REVIEW 3 major objections 5 minor 53 references
Mean-Field Limits for Nearly Unstable Hawkes Processes
T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Near-critical Hawkes scaling limits are affine stochastic Volterra diffusions; mean-field particle systems have three regimes set by $n(1-\|\phi^n\|_{L^1})^2$.
desk verdict A solid conditional theorem with a genuinely new mean-field trichotomy, but the abstract's L1-based phase diagram is not supported by the theorem's zeta=lim n beta_n^2 and needs major revision. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying object is the second-kind resolvent $\psi^n=\sum_{k\ge1}(\phi^n)^{*k}$, the sum of all convolution powers of the self-exciting kernel, which encodes the total cumulative effect of repeated excitation. The martingale representation $\lambda^n(t)=\mu^n+\|\psi^n\|_{L^1_t}\mu^n+\psi^n*dM^n(t)$ rewrites the intensity as a deterministic drift plus a stochastic convolution against the compensated point process. The vague convergence of $\beta_n\psi^n(t)\,dt$ fixes the kernel $f$ in the limiting Volterra equation, while the uniform bound on $\beta_n\|\psi^n\|_{L^2_T}$ supplies the tightness estimates that force the limit to be continuous and the martingale to converge to a time-changed Brownian motion. In the mean-field proof, an auxiliary particle system built from the same Poisson measures but with the first $K$ coordinates removed is shown to be asymptotically equivalent to the original system, which makes the first $K$ particles conditionally independent Cox processes time-changed by the aggregate compensator.
What would settle it
Take a family of kernels with $\|\phi^n\|_{L^1}=1-1/n$ whose mass is concentrated on intervals of width $\varepsilon_n\to0$, and check whether any $\beta_n\to0$ keeps $\sup_n\beta_n\|\psi^n\|_{L^2_T}<\infty$ while $\beta_n\psi^n(t)\,dt$ has a non-degenerate vague limit; if only degenerate or unbounded rescalings are possible and the rescaled Hawkes process still converges to a non-affine limit, the blanket claim that $L^1$-criticality forces affine Volterra limits would be refuted.
Extended reading notes
Core claim
The central claim is that near-criticality is a general source of affine stochastic Volterra limits. Under Condition 2.1, meaning there is $\beta_n\to0$ with $\sup_n \beta_n\|\psi^n_{ij}\|_{L^2_T}<\infty$ and $\beta_n\psi^n_{ij}(t)\,dt$ converging vaguely to $F_{ij}(dt)$, together with $\beta_n\mu^n\to a$, the rescaled processes $(\beta_n^2\Lambda^n,\beta_n^2N^n,\beta_nM^n)$ are C-tight, so all limit points are continuous, and every limit $(X,X,Z)$ satisfies $X(t)=\|F\|_{L^1_t}\cdot a+f*Z(t)$ with $Z_i=B_i\circ X_i$ for a Brownian motion $B$. In the mean-field particle system, the aggregate process is itself a univariate Hawkes process, and the limiting law of any fixed block of $K$ particles is governed by $\zeta=\lim n\beta_n^2$: if $\zeta=0$, $X_i=X$ and $Z_i=W_i\circ X$; if $0<\zeta<\infty$, $X_i=\zeta N_i^\circ(X/\zeta)$ and $Z_i=\sqrt{\zeta}\,\widetilde N_i^\circ(X/\zeta)$ for independent unit-rate Poisson processes $N_i^\circ$; and if $\zeta=\infty$, $X_i=Z_i=0$.
Load-bearing premise
The load-bearing premise is that the cumulative self-excitation kernel can be rescaled to converge to a nontrivial limit measure while its rescaled squared size stays uniformly bounded; merely requiring the original kernel mass to approach one does not guarantee this.
Editorial extensions
If this is right
- Under Condition 2.1, every weak limit of the rescaled Hawkes compensator and counting process is an affine stochastic Volterra diffusion, so macroscopic equations of square-root or fractional type arise from microscopic self-excitation.
- At $\zeta=0$ the mean-field network synchronizes: each particle has the same limit $X$ as the aggregate, and its noise is a Brownian motion time-changed by $X$.
