Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Definitions for Strong Converse and Stein's Lemma in the Quantum Hypothesis Testing

Definition
QAlgorithms_Papers_Independent_Gap48_P41

by Goku · Oct 22, 2026 · Mathlib 0df444a (Lean v4.33.1)

prove2quantum

The objects the formalized statements of Strong Converse and Stein's Lemma in the Quantum Hypothesis Testing (arXiv:quant-ph/9906090) are stated in. Each was checked against the paper by a blind read-back and a moderator.

  • blockDifference: PDF p.3, (4): the specific comparison operator on n identical copies.
  • nonnegativeProjection: PDF p.3: projection onto NONNEGATIVE eigenvalues, including the whole kernel. Only Hermitian arguments are used; at B=0 the value is I.
  • petzPsi: PDF p.3, (8): Petz power-product logarithmic moment, distinct from the sandwiched moment in Dependency50.P28. On nested-support density pairs the trace is strictly positive for all real s, so Real.log never sees zero.
  • petzPhi: PDF p.3, (7): the accepted compact Legendre transform specialized to [0,1]. On the stated density/support domain its supremum is a finite attained maximum.
  • petzG: PDF pp.5--6, (15): the compact perspective objective, in nats.
  • petzExponent: PDF p.5, (15): the paper's bound, with orders restricted to s in [0,1]. Nonempty compactness and continuity on the stated domain ensure finiteness; Theorem 3 explicitly supplies maximum attainment.
  • crossing: PDF pp.5--6: the crossing parameter is CONSTRUCTED from rho, sigma, r, not supplied as a variable satisfying the desired crossing equation.
  • betaValues: PDF p.2: all feasible type-II values, over arbitrary effects.
  • betaStar: PDF p.2: optimal type-II error. For density inputs and eps in [0,1], the feasible set contains I, is compact, and has values in [0,1]. The Stein theorem asserts attainment explicitly. Off-domain empty/unbounded real infima have Lean's default value zero and are not used by any statement.
  • quantumValues: PDF p.8, (22): extended objective values over ALL feasible density matrices. Feasibility makes the subtracted divergence finite. The first may be +infinity.
  • quantumMinimum: PDF p.8, (22): extended infimum; minimum attainment is part of the result. For r>D, rho itself is a finite feasible candidate; no trial state is discarded merely for having infinite divergence to rho.
  • classicalPsi: PDF pp.7--8: zero-preserving classical power-product log moment. On valid laws with nested supports it is a finite real analytic function.
  • tilt: PDF p.8, Appendix: the constructed exponential tilt. Zero p coordinates stay zero; no full-support hypothesis is imposed. Validity is concluded below.
  • tiltedLogLikelihood: PDF p.8, Appendix: eta(s), the tilted log-likelihood expectation. Zero p mass contributes zero; support inclusion guards positive p coordinates.
  • classicalG: PDF p.7, (21): the classical perspective expression, used only for s>=0.
  • classicalValues: PDF p.7, (20): ALL feasible laws, retaining infinite KL values to p. The KL value to q is finite by the real-rate feasibility constraint.
  • classicalMinimum: PDF p.7, (20): extended variational infimum; theorems assert attainment.
  • classicalDualValues: PDF p.7, (21): values on the noncompact parameter domain. An extended set avoids any bounded-real-supremum default. Attainment is asserted only in the explicit finite-positive-tilt regime from the Appendix.
Definition code
import Mathlib
import Definitions.Def_QAlgorithms_Defs_Information
import Definitions.Def_QAlgorithms_Papers_Independent_Dependency50_P28
import Definitions.Def_QAlgorithms_Papers_Independent_Claimed50_P15
import Definitions.Def_QAlgorithms_Papers_Independent_Information100_P56

/-! Ogawa--Nagaoka, arXiv:quant-ph/9906090v1. Finite-dimensional states,
natural logarithms, arbitrary effects, and support-restricted powers.
The printed zero-error Stein endpoint and strict-curvature calculation are
corrected explicitly below. The ambiguous combined log-Euclidean assertion
on PDF p.8 is deferred in HARD.md. Compilation is pending batch handoff. -/

noncomputable section

namespace QAlgorithms.Papers.Independent.Gap48.P41

open scoped Matrix ComplexOrder BigOperators
open Filter QAlgorithms
open QAlgorithms.Papers.Independent.Information100.P60 (Mat copies)
open QAlgorithms.Papers.Independent.Information100.P71 (supportPower)
open QAlgorithms.Papers.Independent.Information100.P07 (positiveProjection)
open QAlgorithms.Papers.Independent.Dependency50.P28
  (tr SupportLE relativeEntropy scalarPower classicalD typeI accept TestSequence
    scaled logNonnegativeExtended)
open QAlgorithms.Papers.Independent.Claimed50.P15 (legendre)
open QAlgorithms.Papers.Independent.Information100.P56 (IsLaw)

attribute [local instance] Classical.propDecidable Classical.decEq

variable {I : Type} [Fintype I] [DecidableEq I]

