Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Definitions for Quantum Information Processing with Finite Resources -- Mathematical Foundations

Definition
QAlgorithms_Papers_Independent_Intakenext100_P07

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

prove2quantum

The objects the formalized statements of Quantum Information Processing with Finite Resources -- Mathematical Foundations (arXiv:1504.00233) are stated in. Each was checked against the paper by a blind read-back and a moderator.

  • paperMat
  • paperBelow: PDF p.22: Loewner order.
  • paperSupported: PDF p.22: support inclusion is reverse kernel inclusion.
  • paperSubnormalized: PDF pp.26-28: subnormalized PSD states, including zero.
  • paperPure: PDF p.27: pure means rank one.
  • paperLogE: PDF pp.19-20: binary logarithm with log(0)=-infinity.
  • expE: PDF pp.19-20: binary exponential at finite and infinite arguments.
  • paperPower: PDF p.32: support powers, with zero on the kernel even at exponent zero.
  • matrixLog: PDF p.32: logarithm on the support, zero on the kernel.
  • paperSchatten: PDF p.40: Schatten functional at positive real orders.
  • positiveNorm: PDF p.42, (3.17): norm on all complex operators, with trace modulus.
  • paperDelta: PDF p.43: generalized trace distance for all operators.
  • paperFidelity: PDF p.44: squared algebraic fidelity, distinct from generalized fidelity.
  • paperGeneralizedFidelity: PDF p.46: squared generalized fidelity.
  • purified: PDF p.47: purified distance, using generalized fidelity once.
  • paperBall: PDF p.98: full subnormalized ball, without a support restriction.
  • Radius: PDF p.98: standing trace-dependent smoothing-radius restriction.
  • paperTraceMiddle: PDF p.30: trace out B in (AB)C.
  • paperRegroup: PDF p.30: actual regrouping (AB)C to A(BC).
  • paperMarginalBC: PDF p.30: computed marginal BC.
  • paperHmin: PDF p.92: extended supremum of domination exponents and reference states.
  • hmax: PDF p.94: extended max-entropy from squared algebraic fidelity; Hmax(0)=-infinity.
  • paperSmoothMin: PDF p.98: smooth min-entropy over the full subnormalized ball.
  • paperSmoothMax: PDF p.98: smooth max-entropy is an infimum over the same ball.
  • conditionalVN: PDF p.76: conditional von Neumann entropy in bits.
  • paperJointPower: PDF pp.109-111: iid matrix product, regrouped as A^n B^n.
  • paperTensorPower: PDF p.109: single-system matrix tensor power, with tensor unit at n=0.
  • minRate: PDF p.110: regularization in EReal; the n=0 default is irrelevant to the limit.
  • maxRate: PDF p.110: regularized smooth max-entropy.
  • paperLocalEmbedding: PDF p.100: arbitrary rectangular local isometric embedding.
  • CPTNI: PDF p.26: CPTNI, on every PSD input.
  • paperCPTP: PDF p.26: CPTP, with trace preservation on every operator.
  • paperTensorAction: PDF p.104: tensor action on the actual matrix-unit expansion.
  • paperMix: PDF p.32: binary convex combination of matrices.
  • ClassicalLeft: PDF p.32: classical on the first register.
  • ClassicalXY: PDF p.100: classical X and Y in register order (AX)(BY).
  • classicalProcess: PDF p.105: pushforward of the classical register; quantum registers retained.
  • paperPenalty: PDF p.103: smoothing penalty, used only at 0<e<1.
  • maxD: PDF p.106: max-divergence, including -infinity for a zero numerator.
  • smoothMaxD: PDF p.106: relative smoothing over subnormalized states, also at large radii.
  • paperPositivePart: PDF p.107: actual positive spectral part.
  • excessRadius: PDF p.107: radius computed from the positive excess, not a free witness.
  • paperRelativeD: PDF pp.61,65: normalized-trace Umegaki entropy with support-failure value +infinity.
  • paperVariance: PDF pp.72,109: centered relative-entropy variance on the support domain.
  • MinPrimal: PDF p.93: min-entropy primal feasibility.
  • MinDual: PDF p.93: min-entropy dual feasibility.
  • SmoothPrimal: PDF p.99: smooth min-entropy SDP, with overlap threshold 1-e^2.
  • conditionalBlocks: PDF p.96: classical blocks in the BY register without renormalization.
  • guessObjective: PDF p.97: guessing success probability.
  • paperMeasuredJoint: PDF p.97: classical joint distribution produced by a measurement.
  • paperCqState: PDF p.97: actual CQ assembly.
  • hartleyConditional: PDF p.95: optimized support entropy with the extended zero value.
  • paperRenyiQ: PDF pp.62,69: unnormalized sandwiched or Petz trace functional.
  • renyiFinite: PDF pp.61,69: finite positive-order divergence, normalized by Tr r. The a=1 value is the Umegaki limit, and support/orthogonality guards preserve infinity.
  • paperRenyi: PDF pp.61,69,78: endpoint orders are limits of the increasing finite-order family. For the sandwiched family infinity is the max-divergence.
  • conditionalDown: PDF p.77, Definition 5.2: fixed-reference conditional Renyi entropy.
  • conditionalUp: PDF p.77, Definition 5.2: optimize over actual normalized reference states.
  • paperConditional: PDF p.81: select fixed-reference or optimized entropy without changing the divergence family.
  • DPIOrder: PDF pp.81,86: orders for which this family has data processing, including limiting endpoints.
  • transformedOrder: PDF pp.84-85: real order 2-1/a including a=infinity, with reciprocal endpoint limits.
  • paperIndependentProduct: PDF p.84: regrouped independent product AB,CD to AC,BD.
  • spectralValues: PDF p.36: distinct spectral values; zero is included when present.
  • paperSpectralProjector: PDF p.36: actual spectral projection at l.
  • paperPinch: PDF p.36: spectral pinching of an operator.
  • SpectralRep: PDF p.70: spectral data is tied to the input matrix by reconstruction.
  • paperNsP: PDF p.70: Nussbaum-Szkola probability law from the actual two spectral bases.
  • paperNsQ: PDF p.70: Nussbaum-Szkola reference law.
  • paperProbability: PDF p.12: a finite normalized classical probability distribution.
  • paperSeparable: PDF p.76: finite separable states, with actual normalized product components.
  • basisLaw: PDF p.116: Born law for the orthonormal basis in the columns of U.
  • paperBasisMeasurement: PDF p.117: local basis measurement on A, retaining the side information B.
  • basisOverlap: PDF p.116: maximum SQUARED overlap of the two actual bases.
  • paperTypeI: PDF p.114: type-I error when T accepts rho.
  • paperTypeII: PDF p.114: type-II error when T accepts rho.
  • alphaStar: PDF p.114, (7.6). Printed: minimize alpha(T;sigma) subject to beta(T;rho). Changed: minimize alpha(T;rho) subject to beta(T;sigma). Why: (7.1)-(7.2), the sentence before (7.6), and the Stein threshold all fix these roles.
  • paperBetaStar: PDF p.115: minimum type-II error at prescribed type-I tolerance.
  • paperSymmetricError: PDF p.114: optimal average error with equal priors.
  • paperRenyiReal: PDF p.72: Renyi divergences as real-valued functions only on supported pairs.
  • petzExtendedQ: PDF p.69: extended Petz Q on support failure above one.
  • paperSourceCode: PDF p.12: fixed-cardinality deterministic source code.
  • paperCodeError: PDF p.12: actual probability of decoding failure.
  • paperClassicalRenyi: PDF p.13: classical Renyi entropy at finite positive order other than one.
  • paperShannon: PDF p.13: Shannon entropy, with zero contributions from zero probabilities.
  • variation: PDF p.14: introductory variation distance (different from purified distance).
  • introSmoothMax: PDF p.14: smoothing only over normalized classical probability laws.
  • iidLaw: PDF p.13: iid classical product law.
  • codeCardinality: PDF p.13: integer memory cardinality rounded upward from 2^(nR).
  • Resolution: PDF p.36: finite orthogonal resolution of identity.
  • resolutionPinch: PDF p.36: pinching from a finite resolution.
  • UnitaryInvariantNorm: PDF p.40: all unitarily invariant complex norms, rather than a particular matrix norm.
  • hatState: PDF p.46: normalized hat-state with a single vacuum flag.
  • paperChoi: PDF p.37: Choi matrix from the UNNORMALIZED entangled vector.
  • inverseChoi: PDF p.37, (2.69): inverse in register order A-prime,B; the input transpose is implicit in the pairing.
  • paperEntangled: PDF p.93: normalized maximally entangled density matrix on two copies of the same basis.
  • preparedEntangled: PDF p.94, (6.11): a B-prime-to-A channel applied to the second half of psi_BBprime, then regrouped as AB.
  • paperAdjointAction: PDF pp.91-92: the actual Hilbert-Schmidt adjoint, computed from matrix units.
  • sdpPrimal: PDF p.92: extended primal infimum; infeasibility is +infinity.
  • sdpDual: PDF p.92: extended dual supremum; infeasibility is -infinity.
  • productState: PDF p.109: heterogeneous product states, retaining every factor.
  • paperQuadratic: PDF pp.31,42: the actual quadratic form, with conjugation in the first argument.
  • measureOutput: PDF pp.35,68: actual classical probability matrix of a POVM output.
  • optimalCardinality: PDF p.14: minimum code alphabet cardinality. On probability sources the feasible set is nonempty (lossless coding).
  • paperDivergenceFamily: PDF pp.52-58: a divergence is one family across all finite systems, so classical outputs, tensor powers, direct sums and rectangular channels have the same value rule.
  • paperClassicalAgreement: PDF pp.56,58,61: agreement with the classical divergence on diagonal positive pairs.
  • AdditiveDivergence: PDF p.52, (V): tensor additivity, on arbitrary nonzero positive numerators.
  • paperDataProcessing: PDF p.53, (VIII): every CPTP map, including maps between different dimensions.
  • paperMaxFamily: PDF pp.60-61: the concrete max-divergence family.
  • measuredRenyi: PDF p.66, (4.73): optimization over actual finite-outcome POVMs, in EReal.
  • MeasuredRepresentation: PDF p.66, (4.73): the measured representation for every positive input pair, including attainment of each finite-copy maximum. This predicate does not assume DPI.
  • AxiomDomain: PDF p.52: finite real domain of the six axioms.
  • RealDivergenceFamily: PDF p.52: real-valued divergence across every finite system, before adjoining infinities.
  • paperDirectSum: PDF p.52: direct sum in the disjoint-union basis.
  • paperSixAxioms: PDF p.52, (I)-(VI): the actual unbounded positive cone and the full trace-weighted mean. The accepted Dependency50.P35.SixAxioms restricts the mean to total traces at most one; this book states the mean for arbitrary positive inputs, so it has a distinct local predicate.
  • MeanCurvature: PDF pp.53,57: joint convexity or concavity of g(D) on normalized pairs.
  • RealDataProcessing: PDF pp.53,57: data processing on the finite domain where the axiomatic D is real.
Definition code
import Mathlib
import Definitions.Def_QAlgorithms_Defs_MatrixDistances
import Definitions.Def_QAlgorithms_Defs_MatrixMapPositivity
import Definitions.Def_QAlgorithms_Defs_QuantumInformation
import Definitions.Def_QAlgorithms_Defs_Ensemble

/-! Tomamichel, arXiv:1504.00233v5. PDF pages; binary logarithms.
New definitions are local to this paper. Compilation and independent moderation pending. -/
noncomputable section
open Filter
open scoped BigOperators Matrix ComplexConjugate ComplexOrder Matrix.Norms.L2Operator Topology
namespace QAlgorithms.Papers.Independent.Intakenext100.P07
attribute [local instance] Classical.propDecidable Classical.decEq

abbrev paperMat (I : Type*) := Matrix I I ℂ

variable {I J A B C D X Y Z : Type*}
variable [Fintype I] [DecidableEq I] [Fintype J] [DecidableEq J]
variable [Fintype A] [DecidableEq A] [Fintype B] [DecidableEq B]
variable [Fintype C] [DecidableEq C] [Fintype D] [DecidableEq D]
variable [Fintype X] [DecidableEq X] [Fintype Y] [DecidableEq Y]
variable [Fintype Z] [DecidableEq Z]

/-- PDF p.22: Loewner order. -/
def paperBelow (M N : paperMat I) : Prop := (N-M).PosSemidef

/-- PDF p.22: support inclusion is reverse kernel inclusion. -/
def paperSupported (M N : paperMat I) : Prop := LinearMap.ker N.mulVecLin ≤ LinearMap.ker M.mulVecLin

/-- PDF pp.26-28: subnormalized PSD states, including zero. -/
def paperSubnormalized (r : paperMat I) : Prop := r.PosSemidef ∧ r.trace.re ≤ 1

/-- PDF p.27: pure means rank one. -/
def paperPure (r : paperMat I) : Prop := r.PosSemidef ∧ r.rank = 1

/-- PDF pp.19-20: binary logarithm with log(0)=-infinity. -/
def paperLogE (x : ℝ) : EReal := if x = 0 then ⊥ else (Real.logb 2 x : EReal)

/-- PDF pp.19-20: binary exponential at finite and infinite arguments. -/
def expE (x : EReal) : ENNReal :=
  if x = ⊤ then ⊤ else if x = ⊥ then 0 else ENNReal.ofReal ((2:ℝ)^x.toReal)

/-- PDF p.32: support powers, with zero on the kernel even at exponent zero. -/
def paperPower (M : paperMat I) (p : ℝ) : paperMat I := cfc (fun x : ℝ => if x = 0 then 0 else x^p) M

/-- PDF p.32: logarithm on the support, zero on the kernel. -/
def matrixLog (M : paperMat I) : paperMat I := cfc (fun x : ℝ => Real.logb 2 x) M

/-- PDF p.40: Schatten functional at positive real orders. -/
def paperSchatten (M : paperMat I) (p : ℝ) : ℝ :=
  (paperPower (cfc Real.sqrt (Mᴴ*M)) p).trace.re ^ (1/p)

/-- PDF p.42, (3.17): norm on all complex operators, with trace modulus. -/
def positiveNorm (M : paperMat I) : ℝ :=
  sSup {r : ℝ | ∃ E : paperMat I, IsEffect E ∧ r = ‖(M*E).trace‖}

/-- PDF p.43: generalized trace distance for all operators. -/
def paperDelta (M N : paperMat I) : ℝ := positiveNorm (M-N)

/-- PDF p.44: squared algebraic fidelity, distinct from generalized fidelity. -/
def paperFidelity (r s : paperMat I) : ℝ := rootFidelity r s ^ 2

/-- PDF p.46: squared generalized fidelity. -/
def paperGeneralizedFidelity (r s : paperMat I) : ℝ :=
  (rootFidelity r s + Real.sqrt ((1-r.trace.re)*(1-s.trace.re)))^2

/-- PDF p.47: purified distance, using generalized fidelity once. -/
def purified (r s : paperMat I) : ℝ := Real.sqrt (1-paperGeneralizedFidelity r s)

/-- PDF p.98: full subnormalized ball, without a support restriction. -/
def paperBall (r : paperMat I) (e : ℝ) : Set (paperMat I) := {s | paperSubnormalized s ∧ purified s r ≤ e}

/-- PDF p.98: standing trace-dependent smoothing-radius restriction. -/
def Radius (r : paperMat I) (e : ℝ) : Prop := 0 ≤ e ∧ e < Real.sqrt r.trace.re

/-- PDF p.30: trace out B in (AB)C. -/
def paperTraceMiddle (r : paperMat ((A × B) × C)) : paperMat (A × C) :=
  fun x y => ∑ b, r ((x.1,b),x.2) ((y.1,b),y.2)

/-- PDF p.30: actual regrouping (AB)C to A(BC). -/
def paperRegroup (r : paperMat ((A × B) × C)) : paperMat (A × (B × C)) :=
  fun x y => r ((x.1,x.2.1),x.2.2) ((y.1,y.2.1),y.2.2)

/-- PDF p.30: computed marginal BC. -/
def paperMarginalBC (r : paperMat ((A × B) × C)) : paperMat (B × C) := traceLeft (paperRegroup r)

/-- PDF p.92: extended supremum of domination exponents and reference states. -/
def paperHmin (r : paperMat (A × B)) : EReal :=
  sSup {h : EReal | ∃ l : ℝ, h = (l:EReal) ∧ ∃ s : paperMat B, paperSubnormalized s ∧
    paperBelow r ((((2:ℝ)^(-l) : ℝ) : ℂ) • Matrix.kronecker (1:paperMat A) s)}

/-- PDF p.94: extended max-entropy from squared algebraic fidelity; Hmax(0)=-infinity. -/
def hmax (r : paperMat (A × B)) : EReal :=
  sSup {h : EReal | ∃ s : paperMat B, paperSubnormalized s ∧ h = paperLogE (paperFidelity r (Matrix.kronecker (1:paperMat A) s))}

/-- PDF p.98: smooth min-entropy over the full subnormalized ball. -/
def paperSmoothMin (r : paperMat (A × B)) (e : ℝ) : EReal :=
  sSup {h : EReal | ∃ s ∈ paperBall r e, h = paperHmin s}

/-- PDF p.98: smooth max-entropy is an infimum over the same ball. -/
def paperSmoothMax (r : paperMat (A × B)) (e : ℝ) : EReal :=
  sInf {h : EReal | ∃ s ∈ paperBall r e, h = hmax s}

/-- PDF p.76: conditional von Neumann entropy in bits. -/
def conditionalVN (r : paperMat (A × B)) : ℝ := vnEntropy r-vnEntropy (traceLeft r)

/-- PDF pp.109-111: iid matrix product, regrouped as A^n B^n. -/
def paperJointPower (r : paperMat (A × B)) (n : ℕ) : paperMat ((Fin n → A) × (Fin n → B)) :=
  fun x y => ∏ i, r (x.1 i,x.2 i) (y.1 i,y.2 i)

/-- PDF p.109: single-system matrix tensor power, with tensor unit at n=0. -/
def paperTensorPower (r : paperMat I) (n : ℕ) : paperMat (Fin n → I) := fun x y => ∏ i, r (x i) (y i)

/-- PDF p.110: regularization in EReal; the n=0 default is irrelevant to the limit. -/
def minRate (r : paperMat (A × B)) (e : ℝ) (n : ℕ) : EReal :=
  ((1/(n:ℝ) : ℝ) : EReal)*paperSmoothMin (paperJointPower r n) e

/-- PDF p.110: regularized smooth max-entropy. -/
def maxRate (r : paperMat (A × B)) (e : ℝ) (n : ℕ) : EReal :=
  ((1/(n:ℝ) : ℝ) : EReal)*paperSmoothMax (paperJointPower r n) e

/-- PDF p.100: arbitrary rectangular local isometric embedding. -/
def paperLocalEmbedding (U : Matrix C A ℂ) (V : Matrix D B ℂ) (r : paperMat (A × B)) : paperMat (C × D) :=
  Matrix.kronecker U V*r*(Matrix.kronecker U V)ᴴ

/-- PDF p.26: CPTNI, on every PSD input. -/
def CPTNI (E : paperMat I →ₗ[ℂ] paperMat J) : Prop :=
  QAlgorithms.Entanglement.IsCompletelyPositive E ∧ ∀ r, r.PosSemidef → (E r).trace.re ≤ r.trace.re

/-- PDF p.26: CPTP, with trace preservation on every operator. -/
def paperCPTP (E : paperMat I →ₗ[ℂ] paperMat J) : Prop :=
  QAlgorithms.Entanglement.IsCompletelyPositive E ∧ ∀ M, (E M).trace = M.trace

/-- PDF p.104: tensor action on the actual matrix-unit expansion. -/
def paperTensorAction (E : paperMat A →ₗ[ℂ] paperMat C) (F : paperMat B →ₗ[ℂ] paperMat D) (r : paperMat (A × B)) : paperMat (C × D) :=
  fun x y => ∑ a, ∑ a', ∑ b, ∑ b', r (a,b) (a',b') * E (Matrix.single a a' 1) x.1 y.1 *
    F (Matrix.single b b' 1) x.2 y.2

/-- PDF p.32: binary convex combination of matrices. -/
def paperMix (t : ℝ) (M N : paperMat I) : paperMat I := (t:ℂ) • M + ((1-t:ℝ):ℂ) • N

/-- PDF p.32: classical on the first register. -/
def ClassicalLeft (r : paperMat (X × A)) : Prop := ∀ x y a b, x ≠ y → r (x,a) (y,b) = 0

/-- PDF p.100: classical X and Y in register order (AX)(BY). -/
def ClassicalXY (r : paperMat ((A × X) × (B × Y))) : Prop :=
  ∀ a x b y a' x' b' y', x ≠ x' ∨ y ≠ y' → r ((a,x),(b,y)) ((a',x'),(b',y')) = 0

/-- PDF p.105: pushforward of the classical register; quantum registers retained. -/
def classicalProcess (f : X → Z) (r : paperMat ((X × A) × B)) : paperMat ((Z × A) × B) :=
  fun z w => if z.1.1 = w.1.1 then ∑ x, if f x = z.1.1 then
    r ((x,z.1.2),z.2) ((x,w.1.2),w.2) else 0 else 0

/-- PDF p.103: smoothing penalty, used only at 0<e<1. -/
def paperPenalty (e : ℝ) : ℝ := -Real.logb 2 (1-Real.sqrt (1-e^2))

/-- PDF p.106: max-divergence, including -infinity for a zero numerator. -/
def maxD (r s : paperMat I) : EReal :=
  sInf {d : EReal | ∃ l : ℝ, d = (l:EReal) ∧ paperBelow r (((2:ℝ)^l : ℂ) • s)}

/-- PDF p.106: relative smoothing over subnormalized states, also at large radii. -/
def smoothMaxD (r s : paperMat I) (e : ℝ) : EReal :=
  sInf {d : EReal | ∃ t ∈ paperBall r e, d = maxD t s}

/-- PDF p.107: actual positive spectral part. -/
def paperPositivePart (M : paperMat I) : paperMat I := cfc (fun x : ℝ => max x 0) M

/-- PDF p.107: radius computed from the positive excess, not a free witness. -/
def excessRadius (r s : paperMat I) (l : ℝ) : ℝ :=
  let t := (paperPositivePart (r-(((2:ℝ)^l : ℂ) • s))).trace.re
  Real.sqrt (2*t-t^2)

/-- PDF pp.61,65: normalized-trace Umegaki entropy with support-failure value +infinity. -/
def paperRelativeD (r s : paperMat I) : EReal := if paperSupported r s then
  (((r*(matrixLog r-matrixLog s)).trace.re/r.trace.re : ℝ):EReal) else ⊤

/-- PDF pp.72,109: centered relative-entropy variance on the support domain. -/
def paperVariance (r s : paperMat I) : ℝ :=
  (r*(matrixLog r-matrixLog s-((paperRelativeD r s).toReal:ℂ) • (1:paperMat I))^2).trace.re

/-- PDF p.93: min-entropy primal feasibility. -/
def MinPrimal (r : paperMat (A × B)) (s : paperMat B) : Prop := s.PosSemidef ∧ paperBelow r (Matrix.kronecker (1:paperMat A) s)

