{"id":"b7b34a92-f0bb-4eb0-a700-ee98b49a2fd2","arxiv_id":"1908.09284","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A finite-state Markov chain whose state mean is nonzero has an autocorrelation that does not decay to zero, hence is not Lp integrable; the paper's derivation for the two-state case is incorrect.","lead":"Two researchers analyze when autocorrelation functions of finite-state Markov chains fail to be integrable. Their main proof contains a wrong formula for the two-state case, yet the non-integrability conclusion survives if corrected.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-state equilibrium autocorrelation in eq. (4) is wrong: the Note after eq. (3) drops the exponential term, so Lemma 1's proof of non-integrability is invalid as written.","rationale":"The paper's central observation—that a nonzero constant term in the autocorrelation makes it non-integrable—is correct and elementary, and the generalized decomposition in Section 3 is broadly right for irreducible chains. However, the two-state derivation in Section 2, which is the motivating special case, contains a concrete algebraic error: the equilibrium joint probability P{X(τ)=X(0)} is not the product of the marginals, and the exponential term cannot be discarded. The proposed concrete test settles this by direct computation. Because the proof of Lemma 1 as written is invalid, the paper does not meet the standard of a rigorous mathematical contribution, and the reader's REJECT is appropriate. The reader's weakest assumption identifies exactly this false step, so I agree with the reader's assessment and do not change the verdict.","tokens_in":7798,"tokens_out":12806,"duration_ms":121312,"concrete_test":"For the two-state CTMC with α=1, β=2 and equilibrium initial distribution, compute e^{Qτ} directly: P11 = (2+e^{-3τ})/3, P22 = (1+2e^{-3τ})/3. Then P{X(τ)=X(0)} = (2/3)P11 + (1/3)P22 = (5+4e^{-3τ})/9, so R(τ) = (1+8e^{-3τ})/9. This is not the constant (α-β)²/(α+β)² = 1/9 claimed in eq. (4). Verify that the corrected R(τ) still has Lp norm infinite for every p≥1 because its limit is 1/9 ≠ 0, confirming that Lemma 1's statement is true but its proof is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the Note after eq. (3), the paper claims that when the initial distribution equals the equilibrium distribution, the transient distribution equals the equilibrium distribution and therefore P{X(τ)=X(0)} = (α²+β²)/(α+β)². This does not follow: the transient marginal equality does not make the joint event X(τ)=X(0) independent. Using the paper's own expression (3), the correct joint probability is P{X(τ)=X(0)} = (α²+β²)/(α+β)² + (2αβ/(α+β)²)e^{-(α+β)τ}. Substituting into (1) yields R(τ) = (α-β)²/(α+β)² + (4αβ/(α+β)²)e^{-(α+β)τ}, not the constant in eq. (4). Consequently Lemma 1's proof, which relies on R being 'identically a non-zero constant', is incorrect. The conclusion that R is not in Lp for p≥1 survives because the true R has a nonzero limit, but the stated proof does not establish it. Since Section 2 is the motivating special case for the paper's central claim, this false derivation is a load-bearing defect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies autocorrelation functions of finite-state continuous-time Markov chains (CTMCs), focusing on unit ({+1,−1}) and symmetric state spaces. For the two-state chain it computes R(τ) and claims that, under a stationary initial distribution, the autocorrelation is the nonzero constant (α−β)²/(α+β)². It then uses this claim to prove that the autocorrelation is not in L_p for any p ≥ 1. A general finite-state argument is given via the spectral decomposition of the generator, in which the zero-eigenvalue term contributes a constant c = E[X(0)]E[Z], and the paper concludes non-integrability whenever c ≠ 0. The final section draws inferences about point processes associated with CTMCs and DTMCs.","tokens_in":8124,"tokens_out":11679,"duration_ms":120503,"significance":"The core structural observation—that a finite-state Markov chain autocorrelation with a nonzero limit at infinity cannot be integrable—is valid and worth stating. The identification c = E[X(0)]E[Z] in the zero-eigenvalue term