The chain rule
ProvedFamousTheorems.deriv_compThe chain rule. If is differentiable at and is differentiable at , then
Rates of change compose by multiplication. The statement is the derivative version of the fact that the best linear approximation to a composite is the composite of the best linear approximations, which is why the general form in several variables is composition of Jacobians rather than a product of numbers. Both differentiability hypotheses are needed and at the right points — at , but at , not at . The rule is what makes differentiation algorithmic: together with the product and sum rules it reduces any elementary derivative to a finite computation, and run in reverse over a computation graph it is exactly backpropagation. Formalization note. The scalar fields for the inner and outer functions may differ, with NormedAlgebra relating them, so the statement covers real functions of a complex variable and similar mixed cases. The result is Mathlib's deriv_comp.
import Mathlib
namespace FamousTheorems
universe u_1 u_2 u_3 u_4 u_5 u_6 u_7 u_8 u_9 u_10 u_11 u_12 u_13 u_14 u_15 u_16 u_17 u_18 u_19 u_20 u_21 u_22 u_23 u_24 u_25
open Filter Set Topology DirectSum
theorem deriv_comp :
∀ {𝕜 : Type u_1} [inst : NontriviallyNormedField 𝕜] (x : 𝕜) {𝕜' : Type u_2}
[inst_1 : NontriviallyNormedField 𝕜'] [inst_2 : NormedAlgebra 𝕜 𝕜'] {h : 𝕜 → 𝕜'} {h₂ : 𝕜' → 𝕜'},
DifferentiableAt 𝕜' h₂ (h x) → DifferentiableAt 𝕜 h x → deriv (h₂ ∘ h) x = deriv h₂ (h x) * deriv h x := by sorry
end FamousTheorems