Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Frozen learning: finite dynamics, asymptotic prediction and norm-log rate

Open
QAlgorithms.Papers.Independent.Intakenext100.P88.frozen_learning

by Gohan · Oct 31, 2026 · Mathlib 0df444a (Lean v4.33.1)

prove2quantum

The circuit acts on a nonzero finite-dimensional complex Hilbert space H=ChH=\mathbb C^hH=Ch and has the ordered form U(θ)=∏ℓ=1LWℓexp⁡(iθℓXℓ)U(\theta)=\prod_{\ell=1}^L W_\ell\exp(i\theta_\ell X_\ell)U(θ)=∏ℓ=1L​Wℓ​exp(iθℓ​Xℓ​), with fixed unitary WℓW_\ellWℓ​, Hermitian XℓX_\ellXℓ​, and real angles. Products are written in increasing index order and the rightmost factor acts first. Feature vectors ϕ(x)\phi(x)ϕ(x) have norm one. For Hermitian observables OiO_iOi​, put zx,i(θ)=⟨U(θ)ϕ(x),OiU(θ)ϕ(x)⟩∈Rz_{x,i}(\theta)=\langle U(\theta)\phi(x),O_iU(\theta)\phi(x)\rangle\in\mathbb Rzx,i​(θ)=⟨U(θ)ϕ(x),Oi​U(θ)ϕ(x)⟩∈R.

Absorb the reference angles into the fixed gates and set θ=δφ\theta=\delta\varphiθ=δφ, where δ>0\delta>0δ>0. Write Jaℓ=∂φℓza(0)J_{a\ell}=\partial_{\varphi_\ell}z_a(0)Jaℓ​=∂φℓ​​za​(0) and Ha,kℓ=∂φk∂φℓza(0)H_{a,k\ell}=\partial_{\varphi_k}\partial_{\varphi_\ell}z_a(0)Ha,kℓ​=∂φk​​∂φℓ​​za​(0) for derivatives of the actual circuit. The frozen kernel is Kab=∑ℓJaℓJbℓK_{ab}=\sum_\ell J_{a\ell}J_{b\ell}Kab​=∑ℓ​Jaℓ​Jbℓ​.

Let AAA and III be nonempty finite training and observable index sets. Put KA=K∣(A×I)2K_A=K|_{(A\times I)^2}KA​=K∣(A×I)2​, let z0z^0z0 be the first Taylor output at initial rescaled angles vvv, and set ea=za0−yae_a=z^0_a-y_aea​=za0​−ya​ on training indices. Assume det⁡KA≠0\det K_A\ne0detKA​=0, η>0\eta>0η>0, 1−ηKA≻01-\eta K_A\succ01−ηKA​≻0, and ∥1−ηKA∥2<1\|1-\eta K_A\|_2<1∥1−ηKA​∥2​<1. Equivalently, all eigenvalues obey 0<1−ηλ<10<1-\eta\lambda<10<1−ηλ<1, the positive-step geometric-series and real-log regime used in the source. Under the frozen recurrences, for every t∈Nt\in\mathbb Nt∈N and general data-observable index aaa,

ϵ(t)=(1−ηKA)te⟶0,za(t)⟶za0−∑b,c∈A×IKab(KA−1)bcec.\epsilon(t)=(1-\eta K_A)^t e\longrightarrow0,\qquad z_a(t)\longrightarrow z^0_a-\sum_{b,c\in A\times I}K_{ab}(K_A^{-1})_{bc}e_c.ϵ(t)=(1−ηKA​)te⟶0,za​(t)⟶za0​−b,c∈A×I∑​Kab​(KA−1​)bc​ec​.

On training indices za(t)−ya→0z_a(t)-y_a\to0za​(t)−ya​→0. The paper's statistic τc(s)=∥−log⁡(1−sKA)∥2\tau_c(s)=\|{-\log(1-sK_A)}\|_2τc​(s)=∥−log(1−sKA​)∥2​ satisfies lim⁡s↓0τc(s)/s=∥KA∥2\lim_{s\downarrow0}\tau_c(s)/s=\|K_A\|_2lims↓0​τc​(s)/s=∥KA​∥2​. This packages the finite solution, training interpolation, and prediction on general data. The norm is the L2 operator norm, and the inverse is solely the restricted training inverse.

Preamble
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P88
Formal statement
noncomputable section

namespace QAlgorithms.Papers.Independent.Intakenext100.P88

open scoped BigOperators ComplexConjugate Topology NNReal ComplexOrder

open MeasureTheory ProbabilityTheory Filter

/-- PDF p.6 Theorem 3, SM pp.24-25 Eqs.(98)-(104): frozen finite-time solution,
asymptotic prediction and training interpolation. The inverse is only the training kernel.
Positive definiteness of I-eta K and its strict L2 contraction make the proof's positive
geometric-series/logarithm regime explicit. The norm-log approximation is its one-sided
small-learning-rate leading term; it is not the slowest-mode rate. -/
theorem frozen_learning {D T I : Type*} [Fintype T] [DecidableEq T] [Nonempty T]
    [Fintype I] [DecidableEq I] [Nonempty 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)
    (hinv : (trainingKernel M delta embed).det ≠ 0)
    (hpositive : (1 - eta • trainingKernel M delta embed).PosDef)
    (hcontract : QAlgorithms.specNorm (1 - eta • trainingKernel M delta embed) < 1) :
    let K := trainingKernel M delta embed
    let z := linearOutput M delta v
    let e : EuclideanSpace ℝ (T × I) := WithLp.toLp 2
      (fun a => z (trainingIndex embed a) - y a)
    let Z := prediction (frozenKernel M delta) (trainingIndex embed) eta z e
    (∀ t, frozenResidual K eta e t = l2Action ((1 - eta • K) ^ t) e) ∧
    Tendsto (frozenResidual K eta e) atTop (nhds 0) ∧
    (∀ a, Tendsto (fun t => Z t a) atTop
      (nhds (z a - ∑ b, ∑ c, frozenKernel M delta a (trainingIndex embed b) *
        (K⁻¹) b c * e c))) ∧
    (∀ a, Tendsto (fun t => Z t (trainingIndex embed a) - y a) atTop (nhds 0)) ∧
    Tendsto (fun s : ℝ => logRate K s / s) (nhdsWithin 0 (Set.Ioi 0))
      (nhds (QAlgorithms.specNorm K)) := by sorry

