InformationTheory.Shannon.RateDistortion.AchievabilityGeneralSource
Rate-distortion achievability — general (arbitrary) source #
The unconditional operational achievability theorem
rate_distortion_achievability_operational carries a full-support regularity
precondition hP_supp : ∀ a, 0 < P_X a. This file removes it: the general
version rate_distortion_achievability_operational_general holds for any source
pmf P_X ∈ stdSimplex ℝ α.
Approach #
Restrict to the support subtype α' := {a // 0 < P_X a}, apply the full-support
theorem there, then lift the resulting lossy code back to the whole alphabet.
- A feasible joint pmf
q ∈ RDConstraint P_X d Dvanishes off the support (marginalFst q = P_Xwithq ≥ 0forcesq (a, b) = 0whenP_X a = 0). Hence the restriction/padding maps between joint pmfs onα × βandα' × βare mutually inverse on feasible sets and preservemutualInfoPmf,expectedDistortionPmf, andmarginalFst(negMulLog 0 = 0). - Therefore
rateDistortionFunctionPmfagrees onαandα', the hypotheses transport, and the full-support theorem produces a codec'onα'. - A retraction
r : α → α'(identity on the support, a fixed default off it) liftsc'to a code onα. The source measure never charges the off-support symbols, so the expected block distortion is unchanged.
Support subtype, restriction, and padding #
InformationTheory.Shannon.rdSupportPmf
sourceSource pmf restricted to the support subtype.
Equations
- InformationTheory.Shannon.rdSupportPmf P_X a' = P_X ↑a'
Instances For
Used by
InformationTheory.Shannon.rdSupportDist
sourceDistortion function restricted to the support subtype.
Equations
- InformationTheory.Shannon.rdSupportDist P_X d a' b = d (↑a') b
Instances For
Used by
InformationTheory.Shannon.rdRestrict
sourceRestrict a joint pmf on α × β to the support subtype α' × β.
Equations
- InformationTheory.Shannon.rdRestrict P_X q p = q (↑p.1, p.2)
Instances For
Used by
InformationTheory.Shannon.rdPad
sourcePad a joint pmf on α' × β back to α × β by zero off the support.
Equations
Instances For
Used by
InformationTheory.Shannon.rdRetract
sourceRetraction of the whole alphabet onto the support subtype: the identity on the support and a fixed default off it.
Equations
- InformationTheory.Shannon.rdRetract P_X a₀ a = if h : 0 < P_X a then ⟨a, h⟩ else a₀
Instances For
Used by
G0: support-zero #
InformationTheory.Shannon.rd_support_row_zero
sourceA feasible joint pmf vanishes on rows outside the source support.
Used by
InformationTheory.Shannon.rd_support_nonempty
sourceThere is a symbol in the support of any source pmf.
Used by
G1: finite-sum transfer between α and the support subtype #
InformationTheory.Shannon.rd_sum_support_transfer
sourceSum transfer: a function vanishing off the support has equal sums over α
and over the support subtype.
Used by
InformationTheory.Shannon.rd_sum_prod_support_transfer
sourceProduct sum transfer: a function vanishing on off-support rows has equal
sums over α × β and over α' × β.
Used by
InformationTheory.Shannon.rdRestrict_pad
sourceRestriction then padding is the identity on feasible pmfs.
Used by
InformationTheory.Shannon.marginalFst_rdRestrict
sourceThe first marginal of a restriction is the restriction of the first marginal.
Used by
InformationTheory.Shannon.marginalSnd_rdRestrict
sourceThe second marginal is preserved under restriction of a support-zero pmf.
Used by
InformationTheory.Shannon.mutualInfoPmf_eq_restrict
sourceMutual information is preserved by restriction of a support-zero pmf.
Used by
InformationTheory.Shannon.expectedDistortionPmf_eq_restrict
sourceExpected distortion is preserved by restriction of a support-zero pmf.
Used by
Membership transport #
InformationTheory.Shannon.rdPad_support_zero
sourcePadding is support-zero by construction.
Used by
InformationTheory.Shannon.rdRestrict_mem_stdSimplex
sourceRestriction lands in the simplex.
Used by
InformationTheory.Shannon.rdPad_mem_stdSimplex
sourcePadding lands in the simplex.
Used by
InformationTheory.Shannon.marginalFst_rdRestrict_eq
sourceThe first marginal of a restriction equals the restricted source pmf.
Used by
InformationTheory.Shannon.marginalFst_rdPad_eq
sourceThe first marginal of a padding equals the original source pmf.
Used by
InformationTheory.Shannon.rdSupportPmf_mem_stdSimplex
sourceThe restricted source pmf lies in the simplex of the support subtype.
Used by
G2: rate-distortion function invariance #
InformationTheory.Shannon.rdRestrict_mem_constraint
sourceMembership transports along restriction.
Used by
InformationTheory.Shannon.rdPad_mem_constraint
sourceMembership transports along padding.
Used by
InformationTheory.Shannon.rd_function_eq
sourceThe pmf-direct rate-distortion function is unchanged by support restriction.
Used by
G4: source measure charges only the support #
InformationTheory.Shannon.pmfToMeasure_support_compl_null
sourceThe source measure gives zero mass to the off-support symbols.
Used by
InformationTheory.Shannon.pmfToMeasure_map_retract
sourcePushing the source measure through the support retraction yields the restricted source pmf.
Used by
InformationTheory.Shannon.expectedBlockDistortion_lift_le
sourceLift the expected block distortion bound from the support subtype to the whole alphabet.
Used by
Wrapper: general-source operational achievability #
InformationTheory.Shannon.rate_distortion_achievability_operational_general
sourceRate-distortion theorem (achievability, general source).
Removes the full-support precondition from
rate_distortion_achievability_operational: for any source P_X ∈ stdSimplex
with a feasible constraint set and any rate R > R(D), for every ε > 0 there
is a block length N such that every n ≥ N admits a rate-R lossy code whose
expected block distortion is at most D + ε.
@audit:ok