- At $0<\zeta<\infty$ the particles are conditionally independent Poisson processes time-changed by $X/\zeta$, a propagation of chaos toward a stochastic intensity rather than a deterministic one.
- At $\zeta=\infty$ the rescaled particles vanish: both the counting process and its martingale part converge to zero.
- The empirical measures of the particles converge to the conditional law of a single particle given the aggregate, so the three regimes describe the full distributional behavior of a large network.
Reading between the lines
- A testable consequence is that $n(1-\|\phi^n\|_{L^1})^2$ acts as an order parameter: for a fixed kernel family, simulations should cross from synchronized to conditionally independent to extinct behavior as this single number is swept.
- Because the univariate characterization runs through the Bernstein-function representation, the possible limiting kernels are exactly those with Laplace transform $1/\Phi$ for a Bernstein function $\Phi$, so path regularity could be read off from the corresponding L\'evy triplet rather than from the original kernels.
- The auxiliary-coupling method suggests a broader principle: in near-critical mean-field systems, conditional independence of a finite block from the aggregate is structural and may persist for nonlinear intensities satisfying an analogous resolvent condition.
- The extinction regime warns that the chosen rescaling can hide all idiosyncratic noise; observing residual fluctuations would require a different normalization.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies scaling limits of nearly unstable Hawkes processes. Under Condition 2.1, which postulates the existence of a vanishing sequence β_n such that the rescaled resolvent measures β_n ψ^n_{ij}(t)dt converge vaguely to a σ-finite limit F_{ij} with a uniform L2 bound, Theorem 2.2 shows that the rescaled compensator, counting process, and martingale are C-tight and every limit point satisfies the affine stochastic Volterra equation X = F([0,t])·a + f*Z with Z_i = B_i∘X_i. The paper then considers a mean-field system of n Hawkes particles with a common intensity and, under the same Condition 2.1, establishes propagation of chaos (Theorems 3.1 and 3.2) with three regimes governed by ζ = lim n β_n^2: complete synchronization for ζ=0, conditional independence with Poisson-type idiosyncratic components for ζ∈(0,∞), and extinction for ζ=∞. The univariate case is further characterized in terms of Bernstein functions.
Significance. The main convergence argument is rigorous and economical: the proof of Theorem 2.2 combines moment estimates, Kolmogorov–Chentsov tightness, and the martingale representation theorem, and it does not require the restrictive kernel form φ^n = a_n φ used in earlier work. The mean-field trichotomy is a natural and nontrivial extension of Delattre–Fournier–Hoffmann, and the empirical measure convergence in Theorem 3.2 is a useful strengthening. The paper also gives a clean Bernstein-function parametrization of the possible limits in the univariate case. However, the advertised regime parameter in the abstract is not the one appearing in the theorems, and the abstract's 'mild' condition is weaker than Condition 2.1; these mismatches affect how the results can be applied and need to be fixed.
major comments (3)
- [Abstract and Theorem 3.1] The abstract states that the three mean-field regimes depend on the asymptotics of n(1-||φ^n||_{L1})^2, but Theorem 3.1 and Remark 3.3 state that the trichotomy is governed by ζ = lim n β_n^2, where β_n is the scaling sequence appearing in Condition 2.1. Condition 2.1 does not imply that β_n is comparable to 1-||φ^n||_{L1}. In fact, by Lemma 2.4(3), 1-||φ^n||_{L1} = β_n/LF(0) + o(β_n); when LF(0)=∞ this quantity is o(β_n). A concrete such family is φ_n(t) = a_n b_n φ(b_n t), where φ is the one-sided α-stable density with Laplace transform e^{-c z^α}, α∈(1/2,1), a_n=1-n^{-2/3}, b_n=n^{1/(3α)}, and β_n=n^{-1/3}. For this family Condition 2.1 holds (the rescaled resolvent converges to the density with Laplace transform 1/(c z^α) and lies in L^2 on [0,T]), and with μ_0^n = a n^{1/3} we have β_n μ_0^n → a. But nβ_n^2 = n^{1/3} → ∞, so Theorem 3.1 gives extinction, while n(1-||φ^n||_{L1})^2 = n^{-1/3} → 0, which the abstract labels as the synchronized regime. Thus the abstract's phase diagram is not a consequence of the theorems and is in fact false for this family.
