THE FOLD / CHEAT / GOD MODE / THE SHANNON LIMIT
THE SHANNON LIMIT
error-free, through noise, at a rate that does not vanish
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Before 1948 the assumption was obvious and wrong: to make a noisy channel more reliable, you must send more slowly. Repeat each bit three times, five times, nine times — the errors fall, and so does your rate, toward zero. Reliability looked like something you bought with throughput. Shannon proved that below a specific rate, the channel capacity C, you can make the error probability as small as you like without the rate going to zero. Above C, you cannot, at any price. The wall is sharp and it sits at C = 1 − H(p) for a binary symmetric channel.
LIT verified live: capacity is computed across the range — p = 0.01 gives 0.919207, p = 0.1 gives 0.531004, p = 0.25 gives 0.188722, and p = 0.5 gives exactly 0; capacity is symmetric about p = ½; repetition coding at p = 0.1 is measured exactly by the binomial and drives error from 1.0e-1 down to 6.9e-9 while rate collapses from 1.000 to 0.032; and the counting behind the theorem is checked directly — at n = 200, p = 0.1 the typical set holds 293.8 of 2200 sequences, a fraction of 1.1e-32, and that emptiness is the room a code hides in.
LIT verified live: capacity is computed across the range — p = 0.01 gives 0.919207, p = 0.1 gives 0.531004, p = 0.25 gives 0.188722, and p = 0.5 gives exactly 0; capacity is symmetric about p = ½; repetition coding at p = 0.1 is measured exactly by the binomial and drives error from 1.0e-1 down to 6.9e-9 while rate collapses from 1.000 to 0.032; and the counting behind the theorem is checked directly — at n = 200, p = 0.1 the typical set holds 293.8 of 2200 sequences, a fraction of 1.1e-32, and that emptiness is the room a code hides in.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this at GOD MODE, and it earns the seat more literally than most. Arbitrarily reliable communication across an unreliable channel, at a rate that does not vanish, reads exactly like a cheat code — the noise is still there, every symbol still gets corrupted at rate p, and the message still arrives perfect.
AVAN (AI) wrote a false claim into the first draft and the test caught it. The assertion was that every repetition-coding rate sits below capacity “as it must”. It failed immediately: the n = 1 point has rate 1.000, and capacity at p = 0.1 is 0.531004. Rate 1 is above C. That is not a flaw in the experiment, it is the entire lesson — uncoded transmission is above capacity, which is precisely why its error sticks at 0.1 and cannot be driven down. The page now states it that way. One further boundary, stated plainly: this page does not prove the coding theorem. It computes capacity, measures a real code against it, and verifies the typical-set counting that makes the theorem plausible. The achievability proof itself is Shannon’s, cited and not reproduced here.
AVAN (AI) wrote a false claim into the first draft and the test caught it. The assertion was that every repetition-coding rate sits below capacity “as it must”. It failed immediately: the n = 1 point has rate 1.000, and capacity at p = 0.1 is 0.531004. Rate 1 is above C. That is not a flaw in the experiment, it is the entire lesson — uncoded transmission is above capacity, which is precisely why its error sticks at 0.1 and cannot be driven down. The page now states it that way. One further boundary, stated plainly: this page does not prove the coding theorem. It computes capacity, measures a real code against it, and verifies the typical-set counting that makes the theorem plausible. The achievability proof itself is Shannon’s, cited and not reproduced here.
3 ONE DIMENSION
C = 1 − H(p). Perfect at p = 0, zero at p = ½, and perfect again at p = 1.
4 TWO DIMENSIONS · INTERACTIVE
Move the noise and watch repetition coding buy reliability with rate it cannot afford.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the typical set as a thin shell inside an enormous cube.
AVAN’s addition (the inverse-companion): the forward reading is “the channel has a capacity.” The inverse is that the theorem is not really about channels — it is about how empty high-dimensional spaces are. Almost every long binary sequence is atypical and will simply never occur; the ones that do occur cluster in a vanishing shell, 1.1e-32 of the whole at n = 200. A code works because it can place its words in that overwhelming emptiness far enough apart that noise cannot bridge them. Read backwards, capacity is a measurement of available room, and the surprise is not that reliable communication is possible but that we ever imagined the space was full.
LIT capacity 1-H(p) computed across the range: p=0.01 gives 0.919207, p=0.1 gives 0.531004, p=0.25 gives 0.188722, p=0.5 gives exactly 0, symmetric about one half; repetition coding at p=0.1 measured exactly by the binomial drives error from 1.0e-1 to 6.9e-9 while rate collapses 1.000 to 0.032; at n=200, p=0.1 the typical set holds 2^93.8 of 2^200 sequences, a fraction of 1.1e-32
FIG A first draft asserted every repetition rate sits below capacity and failed its own test: the uncoded n=1 point has rate 1.000 against a capacity of 0.531004. That is the lesson rather than a bug — uncoded transmission is above capacity, which is why its error sticks at 0.1. This page does not prove the coding theorem; it computes capacity, measures a real code against it, and verifies the typical-set counting. Achievability is Shannon's, cited not reproduced.
FIG A first draft asserted every repetition rate sits below capacity and failed its own test: the uncoded n=1 point has rate 1.000 against a capacity of 0.531004. That is the lesson rather than a bug — uncoded transmission is above capacity, which is why its error sticks at 0.1. This page does not prove the coding theorem; it computes capacity, measures a real code against it, and verifies the typical-set counting. Achievability is Shannon's, cited not reproduced.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of GOD MODE · David Lee Wise (ROOT0), with AVAN