{"id":"ad190b5e-c0a3-4111-9cf9-5971d79e170a","arxiv_id":"1908.03762","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Density-dependent Markov chains with finite jump set and Lipschitz rates satisfy a pathwise moderate deviation principle with a rate function that becomes quadratic when the local covariance matrix is invertible.","lead":"This paper proves a moderate deviation principle for the paths of density-dependent Markov chains, covering fluctuations between the law of large numbers and central limit theorem scales. The result gives explicit exponential rates for rare deviations in models such as the contact process, SIR epidemics, and chemical reactions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5.1 applies a minimax theorem to a non-convex compact set in the Skorokhod space; the required equality sup_g min_K H = min_K sup_g H can fail, so the upper bound for closed sets is not established.","rationale":"The reader's verdict CONDITIONAL is supported, but the reader's stated weakest_assumption (infinite A / superlinear rates) is not the load-bearing issue: the paper explicitly assumes A finite and Lipschitz gradients, so that is a boundary of the theorem rather than an internal gap. The reader's rationale, however, identifies the same Sion minimax concern that I find decisive. The minimax failure is load-bearing because Lemma 5.1 is the sole route from the exponential-martingale bound to the upper bound for compact sets, and the closed-set upper bound in Theorem 2.1 follows by exponential tightness from that compact bound. Without the interchange, the proof only gives a weaker bound of the form -sup_g inf_f H, which can be substantially smaller in rate, as the two-point compact set example shows. The result may still be true and repairable by a covering argument or by working with convex compact sets, but the paper as written does not establish it. Other components, including the lower bound, rate-function identification, and examples, appear coherent and receive credit. The verdict remains CONDITIONAL: the proof has a real gap, not a refutation of the theorem.","tokens_in":23865,"tokens_out":36250,"duration_ms":370820,"concrete_test":"Verify the minimax equality used in Lemma 5.1 on the explicit compact set K={f1,f2} subset of D([0,1],R), with f1(t)=t, f2(t)=-t, b=0, sigma=1. Compute H(f_i,g)=f_i(1)g(1)-∫ f_i g' ds - (1/2)∫ g^2 ds = ±∫ g ds - (1/2)∫ g^2 ds. Show that sup_{g in C^2} min(H(f1,g),H(f2,g)) = 0, while inf_{f in K} I(f) = 1/2. If the equality sup_g inf_f H = inf_f sup_g H is false for this K, then Lemma 5.1's minimax step is invalid and the upper bound proof needs a different argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Lemma 5.1 (Section 5) derives limsup (n/a_n^2) log P(ϑ^n in ~K) <= -sup_g inf_{f in ~K} H(f,g), and then invokes Sion's minimax theorem to replace sup_g inf_f by inf_f sup_g = inf_f I(f). This interchange is not justified. Sion's theorem requires the compact set ~K to be convex (or the function to satisfy a convex-like condition) and the ambient space to be a topological vector space. Here ~K is an arbitrary compact subset of the Skorokhod space D([0,T0],R^d), which is not a topological vector space and is generally not convex. The claimed equality is false in general: take b=0, sigma=1, T0=1 and K={f1,f2} with f1(t)=t, f2(t)=-t. Then H(f_i,g)=±∫ g - (1/2)∫ g^2, so sup_g min_i H(f_i,g)=0, whereas min_i I(f_i)=1/2. Thus the upper bound for compact sets is not proven by the argument given, and since the closed-set upper bound in Theorem 2.1 relies on Lemma 5.1 via exponential tightness, the central claim is not established as written. The finiteness/infinite-A limitation noted in the reader's weakest_assumption is a condition of the theorem, not a gap; the load-bearing issue is this minimax failure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a moderate deviation principle (MDP) for the paths of density-dependent Markov chains. With X_t the deterministic LLN limit and a_n an intermediate scale, the rescaled path \\vartheta^n_t = (X^n_t - nX_t)/a_n is shown in Theorem 2.1 to satisfy upper and lower large-deviation bounds with rate function I(f) = sup_g { f(T0)\\cdot g(T0) - \\int f\\cdot g' - \\int (b f)\\cdot g - (1/2)\\int g^T \\sigma g }, and when \\sigma_t is invertible, I(f) is identified as the quadratic action (1/2)\\int (f'-bf)^T \\sigma^{-1}(f'-bf) for absolutely continuous f. The proofs use an exponential martingale built from