is a useful and essentially correct decomposition in the diagonalizable case. However, the two-state calculation that motivates the paper is incorrect, the L_p-limit statements are mathematically confused, and the point-process claims contain a false assertion. The central theorem is salvageable, but the present manuscript would require substantial correction before it could be considered for publication.","major_comments":[{"comment":"The claim that P{X(τ)=X(0)} = (α²+β²)/(α+β)² under a stationary initial distribution does not follow from equality of the transient and equilibrium marginal distributions, because the joint event requires the transition probabilities. Using the paper's own e^{Qτ}, P{X(τ)=X(0)} = (α²+β²)/(α+β)² + 2αβ/(α+β)² e^{-(α+β)τ}, and hence R(τ) = (α−β)²/(α+β)² + 4αβ/(α+β)² e^{-(α+β)τ}, not the constant in Eq. (4). This also creates an internal contradiction: for α=β, Eq. (4) gives R(τ)=0, whereas the arbitrary-initial-distribution calculation immediately below gives R(τ)=e^{-(α+β)τ}, and the correct stationary value is e^{-2ατ}.","section":"§2, Note after Eq. (3) and Eq. (4)"},{"comment":"The paper begins by assuming the process is wide-sense stationary, but then computes R(τ)=E[X(0)X(τ)] for an arbitrary initial distribution q. For q different from the stationary distribution, E[X(t)] varies with t, so the process is not wide-sense stationary and E[X(0)X(τ)] is not a stationary lag autocorrelation. Moreover, R(τ) need not be symmetric in τ, although the paper later symmetrizes to expressions such as e^{-2α|τ|}. The non-integrability analysis for arbitrary q is therefore applied to a different object—a covariance computed from a fixed nonstationary initial law—and this distinction must be addressed explicitly.","section":"§2, first paragraph and arbitrary-initial-distribution paragraph"},{"comment":"The statement 'We now prove that for any finite state space CTMC, the autocorrelation function is not integrable' is false, as the paper itself shows in Section 2: for a unit CTMC with q=1/2 and α=β, R(τ)=e^{-2α|τ|}, which is in L_p for every p ≥ 1. The claimed 'without loss of generality' restriction to a positive state space changes the value of c; for state spaces containing negative values, c=E[X(0)]E[Z] can vanish (e.g., symmetric uniform equilibrium), so Lemma 2's conditional formulation with c ≠ 0 cannot be promoted to an unconditional statement about arbitrary finite-state CTMCs.","section":"§3.1, first paragraph and Lemma 2"},{"comment":"The L_p-limit statements are internally inconsistent. The paper states that for c ≠ 0, ∫(R(τ))^p dτ is infinite for every p ≥ 1, and then adds 'Further if |c| < 1, then ∫(R(τ))^p dτ approaches zero as p → ∞.' An integral that is infinite for every finite p cannot approach zero, and the L_p norm with the 1/p power is likewise infinite for every finite p. The repeated claim that the L_p-norm approaches a finite constant as p → ∞ is therefore unsupported and should be removed or redefined.","section":"§3.1, paragraph after Eq. (5); Lemmas 1 and 2"},{"comment":"The assertion that 'when successive visits to a state of a CTMC are stitched together, a Poisson process naturally results' is not generally true. The point process of visits to a fixed state of an irreducible CTMC is a renewal process whose interarrival times include both the sojourn time in that state and the random time needed to return, and this return time is not exponentially distributed in general. The numbered inferences in Section 4 are merely transient and equilibrium probabilities of the CTMC; they do not establish the claimed point-process characterization.","section":"§4, first paragraph"}],"minor_comments":[{"comment":"The phrase 'has not ben investigated' contains a typo; it should read 'has not been investigated.'","section":"Introduction, first paragraph"},{"comment":"The notation for left and right eigenvectors is inconsistent: Eq. (5) calls f_k a right eigenvector and g_k a left eigenvector, while in Section 2 the roles of f and g are the opposite. This should be reconciled throughout.","section":"§3.1, spectral decomposition"},{"comment":"Lemma 1 states α ≠ β, but the proof also treats the case