Pith. sign in

REVIEW 2 major objections 4 minor 27 references

Moderate deviations of density-dependent Markov chains

T0 review · 2 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 1908.03762 v3 pith:2N5P5BXY submitted 2019-08-10 math.PR

classification math.PR MSC 60F1060J27
keywords moderatedeviationsdensity-dependentMarkovchainpath-spaceratefunctionexponentialmartingaleGirsanovtheoremcontactprocessSIRmodelYule
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

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.

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 (2)
  1. [Section 5, Lemma 5.1] 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.
  2. [Section 5, Lemma 5.1] 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.
minor comments (4)
  1. [Section 3, Proof of Lemma 3.1] 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.
  2. [Section 2, proof of Equation (2.2)] 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.
  3. [Section 5, Lemma 5.2] 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}.
  4. [General] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the rate function and bounds are derived from the generator and model primitives, not from the asserted moderate deviation principle.

full rationale

The rate function I(f) in Equation (2.1) is defined directly from the LLN trajectory X_t and the model matrices b_t = sum_l l (nabla^T F_l)(X_t) and sigma_t = sum_l l F_l(X_t) l^T, which are themselves built from the transition rates F_l and jump vectors l. The lower-bound proof in Section 4 constructs an exponential martingale from the generator of the chain and uses a generalized Girsanov theorem to change measure, then estimates the probability of the event theta^n in O in terms of I(f). The upper-bound proof in Section 5 uses the same martingale, exponential tightness, and a compact-set estimate. No fitted constants, no data-dependent parameters, and no quantity called a prediction is obtained from prior fitting. The external citations are standard tools: Kurtz (1978) for the LLN/CLT and Poisson representation, Schuppen-Wong for Girsanov transformation, Sion for minimax, and Puhalskii for exponential tightness; none of these imports the moderate deviation result itself. The proof's heuristic explanation of the rate function in Section 2 is an illustration, not a circular derivation, because the rigorous rate function is independently defined and then proved to equal the explicit quadratic form when sigma is invertible. Any concern about whether Sion's minimax theorem applies to compact subsets of Skorokhod space in Lemma 5.1, or about the finiteness of A in Assumption (4), is a correctness or scope issue, not circularity: the paper does not assume the conclusion it claims to prove. The manuscript even notes where its assumptions are needed, such as the comment that Lemmas 3.2 and 3.3 require proof because earlier large deviation results assume bounded rates. Thus no self-definitional, fitted-input, self-citation-load-bearing, imported-uniqueness, ansatz-smuggling, or renaming pattern is present.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No free parameters and no invented entities. The rate function is constructed from the model's rates F_l and the deterministic LLN path X_t; the assumptions are standard Lipschitz and finiteness conditions. The proof relies on imported results, including Kurtz's LLN/CLT, Schuppen-Wong Girsanov, Puhalskii exponential tightness, and Sion minimax, which are listed as axioms.

assumptions (5)
  • domain assumption The jump set A is finite and each rate F_l is C^1 with bounded gradient on G and F_l(0)=0.
    Assumptions (4)-(5). They guarantee linear growth, finite sums, exponential moments, and the Yule coupling used in Lemmas 3.1-3.3 and 4.1.
  • domain assumption The initial state is deterministic: X_0^n = n x_0 with x0 ≠ 0 in G.
    Assumptions (1)-(2). Used in the lower bound to cancel the X_0 terms; random initial states are not treated.
  • standard math The limiting ODE dX_t/dt = Σ l F_l(X_t), X_0=x0 has a unique solution on [0,T0].
    Follows from the Lipschitz condition on Σ lF_l; defines X_t, b_t, σ_t and the rate function throughout.
  • standard math Skorokhod space D([0,T0], R^d) with the Skorokhod metric is a complete separable metric space.
    Imported from [24]; provides the topology in which the moderate deviation principle is stated.
  • standard math The generalized Girsanov theorem of Schuppen-Wong and the exponential tightness criterion of Puhalskii are valid in the stated generality.
    Used in Lemmas 4.3 and 5.2 to change measure and derive tightness from pathwise estimates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moderate deviations of density-dependent Markov chains." pith.science (2026). https://pith.science/paper/2N5P5BXY

@misc{pith2026190803762,
  author       = {Pith},
  title        = {Pith review of: Moderate deviations of density-dependent Markov chains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2N5P5BXY}},
  note         = {Machine review of arXiv:1908.03762}
}
read the original abstract

The density-dependent Markov chain (DDMC) introduced in \cite{Kurtz1978} is a continuous time Markov process applied in fields such as epidemics, chemical reactions and so on. In this paper, we give moderate deviation principles of paths of DDMC under some generally satisfied assumptions. The proofs for the lower and upper bounds of our main result utilize an exponential martingale and a generalized version of Girsanov's theorem. The exponential martingale is defined according to the generator of DDMC.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages

  1. [23]