/-- PDF p.3, (4): the specific comparison operator on n identical copies. -/
def blockDifference (rho sigma : Mat I) (n : ℕ) (lambda : ℝ) : Mat (Fin n → I) :=
  copies n rho - (Real.exp ((n : ℝ) * lambda) : ℂ) • copies n sigma

/-- PDF p.3: projection onto NONNEGATIVE eigenvalues, including the whole kernel.
Only Hermitian arguments are used; at B=0 the value is I. -/
def nonnegativeProjection (B : Mat I) : Mat I := 1 - positiveProjection (-B)

/-- PDF p.3, (8): Petz power-product logarithmic moment, distinct from the
sandwiched moment in Dependency50.P28. On nested-support density pairs the
trace is strictly positive for all real s, so Real.log never sees zero. -/
def petzPsi (rho sigma : Mat I) (s : ℝ) : ℝ :=
  Real.log (tr (supportPower rho (1 + s) * supportPower sigma (-s)))

/-- PDF p.3, (7): the accepted compact Legendre transform specialized to [0,1].
On the stated density/support domain its supremum is a finite attained maximum. -/
abbrev petzPhi (rho sigma : Mat I) (lambda : ℝ) : ℝ :=
  legendre 0 1 (petzPsi rho sigma) lambda

/-- PDF pp.5--6, (15): the compact perspective objective, in nats. -/
def petzG (rho sigma : Mat I) (r s : ℝ) : ℝ :=
  s / (1 + s) * r - petzPsi rho sigma s / (1 + s)

/-- PDF p.5, (15): the paper's bound, with orders restricted to s in [0,1].
Nonempty compactness and continuity on the stated domain ensure finiteness;
Theorem 3 explicitly supplies maximum attainment. -/
def petzExponent (rho sigma : Mat I) (r : ℝ) : ℝ :=
  sSup (petzG rho sigma r '' Set.Icc 0 1)

/-- PDF pp.5--6: the crossing parameter is CONSTRUCTED from rho, sigma, r,
not supplied as a variable satisfying the desired crossing equation. -/
def crossing (rho sigma : Mat I) (r : ℝ) : ℝ := r - petzExponent rho sigma r

/-- PDF p.2: all feasible type-II values, over arbitrary effects. -/
def betaValues (rho sigma : Mat I) (n : ℕ) (eps : ℝ) : Set ℝ :=
  {v | ∃ A : Mat (Fin n → I), IsEffect A ∧ typeI rho n A ≤ eps ∧
    v = accept sigma n A}

/-- PDF p.2: optimal type-II error. For density inputs and eps in [0,1], the
feasible set contains I, is compact, and has values in [0,1]. The Stein theorem
asserts attainment explicitly. Off-domain empty/unbounded real infima have
Lean's default value zero and are not used by any statement. -/
def betaStar (rho sigma : Mat I) (n : ℕ) (eps : ℝ) : ℝ :=
  sInf (betaValues rho sigma n eps)

/-- PDF p.8, (22): extended objective values over ALL feasible density matrices.
Feasibility makes the subtracted divergence finite. The first may be +infinity. -/
def quantumValues (rho sigma : Mat I) (r : ℝ) : Set EReal :=
  {v | ∃ theta : Mat I, IsDensityMatrix theta ∧ relativeEntropy theta sigma ≤ (r : EReal) ∧
    v = relativeEntropy theta rho + (r : EReal) - relativeEntropy theta sigma}

/-- PDF p.8, (22): extended infimum; minimum attainment is part of the result.
For r>D, rho itself is a finite feasible candidate; no trial state is discarded
merely for having infinite divergence to rho. -/
def quantumMinimum (rho sigma : Mat I) (r : ℝ) : EReal :=
  sInf (quantumValues rho sigma r)

/-- PDF p.3 tie check: equal hypotheses at lambda=0 have comparison operator zero. -/
example (rho : Mat I) (n : ℕ) : blockDifference rho rho n 0 = 0 := by
  simp [blockDifference]

/-- PDF p.3 tie check: the constructed projector includes the zero eigenspace. -/
example : nonnegativeProjection (0 : Mat I) = 1 := by
  simp [nonnegativeProjection,
    QAlgorithms.Papers.Independent.Information100.P07.positiveProjection,
    QAlgorithms.Papers.Independent.Information100.P07.positivePart, cfc_zero,
    QAlgorithms.Papers.Independent.Information100.P07.support,
    QAlgorithms.Papers.Independent.Information100.P07.projection,
    QAlgorithms.Papers.Independent.Information100.P07.operatorRange,
    Submodule.starProjection_bot]

