Remark 6.2
OpenQAlgorithms.Papers.Independent.Support50.P12.remark_6_2Let be positive semidefinite operators on a nonzero finite-dimensional complex Hilbert space. Let be continuous on with signed extended recession slope . Write for the paper's quantum divergence, , and for the support projection. All powers vanish on the kernel, including exponent zero. Traces are representation traces. Write for the unnormalized Petz Rényi entropy, , , and . Logarithms are natural and . The Petz map is .
Put for and for , and . Then is convex and decreasing in , , , and
For nonorthogonal pairs, is real convex and is lower semicontinuous and convex. Its derivative is , its right derivative at zero is , and its derivative tends to at infinity. If , ; if , . Bipolarity holds on , and the displayed supremum can be restricted to the closed interval , including a singleton. If , increases on , , and . When , . No derivative is assigned to the identically orthogonal case. These identities relate the Rényi and Hoeffding families.
import Definitions.Def_QAlgorithms_Papers_Independent_Support50_P12
noncomputable section
namespace QAlgorithms.Papers.Independent.Support50.P12
open Filter MeasureTheory Set
open QAlgorithms.Papers.Independent.Information100.P60 (Mat copies)
open QAlgorithms.Papers.Independent.Information100.P30 (MatrixMap)
open QAlgorithms.Papers.Independent.Dependency50.P28 (tr LE logNonnegativeExtended pinch)
open QAlgorithms.Papers.Independent.Dependency50.P21 (channelAdjoint spectralFunction petz relativeEntropy)
open QAlgorithms.Papers.Independent.Frontier18.P17 (hs)
open QAlgorithms.Papers.Independent.Next46.P23 (Supports)
open QAlgorithms.Papers.Independent.Support21.P03 (ExtendedConvex)
open QAlgorithms (spectralProjector)
open QAlgorithms.Learning (IsDistribution)
open QAlgorithms.Papers.Independent.Information100.P04 (absOp)
open scoped BigOperators Matrix ComplexConjugate ComplexOrder ENNReal Topology Kronecker
attribute [local instance] Classical.propDecidable Classical.decEq
open QAlgorithms.Papers.Independent.Information100.P71 (supportPower)
variable {I J K : Type} [Fintype I] [DecidableEq I] [Fintype J] [DecidableEq J]
[Fintype K] [DecidableEq K]
/-- Remark 6.2, PDF pp.27–28: polar duality, restricted CLOSED interval, derivatives
on nonorthogonal pairs, and the trace-guarded order-one and right-rate limits.
The separately deferred unnormalized endpoint-entropy derivative formulas are omitted. -/
theorem remark_6_2 [Nonempty I] (X Y : Mat I) (hX : X.PosSemidef) (hY : Y.PosSemidef) :
(∀ r : ℝ, hoeffding X Y r = polar (logOverlapPerspective X Y) (-r)) ∧
(∀ r s : ℝ, r ≤ s → hoeffding X Y s ≤ hoeffding X Y r) ∧
ExtendedConvex (hoeffding X Y) ∧
logOverlap X Y 0 = -petzRenyi 0 X Y ∧ logOverlap X Y 1 = -petzRenyi 0 Y X ∧
(∀ a ∈ Ico (0 : ℝ) 1, -petzRenyi a X Y =
sSup {v | ∃ r : ℝ, v = ((-r * a / (1 - a) : ℝ) : EReal) - hoeffding X Y r}) ∧
(0 < tr (supp X * supp Y) →
ConvexOn ℝ univ (psiReal X Y) ∧ ExtendedConvex (logOverlapPerspective X Y) ∧
LowerSemicontinuous (logOverlapPerspective X Y) ∧
(∀ s > 0, HasDerivAt (fun u => (logOverlapPerspective X Y u).toReal)
(psiReal X Y (s / (1 + s)) + deriv (psiReal X Y) (s / (1 + s)) / (1 + s)) s) ∧
HasDerivWithinAt (fun u => (logOverlapPerspective X Y u).toReal)
(psiReal X Y 0 + deriv (psiReal X Y) 0) (Ici 0) 0 ∧
Tendsto (fun s : ℝ => deriv (fun u => (logOverlapPerspective X Y u).toReal) s)
atTop (𝓝 (psiReal X Y 1)) ∧
(∀ r : ℝ, -r < psiReal X Y 0 + deriv (psiReal X Y) 0 →
hoeffding X Y r = -logOverlap X Y 0) ∧
(∀ r : ℝ, psiReal X Y 1 < -r → hoeffding X Y r = ⊤) ∧
(∀ s ≥ 0, logOverlapPerspective X Y s = polar (polar (logOverlapPerspective X Y)) s) ∧
(∀ a ∈ Ico (0 : ℝ) 1, -petzRenyi a X Y =
sSup {v | ∃ r : ℝ, -psiReal X Y 1 ≤ r ∧
r ≤ -psiReal X Y 0 - deriv (psiReal X Y) 0 ∧
v = ((-r * a / (1 - a) : ℝ) : EReal) - hoeffding X Y r})) ∧
(tr X ≤ 1 →
MonotoneOn (fun a => petzRenyi a X Y) (Ico 0 1) ∧
Tendsto (fun a => petzRenyi a X Y) (𝓝[<] 1) (𝓝 (hoeffding X Y 0)) ∧
Tendsto (hoeffding X Y) (𝓝[>] 0) (𝓝 (hoeffding X Y 0)) ∧
hoeffding X Y 0 = renyiOne X Y) ∧
(tr X = 1 → renyiOne X Y = relativeEntropy X Y) := by sorry
end QAlgorithms.Papers.Independent.Support50.P12
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let be a finite, nonempty index type with decidable equality, and let be positive semidefinite complex matrices indexed by , without any trace normalization assumption. Write , and write for spectral functional calculus with the function that is when and when ; in particular, is the support projection, including when . Define for and for , and set
where sends either infinite extended-real value to , and all extended-real arithmetic uses its total conventions. Define
Then for every real ; whenever ; and is extended convex in the precise sense that, whenever and for real , one has for every . Moreover, , , and, for every ,
If , then is convex on all of , is extended convex in the same finite-upper-bound sense and is lower semicontinuous, its real-valued conversion has derivative
has derivative within at equal to , and its derivatives converge to as ; also,
for every , and the displayed supremum for can be restricted to
If , including , then is nondecreasing on , as , as , and , where is the definition’s totalized left-limit value of . Finally, if , then equals when , and equals otherwise, with matrix logarithms interpreted by spectral functional calculus.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.