/-- PDF p.93: min-entropy dual feasibility. -/
def MinDual (W : paperMat (A × B)) : Prop := W.PosSemidef ∧ paperBelow (traceLeft W) 1

/-- PDF p.99: smooth min-entropy SDP, with overlap threshold 1-e^2. -/
def SmoothPrimal (r : paperMat ((A × B) × C)) (e : ℝ) (t : paperMat ((A × B) × C)) (s : paperMat B) : Prop :=
  paperSubnormalized t ∧ s.PosSemidef ∧ 1-e^2 ≤ (t*r).trace.re ∧
    paperBelow (traceRight t) (Matrix.kronecker (1:paperMat A) s)

/-- PDF p.96: classical blocks in the BY register without renormalization. -/
def conditionalBlocks (r : Y → paperMat (A × B)) : paperMat (A × (B × Y)) :=
  fun x z => if x.2.2 = z.2.2 then r x.2.2 (x.1,x.2.1) (z.1,z.2.1) else 0

/-- PDF p.97: guessing success probability. -/
def guessObjective (r : X → paperMat B) (M : X → paperMat B) : ℝ := ∑ x, (M x*r x).trace.re

/-- PDF p.97: classical joint distribution produced by a measurement. -/
def paperMeasuredJoint (r : X → paperMat B) (M : X → paperMat B) : paperMat (X × X) :=
  Matrix.diagonal (fun xy => ((M xy.2*r xy.1).trace.re:ℂ))

/-- PDF p.97: actual CQ assembly. -/
def paperCqState (r : X → paperMat B) : paperMat (X × B) := fun x y => if x.1 = y.1 then r x.1 x.2 y.2 else 0

/-- PDF p.95: optimized support entropy with the extended zero value. -/
def hartleyConditional (r : paperMat (A × B)) : EReal :=
  sSup {h : EReal | ∃ s : paperMat B, paperSubnormalized s ∧
    h = paperLogE ((paperPower r 0*Matrix.kronecker (1:paperMat A) s).trace.re)}

example (M : paperMat I) : paperMix 0 M M = M := by simp [paperMix]

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

example : paperSubnormalized (0:paperMat I) := ⟨Matrix.PosSemidef.zero, by simp⟩

example (r : paperMat I) (x y : Fin 0 → I) : paperTensorPower r 0 x y = 1 := by simp [paperTensorPower]

example (r : paperMat I) (e : ℝ) (s : paperMat I) : s ∈ paperBall r e ↔ paperSubnormalized s ∧ purified s r ≤ e := Iff.rfl

example (r : paperMat ((A × B) × C)) (a a' : A) (b b' : B) (c c' : C) :
    paperRegroup r (a,(b,c)) (a',(b',c')) = r ((a,b),c) ((a',b'),c') := rfl

example (r : Y → paperMat (A × B)) (a a' : A) (b b' : B) (y : Y) :
    conditionalBlocks r (a,(b,y)) (a',(b',y)) = r y (a,b) (a',b') := by simp [conditionalBlocks]

/-- PDF pp.62,69: unnormalized sandwiched or Petz trace functional. -/
def paperRenyiQ (sand : Bool) (r s : paperMat I) (a : ℝ) : ℝ :=
  if sand then (paperPower (paperPower s ((1-a)/(2*a))*r*paperPower s ((1-a)/(2*a))) a).trace.re
  else (paperPower r a*paperPower s (1-a)).trace.re

/-- PDF pp.61,69: finite positive-order divergence, normalized by Tr r.
The a=1 value is the Umegaki limit, and support/orthogonality guards preserve infinity. -/
def renyiFinite (sand : Bool) (r s : paperMat I) (a : ℝ) : EReal :=
  if a = 1 then paperRelativeD r s
  else if (a < 1 ∧ 0 < paperRenyiQ sand r s a) ∨ paperSupported r s then
    ((1/(a-1):ℝ):EReal)*paperLogE (paperRenyiQ sand r s a/r.trace.re)
  else ⊤

/-- PDF pp.61,69,78: endpoint orders are limits of the increasing finite-order family.
For the sandwiched family infinity is the max-divergence. -/
def paperRenyi (sand : Bool) (r s : paperMat I) (a : ENNReal) : EReal :=
  if a = 0 then sInf {d : EReal | ∃ t : ℝ, 0 < t ∧ d = renyiFinite sand r s t}
  else if a = ⊤ then
    if sand then maxD r s else sSup {d : EReal | ∃ t : ℝ, 0 < t ∧ d = renyiFinite sand r s t}
  else renyiFinite sand r s a.toReal

/-- PDF p.77, Definition 5.2: fixed-reference conditional Renyi entropy. -/
def conditionalDown (sand : Bool) (a : ENNReal) (r : paperMat (A × B)) : EReal :=
  -paperRenyi sand r (Matrix.kronecker (1:paperMat A) (traceLeft r)) a

/-- PDF p.77, Definition 5.2: optimize over actual normalized reference states. -/
def conditionalUp (sand : Bool) (a : ENNReal) (r : paperMat (A × B)) : EReal :=
  sSup {h : EReal | ∃ s : paperMat B, IsDensityMatrix s ∧ h = -paperRenyi sand r (Matrix.kronecker (1:paperMat A) s) a}

/-- PDF p.81: select fixed-reference or optimized entropy without changing the divergence family. -/
def paperConditional (sand up : Bool) (a : ENNReal) (r : paperMat (A × B)) : EReal :=
  if up then conditionalUp sand a r else conditionalDown sand a r

/-- PDF pp.81,86: orders for which this family has data processing, including limiting endpoints. -/
def DPIOrder (sand : Bool) (a : ENNReal) : Prop :=
  if sand then (1/2:ENNReal) ≤ a else a ≤ 2

/-- PDF pp.84-85: real order 2-1/a including a=infinity, with reciprocal endpoint limits. -/
def transformedOrder (a : ENNReal) : ENNReal :=
  if a = ⊤ then 2 else ENNReal.ofReal (2-1/a.toReal)

/-- PDF p.84: regrouped independent product AB,CD to AC,BD. -/
def paperIndependentProduct (r : paperMat (A × B)) (s : paperMat (C × D)) : paperMat ((A × C) × (B × D)) :=
  fun x y => r (x.1.1,x.2.1) (y.1.1,y.2.1)*s (x.1.2,x.2.2) (y.1.2,y.2.2)

/-- PDF p.36: distinct spectral values; zero is included when present. -/
def spectralValues (H : paperMat I) : Finset ℝ := Finset.univ.image (ascEigenvalues H)

/-- PDF p.36: actual spectral projection at l. -/
def paperSpectralProjector (H : paperMat I) (l : ℝ) : paperMat I :=
  cfc (fun x : ℝ => if x = l then (1:ℝ) else 0) H

/-- PDF p.36: spectral pinching of an operator. -/
def paperPinch (H L : paperMat I) : paperMat I :=
  ∑ l ∈ spectralValues H, paperSpectralProjector H l*L*paperSpectralProjector H l

/-- PDF p.70: spectral data is tied to the input matrix by reconstruction. -/
def SpectralRep (r : paperMat I) (v : I → ℝ) (U : paperMat I) : Prop :=
  (∀ i, 0 ≤ v i) ∧ Uᴴ*U = 1 ∧ U*Uᴴ = 1 ∧ r = U*Matrix.diagonal (fun i => (v i:ℂ))*Uᴴ

/-- PDF p.70: Nussbaum-Szkola probability law from the actual two spectral bases. -/
def paperNsP (v : I → ℝ) (U V : paperMat I) : I × I → ℝ :=
  fun xy => v xy.1*‖(Uᴴ*V) xy.1 xy.2‖^2

/-- PDF p.70: Nussbaum-Szkola reference law. -/
def paperNsQ (v : I → ℝ) (U V : paperMat I) : I × I → ℝ :=
  fun xy => v xy.2*‖(Uᴴ*V) xy.1 xy.2‖^2

/-- PDF p.12: a finite normalized classical probability distribution. -/
def paperProbability (p : X → ℝ) : Prop := (∀ x, 0 ≤ p x) ∧ ∑ x, p x = 1

/-- PDF p.76: finite separable states, with actual normalized product components. -/
def paperSeparable (r : paperMat (A × B)) : Prop :=
  ∃ n : ℕ, ∃ w : Fin n → ℝ, ∃ s : Fin n → paperMat A, ∃ t : Fin n → paperMat B,
    paperProbability w ∧ (∀ i, IsDensityMatrix (s i) ∧ IsDensityMatrix (t i)) ∧
      r = ∑ i, (w i:ℂ) • Matrix.kronecker (s i) (t i)

/-- PDF p.116: Born law for the orthonormal basis in the columns of U. -/
def basisLaw (U r : paperMat I) : paperMat I := Matrix.diagonal (fun i => (Uᴴ*r*U) i i)

/-- PDF p.117: local basis measurement on A, retaining the side information B. -/
def paperBasisMeasurement (U : paperMat A) (r : paperMat (A × B)) : paperMat (A × B) :=
  let W := Matrix.kronecker U (1:paperMat B)
  fun x y => if x.1 = y.1 then (Wᴴ*r*W) x y else 0

/-- PDF p.116: maximum SQUARED overlap of the two actual bases. -/
def basisOverlap (U V : paperMat I) : ℝ := sSup {c : ℝ | ∃ i j, c = ‖(Uᴴ*V) i j‖^2}

/-- PDF p.114: type-I error when T accepts rho. -/
def paperTypeI (r T : paperMat I) : ℝ := ((1-T)*r).trace.re

/-- PDF p.114: type-II error when T accepts rho. -/
def paperTypeII (s T : paperMat I) : ℝ := (T*s).trace.re

/-- PDF p.114, (7.6). Printed: minimize alpha(T;sigma) subject to beta(T;rho).
Changed: minimize alpha(T;rho) subject to beta(T;sigma).
Why: (7.1)-(7.2), the sentence before (7.6), and the Stein threshold all fix these roles. -/
def alphaStar (r s : paperMat I) (e : ℝ) : ℝ :=
  sInf {v : ℝ | ∃ T : paperMat I, IsEffect T ∧ paperTypeII s T ≤ e ∧ v = paperTypeI r T}

/-- PDF p.115: minimum type-II error at prescribed type-I tolerance. -/
def paperBetaStar (r s : paperMat I) (e : ℝ) : ℝ :=
  sInf {v : ℝ | ∃ T : paperMat I, IsEffect T ∧ paperTypeI r T ≤ e ∧ v = paperTypeII s T}

/-- PDF p.114: optimal average error with equal priors. -/
def paperSymmetricError (r s : paperMat I) : ℝ :=
  sInf {v : ℝ | ∃ T : paperMat I, IsEffect T ∧ v = (paperTypeI r T+paperTypeII s T)/2}

/-- PDF p.72: Renyi divergences as real-valued functions only on supported pairs. -/
def paperRenyiReal (sand : Bool) (r s : paperMat I) (a : ℝ) : ℝ := (renyiFinite sand r s a).toReal

example (r : paperMat I) (a : ENNReal) :
    conditionalDown true a (fun x y : I × Unit => r x.1 y.1) = 
      -paperRenyi true (fun x y : I × Unit => r x.1 y.1)
        (Matrix.kronecker (1:paperMat I) (traceLeft (fun x y : I × Unit => r x.1 y.1))) a := rfl

example (r s : paperMat I) : renyiFinite false r s 1 = paperRelativeD r s := by simp [renyiFinite]

example (r : paperMat I) (v : I → ℝ) (U : paperMat I) : paperNsP v U U = (fun xy => v xy.1*‖(Uᴴ*U) xy.1 xy.2‖^2) := rfl

example (r : paperMat I) : paperTypeI r 1 = 0 := by simp [paperTypeI]

example (s : paperMat I) : paperTypeII s 0 = 0 := by simp [paperTypeII]

/-- PDF p.69: extended Petz Q on support failure above one. -/
def petzExtendedQ (r s : paperMat I) (a : ℝ) : EReal :=
  if 1 < a ∧ ¬paperSupported r s then ⊤ else (paperRenyiQ false r s a:EReal)

/-- PDF p.12: fixed-cardinality deterministic source code. -/
structure paperSourceCode (X : Type*) (m : ℕ) where
  encode : X → Fin m
  decode : Fin m → X

/-- PDF p.12: actual probability of decoding failure. -/
def paperCodeError (p : X → ℝ) {m : ℕ} (c : paperSourceCode X m) : ℝ :=
  ∑ x, if c.decode (c.encode x) = x then 0 else p x

/-- PDF p.13: classical Renyi entropy at finite positive order other than one. -/
def paperClassicalRenyi (p : X → ℝ) (a : ℝ) : ℝ := Real.logb 2 (∑ x, p x^a)/(1-a)

/-- PDF p.13: Shannon entropy, with zero contributions from zero probabilities. -/
def paperShannon (p : X → ℝ) : ℝ := -∑ x, p x*Real.logb 2 (p x)

/-- PDF p.14: introductory variation distance (different from purified distance). -/
def variation (p q : X → ℝ) : ℝ := (∑ x, |p x-q x|)/2

/-- PDF p.14: smoothing only over normalized classical probability laws. -/
def introSmoothMax (p : X → ℝ) (e : ℝ) : ℝ :=
  sInf {h : ℝ | ∃ q : X → ℝ, paperProbability q ∧ variation p q ≤ e ∧ h = paperClassicalRenyi q (1/2)}

/-- PDF p.13: iid classical product law. -/
def iidLaw (p : X → ℝ) (n : ℕ) : (Fin n → X) → ℝ := fun x => ∏ i, p (x i)

/-- PDF p.13: integer memory cardinality rounded upward from 2^(nR). -/
def codeCardinality (n : ℕ) (R : ℝ) : ℕ := Nat.ceil ((2:ℝ)^((n:ℝ)*R))

/-- PDF p.36: finite orthogonal resolution of identity. -/
def Resolution {m : ℕ} (P : Fin m → paperMat I) : Prop :=
  (∀ i, (P i)ᴴ = P i ∧ P i*P i = P i) ∧ (∀ i j, i ≠ j → P i*P j = 0) ∧ ∑ i, P i = 1

/-- PDF p.36: pinching from a finite resolution. -/
def resolutionPinch {m : ℕ} (P : Fin m → paperMat I) (M : paperMat I) : paperMat I := ∑ i, P i*M*P i

/-- PDF p.40: all unitarily invariant complex norms, rather than a particular matrix norm. -/
def UnitaryInvariantNorm (v : paperMat I → ℝ) : Prop :=
  (∀ M, 0 ≤ v M ∧ (v M = 0 ↔ M = 0)) ∧
  (∀ (z : ℂ) M, v (z • M) = ‖z‖*v M) ∧ (∀ M N, v (M+N) ≤ v M+v N) ∧
  ∀ U V M : paperMat I, Uᴴ*U = 1 → Vᴴ*V = 1 → v (U*M*V) = v M

/-- PDF p.46: normalized hat-state with a single vacuum flag. -/
def hatState (r : paperMat I) : paperMat (Sum I Unit) :=
  fun x y => match x,y with
    | .inl i,.inl j => r i j
    | .inr _,.inr _ => (1-r.trace.re:ℂ)
    | _,_ => 0

/-- PDF p.37: Choi matrix from the UNNORMALIZED entangled vector. -/
def paperChoi (E : paperMat I →ₗ[ℂ] paperMat J) : paperMat (I × J) :=
  fun x y => E (Matrix.single x.1 y.1 1) x.2 y.2

/-- PDF p.37, (2.69): inverse in register order A-prime,B; the input transpose is implicit in the pairing. -/
def inverseChoi (G : paperMat (I × J)) (r : paperMat I) : paperMat J :=
  fun b b' => ∑ i, ∑ j, r i j*G (i,b) (j,b')

/-- PDF p.93: normalized maximally entangled density matrix on two copies of the same basis. -/
def paperEntangled (I : Type*) [Fintype I] [DecidableEq I] : paperMat (I × I) :=
  fun x y => if x.1 = x.2 ∧ y.1 = y.2 then (Fintype.card I:ℂ)⁻¹ else 0

/-- PDF p.94, (6.11): a B-prime-to-A channel applied to the second half of psi_BBprime, then regrouped as AB. -/
def preparedEntangled (E : paperMat B →ₗ[ℂ] paperMat A) : paperMat (A × B) :=
  fun x y => idTensorSuperop E (paperEntangled B) (x.2,x.1) (y.2,y.1)

/-- PDF pp.91-92: the actual Hilbert-Schmidt adjoint, computed from matrix units. -/
def paperAdjointAction (E : paperMat I →ₗ[ℂ] paperMat J) (Y : paperMat J) : paperMat I :=
  fun i j => ∑ a, ∑ b, conj (E (Matrix.single i j 1) a b)*Y a b

/-- PDF p.92: extended primal infimum; infeasibility is +infinity. -/
def sdpPrimal (E : paperMat I →ₗ[ℂ] paperMat J) (K : paperMat I) (L : paperMat J) : EReal :=
  sInf {v : EReal | ∃ X : paperMat I, X.PosSemidef ∧ paperBelow L (E X) ∧ v = ((K*X).trace.re:EReal)}

/-- PDF p.92: extended dual supremum; infeasibility is -infinity. -/
def sdpDual (E : paperMat I →ₗ[ℂ] paperMat J) (K : paperMat I) (L : paperMat J) : EReal :=
  sSup {v : EReal | ∃ Y : paperMat J, Y.PosSemidef ∧ paperBelow (paperAdjointAction E Y) K ∧ v = ((L*Y).trace.re:EReal)}

/-- PDF p.109: heterogeneous product states, retaining every factor. -/
def productState {n : ℕ} (r : Fin n → paperMat I) : paperMat (Fin n → I) :=
  fun x y => ∏ i, r i (x i) (y i)

/-- PDF pp.31,42: the actual quadratic form, with conjugation in the first argument. -/
def paperQuadratic (r : paperMat I) (v : I → ℂ) : ℂ := ∑ i, ∑ j, conj (v i)*r i j*v j

example (r : paperMat I) (i j : I) : hatState r (.inl i) (.inl j) = r i j := rfl

example (E : paperMat I →ₗ[ℂ] paperMat J) (i j : I) (b c : J) :
    paperChoi E (i,b) (j,c) = E (Matrix.single i j 1) b c := rfl

example (p : X → ℝ) (x : Fin 0 → X) : iidLaw p 0 x = 1 := by simp [iidLaw]

example (p : X → ℝ) : variation p p = 0 := by simp [variation]

example (K : paperMat I) (L : paperMat J) :
    paperAdjointAction (0 : paperMat I →ₗ[ℂ] paperMat J) L = 0 := by ext i j; simp [paperAdjointAction]

/-- PDF pp.35,68: actual classical probability matrix of a POVM output. -/
def measureOutput {m : ℕ} (M : Fin m → paperMat I) (r : paperMat I) : paperMat (Fin m) :=
  Matrix.diagonal (fun x => (M x*r).trace)

/-- PDF p.14: minimum code alphabet cardinality. On probability sources the feasible set is nonempty (lossless coding). -/
def optimalCardinality (p : X → ℝ) (e : ℝ) : ℕ :=
  sInf {m : ℕ | 0 < m ∧ ∃ c : paperSourceCode X m, paperCodeError p c ≤ e}

/-- PDF pp.52-58: a divergence is one family across all finite systems,
so classical outputs, tensor powers, direct sums and rectangular channels have the same value rule. -/
structure paperDivergenceFamily where
  value : (I : Type) → [Fintype I] → [DecidableEq I] → paperMat I → paperMat I → EReal

/-- PDF pp.56,58,61: agreement with the classical divergence on diagonal positive pairs. -/
def paperClassicalAgreement (D : paperDivergenceFamily) (a : ENNReal) : Prop :=
  ∀ (I : Type) [Fintype I] [DecidableEq I] [Nonempty I], ∀ p q : I → ℝ,
    (∀ i, 0 ≤ p i) → (∀ i, 0 ≤ q i) → p ≠ 0 →
    D.value I (Matrix.diagonal (fun i => (p i:ℂ))) (Matrix.diagonal (fun i => (q i:ℂ))) = 
      paperRenyi false (Matrix.diagonal (fun i => (p i:ℂ))) (Matrix.diagonal (fun i => (q i:ℂ))) a

/-- PDF p.52, (V): tensor additivity, on arbitrary nonzero positive numerators. -/
def AdditiveDivergence (D : paperDivergenceFamily) : Prop :=
  ∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ r s : paperMat I, ∀ t u : paperMat J,
    r.PosSemidef → s.PosSemidef → t.PosSemidef → u.PosSemidef → r ≠ 0 → t ≠ 0 →
    D.value (I × J) (Matrix.kronecker r t) (Matrix.kronecker s u) = D.value I r s+D.value J t u

/-- PDF p.53, (VIII): every CPTP map, including maps between different dimensions. -/
def paperDataProcessing (D : paperDivergenceFamily) : Prop :=
  ∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ E : paperMat I →ₗ[ℂ] paperMat J,
    paperCPTP E → ∀ r s : paperMat I, r.PosSemidef → s.PosSemidef → r ≠ 0 →
      D.value J (E r) (E s) ≤ D.value I r s

/-- PDF pp.60-61: the concrete max-divergence family. -/
def paperMaxFamily : paperDivergenceFamily := ⟨fun _ _ _ r s => maxD r s⟩

/-- PDF p.66, (4.73): optimization over actual finite-outcome POVMs, in EReal. -/
def measuredRenyi (r s : paperMat I) (a : ENNReal) : EReal :=
  sSup {d : EReal | ∃ m : ℕ, 0 < m ∧ ∃ M : Fin m → paperMat I, IsPOVM M ∧
    d = paperRenyi false (measureOutput M r) (measureOutput M s) a}

/-- PDF p.66, (4.73): the measured representation for every positive input pair,
including attainment of each finite-copy maximum. This predicate does not assume DPI. -/
def MeasuredRepresentation (D : paperDivergenceFamily) (a : ENNReal) : Prop :=
  ∀ (I : Type) [Fintype I] [DecidableEq I] [Nonempty I], ∀ r s : paperMat I,
    r.PosSemidef → s.PosSemidef → r ≠ 0 →
    Tendsto (fun n : ℕ => ((1/(n:ℝ):ℝ):EReal)*measuredRenyi (paperTensorPower r n) (paperTensorPower s n) a)
      atTop (nhds (D.value I r s)) ∧
    ∀ n : ℕ, 0 < n → ∃ m : ℕ, 0 < m ∧ ∃ M : Fin m → paperMat (Fin n → I), IsPOVM M ∧
      measuredRenyi (paperTensorPower r n) (paperTensorPower s n) a = 
        paperRenyi false (measureOutput M (paperTensorPower r n)) (measureOutput M (paperTensorPower s n)) a

example {K : Type} [Fintype K] [DecidableEq K] (r s : paperMat K) :
    paperMaxFamily.value K r s = maxD r s := rfl

example (E : paperMat A →ₗ[ℂ] paperMat C) (F : paperMat B →ₗ[ℂ] paperMat D) :
    paperTensorAction E F (0:paperMat (A × B)) = 0 := by ext x y; simp [paperTensorAction]

