The Hermitian dilation of a rectangular operator, and the induced rectangular Mazur map
DefinitionHlawkaSchatten_HermitianDilationFix finite-dimensional complex inner product spaces . For a -linear map , this bundle builds a single self-adjoint ("Hermitian") operator on the Hilbert direct sum that encodes , then uses it to carry the Hermitian spectral constructions of the HermitianSpectral bundle over to rectangular (non-square) operators.
The dilation itself.
hermitianDilation T, on (the Hilbert -sum of and ), sends — the block operator . Internally it transports a plain product-space map, the privateproductOffDiagonal, to the Hilbert-sum type.hermitianBlockDiagonal A B, for and , sends — the block-diagonal operator (transported from the privateproductDiagonal).hermitianSignaturehermitianBlockDiagonal id (-id), the grading involution : on , on .
The bundle proves, briefly, that hermitianDilation T is symmetric, additive, real-homogeneous and injective in ; that it anticommutes with hermitianSignature; and that its square is block-diagonal with the two Gram operators and on the diagonal. It uses this to show that for any , the Hermitian functional calculus of at the square of hermitianDilation T (hermitianDilation_sq_functionalCalculus) block-diagonalizes as the functional calculus of at direct-summed with that at — this is what ties singularValuePowerSum to traces of Hermitian functional calculus (hermitianFunctionalCalculus, HermitianSpectral bundle) of the domain Gram operator.
Dilated Bregman/Mazur quantities. For and :
dilatedBregmanTrace p S TanddilatedMazurDistanceSq p S Tapply the Hermitian-spectral trace-Bregman divergence and trace-Mazur-distance (hermitianBregmanTrace,hermitianMazurDistanceSq, fromHermitianSpectral) to the two dilationshermitianDilation S,hermitianDilation T. Because the square of a dilation is — two copies of the rectangular data, one per summand — each quantity is exactly twice the corresponding rectangular quantity, not merely its analogue: the bundle proves , and, for , .dilatedGradientPairing p S T: a cross term, with the exactly canceling the factor of above. For , and both on the Schatten- power sphere, it equals (dilatedBregmanTrace_div_two_eq_one_sub_pairingin the source module, not part of this bundle).dilatedMazurMap p ThermitianMazurMap p (hermitianDilation T)— the odd Hermitian functional calculus of the scalar Mazur map at the dilation of ; an operator on .
The rectangular Mazur map. Two private helpers, lowerLeftBlock and upperRightBlock, extract the and blocks of an endomorphism of . Then:
rectangularMazurMap p TlowerLeftBlock (dilatedMazurMap p T)— the induced rectangular Mazur map . The bundle proves that for ,rectangularMazurMap pandrectangularMazurMap (4/p)are mutually inverse (rectangularMazurMap_four_div_comp; in particular each is a bijection), and thatrectangularMazurMap pcarries the Schatten- power sum of to the Schatten-2 (Hilbert–Schmidt) power sum of its image (singularValuePowerSum_two_rectangularMazurMap). That it is literally the identity at is proved in the source module (rectangularMazurMap_two), not in this bundle.HasRectangularMazurDilation p T(aProp)dilatedMazurMap p T = hermitianDilation (rectangularMazurMap p T)— the exact compatibility between the two constructions; the bundle proves this always holds (hasRectangularMazurDilation).rectangularMazurPowerSphereEquiv hp, for , is the resulting equivalence between the Schatten- power sphere of operators and the Schatten-2 power sphere, given by .
import Definitions.Def_HlawkaSchatten_HermitianSpectral
import Definitions.Def_HlawkaSchatten_ScalarBregman
import Definitions.Def_HlawkaSchatten_SchattenNorm
import Mathlib.Analysis.Calculus.LHopital
import Mathlib.Analysis.Convex.Deriv
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.InnerProductSpace.Adjoint
import Mathlib.Analysis.InnerProductSpace.Basic
import Mathlib.Analysis.InnerProductSpace.Dual
import Mathlib.Analysis.InnerProductSpace.NormPow
import Mathlib.Analysis.InnerProductSpace.PiL2
import Mathlib.Analysis.InnerProductSpace.ProdL2
import Mathlib.Analysis.InnerProductSpace.SingularValues
import Mathlib.Analysis.InnerProductSpace.Trace
import Mathlib.Data.Sign.Basic
import Mathlib.LinearAlgebra.Matrix.Charpoly.Basic
import Mathlib.Topology.Compactification.OnePoint.Basic
import Mathlib.Topology.Instances.Sign
/-
Copyright (c) 2026 Ezzeri Esa. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Ezzeri Esa
-/
/-!
# Rectangular Hermitian dilation
This file begins the passage from rectangular maps to the Hermitian spectral
comparison by constructing the standard off-diagonal dilation.
