Fix *(::AbstractJuMPScalar{<:Real}, ::Hermitian) to return Hermitian

[odow]

Closes #3694

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.35%. Comparing base (a9b5ff9) to head (9b0eaee).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3695      +/-   ##
==========================================
- Coverage   98.36%   98.35%   -0.02%     
==========================================
  Files          43       43              
  Lines        5699     5707       +8     
==========================================
+ Hits         5606     5613       +7     
- Misses         93       94       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[araujoms]

Could you also do the Symmetric case? It's actually much easier because you don't need to test whether the scalar is real.

[odow]

This one we can fix in MA: jump-dev/MutableArithmetics.jl#268

[araujoms]

I've just realized that with this definition of multiplication you are generating hidden undefs in the lower triangular, that is, you are relying on the bug JuliaLang/julia#52895 being fixed, which I think is only the case for Julia 1.11 onwards. So the following code will fail for all previous version of Julia:

using JuMP
prob = Model()
@variable(prob,x)
m = x*LinearAlgebra.Hermitian(randn(2,2))
m+m

[odow]

Perhaps we should close this then. The currently behavior is correct. The new PR would introduce the risk of incorrect behavior due to bugs in Base.

If JuMP generates a Matrix, you can always explicitly wrap it in Hermitian.

It's really hard to uniformly support all of LinearAlgebra, SparseArrays, MutableArithmetics, and JuMP.

[araujoms]

Or you could just return instead

LinearAlgebra.Hermitian(Matrix(LinearAlgebra.Hermitian(B, c)),c)

This gets rids of the undefs, so no bugs are possible, you keep the performance gain, and don't require the users to always rewrap everything in Hermitian.

[araujoms]

Thanks a lot for solving all of the problems I pointed out! I'm always amazed by how quick you and @blegat are.

[odow]

No problem. A lot of the problems are because you might be the only person actively using JuMP for complex Hermitian stuff... 😆

[araujoms]

Well, one might say only, but I'd rather say first. 😉

There are also these guys, but they invented their own reformulation of the complex Hermitian PSD cone, so they are only using JuMP for the real one.

[odow]

There are also these guys

You might recognize an author

[araujoms]

Well, I know some of them personally, but there is one that I'm sure you do. 😆