Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Result 4, p. 353 — the Gentle AdaBoost algorithm (population version) uses Newton steps for minimizing Ee^{−yF(x)}

Proved
AddLogReg.Gentle.result_4

by mikedeng1 · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

boostingexponential-lossgentle-adaboostnewton-methodp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Let ν\nuν be a probability measure on X×{−1,1}X\times\{-1,1\}X×{−1,1}, the joint law of (x,y)(x,y)(x,y), let F:X→RF : X\to\mathbb RF:X→R be the current fit and fix x∈Xx\in Xx∈X. Write J(F(x)+f(x))=E(e−y(F(x)+f(x))∣x)J(F(x)+f(x)) = E\big(e^{-y(F(x)+f(x))}\mid x\big)J(F(x)+f(x))=E(e−y(F(x)+f(x))∣x) as a function of the real number f(x)f(x)f(x), and let

s(x)=−E(e−yF(x)y∣x),H(x)=E(e−yF(x)∣x).s(x) = -E\big(e^{-yF(x)}y\mid x\big),\qquad H(x) = E\big(e^{-yF(x)}\mid x\big).s(x)=−E(e−yF(x)y∣x),H(x)=E(e−yF(x)∣x).

Then

  1. s(x)s(x)s(x) is the derivative of JJJ at f(x)=0f(x) = 0f(x)=0;
  2. the derivative of JJJ at every f(x)=tf(x) = tf(x)=t is −E(e−y(F(x)+t)y∣x)-E\big(e^{-y(F(x)+t)}y\mid x\big)−E(e−y(F(x)+t)y∣x), and the derivative of this function at t=0t = 0t=0 is H(x)H(x)H(x);
  3. H(x)>0H(x) > 0H(x)>0, so the Newton step is defined;
  4. the population Gentle AdaBoost update (Algorithm 4: F(x)←F(x)+Ew(y∣x)F(x)\leftarrow F(x)+E_w(y\mid x)F(x)←F(x)+Ew​(y∣x) with w(x,y)=e−yF(x)w(x,y) = e^{-yF(x)}w(x,y)=e−yF(x)) is exactly the Newton step:
F(x)+Ew(y∣x)=F(x)−s(x)H(x).F(x) + E_w(y\mid x) = F(x) - \frac{s(x)}{H(x)}.F(x)+Ew​(y∣x)=F(x)−H(x)s(x)​.

This is Result 4: Gentle AdaBoost replaces the exact minimization of Real AdaBoost by one Newton step on the exponential criterion at each iteration.

Formalization Note "Uses Newton steps" is pinned down as in the paper's Derivation: the update equals F(x)F(x)F(x) minus first derivative over second derivative of the conditional criterion. Derivatives are HasDerivAt statements; positivity of H(x)H(x)H(x) is proved, not assumed, so the division is never Lean's junk division by zero. The statement holds for every xxx; conditional expectations are integrals against ν.condKernel x.

Preamble
import Mathlib
import Definitions.Def_AddLogReg_Gentle_Setting

open MeasureTheory ProbabilityTheory
Formal statement
namespace AddLogReg.Gentle

theorem result_4 {X : Type*} [MeasurableSpace X] (ν : Measure (X × Bool))
    [IsProbabilityMeasure ν] (F : X → ℝ) (x : X) :
    HasDerivAt (condCritAt ν F x)
        (-∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * F x)) * AddLogReg.ExpCrit.sgn b ∂(ν.condKernel x)) 0 ∧
      (∀ t : ℝ, HasDerivAt (condCritAt ν F x)
        (-∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * (F x + t))) * AddLogReg.ExpCrit.sgn b ∂(ν.condKernel x)) t) ∧
      HasDerivAt (fun t : ℝ => -∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * (F x + t))) * AddLogReg.ExpCrit.sgn b ∂(ν.condKernel x))
        (∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * F x)) ∂(ν.condKernel x)) 0 ∧
      0 < ∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * F x)) ∂(ν.condKernel x) ∧
      gentleStep ν F x =
        F x - (-∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * F x)) * AddLogReg.ExpCrit.sgn b ∂(ν.condKernel x)) /
          (∫ b, Real.exp (-(AddLogReg.ExpCrit.sgn b * F x)) ∂(ν.condKernel x)) := by sorry

end AddLogReg.Gentle
Source
Friedman, Hastie and Tibshirani, Additive logistic regression: a statistical view of boosting, Ann. Statist. 28 (2000), p. 353, Result 4 and its Derivation; Algorithm 4
Human review
  • Endorsed by Shuze Chen · Oct 7, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 7, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me