the generator and a generalized Girsanov theorem; the four examples (contact process, SIR, chemical reactions, Yule process) are worked out explicitly.","tokens_in":61,"tokens_out":7542,"duration_ms":275013,"significance":"If the main theorem is correct, the paper gives a clean and fairly general MDP for a widely used class of Markovian models, under weaker boundedness assumptions than the existing large-deviation results and covering unbounded-rate examples such as the Yule process. The lower-bound proof and the identification of the rate function are detailed and internally consistent, and the exponential-martingale construction is transparent and parameter-free. The main weakness is that the proof of the compact-set upper bound, which is the load-bearing step for the closed-set upper bound, relies on an unjustified minimax interchange.","major_comments":[{"comment":"The application of the Minimax Theorem from [23] to an arbitrary compact set ~K in D([0,T0],R^d) is not justified. Sion's minimax theorem requires the compact set in the first variable to be a convex subset of a topological vector space; here ~K is an arbitrary compact subset of the Skorokhod space, which is not a topological vector space under the Skorokhod topology and is generally not convex. The claimed equality sup_g inf_{f in ~K} H(f,g) = inf_{f in ~K} sup_g H(f,g) can fail. For example, with b=0, sigma=1, T0=1 and ~K={f1,f2}, f1(t)=t, f2(t)=-t, one has H(f_i,g)=±\\int g - (1/2)\\int g^2, so sup_g min_i H(f_i,g)=0, while min_i I(f_i)=1/2. Thus the compact-set upper bound is not proven, and since the closed-set upper bound in Theorem 2.1 is derived from Lemma 5.1 via exponential tightness, the main theorem is not established as written. A different upper-bound argument, or a restriction to convex compact sets together with a separately justified approximation step, would be needed.","section":"Section 5, Lemma 5.1"},{"comment":"There is a second issue with the minimax step: the function H(f,g)=L_{1,f}(g)-(1/2)L_2(g) is asserted to be continuous in f, but as a function on the Skorokhod space it is not continuous. For instance, moving a jump time of a simple function toward the endpoint changes \\int f\\cdot g' without converging to the value at the limit in the Skorokhod metric. This further obstructs the direct invocation of a minimax theorem on ~K and reinforces that the proof of Lemma 5.1 needs a genuinely different argument.","section":"Section 5, Lemma 5.1"}],"minor_comments":[{"comment":"The phrase 'without loss of generality, we assume that x0(i)/K7 is an integer' is not explained; a short approximation or rounding argument would make the reduction rigorous.","section":"Section 3, Proof of Lemma 3.1"},{"comment":"In the paragraph following Equation (2.3), the notation L2_sigma([0,T],R^d) is used with T where the horizon T0 is meant; this is a minor but confusing typo.","section":"Section 2, proof of Equation (2.2)"},{"comment":"The symbol T0 is used both for the fixed time horizon and for the set of stopping times in condition (2); this overloads notation and should be changed, for example to \\mathcal{T}.","section":"Section 5, Lemma 5.2"},{"comment":"The paper would benefit from a short discussion of why the rate function I is lower semicontinuous on the Skorokhod space; although I is a sup of affine functions in f, the lack of continuity observed above makes this non-obvious and relevant to the upper bound.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline first: this paper correctly identifies a real gap in the literature—a pathwise moderate deviation principle for density-dependent Markov chains under unbounded rates—and the lower-bound machinery, rate-function identification, and examples are all in good shape. But the proof of the upper bound has a load-bearing gap in Lemma 5.1 that needs to be repaired before Theorem 2.1 is established.\n\nWhat is actually new: previous work on large deviations for DDMC (Agazzi-Dembo-Eckmann, Chan, Pardoux-Samegni-Kepgnou, Schwartz-Weiss) covers LDPs under stronger or bounded-rate conditions; Gao-Quastel's MDP is for the symmetric exclusion process. Adapting their exponential-martingale approach to finite-dimensional jump processes with unbounded linear-growth rates is a natural and worthwhile step, and the paper executes it carefully. The lower bound, the Girsanov change of measure, the exponential tightness lemmas, and the explicit rate functions for contact, SIR, chemical reaction, and Yule models all look consistent and correct on my reading. The citation pattern is unproblematic; the paper builds on the standard references.