gaussian sources, the perception measures (KL divergence and Wasserstein distance) exhibit properties that simplify the analysis of the distortion-perception tradeoff, leading to explicit formulas.
  • Common Randomness Offers a "Free Lunch" (Up to a Point): The availability of common randomness ($R_c$) can significantly improve the distortion-rate-perception tradeoff, especially when perceptual quality is a concern. However, the benefit saturates, meaning a small amount of common randomness can be almost as effective as unlimited amounts.
  • Perception vs. Distortion Tradeoff: The paper demonstrates that for Gaussian sources, the perception constraint can significantly alter the distortion-rate tradeoff, particularly at lower rates. This highlights the importance of considering perceptual quality beyond simple distortion metrics.
  • Two Perception Measures Yield Similar Behavior (for Gaussian): For Gaussian sources, the bounds derived using KL divergence and Wasserstein distance show remarkable similarity, suggesting that these measures might capture similar aspects of perceptual quality in this specific context.

Key Results

  • The paper provides explicit characterizations of the rate-distortion-perception function for Gaussian sources under squared error distortion, using both KL divergence and squared quadratic Wasserstein distance as perception measures.
  • It establishes upper and lower bounds on the RDP tradeoff, showing how these bounds coincide for Gaussian sources.
  • The analysis quantifies the impact of common randomness ($R_c$) on the RDP tradeoff, illustrating that its benefit diminishes as $R_c$ increases.
  • The paper demonstrates that the perception constraint can significantly affect the distortion-rate tradeoff, especially when the rate is low.

Limitations and Future Work

  • Focus on Gaussian Sources: The explicit characterizations are primarily for Gaussian sources. Extending these results to more general source distributions remains an open challenge.
  • Marginal vs. Joint Perception: The paper focuses on marginal-distribution-based perception constraints. Investigating the impact of joint-distribution-based perception constraints, which are more general, is suggested for future work.
  • Tightness of Bounds: While bounds are established, further investigation into their tightness for non-Gaussian sources or more complex scenarios is encouraged.
  • Practical Implementation: The paper focuses on theoretical limits. Bridging the gap to practical implementations of RDP coding with limited common randomness is a natural next step.

Summary updated

Abstract

The distortion-rate function of output-constrained lossy source coding with limited common randomness is analyzed for the special case of the squared error distortion measure. An explicit expression is obtained when both the source and reconstruction distributions are Gaussian. This further leads to a partial characterization of the information-theoretic limit of quadratic Gaussian rate-distortion-perception coding, with the perception measure given by either the Kullback-Leibler divergence or the squared quadratic Wasserstein distance, from which Wagner’s result for the perfect realism setting and Zhang et al.’s result for the unlimited common randomness setting can be recovered as special cases.
BibTeX
@article{xie2025output,
  author = {Li Xie and Liangyan Li and Jun Chen and Zhongshan Zhang},
  title = {{Output-Constrained Lossy Source Coding With Application to Rate-Distortion-Perception Theory}},
  journal = {IEEE Trans. Communications},
  year = {2025},
  doi = {10.1109/tcomm.2024.3462688},
}
Figure