Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Definitions for Min- and Max- Relative Entropies and a New Entanglement Monotone

Definition
QAlgorithms_Papers_Independent_Support50_P21

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

prove2quantum

The objects the formalized statements of Min- and Max- Relative Entropies and a New Entanglement Monotone (arXiv:0803.2770) are stated in. Each was checked against the paper by a blind read-back and a moderator.

  • Le: PDF pp.2–3: Loewner order, expressed by positivity of the difference.
  • dMin: PDF p.3, (9): support overlap, including +infinity at zero overlap.
  • unnormalizedPetz: PDF p.3, (11): unnormalized Petz expression. No division by Tr rho. Only orders in (0,1) are used; zero overlap gives +infinity.
  • weakPositive: PDF pp.2–3: weak positive spectral projector includes the zero eigenspace.
  • strictPositive: PDF pp.2,7: strict positive spectral projector excludes the zero eigenspace.
  • largestEigenvalue: PDF p.3, (7): largest eigenvalue. Uses occur on nonempty finite spectra.
  • thresholdSet: PDF p.3, (8): the positive scalar parameters with vanishing positive-part trace.
  • unnormalizedRelativeBits: PDF p.5, (27): Umegaki entropy in bits, with its support guard. The spectral log on the kernel is harmless only in the guarded trace.
  • traceSmoothingBall: PDF p.7, Definition 4: full trace-norm ball with the source trace cap.
  • traceSmoothMax: PDF p.7, Definition 4: infimum smoothing in the FIRST argument only.
  • traceSmoothMin: PDF p.7, Definition 4: supremum smoothing in the FIRST argument only.
  • hMin: PDF p.3, (12): unconditional min entropy.
  • hMax: PDF p.3, (13): the paper's rank-based max entropy.
  • conditionalMin: PDF p.3, (14): reference-state conditional min entropy. Printed first line uses rho_B; corrected to sigma_B using the preceding substitution and the second line of (14). Independent moderation required.
  • conditionalMax: PDF p.3, (15): reference-state conditional rank-based max entropy.
  • mutualMin: PDF p.3, Definition 3: min mutual information with fixed original marginals.
  • mutualMax: PDF p.3, Definition 3: max mutual information.
  • smoothHMin: PDF pp.4,7: smooth unconditional min entropy with the paper's trace-norm ball.
  • smoothHMax: PDF pp.4,7: smooth unconditional rank-based max entropy.
  • smoothConditionalMin: PDF pp.4,7: smooth reference-state conditional min entropy.
  • smoothConditionalMax: PDF pp.4,7: smooth reference-state conditional rank-based max entropy.
  • smoothMutualMin: PDF pp.4,7: smoothing holds the ORIGINAL product of marginals fixed.
  • smoothMutualMax: PDF pp.4,7: smooth max mutual information, with fixed comparison operator.
  • chernoffCoefficient: PDF p.6: support-limit endpoint convention in the Chernoff coefficient. At s=0 the zero spectral power is the support projector, likewise at s=1.
  • chernoff: PDF p.6: Chernoff distance; the theorem also asserts attainment.
  • PositiveTensorBasis: PDF p.7, Definition 5: tensor register at positive length k+1.
  • spectralMass: PDF p.7, (40)–(41): positive-part trace of the difference operator. Every length is k+1, and the threshold is 2^((k+1)*gamma).
  • spectralSup: PDF p.7, (40): extended infimum of the limsup-zero thresholds.
  • spectralInf: PDF p.7, (41): extended supremum of the liminf-one thresholds.
  • maxAsymptotic: PDF p.8, (53): inner asymptotic limsup, taken before eps tends to zero.
  • minAsymptotic: PDF p.9, (67): inner asymptotic liminf, taken before eps tends to zero.
  • eMax: PDF p.6, (35): optimization over every separable density matrix.
  • eRelative: PDF p.7, Lemma 13: relative entropy of entanglement.
  • bundledMatrix: PDF p.6, (35): convert the accepted dependent two-party basis to a product basis.
  • bundledEMax: PDF pp.6,12: Emax on accepted bundled states. Theorem 1 asserts finiteness on every density, so toReal here is never an infinity fallback in the LOCC claim.
Definition code
import Mathlib
import Definitions.Def_QAlgorithms_Defs_MatrixDistances
import Definitions.Def_QAlgorithms_Defs_BipartiteSeparability
import Definitions.Def_QAlgorithms_Papers_Independent_Claimed50_P33
import Definitions.Def_QAlgorithms_Papers_Independent_Information100_P07
import Definitions.Def_QAlgorithms_Papers_Independent_Information100_P71
import Definitions.Def_QAlgorithms_Papers_Independent_Dependency50_P35
import Definitions.Def_QAlgorithms_Papers_Independent_Information100_P67

/-! Datta, arXiv:0803.2770v3. All logarithms are in bits. New definitions
are paper-local; accepted support projections, powers, channels, separability,
and adaptive finite LOCC are imported without modification.
Compilation is pending the separate mechanical handoff. -/

noncomputable section

namespace QAlgorithms.Papers.Independent.Support50.P21

open QAlgorithms QAlgorithms.Entanglement Filter
open scoped Matrix ComplexConjugate ComplexOrder

attribute [local instance] Classical.propDecidable Classical.decEq

-- Use the accepted declarations directly; no paper-local aliases hide dimension binders.
open QAlgorithms.Papers.Independent.Claimed50.P33
  (FiniteState IsAverageLocalMonotone IsAdaptiveLOCCMonotone)
open QAlgorithms.Papers.Independent.Information100.P63 (Mat)
open QAlgorithms.Papers.Independent.Dependency50.P35 (tr elog maxRelative)
open QAlgorithms.Papers.Independent.Information100.P67 (IsSubnormalized)

open QAlgorithms.Papers.Independent.Information100.P07 (support)
open QAlgorithms.Papers.Independent.Information100.P71 (supportPower SupportIncluded)

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

/-- PDF pp.2–3: Loewner order, expressed by positivity of the difference. -/
def Le (A B : Mat I) : Prop := (B - A).PosSemidef

/-- PDF p.3, (9): support overlap, including +infinity at zero overlap. -/
def dMin (rho sigma : Mat I) : EReal := -elog (tr (support rho * sigma))

/-- PDF p.3, (11): unnormalized Petz expression. No division by Tr rho.
Only orders in (0,1) are used; zero overlap gives +infinity. -/
def unnormalizedPetz (alpha : ℝ) (rho sigma : Mat I) : EReal :=
  (((alpha - 1)⁻¹ : ℝ) : EReal) *
    elog (tr (supportPower rho alpha * supportPower sigma (1 - alpha)))

/-- PDF pp.2–3: weak positive spectral projector includes the zero eigenspace. -/
def weakPositive (A : Mat I) : Mat I :=
  cfc (fun x : ℝ => if 0 ≤ x then 1 else 0) A

/-- PDF pp.2,7: strict positive spectral projector excludes the zero eigenspace. -/
def strictPositive (A : Mat I) : Mat I :=
  cfc (fun x : ℝ => if 0 < x then 1 else 0) A

/-- PDF p.3, (7): largest eigenvalue. Uses occur on nonempty finite spectra. -/
def largestEigenvalue (A : Mat I) : ℝ := sSup (Set.range (ascEigenvalues A))

/-- PDF p.3, (8): the positive scalar parameters with vanishing positive-part trace. -/
def thresholdSet (rho sigma : Mat I) : Set ℝ :=
  {t | 0 < t ∧ tr (weakPositive (rho - (t : ℂ) • sigma) *
    (rho - (t : ℂ) • sigma)) = 0}

/-- PDF p.5, (27): Umegaki entropy in bits, with its support guard.
The spectral log on the kernel is harmless only in the guarded trace. -/
def unnormalizedRelativeBits (rho sigma : Mat I) : EReal :=
  if SupportIncluded rho sigma then
    (tr (rho * (cfc (fun x : ℝ => Real.logb 2 x) rho -
      cfc (fun x : ℝ => Real.logb 2 x) sigma)) : EReal)
  else ⊤

/-- PDF p.7, Definition 4: full trace-norm ball with the source trace cap. -/
def traceSmoothingBall (eps : ℝ) (rho : Mat I) : Set (Mat I) :=
  {r | r.PosSemidef ∧ traceNorm (r - rho) ≤ eps ∧ tr r ≤ tr rho}

/-- PDF p.7, Definition 4: infimum smoothing in the FIRST argument only. -/
def traceSmoothMax (eps : ℝ) (rho sigma : Mat I) : EReal :=
  sInf {v | ∃ r ∈ traceSmoothingBall eps rho, v = maxRelative r sigma}

/-- PDF p.7, Definition 4: supremum smoothing in the FIRST argument only. -/
def traceSmoothMin (eps : ℝ) (rho sigma : Mat I) : EReal :=
  sSup {v | ∃ r ∈ traceSmoothingBall eps rho, v = dMin r sigma}

/-- PDF p.3, (12): unconditional min entropy. -/
def hMin (rho : Mat I) : EReal := -maxRelative rho 1

/-- PDF p.3, (13): the paper's rank-based max entropy. -/
def hMax (rho : Mat I) : EReal := -dMin rho 1

/-- PDF p.3, (14): reference-state conditional min entropy.
Printed first line uses rho_B; corrected to sigma_B using the preceding
substitution and the second line of (14). Independent moderation required. -/
def conditionalMin (rho : Mat (I × J)) (sigmaB : Mat J) : EReal :=
  -maxRelative rho (Matrix.kronecker (1 : Mat I) sigmaB)

/-- PDF p.3, (15): reference-state conditional rank-based max entropy. -/
def conditionalMax (rho : Mat (I × J)) (sigmaB : Mat J) : EReal :=
  -dMin rho (Matrix.kronecker (1 : Mat I) sigmaB)

/-- PDF p.3, Definition 3: min mutual information with fixed original marginals. -/
def mutualMin (rho : Mat (I × J)) : EReal :=
  dMin rho (Matrix.kronecker (traceRight rho) (traceLeft rho))

/-- PDF p.3, Definition 3: max mutual information. -/
def mutualMax (rho : Mat (I × J)) : EReal :=
  maxRelative rho (Matrix.kronecker (traceRight rho) (traceLeft rho))

/-- PDF pp.4,7: smooth unconditional min entropy with the paper's trace-norm ball. -/
def smoothHMin (eps : ℝ) (rho : Mat I) : EReal := -traceSmoothMax eps rho 1

/-- PDF pp.4,7: smooth unconditional rank-based max entropy. -/
def smoothHMax (eps : ℝ) (rho : Mat I) : EReal := -traceSmoothMin eps rho 1

/-- PDF pp.4,7: smooth reference-state conditional min entropy. -/
def smoothConditionalMin (eps : ℝ) (rho : Mat (I × J)) (sigmaB : Mat J) : EReal :=
  -traceSmoothMax eps rho (Matrix.kronecker (1 : Mat I) sigmaB)

/-- PDF pp.4,7: smooth reference-state conditional rank-based max entropy. -/
def smoothConditionalMax (eps : ℝ) (rho : Mat (I × J)) (sigmaB : Mat J) : EReal :=
  -traceSmoothMin eps rho (Matrix.kronecker (1 : Mat I) sigmaB)

/-- PDF pp.4,7: smoothing holds the ORIGINAL product of marginals fixed. -/
def smoothMutualMin (eps : ℝ) (rho : Mat (I × J)) : EReal :=
  traceSmoothMin eps rho (Matrix.kronecker (traceRight rho) (traceLeft rho))

/-- PDF pp.4,7: smooth max mutual information, with fixed comparison operator. -/
def smoothMutualMax (eps : ℝ) (rho : Mat (I × J)) : EReal :=
  traceSmoothMax eps rho (Matrix.kronecker (traceRight rho) (traceLeft rho))

/-- PDF p.6: support-limit endpoint convention in the Chernoff coefficient.
At s=0 the zero spectral power is the support projector, likewise at s=1. -/
def chernoffCoefficient (rho sigma : Mat I) (s : ℝ) : ℝ :=
  tr (supportPower rho s * supportPower sigma (1 - s))

/-- PDF p.6: Chernoff distance; the theorem also asserts attainment. -/
def chernoff (rho sigma : Mat I) : EReal :=
  -elog (sInf (chernoffCoefficient rho sigma '' Set.Icc (0 : ℝ) 1))

/-- PDF p.7, Definition 5: tensor register at positive length k+1. -/
abbrev PositiveTensorBasis (I : Type) (k : ℕ) := Fin (k + 1) → I

/-- PDF p.7, (40)–(41): positive-part trace of the difference operator.
Every length is k+1, and the threshold is 2^((k+1)*gamma). -/
def spectralMass (rho sigma : (k : ℕ) → Mat (PositiveTensorBasis I k)) (gamma : ℝ) (k : ℕ) : ℝ :=
  let A := rho k - (((2 : ℝ) ^ (((k + 1 : ℕ) : ℝ) * gamma)) : ℂ) • sigma k
  tr (weakPositive A * A)

/-- PDF p.7, (40): extended infimum of the limsup-zero thresholds. -/
def spectralSup (rho sigma : (k : ℕ) → Mat (PositiveTensorBasis I k)) : EReal :=
  sInf {v : EReal | ∃ gamma : ℝ, v = (gamma : EReal) ∧
    Filter.limsup (spectralMass rho sigma gamma) atTop = 0}

/-- PDF p.7, (41): extended supremum of the liminf-one thresholds. -/
def spectralInf (rho sigma : (k : ℕ) → Mat (PositiveTensorBasis I k)) : EReal :=
  sSup {v : EReal | ∃ gamma : ℝ, v = (gamma : EReal) ∧
    Filter.liminf (spectralMass rho sigma gamma) atTop = 1}

/-- PDF p.8, (53): inner asymptotic limsup, taken before eps tends to zero. -/
def maxAsymptotic (rho sigma : (k : ℕ) → Mat (PositiveTensorBasis I k)) (eps : ℝ) : EReal :=
  Filter.limsup (fun k => (((1 / ((k + 1 : ℕ) : ℝ)) : ℝ) : EReal) *
    traceSmoothMax eps (rho k) (sigma k)) atTop

/-- PDF p.9, (67): inner asymptotic liminf, taken before eps tends to zero. -/
def minAsymptotic (rho sigma : (k : ℕ) → Mat (PositiveTensorBasis I k)) (eps : ℝ) : EReal :=
  Filter.liminf (fun k => (((1 / ((k + 1 : ℕ) : ℝ)) : ℝ) : EReal) *
    traceSmoothMin eps (rho k) (sigma k)) atTop

section Bipartite

variable [Nonempty I] [Nonempty J]

/-- PDF p.6, (35): optimization over every separable density matrix. -/
def eMax (rho : Mat (I × J)) : EReal :=
  sInf {v | ∃ sigma : Mat (I × J), IsSeparableDensity sigma ∧ v = maxRelative rho sigma}

/-- PDF p.7, Lemma 13: relative entropy of entanglement. -/
def eRelative (rho : Mat (I × J)) : EReal :=
  sInf {v | ∃ sigma : Mat (I × J), IsSeparableDensity sigma ∧ v = unnormalizedRelativeBits rho sigma}

end Bipartite

/-- PDF p.6, (35): convert the accepted dependent two-party basis to a product basis. -/
def bundledMatrix (s : FiniteState 2) :
    Mat (Fin (s.dims 0) × Fin (s.dims 1)) :=
  fun x y => s.rho (Fin.cases x.1 (Fin.cases x.2 (fun i => Fin.elim0 i)))
    (Fin.cases y.1 (Fin.cases y.2 (fun i => Fin.elim0 i)))

