{"id":"3f67ca0e-42bb-4d43-9e89-addf7ffdba00","arxiv_id":"2412.13515","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The measure-current large deviations rate functional of a non-reversible finite-state Markov chain admits a Gamma-expansion whose terms encode the full hierarchy of metastable scales.","lead":"This paper proves that the large deviations rate functional of a sequence of Markov chains splits into a sum of simpler rate functionals, one for each metastable time-scale. It completes a program that reads the hierarchical metastable behavior of finite-state chains directly from Gamma-convergence of rate functionals.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 2.3 is a sound conditional Γ-convergence statement, and the flow-convergence argument in Section 4 is internally consistent.","rationale":"The reader's weakest-assumption analysis identifies (2.5) as the structurally fragile premise, and I agree that this is the key hypothesis on which the tree construction and scale separation rest. However, the theorem explicitly assumes (2.5), and the proof relies on the published tree-construction and DV Γ-expansion results of [25] and [22], which are legitimate independent support under the review rules. The additional flow-convergence argument in Section 4 is new, but I checked its logic: the recovery sequence from [22] satisfies θ_n^(p) I_n(ν_n)=O(1), so I_n(ν_n)→0; then the representation (4.2)–(4.4) of the optimal current forces J*_n→J_{µ,R0} edge by edge. The Γ-liminf argument is a clean induction using Lemma 4.1 and the scale separation θ^(p-1)≺θ^(p). The auxiliary results in Section 5 (determination of the generator from the rate functional and the functional derivative computations) also appear coherent. The presentation issues noted by the reader—the P(V)×F_E typo in Theorem 1.1 and the scope remark in Remark 2.6—are minor and do not affect the main theorem. Since I found no load-bearing gap, the appropriate verdict is unchanged from the reader's conditional acceptance; no substantive revision is required beyond fixing the noted typos and clarifying the abstract's 'finishes the project' claim in light of Remark 2.6.","tokens_in":25114,"tokens_out":24161,"duration_ms":200174,"concrete_test":"Independently verify the hinge of the induction, Lemma 4.1, by re-deriving it from the definitions (2.15)–(2.16) and [22, Lemma 5.1] for a concrete two-level example: take a 3-state chain with rates satisfying (2.5) (e.g., R_n(a,b)=1, R_n(b,c)=1/n, R_n(c,a)=1/n^2, with reverse rates of comparable order) and check symbolically that the finiteness domain of I^(1) is exactly the zero set of I^(0), i.e., that finite limits occur only for J=J_{µ,R0}. If this check fails for some µ, the Γ-limsup recovery sequence construction in Section 4 would collapse.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central claim, Theorem 2.3, is conditional on the comparability hypothesis (2.5). The proof reduces the measure-current Γ-convergence to the published Donsker-Varadhan Γ-expansion of [22] via Lemma 4.1 and Remark 2.4, and the novel part—showing that the optimal current J*_n associated to the recovery sequence ν_n converges to J_{µ,R0}—is carefully argued. A potentially delicate step is the inference that I_n(ν_n)→0; this is justified not merely by ν_n→µ and I^(0)(µ)=0, but by the recovery-sequence bound θ^(p)_n I_n(ν_n)=O(1) together with θ^(p)_n→∞. Once I_n(ν_n)→0, equations (4.2)–(4.4) indeed force J*_n→J_{µ,R0}. The hierarchical zero lemma (Lemma 4.1) is consistent with the recursive construction of the measures π^(p)_j. The main limitation, already acknowledged in Remark 2.6, is that intermediate flow-only time-scales are not captured; this is an honest scope restriction, not a defect in the theorem. I could not locate an internal gap in the proof of Theorem 2.3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies sequences of irreducible continuous-time Markov chains on a fixed finite state space with a common edge set, and proves a Gamma-expansion of the level-2.5 (measure-current) large deviations rate functional I_n. Under the comparability hypothesis (2.5) on products of jump rates, the main theorem (Theorem 2.3) states that for each metastable time-scale theta_n^(p) produced by the recursive construction of [25], the rescaled functional theta_n^(p) I_n Gamma-converges to I^(p), where I^(p) is finite only when the empirical measure is a convex combination of the effective stationary