Unitarily invariant bounds for the Sylvester equation #
This file bounds solutions of the Sylvester equation A X - X B = C in unitarily invariant
seminorms, for operators on finite-dimensional inner product spaces.
The Frobenius bound #
Let A and B be symmetric operators on finite-dimensional inner product spaces E and F, with
orthonormal eigenbases (uᵢ), (vⱼ) and eigenvalues (αᵢ), (βⱼ). The Sylvester operator
X ↦ A X - X B multiplies the matrix coefficient Xᵢⱼ = ⟪uᵢ, X vⱼ⟫ by αᵢ - βⱼ, so the Frobenius
seminorm of A X - X B is
‖A X - X B‖_F² = ∑ᵢ ∑ⱼ (αᵢ - βⱼ)² ‖Xᵢⱼ‖².
If the spectra of A and B are at distance at least δ, every factor (αᵢ - βⱼ)² is at least
δ², and therefore δ ‖X‖_F ≤ ‖A X - X B‖_F. Equivalently, every solution of the Sylvester
equation A X - X B = C satisfies δ ‖X‖_F ≤ ‖C‖_F. For the Frobenius seminorm the constant is
1; the corresponding bound for a general unitarily invariant seminorm under the same pairwise
separation hypothesis carries the constant π / 2 (see
TauCeti.Analysis.InnerProductSpace.UnitarilyInvariantSeminorm.Sylvester.Pairwise).
The interval/exterior bound #
If the spectrum of A lies in the interval [a, b] and that of B outside (a - δ, b + δ), then
δ N(X) ≤ N(A X - X B) for every unitarily invariant seminorm N, again with constant 1; the
same holds with the roles of A and B exchanged. With centre c = (a + b) / 2 and radius
r = (b - a) / 2, for symmetric A and B the hypothesis gives ‖(A - c) x‖ ≤ r ‖x‖ and
(r + δ) ‖y‖ ≤ ‖(B - c) y‖. Since
A X - X B = (A - c) X - X (B - c), the ideal inequality N((A - c) X) ≤ r N(X) and its reverse
(r + δ) N(X) ≤ N(X (B - c)) give (r + δ) N(X) ≤ r N(X) + N(A X - X B). This argument needs
neither symmetry nor a real centre, so it is stated first for an arbitrary centre c : 𝕜 and
norm bounds of this shape.
The spectral hypotheses may also be restricted to subspaces: if S is invariant under A, T
invariant under B, X takes values in S and vanishes on Tᗮ, and the eigenvalues of A on
S lie in [a, b] while those of B on T lie outside (a - δ, b + δ), the same bound holds.
On the left only the values of X matter; on the right the symmetry of B lets the bounds along
T pass through the adjoint of X. This is the form consumed by the sin Θ theorems, where X
is a product of orthogonal projections.
Main statements #
ContinuousLinearMap.sq_frobenius_sylvester_apply: the exact formula for‖A X - X B‖_F²in orthonormal eigenbases ofAandB.ContinuousLinearMap.mul_frobenius_le_frobenius_sylvester_apply: if the point spectra ofAandBare at distance at leastδ, thenδ ‖X‖_F ≤ ‖A X - X B‖_F.
In the namespace TauCeti.UnitarilyInvariantSeminorm:
mul_map_le_map_sylvester_apply_of_norm_sub_smul_le_of_le_norm_sub_smul,mul_map_le_map_sylvester_apply_of_le_norm_sub_smul_of_norm_sub_smul_le: ifA - cis bounded byrandB - cbounded below byr + δ, or the other way round, thenδ N(X) ≤ N(A X - X B).mul_map_le_map_sylvester_apply_of_pointIntervalExteriorGap,mul_map_le_map_sylvester_apply_of_pointIntervalExteriorGap_swap: under an interval/exterior gap between the spectra of symmetricAandB, in either orientation,δ N(X) ≤ N(A X - X B).mul_map_le_map_sylvester_apply_of_pointIntervalExteriorGap_of_mem,mul_map_le_map_sylvester_apply_of_pointIntervalExteriorGap_swap_of_mem: the same bound with the gap hypothesis restricted to invariant subspacesSandT, forXtaking values inSand vanishing onTᗮ.
References #
- R. Bhatia, Matrix Analysis, Graduate Texts in Mathematics 169, Springer (1997), §VII.2.
- C. Davis, W. M. Kahan, The rotation of eigenvectors by a perturbation. III, SIAM J. Numer. Anal. 7 (1970), 1–46.
- R. Bhatia, C. Davis, A. McIntosh, Perturbation of spectral subspaces and solution of linear operator equations, Linear Algebra Appl. 52/53 (1983), 45–67.
The Frobenius seminorm of A X - X B in eigenbases. For symmetric A and B with
ordered orthonormal eigenbases (uᵢ), (vⱼ) and eigenvalues (αᵢ), (βⱼ),
‖A X - X B‖_F² = ∑ᵢ ∑ⱼ (αᵢ - βⱼ)² ‖⟪uᵢ, X vⱼ⟫‖².
The Frobenius Sylvester bound. In finite dimension, if A and B are symmetric and every
eigenvalue of A is at distance at least δ from every eigenvalue of B, then
δ ‖X‖_F ≤ ‖A X - X B‖_F for every X. In particular every solution of the Sylvester equation
A X - X B = C satisfies δ ‖X‖_F ≤ ‖C‖_F.
Sylvester bound from a disc and the exterior of a larger disc. If ‖A x - c x‖ ≤ r ‖x‖
for every x and (r + δ) ‖y‖ ≤ ‖B y - c y‖ for every y, then δ N(X) ≤ N(A X - X B) for
every unitarily invariant seminorm N.
Sylvester bound from the exterior of a disc and a smaller disc. If
(r + δ) ‖x‖ ≤ ‖A x - c x‖ for every x and ‖B y - c y‖ ≤ r ‖y‖ for every y, then
δ N(X) ≤ N(A X - X B) for every unitarily invariant seminorm N.
The interval/exterior Sylvester bound on invariant subspaces. Let A and B be
symmetric operators on finite-dimensional inner product spaces, let S be invariant under A
and T invariant under B, and let X take values in S and vanish on Tᗮ. If the
eigenvalues of A on S lie in [a, b] and those of B on T outside (a - δ, b + δ), then
δ N(X) ≤ N(A X - X B) for every unitarily invariant seminorm N.
The interval/exterior Sylvester bound on invariant subspaces, reversed orientation. Let
A and B be symmetric operators on finite-dimensional inner product spaces, let S be
invariant under A and T invariant under B, and let X take values in S and vanish on
Tᗮ. If the eigenvalues of B on T lie in [a, b] and those of A on S outside
(a - δ, b + δ), then δ N(X) ≤ N(A X - X B) for every unitarily invariant seminorm N.
The interval/exterior Sylvester bound. Let A and B be symmetric operators on
finite-dimensional inner product spaces. If the eigenvalues of A lie in [a, b] and those of B
outside (a - δ, b + δ), then δ N(X) ≤ N(A X - X B) for every unitarily invariant seminorm N.
In particular every solution of the Sylvester equation A X - X B = C satisfies
δ N(X) ≤ N(C).
The interval/exterior Sylvester bound, reversed orientation. Let A and B be symmetric
operators on finite-dimensional inner product spaces. If the eigenvalues of B lie in [a, b]
and those of A outside (a - δ, b + δ), then δ N(X) ≤ N(A X - X B) for every unitarily
invariant seminorm N. In particular every solution of the Sylvester equation A X - X B = C
satisfies δ N(X) ≤ N(C).