-/
namespace HlawkaSchatten
open scoped InnerProductSpace
variable {E F : Type*}
[NormedAddCommGroup E] [InnerProductSpace ℂ E] [FiniteDimensional ℂ E]
[NormedAddCommGroup F] [InnerProductSpace ℂ F] [FiniteDimensional ℂ F]
/-- The off-diagonal Hermitian dilation `(x,y) ↦ (T†y, Tx)` of a rectangular
complex-linear map. -/
private noncomputable def productOffDiagonal (T : E →ₗ[ℂ] F) :
E × F →ₗ[ℂ] E × F :=
LinearMap.prod
(T.adjoint.comp (LinearMap.snd ℂ E F))
(T.comp (LinearMap.fst ℂ E F))
private def productDiagonal (A : E →ₗ[ℂ] E) (B : F →ₗ[ℂ] F) :
E × F →ₗ[ℂ] E × F :=
LinearMap.prod
(A.comp (LinearMap.fst ℂ E F))
(B.comp (LinearMap.snd ℂ E F))
/-- The off-diagonal dilation, transported to the Hilbert `L²` product. -/
noncomputable def hermitianDilation (T : E →ₗ[ℂ] F) :
WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F) :=
(WithLp.linearEquiv 2 ℂ (E × F)).symm.toLinearMap.comp
((productOffDiagonal T).comp
(WithLp.linearEquiv 2 ℂ (E × F)).toLinearMap)
/-- A block-diagonal endomorphism transported to the Hilbert `L²` product. -/
noncomputable def hermitianBlockDiagonal (A : E →ₗ[ℂ] E) (B : F →ₗ[ℂ] F) :
WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F) :=
(WithLp.linearEquiv 2 ℂ (E × F)).symm.toLinearMap.comp
((productDiagonal A B).comp
(WithLp.linearEquiv 2 ℂ (E × F)).toLinearMap)
/-- The grading involution which is positive on the domain summand and
negative on the codomain summand. -/
noncomputable def hermitianSignature :
WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F) :=
hermitianBlockDiagonal LinearMap.id (-LinearMap.id)
@[simp]
theorem hermitianDilation_apply (T : E →ₗ[ℂ] F) (x : E) (y : F) :
hermitianDilation T (WithLp.toLp 2 (x, y)) =
WithLp.toLp 2 (T.adjoint y, T x) := by
rfl
omit [FiniteDimensional ℂ E] [FiniteDimensional ℂ F] in
@[simp]
theorem hermitianBlockDiagonal_apply
(A : E →ₗ[ℂ] E) (B : F →ₗ[ℂ] F) (x : E) (y : F) :
hermitianBlockDiagonal A B (WithLp.toLp 2 (x, y)) =
WithLp.toLp 2 (A x, B y) := by
rfl
omit [FiniteDimensional ℂ E] [FiniteDimensional ℂ F] in
@[simp]
theorem hermitianSignature_apply (x : E) (y : F) :
hermitianSignature (WithLp.toLp 2 (x, y)) =
WithLp.toLp 2 (x, -y) := by
simp [hermitianSignature]
/-- The off-diagonal dilation anticommutes with the product grading. -/
theorem hermitianDilation_comp_signature (T : E →ₗ[ℂ] F) :
(hermitianDilation T).comp hermitianSignature =
-(hermitianSignature.comp (hermitianDilation T)) := by
apply LinearMap.ext
intro z
rw [← WithLp.toLp_ofLp 2 z]
rcases WithLp.ofLp z with ⟨x, y⟩
simp only [LinearMap.comp_apply, hermitianSignature_apply,
hermitianDilation_apply, LinearMap.neg_apply]
rw [← WithLp.toLp_neg]
simp
omit [FiniteDimensional ℂ E] [FiniteDimensional ℂ F] in
/-- Trace of a block-diagonal map is the sum of the block traces. -/
theorem trace_hermitianBlockDiagonal
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(A : E →ₗ[ℂ] E) (B : F →ₗ[ℂ] F) :
(hermitianBlockDiagonal A B).trace ℂ (WithLp 2 (E × F)) =
A.trace ℂ E + B.trace ℂ F := by
rw [LinearMap.trace_eq_sum_inner _ (e.prod f), Fintype.sum_sum_type]
simp only [OrthonormalBasis.prod_apply, Sum.elim_inl, Sum.elim_inr,
Function.comp_apply, LinearMap.inl_apply, LinearMap.inr_apply,
hermitianBlockDiagonal_apply, WithLp.prod_inner_apply,
inner_zero_left, add_zero, zero_add]
rw [← LinearMap.trace_eq_sum_inner A e,
← LinearMap.trace_eq_sum_inner B f]
/-- The rectangular off-diagonal dilation is symmetric. -/
theorem hermitianDilation_isSymmetric (T : E →ₗ[ℂ] F) :
(hermitianDilation T).IsSymmetric := by
intro u v
rw [← WithLp.toLp_ofLp 2 u, ← WithLp.toLp_ofLp 2 v]
rcases WithLp.ofLp u with ⟨x₁, y₁⟩
rcases WithLp.ofLp v with ⟨x₂, y₂⟩
simp only [hermitianDilation_apply, WithLp.prod_inner_apply]
rw [T.adjoint_inner_left, T.adjoint_inner_right]
ring
/-- Hermitian dilation preserves addition. -/
theorem hermitianDilation_add (S T : E →ₗ[ℂ] F) :
hermitianDilation (S + T) = hermitianDilation S + hermitianDilation T := by
apply LinearMap.ext
intro u
rw [← WithLp.toLp_ofLp 2 u]
rcases WithLp.ofLp u with ⟨x, y⟩
simp only [hermitianDilation_apply, map_add, LinearMap.add_apply]
rw [← WithLp.toLp_add]
apply WithLp.toLp_injective 2
rfl
/-- Real scaling commutes with Hermitian dilation. -/
theorem hermitianDilation_real_smul (r : ℝ) (T : E →ₗ[ℂ] F) :
hermitianDilation ((r : ℂ) • T) =
(r : ℂ) • hermitianDilation T := by
apply LinearMap.ext
intro u
rw [← WithLp.toLp_ofLp 2 u]
rcases WithLp.ofLp u with ⟨x, y⟩
simp only [hermitianDilation_apply, LinearMap.smul_apply]
apply WithLp.ofLp_injective 2
simp only [Complex.coe_smul, WithLp.ofLp_smul, Prod.smul_mk,
