Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Definitions for Tight bound on relative entropy by entropy difference

Definition
QAlgorithms_Papers_Independent_Support50_P15

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

prove2quantum

The objects the formalized statements of Tight bound on relative entropy by entropy difference (arXiv:1304.0036) are stated in. Each was checked against the paper by a blind read-back and a moderator.

  • EntropyState: Density matrices, including singular ones (PDF p.3, §1.1).
  • stateRelativeNats: Eq. (3), PDF p.3: audited Umegaki entropy, embedded into the signed extended reals.
  • difference: Eq. (2), PDF p.3, with the argument order of the source.
  • klTerm: One term of classical relative entropy; support failure is infinity (PDF p.3).
  • binaryRelative: Binary relative entropy, Eq. (5), PDF p.3, with endpoint limits.
  • massLimit: Largest admissible small-eigenvalue mass in Eq. (6), PDF p.4.
  • degenerateSpectrumEntropy: Entropy of the spectrum (10), PDF pp.3--4.
  • entropyDifferenceFeasible: The scalar feasible set in Eq. (6), PDF p.4.
  • tightBound: Eq. (6), PDF p.4. Infimum of the empty feasible set is +infinity (PDF p.21).
  • minimizingPair: Attainment of the actual scalar optimization (PDF pp.4,20--22).
  • familyProb: The probabilities in the explicitly computed witness (10), PDF p.4.
  • family: The diagonal witness (10), PDF p.4; meaningful for d>=2 and 0<=s<=1.
  • varianceObjective: The function maximized in Eq. (7), PDF p.4; the optimization excludes endpoints.
  • maxVariance: Eq. (7), PDF p.4. On d>=2 the value set is nonempty and bounded.
  • easyVariance: Eq. (8), PDF p.4.
  • optimalMass: The unique maximizing scalar, when d>=2 (PDF pp.8,25).
  • surprisalVariance: Eq. (19), PDF p.7: zero eigenvalues have zero sampling weight.
  • stateObservableVariance: Variance of a Hermitian observable (PDF pp.7--9).
  • exponentialBound: The exponential lower bound in (11), PDF p.4.
  • cubicBound: The cubic polynomial in (11), PDF p.4.
  • redundancyScale: The common denominator L in (46)--(54), PDF pp.14--16.
  • redundancyBound: The explicit lower bound in (51), (52), (54), PDF pp.14--16.
  • asymmetryConjecture: The unproved asymmetry conjecture in a fixed dimension (PDF pp.14--15).
  • continuityEntropy: h_d in (16), PDF p.6.
  • entropyInverse: Inverse on the increasing branch [0,(d-1)/d], PDF p.6, (18).
  • inverseEstimate: The explicit inverse estimate in (18), with its continuous extension at zero (PDF p.6).
  • thermalFinite: Gibbs state at a positive finite temperature, (23), PDF p.8.
  • groundSpaceDensity: Normalized projector onto the entire ground space, PDF p.9.
  • gibbsAtTemperature: Thermal states for all T in [0,infinity], with both endpoint conventions (PDF p.9).
  • thermalExpectedEnergy: Mean energy E(T), PDF p.9.
  • heatCapacity: Heat capacity defined by dE/dT, not by the variance identity (24), PDF p.9.
  • optimalHamiltonian: One fixed optimal Hamiltonian, PDF p.9, Remark 11.
  • fullRank: Full rank on the already restricted finite-dimensional support, PDF p.10.
  • EntropyPathPartition: An arbitrary strict partition, for (38)--(40), PDF p.12.
  • mesh: Mesh controls all subdivisions, not just uniform ones (PDF p.12).
  • piecewiseC1: Piecewise C1 means globally continuous with C1 restrictions on a finite subdivision (p.12).
  • entropyHeatSum: The right-endpoint Riemann sum in (38), PDF p.12.
  • partitionHeat: Actual heat divided by bath temperature, H_j=-T_j log rho(t_j), (27),(38), pp.10,12.
  • Process: The actual finite sequence of states and step Hamiltonians (PDF pp.10,13).
  • isIsothermal: Isothermal equilibration, with final Hamiltonian equal to the fixed external H (p.13).
  • processHeat: Heat is computed from the Hamiltonians and state increments (27), PDF pp.10,13.
  • extractedWork: Extracted work = heat - internal energy increase (41)--(42), PDF p.13.
  • helmholtzFreeEnergy: Free energy at the fixed Hamiltonian and temperature (41), PDF p.13.
  • availableWork: Available work computed from free energies, not defined as a process optimum (41), p.13.
  • extendedDomainInfimum: A minimum over a specified domain in the complete extended real order (pp.14--16).
  • attainsMinimum: Actual attainment, retained whenever the source writes min (PDF p.14).
  • minimaxRedundancy: Minimal redundancy (47), PDF p.14, over quantum states as also asserted on p.15.
  • minEntropy: Minimal entropy of the finite nonempty source family (PDF p.14).
  • sourceEntropyMaximum: Maximal entropy of the finite nonempty source family (PDF p.14).
  • entropyRadius: The intermediate objective in (48), PDF p.14.
  • isLaw: A normalized classical law in the paper's fixed basis (PDF p.3).
  • shannonEntropyNats: Natural-log Shannon entropy, matching the diagonal density matrix (PDF p.3).
  • FiniteOutputChannel: Classical channel, as in Proposition 13, PDF p.15; X need not be finite.
  • finiteOutputCapacityNats: Nats capacity: supremum of mutual information over all finite-support input laws (p.15). For finite output this covers arbitrary nonempty discrete input alphabets by approximation.
  • entropyConstraintDistance: Eq. (53), PDF p.16, with a strict entropy constraint and empty-set value infinity.
  • chernoffCoefficient: Chernoff coefficient infimum including one-sided endpoint limits, PDF p.16. Using 0<s<1 avoids the spurious 0^0=1 at singular endpoints.
  • chernoff: Classical Chernoff information; disjoint support gives infinity (54), PDF p.16.
Definition code
import Mathlib
import Definitions.Def_QAlgorithms_Defs_ExtendedQuantumEntropy
import Definitions.Def_QAlgorithms_Papers_Independent_Information100_P56

/-!
Reeb and Wolf, arXiv:1304.0036v3. All logarithms below are natural.
The accepted entropy in nats is imported from Information100.P56.
Distinct paper-local definitions have complete bodies; only result statements use `sorry`.
-/

noncomputable section

attribute [local instance] Classical.propDecidable

open scoped Matrix ComplexConjugate ENNReal Matrix.Norms.L2Operator
open Filter MeasureTheory

namespace QAlgorithms.Papers.Independent.Support50.P15

-- Exact accepted entropy in nats; no local redeclaration.
open QAlgorithms.Papers.Independent.Information100.P56 (entropy)

/-- Density matrices, including singular ones (PDF p.3, §1.1). -/
abbrev EntropyState (d : ℕ) := {a : Matrix (Fin d) (Fin d) ℂ // IsDensityMatrix a}

/-- Eq. (3), PDF p.3: audited Umegaki entropy, embedded into the signed extended reals. -/
def stateRelativeNats {d : ℕ} (a b : EntropyState d) : EReal :=
  (QuantumLoss.umegakiEntropy a.val b.val a.property b.property).toEReal

/-- Eq. (2), PDF p.3, with the argument order of the source. -/
def difference {d : ℕ} (a b : Matrix (Fin d) (Fin d) ℂ) : ℝ := entropy a - entropy b

/-- One term of classical relative entropy; support failure is infinity (PDF p.3). -/
def klTerm (p q : ℝ) : EReal :=
  if p = 0 then 0 else if q = 0 then ⊤ else (p * Real.log (p / q) : ℝ)

/-- Binary relative entropy, Eq. (5), PDF p.3, with endpoint limits. -/
def binaryRelative (s r : ℝ) : EReal := klTerm s r + klTerm (1 - s) (1 - r)

/-- Largest admissible small-eigenvalue mass in Eq. (6), PDF p.4. -/
def massLimit (d : ℕ) : ℝ := ((d : ℝ) - 1) / d

/-- Entropy of the spectrum (10), PDF pp.3--4. -/
def degenerateSpectrumEntropy (d : ℕ) (s : ℝ) : ℝ := Real.binEntropy s + s * Real.log ((d : ℝ) - 1)

/-- The scalar feasible set in Eq. (6), PDF p.4. -/
def entropyDifferenceFeasible (d : ℕ) (delta s r : ℝ) : Prop :=
  s ∈ Set.Icc 0 (massLimit d) ∧ r ∈ Set.Icc 0 (massLimit d) ∧
    degenerateSpectrumEntropy d s - degenerateSpectrumEntropy d r = delta

/-- Eq. (6), PDF p.4. Infimum of the empty feasible set is +infinity (PDF p.21). -/
def tightBound (d : ℕ) (delta : ℝ) : EReal :=
  sInf {v | ∃ s r : ℝ, entropyDifferenceFeasible d delta s r ∧ v = binaryRelative s r}

/-- Attainment of the actual scalar optimization (PDF pp.4,20--22). -/
def minimizingPair (d : ℕ) (delta s r : ℝ) : Prop :=
  entropyDifferenceFeasible d delta s r ∧ binaryRelative s r = tightBound d delta

/-- The probabilities in the explicitly computed witness (10), PDF p.4. -/
def familyProb (d : ℕ) (s : ℝ) (i : Fin d) : ℝ :=
  if i.val = 0 then 1 - s else s / ((d : ℝ) - 1)

/-- The diagonal witness (10), PDF p.4; meaningful for d>=2 and 0<=s<=1. -/
def family (d : ℕ) (s : ℝ) : Matrix (Fin d) (Fin d) ℂ :=
  Matrix.diagonal (fun i => (familyProb d s i : ℂ))

/-- The function maximized in Eq. (7), PDF p.4; the optimization excludes endpoints. -/
def varianceObjective (d : ℕ) (r : ℝ) : ℝ :=
  r * (1 - r) * (Real.log ((1 - r) / r * ((d : ℝ) - 1))) ^ 2

