TheoremBase

Relative Entropy over a Finite Partition: the Chain Rule, Images of the Pieces, Refinement and Approximation by Partition Entropies

For a finite measurable partition, the relative entropy splits into the entropy of the cell masses plus the average entropy of the conditioned pieces; images of the pieces can only lose entropy; refining a partition can only increase the entropy of the cell masses; and suitable partitions bring it arbitrarily close to the full relative entropy.

Statement

In the settings of The Real Numbers: Standing Notation and Background and Measure Spaces and the Lebesgue Integral: Standing Notation, the measure space of the latter being instantiated at each use by the measure space named, let (S,S)(S,\mathcal{S}) be a measurable space, let γ\gamma and ν\nu be probability measures on it, and suppose that ν\nu has finite relative entropy H(ν ∣ γ)H(\nu\,|\,\gamma) with respect to γ\gamma. Finite measurable partitions and refinement are those of that definition, conditioned probability measures λ(⋅ ∣ A)\lambda(\cdot\,|\,A) are those of that definition, the partition entropy hA(ν ∣ γ)h_{\mathcal{A}}(\nu\,|\,\gamma) is that of that definition, and image measures are those of claim 1 of Image Measures, Measures with Densities, and Change of Variables. In claims 1 and 2, A=(A1,…,Am)\mathcal{A}=(A_{1},\dots,A_{m}) is a finite measurable partition of SS, pi=ν(Ai)p_{i}=\nu(A_{i}) and qi=γ(Ai)q_{i}=\gamma(A_{i}) for i∈[m]i\in[m], and the pieces of ν\nu and of γ\gamma on AiA_{i} are the conditioned measures νi=ν(⋅ ∣ Ai)\nu_{i}=\nu(\cdot\,|\,A_{i}) whenever 0<pi0<p_{i}, and γi=γ(⋅ ∣ Ai)\gamma_{i}=\gamma(\cdot\,|\,A_{i}) whenever 0<qi0<q_{i}.

1. (Chain rule) For every i∈[m]i\in[m] with 0<pi0<p_{i} one has 0<qi0<q_{i}, so that hA(ν ∣ γ)h_{\mathcal{A}}(\nu\,|\,\gamma) is defined; for every such ii the piece νi\nu_{i} has finite relative entropy with respect to the piece γi\gamma_{i}; and

H(ν ∣ γ)=hA(ν ∣ γ)+∑i∈[m], 0<pipi H(νi ∣ γi).H(\nu\,|\,\gamma)=h_{\mathcal{A}}(\nu\,|\,\gamma)+\sum_{i\in[m],\,0<p_{i}}p_{i}\,H(\nu_{i}\,|\,\gamma_{i}).

2. (Images of the pieces) Let (S′,S′)(S',\mathcal{S}') be a measurable space and R:S→S′R:S\to S' measurable with respect to S\mathcal{S} and S′\mathcal{S}'. Then for every i∈[m]i\in[m] with 0<pi0<p_{i}, so that 0<qi0<q_{i} by claim 1, the measure R#νiR_{\#}\nu_{i} has finite relative entropy with respect to R#γiR_{\#}\gamma_{i}, and

hA(ν ∣ γ)+∑i∈[m], 0<pipi H(R#νi ∣ R#γi)≤H(ν ∣ γ).h_{\mathcal{A}}(\nu\,|\,\gamma)+\sum_{i\in[m],\,0<p_{i}}p_{i}\,H(R_{\#}\nu_{i}\,|\,R_{\#}\gamma_{i})\le H(\nu\,|\,\gamma).

3. (Refinement) Let A\mathcal{A} and B\mathcal{B} be finite measurable partitions of SS such that B\mathcal{B} refines A\mathcal{A}; the partition entropies over A\mathcal{A} and over B\mathcal{B} are defined by claim 1. Then hA(ν ∣ γ)≤hB(ν ∣ γ)≤H(ν ∣ γ)h_{\mathcal{A}}(\nu\,|\,\gamma)\le h_{\mathcal{B}}(\nu\,|\,\gamma)\le H(\nu\,|\,\gamma).

4. (Approximation) For every positive ε∈R\varepsilon\in\mathbb{R} there is a finite measurable partition A′\mathcal{A}' of SS with H(ν ∣ γ)−ε<hA′(ν ∣ γ)H(\nu\,|\,\gamma)-\varepsilon<h_{\mathcal{A}'}(\nu\,|\,\gamma).

Proofs

Log in to submit a proof.

Loading...

Citations

Loading…

Dependencies

Loading…

Related

0 relations

Curated associations between results. These are editable and subjective — they do not replace the dependency graph, which is derived from the references in the text.

No relations recorded yet.

Comments

Log in to comment.

Loading…