failure on v1.11-dev

[stevengj]

Currently failing on v"1.11.0-DEV.745" with:

exp.jl: Error During Test at /Users/stevenj/.julia/dev/SkewLinearAlgebra/test/runtests.jl:224
  Test threw exception
  Expression: tan(B) ≈ tan(A)
  StackOverflowError:
  Stacktrace:
   [1] SkewHermitian{Float64, SkewHermitian{Float64, Matrix{Float64}}}(A::SkewHermitian{Float64, Matrix{Float64}}) (repeats 79984 times)
     @ SkewLinearAlgebra ~/.julia/dev/SkewLinearAlgebra/src/skewhermitian.jl:57
exp.jl: Error During Test at /Users/stevenj/.julia/dev/SkewLinearAlgebra/test/runtests.jl:224
  Test threw exception
  Expression: tan(B) ≈ tan(A)
  StackOverflowError:
  Stacktrace:
   [1] SkewHermitian{Float64, SkewHermitian{Float64, Matrix{Float64}}}(A::SkewHermitian{Float64, Matrix{Float64}}) (repeats 79984 times)

[stevengj]

Errors occur for 1x1 and 2x2 matrices, it looks like:

julia> A = skewhermitian(rand(2,2))
2×2 SkewHermitian{Float64, Matrix{Float64}}:
 0.0       -0.259484
 0.259484   0.0

julia> tan(A)
ERROR: StackOverflowError:
Stacktrace:
 [1] SkewHermitian{Float64, SkewHermitian{Float64, Matrix{Float64}}}(A::SkewHermitian{Float64, Matrix{Float64}}) (repeats 79984 times)
   @ SkewLinearAlgebra ~/.julia/dev/SkewLinearAlgebra/src/skewhermitian.jl:57

[stevengj]

Looks like we just need to add a constructor SkewHermitian(A::SkewHermitian) = A

Though fixing that exposes another failure on 1.11 in this test for (T, n) = (ComplexF32, 1):

exp.jl: Test Failed at /Users/stevenj/.julia/dev/SkewLinearAlgebra/test/runtests.jl:225
  Expression: tan(B) ≈ tan(A)
   Evaluated: ComplexF32[0.0f0 + 0.7081147f0im;;] ≈ ComplexF32[0.0f0 + 0.0f0im;;]

In particular, we have this incorrect result on 1.11:

julia> tan(skewhermitian([0.1im;;]))
1×1 SkewHermitian{ComplexF64, Matrix{ComplexF64}}:
 0.0 + 0.0im

whereas it gives the correct result [0.0 + 0.0996679946249559im;;] on 1.9.

Indeed, SkewHermitian(skewhermitian(rand(4,4))) gives a stack overflow even on earlier versions, so it seems like the tan code is somehow triggering a new code path on 1.11.

[stevengj]

In particular, I'm getting the following wrong result on 1.11 that is the source of the tan error:

julia> C = Hermitian([1.0+0im;;]); S = SkewHermitian([0+1.0im;;])
1×1 SkewHermitian{ComplexF64, Matrix{ComplexF64}}:
 0.0 + 1.0im

julia> C \ S
1×1 SkewHermitian{ComplexF64, Matrix{ComplexF64}}:
 0.0 + 0.0im

In both 1.9 and 1.11 it's calling a Base method here, so this might be an error in Base on 1.11:

julia> @which C \ S
\(A::Union{Hermitian{T, S}, Symmetric{T, S}} where {T, S}, B::AbstractMatrix)
     @ LinearAlgebra ~/Documents/Code/julia/usr/share/julia/stdlib/v1.11/LinearAlgebra/src/symmetric.jl:584

[stevengj]

It looks like the problem was exposed to the Diagonal optimization in JuliaLang/julia#48189, which effectively uses Diagonal(C) \ S in 1.11 (and maybe 1.10)? Even in 1.9, if we explicitly use Diagonal then it fails:

julia> C \ S
1×1 Matrix{ComplexF64}:
 0.0 + 1.0im

julia> Diagonal(C) \ S
1×1 SkewHermitian{ComplexF64, Matrix{ComplexF64}}:
 0.0 + 0.0im

[smataigne]

Thanks!