{"id":"1bcf91d6-f40a-4744-8c58-3576220ac9df","arxiv_id":"2411.15118","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For any ergodic Markov process, asymptotic correlations between occupation times are bounded from below by a reference-independent matrix built from stationary probabilities and configuration lifetimes, with saturation for unidirectional cycles.","lead":"This paper proves a new lower bound on how much occupation times fluctuate in any random hopping process between discrete states. It gives a benchmark for simulations that estimate stationary averages and connects these fluctuations to dynamic critical exponents.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Eq. (3) is internally consistent, and the external level-2.5/CLT input is standard for finite irreducible CTMCs.","rationale":"The reader identified the level-2.5 large-deviation input as the load-bearing premise, and I agree that this is the external theorem on which the derivation rests. However, I do not find it fragile: for finite irreducible continuous-time Markov jump processes, the level-2.5 rate function (Maes-Netocny, Bertini-Faggionato-Gabrielli) and the standard CLT relation between the rate-function curvature and the asymptotic variance are established results. The rest of the proof is elementary quadratic-form manipulation, which I checked step by step. In particular, the passage from Eq. (S9) to Eq. (S27) is valid because the map from (dp, delta) to dpt is surjective, and the Cauchy-Schwarz maximization is legitimate since B is symmetric positive definite. The saturation argument for unidirectional cycles is also internally consistent. The paper has no free parameters and the core result is an analytic derivation rather than a numerical claim, so independent numerical verification would still be worthwhile but is not necessary to accept the argument. I therefore see no load-bearing error and recommend keeping the reader's ACCEPT verdict unchanged.","tokens_in":25037,"tokens_out":17199,"duration_ms":169912,"concrete_test":"For an ensemble of random ergodic CTMC generators with d = 4, 6, 8 states, compute p_ss, Lambda, B = D^{-1} Lambda, the pseudoinverse resolvent R on the Q_ss subspace, and C = -(D R + R^T D); then verify numerically that the smallest eigenvalue of C - Q_ss^T B^{-1} Q_ss is nonnegative to machine precision. Include unidirectional cycles in the ensemble and check that the largest eigenvalue of the difference is numerically zero, confirming saturation. This tests Eq. (3) directly from the definition of C without invoking level-2.5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central mOUR, C >= Q_ss^T B^{-1} Q_ss (main-text Eq. (3)), follows from the level-2.5 large-deviation principle for finite irreducible continuous-time Markov chains, the standard CLT identification I''(<O>_ss) = 1/(o^T C o), and an elementary Cauchy-Schwarz optimization. I checked the pivotal algebra in the supplement: the Schur-complement step in Eq. (S11), the change of variables dpt = dp - p_ss delta with dpt^T Q_ss = dp^T, and the bijectivity of the map (dp, delta) -> dpt are all valid. B = D^{-1} Lambda is positive definite on the relevant subspace, and C is nonsingular on the Q_ss subspace for finite ergodic chains, so the inversion and Loewner-order conclusions are sound. The saturation conditions for unidirectional cycles, Eqs. (S69)-(S73), also check out: for a unidirectional cycle the single flux variable factors out and the remaining sum vanishes by p_ss_j proportional to 1/lambda_j. The only external input is the level-2.5 theorem plus the Hessian-variance identification; both are established results for this class of processes, so I do not regard this as a fragile assumption. Minor presentational issues, such as the main text's 'uncertainty vanishing as epsilon^2' (it should say diverging as 1/epsilon^2), do not affect the derivation of Eq. (3).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies finite irreducible continuous-time Markov chains and derives lower bounds on the asymptotic variance of time-integrated system observables. Because a constant shift of a system observable changes the standard uncertainty, the authors define a reference-dependent 'relative uncertainty' and prove an observable-independent bound (Eq. (1)) from the level-2.5 large-deviation rate function for occupations and fluxes. Optimizing over references yields the central matrix occupation uncertainty relation (mOUR), Eq. (3): C >= Q_ss^T B^{-1} Q_ss, where C is the asymptotic covariance-rate matrix of occupation times, D is the diagonal stationary-probability matrix, and B = D^{-1} Lambda. The paper shows that unidirectional cycles saturate the bound, discusses implications for stochastic estimation of stationary averages, derives non-negative dynamic exponents, and gives two finite-size examples with complementary analytic calculations. The supplemental material contains the derivations, tighter reference-dependent bounds, and saturation conditions.","tokens_in":25352,"tokens_out":18972,"duration_ms":189767,"significance":"If correct, Eq. (3) is a genuinely reference-free and parameter-free limit: every ergodic continuous-time Markov chain has occupation-time