{"id":"ce288199-698d-4a2a-b386-a540e87fbe6f","arxiv_id":"2412.15642","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Below a data-detectable sampling threshold, the full generator of a stroboscopically observed Markov network can be reconstructed exactly; above it, candidate-generator enumeration gives tight bounds on entropy production and affinities.","lead":"Sampling a hidden Markov network at fixed time intervals can, if the sampling is fast enough, reveal the exact transition rates and the full entropy production, not just a lower bound. When sampling is too slow, enumerating all rate matrices consistent with the data still yields tight upper and lower bounds that often beat standard estimators.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-data instability of the matrix logarithm near the π/Δt boundary undermines the practical exactness claim; no error analysis is provided for the estimated propagator or the operational rub_max criterion.","rationale":"After reading the full text, I find the mathematical core of Section 3.1 to be valid: the Gerschgorin-circle argument correctly confines the imaginary parts of eigenvalues, the principal-logarithm formula (20) follows, and the operational bounds of Section 3.2.1 are correctly derived from the propagator. The proof is self-contained and, for the idealized infinite-data case, establishes the claimed uniqueness. I also credit the authors for explicitly stating the infinite-data assumption and for acknowledging the degenerate-eigenvalue caveat in Section 3.3. The most load-bearing weakness is not an internal inconsistency but a gap between the theorem and the abstract's practical claim: exact recovery from 'available data' requires the propagator to be known exactly, and the paper contains no finite-sample error analysis. This matters because the method is advertised as a superior inference tool and the operational threshold is estimated from the same noisy data that feed the matrix logarithm. Near the threshold the principal logarithm is intrinsically ill-conditioned, so finite statistics can push eigenvalues across the branch cut and produce qualitatively wrong generators. A Monte Carlo test with finite trajectory lengths would settle whether this instability is severe in practice. I therefore retain the reader's CONDITIONAL verdict: the mathematical claims are sound under the stated idealization, but the practical exactness claim should be qualified or supplemented with error bounds.","tokens_in":19996,"tokens_out":10867,"duration_ms":98119,"concrete_test":"Simulate a 3- or 4-state Markov network with a fixed true generator L0 chosen so that rmax is close to but below π/Δt (e.g., rub_max/π between 0.9 and 1.0). For trajectory lengths T = 10^3, 10^4, 10^5, and 10^6 transitions, estimate the propagator G from empirical counts, compute the empirical rub_max and the reconstructed generator via Eq. (20), and compare the reconstructed entropy production to the true value. Report the bias, variance, and fraction of runs in which the empirical rub_max criterion is satisfied but the reconstructed σ deviates from the true σ by more than 5%. If this fraction does not vanish as T grows and increases as Δt approaches π/rmax, the exactness claim is not robust to finite data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that stroboscopic data below the operationally checkable threshold π/rub_max recover the generator and all thermodynamic quantities exactly rests on the Section 2 idealization of 'in principle infinite data'. The proof of Section 3.1 is mathematically sound for an exactly known propagator, but the abstract and the numerical comparison present the method as a practical inference tool. With finite data, GΔt is an empirical transition matrix, and every downstream quantity—Eq. (20) for L0, Eqs. (27)-(28) for rub_max, and the candidate set in Section 3.3—is a discontinuous nonlinear function of the empirical propagator. The instability is concentrated precisely in the regime where exact recovery is claimed: when rmax is just below π/Δt, the eigenvalues of ΔtL0 lie near ±iπ, so the eigenvalues of GΔt lie near the negative real axis, where the principal branch of the matrix logarithm has a cut. A small statistical fluctuation of GΔt can move an eigenvalue across the cut, changing the reconstructed generator by 2πi/Δt and producing large errors in rates and entropy production. Moreover, rub_max itself is estimated from the noisy propagator; the criterion rub_max < π/Δt can be satisfied by chance when it should not be, and no confidence interval or finite-sample guarantee is given. The numerical section does not resolve this: the candidate generators are computed from the true L0 'rather than GΔt', so the practical reconstruction pipeline is never tested end-to-end on noisy data. A secondary limitation is acknowledged: degenerate eigenvalues of GΔt produce uncountably many candidate generators, so the finite-enumeration bounds beyond the threshold are not universal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether a continuous-time Markov generator on a finite network can be reconstructed from stroboscopic observations of the propagator G_Δt = exp(Δt L0). Section 3.1 proves that if a permissible generator satisfies rmax(L0) < π/Δt, then it is the unique generator for that propagator and can be obtained constructively via the principal matrix logarithm. Section 3.2 replaces the inaccessible condition on rmax with operationally checkable sufficient criteria, in particular the bound rub_max < π/Δt, and Section 3.3 enumerates finitely many candidate generators in the generic nondegenerate case, yielding upper and lower bounds on thermodynamic quantities such as entropy production and cycle affinities. Section 4 compares this approach with a Kullback-Leibler-based entropy estimator and a conjectured affinity bound, using numerical illustrations on three small networks. The paper explicitly works under the idealized assumption of infinite data, and it also acknowledges several limitations, including the non-generic degenerate case and the absence of a full finite-sample analysis.","tokens_in":20288,"tokens_out":3846,"duration_ms":39293,"significance":"If the main theorem is correct, the paper makes a useful conceptual contribution: below an operationally detectable threshold in Δt, stroboscopic data determine the full generator and hence all thermodynamic quantities, not merely lower bounds. The proof of the uniqueness statement is self-contained, anchored in established results (Gerschgorin's theorem and Higham's matrix-logarithm theorems), and introduces no fitted parameters. The finite-candidate enumeration and the resulting tight bounds are a genuine extension beyond exact reconstruction. The numerical demonstrations support the claims in the idealized setting, but the practical significance is attenuated by the lack of finite-data error analysis and by the fact that the numerical candidate enumeration is performed from the true generator rather than from an estimated propagator.","major_comments":[{"comment":"The paper explicitly assumes 'in principle infinite data' in Section 2, but the abstract and the numerical comparison frame the method as a practical inference tool for stroboscopic measurements. No error analysis is provided for an empirically estimated G_Δt. The matrix logarithm is discontinuous when an eigenvalue of G_Δt approaches the negative real axis, and this is precisely the regime relevant when rmax is close to π/Δt; a small statistical fluctuation can shift an eigenvalue across the branch cut and change the reconstructed generator by 2πi/Δt, with correspondingly large errors in rates, entropy production, and affinities. Moreover, the criterion rub_max < π/Δt is itself evaluated from the noisy propagator, so it can be satisfied or violated by chance without any confidence guarantee. I request either a finite-sample error propagation analysis or an explicit restriction of the exactness and tightness claims to the infinite-data idealization throughout the abstract and conclusions.","section":"Section 2 and Section 4.1"},{"comment":"The numerical validation does not test the proposed inference pipeline end-to-end. The text states that 'we can compute all other candidate generators LJ directly from the randomly generated L0 rather than GΔt, which reduces numerical errors.' Thus the candidate set and the reported bounds are constructed from the true generator, bypassing both the matrix logarithm and the operationally accessible bound rub_max that the method is supposed to use. The comparison against the estimators (36) and (38) therefore illustrates ideal mathematical bounds rather than the performance of the proposed procedure on estimated propagators. Please rerun the numerics with empirical transition matrices, or state clearly in the figure captions and Section 4 that these plots assume exact knowledge of the propagator.","section":"Section 4.1, Figure 3, Figure 4"},{"comment":"The abstract says that beyond the critical interval 'we still obtain tight upper and lower bounds on these quantities that improve on extant methods.' Within the paper this claim is explicitly restricted to the generic case of pairwise nondegenerate eigenvalues of G_Δt; Section 3.3 acknowledges that degenerate eigenvalues can give uncountably many candidate generators and break the finite-enumeration bounds. Since this is a load-bearing qualification, the abstract and the summary of the main results should state the nondegeneracy assumption explicitly rather than presenting the bounds as unconditional. The scaling estimate (39) also relies on rub_max ≤ N rmax and on the number of non-real simple eigenvalues; the text should clarify the sense in which this is only a crude upper estimate.","section":"Section 3.3, Eqs. (34)-(39)"},{"comment":"The improved affinity bound (38) is labeled as conjectured and is supported only by numerical evidence for a particular four-state network. The comparison in Figure 4 uses this conjectured estimator as a benchmark, so the claim that the proposed bounds 'improve on extant methods' is conditional on a statement that has not been proved. If (38) fails for other topologies or parameter ranges, the demonstrated improvement over extant affinity estimators may not hold. At minimum, the comparison should report the dependence of the benchmark on