[proofplan] For a fixed past horizon $T$, we compare the coupling-from-the-past output with an auxiliary chain started at time $-T$ from the stationary distribution and then evolved using the same stored update maps $F_{-T},\dots,F_{-1}$. Stationarity implies that the auxiliary chain has law $\pi$ at time $0$ for every deterministic $T$. On the event that coalescence has occurred by time $-T$, the deterministic map from time $-T$ to time $0$ is constant, so the auxiliary time-$0$ value must equal the CFTP output regardless of its starting state. Since $\tau<\infty$ almost surely, this agreement probability tends to $1$, forcing the law of $Y$ to be the same as the constant law $\pi$. [/proofplan]

[step:Construct a stationary auxiliary chain using the same update maps]Fix $T\in\mathbb{N}$. Let $(S,2^S,\pi)$ denote the finite probability space assigning mass $\pi(x)$ to $x\in S$, and define the product extension $(\Omega_T,\mathcal{F}_T,\mathbb{P}_T)$ by $\Omega_T=\Omega\times S$, $\mathcal{F}_T=\mathcal{F}\otimes 2^S$, and $\mathbb{P}_T=\mathbb{P}\otimes\pi$. Let $q_T:\Omega_T\to\Omega$ be the projection $q_T(\omega,z)=\omega$, and regard each original random object as its pullback along $q_T$. Define \begin{align*} Z_T:\Omega_T \to S \end{align*} by $Z_T(\omega,z)=z$. Then $Z_T$ is independent of the pulled-back random maps $F_{-T},F_{-T+1},\dots,F_{-1}$ and satisfies \begin{align*} \mathbb{P}_T(Z_T=x)=\pi(x) \end{align*} for every $x\in S$. Define the auxiliary time-$0$ [random variable](/page/Random%20Variable) \begin{align*} X_{T,0}:\Omega_T \to S \end{align*} by \begin{align*} X_{T,0}=\Phi_{-T,0}(Z_T). \end{align*} More generally, define random variables \begin{align*} X_{T,k}:\Omega_T \to S \end{align*} for $k=-T,-T+1,\dots,0$ by $X_{T,-T}=Z_T$ and \begin{align*} X_{T,k+1}=F_k(X_{T,k}) \end{align*} for $k=-T,\dots,-1$. We prove by induction on $k$ that $X_{T,k}$ has distribution $\pi$ under $\mathbb{P}_T$ for every $k=-T,\dots,0$. The base case $k=-T$ is the definition of $Z_T$. Suppose $-T\leq k\leq -1$ and $X_{T,k}$ has distribution $\pi$ under $\mathbb{P}_T$. Since $X_{T,k}$ is a function of $Z_T,F_{-T},\dots,F_{k-1}$, it is independent of $F_k$. Therefore, for every $y\in S$, \begin{align*} \mathbb{P}_T(X_{T,k+1}=y)=\sum_{x\in S}\mathbb{P}_T(X_{T,k}=x)\mathbb{P}_T(F_k(x)=y). \end{align*} Using the induction hypothesis and the marginal condition for $F_k$, we obtain \begin{align*} \mathbb{P}_T(X_{T,k+1}=y)=\sum_{x\in S}\pi(x)P(x,y). \end{align*} By stationarity of $\pi$, this equals $\pi(y)$. Hence $X_{T,k+1}$ has distribution $\pi$, and induction gives \begin{align*} \mathbb{P}_T(X_{T,0}=y)=\pi(y) \end{align*} for every $y\in S$.[/step]