/-- Eq. (7), PDF p.4. On d>=2 the value set is nonempty and bounded. -/
def maxVariance (d : ℕ) : ℝ :=
  sSup (varianceObjective d '' Set.Ioo 0 (1 / 2))

/-- Eq. (8), PDF p.4. -/
def easyVariance (d : ℕ) : ℝ := (Real.log ((d : ℝ) - 1)) ^ 2 / 4 + 1

/-- The unique maximizing scalar, when d>=2 (PDF pp.8,25). -/
def optimalMass (d : ℕ) : ℝ :=
  Classical.epsilon (fun r : ℝ => r ∈ Set.Ioo 0 (1 / 2) ∧ varianceObjective d r = maxVariance d)

/-- Eq. (19), PDF p.7: zero eigenvalues have zero sampling weight. -/
def surprisalVariance {d : ℕ} (a : Matrix (Fin d) (Fin d) ℂ) : ℝ :=
  (cfc (fun x : ℝ => x * (Real.log x) ^ 2) a).trace.re - (entropy a) ^ 2

/-- Variance of a Hermitian observable (PDF pp.7--9). -/
def stateObservableVariance {d : ℕ} (a h : Matrix (Fin d) (Fin d) ℂ) : ℝ :=
  (a * h ^ 2).trace.re - ((a * h).trace.re) ^ 2

/-- The exponential lower bound in (11), PDF p.4. -/
def exponentialBound (n delta : ℝ) : ℝ := n * (Real.exp (delta / n) - 1 - delta / n)

/-- The cubic polynomial in (11), PDF p.4. -/
def cubicBound (n delta : ℝ) : ℝ := delta ^ 2 / (2 * n) + delta ^ 3 / (6 * n ^ 2)

/-- The common denominator L in (46)--(54), PDF pp.14--16. -/
def redundancyScale (d : ℕ) : ℝ := (Real.log ((d : ℝ) - 1)) ^ 2 + 4

/-- The explicit lower bound in (51), (52), (54), PDF pp.14--16. -/
def redundancyBound (d : ℕ) (a : ℝ) : ℝ :=
  a ^ 2 / (2 * redundancyScale d) - a ^ 3 / (3 * (redundancyScale d) ^ 2)

/-- The unproved asymmetry conjecture in a fixed dimension (PDF pp.14--15). -/
def asymmetryConjecture (d : ℕ) : Prop :=
  ∀ a ∈ Set.Icc 0 (Real.log d), tightBound d (-a) ≤ tightBound d a

/-- h_d in (16), PDF p.6. -/
def continuityEntropy (d : ℕ) (t : ℝ) : ℝ := t * Real.log ((d : ℝ) - 1) + Real.binEntropy t

/-- Inverse on the increasing branch [0,(d-1)/d], PDF p.6, (18). -/
def entropyInverse (d : ℕ) (a : ℝ) : ℝ :=
  sInf {t | t ∈ Set.Icc 0 (massLimit d) ∧ a ≤ continuityEntropy d t}

/-- The explicit inverse estimate in (18), with its continuous extension at zero (PDF p.6). -/
def inverseEstimate (d : ℕ) (a : ℝ) : ℝ :=
  if a = 0 then 0 else (Real.exp 1 - 1) / Real.exp 1 * a /
    (1 + Real.log ((d : ℝ) - 1) - Real.log a)

/-- Gibbs state at a positive finite temperature, (23), PDF p.8. -/
def thermalFinite {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ) : Matrix (Fin d) (Fin d) ℂ :=
  gibbsState (-((t : ℂ)⁻¹) • h)

/-- Normalized projector onto the entire ground space, PDF p.9. -/
def groundSpaceDensity {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) : Matrix (Fin d) (Fin d) ℂ :=
  if hd : 0 < d then
    let e := ascEigenvalues h ⟨0, by simpa only [Fintype.card_fin] using hd⟩
    let p := cfc (fun x : ℝ => if x = e then 1 else 0) h
    p.trace⁻¹ • p
  else 0

/-- Thermal states for all T in [0,infinity], with both endpoint conventions (PDF p.9). -/
def gibbsAtTemperature {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ≥0∞) : Matrix (Fin d) (Fin d) ℂ :=
  if t = ⊤ then (d : ℂ)⁻¹ • 1 else if t = 0 then groundSpaceDensity h else thermalFinite h t.toReal

/-- Mean energy E(T), PDF p.9. -/
def thermalExpectedEnergy {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ) : ℝ :=
  (h * thermalFinite h t).trace.re

/-- Heat capacity defined by dE/dT, not by the variance identity (24), PDF p.9. -/
def heatCapacity {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ≥0∞) : ℝ :=
  if t = 0 ∨ t = ⊤ then 0 else deriv (thermalExpectedEnergy h) t.toReal

/-- One fixed optimal Hamiltonian, PDF p.9, Remark 11. -/
def optimalHamiltonian (d : ℕ) : Matrix (Fin d) (Fin d) ℂ :=
  Matrix.diagonal (fun i => if i.val = 0 then -1 else 0)

/-- Full rank on the already restricted finite-dimensional support, PDF p.10. -/
def fullRank {d : ℕ} (a : Matrix (Fin d) (Fin d) ℂ) : Prop :=
  LinearMap.ker a.mulVecLin = ⊥

/-- An arbitrary strict partition, for (38)--(40), PDF p.12. -/
structure EntropyPathPartition where
  k : ℕ
  positive : 0 < k
  point : Fin (k + 1) → ℝ
  strict : StrictMono point
  first : point 0 = 0
  last : point (Fin.last k) = 1

/-- Mesh controls all subdivisions, not just uniform ones (PDF p.12). -/
def mesh (p : EntropyPathPartition) : ℝ :=
  sSup (Set.range (fun j : Fin p.k => p.point j.succ - p.point j.castSucc))

/-- Piecewise C1 means globally continuous with C1 restrictions on a finite subdivision (p.12). -/
def piecewiseC1 {d : ℕ} (a : ℝ → Matrix (Fin d) (Fin d) ℂ) : Prop :=
  ContinuousOn a (Set.Icc 0 1) ∧ ∃ p : EntropyPathPartition, ∀ j : Fin p.k,
    ContDiffOn ℝ 1 a (Set.Icc (p.point j.castSucc) (p.point j.succ))

/-- The right-endpoint Riemann sum in (38), PDF p.12. -/
def entropyHeatSum {d : ℕ} (a : ℝ → Matrix (Fin d) (Fin d) ℂ) (p : EntropyPathPartition) : ℝ :=
  ∑ j : Fin p.k, ((-cfc Real.log (a (p.point j.succ))) *
    (a (p.point j.succ) - a (p.point j.castSucc))).trace.re

/-- Actual heat divided by bath temperature, H_j=-T_j log rho(t_j), (27),(38), pp.10,12. -/
def partitionHeat {d : ℕ} (a : ℝ → Matrix (Fin d) (Fin d) ℂ) (p : EntropyPathPartition)
    (t : Fin p.k → ℝ) : ℝ :=
  ∑ j : Fin p.k, ((-((t j : ℂ)) • cfc Real.log (a (p.point j.succ))) *
    (a (p.point j.succ) - a (p.point j.castSucc))).trace.re / t j

/-- The actual finite sequence of states and step Hamiltonians (PDF pp.10,13). -/
structure Process (d : ℕ) where
  k : ℕ
  positive : 0 < k
  state : Fin (k + 1) → EntropyState d
  hamiltonian : Fin k → Matrix (Fin d) (Fin d) ℂ

/-- Isothermal equilibration, with final Hamiltonian equal to the fixed external H (p.13). -/
def isIsothermal {d : ℕ} (p : Process d) (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ)
    (a : EntropyState d) : Prop :=
  p.state 0 = a ∧ (∀ j, fullRank (p.state j).val) ∧
  (∀ j, (p.hamiltonian j).IsHermitian ∧
    (p.state j.succ).val = thermalFinite (p.hamiltonian j) t) ∧
  p.hamiltonian ⟨p.k - 1, Nat.sub_lt p.positive (by decide)⟩ = h

/-- Heat is computed from the Hamiltonians and state increments (27), PDF pp.10,13. -/
def processHeat {d : ℕ} (p : Process d) : ℝ :=
  ∑ j : Fin p.k, (p.hamiltonian j * ((p.state j.succ).val - (p.state j.castSucc).val)).trace.re

/-- Extracted work = heat - internal energy increase (41)--(42), PDF p.13. -/
def extractedWork {d : ℕ} (p : Process d) (h : Matrix (Fin d) (Fin d) ℂ) : ℝ :=
  processHeat p - (h * ((p.state (Fin.last p.k)).val - (p.state 0).val)).trace.re

/-- Free energy at the fixed Hamiltonian and temperature (41), PDF p.13. -/
def helmholtzFreeEnergy {d : ℕ} (h a : Matrix (Fin d) (Fin d) ℂ) (t : ℝ) : ℝ :=
  (h * a).trace.re - t * entropy a

/-- Available work computed from free energies, not defined as a process optimum (41), p.13. -/
def availableWork {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) (t : ℝ) (a : EntropyState d) : ℝ :=
  helmholtzFreeEnergy h a.val t - helmholtzFreeEnergy h (thermalFinite h t) t

/-- A minimum over a specified domain in the complete extended real order (pp.14--16). -/
def extendedDomainInfimum {α : Type*} (a : Set α) (f : α → EReal) : EReal := sInf (f '' a)

/-- Actual attainment, retained whenever the source writes min (PDF p.14). -/
def attainsMinimum {α : Type*} (a : Set α) (f : α → EReal) : Prop :=
  ∃ x ∈ a, f x = extendedDomainInfimum a f

/-- Minimal redundancy (47), PDF p.14, over quantum states as also asserted on p.15. -/
def minimaxRedundancy {d m : ℕ} (a : Fin m → EntropyState d) : EReal :=
  ⨅ b : EntropyState d, ⨆ j, stateRelativeNats (a j) b

/-- Minimal entropy of the finite nonempty source family (PDF p.14). -/
def minEntropy {d m : ℕ} (a : Fin m → EntropyState d) : ℝ := sInf (Set.range (fun j => entropy (a j).val))