/-- PDF p.3: the one-dimensional identical state has a zero moment at every order. -/
example (s : ℝ) : petzPsi (1 : Mat (Fin 1)) 1 s = 0 := by
  simp [petzPsi, supportPower, tr, cfc_one]

/-- PDF p.5: the compact perspective at its left endpoint is zero in the one-state case. -/
example (r : ℝ) : petzG (1 : Mat (Fin 1)) 1 r 0 = 0 := by
  simp [petzG, petzPsi, supportPower, tr, cfc_one]

/-- PDF p.8: the null density is a feasible variational candidate when its divergence fits. -/
example (rho sigma : Mat I) (hrho : IsDensityMatrix rho) (r : ℝ)
    (hr : relativeEntropy rho sigma ≤ (r : EReal)) :
    relativeEntropy rho rho + (r : EReal) - relativeEntropy rho sigma ∈
      quantumValues rho sigma r := by
  exact ⟨rho, hrho, hr, rfl⟩

/-- PDF p.2: the zero effect has zero success probability, with extended log -infinity. -/
example (rho : Mat I) (n : ℕ) : logNonnegativeExtended (accept rho n 0) = ⊥ := by
  simp [accept, tr, logNonnegativeExtended]

variable {X : Type} [Fintype X]

/-- PDF pp.7--8: zero-preserving classical power-product log moment.
On valid laws with nested supports it is a finite real analytic function. -/
def classicalPsi (p q : X → ℝ) (s : ℝ) : ℝ :=
  Real.log (∑ x, scalarPower (p x) (1 + s) * scalarPower (q x) (-s))

/-- PDF p.8, Appendix: the constructed exponential tilt. Zero p coordinates
stay zero; no full-support hypothesis is imposed. Validity is concluded below. -/
def tilt (p q : X → ℝ) (s : ℝ) : X → ℝ :=
  fun x => Real.exp (-classicalPsi p q s) *
    scalarPower (p x) (1 + s) * scalarPower (q x) (-s)

/-- PDF p.8, Appendix: eta(s), the tilted log-likelihood expectation.
Zero p mass contributes zero; support inclusion guards positive p coordinates. -/
def tiltedLogLikelihood (p q : X → ℝ) (s : ℝ) : ℝ :=
  ∑ x, if 0 < p x then tilt p q s x * Real.log (p x / q x) else 0

/-- PDF p.7, (21): the classical perspective expression, used only for s>=0. -/
def classicalG (p q : X → ℝ) (r s : ℝ) : ℝ :=
  s / (1 + s) * r - classicalPsi p q s / (1 + s)

/-- PDF p.7, (20): ALL feasible laws, retaining infinite KL values to p.
The KL value to q is finite by the real-rate feasibility constraint. -/
def classicalValues (p q : X → ℝ) (r : ℝ) : Set EReal :=
  {v | ∃ phat : X → ℝ, IsLaw phat ∧ classicalD (.finite 1 (by norm_num)) phat q ≤ (r : EReal) ∧
    v = classicalD (.finite 1 (by norm_num)) phat p + (r : EReal) - classicalD (.finite 1 (by norm_num)) phat q}

/-- PDF p.7, (20): extended variational infimum; theorems assert attainment. -/
def classicalMinimum (p q : X → ℝ) (r : ℝ) : EReal :=
  sInf (classicalValues p q r)

/-- PDF p.7, (21): values on the noncompact parameter domain.
An extended set avoids any bounded-real-supremum default. Attainment is asserted
only in the explicit finite-positive-tilt regime from the Appendix. -/
def classicalDualValues (p q : X → ℝ) (r : ℝ) : Set EReal :=
  {v | ∃ s : ℝ, 0 ≤ s ∧ v = (classicalG p q r s : EReal)}

/-- PDF p.7: one-point probability law is allowed. -/
example : IsLaw (fun _ : Fin 1 => (1 : ℝ)) := by simp [IsLaw]

/-- PDF p.8: the one-point moment and its tilted law are normalized at every order. -/
example (s : ℝ) : classicalPsi (fun _ : Fin 1 => 1) (fun _ => 1) s = 0 ∧
    tilt (fun _ : Fin 1 => 1) (fun _ => 1) s = (fun _ => 1) := by
  constructor
  · simp [classicalPsi, scalarPower]
  · funext x
    simp [tilt, classicalPsi, scalarPower]

/-- PDF pp.7--9: equal one-point laws have zero KL divergence. -/
example : classicalD (.finite 1 (by norm_num)) (fun _ : Fin 1 => 1) (fun _ => 1) = 0 := by
  simp [classicalD]