- [Abstract and Section 2.1] Condition 2.1 requires more than ||φ^n||_{L1} → 1: it requires the existence of β_n → 0 with sup_n β_n ||ψ^n_{ij}||_{L^2_T} < ∞ and vague convergence of β_n ψ^n_{ij}(t)dt. The abstract's phrase 'mild asymptotic criticality condition, specifically ||φ^n||_{L1} → 1' is therefore misleading. The paper itself notes in Section 2.2 that a necessary and sufficient condition for the L^2 bound is unknown and only provides a sufficient condition in dimension one (boundedness of ||φ^n||_{L^2_T}). Since Theorem 2.2 is the basis for Theorems 3.1 and 3.2, the abstract overstates the domain of applicability.
- [Remark 3.3] Remark 3.3 contains an incorrect assertion in the ζ=∞ case. The remark defines U_i = U for all i, where U is the aggregate compensator limit from Theorem 2.2, and then concludes 'Xi = Zi = Ui = 0'. But Theorem 2.2, applied to the aggregate Hawkes process, gives U = X with X(t) = F([0,t])a + f*Z(t); when β_n μ_0^n → a>0 and F is non-degenerate, X(t) ≥ F([0,t])a > 0 for t>0. The individual compensator Λ^{(n)}_i equals Λ^{(n)} (the aggregate rescaled compensator), so its limit is U = X, not 0. The theorem's actual conclusion in case (3) concerns only X_i and Z_i, which can vanish even though U is non-zero, because the martingale term in (3.13) is not tight after multiplication by the diverging factor √(nβ_n^2). The remark's heuristic should be corrected to avoid claiming Ui=0.
minor comments (5)
- [Section 2.2, Proposition 2.5] The statement and proof of Proposition 2.5 write ρ(||g||_{L^2}) in the conclusion, but the argument uses ||g||_{L^1}; the L^2 norm is not defined for general g∈L^1. Please replace L^2 by L^1 in the statement and in the last displayed inequality.
- [Theorem 3.1] The third case of Theorem 3.1 is labeled '(2)' instead of '(3)'; the numbering should be corrected.
- [Section 5.2, proof of Proposition 5.3] The proof uses the process '~λ^n(t)' without defining it; this should be defined as the intensity of the auxiliary process for particles K+1,...,n, or the proof should be rewritten using the previously defined θ^n.
- [Section 2.1, Eq. (2.3)] The notation ||F||_{L^1_t} is nonstandard for the cumulative measure F([0,t]) of the σ-finite measure F; consider writing F([0,t]) or F(t) throughout the paper to avoid confusion with a function norm.
- [Section 1.2] The phrase 'abstracted-valued random variables' should be 'abstract-valued random variables'.
Circularity Check
No circularity: the limiting equations are derived from the assumed resolvent scaling via a genuine passage to the limit, not from the target result.
full rationale
The paper's central limit object, the measure F, is not an independently fitted quantity: it is introduced in Condition 2.1 as the vague limit of beta_n psi^n. Theorem 2.2 then passes to the limit in the martingale representation (A.8)/(2.1), so equation (2.3) follows by a bona fide limit identification rather than by restating the assumption. The same holds for the mean-field trichotomy: the regimes are governed by zeta = lim n beta_n^2 and are proved through tightness, the auxiliary-process coupling in Section 5, and standard martingale/Poisson representation theorems; no fitted parameter is renamed as a prediction. Self-citations, notably [34], [33], and [50], are technical or contextual and are not load-bearing: for instance, Lemma 4.1 borrows a Burkholder-Davis-Gundy style estimate from [34] but the relevant uniform bounds are obtained under Condition 2.1 in the present proof. The abstract's statement that the regimes depend on n(1-||phi^n||_L1)^2 is not the parameter used in Theorem 3.1, which uses zeta = lim n beta_n^2; this is a presentation/accuracy concern, not a circular reduction. No step in the derivation chain reduces to its own conclusion or to an unverified self-citation chain.
Assumptions & free parameters
assumptions (6)
- ad hoc to paper Condition 2.1: there exists beta_n -> 0 with sup_n beta_n ||psi^n_ij||_{L^2_T} < infinity and beta_n psi^n_ij(t)dt -> F_ij(dt) vaguely for all i,j,T.
- domain assumption Stability condition rho(||phi^n||_{L1}) < 1 for all n.
- domain assumption beta_n mu^n -> a in R^d_+ (Theorem 2.2), and beta_n mu_0^n -> a > 0 (Theorem 3.1).
- domain assumption Mean-field scaling mu_i^n = mu_0^n/n and phi_ij^n = phi^n/n in the particle system (3.1).
- domain assumption The limit zeta = lim n beta_n^2 exists in [0, infinity] along the subsequence under consideration.
- standard math Standard weak-convergence tools: Skorokhod representation, martingale representation, Kolmogorov-Chentsov, stochastic Fubini, Meyer-Zheng tightness.
Cite this review
Pith. "Pith review of Mean-Field Limits for Nearly Unstable Hawkes Processes." pith.science (2026). https://pith.science/paper/W3LSRYCX
@misc{pith2026250111648,
author = {Pith},
title = {Pith review of: Mean-Field Limits for Nearly Unstable Hawkes Processes},
year = {2026},
howpublished = {\url{https://pith.science/paper/W3LSRYCX}},
note = {Machine review of arXiv:2501.11648}
}
abstract
In this paper, we establish general scaling limits for nearly unstable Hawkes processes in a mean-field regime by extending the method introduced by Jaisson and Rosenbaum. Under a mild asymptotic criticality condition on the self-exciting kernels $\{\phi^n\}$, specifically $\|\phi^n\|_{L^1} \to 1$, we first show that the scaling limits of these Hawkes processes are necessarily stochastic Volterra diffusions of affine type. Moreover, we establish a propagation of chaos result for Hawkes systems with mean-field interactions, highlighting three distinct regimes for the limiting processes, which depend on the asymptotics of $n(1-\|\phi^n\|_{L^1})^2$. These results provide a significant generalization of the findings by Delattre, Fournier and Hoffmann.
Reference graph
Works this paper leans on
- [1]
-
[2]
E. Abi Jaber, M. Larsson, and S. Pulido. Affine volterra pro cesses. Ann. Appl. Probab. , 29(5):3155–3200, 2019
work page 2019
- [3]
- [4]
- [5]
-
[6]
Second order statistics characterization of Hawkes processes and non-parametric estimation
E. Bacry and J. F. Muzy. Second order statistics characte rization of Hawkes processes and non- parametric estimation. arXiv preprint arXiv:1401.0903 , 2014
work page Pith review arXiv 2014
-
[7]
J. Bertoin. L´ evy Processes. Cambridge University Press, 1996
work page 1996
-
[8]
P. Billingsley. Convergence of Probability Measures . Wiley, New York, NY, 2 edition, 1999
work page 1999
Show all 53 references
-
[9]
Billingsley
P. Billingsley. Probability and Measure . John Wiley & Sons, 2008
2008
-
[10]
Blanc, J
P. Blanc, J. Donier, and J. P. Bouchaud. Quadratic Hawke s processes for financial prices. Quan- titative Finance , 17(2):171–188, 2017
2017
-
[11]
Br´ emaud and Massouli´ e L
P. Br´ emaud and Massouli´ e L. Hawkes branching point pr ocesses without ancestors. J. Appl. Probab., 38(1):122–135, 2001
2001
-
[12]
Br´ emaud and L
P. Br´ emaud and L. Massouli´ e. Stability of nonlinear Hawkes processes. Ann. Probab., 24(3):1563– 1588, 1996
1996
-
[13]
Chavez-Demoulin, A
V. Chavez-Demoulin, A. C. Davison, and A. J. McNeil. Est imating value-at-risk: a point process approach. Quant. Finance, 5(2):227–234, 2005
2005
-
[14]
Chevallier