/-- Maximal entropy of the finite nonempty source family (PDF p.14). -/
def sourceEntropyMaximum {d m : ℕ} (a : Fin m → EntropyState d) : ℝ := sSup (Set.range (fun j => entropy (a j).val))

/-- The intermediate objective in (48), PDF p.14. -/
def entropyRadius {d m : ℕ} (a : Fin m → EntropyState d) (s : ℝ) : EReal :=
  ⨆ j, tightBound d (entropy (a j).val - s)

/-- A normalized classical law in the paper's fixed basis (PDF p.3). -/
def isLaw {d : ℕ} (p : Fin d → ℝ) : Prop := (∀ i, 0 ≤ p i) ∧ ∑ i, p i = 1

/-- Natural-log Shannon entropy, matching the diagonal density matrix (PDF p.3). -/
def shannonEntropyNats {d : ℕ} (p : Fin d → ℝ) : ℝ := -∑ i, p i * Real.log (p i)

/-- Classical channel, as in Proposition 13, PDF p.15; X need not be finite. -/
structure FiniteOutputChannel (X : Type*) (d : ℕ) where
  prob : X → Fin d → ℝ
  normalized : ∀ x, isLaw (prob x)

/-- Nats capacity: supremum of mutual information over all finite-support input laws (p.15).
For finite output this covers arbitrary nonempty discrete input alphabets by approximation. -/
def finiteOutputCapacityNats {X : Type*} {d : ℕ} (c : FiniteOutputChannel X d) : ℝ :=
  sSup {v | ∃ (m : ℕ) (x : Fin m → X) (q : Fin m → ℝ), isLaw q ∧
    v = shannonEntropyNats (fun y => ∑ j, q j * c.prob (x j) y) -
      ∑ j, q j * shannonEntropyNats (c.prob (x j))}

/-- Eq. (53), PDF p.16, with a strict entropy constraint and empty-set value infinity. -/
def entropyConstraintDistance {d : ℕ} (r : ℝ) (a : EntropyState d) : EReal :=
  ⨅ b : {b : EntropyState d // r < entropy b.val}, stateRelativeNats b.val a

/-- Chernoff coefficient infimum including one-sided endpoint limits, PDF p.16.
Using 0<s<1 avoids the spurious 0^0=1 at singular endpoints. -/
def chernoffCoefficient {d : ℕ} (p q : Fin d → ℝ) : ℝ :=
  sInf ((fun s : ℝ => ∑ i, (p i) ^ s * (q i) ^ (1 - s)) '' Set.Ioo 0 1)

/-- Classical Chernoff information; disjoint support gives infinity (54), PDF p.16. -/
def chernoff {d : ℕ} (p q : Fin d → ℝ) : EReal :=
  let z := chernoffCoefficient p q
  if z = 0 then ⊤ else (-Real.log z : ℝ)

/-! Definition-level boundary sanity tests. -/

example {d : ℕ} : entropy (1 : Matrix (Fin d) (Fin d) ℂ) = 0 := by
  simp [entropy, cfc_one]

example (q : ℝ) : klTerm 0 q = 0 := by simp [klTerm]

example : binaryRelative 0 0 = 0 := by simp [binaryRelative, klTerm]

example {s : ℝ} (hs : s ≠ 0) : klTerm s 0 = ⊤ := by simp [klTerm, hs]

example : familyProb 2 (1 / 2) 0 = 1 / 2 := by norm_num [familyProb]

example (d : ℕ) : degenerateSpectrumEntropy d 0 = 0 := by simp [degenerateSpectrumEntropy]

example : easyVariance 2 = 1 := by norm_num [easyVariance]

example (n : ℝ) : exponentialBound n 0 = 0 ∧ cubicBound n 0 = 0 := by
  simp [exponentialBound, cubicBound]

example (d : ℕ) : inverseEstimate d 0 = 0 := by simp [inverseEstimate]

example {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) :
    gibbsAtTemperature h ⊤ = (d : ℂ)⁻¹ • 1 ∧ gibbsAtTemperature h 0 = groundSpaceDensity h := by simp [gibbsAtTemperature]

example {d : ℕ} (h : Matrix (Fin d) (Fin d) ℂ) :
    heatCapacity h 0 = 0 ∧ heatCapacity h ⊤ = 0 := by simp [heatCapacity]

example {α : Type*} (f : α → EReal) : extendedDomainInfimum ∅ f = ⊤ := by simp [extendedDomainInfimum]

example (d : ℕ) : redundancyBound d 0 = 0 := by simp [redundancyBound]

open scoped ComplexOrder in
example : IsDensityMatrix (1 : Matrix (Fin 1) (Fin 1) ℂ) :=
  ⟨Matrix.PosSemidef.one, by simp⟩

example {d : ℕ} (a : EntropyState d) : difference a.val a.val = 0 := by
  simp [difference]

example : surprisalVariance (1 : Matrix (Fin 1) (Fin 1) ℂ) = 0 := by
  simp [surprisalVariance, entropy, cfc_one]

example {d : ℕ} (a : Matrix (Fin d) (Fin d) ℂ) (p : EntropyPathPartition) :
    entropyHeatSum (fun _ => a) p = 0 := by
  simp [entropyHeatSum]

example : isLaw (fun _ : Fin 1 => (1 : ℝ)) := by simp [isLaw]

example : ¬ isLaw (fun i : Fin 0 => (Fin.elim0 i : ℝ)) := by simp [isLaw]

example : shannonEntropyNats (fun _ : Fin 1 => (1 : ℝ)) = 0 := by
  simp [shannonEntropyNats]

end QAlgorithms.Papers.Independent.Support50.P15
Source
Tight bound on relative entropy by entropy difference, arXiv:1304.0036v3, https://arxiv.org/abs/1304.0036v3
Read-back

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

QAlgorithms.Papers.Independent.Support50.P15.EntropyPathPartition

This structure consists of a natural number kkk, a proof that k>0k>0k>0, and real points p0,…,pkp_0,\ldots,p_kp0​,…,pk​ satisfying pi<pjp_i<p_jpi​<pj​ whenever i<ji<ji<j, together with p0=0p_0=0p0​=0 and pk=1p_k=1pk​=1. Thus it specifies finitely many strictly increasing partition points with at least one interval; k=0k=0k=0 is excluded.

QAlgorithms.Papers.Independent.Support50.P15.EntropyState

For every natural number ddd, EntropyState(d)\mathrm{EntropyState}(d)EntropyState(d) is the type of complex d×dd\times dd×d matrices aaa, indexed by {0,…,d−1}\{0,\ldots,d-1\}{0,…,d−1}, together with the requirements that aaa is positive semidefinite and Tr⁡(a)=1\operatorname{Tr}(a)=1Tr(a)=1. When d=0d=0d=0, this type is empty, since the empty matrix has trace zero.

QAlgorithms.Papers.Independent.Support50.P15.FiniteOutputChannel

For any type XXX and natural number ddd, this structure consists of a function c:X×{0,…,d−1}→Rc:X\times\{0,\ldots,d-1\}\to\mathbb Rc:X×{0,…,d−1}→R together with the requirements that, for every x∈Xx\in Xx∈X,

c(x,i)≥0 for every i,∑i=0d−1c(x,i)=1.c(x,i)\ge0\ \text{for every }i, \qquad \sum_{i=0}^{d-1}c(x,i)=1.c(x,i)≥0 for every i,i=0∑d−1​c(x,i)=1.

There is no finiteness or nonemptiness requirement on XXX. If d=0d=0d=0 and XXX is nonempty, these requirements cannot be satisfied; if XXX is empty, they are vacuous.

QAlgorithms.Papers.Independent.Support50.P15.Process

For a natural number ddd, a process consists of a natural number k>0k>0k>0, a sequence ρ0,…,ρk\rho_0,\ldots,\rho_kρ0​,…,ρk​ of complex d×dd\times dd×d positive semidefinite matrices of trace one, and a sequence H0,…,Hk−1H_0,\ldots,H_{k-1}H0​,…,Hk−1​ of arbitrary complex d×dd\times dd×d matrices. No Hermiticity or other condition is imposed on the HjH_jHj​, and no relation between successive states is required. For d=0d=0d=0, no such process exists because a matrix of trace one does not exist.

QAlgorithms.Papers.Independent.Support50.P15.asymmetryConjecture

For every natural number ddd, this defines the proposition that, for every real aaa with 0≤a≤log⁡d0\le a\le\log d0≤a≤logd, one has Td(−a)≤Td(a)T_d(-a)\le T_d(a)Td​(−a)≤Td​(a) in the extended real numbers, where

Td(δ)=inf⁡{K(s,r)+K(1−s,1−r):0≤s,r≤d−1d, Hd(s)−Hd(r)=δ},T_d(\delta)= \inf\left\{ K(s,r)+K(1-s,1-r): 0\le s,r\le\frac{d-1}{d},\ H_d(s)-H_d(r)=\delta \right\},Td​(δ)=inf{K(s,r)+K(1−s,1−r):0≤s,r≤dd−1​, Hd​(s)−Hd​(r)=δ}, Hd(u)=−ulog⁡u−(1−u)log⁡(1−u)+ulog⁡(d−1),K(p,q)={0,p=0,+∞,p≠0 and q=0,plog⁡(p/q),p≠0 and q≠0.H_d(u)=-u\log u-(1-u)\log(1-u)+u\log(d-1), \qquad K(p,q)= \begin{cases} 0,&p=0,\\ +\infty,&p\ne0\text{ and }q=0,\\ p\log(p/q),&p\ne0\text{ and }q\ne0. \end{cases}Hd​(u)=−ulogu−(1−u)log(1−u)+ulog(d−1),K(p,q)=⎩⎨⎧​0,+∞,plog(p/q),​p=0,p=0 and q=0,p=0 and q=0.​

Finite values of KKK are regarded as extended real numbers, and the infimum of an empty feasible set is +∞+\infty+∞. All subtraction involving ddd here is real subtraction; division by zero is zero, log⁡0=0\log0=0log0=0, and the logarithm of a negative real is the logarithm of its absolute value. For d=0d=0d=0 or d=1d=1d=1, the quantified interval for aaa consists only of zero. This declaration defines the proposition; it does not assert that it holds.

