LinearAlgebra: move alg to a keyword argument in symmetric eigen #55481
Conversation
|
Losing the ability to specialize on a custom alg type seems like a pretty big loss. How would one write EigenAlgorithms.jl? FWIW, specialization on algorithms in sorting using the public API is a mess. |
|
Aren't eigenvalue computations large and slow enough that there wouldn't be gains from specialization? Is the argument that it would make the rest of the computation type unstable in some way? |
|
Perhaps we may have an internal function that specialises on |
|
@ViralBShah, the argument is that the extension API is not as discoverable. Performance and type stability should be fine. # LinearAlgebra.jl
public eigen
eigen(::AbstractMatrix, ::Algorithm) = ...
---
# EigenAlgorithms.jl
eigen(::AbstractMatrix, ::MyFancyAlgorithm) = ...vs # LinearAlgebra.jl
public eigen, eigen_dispatch
eigen(x::AbstractMatrix; alg::Algorithm) = eigen_dispatch(x, a)
eigen_dispatch(::AbstracMatrix, ::Algorithm) = ...
---
# EigenAlgorithms.jl
eigen_dispatch(::AbstractMatrix, ::MyFancyAlgorithm) = ... |
|
I don't think this needs to be particularly discoverable, as this isn't user-facing. We may add a note to developers in the docstring of eigen to point to this function. |
|
We have moved the LinearAlgebra stdlib to an external repo: https://github.com/JuliaLang/LinearAlgebra.jl @jishnub If you think that this PR is still relevant, please open a new PR on the LinearAlgebra.jl repo. |
This reopens JuliaLang/julia#55481 One argument there was that packages may wish to define custom eigenvalue algorithms. I'm unsure if this is common, but in principle, we may provide an internal method `eigen(A, alg)` that packages extend, and users may call the method with `alg` as a keyword argument. Fixes #1208 --------- Co-authored-by: Viral B. Shah <[email protected]>
This reopens JuliaLang/julia#55481 One argument there was that packages may wish to define custom eigenvalue algorithms. I'm unsure if this is common, but in principle, we may provide an internal method `eigen(A, alg)` that packages extend, and users may call the method with `alg` as a keyword argument. Fixes #1208 --------- Co-authored-by: Viral B. Shah <[email protected]>
Currently,
eigenfor symmetric matrices accepts analgas the second positional argument, but this departs from the pattern ofeigen(A, B)representing a generalized eigenvalue problem. This PR movesalgto a keyword argument instead. This prevents packages from specializing onalg, but I'm unsure if there's demand for that.Incidentally,
algas a keyword argument ineigenis also consistent with how it is passed to thesortfamily of functions, although there's no direct relation between these.