Data-processing inequality for the trade-off function
ProvedStatUnlearn.tof_le_tof_compLet be a Markov kernel from a measurable space to a measurable space , let be probability distributions on , and let , be their images under . Then for every type-I level ,
Post-processing cannot make two distributions easier to tell apart: a test on the output space pulls back to the test on the input space, with exactly the same type-I and type-II errors, so every error pair achievable for the post-processed distributions is already achievable for the originals and the infimum over the larger set is no larger.
This is the data-processing inequality for trade-off functions, the direction of Blackwell's theorem that governs every information-theoretic statement in the paper: it converts " maps one experiment to another" into an inequality between trade-off functions, and it is what makes the Gaussian baselines comparable in Proposition 3.1 and Proposition 2.1.
import Mathlib.Probability.Kernel.Composition.IntegralCompProd import Mathlib.Probability.Kernel.Composition.CompNotation import Mathlib.MeasureTheory.Measure.MeasureSpaceDef import Definitions.Def_su_tof open MeasureTheory ProbabilityTheory
namespace StatUnlearn
theorem tof_le_tof_comp {W V : Type*} [MeasurableSpace W] [MeasurableSpace V]
(κ : Kernel W V) [IsMarkovKernel κ] (P Q : Measure W) [IsProbabilityMeasure P]
[IsProbabilityMeasure Q] (t : ℝ) (ht : 0 ≤ t) :
tof P Q t ≤ tof (κ ∘ₘ P) (κ ∘ₘ Q) t := by
sorry
end StatUnlearn