Adaptive Run-Length / Golomb-Rice Encoding of Quantized Generalized Gaussian Sources with Unknown Statistics

Basically looks at a short window of history from a Golomb-Rice process. If longer code words have become dominant in the output, adjust the K factor up to be more efficient with longer words. If short words are dominant, adjust the K factor back down where short words are efficient.

System is always less efficient (by 5-10% on average?) than an optimal encoder--however the encoder tunes itself without additional preprocessing burdens.

* * *

When the alphabet to be compressed follows a geometric distribution of symbols, Golomb codes are approximately equivalent to Huffman tables. But a decoding table does not have to be provided as it can be derived from a single integer tuning value.

Backwards adaption: when a rule only depends on previous outputs of the algorithm.

Supports run length encoding for repetitions of the same symbols.

Uses some heuristics to look at previously emitted symbols and adjust the K-factor for future Golomb-Rice codes.