/-- PDF p.8: the one-point log-likelihood expectation is zero. -/
example (s : ℝ) : tiltedLogLikelihood (fun _ : Fin 1 => 1) (fun _ => 1) s = 0 := by
  simp [tiltedLogLikelihood]

/-- PDF p.7: the original law is a feasible variational candidate when its divergence fits. -/
example (p q : X → ℝ) (hp : IsLaw p) (r : ℝ)
    (hr : classicalD (.finite 1 (by norm_num)) p q ≤ (r : EReal)) :
    classicalD (.finite 1 (by norm_num)) p p + (r : EReal) - classicalD (.finite 1 (by norm_num)) p q ∈ classicalValues p q r := by
  exact ⟨p, hp, hr, rfl⟩

end QAlgorithms.Papers.Independent.Gap48.P41
Source
Strong Converse and Stein's Lemma in the Quantum Hypothesis Testing, arXiv:quant-ph/9906090v1, https://arxiv.org/abs/quant-ph/9906090v1
Read-back

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

QAlgorithms.Papers.Independent.Gap48.P41.betaStar

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, natural number nnn, and real number ε\varepsilonε, this defines the real infimum

inf⁡{Re⁡Tr⁡(σ⊗nA)  |  A⪰0,In−A⪰0,Re⁡Tr⁡(ρ⊗n(In−A))≤ε}.\inf\left\{ \operatorname{Re}\operatorname{Tr}(\sigma^{\otimes n}A) \;\middle|\; A\succeq0,\quad I_n-A\succeq0,\quad \operatorname{Re}\operatorname{Tr}\bigl(\rho^{\otimes n}(I_n-A)\bigr)\le\varepsilon \right\}.inf{ReTr(σ⊗nA)​A⪰0,In​−A⪰0,ReTr(ρ⊗n(In​−A))≤ε}.

Here all matrices in the optimization are indexed by functions {0,…,n−1}→I\{0,\ldots,n-1\}\to I{0,…,n−1}→I, InI_nIn​ is their identity, and (X⊗n)x,y=∏i=0n−1Xxi,yi(X^{\otimes n})_{x,y}=\prod_{i=0}^{n-1}X_{x_i,y_i}(X⊗n)x,y​=∏i=0n−1​Xxi​,yi​​. No positivity or normalization of ρ,σ\rho,\sigmaρ,σ, or restriction on ε\varepsilonε, is assumed. An empty feasible set gives the totalized real infimum 000; the definition does not assert attainment. For n=0n=0n=0, both tensor powers are [1][1][1].

QAlgorithms.Papers.Independent.Gap48.P41.betaValues

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, natural number nnn, and real number ε\varepsilonε, this defines the set

{Re⁡Tr⁡(σ⊗nA)  |  A⪰0,In−A⪰0,Re⁡Tr⁡(ρ⊗n(In−A))≤ε},\left\{ \operatorname{Re}\operatorname{Tr}(\sigma^{\otimes n}A) \;\middle|\; A\succeq0,\quad I_n-A\succeq0,\quad \operatorname{Re}\operatorname{Tr}\bigl(\rho^{\otimes n}(I_n-A)\bigr)\le\varepsilon \right\},{ReTr(σ⊗nA)​A⪰0,In​−A⪰0,ReTr(ρ⊗n(In​−A))≤ε},

where AAA and the identity InI_nIn​ are indexed by functions {0,…,n−1}→I\{0,\ldots,n-1\}\to I{0,…,n−1}→I, and (X⊗n)x,y=∏i=0n−1Xxi,yi(X^{\otimes n})_{x,y}=\prod_{i=0}^{n-1}X_{x_i,y_i}(X⊗n)x,y​=∏i=0n−1​Xxi​,yi​​. No conditions are imposed on ρ,σ,ε\rho,\sigma,\varepsilonρ,σ,ε, so the set may be empty. The case n=0n=0n=0 is included, with each tensor power the 1×11\times11×1 matrix [1][1][1].

QAlgorithms.Papers.Independent.Gap48.P41.blockDifference

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, natural number nnn, and real number λ\lambdaλ, this defines the matrix Bn(ρ,σ,λ)B_n(\rho,\sigma,\lambda)Bn​(ρ,σ,λ), indexed by functions {0,…,n−1}→I\{0,\ldots,n-1\}\to I{0,…,n−1}→I, by

Bn(ρ,σ,λ)x,y=∏i=0n−1ρxi,yi−enλ∏i=0n−1σxi,yi.B_n(\rho,\sigma,\lambda)_{x,y} =\prod_{i=0}^{n-1}\rho_{x_i,y_i} -e^{n\lambda}\prod_{i=0}^{n-1}\sigma_{x_i,y_i}.Bn​(ρ,σ,λ)x,y​=i=0∏n−1​ρxi​,yi​​−enλi=0∏n−1​σxi​,yi​​.