/-- PDF pp.6,12: Emax on accepted bundled states. Theorem 1 asserts finiteness
on every density, so toReal here is never an infinity fallback in the LOCC claim. -/
def bundledEMax (s : FiniteState 2) : ℝ :=
  letI : Nonempty (Fin (s.dims 0)) := ⟨⟨0, s.positive 0⟩⟩
  letI : Nonempty (Fin (s.dims 1)) := ⟨⟨0, s.positive 1⟩⟩
  (eMax (bundledMatrix s)).toReal

/-! Elementary sanity checks for the new definitions. -/

example : elog 0 = ⊥ := by simp [elog]

example : elog 1 = 0 := by simp [elog]

example : elog 2 = 1 := by norm_num [elog, Real.logb_self_eq_one]

example {rho : Mat I} : IsSubnormalized rho → rho.PosSemidef := fun h => h.1

example (A : Mat I) : Le A A := by
  simpa only [Le, sub_self] using (Matrix.PosSemidef.zero : (0 : Mat I).PosSemidef)

example (rho : Mat I) : hMin rho = -maxRelative rho 1 := rfl

example (rho : Mat (I × J)) (sigma : Mat J) (eps : ℝ) :
    smoothConditionalMax eps rho sigma =
      -traceSmoothMin eps rho (Matrix.kronecker (1 : Mat I) sigma) := rfl

example (rho : Mat (I × J)) (eps : ℝ) :
    smoothMutualMax eps rho =
      traceSmoothMax eps rho (Matrix.kronecker (traceRight rho) (traceLeft rho)) := rfl

example (rho sigma : Mat I) :
    chernoffCoefficient rho sigma 0 = tr (supportPower rho 0 * supportPower sigma 1) := by
  simp only [chernoffCoefficient, sub_zero]

example (rho : Mat I) : dMin rho 0 = ⊤ := by
  simp [dMin, tr, elog]

example (rho sigma : Mat I) (h : ¬ SupportIncluded rho sigma) :
    unnormalizedRelativeBits rho sigma = ⊤ := by simp [unnormalizedRelativeBits, h]

example : strictPositive (0 : Mat I) = 0 := by
  simp [strictPositive, cfc_zero]

example (rho : Mat I) (h : rho.PosSemidef) : rho ∈ traceSmoothingBall 0 rho := by
  simp [traceSmoothingBall, h, traceNorm, cfc_zero]

example (rho : Mat I) : maxRelative rho rho ≤ 0 := by
  unfold maxRelative
  apply sInf_le
  refine ⟨0, by simp, ?_⟩
  simpa only [Real.rpow_zero, Complex.ofReal_one, one_smul, sub_self] using
    (Matrix.PosSemidef.zero : (0 : Mat I).PosSemidef)

example (rho sigma : Mat I) (h : rho ∈ traceSmoothingBall 0 rho) :
    traceSmoothMax 0 rho sigma ≤ maxRelative rho sigma := by
  unfold traceSmoothMax
  exact sInf_le ⟨rho, h, rfl⟩

example (rho sigma : Mat I) (h : rho ∈ traceSmoothingBall 0 rho) :
    dMin rho sigma ≤ traceSmoothMin 0 rho sigma := by
  unfold traceSmoothMin
  exact le_sSup ⟨rho, h, rfl⟩

example (rho : Mat I) : 1 ∈ thresholdSet rho rho := by
  simp [thresholdSet, tr]

example (s : FiniteState 2) (x y : Fin (s.dims 0) × Fin (s.dims 1)) :
    bundledMatrix s x y = s.rho (Fin.cases x.1 (Fin.cases x.2 (fun i => Fin.elim0 i)))
      (Fin.cases y.1 (Fin.cases y.2 (fun i => Fin.elim0 i))) := rfl

section Statements

variable [Nonempty I] [Nonempty J]

end Statements

end QAlgorithms.Papers.Independent.Support50.P21
Source
Min-and Max-Relative Entropies and a New Entanglement Monotone, arXiv:0803.2770v3, https://arxiv.org/abs/0803.2770v3
Read-back

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

QAlgorithms.Papers.Independent.Support50.P21.Le

For a finite index type III and arbitrary complex square matrices A,BA,BA,B indexed by III, A≤BA\le BA≤B means that B−AB-AB−A is positive semidefinite. Neither matrix is separately required to be Hermitian, positive semidefinite, or normalized. The index type may be empty.

QAlgorithms.Papers.Independent.Support50.P21.PositiveTensorBasis

For any type III and natural number kkk, this is the type of functions from {0,…,k}\{0,\ldots,k\}{0,…,k} to III, equivalently ordered tuples of k+1k+1k+1 elements of III. Thus k=0k=0k=0 gives tuples with one coordinate, rather than empty tuples; III is not required to be finite or nonempty.

QAlgorithms.Papers.Independent.Support50.P21.bundledEMax

Given a two-party finite state sss with positive local dimensions d0,d1d_0,d_1d0​,d1​ and density matrix ρs\rho_sρs​, first reindex its function basis as pairs, obtaining

ρ~s((x0,x1),(y0,y1))=ρs(zx,zy),zx(0)=x0,zx(1)=x1.\widetilde\rho_s((x_0,x_1),(y_0,y_1))=\rho_s(z_x,z_y), \quad z_x(0)=x_0,\quad z_x(1)=x_1.ρ​s​((x0​,x1​),(y0​,y1​))=ρs​(zx​,zy​),zx​(0)=x0​,zx​(1)=x1​.

Form the extended-real infimum

E=inf⁡σinf⁡{ℓ∈R:2ℓσ−ρ~s≥0},E=\inf_{\sigma} \inf\{\ell\in\mathbb R:2^\ell\sigma-\widetilde\rho_s\ge0\},E=σinf​inf{ℓ∈R:2ℓσ−ρ​s​≥0},

where σ\sigmaσ ranges over finite convex combinations of tensor products of local positive semidefinite, trace-one matrices. The defined real number is the conversion of EEE to R\mathbb RR: it equals EEE when EEE is finite and equals zero when E=+∞E=+\inftyE=+∞ or E=−∞E=-\inftyE=−∞. The positive local dimensions supply the nonempty local basis sets required for the outer infimum.

QAlgorithms.Papers.Independent.Support50.P21.bundledMatrix

Given a two-party finite state sss, consisting of positive natural dimensions d0,d1d_0,d_1d0​,d1​ and a positive semidefinite, trace-one matrix ρs\rho_sρs​ indexed by functions zzz with z(0)∈{0,…,d0−1}z(0)\in\{0,\ldots,d_0-1\}z(0)∈{0,…,d0​−1} and z(1)∈{0,…,d1−1}z(1)\in\{0,\ldots,d_1-1\}z(1)∈{0,…,d1​−1}, this defines a matrix indexed by pairs from those two sets. Its entries are

ρ~s((x0,x1),(y0,y1))=ρs(zx,zy),zx(0)=x0,zx(1)=x1,zy(0)=y0,zy(1)=y1.\widetilde\rho_s\bigl((x_0,x_1),(y_0,y_1)\bigr) =\rho_s(z_x,z_y), \qquad z_x(0)=x_0,\quad z_x(1)=x_1,\quad z_y(0)=y_0,\quad z_y(1)=y_1.ρ​s​((x0​,x1​),(y0​,y1​))=ρs​(zx​,zy​),zx​(0)=x0​,zx​(1)=x1​,zy​(0)=y0​,zy​(1)=y1​.

Thus the definition reindexes the two-coordinate function basis as a pair basis; both dimensions are required to be strictly positive by the data carried by sss.

