The averaged loss (28) is -Lipschitz for the norm
ProvedStabGen.Entropy.avgLoss_lipschitzalgorithmic-stabilitylearning-theoryp2o-batch-pfp1ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let carry a reference measure and let the base loss satisfy for all and , with measurable. Let be the averaged loss (28). For all integrable and every example ,
Since is linear in , this is the statement that is -admissible with respect to the class of densities; it is used twice in the proof of Theorem 24.
Formalization Note The page states the inequality for elements of (densities); it is stated here for all integrable , which contains that case. Integrability makes both Bochner integrals genuine.
Preamble
import Mathlib import Definitions.Def_StabGen_Entropy_Model
Formal statement
namespace StabGen.Entropy
open MeasureTheory
/-- §5.2.3, p. 518: the averaged loss (28) is `M`-Lipschitz in `g` for the `L¹(ν)` norm,
`|ℓ(g, z) − ℓ(g', z)| ≤ M ∫_Θ |g(θ) − g'(θ)| dθ`, when the base loss `r` is bounded by `M`. -/
theorem avgLoss_lipschitz {Θ Z : Type*} [MeasurableSpace Θ] (ν : Measure Θ)
(r : Θ → Z → ℝ) (M : ℝ)
(hr_meas : ∀ z, Measurable (fun θ => r θ z))
(hr : ∀ θ z, 0 ≤ r θ z ∧ r θ z ≤ M)
(g g' : Θ → ℝ) (hg : Integrable g ν) (hg' : Integrable g' ν) (z : Z) :
|avgLoss ν r g z - avgLoss ν r g' z| ≤ M * ∫ θ, |g θ - g' θ| ∂ν := by sorry
end StabGen.Entropy
Source
Bousquet & Elisseeff, Stability and Generalization, JMLR 2 (2002), p. 518, §5.2.3 (the display before Theorem 24)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.