    Sion, M. (1958). On general minimax theorems. Pacific Journal of Mathematics 8, 171-176

  2. [1]

    and Eckmann, J-P

    Agazzi, A., Dembo, A. and Eckmann, J-P. (2018). Large deviatio ns theory for Markov jump models of chemical reaction networks. The Annals of Applied Probability 28, 1821- 1855

  3. [2]

    Borovkov, A. A. and Mogulskii, A. A. (1978). Probabilities of large deviations in topo- logical spaces I. Siberian Mathematical Journal 19, 697-709

  4. [3]

    Borovkov, A. A. and Mogulskii, A. A. (1980). Probabilities of large deviations in topo- logical spaces II. Siberian Mathematical Journal 21, 653-664

  5. [4]

    Chan, T. (1998). Large deviations and quasi-stationarity for d ensity-dependent birth- death processes. Australian Mathematical Society. Journal. Series B. Appli ed Mathematics 40, 238-256

  6. [5]

    and Zeitouni, O

    Dembo, A. and Zeitouni, O. (1997). Large Deviations: Techniques and Applications. Springer, Berlin

  7. [6]

    Deuschel, J. D. and Stroock, D. W. (1989). Large deviations (Pure and Applied Mathe- matics). 137, Academic Press

  8. [7]

    and Chen, X

    de Acosta, A. and Chen, X. (1998). Moderate deviations for em pirical measures of Markov chains. Journal of Theoretical Probability 11, 1075-1110

Show all 27 references
  1. [8]

    and Kurtz, T

    Ethier, N. and Kurtz, T. (1986). Markov Processes: Characterization and Convergence. John Wiley and Sons, Hoboken, NJ, USA

  2. [9]

    Gao, FQ. (1996). Moderate deviations for martingales and mixing random processes. Stochastic Processes and their Applications , 61, 263-275

  3. [10]

    and Quastel, J

    Gao, FQ. and Quastel, J. (2003). Moderate deviations from th e hydrodynamic limit of the symmetric exclusion process. Science in China (Series A) 5, 577-592

  4. [11]

    Gao, FQ., Jiang, H., and Wang, BB. (2010). Moderate deviations for parameter esti- mators in fractional Ornstein-Uhlenbeck process. Acta Mathematica Scientia. Series B. English Edition 30, 1125-1133

  5. [12]

    Gao, FQ. (2017). Long time asymptotics of unbounded additive functionals of Markov processes. Electronic Journal of Probability 22, No. 94, 1-21

  6. [13]

    and Qian, H

    Ge, H. and Qian, H. (2017). Mathematical formalism of nonequilib rium thermodynam- ics for nonlinear chemical reaction systems with general rate law. Journal of Statistical Physics 166, 190-209

  7. [14]

    and Varadhan, S

    Kipnis, C., Olla, S. and Varadhan, S. R. S. (1989). Hydrodynamic s and large deviations for simple exclusion processes. Communications on Pure and Applied Mathematics 42, 115-137

  8. [15]

    and Landim, C

    Kipnis, C. and Landim, C. (1999). Scaling Limits of Interacting Particle Systems. Springer-Verlag, Berlin

  9. [16]

    Kurtz, T. (1978). Strong approximation theorems for densit y dependent Markov chains. Stochastic Processes and their Applications 6, 223-240

  10. [17]

    Liggett, T. M. (1985). Interacting Particle Systems. Springer, New York. 26

  11. [18]

    and Samegni-Kepgnou, B

    Pardoux, ´E. and Samegni-Kepgnou, B. (2017). Large deviation principle for e pidemic models. Journal of Applied Probability 54, 905-920

  12. [19]

    Puhalskii, A. (1994). The method of stochastic exponentials fo r large deviations. Stochastic Processes and their Applications 54, 45-70

  13. [20]

    and Williams, D

    Rogers, C. and Williams, D. (1986). Diffusions, Markov Processes and Martingales. Cambridge

  14. [21]

    Schuppen, V. J. and Wong, E. (1974). Transformation of loca l martingales under a change of law. The Annals of Probability 2, 879-888

  15. [22]

    and Weiss, A

    Schwartz, A. and Weiss, A. (1995). Large Deviations for Performance Analysis. Chap- man and Hall, London

  16. [24]

    Skorokhod, A. V. (1956). Limit theorems for stochastic proc esses. Theory of Probability and Its Applications 1-3, 261-290

  17. [25]

    and Xu, LH

    Wang, FY., Xiong, J. and Xu, LH. (2016). Asymptotics of sample entropy production rate for stochastic differential equations. Journal of Statistical Physics 163, 1211-1234

  18. [26]

    and Zhang, TS

    Wang, R. and Zhang, TS. (2015). Moderate deviations for sto chastic reaction-diffusion equations with multiplicative noise. Potential Analysis 42, 99-113

  19. [27]

    Wu, L. (1995). Moderate deviations of dependent random var iables related to CLT. The Annals of Probability 23, 420-445. 27

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.