Did you consider structuring Walrus as a plain arithmetic coder, where the model knows three symbols? Namely, A: corresponding to symbol A, B: corresponding to symbol B, and F: corresponding to the event of flushing bits down the bit stream When the decoder sees F, it has the effect of advancing its bit buffer, but of course the symbol F is omitted from the output. I think the probability of F can be set at each point such that it can always be decoded with the bits available to the decoder. On 11/20/20 20:34, Steve Witham wrote:

"ABSTRACT: The Walrus coder does the same job as a binary arithmetic coder-- "binary" meaning having an input alphabet of just two symbols. The Walrus has the same API as arithmetic coders, and similar coding efficiency and latency, but the Walrus maps symbols to bits differently: at each step it designates one of the two symbols "walrus" and the other "eggman." The walrus is assigned an available bit string, which is then either output (if walrus is sent) or made unavailable (if eggman is sent). ... This file is "literate executable pseudocode" in Python... An appendix contains test code... A separate Jupyter notebook [N] does some analysis [and graphs] of the Walrus's efficiency both empirically and theoretically."

http://www.mac-guyver.com/switham/2020/11/Walrus

 --Steve

_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun