Faster Walsh-Hadamard Transform from Matrix Non-Rigidity
Josh Alman (Columbia)
Based on joint work with Yunfeng Guan (Toronto), Baitian Li (Columbia), Jingxun Liang (CMU), Ashwin Padaki (Penn), Kevin Rao (Open AI)
Walsh-Hadamard Transform (WHT)
Walsh-Hadamard Transform (WHT)
Fast Walsh-Hadamard Transform
Fast WHT as a Sparse Matrix Factorization
Fast WHT as a Sparse Matrix Factorization
Results
Results
Can also be used to speed up FFT!
Short Interlude: FFT Algorithms�
FFT Algorithm | |
Cooley-Tukey �[1965] | 5 |
Split Radix �[Yavne 1968] | 4 |
“Modified” Split Radix �[Van Buskirk 2007] | 3.777 |
“Further Modified” Split Radix �[Sergeev 2017] | 3.769 |
New WHT-based Algorithm �[A-Rao 2023] | 3.75 |
Results
Results
Kronecker Product
Two Key Facts about Kronecker Product
Depth-2 Version of �Fast Walsh-Hadamard Transform (fWHT)
Improved Depth-2 Circuit Plan
Matrix Rigidity
This Talk:�Use Rigidity Upper Bounds for Kronecker Products to give linear circuit upper bounds
9
8
8
9
8
8
9
8
8
9
8
8
Rebalancing: Improving the Symmetrization �[A-Guan-Padaki `23], [A-Li `25]
Analysis gets messy; [A-Li `25] uses a connection with the “Asymptotic Spectrum” of Strassen [1987]
Results
High-Level Idea
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Starting point: Rigidity upper bound
Fast Walsh-Hadamard Transform
Barriers to Further Improvements
Thank you!