end QAlgorithms.Papers.Independent.Intakenext100.P88

end
Source
Representation Learning via Quantum Neural Tangent Kernels, arXiv:2111.04225v3, PDF p. 6, Main Theorem 3; SM Eqs.(98)-(104)
Read-back

What the Lean code literally says, in plain math · gpt-6.1-sol

Let D,T,ID,T,ID,T,I be types, with TTT and III finite and nonempty and with decidable equality, and let h,Lh,Lh,L be natural numbers. Let MMM consist of h>0h>0h>0, unitary complex h×hh\times hh×h matrices WlW_lWl​, Hermitian matrices XlX_lXl​, unit vectors ϕx∈Ch\phi_x\in\mathbb C^hϕx​∈Ch for x∈Dx\in Dx∈D, and Hermitian observables OiO_iOi​ for i∈Ii\in Ii∈I. Set U(θ)=∏l=0L−1Wlexp⁡(iθlXl)U(\theta)=\prod_{l=0}^{L-1}W_l\exp(i\theta_lX_l)U(θ)=∏l=0L−1​Wl​exp(iθl​Xl​) in increasing index order and f(x,i)(θ)=Re⁡⟨U(θ)ϕx,OiU(θ)ϕx⟩f_{(x,i)}(\theta)=\operatorname{Re}\langle U(\theta)\phi_x,O_iU(\theta)\phi_x\ranglef(x,i)​(θ)=Re⟨U(θ)ϕx​,Oi​U(θ)ϕx​⟩. Fix an injection ι:T↪D\iota:T\hookrightarrow Dι:T↪D, positive reals δ,η\delta,\etaδ,η, v∈RLv\in\mathbb R^Lv∈RL, and labels y:T×I→Ry:T\times I\to\mathbb Ry:T×I→R. Write B=T×IB=T\times IB=T×I, (t,i)^=(ι(t),i)\widehat{(t,i)}=(\iota(t),i)(t,i)​=(ι(t),i), Jl(a)=∂ulfa(δu)∣u=0J_l(a)=\left.\partial_{u_l}f_a(\delta u)\right|_{u=0}Jl​(a)=∂ul​​fa​(δu)∣u=0​, k(a,c)=∑lJl(a)Jl(c)k(a,c)=\sum_lJ_l(a)J_l(c)k(a,c)=∑l​Jl​(a)Jl​(c), Kbc=k(b^,c^)K_{bc}=k(\widehat b,\widehat c)Kbc​=k(b,c), z(a)=fa(0)+∑lJl(a)vlz(a)=f_a(0)+\sum_lJ_l(a)v_lz(a)=fa​(0)+∑l​Jl​(a)vl​, and eb=z(b^)−ybe_b=z(\widehat b)-y_beb​=z(b)−yb​. Assume det⁡K≠0\det K\ne0detK=0, that A=IB−ηKA=I_B-\eta KA=IB​−ηK is positive definite, and that its operator norm for the Euclidean norm satisfies ∥A∥op<1\|A\|_{\mathrm{op}}<1∥A∥op​<1. Define r0=er_0=er0​=e, rt+1=Artr_{t+1}=Ar_trt+1​=Art​, Z0=zZ_0=zZ0​=z, and Zt+1(a)=Zt(a)−η∑b∈Bk(a,b^)(rt)bZ_{t+1}(a)=Z_t(a)-\eta\sum_{b\in B}k(a,\widehat b)(r_t)_bZt+1​(a)=Zt​(a)−η∑b∈B​k(a,b)(rt​)b​. Then rt=Ater_t=A^tert​=Ate for every t∈Nt\in\mathbb Nt∈N, rt→0r_t\to0rt​→0 in Euclidean norm, and for every a∈D×Ia\in D\times Ia∈D×I,

Zt(a)⟶z(a)−∑b,c∈Bk(a,b^)(K−1)bcec.Z_t(a)\longrightarrow z(a)-\sum_{b,c\in B}k(a,\widehat b)(K^{-1})_{bc}e_c.Zt​(a)⟶z(a)−b,c∈B∑​k(a,b)(K−1)bc​ec​.

For every b∈Bb\in Bb∈B, Zt(b^)−yb→0Z_t(\widehat b)-y_b\to0Zt​(b)−yb​→0. Finally, defining ρ(s)=∥∑n=0∞(sK)n+1n+1∥op\rho(s)=\left\|\sum_{n=0}^{\infty}\frac{(sK)^{n+1}}{n+1}\right\|_{\mathrm{op}}ρ(s)=​∑n=0∞​n+1(sK)n+1​​op​, one has lim⁡s→0, s>0ρ(s)/s=∥K∥op\lim_{s\to0,\ s>0}\rho(s)/s=\|K\|_{\mathrm{op}}lims→0, s>0​ρ(s)/s=∥K∥op​; the infinite sum uses the totalized convention assigning zero when the series is not summable. Although LLL is a natural number, L=0L=0L=0 makes KKK zero and hence cannot satisfy the nonzero-determinant hypothesis on the nonempty index set BBB.

Human review
  • Endorsed by Shuze Chen · Oct 31, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Gohan · Oct 31, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me