QAlgorithms.Papers.Independent.Support50.P21.chernoff

For a finite basis type III with decidable equality and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, define Pt(X)P_t(X)Pt​(X) by real continuous functional calculus using the scalar function x↦0x\mapsto 0x↦0 when x=0x=0x=0 and x↦xtx\mapsto x^tx↦xt otherwise. In particular, P0(X)P_0(X)P0​(X) uses value 000, rather than 111, at a zero eigenvalue. Set

c=inf⁡0≤s≤1Re⁡Tr⁡ ⁣(Ps(ρ)P1−s(σ)),c=\inf_{0\le s\le1}\operatorname{Re}\operatorname{Tr}\!\bigl(P_s(\rho)P_{1-s}(\sigma)\bigr),c=0≤s≤1inf​ReTr(Ps​(ρ)P1−s​(σ)),

where this infimum is taken in R\mathbb RR. The defined value is −log⁡2c-\log_2 c−log2​c, regarded as an extended real number, if c>0c>0c>0, and is +∞+\infty+∞ if c≤0c\le0c≤0. Neither positivity nor trace normalization of either input matrix is assumed.

QAlgorithms.Papers.Independent.Support50.P21.chernoffCoefficient

For any finite type III with decidable equality, possibly empty, arbitrary complex matrices ρ,σ\rho,\sigmaρ,σ on III, and any real sss, define ft(x)=0f_t(x)=0ft​(x)=0 when x=0x=0x=0 and ft(x)=xtf_t(x)=x^tft​(x)=xt otherwise, using the total real-power operation. The real-valued coefficient is

Re⁡Tr⁡(fs(ρ) f1−s(σ)),\operatorname{Re}\operatorname{Tr} \bigl(f_s(\rho)\,f_{1-s}(\sigma)\bigr),ReTr(fs​(ρ)f1−s​(σ)),

where each matrix function is evaluated by the real continuous functional calculus. There are no positivity, normalization, or Hermitian assumptions on the matrices, and no restriction 0≤s≤10\le s\le10≤s≤1. Zero eigenvalues contribute zero even when the exponent is zero or negative; on a non-Hermitian matrix this functional calculus returns zero. For empty III, the coefficient is zero.

QAlgorithms.Papers.Independent.Support50.P21.conditionalMax

For any finite types I,JI,JI,J with decidable equality, possibly empty, any complex matrix ρ\rhoρ on I×JI\times JI×J, and any complex matrix σB\sigma_BσB​ on JJJ, let PρP_\rhoPρ​ be the orthogonal projection onto the range of ρ\rhoρ. Define L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x for x>0x>0x>0, and L(x)=−∞L(x)=-\inftyL(x)=−∞ for x≤0x\le0x≤0. The extended-real value defined here is

L ⁣(Re⁡Tr⁡(Pρ(II⊗σB))).L\!\left(\operatorname{Re}\operatorname{Tr} \bigl(P_\rho(I_I\otimes\sigma_B)\bigr)\right).L(ReTr(Pρ​(II​⊗σB​))).

There are no positivity or normalization assumptions and no optimization over σB\sigma_BσB​. In particular, the value is −∞-\infty−∞ when ρ=0\rho=0ρ=0.

QAlgorithms.Papers.Independent.Support50.P21.conditionalMin

For any finite types I,JI,JI,J with decidable equality, possibly empty, any complex matrix ρ\rhoρ indexed by (I×J)×(I×J)(I\times J)\times(I\times J)(I×J)×(I×J), and any complex matrix σB\sigma_BσB​ indexed by J×JJ\times JJ×J, the definition assigns the extended-real value

−inf⁡{λ∈R:2λ(II⊗σB)−ρ is positive semidefinite},-\inf\left\{\lambda\in\mathbb R: 2^\lambda(I_I\otimes\sigma_B)-\rho \text{ is positive semidefinite}\right\},−inf{λ∈R:2λ(II​⊗σB​)−ρ is positive semidefinite},

where III_III​ is the identity matrix on III, and the infimum is taken in the extended reals. Neither matrix is required to be positive or normalized, and there is no optimization over σB\sigma_BσB​. If the displayed set is empty, the value is −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.dMin

For a finite index type III with decidable equality and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, let PρP_\rhoPρ​ be the orthogonal projection onto the range of ρ\rhoρ, and put x=Re⁡Tr⁡(Pρσ)x=\operatorname{Re}\operatorname{Tr}(P_\rho\sigma)x=ReTr(Pρ​σ). The extended-real quantity defined here is

dmin⁡(ρ,σ)={−log⁡2x,x>0,+∞,x≤0.d_{\min}(\rho,\sigma)= \begin{cases} -\log_2 x,&x>0,\\ +\infty,&x\le0. \end{cases}dmin​(ρ,σ)={−log2​x,+∞,​x>0,x≤0.​

No positivity or normalization is assumed. In particular, ρ=0\rho=0ρ=0 gives +∞+\infty+∞, including on an empty index type.

QAlgorithms.Papers.Independent.Support50.P21.eMax

For finite nonempty basis types I,JI,JI,J with decidable equality and an arbitrary complex square matrix ρ\rhoρ indexed by I×JI\times JI×J, this is the extended-real quantity

inf⁡σinf⁡{ℓ∈R:2ℓσ−ρ≥0}.\inf_{\sigma} \inf\left\{\ell\in\mathbb R:2^\ell\sigma-\rho\ge0\right\}.σinf​inf{ℓ∈R:2ℓσ−ρ≥0}.

The outer infimum ranges over precisely the matrices admitting a finite decomposition

σ=∑j=0q−1wj(Aj⊗Bj),wj≥0,∑jwj=1,\sigma=\sum_{j=0}^{q-1}w_j(A_j\otimes B_j), \qquad w_j\ge0,\quad \sum_jw_j=1,σ=j=0∑q−1​wj​(Aj​⊗Bj​),wj​≥0,j∑​wj​=1,

where every AjA_jAj​ and BjB_jBj​ is positive semidefinite with complex trace equal to one on III and JJJ, respectively. Zero weights are permitted, while q=0q=0q=0 cannot satisfy the weight-sum condition. The matrix inequality in the inner infimum means positive semidefiniteness. No density-matrix assumption is made on ρ\rhoρ; an empty inner set has infimum +∞+\infty+∞, and an inner set unbounded below has infimum −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.eRelative

For finite nonempty basis types I,JI,JI,J with decidable equality and an arbitrary complex square matrix ρ\rhoρ indexed by I×JI\times JI×J, this is the extended-real infimum, over matrices σ\sigmaσ of the form

σ=∑j=0q−1wj(Aj⊗Bj),wj≥0,∑jwj=1,\sigma=\sum_{j=0}^{q-1}w_j(A_j\otimes B_j), \qquad w_j\ge0,\quad \sum_jw_j=1,σ=j=0∑q−1​wj​(Aj​⊗Bj​),wj​≥0,j∑​wj​=1,

with every Aj,BjA_j,B_jAj​,Bj​ positive semidefinite and of trace one, of the value

{Re⁡Tr⁡ ⁣[ρ(log⁡2(ρ)−log⁡2(σ))],ker⁡σ⊆ker⁡ρ,+∞,ker⁡σ⊈ker⁡ρ.\begin{cases} \operatorname{Re}\operatorname{Tr}\!\left[ \rho\bigl(\log_2(\rho)-\log_2(\sigma)\bigr)\right], &\ker\sigma\subseteq\ker\rho,\\ +\infty,&\ker\sigma\not\subseteq\ker\rho. \end{cases}{ReTr[ρ(log2​(ρ)−log2​(σ))],+∞,​kerσ⊆kerρ,kerσ⊆kerρ.​

Here each matrix logarithm is formed by real continuous functional calculus using the total real function x↦log⁡2xx\mapsto\log_2 xx↦log2​x, including its value zero at x=0x=0x=0. The trace expression is not divided by Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}\rhoReTrρ. No positivity or trace-normalization hypothesis is imposed on ρ\rhoρ. The decomposition is finite, allows zero weights, and cannot have q=0q=0q=0, because its weights must sum to one.

QAlgorithms.Papers.Independent.Support50.P21.hMax

For any finite type III with decidable equality, possibly empty, and any complex matrix ρ\rhoρ indexed by I×II\times II×I, let PρP_\rhoPρ​ be the orthogonal projection onto the range of ρ\rhoρ, and define L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x when x>0x>0x>0 and L(x)=−∞L(x)=-\inftyL(x)=−∞ otherwise. The extended-real value defined here is

hmax⁡(ρ)=L ⁣(Re⁡Tr⁡Pρ).h_{\max}(\rho)=L\!\left(\operatorname{Re}\operatorname{Tr}P_\rho\right).hmax​(ρ)=L(ReTrPρ​).