correlation rates bounded below by Q_ss^T B^{-1} Q_ss, a matrix determined only by stationary probabilities and configuration lifetimes. This is both a fundamental statement and a practical benchmark for Monte Carlo sampling. The central derivation is standard level-2.5 large-deviation input followed by an elementary Cauchy-Schwarz optimization; I checked the pivotal algebra in the supplement, including the expansion in Eq. (S4), the Schur-complement step, and the change of variables that eliminates delta, and found no error. The saturation proof for unidirectional cycles is given in three independent ways, and the counting argument excluding equilibrium saturation for d >= 3 is convincing. The paper ships no code, but the proofs are explicit enough to be checked by hand. The only fragile-looking input is the identification of the Hessian of the level-1 rate function with the inverse asymptotic variance; for finite irreducible CTMCs this is an established result, so I do not regard it as a gap.","major_comments":[],"minor_comments":[{"comment":"The sentence 'with both the bound and the uncertainty vanishing as epsilon^2' should say 'diverging as 1/epsilon^2', since the denominator [(p_ss - p)^T o]^2 is of order epsilon^2. This is a typographical inversion, but it appears precisely where the tightness of the bound is first discussed.","section":"OURs, after Eq. (1)"},{"comment":"The sentence 'Redefining p = p_ss + delta p, we then find Eq. (1) holds for any p...' is terse. The perturbative derivation gives a quadratic-form inequality in delta p after dividing by epsilon^2 and taking the limit; making this step explicit would prevent a reader from thinking Eq. (1) is an analytic continuation of an infinitesimal statement.","section":"OURs, same paragraph"},{"comment":"'As delta B is positive definite' should read 'positive semi-definite'; for unidirectional cycles delta B = 0, and only semi-definiteness is needed for the bounds to be tighter. The later text in Sec. II.B uses the correct wording.","section":"Supplement I.B, Eq. (S14)"},{"comment":"The sentence 'This yields Eq. (S69)' is a typo; it should refer to Eq. (S81).","section":"Supplement III.B, end of 'Derivation of Eq. (S81)'"},{"comment":"The claim that saturation of Eq. (3) forces breaking of local detailed balance is stated compactly and delegated to Refs. [26,27]. Since this is one of the paper's interpretive claims, a short precise statement of the external theorem and its hypotheses would be helpful.","section":"Saturating OURs and mOUR"}],"recommendation":"minor_revision","confidential_remarks":"None beyond the above. The paper fits a statistical-mechanics journal; the single self-citation (Ref. [24]) is contextual and does not raise concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Eq. (3) is the real thing. The matrix bound C >= Q_ss^T B^{-1} Q_ss is genuinely reference-free, which distinguishes it from earlier TUR/KUR/uTUR matrix bounds, and the derivation from the level-2.5 rate function is explicit and checkable. I went through the supplement's Schur-complement step and the delta-tilde change of variables; they work. The Cauchy-Schwarz optimization is elementary, and the saturation proof for unidirectional cycles is convincing in all three forms. This is a solid theoretical contribution.\n\nThe paper also does something useful with the bound: it recovers the reference-dependent OUR as a worst-case, it gives a criterion for optimal sampling of stationary distributions by occupation times, and it connects the matrix bound to non-negative dynamic exponents. Those applications are less developed but reasonable. The comparison with prior work is fair; the claim that earlier matrix bounds are rank-one or reference-dependent holds up against the cited references.\n\nSoft spots: the tighter bounds in the supplement (B - delta B) are formal. The paper states B - delta B > 0 follows from strict positivity of the rate function, but doesn't give a self-contained proof, and the saturation claims for general cycles and equilibrium dynamics are supported by variable-counting arguments rather than fully rigorous derivations. That is clearly labeled in the text, so it's a minor issue. There is a small slip in the main text: the uncertainty does not vanish as epsilon^2 when the reference approaches the stationary distribution; it diverges as 1/epsilon^2. The stress-test note caught this; it doesn't affect Eq. (3). Finally, the dynamic-exponent part is more of a framework than a theorem, and Example II is explicitly trivial. None of this threatens the central result.\n\nThe citation pattern is clean; self-citations are contextual and the key external input (level-2.5 LDP) is standard. No fitted parameters, no circularity.\n\nWho is this for? Stochastic thermodynamics, large deviation theory, and Monte Carlo sampling communities. A serious referee should engage with it. My recommendation: send to peer review; with minor presentational fixes it should be publishable.","headline":"A genuine reference-free matrix bound on occupation-time fluctuations, derived cleanly from level-2.5 large deviations; minor presentational slips don't undermine the core result.","tokens_in":25855,"tokens_out":2877,"would_cite":true,"duration_ms":27388,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J27","60F10","82C31"],"pacs":["05.40.