Int / Float type inference in constructor

General Usage
1

I defined a struct as follows:

struct inverse{T<:Number}
x::Array{T}
y::Array{T}
function inverse{T}(x0::Tuple{...,T,...}) where T<:Number
x = convert_to_x(x0)
y = inv(x)
new(promote(x,y)...)
end
end

where the constructor is incomplete. The idea is that y=inv(x), both are square matrices with the same eltype. The input is some nested tuple that contains column vectors (or tuples), I first convert it to a matrix then take the inverse. The problem is when x0 is Int-valued, x becomes Int-valued, and y is float-valued. I can attempt to promote (x,y) to float-valued, but the difficulty is the type T will be inferenced as Int, so the struct object still contains Int-valued arrays. How am I supposed to force the promotion without knowing a-priori the promoted type T?

2

Does it have to be an inner constructor? If not, there is no problem, I think.

function inverse(x0)
  x = [...]
  y = [...]
  inverse(promote(x,y)...)
end
3

Thanks. I had problems with outer constructor. It looks like if this struct is nested in another struct as a field, only the inner constructor is called when the nesting struct is created, in which case I get an error complaining “no matching method for blah blah …” even though that particular method is defined as an outer constructor and works from the REPL.

4

This is mostly unrelated to your question, but you want Matrix{T} not Array{T}

1 Like
5

Why do you want this? I’m very curious to hear a good reason for a numerical matrix inverse.

1 Like
6

You are right. Mostly we don’t want to explicitly compute matrix inverse. My situation is that those are 2x2 SMatrix which are used frequently. It makes the code look cleaner w/o much affecting performance.