MethodError bounds for CubicSplineInterpolation

[hendri54]

Calling bounds on a CubicSplineInterpolation throws a MethodError:

julia> xV = LinRange(-20.0, 10.0, 50);
julia> yV = LinRange(-10.0, 30.0, 40);
julia> zM = xV .+ xV .* yV';
julia> cubic1 = CubicSplineInterpolation((xV, yV), zM);
julia> bounds(cubic1)
ERROR: MethodError: no method matching lbounds(::Interpolations.Extrapolation{Float64, 2, ScaledInterpolation{Float64, 2, Interpolations.BSplineInterpolation{Float64, 2, OffsetArrays.OffsetMatrix{Float64, Matrix{Float64}}, BSpline{Cubic{Line{OnGrid}}}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}}, BSpline{Cubic{Line{OnGrid}}}, Tuple{LinRange{Float64}, LinRange{Float64}}}, BSpline{Cubic{Line{OnGrid}}}, Throw{Nothing}})
Closest candidates are:
  lbounds(::Interpolations.BSplineInterpolation) at /Users/lutz/.julia/packages/Interpolations/GIn2o/src/b-splines/b-splines.jl:133
  lbounds(::Interpolations.GriddedInterpolation) at /Users/lutz/.julia/packages/Interpolations/GIn2o/src/gridded/gridded.jl:92
  lbounds(::Interpolations.MonotonicInterpolation) at /Users/lutz/.julia/packages/Interpolations/GIn2o/src/monotonic/monotonic.jl:387
  ...
Stacktrace:
 [1] bounds(itp::Interpolations.Extrapolation{Float64, 2, ScaledInterpolation{Float64, 2, Interpolations.BSplineInterpolation{Float64, 2, OffsetArrays.OffsetMatrix{Float64, Matrix{Float64}}, BSpline{Cubic{Line{OnGrid}}}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}}, BSpline{Cubic{Line{OnGrid}}}, Tuple{LinRange{Float64}, LinRange{Float64}}}, BSpline{Cubic{Line{OnGrid}}}, Throw{Nothing}})
   @ Interpolations ~/.julia/packages/Interpolations/GIn2o/src/Interpolations.jl:146
 [2] top-level scope

Constructing what appears to be the same interpolation the long way round does not throw a MethodError:

julia> cubic2 = interpolate(zM, BSpline(Cubic(Line(OnGrid()))));
julia> cubic2 = scale(cubic2, xV, yV);
julia> cubic1(-10.5, 7.5)
-89.25
julia> cubic2(-10.5, 7.5)
-89.25
julia> bounds(cubic2)
((-20.0, 10.0), (-10.0, 30.0))

I thought both ways of constructing the CubicSplineInterpolation were the same. What am I missing?

[mkitti]

In the first case, it seems to be producing an Extrapolation for some reason.

[hendri54]

Yes, but still throws a BoundsError when out of range. So the Extrapolation seems to be Throw in both cases.

[mkitti]

Can you obtain the bounds via bounds(parent(cubic1))?

[hendri54]

Yes, that works.

[axsk]

For me I need to use bounds(parent(itp)) even in the one dimensional case. Not quite so intuitive..

[mkitti]

The problem is that the "convenience" methods produce a fully wrapped, with extrapolation, instance.

Would bounds returning the extrapolation wrapped internal bounds of the interpolation be intuitive?

[axsk]

I guess I agree that the convenciene linear_interpolation returning an extrapolation is a problem / counterintuitive :)

In any case, I think bounds on an extrapolation with Throw returning the wrapped internal bounds makes sense, as these indeed are the bounds even of the extrapolation:

bounds(itp::Extrapolation{<:Any, <:Any, <:Any, <:Any, <:Throw})=bounds(parent(itp))

[mkitti]

Ok, that is helpful. Let's do that.

I will also look into intercepting the bounds call and presenting a suggestion of what the user might want.