Publications scientifiques
Stéphane Gaël R. Ekodeck
Ebele, Serge, Alain
Chantal Marguerite Mveh-Abia
Hervé Talé Kalachi
Norbert Djong Wang
HAL (Le Centre pour la Communication Scientifique Directe)
We study complete-stream accounting for static additive block formats through FISA 1.0 (Fibonacci Index-gap Shared Alphabet), a byte-exact test format that amortises one gap alphabet over a stream of blocks. The accounting criterion keeps stream-level framing and model bytes, per-block records, and input size in one auditable inequality; we use it as a measurement discipline rather than as a new coding theorem. FISA, a zero-block bitmap, and a boundary-trim format are evaluated as fully serialized, exactly decodable instances. A retrospective offline portfolio bound reports the smallest measured self-contained stream among these formats and raw storage, yielding ratios of 0.8936 on Canterbury and 0.9774 on Silesia. The FISA transform is proved bijective; a compiled implementation verifies decode round-trips on Canterbury, Silesia, and the 100 MB enwik8 benchmark, and the supplied test suite passes 10,119 cases. Corpus experiments show that 256-byte blocks average about 815.63 Zeckendorf terms, well above the measured favorable range of 326-392 terms: only 3.70% and 1.94% of blocks satisfy the sufficient local gain condition. A controlled Fibonacci-sparsity threshold experiment provides an internal consistency check on that cost model. For streams whose blocks contain at most 200 Zeckendorf terms, FISA reaches ratios of 0.04-0.62 and is smaller than Zstandard level 19, the tested zero-structure controls, and a new complete binary position-gap baseline. Structural controls on standard corpora attribute the observed savings primarily to zero structure rather than to the Fibonacci basis. The resulting contribution is therefore a reproducible account of where an apparently compact positional transform does and does not survive complete serialization.
