Additivity of sandwiched channel mutual information
OpenQAlgorithms.Papers.Independent.Intakenext100.P15.sandwiched_channel_additivityAll Hilbert spaces are finite-dimensional and nonzero. A density operator is positive semidefinite with trace one; singular states are included. All logarithms are base two. For real , denotes the sandwiched Rényi divergence and the traditional Petz divergence, with value on support failure and negative spectral powers taken on the support.
Write , where is unnormalized, and . Its reference marginal is and its channel-input marginal is . Define over all normalized densities. The finite attained extrema are retained as conclusions in the minimax statement.
For arbitrary CPTP maps and , and every ,
The tensor-power specialization, for every integer , is
Input and output dimensions may all differ. The joint optimization includes correlated states on both the reference inputs and comparison outputs. This additivity converts the multiletter converse into a single-letter bound.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P15
noncomputable section
open scoped Matrix ComplexOrder ComplexConjugate Topology
open Filter
namespace QAlgorithms.Papers.Independent.Intakenext100.P15
attribute [local instance] Classical.propDecidable
attribute [local instance] Classical.decEq
-- These are pinned accepted definitions, not other chunks from the current run.
open QAlgorithms.Papers.Independent.Bridge50.P43 (NormalizedDensity payoff)
open QAlgorithms.Papers.Independent.Dependency50.P09 (Divergence logExt sandwichedD petzD)
open QAlgorithms.Papers.Independent.Information100.P34 (choi)
open QAlgorithms.Papers.Independent.Information100.P63 (schatten)
open QAlgorithms.Papers.Independent.Information100.P71
(supportPower SupportIncluded stateRelative channelPower diamond)
/-- PDF pp.12--13, Lemma 6, (5.23). Joint optimizations include correlated inputs
and comparisons. The two channels may have four different nonzero dimensions. -/
theorem sandwiched_channel_additivity
{A1 B1 A2 B2 : Type} [Fintype A1] [Fintype B1] [Fintype A2] [Fintype B2]
[DecidableEq A1] [DecidableEq B1] [DecidableEq A2] [DecidableEq B2]
[Nonempty A1] [Nonempty B1] [Nonempty A2] [Nonempty B2]
(N1 : Superop A1 B1) (hN1 : IsChannel N1)
(N2 : Superop A2 B2) (hN2 : IsChannel N2) (alpha : ℝ) (ha : 1 < alpha) :
channelInfo (sandwichedD alpha) (unbundledTensorMap N1 N2) =
channelInfo (sandwichedD alpha) N1 + channelInfo (sandwichedD alpha) N2 ∧
(∀ n : ℕ, 1 ≤ n → channelInfo (sandwichedD alpha) (channelPower N1 n) =
((n : ℝ) : EReal) * channelInfo (sandwichedD alpha) N1) := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P15
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For nonempty finite types , all with decidable equality, let be matrix functions admitting finite Kraus representations with , and let be real. For any such matrix function , define , , and . For positive semidefinite trace-one , let if , and otherwise let it equal
where is at , for , and otherwise. Set , taking the extrema over all positive semidefinite trace-one input and output matrices and using extended-real arithmetic. Define
and define . The theorem asserts , and, for every natural , ; the power assertion does not include .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.