Finite leading dQNTK learning response and bound
OpenQAlgorithms.Papers.Independent.Intakenext100.P88.learning_responseThe 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 .
Absorb the reference angles into the fixed gates and set , where . Write and for derivatives of the actual circuit. The frozen kernel is .
Let be a finite training subset of the data space and let be the finite observable index set. All matrices in the training recurrence are restricted to . For initial rescaled angles , put
The initial residual is the quadratic Taylor output at minus the real training labels. The retained coefficient dynamics are , , , and . Thus the retained residual is .
For , set . Then
This is the fixed-time leading response in the retained perturbative model. It supplies no higher-order remainder estimate for exact nonlinear training.
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.6 Theorem 4 Eqs.(35)-(39), SM pp.26-27 Eqs.(114)-(118): leading retained
coefficient response and L2 bound at each finite time. No exact nonlinear GD equality,
uniform-in-time remainder or unsupported source component shorthand is used. -/
theorem learning_response {D T I : Type*} [Fintype T] [DecidableEq T]
[Fintype I] [DecidableEq I] {h L : ℕ}
(M : ExpectationNetwork D I h L) (embed : T ↪ D)
(delta eta : ℝ) (hdelta : 0 < delta) (heta : 0 < eta)
(v : Fin L → ℝ) (y : T × I → ℝ) :
let K := trainingKernel M delta embed
let D : Matrix (T × I) (T × I) ℝ := fun a b =>
deltaKernel M delta v (trainingIndex embed a) (trainingIndex embed b)
let e : EuclideanSpace ℝ (T × I) := WithLp.toLp 2
(fun a => quadraticOutput M delta v (trainingIndex embed a) - y a)
let R := retainedResidual K D eta e
(∀ t, (R t).1 = l2Action ((1 - eta • K) ^ t) e) ∧
(R 0).2 = 0 ∧
(∀ t : ℕ, (R (t + 1)).2 = -eta •
(∑ s ∈ Finset.range (t + 1),
l2Action (((1 - eta • K) ^ (t - s)) * D * ((1 - eta • K) ^ s)) e)) ∧
∀ t : ℕ, ‖(R (t + 1)).2‖ ≤ eta * ((t : ℝ) + 1) *
QAlgorithms.specNorm (1 - eta • 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 and finite and with decidable equality, let be natural numbers, and let consist of , unitary matrices , Hermitian matrices , unit vectors , and Hermitian observables . Define in increasing index order and . Fix an injection , , , and . Put , , , , , , and
Let , , and define real Euclidean vectors , , , . Then for every , , and, for every ,
Here matrix operator norms are induced by Euclidean norms. The statement allows and empty or ; empty sums and vectors have their usual zero values. In the displayed sum , so is always a nonnegative integer.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.