Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 9 low-rate branch (A9), corrected (A20)–(A21)

Open
QAlgorithms.Papers.Independent.Intakenext100.P20.low_rate_variational

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

prove2quantum

Let ρ\rhoρ be a normalized density matrix on a complex space of dimension d>0d>0d>0, with decreasing eigenvalues a1≥⋯≥ad≥0a_1\ge\cdots\ge a_d\ge0a1​≥⋯≥ad​≥0. Suppose the top eigenvalue has multiplicity 1≤k<d1\le k<d1≤k<d, so a1=⋯=ak>ak+1a_1=\cdots=a_k>a_{k+1}a1​=⋯=ak​>ak+1​. For 0≤R≤log⁡k0\le R\le\log k0≤R≤logk set v=−log⁡a1−Rv=-\log a_1-Rv=−loga1​−R and ψ(s)=log⁡Tr⁡ρs\psi(s)=\log\operatorname{Tr}\rho^sψ(s)=logTrρs. Write HHH for Shannon or von Neumann entropy in nats and DDD for classical or quantum relative entropy, with value +∞+\infty+∞ on support violation. Then

min⁡H(b)=RD(b∥a)=min⁡H(σ)=RD(σ∥ρ)=min⁡H(b)≤RD(b∥a)=min⁡H(σ)≤RD(σ∥ρ)=v,\min_{H(b)=R}D(b\Vert a)=\min_{H(\sigma)=R}D(\sigma\Vert\rho)=\min_{H(b)\le R}D(b\Vert a)=\min_{H(\sigma)\le R}D(\sigma\Vert\rho)=v,H(b)=Rmin​D(b∥a)=H(σ)=Rmin​D(σ∥ρ)=H(b)≤Rmin​D(b∥a)=H(σ)≤Rmin​D(σ∥ρ)=v, min⁡H(b)≥RD(b∥a)=min⁡H(σ)≥RD(σ∥ρ)=max⁡0<s≤1(1−s)R−ψ(s)s=0,\min_{H(b)\ge R}D(b\Vert a)=\min_{H(\sigma)\ge R}D(\sigma\Vert\rho)=\max_{0<s\le1}\frac{(1-s)R-\psi(s)}s=0,H(b)≥Rmin​D(b∥a)=H(σ)≥Rmin​D(σ∥ρ)=0<s≤1max​s(1−s)R−ψ(s)​=0, sup⁡s≥1(1−s)R−ψ(s)s=v.\sup_{s\ge1}\frac{(1-s)R-\psi(s)}s=v.s≥1sup​s(1−s)R−ψ(s)​=v.

Here bbb ranges over all probability vectors of length ddd, and σ\sigmaσ over all normalized density matrices on the same space. Every minimum and the displayed maximum is attained; the final supremum need not be. This gives the low-rate variational formulas without inverse-tilt or full-rank hypotheses.

Source correction. Equation (A20) prints log⁡a1−R\log a_1-Rloga1​−R. Its missing minus is fixed by (A9) and the attaining-state calculation on PDF p.9; the correction was independently checked in moderation round 1; a fresh no-edit round is required after this round’s repairs.

Preamble
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P20
Formal statement
noncomputable section

open scoped Matrix ENNReal ComplexOrder Topology

namespace QAlgorithms.Papers.Independent.Intakenext100.P20

open QAlgorithms.Papers.Independent.Information100.P08 (IsPure)

open QAlgorithms.Papers.Independent.Information100.P47 (IsProbability)

open QAlgorithms.Papers.Independent.Information100.P53 (SpectralData)

open QAlgorithms.Papers.Independent.Gap48.P19 (shannonNats)

open QAlgorithms.Papers.Independent.Dependency50.P28 (classicalD)

attribute [local instance] Classical.propDecidable

/-- Lemma 9, PDF pp.8-9, low-rate (A9), (A18)-(A21).
Correction: (A20) prints log a1-R; (A9) and the p.9 attaining-state calculation
uniquely give -log a1-R. No full-rank assumption or inverse tilt is used.
Every minimum is stated as an attained least value; the s>=1 expression is a supremum. -/
theorem low_rate_variational {d : ℕ} (hd : 0 < d)
    (rho : {M : Matrix (Fin d) (Fin d) ℂ // IsDensityMatrix M})
    (spectral : SpectralData rho.val) (k : ℕ) (hk : 0 < k) (hkd : k < d)
    (htop : ∀ i : Fin d, i.val < k → spectral.value i = spectral.value ⟨0, hd⟩)
    (hgap : spectral.value ⟨k, hkd⟩ < spectral.value ⟨0, hd⟩)
    (R : ℝ) (hR : 0 ≤ R) (hRk : R ≤ Real.log (k : ℝ)) :
    let a := spectral.value
    let value := ENNReal.ofReal (-Real.log (a ⟨0, hd⟩) - R)
    IsLeast (classicalDValues a (fun h => h = R)) value ∧
    IsLeast (quantumDValues rho (fun h => h = R)) value ∧
    IsLeast (classicalDValues a (fun h => h ≤ R)) value ∧
    IsLeast (quantumDValues rho (fun h => h ≤ R)) value ∧
    IsLeast (classicalDValues a (fun h => R ≤ h)) 0 ∧
    IsLeast (quantumDValues rho (fun h => R ≤ h)) 0 ∧
    IsGreatest {v : ℝ | ∃ s : ℝ, 0 < s ∧ s ≤ 1 ∧
      v = ((1 - s) * R - logPowerTrace rho.val s) / s} 0 ∧
    (⨆ s : {s : ℝ // 1 ≤ s},
      ENNReal.ofReal (((1 - s.val) * R - logPowerTrace rho.val s.val) / s.val)) = value := by sorry

end QAlgorithms.Papers.Independent.Intakenext100.P20

end
Source
Exponents of quantum fixed-length pure state source coding, arXiv:quant-ph/0202002v6, PDF p. 8, Lemma 9 low-rate branch (A9), corrected (A20)–(A21)
Read-back

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

Let d>0d>0d>0 be a natural number, let ρ∈Cd×d\rho\in\mathbb C^{d\times d}ρ∈Cd×d be positive semidefinite with trace 111, and suppose supplied spectral data consist of real numbers a0≥⋯≥ad−1a_0\ge\cdots\ge a_{d-1}a0​≥⋯≥ad−1​ and an orthonormal basis v0,…,vd−1v_0,\ldots,v_{d-1}v0​,…,vd−1​ of Cd\mathbb C^dCd satisfying ρ=∑iaivivi†\rho=\sum_i a_i v_iv_i^\daggerρ=∑i​ai​vi​vi†​. Let kkk be a natural number with 0<k<d0<k<d0<k<d, assume ai=a0a_i=a_0ai​=a0​ whenever i<ki<ki<k and ak<a0a_k<a_0ak​<a0​, and let R∈RR\in\mathbb RR∈R satisfy 0≤R≤log⁡k0\le R\le\log k0≤R≤logk, using natural logarithms. For a probability vector bbb, meaning bi≥0b_i\ge0bi​≥0 and ∑ibi=1\sum_i b_i=1∑i​bi​=1, put H(b)=−∑ibilog⁡biH(b)=-\sum_i b_i\log b_iH(b)=−∑i​bi​logbi​, and define

Dc(b∥a)={max⁡ ⁣{0,∑i:bi>0bilog⁡biai},ai>0 whenever bi>0,+∞,otherwise.D_{\mathrm c}(b\Vert a)= \begin{cases} \displaystyle\max\!\left\{0,\sum_{i:b_i>0}b_i\log\frac{b_i}{a_i}\right\}, &a_i>0\text{ whenever }b_i>0,\\ +\infty,&\text{otherwise}. \end{cases}Dc​(b∥a)=⎩⎨⎧​max{0,i:bi​>0∑​bi​logai​bi​​},+∞,​ai​>0 whenever bi​>0,otherwise.​

For every density matrix σ\sigmaσ, put H(σ)=−Re⁡Tr⁡f(σ)H(\sigma)=-\operatorname{Re}\operatorname{Tr}f(\sigma)H(σ)=−ReTrf(σ), where f(x)=xlog⁡xf(x)=x\log xf(x)=xlogx, and define

Dq(σ∥ρ)={max⁡{0,Re⁡Tr⁡[f(σ)−σlog⁡ρ]},ker⁡ρ⊆ker⁡σ,+∞,otherwise,D_{\mathrm q}(\sigma\Vert\rho)= \begin{cases} \max\{0,\operatorname{Re}\operatorname{Tr}[f(\sigma)-\sigma\log\rho]\}, &\ker\rho\subseteq\ker\sigma,\\ +\infty,&\text{otherwise}, \end{cases}Dq​(σ∥ρ)={max{0,ReTr[f(σ)−σlogρ]},+∞,​kerρ⊆kerσ,otherwise,​

with the matrix functions defined by functional calculus and log⁡0=0\log0=0log0=0. Let V=max⁡{0,−log⁡a0−R}V=\max\{0,-\log a_0-R\}V=max{0,−loga0​−R}, viewed in [0,∞][0,\infty][0,∞]. Then VVV is the least element of each of the four sets of divergence values obtained from Dc(b∥a)D_{\mathrm c}(b\Vert a)Dc​(b∥a) with H(b)=RH(b)=RH(b)=R, from Dq(σ∥ρ)D_{\mathrm q}(\sigma\Vert\rho)Dq​(σ∥ρ) with H(σ)=RH(\sigma)=RH(σ)=R, and from the same two divergences with the corresponding entropy condition replaced by entropy at most RRR. The least element is 000 for each of the two sets with entropy at least RRR; these least-element assertions include attainment. Finally, writing L(s)=log⁡Re⁡Tr⁡(ρs)L(s)=\log\operatorname{Re}\operatorname{Tr}(\rho^s)L(s)=logReTr(ρs), with powers defined by functional calculus, 000 is the greatest element of

{(1−s)R−L(s)s:0<s≤1},\left\{\frac{(1-s)R-L(s)}s:0<s\le1\right\},{s(1−s)R−L(s)​:0<s≤1},

and

sup⁡s≥1max⁡ ⁣{0,(1−s)R−L(s)s}=Vin [0,∞].\sup_{s\ge1}\max\!\left\{0,\frac{(1-s)R-L(s)}s\right\}=V \quad\text{in }[0,\infty].s≥1sup​max{0,s(1−s)R−L(s)​}=Vin [0,∞].

The assumptions allow k=1k=1k=1, in which case R=0R=0R=0.

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