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 2501–2600 of 4249. Source text is displayed literally.
2501$f(x)=1_C(x)\,\nu_x(D)$ depends on finitely many coordinates of $x$, and likewise $f'$. For large $k$ the output2502blocks defining $D$ and $T^{-k}D'$ are more than $m$ apart, so by finite output memory and stationarity2503\[ \omega(E\cap T^{-k}E')=\int f(x)\,f'(T^kx)\,d\mu(x). \]2504By the mean ergodic theorem the Ces\`aro averages of the right side converge to $\int f\,d\mu\int f'\,d\mu=\omega(E)\,\omega(E')$,2505and this criterion on cylinders is equivalent to ergodicity (\cref{exer:cylinders}).2506\end{proof}2507 2508The categorical picture explains the shape of the gap. By \cref{prop:cofree}, an equivariant map into $\BZ$ is2509determined by its time-zero coordinate, but an equivariant kernel is not. A stationary channel consists of2510the conditional laws of finite output blocks given the input, and of the way these fit together across2511time. Khinchin's memory condition bounds the first, the dependence on the input. Ergodicity of the2512composite is a statement about the second, and \cref{ex:gap} is a channel that is as trivial as possible in2513the first respect and as bad as possible in the second. In $\Dyn$ the ergodic states are not preserved by2514arbitrary morphisms; \cref{thm:ergodicchannel} describes a class of morphisms that do preserve them.2515Deterministic morphisms always do, since a factor of an ergodic source is ergodic.2516 2517\section{Rate and capacity}2518 2519\begin{definition}[\cite{Khi56}, \S10]2520For a stationary source and channel, the \term{rate of transmission} is2521\[ R(\mu,\nu)=H(\mu)+H(\eta)-H(\omega)=\lim_{n\to\infty}\frac1n\,I(x_1\dots x_n;\,y_1\dots y_n). \]2522The \term{ergodic capacity} $C$ of the channel is the supremum of $R(\mu,\nu)$ over ergodic sources $\mu$.2523\end{definition}2524 2525Since $\mu\mapsto\eta$ and $\mu\mapsto\omega$ are affine, \cref{thm:affine} gives at once:2526 2527\begin{proposition}\label{prop:rateaffine}2528\Scl[\cite{Par61}] For every stationary channel, $R(\,\cdot\,,\nu)$ is an affine function on the simplex of2529stationary sources.2530\end{proposition}2531 2532\begin{theorem}\label{thm:capacities}2533\Scl[\cite{Bre60,Par61}] For a stationary channel of finite memory and finite output memory,2534$R(\,\cdot\,,\nu)$ is upper semicontinuous. Hence the supremum over all stationary sources equals the ergodic2535capacity, and it is attained by an ergodic source.2536\end{theorem}2537 2538The second statement follows from the first by \cref{cor:sup}. It answers the question Khinchin leaves open2539in his conclusion, where he remarks that capacity is usually defined carelessly and that his own definition2540may give a smaller number than a supremum over all stationary sources. It does not.2541 2542\section{Memoryless channels}2543 2544For a stochastic matrix $W\colon A\to B$ and a distribution $p$ on $A$, write $I(p,W)$ for the mutual2545information of the joint distribution $p\rtimes W$.2546 2547\begin{theorem}\label{thm:dmc}2548\Scl[\cite{Sha48}] The capacity of the memoryless channel with matrix $W$ is2549\[ C(W)=\max_pI(p,W), \]2550attained by the source of independent letters with a maximizing distribution.2551\end{theorem}2552 2553\begin{proof}2554Given the input block, the output letters are independent, so2555$I(x_1^n;y_1^n)=H(y_1^n)-\sum_kH(y_k\st x_k)\le\sum_kI(x_k;y_k)\le n\max_pI(p,W)$. Independent inputs give2556equality.2557\end{proof}2558 2559\begin{example}\label{ex:channels}2560The binary symmetric channel with flip probability $\delta$ has $C=1-\hb(\delta)$; for $\delta=0.1$ this is2561$0.531$. The binary erasure channel, which replaces its input by an erasure symbol with probability $e$,2562has $C=1-e$. The Z-channel, which sends $0$ to $0$ and flips $1$ to $0$ with probability $\delta$, has2563$C=\log\bigl(1+(1-\delta)\,\delta^{\delta/(1-\delta)}\bigr)$, which is $0.763$ for $\delta=0.1$.2564\end{example}2565 2566\section{Capacity as a function on morphisms}2567 2568Restricted to memoryless channels, capacity is a function on the morphisms of $\FinStoch$.2569 2570\begin{proposition}\label{prop:capacity}2571\Scl2572\begin{enumerate}[label=(\roman*)]2573\item $C(\id_X)=\log|X|$, and $C(W)=0$ iff $W$ factors through the one-point set.2574\item $C(V\circ W)\le\min\{C(V),C(W)\}$.2575\item $C(W\otimes W')=C(W)+C(W')$.2576\item If $W$ factors through a set with $r$ elements, then $C(W)\le\log r$.2577\end{enumerate}2578\end{proposition}2579 2580\begin{proof}2581(ii) For $X\to Y\to Z$ the data-processing inequality gives $I(X;Z)\le\min\{I(X;Y),I(Y;Z)\}$. (iii) Product2582inputs give $\ge$; conversely for the product channel2583$I(XX';YY')=H(YY')-H(Y\st X)-H(Y'\st X')\le I(X;Y)+I(X';Y')$. (iv) follows from (i) and (ii).2584\end{proof}2585 2586These are the properties of the logarithm of the rank of a matrix: it is $\log n$ on the identity of an2587$n$-dimensional space, it does not increase under composition, it is additive under tensor product, and2588it is bounded by the logarithm of the dimension of anything the matrix factors through. Capacity is a2589rank for stochastic matrices that is stable under small perturbations. The next chapter makes this exact:2590capacity is the growth rate of the largest identity onto which the tensor powers of $W$ retract, up to2591$\varepsilon$.2592 2593\begin{example}[A channel with memory]\label{ex:isi}2594Let $A=B=\{0,1\}$ and $y_n=x_n\oplus x_{n-1}\oplus z_n$, where the $z_n$ are independent bits equal to $1$ with2595probability $\delta$. This channel is stationary and without anticipation, has memory $1$, and has output2596memory $0$, since the outputs are independent given the input. It is the composite of the sliding block2597code $\phi(x)_n=x_n\oplus x_{n-1}$ with the memoryless binary symmetric channel, so its capacity is at most2598$1-\hb(\delta)$ by the inequality above. Equality holds: $\phi$ maps independent fair bits to independent2599fair bits, so the source of independent fair bits has rate of transmission $1-\hb(\delta)$. The memory costs2600nothing here because $\phi$ is invertible up to one bit.