Chapter 1: From failure configurations to a zero-entropy witness and ordered arrays

This chapter follows R Proposition 2.1 (pp.4–6) and §3.1 (pp.6–7). Starting from a minimal failure order on an arbitrary probability space, the source obtains a finite-alphabet zero-entropy factor and then ordered arrays. We develop the successive past-tail replacements, finite-entropy calculation, standard-factor coding, and compatibility of the probability laws at every level.

Fixed source, from p.4

1. Retaining actual times before taking a joining

If mixing fails at some finite order, choose the least such order $k\ge 3$. Ordinary mixing supplies order two. Failure at order $k$ gives fixed sets $A_{1},\ldots ,A_k$, a number $\varepsilon >0$, and time configurations, called layouts in the source,

For each of these configurations, the set correlation differs in absolute value from the product of the means by at least $\varepsilon$. Set $g_i=\mathbf1_{A_i}$. The discrepancy is real, so pass to a subsequence with fixed sign $\sigma\in\{-1,1\}$:

The sets and functions remain fixed; only the time configurations vary. Mixing already holds for each $s<k$. Finite simple-function expansion and an $L^{1}$ product-error bound extend it to fixed bounded functions.

A joining here is a joint probability law with the specified coordinate marginals, invariant under the common time action. Independence of every proper coordinate subfamily of an arbitrary joining does not force all nonzero correlations into a zero-entropy factor. For example, let $U$ and $V$ be independent two-sided fair $\pm1$ Bernoulli processes and set $W_n=U_nV_n$. Each pair of processes $U,V$, $U,W$, and $V,W$ is independent, yet $U_{0}V_{0}W_{0}=1$ identically, so the threefold joining is not a product.

Each individual Bernoulli process has trivial past tail. A tail event is independent of every finite-coordinate cylinder; a monotone-class argument extends this to the full process $\sigma$-algebra. The event is therefore independent of itself and has probability zero or one. The conditional expectation of each process's zeroth sign given its tail is zero. Tail replacement based only on the joining's independent proper marginals would incorrectly turn correlation one into zero.

This example is not a counterexample to R. It shows that the reduction must also use actual time separation within one system. R estimates the original time configurations directly before passing to joining or array limits.

2. Working directly on an original space that need not be standard

Let $\alpha$ be the finite partition generated by $A_{1},\ldots ,A_k$, and set $\alpha _i=T^{-i}\alpha$. It suffices to consider

and its probability completion $\Gamma$. The algebra $\Gamma _{0}$ is countably generated; its completion $\Gamma$ is not asserted to be countably generated as a literal collection of sets. Modulo null sets, they describe the same countably generated factor at the measure-algebra level.

To realize this factor concretely, code $\omega$ by its complete $\alpha$-name: $\pi _\alpha (\omega)_i$ is the atom containing $T^i\omega$ in the partition $\alpha$. This map into the two-sided finite-alphabet sequence space is measurable and intertwines the invertible left shift. The original space need not be isomorphic to a standard space, and no regular conditional distributions on it are used.

Tail replacement and entropy calculations can still be made directly on $(\Omega ,\Gamma ,\mu )$. Conditional expectation and forward and reverse martingale convergence hold on arbitrary probability spaces.

Complete all the following $\sigma$-algebras first, and define

The tail $\mathcal C$ is a completed sub-$\sigma$-algebra. No separate assertion that completion commutes with infinite intersections is needed; apply the decreasing martingale theorem directly to this definition.

By $\alpha _i=T^{-i}\alpha$, pullback under $T$ shifts the cutoffs of these past $\sigma$-algebras by one. Invertibility makes pullback compatible with intersection, while adding or removing finitely many initial cutoffs leaves the tail intersection unchanged. Thus $\mathcal C$ is invariant under both $T$ and $T^{-1}$: every integer translate of a $\mathcal C$-measurable function is again $\mathcal C$-measurable.

