Documentation

Oseledets.Entropy.ProductIdEntropy

Frozen-factor product entropy: h(T × id) = h(T) #

For a measure-preserving system (X, T, μ) on a standard-Borel space and a standard-Borel probability space (Y, ν), the frozen product transformation T × id on (X × Y, μ ⊗ ν) has the same Kolmogorov–Sinai entropy as the base system:

h(T × id) = h(T) (Oseledets.Entropy.ksEntropy_prod_id_eq).

This is Walters' product-entropy theorem (Walters, An Introduction to Ergodic Theory, Theorem 4.23), the final assembly of GitHub issue #20. It is proved by le_antisymm from the two inequalities, each supplied by an already-proved sub-lemma module:

The reverse bound #

ksEntropy (T × id) is the supremum over finite partitions P of X × Y of the partition entropies h(P, T × id). For each P we squeeze with an increasing rectangle sequence rectₘ = ξₘ ⊠ ηₘ, where ξₘ and ηₘ are the standard countably-generated atom partitions of X and Y (MeasurableSpace.countablePartition). Le Maître's inequality gives h(P, T × id) ≤ h(rectₘ, T × id) + H(P | σ(rectₘ)); the rectangle identity rewrites h(rectₘ, T × id) = h(ξₘ, T) ≤ h(T); and the rectangle σ-algebras saturate m(X × Y), so the static conditional entropies H(P | σ(rectₘ)) tend to 0. The limit is taken on (EReal addition is not jointly continuous), handling h(T) = ⊤ separately by le_top.

Main results #

References #

@[implicit_reducible]

Fintype on a countably-generated finite partition. Each countablePartition α n is a finite set of sets (finite_countablePartition); we promote that Finite-ness to a Fintype so the set can index a MeasurePartition.

Equations

The countably-generated atom partition of a standard-Borel space, packaged as a MeasurePartition indexed by the (finite) set countablePartition α n itself. Its cells are the atoms (Subtype.val); they are measurable, genuinely disjoint and cover the space.

Equations
Instances For

    The same atom partition, reindexed to a Fin-indexed partition (so it can feed le_ksEntropy, whose supremum ranges over Fin-indexed partitions).

    Equations
    Instances For

      The Fin-reindexed atom partition generates the same σ-algebra as countablePartition.

      The generated σ-algebras of countablePartition are increasing in refinement.

      The countably-generated atom partitions saturate the ambient σ-algebra: ⨆ n, σ(countablePartition α n) = mα.

      Rectangle σ-algebra. For partitions ξ of (X, μ) and η of (Y, ν), the σ-algebra generated by the rectangle partition (fst⁻¹ ξ) ∨ (snd⁻¹ η) of X × Y is the join of the pulled- back coordinate σ-algebras comap fst σ(ξ) ⊔ comap snd σ(η). The step writes each rectangle ξᵢ × ηⱼ = fst⁻¹ ξᵢ ∩ snd⁻¹ ηⱼ; the direction writes each fst⁻¹ ξᵢ = ⋃ⱼ ξᵢ × ηⱼ using that η covers Y (and symmetrically for snd).

      The Fin-indexed rectangle sequence feeding the conditional-entropy glue: the rectangle partition ξₙ ⊠ ηₙ (atom partitions of X and Y), reindexed to a Fin-indexed partition.

      Equations
      • One or more equations did not get rendered due to their size.
      Instances For

        Per-partition reverse bound. For each finite partition P of X × Y with nonempty index, h(P, T × id) ≤ h(T). The argument squeezes via the rectangle sequence: Le Maître's inequality bounds h(P, T × id) by h(ξₙ, T) + H(P | σ(rectₙ)), the first term is ≤ h(T), and the second tends to 0.

        Walters' product-entropy theorem (frozen identity factor): h(T × id) = h(T).

        The free bound h(T) ≤ h(T × id) is ksEntropy_le_prod (the base is a factor of the product). The reverse bound h(T × id) ≤ h(T) is ksEntropyPartition_prod_le for each finite partition P of X × Y: partitions over an empty index do not exist on the nonempty space X × Y, so the supremum reduces to the per-partition bounds.