Prod.mk.injEq, and_true]
apply ext_inner_right ℂ
intro z
rw [LinearMap.adjoint_inner_left]
change ⟪y, (r : ℂ) • T z⟫_ℂ =
⟪(r : ℂ) • T.adjoint y, z⟫_ℂ
rw [inner_smul_right, inner_smul_left]
rw [LinearMap.adjoint_inner_left]
have hr : (starRingEnd ℂ) (r : ℂ) = (r : ℂ) := by
apply Complex.ext <;> simp
rw [hr]
@[simp]
theorem hermitianDilation_neg (T : E →ₗ[ℂ] F) :
hermitianDilation (-T) = -hermitianDilation T := by
simpa using hermitianDilation_real_smul (-1) T
theorem hermitianDilation_sub (S T : E →ₗ[ℂ] F) :
hermitianDilation (S - T) = hermitianDilation S - hermitianDilation T := by
rw [sub_eq_add_neg, hermitianDilation_add, hermitianDilation_neg, sub_eq_add_neg]
@[simp]
theorem hermitianDilation_zero :
hermitianDilation (0 : E →ₗ[ℂ] F) = 0 := by
apply LinearMap.ext
intro u
rw [← WithLp.toLp_ofLp 2 u]
rcases WithLp.ofLp u with ⟨x, y⟩
simp
/-- The off-diagonal dilation remembers the original rectangular map. -/
theorem hermitianDilation_injective :
Function.Injective (hermitianDilation :
(E →ₗ[ℂ] F) → WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F)) := by
intro S T h
apply LinearMap.ext
intro x
have hx := DFunLike.congr_fun h (WithLp.toLp 2 (x, 0))
have hy := congrArg
(fun u : WithLp 2 (E × F) ↦ (WithLp.ofLp u).2) hx
simpa using hy
/-- Squaring the dilation produces the two Gram operators on the diagonal. -/
@[simp]
theorem hermitianDilation_sq_apply (T : E →ₗ[ℂ] F) (x : E) (y : F) :
((hermitianDilation T).comp (hermitianDilation T))
(WithLp.toLp 2 (x, y)) =
WithLp.toLp 2 (T.adjoint (T x), T (T.adjoint y)) := by
rw [LinearMap.comp_apply, hermitianDilation_apply,
hermitianDilation_apply]
/-- Functional calculus of the squared dilation splits into the functional
calculi of its two Gram blocks. -/
theorem hermitianDilation_sq_functionalCalculus
(g : ℝ → ℝ) (T : E →ₗ[ℂ] F) :
hermitianFunctionalCalculus g
((hermitianDilation T).comp (hermitianDilation T))
(isSymmetric_comp_self (hermitianDilation T)
(hermitianDilation_isSymmetric T)) =
hermitianBlockDiagonal
(hermitianFunctionalCalculus g (T.adjoint.comp T)
T.isSymmetric_adjoint_comp_self)
(hermitianFunctionalCalculus g (T.comp T.adjoint)
T.isSymmetric_self_comp_adjoint) := by
let hD := hermitianDilation_isSymmetric T
let hG := T.isSymmetric_adjoint_comp_self
let hK := T.isSymmetric_self_comp_adjoint
apply ((hG.eigenvectorBasis rfl).prod
(hK.eigenvectorBasis rfl)).toBasis.ext
intro ij
rcases ij with i | j
· rw [OrthonormalBasis.coe_toBasis, OrthonormalBasis.prod_apply]
simp only [Sum.elim_inl, Function.comp_apply, LinearMap.inl_apply,
hermitianBlockDiagonal_apply]
have hx :
((hermitianDilation T).comp (hermitianDilation T))
(WithLp.toLp 2 (hG.eigenvectorBasis rfl i, 0)) =
(hG.eigenvalues rfl i : ℂ) •
WithLp.toLp 2 (hG.eigenvectorBasis rfl i, 0) := by
rw [hermitianDilation_sq_apply]
simp only [map_zero]
change WithLp.toLp 2
((T.adjoint.comp T) (hG.eigenvectorBasis rfl i), 0) = _
rw [hG.apply_eigenvectorBasis, ← WithLp.toLp_smul]
apply WithLp.toLp_injective 2
simp
rw [hermitianFunctionalCalculus_apply_of_apply_eq_smul g _ _ _ _ hx,
hermitianFunctionalCalculus_apply_eigenvectorBasis, map_zero]
rw [← WithLp.toLp_smul]
apply WithLp.toLp_injective 2
simp
· rw [OrthonormalBasis.coe_toBasis, OrthonormalBasis.prod_apply]
simp only [Sum.elim_inr, Function.comp_apply, LinearMap.inr_apply,
hermitianBlockDiagonal_apply]
have hx :
((hermitianDilation T).comp (hermitianDilation T))
(WithLp.toLp 2 (0, hK.eigenvectorBasis rfl j)) =
(hK.eigenvalues rfl j : ℂ) •
WithLp.toLp 2 (0, hK.eigenvectorBasis rfl j) := by
rw [hermitianDilation_sq_apply]
simp only [map_zero]
change WithLp.toLp 2
(0, (T.comp T.adjoint) (hK.eigenvectorBasis rfl j)) = _
rw [hK.apply_eigenvectorBasis, ← WithLp.toLp_smul]
apply WithLp.toLp_injective 2
simp
rw [hermitianFunctionalCalculus_apply_of_apply_eq_smul g _ _ _ _ hx,
hermitianFunctionalCalculus_apply_eigenvectorBasis, map_zero]
rw [← WithLp.toLp_smul]
apply WithLp.toLp_injective 2
simp
/-- Trace-level form of the Gram-block functional-calculus split. -/
theorem re_trace_hermitianDilation_sq_functionalCalculus
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(g : ℝ → ℝ) (T : E →ₗ[ℂ] F) :
((hermitianFunctionalCalculus g
((hermitianDilation T).comp (hermitianDilation T))
(isSymmetric_comp_self (hermitianDilation T)
(hermitianDilation_isSymmetric T))).trace ℂ
(WithLp 2 (E × F))).re =
((hermitianFunctionalCalculus g (T.adjoint.comp T)
T.isSymmetric_adjoint_comp_self).trace ℂ E).re +
((hermitianFunctionalCalculus g (T.comp T.adjoint)
T.isSymmetric_self_comp_adjoint).trace ℂ F).re := by