QAlgorithms.Papers.Independent.Support50.P15.attainsMinimum

For any type α\alphaα, any subset A⊆αA\subseteq\alphaA⊆α, and any extended-real-valued function f:α→[−∞,+∞]f:\alpha\to[-\infty,+\infty]f:α→[−∞,+∞], this predicate states that there exists x∈Ax\in Ax∈A such that

f(x)=inf⁡{f(y):y∈A}.f(x)=\inf\{f(y):y\in A\}.f(x)=inf{f(y):y∈A}.

The attained value may be either infinity. The predicate is false when AAA is empty.

QAlgorithms.Papers.Independent.Support50.P15.availableWork

For any natural number ddd, any complex d×dd\times dd×d matrix hhh, any real number ttt, and any positive semidefinite trace-one matrix aaa, set

τ=(tr⁡exp⁡(−t−1h))−1exp⁡(−t−1h),f(x)=xlog⁡x.\tau=\bigl(\operatorname{tr}\exp(-t^{-1}h)\bigr)^{-1}\exp(-t^{-1}h), \qquad f(x)=x\log x.τ=(trexp(−t−1h))−1exp(−t−1h),f(x)=xlogx.

This defines

Re⁡tr⁡(ha)+t Re⁡tr⁡f(a)−Re⁡tr⁡(hτ)−t Re⁡tr⁡f(τ),\operatorname{Re}\operatorname{tr}(ha) +t\,\operatorname{Re}\operatorname{tr}f(a) -\operatorname{Re}\operatorname{tr}(h\tau) -t\,\operatorname{Re}\operatorname{tr}f(\tau),Retr(ha)+tRetrf(a)−Retr(hτ)−tRetrf(τ),

using real functional calculus for fff, the convention log⁡0=0\log 0=0log0=0, and total scalar inversion, including inverse zero equal to zero. No Hermiticity of hhh or positivity of ttt is assumed; at t=0t=0t=0, τ=Id/d\tau=I_d/dτ=Id​/d. The argument aaa cannot exist for d=0d=0d=0.

QAlgorithms.Papers.Independent.Support50.P15.binaryRelative

For arbitrary real numbers s,rs,rs,r, this defines

binaryRelative(s,r)=K(s,r)+K(1−s,1−r),K(p,q)={0,p=0,+∞,p≠0 and q=0,plog⁡(p/q),p≠0 and q≠0,\mathrm{binaryRelative}(s,r)=K(s,r)+K(1-s,1-r), \qquad K(p,q)= \begin{cases} 0,&p=0,\\ +\infty,&p\ne0\ \text{and}\ q=0,\\ p\log(p/q),&p\ne0\ \text{and}\ q\ne0, \end{cases}binaryRelative(s,r)=K(s,r)+K(1−s,1−r),K(p,q)=⎩⎨⎧​0,+∞,plog(p/q),​p=0,p=0 and q=0,p=0 and q=0,​

with addition in the extended reals and the natural logarithm, interpreted on negative arguments as log⁡∣x∣\log|x|log∣x∣. Neither sss nor rrr is restricted to [0,1][0,1][0,1]. In particular, the value is +∞+\infty+∞ if r=0r=0r=0 and s≠0s\ne0s=0, or if r=1r=1r=1 and s≠1s\ne1s=1; it is zero at (0,0)(0,0)(0,0) and (1,1)(1,1)(1,1).

QAlgorithms.Papers.Independent.Support50.P15.chernoff

For any natural number ddd and arbitrary real-valued vectors p,qp,qp,q indexed by 0,…,d−10,\ldots,d-10,…,d−1, let

z=inf⁡0<s<1∑i=0d−1pi sqi 1−s,z=\inf_{0<s<1}\sum_{i=0}^{d-1}p_i^{\,s}q_i^{\,1-s},z=0<s<1inf​i=0∑d−1​pis​qi1−s​,

using the total real-power operation and the real infimum, which returns 000 for a set unbounded below. This defines an extended-real value equal to +∞+\infty+∞ if z=0z=0z=0, and otherwise equal to the finite real number −log⁡z-\log z−logz. No nonnegativity or normalization assumptions are imposed on p,qp,qp,q; if z<0z<0z<0, the total real logarithm uses log⁡z=log⁡∣z∣\log z=\log|z|logz=log∣z∣. In particular, when d=0d=0d=0, the value is +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P15.chernoffCoefficient

For any natural number ddd and arbitrary real-valued vectors p,qp,qp,q indexed by 0,…,d−10,\ldots,d-10,…,d−1, this defines the real infimum

inf⁡0<s<1 ∑i=0d−1pi sqi 1−s.\inf_{0<s<1}\ \sum_{i=0}^{d-1}p_i^{\,s}q_i^{\,1-s}.0<s<1inf​ i=0∑d−1​pis​qi1−s​.

Neither nonnegativity nor normalization of ppp or qqq is assumed. Powers are the total real-power operation, including its defined values for zero or negative bases. The endpoints s=0s=0s=0 and s=1s=1s=1 are excluded, and no attainment is asserted. When d=0d=0d=0, every sum is 000, so the coefficient is 000; if the set of values is unbounded below, the real infimum operation also returns 000.

QAlgorithms.Papers.Independent.Support50.P15.continuityEntropy

For every natural number ddd and every real number ttt, this defines

tlog⁡(d−1)−tlog⁡t−(1−t)log⁡(1−t).t\log(d-1)-t\log t-(1-t)\log(1-t).tlog(d−1)−tlogt−(1−t)log(1−t).

There is no restriction that d≥2d\ge2d≥2 or that t∈[0,1]t\in[0,1]t∈[0,1]. The subtraction d−1d-1d−1 is real subtraction, and every logarithm uses the total real convention log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx.

QAlgorithms.Papers.Independent.Support50.P15.cubicBound

For all real numbers nnn and δ\deltaδ, with no sign or nonzero restrictions, this defines

δ22n+δ36n2.\frac{\delta^2}{2n}+\frac{\delta^3}{6n^2}.2nδ2​+6n2δ3​.

Division by zero is defined as zero, so the value is zero when n=0n=0n=0, for every δ\deltaδ.

QAlgorithms.Papers.Independent.Support50.P15.degenerateSpectrumEntropy

For every natural number ddd and arbitrary real number sss, this defines

degenerateSpectrumEntropy(d,s)=−slog⁡s−(1−s)log⁡(1−s)+slog⁡(d−1).\mathrm{degenerateSpectrumEntropy}(d,s) =-s\log s-(1-s)\log(1-s)+s\log(d-1).degenerateSpectrumEntropy(d,s)=−slogs−(1−s)log(1−s)+slog(d−1).

All logarithms are natural and use the total real logarithm, with log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. No restriction on sss or lower bound on ddd is imposed; at both d=0d=0d=0 and d=1d=1d=1, the final term is zero.

QAlgorithms.Papers.Independent.Support50.P15.difference

For every natural number ddd and arbitrary complex d×dd\times dd×d matrices a,ba,ba,b, this defines the real number

difference(a,b)=−Re⁡Tr⁡(f(a))+Re⁡Tr⁡(f(b)),f(x)=xlog⁡x,\mathrm{difference}(a,b) =-\operatorname{Re}\operatorname{Tr}(f(a)) +\operatorname{Re}\operatorname{Tr}(f(b)), \qquad f(x)=x\log x,difference(a,b)=−ReTr(f(a))+ReTr(f(b)),f(x)=xlogx,

where f(a)f(a)f(a) and f(b)f(b)f(b) use continuous functional calculus and log⁡\loglog is the natural logarithm with log⁡0=0\log 0=0log0=0. Neither matrix is required to be Hermitian, positive semidefinite, or of trace one. For d=0d=0d=0, the value is zero.

QAlgorithms.Papers.Independent.Support50.P15.easyVariance

For every natural number ddd, including 000 and 111, this defines

[log⁡(d−1)]24+1,\frac{[\log(d-1)]^2}{4}+1,4[log(d−1)]2​+1,

where d−1d-1d−1 is subtraction in the real numbers. The real logarithm satisfies log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx, so the value is 111 when d=0d=0d=0 or d=1d=1d=1.

QAlgorithms.Papers.Independent.Support50.P15.entropyConstraintDistance

For any natural number ddd, real number rrr, and complex d×dd\times dd×d positive semidefinite matrix aaa of trace 111, this defines

inf⁡b positive semidefinitetr⁡b=1, S(b)>rD(b∥a)\inf_{\substack{b\ \text{positive semidefinite}\\ \operatorname{tr}b=1,\ S(b)>r}} D(b\Vert a)b positive semidefinitetrb=1, S(b)>r​inf​D(b∥a)

in the extended reals. Here S(b)=−Re⁡tr⁡(f(b))S(b)=-\operatorname{Re}\operatorname{tr}(f(b))S(b)=−Retr(f(b)) with f(x)=xlog⁡xf(x)=x\log xf(x)=xlogx, and

D(b∥a)={max⁡ ⁣{0,Re⁡tr⁡(f(b)−blog⁡a)},ker⁡a⊆ker⁡b,+∞,ker⁡a⊈ker⁡b,D(b\Vert a)= \begin{cases} \max\!\left\{0,\operatorname{Re}\operatorname{tr}\bigl(f(b)-b\log a\bigr)\right\}, &\ker a\subseteq\ker b,\\ +\infty,&\ker a\not\subseteq\ker b, \end{cases}D(b∥a)={max{0,Retr(f(b)−bloga)},+∞,​kera⊆kerb,kera⊆kerb,​

with matrix functions interpreted by continuous functional calculus and log⁡0=0\log0=0log0=0. The entropy constraint is strict. If no such bbb exists, the infimum is +∞+\infty+∞. The parameter d=0d=0d=0 is allowed syntactically, but no input state aaa then exists.