\n\nThe soft spot is exactly where the stress-test note lands. Lemma 5.1 invokes Sion's minimax theorem to interchange sup over g and inf over an arbitrary compact set K in D([0,T0],R^d). Sion's theorem requires the compact factor to be convex, and here K is not assumed convex; D is not a topological vector space under the Skorokhod metric. More concretely, the equality sup_g inf_K H = inf_K sup_g H can fail. Take b=0, sigma=1, K={t, -t}; then sup_g min_i H = 0 while min_i I = 1/2. So the compact upper bound is not proven as written, and the closed-set upper bound inherits the defect. This is a central issue, not a cosmetic one. The finiteness of A and the linear-growth conditions in Assumptions (4)-(5) are explicit hypotheses satisfied by all four examples; they are not gaps.\n\nWho is this for: applied probabilists and modelers in epidemics and chemical reactions who want moderate-deviation rates beyond the CLT. I would send it to a serious referee, but specifically ask them to verify whether the upper bound can be repaired—for instance, by proving the compact upper bound through convexification or by a different covering argument. The result is plausibly true; the paper as written is not fully established.","headline":"A well-aimed pathwise MDP for density-dependent Markov chains with a real minimax gap in the upper bound that needs repair before the main theorem stands.","tokens_in":24680,"tokens_out":3800,"would_cite":false,"duration_ms":39681,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F10","60J27"],"pacs":[],"model":"deepseek-v4-flash","headline":"For density-dependent Markov chains, a path-level moderate deviation principle holds with an explicit rate function derived from the linearized drift and jump noise.","keywords":["moderate deviations","density-dependent Markov chain","path-space rate function","exponential martingale","Girsanov theorem","contact process","SIR model","Yule process"],"falsifier":"Simulate the Yule-process example at several population sizes with $a_n=n^{3/4}$, count trajectories whose normalized fluctuation path stays inside a small tube around a test path $f$, and plot $(a_n^2/n)\\log P$ against the tube radius; as the radius shrinks, the value should approach $-\\int_0^{T_0}(f'-\\lambda f)^2/(2\\lambda x_0 e^{\\lambda t})\\,dt$. A systematically different limit would show that the exponential martingale calculation misses a term.","tokens_in":23613,"feed_emoji":"📈","tokens_out":10970,"duration_ms":119640,"temperature":0.7,"pith_summary":"Density-dependent Markov chains model large-population systems—contact processes, SIR epidemics, chemical reaction networks, and Yule processes—where each jump changes the count by a fixed vector and the jump rate is n times a smooth function of the current density. The paper proves that for every scaling sequence $a_n$ with $a_n/n\\to 0$ and $a_n/\\sqrt n\\to+\\infty$, the rescaled fluctuation path $\\vartheta^n_t=(X^n_t-nX_t)/a_n$ obeys a moderate deviation principle: probabilities of path sets have exponential rate $n/a_n^2$ governed by an explicit rate function $I$. When the noise covariance matrix $\\sigma_t$ is invertible, $I$ collapses to the quadratic action $\\frac12\\int_0^{T_0}(f'-b f)^T\\sigma^{-1}(f'-b f)\\,dt$, so the exponential cost of a deviation is controlled by how far the fluctuation path departs from the linearized drift. The proof builds an exponential martingale from the generator, tilts the path measure with it, and uses a generalized Girsanov theorem to identify the tilted dynamics. Because the assumptions allow unbounded rate functions with linear growth, the result covers models outside earlier bounded-rate large-deviation treatments.","feed_headline":"Moderate deviations proven for density-dependent Markov chains","feed_subtitle":"Fluctuations between CLT and large-deviation scales obey a quadratic path rate function tied to drift and noise.","key_machinery":"The central object is the exponential martingale $\\omega^n_t(g)$, constructed from $H^n_g(t,x)=\\exp\\{(a_n/n)\\,g_t\\cdot(x-nX_t)\\}$ and the generator $\\Omega^n$ of the density-dependent chain. For large $n$ it is a martingale with expectation one and serves as the Radon--Nikodym derivative of a tilted measure $\\mathbb{P}^n_g$. Under that tilted measure, Lemmas 4.2 and 4.3 show that $X^n_t/n$ converges back to the law-of-large-numbers path $X_t$ and that the fluctuation path converges to the solution of $y'=b_t y+\\sigma_t g$, $y_0=0$. This identifies the rate function as the supremum over $g$ of the associated quadratic form, with the quadratic reduction obtained through Riesz representation and Cauchy--Schwarz. The upper bound uses the same martingale on compact sets, Sion's minimax theorem to interchange supremum and infimum, and exponential tightness (Lemma 5.2) derived from Poisson-process and Yule-process tail estimates.","core_discovery":"Theorem 2.1 states that, under Assumptions (1)--(5)—a deterministic initial density $x_0$, the intermediate scale condition on $a_n$, a finite jump set $A$, and $C^1$ rate functions $F_l$ with $F_l(0)=0$ and globally Lipschitz gradient—the path $\\vartheta^n=(X^n_t-nX_t)/a_n$ satisfies the moderate deviation upper and lower bounds: for every open $O$, $\\liminf (n/a_n^2)\\log P(\\vartheta^n\\in O)\\ge -\\inf_{f\\in O} I(f)$, and for every closed $C$, $\\limsup (n/a_n^2)\\log P(\\vartheta^n\\in C)\\le -\\inf_{f\\in C} I(f)$. The rate function is $I(f)=\\sup_g\\{f(T_0)\\cdot g(T_0)-\\int_0^{T_0} f\\cdot g'\\,ds-\\int_0^{T_0}(b_s f_s)\\cdot g_s\\,ds-\\frac12\\int_0^{T_0} g_s^T\\sigma_s g_s\\,ds\\}$, where $b_t=\\sum_{l\\in A} l(\\nabla^T F_l)(X_t)$ is the linearized drift and $\\sigma_t=\\sum_{l\\in A} l F_l(X_t)l^T$ is the jump covariance of the limiting Ornstein--Uhlenbeck process. If $\\sigma_t$ is invertible, $I(f)=\\frac12\\int_0^{T_0}(f'_s-b_s f_s)^T\\sigma_s^{-1}(f'_s-b_s f_s)\\,ds$ for absolutely continuous $f$, and $+\\infty$ otherwise; a degenerate version (Lemma 4.4) represents every finite-cost path as solving $f'=bf+\\sigma\\psi$ with cost $\\frac12\\int\\psi^T\\sigma\\psi\\,ds$. The lower bound tilts the measure so that the fluctuation path converges to the ODE $y'=b_t y+\\sigma_t g$, and the upper bound combines compact-set estimates, a minimax interchange, and exponential tightness.","pith_inferences":["A natural extension beyond the paper is to apply the same tilt-by-generator construction to other population processes whose jump rates factor as $n$ times a smooth density-dependent rate, such as open chemical reaction networks or metapopulation models; the rate function should again be the action of the linearized Ornstein--Uhlenbeck process whenever a Yule-type domination supplies exponential m","The degenerate-covariance representation suggests a fluctuation-level conservation law: when $\\sigma_t$ is singular, finite-cost fluctuation paths are confined to the linear span of the jump directions, so moderate deviations are exponentially suppressed in conserved components and the effective action involves only the fluctuating coordinates.","One could test the SIR or Yule rate functions by rare-event simulation: fix a tube around a chosen path $f$, estimate $\\log P(\\vartheta^n\\in \\text{tube})$ across increasing $n$, and check that $(a_n^2/n)\\log P$ converges to $-I(f)$; a systematic mismatch would indicate a missing jump-correction term in the exponential martingale calculation.","Letting $a_n$ approach $\\sqrt n$ from above should recover fluctuation-scale asymptotics, while letting $a_n$ approach $n$ should connect to the large-deviation regime; the paper does not prove these endpoint limits, but its scale-uniform statement invites such an interpolation."],"forward_implications":["For the contact process on the complete graph, the moderate deviation rate is $I(f)=\\int_0^{T_0}(f'-b f)^2/(2\\sigma)\\,dt$, giving explicit exponential asymptotics for deviations of the infected count at all intermediate scales.","For the SIR model, the joint deviation of susceptible and infected paths is governed by the two-dimensional quadratic form with explicit $\\sigma^{-1}$, so correlation between the two coordinates is built into the rate.","For the reversible chemical reaction $R_1+R_2\\rightleftharpoons R_3$, the rate function is finite only on fluctuation paths lying in the one-dimensional stoichiometric subspace, forcing the three coordinates to deviate in a fixed proportion.","For the Yule process with rate $\\lambda$, the result gives $I(f)=\\int_0^{T_0}(f'-\\lambda f)^2/(2\\lambda x_0 e^{\\lambda t})\\,dt$, so even with unbounded population size the moderate deviation principle holds with an explicit quadratic rate.","The theorem provides a single path-space moderate deviation principle covering all four canonical examples, so