rw [hermitianDilation_sq_functionalCalculus,
trace_hermitianBlockDiagonal e f, Complex.add_re]
/-- The characteristic roots of the two Gram operators agree after padding
the smaller side with zero roots. This is the rectangular `AB`/`BA`
characteristic-polynomial identity in basis-free operator form. -/
theorem gram_charpoly_roots_padded
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(T : E →ₗ[ℂ] F) :
Fintype.card κ • ({0} : Multiset ℂ) +
(T.adjoint.comp T).charpoly.roots =
Fintype.card ι • ({0} : Multiset ℂ) +
(T.comp T.adjoint).charpoly.roots := by
classical
open Polynomial in
have hchar :
X ^ Fintype.card κ * (T.adjoint.comp T).charpoly =
X ^ Fintype.card ι * (T.comp T.adjoint).charpoly := by
rw [← LinearMap.charpoly_toMatrix (T.adjoint.comp T) e.toBasis,
← LinearMap.charpoly_toMatrix (T.comp T.adjoint) f.toBasis,
LinearMap.toMatrix_comp e.toBasis f.toBasis e.toBasis,
LinearMap.toMatrix_comp f.toBasis e.toBasis f.toBasis]
exact Matrix.charpoly_mul_comm'
(LinearMap.toMatrix f.toBasis e.toBasis T.adjoint)
(LinearMap.toMatrix e.toBasis f.toBasis T)
have hroots := congrArg Polynomial.roots hchar
rw [Polynomial.roots_mul (mul_ne_zero
(pow_ne_zero _ Polynomial.X_ne_zero)
(LinearMap.charpoly_monic _).ne_zero),
Polynomial.roots_mul (mul_ne_zero
(pow_ne_zero _ Polynomial.X_ne_zero)
(LinearMap.charpoly_monic _).ne_zero),
Polynomial.roots_X_pow, Polynomial.roots_X_pow] at hroots
exact hroots
/-- Positive real powers have equal total spectral mass on the two Gram
operators, since the only discrepancy in their spectra consists of zeros. -/
theorem sum_gram_eigenvalues_rpow_eq
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
(∑ i : Fin (Module.finrank ℂ E),
T.isSymmetric_adjoint_comp_self.eigenvalues rfl i ^ (p / 2)) =
∑ j : Fin (Module.finrank ℂ F),
T.isSymmetric_self_comp_adjoint.eigenvalues rfl j ^ (p / 2) := by
let hG := T.isSymmetric_adjoint_comp_self
let hK := T.isSymmetric_self_comp_adjoint
have hroots := gram_charpoly_roots_padded (hG.eigenvectorBasis rfl)
(hK.eigenvectorBasis rfl) T
rw [hG.roots_charpoly_eq_eigenvalues rfl,
hK.roots_charpoly_eq_eigenvalues rfl] at hroots
have hsum := congrArg
(fun s : Multiset ℂ ↦ (s.map (fun z ↦ z.re ^ (p / 2))).sum) hroots
have hp2 : 0 < p / 2 := div_pos hp (by norm_num)
rw [← List.sum_ofFn, ← List.sum_ofFn]
simpa [Multiset.map_add, Multiset.map_nsmul, Multiset.sum_add,
Multiset.sum_nsmul, hp2.ne', Function.comp_def] using hsum
/-- The singular-value power sum is the trace of the corresponding positive
functional calculus of the domain Gram operator. -/
theorem singularValuePowerSum_eq_re_trace_gramFunctionalCalculus
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
singularValuePowerSum p T =
((hermitianFunctionalCalculus (fun y ↦ y ^ (p / 2))
(T.adjoint.comp T) T.isSymmetric_adjoint_comp_self).trace ℂ E).re := by
rw [re_trace_hermitianFunctionalCalculus]
have hrank : Module.finrank ℂ T.range ≤ Module.finrank ℂ E :=
T.finrank_range_le
have hsum :
(∑ i ∈ Finset.range (Module.finrank ℂ T.range),
T.singularValues i ^ p) =
∑ i ∈ Finset.range (Module.finrank ℂ E),
T.singularValues i ^ p := by
apply Finset.sum_subset (Finset.range_mono hrank)
intro i hi hi'
have hirank : Module.finrank ℂ T.range ≤ i := by
exact Nat.le_of_not_gt (fun hlt ↦ hi' (Finset.mem_range.mpr hlt))
rw [T.singularValues_eq_zero_iff_le_finrank_range.mpr hirank]
exact Real.zero_rpow hp.ne'
unfold singularValuePowerSum
rw [T.support_singularValues]
calc
(∑ i ∈ Finset.range (Module.finrank ℂ T.range),
T.singularValues i ^ p) =
∑ i ∈ Finset.range (Module.finrank ℂ E),
T.singularValues i ^ p := hsum
_ = ∑ i : Fin (Module.finrank ℂ E), T.singularValues i ^ p :=
(Fin.sum_univ_eq_sum_range
(fun i : ℕ ↦ T.singularValues i ^ p) (Module.finrank ℂ E)).symm
_ = ∑ i : Fin (Module.finrank ℂ E),
T.isSymmetric_adjoint_comp_self.eigenvalues rfl i ^ (p / 2) := by
apply Fintype.sum_congr
intro i
rw [T.singularValues_fin rfl]
exact (Real.rpow_div_two_eq_sqrt p
(T.isPositive_adjoint_comp_self.nonneg_eigenvalues rfl i)).symm
/-- Normalized power-potential trace of the domain Gram block. -/
theorem re_trace_gramPowerPotential_eq_singularValuePowerSum_div
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
((hermitianFunctionalCalculus (fun y ↦ y ^ (p / 2) / p)
(T.adjoint.comp T) T.isSymmetric_adjoint_comp_self).trace ℂ E).re =
singularValuePowerSum p T / p := by
rw [singularValuePowerSum_eq_re_trace_gramFunctionalCalculus hp,
re_trace_hermitianFunctionalCalculus,
re_trace_hermitianFunctionalCalculus, Finset.sum_div]