3. Replacing the latest time first: the key $L^{1}$ estimate

Set $\bar g_i=\mathbb E[g_i|\mathcal C]$, with $0\le \bar g_i\le 1$. In the order $i=k,k-1,\ldots ,1$, replace $g_i\circ T^{d_i(b)}$ by $\bar g_i\circ T^{d_i(b)}$.

At replacement $i$, functions earlier than $i$ are still the original $g_j$, while those later than $i$ are already tail functions $\bar g_j$. Using measure preservation and invertibility of $T$, shift the current time $d_i(b)$ to zero. The difference at this step becomes

where

For $j<i$, we have $d_j-d_i\le -g(b)$, and $g_j$ is $\alpha _{0}$-measurable, so this factor is $\mathcal F_{-g(b)}$-measurable. For $j>i$, two-sided invariance of the tail $\mathcal C$ makes the factor $\mathcal C$-measurable and hence $\mathcal F_{-g(b)}$-measurable. Thus $H_{i,b}$ is $\mathcal F_{-g(b)}$-measurable and satisfies $|H_{i,b}|\le 1$.

Conditional expectation gives

The last step is reverse martingale convergence, since $\mathcal F_{-m}\downarrow \mathcal C$ and $g_i$ is integrable.

The test $H_{i,b}$ may vary with $b$. The estimate is uniform over all tests bounded in absolute value by $1$ and measurable with respect to $\mathcal F_{-g(b)}$. It requires neither convergence of $H_{i,b}$ nor a mixing rate uniform over varying functions. This is why the actual time configurations matter.

At $i=1$, all other factors already belong to $\mathcal C$, so the error is in fact zero, though the same bound suffices. Summing the finitely many $k$ steps gives total error tending to zero. Conditional expectation preserves means, so the tail functions retain a nonzero discrepancy of fixed sign between their product correlation and the product of their means.

Replacing in reverse chronological order is essential. Replacing an earlier factor while retaining original future factors would generally leave the remaining product outside the remote-past $\sigma$-algebra, invalidating the conditional-expectation identity.

4. Why every finite tail partition has zero entropy rate

There is no initial appeal to a Pinsker theorem, nor any need to identify $\mathcal C$ with the full Pinsker $\sigma$-algebra. R proves the precise statement that every finite $\mathcal C$-measurable partition $\beta$ has symbol entropy rate $h(\beta )=0$.

4.1 Past conditional entropy for a finite partition

For any finite partition $\gamma$, set $\gamma _i=T^{-i}\gamma$. Stationarity and the entropy chain rule give

The conditional entropies on the right decrease. Apply increasing martingale convergence to the conditional probabilities of the finitely many $\gamma$ atoms, then bounded continuity of Shannon entropy on the finite probability simplex. Their limit is

Cesàro averaging gives existence of the entropy rate and source equation (2):

Every entropy used is Shannon entropy of a finite partition, with convention $0\log0=0$ and upper bound $\log|\gamma|$. No infinite entropies are subtracted.

4.2 Adding $\beta$ does not increase the entropy rate of $\alpha$

The partition $\beta$ is finite and $\Gamma$-measurable. Finite two-sided $\alpha$-symbol windows generate increasing $\sigma$-algebras that together generate $\Gamma _{0}$. Completion changes only null sets, so the conditional probabilities of the $\beta$ atoms converge to their indicators. For any $\varepsilon' >0$, choose a fixed $M$ such that

Stationarity gives the same inequality at each shift $i$. Hence

Fix $M$ before taking $n\to \infty$. Divide by $n$. The first term still converges to $h(\alpha )$, so $h(\alpha \vee \beta )\le h(\alpha )+\varepsilon'$. Refinement gives the reverse inequality. Letting $\varepsilon' \downarrow 0$ yields

4.3 The entropy of $\beta$ must also enter that same joint rate

Every integer translate of $\beta$ belongs to the invariant tail $\mathcal C$. At any time $i$, the tail $\mathcal C$ lies in the full past of $\alpha$, namely $\sigma (\alpha _j:j<i)$. Therefore

Although $\beta _{1},\ldots ,\beta _n$ apparently includes future times, tail measurability puts all these partitions in the full past before any fixed time $i$ of the process generated by $\alpha$. This determines the direction of the next inequality.