QAlgorithms.Papers.Independent.Support50.P15.entropyDifferenceFeasible

For every natural number ddd and real numbers δ,s,r\delta,s,rδ,s,r, this predicate means precisely

0≤s≤d−1d,0≤r≤d−1d,hd(s)−hd(r)=δ,0\le s\le\frac{d-1}{d},\qquad 0\le r\le\frac{d-1}{d},\qquad h_d(s)-h_d(r)=\delta,0≤s≤dd−1​,0≤r≤dd−1​,hd​(s)−hd​(r)=δ,

where hd(u)=−ulog⁡u−(1−u)log⁡(1−u)+ulog⁡(d−1)h_d(u)=-u\log u-(1-u)\log(1-u)+u\log(d-1)hd​(u)=−ulogu−(1−u)log(1−u)+ulog(d−1). Subtraction in d−1d-1d−1 is real subtraction, division by zero is zero, and the natural logarithm has log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. In particular, when d=0d=0d=0 or d=1d=1d=1, the predicate holds exactly when s=r=0s=r=0s=r=0 and δ=0\delta=0δ=0.

QAlgorithms.Papers.Independent.Support50.P15.entropyHeatSum

For any natural number ddd, any function a:R→Md(C)a:\mathbb R\to M_d(\mathbb C)a:R→Md​(C), and any partition ppp consisting of a positive integer kkk and strictly increasing real points x0,…,xkx_0,\ldots,x_kx0​,…,xk​ with x0=0x_0=0x0​=0 and xk=1x_k=1xk​=1, this defines the real number

∑j=0k−1Re⁡tr⁡ ⁣[−log⁡ ⁣(a(xj+1))(a(xj+1)−a(xj))].\sum_{j=0}^{k-1}\operatorname{Re}\operatorname{tr}\!\left[ -\log\!\bigl(a(x_{j+1})\bigr)\bigl(a(x_{j+1})-a(x_j)\bigr) \right].j=0∑k−1​Retr[−log(a(xj+1​))(a(xj+1​)−a(xj​))].

Here the matrix logarithm means the real functional calculus applied to the total real logarithm, for which log⁡0=0\log 0=0log0=0. No continuity, positivity, Hermiticity, normalization, or invertibility is required of aaa. The dimension d=0d=0d=0 is included, in which case the value is zero.

QAlgorithms.Papers.Independent.Support50.P15.entropyInverse

For every natural number ddd, including 000, and real number aaa, this defines

entropyInverse⁡(d,a)=inf⁡{t∈R:0≤t≤d−1d,a≤tlog⁡(d−1)−tlog⁡t−(1−t)log⁡(1−t)}.\operatorname{entropyInverse}(d,a) =\inf\left\{t\in\mathbb R: 0\le t\le\frac{d-1}{d},\quad a\le t\log(d-1)-t\log t-(1-t)\log(1-t)\right\}.entropyInverse(d,a)=inf{t∈R:0≤t≤dd−1​,a≤tlog(d−1)−tlogt−(1−t)log(1−t)}.

Here d−1d-1d−1 is subtraction in R\mathbb RR, division by zero is 000, and the real logarithm satisfies log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. There are no restrictions on aaa or ddd. The infimum of an empty set is assigned the real value 000; in particular, the definition gives 000 for both d=0d=0d=0 and d=1d=1d=1, for every aaa.

QAlgorithms.Papers.Independent.Support50.P15.entropyRadius

For natural numbers d,md,md,m, a family (aj)j=0m−1(a_j)_{j=0}^{m-1}(aj​)j=0m−1​ of complex d×dd\times dd×d positive semidefinite matrices of trace 111, and any real number sss, this defines the extended-real quantity

sup⁡0≤j<mBd(S(aj)−s).\sup_{0\le j<m} B_d\bigl(S(a_j)-s\bigr).0≤j<msup​Bd​(S(aj​)−s).

Here S(ρ)=−Re⁡tr⁡(f(ρ))S(\rho)=-\operatorname{Re}\operatorname{tr}(f(\rho))S(ρ)=−Retr(f(ρ)), with f(x)=xlog⁡xf(x)=x\log xf(x)=xlogx applied by continuous functional calculus, and

Bd(δ)=inf⁡{K(u,v):0≤u,v≤d−1d, Hd(u)−Hd(v)=δ},Hd(u)=binEntropy⁡(u)+ulog⁡(d−1),B_d(\delta)= \inf\left\{ K(u,v): 0\le u,v\le \frac{d-1}{d},\ H_d(u)-H_d(v)=\delta \right\}, \qquad H_d(u)=\operatorname{binEntropy}(u)+u\log(d-1),Bd​(δ)=inf{K(u,v):0≤u,v≤dd−1​, Hd​(u)−Hd​(v)=δ},Hd​(u)=binEntropy(u)+ulog(d−1),

where d−1d-1d−1 in these real expressions is real subtraction, binEntropy⁡(u)=−ulog⁡u−(1−u)log⁡(1−u)\operatorname{binEntropy}(u)=-u\log u-(1-u)\log(1-u)binEntropy(u)=−ulogu−(1−u)log(1−u), and

K(u,v)=k(u,v)+k(1−u,1−v),k(p,q)={0,p=0,+∞,p≠0, q=0,plog⁡(p/q),p≠0, q≠0.K(u,v)=k(u,v)+k(1-u,1-v), \qquad k(p,q)= \begin{cases} 0,&p=0,\\ +\infty,&p\ne0,\ q=0,\\ p\log(p/q),&p\ne0,\ q\ne0. \end{cases}K(u,v)=k(u,v)+k(1−u,1−v),k(p,q)=⎩⎨⎧​0,+∞,plog(p/q),​p=0,p=0, q=0,p=0, q=0.​

The infimum and outer supremum are in the extended reals: an empty feasible set gives Bd(δ)=+∞B_d(\delta)=+\inftyBd​(δ)=+∞, and m=0m=0m=0 gives an entropy radius of −∞-\infty−∞. Real division by 000 and log⁡0\log0log0 both have value 000. The parameters include d=0d=0d=0, for which a family of the required states can exist only if m=0m=0m=0.

QAlgorithms.Papers.Independent.Support50.P15.exponentialBound

For all real numbers nnn and δ\deltaδ, with no sign or nonzero restrictions, this defines

n(eδ/n−1−δn).n\left(e^{\delta/n}-1-\frac{\delta}{n}\right).n(eδ/n−1−nδ​).

Division by zero is defined as zero, so the value is zero when n=0n=0n=0, for every δ\deltaδ.

QAlgorithms.Papers.Independent.Support50.P15.extendedDomainInfimum

For any type α\alphaα, any subset A⊆αA\subseteq\alphaA⊆α, and any function f:α→[−∞,+∞]f:\alpha\to[-\infty,+\infty]f:α→[−∞,+∞], this defines the extended-real infimum

inf⁡{f(x):x∈A}.\inf\{f(x):x\in A\}.inf{f(x):x∈A}.

The set AAA may be empty, in which case the value is +∞+\infty+∞; no attainment or finiteness is required.

QAlgorithms.Papers.Independent.Support50.P15.extractedWork

For any natural number ddd, any process ppp consisting of k>0k>0k>0, positive semidefinite trace-one matrices ρ0,…,ρk\rho_0,\ldots,\rho_kρ0​,…,ρk​, and arbitrary complex matrices H0,…,Hk−1H_0,\ldots,H_{k-1}H0​,…,Hk−1​, and any further complex d×dd\times dd×d matrix hhh, this defines

∑j=0k−1Re⁡tr⁡ ⁣[Hj(ρj+1−ρj)]−Re⁡tr⁡ ⁣[h(ρk−ρ0)].\sum_{j=0}^{k-1}\operatorname{Re}\operatorname{tr}\!\left[H_j(\rho_{j+1}-\rho_j)\right] -\operatorname{Re}\operatorname{tr}\!\left[h(\rho_k-\rho_0)\right].j=0∑k−1​Retr[Hj​(ρj+1​−ρj​)]−Retr[h(ρk​−ρ0​)].

There is no requirement that h=Hk−1h=H_{k-1}h=Hk−1​, that these matrices be Hermitian, or that the process satisfy the isothermal predicate. A process cannot exist when d=0d=0d=0.

QAlgorithms.Papers.Independent.Support50.P15.family

For every natural number ddd and arbitrary real number sss, this defines the complex d×dd\times dd×d diagonal matrix

family(d,s)ij={1−s,i=j=0,sd−1,i=j≠0,0,i≠j.\mathrm{family}(d,s)_{ij}= \begin{cases} 1-s,&i=j=0,\\ \dfrac{s}{d-1},&i=j\ne0,\\ 0,&i\ne j. \end{cases}family(d,s)ij​=⎩⎨⎧​1−s,d−1s​,0,​i=j=0,i=j=0,i=j.​

The real entries are embedded into the complex numbers, and no positivity or trace-one requirement is imposed. For d=0d=0d=0, this is the empty matrix; for d=1d=1d=1, it is the one-entry matrix [1−s][1-s][1−s].

QAlgorithms.Papers.Independent.Support50.P15.familyProb

For every natural number ddd, arbitrary real number sss, and index i∈{0,…,d−1}i\in\{0,\ldots,d-1\}i∈{0,…,d−1}, this defines

familyProb(d,s,i)={1−s,i=0,sd−1,i≠0.\mathrm{familyProb}(d,s,i)= \begin{cases} 1-s,&i=0,\\ \dfrac{s}{d-1},&i\ne0. \end{cases}familyProb(d,s,i)={1−s,d−1s​,​i=0,i=0.​

There is no restriction on sss, so these values are not required to form a probability distribution. When d=0d=0d=0, there is no index iii; when d=1d=1d=1, only the first branch occurs, giving 1−s1-s1−s.