No positivity, Hermitian symmetry, or trace normalization is assumed. The definition includes n=0n=0n=0: both empty products equal 111, so the resulting matrix on the singleton set of empty functions is zero.

QAlgorithms.Papers.Independent.Gap48.P41.classicalDualValues

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real rrr, put Pt(u)=utP_t(u)=u^tPt​(u)=ut when u>0u>0u>0 and Pt(u)=0P_t(u)=0Pt​(u)=0 otherwise. The declaration defines the following subset of the extended real line, whose displayed values are embedded real numbers:

{s1+sr−11+slog⁡ ⁣(∑x∈XP1+s(p(x))P−s(q(x))):s∈R, s≥0}.\left\{ \frac{s}{1+s}r- \frac{1}{1+s}\log\!\left(\sum_{x\in X}P_{1+s}(p(x))P_{-s}(q(x))\right): s\in\mathbb R,\ s\ge0 \right\}.{1+ss​r−1+s1​log(x∈X∑​P1+s​(p(x))P−s​(q(x))):s∈R, s≥0}.

There is no upper bound on sss, and no probability-law assumption on p,qp,qp,q. The logarithm is total with log⁡0=0\log0=0log0=0, truncated powers vanish at every nonpositive input, and XXX may be empty.

QAlgorithms.Papers.Independent.Gap48.P41.classicalG

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any reals r,sr,sr,s, let Pt(u)=utP_t(u)=u^tPt​(u)=ut for u>0u>0u>0 and 000 otherwise. The declaration defines

Gp,q(r,s)=s1+sr−11+slog⁡ ⁣(∑x∈XP1+s(p(x))P−s(q(x))).G_{p,q}(r,s)= \frac{s}{1+s}r- \frac{1}{1+s}\log\!\left(\sum_{x\in X}P_{1+s}(p(x))P_{-s}(q(x))\right).Gp,q​(r,s)=1+ss​r−1+s1​log(x∈X∑​P1+s​(p(x))P−s​(q(x))).

The real logarithm has log⁡0=0\log0=0log0=0, and division by zero is zero; consequently this definition includes s=−1s=-1s=−1, where its value is 000. The functions need not be probability laws, and XXX may be empty.

QAlgorithms.Papers.Independent.Gap48.P41.classicalMinimum

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real rrr, define D1(h∥a)D_1(h\Vert a)D1​(h∥a) to be ∑x:h(x)>0h(x)log⁡(h(x)/a(x))\sum_{x:h(x)>0}h(x)\log(h(x)/a(x))∑x:h(x)>0​h(x)log(h(x)/a(x)) when every xxx with h(x)>0h(x)>0h(x)>0 also has a(x)>0a(x)>0a(x)>0, and +∞+\infty+∞ otherwise. The declaration defines the extended real number

inf⁡p^:X→R, p^(x)≥0 ∀x∑xp^(x)=1, D1(p^∥q)≤r[D1(p^∥p)+r−D1(p^∥q)].\inf_{\substack{\widehat p:X\to\mathbb R,\ \widehat p(x)\ge0\ \forall x\\ \sum_x\widehat p(x)=1,\ D_1(\widehat p\Vert q)\le r}} \left[D_1(\widehat p\Vert p)+r-D_1(\widehat p\Vert q)\right].p​:X→R, p​(x)≥0 ∀x∑x​p​(x)=1, D1​(p​∥q)≤r​inf​[D1​(p​∥p)+r−D1​(p​∥q)].

There are no probability-law assumptions on p,qp,qp,q. This is an infimum without an assertion of attainment; if there is no admissible p^\widehat pp​, including when XXX is empty, its value is +∞+\infty+∞.

QAlgorithms.Papers.Independent.Gap48.P41.classicalPsi

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real sss, define Pt(x)=xtP_t(x)=x^tPt​(x)=xt when x>0x>0x>0 and Pt(x)=0P_t(x)=0Pt​(x)=0 when x≤0x\le0x≤0. The declaration defines

Ψp,q(s)=log⁡ ⁣(∑x∈XP1+s(p(x))P−s(q(x))).\Psi_{p,q}(s)=\log\!\left(\sum_{x\in X}P_{1+s}(p(x))P_{-s}(q(x))\right).Ψp,q​(s)=log(x∈X∑​P1+s​(p(x))P−s​(q(x))).

Neither ppp nor qqq is required to be a probability law. The truncated powers are zero at zero even for exponent zero or a negative exponent. The real logarithm is total with log⁡0=0\log0=0log0=0, so the result is 000 if the sum vanishes, including when XXX is empty.