No positivity or normalization of ρ\rhoρ is assumed. In particular, this is −∞-\infty−∞ when ρ=0\rho=0ρ=0, including when III is empty.

QAlgorithms.Papers.Independent.Support50.P21.hMin

For a finite index type III with decidable equality and any complex square matrix ρ\rhoρ indexed by III, this defines

hmin⁡(ρ)=−inf⁡{λ∈R:2λI−ρ⪰0},h_{\min}(\rho)= -\inf\left\{\lambda\in\mathbb R: 2^\lambda I-\rho\succeq0\right\},hmin​(ρ)=−inf{λ∈R:2λI−ρ⪰0},

where the infimum is taken in the extended reals and each real λ\lambdaλ is embedded there. No positivity or normalization is assumed. If the feasible set is empty, the value is −∞-\infty−∞; if its infimum is −∞-\infty−∞, the value is +∞+\infty+∞. In particular, ρ=0\rho=0ρ=0, including on an empty index type, gives +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P21.largestEigenvalue

For a finite index type III with decidable equality and an arbitrary complex square matrix AAA indexed by III, this defines the real supremum of the values in its ascending eigenvalue list. For Hermitian AAA, that list contains its real eigenvalues in ascending order, with multiplicities; for non-Hermitian AAA, every entry is defined to be 000. Thus, in positive dimension the value is the greatest eigenvalue when AAA is Hermitian and 000 otherwise. If III is empty, the list is empty and its real supremum is defined to be 000.

QAlgorithms.Papers.Independent.Support50.P21.maxAsymptotic

For a finite basis type III with decidable equality, arbitrary matrix sequences ρk,σk\rho_k,\sigma_kρk​,σk​ indexed by Ik+1I^{k+1}Ik+1, and any real ε\varepsilonε, let

Bε(ρk)={r:r≥0, ∥r−ρk∥1≤ε, Re⁡Tr⁡r≤Re⁡Tr⁡ρk},\mathcal B_\varepsilon(\rho_k)= \left\{r:r\ge0,\ \|r-\rho_k\|_1\le\varepsilon,\ \operatorname{Re}\operatorname{Tr}r\le \operatorname{Re}\operatorname{Tr}\rho_k\right\},Bε​(ρk​)={r:r≥0, ∥r−ρk​∥1​≤ε, ReTrr≤ReTrρk​},

where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. The defined value is

lim sup⁡k→∞1k+1inf⁡r∈Bε(ρk)inf⁡{ℓ∈R:2ℓσk−r≥0},\limsup_{k\to\infty}\frac1{k+1} \inf_{r\in\mathcal B_\varepsilon(\rho_k)} \inf\left\{\ell\in\mathbb R:2^\ell\sigma_k-r\ge0\right\},k→∞limsup​k+11​r∈Bε​(ρk​)inf​inf{ℓ∈R:2ℓσk​−r≥0},

with both infima and the limit superior taken in the extended real numbers, and matrix inequalities meaning positive semidefiniteness. The smoothing matrices need not have trace one or be nonzero. Neither input sequence is assumed positive or normalized, and its terms need not be tensor powers. The parameter ε\varepsilonε has no sign restriction; if it is negative, every smoothing ball is empty and the defined value is +∞+\infty+∞. The denominator k+1k+1k+1 is always positive, including at k=0k=0k=0.

QAlgorithms.Papers.Independent.Support50.P21.minAsymptotic

For a finite basis type III with decidable equality, arbitrary matrix sequences ρk,σk\rho_k,\sigma_kρk​,σk​ indexed by Ik+1I^{k+1}Ik+1, and any real ε\varepsilonε, let

Bε(ρk)={r:r≥0, ∥r−ρk∥1≤ε, Re⁡Tr⁡r≤Re⁡Tr⁡ρk},∥X∥1=Re⁡Tr⁡X†X.\mathcal B_\varepsilon(\rho_k)= \left\{r:r\ge0,\ \|r-\rho_k\|_1\le\varepsilon,\ \operatorname{Re}\operatorname{Tr}r\le \operatorname{Re}\operatorname{Tr}\rho_k\right\}, \qquad \|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}.Bε​(ρk​)={r:r≥0, ∥r−ρk​∥1​≤ε, ReTrr≤ReTrρk​},∥X∥1​=ReTrX†X​.

For such an rrr, let PrP_rPr​ be the orthogonal projection onto the range of the linear operator represented by rrr, and define

d(r,σk)={−log⁡2 ⁣(Re⁡Tr⁡(Prσk)),Re⁡Tr⁡(Prσk)>0,+∞,Re⁡Tr⁡(Prσk)≤0.d(r,\sigma_k)= \begin{cases} -\log_2\!\bigl(\operatorname{Re}\operatorname{Tr}(P_r\sigma_k)\bigr), &\operatorname{Re}\operatorname{Tr}(P_r\sigma_k)>0,\\ +\infty,&\operatorname{Re}\operatorname{Tr}(P_r\sigma_k)\le0. \end{cases}d(r,σk​)={−log2​(ReTr(Pr​σk​)),+∞,​ReTr(Pr​σk​)>0,ReTr(Pr​σk​)≤0.​

The defined value is

lim inf⁡k→∞1k+1sup⁡r∈Bε(ρk)d(r,σk),\liminf_{k\to\infty}\frac1{k+1} \sup_{r\in\mathcal B_\varepsilon(\rho_k)}d(r,\sigma_k),k→∞liminf​k+11​r∈Bε​(ρk​)sup​d(r,σk​),

using extended-real suprema and limit inferior. No positivity or normalization of the input sequences, tensor-power relationship, or sign condition on ε\varepsilonε is assumed. A negative ε\varepsilonε makes every smoothing ball empty and gives value −∞-\infty−∞. Zero smoothing matrices are allowed; whenever zero belongs to a smoothing ball, its contribution to the inner supremum is +∞+\infty+∞. The denominator k+1k+1k+1 never vanishes.

QAlgorithms.Papers.Independent.Support50.P21.mutualMax

For any finite types I,JI,JI,J with decidable equality, possibly empty, and any complex matrix ρ\rhoρ on I×JI\times JI×J, define (ρI)ij=∑b∈Jρ(i,b),(j,b)(\rho_I)_{ij}=\sum_{b\in J}\rho_{(i,b),(j,b)}(ρI​)ij​=∑b∈J​ρ(i,b),(j,b)​ and (ρJ)ab=∑i∈Iρ(i,a),(i,b)(\rho_J)_{ab}=\sum_{i\in I}\rho_{(i,a),(i,b)}(ρJ​)ab​=∑i∈I​ρ(i,a),(i,b)​. The extended-real value defined here is

inf⁡{λ∈R:2λ(ρI⊗ρJ)−ρ is positive semidefinite}.\inf\left\{\lambda\in\mathbb R: 2^\lambda(\rho_I\otimes\rho_J)-\rho \text{ is positive semidefinite}\right\}.inf{λ∈R:2λ(ρI​⊗ρJ​)−ρ is positive semidefinite}.