More conditioning information decreases conditional entropy, so

The joint entropy chain rule then gives

Passing to entropy rates gives $h(\alpha \vee \beta )\ge h(\beta )+h(\alpha )$. The preceding equality $h(\alpha \vee \beta )=h(\alpha )$ forces $h(\beta )=0$.

This uses only finite partitions, inclusion in the past tail, and the entropy chain rule. It does not follow from independent proper coordinate marginals of a joining.

5. A finite partition preserving the witness is enough

For each $\bar g_i$, choose a finite-valued $\mathcal C$-measurable approximation $h_i$ taking values in $[0,1]$, for example by rounding downward on a fine grid. Once chosen, these functions are independent of the configuration index $b$.

For two families of functions bounded in modulus by $1$, denoted by $a_i,b_i$, telescoping the product gives

This bound is uniform over all integer time configurations. The products of means obey the same bound. Replacing $\bar g_i$ by $h_i$ therefore changes the correlation discrepancy by at most $2\sum _i\|h_i-\bar g_i\|_{1}$.

The parameter order can be explicit: discard finitely many configurations so that the tail functions' signed discrepancy is at least $\varepsilon /2$; then fix approximations with $\sum _i\|h_i-\bar g_i\|_{1}\le \varepsilon /8$. The signed discrepancy for $h_i$ is at least $\varepsilon /4$. This approximation is independent of $b$.

Let $\beta$ be the common finite partition of all $h_i$. Section 4 gives $h(\beta )=0$. We do not require $\beta$ to generate the entire $\mathcal C$; the partition $\beta$ need only distinguish these finitely many witnesses. Krieger's finite-generator theorem and its additional hypotheses are therefore unnecessary.

The completed atoms of $\beta$ have representatives in $\Gamma _{0}$. Modify these on null sets to form an actual disjoint finite partition of the whole space. Only almost-everywhere equality with the original $\mathcal C$-measurable atoms is used; the new representatives are not asserted to belong to an uncompleted tail $\sigma$-algebra. The union of all integer translates of the changed null sets is still null. Correlations, entropies, means, and subsequent name readings are unchanged. The $\beta$-name can thus use original measurable representatives without requiring regular conditional kernels on the nonstandard original space.

Take the atoms of $\beta$ as a finite alphabet $\mathcal A$. Define the name map by letting $\pi _\beta (\omega)_n$ be the atom containing $T^{n}\omega$ in $\beta$, for $n\in \mathbb Z$.

This is a measurable map into $X=\mathcal A^{\mathbb Z}$ intertwining the left shift $T_X$. Its pushforward $\mu _X$ makes the shift probability-preserving. The space $X$ is compact metrizable and standard Borel; probability completion gives a standard probability system. The shift itself is invertible and bimeasurable.

Each $h_i$ becomes a finite-valued function on $X$ depending only on the zeroth symbol. On null symbols, choose any value still in $[0,1]$. Symbol-block distributions agree with the distributions of $\beta$-name blocks, so

The normalized limit is zero. Pulling Borel sets and their time translates back through $\pi _\beta$ shows that every established mixing order $s<k$ passes to $X$. Null-set representatives extend this to the completion. Standardness of the original space is not required.

6. Centering leaves only the full $k$-point discrepancy

Set $m_i=\int h_i \,\mathrm{d}\mu _X$ and $f_i=h_i-m_i$. These are real, centered, finite-valued one-site functions satisfying $\|f_i\|_{\infty }\le 1$.

