Finite amplitude dQNTK response and bound
OpenQAlgorithms.Papers.Independent.Intakenext100.P88.amplitude_responseUse the same ordered unitary circuit and normalized feature states, but output all complex amplitudes , . Duplicate each output as and . The worldsheet index therefore records whether an amplitude is conjugated. For real parameters, write .
Absorb the reference angles, rescale by , and compute doubled jets and at zero directly from the actual circuit. For initial angles , set , , and , all restricted to doubled training indices. Let be the quadratic Taylor amplitude output minus the complex labels and their conjugates. With and , the retained coefficient recurrences and , initialized by , satisfy
The retained residual is . This is a fixed-time coefficient response and requires no impossible full doubled inverse or nonlinear remainder estimate.
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 p.33 Eqs.(170)-(173): finite doubled-amplitude response and L2 norm bound.
All jets come from the original U; the response retains conjugation symmetry and eta/2.
The coefficient recurrence needs no inverse of the necessarily singular doubled kernel. -/
theorem amplitude_response {D T : Type*} [Fintype T] [DecidableEq T] {h L : ℕ}
(M : AmplitudeModel D h L) (embed : T ↪ D)
(delta eta : ℝ) (hdelta : 0 < delta) (heta : 0 < eta)
(v : Fin L → ℝ) (y : T × Fin h → ℂ) :
let incl := fun a : T × Bool × Fin h => (embed a.1, a.2.1, a.2.2)
let K : Matrix (T × Bool × Fin h) (T × Bool × Fin h) ℂ := fun a b =>
frozenAmplitudeKernel M delta (incl a) (incl b)
let D : Matrix (T × Bool × Fin h) (T × Bool × Fin h) ℂ := fun a b =>
deltaAmplitudeKernel M delta v (incl a) (incl b)
let e : EuclideanSpace ℂ (T × Bool × Fin h) := WithLp.toLp 2 (fun a =>
quadraticAmplitude M delta v (incl a) -
(if a.2.1 then star (y (a.1, a.2.2)) else y (a.1, a.2.2)))
let R := retainedResidual K D (eta / 2) e
(∀ t, (R t).1 = l2Action ((1 - (eta / 2) • K) ^ t) e) ∧
(R 0).2 = 0 ∧
(∀ t : ℕ, (R (t + 1)).2 = -(eta / 2 : ℝ) •
(∑ s ∈ Finset.range (t + 1),
l2Action (((1 - (eta / 2) • K) ^ (t - s)) * D * ((1 - (eta / 2) • K) ^ s)) e)) ∧
∀ t : ℕ, ‖(R (t + 1)).2‖ ≤ (eta / 2) * ((t : ℝ) + 1) *
QAlgorithms.specNorm (1 - (eta / 2) • K) ^ t * QAlgorithms.specNorm D * ‖e‖ := 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 finite and with decidable equality, let be natural numbers, and let consist of , unitary matrices , Hermitian matrices , and unit vectors . Define in increasing index order, , and . Fix an injection , , , and complex labels . Let , , , and , using real coordinate derivatives. Define , , and
Set , , , and . For complex Euclidean vectors defined by , , , and , one has for every , , and
The operator norms are induced by complex Euclidean norms. Empty and are allowed, with empty sums equal to zero and empty products equal to the identity; every exponent in the sum is nonnegative.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.