No positivity or normalization of ρ\rhoρ is assumed. The infimum is +∞+\infty+∞ for an empty set and −∞-\infty−∞ for a set unbounded below; in particular, ρ=0\rho=0ρ=0 gives −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.mutualMin

For any finite types I,JI,JI,J with decidable equality, possibly empty, and any complex matrix ρ\rhoρ on I×JI\times JI×J, define matrices ρI,ρJ\rho_I,\rho_JρI​,ρJ​ by (ρI)ij=∑b∈Jρ(i,b),(j,b)(\rho_I)_{ij}=\sum_{b\in J}\rho_{(i,b),(j,b)}(ρI​)ij​=∑b∈J​ρ(i,b),(j,b)​ and (ρJ)ab=∑i∈Iρ(i,a),(i,b)(\rho_J)_{ab}=\sum_{i\in I}\rho_{(i,a),(i,b)}(ρJ​)ab​=∑i∈I​ρ(i,a),(i,b)​. Let PρP_\rhoPρ​ be the orthogonal projection onto the range of ρ\rhoρ, and let L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x for x>0x>0x>0, with L(x)=−∞L(x)=-\inftyL(x)=−∞ otherwise. The definition is

−L ⁣(Re⁡Tr⁡(Pρ(ρI⊗ρJ))).-L\!\left(\operatorname{Re}\operatorname{Tr} \bigl(P_\rho(\rho_I\otimes\rho_J)\bigr)\right).−L(ReTr(Pρ​(ρI​⊗ρJ​))).

No positivity or normalization of ρ\rhoρ is assumed. A nonpositive argument of LLL, in particular the argument obtained when ρ=0\rho=0ρ=0, gives the value +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P21.smoothConditionalMax

For any finite types I,JI,JI,J with decidable equality, possibly empty, any real ε\varepsilonε, any complex matrix ρ\rhoρ on I×JI\times JI×J, and any complex matrix σB\sigma_BσB​ on JJJ, let Bε(ρ)\mathcal B_\varepsilon(\rho)Bε​(ρ) consist of the positive semidefinite matrices rrr satisfying ∥r−ρ∥1≤ε\|r-\rho\|_1\le\varepsilon∥r−ρ∥1​≤ε and Re⁡Tr⁡r≤Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rhoReTrr≤ReTrρ, where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. Write PrP_rPr​ for the orthogonal projection onto the range of rrr, and define L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x for x>0x>0x>0, with L(x)=−∞L(x)=-\inftyL(x)=−∞ otherwise. The definition is

−sup⁡r∈Bε(ρ)[−L ⁣(Re⁡Tr⁡(Pr(II⊗σB)))],-\sup_{r\in\mathcal B_\varepsilon(\rho)} \left[-L\!\left(\operatorname{Re}\operatorname{Tr} \bigl(P_r(I_I\otimes\sigma_B)\bigr)\right)\right],−r∈Bε​(ρ)sup​[−L(ReTr(Pr​(II​⊗σB​)))],

with an extended-real supremum. No positivity or normalization of either supplied matrix is assumed, and σB\sigma_BσB​ is fixed. An empty smoothing set gives +∞+\infty+∞, including for ε<0\varepsilon<0ε<0; inclusion of zero gives −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.smoothConditionalMin

For any finite types I,JI,JI,J with decidable equality, possibly empty, any real ε\varepsilonε, any complex matrix ρ\rhoρ on I×JI\times JI×J, and any complex matrix σB\sigma_BσB​ on JJJ, let Bε(ρ)\mathcal B_\varepsilon(\rho)Bε​(ρ) be the positive semidefinite matrices rrr on I×JI\times JI×J satisfying ∥r−ρ∥1≤ε\|r-\rho\|_1\le\varepsilon∥r−ρ∥1​≤ε and Re⁡Tr⁡r≤Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rhoReTrr≤ReTrρ, where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. The definition assigns

−inf⁡r∈Bε(ρ)inf⁡{λ∈R:2λ(II⊗σB)−r is positive semidefinite},-\inf_{r\in\mathcal B_\varepsilon(\rho)} \inf\left\{\lambda\in\mathbb R: 2^\lambda(I_I\otimes\sigma_B)-r \text{ is positive semidefinite}\right\},−r∈Bε​(ρ)inf​inf{λ∈R:2λ(II​⊗σB​)−r is positive semidefinite},

using extended-real infima. Neither ρ\rhoρ nor σB\sigma_BσB​ is required to be positive or normalized, and σB\sigma_BσB​ remains fixed. An empty smoothing set gives −∞-\infty−∞, including whenever ε<0\varepsilon<0ε<0.

QAlgorithms.Papers.Independent.Support50.P21.smoothHMax

For any finite type III with decidable equality, possibly empty, any real ε\varepsilonε, and any complex matrix ρ\rhoρ on III, let Bε(ρ)\mathcal B_\varepsilon(\rho)Bε​(ρ) consist of all positive semidefinite matrices rrr satisfying ∥r−ρ∥1≤ε\|r-\rho\|_1\le\varepsilon∥r−ρ∥1​≤ε and Re⁡Tr⁡r≤Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rhoReTrr≤ReTrρ, where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. For each rrr, let PrP_rPr​ be the orthogonal projection onto its range, and set L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x for x>0x>0x>0 and L(x)=−∞L(x)=-\inftyL(x)=−∞ otherwise. The definition is

−sup⁡r∈Bε(ρ)[−L ⁣(Re⁡Tr⁡Pr)],-\sup_{r\in\mathcal B_\varepsilon(\rho)} \left[-L\!\left(\operatorname{Re}\operatorname{Tr}P_r\right)\right],−r∈Bε​(ρ)sup​[−L(ReTrPr​)],

with the supremum in the extended reals. No positivity or normalization of ρ\rhoρ is required. An empty smoothing set gives +∞+\infty+∞, including for ε<0\varepsilon<0ε<0; if zero belongs to the smoothing set, the value is −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.smoothHMin

For any finite type III with decidable equality, possibly empty, any real ε\varepsilonε, and any complex matrix ρ\rhoρ on III, put

Bε(ρ)={r: r is positive semidefinite, ∥r−ρ∥1≤ε, Re⁡Tr⁡r≤Re⁡Tr⁡ρ},∥X∥1=Re⁡Tr⁡X†X.\mathcal B_\varepsilon(\rho)= \left\{r:\ r\text{ is positive semidefinite},\ \|r-\rho\|_1\le\varepsilon,\ \operatorname{Re}\operatorname{Tr}r\le \operatorname{Re}\operatorname{Tr}\rho\right\}, \qquad \|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}.Bε​(ρ)={r: r is positive semidefinite, ∥r−ρ∥1​≤ε, ReTrr≤ReTrρ},∥X∥1​=ReTrX†X​.

The definition is the negative of the extended-real infimum, over r∈Bε(ρ)r\in\mathcal B_\varepsilon(\rho)r∈Bε​(ρ), of inf⁡{λ∈R:2λII−r is positive semidefinite}\inf\{\lambda\in\mathbb R:2^\lambda I_I-r\text{ is positive semidefinite}\}inf{λ∈R:2λII​−r is positive semidefinite}. The matrix ρ\rhoρ need not be positive or normalized, and zero matrices are permitted in the smoothing set. An empty smoothing set gives −∞-\infty−∞; this occurs whenever ε<0\varepsilon<0ε<0.

QAlgorithms.Papers.Independent.Support50.P21.smoothMutualMax

For any finite types I,JI,JI,J with decidable equality, possibly empty, any real ε\varepsilonε, and any complex matrix ρ\rhoρ on I×JI\times JI×J, define (ρI)ij=∑b∈Jρ(i,b),(j,b)(\rho_I)_{ij}=\sum_{b\in J}\rho_{(i,b),(j,b)}(ρI​)ij​=∑b∈J​ρ(i,b),(j,b)​ and (ρJ)ab=∑i∈Iρ(i,a),(i,b)(\rho_J)_{ab}=\sum_{i\in I}\rho_{(i,a),(i,b)}(ρJ​)ab​=∑i∈I​ρ(i,a),(i,b)​. Let Bε(ρ)\mathcal B_\varepsilon(\rho)Bε​(ρ) consist of all positive semidefinite matrices rrr satisfying ∥r−ρ∥1≤ε\|r-\rho\|_1\le\varepsilon∥r−ρ∥1​≤ε and Re⁡Tr⁡r≤Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rhoReTrr≤ReTrρ, where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. The definition is

inf⁡r∈Bε(ρ)inf⁡{λ∈R:2λ(ρI⊗ρJ)−r is positive semidefinite},\inf_{r\in\mathcal B_\varepsilon(\rho)} \inf\left\{\lambda\in\mathbb R: 2^\lambda(\rho_I\otimes\rho_J)-r \text{ is positive semidefinite}\right\},r∈Bε​(ρ)inf​inf{λ∈R:2λ(ρI​⊗ρJ​)−r is positive semidefinite},

with extended-real infima. The marginal matrices remain those of the original ρ\rhoρ, which need not be positive or normalized. An empty smoothing set gives +∞+\infty+∞, including whenever ε<0\varepsilon<0ε<0.

