Pinsker inequality in the source full-L1/nats convention
OpenQAlgorithms.Papers.Independent.Intake100.P73.pinsker_full_l1For any probability mass functions p and q on an arbitrary discrete alphabet, full L1 distance d(p,q)=sum_y |p(y)-q(y)| and extended KL divergence D(p||q) in natural logarithms satisfy d(p,q)^2/2 <= D(p||q). Support mismatch gives infinite divergence; zero-mass summands follow the standard continuous convention. The printed coefficient 1 is corrected to 1/2.
import Definitions.Def_QAlgorithms_Papers_Independent_Intake100_P73
noncomputable section
namespace QAlgorithms.Papers.Independent.Intake100.P73
open Filter
open scoped BigOperators ENNReal Topology
attribute [local instance] Classical.propDecidable Classical.decEq
universe u
/-- Hayashi PDF p.10, background Pinsker inequality. Printed: D>=full-L1^2.
Changed: D>=full-L1^2/2, because the source uses full L1 and natural logarithms.
No finite-alphabet or absolute-continuity restriction is added. -/
theorem pinsker_full_l1 {Omega : Type*} (p q : PMF Omega) :
fullL1 p q ^ 2 / 2 ≤ discreteKLDivergence p q := by sorry
end QAlgorithms.Papers.Independent.Intake100.P73
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For any type and any two probability mass functions , one has
The logarithm is the natural logarithm, and the inequality and sums are interpreted in the extended nonnegative reals. Each value of a probability mass function is finite, so the absolute difference here is exactly the sum of the two truncated differences used in the definition. No finiteness assumption is imposed on , and zero masses are allowed: , whereas makes the right-hand side infinite. The existence of excludes an empty .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.