# Review: does matching the 2‑sum and 3‑sum laws force independence in a reversible HMM? **Verdict in brief:** the claim is correct and the supporting proof is valid. The only unstated condition, positive stationary probabilities, costs nothing. ## 1. Precise restatement **Setting.** - Z_t (t ∈ ℤ or t ≥ 0) is a stationary Markov chain on a finite set S, with transition matrix P and stationary law π. - The chain is reversible: π(x)P(x,y) = π(y)P(y,x). - f : S → ℝ, and X_t = f(Z_t). - Irreducibility and aperiodicity are not assumed. **Hypotheses.** - (H2) X_0 + X_1 has the same law as X_0 + X_0′, where X_0′ is an independent copy of X_0. - (H3) X_0 + X_1 + X_2 has the same law as the sum of three independent copies of X_0. **Conclusion.** (X_t) is i.i.d. **Hidden assumption.** The text assumes π > 0 everywhere. This is harmless: states with π = 0 are never visited under the stationary law, so S can be restricted to supp π. That set is closed under P almost surely, and the restricted chain is still stationary and reversible. ## 2. Step-by-step check and independent re-derivation Fix t ∈ ℝ. Put m(t) = E e^{tX_0} and W(z) = Σ_y P(z,y) e^{t f(y)}. **Two-term identity.** By the Markov property, E e^{t(X_0+X_1)} = Σ_z π(z) e^{t f(z)} W(z). This is correct. **Three-term identity.** Condition on Z_1 = z. - By the Markov property, Z_0 and Z_2 are conditionally independent given Z_1. - Z_2 has conditional law P(z,·). - Z_0 has conditional law π(y)P(y,z)/π(z), which equals P(z,y) by reversibility. Hence E[e^{t(X_0+X_2)} | Z_1 = z] = W(z)². Stationarity then gives E e^{t(X_0+X_1+X_2)} = Σ_z π(z) e^{t f(z)} W(z)². This is correct, and it is exactly where reversibility enters. **From the hypotheses to MGFs.** (H2) and (H3) give equality of laws, hence equality of the bounded-support MGFs: - the two-term expression equals m², and - the three-term expression equals m³. **The square.** Expand: Σ_z π(z) e^{tf(z)} (W(z) − m)² = m³ − 2m·m² + m²·m = 0. The last term uses Σ π e^{tf} = m. Each weight π(z)e^{tf(z)} is strictly positive, so W(z) = m(t) for every z and every t. **From MGFs to laws.** - The conditional law of X_1 given Z_0 = z has finite support. - An MGF on ℝ determines a finitely supported law. - Therefore that conditional law equals the law of X_0, for every z. **From one step to the whole sequence.** - By the Markov property and stationarity, P(X_{n+1} ∈ · | Z_0, …, Z_n) = μ, where μ is the law of X_0. This does not depend on the past. - So X_{n+1} is independent of σ(Z_0, …, Z_n), which contains σ(X_0, …, X_n). - Induction gives that (X_0, …, X_n) is i.i.d. for every n, so the whole sequence is i.i.d. In fact the argument shows something stronger: X_{n+1} is independent of the entire hidden past. **Status.** I re-derived every step and found no gap. ## 3. Attempts to break it **Reducible chains.** Take two closed classes with different output laws. Then W varies with the class, and the argument shows (H2)–(H3) must fail. This is consistent with the theorem; there is no counterexample here. **Zero-probability states.** These are handled by the restriction described in §1. **Periodic or deterministic chains.** Example: a 2-cycle with f(0) = a and f(1) = b, where a ≠ b. Then X_0 + X_1 = a + b almost surely, which is not the law of an independent sum. So (H2) fails. This is again consistent. **Necessity of reversibility.** This is not proved in the text, but the claim does not assert it. Reversibility is exactly what makes the past and future of the middle state identically distributed. Without it the three-term expression becomes Σ π e^{tf} W·W̃, where W̃ is built from the reversed kernel, and the perfect square disappears. **Necessity of the three-term condition.** Whether (H2) alone suffices is not addressed, and it is not needed for the claim. For binary outputs with two hidden states, (H2) alone already forces independence, so a counterexample would need a richer construction. I did not settle the general question. **Conclusion.** No counterexample exists; the proof is complete. ## 4. Novelty The result is related to the literature on "subindependence", that is, dependent variables whose sums have the law they would have under independence. It also echoes classical moment and quadratic-form identities for reversible chains, where E[g(Z_0)h(Z_2)] is a quadratic form in P. I do not know a specific published statement of this exact result. I would classify it as new as far as I know, while noting the technique is elementary and natural. ## 5. Fairness of the one-sentence claim The sentence matches what is proved: - "finite-state" gives the finite output support used in the MGF step; - "deterministic outputs" means X_t = f(Z_t); - "stationary reversible" is exactly what the argument uses. The positivity of π is the only omitted condition, and it is harmless. The phrase "just the two-term and three-term" is accurate. {"verdict": "correct", "novelty": "new as far as I know", "fair_summary": "yes", "confidence": 4, "one_line": "The reversibility-based perfect-square argument is valid: matching the 2-sum and 3-sum laws forces E[e^{tX_1}|Z_0=z] to be constant in z for all t, which makes the observed sequence i.i.d."}