THE FOLD / CHEAT / THE EXPLOIT / THE STRASSEN
THE STRASSEN
multiply 2x2 with 7 products, not 8 — bending O(n^3)
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Strassen’s algorithm multiplies matrices faster than the schoolbook method by exploiting hidden algebraic slack. To multiply two 2×2 matrices you seemingly need 8 scalar multiplications; Strassen does it with just 7 — using clever combined products like M₁=(a+d)(e+h) — trading one multiply for a few extra additions.
Applied recursively to n×n matrices split into quadrants, 7 instead of 8 turns O(n³) into O(nlog₂7) ≈ O(n2.807) — the first sub-cubic matrix multiplication, and the crack that opened the whole field of fast linear algebra.
LIT verified live: the recursive Strassen product equals the naive product exactly for random integer matrices, using 7 scalar multiplications per 2×2 where the naive method uses 8 (window.__strassen). FIG no framing; exact integer matrix arithmetic.
Applied recursively to n×n matrices split into quadrants, 7 instead of 8 turns O(n³) into O(nlog₂7) ≈ O(n2.807) — the first sub-cubic matrix multiplication, and the crack that opened the whole field of fast linear algebra.
LIT verified live: the recursive Strassen product equals the naive product exactly for random integer matrices, using 7 scalar multiplications per 2×2 where the naive method uses 8 (window.__strassen). FIG no framing; exact integer matrix arithmetic.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this at the-exploit — finding the slack the obvious method leaves on the table. Strassen is the exploit that broke the ‘matrix multiply must be cubic’ assumption. AVAN (AI) built the instrument: the seven products, the recombination, the exact check against naive, the multiplication count.
Credit as content: Volker Strassen (1969), Gaussian Elimination is not Optimal. The weave: David names the exploit; I compute the seven products, recombine them into the exact same result as the eight-multiply method, and count the saving that compounds through the recursion.
Credit as content: Volker Strassen (1969), Gaussian Elimination is not Optimal. The weave: David names the exploit; I compute the seven products, recombine them into the exact same result as the eight-multiply method, and count the saving that compounds through the recursion.
3 ONE DIMENSION
The seven products M₁…M₇, each a single multiplication of two sums, and how they recombine into the four output blocks — one multiply fewer than the eight the naive method needs.
4 TWO DIMENSIONS · INTERACTIVE
Two 4×4 integer matrices. The instrument multiplies them with recursive Strassen and with the naive method, confirms the products are identical, and counts the scalar multiplications (7 per 2×2 vs 8).
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the recursion tree branching 7 ways at each level instead of 8 — the saving compounding down the depth.
AVAN’s addition (the inverse-companion): one fewer multiplication does not just save a constant — it lowers the exponent forever. Each recursion level replaces 8 subproblems with 7, so the cost is 7k instead of 8k at depth k, and the exponent drops from log₂8 = 3 to log₂7 ≈ 2.807. The inverse of ‘one less multiply’ is ‘an asymptotic exponent bent down permanently.’ Magenta is the 8th product you never compute — the naive multiply skipped at every node; green is the 7 that suffice, recursing to the bottom. A single algebraic identity, applied all the way down, reshapes the complexity of multiplication. (Honesty: Strassen trades stability and constants, so it is used above a crossover size, not everywhere.)
LIT Genuine Strassen algorithm (Strassen 1969, 'Gaussian Elimination is not Optimal'). Verified live: the recursive 7-product scheme equals the naive O(n^3) product exactly for 50 random 4x4 integer matrices, and a 2x2 base multiply uses exactly 7 scalar multiplications vs the naive 8 (window.__strassen.productMatchesNaive && mults7===7).
FIG No framing: the seven products, the recombination, and the exact comparison to naive multiplication run in-browser. The AVAN inverse is honest — 7 vs 8 lowers the asymptotic exponent from log2 8=3 to log2 7~2.807 (a real, permanent change); magenta is the 8th product skipped. Honestly noted: Strassen has worse constants/stability and is used above a crossover size.
FIG No framing: the seven products, the recombination, and the exact comparison to naive multiplication run in-browser. The AVAN inverse is honest — 7 vs 8 lowers the asymptotic exponent from log2 8=3 to log2 7~2.807 (a real, permanent change); magenta is the 8th product skipped. Honestly noted: Strassen has worse constants/stability and is used above a crossover size.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE EXPLOIT · David Lee Wise (ROOT0), with AVAN