q=1/2 and α=β; the statement should be aligned with the cases actually considered, and it should explicitly exclude or handle the integrable case q=1/2.","section":"Lemma 1 statement and proof"}],"recommendation":"major_revision","confidential_remarks":"The two-state calculation in Section 2 is the paper's motivating example, and it is wrong; however, the general c ≠ 0 non-integrability result is correct in essence and can be repaired. The L_p-limit and point-process claims need either correction or removal. In its current form the paper would not be suitable for publication, but I do not see the errors as irreparable within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my honest read of 1908.09284. The paper's main message is that for a finite-state CTMC, when the autocorrelation R(τ) has a nonzero limit as τ→∞, R cannot be in Lp for any p≥1. That statement is true. The explicit computation c=E[X(0)]E[Z] in Section 3 is a clean way to identify the limit, and the spectral decomposition used there is standard. I also think the two-state calculation for a symmetric initial distribution (q=1/2) gives the familiar random telegraph autocorrelation correctly.\n\nThe soft spot is not in the conclusion but in the key demonstration. In Section 2, when the initial distribution is the equilibrium distribution, the paper claims P{X(τ)=X(0)} = (α²+β²)/(α+β)² and hence R(τ) is a constant. That drops the exponential term in the transition matrix. The correct joint probability has an extra 2αβ e^{-(α+β)τ}/(α+β)², so R(τ) = (α-β)²/(α+β)² + 4αβ e^{-(α+β)τ}/(α+β)². Lemma 1's proof relies on R being identically a nonzero constant; as written, the proof is invalid. The lemma's conclusion survives because the limit is still nonzero when α≠β, but the derivation is wrong. The paper also contradicts itself: for α=β the constant formula gives R=0, while later the same paper uses R=e^{-2α|τ|}.\n\nSection 4's point-process \"inferences\" are paraphrases of Markov chain transition probabilities and carry no new content. There is also a notational slip in Section 3 where left and right eigenvectors are swapped, though the final c formula is correct.\n\nWho is this for? A reader looking for a one-line reason why some Markov chain autocorrelations are not integrable would get that from Section 3. But the paper is not a rigorous mathematical contribution in its current form. The authors could fix the two-state computation and reframe the paper as a short note on nonzero asymptotic autocorrelation. As is, I would not encourage the editor to spend referee time on it. Desk reject seems right; the flaw sits in the motivating example and the remaining content is too thin to justify a full review.","headline":"The general conclusion is true but nearly tautological, and the two-state calculation that is supposed to motivate it is wrong, so the paper is not publishable as it stands.","tokens_in":8599,"tokens_out":6085,"would_cite":false,"duration_ms":54119,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J27","60G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a finite-state Markov chain, any nonzero constant term in the autocorrelation function forces every finite Lp norm to be infinite.","keywords":["autocorrelation function","continuous-time Markov chain","Lp norm","integrability","unit stochastic process","equilibrium distribution","point process superposition","spectral decomposition"],"falsifier":"For $\\alpha=1$, $\\beta=3$, $q=1/4$, compute the exact two-state autocorrelation without dropping any terms: the correct expression is $R(\\tau)=1.25e^{-4\\tau}-0.25$. Since the constant $-0.25$ makes $\\int |R(\\tau)|^p d\\tau$ infinite for every $p\\geq 1$, this confirms the claim; if a full calculation for some finite-state chain with nonzero $c$ instead produced an integrable autocorrelation, the claim would be falsified.","tokens_in":7621,"feed_emoji":"♾️","tokens_out":13470,"duration_ms":117629,"temperature":0.7,"pith_summary":"This paper establishes a structural fact about continuous-time Markov chains with finitely many states: their autocorrelation function splits into a decaying exponential part and a constant offset, and whenever that offset