/-- PDF p.52: finite real domain of the six axioms. -/
def AxiomDomain (r s : paperMat I) : Prop := r.PosSemidef ∧ s.PosSemidef ∧ r ≠ 0 ∧ paperSupported r s

/-- PDF p.52: real-valued divergence across every finite system, before adjoining infinities. -/
abbrev RealDivergenceFamily := (I : Type) → [Fintype I] → [DecidableEq I] → paperMat I → paperMat I → ℝ

/-- PDF p.52: direct sum in the disjoint-union basis. -/
def paperDirectSum (r : paperMat I) (s : paperMat J) : paperMat (I ⊕ J) :=
  fun x y => match x,y with
    | .inl i,.inl j => r i j
    | .inr i,.inr j => s i j
    | _,_ => 0

/-- PDF p.52, (I)-(VI): the actual unbounded positive cone and the full trace-weighted mean.
The accepted Dependency50.P35.SixAxioms restricts the mean to total traces at most one;
this book states the mean for arbitrary positive inputs, so it has a distinct local predicate. -/
def paperSixAxioms (D : RealDivergenceFamily) (g : ℝ → ℝ) : Prop :=
  Continuous g ∧ (StrictMono g ∨ StrictAnti g) ∧
  (∀ (I : Type) [Fintype I] [DecidableEq I] [Nonempty I],
    ContinuousOn (fun x : paperMat I × paperMat I => D I x.1 x.2) {x | AxiomDomain x.1 x.2}) ∧
  (∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ r s : paperMat I, ∀ U : Matrix J I ℂ,
    AxiomDomain r s → Uᴴ*U = 1 → U*Uᴴ = 1 → D I r s = D J (U*r*Uᴴ) (U*s*Uᴴ)) ∧
  D (Fin 1) 1 ((1/2:ℂ) • (1:paperMat (Fin 1))) = Real.logb 2 2 ∧
  (∀ (I : Type) [Fintype I] [DecidableEq I] [Nonempty I], ∀ r s : paperMat I,
    AxiomDomain r s → (paperBelow s r → 0 ≤ D I r s) ∧ (paperBelow r s → D I r s ≤ 0)) ∧
  (∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ r s : paperMat I, ∀ t u : paperMat J,
    AxiomDomain r s → AxiomDomain t u →
    D (I × J) (Matrix.kronecker r t) (Matrix.kronecker s u) = D I r s+D J t u) ∧
  (∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ r s : paperMat I, ∀ t u : paperMat J,
    AxiomDomain r s → AxiomDomain t u →
    g (D (I ⊕ J) (paperDirectSum r t) (paperDirectSum s u)) = 
      r.trace.re/(r.trace.re+t.trace.re)*g (D I r s)+t.trace.re/(r.trace.re+t.trace.re)*g (D J t u))

/-- PDF pp.53,57: joint convexity or concavity of g(D) on normalized pairs. -/
def MeanCurvature (D : RealDivergenceFamily) (g : ℝ → ℝ) : Prop :=
  ∀ (I : Type) [Fintype I] [DecidableEq I] [Nonempty I], ∀ r s t u : paperMat I,
    AxiomDomain r s → AxiomDomain t u →
    IsDensityMatrix r → IsDensityMatrix s → IsDensityMatrix t → IsDensityMatrix u →
    ∀ l : ℝ, l ∈ Set.Icc 0 1 →
      (StrictMono g → g (D I (paperMix l r t) (paperMix l s u)) ≤ l*g (D I r s)+(1-l)*g (D I t u)) ∧
      (StrictAnti g → l*g (D I r s)+(1-l)*g (D I t u) ≤ g (D I (paperMix l r t) (paperMix l s u)))

/-- PDF pp.53,57: data processing on the finite domain where the axiomatic D is real. -/
def RealDataProcessing (D : RealDivergenceFamily) : Prop :=
  ∀ (I J : Type) [Fintype I] [DecidableEq I] [Nonempty I]
    [Fintype J] [DecidableEq J] [Nonempty J], ∀ E : paperMat I →ₗ[ℂ] paperMat J,
    paperCPTP E → ∀ r s : paperMat I, AxiomDomain r s → D J (E r) (E s) ≤ D I r s

example (r : paperMat I) (s : paperMat J) (i j : I) : paperDirectSum r s (.inl i) (.inl j) = r i j := rfl

example (r : paperMat I) (h : AxiomDomain r r) : r ≠ 0 := h.2.2.1

/-- Definition sanity: the exponential has the extended endpoint values. -/
example : expE ⊥ = 0 ∧ expE ⊤ = ⊤ := by simp [expE]

/-- Definition sanity: the zero matrix remains zero at support exponent zero. -/
example : paperPower (0:paperMat I) 0 = 0 := by simp [paperPower, cfc_zero]

/-- Definition sanity: zero radius around the zero center is outside the strict chapter domain. -/
example : ¬Radius (0:paperMat I) 0 := by simp [Radius]

/-- Definition sanity: the order transformation maps infinity to collision order two. -/
example : transformedOrder ⊤ = 2 := by simp [transformedOrder]

/-- Definition sanity: two zero subnormalized states have generalized fidelity one. -/
example : paperGeneralizedFidelity (0:paperMat I) 0 = 1 := by
  simp [paperGeneralizedFidelity, rootFidelity, cfc_zero]

/-- Definition sanity: algebraic fidelity at zero gives extended max-entropy minus infinity. -/
example : hmax (0:paperMat (A × B)) = ⊥ := by
  apply le_antisymm
  · apply sSup_le
    rintro h ⟨s, hs, rfl⟩
    simp [paperLogE, paperFidelity, rootFidelity, cfc_zero]
  · exact bot_le

/-- Definition sanity: a one-symbol source is decoded exactly by a one-symbol code. -/
example : paperCodeError (fun _ : Unit => (1:ℝ))
    (⟨fun _ => 0, fun _ => ()⟩ : paperSourceCode Unit 1) = 0 := by
  simp [paperCodeError]

end QAlgorithms.Papers.Independent.Intakenext100.P07
Source
Quantum Information Processing with Finite Resources -- Mathematical Foundations, arXiv:1504.00233v5, https://arxiv.org/abs/1504.00233v5
Read-back

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

QAlgorithms.Papers.Independent.Intakenext100.P07.AdditiveDivergence

For an assignment DDD taking pairs of complex square matrices on finite types with decidable equality to extended real numbers, this property says that for every pair of nonempty finite types I,JI,JI,J with decidable equality, every positive semidefinite III-indexed matrices r,sr,sr,s, and every positive semidefinite JJJ-indexed matrices t,ut,ut,u, if r≠0r\ne0r=0 and t≠0t\ne0t=0, then

DI×J(r⊗t, s⊗u)=DI(r,s)+DJ(t,u).D_{I\times J}(r\otimes t,\ s\otimes u) =D_I(r,s)+D_J(t,u).DI×J​(r⊗t, s⊗u)=DI​(r,s)+DJ​(t,u).

The tensor products are Kronecker products, and the sum on the right uses extended real addition. There is no trace normalization or support hypothesis; sss and uuu are allowed to be zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.AxiomDomain

For square complex matrices r,sr,sr,s indexed by a finite type III, this predicate means exactly that rrr and sss are positive semidefinite, r≠0r\ne0r=0, and ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr, where the kernels are those of their actions on complex coordinate vectors. It imposes no trace normalization and does not require sss to be invertible; on an empty index type its requirement r≠0r\ne0r=0 cannot hold.

QAlgorithms.Papers.Independent.Intakenext100.P07.CPTNI

For finite index types I,JI,JI,J and a complex linear map EEE from complex matrices indexed by III to complex matrices indexed by JJJ, this predicate means that both of the following hold: for every natural number kkk, including 000, and every positive semidefinite matrix XXX indexed by {0,…,k−1}×I\{0,\ldots,k-1\}\times I{0,…,k−1}×I, the matrix with entries Y(u,j),(v,j′)=[E ⁣((X(u,i),(v,i′))i,i′∈I)]j,j′\displaystyle Y_{(u,j),(v,j')}=\left[E\!\left((X_{(u,i),(v,i')})_{i,i'\in I}\right)\right]_{j,j'}Y(u,j),(v,j′)​=[E((X(u,i),(v,i′)​)i,i′∈I​)]j,j′​ is positive semidefinite; and for every positive semidefinite matrix rrr indexed by III, Re⁡tr⁡(E(r))≤Re⁡tr⁡r\operatorname{Re}\operatorname{tr}(E(r))\le\operatorname{Re}\operatorname{tr}rRetr(E(r))≤Retrr. The trace inequality applies to all positive semidefinite inputs, without a trace normalization requirement, and empty index types are allowed.

QAlgorithms.Papers.Independent.Intakenext100.P07.ClassicalLeft

For index types X,AX,AX,A and any complex matrix rrr indexed by X×AX\times AX×A, this predicate means that, for every x,y∈Xx,y\in Xx,y∈X and a,b∈Aa,b\in Aa,b∈A, x≠yx\ne yx=y implies r(x,a),(y,b)=0r_{(x,a),(y,b)}=0r(x,a),(y,b)​=0. Entries within each fixed XXX-block are unrestricted, and no positivity or normalization is required. The condition is vacuous if AAA is empty or XXX has at most one element.

QAlgorithms.Papers.Independent.Intakenext100.P07.ClassicalXY

For index types A,X,B,YA,X,B,YA,X,B,Y and any complex matrix rrr indexed by (A×X)×(B×Y)(A\times X)\times(B\times Y)(A×X)×(B×Y), this predicate means that, for every a,a′∈Aa,a'\in Aa,a′∈A, x,x′∈Xx,x'\in Xx,x′∈X, b,b′∈Bb,b'\in Bb,b′∈B, and y,y′∈Yy,y'\in Yy,y′∈Y, the condition x≠x′x\ne x'x=x′ or y≠y′y\ne y'y=y′ implies r((a,x),(b,y)),((a′,x′),(b′,y′))=0\displaystyle r_{((a,x),(b,y)),((a',x'),(b',y'))}=0r((a,x),(b,y)),((a′,x′),(b′,y′))​=0. Entries with x=x′x=x'x=x′ and y=y′y=y'y=y′ are unrestricted, and no positivity or normalization is assumed. The condition is vacuous if any indexing factor is empty, or if both XXX and YYY have at most one element.

QAlgorithms.Papers.Independent.Intakenext100.P07.DPIOrder

For a Boolean bbb and an extended nonnegative real α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞], this predicate means 12≤α\frac12\le\alpha21​≤α when b=trueb=\mathrm{true}b=true, and α≤2\alpha\le2α≤2 when b=falseb=\mathrm{false}b=false. Thus the true branch includes α=+∞\alpha=+\inftyα=+∞ and excludes 000, while the false branch includes 000 and excludes +∞+\infty+∞; both bounds are inclusive.

QAlgorithms.Papers.Independent.Intakenext100.P07.MeanCurvature

Let DI(r,s)D_I(r,s)DI​(r,s) be a real-valued rule on pairs of square complex matrices for every finite index type III with decidable equality, and let g:R→Rg:\mathbb R\to\mathbb Rg:R→R. This predicate says that for every such nonempty III, every four positive semidefinite matrices r,s,t,ur,s,t,ur,s,t,u on III each having complex trace 111, with r≠0r\ne0r=0, t≠0t\ne0t=0, ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr, and ker⁡u⊆ker⁡t\ker u\subseteq\ker tkeru⊆kert, and every real l∈[0,1]l\in[0,1]l∈[0,1], both implications

g strictly increasing⟹g ⁣(DI(lr+(1−l)t,  ls+(1−l)u))≤l g(DI(r,s))+(1−l) g(DI(t,u)),g strictly decreasing⟹l g(DI(r,s))+(1−l) g(DI(t,u))≤g ⁣(DI(lr+(1−l)t,  ls+(1−l)u))\begin{aligned} g\text{ strictly increasing}\quad&\Longrightarrow\quad g\!\left(D_I(lr+(1-l)t,\;ls+(1-l)u)\right) \le l\,g(D_I(r,s))+(1-l)\,g(D_I(t,u)),\\ g\text{ strictly decreasing}\quad&\Longrightarrow\quad l\,g(D_I(r,s))+(1-l)\,g(D_I(t,u)) \le g\!\left(D_I(lr+(1-l)t,\;ls+(1-l)u)\right) \end{aligned}g strictly increasingg strictly decreasing​⟹g(DI​(lr+(1−l)t,ls+(1−l)u))≤lg(DI​(r,s))+(1−l)g(DI​(t,u)),⟹lg(DI​(r,s))+(1−l)g(DI​(t,u))≤g(DI​(lr+(1−l)t,ls+(1−l)u))​

hold. The endpoints l=0,1l=0,1l=0,1 are included. Continuity or monotonicity of ggg is not itself required by this predicate; if ggg is neither strictly increasing nor strictly decreasing, both implications hold vacuously.

QAlgorithms.Papers.Independent.Intakenext100.P07.MeasuredRepresentation

For an extended real-valued assignment DDD on pairs of complex square matrices over finite types with decidable equality, and an order a∈[0,∞]a\in[0,\infty]a∈[0,∞], this property says the following for every nonempty finite type III with decidable equality and every positive semidefinite pair r,sr,sr,s indexed by III, with r≠0r\ne0r=0. Define r[n]r^{[n]}r[n], indexed by functions {0,…,n−1}→I\{0,\ldots,n-1\}\to I{0,…,n−1}→I, by (r[n])x,y=∏i<nrxi,yi(r^{[n]})_{x,y}=\prod_{i<n}r_{x_i,y_i}(r[n])x,y​=∏i<n​rxi​,yi​​, and define s[n]s^{[n]}s[n] similarly. For matrices A,BA,BA,B, define Ma(A,B)\mathfrak M_a(A,B)Ma​(A,B) as the extended real supremum, over all integers m>0m>0m>0 and all positive semidefinite families M0,…,Mm−1M_0,\ldots,M_{m-1}M0​,…,Mm−1​ summing to the identity, of Ra(M(A),M(B))R_a(\mathcal M(A),\mathcal M(B))Ra​(M(A),M(B)), where M(A)=diag⁡(tr⁡(MiA))i<m\mathcal M(A)=\operatorname{diag}(\operatorname{tr}(M_iA))_{i<m}M(A)=diag(tr(Mi​A))i<m​. Here RaR_aRa​ is defined by the following formulas: Pt(A)P_t(A)Pt​(A) is functional calculus for the function 000 at 000 and xtx^txt elsewhere; L(A)L(A)L(A) is functional calculus for the total real logarithm log⁡2x\log_2xlog2​x, with value 000 at 000; A≪BA\ll BA≪B means ker⁡B⊆ker⁡A\ker B\subseteq\ker AkerB⊆kerA; and ℓ(0)=−∞\ell(0)=-\inftyℓ(0)=−∞, ℓ(x)=log⁡2x\ell(x)=\log_2xℓ(x)=log2​x otherwise. For t>0t>0t>0, put Qt=Re⁡tr⁡(Pt(A)P1−t(B))Q_t=\operatorname{Re}\operatorname{tr}(P_t(A)P_{1-t}(B))Qt​=Retr(Pt​(A)P1−t​(B)); set F1(A,B)=Re⁡tr⁡(A(L(A)−L(B)))/Re⁡tr⁡AF_1(A,B)=\operatorname{Re}\operatorname{tr}(A(L(A)-L(B)))/\operatorname{Re}\operatorname{tr}AF1​(A,B)=Retr(A(L(A)−L(B)))/RetrA when A≪BA\ll BA≪B, and +∞+\infty+∞ otherwise; for t≠1t\ne1t=1, set Ft(A,B)=(t−1)−1ℓ(Qt/Re⁡tr⁡A)F_t(A,B)=(t-1)^{-1}\ell(Q_t/\operatorname{Re}\operatorname{tr}A)Ft​(A,B)=(t−1)−1ℓ(Qt​/RetrA) when (t<1 and Qt>0)(t<1\text{ and }Q_t>0)(t<1 and Qt​>0) or A≪BA\ll BA≪B, and +∞+\infty+∞ otherwise. Take Ra=FaR_a=F_aRa​=Fa​ for finite positive aaa, R0=inf⁡t>0FtR_0=\inf_{t>0}F_tR0​=inft>0​Ft​, and R∞=sup⁡t>0FtR_\infty=\sup_{t>0}F_tR∞​=supt>0​Ft​, using total real division and extended real arithmetic. The required conclusions are

1n Ma(r[n],s[n])⟶DI(r,s)(n→∞)\frac1n\,\mathfrak M_a(r^{[n]},s^{[n]}) \longrightarrow D_I(r,s) \qquad(n\to\infty)n1​Ma​(r[n],s[n])⟶DI​(r,s)(n→∞)

in the topology of the extended reals, and, for every natural number n>0n>0n>0, the supremum defining Ma(r[n],s[n])\mathfrak M_a(r^{[n]},s^{[n]})Ma​(r[n],s[n]) is attained by some finite measurement with a positive number of outcomes. The sequence is defined also at n=0n=0n=0: the tensor products then have every entry equal to the empty product 111, and the real factor 1/01/01/0 is 000. No normalization of r,sr,sr,s is required.

QAlgorithms.Papers.Independent.Intakenext100.P07.MinDual

For finite index types A,BA,BA,B with decidable equality, allowing either to be empty, and a complex matrix WWW indexed by A×BA\times BA×B, this predicate means precisely that WWW is positive semidefinite and

IB−Tr⁡AWI_B-\operatorname{Tr}_A WIB​−TrA​W

is positive semidefinite, where the partial trace is defined entrywise by

(Tr⁡AW)b,b′=∑a∈AW(a,b),(a,b′).(\operatorname{Tr}_A W)_{b,b'}=\sum_{a\in A}W_{(a,b),(a,b')}.(TrA​W)b,b′​=a∈A∑​W(a,b),(a,b′)​.

No trace normalization of WWW is required. If AAA is empty, the displayed partial trace is the zero matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.MinPrimal

For finite index types A,BA,BA,B with decidable equality, allowing either to be empty, a complex matrix rrr indexed by A×BA\times BA×B, and a complex matrix sss indexed by BBB, this predicate means precisely that sss is positive semidefinite and

IA⊗s−rI_A\otimes s-rIA​⊗s−r

is positive semidefinite, where IAI_AIA​ is the identity matrix and ⊗\otimes⊗ is the Kronecker product. It imposes no separate positivity or trace condition on rrr and no trace bound on sss.

QAlgorithms.Papers.Independent.Intakenext100.P07.Radius

For a finite index type III, an arbitrary complex matrix r∈CI×Ir\in\mathbb C^{I\times I}r∈CI×I, and a real number eee, this defines the proposition

0≤eande<Re⁡tr⁡r.0\le e \quad\text{and}\quad e<\sqrt{\operatorname{Re}\operatorname{tr}r}.0≤eande<Retrr​.

No positivity or normalization hypothesis is imposed on rrr. The real square root is 000 on nonpositive inputs, so this proposition cannot hold if Re⁡tr⁡r≤0\operatorname{Re}\operatorname{tr}r\le0Retrr≤0, including when III is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.RealDataProcessing

For a real-valued rule DI(r,s)D_I(r,s)DI​(r,s) on pairs of square complex matrices on every finite index type with decidable equality, this predicate says the following: for every pair of finite nonempty types I,JI,JI,J with decidable equality, every complex linear map EEE from matrices on III to matrices on JJJ that is completely positive and preserves the complex trace of every matrix, and every pair of positive semidefinite matrices r,sr,sr,s on III satisfying r≠0r\ne0r=0 and ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr,

DJ(E(r),E(s))≤DI(r,s).D_J(E(r),E(s))\le D_I(r,s).DJ​(E(r),E(s))≤DI​(r,s).

Complete positivity here means that for every natural number kkk, including 000, and every positive semidefinite matrix XXX on Fin⁡k×I\operatorname{Fin}k\times IFink×I, the matrix on Fin⁡k×J\operatorname{Fin}k\times JFink×J whose ((p,j),(q,j′))((p,j),(q,j'))((p,j),(q,j′))-entry is

[E ⁣((X(p,i),(q,i′))i,i′∈I)]j,j′\left[E\!\left((X_{(p,i),(q,i')})_{i,i'\in I}\right)\right]_{j,j'}[E((X(p,i),(q,i′)​)i,i′∈I​)]j,j′​

is positive semidefinite. The inputs r,sr,sr,s need not have trace one.

QAlgorithms.Papers.Independent.Intakenext100.P07.RealDivergenceFamily

This is the type of rules DDD that, for every type III, every specified finite enumeration of III, and every specified decidable equality on III, take two square complex matrices indexed by III and return a real number. Empty index types are included. No positivity, normalization, continuity, or other property of the rule or its matrix arguments is required by this abbreviation.

QAlgorithms.Papers.Independent.Intakenext100.P07.Resolution

For a finite index type III with decidable equality, a natural number mmm, and complex matrices P0,…,Pm−1P_0,\ldots,P_{m-1}P0​,…,Pm−1​ indexed by III, this predicate requires Pi†=PiP_i^\dagger=P_iPi†​=Pi​ and Pi2=PiP_i^2=P_iPi2​=Pi​ for every iii, PiPj=0P_iP_j=0Pi​Pj​=0 for every pair of distinct indices i,ji,ji,j, and ∑i=0m−1Pi=I\sum_{i=0}^{m-1}P_i=I∑i=0m−1​Pi​=I. Individual matrices may be zero. For m=0m=0m=0, the requirement reduces to 0=I0=I0=I, which holds for empty III and fails for nonempty III.

QAlgorithms.Papers.Independent.Intakenext100.P07.SmoothPrimal

For finite index types A,B,CA,B,CA,B,C with decidable equality, allowing empty types, complex matrices r,tr,tr,t indexed by (A×B)×C(A\times B)\times C(A×B)×C, a complex matrix sss indexed by BBB, and every real eee, this predicate means that all four conditions hold:

t is positive semidefinite,Re⁡tr⁡t≤1,s is positive semidefinite,1−e2≤Re⁡tr⁡(tr),t\text{ is positive semidefinite},\qquad \operatorname{Re}\operatorname{tr}t\le1,\qquad s\text{ is positive semidefinite},\qquad 1-e^2\le\operatorname{Re}\operatorname{tr}(tr),t is positive semidefinite,Retrt≤1,s is positive semidefinite,1−e2≤Retr(tr),

and, additionally, IA⊗s−Tr⁡CtI_A\otimes s-\operatorname{Tr}_C tIA​⊗s−TrC​t is positive semidefinite, where

(Tr⁡Ct)(a,b),(a′,b′)=∑c∈Ct((a,b),c),((a′,b′),c).(\operatorname{Tr}_C t)_{(a,b),(a',b')} =\sum_{c\in C}t_{((a,b),c),((a',b'),c)}.(TrC​t)(a,b),(a′,b′)​=c∈C∑​t((a,b),c),((a′,b′),c)​.

