On Nonlinear Compression Costs: When Shannon Meets Rényi

In compression problems, the minimum average codeword length is achieved by Shannon entropy, and efficient coding schemes such as Arithmetic Coding (AC) achieve optimal compression. In contrast, when minimizing the exponential average length, Rényi entropy emerges as a compression lower bound. This paper presents a novel approach that extends and applies the AC model to achieve results that are arbitrarily close to Rényi’s lower bound. While rooted in the theoretical framework assuming independent and identically distributed symbols, the empirical testing of this generalized AC model on a Wikipedia dataset with correlated symbols reveals significant performance enhancements over its classical counterpart, when considering the exponential average. The paper also demonstrates an intriguing equivalence between minimizing the exponential average and minimizing the likelihood of exceeding a predetermined threshold in codewords’ length. An extensive experimental comparison between generalized and classical AC unveils a remarkable reduction, by several orders of magnitude, in the fraction of codewords surpassing the specified threshold in the Wikipedia dataset.