InformationTheory

InformationTheory.Shannon.WynerZiv.Converse

source

Wyner–Ziv converse (operational lower bound on the rate) #

This file provides the converse leg of the Wyner–Ziv operational main theorem (Cover–Thomas): every achievable rate R at distortion D for the i.i.d. source P_XY with decoder side information satisfies R_WZ(D) ≤ R, where R_WZ is the reshaped Wyner–Ziv rate function wynerZivRate — the infimum of the objective over feasible factorizable points at every finite auxiliary alphabet (FactorizableRate.lean §10).

Proof outline (steps 6–10 of the plan) #

For a block Wyner–Ziv code with deterministic encoder J : (Fin n → α) → Fin M and side-information decoder on an i.i.d. source (Xⁿ, Yⁿ):

  1. n·R ≥ H(J) ≥ I(J; Xⁿ) − I(J; Yⁿ) (deterministic encoder + data processing).
  2. The single-letter auxiliary is Uᵢ := (J, Y_{\i}) — the encoder output J together with all the other side-information symbols Y_{\i} = (Yⱼ)_{j≠i}. The full block Yⁿ = (Y_{\i}, Yᵢ) is forced onto Uᵢ because the per-letter reconstruction X̂ᵢ = (decoder (J, Yⁿ))ᵢ depends on the entire Yⁿ; a one-sided Y^{i-1} auxiliary is therefore ruled out (distortion-hostile).
  3. Memorylessness gives the per-letter Markov chain Uᵢ − Xᵢ − Yᵢ (wz_perletter_markov, proved sorry-free from iIndepFun). Together with the conditional mutual-information chain — not the heterogeneous Csiszár sum identity, which is orphaned on this route — ∑ᵢ [I(Xᵢ; Uᵢ) − I(Yᵢ; Uᵢ)] = ∑ᵢ I(Xᵢ; Uᵢ | Yᵢ) (Markov ⟹ I(Yᵢ; Uᵢ | Xᵢ) = 0) = ∑ᵢ I(Xᵢ; J | Yⁿ) ((Y_{\i}, Yᵢ) = Yⁿ + memoryless collapse) ≤ I(Xⁿ; J | Yⁿ) = I(J; Xⁿ) − I(J; Yⁿ) (conditional chain rule + J − Xⁿ − Yⁿ).
  4. Per-letter feasibility (each empirical (Xᵢ, Yᵢ, Uᵢ) is IsWynerZivFactorizable via the Markov chain) lands each objective as a value of wzRateValueSet at its own budget Dᵢ; time-sharing (wzRateValueSet_avg_mem) averages them.
  5. The average distortion budget (1/n) ∑ᵢ Dᵢ ≤ D (from hD) with wzRateValueSet_mono_in_D and the reshaped landing wynerZivRate_le_of_feasible reaches R_WZ(D) ≤ (1/n)(I(J; Xⁿ) − I(J; Yⁿ)) ≤ (1/n) log M.

The per-letter measure-form mutual informations are landed onto the pmf-form wzMutualInfoXU / wzMutualInfoYU via the proved bridges wzMutualInfoXU_eq_mutualInfo / wzMutualInfoYU_eq_mutualInfo.

Auxiliary-alphabet quantification (reshape rationale) #

The single-letterized auxiliary Uᵢ := (J, Y_{\i}) constructed in the proof has a type that varies with i and n and a cardinality that grows with the block length. The fixed-U rate wynerZivRateFactorizable U cannot receive such an auxiliary without a Carathéodory cardinality reduction (embedding the rate-optimal auxiliary into a U with |α| + 1 ≤ |U|) — a hard support lemma plus a shared-decoder n-ary Jensen on the converse's critical path.

The reshape (proposal A) removes both: the converse concludes against wynerZivRate, the infimum of the objective over feasible factorizable points at every finite auxiliary alphabet Fin k at once (FactorizableRate.lean §10). A large single-letterization auxiliary of any finite type (here Uᵢ of type Fin M × ({j // j ≠ i} → β)) then lands directly as a feasible point of the reshaped infimum via wynerZivRate_le_of_feasible, with no cardinality bound and no support lemma. The reshaped statement is -clean: it carries no auxiliary sizing precondition.

Non-degeneracy (junk-sInf guard): wynerZivRate = sInf (wzRateValueSet …) and, in , sInf ∅ = 0. The union-of-images form of wzRateValueSet injects no junk (empty constraints contribute the empty image), and the objective's data-processing non-negativity I(X;U) − I(Y;U) ≥ 0 (Markov chain U − X − Y) bounds the value set below by 0 uniformly in the auxiliary size (wzRateValueSet_bddBelow_of_pmf), so the sInf is a genuine non-negative rate, not a vacuous ≤ 0.

The single-letterization sub-lemmas — per-letter factorizability wz_perletter_factorizable (with its empirical-factorizable crux wz_perletter_empirical_factorizable), the conditional-MI collapse / rate atoms, and the distortion average wz_perletter_distortion_avg — are now closed sorryAx-free; the data-processing non-negativity wzObjective_nonneg_of_factorizable is likewise discharged genuinely (sorryAx-free) via the measure-form DPI + the pmf↔measure bridges + a discrete Markov-chain realization (wzFactorizable_isMarkovChain), so wzRateValueSet_bddBelow_of_pmf (the reshaped rate's non-degeneracy BddBelow guard) is likewise unconditional. The single-letterization witness wz_converse_feasible_point is itself closed sorryAx-free (machine-checked #print axioms). The L1 Carathéodory fixed-K identification wynerZivRate_eq_factorizable_finK and its core wz_support_reduce (the support-cardinality reduction to Fin (|α|+3)) are now closed sorry-free, so the entire converse headline wyner_ziv_converse is sorryAx-free.

Module structure #

Umbrella of the Shannon/WynerZiv/Converse/ family, re-exporting:

  • Converse.Prelim — the n-letter converse, the reshaped-rate non-degeneracy, the local pmf → measure realization, and the append form of IsMarkovChain.
  • Converse.SingleLetter — the per-letter Markov gateway atom, the single-letterization sub-lemmas, and the single-letter rate bound wynerZivRate_le_of_code.
  • Converse.Headline — the endpoint right-continuity infrastructure and the operational converse headline wyner_ziv_converse.