Here trtrtr denotes the matrix product of ttt and rrr. No positivity or normalization of rrr, trace bound on sss, or restriction on eee is imposed; eee enters only through e2e^2e2, and an empty CCC makes the partial trace zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.SpectralRep

For a finite index type III with decidable equality, complex square matrices r,Ur,Ur,U indexed by III, and a real function v:I→Rv:I\to\mathbb Rv:I→R, this predicate means precisely that vi≥0v_i\geq0vi​≥0 for every iii, U†U=1U^\dagger U=1U†U=1, UU†=1UU^\dagger=1UU†=1, and

r=U diag⁡(vi) U†,r=U\,\operatorname{diag}(v_i)\,U^\dagger,r=Udiag(vi​)U†,

where †\dagger† denotes conjugate transpose and the real diagonal entries are regarded as complex numbers. No condition requires ∑ivi=1\sum_i v_i=1∑i​vi​=1. Empty III is allowed, in which case the entrywise condition is vacuous and all the displayed matrix equalities hold for the unique empty matrices.

QAlgorithms.Papers.Independent.Intakenext100.P07.UnitaryInvariantNorm

For a finite index type III with decidable equality and a function v:CI×I→Rv:\mathbb C^{I\times I}\to\mathbb Rv:CI×I→R, this predicate means that all four conditions hold: every matrix MMM satisfies v(M)≥0v(M)\ge0v(M)≥0 and v(M)=0v(M)=0v(M)=0 if and only if M=0M=0M=0; every complex scalar zzz and matrix MMM satisfy v(zM)=∣z∣v(M)v(zM)=|z|v(M)v(zM)=∣z∣v(M); every pair of matrices M,NM,NM,N satisfies v(M+N)≤v(M)+v(N)v(M+N)\le v(M)+v(N)v(M+N)≤v(M)+v(N); and every triple of matrices U,V,MU,V,MU,V,M satisfying U†U=1U^\dagger U=1U†U=1 and V†V=1V^\dagger V=1V†V=1 satisfies v(UMV)=v(M)v(UMV)=v(M)v(UMV)=v(M), where †\dagger† denotes conjugate transpose and 111 is the identity matrix. The index type III is allowed to be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.alphaStar

For any finite index type III with decidable equality, any complex matrices r,sr,sr,s indexed by I×II\times II×I, and any real eee, this defines

α∗(r,s,e)=inf⁡{Re⁡tr⁡((1−T)r):T⪰0, 1−T⪰0, Re⁡tr⁡(Ts)≤e}.\alpha_*(r,s,e)= \inf\left\{ \operatorname{Re}\operatorname{tr}((\mathbf 1-T)r): T\succeq0,\ \mathbf 1-T\succeq0,\ \operatorname{Re}\operatorname{tr}(Ts)\le e \right\}.α∗​(r,s,e)=inf{Retr((1−T)r):T⪰0, 1−T⪰0, Retr(Ts)≤e}.

Here ⪰0\succeq0⪰0 means positive semidefinite, and the infimum is taken in the real numbers, with the convention that an empty set or a set unbounded below has infimum 000. There is no assumption that rrr or sss is positive semidefinite or normalized, and eee is unrestricted; thus the feasible set can be empty. Empty III is allowed and gives value 000, including when e<0e<0e<0 makes the feasible set empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.basisLaw

For a finite index type III with decidable equality and arbitrary complex square matrices U,rU,rU,r indexed by III, this defines the diagonal complex matrix DDD with

Dij={(U†rU)ii,i=j,0,i≠j.D_{ij}= \begin{cases} (U^\dagger rU)_{ii},&i=j,\\ 0,&i\ne j. \end{cases}Dij​={(U†rU)ii​,0,​i=j,i=j.​

Here †\dagger† denotes conjugate transpose. No unitarity, positivity, or normalization requirement is imposed. Empty III gives the unique empty matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.basisOverlap

For a finite index type III and arbitrary complex square matrices U,VU,VU,V indexed by III, this defines

sup⁡{∣(U†V)ij∣2:i,j∈I},\sup\left\{\bigl|(U^\dagger V)_{ij}\bigr|^2:i,j\in I\right\},sup{​(U†V)ij​​2:i,j∈I},

where †\dagger† denotes conjugate transpose and absolute value denotes complex modulus. When III is nonempty, this is the maximum of these finitely many real numbers. When III is empty, the real supremum convention used here gives 000. No unitarity requirement is imposed on either matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.classicalProcess

For index types X,Z,A,BX,Z,A,BX,Z,A,B, with XXX finite and equality on ZZZ decidable, a function f:X→Zf:X\to Zf:X→Z, and a complex matrix rrr indexed by (X×A)×B(X\times A)\times B(X×A)×B, this defines a complex matrix indexed by (Z×A)×B(Z\times A)\times B(Z×A)×B whose entries are

(classicalProcess⁡(f,r))((z,a),b),((z′,a′),b′)={∑x∈Xf(x)=zr((x,a),b),((x,a′),b′),z=z′,0,z≠z′.\bigl(\operatorname{classicalProcess}(f,r)\bigr)_{((z,a),b),((z',a'),b')} = \begin{cases} \displaystyle\sum_{\substack{x\in X\\f(x)=z}} r_{((x,a),b),((x,a'),b')},&z=z',\\ 0,&z\ne z'. \end{cases}(classicalProcess(f,r))((z,a),b),((z′,a′),b′)​=⎩⎨⎧​x∈Xf(x)=z​∑​r((x,a),b),((x,a′),b′)​,0,​z=z′,z=z′.​

No positivity, normalization, or block structure of rrr is assumed. An empty fiber of fff contributes zero; in particular, if XXX is empty, every output entry is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.codeCardinality

For every natural number nnn and real number RRR, this defines the natural-number ceiling ⌈2nR⌉N\lceil 2^{nR}\rceil_{\mathbb N}⌈2nR⌉N​. Negative and zero values of RRR, and n=0n=0n=0, are included. The value is one whenever n=0n=0n=0 or R≤0R\le0R≤0.

QAlgorithms.Papers.Independent.Intakenext100.P07.conditionalBlocks

For index types Y,A,BY,A,BY,A,B, with equality on YYY decidable, and any family rrr assigning a complex matrix indexed by A×BA\times BA×B to each y∈Yy\in Yy∈Y, this defines a complex matrix indexed by A×(B×Y)A\times(B\times Y)A×(B×Y) through

(conditionalBlocks⁡(r))(a,(b,y)),(a′,(b′,y′))={(ry)(a,b),(a′,b′),y=y′,0,y≠y′.\bigl(\operatorname{conditionalBlocks}(r)\bigr)_{(a,(b,y)),(a',(b',y'))} = \begin{cases} (r_y)_{(a,b),(a',b')},&y=y',\\ 0,&y\ne y'. \end{cases}(conditionalBlocks(r))(a,(b,y)),(a′,(b′,y′))​={(ry​)(a,b),(a′,b′)​,0,​y=y′,y=y′.​

No positivity, normalization, or weighting of the matrices ryr_yry​ is required. Empty index types are allowed; if YYY is empty, the output matrix has an empty index type.

QAlgorithms.Papers.Independent.Intakenext100.P07.conditionalDown

Let A,BA,BA,B be finite types with decidable equality, possibly empty, rrr any complex matrix on A×BA\times BA×B, bbb a Boolean, and α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞]. Define rBr_BrB​ by (rB)j,k=∑i∈Ar(i,j),(i,k)(r_B)_{j,k}=\sum_{i\in A}r_{(i,j),(i,k)}(rB​)j,k​=∑i∈A​r(i,j),(i,k)​. To specify the extended-real function used here, let Pp(M)P_p(M)Pp​(M) be real functional calculus for the scalar function equal to 000 at 000 and to xpx^pxp elsewhere; put Qb(u,v;t)=Re⁡tr⁡Pt(P(1−t)/(2t)(v) u P(1−t)/(2t)(v))Q_b(u,v;t)=\operatorname{Re}\operatorname{tr}P_t(P_{(1-t)/(2t)}(v)\,u\,P_{(1-t)/(2t)}(v))Qb​(u,v;t)=RetrPt​(P(1−t)/(2t)​(v)uP(1−t)/(2t)​(v)) for true bbb, and Qb(u,v;t)=Re⁡tr⁡(Pt(u)P1−t(v))Q_b(u,v;t)=\operatorname{Re}\operatorname{tr}(P_t(u)P_{1-t}(v))Qb​(u,v;t)=Retr(Pt​(u)P1−t​(v)) for false bbb. Write S(u,v)S(u,v)S(u,v) for ker⁡v⊆ker⁡u\ker v\subseteq\ker ukerv⊆keru, L(0)=−∞L(0)=-\inftyL(0)=−∞, L(x)=log⁡2xL(x)=\log_2xL(x)=log2​x otherwise, and let matrix logarithms use real functional calculus. Set db(u,v;1)=Re⁡tr⁡[u(log⁡2u−log⁡2v)]/Re⁡tr⁡ud_b(u,v;1)=\operatorname{Re}\operatorname{tr}[u(\log_2u-\log_2v)]/\operatorname{Re}\operatorname{tr}udb​(u,v;1)=Retr[u(log2​u−log2​v)]/Retru if S(u,v)S(u,v)S(u,v), and +∞+\infty+∞ otherwise; for t≠1t\ne1t=1, set db(u,v;t)=1t−1L(Qb(u,v;t)/Re⁡tr⁡u)d_b(u,v;t)=\frac1{t-1}L(Q_b(u,v;t)/\operatorname{Re}\operatorname{tr}u)db​(u,v;t)=t−11​L(Qb​(u,v;t)/Retru) if (t<1 and Qb(u,v;t)>0)(t<1\text{ and }Q_b(u,v;t)>0)(t<1 and Qb​(u,v;t)>0) or S(u,v)S(u,v)S(u,v), and +∞+\infty+∞ otherwise. Define Rb(u,v;0)=inf⁡t>0db(u,v;t)\mathcal R_b(u,v;0)=\inf_{t>0}d_b(u,v;t)Rb​(u,v;0)=inft>0​db​(u,v;t); define Rtrue(u,v;+∞)=inf⁡{ℓ∈R:2ℓv−u⪰0}\mathcal R_{\mathrm{true}}(u,v;+\infty)=\inf\{\ell\in\mathbb R:2^\ell v-u\succeq0\}Rtrue​(u,v;+∞)=inf{ℓ∈R:2ℓv−u⪰0} and Rfalse(u,v;+∞)=sup⁡t>0db(u,v;t)\mathcal R_{\mathrm{false}}(u,v;+\infty)=\sup_{t>0}d_b(u,v;t)Rfalse​(u,v;+∞)=supt>0​db​(u,v;t); and at finite positive α\alphaα, define Rb(u,v;α)=db(u,v;α)\mathcal R_b(u,v;\alpha)=d_b(u,v;\alpha)Rb​(u,v;α)=db​(u,v;α). The declaration defines the value to be

−Rb(r,IA⊗rB;α).-\mathcal R_b(r,I_A\otimes r_B;\alpha).−Rb​(r,IA​⊗rB​;α).

All extrema and the final negation are extended-real. There is no positivity, normalization, or nonzero-trace requirement; real divisions and logarithms are total, with division by zero equal to 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.conditionalUp

Let A,BA,BA,B be finite types with decidable equality, possibly empty, rrr any complex matrix on A×BA\times BA×B, bbb a Boolean, and α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞]. Define Pp(M)P_p(M)Pp​(M) by real functional calculus for the scalar function equal to 000 at 000 and to xpx^pxp elsewhere, and set Qb(u,v;t)=Re⁡tr⁡Pt(P(1−t)/(2t)(v) u P(1−t)/(2t)(v))Q_b(u,v;t)=\operatorname{Re}\operatorname{tr}P_t(P_{(1-t)/(2t)}(v)\,u\,P_{(1-t)/(2t)}(v))Qb​(u,v;t)=RetrPt​(P(1−t)/(2t)​(v)uP(1−t)/(2t)​(v)) when bbb is true, and Qb(u,v;t)=Re⁡tr⁡(Pt(u)P1−t(v))Q_b(u,v;t)=\operatorname{Re}\operatorname{tr}(P_t(u)P_{1-t}(v))Qb​(u,v;t)=Retr(Pt​(u)P1−t​(v)) otherwise. Write S(u,v)S(u,v)S(u,v) for ker⁡v⊆ker⁡u\ker v\subseteq\ker ukerv⊆keru, let L(0)=−∞L(0)=-\inftyL(0)=−∞, L(x)=log⁡2xL(x)=\log_2xL(x)=log2​x otherwise, and interpret matrix logarithms by real functional calculus. Define db(u,v;1)=Re⁡tr⁡[u(log⁡2u−log⁡2v)]/Re⁡tr⁡ud_b(u,v;1)=\operatorname{Re}\operatorname{tr}[u(\log_2u-\log_2v)]/\operatorname{Re}\operatorname{tr}udb​(u,v;1)=Retr[u(log2​u−log2​v)]/Retru if S(u,v)S(u,v)S(u,v), and +∞+\infty+∞ otherwise; for t≠1t\ne1t=1, define db(u,v;t)=1t−1L(Qb(u,v;t)/Re⁡tr⁡u)d_b(u,v;t)=\frac1{t-1}L(Q_b(u,v;t)/\operatorname{Re}\operatorname{tr}u)db​(u,v;t)=t−11​L(Qb​(u,v;t)/Retru) if (t<1 and Qb(u,v;t)>0)(t<1\text{ and }Q_b(u,v;t)>0)(t<1 and Qb​(u,v;t)>0) or S(u,v)S(u,v)S(u,v), and +∞+\infty+∞ otherwise. Define Rb(u,v;0)=inf⁡t>0db(u,v;t)\mathcal R_b(u,v;0)=\inf_{t>0}d_b(u,v;t)Rb​(u,v;0)=inft>0​db​(u,v;t), Rtrue(u,v;+∞)=inf⁡{ℓ∈R:2ℓv−u⪰0}\mathcal R_{\mathrm{true}}(u,v;+\infty)=\inf\{\ell\in\mathbb R:2^\ell v-u\succeq0\}Rtrue​(u,v;+∞)=inf{ℓ∈R:2ℓv−u⪰0}, Rfalse(u,v;+∞)=sup⁡t>0db(u,v;t)\mathcal R_{\mathrm{false}}(u,v;+\infty)=\sup_{t>0}d_b(u,v;t)Rfalse​(u,v;+∞)=supt>0​db​(u,v;t), and Rb(u,v;α)=db(u,v;α)\mathcal R_b(u,v;\alpha)=d_b(u,v;\alpha)Rb​(u,v;α)=db​(u,v;α) at finite positive α\alphaα. The declaration defines

sup⁡s on Bs⪰0, tr⁡s=1[−Rb(r,IA⊗s;α)].\sup_{\substack{s\text{ on }B\\s\succeq0,\ \operatorname{tr}s=1}} \bigl[-\mathcal R_b(r,I_A\otimes s;\alpha)\bigr].s on Bs⪰0, trs=1​sup​[−Rb​(r,IA​⊗s;α)].

This is an extended-real supremum, equal to −∞-\infty−∞ if there is no such sss, in particular when BBB is empty. No positivity, normalization, or nonzero-trace assumption is imposed on rrr; real division by zero is 000, and real powers and logarithms use their total real definitions.

QAlgorithms.Papers.Independent.Intakenext100.P07.conditionalVN

For finite index types A,BA,BA,B with decidable equality and any complex matrix rrr indexed by A×BA\times BA×B, this defines the real number S(r)−S(rB)S(r)-S(r_B)S(r)−S(rB​), where (rB)b,b′=∑a∈Ar(a,b),(a,b′)(r_B)_{b,b'}=\sum_{a\in A}r_{(a,b),(a,b')}(rB​)b,b′​=∑a∈A​r(a,b),(a,b′)​ and S(M)=−Re⁡tr⁡(f(M))S(M)=-\operatorname{Re}\operatorname{tr}(f(M))S(M)=−Retr(f(M)), with f(x)=xlog⁡2xf(x)=x\log_2 xf(x)=xlog2​x applied through the real continuous functional calculus. No positivity, Hermitianity, or trace normalization of rrr is assumed.

QAlgorithms.Papers.Independent.Intakenext100.P07.excessRadius

For a finite index type III with decidable equality, including the empty type, arbitrary complex I×II\times II×I matrices r,sr,sr,s, and every real ℓ\ellℓ, this defines

t=Re⁡tr⁡((r−2ℓs)+),excessRadius⁡(r,s,ℓ)=2t−t2,t=\operatorname{Re}\operatorname{tr}\bigl((r-2^\ell s)_+\bigr), \qquad \operatorname{excessRadius}(r,s,\ell)=\sqrt{2t-t^2},t=Retr((r−2ℓs)+​),excessRadius(r,s,ℓ)=2t−t2​,

where H+H_+H+​ is the real continuous functional calculus of x↦max⁡(x,0)x\mapsto\max(x,0)x↦max(x,0), retaining the positive eigenvalues of Hermitian HHH and returning zero for non-Hermitian HHH. No positivity, normalization, or bound on ttt is assumed. The real square root returns zero when its argument is negative, so the result is zero whenever t∉[0,2]t\notin[0,2]t∈/[0,2], as well as at t=0t=0t=0 and t=2t=2t=2.

QAlgorithms.Papers.Independent.Intakenext100.P07.expE

For every extended real number xxx, expE(x)\mathrm{expE}(x)expE(x) is an extended nonnegative real number. It equals +∞+\infty+∞ if x=+∞x=+\inftyx=+∞, equals 000 if x=−∞x=-\inftyx=−∞, and otherwise equals 2t2^t2t, regarded as an extended nonnegative real, where ttt is the finite real value of xxx.

QAlgorithms.Papers.Independent.Intakenext100.P07.guessObjective

For finite index types XXX and BBB, and arbitrary families rx,Mx∈CB×Br_x,M_x\in\mathbb C^{B\times B}rx​,Mx​∈CB×B, the guessing objective is the real number

∑x∈XRe⁡tr⁡(Mxrx).\sum_{x\in X}\operatorname{Re}\operatorname{tr}(M_xr_x).x∈X∑​Retr(Mx​rx​).

There are no positivity, normalization, or measurement assumptions on either family. If XXX is empty, the sum is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.hartleyConditional

For finite index types A,BA,BA,B with decidable equality and an arbitrary complex matrix rrr indexed by A×BA\times BA×B, define P0(r)=f0(r)P_0(r)=f_0(r)P0​(r)=f0​(r) by functional calculus, where f0(0)=0f_0(0)=0f0​(0)=0 and f0(x)=1f_0(x)=1f0​(x)=1 for every nonzero real xxx. The conditional Hartley quantity is the supremum in the extended real line of

{ℓ ⁣(Re⁡tr⁡ ⁣(P0(r)(IA⊗s))): s⪰0, Re⁡tr⁡s≤1},\left\{ \ell\!\left(\operatorname{Re}\operatorname{tr}\!\left(P_0(r)(I_A\otimes s)\right)\right) :\ s\succeq0,\ \operatorname{Re}\operatorname{tr}s\le1 \right\},{ℓ(Retr(P0​(r)(IA​⊗s))): s⪰0, Retrs≤1},

where ℓ(0)=−∞\ell(0)=-\inftyℓ(0)=−∞ and ℓ(z)=log⁡2z\ell(z)=\log_2 zℓ(z)=log2​z, interpreted using the total real logarithm, for z≠0z\ne0z=0. No positivity or normalization condition is imposed on rrr; s=0s=0s=0 is allowed and contributes −∞-\infty−∞.

QAlgorithms.Papers.Independent.Intakenext100.P07.hatState

For a finite index type III and any complex matrix rrr indexed by III, this definition gives a matrix indexed by the disjoint union of III and a singleton, with block form

r^=(r001−Re⁡tr⁡(r)).\widehat r= \begin{pmatrix} r&0\\ 0&1-\operatorname{Re}\operatorname{tr}(r) \end{pmatrix}.r=(r0​01−Retr(r)​).

The lower-right entry is regarded as a complex number with zero imaginary part. There is no positivity or trace assumption on rrr; if III is empty, the resulting matrix is the 1×11\times11×1 matrix with entry 111.

QAlgorithms.Papers.Independent.Intakenext100.P07.hmax

For finite index types A,BA,BA,B with decidable equality and any complex matrix rrr indexed by A×BA\times BA×B, this defines

sup⁡s∈CB×Bs⪰0, Re⁡tr⁡s≤1L ⁣([Re⁡tr⁡S(S(r)(IA⊗s)S(r))]2)\sup_{\substack{s\in\mathbb C^{B\times B}\\ s\succeq0,\ \operatorname{Re}\operatorname{tr}s\le1}} L\!\left( \left[\operatorname{Re}\operatorname{tr} S\bigl(S(r)(I_A\otimes s)S(r)\bigr)\right]^2 \right)s∈CB×Bs⪰0, Retrs≤1​sup​L([RetrS(S(r)(IA​⊗s)S(r))]2)

in the extended reals, where S(M)S(M)S(M) is the total real continuous functional calculus for x↦xx\mapsto\sqrt{x}x↦x​, and L(x)=−∞L(x)=-\inftyL(x)=−∞ if x=0x=0x=0, while L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x, embedded into the extended reals, otherwise. No positivity or normalization assumption is imposed on rrr. The witness s=0s=0s=0 is permitted and contributes −∞-\infty−∞; empty index types are allowed, and the supremum may take an infinite value.

QAlgorithms.Papers.Independent.Intakenext100.P07.iidLaw

For any type XXX, any function p:X→Rp:X\to\mathbb Rp:X→R, and any natural number nnn, this defines the function on tuples x:{0,…,n−1}→Xx:\{0,\ldots,n-1\}\to Xx:{0,…,n−1}→X whose value is ∏i=0n−1p(x(i))\prod_{i=0}^{n-1}p(x(i))∏i=0n−1​p(x(i)). No finiteness, nonnegativity, or normalization assumption on XXX or ppp is required. At n=0n=0n=0, the unique empty tuple has value one.

QAlgorithms.Papers.Independent.Intakenext100.P07.introSmoothMax

For a finite type XXX, any function p:X→Rp:X\to\mathbb Rp:X→R, and any real eee, this defines the real infimum of 2log⁡2(∑x∈Xq(x))2\log_2(\sum_{x\in X}\sqrt{q(x)})2log2​(∑x∈X​q(x)​) over all functions q:X→Rq:X\to\mathbb Rq:X→R satisfying q(x)≥0q(x)\ge0q(x)≥0 for every xxx, ∑xq(x)=1\sum_xq(x)=1∑x​q(x)=1, and 12∑x∣p(x)−q(x)∣≤e\tfrac12\sum_x|p(x)-q(x)|\le e21​∑x​∣p(x)−q(x)∣≤e. There is no probability assumption on ppp and no restriction on eee. If no such qqq exists—including when XXX is empty or e<0e<0e<0—the real infimum of the empty set has the defined value zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.inverseChoi

For a finite index type III, an index type JJJ, any complex matrix GGG indexed by I×JI\times JI×J, and any complex matrix rrr indexed by III, this definition gives the matrix on JJJ with entries

(inverseChoi⁡(G,r))b,b′=∑i∈I∑j∈Iri,j G(i,b),(j,b′)(b,b′∈J).\bigl(\operatorname{inverseChoi}(G,r)\bigr)_{b,b'} =\sum_{i\in I}\sum_{j\in I}r_{i,j}\,G_{(i,b),(j,b')} \qquad(b,b'\in J).(inverseChoi(G,r))b,b′​=i∈I∑​j∈I∑​ri,j​G(i,b),(j,b′)​(b,b′∈J).