QAlgorithms.Papers.Independent.Support50.P15.finiteOutputCapacityNats

For any type XXX, natural number ddd, and function c(x,y)c(x,y)c(x,y) whose values are nonnegative and sum to 111 over 0≤y<d0\le y<d0≤y<d for each x∈Xx\in Xx∈X, this defines the real supremum of

H ⁣((∑j=0m−1qjc(xj,y))y=0d−1)−∑j=0m−1qjH ⁣((c(xj,y))y=0d−1)H\!\left(\left(\sum_{j=0}^{m-1}q_jc(x_j,y)\right)_{y=0}^{d-1}\right) -\sum_{j=0}^{m-1}q_j H\!\left((c(x_j,y))_{y=0}^{d-1}\right)H​(j=0∑m−1​qj​c(xj​,y))y=0d−1​​−j=0∑m−1​qj​H((c(xj​,y))y=0d−1​)

over every natural number mmm, every list of inputs (xj)j=0m−1(x_j)_{j=0}^{m-1}(xj​)j=0m−1​ from XXX, and every real vector qqq satisfying qj≥0q_j\ge0qj​≥0 and ∑jqj=1\sum_jq_j=1∑j​qj​=1, where H(p)=−∑ipilog⁡piH(p)=-\sum_i p_i\log p_iH(p)=−∑i​pi​logpi​, using natural logarithms and the convention log⁡0=0\log0=0log0=0. Inputs in the list may repeat. Although m=0m=0m=0 is quantified over, it supplies no candidate because its normalization condition would be 0=10=10=1. If XXX is empty, the candidate set is empty and this real supremum is 000.

QAlgorithms.Papers.Independent.Support50.P15.fullRank

For every natural number ddd and complex d×dd\times dd×d matrix aaa, this predicate means that the kernel of the linear map v↦avv\mapsto avv↦av is the zero subspace; equivalently, av=0av=0av=0 implies v=0v=0v=0 for every v∈Cdv\in\mathbb C^dv∈Cd. No Hermitian or density-matrix assumption is required. For d=0d=0d=0, the predicate holds for the unique empty matrix.

QAlgorithms.Papers.Independent.Support50.P15.gibbsAtTemperature

For every natural number ddd, arbitrary complex d×dd\times dd×d matrix hhh, and extended nonnegative real t∈[0,∞]t\in[0,\infty]t∈[0,∞], this defines d−1Idd^{-1}I_dd−1Id​ when t=∞t=\inftyt=∞; when t=0t=0t=0, it defines zero if d=0d=0d=0, and otherwise (Tr⁡p)−1p(\operatorname{Tr}p)^{-1}p(Trp)−1p, where ppp is the real functional calculus of hhh for the indicator of its smallest eigenvalue if hhh is Hermitian, and is zero if hhh is non-Hermitian. For finite positive ttt, it defines

(Tr⁡exp⁡(−t−1h))−1exp⁡(−t−1h).\bigl(\operatorname{Tr}\exp(-t^{-1}h)\bigr)^{-1}\exp(-t^{-1}h).(Trexp(−t−1h))−1exp(−t−1h).

All scalar inverses are total, with 0−1=00^{-1}=00−1=0; no Hermitian assumption is imposed, and d=0d=0d=0 is included.

QAlgorithms.Papers.Independent.Support50.P15.groundSpaceDensity

For every natural number ddd and arbitrary complex d×dd\times dd×d matrix hhh, this defines the zero matrix when d=0d=0d=0. When d>0d>0d>0, let eee be the first entry of the ascending eigenvalue list of hhh: for Hermitian hhh, this is its smallest eigenvalue, while for non-Hermitian hhh, that list is defined to be identically zero. Apply the real functional calculus to hhh with the function x↦1x\mapsto1x↦1 if x=ex=ex=e and x↦0x\mapsto0x↦0 otherwise, obtaining ppp; the result is

(Tr⁡p)−1p.(\operatorname{Tr}p)^{-1}p.(Trp)−1p.

For Hermitian hhh, ppp is the orthogonal projection onto its smallest-eigenvalue eigenspace. The functional calculus gives zero for non-Hermitian hhh, and a zero trace also makes the normalization zero.

QAlgorithms.Papers.Independent.Support50.P15.heatCapacity

For every natural number ddd, arbitrary complex d×dd\times dd×d matrix hhh, and extended nonnegative real t∈[0,∞]t\in[0,\infty]t∈[0,∞], this defines 000 when t=0t=0t=0 or t=∞t=\inftyt=∞. For finite positive ttt, it is the ordinary real derivative at u=tu=tu=t of the function

u⟼Re⁡Tr⁡[h (Tr⁡exp⁡(−u−1h))−1exp⁡(−u−1h)].u\longmapsto \operatorname{Re}\operatorname{Tr} \left[h\,\bigl(\operatorname{Tr}\exp(-u^{-1}h)\bigr)^{-1} \exp(-u^{-1}h)\right].u⟼ReTr[h(Trexp(−u−1h))−1exp(−u−1h)].

No Hermitian assumption or differentiability hypothesis is imposed; the derivative operator returns 000 where the function is not differentiable. All inverses send zero to zero, and d=0d=0d=0 is included.

QAlgorithms.Papers.Independent.Support50.P15.helmholtzFreeEnergy

For any natural number ddd, any complex d×dd\times dd×d matrices h,ah,ah,a, and any real number ttt, this defines

Re⁡tr⁡(ha)+t Re⁡tr⁡ ⁣[f(a)],f(x)=xlog⁡x,\operatorname{Re}\operatorname{tr}(ha) +t\,\operatorname{Re}\operatorname{tr}\!\left[f(a)\right], \qquad f(x)=x\log x,Retr(ha)+tRetr[f(a)],f(x)=xlogx,

where f(a)f(a)f(a) is obtained by real functional calculus and the real logarithm has log⁡0=0\log 0=0log0=0. Neither matrix is required to be Hermitian, positive, normalized, or invertible, and ttt may be zero or negative. For d=0d=0d=0, the value is zero.

QAlgorithms.Papers.Independent.Support50.P15.inverseEstimate

For every natural number ddd, including 000, and every real aaa, this defines

inverseEstimate⁡(d,a)={0,a=0,(e−1)ae(1+log⁡(d−1)−log⁡a),a≠0.\operatorname{inverseEstimate}(d,a)= \begin{cases} 0,&a=0,\\[2mm] \displaystyle\frac{(e-1)a}{e\bigl(1+\log(d-1)-\log a\bigr)},&a\ne0. \end{cases}inverseEstimate(d,a)=⎩⎨⎧​0,e(1+log(d−1)−loga)(e−1)a​,​a=0,a=0.​

The subtraction d−1d-1d−1 is in R\mathbb RR. No positivity assumptions are imposed: the logarithms are total real logarithms, with log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx, and a zero denominator makes the quotient 000.

QAlgorithms.Papers.Independent.Support50.P15.isIsothermal

For any natural number ddd, a process ppp with k>0k>0k>0, density matrices ρ0,…,ρk\rho_0,\ldots,\rho_kρ0​,…,ρk​, and matrices H0,…,Hk−1H_0,\ldots,H_{k-1}H0​,…,Hk−1​, an arbitrary matrix h∈Md(C)h\in M_d(\mathbb C)h∈Md​(C), a real number ttt, and a density matrix aaa, this predicate means that ρ0=a\rho_0=aρ0​=a, every ρj\rho_jρj​, including both endpoints, has zero kernel, every HjH_jHj​ is Hermitian, and

ρj+1=(tr⁡exp⁡(−t−1Hj))−1exp⁡(−t−1Hj)(0≤j<k),Hk−1=h.\rho_{j+1} =\bigl(\operatorname{tr}\exp(-t^{-1}H_j)\bigr)^{-1}\exp(-t^{-1}H_j) \quad(0\le j<k), \qquad H_{k-1}=h.ρj+1​=(trexp(−t−1Hj​))−1exp(−t−1Hj​)(0≤j<k),Hk−1​=h.

There is no requirement that t>0t>0t>0. In particular, at t=0t=0t=0, the convention 0−1=00^{-1}=00−1=0 makes every displayed thermal state equal to Id/dI_d/dId​/d. For d=0d=0d=0, the process and density-matrix arguments cannot exist.

QAlgorithms.Papers.Independent.Support50.P15.isLaw

For any natural number ddd and any real-valued vector p=(pi)i=0d−1p=(p_i)_{i=0}^{d-1}p=(pi​)i=0d−1​, this is the proposition

(∀i, pi≥0)and∑i=0d−1pi=1.(\forall i,\ p_i\ge0)\quad\text{and}\quad \sum_{i=0}^{d-1}p_i=1.(∀i, pi​≥0)andi=0∑d−1​pi​=1.

When d=0d=0d=0, the sum is 000, so no vector satisfies this proposition.

QAlgorithms.Papers.Independent.Support50.P15.klTerm

For arbitrary real numbers p,qp,qp,q, this defines the extended real number

klTerm(p,q)={0,p=0,+∞,p≠0 and q=0,plog⁡(p/q),p≠0 and q≠0.\mathrm{klTerm}(p,q)= \begin{cases} 0,&p=0,\\ +\infty,&p\ne0\ \text{and}\ q=0,\\ p\log(p/q),&p\ne0\ \text{and}\ q\ne0. \end{cases}klTerm(p,q)=⎩⎨⎧​0,+∞,plog(p/q),​p=0,p=0 and q=0,p=0 and q=0.​

The first branch takes precedence, so the value at (0,0)(0,0)(0,0) is zero. There are no nonnegativity or probability assumptions on p,qp,qp,q; for negative arguments, the total real logarithm satisfies log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣.

QAlgorithms.Papers.Independent.Support50.P15.massLimit

For every natural number ddd, this defines the real number

massLimit(d)=d−1d,\mathrm{massLimit}(d)=\frac{d-1}{d},massLimit(d)=dd−1​,

where subtraction and division are performed in the reals after converting ddd. Thus the numerator at d=0d=0d=0 is −1-1−1, and total real division by zero makes the value zero; the value is also zero at d=1d=1d=1.

QAlgorithms.Papers.Independent.Support50.P15.maxVariance

For every natural number ddd, including 000 and 111, this defines the real number

sup⁡{r(1−r)[log⁡ ⁣(1−rr(d−1))]2: 0<r<12}.\sup\left\{ r(1-r)\left[\log\!\left(\frac{1-r}{r}(d-1)\right)\right]^2 :\ 0<r<\frac12 \right\}.sup{r(1−r)[log(r1−r​(d−1))]2: 0<r<21​}.

