I am interested in taking the derivative of a polynomial with respect to its coefficients. When the coefficients are such that all terms of order $n \ge k$ for some $k$ are 0, automatic differentiation packages return 0 as the derivative with respect to those coefficients. Of course, the true derivative of $\beta_0 + \beta_1 x + \cdots + \beta_K x^K$ with respect to $\beta_n$ should be $x^n$.
Here is an example:
using Polynomials
using ForwardDiff
f(x) = Polynomial(x)(0.5)
∇f(x) = ForwardDiff.gradient(f, x)
true_∇f(x) = 0.5.^(0:length(x)-1)
all_params = [[1.0, 2.0, 3.0], [1.0, 0.0, 3.0], [1.0, 0.0, 0.0]]
for param in all_params
println("Params are $(param).")
println("Gradient from ForwardDiff: $(∇f(param))")
println("True gradient: $(true_∇f(param))")
println("")
end
The output from this example is below. Note the issue in the final example.
Params are [1.0, 2.0, 3.0].
Gradient from ForwardDiff: [1.0, 0.5, 0.25]
True gradient: [1.0, 0.5, 0.25]
Params are [1.0, 0.0, 3.0].
Gradient from ForwardDiff: [1.0, 0.5, 0.25]
True gradient: [1.0, 0.5, 0.25]
Params are [1.0, 0.0, 0.0].
Gradient from ForwardDiff: [1.0, 0.0, 0.0]
True gradient: [1.0, 0.5, 0.25]
I am interested in taking the derivative of a polynomial with respect to its coefficients. When the coefficients are such that all terms of order $n \ge k$ for some $k$ are 0, automatic differentiation packages return 0 as the derivative with respect to those coefficients. Of course, the true derivative of $\beta_0 + \beta_1 x + \cdots + \beta_K x^K$ with respect to $\beta_n$ should be $x^n$.
Here is an example:
The output from this example is below. Note the issue in the final example.
Is there a way to avoid this issue?