measures pi_j^(p) and the empirical current equals J_{mu,R0}; in that case I^(p) reduces to the Donsker-Varadhan rate functional of the effective Markov chain X^(p)_t. The paper also proves the p=0 case (Proposition 2.1), characterizes the zero set of I^(0) (Lemma 3.1), and in Section 5 gives conditions under which the Donsker-Varadhan or measure-current rate functional determines the generator, together with counterexamples showing sharpness and formulas for the first and second functional derivatives of the Donsker-Varadhan functional.","tokens_in":25239,"tokens_out":13758,"duration_ms":113998,"significance":"If correct, the main result completes the program initiated in [8] and [22] by extending the Gamma-expansion of level-2 rate functionals to the level-2.5 measure-current setting, thereby providing a hierarchical description of metastable behavior in terms of large deviations rate functionals. The proof is conditional on the comparability hypothesis (2.5) and relies heavily on prior published results for the tree construction and for the Donsker-Varadhan Gamma-expansion, but the reduction is explicit and the novel step--showing that the optimal current J_n^* associated to the recovery sequence converges to J_{mu,R0}--is carefully argued. The paper also contains clean, falsifiable statements in Section 5: reversibility is sufficient for the Donsker-Varadhan functional to determine the generator and is not necessary in general (Example 5.4), and recurrence is sufficient for the measure-current functional and not necessary in general (Example 5.9). The derivative identities in Lemmas 5.11 and 5.14 and Proposition 5.13 generalize the classical Cramer-type formulas and are of independent interest.","major_comments":[],"minor_comments":[{"comment":"In the paragraph following Eq. (4.4), the sentence \"As nu_n -> mu and I^(0)(mu)=0 by Lemma 3.1, lim_n I_n(nu_n)=I^(0)(mu)=0\" is not a valid consequence of Gamma-convergence, since Proposition 2.1 is only a Gamma-limit and does not give pointwise convergence along arbitrary recovery sequences. The desired conclusion I_n(nu_n)->0 does follow from the preceding display, which gives theta_n^(p) I_n(nu_n) <= I^(p)(mu)<infinity together with theta_n^(p)->infinity; please rephrase the justification accordingly.","section":"Section 4, proof of Theorem 2.3"},{"comment":"The symbol n is used both as the sequence index and as the number of closed irreducible classes V_1,...,V_n of the limiting chain; this collision is confusing, especially since n_p is later introduced for the number of wells at level p. Please rename the number of classes, for instance to N or m.","section":"Section 2, Eq. (2.3)"},{"comment":"The displayed stationary weights contain a typo: the text reads \"pi_n(-1)=pi_n(-1)=a_n/n\", but it should be \"pi_n(-1)=pi_n(1)=a_n/n\". In the same remark, \"metastabe\" should be \"metastable\".","section":"Remark 2.6"},{"comment":"The compiled text contains spacing artifacts such as \"Gamma -exp ansion\" and \"Mark ovian\" in the abstract and title; these should be cleaned in the final version.","section":"Abstract and Introduction"},{"comment":"The statement \"One can derive from I(.) the values of lambda(z) and R(x,y)R(y,x)\" could be made more precise by saying \"from the full rate functional I on P(V)\"; otherwise the reader might wonder whether the claim is about the values at a single measure rather than the whole functional.","section":"Section 5, Lemma 5.1"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a strong fit for a probability journal. The dependence on the authors' own prior work [22] and [25] is substantial, but it is openly acknowledged and the hypotheses are stated explicitly, so I do not see a novelty-disclosure concern. The comparability hypothesis (2.5) is the main structural assumption; if future work could exhibit a natural example where (2.5) fails but the Gamma-expansion still holds, that would strengthen the paper, but this is not necessary for the present submission. The only technical point that needs attention is the justification of I_n(nu_n)->0 in the proof of Theorem 2.3, which is repairable in a sentence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this paper proves the Γ-expansion of the level-2.5 measure-current rate functional for non-reversible finite-state Markov chains, extending Landim's earlier Donsker-Varadhan Γ-expansion. It is a real result, but its strength is tied to the comparability hypothesis (2.5) and to heavy results from earlier papers, mostly by the same group.