Skip to content

Third-order derivatives of x / y are Inf or NaN for small y, e.g. in gradients of Laplace approximations #866

Closed

Description

@devmotion

Gradients of Laplace approximations of marginal likelihoods require third derivatives of the log-density: the Laplace approximation contains the log-determinant of the Hessian with respect to the latent variables, and its gradient with respect to the parameters is computed by differentiating that Hessian once more. With ForwardDiff, these third derivatives are Inf or NaN when the log-density contains a division by a small number, even though the derivatives themselves are finite and well within the range of Float64.

For example, in a latent-variable model with an exponential decay and multiplicative noise, the scale of the noise becomes small for observations at later times:

using ForwardDiff, Distributions

a, ω, σ = 30.0, 1.0, 0.3
t = [1.0, 11.0]
y = [0.0014, 4.7e-47]

decay(k, z) = @. a * exp(-k * exp(z) * t)
function logjoint(k, z)
    μ = decay(k, z)
    return sum(logpdf.(Normal.(μ, σ .* μ), y)) + logpdf(Normal(0, ω), z)
end

# Laplace approximation of log ∫ p(y, z | k) dz at the mode ẑ of p(y, z | k)
∂z(k, z) = ForwardDiff.derivative(z -> logjoint(k, z), z)
∂²z(k, z) = ForwardDiff.derivative(z -> ∂z(k, z), z)
laplace(k, ẑ) = logjoint(k, ẑ) - log(-∂²z(k, ẑ)) / 2 + log(2π) / 2

function mode(k)
    lo, hi = 3.0, 3.5 # ∂z(k, lo) > 0 > ∂z(k, hi)
    for _ in 1:60
        mid = (lo + hi) / 2
        ∂z(k, mid) > 0 ? (lo = mid) : (hi = mid)
    end
    return (lo + hi) / 2
end
julia> k = 0.4;

julia> ẑ = mode(k)
3.2202943922395297

julia> laplace(k, ẑ)
102.67519619445666

julia> ForwardDiff.derivative(k -> laplace(k, ẑ), k)
-Inf

julia> h = 1e-6; (laplace(k + h, ẑ) - laplace(k - h, ẑ)) / 2h
-386.2937856666804

Cause

logpdf(Normal(μ, σ), y) computes (y - μ) / σ, here with σ ≈ 1.3e-47 for the second observation. For dual numbers x and y, the partials of x / y are computed as inv(bval) * a - (aval / (bval * bval)) * b:

https://github.com/JuliaDiff/ForwardDiff.jl/blob/v1.4.6/src/partials.jl#L99-L101

With nested duals, bval is itself a dual number, so aval / (bval * bval) divides by a dual number again, and the n-th derivative divides by y^(2^n). In the example, the third derivative divides by σ^8 ≈ (1.3e-47)^8, which underflows to 0.

The same happens without any statistics:

julia> f(x) = x / (x^2 + x^3); # = 1 / (x + x^2)

julia> g(x) = 1 / (x + x^2);

julia> d3(f, x) = ForwardDiff.derivative(x -> ForwardDiff.derivative(x -> ForwardDiff.derivative(f, x), x), x);

julia> for x in (1e-20, 1e-30, 1e-40)
           println("x = $x: ", d3(f, x), " vs ", d3(g, x))
       end
x = 1.0e-20: -6.000712508240443e80 vs -6.000000000000001e80
x = 1.0e-30: -Inf vs -5.999999999999998e120
x = 1.0e-40: NaN vs -6.000000000000002e160

Even at x = 1e-20 the third derivative of f has a relative error of 1e-4. The derivative of g is correct since a non-dual numerator is handled without the square (-(divv / v) * partials(y)):

https://github.com/JuliaDiff/ForwardDiff.jl/blob/v1.4.6/src/dual.jl#L539-L543

Computing the partials as (a - (aval / bval) * b) / bval, i.e., dividing by bval twice instead of by bval * bval, gives -386.2936383633966 in the first example and -6.000000000000001e80, -6.000000000000002e120 and -6.000000000000005e160 in the second one.

ForwardDiff v1.4.6, Distributions v0.25.132, Julia 1.13.2.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions