Large deviations for stochastic nonlinear systems of slow-fast diffusions with non-Gaussian L\'evy noises

Jinqiao Duan; Ren\'e Schilling; Shenglan Yuan

REVIEW 4 major objections 5 minor 34 references

Large deviations for stochastic nonlinear systems of slow-fast diffusions with non-Gaussian L\'evy noises

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Slow variables in Lévy-driven slow-fast diffusions obey a large deviation principle with explicit rate function.

desk verdict The three-regime picture is plausible and the critical-case proof is real, but the new regimes ride on an asserted operator inequality, so the main theorem is not proven as written. read the letter →

arxiv 1908.03481 v2 pith:4N246LYW submitted 2019-08-09 math.DS

Shenglan Yuan , René Schilling , Jinqiao Duan This is my paper · ORCID

classification math.DS MSC 60F1037H10

keywords largedeviationsslow-fastsystemsLévynoiseviscositysolutionscomparisonprinciplenon-Gaussianmultiscalestochasticdynamicsrareevents

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

Large deviations describe how likely the slow variable is to wander far from its typical path, on the exponential scale. This paper proves that in a coupled slow-fast diffusion with both Brownian and Lévy noise, these probabilities decay at rate $1/\varepsilon$ with a cost $I(x,x_0,t)$ determined by a limiting Hamilton-Jacobi equation. The limiting Hamiltonian takes one of three explicit forms depending on the time-scale exponent $\alpha$: an average over fast states, a principal eigenvalue, or a maximum over fast states. If the claim holds, modelers can compute rare-event asymptotics in multiscale systems with non-Gaussian noise without solving the full two-scale dynamics.

What carries the argument

The load-bearing identity is the logarithmic transform $U^\varepsilon(t,x,y)=\varepsilon\ln\mathbb{E}[e^{h(X^\varepsilon_t)/\varepsilon}\mid X^\varepsilon_0=x,Y^\varepsilon_0=y]$, which converts rare-event expectations into viscosity solutions of a nonlinear partial integro-differential equation of Hamilton-Jacobi type. The limit is controlled by the Hamiltonian $H^0$; its correct form is forced by an operator inequality $\inf_\lambda \hat{H}^\lambda \le H^0 \le \sup_\lambda \check{H}^\lambda$, where $\hat{H}^\lambda$ and $\check{H}^\lambda$ are regularized operators built from test functions, a Lyapunov function, and an indexing set. The comparison principle for the limit equation then identifies the upper and lower semilimits, yielding uniform convergence of $U^\varepsilon$ to $U^0$. This uniform convergence, plus exponential tightness of the slow process, is what produces the large deviation principle.

What would settle it

For a concrete two-scale Ornstein-Uhlenbeck-type model with Lévy jumps, compute $H^0$ in each regime and compare Monte Carlo estimates of $-\varepsilon\log P(X^\varepsilon_t\in A)$ with the predicted $\inf_{x\in A}I(x,x_0,t)$ over decreasing $\varepsilon$; a mismatch would falsify the large deviation claim. A cheaper check is whether the three $H^0$ forms are uniformly continuous on compact sets, the condition the comparison-principle proof needs.

Watch

Extended reading notes

Core claim

The central claim is that the slow variable $X^\varepsilon_t$ satisfies a large deviation principle (Theorem 2). Under conditions (C1)-(C5), the probability that $X^\varepsilon_t$ lies near $x$ decays as $\exp(-I(x,x_0,t)/\varepsilon)$, where $I(x,x_0,t)=\sup_{h\in C_b(\mathbb{R})}\{h(x)-U^0(t,x_0)\}$ and $U^0$ is the unique viscosity solution of the limiting Cauchy problem $\partial_t U=H^0(x,\partial_x U)$ with $U(0,x)=h(x)$. The paper identifies $H^0$ separately for $\alpha>2$, $\alpha=2$, and $\alpha<2$: respectively, the integral of the local Hamiltonian density $V_{x,p}(y)$ against the invariant measure of the fast motion; the principal eigenvalue of the generator $L_{x,p}+V_{x,p}$ expressed through a variational formula over probability measures; and $\max_y V_{x,p}(y)$. The derivation passes through the convergence of the logarithmic functionals $U^\varepsilon$ to $U^0$ on compact sets, obtained by viscosity sub- and supersolution methods and the comparison principle.

Load-bearing premise

The proof assumes the comparison principle for the limit Cauchy problem (3.16), and the paper proves that principle only under a uniform-continuity condition on $H^0$ that is not verified for the three explicit Hamiltonians.

Editorial extensions

If this is right

If the theorem is correct, rare-event probabilities for the slow variable are asymptotically computable from the limiting equation, bypassing the two-scale dynamics.
In the supercritical regime ($\alpha>2$) the rate is an average over the fast invariant measure; in the subcritical regime ($\alpha<2$) it is a maximum over fast states, so the two regimes can have quite different rare-event costs.
When the coefficients do not depend on $x$, the rate function simplifies to $t Q^0((x_0-x)/t)$, an explicit Legendre-transform form.
The three-regime distinction gives a practical rule: the time-scale ratio $\delta=\varepsilon^\alpha$ changes not just the averaging but the structure of the effective Hamiltonian.

Reading between the lines

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

The paper leaves uniform continuity of the three $H^0$ forms as an open verification; a direct check for each regime would turn the conditional convergence theorem into an unconditional one.
The subcritical form $H^0=\max_y V_{x,p}(y)$ suggests that, in that regime, the rare-event cost is controlled by the single fast state most favourable to the fluctuation, so model reduction could be possible by maximizing over $y$ alone.
The same semilimit-plus-comparison route should extend to multidimensional slow variables, with the Hamiltonian forms re-derived coordinate-wise, but the Lyapunov conditions would need to be adapted.

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.

Desk Editor's Note

read the letter

Two things you should know up front. First, the paper's genuinely new results—the supercritical and subcritical large deviations—are not actually proved. The operator inequality (4.31) that identifies the effective Hamiltonian is proved only in the critical case alpha=2, which is exactly the case already handled by Kumar-Popovic. For the new regimes the paper says it "can be proven with similar ideas." That is a placeholder, not a proof. Second, the formal picture—average over the invariant measure for alpha>2, principal eigenvalue for alpha=2, max over y of V_{x,p} for alpha<2—is natural and probably right. So the paper is worth reading for the heuristic, but Theorem 2 should not be accepted as established.

What the paper does well: the three-regime formal asymptotic derivation is clean, and the distinction between the regimes is interesting. The critical-case operator inequality is a real argument using Donsker-Varadhan variational formulas, with the occupation-measure lower bound spelled out. The exponential tightness lemma is also in place. That is substantive work.

The soft spots are exactly the load-bearing ones. (i) Theorem 1 proves comparison only under uniform continuity of H0 on compact sets; Remark 2 defers verification, but no verification appears for any of the three Hamiltonians. Lemma 4 then assumes comparison, and its proof immediately invokes Theorem 1, so the gap is internal. (ii) The subcritical derivation of H0 = max_y V hides a cell problem: one has to show that a periodic corrector exists exactly at that level and that no lower level gives a subsolution. The one-line completion in (3.15) does not supply that. (iii) Lemma 1, the viscosity property of U^epsilon, is stated without proof; this is probably standard and fixable, so I treat it as minor. The abstract's claim that the comparison principle is "verified" is too strong—it is verified conditionally.

None of this means the paper is nonsense. The framework is the well-trodden Bryc/Varadhan route, and the critical case shows the authors can do the hard estimate when they need to. But as written, the central theorem relies on exactly the cases they did not prove.

Who it is for: people working on averaging and large deviations for jump processes will want to know the three-regime classification. It deserves a serious referee, and I would engage with it. My recommendation to the editor: send it to review, but tell the referee the priorities are the missing verification of (4.31) for alpha>2 and alpha<2, and the uniform continuity needed by Theorem 1. A major revision could make this a solid paper.

Referee Report

4 major / 5 minor

Summary. The paper aims to establish a large deviation principle for the slow component X^ε_t of a fully coupled slow-fast stochastic system driven by both Brownian and Lévy noises, in the scaling δ = ε^α with α > 1. The proof follows the Bryc–Varadhan/viscosity-solution route: the log-moment functionals U^ε are characterized as viscosity solutions of a nonlinear partial integro-differential equation (2.12), their relaxed semilimits are shown to converge to the unique solution of a limit Hamilton–Jacobi equation (3.16), and the rate function is given by the Legendre-type transform (5.3). Three regimes are treated: supercritical α > 2, critical α = 2, and subcritical α < 2, with explicit limit Hamiltonians (3.4), (3.11), and (3.15).

Significance. If fully established, the result would be a natural extension of Kumar–Popovic [26] to all α > 1 in the fully coupled case, and the explicit Hamiltonians would make the LDP potentially applicable to multiscale models with non-Gaussian noise. The paper contains a substantial rigorous argument in the critical case, including the proof of the operator inequality (4.31) via Donsker–Varadhan variational formulas and occupation-measure estimates, and the exponential-tightness argument in Lemma 5. However, the most novel parts—the supercritical and subcritical regimes—rest on unproved assertions, and the comparison principle is conditional on an unverified regularity hypothesis.

major comments (4)

[§4.4, inequality (4.31)] The operator inequality (4.31) is proved only for the critical case α = 2. For the supercritical (α > 2) and subcritical (α < 2) cases, the text states that (4.31) 'can be proven with similar ideas', but no proof is given. Since (4.31) is the step that identifies the effective Hamiltonian H0 in the comparison argument of Lemma 4, Theorem 2 is not established for the two regimes that constitute the paper's main new contribution. A proof cannot be replaced by an assertion, especially because the Hamiltonians in (3.4) and (3.15) are not principal eigenvalues and do not obviously follow from the critical-case reasoning.
[Theorem 1 and Remark 2] Theorem 1 proves the comparison principle for (3.16) only under the explicit assumption that H0 is uniformly continuous on compact sets. No verification of this assumption is supplied for any of the three Hamiltonians (3.4), (3.11), (3.15); Remark 2 merely defers the verification, and Remark 3 covers only a special x-independent case with ρ = 0. Because Lemma 4 and hence Theorem 2 rely on this comparison principle, the main theorem is conditional on an unverified regularity property.
[§3, subcritical case, Eq. (3.15)] The derivation of H0(x,p) = max_y V_{x,p}(y) in the subcritical regime is not a proof. The line 'If H0 ≥ V, we have ... then we obtain' asserts the equality of the square-root expression without justification; one must prove existence of a periodic solution W to the cell problem at exactly that level and exclude smaller levels. As written, (3.15) is an unproved ansatz that feeds directly into the rate function of Theorem 2.
[Lemma 4] Lemma 4 states 'Suppose the comparison principle holds for the Cauchy problem (3.16)', but its proof proceeds by invoking Theorem 1 to obtain the comparison principle. This makes the lemma logically inconsistent: either the hypothesis is redundant and should be removed with the uniform-continuity condition verified, or the proof does not use the stated hypothesis. In either case, the convergence result is not established under the assumptions (C1)-(C5) alone, contrary to the statement of Theorem 2.

minor comments (5)

[§1, Introduction] The word 'sloutions' should be 'solutions'.
[Lemma 5] The word 'supermatingale' should be 'supermartingale'.
[Remark 2] The phrase 'the expression as on the right-hand side of (4.30)' is awkward; it should be 'the expression on the right-hand side of (4.30)'.
[Lemma 1] The proof of Lemma 1 is omitted with a reference to 'similar arguments' in [17,19,26]; for a self-contained journal paper, a precise citation to the exact theorem or a short proof should be included.
[Example 1] The claim that the proof for system (5.6) is 'similar to that of system (1.1)' needs more detail, since the drift modification Kν1, Kν2 changes the Hamiltonian in the formal expansions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard Bryc/viscosity-solution argument with independent derivation of H0; unproved operator inequality is a completeness gap, not a circular step.

full rationale

The derivation chain is the standard Bryc/inverse-Varadhan viscosity argument and is not circular. U^epsilon is defined as epsilon log E[exp(h(X^epsilon_t)/epsilon)], so the limit U0 is a log-moment functional, and the rate function I(x,x0,t)=sup_h {h(x)-U0(t,x0)} is the Bryc transform, not an input fitted to the desired LDP. The three Hamiltonians H0 are obtained by formal asymptotic expansions in Section 3 and then are meant to be identified with the inf/sup of the approximate Hamiltonians via the operator inequality (4.31); they are not constructed from the rate function. The methodological citations [17,19,26] are prior works by other authors, and the only self-citations ([12] Duan, [32] Schilling) are standard textbook facts that are not load-bearing. The genuine weaknesses are verification gaps: Section 4.4 proves (4.31) for the critical case and then states that the supercritical and subcritical cases 'can be proven with similar ideas', and Theorem 1 assumes uniform continuity of H0 on compact sets with verification deferred to Remark 2. These are correctness/completeness issues, not circularity: no equation reduces by construction to its own input, and no fitted parameter is renamed as a prediction. Score 0.

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

The paper relies on conditions C1-C5 on the coefficients, on the viscosity solution property of U^{epsilon,delta} (Lemma 1, proof omitted), on the comparison principle for the limit equation (3.16), on the operator inequality (4.31) for all regimes, and on the uniform continuity of H0. None of these except C1-C5 are stated as hypotheses in Theorem 2, yet the proof requires them.

assumptions (5)

domain assumption C1-C5: Lipschitz, growth, periodic, ergodicity, and Lyapunov conditions on the coefficients.
These hypotheses define the model class and are used throughout (Section 2).
ad hoc to paper U^{epsilon,delta} is a viscosity solution of (2.12) for each epsilon, delta.
Lemma 1 asserts this with 'proof omitted', citing [17,19,26].
ad hoc to paper Comparison principle holds for limit Cauchy problem (3.16).
Lemma 4 assumes it; Theorem 1 only proves it under unverified uniform continuity of H0.
ad hoc to paper Operator inequality (4.31) holds for supercritical and subcritical cases.
Proved only for critical case; for alpha>2 and alpha<2 stated without proof.
ad hoc to paper H0 is uniformly continuous on compact sets.
Required by Theorem 1; Remark 2 admits no general verification is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large deviations for stochastic nonlinear systems of slow-fast diffusions with non-Gaussian L\'evy noises." pith.science (2026). https://pith.science/paper/4N246LYW

@misc{pith2026190803481,
  author       = {Pith},
  title        = {Pith review of: Large deviations for stochastic nonlinear systems of slow-fast diffusions with non-Gaussian L\'evy noises},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4N246LYW}},
  note         = {Machine review of arXiv:1908.03481}
}

read the original abstract

We establish the large deviation principle for the slow variables in slow-fast dynamical system driven by both Brownian noises and L\'evy noises. The fast variables evolve at much faster time scale than the slow variables, but they are fully inter-dependent. We study the asymptotics of the logarithmic functionals of the slow variables in the three regimes based on viscosity solutions to the Cauchy problem for a sequence of partial integro-differential equations. We also verify the comparison principle for the related Cauchy problem to show the existence and uniqueness of the limit for viscosity solutions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[26]

Kumar, L

R. Kumar, L. Popovic, Large deviations for multi-scale jump-diﬀ usion processes. Stochastic Processes and their Applications , 127(2017)1297- 1320

work page 2017
[1]

Applebaum, L´ evy Processes and Stochastic Calculus

D. Applebaum, L´ evy Processes and Stochastic Calculus. Cambridge Uni- versity Press, Cambridge, 2009

work page 2009
[2]

Awatif, Equqtions D’Hamilton-Jacobi Du Premier Ordre Avec Te r- mes Intgro-Diﬀrentiels

S. Awatif, Equqtions D’Hamilton-Jacobi Du Premier Ordre Avec Te r- mes Intgro-Diﬀrentiels. Communications in partial diﬀerential equa- tions, 16(1991)1057-1074

work page 1991
[3]

Alvarez, A

O. Alvarez, A. Tourin, Viscosity solutions of nonlinear integro- diﬀerential equations. Nonlinear Analysis. Elsevier Masson , 13(2016)293-317

work page 2016
[4]

V. I. Bakhtin, Cram´ er’s asymptotics in systems with fast and slo w mo- tions. Stochastics and Stochastic Reports , 75(2003)319-341

work page 2003
[5]

Bardi, I

M. Bardi, I. Capuzzo Dolcetta, Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations. Springer Science & Business Media, 2008

work page 2008
[6]

Barles, C

G. Barles, C. Imbert, Second-order elliptic integro-diﬀerential equations: viscosity solutions’ theory revisited.Annales de l’Institut Henri Poincare (C) Non Linear Analysis . Elsevier Masson, 25(2008)567-585

work page 2008
[7]

Berestycki, J

H. Berestycki, J. Busca, I. Florent, Asymptotics and calibratio n of local volatility models. Quant. Finance , 2(2002): 61-69

work page 2002

Show all 34 references

[8]

Bardi, A

M. Bardi, A. Cesaroni, D. Ghilli, Large deviations for some fast sto chas- tic volatility models by viscosity methods. Discrete & Continuous Dy- namical Systems-A , 35(2015)3965-3988

2015
[9]

Bardi, A

M. Bardi, A. Cesaroni, A. Scotti, Convergence in multiscale ﬁnanc ial models with non-Gaussian stochastic volatility. ESAIM: Control, Opti- misation and Calculus of Variations , 22(2016)500-518. 32

2016
[10]

Bouchet, T

F. Bouchet, T. Grafke, T. Tangarife, and E. Vanden-Eijnden , Large devi- ations in fast-slow systems. Journal of Statistical Physics , 162(2016)793- 812

2016
[11]

M. F. Chen, Eigenvalues, inequalities, and ergodic theory . Springer Sci- ence and Business Media, 2006

2006
[12]

Duan, An introduction to stochastic dynamics

J. Duan, An introduction to stochastic dynamics . Cambridge University Press, 2015

2015
[13]

J. D. Deuschel, D. W. Stroock, Large deviations . Boston: Academic press, 1989

1989
[14]

Dembo, O

A. Dembo, O. Zeitouni, Large deviations techniques and applications , volume 38 of Stochastic Modelling and Applied Probability, 2010

2010
[15]

M. D. Donsker, S. R. S. Varadhan, On a variational formula for the principal eigenvalue for operators with maximum principle. Proceedings of the National Academy of Sciences , 72(1975)780-783

1975
[16]

M. D. Donsker, S. R. S. Varadhan, Asymptotic evaluation of ce rtain Markov process expectations for large time, IV. Communications on Pure and Applied Mathematics , 36(1983)183-212

1983
[17]

J. Feng, T. G. Kurtz, Large deviations for stochastic processes. American Mathematical Soc., 2006

2006
[18]

W. H. Fleming, H. M. Soner, Controlled Markov processes and viscosity solutions. Springer Science and Business Media, 2006

2006
[19]

J. Feng, J. P. Fouque, R. Kumar, Small-time asymptotics for fa st mean- reverting stochastic volatility models. The Annals of Applied Probability , 22(2012)1541-1575

2012
[20]

F. D. Hollander, Large deviations. American Mathematical Soc., 2008

2008
[21]

Hollender, L´ evy-Type Processes under Uncertainty and R elated Non- local Equations, 2016

J. Hollender, L´ evy-Type Processes under Uncertainty and R elated Non- local Equations, 2016

2016
[22]

E. R. Jakobsen, K. H. Karlsen, A maximum principle for semicontin uous functions applicable to integro-partial diﬀerential equations. Nonlinear Diﬀerential Equations and Applications , 13( 2006)137-165. 33

2006
[23]

Kifer, Large deviations and adiabatic transitions for dynamical s ys- tems and Markov processes in fully coupled averaging

Y. Kifer, Large deviations and adiabatic transitions for dynamical s ys- tems and Markov processes in fully coupled averaging . American Math- ematical Soc., 2009