/-- A rectangular map and its adjoint have the same positive singular-value
power sum. -/
theorem singularValuePowerSum_adjoint
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
singularValuePowerSum p T.adjoint = singularValuePowerSum p T := by
rw [singularValuePowerSum_eq_re_trace_gramFunctionalCalculus hp T.adjoint,
singularValuePowerSum_eq_re_trace_gramFunctionalCalculus hp T,
re_trace_hermitianFunctionalCalculus,
re_trace_hermitianFunctionalCalculus]
simp only [LinearMap.adjoint_adjoint]
exact (sum_gram_eigenvalues_rpow_eq hp T).symm
/-- Exact general-p trace formula for the dilated power potential, before
identifying the singular-value sums of a map and its adjoint. -/
theorem re_trace_dilated_powerPotential_eq_self_add_adjoint
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
((hermitianFunctionalCalculus (powerPotential p) (hermitianDilation T)
(hermitianDilation_isSymmetric T)).trace ℂ
(WithLp 2 (E × F))).re =
(singularValuePowerSum p T + singularValuePowerSum p T.adjoint) / p := by
rw [hermitianFunctionalCalculus_powerPotential_comp_self,
re_trace_hermitianDilation_sq_functionalCalculus e f]
rw [re_trace_gramPowerPotential_eq_singularValuePowerSum_div hp T]
have hAdj :=
re_trace_gramPowerPotential_eq_singularValuePowerSum_div hp T.adjoint
simp only [LinearMap.adjoint_adjoint] at hAdj
rw [hAdj]
ring
/-- The power-potential trace of the Hermitian dilation is exactly twice the
rectangular singular-value power sum, with the potential's normalization. -/
theorem re_trace_dilated_powerPotential_eq_two_mul_singularValuePowerSum
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
((hermitianFunctionalCalculus (powerPotential p) (hermitianDilation T)
(hermitianDilation_isSymmetric T)).trace ℂ
(WithLp 2 (E × F))).re =
2 * singularValuePowerSum p T / p := by
rw [re_trace_dilated_powerPotential_eq_self_add_adjoint
(T.isSymmetric_adjoint_comp_self.eigenvectorBasis rfl)
(T.isSymmetric_self_comp_adjoint.eigenvectorBasis rfl) hp T,
singularValuePowerSum_adjoint hp T]
ring
/-- Positive real scaling raises the rectangular singular-value power sum
to the same exponent. -/
theorem singularValuePowerSum_real_smul
{p c : ℝ} (hp : 0 < p) (hc : 0 < c) (T : E →ₗ[ℂ] F) :
singularValuePowerSum p ((c : ℂ) • T) =
c ^ p * singularValuePowerSum p T := by
have hct := re_trace_dilated_powerPotential_eq_two_mul_singularValuePowerSum
hp ((c : ℂ) • T)
have ht := re_trace_dilated_powerPotential_eq_two_mul_singularValuePowerSum hp T
have hct' :
((hermitianFunctionalCalculus (powerPotential p)
((c : ℂ) • hermitianDilation T)
((hermitianDilation_isSymmetric T).smul (by simp))).trace ℂ
(WithLp 2 (E × F))).re =
2 * singularValuePowerSum p ((c : ℂ) • T) / p := by
simpa only [hermitianDilation_real_smul] using hct
have hs := re_trace_powerPotential_real_smul (p := p) hc
(hermitianDilation T) (hermitianDilation_isSymmetric T)
rw [hs, ht] at hct'
field_simp [hp.ne'] at hct'
linarith
/-- Positive real homogeneity of the finite-dimensional Schatten quantity. -/
theorem schattenPNorm_real_smul
{p c : ℝ} (hp : 0 < p) (hc : 0 < c) (T : E →ₗ[ℂ] F) :
schattenPNorm p ((c : ℂ) • T) = c * schattenPNorm p T := by
unfold schattenPNorm
rw [singularValuePowerSum_real_smul hp hc]
rw [Real.mul_rpow (Real.rpow_nonneg hc.le p)
(singularValuePowerSum_nonneg p T)]
rw [← Real.rpow_mul hc.le]
rw [show p * (1 / p) = 1 by field_simp [hp.ne'], Real.rpow_one]
/-- Odd functional calculus of the dilation still anticommutes with the
product grading. -/
theorem hermitianFunctionalCalculus_odd_comp_signature
(g : ℝ → ℝ) (hodd : ∀ x, g (-x) = -g x) (T : E →ₗ[ℂ] F) :
(hermitianFunctionalCalculus g (hermitianDilation T)
(hermitianDilation_isSymmetric T)).comp hermitianSignature =
-(hermitianSignature.comp
(hermitianFunctionalCalculus g (hermitianDilation T)
(hermitianDilation_isSymmetric T))) := by
let hD := hermitianDilation_isSymmetric T
apply (hD.eigenvectorBasis rfl).toBasis.ext
intro i
rw [OrthonormalBasis.coe_toBasis]
have hJi :
hermitianDilation T (hermitianSignature (hD.eigenvectorBasis rfl i)) =
((-hD.eigenvalues rfl i : ℝ) : ℂ) •
hermitianSignature (hD.eigenvectorBasis rfl i) := by
calc
hermitianDilation T (hermitianSignature (hD.eigenvectorBasis rfl i)) =
((hermitianDilation T).comp hermitianSignature)
(hD.eigenvectorBasis rfl i) := rfl
_ = (-(hermitianSignature.comp (hermitianDilation T)))
(hD.eigenvectorBasis rfl i) := by
rw [hermitianDilation_comp_signature]
_ = ((-hD.eigenvalues rfl i : ℝ) : ℂ) •
hermitianSignature (hD.eigenvectorBasis rfl i) := by
rw [LinearMap.neg_apply, LinearMap.comp_apply,
hD.apply_eigenvectorBasis, map_smul]
