Closed dlfivefifty closed 6 years ago
No objections. Probably the redesign would involve separating the hierarchical factorizations and solves from the encapsulation of such representations as coefficients of functions.
There’s two tracks to pursue: (1) the one we originally went with and Alex is pursuing, where linear algebra works with general types that may be Fun or Array, (2) Fun is a wrapper and all linear algebra happens with AbstractArrays. I started thinking (1) is better, but the nature of implementing algorithms and the efficiency of BLAS Level 3 has lead me to the latter: first step is reduce everything to linear algebra with (possibly infinite) arrays,
Done!
The Hierarchical stuff seems a bit crusty, and also looks like it is fairly self-contained. I propose moving it to a separate package, say HierarchicalSingularIntegralEquations.jl.
This would be left unregistered, and unmaintained (by me) for the time being.
There's some questions on whether the design needs to be completely redone anyways: I've come around to the idea that
LowRankFun
should be replaced by aFun
with a low rank matrix of coefficients. That is: everything in ApproxFun should be wrappers around linear algebra. Not sure how hierarchical GreenFun's fit in to this view, perhaps the coefficients would be aVector{<:HierarchicalMatrix}
.@MikaelSlevinsky Any objections with moving this out? I'm not sure you are using it at the moment.