Add natural gamma manifold

[Nimrais]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.28%. Comparing base (d59a096) to head (f9beac2).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #26      +/-   ##
==========================================
+ Coverage   96.69%   97.28%   +0.58%     
==========================================
  Files          25       28       +3     
  Lines         333      405      +72     
==========================================
+ Hits          322      394      +72     
  Misses         11       11              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ismailsenoz]

A more generic implementation of Christoffel symbol.

function christoffel(::Type{D}, ::NaturalParametersSpace, η) where D
    # Pre-compute values used multiple times
    ef = ExponentialFamilyDistribution(D, η, nothing)
    G = fisherinformation(ef)
    G⁻¹ = cholinv(G)
    dG = partial_metric(D, NaturalParametersSpace(), η)
    
    T = eltype(G)  # Get the element type from G
    
    # Pre-allocate output array
    rep_size = sum(representation_size(get_natural_manifold(D, size(η))))
    Γ = zeros(T, rep_size, rep_size, rep_size)
        
    # Pre-compute the sum of dG terms
    dG_sum = similar(dG)
    
    @tensor begin
        dG_sum[ℓ,j,i] = dG[ℓ,j,i] + dG[ℓ,i,j] - dG[i,j,ℓ]
    end
    
    # Use tensor operations with pre-computed sum
    @tensor begin
        Γ[k,i,j] = T(0.5) * G⁻¹[k,ℓ] * dG_sum[ℓ,j,i]
    end
    
    return Γ
end

[ismailsenoz]

Is there an ETA on this one?

[Nimrais]

@ismailsenoz I think there are nuances, but more or less all things are addressed.

I don't want to change the partial code for now, I will think about how to do it properly when I address a review for Juliacon, implementing all ideas from this PR but generically for all exponential families.

[ismailsenoz]

ForwardDiff instead of symbolic partial derivatives.