This question was initially posted a few weeks back as an issue in LinearSolve.jl but got no attention there. By reposting it here I’m hoping to generate a bit of discussion about this.
As far as I understand the purpose of LinearSolve, it is an attemp at providing downstream users and developpers a unified interface between multiple solvers.
As highlighted by the documentation, matrix factorizations play a significant role in the different solvers. However, it does not seem that a unified interface for these factorizations is provided by LinearSolve.
I’m just thinking about this because I’m myself in the need of finding the best QR factorization (to retrieve the R term only) for one of my projects and LinearSolve seemed a nice place to start from. I thought I would find something like
fact_problem = FactorizationProblem(<Matrix>,<FactorizationMethod>)
fact = solve(fact_problem)
Would such an approach be relevant for LinearSolve or is it useless ? I can’t really make my mind… it seems completely overkill, yet achieves somewhat the same pupose as LinearProblem.