Trace pairing lies between permutation pairings of the spectra
ProvedBhatia.trace_mul_perm_boundsFor two real symmetric matrices with eigenvalues and , the Frobenius pairing is always sandwiched between two pairings of the spectra: there is a permutation whose pairing underestimates the trace, and a permutation whose pairing overestimates it,
Composed with the rearrangement inequality, the two extremes are the reverse-ordered and like-ordered sums, which is the classical statement
for eigenvalues listed in decreasing order — a member of the von Neumann trace-inequality family.
Why this form. Stating the bounds with an existentially quantified permutation avoids having to sort the eigenvalues, which Mathlib's IsHermitian.eigenvalues does not do. The sorted statement follows immediately by applying the rearrangement inequality to each permutation sum.
Proof sketch. Diagonalise both matrices, and . Cyclicity of the trace turns into with , which expands to . Because for two orthonormal eigenbases, the coefficient matrix has all row and column sums equal to by Parseval, so it is doubly stochastic. Birkhoff's theorem writes it as a convex combination of permutation matrices, exhibiting as a convex combination of the sums — and a convex combination always lies between the least and greatest of the values combined.
import Mathlib.Analysis.Matrix.Spectrum open Matrix
theorem Bhatia.trace_mul_perm_bounds {n : ℕ} (A B : Matrix (Fin n) (Fin n) ℝ)
(hA : A.IsHermitian) (hB : B.IsHermitian) :
(∃ σ : Equiv.Perm (Fin n),
∑ i, hA.eigenvalues i * hB.eigenvalues (σ i) ≤ (A * B).trace) ∧
(∃ σ : Equiv.Perm (Fin n),
(A * B).trace ≤ ∑ i, hA.eigenvalues i * hB.eigenvalues (σ i)) := by sorry