Theorem D.2 — dimension-free spectral modulus:
OpenSpectralFeedback.Supplement.theorem_D_2_spectral_moduluslinear-algebramatrix-analysisoptimizationp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1quantum-information
Let and extend to all positive matrices by (C.6), without trace normalization.
- If are complex matrices, and , then
- For probability vectors with , writing for ,
In both parts the maxima defining the values exist. The modulus does not depend on the dimension, which is what makes certified value enclosures possible from approximate spectra.
Differences from the source:
- generalization: probability spectra λ, ν → any probability vectors λ, ν ∈ ℝ^n (not required to be sorted), G evaluated at diag λ; reason: G is invariant under unitary conjugation, so sorted vectors are the special case; the source's argument (the common lower bound diag(min(λ_i, ν_i))) covers unsorted vectors
- encoding: the value G_{m,r}(ρ) (a maximum, (C.6)) → IsGreatest of the published rewardSet with F = χ_r; reason: the base mission's convention; IsGreatest asserts the maximum exists and equals the value, never a junk sSup
- encoding: 0 ⪯ σ ⪯ ρ, T_0 = tr ρ, τ = tr(ρ − σ) → σ ⪰ 0 and ρ − σ ⪰ 0; traces are real parts of the complex traces; reason: the traces of Hermitian matrices are real
Preamble
import Mathlib import Definitions.Def_SpectralFeedback_rewardSet import Definitions.Def_SpectralFeedback_Supplement_chi open scoped ComplexOrder
Formal statement
namespace SpectralFeedback
namespace Supplement
/-- **Theorem D.2** (A dimension-free spectral modulus; Qi, *Spectral Structure and Exact Arithmetic
of Finite-Feedback Measurement Discrimination*, manuscript (Nov. 2026), p.55 (PDF p.57)). Extend
`G_{m,r}` to positive matrices by (C.6). If `0 ⪯ σ ⪯ ρ`, `T_0 = tr ρ` and `τ = tr(ρ − σ)`, then
`0 ≤ G_{m,r}(ρ) − G_{m,r}(σ) ≤ 2√(T_0 τ) + τ` (D.10). For probability spectra with
`δ = ‖λ − ν‖_1`, `|G_{m,r}(λ) − G_{m,r}(ν)| ≤ √(2δ) + δ/2` (D.11).
**Formalization Note** `G_{m,r}(·) = v` is `IsGreatest (rewardSet r (chi r) m ·) v`; each claim
asserts that the maxima exist and satisfy the bound. `0 ⪯ σ ⪯ ρ` is `σ ⪰ 0` and `ρ − σ ⪰ 0` (same
dimension `d`, any `d`, no trace normalization); traces are real parts of the complex traces.
`G_{m,r}(λ)` for a probability vector `λ ∈ ℝ^n` is `G_{m,r}(diag λ)`. The vectors `λ, ν` are any
probability vectors of the same length (not required to be sorted). -/
theorem theorem_D_2_spectral_modulus (m r : ℕ) (hm : 1 ≤ m) (hr : 1 ≤ r) :
(∀ {d : ℕ} (ρ σ : Matrix (Fin d) (Fin d) ℂ), σ.PosSemidef → (ρ - σ).PosSemidef →
∃ v w : ℝ, IsGreatest (rewardSet r (chi r) m ρ) v ∧ IsGreatest (rewardSet r (chi r) m σ) w ∧
0 ≤ v - w ∧
v - w ≤ 2 * Real.sqrt (ρ.trace.re * (ρ - σ).trace.re) + (ρ - σ).trace.re) ∧
(∀ {n : ℕ} (lam nu : Fin n → ℝ), 0 ≤ lam → ∑ i, lam i = 1 → 0 ≤ nu → ∑ i, nu i = 1 →
∃ v w : ℝ,
IsGreatest (rewardSet r (chi r) m (Matrix.diagonal (fun i => ((lam i : ℝ) : ℂ)))) v ∧
IsGreatest (rewardSet r (chi r) m (Matrix.diagonal (fun i => ((nu i : ℝ) : ℂ)))) w ∧
|v - w| ≤ Real.sqrt (2 * ∑ i, |lam i - nu i|) + (∑ i, |lam i - nu i|) / 2) := by sorry
end Supplement
end SpectralFeedback
Source
Qi, Spectral Structure and Exact Arithmetic of Finite-Feedback Measurement Discrimination, manuscript (Nov. 2026; PDF sha256 a5b5cf40…eabde9cf), Theorem D.2, (D.10)–(D.11), p.55 (PDF p.57)