THE FOLD / RESPAWN / THE RESURRECT / THE GILBERT-VARSHAMOV
THE GILBERT-VARSHAMOV
the code is there; nobody has to find it
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
An error-correcting code is a set of binary words kept far enough apart that noise cannot carry one into another. The question is how many such words fit. Two bounds answer it from opposite sides. The Hamming bound says: draw a ball of radius t around each codeword, the balls cannot overlap, so you cannot have more codewords than balls fit — an upper limit. The Gilbert–Varshamov bound argues the other way, and it is the stranger argument: keep greedily picking any word at distance ≥ d from everything chosen; you can only be stopped when the balls of radius d−1 cover the whole space; therefore a code of size at least 2n/|ball(n, d−1)| must exist. It proves the code is there without ever exhibiting one.
LIT verified live: the two bounds bracket the truth in every case tested — (8,3): 7 ≤ A ≤ 28, (10,3): 19 ≤ A ≤ 93, (12,5): 6 ≤ A ≤ 51, (15,3): 271 ≤ A ≤ 2048, (16,5): 27 ≤ A ≤ 478 — and GV never exceeds Hamming; a greedy construction actually run here meets or beats GV every time, building 16 words for (8,3) against a guarantee of 7, and 64 for (10,3) against 19; and the built (10,3) code is re-checked pairwise, confirming minimum distance 3 across all 2016 pairs.
LIT verified live: the two bounds bracket the truth in every case tested — (8,3): 7 ≤ A ≤ 28, (10,3): 19 ≤ A ≤ 93, (12,5): 6 ≤ A ≤ 51, (15,3): 271 ≤ A ≤ 2048, (16,5): 27 ≤ A ≤ 478 — and GV never exceeds Hamming; a greedy construction actually run here meets or beats GV every time, building 16 words for (8,3) against a guarantee of 7, and 64 for (10,3) against 19; and the built (10,3) code is re-checked pairwise, confirming minimum distance 3 across all 2016 pairs.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this at THE RESURRECT. A code is a promise that a corrupted message can be brought back — not patched, not approximated, but returned to exactly what was sent. The distance is the size of the wound it can survive.
AVAN (AI) wants to be plain about what the greedy result does and does not show. Greedy building 64 words where GV guarantees 19 is not evidence that GV is weak. GV is a worst-case existence floor derived by counting alone; any actual construction should beat it, and the interesting fact is the opposite one — that for thirty years nothing beat GV asymptotically, for any family of codes, by any method. It stood as the best known lower bound on the achievable rate until Tsfasman, Vlăduţ and Zink got past it in 1982 using algebraic geometry over function fields, and only for alphabets of size 49 and above. That result is cited here, not verified here: nothing on this page tests it. What this page tests is the counting argument itself, at small n, where it can be checked exhaustively.
AVAN (AI) wants to be plain about what the greedy result does and does not show. Greedy building 64 words where GV guarantees 19 is not evidence that GV is weak. GV is a worst-case existence floor derived by counting alone; any actual construction should beat it, and the interesting fact is the opposite one — that for thirty years nothing beat GV asymptotically, for any family of codes, by any method. It stood as the best known lower bound on the achievable rate until Tsfasman, Vlăduţ and Zink got past it in 1982 using algebraic geometry over function fields, and only for alphabets of size 49 and above. That result is cited here, not verified here: nothing on this page tests it. What this page tests is the counting argument itself, at small n, where it can be checked exhaustively.
3 ONE DIMENSION
The bracket. GV guarantees the floor; Hamming forbids the ceiling; the truth lives between.
4 TWO DIMENSIONS · INTERACTIVE
Run the greedy construction and watch a real code assemble.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the 4-cube with a distance-2 code lit up inside it.
AVAN’s addition (the inverse-companion): the forward reading is “a good code exists.” The inverse is that GV is really a statement about covering, not packing — the greedy process can only halt when the balls of radius d−1 have covered every point in the space. So the same inequality read the other way is a covering bound, and the existence of a good error-correcting code is the shadow of the impossibility of an efficient covering. Nothing is constructed in either direction. The argument works entirely by making a construction impossible to stop early, which is a different kind of proof from a recipe — and it is why the bound is easy to state, easy to verify, and was very nearly impossible to beat.
LIT the GV and Hamming bounds bracket the truth in every case tested, (8,3): 7 <= A <= 28 through (16,5): 27 <= A <= 478, with GV never exceeding Hamming; a greedy construction run here meets or beats GV every time, building 16 words for (8,3) against a guarantee of 7 and 64 for (10,3) against 19; the built (10,3) code is re-checked pairwise at minimum distance 3 across all 2016 pairs
FIG Greedy beating GV by a wide margin is expected and is not evidence the bound is weak — GV is a worst-case existence floor derived by counting alone. The genuinely remarkable fact, that nothing beat GV asymptotically until Tsfasman, Vladut and Zink in 1982 using algebraic geometry over function fields, is cited and NOT verified here; nothing on the page tests it.
FIG Greedy beating GV by a wide margin is expected and is not evidence the bound is weak — GV is a worst-case existence floor derived by counting alone. The genuinely remarkable fact, that nothing beat GV asymptotically until Tsfasman, Vladut and Zink in 1982 using algebraic geometry over function fields, is cited and NOT verified here; nothing on the page tests it.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE RESURRECT · David Lee Wise (ROOT0), with AVAN