simp
rw [LinearMap.comp_apply,
hermitianFunctionalCalculus_apply_of_apply_eq_smul g _ _ _
(-hD.eigenvalues rfl i) hJi,
LinearMap.neg_apply, LinearMap.comp_apply,
hermitianFunctionalCalculus_apply_eigenvectorBasis, map_smul]
rw [hodd]
simp
/-- The trace of the squared dilation is twice the trace of the domain Gram
operator. This is the Hilbert-point instance of the two-copy singular-value
normalization. -/
theorem trace_hermitianDilation_sq
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(T : E →ₗ[ℂ] F) :
(((hermitianDilation T).comp (hermitianDilation T)).trace ℂ
(WithLp 2 (E × F))) =
2 * ((T.adjoint.comp T).trace ℂ E) := by
rw [LinearMap.trace_eq_sum_inner _ (e.prod f), Fintype.sum_sum_type]
simp only [OrthonormalBasis.prod_apply, Sum.elim_inl, Sum.elim_inr,
Function.comp_apply, LinearMap.inl_apply, LinearMap.inr_apply,
hermitianDilation_sq_apply, WithLp.prod_inner_apply,
inner_zero_left, add_zero, zero_add]
change (∑ i, ⟪e i, (T.adjoint.comp T) (e i)⟫_ℂ) +
(∑ j, ⟪f j, (T.comp T.adjoint) (f j)⟫_ℂ) =
2 * ((T.adjoint.comp T).trace ℂ E)
rw [← LinearMap.trace_eq_sum_inner (T.adjoint.comp T) e]
rw [← LinearMap.trace_eq_sum_inner (T.comp T.adjoint) f]
rw [LinearMap.trace_comp_comm' T T.adjoint]
ring
/-- The squared dilation has exactly twice the Schatten-2 power sum. -/
theorem re_trace_hermitianDilation_sq_eq_two_mul_singularValuePowerSum
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(T : E →ₗ[ℂ] F) :
((((hermitianDilation T).comp (hermitianDilation T)).trace ℂ
(WithLp 2 (E × F))).re) = 2 * singularValuePowerSum 2 T := by
rw [trace_hermitianDilation_sq e f T, Complex.mul_re]
norm_num
rw [singularValuePowerSum_two_eq_re_trace]
exact RCLike.re_to_complex
/-- Hermitian Bregman trace quantity after rectangular dilation. -/
noncomputable def dilatedBregmanTrace
(p : ℝ) (S T : E →ₗ[ℂ] F) : ℝ :=
hermitianBregmanTrace p (hermitianDilation S) (hermitianDilation T)
(hermitianDilation_isSymmetric S) (hermitianDilation_isSymmetric T)
/-- The real trace pairing between a rectangular map and the dilated power
gradient of a second map, normalized by the two dilation copies. -/
noncomputable def dilatedGradientPairing
(p : ℝ) (S T : E →ₗ[ℂ] F) : ℝ :=
(((hermitianDilation S).comp
(hermitianFunctionalCalculus (powerGradient p) (hermitianDilation T)
(hermitianDilation_isSymmetric T))).trace ℂ
(WithLp 2 (E × F))).re / 2
/-- Hermitian Mazur Hilbert--Schmidt square after rectangular dilation. -/
noncomputable def dilatedMazurDistanceSq
(p : ℝ) (S T : E →ₗ[ℂ] F) : ℝ :=
hermitianMazurDistanceSq p (hermitianDilation S) (hermitianDilation T)
(hermitianDilation_isSymmetric S) (hermitianDilation_isSymmetric T)
/-- The Hermitian Mazur map of a rectangular operator's dilation. -/
noncomputable def dilatedMazurMap (p : ℝ) (T : E →ₗ[ℂ] F) :
WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F) :=
hermitianMazurMap p (hermitianDilation T) (hermitianDilation_isSymmetric T)
/-- Extract the lower-left rectangular block of an endomorphism of the
Hilbert `L²` product. -/
private noncomputable def lowerLeftBlock
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F)) : E →ₗ[ℂ] F :=
(LinearMap.snd ℂ E F).comp
((WithLp.linearEquiv 2 ℂ (E × F)).toLinearMap.comp
(A.comp
((WithLp.linearEquiv 2 ℂ (E × F)).symm.toLinearMap.comp
(LinearMap.inl ℂ E F))))
omit [FiniteDimensional ℂ E] [FiniteDimensional ℂ F] in
@[simp]
private theorem lowerLeftBlock_apply
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F)) (x : E) :
lowerLeftBlock A x = (WithLp.ofLp (A (WithLp.toLp 2 (x, 0)))).2 := by
rfl
/-- Extract the upper-right rectangular block of an endomorphism of the
Hilbert `L²` product. -/
private noncomputable def upperRightBlock
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F)) : F →ₗ[ℂ] E :=
(LinearMap.fst ℂ E F).comp
((WithLp.linearEquiv 2 ℂ (E × F)).toLinearMap.comp
(A.comp
((WithLp.linearEquiv 2 ℂ (E × F)).symm.toLinearMap.comp
(LinearMap.inr ℂ E F))))
omit [FiniteDimensional ℂ E] [FiniteDimensional ℂ F] in
@[simp]
private theorem upperRightBlock_apply
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F)) (y : F) :
upperRightBlock A y =
(WithLp.ofLp (A (WithLp.toLp 2 (0, y)))).1 := by
rfl
/-- The two off-diagonal blocks of a symmetric product endomorphism are
adjoints of one another. -/
private theorem upperRightBlock_eq_adjoint_lowerLeftBlock
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F))
(hA : A.IsSymmetric) :
upperRightBlock A = (lowerLeftBlock A).adjoint := by
rw [LinearMap.eq_adjoint_iff]
intro y x
have h := hA (WithLp.toLp 2 (0, y)) (WithLp.toLp 2 (x, 0))
