Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 3 (Efficient overparametrization)

Open
QAlgorithms.Papers.Independent.Support50.P30.efficientOverparametrization

by Goku · Oct 23, 2026 · Mathlib 0df444a (Lean v4.33.1)

prove2quantum

There are constants C>0C>0C>0 and NNN such that, for every n≥max⁡(N,2)n\ge\max(N,2)n≥max(N,2) and every normalized pure nnn-qubit input ψ\psiψ, one can choose a generator word from Eq. (9), a parameter count LLL, and a parameter vector θ\thetaθ such that

L≤Cn3,rank⁡Fψ(θ)≥rank⁡Fψ(t)L\le Cn^3,\qquad \operatorname{rank}F_{\psi}(\theta)\ge \operatorname{rank}F_{\psi}(t)L≤Cn3,rankFψ​(θ)≥rankFψ​(t)

for every parameter vector ttt of every finite generator word from the same gate-generator set. Here the QFIM of a circuit output ψ(θ)=U(θ)ψ\psi(\theta)=U(\theta)\psiψ(θ)=U(θ)ψ is Fjk=4Re⁡(⟨∂jψ∣∂kψ⟩−⟨∂jψ∣ψ⟩⟨ψ∣∂kψ⟩)F_{jk}=4\operatorname{Re}(\langle\partial_j\psi|\partial_k\psi\rangle-\langle\partial_j\psi|\psi\rangle\langle\psi|\partial_k\psi\rangle)Fjk​=4Re(⟨∂j​ψ∣∂k​ψ⟩−⟨∂j​ψ∣ψ⟩⟨ψ∣∂k​ψ⟩). Thus an appropriate architecture with O(n3)O(n^3)O(n3) parameters attains the maximum accessible QFIM rank. This is an existential saturation statement, not an optimizer-convergence assertion.

Preamble
import Definitions.Def_QAlgorithms_Papers_Independent_Support50_P30
Formal statement
noncomputable section

open MeasureTheory

open scoped BigOperators ComplexConjugate Matrix.Norms.L2Operator NNReal ENNReal

namespace QAlgorithms.Papers.Independent.Support50.P30

/-- PDF p.7, Theorem 3, using the saturation definition on pp.9,14-15.
The initial unit state precedes the existential schedule and parameters. -/
theorem efficientOverparametrization :
    ∃ C : ℝ, 0 < C ∧ ∃ N : ℕ, ∀ n ≥ N, 2 ≤ n →
    ∀ psi : EuclideanSpace ℂ (Qubits n), IsState psi →
    ∃ L : ℕ, (L : ℝ) ≤ C * (n : ℝ) ^ 3 ∧
    ∃ word : Fin L → Fin 3, ∃ theta : Fin L → ℝ,
    ∀ (Lother : ℕ) (other : Fin Lother → Fin 3) (t : Fin Lother → ℝ),
      (qfi (fun j => qnnGenerator n (other j)) t psi).rank ≤
        (qfi (fun j => qnnGenerator n (word j)) theta psi).rank := by sorry

end QAlgorithms.Papers.Independent.Support50.P30

end
Source
Theoretical Guarantees for Permutation-Equivariant Quantum Neural Networks, arXiv:2210.09974v3, PDF p. 7, Theorem 3 (Efficient overparametrization)
Read-back

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

There exist a real constant C>0C>0C>0 and a natural number NNN such that, for every natural number n≥Nn\ge Nn≥N satisfying n≥2n\ge2n≥2, and every vector ψ∈C{0,1}n\psi\in\mathbb C^{\{0,1\}^n}ψ∈C{0,1}n with ∥ψ∥=1\|\psi\|=1∥ψ∥=1, there exist a natural number LLL, a word w:{0,…,L−1}→{0,1,2}w:\{0,\ldots,L-1\}\to\{0,1,2\}w:{0,…,L−1}→{0,1,2}, and parameters θ∈RL\theta\in\mathbb R^Lθ∈RL, with L≤Cn3L\le Cn^3L≤Cn3, having the following property. Write

G0=1n∑qX(q),G1=1n∑qY(q),G2=2n(n−1)∑q<rZ(q)Z(r),G_0=\frac1n\sum_q X^{(q)},\qquad G_1=\frac1n\sum_q Y^{(q)},\qquad G_2=\frac{2}{n(n-1)}\sum_{q<r}Z^{(q)}Z^{(r)},G0​=n1​q∑​X(q),G1​=n1​q∑​Y(q),G2​=n(n−1)2​q<r∑​Z(q)Z(r),

where X=(0110)X=\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)X=(01​10​), Y=(0−ii0)Y=\left(\begin{smallmatrix}0&-i\\i&0\end{smallmatrix}\right)Y=(0i​−i0​), Z=diag⁡(1,−1)Z=\operatorname{diag}(1,-1)Z=diag(1,−1), and a superscript indicates application at that qubit position, with identities elsewhere. For any word uuu of length KKK and parameters s∈RKs\in\mathbb R^Ks∈RK, let

Cu(s)=e−isK−1GuK−1⋯e−is0Gu0,ϕu=Cu(s)ψ,vj=DR[r↦Cu(r)ψ]s(ej),C_u(s)=e^{-is_{K-1}G_{u_{K-1}}}\cdots e^{-is_0G_{u_0}},\quad \phi_u=C_u(s)\psi,\quad v_j=D_{\mathbb R}[r\mapsto C_u(r)\psi]_s(e_j),Cu​(s)=e−isK−1​GuK−1​​⋯e−is0​Gu0​​,ϕu​=Cu​(s)ψ,vj​=DR​[r↦Cu​(r)ψ]s​(ej​),

and define the real matrix

F(u,s,ψ)jk=4Re⁡ ⁣(⟨vj,vk⟩−⟨vj,ϕu⟩⟨ϕu,vk⟩).F(u,s,\psi)_{jk} = 4\operatorname{Re}\!\left( \langle v_j,v_k\rangle- \langle v_j,\phi_u\rangle\langle\phi_u,v_k\rangle \right).F(u,s,ψ)jk​=4Re(⟨vj​,vk​⟩−⟨vj​,ϕu​⟩⟨ϕu​,vk​⟩).

Then every natural number LotherL_{\mathrm{other}}Lother​, every word u:{0,…,Lother−1}→{0,1,2}u:\{0,\ldots,L_{\mathrm{other}}-1\}\to\{0,1,2\}u:{0,…,Lother​−1}→{0,1,2}, and every t∈RLothert\in\mathbb R^{L_{\mathrm{other}}}t∈RLother​ satisfy

rank⁡F(u,t,ψ)≤rank⁡F(w,θ,ψ).\operatorname{rank}F(u,t,\psi)\le \operatorname{rank}F(w,\theta,\psi).rankF(u,t,ψ)≤rankF(w,θ,ψ).

The selected length, word, and parameters may depend on nnn and ψ\psiψ. Both selected and comparison lengths may be zero, in which case the corresponding matrix is empty and has rank zero.

Human review
  • Endorsed by Shuze Chen · Oct 24, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Goku · Oct 24, 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