Definition 1, p. 354 — for J = 2 the symmetric multiple logistic transformation is the two-class ½ log[p/(1−p)] and (40) is (13)
ProvedAddLogReg.Multiclass.def_1_two_classlogistic-regressionmulticlass-classificationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1softmax
Take classes. Let with and , and let be a real vector with . Then:
- the transformation (39) is the two-class symmetric log-odds of (12), and is antisymmetric:
where ; 2. the probabilities (40) are the two-class symmetric logistic (13):
This is "the equivalence with the two-class case" asserted after Definition 1: with class 1 as , the multiclass parametrization reduces to and of §4.
Formalization Note Classes are Fin 2 with playing class 1 and playing class 2.
Preamble
import Mathlib import Definitions.Def_AddLogReg_Multiclass_Setting open MeasureTheory ProbabilityTheory ENNReal
Formal statement
namespace AddLogReg.Multiclass
theorem def_1_two_class (p : Fin 2 → ℝ) (hp_pos : ∀ j, 0 < p j) (hp_sum : ∑ j, p j = 1)
(F : Fin 2 → ℝ) (hF : F 0 + F 1 = 0) :
symMultiLogit p 0 = (1 / 2) * Real.log (p 0 / p 1) ∧
symMultiLogit p 1 = -symMultiLogit p 0 ∧
softmax F 0 = Real.exp (F 0) / (Real.exp (F 0) + Real.exp (-F 0)) ∧
softmax F 1 = 1 - softmax F 0 := by sorry
end AddLogReg.Multiclass
Source
Friedman, Hastie and Tibshirani, Additive logistic regression: a statistical view of boosting, Ann. Statist. 28 (2000), p. 354, Definition 1, last sentence ("as well as the equivalence with the two-class case"); p. 345, (12)–(13)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.