is nonzero the whole function is not in $L_p$ for any $p \\geq 1$. The two-state $\\{-1,+1\\}$ case is worked out first, showing that the offset is the product of the mean initial state and the mean equilibrium state, and that it vanishes only for special symmetric choices such as equal transition rates with a $q=1/2$ start. The same decomposition is then derived for arbitrary finite state spaces from the spectral decomposition of the generator matrix. If the claim is right, standard tools that assume integrable autocorrelations, such as power spectral densities and some simulation variance estimates, need to be applied to such chains with extra care or replaced by generalized formulations.","feed_headline":"Constant autocorrelation term makes Markov Lp norms infinite","feed_subtitle":"If the constant offset is nonzero, correlation is not integrable, so spectral and simulation analysis needs care.","key_machinery":"The load-bearing object is the spectral expansion of the transition semigroup, $e^{Q\\tau}=\\sum_{k=1}^{N} e^{\\gamma_k \\tau} E_k$ with $E_k=\\bar{f}_k \\bar{g}_k$, where the zero eigenvalue contributes a time-independent residue matrix $\\bar{f}_N \\bar{g}_N$. Feeding this expansion into $R(\\tau)=\\sum_{i,j} i\\,j\\,q_i\\,(e^{Q\\tau})_{ij}$ isolates the constant term $c=(\\sum_i i q_i)(\\sum_j j\\pi_j)=E[X(0)]E[Z]$. Because the nonzero eigenvalues of a generator have negative real parts, their contribution $f(\\tau)$ is integrable; the zero-eigenvalue contribution is what decides integrability of the whole autocorrelation.","core_discovery":"The central claim is that for a homogeneous finite-state continuous-time Markov chain, $R(\\tau)=E[X(0)X(\\tau)]$ has the form $f(\\tau)+c$, where $f(\\tau)$ is a sum of decaying exponentials coming from the nonzero eigenvalues of the generator and $c=E[X(0)]E[Z]$ is the product of the initial mean and the equilibrium mean. When $c\\neq 0$, the constant term alone makes $R(\\tau)$ fail to belong to $L_p(\\mathbb{R})$ for every $p\\geq 1$, so the autocorrelation is not integrable in any Lebesgue sense; the $L_p$ norm instead approaches a finite constant as $p\\to\\infty$. The paper derives this explicitly for a unit two-state chain with rates $\\alpha,\\beta$: with a general initial distribution $q$, $c=2[(\\beta-\\alpha)/(\\alpha+\\beta)]q+(\\alpha-\\beta)/(\\alpha+\\beta)$, and only the choice $q=1/2$ removes the constant, leaving a purely exponential autocorrelation. It also observes the asymptotic identity $\\lim_{\\tau\\to\\infty}R(\\tau)=E[X(0)]E[Z]$, which it reads as asymptotic independence between the initial and equilibrium variables.","pith_inferences":["The non-integrability conclusion survives even if the paper's simplification that drops the exponential term when the chain starts in equilibrium is corrected: the exact equilibrium-start autocorrelation contains the same constant $c$ plus a decaying exponential, and the constant alone already forces infinite $L_p$ norm.","A natural sharpening, suggested by the paper's conjecture on symmetric state spaces, is that integrable autocorrelations occur exactly when the equilibrium mean is zero; a testable extension would search for a finite-state chain with zero equilibrium mean but nonzero initial mean whose exact $R(\\tau)$ is not integrable, which would disprove the sufficiency part of that conjecture.","If the constant offset is present, standard spectral density estimators should show a growing spike near zero frequency as the observation window lengthens; detecting that empirically would confirm non-integrability without computing $L_p$ norms.","For countably infinite CTMCs the paper states the same decomposition should generalize; a testable extension is to check that a nonzero limit $E[X(0)]E[Z]$ with finite second moments still yields non-integrability when the remaining spectrum has infinitely many negative eigenvalues."],"forward_implications":["Any finite-state chain started in equilibrium with a nonzero mean state yields an autocorrelation with infinite $L_p$ norm for every finite $p$, so its power spectral density is not defined in the usual Fourier sense.","For a two-state unit chain, the only