On distributionally robust lossy source coding

This talk introduces a distributionally robust framework for lossy source coding under model uncertainty. Classical rate-distortion theory assumes known source statistics, which is often unrealistic. We develop a robust extension incorporating worst-case guarantees over ambiguity sets defined by KL-balls. Key results include robust versions of the Strong Functional Representation Lemma and new one-shot and asymptotic achievability theorems. The theory naturally extends to perceptual distortion metrics, yielding a robust rate-distortion-perception function. We illustrate analytical results for Bernoulli sources with Hamming distance and discuss optimization formulations, duality, and future research directions.