Scaling Limits of Bivariate Nearly-Unstable Hawkes Processes and Applications to Rough Volatility

Sohaib El Karmi · 2026

Read Scaling Limits of Bivariate Nearly-Unstable Hawkes Processes and Applications to Rough Volatility by Sohaib El Karmi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03703.txt.

Introduction

Sohaib El Karmi studies scaling limits of a pair of nearly unstable Hawkes processes, self-exciting point processes used to model clustered events, coupled in one direction only. The first process evolves on its own and excites the second, but not the reverse. Each has a heavy-tailed memory kernel, and the two tails may differ, giving different degrees of roughness.

As the system nears criticality, El Karmi proves the rescaled intensity converges to the unique solution of coupled stochastic Volterra equations of rough-volatility type. The second component inherits noise from the first through an effective cross-kernel built by convolving the two Mittag-Leffler kernels. A short-time cross-decorrelation law follows: functional correlation between the components vanishes at small scales at an explicit polynomial rate.

Catalog record

CREATORSohaib El Karmi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE113 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03703

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Scaling Limits of Bivariate Nearly-Unstable Hawkes Processes and Applications to Rough Volatility Sohaib El Karmi Abstract . We prove a functional limit theorem for a pair of nearly-unstable Hawkes processes interacting through a triangular cross-excitation mechanism with kernels with distinct heavy tails. To our knowledge this is the first functional limit theorem of this type in the heterogeneous case, where the two components have different degrees of roughness. As the system approaches criticality, the renormalized intensity processes converge weakly to the unique solution of a coupled stochastic Volterra system driven by two independent Brownian motions. The first component evolves autonomously as a rough fractional diffusion with Hurst index H 1 = α 1 − 1 / 2 H_{1}=\alpha_{1}-1/2 . The second component is driven both by its own noise and, through a convolution cross-kernel, by the noise of the first: it is this cross-kernel, the convolution of the two individual Mittag-Leffler kernels, that encodes both Hurst parameters simultaneously and distinguishes the bivariate limit from any product of independent univariate limits or any classical bivariate Brownian model. A key consequence is that the two components are decorrelated at short time scales : their functional correlation ϱ ⁡ ( t ) \varrho(t) vanishes as C ϱ ​ t α 1 C_{\varrho}\,t^{\alpha_{1}} as t ↓ 0 t\downarrow 0 , with the exponent governed exclusively by the rougher of the two components and the constant C ϱ C_{\varrho} given in closed form. We also show that hypothesis (A5) (scale matching) is not a structural restriction: dropping it replaces K 1 ∗ K 2 K_{1}*K_{2} by a time-rescaled version K 1 ∗ K 2 ( ρ 12 ) K_{1}*K_{2}^{(\rho_{12})} with an explicit correction factor. The proof is self-contained and combines four main ingredients: convergence of the renormalized kernels in the space of square-integrable functions, tightness of the joint intensity process in the space C ⁡ ( [ 0 , 1 ] , ℝ 2 ) C([0,1];\mathbb{R}^{2}) , identification of every limit point as a martingale solution via Rebolledo’s theorem, and uniqueness from the affine Volterra theory of Abi Jaber, Larsson, and Pulido, which we extend to a new theorem (Theorem 7.3 ) covering equations with random adapted drifts. We also identify precisely the three obstructions that prevent extending the proof to the fully symmetric case ψ 21 ≠ 0 \psi^{21}\neq 0 . 2020 Mathematics Subject Classification 60F17 (primary); 60G55, 91G99, 60H20 (secondary) 1. Introduction 1.1. Background Hawkes processes [ 9 ] are self-exciting point processes whose intensity is driven by the history of past events through a memory kernel. When this kernel has heavy tails and the process operates near criticality (branching ratio close to one), the macroscopic fluctuations of the intensity exhibit long memory and fractional scaling. The univariate limit theory is well developed. Jaisson and Rosenbaum [ 16 ] proved that a nearly-unstable Hawkes process with a power-law kernel of exponent \alpha\in(1/2,1) converges, after appropriate renormalization, to a rough fractional diffusion with Hurst parameter H=\alpha-1/2\in(0,1/2) . This gives a microscopic derivation of rough volatility models, which are empirically supported by high-frequency equity data [ 8 ] . Subsequent work by Horst, Xu, and Zhang [ 15 ] and Horst and Xu [ 10 ] extended these results to more general kernels and established C -tightness of the intensity process in C([0,1]) . The present paper addresses what happens when two nearly-unstable Hawkes components cross-excite each other. 