Trace norm by duality: if for all contractions then (self-adjoint on )
ProvedBosonicLearn.traceSize_le_of_pairingbosonic-gaussian-statesquantum-informationquantum-learning
The variational characterization of the trace norm, in the direction needed to bound it, for the nuclear-expansion definition traceSize.
Let be a bounded self-adjoint operator on ( countable) and . If for every operator with and every finite set of basis indices
then the matrix is trace class with trace norm at most .
Proof idea: with the continuous functional calculus, and ; the contractions give , so , and a positive operator with finite trace has a nuclear expansion of total weight .
Formalization Note traceSize is the infimum over nuclear expansions ( if there is none), so the conclusion includes that is trace class. is lp.single 2 u 1 in Hilbert I.
Preamble
import Mathlib import Definitions.Def_QAlgorithms_Papers_Independent_Dependency50_P33 open QAlgorithms.Papers.Independent.Dependency50.P33 (State IsDensity traceSize) open QAlgorithms.RealRegisters (Hilbert) open scoped ENNReal
Formal statement
namespace BosonicLearn
/-- **Trace norm by duality: `‖A‖₁ ≤ sup_{‖V‖ ≤ 1} |Tr(AV)|` for self-adjoint `A`.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, footnote 1, p. 3 (trace distance `½‖ρ − σ‖₁`), and §7.1, p. 55 (trace-distance bounds between Gaussian states); this is the
standard variational characterization of the trace norm used to bound it.
Let `A` be a bounded self-adjoint operator on `ℓ²(I)` and `c ≥ 0`. If for every contraction `V` and every finite set
`F` of basis indices the partial trace `∑_{u ∈ F} ⟨e_u, A V e_u⟩` has modulus at most `c`, then the matrix of `A` in
the standard basis is trace class with trace norm at most `c`.
**Formalization Note.** `traceSize` is the infimum over nuclear expansions (`∞` if there is none), so the conclusion
includes that `A` is trace class. The matrix of `A` is `⟨e_u, A e_v⟩`, `e_u = lp.single 2 u 1`. -/
theorem traceSize_le_of_pairing {I : Type} [DecidableEq I] [Countable I] (A : Hilbert I →L[ℂ] Hilbert I)
(hA : IsSelfAdjoint A) (c : ℝ) (hc : 0 ≤ c)
(h : ∀ V : Hilbert I →L[ℂ] Hilbert I, ‖V‖ ≤ 1 → ∀ F : Finset I,
‖∑ u ∈ F, inner ℂ (lp.single 2 u (1 : ℂ) : Hilbert I) (A (V (lp.single 2 u (1 : ℂ))))‖ ≤ c) :
traceSize (fun u v : I => inner ℂ (lp.single 2 u (1 : ℂ) : Hilbert I) (A (lp.single 2 v (1 : ℂ)))) ≤
ENNReal.ofReal c := by sorry
end BosonicLearn
Source
Chen, Anna Mele, Anna Mele, Preskill, When are bosonic Gaussian states classical to learn?, arXiv:2609.26705, footnote 1, p. 3 (trace distance ); §7.1, p. 55 (trace-distance bounds between Gaussian states)