2009
[24]

Kifer, Averaging principle for fully coupled dynamical systems and large deviations

Y. Kifer, Averaging principle for fully coupled dynamical systems and large deviations. Ergodic Theory and Dynamical Systems , 24(2004)847- 871

2004
[25]

Karlin, H

S. Karlin, H. E. Taylor, A second course in stochastic processes. Elsevier, 1981

1981
[27]

Liptser, Large deviations for two scaled diﬀusions

R. Liptser, Large deviations for two scaled diﬀusions. Probability theory and related ﬁelds , 106(1996)71-104

1996
[28]

Majda, X

A. Majda, X. Wang, Nonlinear dynamics and statistical theories for basic geophysical ﬂows . Cambridge University Press, 2006

2006
[29]

Pham, Optimal stopping of controlled jump diﬀusion processe s: a vis- cosity solution approach

H. Pham, Optimal stopping of controlled jump diﬀusion processe s: a vis- cosity solution approach. Journal of Mathematical Systems, Estimation and Control , 1998

1998
[30]

T. N. Palmer, A nonlinear dynamical perspective on model error : A pro- posal for non-local stochastic-dynamic parametrization in weath er and climate prediction models. Quarterly Journal of the Royal Meteorological Society, 127(2001)279-304

2001
[31]

Qian, Mesoscopic nonequilibrium thermodynamics of single macr o- molecules and dynamic entropy-energy compensation

H. Qian, Mesoscopic nonequilibrium thermodynamics of single macr o- molecules and dynamic entropy-energy compensation. Physical Review E, 65(2001)016102

2001
[32]

R. L. Schilling, An Introduction to L´ evy and Feller Processes. Dresden, 2015

2015
[33]

H. M. Soner, Optimal control with state-space constraint. SIAM journal on control and optimization , 24(1986)1110-1122. 34

1986
[34]

Spiliopoulos, Large deviations and importance sampling for sys tems of slow-fast motion

K. Spiliopoulos, Large deviations and importance sampling for sys tems of slow-fast motion. Applied Mathematics & Optimization, 67(2013)123- 161. 35

2013

Pith tools

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

[1] [26]

Kumar, L

R. Kumar, L. Popovic, Large deviations for multi-scale jump-diﬀ usion processes. Stochastic Processes and their Applications , 127(2017)1297- 1320

work page 2017

[2] [1]

Applebaum, L´ evy Processes and Stochastic Calculus

D. Applebaum, L´ evy Processes and Stochastic Calculus. Cambridge Uni- versity Press, Cambridge, 2009

work page 2009

[3] [2]

Awatif, Equqtions D’Hamilton-Jacobi Du Premier Ordre Avec Te r- mes Intgro-Diﬀrentiels

S. Awatif, Equqtions D’Hamilton-Jacobi Du Premier Ordre Avec Te r- mes Intgro-Diﬀrentiels. Communications in partial diﬀerential equa- tions, 16(1991)1057-1074

work page 1991

[4] [3]

Alvarez, A

O. Alvarez, A. Tourin, Viscosity solutions of nonlinear integro- diﬀerential equations. Nonlinear Analysis. Elsevier Masson , 13(2016)293-317

work page 2016

[5] [4]

V. I. Bakhtin, Cram´ er’s asymptotics in systems with fast and slo w mo- tions. Stochastics and Stochastic Reports , 75(2003)319-341

work page 2003

[6] [5]

Bardi, I

M. Bardi, I. Capuzzo Dolcetta, Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations. Springer Science & Business Media, 2008

work page 2008

[7] [6]

Barles, C

G. Barles, C. Imbert, Second-order elliptic integro-diﬀerential equations: viscosity solutions’ theory revisited.Annales de l’Institut Henri Poincare (C) Non Linear Analysis . Elsevier Masson, 25(2008)567-585

work page 2008

[8] [7]

Berestycki, J

H. Berestycki, J. Busca, I. Florent, Asymptotics and calibratio n of local volatility models. Quant. Finance , 2(2002): 61-69

work page 2002