## 1. Precise statement Let \(E\) be a finite state space and \(P\) a time-homogeneous Markov transition matrix. Assume: - \(Z=(Z_n)_{n\ge 0}\) starts in a stationary distribution \(\pi\), with \(\pi(z)>0\) for every \(z\in E\). - The chain is reversible: \[ \pi(z)P(z,y)=\pi(y)P(y,z). \] - \(f:E\to\mathbb R\) is an arbitrary deterministic function, and \(X_n=f(Z_n)\). - Writing \(\mu=\mathcal L(X_0)\), and taking independent \(Y_1,Y_2,Y_3\sim\mu\), \[ X_0+X_1\overset d=Y_1+Y_2, \qquad X_0+X_1+X_2\overset d=Y_1+Y_2+Y_3. \] Then the conclusion is that every finite-dimensional distribution of \(X\) is the corresponding product of \(\mu\). In other words, \(X\) is iid. No irreducibility, aperiodicity, injectivity of \(f\), or independence of the hidden states is assumed. The hypotheses concern **exact distributions**, not merely means and variances. Finite state space ensures that all exponential moments used below exist. ## 2. Checking the argument, followed by an independent derivation ### Checking the supplied proof Fix a real number \(\theta\), and write \[ g(z)=e^{\theta f(z)},\qquad m=\sum_z\pi(z)g(z),\qquad W=Pg. \] **The two-term identity is correct:** \[ \mathbb E[e^{\theta(X_0+X_1)}] =\sum_z\pi(z)g(z)(Pg)(z). \] **The conditional-neighbor assertion is correct.** Indeed, \[ \begin{aligned} \Pr(Z_0=i,Z_2=k\mid Z_1=j) &=\frac{\pi(i)P(i,j)P(j,k)}{\pi(j)}\\ &=P(j,i)P(j,k), \end{aligned} \] where the second equality uses detailed balance. Thus, conditional on \(Z_1=j\), the two neighbors are independent and both have transition law \(P(j,\cdot)\). Consequently, \[ \mathbb E[e^{\theta(X_0+X_1+X_2)}] =\mathbb E[g(Z_1)W(Z_1)^2] =\mathbb E[g(Z_0)W(Z_0)^2]. \] The last equality uses stationarity. This explains the otherwise slightly implicit reindexing in the text. **The transform equalities and square expansion are correct.** The assumed sum laws give \[ \mathbb E[gW]=m^2,\qquad \mathbb E[gW^2]=m^3,\qquad \mathbb E[g]=m. \] Therefore \[ \mathbb E[g(W-m)^2] =m^3-2m\,m^2+m^2m=0. \] Since every \(\pi(z)g(z)\) is strictly positive, \(W(z)=m\) at every state. **The remaining conclusions are also valid.** Equality of all these exponential transforms determines the finite-support output distributions. The final independence step should explicitly use the Markov property of the *hidden chain*, rather than assume that the output process itself is Markov. Here is a complete derivation that supplies that detail. ### Independent derivation using self-adjointness Use the inner product \[ \langle u,v\rangle_\pi=\sum_z\pi(z)u(z)v(z). \] Reversibility says that \(P\) is self-adjoint for this inner product. For fixed \(\theta\), the two transform identities are \[ A:=\langle g,Pg\rangle_\pi=m^2 \] and \[ \begin{aligned} B &:=\langle g,P(g\,Pg)\rangle_\pi\\ &=\langle Pg,g\,Pg\rangle_\pi\\ &=\sum_z\pi(z)g(z)(Pg(z))^2 =m^3. \end{aligned} \] Cauchy–Schwarz gives \[ A^2 = \left(\sum_z\pi(z)\sqrt{g(z)} \sqrt{g(z)}Pg(z)\right)^2 \le \left(\sum_z\pi(z)g(z)\right) \left(\sum_z\pi(z)g(z)(Pg(z))^2\right) =mB. \] The assumed values give equality: \(A^2=m^4=mB\). Its equality condition implies that \(Pg\) is constant. Since \(A=m^2\), that constant is \(m\). Hence \[ P(e^{\theta f})(z)=\mathbb E[e^{\theta X_0}] \quad\text{for every }z,\theta. \] Now list the distinct output values as \(a_1,\ldots,a_r\), and put \[ C_j=f^{-1}(\{a_j\}),\qquad p_j=\pi(C_j). \] The preceding identity says \[ \sum_{j=1}^r \left(P(z,C_j)-p_j\right)e^{\theta a_j}=0 \quad\text{for all real }\theta. \] The functions \(e^{\theta a_j}\) are linearly independent. For example, evaluating at \(\theta=0,1,\ldots,r-1\) gives an invertible Vandermonde matrix with distinct nodes \(e^{a_j}\). Thus \[ P(z,C_j)=p_j\qquad\text{for every }z,j. \] Let \(\mathcal F_n^Z=\sigma(Z_0,\ldots,Z_n)\). By the hidden-chain Markov property, \[ \Pr(X_{n+1}=a_j\mid\mathcal F_n^Z) =P(Z_n,C_j)=p_j. \] Conditioning further on \(\mathcal F_n^X=\sigma(X_0,\ldots,X_n)\) gives \[ \Pr(X_{n+1}=a_j\mid\mathcal F_n^X)=p_j. \] Each next output is therefore independent of the complete output past, with marginal \(\mu\). Induction proves iid. **The proof is complete.** ## 3. Attempts to break the result ### Reversibility really matters Take the hidden chain to be deterministic rotation around an eight-state directed cycle, started uniformly. Label its states, in cyclic order, by \[ 00010111. \] This is a binary de Bruijn cycle of order three: its eight consecutive cyclic triples are \[ 000,\ 001,\ 010,\ 101,\ 011,\ 111,\ 110,\ 100. \] Thus every binary triple occurs with probability \(1/8\). In particular, the two-term and three-term sums have exactly the distributions obtained from independent fair bits. Nevertheless, the output process is not iid: four consecutive zeros never occur, whereas their iid probability would be \(1/16\). This is a stationary finite-state deterministic-output counterexample **without reversibility**. It confirms that the proof uses a substantive hypothesis. ### Neither sum condition alone suffices There is a simple stationary reversible four-state family. Choose a mode once: - With probability \(r\), output a constant fair bit forever. - With probability \(1-r\), output alternating bits, with a fair initial phase. A hidden representation consists of two absorbing states and a separate two-cycle. Give the absorbing states stationary probabilities \(r/2\) each and the cycle states probabilities \((1-r)/2\) each. This chain is reversible, and all four probabilities are positive when \(0