Generally, low-rank and block-diagonal matrices are both great strategies for producing expressive matmuls with fewer parameters. We can view the FFT as a particularly deft example of factorizing one big matmul into a number of block-diagonal matmuls, greatly reducing the overall number of multiplications by minimizing the block size. However, on a G/TPU, we have a lot more parallelism available, so the sweet spot for size of the blocks may be larger than 2x2...
We can also mix low-rank, block diagonal, and residual connections to get the best of both worlds:
x' = (L@x + B@x + x)
The block-diagonal matrix does 'local' work, and the low-rank matrix does 'broadcast' work. I find it pretty typical to be able to replace a single dense matmul with this kind of structure and save ~90% of the params with no quality cost... (and sometimes the regularization actually helps!)