Proposition 2.7: Classical loss comparisons
ProvedQAlgorithms.Papers.Independent.Extend50.P32.classicalLossComparisonsLet and let be nonnegative probability vectors on , and let . For , put . Hellinger squared is , chi-squared is , and KL is . The estimate is the first argument, zero-over-zero terms contribute zero, and positive estimated mass at zero true mass gives infinite divergence. A superscript retains only the first terms, without renormalization.
This isolates the source comparison or construction with all of its domain and occurrence conventions retained.
import Definitions.Def_QAlgorithms_Papers_Independent_Extend50_P32
noncomputable section
open MeasureTheory Filter
open scoped BigOperators ENNReal Matrix ComplexOrder Topology
namespace QAlgorithms.Papers.Independent.Extend50.P32
open QAlgorithms.Papers.Independent.Support50.P38 (wordMass tv)
-- For the density matrices below, the characteristic-polynomial identity
-- preserves all eigenvalues with multiplicity and is equivalent to unitary
-- diagonalization with these real eigenvalues. No accuracy is assumed.
open QAlgorithms.Papers.Independent.Gap20.P17 (HasSpectrum)
/-- Proposition 2.7, PDF p.12: full/full, truncated/truncated, and full KL/chi.
No probability support is excluded. -/
theorem classicalLossComparisons {d : ℕ} (hd : 0 < d) (alpha beta : Fin d → ℝ)
(hAlpha : Learning.IsDistribution alpha) (hBeta : Learning.IsDistribution beta)
(k : ℕ) (hk : 1 ≤ k) (hkd : k ≤ d) :
ENNReal.ofReal (h2 d alpha beta) ≤ chi2 d alpha beta ∧
ENNReal.ofReal (h2 k alpha beta) ≤ chi2 k alpha beta ∧
kl alpha beta ≤ chi2 d alpha beta := by sorry
end QAlgorithms.Papers.Independent.Extend50.P32
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let , let have nonnegative coordinates summing to one, and let . For , define
Define if some coordinate satisfies and ; otherwise define
using the natural logarithm and viewing finite nonnegative values as extended nonnegative reals. The three assertions are
Zero coordinates are permitted, with the infinite-value cases specified above; no sorting assumption is made.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.