-a","05.10.Ln"],"model":"deepseek-v4-flash","headline":"Every ergodic Markov jump process obeys a reference-free lower bound on occupation-time fluctuation rates.","keywords":["occupation uncertainty relations","Markov jump processes","level-2.5 large deviations","occupation times","stochastic simulation","dynamic exponent","detailed balance","irreversible Markov chains"],"falsifier":"Take any ergodic continuous-time Markov chain with three or more configurations, diagonalize or simulate it long enough to estimate the covariance matrix $C$ exactly, and check whether any observable $o$ gives $o^{\\mathrm{T}} C o < (Q_{\\mathrm{ss}}o)^{\\mathrm{T}} B^{-1}(Q_{\\mathrm{ss}}o)$; one clean numerical counterexample would refute the bound.","tokens_in":24817,"feed_emoji":"⏱️","tokens_out":6165,"duration_ms":54451,"temperature":0.7,"pith_summary":"Stochastic trajectories spend fluctuating amounts of time in each configuration of a Markov process, and this paper asks how small those fluctuations can be. The answer it proposes is a matrix inequality: for any ergodic continuous-time Markov jump process, the asymptotic covariance matrix of occupation times is bounded from below by a matrix built only from the stationary distribution and the lifetimes of individual configurations. That bound is independent of any reference or observable choice, and it implies ordinary observable uncertainties as special cases. If correct, it gives a universal error benchmark for simulations that estimate stationary averages by occupation times, and it ties dynamic correlations to static ones, making dynamic exponents nonnegative.","feed_headline":"Markov systems obey a universal floor on occupation-time fluctuations","feed_subtitle":"The bound depends only on stationary probabilities and lifetimes, setting the error floor for stochastic simulations.","key_machinery":"The load-bearing mechanism is the level-2.5 large-deviation rate function for occupation times and transition fluxes of a continuous-time Markov chain. Expanding that rate function to second order around the stationary averages and using the standard identity that the second derivative of the rate function is the inverse asymptotic variance rate, the paper reduces the fluctuation bound to a Cauchy-Schwarz optimization in the space weighted by $B$, the diagonal matrix of decay rates divided by stationary probabilities. The resulting matrix inequality is reference-free because optimizing over the reference distribution is equivalent to projecting with $Q_{\\mathrm{ss}}$ and inverting $B$ on the orthogonal complement of the stationary vector.","core_discovery":"The paper's central discovery is that the covariance matrix $C$ of occupation-time fluctuations obeys $C \\ge Q_{\\mathrm{ss}}^{\\mathrm{T}} B^{-1} Q_{\\mathrm{ss}}$, where $B = D^{-1}\\Lambda = \\Lambda D^{-1}$ is the diagonal matrix of configuration decay rates weighted by stationary probabilities and $Q_{\\mathrm{ss}} = I - \\mathbf{1} p_{\\mathrm{ss}}^{\\mathrm{T}}$ projects off the stationary distribution. Equivalently, for any observable $o$, the asymptotic variance rate satisfies $o^{\\mathrm{T}} C o \\ge (Q_{\\mathrm{ss}}o)^{\\mathrm{T}} B^{-1}(Q_{\\mathrm{ss}}o)$. The bound is faithful for non-constant observables, invariant under constant shifts, and independent of the reference distribution that ordinary occupation uncertainty relations require. The paper proves that unidirectional cycles saturate the bound, that equilibrium dynamics generally cannot saturate it at finite size, and that detailed-balance processes can saturate it asymptotically in the thermodynamic limit, yielding zero dynamic exponents.","pith_inferences":["One could test the matrix bound directly in experiments or simulations by estimating the occupation-time covariance matrix $C$ from long trajectories of a Markov jump process and comparing it with $Q_{\\mathrm{ss}}^{\\mathrm{T}}B^{-1}Q_{\\mathrm{ss}}$ computed from the generator; violations would signal non-Markovian effects.","Because the bound depends only on $p_{\\mathrm{ss}}$ and lifetimes, it suggests a practical design criterion for irreversible samplers: tune the generator to saturate the inequality for a chosen set of observables, not just for the stationary distribution.","The same level-2.5 expansion might extend the matrix bound to time-dependent or periodically driven generators, replacing the stationary distribution by the instantaneous one, though the paper does not make that claim.","The inequality connecting global eigenmode timescales and local lifetimes could serve as a diagnostic for decoupling phenomena such as Stokes-Einstein breakdown, where the ratio of these timescales diverges."],"forward_implications":["Occupation-time estimators of stationary averages have a universal lower bound on their asymptotic variance, fixed only by $p_{\\mathrm{ss}}$ and the configuration lifetimes.","For a given stationary distribution and given decay rates, any dynamics saturating the matrix bound is an optimal sampler of the stationary distribution via occupation times; unidirectional cycles