On the rate-distortion-perception trade-off and remote coordination: the role of randomness
File(s)
Author(s)
Yassine, Hamdi
Type
Thesis or dissertation
Abstract
We collect, process, store and communicate ever increasing amounts of data. Lossy compression discards less relevant data under computational, latency, bandwidth or storage constraints. This thesis addresses fundamental problems in lossy compression, focusing on two areas: compression with realism constraints, and distributed coordination.
Classical rate-distortion (RD) theory characterizes the trade-off between compression level (rate) and difference from the original data (distortion), but fails to capture perceptual quality, or realism. Recent work has proposed formulations of a three-way rate-distortion-perception (RDP) trade-off; several of these require randomized algorithms, unlike classical RD. This randomness may be private to the encoder or the decoder, or common to both. First, we characterize the role of private randomness across RDP formulations, showing that encoder private randomness does not improve performance, even when conveyed to the decoder. Second, we analyze the trade-off under the strongest realism constraints in the presence of side information, characterizing its dual role as both a second observation of the source and a source of common randomness. Third, we identify important limitations of existing RDP formulations and propose an alternative framework, based on algorithmic information theory and involving critics of realism. We show that deterministic schemes are optimal when the number of samples a critic observes at a time is practical, and that our formulation is equivalent to previous ones when that number is impractically large.
The second part addresses coordination in distributed systems, where some agents must coordinate without direct communication while others communicate under latency or bandwidth constraints. Coordination requirements admit a mathematical formulation closely resembling the realism constraints above. We study non-interactive remote coordination, in which two agents receive information from two controllers that themselves coordinate without communicating. We develop remote channel synthesis, a technique extending classical channel synthesis, yielding a scheme with improved performance.
Classical rate-distortion (RD) theory characterizes the trade-off between compression level (rate) and difference from the original data (distortion), but fails to capture perceptual quality, or realism. Recent work has proposed formulations of a three-way rate-distortion-perception (RDP) trade-off; several of these require randomized algorithms, unlike classical RD. This randomness may be private to the encoder or the decoder, or common to both. First, we characterize the role of private randomness across RDP formulations, showing that encoder private randomness does not improve performance, even when conveyed to the decoder. Second, we analyze the trade-off under the strongest realism constraints in the presence of side information, characterizing its dual role as both a second observation of the source and a source of common randomness. Third, we identify important limitations of existing RDP formulations and propose an alternative framework, based on algorithmic information theory and involving critics of realism. We show that deterministic schemes are optimal when the number of samples a critic observes at a time is practical, and that our formulation is equivalent to previous ones when that number is impractically large.
The second part addresses coordination in distributed systems, where some agents must coordinate without direct communication while others communicate under latency or bandwidth constraints. Coordination requirements admit a mathematical formulation closely resembling the realism constraints above. We study non-interactive remote coordination, in which two agents receive information from two controllers that themselves coordinate without communicating. We develop remote channel synthesis, a technique extending classical channel synthesis, yielding a scheme with improved performance.
Version
Open Access
Date Issued
2025-11-04
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Gündüz, Deniz
Sponsor
European Commission
Grant Number
GA. No. 953775
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)