1.2. The bivariate problem and its difficulty Bivariate Hawkes processes arise naturally in financial microstructure, for instance as models of the joint order flow of two related assets or the co-excitation structure of a bid-ask spread [ 5 , 6 , 4 ] . In such applications the two components typically exhibit different degrees of roughness, reflecting different autocorrelation structures in the underlying flows. Empirical calibrations on high-frequency equity data indicate that the nearly-unstable subcritical regime is essential for reproducing the clustering and autocorrelation patterns observed in order flow [ 5 , 8 ] ; the present paper provides a rigorous functional limit theorem underpinning this observation in the bivariate heavy-tailed case. The essential difficulty in the heterogeneous case is the following. When the two self-exciting kernels have distinct heavy-tail exponents 1/2<\alpha_{1}<\alpha_{2}<1 , the two components must be renormalized on different time scales simultaneously: the first by T^{\alpha_{1}-1} , the second by T^{\alpha_{2}-1} . Because these scales are different, the cross-excitation between the components does not vanish in the limit nor reduce to a simple Brownian correlation. Instead it contributes a genuinely functional dependence: the limiting correlation structure is carried by a new kernel, the convolution L_{12}=b_{\infty}^{12}\|\psi^{12}\|_{1}\cdot(K_{1}*K_{2}) , which encodes both exponents simultaneously. No existing limit theorem for bivariate Hawkes processes addresses this simultaneous renormalization problem. Multivariate Hawkes processes have been studied in the finite-dimensional subcritical regime [ 2 , 9 ] , and Bacry–Muzy [ 5 ] have calibrated bivariate models to financial data, but neither paper addresses the near-critical scaling limit with heterogeneous tail indices. 1.3. Main novelty To our knowledge, this is the first functional limit theorem for a bivariate nearly-unstable Hawkes process with heterogeneous heavy-tail indices 1/2<\alpha_{1}<\alpha_{2}<1 . The cross-kernel. The cross-excitation survives renormalization and is captured in the limit by the effective kernel L_{12}=b_{\infty}^{12}\|\psi^{12}\|_{1}\cdot(K_{1}*K_{2}) , where K_{i} are the individual Mittag-Leffler kernels of the two components (Proposition 3.5 ). The convolution K_{1}*K_{2} has Laplace exponent \alpha_{1}+\alpha_{2} , strictly larger than either individual exponent, so the cross-kernel is smoother than either self-kernel. The correlation between the two limit processes is therefore not a constant Brownian correlation but a time-dependent functional dependence whose strength decreases with the lag. Short-time decorrelation. A direct consequence is that the two components are decorrelated at short time scales (i.e. as the time lag t\downarrow 0 , not in the sense of trading frequency). Their functional correlation satisfies \varrho(t)\sim C_{\varrho}\,t^{\alpha_{1}} as t\downarrow 0 , where \alpha_{1} is the exponent of the rougher component and C_{\varrho} is an explicit constant (Corollary 3.8 ). The rougher component governs the rate at which correlation is lost at fine scales. This power-law decay of correlation at short lags is consistent with the empirical Epps effect [ 3 ] , which documents decreasing cross-correlations between asset returns as the sampling frequency increases. 1.4. Main contributions The core of the paper (Sections 3 – 7 ) establishes the following results for the triangular case ( \psi^{12}\neq 0 , \psi^{21}=0 ), in which component 1 excites component 2 but not conversely. (1) Functional limit theorem (Theorem 7.8 ). The bivariate renormalized intensity process (V^{T,1},V^{T,2}) converges weakly in C([0,1];\mathbb{R}^{2}) to the unique weak solution of a coupled stochastic Volterra system in which V^{1} is autonomous and V^{2} is driven by both its own noise and the noise of V^{1} through the cross-kernel L_{12} . The two components have distinct Hurst parameters H_{i}=\alpha_{i}-1/2 . (2) Cross-kernel identification and covariance formula (Proposition 3.5 ). The effective cross-kernel is L_{12}=b_{\infty}^{12}\|\psi^{12}\|_{1}\cdot(K_{1}*K_{2}) , and the covariance between the two limit processes is given by an explicit integral formula involving K_{1} and K_{1}*K_{2} . (3) Short-time decorrelation law (Corollary 3.8 ). The functional correlation satisfies \varrho(t)\sim C_{\varrho}\,t^{\alpha_{1}} as t\downarrow 0 , with an explicit constant C_{\varrho} . (4) Bivariate C -tightness (Proposition 5.5 ). The cross-term in V^{T,2} has strictly better Hölder regularity

Read the complete text