paq8px v206 [1] with option "-8art" compresses it to 9,145 bytes using ~2GB of memory in 82s, mostly due to a better pre-training stage.
Hmm. My own entropy estimating code suggests a minimum entropy of 17443 bits with a 63 bit word size (I would need to change how I do things to search above 64 bits, which I would like to do and probably will one day). I used this list:
https://gist.github.com/cfreshman/a03ef2cba789d8cf00c08f767e...
Here is the code:
X = fread(fid, sprintf('*ubit%d', nBits));
M = numel(X);
words = double(unique(X))';
binEdges = [words-0.5; words+0.5];
binEdges = unique(binEdges(:));
P = histcounts(X, binEdges) / M;
P(P == 0) = [];
H = -sum(P .* log(P) ./ log(max([nBits, 2]))) / nBits * log2(max([nBits, 2]));
E = H * M * nBits;
This is a textbook entropy implementation and I've been under the impression that you'll never do better than this (for a given word size).
Re:Dictionary Compression
by lifthrasiir on Tuesday March 01, 2022 @09:03PM
(#30522764)
You calculated the sum of entropys of each word, not the entropy of the entire word list as a whole (which would be much lower due to the cross-correlation).
[1] https://github.com/byronknoll/cmix
[2] http://golf.horse/wordle/list