\n\nWhat's new and good: the main theorem, Theorem 2.3, states that θ_n^(p) I_n Γ-converges to I^(p), with I^(p) built from the DV functional of the effective chain X^(p)_t and the current forced to the stationary flow J_{µ,R0}. The reduction in Remark 2.4 is elegant: at scales p≥1 the current is too costly to deviate, so the measure-current functional collapses to the DV functional. The proof of the limsup is careful about the optimal current J*_n, using the recovery bound to get I_n(ν_n)→0 and then (4.2)–(4.4) to force J*_n→J_{µ,R0}. The hierarchical zero lemma (Lemma 4.1) is a clean structural observation. Section 5 is a genuine bonus: optimal conditions for the DV and BFG functionals to determine the generator, with explicit counterexamples, and derivative formulas (Lemma 5.11, Proposition 5.13, Lemma 5.14) that generalize the i.i.d. results. Proposition 2.1, the p=0 case, is self-contained and well argued.\n\nSoft spots: the main theorem is conditional on (2.5), comparability of all products of rates. That is a strong structural assumption; the paper cites many models that satisfy it, but it is the place where the expansion could fail. The proof also imports the entire tree construction and the DV Γ-expansion from [22,25,2], so a referee has to trust several heavy lemmas not reproduced here. There is a minor typo in Theorem 1.1: it writes P(V)×E where F_E should be intended. The abstract's 'finishes the project' claim is a little overconfident, since Remark 2.6 honestly says intermediate flow-only time-scales are not captured. That is a scope restriction, not a flaw.\n\nThis paper is for probabilists working in metastability and large deviations. The reader's take and the stress-test note align with my reading: I could not locate an internal gap, and the flow-convergence argument in Section 4 is internally consistent. I would send this to a serious referee. The main things to check are the uses of [22]'s lemmas and whether (2.5) is as benign as claimed. I'd cite it if I worked in this area.","headline":"A careful, conditional Γ-expansion theorem that completes a program; the proof leans on prior work from the same group, but the reduction is explicit and the honest scope restrictions keep it credible.","tokens_in":25921,"tokens_out":3096,"would_cite":true,"duration_ms":26295,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60F10","82C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the measure-current large deviations rate functional of a finite-state Markov chain admits a hierarchical Γ-expansion whose terms are the Donsker–Varadhan rate functionals of the effective chains at each metastable…","keywords":["large deviations","Γ-convergence","empirical current","empirical measure","metastability","Donsker–Varadhan rate functional","non-reversible Markov chains","finite-state Markov chains"],"falsifier":"Take $V=\\{1,2,3\\}$ with rates $R_n(1,2)=R_n(2,1)=R_n(2,3)=R_n(3,2)=1$, $R_n(3,1)=a_n$, $R_n(1,3)=b_n$, where $a_n/b_n$ oscillates between two values; the products $a_n$ and $b_n$ appearing in (2.5) for $m=2$ are then not comparable. If the trace-rate expressions (2.18) oscillate and $\\theta_n^{(1)} I_n$ has no $\\Gamma$-limit, the comparability hypothesis is doing the claimed work; if a $\\Gamma$-limit still exists, the theorem's hypothesis is not necessary.","tokens_in":24760,"feed_emoji":"⏳","tokens_out":11762,"duration_ms":93345,"temperature":0.7,"pith_summary":"This paper studies sequences of continuous-time Markov chains on a fixed finite state space and asks what their large deviations rate functionals reveal about metastability. Its central claim is that, under a comparability condition on the jump rates, the measure-current rate functional $I_n$ admits the hierarchical $\\Gamma$-expansion $I_n = I^{(0)} + \\sum_{p=1}^{q} \\theta_n^{(p)-1} I^{(p)}$, where the weights $\\theta_n^{(p)}$ are the metastable time scales and each $I^{(p)}$ is the Donsker–Varadhan rate functional of an effective Markov chain describing transitions among metastable wells at that scale. If the claim is right, the entire metastable hierarchy — the wells, their relative weights, and the effective dynamics at every time scale — is encoded in a single functional, and the empirical current is forced to be the stationary flow of the limiting chain on all scales beyond the first. The paper also proves optimal conditions for recovering