always achieve this.","The bound implies $o^{\\mathrm{T}} C o \\ge \\lambda_o^{-1} o^{\\mathrm{T}} \\Delta^2 o$, so the dynamic variance rate measured in units of the local lifetime always exceeds the static variance.","Integrated autocorrelation times satisfy $\\lambda_{\\max} \\tau_o \\ge \\lambda_o \\tau_o \\ge 1/2$, so dynamic exponents defined in the thermodynamic limit are nonnegative.","Equilibrium dynamics can saturate the bound only asymptotically, which corresponds to vanishing dynamic exponents and optimal simulation performance in that limit."],"supporting_citations":[{"why":"Provides the large-deviation framework and contraction principle used to write the rate function for occupation-time observables.","marker":"[15]"},{"why":"Supplies the standard identification of the second derivative of the rate function with the inverse asymptotic variance rate.","marker":"[16]"},{"why":"Gives the level-2.5 large-deviation rate function for occupations and fluxes that anchors the derivation of Eq. (1).","marker":"[17]"},{"why":"Provides the level-2.5 large-deviation result for continuous-time Markov chains used in the same derivation.","marker":"[18]"},{"why":"Contains the proofs of Eqs. (1)-(3) and the saturation conditions for unidirectional cycles.","marker":"[19]"},{"why":"Contributes the perturbation-based saturation analysis that the paper adapts to occupation uncertainty relations.","marker":"[25]"},{"why":"Supplies the upper bound on variance rates relative to detailed-balance dynamics, used to argue the matrix bound can saturate only away from equilibrium.","marker":"[26]"},{"why":"Extends the variance-rate comparison for non-reversible samplers, supporting the same saturation argument.","marker":"[27]"}],"fun_headline_variants":["Universal bound on occupation-time fluctuations in Markov dynamics","Occupation time uncertainties have a universal floor for Markov processes","Stochastic simulations' error floor set by universal occupation-time bound","Markov chains: universal floor on occupation time fluctuations","Occupation time variance lower bounded by a universal matrix inequality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the standard large-deviation law for Markov jump processes saying that the rate at which each configuration is visited and each transition occurs obeys the known level-2.5 rate function, whose second derivative equals the inverse of the occupation-time variance rate.","fun_headline_variants_meta":{"raw":{"variants":["Universal bound on occupation-time fluctuations in Markov dynamics","Occupation time uncertainties have a universal floor for Markov processes","Stochastic simulations' error floor set by universal occupation-time bound","Markov chains: universal floor on occupation time fluctuations","Occupation time variance lower bounded by a universal matrix inequality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3349,"prompt_tokens":959,"completion_tokens":2390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2323}},"tokens_in":575,"tokens_out":2390,"duration_ms":16294,"temperature":1.0,"reasoning_tokens":2323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:28:45.687750+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any ergodic continuous-time Markov chain with three or more configurations, diagonalize or simulate it long enough to estimate the covariance matrix $C$ exactly, and check whether any observable $o$ gives $o^{\\mathrm{T}} C o < (Q_{\\mathrm{ss}}o)^{\\mathrm{T}} B^{-1}(Q_{\\mathrm{ss}}o)$; one clean numerical counterexample would refute the bound.","supporting_citations":[{"cited_title":"Sokal, Monte Carlo Methods in Statistical Mechan- ics: Foundations and New Algorithms, inFunctional In- tegration: Basics and Applications , edited by C","cited_arxiv_id":null,"evidence_quote":"Provides the large-deviation framework and contraction principle used to write the rate function for occupation-time observables."},{"cited_title":"den Hollander, Large Deviation Theory (American Mathematical Society, 2000)","cited_arxiv_id":null,"evidence_quote":"Supplies the standard identification of the second derivative of the rate function with the inverse asymptotic variance rate."},{"cited_title":"Maes and K","cited_arxiv_id":null,"evidence_quote":"Provides the level-2.5 large-deviation result for continuous-time Markov chains used in the same derivation."},{"cited_title":"Bertini, A","cited_arxiv_id":null,"evidence_quote":"Contains the proofs of Eqs. (1)-(3) and the saturation conditions for unidirectional cycles."},{"cited_title":"Macieszczak, K","cited_arxiv_id":null,"evidence_quote":"Contributes the perturbation-based saturation analysis that the paper adapts to occupation uncertainty relations."},{"cited_title":"Shiraishi, Optimal Thermodynamic Uncertainty Re- lation in Markov Jump Processes, J","cited_arxiv_id":null,"evidence_quote":"Supplies the upper bound on variance rates relative to detailed-balance dynamics, used to argue the matrix bound can saturate only away from equilibrium."},{"cited_title":"Rey-Bellet and K","cited_arxiv_id":null,"evidence_quote":"Extends the variance-rate comparison for non-reversible samplers, supporting the same saturation argument."}],"review_version":1}