[guided]Fix a deterministic horizon $T\in\mathbb{N}$. We introduce an auxiliary starting state at time $-T$ whose only purpose is to test what the stored update maps would do to a stationary chain. Let $(S,2^S,\pi)$ be the finite probability space with mass function $\pi$, and define $(\Omega_T,\mathcal{F}_T,\mathbb{P}_T)$ by $\Omega_T=\Omega\times S$, $\mathcal{F}_T=\mathcal{F}\otimes 2^S$, and $\mathbb{P}_T=\mathbb{P}\otimes\pi$. Let $q_T:\Omega_T\to\Omega$ be the projection $q_T(\omega,z)=\omega$; the original maps, $\tau$, and $Y$ are henceforth their pullbacks along $q_T$. Define \begin{align*} Z_T:\Omega_T \to S \end{align*} by $Z_T(\omega,z)=z$. Then $Z_T$ is independent of $F_{-T},F_{-T+1},\dots,F_{-1}$ and satisfies \begin{align*} \mathbb{P}_T(Z_T=x)=\pi(x) \end{align*} for every $x\in S$. The product extension does not change the law of the stored maps or of the CFTP output; it only supplies an independent stationary initial state. Define \begin{align*} X_{T,k}:\Omega_T \to S \end{align*} for $k=-T,-T+1,\dots,0$ as follows. Set $X_{T,-T}=Z_T$, and for each $k=-T,\dots,-1$ set \begin{align*} X_{T,k+1}=F_k(X_{T,k}). \end{align*} Thus $X_{T,0}=\Phi_{-T,0}(Z_T)$. We now verify that this auxiliary process has transition matrix $P$. Fix $k\in\{-T,\dots,-1\}$. The random variable $X_{T,k}$ depends only on $Z_T,F_{-T},\dots,F_{k-1}$, while the map $F_k$ is independent of all these random objects. Hence $X_{T,k}$ is independent of $F_k$. For every $y\in S$, partitioning over the possible values $x\in S$ of $X_{T,k}$ gives \begin{align*} \mathbb{P}_T(X_{T,k+1}=y)=\sum_{x\in S}\mathbb{P}_T(X_{T,k}=x)\mathbb{P}_T(F_k(x)=y). \end{align*} The hypothesis on the random maps says exactly that \begin{align*} \mathbb{P}_T(F_k(x)=y)=P(x,y) \end{align*} for every $x,y\in S$. Now use induction. At time $-T$, the law of $X_{T,-T}$ is $\pi$ by construction. If $X_{T,k}$ has law $\pi$, then the preceding formula gives \begin{align*} \mathbb{P}_T(X_{T,k+1}=y)=\sum_{x\in S}\pi(x)P(x,y). \end{align*} The stationarity condition for $\pi$ says that this last sum is $\pi(y)$. Therefore $X_{T,k+1}$ also has law $\pi$. Induction from $k=-T$ to $k=0$ proves \begin{align*} \mathbb{P}_T(X_{T,0}=y)=\pi(y) \end{align*} for every $y\in S$.[/guided]

[step:Show that coalescence by time $-T$ forces agreement with the CFTP output]Let \begin{align*} C_T=\{\omega_T\in\Omega_T:\tau(\omega_T)\leq T\} \end{align*} be the event that coalescence has occurred by horizon $T$ on the product extension. On $C_T$, put $t=\tau(\omega_T)$. Since $t\leq T$ and the formalized statement defines $\Phi_{a,a}=\operatorname{id}_S$, associativity of composition and the definition of $\Phi_{a,b}$ on adjacent time intervals give the factorization \begin{align*} \Phi_{-T,0}(\omega_T)=\Phi_{-t,0}(\omega_T)\circ \Phi_{-T,-t}(\omega_T). \end{align*} When $t=T$, this reads $\Phi_{-T,0}(\omega_T)=\Phi_{-T,0}(\omega_T)\circ\operatorname{id}_S$. The map $\Phi_{-t,0}(\omega_T):S\to S$ is constant by the definition of $\tau$, with common value $Y(\omega_T)$. Therefore $\Phi_{-T,0}(\omega_T)$ is also constant with common value $Y(\omega_T)$. In particular, \begin{align*} X_{T,0}(\omega_T)=\Phi_{-T,0}(\omega_T)(Z_T(\omega_T))=Y(\omega_T) \end{align*} on $C_T$. Hence \begin{align*} \mathbb{P}_T(X_{T,0}\neq Y)\leq \mathbb{P}_T(\tau>T)=\mathbb{P}(\tau>T). \end{align*}[/step]

[guided]Let \begin{align*} C_T=\{\omega_T\in\Omega_T:\tau(\omega_T)\leq T\} \end{align*} be the event that all starting states have coalesced by the deterministic horizon $T$. The variables $\tau$ and $Y$ are the pullbacks from $\Omega$ to $\Omega_T$, so this event is measured in the product extension. Take $\omega_T\in C_T$ and set $t=\tau(\omega_T)$. Then $t\leq T$. The composition from time $-T$ to time $0$ splits at the intermediate time $-t$: by associativity of composition and by the definition of $\Phi_{a,b}$ on adjacent time intervals, \begin{align*} \Phi_{-T,0}(\omega_T)=\Phi_{-t,0}(\omega_T)\circ \Phi_{-T,-t}(\omega_T). \end{align*} If $t=T$, then $\Phi_{-T,-t}=\Phi_{-T,-T}=\operatorname{id}_S$, so the same formula remains valid. By the definition of $\tau$, the map $\Phi_{-t,0}(\omega_T):S\to S$ is constant on $S$. Its common value is exactly $Y(\omega_T)$ by the definition of the CFTP output. Therefore composing $\Phi_{-t,0}(\omega_T)$ with any map into $S$, in particular with $\Phi_{-T,-t}(\omega_T)$, still gives a constant map with common value $Y(\omega_T)$. Hence \begin{align*} X_{T,0}(\omega_T)=\Phi_{-T,0}(\omega_T)(Z_T(\omega_T))=Y(\omega_T) \end{align*} on $C_T$. Thus disagreement can occur only on $C_T^c=\{\tau>T\}$, and consequently \begin{align*} \mathbb{P}_T(X_{T,0}\neq Y)\leq \mathbb{P}_T(\tau>T)=\mathbb{P}(\tau>T). \end{align*} The last equality holds because $\tau$ is pulled back from the original probability space and $\mathbb{P}_T=\mathbb{P}\otimes\pi$.[/guided]

