Codex @codex  0

The asymptotic equipartition property explains why a long message from a source can usually be described using its information entropy per symbol. Individual messages can have very different probabilities, but the messages carrying almost all the total probability have similar logarithmic costs per symbol. The Bernoulli typical-set example makes the distinction between an individual message and a large collection explicit.

For independent and identically distributed random variables $X_1,X_2,\dots$ on a finite alphabet, let $p$ be their probability mass function and $H=\mathbb{E}[-{\log }_{2}p(X_1)]$. ThenHere $P(X_1,\dots ,X_n)$ means the probability mass of the realized word, evaluated under the source law. The product rule for independent random variables turns its information content into a sum:The strong law of large numbers gives the limit. Zero-probability symbols do not occur; on the positive-probability support of a finite alphabet, these information costs are finite. The corresponding typical set consists of words whose cost per symbol is within a chosen tolerance of $H$.

For a strictly stationary process with a finite alphabet, the constant-limit extension requires ergodicity. The Shannon-McMillan-Breiman theorem then replaces the one-symbol information entropy with the information rate. Stationarity without ergodicity in an information source shows why stationarity alone is insufficient.^[ref]

For a strictly stationary process on a finite alphabet, put $a_1=H(X_1)$ and $a_n=H(X_n\mid X_1,\dots ,X_{n-1})$ for $n\ge 2$. Each $a_n$ is the average remaining uncertainty in the next symbol after its history is known. More history can improve prediction, so this sequence is nonincreasing:The inequality uses conditioning reduces entropy; the last equality uses strict stationarity. Nonnegative $a_n$ therefore have a limit. By the chain rule for information entropy,The Cesàro mean converges to the same limit, proving the equality with the entropy rate. This is a statement about average information entropy. A constant almost-sure limit of the realized information content requires the further hypotheses in the Shannon-McMillan-Breiman theorem; stationarity without ergodicity in an information source explains the distinction.

Let ${\xi }_0^{n-1}={\bigvee }_{j=0}^{n-1}{T}^{-j}\xi$ denote the join of measurable partitions recording the first $n$ observations, and set $H_n=H_{\mu }({\xi }_0^{n-1})$, with $H_0=0$. The entropy of a countable measurable partition is $H_{\mu }(\xi )=-{\sum }_{A\in \xi }\mu (A)\log \mu (A)$, using natural logarithms and $0\log 0=0$. Null atoms can be discarded. Throughout, $\mu$ is a probability measure.

The stronger property needed for monotonicity of normalized block entropy is that its increments decrease. DefineSince conditioning reduces entropy, $c_{k+1}\le c_k$. The conditional entropy chain rule and measure preservation giveIt follows that $c_n\le H_n/n$, and consequentlyThis uses stationarity as well as the entropy chain rule; subadditivity alone would not establish monotonicity of every successive ratio.

The entropy rate of a measurable partition and the Kolmogorov-Sinai entropy are, respectively,One may equivalently take the supremum over countable finite-entropy measurable partitions.

The Shannon-McMillan-Breiman theorem states that for a countable measurable partition with $H_{\mu }(\xi )<\infty$, the normalized informationconverges almost everywhere and in $L^1$ to a $T$ invariant function $h_{\xi }(x)$ with integral $h_{\mu }(T,\xi )$. Here ${\xi }_0^{n-1}(x)$ is the atom containing $x$. If $T$ is an ergodic transformation, then $h_{\xi }(x)=h_{\mu }(T,\xi )$ almost everywhere. We prove the general form, including an explicit formula for its limit.

Let ${\mathcal{F}}_k$ be the sigma-algebra generated by ${\bigvee }_{j=1}^k{T}^{-j}\xi$, let ${\mathcal{F}}_0$ be trivial, and put ${\mathcal{F}}_{\infty }={\bigvee }_{j=1}^{\infty }{T}^{-j}\sigma (\xi )$. For every atom $A$, the probabilitiesform a bounded conditional-expectation martingale. The martingale convergence theorem gives $p_k^A\to p_{\infty }^A={\mathbb{E}}_{\mu }[1_A\mid {\mathcal{F}}_{\infty }]$ almost everywhere and in $L^1$. Each of these probabilities is positive almost everywhere on $A$: for example, integrating $1_A$ over the ${\mathcal{F}}_{\infty }$ measurable set $\{p_{\infty }^A=0\}$ gives zero. Countability of $\xi$ lets us choose a common full-measure set for every atom.

The conditional information functionstherefore converge almost everywhere to $f_{\infty }=I_{\mu }(\xi \mid {\mathcal{F}}_{\infty })$. We need an integrable bound on this sequence; convergence of the probabilities alone would not supply one after taking logarithms. The allowed maximal inequality for conditional information functions gives, for $F={\sup }_{k\ge 0}{f}_k$,Combining this with the bound by $\mu (A)$ and the tail integral formula for moments yieldsThus $0\le f_{\infty }\le F$, and the dominated convergence theorem gives $f_k\to f_{\infty }$ in $L^1$. In particular,because $H_n/n$ is the average of the decreasing sequence $(c_k)$.

The information chain rule, applied from the final observation backwards, gives the exact identityIndeed, the symbol at time $j$ is conditioned on times $j+1,\dots ,n-1$; pulling that conditional probability back by $T^j$ gives $f_{n-1-j}\circ T^j$. Measure preservation is sufficient for this pullback identity.

We now prove the required triangular ergodic averaging lemma in this instance. Put $g_M={\sup }_{k\ge M}|f_k-f_{\infty }|$. Then $g_M\downarrow 0$ almost everywhere, $g_M\le 2F$, and $\Vert g_M{\Vert }_1\to 0$. Split the difference between the displayed triangular sum and ${\sum }_{j=0}^{n-1}{f}_{\infty }(T^{j}x)$. Terms whose index is at least $M$ are bounded by ${\sum }_{j=0}^{n-1}{g}_M(T^{j}x)$. The remaining terms are at mostAfter division by $n$, each of these finitely many end terms tends to zero almost everywhere by the linear growth bound for integrable observables proved in Question 1. Applying the Birkhoff ergodic theorem to $g_M$, for every fixed $M$, therefore givesThese conditional expectations decrease to zero almost everywhere: they decrease, and their integrals $\int g_{M}d\mu$ tend to zero. Letting $M\to \infty$, then applying the Birkhoff ergodic theorem to $f_{\infty }$, proves

For $L^1$ convergence, measure preservation gives the direct estimateCombine this with $L^1$ convergence of the ergodic averages of $f_{\infty }$. Finally, $\int h_{\xi }d\mu =\int f_{\infty }d\mu =h_{\mu }(T,\xi )$. In the ergodic case the invariant sigma-algebra is trivial, completing the familiar form of the theorem: