Lemma 3.11 — finiteness of the minimal inner risk characterizes existence of approximate minimizers (measurable loss)
ProvedSupportVectorMachines.Calibration.lemma_3_11_existence_of_approximate_minimizers_v2This is Lemma 3.11 (Existence of approximate minimizers) of Steinwart & Christmann, Support Vector Machines (Springer 2008, p. 57), cited in the proof of Theorem 3.17.
Let be a complete measurable space, a loss (Definition 2.1: measurable and nonnegative), a distribution on represented by , and . Then the following are equivalent:
- for -almost all ;
- there exists a measurable with for -almost all .
Formalization Note. The retired version quantified over the bare function type for ; a loss whose unique -minimizer depends non-measurably on admits no measurable selection, which refuted (i) ⇒ (ii). The corrected statement takes in the bundled Loss X (measurable and nonnegative), as Definition 2.1 and the measurable-selection step of the proof require, and uses the book's definition of a complete measurable space. ranges over ENNReal with , i.e. over exactly. Conventions made explicit (common to the corrected Chapter 3 milestones): a loss is the bundled Loss X of Definition 2.1 — measurable on and nonnegative — with the closed label set embedded in (a loss on extends by outside , a distribution on is one on supported on , and a set of distributions on is a set of measures on ); a distribution on is represented by its marginal (a probability measure) together with a measurable family of probability measures, the regular conditional probabilities (which exist since is Polish, Lemma A.3.16; every statement is invariant under the choice of version), so that risks are written in the form of Eq. (3.5); all risks are -valued Lebesgue integrals, as in the book; and IsCompleteMeasurableSpace is now the book's own notion (the -algebra equals its universal completion) instead of the retired stronger sufficient condition.
import Mathlib import Definitions.Def_SupportVectorMachines_Calibration_Loss_v2 import Definitions.Def_SupportVectorMachines_Calibration_InnerRisks_v2 import Definitions.Def_SupportVectorMachines_Calibration_OuterRisks_v2 import Definitions.Def_SupportVectorMachines_Calibration_IsCompleteMeasurableSpace_v2 open MeasureTheory
namespace SupportVectorMachines.Calibration
/-- Lemma 3.11 (Existence of approximate minimizers), Steinwart & Christmann, *Support Vector
Machines*, Springer 2008, p. 57: let `X` be a complete measurable space,
`L : X × Y × ℝ → [0,∞)` be a loss (Definition 2.1: measurable and nonnegative, bundled in
`Loss X`), `P` (represented by `(PX, κ)`, `κ` a measurable family of conditional distributions
`P(·|x)`) be a distribution on `X × ℝ`, and `ε ∈ (0,∞]`. Then the following are equivalent:
i) `C*_{L,P(·|x),x} < ∞` for `PX`-almost all `x ∈ X`; ii) there exists a measurable `f : X → ℝ`
such that `f(x) ∈ M_{L,P(·|x),x}(ε)` for `PX`-almost all `x ∈ X`.
Corrected version of `lemma_3_11_existence_of_approximate_minimizers`, whose `Loss X` was the
bare function type (no measurability, so no measurable selection existed), and whose
completeness notion was a stronger sufficient condition. -/
theorem lemma_3_11_existence_of_approximate_minimizers_v2 {X : Type*} [MeasurableSpace X]
(hX : IsCompleteMeasurableSpace X) (L : Loss X)
(PX : Measure X) [IsProbabilityMeasure PX] (κ : X → Measure ℝ) (hκ : Measurable κ)
(hκprob : ∀ x, IsProbabilityMeasure (κ x))
(ε : ENNReal) (hε : 0 < ε) :
(∀ᵐ x ∂PX, minInnerRisk L (κ x) x < ⊤) ↔
∃ f : X → ℝ, Measurable f ∧ ∀ᵐ x ∂PX, f x ∈ approxMinimizers L (κ x) x ε := by sorry
end SupportVectorMachines.Calibration