Documentation

LeanPool.DavisKahan.ForTauCeti.Analysis.InnerProductSpace.DoubleAngle.UnboundedReflection

The reflection block system for an unbounded operator #

Let A be a partial linear map reduced by U, let B be a bounded operator that is odd for the splitting U ⊕ Uᗮ (it exchanges the two summands), and let Z be a bounded operator that commutes with A + B on D(A) and preserves D(A). Writing

this module proves the two facts the unbounded double-angle theory rests on.

Domain control #

C and S preserve D(A). In blocks, with Z = [[D₀, G⋆], [G, -D₁]], this is exactly

D₀ D(A₀) ⊆ D(A₀), G D(A₀) ⊆ D(A₁), G⋆ D(A₁) ⊆ D(A₀), D₁ D(A₁) ⊆ D(A₁),

the compatibility that is awkward to obtain when the four blocks are handled separately. It follows from Z D(A) ⊆ D(A) together with the reducing projections preserving D(A); nothing about the spectrum of A is used.

The Sylvester equation #

On the whole of D(A),

A (S x) - S (A x) = C (B x) - B (C x).

This is the domain-correct, branch-free form of Davis--Kahan equation (7.6). Blockwise, on D(A₀), it reads

A₁ G - G A₀ = -D₁ B - B D₀,

an honest operator identity: every term is defined, by the domain inclusions above. No global |A|, no indefinite closed form, and no sign selection: the sign of cos 2θ is inside C x, and only C appears.

Main results #

References #

def TauCeti.IsOddFor {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] (U : Submodule 𝕜 H) (B : H →L[𝕜] H) :

B exchanges the two summands. This is the Davis--Kahan hypothesis H₀ = H₁ = 0: the perturbation is fully off-diagonal.

Equations
Instances For
    theorem TauCeti.mem_domain_diagonalPart {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] {A : H →ₗ.[𝕜] H} {U : Submodule 𝕜 H} [U.HasOrthogonalProjection] {Z : H →L[𝕜] H} (hred : LinearPMap.ReducesSubspace A U) (hZ : LinearPMap.MapsDomainTo A A Z) (x : ↥A.domain) :
    (U.diagonalPart Z) ↑x ∈ A.domain

    The even block preserves the domain.

    theorem TauCeti.mem_domain_offDiagonalPart {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] {A : H →ₗ.[𝕜] H} {U : Submodule 𝕜 H} [U.HasOrthogonalProjection] {Z : H →L[𝕜] H} (hred : LinearPMap.ReducesSubspace A U) (hZ : LinearPMap.MapsDomainTo A A Z) (x : ↥A.domain) :

    The odd block preserves the domain.

    theorem TauCeti.sylvester_offDiagonalPart_of_mem {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] {A : H →ₗ.[𝕜] H} {U : Submodule 𝕜 H} [U.HasOrthogonalProjection] {B Z : H →L[𝕜] H} (hred : LinearPMap.ReducesSubspace A U) (hB : IsOddFor U B) (hZdom : LinearPMap.MapsDomainTo A A Z) (hZcomm : ∀ (x : ↥A.domain), ↑A ⟨Z ↑x, ⋯⟩ + B (Z ↑x) = Z (↑A x) + Z (B ↑x)) (x : ↥A.domain) (hx : ↑x ∈ U) :
    ↑A ⟨(U.offDiagonalPart Z) ↑x, ⋯⟩ + B ((U.diagonalPart Z) ↑x) = (U.offDiagonalPart Z) (↑A x) + (U.diagonalPart Z) (B ↑x)

    Equation (7.6) for an unbounded operator, on D(A₀).

    For x in the domain and in U,

    A (S x) - S (A x) = C (B x) - B (C x),

    which in blocks is A₁ G x - G A₀ x = -D₁ B x - B D₀ x. Every term is defined: S x ∈ D(A) ∩ Uᗮ and C x ∈ D(A) ∩ U by mem_domain_offDiagonalPart and mem_domain_diagonalPart.

    theorem TauCeti.sylvester_offDiagonalPart_of_mem_orthogonal {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] {A : H →ₗ.[𝕜] H} {U : Submodule 𝕜 H} [U.HasOrthogonalProjection] {B Z : H →L[𝕜] H} (hred : LinearPMap.ReducesSubspace A U) (hB : IsOddFor U B) (hZdom : LinearPMap.MapsDomainTo A A Z) (hZcomm : ∀ (x : ↥A.domain), ↑A ⟨Z ↑x, ⋯⟩ + B (Z ↑x) = Z (↑A x) + Z (B ↑x)) (x : ↥A.domain) (hx : ↑x ∈ Uᗮ) :
    ↑A ⟨(U.offDiagonalPart Z) ↑x, ⋯⟩ + B ((U.diagonalPart Z) ↑x) = (U.offDiagonalPart Z) (↑A x) + (U.diagonalPart Z) (B ↑x)

    Equation (7.6) for an unbounded operator, on D(A₁). The mirror of sylvester_offDiagonalPart_of_mem, with the same conclusion.

    theorem TauCeti.sylvester_offDiagonalPart {𝕜 : Type u_1} {H : Type u_2} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] {A : H →ₗ.[𝕜] H} {U : Submodule 𝕜 H} [U.HasOrthogonalProjection] {B Z : H →L[𝕜] H} (hred : LinearPMap.ReducesSubspace A U) (hB : IsOddFor U B) (hZdom : LinearPMap.MapsDomainTo A A Z) (hZcomm : ∀ (x : ↥A.domain), ↑A ⟨Z ↑x, ⋯⟩ + B (Z ↑x) = Z (↑A x) + Z (B ↑x)) (x : ↥A.domain) :
    ↑A ⟨(U.offDiagonalPart Z) ↑x, ⋯⟩ + B ((U.diagonalPart Z) ↑x) = (U.offDiagonalPart Z) (↑A x) + (U.diagonalPart Z) (B ↑x)

    Equation (7.6) for an unbounded operator, on all of D(A).

    A (S x) - S (A x) = C (B x) - B (C x), with every term defined by the domain inclusions. Branch-free: no sign of cos 2θ is selected anywhere, because the sign is absorbed into C x.