the conjecture, or use a provable bound instead.","section":"Section 4.2, Eq. (38)"}],"minor_comments":[{"comment":"There is a typo: 'the state of the system system is measured' should read 'the state of the system is measured'.","section":"Section 2, after Eq. (7)"},{"comment":"The criterion (27) is derived in the text through Eqs. (21)-(26), but the improved criterion (28) is cited to Cuthbert without proof. Since the paper aims to be self-contained, a proof or a statement that (28) is used only as an optional refinement would help; as written, the reader cannot verify the claimed improvement without consulting Ref. [45].","section":"Section 3.2.1, Eqs. (27)-(28)"},{"comment":"The phrase '10 5 configurations' appears with a missing superscript; it should be '10^5 configurations'. The same issue occurs in the caption of Figure 4.","section":"Figure caption of Figure 3"},{"comment":"The spelling 'Gerschgorin' is used consistently, but the standard English transliteration is 'Gershgorin'; the authors may wish to align with the reference list or add a parenthetical note.","section":"Throughout, especially Section 3.1"},{"comment":"The quantity (36) is first introduced as a Kullback-Leibler divergence but then called the 'quality factor' in the figures; please define the quality factor explicitly in the main text rather than only in the figure captions.","section":"Section 4.1, paragraph after Eq. (36)"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core appears sound and the paper is well positioned within the thermodynamic-inference literature. My main concern is the gap between the infinite-data theorems and the practical claims made in the abstract and numerical comparisons: the absence of finite-sample error analysis, and the fact that the numerical bounds are generated from the true generator rather than from an estimated propagator, mean that the end-to-end method has not yet been demonstrated. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The reliance on a conjectured affinity benchmark in Eq. (38) should also be addressed, either by proving the bound for the stated topology or by softening the comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mathias — quick read of Bauer, Seifert, van der Meer. The mathematical core is solid. The uniqueness theorem in Sec. 3.1 is a clean, self-contained proof of Cuthbert's result, and the authors say so plainly. The genuinely new piece is Sec. 3.3: when the generator is not unique, enumerate the finite candidate set and get exact upper/lower bounds on entropy production and cycle affinities. That is a real improvement over one-sided lower bounds, and it works in the generic non-degenerate case. I also like the explicit example showing the bound π/rmax is tight.\n\nThe paper is honest about provenance, which matters. It cites Cuthbert and Singer–Spilerman for the core theorems, and the self-contained presentation makes the logic easy to check. I did not find circularity; the original bounds are derived from the candidate set, not fitted.\n\nThe soft spots are real but not fatal. The biggest is that everything is done in the infinite-data idealization. With finite statistics, GΔt is an empirical matrix, and the matrix logarithm is unstable exactly when rmax is close to π/Δt: eigenvalues near the negative real axis can cross the branch cut, and the reconstructed generator changes by 2πi/Δt. The paper says it assumes infinite data, but the abstract and the numerics sell this as a practical inference method, and the numerical section computes candidates from the true L0 rather than from an estimated GΔt. That leaves the end-to-end noisy pipeline untested. The operational criterion (27) is cited, not proved, and the refined bound around Eq. (29) is only a sketch. All of that is addressable; it is a completeness gap, not a wrong result.\n\nOne more minor point: the 'tight bounds' claim is fine in the model-class sense, since the extremal candidate is a permissible generator consistent with the data. But with degenerate eigenvalues the candidate set can be uncountable and the enumeration breaks down; the authors acknowledge this.\n\nWho should read it: anyone doing thermodynamic inference from discrete-time data, especially if they care about when the generator is identifiable. It deserves a serious referee. I'd send it out, with a request that the authors either add a finite-sample error analysis or at least clearly demote the practical claims to the infinite-data regime.","headline":"A sound and honest compilation of known generator-uniqueness results with a genuinely useful new bounding method, weakened mainly by the infinite-data idealization and the absence of finite-sample error analysis.","tokens_in":20834,"tokens_out":2145,"would_cite":true,"duration_ms":19304,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A16","60J27","82C31"],"pacs":["05.70.Ln","05.40.