The model is part of the message
The Gettysburg Address is 1,458 characters, which is 11,664 bits
if you send it as bytes. Count the letters and it is 6,008. Look at the previous
letter too and it is 4,360. Look at the previous four and it is 509 — which
sounds like a triumph, and is a model that has memorised the text. To decode those 509
bits you would need the table that produced them, and that table is 11,466 bits.
Nothing was compressed. Everything was moved.
What the letters cost
Thirty-five distinct characters, so five bits each would do it. But they are not equally likely — the space appears 270 times and half the alphabet under ten — and a code that gives the common ones short strings does better. The best possible average is the entropy; Huffman gets within a twentieth of a bit of it, and cannot do worse than one.
Look at the previous letter
A q is followed by a u, a space is rarely followed by a space. Counting pairs rather than letters cuts the address by a quarter. The control is the same letters in a shuffled order: it has exactly the same letter frequencies, so exactly the same first number — and none of the second. The gap between the two is the structure.
·
Keep going and it falls to nothing
Each extra character of context divides the text into more contexts, and each context is seen fewer times, until every one of them has been seen once and predicts its single successor perfectly. The message goes to zero. The table needed to decode it does not.
·
What this page checks about itself
The entropy, the code, the Kraft sum and every bit count are computed here from the text above, which is in the page. Nothing below is quoted from anywhere.