integrable-autocorrelation cases are those with $q=1/2$; in the symmetric-rate case $\\alpha=\\beta$ this gives $R(\\tau)=e^{-2\\alpha|\\tau|}$, the classical telegraph-signal form.","The asymptotic limit $\\lim_{\\tau\\to\\infty}R(\\tau)=E[X(0)]E[Z]$ gives a practical way to detect the constant offset: a nonzero long-lag autocorrelation estimate indicates non-integrability.","When a CTMC is viewed as a superposition of point processes, the equilibrium probability of the $j$-th constituent process is $\\pi_j$ and its transient probability at time $\\tau$ is $[\\pi(0)e^{Q\\tau}]_j$, so the chain's transient and equilibrium computations transfer directly to that setting."],"supporting_citations":[{"why":"Gives the prior characterization of covariance functions of stationary unit processes that this paper extends to Lp integrability.","marker":"[3]"},{"why":"Derives covariance properties of unit processes that motivate the two-state model and the constant-offset analysis.","marker":"[2]"},{"why":"Frames the standing problem of characterizing autocorrelation functions of unit stochastic processes.","marker":"[1]"},{"why":"Supplies the result that successive visits to a CTMC state form a Poisson process, which underlies the point-process interpretation in Section 4.","marker":"[13]"}],"fun_headline_variants":["Markov autocorrelation: L^p infinite unless constant vanishes","Constant term in autocorrelation blows up L^p norms","Markov chains: autocorrelation L^p infinite for most initial states","Autocorrelation of CTMCs: L^p infinity unless q=1/2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on assuming the autocorrelation depends only on the time lag even when the chain is started away from equilibrium, and on treating the zero-eigenvalue projection as the only long-time part; if wide-sense stationarity or that projection step fails, the constant-offset formula for those cases needs re-examination.","fun_headline_variants_meta":{"raw":{"variants":["Markov autocorrelation: L^p infinite unless constant vanishes","Constant term in autocorrelation blows up L^p norms","Markov chains: autocorrelation L^p infinite for most initial states","Autocorrelation of CTMCs: L^p infinity unless q=1/2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3447,"prompt_tokens":846,"completion_tokens":2601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":2524}},"tokens_in":462,"tokens_out":2601,"duration_ms":19221,"temperature":1.0,"reasoning_tokens":2524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:51.777888+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $\\alpha=1$, $\\beta=3$, $q=1/4$, compute the exact two-state autocorrelation without dropping any terms: the correct expression is $R(\\tau)=1.25e^{-4\\tau}-0.25$. Since the constant $-0.25$ makes $\\int |R(\\tau)|^p d\\tau$ infinite for every $p\\geq 1$, this confirms the claim; if a full calculation for some finite-state chain with nonzero $c$ instead produced an integrable autocorrelation, the claim would be falsified.","supporting_citations":[{"cited_title":"Masry, On covariance functions of unit processes, SIA M Journal of Applied Mathematics 23 (1972) 28–33","cited_arxiv_id":null,"evidence_quote":"Gives the prior characterization of covariance functions of stationary unit processes that this paper extends to Lp integrability."},{"cited_title":"Shepp, Covariance of unit processes, in: Working Conf erence on Stochastic Processes, Santa Barbara, CA, 1967, pp","cited_arxiv_id":null,"evidence_quote":"Derives covariance properties of unit processes that motivate the two-state model and the constant-offset analysis."},{"cited_title":"McMillan, History of a problem, SIAM Journal of Applie d Mathematics 3 (1955) 114–128","cited_arxiv_id":null,"evidence_quote":"Frames the standing problem of characterizing autocorrelation functions of unit stochastic processes."},{"cited_title":"C ¸ inlar, Introduction to Stochastic Processes, Prentice-Hall, Inc., Englewood Cliffs, New Jersey, 1975","cited_arxiv_id":null,"evidence_quote":"Supplies the result that successive visits to a CTMC state form a Poisson process, which underlies the point-process interpretation in Section 4."}],"review_version":1}