-a"],"model":"deepseek-v4-flash","headline":"Below a critical sampling rate, stroboscopic snapshots recover the full generator.","keywords":["stochastic thermodynamics","thermodynamic inference","Markov networks","generator reconstruction","matrix logarithm","entropy production","cycle affinities","stroboscopic measurements"],"falsifier":"A decisive numerical check of the claimed threshold: for the four-state cyclic generator family of Appendix A, set $\\Delta t=\\pi/r+\\varepsilon$ with $\\varepsilon>0$ and verify that $L_0$ and $L_0^T$ are distinct permissible generators with the same propagator; then set $\\Delta t=\\pi/r-\\varepsilon$ and exhaustively search the integer-shift family for a second permissible generator with $r_{\\max}<\\pi/\\Delta t$. Finding none below the threshold confirms the bound is tight; finding one would refute the uniqueness theorem.","tokens_in":19821,"feed_emoji":"⏱️","tokens_out":8707,"duration_ms":73413,"temperature":0.7,"pith_summary":"This paper argues that regularly spaced snapshots of a Markov network can sometimes be enough to reconstruct the underlying continuous-time dynamics exactly, instead of only bounding its dissipation. Its central claim is that when the observation interval $\\Delta t$ is below a critical value that can be computed from the data alone, the propagator $G_{\\Delta t}=e^{\\Delta t L_0}$ determines the generator $L_0$ uniquely via the matrix logarithm, so entropy production, cycle affinities, and all transition rates are recovered exactly. Above that value, the paper shows how to enumerate the finitely many candidate generators and turn them into tight upper and lower bounds that improve on the usual Kullback-Leibler-based estimators. The practical point is that an experimentalist can check which regime they are in from the data and, whenever the check is passed, obtain exact thermodynamics from coarse temporal data.","feed_headline":"Fast enough sampling turns sparse snapshots into exact thermodynamics","feed_subtitle":"A data-checkable threshold separates exact generator recovery from tight bounds that beat standard estimators.","key_machinery":"The load-bearing identity is the matrix exponential $G_{\\Delta t}=e^{\\Delta t L_0}$ together with the classification of all matrices whose exponential equals $G_{\\Delta t}$. The proof uses the Gerschgorin circle theorem to show that any permissible generator has all eigenvalues inside a disc of radius $r_{\\max}$ centered at $-r_{\\max}$, so eigenvalues shifted by $2\\pi i/\\Delta t$ leave the allowed region whenever $r_{\\max}<\\pi/\\Delta t$; this forces $J=0$ in the logarithm formula and makes the principal-branch logarithm the unique answer. The operationally accessible version replaces the unknown $r_{\\max}$ with an upper bound $r_{\\max}^{\\mathrm{ub}}$ derived from the propagator's diagonal entries and the trace identity. This machinery both proves exact recovery below the threshold and parametrizes the finite candidate set above it.","core_discovery":"For a continuous-time Markov jump process on a connected network observed at fixed intervals $\\Delta t$, the paper proves that a permissible generator $L_0$ satisfying $\\exp(\\Delta t L_0)=G_{\\Delta t}$ and $r_{\\max}(L_0)<\\pi/\\Delta t$ is unique and is constructively recovered as $L_0 = \\frac{1}{\\Delta t} Z \\ln(D) Z^{-1}$ from a diagonalization $G_{\\Delta t}=Z D Z^{-1}$. Since $r_{\\max}$ is not directly observable, the paper derives the operational criterion $r_{\\max}^{\\mathrm{ub}}<\\pi/\\Delta t$, computed from the diagonal of $G_{\\Delta t}$ and the trace identity $\\sum_i r_i=-\\ln\\det(G_{\\Delta t})/\\Delta t$, that guarantees the uniqueness condition is met. When the criterion fails and the eigenvalues of $G_{\\Delta t}$ are non-degenerate, all candidate generators take the form $L_J = Z\\big(\\frac{1}{\\Delta t}\\ln D + \\frac{2\\pi i}{\\Delta t}\\operatorname{diag}(j_1,\\dots,j_N)\\big)Z^{-1}$ with finitely many integer vectors $J$; evaluating thermodynamic quantities over the permissible candidates gives tight upper and lower bounds, which is an original contribution relative to the existing literature.","pith_inferences":["The paper's exactness claim assumes an exactly known propagator; for finite data, errors in $G_{\\Delta t}$ propagate through the matrix logarithm and are likely to be severe near $\\Delta t=\\pi/r_{\\max}$, so practical implementations will need regularization or uncertainty quantification that the paper does not provide.","Because the full generator is recovered below threshold, the same reconstruction could feed any generator-dependent thermodynamic or kinetic quantity, not only entropy production and affinities, for example current fluctuations or average traffic.","Degenerate eigenvalues break the finite-enumeration argument and can yield uncountably many candidate generators; detecting near-degeneracies from data and deciding how to bound thermodynamics in that case is an open problem the paper only flags implicitly.","The threshold has a physical reading: strobes must be faster than the fastest escape timescale. This suggests a practical protocol, shorten $\\Delta t$ until $r_{\\max}^{\\mathrm{ub}}\\Delta t/\\pi<1$, that makes exact inference a tunable experimental resource rather than a fixed property of the system."],"forward_implications":["For $\\Delta t<\\pi/r_{\\max}^{\\mathrm{ub}}$, all transition rates, the mean entropy production rate, and every cycle affinity are determined exactly from stroboscopic