model-specific checks for the intermediate asymptotic regime are no longer needed in these systems."],"supporting_citations":[{"why":"Introduces density-dependent Markov chains and supplies the LLN, the CLT, the random time-change representation, and the coupling estimates that underpin Lemmas 3.1--3.3.","marker":"[16]"},{"why":"Develops the moderate-deviation-from-hydrodynamic-limit method with exponential martingale and tilted measure; the paper adapts this template from symmetric exclusion to finite-dimensional density-dependent chains.","marker":"[10]"},{"why":"Provides the generalized Girsanov theorem used to transform local martingales under the tilted measure, yielding the limit ODE $y'=b_t y+\\sigma_t g$ in Lemma 4.3.","marker":"[21]"},{"why":"Supplies the criterion for exponential tightness of the fluctuating paths, which extends the upper bound from compact sets to all closed sets.","marker":"[19]"},{"why":"Gives the minimax theorem used in Lemma 5.1 to interchange the supremum over $g$ and the infimum over $f$ on compact sets.","marker":"[23]"},{"why":"Provides the i.i.d. large-deviation bound used to control the tails of the dominating Yule process in Lemmas 3.1--3.2 and 4.3.","marker":"[5]"},{"why":"Serves as the standard reference for generator martingales and the time-change representation of the Markov jump processes used in the examples and preliminaries.","marker":"[8]"}],"fun_headline_variants":["Quadratic-rate moderate deviations for density-dependent chains","Moderate deviation principle for density-dependent Markov paths","Tight moderate deviation bounds for density-dependent chains","Exponential-rate deviations for density-dependent Markov dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's exponential estimates, and therefore the theorem, depend on the jump set $A$ being finite and every rate function $F_l$ being $C^1$ with $F_l(0)=0$ and globally Lipschitz gradient; if a model has infinitely many jump directions or superlinearly growing rates, the moderate deviation principle as stated is not proven.","fun_headline_variants_meta":{"raw":{"variants":["Quadratic-rate moderate deviations for density-dependent chains","Moderate deviation principle for density-dependent Markov paths","Tight moderate deviation bounds for density-dependent chains","Exponential-rate deviations for density-dependent Markov dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2711,"prompt_tokens":1063,"completion_tokens":1648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":1589}},"tokens_in":679,"tokens_out":1648,"duration_ms":13015,"temperature":1.0,"reasoning_tokens":1589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:04:44.588526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the Yule-process example at several population sizes with $a_n=n^{3/4}$, count trajectories whose normalized fluctuation path stays inside a small tube around a test path $f$, and plot $(a_n^2/n)\\log P$ against the tube radius; as the radius shrinks, the value should approach $-\\int_0^{T_0}(f'-\\lambda f)^2/(2\\lambda x_0 e^{\\lambda t})\\,dt$. A systematically different limit would show that the exponential martingale calculation misses a term.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces density-dependent Markov chains and supplies the LLN, the CLT, the random time-change representation, and the coupling estimates that underpin Lemmas 3.1--3.3."},{"cited_title":"and Quastel, J","cited_arxiv_id":null,"evidence_quote":"Develops the moderate-deviation-from-hydrodynamic-limit method with exponential martingale and tilted measure; the paper adapts this template from symmetric exclusion to finite-dimensional density-dependent chains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generalized Girsanov theorem used to transform local martingales under the tilted measure, yielding the limit ODE $y'=b_t y+\\sigma_t g$ in Lemma 4.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the criterion for exponential tightness of the fluctuating paths, which extends the upper bound from compact sets to all closed sets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the minimax theorem used in Lemma 5.1 to interchange the supremum over $g$ and the infimum over $f$ on compact sets."},{"cited_title":"and Zeitouni, O","cited_arxiv_id":null,"evidence_quote":"Provides the i.i.d. large-deviation bound used to control the tails of the dominating Yule process in Lemmas 3.1--3.2 and 4.3."}],"review_version":1}