QAlgorithms.Papers.Independent.Gap48.P41.classicalValues

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real rrr, define, for any h,a:X→Rh,a:X\to\mathbb Rh,a:X→R,

D1(h∥a)={∑x∈Xh(x)>0h(x)log⁡ ⁣(h(x)a(x)),if h(x)>0⇒a(x)>0 for every x,+∞,otherwise,D_1(h\Vert a)= \begin{cases} \displaystyle\sum_{\substack{x\in X\\h(x)>0}} h(x)\log\!\left(\frac{h(x)}{a(x)}\right), &\text{if }h(x)>0\Rightarrow a(x)>0\text{ for every }x,\\ +\infty,&\text{otherwise}, \end{cases}D1​(h∥a)=⎩⎨⎧​x∈Xh(x)>0​∑​h(x)log(a(x)h(x)​),+∞,​if h(x)>0⇒a(x)>0 for every x,otherwise,​

with the finite sum regarded as an extended real number. The declaration defines the subset of the extended real line

{D1(p^∥p)+r−D1(p^∥q):p^:X→R, p^(x)≥0 ∀x, ∑xp^(x)=1, D1(p^∥q)≤r}.\left\{ D_1(\widehat p\Vert p)+r-D_1(\widehat p\Vert q): \widehat p:X\to\mathbb R,\ \widehat p(x)\ge0\ \forall x,\ \sum_x\widehat p(x)=1,\ D_1(\widehat p\Vert q)\le r \right\}.{D1​(p​∥p)+r−D1​(p​∥q):p​:X→R, p​(x)≥0 ∀x, x∑​p​(x)=1, D1​(p​∥q)≤r}.

The functions p,qp,qp,q themselves have no positivity or normalization requirements. The set may be empty; in particular it is empty when XXX is empty, because no p^\widehat pp​ then has sum 111.

QAlgorithms.Papers.Independent.Gap48.P41.crossing

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number rrr, this defines

r−sup⁡0≤s≤1{s1+s r−log⁡ ⁣(Re⁡Tr⁡(H1+s(ρ)H−s(σ)))1+s}.r-\sup_{0\le s\le1} \left\{ \frac{s}{1+s}\,r- \frac{\log\!\left(\operatorname{Re}\operatorname{Tr} \bigl(H_{1+s}(\rho)H_{-s}(\sigma)\bigr)\right)}{1+s} \right\}.r−0≤s≤1sup​{1+ss​r−1+slog(ReTr(H1+s​(ρ)H−s​(σ)))​}.

Here HtH_tHt​ is real continuous functional calculus for the function equal to 000 at zero and to the total real power xtx^txt elsewhere; the real logarithm is 000 at zero and log⁡∣x∣\log|x|log∣x∣ at negative arguments. The real supremum has totalized value 000 if its set is unbounded above. Neither matrix is required to be positive or normalized, and rrr is unrestricted.

QAlgorithms.Papers.Independent.Gap48.P41.nonnegativeProjection

For any finite index type III with decidable equality and any complex matrix BBB indexed by III, this defines

II−Pran⁡(f(−B)),f(t)=max⁡(t,0),I_I-P_{\operatorname{ran}(f(-B))}, \qquad f(t)=\max(t,0),II​−Pran(f(−B))​,f(t)=max(t,0),

where f(−B)f(-B)f(−B) is formed by real continuous functional calculus, ran⁡\operatorname{ran}ran denotes the range of the corresponding linear operator on CI\mathbb C^ICI, and PSP_SPS​ is the orthogonal projection onto a subspace SSS. No Hermitian hypothesis on BBB or nonemptiness hypothesis on III is imposed.

QAlgorithms.Papers.Independent.Gap48.P41.petzExponent

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number rrr, this defines

Eρ,σ(r)=sup⁡0≤s≤1{s1+s r−log⁡ ⁣(Re⁡Tr⁡(H1+s(ρ)H−s(σ)))1+s},E_{\rho,\sigma}(r) =\sup_{0\le s\le1} \left\{ \frac{s}{1+s}\,r- \frac{\log\!\left(\operatorname{Re}\operatorname{Tr} \bigl(H_{1+s}(\rho)H_{-s}(\sigma)\bigr)\right)}{1+s} \right\},Eρ,σ​(r)=0≤s≤1sup​{1+ss​r−1+slog(ReTr(H1+s​(ρ)H−s​(σ)))​},

where Ht(X)H_t(X)Ht​(X) is obtained by real continuous functional calculus from the function equal to 000 at zero and to the total real power xtx^txt elsewhere. The logarithm is total, with log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. No positivity or normalization of either matrix is assumed. Both endpoints are included, and this is a real supremum, taking the totalized value 000 if the set is unbounded above.