J. Chevallier. Mean-field limit of generalized Hawkes p rocesses. Stoch. Process. Appl. , 127(12):3870–3912, 2017
2017
-
[15]
Chevallier, A
J. Chevallier, A. Duarte, E. L¨ ocherbach, and G. Ost. Me an field limits for nonlinear spatially extended Hawkes processes with exponential memory kernels . Stoch. Process. Appl., 129(1):1–27, 2019
2019
-
[16]
Costa, C
M. Costa, C. Graham, L. Marsalle, and V. C. Tran. Renewal in Hawkes processes with self- excitation and inhibition. Adv. Appl. Probab. , 52(3):879–915, 2020
2020
-
[17]
Dandapani, P
A. Dandapani, P. Jusselin, and M. Rosenbaum. From quadr atic Hawkes processes to super-Heston rough volatility models with Zumbach effect. Quantitative Finance , pages 1–13, 2021
2021
-
[18]
Delattre and N
S. Delattre and N. Fournier. Statistical inference ver sus mean field limit for Hawkes processes. Electron. J. Statist. , 10(1):1223–1295, 2016
2016
-
[19]
Delattre, N
S. Delattre, N. Fournier, and M. Hoffmann. Hawkes process es on large networks. Ann. Appl. Probab., 26(1):216–261, 2016
2016
-
[20]
Durin, M
B. Durin, M. Rosenbaum, and G. Szymanski. The two square root laws of market impact and the role of sophisticated market participants. arXiv preprint arXiv:2311.18283 , 2023. 26
2023 arXiv
-
[21]
El Euch, M
O. El Euch, M. Fukasawa, and M. Rosenbaum. The microstru ctural foundations of leverage effect and rough volatility. Finance Stoch., 22(2):241–280, 2018
2018
-
[22]
El Euch and M
O. El Euch and M. Rosenbaum. The characteristic functio n of rough Heston models. Math. Finance, 29(1):3–38, 2019
2019
-
[23]
X. Erny, E. L¨ ocherbach, and D. Loukianova. Conditiona l propagation of chaos for mean field systems of interacting neurons. Electron. J. Probab. , 26:1–25, 2021
2021
-
[24]
F. Gao, Y. Gao, and L. Zhu. Fluctuations and moderate dev iations for the mean fields of Hawkes processes. arXiv: 2307.15903 , 2023
2023 arXiv
-
[25]
Gao and L
F. Gao and L. Zhu. Large deviations for the mean-field lim it of Hawkes processes. arXiv:2301.07585, 2023
2023 arXiv
-
[26]
Gatheral, T
J. Gatheral, T. Jaisson, and M. Rosenbaum. Volatility i s rough. Quant. Financ. , 18:933–949, 2018
2018
-
[27]
Giesecke, L
K. Giesecke, L. Goldberg, and X. Ding. A top-down approa ch to multi-name credit. J. Credit Risk, 2009
2009
-
[28]
Gr¨ un and S
S. Gr¨ un and S. Rotter. Analysis of Parallel Spike Trains , volume 7. Springer, 2010
2010
-
[29]
A. G. Hawkes. Point spectra of some mutually exciting po int processes. J. Roy. Statist. Soc. Ser. B, 33(3):438–443, 1971
1971
-
[30]
A. G. Hawkes. Spectra of some self-exciting and mutuall y exciting point processes. Biometrika, 58(1):83–90, 1971
1971
-
[31]
A. G. Hawkes and D. Oakes. A cluster process representat ion of a self-exciting process. J. Appl. Probab., 11(3):493–503, 1974
1974
-
[32]
P. Hewlett. Clustering of order arrivals, price impact and trade path optimisation. In Workshop on Financial Modeling with Jump processes , pages 6–8, 2006
2006
-
[33]
Horst and W
U. Horst and W. Xu. Functional limit theorems for Hawkes processes. To appear in Probab. Theory Related Fields , 2025
2025
-
[34]
Horst, W
U. Horst, W. Xu, and R. Zhang. Convergence of heavy-tail ed Hawkes processes and the mi- crostructure of rough volatility. arXiv: 2312.08784 , 2023
2023
-
[35]
Ikeda and S
N. Ikeda and S. Watanabe. Stochastic Differential Equations and Diffusion Processes . North- Holland/Kodansha, Amsterdam/Tokyo, 1989
1989
-
[36]
Jacod and A
J. Jacod and A. N. Shiryaev. Limit Theorems for Stochastic Processes . Springer, Berlin, 2003
2003
-
[37]
Jaisson and M
T. Jaisson and M. Rosenbaum. Limit theorems for nearly u nstable Hawkes processes. Ann. Appl. Probab., 25(2):600–631, 2015
2015
-
[38]
Jaisson and M
T. Jaisson and M. Rosenbaum. Rough fractional diffusions as scaling limits of nearly unstable heavy tailed Hawkes processes. Ann. Appl. Probab. , 26(5):2860–2882, 2016
2016
-
[39]
Jusselin and M
P. Jusselin and M. Rosenbaum. No-arbitrage implies pow er-law market impact and rough volatil- ity. Math. Finance, 30(4):1309–1336, 2020. 27
2020
-
[40]
Kallenberg
O. Kallenberg. Random Measures, Theory and Applications . Springer, 2017
2017
-
[41]
Lewis, G
E. Lewis, G. Mohler, P. Brantingham, and A. Bertozzi. Se lf-exciting point process models of insurgency in Iraq. Security Journal , 25:1–21, 2011
2011
-
[42]
Meyer and W.A
P.A. Meyer and W.A. Zheng. Tightness criteria for laws o f semimartingales. Ann. Inst. Henri Poincar´ e Probab. Statist., 20(4):353–372, 1984
1984
-
[43]
G. O. Mohler, M. B. Short, P. J. Brantingham, F. P. Schoen berg, and G. E. Tita. Self-exciting point process modelling of crime. J. Amer. Statist. Assoc. , 106:100–108, 2011
2011
-
[44]
Okatan, M
M. Okatan, M. A. Wilson, and E. N. Brown. Analyzing funct ional connectivity using a network likelihood model of ensemble neural spiking activity. Neural Computation, 17(9):1927–1961, 2005
1927
-
[45]
Pakdaman, B
K. Pakdaman, B. Perthame, and D. Salort. Dynamics of a st ructured neuron population. Non- linearity, 23(1):55, 2009
2009
-
[46]
Pfaffelhuber, S
P. Pfaffelhuber, S. Rotter, and J. Stiefel. Mean-field lim its for non-linear Hawkes processes with excitation and inhibition. Stoch. Process. Appl. , 153:57–78, 2022
2022
-
[47]
Reynaud-Bouret and S
P. Reynaud-Bouret and S. Schbath. Adaptive estimation for Hawkes processes; application to genome analysis. Ann. Statist. , 38(5):2781 – 2822, 2010
2010
-
[48]
Rosenbaum and M
M. Rosenbaum and M. Tomas. From microscopic price dynam ics to multidimensional rough volatility models. Adv. Appl. Probab. , 53(2):425–462, 2021
2021
-
[49]
R. L. Schilling, R. Song, and Z. Vondracek. Bernstein Functions: theory and applications . De Gruyter, 2012
2012
-
[50]
W. Xu. Diffusion approximations for marked self-excited systems with applications to general branching processes. Ann. Appl. Probab. , 34(3):2733–2798, 2024
2024
-
[51]
K. Zhou, H. Zha, and L. Song. Learning triggering kernel s for multi-dimensional Hawkes processes. In Proceedings of the 30th International Conference on Internati onal Conference on Machine Learning - Volume 28 , 2013
2013
-
[52]
L. Zhu. Central limit theorem for nonlinear Hawkes proc esses. J. Appl. Probab. , 50(3):760–771, 2013. A Summary about Hawkes processes In this section, we summarize some results about multidimen sional Hawkes processes. Definition A.1 A d-dimensional Hawkes process with baseli...
2013
-
[53]
Then we can build on an en- larged probability space(~Ω, ~F, ~Ft, ~P) a family of independent Poisson measures {Πi(ds,dz )}1≤i≤d on R2 + with intensity dsdz such that ( A.2) holds
Suppose that N is a Hawkes process in the sense of Definition A.1. Then we can build on an en- larged probability space(~Ω, ~F, ~Ft, ~P) a family of independent Poisson measures {Πi(ds,dz )}1≤i≤d on R2 + with intensity dsdz such that ( A.2) holds. For convenience, we write ( 1....
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.