Misleading of code in src/approximations/linearization.jl

[Flawless1202]

In the following code in src/approximations/linearization.jl:

ReactiveMP.jl/src/approximations/linearization.jl

Lines 68 to 84 in 308dce4

	# In case if `g(x_hat)` returns a number and inputs are numbers too
	function local_linearization(r::Real, splitg::S, g::G, x_hat) where {S, G}
	return local_linearization(r, splitg, g, (g, lx_hat) -> (ForwardDiff.gradient(splitg, lx_hat)::Vector{eltype(lx_hat)})', x_hat)
	end
	# In case if `g(x_hat)` returns a vector, but inputs are numbers
	function local_linearization(r::AbstractVector, splitg::S, g::G, x_hat) where {S, G}
	return local_linearization(r, splitg, g, (g, lx_hat) -> (ForwardDiff.jacobian(splitg, lx_hat)::Matrix{eltype(lx_hat)}), x_hat)
	end
	function local_linearization(r, splitg::S, g::G, fA::F, x_hat) where {S, G, F}
	lx_hat = __as_vec(x_hat)
	# `fA` calls either `gradient` or `jacobian`, depending on the type of the `r`
	A = fA(splitg, lx_hat)
	b = r - A * lx_hat
	return (A, b)
	end

As far as I am concerned, the lambda function

(g, lx_hat) -> (ForwardDiff.gradient(splitg, lx_hat)::Vector{eltype(lx_hat)})'

and

(g, lx_hat) -> (ForwardDiff.jacobian(splitg, lx_hat)::Matrix{eltype(lx_hat)})

should be

(splitg, lx_hat) -> (ForwardDiff.gradient(splitg, lx_hat)::Vector{eltype(lx_hat)})'

and

(splitg, lx_hat) -> (ForwardDiff.jacobian(splitg, lx_hat)::Matrix{eltype(lx_hat)})

Is it correct?

[bvdmitri]

I believe you're right here, but fortunately, later on in line 81, we pass splitg as the g argument, making it effectively equivalent. The code is indeed a bit tricky to read, but I don't anticipate it being a real bug. We can clean it up, though.

I don't quite remember the original reasoning behind using the lambda function, but it had to do with ForwardDiff requiring a function with a single vector argument, while we wanted to handle multi-argument functions as well."