the generator from the measure or measure-current rate functional, and derives first and second derivatives of the measure rate functional.","feed_headline":"Metastable hierarchy emerges from a single rate functional","feed_subtitle":"Each term of the expansion describes the effective Markov chain among wells; currents freeze at all long scales.","key_machinery":"The carrying object is the rooted tree of partitions of $V$ constructed recursively from the comparability hypothesis: the leaves are the recurrent classes of the limiting chain $X_t^{(0)}$, each generation is a partition of $V$, and each step coarsens the partition by lumping the recurrent classes of the effective chain $X_t^{(p)}$. The construction generates the time scales $\\theta_n^{(p)}$, the effective chains $X_t^{(p)}$, and the measures $\\pi_j^{(p)}$ used in (2.16). The expansion is proved term by term: the $p=0$ term is the rate functional of the limiting chain, and for $p\\ge1$ the argument uses the identity (1.7) that the measure-current functional minimized over currents equals the Donsker–Varadhan functional, so the measure-current $\\Gamma$-limit is inherited from the known level-2 $\\Gamma$-convergence $\\theta_n^{(p)} I_n \\to I^{(p)}$.","core_discovery":"The main theorem states that for each $1\\le p\\le q$ the scaled functional $\\theta_n^{(p)} I_n$ $\\Gamma$-converges to $I^{(p)}$, where $I^{(p)}(\\mu,J)$ is finite exactly when $J=J_{\\mu,R_0}$ and $\\mu=\\sum_j \\omega_j \\pi_j^{(p)}$ for a probability vector $\\omega$, and in that case equals the Donsker–Varadhan rate $I^{(p)}(\\omega)$ of the effective chain $X_t^{(p)}$. Thus the level-2.5 functional determines, at every metastable scale, both the asymptotic stationary weights $\\pi_j^{(p)}$ and the effective chain $X_t^{(p)}$ through its Donsker–Varadhan functional. The proof reduces the $\\Gamma$-limsup to the known $\\Gamma$-convergence of level-2 functionals by projecting onto the unique optimal current, and the $\\Gamma$-liminf follows from the nested zero structure $I^{(p)}(\\mu,J)<\\infty \\iff I^{(p-1)}(\\mu,J)=0$.","pith_inferences":["Because condition (2.5) only involves finite products of rates, one could verify it computationally for concrete model families; models satisfying it are exactly those for which the expansion's time scales are well-defined.","Remark 2.6's example shows there can be intermediate scales where the flow structure changes but the empirical measure does not; a finer expansion that keeps track of currents at every intermediate scale would capture these flow-only transitions.","The recoverability results imply an identifiability criterion for statistical inference: if all states are recurrent, long-time observations of empirical measure and current determine the jump rates uniquely; with transient states, different generators share the same rate functional.","The derivative formulas suggest that the DV rate functional can be used as a generating function for cumulants of additive functionals of the chain, analogous to how the i.i.d. rate function generates cumulants via its derivatives."],"forward_implications":["Corollary 1.2: at each scale the rescaled log-probability of any closed or open set of empirical measure-current pairs is controlled by the infimum of $I^{(p)}$, so the expansion gives quantitative estimates for the time the chain spends near or away from each metastable well.","The zeros of the $I^{(p)}$ form a nested hierarchy: $I^{(p)}(\\mu,J)<\\infty$ if and only if $I^{(p-1)}(\\mu,J)=0$, so each additional scale only refines the set of pairs that are costless at the previous scale.","For $p\\ge1$, finiteness forces $J=J_{\\mu,R_0}$; the empirical current is asymptotically the stationary flow of the limiting chain, and any other current is exponentially unlikely at long time scales.","The rate functional determines the dynamics exactly when the state space is completely recurrent (or reversible for the measure-only functional), with explicit three-state counterexamples showing the conditions cannot be weakened.","The second derivative of the DV rate functional at the stationary state is expressible through the generator's symmetric part and the asymptotic variance, extending the i.i.d. relation between rate curvature and variance to Markov chains."],"supporting_citations":[{"why":"Constructs the recursive tree, partitions, time scales and effective chains that Theorem 2.3 expands.","marker":"[25]"},{"why":"Supplies the