First-order QNTK gradient-descent evolution
ProvedQAlgorithms.Papers.Independent.Intakenext100.P88.linearized_gradient_lawThe circuit acts on a nonzero finite-dimensional complex Hilbert space and has the ordered form , with fixed unitary , Hermitian , and real angles. Products are written in increasing index order and the rightmost factor acts first. Feature vectors have norm one. For Hermitian observables , put .
Let be a finite training subset, a finite observable family, real labels, , and . Use the unaveraged loss . For the real gradient step , its first Taylor output increment satisfies
For complex amplitude outputs, duplicate residuals and outputs into conjugate pairs, and form the complex Gram kernel of the doubled real-parameter derivatives. The analogous real gradient of gives . The scalar formula is the singleton-data, singleton-observable case. These identities describe first Taylor increments and provide the evolution laws used in the frozen models.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P88
noncomputable section
namespace QAlgorithms.Papers.Independent.Intakenext100.P88
open scoped BigOperators ComplexConjugate Topology NNReal ComplexOrder
open MeasureTheory ProbabilityTheory Filter
/-- PDF pp.2,5,23 Eqs.(4),(25)-(28),(88)-(93), and pp.28-29 Eqs.(132)-(140).
This is the first Taylor increment induced by the real gradient of the unaveraged squared
loss. It is not equality with an exact nonlinear one-step circuit increment. Unit data and
Unit observables recover the scalar optimization law. -/
theorem linearized_gradient_law {D T I : Type*} [Fintype T] [Fintype I]
{h L : ℕ} (M : ExpectationNetwork D I h L) (embed : T ↪ D)
(theta : Fin L → ℝ) (y : T × I → ℝ) (eta : ℝ) (heta : 0 < eta) :
(let e := fun b : T × I => expectationOutput M theta (trainingIndex embed b) - y b
let dp := fun l => -eta * ∑ b, e b *
coordinateDerivative (fun u => expectationOutput M u (trainingIndex embed b)) theta l
∀ a, (∑ l, coordinateDerivative (fun u => expectationOutput M u a) theta l * dp l) =
-eta * ∑ b, expectationKernel M theta a (trainingIndex embed b) * e b) ∧
(∀ (N : AmplitudeModel D h L) (labels : T × Fin h → ℂ),
let e := fun b : T × Fin h => basisAmplitude N theta (embed b.1, b.2) - labels b
let dp := fun l => -eta * (∑ b, e b *
star (coordinateDerivative (fun u => basisAmplitude N u (embed b.1, b.2)) theta l)).re
let de := fun b : T × Bool × Fin h =>
if b.2.1 then star (e (b.1, b.2.2)) else e (b.1, b.2.2)
∀ a, (∑ l, coordinateDerivative (fun u => doubledAmplitude N u a) theta l * (dp l : ℂ)) =
-(eta / 2 : ℝ) • (∑ b : T × Bool × Fin h,
amplitudeKernel N theta a (embed b.1, b.2.1, b.2.2) * de b)) := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P88
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let be types with and finite, let be natural numbers, and let consist of , unitary matrices , Hermitian matrices , unit vectors for , and Hermitian observables for . Define in increasing index order, and . Let be injective, write , and fix arbitrary , labels , and . All coordinate derivatives below are with respect to real parameters. Define , , and . Then, for every , . In addition, for every amplitude model of the same dimensions—consisting of its own unitary , Hermitian , and unit vectors —and every complex label function , define , , , , , and , . With , every satisfies . Empty parameter and training sums are allowed; and need not be nonempty.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.