Infinite default bounds for AbstractFloat #1331
Conversation
| _no_lower_bound(::Type{T}) where {T} = zero(T) | ||
| _no_lower_bound(::Type{T}) where {T<:AbstractFloat} = typemin(T) | ||
| _no_upper_bound(::Type{T}) where {T} = zero(T) | ||
| _no_upper_bound(::Type{T}) where {T<:AbstractFloat} = typemax(T) |
There was a problem hiding this comment.
Why is zero(T) necessary and not just typemin(T) and typemax(T)?
There was a problem hiding this comment.
Answered my own question:
julia> typemax(Complex{Int})
ERROR: MethodError: no method matching typemax(::Type{Complex{Int64}})There was a problem hiding this comment.
Also, typemax(Int) is a bit scary with overflows. Maybe we can say that we don't guarantee anything for non-AbstractFloat so that we can change our mind in the future
There was a problem hiding this comment.
The other option is to make the bound vectors Union{Nothing,T}?
There was a problem hiding this comment.
Who uses AbstractModel with something other than an AbstractFloat at present?
There was a problem hiding this comment.
The advantage here is that Clp can directly take the Vector{Float64} and pass it to the inner solver without the need to do any transformation: https://github.com/jump-dev/Clp.jl/pull/111/files#diff-18c6af231f3facc5ba10dd45231da349efda22ecd8d68a4661dee8027bc413a7R279-R280.
We already have a vector of flags that contains the information on whether a bound is set or not.
So the value of the bound if it is not set does not alter the MOI API.
Using Union{Nothing,T} would make things slower, unallow Clp and friends to use it directly and would not bring any advantage.
odow
left a comment
odow
left a comment
There was a problem hiding this comment.
For now, this is at least just a copy of existing behavior for non-floats.
Needed for jump-dev/Clp.jl#111. With this PR,
lower_boundandupper_boundare now part of the API since it's documented so Clp can use it directly.