QAlgorithms.Papers.Independent.Support50.P21.smoothMutualMin

For any finite types I,JI,JI,J with decidable equality, possibly empty, any real ε\varepsilonε, and any complex matrix ρ\rhoρ on I×JI\times JI×J, define (ρI)ij=∑b∈Jρ(i,b),(j,b)(\rho_I)_{ij}=\sum_{b\in J}\rho_{(i,b),(j,b)}(ρI​)ij​=∑b∈J​ρ(i,b),(j,b)​ and (ρJ)ab=∑i∈Iρ(i,a),(i,b)(\rho_J)_{ab}=\sum_{i\in I}\rho_{(i,a),(i,b)}(ρJ​)ab​=∑i∈I​ρ(i,a),(i,b)​. Let Bε(ρ)\mathcal B_\varepsilon(\rho)Bε​(ρ) be the positive semidefinite matrices rrr satisfying ∥r−ρ∥1≤ε\|r-\rho\|_1\le\varepsilon∥r−ρ∥1​≤ε and Re⁡Tr⁡r≤Re⁡Tr⁡ρ\operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rhoReTrr≤ReTrρ, where ∥X∥1=Re⁡Tr⁡X†X\|X\|_1=\operatorname{Re}\operatorname{Tr}\sqrt{X^\dagger X}∥X∥1​=ReTrX†X​. With PrP_rPr​ the orthogonal projection onto the range of rrr, and L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x for x>0x>0x>0 and −∞-\infty−∞ otherwise, the definition is

sup⁡r∈Bε(ρ)[−L ⁣(Re⁡Tr⁡(Pr(ρI⊗ρJ)))].\sup_{r\in\mathcal B_\varepsilon(\rho)} \left[-L\!\left(\operatorname{Re}\operatorname{Tr} \bigl(P_r(\rho_I\otimes\rho_J)\bigr)\right)\right].r∈Bε​(ρ)sup​[−L(ReTr(Pr​(ρI​⊗ρJ​)))].

The supremum is extended-real, and the two marginal matrices are computed from the original ρ\rhoρ throughout. No positivity or normalization of ρ\rhoρ is assumed. An empty smoothing set gives −∞-\infty−∞, including for ε<0\varepsilon<0ε<0; inclusion of zero gives +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P21.spectralInf

For a finite basis type III with decidable equality and arbitrary matrix sequences ρk,σk\rho_k,\sigma_kρk​,σk​ indexed by Ik+1I^{k+1}Ik+1, put

Ak(γ)=ρk−2(k+1)γσk,mk(γ)=Re⁡Tr⁡ ⁣(f(Ak(γ))Ak(γ)),A_k(\gamma)=\rho_k-2^{(k+1)\gamma}\sigma_k, \qquad m_k(\gamma)=\operatorname{Re}\operatorname{Tr}\!\bigl(f(A_k(\gamma))A_k(\gamma)\bigr),Ak​(γ)=ρk​−2(k+1)γσk​,mk​(γ)=ReTr(f(Ak​(γ))Ak​(γ)),

where real continuous functional calculus applies f(x)=1f(x)=1f(x)=1 for x≥0x\ge0x≥0 and f(x)=0f(x)=0f(x)=0 for x<0x<0x<0. The defined value is the extended-real supremum

sup⁡{γ∈R:lim inf⁡k→∞mk(γ)=1}.\sup\left\{\gamma\in\mathbb R: \liminf_{k\to\infty}m_k(\gamma)=1\right\}.sup{γ∈R:k→∞liminf​mk​(γ)=1}.

The inner limit inferior is taken in R\mathbb RR, with its total order-theoretic conventions; equality to one does not itself assert convergence to one. Only finite real thresholds enter the set, whose extended-real supremum is −∞-\infty−∞ if it is empty and +∞+\infty+∞ if it is unbounded above. No positivity, normalization, or tensor-power relationship is assumed.

QAlgorithms.Papers.Independent.Support50.P21.spectralMass

For a finite basis type III with decidable equality, arbitrary sequences of complex square matrices ρk,σk\rho_k,\sigma_kρk​,σk​ indexed by Ik+1I^{k+1}Ik+1, a real number γ\gammaγ, and a natural number kkk, put

Ak=ρk−2(k+1)γσk.A_k=\rho_k-2^{(k+1)\gamma}\sigma_k.Ak​=ρk​−2(k+1)γσk​.

The defined real number is

Re⁡Tr⁡ ⁣(f(Ak)Ak),f(x)={1,0≤x,0,x<0,\operatorname{Re}\operatorname{Tr}\!\bigl(f(A_k)A_k\bigr), \qquad f(x)= \begin{cases} 1,&0\le x,\\ 0,&x<0, \end{cases}ReTr(f(Ak​)Ak​),f(x)={1,0,​0≤x,x<0,​

where f(Ak)f(A_k)f(Ak​) is formed by real continuous functional calculus. The scalar function includes zero in its nonnegative branch. The definition imposes no positivity, Hermiticity, normalization, or tensor-power relationship on the matrices, and k=0k=0k=0 uses matrices indexed by one-coordinate tuples.

QAlgorithms.Papers.Independent.Support50.P21.spectralSup

For a finite basis type III with decidable equality and arbitrary matrix sequences ρk,σk\rho_k,\sigma_kρk​,σk​ indexed by Ik+1I^{k+1}Ik+1, define, for each real γ\gammaγ,

Ak(γ)=ρk−2(k+1)γσk,mk(γ)=Re⁡Tr⁡ ⁣(f(Ak(γ))Ak(γ)),A_k(\gamma)=\rho_k-2^{(k+1)\gamma}\sigma_k, \qquad m_k(\gamma)=\operatorname{Re}\operatorname{Tr}\!\bigl(f(A_k(\gamma))A_k(\gamma)\bigr),Ak​(γ)=ρk​−2(k+1)γσk​,mk​(γ)=ReTr(f(Ak​(γ))Ak​(γ)),

where real continuous functional calculus applies f(x)=1f(x)=1f(x)=1 for x≥0x\ge0x≥0 and f(x)=0f(x)=0f(x)=0 for x<0x<0x<0. The defined value is the extended-real infimum

inf⁡{γ∈R:lim sup⁡k→∞mk(γ)=0}.\inf\left\{\gamma\in\mathbb R: \limsup_{k\to\infty}m_k(\gamma)=0\right\}.inf{γ∈R:k→∞limsup​mk​(γ)=0}.

The inner limit superior is taken in R\mathbb RR, with its total order-theoretic conventions; the condition is equality of that limit superior to zero, rather than an assertion that the sequence converges to zero. Only finite real thresholds enter the set, whose extended-real infimum is +∞+\infty+∞ if it is empty and −∞-\infty−∞ if it is unbounded below. No state assumptions or tensor-power relationships are imposed.

QAlgorithms.Papers.Independent.Support50.P21.strictPositive

For a finite index type III with decidable equality and any complex square matrix AAA indexed by III, this defines the real functional-calculus value

1(0,∞)(A).\mathbf 1_{(0,\infty)}(A).1(0,∞)​(A).

When AAA is Hermitian, it is the orthogonal projection onto the sum of its eigenspaces with strictly positive eigenvalues; zero eigenvalues are excluded. No Hermitian hypothesis is imposed, and the functional calculus returns 000 on non-Hermitian inputs. Its value at the zero matrix is 000.