No conjugation of ri,jr_{i,j}ri,j​ occurs. If III is empty, the resulting matrix is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.matrixLog

For a finite index type III with decidable equality and any complex matrix MMM indexed by III, matrixLog(M)\mathrm{matrixLog}(M)matrixLog(M) applies the real continuous functional calculus to the total real function x↦log⁡2xx\mapsto\log_2 xx↦log2​x. For Hermitian MMM, an eigenvalue x≠0x\neq0x=0 is replaced by log⁡2∣x∣\log_2|x|log2​∣x∣, and a zero eigenvalue is replaced by zero; for non-Hermitian MMM, the result is the zero matrix. Empty index types are included.

QAlgorithms.Papers.Independent.Intakenext100.P07.maxD

For a finite index type III with decidable equality, including the empty type, and arbitrary complex I×II\times II×I matrices r,sr,sr,s, this defines an extended real number by

maxD⁡(r,s)=inf⁡{ℓ∈R:2ℓs−r is positive semidefinite},\operatorname{maxD}(r,s) = \inf\left\{\ell\in\mathbb R: 2^\ell s-r\text{ is positive semidefinite}\right\},maxD(r,s)=inf{ℓ∈R:2ℓs−r is positive semidefinite},

where the real numbers in the set are embedded into the extended reals. No positivity or normalization of rrr or sss is assumed. The infimum is +∞+\infty+∞ if the set is empty and −∞-\infty−∞ if it is unbounded below; for empty III, every real ℓ\ellℓ is admissible, so the value is −∞-\infty−∞.

QAlgorithms.Papers.Independent.Intakenext100.P07.maxRate

For finite index types A,BA,BA,B with decidable equality, any complex matrix rrr indexed by A×BA\times BA×B, any real eee, and any natural number nnn, put An={0,…,n−1}→AA_n=\{0,\ldots,n-1\}\to AAn​={0,…,n−1}→A, Bn={0,…,n−1}→BB_n=\{0,\ldots,n-1\}\to BBn​={0,…,n−1}→B, and R(xA,xB),(yA,yB)=∏i=0n−1r(xA(i),xB(i)),(yA(i),yB(i))R_{(x_A,x_B),(y_A,y_B)}=\prod_{i=0}^{n-1}r_{(x_A(i),x_B(i)),(y_A(i),y_B(i))}R(xA​,xB​),(yA​,yB​)​=∏i=0n−1​r(xA​(i),xB​(i)),(yA​(i),yB​(i))​. This defines the extended real number maxRate⁡(r,e,n)=(1n)inf⁡q∈Be(R)Hmax⁡(q)\displaystyle \operatorname{maxRate}(r,e,n)=\left(\frac1n\right)\inf_{q\in\mathcal B_e(R)}H_{\max}(q)maxRate(r,e,n)=(n1​)q∈Be​(R)inf​Hmax​(q), where Be(R)\mathcal B_e(R)Be​(R) consists of matrices q⪰0q\succeq0q⪰0 indexed by An×BnA_n\times B_nAn​×Bn​ satisfying Re⁡tr⁡q≤1\operatorname{Re}\operatorname{tr}q\le1Retrq≤1 and 1−(F0(q,R)+(1−Re⁡tr⁡q)(1−Re⁡tr⁡R))2≤e\displaystyle \sqrt{1-\left(F_0(q,R)+\sqrt{(1-\operatorname{Re}\operatorname{tr}q)(1-\operatorname{Re}\operatorname{tr}R)}\right)^2}\le e1−(F0​(q,R)+(1−Retrq)(1−RetrR)​)2​≤e. Here F0(M,N)=Re⁡tr⁡M N M\displaystyle F_0(M,N)=\operatorname{Re}\operatorname{tr}\sqrt{\sqrt M\,N\,\sqrt M}F0​(M,N)=RetrM​NM​​, using the real continuous functional calculus for matrix square roots, and Hmax⁡(q)=sup⁡σ⪰0, Re⁡tr⁡σ≤1L ⁣(F0(q,IAn⊗σ)2)\displaystyle H_{\max}(q)=\sup_{\sigma\succeq0,\ \operatorname{Re}\operatorname{tr}\sigma\le1}L\!\left(F_0(q,I_{A_n}\otimes\sigma)^2\right)Hmax​(q)=σ⪰0, Retrσ≤1sup​L(F0​(q,IAn​​⊗σ)2), where σ\sigmaσ is indexed by BnB_nBn​, L(0)=−∞L(0)=-\inftyL(0)=−∞, and L(u)=log⁡2uL(u)=\log_2uL(u)=log2​u for u≠0u\ne0u=0, regarded as an extended real. No positivity or normalization of rrr, restriction on eee, or positivity of nnn is assumed; an empty infimum is +∞+\infty+∞, and at n=0n=0n=0 the real reciprocal is 000, so the rate is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.measureOutput

For a finite index type III with decidable equality, any natural number mmm, any family of complex matrices MxM_xMx​ on III indexed by x∈Fin⁡(m)x\in\operatorname{Fin}(m)x∈Fin(m), and any complex matrix rrr on III, this definition produces the matrix on Fin⁡(m)\operatorname{Fin}(m)Fin(m) whose entries are

(measureOutput⁡(M,r))xy={tr⁡(Mxr)x=y,0x≠y.\bigl(\operatorname{measureOutput}(M,r)\bigr)_{xy} = \begin{cases} \operatorname{tr}(M_xr)&x=y,\\ 0&x\ne y. \end{cases}(measureOutput(M,r))xy​={tr(Mx​r)0​x=y,x=y.​

The diagonal uses the full complex trace, without taking its real part. There are no positivity, normalization, or measurement hypotheses on MMM or rrr; m=0m=0m=0 is allowed and gives the matrix with empty index set.

QAlgorithms.Papers.Independent.Intakenext100.P07.measuredRenyi

For a finite type III with decidable equality, arbitrary complex square matrices r,sr,sr,s indexed by III, and an order a∈[0,∞]a\in[0,\infty]a∈[0,∞], this defines

sup⁡m∈N, m>0M0,…,Mm−1⪰0, ∑iMi=1Ra ⁣(diag⁡(tr⁡(Mir))i<m,diag⁡(tr⁡(Mis))i<m),\sup_{\substack{m\in\mathbb N,\ m>0\\ M_0,\ldots,M_{m-1}\succeq0,\ \sum_iM_i=\mathbf1}} R_a\!\left(\operatorname{diag}(\operatorname{tr}(M_ir))_{i<m}, \operatorname{diag}(\operatorname{tr}(M_is))_{i<m}\right),m∈N, m>0M0​,…,Mm−1​⪰0, ∑i​Mi​=1​sup​Ra​(diag(tr(Mi​r))i<m​,diag(tr(Mi​s))i<m​),

with supremum in the extended reals. To specify RaR_aRa​ literally, let Pt(A)P_t(A)Pt​(A) be real functional calculus for the function equal to 000 at 000 and xtx^txt elsewhere, let L(A)L(A)L(A) be real functional calculus for the total real logarithm log⁡2x\log_2xlog2​x, with log⁡20=0\log_2 0=0log2​0=0, and put Qt(A,B)=Re⁡tr⁡(Pt(A)P1−t(B))Q_t(A,B)=\operatorname{Re}\operatorname{tr}(P_t(A)P_{1-t}(B))Qt​(A,B)=Retr(Pt​(A)P1−t​(B)). Write A≪BA\ll BA≪B for ker⁡B⊆ker⁡A\ker B\subseteq\ker AkerB⊆kerA, and let ℓ(0)=−∞\ell(0)=-\inftyℓ(0)=−∞, ℓ(x)=log⁡2x\ell(x)=\log_2xℓ(x)=log2​x otherwise. For t>0t>0t>0, define

Ft(A,B)={Re⁡tr⁡(A(L(A)−L(B)))Re⁡tr⁡A,t=1, A≪B,+∞,t=1, A≪̸B,1t−1 ℓ ⁣(Qt(A,B)Re⁡tr⁡A),t≠1, ((t<1 and Qt(A,B)>0) or A≪B),+∞,otherwise.F_t(A,B)= \begin{cases} \displaystyle \frac{\operatorname{Re}\operatorname{tr}(A(L(A)-L(B)))} {\operatorname{Re}\operatorname{tr}A}, &t=1,\ A\ll B,\\ +\infty,&t=1,\ A\not\ll B,\\ \displaystyle\frac1{t-1}\, \ell\!\left(\frac{Q_t(A,B)}{\operatorname{Re}\operatorname{tr}A}\right), &t\ne1,\ \bigl((t<1\text{ and }Q_t(A,B)>0)\text{ or }A\ll B\bigr),\\ +\infty,&\text{otherwise}. \end{cases}Ft​(A,B)=⎩⎨⎧​RetrARetr(A(L(A)−L(B)))​,+∞,t−11​ℓ(RetrAQt​(A,B)​),+∞,​t=1, A≪B,t=1, A≪B,t=1, ((t<1 and Qt​(A,B)>0) or A≪B),otherwise.​

Then Ra=FaR_a=F_aRa​=Fa​ for finite positive aaa, R0=inf⁡t>0FtR_0=\inf_{t>0}F_tR0​=inft>0​Ft​, and R∞=sup⁡t>0FtR_\infty=\sup_{t>0}F_tR∞​=supt>0​Ft​. These formulas use total functional calculus and total real division, including division by zero, and extended real arithmetic after applying ℓ\ellℓ. In particular, the measurement outputs above retain the full complex traces as diagonal entries; no positivity or normalization of r,sr,sr,s is assumed. The index type III may be empty, and an empty set of attained values has extended real supremum −∞-\infty−∞.

QAlgorithms.Papers.Independent.Intakenext100.P07.minRate

For finite index types A,BA,BA,B with decidable equality, any complex matrix rrr indexed by A×BA\times BA×B, any real eee, and any natural number nnn, put An={0,…,n−1}→AA_n=\{0,\ldots,n-1\}\to AAn​={0,…,n−1}→A, Bn={0,…,n−1}→BB_n=\{0,\ldots,n-1\}\to BBn​={0,…,n−1}→B, and R(xA,xB),(yA,yB)=∏i=0n−1r(xA(i),xB(i)),(yA(i),yB(i))R_{(x_A,x_B),(y_A,y_B)}=\prod_{i=0}^{n-1}r_{(x_A(i),x_B(i)),(y_A(i),y_B(i))}R(xA​,xB​),(yA​,yB​)​=∏i=0n−1​r(xA​(i),xB​(i)),(yA​(i),yB​(i))​. This defines the extended real number minRate⁡(r,e,n)=(1n)sup⁡q∈Be(R)Hmin⁡(q)\displaystyle \operatorname{minRate}(r,e,n)=\left(\frac1n\right)\sup_{q\in\mathcal B_e(R)}H_{\min}(q)minRate(r,e,n)=(n1​)q∈Be​(R)sup​Hmin​(q), where Be(R)\mathcal B_e(R)Be​(R) consists of matrices qqq indexed by An×BnA_n\times B_nAn​×Bn​ such that qqq is positive semidefinite, Re⁡tr⁡q≤1\operatorname{Re}\operatorname{tr}q\le1Retrq≤1, and 1−(F0(q,R)+(1−Re⁡tr⁡q)(1−Re⁡tr⁡R))2≤e\displaystyle \sqrt{1-\left(F_0(q,R)+\sqrt{(1-\operatorname{Re}\operatorname{tr}q)(1-\operatorname{Re}\operatorname{tr}R)}\right)^2}\le e1−(F0​(q,R)+(1−Retrq)(1−RetrR)​)2​≤e; here F0(q,R)=Re⁡tr⁡q R q\displaystyle F_0(q,R)=\operatorname{Re}\operatorname{tr}\sqrt{\sqrt q\,R\,\sqrt q}F0​(q,R)=Retrq​Rq​​, with matrix square roots taken through the real continuous functional calculus. For such qqq, Hmin⁡(q)=sup⁡{ℓ∈R:∃σ, σ⪰0, Re⁡tr⁡σ≤1, 2−ℓ(IAn⊗σ)−q⪰0}\displaystyle H_{\min}(q)=\sup\{\ell\in\mathbb R:\exists \sigma,\ \sigma\succeq0,\ \operatorname{Re}\operatorname{tr}\sigma\le1,\ 2^{-\ell}(I_{A_n}\otimes\sigma)-q\succeq0\}Hmin​(q)=sup{ℓ∈R:∃σ, σ⪰0, Retrσ≤1, 2−ℓ(IAn​​⊗σ)−q⪰0}, where σ\sigmaσ is indexed by BnB_nBn​ and the supremum is taken in the extended reals. There are no positivity or normalization assumptions on rrr, no restriction on eee, and no requirement that n>0n>0n>0; an empty supremum is −∞-\infty−∞, and at n=0n=0n=0 the real reciprocal is defined as 000, so the rate is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.optimalCardinality

For a finite type XXX with decidable equality, an arbitrary real-valued function p:X→Rp:X\to\mathbb Rp:X→R, and an arbitrary real number eee, this defines the natural number

inf⁡{m∈N:m>0 and there exist f:X→{0,…,m−1}, g:{0,…,m−1}→X with ∑x∈Xg(f(x))≠xp(x)≤e}.\inf\left\{m\in\mathbb N: m>0\ \text{and there exist }f:X\to\{0,\ldots,m-1\},\ g:\{0,\ldots,m-1\}\to X \text{ with } \sum_{\substack{x\in X\\g(f(x))\ne x}}p(x)\le e \right\}.inf⎩⎨⎧​m∈N:m>0 and there exist f:X→{0,…,m−1}, g:{0,…,m−1}→X with x∈Xg(f(x))=x​∑​p(x)≤e⎭⎬⎫​.

Thus, when the set is nonempty, the value is its least element; when it is empty, the natural-number infimum is 000. The definition does not require ppp to be nonnegative or to sum to 111, does not restrict eee, and does not require either coding function to be injective or surjective.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperAdjointAction

For finite index types I,JI,JI,J, with decidable equality on III, any complex-linear map E:CI×I→CJ×JE:\mathbb C^{I\times I}\to\mathbb C^{J\times J}E:CI×I→CJ×J, and any complex matrix YYY indexed by JJJ, this definition gives the matrix on III with entries

(E♯(Y))i,j=∑a∈J∑b∈J(E(eij))a,b‾ Ya,b(i,j∈I),\bigl(E^\sharp(Y)\bigr)_{i,j} =\sum_{a\in J}\sum_{b\in J} \overline{\bigl(E(e_{ij})\bigr)_{a,b}}\,Y_{a,b} \qquad(i,j\in I),(E♯(Y))i,j​=a∈J∑​b∈J∑​(E(eij​))a,b​​Ya,b​(i,j∈I),

where eije_{ij}eij​ has entry 111 at (i,j)(i,j)(i,j) and 000 elsewhere. If JJJ is empty, the resulting matrix is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperBall

For any finite index type III with decidable equality, arbitrary matrix r∈CI×Ir\in\mathbb C^{I\times I}r∈CI×I, and real number eee, this defines the set of all matrices s∈CI×Is\in\mathbb C^{I\times I}s∈CI×I such that sss is positive semidefinite, Re⁡tr⁡s≤1\operatorname{Re}\operatorname{tr}s\le1Retrs≤1, and

1−(Re⁡tr⁡S(S(s) r S(s))+(1−Re⁡tr⁡s)(1−Re⁡tr⁡r))2≤e,\sqrt{ 1-\left( \operatorname{Re}\operatorname{tr} S\bigl(S(s)\,r\,S(s)\bigr) +\sqrt{\bigl(1-\operatorname{Re}\operatorname{tr}s\bigr) \bigl(1-\operatorname{Re}\operatorname{tr}r\bigr)} \right)^2 }\le e,1−(RetrS(S(s)rS(s))+(1−Retrs)(1−Retrr)​)2​≤e,

where S(M)S(M)S(M) denotes the total real continuous functional calculus for x↦xx\mapsto\sqrt{x}x↦x​, and scalar square roots of negative numbers are 000. No positivity or normalization condition is imposed on the center rrr, and no restriction on eee is assumed. The set is empty when e<0e<0e<0; empty index types are allowed.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperBasisMeasurement

For finite index types A,BA,BA,B with decidable equality, any complex square matrix UUU indexed by AAA, and any complex square matrix rrr indexed by A×BA\times BA×B, put W=U⊗1BW=U\otimes1_BW=U⊗1B​, so W(a,b),(a′,b′)=Ua,a′δb,b′W_{(a,b),(a',b')}=U_{a,a'}\delta_{b,b'}W(a,b),(a′,b′)​=Ua,a′​δb,b′​. This defines the matrix RRR by

R(a,b),(a′,b′)={(W†rW)(a,b),(a′,b′),a=a′,0,a≠a′.R_{(a,b),(a',b')}= \begin{cases} (W^\dagger rW)_{(a,b),(a',b')},&a=a',\\ 0,&a\ne a'. \end{cases}R(a,b),(a′,b′)​={(W†rW)(a,b),(a′,b′)​,0,​a=a′,a=a′.​

No unitarity of UUU or positivity or trace condition on rrr is required. If either index type is empty, the result is the unique empty matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperBelow

For a finite index type III and complex matrices M,NM,NM,N indexed by III, paperBelow(M,N)\mathrm{paperBelow}(M,N)paperBelow(M,N) means that N−MN-MN−M is positive semidefinite. Empty index types are included.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperBetaStar

For any finite index type III with decidable equality, any complex matrices r,sr,sr,s indexed by I×II\times II×I, and any real eee, this defines

β∗(r,s,e)=inf⁡{Re⁡tr⁡(Ts):T⪰0, 1−T⪰0, Re⁡tr⁡((1−T)r)≤e}.\beta_*(r,s,e)= \inf\left\{ \operatorname{Re}\operatorname{tr}(Ts): T\succeq0,\ \mathbf 1-T\succeq0,\ \operatorname{Re}\operatorname{tr}((\mathbf 1-T)r)\le e \right\}.β∗​(r,s,e)=inf{Retr(Ts):T⪰0, 1−T⪰0, Retr((1−T)r)≤e}.

Here the infimum is a real infimum, assigned value 000 when the feasible set is empty or the objective values are unbounded below. Neither rrr nor sss is required to be positive semidefinite or normalized, and eee may be any real number. Empty III is allowed and gives value 000, including when e<0e<0e<0 makes the feasible set empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperCPTP

For finite index types I,JI,JI,J and a complex linear map EEE from complex matrices indexed by III to complex matrices indexed by JJJ, this predicate means that, for every natural number kkk, including 000, and every positive semidefinite matrix XXX indexed by {0,…,k−1}×I\{0,\ldots,k-1\}\times I{0,…,k−1}×I, the matrix Y(u,j),(v,j′)=[E ⁣((X(u,i),(v,i′))i,i′∈I)]j,j′\displaystyle Y_{(u,j),(v,j')}=\left[E\!\left((X_{(u,i),(v,i')})_{i,i'\in I}\right)\right]_{j,j'}Y(u,j),(v,j′)​=[E((X(u,i),(v,i′)​)i,i′∈I​)]j,j′​ is positive semidefinite, and that tr⁡(E(M))=tr⁡M\operatorname{tr}(E(M))=\operatorname{tr}Mtr(E(M))=trM for every complex matrix MMM indexed by III. This is equality of complex traces for all inputs, with no positivity or normalization restriction on MMM; empty index types are included.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperChoi

For finite index types I,JI,JI,J, with decidable equality on III, and any complex-linear map E:CI×I→CJ×JE:\mathbb C^{I\times I}\to\mathbb C^{J\times J}E:CI×I→CJ×J, this definition gives the complex matrix CEC_ECE​ indexed by I×JI\times JI×J whose entries are

(CE)(i,a),(j,b)=(E(eij))a,b,(C_E)_{(i,a),(j,b)}=\bigl(E(e_{ij})\bigr)_{a,b},(CE​)(i,a),(j,b)​=(E(eij​))a,b​,

for every i,j∈Ii,j\in Ii,j∈I and a,b∈Ja,b\in Ja,b∈J, where eije_{ij}eij​ is the matrix with entry 111 at (i,j)(i,j)(i,j) and 000 elsewhere. No normalization factor or positivity assumption is included; either index type may be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperClassicalAgreement

For an assignment DDD taking pairs of complex square matrices on each finite type with decidable equality to extended real numbers, and an order a∈[0,∞]a\in[0,\infty]a∈[0,∞], this property requires the following for every nonempty finite type III with decidable equality and every pair of functions p,q:I→Rp,q:I\to\mathbb Rp,q:I→R: if pi,qi≥0p_i,q_i\ge0pi​,qi​≥0 for all iii and ppp is not identically zero, then DI(diag⁡p,diag⁡q)=Ra(p,q)D_I(\operatorname{diag}p,\operatorname{diag}q)=R_a(p,q)DI​(diagp,diagq)=Ra​(p,q), where the extended real quantity RaR_aRa​ is specified as follows. Set zt(x)=0z_t(x)=0zt​(x)=0 if x=0x=0x=0, and zt(x)=xtz_t(x)=x^tzt​(x)=xt otherwise; put Qt=∑izt(pi)z1−t(qi)Q_t=\sum_i z_t(p_i)z_{1-t}(q_i)Qt​=∑i​zt​(pi​)z1−t​(qi​), and write p≪qp\ll qp≪q for the condition qi=0⇒pi=0q_i=0\Rightarrow p_i=0qi​=0⇒pi​=0 for every iii. Let ℓ(0)=−∞\ell(0)=-\inftyℓ(0)=−∞ and ℓ(x)=log⁡2x\ell(x)=\log_2xℓ(x)=log2​x otherwise. For real t>0t>0t>0, set

Ft={∑ipi(log⁡2pi−log⁡2qi)∑ipi,t=1 and p≪q,+∞,t=1 and p≪̸q,1t−1 ℓ ⁣(Qt∑ipi),t≠1 and ((t<1 and Qt>0) or p≪q),+∞,otherwise.F_t= \begin{cases} \displaystyle\frac{\sum_i p_i(\log_2p_i-\log_2q_i)}{\sum_i p_i}, &t=1\text{ and }p\ll q,\\ +\infty,&t=1\text{ and }p\not\ll q,\\ \displaystyle\frac1{t-1}\,\ell\!\left(\frac{Q_t}{\sum_i p_i}\right), &t\ne1\text{ and }\bigl((t<1\text{ and }Q_t>0)\text{ or }p\ll q\bigr),\\ +\infty,&\text{otherwise}. \end{cases}Ft​=⎩⎨⎧​∑i​pi​∑i​pi​(log2​pi​−log2​qi​)​,+∞,t−11​ℓ(∑i​pi​Qt​​),+∞,​t=1 and p≪q,t=1 and p≪q,t=1 and ((t<1 and Qt​>0) or p≪q),otherwise.​