rw [← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (x, 0))),
← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (0, y)))] at h
simpa only [upperRightBlock_apply, lowerLeftBlock_apply,
WithLp.prod_inner_apply, inner_zero_left, inner_zero_right,
zero_add, add_zero] using h
private theorem eq_zero_of_eq_neg_complex_module
{G : Type*} [AddCommGroup G] [Module ℂ G]
{x : G} (h : x = -x) : x = 0 := by
have hz : x + x = 0 := eq_neg_iff_add_eq_zero.mp h
have hs := congrArg (fun z : G ↦ (1 / 2 : ℂ) • z) hz
simpa [smul_add, ← add_smul] using hs
/-- A symmetric product endomorphism which anticommutes with the grading is
the Hermitian dilation of its lower-left block. -/
private theorem eq_hermitianDilation_lowerLeftBlock_of_anticommute
(A : WithLp 2 (E × F) →ₗ[ℂ] WithLp 2 (E × F))
(hA : A.IsSymmetric)
(hanti : A.comp hermitianSignature = -(hermitianSignature.comp A)) :
A = hermitianDilation (lowerLeftBlock A) := by
have hupper : upperRightBlock A = (lowerLeftBlock A).adjoint :=
upperRightBlock_eq_adjoint_lowerLeftBlock A hA
have hAx : ∀ x : E,
A (WithLp.toLp 2 (x, 0)) =
WithLp.toLp 2 (0, lowerLeftBlock A x) := by
intro x
have hu := LinearMap.congr_fun hanti (WithLp.toLp 2 (x, 0))
simp only [LinearMap.comp_apply, hermitianSignature_apply,
neg_zero, LinearMap.neg_apply] at hu
rw [← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (x, 0))),
hermitianSignature_apply, ← WithLp.toLp_neg] at hu
have hp := congrArg WithLp.ofLp hu
simp only [Prod.neg_mk, neg_neg] at hp
have hfirst : (WithLp.ofLp (A (WithLp.toLp 2 (x, 0)))).1 =
-(WithLp.ofLp (A (WithLp.toLp 2 (x, 0)))).1 :=
congrArg Prod.fst hp
rw [← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (x, 0)))]
apply congrArg (WithLp.toLp 2)
apply Prod.ext
· simpa using eq_zero_of_eq_neg_complex_module hfirst
· rfl
have hAy : ∀ y : F,
A (WithLp.toLp 2 (0, y)) =
WithLp.toLp 2 (upperRightBlock A y, 0) := by
intro y
have hu := LinearMap.congr_fun hanti (WithLp.toLp 2 (0, y))
simp only [LinearMap.comp_apply, hermitianSignature_apply,
LinearMap.neg_apply] at hu
rw [show WithLp.toLp 2 ((0 : E), -y) =
-WithLp.toLp 2 ((0 : E), y) by
rw [← WithLp.toLp_neg]
congr 1
simp,
map_neg, ← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (0, y))),
hermitianSignature_apply, ← WithLp.toLp_neg] at hu
have hp := congrArg WithLp.ofLp hu
simp only [WithLp.ofLp_neg, Prod.neg_mk, neg_neg] at hp
have hsecond : -(WithLp.ofLp (A (WithLp.toLp 2 (0, y)))).2 =
(WithLp.ofLp (A (WithLp.toLp 2 (0, y)))).2 :=
congrArg Prod.snd hp
rw [← WithLp.toLp_ofLp 2 (A (WithLp.toLp 2 (0, y)))]
apply congrArg (WithLp.toLp 2)
apply Prod.ext
· rfl
· exact eq_zero_of_eq_neg_complex_module hsecond.symm
apply LinearMap.ext
intro z
rw [← WithLp.toLp_ofLp 2 z]
rcases WithLp.ofLp z with ⟨x, y⟩
rw [show WithLp.toLp 2 (x, y) =
WithLp.toLp 2 (x, 0) + WithLp.toLp 2 (0, y) by
rw [← WithLp.toLp_add]
apply WithLp.toLp_injective 2
simp,
map_add, hAx, hAy, hermitianDilation_apply, hupper]
rw [← WithLp.toLp_add]
simp
/-- The rectangular Mazur map obtained as the lower-left block of the odd
spectral map of the Hermitian dilation. -/
noncomputable def rectangularMazurMap (p : ℝ) (T : E →ₗ[ℂ] F) : E →ₗ[ℂ] F :=
lowerLeftBlock (dilatedMazurMap p T)
@[simp]
theorem dilatedMazurMap_zero (p : ℝ) :
dilatedMazurMap p (0 : E →ₗ[ℂ] F) = 0 := by
unfold dilatedMazurMap hermitianMazurMap
simpa only [hermitianDilation_zero] using
(hermitianFunctionalCalculus_zero (E := WithLp 2 (E × F))
(scalarMazur p) (scalarMazur_zero p))
@[simp]
theorem rectangularMazurMap_zero (p : ℝ) :
rectangularMazurMap p (0 : E →ₗ[ℂ] F) = 0 := by
apply LinearMap.ext
intro x
simp [rectangularMazurMap]
/-- The exact off-diagonal form needed to identify the Hermitian spectral
Mazur map with a rectangular one. The lower-left block is pinned by
`rectangularMazurMap`; the content is that no diagonal blocks remain and the
upper-right block is its adjoint. -/
def HasRectangularMazurDilation (p : ℝ) (T : E →ₗ[ℂ] F) : Prop :=
dilatedMazurMap p T = hermitianDilation (rectangularMazurMap p T)
/-- The odd scalar Mazur functional calculus of a Hermitian dilation is the
Hermitian dilation of its extracted rectangular block. -/
theorem hasRectangularMazurDilation (p : ℝ) (T : E →ₗ[ℂ] F) :
HasRectangularMazurDilation p T := by
unfold HasRectangularMazurDilation rectangularMazurMap
apply eq_hermitianDilation_lowerLeftBlock_of_anticommute
· exact hermitianMazurMap_isSymmetric p (hermitianDilation T)
(hermitianDilation_isSymmetric T)
· exact hermitianFunctionalCalculus_odd_comp_signature
(scalarMazur p) (scalarMazur_neg p) T
