guard against chance singularity in a sparse linalg test #25218
Conversation
Sacha0
added
bugfix
|
(looks like this also happened on https://circleci.com/gh/JuliaLang/julia/12963?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link) |
test/sparse/sparse.jl
Outdated
| @@ -1672,8 +1672,8 @@ end | |||
|
|
|||
| @testset "sparse matrix cond" begin | |||
| local A = sparse(reshape([1.0], 1, 1)) | |||
| Ac = sprandn(20, 20,.5) + im* sprandn(20, 20,.5) | |||
| Ar = sprandn(20, 20,.5) | |||
| Ac = Diagonal(fill(1., 20)) + (sprandn(20, 20,.5) + im*sprandn(20, 20,.5))/40 | |||
There was a problem hiding this comment.
This will probably make it too well condition. I did some simulations locally and
Ar = sprandn(20, 20,.5) + eps()*I
seems sufficient to avoid the singular case while still allowing for some variation in the condition number. It is not necessary to adjust the complex version since adding the two random patterns makes it extremely unlikely that the matrix is singular.
There was a problem hiding this comment.
This solution also seems to take care of the issue reported in #25218 (comment)
There was a problem hiding this comment.
Revised accordingly to worsen conditioning. (Not something I anticipated ever saying 😉.)
|
Thanks all! :) |
Modify a sparse linalg test to guard against chance singularity causing test failure. Failure identified by @vtjnash in https://ci.appveyor.com/project/JuliaLang/julia/build/1.0.22640/job/prc1vdmu7xdix9r3. Best!