Add range indexing to UniformScaling

[dalum]

Ref: #23156. I thought this might be a usable alternative to eye and having callable UniformScaling.

[stevengj]

It's kind of unsatisfying to have this type-unstable getindex. I'm not sure I would encourage people to use an API like this.

Maybe just always return SparseMatrix?

[mbauman]

Just a few minor changes. 👍

[mbauman]

Interestingly, this should also work for negative indices. At first I thought about requesting a bounds check there, but given that I should work similarly well for OffsetArrays… as should this definition as-is… I think we could leave that ability as a feature. Thoughts?

[dalum]

@mbauman Thanks for the review! ❤️ That it works for negative indices was intentional, since integer indexing works that way. Any other infinite-dimensional abstract matrix could also allow arbitrary integer indexing, so I see no good argument for enforcing strictly positive indexing.

[stevengj]

Do we want to resurrect this at some point?

It occurred to me recently that it might be nice to have the related syntax I[1:m, n] for forming the n-th unit vector (as a Vector).

review was addressed!

[dalum]

@stevengj I updated the PR to include the indexing you suggested. I also had to change the behaviour of range indexing to return a dense array, since the separation of LinearAlgebra and SparseArrays meant that sparse functions were not available in the scope. I think it's better not to make LinearAlgebra depend on SparseArrays for this functionality?

The implementation right now is about twice as fast as first instancing a diagonal identity matrix through I(n) and then subsequently indexing into it, at least for small matrices.

[stevengj]

Needs a NEWS item?

[dalum]

@mbauman It is not entirely clear to me where to add the compat annotation. There are no existing docstrings for getindex(::UniformScaling), so should I write one? I'm just a little worried about cluttering ?getindex.

[mbauman]

Ah, very good question. I agree about it not really belonging in getindex, but I think we could say something in the type itself:

julia/stdlib/LinearAlgebra/src/uniformscaling.jl

Lines 9 to 10 in 147d5b1

	Generically sized uniform scaling operator defined as a scalar times the
	identity operator, `λ*I`. See also [`I`](@ref).

It could be nice to say something about it acting somewhat matrix-like with some indexing supported but without an explicit size — and then a compat entry would fit there.

[dalum]

Thanks! Not sure if I worded it right, but I added a docstring example that should give some idea about the kind of indexing that is supported.

[mbauman]

Looks great to me, thank you!

[codecov-commenter]

Codecov Report

Merging #24359 into master will increase coverage by 0.02%.
The diff coverage is 96.20%.

@@            Coverage Diff             @@
##           master   #24359      +/-   ##
==========================================
+ Coverage   86.12%   86.14%   +0.02%     
==========================================
  Files         348      348              
  Lines       64939    64971      +32     
==========================================
+ Hits        55928    55970      +42     
+ Misses       9011     9001      -10     

Impacted Files	Coverage Δ
base/multimedia.jl	50.00% <ø> (ø)
stdlib/Test/src/Test.jl	89.74% <75.00%> (-0.13%)	⬇️
stdlib/SparseArrays/src/sparsevector.jl	95.26% <88.88%> (-0.07%)	⬇️
stdlib/SparseArrays/src/sparsematrix.jl	89.68% <91.66%> (+<0.01%)	⬆️
base/show.jl	88.53% <100.00%> (ø)
stdlib/SparseArrays/src/linalg.jl	95.14% <100.00%> (+0.36%)	⬆️
base/reshapedarray.jl	86.15% <0.00%> (-0.77%)	⬇️
base/bitset.jl	93.20% <0.00%> (ø)
base/missing.jl	88.52% <0.00%> (+0.54%)	⬆️
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7301dc6...e4d4762. Read the comment docs.

[stevengj]

Should we make I(m,n) return I[1:m, 1:n]?

[StefanKarpinski]

Seems like the only sane definition to me, so 👍