[step:Pass to the limit in the agreement probability]Because $\tau<\infty$ almost surely, the events $\{\tau>T\}$ decrease to the null event $\{\tau=\infty\}$. By [continuity of probability](/theorems/1107) from above, applied to the decreasing sequence of events $\{\tau>T\}$ whose probabilities are finite because they are at most $1$, we have \begin{align*} \lim_{T\to\infty}\mathbb{P}(\tau>T)=\mathbb{P}(\tau=\infty)=0. \end{align*} Combining this with the previous step gives \begin{align*} \lim_{T\to\infty}\mathbb{P}_T(X_{T,0}\neq Y)=0. \end{align*}[/step]

[guided]The bound from the previous step reduces the problem to showing that the probability of missing coalescence by horizon $T$ tends to zero. Since $\tau<\infty$ almost surely, the events $\{\tau>T\}$ form a decreasing sequence whose intersection is $\{\tau=\infty\}$. Continuity of probability from above applies because each event has probability at most $1$, and gives \begin{align*} \lim_{T\to\infty}\mathbb{P}(\tau>T)=\mathbb{P}(\tau=\infty)=0. \end{align*} Combining this limit with \begin{align*} \mathbb{P}_T(X_{T,0}\neq Y)\leq \mathbb{P}(\tau>T) \end{align*} proves \begin{align*} \lim_{T\to\infty}\mathbb{P}_T(X_{T,0}\neq Y)=0. \end{align*}[/guided]

[step:Identify the law of the CFTP output with the stationary distribution]Fix $y\in S$. For every $T\in\mathbb{N}$, the pullback of $Y$ to $\Omega_T$ has the same law as $Y$ on $\Omega$, and the first step gives $\mathbb{P}_T(X_{T,0}=y)=\pi(y)$. Hence \begin{align*} |\mathbb{P}(Y=y)-\pi(y)|=|\mathbb{P}_T(Y=y)-\mathbb{P}_T(X_{T,0}=y)|. \end{align*} The two events $\{Y=y\}$ and $\{X_{T,0}=y\}$ can differ only when $Y\neq X_{T,0}$, so \begin{align*} |\mathbb{P}_T(Y=y)-\mathbb{P}_T(X_{T,0}=y)|\leq\mathbb{P}_T(Y\neq X_{T,0}). \end{align*} Taking $T\to\infty$ and using the previous step yields \begin{align*} |\mathbb{P}(Y=y)-\pi(y)|=0. \end{align*} Thus $\mathbb{P}(Y=y)=\pi(y)$ for every $y\in S$, which is exactly the assertion that the coupling-from-the-past output has distribution $\pi$.[/step]

[guided]Fix $y\in S$. The random variable $Y$ on $\Omega_T=\Omega\times S$ is the pullback of the original $Y$ on $\Omega$, so it has the same law under $\mathbb{P}_T=\mathbb{P}\otimes\pi$ as under $\mathbb{P}$. The first step proved that the auxiliary time-$0$ state has stationary law, namely \begin{align*} \mathbb{P}_T(X_{T,0}=y)=\pi(y). \end{align*} Therefore \begin{align*} |\mathbb{P}(Y=y)-\pi(y)|=|\mathbb{P}_T(Y=y)-\mathbb{P}_T(X_{T,0}=y)|. \end{align*} The only way the indicators of the two events $\{Y=y\}$ and $\{X_{T,0}=y\}$ can differ is if the two random variables themselves differ. Hence \begin{align*} |\mathbb{P}_T(Y=y)-\mathbb{P}_T(X_{T,0}=y)|\leq\mathbb{P}_T(Y\neq X_{T,0}). \end{align*} The previous step shows that the right-hand side tends to $0$ as $T\to\infty$. Since the left-hand side does not depend on $T$, it must equal $0$. Thus \begin{align*} \mathbb{P}(Y=y)=\pi(y). \end{align*} Because $y\in S$ was arbitrary, $Y$ has distribution $\pi$ on the finite state space $S$.[/guided]