/-- Rectangular Mazur maps with exponents `p` and `4 / p` are inverse. -/
theorem rectangularMazurMap_four_div_comp
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
rectangularMazurMap (4 / p) (rectangularMazurMap p T) = T := by
apply hermitianDilation_injective
rw [← hasRectangularMazurDilation (4 / p) (rectangularMazurMap p T)]
unfold dilatedMazurMap
have hpD : hermitianDilation (rectangularMazurMap p T) =
dilatedMazurMap p T := (hasRectangularMazurDilation p T).symm
simp only [hpD]
unfold dilatedMazurMap
exact hermitianMazurMap_four_div_comp hp (hermitianDilation T)
(hermitianDilation_isSymmetric T)
/-- The dilated Mazur square is exactly the real trace-square distance
between the corresponding dilated Mazur maps. -/
theorem dilatedMazurDistanceSq_eq_re_trace_sq_sub
(p : ℝ) (S T : E →ₗ[ℂ] F) :
dilatedMazurDistanceSq p S T =
((((dilatedMazurMap p S - dilatedMazurMap p T).comp
(dilatedMazurMap p S - dilatedMazurMap p T)).trace ℂ
(WithLp 2 (E × F))).re) :=
hermitianMazurDistanceSq_eq_re_trace_sq_sub p
(hermitianDilation S) (hermitianDilation T)
(hermitianDilation_isSymmetric S) (hermitianDilation_isSymmetric T)
/-- Once the odd functional calculus is known to retain off-diagonal form,
the dilated Mazur distance has exactly the expected two-copy rectangular
Hilbert--Schmidt normalization. -/
theorem dilatedMazurDistanceSq_eq_two_mul_singularValuePowerSum
{ι κ : Type*} [Fintype ι] [Fintype κ]
(e : OrthonormalBasis ι ℂ E) (f : OrthonormalBasis κ ℂ F)
(p : ℝ) (S T : E →ₗ[ℂ] F) :
dilatedMazurDistanceSq p S T =
2 * singularValuePowerSum 2
(rectangularMazurMap p S - rectangularMazurMap p T) := by
rw [dilatedMazurDistanceSq_eq_re_trace_sq_sub,
hasRectangularMazurDilation p S, hasRectangularMazurDilation p T,
← hermitianDilation_sub]
exact re_trace_hermitianDilation_sq_eq_two_mul_singularValuePowerSum e f _
/-- Against zero, the dilated Mazur square is exactly twice the rectangular
Schatten-`p` power sum. -/
theorem dilatedMazurDistanceSq_zero_eq_two_mul_singularValuePowerSum
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
dilatedMazurDistanceSq p T 0 = 2 * singularValuePowerSum p T := by
unfold dilatedMazurDistanceSq
calc
hermitianMazurDistanceSq p (hermitianDilation T) (hermitianDilation 0)
(hermitianDilation_isSymmetric T) (hermitianDilation_isSymmetric 0) =
p * ((hermitianFunctionalCalculus (powerPotential p)
(hermitianDilation T) (hermitianDilation_isSymmetric T)).trace ℂ
(WithLp 2 (E × F))).re := by
simpa only [hermitianDilation_zero] using
hermitianMazurDistanceSq_zero_eq_powerPotentialTrace hp
(hermitianDilation T) (hermitianDilation_isSymmetric T)
_ = 2 * singularValuePowerSum p T := by
rw [re_trace_dilated_powerPotential_eq_two_mul_singularValuePowerSum hp T]
field_simp
/-- The rectangular Mazur map sends the Schatten-`p` power sum exactly to
the Hilbert--Schmidt power sum. -/
theorem singularValuePowerSum_two_rectangularMazurMap
{p : ℝ} (hp : 0 < p) (T : E →ₗ[ℂ] F) :
singularValuePowerSum 2 (rectangularMazurMap p T) =
singularValuePowerSum p T := by
have htwo :=
dilatedMazurDistanceSq_eq_two_mul_singularValuePowerSum
(T.isSymmetric_adjoint_comp_self.eigenvectorBasis rfl)
(T.isSymmetric_self_comp_adjoint.eigenvectorBasis rfl) p T 0
rw [rectangularMazurMap_zero, sub_zero,
dilatedMazurDistanceSq_zero_eq_two_mul_singularValuePowerSum hp] at htwo
linarith
/-- For every positive `p`, the rectangular Mazur map is an exact
equivalence from the Schatten-`p` power sphere to the Hilbert--Schmidt
power sphere. Its inverse is the Mazur map with exponent `4 / p`. -/
noncomputable def rectangularMazurPowerSphereEquiv
{p : ℝ} (hp : 0 < p) :
schattenPowerSphere (𝕜 := ℂ) (E := E) (F := F) p ≃
schattenPowerSphere (𝕜 := ℂ) (E := E) (F := F) 2 where
toFun T := ⟨rectangularMazurMap p T.1, by
exact (singularValuePowerSum_two_rectangularMazurMap hp T.1).trans T.property⟩
invFun S := ⟨rectangularMazurMap (4 / p) S.1, by
have hnorm := singularValuePowerSum_two_rectangularMazurMap hp
(rectangularMazurMap (4 / p) S.1)
have hcomp := rectangularMazurMap_four_div_comp
(p := 4 / p) (div_pos (by norm_num) hp) S.1
have hquot : 4 / (4 / p) = p := by field_simp
simp only [hquot] at hcomp
rw [hcomp] at hnorm
exact hnorm.symm.trans S.property⟩
left_inv T := Subtype.ext (rectangularMazurMap_four_div_comp hp T.1)
right_inv S := by
apply Subtype.ext
have h := rectangularMazurMap_four_div_comp
(p := 4 / p) (div_pos (by norm_num) hp) S.1
simpa only [show 4 / (4 / p) = p by field_simp] using h
end HlawkaSchatten
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.