Lemma 5.4, (5.10) — KL divergence is at most J-divergence
ProvedDROReview.PhiDiv.kl_le_jdistributionally-robust-optimizationj-divergencekullback-leiblerp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1phi-divergence
For any probability measures and on the same measurable space, the Kullback–Leibler divergence is bounded by the J-divergence:
This is the third comparison in (5.10). It gives the same-radius inclusion of the J-divergence ambiguity ball in the KL ball, including when either divergence is infinite.
Formalization Note The J generator is at zero, as required for this inequality to hold when vanishes on a set charged by .
Preamble
import Definitions.Def_DROReview_PhiDiv_phiDiv import Definitions.Def_DROReview_PhiDiv_generators
Formal statement
namespace DROReview.PhiDiv
open MeasureTheory
/-- Lemma 5.4, (5.10), third inequality, p. 34. -/
theorem kl_le_j {Ξ : Type*} [MeasurableSpace Ξ]
(P P₀ : ProbabilityMeasure Ξ) :
phiDiv phiKL P P₀ ≤
phiDiv phiJ P P₀ := by sorry
end DROReview.PhiDiv
Source
Rahimian, Mehrotra, Distributionally Robust Optimization: A Review, arXiv:1908.05659v1 (2019), p. 34, Lemma 5.4, (5.10), third inequality
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.