Here the real logarithm at 000 in the first line is 000, and the third line uses extended real multiplication. For 0<a<∞0<a<\infty0<a<∞, Ra=FaR_a=F_aRa​=Fa​; R0=inf⁡t>0FtR_0=\inf_{t>0}F_tR0​=inft>0​Ft​; and R∞=sup⁡t>0FtR_\infty=\sup_{t>0}F_tR∞​=supt>0​Ft​. Neither ppp nor qqq is required to sum to 111, and qqq may be identically zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperClassicalRenyi

For a finite type XXX, any function p:X→Rp:X\to\mathbb Rp:X→R, and any real aaa, this defines log⁡2(∑x∈Xp(x)a)/(1−a)\log_2(\sum_{x\in X}p(x)^a)/(1-a)log2​(∑x∈X​p(x)a)/(1−a), using the total real power, logarithm, and division operations. No probability or positivity assumption is imposed on ppp, and no order is excluded. In particular, the value at a=1a=1a=1 is zero because division by zero is defined as zero; an empty XXX also gives zero because log⁡20=0\log_2 0=0log2​0=0.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperCodeError

For a finite type XXX, any real-valued function ppp on XXX, any natural number mmm, and encoding and decoding functions E:X→{0,…,m−1}E:X\to\{0,\ldots,m-1\}E:X→{0,…,m−1} and D:{0,…,m−1}→XD:\{0,\ldots,m-1\}\to XD:{0,…,m−1}→X, this defines the real number ∑x∈X: D(E(x))≠xp(x)\sum_{x\in X:\,D(E(x))\ne x}p(x)∑x∈X:D(E(x))=x​p(x). Neither nonnegativity nor normalization of ppp is assumed, and reconstruction need not succeed anywhere. The sum is zero when XXX is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperConditional

Let A,BA,BA,B be finite types with decidable equality, possibly empty, rrr any complex matrix on A×BA\times BA×B, b,ub,ub,u Booleans, and α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞]. Write (rB)j,k=∑i∈Ar(i,j),(i,k)(r_B)_{j,k}=\sum_{i\in A}r_{(i,j),(i,k)}(rB​)j,k​=∑i∈A​r(i,j),(i,k)​. For real ppp, let Pp(M)P_p(M)Pp​(M) be real functional calculus for the function equal to 000 at 000 and xpx^pxp elsewhere. Put Qb(v,w;t)=Re⁡tr⁡Pt(P(1−t)/(2t)(w) v P(1−t)/(2t)(w))Q_b(v,w;t)=\operatorname{Re}\operatorname{tr}P_t(P_{(1-t)/(2t)}(w)\,v\,P_{(1-t)/(2t)}(w))Qb​(v,w;t)=RetrPt​(P(1−t)/(2t)​(w)vP(1−t)/(2t)​(w)) for true bbb, and Qb(v,w;t)=Re⁡tr⁡(Pt(v)P1−t(w))Q_b(v,w;t)=\operatorname{Re}\operatorname{tr}(P_t(v)P_{1-t}(w))Qb​(v,w;t)=Retr(Pt​(v)P1−t​(w)) for false bbb. Let S(v,w)S(v,w)S(v,w) mean ker⁡w⊆ker⁡v\ker w\subseteq\ker vkerw⊆kerv, let L(0)=−∞L(0)=-\inftyL(0)=−∞ and L(x)=log⁡2xL(x)=\log_2xL(x)=log2​x otherwise, and use real functional calculus for matrix logarithms. Define db(v,w;1)=Re⁡tr⁡[v(log⁡2v−log⁡2w)]/Re⁡tr⁡vd_b(v,w;1)=\operatorname{Re}\operatorname{tr}[v(\log_2v-\log_2w)]/\operatorname{Re}\operatorname{tr}vdb​(v,w;1)=Retr[v(log2​v−log2​w)]/Retrv when S(v,w)S(v,w)S(v,w), and +∞+\infty+∞ otherwise; for t≠1t\ne1t=1, define db(v,w;t)=1t−1L(Qb(v,w;t)/Re⁡tr⁡v)d_b(v,w;t)=\frac1{t-1}L(Q_b(v,w;t)/\operatorname{Re}\operatorname{tr}v)db​(v,w;t)=t−11​L(Qb​(v,w;t)/Retrv) when (t<1 and Qb(v,w;t)>0)(t<1\text{ and }Q_b(v,w;t)>0)(t<1 and Qb​(v,w;t)>0) or S(v,w)S(v,w)S(v,w), and +∞+\infty+∞ otherwise. Define Rb(v,w;0)=inf⁡t>0db(v,w;t)\mathcal R_b(v,w;0)=\inf_{t>0}d_b(v,w;t)Rb​(v,w;0)=inft>0​db​(v,w;t), Rtrue(v,w;+∞)=inf⁡{ℓ∈R:2ℓw−v⪰0}\mathcal R_{\mathrm{true}}(v,w;+\infty)=\inf\{\ell\in\mathbb R:2^\ell w-v\succeq0\}Rtrue​(v,w;+∞)=inf{ℓ∈R:2ℓw−v⪰0}, Rfalse(v,w;+∞)=sup⁡t>0db(v,w;t)\mathcal R_{\mathrm{false}}(v,w;+\infty)=\sup_{t>0}d_b(v,w;t)Rfalse​(v,w;+∞)=supt>0​db​(v,w;t), and Rb(v,w;α)=db(v,w;α)\mathcal R_b(v,w;\alpha)=d_b(v,w;\alpha)Rb​(v,w;α)=db​(v,w;α) for finite positive α\alphaα. The declaration defines

{sup⁡s on Bs⪰0, tr⁡s=1[−Rb(r,IA⊗s;α)],u=true,−Rb(r,IA⊗rB;α),u=false.\begin{cases} \displaystyle\sup_{\substack{s\text{ on }B\\s\succeq0,\ \operatorname{tr}s=1}} [-\mathcal R_b(r,I_A\otimes s;\alpha)],&u=\mathrm{true},\\ -\mathcal R_b(r,I_A\otimes r_B;\alpha),&u=\mathrm{false}. \end{cases}⎩⎨⎧​s on Bs⪰0, trs=1​sup​[−Rb​(r,IA​⊗s;α)],−Rb​(r,IA​⊗rB​;α),​u=true,u=false.​

All extrema and entropy arithmetic are extended-real, so the first branch is −∞-\infty−∞ when its feasible set is empty. There are no positivity or normalization assumptions on rrr; real division by zero equals 000, and real powers and logarithms are total.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperCqState

For index types X,BX,BX,B, with equality on XXX decidable, and an arbitrary family rx∈CB×Br_x\in\mathbb C^{B\times B}rx​∈CB×B, the associated complex matrix on X×BX\times BX×B is defined by

R(x,b),(y,c)={(rx)b,c,x=y,0,x≠y.R_{(x,b),(y,c)}= \begin{cases} (r_x)_{b,c},&x=y,\\ 0,&x\ne y. \end{cases}R(x,b),(y,c)​={(rx​)b,c​,0,​x=y,x=y.​

Thus its blocks indexed by distinct elements of XXX vanish, and its block indexed by xxx is rxr_xrx​. No positivity, trace, or normalization condition is imposed.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperDataProcessing

For an assignment DDD taking pairs of complex square matrices on finite types with decidable equality to extended real numbers, this property requires that for every pair of nonempty finite types I,JI,JI,J with decidable equality and every complex linear map EEE from III-indexed matrices to JJJ-indexed matrices, if EEE preserves the complex trace of every matrix and is completely positive in the explicit sense that, for every natural number kkk, including 000, and every positive semidefinite matrix RRR indexed by {0,…,k−1}×I\{0,\ldots,k-1\}\times I{0,…,k−1}×I, the matrix indexed by {0,…,k−1}×J\{0,\ldots,k-1\}\times J{0,…,k−1}×J with entries

[(id⁡k⊗E)(R)](a,b),(a′,b′)=[E ⁣((R(a,i),(a′,i′))i,i′∈I)]b,b′\bigl[(\operatorname{id}_k\otimes E)(R)\bigr]_{(a,b),(a',b')} = \left[E\!\left((R_{(a,i),(a',i')})_{i,i'\in I}\right)\right]_{b,b'}[(idk​⊗E)(R)](a,b),(a′,b′)​=[E((R(a,i),(a′,i′)​)i,i′∈I​)]b,b′​

is positive semidefinite, then every pair of positive semidefinite III-indexed matrices r,sr,sr,s with r≠0r\ne0r=0 satisfies

DJ(Er,Es)≤DI(r,s).D_J(Er,Es)\le D_I(r,s).DJ​(Er,Es)≤DI​(r,s).

No trace normalization or support relation is required, and s=0s=0s=0 is included.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperDelta

For a finite index type III with decidable equality and any complex matrices M,NM,NM,N indexed by III, paperDelta(M,N)\mathrm{paperDelta}(M,N)paperDelta(M,N) is the real supremum sup⁡{∣tr⁡((M−N)E)∣:E and 1−E are positive semidefinite}\sup\{|\operatorname{tr}((M-N)E)|: E\text{ and }\mathbf1-E\text{ are positive semidefinite}\}sup{∣tr((M−N)E)∣:E and 1−E are positive semidefinite}, where the absolute value is the complex modulus and 1\mathbf11 is the identity matrix. Empty index types are included, in which case the value is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperDirectSum

For index types I,JI,JI,J and square complex matrices rrr on III and sss on JJJ, this defines the square matrix r⊕sr\oplus sr⊕s on the disjoint union I⊔JI\sqcup JI⊔J: its I×II\times II×I block is rrr, its J×JJ\times JJ×J block is sss, and both mixed blocks are zero. No finiteness, nonemptiness, positivity, or normalization condition is imposed; either index type may be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperDivergenceFamily

This structure consists of one assignment DDD that, for every type III equipped with a finite enumeration and decidable equality, takes two complex square matrices indexed by III and returns an extended real number in R∪{−∞,+∞}\mathbb R\cup\{-\infty,+\infty\}R∪{−∞,+∞}. Empty index types are included. The structure imposes no positivity, normalization, continuity, or other condition on either the matrices or the assignment.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperEntangled

For every finite index type III with decidable equality, this definition gives a complex matrix ΩI\Omega_IΩI​ indexed by I×II\times II×I, with entries

(ΩI)(i,j),(k,l)={∣I∣−1,i=j and k=l,0,otherwise.(\Omega_I)_{(i,j),(k,l)} = \begin{cases} |I|^{-1},&i=j\text{ and }k=l,\\ 0,&\text{otherwise}. \end{cases}(ΩI​)(i,j),(k,l)​={∣I∣−1,0,​i=j and k=l,otherwise.​

Here the reciprocal is taken in C\mathbb CC. When III is nonempty, this is ∣I∣−1∑i,k∈Ieik⊗eik|I|^{-1}\sum_{i,k\in I}e_{ik}\otimes e_{ik}∣I∣−1∑i,k∈I​eik​⊗eik​. The empty index type is also permitted: its reciprocal scalar is defined as 0−1=00^{-1}=00−1=0, and the resulting matrix has no entries.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperFidelity

For any finite index type III with decidable equality and arbitrary complex matrices r,s∈CI×Ir,s\in\mathbb C^{I\times I}r,s∈CI×I, this defines the real number

(Re⁡tr⁡S(S(r) s S(r)))2,\left(\operatorname{Re}\operatorname{tr} S\bigl(S(r)\,s\,S(r)\bigr)\right)^2,(RetrS(S(r)sS(r)))2,

where S(M)S(M)S(M) is the total real continuous functional calculus of MMM for the function x↦xx\mapsto\sqrt{x}x↦x​. No positivity, Hermitian, trace, or normalization hypothesis is imposed on either matrix. Scalar real square roots take value 000 on negative inputs, and the empty index type is allowed.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperGeneralizedFidelity

For any finite index type III with decidable equality and arbitrary complex matrices r,s∈CI×Ir,s\in\mathbb C^{I\times I}r,s∈CI×I, this defines

(Re⁡tr⁡S(S(r) s S(r))+(1−Re⁡tr⁡r)(1−Re⁡tr⁡s))2,\left( \operatorname{Re}\operatorname{tr} S\bigl(S(r)\,s\,S(r)\bigr) +\sqrt{\bigl(1-\operatorname{Re}\operatorname{tr}r\bigr) \bigl(1-\operatorname{Re}\operatorname{tr}s\bigr)} \right)^2,(RetrS(S(r)sS(r))+(1−Retrr)(1−Retrs)​)2,

where S(M)S(M)S(M) denotes the total real continuous functional calculus for x↦xx\mapsto\sqrt{x}x↦x​. There are no positivity or trace restrictions on r,sr,sr,s; in particular, the product inside the scalar square root may be negative, in which case that square root is 000. The index type may be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperHmin

For finite index types A,BA,BA,B, with decidable equality on AAA for its identity matrix, and any complex matrix rrr indexed by A×BA\times BA×B, this defines an extended real number as

sup⁡{ℓ∈R:∃s∈CB×B, s⪰0,Re⁡tr⁡s≤1,2−ℓ(IA⊗s)−r⪰0},\sup\left\{ \ell\in\mathbb R: \exists s\in\mathbb C^{B\times B},\ s\succeq0,\quad \operatorname{Re}\operatorname{tr}s\le1,\quad 2^{-\ell}(I_A\otimes s)-r\succeq0 \right\},sup{ℓ∈R:∃s∈CB×B, s⪰0,Retrs≤1,2−ℓ(IA​⊗s)−r⪰0},

where each real ℓ\ellℓ is embedded into the extended reals and ⪰0\succeq0⪰0 means positive semidefinite. The witness sss may be zero, and rrr has no positivity or normalization hypothesis. The supremum is −∞-\infty−∞ if there are no witnesses and may be +∞+\infty+∞; in particular, r=0r=0r=0 makes every real ℓ\ellℓ feasible. Empty index types are allowed.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperIndependentProduct

For arbitrary index types A,B,C,DA,B,C,DA,B,C,D, a complex matrix rrr indexed by A×BA\times BA×B, and a complex matrix sss indexed by C×DC\times DC×D, this defines a complex matrix TTT indexed by (A×C)×(B×D)(A\times C)\times(B\times D)(A×C)×(B×D) through

T((a,c),(b,d)), ((a′,c′),(b′,d′))=r(a,b),(a′,b′) s(c,d),(c′,d′).T_{((a,c),(b,d)),\,((a',c'),(b',d'))} =r_{(a,b),(a',b')}\,s_{(c,d),(c',d')}.T((a,c),(b,d)),((a′,c′),(b′,d′))​=r(a,b),(a′,b′)​s(c,d),(c′,d′)​.

There are no positivity, trace, or nonemptiness requirements on these inputs.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperJointPower

For index types A,BA,BA,B, any complex matrix rrr indexed by A×BA\times BA×B, and any natural number nnn, this defines a matrix indexed by pairs of functions (xA,xB)(x_A,x_B)(xA​,xB​), where xA:{0,…,n−1}→Ax_A:\{0,\ldots,n-1\}\to AxA​:{0,…,n−1}→A and xB:{0,…,n−1}→Bx_B:\{0,\ldots,n-1\}\to BxB​:{0,…,n−1}→B, whose entries are (r[n])(xA,xB),(yA,yB)=∏i=0n−1r(xA(i),xB(i)),(yA(i),yB(i))\displaystyle (r^{[n]})_{(x_A,x_B),(y_A,y_B)}=\prod_{i=0}^{n-1}r_{(x_A(i),x_B(i)),(y_A(i),y_B(i))}(r[n])(xA​,xB​),(yA​,yB​)​=i=0∏n−1​r(xA​(i),xB​(i)),(yA​(i),yB​(i))​. No conditions on rrr are imposed. The case n=0n=0n=0 is included: both function spaces are singletons, and the resulting matrix has its sole entry equal to the empty product 111, even if AAA or BBB is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperLocalEmbedding

For finite input index types A,BA,BA,B, output index types C,DC,DC,D, complex matrices UUU with rows indexed by CCC and columns by AAA, VVV with rows indexed by DDD and columns by BBB, and any complex matrix rrr indexed by A×BA\times BA×B, this defines the matrix (U⊗V) r (U⊗V)†(U\otimes V)\,r\,(U\otimes V)^\dagger(U⊗V)r(U⊗V)†, indexed by C×DC\times DC×D, where (U⊗V)(c,d),(a,b)=Uc,aVd,b(U\otimes V)_{(c,d),(a,b)}=U_{c,a}V_{d,b}(U⊗V)(c,d),(a,b)​=Uc,a​Vd,b​ and †\dagger† denotes conjugate transpose. Neither UUU nor VVV is required to be an isometry, and no positivity or normalization condition is imposed on rrr.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperLogE

For every real number xxx, paperLogE(x)\mathrm{paperLogE}(x)paperLogE(x) is an extended real number: it equals −∞-\infty−∞ when x=0x=0x=0, and otherwise equals the finite real value log⁡2x\log_2 xlog2​x under the total real logarithm convention. In particular, for negative xxx, this finite value is log⁡2∣x∣\log_2|x|log2​∣x∣.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperMarginalBC

For a finite index type AAA, arbitrary index types B,CB,CB,C, and any complex matrix rrr indexed by (A×B)×C(A\times B)\times C(A×B)×C, this defines a matrix indexed by B×CB\times CB×C with entries

(paperMarginalBC⁡(r))(b,c),(b′,c′)=∑a∈Ar((a,b),c),((a,b′),c′).\bigl(\operatorname{paperMarginalBC}(r)\bigr)_{(b,c),(b',c')} =\sum_{a\in A}r_{((a,b),c),((a,b'),c')}.(paperMarginalBC(r))(b,c),(b′,c′)​=a∈A∑​r((a,b),c),((a,b′),c′)​.

No positivity, Hermitian, or normalization assumption is imposed. If AAA is empty, the resulting matrix is zero; empty BBB or CCC is also allowed.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperMat

For any index type III, paperMat(I)\mathrm{paperMat}(I)paperMat(I) denotes the type of complex matrices whose rows and columns are both indexed by III, equivalently functions I×I→CI\times I\to\mathbb CI×I→C. The index type may be empty or infinite.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperMaxFamily

This defines an extended real-valued assignment on pairs of complex square matrices over every finite type III with decidable equality by

DI(r,s)=inf⁡{l∈R:2ls−r is positive semidefinite},D_I(r,s)= \inf\left\{l\in\mathbb R:2^l s-r\text{ is positive semidefinite}\right\},DI​(r,s)=inf{l∈R:2ls−r is positive semidefinite},

where the infimum is taken in the extended reals after embedding each real lll. An empty set therefore gives +∞+\infty+∞, and a set unbounded below gives −∞-\infty−∞. The definition includes empty index types and imposes no positivity or normalization assumptions on rrr or sss.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperMeasuredJoint

For an index type XXX with decidable equality, a finite index type BBB, and arbitrary families rx,Mx∈CB×Br_x,M_x\in\mathbb C^{B\times B}rx​,Mx​∈CB×B, the measured joint matrix is the complex matrix indexed by X×XX\times XX×X whose diagonal entry at (x,y)(x,y)(x,y) is the real number Re⁡tr⁡(Myrx)\operatorname{Re}\operatorname{tr}(M_yr_x)Retr(My​rx​), embedded in C\mathbb CC, and whose off-diagonal entries are zero. Neither positivity nor normalization is assumed, so the definition does not require these diagonal entries to form a probability distribution.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperMix

For any index type III, any real number ttt, and any two complex matrices M,NM,NM,N indexed by III, this defines the matrix tM+(1−t)NtM+(1-t)NtM+(1−t)N, with the real coefficients regarded as complex scalars. There is no requirement that 0≤t≤10\le t\le10≤t≤1, and neither matrix is required to be positive semidefinite or normalized.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperNsP

For a finite index type III, any real function v:I→Rv:I\to\mathbb Rv:I→R, and arbitrary complex square matrices U,VU,VU,V indexed by III, this defines the real function on I×II\times II×I given by

P(i,j)=vi ∣(U†V)ij∣2.P(i,j)=v_i\,\bigl|(U^\dagger V)_{ij}\bigr|^2.P(i,j)=vi​​(U†V)ij​​2.

Here †\dagger† denotes conjugate transpose and absolute value denotes complex modulus. Neither nonnegativity or normalization of vvv nor unitarity of U,VU,VU,V is required. If III is empty, this is the unique function on the empty domain.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperNsQ

For a finite index type III, any real function v:I→Rv:I\to\mathbb Rv:I→R, and arbitrary complex square matrices U,VU,VU,V indexed by III, this defines the real function on I×II\times II×I given by

Q(i,j)=vj ∣(U†V)ij∣2.Q(i,j)=v_j\,\bigl|(U^\dagger V)_{ij}\bigr|^2.Q(i,j)=vj​​(U†V)ij​​2.

Here †\dagger† denotes conjugate transpose and absolute value denotes complex modulus. Neither nonnegativity or normalization of vvv nor unitarity of U,VU,VU,V is required. If III is empty, this is the unique function on the empty domain.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperPenalty

For every real number eee, this defines the real number

paperPenalty⁡(e)=−log⁡2 ⁣(1−1−e2),\operatorname{paperPenalty}(e)=-\log_2\!\left(1-\sqrt{1-e^2}\right),paperPenalty(e)=−log2​(1−1−e2​),

using the total real square root and logarithm: the square root of a negative real number is zero, and log⁡20=0\log_2 0=0log2​0=0. There is no restriction on eee; the value is zero when e=0e=0e=0 and also whenever ∣e∣≥1|e|\ge 1∣e∣≥1.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperPinch

For a finite index type III with decidable equality and arbitrary complex square matrices H,LH,LH,L indexed by III, this defines

∑l∈S(H)Pl(H) L Pl(H),\sum_{l\in S(H)} P_l(H)\,L\,P_l(H),l∈S(H)∑​Pl​(H)LPl​(H),

where S(H)S(H)S(H) is the set of distinct eigenvalues of HHH when HHH is Hermitian, is {0}\{0\}{0} when HHH is non-Hermitian and III is nonempty, and is empty when III is empty; Pl(H)P_l(H)Pl​(H) is the real continuous functional calculus of HHH for the function equal to 111 at lll and 000 elsewhere. For Hermitian HHH, these are its eigenspace projections; for non-Hermitian HHH, they are zero, so the result is zero. An empty index type also gives the empty sum, hence zero. No positivity or trace requirement is imposed on LLL.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperPositivePart

For a finite index type III with decidable equality, including the empty type, and every complex I×II\times II×I matrix MMM, this defines paperPositivePart⁡(M)\operatorname{paperPositivePart}(M)paperPositivePart(M) by applying the real continuous functional calculus to the function x↦max⁡(x,0)x\mapsto\max(x,0)x↦max(x,0). Thus, for Hermitian MMM, it retains each positive eigenvalue and replaces each nonpositive eigenvalue by zero; for non-Hermitian MMM, this functional calculus returns the zero matrix. No Hermitian or positivity hypothesis is imposed on MMM.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperPower

For a finite index type III with decidable equality, a complex matrix MMM indexed by III, and any real number ppp, paperPower(M,p)\mathrm{paperPower}(M,p)paperPower(M,p) applies the real continuous functional calculus to the scalar function fpf_pfp​ defined by fp(0)=0f_p(0)=0fp​(0)=0 and fp(x)=xpf_p(x)=x^pfp​(x)=xp for x≠0x\neq0x=0, using the total real power convention. Thus, for Hermitian MMM, it applies fpf_pfp​ to each eigenvalue; for non-Hermitian MMM, this functional calculus returns the zero matrix. All real exponents are allowed, including zero and negative exponents; a zero eigenvalue is always sent to zero, even when p=0p=0p=0. Empty index types are included.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperProbability

For a finite index type XXX and a real function p:X→Rp:X\to\mathbb Rp:X→R, this predicate means that px≥0p_x\geq0px​≥0 for every x∈Xx\in Xx∈X and ∑x∈Xpx=1\sum_{x\in X}p_x=1∑x∈X​px​=1. If XXX is empty, the sum is zero, so no function satisfies the predicate.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperPure

For a finite index type III and a complex matrix rrr indexed by III, paperPure(r)\mathrm{paperPure}(r)paperPure(r) means that rrr is positive semidefinite and has rank exactly 111. There is no trace normalization condition. For an empty index type, this condition cannot hold.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperQuadratic

For a finite index type III, any complex matrix rrr indexed by III, and any function v:I→Cv:I\to\mathbb Cv:I→C, this definition gives the complex number

∑i∈I∑j∈Ivi‾ ri,j vj.\sum_{i\in I}\sum_{j\in I}\overline{v_i}\,r_{i,j}\,v_j.i∈I∑​j∈I∑​vi​​ri,j​vj​.

