(31), p. 352 — the integrand 2y*t − log(1+e^{2t}) is the Bernoulli log-likelihood of y* under p = e^t/(e^t+e^{−t})
ProvedAddLogReg.LogitBoost.eq_31boostinglog-likelihoodlogistic-regressionlogitboostp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and , and let be the symmetric logistic (30). Then
The right-hand side is the Bernoulli log-likelihood of the response when its success probability is . With the identity shows that the integrand of (31) is the log-likelihood of the symmetric logistic model, so that in (31) is the expected Bernoulli log-likelihood the Derivation of Result 3 differentiates.
Preamble
import Mathlib import Definitions.Def_AddLogReg_LogitBoost_Setting open MeasureTheory ProbabilityTheory
Formal statement
namespace AddLogReg.LogitBoost
theorem eq_31 (b : Bool) (t : ℝ) :
2 * ystar b * t - Real.log (1 + Real.exp (2 * t)) =
ystar b * Real.log (symLogistic t) + (1 - ystar b) * Real.log (1 - symLogistic t) := by sorry
end AddLogReg.LogitBoost
Source
Friedman, Hastie and Tibshirani, Additive logistic regression: a statistical view of boosting, Ann. Statist. 28 (2000), p. 352, (31); p. 351, (30)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.