QAlgorithms.Papers.Independent.Support50.P21.thresholdSet

For a finite index type III with decidable equality and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, this defines the set

{t∈R:t>0,Re⁡Tr⁡ ⁣(1[0,∞)(ρ−tσ)(ρ−tσ))=0}.\left\{t\in\mathbb R: t>0,\quad \operatorname{Re}\operatorname{Tr}\!\left( \mathbf 1_{[0,\infty)}(\rho-t\sigma)(\rho-t\sigma) \right)=0 \right\}.{t∈R:t>0,ReTr(1[0,∞)​(ρ−tσ)(ρ−tσ))=0}.

Here the indicator is applied by real functional calculus: on a Hermitian matrix it projects onto the nonnegative eigenspaces, including the kernel, and on a non-Hermitian matrix it returns 000. No positivity or normalization is assumed for ρ\rhoρ or σ\sigmaσ. Consequently, every positive ttt for which ρ−tσ\rho-t\sigmaρ−tσ is non-Hermitian satisfies the displayed trace condition; on an empty index type the set is all positive reals.

QAlgorithms.Papers.Independent.Support50.P21.traceSmoothMax

For a finite index type III with decidable equality, any real ε\varepsilonε, and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, this defines

inf⁡r⪰0Re⁡Tr⁡(r−ρ)†(r−ρ)≤εRe⁡Tr⁡r≤Re⁡Tr⁡ρ inf⁡{λ∈R:2λσ−r⪰0}.\inf_{\substack{r\succeq0\\ \operatorname{Re}\operatorname{Tr}\sqrt{(r-\rho)^\dagger(r-\rho)}\le\varepsilon\\ \operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rho}} \ \inf\left\{\lambda\in\mathbb R: 2^\lambda\sigma-r\succeq0\right\}.r⪰0ReTr(r−ρ)†(r−ρ)​≤εReTrr≤ReTrρ​inf​ inf{λ∈R:2λσ−r⪰0}.

Both infima are taken in the extended reals, with real λ\lambdaλ embedded there; an empty infimum is +∞+\infty+∞. No positivity or normalization is assumed for ρ,σ\rho,\sigmaρ,σ, and zero smoothing matrices are allowed. In particular, if σ⪰0\sigma\succeq0σ⪰0 and r=0r=0r=0 satisfies the outer constraints, the value is −∞-\infty−∞.

QAlgorithms.Papers.Independent.Support50.P21.traceSmoothMin

For a finite index type III with decidable equality, any real ε\varepsilonε, and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, this defines the extended-real supremum, over all r⪰0r\succeq0r⪰0 satisfying

Re⁡Tr⁡(r−ρ)†(r−ρ)≤ε,Re⁡Tr⁡r≤Re⁡Tr⁡ρ,\operatorname{Re}\operatorname{Tr}\sqrt{(r-\rho)^\dagger(r-\rho)} \le\varepsilon,\qquad \operatorname{Re}\operatorname{Tr}r\le\operatorname{Re}\operatorname{Tr}\rho,ReTr(r−ρ)†(r−ρ)​≤ε,ReTrr≤ReTrρ,

of the following value: writing PrP_rPr​ for the orthogonal projection onto the range of rrr, it is −log⁡2(Re⁡Tr⁡(Prσ))-\log_2(\operatorname{Re}\operatorname{Tr}(P_r\sigma))−log2​(ReTr(Pr​σ)) when that trace is strictly positive, and +∞+\infty+∞ otherwise. No positivity or normalization of ρ,σ\rho,\sigmaρ,σ is assumed. An empty outer supremum is −∞-\infty−∞; if r=0r=0r=0 meets the constraints, the value is +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P21.traceSmoothingBall

For a finite index type III with decidable equality, any real ε\varepsilonε, and an arbitrary complex square matrix ρ\rhoρ indexed by III, this defines the set of complex square matrices rrr satisfying

r⪰0,Re⁡Tr⁡(r−ρ)†(r−ρ)≤ε,Re⁡Tr⁡r≤Re⁡Tr⁡ρ.r\succeq0,\qquad \operatorname{Re}\operatorname{Tr}\sqrt{(r-\rho)^\dagger(r-\rho)} \le\varepsilon,\qquad \operatorname{Re}\operatorname{Tr}r \le\operatorname{Re}\operatorname{Tr}\rho.r⪰0,ReTr(r−ρ)†(r−ρ)​≤ε,ReTrr≤ReTrρ.

The square root is taken by real continuous functional calculus. Neither normalization nor positivity of ρ\rhoρ is required, and r=0r=0r=0 is permitted when it meets the inequalities. The set is empty if ε<0\varepsilon<0ε<0, or if Re⁡Tr⁡ρ<0\operatorname{Re}\operatorname{Tr}\rho<0ReTrρ<0.

QAlgorithms.Papers.Independent.Support50.P21.unnormalizedPetz

For a finite index type III with decidable equality, any real α\alphaα, and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, define Pt(X)P_t(X)Pt​(X) by real continuous functional calculus using the scalar function that is 000 at 000 and xtx^txt otherwise, with real powers interpreted as total real powers. Put x=Re⁡Tr⁡(Pα(ρ)P1−α(σ))x=\operatorname{Re}\operatorname{Tr}(P_\alpha(\rho)P_{1-\alpha}(\sigma))x=ReTr(Pα​(ρ)P1−α​(σ)). The defined extended-real value is

(α−1)−1{log⁡2x,x>0,−∞,x≤0.(\alpha-1)^{-1} \begin{cases} \log_2 x,&x>0,\\ -\infty,&x\le0. \end{cases}(α−1)−1{log2​x,−∞,​x>0,x≤0.​

There are no positivity, normalization, support, or parameter restrictions. Zero eigenvalues contribute 000 even when the exponent is 000; real functional calculus returns the zero matrix on non-Hermitian inputs. At α=1\alpha=1α=1, the reciprocal is 000, so the value is 000 under the extended-real convention 0⋅(±∞)=00\cdot(\pm\infty)=00⋅(±∞)=0.

QAlgorithms.Papers.Independent.Support50.P21.unnormalizedRelativeBits

For a finite index type III with decidable equality and arbitrary complex square matrices ρ,σ\rho,\sigmaρ,σ indexed by III, this defines

{Re⁡Tr⁡ ⁣[ρ((log⁡2)(ρ)−(log⁡2)(σ))],ker⁡σ⊆ker⁡ρ,+∞,otherwise,\begin{cases} \operatorname{Re}\operatorname{Tr}\!\left[ \rho\bigl((\log_2)(\rho)-(\log_2)(\sigma)\bigr) \right],&\ker\sigma\subseteq\ker\rho,\\ +\infty,&\text{otherwise}, \end{cases}{ReTr[ρ((log2​)(ρ)−(log2​)(σ))],+∞,​kerσ⊆kerρ,otherwise,​

with the finite real value embedded in the extended reals. Both matrix logarithms use real continuous functional calculus with the total scalar function log⁡2x=log⁡x/log⁡2\log_2 x=\log x/\log2log2​x=logx/log2, where log⁡0=0\log0=0log0=0 and the real logarithm of a negative argument equals the logarithm of its absolute value; this calculus returns 000 on non-Hermitian inputs. No positivity, nonzero, or normalization hypothesis is imposed, and there is no division by Tr⁡ρ\operatorname{Tr}\rhoTrρ. In particular, ρ=0\rho=0ρ=0 gives 000.

QAlgorithms.Papers.Independent.Support50.P21.weakPositive

For a finite index type III with decidable equality and any complex square matrix AAA indexed by III, this defines the real functional-calculus value

1[0,∞)(A).\mathbf 1_{[0,\infty)}(A).1[0,∞)​(A).

When AAA is Hermitian, it is the orthogonal projection onto the sum of its eigenspaces with nonnegative eigenvalues, including its kernel. No Hermitian hypothesis is imposed; on non-Hermitian inputs the functional calculus returns 000. In particular, the value at the zero matrix is the identity.

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).

  • Endorsed by Gohan · Oct 29, 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