Let $I=\{1,\ldots,k\}$. Expanding the centered product gives

For $S\subsetneq I$, its cardinality is smaller than $k$, and each consecutive gap in the inherited subconfiguration is at least the original minimum gap. Its integral therefore tends to $\prod _{i\in S}m_i$; the integral of the empty product is one. Summing the constant contributions from all proper subsets $S$ and using $(1-1)^k=0$ gives

There are finitely many functions and subsets, so the $o(1)$ errors can be handled simultaneously. Discard finitely many more configurations to make the signed value at least $\varepsilon /8$. Multiply $f_{1}$ by the original fixed sign $\sigma$, put $\delta =\varepsilon /8$, and reindex the configurations. For every $b$ we obtain

This gives each stated output of Proposition 2.1: a standard two-sided finite-alphabet process, zero symbol entropy rate, mixing at every order below $k$, fixed centered finite-valued one-site functions, and a uniform positive correlation on every configuration.

7. Why the ordered arrays in §3.1 define probability laws

Fix the resulting $X$, time configurations, and functions, with $I=\{1,\ldots,k\}$. Regard $I^{0}$ as a singleton and set $\rho _{0}=\mu _X$.

Choose a free ultrafilter $\mathcal U$ on the common configuration-index set $\mathbb N$. For fixed $n\ge 1$ and $b_{1},\ldots ,b_n$, first take the actual probability law on $X^{I^n}$:

Take iterated ultralimits first in $b_n$, then $b_{n-1}$, and so on through $b_{1}$. Use the same $\mathcal U$ and time-configuration data at every level.

Taking ultralimits of probabilities of arbitrary measurable sets would not by itself prove countable additivity. R first defines the limit only on symbol cylinders.

For finitely many symbol coordinates and specified values, the probabilities lie in $[0,1]$ and have iterated ultralimits. Continuity of finite sums preserves nonnegativity, finite normalization, and consistency under deletion of coordinates. These are consistent finite-dimensional distributions on countably many finite-alphabet coordinates, so the probability extension theorem gives a unique law $\rho _n$.

Countable additivity can also be seen directly. The space $X^{I^n}$ is a countable product of finite alphabets, hence compact metrizable. Every set in its cylinder algebra is clopen and compact. A decreasing sequence from this algebra with empty intersection must already have an empty member by compactness. The finitely additive limit is therefore continuous from above at the empty set, hence a premeasure. Carathéodory extension gives a Borel probability. No general interchange of ultralimits with countable sums is assumed.

Bounded scalar ultralimits respect finite sums and products. Freeness also makes them agree with ordinary limits of convergent sequences. These are the ultrafilter properties used below; they do not interchange two limits.

8. Verifying each array property

8.1 Invariance under common time

Each finite-configuration law is invariant under simultaneous application of $T_X$ to all cells: every time is an integer power of the same $T_X$, and $\mu _X$ is invariant under $T_X$. The equality persists on cylinders in the limit. Since cylinders determine probability, $\rho _n$ is invariant under common time.

This is the diagonal action of one time transformation. Array indices are not extra time generators, and invariance under independent shifts of the cells has not been obtained.

8.2 Every fixed coordinate face has law $\rho _{n-1}$

On axis $\ell$, fix position $j_\ell =a$. Every remaining cell's time contains the same extra shift $d_a(b_\ell )$. Before taking any limit, remove it by invariance of $\mu _X$. The finite-configuration law no longer depends on $b_\ell$ and retains precisely the same formula on the other $n-1$ axes.

Now take iterated ultralimits in their original order. The limit in the absent variable $b_\ell$ can simply be omitted. The remaining indices keep their relative order, giving exactly the definition of $\rho _{n-1}$. No limits are interchanged.

Fixing any axis at any position thus gives the same lower-level marginal. There are countably many levels, and the identities come from one finite-configuration formula, rather than unrelated choices of limits at separate levels.

8.3 Joint independence of proper families of last-axis planes

Fix $J\subsetneq I$ with $s=|J|\le k-1$. For each last-axis plane $j\in J$, choose a bounded test $\Phi_j$ depending on finitely many symbol coordinates. It may inspect finitely many symbol readings anywhere in that full $(n-1)$-dimensional plane.

