InformationTheory

InformationTheory.Shannon.BrascampLieb

source

Brascamp–Lieb inequality (combinatorial form) and hypercube product projection bound #

Generalizes the Loomis–Whitney inequality (LoomisWhitney.lean) to an arbitrary cover family S : ι → Finset (Fin n) where each j : Fin n is covered at least k times: |A|ki:ι|πSi(A)|.

Main definitions #

Main statements #

Implementation notes #

The proof routes through shearer_inequality from LoomisWhitney.lean as the entropic engine, then peels off the logarithm via Real.log_le_log_iff.

Loomis–Whitney (loomis_whitney) is the special case S i := univ.filter (· ≠ i) (each j covered n-1 times), but the existing LoomisWhitney.lean shape is preserved to avoid changing projectionExcept and loomis_whitney.

Projection onto a coordinate subset #

def

InformationTheory.Shannon.projectionSubset

source
{n : } {α : Type u_1} [DecidableEq α] (S : Finset (Fin n)) (A : Finset (Fin nα)) :
Finset (Sα)

The image of A : Finset (Fin n → α) under restriction to coordinates in S : Finset (Fin n).

Equations
Instances For
    Used by
      theorem

      InformationTheory.Shannon.jointEntropySubset_le_log_projectionSubset_card

      source
      {n : } {α : Type u_1} [Fintype α] [DecidableEq α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] {A : Finset (Fin nα)} (hA : A.Nonempty) (S : Finset (Fin n)) :
      jointEntropySubset (ProbabilityTheory.uniformOn A) (fun (i : Fin n) (ω : Fin nα) => ω i) S Real.log (projectionSubset S A).card

      For A equipped with the uniform measure, the joint entropy over coordinates S is bounded by log |projectionSubset S A|.

      Used by

        Brascamp–Lieb inequality #

        theorem

        InformationTheory.Shannon.brascamp_lieb_finset

        source
        {n k : } {ι : Type u_1} [Fintype ι] {α : Type u_2} [Fintype α] [DecidableEq α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] {A : Finset (Fin nα)} (hA : A.Nonempty) (S : ιFinset (Fin n)) (hk : ∀ (j : Fin n), k {i : ι | j S i}.card) :
        A.card ^ k i : ι, (projectionSubset (S i) A).card

        Brascamp–Lieb inequality (combinatorial form).

        If S : ι → Finset (Fin n) is a cover family such that each j : Fin n is covered at least k times, then for any nonempty finite set A : Finset (Fin n → α): |A|ki:ι|πSi(A)|.

        Loomis–Whitney is the special case S i := univ.filter (· ≠ i) with k = n - 1. The hypercube product projection bound is the special case S i := {i} with k = 1.

        Used by

          Hypercube product projection bound #

          theorem

          InformationTheory.Shannon.hypercube_product_projection_bound

          source
          {n : } {α : Type u_1} [Fintype α] [DecidableEq α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] {A : Finset (Fin nα)} (hA : A.Nonempty) :
          A.card i : Fin n, (projectionSubset {i} A).card

          Singleton-cover corollary: for any nonempty A : Finset (Fin n → α), |A| ≤ ∏ i, |π_{{i}}(A)|.

          This is brascamp_lieb_finset with S i := {i} and k := 1. Setting α = Bool recovers |A| ≤ 2^n since each singleton projection has cardinality at most 2.

          Used by