The endpoints 000 and 1/21/21/2 are excluded. The logarithm uses its total real convention, including log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. This definition does not assert that the supremum is attained; the real supremum operation returns zero for an empty set or a set unbounded above.

QAlgorithms.Papers.Independent.Support50.P15.mesh

For every partition ppp consisting of a positive integer kkk and strictly increasing real points 0=p0<⋯<pk=10=p_0<\cdots<p_k=10=p0​<⋯<pk​=1, this defines

mesh⁡(p)=sup⁡{pj+1−pj:0≤j<k}.\operatorname{mesh}(p)=\sup\{p_{j+1}-p_j:0\le j<k\}.mesh(p)=sup{pj+1​−pj​:0≤j<k}.

The set is finite and nonempty, so this supremum is its largest interval length.

QAlgorithms.Papers.Independent.Support50.P15.minEntropy

For any natural numbers d,md,md,m and any family a0,…,am−1a_0,\ldots,a_{m-1}a0​,…,am−1​ of complex d×dd\times dd×d positive semidefinite trace-one matrices, this defines the real infimum

inf⁡{−Re⁡tr⁡f(aj):0≤j<m},f(x)=xlog⁡x,\inf\left\{ -\operatorname{Re}\operatorname{tr}f(a_j):0\le j<m \right\}, \qquad f(x)=x\log x,inf{−Retrf(aj​):0≤j<m},f(x)=xlogx,

where f(aj)f(a_j)f(aj​) is given by real functional calculus using the natural logarithm and log⁡0=0\log 0=0log0=0. When m=0m=0m=0, the real infimum of the empty set is defined to be zero. If d=0d=0d=0, such a family exists only for m=0m=0m=0.

QAlgorithms.Papers.Independent.Support50.P15.minimaxRedundancy

For any natural numbers d,md,md,m and any family a0,…,am−1a_0,\ldots,a_{m-1}a0​,…,am−1​ of complex d×dd\times dd×d positive semidefinite trace-one matrices, this defines the extended-real number

inf⁡b⪰0tr⁡b=1  sup⁡0≤j<mD(aj∥b),\inf_{\substack{b\succeq0\\ \operatorname{tr}b=1}}\; \sup_{0\le j<m}D(a_j\Vert b),b⪰0trb=1​inf​0≤j<msup​D(aj​∥b),

where D(ρ∥σ)=+∞D(\rho\Vert\sigma)=+\inftyD(ρ∥σ)=+∞ if ker⁡σ⊈ker⁡ρ\ker\sigma\not\subseteq\ker\rhokerσ⊆kerρ, and otherwise

D(ρ∥σ)=max⁡ ⁣{0,Re⁡tr⁡[f(ρ)−ρlog⁡σ]},f(x)=xlog⁡x.D(\rho\Vert\sigma) =\max\!\left\{0,\operatorname{Re}\operatorname{tr} \left[f(\rho)-\rho\log\sigma\right]\right\}, \qquad f(x)=x\log x.D(ρ∥σ)=max{0,Retr[f(ρ)−ρlogσ]},f(x)=xlogx.

Both matrix functions use real functional calculus with the total real logarithm, including log⁡0=0\log 0=0log0=0. No minimizer is asserted. If m=0m=0m=0 and d>0d>0d>0, the value is −∞-\infty−∞, since the inner supremum is empty. For d=0d=0d=0, a supplied family can exist only when m=0m=0m=0, and the outer infimum is empty, giving +∞+\infty+∞.

QAlgorithms.Papers.Independent.Support50.P15.minimizingPair

For every natural number ddd and real numbers δ,s,r\delta,s,rδ,s,r, this predicate requires that s,r∈[0,(d−1)/d]s,r\in[0,(d-1)/d]s,r∈[0,(d−1)/d], that hd(s)−hd(r)=δh_d(s)-h_d(r)=\deltahd​(s)−hd​(r)=δ, and that

K(s,r)+K(1−s,1−r)=inf⁡u,v∈[0,(d−1)/d]hd(u)−hd(v)=δ(K(u,v)+K(1−u,1−v)),K(s,r)+K(1-s,1-r) = \inf_{\substack{u,v\in[0,(d-1)/d]\\h_d(u)-h_d(v)=\delta}} \bigl(K(u,v)+K(1-u,1-v)\bigr),K(s,r)+K(1−s,1−r)=u,v∈[0,(d−1)/d]hd​(u)−hd​(v)=δ​inf​(K(u,v)+K(1−u,1−v)),

where hd(u)=−ulog⁡u−(1−u)log⁡(1−u)+ulog⁡(d−1)h_d(u)=-u\log u-(1-u)\log(1-u)+u\log(d-1)hd​(u)=−ulogu−(1−u)log(1−u)+ulog(d−1), and K(p,q)K(p,q)K(p,q) equals zero when p=0p=0p=0, equals +∞+\infty+∞ when p≠0p\ne0p=0 and q=0q=0q=0, and otherwise equals plog⁡(p/q)p\log(p/q)plog(p/q). These are extended real values; the natural logarithm has log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx, real division by zero is zero, and an empty infimum is +∞+\infty+∞. This requires attainment but not uniqueness or finiteness. For d=0d=0d=0 or d=1d=1d=1, it holds exactly when δ=s=r=0\delta=s=r=0δ=s=r=0.

QAlgorithms.Papers.Independent.Support50.P15.optimalHamiltonian

For every natural number ddd, this defines the complex diagonal d×dd\times dd×d matrix whose diagonal entry at index 000 is −1-1−1 and whose other diagonal entries are 000, with every off-diagonal entry also 000. When d=0d=0d=0, there is no index 000, and the result is the empty matrix.

QAlgorithms.Papers.Independent.Support50.P15.optimalMass

For every natural number ddd, this selects a real number rrr satisfying

0<r<12,r(1−r)[log⁡ ⁣(1−rr(d−1))]2=sup⁡0<u<1/2u(1−u)[log⁡ ⁣(1−uu(d−1))]2,0<r<\frac12,\qquad r(1-r)\left[\log\!\left(\frac{1-r}{r}(d-1)\right)\right]^2 = \sup_{0<u<1/2} u(1-u)\left[\log\!\left(\frac{1-u}{u}(d-1)\right)\right]^2,0<r<21​,r(1−r)[log(r1−r​(d−1))]2=0<u<1/2sup​u(1−u)[log(u1−u​(d−1))]2,

if such a number exists. The selection is by classical choice and need not be unique; if no such number exists, it selects an unspecified real number without the displayed properties. Existence is not asserted by this definition. 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, and the real supremum returns zero on sets unbounded above. In particular, when d=1d=1d=1, every r∈(0,1/2)r\in(0,1/2)r∈(0,1/2) satisfies the selection condition.

QAlgorithms.Papers.Independent.Support50.P15.partitionHeat

For any natural number ddd, any function a:R→Md(C)a:\mathbb R\to M_d(\mathbb C)a:R→Md​(C), any partition with a positive integer kkk and strictly increasing points 0=x0<⋯<xk=10=x_0<\cdots<x_k=10=x0​<⋯<xk​=1, and any real numbers t0,…,tk−1t_0,\ldots,t_{k-1}t0​,…,tk−1​, this defines

∑j=0k−1Re⁡tr⁡ ⁣[(−tjlog⁡(a(xj+1)))(a(xj+1)−a(xj))]tj.\sum_{j=0}^{k-1} \frac{\operatorname{Re}\operatorname{tr}\!\left[ \bigl(-t_j\log(a(x_{j+1}))\bigr)\bigl(a(x_{j+1})-a(x_j)\bigr) \right]}{t_j}.j=0∑k−1​tj​Retr[(−tj​log(a(xj+1​)))(a(xj+1​)−a(xj​))]​.

The matrix logarithm is the real functional calculus of the total real logarithm, with log⁡0=0\log 0=0log0=0. No regularity or matrix properties are assumed for aaa, and the tjt_jtj​ need not be positive or nonzero: a term with tj=0t_j=0tj​=0 equals zero because division by zero is defined to be zero. Dimension d=0d=0d=0 is allowed and gives zero.

QAlgorithms.Papers.Independent.Support50.P15.piecewiseC1

For every natural number ddd, including 000, and function a:R→Cd×da:\mathbb R\to\mathbb C^{d\times d}a:R→Cd×d, this predicate means that aaa is continuous on [0,1][0,1][0,1] and that there exist a positive integer kkk and strictly increasing real points 0=p0<⋯<pk=10=p_0<\cdots<p_k=10=p0​<⋯<pk​=1 such that aaa is continuously differentiable over R\mathbb RR on each closed interval [pj,pj+1][p_j,p_{j+1}][pj​,pj+1​], for every 0≤j<k0\le j<k0≤j<k, in the within-set sense of differentiability. It imposes no density-matrix, Hermitian, or rank conditions on the values of aaa, and no conditions outside [0,1][0,1][0,1].

QAlgorithms.Papers.Independent.Support50.P15.processHeat

For any natural number ddd and any process ppp consisting of k>0k>0k>0, density matrices ρ0,…,ρk\rho_0,\ldots,\rho_kρ0​,…,ρk​ that are positive semidefinite with trace one, and arbitrary complex matrices H0,…,Hk−1H_0,\ldots,H_{k-1}H0​,…,Hk−1​, this defines

∑j=0k−1Re⁡tr⁡ ⁣[Hj(ρj+1−ρj)].\sum_{j=0}^{k-1}\operatorname{Re}\operatorname{tr}\!\left[H_j(\rho_{j+1}-\rho_j)\right].j=0∑k−1​Retr[Hj​(ρj+1​−ρj​)].

No isothermal condition or Hermiticity of the HjH_jHj​ is assumed. There are no processes in dimension zero.