First fix $b_{1},\ldots ,b_{n-1}$ and define

These are fixed bounded functions as $b_n$ tends to infinity. Last-axis plane $j$ adds only the time $d_j(b_n)$, so the full product integral is

By the established mixing of order $s$, first let $b_n\to \infty$. The limit is

For $s=1$, this is just measure preservation and requires no mixing. Every gap in the chosen subconfiguration still tends to infinity. The threshold may depend on the already fixed $b_{1},\ldots ,b_{n-1}$; no uniform bound is needed here.

Take the remaining ultralimits in the prescribed order. Their finite multiplicativity preserves the product, and each factor is exactly the integral under $\rho _{n-1}$ of $\Phi _j$. Thus

Symbol cylinders generate the full Borel $\sigma$-algebra of each plane. Extend finite product tests by a monotone-class argument to obtain the product joint law of the complete planes. Any family of at most $k-1$ last-axis planes is jointly independent, not merely pairwise independent.

8.4 The witness correlation persists at level one

The level-one test $\prod_i f_i(x_i)$ reads only finitely many symbols, since each $f_i$ is finite-valued and depends only on the zeroth symbol. Its integral under $\rho _{1}$ is therefore the ultralimit of the original correlations, each of which is $\ge \delta$. Hence

Convergence on cylinders or continuous functions has not been promoted to an interchange of limits for arbitrary bounded measurable tests.

9. Ordered limits do not imply axis-permutation symmetry

The finite-configuration formula is formally symmetric under simultaneous relabeling of axes and $b$ indices, but this does not interchange the iterated limits. The elementary scalar example $\mathbf1_{\{b_1<b_2\}}$ gives one when $b_{2}$ tends to infinity first, and zero when $b_{1}$ does. The distinction holds for any free ultrafilter, since the relevant sequences are eventually constant.

Compatibility of faces on arbitrary axes therefore does not make the full $\rho _n$ invariant under axis permutations. Mixing in Section 8.3 must act on the innermost variable $b_n$, whose limit is taken first while the other parameters are fixed. That argument cannot be moved unchanged to another axis.

This is why §§3.2–3.5 still need saturated extensions and conditional locality. The initial arrays supply proper-subfamily product laws in the last-axis direction only.

10. The initial object and the classical tools used

Now take $Y=X$, use the identity as the local $X$ factor, let the parameter $U$ be a singleton, and set $m_n=\rho _n$. We have verified common-time invariance, arbitrary-face compatibility, and joint independence of proper families of last-axis planes. With constant parameters, the full-parameter conditional product law is the unconditional product law; independence of the local $X$ from the parameter is automatic. This is an admissible starting object for Definition 3.1 and the later saturation construction.

The classical inputs used here have the following scope:

1. Conditional expectation and reverse martingale convergence: on the same probability space, $g_i\in L^{1}$ and $\mathcal F_{-m}$ decreases to the defined completed tail $\mathcal C$. Convergence is in $L^{1}$; no standard original space is required.
2. Increasing martingale convergence: conditional probabilities of finite atoms and finite two-sided $\alpha$-symbol windows with their $\sigma$-algebras, used for past conditional entropy and approximation of the conditional entropy of $\beta$.
3. The finite-partition entropy chain rule and monotonicity of conditional entropy: all entropies are finite, and limits use bounded continuity on a finite simplex. The two entropy-rate inequalities in the zero-entropy reduction are written out in R and developed here.
4. Simple-function or finite quantization approximation, product telescoping, and factor pullback: these give mixing for fixed bounded functions, fixed finite-valued approximations, and transfer of mixing, without an additional dynamical structure theorem.
5. A free ultrafilter and the finite algebraic rules for bounded scalar limits; probability extension on countably many finite-alphabet coordinates, or the compact-cylinder premeasure argument above. These are used only to fix compatible array laws nonconstructively.

R's references mention Ryzhikov, Thouvenot, and de la Rue as historical precedents for the Pinsker reduction. The proof here does not use unexplained results from those references to fill a gap. In particular, it requires neither identification of $\mathcal C$ as the maximal zero-entropy factor nor a finite-generator theorem. Historical attribution and the mathematical inputs actually used should be assessed separately.