Neither Hermiticity of rrr nor normalization of vvv is assumed, and the result is not restricted to its real part. If III is empty, the value is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperRegroup

For arbitrary index types A,B,CA,B,CA,B,C and any complex matrix rrr indexed by (A×B)×C(A\times B)\times C(A×B)×C, this defines a matrix indexed by A×(B×C)A\times(B\times C)A×(B×C) by the entrywise rule

(paperRegroup⁡(r))(a,(b,c)),(a′,(b′,c′))=r((a,b),c),((a′,b′),c′).\bigl(\operatorname{paperRegroup}(r)\bigr)_{(a,(b,c)),(a',(b',c'))} =r_{((a,b),c),((a',b'),c')}.(paperRegroup(r))(a,(b,c)),(a′,(b′,c′))​=r((a,b),c),((a′,b′),c′)​.

There are no finiteness, nonemptiness, positivity, or normalization hypotheses; the definition changes the grouping of the indices.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperRelativeD

For a finite index type III with decidable equality, including the empty type, and arbitrary complex I×II\times II×I matrices r,sr,sr,s, this defines

paperRelativeD⁡(r,s)={Re⁡tr⁡(r(L(r)−L(s)))Re⁡tr⁡r,ker⁡s⊆ker⁡r,+∞,otherwise,\operatorname{paperRelativeD}(r,s)= \begin{cases} \displaystyle \frac{\operatorname{Re}\operatorname{tr}\bigl(r(L(r)-L(s))\bigr)} {\operatorname{Re}\operatorname{tr}r}, &\ker s\subseteq\ker r,\\[6pt] +\infty,&\text{otherwise}, \end{cases}paperRelativeD(r,s)=⎩⎨⎧​RetrrRetr(r(L(r)−L(s)))​,+∞,​kers⊆kerr,otherwise,​

where the quotient in the first branch is embedded into the extended reals and L(M)L(M)L(M) is the real continuous functional calculus of the total base-two logarithm. For Hermitian MMM, this replaces a nonzero eigenvalue λ\lambdaλ by log⁡2∣λ∣\log_2|\lambda|log2​∣λ∣ and a zero eigenvalue by zero; for non-Hermitian MMM, it returns zero. No positivity or normalization is assumed. Division by zero in the first branch gives zero, so that branch returns zero whenever Re⁡tr⁡r=0\operatorname{Re}\operatorname{tr}r=0Retrr=0.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperRenyi

Let III be any finite type with decidable equality, r,sr,sr,s arbitrary complex matrices on III, bbb a Boolean, and α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞]. For real ppp, let Pp(M)P_p(M)Pp​(M) be real functional calculus for x↦0x\mapsto0x↦0 at 000 and x↦xpx\mapsto x^px↦xp elsewhere, and define Qb(r,s;t)Q_b(r,s;t)Qb​(r,s;t) as Re⁡tr⁡Pt(P(1−t)/(2t)(s) r P(1−t)/(2t)(s))\operatorname{Re}\operatorname{tr}P_t(P_{(1-t)/(2t)}(s)\,r\,P_{(1-t)/(2t)}(s))RetrPt​(P(1−t)/(2t)​(s)rP(1−t)/(2t)​(s)) when bbb is true, and as Re⁡tr⁡(Pt(r)P1−t(s))\operatorname{Re}\operatorname{tr}(P_t(r)P_{1-t}(s))Retr(Pt​(r)P1−t​(s)) otherwise. Write S(r,s)S(r,s)S(r,s) for ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr, let log⁡2M\log_2Mlog2​M denote real functional calculus, and let L(0)=−∞L(0)=-\inftyL(0)=−∞, L(x)=log⁡2xL(x)=\log_2xL(x)=log2​x otherwise. Define db(r,s;1)d_b(r,s;1)db​(r,s;1) to be Re⁡tr⁡[r(log⁡2r−log⁡2s)]/Re⁡tr⁡r\operatorname{Re}\operatorname{tr}[r(\log_2r-\log_2s)]/\operatorname{Re}\operatorname{tr}rRetr[r(log2​r−log2​s)]/Retrr if S(r,s)S(r,s)S(r,s), and +∞+\infty+∞ otherwise; for t≠1t\ne1t=1, define db(r,s;t)=1t−1L(Qb(r,s;t)/Re⁡tr⁡r)d_b(r,s;t)=\frac1{t-1}L(Q_b(r,s;t)/\operatorname{Re}\operatorname{tr}r)db​(r,s;t)=t−11​L(Qb​(r,s;t)/Retrr) if (t<1 and Qb(r,s;t)>0)(t<1\text{ and }Q_b(r,s;t)>0)(t<1 and Qb​(r,s;t)>0) or S(r,s)S(r,s)S(r,s), and +∞+\infty+∞ otherwise. Also put Dmax⁡(r,s)=inf⁡{ℓ∈R:2ℓs−r⪰0}D_{\max}(r,s)=\inf\{\ell\in\mathbb R:2^\ell s-r\succeq0\}Dmax​(r,s)=inf{ℓ∈R:2ℓs−r⪰0}, as an extended-real infimum. The declaration defines

Rb(r,s;α)={inf⁡t>0db(r,s;t),α=0,Dmax⁡(r,s),α=+∞, b=true,sup⁡t>0db(r,s;t),α=+∞, b=false,db(r,s;α),0<α<+∞.\mathcal R_b(r,s;\alpha)= \begin{cases} \inf_{t>0}d_b(r,s;t),&\alpha=0,\\ D_{\max}(r,s),&\alpha=+\infty,\ b=\mathrm{true},\\ \sup_{t>0}d_b(r,s;t),&\alpha=+\infty,\ b=\mathrm{false},\\ d_b(r,s;\alpha),&0<\alpha<+\infty. \end{cases}Rb​(r,s;α)=⎩⎨⎧​inft>0​db​(r,s;t),Dmax​(r,s),supt>0​db​(r,s;t),db​(r,s;α),​α=0,α=+∞, b=true,α=+∞, b=false,0<α<+∞.​

The endpoint extrema range over every positive real ttt, including t=1t=1t=1. All indicated entropy arithmetic is extended-real; no positivity or normalization assumptions are made, empty index types and zero traces are permitted, and real division by zero is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperRenyiQ

For a finite type III with decidable equality, arbitrary complex square matrices r,sr,sr,s indexed by III, a Boolean bbb, and any real aaa, define Pp(M)P_p(M)Pp​(M) by real functional calculus applied to the function that is 000 at 000 and xpx^pxp at every nonzero real xxx. The declaration defines the real number

Qb(r,s;a)={Re⁡tr⁡Pa ⁣(P(1−a)/(2a)(s) r P(1−a)/(2a)(s)),b=true,Re⁡tr⁡ ⁣(Pa(r)P1−a(s)),b=false.Q_b(r,s;a)= \begin{cases} \operatorname{Re}\operatorname{tr}P_a\!\left(P_{(1-a)/(2a)}(s)\,r\,P_{(1-a)/(2a)}(s)\right),&b=\mathrm{true},\\ \operatorname{Re}\operatorname{tr}\!\left(P_a(r)P_{1-a}(s)\right),&b=\mathrm{false}. \end{cases}Qb​(r,s;a)={RetrPa​(P(1−a)/(2a)​(s)rP(1−a)/(2a)​(s)),Retr(Pa​(r)P1−a​(s)),​b=true,b=false.​

There are no positivity, normalization, nonzero-matrix, or restrictions on aaa. In particular, a=0a=0a=0 and a=1a=1a=1 are included; division by zero uses the total real operation and gives 000, and the scalar function defining P0P_0P0​ still takes value 000 at 000. The index type may be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperRenyiReal

For any finite index type III with decidable equality, any Boolean bbb, any complex matrices r,sr,sr,s indexed by I×II\times II×I, and any real aaa, this defines the real value obtained from the following extended-real expression, sending either +∞+\infty+∞ or −∞-\infty−∞ to 000. Write Pt(M)P_t(M)Pt​(M) for the continuous functional calculus of MMM with the real function equal to 000 at x=0x=0x=0 and to xtx^txt otherwise, write L(M)L(M)L(M) for the continuous functional calculus with x↦log⁡2xx\mapsto\log_2 xx↦log2​x, and let S(r,s)S(r,s)S(r,s) mean ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr. Define

Qb(r,s,a)={Re⁡tr⁡ ⁣(Pa ⁣(P1−a2a(s) r P1−a2a(s))),b=true,Re⁡tr⁡ ⁣(Pa(r)P1−a(s)),b=false.Q_b(r,s,a)= \begin{cases} \operatorname{Re}\operatorname{tr}\!\left( P_a\!\left(P_{\frac{1-a}{2a}}(s)\,r\,P_{\frac{1-a}{2a}}(s)\right)\right), &b=\mathrm{true},\\ \operatorname{Re}\operatorname{tr}\!\left(P_a(r)P_{1-a}(s)\right), &b=\mathrm{false}. \end{cases}Qb​(r,s,a)={Retr(Pa​(P2a1−a​​(s)rP2a1−a​​(s))),Retr(Pa​(r)P1−a​(s)),​b=true,b=false.​

With ℓ(0)=−∞\ell(0)=-\inftyℓ(0)=−∞ and ℓ(x)=log⁡2x\ell(x)=\log_2 xℓ(x)=log2​x as an embedded real otherwise, the extended-real expression is

Db(r,s,a)={Re⁡tr⁡(r(L(r)−L(s)))Re⁡tr⁡r,a=1 and S(r,s),+∞,a=1 and ¬S(r,s),1a−1 ℓ ⁣(Qb(r,s,a)Re⁡tr⁡r),a≠1 and ((a<1 and Qb(r,s,a)>0) or S(r,s)),+∞,otherwise.D_b(r,s,a)= \begin{cases} \displaystyle \frac{\operatorname{Re}\operatorname{tr}(r(L(r)-L(s)))} {\operatorname{Re}\operatorname{tr}r}, &a=1\text{ and }S(r,s),\\ +\infty,&a=1\text{ and }\neg S(r,s),\\ \displaystyle\frac{1}{a-1}\, \ell\!\left(\frac{Q_b(r,s,a)}{\operatorname{Re}\operatorname{tr}r}\right), &a\ne1\text{ and }\bigl((a<1\text{ and }Q_b(r,s,a)>0)\text{ or }S(r,s)\bigr),\\ +\infty,&\text{otherwise}. \end{cases}Db​(r,s,a)=⎩⎨⎧​RetrrRetr(r(L(r)−L(s)))​,+∞,a−11​ℓ(RetrrQb​(r,s,a)​),+∞,​a=1 and S(r,s),a=1 and ¬S(r,s),a=1 and ((a<1 and Qb​(r,s,a)>0) or S(r,s)),otherwise.​

Thus paperRenyiReal⁡(b,r,s,a)=Db(r,s,a).toReal\operatorname{paperRenyiReal}(b,r,s,a)=D_b(r,s,a).\mathrm{toReal}paperRenyiReal(b,r,s,a)=Db​(r,s,a).toReal. There are no positivity, nonzero, normalization, nonemptiness, or parameter-range assumptions: in particular a=0a=0a=0, negative aaa, and zero trace are included, with real division by zero equal to 000 and the supplied total functional-calculus, logarithm, and real-power operations used throughout.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSchatten

For a finite index type III with decidable equality, any complex matrix MMM indexed by III, and any real number ppp, let S=M†MS=\sqrt{M^\dagger M}S=M†M​, using the positive semidefinite matrix square root, and let SpS_pSp​ be obtained by replacing each zero eigenvalue of SSS by zero and each nonzero eigenvalue sss by sps^psp. Then paperSchatten(M,p)\mathrm{paperSchatten}(M,p)paperSchatten(M,p) is the real number (Re⁡tr⁡(Sp))1/p\bigl(\operatorname{Re}\operatorname{tr}(S_p)\bigr)^{1/p}(Retr(Sp​))1/p, with total real division and exponentiation. There is no restriction on ppp; in particular, at p=0p=0p=0, the exponent 1/p1/p1/p is zero and the result is 111, including for the zero matrix and for an empty index type.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSeparable

For finite index types A,BA,BA,B and a complex square matrix rrr indexed by A×BA\times BA×B, this predicate means that there exist a natural number nnn, real weights w0,…,wn−1w_0,\ldots,w_{n-1}w0​,…,wn−1​, complex square matrices sis_isi​ indexed by AAA, and complex square matrices tit_iti​ indexed by BBB, such that every wi≥0w_i\geq0wi​≥0, ∑i=0n−1wi=1\sum_{i=0}^{n-1}w_i=1∑i=0n−1​wi​=1, every si,tis_i,t_isi​,ti​ is positive semidefinite with complex trace equal to 111, and

r=∑i=0n−1wi(si⊗ti),(si⊗ti)(a,b),(a′,b′)=(si)a,a′(ti)b,b′.r=\sum_{i=0}^{n-1}w_i(s_i\otimes t_i), \qquad (s_i\otimes t_i)_{(a,b),(a',b')}=(s_i)_{a,a'}(t_i)_{b,b'}.r=i=0∑n−1​wi​(si​⊗ti​),(si​⊗ti​)(a,b),(a′,b′)​=(si​)a,a′​(ti​)b,b′​.

The quantified nnn includes zero, but zero cannot satisfy the weight-sum condition. Zero weights are allowed. If either AAA or BBB is empty, the required trace-one factor matrices cannot exist.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperShannon

For a finite type XXX and any function p:X→Rp:X\to\mathbb Rp:X→R, this defines −∑x∈Xp(x)log⁡2p(x)-\sum_{x\in X}p(x)\log_2p(x)−∑x∈X​p(x)log2​p(x). No nonnegativity or normalization condition is imposed. The logarithm is the total real logarithm, with log⁡20=0\log_2 0=0log2​0=0, so zero entries contribute zero; an empty XXX gives zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSixAxioms

For a rule DI(r,s)D_I(r,s)DI​(r,s) returning a real number for each pair of square complex matrices on each finite index type with decidable equality, and a function g:R→Rg:\mathbb R\to\mathbb Rg:R→R, this predicate is the conjunction of the following requirements. Define AI={(r,s):r,s are positive semidefinite, r≠0, ker⁡s⊆ker⁡r}\mathcal A_I=\{(r,s):r,s\text{ are positive semidefinite},\ r\ne0,\ \ker s\subseteq\ker r\}AI​={(r,s):r,s are positive semidefinite, r=0, kers⊆kerr}. The function ggg is continuous and is either strictly increasing or strictly decreasing; for every finite nonempty III with decidable equality, (r,s)↦DI(r,s)(r,s)\mapsto D_I(r,s)(r,s)↦DI​(r,s) is continuous on AI\mathcal A_IAI​; for every two such types I,JI,JI,J, every (r,s)∈AI(r,s)\in\mathcal A_I(r,s)∈AI​, and every complex matrix UUU with rows indexed by JJJ and columns by III satisfying both U∗U=1U^*U=1U∗U=1 and UU∗=1UU^*=1UU∗=1,

DI(r,s)=DJ(UrU∗,UsU∗);D_I(r,s)=D_J(UrU^*,UsU^*);DI​(r,s)=DJ​(UrU∗,UsU∗);

on the singleton index type, D(1,12 1)=log⁡22D(1,\tfrac12\,1)=\log_2 2D(1,21​1)=log2​2; for every such III and (r,s)∈AI(r,s)\in\mathcal A_I(r,s)∈AI​, positive semidefiniteness of r−sr-sr−s implies 0≤DI(r,s)0\le D_I(r,s)0≤DI​(r,s), and positive semidefiniteness of s−rs-rs−r implies DI(r,s)≤0D_I(r,s)\le0DI​(r,s)≤0; for every such I,JI,JI,J, (r,s)∈AI(r,s)\in\mathcal A_I(r,s)∈AI​, and (t,u)∈AJ(t,u)\in\mathcal A_J(t,u)∈AJ​,

DI×J(r⊗t,s⊗u)=DI(r,s)+DJ(t,u)D_{I\times J}(r\otimes t,s\otimes u)=D_I(r,s)+D_J(t,u)DI×J​(r⊗t,s⊗u)=DI​(r,s)+DJ​(t,u)

and

g ⁣(DI⊔J(r⊕t,s⊕u))=Re⁡tr⁡rRe⁡tr⁡r+Re⁡tr⁡t g(DI(r,s))+Re⁡tr⁡tRe⁡tr⁡r+Re⁡tr⁡t g(DJ(t,u)).g\!\left(D_{I\sqcup J}(r\oplus t,s\oplus u)\right) = \frac{\operatorname{Re}\operatorname{tr}r} {\operatorname{Re}\operatorname{tr}r+\operatorname{Re}\operatorname{tr}t}\, g(D_I(r,s)) + \frac{\operatorname{Re}\operatorname{tr}t} {\operatorname{Re}\operatorname{tr}r+\operatorname{Re}\operatorname{tr}t}\, g(D_J(t,u)).g(DI⊔J​(r⊕t,s⊕u))=Retrr+RetrtRetrr​g(DI​(r,s))+Retrr+RetrtRetrt​g(DJ​(t,u)).

Here U∗U^*U∗ is the conjugate transpose, ⊗\otimes⊗ is the Kronecker product, and ⊕\oplus⊕ puts its arguments in the two diagonal blocks on the disjoint union and puts zero in the mixed blocks. These requirements do not restrict the matrices to trace one, and the stated continuity, invariance, ordering, and composition requirements quantify only over nonempty index types.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSmoothMax

For finite index types A,BA,BA,B with decidable equality, any complex matrix rrr indexed by A×BA\times BA×B, and any real eee, let S(M)S(M)S(M) denote the total real continuous functional calculus for x↦xx\mapsto\sqrt{x}x↦x​, and let KKK consist of all matrices t⪰0t\succeq0t⪰0 with Re⁡tr⁡t≤1\operatorname{Re}\operatorname{tr}t\le1Retrt≤1 satisfying

1−(Re⁡tr⁡S(S(t)rS(t))+(1−Re⁡tr⁡t)(1−Re⁡tr⁡r))2≤e.\sqrt{1-\left( \operatorname{Re}\operatorname{tr}S\bigl(S(t)rS(t)\bigr) +\sqrt{(1-\operatorname{Re}\operatorname{tr}t)(1-\operatorname{Re}\operatorname{tr}r)} \right)^2}\le e.1−(RetrS(S(t)rS(t))+(1−Retrt)(1−Retrr)​)2​≤e.

This defines the extended real number

inf⁡t∈K sup⁡u∈CB×Bu⪰0, Re⁡tr⁡u≤1L ⁣([Re⁡tr⁡S(S(t)(IA⊗u)S(t))]2),\inf_{t\in K}\ \sup_{\substack{u\in\mathbb C^{B\times B}\\ u\succeq0,\ \operatorname{Re}\operatorname{tr}u\le1}} L\!\left( \left[\operatorname{Re}\operatorname{tr} S\bigl(S(t)(I_A\otimes u)S(t)\bigr)\right]^2 \right),t∈Kinf​ u∈CB×Bu⪰0, Retru≤1​sup​L([RetrS(S(t)(IA​⊗u)S(t))]2),

where L(0)=−∞L(0)=-\inftyL(0)=−∞ and L(x)=log⁡2xL(x)=\log_2 xL(x)=log2​x, embedded into the extended reals, for x≠0x\ne0x=0. Scalar square roots of negative inputs are 000, zero witnesses uuu are permitted, and no positivity condition on rrr or restriction on eee is imposed. Empty index types are allowed. An infimum over empty KKK is +∞+\infty+∞; in particular, this occurs when e<0e<0e<0.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSmoothMin

For finite index types A,BA,BA,B with decidable equality, any complex matrix rrr indexed by A×BA\times BA×B, and any real eee, let S(M)S(M)S(M) denote the total real continuous functional calculus for x↦xx\mapsto\sqrt{x}x↦x​, and let KKK be the set of matrices t⪰0t\succeq0t⪰0 with Re⁡tr⁡t≤1\operatorname{Re}\operatorname{tr}t\le1Retrt≤1 satisfying

1−(Re⁡tr⁡S(S(t)rS(t))+(1−Re⁡tr⁡t)(1−Re⁡tr⁡r))2≤e.\sqrt{1-\left( \operatorname{Re}\operatorname{tr}S\bigl(S(t)rS(t)\bigr) +\sqrt{(1-\operatorname{Re}\operatorname{tr}t)(1-\operatorname{Re}\operatorname{tr}r)} \right)^2}\le e.1−(RetrS(S(t)rS(t))+(1−Retrt)(1−Retrr)​)2​≤e.

