Indexed text attachment
entropy-sources-and-channels-in-higher-categories.tex
Source for Entropy, Sources, and Channels in Higher Categories
View paper · Download original file
Lines 1461–1560 of 4249. Source text is displayed literally.
1461The proposition puts Khinchin's three kinds of object in one place. His remark that a code may be regarded1462as a channel is the inclusion of the deterministic morphisms among all morphisms.1463 1464\section{The entropy rate}1465 1466For a stationary source $\mu$ write $H_n=H(\mu_n)$ for the entropy of a block of length $n$, and $H_0=0$.1467 1468\begin{proposition}\label{prop:rate}1469\Scl[\cite{Sha48,Khi56}] Let $\mu$ be stationary.1470\begin{enumerate}[label=(\roman*)]1471\item $H_{m+n}\le H_m+H_n$.1472\item The increments $c_n=H_{n+1}-H_n=H(x_0\st x_{-n},\dots,x_{-1})$ are nonincreasing in $n$.1473\item The limits $\lim_nH_n/n$ and $\lim_nc_n$ exist and are equal, and equal $\inf_nH_n/n$.1474\end{enumerate}1475\end{proposition}1476 1477\begin{proof}1478(i) is Shannon's inequality for the two halves of a block, whose laws are $\mu_m$ and $\mu_n$ by1479stationarity. (ii) is \cref{lem:SSA}: conditioning on one more letter of the past does not increase the1480entropy; stationarity identifies the conditional entropies with the increments. (iii) $H_n=\sum_{k<n}c_k$ is1481a sum of a nonincreasing nonnegative sequence, so $H_n/n$ converges to $\lim c_k$; the infimum is Fekete's1482lemma for the subadditive sequence $H_n$.1483\end{proof}1484 1485\begin{definition}1486The common value $H(\mu)$ of the limits in \cref{prop:rate} is the \term{entropy} of the source, in bits per1487letter. It satisfies $0\le H(\mu)\le\log|A|$.1488\end{definition}1489 1490For the source that emits independent letters with distribution $p$, the state $p^{\otimes\Z}$, the block1491entropies are $H_n=nH(p)$ and the entropy of the source is $H(p)$. The second expression in (iii) has a1492meaning that the first hides: $H(\mu)$ is the uncertainty of the next letter given the entire past. We make1493this precise in \cref{ch:martingales}.1494 1495\section{Stationary codes and cofreeness}1496 1497A stationary code $\phi\colon\AZ\to\BZ$ is determined by its coordinate at time zero,1498$\phi_0=\pi_0\circ\phi\colon\AZ\to B$, since $\phi(x)_n=\phi_0(T^nx)$. This is a universal property.1499 1500\begin{proposition}\label{prop:cofree}1501\Srf[\cite{Hed69}] In the category of sets, of measurable spaces, or of compact Hausdorff spaces, let $(X,T_X)$ be1502an object with an automorphism. Then composition with $\pi_0$ is a bijection between equivariant maps1503$(X,T_X)\to(\BZ,T)$ and maps $X\to B$. That is, $(\BZ,T)$ is the cofree system on $B$. Consequently:1504\begin{enumerate}[label=(\roman*)]1505\item stationary codes $\AZ\to\BZ$ correspond to measurable functions $\AZ\to B$;1506\item continuous equivariant maps $\AZ\to\BZ$ correspond to continuous functions $\AZ\to B$, which by1507compactness depend on finitely many coordinates: they are the \term{sliding block codes}1508$\phi(x)_n=\Phi(x_{n-r},\dots,x_{n+r})$ (the Curtis--Hedlund--Lyndon theorem).1509\end{enumerate}1510In $\BorelStoch$ the corresponding statement is false.1511\end{proposition}1512 1513\begin{proof}1514Given $g\colon X\to B$ define $\phi(x)_n=g(T_X^nx)$. This is the unique equivariant map with $\pi_0\circ\phi=g$,1515and it is measurable or continuous when $g$ is, because $\BZ$ is the product in those categories. A1516continuous function from the compact space $\AZ$ to the finite discrete space $B$ is locally constant, and1517finitely many cylinders on which it is constant cover $\AZ$. For the last statement let $X=1$: an1518equivariant kernel $1\to\BZ$ is a stationary source, a kernel $1\to B$ is a distribution on $B$, and many1519stationary sources have the same one-letter marginal.1520\end{proof}1521 1522The failure in the stochastic case is not a technicality. An equivariant kernel $\nu\colon\AZ\to\BZ$1523consists of its time-zero marginal $\AZ\to B$ and, in addition, of the dependence among the output letters1524for a fixed input. The first part is what Khinchin's notion of the memory of a channel controls. The1525second is not controlled by it, and we shall see in \cref{ch:channels} that this is exactly where his1526treatment of ergodicity has a gap.1527 1528\section{Markov chains}1529 1530Let $Q=(q_{ab})$ be a stochastic matrix on $A$ and $\pi$ a stationary distribution, $\pi Q=\pi$. The1531\term{stationary Markov source} has block laws $\mu_n(x_1\dots x_n)=\pi(x_1)\,q_{x_1x_2}\cdots q_{x_{n-1}x_n}$.1532 1533\begin{proposition}\label{prop:markov}1534\Scl[\cite{Sha48}; \cite{Khi53}, \S3] The entropy of the stationary Markov source is1535\[ H(\mu)=\sum_a\pi(a)\,H(q_{a\,\cdot})=-\sum_{a,b}\pi(a)\,q_{ab}\log q_{ab}. \]1536\end{proposition}1537 1538\begin{proof}1539By the Markov property $c_n=H(x_0\st x_{-1})$ for all $n\ge1$.1540\end{proof}1541 1542\begin{example}\label{ex:markov}1543Let $A=\{0,1\}$ with $q_{01}=0.1$ and $q_{10}=0.3$. Then $\pi=(0.75,0.25)$ and1544$H(\mu)=0.75\,\hb(0.1)+0.25\,\hb(0.3)\approx0.572$ bits per letter. A source emitting independent letters with the1545same marginal has entropy $\hb(0.25)\approx0.811$. The difference is what the memory of the chain is worth to1546someone predicting it.1547\end{example}1548 1549\section{Finite-state sources as coalgebras}1550 1551A Markov chain is the simplest machine that emits a sequence. The general finite machine has a finite set1552$S$ of internal states and, at each step, emits a letter and moves to a new state, the pair being drawn1553from a distribution that depends on the current state. That is a morphism1554\[ c\colon S\longrightarrow A\times S\quad\text{in }\FinStoch, \]1555a coalgebra for the functor $A\times-$. With an initial distribution on $S$ it generates a process, called1556a hidden Markov source or a function of a Markov chain.1557 1558\begin{theorem}\label{thm:trace}1559\Sli[\cite{KK13}; cf.\ \cite{HJS07}] In $\BorelStoch$ the space $A^{\N}$ of one-sided sequences, with the1560isomorphism $A^{\N}\iso A\times A^{\N}$ splitting off the first letter, is the final coalgebra of the functor