Matrix Multiplication: Coppersmith-Winograd and Beyond
summary
Describes a concrete route from tensors to matrix multiplication exponent: use border rank q+2 bounds on a specific Coppersmith‑Winograd family, symmetrize and take tensor powers, then apply blocking and zeroing‑out combinatorics to extract many small matrix‑products and lower‑bound a value function via multinomial counts and Stirling approximations. That pipeline produces numerical exponents (CW square gives 2.376, higher powers inch down toward the current ~2.373) and also exposes a provable barrier: the standard merging/zeroing framework cannot push omega to 2 without new tensors or restrictions.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:49Tensors and matrix multiplication
- 5:52Border Rank
- 8:34Restrictions of tensors
- 10:48Zeroing out - a special restriction
- 14:39Value of a tensor
- 23:20The CW tensor
- 25:38Blocking of a tensor
- 31:16Zeroing out blocks
- 33:18Fractions for block triples
- 38:36When the linear system doesn't have full rank
- 41:36Example
- 45:07Blockings and tensor powers
- 49:24Computational issues
- 50:05Summary
- 51:00Questions?
- 51:35Limitations