data; no lower-bound estimator is needed.","The threshold $\\pi/r_{\\max}^{\\mathrm{ub}}$ is computable from the observed propagator alone, so an experimenter can know from the data whether exact inference is guaranteed.","Above the threshold, in the generic non-degenerate case, the minimum and maximum of any rate-dependent thermodynamic quantity over the finite candidate set are tight bounds that improve on the KL-divergence estimator $\\hat{\\sigma}$.","If $G_{\\Delta t}$ has real pairwise distinct eigenvalues, the generator is unique for every $\\Delta t$; in particular, two-state systems are always exactly reconstructible.","For cycle affinities the method yields individual lower and upper bounds for every cycle, whereas the compared extant estimator offers only a single lower bound on the largest affinity."],"supporting_citations":[{"why":"Defines the mean entropy production rate and the microscopic reversibility condition used throughout.","marker":"[1]"},{"why":"Supplies the operationally accessible sufficient criterion for uniqueness, Theorem 2 and its corollary, that avoids knowing $r_{\\max}$.","marker":"[45]"},{"why":"Gerschgorin circle theorem bounds the spectrum of permissible generators and yields condition (12b).","marker":"[57]"},{"why":"Theorem 1.27 classifies all matrices whose exponential equals $G_{\\Delta t}$, the starting point for uniqueness and enumeration.","marker":"[58]"},{"why":"Establishes uniqueness for generators with real distinct eigenvalues and discusses the embedding problem.","marker":"[50]"},{"why":"Gives the KL-divergence-based entropy estimator $\\hat{\\sigma}$ that the reconstruction method is compared against.","marker":"[22]"},{"why":"Provides the conjectured bound on cycle affinities that Section 4.2 compares and improves.","marker":"[41]"},{"why":"Introduces the dissipative timescale extracted from $\\hat{\\sigma}$, used to compare thresholds.","marker":"[40]"}],"fun_headline_variants":["A data-checkable threshold separates exact generator recovery from bounds","Fast snapshots unlock exact thermodynamics; slower ones still tight bounds","Matrix logarithm yields exact dissipation when sampling is fast enough","A simple timescale check on data decides exactness vs bounds","Exact thermodynamics from fast enough snapshots, tight bounds otherwise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the propagator $G_{\\Delta t}$ is known exactly, i.e., infinite stroboscopic data; with finite sampling, the matrix logarithm becomes unstable precisely in the regime where exact recovery is claimed, and the paper does not analyze that error propagation.","fun_headline_variants_meta":{"raw":{"variants":["A data-checkable threshold separates exact generator recovery from bounds","Fast snapshots unlock exact thermodynamics; slower ones still tight bounds","Matrix logarithm yields exact dissipation when sampling is fast enough","A simple timescale check on data decides exactness vs bounds","Exact thermodynamics from fast enough snapshots, tight bounds otherwise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002237,"raw_usage":{"total_tokens":8690,"prompt_tokens":1025,"completion_tokens":7665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":7582}},"tokens_in":641,"tokens_out":7665,"duration_ms":39867,"temperature":1.0,"reasoning_tokens":7582,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:13:47.788805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive numerical check of the claimed threshold: for the four-state cyclic generator family of Appendix A, set $\\Delta t=\\pi/r+\\varepsilon$ with $\\varepsilon>0$ and verify that $L_0$ and $L_0^T$ are distinct permissible generators with the same propagator; then set $\\Delta t=\\pi/r-\\varepsilon$ and exhaustively search the integer-shift family for a second permissible generator with $r_{\\max}<\\pi/\\Delta t$. Finding none below the threshold confirms the bound is tight; finding one would refute the uniqueness theorem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the mean entropy production rate and the microscopic reversibility condition used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the operationally accessible sufficient criterion for uniqueness, Theorem 2 and its corollary, that avoids knowing $r_{\\max}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gerschgorin circle theorem bounds the spectrum of permissible generators and yields condition (12b)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Theorem 1.27 classifies all matrices whose exponential equals $G_{\\Delta t}$, the starting point for uniqueness and enumeration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes uniqueness for generators with real distinct eigenvalues and discusses the embedding problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the KL-divergence-based entropy estimator $\\hat{\\sigma}$ that the reconstruction method is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the conjectured bound on cycle affinities that Section 4.2 compares and improves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the dissipative timescale extracted from $\\hat{\\sigma}$, used to compare thresholds."}],"review_version":1}