A Divergence Formula for Randomness and Dimension

TL;DR

This paper establishes a divergence formula linking the dimension of sequences with respect to a probability measure to the Shannon entropy and Kullback-Leibler divergence, providing a new way to measure the similarity of probability measures using randomness.

Contribution

It introduces a divergence formula connecting sequence dimensions, entropy, and divergence, and extends it to finite-state dimensions and normal sequences, with new compression characterizations.

Findings

The divergence formula holds for random sequences with respect to computable measures.

The formula extends to all $eta$-normal sequences using finite-state dimensions.

Finite-state compression characterizations of dimensions are established.

Abstract

If $S$ is an infinite sequence over a finite alphabet $\Sigma$ and $\beta$ is a probability measure on $\Sigma$, then the {\it dimension} of $ S$ with respect to $\beta$, written $\dim^\beta(S)$, is a constructive version of Billingsley dimension that coincides with the (constructive Hausdorff) dimension $\dim(S)$ when $\beta$ is the uniform probability measure. This paper shows that $\dim^\beta(S)$ and its dual $\Dim^\beta(S)$, the {\it strong dimension} of $S$ with respect to $\beta$, can be used in conjunction with randomness to measure the similarity of two probability measures $\alpha$ and $\beta$ on $\Sigma$. Specifically, we prove that the {\it divergence formula} \[ \dim^\beta(R) = \Dim^\beta(R) =\frac{\CH(\alpha)}{\CH(\alpha) + \D(\alpha || \beta)} \] holds whenever $\alpha$ and $\beta$ are computable, positive probability measures on $\Sigma$ and $R \in \Sigma^\infty$ is random with respect to $\alpha$. In this formula, $\CH(\alpha)$ is the Shannon entropy of $\alpha$, and $\D(\alpha||\beta)$ is the Kullback-Leibler divergence between $\alpha$ and $\beta$. We also show that the above formula holds for all sequences $R$ that are $\alpha$-normal (in the sense of Borel) when $\dim^\beta(R)$ and $\Dim^\beta(R)$ are replaced by the more effective finite-state dimensions $\dimfs^\beta(R)$ and $\Dimfs^\beta(R)$. In the course of proving this, we also prove finite-state compression characterizations of $\dimfs^\beta(S)$ and $\Dimfs^\beta(S)$.