Proof of Result 1, p. 347 — −yf(x) = 2 × 1_{[y≠f(x)]} − 1, so e^{−cf(x)y} = e^{−c} exp(2c 1_{[y≠f(x)]})
ProvedAddLogReg.Discrete.neg_y_f_eqadaboostboostingp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
For ,
and consequently, for every real ,
Applied with and the line-search coefficient , so that , this shows that the population weight update differs from Discrete AdaBoost's update only by the constant factor , which renormalization removes.
Preamble
import Mathlib import Definitions.Def_AddLogReg_Discrete_Setting open MeasureTheory ProbabilityTheory
Formal statement
namespace AddLogReg.Discrete
/-- Proof of Result 1, p. 347: for `y, s ∈ {−1, 1}`, `−ys = 2 × 1_{[y ≠ s]} − 1`; hence for every
real `c` the weight factor `e^{−csy}` equals `e^{−c} · exp(2c · 1_{[y ≠ s]})`. -/
theorem neg_y_f_eq (y s : ℝ) (hy : y = 1 ∨ y = -1) (hs : s = 1 ∨ s = -1) :
-(y * s) = 2 * (if y ≠ s then 1 else 0) - 1 ∧
∀ c : ℝ, Real.exp (-(c * s * y)) =
Real.exp (-c) * Real.exp (2 * c * (if y ≠ s then 1 else 0)) := by sorry
end AddLogReg.Discrete
Source
Friedman, Hastie and Tibshirani, Additive logistic regression: a statistical view of boosting, Ann. Statist. 28 (2000), p. 347, proof of Result 1, last two displays
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.