QAlgorithms.Papers.Independent.Gap48.P41.petzG

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real numbers r,sr,sr,s, this defines

Gρ,σ(r,s)=s1+s r−log⁡ ⁣(Re⁡Tr⁡(H1+s(ρ)H−s(σ)))1+s,G_{\rho,\sigma}(r,s) =\frac{s}{1+s}\,r- \frac{\log\!\left(\operatorname{Re}\operatorname{Tr} \bigl(H_{1+s}(\rho)H_{-s}(\sigma)\bigr)\right)}{1+s},Gρ,σ​(r,s)=1+ss​r−1+slog(ReTr(H1+s​(ρ)H−s​(σ)))​,

where HtH_tHt​ applies real continuous functional calculus to the function equal to 000 at zero and to the total real power xtx^txt elsewhere. The real logarithm takes value 000 at zero and log⁡∣x∣\log|x|log∣x∣ at negative arguments. No matrix conditions or restrictions on r,sr,sr,s are imposed. Division is totalized by a/0=0a/0=0a/0=0, so Gρ,σ(r,−1)=0G_{\rho,\sigma}(r,-1)=0Gρ,σ​(r,−1)=0.

QAlgorithms.Papers.Independent.Gap48.P41.petzPhi

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number λ\lambdaλ, this abbreviates the real supremum

sup⁡0≤s≤1{λs−log⁡ ⁣(Re⁡Tr⁡(H1+s(ρ)H−s(σ)))},\sup_{0\le s\le1} \left\{ \lambda s- \log\!\left(\operatorname{Re}\operatorname{Tr} \bigl(H_{1+s}(\rho)H_{-s}(\sigma)\bigr)\right) \right\},0≤s≤1sup​{λs−log(ReTr(H1+s​(ρ)H−s​(σ)))},

where Ht(X)H_t(X)Ht​(X) is real continuous functional calculus applied to the function taking value 000 at x=0x=0x=0 and the total real power xtx^txt elsewhere. The logarithm is total, with log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for x<0x<0x<0. No matrix positivity or normalization is assumed; both endpoints of the interval are included. The supremum is real-valued, with the totalized value 000 if the displayed set is unbounded above.

QAlgorithms.Papers.Independent.Gap48.P41.petzPsi

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number sss, this defines

ψρ,σ(s)=log⁡ ⁣(Re⁡Tr⁡(H1+s(ρ)H−s(σ))),Ht(X)=ft(X),ft(x)={0,x=0,xt,x≠0,\psi_{\rho,\sigma}(s) =\log\!\left(\operatorname{Re}\operatorname{Tr} \bigl(H_{1+s}(\rho)H_{-s}(\sigma)\bigr)\right), \qquad H_t(X)=f_t(X),\quad f_t(x)= \begin{cases} 0,&x=0,\\ x^t,&x\ne0, \end{cases}ψρ,σ​(s)=log(ReTr(H1+s​(ρ)H−s​(σ))),Ht​(X)=ft​(X),ft​(x)={0,xt,​x=0,x=0,​

where matrix functions use real continuous functional calculus and xtx^txt uses the total real-power function. In particular, f0(0)=0f_0(0)=0f0​(0)=0. There are no positivity, normalization, support, or restrictions on sss. The logarithm is the total real logarithm: it is 000 at zero and equals log⁡∣x∣\log|x|log∣x∣ for negative arguments.

QAlgorithms.Papers.Independent.Gap48.P41.quantumMinimum

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number rrr, this defines

inf⁡θ⪰0, Tr⁡θ=1D(θ∥σ)≤r(D(θ∥ρ)+r−D(θ∥σ))\inf_{\substack{\theta\succeq0,\ \operatorname{Tr}\theta=1\\ D(\theta\Vert\sigma)\le r}} \bigl(D(\theta\Vert\rho)+r-D(\theta\Vert\sigma)\bigr)θ⪰0, Trθ=1D(θ∥σ)≤r​inf​(D(θ∥ρ)+r−D(θ∥σ))

in the extended reals, where

D(X∥Y)={Re⁡Tr⁡(X(log⁡X−log⁡Y))Re⁡Tr⁡X,ker⁡Y⊆ker⁡X,+∞,otherwise.D(X\Vert Y)= \begin{cases} \displaystyle \frac{\operatorname{Re}\operatorname{Tr} \bigl(X(\log X-\log Y)\bigr)} {\operatorname{Re}\operatorname{Tr}X}, &\ker Y\subseteq\ker X,\\[6pt] +\infty,&\text{otherwise}. \end{cases}D(X∥Y)=⎩⎨⎧​ReTrXReTr(X(logX−logY))​,+∞,​kerY⊆kerX,otherwise.​

Matrix logarithms use real continuous functional calculus for the total real logarithm, with log⁡0=0\log0=0log0=0, and division and extended-real arithmetic retain their totalized meanings. Neither ρ\rhoρ nor σ\sigmaσ is assumed positive or normalized. The infimum of an empty feasible set is +∞+\infty+∞, including when III is empty, and attainment is not asserted.

