How to derivative quassi likelihood function using ForwardDiff

General Usage
1

Continuing the discussion from How to derivative using ForwardDiff:

I want some help. I want to make more complicated function including more parameters

Please teach me how to apply ForwardDiff.gradient() and ForwardDiff.hessian() to this function.
g is a penalized quassi liklifood funcion

- g(m_{x,t}) = ∑_{x,t} (D_{x,t} \ln m_{x,t} − E_{x,t} m_{x,t} )− \dfrac{1}{2} (k− μ)′ {V}^{−1} (k− μ)

where

m_{x,t} = \exp \{ a_x + b_x k_t + c_{x,t} \pi_t \}

D_{x,t} and E_{x,t} is a fixed matrix

function g(params; X,T)
    a,b,k,c,π,μ_vec,V = params
    R = eltype(params)
    @show R 
    # a = zeros(R,X)
    # b = zeros(R,X)
    # k = zeros(R,T)
    # c = zeros(R,X,T)
    # π = zeros(R,T)
    # μ_vec = zeros(R,T)
    # V = zeros(R,T,T)
    term1 = Dxt .* (a .+ b * k' + c .* π' ) - Ext .* exp.( a .+ b * k' + c .* π') 
    term2 = (k - μ_vec)' * (V \ (k - μ_vec) )
    return sum(term1) - term2 / 2
end

Here is a reference

2

Unlike other autodiff frameworks like Zygote, ForwardDiff only supports functions with a single array argument. I think a good solution is to use a ComponentArray instead of a tuple

3

Thank you for your reply.

Could you teach me how to maximize this quassi likelihood function using Zygote.jl or Optim.jl?

using Optim
using ComponentArrays
using Plots
using LaTeXStrings


function g(params; X,T)
    a,b,k,c,π,μ_vec,V = params
    term1 = Dxt .* (a .+ b * k' + c .* π' ) - Ext .* exp.( a .+ b * k' + c .* π') 
    term2 = (k - μ_vec)' * (V \ (k - μ_vec) )
    return sum(term1) - term2 / 2
end  


params = ComponentArray(a = a0, b = b0, k = k0, c = c0, π = π0, μ_vec = μ0_vec, V = V0)
g(params;X,T)
#jacoian of g
dg(params) = ForwardDiff.gradient(param -> g(param; X = X, T = T),params)
#Hessian of g
ddg(params) = ForwardDiff.hessian(param -> g(param; X = X, T = T),params)

#get solution of Optimization by Newton method
result = Optim.optimize(-g, -dg, -ddg, params, Optim.Newton(), Optim.Options(g_tol = 1e-10, iterations = 1000))

I got this error

MethodError: no method matching -(::typeof(g)) Closest candidates are: -(::Any, ::AbstractThunk) at ~/.julia/packages/ChainRulesCore/a4mIA/src/tangent_types/thunks.jl:35 -(::Any, ::ZeroTangent) at ~/.julia/packages/ChainRulesCore/a4mIA/src/tangent_arithmetic.jl:102 -(::Any, ::NoTangent) at ~/.julia/packages/ChainRulesCore/a4mIA/src/tangent_arithmetic.jl:62 … Stacktrace: [1] top-level scope @ In[122]:23

4

Trying to interpret error stacktraces in Julia is difficult at first, but it’s a worthy skill to acquire! What do you think this error means? Here’s a clue: -(typeof(g)) means that you’re trying to apply the function - (“minus”) to an object that has the same type as g (in this case the function g itself)