level-2 Gamma-convergence theta_n^(p) I_n to I^(p) and the recovery sequence used in the Gamma-limsup.","marker":"[22]"},{"why":"Establishes the level-2 metastable expansion that this paper extends to measure-current functionals and yields the metastable weights via its Theorem 3.1.","marker":"[8]"},{"why":"Gives the projection identity (1.7) and the unique optimal current used to reduce the measure-current Gamma-limsup to the measure case.","marker":"[6]"},{"why":"Provides the large deviations principle for the empirical measure-current pair that defines the rate functional I_n.","marker":"[7]"},{"why":"Introduces the comparability hypothesis (2.5) and the trace-rate convergence needed for the tree construction.","marker":"[2]"},{"why":"Converts convergence of the trace process into convergence of finite-dimensional distributions of the projected chain, used in Theorem 2.2.","marker":"[24]"}],"fun_headline_variants":["Rate functional expands to expose metastable hierarchy","Γ-expansion turns rate function into a ladder of effective chains","Single functional splits into time-scale terms revealing wells","Hierarchical metastability emerges from Γ-expansion of rate functionals","Level-2.5 functional fixes each metastable scale's Markov chain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument requires the comparability hypothesis (2.5): for every $m\\ge1$, the sequences formed by products of $m$ jump rates along directed edges must be pairwise comparable, so none of their ratios may oscillate; without this, the trace rates in (2.18) can fail to converge and the recursive tree, time scales, and effective chains used in Theorem 2.3 may not be well defined.","fun_headline_variants_meta":{"raw":{"variants":["Rate functional expands to expose metastable hierarchy","Γ-expansion turns rate function into a ladder of effective chains","Single functional splits into time-scale terms revealing wells","Hierarchical metastability emerges from Γ-expansion of rate functionals","Level-2.5 functional fixes each metastable scale's Markov chain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1510,"prompt_tokens":1078,"completion_tokens":432,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":350}},"tokens_in":694,"tokens_out":432,"duration_ms":4800,"temperature":1.0,"reasoning_tokens":350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:02:59.610553+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $V=\\{1,2,3\\}$ with rates $R_n(1,2)=R_n(2,1)=R_n(2,3)=R_n(3,2)=1$, $R_n(3,1)=a_n$, $R_n(1,3)=b_n$, where $a_n/b_n$ oscillates between two values; the products $a_n$ and $b_n$ appearing in (2.5) for $m=2$ are then not comparable. If the trace-rate expressions (2.18) oscillate and $\\theta_n^{(1)} I_n$ has no $\\Gamma$-limit, the comparability hypothesis is doing the claimed work; if a $\\Gamma$-limit still exists, the theorem's hypothesis is not necessary.","supporting_citations":[{"cited_title":"Landim, T","cited_arxiv_id":null,"evidence_quote":"Constructs the recursive tree, partitions, time scales and effective chains that Theorem 2.3 expands."},{"cited_title":"Landim: Metastability from the large deviations poi nt of view: A Γ-expansion of the level two large deviations rate functional of non-reversible fini te-state Markov chains","cited_arxiv_id":null,"evidence_quote":"Supplies the level-2 Gamma-convergence theta_n^(p) I_n to I^(p) and the recovery sequence used in the Gamma-limsup."},{"cited_title":"Bertini, D","cited_arxiv_id":null,"evidence_quote":"Establishes the level-2 metastable expansion that this paper extends to measure-current functionals and yields the metastable weights via its Theorem 3.1."},{"cited_title":"Bertini, A","cited_arxiv_id":null,"evidence_quote":"Gives the projection identity (1.7) and the unique optimal current used to reduce the measure-current Gamma-limsup to the measure case."},{"cited_title":"Bertini, A","cited_arxiv_id":null,"evidence_quote":"Provides the large deviations principle for the empirical measure-current pair that defines the rate functional I_n."},{"cited_title":"Beltr´ an, C","cited_arxiv_id":null,"evidence_quote":"Introduces the comparability hypothesis (2.5) and the trace-rate convergence needed for the tree construction."},{"cited_title":"Landim, M","cited_arxiv_id":null,"evidence_quote":"Converts convergence of the trace process into convergence of finite-dimensional distributions of the projected chain, used in Theorem 2.2."}],"review_version":1}