This defines the extended real number

sup⁡t∈K sup⁡{ℓ∈R:∃u∈CB×B, u⪰0,Re⁡tr⁡u≤1,2−ℓ(IA⊗u)−t⪰0},\sup_{t\in K}\ \sup\left\{\ell\in\mathbb R: \exists u\in\mathbb C^{B\times B},\ u\succeq0,\quad \operatorname{Re}\operatorname{tr}u\le1,\quad 2^{-\ell}(I_A\otimes u)-t\succeq0 \right\},t∈Ksup​ sup{ℓ∈R:∃u∈CB×B, u⪰0,Retru≤1,2−ℓ(IA​⊗u)−t⪰0},

with real values embedded into the extended reals. Scalar square roots of negative inputs are 000. Neither positivity of rrr nor a restriction on eee is assumed, and empty index types are allowed. Each empty supremum is −∞-\infty−∞; in particular, e<0e<0e<0 makes KKK empty and the defined value −∞-\infty−∞.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSourceCode

For any type XXX and natural number mmm, this structure consists of two functions: an encoding function X→{0,…,m−1}X\to\{0,\ldots,m-1\}X→{0,…,m−1} and a decoding function {0,…,m−1}→X\{0,\ldots,m-1\}\to X{0,…,m−1}→X. No inverse or reconstruction condition is required. The case m=0m=0m=0 is included, and such a structure can then exist only when XXX is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSpectralProjector

For a finite index type III with decidable equality, any complex square matrix HHH indexed by III, and any real number lll, this defines fl(H)f_l(H)fl​(H) using real continuous functional calculus, where fl(x)=1f_l(x)=1fl​(x)=1 if x=lx=lx=l, and fl(x)=0f_l(x)=0fl​(x)=0 otherwise. For Hermitian HHH, this is the orthogonal projection onto its lll-eigenspace, including the zero projection when lll is not an eigenvalue. For non-Hermitian HHH, this functional-calculus expression is defined to be zero; an empty III gives the unique empty matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSubnormalized

For a finite index type III and a complex matrix rrr indexed by III, paperSubnormalized(r)\mathrm{paperSubnormalized}(r)paperSubnormalized(r) means that rrr is positive semidefinite and Re⁡tr⁡(r)≤1\operatorname{Re}\operatorname{tr}(r)\leq1Retr(r)≤1. Zero matrices are included; for an empty index type, the unique matrix satisfies this condition.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSupported

For a finite index type III and complex matrices M,NM,NM,N indexed by III, paperSupported(M,N)\mathrm{paperSupported}(M,N)paperSupported(M,N) means ker⁡N⊆ker⁡M\ker N\subseteq\ker MkerN⊆kerM, where the kernels are those of matrix multiplication on vectors I→CI\to\mathbb CI→C. Thus every vector annihilated by NNN must also be annihilated by MMM. This includes the empty index type.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperSymmetricError

For any finite index type III with decidable equality and any complex matrices r,sr,sr,s indexed by I×II\times II×I, this defines

paperSymmetricError⁡(r,s)=inf⁡T⪰01−T⪰0Re⁡tr⁡((1−T)r)+Re⁡tr⁡(Ts)2.\operatorname{paperSymmetricError}(r,s)= \inf_{\substack{T\succeq0\\ \mathbf 1-T\succeq0}} \frac{\operatorname{Re}\operatorname{tr}((\mathbf 1-T)r) +\operatorname{Re}\operatorname{tr}(Ts)}{2}.paperSymmetricError(r,s)=T⪰01−T⪰0​inf​2Retr((1−T)r)+Retr(Ts)​.

No positivity or normalization assumptions are imposed on rrr or sss. The infimum is taken in R\mathbb RR, whose infimum convention assigns 000 to an empty set or a set unbounded below; here the feasible set contains T=0T=0T=0. Empty III is permitted and gives value 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperTensorAction

For finite index types A,BA,BA,B with decidable equality, output index types C,DC,DC,D, complex linear maps E:Mat⁡A(C)→Mat⁡C(C)E:\operatorname{Mat}_A(\mathbb C)\to\operatorname{Mat}_C(\mathbb C)E:MatA​(C)→MatC​(C) and F:Mat⁡B(C)→Mat⁡D(C)F:\operatorname{Mat}_B(\mathbb C)\to\operatorname{Mat}_D(\mathbb C)F:MatB​(C)→MatD​(C), and any complex matrix rrr indexed by A×BA\times BA×B, this defines a matrix TTT indexed by C×DC\times DC×D whose entries are T(c,d),(c′,d′)=∑a,a′∈A∑b,b′∈Br(a,b),(a′,b′) [E(ea,a′)]c,c′ [F(eb,b′)]d,d′\displaystyle T_{(c,d),(c',d')}=\sum_{a,a'\in A}\sum_{b,b'\in B}r_{(a,b),(a',b')}\,[E(e_{a,a'})]_{c,c'}\,[F(e_{b,b'})]_{d,d'}T(c,d),(c′,d′)​=a,a′∈A∑​b,b′∈B∑​r(a,b),(a′,b′)​[E(ea,a′​)]c,c′​[F(eb,b′​)]d,d′​, where eu,ve_{u,v}eu,v​ is the matrix with entry 111 at (u,v)(u,v)(u,v) and 000 elsewhere. No positivity or trace preservation is required of either map, and rrr is unrestricted; if either input index type is empty, the sums are empty and the result is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperTensorPower

For any index type III, any complex matrix rrr indexed by III, and any natural number nnn, this defines a matrix indexed by functions x:{0,…,n−1}→Ix:\{0,\ldots,n-1\}\to Ix:{0,…,n−1}→I, with entries (r⊗n)x,y=∏i=0n−1rx(i),y(i)\displaystyle (r^{\otimes n})_{x,y}=\prod_{i=0}^{n-1}r_{x(i),y(i)}(r⊗n)x,y​=i=0∏n−1​rx(i),y(i)​. No positivity or normalization is required. When n=0n=0n=0, the indexing function space is a singleton and the matrix has sole entry 111, including when III is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperTraceMiddle

For arbitrary index types A,CA,CA,C, a finite index type BBB, and any complex matrix rrr indexed by (A×B)×C(A\times B)\times C(A×B)×C, this defines a complex matrix indexed by A×CA\times CA×C whose entries are

(paperTraceMiddle⁡(r))(a,c),(a′,c′)=∑b∈Br((a,b),c),((a′,b),c′).\bigl(\operatorname{paperTraceMiddle}(r)\bigr)_{(a,c),(a',c')} =\sum_{b\in B}r_{((a,b),c),((a',b),c')}.(paperTraceMiddle(r))(a,c),(a′,c′)​=b∈B∑​r((a,b),c),((a′,b),c′)​.

No positivity, Hermitian, or normalization condition is required. If BBB is empty, the sum is 000, so the resulting matrix is zero; AAA and CCC may also be empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperTypeI

For any finite index type III with decidable equality and any complex matrices r,Tr,Tr,T indexed by I×II\times II×I, this defines

paperTypeI⁡(r,T)=Re⁡tr⁡((1−T)r),\operatorname{paperTypeI}(r,T)=\operatorname{Re}\operatorname{tr}((\mathbf 1-T)r),paperTypeI(r,T)=Retr((1−T)r),

where 1\mathbf 11 is the identity matrix. No positivity, normalization, or bound on either matrix is assumed. The index type may be empty, in which case the value is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperTypeII

For any finite index type III with decidable equality and any complex matrices s,Ts,Ts,T indexed by I×II\times II×I, this defines

paperTypeII⁡(s,T)=Re⁡tr⁡(Ts).\operatorname{paperTypeII}(s,T)=\operatorname{Re}\operatorname{tr}(Ts).paperTypeII(s,T)=Retr(Ts).

No positivity, normalization, or bound on either matrix is assumed. The index type may be empty, in which case the value is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.paperVariance

For a finite index type III with decidable equality, including the empty type, and arbitrary complex I×II\times II×I matrices r,sr,sr,s, this defines

paperVariance⁡(r,s)=Re⁡tr⁡[r(L(r)−L(s)−dII)2].\operatorname{paperVariance}(r,s) = \operatorname{Re}\operatorname{tr} \left[r\left(L(r)-L(s)-dI_I\right)^2\right].paperVariance(r,s)=Retr[r(L(r)−L(s)−dII​)2].

Here III_III​ is the identity matrix, and LLL applies the total real base-two logarithm by continuous functional calculus: on Hermitian matrices it sends nonzero eigenvalues λ\lambdaλ to log⁡2∣λ∣\log_2|\lambda|log2​∣λ∣ and zero eigenvalues to zero, while on non-Hermitian matrices it returns zero. If ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr, then

d=Re⁡tr⁡(r(L(r)−L(s)))Re⁡tr⁡r;d=\frac{\operatorname{Re}\operatorname{tr}\bigl(r(L(r)-L(s))\bigr)} {\operatorname{Re}\operatorname{tr}r};d=RetrrRetr(r(L(r)−L(s)))​;

otherwise d=0d=0d=0, because conversion of the extended real value +∞+\infty+∞ to a real number gives zero. Real division by zero also gives zero. No positivity or normalization of either matrix is assumed, and the square is matrix multiplication of the centered expression by itself.

QAlgorithms.Papers.Independent.Intakenext100.P07.petzExtendedQ

For a finite index type III with decidable equality, arbitrary complex matrices r,sr,sr,s indexed by III, and any real aaa, this defines an extended-real value equal to +∞+\infty+∞ when a>1a>1a>1 and ker⁡s⊈ker⁡r\ker s\not\subseteq\ker rkers⊆kerr, and otherwise equal to Re⁡tr⁡(Pa(r)P1−a(s))\operatorname{Re}\operatorname{tr}(P_a(r)P_{1-a}(s))Retr(Pa​(r)P1−a​(s)), embedded into the extended reals. Here Pt(M)P_t(M)Pt​(M) is the continuous functional calculus with the real function taking value zero at zero and xtx^txt at every nonzero xxx. No positivity, normalization, or nonemptiness assumption is imposed; in particular, a=1a=1a=1 uses the second branch, and P0P_0P0​ still has value zero on the zero spectral value.

QAlgorithms.Papers.Independent.Intakenext100.P07.positiveNorm

For a finite index type III with decidable equality and any complex matrix MMM indexed by III, positiveNorm(M)\mathrm{positiveNorm}(M)positiveNorm(M) is the real supremum sup⁡{∣tr⁡(ME)∣:E and 1−E are positive semidefinite}\sup\{|\operatorname{tr}(ME)|: E\text{ and }\mathbf1-E\text{ are positive semidefinite}\}sup{∣tr(ME)∣:E and 1−E are positive semidefinite}, where the absolute value is the complex modulus and 1\mathbf11 is the identity matrix. The supremum ranges over all such complex matrices EEE. For an empty index type, its value is zero.

QAlgorithms.Papers.Independent.Intakenext100.P07.preparedEntangled

For finite index types A,BA,BA,B, with decidable equality on BBB, and any complex-linear map E:CB×B→CA×AE:\mathbb C^{B\times B}\to\mathbb C^{A\times A}E:CB×B→CA×A, this definition gives a matrix on A×BA\times BA×B by applying EEE to the second matrix factor of the matrix on B×BB\times BB×B whose ((b,c),(b′,c′))((b,c),(b',c'))((b,c),(b′,c′))-entry is ∣B∣−1|B|^{-1}∣B∣−1 when b=cb=cb=c and b′=c′b'=c'b′=c′, and 000 otherwise, and then exchanging the two output factors. Explicitly, for all a,a′∈Aa,a'\in Aa,a′∈A and b,b′∈Bb,b'\in Bb,b′∈B,

R(a,b),(a′,b′)=∣B∣−1(E(ebb′))a,a′,R_{(a,b),(a',b')}=|B|^{-1}\bigl(E(e_{bb'})\bigr)_{a,a'},R(a,b),(a′,b′)​=∣B∣−1(E(ebb′​))a,a′​,

where ebb′e_{bb'}ebb′​ has a single entry 111 at (b,b′)(b,b')(b,b′). No positivity or trace-preservation assumption is imposed on EEE; if either index type is empty, the output matrix has no entries.

QAlgorithms.Papers.Independent.Intakenext100.P07.productState

For every natural number nnn, index type III, and family of arbitrary complex matrices r0,…,rn−1r_0,\ldots,r_{n-1}r0​,…,rn−1​ indexed by III, this definition gives a matrix indexed by functions {0,…,n−1}→I\{0,\ldots,n-1\}\to I{0,…,n−1}→I, with entries

Rx,y=∏i=0n−1(ri)x(i),y(i)R_{x,y}=\prod_{i=0}^{n-1}(r_i)_{x(i),y(i)}Rx,y​=i=0∏n−1​(ri​)x(i),y(i)​

for every pair of such functions x,yx,yx,y. No positivity or normalization is assumed. For n=0n=0n=0, the function index type has one element and the empty product makes RRR the 1×11\times11×1 matrix with entry 111, including when III is empty.

QAlgorithms.Papers.Independent.Intakenext100.P07.purified

For any finite index type III with decidable equality and arbitrary complex matrices r,s∈CI×Ir,s\in\mathbb C^{I\times I}r,s∈CI×I, this defines

1−(Re⁡tr⁡S(S(r) s S(r))+(1−Re⁡tr⁡r)(1−Re⁡tr⁡s))2,\sqrt{ 1-\left( \operatorname{Re}\operatorname{tr} S\bigl(S(r)\,s\,S(r)\bigr) +\sqrt{\bigl(1-\operatorname{Re}\operatorname{tr}r\bigr) \bigl(1-\operatorname{Re}\operatorname{tr}s\bigr)} \right)^2 },1−(RetrS(S(r)sS(r))+(1−Retrr)(1−Retrs)​)2​,

where S(M)S(M)S(M) is the total real continuous functional calculus for the real square-root function. Neither matrix is assumed positive or normalized, and the index type may be empty. Every scalar real square root is 000 when its argument is negative; thus the outer square root is also 000 if the squared expression exceeds 111.

QAlgorithms.Papers.Independent.Intakenext100.P07.renyiFinite

For a finite type III with decidable equality, arbitrary complex matrices r,sr,sr,s on III, a Boolean bbb, and any real aaa, let Pp(M)P_p(M)Pp​(M) be real functional calculus for the scalar function x↦0x\mapsto0x↦0 when x=0x=0x=0, and x↦xpx\mapsto x^px↦xp otherwise, and put

Qb(r,s;a)={Re⁡tr⁡Pa(P(1−a)/(2a)(s) r P(1−a)/(2a)(s)),b=true,Re⁡tr⁡(Pa(r)P1−a(s)),b=false.Q_b(r,s;a)= \begin{cases} \operatorname{Re}\operatorname{tr}P_a(P_{(1-a)/(2a)}(s)\,r\,P_{(1-a)/(2a)}(s)),&b=\mathrm{true},\\ \operatorname{Re}\operatorname{tr}(P_a(r)P_{1-a}(s)),&b=\mathrm{false}. \end{cases}Qb​(r,s;a)={RetrPa​(P(1−a)/(2a)​(s)rP(1−a)/(2a)​(s)),Retr(Pa​(r)P1−a​(s)),​b=true,b=false.​

Write S(r,s)S(r,s)S(r,s) for ker⁡s⊆ker⁡r\ker s\subseteq\ker rkers⊆kerr, define log⁡2M\log_2Mlog2​M by real functional calculus, and set D(r,s)=Re⁡tr⁡[r(log⁡2r−log⁡2s)]/Re⁡tr⁡rD(r,s)=\operatorname{Re}\operatorname{tr}[r(\log_2r-\log_2s)]/\operatorname{Re}\operatorname{tr}rD(r,s)=Retr[r(log2​r−log2​s)]/Retrr when S(r,s)S(r,s)S(r,s) holds, and D(r,s)=+∞D(r,s)=+\inftyD(r,s)=+∞ otherwise. With L(0)=−∞L(0)=-\inftyL(0)=−∞ and L(x)=log⁡2xL(x)=\log_2xL(x)=log2​x for nonzero real xxx, the declaration defines

db(r,s;a)={D(r,s),a=1,1a−1 L ⁣(Qb(r,s;a)Re⁡tr⁡r),a≠1 and [(a<1 and Qb(r,s;a)>0) or S(r,s)],+∞,otherwise.d_b(r,s;a)= \begin{cases} D(r,s),&a=1,\\ \displaystyle \frac1{a-1}\,L\!\left(\frac{Q_b(r,s;a)}{\operatorname{Re}\operatorname{tr}r}\right), &a\ne1\ \text{and}\ \bigl[(a<1\ \text{and}\ Q_b(r,s;a)>0)\ \text{or}\ S(r,s)\bigr],\\ +\infty,&\text{otherwise}. \end{cases}db​(r,s;a)=⎩⎨⎧​D(r,s),a−11​L(RetrrQb​(r,s;a)​),+∞,​a=1,a=1 and [(a<1 and Qb​(r,s;a)>0) or S(r,s)],otherwise.​

The product in the middle branch is extended-real multiplication after embedding 1/(a−1)1/(a-1)1/(a−1). No positivity, normalization, or nonzero-trace hypotheses are imposed; an empty III, a=0a=0a=0, and zero denominators are allowed, with real division by zero equal to 000 and the real logarithm interpreted as its total real function.

QAlgorithms.Papers.Independent.Intakenext100.P07.resolutionPinch

For a finite index type III, any natural number mmm, any family of complex matrices P0,…,Pm−1P_0,\ldots,P_{m-1}P0​,…,Pm−1​ indexed by III, and any complex matrix MMM indexed by III, this defines ∑i=0m−1PiMPi\sum_{i=0}^{m-1}P_iMP_i∑i=0m−1​Pi​MPi​. No Hermitian, idempotence, orthogonality, or sum-to-identity condition is assumed on the family. When m=0m=0m=0, the result is the zero matrix.

QAlgorithms.Papers.Independent.Intakenext100.P07.sdpDual

For finite index types I,JI,JI,J, with decidable equality on III, any complex-linear map E:CI×I→CJ×JE:\mathbb C^{I\times I}\to\mathbb C^{J\times J}E:CI×I→CJ×J, and arbitrary complex matrices KKK on III and LLL on JJJ, this definition gives the extended-real value

sup⁡{Re⁡tr⁡(LY):Y∈CJ×J,Y⪰0,K−E♯(Y)⪰0},(E♯(Y))i,j=∑a,b∈J(E(eij))a,b‾ Ya,b,\sup\left\{ \operatorname{Re}\operatorname{tr}(LY): Y\in\mathbb C^{J\times J},\quad Y\succeq0,\quad K-E^\sharp(Y)\succeq0 \right\}, \qquad \bigl(E^\sharp(Y)\bigr)_{i,j} =\sum_{a,b\in J}\overline{\bigl(E(e_{ij})\bigr)_{a,b}}\,Y_{a,b},sup{Retr(LY):Y∈CJ×J,Y⪰0,K−E♯(Y)⪰0},(E♯(Y))i,j​=a,b∈J∑​(E(eij​))a,b​​Ya,b​,

where eije_{ij}eij​ has a single entry 111 at (i,j)(i,j)(i,j), and ⪰0\succeq0⪰0 means positive semidefinite. Each real objective value is embedded into the extended reals; an empty feasible set gives −∞-\infty−∞, and objective values unbounded above give +∞+\infty+∞. Empty index types are permitted.

QAlgorithms.Papers.Independent.Intakenext100.P07.sdpPrimal

For finite index types I,JI,JI,J, any complex-linear map E:CI×I→CJ×JE:\mathbb C^{I\times I}\to\mathbb C^{J\times J}E:CI×I→CJ×J, and arbitrary complex matrices KKK on III and LLL on JJJ, this definition gives the extended-real value

inf⁡{Re⁡tr⁡(KX):X∈CI×I,X⪰0,E(X)−L⪰0},\inf\left\{ \operatorname{Re}\operatorname{tr}(KX): X\in\mathbb C^{I\times I},\quad X\succeq0,\quad E(X)-L\succeq0 \right\},inf{Retr(KX):X∈CI×I,X⪰0,E(X)−L⪰0},

where ⪰0\succeq0⪰0 means positive semidefinite and each real objective value is embedded into the extended reals. An empty feasible set gives +∞+\infty+∞, and objective values unbounded below give −∞-\infty−∞. Empty index types are permitted.

QAlgorithms.Papers.Independent.Intakenext100.P07.smoothMaxD

For a finite index type III with decidable equality, including the empty type, arbitrary complex I×II\times II×I matrices r,sr,sr,s, and every real eee, this defines

smoothMaxD⁡(r,s,e)=inf⁡u positive semidefiniteRe⁡tr⁡u≤1P(u,r)≤e inf⁡{ℓ∈R:2ℓs−u is positive semidefinite},\operatorname{smoothMaxD}(r,s,e) = \inf_{\substack{u\ \text{positive semidefinite}\\ \operatorname{Re}\operatorname{tr}u\le1\\ P(u,r)\le e}} \ \inf\left\{\ell\in\mathbb R:2^\ell s-u\text{ is positive semidefinite}\right\},smoothMaxD(r,s,e)=u positive semidefiniteRetru≤1P(u,r)≤e​inf​ inf{ℓ∈R:2ℓs−u is positive semidefinite},

with both infima taken in the extended reals, where

P(u,r)=1−(Re⁡tr⁡S ⁣(S(u) r S(u))+(1−Re⁡tr⁡u)(1−Re⁡tr⁡r))2.P(u,r)= \sqrt{1-\left( \operatorname{Re}\operatorname{tr} \mathcal S\!\left(\mathcal S(u)\,r\,\mathcal S(u)\right) + \sqrt{(1-\operatorname{Re}\operatorname{tr}u)(1-\operatorname{Re}\operatorname{tr}r)} \right)^2}.P(u,r)=1−(RetrS(S(u)rS(u))+(1−Retru)(1−Retrr)​)2​.

Here S(H)\mathcal S(H)S(H) is the real continuous functional calculus of x↦xx\mapsto\sqrt{x}x↦x​: on a Hermitian matrix it replaces each eigenvalue λ\lambdaλ by max⁡(λ,0)\sqrt{\max(\lambda,0)}max(λ,0)​, and on a non-Hermitian matrix it is zero. All real square roots return zero on negative arguments. No positivity or trace condition is imposed on rrr or sss. An empty outer admissible set gives +∞+\infty+∞; in particular, this occurs for e<0e<0e<0.

QAlgorithms.Papers.Independent.Intakenext100.P07.spectralValues

For a finite index type III with decidable equality and any complex square matrix HHH indexed by III, this defines the finite set of values in the following list: if HHH is Hermitian, its eigenvalues in ascending order, counted with multiplicity; otherwise, a list of ∣I∣|I|∣I∣ zeros. Taking the set removes repetitions. Thus the result is empty when III is empty, and is {0}\{0\}{0} when III is nonempty and HHH is not Hermitian.

QAlgorithms.Papers.Independent.Intakenext100.P07.transformedOrder

For every extended nonnegative real α∈[0,∞]\alpha\in[0,\infty]α∈[0,∞], this definition returns 222 when α=+∞\alpha=+\inftyα=+∞, and otherwise returns the extended nonnegative real obtained by truncating 2−1/αR2-1/\alpha_{\mathbb R}2−1/αR​ below at 000, where αR\alpha_{\mathbb R}αR​ is the real value of finite α\alphaα. Explicitly,

T(α)={2,α=+∞,max⁡(2−1/αR,0),α<+∞.T(\alpha)= \begin{cases} 2,&\alpha=+\infty,\\ \max(2-1/\alpha_{\mathbb R},0),&\alpha<+\infty. \end{cases}T(α)={2,max(2−1/αR​,0),​α=+∞,α<+∞.​

Real division by zero is 000, so T(0)=2T(0)=2T(0)=2; for 0<α≤120<\alpha\le\frac120<α≤21​, the value is 000.

QAlgorithms.Papers.Independent.Intakenext100.P07.variation

For a finite type XXX and arbitrary functions p,q:X→Rp,q:X\to\mathbb Rp,q:X→R, this defines 12∑x∈X∣p(x)−q(x)∣\tfrac12\sum_{x\in X}|p(x)-q(x)|21​∑x∈X​∣p(x)−q(x)∣. Neither function is required to be nonnegative or normalized. If XXX is empty, the value is zero.

Human review
  • Endorsed by Shuze Chen · Oct 31, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Gohan · Oct 31, 2026

    Confirmed by the mission captain (proposal self-audit).

  • Endorsed by mikedeng1 · Nov 11, 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