QAlgorithms.Papers.Independent.Gap48.P41.quantumValues

For any finite index type III with decidable equality, complex matrices ρ,σ\rho,\sigmaρ,σ indexed by III, and real number rrr, this defines the subset of the extended reals

{D(θ∥ρ)+r−D(θ∥σ)  |  θ⪰0,Tr⁡θ=1,D(θ∥σ)≤r},\left\{ D(\theta\Vert\rho)+r-D(\theta\Vert\sigma) \;\middle|\; \theta\succeq0,\quad \operatorname{Tr}\theta=1,\quad D(\theta\Vert\sigma)\le r \right\},{D(θ∥ρ)+r−D(θ∥σ)∣θ⪰0,Trθ=1,D(θ∥σ)≤r},

where

D(X∥Y)={Re⁡Tr⁡(X(log⁡X−log⁡Y))Re⁡Tr⁡X,ker⁡Y⊆ker⁡X,+∞,otherwise.D(X\Vert Y)= \begin{cases} \displaystyle \frac{\operatorname{Re}\operatorname{Tr} \bigl(X(\log X-\log Y)\bigr)} {\operatorname{Re}\operatorname{Tr}X}, &\ker Y\subseteq\ker X,\\[6pt] +\infty,&\text{otherwise}. \end{cases}D(X∥Y)=⎩⎨⎧​ReTrXReTr(X(logX−logY))​,+∞,​kerY⊆kerX,otherwise.​

The matrix logarithms use real continuous functional calculus with the total real logarithm, including log⁡0=0\log0=0log0=0; real division is also totalized. The expression defining each value uses extended-real arithmetic. No density-matrix assumptions are imposed on ρ\rhoρ or σ\sigmaσ. The set can be empty; in particular, it is empty when III is empty because no matrix then has trace 111.

QAlgorithms.Papers.Independent.Gap48.P41.tilt

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real sss, let Pt(u)=utP_t(u)=u^tPt​(u)=ut for u>0u>0u>0 and Pt(u)=0P_t(u)=0Pt​(u)=0 otherwise, and define

Z(s)=∑y∈XP1+s(p(y))P−s(q(y)).Z(s)=\sum_{y\in X}P_{1+s}(p(y))P_{-s}(q(y)).Z(s)=y∈X∑​P1+s​(p(y))P−s​(q(y)).

The declaration defines the function

p~s(x)=e−log⁡Z(s)P1+s(p(x))P−s(q(x))(x∈X).\widetilde p_s(x)=e^{-\log Z(s)}P_{1+s}(p(x))P_{-s}(q(x)) \quad(x\in X).p​s​(x)=e−logZ(s)P1+s​(p(x))P−s​(q(x))(x∈X).

The real logarithm satisfies log⁡0=0\log0=0log0=0, so the exponential factor equals 111 when Z(s)=0Z(s)=0Z(s)=0; in that case every value of this function is zero. The functions p,qp,qp,q are unrestricted, and an empty XXX is allowed.

QAlgorithms.Papers.Independent.Gap48.P41.tiltedLogLikelihood

For a finite type XXX, arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, and any real sss, write Pt(u)=utP_t(u)=u^tPt​(u)=ut when u>0u>0u>0 and 000 otherwise, set Z(s)=∑y∈XP1+s(p(y))P−s(q(y))Z(s)=\sum_{y\in X}P_{1+s}(p(y))P_{-s}(q(y))Z(s)=∑y∈X​P1+s​(p(y))P−s​(q(y)), and set p~s(x)=e−log⁡Z(s)P1+s(p(x))P−s(q(x))\widetilde p_s(x)=e^{-\log Z(s)}P_{1+s}(p(x))P_{-s}(q(x))p​s​(x)=e−logZ(s)P1+s​(p(x))P−s​(q(x)). The declaration defines

Lp,q(s)=∑x∈Xp(x)>0p~s(x)log⁡ ⁣(p(x)q(x)).L_{p,q}(s)= \sum_{\substack{x\in X\\p(x)>0}} \widetilde p_s(x)\log\!\left(\frac{p(x)}{q(x)}\right).Lp,q​(s)=x∈Xp(x)>0​∑​p​s​(x)log(q(x)p(x)​).

No probability-law or support assumptions are imposed. Division and the real logarithm are total: division by zero gives zero and log⁡0=0\log0=0log0=0. Terms with p(x)≤0p(x)\le0p(x)≤0 are assigned zero directly, and the empty sum is zero.

Human review
  • Endorsed by Shuze Chen · Oct 22, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Goku · Oct 22, 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