Sheehan Olver

> If the first one is real and the second one is complex, you can just use complex types That’s roughly 4x slower. Also that was just a model example....

I know. That’s why it’s a “?”

Just to be clear: the syntax would be `conv(u, v’, A) `, right? And it’s consistent because ` conv(u, v’) = u*v’ `

Ah it's because ArrayLayouts.jl defaults to QR factorization, since I believe it's more likely to be stable for structured matrices E.g. for a `BlockBandedMatrix` doing block LU may introduce instability....

Will merge a PR

I think empty blocks are probably useful… so would probably be best to figure out how to fix broadcasting

The problem with that is that we are treating `Array` as a block-array with one block. So whatever behaviour `vcat` and `hcat` have should be consistent.

But I don't think we want this, because we want consistency, that is, we want: ```julia a = 1:5 b = 10:12 blocksizes(vcat(a, b)) == blocksizes(vcat(BlockArray(a), BlockArray(b))) ``` The reason...

yes. There's no such thing as "no block"

unblocked == 1-blocked that's my point