QAlgorithms.Papers.Independent.Support50.P15.redundancyBound

For every natural number ddd and every real number aaa, this defines

a22([log⁡(d−1)]2+4)−a33([log⁡(d−1)]2+4)2.\frac{a^2}{2\bigl([\log(d-1)]^2+4\bigr)} - \frac{a^3}{3\bigl([\log(d-1)]^2+4\bigr)^2}.2([log(d−1)]2+4)a2​−3([log(d−1)]2+4)2a3​.

There is no sign restriction on aaa, and d=0,1d=0,1d=0,1 are included. Here d−1d-1d−1 is real subtraction, log⁡0=0\log0=0log0=0, and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx; both denominators are strictly positive.

QAlgorithms.Papers.Independent.Support50.P15.redundancyScale

For every natural number ddd, including 000 and 111, this defines

[log⁡(d−1)]2+4,[\log(d-1)]^2+4,[log(d−1)]2+4,

where d−1d-1d−1 is real subtraction and the real logarithm satisfies log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. The value is always at least 444, and equals 444 when d=0d=0d=0 or d=1d=1d=1.

QAlgorithms.Papers.Independent.Support50.P15.shannonEntropyNats

For any natural number ddd and any real-valued vector p=(pi)i=0d−1p=(p_i)_{i=0}^{d-1}p=(pi​)i=0d−1​, this defines

−∑i=0d−1pilog⁡pi.-\sum_{i=0}^{d-1}p_i\log p_i.−i=0∑d−1​pi​logpi​.

No nonnegativity or normalization assumption is imposed on ppp. The logarithm is the total real logarithm: log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. Thus zero entries contribute 000, and when d=0d=0d=0 the value is 000.

QAlgorithms.Papers.Independent.Support50.P15.sourceEntropyMaximum

For natural numbers d,md,md,m and a family (aj)j=0m−1(a_j)_{j=0}^{m-1}(aj​)j=0m−1​ of complex d×dd\times dd×d positive semidefinite matrices of trace 111, this defines

sup⁡0≤j<mS(aj),S(ρ)=−Re⁡tr⁡ ⁣(f(ρ)),f(x)=xlog⁡x,\sup_{0\le j<m} S(a_j), \qquad S(\rho)=-\operatorname{Re}\operatorname{tr}\!\bigl(f(\rho)\bigr), \quad f(x)=x\log x,0≤j<msup​S(aj​),S(ρ)=−Retr(f(ρ)),f(x)=xlogx,

where f(ρ)f(\rho)f(ρ) is defined by continuous functional calculus and log⁡\loglog is the natural logarithm, with log⁡0=0\log 0=0log0=0. The supremum is taken in R\mathbb RR; when m=0m=0m=0, its defining set is empty and the value is 000. The parameters permit d=0d=0d=0, but then such a family can exist only when m=0m=0m=0, since a 0×00\times00×0 matrix has trace 000.

QAlgorithms.Papers.Independent.Support50.P15.stateObservableVariance

For every natural number ddd and all complex d×dd\times dd×d matrices aaa and hhh, this defines

Re⁡tr⁡(ah2)−(Re⁡tr⁡(ah))2.\operatorname{Re}\operatorname{tr}(ah^2) - \bigl(\operatorname{Re}\operatorname{tr}(ah)\bigr)^2.Retr(ah2)−(Retr(ah))2.

Here h2h^2h2 is the matrix product hhhhhh. There are no Hermitian, positivity, normalization, or commutation hypotheses on either matrix. The case d=0d=0d=0 is included and gives zero.

QAlgorithms.Papers.Independent.Support50.P15.stateRelativeNats

For every natural number ddd and every pair of positive semidefinite complex d×dd\times dd×d matrices a,ba,ba,b of trace one, this defines an extended real number by

stateRelativeNats(a,b)={max⁡ ⁣{0,Re⁡Tr⁡ ⁣(f(a)−a g(b))},ker⁡b⊆ker⁡a,+∞,otherwise,f(x)=xlog⁡x,g(x)=log⁡x,\mathrm{stateRelativeNats}(a,b)= \begin{cases} \max\!\left\{0,\operatorname{Re}\operatorname{Tr}\!\left(f(a)-a\,g(b)\right)\right\}, & \ker b\subseteq\ker a,\\ +\infty,&\text{otherwise}, \end{cases} \qquad f(x)=x\log x,\quad g(x)=\log x,stateRelativeNats(a,b)={max{0,ReTr(f(a)−ag(b))},+∞,​kerb⊆kera,otherwise,​f(x)=xlogx,g(x)=logx,

where the matrix functions use continuous functional calculus and the logarithm is natural, with log⁡0=0\log 0=0log0=0 in these functions. The maximum with zero comes from conversion to the extended nonnegative reals before embedding into the extended reals. There are no possible arguments a,ba,ba,b when d=0d=0d=0.

QAlgorithms.Papers.Independent.Support50.P15.surprisalVariance

For every natural number ddd and every complex d×dd\times dd×d matrix aaa, this defines

Re⁡tr⁡(f(a))−[−Re⁡tr⁡(g(a))]2,f(x)=x(log⁡x)2,g(x)=xlog⁡x,\operatorname{Re}\operatorname{tr}\bigl(f(a)\bigr) - \left[-\operatorname{Re}\operatorname{tr}\bigl(g(a)\bigr)\right]^2, \qquad f(x)=x(\log x)^2,\quad g(x)=x\log x,Retr(f(a))−[−Retr(g(a))]2,f(x)=x(logx)2,g(x)=xlogx,

where f(a)f(a)f(a) and g(a)g(a)g(a) are formed using the real continuous functional calculus, including its total default convention outside its applicability conditions. The real logarithm has log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. No Hermitian, positivity, or trace-one hypothesis is imposed on aaa. The case d=0d=0d=0 is included and gives zero.

QAlgorithms.Papers.Independent.Support50.P15.thermalExpectedEnergy

For every natural number ddd, arbitrary complex d×dd\times dd×d matrix hhh, and real ttt, this defines

thermalExpectedEnergy⁡(h,t)=Re⁡Tr⁡[h (Tr⁡exp⁡(−t−1h))−1exp⁡(−t−1h)].\operatorname{thermalExpectedEnergy}(h,t) =\operatorname{Re}\operatorname{Tr} \left[h\,\bigl(\operatorname{Tr}\exp(-t^{-1}h)\bigr)^{-1} \exp(-t^{-1}h)\right].thermalExpectedEnergy(h,t)=ReTr[h(Trexp(−t−1h))−1exp(−t−1h)].

There are no Hermitian or temperature restrictions. Scalar inverses, including the complex inverse of the trace, send zero to zero. Thus at t=0t=0t=0 the expression is Re⁡Tr⁡(h d−1Id)\operatorname{Re}\operatorname{Tr}(h\,d^{-1}I_d)ReTr(hd−1Id​), and for d=0d=0d=0 it is 000.

QAlgorithms.Papers.Independent.Support50.P15.thermalFinite

For every natural number ddd, arbitrary complex d×dd\times dd×d matrix hhh, and real ttt, this defines

thermalFinite⁡(h,t)=(Tr⁡exp⁡(−t−1h))−1exp⁡(−t−1h),\operatorname{thermalFinite}(h,t) =\bigl(\operatorname{Tr}\exp(-t^{-1}h)\bigr)^{-1}\exp(-t^{-1}h),thermalFinite(h,t)=(Trexp(−t−1h))−1exp(−t−1h),

where exp⁡\expexp is the matrix exponential and the trace and its inverse are complex-valued. No Hermitian or positivity assumption on hhh, or positivity assumption on ttt, is required. Inverses of zero are 000, so a zero trace gives the zero matrix; at t=0t=0t=0, the value is d−1Idd^{-1}I_dd−1Id​. For d=0d=0d=0, the result is the unique empty matrix.

QAlgorithms.Papers.Independent.Support50.P15.tightBound

For every natural number ddd and real number δ\deltaδ, this defines the extended real infimum

tightBound(d,δ)=inf⁡{K(s,r)+K(1−s,1−r):s,r∈[0,d−1d], hd(s)−hd(r)=δ},\mathrm{tightBound}(d,\delta) =\inf\left\{ K(s,r)+K(1-s,1-r): s,r\in\left[0,\frac{d-1}{d}\right],\ h_d(s)-h_d(r)=\delta \right\},tightBound(d,δ)=inf{K(s,r)+K(1−s,1−r):s,r∈[0,dd−1​], hd​(s)−hd​(r)=δ},

where hd(u)=−ulog⁡u−(1−u)log⁡(1−u)+ulog⁡(d−1)h_d(u)=-u\log u-(1-u)\log(1-u)+u\log(d-1)hd​(u)=−ulogu−(1−u)log(1−u)+ulog(d−1), and K(p,q)K(p,q)K(p,q) is zero if p=0p=0p=0, is +∞+\infty+∞ if p≠0p\ne0p=0 and q=0q=0q=0, and otherwise is plog⁡(p/q)p\log(p/q)plog(p/q). The logarithm is natural, with log⁡0=0\log0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx, and real division by zero is zero. The infimum of an empty feasible set is +∞+\infty+∞; no attainment is required. For d=0d=0d=0 or d=1d=1d=1, the value is zero when δ=0\delta=0δ=0 and +∞+\infty+∞ otherwise.

QAlgorithms.Papers.Independent.Support50.P15.varianceObjective

For every natural number ddd, including 000 and 111, and every real number rrr, this defines

Vd(r)=r(1−r)[log⁡ ⁣(1−rr(d−1))]2.V_d(r)=r(1-r)\left[\log\!\left(\frac{1-r}{r}(d-1)\right)\right]^2.Vd​(r)=r(1−r)[log(r1−r​(d−1))]2.

There is no restriction on rrr. Division by zero is defined as zero, and the real logarithm is total: log⁡0=0\log 0=0log0=0 and log⁡x=log⁡∣x∣\log x=\log|x|logx=log∣x∣ for negative xxx. In particular, Vd(0)=Vd(1)=0V_d(0)=V_d(1)=0Vd​(0)=Vd​(1)=0, and V1(r)=0V_1(r)=0V1​(r)=0 for every rrr.

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