InformationTheory.Shannon.BroadcastChannel.Marton.Achievability
Marton's inner bound — achievability #
The three-tier random-coding ensemble of Marton.ErrorAnalysis is assembled here into a single
broadcast code. A rate pair strictly inside Marton's region is first split into subcodebook rates
by exists_martonRateSplit; the covering lemma then makes the encoder's selection succeed with
high probability, the conditional AEP of Marton.MarkovCore bounds the transmitted-pair term of
each receiver, and the alias estimate bounds every other message row. A pigeonhole over the
ensemble turns the averaged bound into one deterministic code.
Three radii are nested rather than shared: the decoders test weak joint typicality at radius ε,
the transmitted blocks are pinned at martonStrongRadius, and the encoder selects a pair pinned at
martonCoveringRadius, the smallest of the three. The two receivers induce unrelated covering
radii, so the selection runs at their minimum and is reopened to either one by
jointStronglyTypicalSet_mono_radius.
Main statements #
marton_achievability— a rate pair satisfying the three strict Marton inequalities is achievable over a general two-receiver broadcast channel.
Generic finite-sum plumbing #
Aggregating one receiver's Bonferroni decomposition over the ensemble #
The transmitted-pair term over the ensemble #
Per-receiver ensemble averages #
From the ensemble to one code #
Headline #
InformationTheory.Shannon.BroadcastChannel.Marton.marton_achievability
sourceAchievability of Marton's inner bound for a two-receiver broadcast channel with private
messages. For auxiliary variables with joint law pV, an input kernel K and a channel W, a
rate pair obeying the three strict inequalities
R₁ < I(V₁; Y₁),R₂ < I(V₂; Y₂),R₁ + R₂ < I(V₁; Y₁) + I(V₂; Y₂) − I(V₁; V₂)
is achievable: at every large enough block length there is a BroadcastCode whose two average
error probabilities are both below any prescribed ε'.
The input is a general kernel K : Kernel (V₁ × V₂) α rather than a deterministic map
x = f(v₁, v₂), so the textbook statement for a deterministic input is not a direct corollary:
a deterministic kernel puts no mass on inputs off its image and therefore violates the
full-support hypothesis hK that every typicality estimate in this development requires. The
rate region is the same either way.
The hypotheses hpV, hK and hW are full-support regularity preconditions, shared with
marton_strong_mutual_covering and with bc_achievability, and carry no part of the coding
argument. No positivity of the rates is assumed: a nonpositive rate only asks for a single
message, ⌈exp (n R)⌉₊ = 1, which the construction supplies as well.
No typicality radius appears in the statement, and the construction uses three nested ones rather
than a single shared one: the two decoders test weak joint typicality at a radius ε, the
transmitted blocks are pinned at the strictly smaller martonStrongRadius, and the encoder
selects a pair pinned at the smaller martonCoveringRadius again.
@audit:ok