Log-Likelihood Loss for Semantic Compression
Information Theory
2026-01-26 v1 math.IT
Abstract
We study lossy source coding under a distortion measure defined by the negative log-likelihood induced by a prescribed conditional distribution . This \emph{log-likelihood distortion} models compression settings in which the reconstruction is a semantic representation from which the source can be probabilistically generated, rather than a pointwise approximation. We formulate the corresponding rate-distortion problem and characterize fundamental properties of the resulting rate-distortion function, including its connections to lossy compression under log-loss, classical rate-distortion problems with arbitrary distortion measures, and rate-distortion with perfect perception.
Keywords
Cite
@article{arxiv.2601.16461,
title = {Log-Likelihood Loss for Semantic Compression},
author = {Anuj Kumar Yadav and Dan Song and Yanina Shkel and Ayfer Özgür},
journal= {arXiv preprint arXiv:2601.16461},
year = {2026}
}
Comments
18 pages, 4 figures