InformationTheory.Shannon.WynerZiv.ObjectiveConvexity
Wyner–Ziv objective convexity (Cover–Thomas) #
This file discharges the h_obj_convex hypothesis carried by
WynerZiv/FactorizableRate.lean (wynerZivRateFactorizable_convex,
wynerZivRateFactorizable_convex_in_D_of_obj_convex): the convexity of the Wyner–Ziv
objective I(X;U) − I(Y;U) in the joint pmf q, along convex combinations
of factorizable joints q = κ(u|x)·P_XY(x,y).
Approach #
The objective wzMutualInfoXU U q − wzMutualInfoYU U q is expanded through
the entropy form mutualInfoPmf m = H(fst m) + H(snd m) − H(m). Two exact
algebraic cancellations make the deep content explicit and isolate it from
the trivial parts:
H(U)cancellation (anyq): theU-marginal ofwzMarginalXU U qand ofwzMarginalYU U qcoincide — both equalP_U(u) = ∑_{x,y} q(x,y,u)— by Fubini. Hence the∑_u negMulLog(P_U u)terms (theH(U)blocks) cancel betweenI(X;U)andI(Y;U). This needs no factorization.H(X) − H(Y)constancy (factorizableq): theX-marginal ofwzMarginalXU U qequals∑_y q(x,y,·)summed overu, i.e. the(X)-marginal ofwzMarginalXY U q. For factorizableqthe latter isP_XY, so this block is the constantH(P_X)— independent of the kernelκ. Likewise theY-block is the constantH(P_Y). Along a convex combination of two factorizable joints these constants are identical, so they cancel out of the convexity inequality.
After both cancellations the objective reads
objective(q) = (H(P_X) − H(P_Y)) + (H(m_YU) − H(m_XU)),
where m_XU = wzMarginalXU U q, m_YU = wzMarginalYU U q, and
H(m) = ∑ negMulLog(m ·) is the joint Shannon block. The whole
non-trivial content is therefore exactly the convexity of
the conditional-entropy difference κ ↦ H(m_YU) − H(m_XU) (this is the
conditional mutual information I(X;U|Y)).
That residual is published as the primitive predicate
WynerZivCondEntDiffConvex — a genuine ConvexOn-shaped convexity-of-MI
statement, the irreducible Lemma-15.9 core. h_obj_convex is then derived
in full from this single predicate (wzObjective_convex_of_condEntDiff),
and the rate-level wrapper wynerZivRateFactorizable_convex_in_D_of_obj_convex is
re-published with h_obj_convex replaced by the strictly more primitive
WynerZivCondEntDiffConvex.
Implementation notes #
The bare convexity of the conditional-entropy difference
WynerZivCondEntDiffConvex is the irreducible analytic core of Cover–Thomas
(convexity of I(X;U|Y) in κ, a joint-convexity-of-KL argument);
it is carried as a predicate rather than proved here. Everything around it — the
H(U) cancellation, the H(X)−H(Y) constancy, and the assembly into
h_obj_convex — is discharged.
§1 Joint Shannon block of the (·,U) marginals #
InformationTheory.Shannon.wzJointEntXU
sourceJoint Shannon block H(m_XU) = ∑_{x,u} negMulLog(m_XU(x,u)) of the
(X,U)-marginal of a joint pmf q.
Equations
- InformationTheory.Shannon.wzJointEntXU U q = ∑ p : α × U, (InformationTheory.Shannon.wzMarginalXU U q p).negMulLog
Instances For
Used by
InformationTheory.Shannon.wzJointEntYU
sourceJoint Shannon block H(m_YU) = ∑_{y,u} negMulLog(m_YU(y,u)) of the
(Y,U)-marginal of a joint pmf q.
Equations
- InformationTheory.Shannon.wzJointEntYU U q = ∑ p : β × U, (InformationTheory.Shannon.wzMarginalYU U q p).negMulLog
Instances For
Used by
§2 The U-marginals of wzMarginalXU and wzMarginalYU agree #
InformationTheory.Shannon.marginalSnd_wzMarginalXU_eq_marginalSnd_wzMarginalYU
sourceThe U-marginal of wzMarginalXU U q equals the U-marginal of
wzMarginalYU U q: both are P_U(u) = ∑_{x,y} q(x,y,u). Holds for any q
(Fubini swap of the x/y sums).
Used by
§3 The X/Y-marginals of the (·,U) marginals #
InformationTheory.Shannon.marginalFst_wzMarginalXU_eq_marginalFst_wzMarginalXY
sourceThe X-marginal of wzMarginalXU U q equals the X-marginal of
wzMarginalXY U q (sum over (y,u)).
Used by
InformationTheory.Shannon.marginalFst_wzMarginalYU_eq_marginalSnd_wzMarginalXY
sourceThe X-marginal of wzMarginalYU U q equals the Y-marginal of
wzMarginalXY U q (sum over (x,u)).
Used by
§4 Objective decomposition: H(U) cancellation #
InformationTheory.Shannon.wzObjective_decomp
sourceObjective decomposition (any q). After canceling the shared
H(U) block,
I(X;U) − I(Y;U) = [H(marginalFst m_XU) − H(marginalFst m_YU)] + [H(m_YU) − H(m_XU)],
where H(g) = ∑ negMulLog(g ·) over the relevant alphabet.
Used by
§5 The marginal-block constant on factorizable joints #
InformationTheory.Shannon.wzObjective_marginalBlock_factorisable
sourceOn a factorizable joint the X/Y-marginal blocks reduce to the
source-only constant H(P_X) − H(P_Y).
Used by
§6 The primitive convexity predicate (Lemma-15.9 core) #
InformationTheory.Shannon.WynerZivCondEntDiffConvex
sourceThe irreducible Lemma-15.9 core. Convexity of the conditional
entropy difference H(m_YU) − H(m_XU) = I(X;U|Y) along convex combinations
of factorizable joints. This is the genuine analytic content of
Cover–Thomas (a joint-convexity-of-KL argument); it is strictly
more primitive than h_obj_convex because both the H(U) cancellation and
the H(X)−H(Y) constancy have been factored out.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Used by
§7 h_obj_convex from the primitive predicate #
InformationTheory.Shannon.wzObjective_convex_of_condEntDiff
sourceObjective convexity from the conditional-entropy-difference core.
Given WynerZivCondEntDiffConvex (the Lemma-15.9 core), the Wyner–Ziv
objective I(X;U) − I(Y;U) is convex along factorizable convex
combinations — i.e. exactly the h_obj_convex hypothesis consumed by
WynerZiv/FactorizableRate.lean.
@audit:superseded-by(wynerZivCondEntDiffConvex_holds)
Used by
§8 Re-published rate-level convexity wrapper #
InformationTheory.Shannon.wynerZivRateFactorizable_convex_in_D_of_condEntDiff
sourceConvexity of the factorizable rate function in D, with h_obj_convex
replaced by the more primitive conditional-entropy-difference convexity predicate
WynerZivCondEntDiffConvex (discharged into h_obj_convex by
wzObjective_convex_of_condEntDiff).
@audit:superseded-by(wynerZivRateFactorizable_convex_in_D)