divonne and cuhre pass vector of length 2 for one-dimensional integration

[oschulz]

When integrating in one dimension, vegas and suave call the integrand with a vectir of length 1 (as expected), but divonne and cuhre pass a vector of length 2:

julia> vegas((x, f) -> ((@assert length(x) == 1); f[1] = norm(x)), 1)
...

julia> suave((x, f) -> ((@assert length(x) == 1); f[1] = norm(x)), 1)
...

julia> divonne((x, f) -> ((@assert length(x) == 1); f[1] = norm(x)), 1)
ERROR: AssertionError: length(x) == 1
julia> divonne((x, f) -> ((@assert length(x) == 2); f[1] = norm(x)), 1)
...

julia> cuhre((x, f) -> ((@assert length(x) == 1); f[1] = norm(x)), 1)
ERROR: AssertionError: length(x) == 1
julia> cuhre((x, f) -> ((@assert length(x) == 2); f[1] = norm(x)), 1)
...

[giordano]

Hi Oliver. That's correct, those methods require ndim to be at least 2. Did you run into some issues because of this?

[oschulz]

Hi, thanks - should have looked into the Cuba docs. :-) I did run into issues, but it was only in a toy case. But maybe divonne and cuhre should throw an ArgumentError if run for one dimension, just in case?

[giordano]

Currently I forcibly set ndim = 2 if it's equal to 1:

Cuba.jl/src/cuhre.jl

Line 94 in 9d0354a

ndim == 1 && (ndim = 2)

I thought that this would have been quite harmless, but I see that using something like norm can lead to unexpected results.

I will change the default value of ndim for cuhre and divonne to 2, throw an error for lower values, better document this.

[oschulz]

grazie!

[giordano]

Bitte ;-) I've merged PR #15, do let me know if you need a tagged version of this package.

[oschulz]

Thanks, no, I don't need an urgent tag on this.