Dice Question Streamline Icon: https://streamlinehq.com

Achieving full sharing of level-ℓ blocks across all dimensions in laser-method analyses

Develop a laser-method-based procedure for powers of the Coppersmith–Winograd tensor CW_q that allows different remaining level-ℓ subtensors to share level-ℓ variable blocks in all three dimensions (X, Y, Z), while requiring only that each level-1 variable block belongs to a unique remaining level-ℓ subtensor.

Information Square Streamline Icon: https://streamlinehq.com

Background

Recent improvements (Duan–Wu–Zhou; Vassilevska Williams–Xu–Xu–Zhou) introduced asymmetry to mitigate combination loss in recursive laser-method analyses, first permitting sharing in Z and later finer-grained sharing. The authors further advance this by enabling sharing in Y and Z while retaining uniqueness of level-ℓ X-blocks.

They state the natural end goal of the line of work: sharing level-ℓ blocks in all three dimensions with only level-1 uniqueness constraints. They explicitly note that achieving this goal is presently unclear, highlighting a central methodological open direction.

References

Following from the sequence of improvements obtained by [duan2023,VXXZ24], the natural end goal for this line of work is to allow different remaining subtensors to share level-ℓ variable blocks in all three dimensions, and only require level-1 variable blocks to belong to unique remaining level-ℓ subtensors. However, it is unclear how one may achieve this goal since the algorithms of both [duan2023] and [VXXZ24] crucially use the fact that different remaining level-ℓ subtensors use unique level-ℓ variable blocks in the X- and Y-dimension.

More Asymmetry Yields Faster Matrix Multiplication (2404.16349 - Alman et al., 25 Apr 2024) in Section 2.4 (Technical Overview: Our Contributions: More Asymmetry)